Wed, 02 Nov 2011 08:04:23 +0100
7106751: G1: gc/gctests/nativeGC03 crashes VM with SIGSEGV
Summary: _cset_rs_update_cl[] was indexed with values beyond what it is set up to handle.
Reviewed-by: ysr, jmasa, johnc
1 /*
2 * Copyright (c) 2001, 2011, 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/g1MarkSweep.hpp"
36 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
37 #include "gc_implementation/g1/g1RemSet.inline.hpp"
38 #include "gc_implementation/g1/heapRegionRemSet.hpp"
39 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
40 #include "gc_implementation/g1/vm_operations_g1.hpp"
41 #include "gc_implementation/shared/isGCActiveMark.hpp"
42 #include "memory/gcLocker.inline.hpp"
43 #include "memory/genOopClosures.inline.hpp"
44 #include "memory/generationSpec.hpp"
45 #include "memory/referenceProcessor.hpp"
46 #include "oops/oop.inline.hpp"
47 #include "oops/oop.pcgc.inline.hpp"
48 #include "runtime/aprofiler.hpp"
49 #include "runtime/vmThread.hpp"
51 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
53 // turn it on so that the contents of the young list (scan-only /
54 // to-be-collected) are printed at "strategic" points before / during
55 // / after the collection --- this is useful for debugging
56 #define YOUNG_LIST_VERBOSE 0
57 // CURRENT STATUS
58 // This file is under construction. Search for "FIXME".
60 // INVARIANTS/NOTES
61 //
62 // All allocation activity covered by the G1CollectedHeap interface is
63 // serialized by acquiring the HeapLock. This happens in mem_allocate
64 // and allocate_new_tlab, which are the "entry" points to the
65 // allocation code from the rest of the JVM. (Note that this does not
66 // apply to TLAB allocation, which is not part of this interface: it
67 // is done by clients of this interface.)
69 // Local to this file.
71 class RefineCardTableEntryClosure: public CardTableEntryClosure {
72 SuspendibleThreadSet* _sts;
73 G1RemSet* _g1rs;
74 ConcurrentG1Refine* _cg1r;
75 bool _concurrent;
76 public:
77 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
78 G1RemSet* g1rs,
79 ConcurrentG1Refine* cg1r) :
80 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
81 {}
82 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
83 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
84 // This path is executed by the concurrent refine or mutator threads,
85 // concurrently, and so we do not care if card_ptr contains references
86 // that point into the collection set.
87 assert(!oops_into_cset, "should be");
89 if (_concurrent && _sts->should_yield()) {
90 // Caller will actually yield.
91 return false;
92 }
93 // Otherwise, we finished successfully; return true.
94 return true;
95 }
96 void set_concurrent(bool b) { _concurrent = b; }
97 };
100 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
101 int _calls;
102 G1CollectedHeap* _g1h;
103 CardTableModRefBS* _ctbs;
104 int _histo[256];
105 public:
106 ClearLoggedCardTableEntryClosure() :
107 _calls(0)
108 {
109 _g1h = G1CollectedHeap::heap();
110 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
111 for (int i = 0; i < 256; i++) _histo[i] = 0;
112 }
113 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
114 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
115 _calls++;
116 unsigned char* ujb = (unsigned char*)card_ptr;
117 int ind = (int)(*ujb);
118 _histo[ind]++;
119 *card_ptr = -1;
120 }
121 return true;
122 }
123 int calls() { return _calls; }
124 void print_histo() {
125 gclog_or_tty->print_cr("Card table value histogram:");
126 for (int i = 0; i < 256; i++) {
127 if (_histo[i] != 0) {
128 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
129 }
130 }
131 }
132 };
134 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
135 int _calls;
136 G1CollectedHeap* _g1h;
137 CardTableModRefBS* _ctbs;
138 public:
139 RedirtyLoggedCardTableEntryClosure() :
140 _calls(0)
141 {
142 _g1h = G1CollectedHeap::heap();
143 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
144 }
145 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
146 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
147 _calls++;
148 *card_ptr = 0;
149 }
150 return true;
151 }
152 int calls() { return _calls; }
153 };
155 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
156 public:
157 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
158 *card_ptr = CardTableModRefBS::dirty_card_val();
159 return true;
160 }
161 };
163 YoungList::YoungList(G1CollectedHeap* g1h)
164 : _g1h(g1h), _head(NULL),
165 _length(0),
166 _last_sampled_rs_lengths(0),
167 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0)
168 {
169 guarantee( check_list_empty(false), "just making sure..." );
170 }
172 void YoungList::push_region(HeapRegion *hr) {
173 assert(!hr->is_young(), "should not already be young");
174 assert(hr->get_next_young_region() == NULL, "cause it should!");
176 hr->set_next_young_region(_head);
177 _head = hr;
179 hr->set_young();
180 double yg_surv_rate = _g1h->g1_policy()->predict_yg_surv_rate((int)_length);
181 ++_length;
182 }
184 void YoungList::add_survivor_region(HeapRegion* hr) {
185 assert(hr->is_survivor(), "should be flagged as survivor region");
186 assert(hr->get_next_young_region() == NULL, "cause it should!");
188 hr->set_next_young_region(_survivor_head);
189 if (_survivor_head == NULL) {
190 _survivor_tail = hr;
191 }
192 _survivor_head = hr;
194 ++_survivor_length;
195 }
197 void YoungList::empty_list(HeapRegion* list) {
198 while (list != NULL) {
199 HeapRegion* next = list->get_next_young_region();
200 list->set_next_young_region(NULL);
201 list->uninstall_surv_rate_group();
202 list->set_not_young();
203 list = next;
204 }
205 }
207 void YoungList::empty_list() {
208 assert(check_list_well_formed(), "young list should be well formed");
210 empty_list(_head);
211 _head = NULL;
212 _length = 0;
214 empty_list(_survivor_head);
215 _survivor_head = NULL;
216 _survivor_tail = NULL;
217 _survivor_length = 0;
219 _last_sampled_rs_lengths = 0;
221 assert(check_list_empty(false), "just making sure...");
222 }
224 bool YoungList::check_list_well_formed() {
225 bool ret = true;
227 size_t length = 0;
228 HeapRegion* curr = _head;
229 HeapRegion* last = NULL;
230 while (curr != NULL) {
231 if (!curr->is_young()) {
232 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
233 "incorrectly tagged (y: %d, surv: %d)",
234 curr->bottom(), curr->end(),
235 curr->is_young(), curr->is_survivor());
236 ret = false;
237 }
238 ++length;
239 last = curr;
240 curr = curr->get_next_young_region();
241 }
242 ret = ret && (length == _length);
244 if (!ret) {
245 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
246 gclog_or_tty->print_cr("### list has %d entries, _length is %d",
247 length, _length);
248 }
250 return ret;
251 }
253 bool YoungList::check_list_empty(bool check_sample) {
254 bool ret = true;
256 if (_length != 0) {
257 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %d",
258 _length);
259 ret = false;
260 }
261 if (check_sample && _last_sampled_rs_lengths != 0) {
262 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
263 ret = false;
264 }
265 if (_head != NULL) {
266 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
267 ret = false;
268 }
269 if (!ret) {
270 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
271 }
273 return ret;
274 }
276 void
277 YoungList::rs_length_sampling_init() {
278 _sampled_rs_lengths = 0;
279 _curr = _head;
280 }
282 bool
283 YoungList::rs_length_sampling_more() {
284 return _curr != NULL;
285 }
287 void
288 YoungList::rs_length_sampling_next() {
289 assert( _curr != NULL, "invariant" );
290 size_t rs_length = _curr->rem_set()->occupied();
292 _sampled_rs_lengths += rs_length;
294 // The current region may not yet have been added to the
295 // incremental collection set (it gets added when it is
296 // retired as the current allocation region).
297 if (_curr->in_collection_set()) {
298 // Update the collection set policy information for this region
299 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
300 }
302 _curr = _curr->get_next_young_region();
303 if (_curr == NULL) {
304 _last_sampled_rs_lengths = _sampled_rs_lengths;
305 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
306 }
307 }
309 void
310 YoungList::reset_auxilary_lists() {
311 guarantee( is_empty(), "young list should be empty" );
312 assert(check_list_well_formed(), "young list should be well formed");
314 // Add survivor regions to SurvRateGroup.
315 _g1h->g1_policy()->note_start_adding_survivor_regions();
316 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
318 for (HeapRegion* curr = _survivor_head;
319 curr != NULL;
320 curr = curr->get_next_young_region()) {
321 _g1h->g1_policy()->set_region_survivors(curr);
323 // The region is a non-empty survivor so let's add it to
324 // the incremental collection set for the next evacuation
325 // pause.
326 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
327 }
328 _g1h->g1_policy()->note_stop_adding_survivor_regions();
330 _head = _survivor_head;
331 _length = _survivor_length;
332 if (_survivor_head != NULL) {
333 assert(_survivor_tail != NULL, "cause it shouldn't be");
334 assert(_survivor_length > 0, "invariant");
335 _survivor_tail->set_next_young_region(NULL);
336 }
338 // Don't clear the survivor list handles until the start of
339 // the next evacuation pause - we need it in order to re-tag
340 // the survivor regions from this evacuation pause as 'young'
341 // at the start of the next.
343 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
345 assert(check_list_well_formed(), "young list should be well formed");
346 }
348 void YoungList::print() {
349 HeapRegion* lists[] = {_head, _survivor_head};
350 const char* names[] = {"YOUNG", "SURVIVOR"};
352 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
353 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
354 HeapRegion *curr = lists[list];
355 if (curr == NULL)
356 gclog_or_tty->print_cr(" empty");
357 while (curr != NULL) {
358 gclog_or_tty->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
359 "age: %4d, y: %d, surv: %d",
360 curr->bottom(), curr->end(),
361 curr->top(),
362 curr->prev_top_at_mark_start(),
363 curr->next_top_at_mark_start(),
364 curr->top_at_conc_mark_count(),
365 curr->age_in_surv_rate_group_cond(),
366 curr->is_young(),
367 curr->is_survivor());
368 curr = curr->get_next_young_region();
369 }
370 }
372 gclog_or_tty->print_cr("");
373 }
375 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
376 {
377 // Claim the right to put the region on the dirty cards region list
378 // by installing a self pointer.
379 HeapRegion* next = hr->get_next_dirty_cards_region();
380 if (next == NULL) {
381 HeapRegion* res = (HeapRegion*)
382 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
383 NULL);
384 if (res == NULL) {
385 HeapRegion* head;
386 do {
387 // Put the region to the dirty cards region list.
388 head = _dirty_cards_region_list;
389 next = (HeapRegion*)
390 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
391 if (next == head) {
392 assert(hr->get_next_dirty_cards_region() == hr,
393 "hr->get_next_dirty_cards_region() != hr");
394 if (next == NULL) {
395 // The last region in the list points to itself.
396 hr->set_next_dirty_cards_region(hr);
397 } else {
398 hr->set_next_dirty_cards_region(next);
399 }
400 }
401 } while (next != head);
402 }
403 }
404 }
406 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
407 {
408 HeapRegion* head;
409 HeapRegion* hr;
410 do {
411 head = _dirty_cards_region_list;
412 if (head == NULL) {
413 return NULL;
414 }
415 HeapRegion* new_head = head->get_next_dirty_cards_region();
416 if (head == new_head) {
417 // The last region.
418 new_head = NULL;
419 }
420 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
421 head);
422 } while (hr != head);
423 assert(hr != NULL, "invariant");
424 hr->set_next_dirty_cards_region(NULL);
425 return hr;
426 }
428 void G1CollectedHeap::stop_conc_gc_threads() {
429 _cg1r->stop();
430 _cmThread->stop();
431 }
433 #ifdef ASSERT
434 // A region is added to the collection set as it is retired
435 // so an address p can point to a region which will be in the
436 // collection set but has not yet been retired. This method
437 // therefore is only accurate during a GC pause after all
438 // regions have been retired. It is used for debugging
439 // to check if an nmethod has references to objects that can
440 // be move during a partial collection. Though it can be
441 // inaccurate, it is sufficient for G1 because the conservative
442 // implementation of is_scavengable() for G1 will indicate that
443 // all nmethods must be scanned during a partial collection.
444 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
445 HeapRegion* hr = heap_region_containing(p);
446 return hr != NULL && hr->in_collection_set();
447 }
448 #endif
450 // Returns true if the reference points to an object that
451 // can move in an incremental collecction.
452 bool G1CollectedHeap::is_scavengable(const void* p) {
453 G1CollectedHeap* g1h = G1CollectedHeap::heap();
454 G1CollectorPolicy* g1p = g1h->g1_policy();
455 HeapRegion* hr = heap_region_containing(p);
456 if (hr == NULL) {
457 // perm gen (or null)
458 return false;
459 } else {
460 return !hr->isHumongous();
461 }
462 }
464 void G1CollectedHeap::check_ct_logs_at_safepoint() {
465 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
466 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
468 // Count the dirty cards at the start.
469 CountNonCleanMemRegionClosure count1(this);
470 ct_bs->mod_card_iterate(&count1);
471 int orig_count = count1.n();
473 // First clear the logged cards.
474 ClearLoggedCardTableEntryClosure clear;
475 dcqs.set_closure(&clear);
476 dcqs.apply_closure_to_all_completed_buffers();
477 dcqs.iterate_closure_all_threads(false);
478 clear.print_histo();
480 // Now ensure that there's no dirty cards.
481 CountNonCleanMemRegionClosure count2(this);
482 ct_bs->mod_card_iterate(&count2);
483 if (count2.n() != 0) {
484 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
485 count2.n(), orig_count);
486 }
487 guarantee(count2.n() == 0, "Card table should be clean.");
489 RedirtyLoggedCardTableEntryClosure redirty;
490 JavaThread::dirty_card_queue_set().set_closure(&redirty);
491 dcqs.apply_closure_to_all_completed_buffers();
492 dcqs.iterate_closure_all_threads(false);
493 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
494 clear.calls(), orig_count);
495 guarantee(redirty.calls() == clear.calls(),
496 "Or else mechanism is broken.");
498 CountNonCleanMemRegionClosure count3(this);
499 ct_bs->mod_card_iterate(&count3);
500 if (count3.n() != orig_count) {
501 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
502 orig_count, count3.n());
503 guarantee(count3.n() >= orig_count, "Should have restored them all.");
504 }
506 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
507 }
509 // Private class members.
511 G1CollectedHeap* G1CollectedHeap::_g1h;
513 // Private methods.
515 HeapRegion*
516 G1CollectedHeap::new_region_try_secondary_free_list() {
517 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
518 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
519 if (!_secondary_free_list.is_empty()) {
520 if (G1ConcRegionFreeingVerbose) {
521 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
522 "secondary_free_list has "SIZE_FORMAT" entries",
523 _secondary_free_list.length());
524 }
525 // It looks as if there are free regions available on the
526 // secondary_free_list. Let's move them to the free_list and try
527 // again to allocate from it.
528 append_secondary_free_list();
530 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
531 "empty we should have moved at least one entry to the free_list");
532 HeapRegion* res = _free_list.remove_head();
533 if (G1ConcRegionFreeingVerbose) {
534 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
535 "allocated "HR_FORMAT" from secondary_free_list",
536 HR_FORMAT_PARAMS(res));
537 }
538 return res;
539 }
541 // Wait here until we get notifed either when (a) there are no
542 // more free regions coming or (b) some regions have been moved on
543 // the secondary_free_list.
544 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
545 }
547 if (G1ConcRegionFreeingVerbose) {
548 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
549 "could not allocate from secondary_free_list");
550 }
551 return NULL;
552 }
554 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
555 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
556 "the only time we use this to allocate a humongous region is "
557 "when we are allocating a single humongous region");
559 HeapRegion* res;
560 if (G1StressConcRegionFreeing) {
561 if (!_secondary_free_list.is_empty()) {
562 if (G1ConcRegionFreeingVerbose) {
563 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
564 "forced to look at the secondary_free_list");
565 }
566 res = new_region_try_secondary_free_list();
567 if (res != NULL) {
568 return res;
569 }
570 }
571 }
572 res = _free_list.remove_head_or_null();
573 if (res == NULL) {
574 if (G1ConcRegionFreeingVerbose) {
575 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
576 "res == NULL, trying the secondary_free_list");
577 }
578 res = new_region_try_secondary_free_list();
579 }
580 if (res == NULL && do_expand) {
581 ergo_verbose1(ErgoHeapSizing,
582 "attempt heap expansion",
583 ergo_format_reason("region allocation request failed")
584 ergo_format_byte("allocation request"),
585 word_size * HeapWordSize);
586 if (expand(word_size * HeapWordSize)) {
587 // Even though the heap was expanded, it might not have reached
588 // the desired size. So, we cannot assume that the allocation
589 // will succeed.
590 res = _free_list.remove_head_or_null();
591 }
592 }
593 return res;
594 }
596 size_t G1CollectedHeap::humongous_obj_allocate_find_first(size_t num_regions,
597 size_t word_size) {
598 assert(isHumongous(word_size), "word_size should be humongous");
599 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
601 size_t first = G1_NULL_HRS_INDEX;
602 if (num_regions == 1) {
603 // Only one region to allocate, no need to go through the slower
604 // path. The caller will attempt the expasion if this fails, so
605 // let's not try to expand here too.
606 HeapRegion* hr = new_region(word_size, false /* do_expand */);
607 if (hr != NULL) {
608 first = hr->hrs_index();
609 } else {
610 first = G1_NULL_HRS_INDEX;
611 }
612 } else {
613 // We can't allocate humongous regions while cleanupComplete() is
614 // running, since some of the regions we find to be empty might not
615 // yet be added to the free list and it is not straightforward to
616 // know which list they are on so that we can remove them. Note
617 // that we only need to do this if we need to allocate more than
618 // one region to satisfy the current humongous allocation
619 // request. If we are only allocating one region we use the common
620 // region allocation code (see above).
621 wait_while_free_regions_coming();
622 append_secondary_free_list_if_not_empty_with_lock();
624 if (free_regions() >= num_regions) {
625 first = _hrs.find_contiguous(num_regions);
626 if (first != G1_NULL_HRS_INDEX) {
627 for (size_t i = first; i < first + num_regions; ++i) {
628 HeapRegion* hr = region_at(i);
629 assert(hr->is_empty(), "sanity");
630 assert(is_on_master_free_list(hr), "sanity");
631 hr->set_pending_removal(true);
632 }
633 _free_list.remove_all_pending(num_regions);
634 }
635 }
636 }
637 return first;
638 }
640 HeapWord*
641 G1CollectedHeap::humongous_obj_allocate_initialize_regions(size_t first,
642 size_t num_regions,
643 size_t word_size) {
644 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
645 assert(isHumongous(word_size), "word_size should be humongous");
646 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
648 // Index of last region in the series + 1.
649 size_t last = first + num_regions;
651 // We need to initialize the region(s) we just discovered. This is
652 // a bit tricky given that it can happen concurrently with
653 // refinement threads refining cards on these regions and
654 // potentially wanting to refine the BOT as they are scanning
655 // those cards (this can happen shortly after a cleanup; see CR
656 // 6991377). So we have to set up the region(s) carefully and in
657 // a specific order.
659 // The word size sum of all the regions we will allocate.
660 size_t word_size_sum = num_regions * HeapRegion::GrainWords;
661 assert(word_size <= word_size_sum, "sanity");
663 // This will be the "starts humongous" region.
664 HeapRegion* first_hr = region_at(first);
665 // The header of the new object will be placed at the bottom of
666 // the first region.
667 HeapWord* new_obj = first_hr->bottom();
668 // This will be the new end of the first region in the series that
669 // should also match the end of the last region in the seriers.
670 HeapWord* new_end = new_obj + word_size_sum;
671 // This will be the new top of the first region that will reflect
672 // this allocation.
673 HeapWord* new_top = new_obj + word_size;
675 // First, we need to zero the header of the space that we will be
676 // allocating. When we update top further down, some refinement
677 // threads might try to scan the region. By zeroing the header we
678 // ensure that any thread that will try to scan the region will
679 // come across the zero klass word and bail out.
680 //
681 // NOTE: It would not have been correct to have used
682 // CollectedHeap::fill_with_object() and make the space look like
683 // an int array. The thread that is doing the allocation will
684 // later update the object header to a potentially different array
685 // type and, for a very short period of time, the klass and length
686 // fields will be inconsistent. This could cause a refinement
687 // thread to calculate the object size incorrectly.
688 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
690 // We will set up the first region as "starts humongous". This
691 // will also update the BOT covering all the regions to reflect
692 // that there is a single object that starts at the bottom of the
693 // first region.
694 first_hr->set_startsHumongous(new_top, new_end);
696 // Then, if there are any, we will set up the "continues
697 // humongous" regions.
698 HeapRegion* hr = NULL;
699 for (size_t i = first + 1; i < last; ++i) {
700 hr = region_at(i);
701 hr->set_continuesHumongous(first_hr);
702 }
703 // If we have "continues humongous" regions (hr != NULL), then the
704 // end of the last one should match new_end.
705 assert(hr == NULL || hr->end() == new_end, "sanity");
707 // Up to this point no concurrent thread would have been able to
708 // do any scanning on any region in this series. All the top
709 // fields still point to bottom, so the intersection between
710 // [bottom,top] and [card_start,card_end] will be empty. Before we
711 // update the top fields, we'll do a storestore to make sure that
712 // no thread sees the update to top before the zeroing of the
713 // object header and the BOT initialization.
714 OrderAccess::storestore();
716 // Now that the BOT and the object header have been initialized,
717 // we can update top of the "starts humongous" region.
718 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
719 "new_top should be in this region");
720 first_hr->set_top(new_top);
721 if (_hr_printer.is_active()) {
722 HeapWord* bottom = first_hr->bottom();
723 HeapWord* end = first_hr->orig_end();
724 if ((first + 1) == last) {
725 // the series has a single humongous region
726 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
727 } else {
728 // the series has more than one humongous regions
729 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
730 }
731 }
733 // Now, we will update the top fields of the "continues humongous"
734 // regions. The reason we need to do this is that, otherwise,
735 // these regions would look empty and this will confuse parts of
736 // G1. For example, the code that looks for a consecutive number
737 // of empty regions will consider them empty and try to
738 // re-allocate them. We can extend is_empty() to also include
739 // !continuesHumongous(), but it is easier to just update the top
740 // fields here. The way we set top for all regions (i.e., top ==
741 // end for all regions but the last one, top == new_top for the
742 // last one) is actually used when we will free up the humongous
743 // region in free_humongous_region().
744 hr = NULL;
745 for (size_t i = first + 1; i < last; ++i) {
746 hr = region_at(i);
747 if ((i + 1) == last) {
748 // last continues humongous region
749 assert(hr->bottom() < new_top && new_top <= hr->end(),
750 "new_top should fall on this region");
751 hr->set_top(new_top);
752 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
753 } else {
754 // not last one
755 assert(new_top > hr->end(), "new_top should be above this region");
756 hr->set_top(hr->end());
757 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
758 }
759 }
760 // If we have continues humongous regions (hr != NULL), then the
761 // end of the last one should match new_end and its top should
762 // match new_top.
763 assert(hr == NULL ||
764 (hr->end() == new_end && hr->top() == new_top), "sanity");
766 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
767 _summary_bytes_used += first_hr->used();
768 _humongous_set.add(first_hr);
770 return new_obj;
771 }
773 // If could fit into free regions w/o expansion, try.
774 // Otherwise, if can expand, do so.
775 // Otherwise, if using ex regions might help, try with ex given back.
776 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
777 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
779 verify_region_sets_optional();
781 size_t num_regions =
782 round_to(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords;
783 size_t x_size = expansion_regions();
784 size_t fs = _hrs.free_suffix();
785 size_t first = humongous_obj_allocate_find_first(num_regions, word_size);
786 if (first == G1_NULL_HRS_INDEX) {
787 // The only thing we can do now is attempt expansion.
788 if (fs + x_size >= num_regions) {
789 // If the number of regions we're trying to allocate for this
790 // object is at most the number of regions in the free suffix,
791 // then the call to humongous_obj_allocate_find_first() above
792 // should have succeeded and we wouldn't be here.
793 //
794 // We should only be trying to expand when the free suffix is
795 // not sufficient for the object _and_ we have some expansion
796 // room available.
797 assert(num_regions > fs, "earlier allocation should have succeeded");
799 ergo_verbose1(ErgoHeapSizing,
800 "attempt heap expansion",
801 ergo_format_reason("humongous allocation request failed")
802 ergo_format_byte("allocation request"),
803 word_size * HeapWordSize);
804 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
805 // Even though the heap was expanded, it might not have
806 // reached the desired size. So, we cannot assume that the
807 // allocation will succeed.
808 first = humongous_obj_allocate_find_first(num_regions, word_size);
809 }
810 }
811 }
813 HeapWord* result = NULL;
814 if (first != G1_NULL_HRS_INDEX) {
815 result =
816 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
817 assert(result != NULL, "it should always return a valid result");
819 // A successful humongous object allocation changes the used space
820 // information of the old generation so we need to recalculate the
821 // sizes and update the jstat counters here.
822 g1mm()->update_sizes();
823 }
825 verify_region_sets_optional();
827 return result;
828 }
830 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
831 assert_heap_not_locked_and_not_at_safepoint();
832 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
834 unsigned int dummy_gc_count_before;
835 return attempt_allocation(word_size, &dummy_gc_count_before);
836 }
838 HeapWord*
839 G1CollectedHeap::mem_allocate(size_t word_size,
840 bool* gc_overhead_limit_was_exceeded) {
841 assert_heap_not_locked_and_not_at_safepoint();
843 // Loop until the allocation is satisified, or unsatisfied after GC.
844 for (int try_count = 1; /* we'll return */; try_count += 1) {
845 unsigned int gc_count_before;
847 HeapWord* result = NULL;
848 if (!isHumongous(word_size)) {
849 result = attempt_allocation(word_size, &gc_count_before);
850 } else {
851 result = attempt_allocation_humongous(word_size, &gc_count_before);
852 }
853 if (result != NULL) {
854 return result;
855 }
857 // Create the garbage collection operation...
858 VM_G1CollectForAllocation op(gc_count_before, word_size);
859 // ...and get the VM thread to execute it.
860 VMThread::execute(&op);
862 if (op.prologue_succeeded() && op.pause_succeeded()) {
863 // If the operation was successful we'll return the result even
864 // if it is NULL. If the allocation attempt failed immediately
865 // after a Full GC, it's unlikely we'll be able to allocate now.
866 HeapWord* result = op.result();
867 if (result != NULL && !isHumongous(word_size)) {
868 // Allocations that take place on VM operations do not do any
869 // card dirtying and we have to do it here. We only have to do
870 // this for non-humongous allocations, though.
871 dirty_young_block(result, word_size);
872 }
873 return result;
874 } else {
875 assert(op.result() == NULL,
876 "the result should be NULL if the VM op did not succeed");
877 }
879 // Give a warning if we seem to be looping forever.
880 if ((QueuedAllocationWarningCount > 0) &&
881 (try_count % QueuedAllocationWarningCount == 0)) {
882 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
883 }
884 }
886 ShouldNotReachHere();
887 return NULL;
888 }
890 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
891 unsigned int *gc_count_before_ret) {
892 // Make sure you read the note in attempt_allocation_humongous().
894 assert_heap_not_locked_and_not_at_safepoint();
895 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
896 "be called for humongous allocation requests");
898 // We should only get here after the first-level allocation attempt
899 // (attempt_allocation()) failed to allocate.
901 // We will loop until a) we manage to successfully perform the
902 // allocation or b) we successfully schedule a collection which
903 // fails to perform the allocation. b) is the only case when we'll
904 // return NULL.
905 HeapWord* result = NULL;
906 for (int try_count = 1; /* we'll return */; try_count += 1) {
907 bool should_try_gc;
908 unsigned int gc_count_before;
910 {
911 MutexLockerEx x(Heap_lock);
913 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
914 false /* bot_updates */);
915 if (result != NULL) {
916 return result;
917 }
919 // If we reach here, attempt_allocation_locked() above failed to
920 // allocate a new region. So the mutator alloc region should be NULL.
921 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
923 if (GC_locker::is_active_and_needs_gc()) {
924 if (g1_policy()->can_expand_young_list()) {
925 // No need for an ergo verbose message here,
926 // can_expand_young_list() does this when it returns true.
927 result = _mutator_alloc_region.attempt_allocation_force(word_size,
928 false /* bot_updates */);
929 if (result != NULL) {
930 return result;
931 }
932 }
933 should_try_gc = false;
934 } else {
935 // Read the GC count while still holding the Heap_lock.
936 gc_count_before = SharedHeap::heap()->total_collections();
937 should_try_gc = true;
938 }
939 }
941 if (should_try_gc) {
942 bool succeeded;
943 result = do_collection_pause(word_size, gc_count_before, &succeeded);
944 if (result != NULL) {
945 assert(succeeded, "only way to get back a non-NULL result");
946 return result;
947 }
949 if (succeeded) {
950 // If we get here we successfully scheduled a collection which
951 // failed to allocate. No point in trying to allocate
952 // further. We'll just return NULL.
953 MutexLockerEx x(Heap_lock);
954 *gc_count_before_ret = SharedHeap::heap()->total_collections();
955 return NULL;
956 }
957 } else {
958 GC_locker::stall_until_clear();
959 }
961 // We can reach here if we were unsuccessul in scheduling a
962 // collection (because another thread beat us to it) or if we were
963 // stalled due to the GC locker. In either can we should retry the
964 // allocation attempt in case another thread successfully
965 // performed a collection and reclaimed enough space. We do the
966 // first attempt (without holding the Heap_lock) here and the
967 // follow-on attempt will be at the start of the next loop
968 // iteration (after taking the Heap_lock).
969 result = _mutator_alloc_region.attempt_allocation(word_size,
970 false /* bot_updates */);
971 if (result != NULL ){
972 return result;
973 }
975 // Give a warning if we seem to be looping forever.
976 if ((QueuedAllocationWarningCount > 0) &&
977 (try_count % QueuedAllocationWarningCount == 0)) {
978 warning("G1CollectedHeap::attempt_allocation_slow() "
979 "retries %d times", try_count);
980 }
981 }
983 ShouldNotReachHere();
984 return NULL;
985 }
987 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
988 unsigned int * gc_count_before_ret) {
989 // The structure of this method has a lot of similarities to
990 // attempt_allocation_slow(). The reason these two were not merged
991 // into a single one is that such a method would require several "if
992 // allocation is not humongous do this, otherwise do that"
993 // conditional paths which would obscure its flow. In fact, an early
994 // version of this code did use a unified method which was harder to
995 // follow and, as a result, it had subtle bugs that were hard to
996 // track down. So keeping these two methods separate allows each to
997 // be more readable. It will be good to keep these two in sync as
998 // much as possible.
1000 assert_heap_not_locked_and_not_at_safepoint();
1001 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1002 "should only be called for humongous allocations");
1004 // We will loop until a) we manage to successfully perform the
1005 // allocation or b) we successfully schedule a collection which
1006 // fails to perform the allocation. b) is the only case when we'll
1007 // return NULL.
1008 HeapWord* result = NULL;
1009 for (int try_count = 1; /* we'll return */; try_count += 1) {
1010 bool should_try_gc;
1011 unsigned int gc_count_before;
1013 {
1014 MutexLockerEx x(Heap_lock);
1016 // Given that humongous objects are not allocated in young
1017 // regions, we'll first try to do the allocation without doing a
1018 // collection hoping that there's enough space in the heap.
1019 result = humongous_obj_allocate(word_size);
1020 if (result != NULL) {
1021 return result;
1022 }
1024 if (GC_locker::is_active_and_needs_gc()) {
1025 should_try_gc = false;
1026 } else {
1027 // Read the GC count while still holding the Heap_lock.
1028 gc_count_before = SharedHeap::heap()->total_collections();
1029 should_try_gc = true;
1030 }
1031 }
1033 if (should_try_gc) {
1034 // If we failed to allocate the humongous object, we should try to
1035 // do a collection pause (if we're allowed) in case it reclaims
1036 // enough space for the allocation to succeed after the pause.
1038 bool succeeded;
1039 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1040 if (result != NULL) {
1041 assert(succeeded, "only way to get back a non-NULL result");
1042 return result;
1043 }
1045 if (succeeded) {
1046 // If we get here we successfully scheduled a collection which
1047 // failed to allocate. No point in trying to allocate
1048 // further. We'll just return NULL.
1049 MutexLockerEx x(Heap_lock);
1050 *gc_count_before_ret = SharedHeap::heap()->total_collections();
1051 return NULL;
1052 }
1053 } else {
1054 GC_locker::stall_until_clear();
1055 }
1057 // We can reach here if we were unsuccessul in scheduling a
1058 // collection (because another thread beat us to it) or if we were
1059 // stalled due to the GC locker. In either can we should retry the
1060 // allocation attempt in case another thread successfully
1061 // performed a collection and reclaimed enough space. Give a
1062 // warning if we seem to be looping forever.
1064 if ((QueuedAllocationWarningCount > 0) &&
1065 (try_count % QueuedAllocationWarningCount == 0)) {
1066 warning("G1CollectedHeap::attempt_allocation_humongous() "
1067 "retries %d times", try_count);
1068 }
1069 }
1071 ShouldNotReachHere();
1072 return NULL;
1073 }
1075 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1076 bool expect_null_mutator_alloc_region) {
1077 assert_at_safepoint(true /* should_be_vm_thread */);
1078 assert(_mutator_alloc_region.get() == NULL ||
1079 !expect_null_mutator_alloc_region,
1080 "the current alloc region was unexpectedly found to be non-NULL");
1082 if (!isHumongous(word_size)) {
1083 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1084 false /* bot_updates */);
1085 } else {
1086 return humongous_obj_allocate(word_size);
1087 }
1089 ShouldNotReachHere();
1090 }
1092 class PostMCRemSetClearClosure: public HeapRegionClosure {
1093 ModRefBarrierSet* _mr_bs;
1094 public:
1095 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1096 bool doHeapRegion(HeapRegion* r) {
1097 r->reset_gc_time_stamp();
1098 if (r->continuesHumongous())
1099 return false;
1100 HeapRegionRemSet* hrrs = r->rem_set();
1101 if (hrrs != NULL) hrrs->clear();
1102 // You might think here that we could clear just the cards
1103 // corresponding to the used region. But no: if we leave a dirty card
1104 // in a region we might allocate into, then it would prevent that card
1105 // from being enqueued, and cause it to be missed.
1106 // Re: the performance cost: we shouldn't be doing full GC anyway!
1107 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1108 return false;
1109 }
1110 };
1113 class PostMCRemSetInvalidateClosure: public HeapRegionClosure {
1114 ModRefBarrierSet* _mr_bs;
1115 public:
1116 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1117 bool doHeapRegion(HeapRegion* r) {
1118 if (r->continuesHumongous()) return false;
1119 if (r->used_region().word_size() != 0) {
1120 _mr_bs->invalidate(r->used_region(), true /*whole heap*/);
1121 }
1122 return false;
1123 }
1124 };
1126 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1127 G1CollectedHeap* _g1h;
1128 UpdateRSOopClosure _cl;
1129 int _worker_i;
1130 public:
1131 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1132 _cl(g1->g1_rem_set(), worker_i),
1133 _worker_i(worker_i),
1134 _g1h(g1)
1135 { }
1137 bool doHeapRegion(HeapRegion* r) {
1138 if (!r->continuesHumongous()) {
1139 _cl.set_from(r);
1140 r->oop_iterate(&_cl);
1141 }
1142 return false;
1143 }
1144 };
1146 class ParRebuildRSTask: public AbstractGangTask {
1147 G1CollectedHeap* _g1;
1148 public:
1149 ParRebuildRSTask(G1CollectedHeap* g1)
1150 : AbstractGangTask("ParRebuildRSTask"),
1151 _g1(g1)
1152 { }
1154 void work(int i) {
1155 RebuildRSOutOfRegionClosure rebuild_rs(_g1, i);
1156 _g1->heap_region_par_iterate_chunked(&rebuild_rs, i,
1157 HeapRegion::RebuildRSClaimValue);
1158 }
1159 };
1161 class PostCompactionPrinterClosure: public HeapRegionClosure {
1162 private:
1163 G1HRPrinter* _hr_printer;
1164 public:
1165 bool doHeapRegion(HeapRegion* hr) {
1166 assert(!hr->is_young(), "not expecting to find young regions");
1167 // We only generate output for non-empty regions.
1168 if (!hr->is_empty()) {
1169 if (!hr->isHumongous()) {
1170 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1171 } else if (hr->startsHumongous()) {
1172 if (hr->capacity() == HeapRegion::GrainBytes) {
1173 // single humongous region
1174 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1175 } else {
1176 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1177 }
1178 } else {
1179 assert(hr->continuesHumongous(), "only way to get here");
1180 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1181 }
1182 }
1183 return false;
1184 }
1186 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1187 : _hr_printer(hr_printer) { }
1188 };
1190 bool G1CollectedHeap::do_collection(bool explicit_gc,
1191 bool clear_all_soft_refs,
1192 size_t word_size) {
1193 assert_at_safepoint(true /* should_be_vm_thread */);
1195 if (GC_locker::check_active_before_gc()) {
1196 return false;
1197 }
1199 SvcGCMarker sgcm(SvcGCMarker::FULL);
1200 ResourceMark rm;
1202 if (PrintHeapAtGC) {
1203 Universe::print_heap_before_gc();
1204 }
1206 verify_region_sets_optional();
1208 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1209 collector_policy()->should_clear_all_soft_refs();
1211 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1213 {
1214 IsGCActiveMark x;
1216 // Timing
1217 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc);
1218 assert(!system_gc || explicit_gc, "invariant");
1219 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1220 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1221 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC",
1222 PrintGC, true, gclog_or_tty);
1224 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1225 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1227 double start = os::elapsedTime();
1228 g1_policy()->record_full_collection_start();
1230 wait_while_free_regions_coming();
1231 append_secondary_free_list_if_not_empty_with_lock();
1233 gc_prologue(true);
1234 increment_total_collections(true /* full gc */);
1236 size_t g1h_prev_used = used();
1237 assert(used() == recalculate_used(), "Should be equal");
1239 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
1240 HandleMark hm; // Discard invalid handles created during verification
1241 gclog_or_tty->print(" VerifyBeforeGC:");
1242 prepare_for_verify();
1243 Universe::verify(/* allow dirty */ true,
1244 /* silent */ false,
1245 /* option */ VerifyOption_G1UsePrevMarking);
1247 }
1248 pre_full_gc_dump();
1250 COMPILER2_PRESENT(DerivedPointerTable::clear());
1252 // Disable discovery and empty the discovered lists
1253 // for the CM ref processor.
1254 ref_processor_cm()->disable_discovery();
1255 ref_processor_cm()->abandon_partial_discovery();
1256 ref_processor_cm()->verify_no_references_recorded();
1258 // Abandon current iterations of concurrent marking and concurrent
1259 // refinement, if any are in progress.
1260 concurrent_mark()->abort();
1262 // Make sure we'll choose a new allocation region afterwards.
1263 release_mutator_alloc_region();
1264 abandon_gc_alloc_regions();
1265 g1_rem_set()->cleanupHRRS();
1266 tear_down_region_lists();
1268 // We should call this after we retire any currently active alloc
1269 // regions so that all the ALLOC / RETIRE events are generated
1270 // before the start GC event.
1271 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1273 // We may have added regions to the current incremental collection
1274 // set between the last GC or pause and now. We need to clear the
1275 // incremental collection set and then start rebuilding it afresh
1276 // after this full GC.
1277 abandon_collection_set(g1_policy()->inc_cset_head());
1278 g1_policy()->clear_incremental_cset();
1279 g1_policy()->stop_incremental_cset_building();
1281 empty_young_list();
1282 g1_policy()->set_full_young_gcs(true);
1284 // See the comments in g1CollectedHeap.hpp and
1285 // G1CollectedHeap::ref_processing_init() about
1286 // how reference processing currently works in G1.
1288 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1289 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1291 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1292 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1294 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1295 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1297 // Do collection work
1298 {
1299 HandleMark hm; // Discard invalid handles created during gc
1300 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1301 }
1303 assert(free_regions() == 0, "we should not have added any free regions");
1304 rebuild_region_lists();
1306 _summary_bytes_used = recalculate_used();
1308 // Enqueue any discovered reference objects that have
1309 // not been removed from the discovered lists.
1310 ref_processor_stw()->enqueue_discovered_references();
1312 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1314 MemoryService::track_memory_usage();
1316 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
1317 HandleMark hm; // Discard invalid handles created during verification
1318 gclog_or_tty->print(" VerifyAfterGC:");
1319 prepare_for_verify();
1320 Universe::verify(/* allow dirty */ false,
1321 /* silent */ false,
1322 /* option */ VerifyOption_G1UsePrevMarking);
1324 }
1326 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1327 ref_processor_stw()->verify_no_references_recorded();
1329 // Note: since we've just done a full GC, concurrent
1330 // marking is no longer active. Therefore we need not
1331 // re-enable reference discovery for the CM ref processor.
1332 // That will be done at the start of the next marking cycle.
1333 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1334 ref_processor_cm()->verify_no_references_recorded();
1336 reset_gc_time_stamp();
1337 // Since everything potentially moved, we will clear all remembered
1338 // sets, and clear all cards. Later we will rebuild remebered
1339 // sets. We will also reset the GC time stamps of the regions.
1340 PostMCRemSetClearClosure rs_clear(mr_bs());
1341 heap_region_iterate(&rs_clear);
1343 // Resize the heap if necessary.
1344 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1346 if (_hr_printer.is_active()) {
1347 // We should do this after we potentially resize the heap so
1348 // that all the COMMIT / UNCOMMIT events are generated before
1349 // the end GC event.
1351 PostCompactionPrinterClosure cl(hr_printer());
1352 heap_region_iterate(&cl);
1354 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1355 }
1357 if (_cg1r->use_cache()) {
1358 _cg1r->clear_and_record_card_counts();
1359 _cg1r->clear_hot_cache();
1360 }
1362 // Rebuild remembered sets of all regions.
1364 if (G1CollectedHeap::use_parallel_gc_threads()) {
1365 ParRebuildRSTask rebuild_rs_task(this);
1366 assert(check_heap_region_claim_values(
1367 HeapRegion::InitialClaimValue), "sanity check");
1368 set_par_threads(workers()->total_workers());
1369 workers()->run_task(&rebuild_rs_task);
1370 set_par_threads(0);
1371 assert(check_heap_region_claim_values(
1372 HeapRegion::RebuildRSClaimValue), "sanity check");
1373 reset_heap_region_claim_values();
1374 } else {
1375 RebuildRSOutOfRegionClosure rebuild_rs(this);
1376 heap_region_iterate(&rebuild_rs);
1377 }
1379 if (PrintGC) {
1380 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1381 }
1383 if (true) { // FIXME
1384 // Ask the permanent generation to adjust size for full collections
1385 perm()->compute_new_size();
1386 }
1388 // Start a new incremental collection set for the next pause
1389 assert(g1_policy()->collection_set() == NULL, "must be");
1390 g1_policy()->start_incremental_cset_building();
1392 // Clear the _cset_fast_test bitmap in anticipation of adding
1393 // regions to the incremental collection set for the next
1394 // evacuation pause.
1395 clear_cset_fast_test();
1397 init_mutator_alloc_region();
1399 double end = os::elapsedTime();
1400 g1_policy()->record_full_collection_end();
1402 #ifdef TRACESPINNING
1403 ParallelTaskTerminator::print_termination_counts();
1404 #endif
1406 gc_epilogue(true);
1408 // Discard all rset updates
1409 JavaThread::dirty_card_queue_set().abandon_logs();
1410 assert(!G1DeferredRSUpdate
1411 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1412 }
1414 _young_list->reset_sampled_info();
1415 // At this point there should be no regions in the
1416 // entire heap tagged as young.
1417 assert( check_young_list_empty(true /* check_heap */),
1418 "young list should be empty at this point");
1420 // Update the number of full collections that have been completed.
1421 increment_full_collections_completed(false /* concurrent */);
1423 _hrs.verify_optional();
1424 verify_region_sets_optional();
1426 if (PrintHeapAtGC) {
1427 Universe::print_heap_after_gc();
1428 }
1429 g1mm()->update_sizes();
1430 post_full_gc_dump();
1432 return true;
1433 }
1435 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1436 // do_collection() will return whether it succeeded in performing
1437 // the GC. Currently, there is no facility on the
1438 // do_full_collection() API to notify the caller than the collection
1439 // did not succeed (e.g., because it was locked out by the GC
1440 // locker). So, right now, we'll ignore the return value.
1441 bool dummy = do_collection(true, /* explicit_gc */
1442 clear_all_soft_refs,
1443 0 /* word_size */);
1444 }
1446 // This code is mostly copied from TenuredGeneration.
1447 void
1448 G1CollectedHeap::
1449 resize_if_necessary_after_full_collection(size_t word_size) {
1450 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1452 // Include the current allocation, if any, and bytes that will be
1453 // pre-allocated to support collections, as "used".
1454 const size_t used_after_gc = used();
1455 const size_t capacity_after_gc = capacity();
1456 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1458 // This is enforced in arguments.cpp.
1459 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1460 "otherwise the code below doesn't make sense");
1462 // We don't have floating point command-line arguments
1463 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1464 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1465 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1466 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1468 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1469 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1471 // We have to be careful here as these two calculations can overflow
1472 // 32-bit size_t's.
1473 double used_after_gc_d = (double) used_after_gc;
1474 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1475 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1477 // Let's make sure that they are both under the max heap size, which
1478 // by default will make them fit into a size_t.
1479 double desired_capacity_upper_bound = (double) max_heap_size;
1480 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1481 desired_capacity_upper_bound);
1482 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1483 desired_capacity_upper_bound);
1485 // We can now safely turn them into size_t's.
1486 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1487 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1489 // This assert only makes sense here, before we adjust them
1490 // with respect to the min and max heap size.
1491 assert(minimum_desired_capacity <= maximum_desired_capacity,
1492 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1493 "maximum_desired_capacity = "SIZE_FORMAT,
1494 minimum_desired_capacity, maximum_desired_capacity));
1496 // Should not be greater than the heap max size. No need to adjust
1497 // it with respect to the heap min size as it's a lower bound (i.e.,
1498 // we'll try to make the capacity larger than it, not smaller).
1499 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1500 // Should not be less than the heap min size. No need to adjust it
1501 // with respect to the heap max size as it's an upper bound (i.e.,
1502 // we'll try to make the capacity smaller than it, not greater).
1503 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1505 if (capacity_after_gc < minimum_desired_capacity) {
1506 // Don't expand unless it's significant
1507 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1508 ergo_verbose4(ErgoHeapSizing,
1509 "attempt heap expansion",
1510 ergo_format_reason("capacity lower than "
1511 "min desired capacity after Full GC")
1512 ergo_format_byte("capacity")
1513 ergo_format_byte("occupancy")
1514 ergo_format_byte_perc("min desired capacity"),
1515 capacity_after_gc, used_after_gc,
1516 minimum_desired_capacity, (double) MinHeapFreeRatio);
1517 expand(expand_bytes);
1519 // No expansion, now see if we want to shrink
1520 } else if (capacity_after_gc > maximum_desired_capacity) {
1521 // Capacity too large, compute shrinking size
1522 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1523 ergo_verbose4(ErgoHeapSizing,
1524 "attempt heap shrinking",
1525 ergo_format_reason("capacity higher than "
1526 "max desired capacity after Full GC")
1527 ergo_format_byte("capacity")
1528 ergo_format_byte("occupancy")
1529 ergo_format_byte_perc("max desired capacity"),
1530 capacity_after_gc, used_after_gc,
1531 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1532 shrink(shrink_bytes);
1533 }
1534 }
1537 HeapWord*
1538 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1539 bool* succeeded) {
1540 assert_at_safepoint(true /* should_be_vm_thread */);
1542 *succeeded = true;
1543 // Let's attempt the allocation first.
1544 HeapWord* result =
1545 attempt_allocation_at_safepoint(word_size,
1546 false /* expect_null_mutator_alloc_region */);
1547 if (result != NULL) {
1548 assert(*succeeded, "sanity");
1549 return result;
1550 }
1552 // In a G1 heap, we're supposed to keep allocation from failing by
1553 // incremental pauses. Therefore, at least for now, we'll favor
1554 // expansion over collection. (This might change in the future if we can
1555 // do something smarter than full collection to satisfy a failed alloc.)
1556 result = expand_and_allocate(word_size);
1557 if (result != NULL) {
1558 assert(*succeeded, "sanity");
1559 return result;
1560 }
1562 // Expansion didn't work, we'll try to do a Full GC.
1563 bool gc_succeeded = do_collection(false, /* explicit_gc */
1564 false, /* clear_all_soft_refs */
1565 word_size);
1566 if (!gc_succeeded) {
1567 *succeeded = false;
1568 return NULL;
1569 }
1571 // Retry the allocation
1572 result = attempt_allocation_at_safepoint(word_size,
1573 true /* expect_null_mutator_alloc_region */);
1574 if (result != NULL) {
1575 assert(*succeeded, "sanity");
1576 return result;
1577 }
1579 // Then, try a Full GC that will collect all soft references.
1580 gc_succeeded = do_collection(false, /* explicit_gc */
1581 true, /* clear_all_soft_refs */
1582 word_size);
1583 if (!gc_succeeded) {
1584 *succeeded = false;
1585 return NULL;
1586 }
1588 // Retry the allocation once more
1589 result = attempt_allocation_at_safepoint(word_size,
1590 true /* expect_null_mutator_alloc_region */);
1591 if (result != NULL) {
1592 assert(*succeeded, "sanity");
1593 return result;
1594 }
1596 assert(!collector_policy()->should_clear_all_soft_refs(),
1597 "Flag should have been handled and cleared prior to this point");
1599 // What else? We might try synchronous finalization later. If the total
1600 // space available is large enough for the allocation, then a more
1601 // complete compaction phase than we've tried so far might be
1602 // appropriate.
1603 assert(*succeeded, "sanity");
1604 return NULL;
1605 }
1607 // Attempting to expand the heap sufficiently
1608 // to support an allocation of the given "word_size". If
1609 // successful, perform the allocation and return the address of the
1610 // allocated block, or else "NULL".
1612 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1613 assert_at_safepoint(true /* should_be_vm_thread */);
1615 verify_region_sets_optional();
1617 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1618 ergo_verbose1(ErgoHeapSizing,
1619 "attempt heap expansion",
1620 ergo_format_reason("allocation request failed")
1621 ergo_format_byte("allocation request"),
1622 word_size * HeapWordSize);
1623 if (expand(expand_bytes)) {
1624 _hrs.verify_optional();
1625 verify_region_sets_optional();
1626 return attempt_allocation_at_safepoint(word_size,
1627 false /* expect_null_mutator_alloc_region */);
1628 }
1629 return NULL;
1630 }
1632 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1633 HeapWord* new_end) {
1634 assert(old_end != new_end, "don't call this otherwise");
1635 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1637 // Update the committed mem region.
1638 _g1_committed.set_end(new_end);
1639 // Tell the card table about the update.
1640 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1641 // Tell the BOT about the update.
1642 _bot_shared->resize(_g1_committed.word_size());
1643 }
1645 bool G1CollectedHeap::expand(size_t expand_bytes) {
1646 size_t old_mem_size = _g1_storage.committed_size();
1647 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1648 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1649 HeapRegion::GrainBytes);
1650 ergo_verbose2(ErgoHeapSizing,
1651 "expand the heap",
1652 ergo_format_byte("requested expansion amount")
1653 ergo_format_byte("attempted expansion amount"),
1654 expand_bytes, aligned_expand_bytes);
1656 // First commit the memory.
1657 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1658 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1659 if (successful) {
1660 // Then propagate this update to the necessary data structures.
1661 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1662 update_committed_space(old_end, new_end);
1664 FreeRegionList expansion_list("Local Expansion List");
1665 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1666 assert(mr.start() == old_end, "post-condition");
1667 // mr might be a smaller region than what was requested if
1668 // expand_by() was unable to allocate the HeapRegion instances
1669 assert(mr.end() <= new_end, "post-condition");
1671 size_t actual_expand_bytes = mr.byte_size();
1672 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1673 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1674 "post-condition");
1675 if (actual_expand_bytes < aligned_expand_bytes) {
1676 // We could not expand _hrs to the desired size. In this case we
1677 // need to shrink the committed space accordingly.
1678 assert(mr.end() < new_end, "invariant");
1680 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1681 // First uncommit the memory.
1682 _g1_storage.shrink_by(diff_bytes);
1683 // Then propagate this update to the necessary data structures.
1684 update_committed_space(new_end, mr.end());
1685 }
1686 _free_list.add_as_tail(&expansion_list);
1688 if (_hr_printer.is_active()) {
1689 HeapWord* curr = mr.start();
1690 while (curr < mr.end()) {
1691 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1692 _hr_printer.commit(curr, curr_end);
1693 curr = curr_end;
1694 }
1695 assert(curr == mr.end(), "post-condition");
1696 }
1697 g1_policy()->record_new_heap_size(n_regions());
1698 } else {
1699 ergo_verbose0(ErgoHeapSizing,
1700 "did not expand the heap",
1701 ergo_format_reason("heap expansion operation failed"));
1702 // The expansion of the virtual storage space was unsuccessful.
1703 // Let's see if it was because we ran out of swap.
1704 if (G1ExitOnExpansionFailure &&
1705 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1706 // We had head room...
1707 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1708 }
1709 }
1710 return successful;
1711 }
1713 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1714 size_t old_mem_size = _g1_storage.committed_size();
1715 size_t aligned_shrink_bytes =
1716 ReservedSpace::page_align_size_down(shrink_bytes);
1717 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1718 HeapRegion::GrainBytes);
1719 size_t num_regions_deleted = 0;
1720 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1721 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1722 assert(mr.end() == old_end, "post-condition");
1724 ergo_verbose3(ErgoHeapSizing,
1725 "shrink the heap",
1726 ergo_format_byte("requested shrinking amount")
1727 ergo_format_byte("aligned shrinking amount")
1728 ergo_format_byte("attempted shrinking amount"),
1729 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1730 if (mr.byte_size() > 0) {
1731 if (_hr_printer.is_active()) {
1732 HeapWord* curr = mr.end();
1733 while (curr > mr.start()) {
1734 HeapWord* curr_end = curr;
1735 curr -= HeapRegion::GrainWords;
1736 _hr_printer.uncommit(curr, curr_end);
1737 }
1738 assert(curr == mr.start(), "post-condition");
1739 }
1741 _g1_storage.shrink_by(mr.byte_size());
1742 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1743 assert(mr.start() == new_end, "post-condition");
1745 _expansion_regions += num_regions_deleted;
1746 update_committed_space(old_end, new_end);
1747 HeapRegionRemSet::shrink_heap(n_regions());
1748 g1_policy()->record_new_heap_size(n_regions());
1749 } else {
1750 ergo_verbose0(ErgoHeapSizing,
1751 "did not shrink the heap",
1752 ergo_format_reason("heap shrinking operation failed"));
1753 }
1754 }
1756 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1757 verify_region_sets_optional();
1759 // We should only reach here at the end of a Full GC which means we
1760 // should not not be holding to any GC alloc regions. The method
1761 // below will make sure of that and do any remaining clean up.
1762 abandon_gc_alloc_regions();
1764 // Instead of tearing down / rebuilding the free lists here, we
1765 // could instead use the remove_all_pending() method on free_list to
1766 // remove only the ones that we need to remove.
1767 tear_down_region_lists(); // We will rebuild them in a moment.
1768 shrink_helper(shrink_bytes);
1769 rebuild_region_lists();
1771 _hrs.verify_optional();
1772 verify_region_sets_optional();
1773 }
1775 // Public methods.
1777 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1778 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1779 #endif // _MSC_VER
1782 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1783 SharedHeap(policy_),
1784 _g1_policy(policy_),
1785 _dirty_card_queue_set(false),
1786 _into_cset_dirty_card_queue_set(false),
1787 _is_alive_closure_cm(this),
1788 _is_alive_closure_stw(this),
1789 _ref_processor_cm(NULL),
1790 _ref_processor_stw(NULL),
1791 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1792 _bot_shared(NULL),
1793 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1794 _evac_failure_scan_stack(NULL) ,
1795 _mark_in_progress(false),
1796 _cg1r(NULL), _summary_bytes_used(0),
1797 _g1mm(NULL),
1798 _refine_cte_cl(NULL),
1799 _full_collection(false),
1800 _free_list("Master Free List"),
1801 _secondary_free_list("Secondary Free List"),
1802 _humongous_set("Master Humongous Set"),
1803 _free_regions_coming(false),
1804 _young_list(new YoungList(this)),
1805 _gc_time_stamp(0),
1806 _retained_old_gc_alloc_region(NULL),
1807 _surviving_young_words(NULL),
1808 _full_collections_completed(0),
1809 _in_cset_fast_test(NULL),
1810 _in_cset_fast_test_base(NULL),
1811 _dirty_cards_region_list(NULL) {
1812 _g1h = this; // To catch bugs.
1813 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1814 vm_exit_during_initialization("Failed necessary allocation.");
1815 }
1817 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1819 int n_queues = MAX2((int)ParallelGCThreads, 1);
1820 _task_queues = new RefToScanQueueSet(n_queues);
1822 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1823 assert(n_rem_sets > 0, "Invariant.");
1825 HeapRegionRemSetIterator** iter_arr =
1826 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues);
1827 for (int i = 0; i < n_queues; i++) {
1828 iter_arr[i] = new HeapRegionRemSetIterator();
1829 }
1830 _rem_set_iterator = iter_arr;
1832 for (int i = 0; i < n_queues; i++) {
1833 RefToScanQueue* q = new RefToScanQueue();
1834 q->initialize();
1835 _task_queues->register_queue(i, q);
1836 }
1838 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1839 }
1841 jint G1CollectedHeap::initialize() {
1842 CollectedHeap::pre_initialize();
1843 os::enable_vtime();
1845 // Necessary to satisfy locking discipline assertions.
1847 MutexLocker x(Heap_lock);
1849 // We have to initialize the printer before committing the heap, as
1850 // it will be used then.
1851 _hr_printer.set_active(G1PrintHeapRegions);
1853 // While there are no constraints in the GC code that HeapWordSize
1854 // be any particular value, there are multiple other areas in the
1855 // system which believe this to be true (e.g. oop->object_size in some
1856 // cases incorrectly returns the size in wordSize units rather than
1857 // HeapWordSize).
1858 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1860 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1861 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1863 // Ensure that the sizes are properly aligned.
1864 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1865 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1867 _cg1r = new ConcurrentG1Refine();
1869 // Reserve the maximum.
1870 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation();
1871 // Includes the perm-gen.
1873 // When compressed oops are enabled, the preferred heap base
1874 // is calculated by subtracting the requested size from the
1875 // 32Gb boundary and using the result as the base address for
1876 // heap reservation. If the requested size is not aligned to
1877 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1878 // into the ReservedHeapSpace constructor) then the actual
1879 // base of the reserved heap may end up differing from the
1880 // address that was requested (i.e. the preferred heap base).
1881 // If this happens then we could end up using a non-optimal
1882 // compressed oops mode.
1884 // Since max_byte_size is aligned to the size of a heap region (checked
1885 // above), we also need to align the perm gen size as it might not be.
1886 const size_t total_reserved = max_byte_size +
1887 align_size_up(pgs->max_size(), HeapRegion::GrainBytes);
1888 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm");
1890 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop);
1892 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes,
1893 UseLargePages, addr);
1895 if (UseCompressedOops) {
1896 if (addr != NULL && !heap_rs.is_reserved()) {
1897 // Failed to reserve at specified address - the requested memory
1898 // region is taken already, for example, by 'java' launcher.
1899 // Try again to reserver heap higher.
1900 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop);
1902 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes,
1903 UseLargePages, addr);
1905 if (addr != NULL && !heap_rs0.is_reserved()) {
1906 // Failed to reserve at specified address again - give up.
1907 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop);
1908 assert(addr == NULL, "");
1910 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes,
1911 UseLargePages, addr);
1912 heap_rs = heap_rs1;
1913 } else {
1914 heap_rs = heap_rs0;
1915 }
1916 }
1917 }
1919 if (!heap_rs.is_reserved()) {
1920 vm_exit_during_initialization("Could not reserve enough space for object heap");
1921 return JNI_ENOMEM;
1922 }
1924 // It is important to do this in a way such that concurrent readers can't
1925 // temporarily think somethings in the heap. (I've actually seen this
1926 // happen in asserts: DLD.)
1927 _reserved.set_word_size(0);
1928 _reserved.set_start((HeapWord*)heap_rs.base());
1929 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
1931 _expansion_regions = max_byte_size/HeapRegion::GrainBytes;
1933 // Create the gen rem set (and barrier set) for the entire reserved region.
1934 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
1935 set_barrier_set(rem_set()->bs());
1936 if (barrier_set()->is_a(BarrierSet::ModRef)) {
1937 _mr_bs = (ModRefBarrierSet*)_barrier_set;
1938 } else {
1939 vm_exit_during_initialization("G1 requires a mod ref bs.");
1940 return JNI_ENOMEM;
1941 }
1943 // Also create a G1 rem set.
1944 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
1945 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
1946 } else {
1947 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
1948 return JNI_ENOMEM;
1949 }
1951 // Carve out the G1 part of the heap.
1953 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
1954 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
1955 g1_rs.size()/HeapWordSize);
1956 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size);
1958 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set());
1960 _g1_storage.initialize(g1_rs, 0);
1961 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
1962 _hrs.initialize((HeapWord*) _g1_reserved.start(),
1963 (HeapWord*) _g1_reserved.end(),
1964 _expansion_regions);
1966 // 6843694 - ensure that the maximum region index can fit
1967 // in the remembered set structures.
1968 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
1969 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
1971 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
1972 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
1973 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
1974 "too many cards per region");
1976 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
1978 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
1979 heap_word_size(init_byte_size));
1981 _g1h = this;
1983 _in_cset_fast_test_length = max_regions();
1984 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length);
1986 // We're biasing _in_cset_fast_test to avoid subtracting the
1987 // beginning of the heap every time we want to index; basically
1988 // it's the same with what we do with the card table.
1989 _in_cset_fast_test = _in_cset_fast_test_base -
1990 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
1992 // Clear the _cset_fast_test bitmap in anticipation of adding
1993 // regions to the incremental collection set for the first
1994 // evacuation pause.
1995 clear_cset_fast_test();
1997 // Create the ConcurrentMark data structure and thread.
1998 // (Must do this late, so that "max_regions" is defined.)
1999 _cm = new ConcurrentMark(heap_rs, (int) max_regions());
2000 _cmThread = _cm->cmThread();
2002 // Initialize the from_card cache structure of HeapRegionRemSet.
2003 HeapRegionRemSet::init_heap(max_regions());
2005 // Now expand into the initial heap size.
2006 if (!expand(init_byte_size)) {
2007 vm_exit_during_initialization("Failed to allocate initial heap.");
2008 return JNI_ENOMEM;
2009 }
2011 // Perform any initialization actions delegated to the policy.
2012 g1_policy()->init();
2014 _refine_cte_cl =
2015 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2016 g1_rem_set(),
2017 concurrent_g1_refine());
2018 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2020 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2021 SATB_Q_FL_lock,
2022 G1SATBProcessCompletedThreshold,
2023 Shared_SATB_Q_lock);
2025 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2026 DirtyCardQ_FL_lock,
2027 concurrent_g1_refine()->yellow_zone(),
2028 concurrent_g1_refine()->red_zone(),
2029 Shared_DirtyCardQ_lock);
2031 if (G1DeferredRSUpdate) {
2032 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2033 DirtyCardQ_FL_lock,
2034 -1, // never trigger processing
2035 -1, // no limit on length
2036 Shared_DirtyCardQ_lock,
2037 &JavaThread::dirty_card_queue_set());
2038 }
2040 // Initialize the card queue set used to hold cards containing
2041 // references into the collection set.
2042 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2043 DirtyCardQ_FL_lock,
2044 -1, // never trigger processing
2045 -1, // no limit on length
2046 Shared_DirtyCardQ_lock,
2047 &JavaThread::dirty_card_queue_set());
2049 // In case we're keeping closure specialization stats, initialize those
2050 // counts and that mechanism.
2051 SpecializationStats::clear();
2053 // Do later initialization work for concurrent refinement.
2054 _cg1r->init();
2056 // Here we allocate the dummy full region that is required by the
2057 // G1AllocRegion class. If we don't pass an address in the reserved
2058 // space here, lots of asserts fire.
2060 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2061 _g1_reserved.start());
2062 // We'll re-use the same region whether the alloc region will
2063 // require BOT updates or not and, if it doesn't, then a non-young
2064 // region will complain that it cannot support allocations without
2065 // BOT updates. So we'll tag the dummy region as young to avoid that.
2066 dummy_region->set_young();
2067 // Make sure it's full.
2068 dummy_region->set_top(dummy_region->end());
2069 G1AllocRegion::setup(this, dummy_region);
2071 init_mutator_alloc_region();
2073 // Do create of the monitoring and management support so that
2074 // values in the heap have been properly initialized.
2075 _g1mm = new G1MonitoringSupport(this);
2077 return JNI_OK;
2078 }
2080 void G1CollectedHeap::ref_processing_init() {
2081 // Reference processing in G1 currently works as follows:
2082 //
2083 // * There are two reference processor instances. One is
2084 // used to record and process discovered references
2085 // during concurrent marking; the other is used to
2086 // record and process references during STW pauses
2087 // (both full and incremental).
2088 // * Both ref processors need to 'span' the entire heap as
2089 // the regions in the collection set may be dotted around.
2090 //
2091 // * For the concurrent marking ref processor:
2092 // * Reference discovery is enabled at initial marking.
2093 // * Reference discovery is disabled and the discovered
2094 // references processed etc during remarking.
2095 // * Reference discovery is MT (see below).
2096 // * Reference discovery requires a barrier (see below).
2097 // * Reference processing may or may not be MT
2098 // (depending on the value of ParallelRefProcEnabled
2099 // and ParallelGCThreads).
2100 // * A full GC disables reference discovery by the CM
2101 // ref processor and abandons any entries on it's
2102 // discovered lists.
2103 //
2104 // * For the STW processor:
2105 // * Non MT discovery is enabled at the start of a full GC.
2106 // * Processing and enqueueing during a full GC is non-MT.
2107 // * During a full GC, references are processed after marking.
2108 //
2109 // * Discovery (may or may not be MT) is enabled at the start
2110 // of an incremental evacuation pause.
2111 // * References are processed near the end of a STW evacuation pause.
2112 // * For both types of GC:
2113 // * Discovery is atomic - i.e. not concurrent.
2114 // * Reference discovery will not need a barrier.
2116 SharedHeap::ref_processing_init();
2117 MemRegion mr = reserved_region();
2119 // Concurrent Mark ref processor
2120 _ref_processor_cm =
2121 new ReferenceProcessor(mr, // span
2122 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2123 // mt processing
2124 (int) ParallelGCThreads,
2125 // degree of mt processing
2126 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2127 // mt discovery
2128 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2129 // degree of mt discovery
2130 false,
2131 // Reference discovery is not atomic
2132 &_is_alive_closure_cm,
2133 // is alive closure
2134 // (for efficiency/performance)
2135 true);
2136 // Setting next fields of discovered
2137 // lists requires a barrier.
2139 // STW ref processor
2140 _ref_processor_stw =
2141 new ReferenceProcessor(mr, // span
2142 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2143 // mt processing
2144 MAX2((int)ParallelGCThreads, 1),
2145 // degree of mt processing
2146 (ParallelGCThreads > 1),
2147 // mt discovery
2148 MAX2((int)ParallelGCThreads, 1),
2149 // degree of mt discovery
2150 true,
2151 // Reference discovery is atomic
2152 &_is_alive_closure_stw,
2153 // is alive closure
2154 // (for efficiency/performance)
2155 false);
2156 // Setting next fields of discovered
2157 // lists requires a barrier.
2158 }
2160 size_t G1CollectedHeap::capacity() const {
2161 return _g1_committed.byte_size();
2162 }
2164 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2165 DirtyCardQueue* into_cset_dcq,
2166 bool concurrent,
2167 int worker_i) {
2168 // Clean cards in the hot card cache
2169 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2171 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2172 int n_completed_buffers = 0;
2173 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2174 n_completed_buffers++;
2175 }
2176 g1_policy()->record_update_rs_processed_buffers(worker_i,
2177 (double) n_completed_buffers);
2178 dcqs.clear_n_completed_buffers();
2179 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2180 }
2183 // Computes the sum of the storage used by the various regions.
2185 size_t G1CollectedHeap::used() const {
2186 assert(Heap_lock->owner() != NULL,
2187 "Should be owned on this thread's behalf.");
2188 size_t result = _summary_bytes_used;
2189 // Read only once in case it is set to NULL concurrently
2190 HeapRegion* hr = _mutator_alloc_region.get();
2191 if (hr != NULL)
2192 result += hr->used();
2193 return result;
2194 }
2196 size_t G1CollectedHeap::used_unlocked() const {
2197 size_t result = _summary_bytes_used;
2198 return result;
2199 }
2201 class SumUsedClosure: public HeapRegionClosure {
2202 size_t _used;
2203 public:
2204 SumUsedClosure() : _used(0) {}
2205 bool doHeapRegion(HeapRegion* r) {
2206 if (!r->continuesHumongous()) {
2207 _used += r->used();
2208 }
2209 return false;
2210 }
2211 size_t result() { return _used; }
2212 };
2214 size_t G1CollectedHeap::recalculate_used() const {
2215 SumUsedClosure blk;
2216 heap_region_iterate(&blk);
2217 return blk.result();
2218 }
2220 size_t G1CollectedHeap::unsafe_max_alloc() {
2221 if (free_regions() > 0) return HeapRegion::GrainBytes;
2222 // otherwise, is there space in the current allocation region?
2224 // We need to store the current allocation region in a local variable
2225 // here. The problem is that this method doesn't take any locks and
2226 // there may be other threads which overwrite the current allocation
2227 // region field. attempt_allocation(), for example, sets it to NULL
2228 // and this can happen *after* the NULL check here but before the call
2229 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2230 // to be a problem in the optimized build, since the two loads of the
2231 // current allocation region field are optimized away.
2232 HeapRegion* hr = _mutator_alloc_region.get();
2233 if (hr == NULL) {
2234 return 0;
2235 }
2236 return hr->free();
2237 }
2239 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2240 return
2241 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) ||
2242 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent));
2243 }
2245 #ifndef PRODUCT
2246 void G1CollectedHeap::allocate_dummy_regions() {
2247 // Let's fill up most of the region
2248 size_t word_size = HeapRegion::GrainWords - 1024;
2249 // And as a result the region we'll allocate will be humongous.
2250 guarantee(isHumongous(word_size), "sanity");
2252 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2253 // Let's use the existing mechanism for the allocation
2254 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2255 if (dummy_obj != NULL) {
2256 MemRegion mr(dummy_obj, word_size);
2257 CollectedHeap::fill_with_object(mr);
2258 } else {
2259 // If we can't allocate once, we probably cannot allocate
2260 // again. Let's get out of the loop.
2261 break;
2262 }
2263 }
2264 }
2265 #endif // !PRODUCT
2267 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) {
2268 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2270 // We assume that if concurrent == true, then the caller is a
2271 // concurrent thread that was joined the Suspendible Thread
2272 // Set. If there's ever a cheap way to check this, we should add an
2273 // assert here.
2275 // We have already incremented _total_full_collections at the start
2276 // of the GC, so total_full_collections() represents how many full
2277 // collections have been started.
2278 unsigned int full_collections_started = total_full_collections();
2280 // Given that this method is called at the end of a Full GC or of a
2281 // concurrent cycle, and those can be nested (i.e., a Full GC can
2282 // interrupt a concurrent cycle), the number of full collections
2283 // completed should be either one (in the case where there was no
2284 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2285 // behind the number of full collections started.
2287 // This is the case for the inner caller, i.e. a Full GC.
2288 assert(concurrent ||
2289 (full_collections_started == _full_collections_completed + 1) ||
2290 (full_collections_started == _full_collections_completed + 2),
2291 err_msg("for inner caller (Full GC): full_collections_started = %u "
2292 "is inconsistent with _full_collections_completed = %u",
2293 full_collections_started, _full_collections_completed));
2295 // This is the case for the outer caller, i.e. the concurrent cycle.
2296 assert(!concurrent ||
2297 (full_collections_started == _full_collections_completed + 1),
2298 err_msg("for outer caller (concurrent cycle): "
2299 "full_collections_started = %u "
2300 "is inconsistent with _full_collections_completed = %u",
2301 full_collections_started, _full_collections_completed));
2303 _full_collections_completed += 1;
2305 // We need to clear the "in_progress" flag in the CM thread before
2306 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2307 // is set) so that if a waiter requests another System.gc() it doesn't
2308 // incorrectly see that a marking cyle is still in progress.
2309 if (concurrent) {
2310 _cmThread->clear_in_progress();
2311 }
2313 // This notify_all() will ensure that a thread that called
2314 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2315 // and it's waiting for a full GC to finish will be woken up. It is
2316 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2317 FullGCCount_lock->notify_all();
2318 }
2320 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) {
2321 assert_at_safepoint(true /* should_be_vm_thread */);
2322 GCCauseSetter gcs(this, cause);
2323 switch (cause) {
2324 case GCCause::_heap_inspection:
2325 case GCCause::_heap_dump: {
2326 HandleMark hm;
2327 do_full_collection(false); // don't clear all soft refs
2328 break;
2329 }
2330 default: // XXX FIX ME
2331 ShouldNotReachHere(); // Unexpected use of this function
2332 }
2333 }
2335 void G1CollectedHeap::collect(GCCause::Cause cause) {
2336 // The caller doesn't have the Heap_lock
2337 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock");
2339 unsigned int gc_count_before;
2340 unsigned int full_gc_count_before;
2341 {
2342 MutexLocker ml(Heap_lock);
2344 // Read the GC count while holding the Heap_lock
2345 gc_count_before = SharedHeap::heap()->total_collections();
2346 full_gc_count_before = SharedHeap::heap()->total_full_collections();
2347 }
2349 if (should_do_concurrent_full_gc(cause)) {
2350 // Schedule an initial-mark evacuation pause that will start a
2351 // concurrent cycle. We're setting word_size to 0 which means that
2352 // we are not requesting a post-GC allocation.
2353 VM_G1IncCollectionPause op(gc_count_before,
2354 0, /* word_size */
2355 true, /* should_initiate_conc_mark */
2356 g1_policy()->max_pause_time_ms(),
2357 cause);
2358 VMThread::execute(&op);
2359 } else {
2360 if (cause == GCCause::_gc_locker
2361 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2363 // Schedule a standard evacuation pause. We're setting word_size
2364 // to 0 which means that we are not requesting a post-GC allocation.
2365 VM_G1IncCollectionPause op(gc_count_before,
2366 0, /* word_size */
2367 false, /* should_initiate_conc_mark */
2368 g1_policy()->max_pause_time_ms(),
2369 cause);
2370 VMThread::execute(&op);
2371 } else {
2372 // Schedule a Full GC.
2373 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
2374 VMThread::execute(&op);
2375 }
2376 }
2377 }
2379 bool G1CollectedHeap::is_in(const void* p) const {
2380 HeapRegion* hr = _hrs.addr_to_region((HeapWord*) p);
2381 if (hr != NULL) {
2382 return hr->is_in(p);
2383 } else {
2384 return _perm_gen->as_gen()->is_in(p);
2385 }
2386 }
2388 // Iteration functions.
2390 // Iterates an OopClosure over all ref-containing fields of objects
2391 // within a HeapRegion.
2393 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2394 MemRegion _mr;
2395 OopClosure* _cl;
2396 public:
2397 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl)
2398 : _mr(mr), _cl(cl) {}
2399 bool doHeapRegion(HeapRegion* r) {
2400 if (! r->continuesHumongous()) {
2401 r->oop_iterate(_cl);
2402 }
2403 return false;
2404 }
2405 };
2407 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) {
2408 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2409 heap_region_iterate(&blk);
2410 if (do_perm) {
2411 perm_gen()->oop_iterate(cl);
2412 }
2413 }
2415 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) {
2416 IterateOopClosureRegionClosure blk(mr, cl);
2417 heap_region_iterate(&blk);
2418 if (do_perm) {
2419 perm_gen()->oop_iterate(cl);
2420 }
2421 }
2423 // Iterates an ObjectClosure over all objects within a HeapRegion.
2425 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2426 ObjectClosure* _cl;
2427 public:
2428 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2429 bool doHeapRegion(HeapRegion* r) {
2430 if (! r->continuesHumongous()) {
2431 r->object_iterate(_cl);
2432 }
2433 return false;
2434 }
2435 };
2437 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) {
2438 IterateObjectClosureRegionClosure blk(cl);
2439 heap_region_iterate(&blk);
2440 if (do_perm) {
2441 perm_gen()->object_iterate(cl);
2442 }
2443 }
2445 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2446 // FIXME: is this right?
2447 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2448 }
2450 // Calls a SpaceClosure on a HeapRegion.
2452 class SpaceClosureRegionClosure: public HeapRegionClosure {
2453 SpaceClosure* _cl;
2454 public:
2455 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2456 bool doHeapRegion(HeapRegion* r) {
2457 _cl->do_space(r);
2458 return false;
2459 }
2460 };
2462 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2463 SpaceClosureRegionClosure blk(cl);
2464 heap_region_iterate(&blk);
2465 }
2467 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2468 _hrs.iterate(cl);
2469 }
2471 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r,
2472 HeapRegionClosure* cl) const {
2473 _hrs.iterate_from(r, cl);
2474 }
2476 void
2477 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2478 int worker,
2479 jint claim_value) {
2480 const size_t regions = n_regions();
2481 const size_t worker_num = (G1CollectedHeap::use_parallel_gc_threads() ? ParallelGCThreads : 1);
2482 // try to spread out the starting points of the workers
2483 const size_t start_index = regions / worker_num * (size_t) worker;
2485 // each worker will actually look at all regions
2486 for (size_t count = 0; count < regions; ++count) {
2487 const size_t index = (start_index + count) % regions;
2488 assert(0 <= index && index < regions, "sanity");
2489 HeapRegion* r = region_at(index);
2490 // we'll ignore "continues humongous" regions (we'll process them
2491 // when we come across their corresponding "start humongous"
2492 // region) and regions already claimed
2493 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2494 continue;
2495 }
2496 // OK, try to claim it
2497 if (r->claimHeapRegion(claim_value)) {
2498 // success!
2499 assert(!r->continuesHumongous(), "sanity");
2500 if (r->startsHumongous()) {
2501 // If the region is "starts humongous" we'll iterate over its
2502 // "continues humongous" first; in fact we'll do them
2503 // first. The order is important. In on case, calling the
2504 // closure on the "starts humongous" region might de-allocate
2505 // and clear all its "continues humongous" regions and, as a
2506 // result, we might end up processing them twice. So, we'll do
2507 // them first (notice: most closures will ignore them anyway) and
2508 // then we'll do the "starts humongous" region.
2509 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) {
2510 HeapRegion* chr = region_at(ch_index);
2512 // if the region has already been claimed or it's not
2513 // "continues humongous" we're done
2514 if (chr->claim_value() == claim_value ||
2515 !chr->continuesHumongous()) {
2516 break;
2517 }
2519 // Noone should have claimed it directly. We can given
2520 // that we claimed its "starts humongous" region.
2521 assert(chr->claim_value() != claim_value, "sanity");
2522 assert(chr->humongous_start_region() == r, "sanity");
2524 if (chr->claimHeapRegion(claim_value)) {
2525 // we should always be able to claim it; noone else should
2526 // be trying to claim this region
2528 bool res2 = cl->doHeapRegion(chr);
2529 assert(!res2, "Should not abort");
2531 // Right now, this holds (i.e., no closure that actually
2532 // does something with "continues humongous" regions
2533 // clears them). We might have to weaken it in the future,
2534 // but let's leave these two asserts here for extra safety.
2535 assert(chr->continuesHumongous(), "should still be the case");
2536 assert(chr->humongous_start_region() == r, "sanity");
2537 } else {
2538 guarantee(false, "we should not reach here");
2539 }
2540 }
2541 }
2543 assert(!r->continuesHumongous(), "sanity");
2544 bool res = cl->doHeapRegion(r);
2545 assert(!res, "Should not abort");
2546 }
2547 }
2548 }
2550 class ResetClaimValuesClosure: public HeapRegionClosure {
2551 public:
2552 bool doHeapRegion(HeapRegion* r) {
2553 r->set_claim_value(HeapRegion::InitialClaimValue);
2554 return false;
2555 }
2556 };
2558 void
2559 G1CollectedHeap::reset_heap_region_claim_values() {
2560 ResetClaimValuesClosure blk;
2561 heap_region_iterate(&blk);
2562 }
2564 #ifdef ASSERT
2565 // This checks whether all regions in the heap have the correct claim
2566 // value. I also piggy-backed on this a check to ensure that the
2567 // humongous_start_region() information on "continues humongous"
2568 // regions is correct.
2570 class CheckClaimValuesClosure : public HeapRegionClosure {
2571 private:
2572 jint _claim_value;
2573 size_t _failures;
2574 HeapRegion* _sh_region;
2575 public:
2576 CheckClaimValuesClosure(jint claim_value) :
2577 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2578 bool doHeapRegion(HeapRegion* r) {
2579 if (r->claim_value() != _claim_value) {
2580 gclog_or_tty->print_cr("Region ["PTR_FORMAT","PTR_FORMAT"), "
2581 "claim value = %d, should be %d",
2582 r->bottom(), r->end(), r->claim_value(),
2583 _claim_value);
2584 ++_failures;
2585 }
2586 if (!r->isHumongous()) {
2587 _sh_region = NULL;
2588 } else if (r->startsHumongous()) {
2589 _sh_region = r;
2590 } else if (r->continuesHumongous()) {
2591 if (r->humongous_start_region() != _sh_region) {
2592 gclog_or_tty->print_cr("Region ["PTR_FORMAT","PTR_FORMAT"), "
2593 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2594 r->bottom(), r->end(),
2595 r->humongous_start_region(),
2596 _sh_region);
2597 ++_failures;
2598 }
2599 }
2600 return false;
2601 }
2602 size_t failures() {
2603 return _failures;
2604 }
2605 };
2607 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2608 CheckClaimValuesClosure cl(claim_value);
2609 heap_region_iterate(&cl);
2610 return cl.failures() == 0;
2611 }
2612 #endif // ASSERT
2614 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2615 HeapRegion* r = g1_policy()->collection_set();
2616 while (r != NULL) {
2617 HeapRegion* next = r->next_in_collection_set();
2618 if (cl->doHeapRegion(r)) {
2619 cl->incomplete();
2620 return;
2621 }
2622 r = next;
2623 }
2624 }
2626 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2627 HeapRegionClosure *cl) {
2628 if (r == NULL) {
2629 // The CSet is empty so there's nothing to do.
2630 return;
2631 }
2633 assert(r->in_collection_set(),
2634 "Start region must be a member of the collection set.");
2635 HeapRegion* cur = r;
2636 while (cur != NULL) {
2637 HeapRegion* next = cur->next_in_collection_set();
2638 if (cl->doHeapRegion(cur) && false) {
2639 cl->incomplete();
2640 return;
2641 }
2642 cur = next;
2643 }
2644 cur = g1_policy()->collection_set();
2645 while (cur != r) {
2646 HeapRegion* next = cur->next_in_collection_set();
2647 if (cl->doHeapRegion(cur) && false) {
2648 cl->incomplete();
2649 return;
2650 }
2651 cur = next;
2652 }
2653 }
2655 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2656 return n_regions() > 0 ? region_at(0) : NULL;
2657 }
2660 Space* G1CollectedHeap::space_containing(const void* addr) const {
2661 Space* res = heap_region_containing(addr);
2662 if (res == NULL)
2663 res = perm_gen()->space_containing(addr);
2664 return res;
2665 }
2667 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2668 Space* sp = space_containing(addr);
2669 if (sp != NULL) {
2670 return sp->block_start(addr);
2671 }
2672 return NULL;
2673 }
2675 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2676 Space* sp = space_containing(addr);
2677 assert(sp != NULL, "block_size of address outside of heap");
2678 return sp->block_size(addr);
2679 }
2681 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2682 Space* sp = space_containing(addr);
2683 return sp->block_is_obj(addr);
2684 }
2686 bool G1CollectedHeap::supports_tlab_allocation() const {
2687 return true;
2688 }
2690 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2691 return HeapRegion::GrainBytes;
2692 }
2694 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2695 // Return the remaining space in the cur alloc region, but not less than
2696 // the min TLAB size.
2698 // Also, this value can be at most the humongous object threshold,
2699 // since we can't allow tlabs to grow big enough to accomodate
2700 // humongous objects.
2702 HeapRegion* hr = _mutator_alloc_region.get();
2703 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2704 if (hr == NULL) {
2705 return max_tlab_size;
2706 } else {
2707 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2708 }
2709 }
2711 size_t G1CollectedHeap::max_capacity() const {
2712 return _g1_reserved.byte_size();
2713 }
2715 jlong G1CollectedHeap::millis_since_last_gc() {
2716 // assert(false, "NYI");
2717 return 0;
2718 }
2720 void G1CollectedHeap::prepare_for_verify() {
2721 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2722 ensure_parsability(false);
2723 }
2724 g1_rem_set()->prepare_for_verify();
2725 }
2727 class VerifyLivenessOopClosure: public OopClosure {
2728 G1CollectedHeap* _g1h;
2729 VerifyOption _vo;
2730 public:
2731 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
2732 _g1h(g1h), _vo(vo)
2733 { }
2734 void do_oop(narrowOop *p) { do_oop_work(p); }
2735 void do_oop( oop *p) { do_oop_work(p); }
2737 template <class T> void do_oop_work(T *p) {
2738 oop obj = oopDesc::load_decode_heap_oop(p);
2739 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
2740 "Dead object referenced by a not dead object");
2741 }
2742 };
2744 class VerifyObjsInRegionClosure: public ObjectClosure {
2745 private:
2746 G1CollectedHeap* _g1h;
2747 size_t _live_bytes;
2748 HeapRegion *_hr;
2749 VerifyOption _vo;
2750 public:
2751 // _vo == UsePrevMarking -> use "prev" marking information,
2752 // _vo == UseNextMarking -> use "next" marking information,
2753 // _vo == UseMarkWord -> use mark word from object header.
2754 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
2755 : _live_bytes(0), _hr(hr), _vo(vo) {
2756 _g1h = G1CollectedHeap::heap();
2757 }
2758 void do_object(oop o) {
2759 VerifyLivenessOopClosure isLive(_g1h, _vo);
2760 assert(o != NULL, "Huh?");
2761 if (!_g1h->is_obj_dead_cond(o, _vo)) {
2762 // If the object is alive according to the mark word,
2763 // then verify that the marking information agrees.
2764 // Note we can't verify the contra-positive of the
2765 // above: if the object is dead (according to the mark
2766 // word), it may not be marked, or may have been marked
2767 // but has since became dead, or may have been allocated
2768 // since the last marking.
2769 if (_vo == VerifyOption_G1UseMarkWord) {
2770 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
2771 }
2773 o->oop_iterate(&isLive);
2774 if (!_hr->obj_allocated_since_prev_marking(o)) {
2775 size_t obj_size = o->size(); // Make sure we don't overflow
2776 _live_bytes += (obj_size * HeapWordSize);
2777 }
2778 }
2779 }
2780 size_t live_bytes() { return _live_bytes; }
2781 };
2783 class PrintObjsInRegionClosure : public ObjectClosure {
2784 HeapRegion *_hr;
2785 G1CollectedHeap *_g1;
2786 public:
2787 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
2788 _g1 = G1CollectedHeap::heap();
2789 };
2791 void do_object(oop o) {
2792 if (o != NULL) {
2793 HeapWord *start = (HeapWord *) o;
2794 size_t word_sz = o->size();
2795 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
2796 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
2797 (void*) o, word_sz,
2798 _g1->isMarkedPrev(o),
2799 _g1->isMarkedNext(o),
2800 _hr->obj_allocated_since_prev_marking(o));
2801 HeapWord *end = start + word_sz;
2802 HeapWord *cur;
2803 int *val;
2804 for (cur = start; cur < end; cur++) {
2805 val = (int *) cur;
2806 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
2807 }
2808 }
2809 }
2810 };
2812 class VerifyRegionClosure: public HeapRegionClosure {
2813 private:
2814 bool _allow_dirty;
2815 bool _par;
2816 VerifyOption _vo;
2817 bool _failures;
2818 public:
2819 // _vo == UsePrevMarking -> use "prev" marking information,
2820 // _vo == UseNextMarking -> use "next" marking information,
2821 // _vo == UseMarkWord -> use mark word from object header.
2822 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo)
2823 : _allow_dirty(allow_dirty),
2824 _par(par),
2825 _vo(vo),
2826 _failures(false) {}
2828 bool failures() {
2829 return _failures;
2830 }
2832 bool doHeapRegion(HeapRegion* r) {
2833 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue,
2834 "Should be unclaimed at verify points.");
2835 if (!r->continuesHumongous()) {
2836 bool failures = false;
2837 r->verify(_allow_dirty, _vo, &failures);
2838 if (failures) {
2839 _failures = true;
2840 } else {
2841 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
2842 r->object_iterate(¬_dead_yet_cl);
2843 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
2844 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
2845 "max_live_bytes "SIZE_FORMAT" "
2846 "< calculated "SIZE_FORMAT,
2847 r->bottom(), r->end(),
2848 r->max_live_bytes(),
2849 not_dead_yet_cl.live_bytes());
2850 _failures = true;
2851 }
2852 }
2853 }
2854 return false; // stop the region iteration if we hit a failure
2855 }
2856 };
2858 class VerifyRootsClosure: public OopsInGenClosure {
2859 private:
2860 G1CollectedHeap* _g1h;
2861 VerifyOption _vo;
2862 bool _failures;
2863 public:
2864 // _vo == UsePrevMarking -> use "prev" marking information,
2865 // _vo == UseNextMarking -> use "next" marking information,
2866 // _vo == UseMarkWord -> use mark word from object header.
2867 VerifyRootsClosure(VerifyOption vo) :
2868 _g1h(G1CollectedHeap::heap()),
2869 _vo(vo),
2870 _failures(false) { }
2872 bool failures() { return _failures; }
2874 template <class T> void do_oop_nv(T* p) {
2875 T heap_oop = oopDesc::load_heap_oop(p);
2876 if (!oopDesc::is_null(heap_oop)) {
2877 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
2878 if (_g1h->is_obj_dead_cond(obj, _vo)) {
2879 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
2880 "points to dead obj "PTR_FORMAT, p, (void*) obj);
2881 if (_vo == VerifyOption_G1UseMarkWord) {
2882 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
2883 }
2884 obj->print_on(gclog_or_tty);
2885 _failures = true;
2886 }
2887 }
2888 }
2890 void do_oop(oop* p) { do_oop_nv(p); }
2891 void do_oop(narrowOop* p) { do_oop_nv(p); }
2892 };
2894 // This is the task used for parallel heap verification.
2896 class G1ParVerifyTask: public AbstractGangTask {
2897 private:
2898 G1CollectedHeap* _g1h;
2899 bool _allow_dirty;
2900 VerifyOption _vo;
2901 bool _failures;
2903 public:
2904 // _vo == UsePrevMarking -> use "prev" marking information,
2905 // _vo == UseNextMarking -> use "next" marking information,
2906 // _vo == UseMarkWord -> use mark word from object header.
2907 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) :
2908 AbstractGangTask("Parallel verify task"),
2909 _g1h(g1h),
2910 _allow_dirty(allow_dirty),
2911 _vo(vo),
2912 _failures(false) { }
2914 bool failures() {
2915 return _failures;
2916 }
2918 void work(int worker_i) {
2919 HandleMark hm;
2920 VerifyRegionClosure blk(_allow_dirty, true, _vo);
2921 _g1h->heap_region_par_iterate_chunked(&blk, worker_i,
2922 HeapRegion::ParVerifyClaimValue);
2923 if (blk.failures()) {
2924 _failures = true;
2925 }
2926 }
2927 };
2929 void G1CollectedHeap::verify(bool allow_dirty, bool silent) {
2930 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking);
2931 }
2933 void G1CollectedHeap::verify(bool allow_dirty,
2934 bool silent,
2935 VerifyOption vo) {
2936 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2937 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); }
2938 VerifyRootsClosure rootsCl(vo);
2939 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
2941 // We apply the relevant closures to all the oops in the
2942 // system dictionary, the string table and the code cache.
2943 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
2945 process_strong_roots(true, // activate StrongRootsScope
2946 true, // we set "collecting perm gen" to true,
2947 // so we don't reset the dirty cards in the perm gen.
2948 SharedHeap::ScanningOption(so), // roots scanning options
2949 &rootsCl,
2950 &blobsCl,
2951 &rootsCl);
2953 // If we're verifying after the marking phase of a Full GC then we can't
2954 // treat the perm gen as roots into the G1 heap. Some of the objects in
2955 // the perm gen may be dead and hence not marked. If one of these dead
2956 // objects is considered to be a root then we may end up with a false
2957 // "Root location <x> points to dead ob <y>" failure.
2958 if (vo != VerifyOption_G1UseMarkWord) {
2959 // Since we used "collecting_perm_gen" == true above, we will not have
2960 // checked the refs from perm into the G1-collected heap. We check those
2961 // references explicitly below. Whether the relevant cards are dirty
2962 // is checked further below in the rem set verification.
2963 if (!silent) { gclog_or_tty->print("Permgen roots "); }
2964 perm_gen()->oop_iterate(&rootsCl);
2965 }
2966 bool failures = rootsCl.failures();
2968 if (vo != VerifyOption_G1UseMarkWord) {
2969 // If we're verifying during a full GC then the region sets
2970 // will have been torn down at the start of the GC. Therefore
2971 // verifying the region sets will fail. So we only verify
2972 // the region sets when not in a full GC.
2973 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
2974 verify_region_sets();
2975 }
2977 if (!silent) { gclog_or_tty->print("HeapRegions "); }
2978 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
2979 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2980 "sanity check");
2982 G1ParVerifyTask task(this, allow_dirty, vo);
2983 int n_workers = workers()->total_workers();
2984 set_par_threads(n_workers);
2985 workers()->run_task(&task);
2986 set_par_threads(0);
2987 if (task.failures()) {
2988 failures = true;
2989 }
2991 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
2992 "sanity check");
2994 reset_heap_region_claim_values();
2996 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2997 "sanity check");
2998 } else {
2999 VerifyRegionClosure blk(allow_dirty, false, vo);
3000 heap_region_iterate(&blk);
3001 if (blk.failures()) {
3002 failures = true;
3003 }
3004 }
3005 if (!silent) gclog_or_tty->print("RemSet ");
3006 rem_set()->verify();
3008 if (failures) {
3009 gclog_or_tty->print_cr("Heap:");
3010 print_on(gclog_or_tty, true /* extended */);
3011 gclog_or_tty->print_cr("");
3012 #ifndef PRODUCT
3013 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3014 concurrent_mark()->print_reachable("at-verification-failure",
3015 vo, false /* all */);
3016 }
3017 #endif
3018 gclog_or_tty->flush();
3019 }
3020 guarantee(!failures, "there should not have been any failures");
3021 } else {
3022 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3023 }
3024 }
3026 class PrintRegionClosure: public HeapRegionClosure {
3027 outputStream* _st;
3028 public:
3029 PrintRegionClosure(outputStream* st) : _st(st) {}
3030 bool doHeapRegion(HeapRegion* r) {
3031 r->print_on(_st);
3032 return false;
3033 }
3034 };
3036 void G1CollectedHeap::print() const { print_on(tty); }
3038 void G1CollectedHeap::print_on(outputStream* st) const {
3039 print_on(st, PrintHeapAtGCExtended);
3040 }
3042 void G1CollectedHeap::print_on(outputStream* st, bool extended) const {
3043 st->print(" %-20s", "garbage-first heap");
3044 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3045 capacity()/K, used_unlocked()/K);
3046 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3047 _g1_storage.low_boundary(),
3048 _g1_storage.high(),
3049 _g1_storage.high_boundary());
3050 st->cr();
3051 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3052 size_t young_regions = _young_list->length();
3053 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ",
3054 young_regions, young_regions * HeapRegion::GrainBytes / K);
3055 size_t survivor_regions = g1_policy()->recorded_survivor_regions();
3056 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)",
3057 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K);
3058 st->cr();
3059 perm()->as_gen()->print_on(st);
3060 if (extended) {
3061 st->cr();
3062 print_on_extended(st);
3063 }
3064 }
3066 void G1CollectedHeap::print_on_extended(outputStream* st) const {
3067 PrintRegionClosure blk(st);
3068 heap_region_iterate(&blk);
3069 }
3071 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3072 if (G1CollectedHeap::use_parallel_gc_threads()) {
3073 workers()->print_worker_threads_on(st);
3074 }
3075 _cmThread->print_on(st);
3076 st->cr();
3077 _cm->print_worker_threads_on(st);
3078 _cg1r->print_worker_threads_on(st);
3079 st->cr();
3080 }
3082 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3083 if (G1CollectedHeap::use_parallel_gc_threads()) {
3084 workers()->threads_do(tc);
3085 }
3086 tc->do_thread(_cmThread);
3087 _cg1r->threads_do(tc);
3088 }
3090 void G1CollectedHeap::print_tracing_info() const {
3091 // We'll overload this to mean "trace GC pause statistics."
3092 if (TraceGen0Time || TraceGen1Time) {
3093 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3094 // to that.
3095 g1_policy()->print_tracing_info();
3096 }
3097 if (G1SummarizeRSetStats) {
3098 g1_rem_set()->print_summary_info();
3099 }
3100 if (G1SummarizeConcMark) {
3101 concurrent_mark()->print_summary_info();
3102 }
3103 g1_policy()->print_yg_surv_rate_info();
3104 SpecializationStats::print();
3105 }
3107 #ifndef PRODUCT
3108 // Helpful for debugging RSet issues.
3110 class PrintRSetsClosure : public HeapRegionClosure {
3111 private:
3112 const char* _msg;
3113 size_t _occupied_sum;
3115 public:
3116 bool doHeapRegion(HeapRegion* r) {
3117 HeapRegionRemSet* hrrs = r->rem_set();
3118 size_t occupied = hrrs->occupied();
3119 _occupied_sum += occupied;
3121 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3122 HR_FORMAT_PARAMS(r));
3123 if (occupied == 0) {
3124 gclog_or_tty->print_cr(" RSet is empty");
3125 } else {
3126 hrrs->print();
3127 }
3128 gclog_or_tty->print_cr("----------");
3129 return false;
3130 }
3132 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3133 gclog_or_tty->cr();
3134 gclog_or_tty->print_cr("========================================");
3135 gclog_or_tty->print_cr(msg);
3136 gclog_or_tty->cr();
3137 }
3139 ~PrintRSetsClosure() {
3140 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3141 gclog_or_tty->print_cr("========================================");
3142 gclog_or_tty->cr();
3143 }
3144 };
3146 void G1CollectedHeap::print_cset_rsets() {
3147 PrintRSetsClosure cl("Printing CSet RSets");
3148 collection_set_iterate(&cl);
3149 }
3151 void G1CollectedHeap::print_all_rsets() {
3152 PrintRSetsClosure cl("Printing All RSets");;
3153 heap_region_iterate(&cl);
3154 }
3155 #endif // PRODUCT
3157 G1CollectedHeap* G1CollectedHeap::heap() {
3158 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3159 "not a garbage-first heap");
3160 return _g1h;
3161 }
3163 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3164 // always_do_update_barrier = false;
3165 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3166 // Call allocation profiler
3167 AllocationProfiler::iterate_since_last_gc();
3168 // Fill TLAB's and such
3169 ensure_parsability(true);
3170 }
3172 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3173 // FIXME: what is this about?
3174 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3175 // is set.
3176 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3177 "derived pointer present"));
3178 // always_do_update_barrier = true;
3180 // We have just completed a GC. Update the soft reference
3181 // policy with the new heap occupancy
3182 Universe::update_heap_info_at_gc();
3183 }
3185 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3186 unsigned int gc_count_before,
3187 bool* succeeded) {
3188 assert_heap_not_locked_and_not_at_safepoint();
3189 g1_policy()->record_stop_world_start();
3190 VM_G1IncCollectionPause op(gc_count_before,
3191 word_size,
3192 false, /* should_initiate_conc_mark */
3193 g1_policy()->max_pause_time_ms(),
3194 GCCause::_g1_inc_collection_pause);
3195 VMThread::execute(&op);
3197 HeapWord* result = op.result();
3198 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3199 assert(result == NULL || ret_succeeded,
3200 "the result should be NULL if the VM did not succeed");
3201 *succeeded = ret_succeeded;
3203 assert_heap_not_locked();
3204 return result;
3205 }
3207 void
3208 G1CollectedHeap::doConcurrentMark() {
3209 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3210 if (!_cmThread->in_progress()) {
3211 _cmThread->set_started();
3212 CGC_lock->notify();
3213 }
3214 }
3216 // <NEW PREDICTION>
3218 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr,
3219 bool young) {
3220 return _g1_policy->predict_region_elapsed_time_ms(hr, young);
3221 }
3223 void G1CollectedHeap::check_if_region_is_too_expensive(double
3224 predicted_time_ms) {
3225 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms);
3226 }
3228 size_t G1CollectedHeap::pending_card_num() {
3229 size_t extra_cards = 0;
3230 JavaThread *curr = Threads::first();
3231 while (curr != NULL) {
3232 DirtyCardQueue& dcq = curr->dirty_card_queue();
3233 extra_cards += dcq.size();
3234 curr = curr->next();
3235 }
3236 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3237 size_t buffer_size = dcqs.buffer_size();
3238 size_t buffer_num = dcqs.completed_buffers_num();
3239 return buffer_size * buffer_num + extra_cards;
3240 }
3242 size_t G1CollectedHeap::max_pending_card_num() {
3243 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3244 size_t buffer_size = dcqs.buffer_size();
3245 size_t buffer_num = dcqs.completed_buffers_num();
3246 int thread_num = Threads::number_of_threads();
3247 return (buffer_num + thread_num) * buffer_size;
3248 }
3250 size_t G1CollectedHeap::cards_scanned() {
3251 return g1_rem_set()->cardsScanned();
3252 }
3254 void
3255 G1CollectedHeap::setup_surviving_young_words() {
3256 guarantee( _surviving_young_words == NULL, "pre-condition" );
3257 size_t array_length = g1_policy()->young_cset_length();
3258 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length);
3259 if (_surviving_young_words == NULL) {
3260 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3261 "Not enough space for young surv words summary.");
3262 }
3263 memset(_surviving_young_words, 0, array_length * sizeof(size_t));
3264 #ifdef ASSERT
3265 for (size_t i = 0; i < array_length; ++i) {
3266 assert( _surviving_young_words[i] == 0, "memset above" );
3267 }
3268 #endif // !ASSERT
3269 }
3271 void
3272 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3273 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3274 size_t array_length = g1_policy()->young_cset_length();
3275 for (size_t i = 0; i < array_length; ++i)
3276 _surviving_young_words[i] += surv_young_words[i];
3277 }
3279 void
3280 G1CollectedHeap::cleanup_surviving_young_words() {
3281 guarantee( _surviving_young_words != NULL, "pre-condition" );
3282 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words);
3283 _surviving_young_words = NULL;
3284 }
3286 // </NEW PREDICTION>
3288 #ifdef ASSERT
3289 class VerifyCSetClosure: public HeapRegionClosure {
3290 public:
3291 bool doHeapRegion(HeapRegion* hr) {
3292 // Here we check that the CSet region's RSet is ready for parallel
3293 // iteration. The fields that we'll verify are only manipulated
3294 // when the region is part of a CSet and is collected. Afterwards,
3295 // we reset these fields when we clear the region's RSet (when the
3296 // region is freed) so they are ready when the region is
3297 // re-allocated. The only exception to this is if there's an
3298 // evacuation failure and instead of freeing the region we leave
3299 // it in the heap. In that case, we reset these fields during
3300 // evacuation failure handling.
3301 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3303 // Here's a good place to add any other checks we'd like to
3304 // perform on CSet regions.
3305 return false;
3306 }
3307 };
3308 #endif // ASSERT
3310 #if TASKQUEUE_STATS
3311 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3312 st->print_raw_cr("GC Task Stats");
3313 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3314 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3315 }
3317 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3318 print_taskqueue_stats_hdr(st);
3320 TaskQueueStats totals;
3321 const int n = workers() != NULL ? workers()->total_workers() : 1;
3322 for (int i = 0; i < n; ++i) {
3323 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3324 totals += task_queue(i)->stats;
3325 }
3326 st->print_raw("tot "); totals.print(st); st->cr();
3328 DEBUG_ONLY(totals.verify());
3329 }
3331 void G1CollectedHeap::reset_taskqueue_stats() {
3332 const int n = workers() != NULL ? workers()->total_workers() : 1;
3333 for (int i = 0; i < n; ++i) {
3334 task_queue(i)->stats.reset();
3335 }
3336 }
3337 #endif // TASKQUEUE_STATS
3339 bool
3340 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3341 assert_at_safepoint(true /* should_be_vm_thread */);
3342 guarantee(!is_gc_active(), "collection is not reentrant");
3344 if (GC_locker::check_active_before_gc()) {
3345 return false;
3346 }
3348 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3349 ResourceMark rm;
3351 if (PrintHeapAtGC) {
3352 Universe::print_heap_before_gc();
3353 }
3355 verify_region_sets_optional();
3356 verify_dirty_young_regions();
3358 {
3359 // This call will decide whether this pause is an initial-mark
3360 // pause. If it is, during_initial_mark_pause() will return true
3361 // for the duration of this pause.
3362 g1_policy()->decide_on_conc_mark_initiation();
3364 // We do not allow initial-mark to be piggy-backed on a
3365 // partially-young GC.
3366 assert(!g1_policy()->during_initial_mark_pause() ||
3367 g1_policy()->full_young_gcs(), "sanity");
3369 // We also do not allow partially-young GCs during marking.
3370 assert(!mark_in_progress() || g1_policy()->full_young_gcs(), "sanity");
3372 char verbose_str[128];
3373 sprintf(verbose_str, "GC pause ");
3374 if (g1_policy()->full_young_gcs()) {
3375 strcat(verbose_str, "(young)");
3376 } else {
3377 strcat(verbose_str, "(partial)");
3378 }
3379 if (g1_policy()->during_initial_mark_pause()) {
3380 strcat(verbose_str, " (initial-mark)");
3381 // We are about to start a marking cycle, so we increment the
3382 // full collection counter.
3383 increment_total_full_collections();
3384 }
3386 // if PrintGCDetails is on, we'll print long statistics information
3387 // in the collector policy code, so let's not print this as the output
3388 // is messy if we do.
3389 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
3390 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3391 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty);
3393 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3394 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3396 // If the secondary_free_list is not empty, append it to the
3397 // free_list. No need to wait for the cleanup operation to finish;
3398 // the region allocation code will check the secondary_free_list
3399 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3400 // set, skip this step so that the region allocation code has to
3401 // get entries from the secondary_free_list.
3402 if (!G1StressConcRegionFreeing) {
3403 append_secondary_free_list_if_not_empty_with_lock();
3404 }
3406 assert(check_young_list_well_formed(),
3407 "young list should be well formed");
3409 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3410 IsGCActiveMark x;
3412 gc_prologue(false);
3413 increment_total_collections(false /* full gc */);
3414 increment_gc_time_stamp();
3416 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
3417 HandleMark hm; // Discard invalid handles created during verification
3418 gclog_or_tty->print(" VerifyBeforeGC:");
3419 prepare_for_verify();
3420 Universe::verify(/* allow dirty */ false,
3421 /* silent */ false,
3422 /* option */ VerifyOption_G1UsePrevMarking);
3424 }
3426 COMPILER2_PRESENT(DerivedPointerTable::clear());
3428 // Please see comment in g1CollectedHeap.hpp and
3429 // G1CollectedHeap::ref_processing_init() to see how
3430 // reference processing currently works in G1.
3432 // Enable discovery in the STW reference processor
3433 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3434 true /*verify_no_refs*/);
3436 {
3437 // We want to temporarily turn off discovery by the
3438 // CM ref processor, if necessary, and turn it back on
3439 // on again later if we do. Using a scoped
3440 // NoRefDiscovery object will do this.
3441 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3443 // Forget the current alloc region (we might even choose it to be part
3444 // of the collection set!).
3445 release_mutator_alloc_region();
3447 // We should call this after we retire the mutator alloc
3448 // region(s) so that all the ALLOC / RETIRE events are generated
3449 // before the start GC event.
3450 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3452 // The elapsed time induced by the start time below deliberately elides
3453 // the possible verification above.
3454 double start_time_sec = os::elapsedTime();
3455 size_t start_used_bytes = used();
3457 #if YOUNG_LIST_VERBOSE
3458 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3459 _young_list->print();
3460 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3461 #endif // YOUNG_LIST_VERBOSE
3463 g1_policy()->record_collection_pause_start(start_time_sec,
3464 start_used_bytes);
3466 #if YOUNG_LIST_VERBOSE
3467 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3468 _young_list->print();
3469 #endif // YOUNG_LIST_VERBOSE
3471 if (g1_policy()->during_initial_mark_pause()) {
3472 concurrent_mark()->checkpointRootsInitialPre();
3473 }
3474 perm_gen()->save_marks();
3476 // We must do this before any possible evacuation that should propagate
3477 // marks.
3478 if (mark_in_progress()) {
3479 double start_time_sec = os::elapsedTime();
3481 _cm->drainAllSATBBuffers();
3482 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0;
3483 g1_policy()->record_satb_drain_time(finish_mark_ms);
3484 }
3485 // Record the number of elements currently on the mark stack, so we
3486 // only iterate over these. (Since evacuation may add to the mark
3487 // stack, doing more exposes race conditions.) If no mark is in
3488 // progress, this will be zero.
3489 _cm->set_oops_do_bound();
3491 if (mark_in_progress()) {
3492 concurrent_mark()->newCSet();
3493 }
3495 #if YOUNG_LIST_VERBOSE
3496 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3497 _young_list->print();
3498 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3499 #endif // YOUNG_LIST_VERBOSE
3501 g1_policy()->choose_collection_set(target_pause_time_ms);
3503 if (_hr_printer.is_active()) {
3504 HeapRegion* hr = g1_policy()->collection_set();
3505 while (hr != NULL) {
3506 G1HRPrinter::RegionType type;
3507 if (!hr->is_young()) {
3508 type = G1HRPrinter::Old;
3509 } else if (hr->is_survivor()) {
3510 type = G1HRPrinter::Survivor;
3511 } else {
3512 type = G1HRPrinter::Eden;
3513 }
3514 _hr_printer.cset(hr);
3515 hr = hr->next_in_collection_set();
3516 }
3517 }
3519 // We have chosen the complete collection set. If marking is
3520 // active then, we clear the region fields of any of the
3521 // concurrent marking tasks whose region fields point into
3522 // the collection set as these values will become stale. This
3523 // will cause the owning marking threads to claim a new region
3524 // when marking restarts.
3525 if (mark_in_progress()) {
3526 concurrent_mark()->reset_active_task_region_fields_in_cset();
3527 }
3529 #ifdef ASSERT
3530 VerifyCSetClosure cl;
3531 collection_set_iterate(&cl);
3532 #endif // ASSERT
3534 setup_surviving_young_words();
3536 // Initialize the GC alloc regions.
3537 init_gc_alloc_regions();
3539 // Actually do the work...
3540 evacuate_collection_set();
3542 free_collection_set(g1_policy()->collection_set());
3543 g1_policy()->clear_collection_set();
3545 cleanup_surviving_young_words();
3547 // Start a new incremental collection set for the next pause.
3548 g1_policy()->start_incremental_cset_building();
3550 // Clear the _cset_fast_test bitmap in anticipation of adding
3551 // regions to the incremental collection set for the next
3552 // evacuation pause.
3553 clear_cset_fast_test();
3555 _young_list->reset_sampled_info();
3557 // Don't check the whole heap at this point as the
3558 // GC alloc regions from this pause have been tagged
3559 // as survivors and moved on to the survivor list.
3560 // Survivor regions will fail the !is_young() check.
3561 assert(check_young_list_empty(false /* check_heap */),
3562 "young list should be empty");
3564 #if YOUNG_LIST_VERBOSE
3565 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3566 _young_list->print();
3567 #endif // YOUNG_LIST_VERBOSE
3569 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3570 _young_list->first_survivor_region(),
3571 _young_list->last_survivor_region());
3573 _young_list->reset_auxilary_lists();
3575 if (evacuation_failed()) {
3576 _summary_bytes_used = recalculate_used();
3577 } else {
3578 // The "used" of the the collection set have already been subtracted
3579 // when they were freed. Add in the bytes evacuated.
3580 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3581 }
3583 if (g1_policy()->during_initial_mark_pause()) {
3584 concurrent_mark()->checkpointRootsInitialPost();
3585 set_marking_started();
3586 // CAUTION: after the doConcurrentMark() call below,
3587 // the concurrent marking thread(s) could be running
3588 // concurrently with us. Make sure that anything after
3589 // this point does not assume that we are the only GC thread
3590 // running. Note: of course, the actual marking work will
3591 // not start until the safepoint itself is released in
3592 // ConcurrentGCThread::safepoint_desynchronize().
3593 doConcurrentMark();
3594 }
3596 allocate_dummy_regions();
3598 #if YOUNG_LIST_VERBOSE
3599 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3600 _young_list->print();
3601 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3602 #endif // YOUNG_LIST_VERBOSE
3604 init_mutator_alloc_region();
3606 {
3607 size_t expand_bytes = g1_policy()->expansion_amount();
3608 if (expand_bytes > 0) {
3609 size_t bytes_before = capacity();
3610 if (!expand(expand_bytes)) {
3611 // We failed to expand the heap so let's verify that
3612 // committed/uncommitted amount match the backing store
3613 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3614 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3615 }
3616 }
3617 }
3619 double end_time_sec = os::elapsedTime();
3620 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS;
3621 g1_policy()->record_pause_time_ms(pause_time_ms);
3622 g1_policy()->record_collection_pause_end();
3624 MemoryService::track_memory_usage();
3626 // In prepare_for_verify() below we'll need to scan the deferred
3627 // update buffers to bring the RSets up-to-date if
3628 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3629 // the update buffers we'll probably need to scan cards on the
3630 // regions we just allocated to (i.e., the GC alloc
3631 // regions). However, during the last GC we called
3632 // set_saved_mark() on all the GC alloc regions, so card
3633 // scanning might skip the [saved_mark_word()...top()] area of
3634 // those regions (i.e., the area we allocated objects into
3635 // during the last GC). But it shouldn't. Given that
3636 // saved_mark_word() is conditional on whether the GC time stamp
3637 // on the region is current or not, by incrementing the GC time
3638 // stamp here we invalidate all the GC time stamps on all the
3639 // regions and saved_mark_word() will simply return top() for
3640 // all the regions. This is a nicer way of ensuring this rather
3641 // than iterating over the regions and fixing them. In fact, the
3642 // GC time stamp increment here also ensures that
3643 // saved_mark_word() will return top() between pauses, i.e.,
3644 // during concurrent refinement. So we don't need the
3645 // is_gc_active() check to decided which top to use when
3646 // scanning cards (see CR 7039627).
3647 increment_gc_time_stamp();
3649 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
3650 HandleMark hm; // Discard invalid handles created during verification
3651 gclog_or_tty->print(" VerifyAfterGC:");
3652 prepare_for_verify();
3653 Universe::verify(/* allow dirty */ true,
3654 /* silent */ false,
3655 /* option */ VerifyOption_G1UsePrevMarking);
3656 }
3658 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
3659 ref_processor_stw()->verify_no_references_recorded();
3661 // CM reference discovery will be re-enabled if necessary.
3662 }
3664 {
3665 size_t expand_bytes = g1_policy()->expansion_amount();
3666 if (expand_bytes > 0) {
3667 size_t bytes_before = capacity();
3668 // No need for an ergo verbose message here,
3669 // expansion_amount() does this when it returns a value > 0.
3670 if (!expand(expand_bytes)) {
3671 // We failed to expand the heap so let's verify that
3672 // committed/uncommitted amount match the backing store
3673 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3674 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3675 }
3676 }
3677 }
3679 // We should do this after we potentially expand the heap so
3680 // that all the COMMIT events are generated before the end GC
3681 // event, and after we retire the GC alloc regions so that all
3682 // RETIRE events are generated before the end GC event.
3683 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
3685 // We have to do this after we decide whether to expand the heap or not.
3686 g1_policy()->print_heap_transition();
3688 if (mark_in_progress()) {
3689 concurrent_mark()->update_g1_committed();
3690 }
3692 #ifdef TRACESPINNING
3693 ParallelTaskTerminator::print_termination_counts();
3694 #endif
3696 gc_epilogue(false);
3697 }
3699 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) {
3700 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum);
3701 print_tracing_info();
3702 vm_exit(-1);
3703 }
3704 }
3706 _hrs.verify_optional();
3707 verify_region_sets_optional();
3709 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
3710 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
3712 if (PrintHeapAtGC) {
3713 Universe::print_heap_after_gc();
3714 }
3715 g1mm()->update_sizes();
3717 if (G1SummarizeRSetStats &&
3718 (G1SummarizeRSetStatsPeriod > 0) &&
3719 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
3720 g1_rem_set()->print_summary_info();
3721 }
3723 return true;
3724 }
3726 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
3727 {
3728 size_t gclab_word_size;
3729 switch (purpose) {
3730 case GCAllocForSurvived:
3731 gclab_word_size = YoungPLABSize;
3732 break;
3733 case GCAllocForTenured:
3734 gclab_word_size = OldPLABSize;
3735 break;
3736 default:
3737 assert(false, "unknown GCAllocPurpose");
3738 gclab_word_size = OldPLABSize;
3739 break;
3740 }
3741 return gclab_word_size;
3742 }
3744 void G1CollectedHeap::init_mutator_alloc_region() {
3745 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
3746 _mutator_alloc_region.init();
3747 }
3749 void G1CollectedHeap::release_mutator_alloc_region() {
3750 _mutator_alloc_region.release();
3751 assert(_mutator_alloc_region.get() == NULL, "post-condition");
3752 }
3754 void G1CollectedHeap::init_gc_alloc_regions() {
3755 assert_at_safepoint(true /* should_be_vm_thread */);
3757 _survivor_gc_alloc_region.init();
3758 _old_gc_alloc_region.init();
3759 HeapRegion* retained_region = _retained_old_gc_alloc_region;
3760 _retained_old_gc_alloc_region = NULL;
3762 // We will discard the current GC alloc region if:
3763 // a) it's in the collection set (it can happen!),
3764 // b) it's already full (no point in using it),
3765 // c) it's empty (this means that it was emptied during
3766 // a cleanup and it should be on the free list now), or
3767 // d) it's humongous (this means that it was emptied
3768 // during a cleanup and was added to the free list, but
3769 // has been subseqently used to allocate a humongous
3770 // object that may be less than the region size).
3771 if (retained_region != NULL &&
3772 !retained_region->in_collection_set() &&
3773 !(retained_region->top() == retained_region->end()) &&
3774 !retained_region->is_empty() &&
3775 !retained_region->isHumongous()) {
3776 retained_region->set_saved_mark();
3777 _old_gc_alloc_region.set(retained_region);
3778 _hr_printer.reuse(retained_region);
3779 }
3780 }
3782 void G1CollectedHeap::release_gc_alloc_regions() {
3783 _survivor_gc_alloc_region.release();
3784 // If we have an old GC alloc region to release, we'll save it in
3785 // _retained_old_gc_alloc_region. If we don't
3786 // _retained_old_gc_alloc_region will become NULL. This is what we
3787 // want either way so no reason to check explicitly for either
3788 // condition.
3789 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
3790 }
3792 void G1CollectedHeap::abandon_gc_alloc_regions() {
3793 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
3794 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
3795 _retained_old_gc_alloc_region = NULL;
3796 }
3798 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
3799 _drain_in_progress = false;
3800 set_evac_failure_closure(cl);
3801 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
3802 }
3804 void G1CollectedHeap::finalize_for_evac_failure() {
3805 assert(_evac_failure_scan_stack != NULL &&
3806 _evac_failure_scan_stack->length() == 0,
3807 "Postcondition");
3808 assert(!_drain_in_progress, "Postcondition");
3809 delete _evac_failure_scan_stack;
3810 _evac_failure_scan_stack = NULL;
3811 }
3813 class UpdateRSetDeferred : public OopsInHeapRegionClosure {
3814 private:
3815 G1CollectedHeap* _g1;
3816 DirtyCardQueue *_dcq;
3817 CardTableModRefBS* _ct_bs;
3819 public:
3820 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
3821 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {}
3823 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
3824 virtual void do_oop( oop* p) { do_oop_work(p); }
3825 template <class T> void do_oop_work(T* p) {
3826 assert(_from->is_in_reserved(p), "paranoia");
3827 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) &&
3828 !_from->is_survivor()) {
3829 size_t card_index = _ct_bs->index_for(p);
3830 if (_ct_bs->mark_card_deferred(card_index)) {
3831 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
3832 }
3833 }
3834 }
3835 };
3837 class RemoveSelfPointerClosure: public ObjectClosure {
3838 private:
3839 G1CollectedHeap* _g1;
3840 ConcurrentMark* _cm;
3841 HeapRegion* _hr;
3842 size_t _prev_marked_bytes;
3843 size_t _next_marked_bytes;
3844 OopsInHeapRegionClosure *_cl;
3845 public:
3846 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr,
3847 OopsInHeapRegionClosure* cl) :
3848 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0),
3849 _next_marked_bytes(0), _cl(cl) {}
3851 size_t prev_marked_bytes() { return _prev_marked_bytes; }
3852 size_t next_marked_bytes() { return _next_marked_bytes; }
3854 // <original comment>
3855 // The original idea here was to coalesce evacuated and dead objects.
3856 // However that caused complications with the block offset table (BOT).
3857 // In particular if there were two TLABs, one of them partially refined.
3858 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~|
3859 // The BOT entries of the unrefined part of TLAB_2 point to the start
3860 // of TLAB_2. If the last object of the TLAB_1 and the first object
3861 // of TLAB_2 are coalesced, then the cards of the unrefined part
3862 // would point into middle of the filler object.
3863 // The current approach is to not coalesce and leave the BOT contents intact.
3864 // </original comment>
3865 //
3866 // We now reset the BOT when we start the object iteration over the
3867 // region and refine its entries for every object we come across. So
3868 // the above comment is not really relevant and we should be able
3869 // to coalesce dead objects if we want to.
3870 void do_object(oop obj) {
3871 HeapWord* obj_addr = (HeapWord*) obj;
3872 assert(_hr->is_in(obj_addr), "sanity");
3873 size_t obj_size = obj->size();
3874 _hr->update_bot_for_object(obj_addr, obj_size);
3875 if (obj->is_forwarded() && obj->forwardee() == obj) {
3876 // The object failed to move.
3877 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs.");
3878 _cm->markPrev(obj);
3879 assert(_cm->isPrevMarked(obj), "Should be marked!");
3880 _prev_marked_bytes += (obj_size * HeapWordSize);
3881 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) {
3882 _cm->markAndGrayObjectIfNecessary(obj);
3883 }
3884 obj->set_mark(markOopDesc::prototype());
3885 // While we were processing RSet buffers during the
3886 // collection, we actually didn't scan any cards on the
3887 // collection set, since we didn't want to update remebered
3888 // sets with entries that point into the collection set, given
3889 // that live objects fromthe collection set are about to move
3890 // and such entries will be stale very soon. This change also
3891 // dealt with a reliability issue which involved scanning a
3892 // card in the collection set and coming across an array that
3893 // was being chunked and looking malformed. The problem is
3894 // that, if evacuation fails, we might have remembered set
3895 // entries missing given that we skipped cards on the
3896 // collection set. So, we'll recreate such entries now.
3897 obj->oop_iterate(_cl);
3898 assert(_cm->isPrevMarked(obj), "Should be marked!");
3899 } else {
3900 // The object has been either evacuated or is dead. Fill it with a
3901 // dummy object.
3902 MemRegion mr((HeapWord*)obj, obj_size);
3903 CollectedHeap::fill_with_object(mr);
3904 _cm->clearRangeBothMaps(mr);
3905 }
3906 }
3907 };
3909 void G1CollectedHeap::remove_self_forwarding_pointers() {
3910 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set());
3911 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set());
3912 UpdateRSetDeferred deferred_update(_g1h, &dcq);
3913 OopsInHeapRegionClosure *cl;
3914 if (G1DeferredRSUpdate) {
3915 cl = &deferred_update;
3916 } else {
3917 cl = &immediate_update;
3918 }
3919 HeapRegion* cur = g1_policy()->collection_set();
3920 while (cur != NULL) {
3921 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
3922 assert(!cur->isHumongous(), "sanity");
3924 if (cur->evacuation_failed()) {
3925 assert(cur->in_collection_set(), "bad CS");
3926 RemoveSelfPointerClosure rspc(_g1h, cur, cl);
3928 // In the common case we make sure that this is done when the
3929 // region is freed so that it is "ready-to-go" when it's
3930 // re-allocated. However, when evacuation failure happens, a
3931 // region will remain in the heap and might ultimately be added
3932 // to a CSet in the future. So we have to be careful here and
3933 // make sure the region's RSet is ready for parallel iteration
3934 // whenever this might be required in the future.
3935 cur->rem_set()->reset_for_par_iteration();
3936 cur->reset_bot();
3937 cl->set_region(cur);
3938 cur->object_iterate(&rspc);
3940 // A number of manipulations to make the TAMS be the current top,
3941 // and the marked bytes be the ones observed in the iteration.
3942 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) {
3943 // The comments below are the postconditions achieved by the
3944 // calls. Note especially the last such condition, which says that
3945 // the count of marked bytes has been properly restored.
3946 cur->note_start_of_marking(false);
3947 // _next_top_at_mark_start == top, _next_marked_bytes == 0
3948 cur->add_to_marked_bytes(rspc.prev_marked_bytes());
3949 // _next_marked_bytes == prev_marked_bytes.
3950 cur->note_end_of_marking();
3951 // _prev_top_at_mark_start == top(),
3952 // _prev_marked_bytes == prev_marked_bytes
3953 }
3954 // If there is no mark in progress, we modified the _next variables
3955 // above needlessly, but harmlessly.
3956 if (_g1h->mark_in_progress()) {
3957 cur->note_start_of_marking(false);
3958 // _next_top_at_mark_start == top, _next_marked_bytes == 0
3959 // _next_marked_bytes == next_marked_bytes.
3960 }
3961 }
3962 cur = cur->next_in_collection_set();
3963 }
3964 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
3966 // Now restore saved marks, if any.
3967 if (_objs_with_preserved_marks != NULL) {
3968 assert(_preserved_marks_of_objs != NULL, "Both or none.");
3969 guarantee(_objs_with_preserved_marks->length() ==
3970 _preserved_marks_of_objs->length(), "Both or none.");
3971 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
3972 oop obj = _objs_with_preserved_marks->at(i);
3973 markOop m = _preserved_marks_of_objs->at(i);
3974 obj->set_mark(m);
3975 }
3976 // Delete the preserved marks growable arrays (allocated on the C heap).
3977 delete _objs_with_preserved_marks;
3978 delete _preserved_marks_of_objs;
3979 _objs_with_preserved_marks = NULL;
3980 _preserved_marks_of_objs = NULL;
3981 }
3982 }
3984 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
3985 _evac_failure_scan_stack->push(obj);
3986 }
3988 void G1CollectedHeap::drain_evac_failure_scan_stack() {
3989 assert(_evac_failure_scan_stack != NULL, "precondition");
3991 while (_evac_failure_scan_stack->length() > 0) {
3992 oop obj = _evac_failure_scan_stack->pop();
3993 _evac_failure_closure->set_region(heap_region_containing(obj));
3994 obj->oop_iterate_backwards(_evac_failure_closure);
3995 }
3996 }
3998 oop
3999 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4000 oop old,
4001 bool should_mark_root) {
4002 assert(obj_in_cs(old),
4003 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4004 (HeapWord*) old));
4005 markOop m = old->mark();
4006 oop forward_ptr = old->forward_to_atomic(old);
4007 if (forward_ptr == NULL) {
4008 // Forward-to-self succeeded.
4010 // should_mark_root will be true when this routine is called
4011 // from a root scanning closure during an initial mark pause.
4012 // In this case the thread that succeeds in self-forwarding the
4013 // object is also responsible for marking the object.
4014 if (should_mark_root) {
4015 assert(!oopDesc::is_null(old), "shouldn't be");
4016 _cm->grayRoot(old);
4017 }
4019 if (_evac_failure_closure != cl) {
4020 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4021 assert(!_drain_in_progress,
4022 "Should only be true while someone holds the lock.");
4023 // Set the global evac-failure closure to the current thread's.
4024 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4025 set_evac_failure_closure(cl);
4026 // Now do the common part.
4027 handle_evacuation_failure_common(old, m);
4028 // Reset to NULL.
4029 set_evac_failure_closure(NULL);
4030 } else {
4031 // The lock is already held, and this is recursive.
4032 assert(_drain_in_progress, "This should only be the recursive case.");
4033 handle_evacuation_failure_common(old, m);
4034 }
4035 return old;
4036 } else {
4037 // Forward-to-self failed. Either someone else managed to allocate
4038 // space for this object (old != forward_ptr) or they beat us in
4039 // self-forwarding it (old == forward_ptr).
4040 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4041 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4042 "should not be in the CSet",
4043 (HeapWord*) old, (HeapWord*) forward_ptr));
4044 return forward_ptr;
4045 }
4046 }
4048 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4049 set_evacuation_failed(true);
4051 preserve_mark_if_necessary(old, m);
4053 HeapRegion* r = heap_region_containing(old);
4054 if (!r->evacuation_failed()) {
4055 r->set_evacuation_failed(true);
4056 _hr_printer.evac_failure(r);
4057 }
4059 push_on_evac_failure_scan_stack(old);
4061 if (!_drain_in_progress) {
4062 // prevent recursion in copy_to_survivor_space()
4063 _drain_in_progress = true;
4064 drain_evac_failure_scan_stack();
4065 _drain_in_progress = false;
4066 }
4067 }
4069 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4070 assert(evacuation_failed(), "Oversaving!");
4071 // We want to call the "for_promotion_failure" version only in the
4072 // case of a promotion failure.
4073 if (m->must_be_preserved_for_promotion_failure(obj)) {
4074 if (_objs_with_preserved_marks == NULL) {
4075 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4076 _objs_with_preserved_marks =
4077 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
4078 _preserved_marks_of_objs =
4079 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
4080 }
4081 _objs_with_preserved_marks->push(obj);
4082 _preserved_marks_of_objs->push(m);
4083 }
4084 }
4086 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4087 size_t word_size) {
4088 if (purpose == GCAllocForSurvived) {
4089 HeapWord* result = survivor_attempt_allocation(word_size);
4090 if (result != NULL) {
4091 return result;
4092 } else {
4093 // Let's try to allocate in the old gen in case we can fit the
4094 // object there.
4095 return old_attempt_allocation(word_size);
4096 }
4097 } else {
4098 assert(purpose == GCAllocForTenured, "sanity");
4099 HeapWord* result = old_attempt_allocation(word_size);
4100 if (result != NULL) {
4101 return result;
4102 } else {
4103 // Let's try to allocate in the survivors in case we can fit the
4104 // object there.
4105 return survivor_attempt_allocation(word_size);
4106 }
4107 }
4109 ShouldNotReachHere();
4110 // Trying to keep some compilers happy.
4111 return NULL;
4112 }
4114 #ifndef PRODUCT
4115 bool GCLabBitMapClosure::do_bit(size_t offset) {
4116 HeapWord* addr = _bitmap->offsetToHeapWord(offset);
4117 guarantee(_cm->isMarked(oop(addr)), "it should be!");
4118 return true;
4119 }
4120 #endif // PRODUCT
4122 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4123 ParGCAllocBuffer(gclab_word_size),
4124 _should_mark_objects(false),
4125 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size),
4126 _retired(false)
4127 {
4128 //_should_mark_objects is set to true when G1ParCopyHelper needs to
4129 // mark the forwarded location of an evacuated object.
4130 // We set _should_mark_objects to true if marking is active, i.e. when we
4131 // need to propagate a mark, or during an initial mark pause, i.e. when we
4132 // need to mark objects immediately reachable by the roots.
4133 if (G1CollectedHeap::heap()->mark_in_progress() ||
4134 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) {
4135 _should_mark_objects = true;
4136 }
4137 }
4139 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num)
4140 : _g1h(g1h),
4141 _refs(g1h->task_queue(queue_num)),
4142 _dcq(&g1h->dirty_card_queue_set()),
4143 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4144 _g1_rem(g1h->g1_rem_set()),
4145 _hash_seed(17), _queue_num(queue_num),
4146 _term_attempts(0),
4147 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4148 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4149 _age_table(false),
4150 _strong_roots_time(0), _term_time(0),
4151 _alloc_buffer_waste(0), _undo_waste(0)
4152 {
4153 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4154 // we "sacrifice" entry 0 to keep track of surviving bytes for
4155 // non-young regions (where the age is -1)
4156 // We also add a few elements at the beginning and at the end in
4157 // an attempt to eliminate cache contention
4158 size_t real_length = 1 + _g1h->g1_policy()->young_cset_length();
4159 size_t array_length = PADDING_ELEM_NUM +
4160 real_length +
4161 PADDING_ELEM_NUM;
4162 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length);
4163 if (_surviving_young_words_base == NULL)
4164 vm_exit_out_of_memory(array_length * sizeof(size_t),
4165 "Not enough space for young surv histo.");
4166 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4167 memset(_surviving_young_words, 0, real_length * sizeof(size_t));
4169 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4170 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4172 _start = os::elapsedTime();
4173 }
4175 void
4176 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4177 {
4178 st->print_raw_cr("GC Termination Stats");
4179 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4180 " ------waste (KiB)------");
4181 st->print_raw_cr("thr ms ms % ms % attempts"
4182 " total alloc undo");
4183 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4184 " ------- ------- -------");
4185 }
4187 void
4188 G1ParScanThreadState::print_termination_stats(int i,
4189 outputStream* const st) const
4190 {
4191 const double elapsed_ms = elapsed_time() * 1000.0;
4192 const double s_roots_ms = strong_roots_time() * 1000.0;
4193 const double term_ms = term_time() * 1000.0;
4194 st->print_cr("%3d %9.2f %9.2f %6.2f "
4195 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4196 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4197 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4198 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4199 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4200 alloc_buffer_waste() * HeapWordSize / K,
4201 undo_waste() * HeapWordSize / K);
4202 }
4204 #ifdef ASSERT
4205 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4206 assert(ref != NULL, "invariant");
4207 assert(UseCompressedOops, "sanity");
4208 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4209 oop p = oopDesc::load_decode_heap_oop(ref);
4210 assert(_g1h->is_in_g1_reserved(p),
4211 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4212 return true;
4213 }
4215 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4216 assert(ref != NULL, "invariant");
4217 if (has_partial_array_mask(ref)) {
4218 // Must be in the collection set--it's already been copied.
4219 oop p = clear_partial_array_mask(ref);
4220 assert(_g1h->obj_in_cs(p),
4221 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4222 } else {
4223 oop p = oopDesc::load_decode_heap_oop(ref);
4224 assert(_g1h->is_in_g1_reserved(p),
4225 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4226 }
4227 return true;
4228 }
4230 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4231 if (ref.is_narrow()) {
4232 return verify_ref((narrowOop*) ref);
4233 } else {
4234 return verify_ref((oop*) ref);
4235 }
4236 }
4237 #endif // ASSERT
4239 void G1ParScanThreadState::trim_queue() {
4240 assert(_evac_cl != NULL, "not set");
4241 assert(_evac_failure_cl != NULL, "not set");
4242 assert(_partial_scan_cl != NULL, "not set");
4244 StarTask ref;
4245 do {
4246 // Drain the overflow stack first, so other threads can steal.
4247 while (refs()->pop_overflow(ref)) {
4248 deal_with_reference(ref);
4249 }
4251 while (refs()->pop_local(ref)) {
4252 deal_with_reference(ref);
4253 }
4254 } while (!refs()->is_empty());
4255 }
4257 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) :
4258 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4259 _par_scan_state(par_scan_state),
4260 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4261 _mark_in_progress(_g1->mark_in_progress()) { }
4263 template <class T> void G1ParCopyHelper::mark_object(T* p) {
4264 // This is called from do_oop_work for objects that are not
4265 // in the collection set. Objects in the collection set
4266 // are marked after they have been evacuated.
4268 T heap_oop = oopDesc::load_heap_oop(p);
4269 if (!oopDesc::is_null(heap_oop)) {
4270 oop obj = oopDesc::decode_heap_oop(heap_oop);
4271 HeapWord* addr = (HeapWord*)obj;
4272 if (_g1->is_in_g1_reserved(addr)) {
4273 _cm->grayRoot(oop(addr));
4274 }
4275 }
4276 }
4278 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root,
4279 bool should_mark_copy) {
4280 size_t word_sz = old->size();
4281 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4282 // +1 to make the -1 indexes valid...
4283 int young_index = from_region->young_index_in_cset()+1;
4284 assert( (from_region->is_young() && young_index > 0) ||
4285 (!from_region->is_young() && young_index == 0), "invariant" );
4286 G1CollectorPolicy* g1p = _g1->g1_policy();
4287 markOop m = old->mark();
4288 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4289 : m->age();
4290 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4291 word_sz);
4292 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4293 oop obj = oop(obj_ptr);
4295 if (obj_ptr == NULL) {
4296 // This will either forward-to-self, or detect that someone else has
4297 // installed a forwarding pointer.
4298 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4299 return _g1->handle_evacuation_failure_par(cl, old, should_mark_root);
4300 }
4302 // We're going to allocate linearly, so might as well prefetch ahead.
4303 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4305 oop forward_ptr = old->forward_to_atomic(obj);
4306 if (forward_ptr == NULL) {
4307 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4308 if (g1p->track_object_age(alloc_purpose)) {
4309 // We could simply do obj->incr_age(). However, this causes a
4310 // performance issue. obj->incr_age() will first check whether
4311 // the object has a displaced mark by checking its mark word;
4312 // getting the mark word from the new location of the object
4313 // stalls. So, given that we already have the mark word and we
4314 // are about to install it anyway, it's better to increase the
4315 // age on the mark word, when the object does not have a
4316 // displaced mark word. We're not expecting many objects to have
4317 // a displaced marked word, so that case is not optimized
4318 // further (it could be...) and we simply call obj->incr_age().
4320 if (m->has_displaced_mark_helper()) {
4321 // in this case, we have to install the mark word first,
4322 // otherwise obj looks to be forwarded (the old mark word,
4323 // which contains the forward pointer, was copied)
4324 obj->set_mark(m);
4325 obj->incr_age();
4326 } else {
4327 m = m->incr_age();
4328 obj->set_mark(m);
4329 }
4330 _par_scan_state->age_table()->add(obj, word_sz);
4331 } else {
4332 obj->set_mark(m);
4333 }
4335 // Mark the evacuated object or propagate "next" mark bit
4336 if (should_mark_copy) {
4337 if (!use_local_bitmaps ||
4338 !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) {
4339 // if we couldn't mark it on the local bitmap (this happens when
4340 // the object was not allocated in the GCLab), we have to bite
4341 // the bullet and do the standard parallel mark
4342 _cm->markAndGrayObjectIfNecessary(obj);
4343 }
4345 if (_g1->isMarkedNext(old)) {
4346 // Unmark the object's old location so that marking
4347 // doesn't think the old object is alive.
4348 _cm->nextMarkBitMap()->parClear((HeapWord*)old);
4349 }
4350 }
4352 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4353 surv_young_words[young_index] += word_sz;
4355 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4356 arrayOop(old)->set_length(0);
4357 oop* old_p = set_partial_array_mask(old);
4358 _par_scan_state->push_on_queue(old_p);
4359 } else {
4360 // No point in using the slower heap_region_containing() method,
4361 // given that we know obj is in the heap.
4362 _scanner->set_region(_g1->heap_region_containing_raw(obj));
4363 obj->oop_iterate_backwards(_scanner);
4364 }
4365 } else {
4366 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4367 obj = forward_ptr;
4368 }
4369 return obj;
4370 }
4372 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4373 template <class T>
4374 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4375 ::do_oop_work(T* p) {
4376 oop obj = oopDesc::load_decode_heap_oop(p);
4377 assert(barrier != G1BarrierRS || obj != NULL,
4378 "Precondition: G1BarrierRS implies obj is nonNull");
4380 // Marking:
4381 // If the object is in the collection set, then the thread
4382 // that copies the object should mark, or propagate the
4383 // mark to, the evacuated object.
4384 // If the object is not in the collection set then we
4385 // should call the mark_object() method depending on the
4386 // value of the template parameter do_mark_object (which will
4387 // be true for root scanning closures during an initial mark
4388 // pause).
4389 // The mark_object() method first checks whether the object
4390 // is marked and, if not, attempts to mark the object.
4392 // here the null check is implicit in the cset_fast_test() test
4393 if (_g1->in_cset_fast_test(obj)) {
4394 if (obj->is_forwarded()) {
4395 oopDesc::encode_store_heap_oop(p, obj->forwardee());
4396 // If we are a root scanning closure during an initial
4397 // mark pause (i.e. do_mark_object will be true) then
4398 // we also need to handle marking of roots in the
4399 // event of an evacuation failure. In the event of an
4400 // evacuation failure, the object is forwarded to itself
4401 // and not copied. For root-scanning closures, the
4402 // object would be marked after a successful self-forward
4403 // but an object could be pointed to by both a root and non
4404 // root location and be self-forwarded by a non-root-scanning
4405 // closure. Therefore we also have to attempt to mark the
4406 // self-forwarded root object here.
4407 if (do_mark_object && obj->forwardee() == obj) {
4408 mark_object(p);
4409 }
4410 } else {
4411 // During an initial mark pause, objects that are pointed to
4412 // by the roots need to be marked - even in the event of an
4413 // evacuation failure. We pass the template parameter
4414 // do_mark_object (which is true for root scanning closures
4415 // during an initial mark pause) to copy_to_survivor_space
4416 // which will pass it on to the evacuation failure handling
4417 // code. The thread that successfully self-forwards a root
4418 // object to itself is responsible for marking the object.
4419 bool should_mark_root = do_mark_object;
4421 // We need to mark the copied object if we're a root scanning
4422 // closure during an initial mark pause (i.e. do_mark_object
4423 // will be true), or the object is already marked and we need
4424 // to propagate the mark to the evacuated copy.
4425 bool should_mark_copy = do_mark_object ||
4426 _during_initial_mark ||
4427 (_mark_in_progress && !_g1->is_obj_ill(obj));
4429 oop copy_oop = copy_to_survivor_space(obj, should_mark_root,
4430 should_mark_copy);
4431 oopDesc::encode_store_heap_oop(p, copy_oop);
4432 }
4433 // When scanning the RS, we only care about objs in CS.
4434 if (barrier == G1BarrierRS) {
4435 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4436 }
4437 } else {
4438 // The object is not in collection set. If we're a root scanning
4439 // closure during an initial mark pause (i.e. do_mark_object will
4440 // be true) then attempt to mark the object.
4441 if (do_mark_object) {
4442 mark_object(p);
4443 }
4444 }
4446 if (barrier == G1BarrierEvac && obj != NULL) {
4447 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4448 }
4450 if (do_gen_barrier && obj != NULL) {
4451 par_do_barrier(p);
4452 }
4453 }
4455 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4456 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4458 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4459 assert(has_partial_array_mask(p), "invariant");
4460 oop old = clear_partial_array_mask(p);
4461 assert(old->is_objArray(), "must be obj array");
4462 assert(old->is_forwarded(), "must be forwarded");
4463 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
4465 objArrayOop obj = objArrayOop(old->forwardee());
4466 assert((void*)old != (void*)old->forwardee(), "self forwarding here?");
4467 // Process ParGCArrayScanChunk elements now
4468 // and push the remainder back onto queue
4469 int start = arrayOop(old)->length();
4470 int end = obj->length();
4471 int remainder = end - start;
4472 assert(start <= end, "just checking");
4473 if (remainder > 2 * ParGCArrayScanChunk) {
4474 // Test above combines last partial chunk with a full chunk
4475 end = start + ParGCArrayScanChunk;
4476 arrayOop(old)->set_length(end);
4477 // Push remainder.
4478 oop* old_p = set_partial_array_mask(old);
4479 assert(arrayOop(old)->length() < obj->length(), "Empty push?");
4480 _par_scan_state->push_on_queue(old_p);
4481 } else {
4482 // Restore length so that the heap remains parsable in
4483 // case of evacuation failure.
4484 arrayOop(old)->set_length(end);
4485 }
4486 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4487 // process our set of indices (include header in first chunk)
4488 obj->oop_iterate_range(&_scanner, start, end);
4489 }
4491 class G1ParEvacuateFollowersClosure : public VoidClosure {
4492 protected:
4493 G1CollectedHeap* _g1h;
4494 G1ParScanThreadState* _par_scan_state;
4495 RefToScanQueueSet* _queues;
4496 ParallelTaskTerminator* _terminator;
4498 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4499 RefToScanQueueSet* queues() { return _queues; }
4500 ParallelTaskTerminator* terminator() { return _terminator; }
4502 public:
4503 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4504 G1ParScanThreadState* par_scan_state,
4505 RefToScanQueueSet* queues,
4506 ParallelTaskTerminator* terminator)
4507 : _g1h(g1h), _par_scan_state(par_scan_state),
4508 _queues(queues), _terminator(terminator) {}
4510 void do_void();
4512 private:
4513 inline bool offer_termination();
4514 };
4516 bool G1ParEvacuateFollowersClosure::offer_termination() {
4517 G1ParScanThreadState* const pss = par_scan_state();
4518 pss->start_term_time();
4519 const bool res = terminator()->offer_termination();
4520 pss->end_term_time();
4521 return res;
4522 }
4524 void G1ParEvacuateFollowersClosure::do_void() {
4525 StarTask stolen_task;
4526 G1ParScanThreadState* const pss = par_scan_state();
4527 pss->trim_queue();
4529 do {
4530 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4531 assert(pss->verify_task(stolen_task), "sanity");
4532 if (stolen_task.is_narrow()) {
4533 pss->deal_with_reference((narrowOop*) stolen_task);
4534 } else {
4535 pss->deal_with_reference((oop*) stolen_task);
4536 }
4538 // We've just processed a reference and we might have made
4539 // available new entries on the queues. So we have to make sure
4540 // we drain the queues as necessary.
4541 pss->trim_queue();
4542 }
4543 } while (!offer_termination());
4545 pss->retire_alloc_buffers();
4546 }
4548 class G1ParTask : public AbstractGangTask {
4549 protected:
4550 G1CollectedHeap* _g1h;
4551 RefToScanQueueSet *_queues;
4552 ParallelTaskTerminator _terminator;
4553 int _n_workers;
4555 Mutex _stats_lock;
4556 Mutex* stats_lock() { return &_stats_lock; }
4558 size_t getNCards() {
4559 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4560 / G1BlockOffsetSharedArray::N_bytes;
4561 }
4563 public:
4564 G1ParTask(G1CollectedHeap* g1h, int workers, RefToScanQueueSet *task_queues)
4565 : AbstractGangTask("G1 collection"),
4566 _g1h(g1h),
4567 _queues(task_queues),
4568 _terminator(workers, _queues),
4569 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true),
4570 _n_workers(workers)
4571 {}
4573 RefToScanQueueSet* queues() { return _queues; }
4575 RefToScanQueue *work_queue(int i) {
4576 return queues()->queue(i);
4577 }
4579 void work(int i) {
4580 if (i >= _n_workers) return; // no work needed this round
4582 double start_time_ms = os::elapsedTime() * 1000.0;
4583 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms);
4585 ResourceMark rm;
4586 HandleMark hm;
4588 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4590 G1ParScanThreadState pss(_g1h, i);
4591 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4592 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4593 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4595 pss.set_evac_closure(&scan_evac_cl);
4596 pss.set_evac_failure_closure(&evac_failure_cl);
4597 pss.set_partial_scan_closure(&partial_scan_cl);
4599 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4600 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp);
4602 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4603 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp);
4605 OopClosure* scan_root_cl = &only_scan_root_cl;
4606 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl;
4608 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4609 // We also need to mark copied objects.
4610 scan_root_cl = &scan_mark_root_cl;
4611 scan_perm_cl = &scan_mark_perm_cl;
4612 }
4614 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4616 pss.start_strong_roots();
4617 _g1h->g1_process_strong_roots(/* not collecting perm */ false,
4618 SharedHeap::SO_AllClasses,
4619 scan_root_cl,
4620 &push_heap_rs_cl,
4621 scan_perm_cl,
4622 i);
4623 pss.end_strong_roots();
4625 {
4626 double start = os::elapsedTime();
4627 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4628 evac.do_void();
4629 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4630 double term_ms = pss.term_time()*1000.0;
4631 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms);
4632 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts());
4633 }
4634 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4635 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4637 // Clean up any par-expanded rem sets.
4638 HeapRegionRemSet::par_cleanup();
4640 if (ParallelGCVerbose) {
4641 MutexLocker x(stats_lock());
4642 pss.print_termination_stats(i);
4643 }
4645 assert(pss.refs()->is_empty(), "should be empty");
4646 double end_time_ms = os::elapsedTime() * 1000.0;
4647 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms);
4648 }
4649 };
4651 // *** Common G1 Evacuation Stuff
4653 // This method is run in a GC worker.
4655 void
4656 G1CollectedHeap::
4657 g1_process_strong_roots(bool collecting_perm_gen,
4658 SharedHeap::ScanningOption so,
4659 OopClosure* scan_non_heap_roots,
4660 OopsInHeapRegionClosure* scan_rs,
4661 OopsInGenClosure* scan_perm,
4662 int worker_i) {
4664 // First scan the strong roots, including the perm gen.
4665 double ext_roots_start = os::elapsedTime();
4666 double closure_app_time_sec = 0.0;
4668 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4669 BufferingOopsInGenClosure buf_scan_perm(scan_perm);
4670 buf_scan_perm.set_generation(perm_gen());
4672 // Walk the code cache w/o buffering, because StarTask cannot handle
4673 // unaligned oop locations.
4674 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true);
4676 process_strong_roots(false, // no scoping; this is parallel code
4677 collecting_perm_gen, so,
4678 &buf_scan_non_heap_roots,
4679 &eager_scan_code_roots,
4680 &buf_scan_perm);
4682 // Now the CM ref_processor roots.
4683 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
4684 // We need to treat the discovered reference lists of the
4685 // concurrent mark ref processor as roots and keep entries
4686 // (which are added by the marking threads) on them live
4687 // until they can be processed at the end of marking.
4688 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
4689 }
4691 // Finish up any enqueued closure apps (attributed as object copy time).
4692 buf_scan_non_heap_roots.done();
4693 buf_scan_perm.done();
4695 double ext_roots_end = os::elapsedTime();
4697 g1_policy()->reset_obj_copy_time(worker_i);
4698 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() +
4699 buf_scan_non_heap_roots.closure_app_seconds();
4700 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
4702 double ext_root_time_ms =
4703 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0;
4705 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
4707 // Scan strong roots in mark stack.
4708 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) {
4709 concurrent_mark()->oops_do(scan_non_heap_roots);
4710 }
4711 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0;
4712 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms);
4714 // Now scan the complement of the collection set.
4715 if (scan_rs != NULL) {
4716 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
4717 }
4719 _process_strong_tasks->all_tasks_completed();
4720 }
4722 void
4723 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
4724 OopClosure* non_root_closure) {
4725 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
4726 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
4727 }
4729 // Weak Reference Processing support
4731 // An always "is_alive" closure that is used to preserve referents.
4732 // If the object is non-null then it's alive. Used in the preservation
4733 // of referent objects that are pointed to by reference objects
4734 // discovered by the CM ref processor.
4735 class G1AlwaysAliveClosure: public BoolObjectClosure {
4736 G1CollectedHeap* _g1;
4737 public:
4738 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4739 void do_object(oop p) { assert(false, "Do not call."); }
4740 bool do_object_b(oop p) {
4741 if (p != NULL) {
4742 return true;
4743 }
4744 return false;
4745 }
4746 };
4748 bool G1STWIsAliveClosure::do_object_b(oop p) {
4749 // An object is reachable if it is outside the collection set,
4750 // or is inside and copied.
4751 return !_g1->obj_in_cs(p) || p->is_forwarded();
4752 }
4754 // Non Copying Keep Alive closure
4755 class G1KeepAliveClosure: public OopClosure {
4756 G1CollectedHeap* _g1;
4757 public:
4758 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4759 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
4760 void do_oop( oop* p) {
4761 oop obj = *p;
4763 if (_g1->obj_in_cs(obj)) {
4764 assert( obj->is_forwarded(), "invariant" );
4765 *p = obj->forwardee();
4766 }
4767 }
4768 };
4770 // Copying Keep Alive closure - can be called from both
4771 // serial and parallel code as long as different worker
4772 // threads utilize different G1ParScanThreadState instances
4773 // and different queues.
4775 class G1CopyingKeepAliveClosure: public OopClosure {
4776 G1CollectedHeap* _g1h;
4777 OopClosure* _copy_non_heap_obj_cl;
4778 OopsInHeapRegionClosure* _copy_perm_obj_cl;
4779 G1ParScanThreadState* _par_scan_state;
4781 public:
4782 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
4783 OopClosure* non_heap_obj_cl,
4784 OopsInHeapRegionClosure* perm_obj_cl,
4785 G1ParScanThreadState* pss):
4786 _g1h(g1h),
4787 _copy_non_heap_obj_cl(non_heap_obj_cl),
4788 _copy_perm_obj_cl(perm_obj_cl),
4789 _par_scan_state(pss)
4790 {}
4792 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
4793 virtual void do_oop( oop* p) { do_oop_work(p); }
4795 template <class T> void do_oop_work(T* p) {
4796 oop obj = oopDesc::load_decode_heap_oop(p);
4798 if (_g1h->obj_in_cs(obj)) {
4799 // If the referent object has been forwarded (either copied
4800 // to a new location or to itself in the event of an
4801 // evacuation failure) then we need to update the reference
4802 // field and, if both reference and referent are in the G1
4803 // heap, update the RSet for the referent.
4804 //
4805 // If the referent has not been forwarded then we have to keep
4806 // it alive by policy. Therefore we have copy the referent.
4807 //
4808 // If the reference field is in the G1 heap then we can push
4809 // on the PSS queue. When the queue is drained (after each
4810 // phase of reference processing) the object and it's followers
4811 // will be copied, the reference field set to point to the
4812 // new location, and the RSet updated. Otherwise we need to
4813 // use the the non-heap or perm closures directly to copy
4814 // the refernt object and update the pointer, while avoiding
4815 // updating the RSet.
4817 if (_g1h->is_in_g1_reserved(p)) {
4818 _par_scan_state->push_on_queue(p);
4819 } else {
4820 // The reference field is not in the G1 heap.
4821 if (_g1h->perm_gen()->is_in(p)) {
4822 _copy_perm_obj_cl->do_oop(p);
4823 } else {
4824 _copy_non_heap_obj_cl->do_oop(p);
4825 }
4826 }
4827 }
4828 }
4829 };
4831 // Serial drain queue closure. Called as the 'complete_gc'
4832 // closure for each discovered list in some of the
4833 // reference processing phases.
4835 class G1STWDrainQueueClosure: public VoidClosure {
4836 protected:
4837 G1CollectedHeap* _g1h;
4838 G1ParScanThreadState* _par_scan_state;
4840 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4842 public:
4843 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
4844 _g1h(g1h),
4845 _par_scan_state(pss)
4846 { }
4848 void do_void() {
4849 G1ParScanThreadState* const pss = par_scan_state();
4850 pss->trim_queue();
4851 }
4852 };
4854 // Parallel Reference Processing closures
4856 // Implementation of AbstractRefProcTaskExecutor for parallel reference
4857 // processing during G1 evacuation pauses.
4859 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
4860 private:
4861 G1CollectedHeap* _g1h;
4862 RefToScanQueueSet* _queues;
4863 WorkGang* _workers;
4864 int _active_workers;
4866 public:
4867 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
4868 WorkGang* workers,
4869 RefToScanQueueSet *task_queues,
4870 int n_workers) :
4871 _g1h(g1h),
4872 _queues(task_queues),
4873 _workers(workers),
4874 _active_workers(n_workers)
4875 {
4876 assert(n_workers > 0, "shouldn't call this otherwise");
4877 }
4879 // Executes the given task using concurrent marking worker threads.
4880 virtual void execute(ProcessTask& task);
4881 virtual void execute(EnqueueTask& task);
4882 };
4884 // Gang task for possibly parallel reference processing
4886 class G1STWRefProcTaskProxy: public AbstractGangTask {
4887 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
4888 ProcessTask& _proc_task;
4889 G1CollectedHeap* _g1h;
4890 RefToScanQueueSet *_task_queues;
4891 ParallelTaskTerminator* _terminator;
4893 public:
4894 G1STWRefProcTaskProxy(ProcessTask& proc_task,
4895 G1CollectedHeap* g1h,
4896 RefToScanQueueSet *task_queues,
4897 ParallelTaskTerminator* terminator) :
4898 AbstractGangTask("Process reference objects in parallel"),
4899 _proc_task(proc_task),
4900 _g1h(g1h),
4901 _task_queues(task_queues),
4902 _terminator(terminator)
4903 {}
4905 virtual void work(int i) {
4906 // The reference processing task executed by a single worker.
4907 ResourceMark rm;
4908 HandleMark hm;
4910 G1STWIsAliveClosure is_alive(_g1h);
4912 G1ParScanThreadState pss(_g1h, i);
4914 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
4915 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
4916 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
4918 pss.set_evac_closure(&scan_evac_cl);
4919 pss.set_evac_failure_closure(&evac_failure_cl);
4920 pss.set_partial_scan_closure(&partial_scan_cl);
4922 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
4923 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
4925 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
4926 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
4928 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
4929 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
4931 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4932 // We also need to mark copied objects.
4933 copy_non_heap_cl = ©_mark_non_heap_cl;
4934 copy_perm_cl = ©_mark_perm_cl;
4935 }
4937 // Keep alive closure.
4938 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
4940 // Complete GC closure
4941 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
4943 // Call the reference processing task's work routine.
4944 _proc_task.work(i, is_alive, keep_alive, drain_queue);
4946 // Note we cannot assert that the refs array is empty here as not all
4947 // of the processing tasks (specifically phase2 - pp2_work) execute
4948 // the complete_gc closure (which ordinarily would drain the queue) so
4949 // the queue may not be empty.
4950 }
4951 };
4953 // Driver routine for parallel reference processing.
4954 // Creates an instance of the ref processing gang
4955 // task and has the worker threads execute it.
4956 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
4957 assert(_workers != NULL, "Need parallel worker threads.");
4959 ParallelTaskTerminator terminator(_active_workers, _queues);
4960 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
4962 _g1h->set_par_threads(_active_workers);
4963 _workers->run_task(&proc_task_proxy);
4964 _g1h->set_par_threads(0);
4965 }
4967 // Gang task for parallel reference enqueueing.
4969 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
4970 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
4971 EnqueueTask& _enq_task;
4973 public:
4974 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
4975 AbstractGangTask("Enqueue reference objects in parallel"),
4976 _enq_task(enq_task)
4977 { }
4979 virtual void work(int i) {
4980 _enq_task.work(i);
4981 }
4982 };
4984 // Driver routine for parallel reference enqueing.
4985 // Creates an instance of the ref enqueueing gang
4986 // task and has the worker threads execute it.
4988 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
4989 assert(_workers != NULL, "Need parallel worker threads.");
4991 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
4993 _g1h->set_par_threads(_active_workers);
4994 _workers->run_task(&enq_task_proxy);
4995 _g1h->set_par_threads(0);
4996 }
4998 // End of weak reference support closures
5000 // Abstract task used to preserve (i.e. copy) any referent objects
5001 // that are in the collection set and are pointed to by reference
5002 // objects discovered by the CM ref processor.
5004 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5005 protected:
5006 G1CollectedHeap* _g1h;
5007 RefToScanQueueSet *_queues;
5008 ParallelTaskTerminator _terminator;
5009 int _n_workers;
5011 public:
5012 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5013 AbstractGangTask("ParPreserveCMReferents"),
5014 _g1h(g1h),
5015 _queues(task_queues),
5016 _terminator(workers, _queues),
5017 _n_workers(workers)
5018 { }
5020 void work(int i) {
5021 ResourceMark rm;
5022 HandleMark hm;
5024 G1ParScanThreadState pss(_g1h, i);
5025 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5026 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5027 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5029 pss.set_evac_closure(&scan_evac_cl);
5030 pss.set_evac_failure_closure(&evac_failure_cl);
5031 pss.set_partial_scan_closure(&partial_scan_cl);
5033 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5036 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5037 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5039 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5040 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5042 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5043 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5045 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5046 // We also need to mark copied objects.
5047 copy_non_heap_cl = ©_mark_non_heap_cl;
5048 copy_perm_cl = ©_mark_perm_cl;
5049 }
5051 // Is alive closure
5052 G1AlwaysAliveClosure always_alive(_g1h);
5054 // Copying keep alive closure. Applied to referent objects that need
5055 // to be copied.
5056 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5058 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5060 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5061 int stride = MIN2(MAX2(_n_workers, 1), limit);
5063 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5064 // So this must be true - but assert just in case someone decides to
5065 // change the worker ids.
5066 assert(0 <= i && i < limit, "sanity");
5067 assert(!rp->discovery_is_atomic(), "check this code");
5069 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5070 for (int idx = i; idx < limit; idx += stride) {
5071 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5073 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5074 while (iter.has_next()) {
5075 // Since discovery is not atomic for the CM ref processor, we
5076 // can see some null referent objects.
5077 iter.load_ptrs(DEBUG_ONLY(true));
5078 oop ref = iter.obj();
5080 // This will filter nulls.
5081 if (iter.is_referent_alive()) {
5082 iter.make_referent_alive();
5083 }
5084 iter.move_to_next();
5085 }
5086 }
5088 // Drain the queue - which may cause stealing
5089 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5090 drain_queue.do_void();
5091 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5092 assert(pss.refs()->is_empty(), "should be");
5093 }
5094 };
5096 // Weak Reference processing during an evacuation pause (part 1).
5097 void G1CollectedHeap::process_discovered_references() {
5098 double ref_proc_start = os::elapsedTime();
5100 ReferenceProcessor* rp = _ref_processor_stw;
5101 assert(rp->discovery_enabled(), "should have been enabled");
5103 // Any reference objects, in the collection set, that were 'discovered'
5104 // by the CM ref processor should have already been copied (either by
5105 // applying the external root copy closure to the discovered lists, or
5106 // by following an RSet entry).
5107 //
5108 // But some of the referents, that are in the collection set, that these
5109 // reference objects point to may not have been copied: the STW ref
5110 // processor would have seen that the reference object had already
5111 // been 'discovered' and would have skipped discovering the reference,
5112 // but would not have treated the reference object as a regular oop.
5113 // As a reult the copy closure would not have been applied to the
5114 // referent object.
5115 //
5116 // We need to explicitly copy these referent objects - the references
5117 // will be processed at the end of remarking.
5118 //
5119 // We also need to do this copying before we process the reference
5120 // objects discovered by the STW ref processor in case one of these
5121 // referents points to another object which is also referenced by an
5122 // object discovered by the STW ref processor.
5124 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5125 workers()->total_workers() : 1);
5127 set_par_threads(n_workers);
5128 G1ParPreserveCMReferentsTask keep_cm_referents(this, n_workers, _task_queues);
5130 if (G1CollectedHeap::use_parallel_gc_threads()) {
5131 workers()->run_task(&keep_cm_referents);
5132 } else {
5133 keep_cm_referents.work(0);
5134 }
5136 set_par_threads(0);
5138 // Closure to test whether a referent is alive.
5139 G1STWIsAliveClosure is_alive(this);
5141 // Even when parallel reference processing is enabled, the processing
5142 // of JNI refs is serial and performed serially by the current thread
5143 // rather than by a worker. The following PSS will be used for processing
5144 // JNI refs.
5146 // Use only a single queue for this PSS.
5147 G1ParScanThreadState pss(this, 0);
5149 // We do not embed a reference processor in the copying/scanning
5150 // closures while we're actually processing the discovered
5151 // reference objects.
5152 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5153 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5154 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5156 pss.set_evac_closure(&scan_evac_cl);
5157 pss.set_evac_failure_closure(&evac_failure_cl);
5158 pss.set_partial_scan_closure(&partial_scan_cl);
5160 assert(pss.refs()->is_empty(), "pre-condition");
5162 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5163 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL);
5165 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5166 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL);
5168 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5169 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5171 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5172 // We also need to mark copied objects.
5173 copy_non_heap_cl = ©_mark_non_heap_cl;
5174 copy_perm_cl = ©_mark_perm_cl;
5175 }
5177 // Keep alive closure.
5178 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss);
5180 // Serial Complete GC closure
5181 G1STWDrainQueueClosure drain_queue(this, &pss);
5183 // Setup the soft refs policy...
5184 rp->setup_policy(false);
5186 if (!rp->processing_is_mt()) {
5187 // Serial reference processing...
5188 rp->process_discovered_references(&is_alive,
5189 &keep_alive,
5190 &drain_queue,
5191 NULL);
5192 } else {
5193 // Parallel reference processing
5194 int active_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5195 assert(rp->num_q() == active_workers, "sanity");
5196 assert(active_workers <= rp->max_num_q(), "sanity");
5198 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5199 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5200 }
5202 // We have completed copying any necessary live referent objects
5203 // (that were not copied during the actual pause) so we can
5204 // retire any active alloc buffers
5205 pss.retire_alloc_buffers();
5206 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5208 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5209 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0);
5210 }
5212 // Weak Reference processing during an evacuation pause (part 2).
5213 void G1CollectedHeap::enqueue_discovered_references() {
5214 double ref_enq_start = os::elapsedTime();
5216 ReferenceProcessor* rp = _ref_processor_stw;
5217 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5219 // Now enqueue any remaining on the discovered lists on to
5220 // the pending list.
5221 if (!rp->processing_is_mt()) {
5222 // Serial reference processing...
5223 rp->enqueue_discovered_references();
5224 } else {
5225 // Parallel reference enqueuing
5227 int active_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5228 assert(rp->num_q() == active_workers, "sanity");
5229 assert(active_workers <= rp->max_num_q(), "sanity");
5231 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5232 rp->enqueue_discovered_references(&par_task_executor);
5233 }
5235 rp->verify_no_references_recorded();
5236 assert(!rp->discovery_enabled(), "should have been disabled");
5238 // FIXME
5239 // CM's reference processing also cleans up the string and symbol tables.
5240 // Should we do that here also? We could, but it is a serial operation
5241 // and could signicantly increase the pause time.
5243 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5244 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0);
5245 }
5247 void G1CollectedHeap::evacuate_collection_set() {
5248 set_evacuation_failed(false);
5250 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5251 concurrent_g1_refine()->set_use_cache(false);
5252 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5254 int n_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5255 set_par_threads(n_workers);
5256 G1ParTask g1_par_task(this, n_workers, _task_queues);
5258 init_for_evac_failure(NULL);
5260 rem_set()->prepare_for_younger_refs_iterate(true);
5262 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5263 double start_par = os::elapsedTime();
5265 if (G1CollectedHeap::use_parallel_gc_threads()) {
5266 // The individual threads will set their evac-failure closures.
5267 StrongRootsScope srs(this);
5268 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5269 workers()->run_task(&g1_par_task);
5270 } else {
5271 StrongRootsScope srs(this);
5272 g1_par_task.work(0);
5273 }
5275 double par_time = (os::elapsedTime() - start_par) * 1000.0;
5276 g1_policy()->record_par_time(par_time);
5277 set_par_threads(0);
5279 // Process any discovered reference objects - we have
5280 // to do this _before_ we retire the GC alloc regions
5281 // as we may have to copy some 'reachable' referent
5282 // objects (and their reachable sub-graphs) that were
5283 // not copied during the pause.
5284 process_discovered_references();
5286 // Weak root processing.
5287 // Note: when JSR 292 is enabled and code blobs can contain
5288 // non-perm oops then we will need to process the code blobs
5289 // here too.
5290 {
5291 G1STWIsAliveClosure is_alive(this);
5292 G1KeepAliveClosure keep_alive(this);
5293 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5294 }
5296 release_gc_alloc_regions();
5297 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5299 concurrent_g1_refine()->clear_hot_cache();
5300 concurrent_g1_refine()->set_use_cache(true);
5302 finalize_for_evac_failure();
5304 // Must do this before removing self-forwarding pointers, which clears
5305 // the per-region evac-failure flags.
5306 concurrent_mark()->complete_marking_in_collection_set();
5308 if (evacuation_failed()) {
5309 remove_self_forwarding_pointers();
5310 if (PrintGCDetails) {
5311 gclog_or_tty->print(" (to-space overflow)");
5312 } else if (PrintGC) {
5313 gclog_or_tty->print("--");
5314 }
5315 }
5317 // Enqueue any remaining references remaining on the STW
5318 // reference processor's discovered lists. We need to do
5319 // this after the card table is cleaned (and verified) as
5320 // the act of enqueuing entries on to the pending list
5321 // will log these updates (and dirty their associated
5322 // cards). We need these updates logged to update any
5323 // RSets.
5324 enqueue_discovered_references();
5326 if (G1DeferredRSUpdate) {
5327 RedirtyLoggedCardTableEntryFastClosure redirty;
5328 dirty_card_queue_set().set_closure(&redirty);
5329 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5331 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5332 dcq.merge_bufferlists(&dirty_card_queue_set());
5333 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5334 }
5335 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5336 }
5338 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5339 size_t* pre_used,
5340 FreeRegionList* free_list,
5341 HumongousRegionSet* humongous_proxy_set,
5342 HRRSCleanupTask* hrrs_cleanup_task,
5343 bool par) {
5344 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5345 if (hr->isHumongous()) {
5346 assert(hr->startsHumongous(), "we should only see starts humongous");
5347 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5348 } else {
5349 free_region(hr, pre_used, free_list, par);
5350 }
5351 } else {
5352 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5353 }
5354 }
5356 void G1CollectedHeap::free_region(HeapRegion* hr,
5357 size_t* pre_used,
5358 FreeRegionList* free_list,
5359 bool par) {
5360 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5361 assert(!hr->is_empty(), "the region should not be empty");
5362 assert(free_list != NULL, "pre-condition");
5364 *pre_used += hr->used();
5365 hr->hr_clear(par, true /* clear_space */);
5366 free_list->add_as_head(hr);
5367 }
5369 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5370 size_t* pre_used,
5371 FreeRegionList* free_list,
5372 HumongousRegionSet* humongous_proxy_set,
5373 bool par) {
5374 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5375 assert(free_list != NULL, "pre-condition");
5376 assert(humongous_proxy_set != NULL, "pre-condition");
5378 size_t hr_used = hr->used();
5379 size_t hr_capacity = hr->capacity();
5380 size_t hr_pre_used = 0;
5381 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5382 hr->set_notHumongous();
5383 free_region(hr, &hr_pre_used, free_list, par);
5385 size_t i = hr->hrs_index() + 1;
5386 size_t num = 1;
5387 while (i < n_regions()) {
5388 HeapRegion* curr_hr = region_at(i);
5389 if (!curr_hr->continuesHumongous()) {
5390 break;
5391 }
5392 curr_hr->set_notHumongous();
5393 free_region(curr_hr, &hr_pre_used, free_list, par);
5394 num += 1;
5395 i += 1;
5396 }
5397 assert(hr_pre_used == hr_used,
5398 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5399 "should be the same", hr_pre_used, hr_used));
5400 *pre_used += hr_pre_used;
5401 }
5403 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5404 FreeRegionList* free_list,
5405 HumongousRegionSet* humongous_proxy_set,
5406 bool par) {
5407 if (pre_used > 0) {
5408 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5409 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5410 assert(_summary_bytes_used >= pre_used,
5411 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5412 "should be >= pre_used: "SIZE_FORMAT,
5413 _summary_bytes_used, pre_used));
5414 _summary_bytes_used -= pre_used;
5415 }
5416 if (free_list != NULL && !free_list->is_empty()) {
5417 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5418 _free_list.add_as_head(free_list);
5419 }
5420 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5421 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5422 _humongous_set.update_from_proxy(humongous_proxy_set);
5423 }
5424 }
5426 class G1ParCleanupCTTask : public AbstractGangTask {
5427 CardTableModRefBS* _ct_bs;
5428 G1CollectedHeap* _g1h;
5429 HeapRegion* volatile _su_head;
5430 public:
5431 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5432 G1CollectedHeap* g1h) :
5433 AbstractGangTask("G1 Par Cleanup CT Task"),
5434 _ct_bs(ct_bs), _g1h(g1h) { }
5436 void work(int i) {
5437 HeapRegion* r;
5438 while (r = _g1h->pop_dirty_cards_region()) {
5439 clear_cards(r);
5440 }
5441 }
5443 void clear_cards(HeapRegion* r) {
5444 // Cards of the survivors should have already been dirtied.
5445 if (!r->is_survivor()) {
5446 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5447 }
5448 }
5449 };
5451 #ifndef PRODUCT
5452 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5453 G1CollectedHeap* _g1h;
5454 CardTableModRefBS* _ct_bs;
5455 public:
5456 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5457 : _g1h(g1h), _ct_bs(ct_bs) { }
5458 virtual bool doHeapRegion(HeapRegion* r) {
5459 if (r->is_survivor()) {
5460 _g1h->verify_dirty_region(r);
5461 } else {
5462 _g1h->verify_not_dirty_region(r);
5463 }
5464 return false;
5465 }
5466 };
5468 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5469 // All of the region should be clean.
5470 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5471 MemRegion mr(hr->bottom(), hr->end());
5472 ct_bs->verify_not_dirty_region(mr);
5473 }
5475 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5476 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5477 // dirty allocated blocks as they allocate them. The thread that
5478 // retires each region and replaces it with a new one will do a
5479 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5480 // not dirty that area (one less thing to have to do while holding
5481 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5482 // is dirty.
5483 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5484 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5485 ct_bs->verify_dirty_region(mr);
5486 }
5488 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5489 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5490 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5491 verify_dirty_region(hr);
5492 }
5493 }
5495 void G1CollectedHeap::verify_dirty_young_regions() {
5496 verify_dirty_young_list(_young_list->first_region());
5497 verify_dirty_young_list(_young_list->first_survivor_region());
5498 }
5499 #endif
5501 void G1CollectedHeap::cleanUpCardTable() {
5502 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5503 double start = os::elapsedTime();
5505 {
5506 // Iterate over the dirty cards region list.
5507 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5509 if (ParallelGCThreads > 0) {
5510 set_par_threads(workers()->total_workers());
5511 workers()->run_task(&cleanup_task);
5512 set_par_threads(0);
5513 } else {
5514 while (_dirty_cards_region_list) {
5515 HeapRegion* r = _dirty_cards_region_list;
5516 cleanup_task.clear_cards(r);
5517 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5518 if (_dirty_cards_region_list == r) {
5519 // The last region.
5520 _dirty_cards_region_list = NULL;
5521 }
5522 r->set_next_dirty_cards_region(NULL);
5523 }
5524 }
5525 #ifndef PRODUCT
5526 if (G1VerifyCTCleanup || VerifyAfterGC) {
5527 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5528 heap_region_iterate(&cleanup_verifier);
5529 }
5530 #endif
5531 }
5533 double elapsed = os::elapsedTime() - start;
5534 g1_policy()->record_clear_ct_time(elapsed * 1000.0);
5535 }
5537 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5538 size_t pre_used = 0;
5539 FreeRegionList local_free_list("Local List for CSet Freeing");
5541 double young_time_ms = 0.0;
5542 double non_young_time_ms = 0.0;
5544 // Since the collection set is a superset of the the young list,
5545 // all we need to do to clear the young list is clear its
5546 // head and length, and unlink any young regions in the code below
5547 _young_list->clear();
5549 G1CollectorPolicy* policy = g1_policy();
5551 double start_sec = os::elapsedTime();
5552 bool non_young = true;
5554 HeapRegion* cur = cs_head;
5555 int age_bound = -1;
5556 size_t rs_lengths = 0;
5558 while (cur != NULL) {
5559 assert(!is_on_master_free_list(cur), "sanity");
5561 if (non_young) {
5562 if (cur->is_young()) {
5563 double end_sec = os::elapsedTime();
5564 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5565 non_young_time_ms += elapsed_ms;
5567 start_sec = os::elapsedTime();
5568 non_young = false;
5569 }
5570 } else {
5571 double end_sec = os::elapsedTime();
5572 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5573 young_time_ms += elapsed_ms;
5575 start_sec = os::elapsedTime();
5576 non_young = true;
5577 }
5579 rs_lengths += cur->rem_set()->occupied();
5581 HeapRegion* next = cur->next_in_collection_set();
5582 assert(cur->in_collection_set(), "bad CS");
5583 cur->set_next_in_collection_set(NULL);
5584 cur->set_in_collection_set(false);
5586 if (cur->is_young()) {
5587 int index = cur->young_index_in_cset();
5588 guarantee( index != -1, "invariant" );
5589 guarantee( (size_t)index < policy->young_cset_length(), "invariant" );
5590 size_t words_survived = _surviving_young_words[index];
5591 cur->record_surv_words_in_group(words_survived);
5593 // At this point the we have 'popped' cur from the collection set
5594 // (linked via next_in_collection_set()) but it is still in the
5595 // young list (linked via next_young_region()). Clear the
5596 // _next_young_region field.
5597 cur->set_next_young_region(NULL);
5598 } else {
5599 int index = cur->young_index_in_cset();
5600 guarantee( index == -1, "invariant" );
5601 }
5603 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5604 (!cur->is_young() && cur->young_index_in_cset() == -1),
5605 "invariant" );
5607 if (!cur->evacuation_failed()) {
5608 // And the region is empty.
5609 assert(!cur->is_empty(), "Should not have empty regions in a CS.");
5610 free_region(cur, &pre_used, &local_free_list, false /* par */);
5611 } else {
5612 cur->uninstall_surv_rate_group();
5613 if (cur->is_young())
5614 cur->set_young_index_in_cset(-1);
5615 cur->set_not_young();
5616 cur->set_evacuation_failed(false);
5617 }
5618 cur = next;
5619 }
5621 policy->record_max_rs_lengths(rs_lengths);
5622 policy->cset_regions_freed();
5624 double end_sec = os::elapsedTime();
5625 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5626 if (non_young)
5627 non_young_time_ms += elapsed_ms;
5628 else
5629 young_time_ms += elapsed_ms;
5631 update_sets_after_freeing_regions(pre_used, &local_free_list,
5632 NULL /* humongous_proxy_set */,
5633 false /* par */);
5634 policy->record_young_free_cset_time_ms(young_time_ms);
5635 policy->record_non_young_free_cset_time_ms(non_young_time_ms);
5636 }
5638 // This routine is similar to the above but does not record
5639 // any policy statistics or update free lists; we are abandoning
5640 // the current incremental collection set in preparation of a
5641 // full collection. After the full GC we will start to build up
5642 // the incremental collection set again.
5643 // This is only called when we're doing a full collection
5644 // and is immediately followed by the tearing down of the young list.
5646 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
5647 HeapRegion* cur = cs_head;
5649 while (cur != NULL) {
5650 HeapRegion* next = cur->next_in_collection_set();
5651 assert(cur->in_collection_set(), "bad CS");
5652 cur->set_next_in_collection_set(NULL);
5653 cur->set_in_collection_set(false);
5654 cur->set_young_index_in_cset(-1);
5655 cur = next;
5656 }
5657 }
5659 void G1CollectedHeap::set_free_regions_coming() {
5660 if (G1ConcRegionFreeingVerbose) {
5661 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5662 "setting free regions coming");
5663 }
5665 assert(!free_regions_coming(), "pre-condition");
5666 _free_regions_coming = true;
5667 }
5669 void G1CollectedHeap::reset_free_regions_coming() {
5670 {
5671 assert(free_regions_coming(), "pre-condition");
5672 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5673 _free_regions_coming = false;
5674 SecondaryFreeList_lock->notify_all();
5675 }
5677 if (G1ConcRegionFreeingVerbose) {
5678 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5679 "reset free regions coming");
5680 }
5681 }
5683 void G1CollectedHeap::wait_while_free_regions_coming() {
5684 // Most of the time we won't have to wait, so let's do a quick test
5685 // first before we take the lock.
5686 if (!free_regions_coming()) {
5687 return;
5688 }
5690 if (G1ConcRegionFreeingVerbose) {
5691 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5692 "waiting for free regions");
5693 }
5695 {
5696 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5697 while (free_regions_coming()) {
5698 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
5699 }
5700 }
5702 if (G1ConcRegionFreeingVerbose) {
5703 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5704 "done waiting for free regions");
5705 }
5706 }
5708 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
5709 assert(heap_lock_held_for_gc(),
5710 "the heap lock should already be held by or for this thread");
5711 _young_list->push_region(hr);
5712 g1_policy()->set_region_short_lived(hr);
5713 }
5715 class NoYoungRegionsClosure: public HeapRegionClosure {
5716 private:
5717 bool _success;
5718 public:
5719 NoYoungRegionsClosure() : _success(true) { }
5720 bool doHeapRegion(HeapRegion* r) {
5721 if (r->is_young()) {
5722 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
5723 r->bottom(), r->end());
5724 _success = false;
5725 }
5726 return false;
5727 }
5728 bool success() { return _success; }
5729 };
5731 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
5732 bool ret = _young_list->check_list_empty(check_sample);
5734 if (check_heap) {
5735 NoYoungRegionsClosure closure;
5736 heap_region_iterate(&closure);
5737 ret = ret && closure.success();
5738 }
5740 return ret;
5741 }
5743 void G1CollectedHeap::empty_young_list() {
5744 assert(heap_lock_held_for_gc(),
5745 "the heap lock should already be held by or for this thread");
5747 _young_list->empty_list();
5748 }
5750 // Done at the start of full GC.
5751 void G1CollectedHeap::tear_down_region_lists() {
5752 _free_list.remove_all();
5753 }
5755 class RegionResetter: public HeapRegionClosure {
5756 G1CollectedHeap* _g1h;
5757 FreeRegionList _local_free_list;
5759 public:
5760 RegionResetter() : _g1h(G1CollectedHeap::heap()),
5761 _local_free_list("Local Free List for RegionResetter") { }
5763 bool doHeapRegion(HeapRegion* r) {
5764 if (r->continuesHumongous()) return false;
5765 if (r->top() > r->bottom()) {
5766 if (r->top() < r->end()) {
5767 Copy::fill_to_words(r->top(),
5768 pointer_delta(r->end(), r->top()));
5769 }
5770 } else {
5771 assert(r->is_empty(), "tautology");
5772 _local_free_list.add_as_tail(r);
5773 }
5774 return false;
5775 }
5777 void update_free_lists() {
5778 _g1h->update_sets_after_freeing_regions(0, &_local_free_list, NULL,
5779 false /* par */);
5780 }
5781 };
5783 // Done at the end of full GC.
5784 void G1CollectedHeap::rebuild_region_lists() {
5785 // This needs to go at the end of the full GC.
5786 RegionResetter rs;
5787 heap_region_iterate(&rs);
5788 rs.update_free_lists();
5789 }
5791 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
5792 _refine_cte_cl->set_concurrent(concurrent);
5793 }
5795 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
5796 HeapRegion* hr = heap_region_containing(p);
5797 if (hr == NULL) {
5798 return is_in_permanent(p);
5799 } else {
5800 return hr->is_in(p);
5801 }
5802 }
5804 // Methods for the mutator alloc region
5806 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
5807 bool force) {
5808 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5809 assert(!force || g1_policy()->can_expand_young_list(),
5810 "if force is true we should be able to expand the young list");
5811 bool young_list_full = g1_policy()->is_young_list_full();
5812 if (force || !young_list_full) {
5813 HeapRegion* new_alloc_region = new_region(word_size,
5814 false /* do_expand */);
5815 if (new_alloc_region != NULL) {
5816 g1_policy()->update_region_num(true /* next_is_young */);
5817 set_region_short_lived_locked(new_alloc_region);
5818 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
5819 return new_alloc_region;
5820 }
5821 }
5822 return NULL;
5823 }
5825 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
5826 size_t allocated_bytes) {
5827 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5828 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
5830 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
5831 _summary_bytes_used += allocated_bytes;
5832 _hr_printer.retire(alloc_region);
5833 // We update the eden sizes here, when the region is retired,
5834 // instead of when it's allocated, since this is the point that its
5835 // used space has been recored in _summary_bytes_used.
5836 g1mm()->update_eden_size();
5837 }
5839 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
5840 bool force) {
5841 return _g1h->new_mutator_alloc_region(word_size, force);
5842 }
5844 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
5845 size_t allocated_bytes) {
5846 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
5847 }
5849 // Methods for the GC alloc regions
5851 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
5852 size_t count,
5853 GCAllocPurpose ap) {
5854 assert(FreeList_lock->owned_by_self(), "pre-condition");
5856 if (count < g1_policy()->max_regions(ap)) {
5857 HeapRegion* new_alloc_region = new_region(word_size,
5858 true /* do_expand */);
5859 if (new_alloc_region != NULL) {
5860 // We really only need to do this for old regions given that we
5861 // should never scan survivors. But it doesn't hurt to do it
5862 // for survivors too.
5863 new_alloc_region->set_saved_mark();
5864 if (ap == GCAllocForSurvived) {
5865 new_alloc_region->set_survivor();
5866 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
5867 } else {
5868 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
5869 }
5870 return new_alloc_region;
5871 } else {
5872 g1_policy()->note_alloc_region_limit_reached(ap);
5873 }
5874 }
5875 return NULL;
5876 }
5878 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
5879 size_t allocated_bytes,
5880 GCAllocPurpose ap) {
5881 alloc_region->note_end_of_copying();
5882 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
5883 if (ap == GCAllocForSurvived) {
5884 young_list()->add_survivor_region(alloc_region);
5885 }
5886 _hr_printer.retire(alloc_region);
5887 }
5889 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
5890 bool force) {
5891 assert(!force, "not supported for GC alloc regions");
5892 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
5893 }
5895 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
5896 size_t allocated_bytes) {
5897 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
5898 GCAllocForSurvived);
5899 }
5901 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
5902 bool force) {
5903 assert(!force, "not supported for GC alloc regions");
5904 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
5905 }
5907 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
5908 size_t allocated_bytes) {
5909 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
5910 GCAllocForTenured);
5911 }
5912 // Heap region set verification
5914 class VerifyRegionListsClosure : public HeapRegionClosure {
5915 private:
5916 HumongousRegionSet* _humongous_set;
5917 FreeRegionList* _free_list;
5918 size_t _region_count;
5920 public:
5921 VerifyRegionListsClosure(HumongousRegionSet* humongous_set,
5922 FreeRegionList* free_list) :
5923 _humongous_set(humongous_set), _free_list(free_list),
5924 _region_count(0) { }
5926 size_t region_count() { return _region_count; }
5928 bool doHeapRegion(HeapRegion* hr) {
5929 _region_count += 1;
5931 if (hr->continuesHumongous()) {
5932 return false;
5933 }
5935 if (hr->is_young()) {
5936 // TODO
5937 } else if (hr->startsHumongous()) {
5938 _humongous_set->verify_next_region(hr);
5939 } else if (hr->is_empty()) {
5940 _free_list->verify_next_region(hr);
5941 }
5942 return false;
5943 }
5944 };
5946 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index,
5947 HeapWord* bottom) {
5948 HeapWord* end = bottom + HeapRegion::GrainWords;
5949 MemRegion mr(bottom, end);
5950 assert(_g1_reserved.contains(mr), "invariant");
5951 // This might return NULL if the allocation fails
5952 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
5953 }
5955 void G1CollectedHeap::verify_region_sets() {
5956 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5958 // First, check the explicit lists.
5959 _free_list.verify();
5960 {
5961 // Given that a concurrent operation might be adding regions to
5962 // the secondary free list we have to take the lock before
5963 // verifying it.
5964 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5965 _secondary_free_list.verify();
5966 }
5967 _humongous_set.verify();
5969 // If a concurrent region freeing operation is in progress it will
5970 // be difficult to correctly attributed any free regions we come
5971 // across to the correct free list given that they might belong to
5972 // one of several (free_list, secondary_free_list, any local lists,
5973 // etc.). So, if that's the case we will skip the rest of the
5974 // verification operation. Alternatively, waiting for the concurrent
5975 // operation to complete will have a non-trivial effect on the GC's
5976 // operation (no concurrent operation will last longer than the
5977 // interval between two calls to verification) and it might hide
5978 // any issues that we would like to catch during testing.
5979 if (free_regions_coming()) {
5980 return;
5981 }
5983 // Make sure we append the secondary_free_list on the free_list so
5984 // that all free regions we will come across can be safely
5985 // attributed to the free_list.
5986 append_secondary_free_list_if_not_empty_with_lock();
5988 // Finally, make sure that the region accounting in the lists is
5989 // consistent with what we see in the heap.
5990 _humongous_set.verify_start();
5991 _free_list.verify_start();
5993 VerifyRegionListsClosure cl(&_humongous_set, &_free_list);
5994 heap_region_iterate(&cl);
5996 _humongous_set.verify_end();
5997 _free_list.verify_end();
5998 }