Thu, 28 Mar 2013 10:27:28 +0100
7014552: gc/lock/jni/jnilockXXX works too slow on 1-processor machine
Summary: Keep a counter of how many times we were stalled by the GC locker, add a diagnostic flag which sets the limit.
Reviewed-by: brutisso, ehelin, johnc
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 // null
467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p));
468 return false;
469 } else {
470 return !hr->isHumongous();
471 }
472 }
474 void G1CollectedHeap::check_ct_logs_at_safepoint() {
475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
476 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
478 // Count the dirty cards at the start.
479 CountNonCleanMemRegionClosure count1(this);
480 ct_bs->mod_card_iterate(&count1);
481 int orig_count = count1.n();
483 // First clear the logged cards.
484 ClearLoggedCardTableEntryClosure clear;
485 dcqs.set_closure(&clear);
486 dcqs.apply_closure_to_all_completed_buffers();
487 dcqs.iterate_closure_all_threads(false);
488 clear.print_histo();
490 // Now ensure that there's no dirty cards.
491 CountNonCleanMemRegionClosure count2(this);
492 ct_bs->mod_card_iterate(&count2);
493 if (count2.n() != 0) {
494 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
495 count2.n(), orig_count);
496 }
497 guarantee(count2.n() == 0, "Card table should be clean.");
499 RedirtyLoggedCardTableEntryClosure redirty;
500 JavaThread::dirty_card_queue_set().set_closure(&redirty);
501 dcqs.apply_closure_to_all_completed_buffers();
502 dcqs.iterate_closure_all_threads(false);
503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
504 clear.calls(), orig_count);
505 guarantee(redirty.calls() == clear.calls(),
506 "Or else mechanism is broken.");
508 CountNonCleanMemRegionClosure count3(this);
509 ct_bs->mod_card_iterate(&count3);
510 if (count3.n() != orig_count) {
511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
512 orig_count, count3.n());
513 guarantee(count3.n() >= orig_count, "Should have restored them all.");
514 }
516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
517 }
519 // Private class members.
521 G1CollectedHeap* G1CollectedHeap::_g1h;
523 // Private methods.
525 HeapRegion*
526 G1CollectedHeap::new_region_try_secondary_free_list() {
527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
528 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
529 if (!_secondary_free_list.is_empty()) {
530 if (G1ConcRegionFreeingVerbose) {
531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
532 "secondary_free_list has %u entries",
533 _secondary_free_list.length());
534 }
535 // It looks as if there are free regions available on the
536 // secondary_free_list. Let's move them to the free_list and try
537 // again to allocate from it.
538 append_secondary_free_list();
540 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
541 "empty we should have moved at least one entry to the free_list");
542 HeapRegion* res = _free_list.remove_head();
543 if (G1ConcRegionFreeingVerbose) {
544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
545 "allocated "HR_FORMAT" from secondary_free_list",
546 HR_FORMAT_PARAMS(res));
547 }
548 return res;
549 }
551 // Wait here until we get notifed either when (a) there are no
552 // more free regions coming or (b) some regions have been moved on
553 // the secondary_free_list.
554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
555 }
557 if (G1ConcRegionFreeingVerbose) {
558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
559 "could not allocate from secondary_free_list");
560 }
561 return NULL;
562 }
564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
566 "the only time we use this to allocate a humongous region is "
567 "when we are allocating a single humongous region");
569 HeapRegion* res;
570 if (G1StressConcRegionFreeing) {
571 if (!_secondary_free_list.is_empty()) {
572 if (G1ConcRegionFreeingVerbose) {
573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
574 "forced to look at the secondary_free_list");
575 }
576 res = new_region_try_secondary_free_list();
577 if (res != NULL) {
578 return res;
579 }
580 }
581 }
582 res = _free_list.remove_head_or_null();
583 if (res == NULL) {
584 if (G1ConcRegionFreeingVerbose) {
585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
586 "res == NULL, trying the secondary_free_list");
587 }
588 res = new_region_try_secondary_free_list();
589 }
590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
591 // Currently, only attempts to allocate GC alloc regions set
592 // do_expand to true. So, we should only reach here during a
593 // safepoint. If this assumption changes we might have to
594 // reconsider the use of _expand_heap_after_alloc_failure.
595 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
597 ergo_verbose1(ErgoHeapSizing,
598 "attempt heap expansion",
599 ergo_format_reason("region allocation request failed")
600 ergo_format_byte("allocation request"),
601 word_size * HeapWordSize);
602 if (expand(word_size * HeapWordSize)) {
603 // Given that expand() succeeded in expanding the heap, and we
604 // always expand the heap by an amount aligned to the heap
605 // region size, the free list should in theory not be empty. So
606 // it would probably be OK to use remove_head(). But the extra
607 // check for NULL is unlikely to be a performance issue here (we
608 // just expanded the heap!) so let's just be conservative and
609 // use remove_head_or_null().
610 res = _free_list.remove_head_or_null();
611 } else {
612 _expand_heap_after_alloc_failure = false;
613 }
614 }
615 return res;
616 }
618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions,
619 size_t word_size) {
620 assert(isHumongous(word_size), "word_size should be humongous");
621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
623 uint first = G1_NULL_HRS_INDEX;
624 if (num_regions == 1) {
625 // Only one region to allocate, no need to go through the slower
626 // path. The caller will attempt the expasion if this fails, so
627 // let's not try to expand here too.
628 HeapRegion* hr = new_region(word_size, false /* do_expand */);
629 if (hr != NULL) {
630 first = hr->hrs_index();
631 } else {
632 first = G1_NULL_HRS_INDEX;
633 }
634 } else {
635 // We can't allocate humongous regions while cleanupComplete() is
636 // running, since some of the regions we find to be empty might not
637 // yet be added to the free list and it is not straightforward to
638 // know which list they are on so that we can remove them. Note
639 // that we only need to do this if we need to allocate more than
640 // one region to satisfy the current humongous allocation
641 // request. If we are only allocating one region we use the common
642 // region allocation code (see above).
643 wait_while_free_regions_coming();
644 append_secondary_free_list_if_not_empty_with_lock();
646 if (free_regions() >= num_regions) {
647 first = _hrs.find_contiguous(num_regions);
648 if (first != G1_NULL_HRS_INDEX) {
649 for (uint i = first; i < first + num_regions; ++i) {
650 HeapRegion* hr = region_at(i);
651 assert(hr->is_empty(), "sanity");
652 assert(is_on_master_free_list(hr), "sanity");
653 hr->set_pending_removal(true);
654 }
655 _free_list.remove_all_pending(num_regions);
656 }
657 }
658 }
659 return first;
660 }
662 HeapWord*
663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
664 uint num_regions,
665 size_t word_size) {
666 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
667 assert(isHumongous(word_size), "word_size should be humongous");
668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
670 // Index of last region in the series + 1.
671 uint last = first + num_regions;
673 // We need to initialize the region(s) we just discovered. This is
674 // a bit tricky given that it can happen concurrently with
675 // refinement threads refining cards on these regions and
676 // potentially wanting to refine the BOT as they are scanning
677 // those cards (this can happen shortly after a cleanup; see CR
678 // 6991377). So we have to set up the region(s) carefully and in
679 // a specific order.
681 // The word size sum of all the regions we will allocate.
682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
683 assert(word_size <= word_size_sum, "sanity");
685 // This will be the "starts humongous" region.
686 HeapRegion* first_hr = region_at(first);
687 // The header of the new object will be placed at the bottom of
688 // the first region.
689 HeapWord* new_obj = first_hr->bottom();
690 // This will be the new end of the first region in the series that
691 // should also match the end of the last region in the seriers.
692 HeapWord* new_end = new_obj + word_size_sum;
693 // This will be the new top of the first region that will reflect
694 // this allocation.
695 HeapWord* new_top = new_obj + word_size;
697 // First, we need to zero the header of the space that we will be
698 // allocating. When we update top further down, some refinement
699 // threads might try to scan the region. By zeroing the header we
700 // ensure that any thread that will try to scan the region will
701 // come across the zero klass word and bail out.
702 //
703 // NOTE: It would not have been correct to have used
704 // CollectedHeap::fill_with_object() and make the space look like
705 // an int array. The thread that is doing the allocation will
706 // later update the object header to a potentially different array
707 // type and, for a very short period of time, the klass and length
708 // fields will be inconsistent. This could cause a refinement
709 // thread to calculate the object size incorrectly.
710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
712 // We will set up the first region as "starts humongous". This
713 // will also update the BOT covering all the regions to reflect
714 // that there is a single object that starts at the bottom of the
715 // first region.
716 first_hr->set_startsHumongous(new_top, new_end);
718 // Then, if there are any, we will set up the "continues
719 // humongous" regions.
720 HeapRegion* hr = NULL;
721 for (uint i = first + 1; i < last; ++i) {
722 hr = region_at(i);
723 hr->set_continuesHumongous(first_hr);
724 }
725 // If we have "continues humongous" regions (hr != NULL), then the
726 // end of the last one should match new_end.
727 assert(hr == NULL || hr->end() == new_end, "sanity");
729 // Up to this point no concurrent thread would have been able to
730 // do any scanning on any region in this series. All the top
731 // fields still point to bottom, so the intersection between
732 // [bottom,top] and [card_start,card_end] will be empty. Before we
733 // update the top fields, we'll do a storestore to make sure that
734 // no thread sees the update to top before the zeroing of the
735 // object header and the BOT initialization.
736 OrderAccess::storestore();
738 // Now that the BOT and the object header have been initialized,
739 // we can update top of the "starts humongous" region.
740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
741 "new_top should be in this region");
742 first_hr->set_top(new_top);
743 if (_hr_printer.is_active()) {
744 HeapWord* bottom = first_hr->bottom();
745 HeapWord* end = first_hr->orig_end();
746 if ((first + 1) == last) {
747 // the series has a single humongous region
748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
749 } else {
750 // the series has more than one humongous regions
751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
752 }
753 }
755 // Now, we will update the top fields of the "continues humongous"
756 // regions. The reason we need to do this is that, otherwise,
757 // these regions would look empty and this will confuse parts of
758 // G1. For example, the code that looks for a consecutive number
759 // of empty regions will consider them empty and try to
760 // re-allocate them. We can extend is_empty() to also include
761 // !continuesHumongous(), but it is easier to just update the top
762 // fields here. The way we set top for all regions (i.e., top ==
763 // end for all regions but the last one, top == new_top for the
764 // last one) is actually used when we will free up the humongous
765 // region in free_humongous_region().
766 hr = NULL;
767 for (uint i = first + 1; i < last; ++i) {
768 hr = region_at(i);
769 if ((i + 1) == last) {
770 // last continues humongous region
771 assert(hr->bottom() < new_top && new_top <= hr->end(),
772 "new_top should fall on this region");
773 hr->set_top(new_top);
774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
775 } else {
776 // not last one
777 assert(new_top > hr->end(), "new_top should be above this region");
778 hr->set_top(hr->end());
779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
780 }
781 }
782 // If we have continues humongous regions (hr != NULL), then the
783 // end of the last one should match new_end and its top should
784 // match new_top.
785 assert(hr == NULL ||
786 (hr->end() == new_end && hr->top() == new_top), "sanity");
788 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
789 _summary_bytes_used += first_hr->used();
790 _humongous_set.add(first_hr);
792 return new_obj;
793 }
795 // If could fit into free regions w/o expansion, try.
796 // Otherwise, if can expand, do so.
797 // Otherwise, if using ex regions might help, try with ex given back.
798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
801 verify_region_sets_optional();
803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords);
804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords);
805 uint x_num = expansion_regions();
806 uint fs = _hrs.free_suffix();
807 uint first = humongous_obj_allocate_find_first(num_regions, word_size);
808 if (first == G1_NULL_HRS_INDEX) {
809 // The only thing we can do now is attempt expansion.
810 if (fs + x_num >= num_regions) {
811 // If the number of regions we're trying to allocate for this
812 // object is at most the number of regions in the free suffix,
813 // then the call to humongous_obj_allocate_find_first() above
814 // should have succeeded and we wouldn't be here.
815 //
816 // We should only be trying to expand when the free suffix is
817 // not sufficient for the object _and_ we have some expansion
818 // room available.
819 assert(num_regions > fs, "earlier allocation should have succeeded");
821 ergo_verbose1(ErgoHeapSizing,
822 "attempt heap expansion",
823 ergo_format_reason("humongous allocation request failed")
824 ergo_format_byte("allocation request"),
825 word_size * HeapWordSize);
826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
827 // Even though the heap was expanded, it might not have
828 // reached the desired size. So, we cannot assume that the
829 // allocation will succeed.
830 first = humongous_obj_allocate_find_first(num_regions, word_size);
831 }
832 }
833 }
835 HeapWord* result = NULL;
836 if (first != G1_NULL_HRS_INDEX) {
837 result =
838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
839 assert(result != NULL, "it should always return a valid result");
841 // A successful humongous object allocation changes the used space
842 // information of the old generation so we need to recalculate the
843 // sizes and update the jstat counters here.
844 g1mm()->update_sizes();
845 }
847 verify_region_sets_optional();
849 return result;
850 }
852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
853 assert_heap_not_locked_and_not_at_safepoint();
854 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
856 unsigned int dummy_gc_count_before;
857 int dummy_gclocker_retry_count = 0;
858 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count);
859 }
861 HeapWord*
862 G1CollectedHeap::mem_allocate(size_t word_size,
863 bool* gc_overhead_limit_was_exceeded) {
864 assert_heap_not_locked_and_not_at_safepoint();
866 // Loop until the allocation is satisified, or unsatisfied after GC.
867 for (int try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) {
868 unsigned int gc_count_before;
870 HeapWord* result = NULL;
871 if (!isHumongous(word_size)) {
872 result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count);
873 } else {
874 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count);
875 }
876 if (result != NULL) {
877 return result;
878 }
880 // Create the garbage collection operation...
881 VM_G1CollectForAllocation op(gc_count_before, word_size);
882 // ...and get the VM thread to execute it.
883 VMThread::execute(&op);
885 if (op.prologue_succeeded() && op.pause_succeeded()) {
886 // If the operation was successful we'll return the result even
887 // if it is NULL. If the allocation attempt failed immediately
888 // after a Full GC, it's unlikely we'll be able to allocate now.
889 HeapWord* result = op.result();
890 if (result != NULL && !isHumongous(word_size)) {
891 // Allocations that take place on VM operations do not do any
892 // card dirtying and we have to do it here. We only have to do
893 // this for non-humongous allocations, though.
894 dirty_young_block(result, word_size);
895 }
896 return result;
897 } else {
898 if (gclocker_retry_count > GCLockerRetryAllocationCount) {
899 return NULL;
900 }
901 assert(op.result() == NULL,
902 "the result should be NULL if the VM op did not succeed");
903 }
905 // Give a warning if we seem to be looping forever.
906 if ((QueuedAllocationWarningCount > 0) &&
907 (try_count % QueuedAllocationWarningCount == 0)) {
908 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
909 }
910 }
912 ShouldNotReachHere();
913 return NULL;
914 }
916 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
917 unsigned int *gc_count_before_ret,
918 int* gclocker_retry_count_ret) {
919 // Make sure you read the note in attempt_allocation_humongous().
921 assert_heap_not_locked_and_not_at_safepoint();
922 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
923 "be called for humongous allocation requests");
925 // We should only get here after the first-level allocation attempt
926 // (attempt_allocation()) failed to allocate.
928 // We will loop until a) we manage to successfully perform the
929 // allocation or b) we successfully schedule a collection which
930 // fails to perform the allocation. b) is the only case when we'll
931 // return NULL.
932 HeapWord* result = NULL;
933 for (int try_count = 1; /* we'll return */; try_count += 1) {
934 bool should_try_gc;
935 unsigned int gc_count_before;
937 {
938 MutexLockerEx x(Heap_lock);
940 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
941 false /* bot_updates */);
942 if (result != NULL) {
943 return result;
944 }
946 // If we reach here, attempt_allocation_locked() above failed to
947 // allocate a new region. So the mutator alloc region should be NULL.
948 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
950 if (GC_locker::is_active_and_needs_gc()) {
951 if (g1_policy()->can_expand_young_list()) {
952 // No need for an ergo verbose message here,
953 // can_expand_young_list() does this when it returns true.
954 result = _mutator_alloc_region.attempt_allocation_force(word_size,
955 false /* bot_updates */);
956 if (result != NULL) {
957 return result;
958 }
959 }
960 should_try_gc = false;
961 } else {
962 // The GCLocker may not be active but the GCLocker initiated
963 // GC may not yet have been performed (GCLocker::needs_gc()
964 // returns true). In this case we do not try this GC and
965 // wait until the GCLocker initiated GC is performed, and
966 // then retry the allocation.
967 if (GC_locker::needs_gc()) {
968 should_try_gc = false;
969 } else {
970 // Read the GC count while still holding the Heap_lock.
971 gc_count_before = total_collections();
972 should_try_gc = true;
973 }
974 }
975 }
977 if (should_try_gc) {
978 bool succeeded;
979 result = do_collection_pause(word_size, gc_count_before, &succeeded);
980 if (result != NULL) {
981 assert(succeeded, "only way to get back a non-NULL result");
982 return result;
983 }
985 if (succeeded) {
986 // If we get here we successfully scheduled a collection which
987 // failed to allocate. No point in trying to allocate
988 // further. We'll just return NULL.
989 MutexLockerEx x(Heap_lock);
990 *gc_count_before_ret = total_collections();
991 return NULL;
992 }
993 } else {
994 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
995 MutexLockerEx x(Heap_lock);
996 *gc_count_before_ret = total_collections();
997 return NULL;
998 }
999 // The GCLocker is either active or the GCLocker initiated
1000 // GC has not yet been performed. Stall until it is and
1001 // then retry the allocation.
1002 GC_locker::stall_until_clear();
1003 (*gclocker_retry_count_ret) += 1;
1004 }
1006 // We can reach here if we were unsuccessul in scheduling a
1007 // collection (because another thread beat us to it) or if we were
1008 // stalled due to the GC locker. In either can we should retry the
1009 // allocation attempt in case another thread successfully
1010 // performed a collection and reclaimed enough space. We do the
1011 // first attempt (without holding the Heap_lock) here and the
1012 // follow-on attempt will be at the start of the next loop
1013 // iteration (after taking the Heap_lock).
1014 result = _mutator_alloc_region.attempt_allocation(word_size,
1015 false /* bot_updates */);
1016 if (result != NULL) {
1017 return result;
1018 }
1020 // Give a warning if we seem to be looping forever.
1021 if ((QueuedAllocationWarningCount > 0) &&
1022 (try_count % QueuedAllocationWarningCount == 0)) {
1023 warning("G1CollectedHeap::attempt_allocation_slow() "
1024 "retries %d times", try_count);
1025 }
1026 }
1028 ShouldNotReachHere();
1029 return NULL;
1030 }
1032 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1033 unsigned int * gc_count_before_ret,
1034 int* gclocker_retry_count_ret) {
1035 // The structure of this method has a lot of similarities to
1036 // attempt_allocation_slow(). The reason these two were not merged
1037 // into a single one is that such a method would require several "if
1038 // allocation is not humongous do this, otherwise do that"
1039 // conditional paths which would obscure its flow. In fact, an early
1040 // version of this code did use a unified method which was harder to
1041 // follow and, as a result, it had subtle bugs that were hard to
1042 // track down. So keeping these two methods separate allows each to
1043 // be more readable. It will be good to keep these two in sync as
1044 // much as possible.
1046 assert_heap_not_locked_and_not_at_safepoint();
1047 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1048 "should only be called for humongous allocations");
1050 // Humongous objects can exhaust the heap quickly, so we should check if we
1051 // need to start a marking cycle at each humongous object allocation. We do
1052 // the check before we do the actual allocation. The reason for doing it
1053 // before the allocation is that we avoid having to keep track of the newly
1054 // allocated memory while we do a GC.
1055 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1056 word_size)) {
1057 collect(GCCause::_g1_humongous_allocation);
1058 }
1060 // We will loop until a) we manage to successfully perform the
1061 // allocation or b) we successfully schedule a collection which
1062 // fails to perform the allocation. b) is the only case when we'll
1063 // return NULL.
1064 HeapWord* result = NULL;
1065 for (int try_count = 1; /* we'll return */; try_count += 1) {
1066 bool should_try_gc;
1067 unsigned int gc_count_before;
1069 {
1070 MutexLockerEx x(Heap_lock);
1072 // Given that humongous objects are not allocated in young
1073 // regions, we'll first try to do the allocation without doing a
1074 // collection hoping that there's enough space in the heap.
1075 result = humongous_obj_allocate(word_size);
1076 if (result != NULL) {
1077 return result;
1078 }
1080 if (GC_locker::is_active_and_needs_gc()) {
1081 should_try_gc = false;
1082 } else {
1083 // The GCLocker may not be active but the GCLocker initiated
1084 // GC may not yet have been performed (GCLocker::needs_gc()
1085 // returns true). In this case we do not try this GC and
1086 // wait until the GCLocker initiated GC is performed, and
1087 // then retry the allocation.
1088 if (GC_locker::needs_gc()) {
1089 should_try_gc = false;
1090 } else {
1091 // Read the GC count while still holding the Heap_lock.
1092 gc_count_before = total_collections();
1093 should_try_gc = true;
1094 }
1095 }
1096 }
1098 if (should_try_gc) {
1099 // If we failed to allocate the humongous object, we should try to
1100 // do a collection pause (if we're allowed) in case it reclaims
1101 // enough space for the allocation to succeed after the pause.
1103 bool succeeded;
1104 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1105 if (result != NULL) {
1106 assert(succeeded, "only way to get back a non-NULL result");
1107 return result;
1108 }
1110 if (succeeded) {
1111 // If we get here we successfully scheduled a collection which
1112 // failed to allocate. No point in trying to allocate
1113 // further. We'll just return NULL.
1114 MutexLockerEx x(Heap_lock);
1115 *gc_count_before_ret = total_collections();
1116 return NULL;
1117 }
1118 } else {
1119 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
1120 MutexLockerEx x(Heap_lock);
1121 *gc_count_before_ret = total_collections();
1122 return NULL;
1123 }
1124 // The GCLocker is either active or the GCLocker initiated
1125 // GC has not yet been performed. Stall until it is and
1126 // then retry the allocation.
1127 GC_locker::stall_until_clear();
1128 (*gclocker_retry_count_ret) += 1;
1129 }
1131 // We can reach here if we were unsuccessul in scheduling a
1132 // collection (because another thread beat us to it) or if we were
1133 // stalled due to the GC locker. In either can we should retry the
1134 // allocation attempt in case another thread successfully
1135 // performed a collection and reclaimed enough space. Give a
1136 // warning if we seem to be looping forever.
1138 if ((QueuedAllocationWarningCount > 0) &&
1139 (try_count % QueuedAllocationWarningCount == 0)) {
1140 warning("G1CollectedHeap::attempt_allocation_humongous() "
1141 "retries %d times", try_count);
1142 }
1143 }
1145 ShouldNotReachHere();
1146 return NULL;
1147 }
1149 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1150 bool expect_null_mutator_alloc_region) {
1151 assert_at_safepoint(true /* should_be_vm_thread */);
1152 assert(_mutator_alloc_region.get() == NULL ||
1153 !expect_null_mutator_alloc_region,
1154 "the current alloc region was unexpectedly found to be non-NULL");
1156 if (!isHumongous(word_size)) {
1157 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1158 false /* bot_updates */);
1159 } else {
1160 HeapWord* result = humongous_obj_allocate(word_size);
1161 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1162 g1_policy()->set_initiate_conc_mark_if_possible();
1163 }
1164 return result;
1165 }
1167 ShouldNotReachHere();
1168 }
1170 class PostMCRemSetClearClosure: public HeapRegionClosure {
1171 G1CollectedHeap* _g1h;
1172 ModRefBarrierSet* _mr_bs;
1173 public:
1174 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1175 _g1h(g1h), _mr_bs(mr_bs) { }
1176 bool doHeapRegion(HeapRegion* r) {
1177 if (r->continuesHumongous()) {
1178 return false;
1179 }
1180 _g1h->reset_gc_time_stamps(r);
1181 HeapRegionRemSet* hrrs = r->rem_set();
1182 if (hrrs != NULL) hrrs->clear();
1183 // You might think here that we could clear just the cards
1184 // corresponding to the used region. But no: if we leave a dirty card
1185 // in a region we might allocate into, then it would prevent that card
1186 // from being enqueued, and cause it to be missed.
1187 // Re: the performance cost: we shouldn't be doing full GC anyway!
1188 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1189 return false;
1190 }
1191 };
1193 void G1CollectedHeap::clear_rsets_post_compaction() {
1194 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1195 heap_region_iterate(&rs_clear);
1196 }
1198 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1199 G1CollectedHeap* _g1h;
1200 UpdateRSOopClosure _cl;
1201 int _worker_i;
1202 public:
1203 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1204 _cl(g1->g1_rem_set(), worker_i),
1205 _worker_i(worker_i),
1206 _g1h(g1)
1207 { }
1209 bool doHeapRegion(HeapRegion* r) {
1210 if (!r->continuesHumongous()) {
1211 _cl.set_from(r);
1212 r->oop_iterate(&_cl);
1213 }
1214 return false;
1215 }
1216 };
1218 class ParRebuildRSTask: public AbstractGangTask {
1219 G1CollectedHeap* _g1;
1220 public:
1221 ParRebuildRSTask(G1CollectedHeap* g1)
1222 : AbstractGangTask("ParRebuildRSTask"),
1223 _g1(g1)
1224 { }
1226 void work(uint worker_id) {
1227 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1228 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1229 _g1->workers()->active_workers(),
1230 HeapRegion::RebuildRSClaimValue);
1231 }
1232 };
1234 class PostCompactionPrinterClosure: public HeapRegionClosure {
1235 private:
1236 G1HRPrinter* _hr_printer;
1237 public:
1238 bool doHeapRegion(HeapRegion* hr) {
1239 assert(!hr->is_young(), "not expecting to find young regions");
1240 // We only generate output for non-empty regions.
1241 if (!hr->is_empty()) {
1242 if (!hr->isHumongous()) {
1243 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1244 } else if (hr->startsHumongous()) {
1245 if (hr->region_num() == 1) {
1246 // single humongous region
1247 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1248 } else {
1249 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1250 }
1251 } else {
1252 assert(hr->continuesHumongous(), "only way to get here");
1253 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1254 }
1255 }
1256 return false;
1257 }
1259 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1260 : _hr_printer(hr_printer) { }
1261 };
1263 void G1CollectedHeap::print_hrs_post_compaction() {
1264 PostCompactionPrinterClosure cl(hr_printer());
1265 heap_region_iterate(&cl);
1266 }
1268 double G1CollectedHeap::verify(bool guard, const char* msg) {
1269 double verify_time_ms = 0.0;
1271 if (guard && total_collections() >= VerifyGCStartAt) {
1272 double verify_start = os::elapsedTime();
1273 HandleMark hm; // Discard invalid handles created during verification
1274 gclog_or_tty->print(msg);
1275 prepare_for_verify();
1276 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking);
1277 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1278 }
1280 return verify_time_ms;
1281 }
1283 void G1CollectedHeap::verify_before_gc() {
1284 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1285 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1286 }
1288 void G1CollectedHeap::verify_after_gc() {
1289 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1290 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1291 }
1293 bool G1CollectedHeap::do_collection(bool explicit_gc,
1294 bool clear_all_soft_refs,
1295 size_t word_size) {
1296 assert_at_safepoint(true /* should_be_vm_thread */);
1298 if (GC_locker::check_active_before_gc()) {
1299 return false;
1300 }
1302 SvcGCMarker sgcm(SvcGCMarker::FULL);
1303 ResourceMark rm;
1305 print_heap_before_gc();
1307 size_t metadata_prev_used = MetaspaceAux::used_in_bytes();
1309 HRSPhaseSetter x(HRSPhaseFullGC);
1310 verify_region_sets_optional();
1312 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1313 collector_policy()->should_clear_all_soft_refs();
1315 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1317 {
1318 IsGCActiveMark x;
1320 // Timing
1321 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1322 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1323 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1325 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1326 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1327 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1329 double start = os::elapsedTime();
1330 g1_policy()->record_full_collection_start();
1332 // Note: When we have a more flexible GC logging framework that
1333 // allows us to add optional attributes to a GC log record we
1334 // could consider timing and reporting how long we wait in the
1335 // following two methods.
1336 wait_while_free_regions_coming();
1337 // If we start the compaction before the CM threads finish
1338 // scanning the root regions we might trip them over as we'll
1339 // be moving objects / updating references. So let's wait until
1340 // they are done. By telling them to abort, they should complete
1341 // early.
1342 _cm->root_regions()->abort();
1343 _cm->root_regions()->wait_until_scan_finished();
1344 append_secondary_free_list_if_not_empty_with_lock();
1346 gc_prologue(true);
1347 increment_total_collections(true /* full gc */);
1348 increment_old_marking_cycles_started();
1350 size_t g1h_prev_used = used();
1351 assert(used() == recalculate_used(), "Should be equal");
1353 verify_before_gc();
1355 pre_full_gc_dump();
1357 COMPILER2_PRESENT(DerivedPointerTable::clear());
1359 // Disable discovery and empty the discovered lists
1360 // for the CM ref processor.
1361 ref_processor_cm()->disable_discovery();
1362 ref_processor_cm()->abandon_partial_discovery();
1363 ref_processor_cm()->verify_no_references_recorded();
1365 // Abandon current iterations of concurrent marking and concurrent
1366 // refinement, if any are in progress. We have to do this before
1367 // wait_until_scan_finished() below.
1368 concurrent_mark()->abort();
1370 // Make sure we'll choose a new allocation region afterwards.
1371 release_mutator_alloc_region();
1372 abandon_gc_alloc_regions();
1373 g1_rem_set()->cleanupHRRS();
1375 // We should call this after we retire any currently active alloc
1376 // regions so that all the ALLOC / RETIRE events are generated
1377 // before the start GC event.
1378 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1380 // We may have added regions to the current incremental collection
1381 // set between the last GC or pause and now. We need to clear the
1382 // incremental collection set and then start rebuilding it afresh
1383 // after this full GC.
1384 abandon_collection_set(g1_policy()->inc_cset_head());
1385 g1_policy()->clear_incremental_cset();
1386 g1_policy()->stop_incremental_cset_building();
1388 tear_down_region_sets(false /* free_list_only */);
1389 g1_policy()->set_gcs_are_young(true);
1391 // See the comments in g1CollectedHeap.hpp and
1392 // G1CollectedHeap::ref_processing_init() about
1393 // how reference processing currently works in G1.
1395 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1396 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1398 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1399 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1401 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1402 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1404 // Do collection work
1405 {
1406 HandleMark hm; // Discard invalid handles created during gc
1407 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1408 }
1410 assert(free_regions() == 0, "we should not have added any free regions");
1411 rebuild_region_sets(false /* free_list_only */);
1413 // Enqueue any discovered reference objects that have
1414 // not been removed from the discovered lists.
1415 ref_processor_stw()->enqueue_discovered_references();
1417 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1419 MemoryService::track_memory_usage();
1421 verify_after_gc();
1423 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1424 ref_processor_stw()->verify_no_references_recorded();
1426 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1427 ClassLoaderDataGraph::purge();
1429 // Note: since we've just done a full GC, concurrent
1430 // marking is no longer active. Therefore we need not
1431 // re-enable reference discovery for the CM ref processor.
1432 // That will be done at the start of the next marking cycle.
1433 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1434 ref_processor_cm()->verify_no_references_recorded();
1436 reset_gc_time_stamp();
1437 // Since everything potentially moved, we will clear all remembered
1438 // sets, and clear all cards. Later we will rebuild remebered
1439 // sets. We will also reset the GC time stamps of the regions.
1440 clear_rsets_post_compaction();
1441 check_gc_time_stamps();
1443 // Resize the heap if necessary.
1444 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1446 if (_hr_printer.is_active()) {
1447 // We should do this after we potentially resize the heap so
1448 // that all the COMMIT / UNCOMMIT events are generated before
1449 // the end GC event.
1451 print_hrs_post_compaction();
1452 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1453 }
1455 if (_cg1r->use_cache()) {
1456 _cg1r->clear_and_record_card_counts();
1457 _cg1r->clear_hot_cache();
1458 }
1460 // Rebuild remembered sets of all regions.
1461 if (G1CollectedHeap::use_parallel_gc_threads()) {
1462 uint n_workers =
1463 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1464 workers()->active_workers(),
1465 Threads::number_of_non_daemon_threads());
1466 assert(UseDynamicNumberOfGCThreads ||
1467 n_workers == workers()->total_workers(),
1468 "If not dynamic should be using all the workers");
1469 workers()->set_active_workers(n_workers);
1470 // Set parallel threads in the heap (_n_par_threads) only
1471 // before a parallel phase and always reset it to 0 after
1472 // the phase so that the number of parallel threads does
1473 // no get carried forward to a serial phase where there
1474 // may be code that is "possibly_parallel".
1475 set_par_threads(n_workers);
1477 ParRebuildRSTask rebuild_rs_task(this);
1478 assert(check_heap_region_claim_values(
1479 HeapRegion::InitialClaimValue), "sanity check");
1480 assert(UseDynamicNumberOfGCThreads ||
1481 workers()->active_workers() == workers()->total_workers(),
1482 "Unless dynamic should use total workers");
1483 // Use the most recent number of active workers
1484 assert(workers()->active_workers() > 0,
1485 "Active workers not properly set");
1486 set_par_threads(workers()->active_workers());
1487 workers()->run_task(&rebuild_rs_task);
1488 set_par_threads(0);
1489 assert(check_heap_region_claim_values(
1490 HeapRegion::RebuildRSClaimValue), "sanity check");
1491 reset_heap_region_claim_values();
1492 } else {
1493 RebuildRSOutOfRegionClosure rebuild_rs(this);
1494 heap_region_iterate(&rebuild_rs);
1495 }
1497 if (G1Log::fine()) {
1498 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1499 }
1501 if (true) { // FIXME
1502 MetaspaceGC::compute_new_size();
1503 }
1505 // Start a new incremental collection set for the next pause
1506 assert(g1_policy()->collection_set() == NULL, "must be");
1507 g1_policy()->start_incremental_cset_building();
1509 // Clear the _cset_fast_test bitmap in anticipation of adding
1510 // regions to the incremental collection set for the next
1511 // evacuation pause.
1512 clear_cset_fast_test();
1514 init_mutator_alloc_region();
1516 double end = os::elapsedTime();
1517 g1_policy()->record_full_collection_end();
1519 #ifdef TRACESPINNING
1520 ParallelTaskTerminator::print_termination_counts();
1521 #endif
1523 gc_epilogue(true);
1525 // Discard all rset updates
1526 JavaThread::dirty_card_queue_set().abandon_logs();
1527 assert(!G1DeferredRSUpdate
1528 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1530 _young_list->reset_sampled_info();
1531 // At this point there should be no regions in the
1532 // entire heap tagged as young.
1533 assert( check_young_list_empty(true /* check_heap */),
1534 "young list should be empty at this point");
1536 // Update the number of full collections that have been completed.
1537 increment_old_marking_cycles_completed(false /* concurrent */);
1539 _hrs.verify_optional();
1540 verify_region_sets_optional();
1542 print_heap_after_gc();
1544 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1545 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1546 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1547 // before any GC notifications are raised.
1548 g1mm()->update_sizes();
1549 }
1551 post_full_gc_dump();
1553 return true;
1554 }
1556 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1557 // do_collection() will return whether it succeeded in performing
1558 // the GC. Currently, there is no facility on the
1559 // do_full_collection() API to notify the caller than the collection
1560 // did not succeed (e.g., because it was locked out by the GC
1561 // locker). So, right now, we'll ignore the return value.
1562 bool dummy = do_collection(true, /* explicit_gc */
1563 clear_all_soft_refs,
1564 0 /* word_size */);
1565 }
1567 // This code is mostly copied from TenuredGeneration.
1568 void
1569 G1CollectedHeap::
1570 resize_if_necessary_after_full_collection(size_t word_size) {
1571 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1573 // Include the current allocation, if any, and bytes that will be
1574 // pre-allocated to support collections, as "used".
1575 const size_t used_after_gc = used();
1576 const size_t capacity_after_gc = capacity();
1577 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1579 // This is enforced in arguments.cpp.
1580 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1581 "otherwise the code below doesn't make sense");
1583 // We don't have floating point command-line arguments
1584 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1585 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1586 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1587 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1589 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1590 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1592 // We have to be careful here as these two calculations can overflow
1593 // 32-bit size_t's.
1594 double used_after_gc_d = (double) used_after_gc;
1595 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1596 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1598 // Let's make sure that they are both under the max heap size, which
1599 // by default will make them fit into a size_t.
1600 double desired_capacity_upper_bound = (double) max_heap_size;
1601 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1602 desired_capacity_upper_bound);
1603 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1604 desired_capacity_upper_bound);
1606 // We can now safely turn them into size_t's.
1607 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1608 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1610 // This assert only makes sense here, before we adjust them
1611 // with respect to the min and max heap size.
1612 assert(minimum_desired_capacity <= maximum_desired_capacity,
1613 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1614 "maximum_desired_capacity = "SIZE_FORMAT,
1615 minimum_desired_capacity, maximum_desired_capacity));
1617 // Should not be greater than the heap max size. No need to adjust
1618 // it with respect to the heap min size as it's a lower bound (i.e.,
1619 // we'll try to make the capacity larger than it, not smaller).
1620 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1621 // Should not be less than the heap min size. No need to adjust it
1622 // with respect to the heap max size as it's an upper bound (i.e.,
1623 // we'll try to make the capacity smaller than it, not greater).
1624 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1626 if (capacity_after_gc < minimum_desired_capacity) {
1627 // Don't expand unless it's significant
1628 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1629 ergo_verbose4(ErgoHeapSizing,
1630 "attempt heap expansion",
1631 ergo_format_reason("capacity lower than "
1632 "min desired capacity after Full GC")
1633 ergo_format_byte("capacity")
1634 ergo_format_byte("occupancy")
1635 ergo_format_byte_perc("min desired capacity"),
1636 capacity_after_gc, used_after_gc,
1637 minimum_desired_capacity, (double) MinHeapFreeRatio);
1638 expand(expand_bytes);
1640 // No expansion, now see if we want to shrink
1641 } else if (capacity_after_gc > maximum_desired_capacity) {
1642 // Capacity too large, compute shrinking size
1643 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1644 ergo_verbose4(ErgoHeapSizing,
1645 "attempt heap shrinking",
1646 ergo_format_reason("capacity higher than "
1647 "max desired capacity after Full GC")
1648 ergo_format_byte("capacity")
1649 ergo_format_byte("occupancy")
1650 ergo_format_byte_perc("max desired capacity"),
1651 capacity_after_gc, used_after_gc,
1652 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1653 shrink(shrink_bytes);
1654 }
1655 }
1658 HeapWord*
1659 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1660 bool* succeeded) {
1661 assert_at_safepoint(true /* should_be_vm_thread */);
1663 *succeeded = true;
1664 // Let's attempt the allocation first.
1665 HeapWord* result =
1666 attempt_allocation_at_safepoint(word_size,
1667 false /* expect_null_mutator_alloc_region */);
1668 if (result != NULL) {
1669 assert(*succeeded, "sanity");
1670 return result;
1671 }
1673 // In a G1 heap, we're supposed to keep allocation from failing by
1674 // incremental pauses. Therefore, at least for now, we'll favor
1675 // expansion over collection. (This might change in the future if we can
1676 // do something smarter than full collection to satisfy a failed alloc.)
1677 result = expand_and_allocate(word_size);
1678 if (result != NULL) {
1679 assert(*succeeded, "sanity");
1680 return result;
1681 }
1683 // Expansion didn't work, we'll try to do a Full GC.
1684 bool gc_succeeded = do_collection(false, /* explicit_gc */
1685 false, /* clear_all_soft_refs */
1686 word_size);
1687 if (!gc_succeeded) {
1688 *succeeded = false;
1689 return NULL;
1690 }
1692 // Retry the allocation
1693 result = attempt_allocation_at_safepoint(word_size,
1694 true /* expect_null_mutator_alloc_region */);
1695 if (result != NULL) {
1696 assert(*succeeded, "sanity");
1697 return result;
1698 }
1700 // Then, try a Full GC that will collect all soft references.
1701 gc_succeeded = do_collection(false, /* explicit_gc */
1702 true, /* clear_all_soft_refs */
1703 word_size);
1704 if (!gc_succeeded) {
1705 *succeeded = false;
1706 return NULL;
1707 }
1709 // Retry the allocation once more
1710 result = attempt_allocation_at_safepoint(word_size,
1711 true /* expect_null_mutator_alloc_region */);
1712 if (result != NULL) {
1713 assert(*succeeded, "sanity");
1714 return result;
1715 }
1717 assert(!collector_policy()->should_clear_all_soft_refs(),
1718 "Flag should have been handled and cleared prior to this point");
1720 // What else? We might try synchronous finalization later. If the total
1721 // space available is large enough for the allocation, then a more
1722 // complete compaction phase than we've tried so far might be
1723 // appropriate.
1724 assert(*succeeded, "sanity");
1725 return NULL;
1726 }
1728 // Attempting to expand the heap sufficiently
1729 // to support an allocation of the given "word_size". If
1730 // successful, perform the allocation and return the address of the
1731 // allocated block, or else "NULL".
1733 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1734 assert_at_safepoint(true /* should_be_vm_thread */);
1736 verify_region_sets_optional();
1738 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1739 ergo_verbose1(ErgoHeapSizing,
1740 "attempt heap expansion",
1741 ergo_format_reason("allocation request failed")
1742 ergo_format_byte("allocation request"),
1743 word_size * HeapWordSize);
1744 if (expand(expand_bytes)) {
1745 _hrs.verify_optional();
1746 verify_region_sets_optional();
1747 return attempt_allocation_at_safepoint(word_size,
1748 false /* expect_null_mutator_alloc_region */);
1749 }
1750 return NULL;
1751 }
1753 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1754 HeapWord* new_end) {
1755 assert(old_end != new_end, "don't call this otherwise");
1756 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1758 // Update the committed mem region.
1759 _g1_committed.set_end(new_end);
1760 // Tell the card table about the update.
1761 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1762 // Tell the BOT about the update.
1763 _bot_shared->resize(_g1_committed.word_size());
1764 }
1766 bool G1CollectedHeap::expand(size_t expand_bytes) {
1767 size_t old_mem_size = _g1_storage.committed_size();
1768 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1769 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1770 HeapRegion::GrainBytes);
1771 ergo_verbose2(ErgoHeapSizing,
1772 "expand the heap",
1773 ergo_format_byte("requested expansion amount")
1774 ergo_format_byte("attempted expansion amount"),
1775 expand_bytes, aligned_expand_bytes);
1777 // First commit the memory.
1778 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1779 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1780 if (successful) {
1781 // Then propagate this update to the necessary data structures.
1782 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1783 update_committed_space(old_end, new_end);
1785 FreeRegionList expansion_list("Local Expansion List");
1786 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1787 assert(mr.start() == old_end, "post-condition");
1788 // mr might be a smaller region than what was requested if
1789 // expand_by() was unable to allocate the HeapRegion instances
1790 assert(mr.end() <= new_end, "post-condition");
1792 size_t actual_expand_bytes = mr.byte_size();
1793 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1794 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1795 "post-condition");
1796 if (actual_expand_bytes < aligned_expand_bytes) {
1797 // We could not expand _hrs to the desired size. In this case we
1798 // need to shrink the committed space accordingly.
1799 assert(mr.end() < new_end, "invariant");
1801 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1802 // First uncommit the memory.
1803 _g1_storage.shrink_by(diff_bytes);
1804 // Then propagate this update to the necessary data structures.
1805 update_committed_space(new_end, mr.end());
1806 }
1807 _free_list.add_as_tail(&expansion_list);
1809 if (_hr_printer.is_active()) {
1810 HeapWord* curr = mr.start();
1811 while (curr < mr.end()) {
1812 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1813 _hr_printer.commit(curr, curr_end);
1814 curr = curr_end;
1815 }
1816 assert(curr == mr.end(), "post-condition");
1817 }
1818 g1_policy()->record_new_heap_size(n_regions());
1819 } else {
1820 ergo_verbose0(ErgoHeapSizing,
1821 "did not expand the heap",
1822 ergo_format_reason("heap expansion operation failed"));
1823 // The expansion of the virtual storage space was unsuccessful.
1824 // Let's see if it was because we ran out of swap.
1825 if (G1ExitOnExpansionFailure &&
1826 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1827 // We had head room...
1828 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1829 }
1830 }
1831 return successful;
1832 }
1834 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1835 size_t old_mem_size = _g1_storage.committed_size();
1836 size_t aligned_shrink_bytes =
1837 ReservedSpace::page_align_size_down(shrink_bytes);
1838 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1839 HeapRegion::GrainBytes);
1840 uint num_regions_deleted = 0;
1841 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1842 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1843 assert(mr.end() == old_end, "post-condition");
1845 ergo_verbose3(ErgoHeapSizing,
1846 "shrink the heap",
1847 ergo_format_byte("requested shrinking amount")
1848 ergo_format_byte("aligned shrinking amount")
1849 ergo_format_byte("attempted shrinking amount"),
1850 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1851 if (mr.byte_size() > 0) {
1852 if (_hr_printer.is_active()) {
1853 HeapWord* curr = mr.end();
1854 while (curr > mr.start()) {
1855 HeapWord* curr_end = curr;
1856 curr -= HeapRegion::GrainWords;
1857 _hr_printer.uncommit(curr, curr_end);
1858 }
1859 assert(curr == mr.start(), "post-condition");
1860 }
1862 _g1_storage.shrink_by(mr.byte_size());
1863 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1864 assert(mr.start() == new_end, "post-condition");
1866 _expansion_regions += num_regions_deleted;
1867 update_committed_space(old_end, new_end);
1868 HeapRegionRemSet::shrink_heap(n_regions());
1869 g1_policy()->record_new_heap_size(n_regions());
1870 } else {
1871 ergo_verbose0(ErgoHeapSizing,
1872 "did not shrink the heap",
1873 ergo_format_reason("heap shrinking operation failed"));
1874 }
1875 }
1877 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1878 verify_region_sets_optional();
1880 // We should only reach here at the end of a Full GC which means we
1881 // should not not be holding to any GC alloc regions. The method
1882 // below will make sure of that and do any remaining clean up.
1883 abandon_gc_alloc_regions();
1885 // Instead of tearing down / rebuilding the free lists here, we
1886 // could instead use the remove_all_pending() method on free_list to
1887 // remove only the ones that we need to remove.
1888 tear_down_region_sets(true /* free_list_only */);
1889 shrink_helper(shrink_bytes);
1890 rebuild_region_sets(true /* free_list_only */);
1892 _hrs.verify_optional();
1893 verify_region_sets_optional();
1894 }
1896 // Public methods.
1898 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1899 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1900 #endif // _MSC_VER
1903 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1904 SharedHeap(policy_),
1905 _g1_policy(policy_),
1906 _dirty_card_queue_set(false),
1907 _into_cset_dirty_card_queue_set(false),
1908 _is_alive_closure_cm(this),
1909 _is_alive_closure_stw(this),
1910 _ref_processor_cm(NULL),
1911 _ref_processor_stw(NULL),
1912 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1913 _bot_shared(NULL),
1914 _evac_failure_scan_stack(NULL) ,
1915 _mark_in_progress(false),
1916 _cg1r(NULL), _summary_bytes_used(0),
1917 _g1mm(NULL),
1918 _refine_cte_cl(NULL),
1919 _full_collection(false),
1920 _free_list("Master Free List"),
1921 _secondary_free_list("Secondary Free List"),
1922 _old_set("Old Set"),
1923 _humongous_set("Master Humongous Set"),
1924 _free_regions_coming(false),
1925 _young_list(new YoungList(this)),
1926 _gc_time_stamp(0),
1927 _retained_old_gc_alloc_region(NULL),
1928 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1929 _old_plab_stats(OldPLABSize, PLABWeight),
1930 _expand_heap_after_alloc_failure(true),
1931 _surviving_young_words(NULL),
1932 _old_marking_cycles_started(0),
1933 _old_marking_cycles_completed(0),
1934 _in_cset_fast_test(NULL),
1935 _in_cset_fast_test_base(NULL),
1936 _dirty_cards_region_list(NULL),
1937 _worker_cset_start_region(NULL),
1938 _worker_cset_start_region_time_stamp(NULL) {
1939 _g1h = this; // To catch bugs.
1940 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1941 vm_exit_during_initialization("Failed necessary allocation.");
1942 }
1944 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1946 int n_queues = MAX2((int)ParallelGCThreads, 1);
1947 _task_queues = new RefToScanQueueSet(n_queues);
1949 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1950 assert(n_rem_sets > 0, "Invariant.");
1952 HeapRegionRemSetIterator** iter_arr =
1953 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC);
1954 for (int i = 0; i < n_queues; i++) {
1955 iter_arr[i] = new HeapRegionRemSetIterator();
1956 }
1957 _rem_set_iterator = iter_arr;
1959 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1960 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1962 for (int i = 0; i < n_queues; i++) {
1963 RefToScanQueue* q = new RefToScanQueue();
1964 q->initialize();
1965 _task_queues->register_queue(i, q);
1966 }
1968 clear_cset_start_regions();
1970 // Initialize the G1EvacuationFailureALot counters and flags.
1971 NOT_PRODUCT(reset_evacuation_should_fail();)
1973 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1974 }
1976 jint G1CollectedHeap::initialize() {
1977 CollectedHeap::pre_initialize();
1978 os::enable_vtime();
1980 G1Log::init();
1982 // Necessary to satisfy locking discipline assertions.
1984 MutexLocker x(Heap_lock);
1986 // We have to initialize the printer before committing the heap, as
1987 // it will be used then.
1988 _hr_printer.set_active(G1PrintHeapRegions);
1990 // While there are no constraints in the GC code that HeapWordSize
1991 // be any particular value, there are multiple other areas in the
1992 // system which believe this to be true (e.g. oop->object_size in some
1993 // cases incorrectly returns the size in wordSize units rather than
1994 // HeapWordSize).
1995 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1997 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1998 size_t max_byte_size = collector_policy()->max_heap_byte_size();
2000 // Ensure that the sizes are properly aligned.
2001 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
2002 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2004 _cg1r = new ConcurrentG1Refine();
2006 // Reserve the maximum.
2008 // When compressed oops are enabled, the preferred heap base
2009 // is calculated by subtracting the requested size from the
2010 // 32Gb boundary and using the result as the base address for
2011 // heap reservation. If the requested size is not aligned to
2012 // HeapRegion::GrainBytes (i.e. the alignment that is passed
2013 // into the ReservedHeapSpace constructor) then the actual
2014 // base of the reserved heap may end up differing from the
2015 // address that was requested (i.e. the preferred heap base).
2016 // If this happens then we could end up using a non-optimal
2017 // compressed oops mode.
2019 // Since max_byte_size is aligned to the size of a heap region (checked
2020 // above).
2021 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2023 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2024 HeapRegion::GrainBytes);
2026 // It is important to do this in a way such that concurrent readers can't
2027 // temporarily think somethings in the heap. (I've actually seen this
2028 // happen in asserts: DLD.)
2029 _reserved.set_word_size(0);
2030 _reserved.set_start((HeapWord*)heap_rs.base());
2031 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2033 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2035 // Create the gen rem set (and barrier set) for the entire reserved region.
2036 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2037 set_barrier_set(rem_set()->bs());
2038 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2039 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2040 } else {
2041 vm_exit_during_initialization("G1 requires a mod ref bs.");
2042 return JNI_ENOMEM;
2043 }
2045 // Also create a G1 rem set.
2046 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2047 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2048 } else {
2049 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2050 return JNI_ENOMEM;
2051 }
2053 // Carve out the G1 part of the heap.
2055 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2056 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2057 g1_rs.size()/HeapWordSize);
2059 _g1_storage.initialize(g1_rs, 0);
2060 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2061 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2062 (HeapWord*) _g1_reserved.end(),
2063 _expansion_regions);
2065 // 6843694 - ensure that the maximum region index can fit
2066 // in the remembered set structures.
2067 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2068 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2070 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2071 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2072 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2073 "too many cards per region");
2075 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2077 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2078 heap_word_size(init_byte_size));
2080 _g1h = this;
2082 _in_cset_fast_test_length = max_regions();
2083 _in_cset_fast_test_base =
2084 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2086 // We're biasing _in_cset_fast_test to avoid subtracting the
2087 // beginning of the heap every time we want to index; basically
2088 // it's the same with what we do with the card table.
2089 _in_cset_fast_test = _in_cset_fast_test_base -
2090 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2092 // Clear the _cset_fast_test bitmap in anticipation of adding
2093 // regions to the incremental collection set for the first
2094 // evacuation pause.
2095 clear_cset_fast_test();
2097 // Create the ConcurrentMark data structure and thread.
2098 // (Must do this late, so that "max_regions" is defined.)
2099 _cm = new ConcurrentMark(this, heap_rs);
2100 if (_cm == NULL || !_cm->completed_initialization()) {
2101 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2102 return JNI_ENOMEM;
2103 }
2104 _cmThread = _cm->cmThread();
2106 // Initialize the from_card cache structure of HeapRegionRemSet.
2107 HeapRegionRemSet::init_heap(max_regions());
2109 // Now expand into the initial heap size.
2110 if (!expand(init_byte_size)) {
2111 vm_shutdown_during_initialization("Failed to allocate initial heap.");
2112 return JNI_ENOMEM;
2113 }
2115 // Perform any initialization actions delegated to the policy.
2116 g1_policy()->init();
2118 _refine_cte_cl =
2119 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2120 g1_rem_set(),
2121 concurrent_g1_refine());
2122 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2124 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2125 SATB_Q_FL_lock,
2126 G1SATBProcessCompletedThreshold,
2127 Shared_SATB_Q_lock);
2129 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2130 DirtyCardQ_FL_lock,
2131 concurrent_g1_refine()->yellow_zone(),
2132 concurrent_g1_refine()->red_zone(),
2133 Shared_DirtyCardQ_lock);
2135 if (G1DeferredRSUpdate) {
2136 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2137 DirtyCardQ_FL_lock,
2138 -1, // never trigger processing
2139 -1, // no limit on length
2140 Shared_DirtyCardQ_lock,
2141 &JavaThread::dirty_card_queue_set());
2142 }
2144 // Initialize the card queue set used to hold cards containing
2145 // references into the collection set.
2146 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2147 DirtyCardQ_FL_lock,
2148 -1, // never trigger processing
2149 -1, // no limit on length
2150 Shared_DirtyCardQ_lock,
2151 &JavaThread::dirty_card_queue_set());
2153 // In case we're keeping closure specialization stats, initialize those
2154 // counts and that mechanism.
2155 SpecializationStats::clear();
2157 // Do later initialization work for concurrent refinement.
2158 _cg1r->init();
2160 // Here we allocate the dummy full region that is required by the
2161 // G1AllocRegion class. If we don't pass an address in the reserved
2162 // space here, lots of asserts fire.
2164 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2165 _g1_reserved.start());
2166 // We'll re-use the same region whether the alloc region will
2167 // require BOT updates or not and, if it doesn't, then a non-young
2168 // region will complain that it cannot support allocations without
2169 // BOT updates. So we'll tag the dummy region as young to avoid that.
2170 dummy_region->set_young();
2171 // Make sure it's full.
2172 dummy_region->set_top(dummy_region->end());
2173 G1AllocRegion::setup(this, dummy_region);
2175 init_mutator_alloc_region();
2177 // Do create of the monitoring and management support so that
2178 // values in the heap have been properly initialized.
2179 _g1mm = new G1MonitoringSupport(this);
2181 return JNI_OK;
2182 }
2184 void G1CollectedHeap::ref_processing_init() {
2185 // Reference processing in G1 currently works as follows:
2186 //
2187 // * There are two reference processor instances. One is
2188 // used to record and process discovered references
2189 // during concurrent marking; the other is used to
2190 // record and process references during STW pauses
2191 // (both full and incremental).
2192 // * Both ref processors need to 'span' the entire heap as
2193 // the regions in the collection set may be dotted around.
2194 //
2195 // * For the concurrent marking ref processor:
2196 // * Reference discovery is enabled at initial marking.
2197 // * Reference discovery is disabled and the discovered
2198 // references processed etc during remarking.
2199 // * Reference discovery is MT (see below).
2200 // * Reference discovery requires a barrier (see below).
2201 // * Reference processing may or may not be MT
2202 // (depending on the value of ParallelRefProcEnabled
2203 // and ParallelGCThreads).
2204 // * A full GC disables reference discovery by the CM
2205 // ref processor and abandons any entries on it's
2206 // discovered lists.
2207 //
2208 // * For the STW processor:
2209 // * Non MT discovery is enabled at the start of a full GC.
2210 // * Processing and enqueueing during a full GC is non-MT.
2211 // * During a full GC, references are processed after marking.
2212 //
2213 // * Discovery (may or may not be MT) is enabled at the start
2214 // of an incremental evacuation pause.
2215 // * References are processed near the end of a STW evacuation pause.
2216 // * For both types of GC:
2217 // * Discovery is atomic - i.e. not concurrent.
2218 // * Reference discovery will not need a barrier.
2220 SharedHeap::ref_processing_init();
2221 MemRegion mr = reserved_region();
2223 // Concurrent Mark ref processor
2224 _ref_processor_cm =
2225 new ReferenceProcessor(mr, // span
2226 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2227 // mt processing
2228 (int) ParallelGCThreads,
2229 // degree of mt processing
2230 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2231 // mt discovery
2232 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2233 // degree of mt discovery
2234 false,
2235 // Reference discovery is not atomic
2236 &_is_alive_closure_cm,
2237 // is alive closure
2238 // (for efficiency/performance)
2239 true);
2240 // Setting next fields of discovered
2241 // lists requires a barrier.
2243 // STW ref processor
2244 _ref_processor_stw =
2245 new ReferenceProcessor(mr, // span
2246 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2247 // mt processing
2248 MAX2((int)ParallelGCThreads, 1),
2249 // degree of mt processing
2250 (ParallelGCThreads > 1),
2251 // mt discovery
2252 MAX2((int)ParallelGCThreads, 1),
2253 // degree of mt discovery
2254 true,
2255 // Reference discovery is atomic
2256 &_is_alive_closure_stw,
2257 // is alive closure
2258 // (for efficiency/performance)
2259 false);
2260 // Setting next fields of discovered
2261 // lists requires a barrier.
2262 }
2264 size_t G1CollectedHeap::capacity() const {
2265 return _g1_committed.byte_size();
2266 }
2268 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2269 assert(!hr->continuesHumongous(), "pre-condition");
2270 hr->reset_gc_time_stamp();
2271 if (hr->startsHumongous()) {
2272 uint first_index = hr->hrs_index() + 1;
2273 uint last_index = hr->last_hc_index();
2274 for (uint i = first_index; i < last_index; i += 1) {
2275 HeapRegion* chr = region_at(i);
2276 assert(chr->continuesHumongous(), "sanity");
2277 chr->reset_gc_time_stamp();
2278 }
2279 }
2280 }
2282 #ifndef PRODUCT
2283 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2284 private:
2285 unsigned _gc_time_stamp;
2286 bool _failures;
2288 public:
2289 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2290 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2292 virtual bool doHeapRegion(HeapRegion* hr) {
2293 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2294 if (_gc_time_stamp != region_gc_time_stamp) {
2295 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2296 "expected %d", HR_FORMAT_PARAMS(hr),
2297 region_gc_time_stamp, _gc_time_stamp);
2298 _failures = true;
2299 }
2300 return false;
2301 }
2303 bool failures() { return _failures; }
2304 };
2306 void G1CollectedHeap::check_gc_time_stamps() {
2307 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2308 heap_region_iterate(&cl);
2309 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2310 }
2311 #endif // PRODUCT
2313 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2314 DirtyCardQueue* into_cset_dcq,
2315 bool concurrent,
2316 int worker_i) {
2317 // Clean cards in the hot card cache
2318 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2320 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2321 int n_completed_buffers = 0;
2322 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2323 n_completed_buffers++;
2324 }
2325 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2326 dcqs.clear_n_completed_buffers();
2327 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2328 }
2331 // Computes the sum of the storage used by the various regions.
2333 size_t G1CollectedHeap::used() const {
2334 assert(Heap_lock->owner() != NULL,
2335 "Should be owned on this thread's behalf.");
2336 size_t result = _summary_bytes_used;
2337 // Read only once in case it is set to NULL concurrently
2338 HeapRegion* hr = _mutator_alloc_region.get();
2339 if (hr != NULL)
2340 result += hr->used();
2341 return result;
2342 }
2344 size_t G1CollectedHeap::used_unlocked() const {
2345 size_t result = _summary_bytes_used;
2346 return result;
2347 }
2349 class SumUsedClosure: public HeapRegionClosure {
2350 size_t _used;
2351 public:
2352 SumUsedClosure() : _used(0) {}
2353 bool doHeapRegion(HeapRegion* r) {
2354 if (!r->continuesHumongous()) {
2355 _used += r->used();
2356 }
2357 return false;
2358 }
2359 size_t result() { return _used; }
2360 };
2362 size_t G1CollectedHeap::recalculate_used() const {
2363 SumUsedClosure blk;
2364 heap_region_iterate(&blk);
2365 return blk.result();
2366 }
2368 size_t G1CollectedHeap::unsafe_max_alloc() {
2369 if (free_regions() > 0) return HeapRegion::GrainBytes;
2370 // otherwise, is there space in the current allocation region?
2372 // We need to store the current allocation region in a local variable
2373 // here. The problem is that this method doesn't take any locks and
2374 // there may be other threads which overwrite the current allocation
2375 // region field. attempt_allocation(), for example, sets it to NULL
2376 // and this can happen *after* the NULL check here but before the call
2377 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2378 // to be a problem in the optimized build, since the two loads of the
2379 // current allocation region field are optimized away.
2380 HeapRegion* hr = _mutator_alloc_region.get();
2381 if (hr == NULL) {
2382 return 0;
2383 }
2384 return hr->free();
2385 }
2387 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2388 switch (cause) {
2389 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2390 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2391 case GCCause::_g1_humongous_allocation: return true;
2392 default: return false;
2393 }
2394 }
2396 #ifndef PRODUCT
2397 void G1CollectedHeap::allocate_dummy_regions() {
2398 // Let's fill up most of the region
2399 size_t word_size = HeapRegion::GrainWords - 1024;
2400 // And as a result the region we'll allocate will be humongous.
2401 guarantee(isHumongous(word_size), "sanity");
2403 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2404 // Let's use the existing mechanism for the allocation
2405 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2406 if (dummy_obj != NULL) {
2407 MemRegion mr(dummy_obj, word_size);
2408 CollectedHeap::fill_with_object(mr);
2409 } else {
2410 // If we can't allocate once, we probably cannot allocate
2411 // again. Let's get out of the loop.
2412 break;
2413 }
2414 }
2415 }
2416 #endif // !PRODUCT
2418 void G1CollectedHeap::increment_old_marking_cycles_started() {
2419 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2420 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2421 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2422 _old_marking_cycles_started, _old_marking_cycles_completed));
2424 _old_marking_cycles_started++;
2425 }
2427 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2428 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2430 // We assume that if concurrent == true, then the caller is a
2431 // concurrent thread that was joined the Suspendible Thread
2432 // Set. If there's ever a cheap way to check this, we should add an
2433 // assert here.
2435 // Given that this method is called at the end of a Full GC or of a
2436 // concurrent cycle, and those can be nested (i.e., a Full GC can
2437 // interrupt a concurrent cycle), the number of full collections
2438 // completed should be either one (in the case where there was no
2439 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2440 // behind the number of full collections started.
2442 // This is the case for the inner caller, i.e. a Full GC.
2443 assert(concurrent ||
2444 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2445 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2446 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2447 "is inconsistent with _old_marking_cycles_completed = %u",
2448 _old_marking_cycles_started, _old_marking_cycles_completed));
2450 // This is the case for the outer caller, i.e. the concurrent cycle.
2451 assert(!concurrent ||
2452 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2453 err_msg("for outer caller (concurrent cycle): "
2454 "_old_marking_cycles_started = %u "
2455 "is inconsistent with _old_marking_cycles_completed = %u",
2456 _old_marking_cycles_started, _old_marking_cycles_completed));
2458 _old_marking_cycles_completed += 1;
2460 // We need to clear the "in_progress" flag in the CM thread before
2461 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2462 // is set) so that if a waiter requests another System.gc() it doesn't
2463 // incorrectly see that a marking cyle is still in progress.
2464 if (concurrent) {
2465 _cmThread->clear_in_progress();
2466 }
2468 // This notify_all() will ensure that a thread that called
2469 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2470 // and it's waiting for a full GC to finish will be woken up. It is
2471 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2472 FullGCCount_lock->notify_all();
2473 }
2475 void G1CollectedHeap::collect(GCCause::Cause cause) {
2476 assert_heap_not_locked();
2478 unsigned int gc_count_before;
2479 unsigned int old_marking_count_before;
2480 bool retry_gc;
2482 do {
2483 retry_gc = false;
2485 {
2486 MutexLocker ml(Heap_lock);
2488 // Read the GC count while holding the Heap_lock
2489 gc_count_before = total_collections();
2490 old_marking_count_before = _old_marking_cycles_started;
2491 }
2493 if (should_do_concurrent_full_gc(cause)) {
2494 // Schedule an initial-mark evacuation pause that will start a
2495 // concurrent cycle. We're setting word_size to 0 which means that
2496 // we are not requesting a post-GC allocation.
2497 VM_G1IncCollectionPause op(gc_count_before,
2498 0, /* word_size */
2499 true, /* should_initiate_conc_mark */
2500 g1_policy()->max_pause_time_ms(),
2501 cause);
2503 VMThread::execute(&op);
2504 if (!op.pause_succeeded()) {
2505 if (old_marking_count_before == _old_marking_cycles_started) {
2506 retry_gc = op.should_retry_gc();
2507 } else {
2508 // A Full GC happened while we were trying to schedule the
2509 // initial-mark GC. No point in starting a new cycle given
2510 // that the whole heap was collected anyway.
2511 }
2513 if (retry_gc) {
2514 if (GC_locker::is_active_and_needs_gc()) {
2515 GC_locker::stall_until_clear();
2516 }
2517 }
2518 }
2519 } else {
2520 if (cause == GCCause::_gc_locker
2521 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2523 // Schedule a standard evacuation pause. We're setting word_size
2524 // to 0 which means that we are not requesting a post-GC allocation.
2525 VM_G1IncCollectionPause op(gc_count_before,
2526 0, /* word_size */
2527 false, /* should_initiate_conc_mark */
2528 g1_policy()->max_pause_time_ms(),
2529 cause);
2530 VMThread::execute(&op);
2531 } else {
2532 // Schedule a Full GC.
2533 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2534 VMThread::execute(&op);
2535 }
2536 }
2537 } while (retry_gc);
2538 }
2540 bool G1CollectedHeap::is_in(const void* p) const {
2541 if (_g1_committed.contains(p)) {
2542 // Given that we know that p is in the committed space,
2543 // heap_region_containing_raw() should successfully
2544 // return the containing region.
2545 HeapRegion* hr = heap_region_containing_raw(p);
2546 return hr->is_in(p);
2547 } else {
2548 return false;
2549 }
2550 }
2552 // Iteration functions.
2554 // Iterates an OopClosure over all ref-containing fields of objects
2555 // within a HeapRegion.
2557 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2558 MemRegion _mr;
2559 ExtendedOopClosure* _cl;
2560 public:
2561 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2562 : _mr(mr), _cl(cl) {}
2563 bool doHeapRegion(HeapRegion* r) {
2564 if (!r->continuesHumongous()) {
2565 r->oop_iterate(_cl);
2566 }
2567 return false;
2568 }
2569 };
2571 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2572 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2573 heap_region_iterate(&blk);
2574 }
2576 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2577 IterateOopClosureRegionClosure blk(mr, cl);
2578 heap_region_iterate(&blk);
2579 }
2581 // Iterates an ObjectClosure over all objects within a HeapRegion.
2583 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2584 ObjectClosure* _cl;
2585 public:
2586 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2587 bool doHeapRegion(HeapRegion* r) {
2588 if (! r->continuesHumongous()) {
2589 r->object_iterate(_cl);
2590 }
2591 return false;
2592 }
2593 };
2595 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2596 IterateObjectClosureRegionClosure blk(cl);
2597 heap_region_iterate(&blk);
2598 }
2600 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2601 // FIXME: is this right?
2602 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2603 }
2605 // Calls a SpaceClosure on a HeapRegion.
2607 class SpaceClosureRegionClosure: public HeapRegionClosure {
2608 SpaceClosure* _cl;
2609 public:
2610 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2611 bool doHeapRegion(HeapRegion* r) {
2612 _cl->do_space(r);
2613 return false;
2614 }
2615 };
2617 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2618 SpaceClosureRegionClosure blk(cl);
2619 heap_region_iterate(&blk);
2620 }
2622 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2623 _hrs.iterate(cl);
2624 }
2626 void
2627 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2628 uint worker_id,
2629 uint no_of_par_workers,
2630 jint claim_value) {
2631 const uint regions = n_regions();
2632 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2633 no_of_par_workers :
2634 1);
2635 assert(UseDynamicNumberOfGCThreads ||
2636 no_of_par_workers == workers()->total_workers(),
2637 "Non dynamic should use fixed number of workers");
2638 // try to spread out the starting points of the workers
2639 const HeapRegion* start_hr =
2640 start_region_for_worker(worker_id, no_of_par_workers);
2641 const uint start_index = start_hr->hrs_index();
2643 // each worker will actually look at all regions
2644 for (uint count = 0; count < regions; ++count) {
2645 const uint index = (start_index + count) % regions;
2646 assert(0 <= index && index < regions, "sanity");
2647 HeapRegion* r = region_at(index);
2648 // we'll ignore "continues humongous" regions (we'll process them
2649 // when we come across their corresponding "start humongous"
2650 // region) and regions already claimed
2651 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2652 continue;
2653 }
2654 // OK, try to claim it
2655 if (r->claimHeapRegion(claim_value)) {
2656 // success!
2657 assert(!r->continuesHumongous(), "sanity");
2658 if (r->startsHumongous()) {
2659 // If the region is "starts humongous" we'll iterate over its
2660 // "continues humongous" first; in fact we'll do them
2661 // first. The order is important. In on case, calling the
2662 // closure on the "starts humongous" region might de-allocate
2663 // and clear all its "continues humongous" regions and, as a
2664 // result, we might end up processing them twice. So, we'll do
2665 // them first (notice: most closures will ignore them anyway) and
2666 // then we'll do the "starts humongous" region.
2667 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2668 HeapRegion* chr = region_at(ch_index);
2670 // if the region has already been claimed or it's not
2671 // "continues humongous" we're done
2672 if (chr->claim_value() == claim_value ||
2673 !chr->continuesHumongous()) {
2674 break;
2675 }
2677 // Noone should have claimed it directly. We can given
2678 // that we claimed its "starts humongous" region.
2679 assert(chr->claim_value() != claim_value, "sanity");
2680 assert(chr->humongous_start_region() == r, "sanity");
2682 if (chr->claimHeapRegion(claim_value)) {
2683 // we should always be able to claim it; noone else should
2684 // be trying to claim this region
2686 bool res2 = cl->doHeapRegion(chr);
2687 assert(!res2, "Should not abort");
2689 // Right now, this holds (i.e., no closure that actually
2690 // does something with "continues humongous" regions
2691 // clears them). We might have to weaken it in the future,
2692 // but let's leave these two asserts here for extra safety.
2693 assert(chr->continuesHumongous(), "should still be the case");
2694 assert(chr->humongous_start_region() == r, "sanity");
2695 } else {
2696 guarantee(false, "we should not reach here");
2697 }
2698 }
2699 }
2701 assert(!r->continuesHumongous(), "sanity");
2702 bool res = cl->doHeapRegion(r);
2703 assert(!res, "Should not abort");
2704 }
2705 }
2706 }
2708 class ResetClaimValuesClosure: public HeapRegionClosure {
2709 public:
2710 bool doHeapRegion(HeapRegion* r) {
2711 r->set_claim_value(HeapRegion::InitialClaimValue);
2712 return false;
2713 }
2714 };
2716 void G1CollectedHeap::reset_heap_region_claim_values() {
2717 ResetClaimValuesClosure blk;
2718 heap_region_iterate(&blk);
2719 }
2721 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2722 ResetClaimValuesClosure blk;
2723 collection_set_iterate(&blk);
2724 }
2726 #ifdef ASSERT
2727 // This checks whether all regions in the heap have the correct claim
2728 // value. I also piggy-backed on this a check to ensure that the
2729 // humongous_start_region() information on "continues humongous"
2730 // regions is correct.
2732 class CheckClaimValuesClosure : public HeapRegionClosure {
2733 private:
2734 jint _claim_value;
2735 uint _failures;
2736 HeapRegion* _sh_region;
2738 public:
2739 CheckClaimValuesClosure(jint claim_value) :
2740 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2741 bool doHeapRegion(HeapRegion* r) {
2742 if (r->claim_value() != _claim_value) {
2743 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2744 "claim value = %d, should be %d",
2745 HR_FORMAT_PARAMS(r),
2746 r->claim_value(), _claim_value);
2747 ++_failures;
2748 }
2749 if (!r->isHumongous()) {
2750 _sh_region = NULL;
2751 } else if (r->startsHumongous()) {
2752 _sh_region = r;
2753 } else if (r->continuesHumongous()) {
2754 if (r->humongous_start_region() != _sh_region) {
2755 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2756 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2757 HR_FORMAT_PARAMS(r),
2758 r->humongous_start_region(),
2759 _sh_region);
2760 ++_failures;
2761 }
2762 }
2763 return false;
2764 }
2765 uint failures() { return _failures; }
2766 };
2768 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2769 CheckClaimValuesClosure cl(claim_value);
2770 heap_region_iterate(&cl);
2771 return cl.failures() == 0;
2772 }
2774 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2775 private:
2776 jint _claim_value;
2777 uint _failures;
2779 public:
2780 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2781 _claim_value(claim_value), _failures(0) { }
2783 uint failures() { return _failures; }
2785 bool doHeapRegion(HeapRegion* hr) {
2786 assert(hr->in_collection_set(), "how?");
2787 assert(!hr->isHumongous(), "H-region in CSet");
2788 if (hr->claim_value() != _claim_value) {
2789 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2790 "claim value = %d, should be %d",
2791 HR_FORMAT_PARAMS(hr),
2792 hr->claim_value(), _claim_value);
2793 _failures += 1;
2794 }
2795 return false;
2796 }
2797 };
2799 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2800 CheckClaimValuesInCSetHRClosure cl(claim_value);
2801 collection_set_iterate(&cl);
2802 return cl.failures() == 0;
2803 }
2804 #endif // ASSERT
2806 // Clear the cached CSet starting regions and (more importantly)
2807 // the time stamps. Called when we reset the GC time stamp.
2808 void G1CollectedHeap::clear_cset_start_regions() {
2809 assert(_worker_cset_start_region != NULL, "sanity");
2810 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2812 int n_queues = MAX2((int)ParallelGCThreads, 1);
2813 for (int i = 0; i < n_queues; i++) {
2814 _worker_cset_start_region[i] = NULL;
2815 _worker_cset_start_region_time_stamp[i] = 0;
2816 }
2817 }
2819 // Given the id of a worker, obtain or calculate a suitable
2820 // starting region for iterating over the current collection set.
2821 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2822 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2824 HeapRegion* result = NULL;
2825 unsigned gc_time_stamp = get_gc_time_stamp();
2827 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2828 // Cached starting region for current worker was set
2829 // during the current pause - so it's valid.
2830 // Note: the cached starting heap region may be NULL
2831 // (when the collection set is empty).
2832 result = _worker_cset_start_region[worker_i];
2833 assert(result == NULL || result->in_collection_set(), "sanity");
2834 return result;
2835 }
2837 // The cached entry was not valid so let's calculate
2838 // a suitable starting heap region for this worker.
2840 // We want the parallel threads to start their collection
2841 // set iteration at different collection set regions to
2842 // avoid contention.
2843 // If we have:
2844 // n collection set regions
2845 // p threads
2846 // Then thread t will start at region floor ((t * n) / p)
2848 result = g1_policy()->collection_set();
2849 if (G1CollectedHeap::use_parallel_gc_threads()) {
2850 uint cs_size = g1_policy()->cset_region_length();
2851 uint active_workers = workers()->active_workers();
2852 assert(UseDynamicNumberOfGCThreads ||
2853 active_workers == workers()->total_workers(),
2854 "Unless dynamic should use total workers");
2856 uint end_ind = (cs_size * worker_i) / active_workers;
2857 uint start_ind = 0;
2859 if (worker_i > 0 &&
2860 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2861 // Previous workers starting region is valid
2862 // so let's iterate from there
2863 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2864 result = _worker_cset_start_region[worker_i - 1];
2865 }
2867 for (uint i = start_ind; i < end_ind; i++) {
2868 result = result->next_in_collection_set();
2869 }
2870 }
2872 // Note: the calculated starting heap region may be NULL
2873 // (when the collection set is empty).
2874 assert(result == NULL || result->in_collection_set(), "sanity");
2875 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2876 "should be updated only once per pause");
2877 _worker_cset_start_region[worker_i] = result;
2878 OrderAccess::storestore();
2879 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2880 return result;
2881 }
2883 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2884 uint no_of_par_workers) {
2885 uint worker_num =
2886 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2887 assert(UseDynamicNumberOfGCThreads ||
2888 no_of_par_workers == workers()->total_workers(),
2889 "Non dynamic should use fixed number of workers");
2890 const uint start_index = n_regions() * worker_i / worker_num;
2891 return region_at(start_index);
2892 }
2894 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2895 HeapRegion* r = g1_policy()->collection_set();
2896 while (r != NULL) {
2897 HeapRegion* next = r->next_in_collection_set();
2898 if (cl->doHeapRegion(r)) {
2899 cl->incomplete();
2900 return;
2901 }
2902 r = next;
2903 }
2904 }
2906 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2907 HeapRegionClosure *cl) {
2908 if (r == NULL) {
2909 // The CSet is empty so there's nothing to do.
2910 return;
2911 }
2913 assert(r->in_collection_set(),
2914 "Start region must be a member of the collection set.");
2915 HeapRegion* cur = r;
2916 while (cur != NULL) {
2917 HeapRegion* next = cur->next_in_collection_set();
2918 if (cl->doHeapRegion(cur) && false) {
2919 cl->incomplete();
2920 return;
2921 }
2922 cur = next;
2923 }
2924 cur = g1_policy()->collection_set();
2925 while (cur != r) {
2926 HeapRegion* next = cur->next_in_collection_set();
2927 if (cl->doHeapRegion(cur) && false) {
2928 cl->incomplete();
2929 return;
2930 }
2931 cur = next;
2932 }
2933 }
2935 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2936 return n_regions() > 0 ? region_at(0) : NULL;
2937 }
2940 Space* G1CollectedHeap::space_containing(const void* addr) const {
2941 Space* res = heap_region_containing(addr);
2942 return res;
2943 }
2945 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2946 Space* sp = space_containing(addr);
2947 if (sp != NULL) {
2948 return sp->block_start(addr);
2949 }
2950 return NULL;
2951 }
2953 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2954 Space* sp = space_containing(addr);
2955 assert(sp != NULL, "block_size of address outside of heap");
2956 return sp->block_size(addr);
2957 }
2959 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2960 Space* sp = space_containing(addr);
2961 return sp->block_is_obj(addr);
2962 }
2964 bool G1CollectedHeap::supports_tlab_allocation() const {
2965 return true;
2966 }
2968 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2969 return HeapRegion::GrainBytes;
2970 }
2972 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2973 // Return the remaining space in the cur alloc region, but not less than
2974 // the min TLAB size.
2976 // Also, this value can be at most the humongous object threshold,
2977 // since we can't allow tlabs to grow big enough to accomodate
2978 // humongous objects.
2980 HeapRegion* hr = _mutator_alloc_region.get();
2981 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2982 if (hr == NULL) {
2983 return max_tlab_size;
2984 } else {
2985 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2986 }
2987 }
2989 size_t G1CollectedHeap::max_capacity() const {
2990 return _g1_reserved.byte_size();
2991 }
2993 jlong G1CollectedHeap::millis_since_last_gc() {
2994 // assert(false, "NYI");
2995 return 0;
2996 }
2998 void G1CollectedHeap::prepare_for_verify() {
2999 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3000 ensure_parsability(false);
3001 }
3002 g1_rem_set()->prepare_for_verify();
3003 }
3005 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
3006 VerifyOption vo) {
3007 switch (vo) {
3008 case VerifyOption_G1UsePrevMarking:
3009 return hr->obj_allocated_since_prev_marking(obj);
3010 case VerifyOption_G1UseNextMarking:
3011 return hr->obj_allocated_since_next_marking(obj);
3012 case VerifyOption_G1UseMarkWord:
3013 return false;
3014 default:
3015 ShouldNotReachHere();
3016 }
3017 return false; // keep some compilers happy
3018 }
3020 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3021 switch (vo) {
3022 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3023 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3024 case VerifyOption_G1UseMarkWord: return NULL;
3025 default: ShouldNotReachHere();
3026 }
3027 return NULL; // keep some compilers happy
3028 }
3030 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3031 switch (vo) {
3032 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3033 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3034 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3035 default: ShouldNotReachHere();
3036 }
3037 return false; // keep some compilers happy
3038 }
3040 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3041 switch (vo) {
3042 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3043 case VerifyOption_G1UseNextMarking: return "NTAMS";
3044 case VerifyOption_G1UseMarkWord: return "NONE";
3045 default: ShouldNotReachHere();
3046 }
3047 return NULL; // keep some compilers happy
3048 }
3050 class VerifyLivenessOopClosure: public OopClosure {
3051 G1CollectedHeap* _g1h;
3052 VerifyOption _vo;
3053 public:
3054 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3055 _g1h(g1h), _vo(vo)
3056 { }
3057 void do_oop(narrowOop *p) { do_oop_work(p); }
3058 void do_oop( oop *p) { do_oop_work(p); }
3060 template <class T> void do_oop_work(T *p) {
3061 oop obj = oopDesc::load_decode_heap_oop(p);
3062 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3063 "Dead object referenced by a not dead object");
3064 }
3065 };
3067 class VerifyObjsInRegionClosure: public ObjectClosure {
3068 private:
3069 G1CollectedHeap* _g1h;
3070 size_t _live_bytes;
3071 HeapRegion *_hr;
3072 VerifyOption _vo;
3073 public:
3074 // _vo == UsePrevMarking -> use "prev" marking information,
3075 // _vo == UseNextMarking -> use "next" marking information,
3076 // _vo == UseMarkWord -> use mark word from object header.
3077 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3078 : _live_bytes(0), _hr(hr), _vo(vo) {
3079 _g1h = G1CollectedHeap::heap();
3080 }
3081 void do_object(oop o) {
3082 VerifyLivenessOopClosure isLive(_g1h, _vo);
3083 assert(o != NULL, "Huh?");
3084 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3085 // If the object is alive according to the mark word,
3086 // then verify that the marking information agrees.
3087 // Note we can't verify the contra-positive of the
3088 // above: if the object is dead (according to the mark
3089 // word), it may not be marked, or may have been marked
3090 // but has since became dead, or may have been allocated
3091 // since the last marking.
3092 if (_vo == VerifyOption_G1UseMarkWord) {
3093 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3094 }
3096 o->oop_iterate_no_header(&isLive);
3097 if (!_hr->obj_allocated_since_prev_marking(o)) {
3098 size_t obj_size = o->size(); // Make sure we don't overflow
3099 _live_bytes += (obj_size * HeapWordSize);
3100 }
3101 }
3102 }
3103 size_t live_bytes() { return _live_bytes; }
3104 };
3106 class PrintObjsInRegionClosure : public ObjectClosure {
3107 HeapRegion *_hr;
3108 G1CollectedHeap *_g1;
3109 public:
3110 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3111 _g1 = G1CollectedHeap::heap();
3112 };
3114 void do_object(oop o) {
3115 if (o != NULL) {
3116 HeapWord *start = (HeapWord *) o;
3117 size_t word_sz = o->size();
3118 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3119 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3120 (void*) o, word_sz,
3121 _g1->isMarkedPrev(o),
3122 _g1->isMarkedNext(o),
3123 _hr->obj_allocated_since_prev_marking(o));
3124 HeapWord *end = start + word_sz;
3125 HeapWord *cur;
3126 int *val;
3127 for (cur = start; cur < end; cur++) {
3128 val = (int *) cur;
3129 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3130 }
3131 }
3132 }
3133 };
3135 class VerifyRegionClosure: public HeapRegionClosure {
3136 private:
3137 bool _par;
3138 VerifyOption _vo;
3139 bool _failures;
3140 public:
3141 // _vo == UsePrevMarking -> use "prev" marking information,
3142 // _vo == UseNextMarking -> use "next" marking information,
3143 // _vo == UseMarkWord -> use mark word from object header.
3144 VerifyRegionClosure(bool par, VerifyOption vo)
3145 : _par(par),
3146 _vo(vo),
3147 _failures(false) {}
3149 bool failures() {
3150 return _failures;
3151 }
3153 bool doHeapRegion(HeapRegion* r) {
3154 if (!r->continuesHumongous()) {
3155 bool failures = false;
3156 r->verify(_vo, &failures);
3157 if (failures) {
3158 _failures = true;
3159 } else {
3160 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3161 r->object_iterate(¬_dead_yet_cl);
3162 if (_vo != VerifyOption_G1UseNextMarking) {
3163 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3164 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3165 "max_live_bytes "SIZE_FORMAT" "
3166 "< calculated "SIZE_FORMAT,
3167 r->bottom(), r->end(),
3168 r->max_live_bytes(),
3169 not_dead_yet_cl.live_bytes());
3170 _failures = true;
3171 }
3172 } else {
3173 // When vo == UseNextMarking we cannot currently do a sanity
3174 // check on the live bytes as the calculation has not been
3175 // finalized yet.
3176 }
3177 }
3178 }
3179 return false; // stop the region iteration if we hit a failure
3180 }
3181 };
3183 class YoungRefCounterClosure : public OopClosure {
3184 G1CollectedHeap* _g1h;
3185 int _count;
3186 public:
3187 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3188 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3189 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3191 int count() { return _count; }
3192 void reset_count() { _count = 0; };
3193 };
3195 class VerifyKlassClosure: public KlassClosure {
3196 YoungRefCounterClosure _young_ref_counter_closure;
3197 OopClosure *_oop_closure;
3198 public:
3199 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3200 void do_klass(Klass* k) {
3201 k->oops_do(_oop_closure);
3203 _young_ref_counter_closure.reset_count();
3204 k->oops_do(&_young_ref_counter_closure);
3205 if (_young_ref_counter_closure.count() > 0) {
3206 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3207 }
3208 }
3209 };
3211 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3212 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3213 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3214 // we can change this closure to extend the simpler OopClosure.
3215 class VerifyRootsClosure: public OopsInGenClosure {
3216 private:
3217 G1CollectedHeap* _g1h;
3218 VerifyOption _vo;
3219 bool _failures;
3220 public:
3221 // _vo == UsePrevMarking -> use "prev" marking information,
3222 // _vo == UseNextMarking -> use "next" marking information,
3223 // _vo == UseMarkWord -> use mark word from object header.
3224 VerifyRootsClosure(VerifyOption vo) :
3225 _g1h(G1CollectedHeap::heap()),
3226 _vo(vo),
3227 _failures(false) { }
3229 bool failures() { return _failures; }
3231 template <class T> void do_oop_nv(T* p) {
3232 T heap_oop = oopDesc::load_heap_oop(p);
3233 if (!oopDesc::is_null(heap_oop)) {
3234 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3235 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3236 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3237 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3238 if (_vo == VerifyOption_G1UseMarkWord) {
3239 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3240 }
3241 obj->print_on(gclog_or_tty);
3242 _failures = true;
3243 }
3244 }
3245 }
3247 void do_oop(oop* p) { do_oop_nv(p); }
3248 void do_oop(narrowOop* p) { do_oop_nv(p); }
3249 };
3251 // This is the task used for parallel heap verification.
3253 class G1ParVerifyTask: public AbstractGangTask {
3254 private:
3255 G1CollectedHeap* _g1h;
3256 VerifyOption _vo;
3257 bool _failures;
3259 public:
3260 // _vo == UsePrevMarking -> use "prev" marking information,
3261 // _vo == UseNextMarking -> use "next" marking information,
3262 // _vo == UseMarkWord -> use mark word from object header.
3263 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3264 AbstractGangTask("Parallel verify task"),
3265 _g1h(g1h),
3266 _vo(vo),
3267 _failures(false) { }
3269 bool failures() {
3270 return _failures;
3271 }
3273 void work(uint worker_id) {
3274 HandleMark hm;
3275 VerifyRegionClosure blk(true, _vo);
3276 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3277 _g1h->workers()->active_workers(),
3278 HeapRegion::ParVerifyClaimValue);
3279 if (blk.failures()) {
3280 _failures = true;
3281 }
3282 }
3283 };
3285 void G1CollectedHeap::verify(bool silent) {
3286 verify(silent, VerifyOption_G1UsePrevMarking);
3287 }
3289 void G1CollectedHeap::verify(bool silent,
3290 VerifyOption vo) {
3291 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3292 if (!silent) { gclog_or_tty->print("Roots "); }
3293 VerifyRootsClosure rootsCl(vo);
3295 assert(Thread::current()->is_VM_thread(),
3296 "Expected to be executed serially by the VM thread at this point");
3298 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3299 VerifyKlassClosure klassCl(this, &rootsCl);
3301 // We apply the relevant closures to all the oops in the
3302 // system dictionary, the string table and the code cache.
3303 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3305 // Need cleared claim bits for the strong roots processing
3306 ClassLoaderDataGraph::clear_claimed_marks();
3308 process_strong_roots(true, // activate StrongRootsScope
3309 false, // we set "is scavenging" to false,
3310 // so we don't reset the dirty cards.
3311 ScanningOption(so), // roots scanning options
3312 &rootsCl,
3313 &blobsCl,
3314 &klassCl
3315 );
3317 bool failures = rootsCl.failures();
3319 if (vo != VerifyOption_G1UseMarkWord) {
3320 // If we're verifying during a full GC then the region sets
3321 // will have been torn down at the start of the GC. Therefore
3322 // verifying the region sets will fail. So we only verify
3323 // the region sets when not in a full GC.
3324 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3325 verify_region_sets();
3326 }
3328 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3329 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3330 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3331 "sanity check");
3333 G1ParVerifyTask task(this, vo);
3334 assert(UseDynamicNumberOfGCThreads ||
3335 workers()->active_workers() == workers()->total_workers(),
3336 "If not dynamic should be using all the workers");
3337 int n_workers = workers()->active_workers();
3338 set_par_threads(n_workers);
3339 workers()->run_task(&task);
3340 set_par_threads(0);
3341 if (task.failures()) {
3342 failures = true;
3343 }
3345 // Checks that the expected amount of parallel work was done.
3346 // The implication is that n_workers is > 0.
3347 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3348 "sanity check");
3350 reset_heap_region_claim_values();
3352 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3353 "sanity check");
3354 } else {
3355 VerifyRegionClosure blk(false, vo);
3356 heap_region_iterate(&blk);
3357 if (blk.failures()) {
3358 failures = true;
3359 }
3360 }
3361 if (!silent) gclog_or_tty->print("RemSet ");
3362 rem_set()->verify();
3364 if (failures) {
3365 gclog_or_tty->print_cr("Heap:");
3366 // It helps to have the per-region information in the output to
3367 // help us track down what went wrong. This is why we call
3368 // print_extended_on() instead of print_on().
3369 print_extended_on(gclog_or_tty);
3370 gclog_or_tty->print_cr("");
3371 #ifndef PRODUCT
3372 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3373 concurrent_mark()->print_reachable("at-verification-failure",
3374 vo, false /* all */);
3375 }
3376 #endif
3377 gclog_or_tty->flush();
3378 }
3379 guarantee(!failures, "there should not have been any failures");
3380 } else {
3381 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3382 }
3383 }
3385 class PrintRegionClosure: public HeapRegionClosure {
3386 outputStream* _st;
3387 public:
3388 PrintRegionClosure(outputStream* st) : _st(st) {}
3389 bool doHeapRegion(HeapRegion* r) {
3390 r->print_on(_st);
3391 return false;
3392 }
3393 };
3395 void G1CollectedHeap::print_on(outputStream* st) const {
3396 st->print(" %-20s", "garbage-first heap");
3397 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3398 capacity()/K, used_unlocked()/K);
3399 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3400 _g1_storage.low_boundary(),
3401 _g1_storage.high(),
3402 _g1_storage.high_boundary());
3403 st->cr();
3404 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3405 uint young_regions = _young_list->length();
3406 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3407 (size_t) young_regions * HeapRegion::GrainBytes / K);
3408 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3409 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3410 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3411 st->cr();
3412 MetaspaceAux::print_on(st);
3413 }
3415 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3416 print_on(st);
3418 // Print the per-region information.
3419 st->cr();
3420 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3421 "HS=humongous(starts), HC=humongous(continues), "
3422 "CS=collection set, F=free, TS=gc time stamp, "
3423 "PTAMS=previous top-at-mark-start, "
3424 "NTAMS=next top-at-mark-start)");
3425 PrintRegionClosure blk(st);
3426 heap_region_iterate(&blk);
3427 }
3429 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3430 if (G1CollectedHeap::use_parallel_gc_threads()) {
3431 workers()->print_worker_threads_on(st);
3432 }
3433 _cmThread->print_on(st);
3434 st->cr();
3435 _cm->print_worker_threads_on(st);
3436 _cg1r->print_worker_threads_on(st);
3437 }
3439 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3440 if (G1CollectedHeap::use_parallel_gc_threads()) {
3441 workers()->threads_do(tc);
3442 }
3443 tc->do_thread(_cmThread);
3444 _cg1r->threads_do(tc);
3445 }
3447 void G1CollectedHeap::print_tracing_info() const {
3448 // We'll overload this to mean "trace GC pause statistics."
3449 if (TraceGen0Time || TraceGen1Time) {
3450 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3451 // to that.
3452 g1_policy()->print_tracing_info();
3453 }
3454 if (G1SummarizeRSetStats) {
3455 g1_rem_set()->print_summary_info();
3456 }
3457 if (G1SummarizeConcMark) {
3458 concurrent_mark()->print_summary_info();
3459 }
3460 g1_policy()->print_yg_surv_rate_info();
3461 SpecializationStats::print();
3462 }
3464 #ifndef PRODUCT
3465 // Helpful for debugging RSet issues.
3467 class PrintRSetsClosure : public HeapRegionClosure {
3468 private:
3469 const char* _msg;
3470 size_t _occupied_sum;
3472 public:
3473 bool doHeapRegion(HeapRegion* r) {
3474 HeapRegionRemSet* hrrs = r->rem_set();
3475 size_t occupied = hrrs->occupied();
3476 _occupied_sum += occupied;
3478 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3479 HR_FORMAT_PARAMS(r));
3480 if (occupied == 0) {
3481 gclog_or_tty->print_cr(" RSet is empty");
3482 } else {
3483 hrrs->print();
3484 }
3485 gclog_or_tty->print_cr("----------");
3486 return false;
3487 }
3489 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3490 gclog_or_tty->cr();
3491 gclog_or_tty->print_cr("========================================");
3492 gclog_or_tty->print_cr(msg);
3493 gclog_or_tty->cr();
3494 }
3496 ~PrintRSetsClosure() {
3497 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3498 gclog_or_tty->print_cr("========================================");
3499 gclog_or_tty->cr();
3500 }
3501 };
3503 void G1CollectedHeap::print_cset_rsets() {
3504 PrintRSetsClosure cl("Printing CSet RSets");
3505 collection_set_iterate(&cl);
3506 }
3508 void G1CollectedHeap::print_all_rsets() {
3509 PrintRSetsClosure cl("Printing All RSets");;
3510 heap_region_iterate(&cl);
3511 }
3512 #endif // PRODUCT
3514 G1CollectedHeap* G1CollectedHeap::heap() {
3515 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3516 "not a garbage-first heap");
3517 return _g1h;
3518 }
3520 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3521 // always_do_update_barrier = false;
3522 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3523 // Call allocation profiler
3524 AllocationProfiler::iterate_since_last_gc();
3525 // Fill TLAB's and such
3526 ensure_parsability(true);
3527 }
3529 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3530 // FIXME: what is this about?
3531 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3532 // is set.
3533 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3534 "derived pointer present"));
3535 // always_do_update_barrier = true;
3537 // We have just completed a GC. Update the soft reference
3538 // policy with the new heap occupancy
3539 Universe::update_heap_info_at_gc();
3540 }
3542 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3543 unsigned int gc_count_before,
3544 bool* succeeded) {
3545 assert_heap_not_locked_and_not_at_safepoint();
3546 g1_policy()->record_stop_world_start();
3547 VM_G1IncCollectionPause op(gc_count_before,
3548 word_size,
3549 false, /* should_initiate_conc_mark */
3550 g1_policy()->max_pause_time_ms(),
3551 GCCause::_g1_inc_collection_pause);
3552 VMThread::execute(&op);
3554 HeapWord* result = op.result();
3555 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3556 assert(result == NULL || ret_succeeded,
3557 "the result should be NULL if the VM did not succeed");
3558 *succeeded = ret_succeeded;
3560 assert_heap_not_locked();
3561 return result;
3562 }
3564 void
3565 G1CollectedHeap::doConcurrentMark() {
3566 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3567 if (!_cmThread->in_progress()) {
3568 _cmThread->set_started();
3569 CGC_lock->notify();
3570 }
3571 }
3573 size_t G1CollectedHeap::pending_card_num() {
3574 size_t extra_cards = 0;
3575 JavaThread *curr = Threads::first();
3576 while (curr != NULL) {
3577 DirtyCardQueue& dcq = curr->dirty_card_queue();
3578 extra_cards += dcq.size();
3579 curr = curr->next();
3580 }
3581 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3582 size_t buffer_size = dcqs.buffer_size();
3583 size_t buffer_num = dcqs.completed_buffers_num();
3585 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3586 // in bytes - not the number of 'entries'. We need to convert
3587 // into a number of cards.
3588 return (buffer_size * buffer_num + extra_cards) / oopSize;
3589 }
3591 size_t G1CollectedHeap::cards_scanned() {
3592 return g1_rem_set()->cardsScanned();
3593 }
3595 void
3596 G1CollectedHeap::setup_surviving_young_words() {
3597 assert(_surviving_young_words == NULL, "pre-condition");
3598 uint array_length = g1_policy()->young_cset_region_length();
3599 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3600 if (_surviving_young_words == NULL) {
3601 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3602 "Not enough space for young surv words summary.");
3603 }
3604 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3605 #ifdef ASSERT
3606 for (uint i = 0; i < array_length; ++i) {
3607 assert( _surviving_young_words[i] == 0, "memset above" );
3608 }
3609 #endif // !ASSERT
3610 }
3612 void
3613 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3614 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3615 uint array_length = g1_policy()->young_cset_region_length();
3616 for (uint i = 0; i < array_length; ++i) {
3617 _surviving_young_words[i] += surv_young_words[i];
3618 }
3619 }
3621 void
3622 G1CollectedHeap::cleanup_surviving_young_words() {
3623 guarantee( _surviving_young_words != NULL, "pre-condition" );
3624 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3625 _surviving_young_words = NULL;
3626 }
3628 #ifdef ASSERT
3629 class VerifyCSetClosure: public HeapRegionClosure {
3630 public:
3631 bool doHeapRegion(HeapRegion* hr) {
3632 // Here we check that the CSet region's RSet is ready for parallel
3633 // iteration. The fields that we'll verify are only manipulated
3634 // when the region is part of a CSet and is collected. Afterwards,
3635 // we reset these fields when we clear the region's RSet (when the
3636 // region is freed) so they are ready when the region is
3637 // re-allocated. The only exception to this is if there's an
3638 // evacuation failure and instead of freeing the region we leave
3639 // it in the heap. In that case, we reset these fields during
3640 // evacuation failure handling.
3641 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3643 // Here's a good place to add any other checks we'd like to
3644 // perform on CSet regions.
3645 return false;
3646 }
3647 };
3648 #endif // ASSERT
3650 #if TASKQUEUE_STATS
3651 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3652 st->print_raw_cr("GC Task Stats");
3653 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3654 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3655 }
3657 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3658 print_taskqueue_stats_hdr(st);
3660 TaskQueueStats totals;
3661 const int n = workers() != NULL ? workers()->total_workers() : 1;
3662 for (int i = 0; i < n; ++i) {
3663 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3664 totals += task_queue(i)->stats;
3665 }
3666 st->print_raw("tot "); totals.print(st); st->cr();
3668 DEBUG_ONLY(totals.verify());
3669 }
3671 void G1CollectedHeap::reset_taskqueue_stats() {
3672 const int n = workers() != NULL ? workers()->total_workers() : 1;
3673 for (int i = 0; i < n; ++i) {
3674 task_queue(i)->stats.reset();
3675 }
3676 }
3677 #endif // TASKQUEUE_STATS
3679 void G1CollectedHeap::log_gc_header() {
3680 if (!G1Log::fine()) {
3681 return;
3682 }
3684 gclog_or_tty->date_stamp(PrintGCDateStamps);
3685 gclog_or_tty->stamp(PrintGCTimeStamps);
3687 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3688 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)")
3689 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3691 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3692 }
3694 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3695 if (!G1Log::fine()) {
3696 return;
3697 }
3699 if (G1Log::finer()) {
3700 if (evacuation_failed()) {
3701 gclog_or_tty->print(" (to-space exhausted)");
3702 }
3703 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3704 g1_policy()->phase_times()->note_gc_end();
3705 g1_policy()->phase_times()->print(pause_time_sec);
3706 g1_policy()->print_detailed_heap_transition();
3707 } else {
3708 if (evacuation_failed()) {
3709 gclog_or_tty->print("--");
3710 }
3711 g1_policy()->print_heap_transition();
3712 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3713 }
3714 gclog_or_tty->flush();
3715 }
3717 bool
3718 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3719 assert_at_safepoint(true /* should_be_vm_thread */);
3720 guarantee(!is_gc_active(), "collection is not reentrant");
3722 if (GC_locker::check_active_before_gc()) {
3723 return false;
3724 }
3726 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3727 ResourceMark rm;
3729 print_heap_before_gc();
3731 HRSPhaseSetter x(HRSPhaseEvacuation);
3732 verify_region_sets_optional();
3733 verify_dirty_young_regions();
3735 // This call will decide whether this pause is an initial-mark
3736 // pause. If it is, during_initial_mark_pause() will return true
3737 // for the duration of this pause.
3738 g1_policy()->decide_on_conc_mark_initiation();
3740 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3741 assert(!g1_policy()->during_initial_mark_pause() ||
3742 g1_policy()->gcs_are_young(), "sanity");
3744 // We also do not allow mixed GCs during marking.
3745 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3747 // Record whether this pause is an initial mark. When the current
3748 // thread has completed its logging output and it's safe to signal
3749 // the CM thread, the flag's value in the policy has been reset.
3750 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3752 // Inner scope for scope based logging, timers, and stats collection
3753 {
3754 if (g1_policy()->during_initial_mark_pause()) {
3755 // We are about to start a marking cycle, so we increment the
3756 // full collection counter.
3757 increment_old_marking_cycles_started();
3758 }
3759 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3761 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3762 workers()->active_workers() : 1);
3763 double pause_start_sec = os::elapsedTime();
3764 g1_policy()->phase_times()->note_gc_start(active_workers);
3765 log_gc_header();
3767 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3768 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3770 // If the secondary_free_list is not empty, append it to the
3771 // free_list. No need to wait for the cleanup operation to finish;
3772 // the region allocation code will check the secondary_free_list
3773 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3774 // set, skip this step so that the region allocation code has to
3775 // get entries from the secondary_free_list.
3776 if (!G1StressConcRegionFreeing) {
3777 append_secondary_free_list_if_not_empty_with_lock();
3778 }
3780 assert(check_young_list_well_formed(),
3781 "young list should be well formed");
3783 // Don't dynamically change the number of GC threads this early. A value of
3784 // 0 is used to indicate serial work. When parallel work is done,
3785 // it will be set.
3787 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3788 IsGCActiveMark x;
3790 gc_prologue(false);
3791 increment_total_collections(false /* full gc */);
3792 increment_gc_time_stamp();
3794 verify_before_gc();
3796 COMPILER2_PRESENT(DerivedPointerTable::clear());
3798 // Please see comment in g1CollectedHeap.hpp and
3799 // G1CollectedHeap::ref_processing_init() to see how
3800 // reference processing currently works in G1.
3802 // Enable discovery in the STW reference processor
3803 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3804 true /*verify_no_refs*/);
3806 {
3807 // We want to temporarily turn off discovery by the
3808 // CM ref processor, if necessary, and turn it back on
3809 // on again later if we do. Using a scoped
3810 // NoRefDiscovery object will do this.
3811 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3813 // Forget the current alloc region (we might even choose it to be part
3814 // of the collection set!).
3815 release_mutator_alloc_region();
3817 // We should call this after we retire the mutator alloc
3818 // region(s) so that all the ALLOC / RETIRE events are generated
3819 // before the start GC event.
3820 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3822 // This timing is only used by the ergonomics to handle our pause target.
3823 // It is unclear why this should not include the full pause. We will
3824 // investigate this in CR 7178365.
3825 //
3826 // Preserving the old comment here if that helps the investigation:
3827 //
3828 // The elapsed time induced by the start time below deliberately elides
3829 // the possible verification above.
3830 double sample_start_time_sec = os::elapsedTime();
3831 size_t start_used_bytes = used();
3833 #if YOUNG_LIST_VERBOSE
3834 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3835 _young_list->print();
3836 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3837 #endif // YOUNG_LIST_VERBOSE
3839 g1_policy()->record_collection_pause_start(sample_start_time_sec,
3840 start_used_bytes);
3842 double scan_wait_start = os::elapsedTime();
3843 // We have to wait until the CM threads finish scanning the
3844 // root regions as it's the only way to ensure that all the
3845 // objects on them have been correctly scanned before we start
3846 // moving them during the GC.
3847 bool waited = _cm->root_regions()->wait_until_scan_finished();
3848 double wait_time_ms = 0.0;
3849 if (waited) {
3850 double scan_wait_end = os::elapsedTime();
3851 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3852 }
3853 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3855 #if YOUNG_LIST_VERBOSE
3856 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3857 _young_list->print();
3858 #endif // YOUNG_LIST_VERBOSE
3860 if (g1_policy()->during_initial_mark_pause()) {
3861 concurrent_mark()->checkpointRootsInitialPre();
3862 }
3864 #if YOUNG_LIST_VERBOSE
3865 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3866 _young_list->print();
3867 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3868 #endif // YOUNG_LIST_VERBOSE
3870 g1_policy()->finalize_cset(target_pause_time_ms);
3872 _cm->note_start_of_gc();
3873 // We should not verify the per-thread SATB buffers given that
3874 // we have not filtered them yet (we'll do so during the
3875 // GC). We also call this after finalize_cset() to
3876 // ensure that the CSet has been finalized.
3877 _cm->verify_no_cset_oops(true /* verify_stacks */,
3878 true /* verify_enqueued_buffers */,
3879 false /* verify_thread_buffers */,
3880 true /* verify_fingers */);
3882 if (_hr_printer.is_active()) {
3883 HeapRegion* hr = g1_policy()->collection_set();
3884 while (hr != NULL) {
3885 G1HRPrinter::RegionType type;
3886 if (!hr->is_young()) {
3887 type = G1HRPrinter::Old;
3888 } else if (hr->is_survivor()) {
3889 type = G1HRPrinter::Survivor;
3890 } else {
3891 type = G1HRPrinter::Eden;
3892 }
3893 _hr_printer.cset(hr);
3894 hr = hr->next_in_collection_set();
3895 }
3896 }
3898 #ifdef ASSERT
3899 VerifyCSetClosure cl;
3900 collection_set_iterate(&cl);
3901 #endif // ASSERT
3903 setup_surviving_young_words();
3905 // Initialize the GC alloc regions.
3906 init_gc_alloc_regions();
3908 // Actually do the work...
3909 evacuate_collection_set();
3911 // We do this to mainly verify the per-thread SATB buffers
3912 // (which have been filtered by now) since we didn't verify
3913 // them earlier. No point in re-checking the stacks / enqueued
3914 // buffers given that the CSet has not changed since last time
3915 // we checked.
3916 _cm->verify_no_cset_oops(false /* verify_stacks */,
3917 false /* verify_enqueued_buffers */,
3918 true /* verify_thread_buffers */,
3919 true /* verify_fingers */);
3921 free_collection_set(g1_policy()->collection_set());
3922 g1_policy()->clear_collection_set();
3924 cleanup_surviving_young_words();
3926 // Start a new incremental collection set for the next pause.
3927 g1_policy()->start_incremental_cset_building();
3929 // Clear the _cset_fast_test bitmap in anticipation of adding
3930 // regions to the incremental collection set for the next
3931 // evacuation pause.
3932 clear_cset_fast_test();
3934 _young_list->reset_sampled_info();
3936 // Don't check the whole heap at this point as the
3937 // GC alloc regions from this pause have been tagged
3938 // as survivors and moved on to the survivor list.
3939 // Survivor regions will fail the !is_young() check.
3940 assert(check_young_list_empty(false /* check_heap */),
3941 "young list should be empty");
3943 #if YOUNG_LIST_VERBOSE
3944 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3945 _young_list->print();
3946 #endif // YOUNG_LIST_VERBOSE
3948 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3949 _young_list->first_survivor_region(),
3950 _young_list->last_survivor_region());
3952 _young_list->reset_auxilary_lists();
3954 if (evacuation_failed()) {
3955 _summary_bytes_used = recalculate_used();
3956 } else {
3957 // The "used" of the the collection set have already been subtracted
3958 // when they were freed. Add in the bytes evacuated.
3959 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3960 }
3962 if (g1_policy()->during_initial_mark_pause()) {
3963 // We have to do this before we notify the CM threads that
3964 // they can start working to make sure that all the
3965 // appropriate initialization is done on the CM object.
3966 concurrent_mark()->checkpointRootsInitialPost();
3967 set_marking_started();
3968 // Note that we don't actually trigger the CM thread at
3969 // this point. We do that later when we're sure that
3970 // the current thread has completed its logging output.
3971 }
3973 allocate_dummy_regions();
3975 #if YOUNG_LIST_VERBOSE
3976 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3977 _young_list->print();
3978 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3979 #endif // YOUNG_LIST_VERBOSE
3981 init_mutator_alloc_region();
3983 {
3984 size_t expand_bytes = g1_policy()->expansion_amount();
3985 if (expand_bytes > 0) {
3986 size_t bytes_before = capacity();
3987 // No need for an ergo verbose message here,
3988 // expansion_amount() does this when it returns a value > 0.
3989 if (!expand(expand_bytes)) {
3990 // We failed to expand the heap so let's verify that
3991 // committed/uncommitted amount match the backing store
3992 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3993 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3994 }
3995 }
3996 }
3998 // We redo the verificaiton but now wrt to the new CSet which
3999 // has just got initialized after the previous CSet was freed.
4000 _cm->verify_no_cset_oops(true /* verify_stacks */,
4001 true /* verify_enqueued_buffers */,
4002 true /* verify_thread_buffers */,
4003 true /* verify_fingers */);
4004 _cm->note_end_of_gc();
4006 // This timing is only used by the ergonomics to handle our pause target.
4007 // It is unclear why this should not include the full pause. We will
4008 // investigate this in CR 7178365.
4009 double sample_end_time_sec = os::elapsedTime();
4010 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
4011 g1_policy()->record_collection_pause_end(pause_time_ms);
4013 MemoryService::track_memory_usage();
4015 // In prepare_for_verify() below we'll need to scan the deferred
4016 // update buffers to bring the RSets up-to-date if
4017 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4018 // the update buffers we'll probably need to scan cards on the
4019 // regions we just allocated to (i.e., the GC alloc
4020 // regions). However, during the last GC we called
4021 // set_saved_mark() on all the GC alloc regions, so card
4022 // scanning might skip the [saved_mark_word()...top()] area of
4023 // those regions (i.e., the area we allocated objects into
4024 // during the last GC). But it shouldn't. Given that
4025 // saved_mark_word() is conditional on whether the GC time stamp
4026 // on the region is current or not, by incrementing the GC time
4027 // stamp here we invalidate all the GC time stamps on all the
4028 // regions and saved_mark_word() will simply return top() for
4029 // all the regions. This is a nicer way of ensuring this rather
4030 // than iterating over the regions and fixing them. In fact, the
4031 // GC time stamp increment here also ensures that
4032 // saved_mark_word() will return top() between pauses, i.e.,
4033 // during concurrent refinement. So we don't need the
4034 // is_gc_active() check to decided which top to use when
4035 // scanning cards (see CR 7039627).
4036 increment_gc_time_stamp();
4038 verify_after_gc();
4040 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4041 ref_processor_stw()->verify_no_references_recorded();
4043 // CM reference discovery will be re-enabled if necessary.
4044 }
4046 // We should do this after we potentially expand the heap so
4047 // that all the COMMIT events are generated before the end GC
4048 // event, and after we retire the GC alloc regions so that all
4049 // RETIRE events are generated before the end GC event.
4050 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4052 if (mark_in_progress()) {
4053 concurrent_mark()->update_g1_committed();
4054 }
4056 #ifdef TRACESPINNING
4057 ParallelTaskTerminator::print_termination_counts();
4058 #endif
4060 gc_epilogue(false);
4061 }
4063 // Print the remainder of the GC log output.
4064 log_gc_footer(os::elapsedTime() - pause_start_sec);
4066 // It is not yet to safe to tell the concurrent mark to
4067 // start as we have some optional output below. We don't want the
4068 // output from the concurrent mark thread interfering with this
4069 // logging output either.
4071 _hrs.verify_optional();
4072 verify_region_sets_optional();
4074 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4075 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4077 print_heap_after_gc();
4079 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4080 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4081 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4082 // before any GC notifications are raised.
4083 g1mm()->update_sizes();
4084 }
4086 if (G1SummarizeRSetStats &&
4087 (G1SummarizeRSetStatsPeriod > 0) &&
4088 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4089 g1_rem_set()->print_summary_info();
4090 }
4092 // It should now be safe to tell the concurrent mark thread to start
4093 // without its logging output interfering with the logging output
4094 // that came from the pause.
4096 if (should_start_conc_mark) {
4097 // CAUTION: after the doConcurrentMark() call below,
4098 // the concurrent marking thread(s) could be running
4099 // concurrently with us. Make sure that anything after
4100 // this point does not assume that we are the only GC thread
4101 // running. Note: of course, the actual marking work will
4102 // not start until the safepoint itself is released in
4103 // ConcurrentGCThread::safepoint_desynchronize().
4104 doConcurrentMark();
4105 }
4107 return true;
4108 }
4110 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4111 {
4112 size_t gclab_word_size;
4113 switch (purpose) {
4114 case GCAllocForSurvived:
4115 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4116 break;
4117 case GCAllocForTenured:
4118 gclab_word_size = _old_plab_stats.desired_plab_sz();
4119 break;
4120 default:
4121 assert(false, "unknown GCAllocPurpose");
4122 gclab_word_size = _old_plab_stats.desired_plab_sz();
4123 break;
4124 }
4126 // Prevent humongous PLAB sizes for two reasons:
4127 // * PLABs are allocated using a similar paths as oops, but should
4128 // never be in a humongous region
4129 // * Allowing humongous PLABs needlessly churns the region free lists
4130 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4131 }
4133 void G1CollectedHeap::init_mutator_alloc_region() {
4134 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4135 _mutator_alloc_region.init();
4136 }
4138 void G1CollectedHeap::release_mutator_alloc_region() {
4139 _mutator_alloc_region.release();
4140 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4141 }
4143 void G1CollectedHeap::init_gc_alloc_regions() {
4144 assert_at_safepoint(true /* should_be_vm_thread */);
4146 _survivor_gc_alloc_region.init();
4147 _old_gc_alloc_region.init();
4148 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4149 _retained_old_gc_alloc_region = NULL;
4151 // We will discard the current GC alloc region if:
4152 // a) it's in the collection set (it can happen!),
4153 // b) it's already full (no point in using it),
4154 // c) it's empty (this means that it was emptied during
4155 // a cleanup and it should be on the free list now), or
4156 // d) it's humongous (this means that it was emptied
4157 // during a cleanup and was added to the free list, but
4158 // has been subseqently used to allocate a humongous
4159 // object that may be less than the region size).
4160 if (retained_region != NULL &&
4161 !retained_region->in_collection_set() &&
4162 !(retained_region->top() == retained_region->end()) &&
4163 !retained_region->is_empty() &&
4164 !retained_region->isHumongous()) {
4165 retained_region->set_saved_mark();
4166 // The retained region was added to the old region set when it was
4167 // retired. We have to remove it now, since we don't allow regions
4168 // we allocate to in the region sets. We'll re-add it later, when
4169 // it's retired again.
4170 _old_set.remove(retained_region);
4171 bool during_im = g1_policy()->during_initial_mark_pause();
4172 retained_region->note_start_of_copying(during_im);
4173 _old_gc_alloc_region.set(retained_region);
4174 _hr_printer.reuse(retained_region);
4175 }
4176 }
4178 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4179 _survivor_gc_alloc_region.release();
4180 // If we have an old GC alloc region to release, we'll save it in
4181 // _retained_old_gc_alloc_region. If we don't
4182 // _retained_old_gc_alloc_region will become NULL. This is what we
4183 // want either way so no reason to check explicitly for either
4184 // condition.
4185 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4187 if (ResizePLAB) {
4188 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4189 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4190 }
4191 }
4193 void G1CollectedHeap::abandon_gc_alloc_regions() {
4194 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4195 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4196 _retained_old_gc_alloc_region = NULL;
4197 }
4199 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4200 _drain_in_progress = false;
4201 set_evac_failure_closure(cl);
4202 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4203 }
4205 void G1CollectedHeap::finalize_for_evac_failure() {
4206 assert(_evac_failure_scan_stack != NULL &&
4207 _evac_failure_scan_stack->length() == 0,
4208 "Postcondition");
4209 assert(!_drain_in_progress, "Postcondition");
4210 delete _evac_failure_scan_stack;
4211 _evac_failure_scan_stack = NULL;
4212 }
4214 void G1CollectedHeap::remove_self_forwarding_pointers() {
4215 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4217 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4219 if (G1CollectedHeap::use_parallel_gc_threads()) {
4220 set_par_threads();
4221 workers()->run_task(&rsfp_task);
4222 set_par_threads(0);
4223 } else {
4224 rsfp_task.work(0);
4225 }
4227 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4229 // Reset the claim values in the regions in the collection set.
4230 reset_cset_heap_region_claim_values();
4232 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4234 // Now restore saved marks, if any.
4235 assert(_objs_with_preserved_marks.size() ==
4236 _preserved_marks_of_objs.size(), "Both or none.");
4237 while (!_objs_with_preserved_marks.is_empty()) {
4238 oop obj = _objs_with_preserved_marks.pop();
4239 markOop m = _preserved_marks_of_objs.pop();
4240 obj->set_mark(m);
4241 }
4242 _objs_with_preserved_marks.clear(true);
4243 _preserved_marks_of_objs.clear(true);
4244 }
4246 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4247 _evac_failure_scan_stack->push(obj);
4248 }
4250 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4251 assert(_evac_failure_scan_stack != NULL, "precondition");
4253 while (_evac_failure_scan_stack->length() > 0) {
4254 oop obj = _evac_failure_scan_stack->pop();
4255 _evac_failure_closure->set_region(heap_region_containing(obj));
4256 obj->oop_iterate_backwards(_evac_failure_closure);
4257 }
4258 }
4260 oop
4261 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4262 oop old) {
4263 assert(obj_in_cs(old),
4264 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4265 (HeapWord*) old));
4266 markOop m = old->mark();
4267 oop forward_ptr = old->forward_to_atomic(old);
4268 if (forward_ptr == NULL) {
4269 // Forward-to-self succeeded.
4271 if (_evac_failure_closure != cl) {
4272 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4273 assert(!_drain_in_progress,
4274 "Should only be true while someone holds the lock.");
4275 // Set the global evac-failure closure to the current thread's.
4276 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4277 set_evac_failure_closure(cl);
4278 // Now do the common part.
4279 handle_evacuation_failure_common(old, m);
4280 // Reset to NULL.
4281 set_evac_failure_closure(NULL);
4282 } else {
4283 // The lock is already held, and this is recursive.
4284 assert(_drain_in_progress, "This should only be the recursive case.");
4285 handle_evacuation_failure_common(old, m);
4286 }
4287 return old;
4288 } else {
4289 // Forward-to-self failed. Either someone else managed to allocate
4290 // space for this object (old != forward_ptr) or they beat us in
4291 // self-forwarding it (old == forward_ptr).
4292 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4293 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4294 "should not be in the CSet",
4295 (HeapWord*) old, (HeapWord*) forward_ptr));
4296 return forward_ptr;
4297 }
4298 }
4300 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4301 set_evacuation_failed(true);
4303 preserve_mark_if_necessary(old, m);
4305 HeapRegion* r = heap_region_containing(old);
4306 if (!r->evacuation_failed()) {
4307 r->set_evacuation_failed(true);
4308 _hr_printer.evac_failure(r);
4309 }
4311 push_on_evac_failure_scan_stack(old);
4313 if (!_drain_in_progress) {
4314 // prevent recursion in copy_to_survivor_space()
4315 _drain_in_progress = true;
4316 drain_evac_failure_scan_stack();
4317 _drain_in_progress = false;
4318 }
4319 }
4321 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4322 assert(evacuation_failed(), "Oversaving!");
4323 // We want to call the "for_promotion_failure" version only in the
4324 // case of a promotion failure.
4325 if (m->must_be_preserved_for_promotion_failure(obj)) {
4326 _objs_with_preserved_marks.push(obj);
4327 _preserved_marks_of_objs.push(m);
4328 }
4329 }
4331 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4332 size_t word_size) {
4333 if (purpose == GCAllocForSurvived) {
4334 HeapWord* result = survivor_attempt_allocation(word_size);
4335 if (result != NULL) {
4336 return result;
4337 } else {
4338 // Let's try to allocate in the old gen in case we can fit the
4339 // object there.
4340 return old_attempt_allocation(word_size);
4341 }
4342 } else {
4343 assert(purpose == GCAllocForTenured, "sanity");
4344 HeapWord* result = old_attempt_allocation(word_size);
4345 if (result != NULL) {
4346 return result;
4347 } else {
4348 // Let's try to allocate in the survivors in case we can fit the
4349 // object there.
4350 return survivor_attempt_allocation(word_size);
4351 }
4352 }
4354 ShouldNotReachHere();
4355 // Trying to keep some compilers happy.
4356 return NULL;
4357 }
4359 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4360 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4362 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4363 : _g1h(g1h),
4364 _refs(g1h->task_queue(queue_num)),
4365 _dcq(&g1h->dirty_card_queue_set()),
4366 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4367 _g1_rem(g1h->g1_rem_set()),
4368 _hash_seed(17), _queue_num(queue_num),
4369 _term_attempts(0),
4370 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4371 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4372 _age_table(false),
4373 _strong_roots_time(0), _term_time(0),
4374 _alloc_buffer_waste(0), _undo_waste(0) {
4375 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4376 // we "sacrifice" entry 0 to keep track of surviving bytes for
4377 // non-young regions (where the age is -1)
4378 // We also add a few elements at the beginning and at the end in
4379 // an attempt to eliminate cache contention
4380 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4381 uint array_length = PADDING_ELEM_NUM +
4382 real_length +
4383 PADDING_ELEM_NUM;
4384 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4385 if (_surviving_young_words_base == NULL)
4386 vm_exit_out_of_memory(array_length * sizeof(size_t),
4387 "Not enough space for young surv histo.");
4388 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4389 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4391 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4392 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4394 _start = os::elapsedTime();
4395 }
4397 void
4398 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4399 {
4400 st->print_raw_cr("GC Termination Stats");
4401 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4402 " ------waste (KiB)------");
4403 st->print_raw_cr("thr ms ms % ms % attempts"
4404 " total alloc undo");
4405 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4406 " ------- ------- -------");
4407 }
4409 void
4410 G1ParScanThreadState::print_termination_stats(int i,
4411 outputStream* const st) const
4412 {
4413 const double elapsed_ms = elapsed_time() * 1000.0;
4414 const double s_roots_ms = strong_roots_time() * 1000.0;
4415 const double term_ms = term_time() * 1000.0;
4416 st->print_cr("%3d %9.2f %9.2f %6.2f "
4417 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4418 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4419 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4420 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4421 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4422 alloc_buffer_waste() * HeapWordSize / K,
4423 undo_waste() * HeapWordSize / K);
4424 }
4426 #ifdef ASSERT
4427 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4428 assert(ref != NULL, "invariant");
4429 assert(UseCompressedOops, "sanity");
4430 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4431 oop p = oopDesc::load_decode_heap_oop(ref);
4432 assert(_g1h->is_in_g1_reserved(p),
4433 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4434 return true;
4435 }
4437 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4438 assert(ref != NULL, "invariant");
4439 if (has_partial_array_mask(ref)) {
4440 // Must be in the collection set--it's already been copied.
4441 oop p = clear_partial_array_mask(ref);
4442 assert(_g1h->obj_in_cs(p),
4443 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4444 } else {
4445 oop p = oopDesc::load_decode_heap_oop(ref);
4446 assert(_g1h->is_in_g1_reserved(p),
4447 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4448 }
4449 return true;
4450 }
4452 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4453 if (ref.is_narrow()) {
4454 return verify_ref((narrowOop*) ref);
4455 } else {
4456 return verify_ref((oop*) ref);
4457 }
4458 }
4459 #endif // ASSERT
4461 void G1ParScanThreadState::trim_queue() {
4462 assert(_evac_cl != NULL, "not set");
4463 assert(_evac_failure_cl != NULL, "not set");
4464 assert(_partial_scan_cl != NULL, "not set");
4466 StarTask ref;
4467 do {
4468 // Drain the overflow stack first, so other threads can steal.
4469 while (refs()->pop_overflow(ref)) {
4470 deal_with_reference(ref);
4471 }
4473 while (refs()->pop_local(ref)) {
4474 deal_with_reference(ref);
4475 }
4476 } while (!refs()->is_empty());
4477 }
4479 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4480 G1ParScanThreadState* par_scan_state) :
4481 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4482 _par_scan_state(par_scan_state),
4483 _worker_id(par_scan_state->queue_num()),
4484 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4485 _mark_in_progress(_g1->mark_in_progress()) { }
4487 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4488 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4489 #ifdef ASSERT
4490 HeapRegion* hr = _g1->heap_region_containing(obj);
4491 assert(hr != NULL, "sanity");
4492 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4493 #endif // ASSERT
4495 // We know that the object is not moving so it's safe to read its size.
4496 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4497 }
4499 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4500 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4501 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4502 #ifdef ASSERT
4503 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4504 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4505 assert(from_obj != to_obj, "should not be self-forwarded");
4507 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4508 assert(from_hr != NULL, "sanity");
4509 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4511 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4512 assert(to_hr != NULL, "sanity");
4513 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4514 #endif // ASSERT
4516 // The object might be in the process of being copied by another
4517 // worker so we cannot trust that its to-space image is
4518 // well-formed. So we have to read its size from its from-space
4519 // image which we know should not be changing.
4520 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4521 }
4523 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4524 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4525 ::copy_to_survivor_space(oop old) {
4526 size_t word_sz = old->size();
4527 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4528 // +1 to make the -1 indexes valid...
4529 int young_index = from_region->young_index_in_cset()+1;
4530 assert( (from_region->is_young() && young_index > 0) ||
4531 (!from_region->is_young() && young_index == 0), "invariant" );
4532 G1CollectorPolicy* g1p = _g1->g1_policy();
4533 markOop m = old->mark();
4534 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4535 : m->age();
4536 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4537 word_sz);
4538 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4539 #ifndef PRODUCT
4540 // Should this evacuation fail?
4541 if (_g1->evacuation_should_fail()) {
4542 if (obj_ptr != NULL) {
4543 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4544 obj_ptr = NULL;
4545 }
4546 }
4547 #endif // !PRODUCT
4549 if (obj_ptr == NULL) {
4550 // This will either forward-to-self, or detect that someone else has
4551 // installed a forwarding pointer.
4552 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4553 return _g1->handle_evacuation_failure_par(cl, old);
4554 }
4556 oop obj = oop(obj_ptr);
4558 // We're going to allocate linearly, so might as well prefetch ahead.
4559 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4561 oop forward_ptr = old->forward_to_atomic(obj);
4562 if (forward_ptr == NULL) {
4563 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4564 if (g1p->track_object_age(alloc_purpose)) {
4565 // We could simply do obj->incr_age(). However, this causes a
4566 // performance issue. obj->incr_age() will first check whether
4567 // the object has a displaced mark by checking its mark word;
4568 // getting the mark word from the new location of the object
4569 // stalls. So, given that we already have the mark word and we
4570 // are about to install it anyway, it's better to increase the
4571 // age on the mark word, when the object does not have a
4572 // displaced mark word. We're not expecting many objects to have
4573 // a displaced marked word, so that case is not optimized
4574 // further (it could be...) and we simply call obj->incr_age().
4576 if (m->has_displaced_mark_helper()) {
4577 // in this case, we have to install the mark word first,
4578 // otherwise obj looks to be forwarded (the old mark word,
4579 // which contains the forward pointer, was copied)
4580 obj->set_mark(m);
4581 obj->incr_age();
4582 } else {
4583 m = m->incr_age();
4584 obj->set_mark(m);
4585 }
4586 _par_scan_state->age_table()->add(obj, word_sz);
4587 } else {
4588 obj->set_mark(m);
4589 }
4591 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4592 surv_young_words[young_index] += word_sz;
4594 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4595 // We keep track of the next start index in the length field of
4596 // the to-space object. The actual length can be found in the
4597 // length field of the from-space object.
4598 arrayOop(obj)->set_length(0);
4599 oop* old_p = set_partial_array_mask(old);
4600 _par_scan_state->push_on_queue(old_p);
4601 } else {
4602 // No point in using the slower heap_region_containing() method,
4603 // given that we know obj is in the heap.
4604 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4605 obj->oop_iterate_backwards(&_scanner);
4606 }
4607 } else {
4608 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4609 obj = forward_ptr;
4610 }
4611 return obj;
4612 }
4614 template <class T>
4615 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4616 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4617 _scanned_klass->record_modified_oops();
4618 }
4619 }
4621 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4622 template <class T>
4623 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4624 ::do_oop_work(T* p) {
4625 oop obj = oopDesc::load_decode_heap_oop(p);
4626 assert(barrier != G1BarrierRS || obj != NULL,
4627 "Precondition: G1BarrierRS implies obj is non-NULL");
4629 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4631 // here the null check is implicit in the cset_fast_test() test
4632 if (_g1->in_cset_fast_test(obj)) {
4633 oop forwardee;
4634 if (obj->is_forwarded()) {
4635 forwardee = obj->forwardee();
4636 } else {
4637 forwardee = copy_to_survivor_space(obj);
4638 }
4639 assert(forwardee != NULL, "forwardee should not be NULL");
4640 oopDesc::encode_store_heap_oop(p, forwardee);
4641 if (do_mark_object && forwardee != obj) {
4642 // If the object is self-forwarded we don't need to explicitly
4643 // mark it, the evacuation failure protocol will do so.
4644 mark_forwarded_object(obj, forwardee);
4645 }
4647 // When scanning the RS, we only care about objs in CS.
4648 if (barrier == G1BarrierRS) {
4649 _par_scan_state->update_rs(_from, p, _worker_id);
4650 } else if (barrier == G1BarrierKlass) {
4651 do_klass_barrier(p, forwardee);
4652 }
4653 } else {
4654 // The object is not in collection set. If we're a root scanning
4655 // closure during an initial mark pause (i.e. do_mark_object will
4656 // be true) then attempt to mark the object.
4657 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4658 mark_object(obj);
4659 }
4660 }
4662 if (barrier == G1BarrierEvac && obj != NULL) {
4663 _par_scan_state->update_rs(_from, p, _worker_id);
4664 }
4666 if (do_gen_barrier && obj != NULL) {
4667 par_do_barrier(p);
4668 }
4669 }
4671 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4672 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4674 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4675 assert(has_partial_array_mask(p), "invariant");
4676 oop from_obj = clear_partial_array_mask(p);
4678 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4679 assert(from_obj->is_objArray(), "must be obj array");
4680 objArrayOop from_obj_array = objArrayOop(from_obj);
4681 // The from-space object contains the real length.
4682 int length = from_obj_array->length();
4684 assert(from_obj->is_forwarded(), "must be forwarded");
4685 oop to_obj = from_obj->forwardee();
4686 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4687 objArrayOop to_obj_array = objArrayOop(to_obj);
4688 // We keep track of the next start index in the length field of the
4689 // to-space object.
4690 int next_index = to_obj_array->length();
4691 assert(0 <= next_index && next_index < length,
4692 err_msg("invariant, next index: %d, length: %d", next_index, length));
4694 int start = next_index;
4695 int end = length;
4696 int remainder = end - start;
4697 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4698 if (remainder > 2 * ParGCArrayScanChunk) {
4699 end = start + ParGCArrayScanChunk;
4700 to_obj_array->set_length(end);
4701 // Push the remainder before we process the range in case another
4702 // worker has run out of things to do and can steal it.
4703 oop* from_obj_p = set_partial_array_mask(from_obj);
4704 _par_scan_state->push_on_queue(from_obj_p);
4705 } else {
4706 assert(length == end, "sanity");
4707 // We'll process the final range for this object. Restore the length
4708 // so that the heap remains parsable in case of evacuation failure.
4709 to_obj_array->set_length(end);
4710 }
4711 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4712 // Process indexes [start,end). It will also process the header
4713 // along with the first chunk (i.e., the chunk with start == 0).
4714 // Note that at this point the length field of to_obj_array is not
4715 // correct given that we are using it to keep track of the next
4716 // start index. oop_iterate_range() (thankfully!) ignores the length
4717 // field and only relies on the start / end parameters. It does
4718 // however return the size of the object which will be incorrect. So
4719 // we have to ignore it even if we wanted to use it.
4720 to_obj_array->oop_iterate_range(&_scanner, start, end);
4721 }
4723 class G1ParEvacuateFollowersClosure : public VoidClosure {
4724 protected:
4725 G1CollectedHeap* _g1h;
4726 G1ParScanThreadState* _par_scan_state;
4727 RefToScanQueueSet* _queues;
4728 ParallelTaskTerminator* _terminator;
4730 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4731 RefToScanQueueSet* queues() { return _queues; }
4732 ParallelTaskTerminator* terminator() { return _terminator; }
4734 public:
4735 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4736 G1ParScanThreadState* par_scan_state,
4737 RefToScanQueueSet* queues,
4738 ParallelTaskTerminator* terminator)
4739 : _g1h(g1h), _par_scan_state(par_scan_state),
4740 _queues(queues), _terminator(terminator) {}
4742 void do_void();
4744 private:
4745 inline bool offer_termination();
4746 };
4748 bool G1ParEvacuateFollowersClosure::offer_termination() {
4749 G1ParScanThreadState* const pss = par_scan_state();
4750 pss->start_term_time();
4751 const bool res = terminator()->offer_termination();
4752 pss->end_term_time();
4753 return res;
4754 }
4756 void G1ParEvacuateFollowersClosure::do_void() {
4757 StarTask stolen_task;
4758 G1ParScanThreadState* const pss = par_scan_state();
4759 pss->trim_queue();
4761 do {
4762 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4763 assert(pss->verify_task(stolen_task), "sanity");
4764 if (stolen_task.is_narrow()) {
4765 pss->deal_with_reference((narrowOop*) stolen_task);
4766 } else {
4767 pss->deal_with_reference((oop*) stolen_task);
4768 }
4770 // We've just processed a reference and we might have made
4771 // available new entries on the queues. So we have to make sure
4772 // we drain the queues as necessary.
4773 pss->trim_queue();
4774 }
4775 } while (!offer_termination());
4777 pss->retire_alloc_buffers();
4778 }
4780 class G1KlassScanClosure : public KlassClosure {
4781 G1ParCopyHelper* _closure;
4782 bool _process_only_dirty;
4783 int _count;
4784 public:
4785 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4786 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4787 void do_klass(Klass* klass) {
4788 // If the klass has not been dirtied we know that there's
4789 // no references into the young gen and we can skip it.
4790 if (!_process_only_dirty || klass->has_modified_oops()) {
4791 // Clean the klass since we're going to scavenge all the metadata.
4792 klass->clear_modified_oops();
4794 // Tell the closure that this klass is the Klass to scavenge
4795 // and is the one to dirty if oops are left pointing into the young gen.
4796 _closure->set_scanned_klass(klass);
4798 klass->oops_do(_closure);
4800 _closure->set_scanned_klass(NULL);
4801 }
4802 _count++;
4803 }
4804 };
4806 class G1ParTask : public AbstractGangTask {
4807 protected:
4808 G1CollectedHeap* _g1h;
4809 RefToScanQueueSet *_queues;
4810 ParallelTaskTerminator _terminator;
4811 uint _n_workers;
4813 Mutex _stats_lock;
4814 Mutex* stats_lock() { return &_stats_lock; }
4816 size_t getNCards() {
4817 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4818 / G1BlockOffsetSharedArray::N_bytes;
4819 }
4821 public:
4822 G1ParTask(G1CollectedHeap* g1h,
4823 RefToScanQueueSet *task_queues)
4824 : AbstractGangTask("G1 collection"),
4825 _g1h(g1h),
4826 _queues(task_queues),
4827 _terminator(0, _queues),
4828 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4829 {}
4831 RefToScanQueueSet* queues() { return _queues; }
4833 RefToScanQueue *work_queue(int i) {
4834 return queues()->queue(i);
4835 }
4837 ParallelTaskTerminator* terminator() { return &_terminator; }
4839 virtual void set_for_termination(int active_workers) {
4840 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4841 // in the young space (_par_seq_tasks) in the G1 heap
4842 // for SequentialSubTasksDone.
4843 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4844 // both of which need setting by set_n_termination().
4845 _g1h->SharedHeap::set_n_termination(active_workers);
4846 _g1h->set_n_termination(active_workers);
4847 terminator()->reset_for_reuse(active_workers);
4848 _n_workers = active_workers;
4849 }
4851 void work(uint worker_id) {
4852 if (worker_id >= _n_workers) return; // no work needed this round
4854 double start_time_ms = os::elapsedTime() * 1000.0;
4855 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4857 {
4858 ResourceMark rm;
4859 HandleMark hm;
4861 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4863 G1ParScanThreadState pss(_g1h, worker_id);
4864 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4865 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4866 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4868 pss.set_evac_closure(&scan_evac_cl);
4869 pss.set_evac_failure_closure(&evac_failure_cl);
4870 pss.set_partial_scan_closure(&partial_scan_cl);
4872 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4873 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4875 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4876 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4878 bool only_young = _g1h->g1_policy()->gcs_are_young();
4879 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4880 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4882 OopClosure* scan_root_cl = &only_scan_root_cl;
4883 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4885 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4886 // We also need to mark copied objects.
4887 scan_root_cl = &scan_mark_root_cl;
4888 scan_klasses_cl = &scan_mark_klasses_cl_s;
4889 }
4891 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4893 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4895 pss.start_strong_roots();
4896 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4897 SharedHeap::ScanningOption(so),
4898 scan_root_cl,
4899 &push_heap_rs_cl,
4900 scan_klasses_cl,
4901 worker_id);
4902 pss.end_strong_roots();
4904 {
4905 double start = os::elapsedTime();
4906 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4907 evac.do_void();
4908 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4909 double term_ms = pss.term_time()*1000.0;
4910 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4911 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4912 }
4913 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4914 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4916 if (ParallelGCVerbose) {
4917 MutexLocker x(stats_lock());
4918 pss.print_termination_stats(worker_id);
4919 }
4921 assert(pss.refs()->is_empty(), "should be empty");
4923 // Close the inner scope so that the ResourceMark and HandleMark
4924 // destructors are executed here and are included as part of the
4925 // "GC Worker Time".
4926 }
4928 double end_time_ms = os::elapsedTime() * 1000.0;
4929 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4930 }
4931 };
4933 // *** Common G1 Evacuation Stuff
4935 // Closures that support the filtering of CodeBlobs scanned during
4936 // external root scanning.
4938 // Closure applied to reference fields in code blobs (specifically nmethods)
4939 // to determine whether an nmethod contains references that point into
4940 // the collection set. Used as a predicate when walking code roots so
4941 // that only nmethods that point into the collection set are added to the
4942 // 'marked' list.
4944 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4946 class G1PointsIntoCSOopClosure : public OopClosure {
4947 G1CollectedHeap* _g1;
4948 bool _points_into_cs;
4949 public:
4950 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4951 _g1(g1), _points_into_cs(false) { }
4953 bool points_into_cs() const { return _points_into_cs; }
4955 template <class T>
4956 void do_oop_nv(T* p) {
4957 if (!_points_into_cs) {
4958 T heap_oop = oopDesc::load_heap_oop(p);
4959 if (!oopDesc::is_null(heap_oop) &&
4960 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4961 _points_into_cs = true;
4962 }
4963 }
4964 }
4966 virtual void do_oop(oop* p) { do_oop_nv(p); }
4967 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4968 };
4970 G1CollectedHeap* _g1;
4972 public:
4973 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4974 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4976 virtual void do_code_blob(CodeBlob* cb) {
4977 nmethod* nm = cb->as_nmethod_or_null();
4978 if (nm != NULL && !(nm->test_oops_do_mark())) {
4979 G1PointsIntoCSOopClosure predicate_cl(_g1);
4980 nm->oops_do(&predicate_cl);
4982 if (predicate_cl.points_into_cs()) {
4983 // At least one of the reference fields or the oop relocations
4984 // in the nmethod points into the collection set. We have to
4985 // 'mark' this nmethod.
4986 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4987 // or MarkingCodeBlobClosure::do_code_blob() change.
4988 if (!nm->test_set_oops_do_mark()) {
4989 do_newly_marked_nmethod(nm);
4990 }
4991 }
4992 }
4993 }
4994 };
4996 // This method is run in a GC worker.
4998 void
4999 G1CollectedHeap::
5000 g1_process_strong_roots(bool is_scavenging,
5001 ScanningOption so,
5002 OopClosure* scan_non_heap_roots,
5003 OopsInHeapRegionClosure* scan_rs,
5004 G1KlassScanClosure* scan_klasses,
5005 int worker_i) {
5007 // First scan the strong roots
5008 double ext_roots_start = os::elapsedTime();
5009 double closure_app_time_sec = 0.0;
5011 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5013 // Walk the code cache w/o buffering, because StarTask cannot handle
5014 // unaligned oop locations.
5015 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5017 process_strong_roots(false, // no scoping; this is parallel code
5018 is_scavenging, so,
5019 &buf_scan_non_heap_roots,
5020 &eager_scan_code_roots,
5021 scan_klasses
5022 );
5024 // Now the CM ref_processor roots.
5025 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5026 // We need to treat the discovered reference lists of the
5027 // concurrent mark ref processor as roots and keep entries
5028 // (which are added by the marking threads) on them live
5029 // until they can be processed at the end of marking.
5030 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5031 }
5033 // Finish up any enqueued closure apps (attributed as object copy time).
5034 buf_scan_non_heap_roots.done();
5036 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5038 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5040 double ext_root_time_ms =
5041 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5043 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5045 // During conc marking we have to filter the per-thread SATB buffers
5046 // to make sure we remove any oops into the CSet (which will show up
5047 // as implicitly live).
5048 double satb_filtering_ms = 0.0;
5049 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5050 if (mark_in_progress()) {
5051 double satb_filter_start = os::elapsedTime();
5053 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5055 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5056 }
5057 }
5058 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5060 // Now scan the complement of the collection set.
5061 if (scan_rs != NULL) {
5062 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5063 }
5064 _process_strong_tasks->all_tasks_completed();
5065 }
5067 void
5068 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5069 OopClosure* non_root_closure) {
5070 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5071 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5072 }
5074 // Weak Reference Processing support
5076 // An always "is_alive" closure that is used to preserve referents.
5077 // If the object is non-null then it's alive. Used in the preservation
5078 // of referent objects that are pointed to by reference objects
5079 // discovered by the CM ref processor.
5080 class G1AlwaysAliveClosure: public BoolObjectClosure {
5081 G1CollectedHeap* _g1;
5082 public:
5083 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5084 void do_object(oop p) { assert(false, "Do not call."); }
5085 bool do_object_b(oop p) {
5086 if (p != NULL) {
5087 return true;
5088 }
5089 return false;
5090 }
5091 };
5093 bool G1STWIsAliveClosure::do_object_b(oop p) {
5094 // An object is reachable if it is outside the collection set,
5095 // or is inside and copied.
5096 return !_g1->obj_in_cs(p) || p->is_forwarded();
5097 }
5099 // Non Copying Keep Alive closure
5100 class G1KeepAliveClosure: public OopClosure {
5101 G1CollectedHeap* _g1;
5102 public:
5103 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5104 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5105 void do_oop( oop* p) {
5106 oop obj = *p;
5108 if (_g1->obj_in_cs(obj)) {
5109 assert( obj->is_forwarded(), "invariant" );
5110 *p = obj->forwardee();
5111 }
5112 }
5113 };
5115 // Copying Keep Alive closure - can be called from both
5116 // serial and parallel code as long as different worker
5117 // threads utilize different G1ParScanThreadState instances
5118 // and different queues.
5120 class G1CopyingKeepAliveClosure: public OopClosure {
5121 G1CollectedHeap* _g1h;
5122 OopClosure* _copy_non_heap_obj_cl;
5123 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5124 G1ParScanThreadState* _par_scan_state;
5126 public:
5127 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5128 OopClosure* non_heap_obj_cl,
5129 OopsInHeapRegionClosure* metadata_obj_cl,
5130 G1ParScanThreadState* pss):
5131 _g1h(g1h),
5132 _copy_non_heap_obj_cl(non_heap_obj_cl),
5133 _copy_metadata_obj_cl(metadata_obj_cl),
5134 _par_scan_state(pss)
5135 {}
5137 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5138 virtual void do_oop( oop* p) { do_oop_work(p); }
5140 template <class T> void do_oop_work(T* p) {
5141 oop obj = oopDesc::load_decode_heap_oop(p);
5143 if (_g1h->obj_in_cs(obj)) {
5144 // If the referent object has been forwarded (either copied
5145 // to a new location or to itself in the event of an
5146 // evacuation failure) then we need to update the reference
5147 // field and, if both reference and referent are in the G1
5148 // heap, update the RSet for the referent.
5149 //
5150 // If the referent has not been forwarded then we have to keep
5151 // it alive by policy. Therefore we have copy the referent.
5152 //
5153 // If the reference field is in the G1 heap then we can push
5154 // on the PSS queue. When the queue is drained (after each
5155 // phase of reference processing) the object and it's followers
5156 // will be copied, the reference field set to point to the
5157 // new location, and the RSet updated. Otherwise we need to
5158 // use the the non-heap or metadata closures directly to copy
5159 // the refernt object and update the pointer, while avoiding
5160 // updating the RSet.
5162 if (_g1h->is_in_g1_reserved(p)) {
5163 _par_scan_state->push_on_queue(p);
5164 } else {
5165 assert(!ClassLoaderDataGraph::contains((address)p),
5166 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5167 PTR_FORMAT, p));
5168 _copy_non_heap_obj_cl->do_oop(p);
5169 }
5170 }
5171 }
5172 };
5174 // Serial drain queue closure. Called as the 'complete_gc'
5175 // closure for each discovered list in some of the
5176 // reference processing phases.
5178 class G1STWDrainQueueClosure: public VoidClosure {
5179 protected:
5180 G1CollectedHeap* _g1h;
5181 G1ParScanThreadState* _par_scan_state;
5183 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5185 public:
5186 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5187 _g1h(g1h),
5188 _par_scan_state(pss)
5189 { }
5191 void do_void() {
5192 G1ParScanThreadState* const pss = par_scan_state();
5193 pss->trim_queue();
5194 }
5195 };
5197 // Parallel Reference Processing closures
5199 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5200 // processing during G1 evacuation pauses.
5202 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5203 private:
5204 G1CollectedHeap* _g1h;
5205 RefToScanQueueSet* _queues;
5206 FlexibleWorkGang* _workers;
5207 int _active_workers;
5209 public:
5210 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5211 FlexibleWorkGang* workers,
5212 RefToScanQueueSet *task_queues,
5213 int n_workers) :
5214 _g1h(g1h),
5215 _queues(task_queues),
5216 _workers(workers),
5217 _active_workers(n_workers)
5218 {
5219 assert(n_workers > 0, "shouldn't call this otherwise");
5220 }
5222 // Executes the given task using concurrent marking worker threads.
5223 virtual void execute(ProcessTask& task);
5224 virtual void execute(EnqueueTask& task);
5225 };
5227 // Gang task for possibly parallel reference processing
5229 class G1STWRefProcTaskProxy: public AbstractGangTask {
5230 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5231 ProcessTask& _proc_task;
5232 G1CollectedHeap* _g1h;
5233 RefToScanQueueSet *_task_queues;
5234 ParallelTaskTerminator* _terminator;
5236 public:
5237 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5238 G1CollectedHeap* g1h,
5239 RefToScanQueueSet *task_queues,
5240 ParallelTaskTerminator* terminator) :
5241 AbstractGangTask("Process reference objects in parallel"),
5242 _proc_task(proc_task),
5243 _g1h(g1h),
5244 _task_queues(task_queues),
5245 _terminator(terminator)
5246 {}
5248 virtual void work(uint worker_id) {
5249 // The reference processing task executed by a single worker.
5250 ResourceMark rm;
5251 HandleMark hm;
5253 G1STWIsAliveClosure is_alive(_g1h);
5255 G1ParScanThreadState pss(_g1h, worker_id);
5257 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5258 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5259 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5261 pss.set_evac_closure(&scan_evac_cl);
5262 pss.set_evac_failure_closure(&evac_failure_cl);
5263 pss.set_partial_scan_closure(&partial_scan_cl);
5265 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5266 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5268 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5269 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5271 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5272 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5274 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5275 // We also need to mark copied objects.
5276 copy_non_heap_cl = ©_mark_non_heap_cl;
5277 copy_metadata_cl = ©_mark_metadata_cl;
5278 }
5280 // Keep alive closure.
5281 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5283 // Complete GC closure
5284 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5286 // Call the reference processing task's work routine.
5287 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5289 // Note we cannot assert that the refs array is empty here as not all
5290 // of the processing tasks (specifically phase2 - pp2_work) execute
5291 // the complete_gc closure (which ordinarily would drain the queue) so
5292 // the queue may not be empty.
5293 }
5294 };
5296 // Driver routine for parallel reference processing.
5297 // Creates an instance of the ref processing gang
5298 // task and has the worker threads execute it.
5299 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5300 assert(_workers != NULL, "Need parallel worker threads.");
5302 ParallelTaskTerminator terminator(_active_workers, _queues);
5303 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5305 _g1h->set_par_threads(_active_workers);
5306 _workers->run_task(&proc_task_proxy);
5307 _g1h->set_par_threads(0);
5308 }
5310 // Gang task for parallel reference enqueueing.
5312 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5313 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5314 EnqueueTask& _enq_task;
5316 public:
5317 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5318 AbstractGangTask("Enqueue reference objects in parallel"),
5319 _enq_task(enq_task)
5320 { }
5322 virtual void work(uint worker_id) {
5323 _enq_task.work(worker_id);
5324 }
5325 };
5327 // Driver routine for parallel reference enqueing.
5328 // Creates an instance of the ref enqueueing gang
5329 // task and has the worker threads execute it.
5331 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5332 assert(_workers != NULL, "Need parallel worker threads.");
5334 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5336 _g1h->set_par_threads(_active_workers);
5337 _workers->run_task(&enq_task_proxy);
5338 _g1h->set_par_threads(0);
5339 }
5341 // End of weak reference support closures
5343 // Abstract task used to preserve (i.e. copy) any referent objects
5344 // that are in the collection set and are pointed to by reference
5345 // objects discovered by the CM ref processor.
5347 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5348 protected:
5349 G1CollectedHeap* _g1h;
5350 RefToScanQueueSet *_queues;
5351 ParallelTaskTerminator _terminator;
5352 uint _n_workers;
5354 public:
5355 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5356 AbstractGangTask("ParPreserveCMReferents"),
5357 _g1h(g1h),
5358 _queues(task_queues),
5359 _terminator(workers, _queues),
5360 _n_workers(workers)
5361 { }
5363 void work(uint worker_id) {
5364 ResourceMark rm;
5365 HandleMark hm;
5367 G1ParScanThreadState pss(_g1h, worker_id);
5368 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5369 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5370 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5372 pss.set_evac_closure(&scan_evac_cl);
5373 pss.set_evac_failure_closure(&evac_failure_cl);
5374 pss.set_partial_scan_closure(&partial_scan_cl);
5376 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5379 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5380 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5382 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5383 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5385 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5386 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5388 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5389 // We also need to mark copied objects.
5390 copy_non_heap_cl = ©_mark_non_heap_cl;
5391 copy_metadata_cl = ©_mark_metadata_cl;
5392 }
5394 // Is alive closure
5395 G1AlwaysAliveClosure always_alive(_g1h);
5397 // Copying keep alive closure. Applied to referent objects that need
5398 // to be copied.
5399 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5401 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5403 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5404 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5406 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5407 // So this must be true - but assert just in case someone decides to
5408 // change the worker ids.
5409 assert(0 <= worker_id && worker_id < limit, "sanity");
5410 assert(!rp->discovery_is_atomic(), "check this code");
5412 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5413 for (uint idx = worker_id; idx < limit; idx += stride) {
5414 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5416 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5417 while (iter.has_next()) {
5418 // Since discovery is not atomic for the CM ref processor, we
5419 // can see some null referent objects.
5420 iter.load_ptrs(DEBUG_ONLY(true));
5421 oop ref = iter.obj();
5423 // This will filter nulls.
5424 if (iter.is_referent_alive()) {
5425 iter.make_referent_alive();
5426 }
5427 iter.move_to_next();
5428 }
5429 }
5431 // Drain the queue - which may cause stealing
5432 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5433 drain_queue.do_void();
5434 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5435 assert(pss.refs()->is_empty(), "should be");
5436 }
5437 };
5439 // Weak Reference processing during an evacuation pause (part 1).
5440 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5441 double ref_proc_start = os::elapsedTime();
5443 ReferenceProcessor* rp = _ref_processor_stw;
5444 assert(rp->discovery_enabled(), "should have been enabled");
5446 // Any reference objects, in the collection set, that were 'discovered'
5447 // by the CM ref processor should have already been copied (either by
5448 // applying the external root copy closure to the discovered lists, or
5449 // by following an RSet entry).
5450 //
5451 // But some of the referents, that are in the collection set, that these
5452 // reference objects point to may not have been copied: the STW ref
5453 // processor would have seen that the reference object had already
5454 // been 'discovered' and would have skipped discovering the reference,
5455 // but would not have treated the reference object as a regular oop.
5456 // As a reult the copy closure would not have been applied to the
5457 // referent object.
5458 //
5459 // We need to explicitly copy these referent objects - the references
5460 // will be processed at the end of remarking.
5461 //
5462 // We also need to do this copying before we process the reference
5463 // objects discovered by the STW ref processor in case one of these
5464 // referents points to another object which is also referenced by an
5465 // object discovered by the STW ref processor.
5467 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5468 no_of_gc_workers == workers()->active_workers(),
5469 "Need to reset active GC workers");
5471 set_par_threads(no_of_gc_workers);
5472 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5473 no_of_gc_workers,
5474 _task_queues);
5476 if (G1CollectedHeap::use_parallel_gc_threads()) {
5477 workers()->run_task(&keep_cm_referents);
5478 } else {
5479 keep_cm_referents.work(0);
5480 }
5482 set_par_threads(0);
5484 // Closure to test whether a referent is alive.
5485 G1STWIsAliveClosure is_alive(this);
5487 // Even when parallel reference processing is enabled, the processing
5488 // of JNI refs is serial and performed serially by the current thread
5489 // rather than by a worker. The following PSS will be used for processing
5490 // JNI refs.
5492 // Use only a single queue for this PSS.
5493 G1ParScanThreadState pss(this, 0);
5495 // We do not embed a reference processor in the copying/scanning
5496 // closures while we're actually processing the discovered
5497 // reference objects.
5498 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5499 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5500 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5502 pss.set_evac_closure(&scan_evac_cl);
5503 pss.set_evac_failure_closure(&evac_failure_cl);
5504 pss.set_partial_scan_closure(&partial_scan_cl);
5506 assert(pss.refs()->is_empty(), "pre-condition");
5508 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5509 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5511 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5512 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5514 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5515 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5517 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5518 // We also need to mark copied objects.
5519 copy_non_heap_cl = ©_mark_non_heap_cl;
5520 copy_metadata_cl = ©_mark_metadata_cl;
5521 }
5523 // Keep alive closure.
5524 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5526 // Serial Complete GC closure
5527 G1STWDrainQueueClosure drain_queue(this, &pss);
5529 // Setup the soft refs policy...
5530 rp->setup_policy(false);
5532 if (!rp->processing_is_mt()) {
5533 // Serial reference processing...
5534 rp->process_discovered_references(&is_alive,
5535 &keep_alive,
5536 &drain_queue,
5537 NULL);
5538 } else {
5539 // Parallel reference processing
5540 assert(rp->num_q() == no_of_gc_workers, "sanity");
5541 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5543 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5544 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5545 }
5547 // We have completed copying any necessary live referent objects
5548 // (that were not copied during the actual pause) so we can
5549 // retire any active alloc buffers
5550 pss.retire_alloc_buffers();
5551 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5553 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5554 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5555 }
5557 // Weak Reference processing during an evacuation pause (part 2).
5558 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5559 double ref_enq_start = os::elapsedTime();
5561 ReferenceProcessor* rp = _ref_processor_stw;
5562 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5564 // Now enqueue any remaining on the discovered lists on to
5565 // the pending list.
5566 if (!rp->processing_is_mt()) {
5567 // Serial reference processing...
5568 rp->enqueue_discovered_references();
5569 } else {
5570 // Parallel reference enqueuing
5572 assert(no_of_gc_workers == workers()->active_workers(),
5573 "Need to reset active workers");
5574 assert(rp->num_q() == no_of_gc_workers, "sanity");
5575 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5577 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5578 rp->enqueue_discovered_references(&par_task_executor);
5579 }
5581 rp->verify_no_references_recorded();
5582 assert(!rp->discovery_enabled(), "should have been disabled");
5584 // FIXME
5585 // CM's reference processing also cleans up the string and symbol tables.
5586 // Should we do that here also? We could, but it is a serial operation
5587 // and could signicantly increase the pause time.
5589 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5590 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5591 }
5593 void G1CollectedHeap::evacuate_collection_set() {
5594 _expand_heap_after_alloc_failure = true;
5595 set_evacuation_failed(false);
5597 // Should G1EvacuationFailureALot be in effect for this GC?
5598 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5600 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5601 concurrent_g1_refine()->set_use_cache(false);
5602 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5604 uint n_workers;
5605 if (G1CollectedHeap::use_parallel_gc_threads()) {
5606 n_workers =
5607 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5608 workers()->active_workers(),
5609 Threads::number_of_non_daemon_threads());
5610 assert(UseDynamicNumberOfGCThreads ||
5611 n_workers == workers()->total_workers(),
5612 "If not dynamic should be using all the workers");
5613 workers()->set_active_workers(n_workers);
5614 set_par_threads(n_workers);
5615 } else {
5616 assert(n_par_threads() == 0,
5617 "Should be the original non-parallel value");
5618 n_workers = 1;
5619 }
5621 G1ParTask g1_par_task(this, _task_queues);
5623 init_for_evac_failure(NULL);
5625 rem_set()->prepare_for_younger_refs_iterate(true);
5627 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5628 double start_par_time_sec = os::elapsedTime();
5629 double end_par_time_sec;
5631 {
5632 StrongRootsScope srs(this);
5634 if (G1CollectedHeap::use_parallel_gc_threads()) {
5635 // The individual threads will set their evac-failure closures.
5636 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5637 // These tasks use ShareHeap::_process_strong_tasks
5638 assert(UseDynamicNumberOfGCThreads ||
5639 workers()->active_workers() == workers()->total_workers(),
5640 "If not dynamic should be using all the workers");
5641 workers()->run_task(&g1_par_task);
5642 } else {
5643 g1_par_task.set_for_termination(n_workers);
5644 g1_par_task.work(0);
5645 }
5646 end_par_time_sec = os::elapsedTime();
5648 // Closing the inner scope will execute the destructor
5649 // for the StrongRootsScope object. We record the current
5650 // elapsed time before closing the scope so that time
5651 // taken for the SRS destructor is NOT included in the
5652 // reported parallel time.
5653 }
5655 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5656 g1_policy()->phase_times()->record_par_time(par_time_ms);
5658 double code_root_fixup_time_ms =
5659 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5660 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5662 set_par_threads(0);
5664 // Process any discovered reference objects - we have
5665 // to do this _before_ we retire the GC alloc regions
5666 // as we may have to copy some 'reachable' referent
5667 // objects (and their reachable sub-graphs) that were
5668 // not copied during the pause.
5669 process_discovered_references(n_workers);
5671 // Weak root processing.
5672 // Note: when JSR 292 is enabled and code blobs can contain
5673 // non-perm oops then we will need to process the code blobs
5674 // here too.
5675 {
5676 G1STWIsAliveClosure is_alive(this);
5677 G1KeepAliveClosure keep_alive(this);
5678 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5679 }
5681 release_gc_alloc_regions(n_workers);
5682 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5684 concurrent_g1_refine()->clear_hot_cache();
5685 concurrent_g1_refine()->set_use_cache(true);
5687 finalize_for_evac_failure();
5689 if (evacuation_failed()) {
5690 remove_self_forwarding_pointers();
5692 // Reset the G1EvacuationFailureALot counters and flags
5693 // Note: the values are reset only when an actual
5694 // evacuation failure occurs.
5695 NOT_PRODUCT(reset_evacuation_should_fail();)
5696 }
5698 // Enqueue any remaining references remaining on the STW
5699 // reference processor's discovered lists. We need to do
5700 // this after the card table is cleaned (and verified) as
5701 // the act of enqueuing entries on to the pending list
5702 // will log these updates (and dirty their associated
5703 // cards). We need these updates logged to update any
5704 // RSets.
5705 enqueue_discovered_references(n_workers);
5707 if (G1DeferredRSUpdate) {
5708 RedirtyLoggedCardTableEntryFastClosure redirty;
5709 dirty_card_queue_set().set_closure(&redirty);
5710 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5712 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5713 dcq.merge_bufferlists(&dirty_card_queue_set());
5714 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5715 }
5716 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5717 }
5719 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5720 size_t* pre_used,
5721 FreeRegionList* free_list,
5722 OldRegionSet* old_proxy_set,
5723 HumongousRegionSet* humongous_proxy_set,
5724 HRRSCleanupTask* hrrs_cleanup_task,
5725 bool par) {
5726 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5727 if (hr->isHumongous()) {
5728 assert(hr->startsHumongous(), "we should only see starts humongous");
5729 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5730 } else {
5731 _old_set.remove_with_proxy(hr, old_proxy_set);
5732 free_region(hr, pre_used, free_list, par);
5733 }
5734 } else {
5735 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5736 }
5737 }
5739 void G1CollectedHeap::free_region(HeapRegion* hr,
5740 size_t* pre_used,
5741 FreeRegionList* free_list,
5742 bool par) {
5743 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5744 assert(!hr->is_empty(), "the region should not be empty");
5745 assert(free_list != NULL, "pre-condition");
5747 *pre_used += hr->used();
5748 hr->hr_clear(par, true /* clear_space */);
5749 free_list->add_as_head(hr);
5750 }
5752 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5753 size_t* pre_used,
5754 FreeRegionList* free_list,
5755 HumongousRegionSet* humongous_proxy_set,
5756 bool par) {
5757 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5758 assert(free_list != NULL, "pre-condition");
5759 assert(humongous_proxy_set != NULL, "pre-condition");
5761 size_t hr_used = hr->used();
5762 size_t hr_capacity = hr->capacity();
5763 size_t hr_pre_used = 0;
5764 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5765 // We need to read this before we make the region non-humongous,
5766 // otherwise the information will be gone.
5767 uint last_index = hr->last_hc_index();
5768 hr->set_notHumongous();
5769 free_region(hr, &hr_pre_used, free_list, par);
5771 uint i = hr->hrs_index() + 1;
5772 while (i < last_index) {
5773 HeapRegion* curr_hr = region_at(i);
5774 assert(curr_hr->continuesHumongous(), "invariant");
5775 curr_hr->set_notHumongous();
5776 free_region(curr_hr, &hr_pre_used, free_list, par);
5777 i += 1;
5778 }
5779 assert(hr_pre_used == hr_used,
5780 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5781 "should be the same", hr_pre_used, hr_used));
5782 *pre_used += hr_pre_used;
5783 }
5785 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5786 FreeRegionList* free_list,
5787 OldRegionSet* old_proxy_set,
5788 HumongousRegionSet* humongous_proxy_set,
5789 bool par) {
5790 if (pre_used > 0) {
5791 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5792 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5793 assert(_summary_bytes_used >= pre_used,
5794 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5795 "should be >= pre_used: "SIZE_FORMAT,
5796 _summary_bytes_used, pre_used));
5797 _summary_bytes_used -= pre_used;
5798 }
5799 if (free_list != NULL && !free_list->is_empty()) {
5800 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5801 _free_list.add_as_head(free_list);
5802 }
5803 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5804 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5805 _old_set.update_from_proxy(old_proxy_set);
5806 }
5807 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5808 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5809 _humongous_set.update_from_proxy(humongous_proxy_set);
5810 }
5811 }
5813 class G1ParCleanupCTTask : public AbstractGangTask {
5814 CardTableModRefBS* _ct_bs;
5815 G1CollectedHeap* _g1h;
5816 HeapRegion* volatile _su_head;
5817 public:
5818 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5819 G1CollectedHeap* g1h) :
5820 AbstractGangTask("G1 Par Cleanup CT Task"),
5821 _ct_bs(ct_bs), _g1h(g1h) { }
5823 void work(uint worker_id) {
5824 HeapRegion* r;
5825 while (r = _g1h->pop_dirty_cards_region()) {
5826 clear_cards(r);
5827 }
5828 }
5830 void clear_cards(HeapRegion* r) {
5831 // Cards of the survivors should have already been dirtied.
5832 if (!r->is_survivor()) {
5833 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5834 }
5835 }
5836 };
5838 #ifndef PRODUCT
5839 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5840 G1CollectedHeap* _g1h;
5841 CardTableModRefBS* _ct_bs;
5842 public:
5843 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5844 : _g1h(g1h), _ct_bs(ct_bs) { }
5845 virtual bool doHeapRegion(HeapRegion* r) {
5846 if (r->is_survivor()) {
5847 _g1h->verify_dirty_region(r);
5848 } else {
5849 _g1h->verify_not_dirty_region(r);
5850 }
5851 return false;
5852 }
5853 };
5855 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5856 // All of the region should be clean.
5857 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5858 MemRegion mr(hr->bottom(), hr->end());
5859 ct_bs->verify_not_dirty_region(mr);
5860 }
5862 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5863 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5864 // dirty allocated blocks as they allocate them. The thread that
5865 // retires each region and replaces it with a new one will do a
5866 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5867 // not dirty that area (one less thing to have to do while holding
5868 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5869 // is dirty.
5870 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5871 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5872 ct_bs->verify_dirty_region(mr);
5873 }
5875 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5876 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5877 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5878 verify_dirty_region(hr);
5879 }
5880 }
5882 void G1CollectedHeap::verify_dirty_young_regions() {
5883 verify_dirty_young_list(_young_list->first_region());
5884 }
5885 #endif
5887 void G1CollectedHeap::cleanUpCardTable() {
5888 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5889 double start = os::elapsedTime();
5891 {
5892 // Iterate over the dirty cards region list.
5893 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5895 if (G1CollectedHeap::use_parallel_gc_threads()) {
5896 set_par_threads();
5897 workers()->run_task(&cleanup_task);
5898 set_par_threads(0);
5899 } else {
5900 while (_dirty_cards_region_list) {
5901 HeapRegion* r = _dirty_cards_region_list;
5902 cleanup_task.clear_cards(r);
5903 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5904 if (_dirty_cards_region_list == r) {
5905 // The last region.
5906 _dirty_cards_region_list = NULL;
5907 }
5908 r->set_next_dirty_cards_region(NULL);
5909 }
5910 }
5911 #ifndef PRODUCT
5912 if (G1VerifyCTCleanup || VerifyAfterGC) {
5913 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5914 heap_region_iterate(&cleanup_verifier);
5915 }
5916 #endif
5917 }
5919 double elapsed = os::elapsedTime() - start;
5920 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5921 }
5923 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5924 size_t pre_used = 0;
5925 FreeRegionList local_free_list("Local List for CSet Freeing");
5927 double young_time_ms = 0.0;
5928 double non_young_time_ms = 0.0;
5930 // Since the collection set is a superset of the the young list,
5931 // all we need to do to clear the young list is clear its
5932 // head and length, and unlink any young regions in the code below
5933 _young_list->clear();
5935 G1CollectorPolicy* policy = g1_policy();
5937 double start_sec = os::elapsedTime();
5938 bool non_young = true;
5940 HeapRegion* cur = cs_head;
5941 int age_bound = -1;
5942 size_t rs_lengths = 0;
5944 while (cur != NULL) {
5945 assert(!is_on_master_free_list(cur), "sanity");
5946 if (non_young) {
5947 if (cur->is_young()) {
5948 double end_sec = os::elapsedTime();
5949 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5950 non_young_time_ms += elapsed_ms;
5952 start_sec = os::elapsedTime();
5953 non_young = false;
5954 }
5955 } else {
5956 if (!cur->is_young()) {
5957 double end_sec = os::elapsedTime();
5958 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5959 young_time_ms += elapsed_ms;
5961 start_sec = os::elapsedTime();
5962 non_young = true;
5963 }
5964 }
5966 rs_lengths += cur->rem_set()->occupied();
5968 HeapRegion* next = cur->next_in_collection_set();
5969 assert(cur->in_collection_set(), "bad CS");
5970 cur->set_next_in_collection_set(NULL);
5971 cur->set_in_collection_set(false);
5973 if (cur->is_young()) {
5974 int index = cur->young_index_in_cset();
5975 assert(index != -1, "invariant");
5976 assert((uint) index < policy->young_cset_region_length(), "invariant");
5977 size_t words_survived = _surviving_young_words[index];
5978 cur->record_surv_words_in_group(words_survived);
5980 // At this point the we have 'popped' cur from the collection set
5981 // (linked via next_in_collection_set()) but it is still in the
5982 // young list (linked via next_young_region()). Clear the
5983 // _next_young_region field.
5984 cur->set_next_young_region(NULL);
5985 } else {
5986 int index = cur->young_index_in_cset();
5987 assert(index == -1, "invariant");
5988 }
5990 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5991 (!cur->is_young() && cur->young_index_in_cset() == -1),
5992 "invariant" );
5994 if (!cur->evacuation_failed()) {
5995 MemRegion used_mr = cur->used_region();
5997 // And the region is empty.
5998 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5999 free_region(cur, &pre_used, &local_free_list, false /* par */);
6000 } else {
6001 cur->uninstall_surv_rate_group();
6002 if (cur->is_young()) {
6003 cur->set_young_index_in_cset(-1);
6004 }
6005 cur->set_not_young();
6006 cur->set_evacuation_failed(false);
6007 // The region is now considered to be old.
6008 _old_set.add(cur);
6009 }
6010 cur = next;
6011 }
6013 policy->record_max_rs_lengths(rs_lengths);
6014 policy->cset_regions_freed();
6016 double end_sec = os::elapsedTime();
6017 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6019 if (non_young) {
6020 non_young_time_ms += elapsed_ms;
6021 } else {
6022 young_time_ms += elapsed_ms;
6023 }
6025 update_sets_after_freeing_regions(pre_used, &local_free_list,
6026 NULL /* old_proxy_set */,
6027 NULL /* humongous_proxy_set */,
6028 false /* par */);
6029 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6030 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6031 }
6033 // This routine is similar to the above but does not record
6034 // any policy statistics or update free lists; we are abandoning
6035 // the current incremental collection set in preparation of a
6036 // full collection. After the full GC we will start to build up
6037 // the incremental collection set again.
6038 // This is only called when we're doing a full collection
6039 // and is immediately followed by the tearing down of the young list.
6041 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6042 HeapRegion* cur = cs_head;
6044 while (cur != NULL) {
6045 HeapRegion* next = cur->next_in_collection_set();
6046 assert(cur->in_collection_set(), "bad CS");
6047 cur->set_next_in_collection_set(NULL);
6048 cur->set_in_collection_set(false);
6049 cur->set_young_index_in_cset(-1);
6050 cur = next;
6051 }
6052 }
6054 void G1CollectedHeap::set_free_regions_coming() {
6055 if (G1ConcRegionFreeingVerbose) {
6056 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6057 "setting free regions coming");
6058 }
6060 assert(!free_regions_coming(), "pre-condition");
6061 _free_regions_coming = true;
6062 }
6064 void G1CollectedHeap::reset_free_regions_coming() {
6065 assert(free_regions_coming(), "pre-condition");
6067 {
6068 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6069 _free_regions_coming = false;
6070 SecondaryFreeList_lock->notify_all();
6071 }
6073 if (G1ConcRegionFreeingVerbose) {
6074 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6075 "reset free regions coming");
6076 }
6077 }
6079 void G1CollectedHeap::wait_while_free_regions_coming() {
6080 // Most of the time we won't have to wait, so let's do a quick test
6081 // first before we take the lock.
6082 if (!free_regions_coming()) {
6083 return;
6084 }
6086 if (G1ConcRegionFreeingVerbose) {
6087 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6088 "waiting for free regions");
6089 }
6091 {
6092 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6093 while (free_regions_coming()) {
6094 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6095 }
6096 }
6098 if (G1ConcRegionFreeingVerbose) {
6099 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6100 "done waiting for free regions");
6101 }
6102 }
6104 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6105 assert(heap_lock_held_for_gc(),
6106 "the heap lock should already be held by or for this thread");
6107 _young_list->push_region(hr);
6108 }
6110 class NoYoungRegionsClosure: public HeapRegionClosure {
6111 private:
6112 bool _success;
6113 public:
6114 NoYoungRegionsClosure() : _success(true) { }
6115 bool doHeapRegion(HeapRegion* r) {
6116 if (r->is_young()) {
6117 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6118 r->bottom(), r->end());
6119 _success = false;
6120 }
6121 return false;
6122 }
6123 bool success() { return _success; }
6124 };
6126 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6127 bool ret = _young_list->check_list_empty(check_sample);
6129 if (check_heap) {
6130 NoYoungRegionsClosure closure;
6131 heap_region_iterate(&closure);
6132 ret = ret && closure.success();
6133 }
6135 return ret;
6136 }
6138 class TearDownRegionSetsClosure : public HeapRegionClosure {
6139 private:
6140 OldRegionSet *_old_set;
6142 public:
6143 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6145 bool doHeapRegion(HeapRegion* r) {
6146 if (r->is_empty()) {
6147 // We ignore empty regions, we'll empty the free list afterwards
6148 } else if (r->is_young()) {
6149 // We ignore young regions, we'll empty the young list afterwards
6150 } else if (r->isHumongous()) {
6151 // We ignore humongous regions, we're not tearing down the
6152 // humongous region set
6153 } else {
6154 // The rest should be old
6155 _old_set->remove(r);
6156 }
6157 return false;
6158 }
6160 ~TearDownRegionSetsClosure() {
6161 assert(_old_set->is_empty(), "post-condition");
6162 }
6163 };
6165 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6166 assert_at_safepoint(true /* should_be_vm_thread */);
6168 if (!free_list_only) {
6169 TearDownRegionSetsClosure cl(&_old_set);
6170 heap_region_iterate(&cl);
6172 // Need to do this after the heap iteration to be able to
6173 // recognize the young regions and ignore them during the iteration.
6174 _young_list->empty_list();
6175 }
6176 _free_list.remove_all();
6177 }
6179 class RebuildRegionSetsClosure : public HeapRegionClosure {
6180 private:
6181 bool _free_list_only;
6182 OldRegionSet* _old_set;
6183 FreeRegionList* _free_list;
6184 size_t _total_used;
6186 public:
6187 RebuildRegionSetsClosure(bool free_list_only,
6188 OldRegionSet* old_set, FreeRegionList* free_list) :
6189 _free_list_only(free_list_only),
6190 _old_set(old_set), _free_list(free_list), _total_used(0) {
6191 assert(_free_list->is_empty(), "pre-condition");
6192 if (!free_list_only) {
6193 assert(_old_set->is_empty(), "pre-condition");
6194 }
6195 }
6197 bool doHeapRegion(HeapRegion* r) {
6198 if (r->continuesHumongous()) {
6199 return false;
6200 }
6202 if (r->is_empty()) {
6203 // Add free regions to the free list
6204 _free_list->add_as_tail(r);
6205 } else if (!_free_list_only) {
6206 assert(!r->is_young(), "we should not come across young regions");
6208 if (r->isHumongous()) {
6209 // We ignore humongous regions, we left the humongous set unchanged
6210 } else {
6211 // The rest should be old, add them to the old set
6212 _old_set->add(r);
6213 }
6214 _total_used += r->used();
6215 }
6217 return false;
6218 }
6220 size_t total_used() {
6221 return _total_used;
6222 }
6223 };
6225 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6226 assert_at_safepoint(true /* should_be_vm_thread */);
6228 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6229 heap_region_iterate(&cl);
6231 if (!free_list_only) {
6232 _summary_bytes_used = cl.total_used();
6233 }
6234 assert(_summary_bytes_used == recalculate_used(),
6235 err_msg("inconsistent _summary_bytes_used, "
6236 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6237 _summary_bytes_used, recalculate_used()));
6238 }
6240 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6241 _refine_cte_cl->set_concurrent(concurrent);
6242 }
6244 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6245 HeapRegion* hr = heap_region_containing(p);
6246 if (hr == NULL) {
6247 return false;
6248 } else {
6249 return hr->is_in(p);
6250 }
6251 }
6253 // Methods for the mutator alloc region
6255 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6256 bool force) {
6257 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6258 assert(!force || g1_policy()->can_expand_young_list(),
6259 "if force is true we should be able to expand the young list");
6260 bool young_list_full = g1_policy()->is_young_list_full();
6261 if (force || !young_list_full) {
6262 HeapRegion* new_alloc_region = new_region(word_size,
6263 false /* do_expand */);
6264 if (new_alloc_region != NULL) {
6265 set_region_short_lived_locked(new_alloc_region);
6266 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6267 return new_alloc_region;
6268 }
6269 }
6270 return NULL;
6271 }
6273 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6274 size_t allocated_bytes) {
6275 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6276 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6278 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6279 _summary_bytes_used += allocated_bytes;
6280 _hr_printer.retire(alloc_region);
6281 // We update the eden sizes here, when the region is retired,
6282 // instead of when it's allocated, since this is the point that its
6283 // used space has been recored in _summary_bytes_used.
6284 g1mm()->update_eden_size();
6285 }
6287 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6288 bool force) {
6289 return _g1h->new_mutator_alloc_region(word_size, force);
6290 }
6292 void G1CollectedHeap::set_par_threads() {
6293 // Don't change the number of workers. Use the value previously set
6294 // in the workgroup.
6295 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6296 uint n_workers = workers()->active_workers();
6297 assert(UseDynamicNumberOfGCThreads ||
6298 n_workers == workers()->total_workers(),
6299 "Otherwise should be using the total number of workers");
6300 if (n_workers == 0) {
6301 assert(false, "Should have been set in prior evacuation pause.");
6302 n_workers = ParallelGCThreads;
6303 workers()->set_active_workers(n_workers);
6304 }
6305 set_par_threads(n_workers);
6306 }
6308 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6309 size_t allocated_bytes) {
6310 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6311 }
6313 // Methods for the GC alloc regions
6315 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6316 uint count,
6317 GCAllocPurpose ap) {
6318 assert(FreeList_lock->owned_by_self(), "pre-condition");
6320 if (count < g1_policy()->max_regions(ap)) {
6321 HeapRegion* new_alloc_region = new_region(word_size,
6322 true /* do_expand */);
6323 if (new_alloc_region != NULL) {
6324 // We really only need to do this for old regions given that we
6325 // should never scan survivors. But it doesn't hurt to do it
6326 // for survivors too.
6327 new_alloc_region->set_saved_mark();
6328 if (ap == GCAllocForSurvived) {
6329 new_alloc_region->set_survivor();
6330 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6331 } else {
6332 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6333 }
6334 bool during_im = g1_policy()->during_initial_mark_pause();
6335 new_alloc_region->note_start_of_copying(during_im);
6336 return new_alloc_region;
6337 } else {
6338 g1_policy()->note_alloc_region_limit_reached(ap);
6339 }
6340 }
6341 return NULL;
6342 }
6344 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6345 size_t allocated_bytes,
6346 GCAllocPurpose ap) {
6347 bool during_im = g1_policy()->during_initial_mark_pause();
6348 alloc_region->note_end_of_copying(during_im);
6349 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6350 if (ap == GCAllocForSurvived) {
6351 young_list()->add_survivor_region(alloc_region);
6352 } else {
6353 _old_set.add(alloc_region);
6354 }
6355 _hr_printer.retire(alloc_region);
6356 }
6358 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6359 bool force) {
6360 assert(!force, "not supported for GC alloc regions");
6361 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6362 }
6364 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6365 size_t allocated_bytes) {
6366 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6367 GCAllocForSurvived);
6368 }
6370 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6371 bool force) {
6372 assert(!force, "not supported for GC alloc regions");
6373 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6374 }
6376 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6377 size_t allocated_bytes) {
6378 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6379 GCAllocForTenured);
6380 }
6381 // Heap region set verification
6383 class VerifyRegionListsClosure : public HeapRegionClosure {
6384 private:
6385 FreeRegionList* _free_list;
6386 OldRegionSet* _old_set;
6387 HumongousRegionSet* _humongous_set;
6388 uint _region_count;
6390 public:
6391 VerifyRegionListsClosure(OldRegionSet* old_set,
6392 HumongousRegionSet* humongous_set,
6393 FreeRegionList* free_list) :
6394 _old_set(old_set), _humongous_set(humongous_set),
6395 _free_list(free_list), _region_count(0) { }
6397 uint region_count() { return _region_count; }
6399 bool doHeapRegion(HeapRegion* hr) {
6400 _region_count += 1;
6402 if (hr->continuesHumongous()) {
6403 return false;
6404 }
6406 if (hr->is_young()) {
6407 // TODO
6408 } else if (hr->startsHumongous()) {
6409 _humongous_set->verify_next_region(hr);
6410 } else if (hr->is_empty()) {
6411 _free_list->verify_next_region(hr);
6412 } else {
6413 _old_set->verify_next_region(hr);
6414 }
6415 return false;
6416 }
6417 };
6419 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6420 HeapWord* bottom) {
6421 HeapWord* end = bottom + HeapRegion::GrainWords;
6422 MemRegion mr(bottom, end);
6423 assert(_g1_reserved.contains(mr), "invariant");
6424 // This might return NULL if the allocation fails
6425 return new HeapRegion(hrs_index, _bot_shared, mr);
6426 }
6428 void G1CollectedHeap::verify_region_sets() {
6429 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6431 // First, check the explicit lists.
6432 _free_list.verify();
6433 {
6434 // Given that a concurrent operation might be adding regions to
6435 // the secondary free list we have to take the lock before
6436 // verifying it.
6437 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6438 _secondary_free_list.verify();
6439 }
6440 _old_set.verify();
6441 _humongous_set.verify();
6443 // If a concurrent region freeing operation is in progress it will
6444 // be difficult to correctly attributed any free regions we come
6445 // across to the correct free list given that they might belong to
6446 // one of several (free_list, secondary_free_list, any local lists,
6447 // etc.). So, if that's the case we will skip the rest of the
6448 // verification operation. Alternatively, waiting for the concurrent
6449 // operation to complete will have a non-trivial effect on the GC's
6450 // operation (no concurrent operation will last longer than the
6451 // interval between two calls to verification) and it might hide
6452 // any issues that we would like to catch during testing.
6453 if (free_regions_coming()) {
6454 return;
6455 }
6457 // Make sure we append the secondary_free_list on the free_list so
6458 // that all free regions we will come across can be safely
6459 // attributed to the free_list.
6460 append_secondary_free_list_if_not_empty_with_lock();
6462 // Finally, make sure that the region accounting in the lists is
6463 // consistent with what we see in the heap.
6464 _old_set.verify_start();
6465 _humongous_set.verify_start();
6466 _free_list.verify_start();
6468 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6469 heap_region_iterate(&cl);
6471 _old_set.verify_end();
6472 _humongous_set.verify_end();
6473 _free_list.verify_end();
6474 }