Print this page
rev 3856 : 8004170: G1: Verbose GC output is not getting flushed to log file using JDK 8
Summary: Add flushes to G1CollectedHeap::log_gc_footer() and TraceCPUTime destructor.
Reviewed-by:
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp
+++ new/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp
1 1 /*
2 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "code/icBuffer.hpp"
27 27 #include "gc_implementation/g1/bufferingOopClosure.hpp"
28 28 #include "gc_implementation/g1/concurrentG1Refine.hpp"
29 29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
30 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
32 32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
33 33 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
34 34 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
35 35 #include "gc_implementation/g1/g1EvacFailure.hpp"
36 36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
37 37 #include "gc_implementation/g1/g1Log.hpp"
38 38 #include "gc_implementation/g1/g1MarkSweep.hpp"
39 39 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
40 40 #include "gc_implementation/g1/g1RemSet.inline.hpp"
41 41 #include "gc_implementation/g1/heapRegion.inline.hpp"
42 42 #include "gc_implementation/g1/heapRegionRemSet.hpp"
43 43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
44 44 #include "gc_implementation/g1/vm_operations_g1.hpp"
45 45 #include "gc_implementation/shared/isGCActiveMark.hpp"
46 46 #include "memory/gcLocker.inline.hpp"
47 47 #include "memory/genOopClosures.inline.hpp"
48 48 #include "memory/generationSpec.hpp"
49 49 #include "memory/referenceProcessor.hpp"
50 50 #include "oops/oop.inline.hpp"
51 51 #include "oops/oop.pcgc.inline.hpp"
52 52 #include "runtime/aprofiler.hpp"
53 53 #include "runtime/vmThread.hpp"
54 54
55 55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
56 56
57 57 // turn it on so that the contents of the young list (scan-only /
58 58 // to-be-collected) are printed at "strategic" points before / during
59 59 // / after the collection --- this is useful for debugging
60 60 #define YOUNG_LIST_VERBOSE 0
61 61 // CURRENT STATUS
62 62 // This file is under construction. Search for "FIXME".
63 63
64 64 // INVARIANTS/NOTES
65 65 //
66 66 // All allocation activity covered by the G1CollectedHeap interface is
67 67 // serialized by acquiring the HeapLock. This happens in mem_allocate
68 68 // and allocate_new_tlab, which are the "entry" points to the
69 69 // allocation code from the rest of the JVM. (Note that this does not
70 70 // apply to TLAB allocation, which is not part of this interface: it
71 71 // is done by clients of this interface.)
72 72
73 73 // Notes on implementation of parallelism in different tasks.
74 74 //
75 75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
76 76 // The number of GC workers is passed to heap_region_par_iterate_chunked().
77 77 // It does use run_task() which sets _n_workers in the task.
78 78 // G1ParTask executes g1_process_strong_roots() ->
79 79 // SharedHeap::process_strong_roots() which calls eventuall to
80 80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
81 81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
82 82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
83 83 //
84 84
85 85 // Local to this file.
86 86
87 87 class RefineCardTableEntryClosure: public CardTableEntryClosure {
88 88 SuspendibleThreadSet* _sts;
89 89 G1RemSet* _g1rs;
90 90 ConcurrentG1Refine* _cg1r;
91 91 bool _concurrent;
92 92 public:
93 93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
94 94 G1RemSet* g1rs,
95 95 ConcurrentG1Refine* cg1r) :
96 96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
97 97 {}
98 98 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
99 99 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
100 100 // This path is executed by the concurrent refine or mutator threads,
101 101 // concurrently, and so we do not care if card_ptr contains references
102 102 // that point into the collection set.
103 103 assert(!oops_into_cset, "should be");
104 104
105 105 if (_concurrent && _sts->should_yield()) {
106 106 // Caller will actually yield.
107 107 return false;
108 108 }
109 109 // Otherwise, we finished successfully; return true.
110 110 return true;
111 111 }
112 112 void set_concurrent(bool b) { _concurrent = b; }
113 113 };
114 114
115 115
116 116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
117 117 int _calls;
118 118 G1CollectedHeap* _g1h;
119 119 CardTableModRefBS* _ctbs;
120 120 int _histo[256];
121 121 public:
122 122 ClearLoggedCardTableEntryClosure() :
123 123 _calls(0)
124 124 {
125 125 _g1h = G1CollectedHeap::heap();
126 126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
127 127 for (int i = 0; i < 256; i++) _histo[i] = 0;
128 128 }
129 129 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
130 130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
131 131 _calls++;
132 132 unsigned char* ujb = (unsigned char*)card_ptr;
133 133 int ind = (int)(*ujb);
134 134 _histo[ind]++;
135 135 *card_ptr = -1;
136 136 }
137 137 return true;
138 138 }
139 139 int calls() { return _calls; }
140 140 void print_histo() {
141 141 gclog_or_tty->print_cr("Card table value histogram:");
142 142 for (int i = 0; i < 256; i++) {
143 143 if (_histo[i] != 0) {
144 144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
145 145 }
146 146 }
147 147 }
148 148 };
149 149
150 150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
151 151 int _calls;
152 152 G1CollectedHeap* _g1h;
153 153 CardTableModRefBS* _ctbs;
154 154 public:
155 155 RedirtyLoggedCardTableEntryClosure() :
156 156 _calls(0)
157 157 {
158 158 _g1h = G1CollectedHeap::heap();
159 159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
160 160 }
161 161 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
162 162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
163 163 _calls++;
164 164 *card_ptr = 0;
165 165 }
166 166 return true;
167 167 }
168 168 int calls() { return _calls; }
169 169 };
170 170
171 171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
172 172 public:
173 173 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
174 174 *card_ptr = CardTableModRefBS::dirty_card_val();
175 175 return true;
176 176 }
177 177 };
178 178
179 179 YoungList::YoungList(G1CollectedHeap* g1h) :
180 180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
181 181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
182 182 guarantee(check_list_empty(false), "just making sure...");
183 183 }
184 184
185 185 void YoungList::push_region(HeapRegion *hr) {
186 186 assert(!hr->is_young(), "should not already be young");
187 187 assert(hr->get_next_young_region() == NULL, "cause it should!");
188 188
189 189 hr->set_next_young_region(_head);
190 190 _head = hr;
191 191
192 192 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
193 193 ++_length;
194 194 }
195 195
196 196 void YoungList::add_survivor_region(HeapRegion* hr) {
197 197 assert(hr->is_survivor(), "should be flagged as survivor region");
198 198 assert(hr->get_next_young_region() == NULL, "cause it should!");
199 199
200 200 hr->set_next_young_region(_survivor_head);
201 201 if (_survivor_head == NULL) {
202 202 _survivor_tail = hr;
203 203 }
204 204 _survivor_head = hr;
205 205 ++_survivor_length;
206 206 }
207 207
208 208 void YoungList::empty_list(HeapRegion* list) {
209 209 while (list != NULL) {
210 210 HeapRegion* next = list->get_next_young_region();
211 211 list->set_next_young_region(NULL);
212 212 list->uninstall_surv_rate_group();
213 213 list->set_not_young();
214 214 list = next;
215 215 }
216 216 }
217 217
218 218 void YoungList::empty_list() {
219 219 assert(check_list_well_formed(), "young list should be well formed");
220 220
221 221 empty_list(_head);
222 222 _head = NULL;
223 223 _length = 0;
224 224
225 225 empty_list(_survivor_head);
226 226 _survivor_head = NULL;
227 227 _survivor_tail = NULL;
228 228 _survivor_length = 0;
229 229
230 230 _last_sampled_rs_lengths = 0;
231 231
232 232 assert(check_list_empty(false), "just making sure...");
233 233 }
234 234
235 235 bool YoungList::check_list_well_formed() {
236 236 bool ret = true;
237 237
238 238 uint length = 0;
239 239 HeapRegion* curr = _head;
240 240 HeapRegion* last = NULL;
241 241 while (curr != NULL) {
242 242 if (!curr->is_young()) {
243 243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
244 244 "incorrectly tagged (y: %d, surv: %d)",
245 245 curr->bottom(), curr->end(),
246 246 curr->is_young(), curr->is_survivor());
247 247 ret = false;
248 248 }
249 249 ++length;
250 250 last = curr;
251 251 curr = curr->get_next_young_region();
252 252 }
253 253 ret = ret && (length == _length);
254 254
255 255 if (!ret) {
256 256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
257 257 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
258 258 length, _length);
259 259 }
260 260
261 261 return ret;
262 262 }
263 263
264 264 bool YoungList::check_list_empty(bool check_sample) {
265 265 bool ret = true;
266 266
267 267 if (_length != 0) {
268 268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
269 269 _length);
270 270 ret = false;
271 271 }
272 272 if (check_sample && _last_sampled_rs_lengths != 0) {
273 273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
274 274 ret = false;
275 275 }
276 276 if (_head != NULL) {
277 277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
278 278 ret = false;
279 279 }
280 280 if (!ret) {
281 281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
282 282 }
283 283
284 284 return ret;
285 285 }
286 286
287 287 void
288 288 YoungList::rs_length_sampling_init() {
289 289 _sampled_rs_lengths = 0;
290 290 _curr = _head;
291 291 }
292 292
293 293 bool
294 294 YoungList::rs_length_sampling_more() {
295 295 return _curr != NULL;
296 296 }
297 297
298 298 void
299 299 YoungList::rs_length_sampling_next() {
300 300 assert( _curr != NULL, "invariant" );
301 301 size_t rs_length = _curr->rem_set()->occupied();
302 302
303 303 _sampled_rs_lengths += rs_length;
304 304
305 305 // The current region may not yet have been added to the
306 306 // incremental collection set (it gets added when it is
307 307 // retired as the current allocation region).
308 308 if (_curr->in_collection_set()) {
309 309 // Update the collection set policy information for this region
310 310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
311 311 }
312 312
313 313 _curr = _curr->get_next_young_region();
314 314 if (_curr == NULL) {
315 315 _last_sampled_rs_lengths = _sampled_rs_lengths;
316 316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
317 317 }
318 318 }
319 319
320 320 void
321 321 YoungList::reset_auxilary_lists() {
322 322 guarantee( is_empty(), "young list should be empty" );
323 323 assert(check_list_well_formed(), "young list should be well formed");
324 324
325 325 // Add survivor regions to SurvRateGroup.
326 326 _g1h->g1_policy()->note_start_adding_survivor_regions();
327 327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
328 328
329 329 int young_index_in_cset = 0;
330 330 for (HeapRegion* curr = _survivor_head;
331 331 curr != NULL;
332 332 curr = curr->get_next_young_region()) {
333 333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
334 334
335 335 // The region is a non-empty survivor so let's add it to
336 336 // the incremental collection set for the next evacuation
337 337 // pause.
338 338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
339 339 young_index_in_cset += 1;
340 340 }
341 341 assert((uint) young_index_in_cset == _survivor_length, "post-condition");
342 342 _g1h->g1_policy()->note_stop_adding_survivor_regions();
343 343
344 344 _head = _survivor_head;
345 345 _length = _survivor_length;
346 346 if (_survivor_head != NULL) {
347 347 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 348 assert(_survivor_length > 0, "invariant");
349 349 _survivor_tail->set_next_young_region(NULL);
350 350 }
351 351
352 352 // Don't clear the survivor list handles until the start of
353 353 // the next evacuation pause - we need it in order to re-tag
354 354 // the survivor regions from this evacuation pause as 'young'
355 355 // at the start of the next.
356 356
357 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
358 358
359 359 assert(check_list_well_formed(), "young list should be well formed");
360 360 }
361 361
362 362 void YoungList::print() {
363 363 HeapRegion* lists[] = {_head, _survivor_head};
364 364 const char* names[] = {"YOUNG", "SURVIVOR"};
365 365
366 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 368 HeapRegion *curr = lists[list];
369 369 if (curr == NULL)
370 370 gclog_or_tty->print_cr(" empty");
371 371 while (curr != NULL) {
372 372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
373 373 HR_FORMAT_PARAMS(curr),
374 374 curr->prev_top_at_mark_start(),
375 375 curr->next_top_at_mark_start(),
376 376 curr->age_in_surv_rate_group_cond());
377 377 curr = curr->get_next_young_region();
378 378 }
379 379 }
380 380
381 381 gclog_or_tty->print_cr("");
382 382 }
383 383
384 384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
385 385 {
386 386 // Claim the right to put the region on the dirty cards region list
387 387 // by installing a self pointer.
388 388 HeapRegion* next = hr->get_next_dirty_cards_region();
389 389 if (next == NULL) {
390 390 HeapRegion* res = (HeapRegion*)
391 391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
392 392 NULL);
393 393 if (res == NULL) {
394 394 HeapRegion* head;
395 395 do {
396 396 // Put the region to the dirty cards region list.
397 397 head = _dirty_cards_region_list;
398 398 next = (HeapRegion*)
399 399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
400 400 if (next == head) {
401 401 assert(hr->get_next_dirty_cards_region() == hr,
402 402 "hr->get_next_dirty_cards_region() != hr");
403 403 if (next == NULL) {
404 404 // The last region in the list points to itself.
405 405 hr->set_next_dirty_cards_region(hr);
406 406 } else {
407 407 hr->set_next_dirty_cards_region(next);
408 408 }
409 409 }
410 410 } while (next != head);
411 411 }
412 412 }
413 413 }
414 414
415 415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
416 416 {
417 417 HeapRegion* head;
418 418 HeapRegion* hr;
419 419 do {
420 420 head = _dirty_cards_region_list;
421 421 if (head == NULL) {
422 422 return NULL;
423 423 }
424 424 HeapRegion* new_head = head->get_next_dirty_cards_region();
425 425 if (head == new_head) {
426 426 // The last region.
427 427 new_head = NULL;
428 428 }
429 429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
430 430 head);
431 431 } while (hr != head);
432 432 assert(hr != NULL, "invariant");
433 433 hr->set_next_dirty_cards_region(NULL);
434 434 return hr;
435 435 }
436 436
437 437 void G1CollectedHeap::stop_conc_gc_threads() {
438 438 _cg1r->stop();
439 439 _cmThread->stop();
440 440 }
441 441
442 442 #ifdef ASSERT
443 443 // A region is added to the collection set as it is retired
444 444 // so an address p can point to a region which will be in the
445 445 // collection set but has not yet been retired. This method
446 446 // therefore is only accurate during a GC pause after all
447 447 // regions have been retired. It is used for debugging
448 448 // to check if an nmethod has references to objects that can
449 449 // be move during a partial collection. Though it can be
450 450 // inaccurate, it is sufficient for G1 because the conservative
451 451 // implementation of is_scavengable() for G1 will indicate that
452 452 // all nmethods must be scanned during a partial collection.
453 453 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
454 454 HeapRegion* hr = heap_region_containing(p);
455 455 return hr != NULL && hr->in_collection_set();
456 456 }
457 457 #endif
458 458
459 459 // Returns true if the reference points to an object that
460 460 // can move in an incremental collecction.
461 461 bool G1CollectedHeap::is_scavengable(const void* p) {
462 462 G1CollectedHeap* g1h = G1CollectedHeap::heap();
463 463 G1CollectorPolicy* g1p = g1h->g1_policy();
464 464 HeapRegion* hr = heap_region_containing(p);
465 465 if (hr == NULL) {
466 466 // null
467 467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p));
468 468 return false;
469 469 } else {
470 470 return !hr->isHumongous();
471 471 }
472 472 }
473 473
474 474 void G1CollectedHeap::check_ct_logs_at_safepoint() {
475 475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
476 476 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
477 477
478 478 // Count the dirty cards at the start.
479 479 CountNonCleanMemRegionClosure count1(this);
480 480 ct_bs->mod_card_iterate(&count1);
481 481 int orig_count = count1.n();
482 482
483 483 // First clear the logged cards.
484 484 ClearLoggedCardTableEntryClosure clear;
485 485 dcqs.set_closure(&clear);
486 486 dcqs.apply_closure_to_all_completed_buffers();
487 487 dcqs.iterate_closure_all_threads(false);
488 488 clear.print_histo();
489 489
490 490 // Now ensure that there's no dirty cards.
491 491 CountNonCleanMemRegionClosure count2(this);
492 492 ct_bs->mod_card_iterate(&count2);
493 493 if (count2.n() != 0) {
494 494 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
495 495 count2.n(), orig_count);
496 496 }
497 497 guarantee(count2.n() == 0, "Card table should be clean.");
498 498
499 499 RedirtyLoggedCardTableEntryClosure redirty;
500 500 JavaThread::dirty_card_queue_set().set_closure(&redirty);
501 501 dcqs.apply_closure_to_all_completed_buffers();
502 502 dcqs.iterate_closure_all_threads(false);
503 503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
504 504 clear.calls(), orig_count);
505 505 guarantee(redirty.calls() == clear.calls(),
506 506 "Or else mechanism is broken.");
507 507
508 508 CountNonCleanMemRegionClosure count3(this);
509 509 ct_bs->mod_card_iterate(&count3);
510 510 if (count3.n() != orig_count) {
511 511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
512 512 orig_count, count3.n());
513 513 guarantee(count3.n() >= orig_count, "Should have restored them all.");
514 514 }
515 515
516 516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
517 517 }
518 518
519 519 // Private class members.
520 520
521 521 G1CollectedHeap* G1CollectedHeap::_g1h;
522 522
523 523 // Private methods.
524 524
525 525 HeapRegion*
526 526 G1CollectedHeap::new_region_try_secondary_free_list() {
527 527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
528 528 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
529 529 if (!_secondary_free_list.is_empty()) {
530 530 if (G1ConcRegionFreeingVerbose) {
531 531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
532 532 "secondary_free_list has %u entries",
533 533 _secondary_free_list.length());
534 534 }
535 535 // It looks as if there are free regions available on the
536 536 // secondary_free_list. Let's move them to the free_list and try
537 537 // again to allocate from it.
538 538 append_secondary_free_list();
539 539
540 540 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
541 541 "empty we should have moved at least one entry to the free_list");
542 542 HeapRegion* res = _free_list.remove_head();
543 543 if (G1ConcRegionFreeingVerbose) {
544 544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
545 545 "allocated "HR_FORMAT" from secondary_free_list",
546 546 HR_FORMAT_PARAMS(res));
547 547 }
548 548 return res;
549 549 }
550 550
551 551 // Wait here until we get notifed either when (a) there are no
552 552 // more free regions coming or (b) some regions have been moved on
553 553 // the secondary_free_list.
554 554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
555 555 }
556 556
557 557 if (G1ConcRegionFreeingVerbose) {
558 558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
559 559 "could not allocate from secondary_free_list");
560 560 }
561 561 return NULL;
562 562 }
563 563
564 564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
565 565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
566 566 "the only time we use this to allocate a humongous region is "
567 567 "when we are allocating a single humongous region");
568 568
569 569 HeapRegion* res;
570 570 if (G1StressConcRegionFreeing) {
571 571 if (!_secondary_free_list.is_empty()) {
572 572 if (G1ConcRegionFreeingVerbose) {
573 573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
574 574 "forced to look at the secondary_free_list");
575 575 }
576 576 res = new_region_try_secondary_free_list();
577 577 if (res != NULL) {
578 578 return res;
579 579 }
580 580 }
581 581 }
582 582 res = _free_list.remove_head_or_null();
583 583 if (res == NULL) {
584 584 if (G1ConcRegionFreeingVerbose) {
585 585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
586 586 "res == NULL, trying the secondary_free_list");
587 587 }
588 588 res = new_region_try_secondary_free_list();
589 589 }
590 590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
591 591 // Currently, only attempts to allocate GC alloc regions set
592 592 // do_expand to true. So, we should only reach here during a
593 593 // safepoint. If this assumption changes we might have to
594 594 // reconsider the use of _expand_heap_after_alloc_failure.
595 595 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
596 596
597 597 ergo_verbose1(ErgoHeapSizing,
598 598 "attempt heap expansion",
599 599 ergo_format_reason("region allocation request failed")
600 600 ergo_format_byte("allocation request"),
601 601 word_size * HeapWordSize);
602 602 if (expand(word_size * HeapWordSize)) {
603 603 // Given that expand() succeeded in expanding the heap, and we
604 604 // always expand the heap by an amount aligned to the heap
605 605 // region size, the free list should in theory not be empty. So
606 606 // it would probably be OK to use remove_head(). But the extra
607 607 // check for NULL is unlikely to be a performance issue here (we
608 608 // just expanded the heap!) so let's just be conservative and
609 609 // use remove_head_or_null().
610 610 res = _free_list.remove_head_or_null();
611 611 } else {
612 612 _expand_heap_after_alloc_failure = false;
613 613 }
614 614 }
615 615 return res;
616 616 }
617 617
618 618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions,
619 619 size_t word_size) {
620 620 assert(isHumongous(word_size), "word_size should be humongous");
621 621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
622 622
623 623 uint first = G1_NULL_HRS_INDEX;
624 624 if (num_regions == 1) {
625 625 // Only one region to allocate, no need to go through the slower
626 626 // path. The caller will attempt the expasion if this fails, so
627 627 // let's not try to expand here too.
628 628 HeapRegion* hr = new_region(word_size, false /* do_expand */);
629 629 if (hr != NULL) {
630 630 first = hr->hrs_index();
631 631 } else {
632 632 first = G1_NULL_HRS_INDEX;
633 633 }
634 634 } else {
635 635 // We can't allocate humongous regions while cleanupComplete() is
636 636 // running, since some of the regions we find to be empty might not
637 637 // yet be added to the free list and it is not straightforward to
638 638 // know which list they are on so that we can remove them. Note
639 639 // that we only need to do this if we need to allocate more than
640 640 // one region to satisfy the current humongous allocation
641 641 // request. If we are only allocating one region we use the common
642 642 // region allocation code (see above).
643 643 wait_while_free_regions_coming();
644 644 append_secondary_free_list_if_not_empty_with_lock();
645 645
646 646 if (free_regions() >= num_regions) {
647 647 first = _hrs.find_contiguous(num_regions);
648 648 if (first != G1_NULL_HRS_INDEX) {
649 649 for (uint i = first; i < first + num_regions; ++i) {
650 650 HeapRegion* hr = region_at(i);
651 651 assert(hr->is_empty(), "sanity");
652 652 assert(is_on_master_free_list(hr), "sanity");
653 653 hr->set_pending_removal(true);
654 654 }
655 655 _free_list.remove_all_pending(num_regions);
656 656 }
657 657 }
658 658 }
659 659 return first;
660 660 }
661 661
662 662 HeapWord*
663 663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
664 664 uint num_regions,
665 665 size_t word_size) {
666 666 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
667 667 assert(isHumongous(word_size), "word_size should be humongous");
668 668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
669 669
670 670 // Index of last region in the series + 1.
671 671 uint last = first + num_regions;
672 672
673 673 // We need to initialize the region(s) we just discovered. This is
674 674 // a bit tricky given that it can happen concurrently with
675 675 // refinement threads refining cards on these regions and
676 676 // potentially wanting to refine the BOT as they are scanning
677 677 // those cards (this can happen shortly after a cleanup; see CR
678 678 // 6991377). So we have to set up the region(s) carefully and in
679 679 // a specific order.
680 680
681 681 // The word size sum of all the regions we will allocate.
682 682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
683 683 assert(word_size <= word_size_sum, "sanity");
684 684
685 685 // This will be the "starts humongous" region.
686 686 HeapRegion* first_hr = region_at(first);
687 687 // The header of the new object will be placed at the bottom of
688 688 // the first region.
689 689 HeapWord* new_obj = first_hr->bottom();
690 690 // This will be the new end of the first region in the series that
691 691 // should also match the end of the last region in the seriers.
692 692 HeapWord* new_end = new_obj + word_size_sum;
693 693 // This will be the new top of the first region that will reflect
694 694 // this allocation.
695 695 HeapWord* new_top = new_obj + word_size;
696 696
697 697 // First, we need to zero the header of the space that we will be
698 698 // allocating. When we update top further down, some refinement
699 699 // threads might try to scan the region. By zeroing the header we
700 700 // ensure that any thread that will try to scan the region will
701 701 // come across the zero klass word and bail out.
702 702 //
703 703 // NOTE: It would not have been correct to have used
704 704 // CollectedHeap::fill_with_object() and make the space look like
705 705 // an int array. The thread that is doing the allocation will
706 706 // later update the object header to a potentially different array
707 707 // type and, for a very short period of time, the klass and length
708 708 // fields will be inconsistent. This could cause a refinement
709 709 // thread to calculate the object size incorrectly.
710 710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
711 711
712 712 // We will set up the first region as "starts humongous". This
713 713 // will also update the BOT covering all the regions to reflect
714 714 // that there is a single object that starts at the bottom of the
715 715 // first region.
716 716 first_hr->set_startsHumongous(new_top, new_end);
717 717
718 718 // Then, if there are any, we will set up the "continues
719 719 // humongous" regions.
720 720 HeapRegion* hr = NULL;
721 721 for (uint i = first + 1; i < last; ++i) {
722 722 hr = region_at(i);
723 723 hr->set_continuesHumongous(first_hr);
724 724 }
725 725 // If we have "continues humongous" regions (hr != NULL), then the
726 726 // end of the last one should match new_end.
727 727 assert(hr == NULL || hr->end() == new_end, "sanity");
728 728
729 729 // Up to this point no concurrent thread would have been able to
730 730 // do any scanning on any region in this series. All the top
731 731 // fields still point to bottom, so the intersection between
732 732 // [bottom,top] and [card_start,card_end] will be empty. Before we
733 733 // update the top fields, we'll do a storestore to make sure that
734 734 // no thread sees the update to top before the zeroing of the
735 735 // object header and the BOT initialization.
736 736 OrderAccess::storestore();
737 737
738 738 // Now that the BOT and the object header have been initialized,
739 739 // we can update top of the "starts humongous" region.
740 740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
741 741 "new_top should be in this region");
742 742 first_hr->set_top(new_top);
743 743 if (_hr_printer.is_active()) {
744 744 HeapWord* bottom = first_hr->bottom();
745 745 HeapWord* end = first_hr->orig_end();
746 746 if ((first + 1) == last) {
747 747 // the series has a single humongous region
748 748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
749 749 } else {
750 750 // the series has more than one humongous regions
751 751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
752 752 }
753 753 }
754 754
755 755 // Now, we will update the top fields of the "continues humongous"
756 756 // regions. The reason we need to do this is that, otherwise,
757 757 // these regions would look empty and this will confuse parts of
758 758 // G1. For example, the code that looks for a consecutive number
759 759 // of empty regions will consider them empty and try to
760 760 // re-allocate them. We can extend is_empty() to also include
761 761 // !continuesHumongous(), but it is easier to just update the top
762 762 // fields here. The way we set top for all regions (i.e., top ==
763 763 // end for all regions but the last one, top == new_top for the
764 764 // last one) is actually used when we will free up the humongous
765 765 // region in free_humongous_region().
766 766 hr = NULL;
767 767 for (uint i = first + 1; i < last; ++i) {
768 768 hr = region_at(i);
769 769 if ((i + 1) == last) {
770 770 // last continues humongous region
771 771 assert(hr->bottom() < new_top && new_top <= hr->end(),
772 772 "new_top should fall on this region");
773 773 hr->set_top(new_top);
774 774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
775 775 } else {
776 776 // not last one
777 777 assert(new_top > hr->end(), "new_top should be above this region");
778 778 hr->set_top(hr->end());
779 779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
780 780 }
781 781 }
782 782 // If we have continues humongous regions (hr != NULL), then the
783 783 // end of the last one should match new_end and its top should
784 784 // match new_top.
785 785 assert(hr == NULL ||
786 786 (hr->end() == new_end && hr->top() == new_top), "sanity");
787 787
788 788 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
789 789 _summary_bytes_used += first_hr->used();
790 790 _humongous_set.add(first_hr);
791 791
792 792 return new_obj;
793 793 }
794 794
795 795 // If could fit into free regions w/o expansion, try.
796 796 // Otherwise, if can expand, do so.
797 797 // Otherwise, if using ex regions might help, try with ex given back.
798 798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
799 799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
800 800
801 801 verify_region_sets_optional();
802 802
803 803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords);
804 804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords);
805 805 uint x_num = expansion_regions();
806 806 uint fs = _hrs.free_suffix();
807 807 uint first = humongous_obj_allocate_find_first(num_regions, word_size);
808 808 if (first == G1_NULL_HRS_INDEX) {
809 809 // The only thing we can do now is attempt expansion.
810 810 if (fs + x_num >= num_regions) {
811 811 // If the number of regions we're trying to allocate for this
812 812 // object is at most the number of regions in the free suffix,
813 813 // then the call to humongous_obj_allocate_find_first() above
814 814 // should have succeeded and we wouldn't be here.
815 815 //
816 816 // We should only be trying to expand when the free suffix is
817 817 // not sufficient for the object _and_ we have some expansion
818 818 // room available.
819 819 assert(num_regions > fs, "earlier allocation should have succeeded");
820 820
821 821 ergo_verbose1(ErgoHeapSizing,
822 822 "attempt heap expansion",
823 823 ergo_format_reason("humongous allocation request failed")
824 824 ergo_format_byte("allocation request"),
825 825 word_size * HeapWordSize);
826 826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
827 827 // Even though the heap was expanded, it might not have
828 828 // reached the desired size. So, we cannot assume that the
829 829 // allocation will succeed.
830 830 first = humongous_obj_allocate_find_first(num_regions, word_size);
831 831 }
832 832 }
833 833 }
834 834
835 835 HeapWord* result = NULL;
836 836 if (first != G1_NULL_HRS_INDEX) {
837 837 result =
838 838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
839 839 assert(result != NULL, "it should always return a valid result");
840 840
841 841 // A successful humongous object allocation changes the used space
842 842 // information of the old generation so we need to recalculate the
843 843 // sizes and update the jstat counters here.
844 844 g1mm()->update_sizes();
845 845 }
846 846
847 847 verify_region_sets_optional();
848 848
849 849 return result;
850 850 }
851 851
852 852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
853 853 assert_heap_not_locked_and_not_at_safepoint();
854 854 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
855 855
856 856 unsigned int dummy_gc_count_before;
857 857 return attempt_allocation(word_size, &dummy_gc_count_before);
858 858 }
859 859
860 860 HeapWord*
861 861 G1CollectedHeap::mem_allocate(size_t word_size,
862 862 bool* gc_overhead_limit_was_exceeded) {
863 863 assert_heap_not_locked_and_not_at_safepoint();
864 864
865 865 // Loop until the allocation is satisified, or unsatisfied after GC.
866 866 for (int try_count = 1; /* we'll return */; try_count += 1) {
867 867 unsigned int gc_count_before;
868 868
869 869 HeapWord* result = NULL;
870 870 if (!isHumongous(word_size)) {
871 871 result = attempt_allocation(word_size, &gc_count_before);
872 872 } else {
873 873 result = attempt_allocation_humongous(word_size, &gc_count_before);
874 874 }
875 875 if (result != NULL) {
876 876 return result;
877 877 }
878 878
879 879 // Create the garbage collection operation...
880 880 VM_G1CollectForAllocation op(gc_count_before, word_size);
881 881 // ...and get the VM thread to execute it.
882 882 VMThread::execute(&op);
883 883
884 884 if (op.prologue_succeeded() && op.pause_succeeded()) {
885 885 // If the operation was successful we'll return the result even
886 886 // if it is NULL. If the allocation attempt failed immediately
887 887 // after a Full GC, it's unlikely we'll be able to allocate now.
888 888 HeapWord* result = op.result();
889 889 if (result != NULL && !isHumongous(word_size)) {
890 890 // Allocations that take place on VM operations do not do any
891 891 // card dirtying and we have to do it here. We only have to do
892 892 // this for non-humongous allocations, though.
893 893 dirty_young_block(result, word_size);
894 894 }
895 895 return result;
896 896 } else {
897 897 assert(op.result() == NULL,
898 898 "the result should be NULL if the VM op did not succeed");
899 899 }
900 900
901 901 // Give a warning if we seem to be looping forever.
902 902 if ((QueuedAllocationWarningCount > 0) &&
903 903 (try_count % QueuedAllocationWarningCount == 0)) {
904 904 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
905 905 }
906 906 }
907 907
908 908 ShouldNotReachHere();
909 909 return NULL;
910 910 }
911 911
912 912 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
913 913 unsigned int *gc_count_before_ret) {
914 914 // Make sure you read the note in attempt_allocation_humongous().
915 915
916 916 assert_heap_not_locked_and_not_at_safepoint();
917 917 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
918 918 "be called for humongous allocation requests");
919 919
920 920 // We should only get here after the first-level allocation attempt
921 921 // (attempt_allocation()) failed to allocate.
922 922
923 923 // We will loop until a) we manage to successfully perform the
924 924 // allocation or b) we successfully schedule a collection which
925 925 // fails to perform the allocation. b) is the only case when we'll
926 926 // return NULL.
927 927 HeapWord* result = NULL;
928 928 for (int try_count = 1; /* we'll return */; try_count += 1) {
929 929 bool should_try_gc;
930 930 unsigned int gc_count_before;
931 931
932 932 {
933 933 MutexLockerEx x(Heap_lock);
934 934
935 935 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
936 936 false /* bot_updates */);
937 937 if (result != NULL) {
938 938 return result;
939 939 }
940 940
941 941 // If we reach here, attempt_allocation_locked() above failed to
942 942 // allocate a new region. So the mutator alloc region should be NULL.
943 943 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
944 944
945 945 if (GC_locker::is_active_and_needs_gc()) {
946 946 if (g1_policy()->can_expand_young_list()) {
947 947 // No need for an ergo verbose message here,
948 948 // can_expand_young_list() does this when it returns true.
949 949 result = _mutator_alloc_region.attempt_allocation_force(word_size,
950 950 false /* bot_updates */);
951 951 if (result != NULL) {
952 952 return result;
953 953 }
954 954 }
955 955 should_try_gc = false;
956 956 } else {
957 957 // The GCLocker may not be active but the GCLocker initiated
958 958 // GC may not yet have been performed (GCLocker::needs_gc()
959 959 // returns true). In this case we do not try this GC and
960 960 // wait until the GCLocker initiated GC is performed, and
961 961 // then retry the allocation.
962 962 if (GC_locker::needs_gc()) {
963 963 should_try_gc = false;
964 964 } else {
965 965 // Read the GC count while still holding the Heap_lock.
966 966 gc_count_before = total_collections();
967 967 should_try_gc = true;
968 968 }
969 969 }
970 970 }
971 971
972 972 if (should_try_gc) {
973 973 bool succeeded;
974 974 result = do_collection_pause(word_size, gc_count_before, &succeeded);
975 975 if (result != NULL) {
976 976 assert(succeeded, "only way to get back a non-NULL result");
977 977 return result;
978 978 }
979 979
980 980 if (succeeded) {
981 981 // If we get here we successfully scheduled a collection which
982 982 // failed to allocate. No point in trying to allocate
983 983 // further. We'll just return NULL.
984 984 MutexLockerEx x(Heap_lock);
985 985 *gc_count_before_ret = total_collections();
986 986 return NULL;
987 987 }
988 988 } else {
989 989 // The GCLocker is either active or the GCLocker initiated
990 990 // GC has not yet been performed. Stall until it is and
991 991 // then retry the allocation.
992 992 GC_locker::stall_until_clear();
993 993 }
994 994
995 995 // We can reach here if we were unsuccessul in scheduling a
996 996 // collection (because another thread beat us to it) or if we were
997 997 // stalled due to the GC locker. In either can we should retry the
998 998 // allocation attempt in case another thread successfully
999 999 // performed a collection and reclaimed enough space. We do the
1000 1000 // first attempt (without holding the Heap_lock) here and the
1001 1001 // follow-on attempt will be at the start of the next loop
1002 1002 // iteration (after taking the Heap_lock).
1003 1003 result = _mutator_alloc_region.attempt_allocation(word_size,
1004 1004 false /* bot_updates */);
1005 1005 if (result != NULL) {
1006 1006 return result;
1007 1007 }
1008 1008
1009 1009 // Give a warning if we seem to be looping forever.
1010 1010 if ((QueuedAllocationWarningCount > 0) &&
1011 1011 (try_count % QueuedAllocationWarningCount == 0)) {
1012 1012 warning("G1CollectedHeap::attempt_allocation_slow() "
1013 1013 "retries %d times", try_count);
1014 1014 }
1015 1015 }
1016 1016
1017 1017 ShouldNotReachHere();
1018 1018 return NULL;
1019 1019 }
1020 1020
1021 1021 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1022 1022 unsigned int * gc_count_before_ret) {
1023 1023 // The structure of this method has a lot of similarities to
1024 1024 // attempt_allocation_slow(). The reason these two were not merged
1025 1025 // into a single one is that such a method would require several "if
1026 1026 // allocation is not humongous do this, otherwise do that"
1027 1027 // conditional paths which would obscure its flow. In fact, an early
1028 1028 // version of this code did use a unified method which was harder to
1029 1029 // follow and, as a result, it had subtle bugs that were hard to
1030 1030 // track down. So keeping these two methods separate allows each to
1031 1031 // be more readable. It will be good to keep these two in sync as
1032 1032 // much as possible.
1033 1033
1034 1034 assert_heap_not_locked_and_not_at_safepoint();
1035 1035 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1036 1036 "should only be called for humongous allocations");
1037 1037
1038 1038 // Humongous objects can exhaust the heap quickly, so we should check if we
1039 1039 // need to start a marking cycle at each humongous object allocation. We do
1040 1040 // the check before we do the actual allocation. The reason for doing it
1041 1041 // before the allocation is that we avoid having to keep track of the newly
1042 1042 // allocated memory while we do a GC.
1043 1043 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1044 1044 word_size)) {
1045 1045 collect(GCCause::_g1_humongous_allocation);
1046 1046 }
1047 1047
1048 1048 // We will loop until a) we manage to successfully perform the
1049 1049 // allocation or b) we successfully schedule a collection which
1050 1050 // fails to perform the allocation. b) is the only case when we'll
1051 1051 // return NULL.
1052 1052 HeapWord* result = NULL;
1053 1053 for (int try_count = 1; /* we'll return */; try_count += 1) {
1054 1054 bool should_try_gc;
1055 1055 unsigned int gc_count_before;
1056 1056
1057 1057 {
1058 1058 MutexLockerEx x(Heap_lock);
1059 1059
1060 1060 // Given that humongous objects are not allocated in young
1061 1061 // regions, we'll first try to do the allocation without doing a
1062 1062 // collection hoping that there's enough space in the heap.
1063 1063 result = humongous_obj_allocate(word_size);
1064 1064 if (result != NULL) {
1065 1065 return result;
1066 1066 }
1067 1067
1068 1068 if (GC_locker::is_active_and_needs_gc()) {
1069 1069 should_try_gc = false;
1070 1070 } else {
1071 1071 // The GCLocker may not be active but the GCLocker initiated
1072 1072 // GC may not yet have been performed (GCLocker::needs_gc()
1073 1073 // returns true). In this case we do not try this GC and
1074 1074 // wait until the GCLocker initiated GC is performed, and
1075 1075 // then retry the allocation.
1076 1076 if (GC_locker::needs_gc()) {
1077 1077 should_try_gc = false;
1078 1078 } else {
1079 1079 // Read the GC count while still holding the Heap_lock.
1080 1080 gc_count_before = total_collections();
1081 1081 should_try_gc = true;
1082 1082 }
1083 1083 }
1084 1084 }
1085 1085
1086 1086 if (should_try_gc) {
1087 1087 // If we failed to allocate the humongous object, we should try to
1088 1088 // do a collection pause (if we're allowed) in case it reclaims
1089 1089 // enough space for the allocation to succeed after the pause.
1090 1090
1091 1091 bool succeeded;
1092 1092 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1093 1093 if (result != NULL) {
1094 1094 assert(succeeded, "only way to get back a non-NULL result");
1095 1095 return result;
1096 1096 }
1097 1097
1098 1098 if (succeeded) {
1099 1099 // If we get here we successfully scheduled a collection which
1100 1100 // failed to allocate. No point in trying to allocate
1101 1101 // further. We'll just return NULL.
1102 1102 MutexLockerEx x(Heap_lock);
1103 1103 *gc_count_before_ret = total_collections();
1104 1104 return NULL;
1105 1105 }
1106 1106 } else {
1107 1107 // The GCLocker is either active or the GCLocker initiated
1108 1108 // GC has not yet been performed. Stall until it is and
1109 1109 // then retry the allocation.
1110 1110 GC_locker::stall_until_clear();
1111 1111 }
1112 1112
1113 1113 // We can reach here if we were unsuccessul in scheduling a
1114 1114 // collection (because another thread beat us to it) or if we were
1115 1115 // stalled due to the GC locker. In either can we should retry the
1116 1116 // allocation attempt in case another thread successfully
1117 1117 // performed a collection and reclaimed enough space. Give a
1118 1118 // warning if we seem to be looping forever.
1119 1119
1120 1120 if ((QueuedAllocationWarningCount > 0) &&
1121 1121 (try_count % QueuedAllocationWarningCount == 0)) {
1122 1122 warning("G1CollectedHeap::attempt_allocation_humongous() "
1123 1123 "retries %d times", try_count);
1124 1124 }
1125 1125 }
1126 1126
1127 1127 ShouldNotReachHere();
1128 1128 return NULL;
1129 1129 }
1130 1130
1131 1131 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1132 1132 bool expect_null_mutator_alloc_region) {
1133 1133 assert_at_safepoint(true /* should_be_vm_thread */);
1134 1134 assert(_mutator_alloc_region.get() == NULL ||
1135 1135 !expect_null_mutator_alloc_region,
1136 1136 "the current alloc region was unexpectedly found to be non-NULL");
1137 1137
1138 1138 if (!isHumongous(word_size)) {
1139 1139 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1140 1140 false /* bot_updates */);
1141 1141 } else {
1142 1142 HeapWord* result = humongous_obj_allocate(word_size);
1143 1143 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1144 1144 g1_policy()->set_initiate_conc_mark_if_possible();
1145 1145 }
1146 1146 return result;
1147 1147 }
1148 1148
1149 1149 ShouldNotReachHere();
1150 1150 }
1151 1151
1152 1152 class PostMCRemSetClearClosure: public HeapRegionClosure {
1153 1153 G1CollectedHeap* _g1h;
1154 1154 ModRefBarrierSet* _mr_bs;
1155 1155 public:
1156 1156 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1157 1157 _g1h(g1h), _mr_bs(mr_bs) { }
1158 1158 bool doHeapRegion(HeapRegion* r) {
1159 1159 if (r->continuesHumongous()) {
1160 1160 return false;
1161 1161 }
1162 1162 _g1h->reset_gc_time_stamps(r);
1163 1163 HeapRegionRemSet* hrrs = r->rem_set();
1164 1164 if (hrrs != NULL) hrrs->clear();
1165 1165 // You might think here that we could clear just the cards
1166 1166 // corresponding to the used region. But no: if we leave a dirty card
1167 1167 // in a region we might allocate into, then it would prevent that card
1168 1168 // from being enqueued, and cause it to be missed.
1169 1169 // Re: the performance cost: we shouldn't be doing full GC anyway!
1170 1170 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1171 1171 return false;
1172 1172 }
1173 1173 };
1174 1174
1175 1175 void G1CollectedHeap::clear_rsets_post_compaction() {
1176 1176 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1177 1177 heap_region_iterate(&rs_clear);
1178 1178 }
1179 1179
1180 1180 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1181 1181 G1CollectedHeap* _g1h;
1182 1182 UpdateRSOopClosure _cl;
1183 1183 int _worker_i;
1184 1184 public:
1185 1185 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1186 1186 _cl(g1->g1_rem_set(), worker_i),
1187 1187 _worker_i(worker_i),
1188 1188 _g1h(g1)
1189 1189 { }
1190 1190
1191 1191 bool doHeapRegion(HeapRegion* r) {
1192 1192 if (!r->continuesHumongous()) {
1193 1193 _cl.set_from(r);
1194 1194 r->oop_iterate(&_cl);
1195 1195 }
1196 1196 return false;
1197 1197 }
1198 1198 };
1199 1199
1200 1200 class ParRebuildRSTask: public AbstractGangTask {
1201 1201 G1CollectedHeap* _g1;
1202 1202 public:
1203 1203 ParRebuildRSTask(G1CollectedHeap* g1)
1204 1204 : AbstractGangTask("ParRebuildRSTask"),
1205 1205 _g1(g1)
1206 1206 { }
1207 1207
1208 1208 void work(uint worker_id) {
1209 1209 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1210 1210 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1211 1211 _g1->workers()->active_workers(),
1212 1212 HeapRegion::RebuildRSClaimValue);
1213 1213 }
1214 1214 };
1215 1215
1216 1216 class PostCompactionPrinterClosure: public HeapRegionClosure {
1217 1217 private:
1218 1218 G1HRPrinter* _hr_printer;
1219 1219 public:
1220 1220 bool doHeapRegion(HeapRegion* hr) {
1221 1221 assert(!hr->is_young(), "not expecting to find young regions");
1222 1222 // We only generate output for non-empty regions.
1223 1223 if (!hr->is_empty()) {
1224 1224 if (!hr->isHumongous()) {
1225 1225 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1226 1226 } else if (hr->startsHumongous()) {
1227 1227 if (hr->region_num() == 1) {
1228 1228 // single humongous region
1229 1229 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1230 1230 } else {
1231 1231 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1232 1232 }
1233 1233 } else {
1234 1234 assert(hr->continuesHumongous(), "only way to get here");
1235 1235 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1236 1236 }
1237 1237 }
1238 1238 return false;
1239 1239 }
1240 1240
1241 1241 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1242 1242 : _hr_printer(hr_printer) { }
1243 1243 };
1244 1244
1245 1245 void G1CollectedHeap::print_hrs_post_compaction() {
1246 1246 PostCompactionPrinterClosure cl(hr_printer());
1247 1247 heap_region_iterate(&cl);
1248 1248 }
1249 1249
1250 1250 double G1CollectedHeap::verify(bool guard, const char* msg) {
1251 1251 double verify_time_ms = 0.0;
1252 1252
1253 1253 if (guard && total_collections() >= VerifyGCStartAt) {
1254 1254 double verify_start = os::elapsedTime();
1255 1255 HandleMark hm; // Discard invalid handles created during verification
1256 1256 gclog_or_tty->print(msg);
1257 1257 prepare_for_verify();
1258 1258 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking);
1259 1259 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1260 1260 }
1261 1261
1262 1262 return verify_time_ms;
1263 1263 }
1264 1264
1265 1265 void G1CollectedHeap::verify_before_gc() {
1266 1266 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1267 1267 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1268 1268 }
1269 1269
1270 1270 void G1CollectedHeap::verify_after_gc() {
1271 1271 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1272 1272 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1273 1273 }
1274 1274
1275 1275 bool G1CollectedHeap::do_collection(bool explicit_gc,
1276 1276 bool clear_all_soft_refs,
1277 1277 size_t word_size) {
1278 1278 assert_at_safepoint(true /* should_be_vm_thread */);
1279 1279
1280 1280 if (GC_locker::check_active_before_gc()) {
1281 1281 return false;
1282 1282 }
1283 1283
1284 1284 SvcGCMarker sgcm(SvcGCMarker::FULL);
1285 1285 ResourceMark rm;
1286 1286
1287 1287 print_heap_before_gc();
1288 1288
1289 1289 size_t metadata_prev_used = MetaspaceAux::used_in_bytes();
1290 1290
1291 1291 HRSPhaseSetter x(HRSPhaseFullGC);
1292 1292 verify_region_sets_optional();
1293 1293
1294 1294 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1295 1295 collector_policy()->should_clear_all_soft_refs();
1296 1296
1297 1297 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1298 1298
1299 1299 {
1300 1300 IsGCActiveMark x;
1301 1301
1302 1302 // Timing
1303 1303 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1304 1304 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1305 1305 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1306 1306
1307 1307 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1308 1308 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1309 1309 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1310 1310
1311 1311 double start = os::elapsedTime();
1312 1312 g1_policy()->record_full_collection_start();
1313 1313
1314 1314 // Note: When we have a more flexible GC logging framework that
1315 1315 // allows us to add optional attributes to a GC log record we
1316 1316 // could consider timing and reporting how long we wait in the
1317 1317 // following two methods.
1318 1318 wait_while_free_regions_coming();
1319 1319 // If we start the compaction before the CM threads finish
1320 1320 // scanning the root regions we might trip them over as we'll
1321 1321 // be moving objects / updating references. So let's wait until
1322 1322 // they are done. By telling them to abort, they should complete
1323 1323 // early.
1324 1324 _cm->root_regions()->abort();
1325 1325 _cm->root_regions()->wait_until_scan_finished();
1326 1326 append_secondary_free_list_if_not_empty_with_lock();
1327 1327
1328 1328 gc_prologue(true);
1329 1329 increment_total_collections(true /* full gc */);
1330 1330 increment_old_marking_cycles_started();
1331 1331
1332 1332 size_t g1h_prev_used = used();
1333 1333 assert(used() == recalculate_used(), "Should be equal");
1334 1334
1335 1335 verify_before_gc();
1336 1336
1337 1337 pre_full_gc_dump();
1338 1338
1339 1339 COMPILER2_PRESENT(DerivedPointerTable::clear());
1340 1340
1341 1341 // Disable discovery and empty the discovered lists
1342 1342 // for the CM ref processor.
1343 1343 ref_processor_cm()->disable_discovery();
1344 1344 ref_processor_cm()->abandon_partial_discovery();
1345 1345 ref_processor_cm()->verify_no_references_recorded();
1346 1346
1347 1347 // Abandon current iterations of concurrent marking and concurrent
1348 1348 // refinement, if any are in progress. We have to do this before
1349 1349 // wait_until_scan_finished() below.
1350 1350 concurrent_mark()->abort();
1351 1351
1352 1352 // Make sure we'll choose a new allocation region afterwards.
1353 1353 release_mutator_alloc_region();
1354 1354 abandon_gc_alloc_regions();
1355 1355 g1_rem_set()->cleanupHRRS();
1356 1356
1357 1357 // We should call this after we retire any currently active alloc
1358 1358 // regions so that all the ALLOC / RETIRE events are generated
1359 1359 // before the start GC event.
1360 1360 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1361 1361
1362 1362 // We may have added regions to the current incremental collection
1363 1363 // set between the last GC or pause and now. We need to clear the
1364 1364 // incremental collection set and then start rebuilding it afresh
1365 1365 // after this full GC.
1366 1366 abandon_collection_set(g1_policy()->inc_cset_head());
1367 1367 g1_policy()->clear_incremental_cset();
1368 1368 g1_policy()->stop_incremental_cset_building();
1369 1369
1370 1370 tear_down_region_sets(false /* free_list_only */);
1371 1371 g1_policy()->set_gcs_are_young(true);
1372 1372
1373 1373 // See the comments in g1CollectedHeap.hpp and
1374 1374 // G1CollectedHeap::ref_processing_init() about
1375 1375 // how reference processing currently works in G1.
1376 1376
1377 1377 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1378 1378 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1379 1379
1380 1380 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1381 1381 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1382 1382
1383 1383 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1384 1384 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1385 1385
1386 1386 // Do collection work
1387 1387 {
1388 1388 HandleMark hm; // Discard invalid handles created during gc
1389 1389 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1390 1390 }
1391 1391
1392 1392 assert(free_regions() == 0, "we should not have added any free regions");
1393 1393 rebuild_region_sets(false /* free_list_only */);
1394 1394
1395 1395 // Enqueue any discovered reference objects that have
1396 1396 // not been removed from the discovered lists.
1397 1397 ref_processor_stw()->enqueue_discovered_references();
1398 1398
1399 1399 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1400 1400
1401 1401 MemoryService::track_memory_usage();
1402 1402
1403 1403 verify_after_gc();
1404 1404
1405 1405 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1406 1406 ref_processor_stw()->verify_no_references_recorded();
1407 1407
1408 1408 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1409 1409 ClassLoaderDataGraph::purge();
1410 1410
1411 1411 // Note: since we've just done a full GC, concurrent
1412 1412 // marking is no longer active. Therefore we need not
1413 1413 // re-enable reference discovery for the CM ref processor.
1414 1414 // That will be done at the start of the next marking cycle.
1415 1415 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1416 1416 ref_processor_cm()->verify_no_references_recorded();
1417 1417
1418 1418 reset_gc_time_stamp();
1419 1419 // Since everything potentially moved, we will clear all remembered
1420 1420 // sets, and clear all cards. Later we will rebuild remebered
1421 1421 // sets. We will also reset the GC time stamps of the regions.
1422 1422 clear_rsets_post_compaction();
1423 1423 check_gc_time_stamps();
1424 1424
1425 1425 // Resize the heap if necessary.
1426 1426 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1427 1427
1428 1428 if (_hr_printer.is_active()) {
1429 1429 // We should do this after we potentially resize the heap so
1430 1430 // that all the COMMIT / UNCOMMIT events are generated before
1431 1431 // the end GC event.
1432 1432
1433 1433 print_hrs_post_compaction();
1434 1434 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1435 1435 }
1436 1436
1437 1437 if (_cg1r->use_cache()) {
1438 1438 _cg1r->clear_and_record_card_counts();
1439 1439 _cg1r->clear_hot_cache();
1440 1440 }
1441 1441
1442 1442 // Rebuild remembered sets of all regions.
1443 1443 if (G1CollectedHeap::use_parallel_gc_threads()) {
1444 1444 uint n_workers =
1445 1445 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1446 1446 workers()->active_workers(),
1447 1447 Threads::number_of_non_daemon_threads());
1448 1448 assert(UseDynamicNumberOfGCThreads ||
1449 1449 n_workers == workers()->total_workers(),
1450 1450 "If not dynamic should be using all the workers");
1451 1451 workers()->set_active_workers(n_workers);
1452 1452 // Set parallel threads in the heap (_n_par_threads) only
1453 1453 // before a parallel phase and always reset it to 0 after
1454 1454 // the phase so that the number of parallel threads does
1455 1455 // no get carried forward to a serial phase where there
1456 1456 // may be code that is "possibly_parallel".
1457 1457 set_par_threads(n_workers);
1458 1458
1459 1459 ParRebuildRSTask rebuild_rs_task(this);
1460 1460 assert(check_heap_region_claim_values(
1461 1461 HeapRegion::InitialClaimValue), "sanity check");
1462 1462 assert(UseDynamicNumberOfGCThreads ||
1463 1463 workers()->active_workers() == workers()->total_workers(),
1464 1464 "Unless dynamic should use total workers");
1465 1465 // Use the most recent number of active workers
1466 1466 assert(workers()->active_workers() > 0,
1467 1467 "Active workers not properly set");
1468 1468 set_par_threads(workers()->active_workers());
1469 1469 workers()->run_task(&rebuild_rs_task);
1470 1470 set_par_threads(0);
1471 1471 assert(check_heap_region_claim_values(
1472 1472 HeapRegion::RebuildRSClaimValue), "sanity check");
1473 1473 reset_heap_region_claim_values();
1474 1474 } else {
1475 1475 RebuildRSOutOfRegionClosure rebuild_rs(this);
1476 1476 heap_region_iterate(&rebuild_rs);
1477 1477 }
1478 1478
1479 1479 if (G1Log::fine()) {
1480 1480 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1481 1481 }
1482 1482
1483 1483 if (true) { // FIXME
1484 1484 MetaspaceGC::compute_new_size();
1485 1485 }
1486 1486
1487 1487 // Start a new incremental collection set for the next pause
1488 1488 assert(g1_policy()->collection_set() == NULL, "must be");
1489 1489 g1_policy()->start_incremental_cset_building();
1490 1490
1491 1491 // Clear the _cset_fast_test bitmap in anticipation of adding
1492 1492 // regions to the incremental collection set for the next
1493 1493 // evacuation pause.
1494 1494 clear_cset_fast_test();
1495 1495
1496 1496 init_mutator_alloc_region();
1497 1497
1498 1498 double end = os::elapsedTime();
1499 1499 g1_policy()->record_full_collection_end();
1500 1500
1501 1501 #ifdef TRACESPINNING
1502 1502 ParallelTaskTerminator::print_termination_counts();
1503 1503 #endif
1504 1504
1505 1505 gc_epilogue(true);
1506 1506
1507 1507 // Discard all rset updates
1508 1508 JavaThread::dirty_card_queue_set().abandon_logs();
1509 1509 assert(!G1DeferredRSUpdate
1510 1510 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1511 1511
1512 1512 _young_list->reset_sampled_info();
1513 1513 // At this point there should be no regions in the
1514 1514 // entire heap tagged as young.
1515 1515 assert( check_young_list_empty(true /* check_heap */),
1516 1516 "young list should be empty at this point");
1517 1517
1518 1518 // Update the number of full collections that have been completed.
1519 1519 increment_old_marking_cycles_completed(false /* concurrent */);
1520 1520
1521 1521 _hrs.verify_optional();
1522 1522 verify_region_sets_optional();
1523 1523
1524 1524 print_heap_after_gc();
1525 1525
1526 1526 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1527 1527 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1528 1528 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1529 1529 // before any GC notifications are raised.
1530 1530 g1mm()->update_sizes();
1531 1531 }
1532 1532
1533 1533 post_full_gc_dump();
1534 1534
1535 1535 return true;
1536 1536 }
1537 1537
1538 1538 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1539 1539 // do_collection() will return whether it succeeded in performing
1540 1540 // the GC. Currently, there is no facility on the
1541 1541 // do_full_collection() API to notify the caller than the collection
1542 1542 // did not succeed (e.g., because it was locked out by the GC
1543 1543 // locker). So, right now, we'll ignore the return value.
1544 1544 bool dummy = do_collection(true, /* explicit_gc */
1545 1545 clear_all_soft_refs,
1546 1546 0 /* word_size */);
1547 1547 }
1548 1548
1549 1549 // This code is mostly copied from TenuredGeneration.
1550 1550 void
1551 1551 G1CollectedHeap::
1552 1552 resize_if_necessary_after_full_collection(size_t word_size) {
1553 1553 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1554 1554
1555 1555 // Include the current allocation, if any, and bytes that will be
1556 1556 // pre-allocated to support collections, as "used".
1557 1557 const size_t used_after_gc = used();
1558 1558 const size_t capacity_after_gc = capacity();
1559 1559 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1560 1560
1561 1561 // This is enforced in arguments.cpp.
1562 1562 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1563 1563 "otherwise the code below doesn't make sense");
1564 1564
1565 1565 // We don't have floating point command-line arguments
1566 1566 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1567 1567 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1568 1568 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1569 1569 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1570 1570
1571 1571 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1572 1572 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1573 1573
1574 1574 // We have to be careful here as these two calculations can overflow
1575 1575 // 32-bit size_t's.
1576 1576 double used_after_gc_d = (double) used_after_gc;
1577 1577 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1578 1578 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1579 1579
1580 1580 // Let's make sure that they are both under the max heap size, which
1581 1581 // by default will make them fit into a size_t.
1582 1582 double desired_capacity_upper_bound = (double) max_heap_size;
1583 1583 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1584 1584 desired_capacity_upper_bound);
1585 1585 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1586 1586 desired_capacity_upper_bound);
1587 1587
1588 1588 // We can now safely turn them into size_t's.
1589 1589 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1590 1590 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1591 1591
1592 1592 // This assert only makes sense here, before we adjust them
1593 1593 // with respect to the min and max heap size.
1594 1594 assert(minimum_desired_capacity <= maximum_desired_capacity,
1595 1595 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1596 1596 "maximum_desired_capacity = "SIZE_FORMAT,
1597 1597 minimum_desired_capacity, maximum_desired_capacity));
1598 1598
1599 1599 // Should not be greater than the heap max size. No need to adjust
1600 1600 // it with respect to the heap min size as it's a lower bound (i.e.,
1601 1601 // we'll try to make the capacity larger than it, not smaller).
1602 1602 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1603 1603 // Should not be less than the heap min size. No need to adjust it
1604 1604 // with respect to the heap max size as it's an upper bound (i.e.,
1605 1605 // we'll try to make the capacity smaller than it, not greater).
1606 1606 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1607 1607
1608 1608 if (capacity_after_gc < minimum_desired_capacity) {
1609 1609 // Don't expand unless it's significant
1610 1610 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1611 1611 ergo_verbose4(ErgoHeapSizing,
1612 1612 "attempt heap expansion",
1613 1613 ergo_format_reason("capacity lower than "
1614 1614 "min desired capacity after Full GC")
1615 1615 ergo_format_byte("capacity")
1616 1616 ergo_format_byte("occupancy")
1617 1617 ergo_format_byte_perc("min desired capacity"),
1618 1618 capacity_after_gc, used_after_gc,
1619 1619 minimum_desired_capacity, (double) MinHeapFreeRatio);
1620 1620 expand(expand_bytes);
1621 1621
1622 1622 // No expansion, now see if we want to shrink
1623 1623 } else if (capacity_after_gc > maximum_desired_capacity) {
1624 1624 // Capacity too large, compute shrinking size
1625 1625 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1626 1626 ergo_verbose4(ErgoHeapSizing,
1627 1627 "attempt heap shrinking",
1628 1628 ergo_format_reason("capacity higher than "
1629 1629 "max desired capacity after Full GC")
1630 1630 ergo_format_byte("capacity")
1631 1631 ergo_format_byte("occupancy")
1632 1632 ergo_format_byte_perc("max desired capacity"),
1633 1633 capacity_after_gc, used_after_gc,
1634 1634 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1635 1635 shrink(shrink_bytes);
1636 1636 }
1637 1637 }
1638 1638
1639 1639
1640 1640 HeapWord*
1641 1641 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1642 1642 bool* succeeded) {
1643 1643 assert_at_safepoint(true /* should_be_vm_thread */);
1644 1644
1645 1645 *succeeded = true;
1646 1646 // Let's attempt the allocation first.
1647 1647 HeapWord* result =
1648 1648 attempt_allocation_at_safepoint(word_size,
1649 1649 false /* expect_null_mutator_alloc_region */);
1650 1650 if (result != NULL) {
1651 1651 assert(*succeeded, "sanity");
1652 1652 return result;
1653 1653 }
1654 1654
1655 1655 // In a G1 heap, we're supposed to keep allocation from failing by
1656 1656 // incremental pauses. Therefore, at least for now, we'll favor
1657 1657 // expansion over collection. (This might change in the future if we can
1658 1658 // do something smarter than full collection to satisfy a failed alloc.)
1659 1659 result = expand_and_allocate(word_size);
1660 1660 if (result != NULL) {
1661 1661 assert(*succeeded, "sanity");
1662 1662 return result;
1663 1663 }
1664 1664
1665 1665 // Expansion didn't work, we'll try to do a Full GC.
1666 1666 bool gc_succeeded = do_collection(false, /* explicit_gc */
1667 1667 false, /* clear_all_soft_refs */
1668 1668 word_size);
1669 1669 if (!gc_succeeded) {
1670 1670 *succeeded = false;
1671 1671 return NULL;
1672 1672 }
1673 1673
1674 1674 // Retry the allocation
1675 1675 result = attempt_allocation_at_safepoint(word_size,
1676 1676 true /* expect_null_mutator_alloc_region */);
1677 1677 if (result != NULL) {
1678 1678 assert(*succeeded, "sanity");
1679 1679 return result;
1680 1680 }
1681 1681
1682 1682 // Then, try a Full GC that will collect all soft references.
1683 1683 gc_succeeded = do_collection(false, /* explicit_gc */
1684 1684 true, /* clear_all_soft_refs */
1685 1685 word_size);
1686 1686 if (!gc_succeeded) {
1687 1687 *succeeded = false;
1688 1688 return NULL;
1689 1689 }
1690 1690
1691 1691 // Retry the allocation once more
1692 1692 result = attempt_allocation_at_safepoint(word_size,
1693 1693 true /* expect_null_mutator_alloc_region */);
1694 1694 if (result != NULL) {
1695 1695 assert(*succeeded, "sanity");
1696 1696 return result;
1697 1697 }
1698 1698
1699 1699 assert(!collector_policy()->should_clear_all_soft_refs(),
1700 1700 "Flag should have been handled and cleared prior to this point");
1701 1701
1702 1702 // What else? We might try synchronous finalization later. If the total
1703 1703 // space available is large enough for the allocation, then a more
1704 1704 // complete compaction phase than we've tried so far might be
1705 1705 // appropriate.
1706 1706 assert(*succeeded, "sanity");
1707 1707 return NULL;
1708 1708 }
1709 1709
1710 1710 // Attempting to expand the heap sufficiently
1711 1711 // to support an allocation of the given "word_size". If
1712 1712 // successful, perform the allocation and return the address of the
1713 1713 // allocated block, or else "NULL".
1714 1714
1715 1715 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1716 1716 assert_at_safepoint(true /* should_be_vm_thread */);
1717 1717
1718 1718 verify_region_sets_optional();
1719 1719
1720 1720 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1721 1721 ergo_verbose1(ErgoHeapSizing,
1722 1722 "attempt heap expansion",
1723 1723 ergo_format_reason("allocation request failed")
1724 1724 ergo_format_byte("allocation request"),
1725 1725 word_size * HeapWordSize);
1726 1726 if (expand(expand_bytes)) {
1727 1727 _hrs.verify_optional();
1728 1728 verify_region_sets_optional();
1729 1729 return attempt_allocation_at_safepoint(word_size,
1730 1730 false /* expect_null_mutator_alloc_region */);
1731 1731 }
1732 1732 return NULL;
1733 1733 }
1734 1734
1735 1735 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1736 1736 HeapWord* new_end) {
1737 1737 assert(old_end != new_end, "don't call this otherwise");
1738 1738 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1739 1739
1740 1740 // Update the committed mem region.
1741 1741 _g1_committed.set_end(new_end);
1742 1742 // Tell the card table about the update.
1743 1743 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1744 1744 // Tell the BOT about the update.
1745 1745 _bot_shared->resize(_g1_committed.word_size());
1746 1746 }
1747 1747
1748 1748 bool G1CollectedHeap::expand(size_t expand_bytes) {
1749 1749 size_t old_mem_size = _g1_storage.committed_size();
1750 1750 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1751 1751 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1752 1752 HeapRegion::GrainBytes);
1753 1753 ergo_verbose2(ErgoHeapSizing,
1754 1754 "expand the heap",
1755 1755 ergo_format_byte("requested expansion amount")
1756 1756 ergo_format_byte("attempted expansion amount"),
1757 1757 expand_bytes, aligned_expand_bytes);
1758 1758
1759 1759 // First commit the memory.
1760 1760 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1761 1761 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1762 1762 if (successful) {
1763 1763 // Then propagate this update to the necessary data structures.
1764 1764 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1765 1765 update_committed_space(old_end, new_end);
1766 1766
1767 1767 FreeRegionList expansion_list("Local Expansion List");
1768 1768 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1769 1769 assert(mr.start() == old_end, "post-condition");
1770 1770 // mr might be a smaller region than what was requested if
1771 1771 // expand_by() was unable to allocate the HeapRegion instances
1772 1772 assert(mr.end() <= new_end, "post-condition");
1773 1773
1774 1774 size_t actual_expand_bytes = mr.byte_size();
1775 1775 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1776 1776 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1777 1777 "post-condition");
1778 1778 if (actual_expand_bytes < aligned_expand_bytes) {
1779 1779 // We could not expand _hrs to the desired size. In this case we
1780 1780 // need to shrink the committed space accordingly.
1781 1781 assert(mr.end() < new_end, "invariant");
1782 1782
1783 1783 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1784 1784 // First uncommit the memory.
1785 1785 _g1_storage.shrink_by(diff_bytes);
1786 1786 // Then propagate this update to the necessary data structures.
1787 1787 update_committed_space(new_end, mr.end());
1788 1788 }
1789 1789 _free_list.add_as_tail(&expansion_list);
1790 1790
1791 1791 if (_hr_printer.is_active()) {
1792 1792 HeapWord* curr = mr.start();
1793 1793 while (curr < mr.end()) {
1794 1794 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1795 1795 _hr_printer.commit(curr, curr_end);
1796 1796 curr = curr_end;
1797 1797 }
1798 1798 assert(curr == mr.end(), "post-condition");
1799 1799 }
1800 1800 g1_policy()->record_new_heap_size(n_regions());
1801 1801 } else {
1802 1802 ergo_verbose0(ErgoHeapSizing,
1803 1803 "did not expand the heap",
1804 1804 ergo_format_reason("heap expansion operation failed"));
1805 1805 // The expansion of the virtual storage space was unsuccessful.
1806 1806 // Let's see if it was because we ran out of swap.
1807 1807 if (G1ExitOnExpansionFailure &&
1808 1808 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1809 1809 // We had head room...
1810 1810 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1811 1811 }
1812 1812 }
1813 1813 return successful;
1814 1814 }
1815 1815
1816 1816 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1817 1817 size_t old_mem_size = _g1_storage.committed_size();
1818 1818 size_t aligned_shrink_bytes =
1819 1819 ReservedSpace::page_align_size_down(shrink_bytes);
1820 1820 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1821 1821 HeapRegion::GrainBytes);
1822 1822 uint num_regions_deleted = 0;
1823 1823 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1824 1824 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1825 1825 assert(mr.end() == old_end, "post-condition");
1826 1826
1827 1827 ergo_verbose3(ErgoHeapSizing,
1828 1828 "shrink the heap",
1829 1829 ergo_format_byte("requested shrinking amount")
1830 1830 ergo_format_byte("aligned shrinking amount")
1831 1831 ergo_format_byte("attempted shrinking amount"),
1832 1832 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1833 1833 if (mr.byte_size() > 0) {
1834 1834 if (_hr_printer.is_active()) {
1835 1835 HeapWord* curr = mr.end();
1836 1836 while (curr > mr.start()) {
1837 1837 HeapWord* curr_end = curr;
1838 1838 curr -= HeapRegion::GrainWords;
1839 1839 _hr_printer.uncommit(curr, curr_end);
1840 1840 }
1841 1841 assert(curr == mr.start(), "post-condition");
1842 1842 }
1843 1843
1844 1844 _g1_storage.shrink_by(mr.byte_size());
1845 1845 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1846 1846 assert(mr.start() == new_end, "post-condition");
1847 1847
1848 1848 _expansion_regions += num_regions_deleted;
1849 1849 update_committed_space(old_end, new_end);
1850 1850 HeapRegionRemSet::shrink_heap(n_regions());
1851 1851 g1_policy()->record_new_heap_size(n_regions());
1852 1852 } else {
1853 1853 ergo_verbose0(ErgoHeapSizing,
1854 1854 "did not shrink the heap",
1855 1855 ergo_format_reason("heap shrinking operation failed"));
1856 1856 }
1857 1857 }
1858 1858
1859 1859 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1860 1860 verify_region_sets_optional();
1861 1861
1862 1862 // We should only reach here at the end of a Full GC which means we
1863 1863 // should not not be holding to any GC alloc regions. The method
1864 1864 // below will make sure of that and do any remaining clean up.
1865 1865 abandon_gc_alloc_regions();
1866 1866
1867 1867 // Instead of tearing down / rebuilding the free lists here, we
1868 1868 // could instead use the remove_all_pending() method on free_list to
1869 1869 // remove only the ones that we need to remove.
1870 1870 tear_down_region_sets(true /* free_list_only */);
1871 1871 shrink_helper(shrink_bytes);
1872 1872 rebuild_region_sets(true /* free_list_only */);
1873 1873
1874 1874 _hrs.verify_optional();
1875 1875 verify_region_sets_optional();
1876 1876 }
1877 1877
1878 1878 // Public methods.
1879 1879
1880 1880 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1881 1881 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1882 1882 #endif // _MSC_VER
1883 1883
1884 1884
1885 1885 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1886 1886 SharedHeap(policy_),
1887 1887 _g1_policy(policy_),
1888 1888 _dirty_card_queue_set(false),
1889 1889 _into_cset_dirty_card_queue_set(false),
1890 1890 _is_alive_closure_cm(this),
1891 1891 _is_alive_closure_stw(this),
1892 1892 _ref_processor_cm(NULL),
1893 1893 _ref_processor_stw(NULL),
1894 1894 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1895 1895 _bot_shared(NULL),
1896 1896 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1897 1897 _evac_failure_scan_stack(NULL) ,
1898 1898 _mark_in_progress(false),
1899 1899 _cg1r(NULL), _summary_bytes_used(0),
1900 1900 _g1mm(NULL),
1901 1901 _refine_cte_cl(NULL),
1902 1902 _full_collection(false),
1903 1903 _free_list("Master Free List"),
1904 1904 _secondary_free_list("Secondary Free List"),
1905 1905 _old_set("Old Set"),
1906 1906 _humongous_set("Master Humongous Set"),
1907 1907 _free_regions_coming(false),
1908 1908 _young_list(new YoungList(this)),
1909 1909 _gc_time_stamp(0),
1910 1910 _retained_old_gc_alloc_region(NULL),
1911 1911 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1912 1912 _old_plab_stats(OldPLABSize, PLABWeight),
1913 1913 _expand_heap_after_alloc_failure(true),
1914 1914 _surviving_young_words(NULL),
1915 1915 _old_marking_cycles_started(0),
1916 1916 _old_marking_cycles_completed(0),
1917 1917 _in_cset_fast_test(NULL),
1918 1918 _in_cset_fast_test_base(NULL),
1919 1919 _dirty_cards_region_list(NULL),
1920 1920 _worker_cset_start_region(NULL),
1921 1921 _worker_cset_start_region_time_stamp(NULL) {
1922 1922 _g1h = this; // To catch bugs.
1923 1923 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1924 1924 vm_exit_during_initialization("Failed necessary allocation.");
1925 1925 }
1926 1926
1927 1927 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1928 1928
1929 1929 int n_queues = MAX2((int)ParallelGCThreads, 1);
1930 1930 _task_queues = new RefToScanQueueSet(n_queues);
1931 1931
1932 1932 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1933 1933 assert(n_rem_sets > 0, "Invariant.");
1934 1934
1935 1935 HeapRegionRemSetIterator** iter_arr =
1936 1936 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC);
1937 1937 for (int i = 0; i < n_queues; i++) {
1938 1938 iter_arr[i] = new HeapRegionRemSetIterator();
1939 1939 }
1940 1940 _rem_set_iterator = iter_arr;
1941 1941
1942 1942 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1943 1943 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1944 1944
1945 1945 for (int i = 0; i < n_queues; i++) {
1946 1946 RefToScanQueue* q = new RefToScanQueue();
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.
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 2082 _cm = new ConcurrentMark(heap_rs, max_regions());
2083 2083 _cmThread = _cm->cmThread();
2084 2084
2085 2085 // Initialize the from_card cache structure of HeapRegionRemSet.
2086 2086 HeapRegionRemSet::init_heap(max_regions());
2087 2087
2088 2088 // Now expand into the initial heap size.
2089 2089 if (!expand(init_byte_size)) {
2090 2090 vm_exit_during_initialization("Failed to allocate initial heap.");
2091 2091 return JNI_ENOMEM;
2092 2092 }
2093 2093
2094 2094 // Perform any initialization actions delegated to the policy.
2095 2095 g1_policy()->init();
2096 2096
2097 2097 _refine_cte_cl =
2098 2098 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2099 2099 g1_rem_set(),
2100 2100 concurrent_g1_refine());
2101 2101 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2102 2102
2103 2103 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2104 2104 SATB_Q_FL_lock,
2105 2105 G1SATBProcessCompletedThreshold,
2106 2106 Shared_SATB_Q_lock);
2107 2107
2108 2108 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2109 2109 DirtyCardQ_FL_lock,
2110 2110 concurrent_g1_refine()->yellow_zone(),
2111 2111 concurrent_g1_refine()->red_zone(),
2112 2112 Shared_DirtyCardQ_lock);
2113 2113
2114 2114 if (G1DeferredRSUpdate) {
2115 2115 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2116 2116 DirtyCardQ_FL_lock,
2117 2117 -1, // never trigger processing
2118 2118 -1, // no limit on length
2119 2119 Shared_DirtyCardQ_lock,
2120 2120 &JavaThread::dirty_card_queue_set());
2121 2121 }
2122 2122
2123 2123 // Initialize the card queue set used to hold cards containing
2124 2124 // references into the collection set.
2125 2125 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2126 2126 DirtyCardQ_FL_lock,
2127 2127 -1, // never trigger processing
2128 2128 -1, // no limit on length
2129 2129 Shared_DirtyCardQ_lock,
2130 2130 &JavaThread::dirty_card_queue_set());
2131 2131
2132 2132 // In case we're keeping closure specialization stats, initialize those
2133 2133 // counts and that mechanism.
2134 2134 SpecializationStats::clear();
2135 2135
2136 2136 // Do later initialization work for concurrent refinement.
2137 2137 _cg1r->init();
2138 2138
2139 2139 // Here we allocate the dummy full region that is required by the
2140 2140 // G1AllocRegion class. If we don't pass an address in the reserved
2141 2141 // space here, lots of asserts fire.
2142 2142
2143 2143 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2144 2144 _g1_reserved.start());
2145 2145 // We'll re-use the same region whether the alloc region will
2146 2146 // require BOT updates or not and, if it doesn't, then a non-young
2147 2147 // region will complain that it cannot support allocations without
2148 2148 // BOT updates. So we'll tag the dummy region as young to avoid that.
2149 2149 dummy_region->set_young();
2150 2150 // Make sure it's full.
2151 2151 dummy_region->set_top(dummy_region->end());
2152 2152 G1AllocRegion::setup(this, dummy_region);
2153 2153
2154 2154 init_mutator_alloc_region();
2155 2155
2156 2156 // Do create of the monitoring and management support so that
2157 2157 // values in the heap have been properly initialized.
2158 2158 _g1mm = new G1MonitoringSupport(this);
2159 2159
2160 2160 return JNI_OK;
2161 2161 }
2162 2162
2163 2163 void G1CollectedHeap::ref_processing_init() {
2164 2164 // Reference processing in G1 currently works as follows:
2165 2165 //
2166 2166 // * There are two reference processor instances. One is
2167 2167 // used to record and process discovered references
2168 2168 // during concurrent marking; the other is used to
2169 2169 // record and process references during STW pauses
2170 2170 // (both full and incremental).
2171 2171 // * Both ref processors need to 'span' the entire heap as
2172 2172 // the regions in the collection set may be dotted around.
2173 2173 //
2174 2174 // * For the concurrent marking ref processor:
2175 2175 // * Reference discovery is enabled at initial marking.
2176 2176 // * Reference discovery is disabled and the discovered
2177 2177 // references processed etc during remarking.
2178 2178 // * Reference discovery is MT (see below).
2179 2179 // * Reference discovery requires a barrier (see below).
2180 2180 // * Reference processing may or may not be MT
2181 2181 // (depending on the value of ParallelRefProcEnabled
2182 2182 // and ParallelGCThreads).
2183 2183 // * A full GC disables reference discovery by the CM
2184 2184 // ref processor and abandons any entries on it's
2185 2185 // discovered lists.
2186 2186 //
2187 2187 // * For the STW processor:
2188 2188 // * Non MT discovery is enabled at the start of a full GC.
2189 2189 // * Processing and enqueueing during a full GC is non-MT.
2190 2190 // * During a full GC, references are processed after marking.
2191 2191 //
2192 2192 // * Discovery (may or may not be MT) is enabled at the start
2193 2193 // of an incremental evacuation pause.
2194 2194 // * References are processed near the end of a STW evacuation pause.
2195 2195 // * For both types of GC:
2196 2196 // * Discovery is atomic - i.e. not concurrent.
2197 2197 // * Reference discovery will not need a barrier.
2198 2198
2199 2199 SharedHeap::ref_processing_init();
2200 2200 MemRegion mr = reserved_region();
2201 2201
2202 2202 // Concurrent Mark ref processor
2203 2203 _ref_processor_cm =
2204 2204 new ReferenceProcessor(mr, // span
2205 2205 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2206 2206 // mt processing
2207 2207 (int) ParallelGCThreads,
2208 2208 // degree of mt processing
2209 2209 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2210 2210 // mt discovery
2211 2211 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2212 2212 // degree of mt discovery
2213 2213 false,
2214 2214 // Reference discovery is not atomic
2215 2215 &_is_alive_closure_cm,
2216 2216 // is alive closure
2217 2217 // (for efficiency/performance)
2218 2218 true);
2219 2219 // Setting next fields of discovered
2220 2220 // lists requires a barrier.
2221 2221
2222 2222 // STW ref processor
2223 2223 _ref_processor_stw =
2224 2224 new ReferenceProcessor(mr, // span
2225 2225 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2226 2226 // mt processing
2227 2227 MAX2((int)ParallelGCThreads, 1),
2228 2228 // degree of mt processing
2229 2229 (ParallelGCThreads > 1),
2230 2230 // mt discovery
2231 2231 MAX2((int)ParallelGCThreads, 1),
2232 2232 // degree of mt discovery
2233 2233 true,
2234 2234 // Reference discovery is atomic
2235 2235 &_is_alive_closure_stw,
2236 2236 // is alive closure
2237 2237 // (for efficiency/performance)
2238 2238 false);
2239 2239 // Setting next fields of discovered
2240 2240 // lists requires a barrier.
2241 2241 }
2242 2242
2243 2243 size_t G1CollectedHeap::capacity() const {
2244 2244 return _g1_committed.byte_size();
2245 2245 }
2246 2246
2247 2247 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2248 2248 assert(!hr->continuesHumongous(), "pre-condition");
2249 2249 hr->reset_gc_time_stamp();
2250 2250 if (hr->startsHumongous()) {
2251 2251 uint first_index = hr->hrs_index() + 1;
2252 2252 uint last_index = hr->last_hc_index();
2253 2253 for (uint i = first_index; i < last_index; i += 1) {
2254 2254 HeapRegion* chr = region_at(i);
2255 2255 assert(chr->continuesHumongous(), "sanity");
2256 2256 chr->reset_gc_time_stamp();
2257 2257 }
2258 2258 }
2259 2259 }
2260 2260
2261 2261 #ifndef PRODUCT
2262 2262 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2263 2263 private:
2264 2264 unsigned _gc_time_stamp;
2265 2265 bool _failures;
2266 2266
2267 2267 public:
2268 2268 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2269 2269 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2270 2270
2271 2271 virtual bool doHeapRegion(HeapRegion* hr) {
2272 2272 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2273 2273 if (_gc_time_stamp != region_gc_time_stamp) {
2274 2274 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2275 2275 "expected %d", HR_FORMAT_PARAMS(hr),
2276 2276 region_gc_time_stamp, _gc_time_stamp);
2277 2277 _failures = true;
2278 2278 }
2279 2279 return false;
2280 2280 }
2281 2281
2282 2282 bool failures() { return _failures; }
2283 2283 };
2284 2284
2285 2285 void G1CollectedHeap::check_gc_time_stamps() {
2286 2286 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2287 2287 heap_region_iterate(&cl);
2288 2288 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2289 2289 }
2290 2290 #endif // PRODUCT
2291 2291
2292 2292 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2293 2293 DirtyCardQueue* into_cset_dcq,
2294 2294 bool concurrent,
2295 2295 int worker_i) {
2296 2296 // Clean cards in the hot card cache
2297 2297 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2298 2298
2299 2299 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2300 2300 int n_completed_buffers = 0;
2301 2301 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2302 2302 n_completed_buffers++;
2303 2303 }
2304 2304 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2305 2305 dcqs.clear_n_completed_buffers();
2306 2306 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2307 2307 }
2308 2308
2309 2309
2310 2310 // Computes the sum of the storage used by the various regions.
2311 2311
2312 2312 size_t G1CollectedHeap::used() const {
2313 2313 assert(Heap_lock->owner() != NULL,
2314 2314 "Should be owned on this thread's behalf.");
2315 2315 size_t result = _summary_bytes_used;
2316 2316 // Read only once in case it is set to NULL concurrently
2317 2317 HeapRegion* hr = _mutator_alloc_region.get();
2318 2318 if (hr != NULL)
2319 2319 result += hr->used();
2320 2320 return result;
2321 2321 }
2322 2322
2323 2323 size_t G1CollectedHeap::used_unlocked() const {
2324 2324 size_t result = _summary_bytes_used;
2325 2325 return result;
2326 2326 }
2327 2327
2328 2328 class SumUsedClosure: public HeapRegionClosure {
2329 2329 size_t _used;
2330 2330 public:
2331 2331 SumUsedClosure() : _used(0) {}
2332 2332 bool doHeapRegion(HeapRegion* r) {
2333 2333 if (!r->continuesHumongous()) {
2334 2334 _used += r->used();
2335 2335 }
2336 2336 return false;
2337 2337 }
2338 2338 size_t result() { return _used; }
2339 2339 };
2340 2340
2341 2341 size_t G1CollectedHeap::recalculate_used() const {
2342 2342 SumUsedClosure blk;
2343 2343 heap_region_iterate(&blk);
2344 2344 return blk.result();
2345 2345 }
2346 2346
2347 2347 size_t G1CollectedHeap::unsafe_max_alloc() {
2348 2348 if (free_regions() > 0) return HeapRegion::GrainBytes;
2349 2349 // otherwise, is there space in the current allocation region?
2350 2350
2351 2351 // We need to store the current allocation region in a local variable
2352 2352 // here. The problem is that this method doesn't take any locks and
2353 2353 // there may be other threads which overwrite the current allocation
2354 2354 // region field. attempt_allocation(), for example, sets it to NULL
2355 2355 // and this can happen *after* the NULL check here but before the call
2356 2356 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2357 2357 // to be a problem in the optimized build, since the two loads of the
2358 2358 // current allocation region field are optimized away.
2359 2359 HeapRegion* hr = _mutator_alloc_region.get();
2360 2360 if (hr == NULL) {
2361 2361 return 0;
2362 2362 }
2363 2363 return hr->free();
2364 2364 }
2365 2365
2366 2366 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2367 2367 switch (cause) {
2368 2368 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2369 2369 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2370 2370 case GCCause::_g1_humongous_allocation: return true;
2371 2371 default: return false;
2372 2372 }
2373 2373 }
2374 2374
2375 2375 #ifndef PRODUCT
2376 2376 void G1CollectedHeap::allocate_dummy_regions() {
2377 2377 // Let's fill up most of the region
2378 2378 size_t word_size = HeapRegion::GrainWords - 1024;
2379 2379 // And as a result the region we'll allocate will be humongous.
2380 2380 guarantee(isHumongous(word_size), "sanity");
2381 2381
2382 2382 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2383 2383 // Let's use the existing mechanism for the allocation
2384 2384 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2385 2385 if (dummy_obj != NULL) {
2386 2386 MemRegion mr(dummy_obj, word_size);
2387 2387 CollectedHeap::fill_with_object(mr);
2388 2388 } else {
2389 2389 // If we can't allocate once, we probably cannot allocate
2390 2390 // again. Let's get out of the loop.
2391 2391 break;
2392 2392 }
2393 2393 }
2394 2394 }
2395 2395 #endif // !PRODUCT
2396 2396
2397 2397 void G1CollectedHeap::increment_old_marking_cycles_started() {
2398 2398 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2399 2399 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2400 2400 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2401 2401 _old_marking_cycles_started, _old_marking_cycles_completed));
2402 2402
2403 2403 _old_marking_cycles_started++;
2404 2404 }
2405 2405
2406 2406 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2407 2407 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2408 2408
2409 2409 // We assume that if concurrent == true, then the caller is a
2410 2410 // concurrent thread that was joined the Suspendible Thread
2411 2411 // Set. If there's ever a cheap way to check this, we should add an
2412 2412 // assert here.
2413 2413
2414 2414 // Given that this method is called at the end of a Full GC or of a
2415 2415 // concurrent cycle, and those can be nested (i.e., a Full GC can
2416 2416 // interrupt a concurrent cycle), the number of full collections
2417 2417 // completed should be either one (in the case where there was no
2418 2418 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2419 2419 // behind the number of full collections started.
2420 2420
2421 2421 // This is the case for the inner caller, i.e. a Full GC.
2422 2422 assert(concurrent ||
2423 2423 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2424 2424 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2425 2425 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2426 2426 "is inconsistent with _old_marking_cycles_completed = %u",
2427 2427 _old_marking_cycles_started, _old_marking_cycles_completed));
2428 2428
2429 2429 // This is the case for the outer caller, i.e. the concurrent cycle.
2430 2430 assert(!concurrent ||
2431 2431 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2432 2432 err_msg("for outer caller (concurrent cycle): "
2433 2433 "_old_marking_cycles_started = %u "
2434 2434 "is inconsistent with _old_marking_cycles_completed = %u",
2435 2435 _old_marking_cycles_started, _old_marking_cycles_completed));
2436 2436
2437 2437 _old_marking_cycles_completed += 1;
2438 2438
2439 2439 // We need to clear the "in_progress" flag in the CM thread before
2440 2440 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2441 2441 // is set) so that if a waiter requests another System.gc() it doesn't
2442 2442 // incorrectly see that a marking cyle is still in progress.
2443 2443 if (concurrent) {
2444 2444 _cmThread->clear_in_progress();
2445 2445 }
2446 2446
2447 2447 // This notify_all() will ensure that a thread that called
2448 2448 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2449 2449 // and it's waiting for a full GC to finish will be woken up. It is
2450 2450 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2451 2451 FullGCCount_lock->notify_all();
2452 2452 }
2453 2453
2454 2454 void G1CollectedHeap::collect(GCCause::Cause cause) {
2455 2455 assert_heap_not_locked();
2456 2456
2457 2457 unsigned int gc_count_before;
2458 2458 unsigned int old_marking_count_before;
2459 2459 bool retry_gc;
2460 2460
2461 2461 do {
2462 2462 retry_gc = false;
2463 2463
2464 2464 {
2465 2465 MutexLocker ml(Heap_lock);
2466 2466
2467 2467 // Read the GC count while holding the Heap_lock
2468 2468 gc_count_before = total_collections();
2469 2469 old_marking_count_before = _old_marking_cycles_started;
2470 2470 }
2471 2471
2472 2472 if (should_do_concurrent_full_gc(cause)) {
2473 2473 // Schedule an initial-mark evacuation pause that will start a
2474 2474 // concurrent cycle. We're setting word_size to 0 which means that
2475 2475 // we are not requesting a post-GC allocation.
2476 2476 VM_G1IncCollectionPause op(gc_count_before,
2477 2477 0, /* word_size */
2478 2478 true, /* should_initiate_conc_mark */
2479 2479 g1_policy()->max_pause_time_ms(),
2480 2480 cause);
2481 2481
2482 2482 VMThread::execute(&op);
2483 2483 if (!op.pause_succeeded()) {
2484 2484 if (old_marking_count_before == _old_marking_cycles_started) {
2485 2485 retry_gc = op.should_retry_gc();
2486 2486 } else {
2487 2487 // A Full GC happened while we were trying to schedule the
2488 2488 // initial-mark GC. No point in starting a new cycle given
2489 2489 // that the whole heap was collected anyway.
2490 2490 }
2491 2491
2492 2492 if (retry_gc) {
2493 2493 if (GC_locker::is_active_and_needs_gc()) {
2494 2494 GC_locker::stall_until_clear();
2495 2495 }
2496 2496 }
2497 2497 }
2498 2498 } else {
2499 2499 if (cause == GCCause::_gc_locker
2500 2500 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2501 2501
2502 2502 // Schedule a standard evacuation pause. We're setting word_size
2503 2503 // to 0 which means that we are not requesting a post-GC allocation.
2504 2504 VM_G1IncCollectionPause op(gc_count_before,
2505 2505 0, /* word_size */
2506 2506 false, /* should_initiate_conc_mark */
2507 2507 g1_policy()->max_pause_time_ms(),
2508 2508 cause);
2509 2509 VMThread::execute(&op);
2510 2510 } else {
2511 2511 // Schedule a Full GC.
2512 2512 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2513 2513 VMThread::execute(&op);
2514 2514 }
2515 2515 }
2516 2516 } while (retry_gc);
2517 2517 }
2518 2518
2519 2519 bool G1CollectedHeap::is_in(const void* p) const {
2520 2520 if (_g1_committed.contains(p)) {
2521 2521 // Given that we know that p is in the committed space,
2522 2522 // heap_region_containing_raw() should successfully
2523 2523 // return the containing region.
2524 2524 HeapRegion* hr = heap_region_containing_raw(p);
2525 2525 return hr->is_in(p);
2526 2526 } else {
2527 2527 return false;
2528 2528 }
2529 2529 }
2530 2530
2531 2531 // Iteration functions.
2532 2532
2533 2533 // Iterates an OopClosure over all ref-containing fields of objects
2534 2534 // within a HeapRegion.
2535 2535
2536 2536 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2537 2537 MemRegion _mr;
2538 2538 ExtendedOopClosure* _cl;
2539 2539 public:
2540 2540 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2541 2541 : _mr(mr), _cl(cl) {}
2542 2542 bool doHeapRegion(HeapRegion* r) {
2543 2543 if (!r->continuesHumongous()) {
2544 2544 r->oop_iterate(_cl);
2545 2545 }
2546 2546 return false;
2547 2547 }
2548 2548 };
2549 2549
2550 2550 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2551 2551 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2552 2552 heap_region_iterate(&blk);
2553 2553 }
2554 2554
2555 2555 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2556 2556 IterateOopClosureRegionClosure blk(mr, cl);
2557 2557 heap_region_iterate(&blk);
2558 2558 }
2559 2559
2560 2560 // Iterates an ObjectClosure over all objects within a HeapRegion.
2561 2561
2562 2562 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2563 2563 ObjectClosure* _cl;
2564 2564 public:
2565 2565 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2566 2566 bool doHeapRegion(HeapRegion* r) {
2567 2567 if (! r->continuesHumongous()) {
2568 2568 r->object_iterate(_cl);
2569 2569 }
2570 2570 return false;
2571 2571 }
2572 2572 };
2573 2573
2574 2574 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2575 2575 IterateObjectClosureRegionClosure blk(cl);
2576 2576 heap_region_iterate(&blk);
2577 2577 }
2578 2578
2579 2579 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2580 2580 // FIXME: is this right?
2581 2581 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2582 2582 }
2583 2583
2584 2584 // Calls a SpaceClosure on a HeapRegion.
2585 2585
2586 2586 class SpaceClosureRegionClosure: public HeapRegionClosure {
2587 2587 SpaceClosure* _cl;
2588 2588 public:
2589 2589 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2590 2590 bool doHeapRegion(HeapRegion* r) {
2591 2591 _cl->do_space(r);
2592 2592 return false;
2593 2593 }
2594 2594 };
2595 2595
2596 2596 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2597 2597 SpaceClosureRegionClosure blk(cl);
2598 2598 heap_region_iterate(&blk);
2599 2599 }
2600 2600
2601 2601 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2602 2602 _hrs.iterate(cl);
2603 2603 }
2604 2604
2605 2605 void
2606 2606 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2607 2607 uint worker_id,
2608 2608 uint no_of_par_workers,
2609 2609 jint claim_value) {
2610 2610 const uint regions = n_regions();
2611 2611 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2612 2612 no_of_par_workers :
2613 2613 1);
2614 2614 assert(UseDynamicNumberOfGCThreads ||
2615 2615 no_of_par_workers == workers()->total_workers(),
2616 2616 "Non dynamic should use fixed number of workers");
2617 2617 // try to spread out the starting points of the workers
2618 2618 const HeapRegion* start_hr =
2619 2619 start_region_for_worker(worker_id, no_of_par_workers);
2620 2620 const uint start_index = start_hr->hrs_index();
2621 2621
2622 2622 // each worker will actually look at all regions
2623 2623 for (uint count = 0; count < regions; ++count) {
2624 2624 const uint index = (start_index + count) % regions;
2625 2625 assert(0 <= index && index < regions, "sanity");
2626 2626 HeapRegion* r = region_at(index);
2627 2627 // we'll ignore "continues humongous" regions (we'll process them
2628 2628 // when we come across their corresponding "start humongous"
2629 2629 // region) and regions already claimed
2630 2630 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2631 2631 continue;
2632 2632 }
2633 2633 // OK, try to claim it
2634 2634 if (r->claimHeapRegion(claim_value)) {
2635 2635 // success!
2636 2636 assert(!r->continuesHumongous(), "sanity");
2637 2637 if (r->startsHumongous()) {
2638 2638 // If the region is "starts humongous" we'll iterate over its
2639 2639 // "continues humongous" first; in fact we'll do them
2640 2640 // first. The order is important. In on case, calling the
2641 2641 // closure on the "starts humongous" region might de-allocate
2642 2642 // and clear all its "continues humongous" regions and, as a
2643 2643 // result, we might end up processing them twice. So, we'll do
2644 2644 // them first (notice: most closures will ignore them anyway) and
2645 2645 // then we'll do the "starts humongous" region.
2646 2646 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2647 2647 HeapRegion* chr = region_at(ch_index);
2648 2648
2649 2649 // if the region has already been claimed or it's not
2650 2650 // "continues humongous" we're done
2651 2651 if (chr->claim_value() == claim_value ||
2652 2652 !chr->continuesHumongous()) {
2653 2653 break;
2654 2654 }
2655 2655
2656 2656 // Noone should have claimed it directly. We can given
2657 2657 // that we claimed its "starts humongous" region.
2658 2658 assert(chr->claim_value() != claim_value, "sanity");
2659 2659 assert(chr->humongous_start_region() == r, "sanity");
2660 2660
2661 2661 if (chr->claimHeapRegion(claim_value)) {
2662 2662 // we should always be able to claim it; noone else should
2663 2663 // be trying to claim this region
2664 2664
2665 2665 bool res2 = cl->doHeapRegion(chr);
2666 2666 assert(!res2, "Should not abort");
2667 2667
2668 2668 // Right now, this holds (i.e., no closure that actually
2669 2669 // does something with "continues humongous" regions
2670 2670 // clears them). We might have to weaken it in the future,
2671 2671 // but let's leave these two asserts here for extra safety.
2672 2672 assert(chr->continuesHumongous(), "should still be the case");
2673 2673 assert(chr->humongous_start_region() == r, "sanity");
2674 2674 } else {
2675 2675 guarantee(false, "we should not reach here");
2676 2676 }
2677 2677 }
2678 2678 }
2679 2679
2680 2680 assert(!r->continuesHumongous(), "sanity");
2681 2681 bool res = cl->doHeapRegion(r);
2682 2682 assert(!res, "Should not abort");
2683 2683 }
2684 2684 }
2685 2685 }
2686 2686
2687 2687 class ResetClaimValuesClosure: public HeapRegionClosure {
2688 2688 public:
2689 2689 bool doHeapRegion(HeapRegion* r) {
2690 2690 r->set_claim_value(HeapRegion::InitialClaimValue);
2691 2691 return false;
2692 2692 }
2693 2693 };
2694 2694
2695 2695 void G1CollectedHeap::reset_heap_region_claim_values() {
2696 2696 ResetClaimValuesClosure blk;
2697 2697 heap_region_iterate(&blk);
2698 2698 }
2699 2699
2700 2700 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2701 2701 ResetClaimValuesClosure blk;
2702 2702 collection_set_iterate(&blk);
2703 2703 }
2704 2704
2705 2705 #ifdef ASSERT
2706 2706 // This checks whether all regions in the heap have the correct claim
2707 2707 // value. I also piggy-backed on this a check to ensure that the
2708 2708 // humongous_start_region() information on "continues humongous"
2709 2709 // regions is correct.
2710 2710
2711 2711 class CheckClaimValuesClosure : public HeapRegionClosure {
2712 2712 private:
2713 2713 jint _claim_value;
2714 2714 uint _failures;
2715 2715 HeapRegion* _sh_region;
2716 2716
2717 2717 public:
2718 2718 CheckClaimValuesClosure(jint claim_value) :
2719 2719 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2720 2720 bool doHeapRegion(HeapRegion* r) {
2721 2721 if (r->claim_value() != _claim_value) {
2722 2722 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2723 2723 "claim value = %d, should be %d",
2724 2724 HR_FORMAT_PARAMS(r),
2725 2725 r->claim_value(), _claim_value);
2726 2726 ++_failures;
2727 2727 }
2728 2728 if (!r->isHumongous()) {
2729 2729 _sh_region = NULL;
2730 2730 } else if (r->startsHumongous()) {
2731 2731 _sh_region = r;
2732 2732 } else if (r->continuesHumongous()) {
2733 2733 if (r->humongous_start_region() != _sh_region) {
2734 2734 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2735 2735 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2736 2736 HR_FORMAT_PARAMS(r),
2737 2737 r->humongous_start_region(),
2738 2738 _sh_region);
2739 2739 ++_failures;
2740 2740 }
2741 2741 }
2742 2742 return false;
2743 2743 }
2744 2744 uint failures() { return _failures; }
2745 2745 };
2746 2746
2747 2747 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2748 2748 CheckClaimValuesClosure cl(claim_value);
2749 2749 heap_region_iterate(&cl);
2750 2750 return cl.failures() == 0;
2751 2751 }
2752 2752
2753 2753 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2754 2754 private:
2755 2755 jint _claim_value;
2756 2756 uint _failures;
2757 2757
2758 2758 public:
2759 2759 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2760 2760 _claim_value(claim_value), _failures(0) { }
2761 2761
2762 2762 uint failures() { return _failures; }
2763 2763
2764 2764 bool doHeapRegion(HeapRegion* hr) {
2765 2765 assert(hr->in_collection_set(), "how?");
2766 2766 assert(!hr->isHumongous(), "H-region in CSet");
2767 2767 if (hr->claim_value() != _claim_value) {
2768 2768 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2769 2769 "claim value = %d, should be %d",
2770 2770 HR_FORMAT_PARAMS(hr),
2771 2771 hr->claim_value(), _claim_value);
2772 2772 _failures += 1;
2773 2773 }
2774 2774 return false;
2775 2775 }
2776 2776 };
2777 2777
2778 2778 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2779 2779 CheckClaimValuesInCSetHRClosure cl(claim_value);
2780 2780 collection_set_iterate(&cl);
2781 2781 return cl.failures() == 0;
2782 2782 }
2783 2783 #endif // ASSERT
2784 2784
2785 2785 // Clear the cached CSet starting regions and (more importantly)
2786 2786 // the time stamps. Called when we reset the GC time stamp.
2787 2787 void G1CollectedHeap::clear_cset_start_regions() {
2788 2788 assert(_worker_cset_start_region != NULL, "sanity");
2789 2789 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2790 2790
2791 2791 int n_queues = MAX2((int)ParallelGCThreads, 1);
2792 2792 for (int i = 0; i < n_queues; i++) {
2793 2793 _worker_cset_start_region[i] = NULL;
2794 2794 _worker_cset_start_region_time_stamp[i] = 0;
2795 2795 }
2796 2796 }
2797 2797
2798 2798 // Given the id of a worker, obtain or calculate a suitable
2799 2799 // starting region for iterating over the current collection set.
2800 2800 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2801 2801 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2802 2802
2803 2803 HeapRegion* result = NULL;
2804 2804 unsigned gc_time_stamp = get_gc_time_stamp();
2805 2805
2806 2806 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2807 2807 // Cached starting region for current worker was set
2808 2808 // during the current pause - so it's valid.
2809 2809 // Note: the cached starting heap region may be NULL
2810 2810 // (when the collection set is empty).
2811 2811 result = _worker_cset_start_region[worker_i];
2812 2812 assert(result == NULL || result->in_collection_set(), "sanity");
2813 2813 return result;
2814 2814 }
2815 2815
2816 2816 // The cached entry was not valid so let's calculate
2817 2817 // a suitable starting heap region for this worker.
2818 2818
2819 2819 // We want the parallel threads to start their collection
2820 2820 // set iteration at different collection set regions to
2821 2821 // avoid contention.
2822 2822 // If we have:
2823 2823 // n collection set regions
2824 2824 // p threads
2825 2825 // Then thread t will start at region floor ((t * n) / p)
2826 2826
2827 2827 result = g1_policy()->collection_set();
2828 2828 if (G1CollectedHeap::use_parallel_gc_threads()) {
2829 2829 uint cs_size = g1_policy()->cset_region_length();
2830 2830 uint active_workers = workers()->active_workers();
2831 2831 assert(UseDynamicNumberOfGCThreads ||
2832 2832 active_workers == workers()->total_workers(),
2833 2833 "Unless dynamic should use total workers");
2834 2834
2835 2835 uint end_ind = (cs_size * worker_i) / active_workers;
2836 2836 uint start_ind = 0;
2837 2837
2838 2838 if (worker_i > 0 &&
2839 2839 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2840 2840 // Previous workers starting region is valid
2841 2841 // so let's iterate from there
2842 2842 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2843 2843 result = _worker_cset_start_region[worker_i - 1];
2844 2844 }
2845 2845
2846 2846 for (uint i = start_ind; i < end_ind; i++) {
2847 2847 result = result->next_in_collection_set();
2848 2848 }
2849 2849 }
2850 2850
2851 2851 // Note: the calculated starting heap region may be NULL
2852 2852 // (when the collection set is empty).
2853 2853 assert(result == NULL || result->in_collection_set(), "sanity");
2854 2854 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2855 2855 "should be updated only once per pause");
2856 2856 _worker_cset_start_region[worker_i] = result;
2857 2857 OrderAccess::storestore();
2858 2858 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2859 2859 return result;
2860 2860 }
2861 2861
2862 2862 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2863 2863 uint no_of_par_workers) {
2864 2864 uint worker_num =
2865 2865 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2866 2866 assert(UseDynamicNumberOfGCThreads ||
2867 2867 no_of_par_workers == workers()->total_workers(),
2868 2868 "Non dynamic should use fixed number of workers");
2869 2869 const uint start_index = n_regions() * worker_i / worker_num;
2870 2870 return region_at(start_index);
2871 2871 }
2872 2872
2873 2873 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2874 2874 HeapRegion* r = g1_policy()->collection_set();
2875 2875 while (r != NULL) {
2876 2876 HeapRegion* next = r->next_in_collection_set();
2877 2877 if (cl->doHeapRegion(r)) {
2878 2878 cl->incomplete();
2879 2879 return;
2880 2880 }
2881 2881 r = next;
2882 2882 }
2883 2883 }
2884 2884
2885 2885 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2886 2886 HeapRegionClosure *cl) {
2887 2887 if (r == NULL) {
2888 2888 // The CSet is empty so there's nothing to do.
2889 2889 return;
2890 2890 }
2891 2891
2892 2892 assert(r->in_collection_set(),
2893 2893 "Start region must be a member of the collection set.");
2894 2894 HeapRegion* cur = r;
2895 2895 while (cur != NULL) {
2896 2896 HeapRegion* next = cur->next_in_collection_set();
2897 2897 if (cl->doHeapRegion(cur) && false) {
2898 2898 cl->incomplete();
2899 2899 return;
2900 2900 }
2901 2901 cur = next;
2902 2902 }
2903 2903 cur = g1_policy()->collection_set();
2904 2904 while (cur != r) {
2905 2905 HeapRegion* next = cur->next_in_collection_set();
2906 2906 if (cl->doHeapRegion(cur) && false) {
2907 2907 cl->incomplete();
2908 2908 return;
2909 2909 }
2910 2910 cur = next;
2911 2911 }
2912 2912 }
2913 2913
2914 2914 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2915 2915 return n_regions() > 0 ? region_at(0) : NULL;
2916 2916 }
2917 2917
2918 2918
2919 2919 Space* G1CollectedHeap::space_containing(const void* addr) const {
2920 2920 Space* res = heap_region_containing(addr);
2921 2921 return res;
2922 2922 }
2923 2923
2924 2924 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2925 2925 Space* sp = space_containing(addr);
2926 2926 if (sp != NULL) {
2927 2927 return sp->block_start(addr);
2928 2928 }
2929 2929 return NULL;
2930 2930 }
2931 2931
2932 2932 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2933 2933 Space* sp = space_containing(addr);
2934 2934 assert(sp != NULL, "block_size of address outside of heap");
2935 2935 return sp->block_size(addr);
2936 2936 }
2937 2937
2938 2938 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2939 2939 Space* sp = space_containing(addr);
2940 2940 return sp->block_is_obj(addr);
2941 2941 }
2942 2942
2943 2943 bool G1CollectedHeap::supports_tlab_allocation() const {
2944 2944 return true;
2945 2945 }
2946 2946
2947 2947 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2948 2948 return HeapRegion::GrainBytes;
2949 2949 }
2950 2950
2951 2951 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2952 2952 // Return the remaining space in the cur alloc region, but not less than
2953 2953 // the min TLAB size.
2954 2954
2955 2955 // Also, this value can be at most the humongous object threshold,
2956 2956 // since we can't allow tlabs to grow big enough to accomodate
2957 2957 // humongous objects.
2958 2958
2959 2959 HeapRegion* hr = _mutator_alloc_region.get();
2960 2960 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2961 2961 if (hr == NULL) {
2962 2962 return max_tlab_size;
2963 2963 } else {
2964 2964 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2965 2965 }
2966 2966 }
2967 2967
2968 2968 size_t G1CollectedHeap::max_capacity() const {
2969 2969 return _g1_reserved.byte_size();
2970 2970 }
2971 2971
2972 2972 jlong G1CollectedHeap::millis_since_last_gc() {
2973 2973 // assert(false, "NYI");
2974 2974 return 0;
2975 2975 }
2976 2976
2977 2977 void G1CollectedHeap::prepare_for_verify() {
2978 2978 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2979 2979 ensure_parsability(false);
2980 2980 }
2981 2981 g1_rem_set()->prepare_for_verify();
2982 2982 }
2983 2983
2984 2984 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
2985 2985 VerifyOption vo) {
2986 2986 switch (vo) {
2987 2987 case VerifyOption_G1UsePrevMarking:
2988 2988 return hr->obj_allocated_since_prev_marking(obj);
2989 2989 case VerifyOption_G1UseNextMarking:
2990 2990 return hr->obj_allocated_since_next_marking(obj);
2991 2991 case VerifyOption_G1UseMarkWord:
2992 2992 return false;
2993 2993 default:
2994 2994 ShouldNotReachHere();
2995 2995 }
2996 2996 return false; // keep some compilers happy
2997 2997 }
2998 2998
2999 2999 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3000 3000 switch (vo) {
3001 3001 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3002 3002 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3003 3003 case VerifyOption_G1UseMarkWord: return NULL;
3004 3004 default: ShouldNotReachHere();
3005 3005 }
3006 3006 return NULL; // keep some compilers happy
3007 3007 }
3008 3008
3009 3009 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3010 3010 switch (vo) {
3011 3011 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3012 3012 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3013 3013 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3014 3014 default: ShouldNotReachHere();
3015 3015 }
3016 3016 return false; // keep some compilers happy
3017 3017 }
3018 3018
3019 3019 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3020 3020 switch (vo) {
3021 3021 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3022 3022 case VerifyOption_G1UseNextMarking: return "NTAMS";
3023 3023 case VerifyOption_G1UseMarkWord: return "NONE";
3024 3024 default: ShouldNotReachHere();
3025 3025 }
3026 3026 return NULL; // keep some compilers happy
3027 3027 }
3028 3028
3029 3029 class VerifyLivenessOopClosure: public OopClosure {
3030 3030 G1CollectedHeap* _g1h;
3031 3031 VerifyOption _vo;
3032 3032 public:
3033 3033 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3034 3034 _g1h(g1h), _vo(vo)
3035 3035 { }
3036 3036 void do_oop(narrowOop *p) { do_oop_work(p); }
3037 3037 void do_oop( oop *p) { do_oop_work(p); }
3038 3038
3039 3039 template <class T> void do_oop_work(T *p) {
3040 3040 oop obj = oopDesc::load_decode_heap_oop(p);
3041 3041 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3042 3042 "Dead object referenced by a not dead object");
3043 3043 }
3044 3044 };
3045 3045
3046 3046 class VerifyObjsInRegionClosure: public ObjectClosure {
3047 3047 private:
3048 3048 G1CollectedHeap* _g1h;
3049 3049 size_t _live_bytes;
3050 3050 HeapRegion *_hr;
3051 3051 VerifyOption _vo;
3052 3052 public:
3053 3053 // _vo == UsePrevMarking -> use "prev" marking information,
3054 3054 // _vo == UseNextMarking -> use "next" marking information,
3055 3055 // _vo == UseMarkWord -> use mark word from object header.
3056 3056 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3057 3057 : _live_bytes(0), _hr(hr), _vo(vo) {
3058 3058 _g1h = G1CollectedHeap::heap();
3059 3059 }
3060 3060 void do_object(oop o) {
3061 3061 VerifyLivenessOopClosure isLive(_g1h, _vo);
3062 3062 assert(o != NULL, "Huh?");
3063 3063 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3064 3064 // If the object is alive according to the mark word,
3065 3065 // then verify that the marking information agrees.
3066 3066 // Note we can't verify the contra-positive of the
3067 3067 // above: if the object is dead (according to the mark
3068 3068 // word), it may not be marked, or may have been marked
3069 3069 // but has since became dead, or may have been allocated
3070 3070 // since the last marking.
3071 3071 if (_vo == VerifyOption_G1UseMarkWord) {
3072 3072 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3073 3073 }
3074 3074
3075 3075 o->oop_iterate_no_header(&isLive);
3076 3076 if (!_hr->obj_allocated_since_prev_marking(o)) {
3077 3077 size_t obj_size = o->size(); // Make sure we don't overflow
3078 3078 _live_bytes += (obj_size * HeapWordSize);
3079 3079 }
3080 3080 }
3081 3081 }
3082 3082 size_t live_bytes() { return _live_bytes; }
3083 3083 };
3084 3084
3085 3085 class PrintObjsInRegionClosure : public ObjectClosure {
3086 3086 HeapRegion *_hr;
3087 3087 G1CollectedHeap *_g1;
3088 3088 public:
3089 3089 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3090 3090 _g1 = G1CollectedHeap::heap();
3091 3091 };
3092 3092
3093 3093 void do_object(oop o) {
3094 3094 if (o != NULL) {
3095 3095 HeapWord *start = (HeapWord *) o;
3096 3096 size_t word_sz = o->size();
3097 3097 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3098 3098 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3099 3099 (void*) o, word_sz,
3100 3100 _g1->isMarkedPrev(o),
3101 3101 _g1->isMarkedNext(o),
3102 3102 _hr->obj_allocated_since_prev_marking(o));
3103 3103 HeapWord *end = start + word_sz;
3104 3104 HeapWord *cur;
3105 3105 int *val;
3106 3106 for (cur = start; cur < end; cur++) {
3107 3107 val = (int *) cur;
3108 3108 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3109 3109 }
3110 3110 }
3111 3111 }
3112 3112 };
3113 3113
3114 3114 class VerifyRegionClosure: public HeapRegionClosure {
3115 3115 private:
3116 3116 bool _par;
3117 3117 VerifyOption _vo;
3118 3118 bool _failures;
3119 3119 public:
3120 3120 // _vo == UsePrevMarking -> use "prev" marking information,
3121 3121 // _vo == UseNextMarking -> use "next" marking information,
3122 3122 // _vo == UseMarkWord -> use mark word from object header.
3123 3123 VerifyRegionClosure(bool par, VerifyOption vo)
3124 3124 : _par(par),
3125 3125 _vo(vo),
3126 3126 _failures(false) {}
3127 3127
3128 3128 bool failures() {
3129 3129 return _failures;
3130 3130 }
3131 3131
3132 3132 bool doHeapRegion(HeapRegion* r) {
3133 3133 if (!r->continuesHumongous()) {
3134 3134 bool failures = false;
3135 3135 r->verify(_vo, &failures);
3136 3136 if (failures) {
3137 3137 _failures = true;
3138 3138 } else {
3139 3139 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3140 3140 r->object_iterate(¬_dead_yet_cl);
3141 3141 if (_vo != VerifyOption_G1UseNextMarking) {
3142 3142 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3143 3143 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3144 3144 "max_live_bytes "SIZE_FORMAT" "
3145 3145 "< calculated "SIZE_FORMAT,
3146 3146 r->bottom(), r->end(),
3147 3147 r->max_live_bytes(),
3148 3148 not_dead_yet_cl.live_bytes());
3149 3149 _failures = true;
3150 3150 }
3151 3151 } else {
3152 3152 // When vo == UseNextMarking we cannot currently do a sanity
3153 3153 // check on the live bytes as the calculation has not been
3154 3154 // finalized yet.
3155 3155 }
3156 3156 }
3157 3157 }
3158 3158 return false; // stop the region iteration if we hit a failure
3159 3159 }
3160 3160 };
3161 3161
3162 3162 class YoungRefCounterClosure : public OopClosure {
3163 3163 G1CollectedHeap* _g1h;
3164 3164 int _count;
3165 3165 public:
3166 3166 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3167 3167 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3168 3168 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3169 3169
3170 3170 int count() { return _count; }
3171 3171 void reset_count() { _count = 0; };
3172 3172 };
3173 3173
3174 3174 class VerifyKlassClosure: public KlassClosure {
3175 3175 YoungRefCounterClosure _young_ref_counter_closure;
3176 3176 OopClosure *_oop_closure;
3177 3177 public:
3178 3178 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3179 3179 void do_klass(Klass* k) {
3180 3180 k->oops_do(_oop_closure);
3181 3181
3182 3182 _young_ref_counter_closure.reset_count();
3183 3183 k->oops_do(&_young_ref_counter_closure);
3184 3184 if (_young_ref_counter_closure.count() > 0) {
3185 3185 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3186 3186 }
3187 3187 }
3188 3188 };
3189 3189
3190 3190 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3191 3191 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3192 3192 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3193 3193 // we can change this closure to extend the simpler OopClosure.
3194 3194 class VerifyRootsClosure: public OopsInGenClosure {
3195 3195 private:
3196 3196 G1CollectedHeap* _g1h;
3197 3197 VerifyOption _vo;
3198 3198 bool _failures;
3199 3199 public:
3200 3200 // _vo == UsePrevMarking -> use "prev" marking information,
3201 3201 // _vo == UseNextMarking -> use "next" marking information,
3202 3202 // _vo == UseMarkWord -> use mark word from object header.
3203 3203 VerifyRootsClosure(VerifyOption vo) :
3204 3204 _g1h(G1CollectedHeap::heap()),
3205 3205 _vo(vo),
3206 3206 _failures(false) { }
3207 3207
3208 3208 bool failures() { return _failures; }
3209 3209
3210 3210 template <class T> void do_oop_nv(T* p) {
3211 3211 T heap_oop = oopDesc::load_heap_oop(p);
3212 3212 if (!oopDesc::is_null(heap_oop)) {
3213 3213 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3214 3214 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3215 3215 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3216 3216 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3217 3217 if (_vo == VerifyOption_G1UseMarkWord) {
3218 3218 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3219 3219 }
3220 3220 obj->print_on(gclog_or_tty);
3221 3221 _failures = true;
3222 3222 }
3223 3223 }
3224 3224 }
3225 3225
3226 3226 void do_oop(oop* p) { do_oop_nv(p); }
3227 3227 void do_oop(narrowOop* p) { do_oop_nv(p); }
3228 3228 };
3229 3229
3230 3230 // This is the task used for parallel heap verification.
3231 3231
3232 3232 class G1ParVerifyTask: public AbstractGangTask {
3233 3233 private:
3234 3234 G1CollectedHeap* _g1h;
3235 3235 VerifyOption _vo;
3236 3236 bool _failures;
3237 3237
3238 3238 public:
3239 3239 // _vo == UsePrevMarking -> use "prev" marking information,
3240 3240 // _vo == UseNextMarking -> use "next" marking information,
3241 3241 // _vo == UseMarkWord -> use mark word from object header.
3242 3242 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3243 3243 AbstractGangTask("Parallel verify task"),
3244 3244 _g1h(g1h),
3245 3245 _vo(vo),
3246 3246 _failures(false) { }
3247 3247
3248 3248 bool failures() {
3249 3249 return _failures;
3250 3250 }
3251 3251
3252 3252 void work(uint worker_id) {
3253 3253 HandleMark hm;
3254 3254 VerifyRegionClosure blk(true, _vo);
3255 3255 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3256 3256 _g1h->workers()->active_workers(),
3257 3257 HeapRegion::ParVerifyClaimValue);
3258 3258 if (blk.failures()) {
3259 3259 _failures = true;
3260 3260 }
3261 3261 }
3262 3262 };
3263 3263
3264 3264 void G1CollectedHeap::verify(bool silent) {
3265 3265 verify(silent, VerifyOption_G1UsePrevMarking);
3266 3266 }
3267 3267
3268 3268 void G1CollectedHeap::verify(bool silent,
3269 3269 VerifyOption vo) {
3270 3270 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3271 3271 if (!silent) { gclog_or_tty->print("Roots "); }
3272 3272 VerifyRootsClosure rootsCl(vo);
3273 3273
3274 3274 assert(Thread::current()->is_VM_thread(),
3275 3275 "Expected to be executed serially by the VM thread at this point");
3276 3276
3277 3277 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3278 3278 VerifyKlassClosure klassCl(this, &rootsCl);
3279 3279
3280 3280 // We apply the relevant closures to all the oops in the
3281 3281 // system dictionary, the string table and the code cache.
3282 3282 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3283 3283
3284 3284 // Need cleared claim bits for the strong roots processing
3285 3285 ClassLoaderDataGraph::clear_claimed_marks();
3286 3286
3287 3287 process_strong_roots(true, // activate StrongRootsScope
3288 3288 false, // we set "is scavenging" to false,
3289 3289 // so we don't reset the dirty cards.
3290 3290 ScanningOption(so), // roots scanning options
3291 3291 &rootsCl,
3292 3292 &blobsCl,
3293 3293 &klassCl
3294 3294 );
3295 3295
3296 3296 bool failures = rootsCl.failures();
3297 3297
3298 3298 if (vo != VerifyOption_G1UseMarkWord) {
3299 3299 // If we're verifying during a full GC then the region sets
3300 3300 // will have been torn down at the start of the GC. Therefore
3301 3301 // verifying the region sets will fail. So we only verify
3302 3302 // the region sets when not in a full GC.
3303 3303 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3304 3304 verify_region_sets();
3305 3305 }
3306 3306
3307 3307 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3308 3308 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3309 3309 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3310 3310 "sanity check");
3311 3311
3312 3312 G1ParVerifyTask task(this, vo);
3313 3313 assert(UseDynamicNumberOfGCThreads ||
3314 3314 workers()->active_workers() == workers()->total_workers(),
3315 3315 "If not dynamic should be using all the workers");
3316 3316 int n_workers = workers()->active_workers();
3317 3317 set_par_threads(n_workers);
3318 3318 workers()->run_task(&task);
3319 3319 set_par_threads(0);
3320 3320 if (task.failures()) {
3321 3321 failures = true;
3322 3322 }
3323 3323
3324 3324 // Checks that the expected amount of parallel work was done.
3325 3325 // The implication is that n_workers is > 0.
3326 3326 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3327 3327 "sanity check");
3328 3328
3329 3329 reset_heap_region_claim_values();
3330 3330
3331 3331 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3332 3332 "sanity check");
3333 3333 } else {
3334 3334 VerifyRegionClosure blk(false, vo);
3335 3335 heap_region_iterate(&blk);
3336 3336 if (blk.failures()) {
3337 3337 failures = true;
3338 3338 }
3339 3339 }
3340 3340 if (!silent) gclog_or_tty->print("RemSet ");
3341 3341 rem_set()->verify();
3342 3342
3343 3343 if (failures) {
3344 3344 gclog_or_tty->print_cr("Heap:");
3345 3345 // It helps to have the per-region information in the output to
3346 3346 // help us track down what went wrong. This is why we call
3347 3347 // print_extended_on() instead of print_on().
3348 3348 print_extended_on(gclog_or_tty);
3349 3349 gclog_or_tty->print_cr("");
3350 3350 #ifndef PRODUCT
3351 3351 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3352 3352 concurrent_mark()->print_reachable("at-verification-failure",
3353 3353 vo, false /* all */);
3354 3354 }
3355 3355 #endif
3356 3356 gclog_or_tty->flush();
3357 3357 }
3358 3358 guarantee(!failures, "there should not have been any failures");
3359 3359 } else {
3360 3360 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3361 3361 }
3362 3362 }
3363 3363
3364 3364 class PrintRegionClosure: public HeapRegionClosure {
3365 3365 outputStream* _st;
3366 3366 public:
3367 3367 PrintRegionClosure(outputStream* st) : _st(st) {}
3368 3368 bool doHeapRegion(HeapRegion* r) {
3369 3369 r->print_on(_st);
3370 3370 return false;
3371 3371 }
3372 3372 };
3373 3373
3374 3374 void G1CollectedHeap::print_on(outputStream* st) const {
3375 3375 st->print(" %-20s", "garbage-first heap");
3376 3376 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3377 3377 capacity()/K, used_unlocked()/K);
3378 3378 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3379 3379 _g1_storage.low_boundary(),
3380 3380 _g1_storage.high(),
3381 3381 _g1_storage.high_boundary());
3382 3382 st->cr();
3383 3383 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3384 3384 uint young_regions = _young_list->length();
3385 3385 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3386 3386 (size_t) young_regions * HeapRegion::GrainBytes / K);
3387 3387 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3388 3388 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3389 3389 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3390 3390 st->cr();
3391 3391 MetaspaceAux::print_on(st);
3392 3392 }
3393 3393
3394 3394 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3395 3395 print_on(st);
3396 3396
3397 3397 // Print the per-region information.
3398 3398 st->cr();
3399 3399 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3400 3400 "HS=humongous(starts), HC=humongous(continues), "
3401 3401 "CS=collection set, F=free, TS=gc time stamp, "
3402 3402 "PTAMS=previous top-at-mark-start, "
3403 3403 "NTAMS=next top-at-mark-start)");
3404 3404 PrintRegionClosure blk(st);
3405 3405 heap_region_iterate(&blk);
3406 3406 }
3407 3407
3408 3408 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3409 3409 if (G1CollectedHeap::use_parallel_gc_threads()) {
3410 3410 workers()->print_worker_threads_on(st);
3411 3411 }
3412 3412 _cmThread->print_on(st);
3413 3413 st->cr();
3414 3414 _cm->print_worker_threads_on(st);
3415 3415 _cg1r->print_worker_threads_on(st);
3416 3416 }
3417 3417
3418 3418 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3419 3419 if (G1CollectedHeap::use_parallel_gc_threads()) {
3420 3420 workers()->threads_do(tc);
3421 3421 }
3422 3422 tc->do_thread(_cmThread);
3423 3423 _cg1r->threads_do(tc);
3424 3424 }
3425 3425
3426 3426 void G1CollectedHeap::print_tracing_info() const {
3427 3427 // We'll overload this to mean "trace GC pause statistics."
3428 3428 if (TraceGen0Time || TraceGen1Time) {
3429 3429 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3430 3430 // to that.
3431 3431 g1_policy()->print_tracing_info();
3432 3432 }
3433 3433 if (G1SummarizeRSetStats) {
3434 3434 g1_rem_set()->print_summary_info();
3435 3435 }
3436 3436 if (G1SummarizeConcMark) {
3437 3437 concurrent_mark()->print_summary_info();
3438 3438 }
3439 3439 g1_policy()->print_yg_surv_rate_info();
3440 3440 SpecializationStats::print();
3441 3441 }
3442 3442
3443 3443 #ifndef PRODUCT
3444 3444 // Helpful for debugging RSet issues.
3445 3445
3446 3446 class PrintRSetsClosure : public HeapRegionClosure {
3447 3447 private:
3448 3448 const char* _msg;
3449 3449 size_t _occupied_sum;
3450 3450
3451 3451 public:
3452 3452 bool doHeapRegion(HeapRegion* r) {
3453 3453 HeapRegionRemSet* hrrs = r->rem_set();
3454 3454 size_t occupied = hrrs->occupied();
3455 3455 _occupied_sum += occupied;
3456 3456
3457 3457 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3458 3458 HR_FORMAT_PARAMS(r));
3459 3459 if (occupied == 0) {
3460 3460 gclog_or_tty->print_cr(" RSet is empty");
3461 3461 } else {
3462 3462 hrrs->print();
3463 3463 }
3464 3464 gclog_or_tty->print_cr("----------");
3465 3465 return false;
3466 3466 }
3467 3467
3468 3468 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3469 3469 gclog_or_tty->cr();
3470 3470 gclog_or_tty->print_cr("========================================");
3471 3471 gclog_or_tty->print_cr(msg);
3472 3472 gclog_or_tty->cr();
3473 3473 }
3474 3474
3475 3475 ~PrintRSetsClosure() {
3476 3476 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3477 3477 gclog_or_tty->print_cr("========================================");
3478 3478 gclog_or_tty->cr();
3479 3479 }
3480 3480 };
3481 3481
3482 3482 void G1CollectedHeap::print_cset_rsets() {
3483 3483 PrintRSetsClosure cl("Printing CSet RSets");
3484 3484 collection_set_iterate(&cl);
3485 3485 }
3486 3486
3487 3487 void G1CollectedHeap::print_all_rsets() {
3488 3488 PrintRSetsClosure cl("Printing All RSets");;
3489 3489 heap_region_iterate(&cl);
3490 3490 }
3491 3491 #endif // PRODUCT
3492 3492
3493 3493 G1CollectedHeap* G1CollectedHeap::heap() {
3494 3494 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3495 3495 "not a garbage-first heap");
3496 3496 return _g1h;
3497 3497 }
3498 3498
3499 3499 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3500 3500 // always_do_update_barrier = false;
3501 3501 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3502 3502 // Call allocation profiler
3503 3503 AllocationProfiler::iterate_since_last_gc();
3504 3504 // Fill TLAB's and such
3505 3505 ensure_parsability(true);
3506 3506 }
3507 3507
3508 3508 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3509 3509 // FIXME: what is this about?
3510 3510 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3511 3511 // is set.
3512 3512 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3513 3513 "derived pointer present"));
3514 3514 // always_do_update_barrier = true;
3515 3515
3516 3516 // We have just completed a GC. Update the soft reference
3517 3517 // policy with the new heap occupancy
3518 3518 Universe::update_heap_info_at_gc();
3519 3519 }
3520 3520
3521 3521 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3522 3522 unsigned int gc_count_before,
3523 3523 bool* succeeded) {
3524 3524 assert_heap_not_locked_and_not_at_safepoint();
3525 3525 g1_policy()->record_stop_world_start();
3526 3526 VM_G1IncCollectionPause op(gc_count_before,
3527 3527 word_size,
3528 3528 false, /* should_initiate_conc_mark */
3529 3529 g1_policy()->max_pause_time_ms(),
3530 3530 GCCause::_g1_inc_collection_pause);
3531 3531 VMThread::execute(&op);
3532 3532
3533 3533 HeapWord* result = op.result();
3534 3534 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3535 3535 assert(result == NULL || ret_succeeded,
3536 3536 "the result should be NULL if the VM did not succeed");
3537 3537 *succeeded = ret_succeeded;
3538 3538
3539 3539 assert_heap_not_locked();
3540 3540 return result;
3541 3541 }
3542 3542
3543 3543 void
3544 3544 G1CollectedHeap::doConcurrentMark() {
3545 3545 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3546 3546 if (!_cmThread->in_progress()) {
3547 3547 _cmThread->set_started();
3548 3548 CGC_lock->notify();
3549 3549 }
3550 3550 }
3551 3551
3552 3552 size_t G1CollectedHeap::pending_card_num() {
3553 3553 size_t extra_cards = 0;
3554 3554 JavaThread *curr = Threads::first();
3555 3555 while (curr != NULL) {
3556 3556 DirtyCardQueue& dcq = curr->dirty_card_queue();
3557 3557 extra_cards += dcq.size();
3558 3558 curr = curr->next();
3559 3559 }
3560 3560 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3561 3561 size_t buffer_size = dcqs.buffer_size();
3562 3562 size_t buffer_num = dcqs.completed_buffers_num();
3563 3563
3564 3564 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3565 3565 // in bytes - not the number of 'entries'. We need to convert
3566 3566 // into a number of cards.
3567 3567 return (buffer_size * buffer_num + extra_cards) / oopSize;
3568 3568 }
3569 3569
3570 3570 size_t G1CollectedHeap::cards_scanned() {
3571 3571 return g1_rem_set()->cardsScanned();
3572 3572 }
3573 3573
3574 3574 void
3575 3575 G1CollectedHeap::setup_surviving_young_words() {
3576 3576 assert(_surviving_young_words == NULL, "pre-condition");
3577 3577 uint array_length = g1_policy()->young_cset_region_length();
3578 3578 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3579 3579 if (_surviving_young_words == NULL) {
3580 3580 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3581 3581 "Not enough space for young surv words summary.");
3582 3582 }
3583 3583 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3584 3584 #ifdef ASSERT
3585 3585 for (uint i = 0; i < array_length; ++i) {
3586 3586 assert( _surviving_young_words[i] == 0, "memset above" );
3587 3587 }
3588 3588 #endif // !ASSERT
3589 3589 }
3590 3590
3591 3591 void
3592 3592 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3593 3593 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3594 3594 uint array_length = g1_policy()->young_cset_region_length();
3595 3595 for (uint i = 0; i < array_length; ++i) {
3596 3596 _surviving_young_words[i] += surv_young_words[i];
3597 3597 }
3598 3598 }
3599 3599
3600 3600 void
3601 3601 G1CollectedHeap::cleanup_surviving_young_words() {
3602 3602 guarantee( _surviving_young_words != NULL, "pre-condition" );
3603 3603 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3604 3604 _surviving_young_words = NULL;
3605 3605 }
3606 3606
3607 3607 #ifdef ASSERT
3608 3608 class VerifyCSetClosure: public HeapRegionClosure {
3609 3609 public:
3610 3610 bool doHeapRegion(HeapRegion* hr) {
3611 3611 // Here we check that the CSet region's RSet is ready for parallel
3612 3612 // iteration. The fields that we'll verify are only manipulated
3613 3613 // when the region is part of a CSet and is collected. Afterwards,
3614 3614 // we reset these fields when we clear the region's RSet (when the
3615 3615 // region is freed) so they are ready when the region is
3616 3616 // re-allocated. The only exception to this is if there's an
3617 3617 // evacuation failure and instead of freeing the region we leave
3618 3618 // it in the heap. In that case, we reset these fields during
3619 3619 // evacuation failure handling.
3620 3620 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3621 3621
3622 3622 // Here's a good place to add any other checks we'd like to
3623 3623 // perform on CSet regions.
3624 3624 return false;
3625 3625 }
3626 3626 };
3627 3627 #endif // ASSERT
3628 3628
3629 3629 #if TASKQUEUE_STATS
3630 3630 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3631 3631 st->print_raw_cr("GC Task Stats");
3632 3632 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3633 3633 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3634 3634 }
3635 3635
3636 3636 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3637 3637 print_taskqueue_stats_hdr(st);
3638 3638
3639 3639 TaskQueueStats totals;
3640 3640 const int n = workers() != NULL ? workers()->total_workers() : 1;
3641 3641 for (int i = 0; i < n; ++i) {
3642 3642 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3643 3643 totals += task_queue(i)->stats;
3644 3644 }
3645 3645 st->print_raw("tot "); totals.print(st); st->cr();
3646 3646
3647 3647 DEBUG_ONLY(totals.verify());
3648 3648 }
3649 3649
3650 3650 void G1CollectedHeap::reset_taskqueue_stats() {
3651 3651 const int n = workers() != NULL ? workers()->total_workers() : 1;
3652 3652 for (int i = 0; i < n; ++i) {
3653 3653 task_queue(i)->stats.reset();
3654 3654 }
3655 3655 }
3656 3656 #endif // TASKQUEUE_STATS
3657 3657
3658 3658 void G1CollectedHeap::log_gc_header() {
3659 3659 if (!G1Log::fine()) {
3660 3660 return;
3661 3661 }
3662 3662
3663 3663 gclog_or_tty->date_stamp(PrintGCDateStamps);
3664 3664 gclog_or_tty->stamp(PrintGCTimeStamps);
3665 3665
3666 3666 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3667 3667 .append(g1_policy()->gcs_are_young() ? " (young)" : " (mixed)")
3668 3668 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3669 3669
3670 3670 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3671 3671 }
3672 3672
3673 3673 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3674 3674 if (!G1Log::fine()) {
3675 3675 return;
3676 3676 }
3677 3677
3678 3678 if (G1Log::finer()) {
3679 3679 if (evacuation_failed()) {
3680 3680 gclog_or_tty->print(" (to-space exhausted)");
3681 3681 }
3682 3682 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
|
↓ open down ↓ |
3682 lines elided |
↑ open up ↑ |
3683 3683 g1_policy()->phase_times()->note_gc_end();
3684 3684 g1_policy()->phase_times()->print(pause_time_sec);
3685 3685 g1_policy()->print_detailed_heap_transition();
3686 3686 } else {
3687 3687 if (evacuation_failed()) {
3688 3688 gclog_or_tty->print("--");
3689 3689 }
3690 3690 g1_policy()->print_heap_transition();
3691 3691 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3692 3692 }
3693 + gclog_or_tty->flush();
3693 3694 }
3694 3695
3695 3696 bool
3696 3697 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3697 3698 assert_at_safepoint(true /* should_be_vm_thread */);
3698 3699 guarantee(!is_gc_active(), "collection is not reentrant");
3699 3700
3700 3701 if (GC_locker::check_active_before_gc()) {
3701 3702 return false;
3702 3703 }
3703 3704
3704 3705 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3705 3706 ResourceMark rm;
3706 3707
3707 3708 print_heap_before_gc();
3708 3709
3709 3710 HRSPhaseSetter x(HRSPhaseEvacuation);
3710 3711 verify_region_sets_optional();
3711 3712 verify_dirty_young_regions();
3712 3713
3713 3714 // This call will decide whether this pause is an initial-mark
3714 3715 // pause. If it is, during_initial_mark_pause() will return true
3715 3716 // for the duration of this pause.
3716 3717 g1_policy()->decide_on_conc_mark_initiation();
3717 3718
3718 3719 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3719 3720 assert(!g1_policy()->during_initial_mark_pause() ||
3720 3721 g1_policy()->gcs_are_young(), "sanity");
3721 3722
3722 3723 // We also do not allow mixed GCs during marking.
3723 3724 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3724 3725
3725 3726 // Record whether this pause is an initial mark. When the current
3726 3727 // thread has completed its logging output and it's safe to signal
3727 3728 // the CM thread, the flag's value in the policy has been reset.
3728 3729 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3729 3730
3730 3731 // Inner scope for scope based logging, timers, and stats collection
3731 3732 {
3732 3733 if (g1_policy()->during_initial_mark_pause()) {
3733 3734 // We are about to start a marking cycle, so we increment the
3734 3735 // full collection counter.
3735 3736 increment_old_marking_cycles_started();
3736 3737 }
3737 3738 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3738 3739
3739 3740 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3740 3741 workers()->active_workers() : 1);
3741 3742 double pause_start_sec = os::elapsedTime();
3742 3743 g1_policy()->phase_times()->note_gc_start(active_workers);
3743 3744 log_gc_header();
3744 3745
3745 3746 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3746 3747 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3747 3748
3748 3749 // If the secondary_free_list is not empty, append it to the
3749 3750 // free_list. No need to wait for the cleanup operation to finish;
3750 3751 // the region allocation code will check the secondary_free_list
3751 3752 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3752 3753 // set, skip this step so that the region allocation code has to
3753 3754 // get entries from the secondary_free_list.
3754 3755 if (!G1StressConcRegionFreeing) {
3755 3756 append_secondary_free_list_if_not_empty_with_lock();
3756 3757 }
3757 3758
3758 3759 assert(check_young_list_well_formed(),
3759 3760 "young list should be well formed");
3760 3761
3761 3762 // Don't dynamically change the number of GC threads this early. A value of
3762 3763 // 0 is used to indicate serial work. When parallel work is done,
3763 3764 // it will be set.
3764 3765
3765 3766 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3766 3767 IsGCActiveMark x;
3767 3768
3768 3769 gc_prologue(false);
3769 3770 increment_total_collections(false /* full gc */);
3770 3771 increment_gc_time_stamp();
3771 3772
3772 3773 verify_before_gc();
3773 3774
3774 3775 COMPILER2_PRESENT(DerivedPointerTable::clear());
3775 3776
3776 3777 // Please see comment in g1CollectedHeap.hpp and
3777 3778 // G1CollectedHeap::ref_processing_init() to see how
3778 3779 // reference processing currently works in G1.
3779 3780
3780 3781 // Enable discovery in the STW reference processor
3781 3782 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3782 3783 true /*verify_no_refs*/);
3783 3784
3784 3785 {
3785 3786 // We want to temporarily turn off discovery by the
3786 3787 // CM ref processor, if necessary, and turn it back on
3787 3788 // on again later if we do. Using a scoped
3788 3789 // NoRefDiscovery object will do this.
3789 3790 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3790 3791
3791 3792 // Forget the current alloc region (we might even choose it to be part
3792 3793 // of the collection set!).
3793 3794 release_mutator_alloc_region();
3794 3795
3795 3796 // We should call this after we retire the mutator alloc
3796 3797 // region(s) so that all the ALLOC / RETIRE events are generated
3797 3798 // before the start GC event.
3798 3799 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3799 3800
3800 3801 // This timing is only used by the ergonomics to handle our pause target.
3801 3802 // It is unclear why this should not include the full pause. We will
3802 3803 // investigate this in CR 7178365.
3803 3804 //
3804 3805 // Preserving the old comment here if that helps the investigation:
3805 3806 //
3806 3807 // The elapsed time induced by the start time below deliberately elides
3807 3808 // the possible verification above.
3808 3809 double sample_start_time_sec = os::elapsedTime();
3809 3810 size_t start_used_bytes = used();
3810 3811
3811 3812 #if YOUNG_LIST_VERBOSE
3812 3813 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3813 3814 _young_list->print();
3814 3815 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3815 3816 #endif // YOUNG_LIST_VERBOSE
3816 3817
3817 3818 g1_policy()->record_collection_pause_start(sample_start_time_sec,
3818 3819 start_used_bytes);
3819 3820
3820 3821 double scan_wait_start = os::elapsedTime();
3821 3822 // We have to wait until the CM threads finish scanning the
3822 3823 // root regions as it's the only way to ensure that all the
3823 3824 // objects on them have been correctly scanned before we start
3824 3825 // moving them during the GC.
3825 3826 bool waited = _cm->root_regions()->wait_until_scan_finished();
3826 3827 double wait_time_ms = 0.0;
3827 3828 if (waited) {
3828 3829 double scan_wait_end = os::elapsedTime();
3829 3830 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3830 3831 }
3831 3832 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3832 3833
3833 3834 #if YOUNG_LIST_VERBOSE
3834 3835 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3835 3836 _young_list->print();
3836 3837 #endif // YOUNG_LIST_VERBOSE
3837 3838
3838 3839 if (g1_policy()->during_initial_mark_pause()) {
3839 3840 concurrent_mark()->checkpointRootsInitialPre();
3840 3841 }
3841 3842
3842 3843 #if YOUNG_LIST_VERBOSE
3843 3844 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3844 3845 _young_list->print();
3845 3846 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3846 3847 #endif // YOUNG_LIST_VERBOSE
3847 3848
3848 3849 g1_policy()->finalize_cset(target_pause_time_ms);
3849 3850
3850 3851 _cm->note_start_of_gc();
3851 3852 // We should not verify the per-thread SATB buffers given that
3852 3853 // we have not filtered them yet (we'll do so during the
3853 3854 // GC). We also call this after finalize_cset() to
3854 3855 // ensure that the CSet has been finalized.
3855 3856 _cm->verify_no_cset_oops(true /* verify_stacks */,
3856 3857 true /* verify_enqueued_buffers */,
3857 3858 false /* verify_thread_buffers */,
3858 3859 true /* verify_fingers */);
3859 3860
3860 3861 if (_hr_printer.is_active()) {
3861 3862 HeapRegion* hr = g1_policy()->collection_set();
3862 3863 while (hr != NULL) {
3863 3864 G1HRPrinter::RegionType type;
3864 3865 if (!hr->is_young()) {
3865 3866 type = G1HRPrinter::Old;
3866 3867 } else if (hr->is_survivor()) {
3867 3868 type = G1HRPrinter::Survivor;
3868 3869 } else {
3869 3870 type = G1HRPrinter::Eden;
3870 3871 }
3871 3872 _hr_printer.cset(hr);
3872 3873 hr = hr->next_in_collection_set();
3873 3874 }
3874 3875 }
3875 3876
3876 3877 #ifdef ASSERT
3877 3878 VerifyCSetClosure cl;
3878 3879 collection_set_iterate(&cl);
3879 3880 #endif // ASSERT
3880 3881
3881 3882 setup_surviving_young_words();
3882 3883
3883 3884 // Initialize the GC alloc regions.
3884 3885 init_gc_alloc_regions();
3885 3886
3886 3887 // Actually do the work...
3887 3888 evacuate_collection_set();
3888 3889
3889 3890 // We do this to mainly verify the per-thread SATB buffers
3890 3891 // (which have been filtered by now) since we didn't verify
3891 3892 // them earlier. No point in re-checking the stacks / enqueued
3892 3893 // buffers given that the CSet has not changed since last time
3893 3894 // we checked.
3894 3895 _cm->verify_no_cset_oops(false /* verify_stacks */,
3895 3896 false /* verify_enqueued_buffers */,
3896 3897 true /* verify_thread_buffers */,
3897 3898 true /* verify_fingers */);
3898 3899
3899 3900 free_collection_set(g1_policy()->collection_set());
3900 3901 g1_policy()->clear_collection_set();
3901 3902
3902 3903 cleanup_surviving_young_words();
3903 3904
3904 3905 // Start a new incremental collection set for the next pause.
3905 3906 g1_policy()->start_incremental_cset_building();
3906 3907
3907 3908 // Clear the _cset_fast_test bitmap in anticipation of adding
3908 3909 // regions to the incremental collection set for the next
3909 3910 // evacuation pause.
3910 3911 clear_cset_fast_test();
3911 3912
3912 3913 _young_list->reset_sampled_info();
3913 3914
3914 3915 // Don't check the whole heap at this point as the
3915 3916 // GC alloc regions from this pause have been tagged
3916 3917 // as survivors and moved on to the survivor list.
3917 3918 // Survivor regions will fail the !is_young() check.
3918 3919 assert(check_young_list_empty(false /* check_heap */),
3919 3920 "young list should be empty");
3920 3921
3921 3922 #if YOUNG_LIST_VERBOSE
3922 3923 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3923 3924 _young_list->print();
3924 3925 #endif // YOUNG_LIST_VERBOSE
3925 3926
3926 3927 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3927 3928 _young_list->first_survivor_region(),
3928 3929 _young_list->last_survivor_region());
3929 3930
3930 3931 _young_list->reset_auxilary_lists();
3931 3932
3932 3933 if (evacuation_failed()) {
3933 3934 _summary_bytes_used = recalculate_used();
3934 3935 } else {
3935 3936 // The "used" of the the collection set have already been subtracted
3936 3937 // when they were freed. Add in the bytes evacuated.
3937 3938 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3938 3939 }
3939 3940
3940 3941 if (g1_policy()->during_initial_mark_pause()) {
3941 3942 // We have to do this before we notify the CM threads that
3942 3943 // they can start working to make sure that all the
3943 3944 // appropriate initialization is done on the CM object.
3944 3945 concurrent_mark()->checkpointRootsInitialPost();
3945 3946 set_marking_started();
3946 3947 // Note that we don't actually trigger the CM thread at
3947 3948 // this point. We do that later when we're sure that
3948 3949 // the current thread has completed its logging output.
3949 3950 }
3950 3951
3951 3952 allocate_dummy_regions();
3952 3953
3953 3954 #if YOUNG_LIST_VERBOSE
3954 3955 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3955 3956 _young_list->print();
3956 3957 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3957 3958 #endif // YOUNG_LIST_VERBOSE
3958 3959
3959 3960 init_mutator_alloc_region();
3960 3961
3961 3962 {
3962 3963 size_t expand_bytes = g1_policy()->expansion_amount();
3963 3964 if (expand_bytes > 0) {
3964 3965 size_t bytes_before = capacity();
3965 3966 // No need for an ergo verbose message here,
3966 3967 // expansion_amount() does this when it returns a value > 0.
3967 3968 if (!expand(expand_bytes)) {
3968 3969 // We failed to expand the heap so let's verify that
3969 3970 // committed/uncommitted amount match the backing store
3970 3971 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3971 3972 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3972 3973 }
3973 3974 }
3974 3975 }
3975 3976
3976 3977 // We redo the verificaiton but now wrt to the new CSet which
3977 3978 // has just got initialized after the previous CSet was freed.
3978 3979 _cm->verify_no_cset_oops(true /* verify_stacks */,
3979 3980 true /* verify_enqueued_buffers */,
3980 3981 true /* verify_thread_buffers */,
3981 3982 true /* verify_fingers */);
3982 3983 _cm->note_end_of_gc();
3983 3984
3984 3985 // This timing is only used by the ergonomics to handle our pause target.
3985 3986 // It is unclear why this should not include the full pause. We will
3986 3987 // investigate this in CR 7178365.
3987 3988 double sample_end_time_sec = os::elapsedTime();
3988 3989 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
3989 3990 g1_policy()->record_collection_pause_end(pause_time_ms);
3990 3991
3991 3992 MemoryService::track_memory_usage();
3992 3993
3993 3994 // In prepare_for_verify() below we'll need to scan the deferred
3994 3995 // update buffers to bring the RSets up-to-date if
3995 3996 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3996 3997 // the update buffers we'll probably need to scan cards on the
3997 3998 // regions we just allocated to (i.e., the GC alloc
3998 3999 // regions). However, during the last GC we called
3999 4000 // set_saved_mark() on all the GC alloc regions, so card
4000 4001 // scanning might skip the [saved_mark_word()...top()] area of
4001 4002 // those regions (i.e., the area we allocated objects into
4002 4003 // during the last GC). But it shouldn't. Given that
4003 4004 // saved_mark_word() is conditional on whether the GC time stamp
4004 4005 // on the region is current or not, by incrementing the GC time
4005 4006 // stamp here we invalidate all the GC time stamps on all the
4006 4007 // regions and saved_mark_word() will simply return top() for
4007 4008 // all the regions. This is a nicer way of ensuring this rather
4008 4009 // than iterating over the regions and fixing them. In fact, the
4009 4010 // GC time stamp increment here also ensures that
4010 4011 // saved_mark_word() will return top() between pauses, i.e.,
4011 4012 // during concurrent refinement. So we don't need the
4012 4013 // is_gc_active() check to decided which top to use when
4013 4014 // scanning cards (see CR 7039627).
4014 4015 increment_gc_time_stamp();
4015 4016
4016 4017 verify_after_gc();
4017 4018
4018 4019 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4019 4020 ref_processor_stw()->verify_no_references_recorded();
4020 4021
4021 4022 // CM reference discovery will be re-enabled if necessary.
4022 4023 }
4023 4024
4024 4025 // We should do this after we potentially expand the heap so
4025 4026 // that all the COMMIT events are generated before the end GC
4026 4027 // event, and after we retire the GC alloc regions so that all
4027 4028 // RETIRE events are generated before the end GC event.
4028 4029 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
|
↓ open down ↓ |
326 lines elided |
↑ open up ↑ |
4029 4030
4030 4031 if (mark_in_progress()) {
4031 4032 concurrent_mark()->update_g1_committed();
4032 4033 }
4033 4034
4034 4035 #ifdef TRACESPINNING
4035 4036 ParallelTaskTerminator::print_termination_counts();
4036 4037 #endif
4037 4038
4038 4039 gc_epilogue(false);
4039 -
4040 - log_gc_footer(os::elapsedTime() - pause_start_sec);
4041 4040 }
4042 4041
4042 + // Print the remainder of the GC log output.
4043 + log_gc_footer(os::elapsedTime() - pause_start_sec);
4044 +
4043 4045 // It is not yet to safe to tell the concurrent mark to
4044 4046 // start as we have some optional output below. We don't want the
4045 4047 // output from the concurrent mark thread interfering with this
4046 4048 // logging output either.
4047 4049
4048 4050 _hrs.verify_optional();
4049 4051 verify_region_sets_optional();
4050 4052
4051 4053 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4052 4054 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4053 4055
4054 4056 print_heap_after_gc();
4055 4057
4056 4058 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4057 4059 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4058 4060 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4059 4061 // before any GC notifications are raised.
4060 4062 g1mm()->update_sizes();
4061 4063 }
4062 4064
4063 4065 if (G1SummarizeRSetStats &&
4064 4066 (G1SummarizeRSetStatsPeriod > 0) &&
4065 4067 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4066 4068 g1_rem_set()->print_summary_info();
4067 4069 }
4068 4070
4069 4071 // It should now be safe to tell the concurrent mark thread to start
4070 4072 // without its logging output interfering with the logging output
4071 4073 // that came from the pause.
4072 4074
4073 4075 if (should_start_conc_mark) {
4074 4076 // CAUTION: after the doConcurrentMark() call below,
4075 4077 // the concurrent marking thread(s) could be running
4076 4078 // concurrently with us. Make sure that anything after
4077 4079 // this point does not assume that we are the only GC thread
4078 4080 // running. Note: of course, the actual marking work will
4079 4081 // not start until the safepoint itself is released in
4080 4082 // ConcurrentGCThread::safepoint_desynchronize().
4081 4083 doConcurrentMark();
4082 4084 }
4083 4085
4084 4086 return true;
4085 4087 }
4086 4088
4087 4089 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4088 4090 {
4089 4091 size_t gclab_word_size;
4090 4092 switch (purpose) {
4091 4093 case GCAllocForSurvived:
4092 4094 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4093 4095 break;
4094 4096 case GCAllocForTenured:
4095 4097 gclab_word_size = _old_plab_stats.desired_plab_sz();
4096 4098 break;
4097 4099 default:
4098 4100 assert(false, "unknown GCAllocPurpose");
4099 4101 gclab_word_size = _old_plab_stats.desired_plab_sz();
4100 4102 break;
4101 4103 }
4102 4104
4103 4105 // Prevent humongous PLAB sizes for two reasons:
4104 4106 // * PLABs are allocated using a similar paths as oops, but should
4105 4107 // never be in a humongous region
4106 4108 // * Allowing humongous PLABs needlessly churns the region free lists
4107 4109 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4108 4110 }
4109 4111
4110 4112 void G1CollectedHeap::init_mutator_alloc_region() {
4111 4113 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4112 4114 _mutator_alloc_region.init();
4113 4115 }
4114 4116
4115 4117 void G1CollectedHeap::release_mutator_alloc_region() {
4116 4118 _mutator_alloc_region.release();
4117 4119 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4118 4120 }
4119 4121
4120 4122 void G1CollectedHeap::init_gc_alloc_regions() {
4121 4123 assert_at_safepoint(true /* should_be_vm_thread */);
4122 4124
4123 4125 _survivor_gc_alloc_region.init();
4124 4126 _old_gc_alloc_region.init();
4125 4127 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4126 4128 _retained_old_gc_alloc_region = NULL;
4127 4129
4128 4130 // We will discard the current GC alloc region if:
4129 4131 // a) it's in the collection set (it can happen!),
4130 4132 // b) it's already full (no point in using it),
4131 4133 // c) it's empty (this means that it was emptied during
4132 4134 // a cleanup and it should be on the free list now), or
4133 4135 // d) it's humongous (this means that it was emptied
4134 4136 // during a cleanup and was added to the free list, but
4135 4137 // has been subseqently used to allocate a humongous
4136 4138 // object that may be less than the region size).
4137 4139 if (retained_region != NULL &&
4138 4140 !retained_region->in_collection_set() &&
4139 4141 !(retained_region->top() == retained_region->end()) &&
4140 4142 !retained_region->is_empty() &&
4141 4143 !retained_region->isHumongous()) {
4142 4144 retained_region->set_saved_mark();
4143 4145 // The retained region was added to the old region set when it was
4144 4146 // retired. We have to remove it now, since we don't allow regions
4145 4147 // we allocate to in the region sets. We'll re-add it later, when
4146 4148 // it's retired again.
4147 4149 _old_set.remove(retained_region);
4148 4150 bool during_im = g1_policy()->during_initial_mark_pause();
4149 4151 retained_region->note_start_of_copying(during_im);
4150 4152 _old_gc_alloc_region.set(retained_region);
4151 4153 _hr_printer.reuse(retained_region);
4152 4154 }
4153 4155 }
4154 4156
4155 4157 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4156 4158 _survivor_gc_alloc_region.release();
4157 4159 // If we have an old GC alloc region to release, we'll save it in
4158 4160 // _retained_old_gc_alloc_region. If we don't
4159 4161 // _retained_old_gc_alloc_region will become NULL. This is what we
4160 4162 // want either way so no reason to check explicitly for either
4161 4163 // condition.
4162 4164 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4163 4165
4164 4166 if (ResizePLAB) {
4165 4167 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4166 4168 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4167 4169 }
4168 4170 }
4169 4171
4170 4172 void G1CollectedHeap::abandon_gc_alloc_regions() {
4171 4173 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4172 4174 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4173 4175 _retained_old_gc_alloc_region = NULL;
4174 4176 }
4175 4177
4176 4178 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4177 4179 _drain_in_progress = false;
4178 4180 set_evac_failure_closure(cl);
4179 4181 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4180 4182 }
4181 4183
4182 4184 void G1CollectedHeap::finalize_for_evac_failure() {
4183 4185 assert(_evac_failure_scan_stack != NULL &&
4184 4186 _evac_failure_scan_stack->length() == 0,
4185 4187 "Postcondition");
4186 4188 assert(!_drain_in_progress, "Postcondition");
4187 4189 delete _evac_failure_scan_stack;
4188 4190 _evac_failure_scan_stack = NULL;
4189 4191 }
4190 4192
4191 4193 void G1CollectedHeap::remove_self_forwarding_pointers() {
4192 4194 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4193 4195
4194 4196 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4195 4197
4196 4198 if (G1CollectedHeap::use_parallel_gc_threads()) {
4197 4199 set_par_threads();
4198 4200 workers()->run_task(&rsfp_task);
4199 4201 set_par_threads(0);
4200 4202 } else {
4201 4203 rsfp_task.work(0);
4202 4204 }
4203 4205
4204 4206 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4205 4207
4206 4208 // Reset the claim values in the regions in the collection set.
4207 4209 reset_cset_heap_region_claim_values();
4208 4210
4209 4211 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4210 4212
4211 4213 // Now restore saved marks, if any.
4212 4214 if (_objs_with_preserved_marks != NULL) {
4213 4215 assert(_preserved_marks_of_objs != NULL, "Both or none.");
4214 4216 guarantee(_objs_with_preserved_marks->length() ==
4215 4217 _preserved_marks_of_objs->length(), "Both or none.");
4216 4218 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
4217 4219 oop obj = _objs_with_preserved_marks->at(i);
4218 4220 markOop m = _preserved_marks_of_objs->at(i);
4219 4221 obj->set_mark(m);
4220 4222 }
4221 4223
4222 4224 // Delete the preserved marks growable arrays (allocated on the C heap).
4223 4225 delete _objs_with_preserved_marks;
4224 4226 delete _preserved_marks_of_objs;
4225 4227 _objs_with_preserved_marks = NULL;
4226 4228 _preserved_marks_of_objs = NULL;
4227 4229 }
4228 4230 }
4229 4231
4230 4232 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4231 4233 _evac_failure_scan_stack->push(obj);
4232 4234 }
4233 4235
4234 4236 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4235 4237 assert(_evac_failure_scan_stack != NULL, "precondition");
4236 4238
4237 4239 while (_evac_failure_scan_stack->length() > 0) {
4238 4240 oop obj = _evac_failure_scan_stack->pop();
4239 4241 _evac_failure_closure->set_region(heap_region_containing(obj));
4240 4242 obj->oop_iterate_backwards(_evac_failure_closure);
4241 4243 }
4242 4244 }
4243 4245
4244 4246 oop
4245 4247 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4246 4248 oop old) {
4247 4249 assert(obj_in_cs(old),
4248 4250 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4249 4251 (HeapWord*) old));
4250 4252 markOop m = old->mark();
4251 4253 oop forward_ptr = old->forward_to_atomic(old);
4252 4254 if (forward_ptr == NULL) {
4253 4255 // Forward-to-self succeeded.
4254 4256
4255 4257 if (_evac_failure_closure != cl) {
4256 4258 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4257 4259 assert(!_drain_in_progress,
4258 4260 "Should only be true while someone holds the lock.");
4259 4261 // Set the global evac-failure closure to the current thread's.
4260 4262 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4261 4263 set_evac_failure_closure(cl);
4262 4264 // Now do the common part.
4263 4265 handle_evacuation_failure_common(old, m);
4264 4266 // Reset to NULL.
4265 4267 set_evac_failure_closure(NULL);
4266 4268 } else {
4267 4269 // The lock is already held, and this is recursive.
4268 4270 assert(_drain_in_progress, "This should only be the recursive case.");
4269 4271 handle_evacuation_failure_common(old, m);
4270 4272 }
4271 4273 return old;
4272 4274 } else {
4273 4275 // Forward-to-self failed. Either someone else managed to allocate
4274 4276 // space for this object (old != forward_ptr) or they beat us in
4275 4277 // self-forwarding it (old == forward_ptr).
4276 4278 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4277 4279 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4278 4280 "should not be in the CSet",
4279 4281 (HeapWord*) old, (HeapWord*) forward_ptr));
4280 4282 return forward_ptr;
4281 4283 }
4282 4284 }
4283 4285
4284 4286 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4285 4287 set_evacuation_failed(true);
4286 4288
4287 4289 preserve_mark_if_necessary(old, m);
4288 4290
4289 4291 HeapRegion* r = heap_region_containing(old);
4290 4292 if (!r->evacuation_failed()) {
4291 4293 r->set_evacuation_failed(true);
4292 4294 _hr_printer.evac_failure(r);
4293 4295 }
4294 4296
4295 4297 push_on_evac_failure_scan_stack(old);
4296 4298
4297 4299 if (!_drain_in_progress) {
4298 4300 // prevent recursion in copy_to_survivor_space()
4299 4301 _drain_in_progress = true;
4300 4302 drain_evac_failure_scan_stack();
4301 4303 _drain_in_progress = false;
4302 4304 }
4303 4305 }
4304 4306
4305 4307 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4306 4308 assert(evacuation_failed(), "Oversaving!");
4307 4309 // We want to call the "for_promotion_failure" version only in the
4308 4310 // case of a promotion failure.
4309 4311 if (m->must_be_preserved_for_promotion_failure(obj)) {
4310 4312 if (_objs_with_preserved_marks == NULL) {
4311 4313 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4312 4314 _objs_with_preserved_marks =
4313 4315 new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4314 4316 _preserved_marks_of_objs =
4315 4317 new (ResourceObj::C_HEAP, mtGC) GrowableArray<markOop>(40, true);
4316 4318 }
4317 4319 _objs_with_preserved_marks->push(obj);
4318 4320 _preserved_marks_of_objs->push(m);
4319 4321 }
4320 4322 }
4321 4323
4322 4324 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4323 4325 size_t word_size) {
4324 4326 if (purpose == GCAllocForSurvived) {
4325 4327 HeapWord* result = survivor_attempt_allocation(word_size);
4326 4328 if (result != NULL) {
4327 4329 return result;
4328 4330 } else {
4329 4331 // Let's try to allocate in the old gen in case we can fit the
4330 4332 // object there.
4331 4333 return old_attempt_allocation(word_size);
4332 4334 }
4333 4335 } else {
4334 4336 assert(purpose == GCAllocForTenured, "sanity");
4335 4337 HeapWord* result = old_attempt_allocation(word_size);
4336 4338 if (result != NULL) {
4337 4339 return result;
4338 4340 } else {
4339 4341 // Let's try to allocate in the survivors in case we can fit the
4340 4342 // object there.
4341 4343 return survivor_attempt_allocation(word_size);
4342 4344 }
4343 4345 }
4344 4346
4345 4347 ShouldNotReachHere();
4346 4348 // Trying to keep some compilers happy.
4347 4349 return NULL;
4348 4350 }
4349 4351
4350 4352 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4351 4353 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4352 4354
4353 4355 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4354 4356 : _g1h(g1h),
4355 4357 _refs(g1h->task_queue(queue_num)),
4356 4358 _dcq(&g1h->dirty_card_queue_set()),
4357 4359 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4358 4360 _g1_rem(g1h->g1_rem_set()),
4359 4361 _hash_seed(17), _queue_num(queue_num),
4360 4362 _term_attempts(0),
4361 4363 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4362 4364 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4363 4365 _age_table(false),
4364 4366 _strong_roots_time(0), _term_time(0),
4365 4367 _alloc_buffer_waste(0), _undo_waste(0) {
4366 4368 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4367 4369 // we "sacrifice" entry 0 to keep track of surviving bytes for
4368 4370 // non-young regions (where the age is -1)
4369 4371 // We also add a few elements at the beginning and at the end in
4370 4372 // an attempt to eliminate cache contention
4371 4373 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4372 4374 uint array_length = PADDING_ELEM_NUM +
4373 4375 real_length +
4374 4376 PADDING_ELEM_NUM;
4375 4377 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4376 4378 if (_surviving_young_words_base == NULL)
4377 4379 vm_exit_out_of_memory(array_length * sizeof(size_t),
4378 4380 "Not enough space for young surv histo.");
4379 4381 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4380 4382 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4381 4383
4382 4384 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4383 4385 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4384 4386
4385 4387 _start = os::elapsedTime();
4386 4388 }
4387 4389
4388 4390 void
4389 4391 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4390 4392 {
4391 4393 st->print_raw_cr("GC Termination Stats");
4392 4394 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4393 4395 " ------waste (KiB)------");
4394 4396 st->print_raw_cr("thr ms ms % ms % attempts"
4395 4397 " total alloc undo");
4396 4398 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4397 4399 " ------- ------- -------");
4398 4400 }
4399 4401
4400 4402 void
4401 4403 G1ParScanThreadState::print_termination_stats(int i,
4402 4404 outputStream* const st) const
4403 4405 {
4404 4406 const double elapsed_ms = elapsed_time() * 1000.0;
4405 4407 const double s_roots_ms = strong_roots_time() * 1000.0;
4406 4408 const double term_ms = term_time() * 1000.0;
4407 4409 st->print_cr("%3d %9.2f %9.2f %6.2f "
4408 4410 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4409 4411 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4410 4412 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4411 4413 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4412 4414 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4413 4415 alloc_buffer_waste() * HeapWordSize / K,
4414 4416 undo_waste() * HeapWordSize / K);
4415 4417 }
4416 4418
4417 4419 #ifdef ASSERT
4418 4420 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4419 4421 assert(ref != NULL, "invariant");
4420 4422 assert(UseCompressedOops, "sanity");
4421 4423 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4422 4424 oop p = oopDesc::load_decode_heap_oop(ref);
4423 4425 assert(_g1h->is_in_g1_reserved(p),
4424 4426 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4425 4427 return true;
4426 4428 }
4427 4429
4428 4430 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4429 4431 assert(ref != NULL, "invariant");
4430 4432 if (has_partial_array_mask(ref)) {
4431 4433 // Must be in the collection set--it's already been copied.
4432 4434 oop p = clear_partial_array_mask(ref);
4433 4435 assert(_g1h->obj_in_cs(p),
4434 4436 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4435 4437 } else {
4436 4438 oop p = oopDesc::load_decode_heap_oop(ref);
4437 4439 assert(_g1h->is_in_g1_reserved(p),
4438 4440 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4439 4441 }
4440 4442 return true;
4441 4443 }
4442 4444
4443 4445 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4444 4446 if (ref.is_narrow()) {
4445 4447 return verify_ref((narrowOop*) ref);
4446 4448 } else {
4447 4449 return verify_ref((oop*) ref);
4448 4450 }
4449 4451 }
4450 4452 #endif // ASSERT
4451 4453
4452 4454 void G1ParScanThreadState::trim_queue() {
4453 4455 assert(_evac_cl != NULL, "not set");
4454 4456 assert(_evac_failure_cl != NULL, "not set");
4455 4457 assert(_partial_scan_cl != NULL, "not set");
4456 4458
4457 4459 StarTask ref;
4458 4460 do {
4459 4461 // Drain the overflow stack first, so other threads can steal.
4460 4462 while (refs()->pop_overflow(ref)) {
4461 4463 deal_with_reference(ref);
4462 4464 }
4463 4465
4464 4466 while (refs()->pop_local(ref)) {
4465 4467 deal_with_reference(ref);
4466 4468 }
4467 4469 } while (!refs()->is_empty());
4468 4470 }
4469 4471
4470 4472 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4471 4473 G1ParScanThreadState* par_scan_state) :
4472 4474 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4473 4475 _par_scan_state(par_scan_state),
4474 4476 _worker_id(par_scan_state->queue_num()),
4475 4477 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4476 4478 _mark_in_progress(_g1->mark_in_progress()) { }
4477 4479
4478 4480 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4479 4481 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4480 4482 #ifdef ASSERT
4481 4483 HeapRegion* hr = _g1->heap_region_containing(obj);
4482 4484 assert(hr != NULL, "sanity");
4483 4485 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4484 4486 #endif // ASSERT
4485 4487
4486 4488 // We know that the object is not moving so it's safe to read its size.
4487 4489 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4488 4490 }
4489 4491
4490 4492 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4491 4493 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4492 4494 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4493 4495 #ifdef ASSERT
4494 4496 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4495 4497 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4496 4498 assert(from_obj != to_obj, "should not be self-forwarded");
4497 4499
4498 4500 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4499 4501 assert(from_hr != NULL, "sanity");
4500 4502 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4501 4503
4502 4504 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4503 4505 assert(to_hr != NULL, "sanity");
4504 4506 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4505 4507 #endif // ASSERT
4506 4508
4507 4509 // The object might be in the process of being copied by another
4508 4510 // worker so we cannot trust that its to-space image is
4509 4511 // well-formed. So we have to read its size from its from-space
4510 4512 // image which we know should not be changing.
4511 4513 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4512 4514 }
4513 4515
4514 4516 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4515 4517 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4516 4518 ::copy_to_survivor_space(oop old) {
4517 4519 size_t word_sz = old->size();
4518 4520 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4519 4521 // +1 to make the -1 indexes valid...
4520 4522 int young_index = from_region->young_index_in_cset()+1;
4521 4523 assert( (from_region->is_young() && young_index > 0) ||
4522 4524 (!from_region->is_young() && young_index == 0), "invariant" );
4523 4525 G1CollectorPolicy* g1p = _g1->g1_policy();
4524 4526 markOop m = old->mark();
4525 4527 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4526 4528 : m->age();
4527 4529 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4528 4530 word_sz);
4529 4531 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4530 4532 #ifndef PRODUCT
4531 4533 // Should this evacuation fail?
4532 4534 if (_g1->evacuation_should_fail()) {
4533 4535 if (obj_ptr != NULL) {
4534 4536 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4535 4537 obj_ptr = NULL;
4536 4538 }
4537 4539 }
4538 4540 #endif // !PRODUCT
4539 4541
4540 4542 if (obj_ptr == NULL) {
4541 4543 // This will either forward-to-self, or detect that someone else has
4542 4544 // installed a forwarding pointer.
4543 4545 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4544 4546 return _g1->handle_evacuation_failure_par(cl, old);
4545 4547 }
4546 4548
4547 4549 oop obj = oop(obj_ptr);
4548 4550
4549 4551 // We're going to allocate linearly, so might as well prefetch ahead.
4550 4552 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4551 4553
4552 4554 oop forward_ptr = old->forward_to_atomic(obj);
4553 4555 if (forward_ptr == NULL) {
4554 4556 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4555 4557 if (g1p->track_object_age(alloc_purpose)) {
4556 4558 // We could simply do obj->incr_age(). However, this causes a
4557 4559 // performance issue. obj->incr_age() will first check whether
4558 4560 // the object has a displaced mark by checking its mark word;
4559 4561 // getting the mark word from the new location of the object
4560 4562 // stalls. So, given that we already have the mark word and we
4561 4563 // are about to install it anyway, it's better to increase the
4562 4564 // age on the mark word, when the object does not have a
4563 4565 // displaced mark word. We're not expecting many objects to have
4564 4566 // a displaced marked word, so that case is not optimized
4565 4567 // further (it could be...) and we simply call obj->incr_age().
4566 4568
4567 4569 if (m->has_displaced_mark_helper()) {
4568 4570 // in this case, we have to install the mark word first,
4569 4571 // otherwise obj looks to be forwarded (the old mark word,
4570 4572 // which contains the forward pointer, was copied)
4571 4573 obj->set_mark(m);
4572 4574 obj->incr_age();
4573 4575 } else {
4574 4576 m = m->incr_age();
4575 4577 obj->set_mark(m);
4576 4578 }
4577 4579 _par_scan_state->age_table()->add(obj, word_sz);
4578 4580 } else {
4579 4581 obj->set_mark(m);
4580 4582 }
4581 4583
4582 4584 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4583 4585 surv_young_words[young_index] += word_sz;
4584 4586
4585 4587 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4586 4588 // We keep track of the next start index in the length field of
4587 4589 // the to-space object. The actual length can be found in the
4588 4590 // length field of the from-space object.
4589 4591 arrayOop(obj)->set_length(0);
4590 4592 oop* old_p = set_partial_array_mask(old);
4591 4593 _par_scan_state->push_on_queue(old_p);
4592 4594 } else {
4593 4595 // No point in using the slower heap_region_containing() method,
4594 4596 // given that we know obj is in the heap.
4595 4597 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4596 4598 obj->oop_iterate_backwards(&_scanner);
4597 4599 }
4598 4600 } else {
4599 4601 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4600 4602 obj = forward_ptr;
4601 4603 }
4602 4604 return obj;
4603 4605 }
4604 4606
4605 4607 template <class T>
4606 4608 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4607 4609 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4608 4610 _scanned_klass->record_modified_oops();
4609 4611 }
4610 4612 }
4611 4613
4612 4614 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4613 4615 template <class T>
4614 4616 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4615 4617 ::do_oop_work(T* p) {
4616 4618 oop obj = oopDesc::load_decode_heap_oop(p);
4617 4619 assert(barrier != G1BarrierRS || obj != NULL,
4618 4620 "Precondition: G1BarrierRS implies obj is non-NULL");
4619 4621
4620 4622 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4621 4623
4622 4624 // here the null check is implicit in the cset_fast_test() test
4623 4625 if (_g1->in_cset_fast_test(obj)) {
4624 4626 oop forwardee;
4625 4627 if (obj->is_forwarded()) {
4626 4628 forwardee = obj->forwardee();
4627 4629 } else {
4628 4630 forwardee = copy_to_survivor_space(obj);
4629 4631 }
4630 4632 assert(forwardee != NULL, "forwardee should not be NULL");
4631 4633 oopDesc::encode_store_heap_oop(p, forwardee);
4632 4634 if (do_mark_object && forwardee != obj) {
4633 4635 // If the object is self-forwarded we don't need to explicitly
4634 4636 // mark it, the evacuation failure protocol will do so.
4635 4637 mark_forwarded_object(obj, forwardee);
4636 4638 }
4637 4639
4638 4640 // When scanning the RS, we only care about objs in CS.
4639 4641 if (barrier == G1BarrierRS) {
4640 4642 _par_scan_state->update_rs(_from, p, _worker_id);
4641 4643 } else if (barrier == G1BarrierKlass) {
4642 4644 do_klass_barrier(p, forwardee);
4643 4645 }
4644 4646 } else {
4645 4647 // The object is not in collection set. If we're a root scanning
4646 4648 // closure during an initial mark pause (i.e. do_mark_object will
4647 4649 // be true) then attempt to mark the object.
4648 4650 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4649 4651 mark_object(obj);
4650 4652 }
4651 4653 }
4652 4654
4653 4655 if (barrier == G1BarrierEvac && obj != NULL) {
4654 4656 _par_scan_state->update_rs(_from, p, _worker_id);
4655 4657 }
4656 4658
4657 4659 if (do_gen_barrier && obj != NULL) {
4658 4660 par_do_barrier(p);
4659 4661 }
4660 4662 }
4661 4663
4662 4664 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4663 4665 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4664 4666
4665 4667 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4666 4668 assert(has_partial_array_mask(p), "invariant");
4667 4669 oop from_obj = clear_partial_array_mask(p);
4668 4670
4669 4671 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4670 4672 assert(from_obj->is_objArray(), "must be obj array");
4671 4673 objArrayOop from_obj_array = objArrayOop(from_obj);
4672 4674 // The from-space object contains the real length.
4673 4675 int length = from_obj_array->length();
4674 4676
4675 4677 assert(from_obj->is_forwarded(), "must be forwarded");
4676 4678 oop to_obj = from_obj->forwardee();
4677 4679 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4678 4680 objArrayOop to_obj_array = objArrayOop(to_obj);
4679 4681 // We keep track of the next start index in the length field of the
4680 4682 // to-space object.
4681 4683 int next_index = to_obj_array->length();
4682 4684 assert(0 <= next_index && next_index < length,
4683 4685 err_msg("invariant, next index: %d, length: %d", next_index, length));
4684 4686
4685 4687 int start = next_index;
4686 4688 int end = length;
4687 4689 int remainder = end - start;
4688 4690 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4689 4691 if (remainder > 2 * ParGCArrayScanChunk) {
4690 4692 end = start + ParGCArrayScanChunk;
4691 4693 to_obj_array->set_length(end);
4692 4694 // Push the remainder before we process the range in case another
4693 4695 // worker has run out of things to do and can steal it.
4694 4696 oop* from_obj_p = set_partial_array_mask(from_obj);
4695 4697 _par_scan_state->push_on_queue(from_obj_p);
4696 4698 } else {
4697 4699 assert(length == end, "sanity");
4698 4700 // We'll process the final range for this object. Restore the length
4699 4701 // so that the heap remains parsable in case of evacuation failure.
4700 4702 to_obj_array->set_length(end);
4701 4703 }
4702 4704 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4703 4705 // Process indexes [start,end). It will also process the header
4704 4706 // along with the first chunk (i.e., the chunk with start == 0).
4705 4707 // Note that at this point the length field of to_obj_array is not
4706 4708 // correct given that we are using it to keep track of the next
4707 4709 // start index. oop_iterate_range() (thankfully!) ignores the length
4708 4710 // field and only relies on the start / end parameters. It does
4709 4711 // however return the size of the object which will be incorrect. So
4710 4712 // we have to ignore it even if we wanted to use it.
4711 4713 to_obj_array->oop_iterate_range(&_scanner, start, end);
4712 4714 }
4713 4715
4714 4716 class G1ParEvacuateFollowersClosure : public VoidClosure {
4715 4717 protected:
4716 4718 G1CollectedHeap* _g1h;
4717 4719 G1ParScanThreadState* _par_scan_state;
4718 4720 RefToScanQueueSet* _queues;
4719 4721 ParallelTaskTerminator* _terminator;
4720 4722
4721 4723 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4722 4724 RefToScanQueueSet* queues() { return _queues; }
4723 4725 ParallelTaskTerminator* terminator() { return _terminator; }
4724 4726
4725 4727 public:
4726 4728 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4727 4729 G1ParScanThreadState* par_scan_state,
4728 4730 RefToScanQueueSet* queues,
4729 4731 ParallelTaskTerminator* terminator)
4730 4732 : _g1h(g1h), _par_scan_state(par_scan_state),
4731 4733 _queues(queues), _terminator(terminator) {}
4732 4734
4733 4735 void do_void();
4734 4736
4735 4737 private:
4736 4738 inline bool offer_termination();
4737 4739 };
4738 4740
4739 4741 bool G1ParEvacuateFollowersClosure::offer_termination() {
4740 4742 G1ParScanThreadState* const pss = par_scan_state();
4741 4743 pss->start_term_time();
4742 4744 const bool res = terminator()->offer_termination();
4743 4745 pss->end_term_time();
4744 4746 return res;
4745 4747 }
4746 4748
4747 4749 void G1ParEvacuateFollowersClosure::do_void() {
4748 4750 StarTask stolen_task;
4749 4751 G1ParScanThreadState* const pss = par_scan_state();
4750 4752 pss->trim_queue();
4751 4753
4752 4754 do {
4753 4755 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4754 4756 assert(pss->verify_task(stolen_task), "sanity");
4755 4757 if (stolen_task.is_narrow()) {
4756 4758 pss->deal_with_reference((narrowOop*) stolen_task);
4757 4759 } else {
4758 4760 pss->deal_with_reference((oop*) stolen_task);
4759 4761 }
4760 4762
4761 4763 // We've just processed a reference and we might have made
4762 4764 // available new entries on the queues. So we have to make sure
4763 4765 // we drain the queues as necessary.
4764 4766 pss->trim_queue();
4765 4767 }
4766 4768 } while (!offer_termination());
4767 4769
4768 4770 pss->retire_alloc_buffers();
4769 4771 }
4770 4772
4771 4773 class G1KlassScanClosure : public KlassClosure {
4772 4774 G1ParCopyHelper* _closure;
4773 4775 bool _process_only_dirty;
4774 4776 int _count;
4775 4777 public:
4776 4778 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4777 4779 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4778 4780 void do_klass(Klass* klass) {
4779 4781 // If the klass has not been dirtied we know that there's
4780 4782 // no references into the young gen and we can skip it.
4781 4783 if (!_process_only_dirty || klass->has_modified_oops()) {
4782 4784 // Clean the klass since we're going to scavenge all the metadata.
4783 4785 klass->clear_modified_oops();
4784 4786
4785 4787 // Tell the closure that this klass is the Klass to scavenge
4786 4788 // and is the one to dirty if oops are left pointing into the young gen.
4787 4789 _closure->set_scanned_klass(klass);
4788 4790
4789 4791 klass->oops_do(_closure);
4790 4792
4791 4793 _closure->set_scanned_klass(NULL);
4792 4794 }
4793 4795 _count++;
4794 4796 }
4795 4797 };
4796 4798
4797 4799 class G1ParTask : public AbstractGangTask {
4798 4800 protected:
4799 4801 G1CollectedHeap* _g1h;
4800 4802 RefToScanQueueSet *_queues;
4801 4803 ParallelTaskTerminator _terminator;
4802 4804 uint _n_workers;
4803 4805
4804 4806 Mutex _stats_lock;
4805 4807 Mutex* stats_lock() { return &_stats_lock; }
4806 4808
4807 4809 size_t getNCards() {
4808 4810 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4809 4811 / G1BlockOffsetSharedArray::N_bytes;
4810 4812 }
4811 4813
4812 4814 public:
4813 4815 G1ParTask(G1CollectedHeap* g1h,
4814 4816 RefToScanQueueSet *task_queues)
4815 4817 : AbstractGangTask("G1 collection"),
4816 4818 _g1h(g1h),
4817 4819 _queues(task_queues),
4818 4820 _terminator(0, _queues),
4819 4821 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4820 4822 {}
4821 4823
4822 4824 RefToScanQueueSet* queues() { return _queues; }
4823 4825
4824 4826 RefToScanQueue *work_queue(int i) {
4825 4827 return queues()->queue(i);
4826 4828 }
4827 4829
4828 4830 ParallelTaskTerminator* terminator() { return &_terminator; }
4829 4831
4830 4832 virtual void set_for_termination(int active_workers) {
4831 4833 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4832 4834 // in the young space (_par_seq_tasks) in the G1 heap
4833 4835 // for SequentialSubTasksDone.
4834 4836 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4835 4837 // both of which need setting by set_n_termination().
4836 4838 _g1h->SharedHeap::set_n_termination(active_workers);
4837 4839 _g1h->set_n_termination(active_workers);
4838 4840 terminator()->reset_for_reuse(active_workers);
4839 4841 _n_workers = active_workers;
4840 4842 }
4841 4843
4842 4844 void work(uint worker_id) {
4843 4845 if (worker_id >= _n_workers) return; // no work needed this round
4844 4846
4845 4847 double start_time_ms = os::elapsedTime() * 1000.0;
4846 4848 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4847 4849
4848 4850 {
4849 4851 ResourceMark rm;
4850 4852 HandleMark hm;
4851 4853
4852 4854 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4853 4855
4854 4856 G1ParScanThreadState pss(_g1h, worker_id);
4855 4857 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4856 4858 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4857 4859 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4858 4860
4859 4861 pss.set_evac_closure(&scan_evac_cl);
4860 4862 pss.set_evac_failure_closure(&evac_failure_cl);
4861 4863 pss.set_partial_scan_closure(&partial_scan_cl);
4862 4864
4863 4865 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4864 4866 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4865 4867
4866 4868 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4867 4869 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4868 4870
4869 4871 bool only_young = _g1h->g1_policy()->gcs_are_young();
4870 4872 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4871 4873 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4872 4874
4873 4875 OopClosure* scan_root_cl = &only_scan_root_cl;
4874 4876 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4875 4877
4876 4878 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4877 4879 // We also need to mark copied objects.
4878 4880 scan_root_cl = &scan_mark_root_cl;
4879 4881 scan_klasses_cl = &scan_mark_klasses_cl_s;
4880 4882 }
4881 4883
4882 4884 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4883 4885
4884 4886 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4885 4887
4886 4888 pss.start_strong_roots();
4887 4889 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4888 4890 SharedHeap::ScanningOption(so),
4889 4891 scan_root_cl,
4890 4892 &push_heap_rs_cl,
4891 4893 scan_klasses_cl,
4892 4894 worker_id);
4893 4895 pss.end_strong_roots();
4894 4896
4895 4897 {
4896 4898 double start = os::elapsedTime();
4897 4899 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4898 4900 evac.do_void();
4899 4901 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4900 4902 double term_ms = pss.term_time()*1000.0;
4901 4903 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4902 4904 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4903 4905 }
4904 4906 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4905 4907 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4906 4908
4907 4909 if (ParallelGCVerbose) {
4908 4910 MutexLocker x(stats_lock());
4909 4911 pss.print_termination_stats(worker_id);
4910 4912 }
4911 4913
4912 4914 assert(pss.refs()->is_empty(), "should be empty");
4913 4915
4914 4916 // Close the inner scope so that the ResourceMark and HandleMark
4915 4917 // destructors are executed here and are included as part of the
4916 4918 // "GC Worker Time".
4917 4919 }
4918 4920
4919 4921 double end_time_ms = os::elapsedTime() * 1000.0;
4920 4922 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4921 4923 }
4922 4924 };
4923 4925
4924 4926 // *** Common G1 Evacuation Stuff
4925 4927
4926 4928 // Closures that support the filtering of CodeBlobs scanned during
4927 4929 // external root scanning.
4928 4930
4929 4931 // Closure applied to reference fields in code blobs (specifically nmethods)
4930 4932 // to determine whether an nmethod contains references that point into
4931 4933 // the collection set. Used as a predicate when walking code roots so
4932 4934 // that only nmethods that point into the collection set are added to the
4933 4935 // 'marked' list.
4934 4936
4935 4937 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4936 4938
4937 4939 class G1PointsIntoCSOopClosure : public OopClosure {
4938 4940 G1CollectedHeap* _g1;
4939 4941 bool _points_into_cs;
4940 4942 public:
4941 4943 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4942 4944 _g1(g1), _points_into_cs(false) { }
4943 4945
4944 4946 bool points_into_cs() const { return _points_into_cs; }
4945 4947
4946 4948 template <class T>
4947 4949 void do_oop_nv(T* p) {
4948 4950 if (!_points_into_cs) {
4949 4951 T heap_oop = oopDesc::load_heap_oop(p);
4950 4952 if (!oopDesc::is_null(heap_oop) &&
4951 4953 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4952 4954 _points_into_cs = true;
4953 4955 }
4954 4956 }
4955 4957 }
4956 4958
4957 4959 virtual void do_oop(oop* p) { do_oop_nv(p); }
4958 4960 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4959 4961 };
4960 4962
4961 4963 G1CollectedHeap* _g1;
4962 4964
4963 4965 public:
4964 4966 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4965 4967 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4966 4968
4967 4969 virtual void do_code_blob(CodeBlob* cb) {
4968 4970 nmethod* nm = cb->as_nmethod_or_null();
4969 4971 if (nm != NULL && !(nm->test_oops_do_mark())) {
4970 4972 G1PointsIntoCSOopClosure predicate_cl(_g1);
4971 4973 nm->oops_do(&predicate_cl);
4972 4974
4973 4975 if (predicate_cl.points_into_cs()) {
4974 4976 // At least one of the reference fields or the oop relocations
4975 4977 // in the nmethod points into the collection set. We have to
4976 4978 // 'mark' this nmethod.
4977 4979 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4978 4980 // or MarkingCodeBlobClosure::do_code_blob() change.
4979 4981 if (!nm->test_set_oops_do_mark()) {
4980 4982 do_newly_marked_nmethod(nm);
4981 4983 }
4982 4984 }
4983 4985 }
4984 4986 }
4985 4987 };
4986 4988
4987 4989 // This method is run in a GC worker.
4988 4990
4989 4991 void
4990 4992 G1CollectedHeap::
4991 4993 g1_process_strong_roots(bool is_scavenging,
4992 4994 ScanningOption so,
4993 4995 OopClosure* scan_non_heap_roots,
4994 4996 OopsInHeapRegionClosure* scan_rs,
4995 4997 G1KlassScanClosure* scan_klasses,
4996 4998 int worker_i) {
4997 4999
4998 5000 // First scan the strong roots
4999 5001 double ext_roots_start = os::elapsedTime();
5000 5002 double closure_app_time_sec = 0.0;
5001 5003
5002 5004 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5003 5005
5004 5006 // Walk the code cache w/o buffering, because StarTask cannot handle
5005 5007 // unaligned oop locations.
5006 5008 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5007 5009
5008 5010 process_strong_roots(false, // no scoping; this is parallel code
5009 5011 is_scavenging, so,
5010 5012 &buf_scan_non_heap_roots,
5011 5013 &eager_scan_code_roots,
5012 5014 scan_klasses
5013 5015 );
5014 5016
5015 5017 // Now the CM ref_processor roots.
5016 5018 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5017 5019 // We need to treat the discovered reference lists of the
5018 5020 // concurrent mark ref processor as roots and keep entries
5019 5021 // (which are added by the marking threads) on them live
5020 5022 // until they can be processed at the end of marking.
5021 5023 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5022 5024 }
5023 5025
5024 5026 // Finish up any enqueued closure apps (attributed as object copy time).
5025 5027 buf_scan_non_heap_roots.done();
5026 5028
5027 5029 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5028 5030
5029 5031 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5030 5032
5031 5033 double ext_root_time_ms =
5032 5034 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5033 5035
5034 5036 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5035 5037
5036 5038 // During conc marking we have to filter the per-thread SATB buffers
5037 5039 // to make sure we remove any oops into the CSet (which will show up
5038 5040 // as implicitly live).
5039 5041 double satb_filtering_ms = 0.0;
5040 5042 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5041 5043 if (mark_in_progress()) {
5042 5044 double satb_filter_start = os::elapsedTime();
5043 5045
5044 5046 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5045 5047
5046 5048 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5047 5049 }
5048 5050 }
5049 5051 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5050 5052
5051 5053 // Now scan the complement of the collection set.
5052 5054 if (scan_rs != NULL) {
5053 5055 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5054 5056 }
5055 5057 _process_strong_tasks->all_tasks_completed();
5056 5058 }
5057 5059
5058 5060 void
5059 5061 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5060 5062 OopClosure* non_root_closure) {
5061 5063 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5062 5064 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5063 5065 }
5064 5066
5065 5067 // Weak Reference Processing support
5066 5068
5067 5069 // An always "is_alive" closure that is used to preserve referents.
5068 5070 // If the object is non-null then it's alive. Used in the preservation
5069 5071 // of referent objects that are pointed to by reference objects
5070 5072 // discovered by the CM ref processor.
5071 5073 class G1AlwaysAliveClosure: public BoolObjectClosure {
5072 5074 G1CollectedHeap* _g1;
5073 5075 public:
5074 5076 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5075 5077 void do_object(oop p) { assert(false, "Do not call."); }
5076 5078 bool do_object_b(oop p) {
5077 5079 if (p != NULL) {
5078 5080 return true;
5079 5081 }
5080 5082 return false;
5081 5083 }
5082 5084 };
5083 5085
5084 5086 bool G1STWIsAliveClosure::do_object_b(oop p) {
5085 5087 // An object is reachable if it is outside the collection set,
5086 5088 // or is inside and copied.
5087 5089 return !_g1->obj_in_cs(p) || p->is_forwarded();
5088 5090 }
5089 5091
5090 5092 // Non Copying Keep Alive closure
5091 5093 class G1KeepAliveClosure: public OopClosure {
5092 5094 G1CollectedHeap* _g1;
5093 5095 public:
5094 5096 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5095 5097 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5096 5098 void do_oop( oop* p) {
5097 5099 oop obj = *p;
5098 5100
5099 5101 if (_g1->obj_in_cs(obj)) {
5100 5102 assert( obj->is_forwarded(), "invariant" );
5101 5103 *p = obj->forwardee();
5102 5104 }
5103 5105 }
5104 5106 };
5105 5107
5106 5108 // Copying Keep Alive closure - can be called from both
5107 5109 // serial and parallel code as long as different worker
5108 5110 // threads utilize different G1ParScanThreadState instances
5109 5111 // and different queues.
5110 5112
5111 5113 class G1CopyingKeepAliveClosure: public OopClosure {
5112 5114 G1CollectedHeap* _g1h;
5113 5115 OopClosure* _copy_non_heap_obj_cl;
5114 5116 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5115 5117 G1ParScanThreadState* _par_scan_state;
5116 5118
5117 5119 public:
5118 5120 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5119 5121 OopClosure* non_heap_obj_cl,
5120 5122 OopsInHeapRegionClosure* metadata_obj_cl,
5121 5123 G1ParScanThreadState* pss):
5122 5124 _g1h(g1h),
5123 5125 _copy_non_heap_obj_cl(non_heap_obj_cl),
5124 5126 _copy_metadata_obj_cl(metadata_obj_cl),
5125 5127 _par_scan_state(pss)
5126 5128 {}
5127 5129
5128 5130 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5129 5131 virtual void do_oop( oop* p) { do_oop_work(p); }
5130 5132
5131 5133 template <class T> void do_oop_work(T* p) {
5132 5134 oop obj = oopDesc::load_decode_heap_oop(p);
5133 5135
5134 5136 if (_g1h->obj_in_cs(obj)) {
5135 5137 // If the referent object has been forwarded (either copied
5136 5138 // to a new location or to itself in the event of an
5137 5139 // evacuation failure) then we need to update the reference
5138 5140 // field and, if both reference and referent are in the G1
5139 5141 // heap, update the RSet for the referent.
5140 5142 //
5141 5143 // If the referent has not been forwarded then we have to keep
5142 5144 // it alive by policy. Therefore we have copy the referent.
5143 5145 //
5144 5146 // If the reference field is in the G1 heap then we can push
5145 5147 // on the PSS queue. When the queue is drained (after each
5146 5148 // phase of reference processing) the object and it's followers
5147 5149 // will be copied, the reference field set to point to the
5148 5150 // new location, and the RSet updated. Otherwise we need to
5149 5151 // use the the non-heap or metadata closures directly to copy
5150 5152 // the refernt object and update the pointer, while avoiding
5151 5153 // updating the RSet.
5152 5154
5153 5155 if (_g1h->is_in_g1_reserved(p)) {
5154 5156 _par_scan_state->push_on_queue(p);
5155 5157 } else {
5156 5158 assert(!ClassLoaderDataGraph::contains((address)p),
5157 5159 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5158 5160 PTR_FORMAT, p));
5159 5161 _copy_non_heap_obj_cl->do_oop(p);
5160 5162 }
5161 5163 }
5162 5164 }
5163 5165 };
5164 5166
5165 5167 // Serial drain queue closure. Called as the 'complete_gc'
5166 5168 // closure for each discovered list in some of the
5167 5169 // reference processing phases.
5168 5170
5169 5171 class G1STWDrainQueueClosure: public VoidClosure {
5170 5172 protected:
5171 5173 G1CollectedHeap* _g1h;
5172 5174 G1ParScanThreadState* _par_scan_state;
5173 5175
5174 5176 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5175 5177
5176 5178 public:
5177 5179 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5178 5180 _g1h(g1h),
5179 5181 _par_scan_state(pss)
5180 5182 { }
5181 5183
5182 5184 void do_void() {
5183 5185 G1ParScanThreadState* const pss = par_scan_state();
5184 5186 pss->trim_queue();
5185 5187 }
5186 5188 };
5187 5189
5188 5190 // Parallel Reference Processing closures
5189 5191
5190 5192 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5191 5193 // processing during G1 evacuation pauses.
5192 5194
5193 5195 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5194 5196 private:
5195 5197 G1CollectedHeap* _g1h;
5196 5198 RefToScanQueueSet* _queues;
5197 5199 FlexibleWorkGang* _workers;
5198 5200 int _active_workers;
5199 5201
5200 5202 public:
5201 5203 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5202 5204 FlexibleWorkGang* workers,
5203 5205 RefToScanQueueSet *task_queues,
5204 5206 int n_workers) :
5205 5207 _g1h(g1h),
5206 5208 _queues(task_queues),
5207 5209 _workers(workers),
5208 5210 _active_workers(n_workers)
5209 5211 {
5210 5212 assert(n_workers > 0, "shouldn't call this otherwise");
5211 5213 }
5212 5214
5213 5215 // Executes the given task using concurrent marking worker threads.
5214 5216 virtual void execute(ProcessTask& task);
5215 5217 virtual void execute(EnqueueTask& task);
5216 5218 };
5217 5219
5218 5220 // Gang task for possibly parallel reference processing
5219 5221
5220 5222 class G1STWRefProcTaskProxy: public AbstractGangTask {
5221 5223 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5222 5224 ProcessTask& _proc_task;
5223 5225 G1CollectedHeap* _g1h;
5224 5226 RefToScanQueueSet *_task_queues;
5225 5227 ParallelTaskTerminator* _terminator;
5226 5228
5227 5229 public:
5228 5230 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5229 5231 G1CollectedHeap* g1h,
5230 5232 RefToScanQueueSet *task_queues,
5231 5233 ParallelTaskTerminator* terminator) :
5232 5234 AbstractGangTask("Process reference objects in parallel"),
5233 5235 _proc_task(proc_task),
5234 5236 _g1h(g1h),
5235 5237 _task_queues(task_queues),
5236 5238 _terminator(terminator)
5237 5239 {}
5238 5240
5239 5241 virtual void work(uint worker_id) {
5240 5242 // The reference processing task executed by a single worker.
5241 5243 ResourceMark rm;
5242 5244 HandleMark hm;
5243 5245
5244 5246 G1STWIsAliveClosure is_alive(_g1h);
5245 5247
5246 5248 G1ParScanThreadState pss(_g1h, worker_id);
5247 5249
5248 5250 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5249 5251 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5250 5252 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5251 5253
5252 5254 pss.set_evac_closure(&scan_evac_cl);
5253 5255 pss.set_evac_failure_closure(&evac_failure_cl);
5254 5256 pss.set_partial_scan_closure(&partial_scan_cl);
5255 5257
5256 5258 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5257 5259 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5258 5260
5259 5261 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5260 5262 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5261 5263
5262 5264 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5263 5265 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5264 5266
5265 5267 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5266 5268 // We also need to mark copied objects.
5267 5269 copy_non_heap_cl = ©_mark_non_heap_cl;
5268 5270 copy_metadata_cl = ©_mark_metadata_cl;
5269 5271 }
5270 5272
5271 5273 // Keep alive closure.
5272 5274 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5273 5275
5274 5276 // Complete GC closure
5275 5277 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5276 5278
5277 5279 // Call the reference processing task's work routine.
5278 5280 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5279 5281
5280 5282 // Note we cannot assert that the refs array is empty here as not all
5281 5283 // of the processing tasks (specifically phase2 - pp2_work) execute
5282 5284 // the complete_gc closure (which ordinarily would drain the queue) so
5283 5285 // the queue may not be empty.
5284 5286 }
5285 5287 };
5286 5288
5287 5289 // Driver routine for parallel reference processing.
5288 5290 // Creates an instance of the ref processing gang
5289 5291 // task and has the worker threads execute it.
5290 5292 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5291 5293 assert(_workers != NULL, "Need parallel worker threads.");
5292 5294
5293 5295 ParallelTaskTerminator terminator(_active_workers, _queues);
5294 5296 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5295 5297
5296 5298 _g1h->set_par_threads(_active_workers);
5297 5299 _workers->run_task(&proc_task_proxy);
5298 5300 _g1h->set_par_threads(0);
5299 5301 }
5300 5302
5301 5303 // Gang task for parallel reference enqueueing.
5302 5304
5303 5305 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5304 5306 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5305 5307 EnqueueTask& _enq_task;
5306 5308
5307 5309 public:
5308 5310 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5309 5311 AbstractGangTask("Enqueue reference objects in parallel"),
5310 5312 _enq_task(enq_task)
5311 5313 { }
5312 5314
5313 5315 virtual void work(uint worker_id) {
5314 5316 _enq_task.work(worker_id);
5315 5317 }
5316 5318 };
5317 5319
5318 5320 // Driver routine for parallel reference enqueing.
5319 5321 // Creates an instance of the ref enqueueing gang
5320 5322 // task and has the worker threads execute it.
5321 5323
5322 5324 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5323 5325 assert(_workers != NULL, "Need parallel worker threads.");
5324 5326
5325 5327 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5326 5328
5327 5329 _g1h->set_par_threads(_active_workers);
5328 5330 _workers->run_task(&enq_task_proxy);
5329 5331 _g1h->set_par_threads(0);
5330 5332 }
5331 5333
5332 5334 // End of weak reference support closures
5333 5335
5334 5336 // Abstract task used to preserve (i.e. copy) any referent objects
5335 5337 // that are in the collection set and are pointed to by reference
5336 5338 // objects discovered by the CM ref processor.
5337 5339
5338 5340 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5339 5341 protected:
5340 5342 G1CollectedHeap* _g1h;
5341 5343 RefToScanQueueSet *_queues;
5342 5344 ParallelTaskTerminator _terminator;
5343 5345 uint _n_workers;
5344 5346
5345 5347 public:
5346 5348 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5347 5349 AbstractGangTask("ParPreserveCMReferents"),
5348 5350 _g1h(g1h),
5349 5351 _queues(task_queues),
5350 5352 _terminator(workers, _queues),
5351 5353 _n_workers(workers)
5352 5354 { }
5353 5355
5354 5356 void work(uint worker_id) {
5355 5357 ResourceMark rm;
5356 5358 HandleMark hm;
5357 5359
5358 5360 G1ParScanThreadState pss(_g1h, worker_id);
5359 5361 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5360 5362 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5361 5363 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5362 5364
5363 5365 pss.set_evac_closure(&scan_evac_cl);
5364 5366 pss.set_evac_failure_closure(&evac_failure_cl);
5365 5367 pss.set_partial_scan_closure(&partial_scan_cl);
5366 5368
5367 5369 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5368 5370
5369 5371
5370 5372 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5371 5373 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5372 5374
5373 5375 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5374 5376 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5375 5377
5376 5378 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5377 5379 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5378 5380
5379 5381 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5380 5382 // We also need to mark copied objects.
5381 5383 copy_non_heap_cl = ©_mark_non_heap_cl;
5382 5384 copy_metadata_cl = ©_mark_metadata_cl;
5383 5385 }
5384 5386
5385 5387 // Is alive closure
5386 5388 G1AlwaysAliveClosure always_alive(_g1h);
5387 5389
5388 5390 // Copying keep alive closure. Applied to referent objects that need
5389 5391 // to be copied.
5390 5392 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5391 5393
5392 5394 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5393 5395
5394 5396 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5395 5397 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5396 5398
5397 5399 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5398 5400 // So this must be true - but assert just in case someone decides to
5399 5401 // change the worker ids.
5400 5402 assert(0 <= worker_id && worker_id < limit, "sanity");
5401 5403 assert(!rp->discovery_is_atomic(), "check this code");
5402 5404
5403 5405 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5404 5406 for (uint idx = worker_id; idx < limit; idx += stride) {
5405 5407 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5406 5408
5407 5409 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5408 5410 while (iter.has_next()) {
5409 5411 // Since discovery is not atomic for the CM ref processor, we
5410 5412 // can see some null referent objects.
5411 5413 iter.load_ptrs(DEBUG_ONLY(true));
5412 5414 oop ref = iter.obj();
5413 5415
5414 5416 // This will filter nulls.
5415 5417 if (iter.is_referent_alive()) {
5416 5418 iter.make_referent_alive();
5417 5419 }
5418 5420 iter.move_to_next();
5419 5421 }
5420 5422 }
5421 5423
5422 5424 // Drain the queue - which may cause stealing
5423 5425 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5424 5426 drain_queue.do_void();
5425 5427 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5426 5428 assert(pss.refs()->is_empty(), "should be");
5427 5429 }
5428 5430 };
5429 5431
5430 5432 // Weak Reference processing during an evacuation pause (part 1).
5431 5433 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5432 5434 double ref_proc_start = os::elapsedTime();
5433 5435
5434 5436 ReferenceProcessor* rp = _ref_processor_stw;
5435 5437 assert(rp->discovery_enabled(), "should have been enabled");
5436 5438
5437 5439 // Any reference objects, in the collection set, that were 'discovered'
5438 5440 // by the CM ref processor should have already been copied (either by
5439 5441 // applying the external root copy closure to the discovered lists, or
5440 5442 // by following an RSet entry).
5441 5443 //
5442 5444 // But some of the referents, that are in the collection set, that these
5443 5445 // reference objects point to may not have been copied: the STW ref
5444 5446 // processor would have seen that the reference object had already
5445 5447 // been 'discovered' and would have skipped discovering the reference,
5446 5448 // but would not have treated the reference object as a regular oop.
5447 5449 // As a reult the copy closure would not have been applied to the
5448 5450 // referent object.
5449 5451 //
5450 5452 // We need to explicitly copy these referent objects - the references
5451 5453 // will be processed at the end of remarking.
5452 5454 //
5453 5455 // We also need to do this copying before we process the reference
5454 5456 // objects discovered by the STW ref processor in case one of these
5455 5457 // referents points to another object which is also referenced by an
5456 5458 // object discovered by the STW ref processor.
5457 5459
5458 5460 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5459 5461 no_of_gc_workers == workers()->active_workers(),
5460 5462 "Need to reset active GC workers");
5461 5463
5462 5464 set_par_threads(no_of_gc_workers);
5463 5465 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5464 5466 no_of_gc_workers,
5465 5467 _task_queues);
5466 5468
5467 5469 if (G1CollectedHeap::use_parallel_gc_threads()) {
5468 5470 workers()->run_task(&keep_cm_referents);
5469 5471 } else {
5470 5472 keep_cm_referents.work(0);
5471 5473 }
5472 5474
5473 5475 set_par_threads(0);
5474 5476
5475 5477 // Closure to test whether a referent is alive.
5476 5478 G1STWIsAliveClosure is_alive(this);
5477 5479
5478 5480 // Even when parallel reference processing is enabled, the processing
5479 5481 // of JNI refs is serial and performed serially by the current thread
5480 5482 // rather than by a worker. The following PSS will be used for processing
5481 5483 // JNI refs.
5482 5484
5483 5485 // Use only a single queue for this PSS.
5484 5486 G1ParScanThreadState pss(this, 0);
5485 5487
5486 5488 // We do not embed a reference processor in the copying/scanning
5487 5489 // closures while we're actually processing the discovered
5488 5490 // reference objects.
5489 5491 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5490 5492 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5491 5493 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5492 5494
5493 5495 pss.set_evac_closure(&scan_evac_cl);
5494 5496 pss.set_evac_failure_closure(&evac_failure_cl);
5495 5497 pss.set_partial_scan_closure(&partial_scan_cl);
5496 5498
5497 5499 assert(pss.refs()->is_empty(), "pre-condition");
5498 5500
5499 5501 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5500 5502 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5501 5503
5502 5504 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5503 5505 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5504 5506
5505 5507 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5506 5508 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5507 5509
5508 5510 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5509 5511 // We also need to mark copied objects.
5510 5512 copy_non_heap_cl = ©_mark_non_heap_cl;
5511 5513 copy_metadata_cl = ©_mark_metadata_cl;
5512 5514 }
5513 5515
5514 5516 // Keep alive closure.
5515 5517 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5516 5518
5517 5519 // Serial Complete GC closure
5518 5520 G1STWDrainQueueClosure drain_queue(this, &pss);
5519 5521
5520 5522 // Setup the soft refs policy...
5521 5523 rp->setup_policy(false);
5522 5524
5523 5525 if (!rp->processing_is_mt()) {
5524 5526 // Serial reference processing...
5525 5527 rp->process_discovered_references(&is_alive,
5526 5528 &keep_alive,
5527 5529 &drain_queue,
5528 5530 NULL);
5529 5531 } else {
5530 5532 // Parallel reference processing
5531 5533 assert(rp->num_q() == no_of_gc_workers, "sanity");
5532 5534 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5533 5535
5534 5536 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5535 5537 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5536 5538 }
5537 5539
5538 5540 // We have completed copying any necessary live referent objects
5539 5541 // (that were not copied during the actual pause) so we can
5540 5542 // retire any active alloc buffers
5541 5543 pss.retire_alloc_buffers();
5542 5544 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5543 5545
5544 5546 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5545 5547 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5546 5548 }
5547 5549
5548 5550 // Weak Reference processing during an evacuation pause (part 2).
5549 5551 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5550 5552 double ref_enq_start = os::elapsedTime();
5551 5553
5552 5554 ReferenceProcessor* rp = _ref_processor_stw;
5553 5555 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5554 5556
5555 5557 // Now enqueue any remaining on the discovered lists on to
5556 5558 // the pending list.
5557 5559 if (!rp->processing_is_mt()) {
5558 5560 // Serial reference processing...
5559 5561 rp->enqueue_discovered_references();
5560 5562 } else {
5561 5563 // Parallel reference enqueuing
5562 5564
5563 5565 assert(no_of_gc_workers == workers()->active_workers(),
5564 5566 "Need to reset active workers");
5565 5567 assert(rp->num_q() == no_of_gc_workers, "sanity");
5566 5568 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5567 5569
5568 5570 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5569 5571 rp->enqueue_discovered_references(&par_task_executor);
5570 5572 }
5571 5573
5572 5574 rp->verify_no_references_recorded();
5573 5575 assert(!rp->discovery_enabled(), "should have been disabled");
5574 5576
5575 5577 // FIXME
5576 5578 // CM's reference processing also cleans up the string and symbol tables.
5577 5579 // Should we do that here also? We could, but it is a serial operation
5578 5580 // and could signicantly increase the pause time.
5579 5581
5580 5582 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5581 5583 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5582 5584 }
5583 5585
5584 5586 void G1CollectedHeap::evacuate_collection_set() {
5585 5587 _expand_heap_after_alloc_failure = true;
5586 5588 set_evacuation_failed(false);
5587 5589
5588 5590 // Should G1EvacuationFailureALot be in effect for this GC?
5589 5591 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5590 5592
5591 5593 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5592 5594 concurrent_g1_refine()->set_use_cache(false);
5593 5595 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5594 5596
5595 5597 uint n_workers;
5596 5598 if (G1CollectedHeap::use_parallel_gc_threads()) {
5597 5599 n_workers =
5598 5600 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5599 5601 workers()->active_workers(),
5600 5602 Threads::number_of_non_daemon_threads());
5601 5603 assert(UseDynamicNumberOfGCThreads ||
5602 5604 n_workers == workers()->total_workers(),
5603 5605 "If not dynamic should be using all the workers");
5604 5606 workers()->set_active_workers(n_workers);
5605 5607 set_par_threads(n_workers);
5606 5608 } else {
5607 5609 assert(n_par_threads() == 0,
5608 5610 "Should be the original non-parallel value");
5609 5611 n_workers = 1;
5610 5612 }
5611 5613
5612 5614 G1ParTask g1_par_task(this, _task_queues);
5613 5615
5614 5616 init_for_evac_failure(NULL);
5615 5617
5616 5618 rem_set()->prepare_for_younger_refs_iterate(true);
5617 5619
5618 5620 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5619 5621 double start_par_time_sec = os::elapsedTime();
5620 5622 double end_par_time_sec;
5621 5623
5622 5624 {
5623 5625 StrongRootsScope srs(this);
5624 5626
5625 5627 if (G1CollectedHeap::use_parallel_gc_threads()) {
5626 5628 // The individual threads will set their evac-failure closures.
5627 5629 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5628 5630 // These tasks use ShareHeap::_process_strong_tasks
5629 5631 assert(UseDynamicNumberOfGCThreads ||
5630 5632 workers()->active_workers() == workers()->total_workers(),
5631 5633 "If not dynamic should be using all the workers");
5632 5634 workers()->run_task(&g1_par_task);
5633 5635 } else {
5634 5636 g1_par_task.set_for_termination(n_workers);
5635 5637 g1_par_task.work(0);
5636 5638 }
5637 5639 end_par_time_sec = os::elapsedTime();
5638 5640
5639 5641 // Closing the inner scope will execute the destructor
5640 5642 // for the StrongRootsScope object. We record the current
5641 5643 // elapsed time before closing the scope so that time
5642 5644 // taken for the SRS destructor is NOT included in the
5643 5645 // reported parallel time.
5644 5646 }
5645 5647
5646 5648 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5647 5649 g1_policy()->phase_times()->record_par_time(par_time_ms);
5648 5650
5649 5651 double code_root_fixup_time_ms =
5650 5652 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5651 5653 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5652 5654
5653 5655 set_par_threads(0);
5654 5656
5655 5657 // Process any discovered reference objects - we have
5656 5658 // to do this _before_ we retire the GC alloc regions
5657 5659 // as we may have to copy some 'reachable' referent
5658 5660 // objects (and their reachable sub-graphs) that were
5659 5661 // not copied during the pause.
5660 5662 process_discovered_references(n_workers);
5661 5663
5662 5664 // Weak root processing.
5663 5665 // Note: when JSR 292 is enabled and code blobs can contain
5664 5666 // non-perm oops then we will need to process the code blobs
5665 5667 // here too.
5666 5668 {
5667 5669 G1STWIsAliveClosure is_alive(this);
5668 5670 G1KeepAliveClosure keep_alive(this);
5669 5671 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5670 5672 }
5671 5673
5672 5674 release_gc_alloc_regions(n_workers);
5673 5675 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5674 5676
5675 5677 concurrent_g1_refine()->clear_hot_cache();
5676 5678 concurrent_g1_refine()->set_use_cache(true);
5677 5679
5678 5680 finalize_for_evac_failure();
5679 5681
5680 5682 if (evacuation_failed()) {
5681 5683 remove_self_forwarding_pointers();
5682 5684
5683 5685 // Reset the G1EvacuationFailureALot counters and flags
5684 5686 // Note: the values are reset only when an actual
5685 5687 // evacuation failure occurs.
5686 5688 NOT_PRODUCT(reset_evacuation_should_fail();)
5687 5689 }
5688 5690
5689 5691 // Enqueue any remaining references remaining on the STW
5690 5692 // reference processor's discovered lists. We need to do
5691 5693 // this after the card table is cleaned (and verified) as
5692 5694 // the act of enqueuing entries on to the pending list
5693 5695 // will log these updates (and dirty their associated
5694 5696 // cards). We need these updates logged to update any
5695 5697 // RSets.
5696 5698 enqueue_discovered_references(n_workers);
5697 5699
5698 5700 if (G1DeferredRSUpdate) {
5699 5701 RedirtyLoggedCardTableEntryFastClosure redirty;
5700 5702 dirty_card_queue_set().set_closure(&redirty);
5701 5703 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5702 5704
5703 5705 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5704 5706 dcq.merge_bufferlists(&dirty_card_queue_set());
5705 5707 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5706 5708 }
5707 5709 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5708 5710 }
5709 5711
5710 5712 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5711 5713 size_t* pre_used,
5712 5714 FreeRegionList* free_list,
5713 5715 OldRegionSet* old_proxy_set,
5714 5716 HumongousRegionSet* humongous_proxy_set,
5715 5717 HRRSCleanupTask* hrrs_cleanup_task,
5716 5718 bool par) {
5717 5719 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5718 5720 if (hr->isHumongous()) {
5719 5721 assert(hr->startsHumongous(), "we should only see starts humongous");
5720 5722 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5721 5723 } else {
5722 5724 _old_set.remove_with_proxy(hr, old_proxy_set);
5723 5725 free_region(hr, pre_used, free_list, par);
5724 5726 }
5725 5727 } else {
5726 5728 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5727 5729 }
5728 5730 }
5729 5731
5730 5732 void G1CollectedHeap::free_region(HeapRegion* hr,
5731 5733 size_t* pre_used,
5732 5734 FreeRegionList* free_list,
5733 5735 bool par) {
5734 5736 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5735 5737 assert(!hr->is_empty(), "the region should not be empty");
5736 5738 assert(free_list != NULL, "pre-condition");
5737 5739
5738 5740 *pre_used += hr->used();
5739 5741 hr->hr_clear(par, true /* clear_space */);
5740 5742 free_list->add_as_head(hr);
5741 5743 }
5742 5744
5743 5745 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5744 5746 size_t* pre_used,
5745 5747 FreeRegionList* free_list,
5746 5748 HumongousRegionSet* humongous_proxy_set,
5747 5749 bool par) {
5748 5750 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5749 5751 assert(free_list != NULL, "pre-condition");
5750 5752 assert(humongous_proxy_set != NULL, "pre-condition");
5751 5753
5752 5754 size_t hr_used = hr->used();
5753 5755 size_t hr_capacity = hr->capacity();
5754 5756 size_t hr_pre_used = 0;
5755 5757 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5756 5758 // We need to read this before we make the region non-humongous,
5757 5759 // otherwise the information will be gone.
5758 5760 uint last_index = hr->last_hc_index();
5759 5761 hr->set_notHumongous();
5760 5762 free_region(hr, &hr_pre_used, free_list, par);
5761 5763
5762 5764 uint i = hr->hrs_index() + 1;
5763 5765 while (i < last_index) {
5764 5766 HeapRegion* curr_hr = region_at(i);
5765 5767 assert(curr_hr->continuesHumongous(), "invariant");
5766 5768 curr_hr->set_notHumongous();
5767 5769 free_region(curr_hr, &hr_pre_used, free_list, par);
5768 5770 i += 1;
5769 5771 }
5770 5772 assert(hr_pre_used == hr_used,
5771 5773 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5772 5774 "should be the same", hr_pre_used, hr_used));
5773 5775 *pre_used += hr_pre_used;
5774 5776 }
5775 5777
5776 5778 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5777 5779 FreeRegionList* free_list,
5778 5780 OldRegionSet* old_proxy_set,
5779 5781 HumongousRegionSet* humongous_proxy_set,
5780 5782 bool par) {
5781 5783 if (pre_used > 0) {
5782 5784 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5783 5785 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5784 5786 assert(_summary_bytes_used >= pre_used,
5785 5787 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5786 5788 "should be >= pre_used: "SIZE_FORMAT,
5787 5789 _summary_bytes_used, pre_used));
5788 5790 _summary_bytes_used -= pre_used;
5789 5791 }
5790 5792 if (free_list != NULL && !free_list->is_empty()) {
5791 5793 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5792 5794 _free_list.add_as_head(free_list);
5793 5795 }
5794 5796 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5795 5797 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5796 5798 _old_set.update_from_proxy(old_proxy_set);
5797 5799 }
5798 5800 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5799 5801 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5800 5802 _humongous_set.update_from_proxy(humongous_proxy_set);
5801 5803 }
5802 5804 }
5803 5805
5804 5806 class G1ParCleanupCTTask : public AbstractGangTask {
5805 5807 CardTableModRefBS* _ct_bs;
5806 5808 G1CollectedHeap* _g1h;
5807 5809 HeapRegion* volatile _su_head;
5808 5810 public:
5809 5811 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5810 5812 G1CollectedHeap* g1h) :
5811 5813 AbstractGangTask("G1 Par Cleanup CT Task"),
5812 5814 _ct_bs(ct_bs), _g1h(g1h) { }
5813 5815
5814 5816 void work(uint worker_id) {
5815 5817 HeapRegion* r;
5816 5818 while (r = _g1h->pop_dirty_cards_region()) {
5817 5819 clear_cards(r);
5818 5820 }
5819 5821 }
5820 5822
5821 5823 void clear_cards(HeapRegion* r) {
5822 5824 // Cards of the survivors should have already been dirtied.
5823 5825 if (!r->is_survivor()) {
5824 5826 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5825 5827 }
5826 5828 }
5827 5829 };
5828 5830
5829 5831 #ifndef PRODUCT
5830 5832 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5831 5833 G1CollectedHeap* _g1h;
5832 5834 CardTableModRefBS* _ct_bs;
5833 5835 public:
5834 5836 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5835 5837 : _g1h(g1h), _ct_bs(ct_bs) { }
5836 5838 virtual bool doHeapRegion(HeapRegion* r) {
5837 5839 if (r->is_survivor()) {
5838 5840 _g1h->verify_dirty_region(r);
5839 5841 } else {
5840 5842 _g1h->verify_not_dirty_region(r);
5841 5843 }
5842 5844 return false;
5843 5845 }
5844 5846 };
5845 5847
5846 5848 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5847 5849 // All of the region should be clean.
5848 5850 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5849 5851 MemRegion mr(hr->bottom(), hr->end());
5850 5852 ct_bs->verify_not_dirty_region(mr);
5851 5853 }
5852 5854
5853 5855 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5854 5856 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5855 5857 // dirty allocated blocks as they allocate them. The thread that
5856 5858 // retires each region and replaces it with a new one will do a
5857 5859 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5858 5860 // not dirty that area (one less thing to have to do while holding
5859 5861 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5860 5862 // is dirty.
5861 5863 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5862 5864 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5863 5865 ct_bs->verify_dirty_region(mr);
5864 5866 }
5865 5867
5866 5868 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5867 5869 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5868 5870 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5869 5871 verify_dirty_region(hr);
5870 5872 }
5871 5873 }
5872 5874
5873 5875 void G1CollectedHeap::verify_dirty_young_regions() {
5874 5876 verify_dirty_young_list(_young_list->first_region());
5875 5877 }
5876 5878 #endif
5877 5879
5878 5880 void G1CollectedHeap::cleanUpCardTable() {
5879 5881 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5880 5882 double start = os::elapsedTime();
5881 5883
5882 5884 {
5883 5885 // Iterate over the dirty cards region list.
5884 5886 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5885 5887
5886 5888 if (G1CollectedHeap::use_parallel_gc_threads()) {
5887 5889 set_par_threads();
5888 5890 workers()->run_task(&cleanup_task);
5889 5891 set_par_threads(0);
5890 5892 } else {
5891 5893 while (_dirty_cards_region_list) {
5892 5894 HeapRegion* r = _dirty_cards_region_list;
5893 5895 cleanup_task.clear_cards(r);
5894 5896 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5895 5897 if (_dirty_cards_region_list == r) {
5896 5898 // The last region.
5897 5899 _dirty_cards_region_list = NULL;
5898 5900 }
5899 5901 r->set_next_dirty_cards_region(NULL);
5900 5902 }
5901 5903 }
5902 5904 #ifndef PRODUCT
5903 5905 if (G1VerifyCTCleanup || VerifyAfterGC) {
5904 5906 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5905 5907 heap_region_iterate(&cleanup_verifier);
5906 5908 }
5907 5909 #endif
5908 5910 }
5909 5911
5910 5912 double elapsed = os::elapsedTime() - start;
5911 5913 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5912 5914 }
5913 5915
5914 5916 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5915 5917 size_t pre_used = 0;
5916 5918 FreeRegionList local_free_list("Local List for CSet Freeing");
5917 5919
5918 5920 double young_time_ms = 0.0;
5919 5921 double non_young_time_ms = 0.0;
5920 5922
5921 5923 // Since the collection set is a superset of the the young list,
5922 5924 // all we need to do to clear the young list is clear its
5923 5925 // head and length, and unlink any young regions in the code below
5924 5926 _young_list->clear();
5925 5927
5926 5928 G1CollectorPolicy* policy = g1_policy();
5927 5929
5928 5930 double start_sec = os::elapsedTime();
5929 5931 bool non_young = true;
5930 5932
5931 5933 HeapRegion* cur = cs_head;
5932 5934 int age_bound = -1;
5933 5935 size_t rs_lengths = 0;
5934 5936
5935 5937 while (cur != NULL) {
5936 5938 assert(!is_on_master_free_list(cur), "sanity");
5937 5939 if (non_young) {
5938 5940 if (cur->is_young()) {
5939 5941 double end_sec = os::elapsedTime();
5940 5942 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5941 5943 non_young_time_ms += elapsed_ms;
5942 5944
5943 5945 start_sec = os::elapsedTime();
5944 5946 non_young = false;
5945 5947 }
5946 5948 } else {
5947 5949 if (!cur->is_young()) {
5948 5950 double end_sec = os::elapsedTime();
5949 5951 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5950 5952 young_time_ms += elapsed_ms;
5951 5953
5952 5954 start_sec = os::elapsedTime();
5953 5955 non_young = true;
5954 5956 }
5955 5957 }
5956 5958
5957 5959 rs_lengths += cur->rem_set()->occupied();
5958 5960
5959 5961 HeapRegion* next = cur->next_in_collection_set();
5960 5962 assert(cur->in_collection_set(), "bad CS");
5961 5963 cur->set_next_in_collection_set(NULL);
5962 5964 cur->set_in_collection_set(false);
5963 5965
5964 5966 if (cur->is_young()) {
5965 5967 int index = cur->young_index_in_cset();
5966 5968 assert(index != -1, "invariant");
5967 5969 assert((uint) index < policy->young_cset_region_length(), "invariant");
5968 5970 size_t words_survived = _surviving_young_words[index];
5969 5971 cur->record_surv_words_in_group(words_survived);
5970 5972
5971 5973 // At this point the we have 'popped' cur from the collection set
5972 5974 // (linked via next_in_collection_set()) but it is still in the
5973 5975 // young list (linked via next_young_region()). Clear the
5974 5976 // _next_young_region field.
5975 5977 cur->set_next_young_region(NULL);
5976 5978 } else {
5977 5979 int index = cur->young_index_in_cset();
5978 5980 assert(index == -1, "invariant");
5979 5981 }
5980 5982
5981 5983 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5982 5984 (!cur->is_young() && cur->young_index_in_cset() == -1),
5983 5985 "invariant" );
5984 5986
5985 5987 if (!cur->evacuation_failed()) {
5986 5988 MemRegion used_mr = cur->used_region();
5987 5989
5988 5990 // And the region is empty.
5989 5991 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5990 5992 free_region(cur, &pre_used, &local_free_list, false /* par */);
5991 5993 } else {
5992 5994 cur->uninstall_surv_rate_group();
5993 5995 if (cur->is_young()) {
5994 5996 cur->set_young_index_in_cset(-1);
5995 5997 }
5996 5998 cur->set_not_young();
5997 5999 cur->set_evacuation_failed(false);
5998 6000 // The region is now considered to be old.
5999 6001 _old_set.add(cur);
6000 6002 }
6001 6003 cur = next;
6002 6004 }
6003 6005
6004 6006 policy->record_max_rs_lengths(rs_lengths);
6005 6007 policy->cset_regions_freed();
6006 6008
6007 6009 double end_sec = os::elapsedTime();
6008 6010 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6009 6011
6010 6012 if (non_young) {
6011 6013 non_young_time_ms += elapsed_ms;
6012 6014 } else {
6013 6015 young_time_ms += elapsed_ms;
6014 6016 }
6015 6017
6016 6018 update_sets_after_freeing_regions(pre_used, &local_free_list,
6017 6019 NULL /* old_proxy_set */,
6018 6020 NULL /* humongous_proxy_set */,
6019 6021 false /* par */);
6020 6022 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6021 6023 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6022 6024 }
6023 6025
6024 6026 // This routine is similar to the above but does not record
6025 6027 // any policy statistics or update free lists; we are abandoning
6026 6028 // the current incremental collection set in preparation of a
6027 6029 // full collection. After the full GC we will start to build up
6028 6030 // the incremental collection set again.
6029 6031 // This is only called when we're doing a full collection
6030 6032 // and is immediately followed by the tearing down of the young list.
6031 6033
6032 6034 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6033 6035 HeapRegion* cur = cs_head;
6034 6036
6035 6037 while (cur != NULL) {
6036 6038 HeapRegion* next = cur->next_in_collection_set();
6037 6039 assert(cur->in_collection_set(), "bad CS");
6038 6040 cur->set_next_in_collection_set(NULL);
6039 6041 cur->set_in_collection_set(false);
6040 6042 cur->set_young_index_in_cset(-1);
6041 6043 cur = next;
6042 6044 }
6043 6045 }
6044 6046
6045 6047 void G1CollectedHeap::set_free_regions_coming() {
6046 6048 if (G1ConcRegionFreeingVerbose) {
6047 6049 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6048 6050 "setting free regions coming");
6049 6051 }
6050 6052
6051 6053 assert(!free_regions_coming(), "pre-condition");
6052 6054 _free_regions_coming = true;
6053 6055 }
6054 6056
6055 6057 void G1CollectedHeap::reset_free_regions_coming() {
6056 6058 assert(free_regions_coming(), "pre-condition");
6057 6059
6058 6060 {
6059 6061 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6060 6062 _free_regions_coming = false;
6061 6063 SecondaryFreeList_lock->notify_all();
6062 6064 }
6063 6065
6064 6066 if (G1ConcRegionFreeingVerbose) {
6065 6067 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6066 6068 "reset free regions coming");
6067 6069 }
6068 6070 }
6069 6071
6070 6072 void G1CollectedHeap::wait_while_free_regions_coming() {
6071 6073 // Most of the time we won't have to wait, so let's do a quick test
6072 6074 // first before we take the lock.
6073 6075 if (!free_regions_coming()) {
6074 6076 return;
6075 6077 }
6076 6078
6077 6079 if (G1ConcRegionFreeingVerbose) {
6078 6080 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6079 6081 "waiting for free regions");
6080 6082 }
6081 6083
6082 6084 {
6083 6085 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6084 6086 while (free_regions_coming()) {
6085 6087 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6086 6088 }
6087 6089 }
6088 6090
6089 6091 if (G1ConcRegionFreeingVerbose) {
6090 6092 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6091 6093 "done waiting for free regions");
6092 6094 }
6093 6095 }
6094 6096
6095 6097 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6096 6098 assert(heap_lock_held_for_gc(),
6097 6099 "the heap lock should already be held by or for this thread");
6098 6100 _young_list->push_region(hr);
6099 6101 }
6100 6102
6101 6103 class NoYoungRegionsClosure: public HeapRegionClosure {
6102 6104 private:
6103 6105 bool _success;
6104 6106 public:
6105 6107 NoYoungRegionsClosure() : _success(true) { }
6106 6108 bool doHeapRegion(HeapRegion* r) {
6107 6109 if (r->is_young()) {
6108 6110 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6109 6111 r->bottom(), r->end());
6110 6112 _success = false;
6111 6113 }
6112 6114 return false;
6113 6115 }
6114 6116 bool success() { return _success; }
6115 6117 };
6116 6118
6117 6119 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6118 6120 bool ret = _young_list->check_list_empty(check_sample);
6119 6121
6120 6122 if (check_heap) {
6121 6123 NoYoungRegionsClosure closure;
6122 6124 heap_region_iterate(&closure);
6123 6125 ret = ret && closure.success();
6124 6126 }
6125 6127
6126 6128 return ret;
6127 6129 }
6128 6130
6129 6131 class TearDownRegionSetsClosure : public HeapRegionClosure {
6130 6132 private:
6131 6133 OldRegionSet *_old_set;
6132 6134
6133 6135 public:
6134 6136 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6135 6137
6136 6138 bool doHeapRegion(HeapRegion* r) {
6137 6139 if (r->is_empty()) {
6138 6140 // We ignore empty regions, we'll empty the free list afterwards
6139 6141 } else if (r->is_young()) {
6140 6142 // We ignore young regions, we'll empty the young list afterwards
6141 6143 } else if (r->isHumongous()) {
6142 6144 // We ignore humongous regions, we're not tearing down the
6143 6145 // humongous region set
6144 6146 } else {
6145 6147 // The rest should be old
6146 6148 _old_set->remove(r);
6147 6149 }
6148 6150 return false;
6149 6151 }
6150 6152
6151 6153 ~TearDownRegionSetsClosure() {
6152 6154 assert(_old_set->is_empty(), "post-condition");
6153 6155 }
6154 6156 };
6155 6157
6156 6158 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6157 6159 assert_at_safepoint(true /* should_be_vm_thread */);
6158 6160
6159 6161 if (!free_list_only) {
6160 6162 TearDownRegionSetsClosure cl(&_old_set);
6161 6163 heap_region_iterate(&cl);
6162 6164
6163 6165 // Need to do this after the heap iteration to be able to
6164 6166 // recognize the young regions and ignore them during the iteration.
6165 6167 _young_list->empty_list();
6166 6168 }
6167 6169 _free_list.remove_all();
6168 6170 }
6169 6171
6170 6172 class RebuildRegionSetsClosure : public HeapRegionClosure {
6171 6173 private:
6172 6174 bool _free_list_only;
6173 6175 OldRegionSet* _old_set;
6174 6176 FreeRegionList* _free_list;
6175 6177 size_t _total_used;
6176 6178
6177 6179 public:
6178 6180 RebuildRegionSetsClosure(bool free_list_only,
6179 6181 OldRegionSet* old_set, FreeRegionList* free_list) :
6180 6182 _free_list_only(free_list_only),
6181 6183 _old_set(old_set), _free_list(free_list), _total_used(0) {
6182 6184 assert(_free_list->is_empty(), "pre-condition");
6183 6185 if (!free_list_only) {
6184 6186 assert(_old_set->is_empty(), "pre-condition");
6185 6187 }
6186 6188 }
6187 6189
6188 6190 bool doHeapRegion(HeapRegion* r) {
6189 6191 if (r->continuesHumongous()) {
6190 6192 return false;
6191 6193 }
6192 6194
6193 6195 if (r->is_empty()) {
6194 6196 // Add free regions to the free list
6195 6197 _free_list->add_as_tail(r);
6196 6198 } else if (!_free_list_only) {
6197 6199 assert(!r->is_young(), "we should not come across young regions");
6198 6200
6199 6201 if (r->isHumongous()) {
6200 6202 // We ignore humongous regions, we left the humongous set unchanged
6201 6203 } else {
6202 6204 // The rest should be old, add them to the old set
6203 6205 _old_set->add(r);
6204 6206 }
6205 6207 _total_used += r->used();
6206 6208 }
6207 6209
6208 6210 return false;
6209 6211 }
6210 6212
6211 6213 size_t total_used() {
6212 6214 return _total_used;
6213 6215 }
6214 6216 };
6215 6217
6216 6218 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6217 6219 assert_at_safepoint(true /* should_be_vm_thread */);
6218 6220
6219 6221 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6220 6222 heap_region_iterate(&cl);
6221 6223
6222 6224 if (!free_list_only) {
6223 6225 _summary_bytes_used = cl.total_used();
6224 6226 }
6225 6227 assert(_summary_bytes_used == recalculate_used(),
6226 6228 err_msg("inconsistent _summary_bytes_used, "
6227 6229 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6228 6230 _summary_bytes_used, recalculate_used()));
6229 6231 }
6230 6232
6231 6233 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6232 6234 _refine_cte_cl->set_concurrent(concurrent);
6233 6235 }
6234 6236
6235 6237 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6236 6238 HeapRegion* hr = heap_region_containing(p);
6237 6239 if (hr == NULL) {
6238 6240 return false;
6239 6241 } else {
6240 6242 return hr->is_in(p);
6241 6243 }
6242 6244 }
6243 6245
6244 6246 // Methods for the mutator alloc region
6245 6247
6246 6248 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6247 6249 bool force) {
6248 6250 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6249 6251 assert(!force || g1_policy()->can_expand_young_list(),
6250 6252 "if force is true we should be able to expand the young list");
6251 6253 bool young_list_full = g1_policy()->is_young_list_full();
6252 6254 if (force || !young_list_full) {
6253 6255 HeapRegion* new_alloc_region = new_region(word_size,
6254 6256 false /* do_expand */);
6255 6257 if (new_alloc_region != NULL) {
6256 6258 set_region_short_lived_locked(new_alloc_region);
6257 6259 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6258 6260 return new_alloc_region;
6259 6261 }
6260 6262 }
6261 6263 return NULL;
6262 6264 }
6263 6265
6264 6266 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6265 6267 size_t allocated_bytes) {
6266 6268 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6267 6269 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6268 6270
6269 6271 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6270 6272 _summary_bytes_used += allocated_bytes;
6271 6273 _hr_printer.retire(alloc_region);
6272 6274 // We update the eden sizes here, when the region is retired,
6273 6275 // instead of when it's allocated, since this is the point that its
6274 6276 // used space has been recored in _summary_bytes_used.
6275 6277 g1mm()->update_eden_size();
6276 6278 }
6277 6279
6278 6280 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6279 6281 bool force) {
6280 6282 return _g1h->new_mutator_alloc_region(word_size, force);
6281 6283 }
6282 6284
6283 6285 void G1CollectedHeap::set_par_threads() {
6284 6286 // Don't change the number of workers. Use the value previously set
6285 6287 // in the workgroup.
6286 6288 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6287 6289 uint n_workers = workers()->active_workers();
6288 6290 assert(UseDynamicNumberOfGCThreads ||
6289 6291 n_workers == workers()->total_workers(),
6290 6292 "Otherwise should be using the total number of workers");
6291 6293 if (n_workers == 0) {
6292 6294 assert(false, "Should have been set in prior evacuation pause.");
6293 6295 n_workers = ParallelGCThreads;
6294 6296 workers()->set_active_workers(n_workers);
6295 6297 }
6296 6298 set_par_threads(n_workers);
6297 6299 }
6298 6300
6299 6301 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6300 6302 size_t allocated_bytes) {
6301 6303 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6302 6304 }
6303 6305
6304 6306 // Methods for the GC alloc regions
6305 6307
6306 6308 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6307 6309 uint count,
6308 6310 GCAllocPurpose ap) {
6309 6311 assert(FreeList_lock->owned_by_self(), "pre-condition");
6310 6312
6311 6313 if (count < g1_policy()->max_regions(ap)) {
6312 6314 HeapRegion* new_alloc_region = new_region(word_size,
6313 6315 true /* do_expand */);
6314 6316 if (new_alloc_region != NULL) {
6315 6317 // We really only need to do this for old regions given that we
6316 6318 // should never scan survivors. But it doesn't hurt to do it
6317 6319 // for survivors too.
6318 6320 new_alloc_region->set_saved_mark();
6319 6321 if (ap == GCAllocForSurvived) {
6320 6322 new_alloc_region->set_survivor();
6321 6323 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6322 6324 } else {
6323 6325 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6324 6326 }
6325 6327 bool during_im = g1_policy()->during_initial_mark_pause();
6326 6328 new_alloc_region->note_start_of_copying(during_im);
6327 6329 return new_alloc_region;
6328 6330 } else {
6329 6331 g1_policy()->note_alloc_region_limit_reached(ap);
6330 6332 }
6331 6333 }
6332 6334 return NULL;
6333 6335 }
6334 6336
6335 6337 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6336 6338 size_t allocated_bytes,
6337 6339 GCAllocPurpose ap) {
6338 6340 bool during_im = g1_policy()->during_initial_mark_pause();
6339 6341 alloc_region->note_end_of_copying(during_im);
6340 6342 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6341 6343 if (ap == GCAllocForSurvived) {
6342 6344 young_list()->add_survivor_region(alloc_region);
6343 6345 } else {
6344 6346 _old_set.add(alloc_region);
6345 6347 }
6346 6348 _hr_printer.retire(alloc_region);
6347 6349 }
6348 6350
6349 6351 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6350 6352 bool force) {
6351 6353 assert(!force, "not supported for GC alloc regions");
6352 6354 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6353 6355 }
6354 6356
6355 6357 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6356 6358 size_t allocated_bytes) {
6357 6359 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6358 6360 GCAllocForSurvived);
6359 6361 }
6360 6362
6361 6363 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6362 6364 bool force) {
6363 6365 assert(!force, "not supported for GC alloc regions");
6364 6366 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6365 6367 }
6366 6368
6367 6369 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6368 6370 size_t allocated_bytes) {
6369 6371 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6370 6372 GCAllocForTenured);
6371 6373 }
6372 6374 // Heap region set verification
6373 6375
6374 6376 class VerifyRegionListsClosure : public HeapRegionClosure {
6375 6377 private:
6376 6378 FreeRegionList* _free_list;
6377 6379 OldRegionSet* _old_set;
6378 6380 HumongousRegionSet* _humongous_set;
6379 6381 uint _region_count;
6380 6382
6381 6383 public:
6382 6384 VerifyRegionListsClosure(OldRegionSet* old_set,
6383 6385 HumongousRegionSet* humongous_set,
6384 6386 FreeRegionList* free_list) :
6385 6387 _old_set(old_set), _humongous_set(humongous_set),
6386 6388 _free_list(free_list), _region_count(0) { }
6387 6389
6388 6390 uint region_count() { return _region_count; }
6389 6391
6390 6392 bool doHeapRegion(HeapRegion* hr) {
6391 6393 _region_count += 1;
6392 6394
6393 6395 if (hr->continuesHumongous()) {
6394 6396 return false;
6395 6397 }
6396 6398
6397 6399 if (hr->is_young()) {
6398 6400 // TODO
6399 6401 } else if (hr->startsHumongous()) {
6400 6402 _humongous_set->verify_next_region(hr);
6401 6403 } else if (hr->is_empty()) {
6402 6404 _free_list->verify_next_region(hr);
6403 6405 } else {
6404 6406 _old_set->verify_next_region(hr);
6405 6407 }
6406 6408 return false;
6407 6409 }
6408 6410 };
6409 6411
6410 6412 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6411 6413 HeapWord* bottom) {
6412 6414 HeapWord* end = bottom + HeapRegion::GrainWords;
6413 6415 MemRegion mr(bottom, end);
6414 6416 assert(_g1_reserved.contains(mr), "invariant");
6415 6417 // This might return NULL if the allocation fails
6416 6418 return new HeapRegion(hrs_index, _bot_shared, mr);
6417 6419 }
6418 6420
6419 6421 void G1CollectedHeap::verify_region_sets() {
6420 6422 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6421 6423
6422 6424 // First, check the explicit lists.
6423 6425 _free_list.verify();
6424 6426 {
6425 6427 // Given that a concurrent operation might be adding regions to
6426 6428 // the secondary free list we have to take the lock before
6427 6429 // verifying it.
6428 6430 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6429 6431 _secondary_free_list.verify();
6430 6432 }
6431 6433 _old_set.verify();
6432 6434 _humongous_set.verify();
6433 6435
6434 6436 // If a concurrent region freeing operation is in progress it will
6435 6437 // be difficult to correctly attributed any free regions we come
6436 6438 // across to the correct free list given that they might belong to
6437 6439 // one of several (free_list, secondary_free_list, any local lists,
6438 6440 // etc.). So, if that's the case we will skip the rest of the
6439 6441 // verification operation. Alternatively, waiting for the concurrent
6440 6442 // operation to complete will have a non-trivial effect on the GC's
6441 6443 // operation (no concurrent operation will last longer than the
6442 6444 // interval between two calls to verification) and it might hide
6443 6445 // any issues that we would like to catch during testing.
6444 6446 if (free_regions_coming()) {
6445 6447 return;
6446 6448 }
6447 6449
6448 6450 // Make sure we append the secondary_free_list on the free_list so
6449 6451 // that all free regions we will come across can be safely
6450 6452 // attributed to the free_list.
6451 6453 append_secondary_free_list_if_not_empty_with_lock();
6452 6454
6453 6455 // Finally, make sure that the region accounting in the lists is
6454 6456 // consistent with what we see in the heap.
6455 6457 _old_set.verify_start();
6456 6458 _humongous_set.verify_start();
6457 6459 _free_list.verify_start();
6458 6460
6459 6461 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6460 6462 heap_region_iterate(&cl);
6461 6463
6462 6464 _old_set.verify_end();
6463 6465 _humongous_set.verify_end();
6464 6466 _free_list.verify_end();
6465 6467 }
|
↓ open down ↓ |
2413 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX