Print this page
rev 6912 : 8065227: Report allocation context stats at end of cleanup
Summary: Moved allocation context update from remark to the cleanup phase.
Reviewed-by: mgerdin, jmasa
Split |
Split |
Close |
Expand all |
Collapse all |
--- old/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.cpp
+++ new/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.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
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/metadataOnStackMark.hpp"
27 27 #include "classfile/symbolTable.hpp"
28 28 #include "code/codeCache.hpp"
29 29 #include "gc_implementation/g1/concurrentMark.inline.hpp"
30 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 31 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
32 32 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
33 33 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
34 34 #include "gc_implementation/g1/g1Log.hpp"
35 35 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
36 36 #include "gc_implementation/g1/g1RemSet.hpp"
37 37 #include "gc_implementation/g1/heapRegion.inline.hpp"
38 38 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
39 39 #include "gc_implementation/g1/heapRegionRemSet.hpp"
40 40 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
41 41 #include "gc_implementation/shared/vmGCOperations.hpp"
42 42 #include "gc_implementation/shared/gcTimer.hpp"
43 43 #include "gc_implementation/shared/gcTrace.hpp"
44 44 #include "gc_implementation/shared/gcTraceTime.hpp"
45 45 #include "memory/allocation.hpp"
46 46 #include "memory/genOopClosures.inline.hpp"
47 47 #include "memory/referencePolicy.hpp"
48 48 #include "memory/resourceArea.hpp"
49 49 #include "oops/oop.inline.hpp"
50 50 #include "runtime/handles.inline.hpp"
51 51 #include "runtime/java.hpp"
52 52 #include "runtime/prefetch.inline.hpp"
53 53 #include "services/memTracker.hpp"
54 54
55 55 // Concurrent marking bit map wrapper
56 56
57 57 CMBitMapRO::CMBitMapRO(int shifter) :
58 58 _bm(),
59 59 _shifter(shifter) {
60 60 _bmStartWord = 0;
61 61 _bmWordSize = 0;
62 62 }
63 63
64 64 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
65 65 const HeapWord* limit) const {
66 66 // First we must round addr *up* to a possible object boundary.
67 67 addr = (HeapWord*)align_size_up((intptr_t)addr,
68 68 HeapWordSize << _shifter);
69 69 size_t addrOffset = heapWordToOffset(addr);
70 70 if (limit == NULL) {
71 71 limit = _bmStartWord + _bmWordSize;
72 72 }
73 73 size_t limitOffset = heapWordToOffset(limit);
74 74 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
75 75 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
76 76 assert(nextAddr >= addr, "get_next_one postcondition");
77 77 assert(nextAddr == limit || isMarked(nextAddr),
78 78 "get_next_one postcondition");
79 79 return nextAddr;
80 80 }
81 81
82 82 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
83 83 const HeapWord* limit) const {
84 84 size_t addrOffset = heapWordToOffset(addr);
85 85 if (limit == NULL) {
86 86 limit = _bmStartWord + _bmWordSize;
87 87 }
88 88 size_t limitOffset = heapWordToOffset(limit);
89 89 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
90 90 HeapWord* nextAddr = offsetToHeapWord(nextOffset);
91 91 assert(nextAddr >= addr, "get_next_one postcondition");
92 92 assert(nextAddr == limit || !isMarked(nextAddr),
93 93 "get_next_one postcondition");
94 94 return nextAddr;
95 95 }
96 96
97 97 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
98 98 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
99 99 return (int) (diff >> _shifter);
100 100 }
101 101
102 102 #ifndef PRODUCT
103 103 bool CMBitMapRO::covers(MemRegion heap_rs) const {
104 104 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
105 105 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
106 106 "size inconsistency");
107 107 return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
108 108 _bmWordSize == heap_rs.word_size();
109 109 }
110 110 #endif
111 111
112 112 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
113 113 _bm.print_on_error(st, prefix);
114 114 }
115 115
116 116 size_t CMBitMap::compute_size(size_t heap_size) {
117 117 return heap_size / mark_distance();
118 118 }
119 119
120 120 size_t CMBitMap::mark_distance() {
121 121 return MinObjAlignmentInBytes * BitsPerByte;
122 122 }
123 123
124 124 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
125 125 _bmStartWord = heap.start();
126 126 _bmWordSize = heap.word_size();
127 127
128 128 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
129 129 _bm.set_size(_bmWordSize >> _shifter);
130 130
131 131 storage->set_mapping_changed_listener(&_listener);
132 132 }
133 133
134 134 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
135 135 if (zero_filled) {
136 136 return;
137 137 }
138 138 // We need to clear the bitmap on commit, removing any existing information.
139 139 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
140 140 _bm->clearRange(mr);
141 141 }
142 142
143 143 // Closure used for clearing the given mark bitmap.
144 144 class ClearBitmapHRClosure : public HeapRegionClosure {
145 145 private:
146 146 ConcurrentMark* _cm;
147 147 CMBitMap* _bitmap;
148 148 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration.
149 149 public:
150 150 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
151 151 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
152 152 }
153 153
154 154 virtual bool doHeapRegion(HeapRegion* r) {
155 155 size_t const chunk_size_in_words = M / HeapWordSize;
156 156
157 157 HeapWord* cur = r->bottom();
158 158 HeapWord* const end = r->end();
159 159
160 160 while (cur < end) {
161 161 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
162 162 _bitmap->clearRange(mr);
163 163
164 164 cur += chunk_size_in_words;
165 165
166 166 // Abort iteration if after yielding the marking has been aborted.
167 167 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
168 168 return true;
169 169 }
170 170 // Repeat the asserts from before the start of the closure. We will do them
171 171 // as asserts here to minimize their overhead on the product. However, we
172 172 // will have them as guarantees at the beginning / end of the bitmap
173 173 // clearing to get some checking in the product.
174 174 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
175 175 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
176 176 }
177 177
178 178 return false;
179 179 }
180 180 };
181 181
182 182 void CMBitMap::clearAll() {
183 183 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
184 184 G1CollectedHeap::heap()->heap_region_iterate(&cl);
185 185 guarantee(cl.complete(), "Must have completed iteration.");
186 186 return;
187 187 }
188 188
189 189 void CMBitMap::markRange(MemRegion mr) {
190 190 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
191 191 assert(!mr.is_empty(), "unexpected empty region");
192 192 assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
193 193 ((HeapWord *) mr.end())),
194 194 "markRange memory region end is not card aligned");
195 195 // convert address range into offset range
196 196 _bm.at_put_range(heapWordToOffset(mr.start()),
197 197 heapWordToOffset(mr.end()), true);
198 198 }
199 199
200 200 void CMBitMap::clearRange(MemRegion mr) {
201 201 mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
202 202 assert(!mr.is_empty(), "unexpected empty region");
203 203 // convert address range into offset range
204 204 _bm.at_put_range(heapWordToOffset(mr.start()),
205 205 heapWordToOffset(mr.end()), false);
206 206 }
207 207
208 208 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
209 209 HeapWord* end_addr) {
210 210 HeapWord* start = getNextMarkedWordAddress(addr);
211 211 start = MIN2(start, end_addr);
212 212 HeapWord* end = getNextUnmarkedWordAddress(start);
213 213 end = MIN2(end, end_addr);
214 214 assert(start <= end, "Consistency check");
215 215 MemRegion mr(start, end);
216 216 if (!mr.is_empty()) {
217 217 clearRange(mr);
218 218 }
219 219 return mr;
220 220 }
221 221
222 222 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
223 223 _base(NULL), _cm(cm)
224 224 #ifdef ASSERT
225 225 , _drain_in_progress(false)
226 226 , _drain_in_progress_yields(false)
227 227 #endif
228 228 {}
229 229
230 230 bool CMMarkStack::allocate(size_t capacity) {
231 231 // allocate a stack of the requisite depth
232 232 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
233 233 if (!rs.is_reserved()) {
234 234 warning("ConcurrentMark MarkStack allocation failure");
235 235 return false;
236 236 }
237 237 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
238 238 if (!_virtual_space.initialize(rs, rs.size())) {
239 239 warning("ConcurrentMark MarkStack backing store failure");
240 240 // Release the virtual memory reserved for the marking stack
241 241 rs.release();
242 242 return false;
243 243 }
244 244 assert(_virtual_space.committed_size() == rs.size(),
245 245 "Didn't reserve backing store for all of ConcurrentMark stack?");
246 246 _base = (oop*) _virtual_space.low();
247 247 setEmpty();
248 248 _capacity = (jint) capacity;
249 249 _saved_index = -1;
250 250 _should_expand = false;
251 251 NOT_PRODUCT(_max_depth = 0);
252 252 return true;
253 253 }
254 254
255 255 void CMMarkStack::expand() {
256 256 // Called, during remark, if we've overflown the marking stack during marking.
257 257 assert(isEmpty(), "stack should been emptied while handling overflow");
258 258 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
259 259 // Clear expansion flag
260 260 _should_expand = false;
261 261 if (_capacity == (jint) MarkStackSizeMax) {
262 262 if (PrintGCDetails && Verbose) {
263 263 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
264 264 }
265 265 return;
266 266 }
267 267 // Double capacity if possible
268 268 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
269 269 // Do not give up existing stack until we have managed to
270 270 // get the double capacity that we desired.
271 271 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
272 272 sizeof(oop)));
273 273 if (rs.is_reserved()) {
274 274 // Release the backing store associated with old stack
275 275 _virtual_space.release();
276 276 // Reinitialize virtual space for new stack
277 277 if (!_virtual_space.initialize(rs, rs.size())) {
278 278 fatal("Not enough swap for expanded marking stack capacity");
279 279 }
280 280 _base = (oop*)(_virtual_space.low());
281 281 _index = 0;
282 282 _capacity = new_capacity;
283 283 } else {
284 284 if (PrintGCDetails && Verbose) {
285 285 // Failed to double capacity, continue;
286 286 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
287 287 SIZE_FORMAT"K to " SIZE_FORMAT"K",
288 288 _capacity / K, new_capacity / K);
289 289 }
290 290 }
291 291 }
292 292
293 293 void CMMarkStack::set_should_expand() {
294 294 // If we're resetting the marking state because of an
295 295 // marking stack overflow, record that we should, if
296 296 // possible, expand the stack.
297 297 _should_expand = _cm->has_overflown();
298 298 }
299 299
300 300 CMMarkStack::~CMMarkStack() {
301 301 if (_base != NULL) {
302 302 _base = NULL;
303 303 _virtual_space.release();
304 304 }
305 305 }
306 306
307 307 void CMMarkStack::par_push(oop ptr) {
308 308 while (true) {
309 309 if (isFull()) {
310 310 _overflow = true;
311 311 return;
312 312 }
313 313 // Otherwise...
314 314 jint index = _index;
315 315 jint next_index = index+1;
316 316 jint res = Atomic::cmpxchg(next_index, &_index, index);
317 317 if (res == index) {
318 318 _base[index] = ptr;
319 319 // Note that we don't maintain this atomically. We could, but it
320 320 // doesn't seem necessary.
321 321 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
322 322 return;
323 323 }
324 324 // Otherwise, we need to try again.
325 325 }
326 326 }
327 327
328 328 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
329 329 while (true) {
330 330 if (isFull()) {
331 331 _overflow = true;
332 332 return;
333 333 }
334 334 // Otherwise...
335 335 jint index = _index;
336 336 jint next_index = index + n;
337 337 if (next_index > _capacity) {
338 338 _overflow = true;
339 339 return;
340 340 }
341 341 jint res = Atomic::cmpxchg(next_index, &_index, index);
342 342 if (res == index) {
343 343 for (int i = 0; i < n; i++) {
344 344 int ind = index + i;
345 345 assert(ind < _capacity, "By overflow test above.");
346 346 _base[ind] = ptr_arr[i];
347 347 }
348 348 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
349 349 return;
350 350 }
351 351 // Otherwise, we need to try again.
352 352 }
353 353 }
354 354
355 355 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
356 356 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
357 357 jint start = _index;
358 358 jint next_index = start + n;
359 359 if (next_index > _capacity) {
360 360 _overflow = true;
361 361 return;
362 362 }
363 363 // Otherwise.
364 364 _index = next_index;
365 365 for (int i = 0; i < n; i++) {
366 366 int ind = start + i;
367 367 assert(ind < _capacity, "By overflow test above.");
368 368 _base[ind] = ptr_arr[i];
369 369 }
370 370 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
371 371 }
372 372
373 373 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
374 374 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
375 375 jint index = _index;
376 376 if (index == 0) {
377 377 *n = 0;
378 378 return false;
379 379 } else {
380 380 int k = MIN2(max, index);
381 381 jint new_ind = index - k;
382 382 for (int j = 0; j < k; j++) {
383 383 ptr_arr[j] = _base[new_ind + j];
384 384 }
385 385 _index = new_ind;
386 386 *n = k;
387 387 return true;
388 388 }
389 389 }
390 390
391 391 template<class OopClosureClass>
392 392 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
393 393 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
394 394 || SafepointSynchronize::is_at_safepoint(),
395 395 "Drain recursion must be yield-safe.");
396 396 bool res = true;
397 397 debug_only(_drain_in_progress = true);
398 398 debug_only(_drain_in_progress_yields = yield_after);
399 399 while (!isEmpty()) {
400 400 oop newOop = pop();
401 401 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
402 402 assert(newOop->is_oop(), "Expected an oop");
403 403 assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
404 404 "only grey objects on this stack");
405 405 newOop->oop_iterate(cl);
406 406 if (yield_after && _cm->do_yield_check()) {
407 407 res = false;
408 408 break;
409 409 }
410 410 }
411 411 debug_only(_drain_in_progress = false);
412 412 return res;
413 413 }
414 414
415 415 void CMMarkStack::note_start_of_gc() {
416 416 assert(_saved_index == -1,
417 417 "note_start_of_gc()/end_of_gc() bracketed incorrectly");
418 418 _saved_index = _index;
419 419 }
420 420
421 421 void CMMarkStack::note_end_of_gc() {
422 422 // This is intentionally a guarantee, instead of an assert. If we
423 423 // accidentally add something to the mark stack during GC, it
424 424 // will be a correctness issue so it's better if we crash. we'll
425 425 // only check this once per GC anyway, so it won't be a performance
426 426 // issue in any way.
427 427 guarantee(_saved_index == _index,
428 428 err_msg("saved index: %d index: %d", _saved_index, _index));
429 429 _saved_index = -1;
430 430 }
431 431
432 432 void CMMarkStack::oops_do(OopClosure* f) {
433 433 assert(_saved_index == _index,
434 434 err_msg("saved index: %d index: %d", _saved_index, _index));
435 435 for (int i = 0; i < _index; i += 1) {
436 436 f->do_oop(&_base[i]);
437 437 }
438 438 }
439 439
440 440 CMRootRegions::CMRootRegions() :
441 441 _young_list(NULL), _cm(NULL), _scan_in_progress(false),
442 442 _should_abort(false), _next_survivor(NULL) { }
443 443
444 444 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
445 445 _young_list = g1h->young_list();
446 446 _cm = cm;
447 447 }
448 448
449 449 void CMRootRegions::prepare_for_scan() {
450 450 assert(!scan_in_progress(), "pre-condition");
451 451
452 452 // Currently, only survivors can be root regions.
453 453 assert(_next_survivor == NULL, "pre-condition");
454 454 _next_survivor = _young_list->first_survivor_region();
455 455 _scan_in_progress = (_next_survivor != NULL);
456 456 _should_abort = false;
457 457 }
458 458
459 459 HeapRegion* CMRootRegions::claim_next() {
460 460 if (_should_abort) {
461 461 // If someone has set the should_abort flag, we return NULL to
462 462 // force the caller to bail out of their loop.
463 463 return NULL;
464 464 }
465 465
466 466 // Currently, only survivors can be root regions.
467 467 HeapRegion* res = _next_survivor;
468 468 if (res != NULL) {
469 469 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
470 470 // Read it again in case it changed while we were waiting for the lock.
471 471 res = _next_survivor;
472 472 if (res != NULL) {
473 473 if (res == _young_list->last_survivor_region()) {
474 474 // We just claimed the last survivor so store NULL to indicate
475 475 // that we're done.
476 476 _next_survivor = NULL;
477 477 } else {
478 478 _next_survivor = res->get_next_young_region();
479 479 }
480 480 } else {
481 481 // Someone else claimed the last survivor while we were trying
482 482 // to take the lock so nothing else to do.
483 483 }
484 484 }
485 485 assert(res == NULL || res->is_survivor(), "post-condition");
486 486
487 487 return res;
488 488 }
489 489
490 490 void CMRootRegions::scan_finished() {
491 491 assert(scan_in_progress(), "pre-condition");
492 492
493 493 // Currently, only survivors can be root regions.
494 494 if (!_should_abort) {
495 495 assert(_next_survivor == NULL, "we should have claimed all survivors");
496 496 }
497 497 _next_survivor = NULL;
498 498
499 499 {
500 500 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
501 501 _scan_in_progress = false;
502 502 RootRegionScan_lock->notify_all();
503 503 }
504 504 }
505 505
506 506 bool CMRootRegions::wait_until_scan_finished() {
507 507 if (!scan_in_progress()) return false;
508 508
509 509 {
510 510 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
511 511 while (scan_in_progress()) {
512 512 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
513 513 }
514 514 }
515 515 return true;
516 516 }
517 517
518 518 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
519 519 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
520 520 #endif // _MSC_VER
521 521
522 522 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
523 523 return MAX2((n_par_threads + 2) / 4, 1U);
524 524 }
525 525
526 526 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
527 527 _g1h(g1h),
528 528 _markBitMap1(),
529 529 _markBitMap2(),
530 530 _parallel_marking_threads(0),
531 531 _max_parallel_marking_threads(0),
532 532 _sleep_factor(0.0),
533 533 _marking_task_overhead(1.0),
534 534 _cleanup_sleep_factor(0.0),
535 535 _cleanup_task_overhead(1.0),
536 536 _cleanup_list("Cleanup List"),
537 537 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
538 538 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
539 539 CardTableModRefBS::card_shift,
540 540 false /* in_resource_area*/),
541 541
542 542 _prevMarkBitMap(&_markBitMap1),
543 543 _nextMarkBitMap(&_markBitMap2),
544 544
545 545 _markStack(this),
546 546 // _finger set in set_non_marking_state
547 547
548 548 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
549 549 // _active_tasks set in set_non_marking_state
550 550 // _tasks set inside the constructor
551 551 _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
552 552 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
553 553
554 554 _has_overflown(false),
555 555 _concurrent(false),
556 556 _has_aborted(false),
557 557 _aborted_gc_id(GCId::undefined()),
558 558 _restart_for_overflow(false),
559 559 _concurrent_marking_in_progress(false),
560 560
561 561 // _verbose_level set below
562 562
563 563 _init_times(),
564 564 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
565 565 _cleanup_times(),
566 566 _total_counting_time(0.0),
567 567 _total_rs_scrub_time(0.0),
568 568
569 569 _parallel_workers(NULL),
570 570
571 571 _count_card_bitmaps(NULL),
572 572 _count_marked_bytes(NULL),
573 573 _completed_initialization(false) {
574 574 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
575 575 if (verbose_level < no_verbose) {
576 576 verbose_level = no_verbose;
577 577 }
578 578 if (verbose_level > high_verbose) {
579 579 verbose_level = high_verbose;
580 580 }
581 581 _verbose_level = verbose_level;
582 582
583 583 if (verbose_low()) {
584 584 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
585 585 "heap end = " INTPTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
586 586 }
587 587
588 588 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
589 589 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
590 590
591 591 // Create & start a ConcurrentMark thread.
592 592 _cmThread = new ConcurrentMarkThread(this);
593 593 assert(cmThread() != NULL, "CM Thread should have been created");
594 594 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
595 595 if (_cmThread->osthread() == NULL) {
596 596 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
597 597 }
598 598
599 599 assert(CGC_lock != NULL, "Where's the CGC_lock?");
600 600 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
601 601 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
602 602
603 603 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
604 604 satb_qs.set_buffer_size(G1SATBBufferSize);
605 605
606 606 _root_regions.init(_g1h, this);
607 607
608 608 if (ConcGCThreads > ParallelGCThreads) {
609 609 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
610 610 "than ParallelGCThreads (" UINTX_FORMAT ").",
611 611 ConcGCThreads, ParallelGCThreads);
612 612 return;
613 613 }
614 614 if (ParallelGCThreads == 0) {
615 615 // if we are not running with any parallel GC threads we will not
616 616 // spawn any marking threads either
617 617 _parallel_marking_threads = 0;
618 618 _max_parallel_marking_threads = 0;
619 619 _sleep_factor = 0.0;
620 620 _marking_task_overhead = 1.0;
621 621 } else {
622 622 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
623 623 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
624 624 // if both are set
625 625 _sleep_factor = 0.0;
626 626 _marking_task_overhead = 1.0;
627 627 } else if (G1MarkingOverheadPercent > 0) {
628 628 // We will calculate the number of parallel marking threads based
629 629 // on a target overhead with respect to the soft real-time goal
630 630 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
631 631 double overall_cm_overhead =
632 632 (double) MaxGCPauseMillis * marking_overhead /
633 633 (double) GCPauseIntervalMillis;
634 634 double cpu_ratio = 1.0 / (double) os::processor_count();
635 635 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
636 636 double marking_task_overhead =
637 637 overall_cm_overhead / marking_thread_num *
638 638 (double) os::processor_count();
639 639 double sleep_factor =
640 640 (1.0 - marking_task_overhead) / marking_task_overhead;
641 641
642 642 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
643 643 _sleep_factor = sleep_factor;
644 644 _marking_task_overhead = marking_task_overhead;
645 645 } else {
646 646 // Calculate the number of parallel marking threads by scaling
647 647 // the number of parallel GC threads.
648 648 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
649 649 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
650 650 _sleep_factor = 0.0;
651 651 _marking_task_overhead = 1.0;
652 652 }
653 653
654 654 assert(ConcGCThreads > 0, "Should have been set");
655 655 _parallel_marking_threads = (uint) ConcGCThreads;
656 656 _max_parallel_marking_threads = _parallel_marking_threads;
657 657
658 658 if (parallel_marking_threads() > 1) {
659 659 _cleanup_task_overhead = 1.0;
660 660 } else {
661 661 _cleanup_task_overhead = marking_task_overhead();
662 662 }
663 663 _cleanup_sleep_factor =
664 664 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
665 665
666 666 #if 0
667 667 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads());
668 668 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
669 669 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor());
670 670 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
671 671 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor());
672 672 #endif
673 673
674 674 guarantee(parallel_marking_threads() > 0, "peace of mind");
675 675 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
676 676 _max_parallel_marking_threads, false, true);
677 677 if (_parallel_workers == NULL) {
678 678 vm_exit_during_initialization("Failed necessary allocation.");
679 679 } else {
680 680 _parallel_workers->initialize_workers();
681 681 }
682 682 }
683 683
684 684 if (FLAG_IS_DEFAULT(MarkStackSize)) {
685 685 uintx mark_stack_size =
686 686 MIN2(MarkStackSizeMax,
687 687 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
688 688 // Verify that the calculated value for MarkStackSize is in range.
689 689 // It would be nice to use the private utility routine from Arguments.
690 690 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
691 691 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
692 692 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
693 693 mark_stack_size, (uintx) 1, MarkStackSizeMax);
694 694 return;
695 695 }
696 696 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
697 697 } else {
698 698 // Verify MarkStackSize is in range.
699 699 if (FLAG_IS_CMDLINE(MarkStackSize)) {
700 700 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
701 701 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
702 702 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
703 703 "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
704 704 MarkStackSize, (uintx) 1, MarkStackSizeMax);
705 705 return;
706 706 }
707 707 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
708 708 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
709 709 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
710 710 " or for MarkStackSizeMax (" UINTX_FORMAT ")",
711 711 MarkStackSize, MarkStackSizeMax);
712 712 return;
713 713 }
714 714 }
715 715 }
716 716 }
717 717
718 718 if (!_markStack.allocate(MarkStackSize)) {
719 719 warning("Failed to allocate CM marking stack");
720 720 return;
721 721 }
722 722
723 723 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
724 724 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
725 725
726 726 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC);
727 727 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
728 728
729 729 BitMap::idx_t card_bm_size = _card_bm.size();
730 730
731 731 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
732 732 _active_tasks = _max_worker_id;
733 733
734 734 size_t max_regions = (size_t) _g1h->max_regions();
735 735 for (uint i = 0; i < _max_worker_id; ++i) {
736 736 CMTaskQueue* task_queue = new CMTaskQueue();
737 737 task_queue->initialize();
738 738 _task_queues->register_queue(i, task_queue);
739 739
740 740 _count_card_bitmaps[i] = BitMap(card_bm_size, false);
741 741 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
742 742
743 743 _tasks[i] = new CMTask(i, this,
744 744 _count_marked_bytes[i],
745 745 &_count_card_bitmaps[i],
746 746 task_queue, _task_queues);
747 747
748 748 _accum_task_vtime[i] = 0.0;
749 749 }
750 750
751 751 // Calculate the card number for the bottom of the heap. Used
752 752 // in biasing indexes into the accounting card bitmaps.
753 753 _heap_bottom_card_num =
754 754 intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
755 755 CardTableModRefBS::card_shift);
756 756
757 757 // Clear all the liveness counting data
758 758 clear_all_count_data();
759 759
760 760 // so that the call below can read a sensible value
761 761 _heap_start = g1h->reserved_region().start();
762 762 set_non_marking_state();
763 763 _completed_initialization = true;
764 764 }
765 765
766 766 void ConcurrentMark::reset() {
767 767 // Starting values for these two. This should be called in a STW
768 768 // phase.
769 769 MemRegion reserved = _g1h->g1_reserved();
770 770 _heap_start = reserved.start();
771 771 _heap_end = reserved.end();
772 772
773 773 // Separated the asserts so that we know which one fires.
774 774 assert(_heap_start != NULL, "heap bounds should look ok");
775 775 assert(_heap_end != NULL, "heap bounds should look ok");
776 776 assert(_heap_start < _heap_end, "heap bounds should look ok");
777 777
778 778 // Reset all the marking data structures and any necessary flags
779 779 reset_marking_state();
780 780
781 781 if (verbose_low()) {
782 782 gclog_or_tty->print_cr("[global] resetting");
783 783 }
784 784
785 785 // We do reset all of them, since different phases will use
786 786 // different number of active threads. So, it's easiest to have all
787 787 // of them ready.
788 788 for (uint i = 0; i < _max_worker_id; ++i) {
789 789 _tasks[i]->reset(_nextMarkBitMap);
790 790 }
791 791
792 792 // we need this to make sure that the flag is on during the evac
793 793 // pause with initial mark piggy-backed
794 794 set_concurrent_marking_in_progress();
795 795 }
796 796
797 797
798 798 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
799 799 _markStack.set_should_expand();
800 800 _markStack.setEmpty(); // Also clears the _markStack overflow flag
801 801 if (clear_overflow) {
802 802 clear_has_overflown();
803 803 } else {
804 804 assert(has_overflown(), "pre-condition");
805 805 }
806 806 _finger = _heap_start;
807 807
808 808 for (uint i = 0; i < _max_worker_id; ++i) {
809 809 CMTaskQueue* queue = _task_queues->queue(i);
810 810 queue->set_empty();
811 811 }
812 812 }
813 813
814 814 void ConcurrentMark::set_concurrency(uint active_tasks) {
815 815 assert(active_tasks <= _max_worker_id, "we should not have more");
816 816
817 817 _active_tasks = active_tasks;
818 818 // Need to update the three data structures below according to the
819 819 // number of active threads for this phase.
820 820 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
821 821 _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
822 822 _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
823 823 }
824 824
825 825 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
826 826 set_concurrency(active_tasks);
827 827
828 828 _concurrent = concurrent;
829 829 // We propagate this to all tasks, not just the active ones.
830 830 for (uint i = 0; i < _max_worker_id; ++i)
831 831 _tasks[i]->set_concurrent(concurrent);
832 832
833 833 if (concurrent) {
834 834 set_concurrent_marking_in_progress();
835 835 } else {
836 836 // We currently assume that the concurrent flag has been set to
837 837 // false before we start remark. At this point we should also be
838 838 // in a STW phase.
839 839 assert(!concurrent_marking_in_progress(), "invariant");
840 840 assert(out_of_regions(),
841 841 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
842 842 p2i(_finger), p2i(_heap_end)));
843 843 }
844 844 }
845 845
846 846 void ConcurrentMark::set_non_marking_state() {
847 847 // We set the global marking state to some default values when we're
848 848 // not doing marking.
849 849 reset_marking_state();
850 850 _active_tasks = 0;
851 851 clear_concurrent_marking_in_progress();
852 852 }
853 853
854 854 ConcurrentMark::~ConcurrentMark() {
855 855 // The ConcurrentMark instance is never freed.
856 856 ShouldNotReachHere();
857 857 }
858 858
859 859 void ConcurrentMark::clearNextBitmap() {
860 860 G1CollectedHeap* g1h = G1CollectedHeap::heap();
861 861
862 862 // Make sure that the concurrent mark thread looks to still be in
863 863 // the current cycle.
864 864 guarantee(cmThread()->during_cycle(), "invariant");
865 865
866 866 // We are finishing up the current cycle by clearing the next
867 867 // marking bitmap and getting it ready for the next cycle. During
868 868 // this time no other cycle can start. So, let's make sure that this
869 869 // is the case.
870 870 guarantee(!g1h->mark_in_progress(), "invariant");
871 871
872 872 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
873 873 g1h->heap_region_iterate(&cl);
874 874
875 875 // Clear the liveness counting data. If the marking has been aborted, the abort()
876 876 // call already did that.
877 877 if (cl.complete()) {
878 878 clear_all_count_data();
879 879 }
880 880
881 881 // Repeat the asserts from above.
882 882 guarantee(cmThread()->during_cycle(), "invariant");
883 883 guarantee(!g1h->mark_in_progress(), "invariant");
884 884 }
885 885
886 886 class CheckBitmapClearHRClosure : public HeapRegionClosure {
887 887 CMBitMap* _bitmap;
888 888 bool _error;
889 889 public:
890 890 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
891 891 }
892 892
893 893 virtual bool doHeapRegion(HeapRegion* r) {
894 894 // This closure can be called concurrently to the mutator, so we must make sure
895 895 // that the result of the getNextMarkedWordAddress() call is compared to the
896 896 // value passed to it as limit to detect any found bits.
897 897 // We can use the region's orig_end() for the limit and the comparison value
898 898 // as it always contains the "real" end of the region that never changes and
899 899 // has no side effects.
900 900 // Due to the latter, there can also be no problem with the compiler generating
901 901 // reloads of the orig_end() call.
902 902 HeapWord* end = r->orig_end();
903 903 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
904 904 }
905 905 };
906 906
907 907 bool ConcurrentMark::nextMarkBitmapIsClear() {
908 908 CheckBitmapClearHRClosure cl(_nextMarkBitMap);
909 909 _g1h->heap_region_iterate(&cl);
910 910 return cl.complete();
911 911 }
912 912
913 913 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
914 914 public:
915 915 bool doHeapRegion(HeapRegion* r) {
916 916 if (!r->continuesHumongous()) {
917 917 r->note_start_of_marking();
918 918 }
919 919 return false;
920 920 }
921 921 };
922 922
923 923 void ConcurrentMark::checkpointRootsInitialPre() {
924 924 G1CollectedHeap* g1h = G1CollectedHeap::heap();
925 925 G1CollectorPolicy* g1p = g1h->g1_policy();
926 926
927 927 _has_aborted = false;
928 928
929 929 #ifndef PRODUCT
930 930 if (G1PrintReachableAtInitialMark) {
931 931 print_reachable("at-cycle-start",
932 932 VerifyOption_G1UsePrevMarking, true /* all */);
933 933 }
934 934 #endif
935 935
936 936 // Initialise marking structures. This has to be done in a STW phase.
937 937 reset();
938 938
939 939 // For each region note start of marking.
940 940 NoteStartOfMarkHRClosure startcl;
941 941 g1h->heap_region_iterate(&startcl);
942 942 }
943 943
944 944
945 945 void ConcurrentMark::checkpointRootsInitialPost() {
946 946 G1CollectedHeap* g1h = G1CollectedHeap::heap();
947 947
948 948 // If we force an overflow during remark, the remark operation will
949 949 // actually abort and we'll restart concurrent marking. If we always
950 950 // force an oveflow during remark we'll never actually complete the
951 951 // marking phase. So, we initilize this here, at the start of the
952 952 // cycle, so that at the remaining overflow number will decrease at
953 953 // every remark and we'll eventually not need to cause one.
954 954 force_overflow_stw()->init();
955 955
956 956 // Start Concurrent Marking weak-reference discovery.
957 957 ReferenceProcessor* rp = g1h->ref_processor_cm();
958 958 // enable ("weak") refs discovery
959 959 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
960 960 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
961 961
962 962 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
963 963 // This is the start of the marking cycle, we're expected all
964 964 // threads to have SATB queues with active set to false.
965 965 satb_mq_set.set_active_all_threads(true, /* new active value */
966 966 false /* expected_active */);
967 967
968 968 _root_regions.prepare_for_scan();
969 969
970 970 // update_g1_committed() will be called at the end of an evac pause
971 971 // when marking is on. So, it's also called at the end of the
972 972 // initial-mark pause to update the heap end, if the heap expands
973 973 // during it. No need to call it here.
974 974 }
975 975
976 976 /*
977 977 * Notice that in the next two methods, we actually leave the STS
978 978 * during the barrier sync and join it immediately afterwards. If we
979 979 * do not do this, the following deadlock can occur: one thread could
980 980 * be in the barrier sync code, waiting for the other thread to also
981 981 * sync up, whereas another one could be trying to yield, while also
982 982 * waiting for the other threads to sync up too.
983 983 *
984 984 * Note, however, that this code is also used during remark and in
985 985 * this case we should not attempt to leave / enter the STS, otherwise
986 986 * we'll either hit an asseert (debug / fastdebug) or deadlock
987 987 * (product). So we should only leave / enter the STS if we are
988 988 * operating concurrently.
989 989 *
990 990 * Because the thread that does the sync barrier has left the STS, it
991 991 * is possible to be suspended for a Full GC or an evacuation pause
992 992 * could occur. This is actually safe, since the entering the sync
993 993 * barrier is one of the last things do_marking_step() does, and it
994 994 * doesn't manipulate any data structures afterwards.
995 995 */
996 996
997 997 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
998 998 if (verbose_low()) {
999 999 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
1000 1000 }
1001 1001
1002 1002 if (concurrent()) {
1003 1003 SuspendibleThreadSet::leave();
1004 1004 }
1005 1005
1006 1006 bool barrier_aborted = !_first_overflow_barrier_sync.enter();
1007 1007
1008 1008 if (concurrent()) {
1009 1009 SuspendibleThreadSet::join();
1010 1010 }
1011 1011 // at this point everyone should have synced up and not be doing any
1012 1012 // more work
1013 1013
1014 1014 if (verbose_low()) {
1015 1015 if (barrier_aborted) {
1016 1016 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1017 1017 } else {
1018 1018 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1019 1019 }
1020 1020 }
1021 1021
1022 1022 if (barrier_aborted) {
1023 1023 // If the barrier aborted we ignore the overflow condition and
1024 1024 // just abort the whole marking phase as quickly as possible.
1025 1025 return;
1026 1026 }
1027 1027
1028 1028 // If we're executing the concurrent phase of marking, reset the marking
1029 1029 // state; otherwise the marking state is reset after reference processing,
1030 1030 // during the remark pause.
1031 1031 // If we reset here as a result of an overflow during the remark we will
1032 1032 // see assertion failures from any subsequent set_concurrency_and_phase()
1033 1033 // calls.
1034 1034 if (concurrent()) {
1035 1035 // let the task associated with with worker 0 do this
1036 1036 if (worker_id == 0) {
1037 1037 // task 0 is responsible for clearing the global data structures
1038 1038 // We should be here because of an overflow. During STW we should
1039 1039 // not clear the overflow flag since we rely on it being true when
1040 1040 // we exit this method to abort the pause and restart concurent
1041 1041 // marking.
1042 1042 reset_marking_state(true /* clear_overflow */);
1043 1043 force_overflow()->update();
1044 1044
1045 1045 if (G1Log::fine()) {
1046 1046 gclog_or_tty->gclog_stamp(concurrent_gc_id());
1047 1047 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1048 1048 }
1049 1049 }
1050 1050 }
1051 1051
1052 1052 // after this, each task should reset its own data structures then
1053 1053 // then go into the second barrier
1054 1054 }
1055 1055
1056 1056 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1057 1057 if (verbose_low()) {
1058 1058 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1059 1059 }
1060 1060
1061 1061 if (concurrent()) {
1062 1062 SuspendibleThreadSet::leave();
1063 1063 }
1064 1064
1065 1065 bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1066 1066
1067 1067 if (concurrent()) {
1068 1068 SuspendibleThreadSet::join();
1069 1069 }
1070 1070 // at this point everything should be re-initialized and ready to go
1071 1071
1072 1072 if (verbose_low()) {
1073 1073 if (barrier_aborted) {
1074 1074 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1075 1075 } else {
1076 1076 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1077 1077 }
1078 1078 }
1079 1079 }
1080 1080
1081 1081 #ifndef PRODUCT
1082 1082 void ForceOverflowSettings::init() {
1083 1083 _num_remaining = G1ConcMarkForceOverflow;
1084 1084 _force = false;
1085 1085 update();
1086 1086 }
1087 1087
1088 1088 void ForceOverflowSettings::update() {
1089 1089 if (_num_remaining > 0) {
1090 1090 _num_remaining -= 1;
1091 1091 _force = true;
1092 1092 } else {
1093 1093 _force = false;
1094 1094 }
1095 1095 }
1096 1096
1097 1097 bool ForceOverflowSettings::should_force() {
1098 1098 if (_force) {
1099 1099 _force = false;
1100 1100 return true;
1101 1101 } else {
1102 1102 return false;
1103 1103 }
1104 1104 }
1105 1105 #endif // !PRODUCT
1106 1106
1107 1107 class CMConcurrentMarkingTask: public AbstractGangTask {
1108 1108 private:
1109 1109 ConcurrentMark* _cm;
1110 1110 ConcurrentMarkThread* _cmt;
1111 1111
1112 1112 public:
1113 1113 void work(uint worker_id) {
1114 1114 assert(Thread::current()->is_ConcurrentGC_thread(),
1115 1115 "this should only be done by a conc GC thread");
1116 1116 ResourceMark rm;
1117 1117
1118 1118 double start_vtime = os::elapsedVTime();
1119 1119
1120 1120 SuspendibleThreadSet::join();
1121 1121
1122 1122 assert(worker_id < _cm->active_tasks(), "invariant");
1123 1123 CMTask* the_task = _cm->task(worker_id);
1124 1124 the_task->record_start_time();
1125 1125 if (!_cm->has_aborted()) {
1126 1126 do {
1127 1127 double start_vtime_sec = os::elapsedVTime();
1128 1128 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1129 1129
1130 1130 the_task->do_marking_step(mark_step_duration_ms,
1131 1131 true /* do_termination */,
1132 1132 false /* is_serial*/);
1133 1133
1134 1134 double end_vtime_sec = os::elapsedVTime();
1135 1135 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1136 1136 _cm->clear_has_overflown();
1137 1137
1138 1138 _cm->do_yield_check(worker_id);
1139 1139
1140 1140 jlong sleep_time_ms;
1141 1141 if (!_cm->has_aborted() && the_task->has_aborted()) {
1142 1142 sleep_time_ms =
1143 1143 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1144 1144 SuspendibleThreadSet::leave();
1145 1145 os::sleep(Thread::current(), sleep_time_ms, false);
1146 1146 SuspendibleThreadSet::join();
1147 1147 }
1148 1148 } while (!_cm->has_aborted() && the_task->has_aborted());
1149 1149 }
1150 1150 the_task->record_end_time();
1151 1151 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1152 1152
1153 1153 SuspendibleThreadSet::leave();
1154 1154
1155 1155 double end_vtime = os::elapsedVTime();
1156 1156 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1157 1157 }
1158 1158
1159 1159 CMConcurrentMarkingTask(ConcurrentMark* cm,
1160 1160 ConcurrentMarkThread* cmt) :
1161 1161 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1162 1162
1163 1163 ~CMConcurrentMarkingTask() { }
1164 1164 };
1165 1165
1166 1166 // Calculates the number of active workers for a concurrent
1167 1167 // phase.
1168 1168 uint ConcurrentMark::calc_parallel_marking_threads() {
1169 1169 if (G1CollectedHeap::use_parallel_gc_threads()) {
1170 1170 uint n_conc_workers = 0;
1171 1171 if (!UseDynamicNumberOfGCThreads ||
1172 1172 (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1173 1173 !ForceDynamicNumberOfGCThreads)) {
1174 1174 n_conc_workers = max_parallel_marking_threads();
1175 1175 } else {
1176 1176 n_conc_workers =
1177 1177 AdaptiveSizePolicy::calc_default_active_workers(
1178 1178 max_parallel_marking_threads(),
1179 1179 1, /* Minimum workers */
1180 1180 parallel_marking_threads(),
1181 1181 Threads::number_of_non_daemon_threads());
1182 1182 // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1183 1183 // that scaling has already gone into "_max_parallel_marking_threads".
1184 1184 }
1185 1185 assert(n_conc_workers > 0, "Always need at least 1");
1186 1186 return n_conc_workers;
1187 1187 }
1188 1188 // If we are not running with any parallel GC threads we will not
1189 1189 // have spawned any marking threads either. Hence the number of
1190 1190 // concurrent workers should be 0.
1191 1191 return 0;
1192 1192 }
1193 1193
1194 1194 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1195 1195 // Currently, only survivors can be root regions.
1196 1196 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1197 1197 G1RootRegionScanClosure cl(_g1h, this, worker_id);
1198 1198
1199 1199 const uintx interval = PrefetchScanIntervalInBytes;
1200 1200 HeapWord* curr = hr->bottom();
1201 1201 const HeapWord* end = hr->top();
1202 1202 while (curr < end) {
1203 1203 Prefetch::read(curr, interval);
1204 1204 oop obj = oop(curr);
1205 1205 int size = obj->oop_iterate(&cl);
1206 1206 assert(size == obj->size(), "sanity");
1207 1207 curr += size;
1208 1208 }
1209 1209 }
1210 1210
1211 1211 class CMRootRegionScanTask : public AbstractGangTask {
1212 1212 private:
1213 1213 ConcurrentMark* _cm;
1214 1214
1215 1215 public:
1216 1216 CMRootRegionScanTask(ConcurrentMark* cm) :
1217 1217 AbstractGangTask("Root Region Scan"), _cm(cm) { }
1218 1218
1219 1219 void work(uint worker_id) {
1220 1220 assert(Thread::current()->is_ConcurrentGC_thread(),
1221 1221 "this should only be done by a conc GC thread");
1222 1222
1223 1223 CMRootRegions* root_regions = _cm->root_regions();
1224 1224 HeapRegion* hr = root_regions->claim_next();
1225 1225 while (hr != NULL) {
1226 1226 _cm->scanRootRegion(hr, worker_id);
1227 1227 hr = root_regions->claim_next();
1228 1228 }
1229 1229 }
1230 1230 };
1231 1231
1232 1232 void ConcurrentMark::scanRootRegions() {
1233 1233 // Start of concurrent marking.
1234 1234 ClassLoaderDataGraph::clear_claimed_marks();
1235 1235
1236 1236 // scan_in_progress() will have been set to true only if there was
1237 1237 // at least one root region to scan. So, if it's false, we
1238 1238 // should not attempt to do any further work.
1239 1239 if (root_regions()->scan_in_progress()) {
1240 1240 _parallel_marking_threads = calc_parallel_marking_threads();
1241 1241 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1242 1242 "Maximum number of marking threads exceeded");
1243 1243 uint active_workers = MAX2(1U, parallel_marking_threads());
1244 1244
1245 1245 CMRootRegionScanTask task(this);
1246 1246 if (use_parallel_marking_threads()) {
1247 1247 _parallel_workers->set_active_workers((int) active_workers);
1248 1248 _parallel_workers->run_task(&task);
1249 1249 } else {
1250 1250 task.work(0);
1251 1251 }
1252 1252
1253 1253 // It's possible that has_aborted() is true here without actually
1254 1254 // aborting the survivor scan earlier. This is OK as it's
1255 1255 // mainly used for sanity checking.
1256 1256 root_regions()->scan_finished();
1257 1257 }
1258 1258 }
1259 1259
1260 1260 void ConcurrentMark::markFromRoots() {
1261 1261 // we might be tempted to assert that:
1262 1262 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1263 1263 // "inconsistent argument?");
1264 1264 // However that wouldn't be right, because it's possible that
1265 1265 // a safepoint is indeed in progress as a younger generation
1266 1266 // stop-the-world GC happens even as we mark in this generation.
1267 1267
1268 1268 _restart_for_overflow = false;
1269 1269 force_overflow_conc()->init();
1270 1270
1271 1271 // _g1h has _n_par_threads
1272 1272 _parallel_marking_threads = calc_parallel_marking_threads();
1273 1273 assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1274 1274 "Maximum number of marking threads exceeded");
1275 1275
1276 1276 uint active_workers = MAX2(1U, parallel_marking_threads());
1277 1277
1278 1278 // Parallel task terminator is set in "set_concurrency_and_phase()"
1279 1279 set_concurrency_and_phase(active_workers, true /* concurrent */);
1280 1280
1281 1281 CMConcurrentMarkingTask markingTask(this, cmThread());
1282 1282 if (use_parallel_marking_threads()) {
1283 1283 _parallel_workers->set_active_workers((int)active_workers);
1284 1284 // Don't set _n_par_threads because it affects MT in process_roots()
1285 1285 // and the decisions on that MT processing is made elsewhere.
1286 1286 assert(_parallel_workers->active_workers() > 0, "Should have been set");
1287 1287 _parallel_workers->run_task(&markingTask);
1288 1288 } else {
1289 1289 markingTask.work(0);
1290 1290 }
1291 1291 print_stats();
1292 1292 }
1293 1293
1294 1294 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1295 1295 // world is stopped at this checkpoint
1296 1296 assert(SafepointSynchronize::is_at_safepoint(),
1297 1297 "world should be stopped");
1298 1298
1299 1299 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1300 1300
1301 1301 // If a full collection has happened, we shouldn't do this.
1302 1302 if (has_aborted()) {
1303 1303 g1h->set_marking_complete(); // So bitmap clearing isn't confused
1304 1304 return;
1305 1305 }
1306 1306
1307 1307 SvcGCMarker sgcm(SvcGCMarker::OTHER);
1308 1308
1309 1309 if (VerifyDuringGC) {
1310 1310 HandleMark hm; // handle scope
1311 1311 Universe::heap()->prepare_for_verify();
1312 1312 Universe::verify(VerifyOption_G1UsePrevMarking,
1313 1313 " VerifyDuringGC:(before)");
1314 1314 }
1315 1315 g1h->check_bitmaps("Remark Start");
1316 1316
1317 1317 G1CollectorPolicy* g1p = g1h->g1_policy();
1318 1318 g1p->record_concurrent_mark_remark_start();
1319 1319
1320 1320 double start = os::elapsedTime();
1321 1321
1322 1322 checkpointRootsFinalWork();
1323 1323
1324 1324 double mark_work_end = os::elapsedTime();
1325 1325
1326 1326 weakRefsWork(clear_all_soft_refs);
1327 1327
1328 1328 if (has_overflown()) {
1329 1329 // Oops. We overflowed. Restart concurrent marking.
1330 1330 _restart_for_overflow = true;
1331 1331 if (G1TraceMarkStackOverflow) {
1332 1332 gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1333 1333 }
1334 1334
1335 1335 // Verify the heap w.r.t. the previous marking bitmap.
1336 1336 if (VerifyDuringGC) {
1337 1337 HandleMark hm; // handle scope
1338 1338 Universe::heap()->prepare_for_verify();
1339 1339 Universe::verify(VerifyOption_G1UsePrevMarking,
1340 1340 " VerifyDuringGC:(overflow)");
1341 1341 }
1342 1342
1343 1343 // Clear the marking state because we will be restarting
1344 1344 // marking due to overflowing the global mark stack.
1345 1345 reset_marking_state();
1346 1346 } else {
1347 1347 // Aggregate the per-task counting data that we have accumulated
1348 1348 // while marking.
1349 1349 aggregate_count_data();
1350 1350
1351 1351 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1352 1352 // We're done with marking.
1353 1353 // This is the end of the marking cycle, we're expected all
1354 1354 // threads to have SATB queues with active set to true.
1355 1355 satb_mq_set.set_active_all_threads(false, /* new active value */
1356 1356 true /* expected_active */);
1357 1357
1358 1358 if (VerifyDuringGC) {
1359 1359 HandleMark hm; // handle scope
1360 1360 Universe::heap()->prepare_for_verify();
1361 1361 Universe::verify(VerifyOption_G1UseNextMarking,
1362 1362 " VerifyDuringGC:(after)");
1363 1363 }
1364 1364 g1h->check_bitmaps("Remark End");
1365 1365 assert(!restart_for_overflow(), "sanity");
1366 1366 // Completely reset the marking state since marking completed
1367 1367 set_non_marking_state();
1368 1368 }
1369 1369
1370 1370 // Expand the marking stack, if we have to and if we can.
1371 1371 if (_markStack.should_expand()) {
1372 1372 _markStack.expand();
1373 1373 }
1374 1374
1375 1375 // Statistics
1376 1376 double now = os::elapsedTime();
1377 1377 _remark_mark_times.add((mark_work_end - start) * 1000.0);
1378 1378 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1379 1379 _remark_times.add((now - start) * 1000.0);
1380 1380
1381 1381 g1p->record_concurrent_mark_remark_end();
1382 1382
1383 1383 G1CMIsAliveClosure is_alive(g1h);
1384 1384 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1385 1385 }
1386 1386
1387 1387 // Base class of the closures that finalize and verify the
1388 1388 // liveness counting data.
1389 1389 class CMCountDataClosureBase: public HeapRegionClosure {
1390 1390 protected:
1391 1391 G1CollectedHeap* _g1h;
1392 1392 ConcurrentMark* _cm;
1393 1393 CardTableModRefBS* _ct_bs;
1394 1394
1395 1395 BitMap* _region_bm;
1396 1396 BitMap* _card_bm;
1397 1397
1398 1398 // Takes a region that's not empty (i.e., it has at least one
1399 1399 // live object in it and sets its corresponding bit on the region
1400 1400 // bitmap to 1. If the region is "starts humongous" it will also set
1401 1401 // to 1 the bits on the region bitmap that correspond to its
1402 1402 // associated "continues humongous" regions.
1403 1403 void set_bit_for_region(HeapRegion* hr) {
1404 1404 assert(!hr->continuesHumongous(), "should have filtered those out");
1405 1405
1406 1406 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1407 1407 if (!hr->startsHumongous()) {
1408 1408 // Normal (non-humongous) case: just set the bit.
1409 1409 _region_bm->par_at_put(index, true);
1410 1410 } else {
1411 1411 // Starts humongous case: calculate how many regions are part of
1412 1412 // this humongous region and then set the bit range.
1413 1413 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1414 1414 _region_bm->par_at_put_range(index, end_index, true);
1415 1415 }
1416 1416 }
1417 1417
1418 1418 public:
1419 1419 CMCountDataClosureBase(G1CollectedHeap* g1h,
1420 1420 BitMap* region_bm, BitMap* card_bm):
1421 1421 _g1h(g1h), _cm(g1h->concurrent_mark()),
1422 1422 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1423 1423 _region_bm(region_bm), _card_bm(card_bm) { }
1424 1424 };
1425 1425
1426 1426 // Closure that calculates the # live objects per region. Used
1427 1427 // for verification purposes during the cleanup pause.
1428 1428 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1429 1429 CMBitMapRO* _bm;
1430 1430 size_t _region_marked_bytes;
1431 1431
1432 1432 public:
1433 1433 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1434 1434 BitMap* region_bm, BitMap* card_bm) :
1435 1435 CMCountDataClosureBase(g1h, region_bm, card_bm),
1436 1436 _bm(bm), _region_marked_bytes(0) { }
1437 1437
1438 1438 bool doHeapRegion(HeapRegion* hr) {
1439 1439
1440 1440 if (hr->continuesHumongous()) {
1441 1441 // We will ignore these here and process them when their
1442 1442 // associated "starts humongous" region is processed (see
1443 1443 // set_bit_for_heap_region()). Note that we cannot rely on their
1444 1444 // associated "starts humongous" region to have their bit set to
1445 1445 // 1 since, due to the region chunking in the parallel region
1446 1446 // iteration, a "continues humongous" region might be visited
1447 1447 // before its associated "starts humongous".
1448 1448 return false;
1449 1449 }
1450 1450
1451 1451 HeapWord* ntams = hr->next_top_at_mark_start();
1452 1452 HeapWord* start = hr->bottom();
1453 1453
1454 1454 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1455 1455 err_msg("Preconditions not met - "
1456 1456 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1457 1457 p2i(start), p2i(ntams), p2i(hr->end())));
1458 1458
1459 1459 // Find the first marked object at or after "start".
1460 1460 start = _bm->getNextMarkedWordAddress(start, ntams);
1461 1461
1462 1462 size_t marked_bytes = 0;
1463 1463
1464 1464 while (start < ntams) {
1465 1465 oop obj = oop(start);
1466 1466 int obj_sz = obj->size();
1467 1467 HeapWord* obj_end = start + obj_sz;
1468 1468
1469 1469 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1470 1470 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1471 1471
1472 1472 // Note: if we're looking at the last region in heap - obj_end
1473 1473 // could be actually just beyond the end of the heap; end_idx
1474 1474 // will then correspond to a (non-existent) card that is also
1475 1475 // just beyond the heap.
1476 1476 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1477 1477 // end of object is not card aligned - increment to cover
1478 1478 // all the cards spanned by the object
1479 1479 end_idx += 1;
1480 1480 }
1481 1481
1482 1482 // Set the bits in the card BM for the cards spanned by this object.
1483 1483 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1484 1484
1485 1485 // Add the size of this object to the number of marked bytes.
1486 1486 marked_bytes += (size_t)obj_sz * HeapWordSize;
1487 1487
1488 1488 // Find the next marked object after this one.
1489 1489 start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1490 1490 }
1491 1491
1492 1492 // Mark the allocated-since-marking portion...
1493 1493 HeapWord* top = hr->top();
1494 1494 if (ntams < top) {
1495 1495 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1496 1496 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1497 1497
1498 1498 // Note: if we're looking at the last region in heap - top
1499 1499 // could be actually just beyond the end of the heap; end_idx
1500 1500 // will then correspond to a (non-existent) card that is also
1501 1501 // just beyond the heap.
1502 1502 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1503 1503 // end of object is not card aligned - increment to cover
1504 1504 // all the cards spanned by the object
1505 1505 end_idx += 1;
1506 1506 }
1507 1507 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1508 1508
1509 1509 // This definitely means the region has live objects.
1510 1510 set_bit_for_region(hr);
1511 1511 }
1512 1512
1513 1513 // Update the live region bitmap.
1514 1514 if (marked_bytes > 0) {
1515 1515 set_bit_for_region(hr);
1516 1516 }
1517 1517
1518 1518 // Set the marked bytes for the current region so that
1519 1519 // it can be queried by a calling verificiation routine
1520 1520 _region_marked_bytes = marked_bytes;
1521 1521
1522 1522 return false;
1523 1523 }
1524 1524
1525 1525 size_t region_marked_bytes() const { return _region_marked_bytes; }
1526 1526 };
1527 1527
1528 1528 // Heap region closure used for verifying the counting data
1529 1529 // that was accumulated concurrently and aggregated during
1530 1530 // the remark pause. This closure is applied to the heap
1531 1531 // regions during the STW cleanup pause.
1532 1532
1533 1533 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1534 1534 G1CollectedHeap* _g1h;
1535 1535 ConcurrentMark* _cm;
1536 1536 CalcLiveObjectsClosure _calc_cl;
1537 1537 BitMap* _region_bm; // Region BM to be verified
1538 1538 BitMap* _card_bm; // Card BM to be verified
1539 1539 bool _verbose; // verbose output?
1540 1540
1541 1541 BitMap* _exp_region_bm; // Expected Region BM values
1542 1542 BitMap* _exp_card_bm; // Expected card BM values
1543 1543
1544 1544 int _failures;
1545 1545
1546 1546 public:
1547 1547 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1548 1548 BitMap* region_bm,
1549 1549 BitMap* card_bm,
1550 1550 BitMap* exp_region_bm,
1551 1551 BitMap* exp_card_bm,
1552 1552 bool verbose) :
1553 1553 _g1h(g1h), _cm(g1h->concurrent_mark()),
1554 1554 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1555 1555 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1556 1556 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1557 1557 _failures(0) { }
1558 1558
1559 1559 int failures() const { return _failures; }
1560 1560
1561 1561 bool doHeapRegion(HeapRegion* hr) {
1562 1562 if (hr->continuesHumongous()) {
1563 1563 // We will ignore these here and process them when their
1564 1564 // associated "starts humongous" region is processed (see
1565 1565 // set_bit_for_heap_region()). Note that we cannot rely on their
1566 1566 // associated "starts humongous" region to have their bit set to
1567 1567 // 1 since, due to the region chunking in the parallel region
1568 1568 // iteration, a "continues humongous" region might be visited
1569 1569 // before its associated "starts humongous".
1570 1570 return false;
1571 1571 }
1572 1572
1573 1573 int failures = 0;
1574 1574
1575 1575 // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1576 1576 // this region and set the corresponding bits in the expected region
1577 1577 // and card bitmaps.
1578 1578 bool res = _calc_cl.doHeapRegion(hr);
1579 1579 assert(res == false, "should be continuing");
1580 1580
1581 1581 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1582 1582 Mutex::_no_safepoint_check_flag);
1583 1583
1584 1584 // Verify the marked bytes for this region.
1585 1585 size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1586 1586 size_t act_marked_bytes = hr->next_marked_bytes();
1587 1587
1588 1588 // We're not OK if expected marked bytes > actual marked bytes. It means
1589 1589 // we have missed accounting some objects during the actual marking.
1590 1590 if (exp_marked_bytes > act_marked_bytes) {
1591 1591 if (_verbose) {
1592 1592 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1593 1593 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1594 1594 hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1595 1595 }
1596 1596 failures += 1;
1597 1597 }
1598 1598
1599 1599 // Verify the bit, for this region, in the actual and expected
1600 1600 // (which was just calculated) region bit maps.
1601 1601 // We're not OK if the bit in the calculated expected region
1602 1602 // bitmap is set and the bit in the actual region bitmap is not.
1603 1603 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1604 1604
1605 1605 bool expected = _exp_region_bm->at(index);
1606 1606 bool actual = _region_bm->at(index);
1607 1607 if (expected && !actual) {
1608 1608 if (_verbose) {
1609 1609 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1610 1610 "expected: %s, actual: %s",
1611 1611 hr->hrm_index(),
1612 1612 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1613 1613 }
1614 1614 failures += 1;
1615 1615 }
1616 1616
1617 1617 // Verify that the card bit maps for the cards spanned by the current
1618 1618 // region match. We have an error if we have a set bit in the expected
1619 1619 // bit map and the corresponding bit in the actual bitmap is not set.
1620 1620
1621 1621 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1622 1622 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1623 1623
1624 1624 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1625 1625 expected = _exp_card_bm->at(i);
1626 1626 actual = _card_bm->at(i);
1627 1627
1628 1628 if (expected && !actual) {
1629 1629 if (_verbose) {
1630 1630 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1631 1631 "expected: %s, actual: %s",
1632 1632 hr->hrm_index(), i,
1633 1633 BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1634 1634 }
1635 1635 failures += 1;
1636 1636 }
1637 1637 }
1638 1638
1639 1639 if (failures > 0 && _verbose) {
1640 1640 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1641 1641 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1642 1642 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1643 1643 _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1644 1644 }
1645 1645
1646 1646 _failures += failures;
1647 1647
1648 1648 // We could stop iteration over the heap when we
1649 1649 // find the first violating region by returning true.
1650 1650 return false;
1651 1651 }
1652 1652 };
1653 1653
1654 1654 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1655 1655 protected:
1656 1656 G1CollectedHeap* _g1h;
1657 1657 ConcurrentMark* _cm;
1658 1658 BitMap* _actual_region_bm;
1659 1659 BitMap* _actual_card_bm;
1660 1660
1661 1661 uint _n_workers;
1662 1662
1663 1663 BitMap* _expected_region_bm;
1664 1664 BitMap* _expected_card_bm;
1665 1665
1666 1666 int _failures;
1667 1667 bool _verbose;
1668 1668
1669 1669 public:
1670 1670 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1671 1671 BitMap* region_bm, BitMap* card_bm,
1672 1672 BitMap* expected_region_bm, BitMap* expected_card_bm)
1673 1673 : AbstractGangTask("G1 verify final counting"),
1674 1674 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1675 1675 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1676 1676 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1677 1677 _failures(0), _verbose(false),
1678 1678 _n_workers(0) {
1679 1679 assert(VerifyDuringGC, "don't call this otherwise");
1680 1680
1681 1681 // Use the value already set as the number of active threads
1682 1682 // in the call to run_task().
1683 1683 if (G1CollectedHeap::use_parallel_gc_threads()) {
1684 1684 assert( _g1h->workers()->active_workers() > 0,
1685 1685 "Should have been previously set");
1686 1686 _n_workers = _g1h->workers()->active_workers();
1687 1687 } else {
1688 1688 _n_workers = 1;
1689 1689 }
1690 1690
1691 1691 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1692 1692 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1693 1693
1694 1694 _verbose = _cm->verbose_medium();
1695 1695 }
1696 1696
1697 1697 void work(uint worker_id) {
1698 1698 assert(worker_id < _n_workers, "invariant");
1699 1699
1700 1700 VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1701 1701 _actual_region_bm, _actual_card_bm,
1702 1702 _expected_region_bm,
1703 1703 _expected_card_bm,
1704 1704 _verbose);
1705 1705
1706 1706 if (G1CollectedHeap::use_parallel_gc_threads()) {
1707 1707 _g1h->heap_region_par_iterate_chunked(&verify_cl,
1708 1708 worker_id,
1709 1709 _n_workers,
1710 1710 HeapRegion::VerifyCountClaimValue);
1711 1711 } else {
1712 1712 _g1h->heap_region_iterate(&verify_cl);
1713 1713 }
1714 1714
1715 1715 Atomic::add(verify_cl.failures(), &_failures);
1716 1716 }
1717 1717
1718 1718 int failures() const { return _failures; }
1719 1719 };
1720 1720
1721 1721 // Closure that finalizes the liveness counting data.
1722 1722 // Used during the cleanup pause.
1723 1723 // Sets the bits corresponding to the interval [NTAMS, top]
1724 1724 // (which contains the implicitly live objects) in the
1725 1725 // card liveness bitmap. Also sets the bit for each region,
1726 1726 // containing live data, in the region liveness bitmap.
1727 1727
1728 1728 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1729 1729 public:
1730 1730 FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1731 1731 BitMap* region_bm,
1732 1732 BitMap* card_bm) :
1733 1733 CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1734 1734
1735 1735 bool doHeapRegion(HeapRegion* hr) {
1736 1736
1737 1737 if (hr->continuesHumongous()) {
1738 1738 // We will ignore these here and process them when their
1739 1739 // associated "starts humongous" region is processed (see
1740 1740 // set_bit_for_heap_region()). Note that we cannot rely on their
1741 1741 // associated "starts humongous" region to have their bit set to
1742 1742 // 1 since, due to the region chunking in the parallel region
1743 1743 // iteration, a "continues humongous" region might be visited
1744 1744 // before its associated "starts humongous".
1745 1745 return false;
1746 1746 }
1747 1747
1748 1748 HeapWord* ntams = hr->next_top_at_mark_start();
1749 1749 HeapWord* top = hr->top();
1750 1750
1751 1751 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1752 1752
1753 1753 // Mark the allocated-since-marking portion...
1754 1754 if (ntams < top) {
1755 1755 // This definitely means the region has live objects.
1756 1756 set_bit_for_region(hr);
1757 1757
1758 1758 // Now set the bits in the card bitmap for [ntams, top)
1759 1759 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1760 1760 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1761 1761
1762 1762 // Note: if we're looking at the last region in heap - top
1763 1763 // could be actually just beyond the end of the heap; end_idx
1764 1764 // will then correspond to a (non-existent) card that is also
1765 1765 // just beyond the heap.
1766 1766 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1767 1767 // end of object is not card aligned - increment to cover
1768 1768 // all the cards spanned by the object
1769 1769 end_idx += 1;
1770 1770 }
1771 1771
1772 1772 assert(end_idx <= _card_bm->size(),
1773 1773 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1774 1774 end_idx, _card_bm->size()));
1775 1775 assert(start_idx < _card_bm->size(),
1776 1776 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1777 1777 start_idx, _card_bm->size()));
1778 1778
1779 1779 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1780 1780 }
1781 1781
1782 1782 // Set the bit for the region if it contains live data
1783 1783 if (hr->next_marked_bytes() > 0) {
1784 1784 set_bit_for_region(hr);
1785 1785 }
1786 1786
1787 1787 return false;
1788 1788 }
1789 1789 };
1790 1790
1791 1791 class G1ParFinalCountTask: public AbstractGangTask {
1792 1792 protected:
1793 1793 G1CollectedHeap* _g1h;
1794 1794 ConcurrentMark* _cm;
1795 1795 BitMap* _actual_region_bm;
1796 1796 BitMap* _actual_card_bm;
1797 1797
1798 1798 uint _n_workers;
1799 1799
1800 1800 public:
1801 1801 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1802 1802 : AbstractGangTask("G1 final counting"),
1803 1803 _g1h(g1h), _cm(_g1h->concurrent_mark()),
1804 1804 _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1805 1805 _n_workers(0) {
1806 1806 // Use the value already set as the number of active threads
1807 1807 // in the call to run_task().
1808 1808 if (G1CollectedHeap::use_parallel_gc_threads()) {
1809 1809 assert( _g1h->workers()->active_workers() > 0,
1810 1810 "Should have been previously set");
1811 1811 _n_workers = _g1h->workers()->active_workers();
1812 1812 } else {
1813 1813 _n_workers = 1;
1814 1814 }
1815 1815 }
1816 1816
1817 1817 void work(uint worker_id) {
1818 1818 assert(worker_id < _n_workers, "invariant");
1819 1819
1820 1820 FinalCountDataUpdateClosure final_update_cl(_g1h,
1821 1821 _actual_region_bm,
1822 1822 _actual_card_bm);
1823 1823
1824 1824 if (G1CollectedHeap::use_parallel_gc_threads()) {
1825 1825 _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1826 1826 worker_id,
1827 1827 _n_workers,
1828 1828 HeapRegion::FinalCountClaimValue);
1829 1829 } else {
1830 1830 _g1h->heap_region_iterate(&final_update_cl);
1831 1831 }
1832 1832 }
1833 1833 };
1834 1834
1835 1835 class G1ParNoteEndTask;
1836 1836
1837 1837 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1838 1838 G1CollectedHeap* _g1;
1839 1839 size_t _max_live_bytes;
1840 1840 uint _regions_claimed;
1841 1841 size_t _freed_bytes;
1842 1842 FreeRegionList* _local_cleanup_list;
1843 1843 HeapRegionSetCount _old_regions_removed;
1844 1844 HeapRegionSetCount _humongous_regions_removed;
1845 1845 HRRSCleanupTask* _hrrs_cleanup_task;
1846 1846 double _claimed_region_time;
1847 1847 double _max_region_time;
1848 1848
1849 1849 public:
1850 1850 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1851 1851 FreeRegionList* local_cleanup_list,
1852 1852 HRRSCleanupTask* hrrs_cleanup_task) :
1853 1853 _g1(g1),
1854 1854 _max_live_bytes(0), _regions_claimed(0),
1855 1855 _freed_bytes(0),
1856 1856 _claimed_region_time(0.0), _max_region_time(0.0),
1857 1857 _local_cleanup_list(local_cleanup_list),
1858 1858 _old_regions_removed(),
1859 1859 _humongous_regions_removed(),
1860 1860 _hrrs_cleanup_task(hrrs_cleanup_task) { }
1861 1861
1862 1862 size_t freed_bytes() { return _freed_bytes; }
1863 1863 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1864 1864 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1865 1865
1866 1866 bool doHeapRegion(HeapRegion *hr) {
1867 1867 if (hr->continuesHumongous()) {
1868 1868 return false;
1869 1869 }
1870 1870 // We use a claim value of zero here because all regions
1871 1871 // were claimed with value 1 in the FinalCount task.
1872 1872 _g1->reset_gc_time_stamps(hr);
1873 1873 double start = os::elapsedTime();
1874 1874 _regions_claimed++;
1875 1875 hr->note_end_of_marking();
1876 1876 _max_live_bytes += hr->max_live_bytes();
1877 1877
1878 1878 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1879 1879 _freed_bytes += hr->used();
1880 1880 hr->set_containing_set(NULL);
1881 1881 if (hr->isHumongous()) {
1882 1882 assert(hr->startsHumongous(), "we should only see starts humongous");
1883 1883 _humongous_regions_removed.increment(1u, hr->capacity());
1884 1884 _g1->free_humongous_region(hr, _local_cleanup_list, true);
1885 1885 } else {
1886 1886 _old_regions_removed.increment(1u, hr->capacity());
1887 1887 _g1->free_region(hr, _local_cleanup_list, true);
1888 1888 }
1889 1889 } else {
1890 1890 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1891 1891 }
1892 1892
1893 1893 double region_time = (os::elapsedTime() - start);
1894 1894 _claimed_region_time += region_time;
1895 1895 if (region_time > _max_region_time) {
1896 1896 _max_region_time = region_time;
1897 1897 }
1898 1898 return false;
1899 1899 }
1900 1900
1901 1901 size_t max_live_bytes() { return _max_live_bytes; }
1902 1902 uint regions_claimed() { return _regions_claimed; }
1903 1903 double claimed_region_time_sec() { return _claimed_region_time; }
1904 1904 double max_region_time_sec() { return _max_region_time; }
1905 1905 };
1906 1906
1907 1907 class G1ParNoteEndTask: public AbstractGangTask {
1908 1908 friend class G1NoteEndOfConcMarkClosure;
1909 1909
1910 1910 protected:
1911 1911 G1CollectedHeap* _g1h;
1912 1912 size_t _max_live_bytes;
1913 1913 size_t _freed_bytes;
1914 1914 FreeRegionList* _cleanup_list;
1915 1915
1916 1916 public:
1917 1917 G1ParNoteEndTask(G1CollectedHeap* g1h,
1918 1918 FreeRegionList* cleanup_list) :
1919 1919 AbstractGangTask("G1 note end"), _g1h(g1h),
1920 1920 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1921 1921
1922 1922 void work(uint worker_id) {
1923 1923 double start = os::elapsedTime();
1924 1924 FreeRegionList local_cleanup_list("Local Cleanup List");
1925 1925 HRRSCleanupTask hrrs_cleanup_task;
1926 1926 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1927 1927 &hrrs_cleanup_task);
1928 1928 if (G1CollectedHeap::use_parallel_gc_threads()) {
1929 1929 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1930 1930 _g1h->workers()->active_workers(),
1931 1931 HeapRegion::NoteEndClaimValue);
1932 1932 } else {
1933 1933 _g1h->heap_region_iterate(&g1_note_end);
1934 1934 }
1935 1935 assert(g1_note_end.complete(), "Shouldn't have yielded!");
1936 1936
1937 1937 // Now update the lists
1938 1938 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1939 1939 {
1940 1940 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1941 1941 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1942 1942 _max_live_bytes += g1_note_end.max_live_bytes();
1943 1943 _freed_bytes += g1_note_end.freed_bytes();
1944 1944
1945 1945 // If we iterate over the global cleanup list at the end of
1946 1946 // cleanup to do this printing we will not guarantee to only
1947 1947 // generate output for the newly-reclaimed regions (the list
1948 1948 // might not be empty at the beginning of cleanup; we might
1949 1949 // still be working on its previous contents). So we do the
1950 1950 // printing here, before we append the new regions to the global
1951 1951 // cleanup list.
1952 1952
1953 1953 G1HRPrinter* hr_printer = _g1h->hr_printer();
1954 1954 if (hr_printer->is_active()) {
1955 1955 FreeRegionListIterator iter(&local_cleanup_list);
1956 1956 while (iter.more_available()) {
1957 1957 HeapRegion* hr = iter.get_next();
1958 1958 hr_printer->cleanup(hr);
1959 1959 }
1960 1960 }
1961 1961
1962 1962 _cleanup_list->add_ordered(&local_cleanup_list);
1963 1963 assert(local_cleanup_list.is_empty(), "post-condition");
1964 1964
1965 1965 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1966 1966 }
1967 1967 }
1968 1968 size_t max_live_bytes() { return _max_live_bytes; }
1969 1969 size_t freed_bytes() { return _freed_bytes; }
1970 1970 };
1971 1971
1972 1972 class G1ParScrubRemSetTask: public AbstractGangTask {
1973 1973 protected:
1974 1974 G1RemSet* _g1rs;
1975 1975 BitMap* _region_bm;
1976 1976 BitMap* _card_bm;
1977 1977 public:
1978 1978 G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1979 1979 BitMap* region_bm, BitMap* card_bm) :
1980 1980 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1981 1981 _region_bm(region_bm), _card_bm(card_bm) { }
1982 1982
1983 1983 void work(uint worker_id) {
1984 1984 if (G1CollectedHeap::use_parallel_gc_threads()) {
1985 1985 _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1986 1986 HeapRegion::ScrubRemSetClaimValue);
1987 1987 } else {
1988 1988 _g1rs->scrub(_region_bm, _card_bm);
1989 1989 }
1990 1990 }
1991 1991
1992 1992 };
1993 1993
1994 1994 void ConcurrentMark::cleanup() {
1995 1995 // world is stopped at this checkpoint
1996 1996 assert(SafepointSynchronize::is_at_safepoint(),
1997 1997 "world should be stopped");
1998 1998 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1999 1999
2000 2000 // If a full collection has happened, we shouldn't do this.
2001 2001 if (has_aborted()) {
2002 2002 g1h->set_marking_complete(); // So bitmap clearing isn't confused
2003 2003 return;
2004 2004 }
2005 2005
2006 2006 g1h->verify_region_sets_optional();
2007 2007
2008 2008 if (VerifyDuringGC) {
2009 2009 HandleMark hm; // handle scope
2010 2010 Universe::heap()->prepare_for_verify();
2011 2011 Universe::verify(VerifyOption_G1UsePrevMarking,
2012 2012 " VerifyDuringGC:(before)");
2013 2013 }
2014 2014 g1h->check_bitmaps("Cleanup Start");
2015 2015
2016 2016 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2017 2017 g1p->record_concurrent_mark_cleanup_start();
2018 2018
2019 2019 double start = os::elapsedTime();
2020 2020
2021 2021 HeapRegionRemSet::reset_for_cleanup_tasks();
2022 2022
2023 2023 uint n_workers;
2024 2024
2025 2025 // Do counting once more with the world stopped for good measure.
2026 2026 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2027 2027
2028 2028 if (G1CollectedHeap::use_parallel_gc_threads()) {
2029 2029 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2030 2030 "sanity check");
2031 2031
2032 2032 g1h->set_par_threads();
2033 2033 n_workers = g1h->n_par_threads();
2034 2034 assert(g1h->n_par_threads() == n_workers,
2035 2035 "Should not have been reset");
2036 2036 g1h->workers()->run_task(&g1_par_count_task);
2037 2037 // Done with the parallel phase so reset to 0.
2038 2038 g1h->set_par_threads(0);
2039 2039
2040 2040 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2041 2041 "sanity check");
2042 2042 } else {
2043 2043 n_workers = 1;
2044 2044 g1_par_count_task.work(0);
2045 2045 }
2046 2046
2047 2047 if (VerifyDuringGC) {
2048 2048 // Verify that the counting data accumulated during marking matches
2049 2049 // that calculated by walking the marking bitmap.
2050 2050
2051 2051 // Bitmaps to hold expected values
2052 2052 BitMap expected_region_bm(_region_bm.size(), true);
2053 2053 BitMap expected_card_bm(_card_bm.size(), true);
2054 2054
2055 2055 G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2056 2056 &_region_bm,
2057 2057 &_card_bm,
2058 2058 &expected_region_bm,
2059 2059 &expected_card_bm);
2060 2060
2061 2061 if (G1CollectedHeap::use_parallel_gc_threads()) {
2062 2062 g1h->set_par_threads((int)n_workers);
2063 2063 g1h->workers()->run_task(&g1_par_verify_task);
2064 2064 // Done with the parallel phase so reset to 0.
2065 2065 g1h->set_par_threads(0);
2066 2066
2067 2067 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2068 2068 "sanity check");
2069 2069 } else {
2070 2070 g1_par_verify_task.work(0);
2071 2071 }
2072 2072
2073 2073 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2074 2074 }
2075 2075
2076 2076 size_t start_used_bytes = g1h->used();
2077 2077 g1h->set_marking_complete();
2078 2078
2079 2079 double count_end = os::elapsedTime();
2080 2080 double this_final_counting_time = (count_end - start);
2081 2081 _total_counting_time += this_final_counting_time;
2082 2082
2083 2083 if (G1PrintRegionLivenessInfo) {
2084 2084 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2085 2085 _g1h->heap_region_iterate(&cl);
2086 2086 }
2087 2087
2088 2088 // Install newly created mark bitMap as "prev".
2089 2089 swapMarkBitMaps();
2090 2090
2091 2091 g1h->reset_gc_time_stamp();
2092 2092
2093 2093 // Note end of marking in all heap regions.
2094 2094 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2095 2095 if (G1CollectedHeap::use_parallel_gc_threads()) {
2096 2096 g1h->set_par_threads((int)n_workers);
2097 2097 g1h->workers()->run_task(&g1_par_note_end_task);
2098 2098 g1h->set_par_threads(0);
2099 2099
2100 2100 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2101 2101 "sanity check");
2102 2102 } else {
2103 2103 g1_par_note_end_task.work(0);
2104 2104 }
2105 2105 g1h->check_gc_time_stamps();
2106 2106
2107 2107 if (!cleanup_list_is_empty()) {
2108 2108 // The cleanup list is not empty, so we'll have to process it
2109 2109 // concurrently. Notify anyone else that might be wanting free
2110 2110 // regions that there will be more free regions coming soon.
2111 2111 g1h->set_free_regions_coming();
2112 2112 }
2113 2113
2114 2114 // call below, since it affects the metric by which we sort the heap
2115 2115 // regions.
2116 2116 if (G1ScrubRemSets) {
2117 2117 double rs_scrub_start = os::elapsedTime();
2118 2118 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2119 2119 if (G1CollectedHeap::use_parallel_gc_threads()) {
2120 2120 g1h->set_par_threads((int)n_workers);
2121 2121 g1h->workers()->run_task(&g1_par_scrub_rs_task);
2122 2122 g1h->set_par_threads(0);
2123 2123
2124 2124 assert(g1h->check_heap_region_claim_values(
2125 2125 HeapRegion::ScrubRemSetClaimValue),
2126 2126 "sanity check");
2127 2127 } else {
2128 2128 g1_par_scrub_rs_task.work(0);
2129 2129 }
2130 2130
2131 2131 double rs_scrub_end = os::elapsedTime();
2132 2132 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2133 2133 _total_rs_scrub_time += this_rs_scrub_time;
2134 2134 }
2135 2135
2136 2136 // this will also free any regions totally full of garbage objects,
2137 2137 // and sort the regions.
2138 2138 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2139 2139
2140 2140 // Statistics.
2141 2141 double end = os::elapsedTime();
2142 2142 _cleanup_times.add((end - start) * 1000.0);
2143 2143
2144 2144 if (G1Log::fine()) {
2145 2145 g1h->print_size_transition(gclog_or_tty,
2146 2146 start_used_bytes,
2147 2147 g1h->used(),
2148 2148 g1h->capacity());
2149 2149 }
2150 2150
2151 2151 // Clean up will have freed any regions completely full of garbage.
2152 2152 // Update the soft reference policy with the new heap occupancy.
2153 2153 Universe::update_heap_info_at_gc();
2154 2154
2155 2155 if (VerifyDuringGC) {
2156 2156 HandleMark hm; // handle scope
2157 2157 Universe::heap()->prepare_for_verify();
2158 2158 Universe::verify(VerifyOption_G1UsePrevMarking,
2159 2159 " VerifyDuringGC:(after)");
2160 2160 }
2161 2161 g1h->check_bitmaps("Cleanup End");
2162 2162
2163 2163 g1h->verify_region_sets_optional();
2164 2164
2165 2165 // We need to make this be a "collection" so any collection pause that
2166 2166 // races with it goes around and waits for completeCleanup to finish.
2167 2167 g1h->increment_total_collections();
↓ open down ↓ |
2167 lines elided |
↑ open up ↑ |
2168 2168
2169 2169 // Clean out dead classes and update Metaspace sizes.
2170 2170 if (ClassUnloadingWithConcurrentMark) {
2171 2171 ClassLoaderDataGraph::purge();
2172 2172 }
2173 2173 MetaspaceGC::compute_new_size();
2174 2174
2175 2175 // We reclaimed old regions so we should calculate the sizes to make
2176 2176 // sure we update the old gen/space data.
2177 2177 g1h->g1mm()->update_sizes();
2178 + g1h->allocation_context_stats().update_after_mark();
2178 2179
2179 2180 g1h->trace_heap_after_concurrent_cycle();
2180 2181 }
2181 2182
2182 2183 void ConcurrentMark::completeCleanup() {
2183 2184 if (has_aborted()) return;
2184 2185
2185 2186 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2186 2187
2187 2188 _cleanup_list.verify_optional();
2188 2189 FreeRegionList tmp_free_list("Tmp Free List");
2189 2190
2190 2191 if (G1ConcRegionFreeingVerbose) {
2191 2192 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2192 2193 "cleanup list has %u entries",
2193 2194 _cleanup_list.length());
2194 2195 }
2195 2196
2196 2197 // No one else should be accessing the _cleanup_list at this point,
2197 2198 // so it is not necessary to take any locks
2198 2199 while (!_cleanup_list.is_empty()) {
2199 2200 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2200 2201 assert(hr != NULL, "Got NULL from a non-empty list");
2201 2202 hr->par_clear();
2202 2203 tmp_free_list.add_ordered(hr);
2203 2204
2204 2205 // Instead of adding one region at a time to the secondary_free_list,
2205 2206 // we accumulate them in the local list and move them a few at a
2206 2207 // time. This also cuts down on the number of notify_all() calls
2207 2208 // we do during this process. We'll also append the local list when
2208 2209 // _cleanup_list is empty (which means we just removed the last
2209 2210 // region from the _cleanup_list).
2210 2211 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2211 2212 _cleanup_list.is_empty()) {
2212 2213 if (G1ConcRegionFreeingVerbose) {
2213 2214 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2214 2215 "appending %u entries to the secondary_free_list, "
2215 2216 "cleanup list still has %u entries",
2216 2217 tmp_free_list.length(),
2217 2218 _cleanup_list.length());
2218 2219 }
2219 2220
2220 2221 {
2221 2222 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2222 2223 g1h->secondary_free_list_add(&tmp_free_list);
2223 2224 SecondaryFreeList_lock->notify_all();
2224 2225 }
2225 2226
2226 2227 if (G1StressConcRegionFreeing) {
2227 2228 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2228 2229 os::sleep(Thread::current(), (jlong) 1, false);
2229 2230 }
2230 2231 }
2231 2232 }
2232 2233 }
2233 2234 assert(tmp_free_list.is_empty(), "post-condition");
2234 2235 }
2235 2236
2236 2237 // Supporting Object and Oop closures for reference discovery
2237 2238 // and processing in during marking
2238 2239
2239 2240 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2240 2241 HeapWord* addr = (HeapWord*)obj;
2241 2242 return addr != NULL &&
2242 2243 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2243 2244 }
2244 2245
2245 2246 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2246 2247 // Uses the CMTask associated with a worker thread (for serial reference
2247 2248 // processing the CMTask for worker 0 is used) to preserve (mark) and
2248 2249 // trace referent objects.
2249 2250 //
2250 2251 // Using the CMTask and embedded local queues avoids having the worker
2251 2252 // threads operating on the global mark stack. This reduces the risk
2252 2253 // of overflowing the stack - which we would rather avoid at this late
2253 2254 // state. Also using the tasks' local queues removes the potential
2254 2255 // of the workers interfering with each other that could occur if
2255 2256 // operating on the global stack.
2256 2257
2257 2258 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2258 2259 ConcurrentMark* _cm;
2259 2260 CMTask* _task;
2260 2261 int _ref_counter_limit;
2261 2262 int _ref_counter;
2262 2263 bool _is_serial;
2263 2264 public:
2264 2265 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2265 2266 _cm(cm), _task(task), _is_serial(is_serial),
2266 2267 _ref_counter_limit(G1RefProcDrainInterval) {
2267 2268 assert(_ref_counter_limit > 0, "sanity");
2268 2269 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2269 2270 _ref_counter = _ref_counter_limit;
2270 2271 }
2271 2272
2272 2273 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2273 2274 virtual void do_oop( oop* p) { do_oop_work(p); }
2274 2275
2275 2276 template <class T> void do_oop_work(T* p) {
2276 2277 if (!_cm->has_overflown()) {
2277 2278 oop obj = oopDesc::load_decode_heap_oop(p);
2278 2279 if (_cm->verbose_high()) {
2279 2280 gclog_or_tty->print_cr("\t[%u] we're looking at location "
2280 2281 "*"PTR_FORMAT" = "PTR_FORMAT,
2281 2282 _task->worker_id(), p2i(p), p2i((void*) obj));
2282 2283 }
2283 2284
2284 2285 _task->deal_with_reference(obj);
2285 2286 _ref_counter--;
2286 2287
2287 2288 if (_ref_counter == 0) {
2288 2289 // We have dealt with _ref_counter_limit references, pushing them
2289 2290 // and objects reachable from them on to the local stack (and
2290 2291 // possibly the global stack). Call CMTask::do_marking_step() to
2291 2292 // process these entries.
2292 2293 //
2293 2294 // We call CMTask::do_marking_step() in a loop, which we'll exit if
2294 2295 // there's nothing more to do (i.e. we're done with the entries that
2295 2296 // were pushed as a result of the CMTask::deal_with_reference() calls
2296 2297 // above) or we overflow.
2297 2298 //
2298 2299 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2299 2300 // flag while there may still be some work to do. (See the comment at
2300 2301 // the beginning of CMTask::do_marking_step() for those conditions -
2301 2302 // one of which is reaching the specified time target.) It is only
2302 2303 // when CMTask::do_marking_step() returns without setting the
2303 2304 // has_aborted() flag that the marking step has completed.
2304 2305 do {
2305 2306 double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2306 2307 _task->do_marking_step(mark_step_duration_ms,
2307 2308 false /* do_termination */,
2308 2309 _is_serial);
2309 2310 } while (_task->has_aborted() && !_cm->has_overflown());
2310 2311 _ref_counter = _ref_counter_limit;
2311 2312 }
2312 2313 } else {
2313 2314 if (_cm->verbose_high()) {
2314 2315 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2315 2316 }
2316 2317 }
2317 2318 }
2318 2319 };
2319 2320
2320 2321 // 'Drain' oop closure used by both serial and parallel reference processing.
2321 2322 // Uses the CMTask associated with a given worker thread (for serial
2322 2323 // reference processing the CMtask for worker 0 is used). Calls the
2323 2324 // do_marking_step routine, with an unbelievably large timeout value,
2324 2325 // to drain the marking data structures of the remaining entries
2325 2326 // added by the 'keep alive' oop closure above.
2326 2327
2327 2328 class G1CMDrainMarkingStackClosure: public VoidClosure {
2328 2329 ConcurrentMark* _cm;
2329 2330 CMTask* _task;
2330 2331 bool _is_serial;
2331 2332 public:
2332 2333 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2333 2334 _cm(cm), _task(task), _is_serial(is_serial) {
2334 2335 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2335 2336 }
2336 2337
2337 2338 void do_void() {
2338 2339 do {
2339 2340 if (_cm->verbose_high()) {
2340 2341 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2341 2342 _task->worker_id(), BOOL_TO_STR(_is_serial));
2342 2343 }
2343 2344
2344 2345 // We call CMTask::do_marking_step() to completely drain the local
2345 2346 // and global marking stacks of entries pushed by the 'keep alive'
2346 2347 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2347 2348 //
2348 2349 // CMTask::do_marking_step() is called in a loop, which we'll exit
2349 2350 // if there's nothing more to do (i.e. we'completely drained the
2350 2351 // entries that were pushed as a a result of applying the 'keep alive'
2351 2352 // closure to the entries on the discovered ref lists) or we overflow
2352 2353 // the global marking stack.
2353 2354 //
2354 2355 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2355 2356 // flag while there may still be some work to do. (See the comment at
2356 2357 // the beginning of CMTask::do_marking_step() for those conditions -
2357 2358 // one of which is reaching the specified time target.) It is only
2358 2359 // when CMTask::do_marking_step() returns without setting the
2359 2360 // has_aborted() flag that the marking step has completed.
2360 2361
2361 2362 _task->do_marking_step(1000000000.0 /* something very large */,
2362 2363 true /* do_termination */,
2363 2364 _is_serial);
2364 2365 } while (_task->has_aborted() && !_cm->has_overflown());
2365 2366 }
2366 2367 };
2367 2368
2368 2369 // Implementation of AbstractRefProcTaskExecutor for parallel
2369 2370 // reference processing at the end of G1 concurrent marking
2370 2371
2371 2372 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2372 2373 private:
2373 2374 G1CollectedHeap* _g1h;
2374 2375 ConcurrentMark* _cm;
2375 2376 WorkGang* _workers;
2376 2377 int _active_workers;
2377 2378
2378 2379 public:
2379 2380 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2380 2381 ConcurrentMark* cm,
2381 2382 WorkGang* workers,
2382 2383 int n_workers) :
2383 2384 _g1h(g1h), _cm(cm),
2384 2385 _workers(workers), _active_workers(n_workers) { }
2385 2386
2386 2387 // Executes the given task using concurrent marking worker threads.
2387 2388 virtual void execute(ProcessTask& task);
2388 2389 virtual void execute(EnqueueTask& task);
2389 2390 };
2390 2391
2391 2392 class G1CMRefProcTaskProxy: public AbstractGangTask {
2392 2393 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2393 2394 ProcessTask& _proc_task;
2394 2395 G1CollectedHeap* _g1h;
2395 2396 ConcurrentMark* _cm;
2396 2397
2397 2398 public:
2398 2399 G1CMRefProcTaskProxy(ProcessTask& proc_task,
2399 2400 G1CollectedHeap* g1h,
2400 2401 ConcurrentMark* cm) :
2401 2402 AbstractGangTask("Process reference objects in parallel"),
2402 2403 _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2403 2404 ReferenceProcessor* rp = _g1h->ref_processor_cm();
2404 2405 assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2405 2406 }
2406 2407
2407 2408 virtual void work(uint worker_id) {
2408 2409 ResourceMark rm;
2409 2410 HandleMark hm;
2410 2411 CMTask* task = _cm->task(worker_id);
2411 2412 G1CMIsAliveClosure g1_is_alive(_g1h);
2412 2413 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2413 2414 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2414 2415
2415 2416 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2416 2417 }
2417 2418 };
2418 2419
2419 2420 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2420 2421 assert(_workers != NULL, "Need parallel worker threads.");
2421 2422 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2422 2423
2423 2424 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2424 2425
2425 2426 // We need to reset the concurrency level before each
2426 2427 // proxy task execution, so that the termination protocol
2427 2428 // and overflow handling in CMTask::do_marking_step() knows
2428 2429 // how many workers to wait for.
2429 2430 _cm->set_concurrency(_active_workers);
2430 2431 _g1h->set_par_threads(_active_workers);
2431 2432 _workers->run_task(&proc_task_proxy);
2432 2433 _g1h->set_par_threads(0);
2433 2434 }
2434 2435
2435 2436 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2436 2437 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2437 2438 EnqueueTask& _enq_task;
2438 2439
2439 2440 public:
2440 2441 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2441 2442 AbstractGangTask("Enqueue reference objects in parallel"),
2442 2443 _enq_task(enq_task) { }
2443 2444
2444 2445 virtual void work(uint worker_id) {
2445 2446 _enq_task.work(worker_id);
2446 2447 }
2447 2448 };
2448 2449
2449 2450 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2450 2451 assert(_workers != NULL, "Need parallel worker threads.");
2451 2452 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2452 2453
2453 2454 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2454 2455
2455 2456 // Not strictly necessary but...
2456 2457 //
2457 2458 // We need to reset the concurrency level before each
2458 2459 // proxy task execution, so that the termination protocol
2459 2460 // and overflow handling in CMTask::do_marking_step() knows
2460 2461 // how many workers to wait for.
2461 2462 _cm->set_concurrency(_active_workers);
2462 2463 _g1h->set_par_threads(_active_workers);
2463 2464 _workers->run_task(&enq_task_proxy);
2464 2465 _g1h->set_par_threads(0);
2465 2466 }
2466 2467
2467 2468 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2468 2469 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2469 2470 }
2470 2471
2471 2472 // Helper class to get rid of some boilerplate code.
2472 2473 class G1RemarkGCTraceTime : public GCTraceTime {
2473 2474 static bool doit_and_prepend(bool doit) {
2474 2475 if (doit) {
2475 2476 gclog_or_tty->put(' ');
2476 2477 }
2477 2478 return doit;
2478 2479 }
2479 2480
2480 2481 public:
2481 2482 G1RemarkGCTraceTime(const char* title, bool doit)
2482 2483 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2483 2484 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2484 2485 }
2485 2486 };
2486 2487
2487 2488 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2488 2489 if (has_overflown()) {
2489 2490 // Skip processing the discovered references if we have
2490 2491 // overflown the global marking stack. Reference objects
2491 2492 // only get discovered once so it is OK to not
2492 2493 // de-populate the discovered reference lists. We could have,
2493 2494 // but the only benefit would be that, when marking restarts,
2494 2495 // less reference objects are discovered.
2495 2496 return;
2496 2497 }
2497 2498
2498 2499 ResourceMark rm;
2499 2500 HandleMark hm;
2500 2501
2501 2502 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2502 2503
2503 2504 // Is alive closure.
2504 2505 G1CMIsAliveClosure g1_is_alive(g1h);
2505 2506
2506 2507 // Inner scope to exclude the cleaning of the string and symbol
2507 2508 // tables from the displayed time.
2508 2509 {
2509 2510 if (G1Log::finer()) {
2510 2511 gclog_or_tty->put(' ');
2511 2512 }
2512 2513 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2513 2514
2514 2515 ReferenceProcessor* rp = g1h->ref_processor_cm();
2515 2516
2516 2517 // See the comment in G1CollectedHeap::ref_processing_init()
2517 2518 // about how reference processing currently works in G1.
2518 2519
2519 2520 // Set the soft reference policy
2520 2521 rp->setup_policy(clear_all_soft_refs);
2521 2522 assert(_markStack.isEmpty(), "mark stack should be empty");
2522 2523
2523 2524 // Instances of the 'Keep Alive' and 'Complete GC' closures used
2524 2525 // in serial reference processing. Note these closures are also
2525 2526 // used for serially processing (by the the current thread) the
2526 2527 // JNI references during parallel reference processing.
2527 2528 //
2528 2529 // These closures do not need to synchronize with the worker
2529 2530 // threads involved in parallel reference processing as these
2530 2531 // instances are executed serially by the current thread (e.g.
2531 2532 // reference processing is not multi-threaded and is thus
2532 2533 // performed by the current thread instead of a gang worker).
2533 2534 //
2534 2535 // The gang tasks involved in parallel reference procssing create
2535 2536 // their own instances of these closures, which do their own
2536 2537 // synchronization among themselves.
2537 2538 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2538 2539 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2539 2540
2540 2541 // We need at least one active thread. If reference processing
2541 2542 // is not multi-threaded we use the current (VMThread) thread,
2542 2543 // otherwise we use the work gang from the G1CollectedHeap and
2543 2544 // we utilize all the worker threads we can.
2544 2545 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2545 2546 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2546 2547 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2547 2548
2548 2549 // Parallel processing task executor.
2549 2550 G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2550 2551 g1h->workers(), active_workers);
2551 2552 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2552 2553
2553 2554 // Set the concurrency level. The phase was already set prior to
2554 2555 // executing the remark task.
2555 2556 set_concurrency(active_workers);
2556 2557
2557 2558 // Set the degree of MT processing here. If the discovery was done MT,
2558 2559 // the number of threads involved during discovery could differ from
2559 2560 // the number of active workers. This is OK as long as the discovered
2560 2561 // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2561 2562 rp->set_active_mt_degree(active_workers);
2562 2563
2563 2564 // Process the weak references.
2564 2565 const ReferenceProcessorStats& stats =
2565 2566 rp->process_discovered_references(&g1_is_alive,
2566 2567 &g1_keep_alive,
2567 2568 &g1_drain_mark_stack,
2568 2569 executor,
2569 2570 g1h->gc_timer_cm(),
2570 2571 concurrent_gc_id());
2571 2572 g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2572 2573
2573 2574 // The do_oop work routines of the keep_alive and drain_marking_stack
2574 2575 // oop closures will set the has_overflown flag if we overflow the
2575 2576 // global marking stack.
2576 2577
2577 2578 assert(_markStack.overflow() || _markStack.isEmpty(),
2578 2579 "mark stack should be empty (unless it overflowed)");
2579 2580
2580 2581 if (_markStack.overflow()) {
2581 2582 // This should have been done already when we tried to push an
2582 2583 // entry on to the global mark stack. But let's do it again.
2583 2584 set_has_overflown();
2584 2585 }
2585 2586
2586 2587 assert(rp->num_q() == active_workers, "why not");
2587 2588
2588 2589 rp->enqueue_discovered_references(executor);
2589 2590
2590 2591 rp->verify_no_references_recorded();
2591 2592 assert(!rp->discovery_enabled(), "Post condition");
2592 2593 }
2593 2594
2594 2595 if (has_overflown()) {
2595 2596 // We can not trust g1_is_alive if the marking stack overflowed
2596 2597 return;
2597 2598 }
2598 2599
2599 2600 assert(_markStack.isEmpty(), "Marking should have completed");
2600 2601
2601 2602 // Unload Klasses, String, Symbols, Code Cache, etc.
2602 2603 {
2603 2604 G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2604 2605
2605 2606 if (ClassUnloadingWithConcurrentMark) {
2606 2607 // Cleaning of klasses depends on correct information from MetadataMarkOnStack. The CodeCache::mark_on_stack
2607 2608 // part is too slow to be done serially, so it is handled during the weakRefsWorkParallelPart phase.
2608 2609 // Defer the cleaning until we have complete on_stack data.
2609 2610 MetadataOnStackMark md_on_stack(false /* Don't visit the code cache at this point */);
2610 2611
2611 2612 bool purged_classes;
2612 2613
2613 2614 {
2614 2615 G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2615 2616 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2616 2617 }
2617 2618
2618 2619 {
2619 2620 G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2620 2621 weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2621 2622 }
2622 2623
2623 2624 {
2624 2625 G1RemarkGCTraceTime trace("Deallocate Metadata", G1Log::finest());
2625 2626 ClassLoaderDataGraph::free_deallocate_lists();
2626 2627 }
2627 2628 }
2628 2629
2629 2630 if (G1StringDedup::is_enabled()) {
2630 2631 G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2631 2632 G1StringDedup::unlink(&g1_is_alive);
2632 2633 }
2633 2634 }
2634 2635 }
2635 2636
2636 2637 void ConcurrentMark::swapMarkBitMaps() {
2637 2638 CMBitMapRO* temp = _prevMarkBitMap;
2638 2639 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap;
2639 2640 _nextMarkBitMap = (CMBitMap*) temp;
2640 2641 }
2641 2642
2642 2643 class CMObjectClosure;
2643 2644
2644 2645 // Closure for iterating over objects, currently only used for
2645 2646 // processing SATB buffers.
2646 2647 class CMObjectClosure : public ObjectClosure {
2647 2648 private:
2648 2649 CMTask* _task;
2649 2650
2650 2651 public:
2651 2652 void do_object(oop obj) {
2652 2653 _task->deal_with_reference(obj);
2653 2654 }
2654 2655
2655 2656 CMObjectClosure(CMTask* task) : _task(task) { }
2656 2657 };
2657 2658
2658 2659 class G1RemarkThreadsClosure : public ThreadClosure {
2659 2660 CMObjectClosure _cm_obj;
2660 2661 G1CMOopClosure _cm_cl;
2661 2662 MarkingCodeBlobClosure _code_cl;
2662 2663 int _thread_parity;
2663 2664 bool _is_par;
2664 2665
2665 2666 public:
2666 2667 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2667 2668 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2668 2669 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2669 2670
2670 2671 void do_thread(Thread* thread) {
2671 2672 if (thread->is_Java_thread()) {
2672 2673 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2673 2674 JavaThread* jt = (JavaThread*)thread;
2674 2675
2675 2676 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2676 2677 // however the liveness of oops reachable from nmethods have very complex lifecycles:
2677 2678 // * Alive if on the stack of an executing method
2678 2679 // * Weakly reachable otherwise
2679 2680 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2680 2681 // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2681 2682 jt->nmethods_do(&_code_cl);
2682 2683
2683 2684 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2684 2685 }
2685 2686 } else if (thread->is_VM_thread()) {
2686 2687 if (thread->claim_oops_do(_is_par, _thread_parity)) {
2687 2688 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2688 2689 }
2689 2690 }
2690 2691 }
2691 2692 };
2692 2693
2693 2694 class CMRemarkTask: public AbstractGangTask {
2694 2695 private:
2695 2696 ConcurrentMark* _cm;
2696 2697 bool _is_serial;
2697 2698 public:
2698 2699 void work(uint worker_id) {
2699 2700 // Since all available tasks are actually started, we should
2700 2701 // only proceed if we're supposed to be actived.
2701 2702 if (worker_id < _cm->active_tasks()) {
2702 2703 CMTask* task = _cm->task(worker_id);
2703 2704 task->record_start_time();
2704 2705 {
2705 2706 ResourceMark rm;
2706 2707 HandleMark hm;
2707 2708
2708 2709 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2709 2710 Threads::threads_do(&threads_f);
2710 2711 }
2711 2712
2712 2713 do {
2713 2714 task->do_marking_step(1000000000.0 /* something very large */,
2714 2715 true /* do_termination */,
2715 2716 _is_serial);
2716 2717 } while (task->has_aborted() && !_cm->has_overflown());
2717 2718 // If we overflow, then we do not want to restart. We instead
2718 2719 // want to abort remark and do concurrent marking again.
2719 2720 task->record_end_time();
2720 2721 }
2721 2722 }
2722 2723
2723 2724 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2724 2725 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2725 2726 _cm->terminator()->reset_for_reuse(active_workers);
2726 2727 }
2727 2728 };
2728 2729
2729 2730 void ConcurrentMark::checkpointRootsFinalWork() {
2730 2731 ResourceMark rm;
2731 2732 HandleMark hm;
2732 2733 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2733 2734
2734 2735 G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2735 2736
2736 2737 g1h->ensure_parsability(false);
2737 2738
2738 2739 if (G1CollectedHeap::use_parallel_gc_threads()) {
2739 2740 G1CollectedHeap::StrongRootsScope srs(g1h);
2740 2741 // this is remark, so we'll use up all active threads
2741 2742 uint active_workers = g1h->workers()->active_workers();
2742 2743 if (active_workers == 0) {
2743 2744 assert(active_workers > 0, "Should have been set earlier");
2744 2745 active_workers = (uint) ParallelGCThreads;
2745 2746 g1h->workers()->set_active_workers(active_workers);
2746 2747 }
2747 2748 set_concurrency_and_phase(active_workers, false /* concurrent */);
2748 2749 // Leave _parallel_marking_threads at it's
2749 2750 // value originally calculated in the ConcurrentMark
2750 2751 // constructor and pass values of the active workers
2751 2752 // through the gang in the task.
2752 2753
2753 2754 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2754 2755 // We will start all available threads, even if we decide that the
2755 2756 // active_workers will be fewer. The extra ones will just bail out
2756 2757 // immediately.
2757 2758 g1h->set_par_threads(active_workers);
2758 2759 g1h->workers()->run_task(&remarkTask);
2759 2760 g1h->set_par_threads(0);
2760 2761 } else {
2761 2762 G1CollectedHeap::StrongRootsScope srs(g1h);
2762 2763 uint active_workers = 1;
2763 2764 set_concurrency_and_phase(active_workers, false /* concurrent */);
2764 2765
2765 2766 // Note - if there's no work gang then the VMThread will be
2766 2767 // the thread to execute the remark - serially. We have
2767 2768 // to pass true for the is_serial parameter so that
2768 2769 // CMTask::do_marking_step() doesn't enter the sync
2769 2770 // barriers in the event of an overflow. Doing so will
2770 2771 // cause an assert that the current thread is not a
2771 2772 // concurrent GC thread.
2772 2773 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2773 2774 remarkTask.work(0);
2774 2775 }
2775 2776 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2776 2777 guarantee(has_overflown() ||
2777 2778 satb_mq_set.completed_buffers_num() == 0,
2778 2779 err_msg("Invariant: has_overflown = %s, num buffers = %d",
2779 2780 BOOL_TO_STR(has_overflown()),
2780 2781 satb_mq_set.completed_buffers_num()));
2781 2782
2782 2783 print_stats();
2783 2784 }
2784 2785
2785 2786 #ifndef PRODUCT
2786 2787
2787 2788 class PrintReachableOopClosure: public OopClosure {
2788 2789 private:
2789 2790 G1CollectedHeap* _g1h;
2790 2791 outputStream* _out;
2791 2792 VerifyOption _vo;
2792 2793 bool _all;
2793 2794
2794 2795 public:
2795 2796 PrintReachableOopClosure(outputStream* out,
2796 2797 VerifyOption vo,
2797 2798 bool all) :
2798 2799 _g1h(G1CollectedHeap::heap()),
2799 2800 _out(out), _vo(vo), _all(all) { }
2800 2801
2801 2802 void do_oop(narrowOop* p) { do_oop_work(p); }
2802 2803 void do_oop( oop* p) { do_oop_work(p); }
2803 2804
2804 2805 template <class T> void do_oop_work(T* p) {
2805 2806 oop obj = oopDesc::load_decode_heap_oop(p);
2806 2807 const char* str = NULL;
2807 2808 const char* str2 = "";
2808 2809
2809 2810 if (obj == NULL) {
2810 2811 str = "";
2811 2812 } else if (!_g1h->is_in_g1_reserved(obj)) {
2812 2813 str = " O";
2813 2814 } else {
2814 2815 HeapRegion* hr = _g1h->heap_region_containing(obj);
2815 2816 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2816 2817 bool marked = _g1h->is_marked(obj, _vo);
2817 2818
2818 2819 if (over_tams) {
2819 2820 str = " >";
2820 2821 if (marked) {
2821 2822 str2 = " AND MARKED";
2822 2823 }
2823 2824 } else if (marked) {
2824 2825 str = " M";
2825 2826 } else {
2826 2827 str = " NOT";
2827 2828 }
2828 2829 }
2829 2830
2830 2831 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s",
2831 2832 p2i(p), p2i((void*) obj), str, str2);
2832 2833 }
2833 2834 };
2834 2835
2835 2836 class PrintReachableObjectClosure : public ObjectClosure {
2836 2837 private:
2837 2838 G1CollectedHeap* _g1h;
2838 2839 outputStream* _out;
2839 2840 VerifyOption _vo;
2840 2841 bool _all;
2841 2842 HeapRegion* _hr;
2842 2843
2843 2844 public:
2844 2845 PrintReachableObjectClosure(outputStream* out,
2845 2846 VerifyOption vo,
2846 2847 bool all,
2847 2848 HeapRegion* hr) :
2848 2849 _g1h(G1CollectedHeap::heap()),
2849 2850 _out(out), _vo(vo), _all(all), _hr(hr) { }
2850 2851
2851 2852 void do_object(oop o) {
2852 2853 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2853 2854 bool marked = _g1h->is_marked(o, _vo);
2854 2855 bool print_it = _all || over_tams || marked;
2855 2856
2856 2857 if (print_it) {
2857 2858 _out->print_cr(" "PTR_FORMAT"%s",
2858 2859 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2859 2860 PrintReachableOopClosure oopCl(_out, _vo, _all);
2860 2861 o->oop_iterate_no_header(&oopCl);
2861 2862 }
2862 2863 }
2863 2864 };
2864 2865
2865 2866 class PrintReachableRegionClosure : public HeapRegionClosure {
2866 2867 private:
2867 2868 G1CollectedHeap* _g1h;
2868 2869 outputStream* _out;
2869 2870 VerifyOption _vo;
2870 2871 bool _all;
2871 2872
2872 2873 public:
2873 2874 bool doHeapRegion(HeapRegion* hr) {
2874 2875 HeapWord* b = hr->bottom();
2875 2876 HeapWord* e = hr->end();
2876 2877 HeapWord* t = hr->top();
2877 2878 HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2878 2879 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2879 2880 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2880 2881 _out->cr();
2881 2882
2882 2883 HeapWord* from = b;
2883 2884 HeapWord* to = t;
2884 2885
2885 2886 if (to > from) {
2886 2887 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2887 2888 _out->cr();
2888 2889 PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2889 2890 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2890 2891 _out->cr();
2891 2892 }
2892 2893
2893 2894 return false;
2894 2895 }
2895 2896
2896 2897 PrintReachableRegionClosure(outputStream* out,
2897 2898 VerifyOption vo,
2898 2899 bool all) :
2899 2900 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2900 2901 };
2901 2902
2902 2903 void ConcurrentMark::print_reachable(const char* str,
2903 2904 VerifyOption vo,
2904 2905 bool all) {
2905 2906 gclog_or_tty->cr();
2906 2907 gclog_or_tty->print_cr("== Doing heap dump... ");
2907 2908
2908 2909 if (G1PrintReachableBaseFile == NULL) {
2909 2910 gclog_or_tty->print_cr(" #### error: no base file defined");
2910 2911 return;
2911 2912 }
2912 2913
2913 2914 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2914 2915 (JVM_MAXPATHLEN - 1)) {
2915 2916 gclog_or_tty->print_cr(" #### error: file name too long");
2916 2917 return;
2917 2918 }
2918 2919
2919 2920 char file_name[JVM_MAXPATHLEN];
2920 2921 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2921 2922 gclog_or_tty->print_cr(" dumping to file %s", file_name);
2922 2923
2923 2924 fileStream fout(file_name);
2924 2925 if (!fout.is_open()) {
2925 2926 gclog_or_tty->print_cr(" #### error: could not open file");
2926 2927 return;
2927 2928 }
2928 2929
2929 2930 outputStream* out = &fout;
2930 2931 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2931 2932 out->cr();
2932 2933
2933 2934 out->print_cr("--- ITERATING OVER REGIONS");
2934 2935 out->cr();
2935 2936 PrintReachableRegionClosure rcl(out, vo, all);
2936 2937 _g1h->heap_region_iterate(&rcl);
2937 2938 out->cr();
2938 2939
2939 2940 gclog_or_tty->print_cr(" done");
2940 2941 gclog_or_tty->flush();
2941 2942 }
2942 2943
2943 2944 #endif // PRODUCT
2944 2945
2945 2946 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2946 2947 // Note we are overriding the read-only view of the prev map here, via
2947 2948 // the cast.
2948 2949 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2949 2950 }
2950 2951
2951 2952 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2952 2953 _nextMarkBitMap->clearRange(mr);
2953 2954 }
2954 2955
2955 2956 HeapRegion*
2956 2957 ConcurrentMark::claim_region(uint worker_id) {
2957 2958 // "checkpoint" the finger
2958 2959 HeapWord* finger = _finger;
2959 2960
2960 2961 // _heap_end will not change underneath our feet; it only changes at
2961 2962 // yield points.
2962 2963 while (finger < _heap_end) {
2963 2964 assert(_g1h->is_in_g1_reserved(finger), "invariant");
2964 2965
2965 2966 // Note on how this code handles humongous regions. In the
2966 2967 // normal case the finger will reach the start of a "starts
2967 2968 // humongous" (SH) region. Its end will either be the end of the
2968 2969 // last "continues humongous" (CH) region in the sequence, or the
2969 2970 // standard end of the SH region (if the SH is the only region in
2970 2971 // the sequence). That way claim_region() will skip over the CH
2971 2972 // regions. However, there is a subtle race between a CM thread
2972 2973 // executing this method and a mutator thread doing a humongous
2973 2974 // object allocation. The two are not mutually exclusive as the CM
2974 2975 // thread does not need to hold the Heap_lock when it gets
2975 2976 // here. So there is a chance that claim_region() will come across
2976 2977 // a free region that's in the progress of becoming a SH or a CH
2977 2978 // region. In the former case, it will either
2978 2979 // a) Miss the update to the region's end, in which case it will
2979 2980 // visit every subsequent CH region, will find their bitmaps
2980 2981 // empty, and do nothing, or
2981 2982 // b) Will observe the update of the region's end (in which case
2982 2983 // it will skip the subsequent CH regions).
2983 2984 // If it comes across a region that suddenly becomes CH, the
2984 2985 // scenario will be similar to b). So, the race between
2985 2986 // claim_region() and a humongous object allocation might force us
2986 2987 // to do a bit of unnecessary work (due to some unnecessary bitmap
2987 2988 // iterations) but it should not introduce and correctness issues.
2988 2989 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2989 2990
2990 2991 // Above heap_region_containing_raw may return NULL as we always scan claim
2991 2992 // until the end of the heap. In this case, just jump to the next region.
2992 2993 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2993 2994
2994 2995 // Is the gap between reading the finger and doing the CAS too long?
2995 2996 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2996 2997 if (res == finger && curr_region != NULL) {
2997 2998 // we succeeded
2998 2999 HeapWord* bottom = curr_region->bottom();
2999 3000 HeapWord* limit = curr_region->next_top_at_mark_start();
3000 3001
3001 3002 if (verbose_low()) {
3002 3003 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
3003 3004 "["PTR_FORMAT", "PTR_FORMAT"), "
3004 3005 "limit = "PTR_FORMAT,
3005 3006 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
3006 3007 }
3007 3008
3008 3009 // notice that _finger == end cannot be guaranteed here since,
3009 3010 // someone else might have moved the finger even further
3010 3011 assert(_finger >= end, "the finger should have moved forward");
3011 3012
3012 3013 if (verbose_low()) {
3013 3014 gclog_or_tty->print_cr("[%u] we were successful with region = "
3014 3015 PTR_FORMAT, worker_id, p2i(curr_region));
3015 3016 }
3016 3017
3017 3018 if (limit > bottom) {
3018 3019 if (verbose_low()) {
3019 3020 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
3020 3021 "returning it ", worker_id, p2i(curr_region));
3021 3022 }
3022 3023 return curr_region;
3023 3024 } else {
3024 3025 assert(limit == bottom,
3025 3026 "the region limit should be at bottom");
3026 3027 if (verbose_low()) {
3027 3028 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3028 3029 "returning NULL", worker_id, p2i(curr_region));
3029 3030 }
3030 3031 // we return NULL and the caller should try calling
3031 3032 // claim_region() again.
3032 3033 return NULL;
3033 3034 }
3034 3035 } else {
3035 3036 assert(_finger > finger, "the finger should have moved forward");
3036 3037 if (verbose_low()) {
3037 3038 if (curr_region == NULL) {
3038 3039 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
3039 3040 "global finger = "PTR_FORMAT", "
3040 3041 "our finger = "PTR_FORMAT,
3041 3042 worker_id, p2i(_finger), p2i(finger));
3042 3043 } else {
3043 3044 gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3044 3045 "global finger = "PTR_FORMAT", "
3045 3046 "our finger = "PTR_FORMAT,
3046 3047 worker_id, p2i(_finger), p2i(finger));
3047 3048 }
3048 3049 }
3049 3050
3050 3051 // read it again
3051 3052 finger = _finger;
3052 3053 }
3053 3054 }
3054 3055
3055 3056 return NULL;
3056 3057 }
3057 3058
3058 3059 #ifndef PRODUCT
3059 3060 enum VerifyNoCSetOopsPhase {
3060 3061 VerifyNoCSetOopsStack,
3061 3062 VerifyNoCSetOopsQueues,
3062 3063 VerifyNoCSetOopsSATBCompleted,
3063 3064 VerifyNoCSetOopsSATBThread
3064 3065 };
3065 3066
3066 3067 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure {
3067 3068 private:
3068 3069 G1CollectedHeap* _g1h;
3069 3070 VerifyNoCSetOopsPhase _phase;
3070 3071 int _info;
3071 3072
3072 3073 const char* phase_str() {
3073 3074 switch (_phase) {
3074 3075 case VerifyNoCSetOopsStack: return "Stack";
3075 3076 case VerifyNoCSetOopsQueues: return "Queue";
3076 3077 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3077 3078 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers";
3078 3079 default: ShouldNotReachHere();
3079 3080 }
3080 3081 return NULL;
3081 3082 }
3082 3083
3083 3084 void do_object_work(oop obj) {
3084 3085 guarantee(!_g1h->obj_in_cs(obj),
3085 3086 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3086 3087 p2i((void*) obj), phase_str(), _info));
3087 3088 }
3088 3089
3089 3090 public:
3090 3091 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3091 3092
3092 3093 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3093 3094 _phase = phase;
3094 3095 _info = info;
3095 3096 }
3096 3097
3097 3098 virtual void do_oop(oop* p) {
3098 3099 oop obj = oopDesc::load_decode_heap_oop(p);
3099 3100 do_object_work(obj);
3100 3101 }
3101 3102
3102 3103 virtual void do_oop(narrowOop* p) {
3103 3104 // We should not come across narrow oops while scanning marking
3104 3105 // stacks and SATB buffers.
3105 3106 ShouldNotReachHere();
3106 3107 }
3107 3108
3108 3109 virtual void do_object(oop obj) {
3109 3110 do_object_work(obj);
3110 3111 }
3111 3112 };
3112 3113
3113 3114 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3114 3115 bool verify_enqueued_buffers,
3115 3116 bool verify_thread_buffers,
3116 3117 bool verify_fingers) {
3117 3118 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3118 3119 if (!G1CollectedHeap::heap()->mark_in_progress()) {
3119 3120 return;
3120 3121 }
3121 3122
3122 3123 VerifyNoCSetOopsClosure cl;
3123 3124
3124 3125 if (verify_stacks) {
3125 3126 // Verify entries on the global mark stack
3126 3127 cl.set_phase(VerifyNoCSetOopsStack);
3127 3128 _markStack.oops_do(&cl);
3128 3129
3129 3130 // Verify entries on the task queues
3130 3131 for (uint i = 0; i < _max_worker_id; i += 1) {
3131 3132 cl.set_phase(VerifyNoCSetOopsQueues, i);
3132 3133 CMTaskQueue* queue = _task_queues->queue(i);
3133 3134 queue->oops_do(&cl);
3134 3135 }
3135 3136 }
3136 3137
3137 3138 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3138 3139
3139 3140 // Verify entries on the enqueued SATB buffers
3140 3141 if (verify_enqueued_buffers) {
3141 3142 cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3142 3143 satb_qs.iterate_completed_buffers_read_only(&cl);
3143 3144 }
3144 3145
3145 3146 // Verify entries on the per-thread SATB buffers
3146 3147 if (verify_thread_buffers) {
3147 3148 cl.set_phase(VerifyNoCSetOopsSATBThread);
3148 3149 satb_qs.iterate_thread_buffers_read_only(&cl);
3149 3150 }
3150 3151
3151 3152 if (verify_fingers) {
3152 3153 // Verify the global finger
3153 3154 HeapWord* global_finger = finger();
3154 3155 if (global_finger != NULL && global_finger < _heap_end) {
3155 3156 // The global finger always points to a heap region boundary. We
3156 3157 // use heap_region_containing_raw() to get the containing region
3157 3158 // given that the global finger could be pointing to a free region
3158 3159 // which subsequently becomes continues humongous. If that
3159 3160 // happens, heap_region_containing() will return the bottom of the
3160 3161 // corresponding starts humongous region and the check below will
3161 3162 // not hold any more.
3162 3163 // Since we always iterate over all regions, we might get a NULL HeapRegion
3163 3164 // here.
3164 3165 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3165 3166 guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3166 3167 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3167 3168 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3168 3169 }
3169 3170
3170 3171 // Verify the task fingers
3171 3172 assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3172 3173 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3173 3174 CMTask* task = _tasks[i];
3174 3175 HeapWord* task_finger = task->finger();
3175 3176 if (task_finger != NULL && task_finger < _heap_end) {
3176 3177 // See above note on the global finger verification.
3177 3178 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3178 3179 guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3179 3180 !task_hr->in_collection_set(),
3180 3181 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3181 3182 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3182 3183 }
3183 3184 }
3184 3185 }
3185 3186 }
3186 3187 #endif // PRODUCT
3187 3188
3188 3189 // Aggregate the counting data that was constructed concurrently
3189 3190 // with marking.
3190 3191 class AggregateCountDataHRClosure: public HeapRegionClosure {
3191 3192 G1CollectedHeap* _g1h;
3192 3193 ConcurrentMark* _cm;
3193 3194 CardTableModRefBS* _ct_bs;
3194 3195 BitMap* _cm_card_bm;
3195 3196 uint _max_worker_id;
3196 3197
3197 3198 public:
3198 3199 AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3199 3200 BitMap* cm_card_bm,
3200 3201 uint max_worker_id) :
3201 3202 _g1h(g1h), _cm(g1h->concurrent_mark()),
3202 3203 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3203 3204 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3204 3205
3205 3206 bool doHeapRegion(HeapRegion* hr) {
3206 3207 if (hr->continuesHumongous()) {
3207 3208 // We will ignore these here and process them when their
3208 3209 // associated "starts humongous" region is processed.
3209 3210 // Note that we cannot rely on their associated
3210 3211 // "starts humongous" region to have their bit set to 1
3211 3212 // since, due to the region chunking in the parallel region
3212 3213 // iteration, a "continues humongous" region might be visited
3213 3214 // before its associated "starts humongous".
3214 3215 return false;
3215 3216 }
3216 3217
3217 3218 HeapWord* start = hr->bottom();
3218 3219 HeapWord* limit = hr->next_top_at_mark_start();
3219 3220 HeapWord* end = hr->end();
3220 3221
3221 3222 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3222 3223 err_msg("Preconditions not met - "
3223 3224 "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3224 3225 "top: "PTR_FORMAT", end: "PTR_FORMAT,
3225 3226 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3226 3227
3227 3228 assert(hr->next_marked_bytes() == 0, "Precondition");
3228 3229
3229 3230 if (start == limit) {
3230 3231 // NTAMS of this region has not been set so nothing to do.
3231 3232 return false;
3232 3233 }
3233 3234
3234 3235 // 'start' should be in the heap.
3235 3236 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3236 3237 // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3237 3238 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3238 3239
3239 3240 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3240 3241 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3241 3242 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3242 3243
3243 3244 // If ntams is not card aligned then we bump card bitmap index
3244 3245 // for limit so that we get the all the cards spanned by
3245 3246 // the object ending at ntams.
3246 3247 // Note: if this is the last region in the heap then ntams
3247 3248 // could be actually just beyond the end of the the heap;
3248 3249 // limit_idx will then correspond to a (non-existent) card
3249 3250 // that is also outside the heap.
3250 3251 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3251 3252 limit_idx += 1;
3252 3253 }
3253 3254
3254 3255 assert(limit_idx <= end_idx, "or else use atomics");
3255 3256
3256 3257 // Aggregate the "stripe" in the count data associated with hr.
3257 3258 uint hrm_index = hr->hrm_index();
3258 3259 size_t marked_bytes = 0;
3259 3260
3260 3261 for (uint i = 0; i < _max_worker_id; i += 1) {
3261 3262 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3262 3263 BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3263 3264
3264 3265 // Fetch the marked_bytes in this region for task i and
3265 3266 // add it to the running total for this region.
3266 3267 marked_bytes += marked_bytes_array[hrm_index];
3267 3268
3268 3269 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3269 3270 // into the global card bitmap.
3270 3271 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3271 3272
3272 3273 while (scan_idx < limit_idx) {
3273 3274 assert(task_card_bm->at(scan_idx) == true, "should be");
3274 3275 _cm_card_bm->set_bit(scan_idx);
3275 3276 assert(_cm_card_bm->at(scan_idx) == true, "should be");
3276 3277
3277 3278 // BitMap::get_next_one_offset() can handle the case when
3278 3279 // its left_offset parameter is greater than its right_offset
3279 3280 // parameter. It does, however, have an early exit if
3280 3281 // left_offset == right_offset. So let's limit the value
3281 3282 // passed in for left offset here.
3282 3283 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3283 3284 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3284 3285 }
3285 3286 }
3286 3287
3287 3288 // Update the marked bytes for this region.
3288 3289 hr->add_to_marked_bytes(marked_bytes);
3289 3290
3290 3291 // Next heap region
3291 3292 return false;
3292 3293 }
3293 3294 };
3294 3295
3295 3296 class G1AggregateCountDataTask: public AbstractGangTask {
3296 3297 protected:
3297 3298 G1CollectedHeap* _g1h;
3298 3299 ConcurrentMark* _cm;
3299 3300 BitMap* _cm_card_bm;
3300 3301 uint _max_worker_id;
3301 3302 int _active_workers;
3302 3303
3303 3304 public:
3304 3305 G1AggregateCountDataTask(G1CollectedHeap* g1h,
3305 3306 ConcurrentMark* cm,
3306 3307 BitMap* cm_card_bm,
3307 3308 uint max_worker_id,
3308 3309 int n_workers) :
3309 3310 AbstractGangTask("Count Aggregation"),
3310 3311 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3311 3312 _max_worker_id(max_worker_id),
3312 3313 _active_workers(n_workers) { }
3313 3314
3314 3315 void work(uint worker_id) {
3315 3316 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3316 3317
3317 3318 if (G1CollectedHeap::use_parallel_gc_threads()) {
3318 3319 _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3319 3320 _active_workers,
3320 3321 HeapRegion::AggregateCountClaimValue);
3321 3322 } else {
3322 3323 _g1h->heap_region_iterate(&cl);
3323 3324 }
3324 3325 }
3325 3326 };
3326 3327
3327 3328
3328 3329 void ConcurrentMark::aggregate_count_data() {
3329 3330 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3330 3331 _g1h->workers()->active_workers() :
3331 3332 1);
3332 3333
3333 3334 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3334 3335 _max_worker_id, n_workers);
3335 3336
3336 3337 if (G1CollectedHeap::use_parallel_gc_threads()) {
3337 3338 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3338 3339 "sanity check");
↓ open down ↓ |
1151 lines elided |
↑ open up ↑ |
3339 3340 _g1h->set_par_threads(n_workers);
3340 3341 _g1h->workers()->run_task(&g1_par_agg_task);
3341 3342 _g1h->set_par_threads(0);
3342 3343
3343 3344 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3344 3345 "sanity check");
3345 3346 _g1h->reset_heap_region_claim_values();
3346 3347 } else {
3347 3348 g1_par_agg_task.work(0);
3348 3349 }
3349 - _g1h->allocation_context_stats().update_at_remark();
3350 3350 }
3351 3351
3352 3352 // Clear the per-worker arrays used to store the per-region counting data
3353 3353 void ConcurrentMark::clear_all_count_data() {
3354 3354 // Clear the global card bitmap - it will be filled during
3355 3355 // liveness count aggregation (during remark) and the
3356 3356 // final counting task.
3357 3357 _card_bm.clear();
3358 3358
3359 3359 // Clear the global region bitmap - it will be filled as part
3360 3360 // of the final counting task.
3361 3361 _region_bm.clear();
3362 3362
3363 3363 uint max_regions = _g1h->max_regions();
3364 3364 assert(_max_worker_id > 0, "uninitialized");
3365 3365
3366 3366 for (uint i = 0; i < _max_worker_id; i += 1) {
3367 3367 BitMap* task_card_bm = count_card_bitmap_for(i);
3368 3368 size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3369 3369
3370 3370 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3371 3371 assert(marked_bytes_array != NULL, "uninitialized");
3372 3372
3373 3373 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3374 3374 task_card_bm->clear();
3375 3375 }
3376 3376 }
3377 3377
3378 3378 void ConcurrentMark::print_stats() {
3379 3379 if (verbose_stats()) {
3380 3380 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3381 3381 for (size_t i = 0; i < _active_tasks; ++i) {
3382 3382 _tasks[i]->print_stats();
3383 3383 gclog_or_tty->print_cr("---------------------------------------------------------------------");
3384 3384 }
3385 3385 }
3386 3386 }
3387 3387
3388 3388 // abandon current marking iteration due to a Full GC
3389 3389 void ConcurrentMark::abort() {
3390 3390 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3391 3391 // concurrent bitmap clearing.
3392 3392 _nextMarkBitMap->clearAll();
3393 3393
3394 3394 // Note we cannot clear the previous marking bitmap here
3395 3395 // since VerifyDuringGC verifies the objects marked during
3396 3396 // a full GC against the previous bitmap.
3397 3397
3398 3398 // Clear the liveness counting data
3399 3399 clear_all_count_data();
3400 3400 // Empty mark stack
3401 3401 reset_marking_state();
3402 3402 for (uint i = 0; i < _max_worker_id; ++i) {
3403 3403 _tasks[i]->clear_region_fields();
3404 3404 }
3405 3405 _first_overflow_barrier_sync.abort();
3406 3406 _second_overflow_barrier_sync.abort();
3407 3407 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3408 3408 if (!gc_id.is_undefined()) {
3409 3409 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3410 3410 // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3411 3411 _aborted_gc_id = gc_id;
3412 3412 }
3413 3413 _has_aborted = true;
3414 3414
3415 3415 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3416 3416 satb_mq_set.abandon_partial_marking();
3417 3417 // This can be called either during or outside marking, we'll read
3418 3418 // the expected_active value from the SATB queue set.
3419 3419 satb_mq_set.set_active_all_threads(
3420 3420 false, /* new active value */
3421 3421 satb_mq_set.is_active() /* expected_active */);
3422 3422
3423 3423 _g1h->trace_heap_after_concurrent_cycle();
3424 3424 _g1h->register_concurrent_cycle_end();
3425 3425 }
3426 3426
3427 3427 const GCId& ConcurrentMark::concurrent_gc_id() {
3428 3428 if (has_aborted()) {
3429 3429 return _aborted_gc_id;
3430 3430 }
3431 3431 return _g1h->gc_tracer_cm()->gc_id();
3432 3432 }
3433 3433
3434 3434 static void print_ms_time_info(const char* prefix, const char* name,
3435 3435 NumberSeq& ns) {
3436 3436 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3437 3437 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3438 3438 if (ns.num() > 0) {
3439 3439 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]",
3440 3440 prefix, ns.sd(), ns.maximum());
3441 3441 }
3442 3442 }
3443 3443
3444 3444 void ConcurrentMark::print_summary_info() {
3445 3445 gclog_or_tty->print_cr(" Concurrent marking:");
3446 3446 print_ms_time_info(" ", "init marks", _init_times);
3447 3447 print_ms_time_info(" ", "remarks", _remark_times);
3448 3448 {
3449 3449 print_ms_time_info(" ", "final marks", _remark_mark_times);
3450 3450 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times);
3451 3451
3452 3452 }
3453 3453 print_ms_time_info(" ", "cleanups", _cleanup_times);
3454 3454 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).",
3455 3455 _total_counting_time,
3456 3456 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3457 3457 (double)_cleanup_times.num()
3458 3458 : 0.0));
3459 3459 if (G1ScrubRemSets) {
3460 3460 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).",
3461 3461 _total_rs_scrub_time,
3462 3462 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3463 3463 (double)_cleanup_times.num()
3464 3464 : 0.0));
3465 3465 }
3466 3466 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.",
3467 3467 (_init_times.sum() + _remark_times.sum() +
3468 3468 _cleanup_times.sum())/1000.0);
3469 3469 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s "
3470 3470 "(%8.2f s marking).",
3471 3471 cmThread()->vtime_accum(),
3472 3472 cmThread()->vtime_mark_accum());
3473 3473 }
3474 3474
3475 3475 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3476 3476 if (use_parallel_marking_threads()) {
3477 3477 _parallel_workers->print_worker_threads_on(st);
3478 3478 }
3479 3479 }
3480 3480
3481 3481 void ConcurrentMark::print_on_error(outputStream* st) const {
3482 3482 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3483 3483 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3484 3484 _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3485 3485 _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3486 3486 }
3487 3487
3488 3488 // We take a break if someone is trying to stop the world.
3489 3489 bool ConcurrentMark::do_yield_check(uint worker_id) {
3490 3490 if (SuspendibleThreadSet::should_yield()) {
3491 3491 if (worker_id == 0) {
3492 3492 _g1h->g1_policy()->record_concurrent_pause();
3493 3493 }
3494 3494 SuspendibleThreadSet::yield();
3495 3495 return true;
3496 3496 } else {
3497 3497 return false;
3498 3498 }
3499 3499 }
3500 3500
3501 3501 #ifndef PRODUCT
3502 3502 // for debugging purposes
3503 3503 void ConcurrentMark::print_finger() {
3504 3504 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3505 3505 p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3506 3506 for (uint i = 0; i < _max_worker_id; ++i) {
3507 3507 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3508 3508 }
3509 3509 gclog_or_tty->cr();
3510 3510 }
3511 3511 #endif
3512 3512
3513 3513 void CMTask::scan_object(oop obj) {
3514 3514 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3515 3515
3516 3516 if (_cm->verbose_high()) {
3517 3517 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3518 3518 _worker_id, p2i((void*) obj));
3519 3519 }
3520 3520
3521 3521 size_t obj_size = obj->size();
3522 3522 _words_scanned += obj_size;
3523 3523
3524 3524 obj->oop_iterate(_cm_oop_closure);
3525 3525 statsOnly( ++_objs_scanned );
3526 3526 check_limits();
3527 3527 }
3528 3528
3529 3529 // Closure for iteration over bitmaps
3530 3530 class CMBitMapClosure : public BitMapClosure {
3531 3531 private:
3532 3532 // the bitmap that is being iterated over
3533 3533 CMBitMap* _nextMarkBitMap;
3534 3534 ConcurrentMark* _cm;
3535 3535 CMTask* _task;
3536 3536
3537 3537 public:
3538 3538 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3539 3539 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3540 3540
3541 3541 bool do_bit(size_t offset) {
3542 3542 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3543 3543 assert(_nextMarkBitMap->isMarked(addr), "invariant");
3544 3544 assert( addr < _cm->finger(), "invariant");
3545 3545
3546 3546 statsOnly( _task->increase_objs_found_on_bitmap() );
3547 3547 assert(addr >= _task->finger(), "invariant");
3548 3548
3549 3549 // We move that task's local finger along.
3550 3550 _task->move_finger_to(addr);
3551 3551
3552 3552 _task->scan_object(oop(addr));
3553 3553 // we only partially drain the local queue and global stack
3554 3554 _task->drain_local_queue(true);
3555 3555 _task->drain_global_stack(true);
3556 3556
3557 3557 // if the has_aborted flag has been raised, we need to bail out of
3558 3558 // the iteration
3559 3559 return !_task->has_aborted();
3560 3560 }
3561 3561 };
3562 3562
3563 3563 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3564 3564 ConcurrentMark* cm,
3565 3565 CMTask* task)
3566 3566 : _g1h(g1h), _cm(cm), _task(task) {
3567 3567 assert(_ref_processor == NULL, "should be initialized to NULL");
3568 3568
3569 3569 if (G1UseConcMarkReferenceProcessing) {
3570 3570 _ref_processor = g1h->ref_processor_cm();
3571 3571 assert(_ref_processor != NULL, "should not be NULL");
3572 3572 }
3573 3573 }
3574 3574
3575 3575 void CMTask::setup_for_region(HeapRegion* hr) {
3576 3576 assert(hr != NULL,
3577 3577 "claim_region() should have filtered out NULL regions");
3578 3578 assert(!hr->continuesHumongous(),
3579 3579 "claim_region() should have filtered out continues humongous regions");
3580 3580
3581 3581 if (_cm->verbose_low()) {
3582 3582 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3583 3583 _worker_id, p2i(hr));
3584 3584 }
3585 3585
3586 3586 _curr_region = hr;
3587 3587 _finger = hr->bottom();
3588 3588 update_region_limit();
3589 3589 }
3590 3590
3591 3591 void CMTask::update_region_limit() {
3592 3592 HeapRegion* hr = _curr_region;
3593 3593 HeapWord* bottom = hr->bottom();
3594 3594 HeapWord* limit = hr->next_top_at_mark_start();
3595 3595
3596 3596 if (limit == bottom) {
3597 3597 if (_cm->verbose_low()) {
3598 3598 gclog_or_tty->print_cr("[%u] found an empty region "
3599 3599 "["PTR_FORMAT", "PTR_FORMAT")",
3600 3600 _worker_id, p2i(bottom), p2i(limit));
3601 3601 }
3602 3602 // The region was collected underneath our feet.
3603 3603 // We set the finger to bottom to ensure that the bitmap
3604 3604 // iteration that will follow this will not do anything.
3605 3605 // (this is not a condition that holds when we set the region up,
3606 3606 // as the region is not supposed to be empty in the first place)
3607 3607 _finger = bottom;
3608 3608 } else if (limit >= _region_limit) {
3609 3609 assert(limit >= _finger, "peace of mind");
3610 3610 } else {
3611 3611 assert(limit < _region_limit, "only way to get here");
3612 3612 // This can happen under some pretty unusual circumstances. An
3613 3613 // evacuation pause empties the region underneath our feet (NTAMS
3614 3614 // at bottom). We then do some allocation in the region (NTAMS
3615 3615 // stays at bottom), followed by the region being used as a GC
3616 3616 // alloc region (NTAMS will move to top() and the objects
3617 3617 // originally below it will be grayed). All objects now marked in
3618 3618 // the region are explicitly grayed, if below the global finger,
3619 3619 // and we do not need in fact to scan anything else. So, we simply
3620 3620 // set _finger to be limit to ensure that the bitmap iteration
3621 3621 // doesn't do anything.
3622 3622 _finger = limit;
3623 3623 }
3624 3624
3625 3625 _region_limit = limit;
3626 3626 }
3627 3627
3628 3628 void CMTask::giveup_current_region() {
3629 3629 assert(_curr_region != NULL, "invariant");
3630 3630 if (_cm->verbose_low()) {
3631 3631 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3632 3632 _worker_id, p2i(_curr_region));
3633 3633 }
3634 3634 clear_region_fields();
3635 3635 }
3636 3636
3637 3637 void CMTask::clear_region_fields() {
3638 3638 // Values for these three fields that indicate that we're not
3639 3639 // holding on to a region.
3640 3640 _curr_region = NULL;
3641 3641 _finger = NULL;
3642 3642 _region_limit = NULL;
3643 3643 }
3644 3644
3645 3645 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3646 3646 if (cm_oop_closure == NULL) {
3647 3647 assert(_cm_oop_closure != NULL, "invariant");
3648 3648 } else {
3649 3649 assert(_cm_oop_closure == NULL, "invariant");
3650 3650 }
3651 3651 _cm_oop_closure = cm_oop_closure;
3652 3652 }
3653 3653
3654 3654 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3655 3655 guarantee(nextMarkBitMap != NULL, "invariant");
3656 3656
3657 3657 if (_cm->verbose_low()) {
3658 3658 gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3659 3659 }
3660 3660
3661 3661 _nextMarkBitMap = nextMarkBitMap;
3662 3662 clear_region_fields();
3663 3663
3664 3664 _calls = 0;
3665 3665 _elapsed_time_ms = 0.0;
3666 3666 _termination_time_ms = 0.0;
3667 3667 _termination_start_time_ms = 0.0;
3668 3668
3669 3669 #if _MARKING_STATS_
3670 3670 _local_pushes = 0;
3671 3671 _local_pops = 0;
3672 3672 _local_max_size = 0;
3673 3673 _objs_scanned = 0;
3674 3674 _global_pushes = 0;
3675 3675 _global_pops = 0;
3676 3676 _global_max_size = 0;
3677 3677 _global_transfers_to = 0;
3678 3678 _global_transfers_from = 0;
3679 3679 _regions_claimed = 0;
3680 3680 _objs_found_on_bitmap = 0;
3681 3681 _satb_buffers_processed = 0;
3682 3682 _steal_attempts = 0;
3683 3683 _steals = 0;
3684 3684 _aborted = 0;
3685 3685 _aborted_overflow = 0;
3686 3686 _aborted_cm_aborted = 0;
3687 3687 _aborted_yield = 0;
3688 3688 _aborted_timed_out = 0;
3689 3689 _aborted_satb = 0;
3690 3690 _aborted_termination = 0;
3691 3691 #endif // _MARKING_STATS_
3692 3692 }
3693 3693
3694 3694 bool CMTask::should_exit_termination() {
3695 3695 regular_clock_call();
3696 3696 // This is called when we are in the termination protocol. We should
3697 3697 // quit if, for some reason, this task wants to abort or the global
3698 3698 // stack is not empty (this means that we can get work from it).
3699 3699 return !_cm->mark_stack_empty() || has_aborted();
3700 3700 }
3701 3701
3702 3702 void CMTask::reached_limit() {
3703 3703 assert(_words_scanned >= _words_scanned_limit ||
3704 3704 _refs_reached >= _refs_reached_limit ,
3705 3705 "shouldn't have been called otherwise");
3706 3706 regular_clock_call();
3707 3707 }
3708 3708
3709 3709 void CMTask::regular_clock_call() {
3710 3710 if (has_aborted()) return;
3711 3711
3712 3712 // First, we need to recalculate the words scanned and refs reached
3713 3713 // limits for the next clock call.
3714 3714 recalculate_limits();
3715 3715
3716 3716 // During the regular clock call we do the following
3717 3717
3718 3718 // (1) If an overflow has been flagged, then we abort.
3719 3719 if (_cm->has_overflown()) {
3720 3720 set_has_aborted();
3721 3721 return;
3722 3722 }
3723 3723
3724 3724 // If we are not concurrent (i.e. we're doing remark) we don't need
3725 3725 // to check anything else. The other steps are only needed during
3726 3726 // the concurrent marking phase.
3727 3727 if (!concurrent()) return;
3728 3728
3729 3729 // (2) If marking has been aborted for Full GC, then we also abort.
3730 3730 if (_cm->has_aborted()) {
3731 3731 set_has_aborted();
3732 3732 statsOnly( ++_aborted_cm_aborted );
3733 3733 return;
3734 3734 }
3735 3735
3736 3736 double curr_time_ms = os::elapsedVTime() * 1000.0;
3737 3737
3738 3738 // (3) If marking stats are enabled, then we update the step history.
3739 3739 #if _MARKING_STATS_
3740 3740 if (_words_scanned >= _words_scanned_limit) {
3741 3741 ++_clock_due_to_scanning;
3742 3742 }
3743 3743 if (_refs_reached >= _refs_reached_limit) {
3744 3744 ++_clock_due_to_marking;
3745 3745 }
3746 3746
3747 3747 double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3748 3748 _interval_start_time_ms = curr_time_ms;
3749 3749 _all_clock_intervals_ms.add(last_interval_ms);
3750 3750
3751 3751 if (_cm->verbose_medium()) {
3752 3752 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3753 3753 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3754 3754 _worker_id, last_interval_ms,
3755 3755 _words_scanned,
3756 3756 (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3757 3757 _refs_reached,
3758 3758 (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3759 3759 }
3760 3760 #endif // _MARKING_STATS_
3761 3761
3762 3762 // (4) We check whether we should yield. If we have to, then we abort.
3763 3763 if (SuspendibleThreadSet::should_yield()) {
3764 3764 // We should yield. To do this we abort the task. The caller is
3765 3765 // responsible for yielding.
3766 3766 set_has_aborted();
3767 3767 statsOnly( ++_aborted_yield );
3768 3768 return;
3769 3769 }
3770 3770
3771 3771 // (5) We check whether we've reached our time quota. If we have,
3772 3772 // then we abort.
3773 3773 double elapsed_time_ms = curr_time_ms - _start_time_ms;
3774 3774 if (elapsed_time_ms > _time_target_ms) {
3775 3775 set_has_aborted();
3776 3776 _has_timed_out = true;
3777 3777 statsOnly( ++_aborted_timed_out );
3778 3778 return;
3779 3779 }
3780 3780
3781 3781 // (6) Finally, we check whether there are enough completed STAB
3782 3782 // buffers available for processing. If there are, we abort.
3783 3783 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3784 3784 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3785 3785 if (_cm->verbose_low()) {
3786 3786 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3787 3787 _worker_id);
3788 3788 }
3789 3789 // we do need to process SATB buffers, we'll abort and restart
3790 3790 // the marking task to do so
3791 3791 set_has_aborted();
3792 3792 statsOnly( ++_aborted_satb );
3793 3793 return;
3794 3794 }
3795 3795 }
3796 3796
3797 3797 void CMTask::recalculate_limits() {
3798 3798 _real_words_scanned_limit = _words_scanned + words_scanned_period;
3799 3799 _words_scanned_limit = _real_words_scanned_limit;
3800 3800
3801 3801 _real_refs_reached_limit = _refs_reached + refs_reached_period;
3802 3802 _refs_reached_limit = _real_refs_reached_limit;
3803 3803 }
3804 3804
3805 3805 void CMTask::decrease_limits() {
3806 3806 // This is called when we believe that we're going to do an infrequent
3807 3807 // operation which will increase the per byte scanned cost (i.e. move
3808 3808 // entries to/from the global stack). It basically tries to decrease the
3809 3809 // scanning limit so that the clock is called earlier.
3810 3810
3811 3811 if (_cm->verbose_medium()) {
3812 3812 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3813 3813 }
3814 3814
3815 3815 _words_scanned_limit = _real_words_scanned_limit -
3816 3816 3 * words_scanned_period / 4;
3817 3817 _refs_reached_limit = _real_refs_reached_limit -
3818 3818 3 * refs_reached_period / 4;
3819 3819 }
3820 3820
3821 3821 void CMTask::move_entries_to_global_stack() {
3822 3822 // local array where we'll store the entries that will be popped
3823 3823 // from the local queue
3824 3824 oop buffer[global_stack_transfer_size];
3825 3825
3826 3826 int n = 0;
3827 3827 oop obj;
3828 3828 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3829 3829 buffer[n] = obj;
3830 3830 ++n;
3831 3831 }
3832 3832
3833 3833 if (n > 0) {
3834 3834 // we popped at least one entry from the local queue
3835 3835
3836 3836 statsOnly( ++_global_transfers_to; _local_pops += n );
3837 3837
3838 3838 if (!_cm->mark_stack_push(buffer, n)) {
3839 3839 if (_cm->verbose_low()) {
3840 3840 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3841 3841 _worker_id);
3842 3842 }
3843 3843 set_has_aborted();
3844 3844 } else {
3845 3845 // the transfer was successful
3846 3846
3847 3847 if (_cm->verbose_medium()) {
3848 3848 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3849 3849 _worker_id, n);
3850 3850 }
3851 3851 statsOnly( int tmp_size = _cm->mark_stack_size();
3852 3852 if (tmp_size > _global_max_size) {
3853 3853 _global_max_size = tmp_size;
3854 3854 }
3855 3855 _global_pushes += n );
3856 3856 }
3857 3857 }
3858 3858
3859 3859 // this operation was quite expensive, so decrease the limits
3860 3860 decrease_limits();
3861 3861 }
3862 3862
3863 3863 void CMTask::get_entries_from_global_stack() {
3864 3864 // local array where we'll store the entries that will be popped
3865 3865 // from the global stack.
3866 3866 oop buffer[global_stack_transfer_size];
3867 3867 int n;
3868 3868 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3869 3869 assert(n <= global_stack_transfer_size,
3870 3870 "we should not pop more than the given limit");
3871 3871 if (n > 0) {
3872 3872 // yes, we did actually pop at least one entry
3873 3873
3874 3874 statsOnly( ++_global_transfers_from; _global_pops += n );
3875 3875 if (_cm->verbose_medium()) {
3876 3876 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3877 3877 _worker_id, n);
3878 3878 }
3879 3879 for (int i = 0; i < n; ++i) {
3880 3880 bool success = _task_queue->push(buffer[i]);
3881 3881 // We only call this when the local queue is empty or under a
3882 3882 // given target limit. So, we do not expect this push to fail.
3883 3883 assert(success, "invariant");
3884 3884 }
3885 3885
3886 3886 statsOnly( int tmp_size = _task_queue->size();
3887 3887 if (tmp_size > _local_max_size) {
3888 3888 _local_max_size = tmp_size;
3889 3889 }
3890 3890 _local_pushes += n );
3891 3891 }
3892 3892
3893 3893 // this operation was quite expensive, so decrease the limits
3894 3894 decrease_limits();
3895 3895 }
3896 3896
3897 3897 void CMTask::drain_local_queue(bool partially) {
3898 3898 if (has_aborted()) return;
3899 3899
3900 3900 // Decide what the target size is, depending whether we're going to
3901 3901 // drain it partially (so that other tasks can steal if they run out
3902 3902 // of things to do) or totally (at the very end).
3903 3903 size_t target_size;
3904 3904 if (partially) {
3905 3905 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3906 3906 } else {
3907 3907 target_size = 0;
3908 3908 }
3909 3909
3910 3910 if (_task_queue->size() > target_size) {
3911 3911 if (_cm->verbose_high()) {
3912 3912 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3913 3913 _worker_id, target_size);
3914 3914 }
3915 3915
3916 3916 oop obj;
3917 3917 bool ret = _task_queue->pop_local(obj);
3918 3918 while (ret) {
3919 3919 statsOnly( ++_local_pops );
3920 3920
3921 3921 if (_cm->verbose_high()) {
3922 3922 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3923 3923 p2i((void*) obj));
3924 3924 }
3925 3925
3926 3926 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3927 3927 assert(!_g1h->is_on_master_free_list(
3928 3928 _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3929 3929
3930 3930 scan_object(obj);
3931 3931
3932 3932 if (_task_queue->size() <= target_size || has_aborted()) {
3933 3933 ret = false;
3934 3934 } else {
3935 3935 ret = _task_queue->pop_local(obj);
3936 3936 }
3937 3937 }
3938 3938
3939 3939 if (_cm->verbose_high()) {
3940 3940 gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3941 3941 _worker_id, _task_queue->size());
3942 3942 }
3943 3943 }
3944 3944 }
3945 3945
3946 3946 void CMTask::drain_global_stack(bool partially) {
3947 3947 if (has_aborted()) return;
3948 3948
3949 3949 // We have a policy to drain the local queue before we attempt to
3950 3950 // drain the global stack.
3951 3951 assert(partially || _task_queue->size() == 0, "invariant");
3952 3952
3953 3953 // Decide what the target size is, depending whether we're going to
3954 3954 // drain it partially (so that other tasks can steal if they run out
3955 3955 // of things to do) or totally (at the very end). Notice that,
3956 3956 // because we move entries from the global stack in chunks or
3957 3957 // because another task might be doing the same, we might in fact
3958 3958 // drop below the target. But, this is not a problem.
3959 3959 size_t target_size;
3960 3960 if (partially) {
3961 3961 target_size = _cm->partial_mark_stack_size_target();
3962 3962 } else {
3963 3963 target_size = 0;
3964 3964 }
3965 3965
3966 3966 if (_cm->mark_stack_size() > target_size) {
3967 3967 if (_cm->verbose_low()) {
3968 3968 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3969 3969 _worker_id, target_size);
3970 3970 }
3971 3971
3972 3972 while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3973 3973 get_entries_from_global_stack();
3974 3974 drain_local_queue(partially);
3975 3975 }
3976 3976
3977 3977 if (_cm->verbose_low()) {
3978 3978 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3979 3979 _worker_id, _cm->mark_stack_size());
3980 3980 }
3981 3981 }
3982 3982 }
3983 3983
3984 3984 // SATB Queue has several assumptions on whether to call the par or
3985 3985 // non-par versions of the methods. this is why some of the code is
3986 3986 // replicated. We should really get rid of the single-threaded version
3987 3987 // of the code to simplify things.
3988 3988 void CMTask::drain_satb_buffers() {
3989 3989 if (has_aborted()) return;
3990 3990
3991 3991 // We set this so that the regular clock knows that we're in the
3992 3992 // middle of draining buffers and doesn't set the abort flag when it
3993 3993 // notices that SATB buffers are available for draining. It'd be
3994 3994 // very counter productive if it did that. :-)
3995 3995 _draining_satb_buffers = true;
3996 3996
3997 3997 CMObjectClosure oc(this);
3998 3998 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3999 3999 if (G1CollectedHeap::use_parallel_gc_threads()) {
4000 4000 satb_mq_set.set_par_closure(_worker_id, &oc);
4001 4001 } else {
4002 4002 satb_mq_set.set_closure(&oc);
4003 4003 }
4004 4004
4005 4005 // This keeps claiming and applying the closure to completed buffers
4006 4006 // until we run out of buffers or we need to abort.
4007 4007 if (G1CollectedHeap::use_parallel_gc_threads()) {
4008 4008 while (!has_aborted() &&
4009 4009 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
4010 4010 if (_cm->verbose_medium()) {
4011 4011 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4012 4012 }
4013 4013 statsOnly( ++_satb_buffers_processed );
4014 4014 regular_clock_call();
4015 4015 }
4016 4016 } else {
4017 4017 while (!has_aborted() &&
4018 4018 satb_mq_set.apply_closure_to_completed_buffer()) {
4019 4019 if (_cm->verbose_medium()) {
4020 4020 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
4021 4021 }
4022 4022 statsOnly( ++_satb_buffers_processed );
4023 4023 regular_clock_call();
4024 4024 }
4025 4025 }
4026 4026
4027 4027 _draining_satb_buffers = false;
4028 4028
4029 4029 assert(has_aborted() ||
4030 4030 concurrent() ||
4031 4031 satb_mq_set.completed_buffers_num() == 0, "invariant");
4032 4032
4033 4033 if (G1CollectedHeap::use_parallel_gc_threads()) {
4034 4034 satb_mq_set.set_par_closure(_worker_id, NULL);
4035 4035 } else {
4036 4036 satb_mq_set.set_closure(NULL);
4037 4037 }
4038 4038
4039 4039 // again, this was a potentially expensive operation, decrease the
4040 4040 // limits to get the regular clock call early
4041 4041 decrease_limits();
4042 4042 }
4043 4043
4044 4044 void CMTask::print_stats() {
4045 4045 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4046 4046 _worker_id, _calls);
4047 4047 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4048 4048 _elapsed_time_ms, _termination_time_ms);
4049 4049 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4050 4050 _step_times_ms.num(), _step_times_ms.avg(),
4051 4051 _step_times_ms.sd());
4052 4052 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4053 4053 _step_times_ms.maximum(), _step_times_ms.sum());
4054 4054
4055 4055 #if _MARKING_STATS_
4056 4056 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4057 4057 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4058 4058 _all_clock_intervals_ms.sd());
4059 4059 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms",
4060 4060 _all_clock_intervals_ms.maximum(),
4061 4061 _all_clock_intervals_ms.sum());
4062 4062 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d",
4063 4063 _clock_due_to_scanning, _clock_due_to_marking);
4064 4064 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d",
4065 4065 _objs_scanned, _objs_found_on_bitmap);
4066 4066 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d",
4067 4067 _local_pushes, _local_pops, _local_max_size);
4068 4068 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d",
4069 4069 _global_pushes, _global_pops, _global_max_size);
4070 4070 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d",
4071 4071 _global_transfers_to,_global_transfers_from);
4072 4072 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed);
4073 4073 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed);
4074 4074 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d",
4075 4075 _steal_attempts, _steals);
4076 4076 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted);
4077 4077 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d",
4078 4078 _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4079 4079 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d",
4080 4080 _aborted_timed_out, _aborted_satb, _aborted_termination);
4081 4081 #endif // _MARKING_STATS_
4082 4082 }
4083 4083
4084 4084 /*****************************************************************************
4085 4085
4086 4086 The do_marking_step(time_target_ms, ...) method is the building
4087 4087 block of the parallel marking framework. It can be called in parallel
4088 4088 with other invocations of do_marking_step() on different tasks
4089 4089 (but only one per task, obviously) and concurrently with the
4090 4090 mutator threads, or during remark, hence it eliminates the need
4091 4091 for two versions of the code. When called during remark, it will
4092 4092 pick up from where the task left off during the concurrent marking
4093 4093 phase. Interestingly, tasks are also claimable during evacuation
4094 4094 pauses too, since do_marking_step() ensures that it aborts before
4095 4095 it needs to yield.
4096 4096
4097 4097 The data structures that it uses to do marking work are the
4098 4098 following:
4099 4099
4100 4100 (1) Marking Bitmap. If there are gray objects that appear only
4101 4101 on the bitmap (this happens either when dealing with an overflow
4102 4102 or when the initial marking phase has simply marked the roots
4103 4103 and didn't push them on the stack), then tasks claim heap
4104 4104 regions whose bitmap they then scan to find gray objects. A
4105 4105 global finger indicates where the end of the last claimed region
4106 4106 is. A local finger indicates how far into the region a task has
4107 4107 scanned. The two fingers are used to determine how to gray an
4108 4108 object (i.e. whether simply marking it is OK, as it will be
4109 4109 visited by a task in the future, or whether it needs to be also
4110 4110 pushed on a stack).
4111 4111
4112 4112 (2) Local Queue. The local queue of the task which is accessed
4113 4113 reasonably efficiently by the task. Other tasks can steal from
4114 4114 it when they run out of work. Throughout the marking phase, a
4115 4115 task attempts to keep its local queue short but not totally
4116 4116 empty, so that entries are available for stealing by other
4117 4117 tasks. Only when there is no more work, a task will totally
4118 4118 drain its local queue.
4119 4119
4120 4120 (3) Global Mark Stack. This handles local queue overflow. During
4121 4121 marking only sets of entries are moved between it and the local
4122 4122 queues, as access to it requires a mutex and more fine-grain
4123 4123 interaction with it which might cause contention. If it
4124 4124 overflows, then the marking phase should restart and iterate
4125 4125 over the bitmap to identify gray objects. Throughout the marking
4126 4126 phase, tasks attempt to keep the global mark stack at a small
4127 4127 length but not totally empty, so that entries are available for
4128 4128 popping by other tasks. Only when there is no more work, tasks
4129 4129 will totally drain the global mark stack.
4130 4130
4131 4131 (4) SATB Buffer Queue. This is where completed SATB buffers are
4132 4132 made available. Buffers are regularly removed from this queue
4133 4133 and scanned for roots, so that the queue doesn't get too
4134 4134 long. During remark, all completed buffers are processed, as
4135 4135 well as the filled in parts of any uncompleted buffers.
4136 4136
4137 4137 The do_marking_step() method tries to abort when the time target
4138 4138 has been reached. There are a few other cases when the
4139 4139 do_marking_step() method also aborts:
4140 4140
4141 4141 (1) When the marking phase has been aborted (after a Full GC).
4142 4142
4143 4143 (2) When a global overflow (on the global stack) has been
4144 4144 triggered. Before the task aborts, it will actually sync up with
4145 4145 the other tasks to ensure that all the marking data structures
4146 4146 (local queues, stacks, fingers etc.) are re-initialized so that
4147 4147 when do_marking_step() completes, the marking phase can
4148 4148 immediately restart.
4149 4149
4150 4150 (3) When enough completed SATB buffers are available. The
4151 4151 do_marking_step() method only tries to drain SATB buffers right
4152 4152 at the beginning. So, if enough buffers are available, the
4153 4153 marking step aborts and the SATB buffers are processed at
4154 4154 the beginning of the next invocation.
4155 4155
4156 4156 (4) To yield. when we have to yield then we abort and yield
4157 4157 right at the end of do_marking_step(). This saves us from a lot
4158 4158 of hassle as, by yielding we might allow a Full GC. If this
4159 4159 happens then objects will be compacted underneath our feet, the
4160 4160 heap might shrink, etc. We save checking for this by just
4161 4161 aborting and doing the yield right at the end.
4162 4162
4163 4163 From the above it follows that the do_marking_step() method should
4164 4164 be called in a loop (or, otherwise, regularly) until it completes.
4165 4165
4166 4166 If a marking step completes without its has_aborted() flag being
4167 4167 true, it means it has completed the current marking phase (and
4168 4168 also all other marking tasks have done so and have all synced up).
4169 4169
4170 4170 A method called regular_clock_call() is invoked "regularly" (in
4171 4171 sub ms intervals) throughout marking. It is this clock method that
4172 4172 checks all the abort conditions which were mentioned above and
4173 4173 decides when the task should abort. A work-based scheme is used to
4174 4174 trigger this clock method: when the number of object words the
4175 4175 marking phase has scanned or the number of references the marking
4176 4176 phase has visited reach a given limit. Additional invocations to
4177 4177 the method clock have been planted in a few other strategic places
4178 4178 too. The initial reason for the clock method was to avoid calling
4179 4179 vtime too regularly, as it is quite expensive. So, once it was in
4180 4180 place, it was natural to piggy-back all the other conditions on it
4181 4181 too and not constantly check them throughout the code.
4182 4182
4183 4183 If do_termination is true then do_marking_step will enter its
4184 4184 termination protocol.
4185 4185
4186 4186 The value of is_serial must be true when do_marking_step is being
4187 4187 called serially (i.e. by the VMThread) and do_marking_step should
4188 4188 skip any synchronization in the termination and overflow code.
4189 4189 Examples include the serial remark code and the serial reference
4190 4190 processing closures.
4191 4191
4192 4192 The value of is_serial must be false when do_marking_step is
4193 4193 being called by any of the worker threads in a work gang.
4194 4194 Examples include the concurrent marking code (CMMarkingTask),
4195 4195 the MT remark code, and the MT reference processing closures.
4196 4196
4197 4197 *****************************************************************************/
4198 4198
4199 4199 void CMTask::do_marking_step(double time_target_ms,
4200 4200 bool do_termination,
4201 4201 bool is_serial) {
4202 4202 assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4203 4203 assert(concurrent() == _cm->concurrent(), "they should be the same");
4204 4204
4205 4205 G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4206 4206 assert(_task_queues != NULL, "invariant");
4207 4207 assert(_task_queue != NULL, "invariant");
4208 4208 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4209 4209
4210 4210 assert(!_claimed,
4211 4211 "only one thread should claim this task at any one time");
4212 4212
4213 4213 // OK, this doesn't safeguard again all possible scenarios, as it is
4214 4214 // possible for two threads to set the _claimed flag at the same
4215 4215 // time. But it is only for debugging purposes anyway and it will
4216 4216 // catch most problems.
4217 4217 _claimed = true;
4218 4218
4219 4219 _start_time_ms = os::elapsedVTime() * 1000.0;
4220 4220 statsOnly( _interval_start_time_ms = _start_time_ms );
4221 4221
4222 4222 // If do_stealing is true then do_marking_step will attempt to
4223 4223 // steal work from the other CMTasks. It only makes sense to
4224 4224 // enable stealing when the termination protocol is enabled
4225 4225 // and do_marking_step() is not being called serially.
4226 4226 bool do_stealing = do_termination && !is_serial;
4227 4227
4228 4228 double diff_prediction_ms =
4229 4229 g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4230 4230 _time_target_ms = time_target_ms - diff_prediction_ms;
4231 4231
4232 4232 // set up the variables that are used in the work-based scheme to
4233 4233 // call the regular clock method
4234 4234 _words_scanned = 0;
4235 4235 _refs_reached = 0;
4236 4236 recalculate_limits();
4237 4237
4238 4238 // clear all flags
4239 4239 clear_has_aborted();
4240 4240 _has_timed_out = false;
4241 4241 _draining_satb_buffers = false;
4242 4242
4243 4243 ++_calls;
4244 4244
4245 4245 if (_cm->verbose_low()) {
4246 4246 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4247 4247 "target = %1.2lfms >>>>>>>>>>",
4248 4248 _worker_id, _calls, _time_target_ms);
4249 4249 }
4250 4250
4251 4251 // Set up the bitmap and oop closures. Anything that uses them is
4252 4252 // eventually called from this method, so it is OK to allocate these
4253 4253 // statically.
4254 4254 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4255 4255 G1CMOopClosure cm_oop_closure(_g1h, _cm, this);
4256 4256 set_cm_oop_closure(&cm_oop_closure);
4257 4257
4258 4258 if (_cm->has_overflown()) {
4259 4259 // This can happen if the mark stack overflows during a GC pause
4260 4260 // and this task, after a yield point, restarts. We have to abort
4261 4261 // as we need to get into the overflow protocol which happens
4262 4262 // right at the end of this task.
4263 4263 set_has_aborted();
4264 4264 }
4265 4265
4266 4266 // First drain any available SATB buffers. After this, we will not
4267 4267 // look at SATB buffers before the next invocation of this method.
4268 4268 // If enough completed SATB buffers are queued up, the regular clock
4269 4269 // will abort this task so that it restarts.
4270 4270 drain_satb_buffers();
4271 4271 // ...then partially drain the local queue and the global stack
4272 4272 drain_local_queue(true);
4273 4273 drain_global_stack(true);
4274 4274
4275 4275 do {
4276 4276 if (!has_aborted() && _curr_region != NULL) {
4277 4277 // This means that we're already holding on to a region.
4278 4278 assert(_finger != NULL, "if region is not NULL, then the finger "
4279 4279 "should not be NULL either");
4280 4280
4281 4281 // We might have restarted this task after an evacuation pause
4282 4282 // which might have evacuated the region we're holding on to
4283 4283 // underneath our feet. Let's read its limit again to make sure
4284 4284 // that we do not iterate over a region of the heap that
4285 4285 // contains garbage (update_region_limit() will also move
4286 4286 // _finger to the start of the region if it is found empty).
4287 4287 update_region_limit();
4288 4288 // We will start from _finger not from the start of the region,
4289 4289 // as we might be restarting this task after aborting half-way
4290 4290 // through scanning this region. In this case, _finger points to
4291 4291 // the address where we last found a marked object. If this is a
4292 4292 // fresh region, _finger points to start().
4293 4293 MemRegion mr = MemRegion(_finger, _region_limit);
4294 4294
4295 4295 if (_cm->verbose_low()) {
4296 4296 gclog_or_tty->print_cr("[%u] we're scanning part "
4297 4297 "["PTR_FORMAT", "PTR_FORMAT") "
4298 4298 "of region "HR_FORMAT,
4299 4299 _worker_id, p2i(_finger), p2i(_region_limit),
4300 4300 HR_FORMAT_PARAMS(_curr_region));
4301 4301 }
4302 4302
4303 4303 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4304 4304 "humongous regions should go around loop once only");
4305 4305
4306 4306 // Some special cases:
4307 4307 // If the memory region is empty, we can just give up the region.
4308 4308 // If the current region is humongous then we only need to check
4309 4309 // the bitmap for the bit associated with the start of the object,
4310 4310 // scan the object if it's live, and give up the region.
4311 4311 // Otherwise, let's iterate over the bitmap of the part of the region
4312 4312 // that is left.
4313 4313 // If the iteration is successful, give up the region.
4314 4314 if (mr.is_empty()) {
4315 4315 giveup_current_region();
4316 4316 regular_clock_call();
4317 4317 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4318 4318 if (_nextMarkBitMap->isMarked(mr.start())) {
4319 4319 // The object is marked - apply the closure
4320 4320 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4321 4321 bitmap_closure.do_bit(offset);
4322 4322 }
4323 4323 // Even if this task aborted while scanning the humongous object
4324 4324 // we can (and should) give up the current region.
4325 4325 giveup_current_region();
4326 4326 regular_clock_call();
4327 4327 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4328 4328 giveup_current_region();
4329 4329 regular_clock_call();
4330 4330 } else {
4331 4331 assert(has_aborted(), "currently the only way to do so");
4332 4332 // The only way to abort the bitmap iteration is to return
4333 4333 // false from the do_bit() method. However, inside the
4334 4334 // do_bit() method we move the _finger to point to the
4335 4335 // object currently being looked at. So, if we bail out, we
4336 4336 // have definitely set _finger to something non-null.
4337 4337 assert(_finger != NULL, "invariant");
4338 4338
4339 4339 // Region iteration was actually aborted. So now _finger
4340 4340 // points to the address of the object we last scanned. If we
4341 4341 // leave it there, when we restart this task, we will rescan
4342 4342 // the object. It is easy to avoid this. We move the finger by
4343 4343 // enough to point to the next possible object header (the
4344 4344 // bitmap knows by how much we need to move it as it knows its
4345 4345 // granularity).
4346 4346 assert(_finger < _region_limit, "invariant");
4347 4347 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4348 4348 // Check if bitmap iteration was aborted while scanning the last object
4349 4349 if (new_finger >= _region_limit) {
4350 4350 giveup_current_region();
4351 4351 } else {
4352 4352 move_finger_to(new_finger);
4353 4353 }
4354 4354 }
4355 4355 }
4356 4356 // At this point we have either completed iterating over the
4357 4357 // region we were holding on to, or we have aborted.
4358 4358
4359 4359 // We then partially drain the local queue and the global stack.
4360 4360 // (Do we really need this?)
4361 4361 drain_local_queue(true);
4362 4362 drain_global_stack(true);
4363 4363
4364 4364 // Read the note on the claim_region() method on why it might
4365 4365 // return NULL with potentially more regions available for
4366 4366 // claiming and why we have to check out_of_regions() to determine
4367 4367 // whether we're done or not.
4368 4368 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4369 4369 // We are going to try to claim a new region. We should have
4370 4370 // given up on the previous one.
4371 4371 // Separated the asserts so that we know which one fires.
4372 4372 assert(_curr_region == NULL, "invariant");
4373 4373 assert(_finger == NULL, "invariant");
4374 4374 assert(_region_limit == NULL, "invariant");
4375 4375 if (_cm->verbose_low()) {
4376 4376 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4377 4377 }
4378 4378 HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4379 4379 if (claimed_region != NULL) {
4380 4380 // Yes, we managed to claim one
4381 4381 statsOnly( ++_regions_claimed );
4382 4382
4383 4383 if (_cm->verbose_low()) {
4384 4384 gclog_or_tty->print_cr("[%u] we successfully claimed "
4385 4385 "region "PTR_FORMAT,
4386 4386 _worker_id, p2i(claimed_region));
4387 4387 }
4388 4388
4389 4389 setup_for_region(claimed_region);
4390 4390 assert(_curr_region == claimed_region, "invariant");
4391 4391 }
4392 4392 // It is important to call the regular clock here. It might take
4393 4393 // a while to claim a region if, for example, we hit a large
4394 4394 // block of empty regions. So we need to call the regular clock
4395 4395 // method once round the loop to make sure it's called
4396 4396 // frequently enough.
4397 4397 regular_clock_call();
4398 4398 }
4399 4399
4400 4400 if (!has_aborted() && _curr_region == NULL) {
4401 4401 assert(_cm->out_of_regions(),
4402 4402 "at this point we should be out of regions");
4403 4403 }
4404 4404 } while ( _curr_region != NULL && !has_aborted());
4405 4405
4406 4406 if (!has_aborted()) {
4407 4407 // We cannot check whether the global stack is empty, since other
4408 4408 // tasks might be pushing objects to it concurrently.
4409 4409 assert(_cm->out_of_regions(),
4410 4410 "at this point we should be out of regions");
4411 4411
4412 4412 if (_cm->verbose_low()) {
4413 4413 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4414 4414 }
4415 4415
4416 4416 // Try to reduce the number of available SATB buffers so that
4417 4417 // remark has less work to do.
4418 4418 drain_satb_buffers();
4419 4419 }
4420 4420
4421 4421 // Since we've done everything else, we can now totally drain the
4422 4422 // local queue and global stack.
4423 4423 drain_local_queue(false);
4424 4424 drain_global_stack(false);
4425 4425
4426 4426 // Attempt at work stealing from other task's queues.
4427 4427 if (do_stealing && !has_aborted()) {
4428 4428 // We have not aborted. This means that we have finished all that
4429 4429 // we could. Let's try to do some stealing...
4430 4430
4431 4431 // We cannot check whether the global stack is empty, since other
4432 4432 // tasks might be pushing objects to it concurrently.
4433 4433 assert(_cm->out_of_regions() && _task_queue->size() == 0,
4434 4434 "only way to reach here");
4435 4435
4436 4436 if (_cm->verbose_low()) {
4437 4437 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4438 4438 }
4439 4439
4440 4440 while (!has_aborted()) {
4441 4441 oop obj;
4442 4442 statsOnly( ++_steal_attempts );
4443 4443
4444 4444 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4445 4445 if (_cm->verbose_medium()) {
4446 4446 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4447 4447 _worker_id, p2i((void*) obj));
4448 4448 }
4449 4449
4450 4450 statsOnly( ++_steals );
4451 4451
4452 4452 assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4453 4453 "any stolen object should be marked");
4454 4454 scan_object(obj);
4455 4455
4456 4456 // And since we're towards the end, let's totally drain the
4457 4457 // local queue and global stack.
4458 4458 drain_local_queue(false);
4459 4459 drain_global_stack(false);
4460 4460 } else {
4461 4461 break;
4462 4462 }
4463 4463 }
4464 4464 }
4465 4465
4466 4466 // If we are about to wrap up and go into termination, check if we
4467 4467 // should raise the overflow flag.
4468 4468 if (do_termination && !has_aborted()) {
4469 4469 if (_cm->force_overflow()->should_force()) {
4470 4470 _cm->set_has_overflown();
4471 4471 regular_clock_call();
4472 4472 }
4473 4473 }
4474 4474
4475 4475 // We still haven't aborted. Now, let's try to get into the
4476 4476 // termination protocol.
4477 4477 if (do_termination && !has_aborted()) {
4478 4478 // We cannot check whether the global stack is empty, since other
4479 4479 // tasks might be concurrently pushing objects on it.
4480 4480 // Separated the asserts so that we know which one fires.
4481 4481 assert(_cm->out_of_regions(), "only way to reach here");
4482 4482 assert(_task_queue->size() == 0, "only way to reach here");
4483 4483
4484 4484 if (_cm->verbose_low()) {
4485 4485 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4486 4486 }
4487 4487
4488 4488 _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4489 4489
4490 4490 // The CMTask class also extends the TerminatorTerminator class,
4491 4491 // hence its should_exit_termination() method will also decide
4492 4492 // whether to exit the termination protocol or not.
4493 4493 bool finished = (is_serial ||
4494 4494 _cm->terminator()->offer_termination(this));
4495 4495 double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4496 4496 _termination_time_ms +=
4497 4497 termination_end_time_ms - _termination_start_time_ms;
4498 4498
4499 4499 if (finished) {
4500 4500 // We're all done.
4501 4501
4502 4502 if (_worker_id == 0) {
4503 4503 // let's allow task 0 to do this
4504 4504 if (concurrent()) {
4505 4505 assert(_cm->concurrent_marking_in_progress(), "invariant");
4506 4506 // we need to set this to false before the next
4507 4507 // safepoint. This way we ensure that the marking phase
4508 4508 // doesn't observe any more heap expansions.
4509 4509 _cm->clear_concurrent_marking_in_progress();
4510 4510 }
4511 4511 }
4512 4512
4513 4513 // We can now guarantee that the global stack is empty, since
4514 4514 // all other tasks have finished. We separated the guarantees so
4515 4515 // that, if a condition is false, we can immediately find out
4516 4516 // which one.
4517 4517 guarantee(_cm->out_of_regions(), "only way to reach here");
4518 4518 guarantee(_cm->mark_stack_empty(), "only way to reach here");
4519 4519 guarantee(_task_queue->size() == 0, "only way to reach here");
4520 4520 guarantee(!_cm->has_overflown(), "only way to reach here");
4521 4521 guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4522 4522
4523 4523 if (_cm->verbose_low()) {
4524 4524 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4525 4525 }
4526 4526 } else {
4527 4527 // Apparently there's more work to do. Let's abort this task. It
4528 4528 // will restart it and we can hopefully find more things to do.
4529 4529
4530 4530 if (_cm->verbose_low()) {
4531 4531 gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4532 4532 _worker_id);
4533 4533 }
4534 4534
4535 4535 set_has_aborted();
4536 4536 statsOnly( ++_aborted_termination );
4537 4537 }
4538 4538 }
4539 4539
4540 4540 // Mainly for debugging purposes to make sure that a pointer to the
4541 4541 // closure which was statically allocated in this frame doesn't
4542 4542 // escape it by accident.
4543 4543 set_cm_oop_closure(NULL);
4544 4544 double end_time_ms = os::elapsedVTime() * 1000.0;
4545 4545 double elapsed_time_ms = end_time_ms - _start_time_ms;
4546 4546 // Update the step history.
4547 4547 _step_times_ms.add(elapsed_time_ms);
4548 4548
4549 4549 if (has_aborted()) {
4550 4550 // The task was aborted for some reason.
4551 4551
4552 4552 statsOnly( ++_aborted );
4553 4553
4554 4554 if (_has_timed_out) {
4555 4555 double diff_ms = elapsed_time_ms - _time_target_ms;
4556 4556 // Keep statistics of how well we did with respect to hitting
4557 4557 // our target only if we actually timed out (if we aborted for
4558 4558 // other reasons, then the results might get skewed).
4559 4559 _marking_step_diffs_ms.add(diff_ms);
4560 4560 }
4561 4561
4562 4562 if (_cm->has_overflown()) {
4563 4563 // This is the interesting one. We aborted because a global
4564 4564 // overflow was raised. This means we have to restart the
4565 4565 // marking phase and start iterating over regions. However, in
4566 4566 // order to do this we have to make sure that all tasks stop
4567 4567 // what they are doing and re-initialise in a safe manner. We
4568 4568 // will achieve this with the use of two barrier sync points.
4569 4569
4570 4570 if (_cm->verbose_low()) {
4571 4571 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4572 4572 }
4573 4573
4574 4574 if (!is_serial) {
4575 4575 // We only need to enter the sync barrier if being called
4576 4576 // from a parallel context
4577 4577 _cm->enter_first_sync_barrier(_worker_id);
4578 4578
4579 4579 // When we exit this sync barrier we know that all tasks have
4580 4580 // stopped doing marking work. So, it's now safe to
4581 4581 // re-initialise our data structures. At the end of this method,
4582 4582 // task 0 will clear the global data structures.
4583 4583 }
4584 4584
4585 4585 statsOnly( ++_aborted_overflow );
4586 4586
4587 4587 // We clear the local state of this task...
4588 4588 clear_region_fields();
4589 4589
4590 4590 if (!is_serial) {
4591 4591 // ...and enter the second barrier.
4592 4592 _cm->enter_second_sync_barrier(_worker_id);
4593 4593 }
4594 4594 // At this point, if we're during the concurrent phase of
4595 4595 // marking, everything has been re-initialized and we're
4596 4596 // ready to restart.
4597 4597 }
4598 4598
4599 4599 if (_cm->verbose_low()) {
4600 4600 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4601 4601 "elapsed = %1.2lfms <<<<<<<<<<",
4602 4602 _worker_id, _time_target_ms, elapsed_time_ms);
4603 4603 if (_cm->has_aborted()) {
4604 4604 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4605 4605 _worker_id);
4606 4606 }
4607 4607 }
4608 4608 } else {
4609 4609 if (_cm->verbose_low()) {
4610 4610 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4611 4611 "elapsed = %1.2lfms <<<<<<<<<<",
4612 4612 _worker_id, _time_target_ms, elapsed_time_ms);
4613 4613 }
4614 4614 }
4615 4615
4616 4616 _claimed = false;
4617 4617 }
4618 4618
4619 4619 CMTask::CMTask(uint worker_id,
4620 4620 ConcurrentMark* cm,
4621 4621 size_t* marked_bytes,
4622 4622 BitMap* card_bm,
4623 4623 CMTaskQueue* task_queue,
4624 4624 CMTaskQueueSet* task_queues)
4625 4625 : _g1h(G1CollectedHeap::heap()),
4626 4626 _worker_id(worker_id), _cm(cm),
4627 4627 _claimed(false),
4628 4628 _nextMarkBitMap(NULL), _hash_seed(17),
4629 4629 _task_queue(task_queue),
4630 4630 _task_queues(task_queues),
4631 4631 _cm_oop_closure(NULL),
4632 4632 _marked_bytes_array(marked_bytes),
4633 4633 _card_bm(card_bm) {
4634 4634 guarantee(task_queue != NULL, "invariant");
4635 4635 guarantee(task_queues != NULL, "invariant");
4636 4636
4637 4637 statsOnly( _clock_due_to_scanning = 0;
4638 4638 _clock_due_to_marking = 0 );
4639 4639
4640 4640 _marking_step_diffs_ms.add(0.5);
4641 4641 }
4642 4642
4643 4643 // These are formatting macros that are used below to ensure
4644 4644 // consistent formatting. The *_H_* versions are used to format the
4645 4645 // header for a particular value and they should be kept consistent
4646 4646 // with the corresponding macro. Also note that most of the macros add
4647 4647 // the necessary white space (as a prefix) which makes them a bit
4648 4648 // easier to compose.
4649 4649
4650 4650 // All the output lines are prefixed with this string to be able to
4651 4651 // identify them easily in a large log file.
4652 4652 #define G1PPRL_LINE_PREFIX "###"
4653 4653
4654 4654 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT
4655 4655 #ifdef _LP64
4656 4656 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s"
4657 4657 #else // _LP64
4658 4658 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s"
4659 4659 #endif // _LP64
4660 4660
4661 4661 // For per-region info
4662 4662 #define G1PPRL_TYPE_FORMAT " %-4s"
4663 4663 #define G1PPRL_TYPE_H_FORMAT " %4s"
4664 4664 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9)
4665 4665 #define G1PPRL_BYTE_H_FORMAT " %9s"
4666 4666 #define G1PPRL_DOUBLE_FORMAT " %14.1f"
4667 4667 #define G1PPRL_DOUBLE_H_FORMAT " %14s"
4668 4668
4669 4669 // For summary info
4670 4670 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT
4671 4671 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT
4672 4672 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB"
4673 4673 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4674 4674
4675 4675 G1PrintRegionLivenessInfoClosure::
4676 4676 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4677 4677 : _out(out),
4678 4678 _total_used_bytes(0), _total_capacity_bytes(0),
4679 4679 _total_prev_live_bytes(0), _total_next_live_bytes(0),
4680 4680 _hum_used_bytes(0), _hum_capacity_bytes(0),
4681 4681 _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4682 4682 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4683 4683 G1CollectedHeap* g1h = G1CollectedHeap::heap();
4684 4684 MemRegion g1_reserved = g1h->g1_reserved();
4685 4685 double now = os::elapsedTime();
4686 4686
4687 4687 // Print the header of the output.
4688 4688 _out->cr();
4689 4689 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4690 4690 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4691 4691 G1PPRL_SUM_ADDR_FORMAT("reserved")
4692 4692 G1PPRL_SUM_BYTE_FORMAT("region-size"),
4693 4693 p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4694 4694 HeapRegion::GrainBytes);
4695 4695 _out->print_cr(G1PPRL_LINE_PREFIX);
4696 4696 _out->print_cr(G1PPRL_LINE_PREFIX
4697 4697 G1PPRL_TYPE_H_FORMAT
4698 4698 G1PPRL_ADDR_BASE_H_FORMAT
4699 4699 G1PPRL_BYTE_H_FORMAT
4700 4700 G1PPRL_BYTE_H_FORMAT
4701 4701 G1PPRL_BYTE_H_FORMAT
4702 4702 G1PPRL_DOUBLE_H_FORMAT
4703 4703 G1PPRL_BYTE_H_FORMAT
4704 4704 G1PPRL_BYTE_H_FORMAT,
4705 4705 "type", "address-range",
4706 4706 "used", "prev-live", "next-live", "gc-eff",
4707 4707 "remset", "code-roots");
4708 4708 _out->print_cr(G1PPRL_LINE_PREFIX
4709 4709 G1PPRL_TYPE_H_FORMAT
4710 4710 G1PPRL_ADDR_BASE_H_FORMAT
4711 4711 G1PPRL_BYTE_H_FORMAT
4712 4712 G1PPRL_BYTE_H_FORMAT
4713 4713 G1PPRL_BYTE_H_FORMAT
4714 4714 G1PPRL_DOUBLE_H_FORMAT
4715 4715 G1PPRL_BYTE_H_FORMAT
4716 4716 G1PPRL_BYTE_H_FORMAT,
4717 4717 "", "",
4718 4718 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4719 4719 "(bytes)", "(bytes)");
4720 4720 }
4721 4721
4722 4722 // It takes as a parameter a reference to one of the _hum_* fields, it
4723 4723 // deduces the corresponding value for a region in a humongous region
4724 4724 // series (either the region size, or what's left if the _hum_* field
4725 4725 // is < the region size), and updates the _hum_* field accordingly.
4726 4726 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4727 4727 size_t bytes = 0;
4728 4728 // The > 0 check is to deal with the prev and next live bytes which
4729 4729 // could be 0.
4730 4730 if (*hum_bytes > 0) {
4731 4731 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4732 4732 *hum_bytes -= bytes;
4733 4733 }
4734 4734 return bytes;
4735 4735 }
4736 4736
4737 4737 // It deduces the values for a region in a humongous region series
4738 4738 // from the _hum_* fields and updates those accordingly. It assumes
4739 4739 // that that _hum_* fields have already been set up from the "starts
4740 4740 // humongous" region and we visit the regions in address order.
4741 4741 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4742 4742 size_t* capacity_bytes,
4743 4743 size_t* prev_live_bytes,
4744 4744 size_t* next_live_bytes) {
4745 4745 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4746 4746 *used_bytes = get_hum_bytes(&_hum_used_bytes);
4747 4747 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes);
4748 4748 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4749 4749 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4750 4750 }
4751 4751
4752 4752 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4753 4753 const char* type = r->get_type_str();
4754 4754 HeapWord* bottom = r->bottom();
4755 4755 HeapWord* end = r->end();
4756 4756 size_t capacity_bytes = r->capacity();
4757 4757 size_t used_bytes = r->used();
4758 4758 size_t prev_live_bytes = r->live_bytes();
4759 4759 size_t next_live_bytes = r->next_live_bytes();
4760 4760 double gc_eff = r->gc_efficiency();
4761 4761 size_t remset_bytes = r->rem_set()->mem_size();
4762 4762 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4763 4763
4764 4764 if (r->startsHumongous()) {
4765 4765 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4766 4766 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4767 4767 "they should have been zeroed after the last time we used them");
4768 4768 // Set up the _hum_* fields.
4769 4769 _hum_capacity_bytes = capacity_bytes;
4770 4770 _hum_used_bytes = used_bytes;
4771 4771 _hum_prev_live_bytes = prev_live_bytes;
4772 4772 _hum_next_live_bytes = next_live_bytes;
4773 4773 get_hum_bytes(&used_bytes, &capacity_bytes,
4774 4774 &prev_live_bytes, &next_live_bytes);
4775 4775 end = bottom + HeapRegion::GrainWords;
4776 4776 } else if (r->continuesHumongous()) {
4777 4777 get_hum_bytes(&used_bytes, &capacity_bytes,
4778 4778 &prev_live_bytes, &next_live_bytes);
4779 4779 assert(end == bottom + HeapRegion::GrainWords, "invariant");
4780 4780 }
4781 4781
4782 4782 _total_used_bytes += used_bytes;
4783 4783 _total_capacity_bytes += capacity_bytes;
4784 4784 _total_prev_live_bytes += prev_live_bytes;
4785 4785 _total_next_live_bytes += next_live_bytes;
4786 4786 _total_remset_bytes += remset_bytes;
4787 4787 _total_strong_code_roots_bytes += strong_code_roots_bytes;
4788 4788
4789 4789 // Print a line for this particular region.
4790 4790 _out->print_cr(G1PPRL_LINE_PREFIX
4791 4791 G1PPRL_TYPE_FORMAT
4792 4792 G1PPRL_ADDR_BASE_FORMAT
4793 4793 G1PPRL_BYTE_FORMAT
4794 4794 G1PPRL_BYTE_FORMAT
4795 4795 G1PPRL_BYTE_FORMAT
4796 4796 G1PPRL_DOUBLE_FORMAT
4797 4797 G1PPRL_BYTE_FORMAT
4798 4798 G1PPRL_BYTE_FORMAT,
4799 4799 type, p2i(bottom), p2i(end),
4800 4800 used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4801 4801 remset_bytes, strong_code_roots_bytes);
4802 4802
4803 4803 return false;
4804 4804 }
4805 4805
4806 4806 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4807 4807 // add static memory usages to remembered set sizes
4808 4808 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4809 4809 // Print the footer of the output.
4810 4810 _out->print_cr(G1PPRL_LINE_PREFIX);
4811 4811 _out->print_cr(G1PPRL_LINE_PREFIX
4812 4812 " SUMMARY"
4813 4813 G1PPRL_SUM_MB_FORMAT("capacity")
4814 4814 G1PPRL_SUM_MB_PERC_FORMAT("used")
4815 4815 G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4816 4816 G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4817 4817 G1PPRL_SUM_MB_FORMAT("remset")
4818 4818 G1PPRL_SUM_MB_FORMAT("code-roots"),
4819 4819 bytes_to_mb(_total_capacity_bytes),
4820 4820 bytes_to_mb(_total_used_bytes),
4821 4821 perc(_total_used_bytes, _total_capacity_bytes),
4822 4822 bytes_to_mb(_total_prev_live_bytes),
4823 4823 perc(_total_prev_live_bytes, _total_capacity_bytes),
4824 4824 bytes_to_mb(_total_next_live_bytes),
4825 4825 perc(_total_next_live_bytes, _total_capacity_bytes),
4826 4826 bytes_to_mb(_total_remset_bytes),
4827 4827 bytes_to_mb(_total_strong_code_roots_bytes));
4828 4828 _out->cr();
4829 4829 }
↓ open down ↓ |
1470 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX