Print this page
rev 2896 : 6484965: G1: piggy-back liveness accounting phase on marking
Summary: Remove the separate counting phase of concurrent marking by tracking the amount of marked bytes and the cards spanned by marked objects in marking task/worker thread local data structures, which are updated as individual objects are marked.
Reviewed-by: brutisso
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp
+++ new/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp
1 1 /*
2 2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "code/icBuffer.hpp"
27 27 #include "gc_implementation/g1/bufferingOopClosure.hpp"
28 28 #include "gc_implementation/g1/concurrentG1Refine.hpp"
29 29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
30 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
32 32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
33 33 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
34 34 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
35 35 #include "gc_implementation/g1/g1MarkSweep.hpp"
36 36 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
37 37 #include "gc_implementation/g1/g1RemSet.inline.hpp"
38 38 #include "gc_implementation/g1/heapRegionRemSet.hpp"
39 39 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
40 40 #include "gc_implementation/g1/vm_operations_g1.hpp"
41 41 #include "gc_implementation/shared/isGCActiveMark.hpp"
42 42 #include "memory/gcLocker.inline.hpp"
43 43 #include "memory/genOopClosures.inline.hpp"
44 44 #include "memory/generationSpec.hpp"
45 45 #include "memory/referenceProcessor.hpp"
46 46 #include "oops/oop.inline.hpp"
47 47 #include "oops/oop.pcgc.inline.hpp"
48 48 #include "runtime/aprofiler.hpp"
49 49 #include "runtime/vmThread.hpp"
50 50
51 51 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
52 52
53 53 // turn it on so that the contents of the young list (scan-only /
54 54 // to-be-collected) are printed at "strategic" points before / during
55 55 // / after the collection --- this is useful for debugging
56 56 #define YOUNG_LIST_VERBOSE 0
57 57 // CURRENT STATUS
58 58 // This file is under construction. Search for "FIXME".
59 59
60 60 // INVARIANTS/NOTES
61 61 //
62 62 // All allocation activity covered by the G1CollectedHeap interface is
63 63 // serialized by acquiring the HeapLock. This happens in mem_allocate
64 64 // and allocate_new_tlab, which are the "entry" points to the
65 65 // allocation code from the rest of the JVM. (Note that this does not
66 66 // apply to TLAB allocation, which is not part of this interface: it
67 67 // is done by clients of this interface.)
68 68
69 69 // Notes on implementation of parallelism in different tasks.
70 70 //
71 71 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
72 72 // The number of GC workers is passed to heap_region_par_iterate_chunked().
73 73 // It does use run_task() which sets _n_workers in the task.
74 74 // G1ParTask executes g1_process_strong_roots() ->
75 75 // SharedHeap::process_strong_roots() which calls eventuall to
76 76 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
77 77 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
78 78 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
79 79 //
80 80
81 81 // Local to this file.
82 82
83 83 class RefineCardTableEntryClosure: public CardTableEntryClosure {
84 84 SuspendibleThreadSet* _sts;
85 85 G1RemSet* _g1rs;
86 86 ConcurrentG1Refine* _cg1r;
87 87 bool _concurrent;
88 88 public:
89 89 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
90 90 G1RemSet* g1rs,
91 91 ConcurrentG1Refine* cg1r) :
92 92 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
93 93 {}
94 94 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
95 95 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
96 96 // This path is executed by the concurrent refine or mutator threads,
97 97 // concurrently, and so we do not care if card_ptr contains references
98 98 // that point into the collection set.
99 99 assert(!oops_into_cset, "should be");
100 100
101 101 if (_concurrent && _sts->should_yield()) {
102 102 // Caller will actually yield.
103 103 return false;
104 104 }
105 105 // Otherwise, we finished successfully; return true.
106 106 return true;
107 107 }
108 108 void set_concurrent(bool b) { _concurrent = b; }
109 109 };
110 110
111 111
112 112 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
113 113 int _calls;
114 114 G1CollectedHeap* _g1h;
115 115 CardTableModRefBS* _ctbs;
116 116 int _histo[256];
117 117 public:
118 118 ClearLoggedCardTableEntryClosure() :
119 119 _calls(0)
120 120 {
121 121 _g1h = G1CollectedHeap::heap();
122 122 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
123 123 for (int i = 0; i < 256; i++) _histo[i] = 0;
124 124 }
125 125 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
126 126 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
127 127 _calls++;
128 128 unsigned char* ujb = (unsigned char*)card_ptr;
129 129 int ind = (int)(*ujb);
130 130 _histo[ind]++;
131 131 *card_ptr = -1;
132 132 }
133 133 return true;
134 134 }
135 135 int calls() { return _calls; }
136 136 void print_histo() {
137 137 gclog_or_tty->print_cr("Card table value histogram:");
138 138 for (int i = 0; i < 256; i++) {
139 139 if (_histo[i] != 0) {
140 140 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
141 141 }
142 142 }
143 143 }
144 144 };
145 145
146 146 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
147 147 int _calls;
148 148 G1CollectedHeap* _g1h;
149 149 CardTableModRefBS* _ctbs;
150 150 public:
151 151 RedirtyLoggedCardTableEntryClosure() :
152 152 _calls(0)
153 153 {
154 154 _g1h = G1CollectedHeap::heap();
155 155 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
156 156 }
157 157 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
158 158 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
159 159 _calls++;
160 160 *card_ptr = 0;
161 161 }
162 162 return true;
163 163 }
164 164 int calls() { return _calls; }
165 165 };
166 166
167 167 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
168 168 public:
169 169 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
170 170 *card_ptr = CardTableModRefBS::dirty_card_val();
171 171 return true;
172 172 }
173 173 };
174 174
175 175 YoungList::YoungList(G1CollectedHeap* g1h)
176 176 : _g1h(g1h), _head(NULL),
177 177 _length(0),
178 178 _last_sampled_rs_lengths(0),
179 179 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0)
180 180 {
181 181 guarantee( check_list_empty(false), "just making sure..." );
182 182 }
183 183
184 184 void YoungList::push_region(HeapRegion *hr) {
185 185 assert(!hr->is_young(), "should not already be young");
186 186 assert(hr->get_next_young_region() == NULL, "cause it should!");
187 187
188 188 hr->set_next_young_region(_head);
189 189 _head = hr;
190 190
191 191 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
192 192 ++_length;
193 193 }
194 194
195 195 void YoungList::add_survivor_region(HeapRegion* hr) {
196 196 assert(hr->is_survivor(), "should be flagged as survivor region");
197 197 assert(hr->get_next_young_region() == NULL, "cause it should!");
198 198
199 199 hr->set_next_young_region(_survivor_head);
200 200 if (_survivor_head == NULL) {
201 201 _survivor_tail = hr;
202 202 }
203 203 _survivor_head = hr;
204 204 ++_survivor_length;
205 205 }
206 206
207 207 void YoungList::empty_list(HeapRegion* list) {
208 208 while (list != NULL) {
209 209 HeapRegion* next = list->get_next_young_region();
210 210 list->set_next_young_region(NULL);
211 211 list->uninstall_surv_rate_group();
212 212 list->set_not_young();
213 213 list = next;
214 214 }
215 215 }
216 216
217 217 void YoungList::empty_list() {
218 218 assert(check_list_well_formed(), "young list should be well formed");
219 219
220 220 empty_list(_head);
221 221 _head = NULL;
222 222 _length = 0;
223 223
224 224 empty_list(_survivor_head);
225 225 _survivor_head = NULL;
226 226 _survivor_tail = NULL;
227 227 _survivor_length = 0;
228 228
229 229 _last_sampled_rs_lengths = 0;
230 230
231 231 assert(check_list_empty(false), "just making sure...");
232 232 }
233 233
234 234 bool YoungList::check_list_well_formed() {
235 235 bool ret = true;
236 236
237 237 size_t length = 0;
238 238 HeapRegion* curr = _head;
239 239 HeapRegion* last = NULL;
240 240 while (curr != NULL) {
241 241 if (!curr->is_young()) {
242 242 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
243 243 "incorrectly tagged (y: %d, surv: %d)",
244 244 curr->bottom(), curr->end(),
245 245 curr->is_young(), curr->is_survivor());
246 246 ret = false;
247 247 }
248 248 ++length;
249 249 last = curr;
250 250 curr = curr->get_next_young_region();
251 251 }
252 252 ret = ret && (length == _length);
253 253
254 254 if (!ret) {
255 255 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
256 256 gclog_or_tty->print_cr("### list has %d entries, _length is %d",
257 257 length, _length);
258 258 }
259 259
260 260 return ret;
261 261 }
262 262
263 263 bool YoungList::check_list_empty(bool check_sample) {
264 264 bool ret = true;
265 265
266 266 if (_length != 0) {
267 267 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %d",
268 268 _length);
269 269 ret = false;
270 270 }
271 271 if (check_sample && _last_sampled_rs_lengths != 0) {
272 272 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
273 273 ret = false;
274 274 }
275 275 if (_head != NULL) {
276 276 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
277 277 ret = false;
278 278 }
279 279 if (!ret) {
280 280 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
281 281 }
282 282
283 283 return ret;
284 284 }
285 285
286 286 void
287 287 YoungList::rs_length_sampling_init() {
288 288 _sampled_rs_lengths = 0;
289 289 _curr = _head;
290 290 }
291 291
292 292 bool
293 293 YoungList::rs_length_sampling_more() {
294 294 return _curr != NULL;
295 295 }
296 296
297 297 void
298 298 YoungList::rs_length_sampling_next() {
299 299 assert( _curr != NULL, "invariant" );
300 300 size_t rs_length = _curr->rem_set()->occupied();
301 301
302 302 _sampled_rs_lengths += rs_length;
303 303
304 304 // The current region may not yet have been added to the
305 305 // incremental collection set (it gets added when it is
306 306 // retired as the current allocation region).
307 307 if (_curr->in_collection_set()) {
308 308 // Update the collection set policy information for this region
309 309 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
310 310 }
311 311
312 312 _curr = _curr->get_next_young_region();
313 313 if (_curr == NULL) {
314 314 _last_sampled_rs_lengths = _sampled_rs_lengths;
315 315 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
316 316 }
317 317 }
318 318
319 319 void
320 320 YoungList::reset_auxilary_lists() {
321 321 guarantee( is_empty(), "young list should be empty" );
322 322 assert(check_list_well_formed(), "young list should be well formed");
323 323
324 324 // Add survivor regions to SurvRateGroup.
325 325 _g1h->g1_policy()->note_start_adding_survivor_regions();
326 326 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
327 327
328 328 int young_index_in_cset = 0;
329 329 for (HeapRegion* curr = _survivor_head;
330 330 curr != NULL;
331 331 curr = curr->get_next_young_region()) {
332 332 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
333 333
334 334 // The region is a non-empty survivor so let's add it to
335 335 // the incremental collection set for the next evacuation
336 336 // pause.
337 337 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
338 338 young_index_in_cset += 1;
339 339 }
340 340 assert((size_t) young_index_in_cset == _survivor_length,
341 341 "post-condition");
342 342 _g1h->g1_policy()->note_stop_adding_survivor_regions();
343 343
344 344 _head = _survivor_head;
345 345 _length = _survivor_length;
346 346 if (_survivor_head != NULL) {
347 347 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 348 assert(_survivor_length > 0, "invariant");
349 349 _survivor_tail->set_next_young_region(NULL);
350 350 }
351 351
352 352 // Don't clear the survivor list handles until the start of
353 353 // the next evacuation pause - we need it in order to re-tag
354 354 // the survivor regions from this evacuation pause as 'young'
355 355 // at the start of the next.
356 356
357 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
358 358
359 359 assert(check_list_well_formed(), "young list should be well formed");
360 360 }
361 361
362 362 void YoungList::print() {
363 363 HeapRegion* lists[] = {_head, _survivor_head};
364 364 const char* names[] = {"YOUNG", "SURVIVOR"};
365 365
366 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 368 HeapRegion *curr = lists[list];
369 369 if (curr == NULL)
370 370 gclog_or_tty->print_cr(" empty");
371 371 while (curr != NULL) {
372 372 gclog_or_tty->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
373 373 "age: %4d, y: %d, surv: %d",
374 374 curr->bottom(), curr->end(),
375 375 curr->top(),
376 376 curr->prev_top_at_mark_start(),
377 377 curr->next_top_at_mark_start(),
378 378 curr->top_at_conc_mark_count(),
379 379 curr->age_in_surv_rate_group_cond(),
380 380 curr->is_young(),
381 381 curr->is_survivor());
382 382 curr = curr->get_next_young_region();
383 383 }
384 384 }
385 385
386 386 gclog_or_tty->print_cr("");
387 387 }
388 388
389 389 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
390 390 {
391 391 // Claim the right to put the region on the dirty cards region list
392 392 // by installing a self pointer.
393 393 HeapRegion* next = hr->get_next_dirty_cards_region();
394 394 if (next == NULL) {
395 395 HeapRegion* res = (HeapRegion*)
396 396 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
397 397 NULL);
398 398 if (res == NULL) {
399 399 HeapRegion* head;
400 400 do {
401 401 // Put the region to the dirty cards region list.
402 402 head = _dirty_cards_region_list;
403 403 next = (HeapRegion*)
404 404 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
405 405 if (next == head) {
406 406 assert(hr->get_next_dirty_cards_region() == hr,
407 407 "hr->get_next_dirty_cards_region() != hr");
408 408 if (next == NULL) {
409 409 // The last region in the list points to itself.
410 410 hr->set_next_dirty_cards_region(hr);
411 411 } else {
412 412 hr->set_next_dirty_cards_region(next);
413 413 }
414 414 }
415 415 } while (next != head);
416 416 }
417 417 }
418 418 }
419 419
420 420 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
421 421 {
422 422 HeapRegion* head;
423 423 HeapRegion* hr;
424 424 do {
425 425 head = _dirty_cards_region_list;
426 426 if (head == NULL) {
427 427 return NULL;
428 428 }
429 429 HeapRegion* new_head = head->get_next_dirty_cards_region();
430 430 if (head == new_head) {
431 431 // The last region.
432 432 new_head = NULL;
433 433 }
434 434 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
435 435 head);
436 436 } while (hr != head);
437 437 assert(hr != NULL, "invariant");
438 438 hr->set_next_dirty_cards_region(NULL);
439 439 return hr;
440 440 }
441 441
442 442 void G1CollectedHeap::stop_conc_gc_threads() {
443 443 _cg1r->stop();
444 444 _cmThread->stop();
445 445 }
446 446
447 447 #ifdef ASSERT
448 448 // A region is added to the collection set as it is retired
449 449 // so an address p can point to a region which will be in the
450 450 // collection set but has not yet been retired. This method
451 451 // therefore is only accurate during a GC pause after all
452 452 // regions have been retired. It is used for debugging
453 453 // to check if an nmethod has references to objects that can
454 454 // be move during a partial collection. Though it can be
455 455 // inaccurate, it is sufficient for G1 because the conservative
456 456 // implementation of is_scavengable() for G1 will indicate that
457 457 // all nmethods must be scanned during a partial collection.
458 458 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
459 459 HeapRegion* hr = heap_region_containing(p);
460 460 return hr != NULL && hr->in_collection_set();
461 461 }
462 462 #endif
463 463
464 464 // Returns true if the reference points to an object that
465 465 // can move in an incremental collecction.
466 466 bool G1CollectedHeap::is_scavengable(const void* p) {
467 467 G1CollectedHeap* g1h = G1CollectedHeap::heap();
468 468 G1CollectorPolicy* g1p = g1h->g1_policy();
469 469 HeapRegion* hr = heap_region_containing(p);
470 470 if (hr == NULL) {
471 471 // perm gen (or null)
472 472 return false;
473 473 } else {
474 474 return !hr->isHumongous();
475 475 }
476 476 }
477 477
478 478 void G1CollectedHeap::check_ct_logs_at_safepoint() {
479 479 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
480 480 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
481 481
482 482 // Count the dirty cards at the start.
483 483 CountNonCleanMemRegionClosure count1(this);
484 484 ct_bs->mod_card_iterate(&count1);
485 485 int orig_count = count1.n();
486 486
487 487 // First clear the logged cards.
488 488 ClearLoggedCardTableEntryClosure clear;
489 489 dcqs.set_closure(&clear);
490 490 dcqs.apply_closure_to_all_completed_buffers();
491 491 dcqs.iterate_closure_all_threads(false);
492 492 clear.print_histo();
493 493
494 494 // Now ensure that there's no dirty cards.
495 495 CountNonCleanMemRegionClosure count2(this);
496 496 ct_bs->mod_card_iterate(&count2);
497 497 if (count2.n() != 0) {
498 498 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
499 499 count2.n(), orig_count);
500 500 }
501 501 guarantee(count2.n() == 0, "Card table should be clean.");
502 502
503 503 RedirtyLoggedCardTableEntryClosure redirty;
504 504 JavaThread::dirty_card_queue_set().set_closure(&redirty);
505 505 dcqs.apply_closure_to_all_completed_buffers();
506 506 dcqs.iterate_closure_all_threads(false);
507 507 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
508 508 clear.calls(), orig_count);
509 509 guarantee(redirty.calls() == clear.calls(),
510 510 "Or else mechanism is broken.");
511 511
512 512 CountNonCleanMemRegionClosure count3(this);
513 513 ct_bs->mod_card_iterate(&count3);
514 514 if (count3.n() != orig_count) {
515 515 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
516 516 orig_count, count3.n());
517 517 guarantee(count3.n() >= orig_count, "Should have restored them all.");
518 518 }
519 519
520 520 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
521 521 }
522 522
523 523 // Private class members.
524 524
525 525 G1CollectedHeap* G1CollectedHeap::_g1h;
526 526
527 527 // Private methods.
528 528
529 529 HeapRegion*
530 530 G1CollectedHeap::new_region_try_secondary_free_list() {
531 531 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
532 532 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
533 533 if (!_secondary_free_list.is_empty()) {
534 534 if (G1ConcRegionFreeingVerbose) {
535 535 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
536 536 "secondary_free_list has "SIZE_FORMAT" entries",
537 537 _secondary_free_list.length());
538 538 }
539 539 // It looks as if there are free regions available on the
540 540 // secondary_free_list. Let's move them to the free_list and try
541 541 // again to allocate from it.
542 542 append_secondary_free_list();
543 543
544 544 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
545 545 "empty we should have moved at least one entry to the free_list");
546 546 HeapRegion* res = _free_list.remove_head();
547 547 if (G1ConcRegionFreeingVerbose) {
548 548 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
549 549 "allocated "HR_FORMAT" from secondary_free_list",
550 550 HR_FORMAT_PARAMS(res));
551 551 }
552 552 return res;
553 553 }
554 554
555 555 // Wait here until we get notifed either when (a) there are no
556 556 // more free regions coming or (b) some regions have been moved on
557 557 // the secondary_free_list.
558 558 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
559 559 }
560 560
561 561 if (G1ConcRegionFreeingVerbose) {
562 562 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
563 563 "could not allocate from secondary_free_list");
564 564 }
565 565 return NULL;
566 566 }
567 567
568 568 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
569 569 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
570 570 "the only time we use this to allocate a humongous region is "
571 571 "when we are allocating a single humongous region");
572 572
573 573 HeapRegion* res;
574 574 if (G1StressConcRegionFreeing) {
575 575 if (!_secondary_free_list.is_empty()) {
576 576 if (G1ConcRegionFreeingVerbose) {
577 577 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
578 578 "forced to look at the secondary_free_list");
579 579 }
580 580 res = new_region_try_secondary_free_list();
581 581 if (res != NULL) {
582 582 return res;
583 583 }
584 584 }
585 585 }
586 586 res = _free_list.remove_head_or_null();
587 587 if (res == NULL) {
588 588 if (G1ConcRegionFreeingVerbose) {
589 589 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
590 590 "res == NULL, trying the secondary_free_list");
591 591 }
592 592 res = new_region_try_secondary_free_list();
593 593 }
594 594 if (res == NULL && do_expand) {
595 595 ergo_verbose1(ErgoHeapSizing,
596 596 "attempt heap expansion",
597 597 ergo_format_reason("region allocation request failed")
598 598 ergo_format_byte("allocation request"),
599 599 word_size * HeapWordSize);
600 600 if (expand(word_size * HeapWordSize)) {
601 601 // Even though the heap was expanded, it might not have reached
602 602 // the desired size. So, we cannot assume that the allocation
603 603 // will succeed.
604 604 res = _free_list.remove_head_or_null();
605 605 }
606 606 }
607 607 return res;
608 608 }
609 609
610 610 size_t G1CollectedHeap::humongous_obj_allocate_find_first(size_t num_regions,
611 611 size_t word_size) {
612 612 assert(isHumongous(word_size), "word_size should be humongous");
613 613 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
614 614
615 615 size_t first = G1_NULL_HRS_INDEX;
616 616 if (num_regions == 1) {
617 617 // Only one region to allocate, no need to go through the slower
618 618 // path. The caller will attempt the expasion if this fails, so
619 619 // let's not try to expand here too.
620 620 HeapRegion* hr = new_region(word_size, false /* do_expand */);
621 621 if (hr != NULL) {
622 622 first = hr->hrs_index();
623 623 } else {
624 624 first = G1_NULL_HRS_INDEX;
625 625 }
626 626 } else {
627 627 // We can't allocate humongous regions while cleanupComplete() is
628 628 // running, since some of the regions we find to be empty might not
629 629 // yet be added to the free list and it is not straightforward to
630 630 // know which list they are on so that we can remove them. Note
631 631 // that we only need to do this if we need to allocate more than
632 632 // one region to satisfy the current humongous allocation
633 633 // request. If we are only allocating one region we use the common
634 634 // region allocation code (see above).
635 635 wait_while_free_regions_coming();
636 636 append_secondary_free_list_if_not_empty_with_lock();
637 637
638 638 if (free_regions() >= num_regions) {
639 639 first = _hrs.find_contiguous(num_regions);
640 640 if (first != G1_NULL_HRS_INDEX) {
641 641 for (size_t i = first; i < first + num_regions; ++i) {
642 642 HeapRegion* hr = region_at(i);
643 643 assert(hr->is_empty(), "sanity");
644 644 assert(is_on_master_free_list(hr), "sanity");
645 645 hr->set_pending_removal(true);
646 646 }
647 647 _free_list.remove_all_pending(num_regions);
648 648 }
649 649 }
650 650 }
651 651 return first;
652 652 }
653 653
654 654 HeapWord*
655 655 G1CollectedHeap::humongous_obj_allocate_initialize_regions(size_t first,
656 656 size_t num_regions,
657 657 size_t word_size) {
658 658 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
659 659 assert(isHumongous(word_size), "word_size should be humongous");
660 660 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
661 661
662 662 // Index of last region in the series + 1.
663 663 size_t last = first + num_regions;
664 664
665 665 // We need to initialize the region(s) we just discovered. This is
666 666 // a bit tricky given that it can happen concurrently with
667 667 // refinement threads refining cards on these regions and
668 668 // potentially wanting to refine the BOT as they are scanning
669 669 // those cards (this can happen shortly after a cleanup; see CR
670 670 // 6991377). So we have to set up the region(s) carefully and in
671 671 // a specific order.
672 672
673 673 // The word size sum of all the regions we will allocate.
674 674 size_t word_size_sum = num_regions * HeapRegion::GrainWords;
675 675 assert(word_size <= word_size_sum, "sanity");
676 676
677 677 // This will be the "starts humongous" region.
678 678 HeapRegion* first_hr = region_at(first);
679 679 // The header of the new object will be placed at the bottom of
680 680 // the first region.
681 681 HeapWord* new_obj = first_hr->bottom();
682 682 // This will be the new end of the first region in the series that
683 683 // should also match the end of the last region in the seriers.
684 684 HeapWord* new_end = new_obj + word_size_sum;
685 685 // This will be the new top of the first region that will reflect
686 686 // this allocation.
687 687 HeapWord* new_top = new_obj + word_size;
688 688
689 689 // First, we need to zero the header of the space that we will be
690 690 // allocating. When we update top further down, some refinement
691 691 // threads might try to scan the region. By zeroing the header we
692 692 // ensure that any thread that will try to scan the region will
693 693 // come across the zero klass word and bail out.
694 694 //
695 695 // NOTE: It would not have been correct to have used
696 696 // CollectedHeap::fill_with_object() and make the space look like
697 697 // an int array. The thread that is doing the allocation will
698 698 // later update the object header to a potentially different array
699 699 // type and, for a very short period of time, the klass and length
700 700 // fields will be inconsistent. This could cause a refinement
701 701 // thread to calculate the object size incorrectly.
702 702 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
703 703
704 704 // We will set up the first region as "starts humongous". This
705 705 // will also update the BOT covering all the regions to reflect
706 706 // that there is a single object that starts at the bottom of the
707 707 // first region.
708 708 first_hr->set_startsHumongous(new_top, new_end);
709 709
710 710 // Then, if there are any, we will set up the "continues
711 711 // humongous" regions.
712 712 HeapRegion* hr = NULL;
713 713 for (size_t i = first + 1; i < last; ++i) {
714 714 hr = region_at(i);
715 715 hr->set_continuesHumongous(first_hr);
716 716 }
717 717 // If we have "continues humongous" regions (hr != NULL), then the
718 718 // end of the last one should match new_end.
719 719 assert(hr == NULL || hr->end() == new_end, "sanity");
720 720
721 721 // Up to this point no concurrent thread would have been able to
722 722 // do any scanning on any region in this series. All the top
723 723 // fields still point to bottom, so the intersection between
724 724 // [bottom,top] and [card_start,card_end] will be empty. Before we
725 725 // update the top fields, we'll do a storestore to make sure that
726 726 // no thread sees the update to top before the zeroing of the
727 727 // object header and the BOT initialization.
728 728 OrderAccess::storestore();
729 729
730 730 // Now that the BOT and the object header have been initialized,
731 731 // we can update top of the "starts humongous" region.
732 732 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
733 733 "new_top should be in this region");
734 734 first_hr->set_top(new_top);
735 735 if (_hr_printer.is_active()) {
736 736 HeapWord* bottom = first_hr->bottom();
737 737 HeapWord* end = first_hr->orig_end();
738 738 if ((first + 1) == last) {
739 739 // the series has a single humongous region
740 740 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
741 741 } else {
742 742 // the series has more than one humongous regions
743 743 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
744 744 }
745 745 }
746 746
747 747 // Now, we will update the top fields of the "continues humongous"
748 748 // regions. The reason we need to do this is that, otherwise,
749 749 // these regions would look empty and this will confuse parts of
750 750 // G1. For example, the code that looks for a consecutive number
751 751 // of empty regions will consider them empty and try to
752 752 // re-allocate them. We can extend is_empty() to also include
753 753 // !continuesHumongous(), but it is easier to just update the top
754 754 // fields here. The way we set top for all regions (i.e., top ==
755 755 // end for all regions but the last one, top == new_top for the
756 756 // last one) is actually used when we will free up the humongous
757 757 // region in free_humongous_region().
758 758 hr = NULL;
759 759 for (size_t i = first + 1; i < last; ++i) {
760 760 hr = region_at(i);
761 761 if ((i + 1) == last) {
762 762 // last continues humongous region
763 763 assert(hr->bottom() < new_top && new_top <= hr->end(),
764 764 "new_top should fall on this region");
765 765 hr->set_top(new_top);
766 766 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
767 767 } else {
768 768 // not last one
769 769 assert(new_top > hr->end(), "new_top should be above this region");
770 770 hr->set_top(hr->end());
771 771 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
772 772 }
773 773 }
774 774 // If we have continues humongous regions (hr != NULL), then the
775 775 // end of the last one should match new_end and its top should
776 776 // match new_top.
777 777 assert(hr == NULL ||
778 778 (hr->end() == new_end && hr->top() == new_top), "sanity");
779 779
780 780 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
781 781 _summary_bytes_used += first_hr->used();
782 782 _humongous_set.add(first_hr);
783 783
784 784 return new_obj;
785 785 }
786 786
787 787 // If could fit into free regions w/o expansion, try.
788 788 // Otherwise, if can expand, do so.
789 789 // Otherwise, if using ex regions might help, try with ex given back.
790 790 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
791 791 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
792 792
793 793 verify_region_sets_optional();
794 794
795 795 size_t num_regions =
796 796 round_to(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords;
797 797 size_t x_size = expansion_regions();
798 798 size_t fs = _hrs.free_suffix();
799 799 size_t first = humongous_obj_allocate_find_first(num_regions, word_size);
800 800 if (first == G1_NULL_HRS_INDEX) {
801 801 // The only thing we can do now is attempt expansion.
802 802 if (fs + x_size >= num_regions) {
803 803 // If the number of regions we're trying to allocate for this
804 804 // object is at most the number of regions in the free suffix,
805 805 // then the call to humongous_obj_allocate_find_first() above
806 806 // should have succeeded and we wouldn't be here.
807 807 //
808 808 // We should only be trying to expand when the free suffix is
809 809 // not sufficient for the object _and_ we have some expansion
810 810 // room available.
811 811 assert(num_regions > fs, "earlier allocation should have succeeded");
812 812
813 813 ergo_verbose1(ErgoHeapSizing,
814 814 "attempt heap expansion",
815 815 ergo_format_reason("humongous allocation request failed")
816 816 ergo_format_byte("allocation request"),
817 817 word_size * HeapWordSize);
818 818 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
819 819 // Even though the heap was expanded, it might not have
820 820 // reached the desired size. So, we cannot assume that the
821 821 // allocation will succeed.
822 822 first = humongous_obj_allocate_find_first(num_regions, word_size);
823 823 }
824 824 }
825 825 }
826 826
827 827 HeapWord* result = NULL;
828 828 if (first != G1_NULL_HRS_INDEX) {
829 829 result =
830 830 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
831 831 assert(result != NULL, "it should always return a valid result");
832 832
833 833 // A successful humongous object allocation changes the used space
834 834 // information of the old generation so we need to recalculate the
835 835 // sizes and update the jstat counters here.
836 836 g1mm()->update_sizes();
837 837 }
838 838
839 839 verify_region_sets_optional();
840 840
841 841 return result;
842 842 }
843 843
844 844 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
845 845 assert_heap_not_locked_and_not_at_safepoint();
846 846 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
847 847
848 848 unsigned int dummy_gc_count_before;
849 849 return attempt_allocation(word_size, &dummy_gc_count_before);
850 850 }
851 851
852 852 HeapWord*
853 853 G1CollectedHeap::mem_allocate(size_t word_size,
854 854 bool* gc_overhead_limit_was_exceeded) {
855 855 assert_heap_not_locked_and_not_at_safepoint();
856 856
857 857 // Loop until the allocation is satisified, or unsatisfied after GC.
858 858 for (int try_count = 1; /* we'll return */; try_count += 1) {
859 859 unsigned int gc_count_before;
860 860
861 861 HeapWord* result = NULL;
862 862 if (!isHumongous(word_size)) {
863 863 result = attempt_allocation(word_size, &gc_count_before);
864 864 } else {
865 865 result = attempt_allocation_humongous(word_size, &gc_count_before);
866 866 }
867 867 if (result != NULL) {
868 868 return result;
869 869 }
870 870
871 871 // Create the garbage collection operation...
872 872 VM_G1CollectForAllocation op(gc_count_before, word_size);
873 873 // ...and get the VM thread to execute it.
874 874 VMThread::execute(&op);
875 875
876 876 if (op.prologue_succeeded() && op.pause_succeeded()) {
877 877 // If the operation was successful we'll return the result even
878 878 // if it is NULL. If the allocation attempt failed immediately
879 879 // after a Full GC, it's unlikely we'll be able to allocate now.
880 880 HeapWord* result = op.result();
881 881 if (result != NULL && !isHumongous(word_size)) {
882 882 // Allocations that take place on VM operations do not do any
883 883 // card dirtying and we have to do it here. We only have to do
884 884 // this for non-humongous allocations, though.
885 885 dirty_young_block(result, word_size);
886 886 }
887 887 return result;
888 888 } else {
889 889 assert(op.result() == NULL,
890 890 "the result should be NULL if the VM op did not succeed");
891 891 }
892 892
893 893 // Give a warning if we seem to be looping forever.
894 894 if ((QueuedAllocationWarningCount > 0) &&
895 895 (try_count % QueuedAllocationWarningCount == 0)) {
896 896 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
897 897 }
898 898 }
899 899
900 900 ShouldNotReachHere();
901 901 return NULL;
902 902 }
903 903
904 904 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
905 905 unsigned int *gc_count_before_ret) {
906 906 // Make sure you read the note in attempt_allocation_humongous().
907 907
908 908 assert_heap_not_locked_and_not_at_safepoint();
909 909 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
910 910 "be called for humongous allocation requests");
911 911
912 912 // We should only get here after the first-level allocation attempt
913 913 // (attempt_allocation()) failed to allocate.
914 914
915 915 // We will loop until a) we manage to successfully perform the
916 916 // allocation or b) we successfully schedule a collection which
917 917 // fails to perform the allocation. b) is the only case when we'll
918 918 // return NULL.
919 919 HeapWord* result = NULL;
920 920 for (int try_count = 1; /* we'll return */; try_count += 1) {
921 921 bool should_try_gc;
922 922 unsigned int gc_count_before;
923 923
924 924 {
925 925 MutexLockerEx x(Heap_lock);
926 926
927 927 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
928 928 false /* bot_updates */);
929 929 if (result != NULL) {
930 930 return result;
931 931 }
932 932
933 933 // If we reach here, attempt_allocation_locked() above failed to
934 934 // allocate a new region. So the mutator alloc region should be NULL.
935 935 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
936 936
937 937 if (GC_locker::is_active_and_needs_gc()) {
938 938 if (g1_policy()->can_expand_young_list()) {
939 939 // No need for an ergo verbose message here,
940 940 // can_expand_young_list() does this when it returns true.
941 941 result = _mutator_alloc_region.attempt_allocation_force(word_size,
942 942 false /* bot_updates */);
943 943 if (result != NULL) {
944 944 return result;
945 945 }
946 946 }
947 947 should_try_gc = false;
948 948 } else {
949 949 // Read the GC count while still holding the Heap_lock.
950 950 gc_count_before = SharedHeap::heap()->total_collections();
951 951 should_try_gc = true;
952 952 }
953 953 }
954 954
955 955 if (should_try_gc) {
956 956 bool succeeded;
957 957 result = do_collection_pause(word_size, gc_count_before, &succeeded);
958 958 if (result != NULL) {
959 959 assert(succeeded, "only way to get back a non-NULL result");
960 960 return result;
961 961 }
962 962
963 963 if (succeeded) {
964 964 // If we get here we successfully scheduled a collection which
965 965 // failed to allocate. No point in trying to allocate
966 966 // further. We'll just return NULL.
967 967 MutexLockerEx x(Heap_lock);
968 968 *gc_count_before_ret = SharedHeap::heap()->total_collections();
969 969 return NULL;
970 970 }
971 971 } else {
972 972 GC_locker::stall_until_clear();
973 973 }
974 974
975 975 // We can reach here if we were unsuccessul in scheduling a
976 976 // collection (because another thread beat us to it) or if we were
977 977 // stalled due to the GC locker. In either can we should retry the
978 978 // allocation attempt in case another thread successfully
979 979 // performed a collection and reclaimed enough space. We do the
980 980 // first attempt (without holding the Heap_lock) here and the
981 981 // follow-on attempt will be at the start of the next loop
982 982 // iteration (after taking the Heap_lock).
983 983 result = _mutator_alloc_region.attempt_allocation(word_size,
984 984 false /* bot_updates */);
985 985 if (result != NULL ){
986 986 return result;
987 987 }
988 988
989 989 // Give a warning if we seem to be looping forever.
990 990 if ((QueuedAllocationWarningCount > 0) &&
991 991 (try_count % QueuedAllocationWarningCount == 0)) {
992 992 warning("G1CollectedHeap::attempt_allocation_slow() "
993 993 "retries %d times", try_count);
994 994 }
995 995 }
996 996
997 997 ShouldNotReachHere();
998 998 return NULL;
999 999 }
1000 1000
1001 1001 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1002 1002 unsigned int * gc_count_before_ret) {
1003 1003 // The structure of this method has a lot of similarities to
1004 1004 // attempt_allocation_slow(). The reason these two were not merged
1005 1005 // into a single one is that such a method would require several "if
1006 1006 // allocation is not humongous do this, otherwise do that"
1007 1007 // conditional paths which would obscure its flow. In fact, an early
1008 1008 // version of this code did use a unified method which was harder to
1009 1009 // follow and, as a result, it had subtle bugs that were hard to
1010 1010 // track down. So keeping these two methods separate allows each to
1011 1011 // be more readable. It will be good to keep these two in sync as
1012 1012 // much as possible.
1013 1013
1014 1014 assert_heap_not_locked_and_not_at_safepoint();
1015 1015 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1016 1016 "should only be called for humongous allocations");
1017 1017
1018 1018 // We will loop until a) we manage to successfully perform the
1019 1019 // allocation or b) we successfully schedule a collection which
1020 1020 // fails to perform the allocation. b) is the only case when we'll
1021 1021 // return NULL.
1022 1022 HeapWord* result = NULL;
1023 1023 for (int try_count = 1; /* we'll return */; try_count += 1) {
1024 1024 bool should_try_gc;
1025 1025 unsigned int gc_count_before;
1026 1026
1027 1027 {
1028 1028 MutexLockerEx x(Heap_lock);
1029 1029
1030 1030 // Given that humongous objects are not allocated in young
1031 1031 // regions, we'll first try to do the allocation without doing a
1032 1032 // collection hoping that there's enough space in the heap.
1033 1033 result = humongous_obj_allocate(word_size);
1034 1034 if (result != NULL) {
1035 1035 return result;
1036 1036 }
1037 1037
1038 1038 if (GC_locker::is_active_and_needs_gc()) {
1039 1039 should_try_gc = false;
1040 1040 } else {
1041 1041 // Read the GC count while still holding the Heap_lock.
1042 1042 gc_count_before = SharedHeap::heap()->total_collections();
1043 1043 should_try_gc = true;
1044 1044 }
1045 1045 }
1046 1046
1047 1047 if (should_try_gc) {
1048 1048 // If we failed to allocate the humongous object, we should try to
1049 1049 // do a collection pause (if we're allowed) in case it reclaims
1050 1050 // enough space for the allocation to succeed after the pause.
1051 1051
1052 1052 bool succeeded;
1053 1053 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1054 1054 if (result != NULL) {
1055 1055 assert(succeeded, "only way to get back a non-NULL result");
1056 1056 return result;
1057 1057 }
1058 1058
1059 1059 if (succeeded) {
1060 1060 // If we get here we successfully scheduled a collection which
1061 1061 // failed to allocate. No point in trying to allocate
1062 1062 // further. We'll just return NULL.
1063 1063 MutexLockerEx x(Heap_lock);
1064 1064 *gc_count_before_ret = SharedHeap::heap()->total_collections();
1065 1065 return NULL;
1066 1066 }
1067 1067 } else {
1068 1068 GC_locker::stall_until_clear();
1069 1069 }
1070 1070
1071 1071 // We can reach here if we were unsuccessul in scheduling a
1072 1072 // collection (because another thread beat us to it) or if we were
1073 1073 // stalled due to the GC locker. In either can we should retry the
1074 1074 // allocation attempt in case another thread successfully
1075 1075 // performed a collection and reclaimed enough space. Give a
1076 1076 // warning if we seem to be looping forever.
1077 1077
1078 1078 if ((QueuedAllocationWarningCount > 0) &&
1079 1079 (try_count % QueuedAllocationWarningCount == 0)) {
1080 1080 warning("G1CollectedHeap::attempt_allocation_humongous() "
1081 1081 "retries %d times", try_count);
1082 1082 }
1083 1083 }
1084 1084
1085 1085 ShouldNotReachHere();
1086 1086 return NULL;
1087 1087 }
1088 1088
1089 1089 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1090 1090 bool expect_null_mutator_alloc_region) {
1091 1091 assert_at_safepoint(true /* should_be_vm_thread */);
1092 1092 assert(_mutator_alloc_region.get() == NULL ||
1093 1093 !expect_null_mutator_alloc_region,
1094 1094 "the current alloc region was unexpectedly found to be non-NULL");
1095 1095
1096 1096 if (!isHumongous(word_size)) {
1097 1097 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1098 1098 false /* bot_updates */);
1099 1099 } else {
1100 1100 return humongous_obj_allocate(word_size);
1101 1101 }
1102 1102
1103 1103 ShouldNotReachHere();
1104 1104 }
1105 1105
1106 1106 class PostMCRemSetClearClosure: public HeapRegionClosure {
1107 1107 ModRefBarrierSet* _mr_bs;
1108 1108 public:
1109 1109 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1110 1110 bool doHeapRegion(HeapRegion* r) {
1111 1111 r->reset_gc_time_stamp();
1112 1112 if (r->continuesHumongous())
1113 1113 return false;
1114 1114 HeapRegionRemSet* hrrs = r->rem_set();
1115 1115 if (hrrs != NULL) hrrs->clear();
1116 1116 // You might think here that we could clear just the cards
1117 1117 // corresponding to the used region. But no: if we leave a dirty card
1118 1118 // in a region we might allocate into, then it would prevent that card
1119 1119 // from being enqueued, and cause it to be missed.
1120 1120 // Re: the performance cost: we shouldn't be doing full GC anyway!
1121 1121 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1122 1122 return false;
1123 1123 }
1124 1124 };
1125 1125
1126 1126
1127 1127 class PostMCRemSetInvalidateClosure: public HeapRegionClosure {
1128 1128 ModRefBarrierSet* _mr_bs;
1129 1129 public:
1130 1130 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {}
1131 1131 bool doHeapRegion(HeapRegion* r) {
1132 1132 if (r->continuesHumongous()) return false;
1133 1133 if (r->used_region().word_size() != 0) {
1134 1134 _mr_bs->invalidate(r->used_region(), true /*whole heap*/);
1135 1135 }
1136 1136 return false;
1137 1137 }
1138 1138 };
1139 1139
1140 1140 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1141 1141 G1CollectedHeap* _g1h;
1142 1142 UpdateRSOopClosure _cl;
1143 1143 int _worker_i;
1144 1144 public:
1145 1145 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1146 1146 _cl(g1->g1_rem_set(), worker_i),
1147 1147 _worker_i(worker_i),
1148 1148 _g1h(g1)
1149 1149 { }
1150 1150
1151 1151 bool doHeapRegion(HeapRegion* r) {
1152 1152 if (!r->continuesHumongous()) {
1153 1153 _cl.set_from(r);
1154 1154 r->oop_iterate(&_cl);
1155 1155 }
1156 1156 return false;
1157 1157 }
1158 1158 };
1159 1159
1160 1160 class ParRebuildRSTask: public AbstractGangTask {
1161 1161 G1CollectedHeap* _g1;
1162 1162 public:
1163 1163 ParRebuildRSTask(G1CollectedHeap* g1)
1164 1164 : AbstractGangTask("ParRebuildRSTask"),
1165 1165 _g1(g1)
1166 1166 { }
1167 1167
1168 1168 void work(int i) {
1169 1169 RebuildRSOutOfRegionClosure rebuild_rs(_g1, i);
1170 1170 _g1->heap_region_par_iterate_chunked(&rebuild_rs, i,
1171 1171 _g1->workers()->active_workers(),
1172 1172 HeapRegion::RebuildRSClaimValue);
1173 1173 }
1174 1174 };
1175 1175
1176 1176 class PostCompactionPrinterClosure: public HeapRegionClosure {
1177 1177 private:
1178 1178 G1HRPrinter* _hr_printer;
1179 1179 public:
1180 1180 bool doHeapRegion(HeapRegion* hr) {
1181 1181 assert(!hr->is_young(), "not expecting to find young regions");
1182 1182 // We only generate output for non-empty regions.
1183 1183 if (!hr->is_empty()) {
1184 1184 if (!hr->isHumongous()) {
1185 1185 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1186 1186 } else if (hr->startsHumongous()) {
1187 1187 if (hr->capacity() == HeapRegion::GrainBytes) {
1188 1188 // single humongous region
1189 1189 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1190 1190 } else {
1191 1191 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1192 1192 }
1193 1193 } else {
1194 1194 assert(hr->continuesHumongous(), "only way to get here");
1195 1195 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1196 1196 }
1197 1197 }
1198 1198 return false;
1199 1199 }
1200 1200
1201 1201 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1202 1202 : _hr_printer(hr_printer) { }
1203 1203 };
1204 1204
1205 1205 bool G1CollectedHeap::do_collection(bool explicit_gc,
1206 1206 bool clear_all_soft_refs,
1207 1207 size_t word_size) {
1208 1208 assert_at_safepoint(true /* should_be_vm_thread */);
1209 1209
1210 1210 if (GC_locker::check_active_before_gc()) {
1211 1211 return false;
1212 1212 }
1213 1213
1214 1214 SvcGCMarker sgcm(SvcGCMarker::FULL);
1215 1215 ResourceMark rm;
1216 1216
1217 1217 if (PrintHeapAtGC) {
1218 1218 Universe::print_heap_before_gc();
1219 1219 }
1220 1220
1221 1221 HRSPhaseSetter x(HRSPhaseFullGC);
1222 1222 verify_region_sets_optional();
1223 1223
1224 1224 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1225 1225 collector_policy()->should_clear_all_soft_refs();
1226 1226
1227 1227 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1228 1228
1229 1229 {
1230 1230 IsGCActiveMark x;
1231 1231
1232 1232 // Timing
1233 1233 bool system_gc = (gc_cause() == GCCause::_java_lang_system_gc);
1234 1234 assert(!system_gc || explicit_gc, "invariant");
1235 1235 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1236 1236 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1237 1237 TraceTime t(system_gc ? "Full GC (System.gc())" : "Full GC",
1238 1238 PrintGC, true, gclog_or_tty);
1239 1239
1240 1240 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1241 1241 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1242 1242
1243 1243 double start = os::elapsedTime();
1244 1244 g1_policy()->record_full_collection_start();
1245 1245
1246 1246 wait_while_free_regions_coming();
1247 1247 append_secondary_free_list_if_not_empty_with_lock();
1248 1248
1249 1249 gc_prologue(true);
1250 1250 increment_total_collections(true /* full gc */);
1251 1251
1252 1252 size_t g1h_prev_used = used();
1253 1253 assert(used() == recalculate_used(), "Should be equal");
1254 1254
1255 1255 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
1256 1256 HandleMark hm; // Discard invalid handles created during verification
1257 1257 gclog_or_tty->print(" VerifyBeforeGC:");
1258 1258 prepare_for_verify();
1259 1259 Universe::verify(/* allow dirty */ true,
1260 1260 /* silent */ false,
1261 1261 /* option */ VerifyOption_G1UsePrevMarking);
1262 1262
1263 1263 }
1264 1264 pre_full_gc_dump();
1265 1265
1266 1266 COMPILER2_PRESENT(DerivedPointerTable::clear());
1267 1267
1268 1268 // Disable discovery and empty the discovered lists
1269 1269 // for the CM ref processor.
1270 1270 ref_processor_cm()->disable_discovery();
1271 1271 ref_processor_cm()->abandon_partial_discovery();
1272 1272 ref_processor_cm()->verify_no_references_recorded();
1273 1273
1274 1274 // Abandon current iterations of concurrent marking and concurrent
1275 1275 // refinement, if any are in progress.
1276 1276 concurrent_mark()->abort();
1277 1277
1278 1278 // Make sure we'll choose a new allocation region afterwards.
1279 1279 release_mutator_alloc_region();
1280 1280 abandon_gc_alloc_regions();
1281 1281 g1_rem_set()->cleanupHRRS();
1282 1282
1283 1283 // We should call this after we retire any currently active alloc
1284 1284 // regions so that all the ALLOC / RETIRE events are generated
1285 1285 // before the start GC event.
1286 1286 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1287 1287
1288 1288 // We may have added regions to the current incremental collection
1289 1289 // set between the last GC or pause and now. We need to clear the
1290 1290 // incremental collection set and then start rebuilding it afresh
1291 1291 // after this full GC.
1292 1292 abandon_collection_set(g1_policy()->inc_cset_head());
1293 1293 g1_policy()->clear_incremental_cset();
1294 1294 g1_policy()->stop_incremental_cset_building();
1295 1295
1296 1296 tear_down_region_sets(false /* free_list_only */);
1297 1297 g1_policy()->set_gcs_are_young(true);
1298 1298
1299 1299 // See the comments in g1CollectedHeap.hpp and
1300 1300 // G1CollectedHeap::ref_processing_init() about
1301 1301 // how reference processing currently works in G1.
1302 1302
1303 1303 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1304 1304 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1305 1305
1306 1306 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1307 1307 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1308 1308
1309 1309 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1310 1310 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1311 1311
1312 1312 // Do collection work
1313 1313 {
1314 1314 HandleMark hm; // Discard invalid handles created during gc
1315 1315 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1316 1316 }
1317 1317
1318 1318 assert(free_regions() == 0, "we should not have added any free regions");
1319 1319 rebuild_region_sets(false /* free_list_only */);
1320 1320
1321 1321 // Enqueue any discovered reference objects that have
1322 1322 // not been removed from the discovered lists.
1323 1323 ref_processor_stw()->enqueue_discovered_references();
1324 1324
1325 1325 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1326 1326
1327 1327 MemoryService::track_memory_usage();
1328 1328
1329 1329 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
1330 1330 HandleMark hm; // Discard invalid handles created during verification
1331 1331 gclog_or_tty->print(" VerifyAfterGC:");
1332 1332 prepare_for_verify();
1333 1333 Universe::verify(/* allow dirty */ false,
1334 1334 /* silent */ false,
1335 1335 /* option */ VerifyOption_G1UsePrevMarking);
1336 1336
1337 1337 }
1338 1338
1339 1339 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1340 1340 ref_processor_stw()->verify_no_references_recorded();
1341 1341
1342 1342 // Note: since we've just done a full GC, concurrent
1343 1343 // marking is no longer active. Therefore we need not
1344 1344 // re-enable reference discovery for the CM ref processor.
1345 1345 // That will be done at the start of the next marking cycle.
1346 1346 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1347 1347 ref_processor_cm()->verify_no_references_recorded();
1348 1348
1349 1349 reset_gc_time_stamp();
1350 1350 // Since everything potentially moved, we will clear all remembered
1351 1351 // sets, and clear all cards. Later we will rebuild remebered
1352 1352 // sets. We will also reset the GC time stamps of the regions.
1353 1353 PostMCRemSetClearClosure rs_clear(mr_bs());
1354 1354 heap_region_iterate(&rs_clear);
1355 1355
1356 1356 // Resize the heap if necessary.
1357 1357 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1358 1358
1359 1359 if (_hr_printer.is_active()) {
1360 1360 // We should do this after we potentially resize the heap so
1361 1361 // that all the COMMIT / UNCOMMIT events are generated before
1362 1362 // the end GC event.
1363 1363
1364 1364 PostCompactionPrinterClosure cl(hr_printer());
1365 1365 heap_region_iterate(&cl);
1366 1366
1367 1367 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1368 1368 }
1369 1369
1370 1370 if (_cg1r->use_cache()) {
1371 1371 _cg1r->clear_and_record_card_counts();
1372 1372 _cg1r->clear_hot_cache();
1373 1373 }
1374 1374
1375 1375 // Rebuild remembered sets of all regions.
1376 1376 if (G1CollectedHeap::use_parallel_gc_threads()) {
1377 1377 int n_workers =
1378 1378 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1379 1379 workers()->active_workers(),
1380 1380 Threads::number_of_non_daemon_threads());
1381 1381 assert(UseDynamicNumberOfGCThreads ||
1382 1382 n_workers == workers()->total_workers(),
1383 1383 "If not dynamic should be using all the workers");
1384 1384 workers()->set_active_workers(n_workers);
1385 1385 // Set parallel threads in the heap (_n_par_threads) only
1386 1386 // before a parallel phase and always reset it to 0 after
1387 1387 // the phase so that the number of parallel threads does
1388 1388 // no get carried forward to a serial phase where there
1389 1389 // may be code that is "possibly_parallel".
1390 1390 set_par_threads(n_workers);
1391 1391
1392 1392 ParRebuildRSTask rebuild_rs_task(this);
1393 1393 assert(check_heap_region_claim_values(
1394 1394 HeapRegion::InitialClaimValue), "sanity check");
1395 1395 assert(UseDynamicNumberOfGCThreads ||
1396 1396 workers()->active_workers() == workers()->total_workers(),
1397 1397 "Unless dynamic should use total workers");
1398 1398 // Use the most recent number of active workers
1399 1399 assert(workers()->active_workers() > 0,
1400 1400 "Active workers not properly set");
1401 1401 set_par_threads(workers()->active_workers());
1402 1402 workers()->run_task(&rebuild_rs_task);
1403 1403 set_par_threads(0);
1404 1404 assert(check_heap_region_claim_values(
1405 1405 HeapRegion::RebuildRSClaimValue), "sanity check");
1406 1406 reset_heap_region_claim_values();
1407 1407 } else {
1408 1408 RebuildRSOutOfRegionClosure rebuild_rs(this);
1409 1409 heap_region_iterate(&rebuild_rs);
1410 1410 }
1411 1411
1412 1412 if (PrintGC) {
1413 1413 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1414 1414 }
1415 1415
1416 1416 if (true) { // FIXME
1417 1417 // Ask the permanent generation to adjust size for full collections
1418 1418 perm()->compute_new_size();
1419 1419 }
1420 1420
1421 1421 // Start a new incremental collection set for the next pause
1422 1422 assert(g1_policy()->collection_set() == NULL, "must be");
1423 1423 g1_policy()->start_incremental_cset_building();
1424 1424
1425 1425 // Clear the _cset_fast_test bitmap in anticipation of adding
1426 1426 // regions to the incremental collection set for the next
1427 1427 // evacuation pause.
1428 1428 clear_cset_fast_test();
1429 1429
1430 1430 init_mutator_alloc_region();
1431 1431
1432 1432 double end = os::elapsedTime();
1433 1433 g1_policy()->record_full_collection_end();
1434 1434
1435 1435 #ifdef TRACESPINNING
1436 1436 ParallelTaskTerminator::print_termination_counts();
1437 1437 #endif
1438 1438
1439 1439 gc_epilogue(true);
1440 1440
1441 1441 // Discard all rset updates
1442 1442 JavaThread::dirty_card_queue_set().abandon_logs();
1443 1443 assert(!G1DeferredRSUpdate
1444 1444 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1445 1445 }
1446 1446
1447 1447 _young_list->reset_sampled_info();
1448 1448 // At this point there should be no regions in the
1449 1449 // entire heap tagged as young.
1450 1450 assert( check_young_list_empty(true /* check_heap */),
1451 1451 "young list should be empty at this point");
1452 1452
1453 1453 // Update the number of full collections that have been completed.
1454 1454 increment_full_collections_completed(false /* concurrent */);
1455 1455
1456 1456 _hrs.verify_optional();
1457 1457 verify_region_sets_optional();
1458 1458
1459 1459 if (PrintHeapAtGC) {
1460 1460 Universe::print_heap_after_gc();
1461 1461 }
1462 1462 g1mm()->update_sizes();
1463 1463 post_full_gc_dump();
1464 1464
1465 1465 return true;
1466 1466 }
1467 1467
1468 1468 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1469 1469 // do_collection() will return whether it succeeded in performing
1470 1470 // the GC. Currently, there is no facility on the
1471 1471 // do_full_collection() API to notify the caller than the collection
1472 1472 // did not succeed (e.g., because it was locked out by the GC
1473 1473 // locker). So, right now, we'll ignore the return value.
1474 1474 bool dummy = do_collection(true, /* explicit_gc */
1475 1475 clear_all_soft_refs,
1476 1476 0 /* word_size */);
1477 1477 }
1478 1478
1479 1479 // This code is mostly copied from TenuredGeneration.
1480 1480 void
1481 1481 G1CollectedHeap::
1482 1482 resize_if_necessary_after_full_collection(size_t word_size) {
1483 1483 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1484 1484
1485 1485 // Include the current allocation, if any, and bytes that will be
1486 1486 // pre-allocated to support collections, as "used".
1487 1487 const size_t used_after_gc = used();
1488 1488 const size_t capacity_after_gc = capacity();
1489 1489 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1490 1490
1491 1491 // This is enforced in arguments.cpp.
1492 1492 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1493 1493 "otherwise the code below doesn't make sense");
1494 1494
1495 1495 // We don't have floating point command-line arguments
1496 1496 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1497 1497 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1498 1498 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1499 1499 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1500 1500
1501 1501 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1502 1502 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1503 1503
1504 1504 // We have to be careful here as these two calculations can overflow
1505 1505 // 32-bit size_t's.
1506 1506 double used_after_gc_d = (double) used_after_gc;
1507 1507 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1508 1508 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1509 1509
1510 1510 // Let's make sure that they are both under the max heap size, which
1511 1511 // by default will make them fit into a size_t.
1512 1512 double desired_capacity_upper_bound = (double) max_heap_size;
1513 1513 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1514 1514 desired_capacity_upper_bound);
1515 1515 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1516 1516 desired_capacity_upper_bound);
1517 1517
1518 1518 // We can now safely turn them into size_t's.
1519 1519 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1520 1520 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1521 1521
1522 1522 // This assert only makes sense here, before we adjust them
1523 1523 // with respect to the min and max heap size.
1524 1524 assert(minimum_desired_capacity <= maximum_desired_capacity,
1525 1525 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1526 1526 "maximum_desired_capacity = "SIZE_FORMAT,
1527 1527 minimum_desired_capacity, maximum_desired_capacity));
1528 1528
1529 1529 // Should not be greater than the heap max size. No need to adjust
1530 1530 // it with respect to the heap min size as it's a lower bound (i.e.,
1531 1531 // we'll try to make the capacity larger than it, not smaller).
1532 1532 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1533 1533 // Should not be less than the heap min size. No need to adjust it
1534 1534 // with respect to the heap max size as it's an upper bound (i.e.,
1535 1535 // we'll try to make the capacity smaller than it, not greater).
1536 1536 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1537 1537
1538 1538 if (capacity_after_gc < minimum_desired_capacity) {
1539 1539 // Don't expand unless it's significant
1540 1540 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1541 1541 ergo_verbose4(ErgoHeapSizing,
1542 1542 "attempt heap expansion",
1543 1543 ergo_format_reason("capacity lower than "
1544 1544 "min desired capacity after Full GC")
1545 1545 ergo_format_byte("capacity")
1546 1546 ergo_format_byte("occupancy")
1547 1547 ergo_format_byte_perc("min desired capacity"),
1548 1548 capacity_after_gc, used_after_gc,
1549 1549 minimum_desired_capacity, (double) MinHeapFreeRatio);
1550 1550 expand(expand_bytes);
1551 1551
1552 1552 // No expansion, now see if we want to shrink
1553 1553 } else if (capacity_after_gc > maximum_desired_capacity) {
1554 1554 // Capacity too large, compute shrinking size
1555 1555 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1556 1556 ergo_verbose4(ErgoHeapSizing,
1557 1557 "attempt heap shrinking",
1558 1558 ergo_format_reason("capacity higher than "
1559 1559 "max desired capacity after Full GC")
1560 1560 ergo_format_byte("capacity")
1561 1561 ergo_format_byte("occupancy")
1562 1562 ergo_format_byte_perc("max desired capacity"),
1563 1563 capacity_after_gc, used_after_gc,
1564 1564 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1565 1565 shrink(shrink_bytes);
1566 1566 }
1567 1567 }
1568 1568
1569 1569
1570 1570 HeapWord*
1571 1571 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1572 1572 bool* succeeded) {
1573 1573 assert_at_safepoint(true /* should_be_vm_thread */);
1574 1574
1575 1575 *succeeded = true;
1576 1576 // Let's attempt the allocation first.
1577 1577 HeapWord* result =
1578 1578 attempt_allocation_at_safepoint(word_size,
1579 1579 false /* expect_null_mutator_alloc_region */);
1580 1580 if (result != NULL) {
1581 1581 assert(*succeeded, "sanity");
1582 1582 return result;
1583 1583 }
1584 1584
1585 1585 // In a G1 heap, we're supposed to keep allocation from failing by
1586 1586 // incremental pauses. Therefore, at least for now, we'll favor
1587 1587 // expansion over collection. (This might change in the future if we can
1588 1588 // do something smarter than full collection to satisfy a failed alloc.)
1589 1589 result = expand_and_allocate(word_size);
1590 1590 if (result != NULL) {
1591 1591 assert(*succeeded, "sanity");
1592 1592 return result;
1593 1593 }
1594 1594
1595 1595 // Expansion didn't work, we'll try to do a Full GC.
1596 1596 bool gc_succeeded = do_collection(false, /* explicit_gc */
1597 1597 false, /* clear_all_soft_refs */
1598 1598 word_size);
1599 1599 if (!gc_succeeded) {
1600 1600 *succeeded = false;
1601 1601 return NULL;
1602 1602 }
1603 1603
1604 1604 // Retry the allocation
1605 1605 result = attempt_allocation_at_safepoint(word_size,
1606 1606 true /* expect_null_mutator_alloc_region */);
1607 1607 if (result != NULL) {
1608 1608 assert(*succeeded, "sanity");
1609 1609 return result;
1610 1610 }
1611 1611
1612 1612 // Then, try a Full GC that will collect all soft references.
1613 1613 gc_succeeded = do_collection(false, /* explicit_gc */
1614 1614 true, /* clear_all_soft_refs */
1615 1615 word_size);
1616 1616 if (!gc_succeeded) {
1617 1617 *succeeded = false;
1618 1618 return NULL;
1619 1619 }
1620 1620
1621 1621 // Retry the allocation once more
1622 1622 result = attempt_allocation_at_safepoint(word_size,
1623 1623 true /* expect_null_mutator_alloc_region */);
1624 1624 if (result != NULL) {
1625 1625 assert(*succeeded, "sanity");
1626 1626 return result;
1627 1627 }
1628 1628
1629 1629 assert(!collector_policy()->should_clear_all_soft_refs(),
1630 1630 "Flag should have been handled and cleared prior to this point");
1631 1631
1632 1632 // What else? We might try synchronous finalization later. If the total
1633 1633 // space available is large enough for the allocation, then a more
1634 1634 // complete compaction phase than we've tried so far might be
1635 1635 // appropriate.
1636 1636 assert(*succeeded, "sanity");
1637 1637 return NULL;
1638 1638 }
1639 1639
1640 1640 // Attempting to expand the heap sufficiently
1641 1641 // to support an allocation of the given "word_size". If
1642 1642 // successful, perform the allocation and return the address of the
1643 1643 // allocated block, or else "NULL".
1644 1644
1645 1645 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1646 1646 assert_at_safepoint(true /* should_be_vm_thread */);
1647 1647
1648 1648 verify_region_sets_optional();
1649 1649
1650 1650 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1651 1651 ergo_verbose1(ErgoHeapSizing,
1652 1652 "attempt heap expansion",
1653 1653 ergo_format_reason("allocation request failed")
1654 1654 ergo_format_byte("allocation request"),
1655 1655 word_size * HeapWordSize);
1656 1656 if (expand(expand_bytes)) {
1657 1657 _hrs.verify_optional();
1658 1658 verify_region_sets_optional();
1659 1659 return attempt_allocation_at_safepoint(word_size,
1660 1660 false /* expect_null_mutator_alloc_region */);
1661 1661 }
1662 1662 return NULL;
1663 1663 }
1664 1664
1665 1665 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1666 1666 HeapWord* new_end) {
1667 1667 assert(old_end != new_end, "don't call this otherwise");
1668 1668 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1669 1669
1670 1670 // Update the committed mem region.
1671 1671 _g1_committed.set_end(new_end);
1672 1672 // Tell the card table about the update.
1673 1673 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1674 1674 // Tell the BOT about the update.
1675 1675 _bot_shared->resize(_g1_committed.word_size());
1676 1676 }
1677 1677
1678 1678 bool G1CollectedHeap::expand(size_t expand_bytes) {
1679 1679 size_t old_mem_size = _g1_storage.committed_size();
1680 1680 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1681 1681 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1682 1682 HeapRegion::GrainBytes);
1683 1683 ergo_verbose2(ErgoHeapSizing,
1684 1684 "expand the heap",
1685 1685 ergo_format_byte("requested expansion amount")
1686 1686 ergo_format_byte("attempted expansion amount"),
1687 1687 expand_bytes, aligned_expand_bytes);
1688 1688
1689 1689 // First commit the memory.
1690 1690 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1691 1691 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1692 1692 if (successful) {
1693 1693 // Then propagate this update to the necessary data structures.
1694 1694 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1695 1695 update_committed_space(old_end, new_end);
1696 1696
1697 1697 FreeRegionList expansion_list("Local Expansion List");
1698 1698 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1699 1699 assert(mr.start() == old_end, "post-condition");
1700 1700 // mr might be a smaller region than what was requested if
1701 1701 // expand_by() was unable to allocate the HeapRegion instances
1702 1702 assert(mr.end() <= new_end, "post-condition");
1703 1703
1704 1704 size_t actual_expand_bytes = mr.byte_size();
1705 1705 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1706 1706 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1707 1707 "post-condition");
1708 1708 if (actual_expand_bytes < aligned_expand_bytes) {
1709 1709 // We could not expand _hrs to the desired size. In this case we
1710 1710 // need to shrink the committed space accordingly.
1711 1711 assert(mr.end() < new_end, "invariant");
1712 1712
1713 1713 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1714 1714 // First uncommit the memory.
1715 1715 _g1_storage.shrink_by(diff_bytes);
1716 1716 // Then propagate this update to the necessary data structures.
1717 1717 update_committed_space(new_end, mr.end());
1718 1718 }
1719 1719 _free_list.add_as_tail(&expansion_list);
1720 1720
1721 1721 if (_hr_printer.is_active()) {
1722 1722 HeapWord* curr = mr.start();
1723 1723 while (curr < mr.end()) {
1724 1724 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1725 1725 _hr_printer.commit(curr, curr_end);
1726 1726 curr = curr_end;
1727 1727 }
1728 1728 assert(curr == mr.end(), "post-condition");
1729 1729 }
1730 1730 g1_policy()->record_new_heap_size(n_regions());
1731 1731 } else {
1732 1732 ergo_verbose0(ErgoHeapSizing,
1733 1733 "did not expand the heap",
1734 1734 ergo_format_reason("heap expansion operation failed"));
1735 1735 // The expansion of the virtual storage space was unsuccessful.
1736 1736 // Let's see if it was because we ran out of swap.
1737 1737 if (G1ExitOnExpansionFailure &&
1738 1738 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1739 1739 // We had head room...
1740 1740 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1741 1741 }
1742 1742 }
1743 1743 return successful;
1744 1744 }
1745 1745
1746 1746 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1747 1747 size_t old_mem_size = _g1_storage.committed_size();
1748 1748 size_t aligned_shrink_bytes =
1749 1749 ReservedSpace::page_align_size_down(shrink_bytes);
1750 1750 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1751 1751 HeapRegion::GrainBytes);
1752 1752 size_t num_regions_deleted = 0;
1753 1753 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1754 1754 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1755 1755 assert(mr.end() == old_end, "post-condition");
1756 1756
1757 1757 ergo_verbose3(ErgoHeapSizing,
1758 1758 "shrink the heap",
1759 1759 ergo_format_byte("requested shrinking amount")
1760 1760 ergo_format_byte("aligned shrinking amount")
1761 1761 ergo_format_byte("attempted shrinking amount"),
1762 1762 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1763 1763 if (mr.byte_size() > 0) {
1764 1764 if (_hr_printer.is_active()) {
1765 1765 HeapWord* curr = mr.end();
1766 1766 while (curr > mr.start()) {
1767 1767 HeapWord* curr_end = curr;
1768 1768 curr -= HeapRegion::GrainWords;
1769 1769 _hr_printer.uncommit(curr, curr_end);
1770 1770 }
1771 1771 assert(curr == mr.start(), "post-condition");
1772 1772 }
1773 1773
1774 1774 _g1_storage.shrink_by(mr.byte_size());
1775 1775 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1776 1776 assert(mr.start() == new_end, "post-condition");
1777 1777
1778 1778 _expansion_regions += num_regions_deleted;
1779 1779 update_committed_space(old_end, new_end);
1780 1780 HeapRegionRemSet::shrink_heap(n_regions());
1781 1781 g1_policy()->record_new_heap_size(n_regions());
1782 1782 } else {
1783 1783 ergo_verbose0(ErgoHeapSizing,
1784 1784 "did not shrink the heap",
1785 1785 ergo_format_reason("heap shrinking operation failed"));
1786 1786 }
1787 1787 }
1788 1788
1789 1789 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1790 1790 verify_region_sets_optional();
1791 1791
1792 1792 // We should only reach here at the end of a Full GC which means we
1793 1793 // should not not be holding to any GC alloc regions. The method
1794 1794 // below will make sure of that and do any remaining clean up.
1795 1795 abandon_gc_alloc_regions();
1796 1796
1797 1797 // Instead of tearing down / rebuilding the free lists here, we
1798 1798 // could instead use the remove_all_pending() method on free_list to
1799 1799 // remove only the ones that we need to remove.
1800 1800 tear_down_region_sets(true /* free_list_only */);
1801 1801 shrink_helper(shrink_bytes);
1802 1802 rebuild_region_sets(true /* free_list_only */);
1803 1803
1804 1804 _hrs.verify_optional();
1805 1805 verify_region_sets_optional();
1806 1806 }
1807 1807
1808 1808 // Public methods.
1809 1809
1810 1810 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1811 1811 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1812 1812 #endif // _MSC_VER
1813 1813
1814 1814
1815 1815 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1816 1816 SharedHeap(policy_),
1817 1817 _g1_policy(policy_),
1818 1818 _dirty_card_queue_set(false),
1819 1819 _into_cset_dirty_card_queue_set(false),
1820 1820 _is_alive_closure_cm(this),
1821 1821 _is_alive_closure_stw(this),
1822 1822 _ref_processor_cm(NULL),
1823 1823 _ref_processor_stw(NULL),
1824 1824 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1825 1825 _bot_shared(NULL),
1826 1826 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1827 1827 _evac_failure_scan_stack(NULL) ,
1828 1828 _mark_in_progress(false),
1829 1829 _cg1r(NULL), _summary_bytes_used(0),
1830 1830 _g1mm(NULL),
1831 1831 _refine_cte_cl(NULL),
1832 1832 _full_collection(false),
1833 1833 _free_list("Master Free List"),
1834 1834 _secondary_free_list("Secondary Free List"),
1835 1835 _old_set("Old Set"),
1836 1836 _humongous_set("Master Humongous Set"),
1837 1837 _free_regions_coming(false),
1838 1838 _young_list(new YoungList(this)),
1839 1839 _gc_time_stamp(0),
1840 1840 _retained_old_gc_alloc_region(NULL),
1841 1841 _surviving_young_words(NULL),
1842 1842 _full_collections_completed(0),
1843 1843 _in_cset_fast_test(NULL),
1844 1844 _in_cset_fast_test_base(NULL),
1845 1845 _dirty_cards_region_list(NULL),
1846 1846 _worker_cset_start_region(NULL),
1847 1847 _worker_cset_start_region_time_stamp(NULL) {
1848 1848 _g1h = this; // To catch bugs.
1849 1849 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1850 1850 vm_exit_during_initialization("Failed necessary allocation.");
1851 1851 }
1852 1852
1853 1853 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1854 1854
1855 1855 int n_queues = MAX2((int)ParallelGCThreads, 1);
1856 1856 _task_queues = new RefToScanQueueSet(n_queues);
1857 1857
1858 1858 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1859 1859 assert(n_rem_sets > 0, "Invariant.");
1860 1860
1861 1861 HeapRegionRemSetIterator** iter_arr =
1862 1862 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues);
1863 1863 for (int i = 0; i < n_queues; i++) {
1864 1864 iter_arr[i] = new HeapRegionRemSetIterator();
1865 1865 }
1866 1866 _rem_set_iterator = iter_arr;
1867 1867
1868 1868 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues);
1869 1869 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues);
1870 1870
1871 1871 for (int i = 0; i < n_queues; i++) {
1872 1872 RefToScanQueue* q = new RefToScanQueue();
1873 1873 q->initialize();
1874 1874 _task_queues->register_queue(i, q);
1875 1875 }
1876 1876
1877 1877 clear_cset_start_regions();
1878 1878
1879 1879 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1880 1880 }
1881 1881
1882 1882 jint G1CollectedHeap::initialize() {
1883 1883 CollectedHeap::pre_initialize();
1884 1884 os::enable_vtime();
1885 1885
1886 1886 // Necessary to satisfy locking discipline assertions.
1887 1887
1888 1888 MutexLocker x(Heap_lock);
1889 1889
1890 1890 // We have to initialize the printer before committing the heap, as
1891 1891 // it will be used then.
1892 1892 _hr_printer.set_active(G1PrintHeapRegions);
1893 1893
1894 1894 // While there are no constraints in the GC code that HeapWordSize
1895 1895 // be any particular value, there are multiple other areas in the
1896 1896 // system which believe this to be true (e.g. oop->object_size in some
1897 1897 // cases incorrectly returns the size in wordSize units rather than
1898 1898 // HeapWordSize).
1899 1899 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1900 1900
1901 1901 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1902 1902 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1903 1903
1904 1904 // Ensure that the sizes are properly aligned.
1905 1905 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1906 1906 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1907 1907
1908 1908 _cg1r = new ConcurrentG1Refine();
1909 1909
1910 1910 // Reserve the maximum.
1911 1911 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation();
1912 1912 // Includes the perm-gen.
1913 1913
1914 1914 // When compressed oops are enabled, the preferred heap base
1915 1915 // is calculated by subtracting the requested size from the
1916 1916 // 32Gb boundary and using the result as the base address for
1917 1917 // heap reservation. If the requested size is not aligned to
1918 1918 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1919 1919 // into the ReservedHeapSpace constructor) then the actual
1920 1920 // base of the reserved heap may end up differing from the
1921 1921 // address that was requested (i.e. the preferred heap base).
1922 1922 // If this happens then we could end up using a non-optimal
1923 1923 // compressed oops mode.
1924 1924
1925 1925 // Since max_byte_size is aligned to the size of a heap region (checked
1926 1926 // above), we also need to align the perm gen size as it might not be.
1927 1927 const size_t total_reserved = max_byte_size +
1928 1928 align_size_up(pgs->max_size(), HeapRegion::GrainBytes);
1929 1929 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm");
1930 1930
1931 1931 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop);
1932 1932
1933 1933 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes,
1934 1934 UseLargePages, addr);
1935 1935
1936 1936 if (UseCompressedOops) {
1937 1937 if (addr != NULL && !heap_rs.is_reserved()) {
1938 1938 // Failed to reserve at specified address - the requested memory
1939 1939 // region is taken already, for example, by 'java' launcher.
1940 1940 // Try again to reserver heap higher.
1941 1941 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop);
1942 1942
1943 1943 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes,
1944 1944 UseLargePages, addr);
1945 1945
1946 1946 if (addr != NULL && !heap_rs0.is_reserved()) {
1947 1947 // Failed to reserve at specified address again - give up.
1948 1948 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop);
1949 1949 assert(addr == NULL, "");
1950 1950
1951 1951 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes,
1952 1952 UseLargePages, addr);
1953 1953 heap_rs = heap_rs1;
1954 1954 } else {
1955 1955 heap_rs = heap_rs0;
1956 1956 }
1957 1957 }
1958 1958 }
1959 1959
1960 1960 if (!heap_rs.is_reserved()) {
1961 1961 vm_exit_during_initialization("Could not reserve enough space for object heap");
1962 1962 return JNI_ENOMEM;
1963 1963 }
1964 1964
1965 1965 // It is important to do this in a way such that concurrent readers can't
1966 1966 // temporarily think somethings in the heap. (I've actually seen this
1967 1967 // happen in asserts: DLD.)
1968 1968 _reserved.set_word_size(0);
1969 1969 _reserved.set_start((HeapWord*)heap_rs.base());
1970 1970 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
1971 1971
1972 1972 _expansion_regions = max_byte_size/HeapRegion::GrainBytes;
1973 1973
1974 1974 // Create the gen rem set (and barrier set) for the entire reserved region.
1975 1975 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
1976 1976 set_barrier_set(rem_set()->bs());
1977 1977 if (barrier_set()->is_a(BarrierSet::ModRef)) {
1978 1978 _mr_bs = (ModRefBarrierSet*)_barrier_set;
1979 1979 } else {
1980 1980 vm_exit_during_initialization("G1 requires a mod ref bs.");
1981 1981 return JNI_ENOMEM;
1982 1982 }
1983 1983
1984 1984 // Also create a G1 rem set.
1985 1985 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
1986 1986 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
1987 1987 } else {
1988 1988 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
1989 1989 return JNI_ENOMEM;
1990 1990 }
1991 1991
1992 1992 // Carve out the G1 part of the heap.
1993 1993
1994 1994 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
1995 1995 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
1996 1996 g1_rs.size()/HeapWordSize);
1997 1997 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size);
1998 1998
1999 1999 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set());
2000 2000
2001 2001 _g1_storage.initialize(g1_rs, 0);
2002 2002 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2003 2003 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2004 2004 (HeapWord*) _g1_reserved.end(),
2005 2005 _expansion_regions);
2006 2006
2007 2007 // 6843694 - ensure that the maximum region index can fit
2008 2008 // in the remembered set structures.
2009 2009 const size_t max_region_idx = ((size_t)1 << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2010 2010 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2011 2011
2012 2012 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2013 2013 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2014 2014 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2015 2015 "too many cards per region");
2016 2016
2017 2017 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2018 2018
2019 2019 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2020 2020 heap_word_size(init_byte_size));
2021 2021
2022 2022 _g1h = this;
2023 2023
2024 2024 _in_cset_fast_test_length = max_regions();
2025 2025 _in_cset_fast_test_base = NEW_C_HEAP_ARRAY(bool, _in_cset_fast_test_length);
2026 2026
2027 2027 // We're biasing _in_cset_fast_test to avoid subtracting the
2028 2028 // beginning of the heap every time we want to index; basically
2029 2029 // it's the same with what we do with the card table.
2030 2030 _in_cset_fast_test = _in_cset_fast_test_base -
2031 2031 ((size_t) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2032 2032
2033 2033 // Clear the _cset_fast_test bitmap in anticipation of adding
2034 2034 // regions to the incremental collection set for the first
2035 2035 // evacuation pause.
2036 2036 clear_cset_fast_test();
2037 2037
2038 2038 // Create the ConcurrentMark data structure and thread.
2039 2039 // (Must do this late, so that "max_regions" is defined.)
2040 2040 _cm = new ConcurrentMark(heap_rs, (int) max_regions());
2041 2041 _cmThread = _cm->cmThread();
2042 2042
2043 2043 // Initialize the from_card cache structure of HeapRegionRemSet.
2044 2044 HeapRegionRemSet::init_heap(max_regions());
2045 2045
2046 2046 // Now expand into the initial heap size.
2047 2047 if (!expand(init_byte_size)) {
2048 2048 vm_exit_during_initialization("Failed to allocate initial heap.");
2049 2049 return JNI_ENOMEM;
2050 2050 }
2051 2051
2052 2052 // Perform any initialization actions delegated to the policy.
2053 2053 g1_policy()->init();
2054 2054
2055 2055 _refine_cte_cl =
2056 2056 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2057 2057 g1_rem_set(),
2058 2058 concurrent_g1_refine());
2059 2059 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2060 2060
2061 2061 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2062 2062 SATB_Q_FL_lock,
2063 2063 G1SATBProcessCompletedThreshold,
2064 2064 Shared_SATB_Q_lock);
2065 2065
2066 2066 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2067 2067 DirtyCardQ_FL_lock,
2068 2068 concurrent_g1_refine()->yellow_zone(),
2069 2069 concurrent_g1_refine()->red_zone(),
2070 2070 Shared_DirtyCardQ_lock);
2071 2071
2072 2072 if (G1DeferredRSUpdate) {
2073 2073 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2074 2074 DirtyCardQ_FL_lock,
2075 2075 -1, // never trigger processing
2076 2076 -1, // no limit on length
2077 2077 Shared_DirtyCardQ_lock,
2078 2078 &JavaThread::dirty_card_queue_set());
2079 2079 }
2080 2080
2081 2081 // Initialize the card queue set used to hold cards containing
2082 2082 // references into the collection set.
2083 2083 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2084 2084 DirtyCardQ_FL_lock,
2085 2085 -1, // never trigger processing
2086 2086 -1, // no limit on length
2087 2087 Shared_DirtyCardQ_lock,
2088 2088 &JavaThread::dirty_card_queue_set());
2089 2089
2090 2090 // In case we're keeping closure specialization stats, initialize those
2091 2091 // counts and that mechanism.
2092 2092 SpecializationStats::clear();
2093 2093
2094 2094 // Do later initialization work for concurrent refinement.
2095 2095 _cg1r->init();
2096 2096
2097 2097 // Here we allocate the dummy full region that is required by the
2098 2098 // G1AllocRegion class. If we don't pass an address in the reserved
2099 2099 // space here, lots of asserts fire.
2100 2100
2101 2101 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2102 2102 _g1_reserved.start());
2103 2103 // We'll re-use the same region whether the alloc region will
2104 2104 // require BOT updates or not and, if it doesn't, then a non-young
2105 2105 // region will complain that it cannot support allocations without
2106 2106 // BOT updates. So we'll tag the dummy region as young to avoid that.
2107 2107 dummy_region->set_young();
2108 2108 // Make sure it's full.
2109 2109 dummy_region->set_top(dummy_region->end());
2110 2110 G1AllocRegion::setup(this, dummy_region);
2111 2111
2112 2112 init_mutator_alloc_region();
2113 2113
2114 2114 // Do create of the monitoring and management support so that
2115 2115 // values in the heap have been properly initialized.
2116 2116 _g1mm = new G1MonitoringSupport(this);
2117 2117
2118 2118 return JNI_OK;
2119 2119 }
2120 2120
2121 2121 void G1CollectedHeap::ref_processing_init() {
2122 2122 // Reference processing in G1 currently works as follows:
2123 2123 //
2124 2124 // * There are two reference processor instances. One is
2125 2125 // used to record and process discovered references
2126 2126 // during concurrent marking; the other is used to
2127 2127 // record and process references during STW pauses
2128 2128 // (both full and incremental).
2129 2129 // * Both ref processors need to 'span' the entire heap as
2130 2130 // the regions in the collection set may be dotted around.
2131 2131 //
2132 2132 // * For the concurrent marking ref processor:
2133 2133 // * Reference discovery is enabled at initial marking.
2134 2134 // * Reference discovery is disabled and the discovered
2135 2135 // references processed etc during remarking.
2136 2136 // * Reference discovery is MT (see below).
2137 2137 // * Reference discovery requires a barrier (see below).
2138 2138 // * Reference processing may or may not be MT
2139 2139 // (depending on the value of ParallelRefProcEnabled
2140 2140 // and ParallelGCThreads).
2141 2141 // * A full GC disables reference discovery by the CM
2142 2142 // ref processor and abandons any entries on it's
2143 2143 // discovered lists.
2144 2144 //
2145 2145 // * For the STW processor:
2146 2146 // * Non MT discovery is enabled at the start of a full GC.
2147 2147 // * Processing and enqueueing during a full GC is non-MT.
2148 2148 // * During a full GC, references are processed after marking.
2149 2149 //
2150 2150 // * Discovery (may or may not be MT) is enabled at the start
2151 2151 // of an incremental evacuation pause.
2152 2152 // * References are processed near the end of a STW evacuation pause.
2153 2153 // * For both types of GC:
2154 2154 // * Discovery is atomic - i.e. not concurrent.
2155 2155 // * Reference discovery will not need a barrier.
2156 2156
2157 2157 SharedHeap::ref_processing_init();
2158 2158 MemRegion mr = reserved_region();
2159 2159
2160 2160 // Concurrent Mark ref processor
2161 2161 _ref_processor_cm =
2162 2162 new ReferenceProcessor(mr, // span
2163 2163 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2164 2164 // mt processing
2165 2165 (int) ParallelGCThreads,
2166 2166 // degree of mt processing
2167 2167 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2168 2168 // mt discovery
2169 2169 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2170 2170 // degree of mt discovery
2171 2171 false,
2172 2172 // Reference discovery is not atomic
2173 2173 &_is_alive_closure_cm,
2174 2174 // is alive closure
2175 2175 // (for efficiency/performance)
2176 2176 true);
2177 2177 // Setting next fields of discovered
2178 2178 // lists requires a barrier.
2179 2179
2180 2180 // STW ref processor
2181 2181 _ref_processor_stw =
2182 2182 new ReferenceProcessor(mr, // span
2183 2183 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2184 2184 // mt processing
2185 2185 MAX2((int)ParallelGCThreads, 1),
2186 2186 // degree of mt processing
2187 2187 (ParallelGCThreads > 1),
2188 2188 // mt discovery
2189 2189 MAX2((int)ParallelGCThreads, 1),
2190 2190 // degree of mt discovery
2191 2191 true,
2192 2192 // Reference discovery is atomic
2193 2193 &_is_alive_closure_stw,
2194 2194 // is alive closure
2195 2195 // (for efficiency/performance)
2196 2196 false);
2197 2197 // Setting next fields of discovered
2198 2198 // lists requires a barrier.
2199 2199 }
2200 2200
2201 2201 size_t G1CollectedHeap::capacity() const {
2202 2202 return _g1_committed.byte_size();
2203 2203 }
2204 2204
2205 2205 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2206 2206 DirtyCardQueue* into_cset_dcq,
2207 2207 bool concurrent,
2208 2208 int worker_i) {
2209 2209 // Clean cards in the hot card cache
2210 2210 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2211 2211
2212 2212 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2213 2213 int n_completed_buffers = 0;
2214 2214 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2215 2215 n_completed_buffers++;
2216 2216 }
2217 2217 g1_policy()->record_update_rs_processed_buffers(worker_i,
2218 2218 (double) n_completed_buffers);
2219 2219 dcqs.clear_n_completed_buffers();
2220 2220 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2221 2221 }
2222 2222
2223 2223
2224 2224 // Computes the sum of the storage used by the various regions.
2225 2225
2226 2226 size_t G1CollectedHeap::used() const {
2227 2227 assert(Heap_lock->owner() != NULL,
2228 2228 "Should be owned on this thread's behalf.");
2229 2229 size_t result = _summary_bytes_used;
2230 2230 // Read only once in case it is set to NULL concurrently
2231 2231 HeapRegion* hr = _mutator_alloc_region.get();
2232 2232 if (hr != NULL)
2233 2233 result += hr->used();
2234 2234 return result;
2235 2235 }
2236 2236
2237 2237 size_t G1CollectedHeap::used_unlocked() const {
2238 2238 size_t result = _summary_bytes_used;
2239 2239 return result;
2240 2240 }
2241 2241
2242 2242 class SumUsedClosure: public HeapRegionClosure {
2243 2243 size_t _used;
2244 2244 public:
2245 2245 SumUsedClosure() : _used(0) {}
2246 2246 bool doHeapRegion(HeapRegion* r) {
2247 2247 if (!r->continuesHumongous()) {
2248 2248 _used += r->used();
2249 2249 }
2250 2250 return false;
2251 2251 }
2252 2252 size_t result() { return _used; }
2253 2253 };
2254 2254
2255 2255 size_t G1CollectedHeap::recalculate_used() const {
2256 2256 SumUsedClosure blk;
2257 2257 heap_region_iterate(&blk);
2258 2258 return blk.result();
2259 2259 }
2260 2260
2261 2261 size_t G1CollectedHeap::unsafe_max_alloc() {
2262 2262 if (free_regions() > 0) return HeapRegion::GrainBytes;
2263 2263 // otherwise, is there space in the current allocation region?
2264 2264
2265 2265 // We need to store the current allocation region in a local variable
2266 2266 // here. The problem is that this method doesn't take any locks and
2267 2267 // there may be other threads which overwrite the current allocation
2268 2268 // region field. attempt_allocation(), for example, sets it to NULL
2269 2269 // and this can happen *after* the NULL check here but before the call
2270 2270 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2271 2271 // to be a problem in the optimized build, since the two loads of the
2272 2272 // current allocation region field are optimized away.
2273 2273 HeapRegion* hr = _mutator_alloc_region.get();
2274 2274 if (hr == NULL) {
2275 2275 return 0;
2276 2276 }
2277 2277 return hr->free();
2278 2278 }
2279 2279
2280 2280 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2281 2281 return
2282 2282 ((cause == GCCause::_gc_locker && GCLockerInvokesConcurrent) ||
2283 2283 (cause == GCCause::_java_lang_system_gc && ExplicitGCInvokesConcurrent));
2284 2284 }
2285 2285
2286 2286 #ifndef PRODUCT
2287 2287 void G1CollectedHeap::allocate_dummy_regions() {
2288 2288 // Let's fill up most of the region
2289 2289 size_t word_size = HeapRegion::GrainWords - 1024;
2290 2290 // And as a result the region we'll allocate will be humongous.
2291 2291 guarantee(isHumongous(word_size), "sanity");
2292 2292
2293 2293 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2294 2294 // Let's use the existing mechanism for the allocation
2295 2295 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2296 2296 if (dummy_obj != NULL) {
2297 2297 MemRegion mr(dummy_obj, word_size);
2298 2298 CollectedHeap::fill_with_object(mr);
2299 2299 } else {
2300 2300 // If we can't allocate once, we probably cannot allocate
2301 2301 // again. Let's get out of the loop.
2302 2302 break;
2303 2303 }
2304 2304 }
2305 2305 }
2306 2306 #endif // !PRODUCT
2307 2307
2308 2308 void G1CollectedHeap::increment_full_collections_completed(bool concurrent) {
2309 2309 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2310 2310
2311 2311 // We assume that if concurrent == true, then the caller is a
2312 2312 // concurrent thread that was joined the Suspendible Thread
2313 2313 // Set. If there's ever a cheap way to check this, we should add an
2314 2314 // assert here.
2315 2315
2316 2316 // We have already incremented _total_full_collections at the start
2317 2317 // of the GC, so total_full_collections() represents how many full
2318 2318 // collections have been started.
2319 2319 unsigned int full_collections_started = total_full_collections();
2320 2320
2321 2321 // Given that this method is called at the end of a Full GC or of a
2322 2322 // concurrent cycle, and those can be nested (i.e., a Full GC can
2323 2323 // interrupt a concurrent cycle), the number of full collections
2324 2324 // completed should be either one (in the case where there was no
2325 2325 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2326 2326 // behind the number of full collections started.
2327 2327
2328 2328 // This is the case for the inner caller, i.e. a Full GC.
2329 2329 assert(concurrent ||
2330 2330 (full_collections_started == _full_collections_completed + 1) ||
2331 2331 (full_collections_started == _full_collections_completed + 2),
2332 2332 err_msg("for inner caller (Full GC): full_collections_started = %u "
2333 2333 "is inconsistent with _full_collections_completed = %u",
2334 2334 full_collections_started, _full_collections_completed));
2335 2335
2336 2336 // This is the case for the outer caller, i.e. the concurrent cycle.
2337 2337 assert(!concurrent ||
2338 2338 (full_collections_started == _full_collections_completed + 1),
2339 2339 err_msg("for outer caller (concurrent cycle): "
2340 2340 "full_collections_started = %u "
2341 2341 "is inconsistent with _full_collections_completed = %u",
2342 2342 full_collections_started, _full_collections_completed));
2343 2343
2344 2344 _full_collections_completed += 1;
2345 2345
2346 2346 // We need to clear the "in_progress" flag in the CM thread before
2347 2347 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2348 2348 // is set) so that if a waiter requests another System.gc() it doesn't
2349 2349 // incorrectly see that a marking cyle is still in progress.
2350 2350 if (concurrent) {
2351 2351 _cmThread->clear_in_progress();
2352 2352 }
2353 2353
2354 2354 // This notify_all() will ensure that a thread that called
2355 2355 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2356 2356 // and it's waiting for a full GC to finish will be woken up. It is
2357 2357 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2358 2358 FullGCCount_lock->notify_all();
2359 2359 }
2360 2360
2361 2361 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) {
2362 2362 assert_at_safepoint(true /* should_be_vm_thread */);
2363 2363 GCCauseSetter gcs(this, cause);
2364 2364 switch (cause) {
2365 2365 case GCCause::_heap_inspection:
2366 2366 case GCCause::_heap_dump: {
2367 2367 HandleMark hm;
2368 2368 do_full_collection(false); // don't clear all soft refs
2369 2369 break;
2370 2370 }
2371 2371 default: // XXX FIX ME
2372 2372 ShouldNotReachHere(); // Unexpected use of this function
2373 2373 }
2374 2374 }
2375 2375
2376 2376 void G1CollectedHeap::collect(GCCause::Cause cause) {
2377 2377 // The caller doesn't have the Heap_lock
2378 2378 assert(!Heap_lock->owned_by_self(), "this thread should not own the Heap_lock");
2379 2379
2380 2380 unsigned int gc_count_before;
2381 2381 unsigned int full_gc_count_before;
2382 2382 {
2383 2383 MutexLocker ml(Heap_lock);
2384 2384
2385 2385 // Read the GC count while holding the Heap_lock
2386 2386 gc_count_before = SharedHeap::heap()->total_collections();
2387 2387 full_gc_count_before = SharedHeap::heap()->total_full_collections();
2388 2388 }
2389 2389
2390 2390 if (should_do_concurrent_full_gc(cause)) {
2391 2391 // Schedule an initial-mark evacuation pause that will start a
2392 2392 // concurrent cycle. We're setting word_size to 0 which means that
2393 2393 // we are not requesting a post-GC allocation.
2394 2394 VM_G1IncCollectionPause op(gc_count_before,
2395 2395 0, /* word_size */
2396 2396 true, /* should_initiate_conc_mark */
2397 2397 g1_policy()->max_pause_time_ms(),
2398 2398 cause);
2399 2399 VMThread::execute(&op);
2400 2400 } else {
2401 2401 if (cause == GCCause::_gc_locker
2402 2402 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2403 2403
2404 2404 // Schedule a standard evacuation pause. We're setting word_size
2405 2405 // to 0 which means that we are not requesting a post-GC allocation.
2406 2406 VM_G1IncCollectionPause op(gc_count_before,
2407 2407 0, /* word_size */
2408 2408 false, /* should_initiate_conc_mark */
2409 2409 g1_policy()->max_pause_time_ms(),
2410 2410 cause);
2411 2411 VMThread::execute(&op);
2412 2412 } else {
2413 2413 // Schedule a Full GC.
2414 2414 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
2415 2415 VMThread::execute(&op);
2416 2416 }
2417 2417 }
2418 2418 }
2419 2419
2420 2420 bool G1CollectedHeap::is_in(const void* p) const {
2421 2421 if (_g1_committed.contains(p)) {
2422 2422 // Given that we know that p is in the committed space,
2423 2423 // heap_region_containing_raw() should successfully
2424 2424 // return the containing region.
2425 2425 HeapRegion* hr = heap_region_containing_raw(p);
2426 2426 return hr->is_in(p);
2427 2427 } else {
2428 2428 return _perm_gen->as_gen()->is_in(p);
2429 2429 }
2430 2430 }
2431 2431
2432 2432 // Iteration functions.
2433 2433
2434 2434 // Iterates an OopClosure over all ref-containing fields of objects
2435 2435 // within a HeapRegion.
2436 2436
2437 2437 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2438 2438 MemRegion _mr;
2439 2439 OopClosure* _cl;
2440 2440 public:
2441 2441 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl)
2442 2442 : _mr(mr), _cl(cl) {}
2443 2443 bool doHeapRegion(HeapRegion* r) {
2444 2444 if (! r->continuesHumongous()) {
2445 2445 r->oop_iterate(_cl);
2446 2446 }
2447 2447 return false;
2448 2448 }
2449 2449 };
2450 2450
2451 2451 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) {
2452 2452 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2453 2453 heap_region_iterate(&blk);
2454 2454 if (do_perm) {
2455 2455 perm_gen()->oop_iterate(cl);
2456 2456 }
2457 2457 }
2458 2458
2459 2459 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) {
2460 2460 IterateOopClosureRegionClosure blk(mr, cl);
2461 2461 heap_region_iterate(&blk);
2462 2462 if (do_perm) {
2463 2463 perm_gen()->oop_iterate(cl);
2464 2464 }
2465 2465 }
2466 2466
2467 2467 // Iterates an ObjectClosure over all objects within a HeapRegion.
2468 2468
2469 2469 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2470 2470 ObjectClosure* _cl;
2471 2471 public:
2472 2472 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2473 2473 bool doHeapRegion(HeapRegion* r) {
2474 2474 if (! r->continuesHumongous()) {
2475 2475 r->object_iterate(_cl);
2476 2476 }
2477 2477 return false;
2478 2478 }
2479 2479 };
2480 2480
2481 2481 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) {
2482 2482 IterateObjectClosureRegionClosure blk(cl);
2483 2483 heap_region_iterate(&blk);
2484 2484 if (do_perm) {
2485 2485 perm_gen()->object_iterate(cl);
2486 2486 }
2487 2487 }
2488 2488
2489 2489 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2490 2490 // FIXME: is this right?
2491 2491 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2492 2492 }
2493 2493
2494 2494 // Calls a SpaceClosure on a HeapRegion.
2495 2495
2496 2496 class SpaceClosureRegionClosure: public HeapRegionClosure {
2497 2497 SpaceClosure* _cl;
2498 2498 public:
2499 2499 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2500 2500 bool doHeapRegion(HeapRegion* r) {
2501 2501 _cl->do_space(r);
2502 2502 return false;
2503 2503 }
2504 2504 };
2505 2505
2506 2506 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2507 2507 SpaceClosureRegionClosure blk(cl);
2508 2508 heap_region_iterate(&blk);
2509 2509 }
2510 2510
2511 2511 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2512 2512 _hrs.iterate(cl);
2513 2513 }
2514 2514
2515 2515 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r,
2516 2516 HeapRegionClosure* cl) const {
2517 2517 _hrs.iterate_from(r, cl);
2518 2518 }
2519 2519
2520 2520 void
2521 2521 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2522 2522 int worker,
2523 2523 int no_of_par_workers,
2524 2524 jint claim_value) {
2525 2525 const size_t regions = n_regions();
2526 2526 const size_t max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2527 2527 no_of_par_workers :
2528 2528 1);
2529 2529 assert(UseDynamicNumberOfGCThreads ||
2530 2530 no_of_par_workers == workers()->total_workers(),
2531 2531 "Non dynamic should use fixed number of workers");
2532 2532 // try to spread out the starting points of the workers
2533 2533 const size_t start_index = regions / max_workers * (size_t) worker;
2534 2534
2535 2535 // each worker will actually look at all regions
2536 2536 for (size_t count = 0; count < regions; ++count) {
2537 2537 const size_t index = (start_index + count) % regions;
2538 2538 assert(0 <= index && index < regions, "sanity");
2539 2539 HeapRegion* r = region_at(index);
2540 2540 // we'll ignore "continues humongous" regions (we'll process them
2541 2541 // when we come across their corresponding "start humongous"
2542 2542 // region) and regions already claimed
2543 2543 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2544 2544 continue;
2545 2545 }
2546 2546 // OK, try to claim it
2547 2547 if (r->claimHeapRegion(claim_value)) {
2548 2548 // success!
2549 2549 assert(!r->continuesHumongous(), "sanity");
2550 2550 if (r->startsHumongous()) {
2551 2551 // If the region is "starts humongous" we'll iterate over its
2552 2552 // "continues humongous" first; in fact we'll do them
2553 2553 // first. The order is important. In on case, calling the
2554 2554 // closure on the "starts humongous" region might de-allocate
2555 2555 // and clear all its "continues humongous" regions and, as a
2556 2556 // result, we might end up processing them twice. So, we'll do
2557 2557 // them first (notice: most closures will ignore them anyway) and
2558 2558 // then we'll do the "starts humongous" region.
2559 2559 for (size_t ch_index = index + 1; ch_index < regions; ++ch_index) {
2560 2560 HeapRegion* chr = region_at(ch_index);
2561 2561
2562 2562 // if the region has already been claimed or it's not
2563 2563 // "continues humongous" we're done
2564 2564 if (chr->claim_value() == claim_value ||
2565 2565 !chr->continuesHumongous()) {
2566 2566 break;
2567 2567 }
2568 2568
2569 2569 // Noone should have claimed it directly. We can given
2570 2570 // that we claimed its "starts humongous" region.
2571 2571 assert(chr->claim_value() != claim_value, "sanity");
2572 2572 assert(chr->humongous_start_region() == r, "sanity");
2573 2573
2574 2574 if (chr->claimHeapRegion(claim_value)) {
2575 2575 // we should always be able to claim it; noone else should
2576 2576 // be trying to claim this region
2577 2577
2578 2578 bool res2 = cl->doHeapRegion(chr);
2579 2579 assert(!res2, "Should not abort");
2580 2580
2581 2581 // Right now, this holds (i.e., no closure that actually
2582 2582 // does something with "continues humongous" regions
2583 2583 // clears them). We might have to weaken it in the future,
2584 2584 // but let's leave these two asserts here for extra safety.
2585 2585 assert(chr->continuesHumongous(), "should still be the case");
2586 2586 assert(chr->humongous_start_region() == r, "sanity");
2587 2587 } else {
2588 2588 guarantee(false, "we should not reach here");
2589 2589 }
2590 2590 }
2591 2591 }
2592 2592
2593 2593 assert(!r->continuesHumongous(), "sanity");
2594 2594 bool res = cl->doHeapRegion(r);
2595 2595 assert(!res, "Should not abort");
2596 2596 }
2597 2597 }
2598 2598 }
2599 2599
2600 2600 class ResetClaimValuesClosure: public HeapRegionClosure {
2601 2601 public:
2602 2602 bool doHeapRegion(HeapRegion* r) {
2603 2603 r->set_claim_value(HeapRegion::InitialClaimValue);
2604 2604 return false;
2605 2605 }
2606 2606 };
2607 2607
2608 2608 void
2609 2609 G1CollectedHeap::reset_heap_region_claim_values() {
2610 2610 ResetClaimValuesClosure blk;
2611 2611 heap_region_iterate(&blk);
2612 2612 }
2613 2613
2614 2614 #ifdef ASSERT
2615 2615 // This checks whether all regions in the heap have the correct claim
2616 2616 // value. I also piggy-backed on this a check to ensure that the
2617 2617 // humongous_start_region() information on "continues humongous"
2618 2618 // regions is correct.
2619 2619
2620 2620 class CheckClaimValuesClosure : public HeapRegionClosure {
2621 2621 private:
2622 2622 jint _claim_value;
2623 2623 size_t _failures;
2624 2624 HeapRegion* _sh_region;
2625 2625 public:
2626 2626 CheckClaimValuesClosure(jint claim_value) :
2627 2627 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2628 2628 bool doHeapRegion(HeapRegion* r) {
2629 2629 if (r->claim_value() != _claim_value) {
2630 2630 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2631 2631 "claim value = %d, should be %d",
2632 2632 HR_FORMAT_PARAMS(r),
2633 2633 r->claim_value(), _claim_value);
2634 2634 ++_failures;
2635 2635 }
2636 2636 if (!r->isHumongous()) {
2637 2637 _sh_region = NULL;
2638 2638 } else if (r->startsHumongous()) {
2639 2639 _sh_region = r;
2640 2640 } else if (r->continuesHumongous()) {
2641 2641 if (r->humongous_start_region() != _sh_region) {
2642 2642 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2643 2643 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2644 2644 HR_FORMAT_PARAMS(r),
2645 2645 r->humongous_start_region(),
2646 2646 _sh_region);
2647 2647 ++_failures;
2648 2648 }
2649 2649 }
2650 2650 return false;
2651 2651 }
2652 2652 size_t failures() {
2653 2653 return _failures;
2654 2654 }
2655 2655 };
2656 2656
2657 2657 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2658 2658 CheckClaimValuesClosure cl(claim_value);
2659 2659 heap_region_iterate(&cl);
2660 2660 return cl.failures() == 0;
2661 2661 }
2662 2662
2663 2663 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2664 2664 jint _claim_value;
2665 2665 size_t _failures;
2666 2666
2667 2667 public:
2668 2668 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2669 2669 _claim_value(claim_value),
2670 2670 _failures(0) { }
2671 2671
2672 2672 size_t failures() {
2673 2673 return _failures;
2674 2674 }
2675 2675
2676 2676 bool doHeapRegion(HeapRegion* hr) {
2677 2677 assert(hr->in_collection_set(), "how?");
2678 2678 assert(!hr->isHumongous(), "H-region in CSet");
2679 2679 if (hr->claim_value() != _claim_value) {
2680 2680 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2681 2681 "claim value = %d, should be %d",
2682 2682 HR_FORMAT_PARAMS(hr),
2683 2683 hr->claim_value(), _claim_value);
2684 2684 _failures += 1;
2685 2685 }
2686 2686 return false;
2687 2687 }
2688 2688 };
2689 2689
2690 2690 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2691 2691 CheckClaimValuesInCSetHRClosure cl(claim_value);
2692 2692 collection_set_iterate(&cl);
2693 2693 return cl.failures() == 0;
2694 2694 }
2695 2695 #endif // ASSERT
2696 2696
2697 2697 // Clear the cached CSet starting regions and (more importantly)
2698 2698 // the time stamps. Called when we reset the GC time stamp.
2699 2699 void G1CollectedHeap::clear_cset_start_regions() {
2700 2700 assert(_worker_cset_start_region != NULL, "sanity");
2701 2701 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2702 2702
2703 2703 int n_queues = MAX2((int)ParallelGCThreads, 1);
2704 2704 for (int i = 0; i < n_queues; i++) {
2705 2705 _worker_cset_start_region[i] = NULL;
2706 2706 _worker_cset_start_region_time_stamp[i] = 0;
2707 2707 }
2708 2708 }
2709 2709
2710 2710 // Given the id of a worker, obtain or calculate a suitable
2711 2711 // starting region for iterating over the current collection set.
2712 2712 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2713 2713 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2714 2714
2715 2715 HeapRegion* result = NULL;
2716 2716 unsigned gc_time_stamp = get_gc_time_stamp();
2717 2717
2718 2718 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2719 2719 // Cached starting region for current worker was set
2720 2720 // during the current pause - so it's valid.
2721 2721 // Note: the cached starting heap region may be NULL
2722 2722 // (when the collection set is empty).
2723 2723 result = _worker_cset_start_region[worker_i];
2724 2724 assert(result == NULL || result->in_collection_set(), "sanity");
2725 2725 return result;
2726 2726 }
2727 2727
2728 2728 // The cached entry was not valid so let's calculate
2729 2729 // a suitable starting heap region for this worker.
2730 2730
2731 2731 // We want the parallel threads to start their collection
2732 2732 // set iteration at different collection set regions to
2733 2733 // avoid contention.
2734 2734 // If we have:
2735 2735 // n collection set regions
2736 2736 // p threads
2737 2737 // Then thread t will start at region floor ((t * n) / p)
2738 2738
2739 2739 result = g1_policy()->collection_set();
2740 2740 if (G1CollectedHeap::use_parallel_gc_threads()) {
2741 2741 size_t cs_size = g1_policy()->cset_region_length();
2742 2742 int active_workers = workers()->active_workers();
2743 2743 assert(UseDynamicNumberOfGCThreads ||
2744 2744 active_workers == workers()->total_workers(),
2745 2745 "Unless dynamic should use total workers");
2746 2746
2747 2747 size_t end_ind = (cs_size * worker_i) / active_workers;
2748 2748 size_t start_ind = 0;
2749 2749
2750 2750 if (worker_i > 0 &&
2751 2751 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2752 2752 // Previous workers starting region is valid
2753 2753 // so let's iterate from there
2754 2754 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2755 2755 result = _worker_cset_start_region[worker_i - 1];
2756 2756 }
2757 2757
2758 2758 for (size_t i = start_ind; i < end_ind; i++) {
2759 2759 result = result->next_in_collection_set();
2760 2760 }
2761 2761 }
2762 2762
2763 2763 // Note: the calculated starting heap region may be NULL
2764 2764 // (when the collection set is empty).
2765 2765 assert(result == NULL || result->in_collection_set(), "sanity");
2766 2766 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2767 2767 "should be updated only once per pause");
2768 2768 _worker_cset_start_region[worker_i] = result;
2769 2769 OrderAccess::storestore();
2770 2770 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2771 2771 return result;
2772 2772 }
2773 2773
2774 2774 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2775 2775 HeapRegion* r = g1_policy()->collection_set();
2776 2776 while (r != NULL) {
2777 2777 HeapRegion* next = r->next_in_collection_set();
2778 2778 if (cl->doHeapRegion(r)) {
2779 2779 cl->incomplete();
2780 2780 return;
2781 2781 }
2782 2782 r = next;
2783 2783 }
2784 2784 }
2785 2785
2786 2786 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2787 2787 HeapRegionClosure *cl) {
2788 2788 if (r == NULL) {
2789 2789 // The CSet is empty so there's nothing to do.
2790 2790 return;
2791 2791 }
2792 2792
2793 2793 assert(r->in_collection_set(),
2794 2794 "Start region must be a member of the collection set.");
2795 2795 HeapRegion* cur = r;
2796 2796 while (cur != NULL) {
2797 2797 HeapRegion* next = cur->next_in_collection_set();
2798 2798 if (cl->doHeapRegion(cur) && false) {
2799 2799 cl->incomplete();
2800 2800 return;
2801 2801 }
2802 2802 cur = next;
2803 2803 }
2804 2804 cur = g1_policy()->collection_set();
2805 2805 while (cur != r) {
2806 2806 HeapRegion* next = cur->next_in_collection_set();
2807 2807 if (cl->doHeapRegion(cur) && false) {
2808 2808 cl->incomplete();
2809 2809 return;
2810 2810 }
2811 2811 cur = next;
2812 2812 }
2813 2813 }
2814 2814
2815 2815 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2816 2816 return n_regions() > 0 ? region_at(0) : NULL;
2817 2817 }
2818 2818
2819 2819
2820 2820 Space* G1CollectedHeap::space_containing(const void* addr) const {
2821 2821 Space* res = heap_region_containing(addr);
2822 2822 if (res == NULL)
2823 2823 res = perm_gen()->space_containing(addr);
2824 2824 return res;
2825 2825 }
2826 2826
2827 2827 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2828 2828 Space* sp = space_containing(addr);
2829 2829 if (sp != NULL) {
2830 2830 return sp->block_start(addr);
2831 2831 }
2832 2832 return NULL;
2833 2833 }
2834 2834
2835 2835 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2836 2836 Space* sp = space_containing(addr);
2837 2837 assert(sp != NULL, "block_size of address outside of heap");
2838 2838 return sp->block_size(addr);
2839 2839 }
2840 2840
2841 2841 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2842 2842 Space* sp = space_containing(addr);
2843 2843 return sp->block_is_obj(addr);
2844 2844 }
2845 2845
2846 2846 bool G1CollectedHeap::supports_tlab_allocation() const {
2847 2847 return true;
2848 2848 }
2849 2849
2850 2850 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2851 2851 return HeapRegion::GrainBytes;
2852 2852 }
2853 2853
2854 2854 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2855 2855 // Return the remaining space in the cur alloc region, but not less than
2856 2856 // the min TLAB size.
2857 2857
2858 2858 // Also, this value can be at most the humongous object threshold,
2859 2859 // since we can't allow tlabs to grow big enough to accomodate
2860 2860 // humongous objects.
2861 2861
2862 2862 HeapRegion* hr = _mutator_alloc_region.get();
2863 2863 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2864 2864 if (hr == NULL) {
2865 2865 return max_tlab_size;
2866 2866 } else {
2867 2867 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2868 2868 }
2869 2869 }
2870 2870
2871 2871 size_t G1CollectedHeap::max_capacity() const {
2872 2872 return _g1_reserved.byte_size();
2873 2873 }
2874 2874
2875 2875 jlong G1CollectedHeap::millis_since_last_gc() {
2876 2876 // assert(false, "NYI");
2877 2877 return 0;
2878 2878 }
2879 2879
2880 2880 void G1CollectedHeap::prepare_for_verify() {
2881 2881 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2882 2882 ensure_parsability(false);
2883 2883 }
2884 2884 g1_rem_set()->prepare_for_verify();
2885 2885 }
2886 2886
2887 2887 class VerifyLivenessOopClosure: public OopClosure {
2888 2888 G1CollectedHeap* _g1h;
2889 2889 VerifyOption _vo;
2890 2890 public:
2891 2891 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
2892 2892 _g1h(g1h), _vo(vo)
2893 2893 { }
2894 2894 void do_oop(narrowOop *p) { do_oop_work(p); }
2895 2895 void do_oop( oop *p) { do_oop_work(p); }
2896 2896
2897 2897 template <class T> void do_oop_work(T *p) {
2898 2898 oop obj = oopDesc::load_decode_heap_oop(p);
2899 2899 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
2900 2900 "Dead object referenced by a not dead object");
2901 2901 }
2902 2902 };
2903 2903
2904 2904 class VerifyObjsInRegionClosure: public ObjectClosure {
2905 2905 private:
2906 2906 G1CollectedHeap* _g1h;
2907 2907 size_t _live_bytes;
2908 2908 HeapRegion *_hr;
2909 2909 VerifyOption _vo;
2910 2910 public:
2911 2911 // _vo == UsePrevMarking -> use "prev" marking information,
2912 2912 // _vo == UseNextMarking -> use "next" marking information,
2913 2913 // _vo == UseMarkWord -> use mark word from object header.
2914 2914 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
2915 2915 : _live_bytes(0), _hr(hr), _vo(vo) {
2916 2916 _g1h = G1CollectedHeap::heap();
2917 2917 }
2918 2918 void do_object(oop o) {
2919 2919 VerifyLivenessOopClosure isLive(_g1h, _vo);
2920 2920 assert(o != NULL, "Huh?");
2921 2921 if (!_g1h->is_obj_dead_cond(o, _vo)) {
2922 2922 // If the object is alive according to the mark word,
2923 2923 // then verify that the marking information agrees.
2924 2924 // Note we can't verify the contra-positive of the
2925 2925 // above: if the object is dead (according to the mark
2926 2926 // word), it may not be marked, or may have been marked
2927 2927 // but has since became dead, or may have been allocated
2928 2928 // since the last marking.
2929 2929 if (_vo == VerifyOption_G1UseMarkWord) {
2930 2930 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
2931 2931 }
2932 2932
2933 2933 o->oop_iterate(&isLive);
2934 2934 if (!_hr->obj_allocated_since_prev_marking(o)) {
2935 2935 size_t obj_size = o->size(); // Make sure we don't overflow
2936 2936 _live_bytes += (obj_size * HeapWordSize);
2937 2937 }
2938 2938 }
2939 2939 }
2940 2940 size_t live_bytes() { return _live_bytes; }
2941 2941 };
2942 2942
2943 2943 class PrintObjsInRegionClosure : public ObjectClosure {
2944 2944 HeapRegion *_hr;
2945 2945 G1CollectedHeap *_g1;
2946 2946 public:
2947 2947 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
2948 2948 _g1 = G1CollectedHeap::heap();
2949 2949 };
2950 2950
2951 2951 void do_object(oop o) {
2952 2952 if (o != NULL) {
2953 2953 HeapWord *start = (HeapWord *) o;
2954 2954 size_t word_sz = o->size();
2955 2955 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
2956 2956 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
2957 2957 (void*) o, word_sz,
2958 2958 _g1->isMarkedPrev(o),
2959 2959 _g1->isMarkedNext(o),
2960 2960 _hr->obj_allocated_since_prev_marking(o));
2961 2961 HeapWord *end = start + word_sz;
2962 2962 HeapWord *cur;
2963 2963 int *val;
2964 2964 for (cur = start; cur < end; cur++) {
2965 2965 val = (int *) cur;
2966 2966 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
2967 2967 }
2968 2968 }
2969 2969 }
2970 2970 };
2971 2971
2972 2972 class VerifyRegionClosure: public HeapRegionClosure {
2973 2973 private:
2974 2974 bool _allow_dirty;
2975 2975 bool _par;
2976 2976 VerifyOption _vo;
2977 2977 bool _failures;
2978 2978 public:
2979 2979 // _vo == UsePrevMarking -> use "prev" marking information,
2980 2980 // _vo == UseNextMarking -> use "next" marking information,
2981 2981 // _vo == UseMarkWord -> use mark word from object header.
2982 2982 VerifyRegionClosure(bool allow_dirty, bool par, VerifyOption vo)
2983 2983 : _allow_dirty(allow_dirty),
2984 2984 _par(par),
2985 2985 _vo(vo),
2986 2986 _failures(false) {}
2987 2987
2988 2988 bool failures() {
2989 2989 return _failures;
2990 2990 }
2991 2991
2992 2992 bool doHeapRegion(HeapRegion* r) {
2993 2993 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue,
2994 2994 "Should be unclaimed at verify points.");
2995 2995 if (!r->continuesHumongous()) {
2996 2996 bool failures = false;
2997 2997 r->verify(_allow_dirty, _vo, &failures);
2998 2998 if (failures) {
2999 2999 _failures = true;
3000 3000 } else {
3001 3001 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3002 3002 r->object_iterate(¬_dead_yet_cl);
3003 3003 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3004 3004 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3005 3005 "max_live_bytes "SIZE_FORMAT" "
3006 3006 "< calculated "SIZE_FORMAT,
3007 3007 r->bottom(), r->end(),
3008 3008 r->max_live_bytes(),
3009 3009 not_dead_yet_cl.live_bytes());
3010 3010 _failures = true;
3011 3011 }
3012 3012 }
3013 3013 }
3014 3014 return false; // stop the region iteration if we hit a failure
3015 3015 }
3016 3016 };
3017 3017
3018 3018 class VerifyRootsClosure: public OopsInGenClosure {
3019 3019 private:
3020 3020 G1CollectedHeap* _g1h;
3021 3021 VerifyOption _vo;
3022 3022 bool _failures;
3023 3023 public:
3024 3024 // _vo == UsePrevMarking -> use "prev" marking information,
3025 3025 // _vo == UseNextMarking -> use "next" marking information,
3026 3026 // _vo == UseMarkWord -> use mark word from object header.
3027 3027 VerifyRootsClosure(VerifyOption vo) :
3028 3028 _g1h(G1CollectedHeap::heap()),
3029 3029 _vo(vo),
3030 3030 _failures(false) { }
3031 3031
3032 3032 bool failures() { return _failures; }
3033 3033
3034 3034 template <class T> void do_oop_nv(T* p) {
3035 3035 T heap_oop = oopDesc::load_heap_oop(p);
3036 3036 if (!oopDesc::is_null(heap_oop)) {
3037 3037 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3038 3038 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3039 3039 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3040 3040 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3041 3041 if (_vo == VerifyOption_G1UseMarkWord) {
3042 3042 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3043 3043 }
3044 3044 obj->print_on(gclog_or_tty);
3045 3045 _failures = true;
3046 3046 }
3047 3047 }
3048 3048 }
3049 3049
3050 3050 void do_oop(oop* p) { do_oop_nv(p); }
3051 3051 void do_oop(narrowOop* p) { do_oop_nv(p); }
3052 3052 };
3053 3053
3054 3054 // This is the task used for parallel heap verification.
3055 3055
3056 3056 class G1ParVerifyTask: public AbstractGangTask {
3057 3057 private:
3058 3058 G1CollectedHeap* _g1h;
3059 3059 bool _allow_dirty;
3060 3060 VerifyOption _vo;
3061 3061 bool _failures;
3062 3062
3063 3063 public:
3064 3064 // _vo == UsePrevMarking -> use "prev" marking information,
3065 3065 // _vo == UseNextMarking -> use "next" marking information,
3066 3066 // _vo == UseMarkWord -> use mark word from object header.
3067 3067 G1ParVerifyTask(G1CollectedHeap* g1h, bool allow_dirty, VerifyOption vo) :
3068 3068 AbstractGangTask("Parallel verify task"),
3069 3069 _g1h(g1h),
3070 3070 _allow_dirty(allow_dirty),
3071 3071 _vo(vo),
3072 3072 _failures(false) { }
3073 3073
3074 3074 bool failures() {
3075 3075 return _failures;
3076 3076 }
3077 3077
3078 3078 void work(int worker_i) {
3079 3079 HandleMark hm;
3080 3080 VerifyRegionClosure blk(_allow_dirty, true, _vo);
3081 3081 _g1h->heap_region_par_iterate_chunked(&blk, worker_i,
3082 3082 _g1h->workers()->active_workers(),
3083 3083 HeapRegion::ParVerifyClaimValue);
3084 3084 if (blk.failures()) {
3085 3085 _failures = true;
3086 3086 }
3087 3087 }
3088 3088 };
3089 3089
3090 3090 void G1CollectedHeap::verify(bool allow_dirty, bool silent) {
3091 3091 verify(allow_dirty, silent, VerifyOption_G1UsePrevMarking);
3092 3092 }
3093 3093
3094 3094 void G1CollectedHeap::verify(bool allow_dirty,
3095 3095 bool silent,
3096 3096 VerifyOption vo) {
3097 3097 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3098 3098 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); }
3099 3099 VerifyRootsClosure rootsCl(vo);
3100 3100
3101 3101 assert(Thread::current()->is_VM_thread(),
3102 3102 "Expected to be executed serially by the VM thread at this point");
3103 3103
3104 3104 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3105 3105
3106 3106 // We apply the relevant closures to all the oops in the
3107 3107 // system dictionary, the string table and the code cache.
3108 3108 const int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3109 3109
3110 3110 process_strong_roots(true, // activate StrongRootsScope
3111 3111 true, // we set "collecting perm gen" to true,
3112 3112 // so we don't reset the dirty cards in the perm gen.
3113 3113 SharedHeap::ScanningOption(so), // roots scanning options
3114 3114 &rootsCl,
3115 3115 &blobsCl,
3116 3116 &rootsCl);
3117 3117
3118 3118 // If we're verifying after the marking phase of a Full GC then we can't
3119 3119 // treat the perm gen as roots into the G1 heap. Some of the objects in
3120 3120 // the perm gen may be dead and hence not marked. If one of these dead
3121 3121 // objects is considered to be a root then we may end up with a false
3122 3122 // "Root location <x> points to dead ob <y>" failure.
3123 3123 if (vo != VerifyOption_G1UseMarkWord) {
3124 3124 // Since we used "collecting_perm_gen" == true above, we will not have
3125 3125 // checked the refs from perm into the G1-collected heap. We check those
3126 3126 // references explicitly below. Whether the relevant cards are dirty
3127 3127 // is checked further below in the rem set verification.
3128 3128 if (!silent) { gclog_or_tty->print("Permgen roots "); }
3129 3129 perm_gen()->oop_iterate(&rootsCl);
3130 3130 }
3131 3131 bool failures = rootsCl.failures();
3132 3132
3133 3133 if (vo != VerifyOption_G1UseMarkWord) {
3134 3134 // If we're verifying during a full GC then the region sets
3135 3135 // will have been torn down at the start of the GC. Therefore
3136 3136 // verifying the region sets will fail. So we only verify
3137 3137 // the region sets when not in a full GC.
3138 3138 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3139 3139 verify_region_sets();
3140 3140 }
3141 3141
3142 3142 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3143 3143 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3144 3144 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3145 3145 "sanity check");
3146 3146
3147 3147 G1ParVerifyTask task(this, allow_dirty, vo);
3148 3148 assert(UseDynamicNumberOfGCThreads ||
3149 3149 workers()->active_workers() == workers()->total_workers(),
3150 3150 "If not dynamic should be using all the workers");
3151 3151 int n_workers = workers()->active_workers();
3152 3152 set_par_threads(n_workers);
3153 3153 workers()->run_task(&task);
3154 3154 set_par_threads(0);
3155 3155 if (task.failures()) {
3156 3156 failures = true;
3157 3157 }
3158 3158
3159 3159 // Checks that the expected amount of parallel work was done.
3160 3160 // The implication is that n_workers is > 0.
3161 3161 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3162 3162 "sanity check");
3163 3163
3164 3164 reset_heap_region_claim_values();
3165 3165
3166 3166 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3167 3167 "sanity check");
3168 3168 } else {
3169 3169 VerifyRegionClosure blk(allow_dirty, false, vo);
3170 3170 heap_region_iterate(&blk);
3171 3171 if (blk.failures()) {
3172 3172 failures = true;
3173 3173 }
3174 3174 }
3175 3175 if (!silent) gclog_or_tty->print("RemSet ");
3176 3176 rem_set()->verify();
3177 3177
3178 3178 if (failures) {
3179 3179 gclog_or_tty->print_cr("Heap:");
3180 3180 // It helps to have the per-region information in the output to
3181 3181 // help us track down what went wrong. This is why we call
3182 3182 // print_extended_on() instead of print_on().
3183 3183 print_extended_on(gclog_or_tty);
3184 3184 gclog_or_tty->print_cr("");
3185 3185 #ifndef PRODUCT
3186 3186 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3187 3187 concurrent_mark()->print_reachable("at-verification-failure",
3188 3188 vo, false /* all */);
3189 3189 }
3190 3190 #endif
3191 3191 gclog_or_tty->flush();
3192 3192 }
3193 3193 guarantee(!failures, "there should not have been any failures");
3194 3194 } else {
3195 3195 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3196 3196 }
3197 3197 }
3198 3198
3199 3199 class PrintRegionClosure: public HeapRegionClosure {
3200 3200 outputStream* _st;
3201 3201 public:
3202 3202 PrintRegionClosure(outputStream* st) : _st(st) {}
3203 3203 bool doHeapRegion(HeapRegion* r) {
3204 3204 r->print_on(_st);
3205 3205 return false;
3206 3206 }
3207 3207 };
3208 3208
3209 3209 void G1CollectedHeap::print_on(outputStream* st) const {
3210 3210 st->print(" %-20s", "garbage-first heap");
3211 3211 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3212 3212 capacity()/K, used_unlocked()/K);
3213 3213 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3214 3214 _g1_storage.low_boundary(),
3215 3215 _g1_storage.high(),
3216 3216 _g1_storage.high_boundary());
3217 3217 st->cr();
3218 3218 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3219 3219 size_t young_regions = _young_list->length();
3220 3220 st->print(SIZE_FORMAT " young (" SIZE_FORMAT "K), ",
3221 3221 young_regions, young_regions * HeapRegion::GrainBytes / K);
3222 3222 size_t survivor_regions = g1_policy()->recorded_survivor_regions();
3223 3223 st->print(SIZE_FORMAT " survivors (" SIZE_FORMAT "K)",
3224 3224 survivor_regions, survivor_regions * HeapRegion::GrainBytes / K);
3225 3225 st->cr();
3226 3226 perm()->as_gen()->print_on(st);
3227 3227 }
3228 3228
3229 3229 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3230 3230 print_on(st);
3231 3231
3232 3232 // Print the per-region information.
3233 3233 st->cr();
3234 3234 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), HS=humongous(starts), HC=humongous(continues), CS=collection set, F=free, TS=gc time stamp, PTAMS=previous top-at-mark-start, NTAMS=next top-at-mark-start)");
3235 3235 PrintRegionClosure blk(st);
3236 3236 heap_region_iterate(&blk);
3237 3237 }
3238 3238
3239 3239 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3240 3240 if (G1CollectedHeap::use_parallel_gc_threads()) {
3241 3241 workers()->print_worker_threads_on(st);
3242 3242 }
3243 3243 _cmThread->print_on(st);
3244 3244 st->cr();
3245 3245 _cm->print_worker_threads_on(st);
3246 3246 _cg1r->print_worker_threads_on(st);
3247 3247 st->cr();
3248 3248 }
3249 3249
3250 3250 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3251 3251 if (G1CollectedHeap::use_parallel_gc_threads()) {
3252 3252 workers()->threads_do(tc);
3253 3253 }
3254 3254 tc->do_thread(_cmThread);
3255 3255 _cg1r->threads_do(tc);
3256 3256 }
3257 3257
3258 3258 void G1CollectedHeap::print_tracing_info() const {
3259 3259 // We'll overload this to mean "trace GC pause statistics."
3260 3260 if (TraceGen0Time || TraceGen1Time) {
3261 3261 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3262 3262 // to that.
3263 3263 g1_policy()->print_tracing_info();
3264 3264 }
3265 3265 if (G1SummarizeRSetStats) {
3266 3266 g1_rem_set()->print_summary_info();
3267 3267 }
3268 3268 if (G1SummarizeConcMark) {
3269 3269 concurrent_mark()->print_summary_info();
3270 3270 }
3271 3271 g1_policy()->print_yg_surv_rate_info();
3272 3272 SpecializationStats::print();
3273 3273 }
3274 3274
3275 3275 #ifndef PRODUCT
3276 3276 // Helpful for debugging RSet issues.
3277 3277
3278 3278 class PrintRSetsClosure : public HeapRegionClosure {
3279 3279 private:
3280 3280 const char* _msg;
3281 3281 size_t _occupied_sum;
3282 3282
3283 3283 public:
3284 3284 bool doHeapRegion(HeapRegion* r) {
3285 3285 HeapRegionRemSet* hrrs = r->rem_set();
3286 3286 size_t occupied = hrrs->occupied();
3287 3287 _occupied_sum += occupied;
3288 3288
3289 3289 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3290 3290 HR_FORMAT_PARAMS(r));
3291 3291 if (occupied == 0) {
3292 3292 gclog_or_tty->print_cr(" RSet is empty");
3293 3293 } else {
3294 3294 hrrs->print();
3295 3295 }
3296 3296 gclog_or_tty->print_cr("----------");
3297 3297 return false;
3298 3298 }
3299 3299
3300 3300 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3301 3301 gclog_or_tty->cr();
3302 3302 gclog_or_tty->print_cr("========================================");
3303 3303 gclog_or_tty->print_cr(msg);
3304 3304 gclog_or_tty->cr();
3305 3305 }
3306 3306
3307 3307 ~PrintRSetsClosure() {
3308 3308 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3309 3309 gclog_or_tty->print_cr("========================================");
3310 3310 gclog_or_tty->cr();
3311 3311 }
3312 3312 };
3313 3313
3314 3314 void G1CollectedHeap::print_cset_rsets() {
3315 3315 PrintRSetsClosure cl("Printing CSet RSets");
3316 3316 collection_set_iterate(&cl);
3317 3317 }
3318 3318
3319 3319 void G1CollectedHeap::print_all_rsets() {
3320 3320 PrintRSetsClosure cl("Printing All RSets");;
3321 3321 heap_region_iterate(&cl);
3322 3322 }
3323 3323 #endif // PRODUCT
3324 3324
3325 3325 G1CollectedHeap* G1CollectedHeap::heap() {
3326 3326 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3327 3327 "not a garbage-first heap");
3328 3328 return _g1h;
3329 3329 }
3330 3330
3331 3331 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3332 3332 // always_do_update_barrier = false;
3333 3333 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3334 3334 // Call allocation profiler
3335 3335 AllocationProfiler::iterate_since_last_gc();
3336 3336 // Fill TLAB's and such
3337 3337 ensure_parsability(true);
3338 3338 }
3339 3339
3340 3340 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3341 3341 // FIXME: what is this about?
3342 3342 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3343 3343 // is set.
3344 3344 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3345 3345 "derived pointer present"));
3346 3346 // always_do_update_barrier = true;
3347 3347
3348 3348 // We have just completed a GC. Update the soft reference
3349 3349 // policy with the new heap occupancy
3350 3350 Universe::update_heap_info_at_gc();
3351 3351 }
3352 3352
3353 3353 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3354 3354 unsigned int gc_count_before,
3355 3355 bool* succeeded) {
3356 3356 assert_heap_not_locked_and_not_at_safepoint();
3357 3357 g1_policy()->record_stop_world_start();
3358 3358 VM_G1IncCollectionPause op(gc_count_before,
3359 3359 word_size,
3360 3360 false, /* should_initiate_conc_mark */
3361 3361 g1_policy()->max_pause_time_ms(),
3362 3362 GCCause::_g1_inc_collection_pause);
3363 3363 VMThread::execute(&op);
3364 3364
3365 3365 HeapWord* result = op.result();
3366 3366 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3367 3367 assert(result == NULL || ret_succeeded,
3368 3368 "the result should be NULL if the VM did not succeed");
3369 3369 *succeeded = ret_succeeded;
3370 3370
3371 3371 assert_heap_not_locked();
3372 3372 return result;
3373 3373 }
3374 3374
3375 3375 void
3376 3376 G1CollectedHeap::doConcurrentMark() {
3377 3377 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3378 3378 if (!_cmThread->in_progress()) {
3379 3379 _cmThread->set_started();
3380 3380 CGC_lock->notify();
3381 3381 }
3382 3382 }
3383 3383
3384 3384 double G1CollectedHeap::predict_region_elapsed_time_ms(HeapRegion *hr,
3385 3385 bool young) {
3386 3386 return _g1_policy->predict_region_elapsed_time_ms(hr, young);
3387 3387 }
3388 3388
3389 3389 void G1CollectedHeap::check_if_region_is_too_expensive(double
3390 3390 predicted_time_ms) {
3391 3391 _g1_policy->check_if_region_is_too_expensive(predicted_time_ms);
3392 3392 }
3393 3393
3394 3394 size_t G1CollectedHeap::pending_card_num() {
3395 3395 size_t extra_cards = 0;
3396 3396 JavaThread *curr = Threads::first();
3397 3397 while (curr != NULL) {
3398 3398 DirtyCardQueue& dcq = curr->dirty_card_queue();
3399 3399 extra_cards += dcq.size();
3400 3400 curr = curr->next();
3401 3401 }
3402 3402 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3403 3403 size_t buffer_size = dcqs.buffer_size();
3404 3404 size_t buffer_num = dcqs.completed_buffers_num();
3405 3405 return buffer_size * buffer_num + extra_cards;
3406 3406 }
3407 3407
3408 3408 size_t G1CollectedHeap::max_pending_card_num() {
3409 3409 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3410 3410 size_t buffer_size = dcqs.buffer_size();
3411 3411 size_t buffer_num = dcqs.completed_buffers_num();
3412 3412 int thread_num = Threads::number_of_threads();
3413 3413 return (buffer_num + thread_num) * buffer_size;
3414 3414 }
3415 3415
3416 3416 size_t G1CollectedHeap::cards_scanned() {
3417 3417 return g1_rem_set()->cardsScanned();
3418 3418 }
3419 3419
3420 3420 void
3421 3421 G1CollectedHeap::setup_surviving_young_words() {
3422 3422 guarantee( _surviving_young_words == NULL, "pre-condition" );
3423 3423 size_t array_length = g1_policy()->young_cset_region_length();
3424 3424 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, array_length);
3425 3425 if (_surviving_young_words == NULL) {
3426 3426 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3427 3427 "Not enough space for young surv words summary.");
3428 3428 }
3429 3429 memset(_surviving_young_words, 0, array_length * sizeof(size_t));
3430 3430 #ifdef ASSERT
3431 3431 for (size_t i = 0; i < array_length; ++i) {
3432 3432 assert( _surviving_young_words[i] == 0, "memset above" );
3433 3433 }
3434 3434 #endif // !ASSERT
3435 3435 }
3436 3436
3437 3437 void
3438 3438 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3439 3439 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3440 3440 size_t array_length = g1_policy()->young_cset_region_length();
3441 3441 for (size_t i = 0; i < array_length; ++i)
3442 3442 _surviving_young_words[i] += surv_young_words[i];
3443 3443 }
3444 3444
3445 3445 void
3446 3446 G1CollectedHeap::cleanup_surviving_young_words() {
3447 3447 guarantee( _surviving_young_words != NULL, "pre-condition" );
3448 3448 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words);
3449 3449 _surviving_young_words = NULL;
3450 3450 }
3451 3451
3452 3452 #ifdef ASSERT
3453 3453 class VerifyCSetClosure: public HeapRegionClosure {
3454 3454 public:
3455 3455 bool doHeapRegion(HeapRegion* hr) {
3456 3456 // Here we check that the CSet region's RSet is ready for parallel
3457 3457 // iteration. The fields that we'll verify are only manipulated
3458 3458 // when the region is part of a CSet and is collected. Afterwards,
3459 3459 // we reset these fields when we clear the region's RSet (when the
3460 3460 // region is freed) so they are ready when the region is
3461 3461 // re-allocated. The only exception to this is if there's an
3462 3462 // evacuation failure and instead of freeing the region we leave
3463 3463 // it in the heap. In that case, we reset these fields during
3464 3464 // evacuation failure handling.
3465 3465 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3466 3466
3467 3467 // Here's a good place to add any other checks we'd like to
3468 3468 // perform on CSet regions.
3469 3469 return false;
3470 3470 }
3471 3471 };
3472 3472 #endif // ASSERT
3473 3473
3474 3474 #if TASKQUEUE_STATS
3475 3475 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3476 3476 st->print_raw_cr("GC Task Stats");
3477 3477 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3478 3478 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3479 3479 }
3480 3480
3481 3481 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3482 3482 print_taskqueue_stats_hdr(st);
3483 3483
3484 3484 TaskQueueStats totals;
3485 3485 const int n = workers() != NULL ? workers()->total_workers() : 1;
3486 3486 for (int i = 0; i < n; ++i) {
3487 3487 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3488 3488 totals += task_queue(i)->stats;
3489 3489 }
3490 3490 st->print_raw("tot "); totals.print(st); st->cr();
3491 3491
3492 3492 DEBUG_ONLY(totals.verify());
3493 3493 }
3494 3494
3495 3495 void G1CollectedHeap::reset_taskqueue_stats() {
3496 3496 const int n = workers() != NULL ? workers()->total_workers() : 1;
3497 3497 for (int i = 0; i < n; ++i) {
3498 3498 task_queue(i)->stats.reset();
3499 3499 }
3500 3500 }
3501 3501 #endif // TASKQUEUE_STATS
3502 3502
3503 3503 bool
3504 3504 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3505 3505 assert_at_safepoint(true /* should_be_vm_thread */);
3506 3506 guarantee(!is_gc_active(), "collection is not reentrant");
3507 3507
3508 3508 if (GC_locker::check_active_before_gc()) {
3509 3509 return false;
3510 3510 }
3511 3511
3512 3512 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3513 3513 ResourceMark rm;
3514 3514
3515 3515 if (PrintHeapAtGC) {
3516 3516 Universe::print_heap_before_gc();
3517 3517 }
3518 3518
3519 3519 HRSPhaseSetter x(HRSPhaseEvacuation);
3520 3520 verify_region_sets_optional();
3521 3521 verify_dirty_young_regions();
3522 3522
3523 3523 {
3524 3524 // This call will decide whether this pause is an initial-mark
3525 3525 // pause. If it is, during_initial_mark_pause() will return true
3526 3526 // for the duration of this pause.
3527 3527 g1_policy()->decide_on_conc_mark_initiation();
3528 3528
3529 3529 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3530 3530 assert(!g1_policy()->during_initial_mark_pause() ||
3531 3531 g1_policy()->gcs_are_young(), "sanity");
3532 3532
3533 3533 // We also do not allow mixed GCs during marking.
3534 3534 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3535 3535
3536 3536 char verbose_str[128];
3537 3537 sprintf(verbose_str, "GC pause ");
3538 3538 if (g1_policy()->gcs_are_young()) {
3539 3539 strcat(verbose_str, "(young)");
3540 3540 } else {
3541 3541 strcat(verbose_str, "(mixed)");
3542 3542 }
3543 3543 if (g1_policy()->during_initial_mark_pause()) {
3544 3544 strcat(verbose_str, " (initial-mark)");
3545 3545 // We are about to start a marking cycle, so we increment the
3546 3546 // full collection counter.
3547 3547 increment_total_full_collections();
3548 3548 }
3549 3549
3550 3550 // if PrintGCDetails is on, we'll print long statistics information
3551 3551 // in the collector policy code, so let's not print this as the output
3552 3552 // is messy if we do.
3553 3553 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
3554 3554 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3555 3555 TraceTime t(verbose_str, PrintGC && !PrintGCDetails, true, gclog_or_tty);
3556 3556
3557 3557 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3558 3558 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3559 3559
3560 3560 // If the secondary_free_list is not empty, append it to the
3561 3561 // free_list. No need to wait for the cleanup operation to finish;
3562 3562 // the region allocation code will check the secondary_free_list
3563 3563 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3564 3564 // set, skip this step so that the region allocation code has to
3565 3565 // get entries from the secondary_free_list.
3566 3566 if (!G1StressConcRegionFreeing) {
3567 3567 append_secondary_free_list_if_not_empty_with_lock();
3568 3568 }
3569 3569
3570 3570 assert(check_young_list_well_formed(),
3571 3571 "young list should be well formed");
3572 3572
3573 3573 // Don't dynamically change the number of GC threads this early. A value of
3574 3574 // 0 is used to indicate serial work. When parallel work is done,
3575 3575 // it will be set.
3576 3576
3577 3577 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3578 3578 IsGCActiveMark x;
3579 3579
3580 3580 gc_prologue(false);
3581 3581 increment_total_collections(false /* full gc */);
3582 3582 increment_gc_time_stamp();
3583 3583
3584 3584 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) {
3585 3585 HandleMark hm; // Discard invalid handles created during verification
3586 3586 gclog_or_tty->print(" VerifyBeforeGC:");
3587 3587 prepare_for_verify();
3588 3588 Universe::verify(/* allow dirty */ false,
3589 3589 /* silent */ false,
3590 3590 /* option */ VerifyOption_G1UsePrevMarking);
3591 3591
3592 3592 }
3593 3593
3594 3594 COMPILER2_PRESENT(DerivedPointerTable::clear());
3595 3595
3596 3596 // Please see comment in g1CollectedHeap.hpp and
3597 3597 // G1CollectedHeap::ref_processing_init() to see how
3598 3598 // reference processing currently works in G1.
3599 3599
3600 3600 // Enable discovery in the STW reference processor
3601 3601 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3602 3602 true /*verify_no_refs*/);
3603 3603
3604 3604 {
3605 3605 // We want to temporarily turn off discovery by the
3606 3606 // CM ref processor, if necessary, and turn it back on
3607 3607 // on again later if we do. Using a scoped
3608 3608 // NoRefDiscovery object will do this.
3609 3609 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3610 3610
3611 3611 // Forget the current alloc region (we might even choose it to be part
3612 3612 // of the collection set!).
3613 3613 release_mutator_alloc_region();
3614 3614
3615 3615 // We should call this after we retire the mutator alloc
3616 3616 // region(s) so that all the ALLOC / RETIRE events are generated
3617 3617 // before the start GC event.
3618 3618 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3619 3619
3620 3620 // The elapsed time induced by the start time below deliberately elides
3621 3621 // the possible verification above.
3622 3622 double start_time_sec = os::elapsedTime();
3623 3623 size_t start_used_bytes = used();
3624 3624
3625 3625 #if YOUNG_LIST_VERBOSE
3626 3626 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3627 3627 _young_list->print();
3628 3628 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3629 3629 #endif // YOUNG_LIST_VERBOSE
3630 3630
3631 3631 g1_policy()->record_collection_pause_start(start_time_sec,
3632 3632 start_used_bytes);
3633 3633
3634 3634 #if YOUNG_LIST_VERBOSE
3635 3635 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3636 3636 _young_list->print();
3637 3637 #endif // YOUNG_LIST_VERBOSE
3638 3638
3639 3639 if (g1_policy()->during_initial_mark_pause()) {
3640 3640 concurrent_mark()->checkpointRootsInitialPre();
3641 3641 }
3642 3642 perm_gen()->save_marks();
3643 3643
3644 3644 // We must do this before any possible evacuation that should propagate
3645 3645 // marks.
3646 3646 if (mark_in_progress()) {
3647 3647 double start_time_sec = os::elapsedTime();
3648 3648
3649 3649 _cm->drainAllSATBBuffers();
3650 3650 double finish_mark_ms = (os::elapsedTime() - start_time_sec) * 1000.0;
3651 3651 g1_policy()->record_satb_drain_time(finish_mark_ms);
3652 3652 }
3653 3653 // Record the number of elements currently on the mark stack, so we
3654 3654 // only iterate over these. (Since evacuation may add to the mark
3655 3655 // stack, doing more exposes race conditions.) If no mark is in
3656 3656 // progress, this will be zero.
3657 3657 _cm->set_oops_do_bound();
3658 3658
3659 3659 if (mark_in_progress()) {
3660 3660 concurrent_mark()->newCSet();
3661 3661 }
3662 3662
3663 3663 #if YOUNG_LIST_VERBOSE
3664 3664 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3665 3665 _young_list->print();
3666 3666 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3667 3667 #endif // YOUNG_LIST_VERBOSE
3668 3668
3669 3669 g1_policy()->choose_collection_set(target_pause_time_ms);
3670 3670
3671 3671 if (_hr_printer.is_active()) {
3672 3672 HeapRegion* hr = g1_policy()->collection_set();
3673 3673 while (hr != NULL) {
3674 3674 G1HRPrinter::RegionType type;
3675 3675 if (!hr->is_young()) {
3676 3676 type = G1HRPrinter::Old;
3677 3677 } else if (hr->is_survivor()) {
3678 3678 type = G1HRPrinter::Survivor;
3679 3679 } else {
3680 3680 type = G1HRPrinter::Eden;
3681 3681 }
3682 3682 _hr_printer.cset(hr);
3683 3683 hr = hr->next_in_collection_set();
3684 3684 }
3685 3685 }
3686 3686
3687 3687 // We have chosen the complete collection set. If marking is
3688 3688 // active then, we clear the region fields of any of the
3689 3689 // concurrent marking tasks whose region fields point into
3690 3690 // the collection set as these values will become stale. This
3691 3691 // will cause the owning marking threads to claim a new region
3692 3692 // when marking restarts.
3693 3693 if (mark_in_progress()) {
3694 3694 concurrent_mark()->reset_active_task_region_fields_in_cset();
3695 3695 }
3696 3696
3697 3697 #ifdef ASSERT
3698 3698 VerifyCSetClosure cl;
3699 3699 collection_set_iterate(&cl);
3700 3700 #endif // ASSERT
3701 3701
3702 3702 setup_surviving_young_words();
3703 3703
3704 3704 // Initialize the GC alloc regions.
3705 3705 init_gc_alloc_regions();
3706 3706
3707 3707 // Actually do the work...
3708 3708 evacuate_collection_set();
3709 3709
3710 3710 free_collection_set(g1_policy()->collection_set());
3711 3711 g1_policy()->clear_collection_set();
3712 3712
3713 3713 cleanup_surviving_young_words();
3714 3714
3715 3715 // Start a new incremental collection set for the next pause.
3716 3716 g1_policy()->start_incremental_cset_building();
3717 3717
3718 3718 // Clear the _cset_fast_test bitmap in anticipation of adding
3719 3719 // regions to the incremental collection set for the next
3720 3720 // evacuation pause.
3721 3721 clear_cset_fast_test();
3722 3722
3723 3723 _young_list->reset_sampled_info();
3724 3724
3725 3725 // Don't check the whole heap at this point as the
3726 3726 // GC alloc regions from this pause have been tagged
3727 3727 // as survivors and moved on to the survivor list.
3728 3728 // Survivor regions will fail the !is_young() check.
3729 3729 assert(check_young_list_empty(false /* check_heap */),
3730 3730 "young list should be empty");
3731 3731
3732 3732 #if YOUNG_LIST_VERBOSE
3733 3733 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3734 3734 _young_list->print();
3735 3735 #endif // YOUNG_LIST_VERBOSE
3736 3736
3737 3737 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3738 3738 _young_list->first_survivor_region(),
3739 3739 _young_list->last_survivor_region());
3740 3740
3741 3741 _young_list->reset_auxilary_lists();
3742 3742
3743 3743 if (evacuation_failed()) {
3744 3744 _summary_bytes_used = recalculate_used();
3745 3745 } else {
3746 3746 // The "used" of the the collection set have already been subtracted
3747 3747 // when they were freed. Add in the bytes evacuated.
3748 3748 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3749 3749 }
3750 3750
3751 3751 if (g1_policy()->during_initial_mark_pause()) {
3752 3752 concurrent_mark()->checkpointRootsInitialPost();
3753 3753 set_marking_started();
3754 3754 // CAUTION: after the doConcurrentMark() call below,
3755 3755 // the concurrent marking thread(s) could be running
3756 3756 // concurrently with us. Make sure that anything after
3757 3757 // this point does not assume that we are the only GC thread
3758 3758 // running. Note: of course, the actual marking work will
3759 3759 // not start until the safepoint itself is released in
3760 3760 // ConcurrentGCThread::safepoint_desynchronize().
3761 3761 doConcurrentMark();
3762 3762 }
3763 3763
3764 3764 allocate_dummy_regions();
3765 3765
3766 3766 #if YOUNG_LIST_VERBOSE
3767 3767 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3768 3768 _young_list->print();
3769 3769 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3770 3770 #endif // YOUNG_LIST_VERBOSE
3771 3771
3772 3772 init_mutator_alloc_region();
3773 3773
3774 3774 {
3775 3775 size_t expand_bytes = g1_policy()->expansion_amount();
3776 3776 if (expand_bytes > 0) {
3777 3777 size_t bytes_before = capacity();
3778 3778 if (!expand(expand_bytes)) {
3779 3779 // We failed to expand the heap so let's verify that
3780 3780 // committed/uncommitted amount match the backing store
3781 3781 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3782 3782 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3783 3783 }
3784 3784 }
3785 3785 }
3786 3786
3787 3787 double end_time_sec = os::elapsedTime();
3788 3788 double pause_time_ms = (end_time_sec - start_time_sec) * MILLIUNITS;
3789 3789 g1_policy()->record_pause_time_ms(pause_time_ms);
3790 3790 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3791 3791 workers()->active_workers() : 1);
3792 3792 g1_policy()->record_collection_pause_end(active_workers);
3793 3793
3794 3794 MemoryService::track_memory_usage();
3795 3795
3796 3796 // In prepare_for_verify() below we'll need to scan the deferred
3797 3797 // update buffers to bring the RSets up-to-date if
3798 3798 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
3799 3799 // the update buffers we'll probably need to scan cards on the
3800 3800 // regions we just allocated to (i.e., the GC alloc
3801 3801 // regions). However, during the last GC we called
3802 3802 // set_saved_mark() on all the GC alloc regions, so card
3803 3803 // scanning might skip the [saved_mark_word()...top()] area of
3804 3804 // those regions (i.e., the area we allocated objects into
3805 3805 // during the last GC). But it shouldn't. Given that
3806 3806 // saved_mark_word() is conditional on whether the GC time stamp
3807 3807 // on the region is current or not, by incrementing the GC time
3808 3808 // stamp here we invalidate all the GC time stamps on all the
3809 3809 // regions and saved_mark_word() will simply return top() for
3810 3810 // all the regions. This is a nicer way of ensuring this rather
3811 3811 // than iterating over the regions and fixing them. In fact, the
3812 3812 // GC time stamp increment here also ensures that
3813 3813 // saved_mark_word() will return top() between pauses, i.e.,
3814 3814 // during concurrent refinement. So we don't need the
3815 3815 // is_gc_active() check to decided which top to use when
3816 3816 // scanning cards (see CR 7039627).
3817 3817 increment_gc_time_stamp();
3818 3818
3819 3819 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) {
3820 3820 HandleMark hm; // Discard invalid handles created during verification
3821 3821 gclog_or_tty->print(" VerifyAfterGC:");
3822 3822 prepare_for_verify();
3823 3823 Universe::verify(/* allow dirty */ true,
3824 3824 /* silent */ false,
3825 3825 /* option */ VerifyOption_G1UsePrevMarking);
3826 3826 }
3827 3827
3828 3828 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
3829 3829 ref_processor_stw()->verify_no_references_recorded();
3830 3830
3831 3831 // CM reference discovery will be re-enabled if necessary.
3832 3832 }
3833 3833
3834 3834 {
3835 3835 size_t expand_bytes = g1_policy()->expansion_amount();
3836 3836 if (expand_bytes > 0) {
3837 3837 size_t bytes_before = capacity();
3838 3838 // No need for an ergo verbose message here,
3839 3839 // expansion_amount() does this when it returns a value > 0.
3840 3840 if (!expand(expand_bytes)) {
3841 3841 // We failed to expand the heap so let's verify that
3842 3842 // committed/uncommitted amount match the backing store
3843 3843 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3844 3844 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3845 3845 }
3846 3846 }
3847 3847 }
3848 3848
3849 3849 // We should do this after we potentially expand the heap so
3850 3850 // that all the COMMIT events are generated before the end GC
3851 3851 // event, and after we retire the GC alloc regions so that all
3852 3852 // RETIRE events are generated before the end GC event.
3853 3853 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
3854 3854
3855 3855 // We have to do this after we decide whether to expand the heap or not.
3856 3856 g1_policy()->print_heap_transition();
3857 3857
3858 3858 if (mark_in_progress()) {
3859 3859 concurrent_mark()->update_g1_committed();
3860 3860 }
3861 3861
3862 3862 #ifdef TRACESPINNING
3863 3863 ParallelTaskTerminator::print_termination_counts();
3864 3864 #endif
3865 3865
3866 3866 gc_epilogue(false);
3867 3867 }
3868 3868
3869 3869 if (ExitAfterGCNum > 0 && total_collections() == ExitAfterGCNum) {
3870 3870 gclog_or_tty->print_cr("Stopping after GC #%d", ExitAfterGCNum);
3871 3871 print_tracing_info();
3872 3872 vm_exit(-1);
3873 3873 }
3874 3874 }
3875 3875
3876 3876 _hrs.verify_optional();
3877 3877 verify_region_sets_optional();
3878 3878
3879 3879 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
3880 3880 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
3881 3881
3882 3882 if (PrintHeapAtGC) {
3883 3883 Universe::print_heap_after_gc();
3884 3884 }
3885 3885 g1mm()->update_sizes();
3886 3886
3887 3887 if (G1SummarizeRSetStats &&
3888 3888 (G1SummarizeRSetStatsPeriod > 0) &&
3889 3889 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
3890 3890 g1_rem_set()->print_summary_info();
3891 3891 }
3892 3892
3893 3893 return true;
3894 3894 }
3895 3895
3896 3896 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
3897 3897 {
3898 3898 size_t gclab_word_size;
3899 3899 switch (purpose) {
3900 3900 case GCAllocForSurvived:
3901 3901 gclab_word_size = YoungPLABSize;
3902 3902 break;
3903 3903 case GCAllocForTenured:
3904 3904 gclab_word_size = OldPLABSize;
3905 3905 break;
3906 3906 default:
3907 3907 assert(false, "unknown GCAllocPurpose");
3908 3908 gclab_word_size = OldPLABSize;
3909 3909 break;
3910 3910 }
3911 3911 return gclab_word_size;
3912 3912 }
3913 3913
3914 3914 void G1CollectedHeap::init_mutator_alloc_region() {
3915 3915 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
3916 3916 _mutator_alloc_region.init();
3917 3917 }
3918 3918
3919 3919 void G1CollectedHeap::release_mutator_alloc_region() {
3920 3920 _mutator_alloc_region.release();
3921 3921 assert(_mutator_alloc_region.get() == NULL, "post-condition");
3922 3922 }
3923 3923
3924 3924 void G1CollectedHeap::init_gc_alloc_regions() {
3925 3925 assert_at_safepoint(true /* should_be_vm_thread */);
3926 3926
3927 3927 _survivor_gc_alloc_region.init();
3928 3928 _old_gc_alloc_region.init();
3929 3929 HeapRegion* retained_region = _retained_old_gc_alloc_region;
3930 3930 _retained_old_gc_alloc_region = NULL;
3931 3931
3932 3932 // We will discard the current GC alloc region if:
3933 3933 // a) it's in the collection set (it can happen!),
3934 3934 // b) it's already full (no point in using it),
3935 3935 // c) it's empty (this means that it was emptied during
3936 3936 // a cleanup and it should be on the free list now), or
3937 3937 // d) it's humongous (this means that it was emptied
3938 3938 // during a cleanup and was added to the free list, but
3939 3939 // has been subseqently used to allocate a humongous
3940 3940 // object that may be less than the region size).
3941 3941 if (retained_region != NULL &&
3942 3942 !retained_region->in_collection_set() &&
3943 3943 !(retained_region->top() == retained_region->end()) &&
3944 3944 !retained_region->is_empty() &&
3945 3945 !retained_region->isHumongous()) {
3946 3946 retained_region->set_saved_mark();
3947 3947 // The retained region was added to the old region set when it was
3948 3948 // retired. We have to remove it now, since we don't allow regions
3949 3949 // we allocate to in the region sets. We'll re-add it later, when
3950 3950 // it's retired again.
3951 3951 _old_set.remove(retained_region);
3952 3952 _old_gc_alloc_region.set(retained_region);
3953 3953 _hr_printer.reuse(retained_region);
3954 3954 }
3955 3955 }
3956 3956
3957 3957 void G1CollectedHeap::release_gc_alloc_regions() {
3958 3958 _survivor_gc_alloc_region.release();
3959 3959 // If we have an old GC alloc region to release, we'll save it in
3960 3960 // _retained_old_gc_alloc_region. If we don't
3961 3961 // _retained_old_gc_alloc_region will become NULL. This is what we
3962 3962 // want either way so no reason to check explicitly for either
3963 3963 // condition.
3964 3964 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
3965 3965 }
3966 3966
3967 3967 void G1CollectedHeap::abandon_gc_alloc_regions() {
3968 3968 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
3969 3969 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
3970 3970 _retained_old_gc_alloc_region = NULL;
3971 3971 }
3972 3972
3973 3973 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
3974 3974 _drain_in_progress = false;
3975 3975 set_evac_failure_closure(cl);
3976 3976 _evac_failure_scan_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
3977 3977 }
3978 3978
3979 3979 void G1CollectedHeap::finalize_for_evac_failure() {
3980 3980 assert(_evac_failure_scan_stack != NULL &&
3981 3981 _evac_failure_scan_stack->length() == 0,
3982 3982 "Postcondition");
3983 3983 assert(!_drain_in_progress, "Postcondition");
3984 3984 delete _evac_failure_scan_stack;
3985 3985 _evac_failure_scan_stack = NULL;
3986 3986 }
3987 3987
3988 3988 class UpdateRSetDeferred : public OopsInHeapRegionClosure {
3989 3989 private:
3990 3990 G1CollectedHeap* _g1;
3991 3991 DirtyCardQueue *_dcq;
3992 3992 CardTableModRefBS* _ct_bs;
3993 3993
3994 3994 public:
3995 3995 UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
3996 3996 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {}
3997 3997
3998 3998 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
3999 3999 virtual void do_oop( oop* p) { do_oop_work(p); }
4000 4000 template <class T> void do_oop_work(T* p) {
4001 4001 assert(_from->is_in_reserved(p), "paranoia");
4002 4002 if (!_from->is_in_reserved(oopDesc::load_decode_heap_oop(p)) &&
4003 4003 !_from->is_survivor()) {
4004 4004 size_t card_index = _ct_bs->index_for(p);
4005 4005 if (_ct_bs->mark_card_deferred(card_index)) {
4006 4006 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
4007 4007 }
4008 4008 }
4009 4009 }
4010 4010 };
4011 4011
4012 4012 class RemoveSelfPointerClosure: public ObjectClosure {
4013 4013 private:
4014 4014 G1CollectedHeap* _g1;
4015 4015 ConcurrentMark* _cm;
4016 4016 HeapRegion* _hr;
4017 4017 size_t _prev_marked_bytes;
4018 4018 size_t _next_marked_bytes;
4019 4019 OopsInHeapRegionClosure *_cl;
4020 4020 public:
4021 4021 RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr,
4022 4022 OopsInHeapRegionClosure* cl) :
4023 4023 _g1(g1), _hr(hr), _cm(_g1->concurrent_mark()), _prev_marked_bytes(0),
4024 4024 _next_marked_bytes(0), _cl(cl) {}
4025 4025
4026 4026 size_t prev_marked_bytes() { return _prev_marked_bytes; }
4027 4027 size_t next_marked_bytes() { return _next_marked_bytes; }
4028 4028
4029 4029 // <original comment>
4030 4030 // The original idea here was to coalesce evacuated and dead objects.
4031 4031 // However that caused complications with the block offset table (BOT).
4032 4032 // In particular if there were two TLABs, one of them partially refined.
4033 4033 // |----- TLAB_1--------|----TLAB_2-~~~(partially refined part)~~~|
4034 4034 // The BOT entries of the unrefined part of TLAB_2 point to the start
4035 4035 // of TLAB_2. If the last object of the TLAB_1 and the first object
4036 4036 // of TLAB_2 are coalesced, then the cards of the unrefined part
4037 4037 // would point into middle of the filler object.
4038 4038 // The current approach is to not coalesce and leave the BOT contents intact.
4039 4039 // </original comment>
4040 4040 //
4041 4041 // We now reset the BOT when we start the object iteration over the
4042 4042 // region and refine its entries for every object we come across. So
4043 4043 // the above comment is not really relevant and we should be able
4044 4044 // to coalesce dead objects if we want to.
4045 4045 void do_object(oop obj) {
4046 4046 HeapWord* obj_addr = (HeapWord*) obj;
↓ open down ↓ |
4046 lines elided |
↑ open up ↑ |
4047 4047 assert(_hr->is_in(obj_addr), "sanity");
4048 4048 size_t obj_size = obj->size();
4049 4049 _hr->update_bot_for_object(obj_addr, obj_size);
4050 4050 if (obj->is_forwarded() && obj->forwardee() == obj) {
4051 4051 // The object failed to move.
4052 4052 assert(!_g1->is_obj_dead(obj), "We should not be preserving dead objs.");
4053 4053 _cm->markPrev(obj);
4054 4054 assert(_cm->isPrevMarked(obj), "Should be marked!");
4055 4055 _prev_marked_bytes += (obj_size * HeapWordSize);
4056 4056 if (_g1->mark_in_progress() && !_g1->is_obj_ill(obj)) {
4057 - _cm->markAndGrayObjectIfNecessary(obj);
4057 + _cm->markAndGrayObjectIfNecessary(obj, 0 /* worker_i */);
4058 4058 }
4059 4059 obj->set_mark(markOopDesc::prototype());
4060 4060 // While we were processing RSet buffers during the
4061 4061 // collection, we actually didn't scan any cards on the
4062 4062 // collection set, since we didn't want to update remebered
4063 4063 // sets with entries that point into the collection set, given
4064 4064 // that live objects fromthe collection set are about to move
4065 4065 // and such entries will be stale very soon. This change also
4066 4066 // dealt with a reliability issue which involved scanning a
4067 4067 // card in the collection set and coming across an array that
4068 4068 // was being chunked and looking malformed. The problem is
4069 4069 // that, if evacuation fails, we might have remembered set
4070 4070 // entries missing given that we skipped cards on the
4071 4071 // collection set. So, we'll recreate such entries now.
4072 4072 obj->oop_iterate(_cl);
4073 4073 assert(_cm->isPrevMarked(obj), "Should be marked!");
4074 4074 } else {
4075 4075 // The object has been either evacuated or is dead. Fill it with a
4076 4076 // dummy object.
4077 4077 MemRegion mr((HeapWord*)obj, obj_size);
4078 4078 CollectedHeap::fill_with_object(mr);
4079 4079 _cm->clearRangeBothMaps(mr);
4080 4080 }
4081 4081 }
4082 4082 };
4083 4083
4084 4084 void G1CollectedHeap::remove_self_forwarding_pointers() {
4085 4085 UpdateRSetImmediate immediate_update(_g1h->g1_rem_set());
4086 4086 DirtyCardQueue dcq(&_g1h->dirty_card_queue_set());
4087 4087 UpdateRSetDeferred deferred_update(_g1h, &dcq);
4088 4088 OopsInHeapRegionClosure *cl;
4089 4089 if (G1DeferredRSUpdate) {
4090 4090 cl = &deferred_update;
4091 4091 } else {
4092 4092 cl = &immediate_update;
4093 4093 }
4094 4094 HeapRegion* cur = g1_policy()->collection_set();
4095 4095 while (cur != NULL) {
4096 4096 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
4097 4097 assert(!cur->isHumongous(), "sanity");
4098 4098
4099 4099 if (cur->evacuation_failed()) {
4100 4100 assert(cur->in_collection_set(), "bad CS");
4101 4101 RemoveSelfPointerClosure rspc(_g1h, cur, cl);
4102 4102
4103 4103 // In the common case we make sure that this is done when the
4104 4104 // region is freed so that it is "ready-to-go" when it's
4105 4105 // re-allocated. However, when evacuation failure happens, a
4106 4106 // region will remain in the heap and might ultimately be added
4107 4107 // to a CSet in the future. So we have to be careful here and
4108 4108 // make sure the region's RSet is ready for parallel iteration
4109 4109 // whenever this might be required in the future.
4110 4110 cur->rem_set()->reset_for_par_iteration();
4111 4111 cur->reset_bot();
4112 4112 cl->set_region(cur);
4113 4113 cur->object_iterate(&rspc);
4114 4114
4115 4115 // A number of manipulations to make the TAMS be the current top,
4116 4116 // and the marked bytes be the ones observed in the iteration.
4117 4117 if (_g1h->concurrent_mark()->at_least_one_mark_complete()) {
4118 4118 // The comments below are the postconditions achieved by the
4119 4119 // calls. Note especially the last such condition, which says that
4120 4120 // the count of marked bytes has been properly restored.
4121 4121 cur->note_start_of_marking(false);
4122 4122 // _next_top_at_mark_start == top, _next_marked_bytes == 0
4123 4123 cur->add_to_marked_bytes(rspc.prev_marked_bytes());
4124 4124 // _next_marked_bytes == prev_marked_bytes.
4125 4125 cur->note_end_of_marking();
4126 4126 // _prev_top_at_mark_start == top(),
4127 4127 // _prev_marked_bytes == prev_marked_bytes
4128 4128 }
4129 4129 // If there is no mark in progress, we modified the _next variables
4130 4130 // above needlessly, but harmlessly.
4131 4131 if (_g1h->mark_in_progress()) {
4132 4132 cur->note_start_of_marking(false);
4133 4133 // _next_top_at_mark_start == top, _next_marked_bytes == 0
4134 4134 // _next_marked_bytes == next_marked_bytes.
4135 4135 }
4136 4136 }
4137 4137 cur = cur->next_in_collection_set();
4138 4138 }
4139 4139 assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
4140 4140
4141 4141 // Now restore saved marks, if any.
4142 4142 if (_objs_with_preserved_marks != NULL) {
4143 4143 assert(_preserved_marks_of_objs != NULL, "Both or none.");
4144 4144 guarantee(_objs_with_preserved_marks->length() ==
4145 4145 _preserved_marks_of_objs->length(), "Both or none.");
4146 4146 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
4147 4147 oop obj = _objs_with_preserved_marks->at(i);
4148 4148 markOop m = _preserved_marks_of_objs->at(i);
4149 4149 obj->set_mark(m);
4150 4150 }
4151 4151 // Delete the preserved marks growable arrays (allocated on the C heap).
4152 4152 delete _objs_with_preserved_marks;
4153 4153 delete _preserved_marks_of_objs;
4154 4154 _objs_with_preserved_marks = NULL;
4155 4155 _preserved_marks_of_objs = NULL;
4156 4156 }
4157 4157 }
4158 4158
4159 4159 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4160 4160 _evac_failure_scan_stack->push(obj);
4161 4161 }
4162 4162
4163 4163 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4164 4164 assert(_evac_failure_scan_stack != NULL, "precondition");
4165 4165
↓ open down ↓ |
98 lines elided |
↑ open up ↑ |
4166 4166 while (_evac_failure_scan_stack->length() > 0) {
4167 4167 oop obj = _evac_failure_scan_stack->pop();
4168 4168 _evac_failure_closure->set_region(heap_region_containing(obj));
4169 4169 obj->oop_iterate_backwards(_evac_failure_closure);
4170 4170 }
4171 4171 }
4172 4172
4173 4173 oop
4174 4174 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4175 4175 oop old,
4176 - bool should_mark_root) {
4176 + bool should_mark_root,
4177 + int worker_i) {
4177 4178 assert(obj_in_cs(old),
4178 4179 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4179 4180 (HeapWord*) old));
4180 4181 markOop m = old->mark();
4181 4182 oop forward_ptr = old->forward_to_atomic(old);
4182 4183 if (forward_ptr == NULL) {
4183 4184 // Forward-to-self succeeded.
4184 4185
4185 4186 // should_mark_root will be true when this routine is called
4186 4187 // from a root scanning closure during an initial mark pause.
4187 4188 // In this case the thread that succeeds in self-forwarding the
4188 4189 // object is also responsible for marking the object.
4189 4190 if (should_mark_root) {
4190 4191 assert(!oopDesc::is_null(old), "shouldn't be");
4191 - _cm->grayRoot(old);
4192 + _cm->grayRoot(old, worker_i);
4192 4193 }
4193 4194
4194 4195 if (_evac_failure_closure != cl) {
4195 4196 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4196 4197 assert(!_drain_in_progress,
4197 4198 "Should only be true while someone holds the lock.");
4198 4199 // Set the global evac-failure closure to the current thread's.
4199 4200 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4200 4201 set_evac_failure_closure(cl);
4201 4202 // Now do the common part.
4202 4203 handle_evacuation_failure_common(old, m);
4203 4204 // Reset to NULL.
4204 4205 set_evac_failure_closure(NULL);
4205 4206 } else {
4206 4207 // The lock is already held, and this is recursive.
4207 4208 assert(_drain_in_progress, "This should only be the recursive case.");
4208 4209 handle_evacuation_failure_common(old, m);
4209 4210 }
4210 4211 return old;
4211 4212 } else {
4212 4213 // Forward-to-self failed. Either someone else managed to allocate
4213 4214 // space for this object (old != forward_ptr) or they beat us in
4214 4215 // self-forwarding it (old == forward_ptr).
4215 4216 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4216 4217 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4217 4218 "should not be in the CSet",
4218 4219 (HeapWord*) old, (HeapWord*) forward_ptr));
4219 4220 return forward_ptr;
4220 4221 }
4221 4222 }
4222 4223
4223 4224 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4224 4225 set_evacuation_failed(true);
4225 4226
4226 4227 preserve_mark_if_necessary(old, m);
4227 4228
4228 4229 HeapRegion* r = heap_region_containing(old);
4229 4230 if (!r->evacuation_failed()) {
4230 4231 r->set_evacuation_failed(true);
4231 4232 _hr_printer.evac_failure(r);
4232 4233 }
4233 4234
4234 4235 push_on_evac_failure_scan_stack(old);
4235 4236
4236 4237 if (!_drain_in_progress) {
4237 4238 // prevent recursion in copy_to_survivor_space()
4238 4239 _drain_in_progress = true;
4239 4240 drain_evac_failure_scan_stack();
4240 4241 _drain_in_progress = false;
4241 4242 }
4242 4243 }
4243 4244
4244 4245 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4245 4246 assert(evacuation_failed(), "Oversaving!");
4246 4247 // We want to call the "for_promotion_failure" version only in the
4247 4248 // case of a promotion failure.
4248 4249 if (m->must_be_preserved_for_promotion_failure(obj)) {
4249 4250 if (_objs_with_preserved_marks == NULL) {
4250 4251 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4251 4252 _objs_with_preserved_marks =
4252 4253 new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
4253 4254 _preserved_marks_of_objs =
4254 4255 new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
4255 4256 }
4256 4257 _objs_with_preserved_marks->push(obj);
4257 4258 _preserved_marks_of_objs->push(m);
4258 4259 }
4259 4260 }
4260 4261
4261 4262 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4262 4263 size_t word_size) {
4263 4264 if (purpose == GCAllocForSurvived) {
4264 4265 HeapWord* result = survivor_attempt_allocation(word_size);
4265 4266 if (result != NULL) {
4266 4267 return result;
4267 4268 } else {
4268 4269 // Let's try to allocate in the old gen in case we can fit the
4269 4270 // object there.
4270 4271 return old_attempt_allocation(word_size);
4271 4272 }
4272 4273 } else {
4273 4274 assert(purpose == GCAllocForTenured, "sanity");
4274 4275 HeapWord* result = old_attempt_allocation(word_size);
4275 4276 if (result != NULL) {
4276 4277 return result;
4277 4278 } else {
4278 4279 // Let's try to allocate in the survivors in case we can fit the
4279 4280 // object there.
4280 4281 return survivor_attempt_allocation(word_size);
4281 4282 }
4282 4283 }
4283 4284
4284 4285 ShouldNotReachHere();
4285 4286 // Trying to keep some compilers happy.
4286 4287 return NULL;
↓ open down ↓ |
85 lines elided |
↑ open up ↑ |
4287 4288 }
4288 4289
4289 4290 #ifndef PRODUCT
4290 4291 bool GCLabBitMapClosure::do_bit(size_t offset) {
4291 4292 HeapWord* addr = _bitmap->offsetToHeapWord(offset);
4292 4293 guarantee(_cm->isMarked(oop(addr)), "it should be!");
4293 4294 return true;
4294 4295 }
4295 4296 #endif // PRODUCT
4296 4297
4297 -G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4298 +void GCLabBitMap::retire(int worker_i) {
4299 + guarantee(use_local_bitmaps, "invariant");
4300 + assert(fields_well_formed(), "invariant");
4301 +
4302 + if (_start_word != NULL) {
4303 + CMBitMap* mark_bitmap = _cm->nextMarkBitMap();
4304 +
4305 + // this means that the bitmap was set up for the GCLab
4306 + assert(_real_start_word != NULL && _real_end_word != NULL, "invariant");
4307 +
4308 + mark_bitmap->mostly_disjoint_range_union(this,
4309 + 0, // always start from the start of the bitmap
4310 + _start_word,
4311 + gclab_real_word_size());
4312 +
4313 + // Note: Even though that not all objects copied into the LAB will
4314 + // have their bit set in the LAB bitmap (the LAB bitmap is used to
4315 + // propagate marks), we can just add the entire lab and its bitmap
4316 + // to the count of the marked data. It's OK (but inaccurate) to
4317 + // count a dead object but we can't miss counting a live object.
4318 + MemRegion lab_region(_real_start_word, _real_end_word);
4319 + _cm->count_region(lab_region, worker_i);
4320 + _cm->grayRegionIfNecessary(lab_region);
4321 +
4322 +#ifndef PRODUCT
4323 + if (use_local_bitmaps && verify_local_bitmaps) {
4324 + verify();
4325 + }
4326 +#endif // PRODUCT
4327 + } else {
4328 + assert(_real_start_word == NULL && _real_end_word == NULL, "invariant");
4329 + }
4330 +}
4331 +
4332 +G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size, int worker_i) :
4298 4333 ParGCAllocBuffer(gclab_word_size),
4299 4334 _should_mark_objects(false),
4300 4335 _bitmap(G1CollectedHeap::heap()->reserved_region().start(), gclab_word_size),
4336 + _worker_i(worker_i),
4301 4337 _retired(false)
4302 4338 {
4303 4339 //_should_mark_objects is set to true when G1ParCopyHelper needs to
4304 4340 // mark the forwarded location of an evacuated object.
4305 4341 // We set _should_mark_objects to true if marking is active, i.e. when we
4306 4342 // need to propagate a mark, or during an initial mark pause, i.e. when we
4307 4343 // need to mark objects immediately reachable by the roots.
4308 4344 if (G1CollectedHeap::heap()->mark_in_progress() ||
4309 4345 G1CollectedHeap::heap()->g1_policy()->during_initial_mark_pause()) {
4310 4346 _should_mark_objects = true;
4311 4347 }
↓ open down ↓ |
1 lines elided |
↑ open up ↑ |
4312 4348 }
4313 4349
4314 4350 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num)
4315 4351 : _g1h(g1h),
4316 4352 _refs(g1h->task_queue(queue_num)),
4317 4353 _dcq(&g1h->dirty_card_queue_set()),
4318 4354 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4319 4355 _g1_rem(g1h->g1_rem_set()),
4320 4356 _hash_seed(17), _queue_num(queue_num),
4321 4357 _term_attempts(0),
4322 - _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4323 - _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4358 + _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived), queue_num),
4359 + _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured), queue_num),
4324 4360 _age_table(false),
4325 4361 _strong_roots_time(0), _term_time(0),
4326 4362 _alloc_buffer_waste(0), _undo_waste(0)
4327 4363 {
4328 4364 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4329 4365 // we "sacrifice" entry 0 to keep track of surviving bytes for
4330 4366 // non-young regions (where the age is -1)
4331 4367 // We also add a few elements at the beginning and at the end in
4332 4368 // an attempt to eliminate cache contention
4333 4369 size_t real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4334 4370 size_t array_length = PADDING_ELEM_NUM +
4335 4371 real_length +
4336 4372 PADDING_ELEM_NUM;
4337 4373 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length);
4338 4374 if (_surviving_young_words_base == NULL)
4339 4375 vm_exit_out_of_memory(array_length * sizeof(size_t),
4340 4376 "Not enough space for young surv histo.");
4341 4377 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4342 4378 memset(_surviving_young_words, 0, real_length * sizeof(size_t));
4343 4379
4344 4380 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4345 4381 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4346 4382
4347 4383 _start = os::elapsedTime();
4348 4384 }
4349 4385
4350 4386 void
4351 4387 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4352 4388 {
4353 4389 st->print_raw_cr("GC Termination Stats");
4354 4390 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4355 4391 " ------waste (KiB)------");
4356 4392 st->print_raw_cr("thr ms ms % ms % attempts"
4357 4393 " total alloc undo");
4358 4394 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4359 4395 " ------- ------- -------");
4360 4396 }
4361 4397
4362 4398 void
4363 4399 G1ParScanThreadState::print_termination_stats(int i,
4364 4400 outputStream* const st) const
4365 4401 {
4366 4402 const double elapsed_ms = elapsed_time() * 1000.0;
4367 4403 const double s_roots_ms = strong_roots_time() * 1000.0;
4368 4404 const double term_ms = term_time() * 1000.0;
4369 4405 st->print_cr("%3d %9.2f %9.2f %6.2f "
4370 4406 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4371 4407 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4372 4408 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4373 4409 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4374 4410 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4375 4411 alloc_buffer_waste() * HeapWordSize / K,
4376 4412 undo_waste() * HeapWordSize / K);
4377 4413 }
4378 4414
4379 4415 #ifdef ASSERT
4380 4416 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4381 4417 assert(ref != NULL, "invariant");
4382 4418 assert(UseCompressedOops, "sanity");
4383 4419 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4384 4420 oop p = oopDesc::load_decode_heap_oop(ref);
4385 4421 assert(_g1h->is_in_g1_reserved(p),
4386 4422 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4387 4423 return true;
4388 4424 }
4389 4425
4390 4426 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4391 4427 assert(ref != NULL, "invariant");
4392 4428 if (has_partial_array_mask(ref)) {
4393 4429 // Must be in the collection set--it's already been copied.
4394 4430 oop p = clear_partial_array_mask(ref);
4395 4431 assert(_g1h->obj_in_cs(p),
4396 4432 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4397 4433 } else {
4398 4434 oop p = oopDesc::load_decode_heap_oop(ref);
4399 4435 assert(_g1h->is_in_g1_reserved(p),
4400 4436 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4401 4437 }
4402 4438 return true;
4403 4439 }
4404 4440
4405 4441 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4406 4442 if (ref.is_narrow()) {
4407 4443 return verify_ref((narrowOop*) ref);
4408 4444 } else {
4409 4445 return verify_ref((oop*) ref);
4410 4446 }
4411 4447 }
4412 4448 #endif // ASSERT
4413 4449
4414 4450 void G1ParScanThreadState::trim_queue() {
4415 4451 assert(_evac_cl != NULL, "not set");
4416 4452 assert(_evac_failure_cl != NULL, "not set");
4417 4453 assert(_partial_scan_cl != NULL, "not set");
4418 4454
4419 4455 StarTask ref;
4420 4456 do {
4421 4457 // Drain the overflow stack first, so other threads can steal.
4422 4458 while (refs()->pop_overflow(ref)) {
4423 4459 deal_with_reference(ref);
4424 4460 }
↓ open down ↓ |
91 lines elided |
↑ open up ↑ |
4425 4461
4426 4462 while (refs()->pop_local(ref)) {
4427 4463 deal_with_reference(ref);
4428 4464 }
4429 4465 } while (!refs()->is_empty());
4430 4466 }
4431 4467
4432 4468 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) :
4433 4469 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4434 4470 _par_scan_state(par_scan_state),
4471 + _worker_i(par_scan_state->queue_num()),
4435 4472 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4436 4473 _mark_in_progress(_g1->mark_in_progress()) { }
4437 4474
4438 4475 template <class T> void G1ParCopyHelper::mark_object(T* p) {
4439 4476 // This is called from do_oop_work for objects that are not
4440 4477 // in the collection set. Objects in the collection set
4441 4478 // are marked after they have been evacuated.
4442 4479
4443 4480 T heap_oop = oopDesc::load_heap_oop(p);
4444 4481 if (!oopDesc::is_null(heap_oop)) {
4445 4482 oop obj = oopDesc::decode_heap_oop(heap_oop);
4446 4483 HeapWord* addr = (HeapWord*)obj;
4447 4484 if (_g1->is_in_g1_reserved(addr)) {
4448 - _cm->grayRoot(oop(addr));
4485 + _cm->grayRoot(oop(addr), _worker_i);
4449 4486 }
4450 4487 }
4451 4488 }
4452 4489
4453 4490 oop G1ParCopyHelper::copy_to_survivor_space(oop old, bool should_mark_root,
4454 4491 bool should_mark_copy) {
4455 4492 size_t word_sz = old->size();
4456 4493 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4457 4494 // +1 to make the -1 indexes valid...
4458 4495 int young_index = from_region->young_index_in_cset()+1;
4459 4496 assert( (from_region->is_young() && young_index > 0) ||
4460 4497 (!from_region->is_young() && young_index == 0), "invariant" );
4461 4498 G1CollectorPolicy* g1p = _g1->g1_policy();
4462 4499 markOop m = old->mark();
4463 4500 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
↓ open down ↓ |
5 lines elided |
↑ open up ↑ |
4464 4501 : m->age();
4465 4502 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4466 4503 word_sz);
4467 4504 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4468 4505 oop obj = oop(obj_ptr);
4469 4506
4470 4507 if (obj_ptr == NULL) {
4471 4508 // This will either forward-to-self, or detect that someone else has
4472 4509 // installed a forwarding pointer.
4473 4510 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4474 - return _g1->handle_evacuation_failure_par(cl, old, should_mark_root);
4511 + return _g1->handle_evacuation_failure_par(cl, old, should_mark_root, _worker_i);
4475 4512 }
4476 4513
4477 4514 // We're going to allocate linearly, so might as well prefetch ahead.
4478 4515 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4479 4516
4480 4517 oop forward_ptr = old->forward_to_atomic(obj);
4481 4518 if (forward_ptr == NULL) {
4482 4519 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4483 4520 if (g1p->track_object_age(alloc_purpose)) {
4484 4521 // We could simply do obj->incr_age(). However, this causes a
4485 4522 // performance issue. obj->incr_age() will first check whether
4486 4523 // the object has a displaced mark by checking its mark word;
4487 4524 // getting the mark word from the new location of the object
4488 4525 // stalls. So, given that we already have the mark word and we
4489 4526 // are about to install it anyway, it's better to increase the
4490 4527 // age on the mark word, when the object does not have a
4491 4528 // displaced mark word. We're not expecting many objects to have
4492 4529 // a displaced marked word, so that case is not optimized
4493 4530 // further (it could be...) and we simply call obj->incr_age().
4494 4531
4495 4532 if (m->has_displaced_mark_helper()) {
4496 4533 // in this case, we have to install the mark word first,
4497 4534 // otherwise obj looks to be forwarded (the old mark word,
4498 4535 // which contains the forward pointer, was copied)
4499 4536 obj->set_mark(m);
4500 4537 obj->incr_age();
4501 4538 } else {
4502 4539 m = m->incr_age();
4503 4540 obj->set_mark(m);
4504 4541 }
4505 4542 _par_scan_state->age_table()->add(obj, word_sz);
4506 4543 } else {
↓ open down ↓ |
22 lines elided |
↑ open up ↑ |
4507 4544 obj->set_mark(m);
4508 4545 }
4509 4546
4510 4547 // Mark the evacuated object or propagate "next" mark bit
4511 4548 if (should_mark_copy) {
4512 4549 if (!use_local_bitmaps ||
4513 4550 !_par_scan_state->alloc_buffer(alloc_purpose)->mark(obj_ptr)) {
4514 4551 // if we couldn't mark it on the local bitmap (this happens when
4515 4552 // the object was not allocated in the GCLab), we have to bite
4516 4553 // the bullet and do the standard parallel mark
4517 - _cm->markAndGrayObjectIfNecessary(obj);
4554 + _cm->markAndGrayObjectIfNecessary(obj, _worker_i);
4518 4555 }
4519 4556
4520 4557 if (_g1->isMarkedNext(old)) {
4521 4558 // Unmark the object's old location so that marking
4522 4559 // doesn't think the old object is alive.
4523 4560 _cm->nextMarkBitMap()->parClear((HeapWord*)old);
4561 +
4562 + // We could clear the count data for the old object here but
4563 + // currently we do not. Why don't we do this? The thread/task
4564 + // that marks a newly copied object is likely _not_ the thread/task
4565 + // that originally marked the old object. So, to clear the count
4566 + // data for the old object, we would have to scan the count
4567 + // data for all of the tasks (and clear the data for the old object
4568 + // in parallel with other threads adding to the count data). Even
4569 + // then we could clear a bit incorrectly (e.g. if the old object
4570 + // does not start or end on a card boundary). It's more important
4571 + // that we don't have missed bits that should've been set than
4572 + // having extra bits set.
4573 + //
4574 + // As a result the accumulated count data could be a superset
4575 + // of the data that is/would have been calculated by walking
4576 + // the marking bitmap.
4524 4577 }
4525 4578 }
4526 4579
4527 4580 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4528 4581 surv_young_words[young_index] += word_sz;
4529 4582
4530 4583 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4531 4584 arrayOop(old)->set_length(0);
4532 4585 oop* old_p = set_partial_array_mask(old);
4533 4586 _par_scan_state->push_on_queue(old_p);
4534 4587 } else {
4535 4588 // No point in using the slower heap_region_containing() method,
4536 4589 // given that we know obj is in the heap.
4537 4590 _scanner->set_region(_g1->heap_region_containing_raw(obj));
4538 4591 obj->oop_iterate_backwards(_scanner);
4539 4592 }
4540 4593 } else {
4541 4594 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4542 4595 obj = forward_ptr;
4543 4596 }
4544 4597 return obj;
4545 4598 }
4546 4599
4547 4600 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4548 4601 template <class T>
4549 4602 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4550 4603 ::do_oop_work(T* p) {
4551 4604 oop obj = oopDesc::load_decode_heap_oop(p);
4552 4605 assert(barrier != G1BarrierRS || obj != NULL,
4553 4606 "Precondition: G1BarrierRS implies obj is nonNull");
4554 4607
4555 4608 // Marking:
4556 4609 // If the object is in the collection set, then the thread
4557 4610 // that copies the object should mark, or propagate the
4558 4611 // mark to, the evacuated object.
4559 4612 // If the object is not in the collection set then we
4560 4613 // should call the mark_object() method depending on the
4561 4614 // value of the template parameter do_mark_object (which will
4562 4615 // be true for root scanning closures during an initial mark
4563 4616 // pause).
4564 4617 // The mark_object() method first checks whether the object
4565 4618 // is marked and, if not, attempts to mark the object.
4566 4619
4567 4620 // here the null check is implicit in the cset_fast_test() test
4568 4621 if (_g1->in_cset_fast_test(obj)) {
4569 4622 if (obj->is_forwarded()) {
4570 4623 oopDesc::encode_store_heap_oop(p, obj->forwardee());
4571 4624 // If we are a root scanning closure during an initial
4572 4625 // mark pause (i.e. do_mark_object will be true) then
4573 4626 // we also need to handle marking of roots in the
4574 4627 // event of an evacuation failure. In the event of an
4575 4628 // evacuation failure, the object is forwarded to itself
4576 4629 // and not copied. For root-scanning closures, the
4577 4630 // object would be marked after a successful self-forward
4578 4631 // but an object could be pointed to by both a root and non
4579 4632 // root location and be self-forwarded by a non-root-scanning
4580 4633 // closure. Therefore we also have to attempt to mark the
4581 4634 // self-forwarded root object here.
4582 4635 if (do_mark_object && obj->forwardee() == obj) {
4583 4636 mark_object(p);
4584 4637 }
4585 4638 } else {
4586 4639 // During an initial mark pause, objects that are pointed to
4587 4640 // by the roots need to be marked - even in the event of an
4588 4641 // evacuation failure. We pass the template parameter
4589 4642 // do_mark_object (which is true for root scanning closures
4590 4643 // during an initial mark pause) to copy_to_survivor_space
4591 4644 // which will pass it on to the evacuation failure handling
4592 4645 // code. The thread that successfully self-forwards a root
4593 4646 // object to itself is responsible for marking the object.
4594 4647 bool should_mark_root = do_mark_object;
4595 4648
4596 4649 // We need to mark the copied object if we're a root scanning
4597 4650 // closure during an initial mark pause (i.e. do_mark_object
4598 4651 // will be true), or the object is already marked and we need
4599 4652 // to propagate the mark to the evacuated copy.
↓ open down ↓ |
66 lines elided |
↑ open up ↑ |
4600 4653 bool should_mark_copy = do_mark_object ||
4601 4654 _during_initial_mark ||
4602 4655 (_mark_in_progress && !_g1->is_obj_ill(obj));
4603 4656
4604 4657 oop copy_oop = copy_to_survivor_space(obj, should_mark_root,
4605 4658 should_mark_copy);
4606 4659 oopDesc::encode_store_heap_oop(p, copy_oop);
4607 4660 }
4608 4661 // When scanning the RS, we only care about objs in CS.
4609 4662 if (barrier == G1BarrierRS) {
4610 - _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4663 + assert(_worker_i == _par_scan_state->queue_num(), "sanity");
4664 + _par_scan_state->update_rs(_from, p, _worker_i);
4611 4665 }
4612 4666 } else {
4613 4667 // The object is not in collection set. If we're a root scanning
4614 4668 // closure during an initial mark pause (i.e. do_mark_object will
4615 4669 // be true) then attempt to mark the object.
4616 4670 if (do_mark_object) {
4617 4671 mark_object(p);
4618 4672 }
4619 4673 }
4620 4674
4621 4675 if (barrier == G1BarrierEvac && obj != NULL) {
4622 - _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
4676 + assert(_worker_i == _par_scan_state->queue_num(), "sanity");
4677 + _par_scan_state->update_rs(_from, p, _worker_i);
4623 4678 }
4624 4679
4625 4680 if (do_gen_barrier && obj != NULL) {
4626 4681 par_do_barrier(p);
4627 4682 }
4628 4683 }
4629 4684
4630 4685 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4631 4686 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4632 4687
4633 4688 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4634 4689 assert(has_partial_array_mask(p), "invariant");
4635 4690 oop old = clear_partial_array_mask(p);
4636 4691 assert(old->is_objArray(), "must be obj array");
4637 4692 assert(old->is_forwarded(), "must be forwarded");
4638 4693 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
4639 4694
4640 4695 objArrayOop obj = objArrayOop(old->forwardee());
4641 4696 assert((void*)old != (void*)old->forwardee(), "self forwarding here?");
4642 4697 // Process ParGCArrayScanChunk elements now
4643 4698 // and push the remainder back onto queue
4644 4699 int start = arrayOop(old)->length();
4645 4700 int end = obj->length();
4646 4701 int remainder = end - start;
4647 4702 assert(start <= end, "just checking");
4648 4703 if (remainder > 2 * ParGCArrayScanChunk) {
4649 4704 // Test above combines last partial chunk with a full chunk
4650 4705 end = start + ParGCArrayScanChunk;
4651 4706 arrayOop(old)->set_length(end);
4652 4707 // Push remainder.
4653 4708 oop* old_p = set_partial_array_mask(old);
4654 4709 assert(arrayOop(old)->length() < obj->length(), "Empty push?");
4655 4710 _par_scan_state->push_on_queue(old_p);
4656 4711 } else {
4657 4712 // Restore length so that the heap remains parsable in
4658 4713 // case of evacuation failure.
4659 4714 arrayOop(old)->set_length(end);
4660 4715 }
4661 4716 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4662 4717 // process our set of indices (include header in first chunk)
4663 4718 obj->oop_iterate_range(&_scanner, start, end);
4664 4719 }
4665 4720
4666 4721 class G1ParEvacuateFollowersClosure : public VoidClosure {
4667 4722 protected:
4668 4723 G1CollectedHeap* _g1h;
4669 4724 G1ParScanThreadState* _par_scan_state;
4670 4725 RefToScanQueueSet* _queues;
4671 4726 ParallelTaskTerminator* _terminator;
4672 4727
4673 4728 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4674 4729 RefToScanQueueSet* queues() { return _queues; }
4675 4730 ParallelTaskTerminator* terminator() { return _terminator; }
4676 4731
4677 4732 public:
4678 4733 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4679 4734 G1ParScanThreadState* par_scan_state,
4680 4735 RefToScanQueueSet* queues,
4681 4736 ParallelTaskTerminator* terminator)
4682 4737 : _g1h(g1h), _par_scan_state(par_scan_state),
4683 4738 _queues(queues), _terminator(terminator) {}
4684 4739
4685 4740 void do_void();
4686 4741
4687 4742 private:
4688 4743 inline bool offer_termination();
4689 4744 };
4690 4745
4691 4746 bool G1ParEvacuateFollowersClosure::offer_termination() {
4692 4747 G1ParScanThreadState* const pss = par_scan_state();
4693 4748 pss->start_term_time();
4694 4749 const bool res = terminator()->offer_termination();
4695 4750 pss->end_term_time();
4696 4751 return res;
4697 4752 }
4698 4753
4699 4754 void G1ParEvacuateFollowersClosure::do_void() {
4700 4755 StarTask stolen_task;
4701 4756 G1ParScanThreadState* const pss = par_scan_state();
4702 4757 pss->trim_queue();
4703 4758
4704 4759 do {
4705 4760 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4706 4761 assert(pss->verify_task(stolen_task), "sanity");
4707 4762 if (stolen_task.is_narrow()) {
4708 4763 pss->deal_with_reference((narrowOop*) stolen_task);
4709 4764 } else {
4710 4765 pss->deal_with_reference((oop*) stolen_task);
4711 4766 }
4712 4767
4713 4768 // We've just processed a reference and we might have made
4714 4769 // available new entries on the queues. So we have to make sure
4715 4770 // we drain the queues as necessary.
4716 4771 pss->trim_queue();
4717 4772 }
4718 4773 } while (!offer_termination());
4719 4774
4720 4775 pss->retire_alloc_buffers();
4721 4776 }
4722 4777
4723 4778 class G1ParTask : public AbstractGangTask {
4724 4779 protected:
4725 4780 G1CollectedHeap* _g1h;
4726 4781 RefToScanQueueSet *_queues;
4727 4782 ParallelTaskTerminator _terminator;
4728 4783 int _n_workers;
4729 4784
4730 4785 Mutex _stats_lock;
4731 4786 Mutex* stats_lock() { return &_stats_lock; }
4732 4787
4733 4788 size_t getNCards() {
4734 4789 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4735 4790 / G1BlockOffsetSharedArray::N_bytes;
4736 4791 }
4737 4792
4738 4793 public:
4739 4794 G1ParTask(G1CollectedHeap* g1h,
4740 4795 RefToScanQueueSet *task_queues)
4741 4796 : AbstractGangTask("G1 collection"),
4742 4797 _g1h(g1h),
4743 4798 _queues(task_queues),
4744 4799 _terminator(0, _queues),
4745 4800 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4746 4801 {}
4747 4802
4748 4803 RefToScanQueueSet* queues() { return _queues; }
4749 4804
4750 4805 RefToScanQueue *work_queue(int i) {
4751 4806 return queues()->queue(i);
4752 4807 }
4753 4808
4754 4809 ParallelTaskTerminator* terminator() { return &_terminator; }
4755 4810
4756 4811 virtual void set_for_termination(int active_workers) {
4757 4812 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4758 4813 // in the young space (_par_seq_tasks) in the G1 heap
4759 4814 // for SequentialSubTasksDone.
4760 4815 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4761 4816 // both of which need setting by set_n_termination().
4762 4817 _g1h->SharedHeap::set_n_termination(active_workers);
4763 4818 _g1h->set_n_termination(active_workers);
4764 4819 terminator()->reset_for_reuse(active_workers);
4765 4820 _n_workers = active_workers;
4766 4821 }
4767 4822
4768 4823 void work(int i) {
4769 4824 if (i >= _n_workers) return; // no work needed this round
4770 4825
4771 4826 double start_time_ms = os::elapsedTime() * 1000.0;
4772 4827 _g1h->g1_policy()->record_gc_worker_start_time(i, start_time_ms);
4773 4828
4774 4829 ResourceMark rm;
4775 4830 HandleMark hm;
4776 4831
4777 4832 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4778 4833
4779 4834 G1ParScanThreadState pss(_g1h, i);
4780 4835 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4781 4836 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4782 4837 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4783 4838
4784 4839 pss.set_evac_closure(&scan_evac_cl);
4785 4840 pss.set_evac_failure_closure(&evac_failure_cl);
4786 4841 pss.set_partial_scan_closure(&partial_scan_cl);
4787 4842
4788 4843 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4789 4844 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp);
4790 4845
4791 4846 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4792 4847 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp);
4793 4848
4794 4849 OopClosure* scan_root_cl = &only_scan_root_cl;
4795 4850 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl;
4796 4851
4797 4852 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4798 4853 // We also need to mark copied objects.
4799 4854 scan_root_cl = &scan_mark_root_cl;
4800 4855 scan_perm_cl = &scan_mark_perm_cl;
4801 4856 }
4802 4857
4803 4858 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4804 4859
4805 4860 pss.start_strong_roots();
4806 4861 _g1h->g1_process_strong_roots(/* not collecting perm */ false,
4807 4862 SharedHeap::SO_AllClasses,
4808 4863 scan_root_cl,
4809 4864 &push_heap_rs_cl,
4810 4865 scan_perm_cl,
4811 4866 i);
4812 4867 pss.end_strong_roots();
4813 4868
4814 4869 {
4815 4870 double start = os::elapsedTime();
4816 4871 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4817 4872 evac.do_void();
4818 4873 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4819 4874 double term_ms = pss.term_time()*1000.0;
4820 4875 _g1h->g1_policy()->record_obj_copy_time(i, elapsed_ms-term_ms);
4821 4876 _g1h->g1_policy()->record_termination(i, term_ms, pss.term_attempts());
4822 4877 }
4823 4878 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4824 4879 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4825 4880
4826 4881 // Clean up any par-expanded rem sets.
4827 4882 HeapRegionRemSet::par_cleanup();
4828 4883
4829 4884 if (ParallelGCVerbose) {
4830 4885 MutexLocker x(stats_lock());
4831 4886 pss.print_termination_stats(i);
4832 4887 }
4833 4888
4834 4889 assert(pss.refs()->is_empty(), "should be empty");
4835 4890 double end_time_ms = os::elapsedTime() * 1000.0;
4836 4891 _g1h->g1_policy()->record_gc_worker_end_time(i, end_time_ms);
4837 4892 }
4838 4893 };
4839 4894
4840 4895 // *** Common G1 Evacuation Stuff
4841 4896
4842 4897 // This method is run in a GC worker.
4843 4898
4844 4899 void
4845 4900 G1CollectedHeap::
4846 4901 g1_process_strong_roots(bool collecting_perm_gen,
4847 4902 SharedHeap::ScanningOption so,
4848 4903 OopClosure* scan_non_heap_roots,
4849 4904 OopsInHeapRegionClosure* scan_rs,
4850 4905 OopsInGenClosure* scan_perm,
4851 4906 int worker_i) {
4852 4907
4853 4908 // First scan the strong roots, including the perm gen.
4854 4909 double ext_roots_start = os::elapsedTime();
4855 4910 double closure_app_time_sec = 0.0;
4856 4911
4857 4912 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4858 4913 BufferingOopsInGenClosure buf_scan_perm(scan_perm);
4859 4914 buf_scan_perm.set_generation(perm_gen());
4860 4915
4861 4916 // Walk the code cache w/o buffering, because StarTask cannot handle
4862 4917 // unaligned oop locations.
4863 4918 CodeBlobToOopClosure eager_scan_code_roots(scan_non_heap_roots, /*do_marking=*/ true);
4864 4919
4865 4920 process_strong_roots(false, // no scoping; this is parallel code
4866 4921 collecting_perm_gen, so,
4867 4922 &buf_scan_non_heap_roots,
4868 4923 &eager_scan_code_roots,
4869 4924 &buf_scan_perm);
4870 4925
4871 4926 // Now the CM ref_processor roots.
4872 4927 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
4873 4928 // We need to treat the discovered reference lists of the
4874 4929 // concurrent mark ref processor as roots and keep entries
4875 4930 // (which are added by the marking threads) on them live
4876 4931 // until they can be processed at the end of marking.
4877 4932 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
4878 4933 }
4879 4934
4880 4935 // Finish up any enqueued closure apps (attributed as object copy time).
4881 4936 buf_scan_non_heap_roots.done();
4882 4937 buf_scan_perm.done();
4883 4938
4884 4939 double ext_roots_end = os::elapsedTime();
4885 4940
4886 4941 g1_policy()->reset_obj_copy_time(worker_i);
4887 4942 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() +
4888 4943 buf_scan_non_heap_roots.closure_app_seconds();
4889 4944 g1_policy()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
4890 4945
4891 4946 double ext_root_time_ms =
4892 4947 ((ext_roots_end - ext_roots_start) - obj_copy_time_sec) * 1000.0;
4893 4948
4894 4949 g1_policy()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
4895 4950
4896 4951 // Scan strong roots in mark stack.
4897 4952 if (!_process_strong_tasks->is_task_claimed(G1H_PS_mark_stack_oops_do)) {
4898 4953 concurrent_mark()->oops_do(scan_non_heap_roots);
4899 4954 }
4900 4955 double mark_stack_scan_ms = (os::elapsedTime() - ext_roots_end) * 1000.0;
4901 4956 g1_policy()->record_mark_stack_scan_time(worker_i, mark_stack_scan_ms);
4902 4957
4903 4958 // Now scan the complement of the collection set.
4904 4959 if (scan_rs != NULL) {
4905 4960 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
4906 4961 }
4907 4962
4908 4963 _process_strong_tasks->all_tasks_completed();
4909 4964 }
4910 4965
4911 4966 void
4912 4967 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
4913 4968 OopClosure* non_root_closure) {
4914 4969 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
4915 4970 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
4916 4971 }
4917 4972
4918 4973 // Weak Reference Processing support
4919 4974
4920 4975 // An always "is_alive" closure that is used to preserve referents.
4921 4976 // If the object is non-null then it's alive. Used in the preservation
4922 4977 // of referent objects that are pointed to by reference objects
4923 4978 // discovered by the CM ref processor.
4924 4979 class G1AlwaysAliveClosure: public BoolObjectClosure {
4925 4980 G1CollectedHeap* _g1;
4926 4981 public:
4927 4982 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4928 4983 void do_object(oop p) { assert(false, "Do not call."); }
4929 4984 bool do_object_b(oop p) {
4930 4985 if (p != NULL) {
4931 4986 return true;
4932 4987 }
4933 4988 return false;
4934 4989 }
4935 4990 };
4936 4991
4937 4992 bool G1STWIsAliveClosure::do_object_b(oop p) {
4938 4993 // An object is reachable if it is outside the collection set,
4939 4994 // or is inside and copied.
4940 4995 return !_g1->obj_in_cs(p) || p->is_forwarded();
4941 4996 }
4942 4997
4943 4998 // Non Copying Keep Alive closure
4944 4999 class G1KeepAliveClosure: public OopClosure {
4945 5000 G1CollectedHeap* _g1;
4946 5001 public:
4947 5002 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
4948 5003 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
4949 5004 void do_oop( oop* p) {
4950 5005 oop obj = *p;
4951 5006
4952 5007 if (_g1->obj_in_cs(obj)) {
4953 5008 assert( obj->is_forwarded(), "invariant" );
4954 5009 *p = obj->forwardee();
4955 5010 }
4956 5011 }
4957 5012 };
4958 5013
4959 5014 // Copying Keep Alive closure - can be called from both
4960 5015 // serial and parallel code as long as different worker
4961 5016 // threads utilize different G1ParScanThreadState instances
4962 5017 // and different queues.
4963 5018
4964 5019 class G1CopyingKeepAliveClosure: public OopClosure {
4965 5020 G1CollectedHeap* _g1h;
4966 5021 OopClosure* _copy_non_heap_obj_cl;
4967 5022 OopsInHeapRegionClosure* _copy_perm_obj_cl;
4968 5023 G1ParScanThreadState* _par_scan_state;
4969 5024
4970 5025 public:
4971 5026 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
4972 5027 OopClosure* non_heap_obj_cl,
4973 5028 OopsInHeapRegionClosure* perm_obj_cl,
4974 5029 G1ParScanThreadState* pss):
4975 5030 _g1h(g1h),
4976 5031 _copy_non_heap_obj_cl(non_heap_obj_cl),
4977 5032 _copy_perm_obj_cl(perm_obj_cl),
4978 5033 _par_scan_state(pss)
4979 5034 {}
4980 5035
4981 5036 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
4982 5037 virtual void do_oop( oop* p) { do_oop_work(p); }
4983 5038
4984 5039 template <class T> void do_oop_work(T* p) {
4985 5040 oop obj = oopDesc::load_decode_heap_oop(p);
4986 5041
4987 5042 if (_g1h->obj_in_cs(obj)) {
4988 5043 // If the referent object has been forwarded (either copied
4989 5044 // to a new location or to itself in the event of an
4990 5045 // evacuation failure) then we need to update the reference
4991 5046 // field and, if both reference and referent are in the G1
4992 5047 // heap, update the RSet for the referent.
4993 5048 //
4994 5049 // If the referent has not been forwarded then we have to keep
4995 5050 // it alive by policy. Therefore we have copy the referent.
4996 5051 //
4997 5052 // If the reference field is in the G1 heap then we can push
4998 5053 // on the PSS queue. When the queue is drained (after each
4999 5054 // phase of reference processing) the object and it's followers
5000 5055 // will be copied, the reference field set to point to the
5001 5056 // new location, and the RSet updated. Otherwise we need to
5002 5057 // use the the non-heap or perm closures directly to copy
5003 5058 // the refernt object and update the pointer, while avoiding
5004 5059 // updating the RSet.
5005 5060
5006 5061 if (_g1h->is_in_g1_reserved(p)) {
5007 5062 _par_scan_state->push_on_queue(p);
5008 5063 } else {
5009 5064 // The reference field is not in the G1 heap.
5010 5065 if (_g1h->perm_gen()->is_in(p)) {
5011 5066 _copy_perm_obj_cl->do_oop(p);
5012 5067 } else {
5013 5068 _copy_non_heap_obj_cl->do_oop(p);
5014 5069 }
5015 5070 }
5016 5071 }
5017 5072 }
5018 5073 };
5019 5074
5020 5075 // Serial drain queue closure. Called as the 'complete_gc'
5021 5076 // closure for each discovered list in some of the
5022 5077 // reference processing phases.
5023 5078
5024 5079 class G1STWDrainQueueClosure: public VoidClosure {
5025 5080 protected:
5026 5081 G1CollectedHeap* _g1h;
5027 5082 G1ParScanThreadState* _par_scan_state;
5028 5083
5029 5084 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5030 5085
5031 5086 public:
5032 5087 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5033 5088 _g1h(g1h),
5034 5089 _par_scan_state(pss)
5035 5090 { }
5036 5091
5037 5092 void do_void() {
5038 5093 G1ParScanThreadState* const pss = par_scan_state();
5039 5094 pss->trim_queue();
5040 5095 }
5041 5096 };
5042 5097
5043 5098 // Parallel Reference Processing closures
5044 5099
5045 5100 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5046 5101 // processing during G1 evacuation pauses.
5047 5102
5048 5103 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5049 5104 private:
5050 5105 G1CollectedHeap* _g1h;
5051 5106 RefToScanQueueSet* _queues;
5052 5107 FlexibleWorkGang* _workers;
5053 5108 int _active_workers;
5054 5109
5055 5110 public:
5056 5111 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5057 5112 FlexibleWorkGang* workers,
5058 5113 RefToScanQueueSet *task_queues,
5059 5114 int n_workers) :
5060 5115 _g1h(g1h),
5061 5116 _queues(task_queues),
5062 5117 _workers(workers),
5063 5118 _active_workers(n_workers)
5064 5119 {
5065 5120 assert(n_workers > 0, "shouldn't call this otherwise");
5066 5121 }
5067 5122
5068 5123 // Executes the given task using concurrent marking worker threads.
5069 5124 virtual void execute(ProcessTask& task);
5070 5125 virtual void execute(EnqueueTask& task);
5071 5126 };
5072 5127
5073 5128 // Gang task for possibly parallel reference processing
5074 5129
5075 5130 class G1STWRefProcTaskProxy: public AbstractGangTask {
5076 5131 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5077 5132 ProcessTask& _proc_task;
5078 5133 G1CollectedHeap* _g1h;
5079 5134 RefToScanQueueSet *_task_queues;
5080 5135 ParallelTaskTerminator* _terminator;
5081 5136
5082 5137 public:
5083 5138 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5084 5139 G1CollectedHeap* g1h,
5085 5140 RefToScanQueueSet *task_queues,
5086 5141 ParallelTaskTerminator* terminator) :
5087 5142 AbstractGangTask("Process reference objects in parallel"),
5088 5143 _proc_task(proc_task),
5089 5144 _g1h(g1h),
5090 5145 _task_queues(task_queues),
5091 5146 _terminator(terminator)
5092 5147 {}
5093 5148
5094 5149 virtual void work(int i) {
5095 5150 // The reference processing task executed by a single worker.
5096 5151 ResourceMark rm;
5097 5152 HandleMark hm;
5098 5153
5099 5154 G1STWIsAliveClosure is_alive(_g1h);
5100 5155
5101 5156 G1ParScanThreadState pss(_g1h, i);
5102 5157
5103 5158 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5104 5159 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5105 5160 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5106 5161
5107 5162 pss.set_evac_closure(&scan_evac_cl);
5108 5163 pss.set_evac_failure_closure(&evac_failure_cl);
5109 5164 pss.set_partial_scan_closure(&partial_scan_cl);
5110 5165
5111 5166 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5112 5167 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5113 5168
5114 5169 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5115 5170 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5116 5171
5117 5172 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5118 5173 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5119 5174
5120 5175 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5121 5176 // We also need to mark copied objects.
5122 5177 copy_non_heap_cl = ©_mark_non_heap_cl;
5123 5178 copy_perm_cl = ©_mark_perm_cl;
5124 5179 }
5125 5180
5126 5181 // Keep alive closure.
5127 5182 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5128 5183
5129 5184 // Complete GC closure
5130 5185 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5131 5186
5132 5187 // Call the reference processing task's work routine.
5133 5188 _proc_task.work(i, is_alive, keep_alive, drain_queue);
5134 5189
5135 5190 // Note we cannot assert that the refs array is empty here as not all
5136 5191 // of the processing tasks (specifically phase2 - pp2_work) execute
5137 5192 // the complete_gc closure (which ordinarily would drain the queue) so
5138 5193 // the queue may not be empty.
5139 5194 }
5140 5195 };
5141 5196
5142 5197 // Driver routine for parallel reference processing.
5143 5198 // Creates an instance of the ref processing gang
5144 5199 // task and has the worker threads execute it.
5145 5200 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5146 5201 assert(_workers != NULL, "Need parallel worker threads.");
5147 5202
5148 5203 ParallelTaskTerminator terminator(_active_workers, _queues);
5149 5204 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5150 5205
5151 5206 _g1h->set_par_threads(_active_workers);
5152 5207 _workers->run_task(&proc_task_proxy);
5153 5208 _g1h->set_par_threads(0);
5154 5209 }
5155 5210
5156 5211 // Gang task for parallel reference enqueueing.
5157 5212
5158 5213 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5159 5214 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5160 5215 EnqueueTask& _enq_task;
5161 5216
5162 5217 public:
5163 5218 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5164 5219 AbstractGangTask("Enqueue reference objects in parallel"),
5165 5220 _enq_task(enq_task)
5166 5221 { }
5167 5222
5168 5223 virtual void work(int i) {
5169 5224 _enq_task.work(i);
5170 5225 }
5171 5226 };
5172 5227
5173 5228 // Driver routine for parallel reference enqueing.
5174 5229 // Creates an instance of the ref enqueueing gang
5175 5230 // task and has the worker threads execute it.
5176 5231
5177 5232 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5178 5233 assert(_workers != NULL, "Need parallel worker threads.");
5179 5234
5180 5235 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5181 5236
5182 5237 _g1h->set_par_threads(_active_workers);
5183 5238 _workers->run_task(&enq_task_proxy);
5184 5239 _g1h->set_par_threads(0);
5185 5240 }
5186 5241
5187 5242 // End of weak reference support closures
5188 5243
5189 5244 // Abstract task used to preserve (i.e. copy) any referent objects
5190 5245 // that are in the collection set and are pointed to by reference
5191 5246 // objects discovered by the CM ref processor.
5192 5247
5193 5248 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5194 5249 protected:
5195 5250 G1CollectedHeap* _g1h;
5196 5251 RefToScanQueueSet *_queues;
5197 5252 ParallelTaskTerminator _terminator;
5198 5253 int _n_workers;
5199 5254
5200 5255 public:
5201 5256 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5202 5257 AbstractGangTask("ParPreserveCMReferents"),
5203 5258 _g1h(g1h),
5204 5259 _queues(task_queues),
5205 5260 _terminator(workers, _queues),
5206 5261 _n_workers(workers)
5207 5262 { }
5208 5263
5209 5264 void work(int i) {
5210 5265 ResourceMark rm;
5211 5266 HandleMark hm;
5212 5267
5213 5268 G1ParScanThreadState pss(_g1h, i);
5214 5269 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5215 5270 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5216 5271 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5217 5272
5218 5273 pss.set_evac_closure(&scan_evac_cl);
5219 5274 pss.set_evac_failure_closure(&evac_failure_cl);
5220 5275 pss.set_partial_scan_closure(&partial_scan_cl);
5221 5276
5222 5277 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5223 5278
5224 5279
5225 5280 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5226 5281 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5227 5282
5228 5283 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5229 5284 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5230 5285
5231 5286 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5232 5287 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5233 5288
5234 5289 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5235 5290 // We also need to mark copied objects.
5236 5291 copy_non_heap_cl = ©_mark_non_heap_cl;
5237 5292 copy_perm_cl = ©_mark_perm_cl;
5238 5293 }
5239 5294
5240 5295 // Is alive closure
5241 5296 G1AlwaysAliveClosure always_alive(_g1h);
5242 5297
5243 5298 // Copying keep alive closure. Applied to referent objects that need
5244 5299 // to be copied.
5245 5300 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5246 5301
5247 5302 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5248 5303
5249 5304 int limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5250 5305 int stride = MIN2(MAX2(_n_workers, 1), limit);
5251 5306
5252 5307 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5253 5308 // So this must be true - but assert just in case someone decides to
5254 5309 // change the worker ids.
5255 5310 assert(0 <= i && i < limit, "sanity");
5256 5311 assert(!rp->discovery_is_atomic(), "check this code");
5257 5312
5258 5313 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5259 5314 for (int idx = i; idx < limit; idx += stride) {
5260 5315 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5261 5316
5262 5317 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5263 5318 while (iter.has_next()) {
5264 5319 // Since discovery is not atomic for the CM ref processor, we
5265 5320 // can see some null referent objects.
5266 5321 iter.load_ptrs(DEBUG_ONLY(true));
5267 5322 oop ref = iter.obj();
5268 5323
5269 5324 // This will filter nulls.
5270 5325 if (iter.is_referent_alive()) {
5271 5326 iter.make_referent_alive();
5272 5327 }
5273 5328 iter.move_to_next();
5274 5329 }
5275 5330 }
5276 5331
5277 5332 // Drain the queue - which may cause stealing
5278 5333 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5279 5334 drain_queue.do_void();
5280 5335 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5281 5336 assert(pss.refs()->is_empty(), "should be");
5282 5337 }
5283 5338 };
5284 5339
5285 5340 // Weak Reference processing during an evacuation pause (part 1).
5286 5341 void G1CollectedHeap::process_discovered_references() {
5287 5342 double ref_proc_start = os::elapsedTime();
5288 5343
5289 5344 ReferenceProcessor* rp = _ref_processor_stw;
5290 5345 assert(rp->discovery_enabled(), "should have been enabled");
5291 5346
5292 5347 // Any reference objects, in the collection set, that were 'discovered'
5293 5348 // by the CM ref processor should have already been copied (either by
5294 5349 // applying the external root copy closure to the discovered lists, or
5295 5350 // by following an RSet entry).
5296 5351 //
5297 5352 // But some of the referents, that are in the collection set, that these
5298 5353 // reference objects point to may not have been copied: the STW ref
5299 5354 // processor would have seen that the reference object had already
5300 5355 // been 'discovered' and would have skipped discovering the reference,
5301 5356 // but would not have treated the reference object as a regular oop.
5302 5357 // As a reult the copy closure would not have been applied to the
5303 5358 // referent object.
5304 5359 //
5305 5360 // We need to explicitly copy these referent objects - the references
5306 5361 // will be processed at the end of remarking.
5307 5362 //
5308 5363 // We also need to do this copying before we process the reference
5309 5364 // objects discovered by the STW ref processor in case one of these
5310 5365 // referents points to another object which is also referenced by an
5311 5366 // object discovered by the STW ref processor.
5312 5367
5313 5368 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5314 5369 workers()->active_workers() : 1);
5315 5370
5316 5371 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5317 5372 active_workers == workers()->active_workers(),
5318 5373 "Need to reset active_workers");
5319 5374
5320 5375 set_par_threads(active_workers);
5321 5376 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues);
5322 5377
5323 5378 if (G1CollectedHeap::use_parallel_gc_threads()) {
5324 5379 workers()->run_task(&keep_cm_referents);
5325 5380 } else {
5326 5381 keep_cm_referents.work(0);
5327 5382 }
5328 5383
5329 5384 set_par_threads(0);
5330 5385
5331 5386 // Closure to test whether a referent is alive.
5332 5387 G1STWIsAliveClosure is_alive(this);
5333 5388
5334 5389 // Even when parallel reference processing is enabled, the processing
5335 5390 // of JNI refs is serial and performed serially by the current thread
5336 5391 // rather than by a worker. The following PSS will be used for processing
5337 5392 // JNI refs.
5338 5393
5339 5394 // Use only a single queue for this PSS.
5340 5395 G1ParScanThreadState pss(this, 0);
5341 5396
5342 5397 // We do not embed a reference processor in the copying/scanning
5343 5398 // closures while we're actually processing the discovered
5344 5399 // reference objects.
5345 5400 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5346 5401 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5347 5402 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5348 5403
5349 5404 pss.set_evac_closure(&scan_evac_cl);
5350 5405 pss.set_evac_failure_closure(&evac_failure_cl);
5351 5406 pss.set_partial_scan_closure(&partial_scan_cl);
5352 5407
5353 5408 assert(pss.refs()->is_empty(), "pre-condition");
5354 5409
5355 5410 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5356 5411 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL);
5357 5412
5358 5413 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5359 5414 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL);
5360 5415
5361 5416 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5362 5417 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5363 5418
5364 5419 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5365 5420 // We also need to mark copied objects.
5366 5421 copy_non_heap_cl = ©_mark_non_heap_cl;
5367 5422 copy_perm_cl = ©_mark_perm_cl;
5368 5423 }
5369 5424
5370 5425 // Keep alive closure.
5371 5426 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss);
5372 5427
5373 5428 // Serial Complete GC closure
5374 5429 G1STWDrainQueueClosure drain_queue(this, &pss);
5375 5430
5376 5431 // Setup the soft refs policy...
5377 5432 rp->setup_policy(false);
5378 5433
5379 5434 if (!rp->processing_is_mt()) {
5380 5435 // Serial reference processing...
5381 5436 rp->process_discovered_references(&is_alive,
5382 5437 &keep_alive,
5383 5438 &drain_queue,
5384 5439 NULL);
5385 5440 } else {
5386 5441 // Parallel reference processing
5387 5442 assert(rp->num_q() == active_workers, "sanity");
5388 5443 assert(active_workers <= rp->max_num_q(), "sanity");
5389 5444
5390 5445 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5391 5446 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5392 5447 }
5393 5448
5394 5449 // We have completed copying any necessary live referent objects
5395 5450 // (that were not copied during the actual pause) so we can
5396 5451 // retire any active alloc buffers
5397 5452 pss.retire_alloc_buffers();
5398 5453 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5399 5454
5400 5455 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5401 5456 g1_policy()->record_ref_proc_time(ref_proc_time * 1000.0);
5402 5457 }
5403 5458
5404 5459 // Weak Reference processing during an evacuation pause (part 2).
5405 5460 void G1CollectedHeap::enqueue_discovered_references() {
5406 5461 double ref_enq_start = os::elapsedTime();
5407 5462
5408 5463 ReferenceProcessor* rp = _ref_processor_stw;
5409 5464 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5410 5465
5411 5466 // Now enqueue any remaining on the discovered lists on to
5412 5467 // the pending list.
5413 5468 if (!rp->processing_is_mt()) {
5414 5469 // Serial reference processing...
5415 5470 rp->enqueue_discovered_references();
5416 5471 } else {
5417 5472 // Parallel reference enqueuing
5418 5473
5419 5474 int active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1);
5420 5475 assert(active_workers == workers()->active_workers(),
5421 5476 "Need to reset active_workers");
5422 5477 assert(rp->num_q() == active_workers, "sanity");
5423 5478 assert(active_workers <= rp->max_num_q(), "sanity");
5424 5479
5425 5480 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5426 5481 rp->enqueue_discovered_references(&par_task_executor);
5427 5482 }
5428 5483
5429 5484 rp->verify_no_references_recorded();
5430 5485 assert(!rp->discovery_enabled(), "should have been disabled");
5431 5486
5432 5487 // FIXME
5433 5488 // CM's reference processing also cleans up the string and symbol tables.
5434 5489 // Should we do that here also? We could, but it is a serial operation
5435 5490 // and could signicantly increase the pause time.
5436 5491
5437 5492 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5438 5493 g1_policy()->record_ref_enq_time(ref_enq_time * 1000.0);
5439 5494 }
5440 5495
5441 5496 void G1CollectedHeap::evacuate_collection_set() {
5442 5497 set_evacuation_failed(false);
5443 5498
5444 5499 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5445 5500 concurrent_g1_refine()->set_use_cache(false);
5446 5501 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5447 5502
5448 5503 int n_workers;
5449 5504 if (G1CollectedHeap::use_parallel_gc_threads()) {
5450 5505 n_workers =
5451 5506 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5452 5507 workers()->active_workers(),
5453 5508 Threads::number_of_non_daemon_threads());
5454 5509 assert(UseDynamicNumberOfGCThreads ||
5455 5510 n_workers == workers()->total_workers(),
5456 5511 "If not dynamic should be using all the workers");
5457 5512 workers()->set_active_workers(n_workers);
5458 5513 set_par_threads(n_workers);
5459 5514 } else {
5460 5515 assert(n_par_threads() == 0,
5461 5516 "Should be the original non-parallel value");
5462 5517 n_workers = 1;
5463 5518 }
5464 5519
5465 5520 G1ParTask g1_par_task(this, _task_queues);
5466 5521
5467 5522 init_for_evac_failure(NULL);
5468 5523
5469 5524 rem_set()->prepare_for_younger_refs_iterate(true);
5470 5525
5471 5526 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5472 5527 double start_par = os::elapsedTime();
5473 5528
5474 5529 if (G1CollectedHeap::use_parallel_gc_threads()) {
5475 5530 // The individual threads will set their evac-failure closures.
5476 5531 StrongRootsScope srs(this);
5477 5532 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5478 5533 // These tasks use ShareHeap::_process_strong_tasks
5479 5534 assert(UseDynamicNumberOfGCThreads ||
5480 5535 workers()->active_workers() == workers()->total_workers(),
5481 5536 "If not dynamic should be using all the workers");
5482 5537 workers()->run_task(&g1_par_task);
5483 5538 } else {
5484 5539 StrongRootsScope srs(this);
5485 5540 g1_par_task.set_for_termination(n_workers);
5486 5541 g1_par_task.work(0);
5487 5542 }
5488 5543
5489 5544 double par_time = (os::elapsedTime() - start_par) * 1000.0;
5490 5545 g1_policy()->record_par_time(par_time);
5491 5546
5492 5547 set_par_threads(0);
5493 5548
5494 5549 // Process any discovered reference objects - we have
5495 5550 // to do this _before_ we retire the GC alloc regions
5496 5551 // as we may have to copy some 'reachable' referent
5497 5552 // objects (and their reachable sub-graphs) that were
5498 5553 // not copied during the pause.
5499 5554 process_discovered_references();
5500 5555
5501 5556 // Weak root processing.
5502 5557 // Note: when JSR 292 is enabled and code blobs can contain
5503 5558 // non-perm oops then we will need to process the code blobs
5504 5559 // here too.
5505 5560 {
5506 5561 G1STWIsAliveClosure is_alive(this);
5507 5562 G1KeepAliveClosure keep_alive(this);
5508 5563 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5509 5564 }
5510 5565
5511 5566 release_gc_alloc_regions();
5512 5567 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5513 5568
5514 5569 concurrent_g1_refine()->clear_hot_cache();
5515 5570 concurrent_g1_refine()->set_use_cache(true);
5516 5571
5517 5572 finalize_for_evac_failure();
5518 5573
5519 5574 // Must do this before clearing the per-region evac-failure flags
5520 5575 // (which is currently done when we free the collection set).
5521 5576 // We also only do this if marking is actually in progress and so
5522 5577 // have to do this before we set the mark_in_progress flag at the
5523 5578 // end of an initial mark pause.
5524 5579 concurrent_mark()->complete_marking_in_collection_set();
5525 5580
5526 5581 if (evacuation_failed()) {
5527 5582 remove_self_forwarding_pointers();
5528 5583 if (PrintGCDetails) {
5529 5584 gclog_or_tty->print(" (to-space overflow)");
5530 5585 } else if (PrintGC) {
5531 5586 gclog_or_tty->print("--");
5532 5587 }
5533 5588 }
5534 5589
5535 5590 // Enqueue any remaining references remaining on the STW
5536 5591 // reference processor's discovered lists. We need to do
5537 5592 // this after the card table is cleaned (and verified) as
5538 5593 // the act of enqueuing entries on to the pending list
5539 5594 // will log these updates (and dirty their associated
5540 5595 // cards). We need these updates logged to update any
5541 5596 // RSets.
5542 5597 enqueue_discovered_references();
5543 5598
5544 5599 if (G1DeferredRSUpdate) {
5545 5600 RedirtyLoggedCardTableEntryFastClosure redirty;
5546 5601 dirty_card_queue_set().set_closure(&redirty);
5547 5602 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5548 5603
5549 5604 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5550 5605 dcq.merge_bufferlists(&dirty_card_queue_set());
5551 5606 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5552 5607 }
5553 5608 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5554 5609 }
5555 5610
5556 5611 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5557 5612 size_t* pre_used,
5558 5613 FreeRegionList* free_list,
5559 5614 OldRegionSet* old_proxy_set,
5560 5615 HumongousRegionSet* humongous_proxy_set,
5561 5616 HRRSCleanupTask* hrrs_cleanup_task,
5562 5617 bool par) {
5563 5618 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5564 5619 if (hr->isHumongous()) {
5565 5620 assert(hr->startsHumongous(), "we should only see starts humongous");
5566 5621 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5567 5622 } else {
5568 5623 _old_set.remove_with_proxy(hr, old_proxy_set);
5569 5624 free_region(hr, pre_used, free_list, par);
5570 5625 }
5571 5626 } else {
5572 5627 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5573 5628 }
5574 5629 }
5575 5630
5576 5631 void G1CollectedHeap::free_region(HeapRegion* hr,
5577 5632 size_t* pre_used,
5578 5633 FreeRegionList* free_list,
5579 5634 bool par) {
5580 5635 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5581 5636 assert(!hr->is_empty(), "the region should not be empty");
5582 5637 assert(free_list != NULL, "pre-condition");
5583 5638
5584 5639 *pre_used += hr->used();
5585 5640 hr->hr_clear(par, true /* clear_space */);
5586 5641 free_list->add_as_head(hr);
5587 5642 }
5588 5643
5589 5644 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5590 5645 size_t* pre_used,
5591 5646 FreeRegionList* free_list,
5592 5647 HumongousRegionSet* humongous_proxy_set,
5593 5648 bool par) {
5594 5649 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5595 5650 assert(free_list != NULL, "pre-condition");
5596 5651 assert(humongous_proxy_set != NULL, "pre-condition");
5597 5652
5598 5653 size_t hr_used = hr->used();
5599 5654 size_t hr_capacity = hr->capacity();
5600 5655 size_t hr_pre_used = 0;
5601 5656 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5602 5657 hr->set_notHumongous();
5603 5658 free_region(hr, &hr_pre_used, free_list, par);
5604 5659
5605 5660 size_t i = hr->hrs_index() + 1;
5606 5661 size_t num = 1;
5607 5662 while (i < n_regions()) {
5608 5663 HeapRegion* curr_hr = region_at(i);
5609 5664 if (!curr_hr->continuesHumongous()) {
5610 5665 break;
5611 5666 }
5612 5667 curr_hr->set_notHumongous();
5613 5668 free_region(curr_hr, &hr_pre_used, free_list, par);
5614 5669 num += 1;
5615 5670 i += 1;
5616 5671 }
5617 5672 assert(hr_pre_used == hr_used,
5618 5673 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5619 5674 "should be the same", hr_pre_used, hr_used));
5620 5675 *pre_used += hr_pre_used;
5621 5676 }
5622 5677
5623 5678 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5624 5679 FreeRegionList* free_list,
5625 5680 OldRegionSet* old_proxy_set,
5626 5681 HumongousRegionSet* humongous_proxy_set,
5627 5682 bool par) {
5628 5683 if (pre_used > 0) {
5629 5684 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5630 5685 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5631 5686 assert(_summary_bytes_used >= pre_used,
5632 5687 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5633 5688 "should be >= pre_used: "SIZE_FORMAT,
5634 5689 _summary_bytes_used, pre_used));
5635 5690 _summary_bytes_used -= pre_used;
5636 5691 }
5637 5692 if (free_list != NULL && !free_list->is_empty()) {
5638 5693 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5639 5694 _free_list.add_as_head(free_list);
5640 5695 }
5641 5696 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5642 5697 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5643 5698 _old_set.update_from_proxy(old_proxy_set);
5644 5699 }
5645 5700 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5646 5701 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5647 5702 _humongous_set.update_from_proxy(humongous_proxy_set);
5648 5703 }
5649 5704 }
5650 5705
5651 5706 class G1ParCleanupCTTask : public AbstractGangTask {
5652 5707 CardTableModRefBS* _ct_bs;
5653 5708 G1CollectedHeap* _g1h;
5654 5709 HeapRegion* volatile _su_head;
5655 5710 public:
5656 5711 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5657 5712 G1CollectedHeap* g1h) :
5658 5713 AbstractGangTask("G1 Par Cleanup CT Task"),
5659 5714 _ct_bs(ct_bs), _g1h(g1h) { }
5660 5715
5661 5716 void work(int i) {
5662 5717 HeapRegion* r;
5663 5718 while (r = _g1h->pop_dirty_cards_region()) {
5664 5719 clear_cards(r);
5665 5720 }
5666 5721 }
5667 5722
5668 5723 void clear_cards(HeapRegion* r) {
5669 5724 // Cards of the survivors should have already been dirtied.
5670 5725 if (!r->is_survivor()) {
5671 5726 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5672 5727 }
5673 5728 }
5674 5729 };
5675 5730
5676 5731 #ifndef PRODUCT
5677 5732 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5678 5733 G1CollectedHeap* _g1h;
5679 5734 CardTableModRefBS* _ct_bs;
5680 5735 public:
5681 5736 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5682 5737 : _g1h(g1h), _ct_bs(ct_bs) { }
5683 5738 virtual bool doHeapRegion(HeapRegion* r) {
5684 5739 if (r->is_survivor()) {
5685 5740 _g1h->verify_dirty_region(r);
5686 5741 } else {
5687 5742 _g1h->verify_not_dirty_region(r);
5688 5743 }
5689 5744 return false;
5690 5745 }
5691 5746 };
5692 5747
5693 5748 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5694 5749 // All of the region should be clean.
5695 5750 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5696 5751 MemRegion mr(hr->bottom(), hr->end());
5697 5752 ct_bs->verify_not_dirty_region(mr);
5698 5753 }
5699 5754
5700 5755 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5701 5756 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5702 5757 // dirty allocated blocks as they allocate them. The thread that
5703 5758 // retires each region and replaces it with a new one will do a
5704 5759 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5705 5760 // not dirty that area (one less thing to have to do while holding
5706 5761 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5707 5762 // is dirty.
5708 5763 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5709 5764 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5710 5765 ct_bs->verify_dirty_region(mr);
5711 5766 }
5712 5767
5713 5768 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5714 5769 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5715 5770 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5716 5771 verify_dirty_region(hr);
5717 5772 }
5718 5773 }
5719 5774
5720 5775 void G1CollectedHeap::verify_dirty_young_regions() {
5721 5776 verify_dirty_young_list(_young_list->first_region());
5722 5777 verify_dirty_young_list(_young_list->first_survivor_region());
5723 5778 }
5724 5779 #endif
5725 5780
5726 5781 void G1CollectedHeap::cleanUpCardTable() {
5727 5782 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5728 5783 double start = os::elapsedTime();
5729 5784
5730 5785 {
5731 5786 // Iterate over the dirty cards region list.
5732 5787 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5733 5788
5734 5789 if (G1CollectedHeap::use_parallel_gc_threads()) {
5735 5790 set_par_threads();
5736 5791 workers()->run_task(&cleanup_task);
5737 5792 set_par_threads(0);
5738 5793 } else {
5739 5794 while (_dirty_cards_region_list) {
5740 5795 HeapRegion* r = _dirty_cards_region_list;
5741 5796 cleanup_task.clear_cards(r);
5742 5797 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5743 5798 if (_dirty_cards_region_list == r) {
5744 5799 // The last region.
5745 5800 _dirty_cards_region_list = NULL;
5746 5801 }
5747 5802 r->set_next_dirty_cards_region(NULL);
5748 5803 }
5749 5804 }
5750 5805 #ifndef PRODUCT
5751 5806 if (G1VerifyCTCleanup || VerifyAfterGC) {
5752 5807 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5753 5808 heap_region_iterate(&cleanup_verifier);
5754 5809 }
5755 5810 #endif
5756 5811 }
5757 5812
5758 5813 double elapsed = os::elapsedTime() - start;
5759 5814 g1_policy()->record_clear_ct_time(elapsed * 1000.0);
5760 5815 }
5761 5816
5762 5817 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5763 5818 size_t pre_used = 0;
5764 5819 FreeRegionList local_free_list("Local List for CSet Freeing");
5765 5820
5766 5821 double young_time_ms = 0.0;
5767 5822 double non_young_time_ms = 0.0;
5768 5823
5769 5824 // Since the collection set is a superset of the the young list,
5770 5825 // all we need to do to clear the young list is clear its
5771 5826 // head and length, and unlink any young regions in the code below
5772 5827 _young_list->clear();
5773 5828
5774 5829 G1CollectorPolicy* policy = g1_policy();
5775 5830
5776 5831 double start_sec = os::elapsedTime();
5777 5832 bool non_young = true;
5778 5833
5779 5834 HeapRegion* cur = cs_head;
5780 5835 int age_bound = -1;
5781 5836 size_t rs_lengths = 0;
5782 5837
5783 5838 while (cur != NULL) {
5784 5839 assert(!is_on_master_free_list(cur), "sanity");
5785 5840 if (non_young) {
5786 5841 if (cur->is_young()) {
5787 5842 double end_sec = os::elapsedTime();
5788 5843 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5789 5844 non_young_time_ms += elapsed_ms;
5790 5845
5791 5846 start_sec = os::elapsedTime();
5792 5847 non_young = false;
5793 5848 }
5794 5849 } else {
5795 5850 if (!cur->is_young()) {
5796 5851 double end_sec = os::elapsedTime();
5797 5852 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5798 5853 young_time_ms += elapsed_ms;
5799 5854
5800 5855 start_sec = os::elapsedTime();
5801 5856 non_young = true;
5802 5857 }
5803 5858 }
5804 5859
5805 5860 rs_lengths += cur->rem_set()->occupied();
5806 5861
5807 5862 HeapRegion* next = cur->next_in_collection_set();
5808 5863 assert(cur->in_collection_set(), "bad CS");
5809 5864 cur->set_next_in_collection_set(NULL);
5810 5865 cur->set_in_collection_set(false);
5811 5866
5812 5867 if (cur->is_young()) {
5813 5868 int index = cur->young_index_in_cset();
5814 5869 assert(index != -1, "invariant");
5815 5870 assert((size_t) index < policy->young_cset_region_length(), "invariant");
5816 5871 size_t words_survived = _surviving_young_words[index];
5817 5872 cur->record_surv_words_in_group(words_survived);
5818 5873
5819 5874 // At this point the we have 'popped' cur from the collection set
5820 5875 // (linked via next_in_collection_set()) but it is still in the
5821 5876 // young list (linked via next_young_region()). Clear the
5822 5877 // _next_young_region field.
5823 5878 cur->set_next_young_region(NULL);
5824 5879 } else {
5825 5880 int index = cur->young_index_in_cset();
5826 5881 assert(index == -1, "invariant");
5827 5882 }
5828 5883
5829 5884 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5830 5885 (!cur->is_young() && cur->young_index_in_cset() == -1),
5831 5886 "invariant" );
5832 5887
5833 5888 if (!cur->evacuation_failed()) {
5834 5889 MemRegion used_mr = cur->used_region();
5835 5890
↓ open down ↓ |
1203 lines elided |
↑ open up ↑ |
5836 5891 // And the region is empty.
5837 5892 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5838 5893
5839 5894 // If marking is in progress then clear any objects marked in
5840 5895 // the current region. Note mark_in_progress() returns false,
5841 5896 // even during an initial mark pause, until the set_marking_started()
5842 5897 // call which takes place later in the pause.
5843 5898 if (mark_in_progress()) {
5844 5899 assert(!g1_policy()->during_initial_mark_pause(), "sanity");
5845 5900 _cm->nextMarkBitMap()->clearRange(used_mr);
5901 + // Need to remove values from the count info
5902 + _cm->clear_count_data_for_heap_region(cur);
5846 5903 }
5847 -
5848 5904 free_region(cur, &pre_used, &local_free_list, false /* par */);
5849 5905 } else {
5850 5906 cur->uninstall_surv_rate_group();
5851 5907 if (cur->is_young()) {
5852 5908 cur->set_young_index_in_cset(-1);
5853 5909 }
5854 5910 cur->set_not_young();
5855 5911 cur->set_evacuation_failed(false);
5856 5912 // The region is now considered to be old.
5857 5913 _old_set.add(cur);
5858 5914 }
5859 5915 cur = next;
5860 5916 }
5861 5917
5862 5918 policy->record_max_rs_lengths(rs_lengths);
5863 5919 policy->cset_regions_freed();
5864 5920
5865 5921 double end_sec = os::elapsedTime();
5866 5922 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5867 5923
5868 5924 if (non_young) {
5869 5925 non_young_time_ms += elapsed_ms;
5870 5926 } else {
5871 5927 young_time_ms += elapsed_ms;
5872 5928 }
5873 5929
5874 5930 update_sets_after_freeing_regions(pre_used, &local_free_list,
5875 5931 NULL /* old_proxy_set */,
5876 5932 NULL /* humongous_proxy_set */,
5877 5933 false /* par */);
5878 5934 policy->record_young_free_cset_time_ms(young_time_ms);
5879 5935 policy->record_non_young_free_cset_time_ms(non_young_time_ms);
5880 5936 }
5881 5937
5882 5938 // This routine is similar to the above but does not record
5883 5939 // any policy statistics or update free lists; we are abandoning
5884 5940 // the current incremental collection set in preparation of a
5885 5941 // full collection. After the full GC we will start to build up
5886 5942 // the incremental collection set again.
5887 5943 // This is only called when we're doing a full collection
5888 5944 // and is immediately followed by the tearing down of the young list.
5889 5945
5890 5946 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
5891 5947 HeapRegion* cur = cs_head;
5892 5948
5893 5949 while (cur != NULL) {
5894 5950 HeapRegion* next = cur->next_in_collection_set();
5895 5951 assert(cur->in_collection_set(), "bad CS");
5896 5952 cur->set_next_in_collection_set(NULL);
5897 5953 cur->set_in_collection_set(false);
5898 5954 cur->set_young_index_in_cset(-1);
5899 5955 cur = next;
5900 5956 }
5901 5957 }
5902 5958
5903 5959 void G1CollectedHeap::set_free_regions_coming() {
5904 5960 if (G1ConcRegionFreeingVerbose) {
5905 5961 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5906 5962 "setting free regions coming");
5907 5963 }
5908 5964
5909 5965 assert(!free_regions_coming(), "pre-condition");
5910 5966 _free_regions_coming = true;
5911 5967 }
5912 5968
5913 5969 void G1CollectedHeap::reset_free_regions_coming() {
5914 5970 {
5915 5971 assert(free_regions_coming(), "pre-condition");
5916 5972 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5917 5973 _free_regions_coming = false;
5918 5974 SecondaryFreeList_lock->notify_all();
5919 5975 }
5920 5976
5921 5977 if (G1ConcRegionFreeingVerbose) {
5922 5978 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
5923 5979 "reset free regions coming");
5924 5980 }
5925 5981 }
5926 5982
5927 5983 void G1CollectedHeap::wait_while_free_regions_coming() {
5928 5984 // Most of the time we won't have to wait, so let's do a quick test
5929 5985 // first before we take the lock.
5930 5986 if (!free_regions_coming()) {
5931 5987 return;
5932 5988 }
5933 5989
5934 5990 if (G1ConcRegionFreeingVerbose) {
5935 5991 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5936 5992 "waiting for free regions");
5937 5993 }
5938 5994
5939 5995 {
5940 5996 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
5941 5997 while (free_regions_coming()) {
5942 5998 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
5943 5999 }
5944 6000 }
5945 6001
5946 6002 if (G1ConcRegionFreeingVerbose) {
5947 6003 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
5948 6004 "done waiting for free regions");
5949 6005 }
5950 6006 }
5951 6007
5952 6008 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
5953 6009 assert(heap_lock_held_for_gc(),
5954 6010 "the heap lock should already be held by or for this thread");
5955 6011 _young_list->push_region(hr);
5956 6012 }
5957 6013
5958 6014 class NoYoungRegionsClosure: public HeapRegionClosure {
5959 6015 private:
5960 6016 bool _success;
5961 6017 public:
5962 6018 NoYoungRegionsClosure() : _success(true) { }
5963 6019 bool doHeapRegion(HeapRegion* r) {
5964 6020 if (r->is_young()) {
5965 6021 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
5966 6022 r->bottom(), r->end());
5967 6023 _success = false;
5968 6024 }
5969 6025 return false;
5970 6026 }
5971 6027 bool success() { return _success; }
5972 6028 };
5973 6029
5974 6030 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
5975 6031 bool ret = _young_list->check_list_empty(check_sample);
5976 6032
5977 6033 if (check_heap) {
5978 6034 NoYoungRegionsClosure closure;
5979 6035 heap_region_iterate(&closure);
5980 6036 ret = ret && closure.success();
5981 6037 }
5982 6038
5983 6039 return ret;
5984 6040 }
5985 6041
5986 6042 class TearDownRegionSetsClosure : public HeapRegionClosure {
5987 6043 private:
5988 6044 OldRegionSet *_old_set;
5989 6045
5990 6046 public:
5991 6047 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
5992 6048
5993 6049 bool doHeapRegion(HeapRegion* r) {
5994 6050 if (r->is_empty()) {
5995 6051 // We ignore empty regions, we'll empty the free list afterwards
5996 6052 } else if (r->is_young()) {
5997 6053 // We ignore young regions, we'll empty the young list afterwards
5998 6054 } else if (r->isHumongous()) {
5999 6055 // We ignore humongous regions, we're not tearing down the
6000 6056 // humongous region set
6001 6057 } else {
6002 6058 // The rest should be old
6003 6059 _old_set->remove(r);
6004 6060 }
6005 6061 return false;
6006 6062 }
6007 6063
6008 6064 ~TearDownRegionSetsClosure() {
6009 6065 assert(_old_set->is_empty(), "post-condition");
6010 6066 }
6011 6067 };
6012 6068
6013 6069 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6014 6070 assert_at_safepoint(true /* should_be_vm_thread */);
6015 6071
6016 6072 if (!free_list_only) {
6017 6073 TearDownRegionSetsClosure cl(&_old_set);
6018 6074 heap_region_iterate(&cl);
6019 6075
6020 6076 // Need to do this after the heap iteration to be able to
6021 6077 // recognize the young regions and ignore them during the iteration.
6022 6078 _young_list->empty_list();
6023 6079 }
6024 6080 _free_list.remove_all();
6025 6081 }
6026 6082
6027 6083 class RebuildRegionSetsClosure : public HeapRegionClosure {
6028 6084 private:
6029 6085 bool _free_list_only;
6030 6086 OldRegionSet* _old_set;
6031 6087 FreeRegionList* _free_list;
6032 6088 size_t _total_used;
6033 6089
6034 6090 public:
6035 6091 RebuildRegionSetsClosure(bool free_list_only,
6036 6092 OldRegionSet* old_set, FreeRegionList* free_list) :
6037 6093 _free_list_only(free_list_only),
6038 6094 _old_set(old_set), _free_list(free_list), _total_used(0) {
6039 6095 assert(_free_list->is_empty(), "pre-condition");
6040 6096 if (!free_list_only) {
6041 6097 assert(_old_set->is_empty(), "pre-condition");
6042 6098 }
6043 6099 }
6044 6100
6045 6101 bool doHeapRegion(HeapRegion* r) {
6046 6102 if (r->continuesHumongous()) {
6047 6103 return false;
6048 6104 }
6049 6105
6050 6106 if (r->is_empty()) {
6051 6107 // Add free regions to the free list
6052 6108 _free_list->add_as_tail(r);
6053 6109 } else if (!_free_list_only) {
6054 6110 assert(!r->is_young(), "we should not come across young regions");
6055 6111
6056 6112 if (r->isHumongous()) {
6057 6113 // We ignore humongous regions, we left the humongous set unchanged
6058 6114 } else {
6059 6115 // The rest should be old, add them to the old set
6060 6116 _old_set->add(r);
6061 6117 }
6062 6118 _total_used += r->used();
6063 6119 }
6064 6120
6065 6121 return false;
6066 6122 }
6067 6123
6068 6124 size_t total_used() {
6069 6125 return _total_used;
6070 6126 }
6071 6127 };
6072 6128
6073 6129 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6074 6130 assert_at_safepoint(true /* should_be_vm_thread */);
6075 6131
6076 6132 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6077 6133 heap_region_iterate(&cl);
6078 6134
6079 6135 if (!free_list_only) {
6080 6136 _summary_bytes_used = cl.total_used();
6081 6137 }
6082 6138 assert(_summary_bytes_used == recalculate_used(),
6083 6139 err_msg("inconsistent _summary_bytes_used, "
6084 6140 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6085 6141 _summary_bytes_used, recalculate_used()));
6086 6142 }
6087 6143
6088 6144 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6089 6145 _refine_cte_cl->set_concurrent(concurrent);
6090 6146 }
6091 6147
6092 6148 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6093 6149 HeapRegion* hr = heap_region_containing(p);
6094 6150 if (hr == NULL) {
6095 6151 return is_in_permanent(p);
6096 6152 } else {
6097 6153 return hr->is_in(p);
6098 6154 }
6099 6155 }
6100 6156
6101 6157 // Methods for the mutator alloc region
6102 6158
6103 6159 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6104 6160 bool force) {
6105 6161 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6106 6162 assert(!force || g1_policy()->can_expand_young_list(),
6107 6163 "if force is true we should be able to expand the young list");
6108 6164 bool young_list_full = g1_policy()->is_young_list_full();
6109 6165 if (force || !young_list_full) {
6110 6166 HeapRegion* new_alloc_region = new_region(word_size,
6111 6167 false /* do_expand */);
6112 6168 if (new_alloc_region != NULL) {
6113 6169 set_region_short_lived_locked(new_alloc_region);
6114 6170 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6115 6171 return new_alloc_region;
6116 6172 }
6117 6173 }
6118 6174 return NULL;
6119 6175 }
6120 6176
6121 6177 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6122 6178 size_t allocated_bytes) {
6123 6179 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6124 6180 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6125 6181
6126 6182 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6127 6183 _summary_bytes_used += allocated_bytes;
6128 6184 _hr_printer.retire(alloc_region);
6129 6185 // We update the eden sizes here, when the region is retired,
6130 6186 // instead of when it's allocated, since this is the point that its
6131 6187 // used space has been recored in _summary_bytes_used.
6132 6188 g1mm()->update_eden_size();
6133 6189 }
6134 6190
6135 6191 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6136 6192 bool force) {
6137 6193 return _g1h->new_mutator_alloc_region(word_size, force);
6138 6194 }
6139 6195
6140 6196 void G1CollectedHeap::set_par_threads() {
6141 6197 // Don't change the number of workers. Use the value previously set
6142 6198 // in the workgroup.
6143 6199 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6144 6200 int n_workers = workers()->active_workers();
6145 6201 assert(UseDynamicNumberOfGCThreads ||
6146 6202 n_workers == workers()->total_workers(),
6147 6203 "Otherwise should be using the total number of workers");
6148 6204 if (n_workers == 0) {
6149 6205 assert(false, "Should have been set in prior evacuation pause.");
6150 6206 n_workers = ParallelGCThreads;
6151 6207 workers()->set_active_workers(n_workers);
6152 6208 }
6153 6209 set_par_threads(n_workers);
6154 6210 }
6155 6211
6156 6212 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6157 6213 size_t allocated_bytes) {
6158 6214 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6159 6215 }
6160 6216
6161 6217 // Methods for the GC alloc regions
6162 6218
6163 6219 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6164 6220 size_t count,
6165 6221 GCAllocPurpose ap) {
6166 6222 assert(FreeList_lock->owned_by_self(), "pre-condition");
6167 6223
6168 6224 if (count < g1_policy()->max_regions(ap)) {
6169 6225 HeapRegion* new_alloc_region = new_region(word_size,
6170 6226 true /* do_expand */);
6171 6227 if (new_alloc_region != NULL) {
6172 6228 // We really only need to do this for old regions given that we
6173 6229 // should never scan survivors. But it doesn't hurt to do it
6174 6230 // for survivors too.
6175 6231 new_alloc_region->set_saved_mark();
6176 6232 if (ap == GCAllocForSurvived) {
6177 6233 new_alloc_region->set_survivor();
6178 6234 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6179 6235 } else {
6180 6236 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6181 6237 }
6182 6238 return new_alloc_region;
6183 6239 } else {
6184 6240 g1_policy()->note_alloc_region_limit_reached(ap);
6185 6241 }
6186 6242 }
6187 6243 return NULL;
6188 6244 }
6189 6245
6190 6246 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6191 6247 size_t allocated_bytes,
6192 6248 GCAllocPurpose ap) {
6193 6249 alloc_region->note_end_of_copying();
6194 6250 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6195 6251 if (ap == GCAllocForSurvived) {
6196 6252 young_list()->add_survivor_region(alloc_region);
6197 6253 } else {
6198 6254 _old_set.add(alloc_region);
6199 6255 }
6200 6256 _hr_printer.retire(alloc_region);
6201 6257 }
6202 6258
6203 6259 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6204 6260 bool force) {
6205 6261 assert(!force, "not supported for GC alloc regions");
6206 6262 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6207 6263 }
6208 6264
6209 6265 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6210 6266 size_t allocated_bytes) {
6211 6267 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6212 6268 GCAllocForSurvived);
6213 6269 }
6214 6270
6215 6271 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6216 6272 bool force) {
6217 6273 assert(!force, "not supported for GC alloc regions");
6218 6274 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6219 6275 }
6220 6276
6221 6277 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6222 6278 size_t allocated_bytes) {
6223 6279 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6224 6280 GCAllocForTenured);
6225 6281 }
6226 6282 // Heap region set verification
6227 6283
6228 6284 class VerifyRegionListsClosure : public HeapRegionClosure {
6229 6285 private:
6230 6286 FreeRegionList* _free_list;
6231 6287 OldRegionSet* _old_set;
6232 6288 HumongousRegionSet* _humongous_set;
6233 6289 size_t _region_count;
6234 6290
6235 6291 public:
6236 6292 VerifyRegionListsClosure(OldRegionSet* old_set,
6237 6293 HumongousRegionSet* humongous_set,
6238 6294 FreeRegionList* free_list) :
6239 6295 _old_set(old_set), _humongous_set(humongous_set),
6240 6296 _free_list(free_list), _region_count(0) { }
6241 6297
6242 6298 size_t region_count() { return _region_count; }
6243 6299
6244 6300 bool doHeapRegion(HeapRegion* hr) {
6245 6301 _region_count += 1;
6246 6302
6247 6303 if (hr->continuesHumongous()) {
6248 6304 return false;
6249 6305 }
6250 6306
6251 6307 if (hr->is_young()) {
6252 6308 // TODO
6253 6309 } else if (hr->startsHumongous()) {
6254 6310 _humongous_set->verify_next_region(hr);
6255 6311 } else if (hr->is_empty()) {
6256 6312 _free_list->verify_next_region(hr);
6257 6313 } else {
6258 6314 _old_set->verify_next_region(hr);
6259 6315 }
6260 6316 return false;
6261 6317 }
6262 6318 };
6263 6319
6264 6320 HeapRegion* G1CollectedHeap::new_heap_region(size_t hrs_index,
6265 6321 HeapWord* bottom) {
6266 6322 HeapWord* end = bottom + HeapRegion::GrainWords;
6267 6323 MemRegion mr(bottom, end);
6268 6324 assert(_g1_reserved.contains(mr), "invariant");
6269 6325 // This might return NULL if the allocation fails
6270 6326 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
6271 6327 }
6272 6328
6273 6329 void G1CollectedHeap::verify_region_sets() {
6274 6330 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6275 6331
6276 6332 // First, check the explicit lists.
6277 6333 _free_list.verify();
6278 6334 {
6279 6335 // Given that a concurrent operation might be adding regions to
6280 6336 // the secondary free list we have to take the lock before
6281 6337 // verifying it.
6282 6338 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6283 6339 _secondary_free_list.verify();
6284 6340 }
6285 6341 _old_set.verify();
6286 6342 _humongous_set.verify();
6287 6343
6288 6344 // If a concurrent region freeing operation is in progress it will
6289 6345 // be difficult to correctly attributed any free regions we come
6290 6346 // across to the correct free list given that they might belong to
6291 6347 // one of several (free_list, secondary_free_list, any local lists,
6292 6348 // etc.). So, if that's the case we will skip the rest of the
6293 6349 // verification operation. Alternatively, waiting for the concurrent
6294 6350 // operation to complete will have a non-trivial effect on the GC's
6295 6351 // operation (no concurrent operation will last longer than the
6296 6352 // interval between two calls to verification) and it might hide
6297 6353 // any issues that we would like to catch during testing.
6298 6354 if (free_regions_coming()) {
6299 6355 return;
6300 6356 }
6301 6357
6302 6358 // Make sure we append the secondary_free_list on the free_list so
6303 6359 // that all free regions we will come across can be safely
6304 6360 // attributed to the free_list.
6305 6361 append_secondary_free_list_if_not_empty_with_lock();
6306 6362
6307 6363 // Finally, make sure that the region accounting in the lists is
6308 6364 // consistent with what we see in the heap.
6309 6365 _old_set.verify_start();
6310 6366 _humongous_set.verify_start();
6311 6367 _free_list.verify_start();
6312 6368
6313 6369 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6314 6370 heap_region_iterate(&cl);
6315 6371
6316 6372 _old_set.verify_end();
6317 6373 _humongous_set.verify_end();
6318 6374 _free_list.verify_end();
6319 6375 }
↓ open down ↓ |
462 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX