Print this page
rev 2869 : 7119908: G1: Cache CSet start region for each worker for subsequent reuse
Summary: Cache workers' calculated starting heap region, used for parallel iteration over the collcection set, for subsequent reuse.
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 // Notes on implementation of parallelism in different tasks.
70 70 //
71 71 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
72 72 // The number of GC workers is passed to heap_region_par_iterate_chunked().
73 73 // It does use run_task() which sets _n_workers in the task.
74 74 // G1ParTask executes g1_process_strong_roots() ->
75 75 // SharedHeap::process_strong_roots() which calls eventuall to
76 76 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
77 77 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
78 78 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
79 79 //
80 80
81 81 // Local to this file.
82 82
83 83 class RefineCardTableEntryClosure: public CardTableEntryClosure {
84 84 SuspendibleThreadSet* _sts;
85 85 G1RemSet* _g1rs;
86 86 ConcurrentG1Refine* _cg1r;
87 87 bool _concurrent;
88 88 public:
89 89 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
90 90 G1RemSet* g1rs,
91 91 ConcurrentG1Refine* cg1r) :
92 92 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
93 93 {}
94 94 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
95 95 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
96 96 // This path is executed by the concurrent refine or mutator threads,
97 97 // concurrently, and so we do not care if card_ptr contains references
98 98 // that point into the collection set.
99 99 assert(!oops_into_cset, "should be");
100 100
101 101 if (_concurrent && _sts->should_yield()) {
102 102 // Caller will actually yield.
103 103 return false;
104 104 }
105 105 // Otherwise, we finished successfully; return true.
106 106 return true;
107 107 }
108 108 void set_concurrent(bool b) { _concurrent = b; }
109 109 };
110 110
111 111
112 112 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
113 113 int _calls;
114 114 G1CollectedHeap* _g1h;
115 115 CardTableModRefBS* _ctbs;
116 116 int _histo[256];
117 117 public:
118 118 ClearLoggedCardTableEntryClosure() :
119 119 _calls(0)
120 120 {
121 121 _g1h = G1CollectedHeap::heap();
122 122 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
123 123 for (int i = 0; i < 256; i++) _histo[i] = 0;
124 124 }
125 125 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
126 126 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
127 127 _calls++;
128 128 unsigned char* ujb = (unsigned char*)card_ptr;
129 129 int ind = (int)(*ujb);
130 130 _histo[ind]++;
131 131 *card_ptr = -1;
132 132 }
133 133 return true;
134 134 }
135 135 int calls() { return _calls; }
136 136 void print_histo() {
137 137 gclog_or_tty->print_cr("Card table value histogram:");
138 138 for (int i = 0; i < 256; i++) {
139 139 if (_histo[i] != 0) {
140 140 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
141 141 }
142 142 }
143 143 }
144 144 };
145 145
146 146 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
147 147 int _calls;
148 148 G1CollectedHeap* _g1h;
149 149 CardTableModRefBS* _ctbs;
150 150 public:
151 151 RedirtyLoggedCardTableEntryClosure() :
152 152 _calls(0)
153 153 {
154 154 _g1h = G1CollectedHeap::heap();
155 155 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
156 156 }
157 157 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
158 158 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
159 159 _calls++;
160 160 *card_ptr = 0;
161 161 }
162 162 return true;
163 163 }
164 164 int calls() { return _calls; }
165 165 };
166 166
167 167 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
168 168 public:
169 169 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
170 170 *card_ptr = CardTableModRefBS::dirty_card_val();
171 171 return true;
172 172 }
173 173 };
174 174
175 175 YoungList::YoungList(G1CollectedHeap* g1h)
176 176 : _g1h(g1h), _head(NULL),
177 177 _length(0),
178 178 _last_sampled_rs_lengths(0),
179 179 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0)
180 180 {
181 181 guarantee( check_list_empty(false), "just making sure..." );
182 182 }
183 183
184 184 void YoungList::push_region(HeapRegion *hr) {
185 185 assert(!hr->is_young(), "should not already be young");
186 186 assert(hr->get_next_young_region() == NULL, "cause it should!");
187 187
188 188 hr->set_next_young_region(_head);
189 189 _head = hr;
190 190
191 191 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
192 192 ++_length;
193 193 }
194 194
195 195 void YoungList::add_survivor_region(HeapRegion* hr) {
196 196 assert(hr->is_survivor(), "should be flagged as survivor region");
197 197 assert(hr->get_next_young_region() == NULL, "cause it should!");
198 198
199 199 hr->set_next_young_region(_survivor_head);
200 200 if (_survivor_head == NULL) {
201 201 _survivor_tail = hr;
202 202 }
203 203 _survivor_head = hr;
204 204 ++_survivor_length;
205 205 }
206 206
207 207 void YoungList::empty_list(HeapRegion* list) {
208 208 while (list != NULL) {
209 209 HeapRegion* next = list->get_next_young_region();
210 210 list->set_next_young_region(NULL);
211 211 list->uninstall_surv_rate_group();
212 212 list->set_not_young();
213 213 list = next;
214 214 }
215 215 }
216 216
217 217 void YoungList::empty_list() {
218 218 assert(check_list_well_formed(), "young list should be well formed");
219 219
220 220 empty_list(_head);
221 221 _head = NULL;
222 222 _length = 0;
223 223
224 224 empty_list(_survivor_head);
225 225 _survivor_head = NULL;
226 226 _survivor_tail = NULL;
227 227 _survivor_length = 0;
228 228
229 229 _last_sampled_rs_lengths = 0;
230 230
231 231 assert(check_list_empty(false), "just making sure...");
232 232 }
233 233
234 234 bool YoungList::check_list_well_formed() {
235 235 bool ret = true;
236 236
237 237 size_t length = 0;
238 238 HeapRegion* curr = _head;
239 239 HeapRegion* last = NULL;
240 240 while (curr != NULL) {
241 241 if (!curr->is_young()) {
242 242 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
243 243 "incorrectly tagged (y: %d, surv: %d)",
244 244 curr->bottom(), curr->end(),
245 245 curr->is_young(), curr->is_survivor());
246 246 ret = false;
247 247 }
248 248 ++length;
249 249 last = curr;
250 250 curr = curr->get_next_young_region();
251 251 }
252 252 ret = ret && (length == _length);
253 253
254 254 if (!ret) {
255 255 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
256 256 gclog_or_tty->print_cr("### list has %d entries, _length is %d",
257 257 length, _length);
258 258 }
259 259
260 260 return ret;
261 261 }
262 262
263 263 bool YoungList::check_list_empty(bool check_sample) {
264 264 bool ret = true;
265 265
266 266 if (_length != 0) {
267 267 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %d",
268 268 _length);
269 269 ret = false;
270 270 }
271 271 if (check_sample && _last_sampled_rs_lengths != 0) {
272 272 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
273 273 ret = false;
274 274 }
275 275 if (_head != NULL) {
276 276 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
277 277 ret = false;
278 278 }
279 279 if (!ret) {
280 280 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
281 281 }
282 282
283 283 return ret;
284 284 }
285 285
286 286 void
287 287 YoungList::rs_length_sampling_init() {
288 288 _sampled_rs_lengths = 0;
289 289 _curr = _head;
290 290 }
291 291
292 292 bool
293 293 YoungList::rs_length_sampling_more() {
294 294 return _curr != NULL;
295 295 }
296 296
297 297 void
298 298 YoungList::rs_length_sampling_next() {
299 299 assert( _curr != NULL, "invariant" );
300 300 size_t rs_length = _curr->rem_set()->occupied();
301 301
302 302 _sampled_rs_lengths += rs_length;
303 303
304 304 // The current region may not yet have been added to the
305 305 // incremental collection set (it gets added when it is
306 306 // retired as the current allocation region).
307 307 if (_curr->in_collection_set()) {
308 308 // Update the collection set policy information for this region
309 309 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
310 310 }
311 311
312 312 _curr = _curr->get_next_young_region();
313 313 if (_curr == NULL) {
314 314 _last_sampled_rs_lengths = _sampled_rs_lengths;
315 315 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
316 316 }
317 317 }
318 318
319 319 void
320 320 YoungList::reset_auxilary_lists() {
321 321 guarantee( is_empty(), "young list should be empty" );
322 322 assert(check_list_well_formed(), "young list should be well formed");
323 323
324 324 // Add survivor regions to SurvRateGroup.
325 325 _g1h->g1_policy()->note_start_adding_survivor_regions();
326 326 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
327 327
328 328 int young_index_in_cset = 0;
329 329 for (HeapRegion* curr = _survivor_head;
330 330 curr != NULL;
331 331 curr = curr->get_next_young_region()) {
332 332 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
333 333
334 334 // The region is a non-empty survivor so let's add it to
335 335 // the incremental collection set for the next evacuation
336 336 // pause.
337 337 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
338 338 young_index_in_cset += 1;
339 339 }
340 340 assert((size_t) young_index_in_cset == _survivor_length,
341 341 "post-condition");
342 342 _g1h->g1_policy()->note_stop_adding_survivor_regions();
343 343
344 344 _head = _survivor_head;
345 345 _length = _survivor_length;
346 346 if (_survivor_head != NULL) {
347 347 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 348 assert(_survivor_length > 0, "invariant");
349 349 _survivor_tail->set_next_young_region(NULL);
350 350 }
351 351
352 352 // Don't clear the survivor list handles until the start of
353 353 // the next evacuation pause - we need it in order to re-tag
354 354 // the survivor regions from this evacuation pause as 'young'
355 355 // at the start of the next.
356 356
357 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
358 358
359 359 assert(check_list_well_formed(), "young list should be well formed");
360 360 }
361 361
362 362 void YoungList::print() {
363 363 HeapRegion* lists[] = {_head, _survivor_head};
364 364 const char* names[] = {"YOUNG", "SURVIVOR"};
365 365
366 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 368 HeapRegion *curr = lists[list];
369 369 if (curr == NULL)
370 370 gclog_or_tty->print_cr(" empty");
371 371 while (curr != NULL) {
372 372 gclog_or_tty->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
373 373 "age: %4d, y: %d, surv: %d",
374 374 curr->bottom(), curr->end(),
375 375 curr->top(),
376 376 curr->prev_top_at_mark_start(),
377 377 curr->next_top_at_mark_start(),
378 378 curr->top_at_conc_mark_count(),
379 379 curr->age_in_surv_rate_group_cond(),
380 380 curr->is_young(),
381 381 curr->is_survivor());
382 382 curr = curr->get_next_young_region();
383 383 }
384 384 }
385 385
386 386 gclog_or_tty->print_cr("");
387 387 }
388 388
389 389 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
390 390 {
391 391 // Claim the right to put the region on the dirty cards region list
392 392 // by installing a self pointer.
393 393 HeapRegion* next = hr->get_next_dirty_cards_region();
394 394 if (next == NULL) {
395 395 HeapRegion* res = (HeapRegion*)
396 396 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
397 397 NULL);
398 398 if (res == NULL) {
399 399 HeapRegion* head;
400 400 do {
401 401 // Put the region to the dirty cards region list.
402 402 head = _dirty_cards_region_list;
403 403 next = (HeapRegion*)
404 404 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
405 405 if (next == head) {
406 406 assert(hr->get_next_dirty_cards_region() == hr,
407 407 "hr->get_next_dirty_cards_region() != hr");
408 408 if (next == NULL) {
409 409 // The last region in the list points to itself.
410 410 hr->set_next_dirty_cards_region(hr);
411 411 } else {
412 412 hr->set_next_dirty_cards_region(next);
413 413 }
414 414 }
415 415 } while (next != head);
416 416 }
417 417 }
418 418 }
419 419
420 420 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
421 421 {
422 422 HeapRegion* head;
423 423 HeapRegion* hr;
424 424 do {
425 425 head = _dirty_cards_region_list;
426 426 if (head == NULL) {
427 427 return NULL;
428 428 }
429 429 HeapRegion* new_head = head->get_next_dirty_cards_region();
430 430 if (head == new_head) {
431 431 // The last region.
432 432 new_head = NULL;
433 433 }
434 434 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
435 435 head);
436 436 } while (hr != head);
437 437 assert(hr != NULL, "invariant");
438 438 hr->set_next_dirty_cards_region(NULL);
439 439 return hr;
440 440 }
441 441
442 442 void G1CollectedHeap::stop_conc_gc_threads() {
443 443 _cg1r->stop();
444 444 _cmThread->stop();
445 445 }
446 446
447 447 #ifdef ASSERT
448 448 // A region is added to the collection set as it is retired
449 449 // so an address p can point to a region which will be in the
450 450 // collection set but has not yet been retired. This method
451 451 // therefore is only accurate during a GC pause after all
452 452 // regions have been retired. It is used for debugging
453 453 // to check if an nmethod has references to objects that can
454 454 // be move during a partial collection. Though it can be
455 455 // inaccurate, it is sufficient for G1 because the conservative
456 456 // implementation of is_scavengable() for G1 will indicate that
457 457 // all nmethods must be scanned during a partial collection.
458 458 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
459 459 HeapRegion* hr = heap_region_containing(p);
460 460 return hr != NULL && hr->in_collection_set();
461 461 }
462 462 #endif
463 463
464 464 // Returns true if the reference points to an object that
465 465 // can move in an incremental collecction.
466 466 bool G1CollectedHeap::is_scavengable(const void* p) {
467 467 G1CollectedHeap* g1h = G1CollectedHeap::heap();
468 468 G1CollectorPolicy* g1p = g1h->g1_policy();
469 469 HeapRegion* hr = heap_region_containing(p);
470 470 if (hr == NULL) {
471 471 // perm gen (or null)
472 472 return false;
473 473 } else {
474 474 return !hr->isHumongous();
475 475 }
476 476 }
477 477
478 478 void G1CollectedHeap::check_ct_logs_at_safepoint() {
479 479 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
480 480 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
481 481
482 482 // Count the dirty cards at the start.
483 483 CountNonCleanMemRegionClosure count1(this);
484 484 ct_bs->mod_card_iterate(&count1);
485 485 int orig_count = count1.n();
486 486
487 487 // First clear the logged cards.
488 488 ClearLoggedCardTableEntryClosure clear;
489 489 dcqs.set_closure(&clear);
490 490 dcqs.apply_closure_to_all_completed_buffers();
491 491 dcqs.iterate_closure_all_threads(false);
492 492 clear.print_histo();
493 493
494 494 // Now ensure that there's no dirty cards.
495 495 CountNonCleanMemRegionClosure count2(this);
496 496 ct_bs->mod_card_iterate(&count2);
497 497 if (count2.n() != 0) {
498 498 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
499 499 count2.n(), orig_count);
500 500 }
501 501 guarantee(count2.n() == 0, "Card table should be clean.");
502 502
503 503 RedirtyLoggedCardTableEntryClosure redirty;
504 504 JavaThread::dirty_card_queue_set().set_closure(&redirty);
505 505 dcqs.apply_closure_to_all_completed_buffers();
506 506 dcqs.iterate_closure_all_threads(false);
507 507 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
508 508 clear.calls(), orig_count);
509 509 guarantee(redirty.calls() == clear.calls(),
510 510 "Or else mechanism is broken.");
511 511
512 512 CountNonCleanMemRegionClosure count3(this);
513 513 ct_bs->mod_card_iterate(&count3);
514 514 if (count3.n() != orig_count) {
515 515 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
516 516 orig_count, count3.n());
517 517 guarantee(count3.n() >= orig_count, "Should have restored them all.");
518 518 }
519 519
520 520 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
521 521 }
522 522
523 523 // Private class members.
524 524
525 525 G1CollectedHeap* G1CollectedHeap::_g1h;
526 526
527 527 // Private methods.
528 528
529 529 HeapRegion*
530 530 G1CollectedHeap::new_region_try_secondary_free_list() {
531 531 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
532 532 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
533 533 if (!_secondary_free_list.is_empty()) {
534 534 if (G1ConcRegionFreeingVerbose) {
535 535 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
536 536 "secondary_free_list has "SIZE_FORMAT" entries",
537 537 _secondary_free_list.length());
538 538 }
539 539 // It looks as if there are free regions available on the
540 540 // secondary_free_list. Let's move them to the free_list and try
541 541 // again to allocate from it.
542 542 append_secondary_free_list();
543 543
544 544 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
545 545 "empty we should have moved at least one entry to the free_list");
546 546 HeapRegion* res = _free_list.remove_head();
547 547 if (G1ConcRegionFreeingVerbose) {
548 548 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
549 549 "allocated "HR_FORMAT" from secondary_free_list",
550 550 HR_FORMAT_PARAMS(res));
551 551 }
552 552 return res;
553 553 }
554 554
555 555 // Wait here until we get notifed either when (a) there are no
556 556 // more free regions coming or (b) some regions have been moved on
557 557 // the secondary_free_list.
558 558 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
559 559 }
560 560
561 561 if (G1ConcRegionFreeingVerbose) {
562 562 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
563 563 "could not allocate from secondary_free_list");
564 564 }
565 565 return NULL;
566 566 }
567 567
568 568 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
569 569 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
570 570 "the only time we use this to allocate a humongous region is "
571 571 "when we are allocating a single humongous region");
572 572
573 573 HeapRegion* res;
574 574 if (G1StressConcRegionFreeing) {
575 575 if (!_secondary_free_list.is_empty()) {
576 576 if (G1ConcRegionFreeingVerbose) {
577 577 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
578 578 "forced to look at the secondary_free_list");
579 579 }
580 580 res = new_region_try_secondary_free_list();
581 581 if (res != NULL) {
582 582 return res;
583 583 }
584 584 }
585 585 }
586 586 res = _free_list.remove_head_or_null();
587 587 if (res == NULL) {
588 588 if (G1ConcRegionFreeingVerbose) {
589 589 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
590 590 "res == NULL, trying the secondary_free_list");
591 591 }
592 592 res = new_region_try_secondary_free_list();
593 593 }
594 594 if (res == NULL && do_expand) {
595 595 ergo_verbose1(ErgoHeapSizing,
596 596 "attempt heap expansion",
597 597 ergo_format_reason("region allocation request failed")
598 598 ergo_format_byte("allocation request"),
599 599 word_size * HeapWordSize);
600 600 if (expand(word_size * HeapWordSize)) {
601 601 // Even though the heap was expanded, it might not have reached
602 602 // the desired size. So, we cannot assume that the allocation
603 603 // will succeed.
604 604 res = _free_list.remove_head_or_null();
605 605 }
606 606 }
607 607 return res;
608 608 }
609 609
610 610 size_t G1CollectedHeap::humongous_obj_allocate_find_first(size_t num_regions,
611 611 size_t word_size) {
612 612 assert(isHumongous(word_size), "word_size should be humongous");
613 613 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
614 614
615 615 size_t first = G1_NULL_HRS_INDEX;
616 616 if (num_regions == 1) {
617 617 // Only one region to allocate, no need to go through the slower
618 618 // path. The caller will attempt the expasion if this fails, so
619 619 // let's not try to expand here too.
620 620 HeapRegion* hr = new_region(word_size, false /* do_expand */);
621 621 if (hr != NULL) {
622 622 first = hr->hrs_index();
623 623 } else {
624 624 first = G1_NULL_HRS_INDEX;
625 625 }
626 626 } else {
627 627 // We can't allocate humongous regions while cleanupComplete() is
628 628 // running, since some of the regions we find to be empty might not
629 629 // yet be added to the free list and it is not straightforward to
630 630 // know which list they are on so that we can remove them. Note
631 631 // that we only need to do this if we need to allocate more than
632 632 // one region to satisfy the current humongous allocation
633 633 // request. If we are only allocating one region we use the common
634 634 // region allocation code (see above).
635 635 wait_while_free_regions_coming();
636 636 append_secondary_free_list_if_not_empty_with_lock();
637 637
638 638 if (free_regions() >= num_regions) {
639 639 first = _hrs.find_contiguous(num_regions);
640 640 if (first != G1_NULL_HRS_INDEX) {
641 641 for (size_t i = first; i < first + num_regions; ++i) {
642 642 HeapRegion* hr = region_at(i);
643 643 assert(hr->is_empty(), "sanity");
644 644 assert(is_on_master_free_list(hr), "sanity");
645 645 hr->set_pending_removal(true);
646 646 }
647 647 _free_list.remove_all_pending(num_regions);
648 648 }
649 649 }
650 650 }
651 651 return first;
652 652 }
653 653
654 654 HeapWord*
655 655 G1CollectedHeap::humongous_obj_allocate_initialize_regions(size_t first,
656 656 size_t num_regions,
657 657 size_t word_size) {
658 658 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
659 659 assert(isHumongous(word_size), "word_size should be humongous");
660 660 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
661 661
662 662 // Index of last region in the series + 1.
663 663 size_t last = first + num_regions;
664 664
665 665 // We need to initialize the region(s) we just discovered. This is
666 666 // a bit tricky given that it can happen concurrently with
667 667 // refinement threads refining cards on these regions and
668 668 // potentially wanting to refine the BOT as they are scanning
669 669 // those cards (this can happen shortly after a cleanup; see CR
670 670 // 6991377). So we have to set up the region(s) carefully and in
671 671 // a specific order.
672 672
673 673 // The word size sum of all the regions we will allocate.
674 674 size_t word_size_sum = num_regions * HeapRegion::GrainWords;
675 675 assert(word_size <= word_size_sum, "sanity");
676 676
677 677 // This will be the "starts humongous" region.
678 678 HeapRegion* first_hr = region_at(first);
679 679 // The header of the new object will be placed at the bottom of
680 680 // the first region.
681 681 HeapWord* new_obj = first_hr->bottom();
682 682 // This will be the new end of the first region in the series that
683 683 // should also match the end of the last region in the seriers.
684 684 HeapWord* new_end = new_obj + word_size_sum;
685 685 // This will be the new top of the first region that will reflect
686 686 // this allocation.
687 687 HeapWord* new_top = new_obj + word_size;
688 688
689 689 // First, we need to zero the header of the space that we will be
690 690 // allocating. When we update top further down, some refinement
691 691 // threads might try to scan the region. By zeroing the header we
692 692 // ensure that any thread that will try to scan the region will
693 693 // come across the zero klass word and bail out.
694 694 //
695 695 // NOTE: It would not have been correct to have used
696 696 // CollectedHeap::fill_with_object() and make the space look like
697 697 // an int array. The thread that is doing the allocation will
698 698 // later update the object header to a potentially different array
699 699 // type and, for a very short period of time, the klass and length
700 700 // fields will be inconsistent. This could cause a refinement
701 701 // thread to calculate the object size incorrectly.
702 702 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
703 703
704 704 // We will set up the first region as "starts humongous". This
705 705 // will also update the BOT covering all the regions to reflect
706 706 // that there is a single object that starts at the bottom of the
707 707 // first region.
708 708 first_hr->set_startsHumongous(new_top, new_end);
709 709
710 710 // Then, if there are any, we will set up the "continues
711 711 // humongous" regions.
712 712 HeapRegion* hr = NULL;
713 713 for (size_t i = first + 1; i < last; ++i) {
714 714 hr = region_at(i);
715 715 hr->set_continuesHumongous(first_hr);
716 716 }
717 717 // If we have "continues humongous" regions (hr != NULL), then the
718 718 // end of the last one should match new_end.
719 719 assert(hr == NULL || hr->end() == new_end, "sanity");
720 720
721 721 // Up to this point no concurrent thread would have been able to
722 722 // do any scanning on any region in this series. All the top
723 723 // fields still point to bottom, so the intersection between
724 724 // [bottom,top] and [card_start,card_end] will be empty. Before we
725 725 // update the top fields, we'll do a storestore to make sure that
726 726 // no thread sees the update to top before the zeroing of the
727 727 // object header and the BOT initialization.
728 728 OrderAccess::storestore();
729 729
730 730 // Now that the BOT and the object header have been initialized,
731 731 // we can update top of the "starts humongous" region.
732 732 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
733 733 "new_top should be in this region");
734 734 first_hr->set_top(new_top);
735 735 if (_hr_printer.is_active()) {
736 736 HeapWord* bottom = first_hr->bottom();
737 737 HeapWord* end = first_hr->orig_end();
738 738 if ((first + 1) == last) {
739 739 // the series has a single humongous region
740 740 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
741 741 } else {
742 742 // the series has more than one humongous regions
743 743 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
744 744 }
745 745 }
746 746
747 747 // Now, we will update the top fields of the "continues humongous"
748 748 // regions. The reason we need to do this is that, otherwise,
749 749 // these regions would look empty and this will confuse parts of
750 750 // G1. For example, the code that looks for a consecutive number
751 751 // of empty regions will consider them empty and try to
752 752 // re-allocate them. We can extend is_empty() to also include
753 753 // !continuesHumongous(), but it is easier to just update the top
754 754 // fields here. The way we set top for all regions (i.e., top ==
755 755 // end for all regions but the last one, top == new_top for the
756 756 // last one) is actually used when we will free up the humongous
757 757 // region in free_humongous_region().
758 758 hr = NULL;
759 759 for (size_t i = first + 1; i < last; ++i) {
760 760 hr = region_at(i);
761 761 if ((i + 1) == last) {
762 762 // last continues humongous region
763 763 assert(hr->bottom() < new_top && new_top <= hr->end(),
764 764 "new_top should fall on this region");
765 765 hr->set_top(new_top);
766 766 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
767 767 } else {
768 768 // not last one
769 769 assert(new_top > hr->end(), "new_top should be above this region");
770 770 hr->set_top(hr->end());
771 771 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
772 772 }
773 773 }
774 774 // If we have continues humongous regions (hr != NULL), then the
775 775 // end of the last one should match new_end and its top should
776 776 // match new_top.
777 777 assert(hr == NULL ||
778 778 (hr->end() == new_end && hr->top() == new_top), "sanity");
779 779
780 780 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
781 781 _summary_bytes_used += first_hr->used();
782 782 _humongous_set.add(first_hr);
783 783
784 784 return new_obj;
785 785 }
786 786
787 787 // If could fit into free regions w/o expansion, try.
788 788 // Otherwise, if can expand, do so.
789 789 // Otherwise, if using ex regions might help, try with ex given back.
790 790 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
791 791 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
792 792
793 793 verify_region_sets_optional();
794 794
795 795 size_t num_regions =
796 796 round_to(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords;
797 797 size_t x_size = expansion_regions();
798 798 size_t fs = _hrs.free_suffix();
799 799 size_t first = humongous_obj_allocate_find_first(num_regions, word_size);
800 800 if (first == G1_NULL_HRS_INDEX) {
801 801 // The only thing we can do now is attempt expansion.
802 802 if (fs + x_size >= num_regions) {
803 803 // If the number of regions we're trying to allocate for this
804 804 // object is at most the number of regions in the free suffix,
805 805 // then the call to humongous_obj_allocate_find_first() above
806 806 // should have succeeded and we wouldn't be here.
807 807 //
808 808 // We should only be trying to expand when the free suffix is
809 809 // not sufficient for the object _and_ we have some expansion
810 810 // room available.
811 811 assert(num_regions > fs, "earlier allocation should have succeeded");
812 812
813 813 ergo_verbose1(ErgoHeapSizing,
814 814 "attempt heap expansion",
815 815 ergo_format_reason("humongous allocation request failed")
816 816 ergo_format_byte("allocation request"),
817 817 word_size * HeapWordSize);
818 818 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
819 819 // Even though the heap was expanded, it might not have
820 820 // reached the desired size. So, we cannot assume that the
821 821 // allocation will succeed.
822 822 first = humongous_obj_allocate_find_first(num_regions, word_size);
823 823 }
824 824 }
825 825 }
826 826
827 827 HeapWord* result = NULL;
828 828 if (first != G1_NULL_HRS_INDEX) {
829 829 result =
830 830 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
831 831 assert(result != NULL, "it should always return a valid result");
832 832
833 833 // A successful humongous object allocation changes the used space
834 834 // information of the old generation so we need to recalculate the
835 835 // sizes and update the jstat counters here.
836 836 g1mm()->update_sizes();
837 837 }
838 838
839 839 verify_region_sets_optional();
840 840
841 841 return result;
842 842 }
843 843
844 844 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
845 845 assert_heap_not_locked_and_not_at_safepoint();
846 846 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
847 847
848 848 unsigned int dummy_gc_count_before;
849 849 return attempt_allocation(word_size, &dummy_gc_count_before);
850 850 }
851 851
852 852 HeapWord*
853 853 G1CollectedHeap::mem_allocate(size_t word_size,
854 854 bool* gc_overhead_limit_was_exceeded) {
855 855 assert_heap_not_locked_and_not_at_safepoint();
856 856
857 857 // Loop until the allocation is satisified, or unsatisfied after GC.
858 858 for (int try_count = 1; /* we'll return */; try_count += 1) {
859 859 unsigned int gc_count_before;
860 860
861 861 HeapWord* result = NULL;
862 862 if (!isHumongous(word_size)) {
863 863 result = attempt_allocation(word_size, &gc_count_before);
864 864 } else {
865 865 result = attempt_allocation_humongous(word_size, &gc_count_before);
866 866 }
867 867 if (result != NULL) {
868 868 return result;
869 869 }
870 870
871 871 // Create the garbage collection operation...
872 872 VM_G1CollectForAllocation op(gc_count_before, word_size);
873 873 // ...and get the VM thread to execute it.
874 874 VMThread::execute(&op);
875 875
876 876 if (op.prologue_succeeded() && op.pause_succeeded()) {
877 877 // If the operation was successful we'll return the result even
878 878 // if it is NULL. If the allocation attempt failed immediately
879 879 // after a Full GC, it's unlikely we'll be able to allocate now.
880 880 HeapWord* result = op.result();
881 881 if (result != NULL && !isHumongous(word_size)) {
882 882 // Allocations that take place on VM operations do not do any
883 883 // card dirtying and we have to do it here. We only have to do
884 884 // this for non-humongous allocations, though.
885 885 dirty_young_block(result, word_size);
886 886 }
887 887 return result;
888 888 } else {
889 889 assert(op.result() == NULL,
890 890 "the result should be NULL if the VM op did not succeed");
891 891 }
892 892
893 893 // Give a warning if we seem to be looping forever.
894 894 if ((QueuedAllocationWarningCount > 0) &&
895 895 (try_count % QueuedAllocationWarningCount == 0)) {
896 896 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
897 897 }
898 898 }
899 899
900 900 ShouldNotReachHere();
901 901 return NULL;
902 902 }
903 903
904 904 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
905 905 unsigned int *gc_count_before_ret) {
906 906 // Make sure you read the note in attempt_allocation_humongous().
907 907
908 908 assert_heap_not_locked_and_not_at_safepoint();
909 909 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
910 910 "be called for humongous allocation requests");
911 911
912 912 // We should only get here after the first-level allocation attempt
913 913 // (attempt_allocation()) failed to allocate.
914 914
915 915 // We will loop until a) we manage to successfully perform the
916 916 // allocation or b) we successfully schedule a collection which
917 917 // fails to perform the allocation. b) is the only case when we'll
918 918 // return NULL.
919 919 HeapWord* result = NULL;
920 920 for (int try_count = 1; /* we'll return */; try_count += 1) {
921 921 bool should_try_gc;
922 922 unsigned int gc_count_before;
923 923
924 924 {
925 925 MutexLockerEx x(Heap_lock);
926 926
927 927 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
928 928 false /* bot_updates */);
929 929 if (result != NULL) {
930 930 return result;
931 931 }
932 932
933 933 // If we reach here, attempt_allocation_locked() above failed to
934 934 // allocate a new region. So the mutator alloc region should be NULL.
935 935 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
936 936
937 937 if (GC_locker::is_active_and_needs_gc()) {
938 938 if (g1_policy()->can_expand_young_list()) {
939 939 // No need for an ergo verbose message here,
940 940 // can_expand_young_list() does this when it returns true.
941 941 result = _mutator_alloc_region.attempt_allocation_force(word_size,
942 942 false /* bot_updates */);
943 943 if (result != NULL) {
944 944 return result;
945 945 }
946 946 }
947 947 should_try_gc = false;
948 948 } else {
949 949 // Read the GC count while still holding the Heap_lock.
950 950 gc_count_before = SharedHeap::heap()->total_collections();
951 951 should_try_gc = true;
952 952 }
953 953 }
954 954
955 955 if (should_try_gc) {
956 956 bool succeeded;
957 957 result = do_collection_pause(word_size, gc_count_before, &succeeded);
958 958 if (result != NULL) {
959 959 assert(succeeded, "only way to get back a non-NULL result");
960 960 return result;
961 961 }
962 962
963 963 if (succeeded) {
964 964 // If we get here we successfully scheduled a collection which
965 965 // failed to allocate. No point in trying to allocate
966 966 // further. We'll just return NULL.
967 967 MutexLockerEx x(Heap_lock);
968 968 *gc_count_before_ret = SharedHeap::heap()->total_collections();
969 969 return NULL;
970 970 }
971 971 } else {
972 972 GC_locker::stall_until_clear();
973 973 }
974 974
975 975 // We can reach here if we were unsuccessul in scheduling a
976 976 // collection (because another thread beat us to it) or if we were
977 977 // stalled due to the GC locker. In either can we should retry the
978 978 // allocation attempt in case another thread successfully
979 979 // performed a collection and reclaimed enough space. We do the
980 980 // first attempt (without holding the Heap_lock) here and the
981 981 // follow-on attempt will be at the start of the next loop
982 982 // iteration (after taking the Heap_lock).
983 983 result = _mutator_alloc_region.attempt_allocation(word_size,
984 984 false /* bot_updates */);
985 985 if (result != NULL ){
986 986 return result;
987 987 }
988 988
989 989 // Give a warning if we seem to be looping forever.
990 990 if ((QueuedAllocationWarningCount > 0) &&
991 991 (try_count % QueuedAllocationWarningCount == 0)) {
992 992 warning("G1CollectedHeap::attempt_allocation_slow() "
993 993 "retries %d times", try_count);
994 994 }
995 995 }
996 996
997 997 ShouldNotReachHere();
998 998 return NULL;
999 999 }
1000 1000
1001 1001 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1002 1002 unsigned int * gc_count_before_ret) {
1003 1003 // The structure of this method has a lot of similarities to
1004 1004 // attempt_allocation_slow(). The reason these two were not merged
1005 1005 // into a single one is that such a method would require several "if
1006 1006 // allocation is not humongous do this, otherwise do that"
1007 1007 // conditional paths which would obscure its flow. In fact, an early
1008 1008 // version of this code did use a unified method which was harder to
1009 1009 // follow and, as a result, it had subtle bugs that were hard to
1010 1010 // track down. So keeping these two methods separate allows each to
1011 1011 // be more readable. It will be good to keep these two in sync as
1012 1012 // much as possible.
1013 1013
1014 1014 assert_heap_not_locked_and_not_at_safepoint();
1015 1015 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1016 1016 "should only be called for humongous allocations");
1017 1017
1018 1018 // We will loop until a) we manage to successfully perform the
1019 1019 // allocation or b) we successfully schedule a collection which
1020 1020 // fails to perform the allocation. b) is the only case when we'll
1021 1021 // return NULL.
1022 1022 HeapWord* result = NULL;
1023 1023 for (int try_count = 1; /* we'll return */; try_count += 1) {
1024 1024 bool should_try_gc;
1025 1025 unsigned int gc_count_before;
1026 1026
1027 1027 {
1028 1028 MutexLockerEx x(Heap_lock);
1029 1029
1030 1030 // Given that humongous objects are not allocated in young
1031 1031 // regions, we'll first try to do the allocation without doing a
1032 1032 // collection hoping that there's enough space in the heap.
1033 1033 result = humongous_obj_allocate(word_size);
1034 1034 if (result != NULL) {
1035 1035 return result;
1036 1036 }
1037 1037
1038 1038 if (GC_locker::is_active_and_needs_gc()) {
1039 1039 should_try_gc = false;
1040 1040 } else {
1041 1041 // Read the GC count while still holding the Heap_lock.
1042 1042 gc_count_before = SharedHeap::heap()->total_collections();
1043 1043 should_try_gc = true;
1044 1044 }
1045 1045 }
1046 1046
1047 1047 if (should_try_gc) {
1048 1048 // If we failed to allocate the humongous object, we should try to
1049 1049 // do a collection pause (if we're allowed) in case it reclaims
1050 1050 // enough space for the allocation to succeed after the pause.
1051 1051
1052 1052 bool succeeded;
1053 1053 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1054 1054 if (result != NULL) {
1055 1055 assert(succeeded, "only way to get back a non-NULL result");
1056 1056 return result;
1057 1057 }
1058 1058
1059 1059 if (succeeded) {
1060 1060 // If we get here we successfully scheduled a collection which
1061 1061 // failed to allocate. No point in trying to allocate
1062 1062 // further. We'll just return NULL.
1063 1063 MutexLockerEx x(Heap_lock);
1064 1064 *gc_count_before_ret = SharedHeap::heap()->total_collections();
1065 1065 return NULL;
1066 1066 }
1067 1067 } else {
1068 1068 GC_locker::stall_until_clear();
1069 1069 }
1070 1070
1071 1071 // We can reach here if we were unsuccessul in scheduling a
1072 1072 // collection (because another thread beat us to it) or if we were
1073 1073 // stalled due to the GC locker. In either can we should retry the
1074 1074 // allocation attempt in case another thread successfully
1075 1075 // performed a collection and reclaimed enough space. Give a
1076 1076 // warning if we seem to be looping forever.
1077 1077
1078 1078 if ((QueuedAllocationWarningCount > 0) &&
1079 1079 (try_count % QueuedAllocationWarningCount == 0)) {
1080 1080 warning("G1CollectedHeap::attempt_allocation_humongous() "
1081 1081 "retries %d times", try_count);
1082 1082 }
1083 1083 }
1084 1084
1085 1085 ShouldNotReachHere();
1086 1086 return NULL;
1087 1087 }
1088 1088
1089 1089 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1090 1090 bool expect_null_mutator_alloc_region) {
1091 1091 assert_at_safepoint(true /* should_be_vm_thread */);
1092 1092 assert(_mutator_alloc_region.get() == NULL ||
1093 1093 !expect_null_mutator_alloc_region,
1094 1094 "the current alloc region was unexpectedly found to be non-NULL");
1095 1095
1096 1096 if (!isHumongous(word_size)) {
1097 1097 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1098 1098 false /* bot_updates */);
1099 1099 } else {
1100 1100 return humongous_obj_allocate(word_size);
1101 1101 }
1102 1102
1103 1103 ShouldNotReachHere();
1104 1104 }
1105 1105
1106 1106 class PostMCRemSetClearClosure: public HeapRegionClosure {
1107 1107 ModRefBarrierSet* _mr_bs;
1108 1108 public:
1109 1109 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1110 1110 bool doHeapRegion(HeapRegion* r) {
1111 1111 r->reset_gc_time_stamp();
1112 1112 if (r->continuesHumongous())
1113 1113 return false;
1114 1114 HeapRegionRemSet* hrrs = r->rem_set();
1115 1115 if (hrrs != NULL) hrrs->clear();
1116 1116 // You might think here that we could clear just the cards
1117 1117 // corresponding to the used region. But no: if we leave a dirty card
1118 1118 // in a region we might allocate into, then it would prevent that card
1119 1119 // from being enqueued, and cause it to be missed.
1120 1120 // Re: the performance cost: we shouldn't be doing full GC anyway!
1121 1121 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1122 1122 return false;
1123 1123 }
1124 1124 };
1125 1125
1126 1126
1127 1127 class PostMCRemSetInvalidateClosure: public HeapRegionClosure {
1128 1128 ModRefBarrierSet* _mr_bs;
1129 1129 public:
1130 1130 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1131 1131 bool doHeapRegion(HeapRegion* r) {
1132 1132 if (r->continuesHumongous()) return false;
1133 1133 if (r->used_region().word_size() != 0) {
1134 1134 _mr_bs->invalidate(r->used_region(), true /*whole heap*/);
1135 1135 }
1136 1136 return false;
1137 1137 }
1138 1138 };
1139 1139
1140 1140 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1141 1141 G1CollectedHeap* _g1h;
1142 1142 UpdateRSOopClosure _cl;
1143 1143 int _worker_i;
1144 1144 public:
1145 1145 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1146 1146 _cl(g1->g1_rem_set(), worker_i),
1147 1147 _worker_i(worker_i),
1148 1148 _g1h(g1)
1149 1149 { }
1150 1150
1151 1151 bool doHeapRegion(HeapRegion* r) {
1152 1152 if (!r->continuesHumongous()) {
1153 1153 _cl.set_from(r);
1154 1154 r->oop_iterate(&_cl);
1155 1155 }
1156 1156 return false;
1157 1157 }
1158 1158 };
1159 1159
1160 1160 class ParRebuildRSTask: public AbstractGangTask {
1161 1161 G1CollectedHeap* _g1;
1162 1162 public:
1163 1163 ParRebuildRSTask(G1CollectedHeap* g1)
1164 1164 : AbstractGangTask("ParRebuildRSTask"),
1165 1165 _g1(g1)
1166 1166 { }
1167 1167
1168 1168 void work(int i) {
1169 1169 RebuildRSOutOfRegionClosure rebuild_rs(_g1, i);
1170 1170 _g1->heap_region_par_iterate_chunked(&rebuild_rs, i,
1171 1171 _g1->workers()->active_workers(),
1172 1172 HeapRegion::RebuildRSClaimValue);
1173 1173 }
1174 1174 };
1175 1175
1176 1176 class PostCompactionPrinterClosure: public HeapRegionClosure {
1177 1177 private:
1178 1178 G1HRPrinter* _hr_printer;
1179 1179 public:
1180 1180 bool doHeapRegion(HeapRegion* hr) {
1181 1181 assert(!hr->is_young(), "not expecting to find young regions");
1182 1182 // We only generate output for non-empty regions.
1183 1183 if (!hr->is_empty()) {
1184 1184 if (!hr->isHumongous()) {
1185 1185 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1186 1186 } else if (hr->startsHumongous()) {
1187 1187 if (hr->capacity() == HeapRegion::GrainBytes) {
1188 1188 // single humongous region
1189 1189 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1190 1190 } else {
1191 1191 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1192 1192 }
1193 1193 } else {
1194 1194 assert(hr->continuesHumongous(), "only way to get here");
1195 1195 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1196 1196 }
1197 1197 }
1198 1198 return false;
1199 1199 }
1200 1200
1201 1201 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1202 1202 : _hr_printer(hr_printer) { }
1203 1203 };
1204 1204
1205 1205 bool G1CollectedHeap::do_collection(bool explicit_gc,
1206 1206 bool clear_all_soft_refs,
1207 1207 size_t word_size) {
1208 1208 assert_at_safepoint(true /* should_be_vm_thread */);
1209 1209
1210 1210 if (GC_locker::check_active_before_gc()) {
1211 1211 return false;
1212 1212 }
1213 1213
1214 1214 SvcGCMarker sgcm(SvcGCMarker::FULL);
1215 1215 ResourceMark rm;
1216 1216
1217 1217 if (PrintHeapAtGC) {
1218 1218 Universe::print_heap_before_gc();
1219 1219 }
1220 1220
1221 1221 HRSPhaseSetter x(HRSPhaseFullGC);
1222 1222 verify_region_sets_optional();
1223 1223
1224 1224 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1225 1225 collector_policy()->should_clear_all_soft_refs();
1226 1226
1227 1227 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1228 1228
1229 1229 {
1230 1230 IsGCActiveMark x;
1231 1231
1232 1232 // Timing
1233 1233 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc);
1234 1234 assert(!system_gc || explicit_gc, "invariant");
1235 1235 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1236 1236 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1237 1237 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC",
1238 1238 PrintGC, true, gclog_or_tty);
1239 1239
1240 1240 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1241 1241 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1242 1242
1243 1243 double start = os::elapsedTime();
1244 1244 g1_policy()->record_full_collection_start();
1245 1245
1246 1246 wait_while_free_regions_coming();
1247 1247 append_secondary_free_list_if_not_empty_with_lock();
1248 1248
1249 1249 gc_prologue(true);
1250 1250 increment_total_collections(true /* full gc */);
1251 1251
1252 1252 size_t g1h_prev_used = used();
1253 1253 assert(used() == recalculate_used(), "Should be equal");
1254 1254
1255 1255 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
1256 1256 HandleMark hm; // Discard invalid handles created during verification
1257 1257 gclog_or_tty->print(" VerifyBeforeGC:");
1258 1258 prepare_for_verify();
1259 1259 Universe::verify(/* allow dirty */ true,
1260 1260 /* silent */ false,
1261 1261 /* option */ VerifyOption_G1UsePrevMarking);
1262 1262
1263 1263 }
1264 1264 pre_full_gc_dump();
1265 1265
1266 1266 COMPILER2_PRESENT(DerivedPointerTable::clear());
1267 1267
1268 1268 // Disable discovery and empty the discovered lists
1269 1269 // for the CM ref processor.
1270 1270 ref_processor_cm()->disable_discovery();
1271 1271 ref_processor_cm()->abandon_partial_discovery();
1272 1272 ref_processor_cm()->verify_no_references_recorded();
1273 1273
1274 1274 // Abandon current iterations of concurrent marking and concurrent
1275 1275 // refinement, if any are in progress.
1276 1276 concurrent_mark()->abort();
1277 1277
1278 1278 // Make sure we'll choose a new allocation region afterwards.
1279 1279 release_mutator_alloc_region();
1280 1280 abandon_gc_alloc_regions();
1281 1281 g1_rem_set()->cleanupHRRS();
1282 1282
1283 1283 // We should call this after we retire any currently active alloc
1284 1284 // regions so that all the ALLOC / RETIRE events are generated
1285 1285 // before the start GC event.
1286 1286 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1287 1287
1288 1288 // We may have added regions to the current incremental collection
1289 1289 // set between the last GC or pause and now. We need to clear the
1290 1290 // incremental collection set and then start rebuilding it afresh
1291 1291 // after this full GC.
1292 1292 abandon_collection_set(g1_policy()->inc_cset_head());
1293 1293 g1_policy()->clear_incremental_cset();
1294 1294 g1_policy()->stop_incremental_cset_building();
1295 1295
1296 1296 tear_down_region_sets(false /* free_list_only */);
1297 1297 g1_policy()->set_full_young_gcs(true);
1298 1298
1299 1299 // See the comments in g1CollectedHeap.hpp and
1300 1300 // G1CollectedHeap::ref_processing_init() about
1301 1301 // how reference processing currently works in G1.
1302 1302
1303 1303 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1304 1304 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1305 1305
1306 1306 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1307 1307 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1308 1308
1309 1309 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1310 1310 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1311 1311
1312 1312 // Do collection work
1313 1313 {
1314 1314 HandleMark hm; // Discard invalid handles created during gc
1315 1315 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1316 1316 }
1317 1317
1318 1318 assert(free_regions() == 0, "we should not have added any free regions");
1319 1319 rebuild_region_sets(false /* free_list_only */);
1320 1320
1321 1321 // Enqueue any discovered reference objects that have
1322 1322 // not been removed from the discovered lists.
1323 1323 ref_processor_stw()->enqueue_discovered_references();
1324 1324
1325 1325 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1326 1326
1327 1327 MemoryService::track_memory_usage();
1328 1328
1329 1329 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
1330 1330 HandleMark hm; // Discard invalid handles created during verification
1331 1331 gclog_or_tty->print(" VerifyAfterGC:");
1332 1332 prepare_for_verify();
1333 1333 Universe::verify(/* allow dirty */ false,
1334 1334 /* silent */ false,
1335 1335 /* option */ VerifyOption_G1UsePrevMarking);
1336 1336
1337 1337 }
1338 1338
1339 1339 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1340 1340 ref_processor_stw()->verify_no_references_recorded();
1341 1341
1342 1342 // Note: since we've just done a full GC, concurrent
1343 1343 // marking is no longer active. Therefore we need not
1344 1344 // re-enable reference discovery for the CM ref processor.
1345 1345 // That will be done at the start of the next marking cycle.
1346 1346 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1347 1347 ref_processor_cm()->verify_no_references_recorded();
1348 1348
1349 1349 reset_gc_time_stamp();
1350 1350 // Since everything potentially moved, we will clear all remembered
1351 1351 // sets, and clear all cards. Later we will rebuild remebered
1352 1352 // sets. We will also reset the GC time stamps of the regions.
1353 1353 PostMCRemSetClearClosure rs_clear(mr_bs());
1354 1354 heap_region_iterate(&rs_clear);
1355 1355
1356 1356 // Resize the heap if necessary.
1357 1357 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1358 1358
1359 1359 if (_hr_printer.is_active()) {
1360 1360 // We should do this after we potentially resize the heap so
1361 1361 // that all the COMMIT / UNCOMMIT events are generated before
1362 1362 // the end GC event.
1363 1363
1364 1364 PostCompactionPrinterClosure cl(hr_printer());
1365 1365 heap_region_iterate(&cl);
1366 1366
1367 1367 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1368 1368 }
1369 1369
1370 1370 if (_cg1r->use_cache()) {
1371 1371 _cg1r->clear_and_record_card_counts();
1372 1372 _cg1r->clear_hot_cache();
1373 1373 }
1374 1374
1375 1375 // Rebuild remembered sets of all regions.
1376 1376 if (G1CollectedHeap::use_parallel_gc_threads()) {
1377 1377 int n_workers =
1378 1378 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1379 1379 workers()->active_workers(),
1380 1380 Threads::number_of_non_daemon_threads());
1381 1381 assert(UseDynamicNumberOfGCThreads ||
1382 1382 n_workers == workers()->total_workers(),
1383 1383 "If not dynamic should be using all the workers");
1384 1384 workers()->set_active_workers(n_workers);
1385 1385 // Set parallel threads in the heap (_n_par_threads) only
1386 1386 // before a parallel phase and always reset it to 0 after
1387 1387 // the phase so that the number of parallel threads does
1388 1388 // no get carried forward to a serial phase where there
1389 1389 // may be code that is "possibly_parallel".
1390 1390 set_par_threads(n_workers);
1391 1391
1392 1392 ParRebuildRSTask rebuild_rs_task(this);
1393 1393 assert(check_heap_region_claim_values(
1394 1394 HeapRegion::InitialClaimValue), "sanity check");
1395 1395 assert(UseDynamicNumberOfGCThreads ||
1396 1396 workers()->active_workers() == workers()->total_workers(),
1397 1397 "Unless dynamic should use total workers");
1398 1398 // Use the most recent number of active workers
1399 1399 assert(workers()->active_workers() > 0,
1400 1400 "Active workers not properly set");
1401 1401 set_par_threads(workers()->active_workers());
1402 1402 workers()->run_task(&rebuild_rs_task);
1403 1403 set_par_threads(0);
1404 1404 assert(check_heap_region_claim_values(
1405 1405 HeapRegion::RebuildRSClaimValue), "sanity check");
1406 1406 reset_heap_region_claim_values();
1407 1407 } else {
1408 1408 RebuildRSOutOfRegionClosure rebuild_rs(this);
1409 1409 heap_region_iterate(&rebuild_rs);
1410 1410 }
1411 1411
1412 1412 if (PrintGC) {
1413 1413 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1414 1414 }
1415 1415
1416 1416 if (true) { // FIXME
1417 1417 // Ask the permanent generation to adjust size for full collections
1418 1418 perm()->compute_new_size();
1419 1419 }
1420 1420
1421 1421 // Start a new incremental collection set for the next pause
1422 1422 assert(g1_policy()->collection_set() == NULL, "must be");
1423 1423 g1_policy()->start_incremental_cset_building();
1424 1424
1425 1425 // Clear the _cset_fast_test bitmap in anticipation of adding
1426 1426 // regions to the incremental collection set for the next
1427 1427 // evacuation pause.
1428 1428 clear_cset_fast_test();
1429 1429
1430 1430 init_mutator_alloc_region();
1431 1431
1432 1432 double end = os::elapsedTime();
1433 1433 g1_policy()->record_full_collection_end();
1434 1434
1435 1435 #ifdef TRACESPINNING
1436 1436 ParallelTaskTerminator::print_termination_counts();
1437 1437 #endif
1438 1438
1439 1439 gc_epilogue(true);
1440 1440
1441 1441 // Discard all rset updates
1442 1442 JavaThread::dirty_card_queue_set().abandon_logs();
1443 1443 assert(!G1DeferredRSUpdate
1444 1444 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1445 1445 }
1446 1446
1447 1447 _young_list->reset_sampled_info();
1448 1448 // At this point there should be no regions in the
1449 1449 // entire heap tagged as young.
1450 1450 assert( check_young_list_empty(true /* check_heap */),
1451 1451 "young list should be empty at this point");
1452 1452
1453 1453 // Update the number of full collections that have been completed.
1454 1454 increment_full_collections_completed(false /* concurrent */);
1455 1455
1456 1456 _hrs.verify_optional();
1457 1457 verify_region_sets_optional();
1458 1458
1459 1459 if (PrintHeapAtGC) {
1460 1460 Universe::print_heap_after_gc();
1461 1461 }
1462 1462 g1mm()->update_sizes();
1463 1463 post_full_gc_dump();
1464 1464
1465 1465 return true;
1466 1466 }
1467 1467
1468 1468 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1469 1469 // do_collection() will return whether it succeeded in performing
1470 1470 // the GC. Currently, there is no facility on the
1471 1471 // do_full_collection() API to notify the caller than the collection
1472 1472 // did not succeed (e.g., because it was locked out by the GC
1473 1473 // locker). So, right now, we'll ignore the return value.
1474 1474 bool dummy = do_collection(true, /* explicit_gc */
1475 1475 clear_all_soft_refs,
1476 1476 0 /* word_size */);
1477 1477 }
1478 1478
1479 1479 // This code is mostly copied from TenuredGeneration.
1480 1480 void
1481 1481 G1CollectedHeap::
1482 1482 resize_if_necessary_after_full_collection(size_t word_size) {
1483 1483 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1484 1484
1485 1485 // Include the current allocation, if any, and bytes that will be
1486 1486 // pre-allocated to support collections, as "used".
1487 1487 const size_t used_after_gc = used();
1488 1488 const size_t capacity_after_gc = capacity();
1489 1489 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1490 1490
1491 1491 // This is enforced in arguments.cpp.
1492 1492 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1493 1493 "otherwise the code below doesn't make sense");
1494 1494
1495 1495 // We don't have floating point command-line arguments
1496 1496 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1497 1497 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1498 1498 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1499 1499 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1500 1500
1501 1501 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1502 1502 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1503 1503
1504 1504 // We have to be careful here as these two calculations can overflow
1505 1505 // 32-bit size_t's.
1506 1506 double used_after_gc_d = (double) used_after_gc;
1507 1507 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1508 1508 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1509 1509
1510 1510 // Let's make sure that they are both under the max heap size, which
1511 1511 // by default will make them fit into a size_t.
1512 1512 double desired_capacity_upper_bound = (double) max_heap_size;
1513 1513 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1514 1514 desired_capacity_upper_bound);
1515 1515 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1516 1516 desired_capacity_upper_bound);
1517 1517
1518 1518 // We can now safely turn them into size_t's.
1519 1519 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1520 1520 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1521 1521
1522 1522 // This assert only makes sense here, before we adjust them
1523 1523 // with respect to the min and max heap size.
1524 1524 assert(minimum_desired_capacity <= maximum_desired_capacity,
1525 1525 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1526 1526 "maximum_desired_capacity = "SIZE_FORMAT,
1527 1527 minimum_desired_capacity, maximum_desired_capacity));
1528 1528
1529 1529 // Should not be greater than the heap max size. No need to adjust
1530 1530 // it with respect to the heap min size as it's a lower bound (i.e.,
1531 1531 // we'll try to make the capacity larger than it, not smaller).
1532 1532 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1533 1533 // Should not be less than the heap min size. No need to adjust it
1534 1534 // with respect to the heap max size as it's an upper bound (i.e.,
1535 1535 // we'll try to make the capacity smaller than it, not greater).
1536 1536 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1537 1537
1538 1538 if (capacity_after_gc < minimum_desired_capacity) {
1539 1539 // Don't expand unless it's significant
1540 1540 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1541 1541 ergo_verbose4(ErgoHeapSizing,
1542 1542 "attempt heap expansion",
1543 1543 ergo_format_reason("capacity lower than "
1544 1544 "min desired capacity after Full GC")
1545 1545 ergo_format_byte("capacity")
1546 1546 ergo_format_byte("occupancy")
1547 1547 ergo_format_byte_perc("min desired capacity"),
1548 1548 capacity_after_gc, used_after_gc,
1549 1549 minimum_desired_capacity, (double) MinHeapFreeRatio);
1550 1550 expand(expand_bytes);
1551 1551
1552 1552 // No expansion, now see if we want to shrink
1553 1553 } else if (capacity_after_gc > maximum_desired_capacity) {
1554 1554 // Capacity too large, compute shrinking size
1555 1555 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1556 1556 ergo_verbose4(ErgoHeapSizing,
1557 1557 "attempt heap shrinking",
1558 1558 ergo_format_reason("capacity higher than "
1559 1559 "max desired capacity after Full GC")
1560 1560 ergo_format_byte("capacity")
1561 1561 ergo_format_byte("occupancy")
1562 1562 ergo_format_byte_perc("max desired capacity"),
1563 1563 capacity_after_gc, used_after_gc,
1564 1564 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1565 1565 shrink(shrink_bytes);
1566 1566 }
1567 1567 }
1568 1568
1569 1569
1570 1570 HeapWord*
1571 1571 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1572 1572 bool* succeeded) {
1573 1573 assert_at_safepoint(true /* should_be_vm_thread */);
1574 1574
1575 1575 *succeeded = true;
1576 1576 // Let's attempt the allocation first.
1577 1577 HeapWord* result =
1578 1578 attempt_allocation_at_safepoint(word_size,
1579 1579 false /* expect_null_mutator_alloc_region */);
1580 1580 if (result != NULL) {
1581 1581 assert(*succeeded, "sanity");
1582 1582 return result;
1583 1583 }
1584 1584
1585 1585 // In a G1 heap, we're supposed to keep allocation from failing by
1586 1586 // incremental pauses. Therefore, at least for now, we'll favor
1587 1587 // expansion over collection. (This might change in the future if we can
1588 1588 // do something smarter than full collection to satisfy a failed alloc.)
1589 1589 result = expand_and_allocate(word_size);
1590 1590 if (result != NULL) {
1591 1591 assert(*succeeded, "sanity");
1592 1592 return result;
1593 1593 }
1594 1594
1595 1595 // Expansion didn't work, we'll try to do a Full GC.
1596 1596 bool gc_succeeded = do_collection(false, /* explicit_gc */
1597 1597 false, /* clear_all_soft_refs */
1598 1598 word_size);
1599 1599 if (!gc_succeeded) {
1600 1600 *succeeded = false;
1601 1601 return NULL;
1602 1602 }
1603 1603
1604 1604 // Retry the allocation
1605 1605 result = attempt_allocation_at_safepoint(word_size,
1606 1606 true /* expect_null_mutator_alloc_region */);
1607 1607 if (result != NULL) {
1608 1608 assert(*succeeded, "sanity");
1609 1609 return result;
1610 1610 }
1611 1611
1612 1612 // Then, try a Full GC that will collect all soft references.
1613 1613 gc_succeeded = do_collection(false, /* explicit_gc */
1614 1614 true, /* clear_all_soft_refs */
1615 1615 word_size);
1616 1616 if (!gc_succeeded) {
1617 1617 *succeeded = false;
1618 1618 return NULL;
1619 1619 }
1620 1620
1621 1621 // Retry the allocation once more
1622 1622 result = attempt_allocation_at_safepoint(word_size,
1623 1623 true /* expect_null_mutator_alloc_region */);
1624 1624 if (result != NULL) {
1625 1625 assert(*succeeded, "sanity");
1626 1626 return result;
1627 1627 }
1628 1628
1629 1629 assert(!collector_policy()->should_clear_all_soft_refs(),
1630 1630 "Flag should have been handled and cleared prior to this point");
1631 1631
1632 1632 // What else? We might try synchronous finalization later. If the total
1633 1633 // space available is large enough for the allocation, then a more
1634 1634 // complete compaction phase than we've tried so far might be
1635 1635 // appropriate.
1636 1636 assert(*succeeded, "sanity");
1637 1637 return NULL;
1638 1638 }
1639 1639
1640 1640 // Attempting to expand the heap sufficiently
1641 1641 // to support an allocation of the given "word_size". If
1642 1642 // successful, perform the allocation and return the address of the
1643 1643 // allocated block, or else "NULL".
1644 1644
1645 1645 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1646 1646 assert_at_safepoint(true /* should_be_vm_thread */);
1647 1647
1648 1648 verify_region_sets_optional();
1649 1649
1650 1650 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1651 1651 ergo_verbose1(ErgoHeapSizing,
1652 1652 "attempt heap expansion",
1653 1653 ergo_format_reason("allocation request failed")
1654 1654 ergo_format_byte("allocation request"),
1655 1655 word_size * HeapWordSize);
1656 1656 if (expand(expand_bytes)) {
1657 1657 _hrs.verify_optional();
1658 1658 verify_region_sets_optional();
1659 1659 return attempt_allocation_at_safepoint(word_size,
1660 1660 false /* expect_null_mutator_alloc_region */);
1661 1661 }
1662 1662 return NULL;
1663 1663 }
1664 1664
1665 1665 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1666 1666 HeapWord* new_end) {
1667 1667 assert(old_end != new_end, "don't call this otherwise");
1668 1668 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1669 1669
1670 1670 // Update the committed mem region.
1671 1671 _g1_committed.set_end(new_end);
1672 1672 // Tell the card table about the update.
1673 1673 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1674 1674 // Tell the BOT about the update.
1675 1675 _bot_shared->resize(_g1_committed.word_size());
1676 1676 }
1677 1677
1678 1678 bool G1CollectedHeap::expand(size_t expand_bytes) {
1679 1679 size_t old_mem_size = _g1_storage.committed_size();
1680 1680 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1681 1681 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1682 1682 HeapRegion::GrainBytes);
1683 1683 ergo_verbose2(ErgoHeapSizing,
1684 1684 "expand the heap",
1685 1685 ergo_format_byte("requested expansion amount")
1686 1686 ergo_format_byte("attempted expansion amount"),
1687 1687 expand_bytes, aligned_expand_bytes);
1688 1688
1689 1689 // First commit the memory.
1690 1690 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1691 1691 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1692 1692 if (successful) {
1693 1693 // Then propagate this update to the necessary data structures.
1694 1694 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1695 1695 update_committed_space(old_end, new_end);
1696 1696
1697 1697 FreeRegionList expansion_list("Local Expansion List");
1698 1698 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1699 1699 assert(mr.start() == old_end, "post-condition");
1700 1700 // mr might be a smaller region than what was requested if
1701 1701 // expand_by() was unable to allocate the HeapRegion instances
1702 1702 assert(mr.end() <= new_end, "post-condition");
1703 1703
1704 1704 size_t actual_expand_bytes = mr.byte_size();
1705 1705 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1706 1706 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1707 1707 "post-condition");
1708 1708 if (actual_expand_bytes < aligned_expand_bytes) {
1709 1709 // We could not expand _hrs to the desired size. In this case we
1710 1710 // need to shrink the committed space accordingly.
1711 1711 assert(mr.end() < new_end, "invariant");
1712 1712
1713 1713 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1714 1714 // First uncommit the memory.
1715 1715 _g1_storage.shrink_by(diff_bytes);
1716 1716 // Then propagate this update to the necessary data structures.
1717 1717 update_committed_space(new_end, mr.end());
1718 1718 }
1719 1719 _free_list.add_as_tail(&expansion_list);
1720 1720
1721 1721 if (_hr_printer.is_active()) {
1722 1722 HeapWord* curr = mr.start();
1723 1723 while (curr < mr.end()) {
1724 1724 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1725 1725 _hr_printer.commit(curr, curr_end);
1726 1726 curr = curr_end;
1727 1727 }
1728 1728 assert(curr == mr.end(), "post-condition");
1729 1729 }
1730 1730 g1_policy()->record_new_heap_size(n_regions());
1731 1731 } else {
1732 1732 ergo_verbose0(ErgoHeapSizing,
1733 1733 "did not expand the heap",
1734 1734 ergo_format_reason("heap expansion operation failed"));
1735 1735 // The expansion of the virtual storage space was unsuccessful.
1736 1736 // Let's see if it was because we ran out of swap.
1737 1737 if (G1ExitOnExpansionFailure &&
1738 1738 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1739 1739 // We had head room...
1740 1740 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1741 1741 }
1742 1742 }
1743 1743 return successful;
1744 1744 }
1745 1745
1746 1746 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1747 1747 size_t old_mem_size = _g1_storage.committed_size();
1748 1748 size_t aligned_shrink_bytes =
1749 1749 ReservedSpace::page_align_size_down(shrink_bytes);
1750 1750 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1751 1751 HeapRegion::GrainBytes);
1752 1752 size_t num_regions_deleted = 0;
1753 1753 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1754 1754 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1755 1755 assert(mr.end() == old_end, "post-condition");
1756 1756
1757 1757 ergo_verbose3(ErgoHeapSizing,
1758 1758 "shrink the heap",
1759 1759 ergo_format_byte("requested shrinking amount")
1760 1760 ergo_format_byte("aligned shrinking amount")
1761 1761 ergo_format_byte("attempted shrinking amount"),
1762 1762 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1763 1763 if (mr.byte_size() > 0) {
1764 1764 if (_hr_printer.is_active()) {
1765 1765 HeapWord* curr = mr.end();
1766 1766 while (curr > mr.start()) {
1767 1767 HeapWord* curr_end = curr;
1768 1768 curr -= HeapRegion::GrainWords;
1769 1769 _hr_printer.uncommit(curr, curr_end);
1770 1770 }
1771 1771 assert(curr == mr.start(), "post-condition");
1772 1772 }
1773 1773
1774 1774 _g1_storage.shrink_by(mr.byte_size());
1775 1775 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1776 1776 assert(mr.start() == new_end, "post-condition");
1777 1777
1778 1778 _expansion_regions += num_regions_deleted;
1779 1779 update_committed_space(old_end, new_end);
1780 1780 HeapRegionRemSet::shrink_heap(n_regions());
1781 1781 g1_policy()->record_new_heap_size(n_regions());
1782 1782 } else {
1783 1783 ergo_verbose0(ErgoHeapSizing,
1784 1784 "did not shrink the heap",
1785 1785 ergo_format_reason("heap shrinking operation failed"));
1786 1786 }
1787 1787 }
1788 1788
1789 1789 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1790 1790 verify_region_sets_optional();
1791 1791
1792 1792 // We should only reach here at the end of a Full GC which means we
1793 1793 // should not not be holding to any GC alloc regions. The method
1794 1794 // below will make sure of that and do any remaining clean up.
1795 1795 abandon_gc_alloc_regions();
1796 1796
1797 1797 // Instead of tearing down / rebuilding the free lists here, we
1798 1798 // could instead use the remove_all_pending() method on free_list to
1799 1799 // remove only the ones that we need to remove.
1800 1800 tear_down_region_sets(true /* free_list_only */);
1801 1801 shrink_helper(shrink_bytes);
1802 1802 rebuild_region_sets(true /* free_list_only */);
1803 1803
1804 1804 _hrs.verify_optional();
1805 1805 verify_region_sets_optional();
1806 1806 }
1807 1807
1808 1808 // Public methods.
1809 1809
1810 1810 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1811 1811 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1812 1812 #endif // _MSC_VER
1813 1813
1814 1814
1815 1815 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1816 1816 SharedHeap(policy_),
1817 1817 _g1_policy(policy_),
1818 1818 _dirty_card_queue_set(false),
1819 1819 _into_cset_dirty_card_queue_set(false),
1820 1820 _is_alive_closure_cm(this),
1821 1821 _is_alive_closure_stw(this),
1822 1822 _ref_processor_cm(NULL),
1823 1823 _ref_processor_stw(NULL),
1824 1824 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1825 1825 _bot_shared(NULL),
1826 1826 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1827 1827 _evac_failure_scan_stack(NULL) ,
1828 1828 _mark_in_progress(false),
1829 1829 _cg1r(NULL), _summary_bytes_used(0),
1830 1830 _g1mm(NULL),
1831 1831 _refine_cte_cl(NULL),
1832 1832 _full_collection(false),
1833 1833 _free_list("Master Free List"),
1834 1834 _secondary_free_list("Secondary Free List"),
↓ open down ↓ |
1834 lines elided |
↑ open up ↑ |
1835 1835 _old_set("Old Set"),
1836 1836 _humongous_set("Master Humongous Set"),
1837 1837 _free_regions_coming(false),
1838 1838 _young_list(new YoungList(this)),
1839 1839 _gc_time_stamp(0),
1840 1840 _retained_old_gc_alloc_region(NULL),
1841 1841 _surviving_young_words(NULL),
1842 1842 _full_collections_completed(0),
1843 1843 _in_cset_fast_test(NULL),
1844 1844 _in_cset_fast_test_base(NULL),
1845 - _dirty_cards_region_list(NULL) {
1845 + _dirty_cards_region_list(NULL),
1846 + _worker_cset_start_region(NULL),
1847 + _worker_cset_start_region_stamp(NULL) {
1846 1848 _g1h = this; // To catch bugs.
1847 1849 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1848 1850 vm_exit_during_initialization("Failed necessary allocation.");
1849 1851 }
1850 1852
1851 1853 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1852 1854
1853 1855 int n_queues = MAX2((int)ParallelGCThreads, 1);
1854 1856 _task_queues = new RefToScanQueueSet(n_queues);
1855 1857
1856 1858 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1857 1859 assert(n_rem_sets > 0, "Invariant.");
1858 1860
1859 1861 HeapRegionRemSetIterator** iter_arr =
1860 1862 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues);
1861 1863 for (int i = 0; i < n_queues; i++) {
1862 1864 iter_arr[i] = new HeapRegionRemSetIterator();
1863 1865 }
1864 1866 _rem_set_iterator = iter_arr;
1865 1867
1868 + _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues);
1869 + _worker_cset_start_region_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues);
1870 +
1866 1871 for (int i = 0; i < n_queues; i++) {
1867 1872 RefToScanQueue* q = new RefToScanQueue();
1868 1873 q->initialize();
1869 1874 _task_queues->register_queue(i, q);
1875 +
1876 + _worker_cset_start_region[i] = NULL;
1877 + _worker_cset_start_region_stamp[i] = 0;
1870 1878 }
1871 1879
1872 1880 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1873 1881 }
1874 1882
1875 1883 jint G1CollectedHeap::initialize() {
1876 1884 CollectedHeap::pre_initialize();
1877 1885 os::enable_vtime();
1878 1886
1879 1887 // Necessary to satisfy locking discipline assertions.
1880 1888
1881 1889 MutexLocker x(Heap_lock);
1882 1890
1883 1891 // We have to initialize the printer before committing the heap, as
1884 1892 // it will be used then.
1885 1893 _hr_printer.set_active(G1PrintHeapRegions);
1886 1894
1887 1895 // While there are no constraints in the GC code that HeapWordSize
1888 1896 // be any particular value, there are multiple other areas in the
1889 1897 // system which believe this to be true (e.g. oop->object_size in some
1890 1898 // cases incorrectly returns the size in wordSize units rather than
1891 1899 // HeapWordSize).
1892 1900 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1893 1901
1894 1902 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1895 1903 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1896 1904
1897 1905 // Ensure that the sizes are properly aligned.
1898 1906 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1899 1907 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1900 1908
1901 1909 _cg1r = new ConcurrentG1Refine();
1902 1910
1903 1911 // Reserve the maximum.
1904 1912 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation();
1905 1913 // Includes the perm-gen.
1906 1914
1907 1915 // When compressed oops are enabled, the preferred heap base
1908 1916 // is calculated by subtracting the requested size from the
1909 1917 // 32Gb boundary and using the result as the base address for
1910 1918 // heap reservation. If the requested size is not aligned to
1911 1919 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1912 1920 // into the ReservedHeapSpace constructor) then the actual
1913 1921 // base of the reserved heap may end up differing from the
1914 1922 // address that was requested (i.e. the preferred heap base).
1915 1923 // If this happens then we could end up using a non-optimal
1916 1924 // compressed oops mode.
1917 1925
1918 1926 // Since max_byte_size is aligned to the size of a heap region (checked
1919 1927 // above), we also need to align the perm gen size as it might not be.
1920 1928 const size_t total_reserved = max_byte_size +
1921 1929 align_size_up(pgs->max_size(), HeapRegion::GrainBytes);
1922 1930 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm");
1923 1931
1924 1932 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop);
1925 1933
1926 1934 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes,
1927 1935 UseLargePages, addr);
1928 1936
1929 1937 if (UseCompressedOops) {
1930 1938 if (addr != NULL && !heap_rs.is_reserved()) {
1931 1939 // Failed to reserve at specified address - the requested memory
1932 1940 // region is taken already, for example, by 'java' launcher.
1933 1941 // Try again to reserver heap higher.
1934 1942 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop);
1935 1943
1936 1944 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes,
1937 1945 UseLargePages, addr);
1938 1946
1939 1947 if (addr != NULL && !heap_rs0.is_reserved()) {
1940 1948 // Failed to reserve at specified address again - give up.
1941 1949 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop);
1942 1950 assert(addr == NULL, "");
1943 1951
1944 1952 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes,
1945 1953 UseLargePages, addr);
1946 1954 heap_rs = heap_rs1;
1947 1955 } else {
1948 1956 heap_rs = heap_rs0;
1949 1957 }
1950 1958 }
1951 1959 }
1952 1960
1953 1961 if (!heap_rs.is_reserved()) {
1954 1962 vm_exit_during_initialization("Could not reserve enough space for object heap");
1955 1963 return JNI_ENOMEM;
1956 1964 }
1957 1965
1958 1966 // It is important to do this in a way such that concurrent readers can't
1959 1967 // temporarily think somethings in the heap. (I've actually seen this
1960 1968 // happen in asserts: DLD.)
1961 1969 _reserved.set_word_size(0);
1962 1970 _reserved.set_start((HeapWord*)heap_rs.base());
1963 1971 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
1964 1972
1965 1973 _expansion_regions = max_byte_size/HeapRegion::GrainBytes;
1966 1974
1967 1975 // Create the gen rem set (and barrier set) for the entire reserved region.
1968 1976 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
1969 1977 set_barrier_set(rem_set()->bs());
1970 1978 if (barrier_set()->is_a(BarrierSet::ModRef)) {
1971 1979 _mr_bs = (ModRefBarrierSet*)_barrier_set;
1972 1980 } else {
1973 1981 vm_exit_during_initialization("G1 requires a mod ref bs.");
1974 1982 return JNI_ENOMEM;
1975 1983 }
1976 1984
1977 1985 // Also create a G1 rem set.
1978 1986 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
1979 1987 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
1980 1988 } else {
1981 1989 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
1982 1990 return JNI_ENOMEM;
1983 1991 }
1984 1992
1985 1993 // Carve out the G1 part of the heap.
1986 1994
1987 1995 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
1988 1996 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
1989 1997 g1_rs.size()/HeapWordSize);
1990 1998 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size);
1991 1999
1992 2000 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set());
1993 2001
1994 2002 _g1_storage.initialize(g1_rs, 0);
1995 2003 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
1996 2004 _hrs.initialize((HeapWord*) _g1_reserved.start(),
1997 2005 (HeapWord*) _g1_reserved.end(),
1998 2006 _expansion_regions);
1999 2007
2000 2008 // 6843694 - ensure that the maximum region index can fit
2001 2009 // in the remembered set structures.
2002 2010 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2003 2011 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2004 2012
2005 2013 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2006 2014 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2007 2015 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2008 2016 "too many cards per region");
2009 2017
2010 2018 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2011 2019
2012 2020 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2013 2021 heap_word_size(init_byte_size));
2014 2022
2015 2023 _g1h = this;
2016 2024
2017 2025 _in_cset_fast_test_length = max_regions();
2018 2026 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length);
2019 2027
2020 2028 // We're biasing _in_cset_fast_test to avoid subtracting the
2021 2029 // beginning of the heap every time we want to index; basically
2022 2030 // it's the same with what we do with the card table.
2023 2031 _in_cset_fast_test = _in_cset_fast_test_base -
2024 2032 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2025 2033
2026 2034 // Clear the _cset_fast_test bitmap in anticipation of adding
2027 2035 // regions to the incremental collection set for the first
2028 2036 // evacuation pause.
2029 2037 clear_cset_fast_test();
2030 2038
2031 2039 // Create the ConcurrentMark data structure and thread.
2032 2040 // (Must do this late, so that "max_regions" is defined.)
2033 2041 _cm = new ConcurrentMark(heap_rs, (int) max_regions());
2034 2042 _cmThread = _cm->cmThread();
2035 2043
2036 2044 // Initialize the from_card cache structure of HeapRegionRemSet.
2037 2045 HeapRegionRemSet::init_heap(max_regions());
2038 2046
2039 2047 // Now expand into the initial heap size.
2040 2048 if (!expand(init_byte_size)) {
2041 2049 vm_exit_during_initialization("Failed to allocate initial heap.");
2042 2050 return JNI_ENOMEM;
2043 2051 }
2044 2052
2045 2053 // Perform any initialization actions delegated to the policy.
2046 2054 g1_policy()->init();
2047 2055
2048 2056 _refine_cte_cl =
2049 2057 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2050 2058 g1_rem_set(),
2051 2059 concurrent_g1_refine());
2052 2060 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2053 2061
2054 2062 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2055 2063 SATB_Q_FL_lock,
2056 2064 G1SATBProcessCompletedThreshold,
2057 2065 Shared_SATB_Q_lock);
2058 2066
2059 2067 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2060 2068 DirtyCardQ_FL_lock,
2061 2069 concurrent_g1_refine()->yellow_zone(),
2062 2070 concurrent_g1_refine()->red_zone(),
2063 2071 Shared_DirtyCardQ_lock);
2064 2072
2065 2073 if (G1DeferredRSUpdate) {
2066 2074 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2067 2075 DirtyCardQ_FL_lock,
2068 2076 -1, // never trigger processing
2069 2077 -1, // no limit on length
2070 2078 Shared_DirtyCardQ_lock,
2071 2079 &JavaThread::dirty_card_queue_set());
2072 2080 }
2073 2081
2074 2082 // Initialize the card queue set used to hold cards containing
2075 2083 // references into the collection set.
2076 2084 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2077 2085 DirtyCardQ_FL_lock,
2078 2086 -1, // never trigger processing
2079 2087 -1, // no limit on length
2080 2088 Shared_DirtyCardQ_lock,
2081 2089 &JavaThread::dirty_card_queue_set());
2082 2090
2083 2091 // In case we're keeping closure specialization stats, initialize those
2084 2092 // counts and that mechanism.
2085 2093 SpecializationStats::clear();
2086 2094
2087 2095 // Do later initialization work for concurrent refinement.
2088 2096 _cg1r->init();
2089 2097
2090 2098 // Here we allocate the dummy full region that is required by the
2091 2099 // G1AllocRegion class. If we don't pass an address in the reserved
2092 2100 // space here, lots of asserts fire.
2093 2101
2094 2102 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2095 2103 _g1_reserved.start());
2096 2104 // We'll re-use the same region whether the alloc region will
2097 2105 // require BOT updates or not and, if it doesn't, then a non-young
2098 2106 // region will complain that it cannot support allocations without
2099 2107 // BOT updates. So we'll tag the dummy region as young to avoid that.
2100 2108 dummy_region->set_young();
2101 2109 // Make sure it's full.
2102 2110 dummy_region->set_top(dummy_region->end());
2103 2111 G1AllocRegion::setup(this, dummy_region);
2104 2112
2105 2113 init_mutator_alloc_region();
2106 2114
2107 2115 // Do create of the monitoring and management support so that
2108 2116 // values in the heap have been properly initialized.
2109 2117 _g1mm = new G1MonitoringSupport(this);
2110 2118
2111 2119 return JNI_OK;
2112 2120 }
2113 2121
2114 2122 void G1CollectedHeap::ref_processing_init() {
2115 2123 // Reference processing in G1 currently works as follows:
2116 2124 //
2117 2125 // * There are two reference processor instances. One is
2118 2126 // used to record and process discovered references
2119 2127 // during concurrent marking; the other is used to
2120 2128 // record and process references during STW pauses
2121 2129 // (both full and incremental).
2122 2130 // * Both ref processors need to 'span' the entire heap as
2123 2131 // the regions in the collection set may be dotted around.
2124 2132 //
2125 2133 // * For the concurrent marking ref processor:
2126 2134 // * Reference discovery is enabled at initial marking.
2127 2135 // * Reference discovery is disabled and the discovered
2128 2136 // references processed etc during remarking.
2129 2137 // * Reference discovery is MT (see below).
2130 2138 // * Reference discovery requires a barrier (see below).
2131 2139 // * Reference processing may or may not be MT
2132 2140 // (depending on the value of ParallelRefProcEnabled
2133 2141 // and ParallelGCThreads).
2134 2142 // * A full GC disables reference discovery by the CM
2135 2143 // ref processor and abandons any entries on it's
2136 2144 // discovered lists.
2137 2145 //
2138 2146 // * For the STW processor:
2139 2147 // * Non MT discovery is enabled at the start of a full GC.
2140 2148 // * Processing and enqueueing during a full GC is non-MT.
2141 2149 // * During a full GC, references are processed after marking.
2142 2150 //
2143 2151 // * Discovery (may or may not be MT) is enabled at the start
2144 2152 // of an incremental evacuation pause.
2145 2153 // * References are processed near the end of a STW evacuation pause.
2146 2154 // * For both types of GC:
2147 2155 // * Discovery is atomic - i.e. not concurrent.
2148 2156 // * Reference discovery will not need a barrier.
2149 2157
2150 2158 SharedHeap::ref_processing_init();
2151 2159 MemRegion mr = reserved_region();
2152 2160
2153 2161 // Concurrent Mark ref processor
2154 2162 _ref_processor_cm =
2155 2163 new ReferenceProcessor(mr, // span
2156 2164 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2157 2165 // mt processing
2158 2166 (int) ParallelGCThreads,
2159 2167 // degree of mt processing
2160 2168 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2161 2169 // mt discovery
2162 2170 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2163 2171 // degree of mt discovery
2164 2172 false,
2165 2173 // Reference discovery is not atomic
2166 2174 &_is_alive_closure_cm,
2167 2175 // is alive closure
2168 2176 // (for efficiency/performance)
2169 2177 true);
2170 2178 // Setting next fields of discovered
2171 2179 // lists requires a barrier.
2172 2180
2173 2181 // STW ref processor
2174 2182 _ref_processor_stw =
2175 2183 new ReferenceProcessor(mr, // span
2176 2184 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2177 2185 // mt processing
2178 2186 MAX2((int)ParallelGCThreads, 1),
2179 2187 // degree of mt processing
2180 2188 (ParallelGCThreads > 1),
2181 2189 // mt discovery
2182 2190 MAX2((int)ParallelGCThreads, 1),
2183 2191 // degree of mt discovery
2184 2192 true,
2185 2193 // Reference discovery is atomic
2186 2194 &_is_alive_closure_stw,
2187 2195 // is alive closure
2188 2196 // (for efficiency/performance)
2189 2197 false);
2190 2198 // Setting next fields of discovered
2191 2199 // lists requires a barrier.
2192 2200 }
2193 2201
2194 2202 size_t G1CollectedHeap::capacity() const {
2195 2203 return _g1_committed.byte_size();
2196 2204 }
2197 2205
2198 2206 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2199 2207 DirtyCardQueue* into_cset_dcq,
2200 2208 bool concurrent,
2201 2209 int worker_i) {
2202 2210 // Clean cards in the hot card cache
2203 2211 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2204 2212
2205 2213 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2206 2214 int n_completed_buffers = 0;
2207 2215 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2208 2216 n_completed_buffers++;
2209 2217 }
2210 2218 g1_policy()->record_update_rs_processed_buffers(worker_i,
2211 2219 (double) n_completed_buffers);
2212 2220 dcqs.clear_n_completed_buffers();
2213 2221 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2214 2222 }
2215 2223
2216 2224
2217 2225 // Computes the sum of the storage used by the various regions.
2218 2226
2219 2227 size_t G1CollectedHeap::used() const {
2220 2228 assert(Heap_lock->owner() != NULL,
2221 2229 "Should be owned on this thread's behalf.");
2222 2230 size_t result = _summary_bytes_used;
2223 2231 // Read only once in case it is set to NULL concurrently
2224 2232 HeapRegion* hr = _mutator_alloc_region.get();
2225 2233 if (hr != NULL)
2226 2234 result += hr->used();
2227 2235 return result;
2228 2236 }
2229 2237
2230 2238 size_t G1CollectedHeap::used_unlocked() const {
2231 2239 size_t result = _summary_bytes_used;
2232 2240 return result;
2233 2241 }
2234 2242
2235 2243 class SumUsedClosure: public HeapRegionClosure {
2236 2244 size_t _used;
2237 2245 public:
2238 2246 SumUsedClosure() : _used(0) {}
2239 2247 bool doHeapRegion(HeapRegion* r) {
2240 2248 if (!r->continuesHumongous()) {
2241 2249 _used += r->used();
2242 2250 }
2243 2251 return false;
2244 2252 }
2245 2253 size_t result() { return _used; }
2246 2254 };
2247 2255
2248 2256 size_t G1CollectedHeap::recalculate_used() const {
2249 2257 SumUsedClosure blk;
2250 2258 heap_region_iterate(&blk);
2251 2259 return blk.result();
2252 2260 }
2253 2261
2254 2262 size_t G1CollectedHeap::unsafe_max_alloc() {
2255 2263 if (free_regions() > 0) return HeapRegion::GrainBytes;
2256 2264 // otherwise, is there space in the current allocation region?
2257 2265
2258 2266 // We need to store the current allocation region in a local variable
2259 2267 // here. The problem is that this method doesn't take any locks and
2260 2268 // there may be other threads which overwrite the current allocation
2261 2269 // region field. attempt_allocation(), for example, sets it to NULL
2262 2270 // and this can happen *after* the NULL check here but before the call
2263 2271 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2264 2272 // to be a problem in the optimized build, since the two loads of the
2265 2273 // current allocation region field are optimized away.
2266 2274 HeapRegion* hr = _mutator_alloc_region.get();
2267 2275 if (hr == NULL) {
2268 2276 return 0;
2269 2277 }
2270 2278 return hr->free();
2271 2279 }
2272 2280
2273 2281 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2274 2282 return
2275 2283 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) ||
2276 2284 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent));
2277 2285 }
2278 2286
2279 2287 #ifndef PRODUCT
2280 2288 void G1CollectedHeap::allocate_dummy_regions() {
2281 2289 // Let's fill up most of the region
2282 2290 size_t word_size = HeapRegion::GrainWords - 1024;
2283 2291 // And as a result the region we'll allocate will be humongous.
2284 2292 guarantee(isHumongous(word_size), "sanity");
2285 2293
2286 2294 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2287 2295 // Let's use the existing mechanism for the allocation
2288 2296 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2289 2297 if (dummy_obj != NULL) {
2290 2298 MemRegion mr(dummy_obj, word_size);
2291 2299 CollectedHeap::fill_with_object(mr);
2292 2300 } else {
2293 2301 // If we can't allocate once, we probably cannot allocate
2294 2302 // again. Let's get out of the loop.
2295 2303 break;
2296 2304 }
2297 2305 }
2298 2306 }
2299 2307 #endif // !PRODUCT
2300 2308
2301 2309 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) {
2302 2310 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2303 2311
2304 2312 // We assume that if concurrent == true, then the caller is a
2305 2313 // concurrent thread that was joined the Suspendible Thread
2306 2314 // Set. If there's ever a cheap way to check this, we should add an
2307 2315 // assert here.
2308 2316
2309 2317 // We have already incremented _total_full_collections at the start
2310 2318 // of the GC, so total_full_collections() represents how many full
2311 2319 // collections have been started.
2312 2320 unsigned int full_collections_started = total_full_collections();
2313 2321
2314 2322 // Given that this method is called at the end of a Full GC or of a
2315 2323 // concurrent cycle, and those can be nested (i.e., a Full GC can
2316 2324 // interrupt a concurrent cycle), the number of full collections
2317 2325 // completed should be either one (in the case where there was no
2318 2326 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2319 2327 // behind the number of full collections started.
2320 2328
2321 2329 // This is the case for the inner caller, i.e. a Full GC.
2322 2330 assert(concurrent ||
2323 2331 (full_collections_started == _full_collections_completed + 1) ||
2324 2332 (full_collections_started == _full_collections_completed + 2),
2325 2333 err_msg("for inner caller (Full GC): full_collections_started = %u "
2326 2334 "is inconsistent with _full_collections_completed = %u",
2327 2335 full_collections_started, _full_collections_completed));
2328 2336
2329 2337 // This is the case for the outer caller, i.e. the concurrent cycle.
2330 2338 assert(!concurrent ||
2331 2339 (full_collections_started == _full_collections_completed + 1),
2332 2340 err_msg("for outer caller (concurrent cycle): "
2333 2341 "full_collections_started = %u "
2334 2342 "is inconsistent with _full_collections_completed = %u",
2335 2343 full_collections_started, _full_collections_completed));
2336 2344
2337 2345 _full_collections_completed += 1;
2338 2346
2339 2347 // We need to clear the "in_progress" flag in the CM thread before
2340 2348 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2341 2349 // is set) so that if a waiter requests another System.gc() it doesn't
2342 2350 // incorrectly see that a marking cyle is still in progress.
2343 2351 if (concurrent) {
2344 2352 _cmThread->clear_in_progress();
2345 2353 }
2346 2354
2347 2355 // This notify_all() will ensure that a thread that called
2348 2356 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2349 2357 // and it's waiting for a full GC to finish will be woken up. It is
2350 2358 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2351 2359 FullGCCount_lock->notify_all();
2352 2360 }
2353 2361
2354 2362 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) {
2355 2363 assert_at_safepoint(true /* should_be_vm_thread */);
2356 2364 GCCauseSetter gcs(this, cause);
2357 2365 switch (cause) {
2358 2366 case GCCause::_heap_inspection:
2359 2367 case GCCause::_heap_dump: {
2360 2368 HandleMark hm;
2361 2369 do_full_collection(false); // don't clear all soft refs
2362 2370 break;
2363 2371 }
2364 2372 default: // XXX FIX ME
2365 2373 ShouldNotReachHere(); // Unexpected use of this function
2366 2374 }
2367 2375 }
2368 2376
2369 2377 void G1CollectedHeap::collect(GCCause::Cause cause) {
2370 2378 // The caller doesn't have the Heap_lock
2371 2379 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock");
2372 2380
2373 2381 unsigned int gc_count_before;
2374 2382 unsigned int full_gc_count_before;
2375 2383 {
2376 2384 MutexLocker ml(Heap_lock);
2377 2385
2378 2386 // Read the GC count while holding the Heap_lock
2379 2387 gc_count_before = SharedHeap::heap()->total_collections();
2380 2388 full_gc_count_before = SharedHeap::heap()->total_full_collections();
2381 2389 }
2382 2390
2383 2391 if (should_do_concurrent_full_gc(cause)) {
2384 2392 // Schedule an initial-mark evacuation pause that will start a
2385 2393 // concurrent cycle. We're setting word_size to 0 which means that
2386 2394 // we are not requesting a post-GC allocation.
2387 2395 VM_G1IncCollectionPause op(gc_count_before,
2388 2396 0, /* word_size */
2389 2397 true, /* should_initiate_conc_mark */
2390 2398 g1_policy()->max_pause_time_ms(),
2391 2399 cause);
2392 2400 VMThread::execute(&op);
2393 2401 } else {
2394 2402 if (cause == GCCause::_gc_locker
2395 2403 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2396 2404
2397 2405 // Schedule a standard evacuation pause. We're setting word_size
2398 2406 // to 0 which means that we are not requesting a post-GC allocation.
2399 2407 VM_G1IncCollectionPause op(gc_count_before,
2400 2408 0, /* word_size */
2401 2409 false, /* should_initiate_conc_mark */
2402 2410 g1_policy()->max_pause_time_ms(),
2403 2411 cause);
2404 2412 VMThread::execute(&op);
2405 2413 } else {
2406 2414 // Schedule a Full GC.
2407 2415 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
2408 2416 VMThread::execute(&op);
2409 2417 }
2410 2418 }
2411 2419 }
2412 2420
2413 2421 bool G1CollectedHeap::is_in(const void* p) const {
2414 2422 HeapRegion* hr = _hrs.addr_to_region((HeapWord*) p);
2415 2423 if (hr != NULL) {
2416 2424 return hr->is_in(p);
2417 2425 } else {
2418 2426 return _perm_gen->as_gen()->is_in(p);
2419 2427 }
2420 2428 }
2421 2429
2422 2430 // Iteration functions.
2423 2431
2424 2432 // Iterates an OopClosure over all ref-containing fields of objects
2425 2433 // within a HeapRegion.
2426 2434
2427 2435 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2428 2436 MemRegion _mr;
2429 2437 OopClosure* _cl;
2430 2438 public:
2431 2439 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl)
2432 2440 : _mr(mr), _cl(cl) {}
2433 2441 bool doHeapRegion(HeapRegion* r) {
2434 2442 if (! r->continuesHumongous()) {
2435 2443 r->oop_iterate(_cl);
2436 2444 }
2437 2445 return false;
2438 2446 }
2439 2447 };
2440 2448
2441 2449 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) {
2442 2450 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2443 2451 heap_region_iterate(&blk);
2444 2452 if (do_perm) {
2445 2453 perm_gen()->oop_iterate(cl);
2446 2454 }
2447 2455 }
2448 2456
2449 2457 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) {
2450 2458 IterateOopClosureRegionClosure blk(mr, cl);
2451 2459 heap_region_iterate(&blk);
2452 2460 if (do_perm) {
2453 2461 perm_gen()->oop_iterate(cl);
2454 2462 }
2455 2463 }
2456 2464
2457 2465 // Iterates an ObjectClosure over all objects within a HeapRegion.
2458 2466
2459 2467 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2460 2468 ObjectClosure* _cl;
2461 2469 public:
2462 2470 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2463 2471 bool doHeapRegion(HeapRegion* r) {
2464 2472 if (! r->continuesHumongous()) {
2465 2473 r->object_iterate(_cl);
2466 2474 }
2467 2475 return false;
2468 2476 }
2469 2477 };
2470 2478
2471 2479 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) {
2472 2480 IterateObjectClosureRegionClosure blk(cl);
2473 2481 heap_region_iterate(&blk);
2474 2482 if (do_perm) {
2475 2483 perm_gen()->object_iterate(cl);
2476 2484 }
2477 2485 }
2478 2486
2479 2487 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2480 2488 // FIXME: is this right?
2481 2489 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2482 2490 }
2483 2491
2484 2492 // Calls a SpaceClosure on a HeapRegion.
2485 2493
2486 2494 class SpaceClosureRegionClosure: public HeapRegionClosure {
2487 2495 SpaceClosure* _cl;
2488 2496 public:
2489 2497 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2490 2498 bool doHeapRegion(HeapRegion* r) {
2491 2499 _cl->do_space(r);
2492 2500 return false;
2493 2501 }
2494 2502 };
2495 2503
2496 2504 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2497 2505 SpaceClosureRegionClosure blk(cl);
2498 2506 heap_region_iterate(&blk);
2499 2507 }
2500 2508
2501 2509 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2502 2510 _hrs.iterate(cl);
2503 2511 }
2504 2512
2505 2513 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r,
2506 2514 HeapRegionClosure* cl) const {
2507 2515 _hrs.iterate_from(r, cl);
2508 2516 }
2509 2517
2510 2518 void
2511 2519 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2512 2520 int worker,
2513 2521 int no_of_par_workers,
2514 2522 jint claim_value) {
2515 2523 const size_t regions = n_regions();
2516 2524 const size_t max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2517 2525 no_of_par_workers :
2518 2526 1);
2519 2527 assert(UseDynamicNumberOfGCThreads ||
2520 2528 no_of_par_workers == workers()->total_workers(),
2521 2529 "Non dynamic should use fixed number of workers");
2522 2530 // try to spread out the starting points of the workers
2523 2531 const size_t start_index = regions / max_workers * (size_t) worker;
2524 2532
2525 2533 // each worker will actually look at all regions
2526 2534 for (size_t count = 0; count < regions; ++count) {
2527 2535 const size_t index = (start_index + count) % regions;
2528 2536 assert(0 <= index && index < regions, "sanity");
2529 2537 HeapRegion* r = region_at(index);
2530 2538 // we'll ignore "continues humongous" regions (we'll process them
2531 2539 // when we come across their corresponding "start humongous"
2532 2540 // region) and regions already claimed
2533 2541 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2534 2542 continue;
2535 2543 }
2536 2544 // OK, try to claim it
2537 2545 if (r->claimHeapRegion(claim_value)) {
2538 2546 // success!
2539 2547 assert(!r->continuesHumongous(), "sanity");
2540 2548 if (r->startsHumongous()) {
2541 2549 // If the region is "starts humongous" we'll iterate over its
2542 2550 // "continues humongous" first; in fact we'll do them
2543 2551 // first. The order is important. In on case, calling the
2544 2552 // closure on the "starts humongous" region might de-allocate
2545 2553 // and clear all its "continues humongous" regions and, as a
2546 2554 // result, we might end up processing them twice. So, we'll do
2547 2555 // them first (notice: most closures will ignore them anyway) and
2548 2556 // then we'll do the "starts humongous" region.
2549 2557 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) {
2550 2558 HeapRegion* chr = region_at(ch_index);
2551 2559
2552 2560 // if the region has already been claimed or it's not
2553 2561 // "continues humongous" we're done
2554 2562 if (chr->claim_value() == claim_value ||
2555 2563 !chr->continuesHumongous()) {
2556 2564 break;
2557 2565 }
2558 2566
2559 2567 // Noone should have claimed it directly. We can given
2560 2568 // that we claimed its "starts humongous" region.
2561 2569 assert(chr->claim_value() != claim_value, "sanity");
2562 2570 assert(chr->humongous_start_region() == r, "sanity");
2563 2571
2564 2572 if (chr->claimHeapRegion(claim_value)) {
2565 2573 // we should always be able to claim it; noone else should
2566 2574 // be trying to claim this region
2567 2575
2568 2576 bool res2 = cl->doHeapRegion(chr);
2569 2577 assert(!res2, "Should not abort");
2570 2578
2571 2579 // Right now, this holds (i.e., no closure that actually
2572 2580 // does something with "continues humongous" regions
2573 2581 // clears them). We might have to weaken it in the future,
2574 2582 // but let's leave these two asserts here for extra safety.
2575 2583 assert(chr->continuesHumongous(), "should still be the case");
2576 2584 assert(chr->humongous_start_region() == r, "sanity");
2577 2585 } else {
2578 2586 guarantee(false, "we should not reach here");
2579 2587 }
2580 2588 }
2581 2589 }
2582 2590
2583 2591 assert(!r->continuesHumongous(), "sanity");
2584 2592 bool res = cl->doHeapRegion(r);
2585 2593 assert(!res, "Should not abort");
2586 2594 }
2587 2595 }
2588 2596 }
2589 2597
2590 2598 class ResetClaimValuesClosure: public HeapRegionClosure {
2591 2599 public:
2592 2600 bool doHeapRegion(HeapRegion* r) {
2593 2601 r->set_claim_value(HeapRegion::InitialClaimValue);
2594 2602 return false;
2595 2603 }
2596 2604 };
2597 2605
2598 2606 void
2599 2607 G1CollectedHeap::reset_heap_region_claim_values() {
2600 2608 ResetClaimValuesClosure blk;
2601 2609 heap_region_iterate(&blk);
2602 2610 }
2603 2611
2604 2612 #ifdef ASSERT
2605 2613 // This checks whether all regions in the heap have the correct claim
2606 2614 // value. I also piggy-backed on this a check to ensure that the
2607 2615 // humongous_start_region() information on "continues humongous"
2608 2616 // regions is correct.
2609 2617
2610 2618 class CheckClaimValuesClosure : public HeapRegionClosure {
2611 2619 private:
2612 2620 jint _claim_value;
2613 2621 size_t _failures;
2614 2622 HeapRegion* _sh_region;
2615 2623 public:
2616 2624 CheckClaimValuesClosure(jint claim_value) :
2617 2625 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2618 2626 bool doHeapRegion(HeapRegion* r) {
2619 2627 if (r->claim_value() != _claim_value) {
2620 2628 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2621 2629 "claim value = %d, should be %d",
2622 2630 HR_FORMAT_PARAMS(r),
2623 2631 r->claim_value(), _claim_value);
2624 2632 ++_failures;
2625 2633 }
2626 2634 if (!r->isHumongous()) {
2627 2635 _sh_region = NULL;
2628 2636 } else if (r->startsHumongous()) {
2629 2637 _sh_region = r;
2630 2638 } else if (r->continuesHumongous()) {
2631 2639 if (r->humongous_start_region() != _sh_region) {
2632 2640 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2633 2641 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2634 2642 HR_FORMAT_PARAMS(r),
2635 2643 r->humongous_start_region(),
2636 2644 _sh_region);
2637 2645 ++_failures;
2638 2646 }
2639 2647 }
2640 2648 return false;
2641 2649 }
2642 2650 size_t failures() {
2643 2651 return _failures;
2644 2652 }
2645 2653 };
2646 2654
2647 2655 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2648 2656 CheckClaimValuesClosure cl(claim_value);
2649 2657 heap_region_iterate(&cl);
2650 2658 return cl.failures() == 0;
2651 2659 }
2652 2660
2653 2661 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2654 2662 jint _claim_value;
2655 2663 size_t _failures;
2656 2664
2657 2665 public:
2658 2666 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2659 2667 _claim_value(claim_value),
2660 2668 _failures(0) { }
2661 2669
2662 2670 size_t failures() {
2663 2671 return _failures;
2664 2672 }
2665 2673
2666 2674 bool doHeapRegion(HeapRegion* hr) {
2667 2675 assert(hr->in_collection_set(), "how?");
2668 2676 assert(!hr->isHumongous(), "H-region in CSet");
2669 2677 if (hr->claim_value() != _claim_value) {
2670 2678 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2671 2679 "claim value = %d, should be %d",
2672 2680 HR_FORMAT_PARAMS(hr),
2673 2681 hr->claim_value(), _claim_value);
2674 2682 _failures += 1;
2675 2683 }
2676 2684 return false;
↓ open down ↓ |
797 lines elided |
↑ open up ↑ |
2677 2685 }
2678 2686 };
2679 2687
2680 2688 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2681 2689 CheckClaimValuesInCSetHRClosure cl(claim_value);
2682 2690 collection_set_iterate(&cl);
2683 2691 return cl.failures() == 0;
2684 2692 }
2685 2693 #endif // ASSERT
2686 2694
2687 -// We want the parallel threads to start their collection
2688 -// set iteration at different collection set regions to
2689 -// avoid contention.
2690 -// If we have:
2691 -// n collection set regions
2692 -// p threads
2693 -// Then thread t will start at region t * floor (n/p)
2694 -
2695 +// Given the id of a worker, obtain or calculate a suitable
2696 +// starting region for iterating over the current collection set.
2695 2697 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2696 - HeapRegion* result = g1_policy()->collection_set();
2698 + assert(_worker_cset_start_region != NULL, "sanity");
2699 + assert(_worker_cset_start_region_stamp != NULL, "sanity");
2700 + assert(total_collections() > 0, "should have been updated by now");
2701 +
2702 + HeapRegion* result = NULL;
2703 +
2704 + if (_worker_cset_start_region_stamp[worker_i] == total_collections()) {
2705 + // Cached starting region for current worker was set
2706 + // during the current pause - so it's valid.
2707 + // Note: the cached starting heap region may be NULL
2708 + // (when the collection set is empty).
2709 + result = _worker_cset_start_region[worker_i];
2710 + assert(result == NULL || result->in_collection_set(), "sanity");
2711 + return result;
2712 + }
2713 +
2714 + // The cached entry was not valid so let's calculate
2715 + // a suitable starting heap region for this worker.
2716 +
2717 + // We want the parallel threads to start their collection
2718 + // set iteration at different collection set regions to
2719 + // avoid contention.
2720 + // If we have:
2721 + // n collection set regions
2722 + // p threads
2723 + // Then thread t will start at region t * floor (n/p)
2724 +
2725 + result = g1_policy()->collection_set();
2697 2726 if (G1CollectedHeap::use_parallel_gc_threads()) {
2698 2727 size_t cs_size = g1_policy()->cset_region_length();
2699 2728 int n_workers = workers()->total_workers();
2700 2729 size_t cs_spans = cs_size / n_workers;
2701 2730 size_t ind = cs_spans * worker_i;
2702 2731 for (size_t i = 0; i < ind; i++) {
2703 2732 result = result->next_in_collection_set();
2704 2733 }
2705 2734 }
2735 +
2736 + // Note: the calculated starting heap region may be NULL
2737 + // (when the collection set is empty).
2738 + assert(result == NULL || result->in_collection_set(), "sanity");
2739 + assert(_worker_cset_start_region_stamp[worker_i] != total_collections(),
2740 + "should be updated only once per pause");
2741 + _worker_cset_start_region[worker_i] = result;
2742 + _worker_cset_start_region_stamp[worker_i] = total_collections();
2706 2743 return result;
2707 2744 }
2708 2745
2709 2746 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2710 2747 HeapRegion* r = g1_policy()->collection_set();
2711 2748 while (r != NULL) {
2712 2749 HeapRegion* next = r->next_in_collection_set();
2713 2750 if (cl->doHeapRegion(r)) {
2714 2751 cl->incomplete();
2715 2752 return;
2716 2753 }
2717 2754 r = next;
2718 2755 }
2719 2756 }
2720 2757
2721 2758 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2722 2759 HeapRegionClosure *cl) {
2723 2760 if (r == NULL) {
2724 2761 // The CSet is empty so there's nothing to do.
2725 2762 return;
2726 2763 }
2727 2764
2728 2765 assert(r->in_collection_set(),
2729 2766 "Start region must be a member of the collection set.");
2730 2767 HeapRegion* cur = r;
2731 2768 while (cur != NULL) {
2732 2769 HeapRegion* next = cur->next_in_collection_set();
2733 2770 if (cl->doHeapRegion(cur) && false) {
2734 2771 cl->incomplete();
2735 2772 return;
2736 2773 }
2737 2774 cur = next;
2738 2775 }
2739 2776 cur = g1_policy()->collection_set();
2740 2777 while (cur != r) {
2741 2778 HeapRegion* next = cur->next_in_collection_set();
2742 2779 if (cl->doHeapRegion(cur) && false) {
2743 2780 cl->incomplete();
2744 2781 return;
2745 2782 }
2746 2783 cur = next;
2747 2784 }
2748 2785 }
2749 2786
2750 2787 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2751 2788 return n_regions() > 0 ? region_at(0) : NULL;
2752 2789 }
2753 2790
2754 2791
2755 2792 Space* G1CollectedHeap::space_containing(const void* addr) const {
2756 2793 Space* res = heap_region_containing(addr);
2757 2794 if (res == NULL)
2758 2795 res = perm_gen()->space_containing(addr);
2759 2796 return res;
2760 2797 }
2761 2798
2762 2799 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2763 2800 Space* sp = space_containing(addr);
2764 2801 if (sp != NULL) {
2765 2802 return sp->block_start(addr);
2766 2803 }
2767 2804 return NULL;
2768 2805 }
2769 2806
2770 2807 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2771 2808 Space* sp = space_containing(addr);
2772 2809 assert(sp != NULL, "block_size of address outside of heap");
2773 2810 return sp->block_size(addr);
2774 2811 }
2775 2812
2776 2813 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2777 2814 Space* sp = space_containing(addr);
2778 2815 return sp->block_is_obj(addr);
2779 2816 }
2780 2817
2781 2818 bool G1CollectedHeap::supports_tlab_allocation() const {
2782 2819 return true;
2783 2820 }
2784 2821
2785 2822 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2786 2823 return HeapRegion::GrainBytes;
2787 2824 }
2788 2825
2789 2826 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2790 2827 // Return the remaining space in the cur alloc region, but not less than
2791 2828 // the min TLAB size.
2792 2829
2793 2830 // Also, this value can be at most the humongous object threshold,
2794 2831 // since we can't allow tlabs to grow big enough to accomodate
2795 2832 // humongous objects.
2796 2833
2797 2834 HeapRegion* hr = _mutator_alloc_region.get();
2798 2835 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2799 2836 if (hr == NULL) {
2800 2837 return max_tlab_size;
2801 2838 } else {
2802 2839 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2803 2840 }
2804 2841 }
2805 2842
2806 2843 size_t G1CollectedHeap::max_capacity() const {
2807 2844 return _g1_reserved.byte_size();
2808 2845 }
2809 2846
2810 2847 jlong G1CollectedHeap::millis_since_last_gc() {
2811 2848 // assert(false, "NYI");
2812 2849 return 0;
2813 2850 }
2814 2851
2815 2852 void G1CollectedHeap::prepare_for_verify() {
2816 2853 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2817 2854 ensure_parsability(false);
2818 2855 }
2819 2856 g1_rem_set()->prepare_for_verify();
2820 2857 }
2821 2858
2822 2859 class VerifyLivenessOopClosure: public OopClosure {
2823 2860 G1CollectedHeap* _g1h;
2824 2861 VerifyOption _vo;
2825 2862 public:
2826 2863 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
2827 2864 _g1h(g1h), _vo(vo)
2828 2865 { }
2829 2866 void do_oop(narrowOop *p) { do_oop_work(p); }
2830 2867 void do_oop( oop *p) { do_oop_work(p); }
2831 2868
2832 2869 template <class T> void do_oop_work(T *p) {
2833 2870 oop obj = oopDesc::load_decode_heap_oop(p);
2834 2871 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
2835 2872 "Dead object referenced by a not dead object");
2836 2873 }
2837 2874 };
2838 2875
2839 2876 class VerifyObjsInRegionClosure: public ObjectClosure {
2840 2877 private:
2841 2878 G1CollectedHeap* _g1h;
2842 2879 size_t _live_bytes;
2843 2880 HeapRegion *_hr;
2844 2881 VerifyOption _vo;
2845 2882 public:
2846 2883 // _vo == UsePrevMarking -> use "prev" marking information,
2847 2884 // _vo == UseNextMarking -> use "next" marking information,
2848 2885 // _vo == UseMarkWord -> use mark word from object header.
2849 2886 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
2850 2887 : _live_bytes(0), _hr(hr), _vo(vo) {
2851 2888 _g1h = G1CollectedHeap::heap();
2852 2889 }
2853 2890 void do_object(oop o) {
2854 2891 VerifyLivenessOopClosure isLive(_g1h, _vo);
2855 2892 assert(o != NULL, "Huh?");
2856 2893 if (!_g1h->is_obj_dead_cond(o, _vo)) {
2857 2894 // If the object is alive according to the mark word,
2858 2895 // then verify that the marking information agrees.
2859 2896 // Note we can't verify the contra-positive of the
2860 2897 // above: if the object is dead (according to the mark
2861 2898 // word), it may not be marked, or may have been marked
2862 2899 // but has since became dead, or may have been allocated
2863 2900 // since the last marking.
2864 2901 if (_vo == VerifyOption_G1UseMarkWord) {
2865 2902 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
2866 2903 }
2867 2904
2868 2905 o->oop_iterate(&isLive);
2869 2906 if (!_hr->obj_allocated_since_prev_marking(o)) {
2870 2907 size_t obj_size = o->size(); // Make sure we don't overflow
2871 2908 _live_bytes += (obj_size * HeapWordSize);
2872 2909 }
2873 2910 }
2874 2911 }
2875 2912 size_t live_bytes() { return _live_bytes; }
2876 2913 };
2877 2914
2878 2915 class PrintObjsInRegionClosure : public ObjectClosure {
2879 2916 HeapRegion *_hr;
2880 2917 G1CollectedHeap *_g1;
2881 2918 public:
2882 2919 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
2883 2920 _g1 = G1CollectedHeap::heap();
2884 2921 };
2885 2922
2886 2923 void do_object(oop o) {
2887 2924 if (o != NULL) {
2888 2925 HeapWord *start = (HeapWord *) o;
2889 2926 size_t word_sz = o->size();
2890 2927 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
2891 2928 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
2892 2929 (void*) o, word_sz,
2893 2930 _g1->isMarkedPrev(o),
2894 2931 _g1->isMarkedNext(o),
2895 2932 _hr->obj_allocated_since_prev_marking(o));
2896 2933 HeapWord *end = start + word_sz;
2897 2934 HeapWord *cur;
2898 2935 int *val;
2899 2936 for (cur = start; cur < end; cur++) {
2900 2937 val = (int *) cur;
2901 2938 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
2902 2939 }
2903 2940 }
2904 2941 }
2905 2942 };
2906 2943
2907 2944 class VerifyRegionClosure: public HeapRegionClosure {
2908 2945 private:
2909 2946 bool _allow_dirty;
2910 2947 bool _par;
2911 2948 VerifyOption _vo;
2912 2949 bool _failures;
2913 2950 public:
2914 2951 // _vo == UsePrevMarking -> use "prev" marking information,
2915 2952 // _vo == UseNextMarking -> use "next" marking information,
2916 2953 // _vo == UseMarkWord -> use mark word from object header.
2917 2954 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo)
2918 2955 : _allow_dirty(allow_dirty),
2919 2956 _par(par),
2920 2957 _vo(vo),
2921 2958 _failures(false) {}
2922 2959
2923 2960 bool failures() {
2924 2961 return _failures;
2925 2962 }
2926 2963
2927 2964 bool doHeapRegion(HeapRegion* r) {
2928 2965 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue,
2929 2966 "Should be unclaimed at verify points.");
2930 2967 if (!r->continuesHumongous()) {
2931 2968 bool failures = false;
2932 2969 r->verify(_allow_dirty, _vo, &failures);
2933 2970 if (failures) {
2934 2971 _failures = true;
2935 2972 } else {
2936 2973 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
2937 2974 r->object_iterate(¬_dead_yet_cl);
2938 2975 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
2939 2976 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
2940 2977 "max_live_bytes "SIZE_FORMAT" "
2941 2978 "< calculated "SIZE_FORMAT,
2942 2979 r->bottom(), r->end(),
2943 2980 r->max_live_bytes(),
2944 2981 not_dead_yet_cl.live_bytes());
2945 2982 _failures = true;
2946 2983 }
2947 2984 }
2948 2985 }
2949 2986 return false; // stop the region iteration if we hit a failure
2950 2987 }
2951 2988 };
2952 2989
2953 2990 class VerifyRootsClosure: public OopsInGenClosure {
2954 2991 private:
2955 2992 G1CollectedHeap* _g1h;
2956 2993 VerifyOption _vo;
2957 2994 bool _failures;
2958 2995 public:
2959 2996 // _vo == UsePrevMarking -> use "prev" marking information,
2960 2997 // _vo == UseNextMarking -> use "next" marking information,
2961 2998 // _vo == UseMarkWord -> use mark word from object header.
2962 2999 VerifyRootsClosure(VerifyOption vo) :
2963 3000 _g1h(G1CollectedHeap::heap()),
2964 3001 _vo(vo),
2965 3002 _failures(false) { }
2966 3003
2967 3004 bool failures() { return _failures; }
2968 3005
2969 3006 template <class T> void do_oop_nv(T* p) {
2970 3007 T heap_oop = oopDesc::load_heap_oop(p);
2971 3008 if (!oopDesc::is_null(heap_oop)) {
2972 3009 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
2973 3010 if (_g1h->is_obj_dead_cond(obj, _vo)) {
2974 3011 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
2975 3012 "points to dead obj "PTR_FORMAT, p, (void*) obj);
2976 3013 if (_vo == VerifyOption_G1UseMarkWord) {
2977 3014 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
2978 3015 }
2979 3016 obj->print_on(gclog_or_tty);
2980 3017 _failures = true;
2981 3018 }
2982 3019 }
2983 3020 }
2984 3021
2985 3022 void do_oop(oop* p) { do_oop_nv(p); }
2986 3023 void do_oop(narrowOop* p) { do_oop_nv(p); }
2987 3024 };
2988 3025
2989 3026 // This is the task used for parallel heap verification.
2990 3027
2991 3028 class G1ParVerifyTask: public AbstractGangTask {
2992 3029 private:
2993 3030 G1CollectedHeap* _g1h;
2994 3031 bool _allow_dirty;
2995 3032 VerifyOption _vo;
2996 3033 bool _failures;
2997 3034
2998 3035 public:
2999 3036 // _vo == UsePrevMarking -> use "prev" marking information,
3000 3037 // _vo == UseNextMarking -> use "next" marking information,
3001 3038 // _vo == UseMarkWord -> use mark word from object header.
3002 3039 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) :
3003 3040 AbstractGangTask("Parallel verify task"),
3004 3041 _g1h(g1h),
3005 3042 _allow_dirty(allow_dirty),
3006 3043 _vo(vo),
3007 3044 _failures(false) { }
3008 3045
3009 3046 bool failures() {
3010 3047 return _failures;
3011 3048 }
3012 3049
3013 3050 void work(int worker_i) {
3014 3051 HandleMark hm;
3015 3052 VerifyRegionClosure blk(_allow_dirty, true, _vo);
3016 3053 _g1h->heap_region_par_iterate_chunked(&blk, worker_i,
3017 3054 _g1h->workers()->active_workers(),
3018 3055 HeapRegion::ParVerifyClaimValue);
3019 3056 if (blk.failures()) {
3020 3057 _failures = true;
3021 3058 }
3022 3059 }
3023 3060 };
3024 3061
3025 3062 void G1CollectedHeap::verify(bool allow_dirty, bool silent) {
3026 3063 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking);
3027 3064 }
3028 3065
3029 3066 void G1CollectedHeap::verify(bool allow_dirty,
3030 3067 bool silent,
3031 3068 VerifyOption vo) {
3032 3069 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3033 3070 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); }
3034 3071 VerifyRootsClosure rootsCl(vo);
3035 3072
3036 3073 assert(Thread::current()->is_VM_thread(),
3037 3074 "Expected to be executed serially by the VM thread at this point");
3038 3075
3039 3076 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3040 3077
3041 3078 // We apply the relevant closures to all the oops in the
3042 3079 // system dictionary, the string table and the code cache.
3043 3080 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3044 3081
3045 3082 process_strong_roots(true, // activate StrongRootsScope
3046 3083 true, // we set "collecting perm gen" to true,
3047 3084 // so we don't reset the dirty cards in the perm gen.
3048 3085 SharedHeap::ScanningOption(so), // roots scanning options
3049 3086 &rootsCl,
3050 3087 &blobsCl,
3051 3088 &rootsCl);
3052 3089
3053 3090 // If we're verifying after the marking phase of a Full GC then we can't
3054 3091 // treat the perm gen as roots into the G1 heap. Some of the objects in
3055 3092 // the perm gen may be dead and hence not marked. If one of these dead
3056 3093 // objects is considered to be a root then we may end up with a false
3057 3094 // "Root location <x> points to dead ob <y>" failure.
3058 3095 if (vo != VerifyOption_G1UseMarkWord) {
3059 3096 // Since we used "collecting_perm_gen" == true above, we will not have
3060 3097 // checked the refs from perm into the G1-collected heap. We check those
3061 3098 // references explicitly below. Whether the relevant cards are dirty
3062 3099 // is checked further below in the rem set verification.
3063 3100 if (!silent) { gclog_or_tty->print("Permgen roots "); }
3064 3101 perm_gen()->oop_iterate(&rootsCl);
3065 3102 }
3066 3103 bool failures = rootsCl.failures();
3067 3104
3068 3105 if (vo != VerifyOption_G1UseMarkWord) {
3069 3106 // If we're verifying during a full GC then the region sets
3070 3107 // will have been torn down at the start of the GC. Therefore
3071 3108 // verifying the region sets will fail. So we only verify
3072 3109 // the region sets when not in a full GC.
3073 3110 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3074 3111 verify_region_sets();
3075 3112 }
3076 3113
3077 3114 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3078 3115 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3079 3116 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3080 3117 "sanity check");
3081 3118
3082 3119 G1ParVerifyTask task(this, allow_dirty, vo);
3083 3120 assert(UseDynamicNumberOfGCThreads ||
3084 3121 workers()->active_workers() == workers()->total_workers(),
3085 3122 "If not dynamic should be using all the workers");
3086 3123 int n_workers = workers()->active_workers();
3087 3124 set_par_threads(n_workers);
3088 3125 workers()->run_task(&task);
3089 3126 set_par_threads(0);
3090 3127 if (task.failures()) {
3091 3128 failures = true;
3092 3129 }
3093 3130
3094 3131 // Checks that the expected amount of parallel work was done.
3095 3132 // The implication is that n_workers is > 0.
3096 3133 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3097 3134 "sanity check");
3098 3135
3099 3136 reset_heap_region_claim_values();
3100 3137
3101 3138 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3102 3139 "sanity check");
3103 3140 } else {
3104 3141 VerifyRegionClosure blk(allow_dirty, false, vo);
3105 3142 heap_region_iterate(&blk);
3106 3143 if (blk.failures()) {
3107 3144 failures = true;
3108 3145 }
3109 3146 }
3110 3147 if (!silent) gclog_or_tty->print("RemSet ");
3111 3148 rem_set()->verify();
3112 3149
3113 3150 if (failures) {
3114 3151 gclog_or_tty->print_cr("Heap:");
3115 3152 // It helps to have the per-region information in the output to
3116 3153 // help us track down what went wrong. This is why we call
3117 3154 // print_extended_on() instead of print_on().
3118 3155 print_extended_on(gclog_or_tty);
3119 3156 gclog_or_tty->print_cr("");
3120 3157 #ifndef PRODUCT
3121 3158 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3122 3159 concurrent_mark()->print_reachable("at-verification-failure",
3123 3160 vo, false /* all */);
3124 3161 }
3125 3162 #endif
3126 3163 gclog_or_tty->flush();
3127 3164 }
3128 3165 guarantee(!failures, "there should not have been any failures");
3129 3166 } else {
3130 3167 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3131 3168 }
3132 3169 }
3133 3170
3134 3171 class PrintRegionClosure: public HeapRegionClosure {
3135 3172 outputStream* _st;
3136 3173 public:
3137 3174 PrintRegionClosure(outputStream* st) : _st(st) {}
3138 3175 bool doHeapRegion(HeapRegion* r) {
3139 3176 r->print_on(_st);
3140 3177 return false;
3141 3178 }
3142 3179 };
3143 3180
3144 3181 void G1CollectedHeap::print_on(outputStream* st) const {
3145 3182 st->print(" %-20s", "garbage-first heap");
3146 3183 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3147 3184 capacity()/K, used_unlocked()/K);
3148 3185 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3149 3186 _g1_storage.low_boundary(),
3150 3187 _g1_storage.high(),
3151 3188 _g1_storage.high_boundary());
3152 3189 st->cr();
3153 3190 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3154 3191 size_t young_regions = _young_list->length();
3155 3192 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ",
3156 3193 young_regions, young_regions * HeapRegion::GrainBytes / K);
3157 3194 size_t survivor_regions = g1_policy()->recorded_survivor_regions();
3158 3195 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)",
3159 3196 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K);
3160 3197 st->cr();
3161 3198 perm()->as_gen()->print_on(st);
3162 3199 }
3163 3200
3164 3201 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3165 3202 print_on(st);
3166 3203
3167 3204 // Print the per-region information.
3168 3205 st->cr();
3169 3206 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), HS=humongous(starts), HC=humongous(continues), CS=collection set, F=free, TS=gc time stamp, PTAMS=previous top-at-mark-start, NTAMS=next top-at-mark-start)");
3170 3207 PrintRegionClosure blk(st);
3171 3208 heap_region_iterate(&blk);
3172 3209 }
3173 3210
3174 3211 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3175 3212 if (G1CollectedHeap::use_parallel_gc_threads()) {
3176 3213 workers()->print_worker_threads_on(st);
3177 3214 }
3178 3215 _cmThread->print_on(st);
3179 3216 st->cr();
3180 3217 _cm->print_worker_threads_on(st);
3181 3218 _cg1r->print_worker_threads_on(st);
3182 3219 st->cr();
3183 3220 }
3184 3221
3185 3222 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3186 3223 if (G1CollectedHeap::use_parallel_gc_threads()) {
3187 3224 workers()->threads_do(tc);
3188 3225 }
3189 3226 tc->do_thread(_cmThread);
3190 3227 _cg1r->threads_do(tc);
3191 3228 }
3192 3229
3193 3230 void G1CollectedHeap::print_tracing_info() const {
3194 3231 // We'll overload this to mean "trace GC pause statistics."
3195 3232 if (TraceGen0Time || TraceGen1Time) {
3196 3233 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3197 3234 // to that.
3198 3235 g1_policy()->print_tracing_info();
3199 3236 }
3200 3237 if (G1SummarizeRSetStats) {
3201 3238 g1_rem_set()->print_summary_info();
3202 3239 }
3203 3240 if (G1SummarizeConcMark) {
3204 3241 concurrent_mark()->print_summary_info();
3205 3242 }
3206 3243 g1_policy()->print_yg_surv_rate_info();
3207 3244 SpecializationStats::print();
3208 3245 }
3209 3246
3210 3247 #ifndef PRODUCT
3211 3248 // Helpful for debugging RSet issues.
3212 3249
3213 3250 class PrintRSetsClosure : public HeapRegionClosure {
3214 3251 private:
3215 3252 const char* _msg;
3216 3253 size_t _occupied_sum;
3217 3254
3218 3255 public:
3219 3256 bool doHeapRegion(HeapRegion* r) {
3220 3257 HeapRegionRemSet* hrrs = r->rem_set();
3221 3258 size_t occupied = hrrs->occupied();
3222 3259 _occupied_sum += occupied;
3223 3260
3224 3261 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3225 3262 HR_FORMAT_PARAMS(r));
3226 3263 if (occupied == 0) {
3227 3264 gclog_or_tty->print_cr(" RSet is empty");
3228 3265 } else {
3229 3266 hrrs->print();
3230 3267 }
3231 3268 gclog_or_tty->print_cr("----------");
3232 3269 return false;
3233 3270 }
3234 3271
3235 3272 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3236 3273 gclog_or_tty->cr();
3237 3274 gclog_or_tty->print_cr("========================================");
3238 3275 gclog_or_tty->print_cr(msg);
3239 3276 gclog_or_tty->cr();
3240 3277 }
3241 3278
3242 3279 ~PrintRSetsClosure() {
3243 3280 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3244 3281 gclog_or_tty->print_cr("========================================");
3245 3282 gclog_or_tty->cr();
3246 3283 }
3247 3284 };
3248 3285
3249 3286 void G1CollectedHeap::print_cset_rsets() {
3250 3287 PrintRSetsClosure cl("Printing CSet RSets");
3251 3288 collection_set_iterate(&cl);
3252 3289 }
3253 3290
3254 3291 void G1CollectedHeap::print_all_rsets() {
3255 3292 PrintRSetsClosure cl("Printing All RSets");;
3256 3293 heap_region_iterate(&cl);
3257 3294 }
3258 3295 #endif // PRODUCT
3259 3296
3260 3297 G1CollectedHeap* G1CollectedHeap::heap() {
3261 3298 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3262 3299 "not a garbage-first heap");
3263 3300 return _g1h;
3264 3301 }
3265 3302
3266 3303 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3267 3304 // always_do_update_barrier = false;
3268 3305 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3269 3306 // Call allocation profiler
3270 3307 AllocationProfiler::iterate_since_last_gc();
3271 3308 // Fill TLAB's and such
3272 3309 ensure_parsability(true);
3273 3310 }
3274 3311
3275 3312 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3276 3313 // FIXME: what is this about?
3277 3314 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3278 3315 // is set.
3279 3316 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3280 3317 "derived pointer present"));
3281 3318 // always_do_update_barrier = true;
3282 3319
3283 3320 // We have just completed a GC. Update the soft reference
3284 3321 // policy with the new heap occupancy
3285 3322 Universe::update_heap_info_at_gc();
3286 3323 }
3287 3324
3288 3325 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3289 3326 unsigned int gc_count_before,
3290 3327 bool* succeeded) {
3291 3328 assert_heap_not_locked_and_not_at_safepoint();
3292 3329 g1_policy()->record_stop_world_start();
3293 3330 VM_G1IncCollectionPause op(gc_count_before,
3294 3331 word_size,
3295 3332 false, /* should_initiate_conc_mark */
3296 3333 g1_policy()->max_pause_time_ms(),
3297 3334 GCCause::_g1_inc_collection_pause);
3298 3335 VMThread::execute(&op);
3299 3336
3300 3337 HeapWord* result = op.result();
3301 3338 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3302 3339 assert(result == NULL || ret_succeeded,
3303 3340 "the result should be NULL if the VM did not succeed");
3304 3341 *succeeded = ret_succeeded;
3305 3342
3306 3343 assert_heap_not_locked();
3307 3344 return result;
3308 3345 }
3309 3346
3310 3347 void
3311 3348 G1CollectedHeap::doConcurrentMark() {
3312 3349 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3313 3350 if (!_cmThread->in_progress()) {
3314 3351 _cmThread->set_started();
3315 3352 CGC_lock->notify();
3316 3353 }
3317 3354 }
3318 3355
3319 3356 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr,
3320 3357 bool young) {
3321 3358 return _g1_policy->predict_region_elapsed_time_ms(hr, young);
3322 3359 }
3323 3360
3324 3361 void G1CollectedHeap::check_if_region_is_too_expensive(double
3325 3362 predicted_time_ms) {
3326 3363 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms);
3327 3364 }
3328 3365
3329 3366 size_t G1CollectedHeap::pending_card_num() {
3330 3367 size_t extra_cards = 0;
3331 3368 JavaThread *curr = Threads::first();
3332 3369 while (curr != NULL) {
3333 3370 DirtyCardQueue& dcq = curr->dirty_card_queue();
3334 3371 extra_cards += dcq.size();
3335 3372 curr = curr->next();
3336 3373 }
3337 3374 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3338 3375 size_t buffer_size = dcqs.buffer_size();
3339 3376 size_t buffer_num = dcqs.completed_buffers_num();
3340 3377 return buffer_size * buffer_num + extra_cards;
3341 3378 }
3342 3379
3343 3380 size_t G1CollectedHeap::max_pending_card_num() {
3344 3381 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3345 3382 size_t buffer_size = dcqs.buffer_size();
3346 3383 size_t buffer_num = dcqs.completed_buffers_num();
3347 3384 int thread_num = Threads::number_of_threads();
3348 3385 return (buffer_num + thread_num) * buffer_size;
3349 3386 }
3350 3387
3351 3388 size_t G1CollectedHeap::cards_scanned() {
3352 3389 return g1_rem_set()->cardsScanned();
3353 3390 }
3354 3391
3355 3392 void
3356 3393 G1CollectedHeap::setup_surviving_young_words() {
3357 3394 guarantee( _surviving_young_words == NULL, "pre-condition" );
3358 3395 size_t array_length = g1_policy()->young_cset_region_length();
3359 3396 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length);
3360 3397 if (_surviving_young_words == NULL) {
3361 3398 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3362 3399 "Not enough space for young surv words summary.");
3363 3400 }
3364 3401 memset(_surviving_young_words, 0, array_length * sizeof(size_t));
3365 3402 #ifdef ASSERT
3366 3403 for (size_t i = 0; i < array_length; ++i) {
3367 3404 assert( _surviving_young_words[i] == 0, "memset above" );
3368 3405 }
3369 3406 #endif // !ASSERT
3370 3407 }
3371 3408
3372 3409 void
3373 3410 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3374 3411 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3375 3412 size_t array_length = g1_policy()->young_cset_region_length();
3376 3413 for (size_t i = 0; i < array_length; ++i)
3377 3414 _surviving_young_words[i] += surv_young_words[i];
3378 3415 }
3379 3416
3380 3417 void
3381 3418 G1CollectedHeap::cleanup_surviving_young_words() {
3382 3419 guarantee( _surviving_young_words != NULL, "pre-condition" );
3383 3420 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words);
3384 3421 _surviving_young_words = NULL;
3385 3422 }
3386 3423
3387 3424 #ifdef ASSERT
3388 3425 class VerifyCSetClosure: public HeapRegionClosure {
3389 3426 public:
3390 3427 bool doHeapRegion(HeapRegion* hr) {
3391 3428 // Here we check that the CSet region's RSet is ready for parallel
3392 3429 // iteration. The fields that we'll verify are only manipulated
3393 3430 // when the region is part of a CSet and is collected. Afterwards,
3394 3431 // we reset these fields when we clear the region's RSet (when the
3395 3432 // region is freed) so they are ready when the region is
3396 3433 // re-allocated. The only exception to this is if there's an
3397 3434 // evacuation failure and instead of freeing the region we leave
3398 3435 // it in the heap. In that case, we reset these fields during
3399 3436 // evacuation failure handling.
3400 3437 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3401 3438
3402 3439 // Here's a good place to add any other checks we'd like to
3403 3440 // perform on CSet regions.
3404 3441 return false;
3405 3442 }
3406 3443 };
3407 3444 #endif // ASSERT
3408 3445
3409 3446 #if TASKQUEUE_STATS
3410 3447 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3411 3448 st->print_raw_cr("GC Task Stats");
3412 3449 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3413 3450 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3414 3451 }
3415 3452
3416 3453 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3417 3454 print_taskqueue_stats_hdr(st);
3418 3455
3419 3456 TaskQueueStats totals;
3420 3457 const int n = workers() != NULL ? workers()->total_workers() : 1;
3421 3458 for (int i = 0; i < n; ++i) {
3422 3459 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3423 3460 totals += task_queue(i)->stats;
3424 3461 }
3425 3462 st->print_raw("tot "); totals.print(st); st->cr();
3426 3463
3427 3464 DEBUG_ONLY(totals.verify());
3428 3465 }
3429 3466
3430 3467 void G1CollectedHeap::reset_taskqueue_stats() {
3431 3468 const int n = workers() != NULL ? workers()->total_workers() : 1;
3432 3469 for (int i = 0; i < n; ++i) {
3433 3470 task_queue(i)->stats.reset();
3434 3471 }
3435 3472 }
3436 3473 #endif // TASKQUEUE_STATS
3437 3474
3438 3475 bool
3439 3476 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3440 3477 assert_at_safepoint(true /* should_be_vm_thread */);
3441 3478 guarantee(!is_gc_active(), "collection is not reentrant");
3442 3479
3443 3480 if (GC_locker::check_active_before_gc()) {
3444 3481 return false;
3445 3482 }
3446 3483
3447 3484 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3448 3485 ResourceMark rm;
3449 3486
3450 3487 if (PrintHeapAtGC) {
3451 3488 Universe::print_heap_before_gc();
3452 3489 }
3453 3490
3454 3491 HRSPhaseSetter x(HRSPhaseEvacuation);
3455 3492 verify_region_sets_optional();
3456 3493 verify_dirty_young_regions();
3457 3494
3458 3495 {
3459 3496 // This call will decide whether this pause is an initial-mark
3460 3497 // pause. If it is, during_initial_mark_pause() will return true
3461 3498 // for the duration of this pause.
3462 3499 g1_policy()->decide_on_conc_mark_initiation();
3463 3500
3464 3501 // We do not allow initial-mark to be piggy-backed on a
3465 3502 // partially-young GC.
3466 3503 assert(!g1_policy()->during_initial_mark_pause() ||
3467 3504 g1_policy()->full_young_gcs(), "sanity");
3468 3505
3469 3506 // We also do not allow partially-young GCs during marking.
3470 3507 assert(!mark_in_progress() || g1_policy()->full_young_gcs(), "sanity");
3471 3508
3472 3509 char verbose_str[128];
3473 3510 sprintf(verbose_str, "GC pause ");
3474 3511 if (g1_policy()->full_young_gcs()) {
3475 3512 strcat(verbose_str, "(young)");
3476 3513 } else {
3477 3514 strcat(verbose_str, "(partial)");
3478 3515 }
3479 3516 if (g1_policy()->during_initial_mark_pause()) {
3480 3517 strcat(verbose_str, " (initial-mark)");
3481 3518 // We are about to start a marking cycle, so we increment the
3482 3519 // full collection counter.
3483 3520 increment_total_full_collections();
3484 3521 }
3485 3522
3486 3523 // if PrintGCDetails is on, we'll print long statistics information
3487 3524 // in the collector policy code, so let's not print this as the output
3488 3525 // is messy if we do.
3489 3526 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
3490 3527 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3491 3528 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty);
3492 3529
3493 3530 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3494 3531 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3495 3532
3496 3533 // If the secondary_free_list is not empty, append it to the
3497 3534 // free_list. No need to wait for the cleanup operation to finish;
3498 3535 // the region allocation code will check the secondary_free_list
3499 3536 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3500 3537 // set, skip this step so that the region allocation code has to
3501 3538 // get entries from the secondary_free_list.
3502 3539 if (!G1StressConcRegionFreeing) {
3503 3540 append_secondary_free_list_if_not_empty_with_lock();
3504 3541 }
3505 3542
3506 3543 assert(check_young_list_well_formed(),
3507 3544 "young list should be well formed");
3508 3545
3509 3546 // Don't dynamically change the number of GC threads this early. A value of
3510 3547 // 0 is used to indicate serial work. When parallel work is done,
3511 3548 // it will be set.
3512 3549
3513 3550 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3514 3551 IsGCActiveMark x;
3515 3552
3516 3553 gc_prologue(false);
3517 3554 increment_total_collections(false /* full gc */);
3518 3555 increment_gc_time_stamp();
3519 3556
3520 3557 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
3521 3558 HandleMark hm; // Discard invalid handles created during verification
3522 3559 gclog_or_tty->print(" VerifyBeforeGC:");
3523 3560 prepare_for_verify();
3524 3561 Universe::verify(/* allow dirty */ false,
↓ open down ↓ |
809 lines elided |
↑ open up ↑ |
3525 3562 /* silent */ false,
3526 3563 /* option */ VerifyOption_G1UsePrevMarking);
3527 3564
3528 3565 }
3529 3566
3530 3567 COMPILER2_PRESENT(DerivedPointerTable::clear());
3531 3568
3532 3569 // Please see comment in g1CollectedHeap.hpp and
3533 3570 // G1CollectedHeap::ref_processing_init() to see how
3534 3571 // reference processing currently works in G1.
3535 -
3572 +
3536 3573 // Enable discovery in the STW reference processor
3537 3574 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3538 3575 true /*verify_no_refs*/);
3539 3576
3540 3577 {
3541 3578 // We want to temporarily turn off discovery by the
3542 3579 // CM ref processor, if necessary, and turn it back on
3543 3580 // on again later if we do. Using a scoped
3544 3581 // NoRefDiscovery object will do this.
3545 3582 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3546 3583
3547 3584 // Forget the current alloc region (we might even choose it to be part
3548 3585 // of the collection set!).
3549 3586 release_mutator_alloc_region();
3550 3587
3551 3588 // We should call this after we retire the mutator alloc
3552 3589 // region(s) so that all the ALLOC / RETIRE events are generated
3553 3590 // before the start GC event.
3554 3591 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3555 3592
3556 3593 // The elapsed time induced by the start time below deliberately elides
3557 3594 // the possible verification above.
3558 3595 double start_time_sec = os::elapsedTime();
3559 3596 size_t start_used_bytes = used();
3560 3597
3561 3598 #if YOUNG_LIST_VERBOSE
3562 3599 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3563 3600 _young_list->print();
3564 3601 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3565 3602 #endif // YOUNG_LIST_VERBOSE
3566 3603
3567 3604 g1_policy()->record_collection_pause_start(start_time_sec,
3568 3605 start_used_bytes);
3569 3606
3570 3607 #if YOUNG_LIST_VERBOSE
3571 3608 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3572 3609 _young_list->print();
3573 3610 #endif // YOUNG_LIST_VERBOSE
3574 3611
3575 3612 if (g1_policy()->during_initial_mark_pause()) {
3576 3613 concurrent_mark()->checkpointRootsInitialPre();
3577 3614 }
3578 3615 perm_gen()->save_marks();
3579 3616
3580 3617 // We must do this before any possible evacuation that should propagate
3581 3618 // marks.
3582 3619 if (mark_in_progress()) {
3583 3620 double start_time_sec = os::elapsedTime();
3584 3621
3585 3622 _cm->drainAllSATBBuffers();
3586 3623 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0;
3587 3624 g1_policy()->record_satb_drain_time(finish_mark_ms);
3588 3625 }
3589 3626 // Record the number of elements currently on the mark stack, so we
3590 3627 // only iterate over these. (Since evacuation may add to the mark
3591 3628 // stack, doing more exposes race conditions.) If no mark is in
3592 3629 // progress, this will be zero.
3593 3630 _cm->set_oops_do_bound();
3594 3631
3595 3632 if (mark_in_progress()) {
3596 3633 concurrent_mark()->newCSet();
3597 3634 }
3598 3635
3599 3636 #if YOUNG_LIST_VERBOSE
3600 3637 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3601 3638 _young_list->print();
3602 3639 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3603 3640 #endif // YOUNG_LIST_VERBOSE
3604 3641
3605 3642 g1_policy()->choose_collection_set(target_pause_time_ms);
3606 3643
3607 3644 if (_hr_printer.is_active()) {
3608 3645 HeapRegion* hr = g1_policy()->collection_set();
3609 3646 while (hr != NULL) {
3610 3647 G1HRPrinter::RegionType type;
3611 3648 if (!hr->is_young()) {
3612 3649 type = G1HRPrinter::Old;
3613 3650 } else if (hr->is_survivor()) {
3614 3651 type = G1HRPrinter::Survivor;
3615 3652 } else {
3616 3653 type = G1HRPrinter::Eden;
3617 3654 }
3618 3655 _hr_printer.cset(hr);
3619 3656 hr = hr->next_in_collection_set();
3620 3657 }
3621 3658 }
3622 3659
3623 3660 // We have chosen the complete collection set. If marking is
3624 3661 // active then, we clear the region fields of any of the
3625 3662 // concurrent marking tasks whose region fields point into
3626 3663 // the collection set as these values will become stale. This
3627 3664 // will cause the owning marking threads to claim a new region
3628 3665 // when marking restarts.
3629 3666 if (mark_in_progress()) {
3630 3667 concurrent_mark()->reset_active_task_region_fields_in_cset();
3631 3668 }
3632 3669
3633 3670 #ifdef ASSERT
3634 3671 VerifyCSetClosure cl;
3635 3672 collection_set_iterate(&cl);
3636 3673 #endif // ASSERT
3637 3674
3638 3675 setup_surviving_young_words();
3639 3676
3640 3677 // Initialize the GC alloc regions.
3641 3678 init_gc_alloc_regions();
3642 3679
3643 3680 // Actually do the work...
3644 3681 evacuate_collection_set();
3645 3682
3646 3683 free_collection_set(g1_policy()->collection_set());
3647 3684 g1_policy()->clear_collection_set();
3648 3685
3649 3686 cleanup_surviving_young_words();
3650 3687
3651 3688 // Start a new incremental collection set for the next pause.
3652 3689 g1_policy()->start_incremental_cset_building();
3653 3690
3654 3691 // Clear the _cset_fast_test bitmap in anticipation of adding
3655 3692 // regions to the incremental collection set for the next
3656 3693 // evacuation pause.
3657 3694 clear_cset_fast_test();
3658 3695
3659 3696 _young_list->reset_sampled_info();
3660 3697
3661 3698 // Don't check the whole heap at this point as the
3662 3699 // GC alloc regions from this pause have been tagged
3663 3700 // as survivors and moved on to the survivor list.
3664 3701 // Survivor regions will fail the !is_young() check.
3665 3702 assert(check_young_list_empty(false /* check_heap */),
3666 3703 "young list should be empty");
3667 3704
3668 3705 #if YOUNG_LIST_VERBOSE
3669 3706 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3670 3707 _young_list->print();
3671 3708 #endif // YOUNG_LIST_VERBOSE
3672 3709
3673 3710 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3674 3711 _young_list->first_survivor_region(),
3675 3712 _young_list->last_survivor_region());
3676 3713
3677 3714 _young_list->reset_auxilary_lists();
3678 3715
3679 3716 if (evacuation_failed()) {
3680 3717 _summary_bytes_used = recalculate_used();
3681 3718 } else {
3682 3719 // The "used" of the the collection set have already been subtracted
3683 3720 // when they were freed. Add in the bytes evacuated.
3684 3721 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3685 3722 }
3686 3723
3687 3724 if (g1_policy()->during_initial_mark_pause()) {
3688 3725 concurrent_mark()->checkpointRootsInitialPost();
3689 3726 set_marking_started();
3690 3727 // CAUTION: after the doConcurrentMark() call below,
3691 3728 // the concurrent marking thread(s) could be running
3692 3729 // concurrently with us. Make sure that anything after
3693 3730 // this point does not assume that we are the only GC thread
3694 3731 // running. Note: of course, the actual marking work will
3695 3732 // not start until the safepoint itself is released in
3696 3733 // ConcurrentGCThread::safepoint_desynchronize().
3697 3734 doConcurrentMark();
3698 3735 }
3699 3736
3700 3737 allocate_dummy_regions();
3701 3738
3702 3739 #if YOUNG_LIST_VERBOSE
3703 3740 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3704 3741 _young_list->print();
3705 3742 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3706 3743 #endif // YOUNG_LIST_VERBOSE
3707 3744
3708 3745 init_mutator_alloc_region();
3709 3746
3710 3747 {
3711 3748 size_t expand_bytes = g1_policy()->expansion_amount();
3712 3749 if (expand_bytes > 0) {
3713 3750 size_t bytes_before = capacity();
3714 3751 if (!expand(expand_bytes)) {
3715 3752 // We failed to expand the heap so let's verify that
3716 3753 // committed/uncommitted amount match the backing store
3717 3754 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3718 3755 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3719 3756 }
3720 3757 }
3721 3758 }
3722 3759
3723 3760 double end_time_sec = os::elapsedTime();
3724 3761 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS;
3725 3762 g1_policy()->record_pause_time_ms(pause_time_ms);
3726 3763 int active_gc_threads = workers()->active_workers();
3727 3764 g1_policy()->record_collection_pause_end(active_gc_threads);
3728 3765
3729 3766 MemoryService::track_memory_usage();
3730 3767
3731 3768 // In prepare_for_verify() below we'll need to scan the deferred
3732 3769 // update buffers to bring the RSets up-to-date if
3733 3770 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3734 3771 // the update buffers we'll probably need to scan cards on the
3735 3772 // regions we just allocated to (i.e., the GC alloc
3736 3773 // regions). However, during the last GC we called
3737 3774 // set_saved_mark() on all the GC alloc regions, so card
3738 3775 // scanning might skip the [saved_mark_word()...top()] area of
3739 3776 // those regions (i.e., the area we allocated objects into
3740 3777 // during the last GC). But it shouldn't. Given that
3741 3778 // saved_mark_word() is conditional on whether the GC time stamp
3742 3779 // on the region is current or not, by incrementing the GC time
3743 3780 // stamp here we invalidate all the GC time stamps on all the
3744 3781 // regions and saved_mark_word() will simply return top() for
3745 3782 // all the regions. This is a nicer way of ensuring this rather
3746 3783 // than iterating over the regions and fixing them. In fact, the
3747 3784 // GC time stamp increment here also ensures that
3748 3785 // saved_mark_word() will return top() between pauses, i.e.,
3749 3786 // during concurrent refinement. So we don't need the
3750 3787 // is_gc_active() check to decided which top to use when
3751 3788 // scanning cards (see CR 7039627).
3752 3789 increment_gc_time_stamp();
3753 3790
3754 3791 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
3755 3792 HandleMark hm; // Discard invalid handles created during verification
3756 3793 gclog_or_tty->print(" VerifyAfterGC:");
3757 3794 prepare_for_verify();
3758 3795 Universe::verify(/* allow dirty */ true,
3759 3796 /* silent */ false,
3760 3797 /* option */ VerifyOption_G1UsePrevMarking);
3761 3798 }
3762 3799
3763 3800 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
3764 3801 ref_processor_stw()->verify_no_references_recorded();
3765 3802
3766 3803 // CM reference discovery will be re-enabled if necessary.
3767 3804 }
3768 3805
3769 3806 {
3770 3807 size_t expand_bytes = g1_policy()->expansion_amount();
3771 3808 if (expand_bytes > 0) {
3772 3809 size_t bytes_before = capacity();
3773 3810 // No need for an ergo verbose message here,
3774 3811 // expansion_amount() does this when it returns a value > 0.
3775 3812 if (!expand(expand_bytes)) {
3776 3813 // We failed to expand the heap so let's verify that
3777 3814 // committed/uncommitted amount match the backing store
3778 3815 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3779 3816 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3780 3817 }
3781 3818 }
3782 3819 }
3783 3820
3784 3821 // We should do this after we potentially expand the heap so
3785 3822 // that all the COMMIT events are generated before the end GC
3786 3823 // event, and after we retire the GC alloc regions so that all
3787 3824 // RETIRE events are generated before the end GC event.
3788 3825 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
3789 3826
3790 3827 // We have to do this after we decide whether to expand the heap or not.
3791 3828 g1_policy()->print_heap_transition();
3792 3829
3793 3830 if (mark_in_progress()) {
3794 3831 concurrent_mark()->update_g1_committed();
3795 3832 }
3796 3833
3797 3834 #ifdef TRACESPINNING
3798 3835 ParallelTaskTerminator::print_termination_counts();
3799 3836 #endif
3800 3837
3801 3838 gc_epilogue(false);
3802 3839 }
3803 3840
3804 3841 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) {
3805 3842 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum);
3806 3843 print_tracing_info();
3807 3844 vm_exit(-1);
3808 3845 }
3809 3846 }
3810 3847
3811 3848 _hrs.verify_optional();
3812 3849 verify_region_sets_optional();
3813 3850
3814 3851 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
3815 3852 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
3816 3853
3817 3854 if (PrintHeapAtGC) {
3818 3855 Universe::print_heap_after_gc();
3819 3856 }
3820 3857 g1mm()->update_sizes();
3821 3858
3822 3859 if (G1SummarizeRSetStats &&
3823 3860 (G1SummarizeRSetStatsPeriod > 0) &&
3824 3861 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
3825 3862 g1_rem_set()->print_summary_info();
3826 3863 }
3827 3864
3828 3865 return true;
3829 3866 }
3830 3867
3831 3868 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
3832 3869 {
3833 3870 size_t gclab_word_size;
3834 3871 switch (purpose) {
3835 3872 case GCAllocForSurvived:
3836 3873 gclab_word_size = YoungPLABSize;
3837 3874 break;
3838 3875 case GCAllocForTenured:
3839 3876 gclab_word_size = OldPLABSize;
3840 3877 break;
3841 3878 default:
3842 3879 assert(false, "unknown GCAllocPurpose");
3843 3880 gclab_word_size = OldPLABSize;
3844 3881 break;
3845 3882 }
3846 3883 return gclab_word_size;
3847 3884 }
3848 3885
3849 3886 void G1CollectedHeap::init_mutator_alloc_region() {
3850 3887 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
3851 3888 _mutator_alloc_region.init();
3852 3889 }
3853 3890
3854 3891 void G1CollectedHeap::release_mutator_alloc_region() {
3855 3892 _mutator_alloc_region.release();
3856 3893 assert(_mutator_alloc_region.get() == NULL, "post-condition");
3857 3894 }
3858 3895
3859 3896 void G1CollectedHeap::init_gc_alloc_regions() {
3860 3897 assert_at_safepoint(true /* should_be_vm_thread */);
3861 3898
3862 3899 _survivor_gc_alloc_region.init();
3863 3900 _old_gc_alloc_region.init();
3864 3901 HeapRegion* retained_region = _retained_old_gc_alloc_region;
3865 3902 _retained_old_gc_alloc_region = NULL;
3866 3903
3867 3904 // We will discard the current GC alloc region if:
3868 3905 // a) it's in the collection set (it can happen!),
3869 3906 // b) it's already full (no point in using it),
3870 3907 // c) it's empty (this means that it was emptied during
3871 3908 // a cleanup and it should be on the free list now), or
3872 3909 // d) it's humongous (this means that it was emptied
3873 3910 // during a cleanup and was added to the free list, but
3874 3911 // has been subseqently used to allocate a humongous
3875 3912 // object that may be less than the region size).
3876 3913 if (retained_region != NULL &&
3877 3914 !retained_region->in_collection_set() &&
3878 3915 !(retained_region->top() == retained_region->end()) &&
3879 3916 !retained_region->is_empty() &&
3880 3917 !retained_region->isHumongous()) {
3881 3918 retained_region->set_saved_mark();
3882 3919 // The retained region was added to the old region set when it was
3883 3920 // retired. We have to remove it now, since we don't allow regions
3884 3921 // we allocate to in the region sets. We'll re-add it later, when
3885 3922 // it's retired again.
3886 3923 _old_set.remove(retained_region);
3887 3924 _old_gc_alloc_region.set(retained_region);
3888 3925 _hr_printer.reuse(retained_region);
3889 3926 }
3890 3927 }
3891 3928
3892 3929 void G1CollectedHeap::release_gc_alloc_regions() {
3893 3930 _survivor_gc_alloc_region.release();
3894 3931 // If we have an old GC alloc region to release, we'll save it in
3895 3932 // _retained_old_gc_alloc_region. If we don't
3896 3933 // _retained_old_gc_alloc_region will become NULL. This is what we
3897 3934 // want either way so no reason to check explicitly for either
3898 3935 // condition.
3899 3936 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
3900 3937 }
3901 3938
3902 3939 void G1CollectedHeap::abandon_gc_alloc_regions() {
3903 3940 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
3904 3941 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
3905 3942 _retained_old_gc_alloc_region = NULL;
3906 3943 }
3907 3944
3908 3945 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
3909 3946 _drain_in_progress = false;
3910 3947 set_evac_failure_closure(cl);
3911 3948 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
3912 3949 }
3913 3950
3914 3951 void G1CollectedHeap::finalize_for_evac_failure() {
3915 3952 assert(_evac_failure_scan_stack != NULL &&
3916 3953 _evac_failure_scan_stack->length() == 0,
3917 3954 "Postcondition");
3918 3955 assert(!_drain_in_progress, "Postcondition");
3919 3956 delete _evac_failure_scan_stack;
3920 3957 _evac_failure_scan_stack = NULL;
3921 3958 }
3922 3959
3923 3960 class UpdateRSetDeferred : public OopsInHeapRegionClosure {
3924 3961 private:
3925 3962 G1CollectedHeap* _g1;
3926 3963 DirtyCardQueue *_dcq;
3927 3964 CardTableModRefBS* _ct_bs;
3928 3965
3929 3966 public:
3930 3967 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
3931 3968 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {}
3932 3969
3933 3970 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
3934 3971 virtual void do_oop( oop* p) { do_oop_work(p); }
3935 3972 template <class T> void do_oop_work(T* p) {
3936 3973 assert(_from->is_in_reserved(p), "paranoia");
3937 3974 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) &&
3938 3975 !_from->is_survivor()) {
3939 3976 size_t card_index = _ct_bs->index_for(p);
3940 3977 if (_ct_bs->mark_card_deferred(card_index)) {
3941 3978 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
3942 3979 }
3943 3980 }
3944 3981 }
3945 3982 };
3946 3983
3947 3984 class RemoveSelfPointerClosure: public ObjectClosure {
3948 3985 private:
3949 3986 G1CollectedHeap* _g1;
3950 3987 ConcurrentMark* _cm;
3951 3988 HeapRegion* _hr;
3952 3989 size_t _prev_marked_bytes;
3953 3990 size_t _next_marked_bytes;
3954 3991 OopsInHeapRegionClosure *_cl;
3955 3992 public:
3956 3993 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr,
3957 3994 OopsInHeapRegionClosure* cl) :
3958 3995 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0),
3959 3996 _next_marked_bytes(0), _cl(cl) {}
3960 3997
3961 3998 size_t prev_marked_bytes() { return _prev_marked_bytes; }
3962 3999 size_t next_marked_bytes() { return _next_marked_bytes; }
3963 4000
3964 4001 // <original comment>
3965 4002 // The original idea here was to coalesce evacuated and dead objects.
3966 4003 // However that caused complications with the block offset table (BOT).
3967 4004 // In particular if there were two TLABs, one of them partially refined.
3968 4005 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~|
3969 4006 // The BOT entries of the unrefined part of TLAB_2 point to the start
3970 4007 // of TLAB_2. If the last object of the TLAB_1 and the first object
3971 4008 // of TLAB_2 are coalesced, then the cards of the unrefined part
3972 4009 // would point into middle of the filler object.
3973 4010 // The current approach is to not coalesce and leave the BOT contents intact.
3974 4011 // </original comment>
3975 4012 //
3976 4013 // We now reset the BOT when we start the object iteration over the
3977 4014 // region and refine its entries for every object we come across. So
3978 4015 // the above comment is not really relevant and we should be able
3979 4016 // to coalesce dead objects if we want to.
3980 4017 void do_object(oop obj) {
3981 4018 HeapWord* obj_addr = (HeapWord*) obj;
3982 4019 assert(_hr->is_in(obj_addr), "sanity");
3983 4020 size_t obj_size = obj->size();
3984 4021 _hr->update_bot_for_object(obj_addr, obj_size);
3985 4022 if (obj->is_forwarded() && obj->forwardee() == obj) {
3986 4023 // The object failed to move.
3987 4024 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs.");
3988 4025 _cm->markPrev(obj);
3989 4026 assert(_cm->isPrevMarked(obj), "Should be marked!");
3990 4027 _prev_marked_bytes += (obj_size * HeapWordSize);
3991 4028 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) {
3992 4029 _cm->markAndGrayObjectIfNecessary(obj);
3993 4030 }
3994 4031 obj->set_mark(markOopDesc::prototype());
3995 4032 // While we were processing RSet buffers during the
3996 4033 // collection, we actually didn't scan any cards on the
3997 4034 // collection set, since we didn't want to update remebered
3998 4035 // sets with entries that point into the collection set, given
3999 4036 // that live objects fromthe collection set are about to move
4000 4037 // and such entries will be stale very soon. This change also
4001 4038 // dealt with a reliability issue which involved scanning a
4002 4039 // card in the collection set and coming across an array that
4003 4040 // was being chunked and looking malformed. The problem is
4004 4041 // that, if evacuation fails, we might have remembered set
4005 4042 // entries missing given that we skipped cards on the
4006 4043 // collection set. So, we'll recreate such entries now.
4007 4044 obj->oop_iterate(_cl);
4008 4045 assert(_cm->isPrevMarked(obj), "Should be marked!");
4009 4046 } else {
4010 4047 // The object has been either evacuated or is dead. Fill it with a
4011 4048 // dummy object.
4012 4049 MemRegion mr((HeapWord*)obj, obj_size);
4013 4050 CollectedHeap::fill_with_object(mr);
4014 4051 _cm->clearRangeBothMaps(mr);
4015 4052 }
4016 4053 }
4017 4054 };
4018 4055
4019 4056 void G1CollectedHeap::remove_self_forwarding_pointers() {
4020 4057 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set());
4021 4058 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set());
4022 4059 UpdateRSetDeferred deferred_update(_g1h, &dcq);
4023 4060 OopsInHeapRegionClosure *cl;
4024 4061 if (G1DeferredRSUpdate) {
4025 4062 cl = &deferred_update;
4026 4063 } else {
4027 4064 cl = &immediate_update;
4028 4065 }
4029 4066 HeapRegion* cur = g1_policy()->collection_set();
4030 4067 while (cur != NULL) {
4031 4068 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
4032 4069 assert(!cur->isHumongous(), "sanity");
4033 4070
4034 4071 if (cur->evacuation_failed()) {
4035 4072 assert(cur->in_collection_set(), "bad CS");
4036 4073 RemoveSelfPointerClosure rspc(_g1h, cur, cl);
4037 4074
4038 4075 // In the common case we make sure that this is done when the
4039 4076 // region is freed so that it is "ready-to-go" when it's
4040 4077 // re-allocated. However, when evacuation failure happens, a
4041 4078 // region will remain in the heap and might ultimately be added
4042 4079 // to a CSet in the future. So we have to be careful here and
4043 4080 // make sure the region's RSet is ready for parallel iteration
4044 4081 // whenever this might be required in the future.
4045 4082 cur->rem_set()->reset_for_par_iteration();
4046 4083 cur->reset_bot();
4047 4084 cl->set_region(cur);
4048 4085 cur->object_iterate(&rspc);
4049 4086
4050 4087 // A number of manipulations to make the TAMS be the current top,
4051 4088 // and the marked bytes be the ones observed in the iteration.
4052 4089 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) {
4053 4090 // The comments below are the postconditions achieved by the
4054 4091 // calls. Note especially the last such condition, which says that
4055 4092 // the count of marked bytes has been properly restored.
4056 4093 cur->note_start_of_marking(false);
4057 4094 // _next_top_at_mark_start == top, _next_marked_bytes == 0
4058 4095 cur->add_to_marked_bytes(rspc.prev_marked_bytes());
4059 4096 // _next_marked_bytes == prev_marked_bytes.
4060 4097 cur->note_end_of_marking();
4061 4098 // _prev_top_at_mark_start == top(),
4062 4099 // _prev_marked_bytes == prev_marked_bytes
4063 4100 }
4064 4101 // If there is no mark in progress, we modified the _next variables
4065 4102 // above needlessly, but harmlessly.
4066 4103 if (_g1h->mark_in_progress()) {
4067 4104 cur->note_start_of_marking(false);
4068 4105 // _next_top_at_mark_start == top, _next_marked_bytes == 0
4069 4106 // _next_marked_bytes == next_marked_bytes.
4070 4107 }
4071 4108 }
4072 4109 cur = cur->next_in_collection_set();
4073 4110 }
4074 4111 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
4075 4112
4076 4113 // Now restore saved marks, if any.
4077 4114 if (_objs_with_preserved_marks != NULL) {
4078 4115 assert(_preserved_marks_of_objs != NULL, "Both or none.");
4079 4116 guarantee(_objs_with_preserved_marks->length() ==
4080 4117 _preserved_marks_of_objs->length(), "Both or none.");
4081 4118 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
4082 4119 oop obj = _objs_with_preserved_marks->at(i);
4083 4120 markOop m = _preserved_marks_of_objs->at(i);
4084 4121 obj->set_mark(m);
4085 4122 }
4086 4123 // Delete the preserved marks growable arrays (allocated on the C heap).
4087 4124 delete _objs_with_preserved_marks;
4088 4125 delete _preserved_marks_of_objs;
4089 4126 _objs_with_preserved_marks = NULL;
4090 4127 _preserved_marks_of_objs = NULL;
4091 4128 }
4092 4129 }
4093 4130
4094 4131 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4095 4132 _evac_failure_scan_stack->push(obj);
4096 4133 }
4097 4134
4098 4135 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4099 4136 assert(_evac_failure_scan_stack != NULL, "precondition");
4100 4137
4101 4138 while (_evac_failure_scan_stack->length() > 0) {
4102 4139 oop obj = _evac_failure_scan_stack->pop();
4103 4140 _evac_failure_closure->set_region(heap_region_containing(obj));
4104 4141 obj->oop_iterate_backwards(_evac_failure_closure);
4105 4142 }
4106 4143 }
4107 4144
4108 4145 oop
4109 4146 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4110 4147 oop old,
4111 4148 bool should_mark_root) {
4112 4149 assert(obj_in_cs(old),
4113 4150 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4114 4151 (HeapWord*) old));
4115 4152 markOop m = old->mark();
4116 4153 oop forward_ptr = old->forward_to_atomic(old);
4117 4154 if (forward_ptr == NULL) {
4118 4155 // Forward-to-self succeeded.
4119 4156
4120 4157 // should_mark_root will be true when this routine is called
4121 4158 // from a root scanning closure during an initial mark pause.
4122 4159 // In this case the thread that succeeds in self-forwarding the
4123 4160 // object is also responsible for marking the object.
4124 4161 if (should_mark_root) {
4125 4162 assert(!oopDesc::is_null(old), "shouldn't be");
4126 4163 _cm->grayRoot(old);
4127 4164 }
4128 4165
4129 4166 if (_evac_failure_closure != cl) {
4130 4167 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4131 4168 assert(!_drain_in_progress,
4132 4169 "Should only be true while someone holds the lock.");
4133 4170 // Set the global evac-failure closure to the current thread's.
4134 4171 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4135 4172 set_evac_failure_closure(cl);
4136 4173 // Now do the common part.
4137 4174 handle_evacuation_failure_common(old, m);
4138 4175 // Reset to NULL.
4139 4176 set_evac_failure_closure(NULL);
4140 4177 } else {
4141 4178 // The lock is already held, and this is recursive.
4142 4179 assert(_drain_in_progress, "This should only be the recursive case.");
4143 4180 handle_evacuation_failure_common(old, m);
4144 4181 }
4145 4182 return old;
4146 4183 } else {
4147 4184 // Forward-to-self failed. Either someone else managed to allocate
4148 4185 // space for this object (old != forward_ptr) or they beat us in
4149 4186 // self-forwarding it (old == forward_ptr).
4150 4187 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4151 4188 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4152 4189 "should not be in the CSet",
4153 4190 (HeapWord*) old, (HeapWord*) forward_ptr));
4154 4191 return forward_ptr;
4155 4192 }
4156 4193 }
4157 4194
4158 4195 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4159 4196 set_evacuation_failed(true);
4160 4197
4161 4198 preserve_mark_if_necessary(old, m);
4162 4199
4163 4200 HeapRegion* r = heap_region_containing(old);
4164 4201 if (!r->evacuation_failed()) {
4165 4202 r->set_evacuation_failed(true);
4166 4203 _hr_printer.evac_failure(r);
4167 4204 }
4168 4205
4169 4206 push_on_evac_failure_scan_stack(old);
4170 4207
4171 4208 if (!_drain_in_progress) {
4172 4209 // prevent recursion in copy_to_survivor_space()
4173 4210 _drain_in_progress = true;
4174 4211 drain_evac_failure_scan_stack();
4175 4212 _drain_in_progress = false;
4176 4213 }
4177 4214 }
4178 4215
4179 4216 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4180 4217 assert(evacuation_failed(), "Oversaving!");
4181 4218 // We want to call the "for_promotion_failure" version only in the
4182 4219 // case of a promotion failure.
4183 4220 if (m->must_be_preserved_for_promotion_failure(obj)) {
4184 4221 if (_objs_with_preserved_marks == NULL) {
4185 4222 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4186 4223 _objs_with_preserved_marks =
4187 4224 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
4188 4225 _preserved_marks_of_objs =
4189 4226 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
4190 4227 }
4191 4228 _objs_with_preserved_marks->push(obj);
4192 4229 _preserved_marks_of_objs->push(m);
4193 4230 }
4194 4231 }
4195 4232
4196 4233 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4197 4234 size_t word_size) {
4198 4235 if (purpose == GCAllocForSurvived) {
4199 4236 HeapWord* result = survivor_attempt_allocation(word_size);
4200 4237 if (result != NULL) {
4201 4238 return result;
4202 4239 } else {
4203 4240 // Let's try to allocate in the old gen in case we can fit the
4204 4241 // object there.
4205 4242 return old_attempt_allocation(word_size);
4206 4243 }
4207 4244 } else {
4208 4245 assert(purpose == GCAllocForTenured, "sanity");
4209 4246 HeapWord* result = old_attempt_allocation(word_size);
4210 4247 if (result != NULL) {
4211 4248 return result;
4212 4249 } else {
4213 4250 // Let's try to allocate in the survivors in case we can fit the
4214 4251 // object there.
4215 4252 return survivor_attempt_allocation(word_size);
4216 4253 }
4217 4254 }
4218 4255
4219 4256 ShouldNotReachHere();
4220 4257 // Trying to keep some compilers happy.
4221 4258 return NULL;
4222 4259 }
4223 4260
4224 4261 #ifndef PRODUCT
4225 4262 bool GCLabBitMapClosure::do_bit(size_t offset) {
4226 4263 HeapWord* addr = _bitmap->offsetToHeapWord(offset);
4227 4264 guarantee(_cm->isMarked(oop(addr)), "it should be!");
4228 4265 return true;
4229 4266 }
4230 4267 #endif // PRODUCT
4231 4268
4232 4269 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4233 4270 ParGCAllocBuffer(gclab_word_size),
4234 4271 _should_mark_objects(false),
4235 4272 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size),
4236 4273 _retired(false)
4237 4274 {
4238 4275 //_should_mark_objects is set to true when G1ParCopyHelper needs to
4239 4276 // mark the forwarded location of an evacuated object.
4240 4277 // We set _should_mark_objects to true if marking is active, i.e. when we
4241 4278 // need to propagate a mark, or during an initial mark pause, i.e. when we
4242 4279 // need to mark objects immediately reachable by the roots.
4243 4280 if (G1CollectedHeap::heap()->mark_in_progress() ||
4244 4281 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) {
4245 4282 _should_mark_objects = true;
4246 4283 }
4247 4284 }
4248 4285
4249 4286 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num)
4250 4287 : _g1h(g1h),
4251 4288 _refs(g1h->task_queue(queue_num)),
4252 4289 _dcq(&g1h->dirty_card_queue_set()),
4253 4290 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4254 4291 _g1_rem(g1h->g1_rem_set()),
4255 4292 _hash_seed(17), _queue_num(queue_num),
4256 4293 _term_attempts(0),
4257 4294 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4258 4295 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4259 4296 _age_table(false),
4260 4297 _strong_roots_time(0), _term_time(0),
4261 4298 _alloc_buffer_waste(0), _undo_waste(0)
4262 4299 {
4263 4300 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4264 4301 // we "sacrifice" entry 0 to keep track of surviving bytes for
4265 4302 // non-young regions (where the age is -1)
4266 4303 // We also add a few elements at the beginning and at the end in
4267 4304 // an attempt to eliminate cache contention
4268 4305 size_t real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4269 4306 size_t array_length = PADDING_ELEM_NUM +
4270 4307 real_length +
4271 4308 PADDING_ELEM_NUM;
4272 4309 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length);
4273 4310 if (_surviving_young_words_base == NULL)
4274 4311 vm_exit_out_of_memory(array_length * sizeof(size_t),
4275 4312 "Not enough space for young surv histo.");
4276 4313 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4277 4314 memset(_surviving_young_words, 0, real_length * sizeof(size_t));
4278 4315
4279 4316 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4280 4317 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4281 4318
4282 4319 _start = os::elapsedTime();
4283 4320 }
4284 4321
4285 4322 void
4286 4323 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4287 4324 {
4288 4325 st->print_raw_cr("GC Termination Stats");
4289 4326 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4290 4327 " ------waste (KiB)------");
4291 4328 st->print_raw_cr("thr ms ms % ms % attempts"
4292 4329 " total alloc undo");
4293 4330 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4294 4331 " ------- ------- -------");
4295 4332 }
4296 4333
4297 4334 void
4298 4335 G1ParScanThreadState::print_termination_stats(int i,
4299 4336 outputStream* const st) const
4300 4337 {
4301 4338 const double elapsed_ms = elapsed_time() * 1000.0;
4302 4339 const double s_roots_ms = strong_roots_time() * 1000.0;
4303 4340 const double term_ms = term_time() * 1000.0;
4304 4341 st->print_cr("%3d %9.2f %9.2f %6.2f "
4305 4342 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4306 4343 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4307 4344 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4308 4345 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4309 4346 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4310 4347 alloc_buffer_waste() * HeapWordSize / K,
4311 4348 undo_waste() * HeapWordSize / K);
4312 4349 }
4313 4350
4314 4351 #ifdef ASSERT
4315 4352 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4316 4353 assert(ref != NULL, "invariant");
4317 4354 assert(UseCompressedOops, "sanity");
4318 4355 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4319 4356 oop p = oopDesc::load_decode_heap_oop(ref);
4320 4357 assert(_g1h->is_in_g1_reserved(p),
4321 4358 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4322 4359 return true;
4323 4360 }
4324 4361
4325 4362 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4326 4363 assert(ref != NULL, "invariant");
4327 4364 if (has_partial_array_mask(ref)) {
4328 4365 // Must be in the collection set--it's already been copied.
4329 4366 oop p = clear_partial_array_mask(ref);
4330 4367 assert(_g1h->obj_in_cs(p),
4331 4368 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4332 4369 } else {
4333 4370 oop p = oopDesc::load_decode_heap_oop(ref);
4334 4371 assert(_g1h->is_in_g1_reserved(p),
4335 4372 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4336 4373 }
4337 4374 return true;
4338 4375 }
4339 4376
4340 4377 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4341 4378 if (ref.is_narrow()) {
4342 4379 return verify_ref((narrowOop*) ref);
4343 4380 } else {
4344 4381 return verify_ref((oop*) ref);
4345 4382 }
4346 4383 }
4347 4384 #endif // ASSERT
4348 4385
4349 4386 void G1ParScanThreadState::trim_queue() {
4350 4387 assert(_evac_cl != NULL, "not set");
4351 4388 assert(_evac_failure_cl != NULL, "not set");
4352 4389 assert(_partial_scan_cl != NULL, "not set");
4353 4390
4354 4391 StarTask ref;
4355 4392 do {
4356 4393 // Drain the overflow stack first, so other threads can steal.
4357 4394 while (refs()->pop_overflow(ref)) {
4358 4395 deal_with_reference(ref);
4359 4396 }
4360 4397
4361 4398 while (refs()->pop_local(ref)) {
4362 4399 deal_with_reference(ref);
4363 4400 }
4364 4401 } while (!refs()->is_empty());
4365 4402 }
4366 4403
4367 4404 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) :
4368 4405 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4369 4406 _par_scan_state(par_scan_state),
4370 4407 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4371 4408 _mark_in_progress(_g1->mark_in_progress()) { }
4372 4409
4373 4410 template <class T> void G1ParCopyHelper::mark_object(T* p) {
4374 4411 // This is called from do_oop_work for objects that are not
4375 4412 // in the collection set. Objects in the collection set
4376 4413 // are marked after they have been evacuated.
4377 4414
4378 4415 T heap_oop = oopDesc::load_heap_oop(p);
4379 4416 if (!oopDesc::is_null(heap_oop)) {
4380 4417 oop obj = oopDesc::decode_heap_oop(heap_oop);
4381 4418 HeapWord* addr = (HeapWord*)obj;
4382 4419 if (_g1->is_in_g1_reserved(addr)) {
4383 4420 _cm->grayRoot(oop(addr));
4384 4421 }
4385 4422 }
4386 4423 }
4387 4424
4388 4425 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root,
4389 4426 bool should_mark_copy) {
4390 4427 size_t word_sz = old->size();
4391 4428 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4392 4429 // +1 to make the -1 indexes valid...
4393 4430 int young_index = from_region->young_index_in_cset()+1;
4394 4431 assert( (from_region->is_young() && young_index > 0) ||
4395 4432 (!from_region->is_young() && young_index == 0), "invariant" );
4396 4433 G1CollectorPolicy* g1p = _g1->g1_policy();
4397 4434 markOop m = old->mark();
4398 4435 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4399 4436 : m->age();
4400 4437 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4401 4438 word_sz);
4402 4439 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4403 4440 oop obj = oop(obj_ptr);
4404 4441
4405 4442 if (obj_ptr == NULL) {
4406 4443 // This will either forward-to-self, or detect that someone else has
4407 4444 // installed a forwarding pointer.
4408 4445 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4409 4446 return _g1->handle_evacuation_failure_par(cl, old, should_mark_root);
4410 4447 }
4411 4448
4412 4449 // We're going to allocate linearly, so might as well prefetch ahead.
4413 4450 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4414 4451
4415 4452 oop forward_ptr = old->forward_to_atomic(obj);
4416 4453 if (forward_ptr == NULL) {
4417 4454 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4418 4455 if (g1p->track_object_age(alloc_purpose)) {
4419 4456 // We could simply do obj->incr_age(). However, this causes a
4420 4457 // performance issue. obj->incr_age() will first check whether
4421 4458 // the object has a displaced mark by checking its mark word;
4422 4459 // getting the mark word from the new location of the object
4423 4460 // stalls. So, given that we already have the mark word and we
4424 4461 // are about to install it anyway, it's better to increase the
4425 4462 // age on the mark word, when the object does not have a
4426 4463 // displaced mark word. We're not expecting many objects to have
4427 4464 // a displaced marked word, so that case is not optimized
4428 4465 // further (it could be...) and we simply call obj->incr_age().
4429 4466
4430 4467 if (m->has_displaced_mark_helper()) {
4431 4468 // in this case, we have to install the mark word first,
4432 4469 // otherwise obj looks to be forwarded (the old mark word,
4433 4470 // which contains the forward pointer, was copied)
4434 4471 obj->set_mark(m);
4435 4472 obj->incr_age();
4436 4473 } else {
4437 4474 m = m->incr_age();
4438 4475 obj->set_mark(m);
4439 4476 }
4440 4477 _par_scan_state->age_table()->add(obj, word_sz);
4441 4478 } else {
4442 4479 obj->set_mark(m);
4443 4480 }
4444 4481
4445 4482 // Mark the evacuated object or propagate "next" mark bit
4446 4483 if (should_mark_copy) {
4447 4484 if (!use_local_bitmaps ||
4448 4485 !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) {
4449 4486 // if we couldn't mark it on the local bitmap (this happens when
4450 4487 // the object was not allocated in the GCLab), we have to bite
4451 4488 // the bullet and do the standard parallel mark
4452 4489 _cm->markAndGrayObjectIfNecessary(obj);
4453 4490 }
4454 4491
4455 4492 if (_g1->isMarkedNext(old)) {
4456 4493 // Unmark the object's old location so that marking
4457 4494 // doesn't think the old object is alive.
4458 4495 _cm->nextMarkBitMap()->parClear((HeapWord*)old);
4459 4496 }
4460 4497 }
4461 4498
4462 4499 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4463 4500 surv_young_words[young_index] += word_sz;
4464 4501
4465 4502 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4466 4503 arrayOop(old)->set_length(0);
4467 4504 oop* old_p = set_partial_array_mask(old);
4468 4505 _par_scan_state->push_on_queue(old_p);
4469 4506 } else {
4470 4507 // No point in using the slower heap_region_containing() method,
4471 4508 // given that we know obj is in the heap.
4472 4509 _scanner->set_region(_g1->heap_region_containing_raw(obj));
4473 4510 obj->oop_iterate_backwards(_scanner);
4474 4511 }
4475 4512 } else {
4476 4513 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4477 4514 obj = forward_ptr;
4478 4515 }
4479 4516 return obj;
4480 4517 }
4481 4518
4482 4519 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4483 4520 template <class T>
4484 4521 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4485 4522 ::do_oop_work(T* p) {
4486 4523 oop obj = oopDesc::load_decode_heap_oop(p);
4487 4524 assert(barrier != G1BarrierRS || obj != NULL,
4488 4525 "Precondition: G1BarrierRS implies obj is nonNull");
4489 4526
4490 4527 // Marking:
4491 4528 // If the object is in the collection set, then the thread
4492 4529 // that copies the object should mark, or propagate the
4493 4530 // mark to, the evacuated object.
4494 4531 // If the object is not in the collection set then we
4495 4532 // should call the mark_object() method depending on the
4496 4533 // value of the template parameter do_mark_object (which will
4497 4534 // be true for root scanning closures during an initial mark
4498 4535 // pause).
4499 4536 // The mark_object() method first checks whether the object
4500 4537 // is marked and, if not, attempts to mark the object.
4501 4538
4502 4539 // here the null check is implicit in the cset_fast_test() test
4503 4540 if (_g1->in_cset_fast_test(obj)) {
4504 4541 if (obj->is_forwarded()) {
4505 4542 oopDesc::encode_store_heap_oop(p, obj->forwardee());
4506 4543 // If we are a root scanning closure during an initial
4507 4544 // mark pause (i.e. do_mark_object will be true) then
4508 4545 // we also need to handle marking of roots in the
4509 4546 // event of an evacuation failure. In the event of an
4510 4547 // evacuation failure, the object is forwarded to itself
4511 4548 // and not copied. For root-scanning closures, the
4512 4549 // object would be marked after a successful self-forward
4513 4550 // but an object could be pointed to by both a root and non
4514 4551 // root location and be self-forwarded by a non-root-scanning
4515 4552 // closure. Therefore we also have to attempt to mark the
4516 4553 // self-forwarded root object here.
4517 4554 if (do_mark_object && obj->forwardee() == obj) {
4518 4555 mark_object(p);
4519 4556 }
4520 4557 } else {
4521 4558 // During an initial mark pause, objects that are pointed to
4522 4559 // by the roots need to be marked - even in the event of an
4523 4560 // evacuation failure. We pass the template parameter
4524 4561 // do_mark_object (which is true for root scanning closures
4525 4562 // during an initial mark pause) to copy_to_survivor_space
4526 4563 // which will pass it on to the evacuation failure handling
4527 4564 // code. The thread that successfully self-forwards a root
4528 4565 // object to itself is responsible for marking the object.
4529 4566 bool should_mark_root = do_mark_object;
4530 4567
4531 4568 // We need to mark the copied object if we're a root scanning
4532 4569 // closure during an initial mark pause (i.e. do_mark_object
4533 4570 // will be true), or the object is already marked and we need
4534 4571 // to propagate the mark to the evacuated copy.
4535 4572 bool should_mark_copy = do_mark_object ||
4536 4573 _during_initial_mark ||
4537 4574 (_mark_in_progress && !_g1->is_obj_ill(obj));
4538 4575
4539 4576 oop copy_oop = copy_to_survivor_space(obj, should_mark_root,
4540 4577 should_mark_copy);
4541 4578 oopDesc::encode_store_heap_oop(p, copy_oop);
4542 4579 }
4543 4580 // When scanning the RS, we only care about objs in CS.
4544 4581 if (barrier == G1BarrierRS) {
4545 4582 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4546 4583 }
4547 4584 } else {
4548 4585 // The object is not in collection set. If we're a root scanning
4549 4586 // closure during an initial mark pause (i.e. do_mark_object will
4550 4587 // be true) then attempt to mark the object.
4551 4588 if (do_mark_object) {
4552 4589 mark_object(p);
4553 4590 }
4554 4591 }
4555 4592
4556 4593 if (barrier == G1BarrierEvac && obj != NULL) {
4557 4594 _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4558 4595 }
4559 4596
4560 4597 if (do_gen_barrier && obj != NULL) {
4561 4598 par_do_barrier(p);
4562 4599 }
4563 4600 }
4564 4601
4565 4602 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4566 4603 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4567 4604
4568 4605 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4569 4606 assert(has_partial_array_mask(p), "invariant");
4570 4607 oop old = clear_partial_array_mask(p);
4571 4608 assert(old->is_objArray(), "must be obj array");
4572 4609 assert(old->is_forwarded(), "must be forwarded");
4573 4610 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
4574 4611
4575 4612 objArrayOop obj = objArrayOop(old->forwardee());
4576 4613 assert((void*)old != (void*)old->forwardee(), "self forwarding here?");
4577 4614 // Process ParGCArrayScanChunk elements now
4578 4615 // and push the remainder back onto queue
4579 4616 int start = arrayOop(old)->length();
4580 4617 int end = obj->length();
4581 4618 int remainder = end - start;
4582 4619 assert(start <= end, "just checking");
4583 4620 if (remainder > 2 * ParGCArrayScanChunk) {
4584 4621 // Test above combines last partial chunk with a full chunk
4585 4622 end = start + ParGCArrayScanChunk;
4586 4623 arrayOop(old)->set_length(end);
4587 4624 // Push remainder.
4588 4625 oop* old_p = set_partial_array_mask(old);
4589 4626 assert(arrayOop(old)->length() < obj->length(), "Empty push?");
4590 4627 _par_scan_state->push_on_queue(old_p);
4591 4628 } else {
4592 4629 // Restore length so that the heap remains parsable in
4593 4630 // case of evacuation failure.
4594 4631 arrayOop(old)->set_length(end);
4595 4632 }
4596 4633 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4597 4634 // process our set of indices (include header in first chunk)
4598 4635 obj->oop_iterate_range(&_scanner, start, end);
4599 4636 }
4600 4637
4601 4638 class G1ParEvacuateFollowersClosure : public VoidClosure {
4602 4639 protected:
4603 4640 G1CollectedHeap* _g1h;
4604 4641 G1ParScanThreadState* _par_scan_state;
4605 4642 RefToScanQueueSet* _queues;
4606 4643 ParallelTaskTerminator* _terminator;
4607 4644
4608 4645 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4609 4646 RefToScanQueueSet* queues() { return _queues; }
4610 4647 ParallelTaskTerminator* terminator() { return _terminator; }
4611 4648
4612 4649 public:
4613 4650 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4614 4651 G1ParScanThreadState* par_scan_state,
4615 4652 RefToScanQueueSet* queues,
4616 4653 ParallelTaskTerminator* terminator)
4617 4654 : _g1h(g1h), _par_scan_state(par_scan_state),
4618 4655 _queues(queues), _terminator(terminator) {}
4619 4656
4620 4657 void do_void();
4621 4658
4622 4659 private:
4623 4660 inline bool offer_termination();
4624 4661 };
4625 4662
4626 4663 bool G1ParEvacuateFollowersClosure::offer_termination() {
4627 4664 G1ParScanThreadState* const pss = par_scan_state();
4628 4665 pss->start_term_time();
4629 4666 const bool res = terminator()->offer_termination();
4630 4667 pss->end_term_time();
4631 4668 return res;
4632 4669 }
4633 4670
4634 4671 void G1ParEvacuateFollowersClosure::do_void() {
4635 4672 StarTask stolen_task;
4636 4673 G1ParScanThreadState* const pss = par_scan_state();
4637 4674 pss->trim_queue();
4638 4675
4639 4676 do {
4640 4677 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4641 4678 assert(pss->verify_task(stolen_task), "sanity");
4642 4679 if (stolen_task.is_narrow()) {
4643 4680 pss->deal_with_reference((narrowOop*) stolen_task);
4644 4681 } else {
4645 4682 pss->deal_with_reference((oop*) stolen_task);
4646 4683 }
4647 4684
4648 4685 // We've just processed a reference and we might have made
4649 4686 // available new entries on the queues. So we have to make sure
4650 4687 // we drain the queues as necessary.
4651 4688 pss->trim_queue();
4652 4689 }
4653 4690 } while (!offer_termination());
4654 4691
4655 4692 pss->retire_alloc_buffers();
4656 4693 }
4657 4694
4658 4695 class G1ParTask : public AbstractGangTask {
4659 4696 protected:
4660 4697 G1CollectedHeap* _g1h;
4661 4698 RefToScanQueueSet *_queues;
4662 4699 ParallelTaskTerminator _terminator;
4663 4700 int _n_workers;
4664 4701
4665 4702 Mutex _stats_lock;
4666 4703 Mutex* stats_lock() { return &_stats_lock; }
4667 4704
4668 4705 size_t getNCards() {
4669 4706 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4670 4707 / G1BlockOffsetSharedArray::N_bytes;
4671 4708 }
4672 4709
4673 4710 public:
4674 4711 G1ParTask(G1CollectedHeap* g1h,
4675 4712 RefToScanQueueSet *task_queues)
4676 4713 : AbstractGangTask("G1 collection"),
4677 4714 _g1h(g1h),
4678 4715 _queues(task_queues),
4679 4716 _terminator(0, _queues),
4680 4717 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4681 4718 {}
4682 4719
4683 4720 RefToScanQueueSet* queues() { return _queues; }
4684 4721
4685 4722 RefToScanQueue *work_queue(int i) {
4686 4723 return queues()->queue(i);
4687 4724 }
4688 4725
4689 4726 ParallelTaskTerminator* terminator() { return &_terminator; }
4690 4727
4691 4728 virtual void set_for_termination(int active_workers) {
4692 4729 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4693 4730 // in the young space (_par_seq_tasks) in the G1 heap
4694 4731 // for SequentialSubTasksDone.
4695 4732 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4696 4733 // both of which need setting by set_n_termination().
4697 4734 _g1h->SharedHeap::set_n_termination(active_workers);
4698 4735 _g1h->set_n_termination(active_workers);
4699 4736 terminator()->reset_for_reuse(active_workers);
4700 4737 _n_workers = active_workers;
4701 4738 }
4702 4739
4703 4740 void work(int i) {
4704 4741 if (i >= _n_workers) return; // no work needed this round
4705 4742
4706 4743 double start_time_ms = os::elapsedTime() * 1000.0;
4707 4744 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms);
4708 4745
4709 4746 ResourceMark rm;
4710 4747 HandleMark hm;
4711 4748
4712 4749 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4713 4750
4714 4751 G1ParScanThreadState pss(_g1h, i);
4715 4752 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4716 4753 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4717 4754 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4718 4755
4719 4756 pss.set_evac_closure(&scan_evac_cl);
4720 4757 pss.set_evac_failure_closure(&evac_failure_cl);
4721 4758 pss.set_partial_scan_closure(&partial_scan_cl);
4722 4759
4723 4760 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4724 4761 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp);
4725 4762
4726 4763 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4727 4764 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp);
4728 4765
4729 4766 OopClosure* scan_root_cl = &only_scan_root_cl;
4730 4767 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl;
4731 4768
4732 4769 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4733 4770 // We also need to mark copied objects.
4734 4771 scan_root_cl = &scan_mark_root_cl;
4735 4772 scan_perm_cl = &scan_mark_perm_cl;
4736 4773 }
4737 4774
4738 4775 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4739 4776
4740 4777 pss.start_strong_roots();
4741 4778 _g1h->g1_process_strong_roots(/* not collecting perm */ false,
4742 4779 SharedHeap::SO_AllClasses,
4743 4780 scan_root_cl,
4744 4781 &push_heap_rs_cl,
4745 4782 scan_perm_cl,
4746 4783 i);
4747 4784 pss.end_strong_roots();
4748 4785
4749 4786 {
4750 4787 double start = os::elapsedTime();
4751 4788 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4752 4789 evac.do_void();
4753 4790 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4754 4791 double term_ms = pss.term_time()*1000.0;
4755 4792 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms);
4756 4793 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts());
4757 4794 }
4758 4795 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4759 4796 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4760 4797
4761 4798 // Clean up any par-expanded rem sets.
4762 4799 HeapRegionRemSet::par_cleanup();
4763 4800
4764 4801 if (ParallelGCVerbose) {
4765 4802 MutexLocker x(stats_lock());
4766 4803 pss.print_termination_stats(i);
4767 4804 }
4768 4805
4769 4806 assert(pss.refs()->is_empty(), "should be empty");
4770 4807 double end_time_ms = os::elapsedTime() * 1000.0;
4771 4808 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms);
4772 4809 }
4773 4810 };
4774 4811
4775 4812 // *** Common G1 Evacuation Stuff
4776 4813
4777 4814 // This method is run in a GC worker.
4778 4815
4779 4816 void
4780 4817 G1CollectedHeap::
4781 4818 g1_process_strong_roots(bool collecting_perm_gen,
4782 4819 SharedHeap::ScanningOption so,
4783 4820 OopClosure* scan_non_heap_roots,
4784 4821 OopsInHeapRegionClosure* scan_rs,
4785 4822 OopsInGenClosure* scan_perm,
4786 4823 int worker_i) {
4787 4824
4788 4825 // First scan the strong roots, including the perm gen.
4789 4826 double ext_roots_start = os::elapsedTime();
4790 4827 double closure_app_time_sec = 0.0;
4791 4828
4792 4829 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4793 4830 BufferingOopsInGenClosure buf_scan_perm(scan_perm);
4794 4831 buf_scan_perm.set_generation(perm_gen());
4795 4832
4796 4833 // Walk the code cache w/o buffering, because StarTask cannot handle
4797 4834 // unaligned oop locations.
4798 4835 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true);
4799 4836
4800 4837 process_strong_roots(false, // no scoping; this is parallel code
4801 4838 collecting_perm_gen, so,
4802 4839 &buf_scan_non_heap_roots,
4803 4840 &eager_scan_code_roots,
4804 4841 &buf_scan_perm);
4805 4842
4806 4843 // Now the CM ref_processor roots.
4807 4844 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
4808 4845 // We need to treat the discovered reference lists of the
4809 4846 // concurrent mark ref processor as roots and keep entries
4810 4847 // (which are added by the marking threads) on them live
4811 4848 // until they can be processed at the end of marking.
4812 4849 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
4813 4850 }
4814 4851
4815 4852 // Finish up any enqueued closure apps (attributed as object copy time).
4816 4853 buf_scan_non_heap_roots.done();
4817 4854 buf_scan_perm.done();
4818 4855
4819 4856 double ext_roots_end = os::elapsedTime();
4820 4857
4821 4858 g1_policy()->reset_obj_copy_time(worker_i);
4822 4859 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() +
4823 4860 buf_scan_non_heap_roots.closure_app_seconds();
4824 4861 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
4825 4862
4826 4863 double ext_root_time_ms =
4827 4864 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0;
4828 4865
4829 4866 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
4830 4867
4831 4868 // Scan strong roots in mark stack.
4832 4869 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) {
4833 4870 concurrent_mark()->oops_do(scan_non_heap_roots);
4834 4871 }
4835 4872 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0;
4836 4873 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms);
4837 4874
4838 4875 // Now scan the complement of the collection set.
4839 4876 if (scan_rs != NULL) {
4840 4877 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
4841 4878 }
4842 4879
4843 4880 _process_strong_tasks->all_tasks_completed();
4844 4881 }
4845 4882
4846 4883 void
4847 4884 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
4848 4885 OopClosure* non_root_closure) {
4849 4886 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
4850 4887 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
4851 4888 }
4852 4889
4853 4890 // Weak Reference Processing support
4854 4891
4855 4892 // An always "is_alive" closure that is used to preserve referents.
4856 4893 // If the object is non-null then it's alive. Used in the preservation
4857 4894 // of referent objects that are pointed to by reference objects
4858 4895 // discovered by the CM ref processor.
4859 4896 class G1AlwaysAliveClosure: public BoolObjectClosure {
4860 4897 G1CollectedHeap* _g1;
4861 4898 public:
4862 4899 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4863 4900 void do_object(oop p) { assert(false, "Do not call."); }
4864 4901 bool do_object_b(oop p) {
4865 4902 if (p != NULL) {
4866 4903 return true;
4867 4904 }
4868 4905 return false;
4869 4906 }
4870 4907 };
4871 4908
4872 4909 bool G1STWIsAliveClosure::do_object_b(oop p) {
4873 4910 // An object is reachable if it is outside the collection set,
4874 4911 // or is inside and copied.
4875 4912 return !_g1->obj_in_cs(p) || p->is_forwarded();
4876 4913 }
4877 4914
4878 4915 // Non Copying Keep Alive closure
4879 4916 class G1KeepAliveClosure: public OopClosure {
4880 4917 G1CollectedHeap* _g1;
4881 4918 public:
4882 4919 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4883 4920 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
4884 4921 void do_oop( oop* p) {
4885 4922 oop obj = *p;
4886 4923
4887 4924 if (_g1->obj_in_cs(obj)) {
4888 4925 assert( obj->is_forwarded(), "invariant" );
4889 4926 *p = obj->forwardee();
4890 4927 }
4891 4928 }
4892 4929 };
4893 4930
4894 4931 // Copying Keep Alive closure - can be called from both
4895 4932 // serial and parallel code as long as different worker
4896 4933 // threads utilize different G1ParScanThreadState instances
4897 4934 // and different queues.
4898 4935
4899 4936 class G1CopyingKeepAliveClosure: public OopClosure {
4900 4937 G1CollectedHeap* _g1h;
4901 4938 OopClosure* _copy_non_heap_obj_cl;
4902 4939 OopsInHeapRegionClosure* _copy_perm_obj_cl;
4903 4940 G1ParScanThreadState* _par_scan_state;
4904 4941
4905 4942 public:
4906 4943 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
4907 4944 OopClosure* non_heap_obj_cl,
4908 4945 OopsInHeapRegionClosure* perm_obj_cl,
4909 4946 G1ParScanThreadState* pss):
4910 4947 _g1h(g1h),
4911 4948 _copy_non_heap_obj_cl(non_heap_obj_cl),
4912 4949 _copy_perm_obj_cl(perm_obj_cl),
4913 4950 _par_scan_state(pss)
4914 4951 {}
4915 4952
4916 4953 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
4917 4954 virtual void do_oop( oop* p) { do_oop_work(p); }
4918 4955
4919 4956 template <class T> void do_oop_work(T* p) {
4920 4957 oop obj = oopDesc::load_decode_heap_oop(p);
4921 4958
4922 4959 if (_g1h->obj_in_cs(obj)) {
4923 4960 // If the referent object has been forwarded (either copied
4924 4961 // to a new location or to itself in the event of an
4925 4962 // evacuation failure) then we need to update the reference
4926 4963 // field and, if both reference and referent are in the G1
4927 4964 // heap, update the RSet for the referent.
4928 4965 //
4929 4966 // If the referent has not been forwarded then we have to keep
4930 4967 // it alive by policy. Therefore we have copy the referent.
4931 4968 //
4932 4969 // If the reference field is in the G1 heap then we can push
4933 4970 // on the PSS queue. When the queue is drained (after each
4934 4971 // phase of reference processing) the object and it's followers
4935 4972 // will be copied, the reference field set to point to the
4936 4973 // new location, and the RSet updated. Otherwise we need to
4937 4974 // use the the non-heap or perm closures directly to copy
4938 4975 // the refernt object and update the pointer, while avoiding
4939 4976 // updating the RSet.
4940 4977
4941 4978 if (_g1h->is_in_g1_reserved(p)) {
4942 4979 _par_scan_state->push_on_queue(p);
4943 4980 } else {
4944 4981 // The reference field is not in the G1 heap.
4945 4982 if (_g1h->perm_gen()->is_in(p)) {
4946 4983 _copy_perm_obj_cl->do_oop(p);
4947 4984 } else {
4948 4985 _copy_non_heap_obj_cl->do_oop(p);
4949 4986 }
4950 4987 }
4951 4988 }
4952 4989 }
4953 4990 };
4954 4991
4955 4992 // Serial drain queue closure. Called as the 'complete_gc'
4956 4993 // closure for each discovered list in some of the
4957 4994 // reference processing phases.
4958 4995
4959 4996 class G1STWDrainQueueClosure: public VoidClosure {
4960 4997 protected:
4961 4998 G1CollectedHeap* _g1h;
4962 4999 G1ParScanThreadState* _par_scan_state;
4963 5000
4964 5001 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4965 5002
4966 5003 public:
4967 5004 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
4968 5005 _g1h(g1h),
4969 5006 _par_scan_state(pss)
4970 5007 { }
4971 5008
4972 5009 void do_void() {
4973 5010 G1ParScanThreadState* const pss = par_scan_state();
4974 5011 pss->trim_queue();
4975 5012 }
4976 5013 };
4977 5014
4978 5015 // Parallel Reference Processing closures
4979 5016
4980 5017 // Implementation of AbstractRefProcTaskExecutor for parallel reference
4981 5018 // processing during G1 evacuation pauses.
4982 5019
4983 5020 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
4984 5021 private:
4985 5022 G1CollectedHeap* _g1h;
4986 5023 RefToScanQueueSet* _queues;
4987 5024 FlexibleWorkGang* _workers;
4988 5025 int _active_workers;
4989 5026
4990 5027 public:
4991 5028 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
4992 5029 FlexibleWorkGang* workers,
4993 5030 RefToScanQueueSet *task_queues,
4994 5031 int n_workers) :
4995 5032 _g1h(g1h),
4996 5033 _queues(task_queues),
4997 5034 _workers(workers),
4998 5035 _active_workers(n_workers)
4999 5036 {
5000 5037 assert(n_workers > 0, "shouldn't call this otherwise");
5001 5038 }
5002 5039
5003 5040 // Executes the given task using concurrent marking worker threads.
5004 5041 virtual void execute(ProcessTask& task);
5005 5042 virtual void execute(EnqueueTask& task);
5006 5043 };
5007 5044
5008 5045 // Gang task for possibly parallel reference processing
5009 5046
5010 5047 class G1STWRefProcTaskProxy: public AbstractGangTask {
5011 5048 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5012 5049 ProcessTask& _proc_task;
5013 5050 G1CollectedHeap* _g1h;
5014 5051 RefToScanQueueSet *_task_queues;
5015 5052 ParallelTaskTerminator* _terminator;
5016 5053
5017 5054 public:
5018 5055 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5019 5056 G1CollectedHeap* g1h,
5020 5057 RefToScanQueueSet *task_queues,
5021 5058 ParallelTaskTerminator* terminator) :
5022 5059 AbstractGangTask("Process reference objects in parallel"),
5023 5060 _proc_task(proc_task),
5024 5061 _g1h(g1h),
5025 5062 _task_queues(task_queues),
5026 5063 _terminator(terminator)
5027 5064 {}
5028 5065
5029 5066 virtual void work(int i) {
5030 5067 // The reference processing task executed by a single worker.
5031 5068 ResourceMark rm;
5032 5069 HandleMark hm;
5033 5070
5034 5071 G1STWIsAliveClosure is_alive(_g1h);
5035 5072
5036 5073 G1ParScanThreadState pss(_g1h, i);
5037 5074
5038 5075 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5039 5076 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5040 5077 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5041 5078
5042 5079 pss.set_evac_closure(&scan_evac_cl);
5043 5080 pss.set_evac_failure_closure(&evac_failure_cl);
5044 5081 pss.set_partial_scan_closure(&partial_scan_cl);
5045 5082
5046 5083 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5047 5084 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5048 5085
5049 5086 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5050 5087 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5051 5088
5052 5089 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5053 5090 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5054 5091
5055 5092 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5056 5093 // We also need to mark copied objects.
5057 5094 copy_non_heap_cl = ©_mark_non_heap_cl;
5058 5095 copy_perm_cl = ©_mark_perm_cl;
5059 5096 }
5060 5097
5061 5098 // Keep alive closure.
5062 5099 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5063 5100
5064 5101 // Complete GC closure
5065 5102 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5066 5103
5067 5104 // Call the reference processing task's work routine.
5068 5105 _proc_task.work(i, is_alive, keep_alive, drain_queue);
5069 5106
5070 5107 // Note we cannot assert that the refs array is empty here as not all
5071 5108 // of the processing tasks (specifically phase2 - pp2_work) execute
5072 5109 // the complete_gc closure (which ordinarily would drain the queue) so
5073 5110 // the queue may not be empty.
5074 5111 }
5075 5112 };
5076 5113
5077 5114 // Driver routine for parallel reference processing.
5078 5115 // Creates an instance of the ref processing gang
5079 5116 // task and has the worker threads execute it.
5080 5117 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5081 5118 assert(_workers != NULL, "Need parallel worker threads.");
5082 5119
5083 5120 ParallelTaskTerminator terminator(_active_workers, _queues);
5084 5121 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5085 5122
5086 5123 _g1h->set_par_threads(_active_workers);
5087 5124 _workers->run_task(&proc_task_proxy);
5088 5125 _g1h->set_par_threads(0);
5089 5126 }
5090 5127
5091 5128 // Gang task for parallel reference enqueueing.
5092 5129
5093 5130 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5094 5131 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5095 5132 EnqueueTask& _enq_task;
5096 5133
5097 5134 public:
5098 5135 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5099 5136 AbstractGangTask("Enqueue reference objects in parallel"),
5100 5137 _enq_task(enq_task)
5101 5138 { }
5102 5139
5103 5140 virtual void work(int i) {
5104 5141 _enq_task.work(i);
5105 5142 }
5106 5143 };
5107 5144
5108 5145 // Driver routine for parallel reference enqueing.
5109 5146 // Creates an instance of the ref enqueueing gang
5110 5147 // task and has the worker threads execute it.
5111 5148
5112 5149 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5113 5150 assert(_workers != NULL, "Need parallel worker threads.");
5114 5151
5115 5152 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5116 5153
5117 5154 _g1h->set_par_threads(_active_workers);
5118 5155 _workers->run_task(&enq_task_proxy);
5119 5156 _g1h->set_par_threads(0);
5120 5157 }
5121 5158
5122 5159 // End of weak reference support closures
5123 5160
5124 5161 // Abstract task used to preserve (i.e. copy) any referent objects
5125 5162 // that are in the collection set and are pointed to by reference
5126 5163 // objects discovered by the CM ref processor.
5127 5164
5128 5165 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5129 5166 protected:
5130 5167 G1CollectedHeap* _g1h;
5131 5168 RefToScanQueueSet *_queues;
5132 5169 ParallelTaskTerminator _terminator;
5133 5170 int _n_workers;
5134 5171
5135 5172 public:
5136 5173 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5137 5174 AbstractGangTask("ParPreserveCMReferents"),
5138 5175 _g1h(g1h),
5139 5176 _queues(task_queues),
5140 5177 _terminator(workers, _queues),
5141 5178 _n_workers(workers)
5142 5179 { }
5143 5180
5144 5181 void work(int i) {
5145 5182 ResourceMark rm;
5146 5183 HandleMark hm;
5147 5184
5148 5185 G1ParScanThreadState pss(_g1h, i);
5149 5186 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5150 5187 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5151 5188 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5152 5189
5153 5190 pss.set_evac_closure(&scan_evac_cl);
5154 5191 pss.set_evac_failure_closure(&evac_failure_cl);
5155 5192 pss.set_partial_scan_closure(&partial_scan_cl);
5156 5193
5157 5194 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5158 5195
5159 5196
5160 5197 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5161 5198 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5162 5199
5163 5200 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5164 5201 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5165 5202
5166 5203 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5167 5204 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5168 5205
5169 5206 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5170 5207 // We also need to mark copied objects.
5171 5208 copy_non_heap_cl = ©_mark_non_heap_cl;
5172 5209 copy_perm_cl = ©_mark_perm_cl;
5173 5210 }
5174 5211
5175 5212 // Is alive closure
5176 5213 G1AlwaysAliveClosure always_alive(_g1h);
5177 5214
5178 5215 // Copying keep alive closure. Applied to referent objects that need
5179 5216 // to be copied.
5180 5217 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5181 5218
5182 5219 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5183 5220
5184 5221 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5185 5222 int stride = MIN2(MAX2(_n_workers, 1), limit);
5186 5223
5187 5224 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5188 5225 // So this must be true - but assert just in case someone decides to
5189 5226 // change the worker ids.
5190 5227 assert(0 <= i && i < limit, "sanity");
5191 5228 assert(!rp->discovery_is_atomic(), "check this code");
5192 5229
5193 5230 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5194 5231 for (int idx = i; idx < limit; idx += stride) {
5195 5232 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5196 5233
5197 5234 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5198 5235 while (iter.has_next()) {
5199 5236 // Since discovery is not atomic for the CM ref processor, we
5200 5237 // can see some null referent objects.
5201 5238 iter.load_ptrs(DEBUG_ONLY(true));
5202 5239 oop ref = iter.obj();
5203 5240
5204 5241 // This will filter nulls.
5205 5242 if (iter.is_referent_alive()) {
5206 5243 iter.make_referent_alive();
5207 5244 }
5208 5245 iter.move_to_next();
5209 5246 }
5210 5247 }
5211 5248
5212 5249 // Drain the queue - which may cause stealing
5213 5250 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5214 5251 drain_queue.do_void();
5215 5252 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5216 5253 assert(pss.refs()->is_empty(), "should be");
5217 5254 }
5218 5255 };
5219 5256
5220 5257 // Weak Reference processing during an evacuation pause (part 1).
5221 5258 void G1CollectedHeap::process_discovered_references() {
5222 5259 double ref_proc_start = os::elapsedTime();
5223 5260
5224 5261 ReferenceProcessor* rp = _ref_processor_stw;
5225 5262 assert(rp->discovery_enabled(), "should have been enabled");
5226 5263
5227 5264 // Any reference objects, in the collection set, that were 'discovered'
5228 5265 // by the CM ref processor should have already been copied (either by
5229 5266 // applying the external root copy closure to the discovered lists, or
5230 5267 // by following an RSet entry).
5231 5268 //
5232 5269 // But some of the referents, that are in the collection set, that these
5233 5270 // reference objects point to may not have been copied: the STW ref
5234 5271 // processor would have seen that the reference object had already
5235 5272 // been 'discovered' and would have skipped discovering the reference,
5236 5273 // but would not have treated the reference object as a regular oop.
5237 5274 // As a reult the copy closure would not have been applied to the
5238 5275 // referent object.
5239 5276 //
5240 5277 // We need to explicitly copy these referent objects - the references
5241 5278 // will be processed at the end of remarking.
↓ open down ↓ |
1696 lines elided |
↑ open up ↑ |
5242 5279 //
5243 5280 // We also need to do this copying before we process the reference
5244 5281 // objects discovered by the STW ref processor in case one of these
5245 5282 // referents points to another object which is also referenced by an
5246 5283 // object discovered by the STW ref processor.
5247 5284
5248 5285 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5249 5286 workers()->active_workers() : 1);
5250 5287
5251 5288 assert(active_workers == workers()->active_workers(),
5252 - "Need to reset active_workers");
5289 + "Need to reset active_workers");
5253 5290 set_par_threads(active_workers);
5254 5291 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues);
5255 5292
5256 5293 if (G1CollectedHeap::use_parallel_gc_threads()) {
5257 5294 workers()->run_task(&keep_cm_referents);
5258 5295 } else {
5259 5296 keep_cm_referents.work(0);
5260 5297 }
5261 5298
5262 5299 set_par_threads(0);
5263 5300
5264 5301 // Closure to test whether a referent is alive.
5265 5302 G1STWIsAliveClosure is_alive(this);
5266 5303
5267 5304 // Even when parallel reference processing is enabled, the processing
5268 5305 // of JNI refs is serial and performed serially by the current thread
5269 5306 // rather than by a worker. The following PSS will be used for processing
5270 5307 // JNI refs.
5271 5308
5272 5309 // Use only a single queue for this PSS.
5273 5310 G1ParScanThreadState pss(this, 0);
5274 5311
5275 5312 // We do not embed a reference processor in the copying/scanning
5276 5313 // closures while we're actually processing the discovered
5277 5314 // reference objects.
5278 5315 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5279 5316 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5280 5317 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5281 5318
5282 5319 pss.set_evac_closure(&scan_evac_cl);
5283 5320 pss.set_evac_failure_closure(&evac_failure_cl);
5284 5321 pss.set_partial_scan_closure(&partial_scan_cl);
5285 5322
5286 5323 assert(pss.refs()->is_empty(), "pre-condition");
5287 5324
5288 5325 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5289 5326 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL);
5290 5327
5291 5328 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5292 5329 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL);
5293 5330
5294 5331 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5295 5332 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5296 5333
5297 5334 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5298 5335 // We also need to mark copied objects.
5299 5336 copy_non_heap_cl = ©_mark_non_heap_cl;
5300 5337 copy_perm_cl = ©_mark_perm_cl;
5301 5338 }
5302 5339
5303 5340 // Keep alive closure.
5304 5341 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss);
5305 5342
5306 5343 // Serial Complete GC closure
5307 5344 G1STWDrainQueueClosure drain_queue(this, &pss);
5308 5345
5309 5346 // Setup the soft refs policy...
5310 5347 rp->setup_policy(false);
5311 5348
5312 5349 if (!rp->processing_is_mt()) {
5313 5350 // Serial reference processing...
5314 5351 rp->process_discovered_references(&is_alive,
5315 5352 &keep_alive,
5316 5353 &drain_queue,
5317 5354 NULL);
5318 5355 } else {
5319 5356 // Parallel reference processing
5320 5357 assert(rp->num_q() == active_workers, "sanity");
5321 5358 assert(active_workers <= rp->max_num_q(), "sanity");
5322 5359
5323 5360 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5324 5361 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5325 5362 }
5326 5363
5327 5364 // We have completed copying any necessary live referent objects
5328 5365 // (that were not copied during the actual pause) so we can
5329 5366 // retire any active alloc buffers
5330 5367 pss.retire_alloc_buffers();
5331 5368 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5332 5369
5333 5370 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5334 5371 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0);
5335 5372 }
5336 5373
5337 5374 // Weak Reference processing during an evacuation pause (part 2).
5338 5375 void G1CollectedHeap::enqueue_discovered_references() {
5339 5376 double ref_enq_start = os::elapsedTime();
5340 5377
5341 5378 ReferenceProcessor* rp = _ref_processor_stw;
5342 5379 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5343 5380
5344 5381 // Now enqueue any remaining on the discovered lists on to
5345 5382 // the pending list.
5346 5383 if (!rp->processing_is_mt()) {
5347 5384 // Serial reference processing...
5348 5385 rp->enqueue_discovered_references();
5349 5386 } else {
5350 5387 // Parallel reference enqueuing
5351 5388
5352 5389 int active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1);
5353 5390 assert(active_workers == workers()->active_workers(),
5354 5391 "Need to reset active_workers");
5355 5392 assert(rp->num_q() == active_workers, "sanity");
5356 5393 assert(active_workers <= rp->max_num_q(), "sanity");
5357 5394
5358 5395 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5359 5396 rp->enqueue_discovered_references(&par_task_executor);
5360 5397 }
5361 5398
5362 5399 rp->verify_no_references_recorded();
5363 5400 assert(!rp->discovery_enabled(), "should have been disabled");
5364 5401
5365 5402 // FIXME
5366 5403 // CM's reference processing also cleans up the string and symbol tables.
5367 5404 // Should we do that here also? We could, but it is a serial operation
5368 5405 // and could signicantly increase the pause time.
5369 5406
5370 5407 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5371 5408 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0);
5372 5409 }
5373 5410
5374 5411 void G1CollectedHeap::evacuate_collection_set() {
5375 5412 set_evacuation_failed(false);
5376 5413
5377 5414 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5378 5415 concurrent_g1_refine()->set_use_cache(false);
5379 5416 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5380 5417
5381 5418 int n_workers;
5382 5419 if (G1CollectedHeap::use_parallel_gc_threads()) {
5383 5420 n_workers =
5384 5421 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5385 5422 workers()->active_workers(),
5386 5423 Threads::number_of_non_daemon_threads());
5387 5424 assert(UseDynamicNumberOfGCThreads ||
5388 5425 n_workers == workers()->total_workers(),
5389 5426 "If not dynamic should be using all the workers");
5390 5427 set_par_threads(n_workers);
5391 5428 } else {
5392 5429 assert(n_par_threads() == 0,
5393 5430 "Should be the original non-parallel value");
5394 5431 n_workers = 1;
5395 5432 }
5396 5433 workers()->set_active_workers(n_workers);
5397 5434
5398 5435 G1ParTask g1_par_task(this, _task_queues);
5399 5436
5400 5437 init_for_evac_failure(NULL);
5401 5438
5402 5439 rem_set()->prepare_for_younger_refs_iterate(true);
5403 5440
5404 5441 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5405 5442 double start_par = os::elapsedTime();
5406 5443
5407 5444 if (G1CollectedHeap::use_parallel_gc_threads()) {
5408 5445 // The individual threads will set their evac-failure closures.
5409 5446 StrongRootsScope srs(this);
5410 5447 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5411 5448 // These tasks use ShareHeap::_process_strong_tasks
5412 5449 assert(UseDynamicNumberOfGCThreads ||
5413 5450 workers()->active_workers() == workers()->total_workers(),
5414 5451 "If not dynamic should be using all the workers");
5415 5452 workers()->run_task(&g1_par_task);
5416 5453 } else {
5417 5454 StrongRootsScope srs(this);
5418 5455 g1_par_task.work(0);
5419 5456 }
5420 5457
5421 5458 double par_time = (os::elapsedTime() - start_par) * 1000.0;
5422 5459 g1_policy()->record_par_time(par_time);
5423 5460
5424 5461 set_par_threads(0);
5425 5462
5426 5463 // Process any discovered reference objects - we have
5427 5464 // to do this _before_ we retire the GC alloc regions
5428 5465 // as we may have to copy some 'reachable' referent
5429 5466 // objects (and their reachable sub-graphs) that were
5430 5467 // not copied during the pause.
5431 5468 process_discovered_references();
5432 5469
5433 5470 // Weak root processing.
5434 5471 // Note: when JSR 292 is enabled and code blobs can contain
5435 5472 // non-perm oops then we will need to process the code blobs
5436 5473 // here too.
5437 5474 {
5438 5475 G1STWIsAliveClosure is_alive(this);
5439 5476 G1KeepAliveClosure keep_alive(this);
5440 5477 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5441 5478 }
5442 5479
5443 5480 release_gc_alloc_regions();
5444 5481 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5445 5482
5446 5483 concurrent_g1_refine()->clear_hot_cache();
5447 5484 concurrent_g1_refine()->set_use_cache(true);
5448 5485
5449 5486 finalize_for_evac_failure();
5450 5487
5451 5488 // Must do this before clearing the per-region evac-failure flags
5452 5489 // (which is currently done when we free the collection set).
5453 5490 // We also only do this if marking is actually in progress and so
5454 5491 // have to do this before we set the mark_in_progress flag at the
5455 5492 // end of an initial mark pause.
5456 5493 concurrent_mark()->complete_marking_in_collection_set();
5457 5494
5458 5495 if (evacuation_failed()) {
5459 5496 remove_self_forwarding_pointers();
5460 5497 if (PrintGCDetails) {
5461 5498 gclog_or_tty->print(" (to-space overflow)");
5462 5499 } else if (PrintGC) {
5463 5500 gclog_or_tty->print("--");
5464 5501 }
5465 5502 }
5466 5503
5467 5504 // Enqueue any remaining references remaining on the STW
5468 5505 // reference processor's discovered lists. We need to do
5469 5506 // this after the card table is cleaned (and verified) as
5470 5507 // the act of enqueuing entries on to the pending list
5471 5508 // will log these updates (and dirty their associated
5472 5509 // cards). We need these updates logged to update any
5473 5510 // RSets.
5474 5511 enqueue_discovered_references();
5475 5512
5476 5513 if (G1DeferredRSUpdate) {
5477 5514 RedirtyLoggedCardTableEntryFastClosure redirty;
5478 5515 dirty_card_queue_set().set_closure(&redirty);
5479 5516 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5480 5517
5481 5518 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5482 5519 dcq.merge_bufferlists(&dirty_card_queue_set());
5483 5520 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5484 5521 }
5485 5522 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5486 5523 }
5487 5524
5488 5525 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5489 5526 size_t* pre_used,
5490 5527 FreeRegionList* free_list,
5491 5528 OldRegionSet* old_proxy_set,
5492 5529 HumongousRegionSet* humongous_proxy_set,
5493 5530 HRRSCleanupTask* hrrs_cleanup_task,
5494 5531 bool par) {
5495 5532 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5496 5533 if (hr->isHumongous()) {
5497 5534 assert(hr->startsHumongous(), "we should only see starts humongous");
5498 5535 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5499 5536 } else {
5500 5537 _old_set.remove_with_proxy(hr, old_proxy_set);
5501 5538 free_region(hr, pre_used, free_list, par);
5502 5539 }
5503 5540 } else {
5504 5541 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5505 5542 }
5506 5543 }
5507 5544
5508 5545 void G1CollectedHeap::free_region(HeapRegion* hr,
5509 5546 size_t* pre_used,
5510 5547 FreeRegionList* free_list,
5511 5548 bool par) {
5512 5549 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5513 5550 assert(!hr->is_empty(), "the region should not be empty");
5514 5551 assert(free_list != NULL, "pre-condition");
5515 5552
5516 5553 *pre_used += hr->used();
5517 5554 hr->hr_clear(par, true /* clear_space */);
5518 5555 free_list->add_as_head(hr);
5519 5556 }
5520 5557
5521 5558 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5522 5559 size_t* pre_used,
5523 5560 FreeRegionList* free_list,
5524 5561 HumongousRegionSet* humongous_proxy_set,
5525 5562 bool par) {
5526 5563 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5527 5564 assert(free_list != NULL, "pre-condition");
5528 5565 assert(humongous_proxy_set != NULL, "pre-condition");
5529 5566
5530 5567 size_t hr_used = hr->used();
5531 5568 size_t hr_capacity = hr->capacity();
5532 5569 size_t hr_pre_used = 0;
5533 5570 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5534 5571 hr->set_notHumongous();
5535 5572 free_region(hr, &hr_pre_used, free_list, par);
5536 5573
5537 5574 size_t i = hr->hrs_index() + 1;
5538 5575 size_t num = 1;
5539 5576 while (i < n_regions()) {
5540 5577 HeapRegion* curr_hr = region_at(i);
5541 5578 if (!curr_hr->continuesHumongous()) {
5542 5579 break;
5543 5580 }
5544 5581 curr_hr->set_notHumongous();
5545 5582 free_region(curr_hr, &hr_pre_used, free_list, par);
5546 5583 num += 1;
5547 5584 i += 1;
5548 5585 }
5549 5586 assert(hr_pre_used == hr_used,
5550 5587 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5551 5588 "should be the same", hr_pre_used, hr_used));
5552 5589 *pre_used += hr_pre_used;
5553 5590 }
5554 5591
5555 5592 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5556 5593 FreeRegionList* free_list,
5557 5594 OldRegionSet* old_proxy_set,
5558 5595 HumongousRegionSet* humongous_proxy_set,
5559 5596 bool par) {
5560 5597 if (pre_used > 0) {
5561 5598 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5562 5599 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5563 5600 assert(_summary_bytes_used >= pre_used,
5564 5601 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5565 5602 "should be >= pre_used: "SIZE_FORMAT,
5566 5603 _summary_bytes_used, pre_used));
5567 5604 _summary_bytes_used -= pre_used;
5568 5605 }
5569 5606 if (free_list != NULL && !free_list->is_empty()) {
5570 5607 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5571 5608 _free_list.add_as_head(free_list);
5572 5609 }
5573 5610 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5574 5611 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5575 5612 _old_set.update_from_proxy(old_proxy_set);
5576 5613 }
5577 5614 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5578 5615 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5579 5616 _humongous_set.update_from_proxy(humongous_proxy_set);
5580 5617 }
5581 5618 }
5582 5619
5583 5620 class G1ParCleanupCTTask : public AbstractGangTask {
5584 5621 CardTableModRefBS* _ct_bs;
5585 5622 G1CollectedHeap* _g1h;
5586 5623 HeapRegion* volatile _su_head;
5587 5624 public:
5588 5625 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5589 5626 G1CollectedHeap* g1h) :
5590 5627 AbstractGangTask("G1 Par Cleanup CT Task"),
5591 5628 _ct_bs(ct_bs), _g1h(g1h) { }
5592 5629
5593 5630 void work(int i) {
5594 5631 HeapRegion* r;
5595 5632 while (r = _g1h->pop_dirty_cards_region()) {
5596 5633 clear_cards(r);
5597 5634 }
5598 5635 }
5599 5636
5600 5637 void clear_cards(HeapRegion* r) {
5601 5638 // Cards of the survivors should have already been dirtied.
5602 5639 if (!r->is_survivor()) {
5603 5640 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5604 5641 }
5605 5642 }
5606 5643 };
5607 5644
5608 5645 #ifndef PRODUCT
5609 5646 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5610 5647 G1CollectedHeap* _g1h;
5611 5648 CardTableModRefBS* _ct_bs;
5612 5649 public:
5613 5650 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5614 5651 : _g1h(g1h), _ct_bs(ct_bs) { }
5615 5652 virtual bool doHeapRegion(HeapRegion* r) {
5616 5653 if (r->is_survivor()) {
5617 5654 _g1h->verify_dirty_region(r);
5618 5655 } else {
5619 5656 _g1h->verify_not_dirty_region(r);
5620 5657 }
5621 5658 return false;
5622 5659 }
5623 5660 };
5624 5661
5625 5662 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5626 5663 // All of the region should be clean.
5627 5664 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5628 5665 MemRegion mr(hr->bottom(), hr->end());
5629 5666 ct_bs->verify_not_dirty_region(mr);
5630 5667 }
5631 5668
5632 5669 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5633 5670 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5634 5671 // dirty allocated blocks as they allocate them. The thread that
5635 5672 // retires each region and replaces it with a new one will do a
5636 5673 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5637 5674 // not dirty that area (one less thing to have to do while holding
5638 5675 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5639 5676 // is dirty.
5640 5677 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5641 5678 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5642 5679 ct_bs->verify_dirty_region(mr);
5643 5680 }
5644 5681
5645 5682 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5646 5683 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5647 5684 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5648 5685 verify_dirty_region(hr);
5649 5686 }
5650 5687 }
5651 5688
5652 5689 void G1CollectedHeap::verify_dirty_young_regions() {
5653 5690 verify_dirty_young_list(_young_list->first_region());
5654 5691 verify_dirty_young_list(_young_list->first_survivor_region());
5655 5692 }
5656 5693 #endif
5657 5694
5658 5695 void G1CollectedHeap::cleanUpCardTable() {
5659 5696 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5660 5697 double start = os::elapsedTime();
5661 5698
5662 5699 {
5663 5700 // Iterate over the dirty cards region list.
5664 5701 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5665 5702
5666 5703 if (ParallelGCThreads > 0) {
5667 5704 set_par_threads(workers()->total_workers());
5668 5705 workers()->run_task(&cleanup_task);
5669 5706 set_par_threads(0);
5670 5707 } else {
5671 5708 while (_dirty_cards_region_list) {
5672 5709 HeapRegion* r = _dirty_cards_region_list;
5673 5710 cleanup_task.clear_cards(r);
5674 5711 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5675 5712 if (_dirty_cards_region_list == r) {
5676 5713 // The last region.
5677 5714 _dirty_cards_region_list = NULL;
5678 5715 }
5679 5716 r->set_next_dirty_cards_region(NULL);
5680 5717 }
5681 5718 }
5682 5719 #ifndef PRODUCT
5683 5720 if (G1VerifyCTCleanup || VerifyAfterGC) {
5684 5721 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5685 5722 heap_region_iterate(&cleanup_verifier);
5686 5723 }
5687 5724 #endif
5688 5725 }
5689 5726
5690 5727 double elapsed = os::elapsedTime() - start;
5691 5728 g1_policy()->record_clear_ct_time(elapsed * 1000.0);
5692 5729 }
5693 5730
5694 5731 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5695 5732 size_t pre_used = 0;
5696 5733 FreeRegionList local_free_list("Local List for CSet Freeing");
5697 5734
5698 5735 double young_time_ms = 0.0;
5699 5736 double non_young_time_ms = 0.0;
5700 5737
5701 5738 // Since the collection set is a superset of the the young list,
5702 5739 // all we need to do to clear the young list is clear its
5703 5740 // head and length, and unlink any young regions in the code below
5704 5741 _young_list->clear();
5705 5742
5706 5743 G1CollectorPolicy* policy = g1_policy();
5707 5744
5708 5745 double start_sec = os::elapsedTime();
5709 5746 bool non_young = true;
5710 5747
5711 5748 HeapRegion* cur = cs_head;
5712 5749 int age_bound = -1;
5713 5750 size_t rs_lengths = 0;
5714 5751
5715 5752 while (cur != NULL) {
5716 5753 assert(!is_on_master_free_list(cur), "sanity");
5717 5754 if (non_young) {
5718 5755 if (cur->is_young()) {
5719 5756 double end_sec = os::elapsedTime();
5720 5757 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5721 5758 non_young_time_ms += elapsed_ms;
5722 5759
5723 5760 start_sec = os::elapsedTime();
5724 5761 non_young = false;
5725 5762 }
5726 5763 } else {
5727 5764 if (!cur->is_young()) {
5728 5765 double end_sec = os::elapsedTime();
5729 5766 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5730 5767 young_time_ms += elapsed_ms;
5731 5768
5732 5769 start_sec = os::elapsedTime();
5733 5770 non_young = true;
5734 5771 }
5735 5772 }
5736 5773
5737 5774 rs_lengths += cur->rem_set()->occupied();
5738 5775
5739 5776 HeapRegion* next = cur->next_in_collection_set();
5740 5777 assert(cur->in_collection_set(), "bad CS");
5741 5778 cur->set_next_in_collection_set(NULL);
5742 5779 cur->set_in_collection_set(false);
5743 5780
5744 5781 if (cur->is_young()) {
5745 5782 int index = cur->young_index_in_cset();
5746 5783 assert(index != -1, "invariant");
5747 5784 assert((size_t) index < policy->young_cset_region_length(), "invariant");
5748 5785 size_t words_survived = _surviving_young_words[index];
5749 5786 cur->record_surv_words_in_group(words_survived);
5750 5787
5751 5788 // At this point the we have 'popped' cur from the collection set
5752 5789 // (linked via next_in_collection_set()) but it is still in the
5753 5790 // young list (linked via next_young_region()). Clear the
5754 5791 // _next_young_region field.
5755 5792 cur->set_next_young_region(NULL);
5756 5793 } else {
5757 5794 int index = cur->young_index_in_cset();
5758 5795 assert(index == -1, "invariant");
5759 5796 }
5760 5797
5761 5798 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5762 5799 (!cur->is_young() && cur->young_index_in_cset() == -1),
5763 5800 "invariant" );
5764 5801
5765 5802 if (!cur->evacuation_failed()) {
5766 5803 MemRegion used_mr = cur->used_region();
5767 5804
5768 5805 // And the region is empty.
5769 5806 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5770 5807
5771 5808 // If marking is in progress then clear any objects marked in
5772 5809 // the current region. Note mark_in_progress() returns false,
5773 5810 // even during an initial mark pause, until the set_marking_started()
5774 5811 // call which takes place later in the pause.
5775 5812 if (mark_in_progress()) {
5776 5813 assert(!g1_policy()->during_initial_mark_pause(), "sanity");
5777 5814 _cm->nextMarkBitMap()->clearRange(used_mr);
5778 5815 }
5779 5816
5780 5817 free_region(cur, &pre_used, &local_free_list, false /* par */);
5781 5818 } else {
5782 5819 cur->uninstall_surv_rate_group();
5783 5820 if (cur->is_young()) {
5784 5821 cur->set_young_index_in_cset(-1);
5785 5822 }
5786 5823 cur->set_not_young();
5787 5824 cur->set_evacuation_failed(false);
5788 5825 // The region is now considered to be old.
5789 5826 _old_set.add(cur);
5790 5827 }
5791 5828 cur = next;
5792 5829 }
5793 5830
5794 5831 policy->record_max_rs_lengths(rs_lengths);
5795 5832 policy->cset_regions_freed();
5796 5833
5797 5834 double end_sec = os::elapsedTime();
5798 5835 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5799 5836
5800 5837 if (non_young) {
5801 5838 non_young_time_ms += elapsed_ms;
5802 5839 } else {
5803 5840 young_time_ms += elapsed_ms;
5804 5841 }
5805 5842
5806 5843 update_sets_after_freeing_regions(pre_used, &local_free_list,
5807 5844 NULL /* old_proxy_set */,
5808 5845 NULL /* humongous_proxy_set */,
5809 5846 false /* par */);
5810 5847 policy->record_young_free_cset_time_ms(young_time_ms);
5811 5848 policy->record_non_young_free_cset_time_ms(non_young_time_ms);
5812 5849 }
5813 5850
5814 5851 // This routine is similar to the above but does not record
5815 5852 // any policy statistics or update free lists; we are abandoning
5816 5853 // the current incremental collection set in preparation of a
5817 5854 // full collection. After the full GC we will start to build up
5818 5855 // the incremental collection set again.
5819 5856 // This is only called when we're doing a full collection
5820 5857 // and is immediately followed by the tearing down of the young list.
5821 5858
5822 5859 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
5823 5860 HeapRegion* cur = cs_head;
5824 5861
5825 5862 while (cur != NULL) {
5826 5863 HeapRegion* next = cur->next_in_collection_set();
5827 5864 assert(cur->in_collection_set(), "bad CS");
5828 5865 cur->set_next_in_collection_set(NULL);
5829 5866 cur->set_in_collection_set(false);
5830 5867 cur->set_young_index_in_cset(-1);
5831 5868 cur = next;
5832 5869 }
5833 5870 }
5834 5871
5835 5872 void G1CollectedHeap::set_free_regions_coming() {
5836 5873 if (G1ConcRegionFreeingVerbose) {
5837 5874 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5838 5875 "setting free regions coming");
5839 5876 }
5840 5877
5841 5878 assert(!free_regions_coming(), "pre-condition");
5842 5879 _free_regions_coming = true;
5843 5880 }
5844 5881
5845 5882 void G1CollectedHeap::reset_free_regions_coming() {
5846 5883 {
5847 5884 assert(free_regions_coming(), "pre-condition");
5848 5885 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5849 5886 _free_regions_coming = false;
5850 5887 SecondaryFreeList_lock->notify_all();
5851 5888 }
5852 5889
5853 5890 if (G1ConcRegionFreeingVerbose) {
5854 5891 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5855 5892 "reset free regions coming");
5856 5893 }
5857 5894 }
5858 5895
5859 5896 void G1CollectedHeap::wait_while_free_regions_coming() {
5860 5897 // Most of the time we won't have to wait, so let's do a quick test
5861 5898 // first before we take the lock.
5862 5899 if (!free_regions_coming()) {
5863 5900 return;
5864 5901 }
5865 5902
5866 5903 if (G1ConcRegionFreeingVerbose) {
5867 5904 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5868 5905 "waiting for free regions");
5869 5906 }
5870 5907
5871 5908 {
5872 5909 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5873 5910 while (free_regions_coming()) {
5874 5911 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
5875 5912 }
5876 5913 }
5877 5914
5878 5915 if (G1ConcRegionFreeingVerbose) {
5879 5916 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5880 5917 "done waiting for free regions");
5881 5918 }
5882 5919 }
5883 5920
5884 5921 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
5885 5922 assert(heap_lock_held_for_gc(),
5886 5923 "the heap lock should already be held by or for this thread");
5887 5924 _young_list->push_region(hr);
5888 5925 }
5889 5926
5890 5927 class NoYoungRegionsClosure: public HeapRegionClosure {
5891 5928 private:
5892 5929 bool _success;
5893 5930 public:
5894 5931 NoYoungRegionsClosure() : _success(true) { }
5895 5932 bool doHeapRegion(HeapRegion* r) {
5896 5933 if (r->is_young()) {
5897 5934 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
5898 5935 r->bottom(), r->end());
5899 5936 _success = false;
5900 5937 }
5901 5938 return false;
5902 5939 }
5903 5940 bool success() { return _success; }
5904 5941 };
5905 5942
5906 5943 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
5907 5944 bool ret = _young_list->check_list_empty(check_sample);
5908 5945
5909 5946 if (check_heap) {
5910 5947 NoYoungRegionsClosure closure;
5911 5948 heap_region_iterate(&closure);
5912 5949 ret = ret && closure.success();
5913 5950 }
5914 5951
5915 5952 return ret;
5916 5953 }
5917 5954
5918 5955 class TearDownRegionSetsClosure : public HeapRegionClosure {
5919 5956 private:
5920 5957 OldRegionSet *_old_set;
5921 5958
5922 5959 public:
5923 5960 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
5924 5961
5925 5962 bool doHeapRegion(HeapRegion* r) {
5926 5963 if (r->is_empty()) {
5927 5964 // We ignore empty regions, we'll empty the free list afterwards
5928 5965 } else if (r->is_young()) {
5929 5966 // We ignore young regions, we'll empty the young list afterwards
5930 5967 } else if (r->isHumongous()) {
5931 5968 // We ignore humongous regions, we're not tearing down the
5932 5969 // humongous region set
5933 5970 } else {
5934 5971 // The rest should be old
5935 5972 _old_set->remove(r);
5936 5973 }
5937 5974 return false;
5938 5975 }
5939 5976
5940 5977 ~TearDownRegionSetsClosure() {
5941 5978 assert(_old_set->is_empty(), "post-condition");
5942 5979 }
5943 5980 };
5944 5981
5945 5982 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
5946 5983 assert_at_safepoint(true /* should_be_vm_thread */);
5947 5984
5948 5985 if (!free_list_only) {
5949 5986 TearDownRegionSetsClosure cl(&_old_set);
5950 5987 heap_region_iterate(&cl);
5951 5988
5952 5989 // Need to do this after the heap iteration to be able to
5953 5990 // recognize the young regions and ignore them during the iteration.
5954 5991 _young_list->empty_list();
5955 5992 }
5956 5993 _free_list.remove_all();
5957 5994 }
5958 5995
5959 5996 class RebuildRegionSetsClosure : public HeapRegionClosure {
5960 5997 private:
5961 5998 bool _free_list_only;
5962 5999 OldRegionSet* _old_set;
5963 6000 FreeRegionList* _free_list;
5964 6001 size_t _total_used;
5965 6002
5966 6003 public:
5967 6004 RebuildRegionSetsClosure(bool free_list_only,
5968 6005 OldRegionSet* old_set, FreeRegionList* free_list) :
5969 6006 _free_list_only(free_list_only),
5970 6007 _old_set(old_set), _free_list(free_list), _total_used(0) {
5971 6008 assert(_free_list->is_empty(), "pre-condition");
5972 6009 if (!free_list_only) {
5973 6010 assert(_old_set->is_empty(), "pre-condition");
5974 6011 }
5975 6012 }
5976 6013
5977 6014 bool doHeapRegion(HeapRegion* r) {
5978 6015 if (r->continuesHumongous()) {
5979 6016 return false;
5980 6017 }
5981 6018
5982 6019 if (r->is_empty()) {
5983 6020 // Add free regions to the free list
5984 6021 _free_list->add_as_tail(r);
5985 6022 } else if (!_free_list_only) {
5986 6023 assert(!r->is_young(), "we should not come across young regions");
5987 6024
5988 6025 if (r->isHumongous()) {
5989 6026 // We ignore humongous regions, we left the humongous set unchanged
5990 6027 } else {
5991 6028 // The rest should be old, add them to the old set
5992 6029 _old_set->add(r);
5993 6030 }
5994 6031 _total_used += r->used();
5995 6032 }
5996 6033
5997 6034 return false;
5998 6035 }
5999 6036
6000 6037 size_t total_used() {
6001 6038 return _total_used;
6002 6039 }
6003 6040 };
6004 6041
6005 6042 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6006 6043 assert_at_safepoint(true /* should_be_vm_thread */);
6007 6044
6008 6045 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6009 6046 heap_region_iterate(&cl);
6010 6047
6011 6048 if (!free_list_only) {
6012 6049 _summary_bytes_used = cl.total_used();
6013 6050 }
6014 6051 assert(_summary_bytes_used == recalculate_used(),
6015 6052 err_msg("inconsistent _summary_bytes_used, "
6016 6053 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6017 6054 _summary_bytes_used, recalculate_used()));
6018 6055 }
6019 6056
6020 6057 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6021 6058 _refine_cte_cl->set_concurrent(concurrent);
6022 6059 }
6023 6060
6024 6061 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6025 6062 HeapRegion* hr = heap_region_containing(p);
6026 6063 if (hr == NULL) {
6027 6064 return is_in_permanent(p);
6028 6065 } else {
6029 6066 return hr->is_in(p);
6030 6067 }
6031 6068 }
6032 6069
6033 6070 // Methods for the mutator alloc region
6034 6071
6035 6072 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6036 6073 bool force) {
6037 6074 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6038 6075 assert(!force || g1_policy()->can_expand_young_list(),
6039 6076 "if force is true we should be able to expand the young list");
6040 6077 bool young_list_full = g1_policy()->is_young_list_full();
6041 6078 if (force || !young_list_full) {
6042 6079 HeapRegion* new_alloc_region = new_region(word_size,
6043 6080 false /* do_expand */);
6044 6081 if (new_alloc_region != NULL) {
6045 6082 set_region_short_lived_locked(new_alloc_region);
6046 6083 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6047 6084 return new_alloc_region;
6048 6085 }
6049 6086 }
6050 6087 return NULL;
6051 6088 }
6052 6089
6053 6090 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6054 6091 size_t allocated_bytes) {
6055 6092 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6056 6093 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6057 6094
6058 6095 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6059 6096 _summary_bytes_used += allocated_bytes;
6060 6097 _hr_printer.retire(alloc_region);
6061 6098 // We update the eden sizes here, when the region is retired,
6062 6099 // instead of when it's allocated, since this is the point that its
6063 6100 // used space has been recored in _summary_bytes_used.
6064 6101 g1mm()->update_eden_size();
6065 6102 }
6066 6103
6067 6104 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6068 6105 bool force) {
6069 6106 return _g1h->new_mutator_alloc_region(word_size, force);
6070 6107 }
6071 6108
6072 6109 void G1CollectedHeap::set_par_threads() {
6073 6110 // Don't change the number of workers. Use the value previously set
6074 6111 // in the workgroup.
6075 6112 int n_workers = workers()->active_workers();
6076 6113 assert(UseDynamicNumberOfGCThreads ||
6077 6114 n_workers == workers()->total_workers(),
6078 6115 "Otherwise should be using the total number of workers");
6079 6116 if (n_workers == 0) {
6080 6117 assert(false, "Should have been set in prior evacuation pause.");
6081 6118 n_workers = ParallelGCThreads;
6082 6119 workers()->set_active_workers(n_workers);
6083 6120 }
6084 6121 set_par_threads(n_workers);
6085 6122 }
6086 6123
6087 6124 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6088 6125 size_t allocated_bytes) {
6089 6126 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6090 6127 }
6091 6128
6092 6129 // Methods for the GC alloc regions
6093 6130
6094 6131 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6095 6132 size_t count,
6096 6133 GCAllocPurpose ap) {
6097 6134 assert(FreeList_lock->owned_by_self(), "pre-condition");
6098 6135
6099 6136 if (count < g1_policy()->max_regions(ap)) {
6100 6137 HeapRegion* new_alloc_region = new_region(word_size,
6101 6138 true /* do_expand */);
6102 6139 if (new_alloc_region != NULL) {
6103 6140 // We really only need to do this for old regions given that we
6104 6141 // should never scan survivors. But it doesn't hurt to do it
6105 6142 // for survivors too.
6106 6143 new_alloc_region->set_saved_mark();
6107 6144 if (ap == GCAllocForSurvived) {
6108 6145 new_alloc_region->set_survivor();
6109 6146 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6110 6147 } else {
6111 6148 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6112 6149 }
6113 6150 return new_alloc_region;
6114 6151 } else {
6115 6152 g1_policy()->note_alloc_region_limit_reached(ap);
6116 6153 }
6117 6154 }
6118 6155 return NULL;
6119 6156 }
6120 6157
6121 6158 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6122 6159 size_t allocated_bytes,
6123 6160 GCAllocPurpose ap) {
6124 6161 alloc_region->note_end_of_copying();
6125 6162 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6126 6163 if (ap == GCAllocForSurvived) {
6127 6164 young_list()->add_survivor_region(alloc_region);
6128 6165 } else {
6129 6166 _old_set.add(alloc_region);
6130 6167 }
6131 6168 _hr_printer.retire(alloc_region);
6132 6169 }
6133 6170
6134 6171 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6135 6172 bool force) {
6136 6173 assert(!force, "not supported for GC alloc regions");
6137 6174 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6138 6175 }
6139 6176
6140 6177 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6141 6178 size_t allocated_bytes) {
6142 6179 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6143 6180 GCAllocForSurvived);
6144 6181 }
6145 6182
6146 6183 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6147 6184 bool force) {
6148 6185 assert(!force, "not supported for GC alloc regions");
6149 6186 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6150 6187 }
6151 6188
6152 6189 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6153 6190 size_t allocated_bytes) {
6154 6191 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6155 6192 GCAllocForTenured);
6156 6193 }
6157 6194 // Heap region set verification
6158 6195
6159 6196 class VerifyRegionListsClosure : public HeapRegionClosure {
6160 6197 private:
6161 6198 FreeRegionList* _free_list;
6162 6199 OldRegionSet* _old_set;
6163 6200 HumongousRegionSet* _humongous_set;
6164 6201 size_t _region_count;
6165 6202
6166 6203 public:
6167 6204 VerifyRegionListsClosure(OldRegionSet* old_set,
6168 6205 HumongousRegionSet* humongous_set,
6169 6206 FreeRegionList* free_list) :
6170 6207 _old_set(old_set), _humongous_set(humongous_set),
6171 6208 _free_list(free_list), _region_count(0) { }
6172 6209
6173 6210 size_t region_count() { return _region_count; }
6174 6211
6175 6212 bool doHeapRegion(HeapRegion* hr) {
6176 6213 _region_count += 1;
6177 6214
6178 6215 if (hr->continuesHumongous()) {
6179 6216 return false;
6180 6217 }
6181 6218
6182 6219 if (hr->is_young()) {
6183 6220 // TODO
6184 6221 } else if (hr->startsHumongous()) {
6185 6222 _humongous_set->verify_next_region(hr);
6186 6223 } else if (hr->is_empty()) {
6187 6224 _free_list->verify_next_region(hr);
6188 6225 } else {
6189 6226 _old_set->verify_next_region(hr);
6190 6227 }
6191 6228 return false;
6192 6229 }
6193 6230 };
6194 6231
6195 6232 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index,
6196 6233 HeapWord* bottom) {
6197 6234 HeapWord* end = bottom + HeapRegion::GrainWords;
6198 6235 MemRegion mr(bottom, end);
6199 6236 assert(_g1_reserved.contains(mr), "invariant");
6200 6237 // This might return NULL if the allocation fails
6201 6238 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
6202 6239 }
6203 6240
6204 6241 void G1CollectedHeap::verify_region_sets() {
6205 6242 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6206 6243
6207 6244 // First, check the explicit lists.
6208 6245 _free_list.verify();
6209 6246 {
6210 6247 // Given that a concurrent operation might be adding regions to
6211 6248 // the secondary free list we have to take the lock before
6212 6249 // verifying it.
6213 6250 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6214 6251 _secondary_free_list.verify();
6215 6252 }
6216 6253 _old_set.verify();
6217 6254 _humongous_set.verify();
6218 6255
6219 6256 // If a concurrent region freeing operation is in progress it will
6220 6257 // be difficult to correctly attributed any free regions we come
6221 6258 // across to the correct free list given that they might belong to
6222 6259 // one of several (free_list, secondary_free_list, any local lists,
6223 6260 // etc.). So, if that's the case we will skip the rest of the
6224 6261 // verification operation. Alternatively, waiting for the concurrent
6225 6262 // operation to complete will have a non-trivial effect on the GC's
6226 6263 // operation (no concurrent operation will last longer than the
6227 6264 // interval between two calls to verification) and it might hide
6228 6265 // any issues that we would like to catch during testing.
6229 6266 if (free_regions_coming()) {
6230 6267 return;
6231 6268 }
6232 6269
6233 6270 // Make sure we append the secondary_free_list on the free_list so
6234 6271 // that all free regions we will come across can be safely
6235 6272 // attributed to the free_list.
6236 6273 append_secondary_free_list_if_not_empty_with_lock();
6237 6274
6238 6275 // Finally, make sure that the region accounting in the lists is
6239 6276 // consistent with what we see in the heap.
6240 6277 _old_set.verify_start();
6241 6278 _humongous_set.verify_start();
6242 6279 _free_list.verify_start();
6243 6280
6244 6281 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6245 6282 heap_region_iterate(&cl);
6246 6283
6247 6284 _old_set.verify_end();
6248 6285 _humongous_set.verify_end();
6249 6286 _free_list.verify_end();
6250 6287 }
↓ open down ↓ |
988 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX