Print this page
rev 3618 : 7193946: Move warnings associated with UseMemSetInBOT flag
Summary: The warnings associated with the UseMemSetInBOT flag are duplicated in CMS and G1. The separate warnings have been removed and single instance of the warning has been placed in a common location.
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, 2012, 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/g1EvacFailure.hpp"
36 36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
37 37 #include "gc_implementation/g1/g1Log.hpp"
38 38 #include "gc_implementation/g1/g1MarkSweep.hpp"
39 39 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
40 40 #include "gc_implementation/g1/g1RemSet.inline.hpp"
41 41 #include "gc_implementation/g1/heapRegion.inline.hpp"
42 42 #include "gc_implementation/g1/heapRegionRemSet.hpp"
43 43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
44 44 #include "gc_implementation/g1/vm_operations_g1.hpp"
45 45 #include "gc_implementation/shared/isGCActiveMark.hpp"
46 46 #include "memory/gcLocker.inline.hpp"
47 47 #include "memory/genOopClosures.inline.hpp"
48 48 #include "memory/generationSpec.hpp"
49 49 #include "memory/referenceProcessor.hpp"
50 50 #include "oops/oop.inline.hpp"
51 51 #include "oops/oop.pcgc.inline.hpp"
52 52 #include "runtime/aprofiler.hpp"
53 53 #include "runtime/vmThread.hpp"
54 54
55 55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
56 56
57 57 // turn it on so that the contents of the young list (scan-only /
58 58 // to-be-collected) are printed at "strategic" points before / during
59 59 // / after the collection --- this is useful for debugging
60 60 #define YOUNG_LIST_VERBOSE 0
61 61 // CURRENT STATUS
62 62 // This file is under construction. Search for "FIXME".
63 63
64 64 // INVARIANTS/NOTES
65 65 //
66 66 // All allocation activity covered by the G1CollectedHeap interface is
67 67 // serialized by acquiring the HeapLock. This happens in mem_allocate
68 68 // and allocate_new_tlab, which are the "entry" points to the
69 69 // allocation code from the rest of the JVM. (Note that this does not
70 70 // apply to TLAB allocation, which is not part of this interface: it
71 71 // is done by clients of this interface.)
72 72
73 73 // Notes on implementation of parallelism in different tasks.
74 74 //
75 75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
76 76 // The number of GC workers is passed to heap_region_par_iterate_chunked().
77 77 // It does use run_task() which sets _n_workers in the task.
78 78 // G1ParTask executes g1_process_strong_roots() ->
79 79 // SharedHeap::process_strong_roots() which calls eventuall to
80 80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
81 81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
82 82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
83 83 //
84 84
85 85 // Local to this file.
86 86
87 87 class RefineCardTableEntryClosure: public CardTableEntryClosure {
88 88 SuspendibleThreadSet* _sts;
89 89 G1RemSet* _g1rs;
90 90 ConcurrentG1Refine* _cg1r;
91 91 bool _concurrent;
92 92 public:
93 93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
94 94 G1RemSet* g1rs,
95 95 ConcurrentG1Refine* cg1r) :
96 96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
97 97 {}
98 98 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
99 99 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
100 100 // This path is executed by the concurrent refine or mutator threads,
101 101 // concurrently, and so we do not care if card_ptr contains references
102 102 // that point into the collection set.
103 103 assert(!oops_into_cset, "should be");
104 104
105 105 if (_concurrent && _sts->should_yield()) {
106 106 // Caller will actually yield.
107 107 return false;
108 108 }
109 109 // Otherwise, we finished successfully; return true.
110 110 return true;
111 111 }
112 112 void set_concurrent(bool b) { _concurrent = b; }
113 113 };
114 114
115 115
116 116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
117 117 int _calls;
118 118 G1CollectedHeap* _g1h;
119 119 CardTableModRefBS* _ctbs;
120 120 int _histo[256];
121 121 public:
122 122 ClearLoggedCardTableEntryClosure() :
123 123 _calls(0)
124 124 {
125 125 _g1h = G1CollectedHeap::heap();
126 126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
127 127 for (int i = 0; i < 256; i++) _histo[i] = 0;
128 128 }
129 129 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
130 130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
131 131 _calls++;
132 132 unsigned char* ujb = (unsigned char*)card_ptr;
133 133 int ind = (int)(*ujb);
134 134 _histo[ind]++;
135 135 *card_ptr = -1;
136 136 }
137 137 return true;
138 138 }
139 139 int calls() { return _calls; }
140 140 void print_histo() {
141 141 gclog_or_tty->print_cr("Card table value histogram:");
142 142 for (int i = 0; i < 256; i++) {
143 143 if (_histo[i] != 0) {
144 144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
145 145 }
146 146 }
147 147 }
148 148 };
149 149
150 150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
151 151 int _calls;
152 152 G1CollectedHeap* _g1h;
153 153 CardTableModRefBS* _ctbs;
154 154 public:
155 155 RedirtyLoggedCardTableEntryClosure() :
156 156 _calls(0)
157 157 {
158 158 _g1h = G1CollectedHeap::heap();
159 159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
160 160 }
161 161 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
162 162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
163 163 _calls++;
164 164 *card_ptr = 0;
165 165 }
166 166 return true;
167 167 }
168 168 int calls() { return _calls; }
169 169 };
170 170
171 171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
172 172 public:
173 173 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
174 174 *card_ptr = CardTableModRefBS::dirty_card_val();
175 175 return true;
176 176 }
177 177 };
178 178
179 179 YoungList::YoungList(G1CollectedHeap* g1h) :
180 180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
181 181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
182 182 guarantee(check_list_empty(false), "just making sure...");
183 183 }
184 184
185 185 void YoungList::push_region(HeapRegion *hr) {
186 186 assert(!hr->is_young(), "should not already be young");
187 187 assert(hr->get_next_young_region() == NULL, "cause it should!");
188 188
189 189 hr->set_next_young_region(_head);
190 190 _head = hr;
191 191
192 192 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
193 193 ++_length;
194 194 }
195 195
196 196 void YoungList::add_survivor_region(HeapRegion* hr) {
197 197 assert(hr->is_survivor(), "should be flagged as survivor region");
198 198 assert(hr->get_next_young_region() == NULL, "cause it should!");
199 199
200 200 hr->set_next_young_region(_survivor_head);
201 201 if (_survivor_head == NULL) {
202 202 _survivor_tail = hr;
203 203 }
204 204 _survivor_head = hr;
205 205 ++_survivor_length;
206 206 }
207 207
208 208 void YoungList::empty_list(HeapRegion* list) {
209 209 while (list != NULL) {
210 210 HeapRegion* next = list->get_next_young_region();
211 211 list->set_next_young_region(NULL);
212 212 list->uninstall_surv_rate_group();
213 213 list->set_not_young();
214 214 list = next;
215 215 }
216 216 }
217 217
218 218 void YoungList::empty_list() {
219 219 assert(check_list_well_formed(), "young list should be well formed");
220 220
221 221 empty_list(_head);
222 222 _head = NULL;
223 223 _length = 0;
224 224
225 225 empty_list(_survivor_head);
226 226 _survivor_head = NULL;
227 227 _survivor_tail = NULL;
228 228 _survivor_length = 0;
229 229
230 230 _last_sampled_rs_lengths = 0;
231 231
232 232 assert(check_list_empty(false), "just making sure...");
233 233 }
234 234
235 235 bool YoungList::check_list_well_formed() {
236 236 bool ret = true;
237 237
238 238 uint length = 0;
239 239 HeapRegion* curr = _head;
240 240 HeapRegion* last = NULL;
241 241 while (curr != NULL) {
242 242 if (!curr->is_young()) {
243 243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
244 244 "incorrectly tagged (y: %d, surv: %d)",
245 245 curr->bottom(), curr->end(),
246 246 curr->is_young(), curr->is_survivor());
247 247 ret = false;
248 248 }
249 249 ++length;
250 250 last = curr;
251 251 curr = curr->get_next_young_region();
252 252 }
253 253 ret = ret && (length == _length);
254 254
255 255 if (!ret) {
256 256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
257 257 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
258 258 length, _length);
259 259 }
260 260
261 261 return ret;
262 262 }
263 263
264 264 bool YoungList::check_list_empty(bool check_sample) {
265 265 bool ret = true;
266 266
267 267 if (_length != 0) {
268 268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
269 269 _length);
270 270 ret = false;
271 271 }
272 272 if (check_sample && _last_sampled_rs_lengths != 0) {
273 273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
274 274 ret = false;
275 275 }
276 276 if (_head != NULL) {
277 277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
278 278 ret = false;
279 279 }
280 280 if (!ret) {
281 281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
282 282 }
283 283
284 284 return ret;
285 285 }
286 286
287 287 void
288 288 YoungList::rs_length_sampling_init() {
289 289 _sampled_rs_lengths = 0;
290 290 _curr = _head;
291 291 }
292 292
293 293 bool
294 294 YoungList::rs_length_sampling_more() {
295 295 return _curr != NULL;
296 296 }
297 297
298 298 void
299 299 YoungList::rs_length_sampling_next() {
300 300 assert( _curr != NULL, "invariant" );
301 301 size_t rs_length = _curr->rem_set()->occupied();
302 302
303 303 _sampled_rs_lengths += rs_length;
304 304
305 305 // The current region may not yet have been added to the
306 306 // incremental collection set (it gets added when it is
307 307 // retired as the current allocation region).
308 308 if (_curr->in_collection_set()) {
309 309 // Update the collection set policy information for this region
310 310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
311 311 }
312 312
313 313 _curr = _curr->get_next_young_region();
314 314 if (_curr == NULL) {
315 315 _last_sampled_rs_lengths = _sampled_rs_lengths;
316 316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
317 317 }
318 318 }
319 319
320 320 void
321 321 YoungList::reset_auxilary_lists() {
322 322 guarantee( is_empty(), "young list should be empty" );
323 323 assert(check_list_well_formed(), "young list should be well formed");
324 324
325 325 // Add survivor regions to SurvRateGroup.
326 326 _g1h->g1_policy()->note_start_adding_survivor_regions();
327 327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
328 328
329 329 int young_index_in_cset = 0;
330 330 for (HeapRegion* curr = _survivor_head;
331 331 curr != NULL;
332 332 curr = curr->get_next_young_region()) {
333 333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
334 334
335 335 // The region is a non-empty survivor so let's add it to
336 336 // the incremental collection set for the next evacuation
337 337 // pause.
338 338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
339 339 young_index_in_cset += 1;
340 340 }
341 341 assert((uint) young_index_in_cset == _survivor_length, "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(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
373 373 HR_FORMAT_PARAMS(curr),
374 374 curr->prev_top_at_mark_start(),
375 375 curr->next_top_at_mark_start(),
376 376 curr->age_in_surv_rate_group_cond());
377 377 curr = curr->get_next_young_region();
378 378 }
379 379 }
380 380
381 381 gclog_or_tty->print_cr("");
382 382 }
383 383
384 384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
385 385 {
386 386 // Claim the right to put the region on the dirty cards region list
387 387 // by installing a self pointer.
388 388 HeapRegion* next = hr->get_next_dirty_cards_region();
389 389 if (next == NULL) {
390 390 HeapRegion* res = (HeapRegion*)
391 391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
392 392 NULL);
393 393 if (res == NULL) {
394 394 HeapRegion* head;
395 395 do {
396 396 // Put the region to the dirty cards region list.
397 397 head = _dirty_cards_region_list;
398 398 next = (HeapRegion*)
399 399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
400 400 if (next == head) {
401 401 assert(hr->get_next_dirty_cards_region() == hr,
402 402 "hr->get_next_dirty_cards_region() != hr");
403 403 if (next == NULL) {
404 404 // The last region in the list points to itself.
405 405 hr->set_next_dirty_cards_region(hr);
406 406 } else {
407 407 hr->set_next_dirty_cards_region(next);
408 408 }
409 409 }
410 410 } while (next != head);
411 411 }
412 412 }
413 413 }
414 414
415 415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
416 416 {
417 417 HeapRegion* head;
418 418 HeapRegion* hr;
419 419 do {
420 420 head = _dirty_cards_region_list;
421 421 if (head == NULL) {
422 422 return NULL;
423 423 }
424 424 HeapRegion* new_head = head->get_next_dirty_cards_region();
425 425 if (head == new_head) {
426 426 // The last region.
427 427 new_head = NULL;
428 428 }
429 429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
430 430 head);
431 431 } while (hr != head);
432 432 assert(hr != NULL, "invariant");
433 433 hr->set_next_dirty_cards_region(NULL);
434 434 return hr;
435 435 }
436 436
437 437 void G1CollectedHeap::stop_conc_gc_threads() {
438 438 _cg1r->stop();
439 439 _cmThread->stop();
440 440 }
441 441
442 442 #ifdef ASSERT
443 443 // A region is added to the collection set as it is retired
444 444 // so an address p can point to a region which will be in the
445 445 // collection set but has not yet been retired. This method
446 446 // therefore is only accurate during a GC pause after all
447 447 // regions have been retired. It is used for debugging
448 448 // to check if an nmethod has references to objects that can
449 449 // be move during a partial collection. Though it can be
450 450 // inaccurate, it is sufficient for G1 because the conservative
451 451 // implementation of is_scavengable() for G1 will indicate that
452 452 // all nmethods must be scanned during a partial collection.
453 453 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
454 454 HeapRegion* hr = heap_region_containing(p);
455 455 return hr != NULL && hr->in_collection_set();
456 456 }
457 457 #endif
458 458
459 459 // Returns true if the reference points to an object that
460 460 // can move in an incremental collecction.
461 461 bool G1CollectedHeap::is_scavengable(const void* p) {
462 462 G1CollectedHeap* g1h = G1CollectedHeap::heap();
463 463 G1CollectorPolicy* g1p = g1h->g1_policy();
464 464 HeapRegion* hr = heap_region_containing(p);
465 465 if (hr == NULL) {
466 466 // null
467 467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p));
468 468 return false;
469 469 } else {
470 470 return !hr->isHumongous();
471 471 }
472 472 }
473 473
474 474 void G1CollectedHeap::check_ct_logs_at_safepoint() {
475 475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
476 476 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
477 477
478 478 // Count the dirty cards at the start.
479 479 CountNonCleanMemRegionClosure count1(this);
480 480 ct_bs->mod_card_iterate(&count1);
481 481 int orig_count = count1.n();
482 482
483 483 // First clear the logged cards.
484 484 ClearLoggedCardTableEntryClosure clear;
485 485 dcqs.set_closure(&clear);
486 486 dcqs.apply_closure_to_all_completed_buffers();
487 487 dcqs.iterate_closure_all_threads(false);
488 488 clear.print_histo();
489 489
490 490 // Now ensure that there's no dirty cards.
491 491 CountNonCleanMemRegionClosure count2(this);
492 492 ct_bs->mod_card_iterate(&count2);
493 493 if (count2.n() != 0) {
494 494 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
495 495 count2.n(), orig_count);
496 496 }
497 497 guarantee(count2.n() == 0, "Card table should be clean.");
498 498
499 499 RedirtyLoggedCardTableEntryClosure redirty;
500 500 JavaThread::dirty_card_queue_set().set_closure(&redirty);
501 501 dcqs.apply_closure_to_all_completed_buffers();
502 502 dcqs.iterate_closure_all_threads(false);
503 503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
504 504 clear.calls(), orig_count);
505 505 guarantee(redirty.calls() == clear.calls(),
506 506 "Or else mechanism is broken.");
507 507
508 508 CountNonCleanMemRegionClosure count3(this);
509 509 ct_bs->mod_card_iterate(&count3);
510 510 if (count3.n() != orig_count) {
511 511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
512 512 orig_count, count3.n());
513 513 guarantee(count3.n() >= orig_count, "Should have restored them all.");
514 514 }
515 515
516 516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
517 517 }
518 518
519 519 // Private class members.
520 520
521 521 G1CollectedHeap* G1CollectedHeap::_g1h;
522 522
523 523 // Private methods.
524 524
525 525 HeapRegion*
526 526 G1CollectedHeap::new_region_try_secondary_free_list() {
527 527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
528 528 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
529 529 if (!_secondary_free_list.is_empty()) {
530 530 if (G1ConcRegionFreeingVerbose) {
531 531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
532 532 "secondary_free_list has %u entries",
533 533 _secondary_free_list.length());
534 534 }
535 535 // It looks as if there are free regions available on the
536 536 // secondary_free_list. Let's move them to the free_list and try
537 537 // again to allocate from it.
538 538 append_secondary_free_list();
539 539
540 540 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
541 541 "empty we should have moved at least one entry to the free_list");
542 542 HeapRegion* res = _free_list.remove_head();
543 543 if (G1ConcRegionFreeingVerbose) {
544 544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
545 545 "allocated "HR_FORMAT" from secondary_free_list",
546 546 HR_FORMAT_PARAMS(res));
547 547 }
548 548 return res;
549 549 }
550 550
551 551 // Wait here until we get notifed either when (a) there are no
552 552 // more free regions coming or (b) some regions have been moved on
553 553 // the secondary_free_list.
554 554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
555 555 }
556 556
557 557 if (G1ConcRegionFreeingVerbose) {
558 558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
559 559 "could not allocate from secondary_free_list");
560 560 }
561 561 return NULL;
562 562 }
563 563
564 564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
565 565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
566 566 "the only time we use this to allocate a humongous region is "
567 567 "when we are allocating a single humongous region");
568 568
569 569 HeapRegion* res;
570 570 if (G1StressConcRegionFreeing) {
571 571 if (!_secondary_free_list.is_empty()) {
572 572 if (G1ConcRegionFreeingVerbose) {
573 573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
574 574 "forced to look at the secondary_free_list");
575 575 }
576 576 res = new_region_try_secondary_free_list();
577 577 if (res != NULL) {
578 578 return res;
579 579 }
580 580 }
581 581 }
582 582 res = _free_list.remove_head_or_null();
583 583 if (res == NULL) {
584 584 if (G1ConcRegionFreeingVerbose) {
585 585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
586 586 "res == NULL, trying the secondary_free_list");
587 587 }
588 588 res = new_region_try_secondary_free_list();
589 589 }
590 590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
591 591 // Currently, only attempts to allocate GC alloc regions set
592 592 // do_expand to true. So, we should only reach here during a
593 593 // safepoint. If this assumption changes we might have to
594 594 // reconsider the use of _expand_heap_after_alloc_failure.
595 595 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
596 596
597 597 ergo_verbose1(ErgoHeapSizing,
598 598 "attempt heap expansion",
599 599 ergo_format_reason("region allocation request failed")
600 600 ergo_format_byte("allocation request"),
601 601 word_size * HeapWordSize);
602 602 if (expand(word_size * HeapWordSize)) {
603 603 // Given that expand() succeeded in expanding the heap, and we
604 604 // always expand the heap by an amount aligned to the heap
605 605 // region size, the free list should in theory not be empty. So
606 606 // it would probably be OK to use remove_head(). But the extra
607 607 // check for NULL is unlikely to be a performance issue here (we
608 608 // just expanded the heap!) so let's just be conservative and
609 609 // use remove_head_or_null().
610 610 res = _free_list.remove_head_or_null();
611 611 } else {
612 612 _expand_heap_after_alloc_failure = false;
613 613 }
614 614 }
615 615 return res;
616 616 }
617 617
618 618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions,
619 619 size_t word_size) {
620 620 assert(isHumongous(word_size), "word_size should be humongous");
621 621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
622 622
623 623 uint first = G1_NULL_HRS_INDEX;
624 624 if (num_regions == 1) {
625 625 // Only one region to allocate, no need to go through the slower
626 626 // path. The caller will attempt the expasion if this fails, so
627 627 // let's not try to expand here too.
628 628 HeapRegion* hr = new_region(word_size, false /* do_expand */);
629 629 if (hr != NULL) {
630 630 first = hr->hrs_index();
631 631 } else {
632 632 first = G1_NULL_HRS_INDEX;
633 633 }
634 634 } else {
635 635 // We can't allocate humongous regions while cleanupComplete() is
636 636 // running, since some of the regions we find to be empty might not
637 637 // yet be added to the free list and it is not straightforward to
638 638 // know which list they are on so that we can remove them. Note
639 639 // that we only need to do this if we need to allocate more than
640 640 // one region to satisfy the current humongous allocation
641 641 // request. If we are only allocating one region we use the common
642 642 // region allocation code (see above).
643 643 wait_while_free_regions_coming();
644 644 append_secondary_free_list_if_not_empty_with_lock();
645 645
646 646 if (free_regions() >= num_regions) {
647 647 first = _hrs.find_contiguous(num_regions);
648 648 if (first != G1_NULL_HRS_INDEX) {
649 649 for (uint i = first; i < first + num_regions; ++i) {
650 650 HeapRegion* hr = region_at(i);
651 651 assert(hr->is_empty(), "sanity");
652 652 assert(is_on_master_free_list(hr), "sanity");
653 653 hr->set_pending_removal(true);
654 654 }
655 655 _free_list.remove_all_pending(num_regions);
656 656 }
657 657 }
658 658 }
659 659 return first;
660 660 }
661 661
662 662 HeapWord*
663 663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
664 664 uint num_regions,
665 665 size_t word_size) {
666 666 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
667 667 assert(isHumongous(word_size), "word_size should be humongous");
668 668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
669 669
670 670 // Index of last region in the series + 1.
671 671 uint last = first + num_regions;
672 672
673 673 // We need to initialize the region(s) we just discovered. This is
674 674 // a bit tricky given that it can happen concurrently with
675 675 // refinement threads refining cards on these regions and
676 676 // potentially wanting to refine the BOT as they are scanning
677 677 // those cards (this can happen shortly after a cleanup; see CR
678 678 // 6991377). So we have to set up the region(s) carefully and in
679 679 // a specific order.
680 680
681 681 // The word size sum of all the regions we will allocate.
682 682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
683 683 assert(word_size <= word_size_sum, "sanity");
684 684
685 685 // This will be the "starts humongous" region.
686 686 HeapRegion* first_hr = region_at(first);
687 687 // The header of the new object will be placed at the bottom of
688 688 // the first region.
689 689 HeapWord* new_obj = first_hr->bottom();
690 690 // This will be the new end of the first region in the series that
691 691 // should also match the end of the last region in the seriers.
692 692 HeapWord* new_end = new_obj + word_size_sum;
693 693 // This will be the new top of the first region that will reflect
694 694 // this allocation.
695 695 HeapWord* new_top = new_obj + word_size;
696 696
697 697 // First, we need to zero the header of the space that we will be
698 698 // allocating. When we update top further down, some refinement
699 699 // threads might try to scan the region. By zeroing the header we
700 700 // ensure that any thread that will try to scan the region will
701 701 // come across the zero klass word and bail out.
702 702 //
703 703 // NOTE: It would not have been correct to have used
704 704 // CollectedHeap::fill_with_object() and make the space look like
705 705 // an int array. The thread that is doing the allocation will
706 706 // later update the object header to a potentially different array
707 707 // type and, for a very short period of time, the klass and length
708 708 // fields will be inconsistent. This could cause a refinement
709 709 // thread to calculate the object size incorrectly.
710 710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
711 711
712 712 // We will set up the first region as "starts humongous". This
713 713 // will also update the BOT covering all the regions to reflect
714 714 // that there is a single object that starts at the bottom of the
715 715 // first region.
716 716 first_hr->set_startsHumongous(new_top, new_end);
717 717
718 718 // Then, if there are any, we will set up the "continues
719 719 // humongous" regions.
720 720 HeapRegion* hr = NULL;
721 721 for (uint i = first + 1; i < last; ++i) {
722 722 hr = region_at(i);
723 723 hr->set_continuesHumongous(first_hr);
724 724 }
725 725 // If we have "continues humongous" regions (hr != NULL), then the
726 726 // end of the last one should match new_end.
727 727 assert(hr == NULL || hr->end() == new_end, "sanity");
728 728
729 729 // Up to this point no concurrent thread would have been able to
730 730 // do any scanning on any region in this series. All the top
731 731 // fields still point to bottom, so the intersection between
732 732 // [bottom,top] and [card_start,card_end] will be empty. Before we
733 733 // update the top fields, we'll do a storestore to make sure that
734 734 // no thread sees the update to top before the zeroing of the
735 735 // object header and the BOT initialization.
736 736 OrderAccess::storestore();
737 737
738 738 // Now that the BOT and the object header have been initialized,
739 739 // we can update top of the "starts humongous" region.
740 740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
741 741 "new_top should be in this region");
742 742 first_hr->set_top(new_top);
743 743 if (_hr_printer.is_active()) {
744 744 HeapWord* bottom = first_hr->bottom();
745 745 HeapWord* end = first_hr->orig_end();
746 746 if ((first + 1) == last) {
747 747 // the series has a single humongous region
748 748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
749 749 } else {
750 750 // the series has more than one humongous regions
751 751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
752 752 }
753 753 }
754 754
755 755 // Now, we will update the top fields of the "continues humongous"
756 756 // regions. The reason we need to do this is that, otherwise,
757 757 // these regions would look empty and this will confuse parts of
758 758 // G1. For example, the code that looks for a consecutive number
759 759 // of empty regions will consider them empty and try to
760 760 // re-allocate them. We can extend is_empty() to also include
761 761 // !continuesHumongous(), but it is easier to just update the top
762 762 // fields here. The way we set top for all regions (i.e., top ==
763 763 // end for all regions but the last one, top == new_top for the
764 764 // last one) is actually used when we will free up the humongous
765 765 // region in free_humongous_region().
766 766 hr = NULL;
767 767 for (uint i = first + 1; i < last; ++i) {
768 768 hr = region_at(i);
769 769 if ((i + 1) == last) {
770 770 // last continues humongous region
771 771 assert(hr->bottom() < new_top && new_top <= hr->end(),
772 772 "new_top should fall on this region");
773 773 hr->set_top(new_top);
774 774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
775 775 } else {
776 776 // not last one
777 777 assert(new_top > hr->end(), "new_top should be above this region");
778 778 hr->set_top(hr->end());
779 779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
780 780 }
781 781 }
782 782 // If we have continues humongous regions (hr != NULL), then the
783 783 // end of the last one should match new_end and its top should
784 784 // match new_top.
785 785 assert(hr == NULL ||
786 786 (hr->end() == new_end && hr->top() == new_top), "sanity");
787 787
788 788 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
789 789 _summary_bytes_used += first_hr->used();
790 790 _humongous_set.add(first_hr);
791 791
792 792 return new_obj;
793 793 }
794 794
795 795 // If could fit into free regions w/o expansion, try.
796 796 // Otherwise, if can expand, do so.
797 797 // Otherwise, if using ex regions might help, try with ex given back.
798 798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
799 799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
800 800
801 801 verify_region_sets_optional();
802 802
803 803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords);
804 804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords);
805 805 uint x_num = expansion_regions();
806 806 uint fs = _hrs.free_suffix();
807 807 uint first = humongous_obj_allocate_find_first(num_regions, word_size);
808 808 if (first == G1_NULL_HRS_INDEX) {
809 809 // The only thing we can do now is attempt expansion.
810 810 if (fs + x_num >= num_regions) {
811 811 // If the number of regions we're trying to allocate for this
812 812 // object is at most the number of regions in the free suffix,
813 813 // then the call to humongous_obj_allocate_find_first() above
814 814 // should have succeeded and we wouldn't be here.
815 815 //
816 816 // We should only be trying to expand when the free suffix is
817 817 // not sufficient for the object _and_ we have some expansion
818 818 // room available.
819 819 assert(num_regions > fs, "earlier allocation should have succeeded");
820 820
821 821 ergo_verbose1(ErgoHeapSizing,
822 822 "attempt heap expansion",
823 823 ergo_format_reason("humongous allocation request failed")
824 824 ergo_format_byte("allocation request"),
825 825 word_size * HeapWordSize);
826 826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
827 827 // Even though the heap was expanded, it might not have
828 828 // reached the desired size. So, we cannot assume that the
829 829 // allocation will succeed.
830 830 first = humongous_obj_allocate_find_first(num_regions, word_size);
831 831 }
832 832 }
833 833 }
834 834
835 835 HeapWord* result = NULL;
836 836 if (first != G1_NULL_HRS_INDEX) {
837 837 result =
838 838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
839 839 assert(result != NULL, "it should always return a valid result");
840 840
841 841 // A successful humongous object allocation changes the used space
842 842 // information of the old generation so we need to recalculate the
843 843 // sizes and update the jstat counters here.
844 844 g1mm()->update_sizes();
845 845 }
846 846
847 847 verify_region_sets_optional();
848 848
849 849 return result;
850 850 }
851 851
852 852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
853 853 assert_heap_not_locked_and_not_at_safepoint();
854 854 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
855 855
856 856 unsigned int dummy_gc_count_before;
857 857 return attempt_allocation(word_size, &dummy_gc_count_before);
858 858 }
859 859
860 860 HeapWord*
861 861 G1CollectedHeap::mem_allocate(size_t word_size,
862 862 bool* gc_overhead_limit_was_exceeded) {
863 863 assert_heap_not_locked_and_not_at_safepoint();
864 864
865 865 // Loop until the allocation is satisified, or unsatisfied after GC.
866 866 for (int try_count = 1; /* we'll return */; try_count += 1) {
867 867 unsigned int gc_count_before;
868 868
869 869 HeapWord* result = NULL;
870 870 if (!isHumongous(word_size)) {
871 871 result = attempt_allocation(word_size, &gc_count_before);
872 872 } else {
873 873 result = attempt_allocation_humongous(word_size, &gc_count_before);
874 874 }
875 875 if (result != NULL) {
876 876 return result;
877 877 }
878 878
879 879 // Create the garbage collection operation...
880 880 VM_G1CollectForAllocation op(gc_count_before, word_size);
881 881 // ...and get the VM thread to execute it.
882 882 VMThread::execute(&op);
883 883
884 884 if (op.prologue_succeeded() && op.pause_succeeded()) {
885 885 // If the operation was successful we'll return the result even
886 886 // if it is NULL. If the allocation attempt failed immediately
887 887 // after a Full GC, it's unlikely we'll be able to allocate now.
888 888 HeapWord* result = op.result();
889 889 if (result != NULL && !isHumongous(word_size)) {
890 890 // Allocations that take place on VM operations do not do any
891 891 // card dirtying and we have to do it here. We only have to do
892 892 // this for non-humongous allocations, though.
893 893 dirty_young_block(result, word_size);
894 894 }
895 895 return result;
896 896 } else {
897 897 assert(op.result() == NULL,
898 898 "the result should be NULL if the VM op did not succeed");
899 899 }
900 900
901 901 // Give a warning if we seem to be looping forever.
902 902 if ((QueuedAllocationWarningCount > 0) &&
903 903 (try_count % QueuedAllocationWarningCount == 0)) {
904 904 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
905 905 }
906 906 }
907 907
908 908 ShouldNotReachHere();
909 909 return NULL;
910 910 }
911 911
912 912 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
913 913 unsigned int *gc_count_before_ret) {
914 914 // Make sure you read the note in attempt_allocation_humongous().
915 915
916 916 assert_heap_not_locked_and_not_at_safepoint();
917 917 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
918 918 "be called for humongous allocation requests");
919 919
920 920 // We should only get here after the first-level allocation attempt
921 921 // (attempt_allocation()) failed to allocate.
922 922
923 923 // We will loop until a) we manage to successfully perform the
924 924 // allocation or b) we successfully schedule a collection which
925 925 // fails to perform the allocation. b) is the only case when we'll
926 926 // return NULL.
927 927 HeapWord* result = NULL;
928 928 for (int try_count = 1; /* we'll return */; try_count += 1) {
929 929 bool should_try_gc;
930 930 unsigned int gc_count_before;
931 931
932 932 {
933 933 MutexLockerEx x(Heap_lock);
934 934
935 935 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
936 936 false /* bot_updates */);
937 937 if (result != NULL) {
938 938 return result;
939 939 }
940 940
941 941 // If we reach here, attempt_allocation_locked() above failed to
942 942 // allocate a new region. So the mutator alloc region should be NULL.
943 943 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
944 944
945 945 if (GC_locker::is_active_and_needs_gc()) {
946 946 if (g1_policy()->can_expand_young_list()) {
947 947 // No need for an ergo verbose message here,
948 948 // can_expand_young_list() does this when it returns true.
949 949 result = _mutator_alloc_region.attempt_allocation_force(word_size,
950 950 false /* bot_updates */);
951 951 if (result != NULL) {
952 952 return result;
953 953 }
954 954 }
955 955 should_try_gc = false;
956 956 } else {
957 957 // The GCLocker may not be active but the GCLocker initiated
958 958 // GC may not yet have been performed (GCLocker::needs_gc()
959 959 // returns true). In this case we do not try this GC and
960 960 // wait until the GCLocker initiated GC is performed, and
961 961 // then retry the allocation.
962 962 if (GC_locker::needs_gc()) {
963 963 should_try_gc = false;
964 964 } else {
965 965 // Read the GC count while still holding the Heap_lock.
966 966 gc_count_before = total_collections();
967 967 should_try_gc = true;
968 968 }
969 969 }
970 970 }
971 971
972 972 if (should_try_gc) {
973 973 bool succeeded;
974 974 result = do_collection_pause(word_size, gc_count_before, &succeeded);
975 975 if (result != NULL) {
976 976 assert(succeeded, "only way to get back a non-NULL result");
977 977 return result;
978 978 }
979 979
980 980 if (succeeded) {
981 981 // If we get here we successfully scheduled a collection which
982 982 // failed to allocate. No point in trying to allocate
983 983 // further. We'll just return NULL.
984 984 MutexLockerEx x(Heap_lock);
985 985 *gc_count_before_ret = total_collections();
986 986 return NULL;
987 987 }
988 988 } else {
989 989 // The GCLocker is either active or the GCLocker initiated
990 990 // GC has not yet been performed. Stall until it is and
991 991 // then retry the allocation.
992 992 GC_locker::stall_until_clear();
993 993 }
994 994
995 995 // We can reach here if we were unsuccessul in scheduling a
996 996 // collection (because another thread beat us to it) or if we were
997 997 // stalled due to the GC locker. In either can we should retry the
998 998 // allocation attempt in case another thread successfully
999 999 // performed a collection and reclaimed enough space. We do the
1000 1000 // first attempt (without holding the Heap_lock) here and the
1001 1001 // follow-on attempt will be at the start of the next loop
1002 1002 // iteration (after taking the Heap_lock).
1003 1003 result = _mutator_alloc_region.attempt_allocation(word_size,
1004 1004 false /* bot_updates */);
1005 1005 if (result != NULL) {
1006 1006 return result;
1007 1007 }
1008 1008
1009 1009 // Give a warning if we seem to be looping forever.
1010 1010 if ((QueuedAllocationWarningCount > 0) &&
1011 1011 (try_count % QueuedAllocationWarningCount == 0)) {
1012 1012 warning("G1CollectedHeap::attempt_allocation_slow() "
1013 1013 "retries %d times", try_count);
1014 1014 }
1015 1015 }
1016 1016
1017 1017 ShouldNotReachHere();
1018 1018 return NULL;
1019 1019 }
1020 1020
1021 1021 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1022 1022 unsigned int * gc_count_before_ret) {
1023 1023 // The structure of this method has a lot of similarities to
1024 1024 // attempt_allocation_slow(). The reason these two were not merged
1025 1025 // into a single one is that such a method would require several "if
1026 1026 // allocation is not humongous do this, otherwise do that"
1027 1027 // conditional paths which would obscure its flow. In fact, an early
1028 1028 // version of this code did use a unified method which was harder to
1029 1029 // follow and, as a result, it had subtle bugs that were hard to
1030 1030 // track down. So keeping these two methods separate allows each to
1031 1031 // be more readable. It will be good to keep these two in sync as
1032 1032 // much as possible.
1033 1033
1034 1034 assert_heap_not_locked_and_not_at_safepoint();
1035 1035 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1036 1036 "should only be called for humongous allocations");
1037 1037
1038 1038 // Humongous objects can exhaust the heap quickly, so we should check if we
1039 1039 // need to start a marking cycle at each humongous object allocation. We do
1040 1040 // the check before we do the actual allocation. The reason for doing it
1041 1041 // before the allocation is that we avoid having to keep track of the newly
1042 1042 // allocated memory while we do a GC.
1043 1043 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1044 1044 word_size)) {
1045 1045 collect(GCCause::_g1_humongous_allocation);
1046 1046 }
1047 1047
1048 1048 // We will loop until a) we manage to successfully perform the
1049 1049 // allocation or b) we successfully schedule a collection which
1050 1050 // fails to perform the allocation. b) is the only case when we'll
1051 1051 // return NULL.
1052 1052 HeapWord* result = NULL;
1053 1053 for (int try_count = 1; /* we'll return */; try_count += 1) {
1054 1054 bool should_try_gc;
1055 1055 unsigned int gc_count_before;
1056 1056
1057 1057 {
1058 1058 MutexLockerEx x(Heap_lock);
1059 1059
1060 1060 // Given that humongous objects are not allocated in young
1061 1061 // regions, we'll first try to do the allocation without doing a
1062 1062 // collection hoping that there's enough space in the heap.
1063 1063 result = humongous_obj_allocate(word_size);
1064 1064 if (result != NULL) {
1065 1065 return result;
1066 1066 }
1067 1067
1068 1068 if (GC_locker::is_active_and_needs_gc()) {
1069 1069 should_try_gc = false;
1070 1070 } else {
1071 1071 // The GCLocker may not be active but the GCLocker initiated
1072 1072 // GC may not yet have been performed (GCLocker::needs_gc()
1073 1073 // returns true). In this case we do not try this GC and
1074 1074 // wait until the GCLocker initiated GC is performed, and
1075 1075 // then retry the allocation.
1076 1076 if (GC_locker::needs_gc()) {
1077 1077 should_try_gc = false;
1078 1078 } else {
1079 1079 // Read the GC count while still holding the Heap_lock.
1080 1080 gc_count_before = total_collections();
1081 1081 should_try_gc = true;
1082 1082 }
1083 1083 }
1084 1084 }
1085 1085
1086 1086 if (should_try_gc) {
1087 1087 // If we failed to allocate the humongous object, we should try to
1088 1088 // do a collection pause (if we're allowed) in case it reclaims
1089 1089 // enough space for the allocation to succeed after the pause.
1090 1090
1091 1091 bool succeeded;
1092 1092 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1093 1093 if (result != NULL) {
1094 1094 assert(succeeded, "only way to get back a non-NULL result");
1095 1095 return result;
1096 1096 }
1097 1097
1098 1098 if (succeeded) {
1099 1099 // If we get here we successfully scheduled a collection which
1100 1100 // failed to allocate. No point in trying to allocate
1101 1101 // further. We'll just return NULL.
1102 1102 MutexLockerEx x(Heap_lock);
1103 1103 *gc_count_before_ret = total_collections();
1104 1104 return NULL;
1105 1105 }
1106 1106 } else {
1107 1107 // The GCLocker is either active or the GCLocker initiated
1108 1108 // GC has not yet been performed. Stall until it is and
1109 1109 // then retry the allocation.
1110 1110 GC_locker::stall_until_clear();
1111 1111 }
1112 1112
1113 1113 // We can reach here if we were unsuccessul in scheduling a
1114 1114 // collection (because another thread beat us to it) or if we were
1115 1115 // stalled due to the GC locker. In either can we should retry the
1116 1116 // allocation attempt in case another thread successfully
1117 1117 // performed a collection and reclaimed enough space. Give a
1118 1118 // warning if we seem to be looping forever.
1119 1119
1120 1120 if ((QueuedAllocationWarningCount > 0) &&
1121 1121 (try_count % QueuedAllocationWarningCount == 0)) {
1122 1122 warning("G1CollectedHeap::attempt_allocation_humongous() "
1123 1123 "retries %d times", try_count);
1124 1124 }
1125 1125 }
1126 1126
1127 1127 ShouldNotReachHere();
1128 1128 return NULL;
1129 1129 }
1130 1130
1131 1131 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1132 1132 bool expect_null_mutator_alloc_region) {
1133 1133 assert_at_safepoint(true /* should_be_vm_thread */);
1134 1134 assert(_mutator_alloc_region.get() == NULL ||
1135 1135 !expect_null_mutator_alloc_region,
1136 1136 "the current alloc region was unexpectedly found to be non-NULL");
1137 1137
1138 1138 if (!isHumongous(word_size)) {
1139 1139 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1140 1140 false /* bot_updates */);
1141 1141 } else {
1142 1142 HeapWord* result = humongous_obj_allocate(word_size);
1143 1143 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1144 1144 g1_policy()->set_initiate_conc_mark_if_possible();
1145 1145 }
1146 1146 return result;
1147 1147 }
1148 1148
1149 1149 ShouldNotReachHere();
1150 1150 }
1151 1151
1152 1152 class PostMCRemSetClearClosure: public HeapRegionClosure {
1153 1153 G1CollectedHeap* _g1h;
1154 1154 ModRefBarrierSet* _mr_bs;
1155 1155 public:
1156 1156 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1157 1157 _g1h(g1h), _mr_bs(mr_bs) { }
1158 1158 bool doHeapRegion(HeapRegion* r) {
1159 1159 if (r->continuesHumongous()) {
1160 1160 return false;
1161 1161 }
1162 1162 _g1h->reset_gc_time_stamps(r);
1163 1163 HeapRegionRemSet* hrrs = r->rem_set();
1164 1164 if (hrrs != NULL) hrrs->clear();
1165 1165 // You might think here that we could clear just the cards
1166 1166 // corresponding to the used region. But no: if we leave a dirty card
1167 1167 // in a region we might allocate into, then it would prevent that card
1168 1168 // from being enqueued, and cause it to be missed.
1169 1169 // Re: the performance cost: we shouldn't be doing full GC anyway!
1170 1170 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1171 1171 return false;
1172 1172 }
1173 1173 };
1174 1174
1175 1175 void G1CollectedHeap::clear_rsets_post_compaction() {
1176 1176 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1177 1177 heap_region_iterate(&rs_clear);
1178 1178 }
1179 1179
1180 1180 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1181 1181 G1CollectedHeap* _g1h;
1182 1182 UpdateRSOopClosure _cl;
1183 1183 int _worker_i;
1184 1184 public:
1185 1185 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1186 1186 _cl(g1->g1_rem_set(), worker_i),
1187 1187 _worker_i(worker_i),
1188 1188 _g1h(g1)
1189 1189 { }
1190 1190
1191 1191 bool doHeapRegion(HeapRegion* r) {
1192 1192 if (!r->continuesHumongous()) {
1193 1193 _cl.set_from(r);
1194 1194 r->oop_iterate(&_cl);
1195 1195 }
1196 1196 return false;
1197 1197 }
1198 1198 };
1199 1199
1200 1200 class ParRebuildRSTask: public AbstractGangTask {
1201 1201 G1CollectedHeap* _g1;
1202 1202 public:
1203 1203 ParRebuildRSTask(G1CollectedHeap* g1)
1204 1204 : AbstractGangTask("ParRebuildRSTask"),
1205 1205 _g1(g1)
1206 1206 { }
1207 1207
1208 1208 void work(uint worker_id) {
1209 1209 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1210 1210 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1211 1211 _g1->workers()->active_workers(),
1212 1212 HeapRegion::RebuildRSClaimValue);
1213 1213 }
1214 1214 };
1215 1215
1216 1216 class PostCompactionPrinterClosure: public HeapRegionClosure {
1217 1217 private:
1218 1218 G1HRPrinter* _hr_printer;
1219 1219 public:
1220 1220 bool doHeapRegion(HeapRegion* hr) {
1221 1221 assert(!hr->is_young(), "not expecting to find young regions");
1222 1222 // We only generate output for non-empty regions.
1223 1223 if (!hr->is_empty()) {
1224 1224 if (!hr->isHumongous()) {
1225 1225 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1226 1226 } else if (hr->startsHumongous()) {
1227 1227 if (hr->region_num() == 1) {
1228 1228 // single humongous region
1229 1229 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1230 1230 } else {
1231 1231 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1232 1232 }
1233 1233 } else {
1234 1234 assert(hr->continuesHumongous(), "only way to get here");
1235 1235 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1236 1236 }
1237 1237 }
1238 1238 return false;
1239 1239 }
1240 1240
1241 1241 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1242 1242 : _hr_printer(hr_printer) { }
1243 1243 };
1244 1244
1245 1245 void G1CollectedHeap::print_hrs_post_compaction() {
1246 1246 PostCompactionPrinterClosure cl(hr_printer());
1247 1247 heap_region_iterate(&cl);
1248 1248 }
1249 1249
1250 1250 double G1CollectedHeap::verify(bool guard, const char* msg) {
1251 1251 double verify_time_ms = 0.0;
1252 1252
1253 1253 if (guard && total_collections() >= VerifyGCStartAt) {
1254 1254 double verify_start = os::elapsedTime();
1255 1255 HandleMark hm; // Discard invalid handles created during verification
1256 1256 gclog_or_tty->print(msg);
1257 1257 prepare_for_verify();
1258 1258 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking);
1259 1259 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1260 1260 }
1261 1261
1262 1262 return verify_time_ms;
1263 1263 }
1264 1264
1265 1265 void G1CollectedHeap::verify_before_gc() {
1266 1266 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1267 1267 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1268 1268 }
1269 1269
1270 1270 void G1CollectedHeap::verify_after_gc() {
1271 1271 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1272 1272 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1273 1273 }
1274 1274
1275 1275 bool G1CollectedHeap::do_collection(bool explicit_gc,
1276 1276 bool clear_all_soft_refs,
1277 1277 size_t word_size) {
1278 1278 assert_at_safepoint(true /* should_be_vm_thread */);
1279 1279
1280 1280 if (GC_locker::check_active_before_gc()) {
1281 1281 return false;
1282 1282 }
1283 1283
1284 1284 SvcGCMarker sgcm(SvcGCMarker::FULL);
1285 1285 ResourceMark rm;
1286 1286
1287 1287 print_heap_before_gc();
1288 1288
1289 1289 size_t metadata_prev_used = MetaspaceAux::used_in_bytes();
1290 1290
1291 1291 HRSPhaseSetter x(HRSPhaseFullGC);
1292 1292 verify_region_sets_optional();
1293 1293
1294 1294 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1295 1295 collector_policy()->should_clear_all_soft_refs();
1296 1296
1297 1297 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1298 1298
1299 1299 {
1300 1300 IsGCActiveMark x;
1301 1301
1302 1302 // Timing
1303 1303 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1304 1304 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1305 1305 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1306 1306
1307 1307 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1308 1308 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1309 1309 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1310 1310
1311 1311 double start = os::elapsedTime();
1312 1312 g1_policy()->record_full_collection_start();
1313 1313
1314 1314 // Note: When we have a more flexible GC logging framework that
1315 1315 // allows us to add optional attributes to a GC log record we
1316 1316 // could consider timing and reporting how long we wait in the
1317 1317 // following two methods.
1318 1318 wait_while_free_regions_coming();
1319 1319 // If we start the compaction before the CM threads finish
1320 1320 // scanning the root regions we might trip them over as we'll
1321 1321 // be moving objects / updating references. So let's wait until
1322 1322 // they are done. By telling them to abort, they should complete
1323 1323 // early.
1324 1324 _cm->root_regions()->abort();
1325 1325 _cm->root_regions()->wait_until_scan_finished();
1326 1326 append_secondary_free_list_if_not_empty_with_lock();
1327 1327
1328 1328 gc_prologue(true);
1329 1329 increment_total_collections(true /* full gc */);
1330 1330 increment_old_marking_cycles_started();
1331 1331
1332 1332 size_t g1h_prev_used = used();
1333 1333 assert(used() == recalculate_used(), "Should be equal");
1334 1334
1335 1335 verify_before_gc();
1336 1336
1337 1337 pre_full_gc_dump();
1338 1338
1339 1339 COMPILER2_PRESENT(DerivedPointerTable::clear());
1340 1340
1341 1341 // Disable discovery and empty the discovered lists
1342 1342 // for the CM ref processor.
1343 1343 ref_processor_cm()->disable_discovery();
1344 1344 ref_processor_cm()->abandon_partial_discovery();
1345 1345 ref_processor_cm()->verify_no_references_recorded();
1346 1346
1347 1347 // Abandon current iterations of concurrent marking and concurrent
1348 1348 // refinement, if any are in progress. We have to do this before
1349 1349 // wait_until_scan_finished() below.
1350 1350 concurrent_mark()->abort();
1351 1351
1352 1352 // Make sure we'll choose a new allocation region afterwards.
1353 1353 release_mutator_alloc_region();
1354 1354 abandon_gc_alloc_regions();
1355 1355 g1_rem_set()->cleanupHRRS();
1356 1356
1357 1357 // We should call this after we retire any currently active alloc
1358 1358 // regions so that all the ALLOC / RETIRE events are generated
1359 1359 // before the start GC event.
1360 1360 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1361 1361
1362 1362 // We may have added regions to the current incremental collection
1363 1363 // set between the last GC or pause and now. We need to clear the
1364 1364 // incremental collection set and then start rebuilding it afresh
1365 1365 // after this full GC.
1366 1366 abandon_collection_set(g1_policy()->inc_cset_head());
1367 1367 g1_policy()->clear_incremental_cset();
1368 1368 g1_policy()->stop_incremental_cset_building();
1369 1369
1370 1370 tear_down_region_sets(false /* free_list_only */);
1371 1371 g1_policy()->set_gcs_are_young(true);
1372 1372
1373 1373 // See the comments in g1CollectedHeap.hpp and
1374 1374 // G1CollectedHeap::ref_processing_init() about
1375 1375 // how reference processing currently works in G1.
1376 1376
1377 1377 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1378 1378 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1379 1379
1380 1380 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1381 1381 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1382 1382
1383 1383 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1384 1384 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1385 1385
1386 1386 // Do collection work
1387 1387 {
1388 1388 HandleMark hm; // Discard invalid handles created during gc
1389 1389 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1390 1390 }
1391 1391
1392 1392 assert(free_regions() == 0, "we should not have added any free regions");
1393 1393 rebuild_region_sets(false /* free_list_only */);
1394 1394
1395 1395 // Enqueue any discovered reference objects that have
1396 1396 // not been removed from the discovered lists.
1397 1397 ref_processor_stw()->enqueue_discovered_references();
1398 1398
1399 1399 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1400 1400
1401 1401 MemoryService::track_memory_usage();
1402 1402
1403 1403 verify_after_gc();
1404 1404
1405 1405 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1406 1406 ref_processor_stw()->verify_no_references_recorded();
1407 1407
1408 1408 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1409 1409 ClassLoaderDataGraph::purge();
1410 1410
1411 1411 // Note: since we've just done a full GC, concurrent
1412 1412 // marking is no longer active. Therefore we need not
1413 1413 // re-enable reference discovery for the CM ref processor.
1414 1414 // That will be done at the start of the next marking cycle.
1415 1415 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1416 1416 ref_processor_cm()->verify_no_references_recorded();
1417 1417
1418 1418 reset_gc_time_stamp();
1419 1419 // Since everything potentially moved, we will clear all remembered
1420 1420 // sets, and clear all cards. Later we will rebuild remebered
1421 1421 // sets. We will also reset the GC time stamps of the regions.
1422 1422 clear_rsets_post_compaction();
1423 1423 check_gc_time_stamps();
1424 1424
1425 1425 // Resize the heap if necessary.
1426 1426 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1427 1427
1428 1428 if (_hr_printer.is_active()) {
1429 1429 // We should do this after we potentially resize the heap so
1430 1430 // that all the COMMIT / UNCOMMIT events are generated before
1431 1431 // the end GC event.
1432 1432
1433 1433 print_hrs_post_compaction();
1434 1434 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1435 1435 }
1436 1436
1437 1437 if (_cg1r->use_cache()) {
1438 1438 _cg1r->clear_and_record_card_counts();
1439 1439 _cg1r->clear_hot_cache();
1440 1440 }
1441 1441
1442 1442 // Rebuild remembered sets of all regions.
1443 1443 if (G1CollectedHeap::use_parallel_gc_threads()) {
1444 1444 uint n_workers =
1445 1445 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1446 1446 workers()->active_workers(),
1447 1447 Threads::number_of_non_daemon_threads());
1448 1448 assert(UseDynamicNumberOfGCThreads ||
1449 1449 n_workers == workers()->total_workers(),
1450 1450 "If not dynamic should be using all the workers");
1451 1451 workers()->set_active_workers(n_workers);
1452 1452 // Set parallel threads in the heap (_n_par_threads) only
1453 1453 // before a parallel phase and always reset it to 0 after
1454 1454 // the phase so that the number of parallel threads does
1455 1455 // no get carried forward to a serial phase where there
1456 1456 // may be code that is "possibly_parallel".
1457 1457 set_par_threads(n_workers);
1458 1458
1459 1459 ParRebuildRSTask rebuild_rs_task(this);
1460 1460 assert(check_heap_region_claim_values(
1461 1461 HeapRegion::InitialClaimValue), "sanity check");
1462 1462 assert(UseDynamicNumberOfGCThreads ||
1463 1463 workers()->active_workers() == workers()->total_workers(),
1464 1464 "Unless dynamic should use total workers");
1465 1465 // Use the most recent number of active workers
1466 1466 assert(workers()->active_workers() > 0,
1467 1467 "Active workers not properly set");
1468 1468 set_par_threads(workers()->active_workers());
1469 1469 workers()->run_task(&rebuild_rs_task);
1470 1470 set_par_threads(0);
1471 1471 assert(check_heap_region_claim_values(
1472 1472 HeapRegion::RebuildRSClaimValue), "sanity check");
1473 1473 reset_heap_region_claim_values();
1474 1474 } else {
1475 1475 RebuildRSOutOfRegionClosure rebuild_rs(this);
1476 1476 heap_region_iterate(&rebuild_rs);
1477 1477 }
1478 1478
1479 1479 if (G1Log::fine()) {
1480 1480 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1481 1481 }
1482 1482
1483 1483 if (true) { // FIXME
1484 1484 MetaspaceGC::compute_new_size();
1485 1485 }
1486 1486
1487 1487 // Start a new incremental collection set for the next pause
1488 1488 assert(g1_policy()->collection_set() == NULL, "must be");
1489 1489 g1_policy()->start_incremental_cset_building();
1490 1490
1491 1491 // Clear the _cset_fast_test bitmap in anticipation of adding
1492 1492 // regions to the incremental collection set for the next
1493 1493 // evacuation pause.
1494 1494 clear_cset_fast_test();
1495 1495
1496 1496 init_mutator_alloc_region();
1497 1497
1498 1498 double end = os::elapsedTime();
1499 1499 g1_policy()->record_full_collection_end();
1500 1500
1501 1501 #ifdef TRACESPINNING
1502 1502 ParallelTaskTerminator::print_termination_counts();
1503 1503 #endif
1504 1504
1505 1505 gc_epilogue(true);
1506 1506
1507 1507 // Discard all rset updates
1508 1508 JavaThread::dirty_card_queue_set().abandon_logs();
1509 1509 assert(!G1DeferredRSUpdate
1510 1510 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1511 1511
1512 1512 _young_list->reset_sampled_info();
1513 1513 // At this point there should be no regions in the
1514 1514 // entire heap tagged as young.
1515 1515 assert( check_young_list_empty(true /* check_heap */),
1516 1516 "young list should be empty at this point");
1517 1517
1518 1518 // Update the number of full collections that have been completed.
1519 1519 increment_old_marking_cycles_completed(false /* concurrent */);
1520 1520
1521 1521 _hrs.verify_optional();
1522 1522 verify_region_sets_optional();
1523 1523
1524 1524 print_heap_after_gc();
1525 1525
1526 1526 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1527 1527 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1528 1528 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1529 1529 // before any GC notifications are raised.
1530 1530 g1mm()->update_sizes();
1531 1531 }
1532 1532
1533 1533 post_full_gc_dump();
1534 1534
1535 1535 return true;
1536 1536 }
1537 1537
1538 1538 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1539 1539 // do_collection() will return whether it succeeded in performing
1540 1540 // the GC. Currently, there is no facility on the
1541 1541 // do_full_collection() API to notify the caller than the collection
1542 1542 // did not succeed (e.g., because it was locked out by the GC
1543 1543 // locker). So, right now, we'll ignore the return value.
1544 1544 bool dummy = do_collection(true, /* explicit_gc */
1545 1545 clear_all_soft_refs,
1546 1546 0 /* word_size */);
1547 1547 }
1548 1548
1549 1549 // This code is mostly copied from TenuredGeneration.
1550 1550 void
1551 1551 G1CollectedHeap::
1552 1552 resize_if_necessary_after_full_collection(size_t word_size) {
1553 1553 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1554 1554
1555 1555 // Include the current allocation, if any, and bytes that will be
1556 1556 // pre-allocated to support collections, as "used".
1557 1557 const size_t used_after_gc = used();
1558 1558 const size_t capacity_after_gc = capacity();
1559 1559 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1560 1560
1561 1561 // This is enforced in arguments.cpp.
1562 1562 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1563 1563 "otherwise the code below doesn't make sense");
1564 1564
1565 1565 // We don't have floating point command-line arguments
1566 1566 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1567 1567 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1568 1568 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1569 1569 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1570 1570
1571 1571 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1572 1572 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1573 1573
1574 1574 // We have to be careful here as these two calculations can overflow
1575 1575 // 32-bit size_t's.
1576 1576 double used_after_gc_d = (double) used_after_gc;
1577 1577 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1578 1578 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1579 1579
1580 1580 // Let's make sure that they are both under the max heap size, which
1581 1581 // by default will make them fit into a size_t.
1582 1582 double desired_capacity_upper_bound = (double) max_heap_size;
1583 1583 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1584 1584 desired_capacity_upper_bound);
1585 1585 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1586 1586 desired_capacity_upper_bound);
1587 1587
1588 1588 // We can now safely turn them into size_t's.
1589 1589 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1590 1590 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1591 1591
1592 1592 // This assert only makes sense here, before we adjust them
1593 1593 // with respect to the min and max heap size.
1594 1594 assert(minimum_desired_capacity <= maximum_desired_capacity,
1595 1595 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1596 1596 "maximum_desired_capacity = "SIZE_FORMAT,
1597 1597 minimum_desired_capacity, maximum_desired_capacity));
1598 1598
1599 1599 // Should not be greater than the heap max size. No need to adjust
1600 1600 // it with respect to the heap min size as it's a lower bound (i.e.,
1601 1601 // we'll try to make the capacity larger than it, not smaller).
1602 1602 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1603 1603 // Should not be less than the heap min size. No need to adjust it
1604 1604 // with respect to the heap max size as it's an upper bound (i.e.,
1605 1605 // we'll try to make the capacity smaller than it, not greater).
1606 1606 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1607 1607
1608 1608 if (capacity_after_gc < minimum_desired_capacity) {
1609 1609 // Don't expand unless it's significant
1610 1610 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1611 1611 ergo_verbose4(ErgoHeapSizing,
1612 1612 "attempt heap expansion",
1613 1613 ergo_format_reason("capacity lower than "
1614 1614 "min desired capacity after Full GC")
1615 1615 ergo_format_byte("capacity")
1616 1616 ergo_format_byte("occupancy")
1617 1617 ergo_format_byte_perc("min desired capacity"),
1618 1618 capacity_after_gc, used_after_gc,
1619 1619 minimum_desired_capacity, (double) MinHeapFreeRatio);
1620 1620 expand(expand_bytes);
1621 1621
1622 1622 // No expansion, now see if we want to shrink
1623 1623 } else if (capacity_after_gc > maximum_desired_capacity) {
1624 1624 // Capacity too large, compute shrinking size
1625 1625 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1626 1626 ergo_verbose4(ErgoHeapSizing,
1627 1627 "attempt heap shrinking",
1628 1628 ergo_format_reason("capacity higher than "
1629 1629 "max desired capacity after Full GC")
1630 1630 ergo_format_byte("capacity")
1631 1631 ergo_format_byte("occupancy")
1632 1632 ergo_format_byte_perc("max desired capacity"),
1633 1633 capacity_after_gc, used_after_gc,
1634 1634 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1635 1635 shrink(shrink_bytes);
1636 1636 }
1637 1637 }
1638 1638
1639 1639
1640 1640 HeapWord*
1641 1641 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1642 1642 bool* succeeded) {
1643 1643 assert_at_safepoint(true /* should_be_vm_thread */);
1644 1644
1645 1645 *succeeded = true;
1646 1646 // Let's attempt the allocation first.
1647 1647 HeapWord* result =
1648 1648 attempt_allocation_at_safepoint(word_size,
1649 1649 false /* expect_null_mutator_alloc_region */);
1650 1650 if (result != NULL) {
1651 1651 assert(*succeeded, "sanity");
1652 1652 return result;
1653 1653 }
1654 1654
1655 1655 // In a G1 heap, we're supposed to keep allocation from failing by
1656 1656 // incremental pauses. Therefore, at least for now, we'll favor
1657 1657 // expansion over collection. (This might change in the future if we can
1658 1658 // do something smarter than full collection to satisfy a failed alloc.)
1659 1659 result = expand_and_allocate(word_size);
1660 1660 if (result != NULL) {
1661 1661 assert(*succeeded, "sanity");
1662 1662 return result;
1663 1663 }
1664 1664
1665 1665 // Expansion didn't work, we'll try to do a Full GC.
1666 1666 bool gc_succeeded = do_collection(false, /* explicit_gc */
1667 1667 false, /* clear_all_soft_refs */
1668 1668 word_size);
1669 1669 if (!gc_succeeded) {
1670 1670 *succeeded = false;
1671 1671 return NULL;
1672 1672 }
1673 1673
1674 1674 // Retry the allocation
1675 1675 result = attempt_allocation_at_safepoint(word_size,
1676 1676 true /* expect_null_mutator_alloc_region */);
1677 1677 if (result != NULL) {
1678 1678 assert(*succeeded, "sanity");
1679 1679 return result;
1680 1680 }
1681 1681
1682 1682 // Then, try a Full GC that will collect all soft references.
1683 1683 gc_succeeded = do_collection(false, /* explicit_gc */
1684 1684 true, /* clear_all_soft_refs */
1685 1685 word_size);
1686 1686 if (!gc_succeeded) {
1687 1687 *succeeded = false;
1688 1688 return NULL;
1689 1689 }
1690 1690
1691 1691 // Retry the allocation once more
1692 1692 result = attempt_allocation_at_safepoint(word_size,
1693 1693 true /* expect_null_mutator_alloc_region */);
1694 1694 if (result != NULL) {
1695 1695 assert(*succeeded, "sanity");
1696 1696 return result;
1697 1697 }
1698 1698
1699 1699 assert(!collector_policy()->should_clear_all_soft_refs(),
1700 1700 "Flag should have been handled and cleared prior to this point");
1701 1701
1702 1702 // What else? We might try synchronous finalization later. If the total
1703 1703 // space available is large enough for the allocation, then a more
1704 1704 // complete compaction phase than we've tried so far might be
1705 1705 // appropriate.
1706 1706 assert(*succeeded, "sanity");
1707 1707 return NULL;
1708 1708 }
1709 1709
1710 1710 // Attempting to expand the heap sufficiently
1711 1711 // to support an allocation of the given "word_size". If
1712 1712 // successful, perform the allocation and return the address of the
1713 1713 // allocated block, or else "NULL".
1714 1714
1715 1715 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1716 1716 assert_at_safepoint(true /* should_be_vm_thread */);
1717 1717
1718 1718 verify_region_sets_optional();
1719 1719
1720 1720 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1721 1721 ergo_verbose1(ErgoHeapSizing,
1722 1722 "attempt heap expansion",
1723 1723 ergo_format_reason("allocation request failed")
1724 1724 ergo_format_byte("allocation request"),
1725 1725 word_size * HeapWordSize);
1726 1726 if (expand(expand_bytes)) {
1727 1727 _hrs.verify_optional();
1728 1728 verify_region_sets_optional();
1729 1729 return attempt_allocation_at_safepoint(word_size,
1730 1730 false /* expect_null_mutator_alloc_region */);
1731 1731 }
1732 1732 return NULL;
1733 1733 }
1734 1734
1735 1735 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1736 1736 HeapWord* new_end) {
1737 1737 assert(old_end != new_end, "don't call this otherwise");
1738 1738 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1739 1739
1740 1740 // Update the committed mem region.
1741 1741 _g1_committed.set_end(new_end);
1742 1742 // Tell the card table about the update.
1743 1743 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1744 1744 // Tell the BOT about the update.
1745 1745 _bot_shared->resize(_g1_committed.word_size());
1746 1746 }
1747 1747
1748 1748 bool G1CollectedHeap::expand(size_t expand_bytes) {
1749 1749 size_t old_mem_size = _g1_storage.committed_size();
1750 1750 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1751 1751 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1752 1752 HeapRegion::GrainBytes);
1753 1753 ergo_verbose2(ErgoHeapSizing,
1754 1754 "expand the heap",
1755 1755 ergo_format_byte("requested expansion amount")
1756 1756 ergo_format_byte("attempted expansion amount"),
1757 1757 expand_bytes, aligned_expand_bytes);
1758 1758
1759 1759 // First commit the memory.
1760 1760 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1761 1761 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1762 1762 if (successful) {
1763 1763 // Then propagate this update to the necessary data structures.
1764 1764 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1765 1765 update_committed_space(old_end, new_end);
1766 1766
1767 1767 FreeRegionList expansion_list("Local Expansion List");
1768 1768 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1769 1769 assert(mr.start() == old_end, "post-condition");
1770 1770 // mr might be a smaller region than what was requested if
1771 1771 // expand_by() was unable to allocate the HeapRegion instances
1772 1772 assert(mr.end() <= new_end, "post-condition");
1773 1773
1774 1774 size_t actual_expand_bytes = mr.byte_size();
1775 1775 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1776 1776 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1777 1777 "post-condition");
1778 1778 if (actual_expand_bytes < aligned_expand_bytes) {
1779 1779 // We could not expand _hrs to the desired size. In this case we
1780 1780 // need to shrink the committed space accordingly.
1781 1781 assert(mr.end() < new_end, "invariant");
1782 1782
1783 1783 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1784 1784 // First uncommit the memory.
1785 1785 _g1_storage.shrink_by(diff_bytes);
1786 1786 // Then propagate this update to the necessary data structures.
1787 1787 update_committed_space(new_end, mr.end());
1788 1788 }
1789 1789 _free_list.add_as_tail(&expansion_list);
1790 1790
1791 1791 if (_hr_printer.is_active()) {
1792 1792 HeapWord* curr = mr.start();
1793 1793 while (curr < mr.end()) {
1794 1794 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1795 1795 _hr_printer.commit(curr, curr_end);
1796 1796 curr = curr_end;
1797 1797 }
1798 1798 assert(curr == mr.end(), "post-condition");
1799 1799 }
1800 1800 g1_policy()->record_new_heap_size(n_regions());
1801 1801 } else {
1802 1802 ergo_verbose0(ErgoHeapSizing,
1803 1803 "did not expand the heap",
1804 1804 ergo_format_reason("heap expansion operation failed"));
1805 1805 // The expansion of the virtual storage space was unsuccessful.
1806 1806 // Let's see if it was because we ran out of swap.
1807 1807 if (G1ExitOnExpansionFailure &&
1808 1808 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1809 1809 // We had head room...
1810 1810 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1811 1811 }
1812 1812 }
1813 1813 return successful;
1814 1814 }
1815 1815
1816 1816 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1817 1817 size_t old_mem_size = _g1_storage.committed_size();
1818 1818 size_t aligned_shrink_bytes =
1819 1819 ReservedSpace::page_align_size_down(shrink_bytes);
1820 1820 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1821 1821 HeapRegion::GrainBytes);
1822 1822 uint num_regions_deleted = 0;
1823 1823 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1824 1824 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1825 1825 assert(mr.end() == old_end, "post-condition");
1826 1826
1827 1827 ergo_verbose3(ErgoHeapSizing,
1828 1828 "shrink the heap",
1829 1829 ergo_format_byte("requested shrinking amount")
1830 1830 ergo_format_byte("aligned shrinking amount")
1831 1831 ergo_format_byte("attempted shrinking amount"),
1832 1832 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1833 1833 if (mr.byte_size() > 0) {
1834 1834 if (_hr_printer.is_active()) {
1835 1835 HeapWord* curr = mr.end();
1836 1836 while (curr > mr.start()) {
1837 1837 HeapWord* curr_end = curr;
1838 1838 curr -= HeapRegion::GrainWords;
1839 1839 _hr_printer.uncommit(curr, curr_end);
1840 1840 }
1841 1841 assert(curr == mr.start(), "post-condition");
1842 1842 }
1843 1843
1844 1844 _g1_storage.shrink_by(mr.byte_size());
1845 1845 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1846 1846 assert(mr.start() == new_end, "post-condition");
1847 1847
1848 1848 _expansion_regions += num_regions_deleted;
1849 1849 update_committed_space(old_end, new_end);
1850 1850 HeapRegionRemSet::shrink_heap(n_regions());
1851 1851 g1_policy()->record_new_heap_size(n_regions());
1852 1852 } else {
1853 1853 ergo_verbose0(ErgoHeapSizing,
1854 1854 "did not shrink the heap",
1855 1855 ergo_format_reason("heap shrinking operation failed"));
1856 1856 }
1857 1857 }
1858 1858
1859 1859 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1860 1860 verify_region_sets_optional();
1861 1861
1862 1862 // We should only reach here at the end of a Full GC which means we
1863 1863 // should not not be holding to any GC alloc regions. The method
1864 1864 // below will make sure of that and do any remaining clean up.
1865 1865 abandon_gc_alloc_regions();
1866 1866
1867 1867 // Instead of tearing down / rebuilding the free lists here, we
1868 1868 // could instead use the remove_all_pending() method on free_list to
1869 1869 // remove only the ones that we need to remove.
1870 1870 tear_down_region_sets(true /* free_list_only */);
1871 1871 shrink_helper(shrink_bytes);
1872 1872 rebuild_region_sets(true /* free_list_only */);
1873 1873
1874 1874 _hrs.verify_optional();
1875 1875 verify_region_sets_optional();
1876 1876 }
1877 1877
1878 1878 // Public methods.
1879 1879
1880 1880 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1881 1881 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1882 1882 #endif // _MSC_VER
1883 1883
1884 1884
1885 1885 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1886 1886 SharedHeap(policy_),
1887 1887 _g1_policy(policy_),
1888 1888 _dirty_card_queue_set(false),
1889 1889 _into_cset_dirty_card_queue_set(false),
1890 1890 _is_alive_closure_cm(this),
1891 1891 _is_alive_closure_stw(this),
1892 1892 _ref_processor_cm(NULL),
1893 1893 _ref_processor_stw(NULL),
1894 1894 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1895 1895 _bot_shared(NULL),
1896 1896 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1897 1897 _evac_failure_scan_stack(NULL) ,
1898 1898 _mark_in_progress(false),
1899 1899 _cg1r(NULL), _summary_bytes_used(0),
1900 1900 _g1mm(NULL),
1901 1901 _refine_cte_cl(NULL),
1902 1902 _full_collection(false),
1903 1903 _free_list("Master Free List"),
1904 1904 _secondary_free_list("Secondary Free List"),
1905 1905 _old_set("Old Set"),
1906 1906 _humongous_set("Master Humongous Set"),
1907 1907 _free_regions_coming(false),
1908 1908 _young_list(new YoungList(this)),
1909 1909 _gc_time_stamp(0),
1910 1910 _retained_old_gc_alloc_region(NULL),
1911 1911 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1912 1912 _old_plab_stats(OldPLABSize, PLABWeight),
1913 1913 _expand_heap_after_alloc_failure(true),
1914 1914 _surviving_young_words(NULL),
1915 1915 _old_marking_cycles_started(0),
1916 1916 _old_marking_cycles_completed(0),
1917 1917 _in_cset_fast_test(NULL),
1918 1918 _in_cset_fast_test_base(NULL),
1919 1919 _dirty_cards_region_list(NULL),
1920 1920 _worker_cset_start_region(NULL),
1921 1921 _worker_cset_start_region_time_stamp(NULL) {
1922 1922 _g1h = this; // To catch bugs.
1923 1923 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1924 1924 vm_exit_during_initialization("Failed necessary allocation.");
1925 1925 }
1926 1926
1927 1927 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1928 1928
1929 1929 int n_queues = MAX2((int)ParallelGCThreads, 1);
1930 1930 _task_queues = new RefToScanQueueSet(n_queues);
1931 1931
1932 1932 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1933 1933 assert(n_rem_sets > 0, "Invariant.");
1934 1934
1935 1935 HeapRegionRemSetIterator** iter_arr =
1936 1936 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC);
1937 1937 for (int i = 0; i < n_queues; i++) {
1938 1938 iter_arr[i] = new HeapRegionRemSetIterator();
1939 1939 }
1940 1940 _rem_set_iterator = iter_arr;
1941 1941
1942 1942 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1943 1943 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1944 1944
1945 1945 for (int i = 0; i < n_queues; i++) {
1946 1946 RefToScanQueue* q = new RefToScanQueue();
|
↓ open down ↓ |
1946 lines elided |
↑ open up ↑ |
1947 1947 q->initialize();
1948 1948 _task_queues->register_queue(i, q);
1949 1949 }
1950 1950
1951 1951 clear_cset_start_regions();
1952 1952
1953 1953 // Initialize the G1EvacuationFailureALot counters and flags.
1954 1954 NOT_PRODUCT(reset_evacuation_should_fail();)
1955 1955
1956 1956 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1957 -#ifdef SPARC
1958 - // Issue a stern warning, but allow use for experimentation and debugging.
1959 - if (VM_Version::is_sun4v() && UseMemSetInBOT) {
1960 - assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
1961 - warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
1962 - " on sun4v; please understand that you are using at your own risk!");
1963 - }
1964 -#endif
1965 1957 }
1966 1958
1967 1959 jint G1CollectedHeap::initialize() {
1968 1960 CollectedHeap::pre_initialize();
1969 1961 os::enable_vtime();
1970 1962
1971 1963 G1Log::init();
1972 1964
1973 1965 // Necessary to satisfy locking discipline assertions.
1974 1966
1975 1967 MutexLocker x(Heap_lock);
1976 1968
1977 1969 // We have to initialize the printer before committing the heap, as
1978 1970 // it will be used then.
1979 1971 _hr_printer.set_active(G1PrintHeapRegions);
1980 1972
1981 1973 // While there are no constraints in the GC code that HeapWordSize
1982 1974 // be any particular value, there are multiple other areas in the
1983 1975 // system which believe this to be true (e.g. oop->object_size in some
1984 1976 // cases incorrectly returns the size in wordSize units rather than
1985 1977 // HeapWordSize).
1986 1978 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1987 1979
1988 1980 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1989 1981 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1990 1982
1991 1983 // Ensure that the sizes are properly aligned.
1992 1984 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1993 1985 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1994 1986
1995 1987 _cg1r = new ConcurrentG1Refine();
1996 1988
1997 1989 // Reserve the maximum.
1998 1990
1999 1991 // When compressed oops are enabled, the preferred heap base
2000 1992 // is calculated by subtracting the requested size from the
2001 1993 // 32Gb boundary and using the result as the base address for
2002 1994 // heap reservation. If the requested size is not aligned to
2003 1995 // HeapRegion::GrainBytes (i.e. the alignment that is passed
2004 1996 // into the ReservedHeapSpace constructor) then the actual
2005 1997 // base of the reserved heap may end up differing from the
2006 1998 // address that was requested (i.e. the preferred heap base).
2007 1999 // If this happens then we could end up using a non-optimal
2008 2000 // compressed oops mode.
2009 2001
2010 2002 // Since max_byte_size is aligned to the size of a heap region (checked
2011 2003 // above).
2012 2004 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2013 2005
2014 2006 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2015 2007 HeapRegion::GrainBytes);
2016 2008
2017 2009 // It is important to do this in a way such that concurrent readers can't
2018 2010 // temporarily think somethings in the heap. (I've actually seen this
2019 2011 // happen in asserts: DLD.)
2020 2012 _reserved.set_word_size(0);
2021 2013 _reserved.set_start((HeapWord*)heap_rs.base());
2022 2014 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2023 2015
2024 2016 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2025 2017
2026 2018 // Create the gen rem set (and barrier set) for the entire reserved region.
2027 2019 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2028 2020 set_barrier_set(rem_set()->bs());
2029 2021 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2030 2022 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2031 2023 } else {
2032 2024 vm_exit_during_initialization("G1 requires a mod ref bs.");
2033 2025 return JNI_ENOMEM;
2034 2026 }
2035 2027
2036 2028 // Also create a G1 rem set.
2037 2029 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2038 2030 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2039 2031 } else {
2040 2032 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2041 2033 return JNI_ENOMEM;
2042 2034 }
2043 2035
2044 2036 // Carve out the G1 part of the heap.
2045 2037
2046 2038 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2047 2039 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2048 2040 g1_rs.size()/HeapWordSize);
2049 2041
2050 2042 _g1_storage.initialize(g1_rs, 0);
2051 2043 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2052 2044 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2053 2045 (HeapWord*) _g1_reserved.end(),
2054 2046 _expansion_regions);
2055 2047
2056 2048 // 6843694 - ensure that the maximum region index can fit
2057 2049 // in the remembered set structures.
2058 2050 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2059 2051 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2060 2052
2061 2053 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2062 2054 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2063 2055 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2064 2056 "too many cards per region");
2065 2057
2066 2058 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2067 2059
2068 2060 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2069 2061 heap_word_size(init_byte_size));
2070 2062
2071 2063 _g1h = this;
2072 2064
2073 2065 _in_cset_fast_test_length = max_regions();
2074 2066 _in_cset_fast_test_base =
2075 2067 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2076 2068
2077 2069 // We're biasing _in_cset_fast_test to avoid subtracting the
2078 2070 // beginning of the heap every time we want to index; basically
2079 2071 // it's the same with what we do with the card table.
2080 2072 _in_cset_fast_test = _in_cset_fast_test_base -
2081 2073 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2082 2074
2083 2075 // Clear the _cset_fast_test bitmap in anticipation of adding
2084 2076 // regions to the incremental collection set for the first
2085 2077 // evacuation pause.
2086 2078 clear_cset_fast_test();
2087 2079
2088 2080 // Create the ConcurrentMark data structure and thread.
2089 2081 // (Must do this late, so that "max_regions" is defined.)
2090 2082 _cm = new ConcurrentMark(heap_rs, max_regions());
2091 2083 _cmThread = _cm->cmThread();
2092 2084
2093 2085 // Initialize the from_card cache structure of HeapRegionRemSet.
2094 2086 HeapRegionRemSet::init_heap(max_regions());
2095 2087
2096 2088 // Now expand into the initial heap size.
2097 2089 if (!expand(init_byte_size)) {
2098 2090 vm_exit_during_initialization("Failed to allocate initial heap.");
2099 2091 return JNI_ENOMEM;
2100 2092 }
2101 2093
2102 2094 // Perform any initialization actions delegated to the policy.
2103 2095 g1_policy()->init();
2104 2096
2105 2097 _refine_cte_cl =
2106 2098 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2107 2099 g1_rem_set(),
2108 2100 concurrent_g1_refine());
2109 2101 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2110 2102
2111 2103 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2112 2104 SATB_Q_FL_lock,
2113 2105 G1SATBProcessCompletedThreshold,
2114 2106 Shared_SATB_Q_lock);
2115 2107
2116 2108 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2117 2109 DirtyCardQ_FL_lock,
2118 2110 concurrent_g1_refine()->yellow_zone(),
2119 2111 concurrent_g1_refine()->red_zone(),
2120 2112 Shared_DirtyCardQ_lock);
2121 2113
2122 2114 if (G1DeferredRSUpdate) {
2123 2115 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2124 2116 DirtyCardQ_FL_lock,
2125 2117 -1, // never trigger processing
2126 2118 -1, // no limit on length
2127 2119 Shared_DirtyCardQ_lock,
2128 2120 &JavaThread::dirty_card_queue_set());
2129 2121 }
2130 2122
2131 2123 // Initialize the card queue set used to hold cards containing
2132 2124 // references into the collection set.
2133 2125 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2134 2126 DirtyCardQ_FL_lock,
2135 2127 -1, // never trigger processing
2136 2128 -1, // no limit on length
2137 2129 Shared_DirtyCardQ_lock,
2138 2130 &JavaThread::dirty_card_queue_set());
2139 2131
2140 2132 // In case we're keeping closure specialization stats, initialize those
2141 2133 // counts and that mechanism.
2142 2134 SpecializationStats::clear();
2143 2135
2144 2136 // Do later initialization work for concurrent refinement.
2145 2137 _cg1r->init();
2146 2138
2147 2139 // Here we allocate the dummy full region that is required by the
2148 2140 // G1AllocRegion class. If we don't pass an address in the reserved
2149 2141 // space here, lots of asserts fire.
2150 2142
2151 2143 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2152 2144 _g1_reserved.start());
2153 2145 // We'll re-use the same region whether the alloc region will
2154 2146 // require BOT updates or not and, if it doesn't, then a non-young
2155 2147 // region will complain that it cannot support allocations without
2156 2148 // BOT updates. So we'll tag the dummy region as young to avoid that.
2157 2149 dummy_region->set_young();
2158 2150 // Make sure it's full.
2159 2151 dummy_region->set_top(dummy_region->end());
2160 2152 G1AllocRegion::setup(this, dummy_region);
2161 2153
2162 2154 init_mutator_alloc_region();
2163 2155
2164 2156 // Do create of the monitoring and management support so that
2165 2157 // values in the heap have been properly initialized.
2166 2158 _g1mm = new G1MonitoringSupport(this);
2167 2159
2168 2160 return JNI_OK;
2169 2161 }
2170 2162
2171 2163 void G1CollectedHeap::ref_processing_init() {
2172 2164 // Reference processing in G1 currently works as follows:
2173 2165 //
2174 2166 // * There are two reference processor instances. One is
2175 2167 // used to record and process discovered references
2176 2168 // during concurrent marking; the other is used to
2177 2169 // record and process references during STW pauses
2178 2170 // (both full and incremental).
2179 2171 // * Both ref processors need to 'span' the entire heap as
2180 2172 // the regions in the collection set may be dotted around.
2181 2173 //
2182 2174 // * For the concurrent marking ref processor:
2183 2175 // * Reference discovery is enabled at initial marking.
2184 2176 // * Reference discovery is disabled and the discovered
2185 2177 // references processed etc during remarking.
2186 2178 // * Reference discovery is MT (see below).
2187 2179 // * Reference discovery requires a barrier (see below).
2188 2180 // * Reference processing may or may not be MT
2189 2181 // (depending on the value of ParallelRefProcEnabled
2190 2182 // and ParallelGCThreads).
2191 2183 // * A full GC disables reference discovery by the CM
2192 2184 // ref processor and abandons any entries on it's
2193 2185 // discovered lists.
2194 2186 //
2195 2187 // * For the STW processor:
2196 2188 // * Non MT discovery is enabled at the start of a full GC.
2197 2189 // * Processing and enqueueing during a full GC is non-MT.
2198 2190 // * During a full GC, references are processed after marking.
2199 2191 //
2200 2192 // * Discovery (may or may not be MT) is enabled at the start
2201 2193 // of an incremental evacuation pause.
2202 2194 // * References are processed near the end of a STW evacuation pause.
2203 2195 // * For both types of GC:
2204 2196 // * Discovery is atomic - i.e. not concurrent.
2205 2197 // * Reference discovery will not need a barrier.
2206 2198
2207 2199 SharedHeap::ref_processing_init();
2208 2200 MemRegion mr = reserved_region();
2209 2201
2210 2202 // Concurrent Mark ref processor
2211 2203 _ref_processor_cm =
2212 2204 new ReferenceProcessor(mr, // span
2213 2205 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2214 2206 // mt processing
2215 2207 (int) ParallelGCThreads,
2216 2208 // degree of mt processing
2217 2209 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2218 2210 // mt discovery
2219 2211 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2220 2212 // degree of mt discovery
2221 2213 false,
2222 2214 // Reference discovery is not atomic
2223 2215 &_is_alive_closure_cm,
2224 2216 // is alive closure
2225 2217 // (for efficiency/performance)
2226 2218 true);
2227 2219 // Setting next fields of discovered
2228 2220 // lists requires a barrier.
2229 2221
2230 2222 // STW ref processor
2231 2223 _ref_processor_stw =
2232 2224 new ReferenceProcessor(mr, // span
2233 2225 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2234 2226 // mt processing
2235 2227 MAX2((int)ParallelGCThreads, 1),
2236 2228 // degree of mt processing
2237 2229 (ParallelGCThreads > 1),
2238 2230 // mt discovery
2239 2231 MAX2((int)ParallelGCThreads, 1),
2240 2232 // degree of mt discovery
2241 2233 true,
2242 2234 // Reference discovery is atomic
2243 2235 &_is_alive_closure_stw,
2244 2236 // is alive closure
2245 2237 // (for efficiency/performance)
2246 2238 false);
2247 2239 // Setting next fields of discovered
2248 2240 // lists requires a barrier.
2249 2241 }
2250 2242
2251 2243 size_t G1CollectedHeap::capacity() const {
2252 2244 return _g1_committed.byte_size();
2253 2245 }
2254 2246
2255 2247 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2256 2248 assert(!hr->continuesHumongous(), "pre-condition");
2257 2249 hr->reset_gc_time_stamp();
2258 2250 if (hr->startsHumongous()) {
2259 2251 uint first_index = hr->hrs_index() + 1;
2260 2252 uint last_index = hr->last_hc_index();
2261 2253 for (uint i = first_index; i < last_index; i += 1) {
2262 2254 HeapRegion* chr = region_at(i);
2263 2255 assert(chr->continuesHumongous(), "sanity");
2264 2256 chr->reset_gc_time_stamp();
2265 2257 }
2266 2258 }
2267 2259 }
2268 2260
2269 2261 #ifndef PRODUCT
2270 2262 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2271 2263 private:
2272 2264 unsigned _gc_time_stamp;
2273 2265 bool _failures;
2274 2266
2275 2267 public:
2276 2268 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2277 2269 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2278 2270
2279 2271 virtual bool doHeapRegion(HeapRegion* hr) {
2280 2272 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2281 2273 if (_gc_time_stamp != region_gc_time_stamp) {
2282 2274 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2283 2275 "expected %d", HR_FORMAT_PARAMS(hr),
2284 2276 region_gc_time_stamp, _gc_time_stamp);
2285 2277 _failures = true;
2286 2278 }
2287 2279 return false;
2288 2280 }
2289 2281
2290 2282 bool failures() { return _failures; }
2291 2283 };
2292 2284
2293 2285 void G1CollectedHeap::check_gc_time_stamps() {
2294 2286 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2295 2287 heap_region_iterate(&cl);
2296 2288 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2297 2289 }
2298 2290 #endif // PRODUCT
2299 2291
2300 2292 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2301 2293 DirtyCardQueue* into_cset_dcq,
2302 2294 bool concurrent,
2303 2295 int worker_i) {
2304 2296 // Clean cards in the hot card cache
2305 2297 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2306 2298
2307 2299 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2308 2300 int n_completed_buffers = 0;
2309 2301 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2310 2302 n_completed_buffers++;
2311 2303 }
2312 2304 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2313 2305 dcqs.clear_n_completed_buffers();
2314 2306 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2315 2307 }
2316 2308
2317 2309
2318 2310 // Computes the sum of the storage used by the various regions.
2319 2311
2320 2312 size_t G1CollectedHeap::used() const {
2321 2313 assert(Heap_lock->owner() != NULL,
2322 2314 "Should be owned on this thread's behalf.");
2323 2315 size_t result = _summary_bytes_used;
2324 2316 // Read only once in case it is set to NULL concurrently
2325 2317 HeapRegion* hr = _mutator_alloc_region.get();
2326 2318 if (hr != NULL)
2327 2319 result += hr->used();
2328 2320 return result;
2329 2321 }
2330 2322
2331 2323 size_t G1CollectedHeap::used_unlocked() const {
2332 2324 size_t result = _summary_bytes_used;
2333 2325 return result;
2334 2326 }
2335 2327
2336 2328 class SumUsedClosure: public HeapRegionClosure {
2337 2329 size_t _used;
2338 2330 public:
2339 2331 SumUsedClosure() : _used(0) {}
2340 2332 bool doHeapRegion(HeapRegion* r) {
2341 2333 if (!r->continuesHumongous()) {
2342 2334 _used += r->used();
2343 2335 }
2344 2336 return false;
2345 2337 }
2346 2338 size_t result() { return _used; }
2347 2339 };
2348 2340
2349 2341 size_t G1CollectedHeap::recalculate_used() const {
2350 2342 SumUsedClosure blk;
2351 2343 heap_region_iterate(&blk);
2352 2344 return blk.result();
2353 2345 }
2354 2346
2355 2347 size_t G1CollectedHeap::unsafe_max_alloc() {
2356 2348 if (free_regions() > 0) return HeapRegion::GrainBytes;
2357 2349 // otherwise, is there space in the current allocation region?
2358 2350
2359 2351 // We need to store the current allocation region in a local variable
2360 2352 // here. The problem is that this method doesn't take any locks and
2361 2353 // there may be other threads which overwrite the current allocation
2362 2354 // region field. attempt_allocation(), for example, sets it to NULL
2363 2355 // and this can happen *after* the NULL check here but before the call
2364 2356 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2365 2357 // to be a problem in the optimized build, since the two loads of the
2366 2358 // current allocation region field are optimized away.
2367 2359 HeapRegion* hr = _mutator_alloc_region.get();
2368 2360 if (hr == NULL) {
2369 2361 return 0;
2370 2362 }
2371 2363 return hr->free();
2372 2364 }
2373 2365
2374 2366 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2375 2367 switch (cause) {
2376 2368 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2377 2369 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2378 2370 case GCCause::_g1_humongous_allocation: return true;
2379 2371 default: return false;
2380 2372 }
2381 2373 }
2382 2374
2383 2375 #ifndef PRODUCT
2384 2376 void G1CollectedHeap::allocate_dummy_regions() {
2385 2377 // Let's fill up most of the region
2386 2378 size_t word_size = HeapRegion::GrainWords - 1024;
2387 2379 // And as a result the region we'll allocate will be humongous.
2388 2380 guarantee(isHumongous(word_size), "sanity");
2389 2381
2390 2382 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2391 2383 // Let's use the existing mechanism for the allocation
2392 2384 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2393 2385 if (dummy_obj != NULL) {
2394 2386 MemRegion mr(dummy_obj, word_size);
2395 2387 CollectedHeap::fill_with_object(mr);
2396 2388 } else {
2397 2389 // If we can't allocate once, we probably cannot allocate
2398 2390 // again. Let's get out of the loop.
2399 2391 break;
2400 2392 }
2401 2393 }
2402 2394 }
2403 2395 #endif // !PRODUCT
2404 2396
2405 2397 void G1CollectedHeap::increment_old_marking_cycles_started() {
2406 2398 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2407 2399 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2408 2400 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2409 2401 _old_marking_cycles_started, _old_marking_cycles_completed));
2410 2402
2411 2403 _old_marking_cycles_started++;
2412 2404 }
2413 2405
2414 2406 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2415 2407 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2416 2408
2417 2409 // We assume that if concurrent == true, then the caller is a
2418 2410 // concurrent thread that was joined the Suspendible Thread
2419 2411 // Set. If there's ever a cheap way to check this, we should add an
2420 2412 // assert here.
2421 2413
2422 2414 // Given that this method is called at the end of a Full GC or of a
2423 2415 // concurrent cycle, and those can be nested (i.e., a Full GC can
2424 2416 // interrupt a concurrent cycle), the number of full collections
2425 2417 // completed should be either one (in the case where there was no
2426 2418 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2427 2419 // behind the number of full collections started.
2428 2420
2429 2421 // This is the case for the inner caller, i.e. a Full GC.
2430 2422 assert(concurrent ||
2431 2423 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2432 2424 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2433 2425 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2434 2426 "is inconsistent with _old_marking_cycles_completed = %u",
2435 2427 _old_marking_cycles_started, _old_marking_cycles_completed));
2436 2428
2437 2429 // This is the case for the outer caller, i.e. the concurrent cycle.
2438 2430 assert(!concurrent ||
2439 2431 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2440 2432 err_msg("for outer caller (concurrent cycle): "
2441 2433 "_old_marking_cycles_started = %u "
2442 2434 "is inconsistent with _old_marking_cycles_completed = %u",
2443 2435 _old_marking_cycles_started, _old_marking_cycles_completed));
2444 2436
2445 2437 _old_marking_cycles_completed += 1;
2446 2438
2447 2439 // We need to clear the "in_progress" flag in the CM thread before
2448 2440 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2449 2441 // is set) so that if a waiter requests another System.gc() it doesn't
2450 2442 // incorrectly see that a marking cyle is still in progress.
2451 2443 if (concurrent) {
2452 2444 _cmThread->clear_in_progress();
2453 2445 }
2454 2446
2455 2447 // This notify_all() will ensure that a thread that called
2456 2448 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2457 2449 // and it's waiting for a full GC to finish will be woken up. It is
2458 2450 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2459 2451 FullGCCount_lock->notify_all();
2460 2452 }
2461 2453
2462 2454 void G1CollectedHeap::collect(GCCause::Cause cause) {
2463 2455 assert_heap_not_locked();
2464 2456
2465 2457 unsigned int gc_count_before;
2466 2458 unsigned int old_marking_count_before;
2467 2459 bool retry_gc;
2468 2460
2469 2461 do {
2470 2462 retry_gc = false;
2471 2463
2472 2464 {
2473 2465 MutexLocker ml(Heap_lock);
2474 2466
2475 2467 // Read the GC count while holding the Heap_lock
2476 2468 gc_count_before = total_collections();
2477 2469 old_marking_count_before = _old_marking_cycles_started;
2478 2470 }
2479 2471
2480 2472 if (should_do_concurrent_full_gc(cause)) {
2481 2473 // Schedule an initial-mark evacuation pause that will start a
2482 2474 // concurrent cycle. We're setting word_size to 0 which means that
2483 2475 // we are not requesting a post-GC allocation.
2484 2476 VM_G1IncCollectionPause op(gc_count_before,
2485 2477 0, /* word_size */
2486 2478 true, /* should_initiate_conc_mark */
2487 2479 g1_policy()->max_pause_time_ms(),
2488 2480 cause);
2489 2481
2490 2482 VMThread::execute(&op);
2491 2483 if (!op.pause_succeeded()) {
2492 2484 if (old_marking_count_before == _old_marking_cycles_started) {
2493 2485 retry_gc = op.should_retry_gc();
2494 2486 } else {
2495 2487 // A Full GC happened while we were trying to schedule the
2496 2488 // initial-mark GC. No point in starting a new cycle given
2497 2489 // that the whole heap was collected anyway.
2498 2490 }
2499 2491
2500 2492 if (retry_gc) {
2501 2493 if (GC_locker::is_active_and_needs_gc()) {
2502 2494 GC_locker::stall_until_clear();
2503 2495 }
2504 2496 }
2505 2497 }
2506 2498 } else {
2507 2499 if (cause == GCCause::_gc_locker
2508 2500 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2509 2501
2510 2502 // Schedule a standard evacuation pause. We're setting word_size
2511 2503 // to 0 which means that we are not requesting a post-GC allocation.
2512 2504 VM_G1IncCollectionPause op(gc_count_before,
2513 2505 0, /* word_size */
2514 2506 false, /* should_initiate_conc_mark */
2515 2507 g1_policy()->max_pause_time_ms(),
2516 2508 cause);
2517 2509 VMThread::execute(&op);
2518 2510 } else {
2519 2511 // Schedule a Full GC.
2520 2512 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2521 2513 VMThread::execute(&op);
2522 2514 }
2523 2515 }
2524 2516 } while (retry_gc);
2525 2517 }
2526 2518
2527 2519 bool G1CollectedHeap::is_in(const void* p) const {
2528 2520 if (_g1_committed.contains(p)) {
2529 2521 // Given that we know that p is in the committed space,
2530 2522 // heap_region_containing_raw() should successfully
2531 2523 // return the containing region.
2532 2524 HeapRegion* hr = heap_region_containing_raw(p);
2533 2525 return hr->is_in(p);
2534 2526 } else {
2535 2527 return false;
2536 2528 }
2537 2529 }
2538 2530
2539 2531 // Iteration functions.
2540 2532
2541 2533 // Iterates an OopClosure over all ref-containing fields of objects
2542 2534 // within a HeapRegion.
2543 2535
2544 2536 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2545 2537 MemRegion _mr;
2546 2538 ExtendedOopClosure* _cl;
2547 2539 public:
2548 2540 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2549 2541 : _mr(mr), _cl(cl) {}
2550 2542 bool doHeapRegion(HeapRegion* r) {
2551 2543 if (!r->continuesHumongous()) {
2552 2544 r->oop_iterate(_cl);
2553 2545 }
2554 2546 return false;
2555 2547 }
2556 2548 };
2557 2549
2558 2550 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2559 2551 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2560 2552 heap_region_iterate(&blk);
2561 2553 }
2562 2554
2563 2555 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2564 2556 IterateOopClosureRegionClosure blk(mr, cl);
2565 2557 heap_region_iterate(&blk);
2566 2558 }
2567 2559
2568 2560 // Iterates an ObjectClosure over all objects within a HeapRegion.
2569 2561
2570 2562 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2571 2563 ObjectClosure* _cl;
2572 2564 public:
2573 2565 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2574 2566 bool doHeapRegion(HeapRegion* r) {
2575 2567 if (! r->continuesHumongous()) {
2576 2568 r->object_iterate(_cl);
2577 2569 }
2578 2570 return false;
2579 2571 }
2580 2572 };
2581 2573
2582 2574 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2583 2575 IterateObjectClosureRegionClosure blk(cl);
2584 2576 heap_region_iterate(&blk);
2585 2577 }
2586 2578
2587 2579 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2588 2580 // FIXME: is this right?
2589 2581 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2590 2582 }
2591 2583
2592 2584 // Calls a SpaceClosure on a HeapRegion.
2593 2585
2594 2586 class SpaceClosureRegionClosure: public HeapRegionClosure {
2595 2587 SpaceClosure* _cl;
2596 2588 public:
2597 2589 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2598 2590 bool doHeapRegion(HeapRegion* r) {
2599 2591 _cl->do_space(r);
2600 2592 return false;
2601 2593 }
2602 2594 };
2603 2595
2604 2596 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2605 2597 SpaceClosureRegionClosure blk(cl);
2606 2598 heap_region_iterate(&blk);
2607 2599 }
2608 2600
2609 2601 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2610 2602 _hrs.iterate(cl);
2611 2603 }
2612 2604
2613 2605 void
2614 2606 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2615 2607 uint worker_id,
2616 2608 uint no_of_par_workers,
2617 2609 jint claim_value) {
2618 2610 const uint regions = n_regions();
2619 2611 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2620 2612 no_of_par_workers :
2621 2613 1);
2622 2614 assert(UseDynamicNumberOfGCThreads ||
2623 2615 no_of_par_workers == workers()->total_workers(),
2624 2616 "Non dynamic should use fixed number of workers");
2625 2617 // try to spread out the starting points of the workers
2626 2618 const HeapRegion* start_hr =
2627 2619 start_region_for_worker(worker_id, no_of_par_workers);
2628 2620 const uint start_index = start_hr->hrs_index();
2629 2621
2630 2622 // each worker will actually look at all regions
2631 2623 for (uint count = 0; count < regions; ++count) {
2632 2624 const uint index = (start_index + count) % regions;
2633 2625 assert(0 <= index && index < regions, "sanity");
2634 2626 HeapRegion* r = region_at(index);
2635 2627 // we'll ignore "continues humongous" regions (we'll process them
2636 2628 // when we come across their corresponding "start humongous"
2637 2629 // region) and regions already claimed
2638 2630 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2639 2631 continue;
2640 2632 }
2641 2633 // OK, try to claim it
2642 2634 if (r->claimHeapRegion(claim_value)) {
2643 2635 // success!
2644 2636 assert(!r->continuesHumongous(), "sanity");
2645 2637 if (r->startsHumongous()) {
2646 2638 // If the region is "starts humongous" we'll iterate over its
2647 2639 // "continues humongous" first; in fact we'll do them
2648 2640 // first. The order is important. In on case, calling the
2649 2641 // closure on the "starts humongous" region might de-allocate
2650 2642 // and clear all its "continues humongous" regions and, as a
2651 2643 // result, we might end up processing them twice. So, we'll do
2652 2644 // them first (notice: most closures will ignore them anyway) and
2653 2645 // then we'll do the "starts humongous" region.
2654 2646 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2655 2647 HeapRegion* chr = region_at(ch_index);
2656 2648
2657 2649 // if the region has already been claimed or it's not
2658 2650 // "continues humongous" we're done
2659 2651 if (chr->claim_value() == claim_value ||
2660 2652 !chr->continuesHumongous()) {
2661 2653 break;
2662 2654 }
2663 2655
2664 2656 // Noone should have claimed it directly. We can given
2665 2657 // that we claimed its "starts humongous" region.
2666 2658 assert(chr->claim_value() != claim_value, "sanity");
2667 2659 assert(chr->humongous_start_region() == r, "sanity");
2668 2660
2669 2661 if (chr->claimHeapRegion(claim_value)) {
2670 2662 // we should always be able to claim it; noone else should
2671 2663 // be trying to claim this region
2672 2664
2673 2665 bool res2 = cl->doHeapRegion(chr);
2674 2666 assert(!res2, "Should not abort");
2675 2667
2676 2668 // Right now, this holds (i.e., no closure that actually
2677 2669 // does something with "continues humongous" regions
2678 2670 // clears them). We might have to weaken it in the future,
2679 2671 // but let's leave these two asserts here for extra safety.
2680 2672 assert(chr->continuesHumongous(), "should still be the case");
2681 2673 assert(chr->humongous_start_region() == r, "sanity");
2682 2674 } else {
2683 2675 guarantee(false, "we should not reach here");
2684 2676 }
2685 2677 }
2686 2678 }
2687 2679
2688 2680 assert(!r->continuesHumongous(), "sanity");
2689 2681 bool res = cl->doHeapRegion(r);
2690 2682 assert(!res, "Should not abort");
2691 2683 }
2692 2684 }
2693 2685 }
2694 2686
2695 2687 class ResetClaimValuesClosure: public HeapRegionClosure {
2696 2688 public:
2697 2689 bool doHeapRegion(HeapRegion* r) {
2698 2690 r->set_claim_value(HeapRegion::InitialClaimValue);
2699 2691 return false;
2700 2692 }
2701 2693 };
2702 2694
2703 2695 void G1CollectedHeap::reset_heap_region_claim_values() {
2704 2696 ResetClaimValuesClosure blk;
2705 2697 heap_region_iterate(&blk);
2706 2698 }
2707 2699
2708 2700 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2709 2701 ResetClaimValuesClosure blk;
2710 2702 collection_set_iterate(&blk);
2711 2703 }
2712 2704
2713 2705 #ifdef ASSERT
2714 2706 // This checks whether all regions in the heap have the correct claim
2715 2707 // value. I also piggy-backed on this a check to ensure that the
2716 2708 // humongous_start_region() information on "continues humongous"
2717 2709 // regions is correct.
2718 2710
2719 2711 class CheckClaimValuesClosure : public HeapRegionClosure {
2720 2712 private:
2721 2713 jint _claim_value;
2722 2714 uint _failures;
2723 2715 HeapRegion* _sh_region;
2724 2716
2725 2717 public:
2726 2718 CheckClaimValuesClosure(jint claim_value) :
2727 2719 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2728 2720 bool doHeapRegion(HeapRegion* r) {
2729 2721 if (r->claim_value() != _claim_value) {
2730 2722 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2731 2723 "claim value = %d, should be %d",
2732 2724 HR_FORMAT_PARAMS(r),
2733 2725 r->claim_value(), _claim_value);
2734 2726 ++_failures;
2735 2727 }
2736 2728 if (!r->isHumongous()) {
2737 2729 _sh_region = NULL;
2738 2730 } else if (r->startsHumongous()) {
2739 2731 _sh_region = r;
2740 2732 } else if (r->continuesHumongous()) {
2741 2733 if (r->humongous_start_region() != _sh_region) {
2742 2734 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2743 2735 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2744 2736 HR_FORMAT_PARAMS(r),
2745 2737 r->humongous_start_region(),
2746 2738 _sh_region);
2747 2739 ++_failures;
2748 2740 }
2749 2741 }
2750 2742 return false;
2751 2743 }
2752 2744 uint failures() { return _failures; }
2753 2745 };
2754 2746
2755 2747 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2756 2748 CheckClaimValuesClosure cl(claim_value);
2757 2749 heap_region_iterate(&cl);
2758 2750 return cl.failures() == 0;
2759 2751 }
2760 2752
2761 2753 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2762 2754 private:
2763 2755 jint _claim_value;
2764 2756 uint _failures;
2765 2757
2766 2758 public:
2767 2759 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2768 2760 _claim_value(claim_value), _failures(0) { }
2769 2761
2770 2762 uint failures() { return _failures; }
2771 2763
2772 2764 bool doHeapRegion(HeapRegion* hr) {
2773 2765 assert(hr->in_collection_set(), "how?");
2774 2766 assert(!hr->isHumongous(), "H-region in CSet");
2775 2767 if (hr->claim_value() != _claim_value) {
2776 2768 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2777 2769 "claim value = %d, should be %d",
2778 2770 HR_FORMAT_PARAMS(hr),
2779 2771 hr->claim_value(), _claim_value);
2780 2772 _failures += 1;
2781 2773 }
2782 2774 return false;
2783 2775 }
2784 2776 };
2785 2777
2786 2778 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2787 2779 CheckClaimValuesInCSetHRClosure cl(claim_value);
2788 2780 collection_set_iterate(&cl);
2789 2781 return cl.failures() == 0;
2790 2782 }
2791 2783 #endif // ASSERT
2792 2784
2793 2785 // Clear the cached CSet starting regions and (more importantly)
2794 2786 // the time stamps. Called when we reset the GC time stamp.
2795 2787 void G1CollectedHeap::clear_cset_start_regions() {
2796 2788 assert(_worker_cset_start_region != NULL, "sanity");
2797 2789 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2798 2790
2799 2791 int n_queues = MAX2((int)ParallelGCThreads, 1);
2800 2792 for (int i = 0; i < n_queues; i++) {
2801 2793 _worker_cset_start_region[i] = NULL;
2802 2794 _worker_cset_start_region_time_stamp[i] = 0;
2803 2795 }
2804 2796 }
2805 2797
2806 2798 // Given the id of a worker, obtain or calculate a suitable
2807 2799 // starting region for iterating over the current collection set.
2808 2800 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2809 2801 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2810 2802
2811 2803 HeapRegion* result = NULL;
2812 2804 unsigned gc_time_stamp = get_gc_time_stamp();
2813 2805
2814 2806 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2815 2807 // Cached starting region for current worker was set
2816 2808 // during the current pause - so it's valid.
2817 2809 // Note: the cached starting heap region may be NULL
2818 2810 // (when the collection set is empty).
2819 2811 result = _worker_cset_start_region[worker_i];
2820 2812 assert(result == NULL || result->in_collection_set(), "sanity");
2821 2813 return result;
2822 2814 }
2823 2815
2824 2816 // The cached entry was not valid so let's calculate
2825 2817 // a suitable starting heap region for this worker.
2826 2818
2827 2819 // We want the parallel threads to start their collection
2828 2820 // set iteration at different collection set regions to
2829 2821 // avoid contention.
2830 2822 // If we have:
2831 2823 // n collection set regions
2832 2824 // p threads
2833 2825 // Then thread t will start at region floor ((t * n) / p)
2834 2826
2835 2827 result = g1_policy()->collection_set();
2836 2828 if (G1CollectedHeap::use_parallel_gc_threads()) {
2837 2829 uint cs_size = g1_policy()->cset_region_length();
2838 2830 uint active_workers = workers()->active_workers();
2839 2831 assert(UseDynamicNumberOfGCThreads ||
2840 2832 active_workers == workers()->total_workers(),
2841 2833 "Unless dynamic should use total workers");
2842 2834
2843 2835 uint end_ind = (cs_size * worker_i) / active_workers;
2844 2836 uint start_ind = 0;
2845 2837
2846 2838 if (worker_i > 0 &&
2847 2839 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2848 2840 // Previous workers starting region is valid
2849 2841 // so let's iterate from there
2850 2842 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2851 2843 result = _worker_cset_start_region[worker_i - 1];
2852 2844 }
2853 2845
2854 2846 for (uint i = start_ind; i < end_ind; i++) {
2855 2847 result = result->next_in_collection_set();
2856 2848 }
2857 2849 }
2858 2850
2859 2851 // Note: the calculated starting heap region may be NULL
2860 2852 // (when the collection set is empty).
2861 2853 assert(result == NULL || result->in_collection_set(), "sanity");
2862 2854 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2863 2855 "should be updated only once per pause");
2864 2856 _worker_cset_start_region[worker_i] = result;
2865 2857 OrderAccess::storestore();
2866 2858 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2867 2859 return result;
2868 2860 }
2869 2861
2870 2862 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2871 2863 uint no_of_par_workers) {
2872 2864 uint worker_num =
2873 2865 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2874 2866 assert(UseDynamicNumberOfGCThreads ||
2875 2867 no_of_par_workers == workers()->total_workers(),
2876 2868 "Non dynamic should use fixed number of workers");
2877 2869 const uint start_index = n_regions() * worker_i / worker_num;
2878 2870 return region_at(start_index);
2879 2871 }
2880 2872
2881 2873 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2882 2874 HeapRegion* r = g1_policy()->collection_set();
2883 2875 while (r != NULL) {
2884 2876 HeapRegion* next = r->next_in_collection_set();
2885 2877 if (cl->doHeapRegion(r)) {
2886 2878 cl->incomplete();
2887 2879 return;
2888 2880 }
2889 2881 r = next;
2890 2882 }
2891 2883 }
2892 2884
2893 2885 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2894 2886 HeapRegionClosure *cl) {
2895 2887 if (r == NULL) {
2896 2888 // The CSet is empty so there's nothing to do.
2897 2889 return;
2898 2890 }
2899 2891
2900 2892 assert(r->in_collection_set(),
2901 2893 "Start region must be a member of the collection set.");
2902 2894 HeapRegion* cur = r;
2903 2895 while (cur != NULL) {
2904 2896 HeapRegion* next = cur->next_in_collection_set();
2905 2897 if (cl->doHeapRegion(cur) && false) {
2906 2898 cl->incomplete();
2907 2899 return;
2908 2900 }
2909 2901 cur = next;
2910 2902 }
2911 2903 cur = g1_policy()->collection_set();
2912 2904 while (cur != r) {
2913 2905 HeapRegion* next = cur->next_in_collection_set();
2914 2906 if (cl->doHeapRegion(cur) && false) {
2915 2907 cl->incomplete();
2916 2908 return;
2917 2909 }
2918 2910 cur = next;
2919 2911 }
2920 2912 }
2921 2913
2922 2914 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2923 2915 return n_regions() > 0 ? region_at(0) : NULL;
2924 2916 }
2925 2917
2926 2918
2927 2919 Space* G1CollectedHeap::space_containing(const void* addr) const {
2928 2920 Space* res = heap_region_containing(addr);
2929 2921 return res;
2930 2922 }
2931 2923
2932 2924 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2933 2925 Space* sp = space_containing(addr);
2934 2926 if (sp != NULL) {
2935 2927 return sp->block_start(addr);
2936 2928 }
2937 2929 return NULL;
2938 2930 }
2939 2931
2940 2932 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2941 2933 Space* sp = space_containing(addr);
2942 2934 assert(sp != NULL, "block_size of address outside of heap");
2943 2935 return sp->block_size(addr);
2944 2936 }
2945 2937
2946 2938 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2947 2939 Space* sp = space_containing(addr);
2948 2940 return sp->block_is_obj(addr);
2949 2941 }
2950 2942
2951 2943 bool G1CollectedHeap::supports_tlab_allocation() const {
2952 2944 return true;
2953 2945 }
2954 2946
2955 2947 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2956 2948 return HeapRegion::GrainBytes;
2957 2949 }
2958 2950
2959 2951 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2960 2952 // Return the remaining space in the cur alloc region, but not less than
2961 2953 // the min TLAB size.
2962 2954
2963 2955 // Also, this value can be at most the humongous object threshold,
2964 2956 // since we can't allow tlabs to grow big enough to accomodate
2965 2957 // humongous objects.
2966 2958
2967 2959 HeapRegion* hr = _mutator_alloc_region.get();
2968 2960 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2969 2961 if (hr == NULL) {
2970 2962 return max_tlab_size;
2971 2963 } else {
2972 2964 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2973 2965 }
2974 2966 }
2975 2967
2976 2968 size_t G1CollectedHeap::max_capacity() const {
2977 2969 return _g1_reserved.byte_size();
2978 2970 }
2979 2971
2980 2972 jlong G1CollectedHeap::millis_since_last_gc() {
2981 2973 // assert(false, "NYI");
2982 2974 return 0;
2983 2975 }
2984 2976
2985 2977 void G1CollectedHeap::prepare_for_verify() {
2986 2978 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2987 2979 ensure_parsability(false);
2988 2980 }
2989 2981 g1_rem_set()->prepare_for_verify();
2990 2982 }
2991 2983
2992 2984 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
2993 2985 VerifyOption vo) {
2994 2986 switch (vo) {
2995 2987 case VerifyOption_G1UsePrevMarking:
2996 2988 return hr->obj_allocated_since_prev_marking(obj);
2997 2989 case VerifyOption_G1UseNextMarking:
2998 2990 return hr->obj_allocated_since_next_marking(obj);
2999 2991 case VerifyOption_G1UseMarkWord:
3000 2992 return false;
3001 2993 default:
3002 2994 ShouldNotReachHere();
3003 2995 }
3004 2996 return false; // keep some compilers happy
3005 2997 }
3006 2998
3007 2999 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3008 3000 switch (vo) {
3009 3001 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3010 3002 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3011 3003 case VerifyOption_G1UseMarkWord: return NULL;
3012 3004 default: ShouldNotReachHere();
3013 3005 }
3014 3006 return NULL; // keep some compilers happy
3015 3007 }
3016 3008
3017 3009 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3018 3010 switch (vo) {
3019 3011 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3020 3012 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3021 3013 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3022 3014 default: ShouldNotReachHere();
3023 3015 }
3024 3016 return false; // keep some compilers happy
3025 3017 }
3026 3018
3027 3019 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3028 3020 switch (vo) {
3029 3021 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3030 3022 case VerifyOption_G1UseNextMarking: return "NTAMS";
3031 3023 case VerifyOption_G1UseMarkWord: return "NONE";
3032 3024 default: ShouldNotReachHere();
3033 3025 }
3034 3026 return NULL; // keep some compilers happy
3035 3027 }
3036 3028
3037 3029 class VerifyLivenessOopClosure: public OopClosure {
3038 3030 G1CollectedHeap* _g1h;
3039 3031 VerifyOption _vo;
3040 3032 public:
3041 3033 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3042 3034 _g1h(g1h), _vo(vo)
3043 3035 { }
3044 3036 void do_oop(narrowOop *p) { do_oop_work(p); }
3045 3037 void do_oop( oop *p) { do_oop_work(p); }
3046 3038
3047 3039 template <class T> void do_oop_work(T *p) {
3048 3040 oop obj = oopDesc::load_decode_heap_oop(p);
3049 3041 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3050 3042 "Dead object referenced by a not dead object");
3051 3043 }
3052 3044 };
3053 3045
3054 3046 class VerifyObjsInRegionClosure: public ObjectClosure {
3055 3047 private:
3056 3048 G1CollectedHeap* _g1h;
3057 3049 size_t _live_bytes;
3058 3050 HeapRegion *_hr;
3059 3051 VerifyOption _vo;
3060 3052 public:
3061 3053 // _vo == UsePrevMarking -> use "prev" marking information,
3062 3054 // _vo == UseNextMarking -> use "next" marking information,
3063 3055 // _vo == UseMarkWord -> use mark word from object header.
3064 3056 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3065 3057 : _live_bytes(0), _hr(hr), _vo(vo) {
3066 3058 _g1h = G1CollectedHeap::heap();
3067 3059 }
3068 3060 void do_object(oop o) {
3069 3061 VerifyLivenessOopClosure isLive(_g1h, _vo);
3070 3062 assert(o != NULL, "Huh?");
3071 3063 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3072 3064 // If the object is alive according to the mark word,
3073 3065 // then verify that the marking information agrees.
3074 3066 // Note we can't verify the contra-positive of the
3075 3067 // above: if the object is dead (according to the mark
3076 3068 // word), it may not be marked, or may have been marked
3077 3069 // but has since became dead, or may have been allocated
3078 3070 // since the last marking.
3079 3071 if (_vo == VerifyOption_G1UseMarkWord) {
3080 3072 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3081 3073 }
3082 3074
3083 3075 o->oop_iterate_no_header(&isLive);
3084 3076 if (!_hr->obj_allocated_since_prev_marking(o)) {
3085 3077 size_t obj_size = o->size(); // Make sure we don't overflow
3086 3078 _live_bytes += (obj_size * HeapWordSize);
3087 3079 }
3088 3080 }
3089 3081 }
3090 3082 size_t live_bytes() { return _live_bytes; }
3091 3083 };
3092 3084
3093 3085 class PrintObjsInRegionClosure : public ObjectClosure {
3094 3086 HeapRegion *_hr;
3095 3087 G1CollectedHeap *_g1;
3096 3088 public:
3097 3089 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3098 3090 _g1 = G1CollectedHeap::heap();
3099 3091 };
3100 3092
3101 3093 void do_object(oop o) {
3102 3094 if (o != NULL) {
3103 3095 HeapWord *start = (HeapWord *) o;
3104 3096 size_t word_sz = o->size();
3105 3097 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3106 3098 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3107 3099 (void*) o, word_sz,
3108 3100 _g1->isMarkedPrev(o),
3109 3101 _g1->isMarkedNext(o),
3110 3102 _hr->obj_allocated_since_prev_marking(o));
3111 3103 HeapWord *end = start + word_sz;
3112 3104 HeapWord *cur;
3113 3105 int *val;
3114 3106 for (cur = start; cur < end; cur++) {
3115 3107 val = (int *) cur;
3116 3108 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3117 3109 }
3118 3110 }
3119 3111 }
3120 3112 };
3121 3113
3122 3114 class VerifyRegionClosure: public HeapRegionClosure {
3123 3115 private:
3124 3116 bool _par;
3125 3117 VerifyOption _vo;
3126 3118 bool _failures;
3127 3119 public:
3128 3120 // _vo == UsePrevMarking -> use "prev" marking information,
3129 3121 // _vo == UseNextMarking -> use "next" marking information,
3130 3122 // _vo == UseMarkWord -> use mark word from object header.
3131 3123 VerifyRegionClosure(bool par, VerifyOption vo)
3132 3124 : _par(par),
3133 3125 _vo(vo),
3134 3126 _failures(false) {}
3135 3127
3136 3128 bool failures() {
3137 3129 return _failures;
3138 3130 }
3139 3131
3140 3132 bool doHeapRegion(HeapRegion* r) {
3141 3133 if (!r->continuesHumongous()) {
3142 3134 bool failures = false;
3143 3135 r->verify(_vo, &failures);
3144 3136 if (failures) {
3145 3137 _failures = true;
3146 3138 } else {
3147 3139 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3148 3140 r->object_iterate(¬_dead_yet_cl);
3149 3141 if (_vo != VerifyOption_G1UseNextMarking) {
3150 3142 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3151 3143 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3152 3144 "max_live_bytes "SIZE_FORMAT" "
3153 3145 "< calculated "SIZE_FORMAT,
3154 3146 r->bottom(), r->end(),
3155 3147 r->max_live_bytes(),
3156 3148 not_dead_yet_cl.live_bytes());
3157 3149 _failures = true;
3158 3150 }
3159 3151 } else {
3160 3152 // When vo == UseNextMarking we cannot currently do a sanity
3161 3153 // check on the live bytes as the calculation has not been
3162 3154 // finalized yet.
3163 3155 }
3164 3156 }
3165 3157 }
3166 3158 return false; // stop the region iteration if we hit a failure
3167 3159 }
3168 3160 };
3169 3161
3170 3162 class YoungRefCounterClosure : public OopClosure {
3171 3163 G1CollectedHeap* _g1h;
3172 3164 int _count;
3173 3165 public:
3174 3166 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3175 3167 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3176 3168 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3177 3169
3178 3170 int count() { return _count; }
3179 3171 void reset_count() { _count = 0; };
3180 3172 };
3181 3173
3182 3174 class VerifyKlassClosure: public KlassClosure {
3183 3175 YoungRefCounterClosure _young_ref_counter_closure;
3184 3176 OopClosure *_oop_closure;
3185 3177 public:
3186 3178 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3187 3179 void do_klass(Klass* k) {
3188 3180 k->oops_do(_oop_closure);
3189 3181
3190 3182 _young_ref_counter_closure.reset_count();
3191 3183 k->oops_do(&_young_ref_counter_closure);
3192 3184 if (_young_ref_counter_closure.count() > 0) {
3193 3185 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3194 3186 }
3195 3187 }
3196 3188 };
3197 3189
3198 3190 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3199 3191 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3200 3192 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3201 3193 // we can change this closure to extend the simpler OopClosure.
3202 3194 class VerifyRootsClosure: public OopsInGenClosure {
3203 3195 private:
3204 3196 G1CollectedHeap* _g1h;
3205 3197 VerifyOption _vo;
3206 3198 bool _failures;
3207 3199 public:
3208 3200 // _vo == UsePrevMarking -> use "prev" marking information,
3209 3201 // _vo == UseNextMarking -> use "next" marking information,
3210 3202 // _vo == UseMarkWord -> use mark word from object header.
3211 3203 VerifyRootsClosure(VerifyOption vo) :
3212 3204 _g1h(G1CollectedHeap::heap()),
3213 3205 _vo(vo),
3214 3206 _failures(false) { }
3215 3207
3216 3208 bool failures() { return _failures; }
3217 3209
3218 3210 template <class T> void do_oop_nv(T* p) {
3219 3211 T heap_oop = oopDesc::load_heap_oop(p);
3220 3212 if (!oopDesc::is_null(heap_oop)) {
3221 3213 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3222 3214 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3223 3215 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3224 3216 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3225 3217 if (_vo == VerifyOption_G1UseMarkWord) {
3226 3218 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3227 3219 }
3228 3220 obj->print_on(gclog_or_tty);
3229 3221 _failures = true;
3230 3222 }
3231 3223 }
3232 3224 }
3233 3225
3234 3226 void do_oop(oop* p) { do_oop_nv(p); }
3235 3227 void do_oop(narrowOop* p) { do_oop_nv(p); }
3236 3228 };
3237 3229
3238 3230 // This is the task used for parallel heap verification.
3239 3231
3240 3232 class G1ParVerifyTask: public AbstractGangTask {
3241 3233 private:
3242 3234 G1CollectedHeap* _g1h;
3243 3235 VerifyOption _vo;
3244 3236 bool _failures;
3245 3237
3246 3238 public:
3247 3239 // _vo == UsePrevMarking -> use "prev" marking information,
3248 3240 // _vo == UseNextMarking -> use "next" marking information,
3249 3241 // _vo == UseMarkWord -> use mark word from object header.
3250 3242 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3251 3243 AbstractGangTask("Parallel verify task"),
3252 3244 _g1h(g1h),
3253 3245 _vo(vo),
3254 3246 _failures(false) { }
3255 3247
3256 3248 bool failures() {
3257 3249 return _failures;
3258 3250 }
3259 3251
3260 3252 void work(uint worker_id) {
3261 3253 HandleMark hm;
3262 3254 VerifyRegionClosure blk(true, _vo);
3263 3255 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3264 3256 _g1h->workers()->active_workers(),
3265 3257 HeapRegion::ParVerifyClaimValue);
3266 3258 if (blk.failures()) {
3267 3259 _failures = true;
3268 3260 }
3269 3261 }
3270 3262 };
3271 3263
3272 3264 void G1CollectedHeap::verify(bool silent) {
3273 3265 verify(silent, VerifyOption_G1UsePrevMarking);
3274 3266 }
3275 3267
3276 3268 void G1CollectedHeap::verify(bool silent,
3277 3269 VerifyOption vo) {
3278 3270 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3279 3271 if (!silent) { gclog_or_tty->print("Roots "); }
3280 3272 VerifyRootsClosure rootsCl(vo);
3281 3273
3282 3274 assert(Thread::current()->is_VM_thread(),
3283 3275 "Expected to be executed serially by the VM thread at this point");
3284 3276
3285 3277 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3286 3278 VerifyKlassClosure klassCl(this, &rootsCl);
3287 3279
3288 3280 // We apply the relevant closures to all the oops in the
3289 3281 // system dictionary, the string table and the code cache.
3290 3282 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3291 3283
3292 3284 // Need cleared claim bits for the strong roots processing
3293 3285 ClassLoaderDataGraph::clear_claimed_marks();
3294 3286
3295 3287 process_strong_roots(true, // activate StrongRootsScope
3296 3288 false, // we set "is scavenging" to false,
3297 3289 // so we don't reset the dirty cards.
3298 3290 ScanningOption(so), // roots scanning options
3299 3291 &rootsCl,
3300 3292 &blobsCl,
3301 3293 &klassCl
3302 3294 );
3303 3295
3304 3296 bool failures = rootsCl.failures();
3305 3297
3306 3298 if (vo != VerifyOption_G1UseMarkWord) {
3307 3299 // If we're verifying during a full GC then the region sets
3308 3300 // will have been torn down at the start of the GC. Therefore
3309 3301 // verifying the region sets will fail. So we only verify
3310 3302 // the region sets when not in a full GC.
3311 3303 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3312 3304 verify_region_sets();
3313 3305 }
3314 3306
3315 3307 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3316 3308 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3317 3309 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3318 3310 "sanity check");
3319 3311
3320 3312 G1ParVerifyTask task(this, vo);
3321 3313 assert(UseDynamicNumberOfGCThreads ||
3322 3314 workers()->active_workers() == workers()->total_workers(),
3323 3315 "If not dynamic should be using all the workers");
3324 3316 int n_workers = workers()->active_workers();
3325 3317 set_par_threads(n_workers);
3326 3318 workers()->run_task(&task);
3327 3319 set_par_threads(0);
3328 3320 if (task.failures()) {
3329 3321 failures = true;
3330 3322 }
3331 3323
3332 3324 // Checks that the expected amount of parallel work was done.
3333 3325 // The implication is that n_workers is > 0.
3334 3326 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3335 3327 "sanity check");
3336 3328
3337 3329 reset_heap_region_claim_values();
3338 3330
3339 3331 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3340 3332 "sanity check");
3341 3333 } else {
3342 3334 VerifyRegionClosure blk(false, vo);
3343 3335 heap_region_iterate(&blk);
3344 3336 if (blk.failures()) {
3345 3337 failures = true;
3346 3338 }
3347 3339 }
3348 3340 if (!silent) gclog_or_tty->print("RemSet ");
3349 3341 rem_set()->verify();
3350 3342
3351 3343 if (failures) {
3352 3344 gclog_or_tty->print_cr("Heap:");
3353 3345 // It helps to have the per-region information in the output to
3354 3346 // help us track down what went wrong. This is why we call
3355 3347 // print_extended_on() instead of print_on().
3356 3348 print_extended_on(gclog_or_tty);
3357 3349 gclog_or_tty->print_cr("");
3358 3350 #ifndef PRODUCT
3359 3351 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3360 3352 concurrent_mark()->print_reachable("at-verification-failure",
3361 3353 vo, false /* all */);
3362 3354 }
3363 3355 #endif
3364 3356 gclog_or_tty->flush();
3365 3357 }
3366 3358 guarantee(!failures, "there should not have been any failures");
3367 3359 } else {
3368 3360 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3369 3361 }
3370 3362 }
3371 3363
3372 3364 class PrintRegionClosure: public HeapRegionClosure {
3373 3365 outputStream* _st;
3374 3366 public:
3375 3367 PrintRegionClosure(outputStream* st) : _st(st) {}
3376 3368 bool doHeapRegion(HeapRegion* r) {
3377 3369 r->print_on(_st);
3378 3370 return false;
3379 3371 }
3380 3372 };
3381 3373
3382 3374 void G1CollectedHeap::print_on(outputStream* st) const {
3383 3375 st->print(" %-20s", "garbage-first heap");
3384 3376 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3385 3377 capacity()/K, used_unlocked()/K);
3386 3378 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3387 3379 _g1_storage.low_boundary(),
3388 3380 _g1_storage.high(),
3389 3381 _g1_storage.high_boundary());
3390 3382 st->cr();
3391 3383 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3392 3384 uint young_regions = _young_list->length();
3393 3385 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3394 3386 (size_t) young_regions * HeapRegion::GrainBytes / K);
3395 3387 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3396 3388 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3397 3389 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3398 3390 st->cr();
3399 3391 }
3400 3392
3401 3393 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3402 3394 print_on(st);
3403 3395
3404 3396 // Print the per-region information.
3405 3397 st->cr();
3406 3398 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3407 3399 "HS=humongous(starts), HC=humongous(continues), "
3408 3400 "CS=collection set, F=free, TS=gc time stamp, "
3409 3401 "PTAMS=previous top-at-mark-start, "
3410 3402 "NTAMS=next top-at-mark-start)");
3411 3403 PrintRegionClosure blk(st);
3412 3404 heap_region_iterate(&blk);
3413 3405 }
3414 3406
3415 3407 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3416 3408 if (G1CollectedHeap::use_parallel_gc_threads()) {
3417 3409 workers()->print_worker_threads_on(st);
3418 3410 }
3419 3411 _cmThread->print_on(st);
3420 3412 st->cr();
3421 3413 _cm->print_worker_threads_on(st);
3422 3414 _cg1r->print_worker_threads_on(st);
3423 3415 st->cr();
3424 3416 }
3425 3417
3426 3418 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3427 3419 if (G1CollectedHeap::use_parallel_gc_threads()) {
3428 3420 workers()->threads_do(tc);
3429 3421 }
3430 3422 tc->do_thread(_cmThread);
3431 3423 _cg1r->threads_do(tc);
3432 3424 }
3433 3425
3434 3426 void G1CollectedHeap::print_tracing_info() const {
3435 3427 // We'll overload this to mean "trace GC pause statistics."
3436 3428 if (TraceGen0Time || TraceGen1Time) {
3437 3429 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3438 3430 // to that.
3439 3431 g1_policy()->print_tracing_info();
3440 3432 }
3441 3433 if (G1SummarizeRSetStats) {
3442 3434 g1_rem_set()->print_summary_info();
3443 3435 }
3444 3436 if (G1SummarizeConcMark) {
3445 3437 concurrent_mark()->print_summary_info();
3446 3438 }
3447 3439 g1_policy()->print_yg_surv_rate_info();
3448 3440 SpecializationStats::print();
3449 3441 }
3450 3442
3451 3443 #ifndef PRODUCT
3452 3444 // Helpful for debugging RSet issues.
3453 3445
3454 3446 class PrintRSetsClosure : public HeapRegionClosure {
3455 3447 private:
3456 3448 const char* _msg;
3457 3449 size_t _occupied_sum;
3458 3450
3459 3451 public:
3460 3452 bool doHeapRegion(HeapRegion* r) {
3461 3453 HeapRegionRemSet* hrrs = r->rem_set();
3462 3454 size_t occupied = hrrs->occupied();
3463 3455 _occupied_sum += occupied;
3464 3456
3465 3457 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3466 3458 HR_FORMAT_PARAMS(r));
3467 3459 if (occupied == 0) {
3468 3460 gclog_or_tty->print_cr(" RSet is empty");
3469 3461 } else {
3470 3462 hrrs->print();
3471 3463 }
3472 3464 gclog_or_tty->print_cr("----------");
3473 3465 return false;
3474 3466 }
3475 3467
3476 3468 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3477 3469 gclog_or_tty->cr();
3478 3470 gclog_or_tty->print_cr("========================================");
3479 3471 gclog_or_tty->print_cr(msg);
3480 3472 gclog_or_tty->cr();
3481 3473 }
3482 3474
3483 3475 ~PrintRSetsClosure() {
3484 3476 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3485 3477 gclog_or_tty->print_cr("========================================");
3486 3478 gclog_or_tty->cr();
3487 3479 }
3488 3480 };
3489 3481
3490 3482 void G1CollectedHeap::print_cset_rsets() {
3491 3483 PrintRSetsClosure cl("Printing CSet RSets");
3492 3484 collection_set_iterate(&cl);
3493 3485 }
3494 3486
3495 3487 void G1CollectedHeap::print_all_rsets() {
3496 3488 PrintRSetsClosure cl("Printing All RSets");;
3497 3489 heap_region_iterate(&cl);
3498 3490 }
3499 3491 #endif // PRODUCT
3500 3492
3501 3493 G1CollectedHeap* G1CollectedHeap::heap() {
3502 3494 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3503 3495 "not a garbage-first heap");
3504 3496 return _g1h;
3505 3497 }
3506 3498
3507 3499 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3508 3500 // always_do_update_barrier = false;
3509 3501 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3510 3502 // Call allocation profiler
3511 3503 AllocationProfiler::iterate_since_last_gc();
3512 3504 // Fill TLAB's and such
3513 3505 ensure_parsability(true);
3514 3506 }
3515 3507
3516 3508 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3517 3509 // FIXME: what is this about?
3518 3510 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3519 3511 // is set.
3520 3512 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3521 3513 "derived pointer present"));
3522 3514 // always_do_update_barrier = true;
3523 3515
3524 3516 // We have just completed a GC. Update the soft reference
3525 3517 // policy with the new heap occupancy
3526 3518 Universe::update_heap_info_at_gc();
3527 3519 }
3528 3520
3529 3521 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3530 3522 unsigned int gc_count_before,
3531 3523 bool* succeeded) {
3532 3524 assert_heap_not_locked_and_not_at_safepoint();
3533 3525 g1_policy()->record_stop_world_start();
3534 3526 VM_G1IncCollectionPause op(gc_count_before,
3535 3527 word_size,
3536 3528 false, /* should_initiate_conc_mark */
3537 3529 g1_policy()->max_pause_time_ms(),
3538 3530 GCCause::_g1_inc_collection_pause);
3539 3531 VMThread::execute(&op);
3540 3532
3541 3533 HeapWord* result = op.result();
3542 3534 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3543 3535 assert(result == NULL || ret_succeeded,
3544 3536 "the result should be NULL if the VM did not succeed");
3545 3537 *succeeded = ret_succeeded;
3546 3538
3547 3539 assert_heap_not_locked();
3548 3540 return result;
3549 3541 }
3550 3542
3551 3543 void
3552 3544 G1CollectedHeap::doConcurrentMark() {
3553 3545 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3554 3546 if (!_cmThread->in_progress()) {
3555 3547 _cmThread->set_started();
3556 3548 CGC_lock->notify();
3557 3549 }
3558 3550 }
3559 3551
3560 3552 size_t G1CollectedHeap::pending_card_num() {
3561 3553 size_t extra_cards = 0;
3562 3554 JavaThread *curr = Threads::first();
3563 3555 while (curr != NULL) {
3564 3556 DirtyCardQueue& dcq = curr->dirty_card_queue();
3565 3557 extra_cards += dcq.size();
3566 3558 curr = curr->next();
3567 3559 }
3568 3560 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3569 3561 size_t buffer_size = dcqs.buffer_size();
3570 3562 size_t buffer_num = dcqs.completed_buffers_num();
3571 3563
3572 3564 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3573 3565 // in bytes - not the number of 'entries'. We need to convert
3574 3566 // into a number of cards.
3575 3567 return (buffer_size * buffer_num + extra_cards) / oopSize;
3576 3568 }
3577 3569
3578 3570 size_t G1CollectedHeap::cards_scanned() {
3579 3571 return g1_rem_set()->cardsScanned();
3580 3572 }
3581 3573
3582 3574 void
3583 3575 G1CollectedHeap::setup_surviving_young_words() {
3584 3576 assert(_surviving_young_words == NULL, "pre-condition");
3585 3577 uint array_length = g1_policy()->young_cset_region_length();
3586 3578 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3587 3579 if (_surviving_young_words == NULL) {
3588 3580 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3589 3581 "Not enough space for young surv words summary.");
3590 3582 }
3591 3583 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3592 3584 #ifdef ASSERT
3593 3585 for (uint i = 0; i < array_length; ++i) {
3594 3586 assert( _surviving_young_words[i] == 0, "memset above" );
3595 3587 }
3596 3588 #endif // !ASSERT
3597 3589 }
3598 3590
3599 3591 void
3600 3592 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3601 3593 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3602 3594 uint array_length = g1_policy()->young_cset_region_length();
3603 3595 for (uint i = 0; i < array_length; ++i) {
3604 3596 _surviving_young_words[i] += surv_young_words[i];
3605 3597 }
3606 3598 }
3607 3599
3608 3600 void
3609 3601 G1CollectedHeap::cleanup_surviving_young_words() {
3610 3602 guarantee( _surviving_young_words != NULL, "pre-condition" );
3611 3603 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3612 3604 _surviving_young_words = NULL;
3613 3605 }
3614 3606
3615 3607 #ifdef ASSERT
3616 3608 class VerifyCSetClosure: public HeapRegionClosure {
3617 3609 public:
3618 3610 bool doHeapRegion(HeapRegion* hr) {
3619 3611 // Here we check that the CSet region's RSet is ready for parallel
3620 3612 // iteration. The fields that we'll verify are only manipulated
3621 3613 // when the region is part of a CSet and is collected. Afterwards,
3622 3614 // we reset these fields when we clear the region's RSet (when the
3623 3615 // region is freed) so they are ready when the region is
3624 3616 // re-allocated. The only exception to this is if there's an
3625 3617 // evacuation failure and instead of freeing the region we leave
3626 3618 // it in the heap. In that case, we reset these fields during
3627 3619 // evacuation failure handling.
3628 3620 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3629 3621
3630 3622 // Here's a good place to add any other checks we'd like to
3631 3623 // perform on CSet regions.
3632 3624 return false;
3633 3625 }
3634 3626 };
3635 3627 #endif // ASSERT
3636 3628
3637 3629 #if TASKQUEUE_STATS
3638 3630 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3639 3631 st->print_raw_cr("GC Task Stats");
3640 3632 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3641 3633 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3642 3634 }
3643 3635
3644 3636 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3645 3637 print_taskqueue_stats_hdr(st);
3646 3638
3647 3639 TaskQueueStats totals;
3648 3640 const int n = workers() != NULL ? workers()->total_workers() : 1;
3649 3641 for (int i = 0; i < n; ++i) {
3650 3642 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3651 3643 totals += task_queue(i)->stats;
3652 3644 }
3653 3645 st->print_raw("tot "); totals.print(st); st->cr();
3654 3646
3655 3647 DEBUG_ONLY(totals.verify());
3656 3648 }
3657 3649
3658 3650 void G1CollectedHeap::reset_taskqueue_stats() {
3659 3651 const int n = workers() != NULL ? workers()->total_workers() : 1;
3660 3652 for (int i = 0; i < n; ++i) {
3661 3653 task_queue(i)->stats.reset();
3662 3654 }
3663 3655 }
3664 3656 #endif // TASKQUEUE_STATS
3665 3657
3666 3658 bool
3667 3659 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3668 3660 assert_at_safepoint(true /* should_be_vm_thread */);
3669 3661 guarantee(!is_gc_active(), "collection is not reentrant");
3670 3662
3671 3663 if (GC_locker::check_active_before_gc()) {
3672 3664 return false;
3673 3665 }
3674 3666
3675 3667 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3676 3668 ResourceMark rm;
3677 3669
3678 3670 print_heap_before_gc();
3679 3671
3680 3672 HRSPhaseSetter x(HRSPhaseEvacuation);
3681 3673 verify_region_sets_optional();
3682 3674 verify_dirty_young_regions();
3683 3675
3684 3676 // This call will decide whether this pause is an initial-mark
3685 3677 // pause. If it is, during_initial_mark_pause() will return true
3686 3678 // for the duration of this pause.
3687 3679 g1_policy()->decide_on_conc_mark_initiation();
3688 3680
3689 3681 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3690 3682 assert(!g1_policy()->during_initial_mark_pause() ||
3691 3683 g1_policy()->gcs_are_young(), "sanity");
3692 3684
3693 3685 // We also do not allow mixed GCs during marking.
3694 3686 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3695 3687
3696 3688 // Record whether this pause is an initial mark. When the current
3697 3689 // thread has completed its logging output and it's safe to signal
3698 3690 // the CM thread, the flag's value in the policy has been reset.
3699 3691 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3700 3692
3701 3693 // Inner scope for scope based logging, timers, and stats collection
3702 3694 {
3703 3695 if (g1_policy()->during_initial_mark_pause()) {
3704 3696 // We are about to start a marking cycle, so we increment the
3705 3697 // full collection counter.
3706 3698 increment_old_marking_cycles_started();
3707 3699 }
3708 3700 // if the log level is "finer" is on, we'll print long statistics information
3709 3701 // in the collector policy code, so let's not print this as the output
3710 3702 // is messy if we do.
3711 3703 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
3712 3704 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3713 3705
3714 3706 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3715 3707 workers()->active_workers() : 1);
3716 3708 double pause_start_sec = os::elapsedTime();
3717 3709 g1_policy()->phase_times()->note_gc_start(active_workers);
3718 3710 bool initial_mark_gc = g1_policy()->during_initial_mark_pause();
3719 3711
3720 3712 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3721 3713 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3722 3714
3723 3715 // If the secondary_free_list is not empty, append it to the
3724 3716 // free_list. No need to wait for the cleanup operation to finish;
3725 3717 // the region allocation code will check the secondary_free_list
3726 3718 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3727 3719 // set, skip this step so that the region allocation code has to
3728 3720 // get entries from the secondary_free_list.
3729 3721 if (!G1StressConcRegionFreeing) {
3730 3722 append_secondary_free_list_if_not_empty_with_lock();
3731 3723 }
3732 3724
3733 3725 assert(check_young_list_well_formed(),
3734 3726 "young list should be well formed");
3735 3727
3736 3728 // Don't dynamically change the number of GC threads this early. A value of
3737 3729 // 0 is used to indicate serial work. When parallel work is done,
3738 3730 // it will be set.
3739 3731
3740 3732 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3741 3733 IsGCActiveMark x;
3742 3734
3743 3735 gc_prologue(false);
3744 3736 increment_total_collections(false /* full gc */);
3745 3737 increment_gc_time_stamp();
3746 3738
3747 3739 verify_before_gc();
3748 3740
3749 3741 COMPILER2_PRESENT(DerivedPointerTable::clear());
3750 3742
3751 3743 // Please see comment in g1CollectedHeap.hpp and
3752 3744 // G1CollectedHeap::ref_processing_init() to see how
3753 3745 // reference processing currently works in G1.
3754 3746
3755 3747 // Enable discovery in the STW reference processor
3756 3748 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3757 3749 true /*verify_no_refs*/);
3758 3750
3759 3751 {
3760 3752 // We want to temporarily turn off discovery by the
3761 3753 // CM ref processor, if necessary, and turn it back on
3762 3754 // on again later if we do. Using a scoped
3763 3755 // NoRefDiscovery object will do this.
3764 3756 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3765 3757
3766 3758 // Forget the current alloc region (we might even choose it to be part
3767 3759 // of the collection set!).
3768 3760 release_mutator_alloc_region();
3769 3761
3770 3762 // We should call this after we retire the mutator alloc
3771 3763 // region(s) so that all the ALLOC / RETIRE events are generated
3772 3764 // before the start GC event.
3773 3765 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3774 3766
3775 3767 // This timing is only used by the ergonomics to handle our pause target.
3776 3768 // It is unclear why this should not include the full pause. We will
3777 3769 // investigate this in CR 7178365.
3778 3770 //
3779 3771 // Preserving the old comment here if that helps the investigation:
3780 3772 //
3781 3773 // The elapsed time induced by the start time below deliberately elides
3782 3774 // the possible verification above.
3783 3775 double sample_start_time_sec = os::elapsedTime();
3784 3776 size_t start_used_bytes = used();
3785 3777
3786 3778 #if YOUNG_LIST_VERBOSE
3787 3779 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3788 3780 _young_list->print();
3789 3781 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3790 3782 #endif // YOUNG_LIST_VERBOSE
3791 3783
3792 3784 g1_policy()->record_collection_pause_start(sample_start_time_sec,
3793 3785 start_used_bytes);
3794 3786
3795 3787 double scan_wait_start = os::elapsedTime();
3796 3788 // We have to wait until the CM threads finish scanning the
3797 3789 // root regions as it's the only way to ensure that all the
3798 3790 // objects on them have been correctly scanned before we start
3799 3791 // moving them during the GC.
3800 3792 bool waited = _cm->root_regions()->wait_until_scan_finished();
3801 3793 double wait_time_ms = 0.0;
3802 3794 if (waited) {
3803 3795 double scan_wait_end = os::elapsedTime();
3804 3796 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3805 3797 }
3806 3798 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3807 3799
3808 3800 #if YOUNG_LIST_VERBOSE
3809 3801 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3810 3802 _young_list->print();
3811 3803 #endif // YOUNG_LIST_VERBOSE
3812 3804
3813 3805 if (g1_policy()->during_initial_mark_pause()) {
3814 3806 concurrent_mark()->checkpointRootsInitialPre();
3815 3807 }
3816 3808
3817 3809 #if YOUNG_LIST_VERBOSE
3818 3810 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3819 3811 _young_list->print();
3820 3812 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3821 3813 #endif // YOUNG_LIST_VERBOSE
3822 3814
3823 3815 g1_policy()->finalize_cset(target_pause_time_ms);
3824 3816
3825 3817 _cm->note_start_of_gc();
3826 3818 // We should not verify the per-thread SATB buffers given that
3827 3819 // we have not filtered them yet (we'll do so during the
3828 3820 // GC). We also call this after finalize_cset() to
3829 3821 // ensure that the CSet has been finalized.
3830 3822 _cm->verify_no_cset_oops(true /* verify_stacks */,
3831 3823 true /* verify_enqueued_buffers */,
3832 3824 false /* verify_thread_buffers */,
3833 3825 true /* verify_fingers */);
3834 3826
3835 3827 if (_hr_printer.is_active()) {
3836 3828 HeapRegion* hr = g1_policy()->collection_set();
3837 3829 while (hr != NULL) {
3838 3830 G1HRPrinter::RegionType type;
3839 3831 if (!hr->is_young()) {
3840 3832 type = G1HRPrinter::Old;
3841 3833 } else if (hr->is_survivor()) {
3842 3834 type = G1HRPrinter::Survivor;
3843 3835 } else {
3844 3836 type = G1HRPrinter::Eden;
3845 3837 }
3846 3838 _hr_printer.cset(hr);
3847 3839 hr = hr->next_in_collection_set();
3848 3840 }
3849 3841 }
3850 3842
3851 3843 #ifdef ASSERT
3852 3844 VerifyCSetClosure cl;
3853 3845 collection_set_iterate(&cl);
3854 3846 #endif // ASSERT
3855 3847
3856 3848 setup_surviving_young_words();
3857 3849
3858 3850 // Initialize the GC alloc regions.
3859 3851 init_gc_alloc_regions();
3860 3852
3861 3853 // Actually do the work...
3862 3854 evacuate_collection_set();
3863 3855
3864 3856 // We do this to mainly verify the per-thread SATB buffers
3865 3857 // (which have been filtered by now) since we didn't verify
3866 3858 // them earlier. No point in re-checking the stacks / enqueued
3867 3859 // buffers given that the CSet has not changed since last time
3868 3860 // we checked.
3869 3861 _cm->verify_no_cset_oops(false /* verify_stacks */,
3870 3862 false /* verify_enqueued_buffers */,
3871 3863 true /* verify_thread_buffers */,
3872 3864 true /* verify_fingers */);
3873 3865
3874 3866 free_collection_set(g1_policy()->collection_set());
3875 3867 g1_policy()->clear_collection_set();
3876 3868
3877 3869 cleanup_surviving_young_words();
3878 3870
3879 3871 // Start a new incremental collection set for the next pause.
3880 3872 g1_policy()->start_incremental_cset_building();
3881 3873
3882 3874 // Clear the _cset_fast_test bitmap in anticipation of adding
3883 3875 // regions to the incremental collection set for the next
3884 3876 // evacuation pause.
3885 3877 clear_cset_fast_test();
3886 3878
3887 3879 _young_list->reset_sampled_info();
3888 3880
3889 3881 // Don't check the whole heap at this point as the
3890 3882 // GC alloc regions from this pause have been tagged
3891 3883 // as survivors and moved on to the survivor list.
3892 3884 // Survivor regions will fail the !is_young() check.
3893 3885 assert(check_young_list_empty(false /* check_heap */),
3894 3886 "young list should be empty");
3895 3887
3896 3888 #if YOUNG_LIST_VERBOSE
3897 3889 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3898 3890 _young_list->print();
3899 3891 #endif // YOUNG_LIST_VERBOSE
3900 3892
3901 3893 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3902 3894 _young_list->first_survivor_region(),
3903 3895 _young_list->last_survivor_region());
3904 3896
3905 3897 _young_list->reset_auxilary_lists();
3906 3898
3907 3899 if (evacuation_failed()) {
3908 3900 _summary_bytes_used = recalculate_used();
3909 3901 } else {
3910 3902 // The "used" of the the collection set have already been subtracted
3911 3903 // when they were freed. Add in the bytes evacuated.
3912 3904 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3913 3905 }
3914 3906
3915 3907 if (g1_policy()->during_initial_mark_pause()) {
3916 3908 // We have to do this before we notify the CM threads that
3917 3909 // they can start working to make sure that all the
3918 3910 // appropriate initialization is done on the CM object.
3919 3911 concurrent_mark()->checkpointRootsInitialPost();
3920 3912 set_marking_started();
3921 3913 // Note that we don't actually trigger the CM thread at
3922 3914 // this point. We do that later when we're sure that
3923 3915 // the current thread has completed its logging output.
3924 3916 }
3925 3917
3926 3918 allocate_dummy_regions();
3927 3919
3928 3920 #if YOUNG_LIST_VERBOSE
3929 3921 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3930 3922 _young_list->print();
3931 3923 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3932 3924 #endif // YOUNG_LIST_VERBOSE
3933 3925
3934 3926 init_mutator_alloc_region();
3935 3927
3936 3928 {
3937 3929 size_t expand_bytes = g1_policy()->expansion_amount();
3938 3930 if (expand_bytes > 0) {
3939 3931 size_t bytes_before = capacity();
3940 3932 // No need for an ergo verbose message here,
3941 3933 // expansion_amount() does this when it returns a value > 0.
3942 3934 if (!expand(expand_bytes)) {
3943 3935 // We failed to expand the heap so let's verify that
3944 3936 // committed/uncommitted amount match the backing store
3945 3937 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3946 3938 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3947 3939 }
3948 3940 }
3949 3941 }
3950 3942
3951 3943 // We redo the verificaiton but now wrt to the new CSet which
3952 3944 // has just got initialized after the previous CSet was freed.
3953 3945 _cm->verify_no_cset_oops(true /* verify_stacks */,
3954 3946 true /* verify_enqueued_buffers */,
3955 3947 true /* verify_thread_buffers */,
3956 3948 true /* verify_fingers */);
3957 3949 _cm->note_end_of_gc();
3958 3950
3959 3951 // This timing is only used by the ergonomics to handle our pause target.
3960 3952 // It is unclear why this should not include the full pause. We will
3961 3953 // investigate this in CR 7178365.
3962 3954 double sample_end_time_sec = os::elapsedTime();
3963 3955 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
3964 3956 g1_policy()->record_collection_pause_end(pause_time_ms);
3965 3957
3966 3958 MemoryService::track_memory_usage();
3967 3959
3968 3960 // In prepare_for_verify() below we'll need to scan the deferred
3969 3961 // update buffers to bring the RSets up-to-date if
3970 3962 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3971 3963 // the update buffers we'll probably need to scan cards on the
3972 3964 // regions we just allocated to (i.e., the GC alloc
3973 3965 // regions). However, during the last GC we called
3974 3966 // set_saved_mark() on all the GC alloc regions, so card
3975 3967 // scanning might skip the [saved_mark_word()...top()] area of
3976 3968 // those regions (i.e., the area we allocated objects into
3977 3969 // during the last GC). But it shouldn't. Given that
3978 3970 // saved_mark_word() is conditional on whether the GC time stamp
3979 3971 // on the region is current or not, by incrementing the GC time
3980 3972 // stamp here we invalidate all the GC time stamps on all the
3981 3973 // regions and saved_mark_word() will simply return top() for
3982 3974 // all the regions. This is a nicer way of ensuring this rather
3983 3975 // than iterating over the regions and fixing them. In fact, the
3984 3976 // GC time stamp increment here also ensures that
3985 3977 // saved_mark_word() will return top() between pauses, i.e.,
3986 3978 // during concurrent refinement. So we don't need the
3987 3979 // is_gc_active() check to decided which top to use when
3988 3980 // scanning cards (see CR 7039627).
3989 3981 increment_gc_time_stamp();
3990 3982
3991 3983 verify_after_gc();
3992 3984
3993 3985 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
3994 3986 ref_processor_stw()->verify_no_references_recorded();
3995 3987
3996 3988 // CM reference discovery will be re-enabled if necessary.
3997 3989 }
3998 3990
3999 3991 // We should do this after we potentially expand the heap so
4000 3992 // that all the COMMIT events are generated before the end GC
4001 3993 // event, and after we retire the GC alloc regions so that all
4002 3994 // RETIRE events are generated before the end GC event.
4003 3995 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4004 3996
4005 3997 if (mark_in_progress()) {
4006 3998 concurrent_mark()->update_g1_committed();
4007 3999 }
4008 4000
4009 4001 #ifdef TRACESPINNING
4010 4002 ParallelTaskTerminator::print_termination_counts();
4011 4003 #endif
4012 4004
4013 4005 gc_epilogue(false);
4014 4006
4015 4007 if (G1Log::fine()) {
4016 4008 if (PrintGCTimeStamps) {
4017 4009 gclog_or_tty->stamp();
4018 4010 gclog_or_tty->print(": ");
4019 4011 }
4020 4012
4021 4013 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
4022 4014 .append(g1_policy()->gcs_are_young() ? " (young)" : " (mixed)")
4023 4015 .append(initial_mark_gc ? " (initial-mark)" : "");
4024 4016
4025 4017 double pause_time_sec = os::elapsedTime() - pause_start_sec;
4026 4018
4027 4019 if (G1Log::finer()) {
4028 4020 if (evacuation_failed()) {
4029 4021 gc_cause_str.append(" (to-space exhausted)");
4030 4022 }
4031 4023 gclog_or_tty->print_cr("[%s, %3.7f secs]", (const char*)gc_cause_str, pause_time_sec);
4032 4024 g1_policy()->phase_times()->note_gc_end();
4033 4025 g1_policy()->phase_times()->print(pause_time_sec);
4034 4026 g1_policy()->print_detailed_heap_transition();
4035 4027 } else {
4036 4028 if (evacuation_failed()) {
4037 4029 gc_cause_str.append("--");
4038 4030 }
4039 4031 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
4040 4032 g1_policy()->print_heap_transition();
4041 4033 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
4042 4034 }
4043 4035 }
4044 4036 }
4045 4037
4046 4038 // It is not yet to safe to tell the concurrent mark to
4047 4039 // start as we have some optional output below. We don't want the
4048 4040 // output from the concurrent mark thread interfering with this
4049 4041 // logging output either.
4050 4042
4051 4043 _hrs.verify_optional();
4052 4044 verify_region_sets_optional();
4053 4045
4054 4046 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4055 4047 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4056 4048
4057 4049 print_heap_after_gc();
4058 4050
4059 4051 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4060 4052 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4061 4053 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4062 4054 // before any GC notifications are raised.
4063 4055 g1mm()->update_sizes();
4064 4056 }
4065 4057
4066 4058 if (G1SummarizeRSetStats &&
4067 4059 (G1SummarizeRSetStatsPeriod > 0) &&
4068 4060 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4069 4061 g1_rem_set()->print_summary_info();
4070 4062 }
4071 4063
4072 4064 // It should now be safe to tell the concurrent mark thread to start
4073 4065 // without its logging output interfering with the logging output
4074 4066 // that came from the pause.
4075 4067
4076 4068 if (should_start_conc_mark) {
4077 4069 // CAUTION: after the doConcurrentMark() call below,
4078 4070 // the concurrent marking thread(s) could be running
4079 4071 // concurrently with us. Make sure that anything after
4080 4072 // this point does not assume that we are the only GC thread
4081 4073 // running. Note: of course, the actual marking work will
4082 4074 // not start until the safepoint itself is released in
4083 4075 // ConcurrentGCThread::safepoint_desynchronize().
4084 4076 doConcurrentMark();
4085 4077 }
4086 4078
4087 4079 return true;
4088 4080 }
4089 4081
4090 4082 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4091 4083 {
4092 4084 size_t gclab_word_size;
4093 4085 switch (purpose) {
4094 4086 case GCAllocForSurvived:
4095 4087 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4096 4088 break;
4097 4089 case GCAllocForTenured:
4098 4090 gclab_word_size = _old_plab_stats.desired_plab_sz();
4099 4091 break;
4100 4092 default:
4101 4093 assert(false, "unknown GCAllocPurpose");
4102 4094 gclab_word_size = _old_plab_stats.desired_plab_sz();
4103 4095 break;
4104 4096 }
4105 4097
4106 4098 // Prevent humongous PLAB sizes for two reasons:
4107 4099 // * PLABs are allocated using a similar paths as oops, but should
4108 4100 // never be in a humongous region
4109 4101 // * Allowing humongous PLABs needlessly churns the region free lists
4110 4102 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4111 4103 }
4112 4104
4113 4105 void G1CollectedHeap::init_mutator_alloc_region() {
4114 4106 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4115 4107 _mutator_alloc_region.init();
4116 4108 }
4117 4109
4118 4110 void G1CollectedHeap::release_mutator_alloc_region() {
4119 4111 _mutator_alloc_region.release();
4120 4112 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4121 4113 }
4122 4114
4123 4115 void G1CollectedHeap::init_gc_alloc_regions() {
4124 4116 assert_at_safepoint(true /* should_be_vm_thread */);
4125 4117
4126 4118 _survivor_gc_alloc_region.init();
4127 4119 _old_gc_alloc_region.init();
4128 4120 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4129 4121 _retained_old_gc_alloc_region = NULL;
4130 4122
4131 4123 // We will discard the current GC alloc region if:
4132 4124 // a) it's in the collection set (it can happen!),
4133 4125 // b) it's already full (no point in using it),
4134 4126 // c) it's empty (this means that it was emptied during
4135 4127 // a cleanup and it should be on the free list now), or
4136 4128 // d) it's humongous (this means that it was emptied
4137 4129 // during a cleanup and was added to the free list, but
4138 4130 // has been subseqently used to allocate a humongous
4139 4131 // object that may be less than the region size).
4140 4132 if (retained_region != NULL &&
4141 4133 !retained_region->in_collection_set() &&
4142 4134 !(retained_region->top() == retained_region->end()) &&
4143 4135 !retained_region->is_empty() &&
4144 4136 !retained_region->isHumongous()) {
4145 4137 retained_region->set_saved_mark();
4146 4138 // The retained region was added to the old region set when it was
4147 4139 // retired. We have to remove it now, since we don't allow regions
4148 4140 // we allocate to in the region sets. We'll re-add it later, when
4149 4141 // it's retired again.
4150 4142 _old_set.remove(retained_region);
4151 4143 bool during_im = g1_policy()->during_initial_mark_pause();
4152 4144 retained_region->note_start_of_copying(during_im);
4153 4145 _old_gc_alloc_region.set(retained_region);
4154 4146 _hr_printer.reuse(retained_region);
4155 4147 }
4156 4148 }
4157 4149
4158 4150 void G1CollectedHeap::release_gc_alloc_regions() {
4159 4151 _survivor_gc_alloc_region.release();
4160 4152 // If we have an old GC alloc region to release, we'll save it in
4161 4153 // _retained_old_gc_alloc_region. If we don't
4162 4154 // _retained_old_gc_alloc_region will become NULL. This is what we
4163 4155 // want either way so no reason to check explicitly for either
4164 4156 // condition.
4165 4157 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4166 4158
4167 4159 if (ResizePLAB) {
4168 4160 _survivor_plab_stats.adjust_desired_plab_sz();
4169 4161 _old_plab_stats.adjust_desired_plab_sz();
4170 4162 }
4171 4163 }
4172 4164
4173 4165 void G1CollectedHeap::abandon_gc_alloc_regions() {
4174 4166 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4175 4167 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4176 4168 _retained_old_gc_alloc_region = NULL;
4177 4169 }
4178 4170
4179 4171 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4180 4172 _drain_in_progress = false;
4181 4173 set_evac_failure_closure(cl);
4182 4174 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4183 4175 }
4184 4176
4185 4177 void G1CollectedHeap::finalize_for_evac_failure() {
4186 4178 assert(_evac_failure_scan_stack != NULL &&
4187 4179 _evac_failure_scan_stack->length() == 0,
4188 4180 "Postcondition");
4189 4181 assert(!_drain_in_progress, "Postcondition");
4190 4182 delete _evac_failure_scan_stack;
4191 4183 _evac_failure_scan_stack = NULL;
4192 4184 }
4193 4185
4194 4186 void G1CollectedHeap::remove_self_forwarding_pointers() {
4195 4187 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4196 4188
4197 4189 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4198 4190
4199 4191 if (G1CollectedHeap::use_parallel_gc_threads()) {
4200 4192 set_par_threads();
4201 4193 workers()->run_task(&rsfp_task);
4202 4194 set_par_threads(0);
4203 4195 } else {
4204 4196 rsfp_task.work(0);
4205 4197 }
4206 4198
4207 4199 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4208 4200
4209 4201 // Reset the claim values in the regions in the collection set.
4210 4202 reset_cset_heap_region_claim_values();
4211 4203
4212 4204 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4213 4205
4214 4206 // Now restore saved marks, if any.
4215 4207 if (_objs_with_preserved_marks != NULL) {
4216 4208 assert(_preserved_marks_of_objs != NULL, "Both or none.");
4217 4209 guarantee(_objs_with_preserved_marks->length() ==
4218 4210 _preserved_marks_of_objs->length(), "Both or none.");
4219 4211 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
4220 4212 oop obj = _objs_with_preserved_marks->at(i);
4221 4213 markOop m = _preserved_marks_of_objs->at(i);
4222 4214 obj->set_mark(m);
4223 4215 }
4224 4216
4225 4217 // Delete the preserved marks growable arrays (allocated on the C heap).
4226 4218 delete _objs_with_preserved_marks;
4227 4219 delete _preserved_marks_of_objs;
4228 4220 _objs_with_preserved_marks = NULL;
4229 4221 _preserved_marks_of_objs = NULL;
4230 4222 }
4231 4223 }
4232 4224
4233 4225 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4234 4226 _evac_failure_scan_stack->push(obj);
4235 4227 }
4236 4228
4237 4229 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4238 4230 assert(_evac_failure_scan_stack != NULL, "precondition");
4239 4231
4240 4232 while (_evac_failure_scan_stack->length() > 0) {
4241 4233 oop obj = _evac_failure_scan_stack->pop();
4242 4234 _evac_failure_closure->set_region(heap_region_containing(obj));
4243 4235 obj->oop_iterate_backwards(_evac_failure_closure);
4244 4236 }
4245 4237 }
4246 4238
4247 4239 oop
4248 4240 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4249 4241 oop old) {
4250 4242 assert(obj_in_cs(old),
4251 4243 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4252 4244 (HeapWord*) old));
4253 4245 markOop m = old->mark();
4254 4246 oop forward_ptr = old->forward_to_atomic(old);
4255 4247 if (forward_ptr == NULL) {
4256 4248 // Forward-to-self succeeded.
4257 4249
4258 4250 if (_evac_failure_closure != cl) {
4259 4251 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4260 4252 assert(!_drain_in_progress,
4261 4253 "Should only be true while someone holds the lock.");
4262 4254 // Set the global evac-failure closure to the current thread's.
4263 4255 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4264 4256 set_evac_failure_closure(cl);
4265 4257 // Now do the common part.
4266 4258 handle_evacuation_failure_common(old, m);
4267 4259 // Reset to NULL.
4268 4260 set_evac_failure_closure(NULL);
4269 4261 } else {
4270 4262 // The lock is already held, and this is recursive.
4271 4263 assert(_drain_in_progress, "This should only be the recursive case.");
4272 4264 handle_evacuation_failure_common(old, m);
4273 4265 }
4274 4266 return old;
4275 4267 } else {
4276 4268 // Forward-to-self failed. Either someone else managed to allocate
4277 4269 // space for this object (old != forward_ptr) or they beat us in
4278 4270 // self-forwarding it (old == forward_ptr).
4279 4271 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4280 4272 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4281 4273 "should not be in the CSet",
4282 4274 (HeapWord*) old, (HeapWord*) forward_ptr));
4283 4275 return forward_ptr;
4284 4276 }
4285 4277 }
4286 4278
4287 4279 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4288 4280 set_evacuation_failed(true);
4289 4281
4290 4282 preserve_mark_if_necessary(old, m);
4291 4283
4292 4284 HeapRegion* r = heap_region_containing(old);
4293 4285 if (!r->evacuation_failed()) {
4294 4286 r->set_evacuation_failed(true);
4295 4287 _hr_printer.evac_failure(r);
4296 4288 }
4297 4289
4298 4290 push_on_evac_failure_scan_stack(old);
4299 4291
4300 4292 if (!_drain_in_progress) {
4301 4293 // prevent recursion in copy_to_survivor_space()
4302 4294 _drain_in_progress = true;
4303 4295 drain_evac_failure_scan_stack();
4304 4296 _drain_in_progress = false;
4305 4297 }
4306 4298 }
4307 4299
4308 4300 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4309 4301 assert(evacuation_failed(), "Oversaving!");
4310 4302 // We want to call the "for_promotion_failure" version only in the
4311 4303 // case of a promotion failure.
4312 4304 if (m->must_be_preserved_for_promotion_failure(obj)) {
4313 4305 if (_objs_with_preserved_marks == NULL) {
4314 4306 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4315 4307 _objs_with_preserved_marks =
4316 4308 new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4317 4309 _preserved_marks_of_objs =
4318 4310 new (ResourceObj::C_HEAP, mtGC) GrowableArray<markOop>(40, true);
4319 4311 }
4320 4312 _objs_with_preserved_marks->push(obj);
4321 4313 _preserved_marks_of_objs->push(m);
4322 4314 }
4323 4315 }
4324 4316
4325 4317 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4326 4318 size_t word_size) {
4327 4319 if (purpose == GCAllocForSurvived) {
4328 4320 HeapWord* result = survivor_attempt_allocation(word_size);
4329 4321 if (result != NULL) {
4330 4322 return result;
4331 4323 } else {
4332 4324 // Let's try to allocate in the old gen in case we can fit the
4333 4325 // object there.
4334 4326 return old_attempt_allocation(word_size);
4335 4327 }
4336 4328 } else {
4337 4329 assert(purpose == GCAllocForTenured, "sanity");
4338 4330 HeapWord* result = old_attempt_allocation(word_size);
4339 4331 if (result != NULL) {
4340 4332 return result;
4341 4333 } else {
4342 4334 // Let's try to allocate in the survivors in case we can fit the
4343 4335 // object there.
4344 4336 return survivor_attempt_allocation(word_size);
4345 4337 }
4346 4338 }
4347 4339
4348 4340 ShouldNotReachHere();
4349 4341 // Trying to keep some compilers happy.
4350 4342 return NULL;
4351 4343 }
4352 4344
4353 4345 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4354 4346 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4355 4347
4356 4348 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4357 4349 : _g1h(g1h),
4358 4350 _refs(g1h->task_queue(queue_num)),
4359 4351 _dcq(&g1h->dirty_card_queue_set()),
4360 4352 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4361 4353 _g1_rem(g1h->g1_rem_set()),
4362 4354 _hash_seed(17), _queue_num(queue_num),
4363 4355 _term_attempts(0),
4364 4356 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4365 4357 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4366 4358 _age_table(false),
4367 4359 _strong_roots_time(0), _term_time(0),
4368 4360 _alloc_buffer_waste(0), _undo_waste(0) {
4369 4361 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4370 4362 // we "sacrifice" entry 0 to keep track of surviving bytes for
4371 4363 // non-young regions (where the age is -1)
4372 4364 // We also add a few elements at the beginning and at the end in
4373 4365 // an attempt to eliminate cache contention
4374 4366 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4375 4367 uint array_length = PADDING_ELEM_NUM +
4376 4368 real_length +
4377 4369 PADDING_ELEM_NUM;
4378 4370 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4379 4371 if (_surviving_young_words_base == NULL)
4380 4372 vm_exit_out_of_memory(array_length * sizeof(size_t),
4381 4373 "Not enough space for young surv histo.");
4382 4374 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4383 4375 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4384 4376
4385 4377 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4386 4378 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4387 4379
4388 4380 _start = os::elapsedTime();
4389 4381 }
4390 4382
4391 4383 void
4392 4384 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4393 4385 {
4394 4386 st->print_raw_cr("GC Termination Stats");
4395 4387 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4396 4388 " ------waste (KiB)------");
4397 4389 st->print_raw_cr("thr ms ms % ms % attempts"
4398 4390 " total alloc undo");
4399 4391 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4400 4392 " ------- ------- -------");
4401 4393 }
4402 4394
4403 4395 void
4404 4396 G1ParScanThreadState::print_termination_stats(int i,
4405 4397 outputStream* const st) const
4406 4398 {
4407 4399 const double elapsed_ms = elapsed_time() * 1000.0;
4408 4400 const double s_roots_ms = strong_roots_time() * 1000.0;
4409 4401 const double term_ms = term_time() * 1000.0;
4410 4402 st->print_cr("%3d %9.2f %9.2f %6.2f "
4411 4403 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4412 4404 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4413 4405 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4414 4406 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4415 4407 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4416 4408 alloc_buffer_waste() * HeapWordSize / K,
4417 4409 undo_waste() * HeapWordSize / K);
4418 4410 }
4419 4411
4420 4412 #ifdef ASSERT
4421 4413 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4422 4414 assert(ref != NULL, "invariant");
4423 4415 assert(UseCompressedOops, "sanity");
4424 4416 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4425 4417 oop p = oopDesc::load_decode_heap_oop(ref);
4426 4418 assert(_g1h->is_in_g1_reserved(p),
4427 4419 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4428 4420 return true;
4429 4421 }
4430 4422
4431 4423 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4432 4424 assert(ref != NULL, "invariant");
4433 4425 if (has_partial_array_mask(ref)) {
4434 4426 // Must be in the collection set--it's already been copied.
4435 4427 oop p = clear_partial_array_mask(ref);
4436 4428 assert(_g1h->obj_in_cs(p),
4437 4429 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4438 4430 } else {
4439 4431 oop p = oopDesc::load_decode_heap_oop(ref);
4440 4432 assert(_g1h->is_in_g1_reserved(p),
4441 4433 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4442 4434 }
4443 4435 return true;
4444 4436 }
4445 4437
4446 4438 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4447 4439 if (ref.is_narrow()) {
4448 4440 return verify_ref((narrowOop*) ref);
4449 4441 } else {
4450 4442 return verify_ref((oop*) ref);
4451 4443 }
4452 4444 }
4453 4445 #endif // ASSERT
4454 4446
4455 4447 void G1ParScanThreadState::trim_queue() {
4456 4448 assert(_evac_cl != NULL, "not set");
4457 4449 assert(_evac_failure_cl != NULL, "not set");
4458 4450 assert(_partial_scan_cl != NULL, "not set");
4459 4451
4460 4452 StarTask ref;
4461 4453 do {
4462 4454 // Drain the overflow stack first, so other threads can steal.
4463 4455 while (refs()->pop_overflow(ref)) {
4464 4456 deal_with_reference(ref);
4465 4457 }
4466 4458
4467 4459 while (refs()->pop_local(ref)) {
4468 4460 deal_with_reference(ref);
4469 4461 }
4470 4462 } while (!refs()->is_empty());
4471 4463 }
4472 4464
4473 4465 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4474 4466 G1ParScanThreadState* par_scan_state) :
4475 4467 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4476 4468 _par_scan_state(par_scan_state),
4477 4469 _worker_id(par_scan_state->queue_num()),
4478 4470 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4479 4471 _mark_in_progress(_g1->mark_in_progress()) { }
4480 4472
4481 4473 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4482 4474 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4483 4475 #ifdef ASSERT
4484 4476 HeapRegion* hr = _g1->heap_region_containing(obj);
4485 4477 assert(hr != NULL, "sanity");
4486 4478 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4487 4479 #endif // ASSERT
4488 4480
4489 4481 // We know that the object is not moving so it's safe to read its size.
4490 4482 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4491 4483 }
4492 4484
4493 4485 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4494 4486 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4495 4487 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4496 4488 #ifdef ASSERT
4497 4489 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4498 4490 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4499 4491 assert(from_obj != to_obj, "should not be self-forwarded");
4500 4492
4501 4493 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4502 4494 assert(from_hr != NULL, "sanity");
4503 4495 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4504 4496
4505 4497 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4506 4498 assert(to_hr != NULL, "sanity");
4507 4499 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4508 4500 #endif // ASSERT
4509 4501
4510 4502 // The object might be in the process of being copied by another
4511 4503 // worker so we cannot trust that its to-space image is
4512 4504 // well-formed. So we have to read its size from its from-space
4513 4505 // image which we know should not be changing.
4514 4506 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4515 4507 }
4516 4508
4517 4509 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4518 4510 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4519 4511 ::copy_to_survivor_space(oop old) {
4520 4512 size_t word_sz = old->size();
4521 4513 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4522 4514 // +1 to make the -1 indexes valid...
4523 4515 int young_index = from_region->young_index_in_cset()+1;
4524 4516 assert( (from_region->is_young() && young_index > 0) ||
4525 4517 (!from_region->is_young() && young_index == 0), "invariant" );
4526 4518 G1CollectorPolicy* g1p = _g1->g1_policy();
4527 4519 markOop m = old->mark();
4528 4520 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4529 4521 : m->age();
4530 4522 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4531 4523 word_sz);
4532 4524 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4533 4525 #ifndef PRODUCT
4534 4526 // Should this evacuation fail?
4535 4527 if (_g1->evacuation_should_fail()) {
4536 4528 if (obj_ptr != NULL) {
4537 4529 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4538 4530 obj_ptr = NULL;
4539 4531 }
4540 4532 }
4541 4533 #endif // !PRODUCT
4542 4534
4543 4535 if (obj_ptr == NULL) {
4544 4536 // This will either forward-to-self, or detect that someone else has
4545 4537 // installed a forwarding pointer.
4546 4538 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4547 4539 return _g1->handle_evacuation_failure_par(cl, old);
4548 4540 }
4549 4541
4550 4542 oop obj = oop(obj_ptr);
4551 4543
4552 4544 // We're going to allocate linearly, so might as well prefetch ahead.
4553 4545 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4554 4546
4555 4547 oop forward_ptr = old->forward_to_atomic(obj);
4556 4548 if (forward_ptr == NULL) {
4557 4549 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4558 4550 if (g1p->track_object_age(alloc_purpose)) {
4559 4551 // We could simply do obj->incr_age(). However, this causes a
4560 4552 // performance issue. obj->incr_age() will first check whether
4561 4553 // the object has a displaced mark by checking its mark word;
4562 4554 // getting the mark word from the new location of the object
4563 4555 // stalls. So, given that we already have the mark word and we
4564 4556 // are about to install it anyway, it's better to increase the
4565 4557 // age on the mark word, when the object does not have a
4566 4558 // displaced mark word. We're not expecting many objects to have
4567 4559 // a displaced marked word, so that case is not optimized
4568 4560 // further (it could be...) and we simply call obj->incr_age().
4569 4561
4570 4562 if (m->has_displaced_mark_helper()) {
4571 4563 // in this case, we have to install the mark word first,
4572 4564 // otherwise obj looks to be forwarded (the old mark word,
4573 4565 // which contains the forward pointer, was copied)
4574 4566 obj->set_mark(m);
4575 4567 obj->incr_age();
4576 4568 } else {
4577 4569 m = m->incr_age();
4578 4570 obj->set_mark(m);
4579 4571 }
4580 4572 _par_scan_state->age_table()->add(obj, word_sz);
4581 4573 } else {
4582 4574 obj->set_mark(m);
4583 4575 }
4584 4576
4585 4577 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4586 4578 surv_young_words[young_index] += word_sz;
4587 4579
4588 4580 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4589 4581 // We keep track of the next start index in the length field of
4590 4582 // the to-space object. The actual length can be found in the
4591 4583 // length field of the from-space object.
4592 4584 arrayOop(obj)->set_length(0);
4593 4585 oop* old_p = set_partial_array_mask(old);
4594 4586 _par_scan_state->push_on_queue(old_p);
4595 4587 } else {
4596 4588 // No point in using the slower heap_region_containing() method,
4597 4589 // given that we know obj is in the heap.
4598 4590 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4599 4591 obj->oop_iterate_backwards(&_scanner);
4600 4592 }
4601 4593 } else {
4602 4594 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4603 4595 obj = forward_ptr;
4604 4596 }
4605 4597 return obj;
4606 4598 }
4607 4599
4608 4600 template <class T>
4609 4601 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4610 4602 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4611 4603 _scanned_klass->record_modified_oops();
4612 4604 }
4613 4605 }
4614 4606
4615 4607 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4616 4608 template <class T>
4617 4609 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4618 4610 ::do_oop_work(T* p) {
4619 4611 oop obj = oopDesc::load_decode_heap_oop(p);
4620 4612 assert(barrier != G1BarrierRS || obj != NULL,
4621 4613 "Precondition: G1BarrierRS implies obj is non-NULL");
4622 4614
4623 4615 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4624 4616
4625 4617 // here the null check is implicit in the cset_fast_test() test
4626 4618 if (_g1->in_cset_fast_test(obj)) {
4627 4619 oop forwardee;
4628 4620 if (obj->is_forwarded()) {
4629 4621 forwardee = obj->forwardee();
4630 4622 } else {
4631 4623 forwardee = copy_to_survivor_space(obj);
4632 4624 }
4633 4625 assert(forwardee != NULL, "forwardee should not be NULL");
4634 4626 oopDesc::encode_store_heap_oop(p, forwardee);
4635 4627 if (do_mark_object && forwardee != obj) {
4636 4628 // If the object is self-forwarded we don't need to explicitly
4637 4629 // mark it, the evacuation failure protocol will do so.
4638 4630 mark_forwarded_object(obj, forwardee);
4639 4631 }
4640 4632
4641 4633 // When scanning the RS, we only care about objs in CS.
4642 4634 if (barrier == G1BarrierRS) {
4643 4635 _par_scan_state->update_rs(_from, p, _worker_id);
4644 4636 } else if (barrier == G1BarrierKlass) {
4645 4637 do_klass_barrier(p, forwardee);
4646 4638 }
4647 4639 } else {
4648 4640 // The object is not in collection set. If we're a root scanning
4649 4641 // closure during an initial mark pause (i.e. do_mark_object will
4650 4642 // be true) then attempt to mark the object.
4651 4643 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4652 4644 mark_object(obj);
4653 4645 }
4654 4646 }
4655 4647
4656 4648 if (barrier == G1BarrierEvac && obj != NULL) {
4657 4649 _par_scan_state->update_rs(_from, p, _worker_id);
4658 4650 }
4659 4651
4660 4652 if (do_gen_barrier && obj != NULL) {
4661 4653 par_do_barrier(p);
4662 4654 }
4663 4655 }
4664 4656
4665 4657 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4666 4658 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4667 4659
4668 4660 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4669 4661 assert(has_partial_array_mask(p), "invariant");
4670 4662 oop from_obj = clear_partial_array_mask(p);
4671 4663
4672 4664 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4673 4665 assert(from_obj->is_objArray(), "must be obj array");
4674 4666 objArrayOop from_obj_array = objArrayOop(from_obj);
4675 4667 // The from-space object contains the real length.
4676 4668 int length = from_obj_array->length();
4677 4669
4678 4670 assert(from_obj->is_forwarded(), "must be forwarded");
4679 4671 oop to_obj = from_obj->forwardee();
4680 4672 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4681 4673 objArrayOop to_obj_array = objArrayOop(to_obj);
4682 4674 // We keep track of the next start index in the length field of the
4683 4675 // to-space object.
4684 4676 int next_index = to_obj_array->length();
4685 4677 assert(0 <= next_index && next_index < length,
4686 4678 err_msg("invariant, next index: %d, length: %d", next_index, length));
4687 4679
4688 4680 int start = next_index;
4689 4681 int end = length;
4690 4682 int remainder = end - start;
4691 4683 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4692 4684 if (remainder > 2 * ParGCArrayScanChunk) {
4693 4685 end = start + ParGCArrayScanChunk;
4694 4686 to_obj_array->set_length(end);
4695 4687 // Push the remainder before we process the range in case another
4696 4688 // worker has run out of things to do and can steal it.
4697 4689 oop* from_obj_p = set_partial_array_mask(from_obj);
4698 4690 _par_scan_state->push_on_queue(from_obj_p);
4699 4691 } else {
4700 4692 assert(length == end, "sanity");
4701 4693 // We'll process the final range for this object. Restore the length
4702 4694 // so that the heap remains parsable in case of evacuation failure.
4703 4695 to_obj_array->set_length(end);
4704 4696 }
4705 4697 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4706 4698 // Process indexes [start,end). It will also process the header
4707 4699 // along with the first chunk (i.e., the chunk with start == 0).
4708 4700 // Note that at this point the length field of to_obj_array is not
4709 4701 // correct given that we are using it to keep track of the next
4710 4702 // start index. oop_iterate_range() (thankfully!) ignores the length
4711 4703 // field and only relies on the start / end parameters. It does
4712 4704 // however return the size of the object which will be incorrect. So
4713 4705 // we have to ignore it even if we wanted to use it.
4714 4706 to_obj_array->oop_iterate_range(&_scanner, start, end);
4715 4707 }
4716 4708
4717 4709 class G1ParEvacuateFollowersClosure : public VoidClosure {
4718 4710 protected:
4719 4711 G1CollectedHeap* _g1h;
4720 4712 G1ParScanThreadState* _par_scan_state;
4721 4713 RefToScanQueueSet* _queues;
4722 4714 ParallelTaskTerminator* _terminator;
4723 4715
4724 4716 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4725 4717 RefToScanQueueSet* queues() { return _queues; }
4726 4718 ParallelTaskTerminator* terminator() { return _terminator; }
4727 4719
4728 4720 public:
4729 4721 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4730 4722 G1ParScanThreadState* par_scan_state,
4731 4723 RefToScanQueueSet* queues,
4732 4724 ParallelTaskTerminator* terminator)
4733 4725 : _g1h(g1h), _par_scan_state(par_scan_state),
4734 4726 _queues(queues), _terminator(terminator) {}
4735 4727
4736 4728 void do_void();
4737 4729
4738 4730 private:
4739 4731 inline bool offer_termination();
4740 4732 };
4741 4733
4742 4734 bool G1ParEvacuateFollowersClosure::offer_termination() {
4743 4735 G1ParScanThreadState* const pss = par_scan_state();
4744 4736 pss->start_term_time();
4745 4737 const bool res = terminator()->offer_termination();
4746 4738 pss->end_term_time();
4747 4739 return res;
4748 4740 }
4749 4741
4750 4742 void G1ParEvacuateFollowersClosure::do_void() {
4751 4743 StarTask stolen_task;
4752 4744 G1ParScanThreadState* const pss = par_scan_state();
4753 4745 pss->trim_queue();
4754 4746
4755 4747 do {
4756 4748 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4757 4749 assert(pss->verify_task(stolen_task), "sanity");
4758 4750 if (stolen_task.is_narrow()) {
4759 4751 pss->deal_with_reference((narrowOop*) stolen_task);
4760 4752 } else {
4761 4753 pss->deal_with_reference((oop*) stolen_task);
4762 4754 }
4763 4755
4764 4756 // We've just processed a reference and we might have made
4765 4757 // available new entries on the queues. So we have to make sure
4766 4758 // we drain the queues as necessary.
4767 4759 pss->trim_queue();
4768 4760 }
4769 4761 } while (!offer_termination());
4770 4762
4771 4763 pss->retire_alloc_buffers();
4772 4764 }
4773 4765
4774 4766 class G1KlassScanClosure : public KlassClosure {
4775 4767 G1ParCopyHelper* _closure;
4776 4768 bool _process_only_dirty;
4777 4769 int _count;
4778 4770 public:
4779 4771 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4780 4772 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4781 4773 void do_klass(Klass* klass) {
4782 4774 // If the klass has not been dirtied we know that there's
4783 4775 // no references into the young gen and we can skip it.
4784 4776 if (!_process_only_dirty || klass->has_modified_oops()) {
4785 4777 // Clean the klass since we're going to scavenge all the metadata.
4786 4778 klass->clear_modified_oops();
4787 4779
4788 4780 // Tell the closure that this klass is the Klass to scavenge
4789 4781 // and is the one to dirty if oops are left pointing into the young gen.
4790 4782 _closure->set_scanned_klass(klass);
4791 4783
4792 4784 klass->oops_do(_closure);
4793 4785
4794 4786 _closure->set_scanned_klass(NULL);
4795 4787 }
4796 4788 _count++;
4797 4789 }
4798 4790 };
4799 4791
4800 4792 class G1ParTask : public AbstractGangTask {
4801 4793 protected:
4802 4794 G1CollectedHeap* _g1h;
4803 4795 RefToScanQueueSet *_queues;
4804 4796 ParallelTaskTerminator _terminator;
4805 4797 uint _n_workers;
4806 4798
4807 4799 Mutex _stats_lock;
4808 4800 Mutex* stats_lock() { return &_stats_lock; }
4809 4801
4810 4802 size_t getNCards() {
4811 4803 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4812 4804 / G1BlockOffsetSharedArray::N_bytes;
4813 4805 }
4814 4806
4815 4807 public:
4816 4808 G1ParTask(G1CollectedHeap* g1h,
4817 4809 RefToScanQueueSet *task_queues)
4818 4810 : AbstractGangTask("G1 collection"),
4819 4811 _g1h(g1h),
4820 4812 _queues(task_queues),
4821 4813 _terminator(0, _queues),
4822 4814 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4823 4815 {}
4824 4816
4825 4817 RefToScanQueueSet* queues() { return _queues; }
4826 4818
4827 4819 RefToScanQueue *work_queue(int i) {
4828 4820 return queues()->queue(i);
4829 4821 }
4830 4822
4831 4823 ParallelTaskTerminator* terminator() { return &_terminator; }
4832 4824
4833 4825 virtual void set_for_termination(int active_workers) {
4834 4826 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4835 4827 // in the young space (_par_seq_tasks) in the G1 heap
4836 4828 // for SequentialSubTasksDone.
4837 4829 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4838 4830 // both of which need setting by set_n_termination().
4839 4831 _g1h->SharedHeap::set_n_termination(active_workers);
4840 4832 _g1h->set_n_termination(active_workers);
4841 4833 terminator()->reset_for_reuse(active_workers);
4842 4834 _n_workers = active_workers;
4843 4835 }
4844 4836
4845 4837 void work(uint worker_id) {
4846 4838 if (worker_id >= _n_workers) return; // no work needed this round
4847 4839
4848 4840 double start_time_ms = os::elapsedTime() * 1000.0;
4849 4841 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4850 4842
4851 4843 {
4852 4844 ResourceMark rm;
4853 4845 HandleMark hm;
4854 4846
4855 4847 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4856 4848
4857 4849 G1ParScanThreadState pss(_g1h, worker_id);
4858 4850 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4859 4851 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4860 4852 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4861 4853
4862 4854 pss.set_evac_closure(&scan_evac_cl);
4863 4855 pss.set_evac_failure_closure(&evac_failure_cl);
4864 4856 pss.set_partial_scan_closure(&partial_scan_cl);
4865 4857
4866 4858 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4867 4859 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4868 4860
4869 4861 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4870 4862 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4871 4863
4872 4864 bool only_young = _g1h->g1_policy()->gcs_are_young();
4873 4865 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4874 4866 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4875 4867
4876 4868 OopClosure* scan_root_cl = &only_scan_root_cl;
4877 4869 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4878 4870
4879 4871 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4880 4872 // We also need to mark copied objects.
4881 4873 scan_root_cl = &scan_mark_root_cl;
4882 4874 scan_klasses_cl = &scan_mark_klasses_cl_s;
4883 4875 }
4884 4876
4885 4877 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4886 4878
4887 4879 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4888 4880
4889 4881 pss.start_strong_roots();
4890 4882 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4891 4883 SharedHeap::ScanningOption(so),
4892 4884 scan_root_cl,
4893 4885 &push_heap_rs_cl,
4894 4886 scan_klasses_cl,
4895 4887 worker_id);
4896 4888 pss.end_strong_roots();
4897 4889
4898 4890 {
4899 4891 double start = os::elapsedTime();
4900 4892 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4901 4893 evac.do_void();
4902 4894 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4903 4895 double term_ms = pss.term_time()*1000.0;
4904 4896 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4905 4897 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4906 4898 }
4907 4899 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4908 4900 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4909 4901
4910 4902 if (ParallelGCVerbose) {
4911 4903 MutexLocker x(stats_lock());
4912 4904 pss.print_termination_stats(worker_id);
4913 4905 }
4914 4906
4915 4907 assert(pss.refs()->is_empty(), "should be empty");
4916 4908
4917 4909 // Close the inner scope so that the ResourceMark and HandleMark
4918 4910 // destructors are executed here and are included as part of the
4919 4911 // "GC Worker Time".
4920 4912 }
4921 4913
4922 4914 double end_time_ms = os::elapsedTime() * 1000.0;
4923 4915 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4924 4916 }
4925 4917 };
4926 4918
4927 4919 // *** Common G1 Evacuation Stuff
4928 4920
4929 4921 // Closures that support the filtering of CodeBlobs scanned during
4930 4922 // external root scanning.
4931 4923
4932 4924 // Closure applied to reference fields in code blobs (specifically nmethods)
4933 4925 // to determine whether an nmethod contains references that point into
4934 4926 // the collection set. Used as a predicate when walking code roots so
4935 4927 // that only nmethods that point into the collection set are added to the
4936 4928 // 'marked' list.
4937 4929
4938 4930 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4939 4931
4940 4932 class G1PointsIntoCSOopClosure : public OopClosure {
4941 4933 G1CollectedHeap* _g1;
4942 4934 bool _points_into_cs;
4943 4935 public:
4944 4936 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4945 4937 _g1(g1), _points_into_cs(false) { }
4946 4938
4947 4939 bool points_into_cs() const { return _points_into_cs; }
4948 4940
4949 4941 template <class T>
4950 4942 void do_oop_nv(T* p) {
4951 4943 if (!_points_into_cs) {
4952 4944 T heap_oop = oopDesc::load_heap_oop(p);
4953 4945 if (!oopDesc::is_null(heap_oop) &&
4954 4946 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4955 4947 _points_into_cs = true;
4956 4948 }
4957 4949 }
4958 4950 }
4959 4951
4960 4952 virtual void do_oop(oop* p) { do_oop_nv(p); }
4961 4953 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4962 4954 };
4963 4955
4964 4956 G1CollectedHeap* _g1;
4965 4957
4966 4958 public:
4967 4959 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4968 4960 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4969 4961
4970 4962 virtual void do_code_blob(CodeBlob* cb) {
4971 4963 nmethod* nm = cb->as_nmethod_or_null();
4972 4964 if (nm != NULL && !(nm->test_oops_do_mark())) {
4973 4965 G1PointsIntoCSOopClosure predicate_cl(_g1);
4974 4966 nm->oops_do(&predicate_cl);
4975 4967
4976 4968 if (predicate_cl.points_into_cs()) {
4977 4969 // At least one of the reference fields or the oop relocations
4978 4970 // in the nmethod points into the collection set. We have to
4979 4971 // 'mark' this nmethod.
4980 4972 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4981 4973 // or MarkingCodeBlobClosure::do_code_blob() change.
4982 4974 if (!nm->test_set_oops_do_mark()) {
4983 4975 do_newly_marked_nmethod(nm);
4984 4976 }
4985 4977 }
4986 4978 }
4987 4979 }
4988 4980 };
4989 4981
4990 4982 // This method is run in a GC worker.
4991 4983
4992 4984 void
4993 4985 G1CollectedHeap::
4994 4986 g1_process_strong_roots(bool is_scavenging,
4995 4987 ScanningOption so,
4996 4988 OopClosure* scan_non_heap_roots,
4997 4989 OopsInHeapRegionClosure* scan_rs,
4998 4990 G1KlassScanClosure* scan_klasses,
4999 4991 int worker_i) {
5000 4992
5001 4993 // First scan the strong roots
5002 4994 double ext_roots_start = os::elapsedTime();
5003 4995 double closure_app_time_sec = 0.0;
5004 4996
5005 4997 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5006 4998
5007 4999 // Walk the code cache w/o buffering, because StarTask cannot handle
5008 5000 // unaligned oop locations.
5009 5001 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5010 5002
5011 5003 process_strong_roots(false, // no scoping; this is parallel code
5012 5004 is_scavenging, so,
5013 5005 &buf_scan_non_heap_roots,
5014 5006 &eager_scan_code_roots,
5015 5007 scan_klasses
5016 5008 );
5017 5009
5018 5010 // Now the CM ref_processor roots.
5019 5011 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5020 5012 // We need to treat the discovered reference lists of the
5021 5013 // concurrent mark ref processor as roots and keep entries
5022 5014 // (which are added by the marking threads) on them live
5023 5015 // until they can be processed at the end of marking.
5024 5016 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5025 5017 }
5026 5018
5027 5019 // Finish up any enqueued closure apps (attributed as object copy time).
5028 5020 buf_scan_non_heap_roots.done();
5029 5021
5030 5022 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5031 5023
5032 5024 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5033 5025
5034 5026 double ext_root_time_ms =
5035 5027 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5036 5028
5037 5029 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5038 5030
5039 5031 // During conc marking we have to filter the per-thread SATB buffers
5040 5032 // to make sure we remove any oops into the CSet (which will show up
5041 5033 // as implicitly live).
5042 5034 double satb_filtering_ms = 0.0;
5043 5035 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5044 5036 if (mark_in_progress()) {
5045 5037 double satb_filter_start = os::elapsedTime();
5046 5038
5047 5039 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5048 5040
5049 5041 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5050 5042 }
5051 5043 }
5052 5044 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5053 5045
5054 5046 // Now scan the complement of the collection set.
5055 5047 if (scan_rs != NULL) {
5056 5048 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5057 5049 }
5058 5050 _process_strong_tasks->all_tasks_completed();
5059 5051 }
5060 5052
5061 5053 void
5062 5054 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5063 5055 OopClosure* non_root_closure) {
5064 5056 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5065 5057 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5066 5058 }
5067 5059
5068 5060 // Weak Reference Processing support
5069 5061
5070 5062 // An always "is_alive" closure that is used to preserve referents.
5071 5063 // If the object is non-null then it's alive. Used in the preservation
5072 5064 // of referent objects that are pointed to by reference objects
5073 5065 // discovered by the CM ref processor.
5074 5066 class G1AlwaysAliveClosure: public BoolObjectClosure {
5075 5067 G1CollectedHeap* _g1;
5076 5068 public:
5077 5069 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5078 5070 void do_object(oop p) { assert(false, "Do not call."); }
5079 5071 bool do_object_b(oop p) {
5080 5072 if (p != NULL) {
5081 5073 return true;
5082 5074 }
5083 5075 return false;
5084 5076 }
5085 5077 };
5086 5078
5087 5079 bool G1STWIsAliveClosure::do_object_b(oop p) {
5088 5080 // An object is reachable if it is outside the collection set,
5089 5081 // or is inside and copied.
5090 5082 return !_g1->obj_in_cs(p) || p->is_forwarded();
5091 5083 }
5092 5084
5093 5085 // Non Copying Keep Alive closure
5094 5086 class G1KeepAliveClosure: public OopClosure {
5095 5087 G1CollectedHeap* _g1;
5096 5088 public:
5097 5089 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5098 5090 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5099 5091 void do_oop( oop* p) {
5100 5092 oop obj = *p;
5101 5093
5102 5094 if (_g1->obj_in_cs(obj)) {
5103 5095 assert( obj->is_forwarded(), "invariant" );
5104 5096 *p = obj->forwardee();
5105 5097 }
5106 5098 }
5107 5099 };
5108 5100
5109 5101 // Copying Keep Alive closure - can be called from both
5110 5102 // serial and parallel code as long as different worker
5111 5103 // threads utilize different G1ParScanThreadState instances
5112 5104 // and different queues.
5113 5105
5114 5106 class G1CopyingKeepAliveClosure: public OopClosure {
5115 5107 G1CollectedHeap* _g1h;
5116 5108 OopClosure* _copy_non_heap_obj_cl;
5117 5109 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5118 5110 G1ParScanThreadState* _par_scan_state;
5119 5111
5120 5112 public:
5121 5113 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5122 5114 OopClosure* non_heap_obj_cl,
5123 5115 OopsInHeapRegionClosure* metadata_obj_cl,
5124 5116 G1ParScanThreadState* pss):
5125 5117 _g1h(g1h),
5126 5118 _copy_non_heap_obj_cl(non_heap_obj_cl),
5127 5119 _copy_metadata_obj_cl(metadata_obj_cl),
5128 5120 _par_scan_state(pss)
5129 5121 {}
5130 5122
5131 5123 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5132 5124 virtual void do_oop( oop* p) { do_oop_work(p); }
5133 5125
5134 5126 template <class T> void do_oop_work(T* p) {
5135 5127 oop obj = oopDesc::load_decode_heap_oop(p);
5136 5128
5137 5129 if (_g1h->obj_in_cs(obj)) {
5138 5130 // If the referent object has been forwarded (either copied
5139 5131 // to a new location or to itself in the event of an
5140 5132 // evacuation failure) then we need to update the reference
5141 5133 // field and, if both reference and referent are in the G1
5142 5134 // heap, update the RSet for the referent.
5143 5135 //
5144 5136 // If the referent has not been forwarded then we have to keep
5145 5137 // it alive by policy. Therefore we have copy the referent.
5146 5138 //
5147 5139 // If the reference field is in the G1 heap then we can push
5148 5140 // on the PSS queue. When the queue is drained (after each
5149 5141 // phase of reference processing) the object and it's followers
5150 5142 // will be copied, the reference field set to point to the
5151 5143 // new location, and the RSet updated. Otherwise we need to
5152 5144 // use the the non-heap or metadata closures directly to copy
5153 5145 // the refernt object and update the pointer, while avoiding
5154 5146 // updating the RSet.
5155 5147
5156 5148 if (_g1h->is_in_g1_reserved(p)) {
5157 5149 _par_scan_state->push_on_queue(p);
5158 5150 } else {
5159 5151 assert(!ClassLoaderDataGraph::contains((address)p),
5160 5152 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5161 5153 PTR_FORMAT, p));
5162 5154 _copy_non_heap_obj_cl->do_oop(p);
5163 5155 }
5164 5156 }
5165 5157 }
5166 5158 };
5167 5159
5168 5160 // Serial drain queue closure. Called as the 'complete_gc'
5169 5161 // closure for each discovered list in some of the
5170 5162 // reference processing phases.
5171 5163
5172 5164 class G1STWDrainQueueClosure: public VoidClosure {
5173 5165 protected:
5174 5166 G1CollectedHeap* _g1h;
5175 5167 G1ParScanThreadState* _par_scan_state;
5176 5168
5177 5169 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5178 5170
5179 5171 public:
5180 5172 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5181 5173 _g1h(g1h),
5182 5174 _par_scan_state(pss)
5183 5175 { }
5184 5176
5185 5177 void do_void() {
5186 5178 G1ParScanThreadState* const pss = par_scan_state();
5187 5179 pss->trim_queue();
5188 5180 }
5189 5181 };
5190 5182
5191 5183 // Parallel Reference Processing closures
5192 5184
5193 5185 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5194 5186 // processing during G1 evacuation pauses.
5195 5187
5196 5188 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5197 5189 private:
5198 5190 G1CollectedHeap* _g1h;
5199 5191 RefToScanQueueSet* _queues;
5200 5192 FlexibleWorkGang* _workers;
5201 5193 int _active_workers;
5202 5194
5203 5195 public:
5204 5196 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5205 5197 FlexibleWorkGang* workers,
5206 5198 RefToScanQueueSet *task_queues,
5207 5199 int n_workers) :
5208 5200 _g1h(g1h),
5209 5201 _queues(task_queues),
5210 5202 _workers(workers),
5211 5203 _active_workers(n_workers)
5212 5204 {
5213 5205 assert(n_workers > 0, "shouldn't call this otherwise");
5214 5206 }
5215 5207
5216 5208 // Executes the given task using concurrent marking worker threads.
5217 5209 virtual void execute(ProcessTask& task);
5218 5210 virtual void execute(EnqueueTask& task);
5219 5211 };
5220 5212
5221 5213 // Gang task for possibly parallel reference processing
5222 5214
5223 5215 class G1STWRefProcTaskProxy: public AbstractGangTask {
5224 5216 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5225 5217 ProcessTask& _proc_task;
5226 5218 G1CollectedHeap* _g1h;
5227 5219 RefToScanQueueSet *_task_queues;
5228 5220 ParallelTaskTerminator* _terminator;
5229 5221
5230 5222 public:
5231 5223 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5232 5224 G1CollectedHeap* g1h,
5233 5225 RefToScanQueueSet *task_queues,
5234 5226 ParallelTaskTerminator* terminator) :
5235 5227 AbstractGangTask("Process reference objects in parallel"),
5236 5228 _proc_task(proc_task),
5237 5229 _g1h(g1h),
5238 5230 _task_queues(task_queues),
5239 5231 _terminator(terminator)
5240 5232 {}
5241 5233
5242 5234 virtual void work(uint worker_id) {
5243 5235 // The reference processing task executed by a single worker.
5244 5236 ResourceMark rm;
5245 5237 HandleMark hm;
5246 5238
5247 5239 G1STWIsAliveClosure is_alive(_g1h);
5248 5240
5249 5241 G1ParScanThreadState pss(_g1h, worker_id);
5250 5242
5251 5243 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5252 5244 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5253 5245 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5254 5246
5255 5247 pss.set_evac_closure(&scan_evac_cl);
5256 5248 pss.set_evac_failure_closure(&evac_failure_cl);
5257 5249 pss.set_partial_scan_closure(&partial_scan_cl);
5258 5250
5259 5251 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5260 5252 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5261 5253
5262 5254 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5263 5255 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5264 5256
5265 5257 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5266 5258 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5267 5259
5268 5260 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5269 5261 // We also need to mark copied objects.
5270 5262 copy_non_heap_cl = ©_mark_non_heap_cl;
5271 5263 copy_metadata_cl = ©_mark_metadata_cl;
5272 5264 }
5273 5265
5274 5266 // Keep alive closure.
5275 5267 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5276 5268
5277 5269 // Complete GC closure
5278 5270 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5279 5271
5280 5272 // Call the reference processing task's work routine.
5281 5273 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5282 5274
5283 5275 // Note we cannot assert that the refs array is empty here as not all
5284 5276 // of the processing tasks (specifically phase2 - pp2_work) execute
5285 5277 // the complete_gc closure (which ordinarily would drain the queue) so
5286 5278 // the queue may not be empty.
5287 5279 }
5288 5280 };
5289 5281
5290 5282 // Driver routine for parallel reference processing.
5291 5283 // Creates an instance of the ref processing gang
5292 5284 // task and has the worker threads execute it.
5293 5285 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5294 5286 assert(_workers != NULL, "Need parallel worker threads.");
5295 5287
5296 5288 ParallelTaskTerminator terminator(_active_workers, _queues);
5297 5289 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5298 5290
5299 5291 _g1h->set_par_threads(_active_workers);
5300 5292 _workers->run_task(&proc_task_proxy);
5301 5293 _g1h->set_par_threads(0);
5302 5294 }
5303 5295
5304 5296 // Gang task for parallel reference enqueueing.
5305 5297
5306 5298 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5307 5299 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5308 5300 EnqueueTask& _enq_task;
5309 5301
5310 5302 public:
5311 5303 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5312 5304 AbstractGangTask("Enqueue reference objects in parallel"),
5313 5305 _enq_task(enq_task)
5314 5306 { }
5315 5307
5316 5308 virtual void work(uint worker_id) {
5317 5309 _enq_task.work(worker_id);
5318 5310 }
5319 5311 };
5320 5312
5321 5313 // Driver routine for parallel reference enqueing.
5322 5314 // Creates an instance of the ref enqueueing gang
5323 5315 // task and has the worker threads execute it.
5324 5316
5325 5317 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5326 5318 assert(_workers != NULL, "Need parallel worker threads.");
5327 5319
5328 5320 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5329 5321
5330 5322 _g1h->set_par_threads(_active_workers);
5331 5323 _workers->run_task(&enq_task_proxy);
5332 5324 _g1h->set_par_threads(0);
5333 5325 }
5334 5326
5335 5327 // End of weak reference support closures
5336 5328
5337 5329 // Abstract task used to preserve (i.e. copy) any referent objects
5338 5330 // that are in the collection set and are pointed to by reference
5339 5331 // objects discovered by the CM ref processor.
5340 5332
5341 5333 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5342 5334 protected:
5343 5335 G1CollectedHeap* _g1h;
5344 5336 RefToScanQueueSet *_queues;
5345 5337 ParallelTaskTerminator _terminator;
5346 5338 uint _n_workers;
5347 5339
5348 5340 public:
5349 5341 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5350 5342 AbstractGangTask("ParPreserveCMReferents"),
5351 5343 _g1h(g1h),
5352 5344 _queues(task_queues),
5353 5345 _terminator(workers, _queues),
5354 5346 _n_workers(workers)
5355 5347 { }
5356 5348
5357 5349 void work(uint worker_id) {
5358 5350 ResourceMark rm;
5359 5351 HandleMark hm;
5360 5352
5361 5353 G1ParScanThreadState pss(_g1h, worker_id);
5362 5354 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5363 5355 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5364 5356 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5365 5357
5366 5358 pss.set_evac_closure(&scan_evac_cl);
5367 5359 pss.set_evac_failure_closure(&evac_failure_cl);
5368 5360 pss.set_partial_scan_closure(&partial_scan_cl);
5369 5361
5370 5362 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5371 5363
5372 5364
5373 5365 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5374 5366 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5375 5367
5376 5368 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5377 5369 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5378 5370
5379 5371 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5380 5372 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5381 5373
5382 5374 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5383 5375 // We also need to mark copied objects.
5384 5376 copy_non_heap_cl = ©_mark_non_heap_cl;
5385 5377 copy_metadata_cl = ©_mark_metadata_cl;
5386 5378 }
5387 5379
5388 5380 // Is alive closure
5389 5381 G1AlwaysAliveClosure always_alive(_g1h);
5390 5382
5391 5383 // Copying keep alive closure. Applied to referent objects that need
5392 5384 // to be copied.
5393 5385 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5394 5386
5395 5387 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5396 5388
5397 5389 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5398 5390 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5399 5391
5400 5392 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5401 5393 // So this must be true - but assert just in case someone decides to
5402 5394 // change the worker ids.
5403 5395 assert(0 <= worker_id && worker_id < limit, "sanity");
5404 5396 assert(!rp->discovery_is_atomic(), "check this code");
5405 5397
5406 5398 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5407 5399 for (uint idx = worker_id; idx < limit; idx += stride) {
5408 5400 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5409 5401
5410 5402 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5411 5403 while (iter.has_next()) {
5412 5404 // Since discovery is not atomic for the CM ref processor, we
5413 5405 // can see some null referent objects.
5414 5406 iter.load_ptrs(DEBUG_ONLY(true));
5415 5407 oop ref = iter.obj();
5416 5408
5417 5409 // This will filter nulls.
5418 5410 if (iter.is_referent_alive()) {
5419 5411 iter.make_referent_alive();
5420 5412 }
5421 5413 iter.move_to_next();
5422 5414 }
5423 5415 }
5424 5416
5425 5417 // Drain the queue - which may cause stealing
5426 5418 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5427 5419 drain_queue.do_void();
5428 5420 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5429 5421 assert(pss.refs()->is_empty(), "should be");
5430 5422 }
5431 5423 };
5432 5424
5433 5425 // Weak Reference processing during an evacuation pause (part 1).
5434 5426 void G1CollectedHeap::process_discovered_references() {
5435 5427 double ref_proc_start = os::elapsedTime();
5436 5428
5437 5429 ReferenceProcessor* rp = _ref_processor_stw;
5438 5430 assert(rp->discovery_enabled(), "should have been enabled");
5439 5431
5440 5432 // Any reference objects, in the collection set, that were 'discovered'
5441 5433 // by the CM ref processor should have already been copied (either by
5442 5434 // applying the external root copy closure to the discovered lists, or
5443 5435 // by following an RSet entry).
5444 5436 //
5445 5437 // But some of the referents, that are in the collection set, that these
5446 5438 // reference objects point to may not have been copied: the STW ref
5447 5439 // processor would have seen that the reference object had already
5448 5440 // been 'discovered' and would have skipped discovering the reference,
5449 5441 // but would not have treated the reference object as a regular oop.
5450 5442 // As a reult the copy closure would not have been applied to the
5451 5443 // referent object.
5452 5444 //
5453 5445 // We need to explicitly copy these referent objects - the references
5454 5446 // will be processed at the end of remarking.
5455 5447 //
5456 5448 // We also need to do this copying before we process the reference
5457 5449 // objects discovered by the STW ref processor in case one of these
5458 5450 // referents points to another object which is also referenced by an
5459 5451 // object discovered by the STW ref processor.
5460 5452
5461 5453 uint active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5462 5454 workers()->active_workers() : 1);
5463 5455
5464 5456 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5465 5457 active_workers == workers()->active_workers(),
5466 5458 "Need to reset active_workers");
5467 5459
5468 5460 set_par_threads(active_workers);
5469 5461 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues);
5470 5462
5471 5463 if (G1CollectedHeap::use_parallel_gc_threads()) {
5472 5464 workers()->run_task(&keep_cm_referents);
5473 5465 } else {
5474 5466 keep_cm_referents.work(0);
5475 5467 }
5476 5468
5477 5469 set_par_threads(0);
5478 5470
5479 5471 // Closure to test whether a referent is alive.
5480 5472 G1STWIsAliveClosure is_alive(this);
5481 5473
5482 5474 // Even when parallel reference processing is enabled, the processing
5483 5475 // of JNI refs is serial and performed serially by the current thread
5484 5476 // rather than by a worker. The following PSS will be used for processing
5485 5477 // JNI refs.
5486 5478
5487 5479 // Use only a single queue for this PSS.
5488 5480 G1ParScanThreadState pss(this, 0);
5489 5481
5490 5482 // We do not embed a reference processor in the copying/scanning
5491 5483 // closures while we're actually processing the discovered
5492 5484 // reference objects.
5493 5485 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5494 5486 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5495 5487 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5496 5488
5497 5489 pss.set_evac_closure(&scan_evac_cl);
5498 5490 pss.set_evac_failure_closure(&evac_failure_cl);
5499 5491 pss.set_partial_scan_closure(&partial_scan_cl);
5500 5492
5501 5493 assert(pss.refs()->is_empty(), "pre-condition");
5502 5494
5503 5495 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5504 5496 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5505 5497
5506 5498 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5507 5499 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5508 5500
5509 5501 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5510 5502 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5511 5503
5512 5504 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5513 5505 // We also need to mark copied objects.
5514 5506 copy_non_heap_cl = ©_mark_non_heap_cl;
5515 5507 copy_metadata_cl = ©_mark_metadata_cl;
5516 5508 }
5517 5509
5518 5510 // Keep alive closure.
5519 5511 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5520 5512
5521 5513 // Serial Complete GC closure
5522 5514 G1STWDrainQueueClosure drain_queue(this, &pss);
5523 5515
5524 5516 // Setup the soft refs policy...
5525 5517 rp->setup_policy(false);
5526 5518
5527 5519 if (!rp->processing_is_mt()) {
5528 5520 // Serial reference processing...
5529 5521 rp->process_discovered_references(&is_alive,
5530 5522 &keep_alive,
5531 5523 &drain_queue,
5532 5524 NULL);
5533 5525 } else {
5534 5526 // Parallel reference processing
5535 5527 assert(rp->num_q() == active_workers, "sanity");
5536 5528 assert(active_workers <= rp->max_num_q(), "sanity");
5537 5529
5538 5530 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5539 5531 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5540 5532 }
5541 5533
5542 5534 // We have completed copying any necessary live referent objects
5543 5535 // (that were not copied during the actual pause) so we can
5544 5536 // retire any active alloc buffers
5545 5537 pss.retire_alloc_buffers();
5546 5538 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5547 5539
5548 5540 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5549 5541 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5550 5542 }
5551 5543
5552 5544 // Weak Reference processing during an evacuation pause (part 2).
5553 5545 void G1CollectedHeap::enqueue_discovered_references() {
5554 5546 double ref_enq_start = os::elapsedTime();
5555 5547
5556 5548 ReferenceProcessor* rp = _ref_processor_stw;
5557 5549 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5558 5550
5559 5551 // Now enqueue any remaining on the discovered lists on to
5560 5552 // the pending list.
5561 5553 if (!rp->processing_is_mt()) {
5562 5554 // Serial reference processing...
5563 5555 rp->enqueue_discovered_references();
5564 5556 } else {
5565 5557 // Parallel reference enqueuing
5566 5558
5567 5559 uint active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1);
5568 5560 assert(active_workers == workers()->active_workers(),
5569 5561 "Need to reset active_workers");
5570 5562 assert(rp->num_q() == active_workers, "sanity");
5571 5563 assert(active_workers <= rp->max_num_q(), "sanity");
5572 5564
5573 5565 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5574 5566 rp->enqueue_discovered_references(&par_task_executor);
5575 5567 }
5576 5568
5577 5569 rp->verify_no_references_recorded();
5578 5570 assert(!rp->discovery_enabled(), "should have been disabled");
5579 5571
5580 5572 // FIXME
5581 5573 // CM's reference processing also cleans up the string and symbol tables.
5582 5574 // Should we do that here also? We could, but it is a serial operation
5583 5575 // and could signicantly increase the pause time.
5584 5576
5585 5577 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5586 5578 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5587 5579 }
5588 5580
5589 5581 void G1CollectedHeap::evacuate_collection_set() {
5590 5582 _expand_heap_after_alloc_failure = true;
5591 5583 set_evacuation_failed(false);
5592 5584
5593 5585 // Should G1EvacuationFailureALot be in effect for this GC?
5594 5586 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5595 5587
5596 5588 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5597 5589 concurrent_g1_refine()->set_use_cache(false);
5598 5590 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5599 5591
5600 5592 uint n_workers;
5601 5593 if (G1CollectedHeap::use_parallel_gc_threads()) {
5602 5594 n_workers =
5603 5595 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5604 5596 workers()->active_workers(),
5605 5597 Threads::number_of_non_daemon_threads());
5606 5598 assert(UseDynamicNumberOfGCThreads ||
5607 5599 n_workers == workers()->total_workers(),
5608 5600 "If not dynamic should be using all the workers");
5609 5601 workers()->set_active_workers(n_workers);
5610 5602 set_par_threads(n_workers);
5611 5603 } else {
5612 5604 assert(n_par_threads() == 0,
5613 5605 "Should be the original non-parallel value");
5614 5606 n_workers = 1;
5615 5607 }
5616 5608
5617 5609 G1ParTask g1_par_task(this, _task_queues);
5618 5610
5619 5611 init_for_evac_failure(NULL);
5620 5612
5621 5613 rem_set()->prepare_for_younger_refs_iterate(true);
5622 5614
5623 5615 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5624 5616 double start_par_time_sec = os::elapsedTime();
5625 5617 double end_par_time_sec;
5626 5618
5627 5619 {
5628 5620 StrongRootsScope srs(this);
5629 5621
5630 5622 if (G1CollectedHeap::use_parallel_gc_threads()) {
5631 5623 // The individual threads will set their evac-failure closures.
5632 5624 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5633 5625 // These tasks use ShareHeap::_process_strong_tasks
5634 5626 assert(UseDynamicNumberOfGCThreads ||
5635 5627 workers()->active_workers() == workers()->total_workers(),
5636 5628 "If not dynamic should be using all the workers");
5637 5629 workers()->run_task(&g1_par_task);
5638 5630 } else {
5639 5631 g1_par_task.set_for_termination(n_workers);
5640 5632 g1_par_task.work(0);
5641 5633 }
5642 5634 end_par_time_sec = os::elapsedTime();
5643 5635
5644 5636 // Closing the inner scope will execute the destructor
5645 5637 // for the StrongRootsScope object. We record the current
5646 5638 // elapsed time before closing the scope so that time
5647 5639 // taken for the SRS destructor is NOT included in the
5648 5640 // reported parallel time.
5649 5641 }
5650 5642
5651 5643 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5652 5644 g1_policy()->phase_times()->record_par_time(par_time_ms);
5653 5645
5654 5646 double code_root_fixup_time_ms =
5655 5647 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5656 5648 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5657 5649
5658 5650 set_par_threads(0);
5659 5651
5660 5652 // Process any discovered reference objects - we have
5661 5653 // to do this _before_ we retire the GC alloc regions
5662 5654 // as we may have to copy some 'reachable' referent
5663 5655 // objects (and their reachable sub-graphs) that were
5664 5656 // not copied during the pause.
5665 5657 process_discovered_references();
5666 5658
5667 5659 // Weak root processing.
5668 5660 // Note: when JSR 292 is enabled and code blobs can contain
5669 5661 // non-perm oops then we will need to process the code blobs
5670 5662 // here too.
5671 5663 {
5672 5664 G1STWIsAliveClosure is_alive(this);
5673 5665 G1KeepAliveClosure keep_alive(this);
5674 5666 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5675 5667 }
5676 5668
5677 5669 release_gc_alloc_regions();
5678 5670 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5679 5671
5680 5672 concurrent_g1_refine()->clear_hot_cache();
5681 5673 concurrent_g1_refine()->set_use_cache(true);
5682 5674
5683 5675 finalize_for_evac_failure();
5684 5676
5685 5677 if (evacuation_failed()) {
5686 5678 remove_self_forwarding_pointers();
5687 5679
5688 5680 // Reset the G1EvacuationFailureALot counters and flags
5689 5681 // Note: the values are reset only when an actual
5690 5682 // evacuation failure occurs.
5691 5683 NOT_PRODUCT(reset_evacuation_should_fail();)
5692 5684 }
5693 5685
5694 5686 // Enqueue any remaining references remaining on the STW
5695 5687 // reference processor's discovered lists. We need to do
5696 5688 // this after the card table is cleaned (and verified) as
5697 5689 // the act of enqueuing entries on to the pending list
5698 5690 // will log these updates (and dirty their associated
5699 5691 // cards). We need these updates logged to update any
5700 5692 // RSets.
5701 5693 enqueue_discovered_references();
5702 5694
5703 5695 if (G1DeferredRSUpdate) {
5704 5696 RedirtyLoggedCardTableEntryFastClosure redirty;
5705 5697 dirty_card_queue_set().set_closure(&redirty);
5706 5698 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5707 5699
5708 5700 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5709 5701 dcq.merge_bufferlists(&dirty_card_queue_set());
5710 5702 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5711 5703 }
5712 5704 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5713 5705 }
5714 5706
5715 5707 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5716 5708 size_t* pre_used,
5717 5709 FreeRegionList* free_list,
5718 5710 OldRegionSet* old_proxy_set,
5719 5711 HumongousRegionSet* humongous_proxy_set,
5720 5712 HRRSCleanupTask* hrrs_cleanup_task,
5721 5713 bool par) {
5722 5714 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5723 5715 if (hr->isHumongous()) {
5724 5716 assert(hr->startsHumongous(), "we should only see starts humongous");
5725 5717 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5726 5718 } else {
5727 5719 _old_set.remove_with_proxy(hr, old_proxy_set);
5728 5720 free_region(hr, pre_used, free_list, par);
5729 5721 }
5730 5722 } else {
5731 5723 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5732 5724 }
5733 5725 }
5734 5726
5735 5727 void G1CollectedHeap::free_region(HeapRegion* hr,
5736 5728 size_t* pre_used,
5737 5729 FreeRegionList* free_list,
5738 5730 bool par) {
5739 5731 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5740 5732 assert(!hr->is_empty(), "the region should not be empty");
5741 5733 assert(free_list != NULL, "pre-condition");
5742 5734
5743 5735 *pre_used += hr->used();
5744 5736 hr->hr_clear(par, true /* clear_space */);
5745 5737 free_list->add_as_head(hr);
5746 5738 }
5747 5739
5748 5740 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5749 5741 size_t* pre_used,
5750 5742 FreeRegionList* free_list,
5751 5743 HumongousRegionSet* humongous_proxy_set,
5752 5744 bool par) {
5753 5745 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5754 5746 assert(free_list != NULL, "pre-condition");
5755 5747 assert(humongous_proxy_set != NULL, "pre-condition");
5756 5748
5757 5749 size_t hr_used = hr->used();
5758 5750 size_t hr_capacity = hr->capacity();
5759 5751 size_t hr_pre_used = 0;
5760 5752 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5761 5753 // We need to read this before we make the region non-humongous,
5762 5754 // otherwise the information will be gone.
5763 5755 uint last_index = hr->last_hc_index();
5764 5756 hr->set_notHumongous();
5765 5757 free_region(hr, &hr_pre_used, free_list, par);
5766 5758
5767 5759 uint i = hr->hrs_index() + 1;
5768 5760 while (i < last_index) {
5769 5761 HeapRegion* curr_hr = region_at(i);
5770 5762 assert(curr_hr->continuesHumongous(), "invariant");
5771 5763 curr_hr->set_notHumongous();
5772 5764 free_region(curr_hr, &hr_pre_used, free_list, par);
5773 5765 i += 1;
5774 5766 }
5775 5767 assert(hr_pre_used == hr_used,
5776 5768 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5777 5769 "should be the same", hr_pre_used, hr_used));
5778 5770 *pre_used += hr_pre_used;
5779 5771 }
5780 5772
5781 5773 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5782 5774 FreeRegionList* free_list,
5783 5775 OldRegionSet* old_proxy_set,
5784 5776 HumongousRegionSet* humongous_proxy_set,
5785 5777 bool par) {
5786 5778 if (pre_used > 0) {
5787 5779 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5788 5780 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5789 5781 assert(_summary_bytes_used >= pre_used,
5790 5782 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5791 5783 "should be >= pre_used: "SIZE_FORMAT,
5792 5784 _summary_bytes_used, pre_used));
5793 5785 _summary_bytes_used -= pre_used;
5794 5786 }
5795 5787 if (free_list != NULL && !free_list->is_empty()) {
5796 5788 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5797 5789 _free_list.add_as_head(free_list);
5798 5790 }
5799 5791 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5800 5792 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5801 5793 _old_set.update_from_proxy(old_proxy_set);
5802 5794 }
5803 5795 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5804 5796 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5805 5797 _humongous_set.update_from_proxy(humongous_proxy_set);
5806 5798 }
5807 5799 }
5808 5800
5809 5801 class G1ParCleanupCTTask : public AbstractGangTask {
5810 5802 CardTableModRefBS* _ct_bs;
5811 5803 G1CollectedHeap* _g1h;
5812 5804 HeapRegion* volatile _su_head;
5813 5805 public:
5814 5806 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5815 5807 G1CollectedHeap* g1h) :
5816 5808 AbstractGangTask("G1 Par Cleanup CT Task"),
5817 5809 _ct_bs(ct_bs), _g1h(g1h) { }
5818 5810
5819 5811 void work(uint worker_id) {
5820 5812 HeapRegion* r;
5821 5813 while (r = _g1h->pop_dirty_cards_region()) {
5822 5814 clear_cards(r);
5823 5815 }
5824 5816 }
5825 5817
5826 5818 void clear_cards(HeapRegion* r) {
5827 5819 // Cards of the survivors should have already been dirtied.
5828 5820 if (!r->is_survivor()) {
5829 5821 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5830 5822 }
5831 5823 }
5832 5824 };
5833 5825
5834 5826 #ifndef PRODUCT
5835 5827 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5836 5828 G1CollectedHeap* _g1h;
5837 5829 CardTableModRefBS* _ct_bs;
5838 5830 public:
5839 5831 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5840 5832 : _g1h(g1h), _ct_bs(ct_bs) { }
5841 5833 virtual bool doHeapRegion(HeapRegion* r) {
5842 5834 if (r->is_survivor()) {
5843 5835 _g1h->verify_dirty_region(r);
5844 5836 } else {
5845 5837 _g1h->verify_not_dirty_region(r);
5846 5838 }
5847 5839 return false;
5848 5840 }
5849 5841 };
5850 5842
5851 5843 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5852 5844 // All of the region should be clean.
5853 5845 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5854 5846 MemRegion mr(hr->bottom(), hr->end());
5855 5847 ct_bs->verify_not_dirty_region(mr);
5856 5848 }
5857 5849
5858 5850 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5859 5851 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5860 5852 // dirty allocated blocks as they allocate them. The thread that
5861 5853 // retires each region and replaces it with a new one will do a
5862 5854 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5863 5855 // not dirty that area (one less thing to have to do while holding
5864 5856 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5865 5857 // is dirty.
5866 5858 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5867 5859 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5868 5860 ct_bs->verify_dirty_region(mr);
5869 5861 }
5870 5862
5871 5863 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5872 5864 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5873 5865 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5874 5866 verify_dirty_region(hr);
5875 5867 }
5876 5868 }
5877 5869
5878 5870 void G1CollectedHeap::verify_dirty_young_regions() {
5879 5871 verify_dirty_young_list(_young_list->first_region());
5880 5872 }
5881 5873 #endif
5882 5874
5883 5875 void G1CollectedHeap::cleanUpCardTable() {
5884 5876 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5885 5877 double start = os::elapsedTime();
5886 5878
5887 5879 {
5888 5880 // Iterate over the dirty cards region list.
5889 5881 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5890 5882
5891 5883 if (G1CollectedHeap::use_parallel_gc_threads()) {
5892 5884 set_par_threads();
5893 5885 workers()->run_task(&cleanup_task);
5894 5886 set_par_threads(0);
5895 5887 } else {
5896 5888 while (_dirty_cards_region_list) {
5897 5889 HeapRegion* r = _dirty_cards_region_list;
5898 5890 cleanup_task.clear_cards(r);
5899 5891 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5900 5892 if (_dirty_cards_region_list == r) {
5901 5893 // The last region.
5902 5894 _dirty_cards_region_list = NULL;
5903 5895 }
5904 5896 r->set_next_dirty_cards_region(NULL);
5905 5897 }
5906 5898 }
5907 5899 #ifndef PRODUCT
5908 5900 if (G1VerifyCTCleanup || VerifyAfterGC) {
5909 5901 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5910 5902 heap_region_iterate(&cleanup_verifier);
5911 5903 }
5912 5904 #endif
5913 5905 }
5914 5906
5915 5907 double elapsed = os::elapsedTime() - start;
5916 5908 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5917 5909 }
5918 5910
5919 5911 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5920 5912 size_t pre_used = 0;
5921 5913 FreeRegionList local_free_list("Local List for CSet Freeing");
5922 5914
5923 5915 double young_time_ms = 0.0;
5924 5916 double non_young_time_ms = 0.0;
5925 5917
5926 5918 // Since the collection set is a superset of the the young list,
5927 5919 // all we need to do to clear the young list is clear its
5928 5920 // head and length, and unlink any young regions in the code below
5929 5921 _young_list->clear();
5930 5922
5931 5923 G1CollectorPolicy* policy = g1_policy();
5932 5924
5933 5925 double start_sec = os::elapsedTime();
5934 5926 bool non_young = true;
5935 5927
5936 5928 HeapRegion* cur = cs_head;
5937 5929 int age_bound = -1;
5938 5930 size_t rs_lengths = 0;
5939 5931
5940 5932 while (cur != NULL) {
5941 5933 assert(!is_on_master_free_list(cur), "sanity");
5942 5934 if (non_young) {
5943 5935 if (cur->is_young()) {
5944 5936 double end_sec = os::elapsedTime();
5945 5937 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5946 5938 non_young_time_ms += elapsed_ms;
5947 5939
5948 5940 start_sec = os::elapsedTime();
5949 5941 non_young = false;
5950 5942 }
5951 5943 } else {
5952 5944 if (!cur->is_young()) {
5953 5945 double end_sec = os::elapsedTime();
5954 5946 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5955 5947 young_time_ms += elapsed_ms;
5956 5948
5957 5949 start_sec = os::elapsedTime();
5958 5950 non_young = true;
5959 5951 }
5960 5952 }
5961 5953
5962 5954 rs_lengths += cur->rem_set()->occupied();
5963 5955
5964 5956 HeapRegion* next = cur->next_in_collection_set();
5965 5957 assert(cur->in_collection_set(), "bad CS");
5966 5958 cur->set_next_in_collection_set(NULL);
5967 5959 cur->set_in_collection_set(false);
5968 5960
5969 5961 if (cur->is_young()) {
5970 5962 int index = cur->young_index_in_cset();
5971 5963 assert(index != -1, "invariant");
5972 5964 assert((uint) index < policy->young_cset_region_length(), "invariant");
5973 5965 size_t words_survived = _surviving_young_words[index];
5974 5966 cur->record_surv_words_in_group(words_survived);
5975 5967
5976 5968 // At this point the we have 'popped' cur from the collection set
5977 5969 // (linked via next_in_collection_set()) but it is still in the
5978 5970 // young list (linked via next_young_region()). Clear the
5979 5971 // _next_young_region field.
5980 5972 cur->set_next_young_region(NULL);
5981 5973 } else {
5982 5974 int index = cur->young_index_in_cset();
5983 5975 assert(index == -1, "invariant");
5984 5976 }
5985 5977
5986 5978 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5987 5979 (!cur->is_young() && cur->young_index_in_cset() == -1),
5988 5980 "invariant" );
5989 5981
5990 5982 if (!cur->evacuation_failed()) {
5991 5983 MemRegion used_mr = cur->used_region();
5992 5984
5993 5985 // And the region is empty.
5994 5986 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5995 5987 free_region(cur, &pre_used, &local_free_list, false /* par */);
5996 5988 } else {
5997 5989 cur->uninstall_surv_rate_group();
5998 5990 if (cur->is_young()) {
5999 5991 cur->set_young_index_in_cset(-1);
6000 5992 }
6001 5993 cur->set_not_young();
6002 5994 cur->set_evacuation_failed(false);
6003 5995 // The region is now considered to be old.
6004 5996 _old_set.add(cur);
6005 5997 }
6006 5998 cur = next;
6007 5999 }
6008 6000
6009 6001 policy->record_max_rs_lengths(rs_lengths);
6010 6002 policy->cset_regions_freed();
6011 6003
6012 6004 double end_sec = os::elapsedTime();
6013 6005 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6014 6006
6015 6007 if (non_young) {
6016 6008 non_young_time_ms += elapsed_ms;
6017 6009 } else {
6018 6010 young_time_ms += elapsed_ms;
6019 6011 }
6020 6012
6021 6013 update_sets_after_freeing_regions(pre_used, &local_free_list,
6022 6014 NULL /* old_proxy_set */,
6023 6015 NULL /* humongous_proxy_set */,
6024 6016 false /* par */);
6025 6017 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6026 6018 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6027 6019 }
6028 6020
6029 6021 // This routine is similar to the above but does not record
6030 6022 // any policy statistics or update free lists; we are abandoning
6031 6023 // the current incremental collection set in preparation of a
6032 6024 // full collection. After the full GC we will start to build up
6033 6025 // the incremental collection set again.
6034 6026 // This is only called when we're doing a full collection
6035 6027 // and is immediately followed by the tearing down of the young list.
6036 6028
6037 6029 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6038 6030 HeapRegion* cur = cs_head;
6039 6031
6040 6032 while (cur != NULL) {
6041 6033 HeapRegion* next = cur->next_in_collection_set();
6042 6034 assert(cur->in_collection_set(), "bad CS");
6043 6035 cur->set_next_in_collection_set(NULL);
6044 6036 cur->set_in_collection_set(false);
6045 6037 cur->set_young_index_in_cset(-1);
6046 6038 cur = next;
6047 6039 }
6048 6040 }
6049 6041
6050 6042 void G1CollectedHeap::set_free_regions_coming() {
6051 6043 if (G1ConcRegionFreeingVerbose) {
6052 6044 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6053 6045 "setting free regions coming");
6054 6046 }
6055 6047
6056 6048 assert(!free_regions_coming(), "pre-condition");
6057 6049 _free_regions_coming = true;
6058 6050 }
6059 6051
6060 6052 void G1CollectedHeap::reset_free_regions_coming() {
6061 6053 assert(free_regions_coming(), "pre-condition");
6062 6054
6063 6055 {
6064 6056 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6065 6057 _free_regions_coming = false;
6066 6058 SecondaryFreeList_lock->notify_all();
6067 6059 }
6068 6060
6069 6061 if (G1ConcRegionFreeingVerbose) {
6070 6062 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6071 6063 "reset free regions coming");
6072 6064 }
6073 6065 }
6074 6066
6075 6067 void G1CollectedHeap::wait_while_free_regions_coming() {
6076 6068 // Most of the time we won't have to wait, so let's do a quick test
6077 6069 // first before we take the lock.
6078 6070 if (!free_regions_coming()) {
6079 6071 return;
6080 6072 }
6081 6073
6082 6074 if (G1ConcRegionFreeingVerbose) {
6083 6075 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6084 6076 "waiting for free regions");
6085 6077 }
6086 6078
6087 6079 {
6088 6080 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6089 6081 while (free_regions_coming()) {
6090 6082 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6091 6083 }
6092 6084 }
6093 6085
6094 6086 if (G1ConcRegionFreeingVerbose) {
6095 6087 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6096 6088 "done waiting for free regions");
6097 6089 }
6098 6090 }
6099 6091
6100 6092 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6101 6093 assert(heap_lock_held_for_gc(),
6102 6094 "the heap lock should already be held by or for this thread");
6103 6095 _young_list->push_region(hr);
6104 6096 }
6105 6097
6106 6098 class NoYoungRegionsClosure: public HeapRegionClosure {
6107 6099 private:
6108 6100 bool _success;
6109 6101 public:
6110 6102 NoYoungRegionsClosure() : _success(true) { }
6111 6103 bool doHeapRegion(HeapRegion* r) {
6112 6104 if (r->is_young()) {
6113 6105 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6114 6106 r->bottom(), r->end());
6115 6107 _success = false;
6116 6108 }
6117 6109 return false;
6118 6110 }
6119 6111 bool success() { return _success; }
6120 6112 };
6121 6113
6122 6114 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6123 6115 bool ret = _young_list->check_list_empty(check_sample);
6124 6116
6125 6117 if (check_heap) {
6126 6118 NoYoungRegionsClosure closure;
6127 6119 heap_region_iterate(&closure);
6128 6120 ret = ret && closure.success();
6129 6121 }
6130 6122
6131 6123 return ret;
6132 6124 }
6133 6125
6134 6126 class TearDownRegionSetsClosure : public HeapRegionClosure {
6135 6127 private:
6136 6128 OldRegionSet *_old_set;
6137 6129
6138 6130 public:
6139 6131 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6140 6132
6141 6133 bool doHeapRegion(HeapRegion* r) {
6142 6134 if (r->is_empty()) {
6143 6135 // We ignore empty regions, we'll empty the free list afterwards
6144 6136 } else if (r->is_young()) {
6145 6137 // We ignore young regions, we'll empty the young list afterwards
6146 6138 } else if (r->isHumongous()) {
6147 6139 // We ignore humongous regions, we're not tearing down the
6148 6140 // humongous region set
6149 6141 } else {
6150 6142 // The rest should be old
6151 6143 _old_set->remove(r);
6152 6144 }
6153 6145 return false;
6154 6146 }
6155 6147
6156 6148 ~TearDownRegionSetsClosure() {
6157 6149 assert(_old_set->is_empty(), "post-condition");
6158 6150 }
6159 6151 };
6160 6152
6161 6153 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6162 6154 assert_at_safepoint(true /* should_be_vm_thread */);
6163 6155
6164 6156 if (!free_list_only) {
6165 6157 TearDownRegionSetsClosure cl(&_old_set);
6166 6158 heap_region_iterate(&cl);
6167 6159
6168 6160 // Need to do this after the heap iteration to be able to
6169 6161 // recognize the young regions and ignore them during the iteration.
6170 6162 _young_list->empty_list();
6171 6163 }
6172 6164 _free_list.remove_all();
6173 6165 }
6174 6166
6175 6167 class RebuildRegionSetsClosure : public HeapRegionClosure {
6176 6168 private:
6177 6169 bool _free_list_only;
6178 6170 OldRegionSet* _old_set;
6179 6171 FreeRegionList* _free_list;
6180 6172 size_t _total_used;
6181 6173
6182 6174 public:
6183 6175 RebuildRegionSetsClosure(bool free_list_only,
6184 6176 OldRegionSet* old_set, FreeRegionList* free_list) :
6185 6177 _free_list_only(free_list_only),
6186 6178 _old_set(old_set), _free_list(free_list), _total_used(0) {
6187 6179 assert(_free_list->is_empty(), "pre-condition");
6188 6180 if (!free_list_only) {
6189 6181 assert(_old_set->is_empty(), "pre-condition");
6190 6182 }
6191 6183 }
6192 6184
6193 6185 bool doHeapRegion(HeapRegion* r) {
6194 6186 if (r->continuesHumongous()) {
6195 6187 return false;
6196 6188 }
6197 6189
6198 6190 if (r->is_empty()) {
6199 6191 // Add free regions to the free list
6200 6192 _free_list->add_as_tail(r);
6201 6193 } else if (!_free_list_only) {
6202 6194 assert(!r->is_young(), "we should not come across young regions");
6203 6195
6204 6196 if (r->isHumongous()) {
6205 6197 // We ignore humongous regions, we left the humongous set unchanged
6206 6198 } else {
6207 6199 // The rest should be old, add them to the old set
6208 6200 _old_set->add(r);
6209 6201 }
6210 6202 _total_used += r->used();
6211 6203 }
6212 6204
6213 6205 return false;
6214 6206 }
6215 6207
6216 6208 size_t total_used() {
6217 6209 return _total_used;
6218 6210 }
6219 6211 };
6220 6212
6221 6213 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6222 6214 assert_at_safepoint(true /* should_be_vm_thread */);
6223 6215
6224 6216 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6225 6217 heap_region_iterate(&cl);
6226 6218
6227 6219 if (!free_list_only) {
6228 6220 _summary_bytes_used = cl.total_used();
6229 6221 }
6230 6222 assert(_summary_bytes_used == recalculate_used(),
6231 6223 err_msg("inconsistent _summary_bytes_used, "
6232 6224 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6233 6225 _summary_bytes_used, recalculate_used()));
6234 6226 }
6235 6227
6236 6228 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6237 6229 _refine_cte_cl->set_concurrent(concurrent);
6238 6230 }
6239 6231
6240 6232 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6241 6233 HeapRegion* hr = heap_region_containing(p);
6242 6234 if (hr == NULL) {
6243 6235 return false;
6244 6236 } else {
6245 6237 return hr->is_in(p);
6246 6238 }
6247 6239 }
6248 6240
6249 6241 // Methods for the mutator alloc region
6250 6242
6251 6243 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6252 6244 bool force) {
6253 6245 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6254 6246 assert(!force || g1_policy()->can_expand_young_list(),
6255 6247 "if force is true we should be able to expand the young list");
6256 6248 bool young_list_full = g1_policy()->is_young_list_full();
6257 6249 if (force || !young_list_full) {
6258 6250 HeapRegion* new_alloc_region = new_region(word_size,
6259 6251 false /* do_expand */);
6260 6252 if (new_alloc_region != NULL) {
6261 6253 set_region_short_lived_locked(new_alloc_region);
6262 6254 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6263 6255 return new_alloc_region;
6264 6256 }
6265 6257 }
6266 6258 return NULL;
6267 6259 }
6268 6260
6269 6261 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6270 6262 size_t allocated_bytes) {
6271 6263 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6272 6264 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6273 6265
6274 6266 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6275 6267 _summary_bytes_used += allocated_bytes;
6276 6268 _hr_printer.retire(alloc_region);
6277 6269 // We update the eden sizes here, when the region is retired,
6278 6270 // instead of when it's allocated, since this is the point that its
6279 6271 // used space has been recored in _summary_bytes_used.
6280 6272 g1mm()->update_eden_size();
6281 6273 }
6282 6274
6283 6275 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6284 6276 bool force) {
6285 6277 return _g1h->new_mutator_alloc_region(word_size, force);
6286 6278 }
6287 6279
6288 6280 void G1CollectedHeap::set_par_threads() {
6289 6281 // Don't change the number of workers. Use the value previously set
6290 6282 // in the workgroup.
6291 6283 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6292 6284 uint n_workers = workers()->active_workers();
6293 6285 assert(UseDynamicNumberOfGCThreads ||
6294 6286 n_workers == workers()->total_workers(),
6295 6287 "Otherwise should be using the total number of workers");
6296 6288 if (n_workers == 0) {
6297 6289 assert(false, "Should have been set in prior evacuation pause.");
6298 6290 n_workers = ParallelGCThreads;
6299 6291 workers()->set_active_workers(n_workers);
6300 6292 }
6301 6293 set_par_threads(n_workers);
6302 6294 }
6303 6295
6304 6296 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6305 6297 size_t allocated_bytes) {
6306 6298 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6307 6299 }
6308 6300
6309 6301 // Methods for the GC alloc regions
6310 6302
6311 6303 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6312 6304 uint count,
6313 6305 GCAllocPurpose ap) {
6314 6306 assert(FreeList_lock->owned_by_self(), "pre-condition");
6315 6307
6316 6308 if (count < g1_policy()->max_regions(ap)) {
6317 6309 HeapRegion* new_alloc_region = new_region(word_size,
6318 6310 true /* do_expand */);
6319 6311 if (new_alloc_region != NULL) {
6320 6312 // We really only need to do this for old regions given that we
6321 6313 // should never scan survivors. But it doesn't hurt to do it
6322 6314 // for survivors too.
6323 6315 new_alloc_region->set_saved_mark();
6324 6316 if (ap == GCAllocForSurvived) {
6325 6317 new_alloc_region->set_survivor();
6326 6318 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6327 6319 } else {
6328 6320 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6329 6321 }
6330 6322 bool during_im = g1_policy()->during_initial_mark_pause();
6331 6323 new_alloc_region->note_start_of_copying(during_im);
6332 6324 return new_alloc_region;
6333 6325 } else {
6334 6326 g1_policy()->note_alloc_region_limit_reached(ap);
6335 6327 }
6336 6328 }
6337 6329 return NULL;
6338 6330 }
6339 6331
6340 6332 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6341 6333 size_t allocated_bytes,
6342 6334 GCAllocPurpose ap) {
6343 6335 bool during_im = g1_policy()->during_initial_mark_pause();
6344 6336 alloc_region->note_end_of_copying(during_im);
6345 6337 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6346 6338 if (ap == GCAllocForSurvived) {
6347 6339 young_list()->add_survivor_region(alloc_region);
6348 6340 } else {
6349 6341 _old_set.add(alloc_region);
6350 6342 }
6351 6343 _hr_printer.retire(alloc_region);
6352 6344 }
6353 6345
6354 6346 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6355 6347 bool force) {
6356 6348 assert(!force, "not supported for GC alloc regions");
6357 6349 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6358 6350 }
6359 6351
6360 6352 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6361 6353 size_t allocated_bytes) {
6362 6354 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6363 6355 GCAllocForSurvived);
6364 6356 }
6365 6357
6366 6358 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6367 6359 bool force) {
6368 6360 assert(!force, "not supported for GC alloc regions");
6369 6361 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6370 6362 }
6371 6363
6372 6364 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6373 6365 size_t allocated_bytes) {
6374 6366 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6375 6367 GCAllocForTenured);
6376 6368 }
6377 6369 // Heap region set verification
6378 6370
6379 6371 class VerifyRegionListsClosure : public HeapRegionClosure {
6380 6372 private:
6381 6373 FreeRegionList* _free_list;
6382 6374 OldRegionSet* _old_set;
6383 6375 HumongousRegionSet* _humongous_set;
6384 6376 uint _region_count;
6385 6377
6386 6378 public:
6387 6379 VerifyRegionListsClosure(OldRegionSet* old_set,
6388 6380 HumongousRegionSet* humongous_set,
6389 6381 FreeRegionList* free_list) :
6390 6382 _old_set(old_set), _humongous_set(humongous_set),
6391 6383 _free_list(free_list), _region_count(0) { }
6392 6384
6393 6385 uint region_count() { return _region_count; }
6394 6386
6395 6387 bool doHeapRegion(HeapRegion* hr) {
6396 6388 _region_count += 1;
6397 6389
6398 6390 if (hr->continuesHumongous()) {
6399 6391 return false;
6400 6392 }
6401 6393
6402 6394 if (hr->is_young()) {
6403 6395 // TODO
6404 6396 } else if (hr->startsHumongous()) {
6405 6397 _humongous_set->verify_next_region(hr);
6406 6398 } else if (hr->is_empty()) {
6407 6399 _free_list->verify_next_region(hr);
6408 6400 } else {
6409 6401 _old_set->verify_next_region(hr);
6410 6402 }
6411 6403 return false;
6412 6404 }
6413 6405 };
6414 6406
6415 6407 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6416 6408 HeapWord* bottom) {
6417 6409 HeapWord* end = bottom + HeapRegion::GrainWords;
6418 6410 MemRegion mr(bottom, end);
6419 6411 assert(_g1_reserved.contains(mr), "invariant");
6420 6412 // This might return NULL if the allocation fails
6421 6413 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
6422 6414 }
6423 6415
6424 6416 void G1CollectedHeap::verify_region_sets() {
6425 6417 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6426 6418
6427 6419 // First, check the explicit lists.
6428 6420 _free_list.verify();
6429 6421 {
6430 6422 // Given that a concurrent operation might be adding regions to
6431 6423 // the secondary free list we have to take the lock before
6432 6424 // verifying it.
6433 6425 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6434 6426 _secondary_free_list.verify();
6435 6427 }
6436 6428 _old_set.verify();
6437 6429 _humongous_set.verify();
6438 6430
6439 6431 // If a concurrent region freeing operation is in progress it will
6440 6432 // be difficult to correctly attributed any free regions we come
6441 6433 // across to the correct free list given that they might belong to
6442 6434 // one of several (free_list, secondary_free_list, any local lists,
6443 6435 // etc.). So, if that's the case we will skip the rest of the
6444 6436 // verification operation. Alternatively, waiting for the concurrent
6445 6437 // operation to complete will have a non-trivial effect on the GC's
6446 6438 // operation (no concurrent operation will last longer than the
6447 6439 // interval between two calls to verification) and it might hide
6448 6440 // any issues that we would like to catch during testing.
6449 6441 if (free_regions_coming()) {
6450 6442 return;
6451 6443 }
6452 6444
6453 6445 // Make sure we append the secondary_free_list on the free_list so
6454 6446 // that all free regions we will come across can be safely
6455 6447 // attributed to the free_list.
6456 6448 append_secondary_free_list_if_not_empty_with_lock();
6457 6449
6458 6450 // Finally, make sure that the region accounting in the lists is
6459 6451 // consistent with what we see in the heap.
6460 6452 _old_set.verify_start();
6461 6453 _humongous_set.verify_start();
6462 6454 _free_list.verify_start();
6463 6455
6464 6456 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6465 6457 heap_region_iterate(&cl);
6466 6458
6467 6459 _old_set.verify_end();
6468 6460 _humongous_set.verify_end();
6469 6461 _free_list.verify_end();
6470 6462 }
|
↓ open down ↓ |
4496 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX