Print this page
rev 3708 : 8000244: G1: Ergonomically set MarkStackSize and use virtual space for global marking stack
Summary: Set the value of MarkStackSize to a value based on the number of parallel marking threads with a reasonable minimum. Expand the marking stack if we have to restart marking due to an overflow up to a reasonable maximum. Allocate the underlying space for the marking stack from virtual memory.
Reviewed-by: jmasa
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();
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 1957 }
1958 1958
1959 1959 jint G1CollectedHeap::initialize() {
1960 1960 CollectedHeap::pre_initialize();
1961 1961 os::enable_vtime();
1962 1962
1963 1963 G1Log::init();
1964 1964
1965 1965 // Necessary to satisfy locking discipline assertions.
1966 1966
1967 1967 MutexLocker x(Heap_lock);
1968 1968
1969 1969 // We have to initialize the printer before committing the heap, as
1970 1970 // it will be used then.
1971 1971 _hr_printer.set_active(G1PrintHeapRegions);
1972 1972
1973 1973 // While there are no constraints in the GC code that HeapWordSize
1974 1974 // be any particular value, there are multiple other areas in the
1975 1975 // system which believe this to be true (e.g. oop->object_size in some
1976 1976 // cases incorrectly returns the size in wordSize units rather than
1977 1977 // HeapWordSize).
1978 1978 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1979 1979
1980 1980 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1981 1981 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1982 1982
1983 1983 // Ensure that the sizes are properly aligned.
1984 1984 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1985 1985 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1986 1986
1987 1987 _cg1r = new ConcurrentG1Refine();
1988 1988
1989 1989 // Reserve the maximum.
1990 1990
1991 1991 // When compressed oops are enabled, the preferred heap base
1992 1992 // is calculated by subtracting the requested size from the
1993 1993 // 32Gb boundary and using the result as the base address for
1994 1994 // heap reservation. If the requested size is not aligned to
1995 1995 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1996 1996 // into the ReservedHeapSpace constructor) then the actual
1997 1997 // base of the reserved heap may end up differing from the
1998 1998 // address that was requested (i.e. the preferred heap base).
1999 1999 // If this happens then we could end up using a non-optimal
2000 2000 // compressed oops mode.
2001 2001
2002 2002 // Since max_byte_size is aligned to the size of a heap region (checked
2003 2003 // above).
2004 2004 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2005 2005
2006 2006 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2007 2007 HeapRegion::GrainBytes);
2008 2008
2009 2009 // It is important to do this in a way such that concurrent readers can't
2010 2010 // temporarily think somethings in the heap. (I've actually seen this
2011 2011 // happen in asserts: DLD.)
2012 2012 _reserved.set_word_size(0);
2013 2013 _reserved.set_start((HeapWord*)heap_rs.base());
2014 2014 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2015 2015
2016 2016 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2017 2017
2018 2018 // Create the gen rem set (and barrier set) for the entire reserved region.
2019 2019 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2020 2020 set_barrier_set(rem_set()->bs());
2021 2021 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2022 2022 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2023 2023 } else {
2024 2024 vm_exit_during_initialization("G1 requires a mod ref bs.");
2025 2025 return JNI_ENOMEM;
2026 2026 }
2027 2027
2028 2028 // Also create a G1 rem set.
2029 2029 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2030 2030 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2031 2031 } else {
2032 2032 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2033 2033 return JNI_ENOMEM;
2034 2034 }
2035 2035
2036 2036 // Carve out the G1 part of the heap.
2037 2037
2038 2038 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2039 2039 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2040 2040 g1_rs.size()/HeapWordSize);
2041 2041
2042 2042 _g1_storage.initialize(g1_rs, 0);
2043 2043 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2044 2044 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2045 2045 (HeapWord*) _g1_reserved.end(),
2046 2046 _expansion_regions);
2047 2047
2048 2048 // 6843694 - ensure that the maximum region index can fit
2049 2049 // in the remembered set structures.
2050 2050 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2051 2051 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2052 2052
2053 2053 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2054 2054 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2055 2055 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2056 2056 "too many cards per region");
2057 2057
2058 2058 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2059 2059
2060 2060 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2061 2061 heap_word_size(init_byte_size));
2062 2062
2063 2063 _g1h = this;
2064 2064
2065 2065 _in_cset_fast_test_length = max_regions();
2066 2066 _in_cset_fast_test_base =
2067 2067 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2068 2068
2069 2069 // We're biasing _in_cset_fast_test to avoid subtracting the
2070 2070 // beginning of the heap every time we want to index; basically
2071 2071 // it's the same with what we do with the card table.
↓ open down ↓ |
2071 lines elided |
↑ open up ↑ |
2072 2072 _in_cset_fast_test = _in_cset_fast_test_base -
2073 2073 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2074 2074
2075 2075 // Clear the _cset_fast_test bitmap in anticipation of adding
2076 2076 // regions to the incremental collection set for the first
2077 2077 // evacuation pause.
2078 2078 clear_cset_fast_test();
2079 2079
2080 2080 // Create the ConcurrentMark data structure and thread.
2081 2081 // (Must do this late, so that "max_regions" is defined.)
2082 - _cm = new ConcurrentMark(heap_rs, max_regions());
2082 + _cm = new ConcurrentMark(this, heap_rs);
2083 + if (_cm == NULL || !_cm->completed_initialization()) {
2084 + vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2085 + return JNI_ENOMEM;
2086 + }
2083 2087 _cmThread = _cm->cmThread();
2084 2088
2085 2089 // Initialize the from_card cache structure of HeapRegionRemSet.
2086 2090 HeapRegionRemSet::init_heap(max_regions());
2087 2091
2088 2092 // Now expand into the initial heap size.
2089 2093 if (!expand(init_byte_size)) {
2090 - vm_exit_during_initialization("Failed to allocate initial heap.");
2094 + vm_shutdown_during_initialization("Failed to allocate initial heap.");
2091 2095 return JNI_ENOMEM;
2092 2096 }
2093 2097
2094 2098 // Perform any initialization actions delegated to the policy.
2095 2099 g1_policy()->init();
2096 2100
2097 2101 _refine_cte_cl =
2098 2102 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2099 2103 g1_rem_set(),
2100 2104 concurrent_g1_refine());
2101 2105 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2102 2106
2103 2107 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2104 2108 SATB_Q_FL_lock,
2105 2109 G1SATBProcessCompletedThreshold,
2106 2110 Shared_SATB_Q_lock);
2107 2111
2108 2112 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2109 2113 DirtyCardQ_FL_lock,
2110 2114 concurrent_g1_refine()->yellow_zone(),
2111 2115 concurrent_g1_refine()->red_zone(),
2112 2116 Shared_DirtyCardQ_lock);
2113 2117
2114 2118 if (G1DeferredRSUpdate) {
2115 2119 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2116 2120 DirtyCardQ_FL_lock,
2117 2121 -1, // never trigger processing
2118 2122 -1, // no limit on length
2119 2123 Shared_DirtyCardQ_lock,
2120 2124 &JavaThread::dirty_card_queue_set());
2121 2125 }
2122 2126
2123 2127 // Initialize the card queue set used to hold cards containing
2124 2128 // references into the collection set.
2125 2129 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2126 2130 DirtyCardQ_FL_lock,
2127 2131 -1, // never trigger processing
2128 2132 -1, // no limit on length
2129 2133 Shared_DirtyCardQ_lock,
2130 2134 &JavaThread::dirty_card_queue_set());
2131 2135
2132 2136 // In case we're keeping closure specialization stats, initialize those
2133 2137 // counts and that mechanism.
2134 2138 SpecializationStats::clear();
2135 2139
2136 2140 // Do later initialization work for concurrent refinement.
2137 2141 _cg1r->init();
2138 2142
2139 2143 // Here we allocate the dummy full region that is required by the
2140 2144 // G1AllocRegion class. If we don't pass an address in the reserved
2141 2145 // space here, lots of asserts fire.
2142 2146
2143 2147 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2144 2148 _g1_reserved.start());
2145 2149 // We'll re-use the same region whether the alloc region will
2146 2150 // require BOT updates or not and, if it doesn't, then a non-young
2147 2151 // region will complain that it cannot support allocations without
2148 2152 // BOT updates. So we'll tag the dummy region as young to avoid that.
2149 2153 dummy_region->set_young();
2150 2154 // Make sure it's full.
2151 2155 dummy_region->set_top(dummy_region->end());
2152 2156 G1AllocRegion::setup(this, dummy_region);
2153 2157
2154 2158 init_mutator_alloc_region();
2155 2159
2156 2160 // Do create of the monitoring and management support so that
2157 2161 // values in the heap have been properly initialized.
2158 2162 _g1mm = new G1MonitoringSupport(this);
2159 2163
2160 2164 return JNI_OK;
2161 2165 }
2162 2166
2163 2167 void G1CollectedHeap::ref_processing_init() {
2164 2168 // Reference processing in G1 currently works as follows:
2165 2169 //
2166 2170 // * There are two reference processor instances. One is
2167 2171 // used to record and process discovered references
2168 2172 // during concurrent marking; the other is used to
2169 2173 // record and process references during STW pauses
2170 2174 // (both full and incremental).
2171 2175 // * Both ref processors need to 'span' the entire heap as
2172 2176 // the regions in the collection set may be dotted around.
2173 2177 //
2174 2178 // * For the concurrent marking ref processor:
2175 2179 // * Reference discovery is enabled at initial marking.
2176 2180 // * Reference discovery is disabled and the discovered
2177 2181 // references processed etc during remarking.
2178 2182 // * Reference discovery is MT (see below).
2179 2183 // * Reference discovery requires a barrier (see below).
2180 2184 // * Reference processing may or may not be MT
2181 2185 // (depending on the value of ParallelRefProcEnabled
2182 2186 // and ParallelGCThreads).
2183 2187 // * A full GC disables reference discovery by the CM
2184 2188 // ref processor and abandons any entries on it's
2185 2189 // discovered lists.
2186 2190 //
2187 2191 // * For the STW processor:
2188 2192 // * Non MT discovery is enabled at the start of a full GC.
2189 2193 // * Processing and enqueueing during a full GC is non-MT.
2190 2194 // * During a full GC, references are processed after marking.
2191 2195 //
2192 2196 // * Discovery (may or may not be MT) is enabled at the start
2193 2197 // of an incremental evacuation pause.
2194 2198 // * References are processed near the end of a STW evacuation pause.
2195 2199 // * For both types of GC:
2196 2200 // * Discovery is atomic - i.e. not concurrent.
2197 2201 // * Reference discovery will not need a barrier.
2198 2202
2199 2203 SharedHeap::ref_processing_init();
2200 2204 MemRegion mr = reserved_region();
2201 2205
2202 2206 // Concurrent Mark ref processor
2203 2207 _ref_processor_cm =
2204 2208 new ReferenceProcessor(mr, // span
2205 2209 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2206 2210 // mt processing
2207 2211 (int) ParallelGCThreads,
2208 2212 // degree of mt processing
2209 2213 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2210 2214 // mt discovery
2211 2215 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2212 2216 // degree of mt discovery
2213 2217 false,
2214 2218 // Reference discovery is not atomic
2215 2219 &_is_alive_closure_cm,
2216 2220 // is alive closure
2217 2221 // (for efficiency/performance)
2218 2222 true);
2219 2223 // Setting next fields of discovered
2220 2224 // lists requires a barrier.
2221 2225
2222 2226 // STW ref processor
2223 2227 _ref_processor_stw =
2224 2228 new ReferenceProcessor(mr, // span
2225 2229 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2226 2230 // mt processing
2227 2231 MAX2((int)ParallelGCThreads, 1),
2228 2232 // degree of mt processing
2229 2233 (ParallelGCThreads > 1),
2230 2234 // mt discovery
2231 2235 MAX2((int)ParallelGCThreads, 1),
2232 2236 // degree of mt discovery
2233 2237 true,
2234 2238 // Reference discovery is atomic
2235 2239 &_is_alive_closure_stw,
2236 2240 // is alive closure
2237 2241 // (for efficiency/performance)
2238 2242 false);
2239 2243 // Setting next fields of discovered
2240 2244 // lists requires a barrier.
2241 2245 }
2242 2246
2243 2247 size_t G1CollectedHeap::capacity() const {
2244 2248 return _g1_committed.byte_size();
2245 2249 }
2246 2250
2247 2251 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2248 2252 assert(!hr->continuesHumongous(), "pre-condition");
2249 2253 hr->reset_gc_time_stamp();
2250 2254 if (hr->startsHumongous()) {
2251 2255 uint first_index = hr->hrs_index() + 1;
2252 2256 uint last_index = hr->last_hc_index();
2253 2257 for (uint i = first_index; i < last_index; i += 1) {
2254 2258 HeapRegion* chr = region_at(i);
2255 2259 assert(chr->continuesHumongous(), "sanity");
2256 2260 chr->reset_gc_time_stamp();
2257 2261 }
2258 2262 }
2259 2263 }
2260 2264
2261 2265 #ifndef PRODUCT
2262 2266 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2263 2267 private:
2264 2268 unsigned _gc_time_stamp;
2265 2269 bool _failures;
2266 2270
2267 2271 public:
2268 2272 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2269 2273 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2270 2274
2271 2275 virtual bool doHeapRegion(HeapRegion* hr) {
2272 2276 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2273 2277 if (_gc_time_stamp != region_gc_time_stamp) {
2274 2278 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2275 2279 "expected %d", HR_FORMAT_PARAMS(hr),
2276 2280 region_gc_time_stamp, _gc_time_stamp);
2277 2281 _failures = true;
2278 2282 }
2279 2283 return false;
2280 2284 }
2281 2285
2282 2286 bool failures() { return _failures; }
2283 2287 };
2284 2288
2285 2289 void G1CollectedHeap::check_gc_time_stamps() {
2286 2290 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2287 2291 heap_region_iterate(&cl);
2288 2292 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2289 2293 }
2290 2294 #endif // PRODUCT
2291 2295
2292 2296 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2293 2297 DirtyCardQueue* into_cset_dcq,
2294 2298 bool concurrent,
2295 2299 int worker_i) {
2296 2300 // Clean cards in the hot card cache
2297 2301 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2298 2302
2299 2303 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2300 2304 int n_completed_buffers = 0;
2301 2305 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2302 2306 n_completed_buffers++;
2303 2307 }
2304 2308 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2305 2309 dcqs.clear_n_completed_buffers();
2306 2310 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2307 2311 }
2308 2312
2309 2313
2310 2314 // Computes the sum of the storage used by the various regions.
2311 2315
2312 2316 size_t G1CollectedHeap::used() const {
2313 2317 assert(Heap_lock->owner() != NULL,
2314 2318 "Should be owned on this thread's behalf.");
2315 2319 size_t result = _summary_bytes_used;
2316 2320 // Read only once in case it is set to NULL concurrently
2317 2321 HeapRegion* hr = _mutator_alloc_region.get();
2318 2322 if (hr != NULL)
2319 2323 result += hr->used();
2320 2324 return result;
2321 2325 }
2322 2326
2323 2327 size_t G1CollectedHeap::used_unlocked() const {
2324 2328 size_t result = _summary_bytes_used;
2325 2329 return result;
2326 2330 }
2327 2331
2328 2332 class SumUsedClosure: public HeapRegionClosure {
2329 2333 size_t _used;
2330 2334 public:
2331 2335 SumUsedClosure() : _used(0) {}
2332 2336 bool doHeapRegion(HeapRegion* r) {
2333 2337 if (!r->continuesHumongous()) {
2334 2338 _used += r->used();
2335 2339 }
2336 2340 return false;
2337 2341 }
2338 2342 size_t result() { return _used; }
2339 2343 };
2340 2344
2341 2345 size_t G1CollectedHeap::recalculate_used() const {
2342 2346 SumUsedClosure blk;
2343 2347 heap_region_iterate(&blk);
2344 2348 return blk.result();
2345 2349 }
2346 2350
2347 2351 size_t G1CollectedHeap::unsafe_max_alloc() {
2348 2352 if (free_regions() > 0) return HeapRegion::GrainBytes;
2349 2353 // otherwise, is there space in the current allocation region?
2350 2354
2351 2355 // We need to store the current allocation region in a local variable
2352 2356 // here. The problem is that this method doesn't take any locks and
2353 2357 // there may be other threads which overwrite the current allocation
2354 2358 // region field. attempt_allocation(), for example, sets it to NULL
2355 2359 // and this can happen *after* the NULL check here but before the call
2356 2360 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2357 2361 // to be a problem in the optimized build, since the two loads of the
2358 2362 // current allocation region field are optimized away.
2359 2363 HeapRegion* hr = _mutator_alloc_region.get();
2360 2364 if (hr == NULL) {
2361 2365 return 0;
2362 2366 }
2363 2367 return hr->free();
2364 2368 }
2365 2369
2366 2370 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2367 2371 switch (cause) {
2368 2372 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2369 2373 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2370 2374 case GCCause::_g1_humongous_allocation: return true;
2371 2375 default: return false;
2372 2376 }
2373 2377 }
2374 2378
2375 2379 #ifndef PRODUCT
2376 2380 void G1CollectedHeap::allocate_dummy_regions() {
2377 2381 // Let's fill up most of the region
2378 2382 size_t word_size = HeapRegion::GrainWords - 1024;
2379 2383 // And as a result the region we'll allocate will be humongous.
2380 2384 guarantee(isHumongous(word_size), "sanity");
2381 2385
2382 2386 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2383 2387 // Let's use the existing mechanism for the allocation
2384 2388 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2385 2389 if (dummy_obj != NULL) {
2386 2390 MemRegion mr(dummy_obj, word_size);
2387 2391 CollectedHeap::fill_with_object(mr);
2388 2392 } else {
2389 2393 // If we can't allocate once, we probably cannot allocate
2390 2394 // again. Let's get out of the loop.
2391 2395 break;
2392 2396 }
2393 2397 }
2394 2398 }
2395 2399 #endif // !PRODUCT
2396 2400
2397 2401 void G1CollectedHeap::increment_old_marking_cycles_started() {
2398 2402 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2399 2403 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2400 2404 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2401 2405 _old_marking_cycles_started, _old_marking_cycles_completed));
2402 2406
2403 2407 _old_marking_cycles_started++;
2404 2408 }
2405 2409
2406 2410 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2407 2411 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2408 2412
2409 2413 // We assume that if concurrent == true, then the caller is a
2410 2414 // concurrent thread that was joined the Suspendible Thread
2411 2415 // Set. If there's ever a cheap way to check this, we should add an
2412 2416 // assert here.
2413 2417
2414 2418 // Given that this method is called at the end of a Full GC or of a
2415 2419 // concurrent cycle, and those can be nested (i.e., a Full GC can
2416 2420 // interrupt a concurrent cycle), the number of full collections
2417 2421 // completed should be either one (in the case where there was no
2418 2422 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2419 2423 // behind the number of full collections started.
2420 2424
2421 2425 // This is the case for the inner caller, i.e. a Full GC.
2422 2426 assert(concurrent ||
2423 2427 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2424 2428 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2425 2429 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2426 2430 "is inconsistent with _old_marking_cycles_completed = %u",
2427 2431 _old_marking_cycles_started, _old_marking_cycles_completed));
2428 2432
2429 2433 // This is the case for the outer caller, i.e. the concurrent cycle.
2430 2434 assert(!concurrent ||
2431 2435 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2432 2436 err_msg("for outer caller (concurrent cycle): "
2433 2437 "_old_marking_cycles_started = %u "
2434 2438 "is inconsistent with _old_marking_cycles_completed = %u",
2435 2439 _old_marking_cycles_started, _old_marking_cycles_completed));
2436 2440
2437 2441 _old_marking_cycles_completed += 1;
2438 2442
2439 2443 // We need to clear the "in_progress" flag in the CM thread before
2440 2444 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2441 2445 // is set) so that if a waiter requests another System.gc() it doesn't
2442 2446 // incorrectly see that a marking cyle is still in progress.
2443 2447 if (concurrent) {
2444 2448 _cmThread->clear_in_progress();
2445 2449 }
2446 2450
2447 2451 // This notify_all() will ensure that a thread that called
2448 2452 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2449 2453 // and it's waiting for a full GC to finish will be woken up. It is
2450 2454 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2451 2455 FullGCCount_lock->notify_all();
2452 2456 }
2453 2457
2454 2458 void G1CollectedHeap::collect(GCCause::Cause cause) {
2455 2459 assert_heap_not_locked();
2456 2460
2457 2461 unsigned int gc_count_before;
2458 2462 unsigned int old_marking_count_before;
2459 2463 bool retry_gc;
2460 2464
2461 2465 do {
2462 2466 retry_gc = false;
2463 2467
2464 2468 {
2465 2469 MutexLocker ml(Heap_lock);
2466 2470
2467 2471 // Read the GC count while holding the Heap_lock
2468 2472 gc_count_before = total_collections();
2469 2473 old_marking_count_before = _old_marking_cycles_started;
2470 2474 }
2471 2475
2472 2476 if (should_do_concurrent_full_gc(cause)) {
2473 2477 // Schedule an initial-mark evacuation pause that will start a
2474 2478 // concurrent cycle. We're setting word_size to 0 which means that
2475 2479 // we are not requesting a post-GC allocation.
2476 2480 VM_G1IncCollectionPause op(gc_count_before,
2477 2481 0, /* word_size */
2478 2482 true, /* should_initiate_conc_mark */
2479 2483 g1_policy()->max_pause_time_ms(),
2480 2484 cause);
2481 2485
2482 2486 VMThread::execute(&op);
2483 2487 if (!op.pause_succeeded()) {
2484 2488 if (old_marking_count_before == _old_marking_cycles_started) {
2485 2489 retry_gc = op.should_retry_gc();
2486 2490 } else {
2487 2491 // A Full GC happened while we were trying to schedule the
2488 2492 // initial-mark GC. No point in starting a new cycle given
2489 2493 // that the whole heap was collected anyway.
2490 2494 }
2491 2495
2492 2496 if (retry_gc) {
2493 2497 if (GC_locker::is_active_and_needs_gc()) {
2494 2498 GC_locker::stall_until_clear();
2495 2499 }
2496 2500 }
2497 2501 }
2498 2502 } else {
2499 2503 if (cause == GCCause::_gc_locker
2500 2504 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2501 2505
2502 2506 // Schedule a standard evacuation pause. We're setting word_size
2503 2507 // to 0 which means that we are not requesting a post-GC allocation.
2504 2508 VM_G1IncCollectionPause op(gc_count_before,
2505 2509 0, /* word_size */
2506 2510 false, /* should_initiate_conc_mark */
2507 2511 g1_policy()->max_pause_time_ms(),
2508 2512 cause);
2509 2513 VMThread::execute(&op);
2510 2514 } else {
2511 2515 // Schedule a Full GC.
2512 2516 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2513 2517 VMThread::execute(&op);
2514 2518 }
2515 2519 }
2516 2520 } while (retry_gc);
2517 2521 }
2518 2522
2519 2523 bool G1CollectedHeap::is_in(const void* p) const {
2520 2524 if (_g1_committed.contains(p)) {
2521 2525 // Given that we know that p is in the committed space,
2522 2526 // heap_region_containing_raw() should successfully
2523 2527 // return the containing region.
2524 2528 HeapRegion* hr = heap_region_containing_raw(p);
2525 2529 return hr->is_in(p);
2526 2530 } else {
2527 2531 return false;
2528 2532 }
2529 2533 }
2530 2534
2531 2535 // Iteration functions.
2532 2536
2533 2537 // Iterates an OopClosure over all ref-containing fields of objects
2534 2538 // within a HeapRegion.
2535 2539
2536 2540 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2537 2541 MemRegion _mr;
2538 2542 ExtendedOopClosure* _cl;
2539 2543 public:
2540 2544 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2541 2545 : _mr(mr), _cl(cl) {}
2542 2546 bool doHeapRegion(HeapRegion* r) {
2543 2547 if (!r->continuesHumongous()) {
2544 2548 r->oop_iterate(_cl);
2545 2549 }
2546 2550 return false;
2547 2551 }
2548 2552 };
2549 2553
2550 2554 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2551 2555 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2552 2556 heap_region_iterate(&blk);
2553 2557 }
2554 2558
2555 2559 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2556 2560 IterateOopClosureRegionClosure blk(mr, cl);
2557 2561 heap_region_iterate(&blk);
2558 2562 }
2559 2563
2560 2564 // Iterates an ObjectClosure over all objects within a HeapRegion.
2561 2565
2562 2566 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2563 2567 ObjectClosure* _cl;
2564 2568 public:
2565 2569 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2566 2570 bool doHeapRegion(HeapRegion* r) {
2567 2571 if (! r->continuesHumongous()) {
2568 2572 r->object_iterate(_cl);
2569 2573 }
2570 2574 return false;
2571 2575 }
2572 2576 };
2573 2577
2574 2578 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2575 2579 IterateObjectClosureRegionClosure blk(cl);
2576 2580 heap_region_iterate(&blk);
2577 2581 }
2578 2582
2579 2583 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2580 2584 // FIXME: is this right?
2581 2585 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2582 2586 }
2583 2587
2584 2588 // Calls a SpaceClosure on a HeapRegion.
2585 2589
2586 2590 class SpaceClosureRegionClosure: public HeapRegionClosure {
2587 2591 SpaceClosure* _cl;
2588 2592 public:
2589 2593 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2590 2594 bool doHeapRegion(HeapRegion* r) {
2591 2595 _cl->do_space(r);
2592 2596 return false;
2593 2597 }
2594 2598 };
2595 2599
2596 2600 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2597 2601 SpaceClosureRegionClosure blk(cl);
2598 2602 heap_region_iterate(&blk);
2599 2603 }
2600 2604
2601 2605 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2602 2606 _hrs.iterate(cl);
2603 2607 }
2604 2608
2605 2609 void
2606 2610 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2607 2611 uint worker_id,
2608 2612 uint no_of_par_workers,
2609 2613 jint claim_value) {
2610 2614 const uint regions = n_regions();
2611 2615 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2612 2616 no_of_par_workers :
2613 2617 1);
2614 2618 assert(UseDynamicNumberOfGCThreads ||
2615 2619 no_of_par_workers == workers()->total_workers(),
2616 2620 "Non dynamic should use fixed number of workers");
2617 2621 // try to spread out the starting points of the workers
2618 2622 const HeapRegion* start_hr =
2619 2623 start_region_for_worker(worker_id, no_of_par_workers);
2620 2624 const uint start_index = start_hr->hrs_index();
2621 2625
2622 2626 // each worker will actually look at all regions
2623 2627 for (uint count = 0; count < regions; ++count) {
2624 2628 const uint index = (start_index + count) % regions;
2625 2629 assert(0 <= index && index < regions, "sanity");
2626 2630 HeapRegion* r = region_at(index);
2627 2631 // we'll ignore "continues humongous" regions (we'll process them
2628 2632 // when we come across their corresponding "start humongous"
2629 2633 // region) and regions already claimed
2630 2634 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2631 2635 continue;
2632 2636 }
2633 2637 // OK, try to claim it
2634 2638 if (r->claimHeapRegion(claim_value)) {
2635 2639 // success!
2636 2640 assert(!r->continuesHumongous(), "sanity");
2637 2641 if (r->startsHumongous()) {
2638 2642 // If the region is "starts humongous" we'll iterate over its
2639 2643 // "continues humongous" first; in fact we'll do them
2640 2644 // first. The order is important. In on case, calling the
2641 2645 // closure on the "starts humongous" region might de-allocate
2642 2646 // and clear all its "continues humongous" regions and, as a
2643 2647 // result, we might end up processing them twice. So, we'll do
2644 2648 // them first (notice: most closures will ignore them anyway) and
2645 2649 // then we'll do the "starts humongous" region.
2646 2650 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2647 2651 HeapRegion* chr = region_at(ch_index);
2648 2652
2649 2653 // if the region has already been claimed or it's not
2650 2654 // "continues humongous" we're done
2651 2655 if (chr->claim_value() == claim_value ||
2652 2656 !chr->continuesHumongous()) {
2653 2657 break;
2654 2658 }
2655 2659
2656 2660 // Noone should have claimed it directly. We can given
2657 2661 // that we claimed its "starts humongous" region.
2658 2662 assert(chr->claim_value() != claim_value, "sanity");
2659 2663 assert(chr->humongous_start_region() == r, "sanity");
2660 2664
2661 2665 if (chr->claimHeapRegion(claim_value)) {
2662 2666 // we should always be able to claim it; noone else should
2663 2667 // be trying to claim this region
2664 2668
2665 2669 bool res2 = cl->doHeapRegion(chr);
2666 2670 assert(!res2, "Should not abort");
2667 2671
2668 2672 // Right now, this holds (i.e., no closure that actually
2669 2673 // does something with "continues humongous" regions
2670 2674 // clears them). We might have to weaken it in the future,
2671 2675 // but let's leave these two asserts here for extra safety.
2672 2676 assert(chr->continuesHumongous(), "should still be the case");
2673 2677 assert(chr->humongous_start_region() == r, "sanity");
2674 2678 } else {
2675 2679 guarantee(false, "we should not reach here");
2676 2680 }
2677 2681 }
2678 2682 }
2679 2683
2680 2684 assert(!r->continuesHumongous(), "sanity");
2681 2685 bool res = cl->doHeapRegion(r);
2682 2686 assert(!res, "Should not abort");
2683 2687 }
2684 2688 }
2685 2689 }
2686 2690
2687 2691 class ResetClaimValuesClosure: public HeapRegionClosure {
2688 2692 public:
2689 2693 bool doHeapRegion(HeapRegion* r) {
2690 2694 r->set_claim_value(HeapRegion::InitialClaimValue);
2691 2695 return false;
2692 2696 }
2693 2697 };
2694 2698
2695 2699 void G1CollectedHeap::reset_heap_region_claim_values() {
2696 2700 ResetClaimValuesClosure blk;
2697 2701 heap_region_iterate(&blk);
2698 2702 }
2699 2703
2700 2704 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2701 2705 ResetClaimValuesClosure blk;
2702 2706 collection_set_iterate(&blk);
2703 2707 }
2704 2708
2705 2709 #ifdef ASSERT
2706 2710 // This checks whether all regions in the heap have the correct claim
2707 2711 // value. I also piggy-backed on this a check to ensure that the
2708 2712 // humongous_start_region() information on "continues humongous"
2709 2713 // regions is correct.
2710 2714
2711 2715 class CheckClaimValuesClosure : public HeapRegionClosure {
2712 2716 private:
2713 2717 jint _claim_value;
2714 2718 uint _failures;
2715 2719 HeapRegion* _sh_region;
2716 2720
2717 2721 public:
2718 2722 CheckClaimValuesClosure(jint claim_value) :
2719 2723 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2720 2724 bool doHeapRegion(HeapRegion* r) {
2721 2725 if (r->claim_value() != _claim_value) {
2722 2726 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2723 2727 "claim value = %d, should be %d",
2724 2728 HR_FORMAT_PARAMS(r),
2725 2729 r->claim_value(), _claim_value);
2726 2730 ++_failures;
2727 2731 }
2728 2732 if (!r->isHumongous()) {
2729 2733 _sh_region = NULL;
2730 2734 } else if (r->startsHumongous()) {
2731 2735 _sh_region = r;
2732 2736 } else if (r->continuesHumongous()) {
2733 2737 if (r->humongous_start_region() != _sh_region) {
2734 2738 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2735 2739 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2736 2740 HR_FORMAT_PARAMS(r),
2737 2741 r->humongous_start_region(),
2738 2742 _sh_region);
2739 2743 ++_failures;
2740 2744 }
2741 2745 }
2742 2746 return false;
2743 2747 }
2744 2748 uint failures() { return _failures; }
2745 2749 };
2746 2750
2747 2751 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2748 2752 CheckClaimValuesClosure cl(claim_value);
2749 2753 heap_region_iterate(&cl);
2750 2754 return cl.failures() == 0;
2751 2755 }
2752 2756
2753 2757 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2754 2758 private:
2755 2759 jint _claim_value;
2756 2760 uint _failures;
2757 2761
2758 2762 public:
2759 2763 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2760 2764 _claim_value(claim_value), _failures(0) { }
2761 2765
2762 2766 uint failures() { return _failures; }
2763 2767
2764 2768 bool doHeapRegion(HeapRegion* hr) {
2765 2769 assert(hr->in_collection_set(), "how?");
2766 2770 assert(!hr->isHumongous(), "H-region in CSet");
2767 2771 if (hr->claim_value() != _claim_value) {
2768 2772 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2769 2773 "claim value = %d, should be %d",
2770 2774 HR_FORMAT_PARAMS(hr),
2771 2775 hr->claim_value(), _claim_value);
2772 2776 _failures += 1;
2773 2777 }
2774 2778 return false;
2775 2779 }
2776 2780 };
2777 2781
2778 2782 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2779 2783 CheckClaimValuesInCSetHRClosure cl(claim_value);
2780 2784 collection_set_iterate(&cl);
2781 2785 return cl.failures() == 0;
2782 2786 }
2783 2787 #endif // ASSERT
2784 2788
2785 2789 // Clear the cached CSet starting regions and (more importantly)
2786 2790 // the time stamps. Called when we reset the GC time stamp.
2787 2791 void G1CollectedHeap::clear_cset_start_regions() {
2788 2792 assert(_worker_cset_start_region != NULL, "sanity");
2789 2793 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2790 2794
2791 2795 int n_queues = MAX2((int)ParallelGCThreads, 1);
2792 2796 for (int i = 0; i < n_queues; i++) {
2793 2797 _worker_cset_start_region[i] = NULL;
2794 2798 _worker_cset_start_region_time_stamp[i] = 0;
2795 2799 }
2796 2800 }
2797 2801
2798 2802 // Given the id of a worker, obtain or calculate a suitable
2799 2803 // starting region for iterating over the current collection set.
2800 2804 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2801 2805 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2802 2806
2803 2807 HeapRegion* result = NULL;
2804 2808 unsigned gc_time_stamp = get_gc_time_stamp();
2805 2809
2806 2810 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2807 2811 // Cached starting region for current worker was set
2808 2812 // during the current pause - so it's valid.
2809 2813 // Note: the cached starting heap region may be NULL
2810 2814 // (when the collection set is empty).
2811 2815 result = _worker_cset_start_region[worker_i];
2812 2816 assert(result == NULL || result->in_collection_set(), "sanity");
2813 2817 return result;
2814 2818 }
2815 2819
2816 2820 // The cached entry was not valid so let's calculate
2817 2821 // a suitable starting heap region for this worker.
2818 2822
2819 2823 // We want the parallel threads to start their collection
2820 2824 // set iteration at different collection set regions to
2821 2825 // avoid contention.
2822 2826 // If we have:
2823 2827 // n collection set regions
2824 2828 // p threads
2825 2829 // Then thread t will start at region floor ((t * n) / p)
2826 2830
2827 2831 result = g1_policy()->collection_set();
2828 2832 if (G1CollectedHeap::use_parallel_gc_threads()) {
2829 2833 uint cs_size = g1_policy()->cset_region_length();
2830 2834 uint active_workers = workers()->active_workers();
2831 2835 assert(UseDynamicNumberOfGCThreads ||
2832 2836 active_workers == workers()->total_workers(),
2833 2837 "Unless dynamic should use total workers");
2834 2838
2835 2839 uint end_ind = (cs_size * worker_i) / active_workers;
2836 2840 uint start_ind = 0;
2837 2841
2838 2842 if (worker_i > 0 &&
2839 2843 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2840 2844 // Previous workers starting region is valid
2841 2845 // so let's iterate from there
2842 2846 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2843 2847 result = _worker_cset_start_region[worker_i - 1];
2844 2848 }
2845 2849
2846 2850 for (uint i = start_ind; i < end_ind; i++) {
2847 2851 result = result->next_in_collection_set();
2848 2852 }
2849 2853 }
2850 2854
2851 2855 // Note: the calculated starting heap region may be NULL
2852 2856 // (when the collection set is empty).
2853 2857 assert(result == NULL || result->in_collection_set(), "sanity");
2854 2858 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2855 2859 "should be updated only once per pause");
2856 2860 _worker_cset_start_region[worker_i] = result;
2857 2861 OrderAccess::storestore();
2858 2862 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2859 2863 return result;
2860 2864 }
2861 2865
2862 2866 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2863 2867 uint no_of_par_workers) {
2864 2868 uint worker_num =
2865 2869 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2866 2870 assert(UseDynamicNumberOfGCThreads ||
2867 2871 no_of_par_workers == workers()->total_workers(),
2868 2872 "Non dynamic should use fixed number of workers");
2869 2873 const uint start_index = n_regions() * worker_i / worker_num;
2870 2874 return region_at(start_index);
2871 2875 }
2872 2876
2873 2877 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2874 2878 HeapRegion* r = g1_policy()->collection_set();
2875 2879 while (r != NULL) {
2876 2880 HeapRegion* next = r->next_in_collection_set();
2877 2881 if (cl->doHeapRegion(r)) {
2878 2882 cl->incomplete();
2879 2883 return;
2880 2884 }
2881 2885 r = next;
2882 2886 }
2883 2887 }
2884 2888
2885 2889 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2886 2890 HeapRegionClosure *cl) {
2887 2891 if (r == NULL) {
2888 2892 // The CSet is empty so there's nothing to do.
2889 2893 return;
2890 2894 }
2891 2895
2892 2896 assert(r->in_collection_set(),
2893 2897 "Start region must be a member of the collection set.");
2894 2898 HeapRegion* cur = r;
2895 2899 while (cur != NULL) {
2896 2900 HeapRegion* next = cur->next_in_collection_set();
2897 2901 if (cl->doHeapRegion(cur) && false) {
2898 2902 cl->incomplete();
2899 2903 return;
2900 2904 }
2901 2905 cur = next;
2902 2906 }
2903 2907 cur = g1_policy()->collection_set();
2904 2908 while (cur != r) {
2905 2909 HeapRegion* next = cur->next_in_collection_set();
2906 2910 if (cl->doHeapRegion(cur) && false) {
2907 2911 cl->incomplete();
2908 2912 return;
2909 2913 }
2910 2914 cur = next;
2911 2915 }
2912 2916 }
2913 2917
2914 2918 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2915 2919 return n_regions() > 0 ? region_at(0) : NULL;
2916 2920 }
2917 2921
2918 2922
2919 2923 Space* G1CollectedHeap::space_containing(const void* addr) const {
2920 2924 Space* res = heap_region_containing(addr);
2921 2925 return res;
2922 2926 }
2923 2927
2924 2928 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2925 2929 Space* sp = space_containing(addr);
2926 2930 if (sp != NULL) {
2927 2931 return sp->block_start(addr);
2928 2932 }
2929 2933 return NULL;
2930 2934 }
2931 2935
2932 2936 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2933 2937 Space* sp = space_containing(addr);
2934 2938 assert(sp != NULL, "block_size of address outside of heap");
2935 2939 return sp->block_size(addr);
2936 2940 }
2937 2941
2938 2942 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2939 2943 Space* sp = space_containing(addr);
2940 2944 return sp->block_is_obj(addr);
2941 2945 }
2942 2946
2943 2947 bool G1CollectedHeap::supports_tlab_allocation() const {
2944 2948 return true;
2945 2949 }
2946 2950
2947 2951 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2948 2952 return HeapRegion::GrainBytes;
2949 2953 }
2950 2954
2951 2955 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2952 2956 // Return the remaining space in the cur alloc region, but not less than
2953 2957 // the min TLAB size.
2954 2958
2955 2959 // Also, this value can be at most the humongous object threshold,
2956 2960 // since we can't allow tlabs to grow big enough to accomodate
2957 2961 // humongous objects.
2958 2962
2959 2963 HeapRegion* hr = _mutator_alloc_region.get();
2960 2964 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2961 2965 if (hr == NULL) {
2962 2966 return max_tlab_size;
2963 2967 } else {
2964 2968 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2965 2969 }
2966 2970 }
2967 2971
2968 2972 size_t G1CollectedHeap::max_capacity() const {
2969 2973 return _g1_reserved.byte_size();
2970 2974 }
2971 2975
2972 2976 jlong G1CollectedHeap::millis_since_last_gc() {
2973 2977 // assert(false, "NYI");
2974 2978 return 0;
2975 2979 }
2976 2980
2977 2981 void G1CollectedHeap::prepare_for_verify() {
2978 2982 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2979 2983 ensure_parsability(false);
2980 2984 }
2981 2985 g1_rem_set()->prepare_for_verify();
2982 2986 }
2983 2987
2984 2988 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
2985 2989 VerifyOption vo) {
2986 2990 switch (vo) {
2987 2991 case VerifyOption_G1UsePrevMarking:
2988 2992 return hr->obj_allocated_since_prev_marking(obj);
2989 2993 case VerifyOption_G1UseNextMarking:
2990 2994 return hr->obj_allocated_since_next_marking(obj);
2991 2995 case VerifyOption_G1UseMarkWord:
2992 2996 return false;
2993 2997 default:
2994 2998 ShouldNotReachHere();
2995 2999 }
2996 3000 return false; // keep some compilers happy
2997 3001 }
2998 3002
2999 3003 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3000 3004 switch (vo) {
3001 3005 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3002 3006 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3003 3007 case VerifyOption_G1UseMarkWord: return NULL;
3004 3008 default: ShouldNotReachHere();
3005 3009 }
3006 3010 return NULL; // keep some compilers happy
3007 3011 }
3008 3012
3009 3013 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3010 3014 switch (vo) {
3011 3015 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3012 3016 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3013 3017 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3014 3018 default: ShouldNotReachHere();
3015 3019 }
3016 3020 return false; // keep some compilers happy
3017 3021 }
3018 3022
3019 3023 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3020 3024 switch (vo) {
3021 3025 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3022 3026 case VerifyOption_G1UseNextMarking: return "NTAMS";
3023 3027 case VerifyOption_G1UseMarkWord: return "NONE";
3024 3028 default: ShouldNotReachHere();
3025 3029 }
3026 3030 return NULL; // keep some compilers happy
3027 3031 }
3028 3032
3029 3033 class VerifyLivenessOopClosure: public OopClosure {
3030 3034 G1CollectedHeap* _g1h;
3031 3035 VerifyOption _vo;
3032 3036 public:
3033 3037 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3034 3038 _g1h(g1h), _vo(vo)
3035 3039 { }
3036 3040 void do_oop(narrowOop *p) { do_oop_work(p); }
3037 3041 void do_oop( oop *p) { do_oop_work(p); }
3038 3042
3039 3043 template <class T> void do_oop_work(T *p) {
3040 3044 oop obj = oopDesc::load_decode_heap_oop(p);
3041 3045 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3042 3046 "Dead object referenced by a not dead object");
3043 3047 }
3044 3048 };
3045 3049
3046 3050 class VerifyObjsInRegionClosure: public ObjectClosure {
3047 3051 private:
3048 3052 G1CollectedHeap* _g1h;
3049 3053 size_t _live_bytes;
3050 3054 HeapRegion *_hr;
3051 3055 VerifyOption _vo;
3052 3056 public:
3053 3057 // _vo == UsePrevMarking -> use "prev" marking information,
3054 3058 // _vo == UseNextMarking -> use "next" marking information,
3055 3059 // _vo == UseMarkWord -> use mark word from object header.
3056 3060 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3057 3061 : _live_bytes(0), _hr(hr), _vo(vo) {
3058 3062 _g1h = G1CollectedHeap::heap();
3059 3063 }
3060 3064 void do_object(oop o) {
3061 3065 VerifyLivenessOopClosure isLive(_g1h, _vo);
3062 3066 assert(o != NULL, "Huh?");
3063 3067 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3064 3068 // If the object is alive according to the mark word,
3065 3069 // then verify that the marking information agrees.
3066 3070 // Note we can't verify the contra-positive of the
3067 3071 // above: if the object is dead (according to the mark
3068 3072 // word), it may not be marked, or may have been marked
3069 3073 // but has since became dead, or may have been allocated
3070 3074 // since the last marking.
3071 3075 if (_vo == VerifyOption_G1UseMarkWord) {
3072 3076 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3073 3077 }
3074 3078
3075 3079 o->oop_iterate_no_header(&isLive);
3076 3080 if (!_hr->obj_allocated_since_prev_marking(o)) {
3077 3081 size_t obj_size = o->size(); // Make sure we don't overflow
3078 3082 _live_bytes += (obj_size * HeapWordSize);
3079 3083 }
3080 3084 }
3081 3085 }
3082 3086 size_t live_bytes() { return _live_bytes; }
3083 3087 };
3084 3088
3085 3089 class PrintObjsInRegionClosure : public ObjectClosure {
3086 3090 HeapRegion *_hr;
3087 3091 G1CollectedHeap *_g1;
3088 3092 public:
3089 3093 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3090 3094 _g1 = G1CollectedHeap::heap();
3091 3095 };
3092 3096
3093 3097 void do_object(oop o) {
3094 3098 if (o != NULL) {
3095 3099 HeapWord *start = (HeapWord *) o;
3096 3100 size_t word_sz = o->size();
3097 3101 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3098 3102 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3099 3103 (void*) o, word_sz,
3100 3104 _g1->isMarkedPrev(o),
3101 3105 _g1->isMarkedNext(o),
3102 3106 _hr->obj_allocated_since_prev_marking(o));
3103 3107 HeapWord *end = start + word_sz;
3104 3108 HeapWord *cur;
3105 3109 int *val;
3106 3110 for (cur = start; cur < end; cur++) {
3107 3111 val = (int *) cur;
3108 3112 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3109 3113 }
3110 3114 }
3111 3115 }
3112 3116 };
3113 3117
3114 3118 class VerifyRegionClosure: public HeapRegionClosure {
3115 3119 private:
3116 3120 bool _par;
3117 3121 VerifyOption _vo;
3118 3122 bool _failures;
3119 3123 public:
3120 3124 // _vo == UsePrevMarking -> use "prev" marking information,
3121 3125 // _vo == UseNextMarking -> use "next" marking information,
3122 3126 // _vo == UseMarkWord -> use mark word from object header.
3123 3127 VerifyRegionClosure(bool par, VerifyOption vo)
3124 3128 : _par(par),
3125 3129 _vo(vo),
3126 3130 _failures(false) {}
3127 3131
3128 3132 bool failures() {
3129 3133 return _failures;
3130 3134 }
3131 3135
3132 3136 bool doHeapRegion(HeapRegion* r) {
3133 3137 if (!r->continuesHumongous()) {
3134 3138 bool failures = false;
3135 3139 r->verify(_vo, &failures);
3136 3140 if (failures) {
3137 3141 _failures = true;
3138 3142 } else {
3139 3143 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3140 3144 r->object_iterate(¬_dead_yet_cl);
3141 3145 if (_vo != VerifyOption_G1UseNextMarking) {
3142 3146 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3143 3147 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3144 3148 "max_live_bytes "SIZE_FORMAT" "
3145 3149 "< calculated "SIZE_FORMAT,
3146 3150 r->bottom(), r->end(),
3147 3151 r->max_live_bytes(),
3148 3152 not_dead_yet_cl.live_bytes());
3149 3153 _failures = true;
3150 3154 }
3151 3155 } else {
3152 3156 // When vo == UseNextMarking we cannot currently do a sanity
3153 3157 // check on the live bytes as the calculation has not been
3154 3158 // finalized yet.
3155 3159 }
3156 3160 }
3157 3161 }
3158 3162 return false; // stop the region iteration if we hit a failure
3159 3163 }
3160 3164 };
3161 3165
3162 3166 class YoungRefCounterClosure : public OopClosure {
3163 3167 G1CollectedHeap* _g1h;
3164 3168 int _count;
3165 3169 public:
3166 3170 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3167 3171 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3168 3172 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3169 3173
3170 3174 int count() { return _count; }
3171 3175 void reset_count() { _count = 0; };
3172 3176 };
3173 3177
3174 3178 class VerifyKlassClosure: public KlassClosure {
3175 3179 YoungRefCounterClosure _young_ref_counter_closure;
3176 3180 OopClosure *_oop_closure;
3177 3181 public:
3178 3182 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3179 3183 void do_klass(Klass* k) {
3180 3184 k->oops_do(_oop_closure);
3181 3185
3182 3186 _young_ref_counter_closure.reset_count();
3183 3187 k->oops_do(&_young_ref_counter_closure);
3184 3188 if (_young_ref_counter_closure.count() > 0) {
3185 3189 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3186 3190 }
3187 3191 }
3188 3192 };
3189 3193
3190 3194 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3191 3195 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3192 3196 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3193 3197 // we can change this closure to extend the simpler OopClosure.
3194 3198 class VerifyRootsClosure: public OopsInGenClosure {
3195 3199 private:
3196 3200 G1CollectedHeap* _g1h;
3197 3201 VerifyOption _vo;
3198 3202 bool _failures;
3199 3203 public:
3200 3204 // _vo == UsePrevMarking -> use "prev" marking information,
3201 3205 // _vo == UseNextMarking -> use "next" marking information,
3202 3206 // _vo == UseMarkWord -> use mark word from object header.
3203 3207 VerifyRootsClosure(VerifyOption vo) :
3204 3208 _g1h(G1CollectedHeap::heap()),
3205 3209 _vo(vo),
3206 3210 _failures(false) { }
3207 3211
3208 3212 bool failures() { return _failures; }
3209 3213
3210 3214 template <class T> void do_oop_nv(T* p) {
3211 3215 T heap_oop = oopDesc::load_heap_oop(p);
3212 3216 if (!oopDesc::is_null(heap_oop)) {
3213 3217 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3214 3218 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3215 3219 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3216 3220 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3217 3221 if (_vo == VerifyOption_G1UseMarkWord) {
3218 3222 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3219 3223 }
3220 3224 obj->print_on(gclog_or_tty);
3221 3225 _failures = true;
3222 3226 }
3223 3227 }
3224 3228 }
3225 3229
3226 3230 void do_oop(oop* p) { do_oop_nv(p); }
3227 3231 void do_oop(narrowOop* p) { do_oop_nv(p); }
3228 3232 };
3229 3233
3230 3234 // This is the task used for parallel heap verification.
3231 3235
3232 3236 class G1ParVerifyTask: public AbstractGangTask {
3233 3237 private:
3234 3238 G1CollectedHeap* _g1h;
3235 3239 VerifyOption _vo;
3236 3240 bool _failures;
3237 3241
3238 3242 public:
3239 3243 // _vo == UsePrevMarking -> use "prev" marking information,
3240 3244 // _vo == UseNextMarking -> use "next" marking information,
3241 3245 // _vo == UseMarkWord -> use mark word from object header.
3242 3246 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3243 3247 AbstractGangTask("Parallel verify task"),
3244 3248 _g1h(g1h),
3245 3249 _vo(vo),
3246 3250 _failures(false) { }
3247 3251
3248 3252 bool failures() {
3249 3253 return _failures;
3250 3254 }
3251 3255
3252 3256 void work(uint worker_id) {
3253 3257 HandleMark hm;
3254 3258 VerifyRegionClosure blk(true, _vo);
3255 3259 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3256 3260 _g1h->workers()->active_workers(),
3257 3261 HeapRegion::ParVerifyClaimValue);
3258 3262 if (blk.failures()) {
3259 3263 _failures = true;
3260 3264 }
3261 3265 }
3262 3266 };
3263 3267
3264 3268 void G1CollectedHeap::verify(bool silent) {
3265 3269 verify(silent, VerifyOption_G1UsePrevMarking);
3266 3270 }
3267 3271
3268 3272 void G1CollectedHeap::verify(bool silent,
3269 3273 VerifyOption vo) {
3270 3274 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3271 3275 if (!silent) { gclog_or_tty->print("Roots "); }
3272 3276 VerifyRootsClosure rootsCl(vo);
3273 3277
3274 3278 assert(Thread::current()->is_VM_thread(),
3275 3279 "Expected to be executed serially by the VM thread at this point");
3276 3280
3277 3281 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3278 3282 VerifyKlassClosure klassCl(this, &rootsCl);
3279 3283
3280 3284 // We apply the relevant closures to all the oops in the
3281 3285 // system dictionary, the string table and the code cache.
3282 3286 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3283 3287
3284 3288 // Need cleared claim bits for the strong roots processing
3285 3289 ClassLoaderDataGraph::clear_claimed_marks();
3286 3290
3287 3291 process_strong_roots(true, // activate StrongRootsScope
3288 3292 false, // we set "is scavenging" to false,
3289 3293 // so we don't reset the dirty cards.
3290 3294 ScanningOption(so), // roots scanning options
3291 3295 &rootsCl,
3292 3296 &blobsCl,
3293 3297 &klassCl
3294 3298 );
3295 3299
3296 3300 bool failures = rootsCl.failures();
3297 3301
3298 3302 if (vo != VerifyOption_G1UseMarkWord) {
3299 3303 // If we're verifying during a full GC then the region sets
3300 3304 // will have been torn down at the start of the GC. Therefore
3301 3305 // verifying the region sets will fail. So we only verify
3302 3306 // the region sets when not in a full GC.
3303 3307 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3304 3308 verify_region_sets();
3305 3309 }
3306 3310
3307 3311 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3308 3312 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3309 3313 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3310 3314 "sanity check");
3311 3315
3312 3316 G1ParVerifyTask task(this, vo);
3313 3317 assert(UseDynamicNumberOfGCThreads ||
3314 3318 workers()->active_workers() == workers()->total_workers(),
3315 3319 "If not dynamic should be using all the workers");
3316 3320 int n_workers = workers()->active_workers();
3317 3321 set_par_threads(n_workers);
3318 3322 workers()->run_task(&task);
3319 3323 set_par_threads(0);
3320 3324 if (task.failures()) {
3321 3325 failures = true;
3322 3326 }
3323 3327
3324 3328 // Checks that the expected amount of parallel work was done.
3325 3329 // The implication is that n_workers is > 0.
3326 3330 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3327 3331 "sanity check");
3328 3332
3329 3333 reset_heap_region_claim_values();
3330 3334
3331 3335 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3332 3336 "sanity check");
3333 3337 } else {
3334 3338 VerifyRegionClosure blk(false, vo);
3335 3339 heap_region_iterate(&blk);
3336 3340 if (blk.failures()) {
3337 3341 failures = true;
3338 3342 }
3339 3343 }
3340 3344 if (!silent) gclog_or_tty->print("RemSet ");
3341 3345 rem_set()->verify();
3342 3346
3343 3347 if (failures) {
3344 3348 gclog_or_tty->print_cr("Heap:");
3345 3349 // It helps to have the per-region information in the output to
3346 3350 // help us track down what went wrong. This is why we call
3347 3351 // print_extended_on() instead of print_on().
3348 3352 print_extended_on(gclog_or_tty);
3349 3353 gclog_or_tty->print_cr("");
3350 3354 #ifndef PRODUCT
3351 3355 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3352 3356 concurrent_mark()->print_reachable("at-verification-failure",
3353 3357 vo, false /* all */);
3354 3358 }
3355 3359 #endif
3356 3360 gclog_or_tty->flush();
3357 3361 }
3358 3362 guarantee(!failures, "there should not have been any failures");
3359 3363 } else {
3360 3364 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3361 3365 }
3362 3366 }
3363 3367
3364 3368 class PrintRegionClosure: public HeapRegionClosure {
3365 3369 outputStream* _st;
3366 3370 public:
3367 3371 PrintRegionClosure(outputStream* st) : _st(st) {}
3368 3372 bool doHeapRegion(HeapRegion* r) {
3369 3373 r->print_on(_st);
3370 3374 return false;
3371 3375 }
3372 3376 };
3373 3377
3374 3378 void G1CollectedHeap::print_on(outputStream* st) const {
3375 3379 st->print(" %-20s", "garbage-first heap");
3376 3380 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3377 3381 capacity()/K, used_unlocked()/K);
3378 3382 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3379 3383 _g1_storage.low_boundary(),
3380 3384 _g1_storage.high(),
3381 3385 _g1_storage.high_boundary());
3382 3386 st->cr();
3383 3387 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3384 3388 uint young_regions = _young_list->length();
3385 3389 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3386 3390 (size_t) young_regions * HeapRegion::GrainBytes / K);
3387 3391 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3388 3392 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3389 3393 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3390 3394 st->cr();
3391 3395 MetaspaceAux::print_on(st);
3392 3396 }
3393 3397
3394 3398 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3395 3399 print_on(st);
3396 3400
3397 3401 // Print the per-region information.
3398 3402 st->cr();
3399 3403 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3400 3404 "HS=humongous(starts), HC=humongous(continues), "
3401 3405 "CS=collection set, F=free, TS=gc time stamp, "
3402 3406 "PTAMS=previous top-at-mark-start, "
3403 3407 "NTAMS=next top-at-mark-start)");
3404 3408 PrintRegionClosure blk(st);
3405 3409 heap_region_iterate(&blk);
3406 3410 }
3407 3411
3408 3412 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3409 3413 if (G1CollectedHeap::use_parallel_gc_threads()) {
3410 3414 workers()->print_worker_threads_on(st);
3411 3415 }
3412 3416 _cmThread->print_on(st);
3413 3417 st->cr();
3414 3418 _cm->print_worker_threads_on(st);
3415 3419 _cg1r->print_worker_threads_on(st);
3416 3420 }
3417 3421
3418 3422 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3419 3423 if (G1CollectedHeap::use_parallel_gc_threads()) {
3420 3424 workers()->threads_do(tc);
3421 3425 }
3422 3426 tc->do_thread(_cmThread);
3423 3427 _cg1r->threads_do(tc);
3424 3428 }
3425 3429
3426 3430 void G1CollectedHeap::print_tracing_info() const {
3427 3431 // We'll overload this to mean "trace GC pause statistics."
3428 3432 if (TraceGen0Time || TraceGen1Time) {
3429 3433 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3430 3434 // to that.
3431 3435 g1_policy()->print_tracing_info();
3432 3436 }
3433 3437 if (G1SummarizeRSetStats) {
3434 3438 g1_rem_set()->print_summary_info();
3435 3439 }
3436 3440 if (G1SummarizeConcMark) {
3437 3441 concurrent_mark()->print_summary_info();
3438 3442 }
3439 3443 g1_policy()->print_yg_surv_rate_info();
3440 3444 SpecializationStats::print();
3441 3445 }
3442 3446
3443 3447 #ifndef PRODUCT
3444 3448 // Helpful for debugging RSet issues.
3445 3449
3446 3450 class PrintRSetsClosure : public HeapRegionClosure {
3447 3451 private:
3448 3452 const char* _msg;
3449 3453 size_t _occupied_sum;
3450 3454
3451 3455 public:
3452 3456 bool doHeapRegion(HeapRegion* r) {
3453 3457 HeapRegionRemSet* hrrs = r->rem_set();
3454 3458 size_t occupied = hrrs->occupied();
3455 3459 _occupied_sum += occupied;
3456 3460
3457 3461 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3458 3462 HR_FORMAT_PARAMS(r));
3459 3463 if (occupied == 0) {
3460 3464 gclog_or_tty->print_cr(" RSet is empty");
3461 3465 } else {
3462 3466 hrrs->print();
3463 3467 }
3464 3468 gclog_or_tty->print_cr("----------");
3465 3469 return false;
3466 3470 }
3467 3471
3468 3472 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3469 3473 gclog_or_tty->cr();
3470 3474 gclog_or_tty->print_cr("========================================");
3471 3475 gclog_or_tty->print_cr(msg);
3472 3476 gclog_or_tty->cr();
3473 3477 }
3474 3478
3475 3479 ~PrintRSetsClosure() {
3476 3480 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3477 3481 gclog_or_tty->print_cr("========================================");
3478 3482 gclog_or_tty->cr();
3479 3483 }
3480 3484 };
3481 3485
3482 3486 void G1CollectedHeap::print_cset_rsets() {
3483 3487 PrintRSetsClosure cl("Printing CSet RSets");
3484 3488 collection_set_iterate(&cl);
3485 3489 }
3486 3490
3487 3491 void G1CollectedHeap::print_all_rsets() {
3488 3492 PrintRSetsClosure cl("Printing All RSets");;
3489 3493 heap_region_iterate(&cl);
3490 3494 }
3491 3495 #endif // PRODUCT
3492 3496
3493 3497 G1CollectedHeap* G1CollectedHeap::heap() {
3494 3498 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3495 3499 "not a garbage-first heap");
3496 3500 return _g1h;
3497 3501 }
3498 3502
3499 3503 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3500 3504 // always_do_update_barrier = false;
3501 3505 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3502 3506 // Call allocation profiler
3503 3507 AllocationProfiler::iterate_since_last_gc();
3504 3508 // Fill TLAB's and such
3505 3509 ensure_parsability(true);
3506 3510 }
3507 3511
3508 3512 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3509 3513 // FIXME: what is this about?
3510 3514 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3511 3515 // is set.
3512 3516 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3513 3517 "derived pointer present"));
3514 3518 // always_do_update_barrier = true;
3515 3519
3516 3520 // We have just completed a GC. Update the soft reference
3517 3521 // policy with the new heap occupancy
3518 3522 Universe::update_heap_info_at_gc();
3519 3523 }
3520 3524
3521 3525 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3522 3526 unsigned int gc_count_before,
3523 3527 bool* succeeded) {
3524 3528 assert_heap_not_locked_and_not_at_safepoint();
3525 3529 g1_policy()->record_stop_world_start();
3526 3530 VM_G1IncCollectionPause op(gc_count_before,
3527 3531 word_size,
3528 3532 false, /* should_initiate_conc_mark */
3529 3533 g1_policy()->max_pause_time_ms(),
3530 3534 GCCause::_g1_inc_collection_pause);
3531 3535 VMThread::execute(&op);
3532 3536
3533 3537 HeapWord* result = op.result();
3534 3538 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3535 3539 assert(result == NULL || ret_succeeded,
3536 3540 "the result should be NULL if the VM did not succeed");
3537 3541 *succeeded = ret_succeeded;
3538 3542
3539 3543 assert_heap_not_locked();
3540 3544 return result;
3541 3545 }
3542 3546
3543 3547 void
3544 3548 G1CollectedHeap::doConcurrentMark() {
3545 3549 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3546 3550 if (!_cmThread->in_progress()) {
3547 3551 _cmThread->set_started();
3548 3552 CGC_lock->notify();
3549 3553 }
3550 3554 }
3551 3555
3552 3556 size_t G1CollectedHeap::pending_card_num() {
3553 3557 size_t extra_cards = 0;
3554 3558 JavaThread *curr = Threads::first();
3555 3559 while (curr != NULL) {
3556 3560 DirtyCardQueue& dcq = curr->dirty_card_queue();
3557 3561 extra_cards += dcq.size();
3558 3562 curr = curr->next();
3559 3563 }
3560 3564 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3561 3565 size_t buffer_size = dcqs.buffer_size();
3562 3566 size_t buffer_num = dcqs.completed_buffers_num();
3563 3567
3564 3568 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3565 3569 // in bytes - not the number of 'entries'. We need to convert
3566 3570 // into a number of cards.
3567 3571 return (buffer_size * buffer_num + extra_cards) / oopSize;
3568 3572 }
3569 3573
3570 3574 size_t G1CollectedHeap::cards_scanned() {
3571 3575 return g1_rem_set()->cardsScanned();
3572 3576 }
3573 3577
3574 3578 void
3575 3579 G1CollectedHeap::setup_surviving_young_words() {
3576 3580 assert(_surviving_young_words == NULL, "pre-condition");
3577 3581 uint array_length = g1_policy()->young_cset_region_length();
3578 3582 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3579 3583 if (_surviving_young_words == NULL) {
3580 3584 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3581 3585 "Not enough space for young surv words summary.");
3582 3586 }
3583 3587 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3584 3588 #ifdef ASSERT
3585 3589 for (uint i = 0; i < array_length; ++i) {
3586 3590 assert( _surviving_young_words[i] == 0, "memset above" );
3587 3591 }
3588 3592 #endif // !ASSERT
3589 3593 }
3590 3594
3591 3595 void
3592 3596 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3593 3597 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3594 3598 uint array_length = g1_policy()->young_cset_region_length();
3595 3599 for (uint i = 0; i < array_length; ++i) {
3596 3600 _surviving_young_words[i] += surv_young_words[i];
3597 3601 }
3598 3602 }
3599 3603
3600 3604 void
3601 3605 G1CollectedHeap::cleanup_surviving_young_words() {
3602 3606 guarantee( _surviving_young_words != NULL, "pre-condition" );
3603 3607 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3604 3608 _surviving_young_words = NULL;
3605 3609 }
3606 3610
3607 3611 #ifdef ASSERT
3608 3612 class VerifyCSetClosure: public HeapRegionClosure {
3609 3613 public:
3610 3614 bool doHeapRegion(HeapRegion* hr) {
3611 3615 // Here we check that the CSet region's RSet is ready for parallel
3612 3616 // iteration. The fields that we'll verify are only manipulated
3613 3617 // when the region is part of a CSet and is collected. Afterwards,
3614 3618 // we reset these fields when we clear the region's RSet (when the
3615 3619 // region is freed) so they are ready when the region is
3616 3620 // re-allocated. The only exception to this is if there's an
3617 3621 // evacuation failure and instead of freeing the region we leave
3618 3622 // it in the heap. In that case, we reset these fields during
3619 3623 // evacuation failure handling.
3620 3624 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3621 3625
3622 3626 // Here's a good place to add any other checks we'd like to
3623 3627 // perform on CSet regions.
3624 3628 return false;
3625 3629 }
3626 3630 };
3627 3631 #endif // ASSERT
3628 3632
3629 3633 #if TASKQUEUE_STATS
3630 3634 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3631 3635 st->print_raw_cr("GC Task Stats");
3632 3636 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3633 3637 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3634 3638 }
3635 3639
3636 3640 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3637 3641 print_taskqueue_stats_hdr(st);
3638 3642
3639 3643 TaskQueueStats totals;
3640 3644 const int n = workers() != NULL ? workers()->total_workers() : 1;
3641 3645 for (int i = 0; i < n; ++i) {
3642 3646 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3643 3647 totals += task_queue(i)->stats;
3644 3648 }
3645 3649 st->print_raw("tot "); totals.print(st); st->cr();
3646 3650
3647 3651 DEBUG_ONLY(totals.verify());
3648 3652 }
3649 3653
3650 3654 void G1CollectedHeap::reset_taskqueue_stats() {
3651 3655 const int n = workers() != NULL ? workers()->total_workers() : 1;
3652 3656 for (int i = 0; i < n; ++i) {
3653 3657 task_queue(i)->stats.reset();
3654 3658 }
3655 3659 }
3656 3660 #endif // TASKQUEUE_STATS
3657 3661
3658 3662 void G1CollectedHeap::log_gc_header() {
3659 3663 if (!G1Log::fine()) {
3660 3664 return;
3661 3665 }
3662 3666
3663 3667 gclog_or_tty->date_stamp(PrintGCDateStamps);
3664 3668 gclog_or_tty->stamp(PrintGCTimeStamps);
3665 3669
3666 3670 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3667 3671 .append(g1_policy()->gcs_are_young() ? " (young)" : " (mixed)")
3668 3672 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3669 3673
3670 3674 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3671 3675 }
3672 3676
3673 3677 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3674 3678 if (!G1Log::fine()) {
3675 3679 return;
3676 3680 }
3677 3681
3678 3682 if (G1Log::finer()) {
3679 3683 if (evacuation_failed()) {
3680 3684 gclog_or_tty->print(" (to-space exhausted)");
3681 3685 }
3682 3686 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3683 3687 g1_policy()->phase_times()->note_gc_end();
3684 3688 g1_policy()->phase_times()->print(pause_time_sec);
3685 3689 g1_policy()->print_detailed_heap_transition();
3686 3690 } else {
3687 3691 if (evacuation_failed()) {
3688 3692 gclog_or_tty->print("--");
3689 3693 }
3690 3694 g1_policy()->print_heap_transition();
3691 3695 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3692 3696 }
3693 3697 }
3694 3698
3695 3699 bool
3696 3700 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3697 3701 assert_at_safepoint(true /* should_be_vm_thread */);
3698 3702 guarantee(!is_gc_active(), "collection is not reentrant");
3699 3703
3700 3704 if (GC_locker::check_active_before_gc()) {
3701 3705 return false;
3702 3706 }
3703 3707
3704 3708 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3705 3709 ResourceMark rm;
3706 3710
3707 3711 print_heap_before_gc();
3708 3712
3709 3713 HRSPhaseSetter x(HRSPhaseEvacuation);
3710 3714 verify_region_sets_optional();
3711 3715 verify_dirty_young_regions();
3712 3716
3713 3717 // This call will decide whether this pause is an initial-mark
3714 3718 // pause. If it is, during_initial_mark_pause() will return true
3715 3719 // for the duration of this pause.
3716 3720 g1_policy()->decide_on_conc_mark_initiation();
3717 3721
3718 3722 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3719 3723 assert(!g1_policy()->during_initial_mark_pause() ||
3720 3724 g1_policy()->gcs_are_young(), "sanity");
3721 3725
3722 3726 // We also do not allow mixed GCs during marking.
3723 3727 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3724 3728
3725 3729 // Record whether this pause is an initial mark. When the current
3726 3730 // thread has completed its logging output and it's safe to signal
3727 3731 // the CM thread, the flag's value in the policy has been reset.
3728 3732 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3729 3733
3730 3734 // Inner scope for scope based logging, timers, and stats collection
3731 3735 {
3732 3736 if (g1_policy()->during_initial_mark_pause()) {
3733 3737 // We are about to start a marking cycle, so we increment the
3734 3738 // full collection counter.
3735 3739 increment_old_marking_cycles_started();
3736 3740 }
3737 3741 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3738 3742
3739 3743 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3740 3744 workers()->active_workers() : 1);
3741 3745 double pause_start_sec = os::elapsedTime();
3742 3746 g1_policy()->phase_times()->note_gc_start(active_workers);
3743 3747 log_gc_header();
3744 3748
3745 3749 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3746 3750 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3747 3751
3748 3752 // If the secondary_free_list is not empty, append it to the
3749 3753 // free_list. No need to wait for the cleanup operation to finish;
3750 3754 // the region allocation code will check the secondary_free_list
3751 3755 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3752 3756 // set, skip this step so that the region allocation code has to
3753 3757 // get entries from the secondary_free_list.
3754 3758 if (!G1StressConcRegionFreeing) {
3755 3759 append_secondary_free_list_if_not_empty_with_lock();
3756 3760 }
3757 3761
3758 3762 assert(check_young_list_well_formed(),
3759 3763 "young list should be well formed");
3760 3764
3761 3765 // Don't dynamically change the number of GC threads this early. A value of
3762 3766 // 0 is used to indicate serial work. When parallel work is done,
3763 3767 // it will be set.
3764 3768
3765 3769 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3766 3770 IsGCActiveMark x;
3767 3771
3768 3772 gc_prologue(false);
3769 3773 increment_total_collections(false /* full gc */);
3770 3774 increment_gc_time_stamp();
3771 3775
3772 3776 verify_before_gc();
3773 3777
3774 3778 COMPILER2_PRESENT(DerivedPointerTable::clear());
3775 3779
3776 3780 // Please see comment in g1CollectedHeap.hpp and
3777 3781 // G1CollectedHeap::ref_processing_init() to see how
3778 3782 // reference processing currently works in G1.
3779 3783
3780 3784 // Enable discovery in the STW reference processor
3781 3785 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3782 3786 true /*verify_no_refs*/);
3783 3787
3784 3788 {
3785 3789 // We want to temporarily turn off discovery by the
3786 3790 // CM ref processor, if necessary, and turn it back on
3787 3791 // on again later if we do. Using a scoped
3788 3792 // NoRefDiscovery object will do this.
3789 3793 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3790 3794
3791 3795 // Forget the current alloc region (we might even choose it to be part
3792 3796 // of the collection set!).
3793 3797 release_mutator_alloc_region();
3794 3798
3795 3799 // We should call this after we retire the mutator alloc
3796 3800 // region(s) so that all the ALLOC / RETIRE events are generated
3797 3801 // before the start GC event.
3798 3802 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3799 3803
3800 3804 // This timing is only used by the ergonomics to handle our pause target.
3801 3805 // It is unclear why this should not include the full pause. We will
3802 3806 // investigate this in CR 7178365.
3803 3807 //
3804 3808 // Preserving the old comment here if that helps the investigation:
3805 3809 //
3806 3810 // The elapsed time induced by the start time below deliberately elides
3807 3811 // the possible verification above.
3808 3812 double sample_start_time_sec = os::elapsedTime();
3809 3813 size_t start_used_bytes = used();
3810 3814
3811 3815 #if YOUNG_LIST_VERBOSE
3812 3816 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3813 3817 _young_list->print();
3814 3818 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3815 3819 #endif // YOUNG_LIST_VERBOSE
3816 3820
3817 3821 g1_policy()->record_collection_pause_start(sample_start_time_sec,
3818 3822 start_used_bytes);
3819 3823
3820 3824 double scan_wait_start = os::elapsedTime();
3821 3825 // We have to wait until the CM threads finish scanning the
3822 3826 // root regions as it's the only way to ensure that all the
3823 3827 // objects on them have been correctly scanned before we start
3824 3828 // moving them during the GC.
3825 3829 bool waited = _cm->root_regions()->wait_until_scan_finished();
3826 3830 double wait_time_ms = 0.0;
3827 3831 if (waited) {
3828 3832 double scan_wait_end = os::elapsedTime();
3829 3833 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3830 3834 }
3831 3835 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3832 3836
3833 3837 #if YOUNG_LIST_VERBOSE
3834 3838 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3835 3839 _young_list->print();
3836 3840 #endif // YOUNG_LIST_VERBOSE
3837 3841
3838 3842 if (g1_policy()->during_initial_mark_pause()) {
3839 3843 concurrent_mark()->checkpointRootsInitialPre();
3840 3844 }
3841 3845
3842 3846 #if YOUNG_LIST_VERBOSE
3843 3847 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3844 3848 _young_list->print();
3845 3849 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3846 3850 #endif // YOUNG_LIST_VERBOSE
3847 3851
3848 3852 g1_policy()->finalize_cset(target_pause_time_ms);
3849 3853
3850 3854 _cm->note_start_of_gc();
3851 3855 // We should not verify the per-thread SATB buffers given that
3852 3856 // we have not filtered them yet (we'll do so during the
3853 3857 // GC). We also call this after finalize_cset() to
3854 3858 // ensure that the CSet has been finalized.
3855 3859 _cm->verify_no_cset_oops(true /* verify_stacks */,
3856 3860 true /* verify_enqueued_buffers */,
3857 3861 false /* verify_thread_buffers */,
3858 3862 true /* verify_fingers */);
3859 3863
3860 3864 if (_hr_printer.is_active()) {
3861 3865 HeapRegion* hr = g1_policy()->collection_set();
3862 3866 while (hr != NULL) {
3863 3867 G1HRPrinter::RegionType type;
3864 3868 if (!hr->is_young()) {
3865 3869 type = G1HRPrinter::Old;
3866 3870 } else if (hr->is_survivor()) {
3867 3871 type = G1HRPrinter::Survivor;
3868 3872 } else {
3869 3873 type = G1HRPrinter::Eden;
3870 3874 }
3871 3875 _hr_printer.cset(hr);
3872 3876 hr = hr->next_in_collection_set();
3873 3877 }
3874 3878 }
3875 3879
3876 3880 #ifdef ASSERT
3877 3881 VerifyCSetClosure cl;
3878 3882 collection_set_iterate(&cl);
3879 3883 #endif // ASSERT
3880 3884
3881 3885 setup_surviving_young_words();
3882 3886
3883 3887 // Initialize the GC alloc regions.
3884 3888 init_gc_alloc_regions();
3885 3889
3886 3890 // Actually do the work...
3887 3891 evacuate_collection_set();
3888 3892
3889 3893 // We do this to mainly verify the per-thread SATB buffers
3890 3894 // (which have been filtered by now) since we didn't verify
3891 3895 // them earlier. No point in re-checking the stacks / enqueued
3892 3896 // buffers given that the CSet has not changed since last time
3893 3897 // we checked.
3894 3898 _cm->verify_no_cset_oops(false /* verify_stacks */,
3895 3899 false /* verify_enqueued_buffers */,
3896 3900 true /* verify_thread_buffers */,
3897 3901 true /* verify_fingers */);
3898 3902
3899 3903 free_collection_set(g1_policy()->collection_set());
3900 3904 g1_policy()->clear_collection_set();
3901 3905
3902 3906 cleanup_surviving_young_words();
3903 3907
3904 3908 // Start a new incremental collection set for the next pause.
3905 3909 g1_policy()->start_incremental_cset_building();
3906 3910
3907 3911 // Clear the _cset_fast_test bitmap in anticipation of adding
3908 3912 // regions to the incremental collection set for the next
3909 3913 // evacuation pause.
3910 3914 clear_cset_fast_test();
3911 3915
3912 3916 _young_list->reset_sampled_info();
3913 3917
3914 3918 // Don't check the whole heap at this point as the
3915 3919 // GC alloc regions from this pause have been tagged
3916 3920 // as survivors and moved on to the survivor list.
3917 3921 // Survivor regions will fail the !is_young() check.
3918 3922 assert(check_young_list_empty(false /* check_heap */),
3919 3923 "young list should be empty");
3920 3924
3921 3925 #if YOUNG_LIST_VERBOSE
3922 3926 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3923 3927 _young_list->print();
3924 3928 #endif // YOUNG_LIST_VERBOSE
3925 3929
3926 3930 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3927 3931 _young_list->first_survivor_region(),
3928 3932 _young_list->last_survivor_region());
3929 3933
3930 3934 _young_list->reset_auxilary_lists();
3931 3935
3932 3936 if (evacuation_failed()) {
3933 3937 _summary_bytes_used = recalculate_used();
3934 3938 } else {
3935 3939 // The "used" of the the collection set have already been subtracted
3936 3940 // when they were freed. Add in the bytes evacuated.
3937 3941 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3938 3942 }
3939 3943
3940 3944 if (g1_policy()->during_initial_mark_pause()) {
3941 3945 // We have to do this before we notify the CM threads that
3942 3946 // they can start working to make sure that all the
3943 3947 // appropriate initialization is done on the CM object.
3944 3948 concurrent_mark()->checkpointRootsInitialPost();
3945 3949 set_marking_started();
3946 3950 // Note that we don't actually trigger the CM thread at
3947 3951 // this point. We do that later when we're sure that
3948 3952 // the current thread has completed its logging output.
3949 3953 }
3950 3954
3951 3955 allocate_dummy_regions();
3952 3956
3953 3957 #if YOUNG_LIST_VERBOSE
3954 3958 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3955 3959 _young_list->print();
3956 3960 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3957 3961 #endif // YOUNG_LIST_VERBOSE
3958 3962
3959 3963 init_mutator_alloc_region();
3960 3964
3961 3965 {
3962 3966 size_t expand_bytes = g1_policy()->expansion_amount();
3963 3967 if (expand_bytes > 0) {
3964 3968 size_t bytes_before = capacity();
3965 3969 // No need for an ergo verbose message here,
3966 3970 // expansion_amount() does this when it returns a value > 0.
3967 3971 if (!expand(expand_bytes)) {
3968 3972 // We failed to expand the heap so let's verify that
3969 3973 // committed/uncommitted amount match the backing store
3970 3974 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3971 3975 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3972 3976 }
3973 3977 }
3974 3978 }
3975 3979
3976 3980 // We redo the verificaiton but now wrt to the new CSet which
3977 3981 // has just got initialized after the previous CSet was freed.
3978 3982 _cm->verify_no_cset_oops(true /* verify_stacks */,
3979 3983 true /* verify_enqueued_buffers */,
3980 3984 true /* verify_thread_buffers */,
3981 3985 true /* verify_fingers */);
3982 3986 _cm->note_end_of_gc();
3983 3987
3984 3988 // This timing is only used by the ergonomics to handle our pause target.
3985 3989 // It is unclear why this should not include the full pause. We will
3986 3990 // investigate this in CR 7178365.
3987 3991 double sample_end_time_sec = os::elapsedTime();
3988 3992 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
3989 3993 g1_policy()->record_collection_pause_end(pause_time_ms);
3990 3994
3991 3995 MemoryService::track_memory_usage();
3992 3996
3993 3997 // In prepare_for_verify() below we'll need to scan the deferred
3994 3998 // update buffers to bring the RSets up-to-date if
3995 3999 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3996 4000 // the update buffers we'll probably need to scan cards on the
3997 4001 // regions we just allocated to (i.e., the GC alloc
3998 4002 // regions). However, during the last GC we called
3999 4003 // set_saved_mark() on all the GC alloc regions, so card
4000 4004 // scanning might skip the [saved_mark_word()...top()] area of
4001 4005 // those regions (i.e., the area we allocated objects into
4002 4006 // during the last GC). But it shouldn't. Given that
4003 4007 // saved_mark_word() is conditional on whether the GC time stamp
4004 4008 // on the region is current or not, by incrementing the GC time
4005 4009 // stamp here we invalidate all the GC time stamps on all the
4006 4010 // regions and saved_mark_word() will simply return top() for
4007 4011 // all the regions. This is a nicer way of ensuring this rather
4008 4012 // than iterating over the regions and fixing them. In fact, the
4009 4013 // GC time stamp increment here also ensures that
4010 4014 // saved_mark_word() will return top() between pauses, i.e.,
4011 4015 // during concurrent refinement. So we don't need the
4012 4016 // is_gc_active() check to decided which top to use when
4013 4017 // scanning cards (see CR 7039627).
4014 4018 increment_gc_time_stamp();
4015 4019
4016 4020 verify_after_gc();
4017 4021
4018 4022 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4019 4023 ref_processor_stw()->verify_no_references_recorded();
4020 4024
4021 4025 // CM reference discovery will be re-enabled if necessary.
4022 4026 }
4023 4027
4024 4028 // We should do this after we potentially expand the heap so
4025 4029 // that all the COMMIT events are generated before the end GC
4026 4030 // event, and after we retire the GC alloc regions so that all
4027 4031 // RETIRE events are generated before the end GC event.
4028 4032 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4029 4033
4030 4034 if (mark_in_progress()) {
4031 4035 concurrent_mark()->update_g1_committed();
4032 4036 }
4033 4037
4034 4038 #ifdef TRACESPINNING
4035 4039 ParallelTaskTerminator::print_termination_counts();
4036 4040 #endif
4037 4041
4038 4042 gc_epilogue(false);
4039 4043
4040 4044 log_gc_footer(os::elapsedTime() - pause_start_sec);
4041 4045 }
4042 4046
4043 4047 // It is not yet to safe to tell the concurrent mark to
4044 4048 // start as we have some optional output below. We don't want the
4045 4049 // output from the concurrent mark thread interfering with this
4046 4050 // logging output either.
4047 4051
4048 4052 _hrs.verify_optional();
4049 4053 verify_region_sets_optional();
4050 4054
4051 4055 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4052 4056 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4053 4057
4054 4058 print_heap_after_gc();
4055 4059
4056 4060 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4057 4061 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4058 4062 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4059 4063 // before any GC notifications are raised.
4060 4064 g1mm()->update_sizes();
4061 4065 }
4062 4066
4063 4067 if (G1SummarizeRSetStats &&
4064 4068 (G1SummarizeRSetStatsPeriod > 0) &&
4065 4069 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4066 4070 g1_rem_set()->print_summary_info();
4067 4071 }
4068 4072
4069 4073 // It should now be safe to tell the concurrent mark thread to start
4070 4074 // without its logging output interfering with the logging output
4071 4075 // that came from the pause.
4072 4076
4073 4077 if (should_start_conc_mark) {
4074 4078 // CAUTION: after the doConcurrentMark() call below,
4075 4079 // the concurrent marking thread(s) could be running
4076 4080 // concurrently with us. Make sure that anything after
4077 4081 // this point does not assume that we are the only GC thread
4078 4082 // running. Note: of course, the actual marking work will
4079 4083 // not start until the safepoint itself is released in
4080 4084 // ConcurrentGCThread::safepoint_desynchronize().
4081 4085 doConcurrentMark();
4082 4086 }
4083 4087
4084 4088 return true;
4085 4089 }
4086 4090
4087 4091 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4088 4092 {
4089 4093 size_t gclab_word_size;
4090 4094 switch (purpose) {
4091 4095 case GCAllocForSurvived:
4092 4096 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4093 4097 break;
4094 4098 case GCAllocForTenured:
4095 4099 gclab_word_size = _old_plab_stats.desired_plab_sz();
4096 4100 break;
4097 4101 default:
4098 4102 assert(false, "unknown GCAllocPurpose");
4099 4103 gclab_word_size = _old_plab_stats.desired_plab_sz();
4100 4104 break;
4101 4105 }
4102 4106
4103 4107 // Prevent humongous PLAB sizes for two reasons:
4104 4108 // * PLABs are allocated using a similar paths as oops, but should
4105 4109 // never be in a humongous region
4106 4110 // * Allowing humongous PLABs needlessly churns the region free lists
4107 4111 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4108 4112 }
4109 4113
4110 4114 void G1CollectedHeap::init_mutator_alloc_region() {
4111 4115 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4112 4116 _mutator_alloc_region.init();
4113 4117 }
4114 4118
4115 4119 void G1CollectedHeap::release_mutator_alloc_region() {
4116 4120 _mutator_alloc_region.release();
4117 4121 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4118 4122 }
4119 4123
4120 4124 void G1CollectedHeap::init_gc_alloc_regions() {
4121 4125 assert_at_safepoint(true /* should_be_vm_thread */);
4122 4126
4123 4127 _survivor_gc_alloc_region.init();
4124 4128 _old_gc_alloc_region.init();
4125 4129 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4126 4130 _retained_old_gc_alloc_region = NULL;
4127 4131
4128 4132 // We will discard the current GC alloc region if:
4129 4133 // a) it's in the collection set (it can happen!),
4130 4134 // b) it's already full (no point in using it),
4131 4135 // c) it's empty (this means that it was emptied during
4132 4136 // a cleanup and it should be on the free list now), or
4133 4137 // d) it's humongous (this means that it was emptied
4134 4138 // during a cleanup and was added to the free list, but
4135 4139 // has been subseqently used to allocate a humongous
4136 4140 // object that may be less than the region size).
4137 4141 if (retained_region != NULL &&
4138 4142 !retained_region->in_collection_set() &&
4139 4143 !(retained_region->top() == retained_region->end()) &&
4140 4144 !retained_region->is_empty() &&
4141 4145 !retained_region->isHumongous()) {
4142 4146 retained_region->set_saved_mark();
4143 4147 // The retained region was added to the old region set when it was
4144 4148 // retired. We have to remove it now, since we don't allow regions
4145 4149 // we allocate to in the region sets. We'll re-add it later, when
4146 4150 // it's retired again.
4147 4151 _old_set.remove(retained_region);
4148 4152 bool during_im = g1_policy()->during_initial_mark_pause();
4149 4153 retained_region->note_start_of_copying(during_im);
4150 4154 _old_gc_alloc_region.set(retained_region);
4151 4155 _hr_printer.reuse(retained_region);
4152 4156 }
4153 4157 }
4154 4158
4155 4159 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4156 4160 _survivor_gc_alloc_region.release();
4157 4161 // If we have an old GC alloc region to release, we'll save it in
4158 4162 // _retained_old_gc_alloc_region. If we don't
4159 4163 // _retained_old_gc_alloc_region will become NULL. This is what we
4160 4164 // want either way so no reason to check explicitly for either
4161 4165 // condition.
4162 4166 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4163 4167
4164 4168 if (ResizePLAB) {
4165 4169 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4166 4170 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4167 4171 }
4168 4172 }
4169 4173
4170 4174 void G1CollectedHeap::abandon_gc_alloc_regions() {
4171 4175 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4172 4176 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4173 4177 _retained_old_gc_alloc_region = NULL;
4174 4178 }
4175 4179
4176 4180 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4177 4181 _drain_in_progress = false;
4178 4182 set_evac_failure_closure(cl);
4179 4183 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4180 4184 }
4181 4185
4182 4186 void G1CollectedHeap::finalize_for_evac_failure() {
4183 4187 assert(_evac_failure_scan_stack != NULL &&
4184 4188 _evac_failure_scan_stack->length() == 0,
4185 4189 "Postcondition");
4186 4190 assert(!_drain_in_progress, "Postcondition");
4187 4191 delete _evac_failure_scan_stack;
4188 4192 _evac_failure_scan_stack = NULL;
4189 4193 }
4190 4194
4191 4195 void G1CollectedHeap::remove_self_forwarding_pointers() {
4192 4196 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4193 4197
4194 4198 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4195 4199
4196 4200 if (G1CollectedHeap::use_parallel_gc_threads()) {
4197 4201 set_par_threads();
4198 4202 workers()->run_task(&rsfp_task);
4199 4203 set_par_threads(0);
4200 4204 } else {
4201 4205 rsfp_task.work(0);
4202 4206 }
4203 4207
4204 4208 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4205 4209
4206 4210 // Reset the claim values in the regions in the collection set.
4207 4211 reset_cset_heap_region_claim_values();
4208 4212
4209 4213 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4210 4214
4211 4215 // Now restore saved marks, if any.
4212 4216 if (_objs_with_preserved_marks != NULL) {
4213 4217 assert(_preserved_marks_of_objs != NULL, "Both or none.");
4214 4218 guarantee(_objs_with_preserved_marks->length() ==
4215 4219 _preserved_marks_of_objs->length(), "Both or none.");
4216 4220 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
4217 4221 oop obj = _objs_with_preserved_marks->at(i);
4218 4222 markOop m = _preserved_marks_of_objs->at(i);
4219 4223 obj->set_mark(m);
4220 4224 }
4221 4225
4222 4226 // Delete the preserved marks growable arrays (allocated on the C heap).
4223 4227 delete _objs_with_preserved_marks;
4224 4228 delete _preserved_marks_of_objs;
4225 4229 _objs_with_preserved_marks = NULL;
4226 4230 _preserved_marks_of_objs = NULL;
4227 4231 }
4228 4232 }
4229 4233
4230 4234 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4231 4235 _evac_failure_scan_stack->push(obj);
4232 4236 }
4233 4237
4234 4238 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4235 4239 assert(_evac_failure_scan_stack != NULL, "precondition");
4236 4240
4237 4241 while (_evac_failure_scan_stack->length() > 0) {
4238 4242 oop obj = _evac_failure_scan_stack->pop();
4239 4243 _evac_failure_closure->set_region(heap_region_containing(obj));
4240 4244 obj->oop_iterate_backwards(_evac_failure_closure);
4241 4245 }
4242 4246 }
4243 4247
4244 4248 oop
4245 4249 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4246 4250 oop old) {
4247 4251 assert(obj_in_cs(old),
4248 4252 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4249 4253 (HeapWord*) old));
4250 4254 markOop m = old->mark();
4251 4255 oop forward_ptr = old->forward_to_atomic(old);
4252 4256 if (forward_ptr == NULL) {
4253 4257 // Forward-to-self succeeded.
4254 4258
4255 4259 if (_evac_failure_closure != cl) {
4256 4260 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4257 4261 assert(!_drain_in_progress,
4258 4262 "Should only be true while someone holds the lock.");
4259 4263 // Set the global evac-failure closure to the current thread's.
4260 4264 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4261 4265 set_evac_failure_closure(cl);
4262 4266 // Now do the common part.
4263 4267 handle_evacuation_failure_common(old, m);
4264 4268 // Reset to NULL.
4265 4269 set_evac_failure_closure(NULL);
4266 4270 } else {
4267 4271 // The lock is already held, and this is recursive.
4268 4272 assert(_drain_in_progress, "This should only be the recursive case.");
4269 4273 handle_evacuation_failure_common(old, m);
4270 4274 }
4271 4275 return old;
4272 4276 } else {
4273 4277 // Forward-to-self failed. Either someone else managed to allocate
4274 4278 // space for this object (old != forward_ptr) or they beat us in
4275 4279 // self-forwarding it (old == forward_ptr).
4276 4280 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4277 4281 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4278 4282 "should not be in the CSet",
4279 4283 (HeapWord*) old, (HeapWord*) forward_ptr));
4280 4284 return forward_ptr;
4281 4285 }
4282 4286 }
4283 4287
4284 4288 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4285 4289 set_evacuation_failed(true);
4286 4290
4287 4291 preserve_mark_if_necessary(old, m);
4288 4292
4289 4293 HeapRegion* r = heap_region_containing(old);
4290 4294 if (!r->evacuation_failed()) {
4291 4295 r->set_evacuation_failed(true);
4292 4296 _hr_printer.evac_failure(r);
4293 4297 }
4294 4298
4295 4299 push_on_evac_failure_scan_stack(old);
4296 4300
4297 4301 if (!_drain_in_progress) {
4298 4302 // prevent recursion in copy_to_survivor_space()
4299 4303 _drain_in_progress = true;
4300 4304 drain_evac_failure_scan_stack();
4301 4305 _drain_in_progress = false;
4302 4306 }
4303 4307 }
4304 4308
4305 4309 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4306 4310 assert(evacuation_failed(), "Oversaving!");
4307 4311 // We want to call the "for_promotion_failure" version only in the
4308 4312 // case of a promotion failure.
4309 4313 if (m->must_be_preserved_for_promotion_failure(obj)) {
4310 4314 if (_objs_with_preserved_marks == NULL) {
4311 4315 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4312 4316 _objs_with_preserved_marks =
4313 4317 new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4314 4318 _preserved_marks_of_objs =
4315 4319 new (ResourceObj::C_HEAP, mtGC) GrowableArray<markOop>(40, true);
4316 4320 }
4317 4321 _objs_with_preserved_marks->push(obj);
4318 4322 _preserved_marks_of_objs->push(m);
4319 4323 }
4320 4324 }
4321 4325
4322 4326 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4323 4327 size_t word_size) {
4324 4328 if (purpose == GCAllocForSurvived) {
4325 4329 HeapWord* result = survivor_attempt_allocation(word_size);
4326 4330 if (result != NULL) {
4327 4331 return result;
4328 4332 } else {
4329 4333 // Let's try to allocate in the old gen in case we can fit the
4330 4334 // object there.
4331 4335 return old_attempt_allocation(word_size);
4332 4336 }
4333 4337 } else {
4334 4338 assert(purpose == GCAllocForTenured, "sanity");
4335 4339 HeapWord* result = old_attempt_allocation(word_size);
4336 4340 if (result != NULL) {
4337 4341 return result;
4338 4342 } else {
4339 4343 // Let's try to allocate in the survivors in case we can fit the
4340 4344 // object there.
4341 4345 return survivor_attempt_allocation(word_size);
4342 4346 }
4343 4347 }
4344 4348
4345 4349 ShouldNotReachHere();
4346 4350 // Trying to keep some compilers happy.
4347 4351 return NULL;
4348 4352 }
4349 4353
4350 4354 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4351 4355 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4352 4356
4353 4357 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4354 4358 : _g1h(g1h),
4355 4359 _refs(g1h->task_queue(queue_num)),
4356 4360 _dcq(&g1h->dirty_card_queue_set()),
4357 4361 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4358 4362 _g1_rem(g1h->g1_rem_set()),
4359 4363 _hash_seed(17), _queue_num(queue_num),
4360 4364 _term_attempts(0),
4361 4365 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4362 4366 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4363 4367 _age_table(false),
4364 4368 _strong_roots_time(0), _term_time(0),
4365 4369 _alloc_buffer_waste(0), _undo_waste(0) {
4366 4370 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4367 4371 // we "sacrifice" entry 0 to keep track of surviving bytes for
4368 4372 // non-young regions (where the age is -1)
4369 4373 // We also add a few elements at the beginning and at the end in
4370 4374 // an attempt to eliminate cache contention
4371 4375 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4372 4376 uint array_length = PADDING_ELEM_NUM +
4373 4377 real_length +
4374 4378 PADDING_ELEM_NUM;
4375 4379 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4376 4380 if (_surviving_young_words_base == NULL)
4377 4381 vm_exit_out_of_memory(array_length * sizeof(size_t),
4378 4382 "Not enough space for young surv histo.");
4379 4383 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4380 4384 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4381 4385
4382 4386 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4383 4387 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4384 4388
4385 4389 _start = os::elapsedTime();
4386 4390 }
4387 4391
4388 4392 void
4389 4393 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4390 4394 {
4391 4395 st->print_raw_cr("GC Termination Stats");
4392 4396 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4393 4397 " ------waste (KiB)------");
4394 4398 st->print_raw_cr("thr ms ms % ms % attempts"
4395 4399 " total alloc undo");
4396 4400 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4397 4401 " ------- ------- -------");
4398 4402 }
4399 4403
4400 4404 void
4401 4405 G1ParScanThreadState::print_termination_stats(int i,
4402 4406 outputStream* const st) const
4403 4407 {
4404 4408 const double elapsed_ms = elapsed_time() * 1000.0;
4405 4409 const double s_roots_ms = strong_roots_time() * 1000.0;
4406 4410 const double term_ms = term_time() * 1000.0;
4407 4411 st->print_cr("%3d %9.2f %9.2f %6.2f "
4408 4412 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4409 4413 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4410 4414 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4411 4415 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4412 4416 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4413 4417 alloc_buffer_waste() * HeapWordSize / K,
4414 4418 undo_waste() * HeapWordSize / K);
4415 4419 }
4416 4420
4417 4421 #ifdef ASSERT
4418 4422 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4419 4423 assert(ref != NULL, "invariant");
4420 4424 assert(UseCompressedOops, "sanity");
4421 4425 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4422 4426 oop p = oopDesc::load_decode_heap_oop(ref);
4423 4427 assert(_g1h->is_in_g1_reserved(p),
4424 4428 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4425 4429 return true;
4426 4430 }
4427 4431
4428 4432 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4429 4433 assert(ref != NULL, "invariant");
4430 4434 if (has_partial_array_mask(ref)) {
4431 4435 // Must be in the collection set--it's already been copied.
4432 4436 oop p = clear_partial_array_mask(ref);
4433 4437 assert(_g1h->obj_in_cs(p),
4434 4438 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4435 4439 } else {
4436 4440 oop p = oopDesc::load_decode_heap_oop(ref);
4437 4441 assert(_g1h->is_in_g1_reserved(p),
4438 4442 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4439 4443 }
4440 4444 return true;
4441 4445 }
4442 4446
4443 4447 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4444 4448 if (ref.is_narrow()) {
4445 4449 return verify_ref((narrowOop*) ref);
4446 4450 } else {
4447 4451 return verify_ref((oop*) ref);
4448 4452 }
4449 4453 }
4450 4454 #endif // ASSERT
4451 4455
4452 4456 void G1ParScanThreadState::trim_queue() {
4453 4457 assert(_evac_cl != NULL, "not set");
4454 4458 assert(_evac_failure_cl != NULL, "not set");
4455 4459 assert(_partial_scan_cl != NULL, "not set");
4456 4460
4457 4461 StarTask ref;
4458 4462 do {
4459 4463 // Drain the overflow stack first, so other threads can steal.
4460 4464 while (refs()->pop_overflow(ref)) {
4461 4465 deal_with_reference(ref);
4462 4466 }
4463 4467
4464 4468 while (refs()->pop_local(ref)) {
4465 4469 deal_with_reference(ref);
4466 4470 }
4467 4471 } while (!refs()->is_empty());
4468 4472 }
4469 4473
4470 4474 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4471 4475 G1ParScanThreadState* par_scan_state) :
4472 4476 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4473 4477 _par_scan_state(par_scan_state),
4474 4478 _worker_id(par_scan_state->queue_num()),
4475 4479 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4476 4480 _mark_in_progress(_g1->mark_in_progress()) { }
4477 4481
4478 4482 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4479 4483 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4480 4484 #ifdef ASSERT
4481 4485 HeapRegion* hr = _g1->heap_region_containing(obj);
4482 4486 assert(hr != NULL, "sanity");
4483 4487 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4484 4488 #endif // ASSERT
4485 4489
4486 4490 // We know that the object is not moving so it's safe to read its size.
4487 4491 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4488 4492 }
4489 4493
4490 4494 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4491 4495 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4492 4496 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4493 4497 #ifdef ASSERT
4494 4498 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4495 4499 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4496 4500 assert(from_obj != to_obj, "should not be self-forwarded");
4497 4501
4498 4502 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4499 4503 assert(from_hr != NULL, "sanity");
4500 4504 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4501 4505
4502 4506 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4503 4507 assert(to_hr != NULL, "sanity");
4504 4508 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4505 4509 #endif // ASSERT
4506 4510
4507 4511 // The object might be in the process of being copied by another
4508 4512 // worker so we cannot trust that its to-space image is
4509 4513 // well-formed. So we have to read its size from its from-space
4510 4514 // image which we know should not be changing.
4511 4515 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4512 4516 }
4513 4517
4514 4518 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4515 4519 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4516 4520 ::copy_to_survivor_space(oop old) {
4517 4521 size_t word_sz = old->size();
4518 4522 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4519 4523 // +1 to make the -1 indexes valid...
4520 4524 int young_index = from_region->young_index_in_cset()+1;
4521 4525 assert( (from_region->is_young() && young_index > 0) ||
4522 4526 (!from_region->is_young() && young_index == 0), "invariant" );
4523 4527 G1CollectorPolicy* g1p = _g1->g1_policy();
4524 4528 markOop m = old->mark();
4525 4529 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4526 4530 : m->age();
4527 4531 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4528 4532 word_sz);
4529 4533 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4530 4534 #ifndef PRODUCT
4531 4535 // Should this evacuation fail?
4532 4536 if (_g1->evacuation_should_fail()) {
4533 4537 if (obj_ptr != NULL) {
4534 4538 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4535 4539 obj_ptr = NULL;
4536 4540 }
4537 4541 }
4538 4542 #endif // !PRODUCT
4539 4543
4540 4544 if (obj_ptr == NULL) {
4541 4545 // This will either forward-to-self, or detect that someone else has
4542 4546 // installed a forwarding pointer.
4543 4547 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4544 4548 return _g1->handle_evacuation_failure_par(cl, old);
4545 4549 }
4546 4550
4547 4551 oop obj = oop(obj_ptr);
4548 4552
4549 4553 // We're going to allocate linearly, so might as well prefetch ahead.
4550 4554 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4551 4555
4552 4556 oop forward_ptr = old->forward_to_atomic(obj);
4553 4557 if (forward_ptr == NULL) {
4554 4558 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4555 4559 if (g1p->track_object_age(alloc_purpose)) {
4556 4560 // We could simply do obj->incr_age(). However, this causes a
4557 4561 // performance issue. obj->incr_age() will first check whether
4558 4562 // the object has a displaced mark by checking its mark word;
4559 4563 // getting the mark word from the new location of the object
4560 4564 // stalls. So, given that we already have the mark word and we
4561 4565 // are about to install it anyway, it's better to increase the
4562 4566 // age on the mark word, when the object does not have a
4563 4567 // displaced mark word. We're not expecting many objects to have
4564 4568 // a displaced marked word, so that case is not optimized
4565 4569 // further (it could be...) and we simply call obj->incr_age().
4566 4570
4567 4571 if (m->has_displaced_mark_helper()) {
4568 4572 // in this case, we have to install the mark word first,
4569 4573 // otherwise obj looks to be forwarded (the old mark word,
4570 4574 // which contains the forward pointer, was copied)
4571 4575 obj->set_mark(m);
4572 4576 obj->incr_age();
4573 4577 } else {
4574 4578 m = m->incr_age();
4575 4579 obj->set_mark(m);
4576 4580 }
4577 4581 _par_scan_state->age_table()->add(obj, word_sz);
4578 4582 } else {
4579 4583 obj->set_mark(m);
4580 4584 }
4581 4585
4582 4586 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4583 4587 surv_young_words[young_index] += word_sz;
4584 4588
4585 4589 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4586 4590 // We keep track of the next start index in the length field of
4587 4591 // the to-space object. The actual length can be found in the
4588 4592 // length field of the from-space object.
4589 4593 arrayOop(obj)->set_length(0);
4590 4594 oop* old_p = set_partial_array_mask(old);
4591 4595 _par_scan_state->push_on_queue(old_p);
4592 4596 } else {
4593 4597 // No point in using the slower heap_region_containing() method,
4594 4598 // given that we know obj is in the heap.
4595 4599 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4596 4600 obj->oop_iterate_backwards(&_scanner);
4597 4601 }
4598 4602 } else {
4599 4603 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4600 4604 obj = forward_ptr;
4601 4605 }
4602 4606 return obj;
4603 4607 }
4604 4608
4605 4609 template <class T>
4606 4610 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4607 4611 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4608 4612 _scanned_klass->record_modified_oops();
4609 4613 }
4610 4614 }
4611 4615
4612 4616 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4613 4617 template <class T>
4614 4618 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4615 4619 ::do_oop_work(T* p) {
4616 4620 oop obj = oopDesc::load_decode_heap_oop(p);
4617 4621 assert(barrier != G1BarrierRS || obj != NULL,
4618 4622 "Precondition: G1BarrierRS implies obj is non-NULL");
4619 4623
4620 4624 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4621 4625
4622 4626 // here the null check is implicit in the cset_fast_test() test
4623 4627 if (_g1->in_cset_fast_test(obj)) {
4624 4628 oop forwardee;
4625 4629 if (obj->is_forwarded()) {
4626 4630 forwardee = obj->forwardee();
4627 4631 } else {
4628 4632 forwardee = copy_to_survivor_space(obj);
4629 4633 }
4630 4634 assert(forwardee != NULL, "forwardee should not be NULL");
4631 4635 oopDesc::encode_store_heap_oop(p, forwardee);
4632 4636 if (do_mark_object && forwardee != obj) {
4633 4637 // If the object is self-forwarded we don't need to explicitly
4634 4638 // mark it, the evacuation failure protocol will do so.
4635 4639 mark_forwarded_object(obj, forwardee);
4636 4640 }
4637 4641
4638 4642 // When scanning the RS, we only care about objs in CS.
4639 4643 if (barrier == G1BarrierRS) {
4640 4644 _par_scan_state->update_rs(_from, p, _worker_id);
4641 4645 } else if (barrier == G1BarrierKlass) {
4642 4646 do_klass_barrier(p, forwardee);
4643 4647 }
4644 4648 } else {
4645 4649 // The object is not in collection set. If we're a root scanning
4646 4650 // closure during an initial mark pause (i.e. do_mark_object will
4647 4651 // be true) then attempt to mark the object.
4648 4652 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4649 4653 mark_object(obj);
4650 4654 }
4651 4655 }
4652 4656
4653 4657 if (barrier == G1BarrierEvac && obj != NULL) {
4654 4658 _par_scan_state->update_rs(_from, p, _worker_id);
4655 4659 }
4656 4660
4657 4661 if (do_gen_barrier && obj != NULL) {
4658 4662 par_do_barrier(p);
4659 4663 }
4660 4664 }
4661 4665
4662 4666 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4663 4667 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4664 4668
4665 4669 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4666 4670 assert(has_partial_array_mask(p), "invariant");
4667 4671 oop from_obj = clear_partial_array_mask(p);
4668 4672
4669 4673 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4670 4674 assert(from_obj->is_objArray(), "must be obj array");
4671 4675 objArrayOop from_obj_array = objArrayOop(from_obj);
4672 4676 // The from-space object contains the real length.
4673 4677 int length = from_obj_array->length();
4674 4678
4675 4679 assert(from_obj->is_forwarded(), "must be forwarded");
4676 4680 oop to_obj = from_obj->forwardee();
4677 4681 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4678 4682 objArrayOop to_obj_array = objArrayOop(to_obj);
4679 4683 // We keep track of the next start index in the length field of the
4680 4684 // to-space object.
4681 4685 int next_index = to_obj_array->length();
4682 4686 assert(0 <= next_index && next_index < length,
4683 4687 err_msg("invariant, next index: %d, length: %d", next_index, length));
4684 4688
4685 4689 int start = next_index;
4686 4690 int end = length;
4687 4691 int remainder = end - start;
4688 4692 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4689 4693 if (remainder > 2 * ParGCArrayScanChunk) {
4690 4694 end = start + ParGCArrayScanChunk;
4691 4695 to_obj_array->set_length(end);
4692 4696 // Push the remainder before we process the range in case another
4693 4697 // worker has run out of things to do and can steal it.
4694 4698 oop* from_obj_p = set_partial_array_mask(from_obj);
4695 4699 _par_scan_state->push_on_queue(from_obj_p);
4696 4700 } else {
4697 4701 assert(length == end, "sanity");
4698 4702 // We'll process the final range for this object. Restore the length
4699 4703 // so that the heap remains parsable in case of evacuation failure.
4700 4704 to_obj_array->set_length(end);
4701 4705 }
4702 4706 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4703 4707 // Process indexes [start,end). It will also process the header
4704 4708 // along with the first chunk (i.e., the chunk with start == 0).
4705 4709 // Note that at this point the length field of to_obj_array is not
4706 4710 // correct given that we are using it to keep track of the next
4707 4711 // start index. oop_iterate_range() (thankfully!) ignores the length
4708 4712 // field and only relies on the start / end parameters. It does
4709 4713 // however return the size of the object which will be incorrect. So
4710 4714 // we have to ignore it even if we wanted to use it.
4711 4715 to_obj_array->oop_iterate_range(&_scanner, start, end);
4712 4716 }
4713 4717
4714 4718 class G1ParEvacuateFollowersClosure : public VoidClosure {
4715 4719 protected:
4716 4720 G1CollectedHeap* _g1h;
4717 4721 G1ParScanThreadState* _par_scan_state;
4718 4722 RefToScanQueueSet* _queues;
4719 4723 ParallelTaskTerminator* _terminator;
4720 4724
4721 4725 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4722 4726 RefToScanQueueSet* queues() { return _queues; }
4723 4727 ParallelTaskTerminator* terminator() { return _terminator; }
4724 4728
4725 4729 public:
4726 4730 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4727 4731 G1ParScanThreadState* par_scan_state,
4728 4732 RefToScanQueueSet* queues,
4729 4733 ParallelTaskTerminator* terminator)
4730 4734 : _g1h(g1h), _par_scan_state(par_scan_state),
4731 4735 _queues(queues), _terminator(terminator) {}
4732 4736
4733 4737 void do_void();
4734 4738
4735 4739 private:
4736 4740 inline bool offer_termination();
4737 4741 };
4738 4742
4739 4743 bool G1ParEvacuateFollowersClosure::offer_termination() {
4740 4744 G1ParScanThreadState* const pss = par_scan_state();
4741 4745 pss->start_term_time();
4742 4746 const bool res = terminator()->offer_termination();
4743 4747 pss->end_term_time();
4744 4748 return res;
4745 4749 }
4746 4750
4747 4751 void G1ParEvacuateFollowersClosure::do_void() {
4748 4752 StarTask stolen_task;
4749 4753 G1ParScanThreadState* const pss = par_scan_state();
4750 4754 pss->trim_queue();
4751 4755
4752 4756 do {
4753 4757 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4754 4758 assert(pss->verify_task(stolen_task), "sanity");
4755 4759 if (stolen_task.is_narrow()) {
4756 4760 pss->deal_with_reference((narrowOop*) stolen_task);
4757 4761 } else {
4758 4762 pss->deal_with_reference((oop*) stolen_task);
4759 4763 }
4760 4764
4761 4765 // We've just processed a reference and we might have made
4762 4766 // available new entries on the queues. So we have to make sure
4763 4767 // we drain the queues as necessary.
4764 4768 pss->trim_queue();
4765 4769 }
4766 4770 } while (!offer_termination());
4767 4771
4768 4772 pss->retire_alloc_buffers();
4769 4773 }
4770 4774
4771 4775 class G1KlassScanClosure : public KlassClosure {
4772 4776 G1ParCopyHelper* _closure;
4773 4777 bool _process_only_dirty;
4774 4778 int _count;
4775 4779 public:
4776 4780 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4777 4781 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4778 4782 void do_klass(Klass* klass) {
4779 4783 // If the klass has not been dirtied we know that there's
4780 4784 // no references into the young gen and we can skip it.
4781 4785 if (!_process_only_dirty || klass->has_modified_oops()) {
4782 4786 // Clean the klass since we're going to scavenge all the metadata.
4783 4787 klass->clear_modified_oops();
4784 4788
4785 4789 // Tell the closure that this klass is the Klass to scavenge
4786 4790 // and is the one to dirty if oops are left pointing into the young gen.
4787 4791 _closure->set_scanned_klass(klass);
4788 4792
4789 4793 klass->oops_do(_closure);
4790 4794
4791 4795 _closure->set_scanned_klass(NULL);
4792 4796 }
4793 4797 _count++;
4794 4798 }
4795 4799 };
4796 4800
4797 4801 class G1ParTask : public AbstractGangTask {
4798 4802 protected:
4799 4803 G1CollectedHeap* _g1h;
4800 4804 RefToScanQueueSet *_queues;
4801 4805 ParallelTaskTerminator _terminator;
4802 4806 uint _n_workers;
4803 4807
4804 4808 Mutex _stats_lock;
4805 4809 Mutex* stats_lock() { return &_stats_lock; }
4806 4810
4807 4811 size_t getNCards() {
4808 4812 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4809 4813 / G1BlockOffsetSharedArray::N_bytes;
4810 4814 }
4811 4815
4812 4816 public:
4813 4817 G1ParTask(G1CollectedHeap* g1h,
4814 4818 RefToScanQueueSet *task_queues)
4815 4819 : AbstractGangTask("G1 collection"),
4816 4820 _g1h(g1h),
4817 4821 _queues(task_queues),
4818 4822 _terminator(0, _queues),
4819 4823 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4820 4824 {}
4821 4825
4822 4826 RefToScanQueueSet* queues() { return _queues; }
4823 4827
4824 4828 RefToScanQueue *work_queue(int i) {
4825 4829 return queues()->queue(i);
4826 4830 }
4827 4831
4828 4832 ParallelTaskTerminator* terminator() { return &_terminator; }
4829 4833
4830 4834 virtual void set_for_termination(int active_workers) {
4831 4835 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4832 4836 // in the young space (_par_seq_tasks) in the G1 heap
4833 4837 // for SequentialSubTasksDone.
4834 4838 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4835 4839 // both of which need setting by set_n_termination().
4836 4840 _g1h->SharedHeap::set_n_termination(active_workers);
4837 4841 _g1h->set_n_termination(active_workers);
4838 4842 terminator()->reset_for_reuse(active_workers);
4839 4843 _n_workers = active_workers;
4840 4844 }
4841 4845
4842 4846 void work(uint worker_id) {
4843 4847 if (worker_id >= _n_workers) return; // no work needed this round
4844 4848
4845 4849 double start_time_ms = os::elapsedTime() * 1000.0;
4846 4850 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4847 4851
4848 4852 {
4849 4853 ResourceMark rm;
4850 4854 HandleMark hm;
4851 4855
4852 4856 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4853 4857
4854 4858 G1ParScanThreadState pss(_g1h, worker_id);
4855 4859 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4856 4860 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4857 4861 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4858 4862
4859 4863 pss.set_evac_closure(&scan_evac_cl);
4860 4864 pss.set_evac_failure_closure(&evac_failure_cl);
4861 4865 pss.set_partial_scan_closure(&partial_scan_cl);
4862 4866
4863 4867 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4864 4868 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4865 4869
4866 4870 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4867 4871 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4868 4872
4869 4873 bool only_young = _g1h->g1_policy()->gcs_are_young();
4870 4874 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4871 4875 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4872 4876
4873 4877 OopClosure* scan_root_cl = &only_scan_root_cl;
4874 4878 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4875 4879
4876 4880 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4877 4881 // We also need to mark copied objects.
4878 4882 scan_root_cl = &scan_mark_root_cl;
4879 4883 scan_klasses_cl = &scan_mark_klasses_cl_s;
4880 4884 }
4881 4885
4882 4886 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4883 4887
4884 4888 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4885 4889
4886 4890 pss.start_strong_roots();
4887 4891 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4888 4892 SharedHeap::ScanningOption(so),
4889 4893 scan_root_cl,
4890 4894 &push_heap_rs_cl,
4891 4895 scan_klasses_cl,
4892 4896 worker_id);
4893 4897 pss.end_strong_roots();
4894 4898
4895 4899 {
4896 4900 double start = os::elapsedTime();
4897 4901 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4898 4902 evac.do_void();
4899 4903 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4900 4904 double term_ms = pss.term_time()*1000.0;
4901 4905 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4902 4906 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4903 4907 }
4904 4908 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4905 4909 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4906 4910
4907 4911 if (ParallelGCVerbose) {
4908 4912 MutexLocker x(stats_lock());
4909 4913 pss.print_termination_stats(worker_id);
4910 4914 }
4911 4915
4912 4916 assert(pss.refs()->is_empty(), "should be empty");
4913 4917
4914 4918 // Close the inner scope so that the ResourceMark and HandleMark
4915 4919 // destructors are executed here and are included as part of the
4916 4920 // "GC Worker Time".
4917 4921 }
4918 4922
4919 4923 double end_time_ms = os::elapsedTime() * 1000.0;
4920 4924 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4921 4925 }
4922 4926 };
4923 4927
4924 4928 // *** Common G1 Evacuation Stuff
4925 4929
4926 4930 // Closures that support the filtering of CodeBlobs scanned during
4927 4931 // external root scanning.
4928 4932
4929 4933 // Closure applied to reference fields in code blobs (specifically nmethods)
4930 4934 // to determine whether an nmethod contains references that point into
4931 4935 // the collection set. Used as a predicate when walking code roots so
4932 4936 // that only nmethods that point into the collection set are added to the
4933 4937 // 'marked' list.
4934 4938
4935 4939 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4936 4940
4937 4941 class G1PointsIntoCSOopClosure : public OopClosure {
4938 4942 G1CollectedHeap* _g1;
4939 4943 bool _points_into_cs;
4940 4944 public:
4941 4945 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4942 4946 _g1(g1), _points_into_cs(false) { }
4943 4947
4944 4948 bool points_into_cs() const { return _points_into_cs; }
4945 4949
4946 4950 template <class T>
4947 4951 void do_oop_nv(T* p) {
4948 4952 if (!_points_into_cs) {
4949 4953 T heap_oop = oopDesc::load_heap_oop(p);
4950 4954 if (!oopDesc::is_null(heap_oop) &&
4951 4955 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4952 4956 _points_into_cs = true;
4953 4957 }
4954 4958 }
4955 4959 }
4956 4960
4957 4961 virtual void do_oop(oop* p) { do_oop_nv(p); }
4958 4962 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4959 4963 };
4960 4964
4961 4965 G1CollectedHeap* _g1;
4962 4966
4963 4967 public:
4964 4968 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4965 4969 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4966 4970
4967 4971 virtual void do_code_blob(CodeBlob* cb) {
4968 4972 nmethod* nm = cb->as_nmethod_or_null();
4969 4973 if (nm != NULL && !(nm->test_oops_do_mark())) {
4970 4974 G1PointsIntoCSOopClosure predicate_cl(_g1);
4971 4975 nm->oops_do(&predicate_cl);
4972 4976
4973 4977 if (predicate_cl.points_into_cs()) {
4974 4978 // At least one of the reference fields or the oop relocations
4975 4979 // in the nmethod points into the collection set. We have to
4976 4980 // 'mark' this nmethod.
4977 4981 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4978 4982 // or MarkingCodeBlobClosure::do_code_blob() change.
4979 4983 if (!nm->test_set_oops_do_mark()) {
4980 4984 do_newly_marked_nmethod(nm);
4981 4985 }
4982 4986 }
4983 4987 }
4984 4988 }
4985 4989 };
4986 4990
4987 4991 // This method is run in a GC worker.
4988 4992
4989 4993 void
4990 4994 G1CollectedHeap::
4991 4995 g1_process_strong_roots(bool is_scavenging,
4992 4996 ScanningOption so,
4993 4997 OopClosure* scan_non_heap_roots,
4994 4998 OopsInHeapRegionClosure* scan_rs,
4995 4999 G1KlassScanClosure* scan_klasses,
4996 5000 int worker_i) {
4997 5001
4998 5002 // First scan the strong roots
4999 5003 double ext_roots_start = os::elapsedTime();
5000 5004 double closure_app_time_sec = 0.0;
5001 5005
5002 5006 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5003 5007
5004 5008 // Walk the code cache w/o buffering, because StarTask cannot handle
5005 5009 // unaligned oop locations.
5006 5010 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5007 5011
5008 5012 process_strong_roots(false, // no scoping; this is parallel code
5009 5013 is_scavenging, so,
5010 5014 &buf_scan_non_heap_roots,
5011 5015 &eager_scan_code_roots,
5012 5016 scan_klasses
5013 5017 );
5014 5018
5015 5019 // Now the CM ref_processor roots.
5016 5020 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5017 5021 // We need to treat the discovered reference lists of the
5018 5022 // concurrent mark ref processor as roots and keep entries
5019 5023 // (which are added by the marking threads) on them live
5020 5024 // until they can be processed at the end of marking.
5021 5025 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5022 5026 }
5023 5027
5024 5028 // Finish up any enqueued closure apps (attributed as object copy time).
5025 5029 buf_scan_non_heap_roots.done();
5026 5030
5027 5031 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5028 5032
5029 5033 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5030 5034
5031 5035 double ext_root_time_ms =
5032 5036 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5033 5037
5034 5038 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5035 5039
5036 5040 // During conc marking we have to filter the per-thread SATB buffers
5037 5041 // to make sure we remove any oops into the CSet (which will show up
5038 5042 // as implicitly live).
5039 5043 double satb_filtering_ms = 0.0;
5040 5044 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5041 5045 if (mark_in_progress()) {
5042 5046 double satb_filter_start = os::elapsedTime();
5043 5047
5044 5048 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5045 5049
5046 5050 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5047 5051 }
5048 5052 }
5049 5053 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5050 5054
5051 5055 // Now scan the complement of the collection set.
5052 5056 if (scan_rs != NULL) {
5053 5057 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5054 5058 }
5055 5059 _process_strong_tasks->all_tasks_completed();
5056 5060 }
5057 5061
5058 5062 void
5059 5063 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5060 5064 OopClosure* non_root_closure) {
5061 5065 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5062 5066 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5063 5067 }
5064 5068
5065 5069 // Weak Reference Processing support
5066 5070
5067 5071 // An always "is_alive" closure that is used to preserve referents.
5068 5072 // If the object is non-null then it's alive. Used in the preservation
5069 5073 // of referent objects that are pointed to by reference objects
5070 5074 // discovered by the CM ref processor.
5071 5075 class G1AlwaysAliveClosure: public BoolObjectClosure {
5072 5076 G1CollectedHeap* _g1;
5073 5077 public:
5074 5078 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5075 5079 void do_object(oop p) { assert(false, "Do not call."); }
5076 5080 bool do_object_b(oop p) {
5077 5081 if (p != NULL) {
5078 5082 return true;
5079 5083 }
5080 5084 return false;
5081 5085 }
5082 5086 };
5083 5087
5084 5088 bool G1STWIsAliveClosure::do_object_b(oop p) {
5085 5089 // An object is reachable if it is outside the collection set,
5086 5090 // or is inside and copied.
5087 5091 return !_g1->obj_in_cs(p) || p->is_forwarded();
5088 5092 }
5089 5093
5090 5094 // Non Copying Keep Alive closure
5091 5095 class G1KeepAliveClosure: public OopClosure {
5092 5096 G1CollectedHeap* _g1;
5093 5097 public:
5094 5098 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5095 5099 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5096 5100 void do_oop( oop* p) {
5097 5101 oop obj = *p;
5098 5102
5099 5103 if (_g1->obj_in_cs(obj)) {
5100 5104 assert( obj->is_forwarded(), "invariant" );
5101 5105 *p = obj->forwardee();
5102 5106 }
5103 5107 }
5104 5108 };
5105 5109
5106 5110 // Copying Keep Alive closure - can be called from both
5107 5111 // serial and parallel code as long as different worker
5108 5112 // threads utilize different G1ParScanThreadState instances
5109 5113 // and different queues.
5110 5114
5111 5115 class G1CopyingKeepAliveClosure: public OopClosure {
5112 5116 G1CollectedHeap* _g1h;
5113 5117 OopClosure* _copy_non_heap_obj_cl;
5114 5118 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5115 5119 G1ParScanThreadState* _par_scan_state;
5116 5120
5117 5121 public:
5118 5122 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5119 5123 OopClosure* non_heap_obj_cl,
5120 5124 OopsInHeapRegionClosure* metadata_obj_cl,
5121 5125 G1ParScanThreadState* pss):
5122 5126 _g1h(g1h),
5123 5127 _copy_non_heap_obj_cl(non_heap_obj_cl),
5124 5128 _copy_metadata_obj_cl(metadata_obj_cl),
5125 5129 _par_scan_state(pss)
5126 5130 {}
5127 5131
5128 5132 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5129 5133 virtual void do_oop( oop* p) { do_oop_work(p); }
5130 5134
5131 5135 template <class T> void do_oop_work(T* p) {
5132 5136 oop obj = oopDesc::load_decode_heap_oop(p);
5133 5137
5134 5138 if (_g1h->obj_in_cs(obj)) {
5135 5139 // If the referent object has been forwarded (either copied
5136 5140 // to a new location or to itself in the event of an
5137 5141 // evacuation failure) then we need to update the reference
5138 5142 // field and, if both reference and referent are in the G1
5139 5143 // heap, update the RSet for the referent.
5140 5144 //
5141 5145 // If the referent has not been forwarded then we have to keep
5142 5146 // it alive by policy. Therefore we have copy the referent.
5143 5147 //
5144 5148 // If the reference field is in the G1 heap then we can push
5145 5149 // on the PSS queue. When the queue is drained (after each
5146 5150 // phase of reference processing) the object and it's followers
5147 5151 // will be copied, the reference field set to point to the
5148 5152 // new location, and the RSet updated. Otherwise we need to
5149 5153 // use the the non-heap or metadata closures directly to copy
5150 5154 // the refernt object and update the pointer, while avoiding
5151 5155 // updating the RSet.
5152 5156
5153 5157 if (_g1h->is_in_g1_reserved(p)) {
5154 5158 _par_scan_state->push_on_queue(p);
5155 5159 } else {
5156 5160 assert(!ClassLoaderDataGraph::contains((address)p),
5157 5161 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5158 5162 PTR_FORMAT, p));
5159 5163 _copy_non_heap_obj_cl->do_oop(p);
5160 5164 }
5161 5165 }
5162 5166 }
5163 5167 };
5164 5168
5165 5169 // Serial drain queue closure. Called as the 'complete_gc'
5166 5170 // closure for each discovered list in some of the
5167 5171 // reference processing phases.
5168 5172
5169 5173 class G1STWDrainQueueClosure: public VoidClosure {
5170 5174 protected:
5171 5175 G1CollectedHeap* _g1h;
5172 5176 G1ParScanThreadState* _par_scan_state;
5173 5177
5174 5178 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5175 5179
5176 5180 public:
5177 5181 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5178 5182 _g1h(g1h),
5179 5183 _par_scan_state(pss)
5180 5184 { }
5181 5185
5182 5186 void do_void() {
5183 5187 G1ParScanThreadState* const pss = par_scan_state();
5184 5188 pss->trim_queue();
5185 5189 }
5186 5190 };
5187 5191
5188 5192 // Parallel Reference Processing closures
5189 5193
5190 5194 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5191 5195 // processing during G1 evacuation pauses.
5192 5196
5193 5197 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5194 5198 private:
5195 5199 G1CollectedHeap* _g1h;
5196 5200 RefToScanQueueSet* _queues;
5197 5201 FlexibleWorkGang* _workers;
5198 5202 int _active_workers;
5199 5203
5200 5204 public:
5201 5205 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5202 5206 FlexibleWorkGang* workers,
5203 5207 RefToScanQueueSet *task_queues,
5204 5208 int n_workers) :
5205 5209 _g1h(g1h),
5206 5210 _queues(task_queues),
5207 5211 _workers(workers),
5208 5212 _active_workers(n_workers)
5209 5213 {
5210 5214 assert(n_workers > 0, "shouldn't call this otherwise");
5211 5215 }
5212 5216
5213 5217 // Executes the given task using concurrent marking worker threads.
5214 5218 virtual void execute(ProcessTask& task);
5215 5219 virtual void execute(EnqueueTask& task);
5216 5220 };
5217 5221
5218 5222 // Gang task for possibly parallel reference processing
5219 5223
5220 5224 class G1STWRefProcTaskProxy: public AbstractGangTask {
5221 5225 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5222 5226 ProcessTask& _proc_task;
5223 5227 G1CollectedHeap* _g1h;
5224 5228 RefToScanQueueSet *_task_queues;
5225 5229 ParallelTaskTerminator* _terminator;
5226 5230
5227 5231 public:
5228 5232 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5229 5233 G1CollectedHeap* g1h,
5230 5234 RefToScanQueueSet *task_queues,
5231 5235 ParallelTaskTerminator* terminator) :
5232 5236 AbstractGangTask("Process reference objects in parallel"),
5233 5237 _proc_task(proc_task),
5234 5238 _g1h(g1h),
5235 5239 _task_queues(task_queues),
5236 5240 _terminator(terminator)
5237 5241 {}
5238 5242
5239 5243 virtual void work(uint worker_id) {
5240 5244 // The reference processing task executed by a single worker.
5241 5245 ResourceMark rm;
5242 5246 HandleMark hm;
5243 5247
5244 5248 G1STWIsAliveClosure is_alive(_g1h);
5245 5249
5246 5250 G1ParScanThreadState pss(_g1h, worker_id);
5247 5251
5248 5252 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5249 5253 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5250 5254 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5251 5255
5252 5256 pss.set_evac_closure(&scan_evac_cl);
5253 5257 pss.set_evac_failure_closure(&evac_failure_cl);
5254 5258 pss.set_partial_scan_closure(&partial_scan_cl);
5255 5259
5256 5260 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5257 5261 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5258 5262
5259 5263 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5260 5264 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5261 5265
5262 5266 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5263 5267 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5264 5268
5265 5269 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5266 5270 // We also need to mark copied objects.
5267 5271 copy_non_heap_cl = ©_mark_non_heap_cl;
5268 5272 copy_metadata_cl = ©_mark_metadata_cl;
5269 5273 }
5270 5274
5271 5275 // Keep alive closure.
5272 5276 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5273 5277
5274 5278 // Complete GC closure
5275 5279 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5276 5280
5277 5281 // Call the reference processing task's work routine.
5278 5282 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5279 5283
5280 5284 // Note we cannot assert that the refs array is empty here as not all
5281 5285 // of the processing tasks (specifically phase2 - pp2_work) execute
5282 5286 // the complete_gc closure (which ordinarily would drain the queue) so
5283 5287 // the queue may not be empty.
5284 5288 }
5285 5289 };
5286 5290
5287 5291 // Driver routine for parallel reference processing.
5288 5292 // Creates an instance of the ref processing gang
5289 5293 // task and has the worker threads execute it.
5290 5294 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5291 5295 assert(_workers != NULL, "Need parallel worker threads.");
5292 5296
5293 5297 ParallelTaskTerminator terminator(_active_workers, _queues);
5294 5298 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5295 5299
5296 5300 _g1h->set_par_threads(_active_workers);
5297 5301 _workers->run_task(&proc_task_proxy);
5298 5302 _g1h->set_par_threads(0);
5299 5303 }
5300 5304
5301 5305 // Gang task for parallel reference enqueueing.
5302 5306
5303 5307 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5304 5308 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5305 5309 EnqueueTask& _enq_task;
5306 5310
5307 5311 public:
5308 5312 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5309 5313 AbstractGangTask("Enqueue reference objects in parallel"),
5310 5314 _enq_task(enq_task)
5311 5315 { }
5312 5316
5313 5317 virtual void work(uint worker_id) {
5314 5318 _enq_task.work(worker_id);
5315 5319 }
5316 5320 };
5317 5321
5318 5322 // Driver routine for parallel reference enqueing.
5319 5323 // Creates an instance of the ref enqueueing gang
5320 5324 // task and has the worker threads execute it.
5321 5325
5322 5326 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5323 5327 assert(_workers != NULL, "Need parallel worker threads.");
5324 5328
5325 5329 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5326 5330
5327 5331 _g1h->set_par_threads(_active_workers);
5328 5332 _workers->run_task(&enq_task_proxy);
5329 5333 _g1h->set_par_threads(0);
5330 5334 }
5331 5335
5332 5336 // End of weak reference support closures
5333 5337
5334 5338 // Abstract task used to preserve (i.e. copy) any referent objects
5335 5339 // that are in the collection set and are pointed to by reference
5336 5340 // objects discovered by the CM ref processor.
5337 5341
5338 5342 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5339 5343 protected:
5340 5344 G1CollectedHeap* _g1h;
5341 5345 RefToScanQueueSet *_queues;
5342 5346 ParallelTaskTerminator _terminator;
5343 5347 uint _n_workers;
5344 5348
5345 5349 public:
5346 5350 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5347 5351 AbstractGangTask("ParPreserveCMReferents"),
5348 5352 _g1h(g1h),
5349 5353 _queues(task_queues),
5350 5354 _terminator(workers, _queues),
5351 5355 _n_workers(workers)
5352 5356 { }
5353 5357
5354 5358 void work(uint worker_id) {
5355 5359 ResourceMark rm;
5356 5360 HandleMark hm;
5357 5361
5358 5362 G1ParScanThreadState pss(_g1h, worker_id);
5359 5363 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5360 5364 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5361 5365 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5362 5366
5363 5367 pss.set_evac_closure(&scan_evac_cl);
5364 5368 pss.set_evac_failure_closure(&evac_failure_cl);
5365 5369 pss.set_partial_scan_closure(&partial_scan_cl);
5366 5370
5367 5371 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5368 5372
5369 5373
5370 5374 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5371 5375 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5372 5376
5373 5377 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5374 5378 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5375 5379
5376 5380 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5377 5381 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5378 5382
5379 5383 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5380 5384 // We also need to mark copied objects.
5381 5385 copy_non_heap_cl = ©_mark_non_heap_cl;
5382 5386 copy_metadata_cl = ©_mark_metadata_cl;
5383 5387 }
5384 5388
5385 5389 // Is alive closure
5386 5390 G1AlwaysAliveClosure always_alive(_g1h);
5387 5391
5388 5392 // Copying keep alive closure. Applied to referent objects that need
5389 5393 // to be copied.
5390 5394 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5391 5395
5392 5396 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5393 5397
5394 5398 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5395 5399 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5396 5400
5397 5401 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5398 5402 // So this must be true - but assert just in case someone decides to
5399 5403 // change the worker ids.
5400 5404 assert(0 <= worker_id && worker_id < limit, "sanity");
5401 5405 assert(!rp->discovery_is_atomic(), "check this code");
5402 5406
5403 5407 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5404 5408 for (uint idx = worker_id; idx < limit; idx += stride) {
5405 5409 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5406 5410
5407 5411 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5408 5412 while (iter.has_next()) {
5409 5413 // Since discovery is not atomic for the CM ref processor, we
5410 5414 // can see some null referent objects.
5411 5415 iter.load_ptrs(DEBUG_ONLY(true));
5412 5416 oop ref = iter.obj();
5413 5417
5414 5418 // This will filter nulls.
5415 5419 if (iter.is_referent_alive()) {
5416 5420 iter.make_referent_alive();
5417 5421 }
5418 5422 iter.move_to_next();
5419 5423 }
5420 5424 }
5421 5425
5422 5426 // Drain the queue - which may cause stealing
5423 5427 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5424 5428 drain_queue.do_void();
5425 5429 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5426 5430 assert(pss.refs()->is_empty(), "should be");
5427 5431 }
5428 5432 };
5429 5433
5430 5434 // Weak Reference processing during an evacuation pause (part 1).
5431 5435 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5432 5436 double ref_proc_start = os::elapsedTime();
5433 5437
5434 5438 ReferenceProcessor* rp = _ref_processor_stw;
5435 5439 assert(rp->discovery_enabled(), "should have been enabled");
5436 5440
5437 5441 // Any reference objects, in the collection set, that were 'discovered'
5438 5442 // by the CM ref processor should have already been copied (either by
5439 5443 // applying the external root copy closure to the discovered lists, or
5440 5444 // by following an RSet entry).
5441 5445 //
5442 5446 // But some of the referents, that are in the collection set, that these
5443 5447 // reference objects point to may not have been copied: the STW ref
5444 5448 // processor would have seen that the reference object had already
5445 5449 // been 'discovered' and would have skipped discovering the reference,
5446 5450 // but would not have treated the reference object as a regular oop.
5447 5451 // As a reult the copy closure would not have been applied to the
5448 5452 // referent object.
5449 5453 //
5450 5454 // We need to explicitly copy these referent objects - the references
5451 5455 // will be processed at the end of remarking.
5452 5456 //
5453 5457 // We also need to do this copying before we process the reference
5454 5458 // objects discovered by the STW ref processor in case one of these
5455 5459 // referents points to another object which is also referenced by an
5456 5460 // object discovered by the STW ref processor.
5457 5461
5458 5462 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5459 5463 no_of_gc_workers == workers()->active_workers(),
5460 5464 "Need to reset active GC workers");
5461 5465
5462 5466 set_par_threads(no_of_gc_workers);
5463 5467 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5464 5468 no_of_gc_workers,
5465 5469 _task_queues);
5466 5470
5467 5471 if (G1CollectedHeap::use_parallel_gc_threads()) {
5468 5472 workers()->run_task(&keep_cm_referents);
5469 5473 } else {
5470 5474 keep_cm_referents.work(0);
5471 5475 }
5472 5476
5473 5477 set_par_threads(0);
5474 5478
5475 5479 // Closure to test whether a referent is alive.
5476 5480 G1STWIsAliveClosure is_alive(this);
5477 5481
5478 5482 // Even when parallel reference processing is enabled, the processing
5479 5483 // of JNI refs is serial and performed serially by the current thread
5480 5484 // rather than by a worker. The following PSS will be used for processing
5481 5485 // JNI refs.
5482 5486
5483 5487 // Use only a single queue for this PSS.
5484 5488 G1ParScanThreadState pss(this, 0);
5485 5489
5486 5490 // We do not embed a reference processor in the copying/scanning
5487 5491 // closures while we're actually processing the discovered
5488 5492 // reference objects.
5489 5493 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5490 5494 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5491 5495 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5492 5496
5493 5497 pss.set_evac_closure(&scan_evac_cl);
5494 5498 pss.set_evac_failure_closure(&evac_failure_cl);
5495 5499 pss.set_partial_scan_closure(&partial_scan_cl);
5496 5500
5497 5501 assert(pss.refs()->is_empty(), "pre-condition");
5498 5502
5499 5503 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5500 5504 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5501 5505
5502 5506 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5503 5507 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5504 5508
5505 5509 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5506 5510 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5507 5511
5508 5512 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5509 5513 // We also need to mark copied objects.
5510 5514 copy_non_heap_cl = ©_mark_non_heap_cl;
5511 5515 copy_metadata_cl = ©_mark_metadata_cl;
5512 5516 }
5513 5517
5514 5518 // Keep alive closure.
5515 5519 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5516 5520
5517 5521 // Serial Complete GC closure
5518 5522 G1STWDrainQueueClosure drain_queue(this, &pss);
5519 5523
5520 5524 // Setup the soft refs policy...
5521 5525 rp->setup_policy(false);
5522 5526
5523 5527 if (!rp->processing_is_mt()) {
5524 5528 // Serial reference processing...
5525 5529 rp->process_discovered_references(&is_alive,
5526 5530 &keep_alive,
5527 5531 &drain_queue,
5528 5532 NULL);
5529 5533 } else {
5530 5534 // Parallel reference processing
5531 5535 assert(rp->num_q() == no_of_gc_workers, "sanity");
5532 5536 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5533 5537
5534 5538 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5535 5539 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5536 5540 }
5537 5541
5538 5542 // We have completed copying any necessary live referent objects
5539 5543 // (that were not copied during the actual pause) so we can
5540 5544 // retire any active alloc buffers
5541 5545 pss.retire_alloc_buffers();
5542 5546 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5543 5547
5544 5548 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5545 5549 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5546 5550 }
5547 5551
5548 5552 // Weak Reference processing during an evacuation pause (part 2).
5549 5553 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5550 5554 double ref_enq_start = os::elapsedTime();
5551 5555
5552 5556 ReferenceProcessor* rp = _ref_processor_stw;
5553 5557 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5554 5558
5555 5559 // Now enqueue any remaining on the discovered lists on to
5556 5560 // the pending list.
5557 5561 if (!rp->processing_is_mt()) {
5558 5562 // Serial reference processing...
5559 5563 rp->enqueue_discovered_references();
5560 5564 } else {
5561 5565 // Parallel reference enqueuing
5562 5566
5563 5567 assert(no_of_gc_workers == workers()->active_workers(),
5564 5568 "Need to reset active workers");
5565 5569 assert(rp->num_q() == no_of_gc_workers, "sanity");
5566 5570 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5567 5571
5568 5572 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5569 5573 rp->enqueue_discovered_references(&par_task_executor);
5570 5574 }
5571 5575
5572 5576 rp->verify_no_references_recorded();
5573 5577 assert(!rp->discovery_enabled(), "should have been disabled");
5574 5578
5575 5579 // FIXME
5576 5580 // CM's reference processing also cleans up the string and symbol tables.
5577 5581 // Should we do that here also? We could, but it is a serial operation
5578 5582 // and could signicantly increase the pause time.
5579 5583
5580 5584 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5581 5585 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5582 5586 }
5583 5587
5584 5588 void G1CollectedHeap::evacuate_collection_set() {
5585 5589 _expand_heap_after_alloc_failure = true;
5586 5590 set_evacuation_failed(false);
5587 5591
5588 5592 // Should G1EvacuationFailureALot be in effect for this GC?
5589 5593 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5590 5594
5591 5595 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5592 5596 concurrent_g1_refine()->set_use_cache(false);
5593 5597 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5594 5598
5595 5599 uint n_workers;
5596 5600 if (G1CollectedHeap::use_parallel_gc_threads()) {
5597 5601 n_workers =
5598 5602 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5599 5603 workers()->active_workers(),
5600 5604 Threads::number_of_non_daemon_threads());
5601 5605 assert(UseDynamicNumberOfGCThreads ||
5602 5606 n_workers == workers()->total_workers(),
5603 5607 "If not dynamic should be using all the workers");
5604 5608 workers()->set_active_workers(n_workers);
5605 5609 set_par_threads(n_workers);
5606 5610 } else {
5607 5611 assert(n_par_threads() == 0,
5608 5612 "Should be the original non-parallel value");
5609 5613 n_workers = 1;
5610 5614 }
5611 5615
5612 5616 G1ParTask g1_par_task(this, _task_queues);
5613 5617
5614 5618 init_for_evac_failure(NULL);
5615 5619
5616 5620 rem_set()->prepare_for_younger_refs_iterate(true);
5617 5621
5618 5622 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5619 5623 double start_par_time_sec = os::elapsedTime();
5620 5624 double end_par_time_sec;
5621 5625
5622 5626 {
5623 5627 StrongRootsScope srs(this);
5624 5628
5625 5629 if (G1CollectedHeap::use_parallel_gc_threads()) {
5626 5630 // The individual threads will set their evac-failure closures.
5627 5631 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5628 5632 // These tasks use ShareHeap::_process_strong_tasks
5629 5633 assert(UseDynamicNumberOfGCThreads ||
5630 5634 workers()->active_workers() == workers()->total_workers(),
5631 5635 "If not dynamic should be using all the workers");
5632 5636 workers()->run_task(&g1_par_task);
5633 5637 } else {
5634 5638 g1_par_task.set_for_termination(n_workers);
5635 5639 g1_par_task.work(0);
5636 5640 }
5637 5641 end_par_time_sec = os::elapsedTime();
5638 5642
5639 5643 // Closing the inner scope will execute the destructor
5640 5644 // for the StrongRootsScope object. We record the current
5641 5645 // elapsed time before closing the scope so that time
5642 5646 // taken for the SRS destructor is NOT included in the
5643 5647 // reported parallel time.
5644 5648 }
5645 5649
5646 5650 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5647 5651 g1_policy()->phase_times()->record_par_time(par_time_ms);
5648 5652
5649 5653 double code_root_fixup_time_ms =
5650 5654 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5651 5655 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5652 5656
5653 5657 set_par_threads(0);
5654 5658
5655 5659 // Process any discovered reference objects - we have
5656 5660 // to do this _before_ we retire the GC alloc regions
5657 5661 // as we may have to copy some 'reachable' referent
5658 5662 // objects (and their reachable sub-graphs) that were
5659 5663 // not copied during the pause.
5660 5664 process_discovered_references(n_workers);
5661 5665
5662 5666 // Weak root processing.
5663 5667 // Note: when JSR 292 is enabled and code blobs can contain
5664 5668 // non-perm oops then we will need to process the code blobs
5665 5669 // here too.
5666 5670 {
5667 5671 G1STWIsAliveClosure is_alive(this);
5668 5672 G1KeepAliveClosure keep_alive(this);
5669 5673 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5670 5674 }
5671 5675
5672 5676 release_gc_alloc_regions(n_workers);
5673 5677 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5674 5678
5675 5679 concurrent_g1_refine()->clear_hot_cache();
5676 5680 concurrent_g1_refine()->set_use_cache(true);
5677 5681
5678 5682 finalize_for_evac_failure();
5679 5683
5680 5684 if (evacuation_failed()) {
5681 5685 remove_self_forwarding_pointers();
5682 5686
5683 5687 // Reset the G1EvacuationFailureALot counters and flags
5684 5688 // Note: the values are reset only when an actual
5685 5689 // evacuation failure occurs.
5686 5690 NOT_PRODUCT(reset_evacuation_should_fail();)
5687 5691 }
5688 5692
5689 5693 // Enqueue any remaining references remaining on the STW
5690 5694 // reference processor's discovered lists. We need to do
5691 5695 // this after the card table is cleaned (and verified) as
5692 5696 // the act of enqueuing entries on to the pending list
5693 5697 // will log these updates (and dirty their associated
5694 5698 // cards). We need these updates logged to update any
5695 5699 // RSets.
5696 5700 enqueue_discovered_references(n_workers);
5697 5701
5698 5702 if (G1DeferredRSUpdate) {
5699 5703 RedirtyLoggedCardTableEntryFastClosure redirty;
5700 5704 dirty_card_queue_set().set_closure(&redirty);
5701 5705 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5702 5706
5703 5707 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5704 5708 dcq.merge_bufferlists(&dirty_card_queue_set());
5705 5709 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5706 5710 }
5707 5711 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5708 5712 }
5709 5713
5710 5714 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5711 5715 size_t* pre_used,
5712 5716 FreeRegionList* free_list,
5713 5717 OldRegionSet* old_proxy_set,
5714 5718 HumongousRegionSet* humongous_proxy_set,
5715 5719 HRRSCleanupTask* hrrs_cleanup_task,
5716 5720 bool par) {
5717 5721 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5718 5722 if (hr->isHumongous()) {
5719 5723 assert(hr->startsHumongous(), "we should only see starts humongous");
5720 5724 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5721 5725 } else {
5722 5726 _old_set.remove_with_proxy(hr, old_proxy_set);
5723 5727 free_region(hr, pre_used, free_list, par);
5724 5728 }
5725 5729 } else {
5726 5730 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5727 5731 }
5728 5732 }
5729 5733
5730 5734 void G1CollectedHeap::free_region(HeapRegion* hr,
5731 5735 size_t* pre_used,
5732 5736 FreeRegionList* free_list,
5733 5737 bool par) {
5734 5738 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5735 5739 assert(!hr->is_empty(), "the region should not be empty");
5736 5740 assert(free_list != NULL, "pre-condition");
5737 5741
5738 5742 *pre_used += hr->used();
5739 5743 hr->hr_clear(par, true /* clear_space */);
5740 5744 free_list->add_as_head(hr);
5741 5745 }
5742 5746
5743 5747 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5744 5748 size_t* pre_used,
5745 5749 FreeRegionList* free_list,
5746 5750 HumongousRegionSet* humongous_proxy_set,
5747 5751 bool par) {
5748 5752 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5749 5753 assert(free_list != NULL, "pre-condition");
5750 5754 assert(humongous_proxy_set != NULL, "pre-condition");
5751 5755
5752 5756 size_t hr_used = hr->used();
5753 5757 size_t hr_capacity = hr->capacity();
5754 5758 size_t hr_pre_used = 0;
5755 5759 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5756 5760 // We need to read this before we make the region non-humongous,
5757 5761 // otherwise the information will be gone.
5758 5762 uint last_index = hr->last_hc_index();
5759 5763 hr->set_notHumongous();
5760 5764 free_region(hr, &hr_pre_used, free_list, par);
5761 5765
5762 5766 uint i = hr->hrs_index() + 1;
5763 5767 while (i < last_index) {
5764 5768 HeapRegion* curr_hr = region_at(i);
5765 5769 assert(curr_hr->continuesHumongous(), "invariant");
5766 5770 curr_hr->set_notHumongous();
5767 5771 free_region(curr_hr, &hr_pre_used, free_list, par);
5768 5772 i += 1;
5769 5773 }
5770 5774 assert(hr_pre_used == hr_used,
5771 5775 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5772 5776 "should be the same", hr_pre_used, hr_used));
5773 5777 *pre_used += hr_pre_used;
5774 5778 }
5775 5779
5776 5780 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5777 5781 FreeRegionList* free_list,
5778 5782 OldRegionSet* old_proxy_set,
5779 5783 HumongousRegionSet* humongous_proxy_set,
5780 5784 bool par) {
5781 5785 if (pre_used > 0) {
5782 5786 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5783 5787 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5784 5788 assert(_summary_bytes_used >= pre_used,
5785 5789 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5786 5790 "should be >= pre_used: "SIZE_FORMAT,
5787 5791 _summary_bytes_used, pre_used));
5788 5792 _summary_bytes_used -= pre_used;
5789 5793 }
5790 5794 if (free_list != NULL && !free_list->is_empty()) {
5791 5795 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5792 5796 _free_list.add_as_head(free_list);
5793 5797 }
5794 5798 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5795 5799 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5796 5800 _old_set.update_from_proxy(old_proxy_set);
5797 5801 }
5798 5802 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5799 5803 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5800 5804 _humongous_set.update_from_proxy(humongous_proxy_set);
5801 5805 }
5802 5806 }
5803 5807
5804 5808 class G1ParCleanupCTTask : public AbstractGangTask {
5805 5809 CardTableModRefBS* _ct_bs;
5806 5810 G1CollectedHeap* _g1h;
5807 5811 HeapRegion* volatile _su_head;
5808 5812 public:
5809 5813 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5810 5814 G1CollectedHeap* g1h) :
5811 5815 AbstractGangTask("G1 Par Cleanup CT Task"),
5812 5816 _ct_bs(ct_bs), _g1h(g1h) { }
5813 5817
5814 5818 void work(uint worker_id) {
5815 5819 HeapRegion* r;
5816 5820 while (r = _g1h->pop_dirty_cards_region()) {
5817 5821 clear_cards(r);
5818 5822 }
5819 5823 }
5820 5824
5821 5825 void clear_cards(HeapRegion* r) {
5822 5826 // Cards of the survivors should have already been dirtied.
5823 5827 if (!r->is_survivor()) {
5824 5828 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5825 5829 }
5826 5830 }
5827 5831 };
5828 5832
5829 5833 #ifndef PRODUCT
5830 5834 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5831 5835 G1CollectedHeap* _g1h;
5832 5836 CardTableModRefBS* _ct_bs;
5833 5837 public:
5834 5838 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5835 5839 : _g1h(g1h), _ct_bs(ct_bs) { }
5836 5840 virtual bool doHeapRegion(HeapRegion* r) {
5837 5841 if (r->is_survivor()) {
5838 5842 _g1h->verify_dirty_region(r);
5839 5843 } else {
5840 5844 _g1h->verify_not_dirty_region(r);
5841 5845 }
5842 5846 return false;
5843 5847 }
5844 5848 };
5845 5849
5846 5850 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5847 5851 // All of the region should be clean.
5848 5852 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5849 5853 MemRegion mr(hr->bottom(), hr->end());
5850 5854 ct_bs->verify_not_dirty_region(mr);
5851 5855 }
5852 5856
5853 5857 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5854 5858 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5855 5859 // dirty allocated blocks as they allocate them. The thread that
5856 5860 // retires each region and replaces it with a new one will do a
5857 5861 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5858 5862 // not dirty that area (one less thing to have to do while holding
5859 5863 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5860 5864 // is dirty.
5861 5865 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5862 5866 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5863 5867 ct_bs->verify_dirty_region(mr);
5864 5868 }
5865 5869
5866 5870 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5867 5871 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5868 5872 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5869 5873 verify_dirty_region(hr);
5870 5874 }
5871 5875 }
5872 5876
5873 5877 void G1CollectedHeap::verify_dirty_young_regions() {
5874 5878 verify_dirty_young_list(_young_list->first_region());
5875 5879 }
5876 5880 #endif
5877 5881
5878 5882 void G1CollectedHeap::cleanUpCardTable() {
5879 5883 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5880 5884 double start = os::elapsedTime();
5881 5885
5882 5886 {
5883 5887 // Iterate over the dirty cards region list.
5884 5888 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5885 5889
5886 5890 if (G1CollectedHeap::use_parallel_gc_threads()) {
5887 5891 set_par_threads();
5888 5892 workers()->run_task(&cleanup_task);
5889 5893 set_par_threads(0);
5890 5894 } else {
5891 5895 while (_dirty_cards_region_list) {
5892 5896 HeapRegion* r = _dirty_cards_region_list;
5893 5897 cleanup_task.clear_cards(r);
5894 5898 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5895 5899 if (_dirty_cards_region_list == r) {
5896 5900 // The last region.
5897 5901 _dirty_cards_region_list = NULL;
5898 5902 }
5899 5903 r->set_next_dirty_cards_region(NULL);
5900 5904 }
5901 5905 }
5902 5906 #ifndef PRODUCT
5903 5907 if (G1VerifyCTCleanup || VerifyAfterGC) {
5904 5908 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5905 5909 heap_region_iterate(&cleanup_verifier);
5906 5910 }
5907 5911 #endif
5908 5912 }
5909 5913
5910 5914 double elapsed = os::elapsedTime() - start;
5911 5915 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5912 5916 }
5913 5917
5914 5918 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5915 5919 size_t pre_used = 0;
5916 5920 FreeRegionList local_free_list("Local List for CSet Freeing");
5917 5921
5918 5922 double young_time_ms = 0.0;
5919 5923 double non_young_time_ms = 0.0;
5920 5924
5921 5925 // Since the collection set is a superset of the the young list,
5922 5926 // all we need to do to clear the young list is clear its
5923 5927 // head and length, and unlink any young regions in the code below
5924 5928 _young_list->clear();
5925 5929
5926 5930 G1CollectorPolicy* policy = g1_policy();
5927 5931
5928 5932 double start_sec = os::elapsedTime();
5929 5933 bool non_young = true;
5930 5934
5931 5935 HeapRegion* cur = cs_head;
5932 5936 int age_bound = -1;
5933 5937 size_t rs_lengths = 0;
5934 5938
5935 5939 while (cur != NULL) {
5936 5940 assert(!is_on_master_free_list(cur), "sanity");
5937 5941 if (non_young) {
5938 5942 if (cur->is_young()) {
5939 5943 double end_sec = os::elapsedTime();
5940 5944 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5941 5945 non_young_time_ms += elapsed_ms;
5942 5946
5943 5947 start_sec = os::elapsedTime();
5944 5948 non_young = false;
5945 5949 }
5946 5950 } else {
5947 5951 if (!cur->is_young()) {
5948 5952 double end_sec = os::elapsedTime();
5949 5953 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5950 5954 young_time_ms += elapsed_ms;
5951 5955
5952 5956 start_sec = os::elapsedTime();
5953 5957 non_young = true;
5954 5958 }
5955 5959 }
5956 5960
5957 5961 rs_lengths += cur->rem_set()->occupied();
5958 5962
5959 5963 HeapRegion* next = cur->next_in_collection_set();
5960 5964 assert(cur->in_collection_set(), "bad CS");
5961 5965 cur->set_next_in_collection_set(NULL);
5962 5966 cur->set_in_collection_set(false);
5963 5967
5964 5968 if (cur->is_young()) {
5965 5969 int index = cur->young_index_in_cset();
5966 5970 assert(index != -1, "invariant");
5967 5971 assert((uint) index < policy->young_cset_region_length(), "invariant");
5968 5972 size_t words_survived = _surviving_young_words[index];
5969 5973 cur->record_surv_words_in_group(words_survived);
5970 5974
5971 5975 // At this point the we have 'popped' cur from the collection set
5972 5976 // (linked via next_in_collection_set()) but it is still in the
5973 5977 // young list (linked via next_young_region()). Clear the
5974 5978 // _next_young_region field.
5975 5979 cur->set_next_young_region(NULL);
5976 5980 } else {
5977 5981 int index = cur->young_index_in_cset();
5978 5982 assert(index == -1, "invariant");
5979 5983 }
5980 5984
5981 5985 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5982 5986 (!cur->is_young() && cur->young_index_in_cset() == -1),
5983 5987 "invariant" );
5984 5988
5985 5989 if (!cur->evacuation_failed()) {
5986 5990 MemRegion used_mr = cur->used_region();
5987 5991
5988 5992 // And the region is empty.
5989 5993 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5990 5994 free_region(cur, &pre_used, &local_free_list, false /* par */);
5991 5995 } else {
5992 5996 cur->uninstall_surv_rate_group();
5993 5997 if (cur->is_young()) {
5994 5998 cur->set_young_index_in_cset(-1);
5995 5999 }
5996 6000 cur->set_not_young();
5997 6001 cur->set_evacuation_failed(false);
5998 6002 // The region is now considered to be old.
5999 6003 _old_set.add(cur);
6000 6004 }
6001 6005 cur = next;
6002 6006 }
6003 6007
6004 6008 policy->record_max_rs_lengths(rs_lengths);
6005 6009 policy->cset_regions_freed();
6006 6010
6007 6011 double end_sec = os::elapsedTime();
6008 6012 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6009 6013
6010 6014 if (non_young) {
6011 6015 non_young_time_ms += elapsed_ms;
6012 6016 } else {
6013 6017 young_time_ms += elapsed_ms;
6014 6018 }
6015 6019
6016 6020 update_sets_after_freeing_regions(pre_used, &local_free_list,
6017 6021 NULL /* old_proxy_set */,
6018 6022 NULL /* humongous_proxy_set */,
6019 6023 false /* par */);
6020 6024 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6021 6025 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6022 6026 }
6023 6027
6024 6028 // This routine is similar to the above but does not record
6025 6029 // any policy statistics or update free lists; we are abandoning
6026 6030 // the current incremental collection set in preparation of a
6027 6031 // full collection. After the full GC we will start to build up
6028 6032 // the incremental collection set again.
6029 6033 // This is only called when we're doing a full collection
6030 6034 // and is immediately followed by the tearing down of the young list.
6031 6035
6032 6036 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6033 6037 HeapRegion* cur = cs_head;
6034 6038
6035 6039 while (cur != NULL) {
6036 6040 HeapRegion* next = cur->next_in_collection_set();
6037 6041 assert(cur->in_collection_set(), "bad CS");
6038 6042 cur->set_next_in_collection_set(NULL);
6039 6043 cur->set_in_collection_set(false);
6040 6044 cur->set_young_index_in_cset(-1);
6041 6045 cur = next;
6042 6046 }
6043 6047 }
6044 6048
6045 6049 void G1CollectedHeap::set_free_regions_coming() {
6046 6050 if (G1ConcRegionFreeingVerbose) {
6047 6051 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6048 6052 "setting free regions coming");
6049 6053 }
6050 6054
6051 6055 assert(!free_regions_coming(), "pre-condition");
6052 6056 _free_regions_coming = true;
6053 6057 }
6054 6058
6055 6059 void G1CollectedHeap::reset_free_regions_coming() {
6056 6060 assert(free_regions_coming(), "pre-condition");
6057 6061
6058 6062 {
6059 6063 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6060 6064 _free_regions_coming = false;
6061 6065 SecondaryFreeList_lock->notify_all();
6062 6066 }
6063 6067
6064 6068 if (G1ConcRegionFreeingVerbose) {
6065 6069 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6066 6070 "reset free regions coming");
6067 6071 }
6068 6072 }
6069 6073
6070 6074 void G1CollectedHeap::wait_while_free_regions_coming() {
6071 6075 // Most of the time we won't have to wait, so let's do a quick test
6072 6076 // first before we take the lock.
6073 6077 if (!free_regions_coming()) {
6074 6078 return;
6075 6079 }
6076 6080
6077 6081 if (G1ConcRegionFreeingVerbose) {
6078 6082 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6079 6083 "waiting for free regions");
6080 6084 }
6081 6085
6082 6086 {
6083 6087 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6084 6088 while (free_regions_coming()) {
6085 6089 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6086 6090 }
6087 6091 }
6088 6092
6089 6093 if (G1ConcRegionFreeingVerbose) {
6090 6094 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6091 6095 "done waiting for free regions");
6092 6096 }
6093 6097 }
6094 6098
6095 6099 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6096 6100 assert(heap_lock_held_for_gc(),
6097 6101 "the heap lock should already be held by or for this thread");
6098 6102 _young_list->push_region(hr);
6099 6103 }
6100 6104
6101 6105 class NoYoungRegionsClosure: public HeapRegionClosure {
6102 6106 private:
6103 6107 bool _success;
6104 6108 public:
6105 6109 NoYoungRegionsClosure() : _success(true) { }
6106 6110 bool doHeapRegion(HeapRegion* r) {
6107 6111 if (r->is_young()) {
6108 6112 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6109 6113 r->bottom(), r->end());
6110 6114 _success = false;
6111 6115 }
6112 6116 return false;
6113 6117 }
6114 6118 bool success() { return _success; }
6115 6119 };
6116 6120
6117 6121 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6118 6122 bool ret = _young_list->check_list_empty(check_sample);
6119 6123
6120 6124 if (check_heap) {
6121 6125 NoYoungRegionsClosure closure;
6122 6126 heap_region_iterate(&closure);
6123 6127 ret = ret && closure.success();
6124 6128 }
6125 6129
6126 6130 return ret;
6127 6131 }
6128 6132
6129 6133 class TearDownRegionSetsClosure : public HeapRegionClosure {
6130 6134 private:
6131 6135 OldRegionSet *_old_set;
6132 6136
6133 6137 public:
6134 6138 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6135 6139
6136 6140 bool doHeapRegion(HeapRegion* r) {
6137 6141 if (r->is_empty()) {
6138 6142 // We ignore empty regions, we'll empty the free list afterwards
6139 6143 } else if (r->is_young()) {
6140 6144 // We ignore young regions, we'll empty the young list afterwards
6141 6145 } else if (r->isHumongous()) {
6142 6146 // We ignore humongous regions, we're not tearing down the
6143 6147 // humongous region set
6144 6148 } else {
6145 6149 // The rest should be old
6146 6150 _old_set->remove(r);
6147 6151 }
6148 6152 return false;
6149 6153 }
6150 6154
6151 6155 ~TearDownRegionSetsClosure() {
6152 6156 assert(_old_set->is_empty(), "post-condition");
6153 6157 }
6154 6158 };
6155 6159
6156 6160 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6157 6161 assert_at_safepoint(true /* should_be_vm_thread */);
6158 6162
6159 6163 if (!free_list_only) {
6160 6164 TearDownRegionSetsClosure cl(&_old_set);
6161 6165 heap_region_iterate(&cl);
6162 6166
6163 6167 // Need to do this after the heap iteration to be able to
6164 6168 // recognize the young regions and ignore them during the iteration.
6165 6169 _young_list->empty_list();
6166 6170 }
6167 6171 _free_list.remove_all();
6168 6172 }
6169 6173
6170 6174 class RebuildRegionSetsClosure : public HeapRegionClosure {
6171 6175 private:
6172 6176 bool _free_list_only;
6173 6177 OldRegionSet* _old_set;
6174 6178 FreeRegionList* _free_list;
6175 6179 size_t _total_used;
6176 6180
6177 6181 public:
6178 6182 RebuildRegionSetsClosure(bool free_list_only,
6179 6183 OldRegionSet* old_set, FreeRegionList* free_list) :
6180 6184 _free_list_only(free_list_only),
6181 6185 _old_set(old_set), _free_list(free_list), _total_used(0) {
6182 6186 assert(_free_list->is_empty(), "pre-condition");
6183 6187 if (!free_list_only) {
6184 6188 assert(_old_set->is_empty(), "pre-condition");
6185 6189 }
6186 6190 }
6187 6191
6188 6192 bool doHeapRegion(HeapRegion* r) {
6189 6193 if (r->continuesHumongous()) {
6190 6194 return false;
6191 6195 }
6192 6196
6193 6197 if (r->is_empty()) {
6194 6198 // Add free regions to the free list
6195 6199 _free_list->add_as_tail(r);
6196 6200 } else if (!_free_list_only) {
6197 6201 assert(!r->is_young(), "we should not come across young regions");
6198 6202
6199 6203 if (r->isHumongous()) {
6200 6204 // We ignore humongous regions, we left the humongous set unchanged
6201 6205 } else {
6202 6206 // The rest should be old, add them to the old set
6203 6207 _old_set->add(r);
6204 6208 }
6205 6209 _total_used += r->used();
6206 6210 }
6207 6211
6208 6212 return false;
6209 6213 }
6210 6214
6211 6215 size_t total_used() {
6212 6216 return _total_used;
6213 6217 }
6214 6218 };
6215 6219
6216 6220 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6217 6221 assert_at_safepoint(true /* should_be_vm_thread */);
6218 6222
6219 6223 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6220 6224 heap_region_iterate(&cl);
6221 6225
6222 6226 if (!free_list_only) {
6223 6227 _summary_bytes_used = cl.total_used();
6224 6228 }
6225 6229 assert(_summary_bytes_used == recalculate_used(),
6226 6230 err_msg("inconsistent _summary_bytes_used, "
6227 6231 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6228 6232 _summary_bytes_used, recalculate_used()));
6229 6233 }
6230 6234
6231 6235 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6232 6236 _refine_cte_cl->set_concurrent(concurrent);
6233 6237 }
6234 6238
6235 6239 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6236 6240 HeapRegion* hr = heap_region_containing(p);
6237 6241 if (hr == NULL) {
6238 6242 return false;
6239 6243 } else {
6240 6244 return hr->is_in(p);
6241 6245 }
6242 6246 }
6243 6247
6244 6248 // Methods for the mutator alloc region
6245 6249
6246 6250 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6247 6251 bool force) {
6248 6252 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6249 6253 assert(!force || g1_policy()->can_expand_young_list(),
6250 6254 "if force is true we should be able to expand the young list");
6251 6255 bool young_list_full = g1_policy()->is_young_list_full();
6252 6256 if (force || !young_list_full) {
6253 6257 HeapRegion* new_alloc_region = new_region(word_size,
6254 6258 false /* do_expand */);
6255 6259 if (new_alloc_region != NULL) {
6256 6260 set_region_short_lived_locked(new_alloc_region);
6257 6261 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6258 6262 return new_alloc_region;
6259 6263 }
6260 6264 }
6261 6265 return NULL;
6262 6266 }
6263 6267
6264 6268 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6265 6269 size_t allocated_bytes) {
6266 6270 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6267 6271 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6268 6272
6269 6273 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6270 6274 _summary_bytes_used += allocated_bytes;
6271 6275 _hr_printer.retire(alloc_region);
6272 6276 // We update the eden sizes here, when the region is retired,
6273 6277 // instead of when it's allocated, since this is the point that its
6274 6278 // used space has been recored in _summary_bytes_used.
6275 6279 g1mm()->update_eden_size();
6276 6280 }
6277 6281
6278 6282 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6279 6283 bool force) {
6280 6284 return _g1h->new_mutator_alloc_region(word_size, force);
6281 6285 }
6282 6286
6283 6287 void G1CollectedHeap::set_par_threads() {
6284 6288 // Don't change the number of workers. Use the value previously set
6285 6289 // in the workgroup.
6286 6290 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6287 6291 uint n_workers = workers()->active_workers();
6288 6292 assert(UseDynamicNumberOfGCThreads ||
6289 6293 n_workers == workers()->total_workers(),
6290 6294 "Otherwise should be using the total number of workers");
6291 6295 if (n_workers == 0) {
6292 6296 assert(false, "Should have been set in prior evacuation pause.");
6293 6297 n_workers = ParallelGCThreads;
6294 6298 workers()->set_active_workers(n_workers);
6295 6299 }
6296 6300 set_par_threads(n_workers);
6297 6301 }
6298 6302
6299 6303 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6300 6304 size_t allocated_bytes) {
6301 6305 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6302 6306 }
6303 6307
6304 6308 // Methods for the GC alloc regions
6305 6309
6306 6310 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6307 6311 uint count,
6308 6312 GCAllocPurpose ap) {
6309 6313 assert(FreeList_lock->owned_by_self(), "pre-condition");
6310 6314
6311 6315 if (count < g1_policy()->max_regions(ap)) {
6312 6316 HeapRegion* new_alloc_region = new_region(word_size,
6313 6317 true /* do_expand */);
6314 6318 if (new_alloc_region != NULL) {
6315 6319 // We really only need to do this for old regions given that we
6316 6320 // should never scan survivors. But it doesn't hurt to do it
6317 6321 // for survivors too.
6318 6322 new_alloc_region->set_saved_mark();
6319 6323 if (ap == GCAllocForSurvived) {
6320 6324 new_alloc_region->set_survivor();
6321 6325 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6322 6326 } else {
6323 6327 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6324 6328 }
6325 6329 bool during_im = g1_policy()->during_initial_mark_pause();
6326 6330 new_alloc_region->note_start_of_copying(during_im);
6327 6331 return new_alloc_region;
6328 6332 } else {
6329 6333 g1_policy()->note_alloc_region_limit_reached(ap);
6330 6334 }
6331 6335 }
6332 6336 return NULL;
6333 6337 }
6334 6338
6335 6339 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6336 6340 size_t allocated_bytes,
6337 6341 GCAllocPurpose ap) {
6338 6342 bool during_im = g1_policy()->during_initial_mark_pause();
6339 6343 alloc_region->note_end_of_copying(during_im);
6340 6344 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6341 6345 if (ap == GCAllocForSurvived) {
6342 6346 young_list()->add_survivor_region(alloc_region);
6343 6347 } else {
6344 6348 _old_set.add(alloc_region);
6345 6349 }
6346 6350 _hr_printer.retire(alloc_region);
6347 6351 }
6348 6352
6349 6353 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6350 6354 bool force) {
6351 6355 assert(!force, "not supported for GC alloc regions");
6352 6356 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6353 6357 }
6354 6358
6355 6359 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6356 6360 size_t allocated_bytes) {
6357 6361 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6358 6362 GCAllocForSurvived);
6359 6363 }
6360 6364
6361 6365 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6362 6366 bool force) {
6363 6367 assert(!force, "not supported for GC alloc regions");
6364 6368 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6365 6369 }
6366 6370
6367 6371 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6368 6372 size_t allocated_bytes) {
6369 6373 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6370 6374 GCAllocForTenured);
6371 6375 }
6372 6376 // Heap region set verification
6373 6377
6374 6378 class VerifyRegionListsClosure : public HeapRegionClosure {
6375 6379 private:
6376 6380 FreeRegionList* _free_list;
6377 6381 OldRegionSet* _old_set;
6378 6382 HumongousRegionSet* _humongous_set;
6379 6383 uint _region_count;
6380 6384
6381 6385 public:
6382 6386 VerifyRegionListsClosure(OldRegionSet* old_set,
6383 6387 HumongousRegionSet* humongous_set,
6384 6388 FreeRegionList* free_list) :
6385 6389 _old_set(old_set), _humongous_set(humongous_set),
6386 6390 _free_list(free_list), _region_count(0) { }
6387 6391
6388 6392 uint region_count() { return _region_count; }
6389 6393
6390 6394 bool doHeapRegion(HeapRegion* hr) {
6391 6395 _region_count += 1;
6392 6396
6393 6397 if (hr->continuesHumongous()) {
6394 6398 return false;
6395 6399 }
6396 6400
6397 6401 if (hr->is_young()) {
6398 6402 // TODO
6399 6403 } else if (hr->startsHumongous()) {
6400 6404 _humongous_set->verify_next_region(hr);
6401 6405 } else if (hr->is_empty()) {
6402 6406 _free_list->verify_next_region(hr);
6403 6407 } else {
6404 6408 _old_set->verify_next_region(hr);
6405 6409 }
6406 6410 return false;
6407 6411 }
6408 6412 };
6409 6413
6410 6414 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6411 6415 HeapWord* bottom) {
6412 6416 HeapWord* end = bottom + HeapRegion::GrainWords;
6413 6417 MemRegion mr(bottom, end);
6414 6418 assert(_g1_reserved.contains(mr), "invariant");
6415 6419 // This might return NULL if the allocation fails
6416 6420 return new HeapRegion(hrs_index, _bot_shared, mr);
6417 6421 }
6418 6422
6419 6423 void G1CollectedHeap::verify_region_sets() {
6420 6424 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6421 6425
6422 6426 // First, check the explicit lists.
6423 6427 _free_list.verify();
6424 6428 {
6425 6429 // Given that a concurrent operation might be adding regions to
6426 6430 // the secondary free list we have to take the lock before
6427 6431 // verifying it.
6428 6432 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6429 6433 _secondary_free_list.verify();
6430 6434 }
6431 6435 _old_set.verify();
6432 6436 _humongous_set.verify();
6433 6437
6434 6438 // If a concurrent region freeing operation is in progress it will
6435 6439 // be difficult to correctly attributed any free regions we come
6436 6440 // across to the correct free list given that they might belong to
6437 6441 // one of several (free_list, secondary_free_list, any local lists,
6438 6442 // etc.). So, if that's the case we will skip the rest of the
6439 6443 // verification operation. Alternatively, waiting for the concurrent
6440 6444 // operation to complete will have a non-trivial effect on the GC's
6441 6445 // operation (no concurrent operation will last longer than the
6442 6446 // interval between two calls to verification) and it might hide
6443 6447 // any issues that we would like to catch during testing.
6444 6448 if (free_regions_coming()) {
6445 6449 return;
6446 6450 }
6447 6451
6448 6452 // Make sure we append the secondary_free_list on the free_list so
6449 6453 // that all free regions we will come across can be safely
6450 6454 // attributed to the free_list.
6451 6455 append_secondary_free_list_if_not_empty_with_lock();
6452 6456
6453 6457 // Finally, make sure that the region accounting in the lists is
6454 6458 // consistent with what we see in the heap.
6455 6459 _old_set.verify_start();
6456 6460 _humongous_set.verify_start();
6457 6461 _free_list.verify_start();
6458 6462
6459 6463 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6460 6464 heap_region_iterate(&cl);
6461 6465
6462 6466 _old_set.verify_end();
6463 6467 _humongous_set.verify_end();
6464 6468 _free_list.verify_end();
6465 6469 }
↓ open down ↓ |
4365 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX