Mercurial > hg > truffle
annotate src/share/vm/gc_implementation/g1/concurrentMark.cpp @ 993:54b3b351d6f9
Merge
| author | jrose |
|---|---|
| date | Wed, 23 Sep 2009 23:56:15 -0700 |
| parents | 148e5441d916 e1fdf4fd34dc |
| children | 035d2e036a9b |
| rev | line source |
|---|---|
| 342 | 1 /* |
| 844 | 2 * Copyright 2001-2009 Sun Microsystems, Inc. All Rights Reserved. |
| 342 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 #include "incls/_precompiled.incl" | |
| 26 #include "incls/_concurrentMark.cpp.incl" | |
| 27 | |
| 28 // | |
| 29 // CMS Bit Map Wrapper | |
| 30 | |
| 31 CMBitMapRO::CMBitMapRO(ReservedSpace rs, int shifter): | |
| 32 _bm((uintptr_t*)NULL,0), | |
| 33 _shifter(shifter) { | |
| 34 _bmStartWord = (HeapWord*)(rs.base()); | |
| 35 _bmWordSize = rs.size()/HeapWordSize; // rs.size() is in bytes | |
| 36 ReservedSpace brs(ReservedSpace::allocation_align_size_up( | |
| 37 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); | |
| 38 | |
| 39 guarantee(brs.is_reserved(), "couldn't allocate CMS bit map"); | |
| 40 // For now we'll just commit all of the bit map up fromt. | |
| 41 // Later on we'll try to be more parsimonious with swap. | |
| 42 guarantee(_virtual_space.initialize(brs, brs.size()), | |
| 43 "couldn't reseve backing store for CMS bit map"); | |
| 44 assert(_virtual_space.committed_size() == brs.size(), | |
| 45 "didn't reserve backing store for all of CMS bit map?"); | |
| 46 _bm.set_map((uintptr_t*)_virtual_space.low()); | |
| 47 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= | |
| 48 _bmWordSize, "inconsistency in bit map sizing"); | |
| 49 _bm.set_size(_bmWordSize >> _shifter); | |
| 50 } | |
| 51 | |
| 52 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr, | |
| 53 HeapWord* limit) const { | |
| 54 // First we must round addr *up* to a possible object boundary. | |
| 55 addr = (HeapWord*)align_size_up((intptr_t)addr, | |
| 56 HeapWordSize << _shifter); | |
| 57 size_t addrOffset = heapWordToOffset(addr); | |
| 58 if (limit == NULL) limit = _bmStartWord + _bmWordSize; | |
| 59 size_t limitOffset = heapWordToOffset(limit); | |
| 60 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); | |
| 61 HeapWord* nextAddr = offsetToHeapWord(nextOffset); | |
| 62 assert(nextAddr >= addr, "get_next_one postcondition"); | |
| 63 assert(nextAddr == limit || isMarked(nextAddr), | |
| 64 "get_next_one postcondition"); | |
| 65 return nextAddr; | |
| 66 } | |
| 67 | |
| 68 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr, | |
| 69 HeapWord* limit) const { | |
| 70 size_t addrOffset = heapWordToOffset(addr); | |
| 71 if (limit == NULL) limit = _bmStartWord + _bmWordSize; | |
| 72 size_t limitOffset = heapWordToOffset(limit); | |
| 73 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset); | |
| 74 HeapWord* nextAddr = offsetToHeapWord(nextOffset); | |
| 75 assert(nextAddr >= addr, "get_next_one postcondition"); | |
| 76 assert(nextAddr == limit || !isMarked(nextAddr), | |
| 77 "get_next_one postcondition"); | |
| 78 return nextAddr; | |
| 79 } | |
| 80 | |
| 81 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const { | |
| 82 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check"); | |
| 83 return (int) (diff >> _shifter); | |
| 84 } | |
| 85 | |
| 86 bool CMBitMapRO::iterate(BitMapClosure* cl, MemRegion mr) { | |
| 87 HeapWord* left = MAX2(_bmStartWord, mr.start()); | |
| 88 HeapWord* right = MIN2(_bmStartWord + _bmWordSize, mr.end()); | |
| 89 if (right > left) { | |
| 90 // Right-open interval [leftOffset, rightOffset). | |
| 91 return _bm.iterate(cl, heapWordToOffset(left), heapWordToOffset(right)); | |
| 92 } else { | |
| 93 return true; | |
| 94 } | |
| 95 } | |
| 96 | |
| 97 void CMBitMapRO::mostly_disjoint_range_union(BitMap* from_bitmap, | |
| 98 size_t from_start_index, | |
| 99 HeapWord* to_start_word, | |
| 100 size_t word_num) { | |
| 101 _bm.mostly_disjoint_range_union(from_bitmap, | |
| 102 from_start_index, | |
| 103 heapWordToOffset(to_start_word), | |
| 104 word_num); | |
| 105 } | |
| 106 | |
| 107 #ifndef PRODUCT | |
| 108 bool CMBitMapRO::covers(ReservedSpace rs) const { | |
| 109 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); | |
|
645
c3a720eefe82
6816308: Changes to allow builds with latest Windows SDK 6.1 on 64bit Windows 2003
kvn
parents:
619
diff
changeset
|
110 assert(((size_t)_bm.size() * (size_t)(1 << _shifter)) == _bmWordSize, |
| 342 | 111 "size inconsistency"); |
| 112 return _bmStartWord == (HeapWord*)(rs.base()) && | |
| 113 _bmWordSize == rs.size()>>LogHeapWordSize; | |
| 114 } | |
| 115 #endif | |
| 116 | |
| 117 void CMBitMap::clearAll() { | |
| 118 _bm.clear(); | |
| 119 return; | |
| 120 } | |
| 121 | |
| 122 void CMBitMap::markRange(MemRegion mr) { | |
| 123 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); | |
| 124 assert(!mr.is_empty(), "unexpected empty region"); | |
| 125 assert((offsetToHeapWord(heapWordToOffset(mr.end())) == | |
| 126 ((HeapWord *) mr.end())), | |
| 127 "markRange memory region end is not card aligned"); | |
| 128 // convert address range into offset range | |
| 129 _bm.at_put_range(heapWordToOffset(mr.start()), | |
| 130 heapWordToOffset(mr.end()), true); | |
| 131 } | |
| 132 | |
| 133 void CMBitMap::clearRange(MemRegion mr) { | |
| 134 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); | |
| 135 assert(!mr.is_empty(), "unexpected empty region"); | |
| 136 // convert address range into offset range | |
| 137 _bm.at_put_range(heapWordToOffset(mr.start()), | |
| 138 heapWordToOffset(mr.end()), false); | |
| 139 } | |
| 140 | |
| 141 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, | |
| 142 HeapWord* end_addr) { | |
| 143 HeapWord* start = getNextMarkedWordAddress(addr); | |
| 144 start = MIN2(start, end_addr); | |
| 145 HeapWord* end = getNextUnmarkedWordAddress(start); | |
| 146 end = MIN2(end, end_addr); | |
| 147 assert(start <= end, "Consistency check"); | |
| 148 MemRegion mr(start, end); | |
| 149 if (!mr.is_empty()) { | |
| 150 clearRange(mr); | |
| 151 } | |
| 152 return mr; | |
| 153 } | |
| 154 | |
| 155 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : | |
| 156 _base(NULL), _cm(cm) | |
| 157 #ifdef ASSERT | |
| 158 , _drain_in_progress(false) | |
| 159 , _drain_in_progress_yields(false) | |
| 160 #endif | |
| 161 {} | |
| 162 | |
| 163 void CMMarkStack::allocate(size_t size) { | |
| 164 _base = NEW_C_HEAP_ARRAY(oop, size); | |
| 165 if (_base == NULL) | |
| 166 vm_exit_during_initialization("Failed to allocate " | |
| 167 "CM region mark stack"); | |
| 168 _index = 0; | |
| 169 // QQQQ cast ... | |
| 170 _capacity = (jint) size; | |
| 171 _oops_do_bound = -1; | |
| 172 NOT_PRODUCT(_max_depth = 0); | |
| 173 } | |
| 174 | |
| 175 CMMarkStack::~CMMarkStack() { | |
| 176 if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base); | |
| 177 } | |
| 178 | |
| 179 void CMMarkStack::par_push(oop ptr) { | |
| 180 while (true) { | |
| 181 if (isFull()) { | |
| 182 _overflow = true; | |
| 183 return; | |
| 184 } | |
| 185 // Otherwise... | |
| 186 jint index = _index; | |
| 187 jint next_index = index+1; | |
| 188 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
| 189 if (res == index) { | |
| 190 _base[index] = ptr; | |
| 191 // Note that we don't maintain this atomically. We could, but it | |
| 192 // doesn't seem necessary. | |
| 193 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); | |
| 194 return; | |
| 195 } | |
| 196 // Otherwise, we need to try again. | |
| 197 } | |
| 198 } | |
| 199 | |
| 200 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { | |
| 201 while (true) { | |
| 202 if (isFull()) { | |
| 203 _overflow = true; | |
| 204 return; | |
| 205 } | |
| 206 // Otherwise... | |
| 207 jint index = _index; | |
| 208 jint next_index = index + n; | |
| 209 if (next_index > _capacity) { | |
| 210 _overflow = true; | |
| 211 return; | |
| 212 } | |
| 213 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
| 214 if (res == index) { | |
| 215 for (int i = 0; i < n; i++) { | |
| 216 int ind = index + i; | |
| 217 assert(ind < _capacity, "By overflow test above."); | |
| 218 _base[ind] = ptr_arr[i]; | |
| 219 } | |
| 220 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); | |
| 221 return; | |
| 222 } | |
| 223 // Otherwise, we need to try again. | |
| 224 } | |
| 225 } | |
| 226 | |
| 227 | |
| 228 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { | |
| 229 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
| 230 jint start = _index; | |
| 231 jint next_index = start + n; | |
| 232 if (next_index > _capacity) { | |
| 233 _overflow = true; | |
| 234 return; | |
| 235 } | |
| 236 // Otherwise. | |
| 237 _index = next_index; | |
| 238 for (int i = 0; i < n; i++) { | |
| 239 int ind = start + i; | |
| 240 guarantee(ind < _capacity, "By overflow test above."); | |
| 241 _base[ind] = ptr_arr[i]; | |
| 242 } | |
| 243 } | |
| 244 | |
| 245 | |
| 246 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { | |
| 247 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
| 248 jint index = _index; | |
| 249 if (index == 0) { | |
| 250 *n = 0; | |
| 251 return false; | |
| 252 } else { | |
| 253 int k = MIN2(max, index); | |
| 254 jint new_ind = index - k; | |
| 255 for (int j = 0; j < k; j++) { | |
| 256 ptr_arr[j] = _base[new_ind + j]; | |
| 257 } | |
| 258 _index = new_ind; | |
| 259 *n = k; | |
| 260 return true; | |
| 261 } | |
| 262 } | |
| 263 | |
| 264 | |
| 265 CMRegionStack::CMRegionStack() : _base(NULL) {} | |
| 266 | |
| 267 void CMRegionStack::allocate(size_t size) { | |
| 268 _base = NEW_C_HEAP_ARRAY(MemRegion, size); | |
| 269 if (_base == NULL) | |
| 270 vm_exit_during_initialization("Failed to allocate " | |
| 271 "CM region mark stack"); | |
| 272 _index = 0; | |
| 273 // QQQQ cast ... | |
| 274 _capacity = (jint) size; | |
| 275 } | |
| 276 | |
| 277 CMRegionStack::~CMRegionStack() { | |
| 278 if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base); | |
| 279 } | |
| 280 | |
| 281 void CMRegionStack::push(MemRegion mr) { | |
| 282 assert(mr.word_size() > 0, "Precondition"); | |
| 283 while (true) { | |
| 284 if (isFull()) { | |
| 285 _overflow = true; | |
| 286 return; | |
| 287 } | |
| 288 // Otherwise... | |
| 289 jint index = _index; | |
| 290 jint next_index = index+1; | |
| 291 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
| 292 if (res == index) { | |
| 293 _base[index] = mr; | |
| 294 return; | |
| 295 } | |
| 296 // Otherwise, we need to try again. | |
| 297 } | |
| 298 } | |
| 299 | |
| 300 MemRegion CMRegionStack::pop() { | |
| 301 while (true) { | |
| 302 // Otherwise... | |
| 303 jint index = _index; | |
| 304 | |
| 305 if (index == 0) { | |
| 306 return MemRegion(); | |
| 307 } | |
| 308 jint next_index = index-1; | |
| 309 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
| 310 if (res == index) { | |
| 311 MemRegion mr = _base[next_index]; | |
| 312 if (mr.start() != NULL) { | |
| 313 tmp_guarantee_CM( mr.end() != NULL, "invariant" ); | |
| 314 tmp_guarantee_CM( mr.word_size() > 0, "invariant" ); | |
| 315 return mr; | |
| 316 } else { | |
| 317 // that entry was invalidated... let's skip it | |
| 318 tmp_guarantee_CM( mr.end() == NULL, "invariant" ); | |
| 319 } | |
| 320 } | |
| 321 // Otherwise, we need to try again. | |
| 322 } | |
| 323 } | |
| 324 | |
| 325 bool CMRegionStack::invalidate_entries_into_cset() { | |
| 326 bool result = false; | |
| 327 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 328 for (int i = 0; i < _oops_do_bound; ++i) { | |
| 329 MemRegion mr = _base[i]; | |
| 330 if (mr.start() != NULL) { | |
| 331 tmp_guarantee_CM( mr.end() != NULL, "invariant"); | |
| 332 tmp_guarantee_CM( mr.word_size() > 0, "invariant" ); | |
| 333 HeapRegion* hr = g1h->heap_region_containing(mr.start()); | |
| 334 tmp_guarantee_CM( hr != NULL, "invariant" ); | |
| 335 if (hr->in_collection_set()) { | |
| 336 // The region points into the collection set | |
| 337 _base[i] = MemRegion(); | |
| 338 result = true; | |
| 339 } | |
| 340 } else { | |
| 341 // that entry was invalidated... let's skip it | |
| 342 tmp_guarantee_CM( mr.end() == NULL, "invariant" ); | |
| 343 } | |
| 344 } | |
| 345 return result; | |
| 346 } | |
| 347 | |
| 348 template<class OopClosureClass> | |
| 349 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { | |
| 350 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after | |
| 351 || SafepointSynchronize::is_at_safepoint(), | |
| 352 "Drain recursion must be yield-safe."); | |
| 353 bool res = true; | |
| 354 debug_only(_drain_in_progress = true); | |
| 355 debug_only(_drain_in_progress_yields = yield_after); | |
| 356 while (!isEmpty()) { | |
| 357 oop newOop = pop(); | |
| 358 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); | |
| 359 assert(newOop->is_oop(), "Expected an oop"); | |
| 360 assert(bm == NULL || bm->isMarked((HeapWord*)newOop), | |
| 361 "only grey objects on this stack"); | |
| 362 // iterate over the oops in this oop, marking and pushing | |
| 363 // the ones in CMS generation. | |
| 364 newOop->oop_iterate(cl); | |
| 365 if (yield_after && _cm->do_yield_check()) { | |
| 366 res = false; break; | |
| 367 } | |
| 368 } | |
| 369 debug_only(_drain_in_progress = false); | |
| 370 return res; | |
| 371 } | |
| 372 | |
| 373 void CMMarkStack::oops_do(OopClosure* f) { | |
| 374 if (_index == 0) return; | |
| 375 assert(_oops_do_bound != -1 && _oops_do_bound <= _index, | |
| 376 "Bound must be set."); | |
| 377 for (int i = 0; i < _oops_do_bound; i++) { | |
| 378 f->do_oop(&_base[i]); | |
| 379 } | |
| 380 _oops_do_bound = -1; | |
| 381 } | |
| 382 | |
| 383 bool ConcurrentMark::not_yet_marked(oop obj) const { | |
| 384 return (_g1h->is_obj_ill(obj) | |
| 385 || (_g1h->is_in_permanent(obj) | |
| 386 && !nextMarkBitMap()->isMarked((HeapWord*)obj))); | |
| 387 } | |
| 388 | |
| 389 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away | |
| 390 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
| 391 #endif // _MSC_VER | |
| 392 | |
| 393 ConcurrentMark::ConcurrentMark(ReservedSpace rs, | |
| 394 int max_regions) : | |
| 395 _markBitMap1(rs, MinObjAlignment - 1), | |
| 396 _markBitMap2(rs, MinObjAlignment - 1), | |
| 397 | |
| 398 _parallel_marking_threads(0), | |
| 399 _sleep_factor(0.0), | |
| 400 _marking_task_overhead(1.0), | |
| 401 _cleanup_sleep_factor(0.0), | |
| 402 _cleanup_task_overhead(1.0), | |
| 403 _region_bm(max_regions, false /* in_resource_area*/), | |
| 404 _card_bm((rs.size() + CardTableModRefBS::card_size - 1) >> | |
| 405 CardTableModRefBS::card_shift, | |
| 406 false /* in_resource_area*/), | |
| 407 _prevMarkBitMap(&_markBitMap1), | |
| 408 _nextMarkBitMap(&_markBitMap2), | |
| 409 _at_least_one_mark_complete(false), | |
| 410 | |
| 411 _markStack(this), | |
| 412 _regionStack(), | |
| 413 // _finger set in set_non_marking_state | |
| 414 | |
| 415 _max_task_num(MAX2(ParallelGCThreads, (size_t)1)), | |
| 416 // _active_tasks set in set_non_marking_state | |
| 417 // _tasks set inside the constructor | |
| 418 _task_queues(new CMTaskQueueSet((int) _max_task_num)), | |
| 419 _terminator(ParallelTaskTerminator((int) _max_task_num, _task_queues)), | |
| 420 | |
| 421 _has_overflown(false), | |
| 422 _concurrent(false), | |
|
619
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
423 _has_aborted(false), |
|
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
424 _restart_for_overflow(false), |
|
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
425 _concurrent_marking_in_progress(false), |
|
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
426 _should_gray_objects(false), |
| 342 | 427 |
| 428 // _verbose_level set below | |
| 429 | |
| 430 _init_times(), | |
| 431 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), | |
| 432 _cleanup_times(), | |
| 433 _total_counting_time(0.0), | |
| 434 _total_rs_scrub_time(0.0), | |
| 435 | |
| 936 | 436 _parallel_workers(NULL) |
| 342 | 437 { |
| 438 CMVerboseLevel verbose_level = | |
| 439 (CMVerboseLevel) G1MarkingVerboseLevel; | |
| 440 if (verbose_level < no_verbose) | |
| 441 verbose_level = no_verbose; | |
| 442 if (verbose_level > high_verbose) | |
| 443 verbose_level = high_verbose; | |
| 444 _verbose_level = verbose_level; | |
| 445 | |
| 446 if (verbose_low()) | |
| 447 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " | |
| 448 "heap end = "PTR_FORMAT, _heap_start, _heap_end); | |
| 449 | |
| 751 | 450 _markStack.allocate(G1MarkStackSize); |
| 451 _regionStack.allocate(G1MarkRegionStackSize); | |
| 342 | 452 |
| 453 // Create & start a ConcurrentMark thread. | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
454 _cmThread = new ConcurrentMarkThread(this); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
455 assert(cmThread() != NULL, "CM Thread should have been created"); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
456 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
457 |
| 342 | 458 _g1h = G1CollectedHeap::heap(); |
| 459 assert(CGC_lock != NULL, "Where's the CGC_lock?"); | |
| 460 assert(_markBitMap1.covers(rs), "_markBitMap1 inconsistency"); | |
| 461 assert(_markBitMap2.covers(rs), "_markBitMap2 inconsistency"); | |
| 462 | |
| 463 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); | |
| 464 satb_qs.set_buffer_size(G1SATBLogBufferSize); | |
| 465 | |
| 466 int size = (int) MAX2(ParallelGCThreads, (size_t)1); | |
| 467 _par_cleanup_thread_state = NEW_C_HEAP_ARRAY(ParCleanupThreadState*, size); | |
| 468 for (int i = 0 ; i < size; i++) { | |
| 469 _par_cleanup_thread_state[i] = new ParCleanupThreadState; | |
| 470 } | |
| 471 | |
| 472 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_task_num); | |
| 473 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_task_num); | |
| 474 | |
| 475 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail | |
| 476 _active_tasks = _max_task_num; | |
| 477 for (int i = 0; i < (int) _max_task_num; ++i) { | |
| 478 CMTaskQueue* task_queue = new CMTaskQueue(); | |
| 479 task_queue->initialize(); | |
| 480 _task_queues->register_queue(i, task_queue); | |
| 481 | |
| 482 _tasks[i] = new CMTask(i, this, task_queue, _task_queues); | |
| 483 _accum_task_vtime[i] = 0.0; | |
| 484 } | |
| 485 | |
| 486 if (ParallelMarkingThreads > ParallelGCThreads) { | |
| 487 vm_exit_during_initialization("Can't have more ParallelMarkingThreads " | |
| 488 "than ParallelGCThreads."); | |
| 489 } | |
| 490 if (ParallelGCThreads == 0) { | |
| 491 // if we are not running with any parallel GC threads we will not | |
| 492 // spawn any marking threads either | |
| 493 _parallel_marking_threads = 0; | |
| 494 _sleep_factor = 0.0; | |
| 495 _marking_task_overhead = 1.0; | |
| 496 } else { | |
| 497 if (ParallelMarkingThreads > 0) { | |
| 751 | 498 // notice that ParallelMarkingThreads overwrites G1MarkingOverheadPercent |
| 342 | 499 // if both are set |
| 500 | |
| 501 _parallel_marking_threads = ParallelMarkingThreads; | |
| 502 _sleep_factor = 0.0; | |
| 503 _marking_task_overhead = 1.0; | |
| 751 | 504 } else if (G1MarkingOverheadPercent > 0) { |
| 342 | 505 // we will calculate the number of parallel marking threads |
| 506 // based on a target overhead with respect to the soft real-time | |
| 507 // goal | |
| 508 | |
| 751 | 509 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; |
| 342 | 510 double overall_cm_overhead = |
| 751 | 511 (double) MaxGCPauseMillis * marking_overhead / |
| 512 (double) GCPauseIntervalMillis; | |
| 342 | 513 double cpu_ratio = 1.0 / (double) os::processor_count(); |
| 514 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); | |
| 515 double marking_task_overhead = | |
| 516 overall_cm_overhead / marking_thread_num * | |
| 517 (double) os::processor_count(); | |
| 518 double sleep_factor = | |
| 519 (1.0 - marking_task_overhead) / marking_task_overhead; | |
| 520 | |
| 521 _parallel_marking_threads = (size_t) marking_thread_num; | |
| 522 _sleep_factor = sleep_factor; | |
| 523 _marking_task_overhead = marking_task_overhead; | |
| 524 } else { | |
| 525 _parallel_marking_threads = MAX2((ParallelGCThreads + 2) / 4, (size_t)1); | |
| 526 _sleep_factor = 0.0; | |
| 527 _marking_task_overhead = 1.0; | |
| 528 } | |
| 529 | |
| 530 if (parallel_marking_threads() > 1) | |
| 531 _cleanup_task_overhead = 1.0; | |
| 532 else | |
| 533 _cleanup_task_overhead = marking_task_overhead(); | |
| 534 _cleanup_sleep_factor = | |
| 535 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); | |
| 536 | |
| 537 #if 0 | |
| 538 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); | |
| 539 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); | |
| 540 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); | |
| 541 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); | |
| 542 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); | |
| 543 #endif | |
| 544 | |
| 545 guarantee( parallel_marking_threads() > 0, "peace of mind" ); | |
| 546 _parallel_workers = new WorkGang("Parallel Marking Threads", | |
| 547 (int) parallel_marking_threads(), false, true); | |
| 548 if (_parallel_workers == NULL) | |
| 549 vm_exit_during_initialization("Failed necessary allocation."); | |
| 550 } | |
| 551 | |
| 552 // so that the call below can read a sensible value | |
| 553 _heap_start = (HeapWord*) rs.base(); | |
| 554 set_non_marking_state(); | |
| 555 } | |
| 556 | |
| 557 void ConcurrentMark::update_g1_committed(bool force) { | |
| 558 // If concurrent marking is not in progress, then we do not need to | |
| 559 // update _heap_end. This has a subtle and important | |
| 560 // side-effect. Imagine that two evacuation pauses happen between | |
| 561 // marking completion and remark. The first one can grow the | |
| 562 // heap (hence now the finger is below the heap end). Then, the | |
| 563 // second one could unnecessarily push regions on the region | |
| 564 // stack. This causes the invariant that the region stack is empty | |
| 565 // at the beginning of remark to be false. By ensuring that we do | |
| 566 // not observe heap expansions after marking is complete, then we do | |
| 567 // not have this problem. | |
| 568 if (!concurrent_marking_in_progress() && !force) | |
| 569 return; | |
| 570 | |
| 571 MemRegion committed = _g1h->g1_committed(); | |
| 572 tmp_guarantee_CM( committed.start() == _heap_start, | |
| 573 "start shouldn't change" ); | |
| 574 HeapWord* new_end = committed.end(); | |
| 575 if (new_end > _heap_end) { | |
| 576 // The heap has been expanded. | |
| 577 | |
| 578 _heap_end = new_end; | |
| 579 } | |
| 580 // Notice that the heap can also shrink. However, this only happens | |
| 581 // during a Full GC (at least currently) and the entire marking | |
| 582 // phase will bail out and the task will not be restarted. So, let's | |
| 583 // do nothing. | |
| 584 } | |
| 585 | |
| 586 void ConcurrentMark::reset() { | |
| 587 // Starting values for these two. This should be called in a STW | |
| 588 // phase. CM will be notified of any future g1_committed expansions | |
| 589 // will be at the end of evacuation pauses, when tasks are | |
| 590 // inactive. | |
| 591 MemRegion committed = _g1h->g1_committed(); | |
| 592 _heap_start = committed.start(); | |
| 593 _heap_end = committed.end(); | |
| 594 | |
| 595 guarantee( _heap_start != NULL && | |
| 596 _heap_end != NULL && | |
| 597 _heap_start < _heap_end, "heap bounds should look ok" ); | |
| 598 | |
| 599 // reset all the marking data structures and any necessary flags | |
| 600 clear_marking_state(); | |
| 601 | |
| 602 if (verbose_low()) | |
| 603 gclog_or_tty->print_cr("[global] resetting"); | |
| 604 | |
| 605 // We do reset all of them, since different phases will use | |
| 606 // different number of active threads. So, it's easiest to have all | |
| 607 // of them ready. | |
| 608 for (int i = 0; i < (int) _max_task_num; ++i) | |
| 609 _tasks[i]->reset(_nextMarkBitMap); | |
| 610 | |
| 611 // we need this to make sure that the flag is on during the evac | |
| 612 // pause with initial mark piggy-backed | |
| 613 set_concurrent_marking_in_progress(); | |
| 614 } | |
| 615 | |
| 616 void ConcurrentMark::set_phase(size_t active_tasks, bool concurrent) { | |
| 617 guarantee( active_tasks <= _max_task_num, "we should not have more" ); | |
| 618 | |
| 619 _active_tasks = active_tasks; | |
| 620 // Need to update the three data structures below according to the | |
| 621 // number of active threads for this phase. | |
| 622 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); | |
| 623 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); | |
| 624 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); | |
| 625 | |
| 626 _concurrent = concurrent; | |
| 627 // We propagate this to all tasks, not just the active ones. | |
| 628 for (int i = 0; i < (int) _max_task_num; ++i) | |
| 629 _tasks[i]->set_concurrent(concurrent); | |
| 630 | |
| 631 if (concurrent) { | |
| 632 set_concurrent_marking_in_progress(); | |
| 633 } else { | |
| 634 // We currently assume that the concurrent flag has been set to | |
| 635 // false before we start remark. At this point we should also be | |
| 636 // in a STW phase. | |
| 637 guarantee( !concurrent_marking_in_progress(), "invariant" ); | |
| 638 guarantee( _finger == _heap_end, "only way to get here" ); | |
| 639 update_g1_committed(true); | |
| 640 } | |
| 641 } | |
| 642 | |
| 643 void ConcurrentMark::set_non_marking_state() { | |
| 644 // We set the global marking state to some default values when we're | |
| 645 // not doing marking. | |
| 646 clear_marking_state(); | |
| 647 _active_tasks = 0; | |
| 648 clear_concurrent_marking_in_progress(); | |
| 649 } | |
| 650 | |
| 651 ConcurrentMark::~ConcurrentMark() { | |
| 652 int size = (int) MAX2(ParallelGCThreads, (size_t)1); | |
| 653 for (int i = 0; i < size; i++) delete _par_cleanup_thread_state[i]; | |
| 654 FREE_C_HEAP_ARRAY(ParCleanupThreadState*, | |
| 655 _par_cleanup_thread_state); | |
| 656 | |
| 657 for (int i = 0; i < (int) _max_task_num; ++i) { | |
| 658 delete _task_queues->queue(i); | |
| 659 delete _tasks[i]; | |
| 660 } | |
| 661 delete _task_queues; | |
| 662 FREE_C_HEAP_ARRAY(CMTask*, _max_task_num); | |
| 663 } | |
| 664 | |
| 665 // This closure is used to mark refs into the g1 generation | |
| 666 // from external roots in the CMS bit map. | |
| 667 // Called at the first checkpoint. | |
| 668 // | |
| 669 | |
| 670 #define PRINT_REACHABLE_AT_INITIAL_MARK 0 | |
| 671 #if PRINT_REACHABLE_AT_INITIAL_MARK | |
| 672 static FILE* reachable_file = NULL; | |
| 673 | |
| 674 class PrintReachableClosure: public OopsInGenClosure { | |
| 675 CMBitMap* _bm; | |
| 676 int _level; | |
| 677 public: | |
| 678 PrintReachableClosure(CMBitMap* bm) : | |
| 679 _bm(bm), _level(0) { | |
| 680 guarantee(reachable_file != NULL, "pre-condition"); | |
| 681 } | |
| 682 void do_oop(oop* p) { | |
| 683 oop obj = *p; | |
| 684 HeapWord* obj_addr = (HeapWord*)obj; | |
| 685 if (obj == NULL) return; | |
| 686 fprintf(reachable_file, "%d: "PTR_FORMAT" -> "PTR_FORMAT" (%d)\n", | |
| 687 _level, p, (void*) obj, _bm->isMarked(obj_addr)); | |
| 688 if (!_bm->isMarked(obj_addr)) { | |
| 689 _bm->mark(obj_addr); | |
| 690 _level++; | |
| 691 obj->oop_iterate(this); | |
| 692 _level--; | |
| 693 } | |
| 694 } | |
| 695 }; | |
| 696 #endif // PRINT_REACHABLE_AT_INITIAL_MARK | |
| 697 | |
| 698 #define SEND_HEAP_DUMP_TO_FILE 0 | |
| 699 #if SEND_HEAP_DUMP_TO_FILE | |
| 700 static FILE* heap_dump_file = NULL; | |
| 701 #endif // SEND_HEAP_DUMP_TO_FILE | |
| 702 | |
| 703 void ConcurrentMark::clearNextBitmap() { | |
| 704 guarantee(!G1CollectedHeap::heap()->mark_in_progress(), "Precondition."); | |
| 705 | |
| 706 // clear the mark bitmap (no grey objects to start with). | |
| 707 // We need to do this in chunks and offer to yield in between | |
| 708 // each chunk. | |
| 709 HeapWord* start = _nextMarkBitMap->startWord(); | |
| 710 HeapWord* end = _nextMarkBitMap->endWord(); | |
| 711 HeapWord* cur = start; | |
| 712 size_t chunkSize = M; | |
| 713 while (cur < end) { | |
| 714 HeapWord* next = cur + chunkSize; | |
| 715 if (next > end) | |
| 716 next = end; | |
| 717 MemRegion mr(cur,next); | |
| 718 _nextMarkBitMap->clearRange(mr); | |
| 719 cur = next; | |
| 720 do_yield_check(); | |
| 721 } | |
| 722 } | |
| 723 | |
| 724 class NoteStartOfMarkHRClosure: public HeapRegionClosure { | |
| 725 public: | |
| 726 bool doHeapRegion(HeapRegion* r) { | |
| 727 if (!r->continuesHumongous()) { | |
| 728 r->note_start_of_marking(true); | |
| 729 } | |
| 730 return false; | |
| 731 } | |
| 732 }; | |
| 733 | |
| 734 void ConcurrentMark::checkpointRootsInitialPre() { | |
| 735 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 736 G1CollectorPolicy* g1p = g1h->g1_policy(); | |
| 737 | |
| 738 _has_aborted = false; | |
| 739 | |
| 740 // Find all the reachable objects... | |
| 741 #if PRINT_REACHABLE_AT_INITIAL_MARK | |
| 742 guarantee(reachable_file == NULL, "Protocol"); | |
| 743 char fn_buf[100]; | |
| 744 sprintf(fn_buf, "/tmp/reachable.txt.%d", os::current_process_id()); | |
| 745 reachable_file = fopen(fn_buf, "w"); | |
| 746 // clear the mark bitmap (no grey objects to start with) | |
| 747 _nextMarkBitMap->clearAll(); | |
| 748 PrintReachableClosure prcl(_nextMarkBitMap); | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
749 g1h->process_strong_roots(true, // activate StrongRootsScope |
| 342 | 750 false, // fake perm gen collection |
| 751 SharedHeap::SO_AllClasses, | |
| 752 &prcl, // Regular roots | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
753 NULL, // do not visit active blobs |
| 342 | 754 &prcl // Perm Gen Roots |
| 755 ); | |
| 756 // The root iteration above "consumed" dirty cards in the perm gen. | |
| 757 // Therefore, as a shortcut, we dirty all such cards. | |
| 758 g1h->rem_set()->invalidate(g1h->perm_gen()->used_region(), false); | |
| 759 fclose(reachable_file); | |
| 760 reachable_file = NULL; | |
| 761 // clear the mark bitmap again. | |
| 762 _nextMarkBitMap->clearAll(); | |
| 763 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); | |
| 764 COMPILER2_PRESENT(DerivedPointerTable::clear()); | |
| 765 #endif // PRINT_REACHABLE_AT_INITIAL_MARK | |
| 766 | |
| 767 // Initialise marking structures. This has to be done in a STW phase. | |
| 768 reset(); | |
| 769 } | |
| 770 | |
| 771 class CMMarkRootsClosure: public OopsInGenClosure { | |
| 772 private: | |
| 773 ConcurrentMark* _cm; | |
| 774 G1CollectedHeap* _g1h; | |
| 775 bool _do_barrier; | |
| 776 | |
| 777 public: | |
| 778 CMMarkRootsClosure(ConcurrentMark* cm, | |
| 779 G1CollectedHeap* g1h, | |
| 780 bool do_barrier) : _cm(cm), _g1h(g1h), | |
| 781 _do_barrier(do_barrier) { } | |
| 782 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
783 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
784 virtual void do_oop( oop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
785 |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
786 template <class T> void do_oop_work(T* p) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
787 T heap_oop = oopDesc::load_heap_oop(p); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
788 if (!oopDesc::is_null(heap_oop)) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
789 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
790 assert(obj->is_oop() || obj->mark() == NULL, |
| 342 | 791 "expected an oop, possibly with mark word displaced"); |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
792 HeapWord* addr = (HeapWord*)obj; |
| 342 | 793 if (_g1h->is_in_g1_reserved(addr)) { |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
794 _cm->grayRoot(obj); |
| 342 | 795 } |
| 796 } | |
| 797 if (_do_barrier) { | |
| 798 assert(!_g1h->is_in_g1_reserved(p), | |
| 799 "Should be called on external roots"); | |
| 800 do_barrier(p); | |
| 801 } | |
| 802 } | |
| 803 }; | |
| 804 | |
| 805 void ConcurrentMark::checkpointRootsInitialPost() { | |
| 806 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 807 | |
| 808 // For each region note start of marking. | |
| 809 NoteStartOfMarkHRClosure startcl; | |
| 810 g1h->heap_region_iterate(&startcl); | |
| 811 | |
| 812 // Start weak-reference discovery. | |
| 813 ReferenceProcessor* rp = g1h->ref_processor(); | |
| 814 rp->verify_no_references_recorded(); | |
| 815 rp->enable_discovery(); // enable ("weak") refs discovery | |
|
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
816 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle |
| 342 | 817 |
| 818 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 819 satb_mq_set.set_process_completed_threshold(G1SATBProcessCompletedThreshold); | |
| 820 satb_mq_set.set_active_all_threads(true); | |
| 821 | |
| 822 // update_g1_committed() will be called at the end of an evac pause | |
| 823 // when marking is on. So, it's also called at the end of the | |
| 824 // initial-mark pause to update the heap end, if the heap expands | |
| 825 // during it. No need to call it here. | |
| 826 } | |
| 827 | |
| 828 // Checkpoint the roots into this generation from outside | |
| 829 // this generation. [Note this initial checkpoint need only | |
| 830 // be approximate -- we'll do a catch up phase subsequently.] | |
| 831 void ConcurrentMark::checkpointRootsInitial() { | |
| 832 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); | |
| 833 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 834 | |
| 835 double start = os::elapsedTime(); | |
| 836 | |
| 837 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); | |
| 838 g1p->record_concurrent_mark_init_start(); | |
| 839 checkpointRootsInitialPre(); | |
| 840 | |
| 841 // YSR: when concurrent precleaning is in place, we'll | |
| 842 // need to clear the cached card table here | |
| 843 | |
| 844 ResourceMark rm; | |
| 845 HandleMark hm; | |
| 846 | |
| 847 g1h->ensure_parsability(false); | |
| 848 g1h->perm_gen()->save_marks(); | |
| 849 | |
| 850 CMMarkRootsClosure notOlder(this, g1h, false); | |
| 851 CMMarkRootsClosure older(this, g1h, true); | |
| 852 | |
| 853 g1h->set_marking_started(); | |
| 854 g1h->rem_set()->prepare_for_younger_refs_iterate(false); | |
| 855 | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
856 g1h->process_strong_roots(true, // activate StrongRootsScope |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
857 false, // fake perm gen collection |
| 342 | 858 SharedHeap::SO_AllClasses, |
| 859 ¬Older, // Regular roots | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
860 NULL, // do not visit active blobs |
| 342 | 861 &older // Perm Gen Roots |
| 862 ); | |
| 863 checkpointRootsInitialPost(); | |
| 864 | |
| 865 // Statistics. | |
| 866 double end = os::elapsedTime(); | |
| 867 _init_times.add((end - start) * 1000.0); | |
| 868 | |
| 869 g1p->record_concurrent_mark_init_end(); | |
| 870 } | |
| 871 | |
| 872 /* | |
| 873 Notice that in the next two methods, we actually leave the STS | |
| 874 during the barrier sync and join it immediately afterwards. If we | |
| 875 do not do this, this then the following deadlock can occur: one | |
| 876 thread could be in the barrier sync code, waiting for the other | |
| 877 thread to also sync up, whereas another one could be trying to | |
| 878 yield, while also waiting for the other threads to sync up too. | |
| 879 | |
| 880 Because the thread that does the sync barrier has left the STS, it | |
| 881 is possible to be suspended for a Full GC or an evacuation pause | |
| 882 could occur. This is actually safe, since the entering the sync | |
| 883 barrier is one of the last things do_marking_step() does, and it | |
| 884 doesn't manipulate any data structures afterwards. | |
| 885 */ | |
| 886 | |
| 887 void ConcurrentMark::enter_first_sync_barrier(int task_num) { | |
| 888 if (verbose_low()) | |
| 889 gclog_or_tty->print_cr("[%d] entering first barrier", task_num); | |
| 890 | |
| 891 ConcurrentGCThread::stsLeave(); | |
| 892 _first_overflow_barrier_sync.enter(); | |
| 893 ConcurrentGCThread::stsJoin(); | |
| 894 // at this point everyone should have synced up and not be doing any | |
| 895 // more work | |
| 896 | |
| 897 if (verbose_low()) | |
| 898 gclog_or_tty->print_cr("[%d] leaving first barrier", task_num); | |
| 899 | |
| 900 // let task 0 do this | |
| 901 if (task_num == 0) { | |
| 902 // task 0 is responsible for clearing the global data structures | |
| 903 clear_marking_state(); | |
| 904 | |
| 905 if (PrintGC) { | |
| 906 gclog_or_tty->date_stamp(PrintGCDateStamps); | |
| 907 gclog_or_tty->stamp(PrintGCTimeStamps); | |
| 908 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); | |
| 909 } | |
| 910 } | |
| 911 | |
| 912 // after this, each task should reset its own data structures then | |
| 913 // then go into the second barrier | |
| 914 } | |
| 915 | |
| 916 void ConcurrentMark::enter_second_sync_barrier(int task_num) { | |
| 917 if (verbose_low()) | |
| 918 gclog_or_tty->print_cr("[%d] entering second barrier", task_num); | |
| 919 | |
| 920 ConcurrentGCThread::stsLeave(); | |
| 921 _second_overflow_barrier_sync.enter(); | |
| 922 ConcurrentGCThread::stsJoin(); | |
| 923 // at this point everything should be re-initialised and ready to go | |
| 924 | |
| 925 if (verbose_low()) | |
| 926 gclog_or_tty->print_cr("[%d] leaving second barrier", task_num); | |
| 927 } | |
| 928 | |
| 929 void ConcurrentMark::grayRoot(oop p) { | |
| 930 HeapWord* addr = (HeapWord*) p; | |
| 931 // We can't really check against _heap_start and _heap_end, since it | |
| 932 // is possible during an evacuation pause with piggy-backed | |
| 933 // initial-mark that the committed space is expanded during the | |
| 934 // pause without CM observing this change. So the assertions below | |
| 935 // is a bit conservative; but better than nothing. | |
| 936 tmp_guarantee_CM( _g1h->g1_committed().contains(addr), | |
| 937 "address should be within the heap bounds" ); | |
| 938 | |
| 939 if (!_nextMarkBitMap->isMarked(addr)) | |
| 940 _nextMarkBitMap->parMark(addr); | |
| 941 } | |
| 942 | |
| 943 void ConcurrentMark::grayRegionIfNecessary(MemRegion mr) { | |
| 944 // The objects on the region have already been marked "in bulk" by | |
| 945 // the caller. We only need to decide whether to push the region on | |
| 946 // the region stack or not. | |
| 947 | |
| 948 if (!concurrent_marking_in_progress() || !_should_gray_objects) | |
| 949 // We're done with marking and waiting for remark. We do not need to | |
| 950 // push anything else on the region stack. | |
| 951 return; | |
| 952 | |
| 953 HeapWord* finger = _finger; | |
| 954 | |
| 955 if (verbose_low()) | |
| 956 gclog_or_tty->print_cr("[global] attempting to push " | |
| 957 "region ["PTR_FORMAT", "PTR_FORMAT"), finger is at " | |
| 958 PTR_FORMAT, mr.start(), mr.end(), finger); | |
| 959 | |
| 960 if (mr.start() < finger) { | |
| 961 // The finger is always heap region aligned and it is not possible | |
| 962 // for mr to span heap regions. | |
| 963 tmp_guarantee_CM( mr.end() <= finger, "invariant" ); | |
| 964 | |
| 965 tmp_guarantee_CM( mr.start() <= mr.end() && | |
| 966 _heap_start <= mr.start() && | |
| 967 mr.end() <= _heap_end, | |
| 968 "region boundaries should fall within the committed space" ); | |
| 969 if (verbose_low()) | |
| 970 gclog_or_tty->print_cr("[global] region ["PTR_FORMAT", "PTR_FORMAT") " | |
| 971 "below the finger, pushing it", | |
| 972 mr.start(), mr.end()); | |
| 973 | |
| 974 if (!region_stack_push(mr)) { | |
| 975 if (verbose_low()) | |
| 976 gclog_or_tty->print_cr("[global] region stack has overflown."); | |
| 977 } | |
| 978 } | |
| 979 } | |
| 980 | |
| 981 void ConcurrentMark::markAndGrayObjectIfNecessary(oop p) { | |
| 982 // The object is not marked by the caller. We need to at least mark | |
| 983 // it and maybe push in on the stack. | |
| 984 | |
| 985 HeapWord* addr = (HeapWord*)p; | |
| 986 if (!_nextMarkBitMap->isMarked(addr)) { | |
| 987 // We definitely need to mark it, irrespective whether we bail out | |
| 988 // because we're done with marking. | |
| 989 if (_nextMarkBitMap->parMark(addr)) { | |
| 990 if (!concurrent_marking_in_progress() || !_should_gray_objects) | |
| 991 // If we're done with concurrent marking and we're waiting for | |
| 992 // remark, then we're not pushing anything on the stack. | |
| 993 return; | |
| 994 | |
| 995 // No OrderAccess:store_load() is needed. It is implicit in the | |
| 996 // CAS done in parMark(addr) above | |
| 997 HeapWord* finger = _finger; | |
| 998 | |
| 999 if (addr < finger) { | |
| 1000 if (!mark_stack_push(oop(addr))) { | |
| 1001 if (verbose_low()) | |
| 1002 gclog_or_tty->print_cr("[global] global stack overflow " | |
| 1003 "during parMark"); | |
| 1004 } | |
| 1005 } | |
| 1006 } | |
| 1007 } | |
| 1008 } | |
| 1009 | |
| 1010 class CMConcurrentMarkingTask: public AbstractGangTask { | |
| 1011 private: | |
| 1012 ConcurrentMark* _cm; | |
| 1013 ConcurrentMarkThread* _cmt; | |
| 1014 | |
| 1015 public: | |
| 1016 void work(int worker_i) { | |
| 1017 guarantee( Thread::current()->is_ConcurrentGC_thread(), | |
| 1018 "this should only be done by a conc GC thread" ); | |
| 1019 | |
| 1020 double start_vtime = os::elapsedVTime(); | |
| 1021 | |
| 1022 ConcurrentGCThread::stsJoin(); | |
| 1023 | |
| 1024 guarantee( (size_t)worker_i < _cm->active_tasks(), "invariant" ); | |
| 1025 CMTask* the_task = _cm->task(worker_i); | |
| 1026 the_task->record_start_time(); | |
| 1027 if (!_cm->has_aborted()) { | |
| 1028 do { | |
| 1029 double start_vtime_sec = os::elapsedVTime(); | |
| 1030 double start_time_sec = os::elapsedTime(); | |
| 1031 the_task->do_marking_step(10.0); | |
| 1032 double end_time_sec = os::elapsedTime(); | |
| 1033 double end_vtime_sec = os::elapsedVTime(); | |
| 1034 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; | |
| 1035 double elapsed_time_sec = end_time_sec - start_time_sec; | |
| 1036 _cm->clear_has_overflown(); | |
| 1037 | |
| 1038 bool ret = _cm->do_yield_check(worker_i); | |
| 1039 | |
| 1040 jlong sleep_time_ms; | |
| 1041 if (!_cm->has_aborted() && the_task->has_aborted()) { | |
| 1042 sleep_time_ms = | |
| 1043 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); | |
| 1044 ConcurrentGCThread::stsLeave(); | |
| 1045 os::sleep(Thread::current(), sleep_time_ms, false); | |
| 1046 ConcurrentGCThread::stsJoin(); | |
| 1047 } | |
| 1048 double end_time2_sec = os::elapsedTime(); | |
| 1049 double elapsed_time2_sec = end_time2_sec - start_time_sec; | |
| 1050 | |
| 1051 #if 0 | |
| 1052 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, " | |
| 1053 "overhead %1.4lf", | |
| 1054 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms, | |
| 1055 the_task->conc_overhead(os::elapsedTime()) * 8.0); | |
| 1056 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms", | |
| 1057 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0); | |
| 1058 #endif | |
| 1059 } while (!_cm->has_aborted() && the_task->has_aborted()); | |
| 1060 } | |
| 1061 the_task->record_end_time(); | |
| 1062 guarantee( !the_task->has_aborted() || _cm->has_aborted(), "invariant" ); | |
| 1063 | |
| 1064 ConcurrentGCThread::stsLeave(); | |
| 1065 | |
| 1066 double end_vtime = os::elapsedVTime(); | |
| 1067 _cm->update_accum_task_vtime(worker_i, end_vtime - start_vtime); | |
| 1068 } | |
| 1069 | |
| 1070 CMConcurrentMarkingTask(ConcurrentMark* cm, | |
| 1071 ConcurrentMarkThread* cmt) : | |
| 1072 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } | |
| 1073 | |
| 1074 ~CMConcurrentMarkingTask() { } | |
| 1075 }; | |
| 1076 | |
| 1077 void ConcurrentMark::markFromRoots() { | |
| 1078 // we might be tempted to assert that: | |
| 1079 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), | |
| 1080 // "inconsistent argument?"); | |
| 1081 // However that wouldn't be right, because it's possible that | |
| 1082 // a safepoint is indeed in progress as a younger generation | |
| 1083 // stop-the-world GC happens even as we mark in this generation. | |
| 1084 | |
| 1085 _restart_for_overflow = false; | |
| 1086 | |
| 1087 set_phase(MAX2((size_t) 1, parallel_marking_threads()), true); | |
| 1088 | |
| 1089 CMConcurrentMarkingTask markingTask(this, cmThread()); | |
| 1090 if (parallel_marking_threads() > 0) | |
| 1091 _parallel_workers->run_task(&markingTask); | |
| 1092 else | |
| 1093 markingTask.work(0); | |
| 1094 print_stats(); | |
| 1095 } | |
| 1096 | |
| 1097 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { | |
| 1098 // world is stopped at this checkpoint | |
| 1099 assert(SafepointSynchronize::is_at_safepoint(), | |
| 1100 "world should be stopped"); | |
| 1101 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 1102 | |
| 1103 // If a full collection has happened, we shouldn't do this. | |
| 1104 if (has_aborted()) { | |
| 1105 g1h->set_marking_complete(); // So bitmap clearing isn't confused | |
| 1106 return; | |
| 1107 } | |
| 1108 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1109 if (VerifyDuringGC) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1110 HandleMark hm; // handle scope |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1111 gclog_or_tty->print(" VerifyDuringGC:(before)"); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1112 Universe::heap()->prepare_for_verify(); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1113 Universe::verify(true, false, true); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1114 } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1115 |
| 342 | 1116 G1CollectorPolicy* g1p = g1h->g1_policy(); |
| 1117 g1p->record_concurrent_mark_remark_start(); | |
| 1118 | |
| 1119 double start = os::elapsedTime(); | |
| 1120 | |
| 1121 checkpointRootsFinalWork(); | |
| 1122 | |
| 1123 double mark_work_end = os::elapsedTime(); | |
| 1124 | |
| 1125 weakRefsWork(clear_all_soft_refs); | |
| 1126 | |
| 1127 if (has_overflown()) { | |
| 1128 // Oops. We overflowed. Restart concurrent marking. | |
| 1129 _restart_for_overflow = true; | |
| 1130 // Clear the flag. We do not need it any more. | |
| 1131 clear_has_overflown(); | |
| 1132 if (G1TraceMarkStackOverflow) | |
| 1133 gclog_or_tty->print_cr("\nRemark led to restart for overflow."); | |
| 1134 } else { | |
| 1135 // We're done with marking. | |
| 1136 JavaThread::satb_mark_queue_set().set_active_all_threads(false); | |
| 811 | 1137 |
| 1138 if (VerifyDuringGC) { | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1139 HandleMark hm; // handle scope |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1140 gclog_or_tty->print(" VerifyDuringGC:(after)"); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1141 Universe::heap()->prepare_for_verify(); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1142 Universe::heap()->verify(/* allow_dirty */ true, |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1143 /* silent */ false, |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1144 /* use_prev_marking */ false); |
| 811 | 1145 } |
| 342 | 1146 } |
| 1147 | |
| 1148 #if VERIFY_OBJS_PROCESSED | |
| 1149 _scan_obj_cl.objs_processed = 0; | |
| 1150 ThreadLocalObjQueue::objs_enqueued = 0; | |
| 1151 #endif | |
| 1152 | |
| 1153 // Statistics | |
| 1154 double now = os::elapsedTime(); | |
| 1155 _remark_mark_times.add((mark_work_end - start) * 1000.0); | |
| 1156 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); | |
| 1157 _remark_times.add((now - start) * 1000.0); | |
| 1158 | |
| 1159 g1p->record_concurrent_mark_remark_end(); | |
| 1160 } | |
| 1161 | |
| 1162 | |
| 1163 #define CARD_BM_TEST_MODE 0 | |
| 1164 | |
| 1165 class CalcLiveObjectsClosure: public HeapRegionClosure { | |
| 1166 | |
| 1167 CMBitMapRO* _bm; | |
| 1168 ConcurrentMark* _cm; | |
| 1169 bool _changed; | |
| 1170 bool _yield; | |
| 1171 size_t _words_done; | |
| 1172 size_t _tot_live; | |
| 1173 size_t _tot_used; | |
| 1174 size_t _regions_done; | |
| 1175 double _start_vtime_sec; | |
| 1176 | |
| 1177 BitMap* _region_bm; | |
| 1178 BitMap* _card_bm; | |
| 1179 intptr_t _bottom_card_num; | |
| 1180 bool _final; | |
| 1181 | |
| 1182 void mark_card_num_range(intptr_t start_card_num, intptr_t last_card_num) { | |
| 1183 for (intptr_t i = start_card_num; i <= last_card_num; i++) { | |
| 1184 #if CARD_BM_TEST_MODE | |
| 1185 guarantee(_card_bm->at(i - _bottom_card_num), | |
| 1186 "Should already be set."); | |
| 1187 #else | |
| 1188 _card_bm->par_at_put(i - _bottom_card_num, 1); | |
| 1189 #endif | |
| 1190 } | |
| 1191 } | |
| 1192 | |
| 1193 public: | |
| 1194 CalcLiveObjectsClosure(bool final, | |
| 1195 CMBitMapRO *bm, ConcurrentMark *cm, | |
| 936 | 1196 BitMap* region_bm, BitMap* card_bm) : |
| 342 | 1197 _bm(bm), _cm(cm), _changed(false), _yield(true), |
| 1198 _words_done(0), _tot_live(0), _tot_used(0), | |
| 936 | 1199 _region_bm(region_bm), _card_bm(card_bm),_final(final), |
| 342 | 1200 _regions_done(0), _start_vtime_sec(0.0) |
| 1201 { | |
| 1202 _bottom_card_num = | |
| 1203 intptr_t(uintptr_t(G1CollectedHeap::heap()->reserved_region().start()) >> | |
| 1204 CardTableModRefBS::card_shift); | |
| 1205 } | |
| 1206 | |
| 829 | 1207 // It takes a region that's not empty (i.e., it has at least one |
| 1208 // live object in it and sets its corresponding bit on the region | |
| 1209 // bitmap to 1. If the region is "starts humongous" it will also set | |
| 1210 // to 1 the bits on the region bitmap that correspond to its | |
| 1211 // associated "continues humongous" regions. | |
| 1212 void set_bit_for_region(HeapRegion* hr) { | |
| 1213 assert(!hr->continuesHumongous(), "should have filtered those out"); | |
| 1214 | |
| 1215 size_t index = hr->hrs_index(); | |
| 1216 if (!hr->startsHumongous()) { | |
| 1217 // Normal (non-humongous) case: just set the bit. | |
| 1218 _region_bm->par_at_put((BitMap::idx_t) index, true); | |
| 1219 } else { | |
| 1220 // Starts humongous case: calculate how many regions are part of | |
| 1221 // this humongous region and then set the bit range. It might | |
| 1222 // have been a bit more efficient to look at the object that | |
| 1223 // spans these humongous regions to calculate their number from | |
| 1224 // the object's size. However, it's a good idea to calculate | |
| 1225 // this based on the metadata itself, and not the region | |
| 1226 // contents, so that this code is not aware of what goes into | |
| 1227 // the humongous regions (in case this changes in the future). | |
| 1228 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 1229 size_t end_index = index + 1; | |
| 831 | 1230 while (end_index < g1h->n_regions()) { |
| 1231 HeapRegion* chr = g1h->region_at(end_index); | |
| 829 | 1232 if (!chr->continuesHumongous()) { |
| 1233 break; | |
| 1234 } | |
| 1235 end_index += 1; | |
| 1236 } | |
| 1237 _region_bm->par_at_put_range((BitMap::idx_t) index, | |
| 1238 (BitMap::idx_t) end_index, true); | |
| 1239 } | |
| 1240 } | |
| 1241 | |
| 342 | 1242 bool doHeapRegion(HeapRegion* hr) { |
| 1243 if (!_final && _regions_done == 0) | |
| 1244 _start_vtime_sec = os::elapsedVTime(); | |
| 1245 | |
| 639 | 1246 if (hr->continuesHumongous()) { |
| 829 | 1247 // We will ignore these here and process them when their |
| 1248 // associated "starts humongous" region is processed (see | |
| 1249 // set_bit_for_heap_region()). Note that we cannot rely on their | |
| 1250 // associated "starts humongous" region to have their bit set to | |
| 1251 // 1 since, due to the region chunking in the parallel region | |
| 1252 // iteration, a "continues humongous" region might be visited | |
| 1253 // before its associated "starts humongous". | |
| 639 | 1254 return false; |
| 1255 } | |
| 342 | 1256 |
| 1257 HeapWord* nextTop = hr->next_top_at_mark_start(); | |
| 1258 HeapWord* start = hr->top_at_conc_mark_count(); | |
| 1259 assert(hr->bottom() <= start && start <= hr->end() && | |
| 1260 hr->bottom() <= nextTop && nextTop <= hr->end() && | |
| 1261 start <= nextTop, | |
| 1262 "Preconditions."); | |
| 1263 // Otherwise, record the number of word's we'll examine. | |
| 1264 size_t words_done = (nextTop - start); | |
| 1265 // Find the first marked object at or after "start". | |
| 1266 start = _bm->getNextMarkedWordAddress(start, nextTop); | |
| 1267 size_t marked_bytes = 0; | |
| 1268 | |
| 1269 // Below, the term "card num" means the result of shifting an address | |
| 1270 // by the card shift -- address 0 corresponds to card number 0. One | |
| 1271 // must subtract the card num of the bottom of the heap to obtain a | |
| 1272 // card table index. | |
| 1273 // The first card num of the sequence of live cards currently being | |
| 1274 // constructed. -1 ==> no sequence. | |
| 1275 intptr_t start_card_num = -1; | |
| 1276 // The last card num of the sequence of live cards currently being | |
| 1277 // constructed. -1 ==> no sequence. | |
| 1278 intptr_t last_card_num = -1; | |
| 1279 | |
| 1280 while (start < nextTop) { | |
| 1281 if (_yield && _cm->do_yield_check()) { | |
| 1282 // We yielded. It might be for a full collection, in which case | |
| 1283 // all bets are off; terminate the traversal. | |
| 1284 if (_cm->has_aborted()) { | |
| 1285 _changed = false; | |
| 1286 return true; | |
| 1287 } else { | |
| 1288 // Otherwise, it might be a collection pause, and the region | |
| 1289 // we're looking at might be in the collection set. We'll | |
| 1290 // abandon this region. | |
| 1291 return false; | |
| 1292 } | |
| 1293 } | |
| 1294 oop obj = oop(start); | |
| 1295 int obj_sz = obj->size(); | |
| 1296 // The card num of the start of the current object. | |
| 1297 intptr_t obj_card_num = | |
| 1298 intptr_t(uintptr_t(start) >> CardTableModRefBS::card_shift); | |
| 1299 | |
| 1300 HeapWord* obj_last = start + obj_sz - 1; | |
| 1301 intptr_t obj_last_card_num = | |
| 1302 intptr_t(uintptr_t(obj_last) >> CardTableModRefBS::card_shift); | |
| 1303 | |
| 1304 if (obj_card_num != last_card_num) { | |
| 1305 if (start_card_num == -1) { | |
| 1306 assert(last_card_num == -1, "Both or neither."); | |
| 1307 start_card_num = obj_card_num; | |
| 1308 } else { | |
| 1309 assert(last_card_num != -1, "Both or neither."); | |
| 1310 assert(obj_card_num >= last_card_num, "Inv"); | |
| 1311 if ((obj_card_num - last_card_num) > 1) { | |
| 1312 // Mark the last run, and start a new one. | |
| 1313 mark_card_num_range(start_card_num, last_card_num); | |
| 1314 start_card_num = obj_card_num; | |
| 1315 } | |
| 1316 } | |
| 1317 #if CARD_BM_TEST_MODE | |
| 1318 /* | |
| 1319 gclog_or_tty->print_cr("Setting bits from %d/%d.", | |
| 1320 obj_card_num - _bottom_card_num, | |
| 1321 obj_last_card_num - _bottom_card_num); | |
| 1322 */ | |
| 1323 for (intptr_t j = obj_card_num; j <= obj_last_card_num; j++) { | |
| 1324 _card_bm->par_at_put(j - _bottom_card_num, 1); | |
| 1325 } | |
| 1326 #endif | |
| 1327 } | |
| 1328 // In any case, we set the last card num. | |
| 1329 last_card_num = obj_last_card_num; | |
| 1330 | |
| 1331 marked_bytes += obj_sz * HeapWordSize; | |
| 1332 // Find the next marked object after this one. | |
| 1333 start = _bm->getNextMarkedWordAddress(start + 1, nextTop); | |
| 1334 _changed = true; | |
| 1335 } | |
| 1336 // Handle the last range, if any. | |
| 1337 if (start_card_num != -1) | |
| 1338 mark_card_num_range(start_card_num, last_card_num); | |
| 1339 if (_final) { | |
| 1340 // Mark the allocated-since-marking portion... | |
| 1341 HeapWord* tp = hr->top(); | |
| 1342 if (nextTop < tp) { | |
| 1343 start_card_num = | |
| 1344 intptr_t(uintptr_t(nextTop) >> CardTableModRefBS::card_shift); | |
| 1345 last_card_num = | |
| 1346 intptr_t(uintptr_t(tp) >> CardTableModRefBS::card_shift); | |
| 1347 mark_card_num_range(start_card_num, last_card_num); | |
| 1348 // This definitely means the region has live objects. | |
| 829 | 1349 set_bit_for_region(hr); |
| 342 | 1350 } |
| 1351 } | |
| 1352 | |
| 1353 hr->add_to_marked_bytes(marked_bytes); | |
| 1354 // Update the live region bitmap. | |
| 1355 if (marked_bytes > 0) { | |
| 829 | 1356 set_bit_for_region(hr); |
| 342 | 1357 } |
| 1358 hr->set_top_at_conc_mark_count(nextTop); | |
| 1359 _tot_live += hr->next_live_bytes(); | |
| 1360 _tot_used += hr->used(); | |
| 1361 _words_done = words_done; | |
| 1362 | |
| 1363 if (!_final) { | |
| 1364 ++_regions_done; | |
| 1365 if (_regions_done % 10 == 0) { | |
| 1366 double end_vtime_sec = os::elapsedVTime(); | |
| 1367 double elapsed_vtime_sec = end_vtime_sec - _start_vtime_sec; | |
| 1368 if (elapsed_vtime_sec > (10.0 / 1000.0)) { | |
| 1369 jlong sleep_time_ms = | |
| 1370 (jlong) (elapsed_vtime_sec * _cm->cleanup_sleep_factor() * 1000.0); | |
| 1371 os::sleep(Thread::current(), sleep_time_ms, false); | |
| 1372 _start_vtime_sec = end_vtime_sec; | |
| 1373 } | |
| 1374 } | |
| 1375 } | |
| 1376 | |
| 1377 return false; | |
| 1378 } | |
| 1379 | |
| 1380 bool changed() { return _changed; } | |
| 1381 void reset() { _changed = false; _words_done = 0; } | |
| 1382 void no_yield() { _yield = false; } | |
| 1383 size_t words_done() { return _words_done; } | |
| 1384 size_t tot_live() { return _tot_live; } | |
| 1385 size_t tot_used() { return _tot_used; } | |
| 1386 }; | |
| 1387 | |
| 1388 | |
| 1389 void ConcurrentMark::calcDesiredRegions() { | |
| 1390 _region_bm.clear(); | |
| 1391 _card_bm.clear(); | |
| 1392 CalcLiveObjectsClosure calccl(false /*final*/, | |
| 1393 nextMarkBitMap(), this, | |
| 936 | 1394 &_region_bm, &_card_bm); |
| 342 | 1395 G1CollectedHeap *g1h = G1CollectedHeap::heap(); |
| 1396 g1h->heap_region_iterate(&calccl); | |
| 1397 | |
| 1398 do { | |
| 1399 calccl.reset(); | |
| 1400 g1h->heap_region_iterate(&calccl); | |
| 1401 } while (calccl.changed()); | |
| 1402 } | |
| 1403 | |
| 1404 class G1ParFinalCountTask: public AbstractGangTask { | |
| 1405 protected: | |
| 1406 G1CollectedHeap* _g1h; | |
| 1407 CMBitMap* _bm; | |
| 1408 size_t _n_workers; | |
| 1409 size_t *_live_bytes; | |
| 1410 size_t *_used_bytes; | |
| 1411 BitMap* _region_bm; | |
| 1412 BitMap* _card_bm; | |
| 1413 public: | |
| 1414 G1ParFinalCountTask(G1CollectedHeap* g1h, CMBitMap* bm, | |
| 1415 BitMap* region_bm, BitMap* card_bm) : | |
| 1416 AbstractGangTask("G1 final counting"), _g1h(g1h), | |
| 1417 _bm(bm), _region_bm(region_bm), _card_bm(card_bm) | |
| 1418 { | |
| 1419 if (ParallelGCThreads > 0) | |
| 1420 _n_workers = _g1h->workers()->total_workers(); | |
| 1421 else | |
| 1422 _n_workers = 1; | |
| 1423 _live_bytes = NEW_C_HEAP_ARRAY(size_t, _n_workers); | |
| 1424 _used_bytes = NEW_C_HEAP_ARRAY(size_t, _n_workers); | |
| 1425 } | |
| 1426 | |
| 1427 ~G1ParFinalCountTask() { | |
| 1428 FREE_C_HEAP_ARRAY(size_t, _live_bytes); | |
| 1429 FREE_C_HEAP_ARRAY(size_t, _used_bytes); | |
| 1430 } | |
| 1431 | |
| 1432 void work(int i) { | |
| 1433 CalcLiveObjectsClosure calccl(true /*final*/, | |
| 1434 _bm, _g1h->concurrent_mark(), | |
| 936 | 1435 _region_bm, _card_bm); |
| 342 | 1436 calccl.no_yield(); |
| 1437 if (ParallelGCThreads > 0) { | |
| 355 | 1438 _g1h->heap_region_par_iterate_chunked(&calccl, i, |
| 1439 HeapRegion::FinalCountClaimValue); | |
| 342 | 1440 } else { |
| 1441 _g1h->heap_region_iterate(&calccl); | |
| 1442 } | |
| 1443 assert(calccl.complete(), "Shouldn't have yielded!"); | |
| 1444 | |
| 1445 guarantee( (size_t)i < _n_workers, "invariant" ); | |
| 1446 _live_bytes[i] = calccl.tot_live(); | |
| 1447 _used_bytes[i] = calccl.tot_used(); | |
| 1448 } | |
| 1449 size_t live_bytes() { | |
| 1450 size_t live_bytes = 0; | |
| 1451 for (size_t i = 0; i < _n_workers; ++i) | |
| 1452 live_bytes += _live_bytes[i]; | |
| 1453 return live_bytes; | |
| 1454 } | |
| 1455 size_t used_bytes() { | |
| 1456 size_t used_bytes = 0; | |
| 1457 for (size_t i = 0; i < _n_workers; ++i) | |
| 1458 used_bytes += _used_bytes[i]; | |
| 1459 return used_bytes; | |
| 1460 } | |
| 1461 }; | |
| 1462 | |
| 1463 class G1ParNoteEndTask; | |
| 1464 | |
| 1465 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { | |
| 1466 G1CollectedHeap* _g1; | |
| 1467 int _worker_num; | |
| 1468 size_t _max_live_bytes; | |
| 1469 size_t _regions_claimed; | |
| 1470 size_t _freed_bytes; | |
| 1471 size_t _cleared_h_regions; | |
| 1472 size_t _freed_regions; | |
| 1473 UncleanRegionList* _unclean_region_list; | |
| 1474 double _claimed_region_time; | |
| 1475 double _max_region_time; | |
| 1476 | |
| 1477 public: | |
| 1478 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, | |
| 1479 UncleanRegionList* list, | |
| 1480 int worker_num); | |
| 1481 size_t freed_bytes() { return _freed_bytes; } | |
| 1482 size_t cleared_h_regions() { return _cleared_h_regions; } | |
| 1483 size_t freed_regions() { return _freed_regions; } | |
| 1484 UncleanRegionList* unclean_region_list() { | |
| 1485 return _unclean_region_list; | |
| 1486 } | |
| 1487 | |
| 1488 bool doHeapRegion(HeapRegion *r); | |
| 1489 | |
| 1490 size_t max_live_bytes() { return _max_live_bytes; } | |
| 1491 size_t regions_claimed() { return _regions_claimed; } | |
| 1492 double claimed_region_time_sec() { return _claimed_region_time; } | |
| 1493 double max_region_time_sec() { return _max_region_time; } | |
| 1494 }; | |
| 1495 | |
| 1496 class G1ParNoteEndTask: public AbstractGangTask { | |
| 1497 friend class G1NoteEndOfConcMarkClosure; | |
| 1498 protected: | |
| 1499 G1CollectedHeap* _g1h; | |
| 1500 size_t _max_live_bytes; | |
| 1501 size_t _freed_bytes; | |
| 1502 ConcurrentMark::ParCleanupThreadState** _par_cleanup_thread_state; | |
| 1503 public: | |
| 1504 G1ParNoteEndTask(G1CollectedHeap* g1h, | |
| 1505 ConcurrentMark::ParCleanupThreadState** | |
| 1506 par_cleanup_thread_state) : | |
| 1507 AbstractGangTask("G1 note end"), _g1h(g1h), | |
| 1508 _max_live_bytes(0), _freed_bytes(0), | |
| 1509 _par_cleanup_thread_state(par_cleanup_thread_state) | |
| 1510 {} | |
| 1511 | |
| 1512 void work(int i) { | |
| 1513 double start = os::elapsedTime(); | |
| 1514 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, | |
| 1515 &_par_cleanup_thread_state[i]->list, | |
| 1516 i); | |
| 1517 if (ParallelGCThreads > 0) { | |
| 355 | 1518 _g1h->heap_region_par_iterate_chunked(&g1_note_end, i, |
| 1519 HeapRegion::NoteEndClaimValue); | |
| 342 | 1520 } else { |
| 1521 _g1h->heap_region_iterate(&g1_note_end); | |
| 1522 } | |
| 1523 assert(g1_note_end.complete(), "Shouldn't have yielded!"); | |
| 1524 | |
| 1525 // Now finish up freeing the current thread's regions. | |
| 1526 _g1h->finish_free_region_work(g1_note_end.freed_bytes(), | |
| 1527 g1_note_end.cleared_h_regions(), | |
| 1528 0, NULL); | |
| 1529 { | |
| 1530 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
| 1531 _max_live_bytes += g1_note_end.max_live_bytes(); | |
| 1532 _freed_bytes += g1_note_end.freed_bytes(); | |
| 1533 } | |
| 1534 double end = os::elapsedTime(); | |
| 1535 if (G1PrintParCleanupStats) { | |
| 1536 gclog_or_tty->print(" Worker thread %d [%8.3f..%8.3f = %8.3f ms] " | |
| 1537 "claimed %d regions (tot = %8.3f ms, max = %8.3f ms).\n", | |
| 1538 i, start, end, (end-start)*1000.0, | |
| 1539 g1_note_end.regions_claimed(), | |
| 1540 g1_note_end.claimed_region_time_sec()*1000.0, | |
| 1541 g1_note_end.max_region_time_sec()*1000.0); | |
| 1542 } | |
| 1543 } | |
| 1544 size_t max_live_bytes() { return _max_live_bytes; } | |
| 1545 size_t freed_bytes() { return _freed_bytes; } | |
| 1546 }; | |
| 1547 | |
| 1548 class G1ParScrubRemSetTask: public AbstractGangTask { | |
| 1549 protected: | |
| 1550 G1RemSet* _g1rs; | |
| 1551 BitMap* _region_bm; | |
| 1552 BitMap* _card_bm; | |
| 1553 public: | |
| 1554 G1ParScrubRemSetTask(G1CollectedHeap* g1h, | |
| 1555 BitMap* region_bm, BitMap* card_bm) : | |
| 1556 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), | |
| 1557 _region_bm(region_bm), _card_bm(card_bm) | |
| 1558 {} | |
| 1559 | |
| 1560 void work(int i) { | |
| 1561 if (ParallelGCThreads > 0) { | |
| 355 | 1562 _g1rs->scrub_par(_region_bm, _card_bm, i, |
| 1563 HeapRegion::ScrubRemSetClaimValue); | |
| 342 | 1564 } else { |
| 1565 _g1rs->scrub(_region_bm, _card_bm); | |
| 1566 } | |
| 1567 } | |
| 1568 | |
| 1569 }; | |
| 1570 | |
| 1571 G1NoteEndOfConcMarkClosure:: | |
| 1572 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, | |
| 1573 UncleanRegionList* list, | |
| 1574 int worker_num) | |
| 1575 : _g1(g1), _worker_num(worker_num), | |
| 1576 _max_live_bytes(0), _regions_claimed(0), | |
| 1577 _freed_bytes(0), _cleared_h_regions(0), _freed_regions(0), | |
| 1578 _claimed_region_time(0.0), _max_region_time(0.0), | |
| 1579 _unclean_region_list(list) | |
| 1580 {} | |
| 1581 | |
| 1582 bool G1NoteEndOfConcMarkClosure::doHeapRegion(HeapRegion *r) { | |
| 1583 // We use a claim value of zero here because all regions | |
| 1584 // were claimed with value 1 in the FinalCount task. | |
| 1585 r->reset_gc_time_stamp(); | |
| 1586 if (!r->continuesHumongous()) { | |
| 1587 double start = os::elapsedTime(); | |
| 1588 _regions_claimed++; | |
| 1589 r->note_end_of_marking(); | |
| 1590 _max_live_bytes += r->max_live_bytes(); | |
| 1591 _g1->free_region_if_totally_empty_work(r, | |
| 1592 _freed_bytes, | |
| 1593 _cleared_h_regions, | |
| 1594 _freed_regions, | |
| 1595 _unclean_region_list, | |
| 1596 true /*par*/); | |
| 1597 double region_time = (os::elapsedTime() - start); | |
| 1598 _claimed_region_time += region_time; | |
| 1599 if (region_time > _max_region_time) _max_region_time = region_time; | |
| 1600 } | |
| 1601 return false; | |
| 1602 } | |
| 1603 | |
| 1604 void ConcurrentMark::cleanup() { | |
| 1605 // world is stopped at this checkpoint | |
| 1606 assert(SafepointSynchronize::is_at_safepoint(), | |
| 1607 "world should be stopped"); | |
| 1608 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 1609 | |
| 1610 // If a full collection has happened, we shouldn't do this. | |
| 1611 if (has_aborted()) { | |
| 1612 g1h->set_marking_complete(); // So bitmap clearing isn't confused | |
| 1613 return; | |
| 1614 } | |
| 1615 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1616 if (VerifyDuringGC) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1617 HandleMark hm; // handle scope |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1618 gclog_or_tty->print(" VerifyDuringGC:(before)"); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1619 Universe::heap()->prepare_for_verify(); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1620 Universe::verify(/* allow dirty */ true, |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1621 /* silent */ false, |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1622 /* prev marking */ true); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1623 } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1624 |
| 342 | 1625 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); |
| 1626 g1p->record_concurrent_mark_cleanup_start(); | |
| 1627 | |
| 1628 double start = os::elapsedTime(); | |
| 1629 | |
| 1630 // Do counting once more with the world stopped for good measure. | |
| 1631 G1ParFinalCountTask g1_par_count_task(g1h, nextMarkBitMap(), | |
| 1632 &_region_bm, &_card_bm); | |
| 1633 if (ParallelGCThreads > 0) { | |
| 355 | 1634 assert(g1h->check_heap_region_claim_values( |
| 1635 HeapRegion::InitialClaimValue), | |
| 1636 "sanity check"); | |
| 1637 | |
| 342 | 1638 int n_workers = g1h->workers()->total_workers(); |
| 1639 g1h->set_par_threads(n_workers); | |
| 1640 g1h->workers()->run_task(&g1_par_count_task); | |
| 1641 g1h->set_par_threads(0); | |
| 355 | 1642 |
| 1643 assert(g1h->check_heap_region_claim_values( | |
| 1644 HeapRegion::FinalCountClaimValue), | |
| 1645 "sanity check"); | |
| 342 | 1646 } else { |
| 1647 g1_par_count_task.work(0); | |
| 1648 } | |
| 1649 | |
| 1650 size_t known_garbage_bytes = | |
| 1651 g1_par_count_task.used_bytes() - g1_par_count_task.live_bytes(); | |
| 1652 #if 0 | |
| 1653 gclog_or_tty->print_cr("used %1.2lf, live %1.2lf, garbage %1.2lf", | |
| 1654 (double) g1_par_count_task.used_bytes() / (double) (1024 * 1024), | |
| 1655 (double) g1_par_count_task.live_bytes() / (double) (1024 * 1024), | |
| 1656 (double) known_garbage_bytes / (double) (1024 * 1024)); | |
| 1657 #endif // 0 | |
| 1658 g1p->set_known_garbage_bytes(known_garbage_bytes); | |
| 1659 | |
| 1660 size_t start_used_bytes = g1h->used(); | |
| 1661 _at_least_one_mark_complete = true; | |
| 1662 g1h->set_marking_complete(); | |
| 1663 | |
| 1664 double count_end = os::elapsedTime(); | |
| 1665 double this_final_counting_time = (count_end - start); | |
| 1666 if (G1PrintParCleanupStats) { | |
| 1667 gclog_or_tty->print_cr("Cleanup:"); | |
| 1668 gclog_or_tty->print_cr(" Finalize counting: %8.3f ms", | |
| 1669 this_final_counting_time*1000.0); | |
| 1670 } | |
| 1671 _total_counting_time += this_final_counting_time; | |
| 1672 | |
| 1673 // Install newly created mark bitMap as "prev". | |
| 1674 swapMarkBitMaps(); | |
| 1675 | |
| 1676 g1h->reset_gc_time_stamp(); | |
| 1677 | |
| 1678 // Note end of marking in all heap regions. | |
| 1679 double note_end_start = os::elapsedTime(); | |
| 1680 G1ParNoteEndTask g1_par_note_end_task(g1h, _par_cleanup_thread_state); | |
| 1681 if (ParallelGCThreads > 0) { | |
| 1682 int n_workers = g1h->workers()->total_workers(); | |
| 1683 g1h->set_par_threads(n_workers); | |
| 1684 g1h->workers()->run_task(&g1_par_note_end_task); | |
| 1685 g1h->set_par_threads(0); | |
| 355 | 1686 |
| 1687 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue), | |
| 1688 "sanity check"); | |
| 342 | 1689 } else { |
| 1690 g1_par_note_end_task.work(0); | |
| 1691 } | |
| 1692 g1h->set_unclean_regions_coming(true); | |
| 1693 double note_end_end = os::elapsedTime(); | |
| 1694 // Tell the mutators that there might be unclean regions coming... | |
| 1695 if (G1PrintParCleanupStats) { | |
| 1696 gclog_or_tty->print_cr(" note end of marking: %8.3f ms.", | |
| 1697 (note_end_end - note_end_start)*1000.0); | |
| 1698 } | |
| 1699 | |
| 355 | 1700 |
| 342 | 1701 // call below, since it affects the metric by which we sort the heap |
| 1702 // regions. | |
| 1703 if (G1ScrubRemSets) { | |
| 1704 double rs_scrub_start = os::elapsedTime(); | |
| 1705 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm); | |
| 1706 if (ParallelGCThreads > 0) { | |
| 1707 int n_workers = g1h->workers()->total_workers(); | |
| 1708 g1h->set_par_threads(n_workers); | |
| 1709 g1h->workers()->run_task(&g1_par_scrub_rs_task); | |
| 1710 g1h->set_par_threads(0); | |
| 355 | 1711 |
| 1712 assert(g1h->check_heap_region_claim_values( | |
| 1713 HeapRegion::ScrubRemSetClaimValue), | |
| 1714 "sanity check"); | |
| 342 | 1715 } else { |
| 1716 g1_par_scrub_rs_task.work(0); | |
| 1717 } | |
| 1718 | |
| 1719 double rs_scrub_end = os::elapsedTime(); | |
| 1720 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); | |
| 1721 _total_rs_scrub_time += this_rs_scrub_time; | |
| 1722 } | |
| 1723 | |
| 1724 // this will also free any regions totally full of garbage objects, | |
| 1725 // and sort the regions. | |
| 1726 g1h->g1_policy()->record_concurrent_mark_cleanup_end( | |
| 1727 g1_par_note_end_task.freed_bytes(), | |
| 1728 g1_par_note_end_task.max_live_bytes()); | |
| 1729 | |
| 1730 // Statistics. | |
| 1731 double end = os::elapsedTime(); | |
| 1732 _cleanup_times.add((end - start) * 1000.0); | |
| 1733 | |
| 1734 // G1CollectedHeap::heap()->print(); | |
| 1735 // gclog_or_tty->print_cr("HEAP GC TIME STAMP : %d", | |
| 1736 // G1CollectedHeap::heap()->get_gc_time_stamp()); | |
| 1737 | |
| 1738 if (PrintGC || PrintGCDetails) { | |
| 1739 g1h->print_size_transition(gclog_or_tty, | |
| 1740 start_used_bytes, | |
| 1741 g1h->used(), | |
| 1742 g1h->capacity()); | |
| 1743 } | |
| 1744 | |
| 1745 size_t cleaned_up_bytes = start_used_bytes - g1h->used(); | |
| 1746 g1p->decrease_known_garbage_bytes(cleaned_up_bytes); | |
| 1747 | |
| 1748 // We need to make this be a "collection" so any collection pause that | |
| 1749 // races with it goes around and waits for completeCleanup to finish. | |
| 1750 g1h->increment_total_collections(); | |
| 1751 | |
| 751 | 1752 if (VerifyDuringGC) { |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1753 HandleMark hm; // handle scope |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1754 gclog_or_tty->print(" VerifyDuringGC:(after)"); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1755 Universe::heap()->prepare_for_verify(); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1756 Universe::verify(/* allow dirty */ true, |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1757 /* silent */ false, |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1758 /* prev marking */ true); |
| 342 | 1759 } |
| 1760 } | |
| 1761 | |
| 1762 void ConcurrentMark::completeCleanup() { | |
| 1763 // A full collection intervened. | |
| 1764 if (has_aborted()) return; | |
| 1765 | |
| 1766 int first = 0; | |
| 1767 int last = (int)MAX2(ParallelGCThreads, (size_t)1); | |
| 1768 for (int t = 0; t < last; t++) { | |
| 1769 UncleanRegionList* list = &_par_cleanup_thread_state[t]->list; | |
| 1770 assert(list->well_formed(), "Inv"); | |
| 1771 HeapRegion* hd = list->hd(); | |
| 1772 while (hd != NULL) { | |
| 1773 // Now finish up the other stuff. | |
| 1774 hd->rem_set()->clear(); | |
| 1775 HeapRegion* next_hd = hd->next_from_unclean_list(); | |
| 1776 (void)list->pop(); | |
| 1777 guarantee(list->hd() == next_hd, "how not?"); | |
| 1778 _g1h->put_region_on_unclean_list(hd); | |
| 1779 if (!hd->isHumongous()) { | |
| 1780 // Add this to the _free_regions count by 1. | |
| 1781 _g1h->finish_free_region_work(0, 0, 1, NULL); | |
| 1782 } | |
| 1783 hd = list->hd(); | |
| 1784 guarantee(hd == next_hd, "how not?"); | |
| 1785 } | |
| 1786 } | |
| 1787 } | |
| 1788 | |
| 1789 | |
| 1790 class G1CMIsAliveClosure: public BoolObjectClosure { | |
| 1791 G1CollectedHeap* _g1; | |
| 1792 public: | |
| 1793 G1CMIsAliveClosure(G1CollectedHeap* g1) : | |
| 1794 _g1(g1) | |
| 1795 {} | |
| 1796 | |
| 1797 void do_object(oop obj) { | |
| 1798 assert(false, "not to be invoked"); | |
| 1799 } | |
| 1800 bool do_object_b(oop obj) { | |
| 1801 HeapWord* addr = (HeapWord*)obj; | |
| 1802 return addr != NULL && | |
| 1803 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); | |
| 1804 } | |
| 1805 }; | |
| 1806 | |
| 1807 class G1CMKeepAliveClosure: public OopClosure { | |
| 1808 G1CollectedHeap* _g1; | |
| 1809 ConcurrentMark* _cm; | |
| 1810 CMBitMap* _bitMap; | |
| 1811 public: | |
| 1812 G1CMKeepAliveClosure(G1CollectedHeap* g1, ConcurrentMark* cm, | |
| 1813 CMBitMap* bitMap) : | |
| 1814 _g1(g1), _cm(cm), | |
| 1815 _bitMap(bitMap) {} | |
| 1816 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1817 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1818 virtual void do_oop( oop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1819 |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1820 template <class T> void do_oop_work(T* p) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1821 oop thisOop = oopDesc::load_decode_heap_oop(p); |
| 342 | 1822 HeapWord* addr = (HeapWord*)thisOop; |
| 1823 if (_g1->is_in_g1_reserved(addr) && _g1->is_obj_ill(thisOop)) { | |
| 1824 _bitMap->mark(addr); | |
| 1825 _cm->mark_stack_push(thisOop); | |
| 1826 } | |
| 1827 } | |
| 1828 }; | |
| 1829 | |
| 1830 class G1CMDrainMarkingStackClosure: public VoidClosure { | |
| 1831 CMMarkStack* _markStack; | |
| 1832 CMBitMap* _bitMap; | |
| 1833 G1CMKeepAliveClosure* _oopClosure; | |
| 1834 public: | |
| 1835 G1CMDrainMarkingStackClosure(CMBitMap* bitMap, CMMarkStack* markStack, | |
| 1836 G1CMKeepAliveClosure* oopClosure) : | |
| 1837 _bitMap(bitMap), | |
| 1838 _markStack(markStack), | |
| 1839 _oopClosure(oopClosure) | |
| 1840 {} | |
| 1841 | |
| 1842 void do_void() { | |
| 1843 _markStack->drain((OopClosure*)_oopClosure, _bitMap, false); | |
| 1844 } | |
| 1845 }; | |
| 1846 | |
| 1847 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { | |
| 1848 ResourceMark rm; | |
| 1849 HandleMark hm; | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1850 G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1851 ReferenceProcessor* rp = g1h->ref_processor(); |
| 342 | 1852 |
| 1853 // Process weak references. | |
|
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
1854 rp->setup_policy(clear_all_soft_refs); |
| 342 | 1855 assert(_markStack.isEmpty(), "mark stack should be empty"); |
| 1856 | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1857 G1CMIsAliveClosure g1IsAliveClosure (g1h); |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1858 G1CMKeepAliveClosure g1KeepAliveClosure(g1h, this, nextMarkBitMap()); |
| 342 | 1859 G1CMDrainMarkingStackClosure |
| 1860 g1DrainMarkingStackClosure(nextMarkBitMap(), &_markStack, | |
| 1861 &g1KeepAliveClosure); | |
| 1862 | |
| 1863 // XXXYYY Also: copy the parallel ref processing code from CMS. | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1864 rp->process_discovered_references(&g1IsAliveClosure, |
| 342 | 1865 &g1KeepAliveClosure, |
| 1866 &g1DrainMarkingStackClosure, | |
| 1867 NULL); | |
| 1868 assert(_markStack.overflow() || _markStack.isEmpty(), | |
| 1869 "mark stack should be empty (unless it overflowed)"); | |
| 1870 if (_markStack.overflow()) { | |
| 1871 set_has_overflown(); | |
| 1872 } | |
| 1873 | |
| 1874 rp->enqueue_discovered_references(); | |
| 1875 rp->verify_no_references_recorded(); | |
| 1876 assert(!rp->discovery_enabled(), "should have been disabled"); | |
| 1877 | |
| 1878 // Now clean up stale oops in SymbolTable and StringTable | |
| 1879 SymbolTable::unlink(&g1IsAliveClosure); | |
| 1880 StringTable::unlink(&g1IsAliveClosure); | |
| 1881 } | |
| 1882 | |
| 1883 void ConcurrentMark::swapMarkBitMaps() { | |
| 1884 CMBitMapRO* temp = _prevMarkBitMap; | |
| 1885 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; | |
| 1886 _nextMarkBitMap = (CMBitMap*) temp; | |
| 1887 } | |
| 1888 | |
| 1889 class CMRemarkTask: public AbstractGangTask { | |
| 1890 private: | |
| 1891 ConcurrentMark *_cm; | |
| 1892 | |
| 1893 public: | |
| 1894 void work(int worker_i) { | |
| 1895 // Since all available tasks are actually started, we should | |
| 1896 // only proceed if we're supposed to be actived. | |
| 1897 if ((size_t)worker_i < _cm->active_tasks()) { | |
| 1898 CMTask* task = _cm->task(worker_i); | |
| 1899 task->record_start_time(); | |
| 1900 do { | |
| 1901 task->do_marking_step(1000000000.0 /* something very large */); | |
| 1902 } while (task->has_aborted() && !_cm->has_overflown()); | |
| 1903 // If we overflow, then we do not want to restart. We instead | |
| 1904 // want to abort remark and do concurrent marking again. | |
| 1905 task->record_end_time(); | |
| 1906 } | |
| 1907 } | |
| 1908 | |
| 1909 CMRemarkTask(ConcurrentMark* cm) : | |
| 1910 AbstractGangTask("Par Remark"), _cm(cm) { } | |
| 1911 }; | |
| 1912 | |
| 1913 void ConcurrentMark::checkpointRootsFinalWork() { | |
| 1914 ResourceMark rm; | |
| 1915 HandleMark hm; | |
| 1916 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 1917 | |
| 1918 g1h->ensure_parsability(false); | |
| 1919 | |
| 1920 if (ParallelGCThreads > 0) { | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
1921 G1CollectedHeap::StrongRootsScope srs(g1h); |
| 342 | 1922 // this is remark, so we'll use up all available threads |
| 1923 int active_workers = ParallelGCThreads; | |
| 1924 set_phase(active_workers, false); | |
| 1925 | |
| 1926 CMRemarkTask remarkTask(this); | |
| 1927 // We will start all available threads, even if we decide that the | |
| 1928 // active_workers will be fewer. The extra ones will just bail out | |
| 1929 // immediately. | |
| 1930 int n_workers = g1h->workers()->total_workers(); | |
| 1931 g1h->set_par_threads(n_workers); | |
| 1932 g1h->workers()->run_task(&remarkTask); | |
| 1933 g1h->set_par_threads(0); | |
| 1934 | |
| 1935 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 1936 guarantee( satb_mq_set.completed_buffers_num() == 0, "invariant" ); | |
| 1937 } else { | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
1938 G1CollectedHeap::StrongRootsScope srs(g1h); |
| 342 | 1939 // this is remark, so we'll use up all available threads |
| 1940 int active_workers = 1; | |
| 1941 set_phase(active_workers, false); | |
| 1942 | |
| 1943 CMRemarkTask remarkTask(this); | |
| 1944 // We will start all available threads, even if we decide that the | |
| 1945 // active_workers will be fewer. The extra ones will just bail out | |
| 1946 // immediately. | |
| 1947 remarkTask.work(0); | |
| 1948 | |
| 1949 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 1950 guarantee( satb_mq_set.completed_buffers_num() == 0, "invariant" ); | |
| 1951 } | |
| 1952 | |
| 1953 print_stats(); | |
| 1954 | |
| 1955 if (!restart_for_overflow()) | |
| 1956 set_non_marking_state(); | |
| 1957 | |
| 1958 #if VERIFY_OBJS_PROCESSED | |
| 1959 if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) { | |
| 1960 gclog_or_tty->print_cr("Processed = %d, enqueued = %d.", | |
| 1961 _scan_obj_cl.objs_processed, | |
| 1962 ThreadLocalObjQueue::objs_enqueued); | |
| 1963 guarantee(_scan_obj_cl.objs_processed == | |
| 1964 ThreadLocalObjQueue::objs_enqueued, | |
| 1965 "Different number of objs processed and enqueued."); | |
| 1966 } | |
| 1967 #endif | |
| 1968 } | |
| 1969 | |
| 1970 class ReachablePrinterOopClosure: public OopClosure { | |
| 1971 private: | |
| 1972 G1CollectedHeap* _g1h; | |
| 1973 CMBitMapRO* _bitmap; | |
| 1974 outputStream* _out; | |
| 1975 | |
| 1976 public: | |
| 1977 ReachablePrinterOopClosure(CMBitMapRO* bitmap, outputStream* out) : | |
| 1978 _bitmap(bitmap), _g1h(G1CollectedHeap::heap()), _out(out) { } | |
| 1979 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1980 void do_oop(narrowOop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1981 void do_oop( oop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1982 |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1983 template <class T> void do_oop_work(T* p) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1984 oop obj = oopDesc::load_decode_heap_oop(p); |
| 342 | 1985 const char* str = NULL; |
| 1986 const char* str2 = ""; | |
| 1987 | |
| 1988 if (!_g1h->is_in_g1_reserved(obj)) | |
| 1989 str = "outside G1 reserved"; | |
| 1990 else { | |
| 1991 HeapRegion* hr = _g1h->heap_region_containing(obj); | |
| 1992 guarantee( hr != NULL, "invariant" ); | |
| 1993 if (hr->obj_allocated_since_prev_marking(obj)) { | |
| 1994 str = "over TAMS"; | |
| 1995 if (_bitmap->isMarked((HeapWord*) obj)) | |
| 1996 str2 = " AND MARKED"; | |
| 1997 } else if (_bitmap->isMarked((HeapWord*) obj)) | |
| 1998 str = "marked"; | |
| 1999 else | |
| 2000 str = "#### NOT MARKED ####"; | |
| 2001 } | |
| 2002 | |
| 2003 _out->print_cr(" "PTR_FORMAT" contains "PTR_FORMAT" %s%s", | |
| 2004 p, (void*) obj, str, str2); | |
| 2005 } | |
| 2006 }; | |
| 2007 | |
| 2008 class ReachablePrinterClosure: public BitMapClosure { | |
| 2009 private: | |
| 2010 CMBitMapRO* _bitmap; | |
| 2011 outputStream* _out; | |
| 2012 | |
| 2013 public: | |
| 2014 ReachablePrinterClosure(CMBitMapRO* bitmap, outputStream* out) : | |
| 2015 _bitmap(bitmap), _out(out) { } | |
| 2016 | |
| 2017 bool do_bit(size_t offset) { | |
| 2018 HeapWord* addr = _bitmap->offsetToHeapWord(offset); | |
| 2019 ReachablePrinterOopClosure oopCl(_bitmap, _out); | |
| 2020 | |
| 2021 _out->print_cr(" obj "PTR_FORMAT", offset %10d (marked)", addr, offset); | |
| 2022 oop(addr)->oop_iterate(&oopCl); | |
| 2023 _out->print_cr(""); | |
| 2024 | |
| 2025 return true; | |
| 2026 } | |
| 2027 }; | |
| 2028 | |
| 2029 class ObjInRegionReachablePrinterClosure : public ObjectClosure { | |
| 2030 private: | |
| 2031 CMBitMapRO* _bitmap; | |
| 2032 outputStream* _out; | |
| 2033 | |
| 2034 public: | |
| 2035 void do_object(oop o) { | |
| 2036 ReachablePrinterOopClosure oopCl(_bitmap, _out); | |
| 2037 | |
| 2038 _out->print_cr(" obj "PTR_FORMAT" (over TAMS)", (void*) o); | |
| 2039 o->oop_iterate(&oopCl); | |
| 2040 _out->print_cr(""); | |
| 2041 } | |
| 2042 | |
| 2043 ObjInRegionReachablePrinterClosure(CMBitMapRO* bitmap, outputStream* out) : | |
| 2044 _bitmap(bitmap), _out(out) { } | |
| 2045 }; | |
| 2046 | |
| 2047 class RegionReachablePrinterClosure : public HeapRegionClosure { | |
| 2048 private: | |
| 2049 CMBitMapRO* _bitmap; | |
| 2050 outputStream* _out; | |
| 2051 | |
| 2052 public: | |
| 2053 bool doHeapRegion(HeapRegion* hr) { | |
| 2054 HeapWord* b = hr->bottom(); | |
| 2055 HeapWord* e = hr->end(); | |
| 2056 HeapWord* t = hr->top(); | |
| 2057 HeapWord* p = hr->prev_top_at_mark_start(); | |
| 2058 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " | |
| 2059 "PTAMS: "PTR_FORMAT, b, e, t, p); | |
| 2060 _out->print_cr(""); | |
| 2061 | |
| 2062 ObjInRegionReachablePrinterClosure ocl(_bitmap, _out); | |
| 2063 hr->object_iterate_mem_careful(MemRegion(p, t), &ocl); | |
| 2064 | |
| 2065 return false; | |
| 2066 } | |
| 2067 | |
| 2068 RegionReachablePrinterClosure(CMBitMapRO* bitmap, | |
| 2069 outputStream* out) : | |
| 2070 _bitmap(bitmap), _out(out) { } | |
| 2071 }; | |
| 2072 | |
| 2073 void ConcurrentMark::print_prev_bitmap_reachable() { | |
| 2074 outputStream* out = gclog_or_tty; | |
| 2075 | |
| 2076 #if SEND_HEAP_DUMP_TO_FILE | |
| 2077 guarantee(heap_dump_file == NULL, "Protocol"); | |
| 2078 char fn_buf[100]; | |
| 2079 sprintf(fn_buf, "/tmp/dump.txt.%d", os::current_process_id()); | |
| 2080 heap_dump_file = fopen(fn_buf, "w"); | |
| 2081 fileStream fstream(heap_dump_file); | |
| 2082 out = &fstream; | |
| 2083 #endif // SEND_HEAP_DUMP_TO_FILE | |
| 2084 | |
| 2085 RegionReachablePrinterClosure rcl(_prevMarkBitMap, out); | |
| 2086 out->print_cr("--- ITERATING OVER REGIONS WITH PTAMS < TOP"); | |
| 2087 _g1h->heap_region_iterate(&rcl); | |
| 2088 out->print_cr(""); | |
| 2089 | |
| 2090 ReachablePrinterClosure cl(_prevMarkBitMap, out); | |
| 2091 out->print_cr("--- REACHABLE OBJECTS ON THE BITMAP"); | |
| 2092 _prevMarkBitMap->iterate(&cl); | |
| 2093 out->print_cr(""); | |
| 2094 | |
| 2095 #if SEND_HEAP_DUMP_TO_FILE | |
| 2096 fclose(heap_dump_file); | |
| 2097 heap_dump_file = NULL; | |
| 2098 #endif // SEND_HEAP_DUMP_TO_FILE | |
| 2099 } | |
| 2100 | |
| 2101 // This note is for drainAllSATBBuffers and the code in between. | |
| 2102 // In the future we could reuse a task to do this work during an | |
| 2103 // evacuation pause (since now tasks are not active and can be claimed | |
| 2104 // during an evacuation pause). This was a late change to the code and | |
| 2105 // is currently not being taken advantage of. | |
| 2106 | |
| 2107 class CMGlobalObjectClosure : public ObjectClosure { | |
| 2108 private: | |
| 2109 ConcurrentMark* _cm; | |
| 2110 | |
| 2111 public: | |
| 2112 void do_object(oop obj) { | |
| 2113 _cm->deal_with_reference(obj); | |
| 2114 } | |
| 2115 | |
| 2116 CMGlobalObjectClosure(ConcurrentMark* cm) : _cm(cm) { } | |
| 2117 }; | |
| 2118 | |
| 2119 void ConcurrentMark::deal_with_reference(oop obj) { | |
| 2120 if (verbose_high()) | |
| 2121 gclog_or_tty->print_cr("[global] we're dealing with reference "PTR_FORMAT, | |
| 2122 (void*) obj); | |
| 2123 | |
| 2124 | |
| 2125 HeapWord* objAddr = (HeapWord*) obj; | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2126 assert(obj->is_oop_or_null(true /* ignore mark word */), "Error"); |
| 342 | 2127 if (_g1h->is_in_g1_reserved(objAddr)) { |
| 2128 tmp_guarantee_CM( obj != NULL, "is_in_g1_reserved should ensure this" ); | |
| 2129 HeapRegion* hr = _g1h->heap_region_containing(obj); | |
| 2130 if (_g1h->is_obj_ill(obj, hr)) { | |
| 2131 if (verbose_high()) | |
| 2132 gclog_or_tty->print_cr("[global] "PTR_FORMAT" is not considered " | |
| 2133 "marked", (void*) obj); | |
| 2134 | |
| 2135 // we need to mark it first | |
| 2136 if (_nextMarkBitMap->parMark(objAddr)) { | |
| 2137 // No OrderAccess:store_load() is needed. It is implicit in the | |
| 2138 // CAS done in parMark(objAddr) above | |
| 2139 HeapWord* finger = _finger; | |
| 2140 if (objAddr < finger) { | |
| 2141 if (verbose_high()) | |
| 2142 gclog_or_tty->print_cr("[global] below the global finger " | |
| 2143 "("PTR_FORMAT"), pushing it", finger); | |
| 2144 if (!mark_stack_push(obj)) { | |
| 2145 if (verbose_low()) | |
| 2146 gclog_or_tty->print_cr("[global] global stack overflow during " | |
| 2147 "deal_with_reference"); | |
| 2148 } | |
| 2149 } | |
| 2150 } | |
| 2151 } | |
| 2152 } | |
| 2153 } | |
| 2154 | |
| 2155 void ConcurrentMark::drainAllSATBBuffers() { | |
| 2156 CMGlobalObjectClosure oc(this); | |
| 2157 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 2158 satb_mq_set.set_closure(&oc); | |
| 2159 | |
| 2160 while (satb_mq_set.apply_closure_to_completed_buffer()) { | |
| 2161 if (verbose_medium()) | |
| 2162 gclog_or_tty->print_cr("[global] processed an SATB buffer"); | |
| 2163 } | |
| 2164 | |
| 2165 // no need to check whether we should do this, as this is only | |
| 2166 // called during an evacuation pause | |
| 2167 satb_mq_set.iterate_closure_all_threads(); | |
| 2168 | |
| 2169 satb_mq_set.set_closure(NULL); | |
| 2170 guarantee( satb_mq_set.completed_buffers_num() == 0, "invariant" ); | |
| 2171 } | |
| 2172 | |
| 2173 void ConcurrentMark::markPrev(oop p) { | |
| 2174 // Note we are overriding the read-only view of the prev map here, via | |
| 2175 // the cast. | |
| 2176 ((CMBitMap*)_prevMarkBitMap)->mark((HeapWord*)p); | |
| 2177 } | |
| 2178 | |
| 2179 void ConcurrentMark::clear(oop p) { | |
| 2180 assert(p != NULL && p->is_oop(), "expected an oop"); | |
| 2181 HeapWord* addr = (HeapWord*)p; | |
| 2182 assert(addr >= _nextMarkBitMap->startWord() || | |
| 2183 addr < _nextMarkBitMap->endWord(), "in a region"); | |
| 2184 | |
| 2185 _nextMarkBitMap->clear(addr); | |
| 2186 } | |
| 2187 | |
| 2188 void ConcurrentMark::clearRangeBothMaps(MemRegion mr) { | |
| 2189 // Note we are overriding the read-only view of the prev map here, via | |
| 2190 // the cast. | |
| 2191 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); | |
| 2192 _nextMarkBitMap->clearRange(mr); | |
| 2193 } | |
| 2194 | |
| 2195 HeapRegion* | |
| 2196 ConcurrentMark::claim_region(int task_num) { | |
| 2197 // "checkpoint" the finger | |
| 2198 HeapWord* finger = _finger; | |
| 2199 | |
| 2200 // _heap_end will not change underneath our feet; it only changes at | |
| 2201 // yield points. | |
| 2202 while (finger < _heap_end) { | |
| 2203 tmp_guarantee_CM( _g1h->is_in_g1_reserved(finger), "invariant" ); | |
| 2204 | |
| 2205 // is the gap between reading the finger and doing the CAS too long? | |
| 2206 | |
| 2207 HeapRegion* curr_region = _g1h->heap_region_containing(finger); | |
| 2208 HeapWord* bottom = curr_region->bottom(); | |
| 2209 HeapWord* end = curr_region->end(); | |
| 2210 HeapWord* limit = curr_region->next_top_at_mark_start(); | |
| 2211 | |
| 2212 if (verbose_low()) | |
| 2213 gclog_or_tty->print_cr("[%d] curr_region = "PTR_FORMAT" " | |
| 2214 "["PTR_FORMAT", "PTR_FORMAT"), " | |
| 2215 "limit = "PTR_FORMAT, | |
| 2216 task_num, curr_region, bottom, end, limit); | |
| 2217 | |
| 2218 HeapWord* res = | |
| 2219 (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); | |
| 2220 if (res == finger) { | |
| 2221 // we succeeded | |
| 2222 | |
| 2223 // notice that _finger == end cannot be guaranteed here since, | |
| 2224 // someone else might have moved the finger even further | |
| 2225 guarantee( _finger >= end, "the finger should have moved forward" ); | |
| 2226 | |
| 2227 if (verbose_low()) | |
| 2228 gclog_or_tty->print_cr("[%d] we were successful with region = " | |
| 2229 PTR_FORMAT, task_num, curr_region); | |
| 2230 | |
| 2231 if (limit > bottom) { | |
| 2232 if (verbose_low()) | |
| 2233 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is not empty, " | |
| 2234 "returning it ", task_num, curr_region); | |
| 2235 return curr_region; | |
| 2236 } else { | |
| 2237 tmp_guarantee_CM( limit == bottom, | |
| 2238 "the region limit should be at bottom" ); | |
| 2239 if (verbose_low()) | |
| 2240 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is empty, " | |
| 2241 "returning NULL", task_num, curr_region); | |
| 2242 // we return NULL and the caller should try calling | |
| 2243 // claim_region() again. | |
| 2244 return NULL; | |
| 2245 } | |
| 2246 } else { | |
| 2247 guarantee( _finger > finger, "the finger should have moved forward" ); | |
| 2248 if (verbose_low()) | |
| 2249 gclog_or_tty->print_cr("[%d] somebody else moved the finger, " | |
| 2250 "global finger = "PTR_FORMAT", " | |
| 2251 "our finger = "PTR_FORMAT, | |
| 2252 task_num, _finger, finger); | |
| 2253 | |
| 2254 // read it again | |
| 2255 finger = _finger; | |
| 2256 } | |
| 2257 } | |
| 2258 | |
| 2259 return NULL; | |
| 2260 } | |
| 2261 | |
| 2262 void ConcurrentMark::oops_do(OopClosure* cl) { | |
| 2263 if (_markStack.size() > 0 && verbose_low()) | |
| 2264 gclog_or_tty->print_cr("[global] scanning the global marking stack, " | |
| 2265 "size = %d", _markStack.size()); | |
| 2266 // we first iterate over the contents of the mark stack... | |
| 2267 _markStack.oops_do(cl); | |
| 2268 | |
| 2269 for (int i = 0; i < (int)_max_task_num; ++i) { | |
| 2270 OopTaskQueue* queue = _task_queues->queue((int)i); | |
| 2271 | |
| 2272 if (queue->size() > 0 && verbose_low()) | |
| 2273 gclog_or_tty->print_cr("[global] scanning task queue of task %d, " | |
| 2274 "size = %d", i, queue->size()); | |
| 2275 | |
| 2276 // ...then over the contents of the all the task queues. | |
| 2277 queue->oops_do(cl); | |
| 2278 } | |
| 2279 | |
| 2280 // finally, invalidate any entries that in the region stack that | |
| 2281 // point into the collection set | |
| 2282 if (_regionStack.invalidate_entries_into_cset()) { | |
| 2283 // otherwise, any gray objects copied during the evacuation pause | |
| 2284 // might not be visited. | |
| 2285 guarantee( _should_gray_objects, "invariant" ); | |
| 2286 } | |
| 2287 } | |
| 2288 | |
| 2289 void ConcurrentMark::clear_marking_state() { | |
| 2290 _markStack.setEmpty(); | |
| 2291 _markStack.clear_overflow(); | |
| 2292 _regionStack.setEmpty(); | |
| 2293 _regionStack.clear_overflow(); | |
| 2294 clear_has_overflown(); | |
| 2295 _finger = _heap_start; | |
| 2296 | |
| 2297 for (int i = 0; i < (int)_max_task_num; ++i) { | |
| 2298 OopTaskQueue* queue = _task_queues->queue(i); | |
| 2299 queue->set_empty(); | |
| 2300 } | |
| 2301 } | |
| 2302 | |
| 2303 void ConcurrentMark::print_stats() { | |
| 2304 if (verbose_stats()) { | |
| 2305 gclog_or_tty->print_cr("---------------------------------------------------------------------"); | |
| 2306 for (size_t i = 0; i < _active_tasks; ++i) { | |
| 2307 _tasks[i]->print_stats(); | |
| 2308 gclog_or_tty->print_cr("---------------------------------------------------------------------"); | |
| 2309 } | |
| 2310 } | |
| 2311 } | |
| 2312 | |
| 2313 class CSMarkOopClosure: public OopClosure { | |
| 2314 friend class CSMarkBitMapClosure; | |
| 2315 | |
| 2316 G1CollectedHeap* _g1h; | |
| 2317 CMBitMap* _bm; | |
| 2318 ConcurrentMark* _cm; | |
| 2319 oop* _ms; | |
| 2320 jint* _array_ind_stack; | |
| 2321 int _ms_size; | |
| 2322 int _ms_ind; | |
| 2323 int _array_increment; | |
| 2324 | |
| 2325 bool push(oop obj, int arr_ind = 0) { | |
| 2326 if (_ms_ind == _ms_size) { | |
| 2327 gclog_or_tty->print_cr("Mark stack is full."); | |
| 2328 return false; | |
| 2329 } | |
| 2330 _ms[_ms_ind] = obj; | |
| 2331 if (obj->is_objArray()) _array_ind_stack[_ms_ind] = arr_ind; | |
| 2332 _ms_ind++; | |
| 2333 return true; | |
| 2334 } | |
| 2335 | |
| 2336 oop pop() { | |
| 2337 if (_ms_ind == 0) return NULL; | |
| 2338 else { | |
| 2339 _ms_ind--; | |
| 2340 return _ms[_ms_ind]; | |
| 2341 } | |
| 2342 } | |
| 2343 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2344 template <class T> bool drain() { |
| 342 | 2345 while (_ms_ind > 0) { |
| 2346 oop obj = pop(); | |
| 2347 assert(obj != NULL, "Since index was non-zero."); | |
| 2348 if (obj->is_objArray()) { | |
| 2349 jint arr_ind = _array_ind_stack[_ms_ind]; | |
| 2350 objArrayOop aobj = objArrayOop(obj); | |
| 2351 jint len = aobj->length(); | |
| 2352 jint next_arr_ind = arr_ind + _array_increment; | |
| 2353 if (next_arr_ind < len) { | |
| 2354 push(obj, next_arr_ind); | |
| 2355 } | |
| 2356 // Now process this portion of this one. | |
| 2357 int lim = MIN2(next_arr_ind, len); | |
| 2358 for (int j = arr_ind; j < lim; j++) { | |
|
912
308762b2bf14
6872000: G1: compilation fails on linux/older gcc
apetrusenko
parents:
866
diff
changeset
|
2359 do_oop(aobj->objArrayOopDesc::obj_at_addr<T>(j)); |
| 342 | 2360 } |
| 2361 | |
| 2362 } else { | |
| 2363 obj->oop_iterate(this); | |
| 2364 } | |
| 2365 if (abort()) return false; | |
| 2366 } | |
| 2367 return true; | |
| 2368 } | |
| 2369 | |
| 2370 public: | |
| 2371 CSMarkOopClosure(ConcurrentMark* cm, int ms_size) : | |
| 2372 _g1h(G1CollectedHeap::heap()), | |
| 2373 _cm(cm), | |
| 2374 _bm(cm->nextMarkBitMap()), | |
| 2375 _ms_size(ms_size), _ms_ind(0), | |
| 2376 _ms(NEW_C_HEAP_ARRAY(oop, ms_size)), | |
| 2377 _array_ind_stack(NEW_C_HEAP_ARRAY(jint, ms_size)), | |
| 2378 _array_increment(MAX2(ms_size/8, 16)) | |
| 2379 {} | |
| 2380 | |
| 2381 ~CSMarkOopClosure() { | |
| 2382 FREE_C_HEAP_ARRAY(oop, _ms); | |
| 2383 FREE_C_HEAP_ARRAY(jint, _array_ind_stack); | |
| 2384 } | |
| 2385 | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2386 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2387 virtual void do_oop( oop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2388 |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2389 template <class T> void do_oop_work(T* p) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2390 T heap_oop = oopDesc::load_heap_oop(p); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2391 if (oopDesc::is_null(heap_oop)) return; |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2392 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
| 342 | 2393 if (obj->is_forwarded()) { |
| 2394 // If the object has already been forwarded, we have to make sure | |
| 2395 // that it's marked. So follow the forwarding pointer. Note that | |
| 2396 // this does the right thing for self-forwarding pointers in the | |
| 2397 // evacuation failure case. | |
| 2398 obj = obj->forwardee(); | |
| 2399 } | |
| 2400 HeapRegion* hr = _g1h->heap_region_containing(obj); | |
| 2401 if (hr != NULL) { | |
| 2402 if (hr->in_collection_set()) { | |
| 2403 if (_g1h->is_obj_ill(obj)) { | |
| 2404 _bm->mark((HeapWord*)obj); | |
| 2405 if (!push(obj)) { | |
| 2406 gclog_or_tty->print_cr("Setting abort in CSMarkOopClosure because push failed."); | |
| 2407 set_abort(); | |
| 2408 } | |
| 2409 } | |
| 2410 } else { | |
| 2411 // Outside the collection set; we need to gray it | |
| 2412 _cm->deal_with_reference(obj); | |
| 2413 } | |
| 2414 } | |
| 2415 } | |
| 2416 }; | |
| 2417 | |
| 2418 class CSMarkBitMapClosure: public BitMapClosure { | |
| 2419 G1CollectedHeap* _g1h; | |
| 2420 CMBitMap* _bitMap; | |
| 2421 ConcurrentMark* _cm; | |
| 2422 CSMarkOopClosure _oop_cl; | |
| 2423 public: | |
| 2424 CSMarkBitMapClosure(ConcurrentMark* cm, int ms_size) : | |
| 2425 _g1h(G1CollectedHeap::heap()), | |
| 2426 _bitMap(cm->nextMarkBitMap()), | |
| 2427 _oop_cl(cm, ms_size) | |
| 2428 {} | |
| 2429 | |
| 2430 ~CSMarkBitMapClosure() {} | |
| 2431 | |
| 2432 bool do_bit(size_t offset) { | |
| 2433 // convert offset into a HeapWord* | |
| 2434 HeapWord* addr = _bitMap->offsetToHeapWord(offset); | |
| 2435 assert(_bitMap->endWord() && addr < _bitMap->endWord(), | |
| 2436 "address out of range"); | |
| 2437 assert(_bitMap->isMarked(addr), "tautology"); | |
| 2438 oop obj = oop(addr); | |
| 2439 if (!obj->is_forwarded()) { | |
| 2440 if (!_oop_cl.push(obj)) return false; | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2441 if (UseCompressedOops) { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2442 if (!_oop_cl.drain<narrowOop>()) return false; |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2443 } else { |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2444 if (!_oop_cl.drain<oop>()) return false; |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2445 } |
| 342 | 2446 } |
| 2447 // Otherwise... | |
| 2448 return true; | |
| 2449 } | |
| 2450 }; | |
| 2451 | |
| 2452 | |
| 2453 class CompleteMarkingInCSHRClosure: public HeapRegionClosure { | |
| 2454 CMBitMap* _bm; | |
| 2455 CSMarkBitMapClosure _bit_cl; | |
| 2456 enum SomePrivateConstants { | |
| 2457 MSSize = 1000 | |
| 2458 }; | |
| 2459 bool _completed; | |
| 2460 public: | |
| 2461 CompleteMarkingInCSHRClosure(ConcurrentMark* cm) : | |
| 2462 _bm(cm->nextMarkBitMap()), | |
| 2463 _bit_cl(cm, MSSize), | |
| 2464 _completed(true) | |
| 2465 {} | |
| 2466 | |
| 2467 ~CompleteMarkingInCSHRClosure() {} | |
| 2468 | |
| 2469 bool doHeapRegion(HeapRegion* r) { | |
| 2470 if (!r->evacuation_failed()) { | |
| 2471 MemRegion mr = MemRegion(r->bottom(), r->next_top_at_mark_start()); | |
| 2472 if (!mr.is_empty()) { | |
| 2473 if (!_bm->iterate(&_bit_cl, mr)) { | |
| 2474 _completed = false; | |
| 2475 return true; | |
| 2476 } | |
| 2477 } | |
| 2478 } | |
| 2479 return false; | |
| 2480 } | |
| 2481 | |
| 2482 bool completed() { return _completed; } | |
| 2483 }; | |
| 2484 | |
| 2485 class ClearMarksInHRClosure: public HeapRegionClosure { | |
| 2486 CMBitMap* _bm; | |
| 2487 public: | |
| 2488 ClearMarksInHRClosure(CMBitMap* bm): _bm(bm) { } | |
| 2489 | |
| 2490 bool doHeapRegion(HeapRegion* r) { | |
| 2491 if (!r->used_region().is_empty() && !r->evacuation_failed()) { | |
| 2492 MemRegion usedMR = r->used_region(); | |
| 2493 _bm->clearRange(r->used_region()); | |
| 2494 } | |
| 2495 return false; | |
| 2496 } | |
| 2497 }; | |
| 2498 | |
| 2499 void ConcurrentMark::complete_marking_in_collection_set() { | |
| 2500 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
| 2501 | |
| 2502 if (!g1h->mark_in_progress()) { | |
| 2503 g1h->g1_policy()->record_mark_closure_time(0.0); | |
| 2504 return; | |
| 2505 } | |
| 2506 | |
| 2507 int i = 1; | |
| 2508 double start = os::elapsedTime(); | |
| 2509 while (true) { | |
| 2510 i++; | |
| 2511 CompleteMarkingInCSHRClosure cmplt(this); | |
| 2512 g1h->collection_set_iterate(&cmplt); | |
| 2513 if (cmplt.completed()) break; | |
| 2514 } | |
| 2515 double end_time = os::elapsedTime(); | |
| 2516 double elapsed_time_ms = (end_time - start) * 1000.0; | |
| 2517 g1h->g1_policy()->record_mark_closure_time(elapsed_time_ms); | |
| 2518 if (PrintGCDetails) { | |
| 2519 gclog_or_tty->print_cr("Mark closure took %5.2f ms.", elapsed_time_ms); | |
| 2520 } | |
| 2521 | |
| 2522 ClearMarksInHRClosure clr(nextMarkBitMap()); | |
| 2523 g1h->collection_set_iterate(&clr); | |
| 2524 } | |
| 2525 | |
| 2526 // The next two methods deal with the following optimisation. Some | |
| 2527 // objects are gray by being marked and located above the finger. If | |
| 2528 // they are copied, during an evacuation pause, below the finger then | |
| 2529 // the need to be pushed on the stack. The observation is that, if | |
| 2530 // there are no regions in the collection set located above the | |
| 2531 // finger, then the above cannot happen, hence we do not need to | |
| 2532 // explicitly gray any objects when copying them to below the | |
| 2533 // finger. The global stack will be scanned to ensure that, if it | |
| 2534 // points to objects being copied, it will update their | |
| 2535 // location. There is a tricky situation with the gray objects in | |
| 2536 // region stack that are being coped, however. See the comment in | |
| 2537 // newCSet(). | |
| 2538 | |
| 2539 void ConcurrentMark::newCSet() { | |
| 2540 if (!concurrent_marking_in_progress()) | |
| 2541 // nothing to do if marking is not in progress | |
| 2542 return; | |
| 2543 | |
| 2544 // find what the lowest finger is among the global and local fingers | |
| 2545 _min_finger = _finger; | |
| 2546 for (int i = 0; i < (int)_max_task_num; ++i) { | |
| 2547 CMTask* task = _tasks[i]; | |
| 2548 HeapWord* task_finger = task->finger(); | |
| 2549 if (task_finger != NULL && task_finger < _min_finger) | |
| 2550 _min_finger = task_finger; | |
| 2551 } | |
| 2552 | |
| 2553 _should_gray_objects = false; | |
| 2554 | |
| 2555 // This fixes a very subtle and fustrating bug. It might be the case | |
| 2556 // that, during en evacuation pause, heap regions that contain | |
| 2557 // objects that are gray (by being in regions contained in the | |
| 2558 // region stack) are included in the collection set. Since such gray | |
| 2559 // objects will be moved, and because it's not easy to redirect | |
| 2560 // region stack entries to point to a new location (because objects | |
| 2561 // in one region might be scattered to multiple regions after they | |
| 2562 // are copied), one option is to ensure that all marked objects | |
| 2563 // copied during a pause are pushed on the stack. Notice, however, | |
| 2564 // that this problem can only happen when the region stack is not | |
| 2565 // empty during an evacuation pause. So, we make the fix a bit less | |
| 2566 // conservative and ensure that regions are pushed on the stack, | |
| 2567 // irrespective whether all collection set regions are below the | |
| 2568 // finger, if the region stack is not empty. This is expected to be | |
| 2569 // a rare case, so I don't think it's necessary to be smarted about it. | |
| 2570 if (!region_stack_empty()) | |
| 2571 _should_gray_objects = true; | |
| 2572 } | |
| 2573 | |
| 2574 void ConcurrentMark::registerCSetRegion(HeapRegion* hr) { | |
| 2575 if (!concurrent_marking_in_progress()) | |
| 2576 return; | |
| 2577 | |
| 2578 HeapWord* region_end = hr->end(); | |
| 2579 if (region_end > _min_finger) | |
| 2580 _should_gray_objects = true; | |
| 2581 } | |
| 2582 | |
| 2583 // abandon current marking iteration due to a Full GC | |
| 2584 void ConcurrentMark::abort() { | |
| 2585 // Clear all marks to force marking thread to do nothing | |
| 2586 _nextMarkBitMap->clearAll(); | |
| 2587 // Empty mark stack | |
| 2588 clear_marking_state(); | |
| 2589 for (int i = 0; i < (int)_max_task_num; ++i) | |
| 2590 _tasks[i]->clear_region_fields(); | |
| 2591 _has_aborted = true; | |
| 2592 | |
| 2593 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 2594 satb_mq_set.abandon_partial_marking(); | |
| 2595 satb_mq_set.set_active_all_threads(false); | |
| 2596 } | |
| 2597 | |
| 2598 static void print_ms_time_info(const char* prefix, const char* name, | |
| 2599 NumberSeq& ns) { | |
| 2600 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", | |
| 2601 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); | |
| 2602 if (ns.num() > 0) { | |
| 2603 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", | |
| 2604 prefix, ns.sd(), ns.maximum()); | |
| 2605 } | |
| 2606 } | |
| 2607 | |
| 2608 void ConcurrentMark::print_summary_info() { | |
| 2609 gclog_or_tty->print_cr(" Concurrent marking:"); | |
| 2610 print_ms_time_info(" ", "init marks", _init_times); | |
| 2611 print_ms_time_info(" ", "remarks", _remark_times); | |
| 2612 { | |
| 2613 print_ms_time_info(" ", "final marks", _remark_mark_times); | |
| 2614 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); | |
| 2615 | |
| 2616 } | |
| 2617 print_ms_time_info(" ", "cleanups", _cleanup_times); | |
| 2618 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", | |
| 2619 _total_counting_time, | |
| 2620 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / | |
| 2621 (double)_cleanup_times.num() | |
| 2622 : 0.0)); | |
| 2623 if (G1ScrubRemSets) { | |
| 2624 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", | |
| 2625 _total_rs_scrub_time, | |
| 2626 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / | |
| 2627 (double)_cleanup_times.num() | |
| 2628 : 0.0)); | |
| 2629 } | |
| 2630 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", | |
| 2631 (_init_times.sum() + _remark_times.sum() + | |
| 2632 _cleanup_times.sum())/1000.0); | |
| 2633 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " | |
| 2634 "(%8.2f s marking, %8.2f s counting).", | |
| 2635 cmThread()->vtime_accum(), | |
| 2636 cmThread()->vtime_mark_accum(), | |
| 2637 cmThread()->vtime_count_accum()); | |
| 2638 } | |
| 2639 | |
| 2640 // Closures | |
| 2641 // XXX: there seems to be a lot of code duplication here; | |
| 2642 // should refactor and consolidate the shared code. | |
| 2643 | |
| 2644 // This closure is used to mark refs into the CMS generation in | |
| 2645 // the CMS bit map. Called at the first checkpoint. | |
| 2646 | |
| 2647 // We take a break if someone is trying to stop the world. | |
| 2648 bool ConcurrentMark::do_yield_check(int worker_i) { | |
| 2649 if (should_yield()) { | |
| 2650 if (worker_i == 0) | |
| 2651 _g1h->g1_policy()->record_concurrent_pause(); | |
| 2652 cmThread()->yield(); | |
| 2653 if (worker_i == 0) | |
| 2654 _g1h->g1_policy()->record_concurrent_pause_end(); | |
| 2655 return true; | |
| 2656 } else { | |
| 2657 return false; | |
| 2658 } | |
| 2659 } | |
| 2660 | |
| 2661 bool ConcurrentMark::should_yield() { | |
| 2662 return cmThread()->should_yield(); | |
| 2663 } | |
| 2664 | |
| 2665 bool ConcurrentMark::containing_card_is_marked(void* p) { | |
| 2666 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1); | |
| 2667 return _card_bm.at(offset >> CardTableModRefBS::card_shift); | |
| 2668 } | |
| 2669 | |
| 2670 bool ConcurrentMark::containing_cards_are_marked(void* start, | |
| 2671 void* last) { | |
| 2672 return | |
| 2673 containing_card_is_marked(start) && | |
| 2674 containing_card_is_marked(last); | |
| 2675 } | |
| 2676 | |
| 2677 #ifndef PRODUCT | |
| 2678 // for debugging purposes | |
| 2679 void ConcurrentMark::print_finger() { | |
| 2680 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, | |
| 2681 _heap_start, _heap_end, _finger); | |
| 2682 for (int i = 0; i < (int) _max_task_num; ++i) { | |
| 2683 gclog_or_tty->print(" %d: "PTR_FORMAT, i, _tasks[i]->finger()); | |
| 2684 } | |
| 2685 gclog_or_tty->print_cr(""); | |
| 2686 } | |
| 2687 #endif | |
| 2688 | |
| 2689 // Closure for iteration over bitmaps | |
| 2690 class CMBitMapClosure : public BitMapClosure { | |
| 2691 private: | |
| 2692 // the bitmap that is being iterated over | |
| 2693 CMBitMap* _nextMarkBitMap; | |
| 2694 ConcurrentMark* _cm; | |
| 2695 CMTask* _task; | |
| 2696 // true if we're scanning a heap region claimed by the task (so that | |
| 2697 // we move the finger along), false if we're not, i.e. currently when | |
| 2698 // scanning a heap region popped from the region stack (so that we | |
| 2699 // do not move the task finger along; it'd be a mistake if we did so). | |
| 2700 bool _scanning_heap_region; | |
| 2701 | |
| 2702 public: | |
| 2703 CMBitMapClosure(CMTask *task, | |
| 2704 ConcurrentMark* cm, | |
| 2705 CMBitMap* nextMarkBitMap) | |
| 2706 : _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } | |
| 2707 | |
| 2708 void set_scanning_heap_region(bool scanning_heap_region) { | |
| 2709 _scanning_heap_region = scanning_heap_region; | |
| 2710 } | |
| 2711 | |
| 2712 bool do_bit(size_t offset) { | |
| 2713 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); | |
| 2714 tmp_guarantee_CM( _nextMarkBitMap->isMarked(addr), "invariant" ); | |
| 2715 tmp_guarantee_CM( addr < _cm->finger(), "invariant" ); | |
| 2716 | |
| 2717 if (_scanning_heap_region) { | |
| 2718 statsOnly( _task->increase_objs_found_on_bitmap() ); | |
| 2719 tmp_guarantee_CM( addr >= _task->finger(), "invariant" ); | |
| 2720 // We move that task's local finger along. | |
| 2721 _task->move_finger_to(addr); | |
| 2722 } else { | |
| 2723 // We move the task's region finger along. | |
| 2724 _task->move_region_finger_to(addr); | |
| 2725 } | |
| 2726 | |
| 2727 _task->scan_object(oop(addr)); | |
| 2728 // we only partially drain the local queue and global stack | |
| 2729 _task->drain_local_queue(true); | |
| 2730 _task->drain_global_stack(true); | |
| 2731 | |
| 2732 // if the has_aborted flag has been raised, we need to bail out of | |
| 2733 // the iteration | |
| 2734 return !_task->has_aborted(); | |
| 2735 } | |
| 2736 }; | |
| 2737 | |
| 2738 // Closure for iterating over objects, currently only used for | |
| 2739 // processing SATB buffers. | |
| 2740 class CMObjectClosure : public ObjectClosure { | |
| 2741 private: | |
| 2742 CMTask* _task; | |
| 2743 | |
| 2744 public: | |
| 2745 void do_object(oop obj) { | |
| 2746 _task->deal_with_reference(obj); | |
| 2747 } | |
| 2748 | |
| 2749 CMObjectClosure(CMTask* task) : _task(task) { } | |
| 2750 }; | |
| 2751 | |
| 2752 // Closure for iterating over object fields | |
| 2753 class CMOopClosure : public OopClosure { | |
| 2754 private: | |
| 2755 G1CollectedHeap* _g1h; | |
| 2756 ConcurrentMark* _cm; | |
| 2757 CMTask* _task; | |
| 2758 | |
| 2759 public: | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2760 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2761 virtual void do_oop( oop* p) { do_oop_work(p); } |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2762 |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2763 template <class T> void do_oop_work(T* p) { |
| 342 | 2764 tmp_guarantee_CM( _g1h->is_in_g1_reserved((HeapWord*) p), "invariant" ); |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2765 tmp_guarantee_CM( !_g1h->heap_region_containing((HeapWord*) p)->is_on_free_list(), "invariant" ); |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2766 |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2767 oop obj = oopDesc::load_decode_heap_oop(p); |
| 342 | 2768 if (_cm->verbose_high()) |
| 2769 gclog_or_tty->print_cr("[%d] we're looking at location " | |
| 2770 "*"PTR_FORMAT" = "PTR_FORMAT, | |
| 2771 _task->task_id(), p, (void*) obj); | |
| 2772 _task->deal_with_reference(obj); | |
| 2773 } | |
| 2774 | |
| 2775 CMOopClosure(G1CollectedHeap* g1h, | |
| 2776 ConcurrentMark* cm, | |
| 2777 CMTask* task) | |
| 2778 : _g1h(g1h), _cm(cm), _task(task) { } | |
| 2779 }; | |
| 2780 | |
| 2781 void CMTask::setup_for_region(HeapRegion* hr) { | |
| 2782 tmp_guarantee_CM( hr != NULL && !hr->continuesHumongous(), | |
| 2783 "claim_region() should have filtered out continues humongous regions" ); | |
| 2784 | |
| 2785 if (_cm->verbose_low()) | |
| 2786 gclog_or_tty->print_cr("[%d] setting up for region "PTR_FORMAT, | |
| 2787 _task_id, hr); | |
| 2788 | |
| 2789 _curr_region = hr; | |
| 2790 _finger = hr->bottom(); | |
| 2791 update_region_limit(); | |
| 2792 } | |
| 2793 | |
| 2794 void CMTask::update_region_limit() { | |
| 2795 HeapRegion* hr = _curr_region; | |
| 2796 HeapWord* bottom = hr->bottom(); | |
| 2797 HeapWord* limit = hr->next_top_at_mark_start(); | |
| 2798 | |
| 2799 if (limit == bottom) { | |
| 2800 if (_cm->verbose_low()) | |
| 2801 gclog_or_tty->print_cr("[%d] found an empty region " | |
| 2802 "["PTR_FORMAT", "PTR_FORMAT")", | |
| 2803 _task_id, bottom, limit); | |
| 2804 // The region was collected underneath our feet. | |
| 2805 // We set the finger to bottom to ensure that the bitmap | |
| 2806 // iteration that will follow this will not do anything. | |
| 2807 // (this is not a condition that holds when we set the region up, | |
| 2808 // as the region is not supposed to be empty in the first place) | |
| 2809 _finger = bottom; | |
| 2810 } else if (limit >= _region_limit) { | |
| 2811 tmp_guarantee_CM( limit >= _finger, "peace of mind" ); | |
| 2812 } else { | |
| 2813 tmp_guarantee_CM( limit < _region_limit, "only way to get here" ); | |
| 2814 // This can happen under some pretty unusual circumstances. An | |
| 2815 // evacuation pause empties the region underneath our feet (NTAMS | |
| 2816 // at bottom). We then do some allocation in the region (NTAMS | |
| 2817 // stays at bottom), followed by the region being used as a GC | |
| 2818 // alloc region (NTAMS will move to top() and the objects | |
| 2819 // originally below it will be grayed). All objects now marked in | |
| 2820 // the region are explicitly grayed, if below the global finger, | |
| 2821 // and we do not need in fact to scan anything else. So, we simply | |
| 2822 // set _finger to be limit to ensure that the bitmap iteration | |
| 2823 // doesn't do anything. | |
| 2824 _finger = limit; | |
| 2825 } | |
| 2826 | |
| 2827 _region_limit = limit; | |
| 2828 } | |
| 2829 | |
| 2830 void CMTask::giveup_current_region() { | |
| 2831 tmp_guarantee_CM( _curr_region != NULL, "invariant" ); | |
| 2832 if (_cm->verbose_low()) | |
| 2833 gclog_or_tty->print_cr("[%d] giving up region "PTR_FORMAT, | |
| 2834 _task_id, _curr_region); | |
| 2835 clear_region_fields(); | |
| 2836 } | |
| 2837 | |
| 2838 void CMTask::clear_region_fields() { | |
| 2839 // Values for these three fields that indicate that we're not | |
| 2840 // holding on to a region. | |
| 2841 _curr_region = NULL; | |
| 2842 _finger = NULL; | |
| 2843 _region_limit = NULL; | |
| 2844 | |
| 2845 _region_finger = NULL; | |
| 2846 } | |
| 2847 | |
| 2848 void CMTask::reset(CMBitMap* nextMarkBitMap) { | |
| 2849 guarantee( nextMarkBitMap != NULL, "invariant" ); | |
| 2850 | |
| 2851 if (_cm->verbose_low()) | |
| 2852 gclog_or_tty->print_cr("[%d] resetting", _task_id); | |
| 2853 | |
| 2854 _nextMarkBitMap = nextMarkBitMap; | |
| 2855 clear_region_fields(); | |
| 2856 | |
| 2857 _calls = 0; | |
| 2858 _elapsed_time_ms = 0.0; | |
| 2859 _termination_time_ms = 0.0; | |
| 2860 _termination_start_time_ms = 0.0; | |
| 2861 | |
| 2862 #if _MARKING_STATS_ | |
| 2863 _local_pushes = 0; | |
| 2864 _local_pops = 0; | |
| 2865 _local_max_size = 0; | |
| 2866 _objs_scanned = 0; | |
| 2867 _global_pushes = 0; | |
| 2868 _global_pops = 0; | |
| 2869 _global_max_size = 0; | |
| 2870 _global_transfers_to = 0; | |
| 2871 _global_transfers_from = 0; | |
| 2872 _region_stack_pops = 0; | |
| 2873 _regions_claimed = 0; | |
| 2874 _objs_found_on_bitmap = 0; | |
| 2875 _satb_buffers_processed = 0; | |
| 2876 _steal_attempts = 0; | |
| 2877 _steals = 0; | |
| 2878 _aborted = 0; | |
| 2879 _aborted_overflow = 0; | |
| 2880 _aborted_cm_aborted = 0; | |
| 2881 _aborted_yield = 0; | |
| 2882 _aborted_timed_out = 0; | |
| 2883 _aborted_satb = 0; | |
| 2884 _aborted_termination = 0; | |
| 2885 #endif // _MARKING_STATS_ | |
| 2886 } | |
| 2887 | |
| 2888 bool CMTask::should_exit_termination() { | |
| 2889 regular_clock_call(); | |
| 2890 // This is called when we are in the termination protocol. We should | |
| 2891 // quit if, for some reason, this task wants to abort or the global | |
| 2892 // stack is not empty (this means that we can get work from it). | |
| 2893 return !_cm->mark_stack_empty() || has_aborted(); | |
| 2894 } | |
| 2895 | |
| 2896 // This determines whether the method below will check both the local | |
| 2897 // and global fingers when determining whether to push on the stack a | |
| 2898 // gray object (value 1) or whether it will only check the global one | |
| 2899 // (value 0). The tradeoffs are that the former will be a bit more | |
| 2900 // accurate and possibly push less on the stack, but it might also be | |
| 2901 // a little bit slower. | |
| 2902 | |
| 2903 #define _CHECK_BOTH_FINGERS_ 1 | |
| 2904 | |
| 2905 void CMTask::deal_with_reference(oop obj) { | |
| 2906 if (_cm->verbose_high()) | |
| 2907 gclog_or_tty->print_cr("[%d] we're dealing with reference = "PTR_FORMAT, | |
| 2908 _task_id, (void*) obj); | |
| 2909 | |
| 2910 ++_refs_reached; | |
| 2911 | |
| 2912 HeapWord* objAddr = (HeapWord*) obj; | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2913 assert(obj->is_oop_or_null(true /* ignore mark word */), "Error"); |
| 342 | 2914 if (_g1h->is_in_g1_reserved(objAddr)) { |
| 2915 tmp_guarantee_CM( obj != NULL, "is_in_g1_reserved should ensure this" ); | |
| 2916 HeapRegion* hr = _g1h->heap_region_containing(obj); | |
| 2917 if (_g1h->is_obj_ill(obj, hr)) { | |
| 2918 if (_cm->verbose_high()) | |
| 2919 gclog_or_tty->print_cr("[%d] "PTR_FORMAT" is not considered marked", | |
| 2920 _task_id, (void*) obj); | |
| 2921 | |
| 2922 // we need to mark it first | |
| 2923 if (_nextMarkBitMap->parMark(objAddr)) { | |
| 2924 // No OrderAccess:store_load() is needed. It is implicit in the | |
| 2925 // CAS done in parMark(objAddr) above | |
| 2926 HeapWord* global_finger = _cm->finger(); | |
| 2927 | |
| 2928 #if _CHECK_BOTH_FINGERS_ | |
| 2929 // we will check both the local and global fingers | |
| 2930 | |
| 2931 if (_finger != NULL && objAddr < _finger) { | |
| 2932 if (_cm->verbose_high()) | |
| 2933 gclog_or_tty->print_cr("[%d] below the local finger ("PTR_FORMAT"), " | |
| 2934 "pushing it", _task_id, _finger); | |
| 2935 push(obj); | |
| 2936 } else if (_curr_region != NULL && objAddr < _region_limit) { | |
| 2937 // do nothing | |
| 2938 } else if (objAddr < global_finger) { | |
| 2939 // Notice that the global finger might be moving forward | |
| 2940 // concurrently. This is not a problem. In the worst case, we | |
| 2941 // mark the object while it is above the global finger and, by | |
| 2942 // the time we read the global finger, it has moved forward | |
| 2943 // passed this object. In this case, the object will probably | |
| 2944 // be visited when a task is scanning the region and will also | |
| 2945 // be pushed on the stack. So, some duplicate work, but no | |
| 2946 // correctness problems. | |
| 2947 | |
| 2948 if (_cm->verbose_high()) | |
| 2949 gclog_or_tty->print_cr("[%d] below the global finger " | |
| 2950 "("PTR_FORMAT"), pushing it", | |
| 2951 _task_id, global_finger); | |
| 2952 push(obj); | |
| 2953 } else { | |
| 2954 // do nothing | |
| 2955 } | |
| 2956 #else // _CHECK_BOTH_FINGERS_ | |
| 2957 // we will only check the global finger | |
| 2958 | |
| 2959 if (objAddr < global_finger) { | |
| 2960 // see long comment above | |
| 2961 | |
| 2962 if (_cm->verbose_high()) | |
| 2963 gclog_or_tty->print_cr("[%d] below the global finger " | |
| 2964 "("PTR_FORMAT"), pushing it", | |
| 2965 _task_id, global_finger); | |
| 2966 push(obj); | |
| 2967 } | |
| 2968 #endif // _CHECK_BOTH_FINGERS_ | |
| 2969 } | |
| 2970 } | |
| 2971 } | |
| 2972 } | |
| 2973 | |
| 2974 void CMTask::push(oop obj) { | |
| 2975 HeapWord* objAddr = (HeapWord*) obj; | |
| 2976 tmp_guarantee_CM( _g1h->is_in_g1_reserved(objAddr), "invariant" ); | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2977 tmp_guarantee_CM( !_g1h->heap_region_containing(objAddr)->is_on_free_list(), "invariant" ); |
| 342 | 2978 tmp_guarantee_CM( !_g1h->is_obj_ill(obj), "invariant" ); |
| 2979 tmp_guarantee_CM( _nextMarkBitMap->isMarked(objAddr), "invariant" ); | |
| 2980 | |
| 2981 if (_cm->verbose_high()) | |
| 2982 gclog_or_tty->print_cr("[%d] pushing "PTR_FORMAT, _task_id, (void*) obj); | |
| 2983 | |
| 2984 if (!_task_queue->push(obj)) { | |
| 2985 // The local task queue looks full. We need to push some entries | |
| 2986 // to the global stack. | |
| 2987 | |
| 2988 if (_cm->verbose_medium()) | |
| 2989 gclog_or_tty->print_cr("[%d] task queue overflow, " | |
| 2990 "moving entries to the global stack", | |
| 2991 _task_id); | |
| 2992 move_entries_to_global_stack(); | |
| 2993 | |
| 2994 // this should succeed since, even if we overflow the global | |
| 2995 // stack, we should have definitely removed some entries from the | |
| 2996 // local queue. So, there must be space on it. | |
| 2997 bool success = _task_queue->push(obj); | |
| 2998 tmp_guarantee_CM( success, "invariant" ); | |
| 2999 } | |
| 3000 | |
| 3001 statsOnly( int tmp_size = _task_queue->size(); | |
| 3002 if (tmp_size > _local_max_size) | |
| 3003 _local_max_size = tmp_size; | |
| 3004 ++_local_pushes ); | |
| 3005 } | |
| 3006 | |
| 3007 void CMTask::reached_limit() { | |
| 3008 tmp_guarantee_CM( _words_scanned >= _words_scanned_limit || | |
| 3009 _refs_reached >= _refs_reached_limit , | |
| 3010 "shouldn't have been called otherwise" ); | |
| 3011 regular_clock_call(); | |
| 3012 } | |
| 3013 | |
| 3014 void CMTask::regular_clock_call() { | |
| 3015 if (has_aborted()) | |
| 3016 return; | |
| 3017 | |
| 3018 // First, we need to recalculate the words scanned and refs reached | |
| 3019 // limits for the next clock call. | |
| 3020 recalculate_limits(); | |
| 3021 | |
| 3022 // During the regular clock call we do the following | |
| 3023 | |
| 3024 // (1) If an overflow has been flagged, then we abort. | |
| 3025 if (_cm->has_overflown()) { | |
| 3026 set_has_aborted(); | |
| 3027 return; | |
| 3028 } | |
| 3029 | |
| 3030 // If we are not concurrent (i.e. we're doing remark) we don't need | |
| 3031 // to check anything else. The other steps are only needed during | |
| 3032 // the concurrent marking phase. | |
| 3033 if (!concurrent()) | |
| 3034 return; | |
| 3035 | |
| 3036 // (2) If marking has been aborted for Full GC, then we also abort. | |
| 3037 if (_cm->has_aborted()) { | |
| 3038 set_has_aborted(); | |
| 3039 statsOnly( ++_aborted_cm_aborted ); | |
| 3040 return; | |
| 3041 } | |
| 3042 | |
| 3043 double curr_time_ms = os::elapsedVTime() * 1000.0; | |
| 3044 | |
| 3045 // (3) If marking stats are enabled, then we update the step history. | |
| 3046 #if _MARKING_STATS_ | |
| 3047 if (_words_scanned >= _words_scanned_limit) | |
| 3048 ++_clock_due_to_scanning; | |
| 3049 if (_refs_reached >= _refs_reached_limit) | |
| 3050 ++_clock_due_to_marking; | |
| 3051 | |
| 3052 double last_interval_ms = curr_time_ms - _interval_start_time_ms; | |
| 3053 _interval_start_time_ms = curr_time_ms; | |
| 3054 _all_clock_intervals_ms.add(last_interval_ms); | |
| 3055 | |
| 3056 if (_cm->verbose_medium()) { | |
| 3057 gclog_or_tty->print_cr("[%d] regular clock, interval = %1.2lfms, " | |
| 3058 "scanned = %d%s, refs reached = %d%s", | |
| 3059 _task_id, last_interval_ms, | |
| 3060 _words_scanned, | |
| 3061 (_words_scanned >= _words_scanned_limit) ? " (*)" : "", | |
| 3062 _refs_reached, | |
| 3063 (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); | |
| 3064 } | |
| 3065 #endif // _MARKING_STATS_ | |
| 3066 | |
| 3067 // (4) We check whether we should yield. If we have to, then we abort. | |
| 3068 if (_cm->should_yield()) { | |
| 3069 // We should yield. To do this we abort the task. The caller is | |
| 3070 // responsible for yielding. | |
| 3071 set_has_aborted(); | |
| 3072 statsOnly( ++_aborted_yield ); | |
| 3073 return; | |
| 3074 } | |
| 3075 | |
| 3076 // (5) We check whether we've reached our time quota. If we have, | |
| 3077 // then we abort. | |
| 3078 double elapsed_time_ms = curr_time_ms - _start_time_ms; | |
| 3079 if (elapsed_time_ms > _time_target_ms) { | |
| 3080 set_has_aborted(); | |
| 3081 _has_aborted_timed_out = true; | |
| 3082 statsOnly( ++_aborted_timed_out ); | |
| 3083 return; | |
| 3084 } | |
| 3085 | |
| 3086 // (6) Finally, we check whether there are enough completed STAB | |
| 3087 // buffers available for processing. If there are, we abort. | |
| 3088 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 3089 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { | |
| 3090 if (_cm->verbose_low()) | |
| 3091 gclog_or_tty->print_cr("[%d] aborting to deal with pending SATB buffers", | |
| 3092 _task_id); | |
| 3093 // we do need to process SATB buffers, we'll abort and restart | |
| 3094 // the marking task to do so | |
| 3095 set_has_aborted(); | |
| 3096 statsOnly( ++_aborted_satb ); | |
| 3097 return; | |
| 3098 } | |
| 3099 } | |
| 3100 | |
| 3101 void CMTask::recalculate_limits() { | |
| 3102 _real_words_scanned_limit = _words_scanned + words_scanned_period; | |
| 3103 _words_scanned_limit = _real_words_scanned_limit; | |
| 3104 | |
| 3105 _real_refs_reached_limit = _refs_reached + refs_reached_period; | |
| 3106 _refs_reached_limit = _real_refs_reached_limit; | |
| 3107 } | |
| 3108 | |
| 3109 void CMTask::decrease_limits() { | |
| 3110 // This is called when we believe that we're going to do an infrequent | |
| 3111 // operation which will increase the per byte scanned cost (i.e. move | |
| 3112 // entries to/from the global stack). It basically tries to decrease the | |
| 3113 // scanning limit so that the clock is called earlier. | |
| 3114 | |
| 3115 if (_cm->verbose_medium()) | |
| 3116 gclog_or_tty->print_cr("[%d] decreasing limits", _task_id); | |
| 3117 | |
| 3118 _words_scanned_limit = _real_words_scanned_limit - | |
| 3119 3 * words_scanned_period / 4; | |
| 3120 _refs_reached_limit = _real_refs_reached_limit - | |
| 3121 3 * refs_reached_period / 4; | |
| 3122 } | |
| 3123 | |
| 3124 void CMTask::move_entries_to_global_stack() { | |
| 3125 // local array where we'll store the entries that will be popped | |
| 3126 // from the local queue | |
| 3127 oop buffer[global_stack_transfer_size]; | |
| 3128 | |
| 3129 int n = 0; | |
| 3130 oop obj; | |
| 3131 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { | |
| 3132 buffer[n] = obj; | |
| 3133 ++n; | |
| 3134 } | |
| 3135 | |
| 3136 if (n > 0) { | |
| 3137 // we popped at least one entry from the local queue | |
| 3138 | |
| 3139 statsOnly( ++_global_transfers_to; _local_pops += n ); | |
| 3140 | |
| 3141 if (!_cm->mark_stack_push(buffer, n)) { | |
| 3142 if (_cm->verbose_low()) | |
| 3143 gclog_or_tty->print_cr("[%d] aborting due to global stack overflow", _task_id); | |
| 3144 set_has_aborted(); | |
| 3145 } else { | |
| 3146 // the transfer was successful | |
| 3147 | |
| 3148 if (_cm->verbose_medium()) | |
| 3149 gclog_or_tty->print_cr("[%d] pushed %d entries to the global stack", | |
| 3150 _task_id, n); | |
| 3151 statsOnly( int tmp_size = _cm->mark_stack_size(); | |
| 3152 if (tmp_size > _global_max_size) | |
| 3153 _global_max_size = tmp_size; | |
| 3154 _global_pushes += n ); | |
| 3155 } | |
| 3156 } | |
| 3157 | |
| 3158 // this operation was quite expensive, so decrease the limits | |
| 3159 decrease_limits(); | |
| 3160 } | |
| 3161 | |
| 3162 void CMTask::get_entries_from_global_stack() { | |
| 3163 // local array where we'll store the entries that will be popped | |
| 3164 // from the global stack. | |
| 3165 oop buffer[global_stack_transfer_size]; | |
| 3166 int n; | |
| 3167 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); | |
| 3168 tmp_guarantee_CM( n <= global_stack_transfer_size, | |
| 3169 "we should not pop more than the given limit" ); | |
| 3170 if (n > 0) { | |
| 3171 // yes, we did actually pop at least one entry | |
| 3172 | |
| 3173 statsOnly( ++_global_transfers_from; _global_pops += n ); | |
| 3174 if (_cm->verbose_medium()) | |
| 3175 gclog_or_tty->print_cr("[%d] popped %d entries from the global stack", | |
| 3176 _task_id, n); | |
| 3177 for (int i = 0; i < n; ++i) { | |
| 3178 bool success = _task_queue->push(buffer[i]); | |
| 3179 // We only call this when the local queue is empty or under a | |
| 3180 // given target limit. So, we do not expect this push to fail. | |
| 3181 tmp_guarantee_CM( success, "invariant" ); | |
| 3182 } | |
| 3183 | |
| 3184 statsOnly( int tmp_size = _task_queue->size(); | |
| 3185 if (tmp_size > _local_max_size) | |
| 3186 _local_max_size = tmp_size; | |
| 3187 _local_pushes += n ); | |
| 3188 } | |
| 3189 | |
| 3190 // this operation was quite expensive, so decrease the limits | |
| 3191 decrease_limits(); | |
| 3192 } | |
| 3193 | |
| 3194 void CMTask::drain_local_queue(bool partially) { | |
| 3195 if (has_aborted()) | |
| 3196 return; | |
| 3197 | |
| 3198 // Decide what the target size is, depending whether we're going to | |
| 3199 // drain it partially (so that other tasks can steal if they run out | |
| 3200 // of things to do) or totally (at the very end). | |
| 3201 size_t target_size; | |
| 3202 if (partially) | |
| 3203 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); | |
| 3204 else | |
| 3205 target_size = 0; | |
| 3206 | |
| 3207 if (_task_queue->size() > target_size) { | |
| 3208 if (_cm->verbose_high()) | |
| 3209 gclog_or_tty->print_cr("[%d] draining local queue, target size = %d", | |
| 3210 _task_id, target_size); | |
| 3211 | |
| 3212 oop obj; | |
| 3213 bool ret = _task_queue->pop_local(obj); | |
| 3214 while (ret) { | |
| 3215 statsOnly( ++_local_pops ); | |
| 3216 | |
| 3217 if (_cm->verbose_high()) | |
| 3218 gclog_or_tty->print_cr("[%d] popped "PTR_FORMAT, _task_id, | |
| 3219 (void*) obj); | |
| 3220 | |
| 3221 tmp_guarantee_CM( _g1h->is_in_g1_reserved((HeapWord*) obj), | |
| 3222 "invariant" ); | |
|
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
3223 tmp_guarantee_CM( !_g1h->heap_region_containing(obj)->is_on_free_list(), |
|
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
3224 "invariant" ); |
| 342 | 3225 |
| 3226 scan_object(obj); | |
| 3227 | |
| 3228 if (_task_queue->size() <= target_size || has_aborted()) | |
| 3229 ret = false; | |
| 3230 else | |
| 3231 ret = _task_queue->pop_local(obj); | |
| 3232 } | |
| 3233 | |
| 3234 if (_cm->verbose_high()) | |
| 3235 gclog_or_tty->print_cr("[%d] drained local queue, size = %d", | |
| 3236 _task_id, _task_queue->size()); | |
| 3237 } | |
| 3238 } | |
| 3239 | |
| 3240 void CMTask::drain_global_stack(bool partially) { | |
| 3241 if (has_aborted()) | |
| 3242 return; | |
| 3243 | |
| 3244 // We have a policy to drain the local queue before we attempt to | |
| 3245 // drain the global stack. | |
| 3246 tmp_guarantee_CM( partially || _task_queue->size() == 0, "invariant" ); | |
| 3247 | |
| 3248 // Decide what the target size is, depending whether we're going to | |
| 3249 // drain it partially (so that other tasks can steal if they run out | |
| 3250 // of things to do) or totally (at the very end). Notice that, | |
| 3251 // because we move entries from the global stack in chunks or | |
| 3252 // because another task might be doing the same, we might in fact | |
| 3253 // drop below the target. But, this is not a problem. | |
| 3254 size_t target_size; | |
| 3255 if (partially) | |
| 3256 target_size = _cm->partial_mark_stack_size_target(); | |
| 3257 else | |
| 3258 target_size = 0; | |
| 3259 | |
| 3260 if (_cm->mark_stack_size() > target_size) { | |
| 3261 if (_cm->verbose_low()) | |
| 3262 gclog_or_tty->print_cr("[%d] draining global_stack, target size %d", | |
| 3263 _task_id, target_size); | |
| 3264 | |
| 3265 while (!has_aborted() && _cm->mark_stack_size() > target_size) { | |
| 3266 get_entries_from_global_stack(); | |
| 3267 drain_local_queue(partially); | |
| 3268 } | |
| 3269 | |
| 3270 if (_cm->verbose_low()) | |
| 3271 gclog_or_tty->print_cr("[%d] drained global stack, size = %d", | |
| 3272 _task_id, _cm->mark_stack_size()); | |
| 3273 } | |
| 3274 } | |
| 3275 | |
| 3276 // SATB Queue has several assumptions on whether to call the par or | |
| 3277 // non-par versions of the methods. this is why some of the code is | |
| 3278 // replicated. We should really get rid of the single-threaded version | |
| 3279 // of the code to simplify things. | |
| 3280 void CMTask::drain_satb_buffers() { | |
| 3281 if (has_aborted()) | |
| 3282 return; | |
| 3283 | |
| 3284 // We set this so that the regular clock knows that we're in the | |
| 3285 // middle of draining buffers and doesn't set the abort flag when it | |
| 3286 // notices that SATB buffers are available for draining. It'd be | |
| 3287 // very counter productive if it did that. :-) | |
| 3288 _draining_satb_buffers = true; | |
| 3289 | |
| 3290 CMObjectClosure oc(this); | |
| 3291 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
| 3292 if (ParallelGCThreads > 0) | |
| 3293 satb_mq_set.set_par_closure(_task_id, &oc); | |
| 3294 else | |
| 3295 satb_mq_set.set_closure(&oc); | |
| 3296 | |
| 3297 // This keeps claiming and applying the closure to completed buffers | |
| 3298 // until we run out of buffers or we need to abort. | |
| 3299 if (ParallelGCThreads > 0) { | |
| 3300 while (!has_aborted() && | |
| 3301 satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) { | |
| 3302 if (_cm->verbose_medium()) | |
| 3303 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); | |
| 3304 statsOnly( ++_satb_buffers_processed ); | |
| 3305 regular_clock_call(); | |
| 3306 } | |
| 3307 } else { | |
| 3308 while (!has_aborted() && | |
| 3309 satb_mq_set.apply_closure_to_completed_buffer()) { | |
| 3310 if (_cm->verbose_medium()) | |
| 3311 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); | |
| 3312 statsOnly( ++_satb_buffers_processed ); | |
| 3313 regular_clock_call(); | |
| 3314 } | |
| 3315 } | |
| 3316 | |
| 3317 if (!concurrent() && !has_aborted()) { | |
| 3318 // We should only do this during remark. | |
| 3319 if (ParallelGCThreads > 0) | |
| 3320 satb_mq_set.par_iterate_closure_all_threads(_task_id); | |
| 3321 else | |
| 3322 satb_mq_set.iterate_closure_all_threads(); | |
| 3323 } | |
| 3324 | |
| 3325 _draining_satb_buffers = false; | |
| 3326 | |
| 3327 tmp_guarantee_CM( has_aborted() || | |
| 3328 concurrent() || | |
| 3329 satb_mq_set.completed_buffers_num() == 0, "invariant" ); | |
| 3330 | |
| 3331 if (ParallelGCThreads > 0) | |
| 3332 satb_mq_set.set_par_closure(_task_id, NULL); | |
| 3333 else | |
| 3334 satb_mq_set.set_closure(NULL); | |
| 3335 | |
| 3336 // again, this was a potentially expensive operation, decrease the | |
| 3337 // limits to get the regular clock call early | |
| 3338 decrease_limits(); | |
| 3339 } | |
| 3340 | |
| 3341 void CMTask::drain_region_stack(BitMapClosure* bc) { | |
| 3342 if (has_aborted()) | |
| 3343 return; | |
| 3344 | |
| 3345 tmp_guarantee_CM( _region_finger == NULL, | |
| 3346 "it should be NULL when we're not scanning a region" ); | |
| 3347 | |
| 3348 if (!_cm->region_stack_empty()) { | |
| 3349 if (_cm->verbose_low()) | |
| 3350 gclog_or_tty->print_cr("[%d] draining region stack, size = %d", | |
| 3351 _task_id, _cm->region_stack_size()); | |
| 3352 | |
| 3353 MemRegion mr = _cm->region_stack_pop(); | |
| 3354 // it returns MemRegion() if the pop fails | |
| 3355 statsOnly(if (mr.start() != NULL) ++_region_stack_pops ); | |
| 3356 | |
| 3357 while (mr.start() != NULL) { | |
| 3358 if (_cm->verbose_medium()) | |
| 3359 gclog_or_tty->print_cr("[%d] we are scanning region " | |
| 3360 "["PTR_FORMAT", "PTR_FORMAT")", | |
| 3361 _task_id, mr.start(), mr.end()); | |
| 3362 tmp_guarantee_CM( mr.end() <= _cm->finger(), | |
| 3363 "otherwise the region shouldn't be on the stack" ); | |
| 3364 assert(!mr.is_empty(), "Only non-empty regions live on the region stack"); | |
| 3365 if (_nextMarkBitMap->iterate(bc, mr)) { | |
| 3366 tmp_guarantee_CM( !has_aborted(), | |
| 3367 "cannot abort the task without aborting the bitmap iteration" ); | |
| 3368 | |
| 3369 // We finished iterating over the region without aborting. | |
| 3370 regular_clock_call(); | |
| 3371 if (has_aborted()) | |
| 3372 mr = MemRegion(); | |
| 3373 else { | |
| 3374 mr = _cm->region_stack_pop(); | |
| 3375 // it returns MemRegion() if the pop fails | |
| 3376 statsOnly(if (mr.start() != NULL) ++_region_stack_pops ); | |
| 3377 } | |
| 3378 } else { | |
| 3379 guarantee( has_aborted(), "currently the only way to do so" ); | |
| 3380 | |
| 3381 // The only way to abort the bitmap iteration is to return | |
| 3382 // false from the do_bit() method. However, inside the | |
| 3383 // do_bit() method we move the _region_finger to point to the | |
| 3384 // object currently being looked at. So, if we bail out, we | |
| 3385 // have definitely set _region_finger to something non-null. | |
| 3386 guarantee( _region_finger != NULL, "invariant" ); | |
| 3387 | |
| 3388 // The iteration was actually aborted. So now _region_finger | |
| 3389 // points to the address of the object we last scanned. If we | |
| 3390 // leave it there, when we restart this task, we will rescan | |
| 3391 // the object. It is easy to avoid this. We move the finger by | |
| 3392 // enough to point to the next possible object header (the | |
| 3393 // bitmap knows by how much we need to move it as it knows its | |
| 3394 // granularity). | |
| 3395 MemRegion newRegion = | |
| 3396 MemRegion(_nextMarkBitMap->nextWord(_region_finger), mr.end()); | |
| 3397 | |
| 3398 if (!newRegion.is_empty()) { | |
| 3399 if (_cm->verbose_low()) { | |
| 3400 gclog_or_tty->print_cr("[%d] pushing unscanned region" | |
| 3401 "[" PTR_FORMAT "," PTR_FORMAT ") on region stack", | |
| 3402 _task_id, | |
| 3403 newRegion.start(), newRegion.end()); | |
| 3404 } | |
| 3405 // Now push the part of the region we didn't scan on the | |
| 3406 // region stack to make sure a task scans it later. | |
| 3407 _cm->region_stack_push(newRegion); | |
| 3408 } | |
| 3409 // break from while | |
| 3410 mr = MemRegion(); | |
| 3411 } | |
| 3412 _region_finger = NULL; | |
| 3413 } | |
| 3414 | |
| 3415 // We only push regions on the region stack during evacuation | |
| 3416 // pauses. So if we come out the above iteration because we region | |
| 3417 // stack is empty, it will remain empty until the next yield | |
| 3418 // point. So, the guarantee below is safe. | |
| 3419 guarantee( has_aborted() || _cm->region_stack_empty(), | |
| 3420 "only way to exit the loop" ); | |
| 3421 | |
| 3422 if (_cm->verbose_low()) | |
| 3423 gclog_or_tty->print_cr("[%d] drained region stack, size = %d", | |
| 3424 _task_id, _cm->region_stack_size()); | |
| 3425 } | |
| 3426 } | |
| 3427 | |
| 3428 void CMTask::print_stats() { | |
| 3429 gclog_or_tty->print_cr("Marking Stats, task = %d, calls = %d", | |
| 3430 _task_id, _calls); | |
| 3431 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", | |
| 3432 _elapsed_time_ms, _termination_time_ms); | |
| 3433 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", | |
| 3434 _step_times_ms.num(), _step_times_ms.avg(), | |
| 3435 _step_times_ms.sd()); | |
| 3436 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", | |
| 3437 _step_times_ms.maximum(), _step_times_ms.sum()); | |
| 3438 | |
| 3439 #if _MARKING_STATS_ | |
| 3440 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", | |
| 3441 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), | |
| 3442 _all_clock_intervals_ms.sd()); | |
| 3443 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", | |
| 3444 _all_clock_intervals_ms.maximum(), | |
| 3445 _all_clock_intervals_ms.sum()); | |
| 3446 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", | |
| 3447 _clock_due_to_scanning, _clock_due_to_marking); | |
| 3448 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", | |
| 3449 _objs_scanned, _objs_found_on_bitmap); | |
| 3450 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", | |
| 3451 _local_pushes, _local_pops, _local_max_size); | |
| 3452 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", | |
| 3453 _global_pushes, _global_pops, _global_max_size); | |
| 3454 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", | |
| 3455 _global_transfers_to,_global_transfers_from); | |
| 3456 gclog_or_tty->print_cr(" Regions: claimed = %d, Region Stack: pops = %d", | |
| 3457 _regions_claimed, _region_stack_pops); | |
| 3458 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); | |
| 3459 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", | |
| 3460 _steal_attempts, _steals); | |
| 3461 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); | |
| 3462 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", | |
| 3463 _aborted_overflow, _aborted_cm_aborted, _aborted_yield); | |
| 3464 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", | |
| 3465 _aborted_timed_out, _aborted_satb, _aborted_termination); | |
| 3466 #endif // _MARKING_STATS_ | |
| 3467 } | |
| 3468 | |
| 3469 /***************************************************************************** | |
| 3470 | |
| 3471 The do_marking_step(time_target_ms) method is the building block | |
| 3472 of the parallel marking framework. It can be called in parallel | |
| 3473 with other invocations of do_marking_step() on different tasks | |
| 3474 (but only one per task, obviously) and concurrently with the | |
| 3475 mutator threads, or during remark, hence it eliminates the need | |
| 3476 for two versions of the code. When called during remark, it will | |
| 3477 pick up from where the task left off during the concurrent marking | |
| 3478 phase. Interestingly, tasks are also claimable during evacuation | |
| 3479 pauses too, since do_marking_step() ensures that it aborts before | |
| 3480 it needs to yield. | |
| 3481 | |
| 3482 The data structures that is uses to do marking work are the | |
| 3483 following: | |
| 3484 | |
| 3485 (1) Marking Bitmap. If there are gray objects that appear only | |
| 3486 on the bitmap (this happens either when dealing with an overflow | |
| 3487 or when the initial marking phase has simply marked the roots | |
| 3488 and didn't push them on the stack), then tasks claim heap | |
| 3489 regions whose bitmap they then scan to find gray objects. A | |
| 3490 global finger indicates where the end of the last claimed region | |
| 3491 is. A local finger indicates how far into the region a task has | |
| 3492 scanned. The two fingers are used to determine how to gray an | |
| 3493 object (i.e. whether simply marking it is OK, as it will be | |
| 3494 visited by a task in the future, or whether it needs to be also | |
| 3495 pushed on a stack). | |
| 3496 | |
| 3497 (2) Local Queue. The local queue of the task which is accessed | |
| 3498 reasonably efficiently by the task. Other tasks can steal from | |
| 3499 it when they run out of work. Throughout the marking phase, a | |
| 3500 task attempts to keep its local queue short but not totally | |
| 3501 empty, so that entries are available for stealing by other | |
| 3502 tasks. Only when there is no more work, a task will totally | |
| 3503 drain its local queue. | |
| 3504 | |
| 3505 (3) Global Mark Stack. This handles local queue overflow. During | |
| 3506 marking only sets of entries are moved between it and the local | |
| 3507 queues, as access to it requires a mutex and more fine-grain | |
| 3508 interaction with it which might cause contention. If it | |
| 3509 overflows, then the marking phase should restart and iterate | |
| 3510 over the bitmap to identify gray objects. Throughout the marking | |
| 3511 phase, tasks attempt to keep the global mark stack at a small | |
| 3512 length but not totally empty, so that entries are available for | |
| 3513 popping by other tasks. Only when there is no more work, tasks | |
| 3514 will totally drain the global mark stack. | |
| 3515 | |
| 3516 (4) Global Region Stack. Entries on it correspond to areas of | |
| 3517 the bitmap that need to be scanned since they contain gray | |
| 3518 objects. Pushes on the region stack only happen during | |
| 3519 evacuation pauses and typically correspond to areas covered by | |
| 3520 GC LABS. If it overflows, then the marking phase should restart | |
| 3521 and iterate over the bitmap to identify gray objects. Tasks will | |
| 3522 try to totally drain the region stack as soon as possible. | |
| 3523 | |
| 3524 (5) SATB Buffer Queue. This is where completed SATB buffers are | |
| 3525 made available. Buffers are regularly removed from this queue | |
| 3526 and scanned for roots, so that the queue doesn't get too | |
| 3527 long. During remark, all completed buffers are processed, as | |
| 3528 well as the filled in parts of any uncompleted buffers. | |
| 3529 | |
| 3530 The do_marking_step() method tries to abort when the time target | |
| 3531 has been reached. There are a few other cases when the | |
| 3532 do_marking_step() method also aborts: | |
| 3533 | |
| 3534 (1) When the marking phase has been aborted (after a Full GC). | |
| 3535 | |
| 3536 (2) When a global overflow (either on the global stack or the | |
| 3537 region stack) has been triggered. Before the task aborts, it | |
| 3538 will actually sync up with the other tasks to ensure that all | |
| 3539 the marking data structures (local queues, stacks, fingers etc.) | |
| 3540 are re-initialised so that when do_marking_step() completes, | |
| 3541 the marking phase can immediately restart. | |
| 3542 | |
| 3543 (3) When enough completed SATB buffers are available. The | |
| 3544 do_marking_step() method only tries to drain SATB buffers right | |
| 3545 at the beginning. So, if enough buffers are available, the | |
| 3546 marking step aborts and the SATB buffers are processed at | |
| 3547 the beginning of the next invocation. | |
| 3548 | |
| 3549 (4) To yield. when we have to yield then we abort and yield | |
| 3550 right at the end of do_marking_step(). This saves us from a lot | |
| 3551 of hassle as, by yielding we might allow a Full GC. If this | |
| 3552 happens then objects will be compacted underneath our feet, the | |
| 3553 heap might shrink, etc. We save checking for this by just | |
| 3554 aborting and doing the yield right at the end. | |
| 3555 | |
| 3556 From the above it follows that the do_marking_step() method should | |
| 3557 be called in a loop (or, otherwise, regularly) until it completes. | |
| 3558 | |
| 3559 If a marking step completes without its has_aborted() flag being | |
| 3560 true, it means it has completed the current marking phase (and | |
| 3561 also all other marking tasks have done so and have all synced up). | |
| 3562 | |
| 3563 A method called regular_clock_call() is invoked "regularly" (in | |
| 3564 sub ms intervals) throughout marking. It is this clock method that | |
| 3565 checks all the abort conditions which were mentioned above and | |
| 3566 decides when the task should abort. A work-based scheme is used to | |
| 3567 trigger this clock method: when the number of object words the | |
| 3568 marking phase has scanned or the number of references the marking | |
| 3569 phase has visited reach a given limit. Additional invocations to | |
| 3570 the method clock have been planted in a few other strategic places | |
| 3571 too. The initial reason for the clock method was to avoid calling | |
| 3572 vtime too regularly, as it is quite expensive. So, once it was in | |
| 3573 place, it was natural to piggy-back all the other conditions on it | |
| 3574 too and not constantly check them throughout the code. | |
| 3575 | |
| 3576 *****************************************************************************/ | |
| 3577 | |
| 3578 void CMTask::do_marking_step(double time_target_ms) { | |
| 3579 guarantee( time_target_ms >= 1.0, "minimum granularity is 1ms" ); | |
| 3580 guarantee( concurrent() == _cm->concurrent(), "they should be the same" ); | |
| 3581 | |
| 3582 guarantee( concurrent() || _cm->region_stack_empty(), | |
| 3583 "the region stack should have been cleared before remark" ); | |
| 3584 guarantee( _region_finger == NULL, | |
| 3585 "this should be non-null only when a region is being scanned" ); | |
| 3586 | |
| 3587 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); | |
| 3588 guarantee( _task_queues != NULL, "invariant" ); | |
| 3589 guarantee( _task_queue != NULL, "invariant" ); | |
| 3590 guarantee( _task_queues->queue(_task_id) == _task_queue, "invariant" ); | |
| 3591 | |
| 3592 guarantee( !_claimed, | |
| 3593 "only one thread should claim this task at any one time" ); | |
| 3594 | |
| 3595 // OK, this doesn't safeguard again all possible scenarios, as it is | |
| 3596 // possible for two threads to set the _claimed flag at the same | |
| 3597 // time. But it is only for debugging purposes anyway and it will | |
| 3598 // catch most problems. | |
| 3599 _claimed = true; | |
| 3600 | |
| 3601 _start_time_ms = os::elapsedVTime() * 1000.0; | |
| 3602 statsOnly( _interval_start_time_ms = _start_time_ms ); | |
| 3603 | |
| 3604 double diff_prediction_ms = | |
| 3605 g1_policy->get_new_prediction(&_marking_step_diffs_ms); | |
| 3606 _time_target_ms = time_target_ms - diff_prediction_ms; | |
| 3607 | |
| 3608 // set up the variables that are used in the work-based scheme to | |
| 3609 // call the regular clock method | |
| 3610 _words_scanned = 0; | |
| 3611 _refs_reached = 0; | |
| 3612 recalculate_limits(); | |
| 3613 | |
| 3614 // clear all flags | |
| 3615 clear_has_aborted(); | |
| 3616 _has_aborted_timed_out = false; | |
| 3617 _draining_satb_buffers = false; | |
| 3618 | |
| 3619 ++_calls; | |
| 3620 | |
| 3621 if (_cm->verbose_low()) | |
| 3622 gclog_or_tty->print_cr("[%d] >>>>>>>>>> START, call = %d, " | |
| 3623 "target = %1.2lfms >>>>>>>>>>", | |
| 3624 _task_id, _calls, _time_target_ms); | |
| 3625 | |
| 3626 // Set up the bitmap and oop closures. Anything that uses them is | |
| 3627 // eventually called from this method, so it is OK to allocate these | |
| 3628 // statically. | |
| 3629 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); | |
| 3630 CMOopClosure oop_closure(_g1h, _cm, this); | |
| 3631 set_oop_closure(&oop_closure); | |
| 3632 | |
| 3633 if (_cm->has_overflown()) { | |
| 3634 // This can happen if the region stack or the mark stack overflows | |
| 3635 // during a GC pause and this task, after a yield point, | |
| 3636 // restarts. We have to abort as we need to get into the overflow | |
| 3637 // protocol which happens right at the end of this task. | |
| 3638 set_has_aborted(); | |
| 3639 } | |
| 3640 | |
| 3641 // First drain any available SATB buffers. After this, we will not | |
| 3642 // look at SATB buffers before the next invocation of this method. | |
| 3643 // If enough completed SATB buffers are queued up, the regular clock | |
| 3644 // will abort this task so that it restarts. | |
| 3645 drain_satb_buffers(); | |
| 3646 // ...then partially drain the local queue and the global stack | |
| 3647 drain_local_queue(true); | |
| 3648 drain_global_stack(true); | |
| 3649 | |
| 3650 // Then totally drain the region stack. We will not look at | |
| 3651 // it again before the next invocation of this method. Entries on | |
| 3652 // the region stack are only added during evacuation pauses, for | |
| 3653 // which we have to yield. When we do, we abort the task anyway so | |
| 3654 // it will look at the region stack again when it restarts. | |
| 3655 bitmap_closure.set_scanning_heap_region(false); | |
| 3656 drain_region_stack(&bitmap_closure); | |
| 3657 // ...then partially drain the local queue and the global stack | |
| 3658 drain_local_queue(true); | |
| 3659 drain_global_stack(true); | |
| 3660 | |
| 3661 do { | |
| 3662 if (!has_aborted() && _curr_region != NULL) { | |
| 3663 // This means that we're already holding on to a region. | |
| 3664 tmp_guarantee_CM( _finger != NULL, | |
| 3665 "if region is not NULL, then the finger " | |
| 3666 "should not be NULL either" ); | |
| 3667 | |
| 3668 // We might have restarted this task after an evacuation pause | |
| 3669 // which might have evacuated the region we're holding on to | |
| 3670 // underneath our feet. Let's read its limit again to make sure | |
| 3671 // that we do not iterate over a region of the heap that | |
| 3672 // contains garbage (update_region_limit() will also move | |
| 3673 // _finger to the start of the region if it is found empty). | |
| 3674 update_region_limit(); | |
| 3675 // We will start from _finger not from the start of the region, | |
| 3676 // as we might be restarting this task after aborting half-way | |
| 3677 // through scanning this region. In this case, _finger points to | |
| 3678 // the address where we last found a marked object. If this is a | |
| 3679 // fresh region, _finger points to start(). | |
| 3680 MemRegion mr = MemRegion(_finger, _region_limit); | |
| 3681 | |
| 3682 if (_cm->verbose_low()) | |
| 3683 gclog_or_tty->print_cr("[%d] we're scanning part " | |
| 3684 "["PTR_FORMAT", "PTR_FORMAT") " | |
| 3685 "of region "PTR_FORMAT, | |
| 3686 _task_id, _finger, _region_limit, _curr_region); | |
| 3687 | |
| 3688 // Let's iterate over the bitmap of the part of the | |
| 3689 // region that is left. | |
| 3690 bitmap_closure.set_scanning_heap_region(true); | |
| 3691 if (mr.is_empty() || | |
| 3692 _nextMarkBitMap->iterate(&bitmap_closure, mr)) { | |
| 3693 // We successfully completed iterating over the region. Now, | |
| 3694 // let's give up the region. | |
| 3695 giveup_current_region(); | |
| 3696 regular_clock_call(); | |
| 3697 } else { | |
| 3698 guarantee( has_aborted(), "currently the only way to do so" ); | |
| 3699 // The only way to abort the bitmap iteration is to return | |
| 3700 // false from the do_bit() method. However, inside the | |
| 3701 // do_bit() method we move the _finger to point to the | |
| 3702 // object currently being looked at. So, if we bail out, we | |
| 3703 // have definitely set _finger to something non-null. | |
| 3704 guarantee( _finger != NULL, "invariant" ); | |
| 3705 | |
| 3706 // Region iteration was actually aborted. So now _finger | |
| 3707 // points to the address of the object we last scanned. If we | |
| 3708 // leave it there, when we restart this task, we will rescan | |
| 3709 // the object. It is easy to avoid this. We move the finger by | |
| 3710 // enough to point to the next possible object header (the | |
| 3711 // bitmap knows by how much we need to move it as it knows its | |
| 3712 // granularity). | |
| 3713 move_finger_to(_nextMarkBitMap->nextWord(_finger)); | |
| 3714 } | |
| 3715 } | |
| 3716 // At this point we have either completed iterating over the | |
| 3717 // region we were holding on to, or we have aborted. | |
| 3718 | |
| 3719 // We then partially drain the local queue and the global stack. | |
| 3720 // (Do we really need this?) | |
| 3721 drain_local_queue(true); | |
| 3722 drain_global_stack(true); | |
| 3723 | |
| 3724 // Read the note on the claim_region() method on why it might | |
| 3725 // return NULL with potentially more regions available for | |
| 3726 // claiming and why we have to check out_of_regions() to determine | |
| 3727 // whether we're done or not. | |
| 3728 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { | |
| 3729 // We are going to try to claim a new region. We should have | |
| 3730 // given up on the previous one. | |
| 3731 tmp_guarantee_CM( _curr_region == NULL && | |
| 3732 _finger == NULL && | |
| 3733 _region_limit == NULL, "invariant" ); | |
| 3734 if (_cm->verbose_low()) | |
| 3735 gclog_or_tty->print_cr("[%d] trying to claim a new region", _task_id); | |
| 3736 HeapRegion* claimed_region = _cm->claim_region(_task_id); | |
| 3737 if (claimed_region != NULL) { | |
| 3738 // Yes, we managed to claim one | |
| 3739 statsOnly( ++_regions_claimed ); | |
| 3740 | |
| 3741 if (_cm->verbose_low()) | |
| 3742 gclog_or_tty->print_cr("[%d] we successfully claimed " | |
| 3743 "region "PTR_FORMAT, | |
| 3744 _task_id, claimed_region); | |
| 3745 | |
| 3746 setup_for_region(claimed_region); | |
| 3747 tmp_guarantee_CM( _curr_region == claimed_region, "invariant" ); | |
| 3748 } | |
| 3749 // It is important to call the regular clock here. It might take | |
| 3750 // a while to claim a region if, for example, we hit a large | |
| 3751 // block of empty regions. So we need to call the regular clock | |
| 3752 // method once round the loop to make sure it's called | |
| 3753 // frequently enough. | |
| 3754 regular_clock_call(); | |
| 3755 } | |
| 3756 | |
| 3757 if (!has_aborted() && _curr_region == NULL) { | |
| 3758 tmp_guarantee_CM( _cm->out_of_regions(), | |
| 3759 "at this point we should be out of regions" ); | |
| 3760 } | |
| 3761 } while ( _curr_region != NULL && !has_aborted()); | |
| 3762 | |
| 3763 if (!has_aborted()) { | |
| 3764 // We cannot check whether the global stack is empty, since other | |
|
343
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3765 // tasks might be pushing objects to it concurrently. We also cannot |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3766 // check if the region stack is empty because if a thread is aborting |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3767 // it can push a partially done region back. |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3768 tmp_guarantee_CM( _cm->out_of_regions(), |
| 342 | 3769 "at this point we should be out of regions" ); |
| 3770 | |
| 3771 if (_cm->verbose_low()) | |
| 3772 gclog_or_tty->print_cr("[%d] all regions claimed", _task_id); | |
| 3773 | |
| 3774 // Try to reduce the number of available SATB buffers so that | |
| 3775 // remark has less work to do. | |
| 3776 drain_satb_buffers(); | |
| 3777 } | |
| 3778 | |
| 3779 // Since we've done everything else, we can now totally drain the | |
| 3780 // local queue and global stack. | |
| 3781 drain_local_queue(false); | |
| 3782 drain_global_stack(false); | |
| 3783 | |
| 3784 // Attempt at work stealing from other task's queues. | |
| 3785 if (!has_aborted()) { | |
| 3786 // We have not aborted. This means that we have finished all that | |
| 3787 // we could. Let's try to do some stealing... | |
| 3788 | |
| 3789 // We cannot check whether the global stack is empty, since other | |
|
343
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3790 // tasks might be pushing objects to it concurrently. We also cannot |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3791 // check if the region stack is empty because if a thread is aborting |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3792 // it can push a partially done region back. |
| 342 | 3793 guarantee( _cm->out_of_regions() && |
| 3794 _task_queue->size() == 0, "only way to reach here" ); | |
| 3795 | |
| 3796 if (_cm->verbose_low()) | |
| 3797 gclog_or_tty->print_cr("[%d] starting to steal", _task_id); | |
| 3798 | |
| 3799 while (!has_aborted()) { | |
| 3800 oop obj; | |
| 3801 statsOnly( ++_steal_attempts ); | |
| 3802 | |
| 3803 if (_cm->try_stealing(_task_id, &_hash_seed, obj)) { | |
| 3804 if (_cm->verbose_medium()) | |
| 3805 gclog_or_tty->print_cr("[%d] stolen "PTR_FORMAT" successfully", | |
| 3806 _task_id, (void*) obj); | |
| 3807 | |
| 3808 statsOnly( ++_steals ); | |
| 3809 | |
| 3810 tmp_guarantee_CM( _nextMarkBitMap->isMarked((HeapWord*) obj), | |
| 3811 "any stolen object should be marked" ); | |
| 3812 scan_object(obj); | |
| 3813 | |
| 3814 // And since we're towards the end, let's totally drain the | |
| 3815 // local queue and global stack. | |
| 3816 drain_local_queue(false); | |
| 3817 drain_global_stack(false); | |
| 3818 } else { | |
| 3819 break; | |
| 3820 } | |
| 3821 } | |
| 3822 } | |
| 3823 | |
| 3824 // We still haven't aborted. Now, let's try to get into the | |
| 3825 // termination protocol. | |
| 3826 if (!has_aborted()) { | |
| 3827 // We cannot check whether the global stack is empty, since other | |
|
343
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3828 // tasks might be concurrently pushing objects on it. We also cannot |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3829 // check if the region stack is empty because if a thread is aborting |
|
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3830 // it can push a partially done region back. |
| 342 | 3831 guarantee( _cm->out_of_regions() && |
| 3832 _task_queue->size() == 0, "only way to reach here" ); | |
| 3833 | |
| 3834 if (_cm->verbose_low()) | |
| 3835 gclog_or_tty->print_cr("[%d] starting termination protocol", _task_id); | |
| 3836 | |
| 3837 _termination_start_time_ms = os::elapsedVTime() * 1000.0; | |
| 3838 // The CMTask class also extends the TerminatorTerminator class, | |
| 3839 // hence its should_exit_termination() method will also decide | |
| 3840 // whether to exit the termination protocol or not. | |
| 3841 bool finished = _cm->terminator()->offer_termination(this); | |
| 3842 double termination_end_time_ms = os::elapsedVTime() * 1000.0; | |
| 3843 _termination_time_ms += | |
| 3844 termination_end_time_ms - _termination_start_time_ms; | |
| 3845 | |
| 3846 if (finished) { | |
| 3847 // We're all done. | |
| 3848 | |
| 3849 if (_task_id == 0) { | |
| 3850 // let's allow task 0 to do this | |
| 3851 if (concurrent()) { | |
| 3852 guarantee( _cm->concurrent_marking_in_progress(), "invariant" ); | |
| 3853 // we need to set this to false before the next | |
| 3854 // safepoint. This way we ensure that the marking phase | |
| 3855 // doesn't observe any more heap expansions. | |
| 3856 _cm->clear_concurrent_marking_in_progress(); | |
| 3857 } | |
| 3858 } | |
| 3859 | |
| 3860 // We can now guarantee that the global stack is empty, since | |
| 3861 // all other tasks have finished. | |
| 3862 guarantee( _cm->out_of_regions() && | |
| 3863 _cm->region_stack_empty() && | |
| 3864 _cm->mark_stack_empty() && | |
| 3865 _task_queue->size() == 0 && | |
| 3866 !_cm->has_overflown() && | |
| 3867 !_cm->mark_stack_overflow() && | |
| 3868 !_cm->region_stack_overflow(), | |
| 3869 "only way to reach here" ); | |
| 3870 | |
| 3871 if (_cm->verbose_low()) | |
| 3872 gclog_or_tty->print_cr("[%d] all tasks terminated", _task_id); | |
| 3873 } else { | |
| 3874 // Apparently there's more work to do. Let's abort this task. It | |
| 3875 // will restart it and we can hopefully find more things to do. | |
| 3876 | |
| 3877 if (_cm->verbose_low()) | |
| 3878 gclog_or_tty->print_cr("[%d] apparently there is more work to do", _task_id); | |
| 3879 | |
| 3880 set_has_aborted(); | |
| 3881 statsOnly( ++_aborted_termination ); | |
| 3882 } | |
| 3883 } | |
| 3884 | |
| 3885 // Mainly for debugging purposes to make sure that a pointer to the | |
| 3886 // closure which was statically allocated in this frame doesn't | |
| 3887 // escape it by accident. | |
| 3888 set_oop_closure(NULL); | |
| 3889 double end_time_ms = os::elapsedVTime() * 1000.0; | |
| 3890 double elapsed_time_ms = end_time_ms - _start_time_ms; | |
| 3891 // Update the step history. | |
| 3892 _step_times_ms.add(elapsed_time_ms); | |
| 3893 | |
| 3894 if (has_aborted()) { | |
| 3895 // The task was aborted for some reason. | |
| 3896 | |
| 3897 statsOnly( ++_aborted ); | |
| 3898 | |
| 3899 if (_has_aborted_timed_out) { | |
| 3900 double diff_ms = elapsed_time_ms - _time_target_ms; | |
| 3901 // Keep statistics of how well we did with respect to hitting | |
| 3902 // our target only if we actually timed out (if we aborted for | |
| 3903 // other reasons, then the results might get skewed). | |
| 3904 _marking_step_diffs_ms.add(diff_ms); | |
| 3905 } | |
| 3906 | |
| 3907 if (_cm->has_overflown()) { | |
| 3908 // This is the interesting one. We aborted because a global | |
| 3909 // overflow was raised. This means we have to restart the | |
| 3910 // marking phase and start iterating over regions. However, in | |
| 3911 // order to do this we have to make sure that all tasks stop | |
| 3912 // what they are doing and re-initialise in a safe manner. We | |
| 3913 // will achieve this with the use of two barrier sync points. | |
| 3914 | |
| 3915 if (_cm->verbose_low()) | |
| 3916 gclog_or_tty->print_cr("[%d] detected overflow", _task_id); | |
| 3917 | |
| 3918 _cm->enter_first_sync_barrier(_task_id); | |
| 3919 // When we exit this sync barrier we know that all tasks have | |
| 3920 // stopped doing marking work. So, it's now safe to | |
| 3921 // re-initialise our data structures. At the end of this method, | |
| 3922 // task 0 will clear the global data structures. | |
| 3923 | |
| 3924 statsOnly( ++_aborted_overflow ); | |
| 3925 | |
| 3926 // We clear the local state of this task... | |
| 3927 clear_region_fields(); | |
| 3928 | |
| 3929 // ...and enter the second barrier. | |
| 3930 _cm->enter_second_sync_barrier(_task_id); | |
| 3931 // At this point everything has bee re-initialised and we're | |
| 3932 // ready to restart. | |
| 3933 } | |
| 3934 | |
| 3935 if (_cm->verbose_low()) { | |
| 3936 gclog_or_tty->print_cr("[%d] <<<<<<<<<< ABORTING, target = %1.2lfms, " | |
| 3937 "elapsed = %1.2lfms <<<<<<<<<<", | |
| 3938 _task_id, _time_target_ms, elapsed_time_ms); | |
| 3939 if (_cm->has_aborted()) | |
| 3940 gclog_or_tty->print_cr("[%d] ========== MARKING ABORTED ==========", | |
| 3941 _task_id); | |
| 3942 } | |
| 3943 } else { | |
| 3944 if (_cm->verbose_low()) | |
| 3945 gclog_or_tty->print_cr("[%d] <<<<<<<<<< FINISHED, target = %1.2lfms, " | |
| 3946 "elapsed = %1.2lfms <<<<<<<<<<", | |
| 3947 _task_id, _time_target_ms, elapsed_time_ms); | |
| 3948 } | |
| 3949 | |
| 3950 _claimed = false; | |
| 3951 } | |
| 3952 | |
| 3953 CMTask::CMTask(int task_id, | |
| 3954 ConcurrentMark* cm, | |
| 3955 CMTaskQueue* task_queue, | |
| 3956 CMTaskQueueSet* task_queues) | |
| 3957 : _g1h(G1CollectedHeap::heap()), | |
| 3958 _task_id(task_id), _cm(cm), | |
| 3959 _claimed(false), | |
| 3960 _nextMarkBitMap(NULL), _hash_seed(17), | |
| 3961 _task_queue(task_queue), | |
| 3962 _task_queues(task_queues), | |
| 3963 _oop_closure(NULL) { | |
| 3964 guarantee( task_queue != NULL, "invariant" ); | |
| 3965 guarantee( task_queues != NULL, "invariant" ); | |
| 3966 | |
| 3967 statsOnly( _clock_due_to_scanning = 0; | |
| 3968 _clock_due_to_marking = 0 ); | |
| 3969 | |
| 3970 _marking_step_diffs_ms.add(0.5); | |
| 3971 } |