Print this page
rev 2724 : 6484965: G1: piggy-back liveness accounting phase on marking
Summary: Remove the separate counting phase of concurrent marking by tracking the amount of marked bytes and the cards spanned by marked objects in marking task/worker thread local data structures, which are updated as individual objects are marked.
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);
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) {
555 555 assert(!isHumongous(word_size) ||
556 556 word_size <= (size_t) HeapRegion::GrainWords,
557 557 "the only time we use this to allocate a humongous region is "
558 558 "when we are allocating a single humongous region");
559 559
560 560 HeapRegion* res;
561 561 if (G1StressConcRegionFreeing) {
562 562 if (!_secondary_free_list.is_empty()) {
563 563 if (G1ConcRegionFreeingVerbose) {
564 564 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
565 565 "forced to look at the secondary_free_list");
566 566 }
567 567 res = new_region_try_secondary_free_list();
568 568 if (res != NULL) {
569 569 return res;
570 570 }
571 571 }
572 572 }
573 573 res = _free_list.remove_head_or_null();
574 574 if (res == NULL) {
575 575 if (G1ConcRegionFreeingVerbose) {
576 576 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
577 577 "res == NULL, trying the secondary_free_list");
578 578 }
579 579 res = new_region_try_secondary_free_list();
580 580 }
581 581 if (res == NULL && do_expand) {
582 582 ergo_verbose1(ErgoHeapSizing,
583 583 "attempt heap expansion",
584 584 ergo_format_reason("region allocation request failed")
585 585 ergo_format_byte("allocation request"),
586 586 word_size * HeapWordSize);
587 587 if (expand(word_size * HeapWordSize)) {
588 588 // Even though the heap was expanded, it might not have reached
589 589 // the desired size. So, we cannot assume that the allocation
590 590 // will succeed.
591 591 res = _free_list.remove_head_or_null();
592 592 }
593 593 }
594 594 return res;
595 595 }
596 596
597 597 size_t G1CollectedHeap::humongous_obj_allocate_find_first(size_t num_regions,
598 598 size_t word_size) {
599 599 assert(isHumongous(word_size), "word_size should be humongous");
600 600 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
601 601
602 602 size_t first = G1_NULL_HRS_INDEX;
603 603 if (num_regions == 1) {
604 604 // Only one region to allocate, no need to go through the slower
605 605 // path. The caller will attempt the expasion if this fails, so
606 606 // let's not try to expand here too.
607 607 HeapRegion* hr = new_region(word_size, false /* do_expand */);
608 608 if (hr != NULL) {
609 609 first = hr->hrs_index();
610 610 } else {
611 611 first = G1_NULL_HRS_INDEX;
612 612 }
613 613 } else {
614 614 // We can't allocate humongous regions while cleanupComplete() is
615 615 // running, since some of the regions we find to be empty might not
616 616 // yet be added to the free list and it is not straightforward to
617 617 // know which list they are on so that we can remove them. Note
618 618 // that we only need to do this if we need to allocate more than
619 619 // one region to satisfy the current humongous allocation
620 620 // request. If we are only allocating one region we use the common
621 621 // region allocation code (see above).
622 622 wait_while_free_regions_coming();
623 623 append_secondary_free_list_if_not_empty_with_lock();
624 624
625 625 if (free_regions() >= num_regions) {
626 626 first = _hrs.find_contiguous(num_regions);
627 627 if (first != G1_NULL_HRS_INDEX) {
628 628 for (size_t i = first; i < first + num_regions; ++i) {
629 629 HeapRegion* hr = region_at(i);
630 630 assert(hr->is_empty(), "sanity");
631 631 assert(is_on_master_free_list(hr), "sanity");
632 632 hr->set_pending_removal(true);
633 633 }
634 634 _free_list.remove_all_pending(num_regions);
635 635 }
636 636 }
637 637 }
638 638 return first;
639 639 }
640 640
641 641 HeapWord*
642 642 G1CollectedHeap::humongous_obj_allocate_initialize_regions(size_t first,
643 643 size_t num_regions,
644 644 size_t word_size) {
645 645 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
646 646 assert(isHumongous(word_size), "word_size should be humongous");
647 647 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
648 648
649 649 // Index of last region in the series + 1.
650 650 size_t last = first + num_regions;
651 651
652 652 // We need to initialize the region(s) we just discovered. This is
653 653 // a bit tricky given that it can happen concurrently with
654 654 // refinement threads refining cards on these regions and
655 655 // potentially wanting to refine the BOT as they are scanning
656 656 // those cards (this can happen shortly after a cleanup; see CR
657 657 // 6991377). So we have to set up the region(s) carefully and in
658 658 // a specific order.
659 659
660 660 // The word size sum of all the regions we will allocate.
661 661 size_t word_size_sum = num_regions * HeapRegion::GrainWords;
662 662 assert(word_size <= word_size_sum, "sanity");
663 663
664 664 // This will be the "starts humongous" region.
665 665 HeapRegion* first_hr = region_at(first);
666 666 // The header of the new object will be placed at the bottom of
667 667 // the first region.
668 668 HeapWord* new_obj = first_hr->bottom();
669 669 // This will be the new end of the first region in the series that
670 670 // should also match the end of the last region in the seriers.
671 671 HeapWord* new_end = new_obj + word_size_sum;
672 672 // This will be the new top of the first region that will reflect
673 673 // this allocation.
674 674 HeapWord* new_top = new_obj + word_size;
675 675
676 676 // First, we need to zero the header of the space that we will be
677 677 // allocating. When we update top further down, some refinement
678 678 // threads might try to scan the region. By zeroing the header we
679 679 // ensure that any thread that will try to scan the region will
680 680 // come across the zero klass word and bail out.
681 681 //
682 682 // NOTE: It would not have been correct to have used
683 683 // CollectedHeap::fill_with_object() and make the space look like
684 684 // an int array. The thread that is doing the allocation will
685 685 // later update the object header to a potentially different array
686 686 // type and, for a very short period of time, the klass and length
687 687 // fields will be inconsistent. This could cause a refinement
688 688 // thread to calculate the object size incorrectly.
689 689 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
690 690
691 691 // We will set up the first region as "starts humongous". This
692 692 // will also update the BOT covering all the regions to reflect
693 693 // that there is a single object that starts at the bottom of the
694 694 // first region.
695 695 first_hr->set_startsHumongous(new_top, new_end);
696 696
697 697 // Then, if there are any, we will set up the "continues
698 698 // humongous" regions.
699 699 HeapRegion* hr = NULL;
700 700 for (size_t i = first + 1; i < last; ++i) {
701 701 hr = region_at(i);
702 702 hr->set_continuesHumongous(first_hr);
703 703 }
704 704 // If we have "continues humongous" regions (hr != NULL), then the
705 705 // end of the last one should match new_end.
706 706 assert(hr == NULL || hr->end() == new_end, "sanity");
707 707
708 708 // Up to this point no concurrent thread would have been able to
709 709 // do any scanning on any region in this series. All the top
710 710 // fields still point to bottom, so the intersection between
711 711 // [bottom,top] and [card_start,card_end] will be empty. Before we
712 712 // update the top fields, we'll do a storestore to make sure that
713 713 // no thread sees the update to top before the zeroing of the
714 714 // object header and the BOT initialization.
715 715 OrderAccess::storestore();
716 716
717 717 // Now that the BOT and the object header have been initialized,
718 718 // we can update top of the "starts humongous" region.
719 719 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
720 720 "new_top should be in this region");
721 721 first_hr->set_top(new_top);
722 722 if (_hr_printer.is_active()) {
723 723 HeapWord* bottom = first_hr->bottom();
724 724 HeapWord* end = first_hr->orig_end();
725 725 if ((first + 1) == last) {
726 726 // the series has a single humongous region
727 727 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
728 728 } else {
729 729 // the series has more than one humongous regions
730 730 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
731 731 }
732 732 }
733 733
734 734 // Now, we will update the top fields of the "continues humongous"
735 735 // regions. The reason we need to do this is that, otherwise,
736 736 // these regions would look empty and this will confuse parts of
737 737 // G1. For example, the code that looks for a consecutive number
738 738 // of empty regions will consider them empty and try to
739 739 // re-allocate them. We can extend is_empty() to also include
740 740 // !continuesHumongous(), but it is easier to just update the top
741 741 // fields here. The way we set top for all regions (i.e., top ==
742 742 // end for all regions but the last one, top == new_top for the
743 743 // last one) is actually used when we will free up the humongous
744 744 // region in free_humongous_region().
745 745 hr = NULL;
746 746 for (size_t i = first + 1; i < last; ++i) {
747 747 hr = region_at(i);
748 748 if ((i + 1) == last) {
749 749 // last continues humongous region
750 750 assert(hr->bottom() < new_top && new_top <= hr->end(),
751 751 "new_top should fall on this region");
752 752 hr->set_top(new_top);
753 753 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
754 754 } else {
755 755 // not last one
756 756 assert(new_top > hr->end(), "new_top should be above this region");
757 757 hr->set_top(hr->end());
758 758 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
759 759 }
760 760 }
761 761 // If we have continues humongous regions (hr != NULL), then the
762 762 // end of the last one should match new_end and its top should
763 763 // match new_top.
764 764 assert(hr == NULL ||
765 765 (hr->end() == new_end && hr->top() == new_top), "sanity");
766 766
767 767 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
768 768 _summary_bytes_used += first_hr->used();
769 769 _humongous_set.add(first_hr);
770 770
771 771 return new_obj;
772 772 }
773 773
774 774 // If could fit into free regions w/o expansion, try.
775 775 // Otherwise, if can expand, do so.
776 776 // Otherwise, if using ex regions might help, try with ex given back.
777 777 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
778 778 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
779 779
780 780 verify_region_sets_optional();
781 781
782 782 size_t num_regions =
783 783 round_to(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords;
784 784 size_t x_size = expansion_regions();
785 785 size_t fs = _hrs.free_suffix();
786 786 size_t first = humongous_obj_allocate_find_first(num_regions, word_size);
787 787 if (first == G1_NULL_HRS_INDEX) {
788 788 // The only thing we can do now is attempt expansion.
789 789 if (fs + x_size >= num_regions) {
790 790 // If the number of regions we're trying to allocate for this
791 791 // object is at most the number of regions in the free suffix,
792 792 // then the call to humongous_obj_allocate_find_first() above
793 793 // should have succeeded and we wouldn't be here.
794 794 //
795 795 // We should only be trying to expand when the free suffix is
796 796 // not sufficient for the object _and_ we have some expansion
797 797 // room available.
798 798 assert(num_regions > fs, "earlier allocation should have succeeded");
799 799
800 800 ergo_verbose1(ErgoHeapSizing,
801 801 "attempt heap expansion",
802 802 ergo_format_reason("humongous allocation request failed")
803 803 ergo_format_byte("allocation request"),
804 804 word_size * HeapWordSize);
805 805 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
806 806 // Even though the heap was expanded, it might not have
807 807 // reached the desired size. So, we cannot assume that the
808 808 // allocation will succeed.
809 809 first = humongous_obj_allocate_find_first(num_regions, word_size);
810 810 }
811 811 }
812 812 }
813 813
814 814 HeapWord* result = NULL;
815 815 if (first != G1_NULL_HRS_INDEX) {
816 816 result =
817 817 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
818 818 assert(result != NULL, "it should always return a valid result");
819 819
820 820 // A successful humongous object allocation changes the used space
821 821 // information of the old generation so we need to recalculate the
822 822 // sizes and update the jstat counters here.
823 823 g1mm()->update_sizes();
824 824 }
825 825
826 826 verify_region_sets_optional();
827 827
828 828 return result;
829 829 }
830 830
831 831 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
832 832 assert_heap_not_locked_and_not_at_safepoint();
833 833 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
834 834
835 835 unsigned int dummy_gc_count_before;
836 836 return attempt_allocation(word_size, &dummy_gc_count_before);
837 837 }
838 838
839 839 HeapWord*
840 840 G1CollectedHeap::mem_allocate(size_t word_size,
841 841 bool* gc_overhead_limit_was_exceeded) {
842 842 assert_heap_not_locked_and_not_at_safepoint();
843 843
844 844 // Loop until the allocation is satisified, or unsatisfied after GC.
845 845 for (int try_count = 1; /* we'll return */; try_count += 1) {
846 846 unsigned int gc_count_before;
847 847
848 848 HeapWord* result = NULL;
849 849 if (!isHumongous(word_size)) {
850 850 result = attempt_allocation(word_size, &gc_count_before);
851 851 } else {
852 852 result = attempt_allocation_humongous(word_size, &gc_count_before);
853 853 }
854 854 if (result != NULL) {
855 855 return result;
856 856 }
857 857
858 858 // Create the garbage collection operation...
859 859 VM_G1CollectForAllocation op(gc_count_before, word_size);
860 860 // ...and get the VM thread to execute it.
861 861 VMThread::execute(&op);
862 862
863 863 if (op.prologue_succeeded() && op.pause_succeeded()) {
864 864 // If the operation was successful we'll return the result even
865 865 // if it is NULL. If the allocation attempt failed immediately
866 866 // after a Full GC, it's unlikely we'll be able to allocate now.
867 867 HeapWord* result = op.result();
868 868 if (result != NULL && !isHumongous(word_size)) {
869 869 // Allocations that take place on VM operations do not do any
870 870 // card dirtying and we have to do it here. We only have to do
871 871 // this for non-humongous allocations, though.
872 872 dirty_young_block(result, word_size);
873 873 }
874 874 return result;
875 875 } else {
876 876 assert(op.result() == NULL,
877 877 "the result should be NULL if the VM op did not succeed");
878 878 }
879 879
880 880 // Give a warning if we seem to be looping forever.
881 881 if ((QueuedAllocationWarningCount > 0) &&
882 882 (try_count % QueuedAllocationWarningCount == 0)) {
883 883 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
884 884 }
885 885 }
886 886
887 887 ShouldNotReachHere();
888 888 return NULL;
889 889 }
890 890
891 891 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
892 892 unsigned int *gc_count_before_ret) {
893 893 // Make sure you read the note in attempt_allocation_humongous().
894 894
895 895 assert_heap_not_locked_and_not_at_safepoint();
896 896 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
897 897 "be called for humongous allocation requests");
898 898
899 899 // We should only get here after the first-level allocation attempt
900 900 // (attempt_allocation()) failed to allocate.
901 901
902 902 // We will loop until a) we manage to successfully perform the
903 903 // allocation or b) we successfully schedule a collection which
904 904 // fails to perform the allocation. b) is the only case when we'll
905 905 // return NULL.
906 906 HeapWord* result = NULL;
907 907 for (int try_count = 1; /* we'll return */; try_count += 1) {
908 908 bool should_try_gc;
909 909 unsigned int gc_count_before;
910 910
911 911 {
912 912 MutexLockerEx x(Heap_lock);
913 913
914 914 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
915 915 false /* bot_updates */);
916 916 if (result != NULL) {
917 917 return result;
918 918 }
919 919
920 920 // If we reach here, attempt_allocation_locked() above failed to
921 921 // allocate a new region. So the mutator alloc region should be NULL.
922 922 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
923 923
924 924 if (GC_locker::is_active_and_needs_gc()) {
925 925 if (g1_policy()->can_expand_young_list()) {
926 926 // No need for an ergo verbose message here,
927 927 // can_expand_young_list() does this when it returns true.
928 928 result = _mutator_alloc_region.attempt_allocation_force(word_size,
929 929 false /* bot_updates */);
930 930 if (result != NULL) {
931 931 return result;
932 932 }
933 933 }
934 934 should_try_gc = false;
935 935 } else {
936 936 // Read the GC count while still holding the Heap_lock.
937 937 gc_count_before = SharedHeap::heap()->total_collections();
938 938 should_try_gc = true;
939 939 }
940 940 }
941 941
942 942 if (should_try_gc) {
943 943 bool succeeded;
944 944 result = do_collection_pause(word_size, gc_count_before, &succeeded);
945 945 if (result != NULL) {
946 946 assert(succeeded, "only way to get back a non-NULL result");
947 947 return result;
948 948 }
949 949
950 950 if (succeeded) {
951 951 // If we get here we successfully scheduled a collection which
952 952 // failed to allocate. No point in trying to allocate
953 953 // further. We'll just return NULL.
954 954 MutexLockerEx x(Heap_lock);
955 955 *gc_count_before_ret = SharedHeap::heap()->total_collections();
956 956 return NULL;
957 957 }
958 958 } else {
959 959 GC_locker::stall_until_clear();
960 960 }
961 961
962 962 // We can reach here if we were unsuccessul in scheduling a
963 963 // collection (because another thread beat us to it) or if we were
964 964 // stalled due to the GC locker. In either can we should retry the
965 965 // allocation attempt in case another thread successfully
966 966 // performed a collection and reclaimed enough space. We do the
967 967 // first attempt (without holding the Heap_lock) here and the
968 968 // follow-on attempt will be at the start of the next loop
969 969 // iteration (after taking the Heap_lock).
970 970 result = _mutator_alloc_region.attempt_allocation(word_size,
971 971 false /* bot_updates */);
972 972 if (result != NULL ){
973 973 return result;
974 974 }
975 975
976 976 // Give a warning if we seem to be looping forever.
977 977 if ((QueuedAllocationWarningCount > 0) &&
978 978 (try_count % QueuedAllocationWarningCount == 0)) {
979 979 warning("G1CollectedHeap::attempt_allocation_slow() "
980 980 "retries %d times", try_count);
981 981 }
982 982 }
983 983
984 984 ShouldNotReachHere();
985 985 return NULL;
986 986 }
987 987
988 988 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
989 989 unsigned int * gc_count_before_ret) {
990 990 // The structure of this method has a lot of similarities to
991 991 // attempt_allocation_slow(). The reason these two were not merged
992 992 // into a single one is that such a method would require several "if
993 993 // allocation is not humongous do this, otherwise do that"
994 994 // conditional paths which would obscure its flow. In fact, an early
995 995 // version of this code did use a unified method which was harder to
996 996 // follow and, as a result, it had subtle bugs that were hard to
997 997 // track down. So keeping these two methods separate allows each to
998 998 // be more readable. It will be good to keep these two in sync as
999 999 // much as possible.
1000 1000
1001 1001 assert_heap_not_locked_and_not_at_safepoint();
1002 1002 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1003 1003 "should only be called for humongous allocations");
1004 1004
1005 1005 // We will loop until a) we manage to successfully perform the
1006 1006 // allocation or b) we successfully schedule a collection which
1007 1007 // fails to perform the allocation. b) is the only case when we'll
1008 1008 // return NULL.
1009 1009 HeapWord* result = NULL;
1010 1010 for (int try_count = 1; /* we'll return */; try_count += 1) {
1011 1011 bool should_try_gc;
1012 1012 unsigned int gc_count_before;
1013 1013
1014 1014 {
1015 1015 MutexLockerEx x(Heap_lock);
1016 1016
1017 1017 // Given that humongous objects are not allocated in young
1018 1018 // regions, we'll first try to do the allocation without doing a
1019 1019 // collection hoping that there's enough space in the heap.
1020 1020 result = humongous_obj_allocate(word_size);
1021 1021 if (result != NULL) {
1022 1022 return result;
1023 1023 }
1024 1024
1025 1025 if (GC_locker::is_active_and_needs_gc()) {
1026 1026 should_try_gc = false;
1027 1027 } else {
1028 1028 // Read the GC count while still holding the Heap_lock.
1029 1029 gc_count_before = SharedHeap::heap()->total_collections();
1030 1030 should_try_gc = true;
1031 1031 }
1032 1032 }
1033 1033
1034 1034 if (should_try_gc) {
1035 1035 // If we failed to allocate the humongous object, we should try to
1036 1036 // do a collection pause (if we're allowed) in case it reclaims
1037 1037 // enough space for the allocation to succeed after the pause.
1038 1038
1039 1039 bool succeeded;
1040 1040 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1041 1041 if (result != NULL) {
1042 1042 assert(succeeded, "only way to get back a non-NULL result");
1043 1043 return result;
1044 1044 }
1045 1045
1046 1046 if (succeeded) {
1047 1047 // If we get here we successfully scheduled a collection which
1048 1048 // failed to allocate. No point in trying to allocate
1049 1049 // further. We'll just return NULL.
1050 1050 MutexLockerEx x(Heap_lock);
1051 1051 *gc_count_before_ret = SharedHeap::heap()->total_collections();
1052 1052 return NULL;
1053 1053 }
1054 1054 } else {
1055 1055 GC_locker::stall_until_clear();
1056 1056 }
1057 1057
1058 1058 // We can reach here if we were unsuccessul in scheduling a
1059 1059 // collection (because another thread beat us to it) or if we were
1060 1060 // stalled due to the GC locker. In either can we should retry the
1061 1061 // allocation attempt in case another thread successfully
1062 1062 // performed a collection and reclaimed enough space. Give a
1063 1063 // warning if we seem to be looping forever.
1064 1064
1065 1065 if ((QueuedAllocationWarningCount > 0) &&
1066 1066 (try_count % QueuedAllocationWarningCount == 0)) {
1067 1067 warning("G1CollectedHeap::attempt_allocation_humongous() "
1068 1068 "retries %d times", try_count);
1069 1069 }
1070 1070 }
1071 1071
1072 1072 ShouldNotReachHere();
1073 1073 return NULL;
1074 1074 }
1075 1075
1076 1076 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1077 1077 bool expect_null_mutator_alloc_region) {
1078 1078 assert_at_safepoint(true /* should_be_vm_thread */);
1079 1079 assert(_mutator_alloc_region.get() == NULL ||
1080 1080 !expect_null_mutator_alloc_region,
1081 1081 "the current alloc region was unexpectedly found to be non-NULL");
1082 1082
1083 1083 if (!isHumongous(word_size)) {
1084 1084 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1085 1085 false /* bot_updates */);
1086 1086 } else {
1087 1087 return humongous_obj_allocate(word_size);
1088 1088 }
1089 1089
1090 1090 ShouldNotReachHere();
1091 1091 }
1092 1092
1093 1093 class PostMCRemSetClearClosure: public HeapRegionClosure {
1094 1094 ModRefBarrierSet* _mr_bs;
1095 1095 public:
1096 1096 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1097 1097 bool doHeapRegion(HeapRegion* r) {
1098 1098 r->reset_gc_time_stamp();
1099 1099 if (r->continuesHumongous())
1100 1100 return false;
1101 1101 HeapRegionRemSet* hrrs = r->rem_set();
1102 1102 if (hrrs != NULL) hrrs->clear();
1103 1103 // You might think here that we could clear just the cards
1104 1104 // corresponding to the used region. But no: if we leave a dirty card
1105 1105 // in a region we might allocate into, then it would prevent that card
1106 1106 // from being enqueued, and cause it to be missed.
1107 1107 // Re: the performance cost: we shouldn't be doing full GC anyway!
1108 1108 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1109 1109 return false;
1110 1110 }
1111 1111 };
1112 1112
1113 1113
1114 1114 class PostMCRemSetInvalidateClosure: public HeapRegionClosure {
1115 1115 ModRefBarrierSet* _mr_bs;
1116 1116 public:
1117 1117 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1118 1118 bool doHeapRegion(HeapRegion* r) {
1119 1119 if (r->continuesHumongous()) return false;
1120 1120 if (r->used_region().word_size() != 0) {
1121 1121 _mr_bs->invalidate(r->used_region(), true /*whole heap*/);
1122 1122 }
1123 1123 return false;
1124 1124 }
1125 1125 };
1126 1126
1127 1127 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1128 1128 G1CollectedHeap* _g1h;
1129 1129 UpdateRSOopClosure _cl;
1130 1130 int _worker_i;
1131 1131 public:
1132 1132 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1133 1133 _cl(g1->g1_rem_set(), worker_i),
1134 1134 _worker_i(worker_i),
1135 1135 _g1h(g1)
1136 1136 { }
1137 1137
1138 1138 bool doHeapRegion(HeapRegion* r) {
1139 1139 if (!r->continuesHumongous()) {
1140 1140 _cl.set_from(r);
1141 1141 r->oop_iterate(&_cl);
1142 1142 }
1143 1143 return false;
1144 1144 }
1145 1145 };
1146 1146
1147 1147 class ParRebuildRSTask: public AbstractGangTask {
1148 1148 G1CollectedHeap* _g1;
1149 1149 public:
1150 1150 ParRebuildRSTask(G1CollectedHeap* g1)
1151 1151 : AbstractGangTask("ParRebuildRSTask"),
1152 1152 _g1(g1)
1153 1153 { }
1154 1154
1155 1155 void work(int i) {
1156 1156 RebuildRSOutOfRegionClosure rebuild_rs(_g1, i);
1157 1157 _g1->heap_region_par_iterate_chunked(&rebuild_rs, i,
1158 1158 HeapRegion::RebuildRSClaimValue);
1159 1159 }
1160 1160 };
1161 1161
1162 1162 class PostCompactionPrinterClosure: public HeapRegionClosure {
1163 1163 private:
1164 1164 G1HRPrinter* _hr_printer;
1165 1165 public:
1166 1166 bool doHeapRegion(HeapRegion* hr) {
1167 1167 assert(!hr->is_young(), "not expecting to find young regions");
1168 1168 // We only generate output for non-empty regions.
1169 1169 if (!hr->is_empty()) {
1170 1170 if (!hr->isHumongous()) {
1171 1171 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1172 1172 } else if (hr->startsHumongous()) {
1173 1173 if (hr->capacity() == (size_t) HeapRegion::GrainBytes) {
1174 1174 // single humongous region
1175 1175 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1176 1176 } else {
1177 1177 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1178 1178 }
1179 1179 } else {
1180 1180 assert(hr->continuesHumongous(), "only way to get here");
1181 1181 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1182 1182 }
1183 1183 }
1184 1184 return false;
1185 1185 }
1186 1186
1187 1187 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1188 1188 : _hr_printer(hr_printer) { }
1189 1189 };
1190 1190
1191 1191 bool G1CollectedHeap::do_collection(bool explicit_gc,
1192 1192 bool clear_all_soft_refs,
1193 1193 size_t word_size) {
1194 1194 assert_at_safepoint(true /* should_be_vm_thread */);
1195 1195
1196 1196 if (GC_locker::check_active_before_gc()) {
1197 1197 return false;
1198 1198 }
1199 1199
1200 1200 SvcGCMarker sgcm(SvcGCMarker::FULL);
1201 1201 ResourceMark rm;
1202 1202
1203 1203 if (PrintHeapAtGC) {
1204 1204 Universe::print_heap_before_gc();
1205 1205 }
1206 1206
1207 1207 verify_region_sets_optional();
1208 1208
1209 1209 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1210 1210 collector_policy()->should_clear_all_soft_refs();
1211 1211
1212 1212 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1213 1213
1214 1214 {
1215 1215 IsGCActiveMark x;
1216 1216
1217 1217 // Timing
1218 1218 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc);
1219 1219 assert(!system_gc || explicit_gc, "invariant");
1220 1220 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1221 1221 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1222 1222 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC",
1223 1223 PrintGC, true, gclog_or_tty);
1224 1224
1225 1225 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1226 1226 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1227 1227
1228 1228 double start = os::elapsedTime();
1229 1229 g1_policy()->record_full_collection_start();
1230 1230
1231 1231 wait_while_free_regions_coming();
1232 1232 append_secondary_free_list_if_not_empty_with_lock();
1233 1233
1234 1234 gc_prologue(true);
1235 1235 increment_total_collections(true /* full gc */);
1236 1236
1237 1237 size_t g1h_prev_used = used();
1238 1238 assert(used() == recalculate_used(), "Should be equal");
1239 1239
1240 1240 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
1241 1241 HandleMark hm; // Discard invalid handles created during verification
1242 1242 gclog_or_tty->print(" VerifyBeforeGC:");
1243 1243 prepare_for_verify();
1244 1244 Universe::verify(/* allow dirty */ true,
1245 1245 /* silent */ false,
1246 1246 /* option */ VerifyOption_G1UsePrevMarking);
1247 1247
1248 1248 }
1249 1249 pre_full_gc_dump();
1250 1250
1251 1251 COMPILER2_PRESENT(DerivedPointerTable::clear());
1252 1252
1253 1253 // Disable discovery and empty the discovered lists
1254 1254 // for the CM ref processor.
1255 1255 ref_processor_cm()->disable_discovery();
1256 1256 ref_processor_cm()->abandon_partial_discovery();
1257 1257 ref_processor_cm()->verify_no_references_recorded();
1258 1258
1259 1259 // Abandon current iterations of concurrent marking and concurrent
1260 1260 // refinement, if any are in progress.
1261 1261 concurrent_mark()->abort();
1262 1262
1263 1263 // Make sure we'll choose a new allocation region afterwards.
1264 1264 release_mutator_alloc_region();
1265 1265 abandon_gc_alloc_regions();
1266 1266 g1_rem_set()->cleanupHRRS();
1267 1267 tear_down_region_lists();
1268 1268
1269 1269 // We should call this after we retire any currently active alloc
1270 1270 // regions so that all the ALLOC / RETIRE events are generated
1271 1271 // before the start GC event.
1272 1272 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1273 1273
1274 1274 // We may have added regions to the current incremental collection
1275 1275 // set between the last GC or pause and now. We need to clear the
1276 1276 // incremental collection set and then start rebuilding it afresh
1277 1277 // after this full GC.
1278 1278 abandon_collection_set(g1_policy()->inc_cset_head());
1279 1279 g1_policy()->clear_incremental_cset();
1280 1280 g1_policy()->stop_incremental_cset_building();
1281 1281
1282 1282 empty_young_list();
1283 1283 g1_policy()->set_full_young_gcs(true);
1284 1284
1285 1285 // See the comments in g1CollectedHeap.hpp and
1286 1286 // G1CollectedHeap::ref_processing_init() about
1287 1287 // how reference processing currently works in G1.
1288 1288
1289 1289 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1290 1290 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1291 1291
1292 1292 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1293 1293 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1294 1294
1295 1295 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1296 1296 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1297 1297
1298 1298 // Do collection work
1299 1299 {
1300 1300 HandleMark hm; // Discard invalid handles created during gc
1301 1301 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1302 1302 }
1303 1303
1304 1304 assert(free_regions() == 0, "we should not have added any free regions");
1305 1305 rebuild_region_lists();
1306 1306
1307 1307 _summary_bytes_used = recalculate_used();
1308 1308
1309 1309 // Enqueue any discovered reference objects that have
1310 1310 // not been removed from the discovered lists.
1311 1311 ref_processor_stw()->enqueue_discovered_references();
1312 1312
1313 1313 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1314 1314
1315 1315 MemoryService::track_memory_usage();
1316 1316
1317 1317 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
1318 1318 HandleMark hm; // Discard invalid handles created during verification
1319 1319 gclog_or_tty->print(" VerifyAfterGC:");
1320 1320 prepare_for_verify();
1321 1321 Universe::verify(/* allow dirty */ false,
1322 1322 /* silent */ false,
1323 1323 /* option */ VerifyOption_G1UsePrevMarking);
1324 1324
1325 1325 }
1326 1326
1327 1327 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1328 1328 ref_processor_stw()->verify_no_references_recorded();
1329 1329
1330 1330 // Note: since we've just done a full GC, concurrent
1331 1331 // marking is no longer active. Therefore we need not
1332 1332 // re-enable reference discovery for the CM ref processor.
1333 1333 // That will be done at the start of the next marking cycle.
1334 1334 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1335 1335 ref_processor_cm()->verify_no_references_recorded();
1336 1336
1337 1337 reset_gc_time_stamp();
1338 1338 // Since everything potentially moved, we will clear all remembered
1339 1339 // sets, and clear all cards. Later we will rebuild remebered
1340 1340 // sets. We will also reset the GC time stamps of the regions.
1341 1341 PostMCRemSetClearClosure rs_clear(mr_bs());
1342 1342 heap_region_iterate(&rs_clear);
1343 1343
1344 1344 // Resize the heap if necessary.
1345 1345 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1346 1346
1347 1347 if (_hr_printer.is_active()) {
1348 1348 // We should do this after we potentially resize the heap so
1349 1349 // that all the COMMIT / UNCOMMIT events are generated before
1350 1350 // the end GC event.
1351 1351
1352 1352 PostCompactionPrinterClosure cl(hr_printer());
1353 1353 heap_region_iterate(&cl);
1354 1354
1355 1355 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1356 1356 }
1357 1357
1358 1358 if (_cg1r->use_cache()) {
1359 1359 _cg1r->clear_and_record_card_counts();
1360 1360 _cg1r->clear_hot_cache();
1361 1361 }
1362 1362
1363 1363 // Rebuild remembered sets of all regions.
1364 1364
1365 1365 if (G1CollectedHeap::use_parallel_gc_threads()) {
1366 1366 ParRebuildRSTask rebuild_rs_task(this);
1367 1367 assert(check_heap_region_claim_values(
1368 1368 HeapRegion::InitialClaimValue), "sanity check");
1369 1369 set_par_threads(workers()->total_workers());
1370 1370 workers()->run_task(&rebuild_rs_task);
1371 1371 set_par_threads(0);
1372 1372 assert(check_heap_region_claim_values(
1373 1373 HeapRegion::RebuildRSClaimValue), "sanity check");
1374 1374 reset_heap_region_claim_values();
1375 1375 } else {
1376 1376 RebuildRSOutOfRegionClosure rebuild_rs(this);
1377 1377 heap_region_iterate(&rebuild_rs);
1378 1378 }
1379 1379
1380 1380 if (PrintGC) {
1381 1381 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1382 1382 }
1383 1383
1384 1384 if (true) { // FIXME
1385 1385 // Ask the permanent generation to adjust size for full collections
1386 1386 perm()->compute_new_size();
1387 1387 }
1388 1388
1389 1389 // Start a new incremental collection set for the next pause
1390 1390 assert(g1_policy()->collection_set() == NULL, "must be");
1391 1391 g1_policy()->start_incremental_cset_building();
1392 1392
1393 1393 // Clear the _cset_fast_test bitmap in anticipation of adding
1394 1394 // regions to the incremental collection set for the next
1395 1395 // evacuation pause.
1396 1396 clear_cset_fast_test();
1397 1397
1398 1398 init_mutator_alloc_region();
1399 1399
1400 1400 double end = os::elapsedTime();
1401 1401 g1_policy()->record_full_collection_end();
1402 1402
1403 1403 #ifdef TRACESPINNING
1404 1404 ParallelTaskTerminator::print_termination_counts();
1405 1405 #endif
1406 1406
1407 1407 gc_epilogue(true);
1408 1408
1409 1409 // Discard all rset updates
1410 1410 JavaThread::dirty_card_queue_set().abandon_logs();
1411 1411 assert(!G1DeferredRSUpdate
1412 1412 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1413 1413 }
1414 1414
1415 1415 _young_list->reset_sampled_info();
1416 1416 // At this point there should be no regions in the
1417 1417 // entire heap tagged as young.
1418 1418 assert( check_young_list_empty(true /* check_heap */),
1419 1419 "young list should be empty at this point");
1420 1420
1421 1421 // Update the number of full collections that have been completed.
1422 1422 increment_full_collections_completed(false /* concurrent */);
1423 1423
1424 1424 _hrs.verify_optional();
1425 1425 verify_region_sets_optional();
1426 1426
1427 1427 if (PrintHeapAtGC) {
1428 1428 Universe::print_heap_after_gc();
1429 1429 }
1430 1430 g1mm()->update_sizes();
1431 1431 post_full_gc_dump();
1432 1432
1433 1433 return true;
1434 1434 }
1435 1435
1436 1436 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1437 1437 // do_collection() will return whether it succeeded in performing
1438 1438 // the GC. Currently, there is no facility on the
1439 1439 // do_full_collection() API to notify the caller than the collection
1440 1440 // did not succeed (e.g., because it was locked out by the GC
1441 1441 // locker). So, right now, we'll ignore the return value.
1442 1442 bool dummy = do_collection(true, /* explicit_gc */
1443 1443 clear_all_soft_refs,
1444 1444 0 /* word_size */);
1445 1445 }
1446 1446
1447 1447 // This code is mostly copied from TenuredGeneration.
1448 1448 void
1449 1449 G1CollectedHeap::
1450 1450 resize_if_necessary_after_full_collection(size_t word_size) {
1451 1451 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1452 1452
1453 1453 // Include the current allocation, if any, and bytes that will be
1454 1454 // pre-allocated to support collections, as "used".
1455 1455 const size_t used_after_gc = used();
1456 1456 const size_t capacity_after_gc = capacity();
1457 1457 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1458 1458
1459 1459 // This is enforced in arguments.cpp.
1460 1460 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1461 1461 "otherwise the code below doesn't make sense");
1462 1462
1463 1463 // We don't have floating point command-line arguments
1464 1464 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1465 1465 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1466 1466 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1467 1467 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1468 1468
1469 1469 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1470 1470 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1471 1471
1472 1472 // We have to be careful here as these two calculations can overflow
1473 1473 // 32-bit size_t's.
1474 1474 double used_after_gc_d = (double) used_after_gc;
1475 1475 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1476 1476 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1477 1477
1478 1478 // Let's make sure that they are both under the max heap size, which
1479 1479 // by default will make them fit into a size_t.
1480 1480 double desired_capacity_upper_bound = (double) max_heap_size;
1481 1481 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1482 1482 desired_capacity_upper_bound);
1483 1483 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1484 1484 desired_capacity_upper_bound);
1485 1485
1486 1486 // We can now safely turn them into size_t's.
1487 1487 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1488 1488 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1489 1489
1490 1490 // This assert only makes sense here, before we adjust them
1491 1491 // with respect to the min and max heap size.
1492 1492 assert(minimum_desired_capacity <= maximum_desired_capacity,
1493 1493 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1494 1494 "maximum_desired_capacity = "SIZE_FORMAT,
1495 1495 minimum_desired_capacity, maximum_desired_capacity));
1496 1496
1497 1497 // Should not be greater than the heap max size. No need to adjust
1498 1498 // it with respect to the heap min size as it's a lower bound (i.e.,
1499 1499 // we'll try to make the capacity larger than it, not smaller).
1500 1500 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1501 1501 // Should not be less than the heap min size. No need to adjust it
1502 1502 // with respect to the heap max size as it's an upper bound (i.e.,
1503 1503 // we'll try to make the capacity smaller than it, not greater).
1504 1504 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1505 1505
1506 1506 if (capacity_after_gc < minimum_desired_capacity) {
1507 1507 // Don't expand unless it's significant
1508 1508 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1509 1509 ergo_verbose4(ErgoHeapSizing,
1510 1510 "attempt heap expansion",
1511 1511 ergo_format_reason("capacity lower than "
1512 1512 "min desired capacity after Full GC")
1513 1513 ergo_format_byte("capacity")
1514 1514 ergo_format_byte("occupancy")
1515 1515 ergo_format_byte_perc("min desired capacity"),
1516 1516 capacity_after_gc, used_after_gc,
1517 1517 minimum_desired_capacity, (double) MinHeapFreeRatio);
1518 1518 expand(expand_bytes);
1519 1519
1520 1520 // No expansion, now see if we want to shrink
1521 1521 } else if (capacity_after_gc > maximum_desired_capacity) {
1522 1522 // Capacity too large, compute shrinking size
1523 1523 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1524 1524 ergo_verbose4(ErgoHeapSizing,
1525 1525 "attempt heap shrinking",
1526 1526 ergo_format_reason("capacity higher than "
1527 1527 "max desired capacity after Full GC")
1528 1528 ergo_format_byte("capacity")
1529 1529 ergo_format_byte("occupancy")
1530 1530 ergo_format_byte_perc("max desired capacity"),
1531 1531 capacity_after_gc, used_after_gc,
1532 1532 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1533 1533 shrink(shrink_bytes);
1534 1534 }
1535 1535 }
1536 1536
1537 1537
1538 1538 HeapWord*
1539 1539 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1540 1540 bool* succeeded) {
1541 1541 assert_at_safepoint(true /* should_be_vm_thread */);
1542 1542
1543 1543 *succeeded = true;
1544 1544 // Let's attempt the allocation first.
1545 1545 HeapWord* result =
1546 1546 attempt_allocation_at_safepoint(word_size,
1547 1547 false /* expect_null_mutator_alloc_region */);
1548 1548 if (result != NULL) {
1549 1549 assert(*succeeded, "sanity");
1550 1550 return result;
1551 1551 }
1552 1552
1553 1553 // In a G1 heap, we're supposed to keep allocation from failing by
1554 1554 // incremental pauses. Therefore, at least for now, we'll favor
1555 1555 // expansion over collection. (This might change in the future if we can
1556 1556 // do something smarter than full collection to satisfy a failed alloc.)
1557 1557 result = expand_and_allocate(word_size);
1558 1558 if (result != NULL) {
1559 1559 assert(*succeeded, "sanity");
1560 1560 return result;
1561 1561 }
1562 1562
1563 1563 // Expansion didn't work, we'll try to do a Full GC.
1564 1564 bool gc_succeeded = do_collection(false, /* explicit_gc */
1565 1565 false, /* clear_all_soft_refs */
1566 1566 word_size);
1567 1567 if (!gc_succeeded) {
1568 1568 *succeeded = false;
1569 1569 return NULL;
1570 1570 }
1571 1571
1572 1572 // Retry the allocation
1573 1573 result = attempt_allocation_at_safepoint(word_size,
1574 1574 true /* expect_null_mutator_alloc_region */);
1575 1575 if (result != NULL) {
1576 1576 assert(*succeeded, "sanity");
1577 1577 return result;
1578 1578 }
1579 1579
1580 1580 // Then, try a Full GC that will collect all soft references.
1581 1581 gc_succeeded = do_collection(false, /* explicit_gc */
1582 1582 true, /* clear_all_soft_refs */
1583 1583 word_size);
1584 1584 if (!gc_succeeded) {
1585 1585 *succeeded = false;
1586 1586 return NULL;
1587 1587 }
1588 1588
1589 1589 // Retry the allocation once more
1590 1590 result = attempt_allocation_at_safepoint(word_size,
1591 1591 true /* expect_null_mutator_alloc_region */);
1592 1592 if (result != NULL) {
1593 1593 assert(*succeeded, "sanity");
1594 1594 return result;
1595 1595 }
1596 1596
1597 1597 assert(!collector_policy()->should_clear_all_soft_refs(),
1598 1598 "Flag should have been handled and cleared prior to this point");
1599 1599
1600 1600 // What else? We might try synchronous finalization later. If the total
1601 1601 // space available is large enough for the allocation, then a more
1602 1602 // complete compaction phase than we've tried so far might be
1603 1603 // appropriate.
1604 1604 assert(*succeeded, "sanity");
1605 1605 return NULL;
1606 1606 }
1607 1607
1608 1608 // Attempting to expand the heap sufficiently
1609 1609 // to support an allocation of the given "word_size". If
1610 1610 // successful, perform the allocation and return the address of the
1611 1611 // allocated block, or else "NULL".
1612 1612
1613 1613 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1614 1614 assert_at_safepoint(true /* should_be_vm_thread */);
1615 1615
1616 1616 verify_region_sets_optional();
1617 1617
1618 1618 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1619 1619 ergo_verbose1(ErgoHeapSizing,
1620 1620 "attempt heap expansion",
1621 1621 ergo_format_reason("allocation request failed")
1622 1622 ergo_format_byte("allocation request"),
1623 1623 word_size * HeapWordSize);
1624 1624 if (expand(expand_bytes)) {
1625 1625 _hrs.verify_optional();
1626 1626 verify_region_sets_optional();
1627 1627 return attempt_allocation_at_safepoint(word_size,
1628 1628 false /* expect_null_mutator_alloc_region */);
1629 1629 }
1630 1630 return NULL;
1631 1631 }
1632 1632
1633 1633 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1634 1634 HeapWord* new_end) {
1635 1635 assert(old_end != new_end, "don't call this otherwise");
1636 1636 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1637 1637
1638 1638 // Update the committed mem region.
1639 1639 _g1_committed.set_end(new_end);
1640 1640 // Tell the card table about the update.
1641 1641 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1642 1642 // Tell the BOT about the update.
1643 1643 _bot_shared->resize(_g1_committed.word_size());
1644 1644 }
1645 1645
1646 1646 bool G1CollectedHeap::expand(size_t expand_bytes) {
1647 1647 size_t old_mem_size = _g1_storage.committed_size();
1648 1648 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1649 1649 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1650 1650 HeapRegion::GrainBytes);
1651 1651 ergo_verbose2(ErgoHeapSizing,
1652 1652 "expand the heap",
1653 1653 ergo_format_byte("requested expansion amount")
1654 1654 ergo_format_byte("attempted expansion amount"),
1655 1655 expand_bytes, aligned_expand_bytes);
1656 1656
1657 1657 // First commit the memory.
1658 1658 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1659 1659 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1660 1660 if (successful) {
1661 1661 // Then propagate this update to the necessary data structures.
1662 1662 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1663 1663 update_committed_space(old_end, new_end);
1664 1664
1665 1665 FreeRegionList expansion_list("Local Expansion List");
1666 1666 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1667 1667 assert(mr.start() == old_end, "post-condition");
1668 1668 // mr might be a smaller region than what was requested if
1669 1669 // expand_by() was unable to allocate the HeapRegion instances
1670 1670 assert(mr.end() <= new_end, "post-condition");
1671 1671
1672 1672 size_t actual_expand_bytes = mr.byte_size();
1673 1673 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1674 1674 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1675 1675 "post-condition");
1676 1676 if (actual_expand_bytes < aligned_expand_bytes) {
1677 1677 // We could not expand _hrs to the desired size. In this case we
1678 1678 // need to shrink the committed space accordingly.
1679 1679 assert(mr.end() < new_end, "invariant");
1680 1680
1681 1681 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1682 1682 // First uncommit the memory.
1683 1683 _g1_storage.shrink_by(diff_bytes);
1684 1684 // Then propagate this update to the necessary data structures.
1685 1685 update_committed_space(new_end, mr.end());
1686 1686 }
1687 1687 _free_list.add_as_tail(&expansion_list);
1688 1688
1689 1689 if (_hr_printer.is_active()) {
1690 1690 HeapWord* curr = mr.start();
1691 1691 while (curr < mr.end()) {
1692 1692 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1693 1693 _hr_printer.commit(curr, curr_end);
1694 1694 curr = curr_end;
1695 1695 }
1696 1696 assert(curr == mr.end(), "post-condition");
1697 1697 }
1698 1698 g1_policy()->record_new_heap_size(n_regions());
1699 1699 } else {
1700 1700 ergo_verbose0(ErgoHeapSizing,
1701 1701 "did not expand the heap",
1702 1702 ergo_format_reason("heap expansion operation failed"));
1703 1703 // The expansion of the virtual storage space was unsuccessful.
1704 1704 // Let's see if it was because we ran out of swap.
1705 1705 if (G1ExitOnExpansionFailure &&
1706 1706 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1707 1707 // We had head room...
1708 1708 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1709 1709 }
1710 1710 }
1711 1711 return successful;
1712 1712 }
1713 1713
1714 1714 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1715 1715 size_t old_mem_size = _g1_storage.committed_size();
1716 1716 size_t aligned_shrink_bytes =
1717 1717 ReservedSpace::page_align_size_down(shrink_bytes);
1718 1718 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1719 1719 HeapRegion::GrainBytes);
1720 1720 size_t num_regions_deleted = 0;
1721 1721 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1722 1722 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1723 1723 assert(mr.end() == old_end, "post-condition");
1724 1724
1725 1725 ergo_verbose3(ErgoHeapSizing,
1726 1726 "shrink the heap",
1727 1727 ergo_format_byte("requested shrinking amount")
1728 1728 ergo_format_byte("aligned shrinking amount")
1729 1729 ergo_format_byte("attempted shrinking amount"),
1730 1730 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1731 1731 if (mr.byte_size() > 0) {
1732 1732 if (_hr_printer.is_active()) {
1733 1733 HeapWord* curr = mr.end();
1734 1734 while (curr > mr.start()) {
1735 1735 HeapWord* curr_end = curr;
1736 1736 curr -= HeapRegion::GrainWords;
1737 1737 _hr_printer.uncommit(curr, curr_end);
1738 1738 }
1739 1739 assert(curr == mr.start(), "post-condition");
1740 1740 }
1741 1741
1742 1742 _g1_storage.shrink_by(mr.byte_size());
1743 1743 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1744 1744 assert(mr.start() == new_end, "post-condition");
1745 1745
1746 1746 _expansion_regions += num_regions_deleted;
1747 1747 update_committed_space(old_end, new_end);
1748 1748 HeapRegionRemSet::shrink_heap(n_regions());
1749 1749 g1_policy()->record_new_heap_size(n_regions());
1750 1750 } else {
1751 1751 ergo_verbose0(ErgoHeapSizing,
1752 1752 "did not shrink the heap",
1753 1753 ergo_format_reason("heap shrinking operation failed"));
1754 1754 }
1755 1755 }
1756 1756
1757 1757 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1758 1758 verify_region_sets_optional();
1759 1759
1760 1760 // We should only reach here at the end of a Full GC which means we
1761 1761 // should not not be holding to any GC alloc regions. The method
1762 1762 // below will make sure of that and do any remaining clean up.
1763 1763 abandon_gc_alloc_regions();
1764 1764
1765 1765 // Instead of tearing down / rebuilding the free lists here, we
1766 1766 // could instead use the remove_all_pending() method on free_list to
1767 1767 // remove only the ones that we need to remove.
1768 1768 tear_down_region_lists(); // We will rebuild them in a moment.
1769 1769 shrink_helper(shrink_bytes);
1770 1770 rebuild_region_lists();
1771 1771
1772 1772 _hrs.verify_optional();
1773 1773 verify_region_sets_optional();
1774 1774 }
1775 1775
1776 1776 // Public methods.
1777 1777
1778 1778 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1779 1779 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1780 1780 #endif // _MSC_VER
1781 1781
1782 1782
1783 1783 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1784 1784 SharedHeap(policy_),
1785 1785 _g1_policy(policy_),
1786 1786 _dirty_card_queue_set(false),
1787 1787 _into_cset_dirty_card_queue_set(false),
1788 1788 _is_alive_closure_cm(this),
1789 1789 _is_alive_closure_stw(this),
1790 1790 _ref_processor_cm(NULL),
1791 1791 _ref_processor_stw(NULL),
1792 1792 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1793 1793 _bot_shared(NULL),
1794 1794 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1795 1795 _evac_failure_scan_stack(NULL) ,
1796 1796 _mark_in_progress(false),
1797 1797 _cg1r(NULL), _summary_bytes_used(0),
1798 1798 _g1mm(NULL),
1799 1799 _refine_cte_cl(NULL),
1800 1800 _full_collection(false),
1801 1801 _free_list("Master Free List"),
1802 1802 _secondary_free_list("Secondary Free List"),
1803 1803 _humongous_set("Master Humongous Set"),
1804 1804 _free_regions_coming(false),
1805 1805 _young_list(new YoungList(this)),
1806 1806 _gc_time_stamp(0),
1807 1807 _retained_old_gc_alloc_region(NULL),
1808 1808 _surviving_young_words(NULL),
1809 1809 _full_collections_completed(0),
1810 1810 _in_cset_fast_test(NULL),
1811 1811 _in_cset_fast_test_base(NULL),
1812 1812 _dirty_cards_region_list(NULL) {
1813 1813 _g1h = this; // To catch bugs.
1814 1814 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1815 1815 vm_exit_during_initialization("Failed necessary allocation.");
1816 1816 }
1817 1817
1818 1818 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1819 1819
1820 1820 int n_queues = MAX2((int)ParallelGCThreads, 1);
1821 1821 _task_queues = new RefToScanQueueSet(n_queues);
1822 1822
1823 1823 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1824 1824 assert(n_rem_sets > 0, "Invariant.");
1825 1825
1826 1826 HeapRegionRemSetIterator** iter_arr =
1827 1827 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues);
1828 1828 for (int i = 0; i < n_queues; i++) {
1829 1829 iter_arr[i] = new HeapRegionRemSetIterator();
1830 1830 }
1831 1831 _rem_set_iterator = iter_arr;
1832 1832
1833 1833 for (int i = 0; i < n_queues; i++) {
1834 1834 RefToScanQueue* q = new RefToScanQueue();
1835 1835 q->initialize();
1836 1836 _task_queues->register_queue(i, q);
1837 1837 }
1838 1838
1839 1839 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1840 1840 }
1841 1841
1842 1842 jint G1CollectedHeap::initialize() {
1843 1843 CollectedHeap::pre_initialize();
1844 1844 os::enable_vtime();
1845 1845
1846 1846 // Necessary to satisfy locking discipline assertions.
1847 1847
1848 1848 MutexLocker x(Heap_lock);
1849 1849
1850 1850 // We have to initialize the printer before committing the heap, as
1851 1851 // it will be used then.
1852 1852 _hr_printer.set_active(G1PrintHeapRegions);
1853 1853
1854 1854 // While there are no constraints in the GC code that HeapWordSize
1855 1855 // be any particular value, there are multiple other areas in the
1856 1856 // system which believe this to be true (e.g. oop->object_size in some
1857 1857 // cases incorrectly returns the size in wordSize units rather than
1858 1858 // HeapWordSize).
1859 1859 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1860 1860
1861 1861 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1862 1862 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1863 1863
1864 1864 // Ensure that the sizes are properly aligned.
1865 1865 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1866 1866 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1867 1867
1868 1868 _cg1r = new ConcurrentG1Refine();
1869 1869
1870 1870 // Reserve the maximum.
1871 1871 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation();
1872 1872 // Includes the perm-gen.
1873 1873
1874 1874 // When compressed oops are enabled, the preferred heap base
1875 1875 // is calculated by subtracting the requested size from the
1876 1876 // 32Gb boundary and using the result as the base address for
1877 1877 // heap reservation. If the requested size is not aligned to
1878 1878 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1879 1879 // into the ReservedHeapSpace constructor) then the actual
1880 1880 // base of the reserved heap may end up differing from the
1881 1881 // address that was requested (i.e. the preferred heap base).
1882 1882 // If this happens then we could end up using a non-optimal
1883 1883 // compressed oops mode.
1884 1884
1885 1885 // Since max_byte_size is aligned to the size of a heap region (checked
1886 1886 // above), we also need to align the perm gen size as it might not be.
1887 1887 const size_t total_reserved = max_byte_size +
1888 1888 align_size_up(pgs->max_size(), HeapRegion::GrainBytes);
1889 1889 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm");
1890 1890
1891 1891 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop);
1892 1892
1893 1893 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes,
1894 1894 UseLargePages, addr);
1895 1895
1896 1896 if (UseCompressedOops) {
1897 1897 if (addr != NULL && !heap_rs.is_reserved()) {
1898 1898 // Failed to reserve at specified address - the requested memory
1899 1899 // region is taken already, for example, by 'java' launcher.
1900 1900 // Try again to reserver heap higher.
1901 1901 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop);
1902 1902
1903 1903 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes,
1904 1904 UseLargePages, addr);
1905 1905
1906 1906 if (addr != NULL && !heap_rs0.is_reserved()) {
1907 1907 // Failed to reserve at specified address again - give up.
1908 1908 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop);
1909 1909 assert(addr == NULL, "");
1910 1910
1911 1911 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes,
1912 1912 UseLargePages, addr);
1913 1913 heap_rs = heap_rs1;
1914 1914 } else {
1915 1915 heap_rs = heap_rs0;
1916 1916 }
1917 1917 }
1918 1918 }
1919 1919
1920 1920 if (!heap_rs.is_reserved()) {
1921 1921 vm_exit_during_initialization("Could not reserve enough space for object heap");
1922 1922 return JNI_ENOMEM;
1923 1923 }
1924 1924
1925 1925 // It is important to do this in a way such that concurrent readers can't
1926 1926 // temporarily think somethings in the heap. (I've actually seen this
1927 1927 // happen in asserts: DLD.)
1928 1928 _reserved.set_word_size(0);
1929 1929 _reserved.set_start((HeapWord*)heap_rs.base());
1930 1930 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
1931 1931
1932 1932 _expansion_regions = max_byte_size/HeapRegion::GrainBytes;
1933 1933
1934 1934 // Create the gen rem set (and barrier set) for the entire reserved region.
1935 1935 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
1936 1936 set_barrier_set(rem_set()->bs());
1937 1937 if (barrier_set()->is_a(BarrierSet::ModRef)) {
1938 1938 _mr_bs = (ModRefBarrierSet*)_barrier_set;
1939 1939 } else {
1940 1940 vm_exit_during_initialization("G1 requires a mod ref bs.");
1941 1941 return JNI_ENOMEM;
1942 1942 }
1943 1943
1944 1944 // Also create a G1 rem set.
1945 1945 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
1946 1946 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
1947 1947 } else {
1948 1948 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
1949 1949 return JNI_ENOMEM;
1950 1950 }
1951 1951
1952 1952 // Carve out the G1 part of the heap.
1953 1953
1954 1954 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
1955 1955 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
1956 1956 g1_rs.size()/HeapWordSize);
1957 1957 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size);
1958 1958
1959 1959 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set());
1960 1960
1961 1961 _g1_storage.initialize(g1_rs, 0);
1962 1962 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
1963 1963 _hrs.initialize((HeapWord*) _g1_reserved.start(),
1964 1964 (HeapWord*) _g1_reserved.end(),
1965 1965 _expansion_regions);
1966 1966
1967 1967 // 6843694 - ensure that the maximum region index can fit
1968 1968 // in the remembered set structures.
1969 1969 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
1970 1970 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
1971 1971
1972 1972 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
1973 1973 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
1974 1974 guarantee((size_t) HeapRegion::CardsPerRegion < max_cards_per_region,
1975 1975 "too many cards per region");
1976 1976
1977 1977 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
1978 1978
1979 1979 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
1980 1980 heap_word_size(init_byte_size));
1981 1981
1982 1982 _g1h = this;
1983 1983
1984 1984 _in_cset_fast_test_length = max_regions();
1985 1985 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length);
1986 1986
1987 1987 // We're biasing _in_cset_fast_test to avoid subtracting the
1988 1988 // beginning of the heap every time we want to index; basically
1989 1989 // it's the same with what we do with the card table.
1990 1990 _in_cset_fast_test = _in_cset_fast_test_base -
1991 1991 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
1992 1992
1993 1993 // Clear the _cset_fast_test bitmap in anticipation of adding
1994 1994 // regions to the incremental collection set for the first
1995 1995 // evacuation pause.
1996 1996 clear_cset_fast_test();
1997 1997
1998 1998 // Create the ConcurrentMark data structure and thread.
1999 1999 // (Must do this late, so that "max_regions" is defined.)
2000 2000 _cm = new ConcurrentMark(heap_rs, (int) max_regions());
2001 2001 _cmThread = _cm->cmThread();
2002 2002
2003 2003 // Initialize the from_card cache structure of HeapRegionRemSet.
2004 2004 HeapRegionRemSet::init_heap(max_regions());
2005 2005
2006 2006 // Now expand into the initial heap size.
2007 2007 if (!expand(init_byte_size)) {
2008 2008 vm_exit_during_initialization("Failed to allocate initial heap.");
2009 2009 return JNI_ENOMEM;
2010 2010 }
2011 2011
2012 2012 // Perform any initialization actions delegated to the policy.
2013 2013 g1_policy()->init();
2014 2014
2015 2015 g1_policy()->note_start_of_mark_thread();
2016 2016
2017 2017 _refine_cte_cl =
2018 2018 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2019 2019 g1_rem_set(),
2020 2020 concurrent_g1_refine());
2021 2021 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2022 2022
2023 2023 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2024 2024 SATB_Q_FL_lock,
2025 2025 G1SATBProcessCompletedThreshold,
2026 2026 Shared_SATB_Q_lock);
2027 2027
2028 2028 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2029 2029 DirtyCardQ_FL_lock,
2030 2030 concurrent_g1_refine()->yellow_zone(),
2031 2031 concurrent_g1_refine()->red_zone(),
2032 2032 Shared_DirtyCardQ_lock);
2033 2033
2034 2034 if (G1DeferredRSUpdate) {
2035 2035 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2036 2036 DirtyCardQ_FL_lock,
2037 2037 -1, // never trigger processing
2038 2038 -1, // no limit on length
2039 2039 Shared_DirtyCardQ_lock,
2040 2040 &JavaThread::dirty_card_queue_set());
2041 2041 }
2042 2042
2043 2043 // Initialize the card queue set used to hold cards containing
2044 2044 // references into the collection set.
2045 2045 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2046 2046 DirtyCardQ_FL_lock,
2047 2047 -1, // never trigger processing
2048 2048 -1, // no limit on length
2049 2049 Shared_DirtyCardQ_lock,
2050 2050 &JavaThread::dirty_card_queue_set());
2051 2051
2052 2052 // In case we're keeping closure specialization stats, initialize those
2053 2053 // counts and that mechanism.
2054 2054 SpecializationStats::clear();
2055 2055
2056 2056 // Do later initialization work for concurrent refinement.
2057 2057 _cg1r->init();
2058 2058
2059 2059 // Here we allocate the dummy full region that is required by the
2060 2060 // G1AllocRegion class. If we don't pass an address in the reserved
2061 2061 // space here, lots of asserts fire.
2062 2062
2063 2063 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2064 2064 _g1_reserved.start());
2065 2065 // We'll re-use the same region whether the alloc region will
2066 2066 // require BOT updates or not and, if it doesn't, then a non-young
2067 2067 // region will complain that it cannot support allocations without
2068 2068 // BOT updates. So we'll tag the dummy region as young to avoid that.
2069 2069 dummy_region->set_young();
2070 2070 // Make sure it's full.
2071 2071 dummy_region->set_top(dummy_region->end());
2072 2072 G1AllocRegion::setup(this, dummy_region);
2073 2073
2074 2074 init_mutator_alloc_region();
2075 2075
2076 2076 // Do create of the monitoring and management support so that
2077 2077 // values in the heap have been properly initialized.
2078 2078 _g1mm = new G1MonitoringSupport(this);
2079 2079
2080 2080 return JNI_OK;
2081 2081 }
2082 2082
2083 2083 void G1CollectedHeap::ref_processing_init() {
2084 2084 // Reference processing in G1 currently works as follows:
2085 2085 //
2086 2086 // * There are two reference processor instances. One is
2087 2087 // used to record and process discovered references
2088 2088 // during concurrent marking; the other is used to
2089 2089 // record and process references during STW pauses
2090 2090 // (both full and incremental).
2091 2091 // * Both ref processors need to 'span' the entire heap as
2092 2092 // the regions in the collection set may be dotted around.
2093 2093 //
2094 2094 // * For the concurrent marking ref processor:
2095 2095 // * Reference discovery is enabled at initial marking.
2096 2096 // * Reference discovery is disabled and the discovered
2097 2097 // references processed etc during remarking.
2098 2098 // * Reference discovery is MT (see below).
2099 2099 // * Reference discovery requires a barrier (see below).
2100 2100 // * Reference processing may or may not be MT
2101 2101 // (depending on the value of ParallelRefProcEnabled
2102 2102 // and ParallelGCThreads).
2103 2103 // * A full GC disables reference discovery by the CM
2104 2104 // ref processor and abandons any entries on it's
2105 2105 // discovered lists.
2106 2106 //
2107 2107 // * For the STW processor:
2108 2108 // * Non MT discovery is enabled at the start of a full GC.
2109 2109 // * Processing and enqueueing during a full GC is non-MT.
2110 2110 // * During a full GC, references are processed after marking.
2111 2111 //
2112 2112 // * Discovery (may or may not be MT) is enabled at the start
2113 2113 // of an incremental evacuation pause.
2114 2114 // * References are processed near the end of a STW evacuation pause.
2115 2115 // * For both types of GC:
2116 2116 // * Discovery is atomic - i.e. not concurrent.
2117 2117 // * Reference discovery will not need a barrier.
2118 2118
2119 2119 SharedHeap::ref_processing_init();
2120 2120 MemRegion mr = reserved_region();
2121 2121
2122 2122 // Concurrent Mark ref processor
2123 2123 _ref_processor_cm =
2124 2124 new ReferenceProcessor(mr, // span
2125 2125 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2126 2126 // mt processing
2127 2127 (int) ParallelGCThreads,
2128 2128 // degree of mt processing
2129 2129 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2130 2130 // mt discovery
2131 2131 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2132 2132 // degree of mt discovery
2133 2133 false,
2134 2134 // Reference discovery is not atomic
2135 2135 &_is_alive_closure_cm,
2136 2136 // is alive closure
2137 2137 // (for efficiency/performance)
2138 2138 true);
2139 2139 // Setting next fields of discovered
2140 2140 // lists requires a barrier.
2141 2141
2142 2142 // STW ref processor
2143 2143 _ref_processor_stw =
2144 2144 new ReferenceProcessor(mr, // span
2145 2145 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2146 2146 // mt processing
2147 2147 MAX2((int)ParallelGCThreads, 1),
2148 2148 // degree of mt processing
2149 2149 (ParallelGCThreads > 1),
2150 2150 // mt discovery
2151 2151 MAX2((int)ParallelGCThreads, 1),
2152 2152 // degree of mt discovery
2153 2153 true,
2154 2154 // Reference discovery is atomic
2155 2155 &_is_alive_closure_stw,
2156 2156 // is alive closure
2157 2157 // (for efficiency/performance)
2158 2158 false);
2159 2159 // Setting next fields of discovered
2160 2160 // lists requires a barrier.
2161 2161 }
2162 2162
2163 2163 size_t G1CollectedHeap::capacity() const {
2164 2164 return _g1_committed.byte_size();
2165 2165 }
2166 2166
2167 2167 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2168 2168 DirtyCardQueue* into_cset_dcq,
2169 2169 bool concurrent,
2170 2170 int worker_i) {
2171 2171 // Clean cards in the hot card cache
2172 2172 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2173 2173
2174 2174 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2175 2175 int n_completed_buffers = 0;
2176 2176 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2177 2177 n_completed_buffers++;
2178 2178 }
2179 2179 g1_policy()->record_update_rs_processed_buffers(worker_i,
2180 2180 (double) n_completed_buffers);
2181 2181 dcqs.clear_n_completed_buffers();
2182 2182 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2183 2183 }
2184 2184
2185 2185
2186 2186 // Computes the sum of the storage used by the various regions.
2187 2187
2188 2188 size_t G1CollectedHeap::used() const {
2189 2189 assert(Heap_lock->owner() != NULL,
2190 2190 "Should be owned on this thread's behalf.");
2191 2191 size_t result = _summary_bytes_used;
2192 2192 // Read only once in case it is set to NULL concurrently
2193 2193 HeapRegion* hr = _mutator_alloc_region.get();
2194 2194 if (hr != NULL)
2195 2195 result += hr->used();
2196 2196 return result;
2197 2197 }
2198 2198
2199 2199 size_t G1CollectedHeap::used_unlocked() const {
2200 2200 size_t result = _summary_bytes_used;
2201 2201 return result;
2202 2202 }
2203 2203
2204 2204 class SumUsedClosure: public HeapRegionClosure {
2205 2205 size_t _used;
2206 2206 public:
2207 2207 SumUsedClosure() : _used(0) {}
2208 2208 bool doHeapRegion(HeapRegion* r) {
2209 2209 if (!r->continuesHumongous()) {
2210 2210 _used += r->used();
2211 2211 }
2212 2212 return false;
2213 2213 }
2214 2214 size_t result() { return _used; }
2215 2215 };
2216 2216
2217 2217 size_t G1CollectedHeap::recalculate_used() const {
2218 2218 SumUsedClosure blk;
2219 2219 heap_region_iterate(&blk);
2220 2220 return blk.result();
2221 2221 }
2222 2222
2223 2223 size_t G1CollectedHeap::unsafe_max_alloc() {
2224 2224 if (free_regions() > 0) return HeapRegion::GrainBytes;
2225 2225 // otherwise, is there space in the current allocation region?
2226 2226
2227 2227 // We need to store the current allocation region in a local variable
2228 2228 // here. The problem is that this method doesn't take any locks and
2229 2229 // there may be other threads which overwrite the current allocation
2230 2230 // region field. attempt_allocation(), for example, sets it to NULL
2231 2231 // and this can happen *after* the NULL check here but before the call
2232 2232 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2233 2233 // to be a problem in the optimized build, since the two loads of the
2234 2234 // current allocation region field are optimized away.
2235 2235 HeapRegion* hr = _mutator_alloc_region.get();
2236 2236 if (hr == NULL) {
2237 2237 return 0;
2238 2238 }
2239 2239 return hr->free();
2240 2240 }
2241 2241
2242 2242 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2243 2243 return
2244 2244 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) ||
2245 2245 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent));
2246 2246 }
2247 2247
2248 2248 #ifndef PRODUCT
2249 2249 void G1CollectedHeap::allocate_dummy_regions() {
2250 2250 // Let's fill up most of the region
2251 2251 size_t word_size = HeapRegion::GrainWords - 1024;
2252 2252 // And as a result the region we'll allocate will be humongous.
2253 2253 guarantee(isHumongous(word_size), "sanity");
2254 2254
2255 2255 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2256 2256 // Let's use the existing mechanism for the allocation
2257 2257 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2258 2258 if (dummy_obj != NULL) {
2259 2259 MemRegion mr(dummy_obj, word_size);
2260 2260 CollectedHeap::fill_with_object(mr);
2261 2261 } else {
2262 2262 // If we can't allocate once, we probably cannot allocate
2263 2263 // again. Let's get out of the loop.
2264 2264 break;
2265 2265 }
2266 2266 }
2267 2267 }
2268 2268 #endif // !PRODUCT
2269 2269
2270 2270 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) {
2271 2271 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2272 2272
2273 2273 // We assume that if concurrent == true, then the caller is a
2274 2274 // concurrent thread that was joined the Suspendible Thread
2275 2275 // Set. If there's ever a cheap way to check this, we should add an
2276 2276 // assert here.
2277 2277
2278 2278 // We have already incremented _total_full_collections at the start
2279 2279 // of the GC, so total_full_collections() represents how many full
2280 2280 // collections have been started.
2281 2281 unsigned int full_collections_started = total_full_collections();
2282 2282
2283 2283 // Given that this method is called at the end of a Full GC or of a
2284 2284 // concurrent cycle, and those can be nested (i.e., a Full GC can
2285 2285 // interrupt a concurrent cycle), the number of full collections
2286 2286 // completed should be either one (in the case where there was no
2287 2287 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2288 2288 // behind the number of full collections started.
2289 2289
2290 2290 // This is the case for the inner caller, i.e. a Full GC.
2291 2291 assert(concurrent ||
2292 2292 (full_collections_started == _full_collections_completed + 1) ||
2293 2293 (full_collections_started == _full_collections_completed + 2),
2294 2294 err_msg("for inner caller (Full GC): full_collections_started = %u "
2295 2295 "is inconsistent with _full_collections_completed = %u",
2296 2296 full_collections_started, _full_collections_completed));
2297 2297
2298 2298 // This is the case for the outer caller, i.e. the concurrent cycle.
2299 2299 assert(!concurrent ||
2300 2300 (full_collections_started == _full_collections_completed + 1),
2301 2301 err_msg("for outer caller (concurrent cycle): "
2302 2302 "full_collections_started = %u "
2303 2303 "is inconsistent with _full_collections_completed = %u",
2304 2304 full_collections_started, _full_collections_completed));
2305 2305
2306 2306 _full_collections_completed += 1;
2307 2307
2308 2308 // We need to clear the "in_progress" flag in the CM thread before
2309 2309 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2310 2310 // is set) so that if a waiter requests another System.gc() it doesn't
2311 2311 // incorrectly see that a marking cyle is still in progress.
2312 2312 if (concurrent) {
2313 2313 _cmThread->clear_in_progress();
2314 2314 }
2315 2315
2316 2316 // This notify_all() will ensure that a thread that called
2317 2317 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2318 2318 // and it's waiting for a full GC to finish will be woken up. It is
2319 2319 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2320 2320 FullGCCount_lock->notify_all();
2321 2321 }
2322 2322
2323 2323 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) {
2324 2324 assert_at_safepoint(true /* should_be_vm_thread */);
2325 2325 GCCauseSetter gcs(this, cause);
2326 2326 switch (cause) {
2327 2327 case GCCause::_heap_inspection:
2328 2328 case GCCause::_heap_dump: {
2329 2329 HandleMark hm;
2330 2330 do_full_collection(false); // don't clear all soft refs
2331 2331 break;
2332 2332 }
2333 2333 default: // XXX FIX ME
2334 2334 ShouldNotReachHere(); // Unexpected use of this function
2335 2335 }
2336 2336 }
2337 2337
2338 2338 void G1CollectedHeap::collect(GCCause::Cause cause) {
2339 2339 // The caller doesn't have the Heap_lock
2340 2340 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock");
2341 2341
2342 2342 unsigned int gc_count_before;
2343 2343 unsigned int full_gc_count_before;
2344 2344 {
2345 2345 MutexLocker ml(Heap_lock);
2346 2346
2347 2347 // Read the GC count while holding the Heap_lock
2348 2348 gc_count_before = SharedHeap::heap()->total_collections();
2349 2349 full_gc_count_before = SharedHeap::heap()->total_full_collections();
2350 2350 }
2351 2351
2352 2352 if (should_do_concurrent_full_gc(cause)) {
2353 2353 // Schedule an initial-mark evacuation pause that will start a
2354 2354 // concurrent cycle. We're setting word_size to 0 which means that
2355 2355 // we are not requesting a post-GC allocation.
2356 2356 VM_G1IncCollectionPause op(gc_count_before,
2357 2357 0, /* word_size */
2358 2358 true, /* should_initiate_conc_mark */
2359 2359 g1_policy()->max_pause_time_ms(),
2360 2360 cause);
2361 2361 VMThread::execute(&op);
2362 2362 } else {
2363 2363 if (cause == GCCause::_gc_locker
2364 2364 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2365 2365
2366 2366 // Schedule a standard evacuation pause. We're setting word_size
2367 2367 // to 0 which means that we are not requesting a post-GC allocation.
2368 2368 VM_G1IncCollectionPause op(gc_count_before,
2369 2369 0, /* word_size */
2370 2370 false, /* should_initiate_conc_mark */
2371 2371 g1_policy()->max_pause_time_ms(),
2372 2372 cause);
2373 2373 VMThread::execute(&op);
2374 2374 } else {
2375 2375 // Schedule a Full GC.
2376 2376 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
2377 2377 VMThread::execute(&op);
2378 2378 }
2379 2379 }
2380 2380 }
2381 2381
2382 2382 bool G1CollectedHeap::is_in(const void* p) const {
2383 2383 HeapRegion* hr = _hrs.addr_to_region((HeapWord*) p);
2384 2384 if (hr != NULL) {
2385 2385 return hr->is_in(p);
2386 2386 } else {
2387 2387 return _perm_gen->as_gen()->is_in(p);
2388 2388 }
2389 2389 }
2390 2390
2391 2391 // Iteration functions.
2392 2392
2393 2393 // Iterates an OopClosure over all ref-containing fields of objects
2394 2394 // within a HeapRegion.
2395 2395
2396 2396 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2397 2397 MemRegion _mr;
2398 2398 OopClosure* _cl;
2399 2399 public:
2400 2400 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl)
2401 2401 : _mr(mr), _cl(cl) {}
2402 2402 bool doHeapRegion(HeapRegion* r) {
2403 2403 if (! r->continuesHumongous()) {
2404 2404 r->oop_iterate(_cl);
2405 2405 }
2406 2406 return false;
2407 2407 }
2408 2408 };
2409 2409
2410 2410 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) {
2411 2411 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2412 2412 heap_region_iterate(&blk);
2413 2413 if (do_perm) {
2414 2414 perm_gen()->oop_iterate(cl);
2415 2415 }
2416 2416 }
2417 2417
2418 2418 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) {
2419 2419 IterateOopClosureRegionClosure blk(mr, cl);
2420 2420 heap_region_iterate(&blk);
2421 2421 if (do_perm) {
2422 2422 perm_gen()->oop_iterate(cl);
2423 2423 }
2424 2424 }
2425 2425
2426 2426 // Iterates an ObjectClosure over all objects within a HeapRegion.
2427 2427
2428 2428 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2429 2429 ObjectClosure* _cl;
2430 2430 public:
2431 2431 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2432 2432 bool doHeapRegion(HeapRegion* r) {
2433 2433 if (! r->continuesHumongous()) {
2434 2434 r->object_iterate(_cl);
2435 2435 }
2436 2436 return false;
2437 2437 }
2438 2438 };
2439 2439
2440 2440 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) {
2441 2441 IterateObjectClosureRegionClosure blk(cl);
2442 2442 heap_region_iterate(&blk);
2443 2443 if (do_perm) {
2444 2444 perm_gen()->object_iterate(cl);
2445 2445 }
2446 2446 }
2447 2447
2448 2448 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2449 2449 // FIXME: is this right?
2450 2450 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2451 2451 }
2452 2452
2453 2453 // Calls a SpaceClosure on a HeapRegion.
2454 2454
2455 2455 class SpaceClosureRegionClosure: public HeapRegionClosure {
2456 2456 SpaceClosure* _cl;
2457 2457 public:
2458 2458 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2459 2459 bool doHeapRegion(HeapRegion* r) {
2460 2460 _cl->do_space(r);
2461 2461 return false;
2462 2462 }
2463 2463 };
2464 2464
2465 2465 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2466 2466 SpaceClosureRegionClosure blk(cl);
2467 2467 heap_region_iterate(&blk);
2468 2468 }
2469 2469
2470 2470 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2471 2471 _hrs.iterate(cl);
2472 2472 }
2473 2473
2474 2474 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r,
2475 2475 HeapRegionClosure* cl) const {
2476 2476 _hrs.iterate_from(r, cl);
2477 2477 }
2478 2478
2479 2479 void
2480 2480 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2481 2481 int worker,
2482 2482 jint claim_value) {
2483 2483 const size_t regions = n_regions();
2484 2484 const size_t worker_num = (G1CollectedHeap::use_parallel_gc_threads() ? ParallelGCThreads : 1);
2485 2485 // try to spread out the starting points of the workers
2486 2486 const size_t start_index = regions / worker_num * (size_t) worker;
2487 2487
2488 2488 // each worker will actually look at all regions
2489 2489 for (size_t count = 0; count < regions; ++count) {
2490 2490 const size_t index = (start_index + count) % regions;
2491 2491 assert(0 <= index && index < regions, "sanity");
2492 2492 HeapRegion* r = region_at(index);
2493 2493 // we'll ignore "continues humongous" regions (we'll process them
2494 2494 // when we come across their corresponding "start humongous"
2495 2495 // region) and regions already claimed
2496 2496 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2497 2497 continue;
2498 2498 }
2499 2499 // OK, try to claim it
2500 2500 if (r->claimHeapRegion(claim_value)) {
2501 2501 // success!
2502 2502 assert(!r->continuesHumongous(), "sanity");
2503 2503 if (r->startsHumongous()) {
2504 2504 // If the region is "starts humongous" we'll iterate over its
2505 2505 // "continues humongous" first; in fact we'll do them
2506 2506 // first. The order is important. In on case, calling the
2507 2507 // closure on the "starts humongous" region might de-allocate
2508 2508 // and clear all its "continues humongous" regions and, as a
2509 2509 // result, we might end up processing them twice. So, we'll do
2510 2510 // them first (notice: most closures will ignore them anyway) and
2511 2511 // then we'll do the "starts humongous" region.
2512 2512 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) {
2513 2513 HeapRegion* chr = region_at(ch_index);
2514 2514
2515 2515 // if the region has already been claimed or it's not
2516 2516 // "continues humongous" we're done
2517 2517 if (chr->claim_value() == claim_value ||
2518 2518 !chr->continuesHumongous()) {
2519 2519 break;
2520 2520 }
2521 2521
2522 2522 // Noone should have claimed it directly. We can given
2523 2523 // that we claimed its "starts humongous" region.
2524 2524 assert(chr->claim_value() != claim_value, "sanity");
2525 2525 assert(chr->humongous_start_region() == r, "sanity");
2526 2526
2527 2527 if (chr->claimHeapRegion(claim_value)) {
2528 2528 // we should always be able to claim it; noone else should
2529 2529 // be trying to claim this region
2530 2530
2531 2531 bool res2 = cl->doHeapRegion(chr);
2532 2532 assert(!res2, "Should not abort");
2533 2533
2534 2534 // Right now, this holds (i.e., no closure that actually
2535 2535 // does something with "continues humongous" regions
2536 2536 // clears them). We might have to weaken it in the future,
2537 2537 // but let's leave these two asserts here for extra safety.
2538 2538 assert(chr->continuesHumongous(), "should still be the case");
2539 2539 assert(chr->humongous_start_region() == r, "sanity");
2540 2540 } else {
2541 2541 guarantee(false, "we should not reach here");
2542 2542 }
2543 2543 }
2544 2544 }
2545 2545
2546 2546 assert(!r->continuesHumongous(), "sanity");
2547 2547 bool res = cl->doHeapRegion(r);
2548 2548 assert(!res, "Should not abort");
2549 2549 }
2550 2550 }
2551 2551 }
2552 2552
2553 2553 class ResetClaimValuesClosure: public HeapRegionClosure {
2554 2554 public:
2555 2555 bool doHeapRegion(HeapRegion* r) {
2556 2556 r->set_claim_value(HeapRegion::InitialClaimValue);
2557 2557 return false;
2558 2558 }
2559 2559 };
2560 2560
2561 2561 void
2562 2562 G1CollectedHeap::reset_heap_region_claim_values() {
2563 2563 ResetClaimValuesClosure blk;
2564 2564 heap_region_iterate(&blk);
2565 2565 }
2566 2566
2567 2567 #ifdef ASSERT
2568 2568 // This checks whether all regions in the heap have the correct claim
2569 2569 // value. I also piggy-backed on this a check to ensure that the
2570 2570 // humongous_start_region() information on "continues humongous"
2571 2571 // regions is correct.
2572 2572
2573 2573 class CheckClaimValuesClosure : public HeapRegionClosure {
2574 2574 private:
2575 2575 jint _claim_value;
2576 2576 size_t _failures;
2577 2577 HeapRegion* _sh_region;
2578 2578 public:
2579 2579 CheckClaimValuesClosure(jint claim_value) :
2580 2580 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2581 2581 bool doHeapRegion(HeapRegion* r) {
2582 2582 if (r->claim_value() != _claim_value) {
2583 2583 gclog_or_tty->print_cr("Region ["PTR_FORMAT","PTR_FORMAT"), "
2584 2584 "claim value = %d, should be %d",
2585 2585 r->bottom(), r->end(), r->claim_value(),
2586 2586 _claim_value);
2587 2587 ++_failures;
2588 2588 }
2589 2589 if (!r->isHumongous()) {
2590 2590 _sh_region = NULL;
2591 2591 } else if (r->startsHumongous()) {
2592 2592 _sh_region = r;
2593 2593 } else if (r->continuesHumongous()) {
2594 2594 if (r->humongous_start_region() != _sh_region) {
2595 2595 gclog_or_tty->print_cr("Region ["PTR_FORMAT","PTR_FORMAT"), "
2596 2596 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2597 2597 r->bottom(), r->end(),
2598 2598 r->humongous_start_region(),
2599 2599 _sh_region);
2600 2600 ++_failures;
2601 2601 }
2602 2602 }
2603 2603 return false;
2604 2604 }
2605 2605 size_t failures() {
2606 2606 return _failures;
2607 2607 }
2608 2608 };
2609 2609
2610 2610 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2611 2611 CheckClaimValuesClosure cl(claim_value);
2612 2612 heap_region_iterate(&cl);
2613 2613 return cl.failures() == 0;
2614 2614 }
2615 2615 #endif // ASSERT
2616 2616
2617 2617 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2618 2618 HeapRegion* r = g1_policy()->collection_set();
2619 2619 while (r != NULL) {
2620 2620 HeapRegion* next = r->next_in_collection_set();
2621 2621 if (cl->doHeapRegion(r)) {
2622 2622 cl->incomplete();
2623 2623 return;
2624 2624 }
2625 2625 r = next;
2626 2626 }
2627 2627 }
2628 2628
2629 2629 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2630 2630 HeapRegionClosure *cl) {
2631 2631 if (r == NULL) {
2632 2632 // The CSet is empty so there's nothing to do.
2633 2633 return;
2634 2634 }
2635 2635
2636 2636 assert(r->in_collection_set(),
2637 2637 "Start region must be a member of the collection set.");
2638 2638 HeapRegion* cur = r;
2639 2639 while (cur != NULL) {
2640 2640 HeapRegion* next = cur->next_in_collection_set();
2641 2641 if (cl->doHeapRegion(cur) && false) {
2642 2642 cl->incomplete();
2643 2643 return;
2644 2644 }
2645 2645 cur = next;
2646 2646 }
2647 2647 cur = g1_policy()->collection_set();
2648 2648 while (cur != r) {
2649 2649 HeapRegion* next = cur->next_in_collection_set();
2650 2650 if (cl->doHeapRegion(cur) && false) {
2651 2651 cl->incomplete();
2652 2652 return;
2653 2653 }
2654 2654 cur = next;
2655 2655 }
2656 2656 }
2657 2657
2658 2658 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2659 2659 return n_regions() > 0 ? region_at(0) : NULL;
2660 2660 }
2661 2661
2662 2662
2663 2663 Space* G1CollectedHeap::space_containing(const void* addr) const {
2664 2664 Space* res = heap_region_containing(addr);
2665 2665 if (res == NULL)
2666 2666 res = perm_gen()->space_containing(addr);
2667 2667 return res;
2668 2668 }
2669 2669
2670 2670 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2671 2671 Space* sp = space_containing(addr);
2672 2672 if (sp != NULL) {
2673 2673 return sp->block_start(addr);
2674 2674 }
2675 2675 return NULL;
2676 2676 }
2677 2677
2678 2678 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2679 2679 Space* sp = space_containing(addr);
2680 2680 assert(sp != NULL, "block_size of address outside of heap");
2681 2681 return sp->block_size(addr);
2682 2682 }
2683 2683
2684 2684 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2685 2685 Space* sp = space_containing(addr);
2686 2686 return sp->block_is_obj(addr);
2687 2687 }
2688 2688
2689 2689 bool G1CollectedHeap::supports_tlab_allocation() const {
2690 2690 return true;
2691 2691 }
2692 2692
2693 2693 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2694 2694 return HeapRegion::GrainBytes;
2695 2695 }
2696 2696
2697 2697 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2698 2698 // Return the remaining space in the cur alloc region, but not less than
2699 2699 // the min TLAB size.
2700 2700
2701 2701 // Also, this value can be at most the humongous object threshold,
2702 2702 // since we can't allow tlabs to grow big enough to accomodate
2703 2703 // humongous objects.
2704 2704
2705 2705 HeapRegion* hr = _mutator_alloc_region.get();
2706 2706 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2707 2707 if (hr == NULL) {
2708 2708 return max_tlab_size;
2709 2709 } else {
2710 2710 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2711 2711 }
2712 2712 }
2713 2713
2714 2714 size_t G1CollectedHeap::max_capacity() const {
2715 2715 return _g1_reserved.byte_size();
2716 2716 }
2717 2717
2718 2718 jlong G1CollectedHeap::millis_since_last_gc() {
2719 2719 // assert(false, "NYI");
2720 2720 return 0;
2721 2721 }
2722 2722
2723 2723 void G1CollectedHeap::prepare_for_verify() {
2724 2724 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2725 2725 ensure_parsability(false);
2726 2726 }
2727 2727 g1_rem_set()->prepare_for_verify();
2728 2728 }
2729 2729
2730 2730 class VerifyLivenessOopClosure: public OopClosure {
2731 2731 G1CollectedHeap* _g1h;
2732 2732 VerifyOption _vo;
2733 2733 public:
2734 2734 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
2735 2735 _g1h(g1h), _vo(vo)
2736 2736 { }
2737 2737 void do_oop(narrowOop *p) { do_oop_work(p); }
2738 2738 void do_oop( oop *p) { do_oop_work(p); }
2739 2739
2740 2740 template <class T> void do_oop_work(T *p) {
2741 2741 oop obj = oopDesc::load_decode_heap_oop(p);
2742 2742 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
2743 2743 "Dead object referenced by a not dead object");
2744 2744 }
2745 2745 };
2746 2746
2747 2747 class VerifyObjsInRegionClosure: public ObjectClosure {
2748 2748 private:
2749 2749 G1CollectedHeap* _g1h;
2750 2750 size_t _live_bytes;
2751 2751 HeapRegion *_hr;
2752 2752 VerifyOption _vo;
2753 2753 public:
2754 2754 // _vo == UsePrevMarking -> use "prev" marking information,
2755 2755 // _vo == UseNextMarking -> use "next" marking information,
2756 2756 // _vo == UseMarkWord -> use mark word from object header.
2757 2757 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
2758 2758 : _live_bytes(0), _hr(hr), _vo(vo) {
2759 2759 _g1h = G1CollectedHeap::heap();
2760 2760 }
2761 2761 void do_object(oop o) {
2762 2762 VerifyLivenessOopClosure isLive(_g1h, _vo);
2763 2763 assert(o != NULL, "Huh?");
2764 2764 if (!_g1h->is_obj_dead_cond(o, _vo)) {
2765 2765 // If the object is alive according to the mark word,
2766 2766 // then verify that the marking information agrees.
2767 2767 // Note we can't verify the contra-positive of the
2768 2768 // above: if the object is dead (according to the mark
2769 2769 // word), it may not be marked, or may have been marked
2770 2770 // but has since became dead, or may have been allocated
2771 2771 // since the last marking.
2772 2772 if (_vo == VerifyOption_G1UseMarkWord) {
2773 2773 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
2774 2774 }
2775 2775
2776 2776 o->oop_iterate(&isLive);
2777 2777 if (!_hr->obj_allocated_since_prev_marking(o)) {
2778 2778 size_t obj_size = o->size(); // Make sure we don't overflow
2779 2779 _live_bytes += (obj_size * HeapWordSize);
2780 2780 }
2781 2781 }
2782 2782 }
2783 2783 size_t live_bytes() { return _live_bytes; }
2784 2784 };
2785 2785
2786 2786 class PrintObjsInRegionClosure : public ObjectClosure {
2787 2787 HeapRegion *_hr;
2788 2788 G1CollectedHeap *_g1;
2789 2789 public:
2790 2790 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
2791 2791 _g1 = G1CollectedHeap::heap();
2792 2792 };
2793 2793
2794 2794 void do_object(oop o) {
2795 2795 if (o != NULL) {
2796 2796 HeapWord *start = (HeapWord *) o;
2797 2797 size_t word_sz = o->size();
2798 2798 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
2799 2799 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
2800 2800 (void*) o, word_sz,
2801 2801 _g1->isMarkedPrev(o),
2802 2802 _g1->isMarkedNext(o),
2803 2803 _hr->obj_allocated_since_prev_marking(o));
2804 2804 HeapWord *end = start + word_sz;
2805 2805 HeapWord *cur;
2806 2806 int *val;
2807 2807 for (cur = start; cur < end; cur++) {
2808 2808 val = (int *) cur;
2809 2809 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
2810 2810 }
2811 2811 }
2812 2812 }
2813 2813 };
2814 2814
2815 2815 class VerifyRegionClosure: public HeapRegionClosure {
2816 2816 private:
2817 2817 bool _allow_dirty;
2818 2818 bool _par;
2819 2819 VerifyOption _vo;
2820 2820 bool _failures;
2821 2821 public:
2822 2822 // _vo == UsePrevMarking -> use "prev" marking information,
2823 2823 // _vo == UseNextMarking -> use "next" marking information,
2824 2824 // _vo == UseMarkWord -> use mark word from object header.
2825 2825 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo)
2826 2826 : _allow_dirty(allow_dirty),
2827 2827 _par(par),
2828 2828 _vo(vo),
2829 2829 _failures(false) {}
2830 2830
2831 2831 bool failures() {
2832 2832 return _failures;
2833 2833 }
2834 2834
2835 2835 bool doHeapRegion(HeapRegion* r) {
2836 2836 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue,
2837 2837 "Should be unclaimed at verify points.");
2838 2838 if (!r->continuesHumongous()) {
2839 2839 bool failures = false;
2840 2840 r->verify(_allow_dirty, _vo, &failures);
2841 2841 if (failures) {
2842 2842 _failures = true;
2843 2843 } else {
2844 2844 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
2845 2845 r->object_iterate(¬_dead_yet_cl);
2846 2846 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
2847 2847 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
2848 2848 "max_live_bytes "SIZE_FORMAT" "
2849 2849 "< calculated "SIZE_FORMAT,
2850 2850 r->bottom(), r->end(),
2851 2851 r->max_live_bytes(),
2852 2852 not_dead_yet_cl.live_bytes());
2853 2853 _failures = true;
2854 2854 }
2855 2855 }
2856 2856 }
2857 2857 return false; // stop the region iteration if we hit a failure
2858 2858 }
2859 2859 };
2860 2860
2861 2861 class VerifyRootsClosure: public OopsInGenClosure {
2862 2862 private:
2863 2863 G1CollectedHeap* _g1h;
2864 2864 VerifyOption _vo;
2865 2865 bool _failures;
2866 2866 public:
2867 2867 // _vo == UsePrevMarking -> use "prev" marking information,
2868 2868 // _vo == UseNextMarking -> use "next" marking information,
2869 2869 // _vo == UseMarkWord -> use mark word from object header.
2870 2870 VerifyRootsClosure(VerifyOption vo) :
2871 2871 _g1h(G1CollectedHeap::heap()),
2872 2872 _vo(vo),
2873 2873 _failures(false) { }
2874 2874
2875 2875 bool failures() { return _failures; }
2876 2876
2877 2877 template <class T> void do_oop_nv(T* p) {
2878 2878 T heap_oop = oopDesc::load_heap_oop(p);
2879 2879 if (!oopDesc::is_null(heap_oop)) {
2880 2880 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
2881 2881 if (_g1h->is_obj_dead_cond(obj, _vo)) {
2882 2882 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
2883 2883 "points to dead obj "PTR_FORMAT, p, (void*) obj);
2884 2884 if (_vo == VerifyOption_G1UseMarkWord) {
2885 2885 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
2886 2886 }
2887 2887 obj->print_on(gclog_or_tty);
2888 2888 _failures = true;
2889 2889 }
2890 2890 }
2891 2891 }
2892 2892
2893 2893 void do_oop(oop* p) { do_oop_nv(p); }
2894 2894 void do_oop(narrowOop* p) { do_oop_nv(p); }
2895 2895 };
2896 2896
2897 2897 // This is the task used for parallel heap verification.
2898 2898
2899 2899 class G1ParVerifyTask: public AbstractGangTask {
2900 2900 private:
2901 2901 G1CollectedHeap* _g1h;
2902 2902 bool _allow_dirty;
2903 2903 VerifyOption _vo;
2904 2904 bool _failures;
2905 2905
2906 2906 public:
2907 2907 // _vo == UsePrevMarking -> use "prev" marking information,
2908 2908 // _vo == UseNextMarking -> use "next" marking information,
2909 2909 // _vo == UseMarkWord -> use mark word from object header.
2910 2910 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) :
2911 2911 AbstractGangTask("Parallel verify task"),
2912 2912 _g1h(g1h),
2913 2913 _allow_dirty(allow_dirty),
2914 2914 _vo(vo),
2915 2915 _failures(false) { }
2916 2916
2917 2917 bool failures() {
2918 2918 return _failures;
2919 2919 }
2920 2920
2921 2921 void work(int worker_i) {
2922 2922 HandleMark hm;
2923 2923 VerifyRegionClosure blk(_allow_dirty, true, _vo);
2924 2924 _g1h->heap_region_par_iterate_chunked(&blk, worker_i,
2925 2925 HeapRegion::ParVerifyClaimValue);
2926 2926 if (blk.failures()) {
2927 2927 _failures = true;
2928 2928 }
2929 2929 }
2930 2930 };
2931 2931
2932 2932 void G1CollectedHeap::verify(bool allow_dirty, bool silent) {
2933 2933 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking);
2934 2934 }
2935 2935
2936 2936 void G1CollectedHeap::verify(bool allow_dirty,
2937 2937 bool silent,
2938 2938 VerifyOption vo) {
2939 2939 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2940 2940 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); }
2941 2941 VerifyRootsClosure rootsCl(vo);
2942 2942 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
2943 2943
2944 2944 // We apply the relevant closures to all the oops in the
2945 2945 // system dictionary, the string table and the code cache.
2946 2946 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
2947 2947
2948 2948 process_strong_roots(true, // activate StrongRootsScope
2949 2949 true, // we set "collecting perm gen" to true,
2950 2950 // so we don't reset the dirty cards in the perm gen.
2951 2951 SharedHeap::ScanningOption(so), // roots scanning options
2952 2952 &rootsCl,
2953 2953 &blobsCl,
2954 2954 &rootsCl);
2955 2955
2956 2956 // If we're verifying after the marking phase of a Full GC then we can't
2957 2957 // treat the perm gen as roots into the G1 heap. Some of the objects in
2958 2958 // the perm gen may be dead and hence not marked. If one of these dead
2959 2959 // objects is considered to be a root then we may end up with a false
2960 2960 // "Root location <x> points to dead ob <y>" failure.
2961 2961 if (vo != VerifyOption_G1UseMarkWord) {
2962 2962 // Since we used "collecting_perm_gen" == true above, we will not have
2963 2963 // checked the refs from perm into the G1-collected heap. We check those
2964 2964 // references explicitly below. Whether the relevant cards are dirty
2965 2965 // is checked further below in the rem set verification.
2966 2966 if (!silent) { gclog_or_tty->print("Permgen roots "); }
2967 2967 perm_gen()->oop_iterate(&rootsCl);
2968 2968 }
2969 2969 bool failures = rootsCl.failures();
2970 2970
2971 2971 if (vo != VerifyOption_G1UseMarkWord) {
2972 2972 // If we're verifying during a full GC then the region sets
2973 2973 // will have been torn down at the start of the GC. Therefore
2974 2974 // verifying the region sets will fail. So we only verify
2975 2975 // the region sets when not in a full GC.
2976 2976 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
2977 2977 verify_region_sets();
2978 2978 }
2979 2979
2980 2980 if (!silent) { gclog_or_tty->print("HeapRegions "); }
2981 2981 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
2982 2982 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2983 2983 "sanity check");
2984 2984
2985 2985 G1ParVerifyTask task(this, allow_dirty, vo);
2986 2986 int n_workers = workers()->total_workers();
2987 2987 set_par_threads(n_workers);
2988 2988 workers()->run_task(&task);
2989 2989 set_par_threads(0);
2990 2990 if (task.failures()) {
2991 2991 failures = true;
2992 2992 }
2993 2993
2994 2994 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
2995 2995 "sanity check");
2996 2996
2997 2997 reset_heap_region_claim_values();
2998 2998
2999 2999 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3000 3000 "sanity check");
3001 3001 } else {
3002 3002 VerifyRegionClosure blk(allow_dirty, false, vo);
3003 3003 heap_region_iterate(&blk);
3004 3004 if (blk.failures()) {
3005 3005 failures = true;
3006 3006 }
3007 3007 }
3008 3008 if (!silent) gclog_or_tty->print("RemSet ");
3009 3009 rem_set()->verify();
3010 3010
3011 3011 if (failures) {
3012 3012 gclog_or_tty->print_cr("Heap:");
3013 3013 print_on(gclog_or_tty, true /* extended */);
3014 3014 gclog_or_tty->print_cr("");
3015 3015 #ifndef PRODUCT
3016 3016 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3017 3017 concurrent_mark()->print_reachable("at-verification-failure",
3018 3018 vo, false /* all */);
3019 3019 }
3020 3020 #endif
3021 3021 gclog_or_tty->flush();
3022 3022 }
3023 3023 guarantee(!failures, "there should not have been any failures");
3024 3024 } else {
3025 3025 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3026 3026 }
3027 3027 }
3028 3028
3029 3029 class PrintRegionClosure: public HeapRegionClosure {
3030 3030 outputStream* _st;
3031 3031 public:
3032 3032 PrintRegionClosure(outputStream* st) : _st(st) {}
3033 3033 bool doHeapRegion(HeapRegion* r) {
3034 3034 r->print_on(_st);
3035 3035 return false;
3036 3036 }
3037 3037 };
3038 3038
3039 3039 void G1CollectedHeap::print() const { print_on(tty); }
3040 3040
3041 3041 void G1CollectedHeap::print_on(outputStream* st) const {
3042 3042 print_on(st, PrintHeapAtGCExtended);
3043 3043 }
3044 3044
3045 3045 void G1CollectedHeap::print_on(outputStream* st, bool extended) const {
3046 3046 st->print(" %-20s", "garbage-first heap");
3047 3047 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3048 3048 capacity()/K, used_unlocked()/K);
3049 3049 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3050 3050 _g1_storage.low_boundary(),
3051 3051 _g1_storage.high(),
3052 3052 _g1_storage.high_boundary());
3053 3053 st->cr();
3054 3054 st->print(" region size " SIZE_FORMAT "K, ",
3055 3055 HeapRegion::GrainBytes/K);
3056 3056 size_t young_regions = _young_list->length();
3057 3057 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ",
3058 3058 young_regions, young_regions * HeapRegion::GrainBytes / K);
3059 3059 size_t survivor_regions = g1_policy()->recorded_survivor_regions();
3060 3060 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)",
3061 3061 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K);
3062 3062 st->cr();
3063 3063 perm()->as_gen()->print_on(st);
3064 3064 if (extended) {
3065 3065 st->cr();
3066 3066 print_on_extended(st);
3067 3067 }
3068 3068 }
3069 3069
3070 3070 void G1CollectedHeap::print_on_extended(outputStream* st) const {
3071 3071 PrintRegionClosure blk(st);
3072 3072 heap_region_iterate(&blk);
3073 3073 }
3074 3074
3075 3075 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3076 3076 if (G1CollectedHeap::use_parallel_gc_threads()) {
3077 3077 workers()->print_worker_threads_on(st);
3078 3078 }
3079 3079 _cmThread->print_on(st);
3080 3080 st->cr();
3081 3081 _cm->print_worker_threads_on(st);
3082 3082 _cg1r->print_worker_threads_on(st);
3083 3083 st->cr();
3084 3084 }
3085 3085
3086 3086 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3087 3087 if (G1CollectedHeap::use_parallel_gc_threads()) {
3088 3088 workers()->threads_do(tc);
3089 3089 }
3090 3090 tc->do_thread(_cmThread);
3091 3091 _cg1r->threads_do(tc);
3092 3092 }
3093 3093
3094 3094 void G1CollectedHeap::print_tracing_info() const {
3095 3095 // We'll overload this to mean "trace GC pause statistics."
3096 3096 if (TraceGen0Time || TraceGen1Time) {
3097 3097 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3098 3098 // to that.
3099 3099 g1_policy()->print_tracing_info();
3100 3100 }
3101 3101 if (G1SummarizeRSetStats) {
3102 3102 g1_rem_set()->print_summary_info();
3103 3103 }
3104 3104 if (G1SummarizeConcMark) {
3105 3105 concurrent_mark()->print_summary_info();
3106 3106 }
3107 3107 g1_policy()->print_yg_surv_rate_info();
3108 3108 SpecializationStats::print();
3109 3109 }
3110 3110
3111 3111 #ifndef PRODUCT
3112 3112 // Helpful for debugging RSet issues.
3113 3113
3114 3114 class PrintRSetsClosure : public HeapRegionClosure {
3115 3115 private:
3116 3116 const char* _msg;
3117 3117 size_t _occupied_sum;
3118 3118
3119 3119 public:
3120 3120 bool doHeapRegion(HeapRegion* r) {
3121 3121 HeapRegionRemSet* hrrs = r->rem_set();
3122 3122 size_t occupied = hrrs->occupied();
3123 3123 _occupied_sum += occupied;
3124 3124
3125 3125 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3126 3126 HR_FORMAT_PARAMS(r));
3127 3127 if (occupied == 0) {
3128 3128 gclog_or_tty->print_cr(" RSet is empty");
3129 3129 } else {
3130 3130 hrrs->print();
3131 3131 }
3132 3132 gclog_or_tty->print_cr("----------");
3133 3133 return false;
3134 3134 }
3135 3135
3136 3136 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3137 3137 gclog_or_tty->cr();
3138 3138 gclog_or_tty->print_cr("========================================");
3139 3139 gclog_or_tty->print_cr(msg);
3140 3140 gclog_or_tty->cr();
3141 3141 }
3142 3142
3143 3143 ~PrintRSetsClosure() {
3144 3144 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3145 3145 gclog_or_tty->print_cr("========================================");
3146 3146 gclog_or_tty->cr();
3147 3147 }
3148 3148 };
3149 3149
3150 3150 void G1CollectedHeap::print_cset_rsets() {
3151 3151 PrintRSetsClosure cl("Printing CSet RSets");
3152 3152 collection_set_iterate(&cl);
3153 3153 }
3154 3154
3155 3155 void G1CollectedHeap::print_all_rsets() {
3156 3156 PrintRSetsClosure cl("Printing All RSets");;
3157 3157 heap_region_iterate(&cl);
3158 3158 }
3159 3159 #endif // PRODUCT
3160 3160
3161 3161 G1CollectedHeap* G1CollectedHeap::heap() {
3162 3162 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3163 3163 "not a garbage-first heap");
3164 3164 return _g1h;
3165 3165 }
3166 3166
3167 3167 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3168 3168 // always_do_update_barrier = false;
3169 3169 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3170 3170 // Call allocation profiler
3171 3171 AllocationProfiler::iterate_since_last_gc();
3172 3172 // Fill TLAB's and such
3173 3173 ensure_parsability(true);
3174 3174 }
3175 3175
3176 3176 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3177 3177 // FIXME: what is this about?
3178 3178 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3179 3179 // is set.
3180 3180 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3181 3181 "derived pointer present"));
3182 3182 // always_do_update_barrier = true;
3183 3183
3184 3184 // We have just completed a GC. Update the soft reference
3185 3185 // policy with the new heap occupancy
3186 3186 Universe::update_heap_info_at_gc();
3187 3187 }
3188 3188
3189 3189 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3190 3190 unsigned int gc_count_before,
3191 3191 bool* succeeded) {
3192 3192 assert_heap_not_locked_and_not_at_safepoint();
3193 3193 g1_policy()->record_stop_world_start();
3194 3194 VM_G1IncCollectionPause op(gc_count_before,
3195 3195 word_size,
3196 3196 false, /* should_initiate_conc_mark */
3197 3197 g1_policy()->max_pause_time_ms(),
3198 3198 GCCause::_g1_inc_collection_pause);
3199 3199 VMThread::execute(&op);
3200 3200
3201 3201 HeapWord* result = op.result();
3202 3202 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3203 3203 assert(result == NULL || ret_succeeded,
3204 3204 "the result should be NULL if the VM did not succeed");
3205 3205 *succeeded = ret_succeeded;
3206 3206
3207 3207 assert_heap_not_locked();
3208 3208 return result;
3209 3209 }
3210 3210
3211 3211 void
3212 3212 G1CollectedHeap::doConcurrentMark() {
3213 3213 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3214 3214 if (!_cmThread->in_progress()) {
3215 3215 _cmThread->set_started();
3216 3216 CGC_lock->notify();
3217 3217 }
3218 3218 }
3219 3219
3220 3220 // <NEW PREDICTION>
3221 3221
3222 3222 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr,
3223 3223 bool young) {
3224 3224 return _g1_policy->predict_region_elapsed_time_ms(hr, young);
3225 3225 }
3226 3226
3227 3227 void G1CollectedHeap::check_if_region_is_too_expensive(double
3228 3228 predicted_time_ms) {
3229 3229 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms);
3230 3230 }
3231 3231
3232 3232 size_t G1CollectedHeap::pending_card_num() {
3233 3233 size_t extra_cards = 0;
3234 3234 JavaThread *curr = Threads::first();
3235 3235 while (curr != NULL) {
3236 3236 DirtyCardQueue& dcq = curr->dirty_card_queue();
3237 3237 extra_cards += dcq.size();
3238 3238 curr = curr->next();
3239 3239 }
3240 3240 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3241 3241 size_t buffer_size = dcqs.buffer_size();
3242 3242 size_t buffer_num = dcqs.completed_buffers_num();
3243 3243 return buffer_size * buffer_num + extra_cards;
3244 3244 }
3245 3245
3246 3246 size_t G1CollectedHeap::max_pending_card_num() {
3247 3247 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3248 3248 size_t buffer_size = dcqs.buffer_size();
3249 3249 size_t buffer_num = dcqs.completed_buffers_num();
3250 3250 int thread_num = Threads::number_of_threads();
3251 3251 return (buffer_num + thread_num) * buffer_size;
3252 3252 }
3253 3253
3254 3254 size_t G1CollectedHeap::cards_scanned() {
3255 3255 return g1_rem_set()->cardsScanned();
3256 3256 }
3257 3257
3258 3258 void
3259 3259 G1CollectedHeap::setup_surviving_young_words() {
3260 3260 guarantee( _surviving_young_words == NULL, "pre-condition" );
3261 3261 size_t array_length = g1_policy()->young_cset_length();
3262 3262 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length);
3263 3263 if (_surviving_young_words == NULL) {
3264 3264 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3265 3265 "Not enough space for young surv words summary.");
3266 3266 }
3267 3267 memset(_surviving_young_words, 0, array_length * sizeof(size_t));
3268 3268 #ifdef ASSERT
3269 3269 for (size_t i = 0; i < array_length; ++i) {
3270 3270 assert( _surviving_young_words[i] == 0, "memset above" );
3271 3271 }
3272 3272 #endif // !ASSERT
3273 3273 }
3274 3274
3275 3275 void
3276 3276 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3277 3277 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3278 3278 size_t array_length = g1_policy()->young_cset_length();
3279 3279 for (size_t i = 0; i < array_length; ++i)
3280 3280 _surviving_young_words[i] += surv_young_words[i];
3281 3281 }
3282 3282
3283 3283 void
3284 3284 G1CollectedHeap::cleanup_surviving_young_words() {
3285 3285 guarantee( _surviving_young_words != NULL, "pre-condition" );
3286 3286 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words);
3287 3287 _surviving_young_words = NULL;
3288 3288 }
3289 3289
3290 3290 // </NEW PREDICTION>
3291 3291
3292 3292 #ifdef ASSERT
3293 3293 class VerifyCSetClosure: public HeapRegionClosure {
3294 3294 public:
3295 3295 bool doHeapRegion(HeapRegion* hr) {
3296 3296 // Here we check that the CSet region's RSet is ready for parallel
3297 3297 // iteration. The fields that we'll verify are only manipulated
3298 3298 // when the region is part of a CSet and is collected. Afterwards,
3299 3299 // we reset these fields when we clear the region's RSet (when the
3300 3300 // region is freed) so they are ready when the region is
3301 3301 // re-allocated. The only exception to this is if there's an
3302 3302 // evacuation failure and instead of freeing the region we leave
3303 3303 // it in the heap. In that case, we reset these fields during
3304 3304 // evacuation failure handling.
3305 3305 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3306 3306
3307 3307 // Here's a good place to add any other checks we'd like to
3308 3308 // perform on CSet regions.
3309 3309 return false;
3310 3310 }
3311 3311 };
3312 3312 #endif // ASSERT
3313 3313
3314 3314 #if TASKQUEUE_STATS
3315 3315 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3316 3316 st->print_raw_cr("GC Task Stats");
3317 3317 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3318 3318 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3319 3319 }
3320 3320
3321 3321 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3322 3322 print_taskqueue_stats_hdr(st);
3323 3323
3324 3324 TaskQueueStats totals;
3325 3325 const int n = workers() != NULL ? workers()->total_workers() : 1;
3326 3326 for (int i = 0; i < n; ++i) {
3327 3327 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3328 3328 totals += task_queue(i)->stats;
3329 3329 }
3330 3330 st->print_raw("tot "); totals.print(st); st->cr();
3331 3331
3332 3332 DEBUG_ONLY(totals.verify());
3333 3333 }
3334 3334
3335 3335 void G1CollectedHeap::reset_taskqueue_stats() {
3336 3336 const int n = workers() != NULL ? workers()->total_workers() : 1;
3337 3337 for (int i = 0; i < n; ++i) {
3338 3338 task_queue(i)->stats.reset();
3339 3339 }
3340 3340 }
3341 3341 #endif // TASKQUEUE_STATS
3342 3342
3343 3343 bool
3344 3344 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3345 3345 assert_at_safepoint(true /* should_be_vm_thread */);
3346 3346 guarantee(!is_gc_active(), "collection is not reentrant");
3347 3347
3348 3348 if (GC_locker::check_active_before_gc()) {
3349 3349 return false;
3350 3350 }
3351 3351
3352 3352 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3353 3353 ResourceMark rm;
3354 3354
3355 3355 if (PrintHeapAtGC) {
3356 3356 Universe::print_heap_before_gc();
3357 3357 }
3358 3358
3359 3359 verify_region_sets_optional();
3360 3360 verify_dirty_young_regions();
3361 3361
3362 3362 {
3363 3363 // This call will decide whether this pause is an initial-mark
3364 3364 // pause. If it is, during_initial_mark_pause() will return true
3365 3365 // for the duration of this pause.
3366 3366 g1_policy()->decide_on_conc_mark_initiation();
3367 3367
3368 3368 // We do not allow initial-mark to be piggy-backed on a
3369 3369 // partially-young GC.
3370 3370 assert(!g1_policy()->during_initial_mark_pause() ||
3371 3371 g1_policy()->full_young_gcs(), "sanity");
3372 3372
3373 3373 // We also do not allow partially-young GCs during marking.
3374 3374 assert(!mark_in_progress() || g1_policy()->full_young_gcs(), "sanity");
3375 3375
3376 3376 char verbose_str[128];
3377 3377 sprintf(verbose_str, "GC pause ");
3378 3378 if (g1_policy()->full_young_gcs()) {
3379 3379 strcat(verbose_str, "(young)");
3380 3380 } else {
3381 3381 strcat(verbose_str, "(partial)");
3382 3382 }
3383 3383 if (g1_policy()->during_initial_mark_pause()) {
3384 3384 strcat(verbose_str, " (initial-mark)");
3385 3385 // We are about to start a marking cycle, so we increment the
3386 3386 // full collection counter.
3387 3387 increment_total_full_collections();
3388 3388 }
3389 3389
3390 3390 // if PrintGCDetails is on, we'll print long statistics information
3391 3391 // in the collector policy code, so let's not print this as the output
3392 3392 // is messy if we do.
3393 3393 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
3394 3394 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3395 3395 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty);
3396 3396
3397 3397 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3398 3398 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3399 3399
3400 3400 // If the secondary_free_list is not empty, append it to the
3401 3401 // free_list. No need to wait for the cleanup operation to finish;
3402 3402 // the region allocation code will check the secondary_free_list
3403 3403 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3404 3404 // set, skip this step so that the region allocation code has to
3405 3405 // get entries from the secondary_free_list.
3406 3406 if (!G1StressConcRegionFreeing) {
3407 3407 append_secondary_free_list_if_not_empty_with_lock();
3408 3408 }
3409 3409
3410 3410 assert(check_young_list_well_formed(),
3411 3411 "young list should be well formed");
3412 3412
3413 3413 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3414 3414 IsGCActiveMark x;
3415 3415
3416 3416 gc_prologue(false);
3417 3417 increment_total_collections(false /* full gc */);
3418 3418 increment_gc_time_stamp();
3419 3419
3420 3420 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
3421 3421 HandleMark hm; // Discard invalid handles created during verification
3422 3422 gclog_or_tty->print(" VerifyBeforeGC:");
3423 3423 prepare_for_verify();
3424 3424 Universe::verify(/* allow dirty */ false,
3425 3425 /* silent */ false,
3426 3426 /* option */ VerifyOption_G1UsePrevMarking);
3427 3427
3428 3428 }
3429 3429
3430 3430 COMPILER2_PRESENT(DerivedPointerTable::clear());
3431 3431
3432 3432 // Please see comment in g1CollectedHeap.hpp and
3433 3433 // G1CollectedHeap::ref_processing_init() to see how
3434 3434 // reference processing currently works in G1.
3435 3435
3436 3436 // Enable discovery in the STW reference processor
3437 3437 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3438 3438 true /*verify_no_refs*/);
3439 3439
3440 3440 {
3441 3441 // We want to temporarily turn off discovery by the
3442 3442 // CM ref processor, if necessary, and turn it back on
3443 3443 // on again later if we do. Using a scoped
3444 3444 // NoRefDiscovery object will do this.
3445 3445 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3446 3446
3447 3447 // Forget the current alloc region (we might even choose it to be part
3448 3448 // of the collection set!).
3449 3449 release_mutator_alloc_region();
3450 3450
3451 3451 // We should call this after we retire the mutator alloc
3452 3452 // region(s) so that all the ALLOC / RETIRE events are generated
3453 3453 // before the start GC event.
3454 3454 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3455 3455
3456 3456 // The elapsed time induced by the start time below deliberately elides
3457 3457 // the possible verification above.
3458 3458 double start_time_sec = os::elapsedTime();
3459 3459 size_t start_used_bytes = used();
3460 3460
3461 3461 #if YOUNG_LIST_VERBOSE
3462 3462 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3463 3463 _young_list->print();
3464 3464 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3465 3465 #endif // YOUNG_LIST_VERBOSE
3466 3466
3467 3467 g1_policy()->record_collection_pause_start(start_time_sec,
3468 3468 start_used_bytes);
3469 3469
3470 3470 #if YOUNG_LIST_VERBOSE
3471 3471 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3472 3472 _young_list->print();
3473 3473 #endif // YOUNG_LIST_VERBOSE
3474 3474
3475 3475 if (g1_policy()->during_initial_mark_pause()) {
3476 3476 concurrent_mark()->checkpointRootsInitialPre();
3477 3477 }
3478 3478 perm_gen()->save_marks();
3479 3479
3480 3480 // We must do this before any possible evacuation that should propagate
3481 3481 // marks.
3482 3482 if (mark_in_progress()) {
3483 3483 double start_time_sec = os::elapsedTime();
3484 3484
3485 3485 _cm->drainAllSATBBuffers();
3486 3486 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0;
3487 3487 g1_policy()->record_satb_drain_time(finish_mark_ms);
3488 3488 }
3489 3489 // Record the number of elements currently on the mark stack, so we
3490 3490 // only iterate over these. (Since evacuation may add to the mark
3491 3491 // stack, doing more exposes race conditions.) If no mark is in
3492 3492 // progress, this will be zero.
3493 3493 _cm->set_oops_do_bound();
3494 3494
3495 3495 if (mark_in_progress()) {
3496 3496 concurrent_mark()->newCSet();
3497 3497 }
3498 3498
3499 3499 #if YOUNG_LIST_VERBOSE
3500 3500 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3501 3501 _young_list->print();
3502 3502 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3503 3503 #endif // YOUNG_LIST_VERBOSE
3504 3504
3505 3505 g1_policy()->choose_collection_set(target_pause_time_ms);
3506 3506
3507 3507 if (_hr_printer.is_active()) {
3508 3508 HeapRegion* hr = g1_policy()->collection_set();
3509 3509 while (hr != NULL) {
3510 3510 G1HRPrinter::RegionType type;
3511 3511 if (!hr->is_young()) {
3512 3512 type = G1HRPrinter::Old;
3513 3513 } else if (hr->is_survivor()) {
3514 3514 type = G1HRPrinter::Survivor;
3515 3515 } else {
3516 3516 type = G1HRPrinter::Eden;
3517 3517 }
3518 3518 _hr_printer.cset(hr);
3519 3519 hr = hr->next_in_collection_set();
3520 3520 }
3521 3521 }
3522 3522
3523 3523 // We have chosen the complete collection set. If marking is
3524 3524 // active then, we clear the region fields of any of the
3525 3525 // concurrent marking tasks whose region fields point into
3526 3526 // the collection set as these values will become stale. This
3527 3527 // will cause the owning marking threads to claim a new region
3528 3528 // when marking restarts.
3529 3529 if (mark_in_progress()) {
3530 3530 concurrent_mark()->reset_active_task_region_fields_in_cset();
3531 3531 }
3532 3532
3533 3533 #ifdef ASSERT
3534 3534 VerifyCSetClosure cl;
3535 3535 collection_set_iterate(&cl);
3536 3536 #endif // ASSERT
3537 3537
3538 3538 setup_surviving_young_words();
3539 3539
3540 3540 // Initialize the GC alloc regions.
3541 3541 init_gc_alloc_regions();
3542 3542
3543 3543 // Actually do the work...
3544 3544 evacuate_collection_set();
3545 3545
3546 3546 free_collection_set(g1_policy()->collection_set());
3547 3547 g1_policy()->clear_collection_set();
3548 3548
3549 3549 cleanup_surviving_young_words();
3550 3550
3551 3551 // Start a new incremental collection set for the next pause.
3552 3552 g1_policy()->start_incremental_cset_building();
3553 3553
3554 3554 // Clear the _cset_fast_test bitmap in anticipation of adding
3555 3555 // regions to the incremental collection set for the next
3556 3556 // evacuation pause.
3557 3557 clear_cset_fast_test();
3558 3558
3559 3559 _young_list->reset_sampled_info();
3560 3560
3561 3561 // Don't check the whole heap at this point as the
3562 3562 // GC alloc regions from this pause have been tagged
3563 3563 // as survivors and moved on to the survivor list.
3564 3564 // Survivor regions will fail the !is_young() check.
3565 3565 assert(check_young_list_empty(false /* check_heap */),
3566 3566 "young list should be empty");
3567 3567
3568 3568 #if YOUNG_LIST_VERBOSE
3569 3569 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3570 3570 _young_list->print();
3571 3571 #endif // YOUNG_LIST_VERBOSE
3572 3572
3573 3573 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3574 3574 _young_list->first_survivor_region(),
3575 3575 _young_list->last_survivor_region());
3576 3576
3577 3577 _young_list->reset_auxilary_lists();
3578 3578
3579 3579 if (evacuation_failed()) {
3580 3580 _summary_bytes_used = recalculate_used();
3581 3581 } else {
3582 3582 // The "used" of the the collection set have already been subtracted
3583 3583 // when they were freed. Add in the bytes evacuated.
3584 3584 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3585 3585 }
3586 3586
3587 3587 if (g1_policy()->during_initial_mark_pause()) {
3588 3588 concurrent_mark()->checkpointRootsInitialPost();
3589 3589 set_marking_started();
3590 3590 // CAUTION: after the doConcurrentMark() call below,
3591 3591 // the concurrent marking thread(s) could be running
3592 3592 // concurrently with us. Make sure that anything after
3593 3593 // this point does not assume that we are the only GC thread
3594 3594 // running. Note: of course, the actual marking work will
3595 3595 // not start until the safepoint itself is released in
3596 3596 // ConcurrentGCThread::safepoint_desynchronize().
3597 3597 doConcurrentMark();
3598 3598 }
3599 3599
3600 3600 allocate_dummy_regions();
3601 3601
3602 3602 #if YOUNG_LIST_VERBOSE
3603 3603 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3604 3604 _young_list->print();
3605 3605 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3606 3606 #endif // YOUNG_LIST_VERBOSE
3607 3607
3608 3608 init_mutator_alloc_region();
3609 3609
3610 3610 {
3611 3611 size_t expand_bytes = g1_policy()->expansion_amount();
3612 3612 if (expand_bytes > 0) {
3613 3613 size_t bytes_before = capacity();
3614 3614 if (!expand(expand_bytes)) {
3615 3615 // We failed to expand the heap so let's verify that
3616 3616 // committed/uncommitted amount match the backing store
3617 3617 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3618 3618 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3619 3619 }
3620 3620 }
3621 3621 }
3622 3622
3623 3623 double end_time_sec = os::elapsedTime();
3624 3624 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS;
3625 3625 g1_policy()->record_pause_time_ms(pause_time_ms);
3626 3626 g1_policy()->record_collection_pause_end();
3627 3627
3628 3628 MemoryService::track_memory_usage();
3629 3629
3630 3630 // In prepare_for_verify() below we'll need to scan the deferred
3631 3631 // update buffers to bring the RSets up-to-date if
3632 3632 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3633 3633 // the update buffers we'll probably need to scan cards on the
3634 3634 // regions we just allocated to (i.e., the GC alloc
3635 3635 // regions). However, during the last GC we called
3636 3636 // set_saved_mark() on all the GC alloc regions, so card
3637 3637 // scanning might skip the [saved_mark_word()...top()] area of
3638 3638 // those regions (i.e., the area we allocated objects into
3639 3639 // during the last GC). But it shouldn't. Given that
3640 3640 // saved_mark_word() is conditional on whether the GC time stamp
3641 3641 // on the region is current or not, by incrementing the GC time
3642 3642 // stamp here we invalidate all the GC time stamps on all the
3643 3643 // regions and saved_mark_word() will simply return top() for
3644 3644 // all the regions. This is a nicer way of ensuring this rather
3645 3645 // than iterating over the regions and fixing them. In fact, the
3646 3646 // GC time stamp increment here also ensures that
3647 3647 // saved_mark_word() will return top() between pauses, i.e.,
3648 3648 // during concurrent refinement. So we don't need the
3649 3649 // is_gc_active() check to decided which top to use when
3650 3650 // scanning cards (see CR 7039627).
3651 3651 increment_gc_time_stamp();
3652 3652
3653 3653 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
3654 3654 HandleMark hm; // Discard invalid handles created during verification
3655 3655 gclog_or_tty->print(" VerifyAfterGC:");
3656 3656 prepare_for_verify();
3657 3657 Universe::verify(/* allow dirty */ true,
3658 3658 /* silent */ false,
3659 3659 /* option */ VerifyOption_G1UsePrevMarking);
3660 3660 }
3661 3661
3662 3662 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
3663 3663 ref_processor_stw()->verify_no_references_recorded();
3664 3664
3665 3665 // CM reference discovery will be re-enabled if necessary.
3666 3666 }
3667 3667
3668 3668 {
3669 3669 size_t expand_bytes = g1_policy()->expansion_amount();
3670 3670 if (expand_bytes > 0) {
3671 3671 size_t bytes_before = capacity();
3672 3672 // No need for an ergo verbose message here,
3673 3673 // expansion_amount() does this when it returns a value > 0.
3674 3674 if (!expand(expand_bytes)) {
3675 3675 // We failed to expand the heap so let's verify that
3676 3676 // committed/uncommitted amount match the backing store
3677 3677 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3678 3678 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3679 3679 }
3680 3680 }
3681 3681 }
3682 3682
3683 3683 // We should do this after we potentially expand the heap so
3684 3684 // that all the COMMIT events are generated before the end GC
3685 3685 // event, and after we retire the GC alloc regions so that all
3686 3686 // RETIRE events are generated before the end GC event.
3687 3687 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
3688 3688
3689 3689 // We have to do this after we decide whether to expand the heap or not.
3690 3690 g1_policy()->print_heap_transition();
3691 3691
3692 3692 if (mark_in_progress()) {
3693 3693 concurrent_mark()->update_g1_committed();
3694 3694 }
3695 3695
3696 3696 #ifdef TRACESPINNING
3697 3697 ParallelTaskTerminator::print_termination_counts();
3698 3698 #endif
3699 3699
3700 3700 gc_epilogue(false);
3701 3701 }
3702 3702
3703 3703 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) {
3704 3704 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum);
3705 3705 print_tracing_info();
3706 3706 vm_exit(-1);
3707 3707 }
3708 3708 }
3709 3709
3710 3710 _hrs.verify_optional();
3711 3711 verify_region_sets_optional();
3712 3712
3713 3713 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
3714 3714 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
3715 3715
3716 3716 if (PrintHeapAtGC) {
3717 3717 Universe::print_heap_after_gc();
3718 3718 }
3719 3719 g1mm()->update_sizes();
3720 3720
3721 3721 if (G1SummarizeRSetStats &&
3722 3722 (G1SummarizeRSetStatsPeriod > 0) &&
3723 3723 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
3724 3724 g1_rem_set()->print_summary_info();
3725 3725 }
3726 3726
3727 3727 return true;
3728 3728 }
3729 3729
3730 3730 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
3731 3731 {
3732 3732 size_t gclab_word_size;
3733 3733 switch (purpose) {
3734 3734 case GCAllocForSurvived:
3735 3735 gclab_word_size = YoungPLABSize;
3736 3736 break;
3737 3737 case GCAllocForTenured:
3738 3738 gclab_word_size = OldPLABSize;
3739 3739 break;
3740 3740 default:
3741 3741 assert(false, "unknown GCAllocPurpose");
3742 3742 gclab_word_size = OldPLABSize;
3743 3743 break;
3744 3744 }
3745 3745 return gclab_word_size;
3746 3746 }
3747 3747
3748 3748 void G1CollectedHeap::init_mutator_alloc_region() {
3749 3749 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
3750 3750 _mutator_alloc_region.init();
3751 3751 }
3752 3752
3753 3753 void G1CollectedHeap::release_mutator_alloc_region() {
3754 3754 _mutator_alloc_region.release();
3755 3755 assert(_mutator_alloc_region.get() == NULL, "post-condition");
3756 3756 }
3757 3757
3758 3758 void G1CollectedHeap::init_gc_alloc_regions() {
3759 3759 assert_at_safepoint(true /* should_be_vm_thread */);
3760 3760
3761 3761 _survivor_gc_alloc_region.init();
3762 3762 _old_gc_alloc_region.init();
3763 3763 HeapRegion* retained_region = _retained_old_gc_alloc_region;
3764 3764 _retained_old_gc_alloc_region = NULL;
3765 3765
3766 3766 // We will discard the current GC alloc region if:
3767 3767 // a) it's in the collection set (it can happen!),
3768 3768 // b) it's already full (no point in using it),
3769 3769 // c) it's empty (this means that it was emptied during
3770 3770 // a cleanup and it should be on the free list now), or
3771 3771 // d) it's humongous (this means that it was emptied
3772 3772 // during a cleanup and was added to the free list, but
3773 3773 // has been subseqently used to allocate a humongous
3774 3774 // object that may be less than the region size).
3775 3775 if (retained_region != NULL &&
3776 3776 !retained_region->in_collection_set() &&
3777 3777 !(retained_region->top() == retained_region->end()) &&
3778 3778 !retained_region->is_empty() &&
3779 3779 !retained_region->isHumongous()) {
3780 3780 retained_region->set_saved_mark();
3781 3781 _old_gc_alloc_region.set(retained_region);
3782 3782 _hr_printer.reuse(retained_region);
3783 3783 }
3784 3784 }
3785 3785
3786 3786 void G1CollectedHeap::release_gc_alloc_regions() {
3787 3787 _survivor_gc_alloc_region.release();
3788 3788 // If we have an old GC alloc region to release, we'll save it in
3789 3789 // _retained_old_gc_alloc_region. If we don't
3790 3790 // _retained_old_gc_alloc_region will become NULL. This is what we
3791 3791 // want either way so no reason to check explicitly for either
3792 3792 // condition.
3793 3793 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
3794 3794 }
3795 3795
3796 3796 void G1CollectedHeap::abandon_gc_alloc_regions() {
3797 3797 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
3798 3798 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
3799 3799 _retained_old_gc_alloc_region = NULL;
3800 3800 }
3801 3801
3802 3802 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
3803 3803 _drain_in_progress = false;
3804 3804 set_evac_failure_closure(cl);
3805 3805 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
3806 3806 }
3807 3807
3808 3808 void G1CollectedHeap::finalize_for_evac_failure() {
3809 3809 assert(_evac_failure_scan_stack != NULL &&
3810 3810 _evac_failure_scan_stack->length() == 0,
3811 3811 "Postcondition");
3812 3812 assert(!_drain_in_progress, "Postcondition");
3813 3813 delete _evac_failure_scan_stack;
3814 3814 _evac_failure_scan_stack = NULL;
3815 3815 }
3816 3816
3817 3817 class UpdateRSetDeferred : public OopsInHeapRegionClosure {
3818 3818 private:
3819 3819 G1CollectedHeap* _g1;
3820 3820 DirtyCardQueue *_dcq;
3821 3821 CardTableModRefBS* _ct_bs;
3822 3822
3823 3823 public:
3824 3824 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
3825 3825 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {}
3826 3826
3827 3827 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
3828 3828 virtual void do_oop( oop* p) { do_oop_work(p); }
3829 3829 template <class T> void do_oop_work(T* p) {
3830 3830 assert(_from->is_in_reserved(p), "paranoia");
3831 3831 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) &&
3832 3832 !_from->is_survivor()) {
3833 3833 size_t card_index = _ct_bs->index_for(p);
3834 3834 if (_ct_bs->mark_card_deferred(card_index)) {
3835 3835 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
3836 3836 }
3837 3837 }
3838 3838 }
3839 3839 };
3840 3840
3841 3841 class RemoveSelfPointerClosure: public ObjectClosure {
3842 3842 private:
3843 3843 G1CollectedHeap* _g1;
3844 3844 ConcurrentMark* _cm;
3845 3845 HeapRegion* _hr;
3846 3846 size_t _prev_marked_bytes;
3847 3847 size_t _next_marked_bytes;
3848 3848 OopsInHeapRegionClosure *_cl;
3849 3849 public:
3850 3850 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr,
3851 3851 OopsInHeapRegionClosure* cl) :
3852 3852 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0),
3853 3853 _next_marked_bytes(0), _cl(cl) {}
3854 3854
3855 3855 size_t prev_marked_bytes() { return _prev_marked_bytes; }
3856 3856 size_t next_marked_bytes() { return _next_marked_bytes; }
3857 3857
3858 3858 // <original comment>
3859 3859 // The original idea here was to coalesce evacuated and dead objects.
3860 3860 // However that caused complications with the block offset table (BOT).
3861 3861 // In particular if there were two TLABs, one of them partially refined.
3862 3862 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~|
3863 3863 // The BOT entries of the unrefined part of TLAB_2 point to the start
3864 3864 // of TLAB_2. If the last object of the TLAB_1 and the first object
3865 3865 // of TLAB_2 are coalesced, then the cards of the unrefined part
3866 3866 // would point into middle of the filler object.
3867 3867 // The current approach is to not coalesce and leave the BOT contents intact.
3868 3868 // </original comment>
3869 3869 //
3870 3870 // We now reset the BOT when we start the object iteration over the
3871 3871 // region and refine its entries for every object we come across. So
3872 3872 // the above comment is not really relevant and we should be able
3873 3873 // to coalesce dead objects if we want to.
3874 3874 void do_object(oop obj) {
3875 3875 HeapWord* obj_addr = (HeapWord*) obj;
↓ open down ↓ |
3875 lines elided |
↑ open up ↑ |
3876 3876 assert(_hr->is_in(obj_addr), "sanity");
3877 3877 size_t obj_size = obj->size();
3878 3878 _hr->update_bot_for_object(obj_addr, obj_size);
3879 3879 if (obj->is_forwarded() && obj->forwardee() == obj) {
3880 3880 // The object failed to move.
3881 3881 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs.");
3882 3882 _cm->markPrev(obj);
3883 3883 assert(_cm->isPrevMarked(obj), "Should be marked!");
3884 3884 _prev_marked_bytes += (obj_size * HeapWordSize);
3885 3885 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) {
3886 - _cm->markAndGrayObjectIfNecessary(obj);
3886 + _cm->markAndGrayObjectIfNecessary(obj, 0 /* worker_i */);
3887 3887 }
3888 3888 obj->set_mark(markOopDesc::prototype());
3889 3889 // While we were processing RSet buffers during the
3890 3890 // collection, we actually didn't scan any cards on the
3891 3891 // collection set, since we didn't want to update remebered
3892 3892 // sets with entries that point into the collection set, given
3893 3893 // that live objects fromthe collection set are about to move
3894 3894 // and such entries will be stale very soon. This change also
3895 3895 // dealt with a reliability issue which involved scanning a
3896 3896 // card in the collection set and coming across an array that
3897 3897 // was being chunked and looking malformed. The problem is
3898 3898 // that, if evacuation fails, we might have remembered set
3899 3899 // entries missing given that we skipped cards on the
3900 3900 // collection set. So, we'll recreate such entries now.
3901 3901 obj->oop_iterate(_cl);
3902 3902 assert(_cm->isPrevMarked(obj), "Should be marked!");
3903 3903 } else {
3904 3904 // The object has been either evacuated or is dead. Fill it with a
3905 3905 // dummy object.
3906 3906 MemRegion mr((HeapWord*)obj, obj_size);
3907 3907 CollectedHeap::fill_with_object(mr);
3908 3908 _cm->clearRangeBothMaps(mr);
3909 3909 }
3910 3910 }
3911 3911 };
3912 3912
3913 3913 void G1CollectedHeap::remove_self_forwarding_pointers() {
3914 3914 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set());
3915 3915 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set());
3916 3916 UpdateRSetDeferred deferred_update(_g1h, &dcq);
3917 3917 OopsInHeapRegionClosure *cl;
3918 3918 if (G1DeferredRSUpdate) {
3919 3919 cl = &deferred_update;
3920 3920 } else {
3921 3921 cl = &immediate_update;
3922 3922 }
3923 3923 HeapRegion* cur = g1_policy()->collection_set();
3924 3924 while (cur != NULL) {
3925 3925 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
3926 3926 assert(!cur->isHumongous(), "sanity");
3927 3927
3928 3928 if (cur->evacuation_failed()) {
3929 3929 assert(cur->in_collection_set(), "bad CS");
3930 3930 RemoveSelfPointerClosure rspc(_g1h, cur, cl);
3931 3931
3932 3932 // In the common case we make sure that this is done when the
3933 3933 // region is freed so that it is "ready-to-go" when it's
3934 3934 // re-allocated. However, when evacuation failure happens, a
3935 3935 // region will remain in the heap and might ultimately be added
3936 3936 // to a CSet in the future. So we have to be careful here and
3937 3937 // make sure the region's RSet is ready for parallel iteration
3938 3938 // whenever this might be required in the future.
3939 3939 cur->rem_set()->reset_for_par_iteration();
3940 3940 cur->reset_bot();
3941 3941 cl->set_region(cur);
3942 3942 cur->object_iterate(&rspc);
3943 3943
3944 3944 // A number of manipulations to make the TAMS be the current top,
3945 3945 // and the marked bytes be the ones observed in the iteration.
3946 3946 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) {
3947 3947 // The comments below are the postconditions achieved by the
3948 3948 // calls. Note especially the last such condition, which says that
3949 3949 // the count of marked bytes has been properly restored.
3950 3950 cur->note_start_of_marking(false);
3951 3951 // _next_top_at_mark_start == top, _next_marked_bytes == 0
3952 3952 cur->add_to_marked_bytes(rspc.prev_marked_bytes());
3953 3953 // _next_marked_bytes == prev_marked_bytes.
3954 3954 cur->note_end_of_marking();
3955 3955 // _prev_top_at_mark_start == top(),
3956 3956 // _prev_marked_bytes == prev_marked_bytes
3957 3957 }
3958 3958 // If there is no mark in progress, we modified the _next variables
3959 3959 // above needlessly, but harmlessly.
3960 3960 if (_g1h->mark_in_progress()) {
3961 3961 cur->note_start_of_marking(false);
3962 3962 // _next_top_at_mark_start == top, _next_marked_bytes == 0
3963 3963 // _next_marked_bytes == next_marked_bytes.
3964 3964 }
3965 3965
3966 3966 // Now make sure the region has the right index in the sorted array.
3967 3967 g1_policy()->note_change_in_marked_bytes(cur);
3968 3968 }
3969 3969 cur = cur->next_in_collection_set();
3970 3970 }
3971 3971 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
3972 3972
3973 3973 // Now restore saved marks, if any.
3974 3974 if (_objs_with_preserved_marks != NULL) {
3975 3975 assert(_preserved_marks_of_objs != NULL, "Both or none.");
3976 3976 guarantee(_objs_with_preserved_marks->length() ==
3977 3977 _preserved_marks_of_objs->length(), "Both or none.");
3978 3978 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
3979 3979 oop obj = _objs_with_preserved_marks->at(i);
3980 3980 markOop m = _preserved_marks_of_objs->at(i);
3981 3981 obj->set_mark(m);
3982 3982 }
3983 3983 // Delete the preserved marks growable arrays (allocated on the C heap).
3984 3984 delete _objs_with_preserved_marks;
3985 3985 delete _preserved_marks_of_objs;
3986 3986 _objs_with_preserved_marks = NULL;
3987 3987 _preserved_marks_of_objs = NULL;
3988 3988 }
3989 3989 }
3990 3990
3991 3991 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
3992 3992 _evac_failure_scan_stack->push(obj);
3993 3993 }
3994 3994
3995 3995 void G1CollectedHeap::drain_evac_failure_scan_stack() {
3996 3996 assert(_evac_failure_scan_stack != NULL, "precondition");
3997 3997
↓ open down ↓ |
101 lines elided |
↑ open up ↑ |
3998 3998 while (_evac_failure_scan_stack->length() > 0) {
3999 3999 oop obj = _evac_failure_scan_stack->pop();
4000 4000 _evac_failure_closure->set_region(heap_region_containing(obj));
4001 4001 obj->oop_iterate_backwards(_evac_failure_closure);
4002 4002 }
4003 4003 }
4004 4004
4005 4005 oop
4006 4006 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4007 4007 oop old,
4008 - bool should_mark_root) {
4008 + bool should_mark_root,
4009 + int worker_i) {
4009 4010 assert(obj_in_cs(old),
4010 4011 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4011 4012 (HeapWord*) old));
4012 4013 markOop m = old->mark();
4013 4014 oop forward_ptr = old->forward_to_atomic(old);
4014 4015 if (forward_ptr == NULL) {
4015 4016 // Forward-to-self succeeded.
4016 4017
4017 4018 // should_mark_root will be true when this routine is called
4018 4019 // from a root scanning closure during an initial mark pause.
4019 4020 // In this case the thread that succeeds in self-forwarding the
4020 4021 // object is also responsible for marking the object.
4021 4022 if (should_mark_root) {
4022 4023 assert(!oopDesc::is_null(old), "shouldn't be");
4023 - _cm->grayRoot(old);
4024 + _cm->grayRoot(old, worker_i);
4024 4025 }
4025 4026
4026 4027 if (_evac_failure_closure != cl) {
4027 4028 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4028 4029 assert(!_drain_in_progress,
4029 4030 "Should only be true while someone holds the lock.");
4030 4031 // Set the global evac-failure closure to the current thread's.
4031 4032 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4032 4033 set_evac_failure_closure(cl);
4033 4034 // Now do the common part.
4034 4035 handle_evacuation_failure_common(old, m);
4035 4036 // Reset to NULL.
4036 4037 set_evac_failure_closure(NULL);
4037 4038 } else {
4038 4039 // The lock is already held, and this is recursive.
4039 4040 assert(_drain_in_progress, "This should only be the recursive case.");
4040 4041 handle_evacuation_failure_common(old, m);
4041 4042 }
4042 4043 return old;
4043 4044 } else {
4044 4045 // Forward-to-self failed. Either someone else managed to allocate
4045 4046 // space for this object (old != forward_ptr) or they beat us in
4046 4047 // self-forwarding it (old == forward_ptr).
4047 4048 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4048 4049 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4049 4050 "should not be in the CSet",
4050 4051 (HeapWord*) old, (HeapWord*) forward_ptr));
4051 4052 return forward_ptr;
4052 4053 }
4053 4054 }
4054 4055
4055 4056 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4056 4057 set_evacuation_failed(true);
4057 4058
4058 4059 preserve_mark_if_necessary(old, m);
4059 4060
4060 4061 HeapRegion* r = heap_region_containing(old);
4061 4062 if (!r->evacuation_failed()) {
4062 4063 r->set_evacuation_failed(true);
4063 4064 _hr_printer.evac_failure(r);
4064 4065 }
4065 4066
4066 4067 push_on_evac_failure_scan_stack(old);
4067 4068
4068 4069 if (!_drain_in_progress) {
4069 4070 // prevent recursion in copy_to_survivor_space()
4070 4071 _drain_in_progress = true;
4071 4072 drain_evac_failure_scan_stack();
4072 4073 _drain_in_progress = false;
4073 4074 }
4074 4075 }
4075 4076
4076 4077 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4077 4078 assert(evacuation_failed(), "Oversaving!");
4078 4079 // We want to call the "for_promotion_failure" version only in the
4079 4080 // case of a promotion failure.
4080 4081 if (m->must_be_preserved_for_promotion_failure(obj)) {
4081 4082 if (_objs_with_preserved_marks == NULL) {
4082 4083 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4083 4084 _objs_with_preserved_marks =
4084 4085 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
4085 4086 _preserved_marks_of_objs =
4086 4087 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
4087 4088 }
4088 4089 _objs_with_preserved_marks->push(obj);
4089 4090 _preserved_marks_of_objs->push(m);
4090 4091 }
4091 4092 }
4092 4093
4093 4094 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4094 4095 size_t word_size) {
4095 4096 if (purpose == GCAllocForSurvived) {
4096 4097 HeapWord* result = survivor_attempt_allocation(word_size);
4097 4098 if (result != NULL) {
4098 4099 return result;
4099 4100 } else {
4100 4101 // Let's try to allocate in the old gen in case we can fit the
4101 4102 // object there.
4102 4103 return old_attempt_allocation(word_size);
4103 4104 }
4104 4105 } else {
4105 4106 assert(purpose == GCAllocForTenured, "sanity");
4106 4107 HeapWord* result = old_attempt_allocation(word_size);
4107 4108 if (result != NULL) {
4108 4109 return result;
4109 4110 } else {
4110 4111 // Let's try to allocate in the survivors in case we can fit the
4111 4112 // object there.
4112 4113 return survivor_attempt_allocation(word_size);
4113 4114 }
4114 4115 }
4115 4116
4116 4117 ShouldNotReachHere();
4117 4118 // Trying to keep some compilers happy.
4118 4119 return NULL;
↓ open down ↓ |
85 lines elided |
↑ open up ↑ |
4119 4120 }
4120 4121
4121 4122 #ifndef PRODUCT
4122 4123 bool GCLabBitMapClosure::do_bit(size_t offset) {
4123 4124 HeapWord* addr = _bitmap->offsetToHeapWord(offset);
4124 4125 guarantee(_cm->isMarked(oop(addr)), "it should be!");
4125 4126 return true;
4126 4127 }
4127 4128 #endif // PRODUCT
4128 4129
4129 -G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4130 +G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size, int worker_i) :
4130 4131 ParGCAllocBuffer(gclab_word_size),
4131 4132 _should_mark_objects(false),
4132 4133 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size),
4134 + _worker_i(worker_i),
4133 4135 _retired(false)
4134 4136 {
4135 4137 //_should_mark_objects is set to true when G1ParCopyHelper needs to
4136 4138 // mark the forwarded location of an evacuated object.
4137 4139 // We set _should_mark_objects to true if marking is active, i.e. when we
4138 4140 // need to propagate a mark, or during an initial mark pause, i.e. when we
4139 4141 // need to mark objects immediately reachable by the roots.
4140 4142 if (G1CollectedHeap::heap()->mark_in_progress() ||
4141 4143 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) {
4142 4144 _should_mark_objects = true;
4143 4145 }
↓ open down ↓ |
1 lines elided |
↑ open up ↑ |
4144 4146 }
4145 4147
4146 4148 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num)
4147 4149 : _g1h(g1h),
4148 4150 _refs(g1h->task_queue(queue_num)),
4149 4151 _dcq(&g1h->dirty_card_queue_set()),
4150 4152 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4151 4153 _g1_rem(g1h->g1_rem_set()),
4152 4154 _hash_seed(17), _queue_num(queue_num),
4153 4155 _term_attempts(0),
4154 - _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4155 - _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4156 + _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived), queue_num),
4157 + _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured), queue_num),
4156 4158 _age_table(false),
4157 4159 _strong_roots_time(0), _term_time(0),
4158 4160 _alloc_buffer_waste(0), _undo_waste(0)
4159 4161 {
4160 4162 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4161 4163 // we "sacrifice" entry 0 to keep track of surviving bytes for
4162 4164 // non-young regions (where the age is -1)
4163 4165 // We also add a few elements at the beginning and at the end in
4164 4166 // an attempt to eliminate cache contention
4165 4167 size_t real_length = 1 + _g1h->g1_policy()->young_cset_length();
4166 4168 size_t array_length = PADDING_ELEM_NUM +
4167 4169 real_length +
4168 4170 PADDING_ELEM_NUM;
4169 4171 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length);
4170 4172 if (_surviving_young_words_base == NULL)
4171 4173 vm_exit_out_of_memory(array_length * sizeof(size_t),
4172 4174 "Not enough space for young surv histo.");
4173 4175 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4174 4176 memset(_surviving_young_words, 0, real_length * sizeof(size_t));
4175 4177
4176 4178 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4177 4179 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4178 4180
4179 4181 _start = os::elapsedTime();
4180 4182 }
4181 4183
4182 4184 void
4183 4185 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4184 4186 {
4185 4187 st->print_raw_cr("GC Termination Stats");
4186 4188 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4187 4189 " ------waste (KiB)------");
4188 4190 st->print_raw_cr("thr ms ms % ms % attempts"
4189 4191 " total alloc undo");
4190 4192 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4191 4193 " ------- ------- -------");
4192 4194 }
4193 4195
4194 4196 void
4195 4197 G1ParScanThreadState::print_termination_stats(int i,
4196 4198 outputStream* const st) const
4197 4199 {
4198 4200 const double elapsed_ms = elapsed_time() * 1000.0;
4199 4201 const double s_roots_ms = strong_roots_time() * 1000.0;
4200 4202 const double term_ms = term_time() * 1000.0;
4201 4203 st->print_cr("%3d %9.2f %9.2f %6.2f "
4202 4204 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4203 4205 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4204 4206 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4205 4207 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4206 4208 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4207 4209 alloc_buffer_waste() * HeapWordSize / K,
4208 4210 undo_waste() * HeapWordSize / K);
4209 4211 }
4210 4212
4211 4213 #ifdef ASSERT
4212 4214 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4213 4215 assert(ref != NULL, "invariant");
4214 4216 assert(UseCompressedOops, "sanity");
4215 4217 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4216 4218 oop p = oopDesc::load_decode_heap_oop(ref);
4217 4219 assert(_g1h->is_in_g1_reserved(p),
4218 4220 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4219 4221 return true;
4220 4222 }
4221 4223
4222 4224 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4223 4225 assert(ref != NULL, "invariant");
4224 4226 if (has_partial_array_mask(ref)) {
4225 4227 // Must be in the collection set--it's already been copied.
4226 4228 oop p = clear_partial_array_mask(ref);
4227 4229 assert(_g1h->obj_in_cs(p),
4228 4230 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4229 4231 } else {
4230 4232 oop p = oopDesc::load_decode_heap_oop(ref);
4231 4233 assert(_g1h->is_in_g1_reserved(p),
4232 4234 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4233 4235 }
4234 4236 return true;
4235 4237 }
4236 4238
4237 4239 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4238 4240 if (ref.is_narrow()) {
4239 4241 return verify_ref((narrowOop*) ref);
4240 4242 } else {
4241 4243 return verify_ref((oop*) ref);
4242 4244 }
4243 4245 }
4244 4246 #endif // ASSERT
4245 4247
4246 4248 void G1ParScanThreadState::trim_queue() {
4247 4249 assert(_evac_cl != NULL, "not set");
4248 4250 assert(_evac_failure_cl != NULL, "not set");
4249 4251 assert(_partial_scan_cl != NULL, "not set");
4250 4252
4251 4253 StarTask ref;
4252 4254 do {
4253 4255 // Drain the overflow stack first, so other threads can steal.
4254 4256 while (refs()->pop_overflow(ref)) {
4255 4257 deal_with_reference(ref);
4256 4258 }
↓ open down ↓ |
91 lines elided |
↑ open up ↑ |
4257 4259
4258 4260 while (refs()->pop_local(ref)) {
4259 4261 deal_with_reference(ref);
4260 4262 }
4261 4263 } while (!refs()->is_empty());
4262 4264 }
4263 4265
4264 4266 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) :
4265 4267 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4266 4268 _par_scan_state(par_scan_state),
4269 + _worker_i(par_scan_state->queue_num()),
4267 4270 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4268 4271 _mark_in_progress(_g1->mark_in_progress()) { }
4269 4272
4270 4273 template <class T> void G1ParCopyHelper::mark_object(T* p) {
4271 4274 // This is called from do_oop_work for objects that are not
4272 4275 // in the collection set. Objects in the collection set
4273 4276 // are marked after they have been evacuated.
4274 4277
4275 4278 T heap_oop = oopDesc::load_heap_oop(p);
4276 4279 if (!oopDesc::is_null(heap_oop)) {
4277 4280 oop obj = oopDesc::decode_heap_oop(heap_oop);
4278 4281 HeapWord* addr = (HeapWord*)obj;
4279 4282 if (_g1->is_in_g1_reserved(addr)) {
4280 - _cm->grayRoot(oop(addr));
4283 + _cm->grayRoot(oop(addr), _worker_i);
4281 4284 }
4282 4285 }
4283 4286 }
4284 4287
4285 4288 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root,
4286 4289 bool should_mark_copy) {
4287 4290 size_t word_sz = old->size();
4288 4291 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4289 4292 // +1 to make the -1 indexes valid...
4290 4293 int young_index = from_region->young_index_in_cset()+1;
4291 4294 assert( (from_region->is_young() && young_index > 0) ||
4292 4295 (!from_region->is_young() && young_index == 0), "invariant" );
4293 4296 G1CollectorPolicy* g1p = _g1->g1_policy();
4294 4297 markOop m = old->mark();
4295 4298 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
↓ open down ↓ |
5 lines elided |
↑ open up ↑ |
4296 4299 : m->age();
4297 4300 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4298 4301 word_sz);
4299 4302 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4300 4303 oop obj = oop(obj_ptr);
4301 4304
4302 4305 if (obj_ptr == NULL) {
4303 4306 // This will either forward-to-self, or detect that someone else has
4304 4307 // installed a forwarding pointer.
4305 4308 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4306 - return _g1->handle_evacuation_failure_par(cl, old, should_mark_root);
4309 + return _g1->handle_evacuation_failure_par(cl, old, should_mark_root, _worker_i);
4307 4310 }
4308 4311
4309 4312 // We're going to allocate linearly, so might as well prefetch ahead.
4310 4313 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4311 4314
4312 4315 oop forward_ptr = old->forward_to_atomic(obj);
4313 4316 if (forward_ptr == NULL) {
4314 4317 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4315 4318 if (g1p->track_object_age(alloc_purpose)) {
4316 4319 // We could simply do obj->incr_age(). However, this causes a
4317 4320 // performance issue. obj->incr_age() will first check whether
4318 4321 // the object has a displaced mark by checking its mark word;
4319 4322 // getting the mark word from the new location of the object
4320 4323 // stalls. So, given that we already have the mark word and we
4321 4324 // are about to install it anyway, it's better to increase the
4322 4325 // age on the mark word, when the object does not have a
4323 4326 // displaced mark word. We're not expecting many objects to have
4324 4327 // a displaced marked word, so that case is not optimized
4325 4328 // further (it could be...) and we simply call obj->incr_age().
4326 4329
4327 4330 if (m->has_displaced_mark_helper()) {
4328 4331 // in this case, we have to install the mark word first,
4329 4332 // otherwise obj looks to be forwarded (the old mark word,
4330 4333 // which contains the forward pointer, was copied)
4331 4334 obj->set_mark(m);
4332 4335 obj->incr_age();
4333 4336 } else {
4334 4337 m = m->incr_age();
↓ open down ↓ |
18 lines elided |
↑ open up ↑ |
4335 4338 obj->set_mark(m);
4336 4339 }
4337 4340 _par_scan_state->age_table()->add(obj, word_sz);
4338 4341 } else {
4339 4342 obj->set_mark(m);
4340 4343 }
4341 4344
4342 4345 // Mark the evacuated object or propagate "next" mark bit
4343 4346 if (should_mark_copy) {
4344 4347 if (!use_local_bitmaps ||
4345 - !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) {
4348 + !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr, word_sz)) {
4346 4349 // if we couldn't mark it on the local bitmap (this happens when
4347 4350 // the object was not allocated in the GCLab), we have to bite
4348 4351 // the bullet and do the standard parallel mark
4349 - _cm->markAndGrayObjectIfNecessary(obj);
4352 + _cm->markAndGrayObjectIfNecessary(obj, _worker_i);
4350 4353 }
4351 4354
4352 4355 if (_g1->isMarkedNext(old)) {
4353 4356 // Unmark the object's old location so that marking
4354 4357 // doesn't think the old object is alive.
4355 4358 _cm->nextMarkBitMap()->parClear((HeapWord*)old);
4359 +
4360 + // We could clear the count data for the old object here but
4361 + // currently we do not. Why don't we do this? The thread/task
4362 + // that marks a newly copied object is likely _not_ the thread/task
4363 + // that originally marked the old object. So, to clear the count
4364 + // data for the old object, we would have to scan the count
4365 + // data for all of the tasks (and clear the data for the old object
4366 + // in parallel with other threads adding to the count data). Even
4367 + // then we could clear a bit incorrectly (e.g. if the old object
4368 + // does not start or end on a card boundary). It's more important
4369 + // that we don't have missed bits that should've been set than
4370 + // having extra bits set.
4371 + //
4372 + // As a result the accumulated count data could be a superset
4373 + // of the data that is/would have been calculated by walking
4374 + // the marking bitmap.
4356 4375 }
4357 4376 }
4358 4377
4359 4378 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4360 4379 surv_young_words[young_index] += word_sz;
4361 4380
4362 4381 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4363 4382 arrayOop(old)->set_length(0);
4364 4383 oop* old_p = set_partial_array_mask(old);
4365 4384 _par_scan_state->push_on_queue(old_p);
4366 4385 } else {
4367 4386 // No point in using the slower heap_region_containing() method,
4368 4387 // given that we know obj is in the heap.
4369 4388 _scanner->set_region(_g1->heap_region_containing_raw(obj));
4370 4389 obj->oop_iterate_backwards(_scanner);
4371 4390 }
4372 4391 } else {
4373 4392 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4374 4393 obj = forward_ptr;
4375 4394 }
4376 4395 return obj;
4377 4396 }
4378 4397
4379 4398 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4380 4399 template <class T>
4381 4400 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4382 4401 ::do_oop_work(T* p) {
4383 4402 oop obj = oopDesc::load_decode_heap_oop(p);
4384 4403 assert(barrier != G1BarrierRS || obj != NULL,
4385 4404 "Precondition: G1BarrierRS implies obj is nonNull");
4386 4405
4387 4406 // Marking:
4388 4407 // If the object is in the collection set, then the thread
4389 4408 // that copies the object should mark, or propagate the
4390 4409 // mark to, the evacuated object.
4391 4410 // If the object is not in the collection set then we
4392 4411 // should call the mark_object() method depending on the
4393 4412 // value of the template parameter do_mark_object (which will
4394 4413 // be true for root scanning closures during an initial mark
4395 4414 // pause).
4396 4415 // The mark_object() method first checks whether the object
4397 4416 // is marked and, if not, attempts to mark the object.
4398 4417
4399 4418 // here the null check is implicit in the cset_fast_test() test
4400 4419 if (_g1->in_cset_fast_test(obj)) {
4401 4420 if (obj->is_forwarded()) {
4402 4421 oopDesc::encode_store_heap_oop(p, obj->forwardee());
4403 4422 // If we are a root scanning closure during an initial
4404 4423 // mark pause (i.e. do_mark_object will be true) then
4405 4424 // we also need to handle marking of roots in the
4406 4425 // event of an evacuation failure. In the event of an
4407 4426 // evacuation failure, the object is forwarded to itself
4408 4427 // and not copied. For root-scanning closures, the
4409 4428 // object would be marked after a successful self-forward
4410 4429 // but an object could be pointed to by both a root and non
4411 4430 // root location and be self-forwarded by a non-root-scanning
4412 4431 // closure. Therefore we also have to attempt to mark the
4413 4432 // self-forwarded root object here.
4414 4433 if (do_mark_object && obj->forwardee() == obj) {
4415 4434 mark_object(p);
4416 4435 }
4417 4436 } else {
4418 4437 // During an initial mark pause, objects that are pointed to
4419 4438 // by the roots need to be marked - even in the event of an
4420 4439 // evacuation failure. We pass the template parameter
4421 4440 // do_mark_object (which is true for root scanning closures
4422 4441 // during an initial mark pause) to copy_to_survivor_space
4423 4442 // which will pass it on to the evacuation failure handling
4424 4443 // code. The thread that successfully self-forwards a root
4425 4444 // object to itself is responsible for marking the object.
4426 4445 bool should_mark_root = do_mark_object;
4427 4446
4428 4447 // We need to mark the copied object if we're a root scanning
4429 4448 // closure during an initial mark pause (i.e. do_mark_object
4430 4449 // will be true), or the object is already marked and we need
4431 4450 // to propagate the mark to the evacuated copy.
↓ open down ↓ |
66 lines elided |
↑ open up ↑ |
4432 4451 bool should_mark_copy = do_mark_object ||
4433 4452 _during_initial_mark ||
4434 4453 (_mark_in_progress && !_g1->is_obj_ill(obj));
4435 4454
4436 4455 oop copy_oop = copy_to_survivor_space(obj, should_mark_root,
4437 4456 should_mark_copy);
4438 4457 oopDesc::encode_store_heap_oop(p, copy_oop);
4439 4458 }
4440 4459 // When scanning the RS, we only care about objs in CS.
4441 4460 if (barrier == G1BarrierRS) {
4442 - _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4461 + assert(_worker_i == _par_scan_state->queue_num(), "sanity");
4462 + _par_scan_state->update_rs(_from, p, _worker_i);
4443 4463 }
4444 4464 } else {
4445 4465 // The object is not in collection set. If we're a root scanning
4446 4466 // closure during an initial mark pause (i.e. do_mark_object will
4447 4467 // be true) then attempt to mark the object.
4448 4468 if (do_mark_object) {
4449 4469 mark_object(p);
4450 4470 }
4451 4471 }
4452 4472
4453 4473 if (barrier == G1BarrierEvac && obj != NULL) {
4454 - _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4474 + assert(_worker_i == _par_scan_state->queue_num(), "sanity");
4475 + _par_scan_state->update_rs(_from, p, _worker_i);
4455 4476 }
4456 4477
4457 4478 if (do_gen_barrier && obj != NULL) {
4458 4479 par_do_barrier(p);
4459 4480 }
4460 4481 }
4461 4482
4462 4483 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4463 4484 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4464 4485
4465 4486 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4466 4487 assert(has_partial_array_mask(p), "invariant");
4467 4488 oop old = clear_partial_array_mask(p);
4468 4489 assert(old->is_objArray(), "must be obj array");
4469 4490 assert(old->is_forwarded(), "must be forwarded");
4470 4491 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
4471 4492
4472 4493 objArrayOop obj = objArrayOop(old->forwardee());
4473 4494 assert((void*)old != (void*)old->forwardee(), "self forwarding here?");
4474 4495 // Process ParGCArrayScanChunk elements now
4475 4496 // and push the remainder back onto queue
4476 4497 int start = arrayOop(old)->length();
4477 4498 int end = obj->length();
4478 4499 int remainder = end - start;
4479 4500 assert(start <= end, "just checking");
4480 4501 if (remainder > 2 * ParGCArrayScanChunk) {
4481 4502 // Test above combines last partial chunk with a full chunk
4482 4503 end = start + ParGCArrayScanChunk;
4483 4504 arrayOop(old)->set_length(end);
4484 4505 // Push remainder.
4485 4506 oop* old_p = set_partial_array_mask(old);
4486 4507 assert(arrayOop(old)->length() < obj->length(), "Empty push?");
4487 4508 _par_scan_state->push_on_queue(old_p);
4488 4509 } else {
4489 4510 // Restore length so that the heap remains parsable in
4490 4511 // case of evacuation failure.
4491 4512 arrayOop(old)->set_length(end);
4492 4513 }
4493 4514 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4494 4515 // process our set of indices (include header in first chunk)
4495 4516 obj->oop_iterate_range(&_scanner, start, end);
4496 4517 }
4497 4518
4498 4519 class G1ParEvacuateFollowersClosure : public VoidClosure {
4499 4520 protected:
4500 4521 G1CollectedHeap* _g1h;
4501 4522 G1ParScanThreadState* _par_scan_state;
4502 4523 RefToScanQueueSet* _queues;
4503 4524 ParallelTaskTerminator* _terminator;
4504 4525
4505 4526 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4506 4527 RefToScanQueueSet* queues() { return _queues; }
4507 4528 ParallelTaskTerminator* terminator() { return _terminator; }
4508 4529
4509 4530 public:
4510 4531 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4511 4532 G1ParScanThreadState* par_scan_state,
4512 4533 RefToScanQueueSet* queues,
4513 4534 ParallelTaskTerminator* terminator)
4514 4535 : _g1h(g1h), _par_scan_state(par_scan_state),
4515 4536 _queues(queues), _terminator(terminator) {}
4516 4537
4517 4538 void do_void();
4518 4539
4519 4540 private:
4520 4541 inline bool offer_termination();
4521 4542 };
4522 4543
4523 4544 bool G1ParEvacuateFollowersClosure::offer_termination() {
4524 4545 G1ParScanThreadState* const pss = par_scan_state();
4525 4546 pss->start_term_time();
4526 4547 const bool res = terminator()->offer_termination();
4527 4548 pss->end_term_time();
4528 4549 return res;
4529 4550 }
4530 4551
4531 4552 void G1ParEvacuateFollowersClosure::do_void() {
4532 4553 StarTask stolen_task;
4533 4554 G1ParScanThreadState* const pss = par_scan_state();
4534 4555 pss->trim_queue();
4535 4556
4536 4557 do {
4537 4558 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4538 4559 assert(pss->verify_task(stolen_task), "sanity");
4539 4560 if (stolen_task.is_narrow()) {
4540 4561 pss->deal_with_reference((narrowOop*) stolen_task);
4541 4562 } else {
4542 4563 pss->deal_with_reference((oop*) stolen_task);
4543 4564 }
4544 4565
4545 4566 // We've just processed a reference and we might have made
4546 4567 // available new entries on the queues. So we have to make sure
4547 4568 // we drain the queues as necessary.
4548 4569 pss->trim_queue();
4549 4570 }
4550 4571 } while (!offer_termination());
4551 4572
4552 4573 pss->retire_alloc_buffers();
4553 4574 }
4554 4575
4555 4576 class G1ParTask : public AbstractGangTask {
4556 4577 protected:
4557 4578 G1CollectedHeap* _g1h;
4558 4579 RefToScanQueueSet *_queues;
4559 4580 ParallelTaskTerminator _terminator;
4560 4581 int _n_workers;
4561 4582
4562 4583 Mutex _stats_lock;
4563 4584 Mutex* stats_lock() { return &_stats_lock; }
4564 4585
4565 4586 size_t getNCards() {
4566 4587 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4567 4588 / G1BlockOffsetSharedArray::N_bytes;
4568 4589 }
4569 4590
4570 4591 public:
4571 4592 G1ParTask(G1CollectedHeap* g1h, int workers, RefToScanQueueSet *task_queues)
4572 4593 : AbstractGangTask("G1 collection"),
4573 4594 _g1h(g1h),
4574 4595 _queues(task_queues),
4575 4596 _terminator(workers, _queues),
4576 4597 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true),
4577 4598 _n_workers(workers)
4578 4599 {}
4579 4600
4580 4601 RefToScanQueueSet* queues() { return _queues; }
4581 4602
4582 4603 RefToScanQueue *work_queue(int i) {
4583 4604 return queues()->queue(i);
4584 4605 }
4585 4606
4586 4607 void work(int i) {
4587 4608 if (i >= _n_workers) return; // no work needed this round
4588 4609
4589 4610 double start_time_ms = os::elapsedTime() * 1000.0;
4590 4611 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms);
4591 4612
4592 4613 ResourceMark rm;
4593 4614 HandleMark hm;
4594 4615
4595 4616 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4596 4617
4597 4618 G1ParScanThreadState pss(_g1h, i);
4598 4619 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4599 4620 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4600 4621 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4601 4622
4602 4623 pss.set_evac_closure(&scan_evac_cl);
4603 4624 pss.set_evac_failure_closure(&evac_failure_cl);
4604 4625 pss.set_partial_scan_closure(&partial_scan_cl);
4605 4626
4606 4627 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4607 4628 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp);
4608 4629
4609 4630 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4610 4631 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp);
4611 4632
4612 4633 OopClosure* scan_root_cl = &only_scan_root_cl;
4613 4634 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl;
4614 4635
4615 4636 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4616 4637 // We also need to mark copied objects.
4617 4638 scan_root_cl = &scan_mark_root_cl;
4618 4639 scan_perm_cl = &scan_mark_perm_cl;
4619 4640 }
4620 4641
4621 4642 // The following closure is used to scan RSets looking for reference
4622 4643 // fields that point into the collection set. The actual field iteration
4623 4644 // is performed by a FilterIntoCSClosure, whose do_oop method calls the
4624 4645 // do_oop method of the following closure.
4625 4646 // Therefore we want to record the reference processor in the
4626 4647 // FilterIntoCSClosure. To do so we record the STW reference
4627 4648 // processor into the following closure and pass it to the
4628 4649 // FilterIntoCSClosure in HeapRegionDCTOC::walk_mem_region_with_cl.
4629 4650 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss, rp);
4630 4651
4631 4652 pss.start_strong_roots();
4632 4653 _g1h->g1_process_strong_roots(/* not collecting perm */ false,
4633 4654 SharedHeap::SO_AllClasses,
4634 4655 scan_root_cl,
4635 4656 &push_heap_rs_cl,
4636 4657 scan_perm_cl,
4637 4658 i);
4638 4659 pss.end_strong_roots();
4639 4660
4640 4661 {
4641 4662 double start = os::elapsedTime();
4642 4663 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4643 4664 evac.do_void();
4644 4665 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4645 4666 double term_ms = pss.term_time()*1000.0;
4646 4667 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms);
4647 4668 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts());
4648 4669 }
4649 4670 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4650 4671 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4651 4672
4652 4673 // Clean up any par-expanded rem sets.
4653 4674 HeapRegionRemSet::par_cleanup();
4654 4675
4655 4676 if (ParallelGCVerbose) {
4656 4677 MutexLocker x(stats_lock());
4657 4678 pss.print_termination_stats(i);
4658 4679 }
4659 4680
4660 4681 assert(pss.refs()->is_empty(), "should be empty");
4661 4682 double end_time_ms = os::elapsedTime() * 1000.0;
4662 4683 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms);
4663 4684 }
4664 4685 };
4665 4686
4666 4687 // *** Common G1 Evacuation Stuff
4667 4688
4668 4689 // This method is run in a GC worker.
4669 4690
4670 4691 void
4671 4692 G1CollectedHeap::
4672 4693 g1_process_strong_roots(bool collecting_perm_gen,
4673 4694 SharedHeap::ScanningOption so,
4674 4695 OopClosure* scan_non_heap_roots,
4675 4696 OopsInHeapRegionClosure* scan_rs,
4676 4697 OopsInGenClosure* scan_perm,
4677 4698 int worker_i) {
4678 4699
4679 4700 // First scan the strong roots, including the perm gen.
4680 4701 double ext_roots_start = os::elapsedTime();
4681 4702 double closure_app_time_sec = 0.0;
4682 4703
4683 4704 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4684 4705 BufferingOopsInGenClosure buf_scan_perm(scan_perm);
4685 4706 buf_scan_perm.set_generation(perm_gen());
4686 4707
4687 4708 // Walk the code cache w/o buffering, because StarTask cannot handle
4688 4709 // unaligned oop locations.
4689 4710 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true);
4690 4711
4691 4712 process_strong_roots(false, // no scoping; this is parallel code
4692 4713 collecting_perm_gen, so,
4693 4714 &buf_scan_non_heap_roots,
4694 4715 &eager_scan_code_roots,
4695 4716 &buf_scan_perm);
4696 4717
4697 4718 // Now the CM ref_processor roots.
4698 4719 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
4699 4720 // We need to treat the discovered reference lists of the
4700 4721 // concurrent mark ref processor as roots and keep entries
4701 4722 // (which are added by the marking threads) on them live
4702 4723 // until they can be processed at the end of marking.
4703 4724 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
4704 4725 }
4705 4726
4706 4727 // Finish up any enqueued closure apps (attributed as object copy time).
4707 4728 buf_scan_non_heap_roots.done();
4708 4729 buf_scan_perm.done();
4709 4730
4710 4731 double ext_roots_end = os::elapsedTime();
4711 4732
4712 4733 g1_policy()->reset_obj_copy_time(worker_i);
4713 4734 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() +
4714 4735 buf_scan_non_heap_roots.closure_app_seconds();
4715 4736 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
4716 4737
4717 4738 double ext_root_time_ms =
4718 4739 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0;
4719 4740
4720 4741 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
4721 4742
4722 4743 // Scan strong roots in mark stack.
4723 4744 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) {
4724 4745 concurrent_mark()->oops_do(scan_non_heap_roots);
4725 4746 }
4726 4747 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0;
4727 4748 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms);
4728 4749
4729 4750 // Now scan the complement of the collection set.
4730 4751 if (scan_rs != NULL) {
4731 4752 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
4732 4753 }
4733 4754
4734 4755 _process_strong_tasks->all_tasks_completed();
4735 4756 }
4736 4757
4737 4758 void
4738 4759 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
4739 4760 OopClosure* non_root_closure) {
4740 4761 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
4741 4762 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
4742 4763 }
4743 4764
4744 4765 // Weak Reference Processing support
4745 4766
4746 4767 // An always "is_alive" closure that is used to preserve referents.
4747 4768 // If the object is non-null then it's alive. Used in the preservation
4748 4769 // of referent objects that are pointed to by reference objects
4749 4770 // discovered by the CM ref processor.
4750 4771 class G1AlwaysAliveClosure: public BoolObjectClosure {
4751 4772 G1CollectedHeap* _g1;
4752 4773 public:
4753 4774 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4754 4775 void do_object(oop p) { assert(false, "Do not call."); }
4755 4776 bool do_object_b(oop p) {
4756 4777 if (p != NULL) {
4757 4778 return true;
4758 4779 }
4759 4780 return false;
4760 4781 }
4761 4782 };
4762 4783
4763 4784 bool G1STWIsAliveClosure::do_object_b(oop p) {
4764 4785 // An object is reachable if it is outside the collection set,
4765 4786 // or is inside and copied.
4766 4787 return !_g1->obj_in_cs(p) || p->is_forwarded();
4767 4788 }
4768 4789
4769 4790 // Non Copying Keep Alive closure
4770 4791 class G1KeepAliveClosure: public OopClosure {
4771 4792 G1CollectedHeap* _g1;
4772 4793 public:
4773 4794 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4774 4795 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
4775 4796 void do_oop( oop* p) {
4776 4797 oop obj = *p;
4777 4798
4778 4799 if (_g1->obj_in_cs(obj)) {
4779 4800 assert( obj->is_forwarded(), "invariant" );
4780 4801 *p = obj->forwardee();
4781 4802 }
4782 4803 }
4783 4804 };
4784 4805
4785 4806 // Copying Keep Alive closure - can be called from both
4786 4807 // serial and parallel code as long as different worker
4787 4808 // threads utilize different G1ParScanThreadState instances
4788 4809 // and different queues.
4789 4810
4790 4811 class G1CopyingKeepAliveClosure: public OopClosure {
4791 4812 G1CollectedHeap* _g1h;
4792 4813 OopClosure* _copy_non_heap_obj_cl;
4793 4814 OopsInHeapRegionClosure* _copy_perm_obj_cl;
4794 4815 G1ParScanThreadState* _par_scan_state;
4795 4816
4796 4817 public:
4797 4818 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
4798 4819 OopClosure* non_heap_obj_cl,
4799 4820 OopsInHeapRegionClosure* perm_obj_cl,
4800 4821 G1ParScanThreadState* pss):
4801 4822 _g1h(g1h),
4802 4823 _copy_non_heap_obj_cl(non_heap_obj_cl),
4803 4824 _copy_perm_obj_cl(perm_obj_cl),
4804 4825 _par_scan_state(pss)
4805 4826 {}
4806 4827
4807 4828 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
4808 4829 virtual void do_oop( oop* p) { do_oop_work(p); }
4809 4830
4810 4831 template <class T> void do_oop_work(T* p) {
4811 4832 oop obj = oopDesc::load_decode_heap_oop(p);
4812 4833
4813 4834 if (_g1h->obj_in_cs(obj)) {
4814 4835 // If the referent object has been forwarded (either copied
4815 4836 // to a new location or to itself in the event of an
4816 4837 // evacuation failure) then we need to update the reference
4817 4838 // field and, if both reference and referent are in the G1
4818 4839 // heap, update the RSet for the referent.
4819 4840 //
4820 4841 // If the referent has not been forwarded then we have to keep
4821 4842 // it alive by policy. Therefore we have copy the referent.
4822 4843 //
4823 4844 // If the reference field is in the G1 heap then we can push
4824 4845 // on the PSS queue. When the queue is drained (after each
4825 4846 // phase of reference processing) the object and it's followers
4826 4847 // will be copied, the reference field set to point to the
4827 4848 // new location, and the RSet updated. Otherwise we need to
4828 4849 // use the the non-heap or perm closures directly to copy
4829 4850 // the refernt object and update the pointer, while avoiding
4830 4851 // updating the RSet.
4831 4852
4832 4853 if (_g1h->is_in_g1_reserved(p)) {
4833 4854 _par_scan_state->push_on_queue(p);
4834 4855 } else {
4835 4856 // The reference field is not in the G1 heap.
4836 4857 if (_g1h->perm_gen()->is_in(p)) {
4837 4858 _copy_perm_obj_cl->do_oop(p);
4838 4859 } else {
4839 4860 _copy_non_heap_obj_cl->do_oop(p);
4840 4861 }
4841 4862 }
4842 4863 }
4843 4864 }
4844 4865 };
4845 4866
4846 4867 // Serial drain queue closure. Called as the 'complete_gc'
4847 4868 // closure for each discovered list in some of the
4848 4869 // reference processing phases.
4849 4870
4850 4871 class G1STWDrainQueueClosure: public VoidClosure {
4851 4872 protected:
4852 4873 G1CollectedHeap* _g1h;
4853 4874 G1ParScanThreadState* _par_scan_state;
4854 4875
4855 4876 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4856 4877
4857 4878 public:
4858 4879 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
4859 4880 _g1h(g1h),
4860 4881 _par_scan_state(pss)
4861 4882 { }
4862 4883
4863 4884 void do_void() {
4864 4885 G1ParScanThreadState* const pss = par_scan_state();
4865 4886 pss->trim_queue();
4866 4887 }
4867 4888 };
4868 4889
4869 4890 // Parallel Reference Processing closures
4870 4891
4871 4892 // Implementation of AbstractRefProcTaskExecutor for parallel reference
4872 4893 // processing during G1 evacuation pauses.
4873 4894
4874 4895 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
4875 4896 private:
4876 4897 G1CollectedHeap* _g1h;
4877 4898 RefToScanQueueSet* _queues;
4878 4899 WorkGang* _workers;
4879 4900 int _active_workers;
4880 4901
4881 4902 public:
4882 4903 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
4883 4904 WorkGang* workers,
4884 4905 RefToScanQueueSet *task_queues,
4885 4906 int n_workers) :
4886 4907 _g1h(g1h),
4887 4908 _queues(task_queues),
4888 4909 _workers(workers),
4889 4910 _active_workers(n_workers)
4890 4911 {
4891 4912 assert(n_workers > 0, "shouldn't call this otherwise");
4892 4913 }
4893 4914
4894 4915 // Executes the given task using concurrent marking worker threads.
4895 4916 virtual void execute(ProcessTask& task);
4896 4917 virtual void execute(EnqueueTask& task);
4897 4918 };
4898 4919
4899 4920 // Gang task for possibly parallel reference processing
4900 4921
4901 4922 class G1STWRefProcTaskProxy: public AbstractGangTask {
4902 4923 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
4903 4924 ProcessTask& _proc_task;
4904 4925 G1CollectedHeap* _g1h;
4905 4926 RefToScanQueueSet *_task_queues;
4906 4927 ParallelTaskTerminator* _terminator;
4907 4928
4908 4929 public:
4909 4930 G1STWRefProcTaskProxy(ProcessTask& proc_task,
4910 4931 G1CollectedHeap* g1h,
4911 4932 RefToScanQueueSet *task_queues,
4912 4933 ParallelTaskTerminator* terminator) :
4913 4934 AbstractGangTask("Process reference objects in parallel"),
4914 4935 _proc_task(proc_task),
4915 4936 _g1h(g1h),
4916 4937 _task_queues(task_queues),
4917 4938 _terminator(terminator)
4918 4939 {}
4919 4940
4920 4941 virtual void work(int i) {
4921 4942 // The reference processing task executed by a single worker.
4922 4943 ResourceMark rm;
4923 4944 HandleMark hm;
4924 4945
4925 4946 G1STWIsAliveClosure is_alive(_g1h);
4926 4947
4927 4948 G1ParScanThreadState pss(_g1h, i);
4928 4949
4929 4950 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
4930 4951 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
4931 4952 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
4932 4953
4933 4954 pss.set_evac_closure(&scan_evac_cl);
4934 4955 pss.set_evac_failure_closure(&evac_failure_cl);
4935 4956 pss.set_partial_scan_closure(&partial_scan_cl);
4936 4957
4937 4958 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
4938 4959 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
4939 4960
4940 4961 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
4941 4962 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
4942 4963
4943 4964 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
4944 4965 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
4945 4966
4946 4967 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4947 4968 // We also need to mark copied objects.
4948 4969 copy_non_heap_cl = ©_mark_non_heap_cl;
4949 4970 copy_perm_cl = ©_mark_perm_cl;
4950 4971 }
4951 4972
4952 4973 // Keep alive closure.
4953 4974 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
4954 4975
4955 4976 // Complete GC closure
4956 4977 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
4957 4978
4958 4979 // Call the reference processing task's work routine.
4959 4980 _proc_task.work(i, is_alive, keep_alive, drain_queue);
4960 4981
4961 4982 // Note we cannot assert that the refs array is empty here as not all
4962 4983 // of the processing tasks (specifically phase2 - pp2_work) execute
4963 4984 // the complete_gc closure (which ordinarily would drain the queue) so
4964 4985 // the queue may not be empty.
4965 4986 }
4966 4987 };
4967 4988
4968 4989 // Driver routine for parallel reference processing.
4969 4990 // Creates an instance of the ref processing gang
4970 4991 // task and has the worker threads execute it.
4971 4992 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
4972 4993 assert(_workers != NULL, "Need parallel worker threads.");
4973 4994
4974 4995 ParallelTaskTerminator terminator(_active_workers, _queues);
4975 4996 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
4976 4997
4977 4998 _g1h->set_par_threads(_active_workers);
4978 4999 _workers->run_task(&proc_task_proxy);
4979 5000 _g1h->set_par_threads(0);
4980 5001 }
4981 5002
4982 5003 // Gang task for parallel reference enqueueing.
4983 5004
4984 5005 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
4985 5006 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
4986 5007 EnqueueTask& _enq_task;
4987 5008
4988 5009 public:
4989 5010 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
4990 5011 AbstractGangTask("Enqueue reference objects in parallel"),
4991 5012 _enq_task(enq_task)
4992 5013 { }
4993 5014
4994 5015 virtual void work(int i) {
4995 5016 _enq_task.work(i);
4996 5017 }
4997 5018 };
4998 5019
4999 5020 // Driver routine for parallel reference enqueing.
5000 5021 // Creates an instance of the ref enqueueing gang
5001 5022 // task and has the worker threads execute it.
5002 5023
5003 5024 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5004 5025 assert(_workers != NULL, "Need parallel worker threads.");
5005 5026
5006 5027 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5007 5028
5008 5029 _g1h->set_par_threads(_active_workers);
5009 5030 _workers->run_task(&enq_task_proxy);
5010 5031 _g1h->set_par_threads(0);
5011 5032 }
5012 5033
5013 5034 // End of weak reference support closures
5014 5035
5015 5036 // Abstract task used to preserve (i.e. copy) any referent objects
5016 5037 // that are in the collection set and are pointed to by reference
5017 5038 // objects discovered by the CM ref processor.
5018 5039
5019 5040 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5020 5041 protected:
5021 5042 G1CollectedHeap* _g1h;
5022 5043 RefToScanQueueSet *_queues;
5023 5044 ParallelTaskTerminator _terminator;
5024 5045 int _n_workers;
5025 5046
5026 5047 public:
5027 5048 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5028 5049 AbstractGangTask("ParPreserveCMReferents"),
5029 5050 _g1h(g1h),
5030 5051 _queues(task_queues),
5031 5052 _terminator(workers, _queues),
5032 5053 _n_workers(workers)
5033 5054 { }
5034 5055
5035 5056 void work(int i) {
5036 5057 ResourceMark rm;
5037 5058 HandleMark hm;
5038 5059
5039 5060 G1ParScanThreadState pss(_g1h, i);
5040 5061 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5041 5062 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5042 5063 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5043 5064
5044 5065 pss.set_evac_closure(&scan_evac_cl);
5045 5066 pss.set_evac_failure_closure(&evac_failure_cl);
5046 5067 pss.set_partial_scan_closure(&partial_scan_cl);
5047 5068
5048 5069 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5049 5070
5050 5071
5051 5072 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5052 5073 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5053 5074
5054 5075 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5055 5076 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5056 5077
5057 5078 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5058 5079 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5059 5080
5060 5081 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5061 5082 // We also need to mark copied objects.
5062 5083 copy_non_heap_cl = ©_mark_non_heap_cl;
5063 5084 copy_perm_cl = ©_mark_perm_cl;
5064 5085 }
5065 5086
5066 5087 // Is alive closure
5067 5088 G1AlwaysAliveClosure always_alive(_g1h);
5068 5089
5069 5090 // Copying keep alive closure. Applied to referent objects that need
5070 5091 // to be copied.
5071 5092 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5072 5093
5073 5094 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5074 5095
5075 5096 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5076 5097 int stride = MIN2(MAX2(_n_workers, 1), limit);
5077 5098
5078 5099 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5079 5100 // So this must be true - but assert just in case someone decides to
5080 5101 // change the worker ids.
5081 5102 assert(0 <= i && i < limit, "sanity");
5082 5103 assert(!rp->discovery_is_atomic(), "check this code");
5083 5104
5084 5105 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5085 5106 for (int idx = i; idx < limit; idx += stride) {
5086 5107 DiscoveredList& ref_list = rp->discovered_soft_refs()[idx];
5087 5108
5088 5109 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5089 5110 while (iter.has_next()) {
5090 5111 // Since discovery is not atomic for the CM ref processor, we
5091 5112 // can see some null referent objects.
5092 5113 iter.load_ptrs(DEBUG_ONLY(true));
5093 5114 oop ref = iter.obj();
5094 5115
5095 5116 // This will filter nulls.
5096 5117 if (iter.is_referent_alive()) {
5097 5118 iter.make_referent_alive();
5098 5119 }
5099 5120 iter.move_to_next();
5100 5121 }
5101 5122 }
5102 5123
5103 5124 // Drain the queue - which may cause stealing
5104 5125 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5105 5126 drain_queue.do_void();
5106 5127 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5107 5128 assert(pss.refs()->is_empty(), "should be");
5108 5129 }
5109 5130 };
5110 5131
5111 5132 // Weak Reference processing during an evacuation pause (part 1).
5112 5133 void G1CollectedHeap::process_discovered_references() {
5113 5134 double ref_proc_start = os::elapsedTime();
5114 5135
5115 5136 ReferenceProcessor* rp = _ref_processor_stw;
5116 5137 assert(rp->discovery_enabled(), "should have been enabled");
5117 5138
5118 5139 // Any reference objects, in the collection set, that were 'discovered'
5119 5140 // by the CM ref processor should have already been copied (either by
5120 5141 // applying the external root copy closure to the discovered lists, or
5121 5142 // by following an RSet entry).
5122 5143 //
5123 5144 // But some of the referents, that are in the collection set, that these
5124 5145 // reference objects point to may not have been copied: the STW ref
5125 5146 // processor would have seen that the reference object had already
5126 5147 // been 'discovered' and would have skipped discovering the reference,
5127 5148 // but would not have treated the reference object as a regular oop.
5128 5149 // As a reult the copy closure would not have been applied to the
5129 5150 // referent object.
5130 5151 //
5131 5152 // We need to explicitly copy these referent objects - the references
5132 5153 // will be processed at the end of remarking.
5133 5154 //
5134 5155 // We also need to do this copying before we process the reference
5135 5156 // objects discovered by the STW ref processor in case one of these
5136 5157 // referents points to another object which is also referenced by an
5137 5158 // object discovered by the STW ref processor.
5138 5159
5139 5160 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5140 5161 workers()->total_workers() : 1);
5141 5162
5142 5163 set_par_threads(n_workers);
5143 5164 G1ParPreserveCMReferentsTask keep_cm_referents(this, n_workers, _task_queues);
5144 5165
5145 5166 if (G1CollectedHeap::use_parallel_gc_threads()) {
5146 5167 workers()->run_task(&keep_cm_referents);
5147 5168 } else {
5148 5169 keep_cm_referents.work(0);
5149 5170 }
5150 5171
5151 5172 set_par_threads(0);
5152 5173
5153 5174 // Closure to test whether a referent is alive.
5154 5175 G1STWIsAliveClosure is_alive(this);
5155 5176
5156 5177 // Even when parallel reference processing is enabled, the processing
5157 5178 // of JNI refs is serial and performed serially by the current thread
5158 5179 // rather than by a worker. The following PSS will be used for processing
5159 5180 // JNI refs.
5160 5181
5161 5182 // Use only a single queue for this PSS.
5162 5183 G1ParScanThreadState pss(this, 0);
5163 5184
5164 5185 // We do not embed a reference processor in the copying/scanning
5165 5186 // closures while we're actually processing the discovered
5166 5187 // reference objects.
5167 5188 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5168 5189 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5169 5190 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5170 5191
5171 5192 pss.set_evac_closure(&scan_evac_cl);
5172 5193 pss.set_evac_failure_closure(&evac_failure_cl);
5173 5194 pss.set_partial_scan_closure(&partial_scan_cl);
5174 5195
5175 5196 assert(pss.refs()->is_empty(), "pre-condition");
5176 5197
5177 5198 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5178 5199 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL);
5179 5200
5180 5201 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5181 5202 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL);
5182 5203
5183 5204 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5184 5205 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5185 5206
5186 5207 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5187 5208 // We also need to mark copied objects.
5188 5209 copy_non_heap_cl = ©_mark_non_heap_cl;
5189 5210 copy_perm_cl = ©_mark_perm_cl;
5190 5211 }
5191 5212
5192 5213 // Keep alive closure.
5193 5214 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss);
5194 5215
5195 5216 // Serial Complete GC closure
5196 5217 G1STWDrainQueueClosure drain_queue(this, &pss);
5197 5218
5198 5219 // Setup the soft refs policy...
5199 5220 rp->setup_policy(false);
5200 5221
5201 5222 if (!rp->processing_is_mt()) {
5202 5223 // Serial reference processing...
5203 5224 rp->process_discovered_references(&is_alive,
5204 5225 &keep_alive,
5205 5226 &drain_queue,
5206 5227 NULL);
5207 5228 } else {
5208 5229 // Parallel reference processing
5209 5230 int active_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5210 5231 assert(rp->num_q() == active_workers, "sanity");
5211 5232 assert(active_workers <= rp->max_num_q(), "sanity");
5212 5233
5213 5234 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5214 5235 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5215 5236 }
5216 5237
5217 5238 // We have completed copying any necessary live referent objects
5218 5239 // (that were not copied during the actual pause) so we can
5219 5240 // retire any active alloc buffers
5220 5241 pss.retire_alloc_buffers();
5221 5242 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5222 5243
5223 5244 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5224 5245 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0);
5225 5246 }
5226 5247
5227 5248 // Weak Reference processing during an evacuation pause (part 2).
5228 5249 void G1CollectedHeap::enqueue_discovered_references() {
5229 5250 double ref_enq_start = os::elapsedTime();
5230 5251
5231 5252 ReferenceProcessor* rp = _ref_processor_stw;
5232 5253 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5233 5254
5234 5255 // Now enqueue any remaining on the discovered lists on to
5235 5256 // the pending list.
5236 5257 if (!rp->processing_is_mt()) {
5237 5258 // Serial reference processing...
5238 5259 rp->enqueue_discovered_references();
5239 5260 } else {
5240 5261 // Parallel reference enqueuing
5241 5262
5242 5263 int active_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5243 5264 assert(rp->num_q() == active_workers, "sanity");
5244 5265 assert(active_workers <= rp->max_num_q(), "sanity");
5245 5266
5246 5267 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5247 5268 rp->enqueue_discovered_references(&par_task_executor);
5248 5269 }
5249 5270
5250 5271 rp->verify_no_references_recorded();
5251 5272 assert(!rp->discovery_enabled(), "should have been disabled");
5252 5273
5253 5274 // FIXME
5254 5275 // CM's reference processing also cleans up the string and symbol tables.
5255 5276 // Should we do that here also? We could, but it is a serial operation
5256 5277 // and could signicantly increase the pause time.
5257 5278
5258 5279 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5259 5280 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0);
5260 5281 }
5261 5282
5262 5283 void G1CollectedHeap::evacuate_collection_set() {
5263 5284 set_evacuation_failed(false);
5264 5285
5265 5286 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5266 5287 concurrent_g1_refine()->set_use_cache(false);
5267 5288 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5268 5289
5269 5290 int n_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
5270 5291 set_par_threads(n_workers);
5271 5292 G1ParTask g1_par_task(this, n_workers, _task_queues);
5272 5293
5273 5294 init_for_evac_failure(NULL);
5274 5295
5275 5296 rem_set()->prepare_for_younger_refs_iterate(true);
5276 5297
5277 5298 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5278 5299 double start_par = os::elapsedTime();
5279 5300
5280 5301 if (G1CollectedHeap::use_parallel_gc_threads()) {
5281 5302 // The individual threads will set their evac-failure closures.
5282 5303 StrongRootsScope srs(this);
5283 5304 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5284 5305 workers()->run_task(&g1_par_task);
5285 5306 } else {
5286 5307 StrongRootsScope srs(this);
5287 5308 g1_par_task.work(0);
5288 5309 }
5289 5310
5290 5311 double par_time = (os::elapsedTime() - start_par) * 1000.0;
5291 5312 g1_policy()->record_par_time(par_time);
5292 5313 set_par_threads(0);
5293 5314
5294 5315 // Process any discovered reference objects - we have
5295 5316 // to do this _before_ we retire the GC alloc regions
5296 5317 // as we may have to copy some 'reachable' referent
5297 5318 // objects (and their reachable sub-graphs) that were
5298 5319 // not copied during the pause.
5299 5320 process_discovered_references();
5300 5321
5301 5322 // Weak root processing.
5302 5323 // Note: when JSR 292 is enabled and code blobs can contain
5303 5324 // non-perm oops then we will need to process the code blobs
5304 5325 // here too.
5305 5326 {
5306 5327 G1STWIsAliveClosure is_alive(this);
5307 5328 G1KeepAliveClosure keep_alive(this);
5308 5329 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5309 5330 }
5310 5331
5311 5332 release_gc_alloc_regions();
5312 5333 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5313 5334
5314 5335 concurrent_g1_refine()->clear_hot_cache();
5315 5336 concurrent_g1_refine()->set_use_cache(true);
5316 5337
5317 5338 finalize_for_evac_failure();
5318 5339
5319 5340 // Must do this before removing self-forwarding pointers, which clears
5320 5341 // the per-region evac-failure flags.
5321 5342 concurrent_mark()->complete_marking_in_collection_set();
5322 5343
5323 5344 if (evacuation_failed()) {
5324 5345 remove_self_forwarding_pointers();
5325 5346 if (PrintGCDetails) {
5326 5347 gclog_or_tty->print(" (to-space overflow)");
5327 5348 } else if (PrintGC) {
5328 5349 gclog_or_tty->print("--");
5329 5350 }
5330 5351 }
5331 5352
5332 5353 // Enqueue any remaining references remaining on the STW
5333 5354 // reference processor's discovered lists. We need to do
5334 5355 // this after the card table is cleaned (and verified) as
5335 5356 // the act of enqueuing entries on to the pending list
5336 5357 // will log these updates (and dirty their associated
5337 5358 // cards). We need these updates logged to update any
5338 5359 // RSets.
5339 5360 enqueue_discovered_references();
5340 5361
5341 5362 if (G1DeferredRSUpdate) {
5342 5363 RedirtyLoggedCardTableEntryFastClosure redirty;
5343 5364 dirty_card_queue_set().set_closure(&redirty);
5344 5365 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5345 5366
5346 5367 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5347 5368 dcq.merge_bufferlists(&dirty_card_queue_set());
5348 5369 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5349 5370 }
5350 5371 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5351 5372 }
5352 5373
5353 5374 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5354 5375 size_t* pre_used,
5355 5376 FreeRegionList* free_list,
5356 5377 HumongousRegionSet* humongous_proxy_set,
5357 5378 HRRSCleanupTask* hrrs_cleanup_task,
5358 5379 bool par) {
5359 5380 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5360 5381 if (hr->isHumongous()) {
5361 5382 assert(hr->startsHumongous(), "we should only see starts humongous");
5362 5383 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5363 5384 } else {
5364 5385 free_region(hr, pre_used, free_list, par);
5365 5386 }
5366 5387 } else {
5367 5388 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5368 5389 }
5369 5390 }
5370 5391
5371 5392 void G1CollectedHeap::free_region(HeapRegion* hr,
5372 5393 size_t* pre_used,
5373 5394 FreeRegionList* free_list,
5374 5395 bool par) {
5375 5396 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5376 5397 assert(!hr->is_empty(), "the region should not be empty");
5377 5398 assert(free_list != NULL, "pre-condition");
5378 5399
5379 5400 *pre_used += hr->used();
5380 5401 hr->hr_clear(par, true /* clear_space */);
5381 5402 free_list->add_as_head(hr);
5382 5403 }
5383 5404
5384 5405 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5385 5406 size_t* pre_used,
5386 5407 FreeRegionList* free_list,
5387 5408 HumongousRegionSet* humongous_proxy_set,
5388 5409 bool par) {
5389 5410 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5390 5411 assert(free_list != NULL, "pre-condition");
5391 5412 assert(humongous_proxy_set != NULL, "pre-condition");
5392 5413
5393 5414 size_t hr_used = hr->used();
5394 5415 size_t hr_capacity = hr->capacity();
5395 5416 size_t hr_pre_used = 0;
5396 5417 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5397 5418 hr->set_notHumongous();
5398 5419 free_region(hr, &hr_pre_used, free_list, par);
5399 5420
5400 5421 size_t i = hr->hrs_index() + 1;
5401 5422 size_t num = 1;
5402 5423 while (i < n_regions()) {
5403 5424 HeapRegion* curr_hr = region_at(i);
5404 5425 if (!curr_hr->continuesHumongous()) {
5405 5426 break;
5406 5427 }
5407 5428 curr_hr->set_notHumongous();
5408 5429 free_region(curr_hr, &hr_pre_used, free_list, par);
5409 5430 num += 1;
5410 5431 i += 1;
5411 5432 }
5412 5433 assert(hr_pre_used == hr_used,
5413 5434 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5414 5435 "should be the same", hr_pre_used, hr_used));
5415 5436 *pre_used += hr_pre_used;
5416 5437 }
5417 5438
5418 5439 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5419 5440 FreeRegionList* free_list,
5420 5441 HumongousRegionSet* humongous_proxy_set,
5421 5442 bool par) {
5422 5443 if (pre_used > 0) {
5423 5444 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5424 5445 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5425 5446 assert(_summary_bytes_used >= pre_used,
5426 5447 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5427 5448 "should be >= pre_used: "SIZE_FORMAT,
5428 5449 _summary_bytes_used, pre_used));
5429 5450 _summary_bytes_used -= pre_used;
5430 5451 }
5431 5452 if (free_list != NULL && !free_list->is_empty()) {
5432 5453 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5433 5454 _free_list.add_as_head(free_list);
5434 5455 }
5435 5456 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5436 5457 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5437 5458 _humongous_set.update_from_proxy(humongous_proxy_set);
5438 5459 }
5439 5460 }
5440 5461
5441 5462 class G1ParCleanupCTTask : public AbstractGangTask {
5442 5463 CardTableModRefBS* _ct_bs;
5443 5464 G1CollectedHeap* _g1h;
5444 5465 HeapRegion* volatile _su_head;
5445 5466 public:
5446 5467 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5447 5468 G1CollectedHeap* g1h) :
5448 5469 AbstractGangTask("G1 Par Cleanup CT Task"),
5449 5470 _ct_bs(ct_bs), _g1h(g1h) { }
5450 5471
5451 5472 void work(int i) {
5452 5473 HeapRegion* r;
5453 5474 while (r = _g1h->pop_dirty_cards_region()) {
5454 5475 clear_cards(r);
5455 5476 }
5456 5477 }
5457 5478
5458 5479 void clear_cards(HeapRegion* r) {
5459 5480 // Cards of the survivors should have already been dirtied.
5460 5481 if (!r->is_survivor()) {
5461 5482 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5462 5483 }
5463 5484 }
5464 5485 };
5465 5486
5466 5487 #ifndef PRODUCT
5467 5488 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5468 5489 G1CollectedHeap* _g1h;
5469 5490 CardTableModRefBS* _ct_bs;
5470 5491 public:
5471 5492 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5472 5493 : _g1h(g1h), _ct_bs(ct_bs) { }
5473 5494 virtual bool doHeapRegion(HeapRegion* r) {
5474 5495 if (r->is_survivor()) {
5475 5496 _g1h->verify_dirty_region(r);
5476 5497 } else {
5477 5498 _g1h->verify_not_dirty_region(r);
5478 5499 }
5479 5500 return false;
5480 5501 }
5481 5502 };
5482 5503
5483 5504 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5484 5505 // All of the region should be clean.
5485 5506 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5486 5507 MemRegion mr(hr->bottom(), hr->end());
5487 5508 ct_bs->verify_not_dirty_region(mr);
5488 5509 }
5489 5510
5490 5511 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5491 5512 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5492 5513 // dirty allocated blocks as they allocate them. The thread that
5493 5514 // retires each region and replaces it with a new one will do a
5494 5515 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5495 5516 // not dirty that area (one less thing to have to do while holding
5496 5517 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5497 5518 // is dirty.
5498 5519 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5499 5520 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5500 5521 ct_bs->verify_dirty_region(mr);
5501 5522 }
5502 5523
5503 5524 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5504 5525 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5505 5526 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5506 5527 verify_dirty_region(hr);
5507 5528 }
5508 5529 }
5509 5530
5510 5531 void G1CollectedHeap::verify_dirty_young_regions() {
5511 5532 verify_dirty_young_list(_young_list->first_region());
5512 5533 verify_dirty_young_list(_young_list->first_survivor_region());
5513 5534 }
5514 5535 #endif
5515 5536
5516 5537 void G1CollectedHeap::cleanUpCardTable() {
5517 5538 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5518 5539 double start = os::elapsedTime();
5519 5540
5520 5541 // Iterate over the dirty cards region list.
5521 5542 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5522 5543
5523 5544 if (ParallelGCThreads > 0) {
5524 5545 set_par_threads(workers()->total_workers());
5525 5546 workers()->run_task(&cleanup_task);
5526 5547 set_par_threads(0);
5527 5548 } else {
5528 5549 while (_dirty_cards_region_list) {
5529 5550 HeapRegion* r = _dirty_cards_region_list;
5530 5551 cleanup_task.clear_cards(r);
5531 5552 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5532 5553 if (_dirty_cards_region_list == r) {
5533 5554 // The last region.
5534 5555 _dirty_cards_region_list = NULL;
5535 5556 }
5536 5557 r->set_next_dirty_cards_region(NULL);
5537 5558 }
5538 5559 }
5539 5560
5540 5561 double elapsed = os::elapsedTime() - start;
5541 5562 g1_policy()->record_clear_ct_time(elapsed * 1000.0);
5542 5563 #ifndef PRODUCT
5543 5564 if (G1VerifyCTCleanup || VerifyAfterGC) {
5544 5565 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5545 5566 heap_region_iterate(&cleanup_verifier);
5546 5567 }
5547 5568 #endif
5548 5569 }
5549 5570
5550 5571 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5551 5572 size_t pre_used = 0;
5552 5573 FreeRegionList local_free_list("Local List for CSet Freeing");
5553 5574
5554 5575 double young_time_ms = 0.0;
5555 5576 double non_young_time_ms = 0.0;
5556 5577
5557 5578 // Since the collection set is a superset of the the young list,
5558 5579 // all we need to do to clear the young list is clear its
5559 5580 // head and length, and unlink any young regions in the code below
5560 5581 _young_list->clear();
5561 5582
5562 5583 G1CollectorPolicy* policy = g1_policy();
5563 5584
5564 5585 double start_sec = os::elapsedTime();
5565 5586 bool non_young = true;
5566 5587
5567 5588 HeapRegion* cur = cs_head;
5568 5589 int age_bound = -1;
5569 5590 size_t rs_lengths = 0;
5570 5591
5571 5592 while (cur != NULL) {
5572 5593 assert(!is_on_master_free_list(cur), "sanity");
5573 5594
5574 5595 if (non_young) {
5575 5596 if (cur->is_young()) {
5576 5597 double end_sec = os::elapsedTime();
5577 5598 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5578 5599 non_young_time_ms += elapsed_ms;
5579 5600
5580 5601 start_sec = os::elapsedTime();
5581 5602 non_young = false;
5582 5603 }
5583 5604 } else {
5584 5605 double end_sec = os::elapsedTime();
5585 5606 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5586 5607 young_time_ms += elapsed_ms;
5587 5608
5588 5609 start_sec = os::elapsedTime();
5589 5610 non_young = true;
5590 5611 }
5591 5612
5592 5613 rs_lengths += cur->rem_set()->occupied();
5593 5614
5594 5615 HeapRegion* next = cur->next_in_collection_set();
5595 5616 assert(cur->in_collection_set(), "bad CS");
5596 5617 cur->set_next_in_collection_set(NULL);
5597 5618 cur->set_in_collection_set(false);
5598 5619
5599 5620 if (cur->is_young()) {
5600 5621 int index = cur->young_index_in_cset();
5601 5622 guarantee( index != -1, "invariant" );
5602 5623 guarantee( (size_t)index < policy->young_cset_length(), "invariant" );
5603 5624 size_t words_survived = _surviving_young_words[index];
5604 5625 cur->record_surv_words_in_group(words_survived);
5605 5626
5606 5627 // At this point the we have 'popped' cur from the collection set
5607 5628 // (linked via next_in_collection_set()) but it is still in the
5608 5629 // young list (linked via next_young_region()). Clear the
5609 5630 // _next_young_region field.
5610 5631 cur->set_next_young_region(NULL);
5611 5632 } else {
5612 5633 int index = cur->young_index_in_cset();
5613 5634 guarantee( index == -1, "invariant" );
5614 5635 }
5615 5636
5616 5637 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5617 5638 (!cur->is_young() && cur->young_index_in_cset() == -1),
5618 5639 "invariant" );
5619 5640
5620 5641 if (!cur->evacuation_failed()) {
5621 5642 // And the region is empty.
5622 5643 assert(!cur->is_empty(), "Should not have empty regions in a CS.");
5623 5644 free_region(cur, &pre_used, &local_free_list, false /* par */);
5624 5645 } else {
5625 5646 cur->uninstall_surv_rate_group();
5626 5647 if (cur->is_young())
5627 5648 cur->set_young_index_in_cset(-1);
5628 5649 cur->set_not_young();
5629 5650 cur->set_evacuation_failed(false);
5630 5651 }
5631 5652 cur = next;
5632 5653 }
5633 5654
5634 5655 policy->record_max_rs_lengths(rs_lengths);
5635 5656 policy->cset_regions_freed();
5636 5657
5637 5658 double end_sec = os::elapsedTime();
5638 5659 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5639 5660 if (non_young)
5640 5661 non_young_time_ms += elapsed_ms;
5641 5662 else
5642 5663 young_time_ms += elapsed_ms;
5643 5664
5644 5665 update_sets_after_freeing_regions(pre_used, &local_free_list,
5645 5666 NULL /* humongous_proxy_set */,
5646 5667 false /* par */);
5647 5668 policy->record_young_free_cset_time_ms(young_time_ms);
5648 5669 policy->record_non_young_free_cset_time_ms(non_young_time_ms);
5649 5670 }
5650 5671
5651 5672 // This routine is similar to the above but does not record
5652 5673 // any policy statistics or update free lists; we are abandoning
5653 5674 // the current incremental collection set in preparation of a
5654 5675 // full collection. After the full GC we will start to build up
5655 5676 // the incremental collection set again.
5656 5677 // This is only called when we're doing a full collection
5657 5678 // and is immediately followed by the tearing down of the young list.
5658 5679
5659 5680 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
5660 5681 HeapRegion* cur = cs_head;
5661 5682
5662 5683 while (cur != NULL) {
5663 5684 HeapRegion* next = cur->next_in_collection_set();
5664 5685 assert(cur->in_collection_set(), "bad CS");
5665 5686 cur->set_next_in_collection_set(NULL);
5666 5687 cur->set_in_collection_set(false);
5667 5688 cur->set_young_index_in_cset(-1);
5668 5689 cur = next;
5669 5690 }
5670 5691 }
5671 5692
5672 5693 void G1CollectedHeap::set_free_regions_coming() {
5673 5694 if (G1ConcRegionFreeingVerbose) {
5674 5695 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5675 5696 "setting free regions coming");
5676 5697 }
5677 5698
5678 5699 assert(!free_regions_coming(), "pre-condition");
5679 5700 _free_regions_coming = true;
5680 5701 }
5681 5702
5682 5703 void G1CollectedHeap::reset_free_regions_coming() {
5683 5704 {
5684 5705 assert(free_regions_coming(), "pre-condition");
5685 5706 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5686 5707 _free_regions_coming = false;
5687 5708 SecondaryFreeList_lock->notify_all();
5688 5709 }
5689 5710
5690 5711 if (G1ConcRegionFreeingVerbose) {
5691 5712 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5692 5713 "reset free regions coming");
5693 5714 }
5694 5715 }
5695 5716
5696 5717 void G1CollectedHeap::wait_while_free_regions_coming() {
5697 5718 // Most of the time we won't have to wait, so let's do a quick test
5698 5719 // first before we take the lock.
5699 5720 if (!free_regions_coming()) {
5700 5721 return;
5701 5722 }
5702 5723
5703 5724 if (G1ConcRegionFreeingVerbose) {
5704 5725 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5705 5726 "waiting for free regions");
5706 5727 }
5707 5728
5708 5729 {
5709 5730 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5710 5731 while (free_regions_coming()) {
5711 5732 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
5712 5733 }
5713 5734 }
5714 5735
5715 5736 if (G1ConcRegionFreeingVerbose) {
5716 5737 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5717 5738 "done waiting for free regions");
5718 5739 }
5719 5740 }
5720 5741
5721 5742 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
5722 5743 assert(heap_lock_held_for_gc(),
5723 5744 "the heap lock should already be held by or for this thread");
5724 5745 _young_list->push_region(hr);
5725 5746 g1_policy()->set_region_short_lived(hr);
5726 5747 }
5727 5748
5728 5749 class NoYoungRegionsClosure: public HeapRegionClosure {
5729 5750 private:
5730 5751 bool _success;
5731 5752 public:
5732 5753 NoYoungRegionsClosure() : _success(true) { }
5733 5754 bool doHeapRegion(HeapRegion* r) {
5734 5755 if (r->is_young()) {
5735 5756 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
5736 5757 r->bottom(), r->end());
5737 5758 _success = false;
5738 5759 }
5739 5760 return false;
5740 5761 }
5741 5762 bool success() { return _success; }
5742 5763 };
5743 5764
5744 5765 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
5745 5766 bool ret = _young_list->check_list_empty(check_sample);
5746 5767
5747 5768 if (check_heap) {
5748 5769 NoYoungRegionsClosure closure;
5749 5770 heap_region_iterate(&closure);
5750 5771 ret = ret && closure.success();
5751 5772 }
5752 5773
5753 5774 return ret;
5754 5775 }
5755 5776
5756 5777 void G1CollectedHeap::empty_young_list() {
5757 5778 assert(heap_lock_held_for_gc(),
5758 5779 "the heap lock should already be held by or for this thread");
5759 5780
5760 5781 _young_list->empty_list();
5761 5782 }
5762 5783
5763 5784 // Done at the start of full GC.
5764 5785 void G1CollectedHeap::tear_down_region_lists() {
5765 5786 _free_list.remove_all();
5766 5787 }
5767 5788
5768 5789 class RegionResetter: public HeapRegionClosure {
5769 5790 G1CollectedHeap* _g1h;
5770 5791 FreeRegionList _local_free_list;
5771 5792
5772 5793 public:
5773 5794 RegionResetter() : _g1h(G1CollectedHeap::heap()),
5774 5795 _local_free_list("Local Free List for RegionResetter") { }
5775 5796
5776 5797 bool doHeapRegion(HeapRegion* r) {
5777 5798 if (r->continuesHumongous()) return false;
5778 5799 if (r->top() > r->bottom()) {
5779 5800 if (r->top() < r->end()) {
5780 5801 Copy::fill_to_words(r->top(),
5781 5802 pointer_delta(r->end(), r->top()));
5782 5803 }
5783 5804 } else {
5784 5805 assert(r->is_empty(), "tautology");
5785 5806 _local_free_list.add_as_tail(r);
5786 5807 }
5787 5808 return false;
5788 5809 }
5789 5810
5790 5811 void update_free_lists() {
5791 5812 _g1h->update_sets_after_freeing_regions(0, &_local_free_list, NULL,
5792 5813 false /* par */);
5793 5814 }
5794 5815 };
5795 5816
5796 5817 // Done at the end of full GC.
5797 5818 void G1CollectedHeap::rebuild_region_lists() {
5798 5819 // This needs to go at the end of the full GC.
5799 5820 RegionResetter rs;
5800 5821 heap_region_iterate(&rs);
5801 5822 rs.update_free_lists();
5802 5823 }
5803 5824
5804 5825 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
5805 5826 _refine_cte_cl->set_concurrent(concurrent);
5806 5827 }
5807 5828
5808 5829 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
5809 5830 HeapRegion* hr = heap_region_containing(p);
5810 5831 if (hr == NULL) {
5811 5832 return is_in_permanent(p);
5812 5833 } else {
5813 5834 return hr->is_in(p);
5814 5835 }
5815 5836 }
5816 5837
5817 5838 // Methods for the mutator alloc region
5818 5839
5819 5840 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
5820 5841 bool force) {
5821 5842 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5822 5843 assert(!force || g1_policy()->can_expand_young_list(),
5823 5844 "if force is true we should be able to expand the young list");
5824 5845 bool young_list_full = g1_policy()->is_young_list_full();
5825 5846 if (force || !young_list_full) {
5826 5847 HeapRegion* new_alloc_region = new_region(word_size,
5827 5848 false /* do_expand */);
5828 5849 if (new_alloc_region != NULL) {
5829 5850 g1_policy()->update_region_num(true /* next_is_young */);
5830 5851 set_region_short_lived_locked(new_alloc_region);
5831 5852 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
5832 5853 return new_alloc_region;
5833 5854 }
5834 5855 }
5835 5856 return NULL;
5836 5857 }
5837 5858
5838 5859 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
5839 5860 size_t allocated_bytes) {
5840 5861 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5841 5862 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
5842 5863
5843 5864 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
5844 5865 _summary_bytes_used += allocated_bytes;
5845 5866 _hr_printer.retire(alloc_region);
5846 5867 // We update the eden sizes here, when the region is retired,
5847 5868 // instead of when it's allocated, since this is the point that its
5848 5869 // used space has been recored in _summary_bytes_used.
5849 5870 g1mm()->update_eden_size();
5850 5871 }
5851 5872
5852 5873 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
5853 5874 bool force) {
5854 5875 return _g1h->new_mutator_alloc_region(word_size, force);
5855 5876 }
5856 5877
5857 5878 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
5858 5879 size_t allocated_bytes) {
5859 5880 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
5860 5881 }
5861 5882
5862 5883 // Methods for the GC alloc regions
5863 5884
5864 5885 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
5865 5886 size_t count,
5866 5887 GCAllocPurpose ap) {
5867 5888 assert(FreeList_lock->owned_by_self(), "pre-condition");
5868 5889
5869 5890 if (count < g1_policy()->max_regions(ap)) {
5870 5891 HeapRegion* new_alloc_region = new_region(word_size,
5871 5892 true /* do_expand */);
5872 5893 if (new_alloc_region != NULL) {
5873 5894 // We really only need to do this for old regions given that we
5874 5895 // should never scan survivors. But it doesn't hurt to do it
5875 5896 // for survivors too.
5876 5897 new_alloc_region->set_saved_mark();
5877 5898 if (ap == GCAllocForSurvived) {
5878 5899 new_alloc_region->set_survivor();
5879 5900 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
5880 5901 } else {
5881 5902 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
5882 5903 }
5883 5904 return new_alloc_region;
5884 5905 } else {
5885 5906 g1_policy()->note_alloc_region_limit_reached(ap);
5886 5907 }
5887 5908 }
5888 5909 return NULL;
5889 5910 }
5890 5911
5891 5912 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
5892 5913 size_t allocated_bytes,
5893 5914 GCAllocPurpose ap) {
5894 5915 alloc_region->note_end_of_copying();
5895 5916 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
5896 5917 if (ap == GCAllocForSurvived) {
5897 5918 young_list()->add_survivor_region(alloc_region);
5898 5919 }
5899 5920 _hr_printer.retire(alloc_region);
5900 5921 }
5901 5922
5902 5923 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
5903 5924 bool force) {
5904 5925 assert(!force, "not supported for GC alloc regions");
5905 5926 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
5906 5927 }
5907 5928
5908 5929 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
5909 5930 size_t allocated_bytes) {
5910 5931 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
5911 5932 GCAllocForSurvived);
5912 5933 }
5913 5934
5914 5935 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
5915 5936 bool force) {
5916 5937 assert(!force, "not supported for GC alloc regions");
5917 5938 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
5918 5939 }
5919 5940
5920 5941 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
5921 5942 size_t allocated_bytes) {
5922 5943 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
5923 5944 GCAllocForTenured);
5924 5945 }
5925 5946 // Heap region set verification
5926 5947
5927 5948 class VerifyRegionListsClosure : public HeapRegionClosure {
5928 5949 private:
5929 5950 HumongousRegionSet* _humongous_set;
5930 5951 FreeRegionList* _free_list;
5931 5952 size_t _region_count;
5932 5953
5933 5954 public:
5934 5955 VerifyRegionListsClosure(HumongousRegionSet* humongous_set,
5935 5956 FreeRegionList* free_list) :
5936 5957 _humongous_set(humongous_set), _free_list(free_list),
5937 5958 _region_count(0) { }
5938 5959
5939 5960 size_t region_count() { return _region_count; }
5940 5961
5941 5962 bool doHeapRegion(HeapRegion* hr) {
5942 5963 _region_count += 1;
5943 5964
5944 5965 if (hr->continuesHumongous()) {
5945 5966 return false;
5946 5967 }
5947 5968
5948 5969 if (hr->is_young()) {
5949 5970 // TODO
5950 5971 } else if (hr->startsHumongous()) {
5951 5972 _humongous_set->verify_next_region(hr);
5952 5973 } else if (hr->is_empty()) {
5953 5974 _free_list->verify_next_region(hr);
5954 5975 }
5955 5976 return false;
5956 5977 }
5957 5978 };
5958 5979
5959 5980 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index,
5960 5981 HeapWord* bottom) {
5961 5982 HeapWord* end = bottom + HeapRegion::GrainWords;
5962 5983 MemRegion mr(bottom, end);
5963 5984 assert(_g1_reserved.contains(mr), "invariant");
5964 5985 // This might return NULL if the allocation fails
5965 5986 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
5966 5987 }
5967 5988
5968 5989 void G1CollectedHeap::verify_region_sets() {
5969 5990 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
5970 5991
5971 5992 // First, check the explicit lists.
5972 5993 _free_list.verify();
5973 5994 {
5974 5995 // Given that a concurrent operation might be adding regions to
5975 5996 // the secondary free list we have to take the lock before
5976 5997 // verifying it.
5977 5998 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5978 5999 _secondary_free_list.verify();
5979 6000 }
5980 6001 _humongous_set.verify();
5981 6002
5982 6003 // If a concurrent region freeing operation is in progress it will
5983 6004 // be difficult to correctly attributed any free regions we come
5984 6005 // across to the correct free list given that they might belong to
5985 6006 // one of several (free_list, secondary_free_list, any local lists,
5986 6007 // etc.). So, if that's the case we will skip the rest of the
5987 6008 // verification operation. Alternatively, waiting for the concurrent
5988 6009 // operation to complete will have a non-trivial effect on the GC's
5989 6010 // operation (no concurrent operation will last longer than the
5990 6011 // interval between two calls to verification) and it might hide
5991 6012 // any issues that we would like to catch during testing.
5992 6013 if (free_regions_coming()) {
5993 6014 return;
5994 6015 }
5995 6016
5996 6017 // Make sure we append the secondary_free_list on the free_list so
5997 6018 // that all free regions we will come across can be safely
5998 6019 // attributed to the free_list.
5999 6020 append_secondary_free_list_if_not_empty_with_lock();
6000 6021
6001 6022 // Finally, make sure that the region accounting in the lists is
6002 6023 // consistent with what we see in the heap.
6003 6024 _humongous_set.verify_start();
6004 6025 _free_list.verify_start();
6005 6026
6006 6027 VerifyRegionListsClosure cl(&_humongous_set, &_free_list);
6007 6028 heap_region_iterate(&cl);
6008 6029
6009 6030 _humongous_set.verify_end();
6010 6031 _free_list.verify_end();
6011 6032 }
↓ open down ↓ |
1547 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX