Print this page
rev 6875 : 8056240: Investigate increased GC remark time after class unloading changes in CRM Fuse
Reviewed-by: mgerdin, coleenp, bdelsart
Split |
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, 2014, 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
↓ open down ↓ |
19 lines elided |
↑ open up ↑ |
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 #if !defined(__clang_major__) && defined(__GNUC__)
26 26 #define ATTRIBUTE_PRINTF(x,y) // FIXME, formats are a mess.
27 27 #endif
28 28
29 29 #include "precompiled.hpp"
30 +#include "classfile/metadataOnStackMark.hpp"
30 31 #include "code/codeCache.hpp"
31 32 #include "code/icBuffer.hpp"
32 33 #include "gc_implementation/g1/bufferingOopClosure.hpp"
33 34 #include "gc_implementation/g1/concurrentG1Refine.hpp"
34 35 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
35 36 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
36 37 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
37 38 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
38 39 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
39 40 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
40 41 #include "gc_implementation/g1/g1EvacFailure.hpp"
41 42 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
42 43 #include "gc_implementation/g1/g1Log.hpp"
43 44 #include "gc_implementation/g1/g1MarkSweep.hpp"
44 45 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
45 46 #include "gc_implementation/g1/g1ParScanThreadState.inline.hpp"
46 47 #include "gc_implementation/g1/g1RegionToSpaceMapper.hpp"
47 48 #include "gc_implementation/g1/g1RemSet.inline.hpp"
48 49 #include "gc_implementation/g1/g1StringDedup.hpp"
49 50 #include "gc_implementation/g1/g1YCTypes.hpp"
50 51 #include "gc_implementation/g1/heapRegion.inline.hpp"
51 52 #include "gc_implementation/g1/heapRegionRemSet.hpp"
52 53 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
53 54 #include "gc_implementation/g1/vm_operations_g1.hpp"
54 55 #include "gc_implementation/shared/gcHeapSummary.hpp"
55 56 #include "gc_implementation/shared/gcTimer.hpp"
56 57 #include "gc_implementation/shared/gcTrace.hpp"
57 58 #include "gc_implementation/shared/gcTraceTime.hpp"
58 59 #include "gc_implementation/shared/isGCActiveMark.hpp"
59 60 #include "memory/allocation.hpp"
60 61 #include "memory/gcLocker.inline.hpp"
61 62 #include "memory/generationSpec.hpp"
62 63 #include "memory/iterator.hpp"
63 64 #include "memory/referenceProcessor.hpp"
64 65 #include "oops/oop.inline.hpp"
65 66 #include "oops/oop.pcgc.inline.hpp"
66 67 #include "runtime/orderAccess.inline.hpp"
67 68 #include "runtime/vmThread.hpp"
68 69
69 70 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
70 71
71 72 // turn it on so that the contents of the young list (scan-only /
72 73 // to-be-collected) are printed at "strategic" points before / during
73 74 // / after the collection --- this is useful for debugging
74 75 #define YOUNG_LIST_VERBOSE 0
75 76 // CURRENT STATUS
76 77 // This file is under construction. Search for "FIXME".
77 78
78 79 // INVARIANTS/NOTES
79 80 //
80 81 // All allocation activity covered by the G1CollectedHeap interface is
81 82 // serialized by acquiring the HeapLock. This happens in mem_allocate
82 83 // and allocate_new_tlab, which are the "entry" points to the
83 84 // allocation code from the rest of the JVM. (Note that this does not
84 85 // apply to TLAB allocation, which is not part of this interface: it
85 86 // is done by clients of this interface.)
86 87
87 88 // Notes on implementation of parallelism in different tasks.
88 89 //
89 90 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
90 91 // The number of GC workers is passed to heap_region_par_iterate_chunked().
91 92 // It does use run_task() which sets _n_workers in the task.
92 93 // G1ParTask executes g1_process_roots() ->
93 94 // SharedHeap::process_roots() which calls eventually to
94 95 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
95 96 // SequentialSubTasksDone. SharedHeap::process_roots() also
96 97 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
97 98 //
98 99
99 100 // Local to this file.
100 101
101 102 class RefineCardTableEntryClosure: public CardTableEntryClosure {
102 103 bool _concurrent;
103 104 public:
104 105 RefineCardTableEntryClosure() : _concurrent(true) { }
105 106
106 107 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
107 108 bool oops_into_cset = G1CollectedHeap::heap()->g1_rem_set()->refine_card(card_ptr, worker_i, false);
108 109 // This path is executed by the concurrent refine or mutator threads,
109 110 // concurrently, and so we do not care if card_ptr contains references
110 111 // that point into the collection set.
111 112 assert(!oops_into_cset, "should be");
112 113
113 114 if (_concurrent && SuspendibleThreadSet::should_yield()) {
114 115 // Caller will actually yield.
115 116 return false;
116 117 }
117 118 // Otherwise, we finished successfully; return true.
118 119 return true;
119 120 }
120 121
121 122 void set_concurrent(bool b) { _concurrent = b; }
122 123 };
123 124
124 125
125 126 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
126 127 size_t _num_processed;
127 128 CardTableModRefBS* _ctbs;
128 129 int _histo[256];
129 130
130 131 public:
131 132 ClearLoggedCardTableEntryClosure() :
132 133 _num_processed(0), _ctbs(G1CollectedHeap::heap()->g1_barrier_set())
133 134 {
134 135 for (int i = 0; i < 256; i++) _histo[i] = 0;
135 136 }
136 137
137 138 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
138 139 unsigned char* ujb = (unsigned char*)card_ptr;
139 140 int ind = (int)(*ujb);
140 141 _histo[ind]++;
141 142
142 143 *card_ptr = (jbyte)CardTableModRefBS::clean_card_val();
143 144 _num_processed++;
144 145
145 146 return true;
146 147 }
147 148
148 149 size_t num_processed() { return _num_processed; }
149 150
150 151 void print_histo() {
151 152 gclog_or_tty->print_cr("Card table value histogram:");
152 153 for (int i = 0; i < 256; i++) {
153 154 if (_histo[i] != 0) {
154 155 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
155 156 }
156 157 }
157 158 }
158 159 };
159 160
160 161 class RedirtyLoggedCardTableEntryClosure : public CardTableEntryClosure {
161 162 private:
162 163 size_t _num_processed;
163 164
164 165 public:
165 166 RedirtyLoggedCardTableEntryClosure() : CardTableEntryClosure(), _num_processed(0) { }
166 167
167 168 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
168 169 *card_ptr = CardTableModRefBS::dirty_card_val();
169 170 _num_processed++;
170 171 return true;
171 172 }
172 173
173 174 size_t num_processed() const { return _num_processed; }
174 175 };
175 176
176 177 YoungList::YoungList(G1CollectedHeap* g1h) :
177 178 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
178 179 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
179 180 guarantee(check_list_empty(false), "just making sure...");
180 181 }
181 182
182 183 void YoungList::push_region(HeapRegion *hr) {
183 184 assert(!hr->is_young(), "should not already be young");
184 185 assert(hr->get_next_young_region() == NULL, "cause it should!");
185 186
186 187 hr->set_next_young_region(_head);
187 188 _head = hr;
188 189
189 190 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
190 191 ++_length;
191 192 }
192 193
193 194 void YoungList::add_survivor_region(HeapRegion* hr) {
194 195 assert(hr->is_survivor(), "should be flagged as survivor region");
195 196 assert(hr->get_next_young_region() == NULL, "cause it should!");
196 197
197 198 hr->set_next_young_region(_survivor_head);
198 199 if (_survivor_head == NULL) {
199 200 _survivor_tail = hr;
200 201 }
201 202 _survivor_head = hr;
202 203 ++_survivor_length;
203 204 }
204 205
205 206 void YoungList::empty_list(HeapRegion* list) {
206 207 while (list != NULL) {
207 208 HeapRegion* next = list->get_next_young_region();
208 209 list->set_next_young_region(NULL);
209 210 list->uninstall_surv_rate_group();
210 211 // This is called before a Full GC and all the non-empty /
211 212 // non-humongous regions at the end of the Full GC will end up as
212 213 // old anyway.
213 214 list->set_old();
214 215 list = next;
215 216 }
216 217 }
217 218
218 219 void YoungList::empty_list() {
219 220 assert(check_list_well_formed(), "young list should be well formed");
220 221
221 222 empty_list(_head);
222 223 _head = NULL;
223 224 _length = 0;
224 225
225 226 empty_list(_survivor_head);
226 227 _survivor_head = NULL;
227 228 _survivor_tail = NULL;
228 229 _survivor_length = 0;
229 230
230 231 _last_sampled_rs_lengths = 0;
231 232
232 233 assert(check_list_empty(false), "just making sure...");
233 234 }
234 235
235 236 bool YoungList::check_list_well_formed() {
236 237 bool ret = true;
237 238
238 239 uint length = 0;
239 240 HeapRegion* curr = _head;
240 241 HeapRegion* last = NULL;
241 242 while (curr != NULL) {
242 243 if (!curr->is_young()) {
243 244 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
244 245 "incorrectly tagged (y: %d, surv: %d)",
245 246 curr->bottom(), curr->end(),
246 247 curr->is_young(), curr->is_survivor());
247 248 ret = false;
248 249 }
249 250 ++length;
250 251 last = curr;
251 252 curr = curr->get_next_young_region();
252 253 }
253 254 ret = ret && (length == _length);
254 255
255 256 if (!ret) {
256 257 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
257 258 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
258 259 length, _length);
259 260 }
260 261
261 262 return ret;
262 263 }
263 264
264 265 bool YoungList::check_list_empty(bool check_sample) {
265 266 bool ret = true;
266 267
267 268 if (_length != 0) {
268 269 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
269 270 _length);
270 271 ret = false;
271 272 }
272 273 if (check_sample && _last_sampled_rs_lengths != 0) {
273 274 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
274 275 ret = false;
275 276 }
276 277 if (_head != NULL) {
277 278 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
278 279 ret = false;
279 280 }
280 281 if (!ret) {
281 282 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
282 283 }
283 284
284 285 return ret;
285 286 }
286 287
287 288 void
288 289 YoungList::rs_length_sampling_init() {
289 290 _sampled_rs_lengths = 0;
290 291 _curr = _head;
291 292 }
292 293
293 294 bool
294 295 YoungList::rs_length_sampling_more() {
295 296 return _curr != NULL;
296 297 }
297 298
298 299 void
299 300 YoungList::rs_length_sampling_next() {
300 301 assert( _curr != NULL, "invariant" );
301 302 size_t rs_length = _curr->rem_set()->occupied();
302 303
303 304 _sampled_rs_lengths += rs_length;
304 305
305 306 // The current region may not yet have been added to the
306 307 // incremental collection set (it gets added when it is
307 308 // retired as the current allocation region).
308 309 if (_curr->in_collection_set()) {
309 310 // Update the collection set policy information for this region
310 311 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
311 312 }
312 313
313 314 _curr = _curr->get_next_young_region();
314 315 if (_curr == NULL) {
315 316 _last_sampled_rs_lengths = _sampled_rs_lengths;
316 317 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
317 318 }
318 319 }
319 320
320 321 void
321 322 YoungList::reset_auxilary_lists() {
322 323 guarantee( is_empty(), "young list should be empty" );
323 324 assert(check_list_well_formed(), "young list should be well formed");
324 325
325 326 // Add survivor regions to SurvRateGroup.
326 327 _g1h->g1_policy()->note_start_adding_survivor_regions();
327 328 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
328 329
329 330 int young_index_in_cset = 0;
330 331 for (HeapRegion* curr = _survivor_head;
331 332 curr != NULL;
332 333 curr = curr->get_next_young_region()) {
333 334 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
334 335
335 336 // The region is a non-empty survivor so let's add it to
336 337 // the incremental collection set for the next evacuation
337 338 // pause.
338 339 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
339 340 young_index_in_cset += 1;
340 341 }
341 342 assert((uint) young_index_in_cset == _survivor_length, "post-condition");
342 343 _g1h->g1_policy()->note_stop_adding_survivor_regions();
343 344
344 345 _head = _survivor_head;
345 346 _length = _survivor_length;
346 347 if (_survivor_head != NULL) {
347 348 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 349 assert(_survivor_length > 0, "invariant");
349 350 _survivor_tail->set_next_young_region(NULL);
350 351 }
351 352
352 353 // Don't clear the survivor list handles until the start of
353 354 // the next evacuation pause - we need it in order to re-tag
354 355 // the survivor regions from this evacuation pause as 'young'
355 356 // at the start of the next.
356 357
357 358 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
358 359
359 360 assert(check_list_well_formed(), "young list should be well formed");
360 361 }
361 362
362 363 void YoungList::print() {
363 364 HeapRegion* lists[] = {_head, _survivor_head};
364 365 const char* names[] = {"YOUNG", "SURVIVOR"};
365 366
366 367 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 368 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 369 HeapRegion *curr = lists[list];
369 370 if (curr == NULL)
370 371 gclog_or_tty->print_cr(" empty");
371 372 while (curr != NULL) {
372 373 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT ", N: "PTR_FORMAT", age: %4d",
373 374 HR_FORMAT_PARAMS(curr),
374 375 curr->prev_top_at_mark_start(),
375 376 curr->next_top_at_mark_start(),
376 377 curr->age_in_surv_rate_group_cond());
377 378 curr = curr->get_next_young_region();
378 379 }
379 380 }
380 381
381 382 gclog_or_tty->cr();
382 383 }
383 384
384 385 void G1RegionMappingChangedListener::reset_from_card_cache(uint start_idx, size_t num_regions) {
385 386 OtherRegionsTable::invalidate(start_idx, num_regions);
386 387 }
387 388
388 389 void G1RegionMappingChangedListener::on_commit(uint start_idx, size_t num_regions, bool zero_filled) {
389 390 // The from card cache is not the memory that is actually committed. So we cannot
390 391 // take advantage of the zero_filled parameter.
391 392 reset_from_card_cache(start_idx, num_regions);
392 393 }
393 394
394 395 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
395 396 {
396 397 // Claim the right to put the region on the dirty cards region list
397 398 // by installing a self pointer.
398 399 HeapRegion* next = hr->get_next_dirty_cards_region();
399 400 if (next == NULL) {
400 401 HeapRegion* res = (HeapRegion*)
401 402 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
402 403 NULL);
403 404 if (res == NULL) {
404 405 HeapRegion* head;
405 406 do {
406 407 // Put the region to the dirty cards region list.
407 408 head = _dirty_cards_region_list;
408 409 next = (HeapRegion*)
409 410 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
410 411 if (next == head) {
411 412 assert(hr->get_next_dirty_cards_region() == hr,
412 413 "hr->get_next_dirty_cards_region() != hr");
413 414 if (next == NULL) {
414 415 // The last region in the list points to itself.
415 416 hr->set_next_dirty_cards_region(hr);
416 417 } else {
417 418 hr->set_next_dirty_cards_region(next);
418 419 }
419 420 }
420 421 } while (next != head);
421 422 }
422 423 }
423 424 }
424 425
425 426 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
426 427 {
427 428 HeapRegion* head;
428 429 HeapRegion* hr;
429 430 do {
430 431 head = _dirty_cards_region_list;
431 432 if (head == NULL) {
432 433 return NULL;
433 434 }
434 435 HeapRegion* new_head = head->get_next_dirty_cards_region();
435 436 if (head == new_head) {
436 437 // The last region.
437 438 new_head = NULL;
438 439 }
439 440 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
440 441 head);
441 442 } while (hr != head);
442 443 assert(hr != NULL, "invariant");
443 444 hr->set_next_dirty_cards_region(NULL);
444 445 return hr;
445 446 }
446 447
447 448 #ifdef ASSERT
448 449 // A region is added to the collection set as it is retired
449 450 // so an address p can point to a region which will be in the
450 451 // collection set but has not yet been retired. This method
451 452 // therefore is only accurate during a GC pause after all
452 453 // regions have been retired. It is used for debugging
453 454 // to check if an nmethod has references to objects that can
454 455 // be move during a partial collection. Though it can be
455 456 // inaccurate, it is sufficient for G1 because the conservative
456 457 // implementation of is_scavengable() for G1 will indicate that
457 458 // all nmethods must be scanned during a partial collection.
458 459 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
459 460 if (p == NULL) {
460 461 return false;
461 462 }
462 463 return heap_region_containing(p)->in_collection_set();
463 464 }
464 465 #endif
465 466
466 467 // Returns true if the reference points to an object that
467 468 // can move in an incremental collection.
468 469 bool G1CollectedHeap::is_scavengable(const void* p) {
469 470 HeapRegion* hr = heap_region_containing(p);
470 471 return !hr->isHumongous();
471 472 }
472 473
473 474 void G1CollectedHeap::check_ct_logs_at_safepoint() {
474 475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
475 476 CardTableModRefBS* ct_bs = g1_barrier_set();
476 477
477 478 // Count the dirty cards at the start.
478 479 CountNonCleanMemRegionClosure count1(this);
479 480 ct_bs->mod_card_iterate(&count1);
480 481 int orig_count = count1.n();
481 482
482 483 // First clear the logged cards.
483 484 ClearLoggedCardTableEntryClosure clear;
484 485 dcqs.apply_closure_to_all_completed_buffers(&clear);
485 486 dcqs.iterate_closure_all_threads(&clear, false);
486 487 clear.print_histo();
487 488
488 489 // Now ensure that there's no dirty cards.
489 490 CountNonCleanMemRegionClosure count2(this);
490 491 ct_bs->mod_card_iterate(&count2);
491 492 if (count2.n() != 0) {
492 493 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
493 494 count2.n(), orig_count);
494 495 }
495 496 guarantee(count2.n() == 0, "Card table should be clean.");
496 497
497 498 RedirtyLoggedCardTableEntryClosure redirty;
498 499 dcqs.apply_closure_to_all_completed_buffers(&redirty);
499 500 dcqs.iterate_closure_all_threads(&redirty, false);
500 501 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
501 502 clear.num_processed(), orig_count);
502 503 guarantee(redirty.num_processed() == clear.num_processed(),
503 504 err_msg("Redirtied "SIZE_FORMAT" cards, bug cleared "SIZE_FORMAT,
504 505 redirty.num_processed(), clear.num_processed()));
505 506
506 507 CountNonCleanMemRegionClosure count3(this);
507 508 ct_bs->mod_card_iterate(&count3);
508 509 if (count3.n() != orig_count) {
509 510 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
510 511 orig_count, count3.n());
511 512 guarantee(count3.n() >= orig_count, "Should have restored them all.");
512 513 }
513 514 }
514 515
515 516 // Private class members.
516 517
517 518 G1CollectedHeap* G1CollectedHeap::_g1h;
518 519
519 520 // Private methods.
520 521
521 522 HeapRegion*
522 523 G1CollectedHeap::new_region_try_secondary_free_list(bool is_old) {
523 524 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
524 525 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
525 526 if (!_secondary_free_list.is_empty()) {
526 527 if (G1ConcRegionFreeingVerbose) {
527 528 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
528 529 "secondary_free_list has %u entries",
529 530 _secondary_free_list.length());
530 531 }
531 532 // It looks as if there are free regions available on the
532 533 // secondary_free_list. Let's move them to the free_list and try
533 534 // again to allocate from it.
534 535 append_secondary_free_list();
535 536
536 537 assert(_hrm.num_free_regions() > 0, "if the secondary_free_list was not "
537 538 "empty we should have moved at least one entry to the free_list");
538 539 HeapRegion* res = _hrm.allocate_free_region(is_old);
539 540 if (G1ConcRegionFreeingVerbose) {
540 541 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
541 542 "allocated "HR_FORMAT" from secondary_free_list",
542 543 HR_FORMAT_PARAMS(res));
543 544 }
544 545 return res;
545 546 }
546 547
547 548 // Wait here until we get notified either when (a) there are no
548 549 // more free regions coming or (b) some regions have been moved on
549 550 // the secondary_free_list.
550 551 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
551 552 }
552 553
553 554 if (G1ConcRegionFreeingVerbose) {
554 555 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
555 556 "could not allocate from secondary_free_list");
556 557 }
557 558 return NULL;
558 559 }
559 560
560 561 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool is_old, bool do_expand) {
561 562 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
562 563 "the only time we use this to allocate a humongous region is "
563 564 "when we are allocating a single humongous region");
564 565
565 566 HeapRegion* res;
566 567 if (G1StressConcRegionFreeing) {
567 568 if (!_secondary_free_list.is_empty()) {
568 569 if (G1ConcRegionFreeingVerbose) {
569 570 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
570 571 "forced to look at the secondary_free_list");
571 572 }
572 573 res = new_region_try_secondary_free_list(is_old);
573 574 if (res != NULL) {
574 575 return res;
575 576 }
576 577 }
577 578 }
578 579
579 580 res = _hrm.allocate_free_region(is_old);
580 581
581 582 if (res == NULL) {
582 583 if (G1ConcRegionFreeingVerbose) {
583 584 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
584 585 "res == NULL, trying the secondary_free_list");
585 586 }
586 587 res = new_region_try_secondary_free_list(is_old);
587 588 }
588 589 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
589 590 // Currently, only attempts to allocate GC alloc regions set
590 591 // do_expand to true. So, we should only reach here during a
591 592 // safepoint. If this assumption changes we might have to
592 593 // reconsider the use of _expand_heap_after_alloc_failure.
593 594 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
594 595
595 596 ergo_verbose1(ErgoHeapSizing,
596 597 "attempt heap expansion",
597 598 ergo_format_reason("region allocation request failed")
598 599 ergo_format_byte("allocation request"),
599 600 word_size * HeapWordSize);
600 601 if (expand(word_size * HeapWordSize)) {
601 602 // Given that expand() succeeded in expanding the heap, and we
602 603 // always expand the heap by an amount aligned to the heap
603 604 // region size, the free list should in theory not be empty.
604 605 // In either case allocate_free_region() will check for NULL.
605 606 res = _hrm.allocate_free_region(is_old);
606 607 } else {
607 608 _expand_heap_after_alloc_failure = false;
608 609 }
609 610 }
610 611 return res;
611 612 }
612 613
613 614 HeapWord*
614 615 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
615 616 uint num_regions,
616 617 size_t word_size,
617 618 AllocationContext_t context) {
618 619 assert(first != G1_NO_HRM_INDEX, "pre-condition");
619 620 assert(isHumongous(word_size), "word_size should be humongous");
620 621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
621 622
622 623 // Index of last region in the series + 1.
623 624 uint last = first + num_regions;
624 625
625 626 // We need to initialize the region(s) we just discovered. This is
626 627 // a bit tricky given that it can happen concurrently with
627 628 // refinement threads refining cards on these regions and
628 629 // potentially wanting to refine the BOT as they are scanning
629 630 // those cards (this can happen shortly after a cleanup; see CR
630 631 // 6991377). So we have to set up the region(s) carefully and in
631 632 // a specific order.
632 633
633 634 // The word size sum of all the regions we will allocate.
634 635 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
635 636 assert(word_size <= word_size_sum, "sanity");
636 637
637 638 // This will be the "starts humongous" region.
638 639 HeapRegion* first_hr = region_at(first);
639 640 // The header of the new object will be placed at the bottom of
640 641 // the first region.
641 642 HeapWord* new_obj = first_hr->bottom();
642 643 // This will be the new end of the first region in the series that
643 644 // should also match the end of the last region in the series.
644 645 HeapWord* new_end = new_obj + word_size_sum;
645 646 // This will be the new top of the first region that will reflect
646 647 // this allocation.
647 648 HeapWord* new_top = new_obj + word_size;
648 649
649 650 // First, we need to zero the header of the space that we will be
650 651 // allocating. When we update top further down, some refinement
651 652 // threads might try to scan the region. By zeroing the header we
652 653 // ensure that any thread that will try to scan the region will
653 654 // come across the zero klass word and bail out.
654 655 //
655 656 // NOTE: It would not have been correct to have used
656 657 // CollectedHeap::fill_with_object() and make the space look like
657 658 // an int array. The thread that is doing the allocation will
658 659 // later update the object header to a potentially different array
659 660 // type and, for a very short period of time, the klass and length
660 661 // fields will be inconsistent. This could cause a refinement
661 662 // thread to calculate the object size incorrectly.
662 663 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
663 664
664 665 // We will set up the first region as "starts humongous". This
665 666 // will also update the BOT covering all the regions to reflect
666 667 // that there is a single object that starts at the bottom of the
667 668 // first region.
668 669 first_hr->set_startsHumongous(new_top, new_end);
669 670 first_hr->set_allocation_context(context);
670 671 // Then, if there are any, we will set up the "continues
671 672 // humongous" regions.
672 673 HeapRegion* hr = NULL;
673 674 for (uint i = first + 1; i < last; ++i) {
674 675 hr = region_at(i);
675 676 hr->set_continuesHumongous(first_hr);
676 677 hr->set_allocation_context(context);
677 678 }
678 679 // If we have "continues humongous" regions (hr != NULL), then the
679 680 // end of the last one should match new_end.
680 681 assert(hr == NULL || hr->end() == new_end, "sanity");
681 682
682 683 // Up to this point no concurrent thread would have been able to
683 684 // do any scanning on any region in this series. All the top
684 685 // fields still point to bottom, so the intersection between
685 686 // [bottom,top] and [card_start,card_end] will be empty. Before we
686 687 // update the top fields, we'll do a storestore to make sure that
687 688 // no thread sees the update to top before the zeroing of the
688 689 // object header and the BOT initialization.
689 690 OrderAccess::storestore();
690 691
691 692 // Now that the BOT and the object header have been initialized,
692 693 // we can update top of the "starts humongous" region.
693 694 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
694 695 "new_top should be in this region");
695 696 first_hr->set_top(new_top);
696 697 if (_hr_printer.is_active()) {
697 698 HeapWord* bottom = first_hr->bottom();
698 699 HeapWord* end = first_hr->orig_end();
699 700 if ((first + 1) == last) {
700 701 // the series has a single humongous region
701 702 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
702 703 } else {
703 704 // the series has more than one humongous regions
704 705 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
705 706 }
706 707 }
707 708
708 709 // Now, we will update the top fields of the "continues humongous"
709 710 // regions. The reason we need to do this is that, otherwise,
710 711 // these regions would look empty and this will confuse parts of
711 712 // G1. For example, the code that looks for a consecutive number
712 713 // of empty regions will consider them empty and try to
713 714 // re-allocate them. We can extend is_empty() to also include
714 715 // !continuesHumongous(), but it is easier to just update the top
715 716 // fields here. The way we set top for all regions (i.e., top ==
716 717 // end for all regions but the last one, top == new_top for the
717 718 // last one) is actually used when we will free up the humongous
718 719 // region in free_humongous_region().
719 720 hr = NULL;
720 721 for (uint i = first + 1; i < last; ++i) {
721 722 hr = region_at(i);
722 723 if ((i + 1) == last) {
723 724 // last continues humongous region
724 725 assert(hr->bottom() < new_top && new_top <= hr->end(),
725 726 "new_top should fall on this region");
726 727 hr->set_top(new_top);
727 728 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
728 729 } else {
729 730 // not last one
730 731 assert(new_top > hr->end(), "new_top should be above this region");
731 732 hr->set_top(hr->end());
732 733 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
733 734 }
734 735 }
735 736 // If we have continues humongous regions (hr != NULL), then the
736 737 // end of the last one should match new_end and its top should
737 738 // match new_top.
738 739 assert(hr == NULL ||
739 740 (hr->end() == new_end && hr->top() == new_top), "sanity");
740 741 check_bitmaps("Humongous Region Allocation", first_hr);
741 742
742 743 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
743 744 _allocator->increase_used(first_hr->used());
744 745 _humongous_set.add(first_hr);
745 746
746 747 return new_obj;
747 748 }
748 749
749 750 // If could fit into free regions w/o expansion, try.
750 751 // Otherwise, if can expand, do so.
751 752 // Otherwise, if using ex regions might help, try with ex given back.
752 753 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size, AllocationContext_t context) {
753 754 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
754 755
755 756 verify_region_sets_optional();
756 757
757 758 uint first = G1_NO_HRM_INDEX;
758 759 uint obj_regions = (uint)(align_size_up_(word_size, HeapRegion::GrainWords) / HeapRegion::GrainWords);
759 760
760 761 if (obj_regions == 1) {
761 762 // Only one region to allocate, try to use a fast path by directly allocating
762 763 // from the free lists. Do not try to expand here, we will potentially do that
763 764 // later.
764 765 HeapRegion* hr = new_region(word_size, true /* is_old */, false /* do_expand */);
765 766 if (hr != NULL) {
766 767 first = hr->hrm_index();
767 768 }
768 769 } else {
769 770 // We can't allocate humongous regions spanning more than one region while
770 771 // cleanupComplete() is running, since some of the regions we find to be
771 772 // empty might not yet be added to the free list. It is not straightforward
772 773 // to know in which list they are on so that we can remove them. We only
773 774 // need to do this if we need to allocate more than one region to satisfy the
774 775 // current humongous allocation request. If we are only allocating one region
775 776 // we use the one-region region allocation code (see above), that already
776 777 // potentially waits for regions from the secondary free list.
777 778 wait_while_free_regions_coming();
778 779 append_secondary_free_list_if_not_empty_with_lock();
779 780
780 781 // Policy: Try only empty regions (i.e. already committed first). Maybe we
781 782 // are lucky enough to find some.
782 783 first = _hrm.find_contiguous_only_empty(obj_regions);
783 784 if (first != G1_NO_HRM_INDEX) {
784 785 _hrm.allocate_free_regions_starting_at(first, obj_regions);
785 786 }
786 787 }
787 788
788 789 if (first == G1_NO_HRM_INDEX) {
789 790 // Policy: We could not find enough regions for the humongous object in the
790 791 // free list. Look through the heap to find a mix of free and uncommitted regions.
791 792 // If so, try expansion.
792 793 first = _hrm.find_contiguous_empty_or_unavailable(obj_regions);
793 794 if (first != G1_NO_HRM_INDEX) {
794 795 // We found something. Make sure these regions are committed, i.e. expand
795 796 // the heap. Alternatively we could do a defragmentation GC.
796 797 ergo_verbose1(ErgoHeapSizing,
797 798 "attempt heap expansion",
798 799 ergo_format_reason("humongous allocation request failed")
799 800 ergo_format_byte("allocation request"),
800 801 word_size * HeapWordSize);
801 802
802 803 _hrm.expand_at(first, obj_regions);
803 804 g1_policy()->record_new_heap_size(num_regions());
804 805
805 806 #ifdef ASSERT
806 807 for (uint i = first; i < first + obj_regions; ++i) {
807 808 HeapRegion* hr = region_at(i);
808 809 assert(hr->is_free(), "sanity");
809 810 assert(hr->is_empty(), "sanity");
810 811 assert(is_on_master_free_list(hr), "sanity");
811 812 }
812 813 #endif
813 814 _hrm.allocate_free_regions_starting_at(first, obj_regions);
814 815 } else {
815 816 // Policy: Potentially trigger a defragmentation GC.
816 817 }
817 818 }
818 819
819 820 HeapWord* result = NULL;
820 821 if (first != G1_NO_HRM_INDEX) {
821 822 result = humongous_obj_allocate_initialize_regions(first, obj_regions,
822 823 word_size, context);
823 824 assert(result != NULL, "it should always return a valid result");
824 825
825 826 // A successful humongous object allocation changes the used space
826 827 // information of the old generation so we need to recalculate the
827 828 // sizes and update the jstat counters here.
828 829 g1mm()->update_sizes();
829 830 }
830 831
831 832 verify_region_sets_optional();
832 833
833 834 return result;
834 835 }
835 836
836 837 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
837 838 assert_heap_not_locked_and_not_at_safepoint();
838 839 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
839 840
840 841 unsigned int dummy_gc_count_before;
841 842 int dummy_gclocker_retry_count = 0;
842 843 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count);
843 844 }
844 845
845 846 HeapWord*
846 847 G1CollectedHeap::mem_allocate(size_t word_size,
847 848 bool* gc_overhead_limit_was_exceeded) {
848 849 assert_heap_not_locked_and_not_at_safepoint();
849 850
850 851 // Loop until the allocation is satisfied, or unsatisfied after GC.
851 852 for (int try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) {
852 853 unsigned int gc_count_before;
853 854
854 855 HeapWord* result = NULL;
855 856 if (!isHumongous(word_size)) {
856 857 result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count);
857 858 } else {
858 859 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count);
859 860 }
860 861 if (result != NULL) {
861 862 return result;
862 863 }
863 864
864 865 // Create the garbage collection operation...
865 866 VM_G1CollectForAllocation op(gc_count_before, word_size);
866 867 op.set_allocation_context(AllocationContext::current());
867 868
868 869 // ...and get the VM thread to execute it.
869 870 VMThread::execute(&op);
870 871
871 872 if (op.prologue_succeeded() && op.pause_succeeded()) {
872 873 // If the operation was successful we'll return the result even
873 874 // if it is NULL. If the allocation attempt failed immediately
874 875 // after a Full GC, it's unlikely we'll be able to allocate now.
875 876 HeapWord* result = op.result();
876 877 if (result != NULL && !isHumongous(word_size)) {
877 878 // Allocations that take place on VM operations do not do any
878 879 // card dirtying and we have to do it here. We only have to do
879 880 // this for non-humongous allocations, though.
880 881 dirty_young_block(result, word_size);
881 882 }
882 883 return result;
883 884 } else {
884 885 if (gclocker_retry_count > GCLockerRetryAllocationCount) {
885 886 return NULL;
886 887 }
887 888 assert(op.result() == NULL,
888 889 "the result should be NULL if the VM op did not succeed");
889 890 }
890 891
891 892 // Give a warning if we seem to be looping forever.
892 893 if ((QueuedAllocationWarningCount > 0) &&
893 894 (try_count % QueuedAllocationWarningCount == 0)) {
894 895 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
895 896 }
896 897 }
897 898
898 899 ShouldNotReachHere();
899 900 return NULL;
900 901 }
901 902
902 903 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
903 904 AllocationContext_t context,
904 905 unsigned int *gc_count_before_ret,
905 906 int* gclocker_retry_count_ret) {
906 907 // Make sure you read the note in attempt_allocation_humongous().
907 908
908 909 assert_heap_not_locked_and_not_at_safepoint();
909 910 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
910 911 "be called for humongous allocation requests");
911 912
912 913 // We should only get here after the first-level allocation attempt
913 914 // (attempt_allocation()) failed to allocate.
914 915
915 916 // We will loop until a) we manage to successfully perform the
916 917 // allocation or b) we successfully schedule a collection which
917 918 // fails to perform the allocation. b) is the only case when we'll
918 919 // return NULL.
919 920 HeapWord* result = NULL;
920 921 for (int try_count = 1; /* we'll return */; try_count += 1) {
921 922 bool should_try_gc;
922 923 unsigned int gc_count_before;
923 924
924 925 {
925 926 MutexLockerEx x(Heap_lock);
926 927 result = _allocator->mutator_alloc_region(context)->attempt_allocation_locked(word_size,
927 928 false /* bot_updates */);
928 929 if (result != NULL) {
929 930 return result;
930 931 }
931 932
932 933 // If we reach here, attempt_allocation_locked() above failed to
933 934 // allocate a new region. So the mutator alloc region should be NULL.
934 935 assert(_allocator->mutator_alloc_region(context)->get() == NULL, "only way to get here");
935 936
936 937 if (GC_locker::is_active_and_needs_gc()) {
937 938 if (g1_policy()->can_expand_young_list()) {
938 939 // No need for an ergo verbose message here,
939 940 // can_expand_young_list() does this when it returns true.
940 941 result = _allocator->mutator_alloc_region(context)->attempt_allocation_force(word_size,
941 942 false /* bot_updates */);
942 943 if (result != NULL) {
943 944 return result;
944 945 }
945 946 }
946 947 should_try_gc = false;
947 948 } else {
948 949 // The GCLocker may not be active but the GCLocker initiated
949 950 // GC may not yet have been performed (GCLocker::needs_gc()
950 951 // returns true). In this case we do not try this GC and
951 952 // wait until the GCLocker initiated GC is performed, and
952 953 // then retry the allocation.
953 954 if (GC_locker::needs_gc()) {
954 955 should_try_gc = false;
955 956 } else {
956 957 // Read the GC count while still holding the Heap_lock.
957 958 gc_count_before = total_collections();
958 959 should_try_gc = true;
959 960 }
960 961 }
961 962 }
962 963
963 964 if (should_try_gc) {
964 965 bool succeeded;
965 966 result = do_collection_pause(word_size, gc_count_before, &succeeded,
966 967 GCCause::_g1_inc_collection_pause);
967 968 if (result != NULL) {
968 969 assert(succeeded, "only way to get back a non-NULL result");
969 970 return result;
970 971 }
971 972
972 973 if (succeeded) {
973 974 // If we get here we successfully scheduled a collection which
974 975 // failed to allocate. No point in trying to allocate
975 976 // further. We'll just return NULL.
976 977 MutexLockerEx x(Heap_lock);
977 978 *gc_count_before_ret = total_collections();
978 979 return NULL;
979 980 }
980 981 } else {
981 982 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
982 983 MutexLockerEx x(Heap_lock);
983 984 *gc_count_before_ret = total_collections();
984 985 return NULL;
985 986 }
986 987 // The GCLocker is either active or the GCLocker initiated
987 988 // GC has not yet been performed. Stall until it is and
988 989 // then retry the allocation.
989 990 GC_locker::stall_until_clear();
990 991 (*gclocker_retry_count_ret) += 1;
991 992 }
992 993
993 994 // We can reach here if we were unsuccessful in scheduling a
994 995 // collection (because another thread beat us to it) or if we were
995 996 // stalled due to the GC locker. In either can we should retry the
996 997 // allocation attempt in case another thread successfully
997 998 // performed a collection and reclaimed enough space. We do the
998 999 // first attempt (without holding the Heap_lock) here and the
999 1000 // follow-on attempt will be at the start of the next loop
1000 1001 // iteration (after taking the Heap_lock).
1001 1002 result = _allocator->mutator_alloc_region(context)->attempt_allocation(word_size,
1002 1003 false /* bot_updates */);
1003 1004 if (result != NULL) {
1004 1005 return result;
1005 1006 }
1006 1007
1007 1008 // Give a warning if we seem to be looping forever.
1008 1009 if ((QueuedAllocationWarningCount > 0) &&
1009 1010 (try_count % QueuedAllocationWarningCount == 0)) {
1010 1011 warning("G1CollectedHeap::attempt_allocation_slow() "
1011 1012 "retries %d times", try_count);
1012 1013 }
1013 1014 }
1014 1015
1015 1016 ShouldNotReachHere();
1016 1017 return NULL;
1017 1018 }
1018 1019
1019 1020 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1020 1021 unsigned int * gc_count_before_ret,
1021 1022 int* gclocker_retry_count_ret) {
1022 1023 // The structure of this method has a lot of similarities to
1023 1024 // attempt_allocation_slow(). The reason these two were not merged
1024 1025 // into a single one is that such a method would require several "if
1025 1026 // allocation is not humongous do this, otherwise do that"
1026 1027 // conditional paths which would obscure its flow. In fact, an early
1027 1028 // version of this code did use a unified method which was harder to
1028 1029 // follow and, as a result, it had subtle bugs that were hard to
1029 1030 // track down. So keeping these two methods separate allows each to
1030 1031 // be more readable. It will be good to keep these two in sync as
1031 1032 // much as possible.
1032 1033
1033 1034 assert_heap_not_locked_and_not_at_safepoint();
1034 1035 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1035 1036 "should only be called for humongous allocations");
1036 1037
1037 1038 // Humongous objects can exhaust the heap quickly, so we should check if we
1038 1039 // need to start a marking cycle at each humongous object allocation. We do
1039 1040 // the check before we do the actual allocation. The reason for doing it
1040 1041 // before the allocation is that we avoid having to keep track of the newly
1041 1042 // allocated memory while we do a GC.
1042 1043 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1043 1044 word_size)) {
1044 1045 collect(GCCause::_g1_humongous_allocation);
1045 1046 }
1046 1047
1047 1048 // We will loop until a) we manage to successfully perform the
1048 1049 // allocation or b) we successfully schedule a collection which
1049 1050 // fails to perform the allocation. b) is the only case when we'll
1050 1051 // return NULL.
1051 1052 HeapWord* result = NULL;
1052 1053 for (int try_count = 1; /* we'll return */; try_count += 1) {
1053 1054 bool should_try_gc;
1054 1055 unsigned int gc_count_before;
1055 1056
1056 1057 {
1057 1058 MutexLockerEx x(Heap_lock);
1058 1059
1059 1060 // Given that humongous objects are not allocated in young
1060 1061 // regions, we'll first try to do the allocation without doing a
1061 1062 // collection hoping that there's enough space in the heap.
1062 1063 result = humongous_obj_allocate(word_size, AllocationContext::current());
1063 1064 if (result != NULL) {
1064 1065 return result;
1065 1066 }
1066 1067
1067 1068 if (GC_locker::is_active_and_needs_gc()) {
1068 1069 should_try_gc = false;
1069 1070 } else {
1070 1071 // The GCLocker may not be active but the GCLocker initiated
1071 1072 // GC may not yet have been performed (GCLocker::needs_gc()
1072 1073 // returns true). In this case we do not try this GC and
1073 1074 // wait until the GCLocker initiated GC is performed, and
1074 1075 // then retry the allocation.
1075 1076 if (GC_locker::needs_gc()) {
1076 1077 should_try_gc = false;
1077 1078 } else {
1078 1079 // Read the GC count while still holding the Heap_lock.
1079 1080 gc_count_before = total_collections();
1080 1081 should_try_gc = true;
1081 1082 }
1082 1083 }
1083 1084 }
1084 1085
1085 1086 if (should_try_gc) {
1086 1087 // If we failed to allocate the humongous object, we should try to
1087 1088 // do a collection pause (if we're allowed) in case it reclaims
1088 1089 // enough space for the allocation to succeed after the pause.
1089 1090
1090 1091 bool succeeded;
1091 1092 result = do_collection_pause(word_size, gc_count_before, &succeeded,
1092 1093 GCCause::_g1_humongous_allocation);
1093 1094 if (result != NULL) {
1094 1095 assert(succeeded, "only way to get back a non-NULL result");
1095 1096 return result;
1096 1097 }
1097 1098
1098 1099 if (succeeded) {
1099 1100 // If we get here we successfully scheduled a collection which
1100 1101 // failed to allocate. No point in trying to allocate
1101 1102 // further. We'll just return NULL.
1102 1103 MutexLockerEx x(Heap_lock);
1103 1104 *gc_count_before_ret = total_collections();
1104 1105 return NULL;
1105 1106 }
1106 1107 } else {
1107 1108 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
1108 1109 MutexLockerEx x(Heap_lock);
1109 1110 *gc_count_before_ret = total_collections();
1110 1111 return NULL;
1111 1112 }
1112 1113 // The GCLocker is either active or the GCLocker initiated
1113 1114 // GC has not yet been performed. Stall until it is and
1114 1115 // then retry the allocation.
1115 1116 GC_locker::stall_until_clear();
1116 1117 (*gclocker_retry_count_ret) += 1;
1117 1118 }
1118 1119
1119 1120 // We can reach here if we were unsuccessful in scheduling a
1120 1121 // collection (because another thread beat us to it) or if we were
1121 1122 // stalled due to the GC locker. In either can we should retry the
1122 1123 // allocation attempt in case another thread successfully
1123 1124 // performed a collection and reclaimed enough space. Give a
1124 1125 // warning if we seem to be looping forever.
1125 1126
1126 1127 if ((QueuedAllocationWarningCount > 0) &&
1127 1128 (try_count % QueuedAllocationWarningCount == 0)) {
1128 1129 warning("G1CollectedHeap::attempt_allocation_humongous() "
1129 1130 "retries %d times", try_count);
1130 1131 }
1131 1132 }
1132 1133
1133 1134 ShouldNotReachHere();
1134 1135 return NULL;
1135 1136 }
1136 1137
1137 1138 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1138 1139 AllocationContext_t context,
1139 1140 bool expect_null_mutator_alloc_region) {
1140 1141 assert_at_safepoint(true /* should_be_vm_thread */);
1141 1142 assert(_allocator->mutator_alloc_region(context)->get() == NULL ||
1142 1143 !expect_null_mutator_alloc_region,
1143 1144 "the current alloc region was unexpectedly found to be non-NULL");
1144 1145
1145 1146 if (!isHumongous(word_size)) {
1146 1147 return _allocator->mutator_alloc_region(context)->attempt_allocation_locked(word_size,
1147 1148 false /* bot_updates */);
1148 1149 } else {
1149 1150 HeapWord* result = humongous_obj_allocate(word_size, context);
1150 1151 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1151 1152 g1_policy()->set_initiate_conc_mark_if_possible();
1152 1153 }
1153 1154 return result;
1154 1155 }
1155 1156
1156 1157 ShouldNotReachHere();
1157 1158 }
1158 1159
1159 1160 class PostMCRemSetClearClosure: public HeapRegionClosure {
1160 1161 G1CollectedHeap* _g1h;
1161 1162 ModRefBarrierSet* _mr_bs;
1162 1163 public:
1163 1164 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1164 1165 _g1h(g1h), _mr_bs(mr_bs) {}
1165 1166
1166 1167 bool doHeapRegion(HeapRegion* r) {
1167 1168 HeapRegionRemSet* hrrs = r->rem_set();
1168 1169
1169 1170 if (r->continuesHumongous()) {
1170 1171 // We'll assert that the strong code root list and RSet is empty
1171 1172 assert(hrrs->strong_code_roots_list_length() == 0, "sanity");
1172 1173 assert(hrrs->occupied() == 0, "RSet should be empty");
1173 1174 return false;
1174 1175 }
1175 1176
1176 1177 _g1h->reset_gc_time_stamps(r);
1177 1178 hrrs->clear();
1178 1179 // You might think here that we could clear just the cards
1179 1180 // corresponding to the used region. But no: if we leave a dirty card
1180 1181 // in a region we might allocate into, then it would prevent that card
1181 1182 // from being enqueued, and cause it to be missed.
1182 1183 // Re: the performance cost: we shouldn't be doing full GC anyway!
1183 1184 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1184 1185
1185 1186 return false;
1186 1187 }
1187 1188 };
1188 1189
1189 1190 void G1CollectedHeap::clear_rsets_post_compaction() {
1190 1191 PostMCRemSetClearClosure rs_clear(this, g1_barrier_set());
1191 1192 heap_region_iterate(&rs_clear);
1192 1193 }
1193 1194
1194 1195 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1195 1196 G1CollectedHeap* _g1h;
1196 1197 UpdateRSOopClosure _cl;
1197 1198 int _worker_i;
1198 1199 public:
1199 1200 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1200 1201 _cl(g1->g1_rem_set(), worker_i),
1201 1202 _worker_i(worker_i),
1202 1203 _g1h(g1)
1203 1204 { }
1204 1205
1205 1206 bool doHeapRegion(HeapRegion* r) {
1206 1207 if (!r->continuesHumongous()) {
1207 1208 _cl.set_from(r);
1208 1209 r->oop_iterate(&_cl);
1209 1210 }
1210 1211 return false;
1211 1212 }
1212 1213 };
1213 1214
1214 1215 class ParRebuildRSTask: public AbstractGangTask {
1215 1216 G1CollectedHeap* _g1;
1216 1217 public:
1217 1218 ParRebuildRSTask(G1CollectedHeap* g1)
1218 1219 : AbstractGangTask("ParRebuildRSTask"),
1219 1220 _g1(g1)
1220 1221 { }
1221 1222
1222 1223 void work(uint worker_id) {
1223 1224 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1224 1225 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1225 1226 _g1->workers()->active_workers(),
1226 1227 HeapRegion::RebuildRSClaimValue);
1227 1228 }
1228 1229 };
1229 1230
1230 1231 class PostCompactionPrinterClosure: public HeapRegionClosure {
1231 1232 private:
1232 1233 G1HRPrinter* _hr_printer;
1233 1234 public:
1234 1235 bool doHeapRegion(HeapRegion* hr) {
1235 1236 assert(!hr->is_young(), "not expecting to find young regions");
1236 1237 if (hr->is_free()) {
1237 1238 // We only generate output for non-empty regions.
1238 1239 } else if (hr->startsHumongous()) {
1239 1240 if (hr->region_num() == 1) {
1240 1241 // single humongous region
1241 1242 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1242 1243 } else {
1243 1244 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1244 1245 }
1245 1246 } else if (hr->continuesHumongous()) {
1246 1247 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1247 1248 } else if (hr->is_old()) {
1248 1249 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1249 1250 } else {
1250 1251 ShouldNotReachHere();
1251 1252 }
1252 1253 return false;
1253 1254 }
1254 1255
1255 1256 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1256 1257 : _hr_printer(hr_printer) { }
1257 1258 };
1258 1259
1259 1260 void G1CollectedHeap::print_hrm_post_compaction() {
1260 1261 PostCompactionPrinterClosure cl(hr_printer());
1261 1262 heap_region_iterate(&cl);
1262 1263 }
1263 1264
1264 1265 bool G1CollectedHeap::do_collection(bool explicit_gc,
1265 1266 bool clear_all_soft_refs,
1266 1267 size_t word_size) {
1267 1268 assert_at_safepoint(true /* should_be_vm_thread */);
1268 1269
1269 1270 if (GC_locker::check_active_before_gc()) {
1270 1271 return false;
1271 1272 }
1272 1273
1273 1274 STWGCTimer* gc_timer = G1MarkSweep::gc_timer();
1274 1275 gc_timer->register_gc_start();
1275 1276
1276 1277 SerialOldTracer* gc_tracer = G1MarkSweep::gc_tracer();
1277 1278 gc_tracer->report_gc_start(gc_cause(), gc_timer->gc_start());
1278 1279
1279 1280 SvcGCMarker sgcm(SvcGCMarker::FULL);
1280 1281 ResourceMark rm;
1281 1282
1282 1283 print_heap_before_gc();
1283 1284 trace_heap_before_gc(gc_tracer);
1284 1285
1285 1286 size_t metadata_prev_used = MetaspaceAux::used_bytes();
1286 1287
1287 1288 verify_region_sets_optional();
1288 1289
1289 1290 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1290 1291 collector_policy()->should_clear_all_soft_refs();
1291 1292
1292 1293 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1293 1294
1294 1295 {
1295 1296 IsGCActiveMark x;
1296 1297
1297 1298 // Timing
1298 1299 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1299 1300 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1300 1301 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1301 1302
1302 1303 {
1303 1304 GCTraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, NULL, gc_tracer->gc_id());
1304 1305 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1305 1306 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1306 1307
1307 1308 double start = os::elapsedTime();
1308 1309 g1_policy()->record_full_collection_start();
1309 1310
1310 1311 // Note: When we have a more flexible GC logging framework that
1311 1312 // allows us to add optional attributes to a GC log record we
1312 1313 // could consider timing and reporting how long we wait in the
1313 1314 // following two methods.
1314 1315 wait_while_free_regions_coming();
1315 1316 // If we start the compaction before the CM threads finish
1316 1317 // scanning the root regions we might trip them over as we'll
1317 1318 // be moving objects / updating references. So let's wait until
1318 1319 // they are done. By telling them to abort, they should complete
1319 1320 // early.
1320 1321 _cm->root_regions()->abort();
1321 1322 _cm->root_regions()->wait_until_scan_finished();
1322 1323 append_secondary_free_list_if_not_empty_with_lock();
1323 1324
1324 1325 gc_prologue(true);
1325 1326 increment_total_collections(true /* full gc */);
1326 1327 increment_old_marking_cycles_started();
1327 1328
1328 1329 assert(used() == recalculate_used(), "Should be equal");
1329 1330
1330 1331 verify_before_gc();
1331 1332
1332 1333 check_bitmaps("Full GC Start");
1333 1334 pre_full_gc_dump(gc_timer);
1334 1335
1335 1336 COMPILER2_PRESENT(DerivedPointerTable::clear());
1336 1337
1337 1338 // Disable discovery and empty the discovered lists
1338 1339 // for the CM ref processor.
1339 1340 ref_processor_cm()->disable_discovery();
1340 1341 ref_processor_cm()->abandon_partial_discovery();
1341 1342 ref_processor_cm()->verify_no_references_recorded();
1342 1343
1343 1344 // Abandon current iterations of concurrent marking and concurrent
1344 1345 // refinement, if any are in progress. We have to do this before
1345 1346 // wait_until_scan_finished() below.
1346 1347 concurrent_mark()->abort();
1347 1348
1348 1349 // Make sure we'll choose a new allocation region afterwards.
1349 1350 _allocator->release_mutator_alloc_region();
1350 1351 _allocator->abandon_gc_alloc_regions();
1351 1352 g1_rem_set()->cleanupHRRS();
1352 1353
1353 1354 // We should call this after we retire any currently active alloc
1354 1355 // regions so that all the ALLOC / RETIRE events are generated
1355 1356 // before the start GC event.
1356 1357 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1357 1358
1358 1359 // We may have added regions to the current incremental collection
1359 1360 // set between the last GC or pause and now. We need to clear the
1360 1361 // incremental collection set and then start rebuilding it afresh
1361 1362 // after this full GC.
1362 1363 abandon_collection_set(g1_policy()->inc_cset_head());
1363 1364 g1_policy()->clear_incremental_cset();
1364 1365 g1_policy()->stop_incremental_cset_building();
1365 1366
1366 1367 tear_down_region_sets(false /* free_list_only */);
1367 1368 g1_policy()->set_gcs_are_young(true);
1368 1369
1369 1370 // See the comments in g1CollectedHeap.hpp and
1370 1371 // G1CollectedHeap::ref_processing_init() about
1371 1372 // how reference processing currently works in G1.
1372 1373
1373 1374 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1374 1375 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1375 1376
1376 1377 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1377 1378 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1378 1379
1379 1380 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1380 1381 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1381 1382
1382 1383 // Do collection work
1383 1384 {
1384 1385 HandleMark hm; // Discard invalid handles created during gc
1385 1386 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1386 1387 }
1387 1388
1388 1389 assert(num_free_regions() == 0, "we should not have added any free regions");
1389 1390 rebuild_region_sets(false /* free_list_only */);
1390 1391
1391 1392 // Enqueue any discovered reference objects that have
1392 1393 // not been removed from the discovered lists.
1393 1394 ref_processor_stw()->enqueue_discovered_references();
1394 1395
1395 1396 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1396 1397
1397 1398 MemoryService::track_memory_usage();
1398 1399
1399 1400 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1400 1401 ref_processor_stw()->verify_no_references_recorded();
1401 1402
1402 1403 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1403 1404 ClassLoaderDataGraph::purge();
1404 1405 MetaspaceAux::verify_metrics();
1405 1406
1406 1407 // Note: since we've just done a full GC, concurrent
1407 1408 // marking is no longer active. Therefore we need not
1408 1409 // re-enable reference discovery for the CM ref processor.
1409 1410 // That will be done at the start of the next marking cycle.
1410 1411 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1411 1412 ref_processor_cm()->verify_no_references_recorded();
1412 1413
1413 1414 reset_gc_time_stamp();
1414 1415 // Since everything potentially moved, we will clear all remembered
1415 1416 // sets, and clear all cards. Later we will rebuild remembered
1416 1417 // sets. We will also reset the GC time stamps of the regions.
1417 1418 clear_rsets_post_compaction();
1418 1419 check_gc_time_stamps();
1419 1420
1420 1421 // Resize the heap if necessary.
1421 1422 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1422 1423
1423 1424 if (_hr_printer.is_active()) {
1424 1425 // We should do this after we potentially resize the heap so
1425 1426 // that all the COMMIT / UNCOMMIT events are generated before
1426 1427 // the end GC event.
1427 1428
1428 1429 print_hrm_post_compaction();
1429 1430 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1430 1431 }
1431 1432
1432 1433 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
1433 1434 if (hot_card_cache->use_cache()) {
1434 1435 hot_card_cache->reset_card_counts();
1435 1436 hot_card_cache->reset_hot_cache();
1436 1437 }
1437 1438
1438 1439 // Rebuild remembered sets of all regions.
1439 1440 if (G1CollectedHeap::use_parallel_gc_threads()) {
1440 1441 uint n_workers =
1441 1442 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1442 1443 workers()->active_workers(),
1443 1444 Threads::number_of_non_daemon_threads());
1444 1445 assert(UseDynamicNumberOfGCThreads ||
1445 1446 n_workers == workers()->total_workers(),
1446 1447 "If not dynamic should be using all the workers");
1447 1448 workers()->set_active_workers(n_workers);
1448 1449 // Set parallel threads in the heap (_n_par_threads) only
1449 1450 // before a parallel phase and always reset it to 0 after
1450 1451 // the phase so that the number of parallel threads does
1451 1452 // no get carried forward to a serial phase where there
1452 1453 // may be code that is "possibly_parallel".
1453 1454 set_par_threads(n_workers);
1454 1455
1455 1456 ParRebuildRSTask rebuild_rs_task(this);
1456 1457 assert(check_heap_region_claim_values(
1457 1458 HeapRegion::InitialClaimValue), "sanity check");
1458 1459 assert(UseDynamicNumberOfGCThreads ||
1459 1460 workers()->active_workers() == workers()->total_workers(),
1460 1461 "Unless dynamic should use total workers");
1461 1462 // Use the most recent number of active workers
1462 1463 assert(workers()->active_workers() > 0,
1463 1464 "Active workers not properly set");
1464 1465 set_par_threads(workers()->active_workers());
1465 1466 workers()->run_task(&rebuild_rs_task);
1466 1467 set_par_threads(0);
1467 1468 assert(check_heap_region_claim_values(
1468 1469 HeapRegion::RebuildRSClaimValue), "sanity check");
1469 1470 reset_heap_region_claim_values();
1470 1471 } else {
1471 1472 RebuildRSOutOfRegionClosure rebuild_rs(this);
1472 1473 heap_region_iterate(&rebuild_rs);
1473 1474 }
1474 1475
1475 1476 // Rebuild the strong code root lists for each region
1476 1477 rebuild_strong_code_roots();
1477 1478
1478 1479 if (true) { // FIXME
1479 1480 MetaspaceGC::compute_new_size();
1480 1481 }
1481 1482
1482 1483 #ifdef TRACESPINNING
1483 1484 ParallelTaskTerminator::print_termination_counts();
1484 1485 #endif
1485 1486
1486 1487 // Discard all rset updates
1487 1488 JavaThread::dirty_card_queue_set().abandon_logs();
1488 1489 assert(dirty_card_queue_set().completed_buffers_num() == 0, "DCQS should be empty");
1489 1490
1490 1491 _young_list->reset_sampled_info();
1491 1492 // At this point there should be no regions in the
1492 1493 // entire heap tagged as young.
1493 1494 assert(check_young_list_empty(true /* check_heap */),
1494 1495 "young list should be empty at this point");
1495 1496
1496 1497 // Update the number of full collections that have been completed.
1497 1498 increment_old_marking_cycles_completed(false /* concurrent */);
1498 1499
1499 1500 _hrm.verify_optional();
1500 1501 verify_region_sets_optional();
1501 1502
1502 1503 verify_after_gc();
1503 1504
1504 1505 // Clear the previous marking bitmap, if needed for bitmap verification.
1505 1506 // Note we cannot do this when we clear the next marking bitmap in
1506 1507 // ConcurrentMark::abort() above since VerifyDuringGC verifies the
1507 1508 // objects marked during a full GC against the previous bitmap.
1508 1509 // But we need to clear it before calling check_bitmaps below since
1509 1510 // the full GC has compacted objects and updated TAMS but not updated
1510 1511 // the prev bitmap.
1511 1512 if (G1VerifyBitmaps) {
1512 1513 ((CMBitMap*) concurrent_mark()->prevMarkBitMap())->clearAll();
1513 1514 }
1514 1515 check_bitmaps("Full GC End");
1515 1516
1516 1517 // Start a new incremental collection set for the next pause
1517 1518 assert(g1_policy()->collection_set() == NULL, "must be");
1518 1519 g1_policy()->start_incremental_cset_building();
1519 1520
1520 1521 clear_cset_fast_test();
1521 1522
1522 1523 _allocator->init_mutator_alloc_region();
1523 1524
1524 1525 double end = os::elapsedTime();
1525 1526 g1_policy()->record_full_collection_end();
1526 1527
1527 1528 if (G1Log::fine()) {
1528 1529 g1_policy()->print_heap_transition();
1529 1530 }
1530 1531
1531 1532 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1532 1533 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1533 1534 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1534 1535 // before any GC notifications are raised.
1535 1536 g1mm()->update_sizes();
1536 1537
1537 1538 gc_epilogue(true);
1538 1539 }
1539 1540
1540 1541 if (G1Log::finer()) {
1541 1542 g1_policy()->print_detailed_heap_transition(true /* full */);
1542 1543 }
1543 1544
1544 1545 print_heap_after_gc();
1545 1546 trace_heap_after_gc(gc_tracer);
1546 1547
1547 1548 post_full_gc_dump(gc_timer);
1548 1549
1549 1550 gc_timer->register_gc_end();
1550 1551 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1551 1552 }
1552 1553
1553 1554 return true;
1554 1555 }
1555 1556
1556 1557 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1557 1558 // do_collection() will return whether it succeeded in performing
1558 1559 // the GC. Currently, there is no facility on the
1559 1560 // do_full_collection() API to notify the caller than the collection
1560 1561 // did not succeed (e.g., because it was locked out by the GC
1561 1562 // locker). So, right now, we'll ignore the return value.
1562 1563 bool dummy = do_collection(true, /* explicit_gc */
1563 1564 clear_all_soft_refs,
1564 1565 0 /* word_size */);
1565 1566 }
1566 1567
1567 1568 // This code is mostly copied from TenuredGeneration.
1568 1569 void
1569 1570 G1CollectedHeap::
1570 1571 resize_if_necessary_after_full_collection(size_t word_size) {
1571 1572 // Include the current allocation, if any, and bytes that will be
1572 1573 // pre-allocated to support collections, as "used".
1573 1574 const size_t used_after_gc = used();
1574 1575 const size_t capacity_after_gc = capacity();
1575 1576 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1576 1577
1577 1578 // This is enforced in arguments.cpp.
1578 1579 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1579 1580 "otherwise the code below doesn't make sense");
1580 1581
1581 1582 // We don't have floating point command-line arguments
1582 1583 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1583 1584 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1584 1585 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1585 1586 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1586 1587
1587 1588 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1588 1589 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1589 1590
1590 1591 // We have to be careful here as these two calculations can overflow
1591 1592 // 32-bit size_t's.
1592 1593 double used_after_gc_d = (double) used_after_gc;
1593 1594 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1594 1595 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1595 1596
1596 1597 // Let's make sure that they are both under the max heap size, which
1597 1598 // by default will make them fit into a size_t.
1598 1599 double desired_capacity_upper_bound = (double) max_heap_size;
1599 1600 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1600 1601 desired_capacity_upper_bound);
1601 1602 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1602 1603 desired_capacity_upper_bound);
1603 1604
1604 1605 // We can now safely turn them into size_t's.
1605 1606 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1606 1607 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1607 1608
1608 1609 // This assert only makes sense here, before we adjust them
1609 1610 // with respect to the min and max heap size.
1610 1611 assert(minimum_desired_capacity <= maximum_desired_capacity,
1611 1612 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1612 1613 "maximum_desired_capacity = "SIZE_FORMAT,
1613 1614 minimum_desired_capacity, maximum_desired_capacity));
1614 1615
1615 1616 // Should not be greater than the heap max size. No need to adjust
1616 1617 // it with respect to the heap min size as it's a lower bound (i.e.,
1617 1618 // we'll try to make the capacity larger than it, not smaller).
1618 1619 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1619 1620 // Should not be less than the heap min size. No need to adjust it
1620 1621 // with respect to the heap max size as it's an upper bound (i.e.,
1621 1622 // we'll try to make the capacity smaller than it, not greater).
1622 1623 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1623 1624
1624 1625 if (capacity_after_gc < minimum_desired_capacity) {
1625 1626 // Don't expand unless it's significant
1626 1627 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1627 1628 ergo_verbose4(ErgoHeapSizing,
1628 1629 "attempt heap expansion",
1629 1630 ergo_format_reason("capacity lower than "
1630 1631 "min desired capacity after Full GC")
1631 1632 ergo_format_byte("capacity")
1632 1633 ergo_format_byte("occupancy")
1633 1634 ergo_format_byte_perc("min desired capacity"),
1634 1635 capacity_after_gc, used_after_gc,
1635 1636 minimum_desired_capacity, (double) MinHeapFreeRatio);
1636 1637 expand(expand_bytes);
1637 1638
1638 1639 // No expansion, now see if we want to shrink
1639 1640 } else if (capacity_after_gc > maximum_desired_capacity) {
1640 1641 // Capacity too large, compute shrinking size
1641 1642 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1642 1643 ergo_verbose4(ErgoHeapSizing,
1643 1644 "attempt heap shrinking",
1644 1645 ergo_format_reason("capacity higher than "
1645 1646 "max desired capacity after Full GC")
1646 1647 ergo_format_byte("capacity")
1647 1648 ergo_format_byte("occupancy")
1648 1649 ergo_format_byte_perc("max desired capacity"),
1649 1650 capacity_after_gc, used_after_gc,
1650 1651 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1651 1652 shrink(shrink_bytes);
1652 1653 }
1653 1654 }
1654 1655
1655 1656
1656 1657 HeapWord*
1657 1658 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1658 1659 AllocationContext_t context,
1659 1660 bool* succeeded) {
1660 1661 assert_at_safepoint(true /* should_be_vm_thread */);
1661 1662
1662 1663 *succeeded = true;
1663 1664 // Let's attempt the allocation first.
1664 1665 HeapWord* result =
1665 1666 attempt_allocation_at_safepoint(word_size,
1666 1667 context,
1667 1668 false /* expect_null_mutator_alloc_region */);
1668 1669 if (result != NULL) {
1669 1670 assert(*succeeded, "sanity");
1670 1671 return result;
1671 1672 }
1672 1673
1673 1674 // In a G1 heap, we're supposed to keep allocation from failing by
1674 1675 // incremental pauses. Therefore, at least for now, we'll favor
1675 1676 // expansion over collection. (This might change in the future if we can
1676 1677 // do something smarter than full collection to satisfy a failed alloc.)
1677 1678 result = expand_and_allocate(word_size, context);
1678 1679 if (result != NULL) {
1679 1680 assert(*succeeded, "sanity");
1680 1681 return result;
1681 1682 }
1682 1683
1683 1684 // Expansion didn't work, we'll try to do a Full GC.
1684 1685 bool gc_succeeded = do_collection(false, /* explicit_gc */
1685 1686 false, /* clear_all_soft_refs */
1686 1687 word_size);
1687 1688 if (!gc_succeeded) {
1688 1689 *succeeded = false;
1689 1690 return NULL;
1690 1691 }
1691 1692
1692 1693 // Retry the allocation
1693 1694 result = attempt_allocation_at_safepoint(word_size,
1694 1695 context,
1695 1696 true /* expect_null_mutator_alloc_region */);
1696 1697 if (result != NULL) {
1697 1698 assert(*succeeded, "sanity");
1698 1699 return result;
1699 1700 }
1700 1701
1701 1702 // Then, try a Full GC that will collect all soft references.
1702 1703 gc_succeeded = do_collection(false, /* explicit_gc */
1703 1704 true, /* clear_all_soft_refs */
1704 1705 word_size);
1705 1706 if (!gc_succeeded) {
1706 1707 *succeeded = false;
1707 1708 return NULL;
1708 1709 }
1709 1710
1710 1711 // Retry the allocation once more
1711 1712 result = attempt_allocation_at_safepoint(word_size,
1712 1713 context,
1713 1714 true /* expect_null_mutator_alloc_region */);
1714 1715 if (result != NULL) {
1715 1716 assert(*succeeded, "sanity");
1716 1717 return result;
1717 1718 }
1718 1719
1719 1720 assert(!collector_policy()->should_clear_all_soft_refs(),
1720 1721 "Flag should have been handled and cleared prior to this point");
1721 1722
1722 1723 // What else? We might try synchronous finalization later. If the total
1723 1724 // space available is large enough for the allocation, then a more
1724 1725 // complete compaction phase than we've tried so far might be
1725 1726 // appropriate.
1726 1727 assert(*succeeded, "sanity");
1727 1728 return NULL;
1728 1729 }
1729 1730
1730 1731 // Attempting to expand the heap sufficiently
1731 1732 // to support an allocation of the given "word_size". If
1732 1733 // successful, perform the allocation and return the address of the
1733 1734 // allocated block, or else "NULL".
1734 1735
1735 1736 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size, AllocationContext_t context) {
1736 1737 assert_at_safepoint(true /* should_be_vm_thread */);
1737 1738
1738 1739 verify_region_sets_optional();
1739 1740
1740 1741 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1741 1742 ergo_verbose1(ErgoHeapSizing,
1742 1743 "attempt heap expansion",
1743 1744 ergo_format_reason("allocation request failed")
1744 1745 ergo_format_byte("allocation request"),
1745 1746 word_size * HeapWordSize);
1746 1747 if (expand(expand_bytes)) {
1747 1748 _hrm.verify_optional();
1748 1749 verify_region_sets_optional();
1749 1750 return attempt_allocation_at_safepoint(word_size,
1750 1751 context,
1751 1752 false /* expect_null_mutator_alloc_region */);
1752 1753 }
1753 1754 return NULL;
1754 1755 }
1755 1756
1756 1757 bool G1CollectedHeap::expand(size_t expand_bytes) {
1757 1758 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1758 1759 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1759 1760 HeapRegion::GrainBytes);
1760 1761 ergo_verbose2(ErgoHeapSizing,
1761 1762 "expand the heap",
1762 1763 ergo_format_byte("requested expansion amount")
1763 1764 ergo_format_byte("attempted expansion amount"),
1764 1765 expand_bytes, aligned_expand_bytes);
1765 1766
1766 1767 if (is_maximal_no_gc()) {
1767 1768 ergo_verbose0(ErgoHeapSizing,
1768 1769 "did not expand the heap",
1769 1770 ergo_format_reason("heap already fully expanded"));
1770 1771 return false;
1771 1772 }
1772 1773
1773 1774 uint regions_to_expand = (uint)(aligned_expand_bytes / HeapRegion::GrainBytes);
1774 1775 assert(regions_to_expand > 0, "Must expand by at least one region");
1775 1776
1776 1777 uint expanded_by = _hrm.expand_by(regions_to_expand);
1777 1778
1778 1779 if (expanded_by > 0) {
1779 1780 size_t actual_expand_bytes = expanded_by * HeapRegion::GrainBytes;
1780 1781 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1781 1782 g1_policy()->record_new_heap_size(num_regions());
1782 1783 } else {
1783 1784 ergo_verbose0(ErgoHeapSizing,
1784 1785 "did not expand the heap",
1785 1786 ergo_format_reason("heap expansion operation failed"));
1786 1787 // The expansion of the virtual storage space was unsuccessful.
1787 1788 // Let's see if it was because we ran out of swap.
1788 1789 if (G1ExitOnExpansionFailure &&
1789 1790 _hrm.available() >= regions_to_expand) {
1790 1791 // We had head room...
1791 1792 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion");
1792 1793 }
1793 1794 }
1794 1795 return regions_to_expand > 0;
1795 1796 }
1796 1797
1797 1798 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1798 1799 size_t aligned_shrink_bytes =
1799 1800 ReservedSpace::page_align_size_down(shrink_bytes);
1800 1801 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1801 1802 HeapRegion::GrainBytes);
1802 1803 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes);
1803 1804
1804 1805 uint num_regions_removed = _hrm.shrink_by(num_regions_to_remove);
1805 1806 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes;
1806 1807
1807 1808 ergo_verbose3(ErgoHeapSizing,
1808 1809 "shrink the heap",
1809 1810 ergo_format_byte("requested shrinking amount")
1810 1811 ergo_format_byte("aligned shrinking amount")
1811 1812 ergo_format_byte("attempted shrinking amount"),
1812 1813 shrink_bytes, aligned_shrink_bytes, shrunk_bytes);
1813 1814 if (num_regions_removed > 0) {
1814 1815 g1_policy()->record_new_heap_size(num_regions());
1815 1816 } else {
1816 1817 ergo_verbose0(ErgoHeapSizing,
1817 1818 "did not shrink the heap",
1818 1819 ergo_format_reason("heap shrinking operation failed"));
1819 1820 }
1820 1821 }
1821 1822
1822 1823 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1823 1824 verify_region_sets_optional();
1824 1825
1825 1826 // We should only reach here at the end of a Full GC which means we
1826 1827 // should not not be holding to any GC alloc regions. The method
1827 1828 // below will make sure of that and do any remaining clean up.
1828 1829 _allocator->abandon_gc_alloc_regions();
1829 1830
1830 1831 // Instead of tearing down / rebuilding the free lists here, we
1831 1832 // could instead use the remove_all_pending() method on free_list to
1832 1833 // remove only the ones that we need to remove.
1833 1834 tear_down_region_sets(true /* free_list_only */);
1834 1835 shrink_helper(shrink_bytes);
1835 1836 rebuild_region_sets(true /* free_list_only */);
1836 1837
1837 1838 _hrm.verify_optional();
1838 1839 verify_region_sets_optional();
1839 1840 }
1840 1841
1841 1842 // Public methods.
1842 1843
1843 1844 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1844 1845 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1845 1846 #endif // _MSC_VER
1846 1847
1847 1848
1848 1849 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1849 1850 SharedHeap(policy_),
1850 1851 _g1_policy(policy_),
1851 1852 _dirty_card_queue_set(false),
1852 1853 _into_cset_dirty_card_queue_set(false),
1853 1854 _is_alive_closure_cm(this),
1854 1855 _is_alive_closure_stw(this),
1855 1856 _ref_processor_cm(NULL),
1856 1857 _ref_processor_stw(NULL),
1857 1858 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1858 1859 _bot_shared(NULL),
1859 1860 _evac_failure_scan_stack(NULL),
1860 1861 _mark_in_progress(false),
1861 1862 _cg1r(NULL),
1862 1863 _g1mm(NULL),
1863 1864 _refine_cte_cl(NULL),
1864 1865 _full_collection(false),
1865 1866 _secondary_free_list("Secondary Free List", new SecondaryFreeRegionListMtSafeChecker()),
1866 1867 _old_set("Old Set", false /* humongous */, new OldRegionSetMtSafeChecker()),
1867 1868 _humongous_set("Master Humongous Set", true /* humongous */, new HumongousRegionSetMtSafeChecker()),
1868 1869 _humongous_is_live(),
1869 1870 _has_humongous_reclaim_candidates(false),
1870 1871 _free_regions_coming(false),
1871 1872 _young_list(new YoungList(this)),
1872 1873 _gc_time_stamp(0),
1873 1874 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1874 1875 _old_plab_stats(OldPLABSize, PLABWeight),
1875 1876 _expand_heap_after_alloc_failure(true),
1876 1877 _surviving_young_words(NULL),
1877 1878 _old_marking_cycles_started(0),
1878 1879 _old_marking_cycles_completed(0),
1879 1880 _concurrent_cycle_started(false),
1880 1881 _in_cset_fast_test(),
1881 1882 _dirty_cards_region_list(NULL),
1882 1883 _worker_cset_start_region(NULL),
1883 1884 _worker_cset_start_region_time_stamp(NULL),
1884 1885 _gc_timer_stw(new (ResourceObj::C_HEAP, mtGC) STWGCTimer()),
1885 1886 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
1886 1887 _gc_tracer_stw(new (ResourceObj::C_HEAP, mtGC) G1NewTracer()),
1887 1888 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()) {
1888 1889
1889 1890 _g1h = this;
1890 1891 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1891 1892 vm_exit_during_initialization("Failed necessary allocation.");
1892 1893 }
1893 1894
1894 1895 _allocator = G1Allocator::create_allocator(_g1h);
1895 1896 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1896 1897
1897 1898 int n_queues = MAX2((int)ParallelGCThreads, 1);
1898 1899 _task_queues = new RefToScanQueueSet(n_queues);
1899 1900
1900 1901 uint n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1901 1902 assert(n_rem_sets > 0, "Invariant.");
1902 1903
1903 1904 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1904 1905 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1905 1906 _evacuation_failed_info_array = NEW_C_HEAP_ARRAY(EvacuationFailedInfo, n_queues, mtGC);
1906 1907
1907 1908 for (int i = 0; i < n_queues; i++) {
1908 1909 RefToScanQueue* q = new RefToScanQueue();
1909 1910 q->initialize();
1910 1911 _task_queues->register_queue(i, q);
1911 1912 ::new (&_evacuation_failed_info_array[i]) EvacuationFailedInfo();
1912 1913 }
1913 1914 clear_cset_start_regions();
1914 1915
1915 1916 // Initialize the G1EvacuationFailureALot counters and flags.
1916 1917 NOT_PRODUCT(reset_evacuation_should_fail();)
1917 1918
1918 1919 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1919 1920 }
1920 1921
1921 1922 jint G1CollectedHeap::initialize() {
1922 1923 CollectedHeap::pre_initialize();
1923 1924 os::enable_vtime();
1924 1925
1925 1926 G1Log::init();
1926 1927
1927 1928 // Necessary to satisfy locking discipline assertions.
1928 1929
1929 1930 MutexLocker x(Heap_lock);
1930 1931
1931 1932 // We have to initialize the printer before committing the heap, as
1932 1933 // it will be used then.
1933 1934 _hr_printer.set_active(G1PrintHeapRegions);
1934 1935
1935 1936 // While there are no constraints in the GC code that HeapWordSize
1936 1937 // be any particular value, there are multiple other areas in the
1937 1938 // system which believe this to be true (e.g. oop->object_size in some
1938 1939 // cases incorrectly returns the size in wordSize units rather than
1939 1940 // HeapWordSize).
1940 1941 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1941 1942
1942 1943 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1943 1944 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1944 1945 size_t heap_alignment = collector_policy()->heap_alignment();
1945 1946
1946 1947 // Ensure that the sizes are properly aligned.
1947 1948 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1948 1949 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1949 1950 Universe::check_alignment(max_byte_size, heap_alignment, "g1 heap");
1950 1951
1951 1952 _refine_cte_cl = new RefineCardTableEntryClosure();
1952 1953
1953 1954 _cg1r = new ConcurrentG1Refine(this, _refine_cte_cl);
1954 1955
1955 1956 // Reserve the maximum.
1956 1957
1957 1958 // When compressed oops are enabled, the preferred heap base
1958 1959 // is calculated by subtracting the requested size from the
1959 1960 // 32Gb boundary and using the result as the base address for
1960 1961 // heap reservation. If the requested size is not aligned to
1961 1962 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1962 1963 // into the ReservedHeapSpace constructor) then the actual
1963 1964 // base of the reserved heap may end up differing from the
1964 1965 // address that was requested (i.e. the preferred heap base).
1965 1966 // If this happens then we could end up using a non-optimal
1966 1967 // compressed oops mode.
1967 1968
1968 1969 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
1969 1970 heap_alignment);
1970 1971
1971 1972 // It is important to do this in a way such that concurrent readers can't
1972 1973 // temporarily think something is in the heap. (I've actually seen this
1973 1974 // happen in asserts: DLD.)
1974 1975 _reserved.set_word_size(0);
1975 1976 _reserved.set_start((HeapWord*)heap_rs.base());
1976 1977 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
1977 1978
1978 1979 // Create the gen rem set (and barrier set) for the entire reserved region.
1979 1980 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
1980 1981 set_barrier_set(rem_set()->bs());
1981 1982 if (!barrier_set()->is_a(BarrierSet::G1SATBCTLogging)) {
1982 1983 vm_exit_during_initialization("G1 requires a G1SATBLoggingCardTableModRefBS");
1983 1984 return JNI_ENOMEM;
1984 1985 }
1985 1986
1986 1987 // Also create a G1 rem set.
1987 1988 _g1_rem_set = new G1RemSet(this, g1_barrier_set());
1988 1989
1989 1990 // Carve out the G1 part of the heap.
1990 1991
1991 1992 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
1992 1993 G1RegionToSpaceMapper* heap_storage =
1993 1994 G1RegionToSpaceMapper::create_mapper(g1_rs,
1994 1995 UseLargePages ? os::large_page_size() : os::vm_page_size(),
1995 1996 HeapRegion::GrainBytes,
1996 1997 1,
1997 1998 mtJavaHeap);
1998 1999 heap_storage->set_mapping_changed_listener(&_listener);
1999 2000
2000 2001 // Reserve space for the block offset table. We do not support automatic uncommit
2001 2002 // for the card table at this time. BOT only.
2002 2003 ReservedSpace bot_rs(G1BlockOffsetSharedArray::compute_size(g1_rs.size() / HeapWordSize));
2003 2004 G1RegionToSpaceMapper* bot_storage =
2004 2005 G1RegionToSpaceMapper::create_mapper(bot_rs,
2005 2006 os::vm_page_size(),
2006 2007 HeapRegion::GrainBytes,
2007 2008 G1BlockOffsetSharedArray::N_bytes,
2008 2009 mtGC);
2009 2010
2010 2011 ReservedSpace cardtable_rs(G1SATBCardTableLoggingModRefBS::compute_size(g1_rs.size() / HeapWordSize));
2011 2012 G1RegionToSpaceMapper* cardtable_storage =
2012 2013 G1RegionToSpaceMapper::create_mapper(cardtable_rs,
2013 2014 os::vm_page_size(),
2014 2015 HeapRegion::GrainBytes,
2015 2016 G1BlockOffsetSharedArray::N_bytes,
2016 2017 mtGC);
2017 2018
2018 2019 // Reserve space for the card counts table.
2019 2020 ReservedSpace card_counts_rs(G1BlockOffsetSharedArray::compute_size(g1_rs.size() / HeapWordSize));
2020 2021 G1RegionToSpaceMapper* card_counts_storage =
2021 2022 G1RegionToSpaceMapper::create_mapper(card_counts_rs,
2022 2023 os::vm_page_size(),
2023 2024 HeapRegion::GrainBytes,
2024 2025 G1BlockOffsetSharedArray::N_bytes,
2025 2026 mtGC);
2026 2027
2027 2028 // Reserve space for prev and next bitmap.
2028 2029 size_t bitmap_size = CMBitMap::compute_size(g1_rs.size());
2029 2030
2030 2031 ReservedSpace prev_bitmap_rs(ReservedSpace::allocation_align_size_up(bitmap_size));
2031 2032 G1RegionToSpaceMapper* prev_bitmap_storage =
2032 2033 G1RegionToSpaceMapper::create_mapper(prev_bitmap_rs,
2033 2034 os::vm_page_size(),
2034 2035 HeapRegion::GrainBytes,
2035 2036 CMBitMap::mark_distance(),
2036 2037 mtGC);
2037 2038
2038 2039 ReservedSpace next_bitmap_rs(ReservedSpace::allocation_align_size_up(bitmap_size));
2039 2040 G1RegionToSpaceMapper* next_bitmap_storage =
2040 2041 G1RegionToSpaceMapper::create_mapper(next_bitmap_rs,
2041 2042 os::vm_page_size(),
2042 2043 HeapRegion::GrainBytes,
2043 2044 CMBitMap::mark_distance(),
2044 2045 mtGC);
2045 2046
2046 2047 _hrm.initialize(heap_storage, prev_bitmap_storage, next_bitmap_storage, bot_storage, cardtable_storage, card_counts_storage);
2047 2048 g1_barrier_set()->initialize(cardtable_storage);
2048 2049 // Do later initialization work for concurrent refinement.
2049 2050 _cg1r->init(card_counts_storage);
2050 2051
2051 2052 // 6843694 - ensure that the maximum region index can fit
2052 2053 // in the remembered set structures.
2053 2054 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2054 2055 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2055 2056
2056 2057 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2057 2058 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2058 2059 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2059 2060 "too many cards per region");
2060 2061
2061 2062 FreeRegionList::set_unrealistically_long_length(max_regions() + 1);
2062 2063
2063 2064 _bot_shared = new G1BlockOffsetSharedArray(_reserved, bot_storage);
2064 2065
2065 2066 _g1h = this;
2066 2067
2067 2068 _in_cset_fast_test.initialize(_hrm.reserved().start(), _hrm.reserved().end(), HeapRegion::GrainBytes);
2068 2069 _humongous_is_live.initialize(_hrm.reserved().start(), _hrm.reserved().end(), HeapRegion::GrainBytes);
2069 2070
2070 2071 // Create the ConcurrentMark data structure and thread.
2071 2072 // (Must do this late, so that "max_regions" is defined.)
2072 2073 _cm = new ConcurrentMark(this, prev_bitmap_storage, next_bitmap_storage);
2073 2074 if (_cm == NULL || !_cm->completed_initialization()) {
2074 2075 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2075 2076 return JNI_ENOMEM;
2076 2077 }
2077 2078 _cmThread = _cm->cmThread();
2078 2079
2079 2080 // Initialize the from_card cache structure of HeapRegionRemSet.
2080 2081 HeapRegionRemSet::init_heap(max_regions());
2081 2082
2082 2083 // Now expand into the initial heap size.
2083 2084 if (!expand(init_byte_size)) {
2084 2085 vm_shutdown_during_initialization("Failed to allocate initial heap.");
2085 2086 return JNI_ENOMEM;
2086 2087 }
2087 2088
2088 2089 // Perform any initialization actions delegated to the policy.
2089 2090 g1_policy()->init();
2090 2091
2091 2092 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2092 2093 SATB_Q_FL_lock,
2093 2094 G1SATBProcessCompletedThreshold,
2094 2095 Shared_SATB_Q_lock);
2095 2096
2096 2097 JavaThread::dirty_card_queue_set().initialize(_refine_cte_cl,
2097 2098 DirtyCardQ_CBL_mon,
2098 2099 DirtyCardQ_FL_lock,
2099 2100 concurrent_g1_refine()->yellow_zone(),
2100 2101 concurrent_g1_refine()->red_zone(),
2101 2102 Shared_DirtyCardQ_lock);
2102 2103
2103 2104 dirty_card_queue_set().initialize(NULL, // Should never be called by the Java code
2104 2105 DirtyCardQ_CBL_mon,
2105 2106 DirtyCardQ_FL_lock,
2106 2107 -1, // never trigger processing
2107 2108 -1, // no limit on length
2108 2109 Shared_DirtyCardQ_lock,
2109 2110 &JavaThread::dirty_card_queue_set());
2110 2111
2111 2112 // Initialize the card queue set used to hold cards containing
2112 2113 // references into the collection set.
2113 2114 _into_cset_dirty_card_queue_set.initialize(NULL, // Should never be called by the Java code
2114 2115 DirtyCardQ_CBL_mon,
2115 2116 DirtyCardQ_FL_lock,
2116 2117 -1, // never trigger processing
2117 2118 -1, // no limit on length
2118 2119 Shared_DirtyCardQ_lock,
2119 2120 &JavaThread::dirty_card_queue_set());
2120 2121
2121 2122 // In case we're keeping closure specialization stats, initialize those
2122 2123 // counts and that mechanism.
2123 2124 SpecializationStats::clear();
2124 2125
2125 2126 // Here we allocate the dummy HeapRegion that is required by the
2126 2127 // G1AllocRegion class.
2127 2128 HeapRegion* dummy_region = _hrm.get_dummy_region();
2128 2129
2129 2130 // We'll re-use the same region whether the alloc region will
2130 2131 // require BOT updates or not and, if it doesn't, then a non-young
2131 2132 // region will complain that it cannot support allocations without
2132 2133 // BOT updates. So we'll tag the dummy region as eden to avoid that.
2133 2134 dummy_region->set_eden();
2134 2135 // Make sure it's full.
2135 2136 dummy_region->set_top(dummy_region->end());
2136 2137 G1AllocRegion::setup(this, dummy_region);
2137 2138
2138 2139 _allocator->init_mutator_alloc_region();
2139 2140
2140 2141 // Do create of the monitoring and management support so that
2141 2142 // values in the heap have been properly initialized.
2142 2143 _g1mm = new G1MonitoringSupport(this);
2143 2144
2144 2145 G1StringDedup::initialize();
2145 2146
2146 2147 return JNI_OK;
2147 2148 }
2148 2149
2149 2150 void G1CollectedHeap::stop() {
2150 2151 // Stop all concurrent threads. We do this to make sure these threads
2151 2152 // do not continue to execute and access resources (e.g. gclog_or_tty)
2152 2153 // that are destroyed during shutdown.
2153 2154 _cg1r->stop();
2154 2155 _cmThread->stop();
2155 2156 if (G1StringDedup::is_enabled()) {
2156 2157 G1StringDedup::stop();
2157 2158 }
2158 2159 }
2159 2160
2160 2161 void G1CollectedHeap::clear_humongous_is_live_table() {
2161 2162 guarantee(G1ReclaimDeadHumongousObjectsAtYoungGC, "Should only be called if true");
2162 2163 _humongous_is_live.clear();
2163 2164 }
2164 2165
2165 2166 size_t G1CollectedHeap::conservative_max_heap_alignment() {
2166 2167 return HeapRegion::max_region_size();
2167 2168 }
2168 2169
2169 2170 void G1CollectedHeap::ref_processing_init() {
2170 2171 // Reference processing in G1 currently works as follows:
2171 2172 //
2172 2173 // * There are two reference processor instances. One is
2173 2174 // used to record and process discovered references
2174 2175 // during concurrent marking; the other is used to
2175 2176 // record and process references during STW pauses
2176 2177 // (both full and incremental).
2177 2178 // * Both ref processors need to 'span' the entire heap as
2178 2179 // the regions in the collection set may be dotted around.
2179 2180 //
2180 2181 // * For the concurrent marking ref processor:
2181 2182 // * Reference discovery is enabled at initial marking.
2182 2183 // * Reference discovery is disabled and the discovered
2183 2184 // references processed etc during remarking.
2184 2185 // * Reference discovery is MT (see below).
2185 2186 // * Reference discovery requires a barrier (see below).
2186 2187 // * Reference processing may or may not be MT
2187 2188 // (depending on the value of ParallelRefProcEnabled
2188 2189 // and ParallelGCThreads).
2189 2190 // * A full GC disables reference discovery by the CM
2190 2191 // ref processor and abandons any entries on it's
2191 2192 // discovered lists.
2192 2193 //
2193 2194 // * For the STW processor:
2194 2195 // * Non MT discovery is enabled at the start of a full GC.
2195 2196 // * Processing and enqueueing during a full GC is non-MT.
2196 2197 // * During a full GC, references are processed after marking.
2197 2198 //
2198 2199 // * Discovery (may or may not be MT) is enabled at the start
2199 2200 // of an incremental evacuation pause.
2200 2201 // * References are processed near the end of a STW evacuation pause.
2201 2202 // * For both types of GC:
2202 2203 // * Discovery is atomic - i.e. not concurrent.
2203 2204 // * Reference discovery will not need a barrier.
2204 2205
2205 2206 SharedHeap::ref_processing_init();
2206 2207 MemRegion mr = reserved_region();
2207 2208
2208 2209 // Concurrent Mark ref processor
2209 2210 _ref_processor_cm =
2210 2211 new ReferenceProcessor(mr, // span
2211 2212 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2212 2213 // mt processing
2213 2214 (int) ParallelGCThreads,
2214 2215 // degree of mt processing
2215 2216 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2216 2217 // mt discovery
2217 2218 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2218 2219 // degree of mt discovery
2219 2220 false,
2220 2221 // Reference discovery is not atomic
2221 2222 &_is_alive_closure_cm);
2222 2223 // is alive closure
2223 2224 // (for efficiency/performance)
2224 2225
2225 2226 // STW ref processor
2226 2227 _ref_processor_stw =
2227 2228 new ReferenceProcessor(mr, // span
2228 2229 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2229 2230 // mt processing
2230 2231 MAX2((int)ParallelGCThreads, 1),
2231 2232 // degree of mt processing
2232 2233 (ParallelGCThreads > 1),
2233 2234 // mt discovery
2234 2235 MAX2((int)ParallelGCThreads, 1),
2235 2236 // degree of mt discovery
2236 2237 true,
2237 2238 // Reference discovery is atomic
2238 2239 &_is_alive_closure_stw);
2239 2240 // is alive closure
2240 2241 // (for efficiency/performance)
2241 2242 }
2242 2243
2243 2244 size_t G1CollectedHeap::capacity() const {
2244 2245 return _hrm.length() * HeapRegion::GrainBytes;
2245 2246 }
2246 2247
2247 2248 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2248 2249 assert(!hr->continuesHumongous(), "pre-condition");
2249 2250 hr->reset_gc_time_stamp();
2250 2251 if (hr->startsHumongous()) {
2251 2252 uint first_index = hr->hrm_index() + 1;
2252 2253 uint last_index = hr->last_hc_index();
2253 2254 for (uint i = first_index; i < last_index; i += 1) {
2254 2255 HeapRegion* chr = region_at(i);
2255 2256 assert(chr->continuesHumongous(), "sanity");
2256 2257 chr->reset_gc_time_stamp();
2257 2258 }
2258 2259 }
2259 2260 }
2260 2261
2261 2262 #ifndef PRODUCT
2262 2263 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2263 2264 private:
2264 2265 unsigned _gc_time_stamp;
2265 2266 bool _failures;
2266 2267
2267 2268 public:
2268 2269 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2269 2270 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2270 2271
2271 2272 virtual bool doHeapRegion(HeapRegion* hr) {
2272 2273 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2273 2274 if (_gc_time_stamp != region_gc_time_stamp) {
2274 2275 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2275 2276 "expected %d", HR_FORMAT_PARAMS(hr),
2276 2277 region_gc_time_stamp, _gc_time_stamp);
2277 2278 _failures = true;
2278 2279 }
2279 2280 return false;
2280 2281 }
2281 2282
2282 2283 bool failures() { return _failures; }
2283 2284 };
2284 2285
2285 2286 void G1CollectedHeap::check_gc_time_stamps() {
2286 2287 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2287 2288 heap_region_iterate(&cl);
2288 2289 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2289 2290 }
2290 2291 #endif // PRODUCT
2291 2292
2292 2293 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2293 2294 DirtyCardQueue* into_cset_dcq,
2294 2295 bool concurrent,
2295 2296 uint worker_i) {
2296 2297 // Clean cards in the hot card cache
2297 2298 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
2298 2299 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq);
2299 2300
2300 2301 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2301 2302 int n_completed_buffers = 0;
2302 2303 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2303 2304 n_completed_buffers++;
2304 2305 }
2305 2306 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2306 2307 dcqs.clear_n_completed_buffers();
2307 2308 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2308 2309 }
2309 2310
2310 2311
2311 2312 // Computes the sum of the storage used by the various regions.
2312 2313 size_t G1CollectedHeap::used() const {
2313 2314 return _allocator->used();
2314 2315 }
2315 2316
2316 2317 size_t G1CollectedHeap::used_unlocked() const {
2317 2318 return _allocator->used_unlocked();
2318 2319 }
2319 2320
2320 2321 class SumUsedClosure: public HeapRegionClosure {
2321 2322 size_t _used;
2322 2323 public:
2323 2324 SumUsedClosure() : _used(0) {}
2324 2325 bool doHeapRegion(HeapRegion* r) {
2325 2326 if (!r->continuesHumongous()) {
2326 2327 _used += r->used();
2327 2328 }
2328 2329 return false;
2329 2330 }
2330 2331 size_t result() { return _used; }
2331 2332 };
2332 2333
2333 2334 size_t G1CollectedHeap::recalculate_used() const {
2334 2335 double recalculate_used_start = os::elapsedTime();
2335 2336
2336 2337 SumUsedClosure blk;
2337 2338 heap_region_iterate(&blk);
2338 2339
2339 2340 g1_policy()->phase_times()->record_evac_fail_recalc_used_time((os::elapsedTime() - recalculate_used_start) * 1000.0);
2340 2341 return blk.result();
2341 2342 }
2342 2343
2343 2344 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2344 2345 switch (cause) {
2345 2346 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2346 2347 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2347 2348 case GCCause::_g1_humongous_allocation: return true;
2348 2349 case GCCause::_update_allocation_context_stats_inc: return true;
2349 2350 default: return false;
2350 2351 }
2351 2352 }
2352 2353
2353 2354 #ifndef PRODUCT
2354 2355 void G1CollectedHeap::allocate_dummy_regions() {
2355 2356 // Let's fill up most of the region
2356 2357 size_t word_size = HeapRegion::GrainWords - 1024;
2357 2358 // And as a result the region we'll allocate will be humongous.
2358 2359 guarantee(isHumongous(word_size), "sanity");
2359 2360
2360 2361 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2361 2362 // Let's use the existing mechanism for the allocation
2362 2363 HeapWord* dummy_obj = humongous_obj_allocate(word_size,
2363 2364 AllocationContext::system());
2364 2365 if (dummy_obj != NULL) {
2365 2366 MemRegion mr(dummy_obj, word_size);
2366 2367 CollectedHeap::fill_with_object(mr);
2367 2368 } else {
2368 2369 // If we can't allocate once, we probably cannot allocate
2369 2370 // again. Let's get out of the loop.
2370 2371 break;
2371 2372 }
2372 2373 }
2373 2374 }
2374 2375 #endif // !PRODUCT
2375 2376
2376 2377 void G1CollectedHeap::increment_old_marking_cycles_started() {
2377 2378 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2378 2379 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2379 2380 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2380 2381 _old_marking_cycles_started, _old_marking_cycles_completed));
2381 2382
2382 2383 _old_marking_cycles_started++;
2383 2384 }
2384 2385
2385 2386 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2386 2387 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2387 2388
2388 2389 // We assume that if concurrent == true, then the caller is a
2389 2390 // concurrent thread that was joined the Suspendible Thread
2390 2391 // Set. If there's ever a cheap way to check this, we should add an
2391 2392 // assert here.
2392 2393
2393 2394 // Given that this method is called at the end of a Full GC or of a
2394 2395 // concurrent cycle, and those can be nested (i.e., a Full GC can
2395 2396 // interrupt a concurrent cycle), the number of full collections
2396 2397 // completed should be either one (in the case where there was no
2397 2398 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2398 2399 // behind the number of full collections started.
2399 2400
2400 2401 // This is the case for the inner caller, i.e. a Full GC.
2401 2402 assert(concurrent ||
2402 2403 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2403 2404 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2404 2405 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2405 2406 "is inconsistent with _old_marking_cycles_completed = %u",
2406 2407 _old_marking_cycles_started, _old_marking_cycles_completed));
2407 2408
2408 2409 // This is the case for the outer caller, i.e. the concurrent cycle.
2409 2410 assert(!concurrent ||
2410 2411 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2411 2412 err_msg("for outer caller (concurrent cycle): "
2412 2413 "_old_marking_cycles_started = %u "
2413 2414 "is inconsistent with _old_marking_cycles_completed = %u",
2414 2415 _old_marking_cycles_started, _old_marking_cycles_completed));
2415 2416
2416 2417 _old_marking_cycles_completed += 1;
2417 2418
2418 2419 // We need to clear the "in_progress" flag in the CM thread before
2419 2420 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2420 2421 // is set) so that if a waiter requests another System.gc() it doesn't
2421 2422 // incorrectly see that a marking cycle is still in progress.
2422 2423 if (concurrent) {
2423 2424 _cmThread->clear_in_progress();
2424 2425 }
2425 2426
2426 2427 // This notify_all() will ensure that a thread that called
2427 2428 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2428 2429 // and it's waiting for a full GC to finish will be woken up. It is
2429 2430 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2430 2431 FullGCCount_lock->notify_all();
2431 2432 }
2432 2433
2433 2434 void G1CollectedHeap::register_concurrent_cycle_start(const Ticks& start_time) {
2434 2435 _concurrent_cycle_started = true;
2435 2436 _gc_timer_cm->register_gc_start(start_time);
2436 2437
2437 2438 _gc_tracer_cm->report_gc_start(gc_cause(), _gc_timer_cm->gc_start());
2438 2439 trace_heap_before_gc(_gc_tracer_cm);
2439 2440 }
2440 2441
2441 2442 void G1CollectedHeap::register_concurrent_cycle_end() {
2442 2443 if (_concurrent_cycle_started) {
2443 2444 if (_cm->has_aborted()) {
2444 2445 _gc_tracer_cm->report_concurrent_mode_failure();
2445 2446 }
2446 2447
2447 2448 _gc_timer_cm->register_gc_end();
2448 2449 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2449 2450
2450 2451 _concurrent_cycle_started = false;
2451 2452 }
2452 2453 }
2453 2454
2454 2455 void G1CollectedHeap::trace_heap_after_concurrent_cycle() {
2455 2456 if (_concurrent_cycle_started) {
2456 2457 trace_heap_after_gc(_gc_tracer_cm);
2457 2458 }
2458 2459 }
2459 2460
2460 2461 G1YCType G1CollectedHeap::yc_type() {
2461 2462 bool is_young = g1_policy()->gcs_are_young();
2462 2463 bool is_initial_mark = g1_policy()->during_initial_mark_pause();
2463 2464 bool is_during_mark = mark_in_progress();
2464 2465
2465 2466 if (is_initial_mark) {
2466 2467 return InitialMark;
2467 2468 } else if (is_during_mark) {
2468 2469 return DuringMark;
2469 2470 } else if (is_young) {
2470 2471 return Normal;
2471 2472 } else {
2472 2473 return Mixed;
2473 2474 }
2474 2475 }
2475 2476
2476 2477 void G1CollectedHeap::collect(GCCause::Cause cause) {
2477 2478 assert_heap_not_locked();
2478 2479
2479 2480 unsigned int gc_count_before;
2480 2481 unsigned int old_marking_count_before;
2481 2482 unsigned int full_gc_count_before;
2482 2483 bool retry_gc;
2483 2484
2484 2485 do {
2485 2486 retry_gc = false;
2486 2487
2487 2488 {
2488 2489 MutexLocker ml(Heap_lock);
2489 2490
2490 2491 // Read the GC count while holding the Heap_lock
2491 2492 gc_count_before = total_collections();
2492 2493 full_gc_count_before = total_full_collections();
2493 2494 old_marking_count_before = _old_marking_cycles_started;
2494 2495 }
2495 2496
2496 2497 if (should_do_concurrent_full_gc(cause)) {
2497 2498 // Schedule an initial-mark evacuation pause that will start a
2498 2499 // concurrent cycle. We're setting word_size to 0 which means that
2499 2500 // we are not requesting a post-GC allocation.
2500 2501 VM_G1IncCollectionPause op(gc_count_before,
2501 2502 0, /* word_size */
2502 2503 true, /* should_initiate_conc_mark */
2503 2504 g1_policy()->max_pause_time_ms(),
2504 2505 cause);
2505 2506 op.set_allocation_context(AllocationContext::current());
2506 2507
2507 2508 VMThread::execute(&op);
2508 2509 if (!op.pause_succeeded()) {
2509 2510 if (old_marking_count_before == _old_marking_cycles_started) {
2510 2511 retry_gc = op.should_retry_gc();
2511 2512 } else {
2512 2513 // A Full GC happened while we were trying to schedule the
2513 2514 // initial-mark GC. No point in starting a new cycle given
2514 2515 // that the whole heap was collected anyway.
2515 2516 }
2516 2517
2517 2518 if (retry_gc) {
2518 2519 if (GC_locker::is_active_and_needs_gc()) {
2519 2520 GC_locker::stall_until_clear();
2520 2521 }
2521 2522 }
2522 2523 }
2523 2524 } else {
2524 2525 if (cause == GCCause::_gc_locker || cause == GCCause::_wb_young_gc
2525 2526 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2526 2527
2527 2528 // Schedule a standard evacuation pause. We're setting word_size
2528 2529 // to 0 which means that we are not requesting a post-GC allocation.
2529 2530 VM_G1IncCollectionPause op(gc_count_before,
2530 2531 0, /* word_size */
2531 2532 false, /* should_initiate_conc_mark */
2532 2533 g1_policy()->max_pause_time_ms(),
2533 2534 cause);
2534 2535 VMThread::execute(&op);
2535 2536 } else {
2536 2537 // Schedule a Full GC.
2537 2538 VM_G1CollectFull op(gc_count_before, full_gc_count_before, cause);
2538 2539 VMThread::execute(&op);
2539 2540 }
2540 2541 }
2541 2542 } while (retry_gc);
2542 2543 }
2543 2544
2544 2545 bool G1CollectedHeap::is_in(const void* p) const {
2545 2546 if (_hrm.reserved().contains(p)) {
2546 2547 // Given that we know that p is in the reserved space,
2547 2548 // heap_region_containing_raw() should successfully
2548 2549 // return the containing region.
2549 2550 HeapRegion* hr = heap_region_containing_raw(p);
2550 2551 return hr->is_in(p);
2551 2552 } else {
2552 2553 return false;
2553 2554 }
2554 2555 }
2555 2556
2556 2557 #ifdef ASSERT
2557 2558 bool G1CollectedHeap::is_in_exact(const void* p) const {
2558 2559 bool contains = reserved_region().contains(p);
2559 2560 bool available = _hrm.is_available(addr_to_region((HeapWord*)p));
2560 2561 if (contains && available) {
2561 2562 return true;
2562 2563 } else {
2563 2564 return false;
2564 2565 }
2565 2566 }
2566 2567 #endif
2567 2568
2568 2569 // Iteration functions.
2569 2570
2570 2571 // Applies an ExtendedOopClosure onto all references of objects within a HeapRegion.
2571 2572
2572 2573 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2573 2574 ExtendedOopClosure* _cl;
2574 2575 public:
2575 2576 IterateOopClosureRegionClosure(ExtendedOopClosure* cl) : _cl(cl) {}
2576 2577 bool doHeapRegion(HeapRegion* r) {
2577 2578 if (!r->continuesHumongous()) {
2578 2579 r->oop_iterate(_cl);
2579 2580 }
2580 2581 return false;
2581 2582 }
2582 2583 };
2583 2584
2584 2585 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2585 2586 IterateOopClosureRegionClosure blk(cl);
2586 2587 heap_region_iterate(&blk);
2587 2588 }
2588 2589
2589 2590 // Iterates an ObjectClosure over all objects within a HeapRegion.
2590 2591
2591 2592 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2592 2593 ObjectClosure* _cl;
2593 2594 public:
2594 2595 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2595 2596 bool doHeapRegion(HeapRegion* r) {
2596 2597 if (! r->continuesHumongous()) {
2597 2598 r->object_iterate(_cl);
2598 2599 }
2599 2600 return false;
2600 2601 }
2601 2602 };
2602 2603
2603 2604 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2604 2605 IterateObjectClosureRegionClosure blk(cl);
2605 2606 heap_region_iterate(&blk);
2606 2607 }
2607 2608
2608 2609 // Calls a SpaceClosure on a HeapRegion.
2609 2610
2610 2611 class SpaceClosureRegionClosure: public HeapRegionClosure {
2611 2612 SpaceClosure* _cl;
2612 2613 public:
2613 2614 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2614 2615 bool doHeapRegion(HeapRegion* r) {
2615 2616 _cl->do_space(r);
2616 2617 return false;
2617 2618 }
2618 2619 };
2619 2620
2620 2621 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2621 2622 SpaceClosureRegionClosure blk(cl);
2622 2623 heap_region_iterate(&blk);
2623 2624 }
2624 2625
2625 2626 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2626 2627 _hrm.iterate(cl);
2627 2628 }
2628 2629
2629 2630 void
2630 2631 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2631 2632 uint worker_id,
2632 2633 uint num_workers,
2633 2634 jint claim_value) const {
2634 2635 _hrm.par_iterate(cl, worker_id, num_workers, claim_value);
2635 2636 }
2636 2637
2637 2638 class ResetClaimValuesClosure: public HeapRegionClosure {
2638 2639 public:
2639 2640 bool doHeapRegion(HeapRegion* r) {
2640 2641 r->set_claim_value(HeapRegion::InitialClaimValue);
2641 2642 return false;
2642 2643 }
2643 2644 };
2644 2645
2645 2646 void G1CollectedHeap::reset_heap_region_claim_values() {
2646 2647 ResetClaimValuesClosure blk;
2647 2648 heap_region_iterate(&blk);
2648 2649 }
2649 2650
2650 2651 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2651 2652 ResetClaimValuesClosure blk;
2652 2653 collection_set_iterate(&blk);
2653 2654 }
2654 2655
2655 2656 #ifdef ASSERT
2656 2657 // This checks whether all regions in the heap have the correct claim
2657 2658 // value. I also piggy-backed on this a check to ensure that the
2658 2659 // humongous_start_region() information on "continues humongous"
2659 2660 // regions is correct.
2660 2661
2661 2662 class CheckClaimValuesClosure : public HeapRegionClosure {
2662 2663 private:
2663 2664 jint _claim_value;
2664 2665 uint _failures;
2665 2666 HeapRegion* _sh_region;
2666 2667
2667 2668 public:
2668 2669 CheckClaimValuesClosure(jint claim_value) :
2669 2670 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2670 2671 bool doHeapRegion(HeapRegion* r) {
2671 2672 if (r->claim_value() != _claim_value) {
2672 2673 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2673 2674 "claim value = %d, should be %d",
2674 2675 HR_FORMAT_PARAMS(r),
2675 2676 r->claim_value(), _claim_value);
2676 2677 ++_failures;
2677 2678 }
2678 2679 if (!r->isHumongous()) {
2679 2680 _sh_region = NULL;
2680 2681 } else if (r->startsHumongous()) {
2681 2682 _sh_region = r;
2682 2683 } else if (r->continuesHumongous()) {
2683 2684 if (r->humongous_start_region() != _sh_region) {
2684 2685 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2685 2686 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2686 2687 HR_FORMAT_PARAMS(r),
2687 2688 r->humongous_start_region(),
2688 2689 _sh_region);
2689 2690 ++_failures;
2690 2691 }
2691 2692 }
2692 2693 return false;
2693 2694 }
2694 2695 uint failures() { return _failures; }
2695 2696 };
2696 2697
2697 2698 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2698 2699 CheckClaimValuesClosure cl(claim_value);
2699 2700 heap_region_iterate(&cl);
2700 2701 return cl.failures() == 0;
2701 2702 }
2702 2703
2703 2704 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2704 2705 private:
2705 2706 jint _claim_value;
2706 2707 uint _failures;
2707 2708
2708 2709 public:
2709 2710 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2710 2711 _claim_value(claim_value), _failures(0) { }
2711 2712
2712 2713 uint failures() { return _failures; }
2713 2714
2714 2715 bool doHeapRegion(HeapRegion* hr) {
2715 2716 assert(hr->in_collection_set(), "how?");
2716 2717 assert(!hr->isHumongous(), "H-region in CSet");
2717 2718 if (hr->claim_value() != _claim_value) {
2718 2719 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2719 2720 "claim value = %d, should be %d",
2720 2721 HR_FORMAT_PARAMS(hr),
2721 2722 hr->claim_value(), _claim_value);
2722 2723 _failures += 1;
2723 2724 }
2724 2725 return false;
2725 2726 }
2726 2727 };
2727 2728
2728 2729 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2729 2730 CheckClaimValuesInCSetHRClosure cl(claim_value);
2730 2731 collection_set_iterate(&cl);
2731 2732 return cl.failures() == 0;
2732 2733 }
2733 2734 #endif // ASSERT
2734 2735
2735 2736 // Clear the cached CSet starting regions and (more importantly)
2736 2737 // the time stamps. Called when we reset the GC time stamp.
2737 2738 void G1CollectedHeap::clear_cset_start_regions() {
2738 2739 assert(_worker_cset_start_region != NULL, "sanity");
2739 2740 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2740 2741
2741 2742 int n_queues = MAX2((int)ParallelGCThreads, 1);
2742 2743 for (int i = 0; i < n_queues; i++) {
2743 2744 _worker_cset_start_region[i] = NULL;
2744 2745 _worker_cset_start_region_time_stamp[i] = 0;
2745 2746 }
2746 2747 }
2747 2748
2748 2749 // Given the id of a worker, obtain or calculate a suitable
2749 2750 // starting region for iterating over the current collection set.
2750 2751 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(uint worker_i) {
2751 2752 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2752 2753
2753 2754 HeapRegion* result = NULL;
2754 2755 unsigned gc_time_stamp = get_gc_time_stamp();
2755 2756
2756 2757 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2757 2758 // Cached starting region for current worker was set
2758 2759 // during the current pause - so it's valid.
2759 2760 // Note: the cached starting heap region may be NULL
2760 2761 // (when the collection set is empty).
2761 2762 result = _worker_cset_start_region[worker_i];
2762 2763 assert(result == NULL || result->in_collection_set(), "sanity");
2763 2764 return result;
2764 2765 }
2765 2766
2766 2767 // The cached entry was not valid so let's calculate
2767 2768 // a suitable starting heap region for this worker.
2768 2769
2769 2770 // We want the parallel threads to start their collection
2770 2771 // set iteration at different collection set regions to
2771 2772 // avoid contention.
2772 2773 // If we have:
2773 2774 // n collection set regions
2774 2775 // p threads
2775 2776 // Then thread t will start at region floor ((t * n) / p)
2776 2777
2777 2778 result = g1_policy()->collection_set();
2778 2779 if (G1CollectedHeap::use_parallel_gc_threads()) {
2779 2780 uint cs_size = g1_policy()->cset_region_length();
2780 2781 uint active_workers = workers()->active_workers();
2781 2782 assert(UseDynamicNumberOfGCThreads ||
2782 2783 active_workers == workers()->total_workers(),
2783 2784 "Unless dynamic should use total workers");
2784 2785
2785 2786 uint end_ind = (cs_size * worker_i) / active_workers;
2786 2787 uint start_ind = 0;
2787 2788
2788 2789 if (worker_i > 0 &&
2789 2790 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2790 2791 // Previous workers starting region is valid
2791 2792 // so let's iterate from there
2792 2793 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2793 2794 result = _worker_cset_start_region[worker_i - 1];
2794 2795 }
2795 2796
2796 2797 for (uint i = start_ind; i < end_ind; i++) {
2797 2798 result = result->next_in_collection_set();
2798 2799 }
2799 2800 }
2800 2801
2801 2802 // Note: the calculated starting heap region may be NULL
2802 2803 // (when the collection set is empty).
2803 2804 assert(result == NULL || result->in_collection_set(), "sanity");
2804 2805 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2805 2806 "should be updated only once per pause");
2806 2807 _worker_cset_start_region[worker_i] = result;
2807 2808 OrderAccess::storestore();
2808 2809 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2809 2810 return result;
2810 2811 }
2811 2812
2812 2813 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2813 2814 HeapRegion* r = g1_policy()->collection_set();
2814 2815 while (r != NULL) {
2815 2816 HeapRegion* next = r->next_in_collection_set();
2816 2817 if (cl->doHeapRegion(r)) {
2817 2818 cl->incomplete();
2818 2819 return;
2819 2820 }
2820 2821 r = next;
2821 2822 }
2822 2823 }
2823 2824
2824 2825 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2825 2826 HeapRegionClosure *cl) {
2826 2827 if (r == NULL) {
2827 2828 // The CSet is empty so there's nothing to do.
2828 2829 return;
2829 2830 }
2830 2831
2831 2832 assert(r->in_collection_set(),
2832 2833 "Start region must be a member of the collection set.");
2833 2834 HeapRegion* cur = r;
2834 2835 while (cur != NULL) {
2835 2836 HeapRegion* next = cur->next_in_collection_set();
2836 2837 if (cl->doHeapRegion(cur) && false) {
2837 2838 cl->incomplete();
2838 2839 return;
2839 2840 }
2840 2841 cur = next;
2841 2842 }
2842 2843 cur = g1_policy()->collection_set();
2843 2844 while (cur != r) {
2844 2845 HeapRegion* next = cur->next_in_collection_set();
2845 2846 if (cl->doHeapRegion(cur) && false) {
2846 2847 cl->incomplete();
2847 2848 return;
2848 2849 }
2849 2850 cur = next;
2850 2851 }
2851 2852 }
2852 2853
2853 2854 HeapRegion* G1CollectedHeap::next_compaction_region(const HeapRegion* from) const {
2854 2855 HeapRegion* result = _hrm.next_region_in_heap(from);
2855 2856 while (result != NULL && result->isHumongous()) {
2856 2857 result = _hrm.next_region_in_heap(result);
2857 2858 }
2858 2859 return result;
2859 2860 }
2860 2861
2861 2862 Space* G1CollectedHeap::space_containing(const void* addr) const {
2862 2863 return heap_region_containing(addr);
2863 2864 }
2864 2865
2865 2866 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2866 2867 Space* sp = space_containing(addr);
2867 2868 return sp->block_start(addr);
2868 2869 }
2869 2870
2870 2871 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2871 2872 Space* sp = space_containing(addr);
2872 2873 return sp->block_size(addr);
2873 2874 }
2874 2875
2875 2876 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2876 2877 Space* sp = space_containing(addr);
2877 2878 return sp->block_is_obj(addr);
2878 2879 }
2879 2880
2880 2881 bool G1CollectedHeap::supports_tlab_allocation() const {
2881 2882 return true;
2882 2883 }
2883 2884
2884 2885 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2885 2886 return (_g1_policy->young_list_target_length() - young_list()->survivor_length()) * HeapRegion::GrainBytes;
2886 2887 }
2887 2888
2888 2889 size_t G1CollectedHeap::tlab_used(Thread* ignored) const {
2889 2890 return young_list()->eden_used_bytes();
2890 2891 }
2891 2892
2892 2893 // For G1 TLABs should not contain humongous objects, so the maximum TLAB size
2893 2894 // must be smaller than the humongous object limit.
2894 2895 size_t G1CollectedHeap::max_tlab_size() const {
2895 2896 return align_size_down(_humongous_object_threshold_in_words - 1, MinObjAlignment);
2896 2897 }
2897 2898
2898 2899 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2899 2900 // Return the remaining space in the cur alloc region, but not less than
2900 2901 // the min TLAB size.
2901 2902
2902 2903 // Also, this value can be at most the humongous object threshold,
2903 2904 // since we can't allow tlabs to grow big enough to accommodate
2904 2905 // humongous objects.
2905 2906
2906 2907 HeapRegion* hr = _allocator->mutator_alloc_region(AllocationContext::current())->get();
2907 2908 size_t max_tlab = max_tlab_size() * wordSize;
2908 2909 if (hr == NULL) {
2909 2910 return max_tlab;
2910 2911 } else {
2911 2912 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab);
2912 2913 }
2913 2914 }
2914 2915
2915 2916 size_t G1CollectedHeap::max_capacity() const {
2916 2917 return _hrm.reserved().byte_size();
2917 2918 }
2918 2919
2919 2920 jlong G1CollectedHeap::millis_since_last_gc() {
2920 2921 // assert(false, "NYI");
2921 2922 return 0;
2922 2923 }
2923 2924
2924 2925 void G1CollectedHeap::prepare_for_verify() {
2925 2926 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2926 2927 ensure_parsability(false);
2927 2928 }
2928 2929 g1_rem_set()->prepare_for_verify();
2929 2930 }
2930 2931
2931 2932 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
2932 2933 VerifyOption vo) {
2933 2934 switch (vo) {
2934 2935 case VerifyOption_G1UsePrevMarking:
2935 2936 return hr->obj_allocated_since_prev_marking(obj);
2936 2937 case VerifyOption_G1UseNextMarking:
2937 2938 return hr->obj_allocated_since_next_marking(obj);
2938 2939 case VerifyOption_G1UseMarkWord:
2939 2940 return false;
2940 2941 default:
2941 2942 ShouldNotReachHere();
2942 2943 }
2943 2944 return false; // keep some compilers happy
2944 2945 }
2945 2946
2946 2947 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
2947 2948 switch (vo) {
2948 2949 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
2949 2950 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
2950 2951 case VerifyOption_G1UseMarkWord: return NULL;
2951 2952 default: ShouldNotReachHere();
2952 2953 }
2953 2954 return NULL; // keep some compilers happy
2954 2955 }
2955 2956
2956 2957 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
2957 2958 switch (vo) {
2958 2959 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
2959 2960 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
2960 2961 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
2961 2962 default: ShouldNotReachHere();
2962 2963 }
2963 2964 return false; // keep some compilers happy
2964 2965 }
2965 2966
2966 2967 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
2967 2968 switch (vo) {
2968 2969 case VerifyOption_G1UsePrevMarking: return "PTAMS";
2969 2970 case VerifyOption_G1UseNextMarking: return "NTAMS";
2970 2971 case VerifyOption_G1UseMarkWord: return "NONE";
2971 2972 default: ShouldNotReachHere();
2972 2973 }
2973 2974 return NULL; // keep some compilers happy
2974 2975 }
2975 2976
2976 2977 class VerifyRootsClosure: public OopClosure {
2977 2978 private:
2978 2979 G1CollectedHeap* _g1h;
2979 2980 VerifyOption _vo;
2980 2981 bool _failures;
2981 2982 public:
2982 2983 // _vo == UsePrevMarking -> use "prev" marking information,
2983 2984 // _vo == UseNextMarking -> use "next" marking information,
2984 2985 // _vo == UseMarkWord -> use mark word from object header.
2985 2986 VerifyRootsClosure(VerifyOption vo) :
2986 2987 _g1h(G1CollectedHeap::heap()),
2987 2988 _vo(vo),
2988 2989 _failures(false) { }
2989 2990
2990 2991 bool failures() { return _failures; }
2991 2992
2992 2993 template <class T> void do_oop_nv(T* p) {
2993 2994 T heap_oop = oopDesc::load_heap_oop(p);
2994 2995 if (!oopDesc::is_null(heap_oop)) {
2995 2996 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
2996 2997 if (_g1h->is_obj_dead_cond(obj, _vo)) {
2997 2998 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
2998 2999 "points to dead obj "PTR_FORMAT, p, (void*) obj);
2999 3000 if (_vo == VerifyOption_G1UseMarkWord) {
3000 3001 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3001 3002 }
3002 3003 obj->print_on(gclog_or_tty);
3003 3004 _failures = true;
3004 3005 }
3005 3006 }
3006 3007 }
3007 3008
3008 3009 void do_oop(oop* p) { do_oop_nv(p); }
3009 3010 void do_oop(narrowOop* p) { do_oop_nv(p); }
3010 3011 };
3011 3012
3012 3013 class G1VerifyCodeRootOopClosure: public OopClosure {
3013 3014 G1CollectedHeap* _g1h;
3014 3015 OopClosure* _root_cl;
3015 3016 nmethod* _nm;
3016 3017 VerifyOption _vo;
3017 3018 bool _failures;
3018 3019
3019 3020 template <class T> void do_oop_work(T* p) {
3020 3021 // First verify that this root is live
3021 3022 _root_cl->do_oop(p);
3022 3023
3023 3024 if (!G1VerifyHeapRegionCodeRoots) {
3024 3025 // We're not verifying the code roots attached to heap region.
3025 3026 return;
3026 3027 }
3027 3028
3028 3029 // Don't check the code roots during marking verification in a full GC
3029 3030 if (_vo == VerifyOption_G1UseMarkWord) {
3030 3031 return;
3031 3032 }
3032 3033
3033 3034 // Now verify that the current nmethod (which contains p) is
3034 3035 // in the code root list of the heap region containing the
3035 3036 // object referenced by p.
3036 3037
3037 3038 T heap_oop = oopDesc::load_heap_oop(p);
3038 3039 if (!oopDesc::is_null(heap_oop)) {
3039 3040 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3040 3041
3041 3042 // Now fetch the region containing the object
3042 3043 HeapRegion* hr = _g1h->heap_region_containing(obj);
3043 3044 HeapRegionRemSet* hrrs = hr->rem_set();
3044 3045 // Verify that the strong code root list for this region
3045 3046 // contains the nmethod
3046 3047 if (!hrrs->strong_code_roots_list_contains(_nm)) {
3047 3048 gclog_or_tty->print_cr("Code root location "PTR_FORMAT" "
3048 3049 "from nmethod "PTR_FORMAT" not in strong "
3049 3050 "code roots for region ["PTR_FORMAT","PTR_FORMAT")",
3050 3051 p, _nm, hr->bottom(), hr->end());
3051 3052 _failures = true;
3052 3053 }
3053 3054 }
3054 3055 }
3055 3056
3056 3057 public:
3057 3058 G1VerifyCodeRootOopClosure(G1CollectedHeap* g1h, OopClosure* root_cl, VerifyOption vo):
3058 3059 _g1h(g1h), _root_cl(root_cl), _vo(vo), _nm(NULL), _failures(false) {}
3059 3060
3060 3061 void do_oop(oop* p) { do_oop_work(p); }
3061 3062 void do_oop(narrowOop* p) { do_oop_work(p); }
3062 3063
3063 3064 void set_nmethod(nmethod* nm) { _nm = nm; }
3064 3065 bool failures() { return _failures; }
3065 3066 };
3066 3067
3067 3068 class G1VerifyCodeRootBlobClosure: public CodeBlobClosure {
3068 3069 G1VerifyCodeRootOopClosure* _oop_cl;
3069 3070
3070 3071 public:
3071 3072 G1VerifyCodeRootBlobClosure(G1VerifyCodeRootOopClosure* oop_cl):
3072 3073 _oop_cl(oop_cl) {}
3073 3074
3074 3075 void do_code_blob(CodeBlob* cb) {
3075 3076 nmethod* nm = cb->as_nmethod_or_null();
3076 3077 if (nm != NULL) {
3077 3078 _oop_cl->set_nmethod(nm);
3078 3079 nm->oops_do(_oop_cl);
3079 3080 }
3080 3081 }
3081 3082 };
3082 3083
3083 3084 class YoungRefCounterClosure : public OopClosure {
3084 3085 G1CollectedHeap* _g1h;
3085 3086 int _count;
3086 3087 public:
3087 3088 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3088 3089 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3089 3090 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3090 3091
3091 3092 int count() { return _count; }
3092 3093 void reset_count() { _count = 0; };
3093 3094 };
3094 3095
3095 3096 class VerifyKlassClosure: public KlassClosure {
3096 3097 YoungRefCounterClosure _young_ref_counter_closure;
3097 3098 OopClosure *_oop_closure;
3098 3099 public:
3099 3100 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3100 3101 void do_klass(Klass* k) {
3101 3102 k->oops_do(_oop_closure);
3102 3103
3103 3104 _young_ref_counter_closure.reset_count();
3104 3105 k->oops_do(&_young_ref_counter_closure);
3105 3106 if (_young_ref_counter_closure.count() > 0) {
3106 3107 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3107 3108 }
3108 3109 }
3109 3110 };
3110 3111
3111 3112 class VerifyLivenessOopClosure: public OopClosure {
3112 3113 G1CollectedHeap* _g1h;
3113 3114 VerifyOption _vo;
3114 3115 public:
3115 3116 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3116 3117 _g1h(g1h), _vo(vo)
3117 3118 { }
3118 3119 void do_oop(narrowOop *p) { do_oop_work(p); }
3119 3120 void do_oop( oop *p) { do_oop_work(p); }
3120 3121
3121 3122 template <class T> void do_oop_work(T *p) {
3122 3123 oop obj = oopDesc::load_decode_heap_oop(p);
3123 3124 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3124 3125 "Dead object referenced by a not dead object");
3125 3126 }
3126 3127 };
3127 3128
3128 3129 class VerifyObjsInRegionClosure: public ObjectClosure {
3129 3130 private:
3130 3131 G1CollectedHeap* _g1h;
3131 3132 size_t _live_bytes;
3132 3133 HeapRegion *_hr;
3133 3134 VerifyOption _vo;
3134 3135 public:
3135 3136 // _vo == UsePrevMarking -> use "prev" marking information,
3136 3137 // _vo == UseNextMarking -> use "next" marking information,
3137 3138 // _vo == UseMarkWord -> use mark word from object header.
3138 3139 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3139 3140 : _live_bytes(0), _hr(hr), _vo(vo) {
3140 3141 _g1h = G1CollectedHeap::heap();
3141 3142 }
3142 3143 void do_object(oop o) {
3143 3144 VerifyLivenessOopClosure isLive(_g1h, _vo);
3144 3145 assert(o != NULL, "Huh?");
3145 3146 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3146 3147 // If the object is alive according to the mark word,
3147 3148 // then verify that the marking information agrees.
3148 3149 // Note we can't verify the contra-positive of the
3149 3150 // above: if the object is dead (according to the mark
3150 3151 // word), it may not be marked, or may have been marked
3151 3152 // but has since became dead, or may have been allocated
3152 3153 // since the last marking.
3153 3154 if (_vo == VerifyOption_G1UseMarkWord) {
3154 3155 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3155 3156 }
3156 3157
3157 3158 o->oop_iterate_no_header(&isLive);
3158 3159 if (!_hr->obj_allocated_since_prev_marking(o)) {
3159 3160 size_t obj_size = o->size(); // Make sure we don't overflow
3160 3161 _live_bytes += (obj_size * HeapWordSize);
3161 3162 }
3162 3163 }
3163 3164 }
3164 3165 size_t live_bytes() { return _live_bytes; }
3165 3166 };
3166 3167
3167 3168 class PrintObjsInRegionClosure : public ObjectClosure {
3168 3169 HeapRegion *_hr;
3169 3170 G1CollectedHeap *_g1;
3170 3171 public:
3171 3172 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3172 3173 _g1 = G1CollectedHeap::heap();
3173 3174 };
3174 3175
3175 3176 void do_object(oop o) {
3176 3177 if (o != NULL) {
3177 3178 HeapWord *start = (HeapWord *) o;
3178 3179 size_t word_sz = o->size();
3179 3180 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3180 3181 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3181 3182 (void*) o, word_sz,
3182 3183 _g1->isMarkedPrev(o),
3183 3184 _g1->isMarkedNext(o),
3184 3185 _hr->obj_allocated_since_prev_marking(o));
3185 3186 HeapWord *end = start + word_sz;
3186 3187 HeapWord *cur;
3187 3188 int *val;
3188 3189 for (cur = start; cur < end; cur++) {
3189 3190 val = (int *) cur;
3190 3191 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3191 3192 }
3192 3193 }
3193 3194 }
3194 3195 };
3195 3196
3196 3197 class VerifyRegionClosure: public HeapRegionClosure {
3197 3198 private:
3198 3199 bool _par;
3199 3200 VerifyOption _vo;
3200 3201 bool _failures;
3201 3202 public:
3202 3203 // _vo == UsePrevMarking -> use "prev" marking information,
3203 3204 // _vo == UseNextMarking -> use "next" marking information,
3204 3205 // _vo == UseMarkWord -> use mark word from object header.
3205 3206 VerifyRegionClosure(bool par, VerifyOption vo)
3206 3207 : _par(par),
3207 3208 _vo(vo),
3208 3209 _failures(false) {}
3209 3210
3210 3211 bool failures() {
3211 3212 return _failures;
3212 3213 }
3213 3214
3214 3215 bool doHeapRegion(HeapRegion* r) {
3215 3216 if (!r->continuesHumongous()) {
3216 3217 bool failures = false;
3217 3218 r->verify(_vo, &failures);
3218 3219 if (failures) {
3219 3220 _failures = true;
3220 3221 } else {
3221 3222 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3222 3223 r->object_iterate(¬_dead_yet_cl);
3223 3224 if (_vo != VerifyOption_G1UseNextMarking) {
3224 3225 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3225 3226 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3226 3227 "max_live_bytes "SIZE_FORMAT" "
3227 3228 "< calculated "SIZE_FORMAT,
3228 3229 r->bottom(), r->end(),
3229 3230 r->max_live_bytes(),
3230 3231 not_dead_yet_cl.live_bytes());
3231 3232 _failures = true;
3232 3233 }
3233 3234 } else {
3234 3235 // When vo == UseNextMarking we cannot currently do a sanity
3235 3236 // check on the live bytes as the calculation has not been
3236 3237 // finalized yet.
3237 3238 }
3238 3239 }
3239 3240 }
3240 3241 return false; // stop the region iteration if we hit a failure
3241 3242 }
3242 3243 };
3243 3244
3244 3245 // This is the task used for parallel verification of the heap regions
3245 3246
3246 3247 class G1ParVerifyTask: public AbstractGangTask {
3247 3248 private:
3248 3249 G1CollectedHeap* _g1h;
3249 3250 VerifyOption _vo;
3250 3251 bool _failures;
3251 3252
3252 3253 public:
3253 3254 // _vo == UsePrevMarking -> use "prev" marking information,
3254 3255 // _vo == UseNextMarking -> use "next" marking information,
3255 3256 // _vo == UseMarkWord -> use mark word from object header.
3256 3257 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3257 3258 AbstractGangTask("Parallel verify task"),
3258 3259 _g1h(g1h),
3259 3260 _vo(vo),
3260 3261 _failures(false) { }
3261 3262
3262 3263 bool failures() {
3263 3264 return _failures;
3264 3265 }
3265 3266
3266 3267 void work(uint worker_id) {
3267 3268 HandleMark hm;
3268 3269 VerifyRegionClosure blk(true, _vo);
3269 3270 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3270 3271 _g1h->workers()->active_workers(),
3271 3272 HeapRegion::ParVerifyClaimValue);
3272 3273 if (blk.failures()) {
3273 3274 _failures = true;
3274 3275 }
3275 3276 }
3276 3277 };
3277 3278
3278 3279 void G1CollectedHeap::verify(bool silent, VerifyOption vo) {
3279 3280 if (SafepointSynchronize::is_at_safepoint()) {
3280 3281 assert(Thread::current()->is_VM_thread(),
3281 3282 "Expected to be executed serially by the VM thread at this point");
3282 3283
3283 3284 if (!silent) { gclog_or_tty->print("Roots "); }
3284 3285 VerifyRootsClosure rootsCl(vo);
3285 3286 VerifyKlassClosure klassCl(this, &rootsCl);
3286 3287 CLDToKlassAndOopClosure cldCl(&klassCl, &rootsCl, false);
3287 3288
3288 3289 // We apply the relevant closures to all the oops in the
3289 3290 // system dictionary, class loader data graph, the string table
3290 3291 // and the nmethods in the code cache.
3291 3292 G1VerifyCodeRootOopClosure codeRootsCl(this, &rootsCl, vo);
3292 3293 G1VerifyCodeRootBlobClosure blobsCl(&codeRootsCl);
3293 3294
3294 3295 process_all_roots(true, // activate StrongRootsScope
3295 3296 SO_AllCodeCache, // roots scanning options
3296 3297 &rootsCl,
3297 3298 &cldCl,
3298 3299 &blobsCl);
3299 3300
3300 3301 bool failures = rootsCl.failures() || codeRootsCl.failures();
3301 3302
3302 3303 if (vo != VerifyOption_G1UseMarkWord) {
3303 3304 // If we're verifying during a full GC then the region sets
3304 3305 // will have been torn down at the start of the GC. Therefore
3305 3306 // verifying the region sets will fail. So we only verify
3306 3307 // the region sets when not in a full GC.
3307 3308 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3308 3309 verify_region_sets();
3309 3310 }
3310 3311
3311 3312 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3312 3313 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3313 3314 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3314 3315 "sanity check");
3315 3316
3316 3317 G1ParVerifyTask task(this, vo);
3317 3318 assert(UseDynamicNumberOfGCThreads ||
3318 3319 workers()->active_workers() == workers()->total_workers(),
3319 3320 "If not dynamic should be using all the workers");
3320 3321 int n_workers = workers()->active_workers();
3321 3322 set_par_threads(n_workers);
3322 3323 workers()->run_task(&task);
3323 3324 set_par_threads(0);
3324 3325 if (task.failures()) {
3325 3326 failures = true;
3326 3327 }
3327 3328
3328 3329 // Checks that the expected amount of parallel work was done.
3329 3330 // The implication is that n_workers is > 0.
3330 3331 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3331 3332 "sanity check");
3332 3333
3333 3334 reset_heap_region_claim_values();
3334 3335
3335 3336 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3336 3337 "sanity check");
3337 3338 } else {
3338 3339 VerifyRegionClosure blk(false, vo);
3339 3340 heap_region_iterate(&blk);
3340 3341 if (blk.failures()) {
3341 3342 failures = true;
3342 3343 }
3343 3344 }
3344 3345 if (!silent) gclog_or_tty->print("RemSet ");
3345 3346 rem_set()->verify();
3346 3347
3347 3348 if (G1StringDedup::is_enabled()) {
3348 3349 if (!silent) gclog_or_tty->print("StrDedup ");
3349 3350 G1StringDedup::verify();
3350 3351 }
3351 3352
3352 3353 if (failures) {
3353 3354 gclog_or_tty->print_cr("Heap:");
3354 3355 // It helps to have the per-region information in the output to
3355 3356 // help us track down what went wrong. This is why we call
3356 3357 // print_extended_on() instead of print_on().
3357 3358 print_extended_on(gclog_or_tty);
3358 3359 gclog_or_tty->cr();
3359 3360 #ifndef PRODUCT
3360 3361 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3361 3362 concurrent_mark()->print_reachable("at-verification-failure",
3362 3363 vo, false /* all */);
3363 3364 }
3364 3365 #endif
3365 3366 gclog_or_tty->flush();
3366 3367 }
3367 3368 guarantee(!failures, "there should not have been any failures");
3368 3369 } else {
3369 3370 if (!silent) {
3370 3371 gclog_or_tty->print("(SKIPPING Roots, HeapRegionSets, HeapRegions, RemSet");
3371 3372 if (G1StringDedup::is_enabled()) {
3372 3373 gclog_or_tty->print(", StrDedup");
3373 3374 }
3374 3375 gclog_or_tty->print(") ");
3375 3376 }
3376 3377 }
3377 3378 }
3378 3379
3379 3380 void G1CollectedHeap::verify(bool silent) {
3380 3381 verify(silent, VerifyOption_G1UsePrevMarking);
3381 3382 }
3382 3383
3383 3384 double G1CollectedHeap::verify(bool guard, const char* msg) {
3384 3385 double verify_time_ms = 0.0;
3385 3386
3386 3387 if (guard && total_collections() >= VerifyGCStartAt) {
3387 3388 double verify_start = os::elapsedTime();
3388 3389 HandleMark hm; // Discard invalid handles created during verification
3389 3390 prepare_for_verify();
3390 3391 Universe::verify(VerifyOption_G1UsePrevMarking, msg);
3391 3392 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
3392 3393 }
3393 3394
3394 3395 return verify_time_ms;
3395 3396 }
3396 3397
3397 3398 void G1CollectedHeap::verify_before_gc() {
3398 3399 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
3399 3400 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
3400 3401 }
3401 3402
3402 3403 void G1CollectedHeap::verify_after_gc() {
3403 3404 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
3404 3405 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
3405 3406 }
3406 3407
3407 3408 class PrintRegionClosure: public HeapRegionClosure {
3408 3409 outputStream* _st;
3409 3410 public:
3410 3411 PrintRegionClosure(outputStream* st) : _st(st) {}
3411 3412 bool doHeapRegion(HeapRegion* r) {
3412 3413 r->print_on(_st);
3413 3414 return false;
3414 3415 }
3415 3416 };
3416 3417
3417 3418 bool G1CollectedHeap::is_obj_dead_cond(const oop obj,
3418 3419 const HeapRegion* hr,
3419 3420 const VerifyOption vo) const {
3420 3421 switch (vo) {
3421 3422 case VerifyOption_G1UsePrevMarking: return is_obj_dead(obj, hr);
3422 3423 case VerifyOption_G1UseNextMarking: return is_obj_ill(obj, hr);
3423 3424 case VerifyOption_G1UseMarkWord: return !obj->is_gc_marked();
3424 3425 default: ShouldNotReachHere();
3425 3426 }
3426 3427 return false; // keep some compilers happy
3427 3428 }
3428 3429
3429 3430 bool G1CollectedHeap::is_obj_dead_cond(const oop obj,
3430 3431 const VerifyOption vo) const {
3431 3432 switch (vo) {
3432 3433 case VerifyOption_G1UsePrevMarking: return is_obj_dead(obj);
3433 3434 case VerifyOption_G1UseNextMarking: return is_obj_ill(obj);
3434 3435 case VerifyOption_G1UseMarkWord: return !obj->is_gc_marked();
3435 3436 default: ShouldNotReachHere();
3436 3437 }
3437 3438 return false; // keep some compilers happy
3438 3439 }
3439 3440
3440 3441 void G1CollectedHeap::print_on(outputStream* st) const {
3441 3442 st->print(" %-20s", "garbage-first heap");
3442 3443 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3443 3444 capacity()/K, used_unlocked()/K);
3444 3445 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3445 3446 _hrm.reserved().start(),
3446 3447 _hrm.reserved().start() + _hrm.length() + HeapRegion::GrainWords,
3447 3448 _hrm.reserved().end());
3448 3449 st->cr();
3449 3450 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3450 3451 uint young_regions = _young_list->length();
3451 3452 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3452 3453 (size_t) young_regions * HeapRegion::GrainBytes / K);
3453 3454 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3454 3455 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3455 3456 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3456 3457 st->cr();
3457 3458 MetaspaceAux::print_on(st);
3458 3459 }
3459 3460
3460 3461 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3461 3462 print_on(st);
3462 3463
3463 3464 // Print the per-region information.
3464 3465 st->cr();
3465 3466 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3466 3467 "HS=humongous(starts), HC=humongous(continues), "
3467 3468 "CS=collection set, F=free, TS=gc time stamp, "
3468 3469 "PTAMS=previous top-at-mark-start, "
3469 3470 "NTAMS=next top-at-mark-start)");
3470 3471 PrintRegionClosure blk(st);
3471 3472 heap_region_iterate(&blk);
3472 3473 }
3473 3474
3474 3475 void G1CollectedHeap::print_on_error(outputStream* st) const {
3475 3476 this->CollectedHeap::print_on_error(st);
3476 3477
3477 3478 if (_cm != NULL) {
3478 3479 st->cr();
3479 3480 _cm->print_on_error(st);
3480 3481 }
3481 3482 }
3482 3483
3483 3484 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3484 3485 if (G1CollectedHeap::use_parallel_gc_threads()) {
3485 3486 workers()->print_worker_threads_on(st);
3486 3487 }
3487 3488 _cmThread->print_on(st);
3488 3489 st->cr();
3489 3490 _cm->print_worker_threads_on(st);
3490 3491 _cg1r->print_worker_threads_on(st);
3491 3492 if (G1StringDedup::is_enabled()) {
3492 3493 G1StringDedup::print_worker_threads_on(st);
3493 3494 }
3494 3495 }
3495 3496
3496 3497 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3497 3498 if (G1CollectedHeap::use_parallel_gc_threads()) {
3498 3499 workers()->threads_do(tc);
3499 3500 }
3500 3501 tc->do_thread(_cmThread);
3501 3502 _cg1r->threads_do(tc);
3502 3503 if (G1StringDedup::is_enabled()) {
3503 3504 G1StringDedup::threads_do(tc);
3504 3505 }
3505 3506 }
3506 3507
3507 3508 void G1CollectedHeap::print_tracing_info() const {
3508 3509 // We'll overload this to mean "trace GC pause statistics."
3509 3510 if (TraceGen0Time || TraceGen1Time) {
3510 3511 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3511 3512 // to that.
3512 3513 g1_policy()->print_tracing_info();
3513 3514 }
3514 3515 if (G1SummarizeRSetStats) {
3515 3516 g1_rem_set()->print_summary_info();
3516 3517 }
3517 3518 if (G1SummarizeConcMark) {
3518 3519 concurrent_mark()->print_summary_info();
3519 3520 }
3520 3521 g1_policy()->print_yg_surv_rate_info();
3521 3522 SpecializationStats::print();
3522 3523 }
3523 3524
3524 3525 #ifndef PRODUCT
3525 3526 // Helpful for debugging RSet issues.
3526 3527
3527 3528 class PrintRSetsClosure : public HeapRegionClosure {
3528 3529 private:
3529 3530 const char* _msg;
3530 3531 size_t _occupied_sum;
3531 3532
3532 3533 public:
3533 3534 bool doHeapRegion(HeapRegion* r) {
3534 3535 HeapRegionRemSet* hrrs = r->rem_set();
3535 3536 size_t occupied = hrrs->occupied();
3536 3537 _occupied_sum += occupied;
3537 3538
3538 3539 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3539 3540 HR_FORMAT_PARAMS(r));
3540 3541 if (occupied == 0) {
3541 3542 gclog_or_tty->print_cr(" RSet is empty");
3542 3543 } else {
3543 3544 hrrs->print();
3544 3545 }
3545 3546 gclog_or_tty->print_cr("----------");
3546 3547 return false;
3547 3548 }
3548 3549
3549 3550 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3550 3551 gclog_or_tty->cr();
3551 3552 gclog_or_tty->print_cr("========================================");
3552 3553 gclog_or_tty->print_cr("%s", msg);
3553 3554 gclog_or_tty->cr();
3554 3555 }
3555 3556
3556 3557 ~PrintRSetsClosure() {
3557 3558 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3558 3559 gclog_or_tty->print_cr("========================================");
3559 3560 gclog_or_tty->cr();
3560 3561 }
3561 3562 };
3562 3563
3563 3564 void G1CollectedHeap::print_cset_rsets() {
3564 3565 PrintRSetsClosure cl("Printing CSet RSets");
3565 3566 collection_set_iterate(&cl);
3566 3567 }
3567 3568
3568 3569 void G1CollectedHeap::print_all_rsets() {
3569 3570 PrintRSetsClosure cl("Printing All RSets");;
3570 3571 heap_region_iterate(&cl);
3571 3572 }
3572 3573 #endif // PRODUCT
3573 3574
3574 3575 G1CollectedHeap* G1CollectedHeap::heap() {
3575 3576 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3576 3577 "not a garbage-first heap");
3577 3578 return _g1h;
3578 3579 }
3579 3580
3580 3581 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3581 3582 // always_do_update_barrier = false;
3582 3583 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3583 3584 // Fill TLAB's and such
3584 3585 accumulate_statistics_all_tlabs();
3585 3586 ensure_parsability(true);
3586 3587
3587 3588 if (G1SummarizeRSetStats && (G1SummarizeRSetStatsPeriod > 0) &&
3588 3589 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
3589 3590 g1_rem_set()->print_periodic_summary_info("Before GC RS summary");
3590 3591 }
3591 3592 }
3592 3593
3593 3594 void G1CollectedHeap::gc_epilogue(bool full) {
3594 3595
3595 3596 if (G1SummarizeRSetStats &&
3596 3597 (G1SummarizeRSetStatsPeriod > 0) &&
3597 3598 // we are at the end of the GC. Total collections has already been increased.
3598 3599 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) {
3599 3600 g1_rem_set()->print_periodic_summary_info("After GC RS summary");
3600 3601 }
3601 3602
3602 3603 // FIXME: what is this about?
3603 3604 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3604 3605 // is set.
3605 3606 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3606 3607 "derived pointer present"));
3607 3608 // always_do_update_barrier = true;
3608 3609
3609 3610 resize_all_tlabs();
3610 3611 allocation_context_stats().update(full);
3611 3612
3612 3613 // We have just completed a GC. Update the soft reference
3613 3614 // policy with the new heap occupancy
3614 3615 Universe::update_heap_info_at_gc();
3615 3616 }
3616 3617
3617 3618 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3618 3619 unsigned int gc_count_before,
3619 3620 bool* succeeded,
3620 3621 GCCause::Cause gc_cause) {
3621 3622 assert_heap_not_locked_and_not_at_safepoint();
3622 3623 g1_policy()->record_stop_world_start();
3623 3624 VM_G1IncCollectionPause op(gc_count_before,
3624 3625 word_size,
3625 3626 false, /* should_initiate_conc_mark */
3626 3627 g1_policy()->max_pause_time_ms(),
3627 3628 gc_cause);
3628 3629
3629 3630 op.set_allocation_context(AllocationContext::current());
3630 3631 VMThread::execute(&op);
3631 3632
3632 3633 HeapWord* result = op.result();
3633 3634 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3634 3635 assert(result == NULL || ret_succeeded,
3635 3636 "the result should be NULL if the VM did not succeed");
3636 3637 *succeeded = ret_succeeded;
3637 3638
3638 3639 assert_heap_not_locked();
3639 3640 return result;
3640 3641 }
3641 3642
3642 3643 void
3643 3644 G1CollectedHeap::doConcurrentMark() {
3644 3645 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3645 3646 if (!_cmThread->in_progress()) {
3646 3647 _cmThread->set_started();
3647 3648 CGC_lock->notify();
3648 3649 }
3649 3650 }
3650 3651
3651 3652 size_t G1CollectedHeap::pending_card_num() {
3652 3653 size_t extra_cards = 0;
3653 3654 JavaThread *curr = Threads::first();
3654 3655 while (curr != NULL) {
3655 3656 DirtyCardQueue& dcq = curr->dirty_card_queue();
3656 3657 extra_cards += dcq.size();
3657 3658 curr = curr->next();
3658 3659 }
3659 3660 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3660 3661 size_t buffer_size = dcqs.buffer_size();
3661 3662 size_t buffer_num = dcqs.completed_buffers_num();
3662 3663
3663 3664 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3664 3665 // in bytes - not the number of 'entries'. We need to convert
3665 3666 // into a number of cards.
3666 3667 return (buffer_size * buffer_num + extra_cards) / oopSize;
3667 3668 }
3668 3669
3669 3670 size_t G1CollectedHeap::cards_scanned() {
3670 3671 return g1_rem_set()->cardsScanned();
3671 3672 }
3672 3673
3673 3674 bool G1CollectedHeap::humongous_region_is_always_live(uint index) {
3674 3675 HeapRegion* region = region_at(index);
3675 3676 assert(region->startsHumongous(), "Must start a humongous object");
3676 3677 return oop(region->bottom())->is_objArray() || !region->rem_set()->is_empty();
3677 3678 }
3678 3679
3679 3680 class RegisterHumongousWithInCSetFastTestClosure : public HeapRegionClosure {
3680 3681 private:
3681 3682 size_t _total_humongous;
3682 3683 size_t _candidate_humongous;
3683 3684 public:
3684 3685 RegisterHumongousWithInCSetFastTestClosure() : _total_humongous(0), _candidate_humongous(0) {
3685 3686 }
3686 3687
3687 3688 virtual bool doHeapRegion(HeapRegion* r) {
3688 3689 if (!r->startsHumongous()) {
3689 3690 return false;
3690 3691 }
3691 3692 G1CollectedHeap* g1h = G1CollectedHeap::heap();
3692 3693
3693 3694 uint region_idx = r->hrm_index();
3694 3695 bool is_candidate = !g1h->humongous_region_is_always_live(region_idx);
3695 3696 // Is_candidate already filters out humongous regions with some remembered set.
3696 3697 // This will not lead to humongous object that we mistakenly keep alive because
3697 3698 // during young collection the remembered sets will only be added to.
3698 3699 if (is_candidate) {
3699 3700 g1h->register_humongous_region_with_in_cset_fast_test(region_idx);
3700 3701 _candidate_humongous++;
3701 3702 }
3702 3703 _total_humongous++;
3703 3704
3704 3705 return false;
3705 3706 }
3706 3707
3707 3708 size_t total_humongous() const { return _total_humongous; }
3708 3709 size_t candidate_humongous() const { return _candidate_humongous; }
3709 3710 };
3710 3711
3711 3712 void G1CollectedHeap::register_humongous_regions_with_in_cset_fast_test() {
3712 3713 if (!G1ReclaimDeadHumongousObjectsAtYoungGC) {
3713 3714 g1_policy()->phase_times()->record_fast_reclaim_humongous_stats(0, 0);
3714 3715 return;
3715 3716 }
3716 3717
3717 3718 RegisterHumongousWithInCSetFastTestClosure cl;
3718 3719 heap_region_iterate(&cl);
3719 3720 g1_policy()->phase_times()->record_fast_reclaim_humongous_stats(cl.total_humongous(),
3720 3721 cl.candidate_humongous());
3721 3722 _has_humongous_reclaim_candidates = cl.candidate_humongous() > 0;
3722 3723
3723 3724 if (_has_humongous_reclaim_candidates) {
3724 3725 clear_humongous_is_live_table();
3725 3726 }
3726 3727 }
3727 3728
3728 3729 void
3729 3730 G1CollectedHeap::setup_surviving_young_words() {
3730 3731 assert(_surviving_young_words == NULL, "pre-condition");
3731 3732 uint array_length = g1_policy()->young_cset_region_length();
3732 3733 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3733 3734 if (_surviving_young_words == NULL) {
3734 3735 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR,
3735 3736 "Not enough space for young surv words summary.");
3736 3737 }
3737 3738 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3738 3739 #ifdef ASSERT
3739 3740 for (uint i = 0; i < array_length; ++i) {
3740 3741 assert( _surviving_young_words[i] == 0, "memset above" );
3741 3742 }
3742 3743 #endif // !ASSERT
3743 3744 }
3744 3745
3745 3746 void
3746 3747 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3747 3748 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3748 3749 uint array_length = g1_policy()->young_cset_region_length();
3749 3750 for (uint i = 0; i < array_length; ++i) {
3750 3751 _surviving_young_words[i] += surv_young_words[i];
3751 3752 }
3752 3753 }
3753 3754
3754 3755 void
3755 3756 G1CollectedHeap::cleanup_surviving_young_words() {
3756 3757 guarantee( _surviving_young_words != NULL, "pre-condition" );
3757 3758 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3758 3759 _surviving_young_words = NULL;
3759 3760 }
3760 3761
3761 3762 #ifdef ASSERT
3762 3763 class VerifyCSetClosure: public HeapRegionClosure {
3763 3764 public:
3764 3765 bool doHeapRegion(HeapRegion* hr) {
3765 3766 // Here we check that the CSet region's RSet is ready for parallel
3766 3767 // iteration. The fields that we'll verify are only manipulated
3767 3768 // when the region is part of a CSet and is collected. Afterwards,
3768 3769 // we reset these fields when we clear the region's RSet (when the
3769 3770 // region is freed) so they are ready when the region is
3770 3771 // re-allocated. The only exception to this is if there's an
3771 3772 // evacuation failure and instead of freeing the region we leave
3772 3773 // it in the heap. In that case, we reset these fields during
3773 3774 // evacuation failure handling.
3774 3775 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3775 3776
3776 3777 // Here's a good place to add any other checks we'd like to
3777 3778 // perform on CSet regions.
3778 3779 return false;
3779 3780 }
3780 3781 };
3781 3782 #endif // ASSERT
3782 3783
3783 3784 #if TASKQUEUE_STATS
3784 3785 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3785 3786 st->print_raw_cr("GC Task Stats");
3786 3787 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3787 3788 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3788 3789 }
3789 3790
3790 3791 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3791 3792 print_taskqueue_stats_hdr(st);
3792 3793
3793 3794 TaskQueueStats totals;
3794 3795 const int n = workers() != NULL ? workers()->total_workers() : 1;
3795 3796 for (int i = 0; i < n; ++i) {
3796 3797 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3797 3798 totals += task_queue(i)->stats;
3798 3799 }
3799 3800 st->print_raw("tot "); totals.print(st); st->cr();
3800 3801
3801 3802 DEBUG_ONLY(totals.verify());
3802 3803 }
3803 3804
3804 3805 void G1CollectedHeap::reset_taskqueue_stats() {
3805 3806 const int n = workers() != NULL ? workers()->total_workers() : 1;
3806 3807 for (int i = 0; i < n; ++i) {
3807 3808 task_queue(i)->stats.reset();
3808 3809 }
3809 3810 }
3810 3811 #endif // TASKQUEUE_STATS
3811 3812
3812 3813 void G1CollectedHeap::log_gc_header() {
3813 3814 if (!G1Log::fine()) {
3814 3815 return;
3815 3816 }
3816 3817
3817 3818 gclog_or_tty->gclog_stamp(_gc_tracer_stw->gc_id());
3818 3819
3819 3820 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3820 3821 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)")
3821 3822 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3822 3823
3823 3824 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3824 3825 }
3825 3826
3826 3827 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3827 3828 if (!G1Log::fine()) {
3828 3829 return;
3829 3830 }
3830 3831
3831 3832 if (G1Log::finer()) {
3832 3833 if (evacuation_failed()) {
3833 3834 gclog_or_tty->print(" (to-space exhausted)");
3834 3835 }
3835 3836 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3836 3837 g1_policy()->phase_times()->note_gc_end();
3837 3838 g1_policy()->phase_times()->print(pause_time_sec);
3838 3839 g1_policy()->print_detailed_heap_transition();
3839 3840 } else {
3840 3841 if (evacuation_failed()) {
3841 3842 gclog_or_tty->print("--");
3842 3843 }
3843 3844 g1_policy()->print_heap_transition();
3844 3845 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3845 3846 }
3846 3847 gclog_or_tty->flush();
3847 3848 }
3848 3849
3849 3850 bool
3850 3851 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3851 3852 assert_at_safepoint(true /* should_be_vm_thread */);
3852 3853 guarantee(!is_gc_active(), "collection is not reentrant");
3853 3854
3854 3855 if (GC_locker::check_active_before_gc()) {
3855 3856 return false;
3856 3857 }
3857 3858
3858 3859 _gc_timer_stw->register_gc_start();
3859 3860
3860 3861 _gc_tracer_stw->report_gc_start(gc_cause(), _gc_timer_stw->gc_start());
3861 3862
3862 3863 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3863 3864 ResourceMark rm;
3864 3865
3865 3866 print_heap_before_gc();
3866 3867 trace_heap_before_gc(_gc_tracer_stw);
3867 3868
3868 3869 verify_region_sets_optional();
3869 3870 verify_dirty_young_regions();
3870 3871
3871 3872 // This call will decide whether this pause is an initial-mark
3872 3873 // pause. If it is, during_initial_mark_pause() will return true
3873 3874 // for the duration of this pause.
3874 3875 g1_policy()->decide_on_conc_mark_initiation();
3875 3876
3876 3877 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3877 3878 assert(!g1_policy()->during_initial_mark_pause() ||
3878 3879 g1_policy()->gcs_are_young(), "sanity");
3879 3880
3880 3881 // We also do not allow mixed GCs during marking.
3881 3882 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3882 3883
3883 3884 // Record whether this pause is an initial mark. When the current
3884 3885 // thread has completed its logging output and it's safe to signal
3885 3886 // the CM thread, the flag's value in the policy has been reset.
3886 3887 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3887 3888
3888 3889 // Inner scope for scope based logging, timers, and stats collection
3889 3890 {
3890 3891 EvacuationInfo evacuation_info;
3891 3892
3892 3893 if (g1_policy()->during_initial_mark_pause()) {
3893 3894 // We are about to start a marking cycle, so we increment the
3894 3895 // full collection counter.
3895 3896 increment_old_marking_cycles_started();
3896 3897 register_concurrent_cycle_start(_gc_timer_stw->gc_start());
3897 3898 }
3898 3899
3899 3900 _gc_tracer_stw->report_yc_type(yc_type());
3900 3901
3901 3902 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3902 3903
3903 3904 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3904 3905 workers()->active_workers() : 1);
3905 3906 double pause_start_sec = os::elapsedTime();
3906 3907 g1_policy()->phase_times()->note_gc_start(active_workers);
3907 3908 log_gc_header();
3908 3909
3909 3910 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3910 3911 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3911 3912
3912 3913 // If the secondary_free_list is not empty, append it to the
3913 3914 // free_list. No need to wait for the cleanup operation to finish;
3914 3915 // the region allocation code will check the secondary_free_list
3915 3916 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3916 3917 // set, skip this step so that the region allocation code has to
3917 3918 // get entries from the secondary_free_list.
3918 3919 if (!G1StressConcRegionFreeing) {
3919 3920 append_secondary_free_list_if_not_empty_with_lock();
3920 3921 }
3921 3922
3922 3923 assert(check_young_list_well_formed(), "young list should be well formed");
3923 3924 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3924 3925 "sanity check");
3925 3926
3926 3927 // Don't dynamically change the number of GC threads this early. A value of
3927 3928 // 0 is used to indicate serial work. When parallel work is done,
3928 3929 // it will be set.
3929 3930
3930 3931 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3931 3932 IsGCActiveMark x;
3932 3933
3933 3934 gc_prologue(false);
3934 3935 increment_total_collections(false /* full gc */);
3935 3936 increment_gc_time_stamp();
3936 3937
3937 3938 verify_before_gc();
3938 3939 check_bitmaps("GC Start");
3939 3940
3940 3941 COMPILER2_PRESENT(DerivedPointerTable::clear());
3941 3942
3942 3943 // Please see comment in g1CollectedHeap.hpp and
3943 3944 // G1CollectedHeap::ref_processing_init() to see how
3944 3945 // reference processing currently works in G1.
3945 3946
3946 3947 // Enable discovery in the STW reference processor
3947 3948 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3948 3949 true /*verify_no_refs*/);
3949 3950
3950 3951 {
3951 3952 // We want to temporarily turn off discovery by the
3952 3953 // CM ref processor, if necessary, and turn it back on
3953 3954 // on again later if we do. Using a scoped
3954 3955 // NoRefDiscovery object will do this.
3955 3956 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3956 3957
3957 3958 // Forget the current alloc region (we might even choose it to be part
3958 3959 // of the collection set!).
3959 3960 _allocator->release_mutator_alloc_region();
3960 3961
3961 3962 // We should call this after we retire the mutator alloc
3962 3963 // region(s) so that all the ALLOC / RETIRE events are generated
3963 3964 // before the start GC event.
3964 3965 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3965 3966
3966 3967 // This timing is only used by the ergonomics to handle our pause target.
3967 3968 // It is unclear why this should not include the full pause. We will
3968 3969 // investigate this in CR 7178365.
3969 3970 //
3970 3971 // Preserving the old comment here if that helps the investigation:
3971 3972 //
3972 3973 // The elapsed time induced by the start time below deliberately elides
3973 3974 // the possible verification above.
3974 3975 double sample_start_time_sec = os::elapsedTime();
3975 3976
3976 3977 #if YOUNG_LIST_VERBOSE
3977 3978 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3978 3979 _young_list->print();
3979 3980 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3980 3981 #endif // YOUNG_LIST_VERBOSE
3981 3982
3982 3983 g1_policy()->record_collection_pause_start(sample_start_time_sec);
3983 3984
3984 3985 double scan_wait_start = os::elapsedTime();
3985 3986 // We have to wait until the CM threads finish scanning the
3986 3987 // root regions as it's the only way to ensure that all the
3987 3988 // objects on them have been correctly scanned before we start
3988 3989 // moving them during the GC.
3989 3990 bool waited = _cm->root_regions()->wait_until_scan_finished();
3990 3991 double wait_time_ms = 0.0;
3991 3992 if (waited) {
3992 3993 double scan_wait_end = os::elapsedTime();
3993 3994 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3994 3995 }
3995 3996 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3996 3997
3997 3998 #if YOUNG_LIST_VERBOSE
3998 3999 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3999 4000 _young_list->print();
4000 4001 #endif // YOUNG_LIST_VERBOSE
4001 4002
4002 4003 if (g1_policy()->during_initial_mark_pause()) {
4003 4004 concurrent_mark()->checkpointRootsInitialPre();
4004 4005 }
4005 4006
4006 4007 #if YOUNG_LIST_VERBOSE
4007 4008 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
4008 4009 _young_list->print();
4009 4010 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
4010 4011 #endif // YOUNG_LIST_VERBOSE
4011 4012
4012 4013 g1_policy()->finalize_cset(target_pause_time_ms, evacuation_info);
4013 4014
4014 4015 register_humongous_regions_with_in_cset_fast_test();
4015 4016
4016 4017 _cm->note_start_of_gc();
4017 4018 // We should not verify the per-thread SATB buffers given that
4018 4019 // we have not filtered them yet (we'll do so during the
4019 4020 // GC). We also call this after finalize_cset() to
4020 4021 // ensure that the CSet has been finalized.
4021 4022 _cm->verify_no_cset_oops(true /* verify_stacks */,
4022 4023 true /* verify_enqueued_buffers */,
4023 4024 false /* verify_thread_buffers */,
4024 4025 true /* verify_fingers */);
4025 4026
4026 4027 if (_hr_printer.is_active()) {
4027 4028 HeapRegion* hr = g1_policy()->collection_set();
4028 4029 while (hr != NULL) {
4029 4030 _hr_printer.cset(hr);
4030 4031 hr = hr->next_in_collection_set();
4031 4032 }
4032 4033 }
4033 4034
4034 4035 #ifdef ASSERT
4035 4036 VerifyCSetClosure cl;
4036 4037 collection_set_iterate(&cl);
4037 4038 #endif // ASSERT
4038 4039
4039 4040 setup_surviving_young_words();
4040 4041
4041 4042 // Initialize the GC alloc regions.
4042 4043 _allocator->init_gc_alloc_regions(evacuation_info);
4043 4044
4044 4045 // Actually do the work...
4045 4046 evacuate_collection_set(evacuation_info);
4046 4047
4047 4048 // We do this to mainly verify the per-thread SATB buffers
4048 4049 // (which have been filtered by now) since we didn't verify
4049 4050 // them earlier. No point in re-checking the stacks / enqueued
4050 4051 // buffers given that the CSet has not changed since last time
4051 4052 // we checked.
4052 4053 _cm->verify_no_cset_oops(false /* verify_stacks */,
4053 4054 false /* verify_enqueued_buffers */,
4054 4055 true /* verify_thread_buffers */,
4055 4056 true /* verify_fingers */);
4056 4057
4057 4058 free_collection_set(g1_policy()->collection_set(), evacuation_info);
4058 4059
4059 4060 eagerly_reclaim_humongous_regions();
4060 4061
4061 4062 g1_policy()->clear_collection_set();
4062 4063
4063 4064 cleanup_surviving_young_words();
4064 4065
4065 4066 // Start a new incremental collection set for the next pause.
4066 4067 g1_policy()->start_incremental_cset_building();
4067 4068
4068 4069 clear_cset_fast_test();
4069 4070
4070 4071 _young_list->reset_sampled_info();
4071 4072
4072 4073 // Don't check the whole heap at this point as the
4073 4074 // GC alloc regions from this pause have been tagged
4074 4075 // as survivors and moved on to the survivor list.
4075 4076 // Survivor regions will fail the !is_young() check.
4076 4077 assert(check_young_list_empty(false /* check_heap */),
4077 4078 "young list should be empty");
4078 4079
4079 4080 #if YOUNG_LIST_VERBOSE
4080 4081 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
4081 4082 _young_list->print();
4082 4083 #endif // YOUNG_LIST_VERBOSE
4083 4084
4084 4085 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
4085 4086 _young_list->first_survivor_region(),
4086 4087 _young_list->last_survivor_region());
4087 4088
4088 4089 _young_list->reset_auxilary_lists();
4089 4090
4090 4091 if (evacuation_failed()) {
4091 4092 _allocator->set_used(recalculate_used());
4092 4093 uint n_queues = MAX2((int)ParallelGCThreads, 1);
4093 4094 for (uint i = 0; i < n_queues; i++) {
4094 4095 if (_evacuation_failed_info_array[i].has_failed()) {
4095 4096 _gc_tracer_stw->report_evacuation_failed(_evacuation_failed_info_array[i]);
4096 4097 }
4097 4098 }
4098 4099 } else {
4099 4100 // The "used" of the the collection set have already been subtracted
4100 4101 // when they were freed. Add in the bytes evacuated.
4101 4102 _allocator->increase_used(g1_policy()->bytes_copied_during_gc());
4102 4103 }
4103 4104
4104 4105 if (g1_policy()->during_initial_mark_pause()) {
4105 4106 // We have to do this before we notify the CM threads that
4106 4107 // they can start working to make sure that all the
4107 4108 // appropriate initialization is done on the CM object.
4108 4109 concurrent_mark()->checkpointRootsInitialPost();
4109 4110 set_marking_started();
4110 4111 // Note that we don't actually trigger the CM thread at
4111 4112 // this point. We do that later when we're sure that
4112 4113 // the current thread has completed its logging output.
4113 4114 }
4114 4115
4115 4116 allocate_dummy_regions();
4116 4117
4117 4118 #if YOUNG_LIST_VERBOSE
4118 4119 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
4119 4120 _young_list->print();
4120 4121 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
4121 4122 #endif // YOUNG_LIST_VERBOSE
4122 4123
4123 4124 _allocator->init_mutator_alloc_region();
4124 4125
4125 4126 {
4126 4127 size_t expand_bytes = g1_policy()->expansion_amount();
4127 4128 if (expand_bytes > 0) {
4128 4129 size_t bytes_before = capacity();
4129 4130 // No need for an ergo verbose message here,
4130 4131 // expansion_amount() does this when it returns a value > 0.
4131 4132 if (!expand(expand_bytes)) {
4132 4133 // We failed to expand the heap. Cannot do anything about it.
4133 4134 }
4134 4135 }
4135 4136 }
4136 4137
4137 4138 // We redo the verification but now wrt to the new CSet which
4138 4139 // has just got initialized after the previous CSet was freed.
4139 4140 _cm->verify_no_cset_oops(true /* verify_stacks */,
4140 4141 true /* verify_enqueued_buffers */,
4141 4142 true /* verify_thread_buffers */,
4142 4143 true /* verify_fingers */);
4143 4144 _cm->note_end_of_gc();
4144 4145
4145 4146 // This timing is only used by the ergonomics to handle our pause target.
4146 4147 // It is unclear why this should not include the full pause. We will
4147 4148 // investigate this in CR 7178365.
4148 4149 double sample_end_time_sec = os::elapsedTime();
4149 4150 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
4150 4151 g1_policy()->record_collection_pause_end(pause_time_ms, evacuation_info);
4151 4152
4152 4153 MemoryService::track_memory_usage();
4153 4154
4154 4155 // In prepare_for_verify() below we'll need to scan the deferred
4155 4156 // update buffers to bring the RSets up-to-date if
4156 4157 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4157 4158 // the update buffers we'll probably need to scan cards on the
4158 4159 // regions we just allocated to (i.e., the GC alloc
4159 4160 // regions). However, during the last GC we called
4160 4161 // set_saved_mark() on all the GC alloc regions, so card
4161 4162 // scanning might skip the [saved_mark_word()...top()] area of
4162 4163 // those regions (i.e., the area we allocated objects into
4163 4164 // during the last GC). But it shouldn't. Given that
4164 4165 // saved_mark_word() is conditional on whether the GC time stamp
4165 4166 // on the region is current or not, by incrementing the GC time
4166 4167 // stamp here we invalidate all the GC time stamps on all the
4167 4168 // regions and saved_mark_word() will simply return top() for
4168 4169 // all the regions. This is a nicer way of ensuring this rather
4169 4170 // than iterating over the regions and fixing them. In fact, the
4170 4171 // GC time stamp increment here also ensures that
4171 4172 // saved_mark_word() will return top() between pauses, i.e.,
4172 4173 // during concurrent refinement. So we don't need the
4173 4174 // is_gc_active() check to decided which top to use when
4174 4175 // scanning cards (see CR 7039627).
4175 4176 increment_gc_time_stamp();
4176 4177
4177 4178 verify_after_gc();
4178 4179 check_bitmaps("GC End");
4179 4180
4180 4181 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4181 4182 ref_processor_stw()->verify_no_references_recorded();
4182 4183
4183 4184 // CM reference discovery will be re-enabled if necessary.
4184 4185 }
4185 4186
4186 4187 // We should do this after we potentially expand the heap so
4187 4188 // that all the COMMIT events are generated before the end GC
4188 4189 // event, and after we retire the GC alloc regions so that all
4189 4190 // RETIRE events are generated before the end GC event.
4190 4191 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4191 4192
4192 4193 #ifdef TRACESPINNING
4193 4194 ParallelTaskTerminator::print_termination_counts();
4194 4195 #endif
4195 4196
4196 4197 gc_epilogue(false);
4197 4198 }
4198 4199
4199 4200 // Print the remainder of the GC log output.
4200 4201 log_gc_footer(os::elapsedTime() - pause_start_sec);
4201 4202
4202 4203 // It is not yet to safe to tell the concurrent mark to
4203 4204 // start as we have some optional output below. We don't want the
4204 4205 // output from the concurrent mark thread interfering with this
4205 4206 // logging output either.
4206 4207
4207 4208 _hrm.verify_optional();
4208 4209 verify_region_sets_optional();
4209 4210
4210 4211 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4211 4212 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4212 4213
4213 4214 print_heap_after_gc();
4214 4215 trace_heap_after_gc(_gc_tracer_stw);
4215 4216
4216 4217 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4217 4218 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4218 4219 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4219 4220 // before any GC notifications are raised.
4220 4221 g1mm()->update_sizes();
4221 4222
4222 4223 _gc_tracer_stw->report_evacuation_info(&evacuation_info);
4223 4224 _gc_tracer_stw->report_tenuring_threshold(_g1_policy->tenuring_threshold());
4224 4225 _gc_timer_stw->register_gc_end();
4225 4226 _gc_tracer_stw->report_gc_end(_gc_timer_stw->gc_end(), _gc_timer_stw->time_partitions());
4226 4227 }
4227 4228 // It should now be safe to tell the concurrent mark thread to start
4228 4229 // without its logging output interfering with the logging output
4229 4230 // that came from the pause.
4230 4231
4231 4232 if (should_start_conc_mark) {
4232 4233 // CAUTION: after the doConcurrentMark() call below,
4233 4234 // the concurrent marking thread(s) could be running
4234 4235 // concurrently with us. Make sure that anything after
4235 4236 // this point does not assume that we are the only GC thread
4236 4237 // running. Note: of course, the actual marking work will
4237 4238 // not start until the safepoint itself is released in
4238 4239 // SuspendibleThreadSet::desynchronize().
4239 4240 doConcurrentMark();
4240 4241 }
4241 4242
4242 4243 return true;
4243 4244 }
4244 4245
4245 4246 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4246 4247 {
4247 4248 size_t gclab_word_size;
4248 4249 switch (purpose) {
4249 4250 case GCAllocForSurvived:
4250 4251 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4251 4252 break;
4252 4253 case GCAllocForTenured:
4253 4254 gclab_word_size = _old_plab_stats.desired_plab_sz();
4254 4255 break;
4255 4256 default:
4256 4257 assert(false, "unknown GCAllocPurpose");
4257 4258 gclab_word_size = _old_plab_stats.desired_plab_sz();
4258 4259 break;
4259 4260 }
4260 4261
4261 4262 // Prevent humongous PLAB sizes for two reasons:
4262 4263 // * PLABs are allocated using a similar paths as oops, but should
4263 4264 // never be in a humongous region
4264 4265 // * Allowing humongous PLABs needlessly churns the region free lists
4265 4266 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4266 4267 }
4267 4268
4268 4269 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4269 4270 _drain_in_progress = false;
4270 4271 set_evac_failure_closure(cl);
4271 4272 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4272 4273 }
4273 4274
4274 4275 void G1CollectedHeap::finalize_for_evac_failure() {
4275 4276 assert(_evac_failure_scan_stack != NULL &&
4276 4277 _evac_failure_scan_stack->length() == 0,
4277 4278 "Postcondition");
4278 4279 assert(!_drain_in_progress, "Postcondition");
4279 4280 delete _evac_failure_scan_stack;
4280 4281 _evac_failure_scan_stack = NULL;
4281 4282 }
4282 4283
4283 4284 void G1CollectedHeap::remove_self_forwarding_pointers() {
4284 4285 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4285 4286
4286 4287 double remove_self_forwards_start = os::elapsedTime();
4287 4288
4288 4289 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4289 4290
4290 4291 if (G1CollectedHeap::use_parallel_gc_threads()) {
4291 4292 set_par_threads();
4292 4293 workers()->run_task(&rsfp_task);
4293 4294 set_par_threads(0);
4294 4295 } else {
4295 4296 rsfp_task.work(0);
4296 4297 }
4297 4298
4298 4299 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4299 4300
4300 4301 // Reset the claim values in the regions in the collection set.
4301 4302 reset_cset_heap_region_claim_values();
4302 4303
4303 4304 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4304 4305
4305 4306 // Now restore saved marks, if any.
4306 4307 assert(_objs_with_preserved_marks.size() ==
4307 4308 _preserved_marks_of_objs.size(), "Both or none.");
4308 4309 while (!_objs_with_preserved_marks.is_empty()) {
4309 4310 oop obj = _objs_with_preserved_marks.pop();
4310 4311 markOop m = _preserved_marks_of_objs.pop();
4311 4312 obj->set_mark(m);
4312 4313 }
4313 4314 _objs_with_preserved_marks.clear(true);
4314 4315 _preserved_marks_of_objs.clear(true);
4315 4316
4316 4317 g1_policy()->phase_times()->record_evac_fail_remove_self_forwards((os::elapsedTime() - remove_self_forwards_start) * 1000.0);
4317 4318 }
4318 4319
4319 4320 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4320 4321 _evac_failure_scan_stack->push(obj);
4321 4322 }
4322 4323
4323 4324 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4324 4325 assert(_evac_failure_scan_stack != NULL, "precondition");
4325 4326
4326 4327 while (_evac_failure_scan_stack->length() > 0) {
4327 4328 oop obj = _evac_failure_scan_stack->pop();
4328 4329 _evac_failure_closure->set_region(heap_region_containing(obj));
4329 4330 obj->oop_iterate_backwards(_evac_failure_closure);
4330 4331 }
4331 4332 }
4332 4333
4333 4334 oop
4334 4335 G1CollectedHeap::handle_evacuation_failure_par(G1ParScanThreadState* _par_scan_state,
4335 4336 oop old) {
4336 4337 assert(obj_in_cs(old),
4337 4338 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4338 4339 (HeapWord*) old));
4339 4340 markOop m = old->mark();
4340 4341 oop forward_ptr = old->forward_to_atomic(old);
4341 4342 if (forward_ptr == NULL) {
4342 4343 // Forward-to-self succeeded.
4343 4344 assert(_par_scan_state != NULL, "par scan state");
4344 4345 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4345 4346 uint queue_num = _par_scan_state->queue_num();
4346 4347
4347 4348 _evacuation_failed = true;
4348 4349 _evacuation_failed_info_array[queue_num].register_copy_failure(old->size());
4349 4350 if (_evac_failure_closure != cl) {
4350 4351 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4351 4352 assert(!_drain_in_progress,
4352 4353 "Should only be true while someone holds the lock.");
4353 4354 // Set the global evac-failure closure to the current thread's.
4354 4355 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4355 4356 set_evac_failure_closure(cl);
4356 4357 // Now do the common part.
4357 4358 handle_evacuation_failure_common(old, m);
4358 4359 // Reset to NULL.
4359 4360 set_evac_failure_closure(NULL);
4360 4361 } else {
4361 4362 // The lock is already held, and this is recursive.
4362 4363 assert(_drain_in_progress, "This should only be the recursive case.");
4363 4364 handle_evacuation_failure_common(old, m);
4364 4365 }
4365 4366 return old;
4366 4367 } else {
4367 4368 // Forward-to-self failed. Either someone else managed to allocate
4368 4369 // space for this object (old != forward_ptr) or they beat us in
4369 4370 // self-forwarding it (old == forward_ptr).
4370 4371 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4371 4372 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4372 4373 "should not be in the CSet",
4373 4374 (HeapWord*) old, (HeapWord*) forward_ptr));
4374 4375 return forward_ptr;
4375 4376 }
4376 4377 }
4377 4378
4378 4379 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4379 4380 preserve_mark_if_necessary(old, m);
4380 4381
4381 4382 HeapRegion* r = heap_region_containing(old);
4382 4383 if (!r->evacuation_failed()) {
4383 4384 r->set_evacuation_failed(true);
4384 4385 _hr_printer.evac_failure(r);
4385 4386 }
4386 4387
4387 4388 push_on_evac_failure_scan_stack(old);
4388 4389
4389 4390 if (!_drain_in_progress) {
4390 4391 // prevent recursion in copy_to_survivor_space()
4391 4392 _drain_in_progress = true;
4392 4393 drain_evac_failure_scan_stack();
4393 4394 _drain_in_progress = false;
4394 4395 }
4395 4396 }
4396 4397
4397 4398 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4398 4399 assert(evacuation_failed(), "Oversaving!");
4399 4400 // We want to call the "for_promotion_failure" version only in the
4400 4401 // case of a promotion failure.
4401 4402 if (m->must_be_preserved_for_promotion_failure(obj)) {
4402 4403 _objs_with_preserved_marks.push(obj);
4403 4404 _preserved_marks_of_objs.push(m);
4404 4405 }
4405 4406 }
4406 4407
4407 4408 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4408 4409 size_t word_size,
4409 4410 AllocationContext_t context) {
4410 4411 if (purpose == GCAllocForSurvived) {
4411 4412 HeapWord* result = survivor_attempt_allocation(word_size, context);
4412 4413 if (result != NULL) {
4413 4414 return result;
4414 4415 } else {
4415 4416 // Let's try to allocate in the old gen in case we can fit the
4416 4417 // object there.
4417 4418 return old_attempt_allocation(word_size, context);
4418 4419 }
4419 4420 } else {
4420 4421 assert(purpose == GCAllocForTenured, "sanity");
4421 4422 HeapWord* result = old_attempt_allocation(word_size, context);
4422 4423 if (result != NULL) {
4423 4424 return result;
4424 4425 } else {
4425 4426 // Let's try to allocate in the survivors in case we can fit the
4426 4427 // object there.
4427 4428 return survivor_attempt_allocation(word_size, context);
4428 4429 }
4429 4430 }
4430 4431
4431 4432 ShouldNotReachHere();
4432 4433 // Trying to keep some compilers happy.
4433 4434 return NULL;
4434 4435 }
4435 4436
4436 4437 void G1ParCopyHelper::mark_object(oop obj) {
4437 4438 assert(!_g1->heap_region_containing(obj)->in_collection_set(), "should not mark objects in the CSet");
4438 4439
4439 4440 // We know that the object is not moving so it's safe to read its size.
4440 4441 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4441 4442 }
4442 4443
4443 4444 void G1ParCopyHelper::mark_forwarded_object(oop from_obj, oop to_obj) {
4444 4445 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4445 4446 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4446 4447 assert(from_obj != to_obj, "should not be self-forwarded");
4447 4448
4448 4449 assert(_g1->heap_region_containing(from_obj)->in_collection_set(), "from obj should be in the CSet");
4449 4450 assert(!_g1->heap_region_containing(to_obj)->in_collection_set(), "should not mark objects in the CSet");
4450 4451
4451 4452 // The object might be in the process of being copied by another
4452 4453 // worker so we cannot trust that its to-space image is
4453 4454 // well-formed. So we have to read its size from its from-space
4454 4455 // image which we know should not be changing.
4455 4456 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4456 4457 }
4457 4458
4458 4459 template <class T>
4459 4460 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4460 4461 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4461 4462 _scanned_klass->record_modified_oops();
4462 4463 }
4463 4464 }
4464 4465
4465 4466 template <G1Barrier barrier, G1Mark do_mark_object>
4466 4467 template <class T>
4467 4468 void G1ParCopyClosure<barrier, do_mark_object>::do_oop_work(T* p) {
4468 4469 T heap_oop = oopDesc::load_heap_oop(p);
4469 4470
4470 4471 if (oopDesc::is_null(heap_oop)) {
4471 4472 return;
4472 4473 }
4473 4474
4474 4475 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
4475 4476
4476 4477 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4477 4478
4478 4479 G1CollectedHeap::in_cset_state_t state = _g1->in_cset_state(obj);
4479 4480
4480 4481 if (state == G1CollectedHeap::InCSet) {
4481 4482 oop forwardee;
4482 4483 if (obj->is_forwarded()) {
4483 4484 forwardee = obj->forwardee();
4484 4485 } else {
4485 4486 forwardee = _par_scan_state->copy_to_survivor_space(obj);
4486 4487 }
4487 4488 assert(forwardee != NULL, "forwardee should not be NULL");
4488 4489 oopDesc::encode_store_heap_oop(p, forwardee);
4489 4490 if (do_mark_object != G1MarkNone && forwardee != obj) {
4490 4491 // If the object is self-forwarded we don't need to explicitly
4491 4492 // mark it, the evacuation failure protocol will do so.
4492 4493 mark_forwarded_object(obj, forwardee);
4493 4494 }
4494 4495
4495 4496 if (barrier == G1BarrierKlass) {
4496 4497 do_klass_barrier(p, forwardee);
4497 4498 }
4498 4499 } else {
4499 4500 if (state == G1CollectedHeap::IsHumongous) {
4500 4501 _g1->set_humongous_is_live(obj);
4501 4502 }
4502 4503 // The object is not in collection set. If we're a root scanning
4503 4504 // closure during an initial mark pause then attempt to mark the object.
4504 4505 if (do_mark_object == G1MarkFromRoot) {
4505 4506 mark_object(obj);
4506 4507 }
4507 4508 }
4508 4509
4509 4510 if (barrier == G1BarrierEvac) {
4510 4511 _par_scan_state->update_rs(_from, p, _worker_id);
4511 4512 }
4512 4513 }
4513 4514
4514 4515 template void G1ParCopyClosure<G1BarrierEvac, G1MarkNone>::do_oop_work(oop* p);
4515 4516 template void G1ParCopyClosure<G1BarrierEvac, G1MarkNone>::do_oop_work(narrowOop* p);
4516 4517
4517 4518 class G1ParEvacuateFollowersClosure : public VoidClosure {
4518 4519 protected:
4519 4520 G1CollectedHeap* _g1h;
4520 4521 G1ParScanThreadState* _par_scan_state;
4521 4522 RefToScanQueueSet* _queues;
4522 4523 ParallelTaskTerminator* _terminator;
4523 4524
4524 4525 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4525 4526 RefToScanQueueSet* queues() { return _queues; }
4526 4527 ParallelTaskTerminator* terminator() { return _terminator; }
4527 4528
4528 4529 public:
4529 4530 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4530 4531 G1ParScanThreadState* par_scan_state,
4531 4532 RefToScanQueueSet* queues,
4532 4533 ParallelTaskTerminator* terminator)
4533 4534 : _g1h(g1h), _par_scan_state(par_scan_state),
4534 4535 _queues(queues), _terminator(terminator) {}
4535 4536
4536 4537 void do_void();
4537 4538
4538 4539 private:
4539 4540 inline bool offer_termination();
4540 4541 };
4541 4542
4542 4543 bool G1ParEvacuateFollowersClosure::offer_termination() {
4543 4544 G1ParScanThreadState* const pss = par_scan_state();
4544 4545 pss->start_term_time();
4545 4546 const bool res = terminator()->offer_termination();
4546 4547 pss->end_term_time();
4547 4548 return res;
4548 4549 }
4549 4550
4550 4551 void G1ParEvacuateFollowersClosure::do_void() {
4551 4552 G1ParScanThreadState* const pss = par_scan_state();
4552 4553 pss->trim_queue();
4553 4554 do {
4554 4555 pss->steal_and_trim_queue(queues());
4555 4556 } while (!offer_termination());
4556 4557 }
4557 4558
4558 4559 class G1KlassScanClosure : public KlassClosure {
4559 4560 G1ParCopyHelper* _closure;
4560 4561 bool _process_only_dirty;
4561 4562 int _count;
4562 4563 public:
4563 4564 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4564 4565 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4565 4566 void do_klass(Klass* klass) {
4566 4567 // If the klass has not been dirtied we know that there's
4567 4568 // no references into the young gen and we can skip it.
4568 4569 if (!_process_only_dirty || klass->has_modified_oops()) {
4569 4570 // Clean the klass since we're going to scavenge all the metadata.
4570 4571 klass->clear_modified_oops();
4571 4572
4572 4573 // Tell the closure that this klass is the Klass to scavenge
4573 4574 // and is the one to dirty if oops are left pointing into the young gen.
4574 4575 _closure->set_scanned_klass(klass);
4575 4576
4576 4577 klass->oops_do(_closure);
4577 4578
4578 4579 _closure->set_scanned_klass(NULL);
4579 4580 }
4580 4581 _count++;
4581 4582 }
4582 4583 };
4583 4584
4584 4585 class G1CodeBlobClosure : public CodeBlobClosure {
4585 4586 class HeapRegionGatheringOopClosure : public OopClosure {
4586 4587 G1CollectedHeap* _g1h;
4587 4588 OopClosure* _work;
4588 4589 nmethod* _nm;
4589 4590
4590 4591 template <typename T>
4591 4592 void do_oop_work(T* p) {
4592 4593 _work->do_oop(p);
4593 4594 T oop_or_narrowoop = oopDesc::load_heap_oop(p);
4594 4595 if (!oopDesc::is_null(oop_or_narrowoop)) {
4595 4596 oop o = oopDesc::decode_heap_oop_not_null(oop_or_narrowoop);
4596 4597 HeapRegion* hr = _g1h->heap_region_containing_raw(o);
4597 4598 assert(!_g1h->obj_in_cs(o) || hr->rem_set()->strong_code_roots_list_contains(_nm), "if o still in CS then evacuation failed and nm must already be in the remset");
4598 4599 hr->add_strong_code_root(_nm);
4599 4600 }
4600 4601 }
4601 4602
4602 4603 public:
4603 4604 HeapRegionGatheringOopClosure(OopClosure* oc) : _g1h(G1CollectedHeap::heap()), _work(oc), _nm(NULL) {}
4604 4605
4605 4606 void do_oop(oop* o) {
4606 4607 do_oop_work(o);
4607 4608 }
4608 4609
4609 4610 void do_oop(narrowOop* o) {
4610 4611 do_oop_work(o);
4611 4612 }
4612 4613
4613 4614 void set_nm(nmethod* nm) {
4614 4615 _nm = nm;
4615 4616 }
4616 4617 };
4617 4618
4618 4619 HeapRegionGatheringOopClosure _oc;
4619 4620 public:
4620 4621 G1CodeBlobClosure(OopClosure* oc) : _oc(oc) {}
4621 4622
4622 4623 void do_code_blob(CodeBlob* cb) {
4623 4624 nmethod* nm = cb->as_nmethod_or_null();
4624 4625 if (nm != NULL) {
4625 4626 if (!nm->test_set_oops_do_mark()) {
4626 4627 _oc.set_nm(nm);
4627 4628 nm->oops_do(&_oc);
4628 4629 nm->fix_oop_relocations();
4629 4630 }
4630 4631 }
4631 4632 }
4632 4633 };
4633 4634
4634 4635 class G1ParTask : public AbstractGangTask {
4635 4636 protected:
4636 4637 G1CollectedHeap* _g1h;
4637 4638 RefToScanQueueSet *_queues;
4638 4639 ParallelTaskTerminator _terminator;
4639 4640 uint _n_workers;
4640 4641
4641 4642 Mutex _stats_lock;
4642 4643 Mutex* stats_lock() { return &_stats_lock; }
4643 4644
4644 4645 public:
4645 4646 G1ParTask(G1CollectedHeap* g1h, RefToScanQueueSet *task_queues)
4646 4647 : AbstractGangTask("G1 collection"),
4647 4648 _g1h(g1h),
4648 4649 _queues(task_queues),
4649 4650 _terminator(0, _queues),
4650 4651 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4651 4652 {}
4652 4653
4653 4654 RefToScanQueueSet* queues() { return _queues; }
4654 4655
4655 4656 RefToScanQueue *work_queue(int i) {
4656 4657 return queues()->queue(i);
4657 4658 }
4658 4659
4659 4660 ParallelTaskTerminator* terminator() { return &_terminator; }
4660 4661
4661 4662 virtual void set_for_termination(int active_workers) {
4662 4663 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4663 4664 // in the young space (_par_seq_tasks) in the G1 heap
4664 4665 // for SequentialSubTasksDone.
4665 4666 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4666 4667 // both of which need setting by set_n_termination().
4667 4668 _g1h->SharedHeap::set_n_termination(active_workers);
4668 4669 _g1h->set_n_termination(active_workers);
4669 4670 terminator()->reset_for_reuse(active_workers);
4670 4671 _n_workers = active_workers;
4671 4672 }
4672 4673
4673 4674 // Helps out with CLD processing.
4674 4675 //
4675 4676 // During InitialMark we need to:
4676 4677 // 1) Scavenge all CLDs for the young GC.
4677 4678 // 2) Mark all objects directly reachable from strong CLDs.
4678 4679 template <G1Mark do_mark_object>
4679 4680 class G1CLDClosure : public CLDClosure {
4680 4681 G1ParCopyClosure<G1BarrierNone, do_mark_object>* _oop_closure;
4681 4682 G1ParCopyClosure<G1BarrierKlass, do_mark_object> _oop_in_klass_closure;
4682 4683 G1KlassScanClosure _klass_in_cld_closure;
4683 4684 bool _claim;
4684 4685
4685 4686 public:
4686 4687 G1CLDClosure(G1ParCopyClosure<G1BarrierNone, do_mark_object>* oop_closure,
4687 4688 bool only_young, bool claim)
4688 4689 : _oop_closure(oop_closure),
4689 4690 _oop_in_klass_closure(oop_closure->g1(),
4690 4691 oop_closure->pss(),
4691 4692 oop_closure->rp()),
4692 4693 _klass_in_cld_closure(&_oop_in_klass_closure, only_young),
4693 4694 _claim(claim) {
4694 4695
4695 4696 }
4696 4697
4697 4698 void do_cld(ClassLoaderData* cld) {
4698 4699 cld->oops_do(_oop_closure, &_klass_in_cld_closure, _claim);
4699 4700 }
4700 4701 };
4701 4702
4702 4703 void work(uint worker_id) {
4703 4704 if (worker_id >= _n_workers) return; // no work needed this round
4704 4705
4705 4706 double start_time_ms = os::elapsedTime() * 1000.0;
4706 4707 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4707 4708
4708 4709 {
4709 4710 ResourceMark rm;
4710 4711 HandleMark hm;
4711 4712
4712 4713 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4713 4714
4714 4715 G1ParScanThreadState pss(_g1h, worker_id, rp);
4715 4716 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4716 4717
4717 4718 pss.set_evac_failure_closure(&evac_failure_cl);
4718 4719
4719 4720 bool only_young = _g1h->g1_policy()->gcs_are_young();
4720 4721
4721 4722 // Non-IM young GC.
4722 4723 G1ParCopyClosure<G1BarrierNone, G1MarkNone> scan_only_root_cl(_g1h, &pss, rp);
4723 4724 G1CLDClosure<G1MarkNone> scan_only_cld_cl(&scan_only_root_cl,
4724 4725 only_young, // Only process dirty klasses.
4725 4726 false); // No need to claim CLDs.
4726 4727 // IM young GC.
4727 4728 // Strong roots closures.
4728 4729 G1ParCopyClosure<G1BarrierNone, G1MarkFromRoot> scan_mark_root_cl(_g1h, &pss, rp);
4729 4730 G1CLDClosure<G1MarkFromRoot> scan_mark_cld_cl(&scan_mark_root_cl,
4730 4731 false, // Process all klasses.
4731 4732 true); // Need to claim CLDs.
4732 4733 // Weak roots closures.
4733 4734 G1ParCopyClosure<G1BarrierNone, G1MarkPromotedFromRoot> scan_mark_weak_root_cl(_g1h, &pss, rp);
4734 4735 G1CLDClosure<G1MarkPromotedFromRoot> scan_mark_weak_cld_cl(&scan_mark_weak_root_cl,
4735 4736 false, // Process all klasses.
4736 4737 true); // Need to claim CLDs.
4737 4738
4738 4739 G1CodeBlobClosure scan_only_code_cl(&scan_only_root_cl);
4739 4740 G1CodeBlobClosure scan_mark_code_cl(&scan_mark_root_cl);
4740 4741 // IM Weak code roots are handled later.
4741 4742
4742 4743 OopClosure* strong_root_cl;
4743 4744 OopClosure* weak_root_cl;
4744 4745 CLDClosure* strong_cld_cl;
4745 4746 CLDClosure* weak_cld_cl;
4746 4747 CodeBlobClosure* strong_code_cl;
4747 4748
4748 4749 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4749 4750 // We also need to mark copied objects.
4750 4751 strong_root_cl = &scan_mark_root_cl;
4751 4752 strong_cld_cl = &scan_mark_cld_cl;
4752 4753 strong_code_cl = &scan_mark_code_cl;
4753 4754 if (ClassUnloadingWithConcurrentMark) {
4754 4755 weak_root_cl = &scan_mark_weak_root_cl;
4755 4756 weak_cld_cl = &scan_mark_weak_cld_cl;
4756 4757 } else {
4757 4758 weak_root_cl = &scan_mark_root_cl;
4758 4759 weak_cld_cl = &scan_mark_cld_cl;
4759 4760 }
4760 4761 } else {
4761 4762 strong_root_cl = &scan_only_root_cl;
4762 4763 weak_root_cl = &scan_only_root_cl;
4763 4764 strong_cld_cl = &scan_only_cld_cl;
4764 4765 weak_cld_cl = &scan_only_cld_cl;
4765 4766 strong_code_cl = &scan_only_code_cl;
4766 4767 }
4767 4768
4768 4769
4769 4770 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4770 4771
4771 4772 pss.start_strong_roots();
4772 4773 _g1h->g1_process_roots(strong_root_cl,
4773 4774 weak_root_cl,
4774 4775 &push_heap_rs_cl,
4775 4776 strong_cld_cl,
4776 4777 weak_cld_cl,
4777 4778 strong_code_cl,
4778 4779 worker_id);
4779 4780
4780 4781 pss.end_strong_roots();
4781 4782
4782 4783 {
4783 4784 double start = os::elapsedTime();
4784 4785 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4785 4786 evac.do_void();
4786 4787 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4787 4788 double term_ms = pss.term_time()*1000.0;
4788 4789 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4789 4790 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4790 4791 }
4791 4792 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4792 4793 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4793 4794
4794 4795 if (ParallelGCVerbose) {
4795 4796 MutexLocker x(stats_lock());
4796 4797 pss.print_termination_stats(worker_id);
4797 4798 }
4798 4799
4799 4800 assert(pss.queue_is_empty(), "should be empty");
4800 4801
4801 4802 // Close the inner scope so that the ResourceMark and HandleMark
4802 4803 // destructors are executed here and are included as part of the
4803 4804 // "GC Worker Time".
4804 4805 }
4805 4806
4806 4807 double end_time_ms = os::elapsedTime() * 1000.0;
4807 4808 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4808 4809 }
4809 4810 };
4810 4811
4811 4812 // *** Common G1 Evacuation Stuff
4812 4813
4813 4814 // This method is run in a GC worker.
4814 4815
4815 4816 void
4816 4817 G1CollectedHeap::
4817 4818 g1_process_roots(OopClosure* scan_non_heap_roots,
4818 4819 OopClosure* scan_non_heap_weak_roots,
4819 4820 OopsInHeapRegionClosure* scan_rs,
4820 4821 CLDClosure* scan_strong_clds,
4821 4822 CLDClosure* scan_weak_clds,
4822 4823 CodeBlobClosure* scan_strong_code,
4823 4824 uint worker_i) {
4824 4825
4825 4826 // First scan the shared roots.
4826 4827 double ext_roots_start = os::elapsedTime();
4827 4828 double closure_app_time_sec = 0.0;
4828 4829
4829 4830 bool during_im = _g1h->g1_policy()->during_initial_mark_pause();
4830 4831 bool trace_metadata = during_im && ClassUnloadingWithConcurrentMark;
4831 4832
4832 4833 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4833 4834 BufferingOopClosure buf_scan_non_heap_weak_roots(scan_non_heap_weak_roots);
4834 4835
4835 4836 process_roots(false, // no scoping; this is parallel code
4836 4837 SharedHeap::SO_None,
4837 4838 &buf_scan_non_heap_roots,
4838 4839 &buf_scan_non_heap_weak_roots,
4839 4840 scan_strong_clds,
4840 4841 // Unloading Initial Marks handle the weak CLDs separately.
4841 4842 (trace_metadata ? NULL : scan_weak_clds),
4842 4843 scan_strong_code);
4843 4844
4844 4845 // Now the CM ref_processor roots.
4845 4846 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
4846 4847 // We need to treat the discovered reference lists of the
4847 4848 // concurrent mark ref processor as roots and keep entries
4848 4849 // (which are added by the marking threads) on them live
4849 4850 // until they can be processed at the end of marking.
4850 4851 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
4851 4852 }
4852 4853
4853 4854 if (trace_metadata) {
4854 4855 // Barrier to make sure all workers passed
4855 4856 // the strong CLD and strong nmethods phases.
4856 4857 active_strong_roots_scope()->wait_until_all_workers_done_with_threads(n_par_threads());
4857 4858
4858 4859 // Now take the complement of the strong CLDs.
4859 4860 ClassLoaderDataGraph::roots_cld_do(NULL, scan_weak_clds);
4860 4861 }
4861 4862
4862 4863 // Finish up any enqueued closure apps (attributed as object copy time).
4863 4864 buf_scan_non_heap_roots.done();
4864 4865 buf_scan_non_heap_weak_roots.done();
4865 4866
4866 4867 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds()
4867 4868 + buf_scan_non_heap_weak_roots.closure_app_seconds();
4868 4869
4869 4870 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
4870 4871
4871 4872 double ext_root_time_ms =
4872 4873 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
4873 4874
4874 4875 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
4875 4876
4876 4877 // During conc marking we have to filter the per-thread SATB buffers
4877 4878 // to make sure we remove any oops into the CSet (which will show up
4878 4879 // as implicitly live).
4879 4880 double satb_filtering_ms = 0.0;
4880 4881 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
4881 4882 if (mark_in_progress()) {
4882 4883 double satb_filter_start = os::elapsedTime();
4883 4884
4884 4885 JavaThread::satb_mark_queue_set().filter_thread_buffers();
4885 4886
4886 4887 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
4887 4888 }
4888 4889 }
4889 4890 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
4890 4891
4891 4892 // Now scan the complement of the collection set.
4892 4893 G1CodeBlobClosure scavenge_cs_nmethods(scan_non_heap_weak_roots);
4893 4894
4894 4895 g1_rem_set()->oops_into_collection_set_do(scan_rs, &scavenge_cs_nmethods, worker_i);
4895 4896
4896 4897 _process_strong_tasks->all_tasks_completed();
4897 4898 }
4898 4899
4899 4900 class G1StringSymbolTableUnlinkTask : public AbstractGangTask {
4900 4901 private:
4901 4902 BoolObjectClosure* _is_alive;
4902 4903 int _initial_string_table_size;
4903 4904 int _initial_symbol_table_size;
4904 4905
4905 4906 bool _process_strings;
4906 4907 int _strings_processed;
4907 4908 int _strings_removed;
4908 4909
4909 4910 bool _process_symbols;
4910 4911 int _symbols_processed;
4911 4912 int _symbols_removed;
4912 4913
4913 4914 bool _do_in_parallel;
4914 4915 public:
4915 4916 G1StringSymbolTableUnlinkTask(BoolObjectClosure* is_alive, bool process_strings, bool process_symbols) :
4916 4917 AbstractGangTask("String/Symbol Unlinking"),
4917 4918 _is_alive(is_alive),
4918 4919 _do_in_parallel(G1CollectedHeap::use_parallel_gc_threads()),
4919 4920 _process_strings(process_strings), _strings_processed(0), _strings_removed(0),
4920 4921 _process_symbols(process_symbols), _symbols_processed(0), _symbols_removed(0) {
4921 4922
4922 4923 _initial_string_table_size = StringTable::the_table()->table_size();
4923 4924 _initial_symbol_table_size = SymbolTable::the_table()->table_size();
4924 4925 if (process_strings) {
4925 4926 StringTable::clear_parallel_claimed_index();
4926 4927 }
4927 4928 if (process_symbols) {
4928 4929 SymbolTable::clear_parallel_claimed_index();
4929 4930 }
4930 4931 }
4931 4932
4932 4933 ~G1StringSymbolTableUnlinkTask() {
4933 4934 guarantee(!_process_strings || !_do_in_parallel || StringTable::parallel_claimed_index() >= _initial_string_table_size,
4934 4935 err_msg("claim value "INT32_FORMAT" after unlink less than initial string table size "INT32_FORMAT,
4935 4936 StringTable::parallel_claimed_index(), _initial_string_table_size));
4936 4937 guarantee(!_process_symbols || !_do_in_parallel || SymbolTable::parallel_claimed_index() >= _initial_symbol_table_size,
4937 4938 err_msg("claim value "INT32_FORMAT" after unlink less than initial symbol table size "INT32_FORMAT,
4938 4939 SymbolTable::parallel_claimed_index(), _initial_symbol_table_size));
4939 4940
4940 4941 if (G1TraceStringSymbolTableScrubbing) {
4941 4942 gclog_or_tty->print_cr("Cleaned string and symbol table, "
4942 4943 "strings: "SIZE_FORMAT" processed, "SIZE_FORMAT" removed, "
4943 4944 "symbols: "SIZE_FORMAT" processed, "SIZE_FORMAT" removed",
4944 4945 strings_processed(), strings_removed(),
4945 4946 symbols_processed(), symbols_removed());
4946 4947 }
4947 4948 }
4948 4949
4949 4950 void work(uint worker_id) {
4950 4951 if (_do_in_parallel) {
4951 4952 int strings_processed = 0;
4952 4953 int strings_removed = 0;
4953 4954 int symbols_processed = 0;
4954 4955 int symbols_removed = 0;
4955 4956 if (_process_strings) {
4956 4957 StringTable::possibly_parallel_unlink(_is_alive, &strings_processed, &strings_removed);
4957 4958 Atomic::add(strings_processed, &_strings_processed);
4958 4959 Atomic::add(strings_removed, &_strings_removed);
4959 4960 }
4960 4961 if (_process_symbols) {
4961 4962 SymbolTable::possibly_parallel_unlink(&symbols_processed, &symbols_removed);
4962 4963 Atomic::add(symbols_processed, &_symbols_processed);
4963 4964 Atomic::add(symbols_removed, &_symbols_removed);
4964 4965 }
4965 4966 } else {
4966 4967 if (_process_strings) {
4967 4968 StringTable::unlink(_is_alive, &_strings_processed, &_strings_removed);
4968 4969 }
4969 4970 if (_process_symbols) {
4970 4971 SymbolTable::unlink(&_symbols_processed, &_symbols_removed);
4971 4972 }
4972 4973 }
4973 4974 }
4974 4975
4975 4976 size_t strings_processed() const { return (size_t)_strings_processed; }
4976 4977 size_t strings_removed() const { return (size_t)_strings_removed; }
4977 4978
4978 4979 size_t symbols_processed() const { return (size_t)_symbols_processed; }
4979 4980 size_t symbols_removed() const { return (size_t)_symbols_removed; }
4980 4981 };
4981 4982
4982 4983 class G1CodeCacheUnloadingTask VALUE_OBJ_CLASS_SPEC {
4983 4984 private:
4984 4985 static Monitor* _lock;
4985 4986
4986 4987 BoolObjectClosure* const _is_alive;
4987 4988 const bool _unloading_occurred;
4988 4989 const uint _num_workers;
4989 4990
4990 4991 // Variables used to claim nmethods.
4991 4992 nmethod* _first_nmethod;
4992 4993 volatile nmethod* _claimed_nmethod;
4993 4994
4994 4995 // The list of nmethods that need to be processed by the second pass.
4995 4996 volatile nmethod* _postponed_list;
4996 4997 volatile uint _num_entered_barrier;
4997 4998
4998 4999 public:
4999 5000 G1CodeCacheUnloadingTask(uint num_workers, BoolObjectClosure* is_alive, bool unloading_occurred) :
5000 5001 _is_alive(is_alive),
5001 5002 _unloading_occurred(unloading_occurred),
5002 5003 _num_workers(num_workers),
5003 5004 _first_nmethod(NULL),
5004 5005 _claimed_nmethod(NULL),
5005 5006 _postponed_list(NULL),
5006 5007 _num_entered_barrier(0)
5007 5008 {
5008 5009 nmethod::increase_unloading_clock();
5009 5010 _first_nmethod = CodeCache::alive_nmethod(CodeCache::first());
5010 5011 _claimed_nmethod = (volatile nmethod*)_first_nmethod;
5011 5012 }
5012 5013
5013 5014 ~G1CodeCacheUnloadingTask() {
5014 5015 CodeCache::verify_clean_inline_caches();
5015 5016
5016 5017 CodeCache::set_needs_cache_clean(false);
5017 5018 guarantee(CodeCache::scavenge_root_nmethods() == NULL, "Must be");
5018 5019
5019 5020 CodeCache::verify_icholder_relocations();
5020 5021 }
5021 5022
5022 5023 private:
5023 5024 void add_to_postponed_list(nmethod* nm) {
5024 5025 nmethod* old;
5025 5026 do {
5026 5027 old = (nmethod*)_postponed_list;
5027 5028 nm->set_unloading_next(old);
5028 5029 } while ((nmethod*)Atomic::cmpxchg_ptr(nm, &_postponed_list, old) != old);
5029 5030 }
5030 5031
5031 5032 void clean_nmethod(nmethod* nm) {
5032 5033 bool postponed = nm->do_unloading_parallel(_is_alive, _unloading_occurred);
5033 5034
5034 5035 if (postponed) {
5035 5036 // This nmethod referred to an nmethod that has not been cleaned/unloaded yet.
5036 5037 add_to_postponed_list(nm);
5037 5038 }
5038 5039
5039 5040 // Mark that this thread has been cleaned/unloaded.
5040 5041 // After this call, it will be safe to ask if this nmethod was unloaded or not.
5041 5042 nm->set_unloading_clock(nmethod::global_unloading_clock());
5042 5043 }
5043 5044
5044 5045 void clean_nmethod_postponed(nmethod* nm) {
5045 5046 nm->do_unloading_parallel_postponed(_is_alive, _unloading_occurred);
5046 5047 }
5047 5048
5048 5049 static const int MaxClaimNmethods = 16;
5049 5050
5050 5051 void claim_nmethods(nmethod** claimed_nmethods, int *num_claimed_nmethods) {
5051 5052 nmethod* first;
5052 5053 nmethod* last;
5053 5054
5054 5055 do {
5055 5056 *num_claimed_nmethods = 0;
5056 5057
5057 5058 first = last = (nmethod*)_claimed_nmethod;
5058 5059
5059 5060 if (first != NULL) {
5060 5061 for (int i = 0; i < MaxClaimNmethods; i++) {
5061 5062 last = CodeCache::alive_nmethod(CodeCache::next(last));
5062 5063
5063 5064 if (last == NULL) {
5064 5065 break;
5065 5066 }
5066 5067
5067 5068 claimed_nmethods[i] = last;
5068 5069 (*num_claimed_nmethods)++;
5069 5070 }
5070 5071 }
5071 5072
5072 5073 } while ((nmethod*)Atomic::cmpxchg_ptr(last, &_claimed_nmethod, first) != first);
5073 5074 }
5074 5075
5075 5076 nmethod* claim_postponed_nmethod() {
5076 5077 nmethod* claim;
5077 5078 nmethod* next;
5078 5079
5079 5080 do {
5080 5081 claim = (nmethod*)_postponed_list;
5081 5082 if (claim == NULL) {
5082 5083 return NULL;
5083 5084 }
5084 5085
5085 5086 next = claim->unloading_next();
5086 5087
5087 5088 } while ((nmethod*)Atomic::cmpxchg_ptr(next, &_postponed_list, claim) != claim);
5088 5089
5089 5090 return claim;
5090 5091 }
5091 5092
5092 5093 public:
5093 5094 // Mark that we're done with the first pass of nmethod cleaning.
5094 5095 void barrier_mark(uint worker_id) {
5095 5096 MonitorLockerEx ml(_lock, Mutex::_no_safepoint_check_flag);
5096 5097 _num_entered_barrier++;
5097 5098 if (_num_entered_barrier == _num_workers) {
5098 5099 ml.notify_all();
5099 5100 }
5100 5101 }
5101 5102
5102 5103 // See if we have to wait for the other workers to
5103 5104 // finish their first-pass nmethod cleaning work.
5104 5105 void barrier_wait(uint worker_id) {
5105 5106 if (_num_entered_barrier < _num_workers) {
5106 5107 MonitorLockerEx ml(_lock, Mutex::_no_safepoint_check_flag);
5107 5108 while (_num_entered_barrier < _num_workers) {
5108 5109 ml.wait(Mutex::_no_safepoint_check_flag, 0, false);
5109 5110 }
5110 5111 }
5111 5112 }
5112 5113
5113 5114 // Cleaning and unloading of nmethods. Some work has to be postponed
5114 5115 // to the second pass, when we know which nmethods survive.
5115 5116 void work_first_pass(uint worker_id) {
5116 5117 // The first nmethods is claimed by the first worker.
5117 5118 if (worker_id == 0 && _first_nmethod != NULL) {
5118 5119 clean_nmethod(_first_nmethod);
5119 5120 _first_nmethod = NULL;
5120 5121 }
5121 5122
5122 5123 int num_claimed_nmethods;
5123 5124 nmethod* claimed_nmethods[MaxClaimNmethods];
5124 5125
5125 5126 while (true) {
↓ open down ↓ |
5086 lines elided |
↑ open up ↑ |
5126 5127 claim_nmethods(claimed_nmethods, &num_claimed_nmethods);
5127 5128
5128 5129 if (num_claimed_nmethods == 0) {
5129 5130 break;
5130 5131 }
5131 5132
5132 5133 for (int i = 0; i < num_claimed_nmethods; i++) {
5133 5134 clean_nmethod(claimed_nmethods[i]);
5134 5135 }
5135 5136 }
5137 +
5138 + // The nmethod cleaning helps out and does the CodeCache part of MetadataOnStackMark.
5139 + // Need to retire the buffers now that this thread has stopped cleaning nmethods.
5140 + MetadataOnStackMark::retire_buffer_for_thread(Thread::current());
5136 5141 }
5137 5142
5138 5143 void work_second_pass(uint worker_id) {
5139 5144 nmethod* nm;
5140 5145 // Take care of postponed nmethods.
5141 5146 while ((nm = claim_postponed_nmethod()) != NULL) {
5142 5147 clean_nmethod_postponed(nm);
5143 5148 }
5144 5149 }
5145 5150 };
5146 5151
5147 5152 Monitor* G1CodeCacheUnloadingTask::_lock = new Monitor(Mutex::leaf, "Code Cache Unload lock");
5148 5153
5149 5154 class G1KlassCleaningTask : public StackObj {
5150 5155 BoolObjectClosure* _is_alive;
5151 5156 volatile jint _clean_klass_tree_claimed;
5152 5157 ClassLoaderDataGraphKlassIteratorAtomic _klass_iterator;
5153 5158
5154 5159 public:
5155 5160 G1KlassCleaningTask(BoolObjectClosure* is_alive) :
5156 5161 _is_alive(is_alive),
5157 5162 _clean_klass_tree_claimed(0),
5158 5163 _klass_iterator() {
5159 5164 }
5160 5165
5161 5166 private:
5162 5167 bool claim_clean_klass_tree_task() {
5163 5168 if (_clean_klass_tree_claimed) {
5164 5169 return false;
5165 5170 }
5166 5171
5167 5172 return Atomic::cmpxchg(1, (jint*)&_clean_klass_tree_claimed, 0) == 0;
5168 5173 }
5169 5174
5170 5175 InstanceKlass* claim_next_klass() {
5171 5176 Klass* klass;
5172 5177 do {
5173 5178 klass =_klass_iterator.next_klass();
5174 5179 } while (klass != NULL && !klass->oop_is_instance());
5175 5180
5176 5181 return (InstanceKlass*)klass;
5177 5182 }
↓ open down ↓ |
32 lines elided |
↑ open up ↑ |
5178 5183
5179 5184 public:
5180 5185
5181 5186 void clean_klass(InstanceKlass* ik) {
5182 5187 ik->clean_implementors_list(_is_alive);
5183 5188 ik->clean_method_data(_is_alive);
5184 5189
5185 5190 // G1 specific cleanup work that has
5186 5191 // been moved here to be done in parallel.
5187 5192 ik->clean_dependent_nmethods();
5193 + if (JvmtiExport::has_redefined_a_class()) {
5194 + InstanceKlass::purge_previous_versions(ik);
5195 + }
5188 5196 }
5189 5197
5190 5198 void work() {
5191 5199 ResourceMark rm;
5192 5200
5193 5201 // One worker will clean the subklass/sibling klass tree.
5194 5202 if (claim_clean_klass_tree_task()) {
5195 5203 Klass::clean_subklass_tree(_is_alive);
5196 5204 }
5197 5205
5198 5206 // All workers will help cleaning the classes,
5199 5207 InstanceKlass* klass;
5200 5208 while ((klass = claim_next_klass()) != NULL) {
5201 5209 clean_klass(klass);
5202 5210 }
5203 5211 }
5204 5212 };
5205 5213
5206 5214 // To minimize the remark pause times, the tasks below are done in parallel.
5207 5215 class G1ParallelCleaningTask : public AbstractGangTask {
5208 5216 private:
5209 5217 G1StringSymbolTableUnlinkTask _string_symbol_task;
5210 5218 G1CodeCacheUnloadingTask _code_cache_task;
5211 5219 G1KlassCleaningTask _klass_cleaning_task;
↓ open down ↓ |
14 lines elided |
↑ open up ↑ |
5212 5220
5213 5221 public:
5214 5222 // The constructor is run in the VMThread.
5215 5223 G1ParallelCleaningTask(BoolObjectClosure* is_alive, bool process_strings, bool process_symbols, uint num_workers, bool unloading_occurred) :
5216 5224 AbstractGangTask("Parallel Cleaning"),
5217 5225 _string_symbol_task(is_alive, process_strings, process_symbols),
5218 5226 _code_cache_task(num_workers, is_alive, unloading_occurred),
5219 5227 _klass_cleaning_task(is_alive) {
5220 5228 }
5221 5229
5230 + void pre_work_verification() {
5231 + assert(!MetadataOnStackMark::has_buffer_for_thread(Thread::current()), "Should be empty");
5232 + }
5233 +
5234 + void post_work_verification() {
5235 + assert(!MetadataOnStackMark::has_buffer_for_thread(Thread::current()), "Should be empty");
5236 + }
5237 +
5222 5238 // The parallel work done by all worker threads.
5223 5239 void work(uint worker_id) {
5240 + pre_work_verification();
5241 +
5224 5242 // Do first pass of code cache cleaning.
5225 5243 _code_cache_task.work_first_pass(worker_id);
5226 5244
5227 5245 // Let the threads mark that the first pass is done.
5228 5246 _code_cache_task.barrier_mark(worker_id);
5229 5247
5230 5248 // Clean the Strings and Symbols.
5231 5249 _string_symbol_task.work(worker_id);
5232 5250
5233 5251 // Wait for all workers to finish the first code cache cleaning pass.
5234 5252 _code_cache_task.barrier_wait(worker_id);
5235 5253
5236 5254 // Do the second code cache cleaning work, which realize on
5237 5255 // the liveness information gathered during the first pass.
5238 5256 _code_cache_task.work_second_pass(worker_id);
5239 5257
5240 5258 // Clean all klasses that were not unloaded.
5241 5259 _klass_cleaning_task.work();
5260 +
5261 + post_work_verification();
5242 5262 }
5243 5263 };
5244 5264
5245 5265
5246 5266 void G1CollectedHeap::parallel_cleaning(BoolObjectClosure* is_alive,
5247 5267 bool process_strings,
5248 5268 bool process_symbols,
5249 5269 bool class_unloading_occurred) {
5250 5270 uint n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5251 5271 workers()->active_workers() : 1);
5252 5272
5253 5273 G1ParallelCleaningTask g1_unlink_task(is_alive, process_strings, process_symbols,
5254 5274 n_workers, class_unloading_occurred);
5255 5275 if (G1CollectedHeap::use_parallel_gc_threads()) {
5256 5276 set_par_threads(n_workers);
5257 5277 workers()->run_task(&g1_unlink_task);
5258 5278 set_par_threads(0);
5259 5279 } else {
5260 5280 g1_unlink_task.work(0);
5261 5281 }
5262 5282 }
5263 5283
5264 5284 void G1CollectedHeap::unlink_string_and_symbol_table(BoolObjectClosure* is_alive,
5265 5285 bool process_strings, bool process_symbols) {
5266 5286 {
5267 5287 uint n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5268 5288 _g1h->workers()->active_workers() : 1);
5269 5289 G1StringSymbolTableUnlinkTask g1_unlink_task(is_alive, process_strings, process_symbols);
5270 5290 if (G1CollectedHeap::use_parallel_gc_threads()) {
5271 5291 set_par_threads(n_workers);
5272 5292 workers()->run_task(&g1_unlink_task);
5273 5293 set_par_threads(0);
5274 5294 } else {
5275 5295 g1_unlink_task.work(0);
5276 5296 }
5277 5297 }
5278 5298
5279 5299 if (G1StringDedup::is_enabled()) {
5280 5300 G1StringDedup::unlink(is_alive);
5281 5301 }
5282 5302 }
5283 5303
5284 5304 class G1RedirtyLoggedCardsTask : public AbstractGangTask {
5285 5305 private:
5286 5306 DirtyCardQueueSet* _queue;
5287 5307 public:
5288 5308 G1RedirtyLoggedCardsTask(DirtyCardQueueSet* queue) : AbstractGangTask("Redirty Cards"), _queue(queue) { }
5289 5309
5290 5310 virtual void work(uint worker_id) {
5291 5311 double start_time = os::elapsedTime();
5292 5312
5293 5313 RedirtyLoggedCardTableEntryClosure cl;
5294 5314 if (G1CollectedHeap::heap()->use_parallel_gc_threads()) {
5295 5315 _queue->par_apply_closure_to_all_completed_buffers(&cl);
5296 5316 } else {
5297 5317 _queue->apply_closure_to_all_completed_buffers(&cl);
5298 5318 }
5299 5319
5300 5320 G1GCPhaseTimes* timer = G1CollectedHeap::heap()->g1_policy()->phase_times();
5301 5321 timer->record_redirty_logged_cards_time_ms(worker_id, (os::elapsedTime() - start_time) * 1000.0);
5302 5322 timer->record_redirty_logged_cards_processed_cards(worker_id, cl.num_processed());
5303 5323 }
5304 5324 };
5305 5325
5306 5326 void G1CollectedHeap::redirty_logged_cards() {
5307 5327 double redirty_logged_cards_start = os::elapsedTime();
5308 5328
5309 5329 uint n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5310 5330 _g1h->workers()->active_workers() : 1);
5311 5331
5312 5332 G1RedirtyLoggedCardsTask redirty_task(&dirty_card_queue_set());
5313 5333 dirty_card_queue_set().reset_for_par_iteration();
5314 5334 if (use_parallel_gc_threads()) {
5315 5335 set_par_threads(n_workers);
5316 5336 workers()->run_task(&redirty_task);
5317 5337 set_par_threads(0);
5318 5338 } else {
5319 5339 redirty_task.work(0);
5320 5340 }
5321 5341
5322 5342 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5323 5343 dcq.merge_bufferlists(&dirty_card_queue_set());
5324 5344 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5325 5345
5326 5346 g1_policy()->phase_times()->record_redirty_logged_cards_time_ms((os::elapsedTime() - redirty_logged_cards_start) * 1000.0);
5327 5347 }
5328 5348
5329 5349 // Weak Reference Processing support
5330 5350
5331 5351 // An always "is_alive" closure that is used to preserve referents.
5332 5352 // If the object is non-null then it's alive. Used in the preservation
5333 5353 // of referent objects that are pointed to by reference objects
5334 5354 // discovered by the CM ref processor.
5335 5355 class G1AlwaysAliveClosure: public BoolObjectClosure {
5336 5356 G1CollectedHeap* _g1;
5337 5357 public:
5338 5358 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5339 5359 bool do_object_b(oop p) {
5340 5360 if (p != NULL) {
5341 5361 return true;
5342 5362 }
5343 5363 return false;
5344 5364 }
5345 5365 };
5346 5366
5347 5367 bool G1STWIsAliveClosure::do_object_b(oop p) {
5348 5368 // An object is reachable if it is outside the collection set,
5349 5369 // or is inside and copied.
5350 5370 return !_g1->obj_in_cs(p) || p->is_forwarded();
5351 5371 }
5352 5372
5353 5373 // Non Copying Keep Alive closure
5354 5374 class G1KeepAliveClosure: public OopClosure {
5355 5375 G1CollectedHeap* _g1;
5356 5376 public:
5357 5377 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5358 5378 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5359 5379 void do_oop(oop* p) {
5360 5380 oop obj = *p;
5361 5381 assert(obj != NULL, "the caller should have filtered out NULL values");
5362 5382
5363 5383 G1CollectedHeap::in_cset_state_t cset_state = _g1->in_cset_state(obj);
5364 5384 if (cset_state == G1CollectedHeap::InNeither) {
5365 5385 return;
5366 5386 }
5367 5387 if (cset_state == G1CollectedHeap::InCSet) {
5368 5388 assert( obj->is_forwarded(), "invariant" );
5369 5389 *p = obj->forwardee();
5370 5390 } else {
5371 5391 assert(!obj->is_forwarded(), "invariant" );
5372 5392 assert(cset_state == G1CollectedHeap::IsHumongous,
5373 5393 err_msg("Only allowed InCSet state is IsHumongous, but is %d", cset_state));
5374 5394 _g1->set_humongous_is_live(obj);
5375 5395 }
5376 5396 }
5377 5397 };
5378 5398
5379 5399 // Copying Keep Alive closure - can be called from both
5380 5400 // serial and parallel code as long as different worker
5381 5401 // threads utilize different G1ParScanThreadState instances
5382 5402 // and different queues.
5383 5403
5384 5404 class G1CopyingKeepAliveClosure: public OopClosure {
5385 5405 G1CollectedHeap* _g1h;
5386 5406 OopClosure* _copy_non_heap_obj_cl;
5387 5407 G1ParScanThreadState* _par_scan_state;
5388 5408
5389 5409 public:
5390 5410 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5391 5411 OopClosure* non_heap_obj_cl,
5392 5412 G1ParScanThreadState* pss):
5393 5413 _g1h(g1h),
5394 5414 _copy_non_heap_obj_cl(non_heap_obj_cl),
5395 5415 _par_scan_state(pss)
5396 5416 {}
5397 5417
5398 5418 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5399 5419 virtual void do_oop( oop* p) { do_oop_work(p); }
5400 5420
5401 5421 template <class T> void do_oop_work(T* p) {
5402 5422 oop obj = oopDesc::load_decode_heap_oop(p);
5403 5423
5404 5424 if (_g1h->is_in_cset_or_humongous(obj)) {
5405 5425 // If the referent object has been forwarded (either copied
5406 5426 // to a new location or to itself in the event of an
5407 5427 // evacuation failure) then we need to update the reference
5408 5428 // field and, if both reference and referent are in the G1
5409 5429 // heap, update the RSet for the referent.
5410 5430 //
5411 5431 // If the referent has not been forwarded then we have to keep
5412 5432 // it alive by policy. Therefore we have copy the referent.
5413 5433 //
5414 5434 // If the reference field is in the G1 heap then we can push
5415 5435 // on the PSS queue. When the queue is drained (after each
5416 5436 // phase of reference processing) the object and it's followers
5417 5437 // will be copied, the reference field set to point to the
5418 5438 // new location, and the RSet updated. Otherwise we need to
5419 5439 // use the the non-heap or metadata closures directly to copy
5420 5440 // the referent object and update the pointer, while avoiding
5421 5441 // updating the RSet.
5422 5442
5423 5443 if (_g1h->is_in_g1_reserved(p)) {
5424 5444 _par_scan_state->push_on_queue(p);
5425 5445 } else {
5426 5446 assert(!Metaspace::contains((const void*)p),
5427 5447 err_msg("Unexpectedly found a pointer from metadata: "
5428 5448 PTR_FORMAT, p));
5429 5449 _copy_non_heap_obj_cl->do_oop(p);
5430 5450 }
5431 5451 }
5432 5452 }
5433 5453 };
5434 5454
5435 5455 // Serial drain queue closure. Called as the 'complete_gc'
5436 5456 // closure for each discovered list in some of the
5437 5457 // reference processing phases.
5438 5458
5439 5459 class G1STWDrainQueueClosure: public VoidClosure {
5440 5460 protected:
5441 5461 G1CollectedHeap* _g1h;
5442 5462 G1ParScanThreadState* _par_scan_state;
5443 5463
5444 5464 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5445 5465
5446 5466 public:
5447 5467 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5448 5468 _g1h(g1h),
5449 5469 _par_scan_state(pss)
5450 5470 { }
5451 5471
5452 5472 void do_void() {
5453 5473 G1ParScanThreadState* const pss = par_scan_state();
5454 5474 pss->trim_queue();
5455 5475 }
5456 5476 };
5457 5477
5458 5478 // Parallel Reference Processing closures
5459 5479
5460 5480 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5461 5481 // processing during G1 evacuation pauses.
5462 5482
5463 5483 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5464 5484 private:
5465 5485 G1CollectedHeap* _g1h;
5466 5486 RefToScanQueueSet* _queues;
5467 5487 FlexibleWorkGang* _workers;
5468 5488 int _active_workers;
5469 5489
5470 5490 public:
5471 5491 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5472 5492 FlexibleWorkGang* workers,
5473 5493 RefToScanQueueSet *task_queues,
5474 5494 int n_workers) :
5475 5495 _g1h(g1h),
5476 5496 _queues(task_queues),
5477 5497 _workers(workers),
5478 5498 _active_workers(n_workers)
5479 5499 {
5480 5500 assert(n_workers > 0, "shouldn't call this otherwise");
5481 5501 }
5482 5502
5483 5503 // Executes the given task using concurrent marking worker threads.
5484 5504 virtual void execute(ProcessTask& task);
5485 5505 virtual void execute(EnqueueTask& task);
5486 5506 };
5487 5507
5488 5508 // Gang task for possibly parallel reference processing
5489 5509
5490 5510 class G1STWRefProcTaskProxy: public AbstractGangTask {
5491 5511 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5492 5512 ProcessTask& _proc_task;
5493 5513 G1CollectedHeap* _g1h;
5494 5514 RefToScanQueueSet *_task_queues;
5495 5515 ParallelTaskTerminator* _terminator;
5496 5516
5497 5517 public:
5498 5518 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5499 5519 G1CollectedHeap* g1h,
5500 5520 RefToScanQueueSet *task_queues,
5501 5521 ParallelTaskTerminator* terminator) :
5502 5522 AbstractGangTask("Process reference objects in parallel"),
5503 5523 _proc_task(proc_task),
5504 5524 _g1h(g1h),
5505 5525 _task_queues(task_queues),
5506 5526 _terminator(terminator)
5507 5527 {}
5508 5528
5509 5529 virtual void work(uint worker_id) {
5510 5530 // The reference processing task executed by a single worker.
5511 5531 ResourceMark rm;
5512 5532 HandleMark hm;
5513 5533
5514 5534 G1STWIsAliveClosure is_alive(_g1h);
5515 5535
5516 5536 G1ParScanThreadState pss(_g1h, worker_id, NULL);
5517 5537 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5518 5538
5519 5539 pss.set_evac_failure_closure(&evac_failure_cl);
5520 5540
5521 5541 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5522 5542
5523 5543 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5524 5544
5525 5545 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5526 5546
5527 5547 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5528 5548 // We also need to mark copied objects.
5529 5549 copy_non_heap_cl = ©_mark_non_heap_cl;
5530 5550 }
5531 5551
5532 5552 // Keep alive closure.
5533 5553 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, &pss);
5534 5554
5535 5555 // Complete GC closure
5536 5556 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5537 5557
5538 5558 // Call the reference processing task's work routine.
5539 5559 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5540 5560
5541 5561 // Note we cannot assert that the refs array is empty here as not all
5542 5562 // of the processing tasks (specifically phase2 - pp2_work) execute
5543 5563 // the complete_gc closure (which ordinarily would drain the queue) so
5544 5564 // the queue may not be empty.
5545 5565 }
5546 5566 };
5547 5567
5548 5568 // Driver routine for parallel reference processing.
5549 5569 // Creates an instance of the ref processing gang
5550 5570 // task and has the worker threads execute it.
5551 5571 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5552 5572 assert(_workers != NULL, "Need parallel worker threads.");
5553 5573
5554 5574 ParallelTaskTerminator terminator(_active_workers, _queues);
5555 5575 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5556 5576
5557 5577 _g1h->set_par_threads(_active_workers);
5558 5578 _workers->run_task(&proc_task_proxy);
5559 5579 _g1h->set_par_threads(0);
5560 5580 }
5561 5581
5562 5582 // Gang task for parallel reference enqueueing.
5563 5583
5564 5584 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5565 5585 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5566 5586 EnqueueTask& _enq_task;
5567 5587
5568 5588 public:
5569 5589 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5570 5590 AbstractGangTask("Enqueue reference objects in parallel"),
5571 5591 _enq_task(enq_task)
5572 5592 { }
5573 5593
5574 5594 virtual void work(uint worker_id) {
5575 5595 _enq_task.work(worker_id);
5576 5596 }
5577 5597 };
5578 5598
5579 5599 // Driver routine for parallel reference enqueueing.
5580 5600 // Creates an instance of the ref enqueueing gang
5581 5601 // task and has the worker threads execute it.
5582 5602
5583 5603 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5584 5604 assert(_workers != NULL, "Need parallel worker threads.");
5585 5605
5586 5606 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5587 5607
5588 5608 _g1h->set_par_threads(_active_workers);
5589 5609 _workers->run_task(&enq_task_proxy);
5590 5610 _g1h->set_par_threads(0);
5591 5611 }
5592 5612
5593 5613 // End of weak reference support closures
5594 5614
5595 5615 // Abstract task used to preserve (i.e. copy) any referent objects
5596 5616 // that are in the collection set and are pointed to by reference
5597 5617 // objects discovered by the CM ref processor.
5598 5618
5599 5619 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5600 5620 protected:
5601 5621 G1CollectedHeap* _g1h;
5602 5622 RefToScanQueueSet *_queues;
5603 5623 ParallelTaskTerminator _terminator;
5604 5624 uint _n_workers;
5605 5625
5606 5626 public:
5607 5627 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5608 5628 AbstractGangTask("ParPreserveCMReferents"),
5609 5629 _g1h(g1h),
5610 5630 _queues(task_queues),
5611 5631 _terminator(workers, _queues),
5612 5632 _n_workers(workers)
5613 5633 { }
5614 5634
5615 5635 void work(uint worker_id) {
5616 5636 ResourceMark rm;
5617 5637 HandleMark hm;
5618 5638
5619 5639 G1ParScanThreadState pss(_g1h, worker_id, NULL);
5620 5640 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5621 5641
5622 5642 pss.set_evac_failure_closure(&evac_failure_cl);
5623 5643
5624 5644 assert(pss.queue_is_empty(), "both queue and overflow should be empty");
5625 5645
5626 5646 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5627 5647
5628 5648 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5629 5649
5630 5650 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5631 5651
5632 5652 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5633 5653 // We also need to mark copied objects.
5634 5654 copy_non_heap_cl = ©_mark_non_heap_cl;
5635 5655 }
5636 5656
5637 5657 // Is alive closure
5638 5658 G1AlwaysAliveClosure always_alive(_g1h);
5639 5659
5640 5660 // Copying keep alive closure. Applied to referent objects that need
5641 5661 // to be copied.
5642 5662 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, &pss);
5643 5663
5644 5664 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5645 5665
5646 5666 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5647 5667 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5648 5668
5649 5669 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5650 5670 // So this must be true - but assert just in case someone decides to
5651 5671 // change the worker ids.
5652 5672 assert(0 <= worker_id && worker_id < limit, "sanity");
5653 5673 assert(!rp->discovery_is_atomic(), "check this code");
5654 5674
5655 5675 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5656 5676 for (uint idx = worker_id; idx < limit; idx += stride) {
5657 5677 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5658 5678
5659 5679 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5660 5680 while (iter.has_next()) {
5661 5681 // Since discovery is not atomic for the CM ref processor, we
5662 5682 // can see some null referent objects.
5663 5683 iter.load_ptrs(DEBUG_ONLY(true));
5664 5684 oop ref = iter.obj();
5665 5685
5666 5686 // This will filter nulls.
5667 5687 if (iter.is_referent_alive()) {
5668 5688 iter.make_referent_alive();
5669 5689 }
5670 5690 iter.move_to_next();
5671 5691 }
5672 5692 }
5673 5693
5674 5694 // Drain the queue - which may cause stealing
5675 5695 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5676 5696 drain_queue.do_void();
5677 5697 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5678 5698 assert(pss.queue_is_empty(), "should be");
5679 5699 }
5680 5700 };
5681 5701
5682 5702 // Weak Reference processing during an evacuation pause (part 1).
5683 5703 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5684 5704 double ref_proc_start = os::elapsedTime();
5685 5705
5686 5706 ReferenceProcessor* rp = _ref_processor_stw;
5687 5707 assert(rp->discovery_enabled(), "should have been enabled");
5688 5708
5689 5709 // Any reference objects, in the collection set, that were 'discovered'
5690 5710 // by the CM ref processor should have already been copied (either by
5691 5711 // applying the external root copy closure to the discovered lists, or
5692 5712 // by following an RSet entry).
5693 5713 //
5694 5714 // But some of the referents, that are in the collection set, that these
5695 5715 // reference objects point to may not have been copied: the STW ref
5696 5716 // processor would have seen that the reference object had already
5697 5717 // been 'discovered' and would have skipped discovering the reference,
5698 5718 // but would not have treated the reference object as a regular oop.
5699 5719 // As a result the copy closure would not have been applied to the
5700 5720 // referent object.
5701 5721 //
5702 5722 // We need to explicitly copy these referent objects - the references
5703 5723 // will be processed at the end of remarking.
5704 5724 //
5705 5725 // We also need to do this copying before we process the reference
5706 5726 // objects discovered by the STW ref processor in case one of these
5707 5727 // referents points to another object which is also referenced by an
5708 5728 // object discovered by the STW ref processor.
5709 5729
5710 5730 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5711 5731 no_of_gc_workers == workers()->active_workers(),
5712 5732 "Need to reset active GC workers");
5713 5733
5714 5734 set_par_threads(no_of_gc_workers);
5715 5735 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5716 5736 no_of_gc_workers,
5717 5737 _task_queues);
5718 5738
5719 5739 if (G1CollectedHeap::use_parallel_gc_threads()) {
5720 5740 workers()->run_task(&keep_cm_referents);
5721 5741 } else {
5722 5742 keep_cm_referents.work(0);
5723 5743 }
5724 5744
5725 5745 set_par_threads(0);
5726 5746
5727 5747 // Closure to test whether a referent is alive.
5728 5748 G1STWIsAliveClosure is_alive(this);
5729 5749
5730 5750 // Even when parallel reference processing is enabled, the processing
5731 5751 // of JNI refs is serial and performed serially by the current thread
5732 5752 // rather than by a worker. The following PSS will be used for processing
5733 5753 // JNI refs.
5734 5754
5735 5755 // Use only a single queue for this PSS.
5736 5756 G1ParScanThreadState pss(this, 0, NULL);
5737 5757
5738 5758 // We do not embed a reference processor in the copying/scanning
5739 5759 // closures while we're actually processing the discovered
5740 5760 // reference objects.
5741 5761 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5742 5762
5743 5763 pss.set_evac_failure_closure(&evac_failure_cl);
5744 5764
5745 5765 assert(pss.queue_is_empty(), "pre-condition");
5746 5766
5747 5767 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5748 5768
5749 5769 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5750 5770
5751 5771 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5752 5772
5753 5773 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5754 5774 // We also need to mark copied objects.
5755 5775 copy_non_heap_cl = ©_mark_non_heap_cl;
5756 5776 }
5757 5777
5758 5778 // Keep alive closure.
5759 5779 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, &pss);
5760 5780
5761 5781 // Serial Complete GC closure
5762 5782 G1STWDrainQueueClosure drain_queue(this, &pss);
5763 5783
5764 5784 // Setup the soft refs policy...
5765 5785 rp->setup_policy(false);
5766 5786
5767 5787 ReferenceProcessorStats stats;
5768 5788 if (!rp->processing_is_mt()) {
5769 5789 // Serial reference processing...
5770 5790 stats = rp->process_discovered_references(&is_alive,
5771 5791 &keep_alive,
5772 5792 &drain_queue,
5773 5793 NULL,
5774 5794 _gc_timer_stw,
5775 5795 _gc_tracer_stw->gc_id());
5776 5796 } else {
5777 5797 // Parallel reference processing
5778 5798 assert(rp->num_q() == no_of_gc_workers, "sanity");
5779 5799 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5780 5800
5781 5801 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5782 5802 stats = rp->process_discovered_references(&is_alive,
5783 5803 &keep_alive,
5784 5804 &drain_queue,
5785 5805 &par_task_executor,
5786 5806 _gc_timer_stw,
5787 5807 _gc_tracer_stw->gc_id());
5788 5808 }
5789 5809
5790 5810 _gc_tracer_stw->report_gc_reference_stats(stats);
5791 5811
5792 5812 // We have completed copying any necessary live referent objects.
5793 5813 assert(pss.queue_is_empty(), "both queue and overflow should be empty");
5794 5814
5795 5815 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5796 5816 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5797 5817 }
5798 5818
5799 5819 // Weak Reference processing during an evacuation pause (part 2).
5800 5820 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5801 5821 double ref_enq_start = os::elapsedTime();
5802 5822
5803 5823 ReferenceProcessor* rp = _ref_processor_stw;
5804 5824 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5805 5825
5806 5826 // Now enqueue any remaining on the discovered lists on to
5807 5827 // the pending list.
5808 5828 if (!rp->processing_is_mt()) {
5809 5829 // Serial reference processing...
5810 5830 rp->enqueue_discovered_references();
5811 5831 } else {
5812 5832 // Parallel reference enqueueing
5813 5833
5814 5834 assert(no_of_gc_workers == workers()->active_workers(),
5815 5835 "Need to reset active workers");
5816 5836 assert(rp->num_q() == no_of_gc_workers, "sanity");
5817 5837 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5818 5838
5819 5839 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5820 5840 rp->enqueue_discovered_references(&par_task_executor);
5821 5841 }
5822 5842
5823 5843 rp->verify_no_references_recorded();
5824 5844 assert(!rp->discovery_enabled(), "should have been disabled");
5825 5845
5826 5846 // FIXME
5827 5847 // CM's reference processing also cleans up the string and symbol tables.
5828 5848 // Should we do that here also? We could, but it is a serial operation
5829 5849 // and could significantly increase the pause time.
5830 5850
5831 5851 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5832 5852 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5833 5853 }
5834 5854
5835 5855 void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) {
5836 5856 _expand_heap_after_alloc_failure = true;
5837 5857 _evacuation_failed = false;
5838 5858
5839 5859 // Should G1EvacuationFailureALot be in effect for this GC?
5840 5860 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5841 5861
5842 5862 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5843 5863
5844 5864 // Disable the hot card cache.
5845 5865 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
5846 5866 hot_card_cache->reset_hot_cache_claimed_index();
5847 5867 hot_card_cache->set_use_cache(false);
5848 5868
5849 5869 uint n_workers;
5850 5870 if (G1CollectedHeap::use_parallel_gc_threads()) {
5851 5871 n_workers =
5852 5872 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5853 5873 workers()->active_workers(),
5854 5874 Threads::number_of_non_daemon_threads());
5855 5875 assert(UseDynamicNumberOfGCThreads ||
5856 5876 n_workers == workers()->total_workers(),
5857 5877 "If not dynamic should be using all the workers");
5858 5878 workers()->set_active_workers(n_workers);
5859 5879 set_par_threads(n_workers);
5860 5880 } else {
5861 5881 assert(n_par_threads() == 0,
5862 5882 "Should be the original non-parallel value");
5863 5883 n_workers = 1;
5864 5884 }
5865 5885
5866 5886 G1ParTask g1_par_task(this, _task_queues);
5867 5887
5868 5888 init_for_evac_failure(NULL);
5869 5889
5870 5890 rem_set()->prepare_for_younger_refs_iterate(true);
5871 5891
5872 5892 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5873 5893 double start_par_time_sec = os::elapsedTime();
5874 5894 double end_par_time_sec;
5875 5895
5876 5896 {
5877 5897 StrongRootsScope srs(this);
5878 5898 // InitialMark needs claim bits to keep track of the marked-through CLDs.
5879 5899 if (g1_policy()->during_initial_mark_pause()) {
5880 5900 ClassLoaderDataGraph::clear_claimed_marks();
5881 5901 }
5882 5902
5883 5903 if (G1CollectedHeap::use_parallel_gc_threads()) {
5884 5904 // The individual threads will set their evac-failure closures.
5885 5905 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5886 5906 // These tasks use ShareHeap::_process_strong_tasks
5887 5907 assert(UseDynamicNumberOfGCThreads ||
5888 5908 workers()->active_workers() == workers()->total_workers(),
5889 5909 "If not dynamic should be using all the workers");
5890 5910 workers()->run_task(&g1_par_task);
5891 5911 } else {
5892 5912 g1_par_task.set_for_termination(n_workers);
5893 5913 g1_par_task.work(0);
5894 5914 }
5895 5915 end_par_time_sec = os::elapsedTime();
5896 5916
5897 5917 // Closing the inner scope will execute the destructor
5898 5918 // for the StrongRootsScope object. We record the current
5899 5919 // elapsed time before closing the scope so that time
5900 5920 // taken for the SRS destructor is NOT included in the
5901 5921 // reported parallel time.
5902 5922 }
5903 5923
5904 5924 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5905 5925 g1_policy()->phase_times()->record_par_time(par_time_ms);
5906 5926
5907 5927 double code_root_fixup_time_ms =
5908 5928 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5909 5929 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5910 5930
5911 5931 set_par_threads(0);
5912 5932
5913 5933 // Process any discovered reference objects - we have
5914 5934 // to do this _before_ we retire the GC alloc regions
5915 5935 // as we may have to copy some 'reachable' referent
5916 5936 // objects (and their reachable sub-graphs) that were
5917 5937 // not copied during the pause.
5918 5938 process_discovered_references(n_workers);
5919 5939
5920 5940 // Weak root processing.
5921 5941 {
5922 5942 G1STWIsAliveClosure is_alive(this);
5923 5943 G1KeepAliveClosure keep_alive(this);
5924 5944 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5925 5945 if (G1StringDedup::is_enabled()) {
5926 5946 G1StringDedup::unlink_or_oops_do(&is_alive, &keep_alive);
5927 5947 }
5928 5948 }
5929 5949
5930 5950 _allocator->release_gc_alloc_regions(n_workers, evacuation_info);
5931 5951 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5932 5952
5933 5953 // Reset and re-enable the hot card cache.
5934 5954 // Note the counts for the cards in the regions in the
5935 5955 // collection set are reset when the collection set is freed.
5936 5956 hot_card_cache->reset_hot_cache();
5937 5957 hot_card_cache->set_use_cache(true);
5938 5958
5939 5959 purge_code_root_memory();
5940 5960
5941 5961 if (g1_policy()->during_initial_mark_pause()) {
5942 5962 // Reset the claim values set during marking the strong code roots
5943 5963 reset_heap_region_claim_values();
5944 5964 }
5945 5965
5946 5966 finalize_for_evac_failure();
5947 5967
5948 5968 if (evacuation_failed()) {
5949 5969 remove_self_forwarding_pointers();
5950 5970
5951 5971 // Reset the G1EvacuationFailureALot counters and flags
5952 5972 // Note: the values are reset only when an actual
5953 5973 // evacuation failure occurs.
5954 5974 NOT_PRODUCT(reset_evacuation_should_fail();)
5955 5975 }
5956 5976
5957 5977 // Enqueue any remaining references remaining on the STW
5958 5978 // reference processor's discovered lists. We need to do
5959 5979 // this after the card table is cleaned (and verified) as
5960 5980 // the act of enqueueing entries on to the pending list
5961 5981 // will log these updates (and dirty their associated
5962 5982 // cards). We need these updates logged to update any
5963 5983 // RSets.
5964 5984 enqueue_discovered_references(n_workers);
5965 5985
5966 5986 redirty_logged_cards();
5967 5987 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5968 5988 }
5969 5989
5970 5990 void G1CollectedHeap::free_region(HeapRegion* hr,
5971 5991 FreeRegionList* free_list,
5972 5992 bool par,
5973 5993 bool locked) {
5974 5994 assert(!hr->is_free(), "the region should not be free");
5975 5995 assert(!hr->is_empty(), "the region should not be empty");
5976 5996 assert(_hrm.is_available(hr->hrm_index()), "region should be committed");
5977 5997 assert(free_list != NULL, "pre-condition");
5978 5998
5979 5999 if (G1VerifyBitmaps) {
5980 6000 MemRegion mr(hr->bottom(), hr->end());
5981 6001 concurrent_mark()->clearRangePrevBitmap(mr);
5982 6002 }
5983 6003
5984 6004 // Clear the card counts for this region.
5985 6005 // Note: we only need to do this if the region is not young
5986 6006 // (since we don't refine cards in young regions).
5987 6007 if (!hr->is_young()) {
5988 6008 _cg1r->hot_card_cache()->reset_card_counts(hr);
5989 6009 }
5990 6010 hr->hr_clear(par, true /* clear_space */, locked /* locked */);
5991 6011 free_list->add_ordered(hr);
5992 6012 }
5993 6013
5994 6014 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5995 6015 FreeRegionList* free_list,
5996 6016 bool par) {
5997 6017 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5998 6018 assert(free_list != NULL, "pre-condition");
5999 6019
6000 6020 size_t hr_capacity = hr->capacity();
6001 6021 // We need to read this before we make the region non-humongous,
6002 6022 // otherwise the information will be gone.
6003 6023 uint last_index = hr->last_hc_index();
6004 6024 hr->clear_humongous();
6005 6025 free_region(hr, free_list, par);
6006 6026
6007 6027 uint i = hr->hrm_index() + 1;
6008 6028 while (i < last_index) {
6009 6029 HeapRegion* curr_hr = region_at(i);
6010 6030 assert(curr_hr->continuesHumongous(), "invariant");
6011 6031 curr_hr->clear_humongous();
6012 6032 free_region(curr_hr, free_list, par);
6013 6033 i += 1;
6014 6034 }
6015 6035 }
6016 6036
6017 6037 void G1CollectedHeap::remove_from_old_sets(const HeapRegionSetCount& old_regions_removed,
6018 6038 const HeapRegionSetCount& humongous_regions_removed) {
6019 6039 if (old_regions_removed.length() > 0 || humongous_regions_removed.length() > 0) {
6020 6040 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
6021 6041 _old_set.bulk_remove(old_regions_removed);
6022 6042 _humongous_set.bulk_remove(humongous_regions_removed);
6023 6043 }
6024 6044
6025 6045 }
6026 6046
6027 6047 void G1CollectedHeap::prepend_to_freelist(FreeRegionList* list) {
6028 6048 assert(list != NULL, "list can't be null");
6029 6049 if (!list->is_empty()) {
6030 6050 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
6031 6051 _hrm.insert_list_into_free_list(list);
6032 6052 }
6033 6053 }
6034 6054
6035 6055 void G1CollectedHeap::decrement_summary_bytes(size_t bytes) {
6036 6056 _allocator->decrease_used(bytes);
6037 6057 }
6038 6058
6039 6059 class G1ParCleanupCTTask : public AbstractGangTask {
6040 6060 G1SATBCardTableModRefBS* _ct_bs;
6041 6061 G1CollectedHeap* _g1h;
6042 6062 HeapRegion* volatile _su_head;
6043 6063 public:
6044 6064 G1ParCleanupCTTask(G1SATBCardTableModRefBS* ct_bs,
6045 6065 G1CollectedHeap* g1h) :
6046 6066 AbstractGangTask("G1 Par Cleanup CT Task"),
6047 6067 _ct_bs(ct_bs), _g1h(g1h) { }
6048 6068
6049 6069 void work(uint worker_id) {
6050 6070 HeapRegion* r;
6051 6071 while (r = _g1h->pop_dirty_cards_region()) {
6052 6072 clear_cards(r);
6053 6073 }
6054 6074 }
6055 6075
6056 6076 void clear_cards(HeapRegion* r) {
6057 6077 // Cards of the survivors should have already been dirtied.
6058 6078 if (!r->is_survivor()) {
6059 6079 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
6060 6080 }
6061 6081 }
6062 6082 };
6063 6083
6064 6084 #ifndef PRODUCT
6065 6085 class G1VerifyCardTableCleanup: public HeapRegionClosure {
6066 6086 G1CollectedHeap* _g1h;
6067 6087 G1SATBCardTableModRefBS* _ct_bs;
6068 6088 public:
6069 6089 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, G1SATBCardTableModRefBS* ct_bs)
6070 6090 : _g1h(g1h), _ct_bs(ct_bs) { }
6071 6091 virtual bool doHeapRegion(HeapRegion* r) {
6072 6092 if (r->is_survivor()) {
6073 6093 _g1h->verify_dirty_region(r);
6074 6094 } else {
6075 6095 _g1h->verify_not_dirty_region(r);
6076 6096 }
6077 6097 return false;
6078 6098 }
6079 6099 };
6080 6100
6081 6101 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
6082 6102 // All of the region should be clean.
6083 6103 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set();
6084 6104 MemRegion mr(hr->bottom(), hr->end());
6085 6105 ct_bs->verify_not_dirty_region(mr);
6086 6106 }
6087 6107
6088 6108 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
6089 6109 // We cannot guarantee that [bottom(),end()] is dirty. Threads
6090 6110 // dirty allocated blocks as they allocate them. The thread that
6091 6111 // retires each region and replaces it with a new one will do a
6092 6112 // maximal allocation to fill in [pre_dummy_top(),end()] but will
6093 6113 // not dirty that area (one less thing to have to do while holding
6094 6114 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
6095 6115 // is dirty.
6096 6116 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set();
6097 6117 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
6098 6118 if (hr->is_young()) {
6099 6119 ct_bs->verify_g1_young_region(mr);
6100 6120 } else {
6101 6121 ct_bs->verify_dirty_region(mr);
6102 6122 }
6103 6123 }
6104 6124
6105 6125 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
6106 6126 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set();
6107 6127 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
6108 6128 verify_dirty_region(hr);
6109 6129 }
6110 6130 }
6111 6131
6112 6132 void G1CollectedHeap::verify_dirty_young_regions() {
6113 6133 verify_dirty_young_list(_young_list->first_region());
6114 6134 }
6115 6135
6116 6136 bool G1CollectedHeap::verify_no_bits_over_tams(const char* bitmap_name, CMBitMapRO* bitmap,
6117 6137 HeapWord* tams, HeapWord* end) {
6118 6138 guarantee(tams <= end,
6119 6139 err_msg("tams: "PTR_FORMAT" end: "PTR_FORMAT, tams, end));
6120 6140 HeapWord* result = bitmap->getNextMarkedWordAddress(tams, end);
6121 6141 if (result < end) {
6122 6142 gclog_or_tty->cr();
6123 6143 gclog_or_tty->print_cr("## wrong marked address on %s bitmap: "PTR_FORMAT,
6124 6144 bitmap_name, result);
6125 6145 gclog_or_tty->print_cr("## %s tams: "PTR_FORMAT" end: "PTR_FORMAT,
6126 6146 bitmap_name, tams, end);
6127 6147 return false;
6128 6148 }
6129 6149 return true;
6130 6150 }
6131 6151
6132 6152 bool G1CollectedHeap::verify_bitmaps(const char* caller, HeapRegion* hr) {
6133 6153 CMBitMapRO* prev_bitmap = concurrent_mark()->prevMarkBitMap();
6134 6154 CMBitMapRO* next_bitmap = (CMBitMapRO*) concurrent_mark()->nextMarkBitMap();
6135 6155
6136 6156 HeapWord* bottom = hr->bottom();
6137 6157 HeapWord* ptams = hr->prev_top_at_mark_start();
6138 6158 HeapWord* ntams = hr->next_top_at_mark_start();
6139 6159 HeapWord* end = hr->end();
6140 6160
6141 6161 bool res_p = verify_no_bits_over_tams("prev", prev_bitmap, ptams, end);
6142 6162
6143 6163 bool res_n = true;
6144 6164 // We reset mark_in_progress() before we reset _cmThread->in_progress() and in this window
6145 6165 // we do the clearing of the next bitmap concurrently. Thus, we can not verify the bitmap
6146 6166 // if we happen to be in that state.
6147 6167 if (mark_in_progress() || !_cmThread->in_progress()) {
6148 6168 res_n = verify_no_bits_over_tams("next", next_bitmap, ntams, end);
6149 6169 }
6150 6170 if (!res_p || !res_n) {
6151 6171 gclog_or_tty->print_cr("#### Bitmap verification failed for "HR_FORMAT,
6152 6172 HR_FORMAT_PARAMS(hr));
6153 6173 gclog_or_tty->print_cr("#### Caller: %s", caller);
6154 6174 return false;
6155 6175 }
6156 6176 return true;
6157 6177 }
6158 6178
6159 6179 void G1CollectedHeap::check_bitmaps(const char* caller, HeapRegion* hr) {
6160 6180 if (!G1VerifyBitmaps) return;
6161 6181
6162 6182 guarantee(verify_bitmaps(caller, hr), "bitmap verification");
6163 6183 }
6164 6184
6165 6185 class G1VerifyBitmapClosure : public HeapRegionClosure {
6166 6186 private:
6167 6187 const char* _caller;
6168 6188 G1CollectedHeap* _g1h;
6169 6189 bool _failures;
6170 6190
6171 6191 public:
6172 6192 G1VerifyBitmapClosure(const char* caller, G1CollectedHeap* g1h) :
6173 6193 _caller(caller), _g1h(g1h), _failures(false) { }
6174 6194
6175 6195 bool failures() { return _failures; }
6176 6196
6177 6197 virtual bool doHeapRegion(HeapRegion* hr) {
6178 6198 if (hr->continuesHumongous()) return false;
6179 6199
6180 6200 bool result = _g1h->verify_bitmaps(_caller, hr);
6181 6201 if (!result) {
6182 6202 _failures = true;
6183 6203 }
6184 6204 return false;
6185 6205 }
6186 6206 };
6187 6207
6188 6208 void G1CollectedHeap::check_bitmaps(const char* caller) {
6189 6209 if (!G1VerifyBitmaps) return;
6190 6210
6191 6211 G1VerifyBitmapClosure cl(caller, this);
6192 6212 heap_region_iterate(&cl);
6193 6213 guarantee(!cl.failures(), "bitmap verification");
6194 6214 }
6195 6215 #endif // PRODUCT
6196 6216
6197 6217 void G1CollectedHeap::cleanUpCardTable() {
6198 6218 G1SATBCardTableModRefBS* ct_bs = g1_barrier_set();
6199 6219 double start = os::elapsedTime();
6200 6220
6201 6221 {
6202 6222 // Iterate over the dirty cards region list.
6203 6223 G1ParCleanupCTTask cleanup_task(ct_bs, this);
6204 6224
6205 6225 if (G1CollectedHeap::use_parallel_gc_threads()) {
6206 6226 set_par_threads();
6207 6227 workers()->run_task(&cleanup_task);
6208 6228 set_par_threads(0);
6209 6229 } else {
6210 6230 while (_dirty_cards_region_list) {
6211 6231 HeapRegion* r = _dirty_cards_region_list;
6212 6232 cleanup_task.clear_cards(r);
6213 6233 _dirty_cards_region_list = r->get_next_dirty_cards_region();
6214 6234 if (_dirty_cards_region_list == r) {
6215 6235 // The last region.
6216 6236 _dirty_cards_region_list = NULL;
6217 6237 }
6218 6238 r->set_next_dirty_cards_region(NULL);
6219 6239 }
6220 6240 }
6221 6241 #ifndef PRODUCT
6222 6242 if (G1VerifyCTCleanup || VerifyAfterGC) {
6223 6243 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
6224 6244 heap_region_iterate(&cleanup_verifier);
6225 6245 }
6226 6246 #endif
6227 6247 }
6228 6248
6229 6249 double elapsed = os::elapsedTime() - start;
6230 6250 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
6231 6251 }
6232 6252
6233 6253 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head, EvacuationInfo& evacuation_info) {
6234 6254 size_t pre_used = 0;
6235 6255 FreeRegionList local_free_list("Local List for CSet Freeing");
6236 6256
6237 6257 double young_time_ms = 0.0;
6238 6258 double non_young_time_ms = 0.0;
6239 6259
6240 6260 // Since the collection set is a superset of the the young list,
6241 6261 // all we need to do to clear the young list is clear its
6242 6262 // head and length, and unlink any young regions in the code below
6243 6263 _young_list->clear();
6244 6264
6245 6265 G1CollectorPolicy* policy = g1_policy();
6246 6266
6247 6267 double start_sec = os::elapsedTime();
6248 6268 bool non_young = true;
6249 6269
6250 6270 HeapRegion* cur = cs_head;
6251 6271 int age_bound = -1;
6252 6272 size_t rs_lengths = 0;
6253 6273
6254 6274 while (cur != NULL) {
6255 6275 assert(!is_on_master_free_list(cur), "sanity");
6256 6276 if (non_young) {
6257 6277 if (cur->is_young()) {
6258 6278 double end_sec = os::elapsedTime();
6259 6279 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6260 6280 non_young_time_ms += elapsed_ms;
6261 6281
6262 6282 start_sec = os::elapsedTime();
6263 6283 non_young = false;
6264 6284 }
6265 6285 } else {
6266 6286 if (!cur->is_young()) {
6267 6287 double end_sec = os::elapsedTime();
6268 6288 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6269 6289 young_time_ms += elapsed_ms;
6270 6290
6271 6291 start_sec = os::elapsedTime();
6272 6292 non_young = true;
6273 6293 }
6274 6294 }
6275 6295
6276 6296 rs_lengths += cur->rem_set()->occupied_locked();
6277 6297
6278 6298 HeapRegion* next = cur->next_in_collection_set();
6279 6299 assert(cur->in_collection_set(), "bad CS");
6280 6300 cur->set_next_in_collection_set(NULL);
6281 6301 cur->set_in_collection_set(false);
6282 6302
6283 6303 if (cur->is_young()) {
6284 6304 int index = cur->young_index_in_cset();
6285 6305 assert(index != -1, "invariant");
6286 6306 assert((uint) index < policy->young_cset_region_length(), "invariant");
6287 6307 size_t words_survived = _surviving_young_words[index];
6288 6308 cur->record_surv_words_in_group(words_survived);
6289 6309
6290 6310 // At this point the we have 'popped' cur from the collection set
6291 6311 // (linked via next_in_collection_set()) but it is still in the
6292 6312 // young list (linked via next_young_region()). Clear the
6293 6313 // _next_young_region field.
6294 6314 cur->set_next_young_region(NULL);
6295 6315 } else {
6296 6316 int index = cur->young_index_in_cset();
6297 6317 assert(index == -1, "invariant");
6298 6318 }
6299 6319
6300 6320 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
6301 6321 (!cur->is_young() && cur->young_index_in_cset() == -1),
6302 6322 "invariant" );
6303 6323
6304 6324 if (!cur->evacuation_failed()) {
6305 6325 MemRegion used_mr = cur->used_region();
6306 6326
6307 6327 // And the region is empty.
6308 6328 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
6309 6329 pre_used += cur->used();
6310 6330 free_region(cur, &local_free_list, false /* par */, true /* locked */);
6311 6331 } else {
6312 6332 cur->uninstall_surv_rate_group();
6313 6333 if (cur->is_young()) {
6314 6334 cur->set_young_index_in_cset(-1);
6315 6335 }
6316 6336 cur->set_evacuation_failed(false);
6317 6337 // The region is now considered to be old.
6318 6338 cur->set_old();
6319 6339 _old_set.add(cur);
6320 6340 evacuation_info.increment_collectionset_used_after(cur->used());
6321 6341 }
6322 6342 cur = next;
6323 6343 }
6324 6344
6325 6345 evacuation_info.set_regions_freed(local_free_list.length());
6326 6346 policy->record_max_rs_lengths(rs_lengths);
6327 6347 policy->cset_regions_freed();
6328 6348
6329 6349 double end_sec = os::elapsedTime();
6330 6350 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6331 6351
6332 6352 if (non_young) {
6333 6353 non_young_time_ms += elapsed_ms;
6334 6354 } else {
6335 6355 young_time_ms += elapsed_ms;
6336 6356 }
6337 6357
6338 6358 prepend_to_freelist(&local_free_list);
6339 6359 decrement_summary_bytes(pre_used);
6340 6360 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6341 6361 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6342 6362 }
6343 6363
6344 6364 class G1FreeHumongousRegionClosure : public HeapRegionClosure {
6345 6365 private:
6346 6366 FreeRegionList* _free_region_list;
6347 6367 HeapRegionSet* _proxy_set;
6348 6368 HeapRegionSetCount _humongous_regions_removed;
6349 6369 size_t _freed_bytes;
6350 6370 public:
6351 6371
6352 6372 G1FreeHumongousRegionClosure(FreeRegionList* free_region_list) :
6353 6373 _free_region_list(free_region_list), _humongous_regions_removed(), _freed_bytes(0) {
6354 6374 }
6355 6375
6356 6376 virtual bool doHeapRegion(HeapRegion* r) {
6357 6377 if (!r->startsHumongous()) {
6358 6378 return false;
6359 6379 }
6360 6380
6361 6381 G1CollectedHeap* g1h = G1CollectedHeap::heap();
6362 6382
6363 6383 oop obj = (oop)r->bottom();
6364 6384 CMBitMap* next_bitmap = g1h->concurrent_mark()->nextMarkBitMap();
6365 6385
6366 6386 // The following checks whether the humongous object is live are sufficient.
6367 6387 // The main additional check (in addition to having a reference from the roots
6368 6388 // or the young gen) is whether the humongous object has a remembered set entry.
6369 6389 //
6370 6390 // A humongous object cannot be live if there is no remembered set for it
6371 6391 // because:
6372 6392 // - there can be no references from within humongous starts regions referencing
6373 6393 // the object because we never allocate other objects into them.
6374 6394 // (I.e. there are no intra-region references that may be missed by the
6375 6395 // remembered set)
6376 6396 // - as soon there is a remembered set entry to the humongous starts region
6377 6397 // (i.e. it has "escaped" to an old object) this remembered set entry will stay
6378 6398 // until the end of a concurrent mark.
6379 6399 //
6380 6400 // It is not required to check whether the object has been found dead by marking
6381 6401 // or not, in fact it would prevent reclamation within a concurrent cycle, as
6382 6402 // all objects allocated during that time are considered live.
6383 6403 // SATB marking is even more conservative than the remembered set.
6384 6404 // So if at this point in the collection there is no remembered set entry,
6385 6405 // nobody has a reference to it.
6386 6406 // At the start of collection we flush all refinement logs, and remembered sets
6387 6407 // are completely up-to-date wrt to references to the humongous object.
6388 6408 //
6389 6409 // Other implementation considerations:
6390 6410 // - never consider object arrays: while they are a valid target, they have not
6391 6411 // been observed to be used as temporary objects.
6392 6412 // - they would also pose considerable effort for cleaning up the the remembered
6393 6413 // sets.
6394 6414 // While this cleanup is not strictly necessary to be done (or done instantly),
6395 6415 // given that their occurrence is very low, this saves us this additional
6396 6416 // complexity.
6397 6417 uint region_idx = r->hrm_index();
6398 6418 if (g1h->humongous_is_live(region_idx) ||
6399 6419 g1h->humongous_region_is_always_live(region_idx)) {
6400 6420
6401 6421 if (G1TraceReclaimDeadHumongousObjectsAtYoungGC) {
6402 6422 gclog_or_tty->print_cr("Live humongous %d region %d with remset "SIZE_FORMAT" code roots "SIZE_FORMAT" is marked %d live-other %d obj array %d",
6403 6423 r->isHumongous(),
6404 6424 region_idx,
6405 6425 r->rem_set()->occupied(),
6406 6426 r->rem_set()->strong_code_roots_list_length(),
6407 6427 next_bitmap->isMarked(r->bottom()),
6408 6428 g1h->humongous_is_live(region_idx),
6409 6429 obj->is_objArray()
6410 6430 );
6411 6431 }
6412 6432
6413 6433 return false;
6414 6434 }
6415 6435
6416 6436 guarantee(!obj->is_objArray(),
6417 6437 err_msg("Eagerly reclaiming object arrays is not supported, but the object "PTR_FORMAT" is.",
6418 6438 r->bottom()));
6419 6439
6420 6440 if (G1TraceReclaimDeadHumongousObjectsAtYoungGC) {
6421 6441 gclog_or_tty->print_cr("Reclaim humongous region %d start "PTR_FORMAT" region %d length "UINT32_FORMAT" with remset "SIZE_FORMAT" code roots "SIZE_FORMAT" is marked %d live-other %d obj array %d",
6422 6442 r->isHumongous(),
6423 6443 r->bottom(),
6424 6444 region_idx,
6425 6445 r->region_num(),
6426 6446 r->rem_set()->occupied(),
6427 6447 r->rem_set()->strong_code_roots_list_length(),
6428 6448 next_bitmap->isMarked(r->bottom()),
6429 6449 g1h->humongous_is_live(region_idx),
6430 6450 obj->is_objArray()
6431 6451 );
6432 6452 }
6433 6453 // Need to clear mark bit of the humongous object if already set.
6434 6454 if (next_bitmap->isMarked(r->bottom())) {
6435 6455 next_bitmap->clear(r->bottom());
6436 6456 }
6437 6457 _freed_bytes += r->used();
6438 6458 r->set_containing_set(NULL);
6439 6459 _humongous_regions_removed.increment(1u, r->capacity());
6440 6460 g1h->free_humongous_region(r, _free_region_list, false);
6441 6461
6442 6462 return false;
6443 6463 }
6444 6464
6445 6465 HeapRegionSetCount& humongous_free_count() {
6446 6466 return _humongous_regions_removed;
6447 6467 }
6448 6468
6449 6469 size_t bytes_freed() const {
6450 6470 return _freed_bytes;
6451 6471 }
6452 6472
6453 6473 size_t humongous_reclaimed() const {
6454 6474 return _humongous_regions_removed.length();
6455 6475 }
6456 6476 };
6457 6477
6458 6478 void G1CollectedHeap::eagerly_reclaim_humongous_regions() {
6459 6479 assert_at_safepoint(true);
6460 6480
6461 6481 if (!G1ReclaimDeadHumongousObjectsAtYoungGC || !_has_humongous_reclaim_candidates) {
6462 6482 g1_policy()->phase_times()->record_fast_reclaim_humongous_time_ms(0.0, 0);
6463 6483 return;
6464 6484 }
6465 6485
6466 6486 double start_time = os::elapsedTime();
6467 6487
6468 6488 FreeRegionList local_cleanup_list("Local Humongous Cleanup List");
6469 6489
6470 6490 G1FreeHumongousRegionClosure cl(&local_cleanup_list);
6471 6491 heap_region_iterate(&cl);
6472 6492
6473 6493 HeapRegionSetCount empty_set;
6474 6494 remove_from_old_sets(empty_set, cl.humongous_free_count());
6475 6495
6476 6496 G1HRPrinter* hr_printer = _g1h->hr_printer();
6477 6497 if (hr_printer->is_active()) {
6478 6498 FreeRegionListIterator iter(&local_cleanup_list);
6479 6499 while (iter.more_available()) {
6480 6500 HeapRegion* hr = iter.get_next();
6481 6501 hr_printer->cleanup(hr);
6482 6502 }
6483 6503 }
6484 6504
6485 6505 prepend_to_freelist(&local_cleanup_list);
6486 6506 decrement_summary_bytes(cl.bytes_freed());
6487 6507
6488 6508 g1_policy()->phase_times()->record_fast_reclaim_humongous_time_ms((os::elapsedTime() - start_time) * 1000.0,
6489 6509 cl.humongous_reclaimed());
6490 6510 }
6491 6511
6492 6512 // This routine is similar to the above but does not record
6493 6513 // any policy statistics or update free lists; we are abandoning
6494 6514 // the current incremental collection set in preparation of a
6495 6515 // full collection. After the full GC we will start to build up
6496 6516 // the incremental collection set again.
6497 6517 // This is only called when we're doing a full collection
6498 6518 // and is immediately followed by the tearing down of the young list.
6499 6519
6500 6520 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6501 6521 HeapRegion* cur = cs_head;
6502 6522
6503 6523 while (cur != NULL) {
6504 6524 HeapRegion* next = cur->next_in_collection_set();
6505 6525 assert(cur->in_collection_set(), "bad CS");
6506 6526 cur->set_next_in_collection_set(NULL);
6507 6527 cur->set_in_collection_set(false);
6508 6528 cur->set_young_index_in_cset(-1);
6509 6529 cur = next;
6510 6530 }
6511 6531 }
6512 6532
6513 6533 void G1CollectedHeap::set_free_regions_coming() {
6514 6534 if (G1ConcRegionFreeingVerbose) {
6515 6535 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6516 6536 "setting free regions coming");
6517 6537 }
6518 6538
6519 6539 assert(!free_regions_coming(), "pre-condition");
6520 6540 _free_regions_coming = true;
6521 6541 }
6522 6542
6523 6543 void G1CollectedHeap::reset_free_regions_coming() {
6524 6544 assert(free_regions_coming(), "pre-condition");
6525 6545
6526 6546 {
6527 6547 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6528 6548 _free_regions_coming = false;
6529 6549 SecondaryFreeList_lock->notify_all();
6530 6550 }
6531 6551
6532 6552 if (G1ConcRegionFreeingVerbose) {
6533 6553 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6534 6554 "reset free regions coming");
6535 6555 }
6536 6556 }
6537 6557
6538 6558 void G1CollectedHeap::wait_while_free_regions_coming() {
6539 6559 // Most of the time we won't have to wait, so let's do a quick test
6540 6560 // first before we take the lock.
6541 6561 if (!free_regions_coming()) {
6542 6562 return;
6543 6563 }
6544 6564
6545 6565 if (G1ConcRegionFreeingVerbose) {
6546 6566 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6547 6567 "waiting for free regions");
6548 6568 }
6549 6569
6550 6570 {
6551 6571 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6552 6572 while (free_regions_coming()) {
6553 6573 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6554 6574 }
6555 6575 }
6556 6576
6557 6577 if (G1ConcRegionFreeingVerbose) {
6558 6578 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6559 6579 "done waiting for free regions");
6560 6580 }
6561 6581 }
6562 6582
6563 6583 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6564 6584 assert(heap_lock_held_for_gc(),
6565 6585 "the heap lock should already be held by or for this thread");
6566 6586 _young_list->push_region(hr);
6567 6587 }
6568 6588
6569 6589 class NoYoungRegionsClosure: public HeapRegionClosure {
6570 6590 private:
6571 6591 bool _success;
6572 6592 public:
6573 6593 NoYoungRegionsClosure() : _success(true) { }
6574 6594 bool doHeapRegion(HeapRegion* r) {
6575 6595 if (r->is_young()) {
6576 6596 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6577 6597 r->bottom(), r->end());
6578 6598 _success = false;
6579 6599 }
6580 6600 return false;
6581 6601 }
6582 6602 bool success() { return _success; }
6583 6603 };
6584 6604
6585 6605 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6586 6606 bool ret = _young_list->check_list_empty(check_sample);
6587 6607
6588 6608 if (check_heap) {
6589 6609 NoYoungRegionsClosure closure;
6590 6610 heap_region_iterate(&closure);
6591 6611 ret = ret && closure.success();
6592 6612 }
6593 6613
6594 6614 return ret;
6595 6615 }
6596 6616
6597 6617 class TearDownRegionSetsClosure : public HeapRegionClosure {
6598 6618 private:
6599 6619 HeapRegionSet *_old_set;
6600 6620
6601 6621 public:
6602 6622 TearDownRegionSetsClosure(HeapRegionSet* old_set) : _old_set(old_set) { }
6603 6623
6604 6624 bool doHeapRegion(HeapRegion* r) {
6605 6625 if (r->is_old()) {
6606 6626 _old_set->remove(r);
6607 6627 } else {
6608 6628 // We ignore free regions, we'll empty the free list afterwards.
6609 6629 // We ignore young regions, we'll empty the young list afterwards.
6610 6630 // We ignore humongous regions, we're not tearing down the
6611 6631 // humongous regions set.
6612 6632 assert(r->is_free() || r->is_young() || r->isHumongous(),
6613 6633 "it cannot be another type");
6614 6634 }
6615 6635 return false;
6616 6636 }
6617 6637
6618 6638 ~TearDownRegionSetsClosure() {
6619 6639 assert(_old_set->is_empty(), "post-condition");
6620 6640 }
6621 6641 };
6622 6642
6623 6643 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6624 6644 assert_at_safepoint(true /* should_be_vm_thread */);
6625 6645
6626 6646 if (!free_list_only) {
6627 6647 TearDownRegionSetsClosure cl(&_old_set);
6628 6648 heap_region_iterate(&cl);
6629 6649
6630 6650 // Note that emptying the _young_list is postponed and instead done as
6631 6651 // the first step when rebuilding the regions sets again. The reason for
6632 6652 // this is that during a full GC string deduplication needs to know if
6633 6653 // a collected region was young or old when the full GC was initiated.
6634 6654 }
6635 6655 _hrm.remove_all_free_regions();
6636 6656 }
6637 6657
6638 6658 class RebuildRegionSetsClosure : public HeapRegionClosure {
6639 6659 private:
6640 6660 bool _free_list_only;
6641 6661 HeapRegionSet* _old_set;
6642 6662 HeapRegionManager* _hrm;
6643 6663 size_t _total_used;
6644 6664
6645 6665 public:
6646 6666 RebuildRegionSetsClosure(bool free_list_only,
6647 6667 HeapRegionSet* old_set, HeapRegionManager* hrm) :
6648 6668 _free_list_only(free_list_only),
6649 6669 _old_set(old_set), _hrm(hrm), _total_used(0) {
6650 6670 assert(_hrm->num_free_regions() == 0, "pre-condition");
6651 6671 if (!free_list_only) {
6652 6672 assert(_old_set->is_empty(), "pre-condition");
6653 6673 }
6654 6674 }
6655 6675
6656 6676 bool doHeapRegion(HeapRegion* r) {
6657 6677 if (r->continuesHumongous()) {
6658 6678 return false;
6659 6679 }
6660 6680
6661 6681 if (r->is_empty()) {
6662 6682 // Add free regions to the free list
6663 6683 r->set_free();
6664 6684 r->set_allocation_context(AllocationContext::system());
6665 6685 _hrm->insert_into_free_list(r);
6666 6686 } else if (!_free_list_only) {
6667 6687 assert(!r->is_young(), "we should not come across young regions");
6668 6688
6669 6689 if (r->isHumongous()) {
6670 6690 // We ignore humongous regions, we left the humongous set unchanged
6671 6691 } else {
6672 6692 // Objects that were compacted would have ended up on regions
6673 6693 // that were previously old or free.
6674 6694 assert(r->is_free() || r->is_old(), "invariant");
6675 6695 // We now consider them old, so register as such.
6676 6696 r->set_old();
6677 6697 _old_set->add(r);
6678 6698 }
6679 6699 _total_used += r->used();
6680 6700 }
6681 6701
6682 6702 return false;
6683 6703 }
6684 6704
6685 6705 size_t total_used() {
6686 6706 return _total_used;
6687 6707 }
6688 6708 };
6689 6709
6690 6710 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6691 6711 assert_at_safepoint(true /* should_be_vm_thread */);
6692 6712
6693 6713 if (!free_list_only) {
6694 6714 _young_list->empty_list();
6695 6715 }
6696 6716
6697 6717 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_hrm);
6698 6718 heap_region_iterate(&cl);
6699 6719
6700 6720 if (!free_list_only) {
6701 6721 _allocator->set_used(cl.total_used());
6702 6722 }
6703 6723 assert(_allocator->used_unlocked() == recalculate_used(),
6704 6724 err_msg("inconsistent _allocator->used_unlocked(), "
6705 6725 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6706 6726 _allocator->used_unlocked(), recalculate_used()));
6707 6727 }
6708 6728
6709 6729 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6710 6730 _refine_cte_cl->set_concurrent(concurrent);
6711 6731 }
6712 6732
6713 6733 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6714 6734 HeapRegion* hr = heap_region_containing(p);
6715 6735 return hr->is_in(p);
6716 6736 }
6717 6737
6718 6738 // Methods for the mutator alloc region
6719 6739
6720 6740 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6721 6741 bool force) {
6722 6742 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6723 6743 assert(!force || g1_policy()->can_expand_young_list(),
6724 6744 "if force is true we should be able to expand the young list");
6725 6745 bool young_list_full = g1_policy()->is_young_list_full();
6726 6746 if (force || !young_list_full) {
6727 6747 HeapRegion* new_alloc_region = new_region(word_size,
6728 6748 false /* is_old */,
6729 6749 false /* do_expand */);
6730 6750 if (new_alloc_region != NULL) {
6731 6751 set_region_short_lived_locked(new_alloc_region);
6732 6752 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6733 6753 check_bitmaps("Mutator Region Allocation", new_alloc_region);
6734 6754 return new_alloc_region;
6735 6755 }
6736 6756 }
6737 6757 return NULL;
6738 6758 }
6739 6759
6740 6760 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6741 6761 size_t allocated_bytes) {
6742 6762 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6743 6763 assert(alloc_region->is_eden(), "all mutator alloc regions should be eden");
6744 6764
6745 6765 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6746 6766 _allocator->increase_used(allocated_bytes);
6747 6767 _hr_printer.retire(alloc_region);
6748 6768 // We update the eden sizes here, when the region is retired,
6749 6769 // instead of when it's allocated, since this is the point that its
6750 6770 // used space has been recored in _summary_bytes_used.
6751 6771 g1mm()->update_eden_size();
6752 6772 }
6753 6773
6754 6774 void G1CollectedHeap::set_par_threads() {
6755 6775 // Don't change the number of workers. Use the value previously set
6756 6776 // in the workgroup.
6757 6777 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6758 6778 uint n_workers = workers()->active_workers();
6759 6779 assert(UseDynamicNumberOfGCThreads ||
6760 6780 n_workers == workers()->total_workers(),
6761 6781 "Otherwise should be using the total number of workers");
6762 6782 if (n_workers == 0) {
6763 6783 assert(false, "Should have been set in prior evacuation pause.");
6764 6784 n_workers = ParallelGCThreads;
6765 6785 workers()->set_active_workers(n_workers);
6766 6786 }
6767 6787 set_par_threads(n_workers);
6768 6788 }
6769 6789
6770 6790 // Methods for the GC alloc regions
6771 6791
6772 6792 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6773 6793 uint count,
6774 6794 GCAllocPurpose ap) {
6775 6795 assert(FreeList_lock->owned_by_self(), "pre-condition");
6776 6796
6777 6797 if (count < g1_policy()->max_regions(ap)) {
6778 6798 bool survivor = (ap == GCAllocForSurvived);
6779 6799 HeapRegion* new_alloc_region = new_region(word_size,
6780 6800 !survivor,
6781 6801 true /* do_expand */);
6782 6802 if (new_alloc_region != NULL) {
6783 6803 // We really only need to do this for old regions given that we
6784 6804 // should never scan survivors. But it doesn't hurt to do it
6785 6805 // for survivors too.
6786 6806 new_alloc_region->record_top_and_timestamp();
6787 6807 if (survivor) {
6788 6808 new_alloc_region->set_survivor();
6789 6809 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6790 6810 check_bitmaps("Survivor Region Allocation", new_alloc_region);
6791 6811 } else {
6792 6812 new_alloc_region->set_old();
6793 6813 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6794 6814 check_bitmaps("Old Region Allocation", new_alloc_region);
6795 6815 }
6796 6816 bool during_im = g1_policy()->during_initial_mark_pause();
6797 6817 new_alloc_region->note_start_of_copying(during_im);
6798 6818 return new_alloc_region;
6799 6819 } else {
6800 6820 g1_policy()->note_alloc_region_limit_reached(ap);
6801 6821 }
6802 6822 }
6803 6823 return NULL;
6804 6824 }
6805 6825
6806 6826 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6807 6827 size_t allocated_bytes,
6808 6828 GCAllocPurpose ap) {
6809 6829 bool during_im = g1_policy()->during_initial_mark_pause();
6810 6830 alloc_region->note_end_of_copying(during_im);
6811 6831 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6812 6832 if (ap == GCAllocForSurvived) {
6813 6833 young_list()->add_survivor_region(alloc_region);
6814 6834 } else {
6815 6835 _old_set.add(alloc_region);
6816 6836 }
6817 6837 _hr_printer.retire(alloc_region);
6818 6838 }
6819 6839
6820 6840 // Heap region set verification
6821 6841
6822 6842 class VerifyRegionListsClosure : public HeapRegionClosure {
6823 6843 private:
6824 6844 HeapRegionSet* _old_set;
6825 6845 HeapRegionSet* _humongous_set;
6826 6846 HeapRegionManager* _hrm;
6827 6847
6828 6848 public:
6829 6849 HeapRegionSetCount _old_count;
6830 6850 HeapRegionSetCount _humongous_count;
6831 6851 HeapRegionSetCount _free_count;
6832 6852
6833 6853 VerifyRegionListsClosure(HeapRegionSet* old_set,
6834 6854 HeapRegionSet* humongous_set,
6835 6855 HeapRegionManager* hrm) :
6836 6856 _old_set(old_set), _humongous_set(humongous_set), _hrm(hrm),
6837 6857 _old_count(), _humongous_count(), _free_count(){ }
6838 6858
6839 6859 bool doHeapRegion(HeapRegion* hr) {
6840 6860 if (hr->continuesHumongous()) {
6841 6861 return false;
6842 6862 }
6843 6863
6844 6864 if (hr->is_young()) {
6845 6865 // TODO
6846 6866 } else if (hr->startsHumongous()) {
6847 6867 assert(hr->containing_set() == _humongous_set, err_msg("Heap region %u is starts humongous but not in humongous set.", hr->hrm_index()));
6848 6868 _humongous_count.increment(1u, hr->capacity());
6849 6869 } else if (hr->is_empty()) {
6850 6870 assert(_hrm->is_free(hr), err_msg("Heap region %u is empty but not on the free list.", hr->hrm_index()));
6851 6871 _free_count.increment(1u, hr->capacity());
6852 6872 } else if (hr->is_old()) {
6853 6873 assert(hr->containing_set() == _old_set, err_msg("Heap region %u is old but not in the old set.", hr->hrm_index()));
6854 6874 _old_count.increment(1u, hr->capacity());
6855 6875 } else {
6856 6876 ShouldNotReachHere();
6857 6877 }
6858 6878 return false;
6859 6879 }
6860 6880
6861 6881 void verify_counts(HeapRegionSet* old_set, HeapRegionSet* humongous_set, HeapRegionManager* free_list) {
6862 6882 guarantee(old_set->length() == _old_count.length(), err_msg("Old set count mismatch. Expected %u, actual %u.", old_set->length(), _old_count.length()));
6863 6883 guarantee(old_set->total_capacity_bytes() == _old_count.capacity(), err_msg("Old set capacity mismatch. Expected " SIZE_FORMAT ", actual " SIZE_FORMAT,
6864 6884 old_set->total_capacity_bytes(), _old_count.capacity()));
6865 6885
6866 6886 guarantee(humongous_set->length() == _humongous_count.length(), err_msg("Hum set count mismatch. Expected %u, actual %u.", humongous_set->length(), _humongous_count.length()));
6867 6887 guarantee(humongous_set->total_capacity_bytes() == _humongous_count.capacity(), err_msg("Hum set capacity mismatch. Expected " SIZE_FORMAT ", actual " SIZE_FORMAT,
6868 6888 humongous_set->total_capacity_bytes(), _humongous_count.capacity()));
6869 6889
6870 6890 guarantee(free_list->num_free_regions() == _free_count.length(), err_msg("Free list count mismatch. Expected %u, actual %u.", free_list->num_free_regions(), _free_count.length()));
6871 6891 guarantee(free_list->total_capacity_bytes() == _free_count.capacity(), err_msg("Free list capacity mismatch. Expected " SIZE_FORMAT ", actual " SIZE_FORMAT,
6872 6892 free_list->total_capacity_bytes(), _free_count.capacity()));
6873 6893 }
6874 6894 };
6875 6895
6876 6896 void G1CollectedHeap::verify_region_sets() {
6877 6897 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6878 6898
6879 6899 // First, check the explicit lists.
6880 6900 _hrm.verify();
6881 6901 {
6882 6902 // Given that a concurrent operation might be adding regions to
6883 6903 // the secondary free list we have to take the lock before
6884 6904 // verifying it.
6885 6905 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6886 6906 _secondary_free_list.verify_list();
6887 6907 }
6888 6908
6889 6909 // If a concurrent region freeing operation is in progress it will
6890 6910 // be difficult to correctly attributed any free regions we come
6891 6911 // across to the correct free list given that they might belong to
6892 6912 // one of several (free_list, secondary_free_list, any local lists,
6893 6913 // etc.). So, if that's the case we will skip the rest of the
6894 6914 // verification operation. Alternatively, waiting for the concurrent
6895 6915 // operation to complete will have a non-trivial effect on the GC's
6896 6916 // operation (no concurrent operation will last longer than the
6897 6917 // interval between two calls to verification) and it might hide
6898 6918 // any issues that we would like to catch during testing.
6899 6919 if (free_regions_coming()) {
6900 6920 return;
6901 6921 }
6902 6922
6903 6923 // Make sure we append the secondary_free_list on the free_list so
6904 6924 // that all free regions we will come across can be safely
6905 6925 // attributed to the free_list.
6906 6926 append_secondary_free_list_if_not_empty_with_lock();
6907 6927
6908 6928 // Finally, make sure that the region accounting in the lists is
6909 6929 // consistent with what we see in the heap.
6910 6930
6911 6931 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_hrm);
6912 6932 heap_region_iterate(&cl);
6913 6933 cl.verify_counts(&_old_set, &_humongous_set, &_hrm);
6914 6934 }
6915 6935
6916 6936 // Optimized nmethod scanning
6917 6937
6918 6938 class RegisterNMethodOopClosure: public OopClosure {
6919 6939 G1CollectedHeap* _g1h;
6920 6940 nmethod* _nm;
6921 6941
6922 6942 template <class T> void do_oop_work(T* p) {
6923 6943 T heap_oop = oopDesc::load_heap_oop(p);
6924 6944 if (!oopDesc::is_null(heap_oop)) {
6925 6945 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
6926 6946 HeapRegion* hr = _g1h->heap_region_containing(obj);
6927 6947 assert(!hr->continuesHumongous(),
6928 6948 err_msg("trying to add code root "PTR_FORMAT" in continuation of humongous region "HR_FORMAT
6929 6949 " starting at "HR_FORMAT,
6930 6950 _nm, HR_FORMAT_PARAMS(hr), HR_FORMAT_PARAMS(hr->humongous_start_region())));
6931 6951
6932 6952 // HeapRegion::add_strong_code_root_locked() avoids adding duplicate entries.
6933 6953 hr->add_strong_code_root_locked(_nm);
6934 6954 }
6935 6955 }
6936 6956
6937 6957 public:
6938 6958 RegisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) :
6939 6959 _g1h(g1h), _nm(nm) {}
6940 6960
6941 6961 void do_oop(oop* p) { do_oop_work(p); }
6942 6962 void do_oop(narrowOop* p) { do_oop_work(p); }
6943 6963 };
6944 6964
6945 6965 class UnregisterNMethodOopClosure: public OopClosure {
6946 6966 G1CollectedHeap* _g1h;
6947 6967 nmethod* _nm;
6948 6968
6949 6969 template <class T> void do_oop_work(T* p) {
6950 6970 T heap_oop = oopDesc::load_heap_oop(p);
6951 6971 if (!oopDesc::is_null(heap_oop)) {
6952 6972 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
6953 6973 HeapRegion* hr = _g1h->heap_region_containing(obj);
6954 6974 assert(!hr->continuesHumongous(),
6955 6975 err_msg("trying to remove code root "PTR_FORMAT" in continuation of humongous region "HR_FORMAT
6956 6976 " starting at "HR_FORMAT,
6957 6977 _nm, HR_FORMAT_PARAMS(hr), HR_FORMAT_PARAMS(hr->humongous_start_region())));
6958 6978
6959 6979 hr->remove_strong_code_root(_nm);
6960 6980 }
6961 6981 }
6962 6982
6963 6983 public:
6964 6984 UnregisterNMethodOopClosure(G1CollectedHeap* g1h, nmethod* nm) :
6965 6985 _g1h(g1h), _nm(nm) {}
6966 6986
6967 6987 void do_oop(oop* p) { do_oop_work(p); }
6968 6988 void do_oop(narrowOop* p) { do_oop_work(p); }
6969 6989 };
6970 6990
6971 6991 void G1CollectedHeap::register_nmethod(nmethod* nm) {
6972 6992 CollectedHeap::register_nmethod(nm);
6973 6993
6974 6994 guarantee(nm != NULL, "sanity");
6975 6995 RegisterNMethodOopClosure reg_cl(this, nm);
6976 6996 nm->oops_do(®_cl);
6977 6997 }
6978 6998
6979 6999 void G1CollectedHeap::unregister_nmethod(nmethod* nm) {
6980 7000 CollectedHeap::unregister_nmethod(nm);
6981 7001
6982 7002 guarantee(nm != NULL, "sanity");
6983 7003 UnregisterNMethodOopClosure reg_cl(this, nm);
6984 7004 nm->oops_do(®_cl, true);
6985 7005 }
6986 7006
6987 7007 void G1CollectedHeap::purge_code_root_memory() {
6988 7008 double purge_start = os::elapsedTime();
6989 7009 G1CodeRootSet::purge();
6990 7010 double purge_time_ms = (os::elapsedTime() - purge_start) * 1000.0;
6991 7011 g1_policy()->phase_times()->record_strong_code_root_purge_time(purge_time_ms);
6992 7012 }
6993 7013
6994 7014 class RebuildStrongCodeRootClosure: public CodeBlobClosure {
6995 7015 G1CollectedHeap* _g1h;
6996 7016
6997 7017 public:
6998 7018 RebuildStrongCodeRootClosure(G1CollectedHeap* g1h) :
6999 7019 _g1h(g1h) {}
7000 7020
7001 7021 void do_code_blob(CodeBlob* cb) {
7002 7022 nmethod* nm = (cb != NULL) ? cb->as_nmethod_or_null() : NULL;
7003 7023 if (nm == NULL) {
7004 7024 return;
7005 7025 }
7006 7026
7007 7027 if (ScavengeRootsInCode) {
7008 7028 _g1h->register_nmethod(nm);
7009 7029 }
7010 7030 }
7011 7031 };
7012 7032
7013 7033 void G1CollectedHeap::rebuild_strong_code_roots() {
7014 7034 RebuildStrongCodeRootClosure blob_cl(this);
7015 7035 CodeCache::blobs_do(&blob_cl);
7016 7036 }
↓ open down ↓ |
1765 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX