Print this page
rev 2722 : 7095194: G1: HeapRegion::GrainBytes, GrainWords, and CardsPerRegion should be size_t
Summary: Declare GrainBytes, GrainWords, and CardsPerRegion as size_t.
Reviewed-by:
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp
+++ new/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp
1 1 /*
2 2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "code/icBuffer.hpp"
27 27 #include "gc_implementation/g1/bufferingOopClosure.hpp"
28 28 #include "gc_implementation/g1/concurrentG1Refine.hpp"
29 29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
30 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
32 32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
33 33 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
34 34 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
35 35 #include "gc_implementation/g1/g1MarkSweep.hpp"
36 36 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
37 37 #include "gc_implementation/g1/g1RemSet.inline.hpp"
38 38 #include "gc_implementation/g1/heapRegionRemSet.hpp"
39 39 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
40 40 #include "gc_implementation/g1/vm_operations_g1.hpp"
41 41 #include "gc_implementation/shared/isGCActiveMark.hpp"
42 42 #include "memory/gcLocker.inline.hpp"
43 43 #include "memory/genOopClosures.inline.hpp"
44 44 #include "memory/generationSpec.hpp"
45 45 #include "memory/referenceProcessor.hpp"
46 46 #include "oops/oop.inline.hpp"
47 47 #include "oops/oop.pcgc.inline.hpp"
48 48 #include "runtime/aprofiler.hpp"
49 49 #include "runtime/vmThread.hpp"
50 50
51 51 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
52 52
53 53 // turn it on so that the contents of the young list (scan-only /
54 54 // to-be-collected) are printed at "strategic" points before / during
55 55 // / after the collection --- this is useful for debugging
56 56 #define YOUNG_LIST_VERBOSE 0
57 57 // CURRENT STATUS
58 58 // This file is under construction. Search for "FIXME".
59 59
60 60 // INVARIANTS/NOTES
61 61 //
62 62 // All allocation activity covered by the G1CollectedHeap interface is
63 63 // serialized by acquiring the HeapLock. This happens in mem_allocate
64 64 // and allocate_new_tlab, which are the "entry" points to the
65 65 // allocation code from the rest of the JVM. (Note that this does not
66 66 // apply to TLAB allocation, which is not part of this interface: it
67 67 // is done by clients of this interface.)
68 68
69 69 // Local to this file.
70 70
71 71 class RefineCardTableEntryClosure: public CardTableEntryClosure {
72 72 SuspendibleThreadSet* _sts;
73 73 G1RemSet* _g1rs;
74 74 ConcurrentG1Refine* _cg1r;
75 75 bool _concurrent;
76 76 public:
77 77 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
78 78 G1RemSet* g1rs,
79 79 ConcurrentG1Refine* cg1r) :
80 80 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
81 81 {}
82 82 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
83 83 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
84 84 // This path is executed by the concurrent refine or mutator threads,
85 85 // concurrently, and so we do not care if card_ptr contains references
86 86 // that point into the collection set.
87 87 assert(!oops_into_cset, "should be");
88 88
89 89 if (_concurrent && _sts->should_yield()) {
90 90 // Caller will actually yield.
91 91 return false;
92 92 }
93 93 // Otherwise, we finished successfully; return true.
94 94 return true;
95 95 }
96 96 void set_concurrent(bool b) { _concurrent = b; }
97 97 };
98 98
99 99
100 100 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
101 101 int _calls;
102 102 G1CollectedHeap* _g1h;
103 103 CardTableModRefBS* _ctbs;
104 104 int _histo[256];
105 105 public:
106 106 ClearLoggedCardTableEntryClosure() :
107 107 _calls(0)
108 108 {
109 109 _g1h = G1CollectedHeap::heap();
110 110 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
111 111 for (int i = 0; i < 256; i++) _histo[i] = 0;
112 112 }
113 113 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
114 114 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
115 115 _calls++;
116 116 unsigned char* ujb = (unsigned char*)card_ptr;
117 117 int ind = (int)(*ujb);
118 118 _histo[ind]++;
119 119 *card_ptr = -1;
120 120 }
121 121 return true;
122 122 }
123 123 int calls() { return _calls; }
124 124 void print_histo() {
125 125 gclog_or_tty->print_cr("Card table value histogram:");
126 126 for (int i = 0; i < 256; i++) {
127 127 if (_histo[i] != 0) {
128 128 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
129 129 }
130 130 }
131 131 }
132 132 };
133 133
134 134 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
135 135 int _calls;
136 136 G1CollectedHeap* _g1h;
137 137 CardTableModRefBS* _ctbs;
138 138 public:
139 139 RedirtyLoggedCardTableEntryClosure() :
140 140 _calls(0)
141 141 {
142 142 _g1h = G1CollectedHeap::heap();
143 143 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
144 144 }
145 145 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
146 146 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
147 147 _calls++;
148 148 *card_ptr = 0;
149 149 }
150 150 return true;
151 151 }
152 152 int calls() { return _calls; }
153 153 };
154 154
155 155 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
156 156 public:
157 157 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
158 158 *card_ptr = CardTableModRefBS::dirty_card_val();
159 159 return true;
160 160 }
161 161 };
162 162
163 163 YoungList::YoungList(G1CollectedHeap* g1h)
164 164 : _g1h(g1h), _head(NULL),
165 165 _length(0),
166 166 _last_sampled_rs_lengths(0),
167 167 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0)
168 168 {
169 169 guarantee( check_list_empty(false), "just making sure..." );
170 170 }
171 171
172 172 void YoungList::push_region(HeapRegion *hr) {
173 173 assert(!hr->is_young(), "should not already be young");
174 174 assert(hr->get_next_young_region() == NULL, "cause it should!");
175 175
176 176 hr->set_next_young_region(_head);
177 177 _head = hr;
178 178
179 179 hr->set_young();
180 180 double yg_surv_rate = _g1h->g1_policy()->predict_yg_surv_rate((int)_length);
181 181 ++_length;
182 182 }
183 183
184 184 void YoungList::add_survivor_region(HeapRegion* hr) {
185 185 assert(hr->is_survivor(), "should be flagged as survivor region");
186 186 assert(hr->get_next_young_region() == NULL, "cause it should!");
187 187
188 188 hr->set_next_young_region(_survivor_head);
189 189 if (_survivor_head == NULL) {
190 190 _survivor_tail = hr;
191 191 }
192 192 _survivor_head = hr;
193 193
194 194 ++_survivor_length;
195 195 }
196 196
197 197 void YoungList::empty_list(HeapRegion* list) {
198 198 while (list != NULL) {
199 199 HeapRegion* next = list->get_next_young_region();
200 200 list->set_next_young_region(NULL);
201 201 list->uninstall_surv_rate_group();
202 202 list->set_not_young();
203 203 list = next;
204 204 }
205 205 }
206 206
207 207 void YoungList::empty_list() {
208 208 assert(check_list_well_formed(), "young list should be well formed");
209 209
210 210 empty_list(_head);
211 211 _head = NULL;
212 212 _length = 0;
213 213
214 214 empty_list(_survivor_head);
215 215 _survivor_head = NULL;
216 216 _survivor_tail = NULL;
217 217 _survivor_length = 0;
218 218
219 219 _last_sampled_rs_lengths = 0;
220 220
221 221 assert(check_list_empty(false), "just making sure...");
222 222 }
223 223
224 224 bool YoungList::check_list_well_formed() {
225 225 bool ret = true;
226 226
227 227 size_t length = 0;
228 228 HeapRegion* curr = _head;
229 229 HeapRegion* last = NULL;
230 230 while (curr != NULL) {
231 231 if (!curr->is_young()) {
232 232 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
233 233 "incorrectly tagged (y: %d, surv: %d)",
234 234 curr->bottom(), curr->end(),
235 235 curr->is_young(), curr->is_survivor());
236 236 ret = false;
237 237 }
238 238 ++length;
239 239 last = curr;
240 240 curr = curr->get_next_young_region();
241 241 }
242 242 ret = ret && (length == _length);
243 243
244 244 if (!ret) {
245 245 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
246 246 gclog_or_tty->print_cr("### list has %d entries, _length is %d",
247 247 length, _length);
248 248 }
249 249
250 250 return ret;
251 251 }
252 252
253 253 bool YoungList::check_list_empty(bool check_sample) {
254 254 bool ret = true;
255 255
256 256 if (_length != 0) {
257 257 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %d",
258 258 _length);
259 259 ret = false;
260 260 }
261 261 if (check_sample && _last_sampled_rs_lengths != 0) {
262 262 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
263 263 ret = false;
264 264 }
265 265 if (_head != NULL) {
266 266 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
267 267 ret = false;
268 268 }
269 269 if (!ret) {
270 270 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
271 271 }
272 272
273 273 return ret;
274 274 }
275 275
276 276 void
277 277 YoungList::rs_length_sampling_init() {
278 278 _sampled_rs_lengths = 0;
279 279 _curr = _head;
280 280 }
281 281
282 282 bool
283 283 YoungList::rs_length_sampling_more() {
284 284 return _curr != NULL;
285 285 }
286 286
287 287 void
288 288 YoungList::rs_length_sampling_next() {
289 289 assert( _curr != NULL, "invariant" );
290 290 size_t rs_length = _curr->rem_set()->occupied();
291 291
292 292 _sampled_rs_lengths += rs_length;
293 293
294 294 // The current region may not yet have been added to the
295 295 // incremental collection set (it gets added when it is
296 296 // retired as the current allocation region).
297 297 if (_curr->in_collection_set()) {
298 298 // Update the collection set policy information for this region
299 299 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
300 300 }
301 301
302 302 _curr = _curr->get_next_young_region();
303 303 if (_curr == NULL) {
304 304 _last_sampled_rs_lengths = _sampled_rs_lengths;
305 305 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
306 306 }
307 307 }
308 308
309 309 void
310 310 YoungList::reset_auxilary_lists() {
311 311 guarantee( is_empty(), "young list should be empty" );
312 312 assert(check_list_well_formed(), "young list should be well formed");
313 313
314 314 // Add survivor regions to SurvRateGroup.
315 315 _g1h->g1_policy()->note_start_adding_survivor_regions();
316 316 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
317 317
318 318 for (HeapRegion* curr = _survivor_head;
319 319 curr != NULL;
320 320 curr = curr->get_next_young_region()) {
321 321 _g1h->g1_policy()->set_region_survivors(curr);
322 322
323 323 // The region is a non-empty survivor so let's add it to
324 324 // the incremental collection set for the next evacuation
325 325 // pause.
326 326 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
327 327 }
328 328 _g1h->g1_policy()->note_stop_adding_survivor_regions();
329 329
330 330 _head = _survivor_head;
331 331 _length = _survivor_length;
332 332 if (_survivor_head != NULL) {
333 333 assert(_survivor_tail != NULL, "cause it shouldn't be");
334 334 assert(_survivor_length > 0, "invariant");
335 335 _survivor_tail->set_next_young_region(NULL);
336 336 }
337 337
338 338 // Don't clear the survivor list handles until the start of
339 339 // the next evacuation pause - we need it in order to re-tag
340 340 // the survivor regions from this evacuation pause as 'young'
341 341 // at the start of the next.
342 342
343 343 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
344 344
345 345 assert(check_list_well_formed(), "young list should be well formed");
346 346 }
347 347
348 348 void YoungList::print() {
349 349 HeapRegion* lists[] = {_head, _survivor_head};
350 350 const char* names[] = {"YOUNG", "SURVIVOR"};
351 351
352 352 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
353 353 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
354 354 HeapRegion *curr = lists[list];
355 355 if (curr == NULL)
356 356 gclog_or_tty->print_cr(" empty");
357 357 while (curr != NULL) {
358 358 gclog_or_tty->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
359 359 "age: %4d, y: %d, surv: %d",
360 360 curr->bottom(), curr->end(),
361 361 curr->top(),
362 362 curr->prev_top_at_mark_start(),
363 363 curr->next_top_at_mark_start(),
364 364 curr->top_at_conc_mark_count(),
365 365 curr->age_in_surv_rate_group_cond(),
366 366 curr->is_young(),
367 367 curr->is_survivor());
368 368 curr = curr->get_next_young_region();
369 369 }
370 370 }
371 371
372 372 gclog_or_tty->print_cr("");
373 373 }
374 374
375 375 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
376 376 {
377 377 // Claim the right to put the region on the dirty cards region list
378 378 // by installing a self pointer.
379 379 HeapRegion* next = hr->get_next_dirty_cards_region();
380 380 if (next == NULL) {
381 381 HeapRegion* res = (HeapRegion*)
382 382 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
383 383 NULL);
384 384 if (res == NULL) {
385 385 HeapRegion* head;
386 386 do {
387 387 // Put the region to the dirty cards region list.
388 388 head = _dirty_cards_region_list;
389 389 next = (HeapRegion*)
390 390 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
391 391 if (next == head) {
392 392 assert(hr->get_next_dirty_cards_region() == hr,
393 393 "hr->get_next_dirty_cards_region() != hr");
394 394 if (next == NULL) {
395 395 // The last region in the list points to itself.
396 396 hr->set_next_dirty_cards_region(hr);
397 397 } else {
398 398 hr->set_next_dirty_cards_region(next);
399 399 }
400 400 }
401 401 } while (next != head);
402 402 }
403 403 }
404 404 }
405 405
406 406 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
407 407 {
408 408 HeapRegion* head;
409 409 HeapRegion* hr;
410 410 do {
411 411 head = _dirty_cards_region_list;
412 412 if (head == NULL) {
413 413 return NULL;
414 414 }
415 415 HeapRegion* new_head = head->get_next_dirty_cards_region();
416 416 if (head == new_head) {
417 417 // The last region.
418 418 new_head = NULL;
419 419 }
420 420 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
421 421 head);
422 422 } while (hr != head);
423 423 assert(hr != NULL, "invariant");
424 424 hr->set_next_dirty_cards_region(NULL);
425 425 return hr;
426 426 }
427 427
428 428 void G1CollectedHeap::stop_conc_gc_threads() {
429 429 _cg1r->stop();
430 430 _cmThread->stop();
431 431 }
432 432
433 433 #ifdef ASSERT
434 434 // A region is added to the collection set as it is retired
435 435 // so an address p can point to a region which will be in the
436 436 // collection set but has not yet been retired. This method
437 437 // therefore is only accurate during a GC pause after all
438 438 // regions have been retired. It is used for debugging
439 439 // to check if an nmethod has references to objects that can
440 440 // be move during a partial collection. Though it can be
441 441 // inaccurate, it is sufficient for G1 because the conservative
442 442 // implementation of is_scavengable() for G1 will indicate that
443 443 // all nmethods must be scanned during a partial collection.
444 444 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
445 445 HeapRegion* hr = heap_region_containing(p);
446 446 return hr != NULL && hr->in_collection_set();
447 447 }
448 448 #endif
449 449
450 450 // Returns true if the reference points to an object that
451 451 // can move in an incremental collecction.
452 452 bool G1CollectedHeap::is_scavengable(const void* p) {
453 453 G1CollectedHeap* g1h = G1CollectedHeap::heap();
454 454 G1CollectorPolicy* g1p = g1h->g1_policy();
455 455 HeapRegion* hr = heap_region_containing(p);
456 456 if (hr == NULL) {
457 457 // perm gen (or null)
458 458 return false;
459 459 } else {
460 460 return !hr->isHumongous();
461 461 }
462 462 }
463 463
464 464 void G1CollectedHeap::check_ct_logs_at_safepoint() {
465 465 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
466 466 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
467 467
468 468 // Count the dirty cards at the start.
469 469 CountNonCleanMemRegionClosure count1(this);
470 470 ct_bs->mod_card_iterate(&count1);
471 471 int orig_count = count1.n();
472 472
473 473 // First clear the logged cards.
474 474 ClearLoggedCardTableEntryClosure clear;
475 475 dcqs.set_closure(&clear);
476 476 dcqs.apply_closure_to_all_completed_buffers();
477 477 dcqs.iterate_closure_all_threads(false);
478 478 clear.print_histo();
479 479
480 480 // Now ensure that there's no dirty cards.
481 481 CountNonCleanMemRegionClosure count2(this);
482 482 ct_bs->mod_card_iterate(&count2);
483 483 if (count2.n() != 0) {
484 484 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
485 485 count2.n(), orig_count);
486 486 }
487 487 guarantee(count2.n() == 0, "Card table should be clean.");
488 488
489 489 RedirtyLoggedCardTableEntryClosure redirty;
490 490 JavaThread::dirty_card_queue_set().set_closure(&redirty);
491 491 dcqs.apply_closure_to_all_completed_buffers();
492 492 dcqs.iterate_closure_all_threads(false);
493 493 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
494 494 clear.calls(), orig_count);
495 495 guarantee(redirty.calls() == clear.calls(),
496 496 "Or else mechanism is broken.");
497 497
498 498 CountNonCleanMemRegionClosure count3(this);
499 499 ct_bs->mod_card_iterate(&count3);
500 500 if (count3.n() != orig_count) {
501 501 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
502 502 orig_count, count3.n());
503 503 guarantee(count3.n() >= orig_count, "Should have restored them all.");
504 504 }
505 505
506 506 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
507 507 }
508 508
509 509 // Private class members.
510 510
511 511 G1CollectedHeap* G1CollectedHeap::_g1h;
512 512
513 513 // Private methods.
514 514
515 515 HeapRegion*
516 516 G1CollectedHeap::new_region_try_secondary_free_list() {
517 517 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
518 518 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
519 519 if (!_secondary_free_list.is_empty()) {
520 520 if (G1ConcRegionFreeingVerbose) {
521 521 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
522 522 "secondary_free_list has "SIZE_FORMAT" entries",
523 523 _secondary_free_list.length());
524 524 }
525 525 // It looks as if there are free regions available on the
526 526 // secondary_free_list. Let's move them to the free_list and try
527 527 // again to allocate from it.
528 528 append_secondary_free_list();
529 529
530 530 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
531 531 "empty we should have moved at least one entry to the free_list");
532 532 HeapRegion* res = _free_list.remove_head();
533 533 if (G1ConcRegionFreeingVerbose) {
534 534 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
535 535 "allocated "HR_FORMAT" from secondary_free_list",
536 536 HR_FORMAT_PARAMS(res));
537 537 }
538 538 return res;
539 539 }
540 540
541 541 // Wait here until we get notifed either when (a) there are no
542 542 // more free regions coming or (b) some regions have been moved on
543 543 // the secondary_free_list.
544 544 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
↓ open down ↓ |
544 lines elided |
↑ open up ↑ |
545 545 }
546 546
547 547 if (G1ConcRegionFreeingVerbose) {
548 548 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
549 549 "could not allocate from secondary_free_list");
550 550 }
551 551 return NULL;
552 552 }
553 553
554 554 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
555 - assert(!isHumongous(word_size) ||
556 - word_size <= (size_t) HeapRegion::GrainWords,
555 + assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
557 556 "the only time we use this to allocate a humongous region is "
558 557 "when we are allocating a single humongous region");
559 558
560 559 HeapRegion* res;
561 560 if (G1StressConcRegionFreeing) {
562 561 if (!_secondary_free_list.is_empty()) {
563 562 if (G1ConcRegionFreeingVerbose) {
564 563 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
565 564 "forced to look at the secondary_free_list");
566 565 }
567 566 res = new_region_try_secondary_free_list();
568 567 if (res != NULL) {
569 568 return res;
570 569 }
571 570 }
572 571 }
573 572 res = _free_list.remove_head_or_null();
574 573 if (res == NULL) {
575 574 if (G1ConcRegionFreeingVerbose) {
576 575 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
577 576 "res == NULL, trying the secondary_free_list");
578 577 }
579 578 res = new_region_try_secondary_free_list();
580 579 }
581 580 if (res == NULL && do_expand) {
582 581 ergo_verbose1(ErgoHeapSizing,
583 582 "attempt heap expansion",
584 583 ergo_format_reason("region allocation request failed")
585 584 ergo_format_byte("allocation request"),
586 585 word_size * HeapWordSize);
587 586 if (expand(word_size * HeapWordSize)) {
588 587 // Even though the heap was expanded, it might not have reached
589 588 // the desired size. So, we cannot assume that the allocation
590 589 // will succeed.
591 590 res = _free_list.remove_head_or_null();
592 591 }
593 592 }
594 593 return res;
595 594 }
596 595
597 596 size_t G1CollectedHeap::humongous_obj_allocate_find_first(size_t num_regions,
598 597 size_t word_size) {
599 598 assert(isHumongous(word_size), "word_size should be humongous");
600 599 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
601 600
602 601 size_t first = G1_NULL_HRS_INDEX;
603 602 if (num_regions == 1) {
604 603 // Only one region to allocate, no need to go through the slower
605 604 // path. The caller will attempt the expasion if this fails, so
606 605 // let's not try to expand here too.
607 606 HeapRegion* hr = new_region(word_size, false /* do_expand */);
608 607 if (hr != NULL) {
609 608 first = hr->hrs_index();
610 609 } else {
611 610 first = G1_NULL_HRS_INDEX;
612 611 }
613 612 } else {
614 613 // We can't allocate humongous regions while cleanupComplete() is
615 614 // running, since some of the regions we find to be empty might not
616 615 // yet be added to the free list and it is not straightforward to
617 616 // know which list they are on so that we can remove them. Note
618 617 // that we only need to do this if we need to allocate more than
619 618 // one region to satisfy the current humongous allocation
620 619 // request. If we are only allocating one region we use the common
621 620 // region allocation code (see above).
622 621 wait_while_free_regions_coming();
623 622 append_secondary_free_list_if_not_empty_with_lock();
624 623
625 624 if (free_regions() >= num_regions) {
626 625 first = _hrs.find_contiguous(num_regions);
627 626 if (first != G1_NULL_HRS_INDEX) {
628 627 for (size_t i = first; i < first + num_regions; ++i) {
629 628 HeapRegion* hr = region_at(i);
630 629 assert(hr->is_empty(), "sanity");
631 630 assert(is_on_master_free_list(hr), "sanity");
632 631 hr->set_pending_removal(true);
633 632 }
634 633 _free_list.remove_all_pending(num_regions);
635 634 }
636 635 }
637 636 }
638 637 return first;
639 638 }
640 639
641 640 HeapWord*
642 641 G1CollectedHeap::humongous_obj_allocate_initialize_regions(size_t first,
643 642 size_t num_regions,
644 643 size_t word_size) {
645 644 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
646 645 assert(isHumongous(word_size), "word_size should be humongous");
647 646 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
648 647
649 648 // Index of last region in the series + 1.
650 649 size_t last = first + num_regions;
651 650
652 651 // We need to initialize the region(s) we just discovered. This is
653 652 // a bit tricky given that it can happen concurrently with
654 653 // refinement threads refining cards on these regions and
655 654 // potentially wanting to refine the BOT as they are scanning
656 655 // those cards (this can happen shortly after a cleanup; see CR
657 656 // 6991377). So we have to set up the region(s) carefully and in
658 657 // a specific order.
659 658
660 659 // The word size sum of all the regions we will allocate.
661 660 size_t word_size_sum = num_regions * HeapRegion::GrainWords;
662 661 assert(word_size <= word_size_sum, "sanity");
663 662
664 663 // This will be the "starts humongous" region.
665 664 HeapRegion* first_hr = region_at(first);
666 665 // The header of the new object will be placed at the bottom of
667 666 // the first region.
668 667 HeapWord* new_obj = first_hr->bottom();
669 668 // This will be the new end of the first region in the series that
670 669 // should also match the end of the last region in the seriers.
671 670 HeapWord* new_end = new_obj + word_size_sum;
672 671 // This will be the new top of the first region that will reflect
673 672 // this allocation.
674 673 HeapWord* new_top = new_obj + word_size;
675 674
676 675 // First, we need to zero the header of the space that we will be
677 676 // allocating. When we update top further down, some refinement
678 677 // threads might try to scan the region. By zeroing the header we
679 678 // ensure that any thread that will try to scan the region will
680 679 // come across the zero klass word and bail out.
681 680 //
682 681 // NOTE: It would not have been correct to have used
683 682 // CollectedHeap::fill_with_object() and make the space look like
684 683 // an int array. The thread that is doing the allocation will
685 684 // later update the object header to a potentially different array
686 685 // type and, for a very short period of time, the klass and length
687 686 // fields will be inconsistent. This could cause a refinement
688 687 // thread to calculate the object size incorrectly.
689 688 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
690 689
691 690 // We will set up the first region as "starts humongous". This
692 691 // will also update the BOT covering all the regions to reflect
693 692 // that there is a single object that starts at the bottom of the
694 693 // first region.
695 694 first_hr->set_startsHumongous(new_top, new_end);
696 695
697 696 // Then, if there are any, we will set up the "continues
698 697 // humongous" regions.
699 698 HeapRegion* hr = NULL;
700 699 for (size_t i = first + 1; i < last; ++i) {
701 700 hr = region_at(i);
702 701 hr->set_continuesHumongous(first_hr);
703 702 }
704 703 // If we have "continues humongous" regions (hr != NULL), then the
705 704 // end of the last one should match new_end.
706 705 assert(hr == NULL || hr->end() == new_end, "sanity");
707 706
708 707 // Up to this point no concurrent thread would have been able to
709 708 // do any scanning on any region in this series. All the top
710 709 // fields still point to bottom, so the intersection between
711 710 // [bottom,top] and [card_start,card_end] will be empty. Before we
712 711 // update the top fields, we'll do a storestore to make sure that
713 712 // no thread sees the update to top before the zeroing of the
714 713 // object header and the BOT initialization.
715 714 OrderAccess::storestore();
716 715
717 716 // Now that the BOT and the object header have been initialized,
718 717 // we can update top of the "starts humongous" region.
719 718 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
720 719 "new_top should be in this region");
721 720 first_hr->set_top(new_top);
722 721 if (_hr_printer.is_active()) {
723 722 HeapWord* bottom = first_hr->bottom();
724 723 HeapWord* end = first_hr->orig_end();
725 724 if ((first + 1) == last) {
726 725 // the series has a single humongous region
727 726 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
728 727 } else {
729 728 // the series has more than one humongous regions
730 729 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
731 730 }
732 731 }
733 732
734 733 // Now, we will update the top fields of the "continues humongous"
735 734 // regions. The reason we need to do this is that, otherwise,
736 735 // these regions would look empty and this will confuse parts of
737 736 // G1. For example, the code that looks for a consecutive number
738 737 // of empty regions will consider them empty and try to
739 738 // re-allocate them. We can extend is_empty() to also include
740 739 // !continuesHumongous(), but it is easier to just update the top
741 740 // fields here. The way we set top for all regions (i.e., top ==
742 741 // end for all regions but the last one, top == new_top for the
743 742 // last one) is actually used when we will free up the humongous
744 743 // region in free_humongous_region().
745 744 hr = NULL;
746 745 for (size_t i = first + 1; i < last; ++i) {
747 746 hr = region_at(i);
748 747 if ((i + 1) == last) {
749 748 // last continues humongous region
750 749 assert(hr->bottom() < new_top && new_top <= hr->end(),
751 750 "new_top should fall on this region");
752 751 hr->set_top(new_top);
753 752 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
754 753 } else {
755 754 // not last one
756 755 assert(new_top > hr->end(), "new_top should be above this region");
757 756 hr->set_top(hr->end());
758 757 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
759 758 }
760 759 }
761 760 // If we have continues humongous regions (hr != NULL), then the
762 761 // end of the last one should match new_end and its top should
763 762 // match new_top.
764 763 assert(hr == NULL ||
765 764 (hr->end() == new_end && hr->top() == new_top), "sanity");
766 765
767 766 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
768 767 _summary_bytes_used += first_hr->used();
769 768 _humongous_set.add(first_hr);
770 769
771 770 return new_obj;
772 771 }
773 772
774 773 // If could fit into free regions w/o expansion, try.
775 774 // Otherwise, if can expand, do so.
776 775 // Otherwise, if using ex regions might help, try with ex given back.
777 776 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
778 777 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
779 778
780 779 verify_region_sets_optional();
781 780
782 781 size_t num_regions =
783 782 round_to(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords;
784 783 size_t x_size = expansion_regions();
785 784 size_t fs = _hrs.free_suffix();
786 785 size_t first = humongous_obj_allocate_find_first(num_regions, word_size);
787 786 if (first == G1_NULL_HRS_INDEX) {
788 787 // The only thing we can do now is attempt expansion.
789 788 if (fs + x_size >= num_regions) {
790 789 // If the number of regions we're trying to allocate for this
791 790 // object is at most the number of regions in the free suffix,
792 791 // then the call to humongous_obj_allocate_find_first() above
793 792 // should have succeeded and we wouldn't be here.
794 793 //
795 794 // We should only be trying to expand when the free suffix is
796 795 // not sufficient for the object _and_ we have some expansion
797 796 // room available.
798 797 assert(num_regions > fs, "earlier allocation should have succeeded");
799 798
800 799 ergo_verbose1(ErgoHeapSizing,
801 800 "attempt heap expansion",
802 801 ergo_format_reason("humongous allocation request failed")
803 802 ergo_format_byte("allocation request"),
804 803 word_size * HeapWordSize);
805 804 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
806 805 // Even though the heap was expanded, it might not have
807 806 // reached the desired size. So, we cannot assume that the
808 807 // allocation will succeed.
809 808 first = humongous_obj_allocate_find_first(num_regions, word_size);
810 809 }
811 810 }
812 811 }
813 812
814 813 HeapWord* result = NULL;
815 814 if (first != G1_NULL_HRS_INDEX) {
816 815 result =
817 816 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
818 817 assert(result != NULL, "it should always return a valid result");
819 818
820 819 // A successful humongous object allocation changes the used space
821 820 // information of the old generation so we need to recalculate the
822 821 // sizes and update the jstat counters here.
823 822 g1mm()->update_sizes();
824 823 }
825 824
826 825 verify_region_sets_optional();
827 826
828 827 return result;
829 828 }
830 829
831 830 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
832 831 assert_heap_not_locked_and_not_at_safepoint();
833 832 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
834 833
835 834 unsigned int dummy_gc_count_before;
836 835 return attempt_allocation(word_size, &dummy_gc_count_before);
837 836 }
838 837
839 838 HeapWord*
840 839 G1CollectedHeap::mem_allocate(size_t word_size,
841 840 bool* gc_overhead_limit_was_exceeded) {
842 841 assert_heap_not_locked_and_not_at_safepoint();
843 842
844 843 // Loop until the allocation is satisified, or unsatisfied after GC.
845 844 for (int try_count = 1; /* we'll return */; try_count += 1) {
846 845 unsigned int gc_count_before;
847 846
848 847 HeapWord* result = NULL;
849 848 if (!isHumongous(word_size)) {
850 849 result = attempt_allocation(word_size, &gc_count_before);
851 850 } else {
852 851 result = attempt_allocation_humongous(word_size, &gc_count_before);
853 852 }
854 853 if (result != NULL) {
855 854 return result;
856 855 }
857 856
858 857 // Create the garbage collection operation...
859 858 VM_G1CollectForAllocation op(gc_count_before, word_size);
860 859 // ...and get the VM thread to execute it.
861 860 VMThread::execute(&op);
862 861
863 862 if (op.prologue_succeeded() && op.pause_succeeded()) {
864 863 // If the operation was successful we'll return the result even
865 864 // if it is NULL. If the allocation attempt failed immediately
866 865 // after a Full GC, it's unlikely we'll be able to allocate now.
867 866 HeapWord* result = op.result();
868 867 if (result != NULL && !isHumongous(word_size)) {
869 868 // Allocations that take place on VM operations do not do any
870 869 // card dirtying and we have to do it here. We only have to do
871 870 // this for non-humongous allocations, though.
872 871 dirty_young_block(result, word_size);
873 872 }
874 873 return result;
875 874 } else {
876 875 assert(op.result() == NULL,
877 876 "the result should be NULL if the VM op did not succeed");
878 877 }
879 878
880 879 // Give a warning if we seem to be looping forever.
881 880 if ((QueuedAllocationWarningCount > 0) &&
882 881 (try_count % QueuedAllocationWarningCount == 0)) {
883 882 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
884 883 }
885 884 }
886 885
887 886 ShouldNotReachHere();
888 887 return NULL;
889 888 }
890 889
891 890 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
892 891 unsigned int *gc_count_before_ret) {
893 892 // Make sure you read the note in attempt_allocation_humongous().
894 893
895 894 assert_heap_not_locked_and_not_at_safepoint();
896 895 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
897 896 "be called for humongous allocation requests");
898 897
899 898 // We should only get here after the first-level allocation attempt
900 899 // (attempt_allocation()) failed to allocate.
901 900
902 901 // We will loop until a) we manage to successfully perform the
903 902 // allocation or b) we successfully schedule a collection which
904 903 // fails to perform the allocation. b) is the only case when we'll
905 904 // return NULL.
906 905 HeapWord* result = NULL;
907 906 for (int try_count = 1; /* we'll return */; try_count += 1) {
908 907 bool should_try_gc;
909 908 unsigned int gc_count_before;
910 909
911 910 {
912 911 MutexLockerEx x(Heap_lock);
913 912
914 913 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
915 914 false /* bot_updates */);
916 915 if (result != NULL) {
917 916 return result;
918 917 }
919 918
920 919 // If we reach here, attempt_allocation_locked() above failed to
921 920 // allocate a new region. So the mutator alloc region should be NULL.
922 921 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
923 922
924 923 if (GC_locker::is_active_and_needs_gc()) {
925 924 if (g1_policy()->can_expand_young_list()) {
926 925 // No need for an ergo verbose message here,
927 926 // can_expand_young_list() does this when it returns true.
928 927 result = _mutator_alloc_region.attempt_allocation_force(word_size,
929 928 false /* bot_updates */);
930 929 if (result != NULL) {
931 930 return result;
932 931 }
933 932 }
934 933 should_try_gc = false;
935 934 } else {
936 935 // Read the GC count while still holding the Heap_lock.
937 936 gc_count_before = SharedHeap::heap()->total_collections();
938 937 should_try_gc = true;
939 938 }
940 939 }
941 940
942 941 if (should_try_gc) {
943 942 bool succeeded;
944 943 result = do_collection_pause(word_size, gc_count_before, &succeeded);
945 944 if (result != NULL) {
946 945 assert(succeeded, "only way to get back a non-NULL result");
947 946 return result;
948 947 }
949 948
950 949 if (succeeded) {
951 950 // If we get here we successfully scheduled a collection which
952 951 // failed to allocate. No point in trying to allocate
953 952 // further. We'll just return NULL.
954 953 MutexLockerEx x(Heap_lock);
955 954 *gc_count_before_ret = SharedHeap::heap()->total_collections();
956 955 return NULL;
957 956 }
958 957 } else {
959 958 GC_locker::stall_until_clear();
960 959 }
961 960
962 961 // We can reach here if we were unsuccessul in scheduling a
963 962 // collection (because another thread beat us to it) or if we were
964 963 // stalled due to the GC locker. In either can we should retry the
965 964 // allocation attempt in case another thread successfully
966 965 // performed a collection and reclaimed enough space. We do the
967 966 // first attempt (without holding the Heap_lock) here and the
968 967 // follow-on attempt will be at the start of the next loop
969 968 // iteration (after taking the Heap_lock).
970 969 result = _mutator_alloc_region.attempt_allocation(word_size,
971 970 false /* bot_updates */);
972 971 if (result != NULL ){
973 972 return result;
974 973 }
975 974
976 975 // Give a warning if we seem to be looping forever.
977 976 if ((QueuedAllocationWarningCount > 0) &&
978 977 (try_count % QueuedAllocationWarningCount == 0)) {
979 978 warning("G1CollectedHeap::attempt_allocation_slow() "
980 979 "retries %d times", try_count);
981 980 }
982 981 }
983 982
984 983 ShouldNotReachHere();
985 984 return NULL;
986 985 }
987 986
988 987 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
989 988 unsigned int * gc_count_before_ret) {
990 989 // The structure of this method has a lot of similarities to
991 990 // attempt_allocation_slow(). The reason these two were not merged
992 991 // into a single one is that such a method would require several "if
993 992 // allocation is not humongous do this, otherwise do that"
994 993 // conditional paths which would obscure its flow. In fact, an early
995 994 // version of this code did use a unified method which was harder to
996 995 // follow and, as a result, it had subtle bugs that were hard to
997 996 // track down. So keeping these two methods separate allows each to
998 997 // be more readable. It will be good to keep these two in sync as
999 998 // much as possible.
1000 999
1001 1000 assert_heap_not_locked_and_not_at_safepoint();
1002 1001 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1003 1002 "should only be called for humongous allocations");
1004 1003
1005 1004 // We will loop until a) we manage to successfully perform the
1006 1005 // allocation or b) we successfully schedule a collection which
1007 1006 // fails to perform the allocation. b) is the only case when we'll
1008 1007 // return NULL.
1009 1008 HeapWord* result = NULL;
1010 1009 for (int try_count = 1; /* we'll return */; try_count += 1) {
1011 1010 bool should_try_gc;
1012 1011 unsigned int gc_count_before;
1013 1012
1014 1013 {
1015 1014 MutexLockerEx x(Heap_lock);
1016 1015
1017 1016 // Given that humongous objects are not allocated in young
1018 1017 // regions, we'll first try to do the allocation without doing a
1019 1018 // collection hoping that there's enough space in the heap.
1020 1019 result = humongous_obj_allocate(word_size);
1021 1020 if (result != NULL) {
1022 1021 return result;
1023 1022 }
1024 1023
1025 1024 if (GC_locker::is_active_and_needs_gc()) {
1026 1025 should_try_gc = false;
1027 1026 } else {
1028 1027 // Read the GC count while still holding the Heap_lock.
1029 1028 gc_count_before = SharedHeap::heap()->total_collections();
1030 1029 should_try_gc = true;
1031 1030 }
1032 1031 }
1033 1032
1034 1033 if (should_try_gc) {
1035 1034 // If we failed to allocate the humongous object, we should try to
1036 1035 // do a collection pause (if we're allowed) in case it reclaims
1037 1036 // enough space for the allocation to succeed after the pause.
1038 1037
1039 1038 bool succeeded;
1040 1039 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1041 1040 if (result != NULL) {
1042 1041 assert(succeeded, "only way to get back a non-NULL result");
1043 1042 return result;
1044 1043 }
1045 1044
1046 1045 if (succeeded) {
1047 1046 // If we get here we successfully scheduled a collection which
1048 1047 // failed to allocate. No point in trying to allocate
1049 1048 // further. We'll just return NULL.
1050 1049 MutexLockerEx x(Heap_lock);
1051 1050 *gc_count_before_ret = SharedHeap::heap()->total_collections();
1052 1051 return NULL;
1053 1052 }
1054 1053 } else {
1055 1054 GC_locker::stall_until_clear();
1056 1055 }
1057 1056
1058 1057 // We can reach here if we were unsuccessul in scheduling a
1059 1058 // collection (because another thread beat us to it) or if we were
1060 1059 // stalled due to the GC locker. In either can we should retry the
1061 1060 // allocation attempt in case another thread successfully
1062 1061 // performed a collection and reclaimed enough space. Give a
1063 1062 // warning if we seem to be looping forever.
1064 1063
1065 1064 if ((QueuedAllocationWarningCount > 0) &&
1066 1065 (try_count % QueuedAllocationWarningCount == 0)) {
1067 1066 warning("G1CollectedHeap::attempt_allocation_humongous() "
1068 1067 "retries %d times", try_count);
1069 1068 }
1070 1069 }
1071 1070
1072 1071 ShouldNotReachHere();
1073 1072 return NULL;
1074 1073 }
1075 1074
1076 1075 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1077 1076 bool expect_null_mutator_alloc_region) {
1078 1077 assert_at_safepoint(true /* should_be_vm_thread */);
1079 1078 assert(_mutator_alloc_region.get() == NULL ||
1080 1079 !expect_null_mutator_alloc_region,
1081 1080 "the current alloc region was unexpectedly found to be non-NULL");
1082 1081
1083 1082 if (!isHumongous(word_size)) {
1084 1083 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1085 1084 false /* bot_updates */);
1086 1085 } else {
1087 1086 return humongous_obj_allocate(word_size);
1088 1087 }
1089 1088
1090 1089 ShouldNotReachHere();
1091 1090 }
1092 1091
1093 1092 class PostMCRemSetClearClosure: public HeapRegionClosure {
1094 1093 ModRefBarrierSet* _mr_bs;
1095 1094 public:
1096 1095 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1097 1096 bool doHeapRegion(HeapRegion* r) {
1098 1097 r->reset_gc_time_stamp();
1099 1098 if (r->continuesHumongous())
1100 1099 return false;
1101 1100 HeapRegionRemSet* hrrs = r->rem_set();
1102 1101 if (hrrs != NULL) hrrs->clear();
1103 1102 // You might think here that we could clear just the cards
1104 1103 // corresponding to the used region. But no: if we leave a dirty card
1105 1104 // in a region we might allocate into, then it would prevent that card
1106 1105 // from being enqueued, and cause it to be missed.
1107 1106 // Re: the performance cost: we shouldn't be doing full GC anyway!
1108 1107 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1109 1108 return false;
1110 1109 }
1111 1110 };
1112 1111
1113 1112
1114 1113 class PostMCRemSetInvalidateClosure: public HeapRegionClosure {
1115 1114 ModRefBarrierSet* _mr_bs;
1116 1115 public:
1117 1116 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1118 1117 bool doHeapRegion(HeapRegion* r) {
1119 1118 if (r->continuesHumongous()) return false;
1120 1119 if (r->used_region().word_size() != 0) {
1121 1120 _mr_bs->invalidate(r->used_region(), true /*whole heap*/);
1122 1121 }
1123 1122 return false;
1124 1123 }
1125 1124 };
1126 1125
1127 1126 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1128 1127 G1CollectedHeap* _g1h;
1129 1128 UpdateRSOopClosure _cl;
1130 1129 int _worker_i;
1131 1130 public:
1132 1131 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1133 1132 _cl(g1->g1_rem_set(), worker_i),
1134 1133 _worker_i(worker_i),
1135 1134 _g1h(g1)
1136 1135 { }
1137 1136
1138 1137 bool doHeapRegion(HeapRegion* r) {
1139 1138 if (!r->continuesHumongous()) {
1140 1139 _cl.set_from(r);
1141 1140 r->oop_iterate(&_cl);
1142 1141 }
1143 1142 return false;
1144 1143 }
1145 1144 };
1146 1145
1147 1146 class ParRebuildRSTask: public AbstractGangTask {
1148 1147 G1CollectedHeap* _g1;
1149 1148 public:
1150 1149 ParRebuildRSTask(G1CollectedHeap* g1)
1151 1150 : AbstractGangTask("ParRebuildRSTask"),
1152 1151 _g1(g1)
1153 1152 { }
1154 1153
1155 1154 void work(int i) {
1156 1155 RebuildRSOutOfRegionClosure rebuild_rs(_g1, i);
1157 1156 _g1->heap_region_par_iterate_chunked(&rebuild_rs, i,
1158 1157 HeapRegion::RebuildRSClaimValue);
1159 1158 }
1160 1159 };
1161 1160
1162 1161 class PostCompactionPrinterClosure: public HeapRegionClosure {
↓ open down ↓ |
596 lines elided |
↑ open up ↑ |
1163 1162 private:
1164 1163 G1HRPrinter* _hr_printer;
1165 1164 public:
1166 1165 bool doHeapRegion(HeapRegion* hr) {
1167 1166 assert(!hr->is_young(), "not expecting to find young regions");
1168 1167 // We only generate output for non-empty regions.
1169 1168 if (!hr->is_empty()) {
1170 1169 if (!hr->isHumongous()) {
1171 1170 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1172 1171 } else if (hr->startsHumongous()) {
1173 - if (hr->capacity() == (size_t) HeapRegion::GrainBytes) {
1172 + if (hr->capacity() == HeapRegion::GrainBytes) {
1174 1173 // single humongous region
1175 1174 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1176 1175 } else {
1177 1176 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1178 1177 }
1179 1178 } else {
1180 1179 assert(hr->continuesHumongous(), "only way to get here");
1181 1180 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1182 1181 }
1183 1182 }
1184 1183 return false;
1185 1184 }
1186 1185
1187 1186 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1188 1187 : _hr_printer(hr_printer) { }
1189 1188 };
1190 1189
1191 1190 bool G1CollectedHeap::do_collection(bool explicit_gc,
1192 1191 bool clear_all_soft_refs,
1193 1192 size_t word_size) {
1194 1193 assert_at_safepoint(true /* should_be_vm_thread */);
1195 1194
1196 1195 if (GC_locker::check_active_before_gc()) {
1197 1196 return false;
1198 1197 }
1199 1198
1200 1199 SvcGCMarker sgcm(SvcGCMarker::FULL);
1201 1200 ResourceMark rm;
1202 1201
1203 1202 if (PrintHeapAtGC) {
1204 1203 Universe::print_heap_before_gc();
1205 1204 }
1206 1205
1207 1206 verify_region_sets_optional();
1208 1207
1209 1208 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1210 1209 collector_policy()->should_clear_all_soft_refs();
1211 1210
1212 1211 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1213 1212
1214 1213 {
1215 1214 IsGCActiveMark x;
1216 1215
1217 1216 // Timing
1218 1217 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc);
1219 1218 assert(!system_gc || explicit_gc, "invariant");
1220 1219 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1221 1220 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1222 1221 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC",
1223 1222 PrintGC, true, gclog_or_tty);
1224 1223
1225 1224 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1226 1225 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1227 1226
1228 1227 double start = os::elapsedTime();
1229 1228 g1_policy()->record_full_collection_start();
1230 1229
1231 1230 wait_while_free_regions_coming();
1232 1231 append_secondary_free_list_if_not_empty_with_lock();
1233 1232
1234 1233 gc_prologue(true);
1235 1234 increment_total_collections(true /* full gc */);
1236 1235
1237 1236 size_t g1h_prev_used = used();
1238 1237 assert(used() == recalculate_used(), "Should be equal");
1239 1238
1240 1239 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
1241 1240 HandleMark hm; // Discard invalid handles created during verification
1242 1241 gclog_or_tty->print(" VerifyBeforeGC:");
1243 1242 prepare_for_verify();
1244 1243 Universe::verify(/* allow dirty */ true,
1245 1244 /* silent */ false,
1246 1245 /* option */ VerifyOption_G1UsePrevMarking);
1247 1246
1248 1247 }
1249 1248 pre_full_gc_dump();
1250 1249
1251 1250 COMPILER2_PRESENT(DerivedPointerTable::clear());
1252 1251
1253 1252 // Disable discovery and empty the discovered lists
1254 1253 // for the CM ref processor.
1255 1254 ref_processor_cm()->disable_discovery();
1256 1255 ref_processor_cm()->abandon_partial_discovery();
1257 1256 ref_processor_cm()->verify_no_references_recorded();
1258 1257
1259 1258 // Abandon current iterations of concurrent marking and concurrent
1260 1259 // refinement, if any are in progress.
1261 1260 concurrent_mark()->abort();
1262 1261
1263 1262 // Make sure we'll choose a new allocation region afterwards.
1264 1263 release_mutator_alloc_region();
1265 1264 abandon_gc_alloc_regions();
1266 1265 g1_rem_set()->cleanupHRRS();
1267 1266 tear_down_region_lists();
1268 1267
1269 1268 // We should call this after we retire any currently active alloc
1270 1269 // regions so that all the ALLOC / RETIRE events are generated
1271 1270 // before the start GC event.
1272 1271 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1273 1272
1274 1273 // We may have added regions to the current incremental collection
1275 1274 // set between the last GC or pause and now. We need to clear the
1276 1275 // incremental collection set and then start rebuilding it afresh
1277 1276 // after this full GC.
1278 1277 abandon_collection_set(g1_policy()->inc_cset_head());
1279 1278 g1_policy()->clear_incremental_cset();
1280 1279 g1_policy()->stop_incremental_cset_building();
1281 1280
1282 1281 empty_young_list();
1283 1282 g1_policy()->set_full_young_gcs(true);
1284 1283
1285 1284 // See the comments in g1CollectedHeap.hpp and
1286 1285 // G1CollectedHeap::ref_processing_init() about
1287 1286 // how reference processing currently works in G1.
1288 1287
1289 1288 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1290 1289 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1291 1290
1292 1291 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1293 1292 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1294 1293
1295 1294 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1296 1295 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1297 1296
1298 1297 // Do collection work
1299 1298 {
1300 1299 HandleMark hm; // Discard invalid handles created during gc
1301 1300 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1302 1301 }
1303 1302
1304 1303 assert(free_regions() == 0, "we should not have added any free regions");
1305 1304 rebuild_region_lists();
1306 1305
1307 1306 _summary_bytes_used = recalculate_used();
1308 1307
1309 1308 // Enqueue any discovered reference objects that have
1310 1309 // not been removed from the discovered lists.
1311 1310 ref_processor_stw()->enqueue_discovered_references();
1312 1311
1313 1312 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1314 1313
1315 1314 MemoryService::track_memory_usage();
1316 1315
1317 1316 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
1318 1317 HandleMark hm; // Discard invalid handles created during verification
1319 1318 gclog_or_tty->print(" VerifyAfterGC:");
1320 1319 prepare_for_verify();
1321 1320 Universe::verify(/* allow dirty */ false,
1322 1321 /* silent */ false,
1323 1322 /* option */ VerifyOption_G1UsePrevMarking);
1324 1323
1325 1324 }
1326 1325
1327 1326 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1328 1327 ref_processor_stw()->verify_no_references_recorded();
1329 1328
1330 1329 // Note: since we've just done a full GC, concurrent
1331 1330 // marking is no longer active. Therefore we need not
1332 1331 // re-enable reference discovery for the CM ref processor.
1333 1332 // That will be done at the start of the next marking cycle.
1334 1333 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1335 1334 ref_processor_cm()->verify_no_references_recorded();
1336 1335
1337 1336 reset_gc_time_stamp();
1338 1337 // Since everything potentially moved, we will clear all remembered
1339 1338 // sets, and clear all cards. Later we will rebuild remebered
1340 1339 // sets. We will also reset the GC time stamps of the regions.
1341 1340 PostMCRemSetClearClosure rs_clear(mr_bs());
1342 1341 heap_region_iterate(&rs_clear);
1343 1342
1344 1343 // Resize the heap if necessary.
1345 1344 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1346 1345
1347 1346 if (_hr_printer.is_active()) {
1348 1347 // We should do this after we potentially resize the heap so
1349 1348 // that all the COMMIT / UNCOMMIT events are generated before
1350 1349 // the end GC event.
1351 1350
1352 1351 PostCompactionPrinterClosure cl(hr_printer());
1353 1352 heap_region_iterate(&cl);
1354 1353
1355 1354 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1356 1355 }
1357 1356
1358 1357 if (_cg1r->use_cache()) {
1359 1358 _cg1r->clear_and_record_card_counts();
1360 1359 _cg1r->clear_hot_cache();
1361 1360 }
1362 1361
1363 1362 // Rebuild remembered sets of all regions.
1364 1363
1365 1364 if (G1CollectedHeap::use_parallel_gc_threads()) {
1366 1365 ParRebuildRSTask rebuild_rs_task(this);
1367 1366 assert(check_heap_region_claim_values(
1368 1367 HeapRegion::InitialClaimValue), "sanity check");
1369 1368 set_par_threads(workers()->total_workers());
1370 1369 workers()->run_task(&rebuild_rs_task);
1371 1370 set_par_threads(0);
1372 1371 assert(check_heap_region_claim_values(
1373 1372 HeapRegion::RebuildRSClaimValue), "sanity check");
1374 1373 reset_heap_region_claim_values();
1375 1374 } else {
1376 1375 RebuildRSOutOfRegionClosure rebuild_rs(this);
1377 1376 heap_region_iterate(&rebuild_rs);
1378 1377 }
1379 1378
1380 1379 if (PrintGC) {
1381 1380 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1382 1381 }
1383 1382
1384 1383 if (true) { // FIXME
1385 1384 // Ask the permanent generation to adjust size for full collections
1386 1385 perm()->compute_new_size();
1387 1386 }
1388 1387
1389 1388 // Start a new incremental collection set for the next pause
1390 1389 assert(g1_policy()->collection_set() == NULL, "must be");
1391 1390 g1_policy()->start_incremental_cset_building();
1392 1391
1393 1392 // Clear the _cset_fast_test bitmap in anticipation of adding
1394 1393 // regions to the incremental collection set for the next
1395 1394 // evacuation pause.
1396 1395 clear_cset_fast_test();
1397 1396
1398 1397 init_mutator_alloc_region();
1399 1398
1400 1399 double end = os::elapsedTime();
1401 1400 g1_policy()->record_full_collection_end();
1402 1401
1403 1402 #ifdef TRACESPINNING
1404 1403 ParallelTaskTerminator::print_termination_counts();
1405 1404 #endif
1406 1405
1407 1406 gc_epilogue(true);
1408 1407
1409 1408 // Discard all rset updates
1410 1409 JavaThread::dirty_card_queue_set().abandon_logs();
1411 1410 assert(!G1DeferredRSUpdate
1412 1411 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1413 1412 }
1414 1413
1415 1414 _young_list->reset_sampled_info();
1416 1415 // At this point there should be no regions in the
1417 1416 // entire heap tagged as young.
1418 1417 assert( check_young_list_empty(true /* check_heap */),
1419 1418 "young list should be empty at this point");
1420 1419
1421 1420 // Update the number of full collections that have been completed.
1422 1421 increment_full_collections_completed(false /* concurrent */);
1423 1422
1424 1423 _hrs.verify_optional();
1425 1424 verify_region_sets_optional();
1426 1425
1427 1426 if (PrintHeapAtGC) {
1428 1427 Universe::print_heap_after_gc();
1429 1428 }
1430 1429 g1mm()->update_sizes();
1431 1430 post_full_gc_dump();
1432 1431
1433 1432 return true;
1434 1433 }
1435 1434
1436 1435 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1437 1436 // do_collection() will return whether it succeeded in performing
1438 1437 // the GC. Currently, there is no facility on the
1439 1438 // do_full_collection() API to notify the caller than the collection
1440 1439 // did not succeed (e.g., because it was locked out by the GC
1441 1440 // locker). So, right now, we'll ignore the return value.
1442 1441 bool dummy = do_collection(true, /* explicit_gc */
1443 1442 clear_all_soft_refs,
1444 1443 0 /* word_size */);
1445 1444 }
1446 1445
1447 1446 // This code is mostly copied from TenuredGeneration.
1448 1447 void
1449 1448 G1CollectedHeap::
1450 1449 resize_if_necessary_after_full_collection(size_t word_size) {
1451 1450 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1452 1451
1453 1452 // Include the current allocation, if any, and bytes that will be
1454 1453 // pre-allocated to support collections, as "used".
1455 1454 const size_t used_after_gc = used();
1456 1455 const size_t capacity_after_gc = capacity();
1457 1456 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1458 1457
1459 1458 // This is enforced in arguments.cpp.
1460 1459 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1461 1460 "otherwise the code below doesn't make sense");
1462 1461
1463 1462 // We don't have floating point command-line arguments
1464 1463 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1465 1464 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1466 1465 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1467 1466 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1468 1467
1469 1468 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1470 1469 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1471 1470
1472 1471 // We have to be careful here as these two calculations can overflow
1473 1472 // 32-bit size_t's.
1474 1473 double used_after_gc_d = (double) used_after_gc;
1475 1474 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1476 1475 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1477 1476
1478 1477 // Let's make sure that they are both under the max heap size, which
1479 1478 // by default will make them fit into a size_t.
1480 1479 double desired_capacity_upper_bound = (double) max_heap_size;
1481 1480 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1482 1481 desired_capacity_upper_bound);
1483 1482 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1484 1483 desired_capacity_upper_bound);
1485 1484
1486 1485 // We can now safely turn them into size_t's.
1487 1486 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1488 1487 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1489 1488
1490 1489 // This assert only makes sense here, before we adjust them
1491 1490 // with respect to the min and max heap size.
1492 1491 assert(minimum_desired_capacity <= maximum_desired_capacity,
1493 1492 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1494 1493 "maximum_desired_capacity = "SIZE_FORMAT,
1495 1494 minimum_desired_capacity, maximum_desired_capacity));
1496 1495
1497 1496 // Should not be greater than the heap max size. No need to adjust
1498 1497 // it with respect to the heap min size as it's a lower bound (i.e.,
1499 1498 // we'll try to make the capacity larger than it, not smaller).
1500 1499 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1501 1500 // Should not be less than the heap min size. No need to adjust it
1502 1501 // with respect to the heap max size as it's an upper bound (i.e.,
1503 1502 // we'll try to make the capacity smaller than it, not greater).
1504 1503 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1505 1504
1506 1505 if (capacity_after_gc < minimum_desired_capacity) {
1507 1506 // Don't expand unless it's significant
1508 1507 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1509 1508 ergo_verbose4(ErgoHeapSizing,
1510 1509 "attempt heap expansion",
1511 1510 ergo_format_reason("capacity lower than "
1512 1511 "min desired capacity after Full GC")
1513 1512 ergo_format_byte("capacity")
1514 1513 ergo_format_byte("occupancy")
1515 1514 ergo_format_byte_perc("min desired capacity"),
1516 1515 capacity_after_gc, used_after_gc,
1517 1516 minimum_desired_capacity, (double) MinHeapFreeRatio);
1518 1517 expand(expand_bytes);
1519 1518
1520 1519 // No expansion, now see if we want to shrink
1521 1520 } else if (capacity_after_gc > maximum_desired_capacity) {
1522 1521 // Capacity too large, compute shrinking size
1523 1522 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1524 1523 ergo_verbose4(ErgoHeapSizing,
1525 1524 "attempt heap shrinking",
1526 1525 ergo_format_reason("capacity higher than "
1527 1526 "max desired capacity after Full GC")
1528 1527 ergo_format_byte("capacity")
1529 1528 ergo_format_byte("occupancy")
1530 1529 ergo_format_byte_perc("max desired capacity"),
1531 1530 capacity_after_gc, used_after_gc,
1532 1531 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1533 1532 shrink(shrink_bytes);
1534 1533 }
1535 1534 }
1536 1535
1537 1536
1538 1537 HeapWord*
1539 1538 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1540 1539 bool* succeeded) {
1541 1540 assert_at_safepoint(true /* should_be_vm_thread */);
1542 1541
1543 1542 *succeeded = true;
1544 1543 // Let's attempt the allocation first.
1545 1544 HeapWord* result =
1546 1545 attempt_allocation_at_safepoint(word_size,
1547 1546 false /* expect_null_mutator_alloc_region */);
1548 1547 if (result != NULL) {
1549 1548 assert(*succeeded, "sanity");
1550 1549 return result;
1551 1550 }
1552 1551
1553 1552 // In a G1 heap, we're supposed to keep allocation from failing by
1554 1553 // incremental pauses. Therefore, at least for now, we'll favor
1555 1554 // expansion over collection. (This might change in the future if we can
1556 1555 // do something smarter than full collection to satisfy a failed alloc.)
1557 1556 result = expand_and_allocate(word_size);
1558 1557 if (result != NULL) {
1559 1558 assert(*succeeded, "sanity");
1560 1559 return result;
1561 1560 }
1562 1561
1563 1562 // Expansion didn't work, we'll try to do a Full GC.
1564 1563 bool gc_succeeded = do_collection(false, /* explicit_gc */
1565 1564 false, /* clear_all_soft_refs */
1566 1565 word_size);
1567 1566 if (!gc_succeeded) {
1568 1567 *succeeded = false;
1569 1568 return NULL;
1570 1569 }
1571 1570
1572 1571 // Retry the allocation
1573 1572 result = attempt_allocation_at_safepoint(word_size,
1574 1573 true /* expect_null_mutator_alloc_region */);
1575 1574 if (result != NULL) {
1576 1575 assert(*succeeded, "sanity");
1577 1576 return result;
1578 1577 }
1579 1578
1580 1579 // Then, try a Full GC that will collect all soft references.
1581 1580 gc_succeeded = do_collection(false, /* explicit_gc */
1582 1581 true, /* clear_all_soft_refs */
1583 1582 word_size);
1584 1583 if (!gc_succeeded) {
1585 1584 *succeeded = false;
1586 1585 return NULL;
1587 1586 }
1588 1587
1589 1588 // Retry the allocation once more
1590 1589 result = attempt_allocation_at_safepoint(word_size,
1591 1590 true /* expect_null_mutator_alloc_region */);
1592 1591 if (result != NULL) {
1593 1592 assert(*succeeded, "sanity");
1594 1593 return result;
1595 1594 }
1596 1595
1597 1596 assert(!collector_policy()->should_clear_all_soft_refs(),
1598 1597 "Flag should have been handled and cleared prior to this point");
1599 1598
1600 1599 // What else? We might try synchronous finalization later. If the total
1601 1600 // space available is large enough for the allocation, then a more
1602 1601 // complete compaction phase than we've tried so far might be
1603 1602 // appropriate.
1604 1603 assert(*succeeded, "sanity");
1605 1604 return NULL;
1606 1605 }
1607 1606
1608 1607 // Attempting to expand the heap sufficiently
1609 1608 // to support an allocation of the given "word_size". If
1610 1609 // successful, perform the allocation and return the address of the
1611 1610 // allocated block, or else "NULL".
1612 1611
1613 1612 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1614 1613 assert_at_safepoint(true /* should_be_vm_thread */);
1615 1614
1616 1615 verify_region_sets_optional();
1617 1616
1618 1617 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1619 1618 ergo_verbose1(ErgoHeapSizing,
1620 1619 "attempt heap expansion",
1621 1620 ergo_format_reason("allocation request failed")
1622 1621 ergo_format_byte("allocation request"),
1623 1622 word_size * HeapWordSize);
1624 1623 if (expand(expand_bytes)) {
1625 1624 _hrs.verify_optional();
1626 1625 verify_region_sets_optional();
1627 1626 return attempt_allocation_at_safepoint(word_size,
1628 1627 false /* expect_null_mutator_alloc_region */);
1629 1628 }
1630 1629 return NULL;
1631 1630 }
1632 1631
1633 1632 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1634 1633 HeapWord* new_end) {
1635 1634 assert(old_end != new_end, "don't call this otherwise");
1636 1635 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1637 1636
1638 1637 // Update the committed mem region.
1639 1638 _g1_committed.set_end(new_end);
1640 1639 // Tell the card table about the update.
1641 1640 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1642 1641 // Tell the BOT about the update.
1643 1642 _bot_shared->resize(_g1_committed.word_size());
1644 1643 }
1645 1644
1646 1645 bool G1CollectedHeap::expand(size_t expand_bytes) {
1647 1646 size_t old_mem_size = _g1_storage.committed_size();
1648 1647 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1649 1648 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1650 1649 HeapRegion::GrainBytes);
1651 1650 ergo_verbose2(ErgoHeapSizing,
1652 1651 "expand the heap",
1653 1652 ergo_format_byte("requested expansion amount")
1654 1653 ergo_format_byte("attempted expansion amount"),
1655 1654 expand_bytes, aligned_expand_bytes);
1656 1655
1657 1656 // First commit the memory.
1658 1657 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1659 1658 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1660 1659 if (successful) {
1661 1660 // Then propagate this update to the necessary data structures.
1662 1661 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1663 1662 update_committed_space(old_end, new_end);
1664 1663
1665 1664 FreeRegionList expansion_list("Local Expansion List");
1666 1665 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1667 1666 assert(mr.start() == old_end, "post-condition");
1668 1667 // mr might be a smaller region than what was requested if
1669 1668 // expand_by() was unable to allocate the HeapRegion instances
1670 1669 assert(mr.end() <= new_end, "post-condition");
1671 1670
1672 1671 size_t actual_expand_bytes = mr.byte_size();
1673 1672 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1674 1673 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1675 1674 "post-condition");
1676 1675 if (actual_expand_bytes < aligned_expand_bytes) {
1677 1676 // We could not expand _hrs to the desired size. In this case we
1678 1677 // need to shrink the committed space accordingly.
1679 1678 assert(mr.end() < new_end, "invariant");
1680 1679
1681 1680 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1682 1681 // First uncommit the memory.
1683 1682 _g1_storage.shrink_by(diff_bytes);
1684 1683 // Then propagate this update to the necessary data structures.
1685 1684 update_committed_space(new_end, mr.end());
1686 1685 }
1687 1686 _free_list.add_as_tail(&expansion_list);
1688 1687
1689 1688 if (_hr_printer.is_active()) {
1690 1689 HeapWord* curr = mr.start();
1691 1690 while (curr < mr.end()) {
1692 1691 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1693 1692 _hr_printer.commit(curr, curr_end);
1694 1693 curr = curr_end;
1695 1694 }
1696 1695 assert(curr == mr.end(), "post-condition");
1697 1696 }
1698 1697 g1_policy()->record_new_heap_size(n_regions());
1699 1698 } else {
1700 1699 ergo_verbose0(ErgoHeapSizing,
1701 1700 "did not expand the heap",
1702 1701 ergo_format_reason("heap expansion operation failed"));
1703 1702 // The expansion of the virtual storage space was unsuccessful.
1704 1703 // Let's see if it was because we ran out of swap.
1705 1704 if (G1ExitOnExpansionFailure &&
1706 1705 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1707 1706 // We had head room...
1708 1707 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1709 1708 }
1710 1709 }
1711 1710 return successful;
1712 1711 }
1713 1712
1714 1713 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1715 1714 size_t old_mem_size = _g1_storage.committed_size();
1716 1715 size_t aligned_shrink_bytes =
1717 1716 ReservedSpace::page_align_size_down(shrink_bytes);
1718 1717 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1719 1718 HeapRegion::GrainBytes);
1720 1719 size_t num_regions_deleted = 0;
1721 1720 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1722 1721 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1723 1722 assert(mr.end() == old_end, "post-condition");
1724 1723
1725 1724 ergo_verbose3(ErgoHeapSizing,
1726 1725 "shrink the heap",
1727 1726 ergo_format_byte("requested shrinking amount")
1728 1727 ergo_format_byte("aligned shrinking amount")
1729 1728 ergo_format_byte("attempted shrinking amount"),
1730 1729 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1731 1730 if (mr.byte_size() > 0) {
1732 1731 if (_hr_printer.is_active()) {
1733 1732 HeapWord* curr = mr.end();
1734 1733 while (curr > mr.start()) {
1735 1734 HeapWord* curr_end = curr;
1736 1735 curr -= HeapRegion::GrainWords;
1737 1736 _hr_printer.uncommit(curr, curr_end);
1738 1737 }
1739 1738 assert(curr == mr.start(), "post-condition");
1740 1739 }
1741 1740
1742 1741 _g1_storage.shrink_by(mr.byte_size());
1743 1742 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1744 1743 assert(mr.start() == new_end, "post-condition");
1745 1744
1746 1745 _expansion_regions += num_regions_deleted;
1747 1746 update_committed_space(old_end, new_end);
1748 1747 HeapRegionRemSet::shrink_heap(n_regions());
1749 1748 g1_policy()->record_new_heap_size(n_regions());
1750 1749 } else {
1751 1750 ergo_verbose0(ErgoHeapSizing,
1752 1751 "did not shrink the heap",
1753 1752 ergo_format_reason("heap shrinking operation failed"));
1754 1753 }
1755 1754 }
1756 1755
1757 1756 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1758 1757 verify_region_sets_optional();
1759 1758
1760 1759 // We should only reach here at the end of a Full GC which means we
1761 1760 // should not not be holding to any GC alloc regions. The method
1762 1761 // below will make sure of that and do any remaining clean up.
1763 1762 abandon_gc_alloc_regions();
1764 1763
1765 1764 // Instead of tearing down / rebuilding the free lists here, we
1766 1765 // could instead use the remove_all_pending() method on free_list to
1767 1766 // remove only the ones that we need to remove.
1768 1767 tear_down_region_lists(); // We will rebuild them in a moment.
1769 1768 shrink_helper(shrink_bytes);
1770 1769 rebuild_region_lists();
1771 1770
1772 1771 _hrs.verify_optional();
1773 1772 verify_region_sets_optional();
1774 1773 }
1775 1774
1776 1775 // Public methods.
1777 1776
1778 1777 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1779 1778 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1780 1779 #endif // _MSC_VER
1781 1780
1782 1781
1783 1782 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1784 1783 SharedHeap(policy_),
1785 1784 _g1_policy(policy_),
1786 1785 _dirty_card_queue_set(false),
1787 1786 _into_cset_dirty_card_queue_set(false),
1788 1787 _is_alive_closure_cm(this),
1789 1788 _is_alive_closure_stw(this),
1790 1789 _ref_processor_cm(NULL),
1791 1790 _ref_processor_stw(NULL),
1792 1791 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1793 1792 _bot_shared(NULL),
1794 1793 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1795 1794 _evac_failure_scan_stack(NULL) ,
1796 1795 _mark_in_progress(false),
1797 1796 _cg1r(NULL), _summary_bytes_used(0),
1798 1797 _g1mm(NULL),
1799 1798 _refine_cte_cl(NULL),
1800 1799 _full_collection(false),
1801 1800 _free_list("Master Free List"),
1802 1801 _secondary_free_list("Secondary Free List"),
1803 1802 _humongous_set("Master Humongous Set"),
1804 1803 _free_regions_coming(false),
1805 1804 _young_list(new YoungList(this)),
1806 1805 _gc_time_stamp(0),
1807 1806 _retained_old_gc_alloc_region(NULL),
1808 1807 _surviving_young_words(NULL),
1809 1808 _full_collections_completed(0),
1810 1809 _in_cset_fast_test(NULL),
1811 1810 _in_cset_fast_test_base(NULL),
1812 1811 _dirty_cards_region_list(NULL) {
1813 1812 _g1h = this; // To catch bugs.
1814 1813 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1815 1814 vm_exit_during_initialization("Failed necessary allocation.");
1816 1815 }
1817 1816
1818 1817 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1819 1818
1820 1819 int n_queues = MAX2((int)ParallelGCThreads, 1);
1821 1820 _task_queues = new RefToScanQueueSet(n_queues);
1822 1821
1823 1822 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1824 1823 assert(n_rem_sets > 0, "Invariant.");
1825 1824
1826 1825 HeapRegionRemSetIterator** iter_arr =
1827 1826 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues);
1828 1827 for (int i = 0; i < n_queues; i++) {
1829 1828 iter_arr[i] = new HeapRegionRemSetIterator();
1830 1829 }
1831 1830 _rem_set_iterator = iter_arr;
1832 1831
1833 1832 for (int i = 0; i < n_queues; i++) {
1834 1833 RefToScanQueue* q = new RefToScanQueue();
1835 1834 q->initialize();
1836 1835 _task_queues->register_queue(i, q);
1837 1836 }
1838 1837
1839 1838 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1840 1839 }
1841 1840
1842 1841 jint G1CollectedHeap::initialize() {
1843 1842 CollectedHeap::pre_initialize();
1844 1843 os::enable_vtime();
1845 1844
1846 1845 // Necessary to satisfy locking discipline assertions.
1847 1846
1848 1847 MutexLocker x(Heap_lock);
1849 1848
1850 1849 // We have to initialize the printer before committing the heap, as
1851 1850 // it will be used then.
1852 1851 _hr_printer.set_active(G1PrintHeapRegions);
1853 1852
1854 1853 // While there are no constraints in the GC code that HeapWordSize
1855 1854 // be any particular value, there are multiple other areas in the
1856 1855 // system which believe this to be true (e.g. oop->object_size in some
1857 1856 // cases incorrectly returns the size in wordSize units rather than
1858 1857 // HeapWordSize).
1859 1858 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1860 1859
1861 1860 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1862 1861 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1863 1862
1864 1863 // Ensure that the sizes are properly aligned.
1865 1864 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1866 1865 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1867 1866
1868 1867 _cg1r = new ConcurrentG1Refine();
1869 1868
1870 1869 // Reserve the maximum.
1871 1870 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation();
1872 1871 // Includes the perm-gen.
1873 1872
1874 1873 // When compressed oops are enabled, the preferred heap base
1875 1874 // is calculated by subtracting the requested size from the
1876 1875 // 32Gb boundary and using the result as the base address for
1877 1876 // heap reservation. If the requested size is not aligned to
1878 1877 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1879 1878 // into the ReservedHeapSpace constructor) then the actual
1880 1879 // base of the reserved heap may end up differing from the
1881 1880 // address that was requested (i.e. the preferred heap base).
1882 1881 // If this happens then we could end up using a non-optimal
1883 1882 // compressed oops mode.
1884 1883
1885 1884 // Since max_byte_size is aligned to the size of a heap region (checked
1886 1885 // above), we also need to align the perm gen size as it might not be.
1887 1886 const size_t total_reserved = max_byte_size +
1888 1887 align_size_up(pgs->max_size(), HeapRegion::GrainBytes);
1889 1888 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm");
1890 1889
1891 1890 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop);
1892 1891
1893 1892 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes,
1894 1893 UseLargePages, addr);
1895 1894
1896 1895 if (UseCompressedOops) {
1897 1896 if (addr != NULL && !heap_rs.is_reserved()) {
1898 1897 // Failed to reserve at specified address - the requested memory
1899 1898 // region is taken already, for example, by 'java' launcher.
1900 1899 // Try again to reserver heap higher.
1901 1900 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop);
1902 1901
1903 1902 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes,
1904 1903 UseLargePages, addr);
1905 1904
1906 1905 if (addr != NULL && !heap_rs0.is_reserved()) {
1907 1906 // Failed to reserve at specified address again - give up.
1908 1907 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop);
1909 1908 assert(addr == NULL, "");
1910 1909
1911 1910 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes,
1912 1911 UseLargePages, addr);
1913 1912 heap_rs = heap_rs1;
1914 1913 } else {
1915 1914 heap_rs = heap_rs0;
1916 1915 }
1917 1916 }
1918 1917 }
1919 1918
1920 1919 if (!heap_rs.is_reserved()) {
1921 1920 vm_exit_during_initialization("Could not reserve enough space for object heap");
1922 1921 return JNI_ENOMEM;
1923 1922 }
1924 1923
1925 1924 // It is important to do this in a way such that concurrent readers can't
1926 1925 // temporarily think somethings in the heap. (I've actually seen this
1927 1926 // happen in asserts: DLD.)
1928 1927 _reserved.set_word_size(0);
1929 1928 _reserved.set_start((HeapWord*)heap_rs.base());
1930 1929 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
1931 1930
1932 1931 _expansion_regions = max_byte_size/HeapRegion::GrainBytes;
1933 1932
1934 1933 // Create the gen rem set (and barrier set) for the entire reserved region.
1935 1934 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
1936 1935 set_barrier_set(rem_set()->bs());
1937 1936 if (barrier_set()->is_a(BarrierSet::ModRef)) {
1938 1937 _mr_bs = (ModRefBarrierSet*)_barrier_set;
1939 1938 } else {
1940 1939 vm_exit_during_initialization("G1 requires a mod ref bs.");
1941 1940 return JNI_ENOMEM;
1942 1941 }
1943 1942
1944 1943 // Also create a G1 rem set.
1945 1944 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
1946 1945 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
1947 1946 } else {
1948 1947 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
1949 1948 return JNI_ENOMEM;
1950 1949 }
1951 1950
1952 1951 // Carve out the G1 part of the heap.
1953 1952
1954 1953 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
1955 1954 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
1956 1955 g1_rs.size()/HeapWordSize);
1957 1956 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size);
1958 1957
1959 1958 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set());
1960 1959
1961 1960 _g1_storage.initialize(g1_rs, 0);
1962 1961 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
1963 1962 _hrs.initialize((HeapWord*) _g1_reserved.start(),
↓ open down ↓ |
780 lines elided |
↑ open up ↑ |
1964 1963 (HeapWord*) _g1_reserved.end(),
1965 1964 _expansion_regions);
1966 1965
1967 1966 // 6843694 - ensure that the maximum region index can fit
1968 1967 // in the remembered set structures.
1969 1968 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
1970 1969 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
1971 1970
1972 1971 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
1973 1972 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
1974 - guarantee((size_t) HeapRegion::CardsPerRegion < max_cards_per_region,
1973 + guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
1975 1974 "too many cards per region");
1976 1975
1977 1976 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
1978 1977
1979 1978 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
1980 1979 heap_word_size(init_byte_size));
1981 1980
1982 1981 _g1h = this;
1983 1982
1984 1983 _in_cset_fast_test_length = max_regions();
1985 1984 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length);
1986 1985
1987 1986 // We're biasing _in_cset_fast_test to avoid subtracting the
1988 1987 // beginning of the heap every time we want to index; basically
1989 1988 // it's the same with what we do with the card table.
1990 1989 _in_cset_fast_test = _in_cset_fast_test_base -
1991 1990 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
1992 1991
1993 1992 // Clear the _cset_fast_test bitmap in anticipation of adding
1994 1993 // regions to the incremental collection set for the first
1995 1994 // evacuation pause.
1996 1995 clear_cset_fast_test();
1997 1996
1998 1997 // Create the ConcurrentMark data structure and thread.
1999 1998 // (Must do this late, so that "max_regions" is defined.)
2000 1999 _cm = new ConcurrentMark(heap_rs, (int) max_regions());
2001 2000 _cmThread = _cm->cmThread();
2002 2001
2003 2002 // Initialize the from_card cache structure of HeapRegionRemSet.
2004 2003 HeapRegionRemSet::init_heap(max_regions());
2005 2004
2006 2005 // Now expand into the initial heap size.
2007 2006 if (!expand(init_byte_size)) {
2008 2007 vm_exit_during_initialization("Failed to allocate initial heap.");
2009 2008 return JNI_ENOMEM;
2010 2009 }
2011 2010
2012 2011 // Perform any initialization actions delegated to the policy.
2013 2012 g1_policy()->init();
2014 2013
2015 2014 g1_policy()->note_start_of_mark_thread();
2016 2015
2017 2016 _refine_cte_cl =
2018 2017 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2019 2018 g1_rem_set(),
2020 2019 concurrent_g1_refine());
2021 2020 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2022 2021
2023 2022 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2024 2023 SATB_Q_FL_lock,
2025 2024 G1SATBProcessCompletedThreshold,
2026 2025 Shared_SATB_Q_lock);
2027 2026
2028 2027 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2029 2028 DirtyCardQ_FL_lock,
2030 2029 concurrent_g1_refine()->yellow_zone(),
2031 2030 concurrent_g1_refine()->red_zone(),
2032 2031 Shared_DirtyCardQ_lock);
2033 2032
2034 2033 if (G1DeferredRSUpdate) {
2035 2034 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2036 2035 DirtyCardQ_FL_lock,
2037 2036 -1, // never trigger processing
2038 2037 -1, // no limit on length
2039 2038 Shared_DirtyCardQ_lock,
2040 2039 &JavaThread::dirty_card_queue_set());
2041 2040 }
2042 2041
2043 2042 // Initialize the card queue set used to hold cards containing
2044 2043 // references into the collection set.
2045 2044 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2046 2045 DirtyCardQ_FL_lock,
2047 2046 -1, // never trigger processing
2048 2047 -1, // no limit on length
2049 2048 Shared_DirtyCardQ_lock,
2050 2049 &JavaThread::dirty_card_queue_set());
2051 2050
2052 2051 // In case we're keeping closure specialization stats, initialize those
2053 2052 // counts and that mechanism.
2054 2053 SpecializationStats::clear();
2055 2054
2056 2055 // Do later initialization work for concurrent refinement.
2057 2056 _cg1r->init();
2058 2057
2059 2058 // Here we allocate the dummy full region that is required by the
2060 2059 // G1AllocRegion class. If we don't pass an address in the reserved
2061 2060 // space here, lots of asserts fire.
2062 2061
2063 2062 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2064 2063 _g1_reserved.start());
2065 2064 // We'll re-use the same region whether the alloc region will
2066 2065 // require BOT updates or not and, if it doesn't, then a non-young
2067 2066 // region will complain that it cannot support allocations without
2068 2067 // BOT updates. So we'll tag the dummy region as young to avoid that.
2069 2068 dummy_region->set_young();
2070 2069 // Make sure it's full.
2071 2070 dummy_region->set_top(dummy_region->end());
2072 2071 G1AllocRegion::setup(this, dummy_region);
2073 2072
2074 2073 init_mutator_alloc_region();
2075 2074
2076 2075 // Do create of the monitoring and management support so that
2077 2076 // values in the heap have been properly initialized.
2078 2077 _g1mm = new G1MonitoringSupport(this);
2079 2078
2080 2079 return JNI_OK;
2081 2080 }
2082 2081
2083 2082 void G1CollectedHeap::ref_processing_init() {
2084 2083 // Reference processing in G1 currently works as follows:
2085 2084 //
2086 2085 // * There are two reference processor instances. One is
2087 2086 // used to record and process discovered references
2088 2087 // during concurrent marking; the other is used to
2089 2088 // record and process references during STW pauses
2090 2089 // (both full and incremental).
2091 2090 // * Both ref processors need to 'span' the entire heap as
2092 2091 // the regions in the collection set may be dotted around.
2093 2092 //
2094 2093 // * For the concurrent marking ref processor:
2095 2094 // * Reference discovery is enabled at initial marking.
2096 2095 // * Reference discovery is disabled and the discovered
2097 2096 // references processed etc during remarking.
2098 2097 // * Reference discovery is MT (see below).
2099 2098 // * Reference discovery requires a barrier (see below).
2100 2099 // * Reference processing may or may not be MT
2101 2100 // (depending on the value of ParallelRefProcEnabled
2102 2101 // and ParallelGCThreads).
2103 2102 // * A full GC disables reference discovery by the CM
2104 2103 // ref processor and abandons any entries on it's
2105 2104 // discovered lists.
2106 2105 //
2107 2106 // * For the STW processor:
2108 2107 // * Non MT discovery is enabled at the start of a full GC.
2109 2108 // * Processing and enqueueing during a full GC is non-MT.
2110 2109 // * During a full GC, references are processed after marking.
2111 2110 //
2112 2111 // * Discovery (may or may not be MT) is enabled at the start
2113 2112 // of an incremental evacuation pause.
2114 2113 // * References are processed near the end of a STW evacuation pause.
2115 2114 // * For both types of GC:
2116 2115 // * Discovery is atomic - i.e. not concurrent.
2117 2116 // * Reference discovery will not need a barrier.
2118 2117
2119 2118 SharedHeap::ref_processing_init();
2120 2119 MemRegion mr = reserved_region();
2121 2120
2122 2121 // Concurrent Mark ref processor
2123 2122 _ref_processor_cm =
2124 2123 new ReferenceProcessor(mr, // span
2125 2124 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2126 2125 // mt processing
2127 2126 (int) ParallelGCThreads,
2128 2127 // degree of mt processing
2129 2128 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2130 2129 // mt discovery
2131 2130 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2132 2131 // degree of mt discovery
2133 2132 false,
2134 2133 // Reference discovery is not atomic
2135 2134 &_is_alive_closure_cm,
2136 2135 // is alive closure
2137 2136 // (for efficiency/performance)
2138 2137 true);
2139 2138 // Setting next fields of discovered
2140 2139 // lists requires a barrier.
2141 2140
2142 2141 // STW ref processor
2143 2142 _ref_processor_stw =
2144 2143 new ReferenceProcessor(mr, // span
2145 2144 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2146 2145 // mt processing
2147 2146 MAX2((int)ParallelGCThreads, 1),
2148 2147 // degree of mt processing
2149 2148 (ParallelGCThreads > 1),
2150 2149 // mt discovery
2151 2150 MAX2((int)ParallelGCThreads, 1),
2152 2151 // degree of mt discovery
2153 2152 true,
2154 2153 // Reference discovery is atomic
2155 2154 &_is_alive_closure_stw,
2156 2155 // is alive closure
2157 2156 // (for efficiency/performance)
2158 2157 false);
2159 2158 // Setting next fields of discovered
2160 2159 // lists requires a barrier.
2161 2160 }
2162 2161
2163 2162 size_t G1CollectedHeap::capacity() const {
2164 2163 return _g1_committed.byte_size();
2165 2164 }
2166 2165
2167 2166 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2168 2167 DirtyCardQueue* into_cset_dcq,
2169 2168 bool concurrent,
2170 2169 int worker_i) {
2171 2170 // Clean cards in the hot card cache
2172 2171 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2173 2172
2174 2173 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2175 2174 int n_completed_buffers = 0;
2176 2175 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2177 2176 n_completed_buffers++;
2178 2177 }
2179 2178 g1_policy()->record_update_rs_processed_buffers(worker_i,
2180 2179 (double) n_completed_buffers);
2181 2180 dcqs.clear_n_completed_buffers();
2182 2181 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2183 2182 }
2184 2183
2185 2184
2186 2185 // Computes the sum of the storage used by the various regions.
2187 2186
2188 2187 size_t G1CollectedHeap::used() const {
2189 2188 assert(Heap_lock->owner() != NULL,
2190 2189 "Should be owned on this thread's behalf.");
2191 2190 size_t result = _summary_bytes_used;
2192 2191 // Read only once in case it is set to NULL concurrently
2193 2192 HeapRegion* hr = _mutator_alloc_region.get();
2194 2193 if (hr != NULL)
2195 2194 result += hr->used();
2196 2195 return result;
2197 2196 }
2198 2197
2199 2198 size_t G1CollectedHeap::used_unlocked() const {
2200 2199 size_t result = _summary_bytes_used;
2201 2200 return result;
2202 2201 }
2203 2202
2204 2203 class SumUsedClosure: public HeapRegionClosure {
2205 2204 size_t _used;
2206 2205 public:
2207 2206 SumUsedClosure() : _used(0) {}
2208 2207 bool doHeapRegion(HeapRegion* r) {
2209 2208 if (!r->continuesHumongous()) {
2210 2209 _used += r->used();
2211 2210 }
2212 2211 return false;
2213 2212 }
2214 2213 size_t result() { return _used; }
2215 2214 };
2216 2215
2217 2216 size_t G1CollectedHeap::recalculate_used() const {
2218 2217 SumUsedClosure blk;
2219 2218 heap_region_iterate(&blk);
2220 2219 return blk.result();
2221 2220 }
2222 2221
2223 2222 size_t G1CollectedHeap::unsafe_max_alloc() {
2224 2223 if (free_regions() > 0) return HeapRegion::GrainBytes;
2225 2224 // otherwise, is there space in the current allocation region?
2226 2225
2227 2226 // We need to store the current allocation region in a local variable
2228 2227 // here. The problem is that this method doesn't take any locks and
2229 2228 // there may be other threads which overwrite the current allocation
2230 2229 // region field. attempt_allocation(), for example, sets it to NULL
2231 2230 // and this can happen *after* the NULL check here but before the call
2232 2231 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2233 2232 // to be a problem in the optimized build, since the two loads of the
2234 2233 // current allocation region field are optimized away.
2235 2234 HeapRegion* hr = _mutator_alloc_region.get();
2236 2235 if (hr == NULL) {
2237 2236 return 0;
2238 2237 }
2239 2238 return hr->free();
2240 2239 }
2241 2240
2242 2241 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2243 2242 return
2244 2243 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) ||
2245 2244 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent));
2246 2245 }
2247 2246
2248 2247 #ifndef PRODUCT
2249 2248 void G1CollectedHeap::allocate_dummy_regions() {
2250 2249 // Let's fill up most of the region
2251 2250 size_t word_size = HeapRegion::GrainWords - 1024;
2252 2251 // And as a result the region we'll allocate will be humongous.
2253 2252 guarantee(isHumongous(word_size), "sanity");
2254 2253
2255 2254 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2256 2255 // Let's use the existing mechanism for the allocation
2257 2256 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2258 2257 if (dummy_obj != NULL) {
2259 2258 MemRegion mr(dummy_obj, word_size);
2260 2259 CollectedHeap::fill_with_object(mr);
2261 2260 } else {
2262 2261 // If we can't allocate once, we probably cannot allocate
2263 2262 // again. Let's get out of the loop.
2264 2263 break;
2265 2264 }
2266 2265 }
2267 2266 }
2268 2267 #endif // !PRODUCT
2269 2268
2270 2269 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) {
2271 2270 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2272 2271
2273 2272 // We assume that if concurrent == true, then the caller is a
2274 2273 // concurrent thread that was joined the Suspendible Thread
2275 2274 // Set. If there's ever a cheap way to check this, we should add an
2276 2275 // assert here.
2277 2276
2278 2277 // We have already incremented _total_full_collections at the start
2279 2278 // of the GC, so total_full_collections() represents how many full
2280 2279 // collections have been started.
2281 2280 unsigned int full_collections_started = total_full_collections();
2282 2281
2283 2282 // Given that this method is called at the end of a Full GC or of a
2284 2283 // concurrent cycle, and those can be nested (i.e., a Full GC can
2285 2284 // interrupt a concurrent cycle), the number of full collections
2286 2285 // completed should be either one (in the case where there was no
2287 2286 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2288 2287 // behind the number of full collections started.
2289 2288
2290 2289 // This is the case for the inner caller, i.e. a Full GC.
2291 2290 assert(concurrent ||
2292 2291 (full_collections_started == _full_collections_completed + 1) ||
2293 2292 (full_collections_started == _full_collections_completed + 2),
2294 2293 err_msg("for inner caller (Full GC): full_collections_started = %u "
2295 2294 "is inconsistent with _full_collections_completed = %u",
2296 2295 full_collections_started, _full_collections_completed));
2297 2296
2298 2297 // This is the case for the outer caller, i.e. the concurrent cycle.
2299 2298 assert(!concurrent ||
2300 2299 (full_collections_started == _full_collections_completed + 1),
2301 2300 err_msg("for outer caller (concurrent cycle): "
2302 2301 "full_collections_started = %u "
2303 2302 "is inconsistent with _full_collections_completed = %u",
2304 2303 full_collections_started, _full_collections_completed));
2305 2304
2306 2305 _full_collections_completed += 1;
2307 2306
2308 2307 // We need to clear the "in_progress" flag in the CM thread before
2309 2308 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2310 2309 // is set) so that if a waiter requests another System.gc() it doesn't
2311 2310 // incorrectly see that a marking cyle is still in progress.
2312 2311 if (concurrent) {
2313 2312 _cmThread->clear_in_progress();
2314 2313 }
2315 2314
2316 2315 // This notify_all() will ensure that a thread that called
2317 2316 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2318 2317 // and it's waiting for a full GC to finish will be woken up. It is
2319 2318 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2320 2319 FullGCCount_lock->notify_all();
2321 2320 }
2322 2321
2323 2322 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) {
2324 2323 assert_at_safepoint(true /* should_be_vm_thread */);
2325 2324 GCCauseSetter gcs(this, cause);
2326 2325 switch (cause) {
2327 2326 case GCCause::_heap_inspection:
2328 2327 case GCCause::_heap_dump: {
2329 2328 HandleMark hm;
2330 2329 do_full_collection(false); // don't clear all soft refs
2331 2330 break;
2332 2331 }
2333 2332 default: // XXX FIX ME
2334 2333 ShouldNotReachHere(); // Unexpected use of this function
2335 2334 }
2336 2335 }
2337 2336
2338 2337 void G1CollectedHeap::collect(GCCause::Cause cause) {
2339 2338 // The caller doesn't have the Heap_lock
2340 2339 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock");
2341 2340
2342 2341 unsigned int gc_count_before;
2343 2342 unsigned int full_gc_count_before;
2344 2343 {
2345 2344 MutexLocker ml(Heap_lock);
2346 2345
2347 2346 // Read the GC count while holding the Heap_lock
2348 2347 gc_count_before = SharedHeap::heap()->total_collections();
2349 2348 full_gc_count_before = SharedHeap::heap()->total_full_collections();
2350 2349 }
2351 2350
2352 2351 if (should_do_concurrent_full_gc(cause)) {
2353 2352 // Schedule an initial-mark evacuation pause that will start a
2354 2353 // concurrent cycle. We're setting word_size to 0 which means that
2355 2354 // we are not requesting a post-GC allocation.
2356 2355 VM_G1IncCollectionPause op(gc_count_before,
2357 2356 0, /* word_size */
2358 2357 true, /* should_initiate_conc_mark */
2359 2358 g1_policy()->max_pause_time_ms(),
2360 2359 cause);
2361 2360 VMThread::execute(&op);
2362 2361 } else {
2363 2362 if (cause == GCCause::_gc_locker
2364 2363 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2365 2364
2366 2365 // Schedule a standard evacuation pause. We're setting word_size
2367 2366 // to 0 which means that we are not requesting a post-GC allocation.
2368 2367 VM_G1IncCollectionPause op(gc_count_before,
2369 2368 0, /* word_size */
2370 2369 false, /* should_initiate_conc_mark */
2371 2370 g1_policy()->max_pause_time_ms(),
2372 2371 cause);
2373 2372 VMThread::execute(&op);
2374 2373 } else {
2375 2374 // Schedule a Full GC.
2376 2375 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
2377 2376 VMThread::execute(&op);
2378 2377 }
2379 2378 }
2380 2379 }
2381 2380
2382 2381 bool G1CollectedHeap::is_in(const void* p) const {
2383 2382 HeapRegion* hr = _hrs.addr_to_region((HeapWord*) p);
2384 2383 if (hr != NULL) {
2385 2384 return hr->is_in(p);
2386 2385 } else {
2387 2386 return _perm_gen->as_gen()->is_in(p);
2388 2387 }
2389 2388 }
2390 2389
2391 2390 // Iteration functions.
2392 2391
2393 2392 // Iterates an OopClosure over all ref-containing fields of objects
2394 2393 // within a HeapRegion.
2395 2394
2396 2395 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2397 2396 MemRegion _mr;
2398 2397 OopClosure* _cl;
2399 2398 public:
2400 2399 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl)
2401 2400 : _mr(mr), _cl(cl) {}
2402 2401 bool doHeapRegion(HeapRegion* r) {
2403 2402 if (! r->continuesHumongous()) {
2404 2403 r->oop_iterate(_cl);
2405 2404 }
2406 2405 return false;
2407 2406 }
2408 2407 };
2409 2408
2410 2409 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) {
2411 2410 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2412 2411 heap_region_iterate(&blk);
2413 2412 if (do_perm) {
2414 2413 perm_gen()->oop_iterate(cl);
2415 2414 }
2416 2415 }
2417 2416
2418 2417 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) {
2419 2418 IterateOopClosureRegionClosure blk(mr, cl);
2420 2419 heap_region_iterate(&blk);
2421 2420 if (do_perm) {
2422 2421 perm_gen()->oop_iterate(cl);
2423 2422 }
2424 2423 }
2425 2424
2426 2425 // Iterates an ObjectClosure over all objects within a HeapRegion.
2427 2426
2428 2427 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2429 2428 ObjectClosure* _cl;
2430 2429 public:
2431 2430 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2432 2431 bool doHeapRegion(HeapRegion* r) {
2433 2432 if (! r->continuesHumongous()) {
2434 2433 r->object_iterate(_cl);
2435 2434 }
2436 2435 return false;
2437 2436 }
2438 2437 };
2439 2438
2440 2439 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) {
2441 2440 IterateObjectClosureRegionClosure blk(cl);
2442 2441 heap_region_iterate(&blk);
2443 2442 if (do_perm) {
2444 2443 perm_gen()->object_iterate(cl);
2445 2444 }
2446 2445 }
2447 2446
2448 2447 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2449 2448 // FIXME: is this right?
2450 2449 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2451 2450 }
2452 2451
2453 2452 // Calls a SpaceClosure on a HeapRegion.
2454 2453
2455 2454 class SpaceClosureRegionClosure: public HeapRegionClosure {
2456 2455 SpaceClosure* _cl;
2457 2456 public:
2458 2457 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2459 2458 bool doHeapRegion(HeapRegion* r) {
2460 2459 _cl->do_space(r);
2461 2460 return false;
2462 2461 }
2463 2462 };
2464 2463
2465 2464 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2466 2465 SpaceClosureRegionClosure blk(cl);
2467 2466 heap_region_iterate(&blk);
2468 2467 }
2469 2468
2470 2469 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2471 2470 _hrs.iterate(cl);
2472 2471 }
2473 2472
2474 2473 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r,
2475 2474 HeapRegionClosure* cl) const {
2476 2475 _hrs.iterate_from(r, cl);
2477 2476 }
2478 2477
2479 2478 void
2480 2479 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2481 2480 int worker,
2482 2481 jint claim_value) {
2483 2482 const size_t regions = n_regions();
2484 2483 const size_t worker_num = (G1CollectedHeap::use_parallel_gc_threads() ? ParallelGCThreads : 1);
2485 2484 // try to spread out the starting points of the workers
2486 2485 const size_t start_index = regions / worker_num * (size_t) worker;
2487 2486
2488 2487 // each worker will actually look at all regions
2489 2488 for (size_t count = 0; count < regions; ++count) {
2490 2489 const size_t index = (start_index + count) % regions;
2491 2490 assert(0 <= index && index < regions, "sanity");
2492 2491 HeapRegion* r = region_at(index);
2493 2492 // we'll ignore "continues humongous" regions (we'll process them
2494 2493 // when we come across their corresponding "start humongous"
2495 2494 // region) and regions already claimed
2496 2495 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2497 2496 continue;
2498 2497 }
2499 2498 // OK, try to claim it
2500 2499 if (r->claimHeapRegion(claim_value)) {
2501 2500 // success!
2502 2501 assert(!r->continuesHumongous(), "sanity");
2503 2502 if (r->startsHumongous()) {
2504 2503 // If the region is "starts humongous" we'll iterate over its
2505 2504 // "continues humongous" first; in fact we'll do them
2506 2505 // first. The order is important. In on case, calling the
2507 2506 // closure on the "starts humongous" region might de-allocate
2508 2507 // and clear all its "continues humongous" regions and, as a
2509 2508 // result, we might end up processing them twice. So, we'll do
2510 2509 // them first (notice: most closures will ignore them anyway) and
2511 2510 // then we'll do the "starts humongous" region.
2512 2511 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) {
2513 2512 HeapRegion* chr = region_at(ch_index);
2514 2513
2515 2514 // if the region has already been claimed or it's not
2516 2515 // "continues humongous" we're done
2517 2516 if (chr->claim_value() == claim_value ||
2518 2517 !chr->continuesHumongous()) {
2519 2518 break;
2520 2519 }
2521 2520
2522 2521 // Noone should have claimed it directly. We can given
2523 2522 // that we claimed its "starts humongous" region.
2524 2523 assert(chr->claim_value() != claim_value, "sanity");
2525 2524 assert(chr->humongous_start_region() == r, "sanity");
2526 2525
2527 2526 if (chr->claimHeapRegion(claim_value)) {
2528 2527 // we should always be able to claim it; noone else should
2529 2528 // be trying to claim this region
2530 2529
2531 2530 bool res2 = cl->doHeapRegion(chr);
2532 2531 assert(!res2, "Should not abort");
2533 2532
2534 2533 // Right now, this holds (i.e., no closure that actually
2535 2534 // does something with "continues humongous" regions
2536 2535 // clears them). We might have to weaken it in the future,
2537 2536 // but let's leave these two asserts here for extra safety.
2538 2537 assert(chr->continuesHumongous(), "should still be the case");
2539 2538 assert(chr->humongous_start_region() == r, "sanity");
2540 2539 } else {
2541 2540 guarantee(false, "we should not reach here");
2542 2541 }
2543 2542 }
2544 2543 }
2545 2544
2546 2545 assert(!r->continuesHumongous(), "sanity");
2547 2546 bool res = cl->doHeapRegion(r);
2548 2547 assert(!res, "Should not abort");
2549 2548 }
2550 2549 }
2551 2550 }
2552 2551
2553 2552 class ResetClaimValuesClosure: public HeapRegionClosure {
2554 2553 public:
2555 2554 bool doHeapRegion(HeapRegion* r) {
2556 2555 r->set_claim_value(HeapRegion::InitialClaimValue);
2557 2556 return false;
2558 2557 }
2559 2558 };
2560 2559
2561 2560 void
2562 2561 G1CollectedHeap::reset_heap_region_claim_values() {
2563 2562 ResetClaimValuesClosure blk;
2564 2563 heap_region_iterate(&blk);
2565 2564 }
2566 2565
2567 2566 #ifdef ASSERT
2568 2567 // This checks whether all regions in the heap have the correct claim
2569 2568 // value. I also piggy-backed on this a check to ensure that the
2570 2569 // humongous_start_region() information on "continues humongous"
2571 2570 // regions is correct.
2572 2571
2573 2572 class CheckClaimValuesClosure : public HeapRegionClosure {
2574 2573 private:
2575 2574 jint _claim_value;
2576 2575 size_t _failures;
2577 2576 HeapRegion* _sh_region;
2578 2577 public:
2579 2578 CheckClaimValuesClosure(jint claim_value) :
2580 2579 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2581 2580 bool doHeapRegion(HeapRegion* r) {
2582 2581 if (r->claim_value() != _claim_value) {
2583 2582 gclog_or_tty->print_cr("Region ["PTR_FORMAT","PTR_FORMAT"), "
2584 2583 "claim value = %d, should be %d",
2585 2584 r->bottom(), r->end(), r->claim_value(),
2586 2585 _claim_value);
2587 2586 ++_failures;
2588 2587 }
2589 2588 if (!r->isHumongous()) {
2590 2589 _sh_region = NULL;
2591 2590 } else if (r->startsHumongous()) {
2592 2591 _sh_region = r;
2593 2592 } else if (r->continuesHumongous()) {
2594 2593 if (r->humongous_start_region() != _sh_region) {
2595 2594 gclog_or_tty->print_cr("Region ["PTR_FORMAT","PTR_FORMAT"), "
2596 2595 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2597 2596 r->bottom(), r->end(),
2598 2597 r->humongous_start_region(),
2599 2598 _sh_region);
2600 2599 ++_failures;
2601 2600 }
2602 2601 }
2603 2602 return false;
2604 2603 }
2605 2604 size_t failures() {
2606 2605 return _failures;
2607 2606 }
2608 2607 };
2609 2608
2610 2609 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2611 2610 CheckClaimValuesClosure cl(claim_value);
2612 2611 heap_region_iterate(&cl);
2613 2612 return cl.failures() == 0;
2614 2613 }
2615 2614 #endif // ASSERT
2616 2615
2617 2616 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2618 2617 HeapRegion* r = g1_policy()->collection_set();
2619 2618 while (r != NULL) {
2620 2619 HeapRegion* next = r->next_in_collection_set();
2621 2620 if (cl->doHeapRegion(r)) {
2622 2621 cl->incomplete();
2623 2622 return;
2624 2623 }
2625 2624 r = next;
2626 2625 }
2627 2626 }
2628 2627
2629 2628 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2630 2629 HeapRegionClosure *cl) {
2631 2630 if (r == NULL) {
2632 2631 // The CSet is empty so there's nothing to do.
2633 2632 return;
2634 2633 }
2635 2634
2636 2635 assert(r->in_collection_set(),
2637 2636 "Start region must be a member of the collection set.");
2638 2637 HeapRegion* cur = r;
2639 2638 while (cur != NULL) {
2640 2639 HeapRegion* next = cur->next_in_collection_set();
2641 2640 if (cl->doHeapRegion(cur) && false) {
2642 2641 cl->incomplete();
2643 2642 return;
2644 2643 }
2645 2644 cur = next;
2646 2645 }
2647 2646 cur = g1_policy()->collection_set();
2648 2647 while (cur != r) {
2649 2648 HeapRegion* next = cur->next_in_collection_set();
2650 2649 if (cl->doHeapRegion(cur) && false) {
2651 2650 cl->incomplete();
2652 2651 return;
2653 2652 }
2654 2653 cur = next;
2655 2654 }
2656 2655 }
2657 2656
2658 2657 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2659 2658 return n_regions() > 0 ? region_at(0) : NULL;
2660 2659 }
2661 2660
2662 2661
2663 2662 Space* G1CollectedHeap::space_containing(const void* addr) const {
2664 2663 Space* res = heap_region_containing(addr);
2665 2664 if (res == NULL)
2666 2665 res = perm_gen()->space_containing(addr);
2667 2666 return res;
2668 2667 }
2669 2668
2670 2669 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2671 2670 Space* sp = space_containing(addr);
2672 2671 if (sp != NULL) {
2673 2672 return sp->block_start(addr);
2674 2673 }
2675 2674 return NULL;
2676 2675 }
2677 2676
2678 2677 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2679 2678 Space* sp = space_containing(addr);
2680 2679 assert(sp != NULL, "block_size of address outside of heap");
2681 2680 return sp->block_size(addr);
2682 2681 }
2683 2682
2684 2683 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2685 2684 Space* sp = space_containing(addr);
2686 2685 return sp->block_is_obj(addr);
2687 2686 }
2688 2687
2689 2688 bool G1CollectedHeap::supports_tlab_allocation() const {
2690 2689 return true;
2691 2690 }
2692 2691
2693 2692 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2694 2693 return HeapRegion::GrainBytes;
2695 2694 }
2696 2695
2697 2696 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2698 2697 // Return the remaining space in the cur alloc region, but not less than
2699 2698 // the min TLAB size.
2700 2699
2701 2700 // Also, this value can be at most the humongous object threshold,
2702 2701 // since we can't allow tlabs to grow big enough to accomodate
2703 2702 // humongous objects.
2704 2703
2705 2704 HeapRegion* hr = _mutator_alloc_region.get();
2706 2705 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2707 2706 if (hr == NULL) {
2708 2707 return max_tlab_size;
2709 2708 } else {
2710 2709 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2711 2710 }
2712 2711 }
2713 2712
2714 2713 size_t G1CollectedHeap::max_capacity() const {
2715 2714 return _g1_reserved.byte_size();
2716 2715 }
2717 2716
2718 2717 jlong G1CollectedHeap::millis_since_last_gc() {
2719 2718 // assert(false, "NYI");
2720 2719 return 0;
2721 2720 }
2722 2721
2723 2722 void G1CollectedHeap::prepare_for_verify() {
2724 2723 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2725 2724 ensure_parsability(false);
2726 2725 }
2727 2726 g1_rem_set()->prepare_for_verify();
2728 2727 }
2729 2728
2730 2729 class VerifyLivenessOopClosure: public OopClosure {
2731 2730 G1CollectedHeap* _g1h;
2732 2731 VerifyOption _vo;
2733 2732 public:
2734 2733 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
2735 2734 _g1h(g1h), _vo(vo)
2736 2735 { }
2737 2736 void do_oop(narrowOop *p) { do_oop_work(p); }
2738 2737 void do_oop( oop *p) { do_oop_work(p); }
2739 2738
2740 2739 template <class T> void do_oop_work(T *p) {
2741 2740 oop obj = oopDesc::load_decode_heap_oop(p);
2742 2741 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
2743 2742 "Dead object referenced by a not dead object");
2744 2743 }
2745 2744 };
2746 2745
2747 2746 class VerifyObjsInRegionClosure: public ObjectClosure {
2748 2747 private:
2749 2748 G1CollectedHeap* _g1h;
2750 2749 size_t _live_bytes;
2751 2750 HeapRegion *_hr;
2752 2751 VerifyOption _vo;
2753 2752 public:
2754 2753 // _vo == UsePrevMarking -> use "prev" marking information,
2755 2754 // _vo == UseNextMarking -> use "next" marking information,
2756 2755 // _vo == UseMarkWord -> use mark word from object header.
2757 2756 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
2758 2757 : _live_bytes(0), _hr(hr), _vo(vo) {
2759 2758 _g1h = G1CollectedHeap::heap();
2760 2759 }
2761 2760 void do_object(oop o) {
2762 2761 VerifyLivenessOopClosure isLive(_g1h, _vo);
2763 2762 assert(o != NULL, "Huh?");
2764 2763 if (!_g1h->is_obj_dead_cond(o, _vo)) {
2765 2764 // If the object is alive according to the mark word,
2766 2765 // then verify that the marking information agrees.
2767 2766 // Note we can't verify the contra-positive of the
2768 2767 // above: if the object is dead (according to the mark
2769 2768 // word), it may not be marked, or may have been marked
2770 2769 // but has since became dead, or may have been allocated
2771 2770 // since the last marking.
2772 2771 if (_vo == VerifyOption_G1UseMarkWord) {
2773 2772 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
2774 2773 }
2775 2774
2776 2775 o->oop_iterate(&isLive);
2777 2776 if (!_hr->obj_allocated_since_prev_marking(o)) {
2778 2777 size_t obj_size = o->size(); // Make sure we don't overflow
2779 2778 _live_bytes += (obj_size * HeapWordSize);
2780 2779 }
2781 2780 }
2782 2781 }
2783 2782 size_t live_bytes() { return _live_bytes; }
2784 2783 };
2785 2784
2786 2785 class PrintObjsInRegionClosure : public ObjectClosure {
2787 2786 HeapRegion *_hr;
2788 2787 G1CollectedHeap *_g1;
2789 2788 public:
2790 2789 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
2791 2790 _g1 = G1CollectedHeap::heap();
2792 2791 };
2793 2792
2794 2793 void do_object(oop o) {
2795 2794 if (o != NULL) {
2796 2795 HeapWord *start = (HeapWord *) o;
2797 2796 size_t word_sz = o->size();
2798 2797 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
2799 2798 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
2800 2799 (void*) o, word_sz,
2801 2800 _g1->isMarkedPrev(o),
2802 2801 _g1->isMarkedNext(o),
2803 2802 _hr->obj_allocated_since_prev_marking(o));
2804 2803 HeapWord *end = start + word_sz;
2805 2804 HeapWord *cur;
2806 2805 int *val;
2807 2806 for (cur = start; cur < end; cur++) {
2808 2807 val = (int *) cur;
2809 2808 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
2810 2809 }
2811 2810 }
2812 2811 }
2813 2812 };
2814 2813
2815 2814 class VerifyRegionClosure: public HeapRegionClosure {
2816 2815 private:
2817 2816 bool _allow_dirty;
2818 2817 bool _par;
2819 2818 VerifyOption _vo;
2820 2819 bool _failures;
2821 2820 public:
2822 2821 // _vo == UsePrevMarking -> use "prev" marking information,
2823 2822 // _vo == UseNextMarking -> use "next" marking information,
2824 2823 // _vo == UseMarkWord -> use mark word from object header.
2825 2824 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo)
2826 2825 : _allow_dirty(allow_dirty),
2827 2826 _par(par),
2828 2827 _vo(vo),
2829 2828 _failures(false) {}
2830 2829
2831 2830 bool failures() {
2832 2831 return _failures;
2833 2832 }
2834 2833
2835 2834 bool doHeapRegion(HeapRegion* r) {
2836 2835 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue,
2837 2836 "Should be unclaimed at verify points.");
2838 2837 if (!r->continuesHumongous()) {
2839 2838 bool failures = false;
2840 2839 r->verify(_allow_dirty, _vo, &failures);
2841 2840 if (failures) {
2842 2841 _failures = true;
2843 2842 } else {
2844 2843 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
2845 2844 r->object_iterate(¬_dead_yet_cl);
2846 2845 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
2847 2846 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
2848 2847 "max_live_bytes "SIZE_FORMAT" "
2849 2848 "< calculated "SIZE_FORMAT,
2850 2849 r->bottom(), r->end(),
2851 2850 r->max_live_bytes(),
2852 2851 not_dead_yet_cl.live_bytes());
2853 2852 _failures = true;
2854 2853 }
2855 2854 }
2856 2855 }
2857 2856 return false; // stop the region iteration if we hit a failure
2858 2857 }
2859 2858 };
2860 2859
2861 2860 class VerifyRootsClosure: public OopsInGenClosure {
2862 2861 private:
2863 2862 G1CollectedHeap* _g1h;
2864 2863 VerifyOption _vo;
2865 2864 bool _failures;
2866 2865 public:
2867 2866 // _vo == UsePrevMarking -> use "prev" marking information,
2868 2867 // _vo == UseNextMarking -> use "next" marking information,
2869 2868 // _vo == UseMarkWord -> use mark word from object header.
2870 2869 VerifyRootsClosure(VerifyOption vo) :
2871 2870 _g1h(G1CollectedHeap::heap()),
2872 2871 _vo(vo),
2873 2872 _failures(false) { }
2874 2873
2875 2874 bool failures() { return _failures; }
2876 2875
2877 2876 template <class T> void do_oop_nv(T* p) {
2878 2877 T heap_oop = oopDesc::load_heap_oop(p);
2879 2878 if (!oopDesc::is_null(heap_oop)) {
2880 2879 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
2881 2880 if (_g1h->is_obj_dead_cond(obj, _vo)) {
2882 2881 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
2883 2882 "points to dead obj "PTR_FORMAT, p, (void*) obj);
2884 2883 if (_vo == VerifyOption_G1UseMarkWord) {
2885 2884 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
2886 2885 }
2887 2886 obj->print_on(gclog_or_tty);
2888 2887 _failures = true;
2889 2888 }
2890 2889 }
2891 2890 }
2892 2891
2893 2892 void do_oop(oop* p) { do_oop_nv(p); }
2894 2893 void do_oop(narrowOop* p) { do_oop_nv(p); }
2895 2894 };
2896 2895
2897 2896 // This is the task used for parallel heap verification.
2898 2897
2899 2898 class G1ParVerifyTask: public AbstractGangTask {
2900 2899 private:
2901 2900 G1CollectedHeap* _g1h;
2902 2901 bool _allow_dirty;
2903 2902 VerifyOption _vo;
2904 2903 bool _failures;
2905 2904
2906 2905 public:
2907 2906 // _vo == UsePrevMarking -> use "prev" marking information,
2908 2907 // _vo == UseNextMarking -> use "next" marking information,
2909 2908 // _vo == UseMarkWord -> use mark word from object header.
2910 2909 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) :
2911 2910 AbstractGangTask("Parallel verify task"),
2912 2911 _g1h(g1h),
2913 2912 _allow_dirty(allow_dirty),
2914 2913 _vo(vo),
2915 2914 _failures(false) { }
2916 2915
2917 2916 bool failures() {
2918 2917 return _failures;
2919 2918 }
2920 2919
2921 2920 void work(int worker_i) {
2922 2921 HandleMark hm;
2923 2922 VerifyRegionClosure blk(_allow_dirty, true, _vo);
2924 2923 _g1h->heap_region_par_iterate_chunked(&blk, worker_i,
2925 2924 HeapRegion::ParVerifyClaimValue);
2926 2925 if (blk.failures()) {
2927 2926 _failures = true;
2928 2927 }
2929 2928 }
2930 2929 };
2931 2930
2932 2931 void G1CollectedHeap::verify(bool allow_dirty, bool silent) {
2933 2932 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking);
2934 2933 }
2935 2934
2936 2935 void G1CollectedHeap::verify(bool allow_dirty,
2937 2936 bool silent,
2938 2937 VerifyOption vo) {
2939 2938 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2940 2939 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); }
2941 2940 VerifyRootsClosure rootsCl(vo);
2942 2941 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
2943 2942
2944 2943 // We apply the relevant closures to all the oops in the
2945 2944 // system dictionary, the string table and the code cache.
2946 2945 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
2947 2946
2948 2947 process_strong_roots(true, // activate StrongRootsScope
2949 2948 true, // we set "collecting perm gen" to true,
2950 2949 // so we don't reset the dirty cards in the perm gen.
2951 2950 SharedHeap::ScanningOption(so), // roots scanning options
2952 2951 &rootsCl,
2953 2952 &blobsCl,
2954 2953 &rootsCl);
2955 2954
2956 2955 // If we're verifying after the marking phase of a Full GC then we can't
2957 2956 // treat the perm gen as roots into the G1 heap. Some of the objects in
2958 2957 // the perm gen may be dead and hence not marked. If one of these dead
2959 2958 // objects is considered to be a root then we may end up with a false
2960 2959 // "Root location <x> points to dead ob <y>" failure.
2961 2960 if (vo != VerifyOption_G1UseMarkWord) {
2962 2961 // Since we used "collecting_perm_gen" == true above, we will not have
2963 2962 // checked the refs from perm into the G1-collected heap. We check those
2964 2963 // references explicitly below. Whether the relevant cards are dirty
2965 2964 // is checked further below in the rem set verification.
2966 2965 if (!silent) { gclog_or_tty->print("Permgen roots "); }
2967 2966 perm_gen()->oop_iterate(&rootsCl);
2968 2967 }
2969 2968 bool failures = rootsCl.failures();
2970 2969
2971 2970 if (vo != VerifyOption_G1UseMarkWord) {
2972 2971 // If we're verifying during a full GC then the region sets
2973 2972 // will have been torn down at the start of the GC. Therefore
2974 2973 // verifying the region sets will fail. So we only verify
2975 2974 // the region sets when not in a full GC.
2976 2975 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
2977 2976 verify_region_sets();
2978 2977 }
2979 2978
2980 2979 if (!silent) { gclog_or_tty->print("HeapRegions "); }
2981 2980 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
2982 2981 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2983 2982 "sanity check");
2984 2983
2985 2984 G1ParVerifyTask task(this, allow_dirty, vo);
2986 2985 int n_workers = workers()->total_workers();
2987 2986 set_par_threads(n_workers);
2988 2987 workers()->run_task(&task);
2989 2988 set_par_threads(0);
2990 2989 if (task.failures()) {
2991 2990 failures = true;
2992 2991 }
2993 2992
2994 2993 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
2995 2994 "sanity check");
2996 2995
2997 2996 reset_heap_region_claim_values();
2998 2997
2999 2998 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3000 2999 "sanity check");
3001 3000 } else {
3002 3001 VerifyRegionClosure blk(allow_dirty, false, vo);
3003 3002 heap_region_iterate(&blk);
3004 3003 if (blk.failures()) {
3005 3004 failures = true;
3006 3005 }
3007 3006 }
3008 3007 if (!silent) gclog_or_tty->print("RemSet ");
3009 3008 rem_set()->verify();
3010 3009
3011 3010 if (failures) {
3012 3011 gclog_or_tty->print_cr("Heap:");
3013 3012 print_on(gclog_or_tty, true /* extended */);
3014 3013 gclog_or_tty->print_cr("");
3015 3014 #ifndef PRODUCT
3016 3015 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3017 3016 concurrent_mark()->print_reachable("at-verification-failure",
3018 3017 vo, false /* all */);
3019 3018 }
3020 3019 #endif
3021 3020 gclog_or_tty->flush();
3022 3021 }
3023 3022 guarantee(!failures, "there should not have been any failures");
3024 3023 } else {
3025 3024 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3026 3025 }
3027 3026 }
3028 3027
3029 3028 class PrintRegionClosure: public HeapRegionClosure {
3030 3029 outputStream* _st;
3031 3030 public:
3032 3031 PrintRegionClosure(outputStream* st) : _st(st) {}
3033 3032 bool doHeapRegion(HeapRegion* r) {
3034 3033 r->print_on(_st);
3035 3034 return false;
3036 3035 }
3037 3036 };
3038 3037
3039 3038 void G1CollectedHeap::print() const { print_on(tty); }
3040 3039
3041 3040 void G1CollectedHeap::print_on(outputStream* st) const {
3042 3041 print_on(st, PrintHeapAtGCExtended);
3043 3042 }
↓ open down ↓ |
1059 lines elided |
↑ open up ↑ |
3044 3043
3045 3044 void G1CollectedHeap::print_on(outputStream* st, bool extended) const {
3046 3045 st->print(" %-20s", "garbage-first heap");
3047 3046 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3048 3047 capacity()/K, used_unlocked()/K);
3049 3048 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3050 3049 _g1_storage.low_boundary(),
3051 3050 _g1_storage.high(),
3052 3051 _g1_storage.high_boundary());
3053 3052 st->cr();
3054 - st->print(" region size " SIZE_FORMAT "K, ",
3055 - HeapRegion::GrainBytes/K);
3053 + st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3056 3054 size_t young_regions = _young_list->length();
3057 3055 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ",
3058 3056 young_regions, young_regions * HeapRegion::GrainBytes / K);
3059 3057 size_t survivor_regions = g1_policy()->recorded_survivor_regions();
3060 3058 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)",
3061 3059 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K);
3062 3060 st->cr();
3063 3061 perm()->as_gen()->print_on(st);
3064 3062 if (extended) {
3065 3063 st->cr();
3066 3064 print_on_extended(st);
3067 3065 }
3068 3066 }
3069 3067
3070 3068 void G1CollectedHeap::print_on_extended(outputStream* st) const {
3071 3069 PrintRegionClosure blk(st);
3072 3070 heap_region_iterate(&blk);
3073 3071 }
3074 3072
3075 3073 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3076 3074 if (G1CollectedHeap::use_parallel_gc_threads()) {
3077 3075 workers()->print_worker_threads_on(st);
3078 3076 }
3079 3077 _cmThread->print_on(st);
3080 3078 st->cr();
3081 3079 _cm->print_worker_threads_on(st);
3082 3080 _cg1r->print_worker_threads_on(st);
3083 3081 st->cr();
3084 3082 }
3085 3083
3086 3084 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3087 3085 if (G1CollectedHeap::use_parallel_gc_threads()) {
3088 3086 workers()->threads_do(tc);
3089 3087 }
3090 3088 tc->do_thread(_cmThread);
3091 3089 _cg1r->threads_do(tc);
3092 3090 }
3093 3091
3094 3092 void G1CollectedHeap::print_tracing_info() const {
3095 3093 // We'll overload this to mean "trace GC pause statistics."
3096 3094 if (TraceGen0Time || TraceGen1Time) {
3097 3095 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3098 3096 // to that.
3099 3097 g1_policy()->print_tracing_info();
3100 3098 }
3101 3099 if (G1SummarizeRSetStats) {
3102 3100 g1_rem_set()->print_summary_info();
3103 3101 }
3104 3102 if (G1SummarizeConcMark) {
3105 3103 concurrent_mark()->print_summary_info();
3106 3104 }
3107 3105 g1_policy()->print_yg_surv_rate_info();
3108 3106 SpecializationStats::print();
3109 3107 }
3110 3108
3111 3109 #ifndef PRODUCT
3112 3110 // Helpful for debugging RSet issues.
3113 3111
3114 3112 class PrintRSetsClosure : public HeapRegionClosure {
3115 3113 private:
3116 3114 const char* _msg;
3117 3115 size_t _occupied_sum;
3118 3116
3119 3117 public:
3120 3118 bool doHeapRegion(HeapRegion* r) {
3121 3119 HeapRegionRemSet* hrrs = r->rem_set();
3122 3120 size_t occupied = hrrs->occupied();
3123 3121 _occupied_sum += occupied;
3124 3122
3125 3123 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3126 3124 HR_FORMAT_PARAMS(r));
3127 3125 if (occupied == 0) {
3128 3126 gclog_or_tty->print_cr(" RSet is empty");
3129 3127 } else {
3130 3128 hrrs->print();
3131 3129 }
3132 3130 gclog_or_tty->print_cr("----------");
3133 3131 return false;
3134 3132 }
3135 3133
3136 3134 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3137 3135 gclog_or_tty->cr();
3138 3136 gclog_or_tty->print_cr("========================================");
3139 3137 gclog_or_tty->print_cr(msg);
3140 3138 gclog_or_tty->cr();
3141 3139 }
3142 3140
3143 3141 ~PrintRSetsClosure() {
3144 3142 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3145 3143 gclog_or_tty->print_cr("========================================");
3146 3144 gclog_or_tty->cr();
3147 3145 }
3148 3146 };
3149 3147
3150 3148 void G1CollectedHeap::print_cset_rsets() {
3151 3149 PrintRSetsClosure cl("Printing CSet RSets");
3152 3150 collection_set_iterate(&cl);
3153 3151 }
3154 3152
3155 3153 void G1CollectedHeap::print_all_rsets() {
3156 3154 PrintRSetsClosure cl("Printing All RSets");;
3157 3155 heap_region_iterate(&cl);
3158 3156 }
3159 3157 #endif // PRODUCT
3160 3158
3161 3159 G1CollectedHeap* G1CollectedHeap::heap() {
3162 3160 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3163 3161 "not a garbage-first heap");
3164 3162 return _g1h;
3165 3163 }
3166 3164
3167 3165 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3168 3166 // always_do_update_barrier = false;
3169 3167 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3170 3168 // Call allocation profiler
3171 3169 AllocationProfiler::iterate_since_last_gc();
3172 3170 // Fill TLAB's and such
3173 3171 ensure_parsability(true);
3174 3172 }
3175 3173
3176 3174 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3177 3175 // FIXME: what is this about?
3178 3176 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3179 3177 // is set.
3180 3178 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3181 3179 "derived pointer present"));
3182 3180 // always_do_update_barrier = true;
3183 3181
3184 3182 // We have just completed a GC. Update the soft reference
3185 3183 // policy with the new heap occupancy
3186 3184 Universe::update_heap_info_at_gc();
3187 3185 }
3188 3186
3189 3187 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3190 3188 unsigned int gc_count_before,
3191 3189 bool* succeeded) {
3192 3190 assert_heap_not_locked_and_not_at_safepoint();
3193 3191 g1_policy()->record_stop_world_start();
3194 3192 VM_G1IncCollectionPause op(gc_count_before,
3195 3193 word_size,
3196 3194 false, /* should_initiate_conc_mark */
3197 3195 g1_policy()->max_pause_time_ms(),
3198 3196 GCCause::_g1_inc_collection_pause);
3199 3197 VMThread::execute(&op);
3200 3198
3201 3199 HeapWord* result = op.result();
3202 3200 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3203 3201 assert(result == NULL || ret_succeeded,
3204 3202 "the result should be NULL if the VM did not succeed");
3205 3203 *succeeded = ret_succeeded;
3206 3204
3207 3205 assert_heap_not_locked();
3208 3206 return result;
3209 3207 }
3210 3208
3211 3209 void
3212 3210 G1CollectedHeap::doConcurrentMark() {
3213 3211 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3214 3212 if (!_cmThread->in_progress()) {
3215 3213 _cmThread->set_started();
3216 3214 CGC_lock->notify();
3217 3215 }
3218 3216 }
3219 3217
3220 3218 // <NEW PREDICTION>
3221 3219
3222 3220 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr,
3223 3221 bool young) {
3224 3222 return _g1_policy->predict_region_elapsed_time_ms(hr, young);
3225 3223 }
3226 3224
3227 3225 void G1CollectedHeap::check_if_region_is_too_expensive(double
3228 3226 predicted_time_ms) {
3229 3227 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms);
3230 3228 }
3231 3229
3232 3230 size_t G1CollectedHeap::pending_card_num() {
3233 3231 size_t extra_cards = 0;
3234 3232 JavaThread *curr = Threads::first();
3235 3233 while (curr != NULL) {
3236 3234 DirtyCardQueue& dcq = curr->dirty_card_queue();
3237 3235 extra_cards += dcq.size();
3238 3236 curr = curr->next();
3239 3237 }
3240 3238 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3241 3239 size_t buffer_size = dcqs.buffer_size();
3242 3240 size_t buffer_num = dcqs.completed_buffers_num();
3243 3241 return buffer_size * buffer_num + extra_cards;
3244 3242 }
3245 3243
3246 3244 size_t G1CollectedHeap::max_pending_card_num() {
3247 3245 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3248 3246 size_t buffer_size = dcqs.buffer_size();
3249 3247 size_t buffer_num = dcqs.completed_buffers_num();
3250 3248 int thread_num = Threads::number_of_threads();
3251 3249 return (buffer_num + thread_num) * buffer_size;
3252 3250 }
3253 3251
3254 3252 size_t G1CollectedHeap::cards_scanned() {
3255 3253 return g1_rem_set()->cardsScanned();
3256 3254 }
3257 3255
3258 3256 void
3259 3257 G1CollectedHeap::setup_surviving_young_words() {
3260 3258 guarantee( _surviving_young_words == NULL, "pre-condition" );
3261 3259 size_t array_length = g1_policy()->young_cset_length();
3262 3260 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length);
3263 3261 if (_surviving_young_words == NULL) {
3264 3262 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3265 3263 "Not enough space for young surv words summary.");
3266 3264 }
3267 3265 memset(_surviving_young_words, 0, array_length * sizeof(size_t));
3268 3266 #ifdef ASSERT
3269 3267 for (size_t i = 0; i < array_length; ++i) {
3270 3268 assert( _surviving_young_words[i] == 0, "memset above" );
3271 3269 }
3272 3270 #endif // !ASSERT
3273 3271 }
3274 3272
3275 3273 void
3276 3274 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3277 3275 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3278 3276 size_t array_length = g1_policy()->young_cset_length();
3279 3277 for (size_t i = 0; i < array_length; ++i)
3280 3278 _surviving_young_words[i] += surv_young_words[i];
3281 3279 }
3282 3280
3283 3281 void
3284 3282 G1CollectedHeap::cleanup_surviving_young_words() {
3285 3283 guarantee( _surviving_young_words != NULL, "pre-condition" );
3286 3284 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words);
3287 3285 _surviving_young_words = NULL;
3288 3286 }
3289 3287
3290 3288 // </NEW PREDICTION>
3291 3289
3292 3290 #ifdef ASSERT
3293 3291 class VerifyCSetClosure: public HeapRegionClosure {
3294 3292 public:
3295 3293 bool doHeapRegion(HeapRegion* hr) {
3296 3294 // Here we check that the CSet region's RSet is ready for parallel
3297 3295 // iteration. The fields that we'll verify are only manipulated
3298 3296 // when the region is part of a CSet and is collected. Afterwards,
3299 3297 // we reset these fields when we clear the region's RSet (when the
3300 3298 // region is freed) so they are ready when the region is
3301 3299 // re-allocated. The only exception to this is if there's an
3302 3300 // evacuation failure and instead of freeing the region we leave
3303 3301 // it in the heap. In that case, we reset these fields during
3304 3302 // evacuation failure handling.
3305 3303 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3306 3304
3307 3305 // Here's a good place to add any other checks we'd like to
3308 3306 // perform on CSet regions.
3309 3307 return false;
3310 3308 }
3311 3309 };
3312 3310 #endif // ASSERT
3313 3311
3314 3312 #if TASKQUEUE_STATS
3315 3313 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3316 3314 st->print_raw_cr("GC Task Stats");
3317 3315 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3318 3316 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3319 3317 }
3320 3318
3321 3319 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3322 3320 print_taskqueue_stats_hdr(st);
3323 3321
3324 3322 TaskQueueStats totals;
3325 3323 const int n = workers() != NULL ? workers()->total_workers() : 1;
3326 3324 for (int i = 0; i < n; ++i) {
3327 3325 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3328 3326 totals += task_queue(i)->stats;
3329 3327 }
3330 3328 st->print_raw("tot "); totals.print(st); st->cr();
3331 3329
3332 3330 DEBUG_ONLY(totals.verify());
3333 3331 }
3334 3332
3335 3333 void G1CollectedHeap::reset_taskqueue_stats() {
3336 3334 const int n = workers() != NULL ? workers()->total_workers() : 1;
3337 3335 for (int i = 0; i < n; ++i) {
3338 3336 task_queue(i)->stats.reset();
3339 3337 }
3340 3338 }
3341 3339 #endif // TASKQUEUE_STATS
3342 3340
3343 3341 bool
3344 3342 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3345 3343 assert_at_safepoint(true /* should_be_vm_thread */);
3346 3344 guarantee(!is_gc_active(), "collection is not reentrant");
3347 3345
3348 3346 if (GC_locker::check_active_before_gc()) {
3349 3347 return false;
3350 3348 }
3351 3349
3352 3350 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3353 3351 ResourceMark rm;
3354 3352
3355 3353 if (PrintHeapAtGC) {
3356 3354 Universe::print_heap_before_gc();
3357 3355 }
3358 3356
3359 3357 verify_region_sets_optional();
3360 3358 verify_dirty_young_regions();
3361 3359
3362 3360 {
3363 3361 // This call will decide whether this pause is an initial-mark
3364 3362 // pause. If it is, during_initial_mark_pause() will return true
3365 3363 // for the duration of this pause.
3366 3364 g1_policy()->decide_on_conc_mark_initiation();
3367 3365
3368 3366 char verbose_str[128];
3369 3367 sprintf(verbose_str, "GC pause ");
3370 3368 if (g1_policy()->full_young_gcs()) {
3371 3369 strcat(verbose_str, "(young)");
3372 3370 } else {
3373 3371 strcat(verbose_str, "(partial)");
3374 3372 }
3375 3373 if (g1_policy()->during_initial_mark_pause()) {
3376 3374 strcat(verbose_str, " (initial-mark)");
3377 3375 // We are about to start a marking cycle, so we increment the
3378 3376 // full collection counter.
3379 3377 increment_total_full_collections();
3380 3378 }
3381 3379
3382 3380 // if PrintGCDetails is on, we'll print long statistics information
3383 3381 // in the collector policy code, so let's not print this as the output
3384 3382 // is messy if we do.
3385 3383 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
3386 3384 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3387 3385 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty);
3388 3386
3389 3387 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3390 3388 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3391 3389
3392 3390 // If the secondary_free_list is not empty, append it to the
3393 3391 // free_list. No need to wait for the cleanup operation to finish;
3394 3392 // the region allocation code will check the secondary_free_list
3395 3393 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3396 3394 // set, skip this step so that the region allocation code has to
3397 3395 // get entries from the secondary_free_list.
3398 3396 if (!G1StressConcRegionFreeing) {
3399 3397 append_secondary_free_list_if_not_empty_with_lock();
3400 3398 }
3401 3399
3402 3400 assert(check_young_list_well_formed(),
3403 3401 "young list should be well formed");
3404 3402
3405 3403 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3406 3404 IsGCActiveMark x;
3407 3405
3408 3406 gc_prologue(false);
3409 3407 increment_total_collections(false /* full gc */);
3410 3408 increment_gc_time_stamp();
3411 3409
3412 3410 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
3413 3411 HandleMark hm; // Discard invalid handles created during verification
3414 3412 gclog_or_tty->print(" VerifyBeforeGC:");
3415 3413 prepare_for_verify();
3416 3414 Universe::verify(/* allow dirty */ false,
3417 3415 /* silent */ false,
3418 3416 /* option */ VerifyOption_G1UsePrevMarking);
3419 3417
3420 3418 }
3421 3419
3422 3420 COMPILER2_PRESENT(DerivedPointerTable::clear());
3423 3421
3424 3422 // Please see comment in g1CollectedHeap.hpp and
3425 3423 // G1CollectedHeap::ref_processing_init() to see how
3426 3424 // reference processing currently works in G1.
3427 3425
3428 3426 // Enable discovery in the STW reference processor
3429 3427 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3430 3428 true /*verify_no_refs*/);
3431 3429
3432 3430 {
3433 3431 // We want to temporarily turn off discovery by the
3434 3432 // CM ref processor, if necessary, and turn it back on
3435 3433 // on again later if we do. Using a scoped
3436 3434 // NoRefDiscovery object will do this.
3437 3435 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3438 3436
3439 3437 // Forget the current alloc region (we might even choose it to be part
3440 3438 // of the collection set!).
3441 3439 release_mutator_alloc_region();
3442 3440
3443 3441 // We should call this after we retire the mutator alloc
3444 3442 // region(s) so that all the ALLOC / RETIRE events are generated
3445 3443 // before the start GC event.
3446 3444 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3447 3445
3448 3446 // The elapsed time induced by the start time below deliberately elides
3449 3447 // the possible verification above.
3450 3448 double start_time_sec = os::elapsedTime();
3451 3449 size_t start_used_bytes = used();
3452 3450
3453 3451 #if YOUNG_LIST_VERBOSE
3454 3452 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3455 3453 _young_list->print();
3456 3454 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3457 3455 #endif // YOUNG_LIST_VERBOSE
3458 3456
3459 3457 g1_policy()->record_collection_pause_start(start_time_sec,
3460 3458 start_used_bytes);
3461 3459
3462 3460 #if YOUNG_LIST_VERBOSE
3463 3461 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3464 3462 _young_list->print();
3465 3463 #endif // YOUNG_LIST_VERBOSE
3466 3464
3467 3465 if (g1_policy()->during_initial_mark_pause()) {
3468 3466 concurrent_mark()->checkpointRootsInitialPre();
3469 3467 }
3470 3468 perm_gen()->save_marks();
3471 3469
3472 3470 // We must do this before any possible evacuation that should propagate
3473 3471 // marks.
3474 3472 if (mark_in_progress()) {
3475 3473 double start_time_sec = os::elapsedTime();
3476 3474
3477 3475 _cm->drainAllSATBBuffers();
3478 3476 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0;
3479 3477 g1_policy()->record_satb_drain_time(finish_mark_ms);
3480 3478 }
3481 3479 // Record the number of elements currently on the mark stack, so we
3482 3480 // only iterate over these. (Since evacuation may add to the mark
3483 3481 // stack, doing more exposes race conditions.) If no mark is in
3484 3482 // progress, this will be zero.
3485 3483 _cm->set_oops_do_bound();
3486 3484
3487 3485 if (mark_in_progress()) {
3488 3486 concurrent_mark()->newCSet();
3489 3487 }
3490 3488
3491 3489 #if YOUNG_LIST_VERBOSE
3492 3490 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3493 3491 _young_list->print();
3494 3492 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3495 3493 #endif // YOUNG_LIST_VERBOSE
3496 3494
3497 3495 g1_policy()->choose_collection_set(target_pause_time_ms);
3498 3496
3499 3497 if (_hr_printer.is_active()) {
3500 3498 HeapRegion* hr = g1_policy()->collection_set();
3501 3499 while (hr != NULL) {
3502 3500 G1HRPrinter::RegionType type;
3503 3501 if (!hr->is_young()) {
3504 3502 type = G1HRPrinter::Old;
3505 3503 } else if (hr->is_survivor()) {
3506 3504 type = G1HRPrinter::Survivor;
3507 3505 } else {
3508 3506 type = G1HRPrinter::Eden;
3509 3507 }
3510 3508 _hr_printer.cset(hr);
3511 3509 hr = hr->next_in_collection_set();
3512 3510 }
3513 3511 }
3514 3512
3515 3513 // We have chosen the complete collection set. If marking is
3516 3514 // active then, we clear the region fields of any of the
3517 3515 // concurrent marking tasks whose region fields point into
3518 3516 // the collection set as these values will become stale. This
3519 3517 // will cause the owning marking threads to claim a new region
3520 3518 // when marking restarts.
3521 3519 if (mark_in_progress()) {
3522 3520 concurrent_mark()->reset_active_task_region_fields_in_cset();
3523 3521 }
3524 3522
3525 3523 #ifdef ASSERT
3526 3524 VerifyCSetClosure cl;
3527 3525 collection_set_iterate(&cl);
3528 3526 #endif // ASSERT
3529 3527
3530 3528 setup_surviving_young_words();
3531 3529
3532 3530 // Initialize the GC alloc regions.
3533 3531 init_gc_alloc_regions();
3534 3532
3535 3533 // Actually do the work...
3536 3534 evacuate_collection_set();
3537 3535
3538 3536 free_collection_set(g1_policy()->collection_set());
3539 3537 g1_policy()->clear_collection_set();
3540 3538
3541 3539 cleanup_surviving_young_words();
3542 3540
3543 3541 // Start a new incremental collection set for the next pause.
3544 3542 g1_policy()->start_incremental_cset_building();
3545 3543
3546 3544 // Clear the _cset_fast_test bitmap in anticipation of adding
3547 3545 // regions to the incremental collection set for the next
3548 3546 // evacuation pause.
3549 3547 clear_cset_fast_test();
3550 3548
3551 3549 _young_list->reset_sampled_info();
3552 3550
3553 3551 // Don't check the whole heap at this point as the
3554 3552 // GC alloc regions from this pause have been tagged
3555 3553 // as survivors and moved on to the survivor list.
3556 3554 // Survivor regions will fail the !is_young() check.
3557 3555 assert(check_young_list_empty(false /* check_heap */),
3558 3556 "young list should be empty");
3559 3557
3560 3558 #if YOUNG_LIST_VERBOSE
3561 3559 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3562 3560 _young_list->print();
3563 3561 #endif // YOUNG_LIST_VERBOSE
3564 3562
3565 3563 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3566 3564 _young_list->first_survivor_region(),
3567 3565 _young_list->last_survivor_region());
3568 3566
3569 3567 _young_list->reset_auxilary_lists();
3570 3568
3571 3569 if (evacuation_failed()) {
3572 3570 _summary_bytes_used = recalculate_used();
3573 3571 } else {
3574 3572 // The "used" of the the collection set have already been subtracted
3575 3573 // when they were freed. Add in the bytes evacuated.
3576 3574 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3577 3575 }
3578 3576
3579 3577 if (g1_policy()->during_initial_mark_pause()) {
3580 3578 concurrent_mark()->checkpointRootsInitialPost();
3581 3579 set_marking_started();
3582 3580 // CAUTION: after the doConcurrentMark() call below,
3583 3581 // the concurrent marking thread(s) could be running
3584 3582 // concurrently with us. Make sure that anything after
3585 3583 // this point does not assume that we are the only GC thread
3586 3584 // running. Note: of course, the actual marking work will
3587 3585 // not start until the safepoint itself is released in
3588 3586 // ConcurrentGCThread::safepoint_desynchronize().
3589 3587 doConcurrentMark();
3590 3588 }
3591 3589
3592 3590 allocate_dummy_regions();
3593 3591
3594 3592 #if YOUNG_LIST_VERBOSE
3595 3593 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3596 3594 _young_list->print();
3597 3595 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3598 3596 #endif // YOUNG_LIST_VERBOSE
3599 3597
3600 3598 init_mutator_alloc_region();
3601 3599
3602 3600 {
3603 3601 size_t expand_bytes = g1_policy()->expansion_amount();
3604 3602 if (expand_bytes > 0) {
3605 3603 size_t bytes_before = capacity();
3606 3604 if (!expand(expand_bytes)) {
3607 3605 // We failed to expand the heap so let's verify that
3608 3606 // committed/uncommitted amount match the backing store
3609 3607 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3610 3608 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3611 3609 }
3612 3610 }
3613 3611 }
3614 3612
3615 3613 double end_time_sec = os::elapsedTime();
3616 3614 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS;
3617 3615 g1_policy()->record_pause_time_ms(pause_time_ms);
3618 3616 g1_policy()->record_collection_pause_end();
3619 3617
3620 3618 MemoryService::track_memory_usage();
3621 3619
3622 3620 // In prepare_for_verify() below we'll need to scan the deferred
3623 3621 // update buffers to bring the RSets up-to-date if
3624 3622 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3625 3623 // the update buffers we'll probably need to scan cards on the
3626 3624 // regions we just allocated to (i.e., the GC alloc
3627 3625 // regions). However, during the last GC we called
3628 3626 // set_saved_mark() on all the GC alloc regions, so card
3629 3627 // scanning might skip the [saved_mark_word()...top()] area of
3630 3628 // those regions (i.e., the area we allocated objects into
3631 3629 // during the last GC). But it shouldn't. Given that
3632 3630 // saved_mark_word() is conditional on whether the GC time stamp
3633 3631 // on the region is current or not, by incrementing the GC time
3634 3632 // stamp here we invalidate all the GC time stamps on all the
3635 3633 // regions and saved_mark_word() will simply return top() for
3636 3634 // all the regions. This is a nicer way of ensuring this rather
3637 3635 // than iterating over the regions and fixing them. In fact, the
3638 3636 // GC time stamp increment here also ensures that
3639 3637 // saved_mark_word() will return top() between pauses, i.e.,
3640 3638 // during concurrent refinement. So we don't need the
3641 3639 // is_gc_active() check to decided which top to use when
3642 3640 // scanning cards (see CR 7039627).
3643 3641 increment_gc_time_stamp();
3644 3642
3645 3643 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
3646 3644 HandleMark hm; // Discard invalid handles created during verification
3647 3645 gclog_or_tty->print(" VerifyAfterGC:");
3648 3646 prepare_for_verify();
3649 3647 Universe::verify(/* allow dirty */ true,
3650 3648 /* silent */ false,
3651 3649 /* option */ VerifyOption_G1UsePrevMarking);
3652 3650 }
3653 3651
3654 3652 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
3655 3653 ref_processor_stw()->verify_no_references_recorded();
3656 3654
3657 3655 // CM reference discovery will be re-enabled if necessary.
3658 3656 }
3659 3657
3660 3658 {
3661 3659 size_t expand_bytes = g1_policy()->expansion_amount();
3662 3660 if (expand_bytes > 0) {
3663 3661 size_t bytes_before = capacity();
3664 3662 // No need for an ergo verbose message here,
3665 3663 // expansion_amount() does this when it returns a value > 0.
3666 3664 if (!expand(expand_bytes)) {
3667 3665 // We failed to expand the heap so let's verify that
3668 3666 // committed/uncommitted amount match the backing store
3669 3667 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3670 3668 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3671 3669 }
3672 3670 }
3673 3671 }
3674 3672
3675 3673 // We should do this after we potentially expand the heap so
3676 3674 // that all the COMMIT events are generated before the end GC
3677 3675 // event, and after we retire the GC alloc regions so that all
3678 3676 // RETIRE events are generated before the end GC event.
3679 3677 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
3680 3678
3681 3679 // We have to do this after we decide whether to expand the heap or not.
3682 3680 g1_policy()->print_heap_transition();
3683 3681
3684 3682 if (mark_in_progress()) {
3685 3683 concurrent_mark()->update_g1_committed();
3686 3684 }
3687 3685
3688 3686 #ifdef TRACESPINNING
3689 3687 ParallelTaskTerminator::print_termination_counts();
3690 3688 #endif
3691 3689
3692 3690 gc_epilogue(false);
3693 3691 }
3694 3692
3695 3693 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) {
3696 3694 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum);
3697 3695 print_tracing_info();
3698 3696 vm_exit(-1);
3699 3697 }
3700 3698 }
3701 3699
3702 3700 _hrs.verify_optional();
3703 3701 verify_region_sets_optional();
3704 3702
3705 3703 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
3706 3704 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
3707 3705
3708 3706 if (PrintHeapAtGC) {
3709 3707 Universe::print_heap_after_gc();
3710 3708 }
3711 3709 g1mm()->update_sizes();
3712 3710
3713 3711 if (G1SummarizeRSetStats &&
3714 3712 (G1SummarizeRSetStatsPeriod > 0) &&
3715 3713 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
3716 3714 g1_rem_set()->print_summary_info();
3717 3715 }
3718 3716
3719 3717 return true;
3720 3718 }
3721 3719
3722 3720 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
3723 3721 {
3724 3722 size_t gclab_word_size;
3725 3723 switch (purpose) {
3726 3724 case GCAllocForSurvived:
3727 3725 gclab_word_size = YoungPLABSize;
3728 3726 break;
3729 3727 case GCAllocForTenured:
3730 3728 gclab_word_size = OldPLABSize;
3731 3729 break;
3732 3730 default:
3733 3731 assert(false, "unknown GCAllocPurpose");
3734 3732 gclab_word_size = OldPLABSize;
3735 3733 break;
3736 3734 }
3737 3735 return gclab_word_size;
3738 3736 }
3739 3737
3740 3738 void G1CollectedHeap::init_mutator_alloc_region() {
3741 3739 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
3742 3740 _mutator_alloc_region.init();
3743 3741 }
3744 3742
3745 3743 void G1CollectedHeap::release_mutator_alloc_region() {
3746 3744 _mutator_alloc_region.release();
3747 3745 assert(_mutator_alloc_region.get() == NULL, "post-condition");
3748 3746 }
3749 3747
3750 3748 void G1CollectedHeap::init_gc_alloc_regions() {
3751 3749 assert_at_safepoint(true /* should_be_vm_thread */);
3752 3750
3753 3751 _survivor_gc_alloc_region.init();
3754 3752 _old_gc_alloc_region.init();
3755 3753 HeapRegion* retained_region = _retained_old_gc_alloc_region;
3756 3754 _retained_old_gc_alloc_region = NULL;
3757 3755
3758 3756 // We will discard the current GC alloc region if:
3759 3757 // a) it's in the collection set (it can happen!),
3760 3758 // b) it's already full (no point in using it),
3761 3759 // c) it's empty (this means that it was emptied during
3762 3760 // a cleanup and it should be on the free list now), or
3763 3761 // d) it's humongous (this means that it was emptied
3764 3762 // during a cleanup and was added to the free list, but
3765 3763 // has been subseqently used to allocate a humongous
3766 3764 // object that may be less than the region size).
3767 3765 if (retained_region != NULL &&
3768 3766 !retained_region->in_collection_set() &&
3769 3767 !(retained_region->top() == retained_region->end()) &&
3770 3768 !retained_region->is_empty() &&
3771 3769 !retained_region->isHumongous()) {
3772 3770 retained_region->set_saved_mark();
3773 3771 _old_gc_alloc_region.set(retained_region);
3774 3772 _hr_printer.reuse(retained_region);
3775 3773 }
3776 3774 }
3777 3775
3778 3776 void G1CollectedHeap::release_gc_alloc_regions() {
3779 3777 _survivor_gc_alloc_region.release();
3780 3778 // If we have an old GC alloc region to release, we'll save it in
3781 3779 // _retained_old_gc_alloc_region. If we don't
3782 3780 // _retained_old_gc_alloc_region will become NULL. This is what we
3783 3781 // want either way so no reason to check explicitly for either
3784 3782 // condition.
3785 3783 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
3786 3784 }
3787 3785
3788 3786 void G1CollectedHeap::abandon_gc_alloc_regions() {
3789 3787 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
3790 3788 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
3791 3789 _retained_old_gc_alloc_region = NULL;
3792 3790 }
3793 3791
3794 3792 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
3795 3793 _drain_in_progress = false;
3796 3794 set_evac_failure_closure(cl);
3797 3795 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
3798 3796 }
3799 3797
3800 3798 void G1CollectedHeap::finalize_for_evac_failure() {
3801 3799 assert(_evac_failure_scan_stack != NULL &&
3802 3800 _evac_failure_scan_stack->length() == 0,
3803 3801 "Postcondition");
3804 3802 assert(!_drain_in_progress, "Postcondition");
3805 3803 delete _evac_failure_scan_stack;
3806 3804 _evac_failure_scan_stack = NULL;
3807 3805 }
3808 3806
3809 3807 class UpdateRSetDeferred : public OopsInHeapRegionClosure {
3810 3808 private:
3811 3809 G1CollectedHeap* _g1;
3812 3810 DirtyCardQueue *_dcq;
3813 3811 CardTableModRefBS* _ct_bs;
3814 3812
3815 3813 public:
3816 3814 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
3817 3815 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {}
3818 3816
3819 3817 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
3820 3818 virtual void do_oop( oop* p) { do_oop_work(p); }
3821 3819 template <class T> void do_oop_work(T* p) {
3822 3820 assert(_from->is_in_reserved(p), "paranoia");
3823 3821 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) &&
3824 3822 !_from->is_survivor()) {
3825 3823 size_t card_index = _ct_bs->index_for(p);
3826 3824 if (_ct_bs->mark_card_deferred(card_index)) {
3827 3825 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
3828 3826 }
3829 3827 }
3830 3828 }
3831 3829 };
3832 3830
3833 3831 class RemoveSelfPointerClosure: public ObjectClosure {
3834 3832 private:
3835 3833 G1CollectedHeap* _g1;
3836 3834 ConcurrentMark* _cm;
3837 3835 HeapRegion* _hr;
3838 3836 size_t _prev_marked_bytes;
3839 3837 size_t _next_marked_bytes;
3840 3838 OopsInHeapRegionClosure *_cl;
3841 3839 public:
3842 3840 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr,
3843 3841 OopsInHeapRegionClosure* cl) :
3844 3842 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0),
3845 3843 _next_marked_bytes(0), _cl(cl) {}
3846 3844
3847 3845 size_t prev_marked_bytes() { return _prev_marked_bytes; }
3848 3846 size_t next_marked_bytes() { return _next_marked_bytes; }
3849 3847
3850 3848 // <original comment>
3851 3849 // The original idea here was to coalesce evacuated and dead objects.
3852 3850 // However that caused complications with the block offset table (BOT).
3853 3851 // In particular if there were two TLABs, one of them partially refined.
3854 3852 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~|
3855 3853 // The BOT entries of the unrefined part of TLAB_2 point to the start
3856 3854 // of TLAB_2. If the last object of the TLAB_1 and the first object
3857 3855 // of TLAB_2 are coalesced, then the cards of the unrefined part
3858 3856 // would point into middle of the filler object.
3859 3857 // The current approach is to not coalesce and leave the BOT contents intact.
3860 3858 // </original comment>
3861 3859 //
3862 3860 // We now reset the BOT when we start the object iteration over the
3863 3861 // region and refine its entries for every object we come across. So
3864 3862 // the above comment is not really relevant and we should be able
3865 3863 // to coalesce dead objects if we want to.
3866 3864 void do_object(oop obj) {
3867 3865 HeapWord* obj_addr = (HeapWord*) obj;
3868 3866 assert(_hr->is_in(obj_addr), "sanity");
3869 3867 size_t obj_size = obj->size();
3870 3868 _hr->update_bot_for_object(obj_addr, obj_size);
3871 3869 if (obj->is_forwarded() && obj->forwardee() == obj) {
3872 3870 // The object failed to move.
3873 3871 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs.");
3874 3872 _cm->markPrev(obj);
3875 3873 assert(_cm->isPrevMarked(obj), "Should be marked!");
3876 3874 _prev_marked_bytes += (obj_size * HeapWordSize);
3877 3875 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) {
3878 3876 _cm->markAndGrayObjectIfNecessary(obj);
3879 3877 }
3880 3878 obj->set_mark(markOopDesc::prototype());
3881 3879 // While we were processing RSet buffers during the
3882 3880 // collection, we actually didn't scan any cards on the
3883 3881 // collection set, since we didn't want to update remebered
3884 3882 // sets with entries that point into the collection set, given
3885 3883 // that live objects fromthe collection set are about to move
3886 3884 // and such entries will be stale very soon. This change also
3887 3885 // dealt with a reliability issue which involved scanning a
3888 3886 // card in the collection set and coming across an array that
3889 3887 // was being chunked and looking malformed. The problem is
3890 3888 // that, if evacuation fails, we might have remembered set
3891 3889 // entries missing given that we skipped cards on the
3892 3890 // collection set. So, we'll recreate such entries now.
3893 3891 obj->oop_iterate(_cl);
3894 3892 assert(_cm->isPrevMarked(obj), "Should be marked!");
3895 3893 } else {
3896 3894 // The object has been either evacuated or is dead. Fill it with a
3897 3895 // dummy object.
3898 3896 MemRegion mr((HeapWord*)obj, obj_size);
3899 3897 CollectedHeap::fill_with_object(mr);
3900 3898 _cm->clearRangeBothMaps(mr);
3901 3899 }
3902 3900 }
3903 3901 };
3904 3902
3905 3903 void G1CollectedHeap::remove_self_forwarding_pointers() {
3906 3904 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set());
3907 3905 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set());
3908 3906 UpdateRSetDeferred deferred_update(_g1h, &dcq);
3909 3907 OopsInHeapRegionClosure *cl;
3910 3908 if (G1DeferredRSUpdate) {
3911 3909 cl = &deferred_update;
3912 3910 } else {
3913 3911 cl = &immediate_update;
3914 3912 }
3915 3913 HeapRegion* cur = g1_policy()->collection_set();
3916 3914 while (cur != NULL) {
3917 3915 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
3918 3916 assert(!cur->isHumongous(), "sanity");
3919 3917
3920 3918 if (cur->evacuation_failed()) {
3921 3919 assert(cur->in_collection_set(), "bad CS");
3922 3920 RemoveSelfPointerClosure rspc(_g1h, cur, cl);
3923 3921
3924 3922 // In the common case we make sure that this is done when the
3925 3923 // region is freed so that it is "ready-to-go" when it's
3926 3924 // re-allocated. However, when evacuation failure happens, a
3927 3925 // region will remain in the heap and might ultimately be added
3928 3926 // to a CSet in the future. So we have to be careful here and
3929 3927 // make sure the region's RSet is ready for parallel iteration
3930 3928 // whenever this might be required in the future.
3931 3929 cur->rem_set()->reset_for_par_iteration();
3932 3930 cur->reset_bot();
3933 3931 cl->set_region(cur);
3934 3932 cur->object_iterate(&rspc);
3935 3933
3936 3934 // A number of manipulations to make the TAMS be the current top,
3937 3935 // and the marked bytes be the ones observed in the iteration.
3938 3936 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) {
3939 3937 // The comments below are the postconditions achieved by the
3940 3938 // calls. Note especially the last such condition, which says that
3941 3939 // the count of marked bytes has been properly restored.
3942 3940 cur->note_start_of_marking(false);
3943 3941 // _next_top_at_mark_start == top, _next_marked_bytes == 0
3944 3942 cur->add_to_marked_bytes(rspc.prev_marked_bytes());
3945 3943 // _next_marked_bytes == prev_marked_bytes.
3946 3944 cur->note_end_of_marking();
3947 3945 // _prev_top_at_mark_start == top(),
3948 3946 // _prev_marked_bytes == prev_marked_bytes
3949 3947 }
3950 3948 // If there is no mark in progress, we modified the _next variables
3951 3949 // above needlessly, but harmlessly.
3952 3950 if (_g1h->mark_in_progress()) {
3953 3951 cur->note_start_of_marking(false);
3954 3952 // _next_top_at_mark_start == top, _next_marked_bytes == 0
3955 3953 // _next_marked_bytes == next_marked_bytes.
3956 3954 }
3957 3955
3958 3956 // Now make sure the region has the right index in the sorted array.
3959 3957 g1_policy()->note_change_in_marked_bytes(cur);
3960 3958 }
3961 3959 cur = cur->next_in_collection_set();
3962 3960 }
3963 3961 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
3964 3962
3965 3963 // Now restore saved marks, if any.
3966 3964 if (_objs_with_preserved_marks != NULL) {
3967 3965 assert(_preserved_marks_of_objs != NULL, "Both or none.");
3968 3966 guarantee(_objs_with_preserved_marks->length() ==
3969 3967 _preserved_marks_of_objs->length(), "Both or none.");
3970 3968 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
3971 3969 oop obj = _objs_with_preserved_marks->at(i);
3972 3970 markOop m = _preserved_marks_of_objs->at(i);
3973 3971 obj->set_mark(m);
3974 3972 }
3975 3973 // Delete the preserved marks growable arrays (allocated on the C heap).
3976 3974 delete _objs_with_preserved_marks;
3977 3975 delete _preserved_marks_of_objs;
3978 3976 _objs_with_preserved_marks = NULL;
3979 3977 _preserved_marks_of_objs = NULL;
3980 3978 }
3981 3979 }
3982 3980
3983 3981 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
3984 3982 _evac_failure_scan_stack->push(obj);
3985 3983 }
3986 3984
3987 3985 void G1CollectedHeap::drain_evac_failure_scan_stack() {
3988 3986 assert(_evac_failure_scan_stack != NULL, "precondition");
3989 3987
3990 3988 while (_evac_failure_scan_stack->length() > 0) {
3991 3989 oop obj = _evac_failure_scan_stack->pop();
3992 3990 _evac_failure_closure->set_region(heap_region_containing(obj));
3993 3991 obj->oop_iterate_backwards(_evac_failure_closure);
3994 3992 }
3995 3993 }
3996 3994
3997 3995 oop
3998 3996 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
3999 3997 oop old,
4000 3998 bool should_mark_root) {
4001 3999 assert(obj_in_cs(old),
4002 4000 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4003 4001 (HeapWord*) old));
4004 4002 markOop m = old->mark();
4005 4003 oop forward_ptr = old->forward_to_atomic(old);
4006 4004 if (forward_ptr == NULL) {
4007 4005 // Forward-to-self succeeded.
4008 4006
4009 4007 // should_mark_root will be true when this routine is called
4010 4008 // from a root scanning closure during an initial mark pause.
4011 4009 // In this case the thread that succeeds in self-forwarding the
4012 4010 // object is also responsible for marking the object.
4013 4011 if (should_mark_root) {
4014 4012 assert(!oopDesc::is_null(old), "shouldn't be");
4015 4013 _cm->grayRoot(old);
4016 4014 }
4017 4015
4018 4016 if (_evac_failure_closure != cl) {
4019 4017 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4020 4018 assert(!_drain_in_progress,
4021 4019 "Should only be true while someone holds the lock.");
4022 4020 // Set the global evac-failure closure to the current thread's.
4023 4021 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4024 4022 set_evac_failure_closure(cl);
4025 4023 // Now do the common part.
4026 4024 handle_evacuation_failure_common(old, m);
4027 4025 // Reset to NULL.
4028 4026 set_evac_failure_closure(NULL);
4029 4027 } else {
4030 4028 // The lock is already held, and this is recursive.
4031 4029 assert(_drain_in_progress, "This should only be the recursive case.");
4032 4030 handle_evacuation_failure_common(old, m);
4033 4031 }
4034 4032 return old;
4035 4033 } else {
4036 4034 // Forward-to-self failed. Either someone else managed to allocate
4037 4035 // space for this object (old != forward_ptr) or they beat us in
4038 4036 // self-forwarding it (old == forward_ptr).
4039 4037 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4040 4038 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4041 4039 "should not be in the CSet",
4042 4040 (HeapWord*) old, (HeapWord*) forward_ptr));
4043 4041 return forward_ptr;
4044 4042 }
4045 4043 }
4046 4044
4047 4045 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4048 4046 set_evacuation_failed(true);
4049 4047
4050 4048 preserve_mark_if_necessary(old, m);
4051 4049
4052 4050 HeapRegion* r = heap_region_containing(old);
4053 4051 if (!r->evacuation_failed()) {
4054 4052 r->set_evacuation_failed(true);
4055 4053 _hr_printer.evac_failure(r);
4056 4054 }
4057 4055
4058 4056 push_on_evac_failure_scan_stack(old);
4059 4057
4060 4058 if (!_drain_in_progress) {
4061 4059 // prevent recursion in copy_to_survivor_space()
4062 4060 _drain_in_progress = true;
4063 4061 drain_evac_failure_scan_stack();
4064 4062 _drain_in_progress = false;
4065 4063 }
4066 4064 }
4067 4065
4068 4066 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4069 4067 assert(evacuation_failed(), "Oversaving!");
4070 4068 // We want to call the "for_promotion_failure" version only in the
4071 4069 // case of a promotion failure.
4072 4070 if (m->must_be_preserved_for_promotion_failure(obj)) {
4073 4071 if (_objs_with_preserved_marks == NULL) {
4074 4072 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4075 4073 _objs_with_preserved_marks =
4076 4074 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
4077 4075 _preserved_marks_of_objs =
4078 4076 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
4079 4077 }
4080 4078 _objs_with_preserved_marks->push(obj);
4081 4079 _preserved_marks_of_objs->push(m);
4082 4080 }
4083 4081 }
4084 4082
4085 4083 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4086 4084 size_t word_size) {
4087 4085 if (purpose == GCAllocForSurvived) {
4088 4086 HeapWord* result = survivor_attempt_allocation(word_size);
4089 4087 if (result != NULL) {
4090 4088 return result;
4091 4089 } else {
4092 4090 // Let's try to allocate in the old gen in case we can fit the
4093 4091 // object there.
4094 4092 return old_attempt_allocation(word_size);
4095 4093 }
4096 4094 } else {
4097 4095 assert(purpose == GCAllocForTenured, "sanity");
4098 4096 HeapWord* result = old_attempt_allocation(word_size);
4099 4097 if (result != NULL) {
4100 4098 return result;
4101 4099 } else {
4102 4100 // Let's try to allocate in the survivors in case we can fit the
4103 4101 // object there.
4104 4102 return survivor_attempt_allocation(word_size);
4105 4103 }
4106 4104 }
4107 4105
4108 4106 ShouldNotReachHere();
4109 4107 // Trying to keep some compilers happy.
4110 4108 return NULL;
4111 4109 }
4112 4110
4113 4111 #ifndef PRODUCT
4114 4112 bool GCLabBitMapClosure::do_bit(size_t offset) {
4115 4113 HeapWord* addr = _bitmap->offsetToHeapWord(offset);
4116 4114 guarantee(_cm->isMarked(oop(addr)), "it should be!");
4117 4115 return true;
4118 4116 }
4119 4117 #endif // PRODUCT
4120 4118
4121 4119 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4122 4120 ParGCAllocBuffer(gclab_word_size),
4123 4121 _should_mark_objects(false),
4124 4122 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size),
4125 4123 _retired(false)
4126 4124 {
4127 4125 //_should_mark_objects is set to true when G1ParCopyHelper needs to
4128 4126 // mark the forwarded location of an evacuated object.
4129 4127 // We set _should_mark_objects to true if marking is active, i.e. when we
4130 4128 // need to propagate a mark, or during an initial mark pause, i.e. when we
4131 4129 // need to mark objects immediately reachable by the roots.
4132 4130 if (G1CollectedHeap::heap()->mark_in_progress() ||
4133 4131 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) {
4134 4132 _should_mark_objects = true;
4135 4133 }
4136 4134 }
4137 4135
4138 4136 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num)
4139 4137 : _g1h(g1h),
4140 4138 _refs(g1h->task_queue(queue_num)),
4141 4139 _dcq(&g1h->dirty_card_queue_set()),
4142 4140 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4143 4141 _g1_rem(g1h->g1_rem_set()),
4144 4142 _hash_seed(17), _queue_num(queue_num),
4145 4143 _term_attempts(0),
4146 4144 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4147 4145 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4148 4146 _age_table(false),
4149 4147 _strong_roots_time(0), _term_time(0),
4150 4148 _alloc_buffer_waste(0), _undo_waste(0)
4151 4149 {
4152 4150 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4153 4151 // we "sacrifice" entry 0 to keep track of surviving bytes for
4154 4152 // non-young regions (where the age is -1)
4155 4153 // We also add a few elements at the beginning and at the end in
4156 4154 // an attempt to eliminate cache contention
4157 4155 size_t real_length = 1 + _g1h->g1_policy()->young_cset_length();
4158 4156 size_t array_length = PADDING_ELEM_NUM +
4159 4157 real_length +
4160 4158 PADDING_ELEM_NUM;
4161 4159 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length);
4162 4160 if (_surviving_young_words_base == NULL)
4163 4161 vm_exit_out_of_memory(array_length * sizeof(size_t),
4164 4162 "Not enough space for young surv histo.");
4165 4163 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4166 4164 memset(_surviving_young_words, 0, real_length * sizeof(size_t));
4167 4165
4168 4166 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4169 4167 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4170 4168
4171 4169 _start = os::elapsedTime();
4172 4170 }
4173 4171
4174 4172 void
4175 4173 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4176 4174 {
4177 4175 st->print_raw_cr("GC Termination Stats");
4178 4176 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4179 4177 " ------waste (KiB)------");
4180 4178 st->print_raw_cr("thr ms ms % ms % attempts"
4181 4179 " total alloc undo");
4182 4180 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4183 4181 " ------- ------- -------");
4184 4182 }
4185 4183
4186 4184 void
4187 4185 G1ParScanThreadState::print_termination_stats(int i,
4188 4186 outputStream* const st) const
4189 4187 {
4190 4188 const double elapsed_ms = elapsed_time() * 1000.0;
4191 4189 const double s_roots_ms = strong_roots_time() * 1000.0;
4192 4190 const double term_ms = term_time() * 1000.0;
4193 4191 st->print_cr("%3d %9.2f %9.2f %6.2f "
4194 4192 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4195 4193 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4196 4194 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4197 4195 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4198 4196 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4199 4197 alloc_buffer_waste() * HeapWordSize / K,
4200 4198 undo_waste() * HeapWordSize / K);
4201 4199 }
4202 4200
4203 4201 #ifdef ASSERT
4204 4202 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4205 4203 assert(ref != NULL, "invariant");
4206 4204 assert(UseCompressedOops, "sanity");
4207 4205 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4208 4206 oop p = oopDesc::load_decode_heap_oop(ref);
4209 4207 assert(_g1h->is_in_g1_reserved(p),
4210 4208 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4211 4209 return true;
4212 4210 }
4213 4211
4214 4212 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4215 4213 assert(ref != NULL, "invariant");
4216 4214 if (has_partial_array_mask(ref)) {
4217 4215 // Must be in the collection set--it's already been copied.
4218 4216 oop p = clear_partial_array_mask(ref);
4219 4217 assert(_g1h->obj_in_cs(p),
4220 4218 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4221 4219 } else {
4222 4220 oop p = oopDesc::load_decode_heap_oop(ref);
4223 4221 assert(_g1h->is_in_g1_reserved(p),
4224 4222 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4225 4223 }
4226 4224 return true;
4227 4225 }
4228 4226
4229 4227 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4230 4228 if (ref.is_narrow()) {
4231 4229 return verify_ref((narrowOop*) ref);
4232 4230 } else {
4233 4231 return verify_ref((oop*) ref);
4234 4232 }
4235 4233 }
4236 4234 #endif // ASSERT
4237 4235
4238 4236 void G1ParScanThreadState::trim_queue() {
4239 4237 assert(_evac_cl != NULL, "not set");
4240 4238 assert(_evac_failure_cl != NULL, "not set");
4241 4239 assert(_partial_scan_cl != NULL, "not set");
4242 4240
4243 4241 StarTask ref;
4244 4242 do {
4245 4243 // Drain the overflow stack first, so other threads can steal.
4246 4244 while (refs()->pop_overflow(ref)) {
4247 4245 deal_with_reference(ref);
4248 4246 }
4249 4247
4250 4248 while (refs()->pop_local(ref)) {
4251 4249 deal_with_reference(ref);
4252 4250 }
4253 4251 } while (!refs()->is_empty());
4254 4252 }
4255 4253
4256 4254 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) :
4257 4255 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4258 4256 _par_scan_state(par_scan_state),
4259 4257 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4260 4258 _mark_in_progress(_g1->mark_in_progress()) { }
4261 4259
4262 4260 template <class T> void G1ParCopyHelper::mark_object(T* p) {
4263 4261 // This is called from do_oop_work for objects that are not
4264 4262 // in the collection set. Objects in the collection set
4265 4263 // are marked after they have been evacuated.
4266 4264
4267 4265 T heap_oop = oopDesc::load_heap_oop(p);
4268 4266 if (!oopDesc::is_null(heap_oop)) {
4269 4267 oop obj = oopDesc::decode_heap_oop(heap_oop);
4270 4268 HeapWord* addr = (HeapWord*)obj;
4271 4269 if (_g1->is_in_g1_reserved(addr)) {
4272 4270 _cm->grayRoot(oop(addr));
4273 4271 }
4274 4272 }
4275 4273 }
4276 4274
4277 4275 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root,
4278 4276 bool should_mark_copy) {
4279 4277 size_t word_sz = old->size();
4280 4278 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4281 4279 // +1 to make the -1 indexes valid...
4282 4280 int young_index = from_region->young_index_in_cset()+1;
4283 4281 assert( (from_region->is_young() && young_index > 0) ||
4284 4282 (!from_region->is_young() && young_index == 0), "invariant" );
4285 4283 G1CollectorPolicy* g1p = _g1->g1_policy();
4286 4284 markOop m = old->mark();
4287 4285 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4288 4286 : m->age();
4289 4287 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4290 4288 word_sz);
4291 4289 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4292 4290 oop obj = oop(obj_ptr);
4293 4291
4294 4292 if (obj_ptr == NULL) {
4295 4293 // This will either forward-to-self, or detect that someone else has
4296 4294 // installed a forwarding pointer.
4297 4295 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4298 4296 return _g1->handle_evacuation_failure_par(cl, old, should_mark_root);
4299 4297 }
4300 4298
4301 4299 // We're going to allocate linearly, so might as well prefetch ahead.
4302 4300 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4303 4301
4304 4302 oop forward_ptr = old->forward_to_atomic(obj);
4305 4303 if (forward_ptr == NULL) {
4306 4304 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4307 4305 if (g1p->track_object_age(alloc_purpose)) {
4308 4306 // We could simply do obj->incr_age(). However, this causes a
4309 4307 // performance issue. obj->incr_age() will first check whether
4310 4308 // the object has a displaced mark by checking its mark word;
4311 4309 // getting the mark word from the new location of the object
4312 4310 // stalls. So, given that we already have the mark word and we
4313 4311 // are about to install it anyway, it's better to increase the
4314 4312 // age on the mark word, when the object does not have a
4315 4313 // displaced mark word. We're not expecting many objects to have
4316 4314 // a displaced marked word, so that case is not optimized
4317 4315 // further (it could be...) and we simply call obj->incr_age().
4318 4316
4319 4317 if (m->has_displaced_mark_helper()) {
4320 4318 // in this case, we have to install the mark word first,
4321 4319 // otherwise obj looks to be forwarded (the old mark word,
4322 4320 // which contains the forward pointer, was copied)
4323 4321 obj->set_mark(m);
4324 4322 obj->incr_age();
4325 4323 } else {
4326 4324 m = m->incr_age();
4327 4325 obj->set_mark(m);
4328 4326 }
4329 4327 _par_scan_state->age_table()->add(obj, word_sz);
4330 4328 } else {
4331 4329 obj->set_mark(m);
4332 4330 }
4333 4331
4334 4332 // Mark the evacuated object or propagate "next" mark bit
4335 4333 if (should_mark_copy) {
4336 4334 if (!use_local_bitmaps ||
4337 4335 !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) {
4338 4336 // if we couldn't mark it on the local bitmap (this happens when
4339 4337 // the object was not allocated in the GCLab), we have to bite
4340 4338 // the bullet and do the standard parallel mark
4341 4339 _cm->markAndGrayObjectIfNecessary(obj);
4342 4340 }
4343 4341
4344 4342 if (_g1->isMarkedNext(old)) {
4345 4343 // Unmark the object's old location so that marking
4346 4344 // doesn't think the old object is alive.
4347 4345 _cm->nextMarkBitMap()->parClear((HeapWord*)old);
4348 4346 }
4349 4347 }
4350 4348
4351 4349 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4352 4350 surv_young_words[young_index] += word_sz;
4353 4351
4354 4352 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4355 4353 arrayOop(old)->set_length(0);
4356 4354 oop* old_p = set_partial_array_mask(old);
4357 4355 _par_scan_state->push_on_queue(old_p);
4358 4356 } else {
4359 4357 // No point in using the slower heap_region_containing() method,
4360 4358 // given that we know obj is in the heap.
4361 4359 _scanner->set_region(_g1->heap_region_containing_raw(obj));
4362 4360 obj->oop_iterate_backwards(_scanner);
4363 4361 }
4364 4362 } else {
4365 4363 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4366 4364 obj = forward_ptr;
4367 4365 }
4368 4366 return obj;
4369 4367 }
4370 4368
4371 4369 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4372 4370 template <class T>
4373 4371 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4374 4372 ::do_oop_work(T* p) {
4375 4373 oop obj = oopDesc::load_decode_heap_oop(p);
4376 4374 assert(barrier != G1BarrierRS || obj != NULL,
4377 4375 "Precondition: G1BarrierRS implies obj is nonNull");
4378 4376
4379 4377 // Marking:
4380 4378 // If the object is in the collection set, then the thread
4381 4379 // that copies the object should mark, or propagate the
4382 4380 // mark to, the evacuated object.
4383 4381 // If the object is not in the collection set then we
4384 4382 // should call the mark_object() method depending on the
4385 4383 // value of the template parameter do_mark_object (which will
4386 4384 // be true for root scanning closures during an initial mark
4387 4385 // pause).
4388 4386 // The mark_object() method first checks whether the object
4389 4387 // is marked and, if not, attempts to mark the object.
4390 4388
4391 4389 // here the null check is implicit in the cset_fast_test() test
4392 4390 if (_g1->in_cset_fast_test(obj)) {
4393 4391 if (obj->is_forwarded()) {
4394 4392 oopDesc::encode_store_heap_oop(p, obj->forwardee());
4395 4393 // If we are a root scanning closure during an initial
4396 4394 // mark pause (i.e. do_mark_object will be true) then
4397 4395 // we also need to handle marking of roots in the
4398 4396 // event of an evacuation failure. In the event of an
4399 4397 // evacuation failure, the object is forwarded to itself
4400 4398 // and not copied. For root-scanning closures, the
4401 4399 // object would be marked after a successful self-forward
4402 4400 // but an object could be pointed to by both a root and non
4403 4401 // root location and be self-forwarded by a non-root-scanning
4404 4402 // closure. Therefore we also have to attempt to mark the
4405 4403 // self-forwarded root object here.
4406 4404 if (do_mark_object && obj->forwardee() == obj) {
4407 4405 mark_object(p);
4408 4406 }
4409 4407 } else {
4410 4408 // During an initial mark pause, objects that are pointed to
4411 4409 // by the roots need to be marked - even in the event of an
4412 4410 // evacuation failure. We pass the template parameter
4413 4411 // do_mark_object (which is true for root scanning closures
4414 4412 // during an initial mark pause) to copy_to_survivor_space
4415 4413 // which will pass it on to the evacuation failure handling
4416 4414 // code. The thread that successfully self-forwards a root
4417 4415 // object to itself is responsible for marking the object.
4418 4416 bool should_mark_root = do_mark_object;
4419 4417
4420 4418 // We need to mark the copied object if we're a root scanning
4421 4419 // closure during an initial mark pause (i.e. do_mark_object
4422 4420 // will be true), or the object is already marked and we need
4423 4421 // to propagate the mark to the evacuated copy.
4424 4422 bool should_mark_copy = do_mark_object ||
4425 4423 _during_initial_mark ||
4426 4424 (_mark_in_progress && !_g1->is_obj_ill(obj));
4427 4425
4428 4426 oop copy_oop = copy_to_survivor_space(obj, should_mark_root,
4429 4427 should_mark_copy);
4430 4428 oopDesc::encode_store_heap_oop(p, copy_oop);
4431 4429 }
4432 4430 // When scanning the RS, we only care about objs in CS.
4433 4431 if (barrier == G1BarrierRS) {
4434 4432 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4435 4433 }
4436 4434 } else {
4437 4435 // The object is not in collection set. If we're a root scanning
4438 4436 // closure during an initial mark pause (i.e. do_mark_object will
4439 4437 // be true) then attempt to mark the object.
4440 4438 if (do_mark_object) {
4441 4439 mark_object(p);
4442 4440 }
4443 4441 }
4444 4442
4445 4443 if (barrier == G1BarrierEvac && obj != NULL) {
4446 4444 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4447 4445 }
4448 4446
4449 4447 if (do_gen_barrier && obj != NULL) {
4450 4448 par_do_barrier(p);
4451 4449 }
4452 4450 }
4453 4451
4454 4452 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4455 4453 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4456 4454
4457 4455 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4458 4456 assert(has_partial_array_mask(p), "invariant");
4459 4457 oop old = clear_partial_array_mask(p);
4460 4458 assert(old->is_objArray(), "must be obj array");
4461 4459 assert(old->is_forwarded(), "must be forwarded");
4462 4460 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
4463 4461
4464 4462 objArrayOop obj = objArrayOop(old->forwardee());
4465 4463 assert((void*)old != (void*)old->forwardee(), "self forwarding here?");
4466 4464 // Process ParGCArrayScanChunk elements now
4467 4465 // and push the remainder back onto queue
4468 4466 int start = arrayOop(old)->length();
4469 4467 int end = obj->length();
4470 4468 int remainder = end - start;
4471 4469 assert(start <= end, "just checking");
4472 4470 if (remainder > 2 * ParGCArrayScanChunk) {
4473 4471 // Test above combines last partial chunk with a full chunk
4474 4472 end = start + ParGCArrayScanChunk;
4475 4473 arrayOop(old)->set_length(end);
4476 4474 // Push remainder.
4477 4475 oop* old_p = set_partial_array_mask(old);
4478 4476 assert(arrayOop(old)->length() < obj->length(), "Empty push?");
4479 4477 _par_scan_state->push_on_queue(old_p);
4480 4478 } else {
4481 4479 // Restore length so that the heap remains parsable in
4482 4480 // case of evacuation failure.
4483 4481 arrayOop(old)->set_length(end);
4484 4482 }
4485 4483 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4486 4484 // process our set of indices (include header in first chunk)
4487 4485 obj->oop_iterate_range(&_scanner, start, end);
4488 4486 }
4489 4487
4490 4488 class G1ParEvacuateFollowersClosure : public VoidClosure {
4491 4489 protected:
4492 4490 G1CollectedHeap* _g1h;
4493 4491 G1ParScanThreadState* _par_scan_state;
4494 4492 RefToScanQueueSet* _queues;
4495 4493 ParallelTaskTerminator* _terminator;
4496 4494
4497 4495 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4498 4496 RefToScanQueueSet* queues() { return _queues; }
4499 4497 ParallelTaskTerminator* terminator() { return _terminator; }
4500 4498
4501 4499 public:
4502 4500 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4503 4501 G1ParScanThreadState* par_scan_state,
4504 4502 RefToScanQueueSet* queues,
4505 4503 ParallelTaskTerminator* terminator)
4506 4504 : _g1h(g1h), _par_scan_state(par_scan_state),
4507 4505 _queues(queues), _terminator(terminator) {}
4508 4506
4509 4507 void do_void();
4510 4508
4511 4509 private:
4512 4510 inline bool offer_termination();
4513 4511 };
4514 4512
4515 4513 bool G1ParEvacuateFollowersClosure::offer_termination() {
4516 4514 G1ParScanThreadState* const pss = par_scan_state();
4517 4515 pss->start_term_time();
4518 4516 const bool res = terminator()->offer_termination();
4519 4517 pss->end_term_time();
4520 4518 return res;
4521 4519 }
4522 4520
4523 4521 void G1ParEvacuateFollowersClosure::do_void() {
4524 4522 StarTask stolen_task;
4525 4523 G1ParScanThreadState* const pss = par_scan_state();
4526 4524 pss->trim_queue();
4527 4525
4528 4526 do {
4529 4527 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4530 4528 assert(pss->verify_task(stolen_task), "sanity");
4531 4529 if (stolen_task.is_narrow()) {
4532 4530 pss->deal_with_reference((narrowOop*) stolen_task);
4533 4531 } else {
4534 4532 pss->deal_with_reference((oop*) stolen_task);
4535 4533 }
4536 4534
4537 4535 // We've just processed a reference and we might have made
4538 4536 // available new entries on the queues. So we have to make sure
4539 4537 // we drain the queues as necessary.
4540 4538 pss->trim_queue();
4541 4539 }
4542 4540 } while (!offer_termination());
4543 4541
4544 4542 pss->retire_alloc_buffers();
4545 4543 }
4546 4544
4547 4545 class G1ParTask : public AbstractGangTask {
4548 4546 protected:
4549 4547 G1CollectedHeap* _g1h;
4550 4548 RefToScanQueueSet *_queues;
4551 4549 ParallelTaskTerminator _terminator;
4552 4550 int _n_workers;
4553 4551
4554 4552 Mutex _stats_lock;
4555 4553 Mutex* stats_lock() { return &_stats_lock; }
4556 4554
4557 4555 size_t getNCards() {
4558 4556 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4559 4557 / G1BlockOffsetSharedArray::N_bytes;
4560 4558 }
4561 4559
4562 4560 public:
4563 4561 G1ParTask(G1CollectedHeap* g1h, int workers, RefToScanQueueSet *task_queues)
4564 4562 : AbstractGangTask("G1 collection"),
4565 4563 _g1h(g1h),
4566 4564 _queues(task_queues),
4567 4565 _terminator(workers, _queues),
4568 4566 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true),
4569 4567 _n_workers(workers)
4570 4568 {}
4571 4569
4572 4570 RefToScanQueueSet* queues() { return _queues; }
4573 4571
4574 4572 RefToScanQueue *work_queue(int i) {
4575 4573 return queues()->queue(i);
4576 4574 }
4577 4575
4578 4576 void work(int i) {
4579 4577 if (i >= _n_workers) return; // no work needed this round
4580 4578
4581 4579 double start_time_ms = os::elapsedTime() * 1000.0;
4582 4580 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms);
4583 4581
4584 4582 ResourceMark rm;
4585 4583 HandleMark hm;
4586 4584
4587 4585 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4588 4586
4589 4587 G1ParScanThreadState pss(_g1h, i);
4590 4588 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4591 4589 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4592 4590 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4593 4591
4594 4592 pss.set_evac_closure(&scan_evac_cl);
4595 4593 pss.set_evac_failure_closure(&evac_failure_cl);
4596 4594 pss.set_partial_scan_closure(&partial_scan_cl);
4597 4595
4598 4596 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4599 4597 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp);
4600 4598
4601 4599 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4602 4600 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp);
4603 4601
4604 4602 OopClosure* scan_root_cl = &only_scan_root_cl;
4605 4603 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl;
4606 4604
4607 4605 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4608 4606 // We also need to mark copied objects.
4609 4607 scan_root_cl = &scan_mark_root_cl;
4610 4608 scan_perm_cl = &scan_mark_perm_cl;
4611 4609 }
4612 4610
4613 4611 // The following closure is used to scan RSets looking for reference
4614 4612 // fields that point into the collection set. The actual field iteration
4615 4613 // is performed by a FilterIntoCSClosure, whose do_oop method calls the
4616 4614 // do_oop method of the following closure.
4617 4615 // Therefore we want to record the reference processor in the
4618 4616 // FilterIntoCSClosure. To do so we record the STW reference
4619 4617 // processor into the following closure and pass it to the
4620 4618 // FilterIntoCSClosure in HeapRegionDCTOC::walk_mem_region_with_cl.
4621 4619 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss, rp);
4622 4620
4623 4621 pss.start_strong_roots();
4624 4622 _g1h->g1_process_strong_roots(/* not collecting perm */ false,
4625 4623 SharedHeap::SO_AllClasses,
4626 4624 scan_root_cl,
4627 4625 &push_heap_rs_cl,
4628 4626 scan_perm_cl,
4629 4627 i);
4630 4628 pss.end_strong_roots();
4631 4629
4632 4630 {
4633 4631 double start = os::elapsedTime();
4634 4632 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4635 4633 evac.do_void();
4636 4634 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4637 4635 double term_ms = pss.term_time()*1000.0;
4638 4636 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms);
4639 4637 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts());
4640 4638 }
4641 4639 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4642 4640 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4643 4641
4644 4642 // Clean up any par-expanded rem sets.
4645 4643 HeapRegionRemSet::par_cleanup();
4646 4644
4647 4645 if (ParallelGCVerbose) {
4648 4646 MutexLocker x(stats_lock());
4649 4647 pss.print_termination_stats(i);
4650 4648 }
4651 4649
4652 4650 assert(pss.refs()->is_empty(), "should be empty");
4653 4651 double end_time_ms = os::elapsedTime() * 1000.0;
4654 4652 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms);
4655 4653 }
4656 4654 };
4657 4655
4658 4656 // *** Common G1 Evacuation Stuff
4659 4657
4660 4658 // This method is run in a GC worker.
4661 4659
4662 4660 void
4663 4661 G1CollectedHeap::
4664 4662 g1_process_strong_roots(bool collecting_perm_gen,
4665 4663 SharedHeap::ScanningOption so,
4666 4664 OopClosure* scan_non_heap_roots,
4667 4665 OopsInHeapRegionClosure* scan_rs,
4668 4666 OopsInGenClosure* scan_perm,
4669 4667 int worker_i) {
4670 4668
4671 4669 // First scan the strong roots, including the perm gen.
4672 4670 double ext_roots_start = os::elapsedTime();
4673 4671 double closure_app_time_sec = 0.0;
4674 4672
4675 4673 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4676 4674 BufferingOopsInGenClosure buf_scan_perm(scan_perm);
4677 4675 buf_scan_perm.set_generation(perm_gen());
4678 4676
4679 4677 // Walk the code cache w/o buffering, because StarTask cannot handle
4680 4678 // unaligned oop locations.
4681 4679 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true);
4682 4680
4683 4681 process_strong_roots(false, // no scoping; this is parallel code
4684 4682 collecting_perm_gen, so,
4685 4683 &buf_scan_non_heap_roots,
4686 4684 &eager_scan_code_roots,
4687 4685 &buf_scan_perm);
4688 4686
4689 4687 // Now the CM ref_processor roots.
4690 4688 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
4691 4689 // We need to treat the discovered reference lists of the
4692 4690 // concurrent mark ref processor as roots and keep entries
4693 4691 // (which are added by the marking threads) on them live
4694 4692 // until they can be processed at the end of marking.
4695 4693 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
4696 4694 }
4697 4695
4698 4696 // Finish up any enqueued closure apps (attributed as object copy time).
4699 4697 buf_scan_non_heap_roots.done();
4700 4698 buf_scan_perm.done();
4701 4699
4702 4700 double ext_roots_end = os::elapsedTime();
4703 4701
4704 4702 g1_policy()->reset_obj_copy_time(worker_i);
4705 4703 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() +
4706 4704 buf_scan_non_heap_roots.closure_app_seconds();
4707 4705 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
4708 4706
4709 4707 double ext_root_time_ms =
4710 4708 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0;
4711 4709
4712 4710 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
4713 4711
4714 4712 // Scan strong roots in mark stack.
4715 4713 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) {
4716 4714 concurrent_mark()->oops_do(scan_non_heap_roots);
4717 4715 }
4718 4716 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0;
4719 4717 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms);
4720 4718
4721 4719 // Now scan the complement of the collection set.
4722 4720 if (scan_rs != NULL) {
4723 4721 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
4724 4722 }
4725 4723
4726 4724 _process_strong_tasks->all_tasks_completed();
4727 4725 }
4728 4726
4729 4727 void
4730 4728 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
4731 4729 OopClosure* non_root_closure) {
4732 4730 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
4733 4731 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
4734 4732 }
4735 4733
4736 4734 // Weak Reference Processing support
4737 4735
4738 4736 // An always "is_alive" closure that is used to preserve referents.
4739 4737 // If the object is non-null then it's alive. Used in the preservation
4740 4738 // of referent objects that are pointed to by reference objects
4741 4739 // discovered by the CM ref processor.
4742 4740 class G1AlwaysAliveClosure: public BoolObjectClosure {
4743 4741 G1CollectedHeap* _g1;
4744 4742 public:
4745 4743 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4746 4744 void do_object(oop p) { assert(false, "Do not call."); }
4747 4745 bool do_object_b(oop p) {
4748 4746 if (p != NULL) {
4749 4747 return true;
4750 4748 }
4751 4749 return false;
4752 4750 }
4753 4751 };
4754 4752
4755 4753 bool G1STWIsAliveClosure::do_object_b(oop p) {
4756 4754 // An object is reachable if it is outside the collection set,
4757 4755 // or is inside and copied.
4758 4756 return !_g1->obj_in_cs(p) || p->is_forwarded();
4759 4757 }
4760 4758
4761 4759 // Non Copying Keep Alive closure
4762 4760 class G1KeepAliveClosure: public OopClosure {
4763 4761 G1CollectedHeap* _g1;
4764 4762 public:
4765 4763 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4766 4764 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
4767 4765 void do_oop( oop* p) {
4768 4766 oop obj = *p;
4769 4767
4770 4768 if (_g1->obj_in_cs(obj)) {
4771 4769 assert( obj->is_forwarded(), "invariant" );
4772 4770 *p = obj->forwardee();
4773 4771 }
4774 4772 }
4775 4773 };
4776 4774
4777 4775 // Copying Keep Alive closure - can be called from both
4778 4776 // serial and parallel code as long as different worker
4779 4777 // threads utilize different G1ParScanThreadState instances
4780 4778 // and different queues.
4781 4779
4782 4780 class G1CopyingKeepAliveClosure: public OopClosure {
4783 4781 G1CollectedHeap* _g1h;
4784 4782 OopClosure* _copy_non_heap_obj_cl;
4785 4783 OopsInHeapRegionClosure* _copy_perm_obj_cl;
4786 4784 G1ParScanThreadState* _par_scan_state;
4787 4785
4788 4786 public:
4789 4787 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
4790 4788 OopClosure* non_heap_obj_cl,
4791 4789 OopsInHeapRegionClosure* perm_obj_cl,
4792 4790 G1ParScanThreadState* pss):
4793 4791 _g1h(g1h),
4794 4792 _copy_non_heap_obj_cl(non_heap_obj_cl),
4795 4793 _copy_perm_obj_cl(perm_obj_cl),
4796 4794 _par_scan_state(pss)
4797 4795 {}
4798 4796
4799 4797 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
4800 4798 virtual void do_oop( oop* p) { do_oop_work(p); }
4801 4799
4802 4800 template <class T> void do_oop_work(T* p) {
4803 4801 oop obj = oopDesc::load_decode_heap_oop(p);
4804 4802
4805 4803 if (_g1h->obj_in_cs(obj)) {
4806 4804 // If the referent object has been forwarded (either copied
4807 4805 // to a new location or to itself in the event of an
4808 4806 // evacuation failure) then we need to update the reference
4809 4807 // field and, if both reference and referent are in the G1
4810 4808 // heap, update the RSet for the referent.
4811 4809 //
4812 4810 // If the referent has not been forwarded then we have to keep
4813 4811 // it alive by policy. Therefore we have copy the referent.
4814 4812 //
4815 4813 // If the reference field is in the G1 heap then we can push
4816 4814 // on the PSS queue. When the queue is drained (after each
4817 4815 // phase of reference processing) the object and it's followers
4818 4816 // will be copied, the reference field set to point to the
4819 4817 // new location, and the RSet updated. Otherwise we need to
4820 4818 // use the the non-heap or perm closures directly to copy
4821 4819 // the refernt object and update the pointer, while avoiding
4822 4820 // updating the RSet.
4823 4821
4824 4822 if (_g1h->is_in_g1_reserved(p)) {
4825 4823 _par_scan_state->push_on_queue(p);
4826 4824 } else {
4827 4825 // The reference field is not in the G1 heap.
4828 4826 if (_g1h->perm_gen()->is_in(p)) {
4829 4827 _copy_perm_obj_cl->do_oop(p);
4830 4828 } else {
4831 4829 _copy_non_heap_obj_cl->do_oop(p);
4832 4830 }
4833 4831 }
4834 4832 }
4835 4833 }
4836 4834 };
4837 4835
4838 4836 // Serial drain queue closure. Called as the 'complete_gc'
4839 4837 // closure for each discovered list in some of the
4840 4838 // reference processing phases.
4841 4839
4842 4840 class G1STWDrainQueueClosure: public VoidClosure {
4843 4841 protected:
4844 4842 G1CollectedHeap* _g1h;
4845 4843 G1ParScanThreadState* _par_scan_state;
4846 4844
4847 4845 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4848 4846
4849 4847 public:
4850 4848 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
4851 4849 _g1h(g1h),
4852 4850 _par_scan_state(pss)
4853 4851 { }
4854 4852
4855 4853 void do_void() {
4856 4854 G1ParScanThreadState* const pss = par_scan_state();
4857 4855 pss->trim_queue();
4858 4856 }
4859 4857 };
4860 4858
4861 4859 // Parallel Reference Processing closures
4862 4860
4863 4861 // Implementation of AbstractRefProcTaskExecutor for parallel reference
4864 4862 // processing during G1 evacuation pauses.
4865 4863
4866 4864 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
4867 4865 private:
4868 4866 G1CollectedHeap* _g1h;
4869 4867 RefToScanQueueSet* _queues;
4870 4868 WorkGang* _workers;
4871 4869 int _active_workers;
4872 4870
4873 4871 public:
4874 4872 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
4875 4873 WorkGang* workers,
4876 4874 RefToScanQueueSet *task_queues,
4877 4875 int n_workers) :
4878 4876 _g1h(g1h),
4879 4877 _queues(task_queues),
4880 4878 _workers(workers),
4881 4879 _active_workers(n_workers)
4882 4880 {
4883 4881 assert(n_workers > 0, "shouldn't call this otherwise");
4884 4882 }
4885 4883
4886 4884 // Executes the given task using concurrent marking worker threads.
4887 4885 virtual void execute(ProcessTask& task);
4888 4886 virtual void execute(EnqueueTask& task);
4889 4887 };
4890 4888
4891 4889 // Gang task for possibly parallel reference processing
4892 4890
4893 4891 class G1STWRefProcTaskProxy: public AbstractGangTask {
4894 4892 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
4895 4893 ProcessTask& _proc_task;
4896 4894 G1CollectedHeap* _g1h;
4897 4895 RefToScanQueueSet *_task_queues;
4898 4896 ParallelTaskTerminator* _terminator;
4899 4897
4900 4898 public:
4901 4899 G1STWRefProcTaskProxy(ProcessTask& proc_task,
4902 4900 G1CollectedHeap* g1h,
4903 4901 RefToScanQueueSet *task_queues,
4904 4902 ParallelTaskTerminator* terminator) :
4905 4903 AbstractGangTask("Process reference objects in parallel"),
4906 4904 _proc_task(proc_task),
4907 4905 _g1h(g1h),
4908 4906 _task_queues(task_queues),
4909 4907 _terminator(terminator)
4910 4908 {}
4911 4909
4912 4910 virtual void work(int i) {
4913 4911 // The reference processing task executed by a single worker.
4914 4912 ResourceMark rm;
4915 4913 HandleMark hm;
4916 4914
4917 4915 G1STWIsAliveClosure is_alive(_g1h);
4918 4916
4919 4917 G1ParScanThreadState pss(_g1h, i);
4920 4918
4921 4919 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
4922 4920 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
4923 4921 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
4924 4922
4925 4923 pss.set_evac_closure(&scan_evac_cl);
4926 4924 pss.set_evac_failure_closure(&evac_failure_cl);
4927 4925 pss.set_partial_scan_closure(&partial_scan_cl);
4928 4926
4929 4927 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
4930 4928 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
4931 4929
4932 4930 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
4933 4931 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
4934 4932
4935 4933 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
4936 4934 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
4937 4935
4938 4936 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4939 4937 // We also need to mark copied objects.
4940 4938 copy_non_heap_cl = ©_mark_non_heap_cl;
4941 4939 copy_perm_cl = ©_mark_perm_cl;
4942 4940 }
4943 4941
4944 4942 // Keep alive closure.
4945 4943 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
4946 4944
4947 4945 // Complete GC closure
4948 4946 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
4949 4947
4950 4948 // Call the reference processing task's work routine.
4951 4949 _proc_task.work(i, is_alive, keep_alive, drain_queue);
4952 4950
4953 4951 // Note we cannot assert that the refs array is empty here as not all
4954 4952 // of the processing tasks (specifically phase2 - pp2_work) execute
4955 4953 // the complete_gc closure (which ordinarily would drain the queue) so
4956 4954 // the queue may not be empty.
4957 4955 }
4958 4956 };
4959 4957
4960 4958 // Driver routine for parallel reference processing.
4961 4959 // Creates an instance of the ref processing gang
4962 4960 // task and has the worker threads execute it.
4963 4961 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
4964 4962 assert(_workers != NULL, "Need parallel worker threads.");
4965 4963
4966 4964 ParallelTaskTerminator terminator(_active_workers, _queues);
4967 4965 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
4968 4966
4969 4967 _g1h->set_par_threads(_active_workers);
4970 4968 _workers->run_task(&proc_task_proxy);
4971 4969 _g1h->set_par_threads(0);
4972 4970 }
4973 4971
4974 4972 // Gang task for parallel reference enqueueing.
4975 4973
4976 4974 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
4977 4975 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
4978 4976 EnqueueTask& _enq_task;
4979 4977
4980 4978 public:
4981 4979 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
4982 4980 AbstractGangTask("Enqueue reference objects in parallel"),
4983 4981 _enq_task(enq_task)
4984 4982 { }
4985 4983
4986 4984 virtual void work(int i) {
4987 4985 _enq_task.work(i);
4988 4986 }
4989 4987 };
4990 4988
4991 4989 // Driver routine for parallel reference enqueing.
4992 4990 // Creates an instance of the ref enqueueing gang
4993 4991 // task and has the worker threads execute it.
4994 4992
4995 4993 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
4996 4994 assert(_workers != NULL, "Need parallel worker threads.");
4997 4995
4998 4996 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
4999 4997
5000 4998 _g1h->set_par_threads(_active_workers);
5001 4999 _workers->run_task(&enq_task_proxy);
5002 5000 _g1h->set_par_threads(0);
5003 5001 }
5004 5002
5005 5003 // End of weak reference support closures
5006 5004
5007 5005 // Abstract task used to preserve (i.e. copy) any referent objects
5008 5006 // that are in the collection set and are pointed to by reference
5009 5007 // objects discovered by the CM ref processor.
5010 5008
5011 5009 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5012 5010 protected:
5013 5011 G1CollectedHeap* _g1h;
5014 5012 RefToScanQueueSet *_queues;
5015 5013 ParallelTaskTerminator _terminator;
5016 5014 int _n_workers;
5017 5015
5018 5016 public:
5019 5017 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5020 5018 AbstractGangTask("ParPreserveCMReferents"),
5021 5019 _g1h(g1h),
5022 5020 _queues(task_queues),
5023 5021 _terminator(workers, _queues),
5024 5022 _n_workers(workers)
5025 5023 { }
5026 5024
5027 5025 void work(int i) {
5028 5026 ResourceMark rm;
5029 5027 HandleMark hm;
5030 5028
5031 5029 G1ParScanThreadState pss(_g1h, i);
5032 5030 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5033 5031 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5034 5032 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5035 5033
5036 5034 pss.set_evac_closure(&scan_evac_cl);
5037 5035 pss.set_evac_failure_closure(&evac_failure_cl);
5038 5036 pss.set_partial_scan_closure(&partial_scan_cl);
5039 5037
5040 5038 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5041 5039
5042 5040
5043 5041 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5044 5042 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5045 5043
5046 5044 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5047 5045 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5048 5046
5049 5047 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5050 5048 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5051 5049
5052 5050 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5053 5051 // We also need to mark copied objects.
5054 5052 copy_non_heap_cl = ©_mark_non_heap_cl;
5055 5053 copy_perm_cl = ©_mark_perm_cl;
5056 5054 }
5057 5055
5058 5056 // Is alive closure
5059 5057 G1AlwaysAliveClosure always_alive(_g1h);
5060 5058
5061 5059 // Copying keep alive closure. Applied to referent objects that need
5062 5060 // to be copied.
5063 5061 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5064 5062
5065 5063 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5066 5064
5067 5065 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5068 5066 int stride = MIN2(MAX2(_n_workers, 1), limit);
5069 5067
5070 5068 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5071 5069 // So this must be true - but assert just in case someone decides to
5072 5070 // change the worker ids.
5073 5071 assert(0 <= i && i < limit, "sanity");
5074 5072 assert(!rp->discovery_is_atomic(), "check this code");
5075 5073
5076 5074 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5077 5075 for (int idx = i; idx < limit; idx += stride) {
5078 5076 DiscoveredList& ref_list = rp->discovered_soft_refs()[idx];
5079 5077
5080 5078 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5081 5079 while (iter.has_next()) {
5082 5080 // Since discovery is not atomic for the CM ref processor, we
5083 5081 // can see some null referent objects.
5084 5082 iter.load_ptrs(DEBUG_ONLY(true));
5085 5083 oop ref = iter.obj();
5086 5084
5087 5085 // This will filter nulls.
5088 5086 if (iter.is_referent_alive()) {
5089 5087 iter.make_referent_alive();
5090 5088 }
5091 5089 iter.move_to_next();
5092 5090 }
5093 5091 }
5094 5092
5095 5093 // Drain the queue - which may cause stealing
5096 5094 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5097 5095 drain_queue.do_void();
5098 5096 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5099 5097 assert(pss.refs()->is_empty(), "should be");
5100 5098 }
5101 5099 };
5102 5100
5103 5101 // Weak Reference processing during an evacuation pause (part 1).
5104 5102 void G1CollectedHeap::process_discovered_references() {
5105 5103 double ref_proc_start = os::elapsedTime();
5106 5104
5107 5105 ReferenceProcessor* rp = _ref_processor_stw;
5108 5106 assert(rp->discovery_enabled(), "should have been enabled");
5109 5107
5110 5108 // Any reference objects, in the collection set, that were 'discovered'
5111 5109 // by the CM ref processor should have already been copied (either by
5112 5110 // applying the external root copy closure to the discovered lists, or
5113 5111 // by following an RSet entry).
5114 5112 //
5115 5113 // But some of the referents, that are in the collection set, that these
5116 5114 // reference objects point to may not have been copied: the STW ref
5117 5115 // processor would have seen that the reference object had already
5118 5116 // been 'discovered' and would have skipped discovering the reference,
5119 5117 // but would not have treated the reference object as a regular oop.
5120 5118 // As a reult the copy closure would not have been applied to the
5121 5119 // referent object.
5122 5120 //
5123 5121 // We need to explicitly copy these referent objects - the references
5124 5122 // will be processed at the end of remarking.
5125 5123 //
5126 5124 // We also need to do this copying before we process the reference
5127 5125 // objects discovered by the STW ref processor in case one of these
5128 5126 // referents points to another object which is also referenced by an
5129 5127 // object discovered by the STW ref processor.
5130 5128
5131 5129 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5132 5130 workers()->total_workers() : 1);
5133 5131
5134 5132 set_par_threads(n_workers);
5135 5133 G1ParPreserveCMReferentsTask keep_cm_referents(this, n_workers, _task_queues);
5136 5134
5137 5135 if (G1CollectedHeap::use_parallel_gc_threads()) {
5138 5136 workers()->run_task(&keep_cm_referents);
5139 5137 } else {
5140 5138 keep_cm_referents.work(0);
5141 5139 }
5142 5140
5143 5141 set_par_threads(0);
5144 5142
5145 5143 // Closure to test whether a referent is alive.
5146 5144 G1STWIsAliveClosure is_alive(this);
5147 5145
5148 5146 // Even when parallel reference processing is enabled, the processing
5149 5147 // of JNI refs is serial and performed serially by the current thread
5150 5148 // rather than by a worker. The following PSS will be used for processing
5151 5149 // JNI refs.
5152 5150
5153 5151 // Use only a single queue for this PSS.
5154 5152 G1ParScanThreadState pss(this, 0);
5155 5153
5156 5154 // We do not embed a reference processor in the copying/scanning
5157 5155 // closures while we're actually processing the discovered
5158 5156 // reference objects.
5159 5157 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5160 5158 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5161 5159 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5162 5160
5163 5161 pss.set_evac_closure(&scan_evac_cl);
5164 5162 pss.set_evac_failure_closure(&evac_failure_cl);
5165 5163 pss.set_partial_scan_closure(&partial_scan_cl);
5166 5164
5167 5165 assert(pss.refs()->is_empty(), "pre-condition");
5168 5166
5169 5167 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5170 5168 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL);
5171 5169
5172 5170 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5173 5171 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL);
5174 5172
5175 5173 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5176 5174 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5177 5175
5178 5176 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5179 5177 // We also need to mark copied objects.
5180 5178 copy_non_heap_cl = ©_mark_non_heap_cl;
5181 5179 copy_perm_cl = ©_mark_perm_cl;
5182 5180 }
5183 5181
5184 5182 // Keep alive closure.
5185 5183 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss);
5186 5184
5187 5185 // Serial Complete GC closure
5188 5186 G1STWDrainQueueClosure drain_queue(this, &pss);
5189 5187
5190 5188 // Setup the soft refs policy...
5191 5189 rp->setup_policy(false);
5192 5190
5193 5191 if (!rp->processing_is_mt()) {
5194 5192 // Serial reference processing...
5195 5193 rp->process_discovered_references(&is_alive,
5196 5194 &keep_alive,
5197 5195 &drain_queue,
5198 5196 NULL);
5199 5197 } else {
5200 5198 // Parallel reference processing
5201 5199 int active_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5202 5200 assert(rp->num_q() == active_workers, "sanity");
5203 5201 assert(active_workers <= rp->max_num_q(), "sanity");
5204 5202
5205 5203 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5206 5204 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5207 5205 }
5208 5206
5209 5207 // We have completed copying any necessary live referent objects
5210 5208 // (that were not copied during the actual pause) so we can
5211 5209 // retire any active alloc buffers
5212 5210 pss.retire_alloc_buffers();
5213 5211 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5214 5212
5215 5213 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5216 5214 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0);
5217 5215 }
5218 5216
5219 5217 // Weak Reference processing during an evacuation pause (part 2).
5220 5218 void G1CollectedHeap::enqueue_discovered_references() {
5221 5219 double ref_enq_start = os::elapsedTime();
5222 5220
5223 5221 ReferenceProcessor* rp = _ref_processor_stw;
5224 5222 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5225 5223
5226 5224 // Now enqueue any remaining on the discovered lists on to
5227 5225 // the pending list.
5228 5226 if (!rp->processing_is_mt()) {
5229 5227 // Serial reference processing...
5230 5228 rp->enqueue_discovered_references();
5231 5229 } else {
5232 5230 // Parallel reference enqueuing
5233 5231
5234 5232 int active_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5235 5233 assert(rp->num_q() == active_workers, "sanity");
5236 5234 assert(active_workers <= rp->max_num_q(), "sanity");
5237 5235
5238 5236 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5239 5237 rp->enqueue_discovered_references(&par_task_executor);
5240 5238 }
5241 5239
5242 5240 rp->verify_no_references_recorded();
5243 5241 assert(!rp->discovery_enabled(), "should have been disabled");
5244 5242
5245 5243 // FIXME
5246 5244 // CM's reference processing also cleans up the string and symbol tables.
5247 5245 // Should we do that here also? We could, but it is a serial operation
5248 5246 // and could signicantly increase the pause time.
5249 5247
5250 5248 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5251 5249 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0);
5252 5250 }
5253 5251
5254 5252 void G1CollectedHeap::evacuate_collection_set() {
5255 5253 set_evacuation_failed(false);
5256 5254
5257 5255 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5258 5256 concurrent_g1_refine()->set_use_cache(false);
5259 5257 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5260 5258
5261 5259 int n_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5262 5260 set_par_threads(n_workers);
5263 5261 G1ParTask g1_par_task(this, n_workers, _task_queues);
5264 5262
5265 5263 init_for_evac_failure(NULL);
5266 5264
5267 5265 rem_set()->prepare_for_younger_refs_iterate(true);
5268 5266
5269 5267 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5270 5268 double start_par = os::elapsedTime();
5271 5269
5272 5270 if (G1CollectedHeap::use_parallel_gc_threads()) {
5273 5271 // The individual threads will set their evac-failure closures.
5274 5272 StrongRootsScope srs(this);
5275 5273 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5276 5274 workers()->run_task(&g1_par_task);
5277 5275 } else {
5278 5276 StrongRootsScope srs(this);
5279 5277 g1_par_task.work(0);
5280 5278 }
5281 5279
5282 5280 double par_time = (os::elapsedTime() - start_par) * 1000.0;
5283 5281 g1_policy()->record_par_time(par_time);
5284 5282 set_par_threads(0);
5285 5283
5286 5284 // Process any discovered reference objects - we have
5287 5285 // to do this _before_ we retire the GC alloc regions
5288 5286 // as we may have to copy some 'reachable' referent
5289 5287 // objects (and their reachable sub-graphs) that were
5290 5288 // not copied during the pause.
5291 5289 process_discovered_references();
5292 5290
5293 5291 // Weak root processing.
5294 5292 // Note: when JSR 292 is enabled and code blobs can contain
5295 5293 // non-perm oops then we will need to process the code blobs
5296 5294 // here too.
5297 5295 {
5298 5296 G1STWIsAliveClosure is_alive(this);
5299 5297 G1KeepAliveClosure keep_alive(this);
5300 5298 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5301 5299 }
5302 5300
5303 5301 release_gc_alloc_regions();
5304 5302 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5305 5303
5306 5304 concurrent_g1_refine()->clear_hot_cache();
5307 5305 concurrent_g1_refine()->set_use_cache(true);
5308 5306
5309 5307 finalize_for_evac_failure();
5310 5308
5311 5309 // Must do this before removing self-forwarding pointers, which clears
5312 5310 // the per-region evac-failure flags.
5313 5311 concurrent_mark()->complete_marking_in_collection_set();
5314 5312
5315 5313 if (evacuation_failed()) {
5316 5314 remove_self_forwarding_pointers();
5317 5315 if (PrintGCDetails) {
5318 5316 gclog_or_tty->print(" (to-space overflow)");
5319 5317 } else if (PrintGC) {
5320 5318 gclog_or_tty->print("--");
5321 5319 }
5322 5320 }
5323 5321
5324 5322 // Enqueue any remaining references remaining on the STW
5325 5323 // reference processor's discovered lists. We need to do
5326 5324 // this after the card table is cleaned (and verified) as
5327 5325 // the act of enqueuing entries on to the pending list
5328 5326 // will log these updates (and dirty their associated
5329 5327 // cards). We need these updates logged to update any
5330 5328 // RSets.
5331 5329 enqueue_discovered_references();
5332 5330
5333 5331 if (G1DeferredRSUpdate) {
5334 5332 RedirtyLoggedCardTableEntryFastClosure redirty;
5335 5333 dirty_card_queue_set().set_closure(&redirty);
5336 5334 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5337 5335
5338 5336 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5339 5337 dcq.merge_bufferlists(&dirty_card_queue_set());
5340 5338 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5341 5339 }
5342 5340 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5343 5341 }
5344 5342
5345 5343 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5346 5344 size_t* pre_used,
5347 5345 FreeRegionList* free_list,
5348 5346 HumongousRegionSet* humongous_proxy_set,
5349 5347 HRRSCleanupTask* hrrs_cleanup_task,
5350 5348 bool par) {
5351 5349 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5352 5350 if (hr->isHumongous()) {
5353 5351 assert(hr->startsHumongous(), "we should only see starts humongous");
5354 5352 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5355 5353 } else {
5356 5354 free_region(hr, pre_used, free_list, par);
5357 5355 }
5358 5356 } else {
5359 5357 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5360 5358 }
5361 5359 }
5362 5360
5363 5361 void G1CollectedHeap::free_region(HeapRegion* hr,
5364 5362 size_t* pre_used,
5365 5363 FreeRegionList* free_list,
5366 5364 bool par) {
5367 5365 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5368 5366 assert(!hr->is_empty(), "the region should not be empty");
5369 5367 assert(free_list != NULL, "pre-condition");
5370 5368
5371 5369 *pre_used += hr->used();
5372 5370 hr->hr_clear(par, true /* clear_space */);
5373 5371 free_list->add_as_head(hr);
5374 5372 }
5375 5373
5376 5374 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5377 5375 size_t* pre_used,
5378 5376 FreeRegionList* free_list,
5379 5377 HumongousRegionSet* humongous_proxy_set,
5380 5378 bool par) {
5381 5379 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5382 5380 assert(free_list != NULL, "pre-condition");
5383 5381 assert(humongous_proxy_set != NULL, "pre-condition");
5384 5382
5385 5383 size_t hr_used = hr->used();
5386 5384 size_t hr_capacity = hr->capacity();
5387 5385 size_t hr_pre_used = 0;
5388 5386 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5389 5387 hr->set_notHumongous();
5390 5388 free_region(hr, &hr_pre_used, free_list, par);
5391 5389
5392 5390 size_t i = hr->hrs_index() + 1;
5393 5391 size_t num = 1;
5394 5392 while (i < n_regions()) {
5395 5393 HeapRegion* curr_hr = region_at(i);
5396 5394 if (!curr_hr->continuesHumongous()) {
5397 5395 break;
5398 5396 }
5399 5397 curr_hr->set_notHumongous();
5400 5398 free_region(curr_hr, &hr_pre_used, free_list, par);
5401 5399 num += 1;
5402 5400 i += 1;
5403 5401 }
5404 5402 assert(hr_pre_used == hr_used,
5405 5403 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5406 5404 "should be the same", hr_pre_used, hr_used));
5407 5405 *pre_used += hr_pre_used;
5408 5406 }
5409 5407
5410 5408 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5411 5409 FreeRegionList* free_list,
5412 5410 HumongousRegionSet* humongous_proxy_set,
5413 5411 bool par) {
5414 5412 if (pre_used > 0) {
5415 5413 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5416 5414 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5417 5415 assert(_summary_bytes_used >= pre_used,
5418 5416 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5419 5417 "should be >= pre_used: "SIZE_FORMAT,
5420 5418 _summary_bytes_used, pre_used));
5421 5419 _summary_bytes_used -= pre_used;
5422 5420 }
5423 5421 if (free_list != NULL && !free_list->is_empty()) {
5424 5422 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5425 5423 _free_list.add_as_head(free_list);
5426 5424 }
5427 5425 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5428 5426 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5429 5427 _humongous_set.update_from_proxy(humongous_proxy_set);
5430 5428 }
5431 5429 }
5432 5430
5433 5431 class G1ParCleanupCTTask : public AbstractGangTask {
5434 5432 CardTableModRefBS* _ct_bs;
5435 5433 G1CollectedHeap* _g1h;
5436 5434 HeapRegion* volatile _su_head;
5437 5435 public:
5438 5436 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5439 5437 G1CollectedHeap* g1h) :
5440 5438 AbstractGangTask("G1 Par Cleanup CT Task"),
5441 5439 _ct_bs(ct_bs), _g1h(g1h) { }
5442 5440
5443 5441 void work(int i) {
5444 5442 HeapRegion* r;
5445 5443 while (r = _g1h->pop_dirty_cards_region()) {
5446 5444 clear_cards(r);
5447 5445 }
5448 5446 }
5449 5447
5450 5448 void clear_cards(HeapRegion* r) {
5451 5449 // Cards of the survivors should have already been dirtied.
5452 5450 if (!r->is_survivor()) {
5453 5451 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5454 5452 }
5455 5453 }
5456 5454 };
5457 5455
5458 5456 #ifndef PRODUCT
5459 5457 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5460 5458 G1CollectedHeap* _g1h;
5461 5459 CardTableModRefBS* _ct_bs;
5462 5460 public:
5463 5461 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5464 5462 : _g1h(g1h), _ct_bs(ct_bs) { }
5465 5463 virtual bool doHeapRegion(HeapRegion* r) {
5466 5464 if (r->is_survivor()) {
5467 5465 _g1h->verify_dirty_region(r);
5468 5466 } else {
5469 5467 _g1h->verify_not_dirty_region(r);
5470 5468 }
5471 5469 return false;
5472 5470 }
5473 5471 };
5474 5472
5475 5473 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5476 5474 // All of the region should be clean.
5477 5475 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5478 5476 MemRegion mr(hr->bottom(), hr->end());
5479 5477 ct_bs->verify_not_dirty_region(mr);
5480 5478 }
5481 5479
5482 5480 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5483 5481 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5484 5482 // dirty allocated blocks as they allocate them. The thread that
5485 5483 // retires each region and replaces it with a new one will do a
5486 5484 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5487 5485 // not dirty that area (one less thing to have to do while holding
5488 5486 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5489 5487 // is dirty.
5490 5488 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5491 5489 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5492 5490 ct_bs->verify_dirty_region(mr);
5493 5491 }
5494 5492
5495 5493 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5496 5494 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5497 5495 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5498 5496 verify_dirty_region(hr);
5499 5497 }
5500 5498 }
5501 5499
5502 5500 void G1CollectedHeap::verify_dirty_young_regions() {
5503 5501 verify_dirty_young_list(_young_list->first_region());
5504 5502 verify_dirty_young_list(_young_list->first_survivor_region());
5505 5503 }
5506 5504 #endif
5507 5505
5508 5506 void G1CollectedHeap::cleanUpCardTable() {
5509 5507 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5510 5508 double start = os::elapsedTime();
5511 5509
5512 5510 // Iterate over the dirty cards region list.
5513 5511 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5514 5512
5515 5513 if (ParallelGCThreads > 0) {
5516 5514 set_par_threads(workers()->total_workers());
5517 5515 workers()->run_task(&cleanup_task);
5518 5516 set_par_threads(0);
5519 5517 } else {
5520 5518 while (_dirty_cards_region_list) {
5521 5519 HeapRegion* r = _dirty_cards_region_list;
5522 5520 cleanup_task.clear_cards(r);
5523 5521 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5524 5522 if (_dirty_cards_region_list == r) {
5525 5523 // The last region.
5526 5524 _dirty_cards_region_list = NULL;
5527 5525 }
5528 5526 r->set_next_dirty_cards_region(NULL);
5529 5527 }
5530 5528 }
5531 5529
5532 5530 double elapsed = os::elapsedTime() - start;
5533 5531 g1_policy()->record_clear_ct_time(elapsed * 1000.0);
5534 5532 #ifndef PRODUCT
5535 5533 if (G1VerifyCTCleanup || VerifyAfterGC) {
5536 5534 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5537 5535 heap_region_iterate(&cleanup_verifier);
5538 5536 }
5539 5537 #endif
5540 5538 }
5541 5539
5542 5540 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5543 5541 size_t pre_used = 0;
5544 5542 FreeRegionList local_free_list("Local List for CSet Freeing");
5545 5543
5546 5544 double young_time_ms = 0.0;
5547 5545 double non_young_time_ms = 0.0;
5548 5546
5549 5547 // Since the collection set is a superset of the the young list,
5550 5548 // all we need to do to clear the young list is clear its
5551 5549 // head and length, and unlink any young regions in the code below
5552 5550 _young_list->clear();
5553 5551
5554 5552 G1CollectorPolicy* policy = g1_policy();
5555 5553
5556 5554 double start_sec = os::elapsedTime();
5557 5555 bool non_young = true;
5558 5556
5559 5557 HeapRegion* cur = cs_head;
5560 5558 int age_bound = -1;
5561 5559 size_t rs_lengths = 0;
5562 5560
5563 5561 while (cur != NULL) {
5564 5562 assert(!is_on_master_free_list(cur), "sanity");
5565 5563
5566 5564 if (non_young) {
5567 5565 if (cur->is_young()) {
5568 5566 double end_sec = os::elapsedTime();
5569 5567 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5570 5568 non_young_time_ms += elapsed_ms;
5571 5569
5572 5570 start_sec = os::elapsedTime();
5573 5571 non_young = false;
5574 5572 }
5575 5573 } else {
5576 5574 double end_sec = os::elapsedTime();
5577 5575 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5578 5576 young_time_ms += elapsed_ms;
5579 5577
5580 5578 start_sec = os::elapsedTime();
5581 5579 non_young = true;
5582 5580 }
5583 5581
5584 5582 rs_lengths += cur->rem_set()->occupied();
5585 5583
5586 5584 HeapRegion* next = cur->next_in_collection_set();
5587 5585 assert(cur->in_collection_set(), "bad CS");
5588 5586 cur->set_next_in_collection_set(NULL);
5589 5587 cur->set_in_collection_set(false);
5590 5588
5591 5589 if (cur->is_young()) {
5592 5590 int index = cur->young_index_in_cset();
5593 5591 guarantee( index != -1, "invariant" );
5594 5592 guarantee( (size_t)index < policy->young_cset_length(), "invariant" );
5595 5593 size_t words_survived = _surviving_young_words[index];
5596 5594 cur->record_surv_words_in_group(words_survived);
5597 5595
5598 5596 // At this point the we have 'popped' cur from the collection set
5599 5597 // (linked via next_in_collection_set()) but it is still in the
5600 5598 // young list (linked via next_young_region()). Clear the
5601 5599 // _next_young_region field.
5602 5600 cur->set_next_young_region(NULL);
5603 5601 } else {
5604 5602 int index = cur->young_index_in_cset();
5605 5603 guarantee( index == -1, "invariant" );
5606 5604 }
5607 5605
5608 5606 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5609 5607 (!cur->is_young() && cur->young_index_in_cset() == -1),
5610 5608 "invariant" );
5611 5609
5612 5610 if (!cur->evacuation_failed()) {
5613 5611 // And the region is empty.
5614 5612 assert(!cur->is_empty(), "Should not have empty regions in a CS.");
5615 5613 free_region(cur, &pre_used, &local_free_list, false /* par */);
5616 5614 } else {
5617 5615 cur->uninstall_surv_rate_group();
5618 5616 if (cur->is_young())
5619 5617 cur->set_young_index_in_cset(-1);
5620 5618 cur->set_not_young();
5621 5619 cur->set_evacuation_failed(false);
5622 5620 }
5623 5621 cur = next;
5624 5622 }
5625 5623
5626 5624 policy->record_max_rs_lengths(rs_lengths);
5627 5625 policy->cset_regions_freed();
5628 5626
5629 5627 double end_sec = os::elapsedTime();
5630 5628 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5631 5629 if (non_young)
5632 5630 non_young_time_ms += elapsed_ms;
5633 5631 else
5634 5632 young_time_ms += elapsed_ms;
5635 5633
5636 5634 update_sets_after_freeing_regions(pre_used, &local_free_list,
5637 5635 NULL /* humongous_proxy_set */,
5638 5636 false /* par */);
5639 5637 policy->record_young_free_cset_time_ms(young_time_ms);
5640 5638 policy->record_non_young_free_cset_time_ms(non_young_time_ms);
5641 5639 }
5642 5640
5643 5641 // This routine is similar to the above but does not record
5644 5642 // any policy statistics or update free lists; we are abandoning
5645 5643 // the current incremental collection set in preparation of a
5646 5644 // full collection. After the full GC we will start to build up
5647 5645 // the incremental collection set again.
5648 5646 // This is only called when we're doing a full collection
5649 5647 // and is immediately followed by the tearing down of the young list.
5650 5648
5651 5649 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
5652 5650 HeapRegion* cur = cs_head;
5653 5651
5654 5652 while (cur != NULL) {
5655 5653 HeapRegion* next = cur->next_in_collection_set();
5656 5654 assert(cur->in_collection_set(), "bad CS");
5657 5655 cur->set_next_in_collection_set(NULL);
5658 5656 cur->set_in_collection_set(false);
5659 5657 cur->set_young_index_in_cset(-1);
5660 5658 cur = next;
5661 5659 }
5662 5660 }
5663 5661
5664 5662 void G1CollectedHeap::set_free_regions_coming() {
5665 5663 if (G1ConcRegionFreeingVerbose) {
5666 5664 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5667 5665 "setting free regions coming");
5668 5666 }
5669 5667
5670 5668 assert(!free_regions_coming(), "pre-condition");
5671 5669 _free_regions_coming = true;
5672 5670 }
5673 5671
5674 5672 void G1CollectedHeap::reset_free_regions_coming() {
5675 5673 {
5676 5674 assert(free_regions_coming(), "pre-condition");
5677 5675 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5678 5676 _free_regions_coming = false;
5679 5677 SecondaryFreeList_lock->notify_all();
5680 5678 }
5681 5679
5682 5680 if (G1ConcRegionFreeingVerbose) {
5683 5681 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5684 5682 "reset free regions coming");
5685 5683 }
5686 5684 }
5687 5685
5688 5686 void G1CollectedHeap::wait_while_free_regions_coming() {
5689 5687 // Most of the time we won't have to wait, so let's do a quick test
5690 5688 // first before we take the lock.
5691 5689 if (!free_regions_coming()) {
5692 5690 return;
5693 5691 }
5694 5692
5695 5693 if (G1ConcRegionFreeingVerbose) {
5696 5694 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5697 5695 "waiting for free regions");
5698 5696 }
5699 5697
5700 5698 {
5701 5699 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5702 5700 while (free_regions_coming()) {
5703 5701 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
5704 5702 }
5705 5703 }
5706 5704
5707 5705 if (G1ConcRegionFreeingVerbose) {
5708 5706 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5709 5707 "done waiting for free regions");
5710 5708 }
5711 5709 }
5712 5710
5713 5711 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
5714 5712 assert(heap_lock_held_for_gc(),
5715 5713 "the heap lock should already be held by or for this thread");
5716 5714 _young_list->push_region(hr);
5717 5715 g1_policy()->set_region_short_lived(hr);
5718 5716 }
5719 5717
5720 5718 class NoYoungRegionsClosure: public HeapRegionClosure {
5721 5719 private:
5722 5720 bool _success;
5723 5721 public:
5724 5722 NoYoungRegionsClosure() : _success(true) { }
5725 5723 bool doHeapRegion(HeapRegion* r) {
5726 5724 if (r->is_young()) {
5727 5725 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
5728 5726 r->bottom(), r->end());
5729 5727 _success = false;
5730 5728 }
5731 5729 return false;
5732 5730 }
5733 5731 bool success() { return _success; }
5734 5732 };
5735 5733
5736 5734 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
5737 5735 bool ret = _young_list->check_list_empty(check_sample);
5738 5736
5739 5737 if (check_heap) {
5740 5738 NoYoungRegionsClosure closure;
5741 5739 heap_region_iterate(&closure);
5742 5740 ret = ret && closure.success();
5743 5741 }
5744 5742
5745 5743 return ret;
5746 5744 }
5747 5745
5748 5746 void G1CollectedHeap::empty_young_list() {
5749 5747 assert(heap_lock_held_for_gc(),
5750 5748 "the heap lock should already be held by or for this thread");
5751 5749
5752 5750 _young_list->empty_list();
5753 5751 }
5754 5752
5755 5753 // Done at the start of full GC.
5756 5754 void G1CollectedHeap::tear_down_region_lists() {
5757 5755 _free_list.remove_all();
5758 5756 }
5759 5757
5760 5758 class RegionResetter: public HeapRegionClosure {
5761 5759 G1CollectedHeap* _g1h;
5762 5760 FreeRegionList _local_free_list;
5763 5761
5764 5762 public:
5765 5763 RegionResetter() : _g1h(G1CollectedHeap::heap()),
5766 5764 _local_free_list("Local Free List for RegionResetter") { }
5767 5765
5768 5766 bool doHeapRegion(HeapRegion* r) {
5769 5767 if (r->continuesHumongous()) return false;
5770 5768 if (r->top() > r->bottom()) {
5771 5769 if (r->top() < r->end()) {
5772 5770 Copy::fill_to_words(r->top(),
5773 5771 pointer_delta(r->end(), r->top()));
5774 5772 }
5775 5773 } else {
5776 5774 assert(r->is_empty(), "tautology");
5777 5775 _local_free_list.add_as_tail(r);
5778 5776 }
5779 5777 return false;
5780 5778 }
5781 5779
5782 5780 void update_free_lists() {
5783 5781 _g1h->update_sets_after_freeing_regions(0, &_local_free_list, NULL,
5784 5782 false /* par */);
5785 5783 }
5786 5784 };
5787 5785
5788 5786 // Done at the end of full GC.
5789 5787 void G1CollectedHeap::rebuild_region_lists() {
5790 5788 // This needs to go at the end of the full GC.
5791 5789 RegionResetter rs;
5792 5790 heap_region_iterate(&rs);
5793 5791 rs.update_free_lists();
5794 5792 }
5795 5793
5796 5794 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
5797 5795 _refine_cte_cl->set_concurrent(concurrent);
5798 5796 }
5799 5797
5800 5798 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
5801 5799 HeapRegion* hr = heap_region_containing(p);
5802 5800 if (hr == NULL) {
5803 5801 return is_in_permanent(p);
5804 5802 } else {
5805 5803 return hr->is_in(p);
5806 5804 }
5807 5805 }
5808 5806
5809 5807 // Methods for the mutator alloc region
5810 5808
5811 5809 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
5812 5810 bool force) {
5813 5811 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5814 5812 assert(!force || g1_policy()->can_expand_young_list(),
5815 5813 "if force is true we should be able to expand the young list");
5816 5814 bool young_list_full = g1_policy()->is_young_list_full();
5817 5815 if (force || !young_list_full) {
5818 5816 HeapRegion* new_alloc_region = new_region(word_size,
5819 5817 false /* do_expand */);
5820 5818 if (new_alloc_region != NULL) {
5821 5819 g1_policy()->update_region_num(true /* next_is_young */);
5822 5820 set_region_short_lived_locked(new_alloc_region);
5823 5821 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
5824 5822 return new_alloc_region;
5825 5823 }
5826 5824 }
5827 5825 return NULL;
5828 5826 }
5829 5827
5830 5828 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
5831 5829 size_t allocated_bytes) {
5832 5830 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5833 5831 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
5834 5832
5835 5833 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
5836 5834 _summary_bytes_used += allocated_bytes;
5837 5835 _hr_printer.retire(alloc_region);
5838 5836 // We update the eden sizes here, when the region is retired,
5839 5837 // instead of when it's allocated, since this is the point that its
5840 5838 // used space has been recored in _summary_bytes_used.
5841 5839 g1mm()->update_eden_size();
5842 5840 }
5843 5841
5844 5842 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
5845 5843 bool force) {
5846 5844 return _g1h->new_mutator_alloc_region(word_size, force);
5847 5845 }
5848 5846
5849 5847 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
5850 5848 size_t allocated_bytes) {
5851 5849 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
5852 5850 }
5853 5851
5854 5852 // Methods for the GC alloc regions
5855 5853
5856 5854 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
5857 5855 size_t count,
5858 5856 GCAllocPurpose ap) {
5859 5857 assert(FreeList_lock->owned_by_self(), "pre-condition");
5860 5858
5861 5859 if (count < g1_policy()->max_regions(ap)) {
5862 5860 HeapRegion* new_alloc_region = new_region(word_size,
5863 5861 true /* do_expand */);
5864 5862 if (new_alloc_region != NULL) {
5865 5863 // We really only need to do this for old regions given that we
5866 5864 // should never scan survivors. But it doesn't hurt to do it
5867 5865 // for survivors too.
5868 5866 new_alloc_region->set_saved_mark();
5869 5867 if (ap == GCAllocForSurvived) {
5870 5868 new_alloc_region->set_survivor();
5871 5869 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
5872 5870 } else {
5873 5871 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
5874 5872 }
5875 5873 return new_alloc_region;
5876 5874 } else {
5877 5875 g1_policy()->note_alloc_region_limit_reached(ap);
5878 5876 }
5879 5877 }
5880 5878 return NULL;
5881 5879 }
5882 5880
5883 5881 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
5884 5882 size_t allocated_bytes,
5885 5883 GCAllocPurpose ap) {
5886 5884 alloc_region->note_end_of_copying();
5887 5885 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
5888 5886 if (ap == GCAllocForSurvived) {
5889 5887 young_list()->add_survivor_region(alloc_region);
5890 5888 }
5891 5889 _hr_printer.retire(alloc_region);
5892 5890 }
5893 5891
5894 5892 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
5895 5893 bool force) {
5896 5894 assert(!force, "not supported for GC alloc regions");
5897 5895 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
5898 5896 }
5899 5897
5900 5898 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
5901 5899 size_t allocated_bytes) {
5902 5900 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
5903 5901 GCAllocForSurvived);
5904 5902 }
5905 5903
5906 5904 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
5907 5905 bool force) {
5908 5906 assert(!force, "not supported for GC alloc regions");
5909 5907 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
5910 5908 }
5911 5909
5912 5910 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
5913 5911 size_t allocated_bytes) {
5914 5912 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
5915 5913 GCAllocForTenured);
5916 5914 }
5917 5915 // Heap region set verification
5918 5916
5919 5917 class VerifyRegionListsClosure : public HeapRegionClosure {
5920 5918 private:
5921 5919 HumongousRegionSet* _humongous_set;
5922 5920 FreeRegionList* _free_list;
5923 5921 size_t _region_count;
5924 5922
5925 5923 public:
5926 5924 VerifyRegionListsClosure(HumongousRegionSet* humongous_set,
5927 5925 FreeRegionList* free_list) :
5928 5926 _humongous_set(humongous_set), _free_list(free_list),
5929 5927 _region_count(0) { }
5930 5928
5931 5929 size_t region_count() { return _region_count; }
5932 5930
5933 5931 bool doHeapRegion(HeapRegion* hr) {
5934 5932 _region_count += 1;
5935 5933
5936 5934 if (hr->continuesHumongous()) {
5937 5935 return false;
5938 5936 }
5939 5937
5940 5938 if (hr->is_young()) {
5941 5939 // TODO
5942 5940 } else if (hr->startsHumongous()) {
5943 5941 _humongous_set->verify_next_region(hr);
5944 5942 } else if (hr->is_empty()) {
5945 5943 _free_list->verify_next_region(hr);
5946 5944 }
5947 5945 return false;
5948 5946 }
5949 5947 };
5950 5948
5951 5949 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index,
5952 5950 HeapWord* bottom) {
5953 5951 HeapWord* end = bottom + HeapRegion::GrainWords;
5954 5952 MemRegion mr(bottom, end);
5955 5953 assert(_g1_reserved.contains(mr), "invariant");
5956 5954 // This might return NULL if the allocation fails
5957 5955 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
5958 5956 }
5959 5957
5960 5958 void G1CollectedHeap::verify_region_sets() {
5961 5959 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5962 5960
5963 5961 // First, check the explicit lists.
5964 5962 _free_list.verify();
5965 5963 {
5966 5964 // Given that a concurrent operation might be adding regions to
5967 5965 // the secondary free list we have to take the lock before
5968 5966 // verifying it.
5969 5967 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5970 5968 _secondary_free_list.verify();
5971 5969 }
5972 5970 _humongous_set.verify();
5973 5971
5974 5972 // If a concurrent region freeing operation is in progress it will
5975 5973 // be difficult to correctly attributed any free regions we come
5976 5974 // across to the correct free list given that they might belong to
5977 5975 // one of several (free_list, secondary_free_list, any local lists,
5978 5976 // etc.). So, if that's the case we will skip the rest of the
5979 5977 // verification operation. Alternatively, waiting for the concurrent
5980 5978 // operation to complete will have a non-trivial effect on the GC's
5981 5979 // operation (no concurrent operation will last longer than the
5982 5980 // interval between two calls to verification) and it might hide
5983 5981 // any issues that we would like to catch during testing.
5984 5982 if (free_regions_coming()) {
5985 5983 return;
5986 5984 }
5987 5985
5988 5986 // Make sure we append the secondary_free_list on the free_list so
5989 5987 // that all free regions we will come across can be safely
5990 5988 // attributed to the free_list.
5991 5989 append_secondary_free_list_if_not_empty_with_lock();
5992 5990
5993 5991 // Finally, make sure that the region accounting in the lists is
5994 5992 // consistent with what we see in the heap.
5995 5993 _humongous_set.verify_start();
5996 5994 _free_list.verify_start();
5997 5995
5998 5996 VerifyRegionListsClosure cl(&_humongous_set, &_free_list);
5999 5997 heap_region_iterate(&cl);
6000 5998
6001 5999 _humongous_set.verify_end();
6002 6000 _free_list.verify_end();
6003 6001 }
↓ open down ↓ |
2938 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX