Fri, 02 Oct 2009 16:12:07 -0400
6885041: G1: inconsistent thread dump
Summary: When G1 is enabled, thread dumps are inconsistent as the info for some of the G1 threads is not formatted properly.
Reviewed-by: ysr, johnc
ysr@777 | 1 | /* |
xdono@1279 | 2 | * Copyright 2001-2009 Sun Microsystems, Inc. All Rights Reserved. |
ysr@777 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
ysr@777 | 4 | * |
ysr@777 | 5 | * This code is free software; you can redistribute it and/or modify it |
ysr@777 | 6 | * under the terms of the GNU General Public License version 2 only, as |
ysr@777 | 7 | * published by the Free Software Foundation. |
ysr@777 | 8 | * |
ysr@777 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
ysr@777 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
ysr@777 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
ysr@777 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
ysr@777 | 13 | * accompanied this code). |
ysr@777 | 14 | * |
ysr@777 | 15 | * You should have received a copy of the GNU General Public License version |
ysr@777 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
ysr@777 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
ysr@777 | 18 | * |
ysr@777 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
ysr@777 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
ysr@777 | 21 | * have any questions. |
ysr@777 | 22 | * |
ysr@777 | 23 | */ |
ysr@777 | 24 | |
ysr@777 | 25 | #include "incls/_precompiled.incl" |
ysr@777 | 26 | #include "incls/_concurrentMark.cpp.incl" |
ysr@777 | 27 | |
ysr@777 | 28 | // |
ysr@777 | 29 | // CMS Bit Map Wrapper |
ysr@777 | 30 | |
ysr@777 | 31 | CMBitMapRO::CMBitMapRO(ReservedSpace rs, int shifter): |
ysr@777 | 32 | _bm((uintptr_t*)NULL,0), |
ysr@777 | 33 | _shifter(shifter) { |
ysr@777 | 34 | _bmStartWord = (HeapWord*)(rs.base()); |
ysr@777 | 35 | _bmWordSize = rs.size()/HeapWordSize; // rs.size() is in bytes |
ysr@777 | 36 | ReservedSpace brs(ReservedSpace::allocation_align_size_up( |
ysr@777 | 37 | (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); |
ysr@777 | 38 | |
ysr@777 | 39 | guarantee(brs.is_reserved(), "couldn't allocate CMS bit map"); |
ysr@777 | 40 | // For now we'll just commit all of the bit map up fromt. |
ysr@777 | 41 | // Later on we'll try to be more parsimonious with swap. |
ysr@777 | 42 | guarantee(_virtual_space.initialize(brs, brs.size()), |
ysr@777 | 43 | "couldn't reseve backing store for CMS bit map"); |
ysr@777 | 44 | assert(_virtual_space.committed_size() == brs.size(), |
ysr@777 | 45 | "didn't reserve backing store for all of CMS bit map?"); |
ysr@777 | 46 | _bm.set_map((uintptr_t*)_virtual_space.low()); |
ysr@777 | 47 | assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= |
ysr@777 | 48 | _bmWordSize, "inconsistency in bit map sizing"); |
ysr@777 | 49 | _bm.set_size(_bmWordSize >> _shifter); |
ysr@777 | 50 | } |
ysr@777 | 51 | |
ysr@777 | 52 | HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr, |
ysr@777 | 53 | HeapWord* limit) const { |
ysr@777 | 54 | // First we must round addr *up* to a possible object boundary. |
ysr@777 | 55 | addr = (HeapWord*)align_size_up((intptr_t)addr, |
ysr@777 | 56 | HeapWordSize << _shifter); |
ysr@777 | 57 | size_t addrOffset = heapWordToOffset(addr); |
ysr@777 | 58 | if (limit == NULL) limit = _bmStartWord + _bmWordSize; |
ysr@777 | 59 | size_t limitOffset = heapWordToOffset(limit); |
ysr@777 | 60 | size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); |
ysr@777 | 61 | HeapWord* nextAddr = offsetToHeapWord(nextOffset); |
ysr@777 | 62 | assert(nextAddr >= addr, "get_next_one postcondition"); |
ysr@777 | 63 | assert(nextAddr == limit || isMarked(nextAddr), |
ysr@777 | 64 | "get_next_one postcondition"); |
ysr@777 | 65 | return nextAddr; |
ysr@777 | 66 | } |
ysr@777 | 67 | |
ysr@777 | 68 | HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr, |
ysr@777 | 69 | HeapWord* limit) const { |
ysr@777 | 70 | size_t addrOffset = heapWordToOffset(addr); |
ysr@777 | 71 | if (limit == NULL) limit = _bmStartWord + _bmWordSize; |
ysr@777 | 72 | size_t limitOffset = heapWordToOffset(limit); |
ysr@777 | 73 | size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset); |
ysr@777 | 74 | HeapWord* nextAddr = offsetToHeapWord(nextOffset); |
ysr@777 | 75 | assert(nextAddr >= addr, "get_next_one postcondition"); |
ysr@777 | 76 | assert(nextAddr == limit || !isMarked(nextAddr), |
ysr@777 | 77 | "get_next_one postcondition"); |
ysr@777 | 78 | return nextAddr; |
ysr@777 | 79 | } |
ysr@777 | 80 | |
ysr@777 | 81 | int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const { |
ysr@777 | 82 | assert((diff & ((1 << _shifter) - 1)) == 0, "argument check"); |
ysr@777 | 83 | return (int) (diff >> _shifter); |
ysr@777 | 84 | } |
ysr@777 | 85 | |
ysr@777 | 86 | bool CMBitMapRO::iterate(BitMapClosure* cl, MemRegion mr) { |
ysr@777 | 87 | HeapWord* left = MAX2(_bmStartWord, mr.start()); |
ysr@777 | 88 | HeapWord* right = MIN2(_bmStartWord + _bmWordSize, mr.end()); |
ysr@777 | 89 | if (right > left) { |
ysr@777 | 90 | // Right-open interval [leftOffset, rightOffset). |
ysr@777 | 91 | return _bm.iterate(cl, heapWordToOffset(left), heapWordToOffset(right)); |
ysr@777 | 92 | } else { |
ysr@777 | 93 | return true; |
ysr@777 | 94 | } |
ysr@777 | 95 | } |
ysr@777 | 96 | |
ysr@777 | 97 | void CMBitMapRO::mostly_disjoint_range_union(BitMap* from_bitmap, |
ysr@777 | 98 | size_t from_start_index, |
ysr@777 | 99 | HeapWord* to_start_word, |
ysr@777 | 100 | size_t word_num) { |
ysr@777 | 101 | _bm.mostly_disjoint_range_union(from_bitmap, |
ysr@777 | 102 | from_start_index, |
ysr@777 | 103 | heapWordToOffset(to_start_word), |
ysr@777 | 104 | word_num); |
ysr@777 | 105 | } |
ysr@777 | 106 | |
ysr@777 | 107 | #ifndef PRODUCT |
ysr@777 | 108 | bool CMBitMapRO::covers(ReservedSpace rs) const { |
ysr@777 | 109 | // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); |
kvn@1080 | 110 | assert(((size_t)_bm.size() * (size_t)(1 << _shifter)) == _bmWordSize, |
ysr@777 | 111 | "size inconsistency"); |
ysr@777 | 112 | return _bmStartWord == (HeapWord*)(rs.base()) && |
ysr@777 | 113 | _bmWordSize == rs.size()>>LogHeapWordSize; |
ysr@777 | 114 | } |
ysr@777 | 115 | #endif |
ysr@777 | 116 | |
ysr@777 | 117 | void CMBitMap::clearAll() { |
ysr@777 | 118 | _bm.clear(); |
ysr@777 | 119 | return; |
ysr@777 | 120 | } |
ysr@777 | 121 | |
ysr@777 | 122 | void CMBitMap::markRange(MemRegion mr) { |
ysr@777 | 123 | mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); |
ysr@777 | 124 | assert(!mr.is_empty(), "unexpected empty region"); |
ysr@777 | 125 | assert((offsetToHeapWord(heapWordToOffset(mr.end())) == |
ysr@777 | 126 | ((HeapWord *) mr.end())), |
ysr@777 | 127 | "markRange memory region end is not card aligned"); |
ysr@777 | 128 | // convert address range into offset range |
ysr@777 | 129 | _bm.at_put_range(heapWordToOffset(mr.start()), |
ysr@777 | 130 | heapWordToOffset(mr.end()), true); |
ysr@777 | 131 | } |
ysr@777 | 132 | |
ysr@777 | 133 | void CMBitMap::clearRange(MemRegion mr) { |
ysr@777 | 134 | mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); |
ysr@777 | 135 | assert(!mr.is_empty(), "unexpected empty region"); |
ysr@777 | 136 | // convert address range into offset range |
ysr@777 | 137 | _bm.at_put_range(heapWordToOffset(mr.start()), |
ysr@777 | 138 | heapWordToOffset(mr.end()), false); |
ysr@777 | 139 | } |
ysr@777 | 140 | |
ysr@777 | 141 | MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, |
ysr@777 | 142 | HeapWord* end_addr) { |
ysr@777 | 143 | HeapWord* start = getNextMarkedWordAddress(addr); |
ysr@777 | 144 | start = MIN2(start, end_addr); |
ysr@777 | 145 | HeapWord* end = getNextUnmarkedWordAddress(start); |
ysr@777 | 146 | end = MIN2(end, end_addr); |
ysr@777 | 147 | assert(start <= end, "Consistency check"); |
ysr@777 | 148 | MemRegion mr(start, end); |
ysr@777 | 149 | if (!mr.is_empty()) { |
ysr@777 | 150 | clearRange(mr); |
ysr@777 | 151 | } |
ysr@777 | 152 | return mr; |
ysr@777 | 153 | } |
ysr@777 | 154 | |
ysr@777 | 155 | CMMarkStack::CMMarkStack(ConcurrentMark* cm) : |
ysr@777 | 156 | _base(NULL), _cm(cm) |
ysr@777 | 157 | #ifdef ASSERT |
ysr@777 | 158 | , _drain_in_progress(false) |
ysr@777 | 159 | , _drain_in_progress_yields(false) |
ysr@777 | 160 | #endif |
ysr@777 | 161 | {} |
ysr@777 | 162 | |
ysr@777 | 163 | void CMMarkStack::allocate(size_t size) { |
ysr@777 | 164 | _base = NEW_C_HEAP_ARRAY(oop, size); |
ysr@777 | 165 | if (_base == NULL) |
ysr@777 | 166 | vm_exit_during_initialization("Failed to allocate " |
ysr@777 | 167 | "CM region mark stack"); |
ysr@777 | 168 | _index = 0; |
ysr@777 | 169 | // QQQQ cast ... |
ysr@777 | 170 | _capacity = (jint) size; |
ysr@777 | 171 | _oops_do_bound = -1; |
ysr@777 | 172 | NOT_PRODUCT(_max_depth = 0); |
ysr@777 | 173 | } |
ysr@777 | 174 | |
ysr@777 | 175 | CMMarkStack::~CMMarkStack() { |
ysr@777 | 176 | if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base); |
ysr@777 | 177 | } |
ysr@777 | 178 | |
ysr@777 | 179 | void CMMarkStack::par_push(oop ptr) { |
ysr@777 | 180 | while (true) { |
ysr@777 | 181 | if (isFull()) { |
ysr@777 | 182 | _overflow = true; |
ysr@777 | 183 | return; |
ysr@777 | 184 | } |
ysr@777 | 185 | // Otherwise... |
ysr@777 | 186 | jint index = _index; |
ysr@777 | 187 | jint next_index = index+1; |
ysr@777 | 188 | jint res = Atomic::cmpxchg(next_index, &_index, index); |
ysr@777 | 189 | if (res == index) { |
ysr@777 | 190 | _base[index] = ptr; |
ysr@777 | 191 | // Note that we don't maintain this atomically. We could, but it |
ysr@777 | 192 | // doesn't seem necessary. |
ysr@777 | 193 | NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); |
ysr@777 | 194 | return; |
ysr@777 | 195 | } |
ysr@777 | 196 | // Otherwise, we need to try again. |
ysr@777 | 197 | } |
ysr@777 | 198 | } |
ysr@777 | 199 | |
ysr@777 | 200 | void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { |
ysr@777 | 201 | while (true) { |
ysr@777 | 202 | if (isFull()) { |
ysr@777 | 203 | _overflow = true; |
ysr@777 | 204 | return; |
ysr@777 | 205 | } |
ysr@777 | 206 | // Otherwise... |
ysr@777 | 207 | jint index = _index; |
ysr@777 | 208 | jint next_index = index + n; |
ysr@777 | 209 | if (next_index > _capacity) { |
ysr@777 | 210 | _overflow = true; |
ysr@777 | 211 | return; |
ysr@777 | 212 | } |
ysr@777 | 213 | jint res = Atomic::cmpxchg(next_index, &_index, index); |
ysr@777 | 214 | if (res == index) { |
ysr@777 | 215 | for (int i = 0; i < n; i++) { |
ysr@777 | 216 | int ind = index + i; |
ysr@777 | 217 | assert(ind < _capacity, "By overflow test above."); |
ysr@777 | 218 | _base[ind] = ptr_arr[i]; |
ysr@777 | 219 | } |
ysr@777 | 220 | NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); |
ysr@777 | 221 | return; |
ysr@777 | 222 | } |
ysr@777 | 223 | // Otherwise, we need to try again. |
ysr@777 | 224 | } |
ysr@777 | 225 | } |
ysr@777 | 226 | |
ysr@777 | 227 | |
ysr@777 | 228 | void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { |
ysr@777 | 229 | MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); |
ysr@777 | 230 | jint start = _index; |
ysr@777 | 231 | jint next_index = start + n; |
ysr@777 | 232 | if (next_index > _capacity) { |
ysr@777 | 233 | _overflow = true; |
ysr@777 | 234 | return; |
ysr@777 | 235 | } |
ysr@777 | 236 | // Otherwise. |
ysr@777 | 237 | _index = next_index; |
ysr@777 | 238 | for (int i = 0; i < n; i++) { |
ysr@777 | 239 | int ind = start + i; |
ysr@777 | 240 | guarantee(ind < _capacity, "By overflow test above."); |
ysr@777 | 241 | _base[ind] = ptr_arr[i]; |
ysr@777 | 242 | } |
ysr@777 | 243 | } |
ysr@777 | 244 | |
ysr@777 | 245 | |
ysr@777 | 246 | bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { |
ysr@777 | 247 | MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); |
ysr@777 | 248 | jint index = _index; |
ysr@777 | 249 | if (index == 0) { |
ysr@777 | 250 | *n = 0; |
ysr@777 | 251 | return false; |
ysr@777 | 252 | } else { |
ysr@777 | 253 | int k = MIN2(max, index); |
ysr@777 | 254 | jint new_ind = index - k; |
ysr@777 | 255 | for (int j = 0; j < k; j++) { |
ysr@777 | 256 | ptr_arr[j] = _base[new_ind + j]; |
ysr@777 | 257 | } |
ysr@777 | 258 | _index = new_ind; |
ysr@777 | 259 | *n = k; |
ysr@777 | 260 | return true; |
ysr@777 | 261 | } |
ysr@777 | 262 | } |
ysr@777 | 263 | |
ysr@777 | 264 | |
ysr@777 | 265 | CMRegionStack::CMRegionStack() : _base(NULL) {} |
ysr@777 | 266 | |
ysr@777 | 267 | void CMRegionStack::allocate(size_t size) { |
ysr@777 | 268 | _base = NEW_C_HEAP_ARRAY(MemRegion, size); |
ysr@777 | 269 | if (_base == NULL) |
ysr@777 | 270 | vm_exit_during_initialization("Failed to allocate " |
ysr@777 | 271 | "CM region mark stack"); |
ysr@777 | 272 | _index = 0; |
ysr@777 | 273 | // QQQQ cast ... |
ysr@777 | 274 | _capacity = (jint) size; |
ysr@777 | 275 | } |
ysr@777 | 276 | |
ysr@777 | 277 | CMRegionStack::~CMRegionStack() { |
ysr@777 | 278 | if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base); |
ysr@777 | 279 | } |
ysr@777 | 280 | |
ysr@777 | 281 | void CMRegionStack::push(MemRegion mr) { |
ysr@777 | 282 | assert(mr.word_size() > 0, "Precondition"); |
ysr@777 | 283 | while (true) { |
ysr@777 | 284 | if (isFull()) { |
ysr@777 | 285 | _overflow = true; |
ysr@777 | 286 | return; |
ysr@777 | 287 | } |
ysr@777 | 288 | // Otherwise... |
ysr@777 | 289 | jint index = _index; |
ysr@777 | 290 | jint next_index = index+1; |
ysr@777 | 291 | jint res = Atomic::cmpxchg(next_index, &_index, index); |
ysr@777 | 292 | if (res == index) { |
ysr@777 | 293 | _base[index] = mr; |
ysr@777 | 294 | return; |
ysr@777 | 295 | } |
ysr@777 | 296 | // Otherwise, we need to try again. |
ysr@777 | 297 | } |
ysr@777 | 298 | } |
ysr@777 | 299 | |
ysr@777 | 300 | MemRegion CMRegionStack::pop() { |
ysr@777 | 301 | while (true) { |
ysr@777 | 302 | // Otherwise... |
ysr@777 | 303 | jint index = _index; |
ysr@777 | 304 | |
ysr@777 | 305 | if (index == 0) { |
ysr@777 | 306 | return MemRegion(); |
ysr@777 | 307 | } |
ysr@777 | 308 | jint next_index = index-1; |
ysr@777 | 309 | jint res = Atomic::cmpxchg(next_index, &_index, index); |
ysr@777 | 310 | if (res == index) { |
ysr@777 | 311 | MemRegion mr = _base[next_index]; |
ysr@777 | 312 | if (mr.start() != NULL) { |
ysr@777 | 313 | tmp_guarantee_CM( mr.end() != NULL, "invariant" ); |
ysr@777 | 314 | tmp_guarantee_CM( mr.word_size() > 0, "invariant" ); |
ysr@777 | 315 | return mr; |
ysr@777 | 316 | } else { |
ysr@777 | 317 | // that entry was invalidated... let's skip it |
ysr@777 | 318 | tmp_guarantee_CM( mr.end() == NULL, "invariant" ); |
ysr@777 | 319 | } |
ysr@777 | 320 | } |
ysr@777 | 321 | // Otherwise, we need to try again. |
ysr@777 | 322 | } |
ysr@777 | 323 | } |
ysr@777 | 324 | |
ysr@777 | 325 | bool CMRegionStack::invalidate_entries_into_cset() { |
ysr@777 | 326 | bool result = false; |
ysr@777 | 327 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 328 | for (int i = 0; i < _oops_do_bound; ++i) { |
ysr@777 | 329 | MemRegion mr = _base[i]; |
ysr@777 | 330 | if (mr.start() != NULL) { |
ysr@777 | 331 | tmp_guarantee_CM( mr.end() != NULL, "invariant"); |
ysr@777 | 332 | tmp_guarantee_CM( mr.word_size() > 0, "invariant" ); |
ysr@777 | 333 | HeapRegion* hr = g1h->heap_region_containing(mr.start()); |
ysr@777 | 334 | tmp_guarantee_CM( hr != NULL, "invariant" ); |
ysr@777 | 335 | if (hr->in_collection_set()) { |
ysr@777 | 336 | // The region points into the collection set |
ysr@777 | 337 | _base[i] = MemRegion(); |
ysr@777 | 338 | result = true; |
ysr@777 | 339 | } |
ysr@777 | 340 | } else { |
ysr@777 | 341 | // that entry was invalidated... let's skip it |
ysr@777 | 342 | tmp_guarantee_CM( mr.end() == NULL, "invariant" ); |
ysr@777 | 343 | } |
ysr@777 | 344 | } |
ysr@777 | 345 | return result; |
ysr@777 | 346 | } |
ysr@777 | 347 | |
ysr@777 | 348 | template<class OopClosureClass> |
ysr@777 | 349 | bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { |
ysr@777 | 350 | assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after |
ysr@777 | 351 | || SafepointSynchronize::is_at_safepoint(), |
ysr@777 | 352 | "Drain recursion must be yield-safe."); |
ysr@777 | 353 | bool res = true; |
ysr@777 | 354 | debug_only(_drain_in_progress = true); |
ysr@777 | 355 | debug_only(_drain_in_progress_yields = yield_after); |
ysr@777 | 356 | while (!isEmpty()) { |
ysr@777 | 357 | oop newOop = pop(); |
ysr@777 | 358 | assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); |
ysr@777 | 359 | assert(newOop->is_oop(), "Expected an oop"); |
ysr@777 | 360 | assert(bm == NULL || bm->isMarked((HeapWord*)newOop), |
ysr@777 | 361 | "only grey objects on this stack"); |
ysr@777 | 362 | // iterate over the oops in this oop, marking and pushing |
ysr@777 | 363 | // the ones in CMS generation. |
ysr@777 | 364 | newOop->oop_iterate(cl); |
ysr@777 | 365 | if (yield_after && _cm->do_yield_check()) { |
ysr@777 | 366 | res = false; break; |
ysr@777 | 367 | } |
ysr@777 | 368 | } |
ysr@777 | 369 | debug_only(_drain_in_progress = false); |
ysr@777 | 370 | return res; |
ysr@777 | 371 | } |
ysr@777 | 372 | |
ysr@777 | 373 | void CMMarkStack::oops_do(OopClosure* f) { |
ysr@777 | 374 | if (_index == 0) return; |
ysr@777 | 375 | assert(_oops_do_bound != -1 && _oops_do_bound <= _index, |
ysr@777 | 376 | "Bound must be set."); |
ysr@777 | 377 | for (int i = 0; i < _oops_do_bound; i++) { |
ysr@777 | 378 | f->do_oop(&_base[i]); |
ysr@777 | 379 | } |
ysr@777 | 380 | _oops_do_bound = -1; |
ysr@777 | 381 | } |
ysr@777 | 382 | |
ysr@777 | 383 | bool ConcurrentMark::not_yet_marked(oop obj) const { |
ysr@777 | 384 | return (_g1h->is_obj_ill(obj) |
ysr@777 | 385 | || (_g1h->is_in_permanent(obj) |
ysr@777 | 386 | && !nextMarkBitMap()->isMarked((HeapWord*)obj))); |
ysr@777 | 387 | } |
ysr@777 | 388 | |
ysr@777 | 389 | #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away |
ysr@777 | 390 | #pragma warning( disable:4355 ) // 'this' : used in base member initializer list |
ysr@777 | 391 | #endif // _MSC_VER |
ysr@777 | 392 | |
ysr@777 | 393 | ConcurrentMark::ConcurrentMark(ReservedSpace rs, |
ysr@777 | 394 | int max_regions) : |
ysr@777 | 395 | _markBitMap1(rs, MinObjAlignment - 1), |
ysr@777 | 396 | _markBitMap2(rs, MinObjAlignment - 1), |
ysr@777 | 397 | |
ysr@777 | 398 | _parallel_marking_threads(0), |
ysr@777 | 399 | _sleep_factor(0.0), |
ysr@777 | 400 | _marking_task_overhead(1.0), |
ysr@777 | 401 | _cleanup_sleep_factor(0.0), |
ysr@777 | 402 | _cleanup_task_overhead(1.0), |
ysr@777 | 403 | _region_bm(max_regions, false /* in_resource_area*/), |
ysr@777 | 404 | _card_bm((rs.size() + CardTableModRefBS::card_size - 1) >> |
ysr@777 | 405 | CardTableModRefBS::card_shift, |
ysr@777 | 406 | false /* in_resource_area*/), |
ysr@777 | 407 | _prevMarkBitMap(&_markBitMap1), |
ysr@777 | 408 | _nextMarkBitMap(&_markBitMap2), |
ysr@777 | 409 | _at_least_one_mark_complete(false), |
ysr@777 | 410 | |
ysr@777 | 411 | _markStack(this), |
ysr@777 | 412 | _regionStack(), |
ysr@777 | 413 | // _finger set in set_non_marking_state |
ysr@777 | 414 | |
ysr@777 | 415 | _max_task_num(MAX2(ParallelGCThreads, (size_t)1)), |
ysr@777 | 416 | // _active_tasks set in set_non_marking_state |
ysr@777 | 417 | // _tasks set inside the constructor |
ysr@777 | 418 | _task_queues(new CMTaskQueueSet((int) _max_task_num)), |
ysr@777 | 419 | _terminator(ParallelTaskTerminator((int) _max_task_num, _task_queues)), |
ysr@777 | 420 | |
ysr@777 | 421 | _has_overflown(false), |
ysr@777 | 422 | _concurrent(false), |
tonyp@1054 | 423 | _has_aborted(false), |
tonyp@1054 | 424 | _restart_for_overflow(false), |
tonyp@1054 | 425 | _concurrent_marking_in_progress(false), |
tonyp@1054 | 426 | _should_gray_objects(false), |
ysr@777 | 427 | |
ysr@777 | 428 | // _verbose_level set below |
ysr@777 | 429 | |
ysr@777 | 430 | _init_times(), |
ysr@777 | 431 | _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), |
ysr@777 | 432 | _cleanup_times(), |
ysr@777 | 433 | _total_counting_time(0.0), |
ysr@777 | 434 | _total_rs_scrub_time(0.0), |
ysr@777 | 435 | |
tonyp@1371 | 436 | _parallel_workers(NULL) |
ysr@777 | 437 | { |
ysr@777 | 438 | CMVerboseLevel verbose_level = |
ysr@777 | 439 | (CMVerboseLevel) G1MarkingVerboseLevel; |
ysr@777 | 440 | if (verbose_level < no_verbose) |
ysr@777 | 441 | verbose_level = no_verbose; |
ysr@777 | 442 | if (verbose_level > high_verbose) |
ysr@777 | 443 | verbose_level = high_verbose; |
ysr@777 | 444 | _verbose_level = verbose_level; |
ysr@777 | 445 | |
ysr@777 | 446 | if (verbose_low()) |
ysr@777 | 447 | gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " |
ysr@777 | 448 | "heap end = "PTR_FORMAT, _heap_start, _heap_end); |
ysr@777 | 449 | |
johnc@1186 | 450 | _markStack.allocate(G1MarkStackSize); |
johnc@1186 | 451 | _regionStack.allocate(G1MarkRegionStackSize); |
ysr@777 | 452 | |
ysr@777 | 453 | // Create & start a ConcurrentMark thread. |
ysr@1280 | 454 | _cmThread = new ConcurrentMarkThread(this); |
ysr@1280 | 455 | assert(cmThread() != NULL, "CM Thread should have been created"); |
ysr@1280 | 456 | assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); |
ysr@1280 | 457 | |
ysr@777 | 458 | _g1h = G1CollectedHeap::heap(); |
ysr@777 | 459 | assert(CGC_lock != NULL, "Where's the CGC_lock?"); |
ysr@777 | 460 | assert(_markBitMap1.covers(rs), "_markBitMap1 inconsistency"); |
ysr@777 | 461 | assert(_markBitMap2.covers(rs), "_markBitMap2 inconsistency"); |
ysr@777 | 462 | |
ysr@777 | 463 | SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); |
ysr@777 | 464 | satb_qs.set_buffer_size(G1SATBLogBufferSize); |
ysr@777 | 465 | |
ysr@777 | 466 | int size = (int) MAX2(ParallelGCThreads, (size_t)1); |
ysr@777 | 467 | _par_cleanup_thread_state = NEW_C_HEAP_ARRAY(ParCleanupThreadState*, size); |
ysr@777 | 468 | for (int i = 0 ; i < size; i++) { |
ysr@777 | 469 | _par_cleanup_thread_state[i] = new ParCleanupThreadState; |
ysr@777 | 470 | } |
ysr@777 | 471 | |
ysr@777 | 472 | _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_task_num); |
ysr@777 | 473 | _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_task_num); |
ysr@777 | 474 | |
ysr@777 | 475 | // so that the assertion in MarkingTaskQueue::task_queue doesn't fail |
ysr@777 | 476 | _active_tasks = _max_task_num; |
ysr@777 | 477 | for (int i = 0; i < (int) _max_task_num; ++i) { |
ysr@777 | 478 | CMTaskQueue* task_queue = new CMTaskQueue(); |
ysr@777 | 479 | task_queue->initialize(); |
ysr@777 | 480 | _task_queues->register_queue(i, task_queue); |
ysr@777 | 481 | |
ysr@777 | 482 | _tasks[i] = new CMTask(i, this, task_queue, _task_queues); |
ysr@777 | 483 | _accum_task_vtime[i] = 0.0; |
ysr@777 | 484 | } |
ysr@777 | 485 | |
ysr@777 | 486 | if (ParallelMarkingThreads > ParallelGCThreads) { |
ysr@777 | 487 | vm_exit_during_initialization("Can't have more ParallelMarkingThreads " |
ysr@777 | 488 | "than ParallelGCThreads."); |
ysr@777 | 489 | } |
ysr@777 | 490 | if (ParallelGCThreads == 0) { |
ysr@777 | 491 | // if we are not running with any parallel GC threads we will not |
ysr@777 | 492 | // spawn any marking threads either |
ysr@777 | 493 | _parallel_marking_threads = 0; |
ysr@777 | 494 | _sleep_factor = 0.0; |
ysr@777 | 495 | _marking_task_overhead = 1.0; |
ysr@777 | 496 | } else { |
ysr@777 | 497 | if (ParallelMarkingThreads > 0) { |
johnc@1186 | 498 | // notice that ParallelMarkingThreads overwrites G1MarkingOverheadPercent |
ysr@777 | 499 | // if both are set |
ysr@777 | 500 | |
ysr@777 | 501 | _parallel_marking_threads = ParallelMarkingThreads; |
ysr@777 | 502 | _sleep_factor = 0.0; |
ysr@777 | 503 | _marking_task_overhead = 1.0; |
johnc@1186 | 504 | } else if (G1MarkingOverheadPercent > 0) { |
ysr@777 | 505 | // we will calculate the number of parallel marking threads |
ysr@777 | 506 | // based on a target overhead with respect to the soft real-time |
ysr@777 | 507 | // goal |
ysr@777 | 508 | |
johnc@1186 | 509 | double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; |
ysr@777 | 510 | double overall_cm_overhead = |
johnc@1186 | 511 | (double) MaxGCPauseMillis * marking_overhead / |
johnc@1186 | 512 | (double) GCPauseIntervalMillis; |
ysr@777 | 513 | double cpu_ratio = 1.0 / (double) os::processor_count(); |
ysr@777 | 514 | double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); |
ysr@777 | 515 | double marking_task_overhead = |
ysr@777 | 516 | overall_cm_overhead / marking_thread_num * |
ysr@777 | 517 | (double) os::processor_count(); |
ysr@777 | 518 | double sleep_factor = |
ysr@777 | 519 | (1.0 - marking_task_overhead) / marking_task_overhead; |
ysr@777 | 520 | |
ysr@777 | 521 | _parallel_marking_threads = (size_t) marking_thread_num; |
ysr@777 | 522 | _sleep_factor = sleep_factor; |
ysr@777 | 523 | _marking_task_overhead = marking_task_overhead; |
ysr@777 | 524 | } else { |
ysr@777 | 525 | _parallel_marking_threads = MAX2((ParallelGCThreads + 2) / 4, (size_t)1); |
ysr@777 | 526 | _sleep_factor = 0.0; |
ysr@777 | 527 | _marking_task_overhead = 1.0; |
ysr@777 | 528 | } |
ysr@777 | 529 | |
ysr@777 | 530 | if (parallel_marking_threads() > 1) |
ysr@777 | 531 | _cleanup_task_overhead = 1.0; |
ysr@777 | 532 | else |
ysr@777 | 533 | _cleanup_task_overhead = marking_task_overhead(); |
ysr@777 | 534 | _cleanup_sleep_factor = |
ysr@777 | 535 | (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); |
ysr@777 | 536 | |
ysr@777 | 537 | #if 0 |
ysr@777 | 538 | gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); |
ysr@777 | 539 | gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); |
ysr@777 | 540 | gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); |
ysr@777 | 541 | gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); |
ysr@777 | 542 | gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); |
ysr@777 | 543 | #endif |
ysr@777 | 544 | |
ysr@777 | 545 | guarantee( parallel_marking_threads() > 0, "peace of mind" ); |
tonyp@1454 | 546 | _parallel_workers = new WorkGang("G1 Parallel Marking Threads", |
ysr@777 | 547 | (int) parallel_marking_threads(), false, true); |
ysr@777 | 548 | if (_parallel_workers == NULL) |
ysr@777 | 549 | vm_exit_during_initialization("Failed necessary allocation."); |
ysr@777 | 550 | } |
ysr@777 | 551 | |
ysr@777 | 552 | // so that the call below can read a sensible value |
ysr@777 | 553 | _heap_start = (HeapWord*) rs.base(); |
ysr@777 | 554 | set_non_marking_state(); |
ysr@777 | 555 | } |
ysr@777 | 556 | |
ysr@777 | 557 | void ConcurrentMark::update_g1_committed(bool force) { |
ysr@777 | 558 | // If concurrent marking is not in progress, then we do not need to |
ysr@777 | 559 | // update _heap_end. This has a subtle and important |
ysr@777 | 560 | // side-effect. Imagine that two evacuation pauses happen between |
ysr@777 | 561 | // marking completion and remark. The first one can grow the |
ysr@777 | 562 | // heap (hence now the finger is below the heap end). Then, the |
ysr@777 | 563 | // second one could unnecessarily push regions on the region |
ysr@777 | 564 | // stack. This causes the invariant that the region stack is empty |
ysr@777 | 565 | // at the beginning of remark to be false. By ensuring that we do |
ysr@777 | 566 | // not observe heap expansions after marking is complete, then we do |
ysr@777 | 567 | // not have this problem. |
ysr@777 | 568 | if (!concurrent_marking_in_progress() && !force) |
ysr@777 | 569 | return; |
ysr@777 | 570 | |
ysr@777 | 571 | MemRegion committed = _g1h->g1_committed(); |
ysr@777 | 572 | tmp_guarantee_CM( committed.start() == _heap_start, |
ysr@777 | 573 | "start shouldn't change" ); |
ysr@777 | 574 | HeapWord* new_end = committed.end(); |
ysr@777 | 575 | if (new_end > _heap_end) { |
ysr@777 | 576 | // The heap has been expanded. |
ysr@777 | 577 | |
ysr@777 | 578 | _heap_end = new_end; |
ysr@777 | 579 | } |
ysr@777 | 580 | // Notice that the heap can also shrink. However, this only happens |
ysr@777 | 581 | // during a Full GC (at least currently) and the entire marking |
ysr@777 | 582 | // phase will bail out and the task will not be restarted. So, let's |
ysr@777 | 583 | // do nothing. |
ysr@777 | 584 | } |
ysr@777 | 585 | |
ysr@777 | 586 | void ConcurrentMark::reset() { |
ysr@777 | 587 | // Starting values for these two. This should be called in a STW |
ysr@777 | 588 | // phase. CM will be notified of any future g1_committed expansions |
ysr@777 | 589 | // will be at the end of evacuation pauses, when tasks are |
ysr@777 | 590 | // inactive. |
ysr@777 | 591 | MemRegion committed = _g1h->g1_committed(); |
ysr@777 | 592 | _heap_start = committed.start(); |
ysr@777 | 593 | _heap_end = committed.end(); |
ysr@777 | 594 | |
ysr@777 | 595 | guarantee( _heap_start != NULL && |
ysr@777 | 596 | _heap_end != NULL && |
ysr@777 | 597 | _heap_start < _heap_end, "heap bounds should look ok" ); |
ysr@777 | 598 | |
ysr@777 | 599 | // reset all the marking data structures and any necessary flags |
ysr@777 | 600 | clear_marking_state(); |
ysr@777 | 601 | |
ysr@777 | 602 | if (verbose_low()) |
ysr@777 | 603 | gclog_or_tty->print_cr("[global] resetting"); |
ysr@777 | 604 | |
ysr@777 | 605 | // We do reset all of them, since different phases will use |
ysr@777 | 606 | // different number of active threads. So, it's easiest to have all |
ysr@777 | 607 | // of them ready. |
ysr@777 | 608 | for (int i = 0; i < (int) _max_task_num; ++i) |
ysr@777 | 609 | _tasks[i]->reset(_nextMarkBitMap); |
ysr@777 | 610 | |
ysr@777 | 611 | // we need this to make sure that the flag is on during the evac |
ysr@777 | 612 | // pause with initial mark piggy-backed |
ysr@777 | 613 | set_concurrent_marking_in_progress(); |
ysr@777 | 614 | } |
ysr@777 | 615 | |
ysr@777 | 616 | void ConcurrentMark::set_phase(size_t active_tasks, bool concurrent) { |
ysr@777 | 617 | guarantee( active_tasks <= _max_task_num, "we should not have more" ); |
ysr@777 | 618 | |
ysr@777 | 619 | _active_tasks = active_tasks; |
ysr@777 | 620 | // Need to update the three data structures below according to the |
ysr@777 | 621 | // number of active threads for this phase. |
ysr@777 | 622 | _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); |
ysr@777 | 623 | _first_overflow_barrier_sync.set_n_workers((int) active_tasks); |
ysr@777 | 624 | _second_overflow_barrier_sync.set_n_workers((int) active_tasks); |
ysr@777 | 625 | |
ysr@777 | 626 | _concurrent = concurrent; |
ysr@777 | 627 | // We propagate this to all tasks, not just the active ones. |
ysr@777 | 628 | for (int i = 0; i < (int) _max_task_num; ++i) |
ysr@777 | 629 | _tasks[i]->set_concurrent(concurrent); |
ysr@777 | 630 | |
ysr@777 | 631 | if (concurrent) { |
ysr@777 | 632 | set_concurrent_marking_in_progress(); |
ysr@777 | 633 | } else { |
ysr@777 | 634 | // We currently assume that the concurrent flag has been set to |
ysr@777 | 635 | // false before we start remark. At this point we should also be |
ysr@777 | 636 | // in a STW phase. |
ysr@777 | 637 | guarantee( !concurrent_marking_in_progress(), "invariant" ); |
ysr@777 | 638 | guarantee( _finger == _heap_end, "only way to get here" ); |
ysr@777 | 639 | update_g1_committed(true); |
ysr@777 | 640 | } |
ysr@777 | 641 | } |
ysr@777 | 642 | |
ysr@777 | 643 | void ConcurrentMark::set_non_marking_state() { |
ysr@777 | 644 | // We set the global marking state to some default values when we're |
ysr@777 | 645 | // not doing marking. |
ysr@777 | 646 | clear_marking_state(); |
ysr@777 | 647 | _active_tasks = 0; |
ysr@777 | 648 | clear_concurrent_marking_in_progress(); |
ysr@777 | 649 | } |
ysr@777 | 650 | |
ysr@777 | 651 | ConcurrentMark::~ConcurrentMark() { |
ysr@777 | 652 | int size = (int) MAX2(ParallelGCThreads, (size_t)1); |
ysr@777 | 653 | for (int i = 0; i < size; i++) delete _par_cleanup_thread_state[i]; |
ysr@777 | 654 | FREE_C_HEAP_ARRAY(ParCleanupThreadState*, |
ysr@777 | 655 | _par_cleanup_thread_state); |
ysr@777 | 656 | |
ysr@777 | 657 | for (int i = 0; i < (int) _max_task_num; ++i) { |
ysr@777 | 658 | delete _task_queues->queue(i); |
ysr@777 | 659 | delete _tasks[i]; |
ysr@777 | 660 | } |
ysr@777 | 661 | delete _task_queues; |
ysr@777 | 662 | FREE_C_HEAP_ARRAY(CMTask*, _max_task_num); |
ysr@777 | 663 | } |
ysr@777 | 664 | |
ysr@777 | 665 | // This closure is used to mark refs into the g1 generation |
ysr@777 | 666 | // from external roots in the CMS bit map. |
ysr@777 | 667 | // Called at the first checkpoint. |
ysr@777 | 668 | // |
ysr@777 | 669 | |
ysr@777 | 670 | #define PRINT_REACHABLE_AT_INITIAL_MARK 0 |
ysr@777 | 671 | #if PRINT_REACHABLE_AT_INITIAL_MARK |
ysr@777 | 672 | static FILE* reachable_file = NULL; |
ysr@777 | 673 | |
ysr@777 | 674 | class PrintReachableClosure: public OopsInGenClosure { |
ysr@777 | 675 | CMBitMap* _bm; |
ysr@777 | 676 | int _level; |
ysr@777 | 677 | public: |
ysr@777 | 678 | PrintReachableClosure(CMBitMap* bm) : |
ysr@777 | 679 | _bm(bm), _level(0) { |
ysr@777 | 680 | guarantee(reachable_file != NULL, "pre-condition"); |
ysr@777 | 681 | } |
ysr@777 | 682 | void do_oop(oop* p) { |
ysr@777 | 683 | oop obj = *p; |
ysr@777 | 684 | HeapWord* obj_addr = (HeapWord*)obj; |
ysr@777 | 685 | if (obj == NULL) return; |
ysr@777 | 686 | fprintf(reachable_file, "%d: "PTR_FORMAT" -> "PTR_FORMAT" (%d)\n", |
ysr@777 | 687 | _level, p, (void*) obj, _bm->isMarked(obj_addr)); |
ysr@777 | 688 | if (!_bm->isMarked(obj_addr)) { |
ysr@777 | 689 | _bm->mark(obj_addr); |
ysr@777 | 690 | _level++; |
ysr@777 | 691 | obj->oop_iterate(this); |
ysr@777 | 692 | _level--; |
ysr@777 | 693 | } |
ysr@777 | 694 | } |
ysr@777 | 695 | }; |
ysr@777 | 696 | #endif // PRINT_REACHABLE_AT_INITIAL_MARK |
ysr@777 | 697 | |
ysr@777 | 698 | #define SEND_HEAP_DUMP_TO_FILE 0 |
ysr@777 | 699 | #if SEND_HEAP_DUMP_TO_FILE |
ysr@777 | 700 | static FILE* heap_dump_file = NULL; |
ysr@777 | 701 | #endif // SEND_HEAP_DUMP_TO_FILE |
ysr@777 | 702 | |
ysr@777 | 703 | void ConcurrentMark::clearNextBitmap() { |
ysr@777 | 704 | guarantee(!G1CollectedHeap::heap()->mark_in_progress(), "Precondition."); |
ysr@777 | 705 | |
ysr@777 | 706 | // clear the mark bitmap (no grey objects to start with). |
ysr@777 | 707 | // We need to do this in chunks and offer to yield in between |
ysr@777 | 708 | // each chunk. |
ysr@777 | 709 | HeapWord* start = _nextMarkBitMap->startWord(); |
ysr@777 | 710 | HeapWord* end = _nextMarkBitMap->endWord(); |
ysr@777 | 711 | HeapWord* cur = start; |
ysr@777 | 712 | size_t chunkSize = M; |
ysr@777 | 713 | while (cur < end) { |
ysr@777 | 714 | HeapWord* next = cur + chunkSize; |
ysr@777 | 715 | if (next > end) |
ysr@777 | 716 | next = end; |
ysr@777 | 717 | MemRegion mr(cur,next); |
ysr@777 | 718 | _nextMarkBitMap->clearRange(mr); |
ysr@777 | 719 | cur = next; |
ysr@777 | 720 | do_yield_check(); |
ysr@777 | 721 | } |
ysr@777 | 722 | } |
ysr@777 | 723 | |
ysr@777 | 724 | class NoteStartOfMarkHRClosure: public HeapRegionClosure { |
ysr@777 | 725 | public: |
ysr@777 | 726 | bool doHeapRegion(HeapRegion* r) { |
ysr@777 | 727 | if (!r->continuesHumongous()) { |
ysr@777 | 728 | r->note_start_of_marking(true); |
ysr@777 | 729 | } |
ysr@777 | 730 | return false; |
ysr@777 | 731 | } |
ysr@777 | 732 | }; |
ysr@777 | 733 | |
ysr@777 | 734 | void ConcurrentMark::checkpointRootsInitialPre() { |
ysr@777 | 735 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 736 | G1CollectorPolicy* g1p = g1h->g1_policy(); |
ysr@777 | 737 | |
ysr@777 | 738 | _has_aborted = false; |
ysr@777 | 739 | |
ysr@777 | 740 | // Find all the reachable objects... |
ysr@777 | 741 | #if PRINT_REACHABLE_AT_INITIAL_MARK |
ysr@777 | 742 | guarantee(reachable_file == NULL, "Protocol"); |
ysr@777 | 743 | char fn_buf[100]; |
ysr@777 | 744 | sprintf(fn_buf, "/tmp/reachable.txt.%d", os::current_process_id()); |
ysr@777 | 745 | reachable_file = fopen(fn_buf, "w"); |
ysr@777 | 746 | // clear the mark bitmap (no grey objects to start with) |
ysr@777 | 747 | _nextMarkBitMap->clearAll(); |
ysr@777 | 748 | PrintReachableClosure prcl(_nextMarkBitMap); |
jrose@1424 | 749 | g1h->process_strong_roots(true, // activate StrongRootsScope |
ysr@777 | 750 | false, // fake perm gen collection |
ysr@777 | 751 | SharedHeap::SO_AllClasses, |
ysr@777 | 752 | &prcl, // Regular roots |
jrose@1424 | 753 | NULL, // do not visit active blobs |
ysr@777 | 754 | &prcl // Perm Gen Roots |
ysr@777 | 755 | ); |
ysr@777 | 756 | // The root iteration above "consumed" dirty cards in the perm gen. |
ysr@777 | 757 | // Therefore, as a shortcut, we dirty all such cards. |
ysr@777 | 758 | g1h->rem_set()->invalidate(g1h->perm_gen()->used_region(), false); |
ysr@777 | 759 | fclose(reachable_file); |
ysr@777 | 760 | reachable_file = NULL; |
ysr@777 | 761 | // clear the mark bitmap again. |
ysr@777 | 762 | _nextMarkBitMap->clearAll(); |
ysr@777 | 763 | COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); |
ysr@777 | 764 | COMPILER2_PRESENT(DerivedPointerTable::clear()); |
ysr@777 | 765 | #endif // PRINT_REACHABLE_AT_INITIAL_MARK |
ysr@777 | 766 | |
ysr@777 | 767 | // Initialise marking structures. This has to be done in a STW phase. |
ysr@777 | 768 | reset(); |
ysr@777 | 769 | } |
ysr@777 | 770 | |
ysr@777 | 771 | class CMMarkRootsClosure: public OopsInGenClosure { |
ysr@777 | 772 | private: |
ysr@777 | 773 | ConcurrentMark* _cm; |
ysr@777 | 774 | G1CollectedHeap* _g1h; |
ysr@777 | 775 | bool _do_barrier; |
ysr@777 | 776 | |
ysr@777 | 777 | public: |
ysr@777 | 778 | CMMarkRootsClosure(ConcurrentMark* cm, |
ysr@777 | 779 | G1CollectedHeap* g1h, |
ysr@777 | 780 | bool do_barrier) : _cm(cm), _g1h(g1h), |
ysr@777 | 781 | _do_barrier(do_barrier) { } |
ysr@777 | 782 | |
ysr@1280 | 783 | virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
ysr@1280 | 784 | virtual void do_oop( oop* p) { do_oop_work(p); } |
ysr@1280 | 785 | |
ysr@1280 | 786 | template <class T> void do_oop_work(T* p) { |
ysr@1280 | 787 | T heap_oop = oopDesc::load_heap_oop(p); |
ysr@1280 | 788 | if (!oopDesc::is_null(heap_oop)) { |
ysr@1280 | 789 | oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
ysr@1280 | 790 | assert(obj->is_oop() || obj->mark() == NULL, |
ysr@777 | 791 | "expected an oop, possibly with mark word displaced"); |
ysr@1280 | 792 | HeapWord* addr = (HeapWord*)obj; |
ysr@777 | 793 | if (_g1h->is_in_g1_reserved(addr)) { |
ysr@1280 | 794 | _cm->grayRoot(obj); |
ysr@777 | 795 | } |
ysr@777 | 796 | } |
ysr@777 | 797 | if (_do_barrier) { |
ysr@777 | 798 | assert(!_g1h->is_in_g1_reserved(p), |
ysr@777 | 799 | "Should be called on external roots"); |
ysr@777 | 800 | do_barrier(p); |
ysr@777 | 801 | } |
ysr@777 | 802 | } |
ysr@777 | 803 | }; |
ysr@777 | 804 | |
ysr@777 | 805 | void ConcurrentMark::checkpointRootsInitialPost() { |
ysr@777 | 806 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 807 | |
ysr@777 | 808 | // For each region note start of marking. |
ysr@777 | 809 | NoteStartOfMarkHRClosure startcl; |
ysr@777 | 810 | g1h->heap_region_iterate(&startcl); |
ysr@777 | 811 | |
ysr@777 | 812 | // Start weak-reference discovery. |
ysr@777 | 813 | ReferenceProcessor* rp = g1h->ref_processor(); |
ysr@777 | 814 | rp->verify_no_references_recorded(); |
ysr@777 | 815 | rp->enable_discovery(); // enable ("weak") refs discovery |
ysr@892 | 816 | rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle |
ysr@777 | 817 | |
ysr@777 | 818 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 819 | satb_mq_set.set_process_completed_threshold(G1SATBProcessCompletedThreshold); |
ysr@777 | 820 | satb_mq_set.set_active_all_threads(true); |
ysr@777 | 821 | |
ysr@777 | 822 | // update_g1_committed() will be called at the end of an evac pause |
ysr@777 | 823 | // when marking is on. So, it's also called at the end of the |
ysr@777 | 824 | // initial-mark pause to update the heap end, if the heap expands |
ysr@777 | 825 | // during it. No need to call it here. |
ysr@777 | 826 | } |
ysr@777 | 827 | |
ysr@777 | 828 | // Checkpoint the roots into this generation from outside |
ysr@777 | 829 | // this generation. [Note this initial checkpoint need only |
ysr@777 | 830 | // be approximate -- we'll do a catch up phase subsequently.] |
ysr@777 | 831 | void ConcurrentMark::checkpointRootsInitial() { |
ysr@777 | 832 | assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); |
ysr@777 | 833 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 834 | |
ysr@777 | 835 | double start = os::elapsedTime(); |
ysr@777 | 836 | |
ysr@777 | 837 | G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); |
ysr@777 | 838 | g1p->record_concurrent_mark_init_start(); |
ysr@777 | 839 | checkpointRootsInitialPre(); |
ysr@777 | 840 | |
ysr@777 | 841 | // YSR: when concurrent precleaning is in place, we'll |
ysr@777 | 842 | // need to clear the cached card table here |
ysr@777 | 843 | |
ysr@777 | 844 | ResourceMark rm; |
ysr@777 | 845 | HandleMark hm; |
ysr@777 | 846 | |
ysr@777 | 847 | g1h->ensure_parsability(false); |
ysr@777 | 848 | g1h->perm_gen()->save_marks(); |
ysr@777 | 849 | |
ysr@777 | 850 | CMMarkRootsClosure notOlder(this, g1h, false); |
ysr@777 | 851 | CMMarkRootsClosure older(this, g1h, true); |
ysr@777 | 852 | |
ysr@777 | 853 | g1h->set_marking_started(); |
ysr@777 | 854 | g1h->rem_set()->prepare_for_younger_refs_iterate(false); |
ysr@777 | 855 | |
jrose@1424 | 856 | g1h->process_strong_roots(true, // activate StrongRootsScope |
jrose@1424 | 857 | false, // fake perm gen collection |
ysr@777 | 858 | SharedHeap::SO_AllClasses, |
ysr@777 | 859 | ¬Older, // Regular roots |
jrose@1424 | 860 | NULL, // do not visit active blobs |
ysr@777 | 861 | &older // Perm Gen Roots |
ysr@777 | 862 | ); |
ysr@777 | 863 | checkpointRootsInitialPost(); |
ysr@777 | 864 | |
ysr@777 | 865 | // Statistics. |
ysr@777 | 866 | double end = os::elapsedTime(); |
ysr@777 | 867 | _init_times.add((end - start) * 1000.0); |
ysr@777 | 868 | |
ysr@777 | 869 | g1p->record_concurrent_mark_init_end(); |
ysr@777 | 870 | } |
ysr@777 | 871 | |
ysr@777 | 872 | /* |
ysr@777 | 873 | Notice that in the next two methods, we actually leave the STS |
ysr@777 | 874 | during the barrier sync and join it immediately afterwards. If we |
ysr@777 | 875 | do not do this, this then the following deadlock can occur: one |
ysr@777 | 876 | thread could be in the barrier sync code, waiting for the other |
ysr@777 | 877 | thread to also sync up, whereas another one could be trying to |
ysr@777 | 878 | yield, while also waiting for the other threads to sync up too. |
ysr@777 | 879 | |
ysr@777 | 880 | Because the thread that does the sync barrier has left the STS, it |
ysr@777 | 881 | is possible to be suspended for a Full GC or an evacuation pause |
ysr@777 | 882 | could occur. This is actually safe, since the entering the sync |
ysr@777 | 883 | barrier is one of the last things do_marking_step() does, and it |
ysr@777 | 884 | doesn't manipulate any data structures afterwards. |
ysr@777 | 885 | */ |
ysr@777 | 886 | |
ysr@777 | 887 | void ConcurrentMark::enter_first_sync_barrier(int task_num) { |
ysr@777 | 888 | if (verbose_low()) |
ysr@777 | 889 | gclog_or_tty->print_cr("[%d] entering first barrier", task_num); |
ysr@777 | 890 | |
ysr@777 | 891 | ConcurrentGCThread::stsLeave(); |
ysr@777 | 892 | _first_overflow_barrier_sync.enter(); |
ysr@777 | 893 | ConcurrentGCThread::stsJoin(); |
ysr@777 | 894 | // at this point everyone should have synced up and not be doing any |
ysr@777 | 895 | // more work |
ysr@777 | 896 | |
ysr@777 | 897 | if (verbose_low()) |
ysr@777 | 898 | gclog_or_tty->print_cr("[%d] leaving first barrier", task_num); |
ysr@777 | 899 | |
ysr@777 | 900 | // let task 0 do this |
ysr@777 | 901 | if (task_num == 0) { |
ysr@777 | 902 | // task 0 is responsible for clearing the global data structures |
ysr@777 | 903 | clear_marking_state(); |
ysr@777 | 904 | |
ysr@777 | 905 | if (PrintGC) { |
ysr@777 | 906 | gclog_or_tty->date_stamp(PrintGCDateStamps); |
ysr@777 | 907 | gclog_or_tty->stamp(PrintGCTimeStamps); |
ysr@777 | 908 | gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); |
ysr@777 | 909 | } |
ysr@777 | 910 | } |
ysr@777 | 911 | |
ysr@777 | 912 | // after this, each task should reset its own data structures then |
ysr@777 | 913 | // then go into the second barrier |
ysr@777 | 914 | } |
ysr@777 | 915 | |
ysr@777 | 916 | void ConcurrentMark::enter_second_sync_barrier(int task_num) { |
ysr@777 | 917 | if (verbose_low()) |
ysr@777 | 918 | gclog_or_tty->print_cr("[%d] entering second barrier", task_num); |
ysr@777 | 919 | |
ysr@777 | 920 | ConcurrentGCThread::stsLeave(); |
ysr@777 | 921 | _second_overflow_barrier_sync.enter(); |
ysr@777 | 922 | ConcurrentGCThread::stsJoin(); |
ysr@777 | 923 | // at this point everything should be re-initialised and ready to go |
ysr@777 | 924 | |
ysr@777 | 925 | if (verbose_low()) |
ysr@777 | 926 | gclog_or_tty->print_cr("[%d] leaving second barrier", task_num); |
ysr@777 | 927 | } |
ysr@777 | 928 | |
ysr@777 | 929 | void ConcurrentMark::grayRoot(oop p) { |
ysr@777 | 930 | HeapWord* addr = (HeapWord*) p; |
ysr@777 | 931 | // We can't really check against _heap_start and _heap_end, since it |
ysr@777 | 932 | // is possible during an evacuation pause with piggy-backed |
ysr@777 | 933 | // initial-mark that the committed space is expanded during the |
ysr@777 | 934 | // pause without CM observing this change. So the assertions below |
ysr@777 | 935 | // is a bit conservative; but better than nothing. |
ysr@777 | 936 | tmp_guarantee_CM( _g1h->g1_committed().contains(addr), |
ysr@777 | 937 | "address should be within the heap bounds" ); |
ysr@777 | 938 | |
ysr@777 | 939 | if (!_nextMarkBitMap->isMarked(addr)) |
ysr@777 | 940 | _nextMarkBitMap->parMark(addr); |
ysr@777 | 941 | } |
ysr@777 | 942 | |
ysr@777 | 943 | void ConcurrentMark::grayRegionIfNecessary(MemRegion mr) { |
ysr@777 | 944 | // The objects on the region have already been marked "in bulk" by |
ysr@777 | 945 | // the caller. We only need to decide whether to push the region on |
ysr@777 | 946 | // the region stack or not. |
ysr@777 | 947 | |
ysr@777 | 948 | if (!concurrent_marking_in_progress() || !_should_gray_objects) |
ysr@777 | 949 | // We're done with marking and waiting for remark. We do not need to |
ysr@777 | 950 | // push anything else on the region stack. |
ysr@777 | 951 | return; |
ysr@777 | 952 | |
ysr@777 | 953 | HeapWord* finger = _finger; |
ysr@777 | 954 | |
ysr@777 | 955 | if (verbose_low()) |
ysr@777 | 956 | gclog_or_tty->print_cr("[global] attempting to push " |
ysr@777 | 957 | "region ["PTR_FORMAT", "PTR_FORMAT"), finger is at " |
ysr@777 | 958 | PTR_FORMAT, mr.start(), mr.end(), finger); |
ysr@777 | 959 | |
ysr@777 | 960 | if (mr.start() < finger) { |
ysr@777 | 961 | // The finger is always heap region aligned and it is not possible |
ysr@777 | 962 | // for mr to span heap regions. |
ysr@777 | 963 | tmp_guarantee_CM( mr.end() <= finger, "invariant" ); |
ysr@777 | 964 | |
ysr@777 | 965 | tmp_guarantee_CM( mr.start() <= mr.end() && |
ysr@777 | 966 | _heap_start <= mr.start() && |
ysr@777 | 967 | mr.end() <= _heap_end, |
ysr@777 | 968 | "region boundaries should fall within the committed space" ); |
ysr@777 | 969 | if (verbose_low()) |
ysr@777 | 970 | gclog_or_tty->print_cr("[global] region ["PTR_FORMAT", "PTR_FORMAT") " |
ysr@777 | 971 | "below the finger, pushing it", |
ysr@777 | 972 | mr.start(), mr.end()); |
ysr@777 | 973 | |
ysr@777 | 974 | if (!region_stack_push(mr)) { |
ysr@777 | 975 | if (verbose_low()) |
ysr@777 | 976 | gclog_or_tty->print_cr("[global] region stack has overflown."); |
ysr@777 | 977 | } |
ysr@777 | 978 | } |
ysr@777 | 979 | } |
ysr@777 | 980 | |
ysr@777 | 981 | void ConcurrentMark::markAndGrayObjectIfNecessary(oop p) { |
ysr@777 | 982 | // The object is not marked by the caller. We need to at least mark |
ysr@777 | 983 | // it and maybe push in on the stack. |
ysr@777 | 984 | |
ysr@777 | 985 | HeapWord* addr = (HeapWord*)p; |
ysr@777 | 986 | if (!_nextMarkBitMap->isMarked(addr)) { |
ysr@777 | 987 | // We definitely need to mark it, irrespective whether we bail out |
ysr@777 | 988 | // because we're done with marking. |
ysr@777 | 989 | if (_nextMarkBitMap->parMark(addr)) { |
ysr@777 | 990 | if (!concurrent_marking_in_progress() || !_should_gray_objects) |
ysr@777 | 991 | // If we're done with concurrent marking and we're waiting for |
ysr@777 | 992 | // remark, then we're not pushing anything on the stack. |
ysr@777 | 993 | return; |
ysr@777 | 994 | |
ysr@777 | 995 | // No OrderAccess:store_load() is needed. It is implicit in the |
ysr@777 | 996 | // CAS done in parMark(addr) above |
ysr@777 | 997 | HeapWord* finger = _finger; |
ysr@777 | 998 | |
ysr@777 | 999 | if (addr < finger) { |
ysr@777 | 1000 | if (!mark_stack_push(oop(addr))) { |
ysr@777 | 1001 | if (verbose_low()) |
ysr@777 | 1002 | gclog_or_tty->print_cr("[global] global stack overflow " |
ysr@777 | 1003 | "during parMark"); |
ysr@777 | 1004 | } |
ysr@777 | 1005 | } |
ysr@777 | 1006 | } |
ysr@777 | 1007 | } |
ysr@777 | 1008 | } |
ysr@777 | 1009 | |
ysr@777 | 1010 | class CMConcurrentMarkingTask: public AbstractGangTask { |
ysr@777 | 1011 | private: |
ysr@777 | 1012 | ConcurrentMark* _cm; |
ysr@777 | 1013 | ConcurrentMarkThread* _cmt; |
ysr@777 | 1014 | |
ysr@777 | 1015 | public: |
ysr@777 | 1016 | void work(int worker_i) { |
ysr@777 | 1017 | guarantee( Thread::current()->is_ConcurrentGC_thread(), |
ysr@777 | 1018 | "this should only be done by a conc GC thread" ); |
ysr@777 | 1019 | |
ysr@777 | 1020 | double start_vtime = os::elapsedVTime(); |
ysr@777 | 1021 | |
ysr@777 | 1022 | ConcurrentGCThread::stsJoin(); |
ysr@777 | 1023 | |
ysr@777 | 1024 | guarantee( (size_t)worker_i < _cm->active_tasks(), "invariant" ); |
ysr@777 | 1025 | CMTask* the_task = _cm->task(worker_i); |
ysr@777 | 1026 | the_task->record_start_time(); |
ysr@777 | 1027 | if (!_cm->has_aborted()) { |
ysr@777 | 1028 | do { |
ysr@777 | 1029 | double start_vtime_sec = os::elapsedVTime(); |
ysr@777 | 1030 | double start_time_sec = os::elapsedTime(); |
ysr@777 | 1031 | the_task->do_marking_step(10.0); |
ysr@777 | 1032 | double end_time_sec = os::elapsedTime(); |
ysr@777 | 1033 | double end_vtime_sec = os::elapsedVTime(); |
ysr@777 | 1034 | double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; |
ysr@777 | 1035 | double elapsed_time_sec = end_time_sec - start_time_sec; |
ysr@777 | 1036 | _cm->clear_has_overflown(); |
ysr@777 | 1037 | |
ysr@777 | 1038 | bool ret = _cm->do_yield_check(worker_i); |
ysr@777 | 1039 | |
ysr@777 | 1040 | jlong sleep_time_ms; |
ysr@777 | 1041 | if (!_cm->has_aborted() && the_task->has_aborted()) { |
ysr@777 | 1042 | sleep_time_ms = |
ysr@777 | 1043 | (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); |
ysr@777 | 1044 | ConcurrentGCThread::stsLeave(); |
ysr@777 | 1045 | os::sleep(Thread::current(), sleep_time_ms, false); |
ysr@777 | 1046 | ConcurrentGCThread::stsJoin(); |
ysr@777 | 1047 | } |
ysr@777 | 1048 | double end_time2_sec = os::elapsedTime(); |
ysr@777 | 1049 | double elapsed_time2_sec = end_time2_sec - start_time_sec; |
ysr@777 | 1050 | |
ysr@777 | 1051 | #if 0 |
ysr@777 | 1052 | gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, " |
ysr@777 | 1053 | "overhead %1.4lf", |
ysr@777 | 1054 | elapsed_vtime_sec * 1000.0, (double) sleep_time_ms, |
ysr@777 | 1055 | the_task->conc_overhead(os::elapsedTime()) * 8.0); |
ysr@777 | 1056 | gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms", |
ysr@777 | 1057 | elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0); |
ysr@777 | 1058 | #endif |
ysr@777 | 1059 | } while (!_cm->has_aborted() && the_task->has_aborted()); |
ysr@777 | 1060 | } |
ysr@777 | 1061 | the_task->record_end_time(); |
ysr@777 | 1062 | guarantee( !the_task->has_aborted() || _cm->has_aborted(), "invariant" ); |
ysr@777 | 1063 | |
ysr@777 | 1064 | ConcurrentGCThread::stsLeave(); |
ysr@777 | 1065 | |
ysr@777 | 1066 | double end_vtime = os::elapsedVTime(); |
ysr@777 | 1067 | _cm->update_accum_task_vtime(worker_i, end_vtime - start_vtime); |
ysr@777 | 1068 | } |
ysr@777 | 1069 | |
ysr@777 | 1070 | CMConcurrentMarkingTask(ConcurrentMark* cm, |
ysr@777 | 1071 | ConcurrentMarkThread* cmt) : |
ysr@777 | 1072 | AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } |
ysr@777 | 1073 | |
ysr@777 | 1074 | ~CMConcurrentMarkingTask() { } |
ysr@777 | 1075 | }; |
ysr@777 | 1076 | |
ysr@777 | 1077 | void ConcurrentMark::markFromRoots() { |
ysr@777 | 1078 | // we might be tempted to assert that: |
ysr@777 | 1079 | // assert(asynch == !SafepointSynchronize::is_at_safepoint(), |
ysr@777 | 1080 | // "inconsistent argument?"); |
ysr@777 | 1081 | // However that wouldn't be right, because it's possible that |
ysr@777 | 1082 | // a safepoint is indeed in progress as a younger generation |
ysr@777 | 1083 | // stop-the-world GC happens even as we mark in this generation. |
ysr@777 | 1084 | |
ysr@777 | 1085 | _restart_for_overflow = false; |
ysr@777 | 1086 | |
ysr@777 | 1087 | set_phase(MAX2((size_t) 1, parallel_marking_threads()), true); |
ysr@777 | 1088 | |
ysr@777 | 1089 | CMConcurrentMarkingTask markingTask(this, cmThread()); |
ysr@777 | 1090 | if (parallel_marking_threads() > 0) |
ysr@777 | 1091 | _parallel_workers->run_task(&markingTask); |
ysr@777 | 1092 | else |
ysr@777 | 1093 | markingTask.work(0); |
ysr@777 | 1094 | print_stats(); |
ysr@777 | 1095 | } |
ysr@777 | 1096 | |
ysr@777 | 1097 | void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { |
ysr@777 | 1098 | // world is stopped at this checkpoint |
ysr@777 | 1099 | assert(SafepointSynchronize::is_at_safepoint(), |
ysr@777 | 1100 | "world should be stopped"); |
ysr@777 | 1101 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 1102 | |
ysr@777 | 1103 | // If a full collection has happened, we shouldn't do this. |
ysr@777 | 1104 | if (has_aborted()) { |
ysr@777 | 1105 | g1h->set_marking_complete(); // So bitmap clearing isn't confused |
ysr@777 | 1106 | return; |
ysr@777 | 1107 | } |
ysr@777 | 1108 | |
ysr@1280 | 1109 | if (VerifyDuringGC) { |
ysr@1280 | 1110 | HandleMark hm; // handle scope |
ysr@1280 | 1111 | gclog_or_tty->print(" VerifyDuringGC:(before)"); |
ysr@1280 | 1112 | Universe::heap()->prepare_for_verify(); |
ysr@1280 | 1113 | Universe::verify(true, false, true); |
ysr@1280 | 1114 | } |
ysr@1280 | 1115 | |
ysr@777 | 1116 | G1CollectorPolicy* g1p = g1h->g1_policy(); |
ysr@777 | 1117 | g1p->record_concurrent_mark_remark_start(); |
ysr@777 | 1118 | |
ysr@777 | 1119 | double start = os::elapsedTime(); |
ysr@777 | 1120 | |
ysr@777 | 1121 | checkpointRootsFinalWork(); |
ysr@777 | 1122 | |
ysr@777 | 1123 | double mark_work_end = os::elapsedTime(); |
ysr@777 | 1124 | |
ysr@777 | 1125 | weakRefsWork(clear_all_soft_refs); |
ysr@777 | 1126 | |
ysr@777 | 1127 | if (has_overflown()) { |
ysr@777 | 1128 | // Oops. We overflowed. Restart concurrent marking. |
ysr@777 | 1129 | _restart_for_overflow = true; |
ysr@777 | 1130 | // Clear the flag. We do not need it any more. |
ysr@777 | 1131 | clear_has_overflown(); |
ysr@777 | 1132 | if (G1TraceMarkStackOverflow) |
ysr@777 | 1133 | gclog_or_tty->print_cr("\nRemark led to restart for overflow."); |
ysr@777 | 1134 | } else { |
ysr@777 | 1135 | // We're done with marking. |
ysr@777 | 1136 | JavaThread::satb_mark_queue_set().set_active_all_threads(false); |
tonyp@1246 | 1137 | |
tonyp@1246 | 1138 | if (VerifyDuringGC) { |
ysr@1280 | 1139 | HandleMark hm; // handle scope |
ysr@1280 | 1140 | gclog_or_tty->print(" VerifyDuringGC:(after)"); |
ysr@1280 | 1141 | Universe::heap()->prepare_for_verify(); |
ysr@1280 | 1142 | Universe::heap()->verify(/* allow_dirty */ true, |
ysr@1280 | 1143 | /* silent */ false, |
ysr@1280 | 1144 | /* use_prev_marking */ false); |
tonyp@1246 | 1145 | } |
ysr@777 | 1146 | } |
ysr@777 | 1147 | |
ysr@777 | 1148 | #if VERIFY_OBJS_PROCESSED |
ysr@777 | 1149 | _scan_obj_cl.objs_processed = 0; |
ysr@777 | 1150 | ThreadLocalObjQueue::objs_enqueued = 0; |
ysr@777 | 1151 | #endif |
ysr@777 | 1152 | |
ysr@777 | 1153 | // Statistics |
ysr@777 | 1154 | double now = os::elapsedTime(); |
ysr@777 | 1155 | _remark_mark_times.add((mark_work_end - start) * 1000.0); |
ysr@777 | 1156 | _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); |
ysr@777 | 1157 | _remark_times.add((now - start) * 1000.0); |
ysr@777 | 1158 | |
ysr@777 | 1159 | g1p->record_concurrent_mark_remark_end(); |
ysr@777 | 1160 | } |
ysr@777 | 1161 | |
ysr@777 | 1162 | |
ysr@777 | 1163 | #define CARD_BM_TEST_MODE 0 |
ysr@777 | 1164 | |
ysr@777 | 1165 | class CalcLiveObjectsClosure: public HeapRegionClosure { |
ysr@777 | 1166 | |
ysr@777 | 1167 | CMBitMapRO* _bm; |
ysr@777 | 1168 | ConcurrentMark* _cm; |
ysr@777 | 1169 | bool _changed; |
ysr@777 | 1170 | bool _yield; |
ysr@777 | 1171 | size_t _words_done; |
ysr@777 | 1172 | size_t _tot_live; |
ysr@777 | 1173 | size_t _tot_used; |
ysr@777 | 1174 | size_t _regions_done; |
ysr@777 | 1175 | double _start_vtime_sec; |
ysr@777 | 1176 | |
ysr@777 | 1177 | BitMap* _region_bm; |
ysr@777 | 1178 | BitMap* _card_bm; |
ysr@777 | 1179 | intptr_t _bottom_card_num; |
ysr@777 | 1180 | bool _final; |
ysr@777 | 1181 | |
ysr@777 | 1182 | void mark_card_num_range(intptr_t start_card_num, intptr_t last_card_num) { |
ysr@777 | 1183 | for (intptr_t i = start_card_num; i <= last_card_num; i++) { |
ysr@777 | 1184 | #if CARD_BM_TEST_MODE |
ysr@777 | 1185 | guarantee(_card_bm->at(i - _bottom_card_num), |
ysr@777 | 1186 | "Should already be set."); |
ysr@777 | 1187 | #else |
ysr@777 | 1188 | _card_bm->par_at_put(i - _bottom_card_num, 1); |
ysr@777 | 1189 | #endif |
ysr@777 | 1190 | } |
ysr@777 | 1191 | } |
ysr@777 | 1192 | |
ysr@777 | 1193 | public: |
ysr@777 | 1194 | CalcLiveObjectsClosure(bool final, |
ysr@777 | 1195 | CMBitMapRO *bm, ConcurrentMark *cm, |
tonyp@1371 | 1196 | BitMap* region_bm, BitMap* card_bm) : |
ysr@777 | 1197 | _bm(bm), _cm(cm), _changed(false), _yield(true), |
ysr@777 | 1198 | _words_done(0), _tot_live(0), _tot_used(0), |
tonyp@1371 | 1199 | _region_bm(region_bm), _card_bm(card_bm),_final(final), |
ysr@777 | 1200 | _regions_done(0), _start_vtime_sec(0.0) |
ysr@777 | 1201 | { |
ysr@777 | 1202 | _bottom_card_num = |
ysr@777 | 1203 | intptr_t(uintptr_t(G1CollectedHeap::heap()->reserved_region().start()) >> |
ysr@777 | 1204 | CardTableModRefBS::card_shift); |
ysr@777 | 1205 | } |
ysr@777 | 1206 | |
tonyp@1264 | 1207 | // It takes a region that's not empty (i.e., it has at least one |
tonyp@1264 | 1208 | // live object in it and sets its corresponding bit on the region |
tonyp@1264 | 1209 | // bitmap to 1. If the region is "starts humongous" it will also set |
tonyp@1264 | 1210 | // to 1 the bits on the region bitmap that correspond to its |
tonyp@1264 | 1211 | // associated "continues humongous" regions. |
tonyp@1264 | 1212 | void set_bit_for_region(HeapRegion* hr) { |
tonyp@1264 | 1213 | assert(!hr->continuesHumongous(), "should have filtered those out"); |
tonyp@1264 | 1214 | |
tonyp@1264 | 1215 | size_t index = hr->hrs_index(); |
tonyp@1264 | 1216 | if (!hr->startsHumongous()) { |
tonyp@1264 | 1217 | // Normal (non-humongous) case: just set the bit. |
tonyp@1264 | 1218 | _region_bm->par_at_put((BitMap::idx_t) index, true); |
tonyp@1264 | 1219 | } else { |
tonyp@1264 | 1220 | // Starts humongous case: calculate how many regions are part of |
tonyp@1264 | 1221 | // this humongous region and then set the bit range. It might |
tonyp@1264 | 1222 | // have been a bit more efficient to look at the object that |
tonyp@1264 | 1223 | // spans these humongous regions to calculate their number from |
tonyp@1264 | 1224 | // the object's size. However, it's a good idea to calculate |
tonyp@1264 | 1225 | // this based on the metadata itself, and not the region |
tonyp@1264 | 1226 | // contents, so that this code is not aware of what goes into |
tonyp@1264 | 1227 | // the humongous regions (in case this changes in the future). |
tonyp@1264 | 1228 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
tonyp@1264 | 1229 | size_t end_index = index + 1; |
tonyp@1266 | 1230 | while (end_index < g1h->n_regions()) { |
tonyp@1266 | 1231 | HeapRegion* chr = g1h->region_at(end_index); |
tonyp@1264 | 1232 | if (!chr->continuesHumongous()) { |
tonyp@1264 | 1233 | break; |
tonyp@1264 | 1234 | } |
tonyp@1264 | 1235 | end_index += 1; |
tonyp@1264 | 1236 | } |
tonyp@1264 | 1237 | _region_bm->par_at_put_range((BitMap::idx_t) index, |
tonyp@1264 | 1238 | (BitMap::idx_t) end_index, true); |
tonyp@1264 | 1239 | } |
tonyp@1264 | 1240 | } |
tonyp@1264 | 1241 | |
ysr@777 | 1242 | bool doHeapRegion(HeapRegion* hr) { |
ysr@777 | 1243 | if (!_final && _regions_done == 0) |
ysr@777 | 1244 | _start_vtime_sec = os::elapsedVTime(); |
ysr@777 | 1245 | |
iveresov@1074 | 1246 | if (hr->continuesHumongous()) { |
tonyp@1264 | 1247 | // We will ignore these here and process them when their |
tonyp@1264 | 1248 | // associated "starts humongous" region is processed (see |
tonyp@1264 | 1249 | // set_bit_for_heap_region()). Note that we cannot rely on their |
tonyp@1264 | 1250 | // associated "starts humongous" region to have their bit set to |
tonyp@1264 | 1251 | // 1 since, due to the region chunking in the parallel region |
tonyp@1264 | 1252 | // iteration, a "continues humongous" region might be visited |
tonyp@1264 | 1253 | // before its associated "starts humongous". |
iveresov@1074 | 1254 | return false; |
iveresov@1074 | 1255 | } |
ysr@777 | 1256 | |
ysr@777 | 1257 | HeapWord* nextTop = hr->next_top_at_mark_start(); |
ysr@777 | 1258 | HeapWord* start = hr->top_at_conc_mark_count(); |
ysr@777 | 1259 | assert(hr->bottom() <= start && start <= hr->end() && |
ysr@777 | 1260 | hr->bottom() <= nextTop && nextTop <= hr->end() && |
ysr@777 | 1261 | start <= nextTop, |
ysr@777 | 1262 | "Preconditions."); |
ysr@777 | 1263 | // Otherwise, record the number of word's we'll examine. |
ysr@777 | 1264 | size_t words_done = (nextTop - start); |
ysr@777 | 1265 | // Find the first marked object at or after "start". |
ysr@777 | 1266 | start = _bm->getNextMarkedWordAddress(start, nextTop); |
ysr@777 | 1267 | size_t marked_bytes = 0; |
ysr@777 | 1268 | |
ysr@777 | 1269 | // Below, the term "card num" means the result of shifting an address |
ysr@777 | 1270 | // by the card shift -- address 0 corresponds to card number 0. One |
ysr@777 | 1271 | // must subtract the card num of the bottom of the heap to obtain a |
ysr@777 | 1272 | // card table index. |
ysr@777 | 1273 | // The first card num of the sequence of live cards currently being |
ysr@777 | 1274 | // constructed. -1 ==> no sequence. |
ysr@777 | 1275 | intptr_t start_card_num = -1; |
ysr@777 | 1276 | // The last card num of the sequence of live cards currently being |
ysr@777 | 1277 | // constructed. -1 ==> no sequence. |
ysr@777 | 1278 | intptr_t last_card_num = -1; |
ysr@777 | 1279 | |
ysr@777 | 1280 | while (start < nextTop) { |
ysr@777 | 1281 | if (_yield && _cm->do_yield_check()) { |
ysr@777 | 1282 | // We yielded. It might be for a full collection, in which case |
ysr@777 | 1283 | // all bets are off; terminate the traversal. |
ysr@777 | 1284 | if (_cm->has_aborted()) { |
ysr@777 | 1285 | _changed = false; |
ysr@777 | 1286 | return true; |
ysr@777 | 1287 | } else { |
ysr@777 | 1288 | // Otherwise, it might be a collection pause, and the region |
ysr@777 | 1289 | // we're looking at might be in the collection set. We'll |
ysr@777 | 1290 | // abandon this region. |
ysr@777 | 1291 | return false; |
ysr@777 | 1292 | } |
ysr@777 | 1293 | } |
ysr@777 | 1294 | oop obj = oop(start); |
ysr@777 | 1295 | int obj_sz = obj->size(); |
ysr@777 | 1296 | // The card num of the start of the current object. |
ysr@777 | 1297 | intptr_t obj_card_num = |
ysr@777 | 1298 | intptr_t(uintptr_t(start) >> CardTableModRefBS::card_shift); |
ysr@777 | 1299 | |
ysr@777 | 1300 | HeapWord* obj_last = start + obj_sz - 1; |
ysr@777 | 1301 | intptr_t obj_last_card_num = |
ysr@777 | 1302 | intptr_t(uintptr_t(obj_last) >> CardTableModRefBS::card_shift); |
ysr@777 | 1303 | |
ysr@777 | 1304 | if (obj_card_num != last_card_num) { |
ysr@777 | 1305 | if (start_card_num == -1) { |
ysr@777 | 1306 | assert(last_card_num == -1, "Both or neither."); |
ysr@777 | 1307 | start_card_num = obj_card_num; |
ysr@777 | 1308 | } else { |
ysr@777 | 1309 | assert(last_card_num != -1, "Both or neither."); |
ysr@777 | 1310 | assert(obj_card_num >= last_card_num, "Inv"); |
ysr@777 | 1311 | if ((obj_card_num - last_card_num) > 1) { |
ysr@777 | 1312 | // Mark the last run, and start a new one. |
ysr@777 | 1313 | mark_card_num_range(start_card_num, last_card_num); |
ysr@777 | 1314 | start_card_num = obj_card_num; |
ysr@777 | 1315 | } |
ysr@777 | 1316 | } |
ysr@777 | 1317 | #if CARD_BM_TEST_MODE |
ysr@777 | 1318 | /* |
ysr@777 | 1319 | gclog_or_tty->print_cr("Setting bits from %d/%d.", |
ysr@777 | 1320 | obj_card_num - _bottom_card_num, |
ysr@777 | 1321 | obj_last_card_num - _bottom_card_num); |
ysr@777 | 1322 | */ |
ysr@777 | 1323 | for (intptr_t j = obj_card_num; j <= obj_last_card_num; j++) { |
ysr@777 | 1324 | _card_bm->par_at_put(j - _bottom_card_num, 1); |
ysr@777 | 1325 | } |
ysr@777 | 1326 | #endif |
ysr@777 | 1327 | } |
ysr@777 | 1328 | // In any case, we set the last card num. |
ysr@777 | 1329 | last_card_num = obj_last_card_num; |
ysr@777 | 1330 | |
ysr@777 | 1331 | marked_bytes += obj_sz * HeapWordSize; |
ysr@777 | 1332 | // Find the next marked object after this one. |
ysr@777 | 1333 | start = _bm->getNextMarkedWordAddress(start + 1, nextTop); |
ysr@777 | 1334 | _changed = true; |
ysr@777 | 1335 | } |
ysr@777 | 1336 | // Handle the last range, if any. |
ysr@777 | 1337 | if (start_card_num != -1) |
ysr@777 | 1338 | mark_card_num_range(start_card_num, last_card_num); |
ysr@777 | 1339 | if (_final) { |
ysr@777 | 1340 | // Mark the allocated-since-marking portion... |
ysr@777 | 1341 | HeapWord* tp = hr->top(); |
ysr@777 | 1342 | if (nextTop < tp) { |
ysr@777 | 1343 | start_card_num = |
ysr@777 | 1344 | intptr_t(uintptr_t(nextTop) >> CardTableModRefBS::card_shift); |
ysr@777 | 1345 | last_card_num = |
ysr@777 | 1346 | intptr_t(uintptr_t(tp) >> CardTableModRefBS::card_shift); |
ysr@777 | 1347 | mark_card_num_range(start_card_num, last_card_num); |
ysr@777 | 1348 | // This definitely means the region has live objects. |
tonyp@1264 | 1349 | set_bit_for_region(hr); |
ysr@777 | 1350 | } |
ysr@777 | 1351 | } |
ysr@777 | 1352 | |
ysr@777 | 1353 | hr->add_to_marked_bytes(marked_bytes); |
ysr@777 | 1354 | // Update the live region bitmap. |
ysr@777 | 1355 | if (marked_bytes > 0) { |
tonyp@1264 | 1356 | set_bit_for_region(hr); |
ysr@777 | 1357 | } |
ysr@777 | 1358 | hr->set_top_at_conc_mark_count(nextTop); |
ysr@777 | 1359 | _tot_live += hr->next_live_bytes(); |
ysr@777 | 1360 | _tot_used += hr->used(); |
ysr@777 | 1361 | _words_done = words_done; |
ysr@777 | 1362 | |
ysr@777 | 1363 | if (!_final) { |
ysr@777 | 1364 | ++_regions_done; |
ysr@777 | 1365 | if (_regions_done % 10 == 0) { |
ysr@777 | 1366 | double end_vtime_sec = os::elapsedVTime(); |
ysr@777 | 1367 | double elapsed_vtime_sec = end_vtime_sec - _start_vtime_sec; |
ysr@777 | 1368 | if (elapsed_vtime_sec > (10.0 / 1000.0)) { |
ysr@777 | 1369 | jlong sleep_time_ms = |
ysr@777 | 1370 | (jlong) (elapsed_vtime_sec * _cm->cleanup_sleep_factor() * 1000.0); |
ysr@777 | 1371 | os::sleep(Thread::current(), sleep_time_ms, false); |
ysr@777 | 1372 | _start_vtime_sec = end_vtime_sec; |
ysr@777 | 1373 | } |
ysr@777 | 1374 | } |
ysr@777 | 1375 | } |
ysr@777 | 1376 | |
ysr@777 | 1377 | return false; |
ysr@777 | 1378 | } |
ysr@777 | 1379 | |
ysr@777 | 1380 | bool changed() { return _changed; } |
ysr@777 | 1381 | void reset() { _changed = false; _words_done = 0; } |
ysr@777 | 1382 | void no_yield() { _yield = false; } |
ysr@777 | 1383 | size_t words_done() { return _words_done; } |
ysr@777 | 1384 | size_t tot_live() { return _tot_live; } |
ysr@777 | 1385 | size_t tot_used() { return _tot_used; } |
ysr@777 | 1386 | }; |
ysr@777 | 1387 | |
ysr@777 | 1388 | |
ysr@777 | 1389 | void ConcurrentMark::calcDesiredRegions() { |
ysr@777 | 1390 | _region_bm.clear(); |
ysr@777 | 1391 | _card_bm.clear(); |
ysr@777 | 1392 | CalcLiveObjectsClosure calccl(false /*final*/, |
ysr@777 | 1393 | nextMarkBitMap(), this, |
tonyp@1371 | 1394 | &_region_bm, &_card_bm); |
ysr@777 | 1395 | G1CollectedHeap *g1h = G1CollectedHeap::heap(); |
ysr@777 | 1396 | g1h->heap_region_iterate(&calccl); |
ysr@777 | 1397 | |
ysr@777 | 1398 | do { |
ysr@777 | 1399 | calccl.reset(); |
ysr@777 | 1400 | g1h->heap_region_iterate(&calccl); |
ysr@777 | 1401 | } while (calccl.changed()); |
ysr@777 | 1402 | } |
ysr@777 | 1403 | |
ysr@777 | 1404 | class G1ParFinalCountTask: public AbstractGangTask { |
ysr@777 | 1405 | protected: |
ysr@777 | 1406 | G1CollectedHeap* _g1h; |
ysr@777 | 1407 | CMBitMap* _bm; |
ysr@777 | 1408 | size_t _n_workers; |
ysr@777 | 1409 | size_t *_live_bytes; |
ysr@777 | 1410 | size_t *_used_bytes; |
ysr@777 | 1411 | BitMap* _region_bm; |
ysr@777 | 1412 | BitMap* _card_bm; |
ysr@777 | 1413 | public: |
ysr@777 | 1414 | G1ParFinalCountTask(G1CollectedHeap* g1h, CMBitMap* bm, |
ysr@777 | 1415 | BitMap* region_bm, BitMap* card_bm) : |
ysr@777 | 1416 | AbstractGangTask("G1 final counting"), _g1h(g1h), |
ysr@777 | 1417 | _bm(bm), _region_bm(region_bm), _card_bm(card_bm) |
ysr@777 | 1418 | { |
ysr@777 | 1419 | if (ParallelGCThreads > 0) |
ysr@777 | 1420 | _n_workers = _g1h->workers()->total_workers(); |
ysr@777 | 1421 | else |
ysr@777 | 1422 | _n_workers = 1; |
ysr@777 | 1423 | _live_bytes = NEW_C_HEAP_ARRAY(size_t, _n_workers); |
ysr@777 | 1424 | _used_bytes = NEW_C_HEAP_ARRAY(size_t, _n_workers); |
ysr@777 | 1425 | } |
ysr@777 | 1426 | |
ysr@777 | 1427 | ~G1ParFinalCountTask() { |
ysr@777 | 1428 | FREE_C_HEAP_ARRAY(size_t, _live_bytes); |
ysr@777 | 1429 | FREE_C_HEAP_ARRAY(size_t, _used_bytes); |
ysr@777 | 1430 | } |
ysr@777 | 1431 | |
ysr@777 | 1432 | void work(int i) { |
ysr@777 | 1433 | CalcLiveObjectsClosure calccl(true /*final*/, |
ysr@777 | 1434 | _bm, _g1h->concurrent_mark(), |
tonyp@1371 | 1435 | _region_bm, _card_bm); |
ysr@777 | 1436 | calccl.no_yield(); |
ysr@777 | 1437 | if (ParallelGCThreads > 0) { |
tonyp@790 | 1438 | _g1h->heap_region_par_iterate_chunked(&calccl, i, |
tonyp@790 | 1439 | HeapRegion::FinalCountClaimValue); |
ysr@777 | 1440 | } else { |
ysr@777 | 1441 | _g1h->heap_region_iterate(&calccl); |
ysr@777 | 1442 | } |
ysr@777 | 1443 | assert(calccl.complete(), "Shouldn't have yielded!"); |
ysr@777 | 1444 | |
ysr@777 | 1445 | guarantee( (size_t)i < _n_workers, "invariant" ); |
ysr@777 | 1446 | _live_bytes[i] = calccl.tot_live(); |
ysr@777 | 1447 | _used_bytes[i] = calccl.tot_used(); |
ysr@777 | 1448 | } |
ysr@777 | 1449 | size_t live_bytes() { |
ysr@777 | 1450 | size_t live_bytes = 0; |
ysr@777 | 1451 | for (size_t i = 0; i < _n_workers; ++i) |
ysr@777 | 1452 | live_bytes += _live_bytes[i]; |
ysr@777 | 1453 | return live_bytes; |
ysr@777 | 1454 | } |
ysr@777 | 1455 | size_t used_bytes() { |
ysr@777 | 1456 | size_t used_bytes = 0; |
ysr@777 | 1457 | for (size_t i = 0; i < _n_workers; ++i) |
ysr@777 | 1458 | used_bytes += _used_bytes[i]; |
ysr@777 | 1459 | return used_bytes; |
ysr@777 | 1460 | } |
ysr@777 | 1461 | }; |
ysr@777 | 1462 | |
ysr@777 | 1463 | class G1ParNoteEndTask; |
ysr@777 | 1464 | |
ysr@777 | 1465 | class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { |
ysr@777 | 1466 | G1CollectedHeap* _g1; |
ysr@777 | 1467 | int _worker_num; |
ysr@777 | 1468 | size_t _max_live_bytes; |
ysr@777 | 1469 | size_t _regions_claimed; |
ysr@777 | 1470 | size_t _freed_bytes; |
ysr@777 | 1471 | size_t _cleared_h_regions; |
ysr@777 | 1472 | size_t _freed_regions; |
ysr@777 | 1473 | UncleanRegionList* _unclean_region_list; |
ysr@777 | 1474 | double _claimed_region_time; |
ysr@777 | 1475 | double _max_region_time; |
ysr@777 | 1476 | |
ysr@777 | 1477 | public: |
ysr@777 | 1478 | G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, |
ysr@777 | 1479 | UncleanRegionList* list, |
ysr@777 | 1480 | int worker_num); |
ysr@777 | 1481 | size_t freed_bytes() { return _freed_bytes; } |
ysr@777 | 1482 | size_t cleared_h_regions() { return _cleared_h_regions; } |
ysr@777 | 1483 | size_t freed_regions() { return _freed_regions; } |
ysr@777 | 1484 | UncleanRegionList* unclean_region_list() { |
ysr@777 | 1485 | return _unclean_region_list; |
ysr@777 | 1486 | } |
ysr@777 | 1487 | |
ysr@777 | 1488 | bool doHeapRegion(HeapRegion *r); |
ysr@777 | 1489 | |
ysr@777 | 1490 | size_t max_live_bytes() { return _max_live_bytes; } |
ysr@777 | 1491 | size_t regions_claimed() { return _regions_claimed; } |
ysr@777 | 1492 | double claimed_region_time_sec() { return _claimed_region_time; } |
ysr@777 | 1493 | double max_region_time_sec() { return _max_region_time; } |
ysr@777 | 1494 | }; |
ysr@777 | 1495 | |
ysr@777 | 1496 | class G1ParNoteEndTask: public AbstractGangTask { |
ysr@777 | 1497 | friend class G1NoteEndOfConcMarkClosure; |
ysr@777 | 1498 | protected: |
ysr@777 | 1499 | G1CollectedHeap* _g1h; |
ysr@777 | 1500 | size_t _max_live_bytes; |
ysr@777 | 1501 | size_t _freed_bytes; |
ysr@777 | 1502 | ConcurrentMark::ParCleanupThreadState** _par_cleanup_thread_state; |
ysr@777 | 1503 | public: |
ysr@777 | 1504 | G1ParNoteEndTask(G1CollectedHeap* g1h, |
ysr@777 | 1505 | ConcurrentMark::ParCleanupThreadState** |
ysr@777 | 1506 | par_cleanup_thread_state) : |
ysr@777 | 1507 | AbstractGangTask("G1 note end"), _g1h(g1h), |
ysr@777 | 1508 | _max_live_bytes(0), _freed_bytes(0), |
ysr@777 | 1509 | _par_cleanup_thread_state(par_cleanup_thread_state) |
ysr@777 | 1510 | {} |
ysr@777 | 1511 | |
ysr@777 | 1512 | void work(int i) { |
ysr@777 | 1513 | double start = os::elapsedTime(); |
ysr@777 | 1514 | G1NoteEndOfConcMarkClosure g1_note_end(_g1h, |
ysr@777 | 1515 | &_par_cleanup_thread_state[i]->list, |
ysr@777 | 1516 | i); |
ysr@777 | 1517 | if (ParallelGCThreads > 0) { |
tonyp@790 | 1518 | _g1h->heap_region_par_iterate_chunked(&g1_note_end, i, |
tonyp@790 | 1519 | HeapRegion::NoteEndClaimValue); |
ysr@777 | 1520 | } else { |
ysr@777 | 1521 | _g1h->heap_region_iterate(&g1_note_end); |
ysr@777 | 1522 | } |
ysr@777 | 1523 | assert(g1_note_end.complete(), "Shouldn't have yielded!"); |
ysr@777 | 1524 | |
ysr@777 | 1525 | // Now finish up freeing the current thread's regions. |
ysr@777 | 1526 | _g1h->finish_free_region_work(g1_note_end.freed_bytes(), |
ysr@777 | 1527 | g1_note_end.cleared_h_regions(), |
ysr@777 | 1528 | 0, NULL); |
ysr@777 | 1529 | { |
ysr@777 | 1530 | MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); |
ysr@777 | 1531 | _max_live_bytes += g1_note_end.max_live_bytes(); |
ysr@777 | 1532 | _freed_bytes += g1_note_end.freed_bytes(); |
ysr@777 | 1533 | } |
ysr@777 | 1534 | double end = os::elapsedTime(); |
ysr@777 | 1535 | if (G1PrintParCleanupStats) { |
ysr@777 | 1536 | gclog_or_tty->print(" Worker thread %d [%8.3f..%8.3f = %8.3f ms] " |
ysr@777 | 1537 | "claimed %d regions (tot = %8.3f ms, max = %8.3f ms).\n", |
ysr@777 | 1538 | i, start, end, (end-start)*1000.0, |
ysr@777 | 1539 | g1_note_end.regions_claimed(), |
ysr@777 | 1540 | g1_note_end.claimed_region_time_sec()*1000.0, |
ysr@777 | 1541 | g1_note_end.max_region_time_sec()*1000.0); |
ysr@777 | 1542 | } |
ysr@777 | 1543 | } |
ysr@777 | 1544 | size_t max_live_bytes() { return _max_live_bytes; } |
ysr@777 | 1545 | size_t freed_bytes() { return _freed_bytes; } |
ysr@777 | 1546 | }; |
ysr@777 | 1547 | |
ysr@777 | 1548 | class G1ParScrubRemSetTask: public AbstractGangTask { |
ysr@777 | 1549 | protected: |
ysr@777 | 1550 | G1RemSet* _g1rs; |
ysr@777 | 1551 | BitMap* _region_bm; |
ysr@777 | 1552 | BitMap* _card_bm; |
ysr@777 | 1553 | public: |
ysr@777 | 1554 | G1ParScrubRemSetTask(G1CollectedHeap* g1h, |
ysr@777 | 1555 | BitMap* region_bm, BitMap* card_bm) : |
ysr@777 | 1556 | AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), |
ysr@777 | 1557 | _region_bm(region_bm), _card_bm(card_bm) |
ysr@777 | 1558 | {} |
ysr@777 | 1559 | |
ysr@777 | 1560 | void work(int i) { |
ysr@777 | 1561 | if (ParallelGCThreads > 0) { |
tonyp@790 | 1562 | _g1rs->scrub_par(_region_bm, _card_bm, i, |
tonyp@790 | 1563 | HeapRegion::ScrubRemSetClaimValue); |
ysr@777 | 1564 | } else { |
ysr@777 | 1565 | _g1rs->scrub(_region_bm, _card_bm); |
ysr@777 | 1566 | } |
ysr@777 | 1567 | } |
ysr@777 | 1568 | |
ysr@777 | 1569 | }; |
ysr@777 | 1570 | |
ysr@777 | 1571 | G1NoteEndOfConcMarkClosure:: |
ysr@777 | 1572 | G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, |
ysr@777 | 1573 | UncleanRegionList* list, |
ysr@777 | 1574 | int worker_num) |
ysr@777 | 1575 | : _g1(g1), _worker_num(worker_num), |
ysr@777 | 1576 | _max_live_bytes(0), _regions_claimed(0), |
ysr@777 | 1577 | _freed_bytes(0), _cleared_h_regions(0), _freed_regions(0), |
ysr@777 | 1578 | _claimed_region_time(0.0), _max_region_time(0.0), |
ysr@777 | 1579 | _unclean_region_list(list) |
ysr@777 | 1580 | {} |
ysr@777 | 1581 | |
ysr@777 | 1582 | bool G1NoteEndOfConcMarkClosure::doHeapRegion(HeapRegion *r) { |
ysr@777 | 1583 | // We use a claim value of zero here because all regions |
ysr@777 | 1584 | // were claimed with value 1 in the FinalCount task. |
ysr@777 | 1585 | r->reset_gc_time_stamp(); |
ysr@777 | 1586 | if (!r->continuesHumongous()) { |
ysr@777 | 1587 | double start = os::elapsedTime(); |
ysr@777 | 1588 | _regions_claimed++; |
ysr@777 | 1589 | r->note_end_of_marking(); |
ysr@777 | 1590 | _max_live_bytes += r->max_live_bytes(); |
ysr@777 | 1591 | _g1->free_region_if_totally_empty_work(r, |
ysr@777 | 1592 | _freed_bytes, |
ysr@777 | 1593 | _cleared_h_regions, |
ysr@777 | 1594 | _freed_regions, |
ysr@777 | 1595 | _unclean_region_list, |
ysr@777 | 1596 | true /*par*/); |
ysr@777 | 1597 | double region_time = (os::elapsedTime() - start); |
ysr@777 | 1598 | _claimed_region_time += region_time; |
ysr@777 | 1599 | if (region_time > _max_region_time) _max_region_time = region_time; |
ysr@777 | 1600 | } |
ysr@777 | 1601 | return false; |
ysr@777 | 1602 | } |
ysr@777 | 1603 | |
ysr@777 | 1604 | void ConcurrentMark::cleanup() { |
ysr@777 | 1605 | // world is stopped at this checkpoint |
ysr@777 | 1606 | assert(SafepointSynchronize::is_at_safepoint(), |
ysr@777 | 1607 | "world should be stopped"); |
ysr@777 | 1608 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 1609 | |
ysr@777 | 1610 | // If a full collection has happened, we shouldn't do this. |
ysr@777 | 1611 | if (has_aborted()) { |
ysr@777 | 1612 | g1h->set_marking_complete(); // So bitmap clearing isn't confused |
ysr@777 | 1613 | return; |
ysr@777 | 1614 | } |
ysr@777 | 1615 | |
ysr@1280 | 1616 | if (VerifyDuringGC) { |
ysr@1280 | 1617 | HandleMark hm; // handle scope |
ysr@1280 | 1618 | gclog_or_tty->print(" VerifyDuringGC:(before)"); |
ysr@1280 | 1619 | Universe::heap()->prepare_for_verify(); |
ysr@1280 | 1620 | Universe::verify(/* allow dirty */ true, |
ysr@1280 | 1621 | /* silent */ false, |
ysr@1280 | 1622 | /* prev marking */ true); |
ysr@1280 | 1623 | } |
ysr@1280 | 1624 | |
ysr@777 | 1625 | G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); |
ysr@777 | 1626 | g1p->record_concurrent_mark_cleanup_start(); |
ysr@777 | 1627 | |
ysr@777 | 1628 | double start = os::elapsedTime(); |
ysr@777 | 1629 | |
ysr@777 | 1630 | // Do counting once more with the world stopped for good measure. |
ysr@777 | 1631 | G1ParFinalCountTask g1_par_count_task(g1h, nextMarkBitMap(), |
ysr@777 | 1632 | &_region_bm, &_card_bm); |
ysr@777 | 1633 | if (ParallelGCThreads > 0) { |
tonyp@790 | 1634 | assert(g1h->check_heap_region_claim_values( |
tonyp@790 | 1635 | HeapRegion::InitialClaimValue), |
tonyp@790 | 1636 | "sanity check"); |
tonyp@790 | 1637 | |
ysr@777 | 1638 | int n_workers = g1h->workers()->total_workers(); |
ysr@777 | 1639 | g1h->set_par_threads(n_workers); |
ysr@777 | 1640 | g1h->workers()->run_task(&g1_par_count_task); |
ysr@777 | 1641 | g1h->set_par_threads(0); |
tonyp@790 | 1642 | |
tonyp@790 | 1643 | assert(g1h->check_heap_region_claim_values( |
tonyp@790 | 1644 | HeapRegion::FinalCountClaimValue), |
tonyp@790 | 1645 | "sanity check"); |
ysr@777 | 1646 | } else { |
ysr@777 | 1647 | g1_par_count_task.work(0); |
ysr@777 | 1648 | } |
ysr@777 | 1649 | |
ysr@777 | 1650 | size_t known_garbage_bytes = |
ysr@777 | 1651 | g1_par_count_task.used_bytes() - g1_par_count_task.live_bytes(); |
ysr@777 | 1652 | #if 0 |
ysr@777 | 1653 | gclog_or_tty->print_cr("used %1.2lf, live %1.2lf, garbage %1.2lf", |
ysr@777 | 1654 | (double) g1_par_count_task.used_bytes() / (double) (1024 * 1024), |
ysr@777 | 1655 | (double) g1_par_count_task.live_bytes() / (double) (1024 * 1024), |
ysr@777 | 1656 | (double) known_garbage_bytes / (double) (1024 * 1024)); |
ysr@777 | 1657 | #endif // 0 |
ysr@777 | 1658 | g1p->set_known_garbage_bytes(known_garbage_bytes); |
ysr@777 | 1659 | |
ysr@777 | 1660 | size_t start_used_bytes = g1h->used(); |
ysr@777 | 1661 | _at_least_one_mark_complete = true; |
ysr@777 | 1662 | g1h->set_marking_complete(); |
ysr@777 | 1663 | |
ysr@777 | 1664 | double count_end = os::elapsedTime(); |
ysr@777 | 1665 | double this_final_counting_time = (count_end - start); |
ysr@777 | 1666 | if (G1PrintParCleanupStats) { |
ysr@777 | 1667 | gclog_or_tty->print_cr("Cleanup:"); |
ysr@777 | 1668 | gclog_or_tty->print_cr(" Finalize counting: %8.3f ms", |
ysr@777 | 1669 | this_final_counting_time*1000.0); |
ysr@777 | 1670 | } |
ysr@777 | 1671 | _total_counting_time += this_final_counting_time; |
ysr@777 | 1672 | |
ysr@777 | 1673 | // Install newly created mark bitMap as "prev". |
ysr@777 | 1674 | swapMarkBitMaps(); |
ysr@777 | 1675 | |
ysr@777 | 1676 | g1h->reset_gc_time_stamp(); |
ysr@777 | 1677 | |
ysr@777 | 1678 | // Note end of marking in all heap regions. |
ysr@777 | 1679 | double note_end_start = os::elapsedTime(); |
ysr@777 | 1680 | G1ParNoteEndTask g1_par_note_end_task(g1h, _par_cleanup_thread_state); |
ysr@777 | 1681 | if (ParallelGCThreads > 0) { |
ysr@777 | 1682 | int n_workers = g1h->workers()->total_workers(); |
ysr@777 | 1683 | g1h->set_par_threads(n_workers); |
ysr@777 | 1684 | g1h->workers()->run_task(&g1_par_note_end_task); |
ysr@777 | 1685 | g1h->set_par_threads(0); |
tonyp@790 | 1686 | |
tonyp@790 | 1687 | assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue), |
tonyp@790 | 1688 | "sanity check"); |
ysr@777 | 1689 | } else { |
ysr@777 | 1690 | g1_par_note_end_task.work(0); |
ysr@777 | 1691 | } |
ysr@777 | 1692 | g1h->set_unclean_regions_coming(true); |
ysr@777 | 1693 | double note_end_end = os::elapsedTime(); |
ysr@777 | 1694 | // Tell the mutators that there might be unclean regions coming... |
ysr@777 | 1695 | if (G1PrintParCleanupStats) { |
ysr@777 | 1696 | gclog_or_tty->print_cr(" note end of marking: %8.3f ms.", |
ysr@777 | 1697 | (note_end_end - note_end_start)*1000.0); |
ysr@777 | 1698 | } |
ysr@777 | 1699 | |
tonyp@790 | 1700 | |
ysr@777 | 1701 | // call below, since it affects the metric by which we sort the heap |
ysr@777 | 1702 | // regions. |
ysr@777 | 1703 | if (G1ScrubRemSets) { |
ysr@777 | 1704 | double rs_scrub_start = os::elapsedTime(); |
ysr@777 | 1705 | G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm); |
ysr@777 | 1706 | if (ParallelGCThreads > 0) { |
ysr@777 | 1707 | int n_workers = g1h->workers()->total_workers(); |
ysr@777 | 1708 | g1h->set_par_threads(n_workers); |
ysr@777 | 1709 | g1h->workers()->run_task(&g1_par_scrub_rs_task); |
ysr@777 | 1710 | g1h->set_par_threads(0); |
tonyp@790 | 1711 | |
tonyp@790 | 1712 | assert(g1h->check_heap_region_claim_values( |
tonyp@790 | 1713 | HeapRegion::ScrubRemSetClaimValue), |
tonyp@790 | 1714 | "sanity check"); |
ysr@777 | 1715 | } else { |
ysr@777 | 1716 | g1_par_scrub_rs_task.work(0); |
ysr@777 | 1717 | } |
ysr@777 | 1718 | |
ysr@777 | 1719 | double rs_scrub_end = os::elapsedTime(); |
ysr@777 | 1720 | double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); |
ysr@777 | 1721 | _total_rs_scrub_time += this_rs_scrub_time; |
ysr@777 | 1722 | } |
ysr@777 | 1723 | |
ysr@777 | 1724 | // this will also free any regions totally full of garbage objects, |
ysr@777 | 1725 | // and sort the regions. |
ysr@777 | 1726 | g1h->g1_policy()->record_concurrent_mark_cleanup_end( |
ysr@777 | 1727 | g1_par_note_end_task.freed_bytes(), |
ysr@777 | 1728 | g1_par_note_end_task.max_live_bytes()); |
ysr@777 | 1729 | |
ysr@777 | 1730 | // Statistics. |
ysr@777 | 1731 | double end = os::elapsedTime(); |
ysr@777 | 1732 | _cleanup_times.add((end - start) * 1000.0); |
ysr@777 | 1733 | |
ysr@777 | 1734 | // G1CollectedHeap::heap()->print(); |
ysr@777 | 1735 | // gclog_or_tty->print_cr("HEAP GC TIME STAMP : %d", |
ysr@777 | 1736 | // G1CollectedHeap::heap()->get_gc_time_stamp()); |
ysr@777 | 1737 | |
ysr@777 | 1738 | if (PrintGC || PrintGCDetails) { |
ysr@777 | 1739 | g1h->print_size_transition(gclog_or_tty, |
ysr@777 | 1740 | start_used_bytes, |
ysr@777 | 1741 | g1h->used(), |
ysr@777 | 1742 | g1h->capacity()); |
ysr@777 | 1743 | } |
ysr@777 | 1744 | |
ysr@777 | 1745 | size_t cleaned_up_bytes = start_used_bytes - g1h->used(); |
ysr@777 | 1746 | g1p->decrease_known_garbage_bytes(cleaned_up_bytes); |
ysr@777 | 1747 | |
ysr@777 | 1748 | // We need to make this be a "collection" so any collection pause that |
ysr@777 | 1749 | // races with it goes around and waits for completeCleanup to finish. |
ysr@777 | 1750 | g1h->increment_total_collections(); |
ysr@777 | 1751 | |
johnc@1186 | 1752 | if (VerifyDuringGC) { |
ysr@1280 | 1753 | HandleMark hm; // handle scope |
ysr@1280 | 1754 | gclog_or_tty->print(" VerifyDuringGC:(after)"); |
ysr@1280 | 1755 | Universe::heap()->prepare_for_verify(); |
ysr@1280 | 1756 | Universe::verify(/* allow dirty */ true, |
ysr@1280 | 1757 | /* silent */ false, |
ysr@1280 | 1758 | /* prev marking */ true); |
ysr@777 | 1759 | } |
ysr@777 | 1760 | } |
ysr@777 | 1761 | |
ysr@777 | 1762 | void ConcurrentMark::completeCleanup() { |
ysr@777 | 1763 | // A full collection intervened. |
ysr@777 | 1764 | if (has_aborted()) return; |
ysr@777 | 1765 | |
ysr@777 | 1766 | int first = 0; |
ysr@777 | 1767 | int last = (int)MAX2(ParallelGCThreads, (size_t)1); |
ysr@777 | 1768 | for (int t = 0; t < last; t++) { |
ysr@777 | 1769 | UncleanRegionList* list = &_par_cleanup_thread_state[t]->list; |
ysr@777 | 1770 | assert(list->well_formed(), "Inv"); |
ysr@777 | 1771 | HeapRegion* hd = list->hd(); |
ysr@777 | 1772 | while (hd != NULL) { |
ysr@777 | 1773 | // Now finish up the other stuff. |
ysr@777 | 1774 | hd->rem_set()->clear(); |
ysr@777 | 1775 | HeapRegion* next_hd = hd->next_from_unclean_list(); |
ysr@777 | 1776 | (void)list->pop(); |
ysr@777 | 1777 | guarantee(list->hd() == next_hd, "how not?"); |
ysr@777 | 1778 | _g1h->put_region_on_unclean_list(hd); |
ysr@777 | 1779 | if (!hd->isHumongous()) { |
ysr@777 | 1780 | // Add this to the _free_regions count by 1. |
ysr@777 | 1781 | _g1h->finish_free_region_work(0, 0, 1, NULL); |
ysr@777 | 1782 | } |
ysr@777 | 1783 | hd = list->hd(); |
ysr@777 | 1784 | guarantee(hd == next_hd, "how not?"); |
ysr@777 | 1785 | } |
ysr@777 | 1786 | } |
ysr@777 | 1787 | } |
ysr@777 | 1788 | |
ysr@777 | 1789 | |
ysr@777 | 1790 | class G1CMIsAliveClosure: public BoolObjectClosure { |
ysr@777 | 1791 | G1CollectedHeap* _g1; |
ysr@777 | 1792 | public: |
ysr@777 | 1793 | G1CMIsAliveClosure(G1CollectedHeap* g1) : |
ysr@777 | 1794 | _g1(g1) |
ysr@777 | 1795 | {} |
ysr@777 | 1796 | |
ysr@777 | 1797 | void do_object(oop obj) { |
ysr@777 | 1798 | assert(false, "not to be invoked"); |
ysr@777 | 1799 | } |
ysr@777 | 1800 | bool do_object_b(oop obj) { |
ysr@777 | 1801 | HeapWord* addr = (HeapWord*)obj; |
ysr@777 | 1802 | return addr != NULL && |
ysr@777 | 1803 | (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); |
ysr@777 | 1804 | } |
ysr@777 | 1805 | }; |
ysr@777 | 1806 | |
ysr@777 | 1807 | class G1CMKeepAliveClosure: public OopClosure { |
ysr@777 | 1808 | G1CollectedHeap* _g1; |
ysr@777 | 1809 | ConcurrentMark* _cm; |
ysr@777 | 1810 | CMBitMap* _bitMap; |
ysr@777 | 1811 | public: |
ysr@777 | 1812 | G1CMKeepAliveClosure(G1CollectedHeap* g1, ConcurrentMark* cm, |
ysr@777 | 1813 | CMBitMap* bitMap) : |
ysr@777 | 1814 | _g1(g1), _cm(cm), |
ysr@777 | 1815 | _bitMap(bitMap) {} |
ysr@777 | 1816 | |
ysr@1280 | 1817 | virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
ysr@1280 | 1818 | virtual void do_oop( oop* p) { do_oop_work(p); } |
ysr@1280 | 1819 | |
ysr@1280 | 1820 | template <class T> void do_oop_work(T* p) { |
ysr@1280 | 1821 | oop thisOop = oopDesc::load_decode_heap_oop(p); |
ysr@777 | 1822 | HeapWord* addr = (HeapWord*)thisOop; |
ysr@777 | 1823 | if (_g1->is_in_g1_reserved(addr) && _g1->is_obj_ill(thisOop)) { |
ysr@777 | 1824 | _bitMap->mark(addr); |
ysr@777 | 1825 | _cm->mark_stack_push(thisOop); |
ysr@777 | 1826 | } |
ysr@777 | 1827 | } |
ysr@777 | 1828 | }; |
ysr@777 | 1829 | |
ysr@777 | 1830 | class G1CMDrainMarkingStackClosure: public VoidClosure { |
ysr@777 | 1831 | CMMarkStack* _markStack; |
ysr@777 | 1832 | CMBitMap* _bitMap; |
ysr@777 | 1833 | G1CMKeepAliveClosure* _oopClosure; |
ysr@777 | 1834 | public: |
ysr@777 | 1835 | G1CMDrainMarkingStackClosure(CMBitMap* bitMap, CMMarkStack* markStack, |
ysr@777 | 1836 | G1CMKeepAliveClosure* oopClosure) : |
ysr@777 | 1837 | _bitMap(bitMap), |
ysr@777 | 1838 | _markStack(markStack), |
ysr@777 | 1839 | _oopClosure(oopClosure) |
ysr@777 | 1840 | {} |
ysr@777 | 1841 | |
ysr@777 | 1842 | void do_void() { |
ysr@777 | 1843 | _markStack->drain((OopClosure*)_oopClosure, _bitMap, false); |
ysr@777 | 1844 | } |
ysr@777 | 1845 | }; |
ysr@777 | 1846 | |
ysr@777 | 1847 | void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { |
ysr@777 | 1848 | ResourceMark rm; |
ysr@777 | 1849 | HandleMark hm; |
ysr@888 | 1850 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@888 | 1851 | ReferenceProcessor* rp = g1h->ref_processor(); |
ysr@777 | 1852 | |
ysr@777 | 1853 | // Process weak references. |
ysr@892 | 1854 | rp->setup_policy(clear_all_soft_refs); |
ysr@777 | 1855 | assert(_markStack.isEmpty(), "mark stack should be empty"); |
ysr@777 | 1856 | |
ysr@888 | 1857 | G1CMIsAliveClosure g1IsAliveClosure (g1h); |
ysr@888 | 1858 | G1CMKeepAliveClosure g1KeepAliveClosure(g1h, this, nextMarkBitMap()); |
ysr@777 | 1859 | G1CMDrainMarkingStackClosure |
ysr@777 | 1860 | g1DrainMarkingStackClosure(nextMarkBitMap(), &_markStack, |
ysr@777 | 1861 | &g1KeepAliveClosure); |
ysr@777 | 1862 | |
ysr@777 | 1863 | // XXXYYY Also: copy the parallel ref processing code from CMS. |
ysr@888 | 1864 | rp->process_discovered_references(&g1IsAliveClosure, |
ysr@777 | 1865 | &g1KeepAliveClosure, |
ysr@777 | 1866 | &g1DrainMarkingStackClosure, |
ysr@777 | 1867 | NULL); |
ysr@777 | 1868 | assert(_markStack.overflow() || _markStack.isEmpty(), |
ysr@777 | 1869 | "mark stack should be empty (unless it overflowed)"); |
ysr@777 | 1870 | if (_markStack.overflow()) { |
ysr@777 | 1871 | set_has_overflown(); |
ysr@777 | 1872 | } |
ysr@777 | 1873 | |
ysr@777 | 1874 | rp->enqueue_discovered_references(); |
ysr@777 | 1875 | rp->verify_no_references_recorded(); |
ysr@777 | 1876 | assert(!rp->discovery_enabled(), "should have been disabled"); |
ysr@777 | 1877 | |
ysr@777 | 1878 | // Now clean up stale oops in SymbolTable and StringTable |
ysr@777 | 1879 | SymbolTable::unlink(&g1IsAliveClosure); |
ysr@777 | 1880 | StringTable::unlink(&g1IsAliveClosure); |
ysr@777 | 1881 | } |
ysr@777 | 1882 | |
ysr@777 | 1883 | void ConcurrentMark::swapMarkBitMaps() { |
ysr@777 | 1884 | CMBitMapRO* temp = _prevMarkBitMap; |
ysr@777 | 1885 | _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; |
ysr@777 | 1886 | _nextMarkBitMap = (CMBitMap*) temp; |
ysr@777 | 1887 | } |
ysr@777 | 1888 | |
ysr@777 | 1889 | class CMRemarkTask: public AbstractGangTask { |
ysr@777 | 1890 | private: |
ysr@777 | 1891 | ConcurrentMark *_cm; |
ysr@777 | 1892 | |
ysr@777 | 1893 | public: |
ysr@777 | 1894 | void work(int worker_i) { |
ysr@777 | 1895 | // Since all available tasks are actually started, we should |
ysr@777 | 1896 | // only proceed if we're supposed to be actived. |
ysr@777 | 1897 | if ((size_t)worker_i < _cm->active_tasks()) { |
ysr@777 | 1898 | CMTask* task = _cm->task(worker_i); |
ysr@777 | 1899 | task->record_start_time(); |
ysr@777 | 1900 | do { |
ysr@777 | 1901 | task->do_marking_step(1000000000.0 /* something very large */); |
ysr@777 | 1902 | } while (task->has_aborted() && !_cm->has_overflown()); |
ysr@777 | 1903 | // If we overflow, then we do not want to restart. We instead |
ysr@777 | 1904 | // want to abort remark and do concurrent marking again. |
ysr@777 | 1905 | task->record_end_time(); |
ysr@777 | 1906 | } |
ysr@777 | 1907 | } |
ysr@777 | 1908 | |
ysr@777 | 1909 | CMRemarkTask(ConcurrentMark* cm) : |
ysr@777 | 1910 | AbstractGangTask("Par Remark"), _cm(cm) { } |
ysr@777 | 1911 | }; |
ysr@777 | 1912 | |
ysr@777 | 1913 | void ConcurrentMark::checkpointRootsFinalWork() { |
ysr@777 | 1914 | ResourceMark rm; |
ysr@777 | 1915 | HandleMark hm; |
ysr@777 | 1916 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 1917 | |
ysr@777 | 1918 | g1h->ensure_parsability(false); |
ysr@777 | 1919 | |
ysr@777 | 1920 | if (ParallelGCThreads > 0) { |
jrose@1424 | 1921 | G1CollectedHeap::StrongRootsScope srs(g1h); |
ysr@777 | 1922 | // this is remark, so we'll use up all available threads |
ysr@777 | 1923 | int active_workers = ParallelGCThreads; |
ysr@777 | 1924 | set_phase(active_workers, false); |
ysr@777 | 1925 | |
ysr@777 | 1926 | CMRemarkTask remarkTask(this); |
ysr@777 | 1927 | // We will start all available threads, even if we decide that the |
ysr@777 | 1928 | // active_workers will be fewer. The extra ones will just bail out |
ysr@777 | 1929 | // immediately. |
ysr@777 | 1930 | int n_workers = g1h->workers()->total_workers(); |
ysr@777 | 1931 | g1h->set_par_threads(n_workers); |
ysr@777 | 1932 | g1h->workers()->run_task(&remarkTask); |
ysr@777 | 1933 | g1h->set_par_threads(0); |
ysr@777 | 1934 | |
ysr@777 | 1935 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 1936 | guarantee( satb_mq_set.completed_buffers_num() == 0, "invariant" ); |
ysr@777 | 1937 | } else { |
jrose@1424 | 1938 | G1CollectedHeap::StrongRootsScope srs(g1h); |
ysr@777 | 1939 | // this is remark, so we'll use up all available threads |
ysr@777 | 1940 | int active_workers = 1; |
ysr@777 | 1941 | set_phase(active_workers, false); |
ysr@777 | 1942 | |
ysr@777 | 1943 | CMRemarkTask remarkTask(this); |
ysr@777 | 1944 | // We will start all available threads, even if we decide that the |
ysr@777 | 1945 | // active_workers will be fewer. The extra ones will just bail out |
ysr@777 | 1946 | // immediately. |
ysr@777 | 1947 | remarkTask.work(0); |
ysr@777 | 1948 | |
ysr@777 | 1949 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 1950 | guarantee( satb_mq_set.completed_buffers_num() == 0, "invariant" ); |
ysr@777 | 1951 | } |
ysr@777 | 1952 | |
ysr@777 | 1953 | print_stats(); |
ysr@777 | 1954 | |
ysr@777 | 1955 | if (!restart_for_overflow()) |
ysr@777 | 1956 | set_non_marking_state(); |
ysr@777 | 1957 | |
ysr@777 | 1958 | #if VERIFY_OBJS_PROCESSED |
ysr@777 | 1959 | if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) { |
ysr@777 | 1960 | gclog_or_tty->print_cr("Processed = %d, enqueued = %d.", |
ysr@777 | 1961 | _scan_obj_cl.objs_processed, |
ysr@777 | 1962 | ThreadLocalObjQueue::objs_enqueued); |
ysr@777 | 1963 | guarantee(_scan_obj_cl.objs_processed == |
ysr@777 | 1964 | ThreadLocalObjQueue::objs_enqueued, |
ysr@777 | 1965 | "Different number of objs processed and enqueued."); |
ysr@777 | 1966 | } |
ysr@777 | 1967 | #endif |
ysr@777 | 1968 | } |
ysr@777 | 1969 | |
ysr@777 | 1970 | class ReachablePrinterOopClosure: public OopClosure { |
ysr@777 | 1971 | private: |
ysr@777 | 1972 | G1CollectedHeap* _g1h; |
ysr@777 | 1973 | CMBitMapRO* _bitmap; |
ysr@777 | 1974 | outputStream* _out; |
ysr@777 | 1975 | |
ysr@777 | 1976 | public: |
ysr@777 | 1977 | ReachablePrinterOopClosure(CMBitMapRO* bitmap, outputStream* out) : |
ysr@777 | 1978 | _bitmap(bitmap), _g1h(G1CollectedHeap::heap()), _out(out) { } |
ysr@777 | 1979 | |
ysr@1280 | 1980 | void do_oop(narrowOop* p) { do_oop_work(p); } |
ysr@1280 | 1981 | void do_oop( oop* p) { do_oop_work(p); } |
ysr@1280 | 1982 | |
ysr@1280 | 1983 | template <class T> void do_oop_work(T* p) { |
ysr@1280 | 1984 | oop obj = oopDesc::load_decode_heap_oop(p); |
ysr@777 | 1985 | const char* str = NULL; |
ysr@777 | 1986 | const char* str2 = ""; |
ysr@777 | 1987 | |
ysr@777 | 1988 | if (!_g1h->is_in_g1_reserved(obj)) |
ysr@777 | 1989 | str = "outside G1 reserved"; |
ysr@777 | 1990 | else { |
ysr@777 | 1991 | HeapRegion* hr = _g1h->heap_region_containing(obj); |
ysr@777 | 1992 | guarantee( hr != NULL, "invariant" ); |
ysr@777 | 1993 | if (hr->obj_allocated_since_prev_marking(obj)) { |
ysr@777 | 1994 | str = "over TAMS"; |
ysr@777 | 1995 | if (_bitmap->isMarked((HeapWord*) obj)) |
ysr@777 | 1996 | str2 = " AND MARKED"; |
ysr@777 | 1997 | } else if (_bitmap->isMarked((HeapWord*) obj)) |
ysr@777 | 1998 | str = "marked"; |
ysr@777 | 1999 | else |
ysr@777 | 2000 | str = "#### NOT MARKED ####"; |
ysr@777 | 2001 | } |
ysr@777 | 2002 | |
ysr@777 | 2003 | _out->print_cr(" "PTR_FORMAT" contains "PTR_FORMAT" %s%s", |
ysr@777 | 2004 | p, (void*) obj, str, str2); |
ysr@777 | 2005 | } |
ysr@777 | 2006 | }; |
ysr@777 | 2007 | |
ysr@777 | 2008 | class ReachablePrinterClosure: public BitMapClosure { |
ysr@777 | 2009 | private: |
ysr@777 | 2010 | CMBitMapRO* _bitmap; |
ysr@777 | 2011 | outputStream* _out; |
ysr@777 | 2012 | |
ysr@777 | 2013 | public: |
ysr@777 | 2014 | ReachablePrinterClosure(CMBitMapRO* bitmap, outputStream* out) : |
ysr@777 | 2015 | _bitmap(bitmap), _out(out) { } |
ysr@777 | 2016 | |
ysr@777 | 2017 | bool do_bit(size_t offset) { |
ysr@777 | 2018 | HeapWord* addr = _bitmap->offsetToHeapWord(offset); |
ysr@777 | 2019 | ReachablePrinterOopClosure oopCl(_bitmap, _out); |
ysr@777 | 2020 | |
ysr@777 | 2021 | _out->print_cr(" obj "PTR_FORMAT", offset %10d (marked)", addr, offset); |
ysr@777 | 2022 | oop(addr)->oop_iterate(&oopCl); |
ysr@777 | 2023 | _out->print_cr(""); |
ysr@777 | 2024 | |
ysr@777 | 2025 | return true; |
ysr@777 | 2026 | } |
ysr@777 | 2027 | }; |
ysr@777 | 2028 | |
ysr@777 | 2029 | class ObjInRegionReachablePrinterClosure : public ObjectClosure { |
ysr@777 | 2030 | private: |
ysr@777 | 2031 | CMBitMapRO* _bitmap; |
ysr@777 | 2032 | outputStream* _out; |
ysr@777 | 2033 | |
ysr@777 | 2034 | public: |
ysr@777 | 2035 | void do_object(oop o) { |
ysr@777 | 2036 | ReachablePrinterOopClosure oopCl(_bitmap, _out); |
ysr@777 | 2037 | |
ysr@777 | 2038 | _out->print_cr(" obj "PTR_FORMAT" (over TAMS)", (void*) o); |
ysr@777 | 2039 | o->oop_iterate(&oopCl); |
ysr@777 | 2040 | _out->print_cr(""); |
ysr@777 | 2041 | } |
ysr@777 | 2042 | |
ysr@777 | 2043 | ObjInRegionReachablePrinterClosure(CMBitMapRO* bitmap, outputStream* out) : |
ysr@777 | 2044 | _bitmap(bitmap), _out(out) { } |
ysr@777 | 2045 | }; |
ysr@777 | 2046 | |
ysr@777 | 2047 | class RegionReachablePrinterClosure : public HeapRegionClosure { |
ysr@777 | 2048 | private: |
ysr@777 | 2049 | CMBitMapRO* _bitmap; |
ysr@777 | 2050 | outputStream* _out; |
ysr@777 | 2051 | |
ysr@777 | 2052 | public: |
ysr@777 | 2053 | bool doHeapRegion(HeapRegion* hr) { |
ysr@777 | 2054 | HeapWord* b = hr->bottom(); |
ysr@777 | 2055 | HeapWord* e = hr->end(); |
ysr@777 | 2056 | HeapWord* t = hr->top(); |
ysr@777 | 2057 | HeapWord* p = hr->prev_top_at_mark_start(); |
ysr@777 | 2058 | _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " |
ysr@777 | 2059 | "PTAMS: "PTR_FORMAT, b, e, t, p); |
ysr@777 | 2060 | _out->print_cr(""); |
ysr@777 | 2061 | |
ysr@777 | 2062 | ObjInRegionReachablePrinterClosure ocl(_bitmap, _out); |
ysr@777 | 2063 | hr->object_iterate_mem_careful(MemRegion(p, t), &ocl); |
ysr@777 | 2064 | |
ysr@777 | 2065 | return false; |
ysr@777 | 2066 | } |
ysr@777 | 2067 | |
ysr@777 | 2068 | RegionReachablePrinterClosure(CMBitMapRO* bitmap, |
ysr@777 | 2069 | outputStream* out) : |
ysr@777 | 2070 | _bitmap(bitmap), _out(out) { } |
ysr@777 | 2071 | }; |
ysr@777 | 2072 | |
ysr@777 | 2073 | void ConcurrentMark::print_prev_bitmap_reachable() { |
ysr@777 | 2074 | outputStream* out = gclog_or_tty; |
ysr@777 | 2075 | |
ysr@777 | 2076 | #if SEND_HEAP_DUMP_TO_FILE |
ysr@777 | 2077 | guarantee(heap_dump_file == NULL, "Protocol"); |
ysr@777 | 2078 | char fn_buf[100]; |
ysr@777 | 2079 | sprintf(fn_buf, "/tmp/dump.txt.%d", os::current_process_id()); |
ysr@777 | 2080 | heap_dump_file = fopen(fn_buf, "w"); |
ysr@777 | 2081 | fileStream fstream(heap_dump_file); |
ysr@777 | 2082 | out = &fstream; |
ysr@777 | 2083 | #endif // SEND_HEAP_DUMP_TO_FILE |
ysr@777 | 2084 | |
ysr@777 | 2085 | RegionReachablePrinterClosure rcl(_prevMarkBitMap, out); |
ysr@777 | 2086 | out->print_cr("--- ITERATING OVER REGIONS WITH PTAMS < TOP"); |
ysr@777 | 2087 | _g1h->heap_region_iterate(&rcl); |
ysr@777 | 2088 | out->print_cr(""); |
ysr@777 | 2089 | |
ysr@777 | 2090 | ReachablePrinterClosure cl(_prevMarkBitMap, out); |
ysr@777 | 2091 | out->print_cr("--- REACHABLE OBJECTS ON THE BITMAP"); |
ysr@777 | 2092 | _prevMarkBitMap->iterate(&cl); |
ysr@777 | 2093 | out->print_cr(""); |
ysr@777 | 2094 | |
ysr@777 | 2095 | #if SEND_HEAP_DUMP_TO_FILE |
ysr@777 | 2096 | fclose(heap_dump_file); |
ysr@777 | 2097 | heap_dump_file = NULL; |
ysr@777 | 2098 | #endif // SEND_HEAP_DUMP_TO_FILE |
ysr@777 | 2099 | } |
ysr@777 | 2100 | |
ysr@777 | 2101 | // This note is for drainAllSATBBuffers and the code in between. |
ysr@777 | 2102 | // In the future we could reuse a task to do this work during an |
ysr@777 | 2103 | // evacuation pause (since now tasks are not active and can be claimed |
ysr@777 | 2104 | // during an evacuation pause). This was a late change to the code and |
ysr@777 | 2105 | // is currently not being taken advantage of. |
ysr@777 | 2106 | |
ysr@777 | 2107 | class CMGlobalObjectClosure : public ObjectClosure { |
ysr@777 | 2108 | private: |
ysr@777 | 2109 | ConcurrentMark* _cm; |
ysr@777 | 2110 | |
ysr@777 | 2111 | public: |
ysr@777 | 2112 | void do_object(oop obj) { |
ysr@777 | 2113 | _cm->deal_with_reference(obj); |
ysr@777 | 2114 | } |
ysr@777 | 2115 | |
ysr@777 | 2116 | CMGlobalObjectClosure(ConcurrentMark* cm) : _cm(cm) { } |
ysr@777 | 2117 | }; |
ysr@777 | 2118 | |
ysr@777 | 2119 | void ConcurrentMark::deal_with_reference(oop obj) { |
ysr@777 | 2120 | if (verbose_high()) |
ysr@777 | 2121 | gclog_or_tty->print_cr("[global] we're dealing with reference "PTR_FORMAT, |
ysr@777 | 2122 | (void*) obj); |
ysr@777 | 2123 | |
ysr@777 | 2124 | |
ysr@777 | 2125 | HeapWord* objAddr = (HeapWord*) obj; |
ysr@1280 | 2126 | assert(obj->is_oop_or_null(true /* ignore mark word */), "Error"); |
ysr@777 | 2127 | if (_g1h->is_in_g1_reserved(objAddr)) { |
ysr@777 | 2128 | tmp_guarantee_CM( obj != NULL, "is_in_g1_reserved should ensure this" ); |
ysr@777 | 2129 | HeapRegion* hr = _g1h->heap_region_containing(obj); |
ysr@777 | 2130 | if (_g1h->is_obj_ill(obj, hr)) { |
ysr@777 | 2131 | if (verbose_high()) |
ysr@777 | 2132 | gclog_or_tty->print_cr("[global] "PTR_FORMAT" is not considered " |
ysr@777 | 2133 | "marked", (void*) obj); |
ysr@777 | 2134 | |
ysr@777 | 2135 | // we need to mark it first |
ysr@777 | 2136 | if (_nextMarkBitMap->parMark(objAddr)) { |
ysr@777 | 2137 | // No OrderAccess:store_load() is needed. It is implicit in the |
ysr@777 | 2138 | // CAS done in parMark(objAddr) above |
ysr@777 | 2139 | HeapWord* finger = _finger; |
ysr@777 | 2140 | if (objAddr < finger) { |
ysr@777 | 2141 | if (verbose_high()) |
ysr@777 | 2142 | gclog_or_tty->print_cr("[global] below the global finger " |
ysr@777 | 2143 | "("PTR_FORMAT"), pushing it", finger); |
ysr@777 | 2144 | if (!mark_stack_push(obj)) { |
ysr@777 | 2145 | if (verbose_low()) |
ysr@777 | 2146 | gclog_or_tty->print_cr("[global] global stack overflow during " |
ysr@777 | 2147 | "deal_with_reference"); |
ysr@777 | 2148 | } |
ysr@777 | 2149 | } |
ysr@777 | 2150 | } |
ysr@777 | 2151 | } |
ysr@777 | 2152 | } |
ysr@777 | 2153 | } |
ysr@777 | 2154 | |
ysr@777 | 2155 | void ConcurrentMark::drainAllSATBBuffers() { |
ysr@777 | 2156 | CMGlobalObjectClosure oc(this); |
ysr@777 | 2157 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 2158 | satb_mq_set.set_closure(&oc); |
ysr@777 | 2159 | |
ysr@777 | 2160 | while (satb_mq_set.apply_closure_to_completed_buffer()) { |
ysr@777 | 2161 | if (verbose_medium()) |
ysr@777 | 2162 | gclog_or_tty->print_cr("[global] processed an SATB buffer"); |
ysr@777 | 2163 | } |
ysr@777 | 2164 | |
ysr@777 | 2165 | // no need to check whether we should do this, as this is only |
ysr@777 | 2166 | // called during an evacuation pause |
ysr@777 | 2167 | satb_mq_set.iterate_closure_all_threads(); |
ysr@777 | 2168 | |
ysr@777 | 2169 | satb_mq_set.set_closure(NULL); |
ysr@777 | 2170 | guarantee( satb_mq_set.completed_buffers_num() == 0, "invariant" ); |
ysr@777 | 2171 | } |
ysr@777 | 2172 | |
ysr@777 | 2173 | void ConcurrentMark::markPrev(oop p) { |
ysr@777 | 2174 | // Note we are overriding the read-only view of the prev map here, via |
ysr@777 | 2175 | // the cast. |
ysr@777 | 2176 | ((CMBitMap*)_prevMarkBitMap)->mark((HeapWord*)p); |
ysr@777 | 2177 | } |
ysr@777 | 2178 | |
ysr@777 | 2179 | void ConcurrentMark::clear(oop p) { |
ysr@777 | 2180 | assert(p != NULL && p->is_oop(), "expected an oop"); |
ysr@777 | 2181 | HeapWord* addr = (HeapWord*)p; |
ysr@777 | 2182 | assert(addr >= _nextMarkBitMap->startWord() || |
ysr@777 | 2183 | addr < _nextMarkBitMap->endWord(), "in a region"); |
ysr@777 | 2184 | |
ysr@777 | 2185 | _nextMarkBitMap->clear(addr); |
ysr@777 | 2186 | } |
ysr@777 | 2187 | |
ysr@777 | 2188 | void ConcurrentMark::clearRangeBothMaps(MemRegion mr) { |
ysr@777 | 2189 | // Note we are overriding the read-only view of the prev map here, via |
ysr@777 | 2190 | // the cast. |
ysr@777 | 2191 | ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); |
ysr@777 | 2192 | _nextMarkBitMap->clearRange(mr); |
ysr@777 | 2193 | } |
ysr@777 | 2194 | |
ysr@777 | 2195 | HeapRegion* |
ysr@777 | 2196 | ConcurrentMark::claim_region(int task_num) { |
ysr@777 | 2197 | // "checkpoint" the finger |
ysr@777 | 2198 | HeapWord* finger = _finger; |
ysr@777 | 2199 | |
ysr@777 | 2200 | // _heap_end will not change underneath our feet; it only changes at |
ysr@777 | 2201 | // yield points. |
ysr@777 | 2202 | while (finger < _heap_end) { |
ysr@777 | 2203 | tmp_guarantee_CM( _g1h->is_in_g1_reserved(finger), "invariant" ); |
ysr@777 | 2204 | |
ysr@777 | 2205 | // is the gap between reading the finger and doing the CAS too long? |
ysr@777 | 2206 | |
ysr@777 | 2207 | HeapRegion* curr_region = _g1h->heap_region_containing(finger); |
ysr@777 | 2208 | HeapWord* bottom = curr_region->bottom(); |
ysr@777 | 2209 | HeapWord* end = curr_region->end(); |
ysr@777 | 2210 | HeapWord* limit = curr_region->next_top_at_mark_start(); |
ysr@777 | 2211 | |
ysr@777 | 2212 | if (verbose_low()) |
ysr@777 | 2213 | gclog_or_tty->print_cr("[%d] curr_region = "PTR_FORMAT" " |
ysr@777 | 2214 | "["PTR_FORMAT", "PTR_FORMAT"), " |
ysr@777 | 2215 | "limit = "PTR_FORMAT, |
ysr@777 | 2216 | task_num, curr_region, bottom, end, limit); |
ysr@777 | 2217 | |
ysr@777 | 2218 | HeapWord* res = |
ysr@777 | 2219 | (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); |
ysr@777 | 2220 | if (res == finger) { |
ysr@777 | 2221 | // we succeeded |
ysr@777 | 2222 | |
ysr@777 | 2223 | // notice that _finger == end cannot be guaranteed here since, |
ysr@777 | 2224 | // someone else might have moved the finger even further |
ysr@777 | 2225 | guarantee( _finger >= end, "the finger should have moved forward" ); |
ysr@777 | 2226 | |
ysr@777 | 2227 | if (verbose_low()) |
ysr@777 | 2228 | gclog_or_tty->print_cr("[%d] we were successful with region = " |
ysr@777 | 2229 | PTR_FORMAT, task_num, curr_region); |
ysr@777 | 2230 | |
ysr@777 | 2231 | if (limit > bottom) { |
ysr@777 | 2232 | if (verbose_low()) |
ysr@777 | 2233 | gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is not empty, " |
ysr@777 | 2234 | "returning it ", task_num, curr_region); |
ysr@777 | 2235 | return curr_region; |
ysr@777 | 2236 | } else { |
ysr@777 | 2237 | tmp_guarantee_CM( limit == bottom, |
ysr@777 | 2238 | "the region limit should be at bottom" ); |
ysr@777 | 2239 | if (verbose_low()) |
ysr@777 | 2240 | gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is empty, " |
ysr@777 | 2241 | "returning NULL", task_num, curr_region); |
ysr@777 | 2242 | // we return NULL and the caller should try calling |
ysr@777 | 2243 | // claim_region() again. |
ysr@777 | 2244 | return NULL; |
ysr@777 | 2245 | } |
ysr@777 | 2246 | } else { |
ysr@777 | 2247 | guarantee( _finger > finger, "the finger should have moved forward" ); |
ysr@777 | 2248 | if (verbose_low()) |
ysr@777 | 2249 | gclog_or_tty->print_cr("[%d] somebody else moved the finger, " |
ysr@777 | 2250 | "global finger = "PTR_FORMAT", " |
ysr@777 | 2251 | "our finger = "PTR_FORMAT, |
ysr@777 | 2252 | task_num, _finger, finger); |
ysr@777 | 2253 | |
ysr@777 | 2254 | // read it again |
ysr@777 | 2255 | finger = _finger; |
ysr@777 | 2256 | } |
ysr@777 | 2257 | } |
ysr@777 | 2258 | |
ysr@777 | 2259 | return NULL; |
ysr@777 | 2260 | } |
ysr@777 | 2261 | |
ysr@777 | 2262 | void ConcurrentMark::oops_do(OopClosure* cl) { |
ysr@777 | 2263 | if (_markStack.size() > 0 && verbose_low()) |
ysr@777 | 2264 | gclog_or_tty->print_cr("[global] scanning the global marking stack, " |
ysr@777 | 2265 | "size = %d", _markStack.size()); |
ysr@777 | 2266 | // we first iterate over the contents of the mark stack... |
ysr@777 | 2267 | _markStack.oops_do(cl); |
ysr@777 | 2268 | |
ysr@777 | 2269 | for (int i = 0; i < (int)_max_task_num; ++i) { |
ysr@777 | 2270 | OopTaskQueue* queue = _task_queues->queue((int)i); |
ysr@777 | 2271 | |
ysr@777 | 2272 | if (queue->size() > 0 && verbose_low()) |
ysr@777 | 2273 | gclog_or_tty->print_cr("[global] scanning task queue of task %d, " |
ysr@777 | 2274 | "size = %d", i, queue->size()); |
ysr@777 | 2275 | |
ysr@777 | 2276 | // ...then over the contents of the all the task queues. |
ysr@777 | 2277 | queue->oops_do(cl); |
ysr@777 | 2278 | } |
ysr@777 | 2279 | |
ysr@777 | 2280 | // finally, invalidate any entries that in the region stack that |
ysr@777 | 2281 | // point into the collection set |
ysr@777 | 2282 | if (_regionStack.invalidate_entries_into_cset()) { |
ysr@777 | 2283 | // otherwise, any gray objects copied during the evacuation pause |
ysr@777 | 2284 | // might not be visited. |
ysr@777 | 2285 | guarantee( _should_gray_objects, "invariant" ); |
ysr@777 | 2286 | } |
ysr@777 | 2287 | } |
ysr@777 | 2288 | |
ysr@777 | 2289 | void ConcurrentMark::clear_marking_state() { |
ysr@777 | 2290 | _markStack.setEmpty(); |
ysr@777 | 2291 | _markStack.clear_overflow(); |
ysr@777 | 2292 | _regionStack.setEmpty(); |
ysr@777 | 2293 | _regionStack.clear_overflow(); |
ysr@777 | 2294 | clear_has_overflown(); |
ysr@777 | 2295 | _finger = _heap_start; |
ysr@777 | 2296 | |
ysr@777 | 2297 | for (int i = 0; i < (int)_max_task_num; ++i) { |
ysr@777 | 2298 | OopTaskQueue* queue = _task_queues->queue(i); |
ysr@777 | 2299 | queue->set_empty(); |
ysr@777 | 2300 | } |
ysr@777 | 2301 | } |
ysr@777 | 2302 | |
ysr@777 | 2303 | void ConcurrentMark::print_stats() { |
ysr@777 | 2304 | if (verbose_stats()) { |
ysr@777 | 2305 | gclog_or_tty->print_cr("---------------------------------------------------------------------"); |
ysr@777 | 2306 | for (size_t i = 0; i < _active_tasks; ++i) { |
ysr@777 | 2307 | _tasks[i]->print_stats(); |
ysr@777 | 2308 | gclog_or_tty->print_cr("---------------------------------------------------------------------"); |
ysr@777 | 2309 | } |
ysr@777 | 2310 | } |
ysr@777 | 2311 | } |
ysr@777 | 2312 | |
ysr@777 | 2313 | class CSMarkOopClosure: public OopClosure { |
ysr@777 | 2314 | friend class CSMarkBitMapClosure; |
ysr@777 | 2315 | |
ysr@777 | 2316 | G1CollectedHeap* _g1h; |
ysr@777 | 2317 | CMBitMap* _bm; |
ysr@777 | 2318 | ConcurrentMark* _cm; |
ysr@777 | 2319 | oop* _ms; |
ysr@777 | 2320 | jint* _array_ind_stack; |
ysr@777 | 2321 | int _ms_size; |
ysr@777 | 2322 | int _ms_ind; |
ysr@777 | 2323 | int _array_increment; |
ysr@777 | 2324 | |
ysr@777 | 2325 | bool push(oop obj, int arr_ind = 0) { |
ysr@777 | 2326 | if (_ms_ind == _ms_size) { |
ysr@777 | 2327 | gclog_or_tty->print_cr("Mark stack is full."); |
ysr@777 | 2328 | return false; |
ysr@777 | 2329 | } |
ysr@777 | 2330 | _ms[_ms_ind] = obj; |
ysr@777 | 2331 | if (obj->is_objArray()) _array_ind_stack[_ms_ind] = arr_ind; |
ysr@777 | 2332 | _ms_ind++; |
ysr@777 | 2333 | return true; |
ysr@777 | 2334 | } |
ysr@777 | 2335 | |
ysr@777 | 2336 | oop pop() { |
ysr@777 | 2337 | if (_ms_ind == 0) return NULL; |
ysr@777 | 2338 | else { |
ysr@777 | 2339 | _ms_ind--; |
ysr@777 | 2340 | return _ms[_ms_ind]; |
ysr@777 | 2341 | } |
ysr@777 | 2342 | } |
ysr@777 | 2343 | |
ysr@1280 | 2344 | template <class T> bool drain() { |
ysr@777 | 2345 | while (_ms_ind > 0) { |
ysr@777 | 2346 | oop obj = pop(); |
ysr@777 | 2347 | assert(obj != NULL, "Since index was non-zero."); |
ysr@777 | 2348 | if (obj->is_objArray()) { |
ysr@777 | 2349 | jint arr_ind = _array_ind_stack[_ms_ind]; |
ysr@777 | 2350 | objArrayOop aobj = objArrayOop(obj); |
ysr@777 | 2351 | jint len = aobj->length(); |
ysr@777 | 2352 | jint next_arr_ind = arr_ind + _array_increment; |
ysr@777 | 2353 | if (next_arr_ind < len) { |
ysr@777 | 2354 | push(obj, next_arr_ind); |
ysr@777 | 2355 | } |
ysr@777 | 2356 | // Now process this portion of this one. |
ysr@777 | 2357 | int lim = MIN2(next_arr_ind, len); |
ysr@777 | 2358 | for (int j = arr_ind; j < lim; j++) { |
apetrusenko@1347 | 2359 | do_oop(aobj->objArrayOopDesc::obj_at_addr<T>(j)); |
ysr@777 | 2360 | } |
ysr@777 | 2361 | |
ysr@777 | 2362 | } else { |
ysr@777 | 2363 | obj->oop_iterate(this); |
ysr@777 | 2364 | } |
ysr@777 | 2365 | if (abort()) return false; |
ysr@777 | 2366 | } |
ysr@777 | 2367 | return true; |
ysr@777 | 2368 | } |
ysr@777 | 2369 | |
ysr@777 | 2370 | public: |
ysr@777 | 2371 | CSMarkOopClosure(ConcurrentMark* cm, int ms_size) : |
ysr@777 | 2372 | _g1h(G1CollectedHeap::heap()), |
ysr@777 | 2373 | _cm(cm), |
ysr@777 | 2374 | _bm(cm->nextMarkBitMap()), |
ysr@777 | 2375 | _ms_size(ms_size), _ms_ind(0), |
ysr@777 | 2376 | _ms(NEW_C_HEAP_ARRAY(oop, ms_size)), |
ysr@777 | 2377 | _array_ind_stack(NEW_C_HEAP_ARRAY(jint, ms_size)), |
ysr@777 | 2378 | _array_increment(MAX2(ms_size/8, 16)) |
ysr@777 | 2379 | {} |
ysr@777 | 2380 | |
ysr@777 | 2381 | ~CSMarkOopClosure() { |
ysr@777 | 2382 | FREE_C_HEAP_ARRAY(oop, _ms); |
ysr@777 | 2383 | FREE_C_HEAP_ARRAY(jint, _array_ind_stack); |
ysr@777 | 2384 | } |
ysr@777 | 2385 | |
ysr@1280 | 2386 | virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
ysr@1280 | 2387 | virtual void do_oop( oop* p) { do_oop_work(p); } |
ysr@1280 | 2388 | |
ysr@1280 | 2389 | template <class T> void do_oop_work(T* p) { |
ysr@1280 | 2390 | T heap_oop = oopDesc::load_heap_oop(p); |
ysr@1280 | 2391 | if (oopDesc::is_null(heap_oop)) return; |
ysr@1280 | 2392 | oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
ysr@777 | 2393 | if (obj->is_forwarded()) { |
ysr@777 | 2394 | // If the object has already been forwarded, we have to make sure |
ysr@777 | 2395 | // that it's marked. So follow the forwarding pointer. Note that |
ysr@777 | 2396 | // this does the right thing for self-forwarding pointers in the |
ysr@777 | 2397 | // evacuation failure case. |
ysr@777 | 2398 | obj = obj->forwardee(); |
ysr@777 | 2399 | } |
ysr@777 | 2400 | HeapRegion* hr = _g1h->heap_region_containing(obj); |
ysr@777 | 2401 | if (hr != NULL) { |
ysr@777 | 2402 | if (hr->in_collection_set()) { |
ysr@777 | 2403 | if (_g1h->is_obj_ill(obj)) { |
ysr@777 | 2404 | _bm->mark((HeapWord*)obj); |
ysr@777 | 2405 | if (!push(obj)) { |
ysr@777 | 2406 | gclog_or_tty->print_cr("Setting abort in CSMarkOopClosure because push failed."); |
ysr@777 | 2407 | set_abort(); |
ysr@777 | 2408 | } |
ysr@777 | 2409 | } |
ysr@777 | 2410 | } else { |
ysr@777 | 2411 | // Outside the collection set; we need to gray it |
ysr@777 | 2412 | _cm->deal_with_reference(obj); |
ysr@777 | 2413 | } |
ysr@777 | 2414 | } |
ysr@777 | 2415 | } |
ysr@777 | 2416 | }; |
ysr@777 | 2417 | |
ysr@777 | 2418 | class CSMarkBitMapClosure: public BitMapClosure { |
ysr@777 | 2419 | G1CollectedHeap* _g1h; |
ysr@777 | 2420 | CMBitMap* _bitMap; |
ysr@777 | 2421 | ConcurrentMark* _cm; |
ysr@777 | 2422 | CSMarkOopClosure _oop_cl; |
ysr@777 | 2423 | public: |
ysr@777 | 2424 | CSMarkBitMapClosure(ConcurrentMark* cm, int ms_size) : |
ysr@777 | 2425 | _g1h(G1CollectedHeap::heap()), |
ysr@777 | 2426 | _bitMap(cm->nextMarkBitMap()), |
ysr@777 | 2427 | _oop_cl(cm, ms_size) |
ysr@777 | 2428 | {} |
ysr@777 | 2429 | |
ysr@777 | 2430 | ~CSMarkBitMapClosure() {} |
ysr@777 | 2431 | |
ysr@777 | 2432 | bool do_bit(size_t offset) { |
ysr@777 | 2433 | // convert offset into a HeapWord* |
ysr@777 | 2434 | HeapWord* addr = _bitMap->offsetToHeapWord(offset); |
ysr@777 | 2435 | assert(_bitMap->endWord() && addr < _bitMap->endWord(), |
ysr@777 | 2436 | "address out of range"); |
ysr@777 | 2437 | assert(_bitMap->isMarked(addr), "tautology"); |
ysr@777 | 2438 | oop obj = oop(addr); |
ysr@777 | 2439 | if (!obj->is_forwarded()) { |
ysr@777 | 2440 | if (!_oop_cl.push(obj)) return false; |
ysr@1280 | 2441 | if (UseCompressedOops) { |
ysr@1280 | 2442 | if (!_oop_cl.drain<narrowOop>()) return false; |
ysr@1280 | 2443 | } else { |
ysr@1280 | 2444 | if (!_oop_cl.drain<oop>()) return false; |
ysr@1280 | 2445 | } |
ysr@777 | 2446 | } |
ysr@777 | 2447 | // Otherwise... |
ysr@777 | 2448 | return true; |
ysr@777 | 2449 | } |
ysr@777 | 2450 | }; |
ysr@777 | 2451 | |
ysr@777 | 2452 | |
ysr@777 | 2453 | class CompleteMarkingInCSHRClosure: public HeapRegionClosure { |
ysr@777 | 2454 | CMBitMap* _bm; |
ysr@777 | 2455 | CSMarkBitMapClosure _bit_cl; |
ysr@777 | 2456 | enum SomePrivateConstants { |
ysr@777 | 2457 | MSSize = 1000 |
ysr@777 | 2458 | }; |
ysr@777 | 2459 | bool _completed; |
ysr@777 | 2460 | public: |
ysr@777 | 2461 | CompleteMarkingInCSHRClosure(ConcurrentMark* cm) : |
ysr@777 | 2462 | _bm(cm->nextMarkBitMap()), |
ysr@777 | 2463 | _bit_cl(cm, MSSize), |
ysr@777 | 2464 | _completed(true) |
ysr@777 | 2465 | {} |
ysr@777 | 2466 | |
ysr@777 | 2467 | ~CompleteMarkingInCSHRClosure() {} |
ysr@777 | 2468 | |
ysr@777 | 2469 | bool doHeapRegion(HeapRegion* r) { |
ysr@777 | 2470 | if (!r->evacuation_failed()) { |
ysr@777 | 2471 | MemRegion mr = MemRegion(r->bottom(), r->next_top_at_mark_start()); |
ysr@777 | 2472 | if (!mr.is_empty()) { |
ysr@777 | 2473 | if (!_bm->iterate(&_bit_cl, mr)) { |
ysr@777 | 2474 | _completed = false; |
ysr@777 | 2475 | return true; |
ysr@777 | 2476 | } |
ysr@777 | 2477 | } |
ysr@777 | 2478 | } |
ysr@777 | 2479 | return false; |
ysr@777 | 2480 | } |
ysr@777 | 2481 | |
ysr@777 | 2482 | bool completed() { return _completed; } |
ysr@777 | 2483 | }; |
ysr@777 | 2484 | |
ysr@777 | 2485 | class ClearMarksInHRClosure: public HeapRegionClosure { |
ysr@777 | 2486 | CMBitMap* _bm; |
ysr@777 | 2487 | public: |
ysr@777 | 2488 | ClearMarksInHRClosure(CMBitMap* bm): _bm(bm) { } |
ysr@777 | 2489 | |
ysr@777 | 2490 | bool doHeapRegion(HeapRegion* r) { |
ysr@777 | 2491 | if (!r->used_region().is_empty() && !r->evacuation_failed()) { |
ysr@777 | 2492 | MemRegion usedMR = r->used_region(); |
ysr@777 | 2493 | _bm->clearRange(r->used_region()); |
ysr@777 | 2494 | } |
ysr@777 | 2495 | return false; |
ysr@777 | 2496 | } |
ysr@777 | 2497 | }; |
ysr@777 | 2498 | |
ysr@777 | 2499 | void ConcurrentMark::complete_marking_in_collection_set() { |
ysr@777 | 2500 | G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
ysr@777 | 2501 | |
ysr@777 | 2502 | if (!g1h->mark_in_progress()) { |
ysr@777 | 2503 | g1h->g1_policy()->record_mark_closure_time(0.0); |
ysr@777 | 2504 | return; |
ysr@777 | 2505 | } |
ysr@777 | 2506 | |
ysr@777 | 2507 | int i = 1; |
ysr@777 | 2508 | double start = os::elapsedTime(); |
ysr@777 | 2509 | while (true) { |
ysr@777 | 2510 | i++; |
ysr@777 | 2511 | CompleteMarkingInCSHRClosure cmplt(this); |
ysr@777 | 2512 | g1h->collection_set_iterate(&cmplt); |
ysr@777 | 2513 | if (cmplt.completed()) break; |
ysr@777 | 2514 | } |
ysr@777 | 2515 | double end_time = os::elapsedTime(); |
ysr@777 | 2516 | double elapsed_time_ms = (end_time - start) * 1000.0; |
ysr@777 | 2517 | g1h->g1_policy()->record_mark_closure_time(elapsed_time_ms); |
ysr@777 | 2518 | if (PrintGCDetails) { |
ysr@777 | 2519 | gclog_or_tty->print_cr("Mark closure took %5.2f ms.", elapsed_time_ms); |
ysr@777 | 2520 | } |
ysr@777 | 2521 | |
ysr@777 | 2522 | ClearMarksInHRClosure clr(nextMarkBitMap()); |
ysr@777 | 2523 | g1h->collection_set_iterate(&clr); |
ysr@777 | 2524 | } |
ysr@777 | 2525 | |
ysr@777 | 2526 | // The next two methods deal with the following optimisation. Some |
ysr@777 | 2527 | // objects are gray by being marked and located above the finger. If |
ysr@777 | 2528 | // they are copied, during an evacuation pause, below the finger then |
ysr@777 | 2529 | // the need to be pushed on the stack. The observation is that, if |
ysr@777 | 2530 | // there are no regions in the collection set located above the |
ysr@777 | 2531 | // finger, then the above cannot happen, hence we do not need to |
ysr@777 | 2532 | // explicitly gray any objects when copying them to below the |
ysr@777 | 2533 | // finger. The global stack will be scanned to ensure that, if it |
ysr@777 | 2534 | // points to objects being copied, it will update their |
ysr@777 | 2535 | // location. There is a tricky situation with the gray objects in |
ysr@777 | 2536 | // region stack that are being coped, however. See the comment in |
ysr@777 | 2537 | // newCSet(). |
ysr@777 | 2538 | |
ysr@777 | 2539 | void ConcurrentMark::newCSet() { |
ysr@777 | 2540 | if (!concurrent_marking_in_progress()) |
ysr@777 | 2541 | // nothing to do if marking is not in progress |
ysr@777 | 2542 | return; |
ysr@777 | 2543 | |
ysr@777 | 2544 | // find what the lowest finger is among the global and local fingers |
ysr@777 | 2545 | _min_finger = _finger; |
ysr@777 | 2546 | for (int i = 0; i < (int)_max_task_num; ++i) { |
ysr@777 | 2547 | CMTask* task = _tasks[i]; |
ysr@777 | 2548 | HeapWord* task_finger = task->finger(); |
ysr@777 | 2549 | if (task_finger != NULL && task_finger < _min_finger) |
ysr@777 | 2550 | _min_finger = task_finger; |
ysr@777 | 2551 | } |
ysr@777 | 2552 | |
ysr@777 | 2553 | _should_gray_objects = false; |
ysr@777 | 2554 | |
ysr@777 | 2555 | // This fixes a very subtle and fustrating bug. It might be the case |
ysr@777 | 2556 | // that, during en evacuation pause, heap regions that contain |
ysr@777 | 2557 | // objects that are gray (by being in regions contained in the |
ysr@777 | 2558 | // region stack) are included in the collection set. Since such gray |
ysr@777 | 2559 | // objects will be moved, and because it's not easy to redirect |
ysr@777 | 2560 | // region stack entries to point to a new location (because objects |
ysr@777 | 2561 | // in one region might be scattered to multiple regions after they |
ysr@777 | 2562 | // are copied), one option is to ensure that all marked objects |
ysr@777 | 2563 | // copied during a pause are pushed on the stack. Notice, however, |
ysr@777 | 2564 | // that this problem can only happen when the region stack is not |
ysr@777 | 2565 | // empty during an evacuation pause. So, we make the fix a bit less |
ysr@777 | 2566 | // conservative and ensure that regions are pushed on the stack, |
ysr@777 | 2567 | // irrespective whether all collection set regions are below the |
ysr@777 | 2568 | // finger, if the region stack is not empty. This is expected to be |
ysr@777 | 2569 | // a rare case, so I don't think it's necessary to be smarted about it. |
ysr@777 | 2570 | if (!region_stack_empty()) |
ysr@777 | 2571 | _should_gray_objects = true; |
ysr@777 | 2572 | } |
ysr@777 | 2573 | |
ysr@777 | 2574 | void ConcurrentMark::registerCSetRegion(HeapRegion* hr) { |
ysr@777 | 2575 | if (!concurrent_marking_in_progress()) |
ysr@777 | 2576 | return; |
ysr@777 | 2577 | |
ysr@777 | 2578 | HeapWord* region_end = hr->end(); |
ysr@777 | 2579 | if (region_end > _min_finger) |
ysr@777 | 2580 | _should_gray_objects = true; |
ysr@777 | 2581 | } |
ysr@777 | 2582 | |
ysr@777 | 2583 | // abandon current marking iteration due to a Full GC |
ysr@777 | 2584 | void ConcurrentMark::abort() { |
ysr@777 | 2585 | // Clear all marks to force marking thread to do nothing |
ysr@777 | 2586 | _nextMarkBitMap->clearAll(); |
ysr@777 | 2587 | // Empty mark stack |
ysr@777 | 2588 | clear_marking_state(); |
ysr@777 | 2589 | for (int i = 0; i < (int)_max_task_num; ++i) |
ysr@777 | 2590 | _tasks[i]->clear_region_fields(); |
ysr@777 | 2591 | _has_aborted = true; |
ysr@777 | 2592 | |
ysr@777 | 2593 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 2594 | satb_mq_set.abandon_partial_marking(); |
ysr@777 | 2595 | satb_mq_set.set_active_all_threads(false); |
ysr@777 | 2596 | } |
ysr@777 | 2597 | |
ysr@777 | 2598 | static void print_ms_time_info(const char* prefix, const char* name, |
ysr@777 | 2599 | NumberSeq& ns) { |
ysr@777 | 2600 | gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", |
ysr@777 | 2601 | prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); |
ysr@777 | 2602 | if (ns.num() > 0) { |
ysr@777 | 2603 | gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", |
ysr@777 | 2604 | prefix, ns.sd(), ns.maximum()); |
ysr@777 | 2605 | } |
ysr@777 | 2606 | } |
ysr@777 | 2607 | |
ysr@777 | 2608 | void ConcurrentMark::print_summary_info() { |
ysr@777 | 2609 | gclog_or_tty->print_cr(" Concurrent marking:"); |
ysr@777 | 2610 | print_ms_time_info(" ", "init marks", _init_times); |
ysr@777 | 2611 | print_ms_time_info(" ", "remarks", _remark_times); |
ysr@777 | 2612 | { |
ysr@777 | 2613 | print_ms_time_info(" ", "final marks", _remark_mark_times); |
ysr@777 | 2614 | print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); |
ysr@777 | 2615 | |
ysr@777 | 2616 | } |
ysr@777 | 2617 | print_ms_time_info(" ", "cleanups", _cleanup_times); |
ysr@777 | 2618 | gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", |
ysr@777 | 2619 | _total_counting_time, |
ysr@777 | 2620 | (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / |
ysr@777 | 2621 | (double)_cleanup_times.num() |
ysr@777 | 2622 | : 0.0)); |
ysr@777 | 2623 | if (G1ScrubRemSets) { |
ysr@777 | 2624 | gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", |
ysr@777 | 2625 | _total_rs_scrub_time, |
ysr@777 | 2626 | (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / |
ysr@777 | 2627 | (double)_cleanup_times.num() |
ysr@777 | 2628 | : 0.0)); |
ysr@777 | 2629 | } |
ysr@777 | 2630 | gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", |
ysr@777 | 2631 | (_init_times.sum() + _remark_times.sum() + |
ysr@777 | 2632 | _cleanup_times.sum())/1000.0); |
ysr@777 | 2633 | gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " |
ysr@777 | 2634 | "(%8.2f s marking, %8.2f s counting).", |
ysr@777 | 2635 | cmThread()->vtime_accum(), |
ysr@777 | 2636 | cmThread()->vtime_mark_accum(), |
ysr@777 | 2637 | cmThread()->vtime_count_accum()); |
ysr@777 | 2638 | } |
ysr@777 | 2639 | |
tonyp@1454 | 2640 | void ConcurrentMark::print_worker_threads_on(outputStream* st) const { |
tonyp@1454 | 2641 | _parallel_workers->print_worker_threads_on(st); |
tonyp@1454 | 2642 | } |
tonyp@1454 | 2643 | |
ysr@777 | 2644 | // Closures |
ysr@777 | 2645 | // XXX: there seems to be a lot of code duplication here; |
ysr@777 | 2646 | // should refactor and consolidate the shared code. |
ysr@777 | 2647 | |
ysr@777 | 2648 | // This closure is used to mark refs into the CMS generation in |
ysr@777 | 2649 | // the CMS bit map. Called at the first checkpoint. |
ysr@777 | 2650 | |
ysr@777 | 2651 | // We take a break if someone is trying to stop the world. |
ysr@777 | 2652 | bool ConcurrentMark::do_yield_check(int worker_i) { |
ysr@777 | 2653 | if (should_yield()) { |
ysr@777 | 2654 | if (worker_i == 0) |
ysr@777 | 2655 | _g1h->g1_policy()->record_concurrent_pause(); |
ysr@777 | 2656 | cmThread()->yield(); |
ysr@777 | 2657 | if (worker_i == 0) |
ysr@777 | 2658 | _g1h->g1_policy()->record_concurrent_pause_end(); |
ysr@777 | 2659 | return true; |
ysr@777 | 2660 | } else { |
ysr@777 | 2661 | return false; |
ysr@777 | 2662 | } |
ysr@777 | 2663 | } |
ysr@777 | 2664 | |
ysr@777 | 2665 | bool ConcurrentMark::should_yield() { |
ysr@777 | 2666 | return cmThread()->should_yield(); |
ysr@777 | 2667 | } |
ysr@777 | 2668 | |
ysr@777 | 2669 | bool ConcurrentMark::containing_card_is_marked(void* p) { |
ysr@777 | 2670 | size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1); |
ysr@777 | 2671 | return _card_bm.at(offset >> CardTableModRefBS::card_shift); |
ysr@777 | 2672 | } |
ysr@777 | 2673 | |
ysr@777 | 2674 | bool ConcurrentMark::containing_cards_are_marked(void* start, |
ysr@777 | 2675 | void* last) { |
ysr@777 | 2676 | return |
ysr@777 | 2677 | containing_card_is_marked(start) && |
ysr@777 | 2678 | containing_card_is_marked(last); |
ysr@777 | 2679 | } |
ysr@777 | 2680 | |
ysr@777 | 2681 | #ifndef PRODUCT |
ysr@777 | 2682 | // for debugging purposes |
ysr@777 | 2683 | void ConcurrentMark::print_finger() { |
ysr@777 | 2684 | gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, |
ysr@777 | 2685 | _heap_start, _heap_end, _finger); |
ysr@777 | 2686 | for (int i = 0; i < (int) _max_task_num; ++i) { |
ysr@777 | 2687 | gclog_or_tty->print(" %d: "PTR_FORMAT, i, _tasks[i]->finger()); |
ysr@777 | 2688 | } |
ysr@777 | 2689 | gclog_or_tty->print_cr(""); |
ysr@777 | 2690 | } |
ysr@777 | 2691 | #endif |
ysr@777 | 2692 | |
ysr@777 | 2693 | // Closure for iteration over bitmaps |
ysr@777 | 2694 | class CMBitMapClosure : public BitMapClosure { |
ysr@777 | 2695 | private: |
ysr@777 | 2696 | // the bitmap that is being iterated over |
ysr@777 | 2697 | CMBitMap* _nextMarkBitMap; |
ysr@777 | 2698 | ConcurrentMark* _cm; |
ysr@777 | 2699 | CMTask* _task; |
ysr@777 | 2700 | // true if we're scanning a heap region claimed by the task (so that |
ysr@777 | 2701 | // we move the finger along), false if we're not, i.e. currently when |
ysr@777 | 2702 | // scanning a heap region popped from the region stack (so that we |
ysr@777 | 2703 | // do not move the task finger along; it'd be a mistake if we did so). |
ysr@777 | 2704 | bool _scanning_heap_region; |
ysr@777 | 2705 | |
ysr@777 | 2706 | public: |
ysr@777 | 2707 | CMBitMapClosure(CMTask *task, |
ysr@777 | 2708 | ConcurrentMark* cm, |
ysr@777 | 2709 | CMBitMap* nextMarkBitMap) |
ysr@777 | 2710 | : _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } |
ysr@777 | 2711 | |
ysr@777 | 2712 | void set_scanning_heap_region(bool scanning_heap_region) { |
ysr@777 | 2713 | _scanning_heap_region = scanning_heap_region; |
ysr@777 | 2714 | } |
ysr@777 | 2715 | |
ysr@777 | 2716 | bool do_bit(size_t offset) { |
ysr@777 | 2717 | HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); |
ysr@777 | 2718 | tmp_guarantee_CM( _nextMarkBitMap->isMarked(addr), "invariant" ); |
ysr@777 | 2719 | tmp_guarantee_CM( addr < _cm->finger(), "invariant" ); |
ysr@777 | 2720 | |
ysr@777 | 2721 | if (_scanning_heap_region) { |
ysr@777 | 2722 | statsOnly( _task->increase_objs_found_on_bitmap() ); |
ysr@777 | 2723 | tmp_guarantee_CM( addr >= _task->finger(), "invariant" ); |
ysr@777 | 2724 | // We move that task's local finger along. |
ysr@777 | 2725 | _task->move_finger_to(addr); |
ysr@777 | 2726 | } else { |
ysr@777 | 2727 | // We move the task's region finger along. |
ysr@777 | 2728 | _task->move_region_finger_to(addr); |
ysr@777 | 2729 | } |
ysr@777 | 2730 | |
ysr@777 | 2731 | _task->scan_object(oop(addr)); |
ysr@777 | 2732 | // we only partially drain the local queue and global stack |
ysr@777 | 2733 | _task->drain_local_queue(true); |
ysr@777 | 2734 | _task->drain_global_stack(true); |
ysr@777 | 2735 | |
ysr@777 | 2736 | // if the has_aborted flag has been raised, we need to bail out of |
ysr@777 | 2737 | // the iteration |
ysr@777 | 2738 | return !_task->has_aborted(); |
ysr@777 | 2739 | } |
ysr@777 | 2740 | }; |
ysr@777 | 2741 | |
ysr@777 | 2742 | // Closure for iterating over objects, currently only used for |
ysr@777 | 2743 | // processing SATB buffers. |
ysr@777 | 2744 | class CMObjectClosure : public ObjectClosure { |
ysr@777 | 2745 | private: |
ysr@777 | 2746 | CMTask* _task; |
ysr@777 | 2747 | |
ysr@777 | 2748 | public: |
ysr@777 | 2749 | void do_object(oop obj) { |
ysr@777 | 2750 | _task->deal_with_reference(obj); |
ysr@777 | 2751 | } |
ysr@777 | 2752 | |
ysr@777 | 2753 | CMObjectClosure(CMTask* task) : _task(task) { } |
ysr@777 | 2754 | }; |
ysr@777 | 2755 | |
ysr@777 | 2756 | // Closure for iterating over object fields |
ysr@777 | 2757 | class CMOopClosure : public OopClosure { |
ysr@777 | 2758 | private: |
ysr@777 | 2759 | G1CollectedHeap* _g1h; |
ysr@777 | 2760 | ConcurrentMark* _cm; |
ysr@777 | 2761 | CMTask* _task; |
ysr@777 | 2762 | |
ysr@777 | 2763 | public: |
ysr@1280 | 2764 | virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
ysr@1280 | 2765 | virtual void do_oop( oop* p) { do_oop_work(p); } |
ysr@1280 | 2766 | |
ysr@1280 | 2767 | template <class T> void do_oop_work(T* p) { |
ysr@777 | 2768 | tmp_guarantee_CM( _g1h->is_in_g1_reserved((HeapWord*) p), "invariant" ); |
ysr@1280 | 2769 | tmp_guarantee_CM( !_g1h->heap_region_containing((HeapWord*) p)->is_on_free_list(), "invariant" ); |
ysr@1280 | 2770 | |
ysr@1280 | 2771 | oop obj = oopDesc::load_decode_heap_oop(p); |
ysr@777 | 2772 | if (_cm->verbose_high()) |
ysr@777 | 2773 | gclog_or_tty->print_cr("[%d] we're looking at location " |
ysr@777 | 2774 | "*"PTR_FORMAT" = "PTR_FORMAT, |
ysr@777 | 2775 | _task->task_id(), p, (void*) obj); |
ysr@777 | 2776 | _task->deal_with_reference(obj); |
ysr@777 | 2777 | } |
ysr@777 | 2778 | |
ysr@777 | 2779 | CMOopClosure(G1CollectedHeap* g1h, |
ysr@777 | 2780 | ConcurrentMark* cm, |
ysr@777 | 2781 | CMTask* task) |
ysr@777 | 2782 | : _g1h(g1h), _cm(cm), _task(task) { } |
ysr@777 | 2783 | }; |
ysr@777 | 2784 | |
ysr@777 | 2785 | void CMTask::setup_for_region(HeapRegion* hr) { |
ysr@777 | 2786 | tmp_guarantee_CM( hr != NULL && !hr->continuesHumongous(), |
ysr@777 | 2787 | "claim_region() should have filtered out continues humongous regions" ); |
ysr@777 | 2788 | |
ysr@777 | 2789 | if (_cm->verbose_low()) |
ysr@777 | 2790 | gclog_or_tty->print_cr("[%d] setting up for region "PTR_FORMAT, |
ysr@777 | 2791 | _task_id, hr); |
ysr@777 | 2792 | |
ysr@777 | 2793 | _curr_region = hr; |
ysr@777 | 2794 | _finger = hr->bottom(); |
ysr@777 | 2795 | update_region_limit(); |
ysr@777 | 2796 | } |
ysr@777 | 2797 | |
ysr@777 | 2798 | void CMTask::update_region_limit() { |
ysr@777 | 2799 | HeapRegion* hr = _curr_region; |
ysr@777 | 2800 | HeapWord* bottom = hr->bottom(); |
ysr@777 | 2801 | HeapWord* limit = hr->next_top_at_mark_start(); |
ysr@777 | 2802 | |
ysr@777 | 2803 | if (limit == bottom) { |
ysr@777 | 2804 | if (_cm->verbose_low()) |
ysr@777 | 2805 | gclog_or_tty->print_cr("[%d] found an empty region " |
ysr@777 | 2806 | "["PTR_FORMAT", "PTR_FORMAT")", |
ysr@777 | 2807 | _task_id, bottom, limit); |
ysr@777 | 2808 | // The region was collected underneath our feet. |
ysr@777 | 2809 | // We set the finger to bottom to ensure that the bitmap |
ysr@777 | 2810 | // iteration that will follow this will not do anything. |
ysr@777 | 2811 | // (this is not a condition that holds when we set the region up, |
ysr@777 | 2812 | // as the region is not supposed to be empty in the first place) |
ysr@777 | 2813 | _finger = bottom; |
ysr@777 | 2814 | } else if (limit >= _region_limit) { |
ysr@777 | 2815 | tmp_guarantee_CM( limit >= _finger, "peace of mind" ); |
ysr@777 | 2816 | } else { |
ysr@777 | 2817 | tmp_guarantee_CM( limit < _region_limit, "only way to get here" ); |
ysr@777 | 2818 | // This can happen under some pretty unusual circumstances. An |
ysr@777 | 2819 | // evacuation pause empties the region underneath our feet (NTAMS |
ysr@777 | 2820 | // at bottom). We then do some allocation in the region (NTAMS |
ysr@777 | 2821 | // stays at bottom), followed by the region being used as a GC |
ysr@777 | 2822 | // alloc region (NTAMS will move to top() and the objects |
ysr@777 | 2823 | // originally below it will be grayed). All objects now marked in |
ysr@777 | 2824 | // the region are explicitly grayed, if below the global finger, |
ysr@777 | 2825 | // and we do not need in fact to scan anything else. So, we simply |
ysr@777 | 2826 | // set _finger to be limit to ensure that the bitmap iteration |
ysr@777 | 2827 | // doesn't do anything. |
ysr@777 | 2828 | _finger = limit; |
ysr@777 | 2829 | } |
ysr@777 | 2830 | |
ysr@777 | 2831 | _region_limit = limit; |
ysr@777 | 2832 | } |
ysr@777 | 2833 | |
ysr@777 | 2834 | void CMTask::giveup_current_region() { |
ysr@777 | 2835 | tmp_guarantee_CM( _curr_region != NULL, "invariant" ); |
ysr@777 | 2836 | if (_cm->verbose_low()) |
ysr@777 | 2837 | gclog_or_tty->print_cr("[%d] giving up region "PTR_FORMAT, |
ysr@777 | 2838 | _task_id, _curr_region); |
ysr@777 | 2839 | clear_region_fields(); |
ysr@777 | 2840 | } |
ysr@777 | 2841 | |
ysr@777 | 2842 | void CMTask::clear_region_fields() { |
ysr@777 | 2843 | // Values for these three fields that indicate that we're not |
ysr@777 | 2844 | // holding on to a region. |
ysr@777 | 2845 | _curr_region = NULL; |
ysr@777 | 2846 | _finger = NULL; |
ysr@777 | 2847 | _region_limit = NULL; |
ysr@777 | 2848 | |
ysr@777 | 2849 | _region_finger = NULL; |
ysr@777 | 2850 | } |
ysr@777 | 2851 | |
ysr@777 | 2852 | void CMTask::reset(CMBitMap* nextMarkBitMap) { |
ysr@777 | 2853 | guarantee( nextMarkBitMap != NULL, "invariant" ); |
ysr@777 | 2854 | |
ysr@777 | 2855 | if (_cm->verbose_low()) |
ysr@777 | 2856 | gclog_or_tty->print_cr("[%d] resetting", _task_id); |
ysr@777 | 2857 | |
ysr@777 | 2858 | _nextMarkBitMap = nextMarkBitMap; |
ysr@777 | 2859 | clear_region_fields(); |
ysr@777 | 2860 | |
ysr@777 | 2861 | _calls = 0; |
ysr@777 | 2862 | _elapsed_time_ms = 0.0; |
ysr@777 | 2863 | _termination_time_ms = 0.0; |
ysr@777 | 2864 | _termination_start_time_ms = 0.0; |
ysr@777 | 2865 | |
ysr@777 | 2866 | #if _MARKING_STATS_ |
ysr@777 | 2867 | _local_pushes = 0; |
ysr@777 | 2868 | _local_pops = 0; |
ysr@777 | 2869 | _local_max_size = 0; |
ysr@777 | 2870 | _objs_scanned = 0; |
ysr@777 | 2871 | _global_pushes = 0; |
ysr@777 | 2872 | _global_pops = 0; |
ysr@777 | 2873 | _global_max_size = 0; |
ysr@777 | 2874 | _global_transfers_to = 0; |
ysr@777 | 2875 | _global_transfers_from = 0; |
ysr@777 | 2876 | _region_stack_pops = 0; |
ysr@777 | 2877 | _regions_claimed = 0; |
ysr@777 | 2878 | _objs_found_on_bitmap = 0; |
ysr@777 | 2879 | _satb_buffers_processed = 0; |
ysr@777 | 2880 | _steal_attempts = 0; |
ysr@777 | 2881 | _steals = 0; |
ysr@777 | 2882 | _aborted = 0; |
ysr@777 | 2883 | _aborted_overflow = 0; |
ysr@777 | 2884 | _aborted_cm_aborted = 0; |
ysr@777 | 2885 | _aborted_yield = 0; |
ysr@777 | 2886 | _aborted_timed_out = 0; |
ysr@777 | 2887 | _aborted_satb = 0; |
ysr@777 | 2888 | _aborted_termination = 0; |
ysr@777 | 2889 | #endif // _MARKING_STATS_ |
ysr@777 | 2890 | } |
ysr@777 | 2891 | |
ysr@777 | 2892 | bool CMTask::should_exit_termination() { |
ysr@777 | 2893 | regular_clock_call(); |
ysr@777 | 2894 | // This is called when we are in the termination protocol. We should |
ysr@777 | 2895 | // quit if, for some reason, this task wants to abort or the global |
ysr@777 | 2896 | // stack is not empty (this means that we can get work from it). |
ysr@777 | 2897 | return !_cm->mark_stack_empty() || has_aborted(); |
ysr@777 | 2898 | } |
ysr@777 | 2899 | |
ysr@777 | 2900 | // This determines whether the method below will check both the local |
ysr@777 | 2901 | // and global fingers when determining whether to push on the stack a |
ysr@777 | 2902 | // gray object (value 1) or whether it will only check the global one |
ysr@777 | 2903 | // (value 0). The tradeoffs are that the former will be a bit more |
ysr@777 | 2904 | // accurate and possibly push less on the stack, but it might also be |
ysr@777 | 2905 | // a little bit slower. |
ysr@777 | 2906 | |
ysr@777 | 2907 | #define _CHECK_BOTH_FINGERS_ 1 |
ysr@777 | 2908 | |
ysr@777 | 2909 | void CMTask::deal_with_reference(oop obj) { |
ysr@777 | 2910 | if (_cm->verbose_high()) |
ysr@777 | 2911 | gclog_or_tty->print_cr("[%d] we're dealing with reference = "PTR_FORMAT, |
ysr@777 | 2912 | _task_id, (void*) obj); |
ysr@777 | 2913 | |
ysr@777 | 2914 | ++_refs_reached; |
ysr@777 | 2915 | |
ysr@777 | 2916 | HeapWord* objAddr = (HeapWord*) obj; |
ysr@1280 | 2917 | assert(obj->is_oop_or_null(true /* ignore mark word */), "Error"); |
ysr@777 | 2918 | if (_g1h->is_in_g1_reserved(objAddr)) { |
ysr@777 | 2919 | tmp_guarantee_CM( obj != NULL, "is_in_g1_reserved should ensure this" ); |
ysr@777 | 2920 | HeapRegion* hr = _g1h->heap_region_containing(obj); |
ysr@777 | 2921 | if (_g1h->is_obj_ill(obj, hr)) { |
ysr@777 | 2922 | if (_cm->verbose_high()) |
ysr@777 | 2923 | gclog_or_tty->print_cr("[%d] "PTR_FORMAT" is not considered marked", |
ysr@777 | 2924 | _task_id, (void*) obj); |
ysr@777 | 2925 | |
ysr@777 | 2926 | // we need to mark it first |
ysr@777 | 2927 | if (_nextMarkBitMap->parMark(objAddr)) { |
ysr@777 | 2928 | // No OrderAccess:store_load() is needed. It is implicit in the |
ysr@777 | 2929 | // CAS done in parMark(objAddr) above |
ysr@777 | 2930 | HeapWord* global_finger = _cm->finger(); |
ysr@777 | 2931 | |
ysr@777 | 2932 | #if _CHECK_BOTH_FINGERS_ |
ysr@777 | 2933 | // we will check both the local and global fingers |
ysr@777 | 2934 | |
ysr@777 | 2935 | if (_finger != NULL && objAddr < _finger) { |
ysr@777 | 2936 | if (_cm->verbose_high()) |
ysr@777 | 2937 | gclog_or_tty->print_cr("[%d] below the local finger ("PTR_FORMAT"), " |
ysr@777 | 2938 | "pushing it", _task_id, _finger); |
ysr@777 | 2939 | push(obj); |
ysr@777 | 2940 | } else if (_curr_region != NULL && objAddr < _region_limit) { |
ysr@777 | 2941 | // do nothing |
ysr@777 | 2942 | } else if (objAddr < global_finger) { |
ysr@777 | 2943 | // Notice that the global finger might be moving forward |
ysr@777 | 2944 | // concurrently. This is not a problem. In the worst case, we |
ysr@777 | 2945 | // mark the object while it is above the global finger and, by |
ysr@777 | 2946 | // the time we read the global finger, it has moved forward |
ysr@777 | 2947 | // passed this object. In this case, the object will probably |
ysr@777 | 2948 | // be visited when a task is scanning the region and will also |
ysr@777 | 2949 | // be pushed on the stack. So, some duplicate work, but no |
ysr@777 | 2950 | // correctness problems. |
ysr@777 | 2951 | |
ysr@777 | 2952 | if (_cm->verbose_high()) |
ysr@777 | 2953 | gclog_or_tty->print_cr("[%d] below the global finger " |
ysr@777 | 2954 | "("PTR_FORMAT"), pushing it", |
ysr@777 | 2955 | _task_id, global_finger); |
ysr@777 | 2956 | push(obj); |
ysr@777 | 2957 | } else { |
ysr@777 | 2958 | // do nothing |
ysr@777 | 2959 | } |
ysr@777 | 2960 | #else // _CHECK_BOTH_FINGERS_ |
ysr@777 | 2961 | // we will only check the global finger |
ysr@777 | 2962 | |
ysr@777 | 2963 | if (objAddr < global_finger) { |
ysr@777 | 2964 | // see long comment above |
ysr@777 | 2965 | |
ysr@777 | 2966 | if (_cm->verbose_high()) |
ysr@777 | 2967 | gclog_or_tty->print_cr("[%d] below the global finger " |
ysr@777 | 2968 | "("PTR_FORMAT"), pushing it", |
ysr@777 | 2969 | _task_id, global_finger); |
ysr@777 | 2970 | push(obj); |
ysr@777 | 2971 | } |
ysr@777 | 2972 | #endif // _CHECK_BOTH_FINGERS_ |
ysr@777 | 2973 | } |
ysr@777 | 2974 | } |
ysr@777 | 2975 | } |
ysr@777 | 2976 | } |
ysr@777 | 2977 | |
ysr@777 | 2978 | void CMTask::push(oop obj) { |
ysr@777 | 2979 | HeapWord* objAddr = (HeapWord*) obj; |
ysr@777 | 2980 | tmp_guarantee_CM( _g1h->is_in_g1_reserved(objAddr), "invariant" ); |
ysr@1280 | 2981 | tmp_guarantee_CM( !_g1h->heap_region_containing(objAddr)->is_on_free_list(), "invariant" ); |
ysr@777 | 2982 | tmp_guarantee_CM( !_g1h->is_obj_ill(obj), "invariant" ); |
ysr@777 | 2983 | tmp_guarantee_CM( _nextMarkBitMap->isMarked(objAddr), "invariant" ); |
ysr@777 | 2984 | |
ysr@777 | 2985 | if (_cm->verbose_high()) |
ysr@777 | 2986 | gclog_or_tty->print_cr("[%d] pushing "PTR_FORMAT, _task_id, (void*) obj); |
ysr@777 | 2987 | |
ysr@777 | 2988 | if (!_task_queue->push(obj)) { |
ysr@777 | 2989 | // The local task queue looks full. We need to push some entries |
ysr@777 | 2990 | // to the global stack. |
ysr@777 | 2991 | |
ysr@777 | 2992 | if (_cm->verbose_medium()) |
ysr@777 | 2993 | gclog_or_tty->print_cr("[%d] task queue overflow, " |
ysr@777 | 2994 | "moving entries to the global stack", |
ysr@777 | 2995 | _task_id); |
ysr@777 | 2996 | move_entries_to_global_stack(); |
ysr@777 | 2997 | |
ysr@777 | 2998 | // this should succeed since, even if we overflow the global |
ysr@777 | 2999 | // stack, we should have definitely removed some entries from the |
ysr@777 | 3000 | // local queue. So, there must be space on it. |
ysr@777 | 3001 | bool success = _task_queue->push(obj); |
ysr@777 | 3002 | tmp_guarantee_CM( success, "invariant" ); |
ysr@777 | 3003 | } |
ysr@777 | 3004 | |
ysr@777 | 3005 | statsOnly( int tmp_size = _task_queue->size(); |
ysr@777 | 3006 | if (tmp_size > _local_max_size) |
ysr@777 | 3007 | _local_max_size = tmp_size; |
ysr@777 | 3008 | ++_local_pushes ); |
ysr@777 | 3009 | } |
ysr@777 | 3010 | |
ysr@777 | 3011 | void CMTask::reached_limit() { |
ysr@777 | 3012 | tmp_guarantee_CM( _words_scanned >= _words_scanned_limit || |
ysr@777 | 3013 | _refs_reached >= _refs_reached_limit , |
ysr@777 | 3014 | "shouldn't have been called otherwise" ); |
ysr@777 | 3015 | regular_clock_call(); |
ysr@777 | 3016 | } |
ysr@777 | 3017 | |
ysr@777 | 3018 | void CMTask::regular_clock_call() { |
ysr@777 | 3019 | if (has_aborted()) |
ysr@777 | 3020 | return; |
ysr@777 | 3021 | |
ysr@777 | 3022 | // First, we need to recalculate the words scanned and refs reached |
ysr@777 | 3023 | // limits for the next clock call. |
ysr@777 | 3024 | recalculate_limits(); |
ysr@777 | 3025 | |
ysr@777 | 3026 | // During the regular clock call we do the following |
ysr@777 | 3027 | |
ysr@777 | 3028 | // (1) If an overflow has been flagged, then we abort. |
ysr@777 | 3029 | if (_cm->has_overflown()) { |
ysr@777 | 3030 | set_has_aborted(); |
ysr@777 | 3031 | return; |
ysr@777 | 3032 | } |
ysr@777 | 3033 | |
ysr@777 | 3034 | // If we are not concurrent (i.e. we're doing remark) we don't need |
ysr@777 | 3035 | // to check anything else. The other steps are only needed during |
ysr@777 | 3036 | // the concurrent marking phase. |
ysr@777 | 3037 | if (!concurrent()) |
ysr@777 | 3038 | return; |
ysr@777 | 3039 | |
ysr@777 | 3040 | // (2) If marking has been aborted for Full GC, then we also abort. |
ysr@777 | 3041 | if (_cm->has_aborted()) { |
ysr@777 | 3042 | set_has_aborted(); |
ysr@777 | 3043 | statsOnly( ++_aborted_cm_aborted ); |
ysr@777 | 3044 | return; |
ysr@777 | 3045 | } |
ysr@777 | 3046 | |
ysr@777 | 3047 | double curr_time_ms = os::elapsedVTime() * 1000.0; |
ysr@777 | 3048 | |
ysr@777 | 3049 | // (3) If marking stats are enabled, then we update the step history. |
ysr@777 | 3050 | #if _MARKING_STATS_ |
ysr@777 | 3051 | if (_words_scanned >= _words_scanned_limit) |
ysr@777 | 3052 | ++_clock_due_to_scanning; |
ysr@777 | 3053 | if (_refs_reached >= _refs_reached_limit) |
ysr@777 | 3054 | ++_clock_due_to_marking; |
ysr@777 | 3055 | |
ysr@777 | 3056 | double last_interval_ms = curr_time_ms - _interval_start_time_ms; |
ysr@777 | 3057 | _interval_start_time_ms = curr_time_ms; |
ysr@777 | 3058 | _all_clock_intervals_ms.add(last_interval_ms); |
ysr@777 | 3059 | |
ysr@777 | 3060 | if (_cm->verbose_medium()) { |
ysr@777 | 3061 | gclog_or_tty->print_cr("[%d] regular clock, interval = %1.2lfms, " |
ysr@777 | 3062 | "scanned = %d%s, refs reached = %d%s", |
ysr@777 | 3063 | _task_id, last_interval_ms, |
ysr@777 | 3064 | _words_scanned, |
ysr@777 | 3065 | (_words_scanned >= _words_scanned_limit) ? " (*)" : "", |
ysr@777 | 3066 | _refs_reached, |
ysr@777 | 3067 | (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); |
ysr@777 | 3068 | } |
ysr@777 | 3069 | #endif // _MARKING_STATS_ |
ysr@777 | 3070 | |
ysr@777 | 3071 | // (4) We check whether we should yield. If we have to, then we abort. |
ysr@777 | 3072 | if (_cm->should_yield()) { |
ysr@777 | 3073 | // We should yield. To do this we abort the task. The caller is |
ysr@777 | 3074 | // responsible for yielding. |
ysr@777 | 3075 | set_has_aborted(); |
ysr@777 | 3076 | statsOnly( ++_aborted_yield ); |
ysr@777 | 3077 | return; |
ysr@777 | 3078 | } |
ysr@777 | 3079 | |
ysr@777 | 3080 | // (5) We check whether we've reached our time quota. If we have, |
ysr@777 | 3081 | // then we abort. |
ysr@777 | 3082 | double elapsed_time_ms = curr_time_ms - _start_time_ms; |
ysr@777 | 3083 | if (elapsed_time_ms > _time_target_ms) { |
ysr@777 | 3084 | set_has_aborted(); |
ysr@777 | 3085 | _has_aborted_timed_out = true; |
ysr@777 | 3086 | statsOnly( ++_aborted_timed_out ); |
ysr@777 | 3087 | return; |
ysr@777 | 3088 | } |
ysr@777 | 3089 | |
ysr@777 | 3090 | // (6) Finally, we check whether there are enough completed STAB |
ysr@777 | 3091 | // buffers available for processing. If there are, we abort. |
ysr@777 | 3092 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 3093 | if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { |
ysr@777 | 3094 | if (_cm->verbose_low()) |
ysr@777 | 3095 | gclog_or_tty->print_cr("[%d] aborting to deal with pending SATB buffers", |
ysr@777 | 3096 | _task_id); |
ysr@777 | 3097 | // we do need to process SATB buffers, we'll abort and restart |
ysr@777 | 3098 | // the marking task to do so |
ysr@777 | 3099 | set_has_aborted(); |
ysr@777 | 3100 | statsOnly( ++_aborted_satb ); |
ysr@777 | 3101 | return; |
ysr@777 | 3102 | } |
ysr@777 | 3103 | } |
ysr@777 | 3104 | |
ysr@777 | 3105 | void CMTask::recalculate_limits() { |
ysr@777 | 3106 | _real_words_scanned_limit = _words_scanned + words_scanned_period; |
ysr@777 | 3107 | _words_scanned_limit = _real_words_scanned_limit; |
ysr@777 | 3108 | |
ysr@777 | 3109 | _real_refs_reached_limit = _refs_reached + refs_reached_period; |
ysr@777 | 3110 | _refs_reached_limit = _real_refs_reached_limit; |
ysr@777 | 3111 | } |
ysr@777 | 3112 | |
ysr@777 | 3113 | void CMTask::decrease_limits() { |
ysr@777 | 3114 | // This is called when we believe that we're going to do an infrequent |
ysr@777 | 3115 | // operation which will increase the per byte scanned cost (i.e. move |
ysr@777 | 3116 | // entries to/from the global stack). It basically tries to decrease the |
ysr@777 | 3117 | // scanning limit so that the clock is called earlier. |
ysr@777 | 3118 | |
ysr@777 | 3119 | if (_cm->verbose_medium()) |
ysr@777 | 3120 | gclog_or_tty->print_cr("[%d] decreasing limits", _task_id); |
ysr@777 | 3121 | |
ysr@777 | 3122 | _words_scanned_limit = _real_words_scanned_limit - |
ysr@777 | 3123 | 3 * words_scanned_period / 4; |
ysr@777 | 3124 | _refs_reached_limit = _real_refs_reached_limit - |
ysr@777 | 3125 | 3 * refs_reached_period / 4; |
ysr@777 | 3126 | } |
ysr@777 | 3127 | |
ysr@777 | 3128 | void CMTask::move_entries_to_global_stack() { |
ysr@777 | 3129 | // local array where we'll store the entries that will be popped |
ysr@777 | 3130 | // from the local queue |
ysr@777 | 3131 | oop buffer[global_stack_transfer_size]; |
ysr@777 | 3132 | |
ysr@777 | 3133 | int n = 0; |
ysr@777 | 3134 | oop obj; |
ysr@777 | 3135 | while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { |
ysr@777 | 3136 | buffer[n] = obj; |
ysr@777 | 3137 | ++n; |
ysr@777 | 3138 | } |
ysr@777 | 3139 | |
ysr@777 | 3140 | if (n > 0) { |
ysr@777 | 3141 | // we popped at least one entry from the local queue |
ysr@777 | 3142 | |
ysr@777 | 3143 | statsOnly( ++_global_transfers_to; _local_pops += n ); |
ysr@777 | 3144 | |
ysr@777 | 3145 | if (!_cm->mark_stack_push(buffer, n)) { |
ysr@777 | 3146 | if (_cm->verbose_low()) |
ysr@777 | 3147 | gclog_or_tty->print_cr("[%d] aborting due to global stack overflow", _task_id); |
ysr@777 | 3148 | set_has_aborted(); |
ysr@777 | 3149 | } else { |
ysr@777 | 3150 | // the transfer was successful |
ysr@777 | 3151 | |
ysr@777 | 3152 | if (_cm->verbose_medium()) |
ysr@777 | 3153 | gclog_or_tty->print_cr("[%d] pushed %d entries to the global stack", |
ysr@777 | 3154 | _task_id, n); |
ysr@777 | 3155 | statsOnly( int tmp_size = _cm->mark_stack_size(); |
ysr@777 | 3156 | if (tmp_size > _global_max_size) |
ysr@777 | 3157 | _global_max_size = tmp_size; |
ysr@777 | 3158 | _global_pushes += n ); |
ysr@777 | 3159 | } |
ysr@777 | 3160 | } |
ysr@777 | 3161 | |
ysr@777 | 3162 | // this operation was quite expensive, so decrease the limits |
ysr@777 | 3163 | decrease_limits(); |
ysr@777 | 3164 | } |
ysr@777 | 3165 | |
ysr@777 | 3166 | void CMTask::get_entries_from_global_stack() { |
ysr@777 | 3167 | // local array where we'll store the entries that will be popped |
ysr@777 | 3168 | // from the global stack. |
ysr@777 | 3169 | oop buffer[global_stack_transfer_size]; |
ysr@777 | 3170 | int n; |
ysr@777 | 3171 | _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); |
ysr@777 | 3172 | tmp_guarantee_CM( n <= global_stack_transfer_size, |
ysr@777 | 3173 | "we should not pop more than the given limit" ); |
ysr@777 | 3174 | if (n > 0) { |
ysr@777 | 3175 | // yes, we did actually pop at least one entry |
ysr@777 | 3176 | |
ysr@777 | 3177 | statsOnly( ++_global_transfers_from; _global_pops += n ); |
ysr@777 | 3178 | if (_cm->verbose_medium()) |
ysr@777 | 3179 | gclog_or_tty->print_cr("[%d] popped %d entries from the global stack", |
ysr@777 | 3180 | _task_id, n); |
ysr@777 | 3181 | for (int i = 0; i < n; ++i) { |
ysr@777 | 3182 | bool success = _task_queue->push(buffer[i]); |
ysr@777 | 3183 | // We only call this when the local queue is empty or under a |
ysr@777 | 3184 | // given target limit. So, we do not expect this push to fail. |
ysr@777 | 3185 | tmp_guarantee_CM( success, "invariant" ); |
ysr@777 | 3186 | } |
ysr@777 | 3187 | |
ysr@777 | 3188 | statsOnly( int tmp_size = _task_queue->size(); |
ysr@777 | 3189 | if (tmp_size > _local_max_size) |
ysr@777 | 3190 | _local_max_size = tmp_size; |
ysr@777 | 3191 | _local_pushes += n ); |
ysr@777 | 3192 | } |
ysr@777 | 3193 | |
ysr@777 | 3194 | // this operation was quite expensive, so decrease the limits |
ysr@777 | 3195 | decrease_limits(); |
ysr@777 | 3196 | } |
ysr@777 | 3197 | |
ysr@777 | 3198 | void CMTask::drain_local_queue(bool partially) { |
ysr@777 | 3199 | if (has_aborted()) |
ysr@777 | 3200 | return; |
ysr@777 | 3201 | |
ysr@777 | 3202 | // Decide what the target size is, depending whether we're going to |
ysr@777 | 3203 | // drain it partially (so that other tasks can steal if they run out |
ysr@777 | 3204 | // of things to do) or totally (at the very end). |
ysr@777 | 3205 | size_t target_size; |
ysr@777 | 3206 | if (partially) |
ysr@777 | 3207 | target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); |
ysr@777 | 3208 | else |
ysr@777 | 3209 | target_size = 0; |
ysr@777 | 3210 | |
ysr@777 | 3211 | if (_task_queue->size() > target_size) { |
ysr@777 | 3212 | if (_cm->verbose_high()) |
ysr@777 | 3213 | gclog_or_tty->print_cr("[%d] draining local queue, target size = %d", |
ysr@777 | 3214 | _task_id, target_size); |
ysr@777 | 3215 | |
ysr@777 | 3216 | oop obj; |
ysr@777 | 3217 | bool ret = _task_queue->pop_local(obj); |
ysr@777 | 3218 | while (ret) { |
ysr@777 | 3219 | statsOnly( ++_local_pops ); |
ysr@777 | 3220 | |
ysr@777 | 3221 | if (_cm->verbose_high()) |
ysr@777 | 3222 | gclog_or_tty->print_cr("[%d] popped "PTR_FORMAT, _task_id, |
ysr@777 | 3223 | (void*) obj); |
ysr@777 | 3224 | |
ysr@777 | 3225 | tmp_guarantee_CM( _g1h->is_in_g1_reserved((HeapWord*) obj), |
ysr@777 | 3226 | "invariant" ); |
ysr@1280 | 3227 | tmp_guarantee_CM( !_g1h->heap_region_containing(obj)->is_on_free_list(), |
ysr@1280 | 3228 | "invariant" ); |
ysr@777 | 3229 | |
ysr@777 | 3230 | scan_object(obj); |
ysr@777 | 3231 | |
ysr@777 | 3232 | if (_task_queue->size() <= target_size || has_aborted()) |
ysr@777 | 3233 | ret = false; |
ysr@777 | 3234 | else |
ysr@777 | 3235 | ret = _task_queue->pop_local(obj); |
ysr@777 | 3236 | } |
ysr@777 | 3237 | |
ysr@777 | 3238 | if (_cm->verbose_high()) |
ysr@777 | 3239 | gclog_or_tty->print_cr("[%d] drained local queue, size = %d", |
ysr@777 | 3240 | _task_id, _task_queue->size()); |
ysr@777 | 3241 | } |
ysr@777 | 3242 | } |
ysr@777 | 3243 | |
ysr@777 | 3244 | void CMTask::drain_global_stack(bool partially) { |
ysr@777 | 3245 | if (has_aborted()) |
ysr@777 | 3246 | return; |
ysr@777 | 3247 | |
ysr@777 | 3248 | // We have a policy to drain the local queue before we attempt to |
ysr@777 | 3249 | // drain the global stack. |
ysr@777 | 3250 | tmp_guarantee_CM( partially || _task_queue->size() == 0, "invariant" ); |
ysr@777 | 3251 | |
ysr@777 | 3252 | // Decide what the target size is, depending whether we're going to |
ysr@777 | 3253 | // drain it partially (so that other tasks can steal if they run out |
ysr@777 | 3254 | // of things to do) or totally (at the very end). Notice that, |
ysr@777 | 3255 | // because we move entries from the global stack in chunks or |
ysr@777 | 3256 | // because another task might be doing the same, we might in fact |
ysr@777 | 3257 | // drop below the target. But, this is not a problem. |
ysr@777 | 3258 | size_t target_size; |
ysr@777 | 3259 | if (partially) |
ysr@777 | 3260 | target_size = _cm->partial_mark_stack_size_target(); |
ysr@777 | 3261 | else |
ysr@777 | 3262 | target_size = 0; |
ysr@777 | 3263 | |
ysr@777 | 3264 | if (_cm->mark_stack_size() > target_size) { |
ysr@777 | 3265 | if (_cm->verbose_low()) |
ysr@777 | 3266 | gclog_or_tty->print_cr("[%d] draining global_stack, target size %d", |
ysr@777 | 3267 | _task_id, target_size); |
ysr@777 | 3268 | |
ysr@777 | 3269 | while (!has_aborted() && _cm->mark_stack_size() > target_size) { |
ysr@777 | 3270 | get_entries_from_global_stack(); |
ysr@777 | 3271 | drain_local_queue(partially); |
ysr@777 | 3272 | } |
ysr@777 | 3273 | |
ysr@777 | 3274 | if (_cm->verbose_low()) |
ysr@777 | 3275 | gclog_or_tty->print_cr("[%d] drained global stack, size = %d", |
ysr@777 | 3276 | _task_id, _cm->mark_stack_size()); |
ysr@777 | 3277 | } |
ysr@777 | 3278 | } |
ysr@777 | 3279 | |
ysr@777 | 3280 | // SATB Queue has several assumptions on whether to call the par or |
ysr@777 | 3281 | // non-par versions of the methods. this is why some of the code is |
ysr@777 | 3282 | // replicated. We should really get rid of the single-threaded version |
ysr@777 | 3283 | // of the code to simplify things. |
ysr@777 | 3284 | void CMTask::drain_satb_buffers() { |
ysr@777 | 3285 | if (has_aborted()) |
ysr@777 | 3286 | return; |
ysr@777 | 3287 | |
ysr@777 | 3288 | // We set this so that the regular clock knows that we're in the |
ysr@777 | 3289 | // middle of draining buffers and doesn't set the abort flag when it |
ysr@777 | 3290 | // notices that SATB buffers are available for draining. It'd be |
ysr@777 | 3291 | // very counter productive if it did that. :-) |
ysr@777 | 3292 | _draining_satb_buffers = true; |
ysr@777 | 3293 | |
ysr@777 | 3294 | CMObjectClosure oc(this); |
ysr@777 | 3295 | SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
ysr@777 | 3296 | if (ParallelGCThreads > 0) |
ysr@777 | 3297 | satb_mq_set.set_par_closure(_task_id, &oc); |
ysr@777 | 3298 | else |
ysr@777 | 3299 | satb_mq_set.set_closure(&oc); |
ysr@777 | 3300 | |
ysr@777 | 3301 | // This keeps claiming and applying the closure to completed buffers |
ysr@777 | 3302 | // until we run out of buffers or we need to abort. |
ysr@777 | 3303 | if (ParallelGCThreads > 0) { |
ysr@777 | 3304 | while (!has_aborted() && |
ysr@777 | 3305 | satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) { |
ysr@777 | 3306 | if (_cm->verbose_medium()) |
ysr@777 | 3307 | gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); |
ysr@777 | 3308 | statsOnly( ++_satb_buffers_processed ); |
ysr@777 | 3309 | regular_clock_call(); |
ysr@777 | 3310 | } |
ysr@777 | 3311 | } else { |
ysr@777 | 3312 | while (!has_aborted() && |
ysr@777 | 3313 | satb_mq_set.apply_closure_to_completed_buffer()) { |
ysr@777 | 3314 | if (_cm->verbose_medium()) |
ysr@777 | 3315 | gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); |
ysr@777 | 3316 | statsOnly( ++_satb_buffers_processed ); |
ysr@777 | 3317 | regular_clock_call(); |
ysr@777 | 3318 | } |
ysr@777 | 3319 | } |
ysr@777 | 3320 | |
ysr@777 | 3321 | if (!concurrent() && !has_aborted()) { |
ysr@777 | 3322 | // We should only do this during remark. |
ysr@777 | 3323 | if (ParallelGCThreads > 0) |
ysr@777 | 3324 | satb_mq_set.par_iterate_closure_all_threads(_task_id); |
ysr@777 | 3325 | else |
ysr@777 | 3326 | satb_mq_set.iterate_closure_all_threads(); |
ysr@777 | 3327 | } |
ysr@777 | 3328 | |
ysr@777 | 3329 | _draining_satb_buffers = false; |
ysr@777 | 3330 | |
ysr@777 | 3331 | tmp_guarantee_CM( has_aborted() || |
ysr@777 | 3332 | concurrent() || |
ysr@777 | 3333 | satb_mq_set.completed_buffers_num() == 0, "invariant" ); |
ysr@777 | 3334 | |
ysr@777 | 3335 | if (ParallelGCThreads > 0) |
ysr@777 | 3336 | satb_mq_set.set_par_closure(_task_id, NULL); |
ysr@777 | 3337 | else |
ysr@777 | 3338 | satb_mq_set.set_closure(NULL); |
ysr@777 | 3339 | |
ysr@777 | 3340 | // again, this was a potentially expensive operation, decrease the |
ysr@777 | 3341 | // limits to get the regular clock call early |
ysr@777 | 3342 | decrease_limits(); |
ysr@777 | 3343 | } |
ysr@777 | 3344 | |
ysr@777 | 3345 | void CMTask::drain_region_stack(BitMapClosure* bc) { |
ysr@777 | 3346 | if (has_aborted()) |
ysr@777 | 3347 | return; |
ysr@777 | 3348 | |
ysr@777 | 3349 | tmp_guarantee_CM( _region_finger == NULL, |
ysr@777 | 3350 | "it should be NULL when we're not scanning a region" ); |
ysr@777 | 3351 | |
ysr@777 | 3352 | if (!_cm->region_stack_empty()) { |
ysr@777 | 3353 | if (_cm->verbose_low()) |
ysr@777 | 3354 | gclog_or_tty->print_cr("[%d] draining region stack, size = %d", |
ysr@777 | 3355 | _task_id, _cm->region_stack_size()); |
ysr@777 | 3356 | |
ysr@777 | 3357 | MemRegion mr = _cm->region_stack_pop(); |
ysr@777 | 3358 | // it returns MemRegion() if the pop fails |
ysr@777 | 3359 | statsOnly(if (mr.start() != NULL) ++_region_stack_pops ); |
ysr@777 | 3360 | |
ysr@777 | 3361 | while (mr.start() != NULL) { |
ysr@777 | 3362 | if (_cm->verbose_medium()) |
ysr@777 | 3363 | gclog_or_tty->print_cr("[%d] we are scanning region " |
ysr@777 | 3364 | "["PTR_FORMAT", "PTR_FORMAT")", |
ysr@777 | 3365 | _task_id, mr.start(), mr.end()); |
ysr@777 | 3366 | tmp_guarantee_CM( mr.end() <= _cm->finger(), |
ysr@777 | 3367 | "otherwise the region shouldn't be on the stack" ); |
ysr@777 | 3368 | assert(!mr.is_empty(), "Only non-empty regions live on the region stack"); |
ysr@777 | 3369 | if (_nextMarkBitMap->iterate(bc, mr)) { |
ysr@777 | 3370 | tmp_guarantee_CM( !has_aborted(), |
ysr@777 | 3371 | "cannot abort the task without aborting the bitmap iteration" ); |
ysr@777 | 3372 | |
ysr@777 | 3373 | // We finished iterating over the region without aborting. |
ysr@777 | 3374 | regular_clock_call(); |
ysr@777 | 3375 | if (has_aborted()) |
ysr@777 | 3376 | mr = MemRegion(); |
ysr@777 | 3377 | else { |
ysr@777 | 3378 | mr = _cm->region_stack_pop(); |
ysr@777 | 3379 | // it returns MemRegion() if the pop fails |
ysr@777 | 3380 | statsOnly(if (mr.start() != NULL) ++_region_stack_pops ); |
ysr@777 | 3381 | } |
ysr@777 | 3382 | } else { |
ysr@777 | 3383 | guarantee( has_aborted(), "currently the only way to do so" ); |
ysr@777 | 3384 | |
ysr@777 | 3385 | // The only way to abort the bitmap iteration is to return |
ysr@777 | 3386 | // false from the do_bit() method. However, inside the |
ysr@777 | 3387 | // do_bit() method we move the _region_finger to point to the |
ysr@777 | 3388 | // object currently being looked at. So, if we bail out, we |
ysr@777 | 3389 | // have definitely set _region_finger to something non-null. |
ysr@777 | 3390 | guarantee( _region_finger != NULL, "invariant" ); |
ysr@777 | 3391 | |
ysr@777 | 3392 | // The iteration was actually aborted. So now _region_finger |
ysr@777 | 3393 | // points to the address of the object we last scanned. If we |
ysr@777 | 3394 | // leave it there, when we restart this task, we will rescan |
ysr@777 | 3395 | // the object. It is easy to avoid this. We move the finger by |
ysr@777 | 3396 | // enough to point to the next possible object header (the |
ysr@777 | 3397 | // bitmap knows by how much we need to move it as it knows its |
ysr@777 | 3398 | // granularity). |
ysr@777 | 3399 | MemRegion newRegion = |
ysr@777 | 3400 | MemRegion(_nextMarkBitMap->nextWord(_region_finger), mr.end()); |
ysr@777 | 3401 | |
ysr@777 | 3402 | if (!newRegion.is_empty()) { |
ysr@777 | 3403 | if (_cm->verbose_low()) { |
ysr@777 | 3404 | gclog_or_tty->print_cr("[%d] pushing unscanned region" |
ysr@777 | 3405 | "[" PTR_FORMAT "," PTR_FORMAT ") on region stack", |
ysr@777 | 3406 | _task_id, |
ysr@777 | 3407 | newRegion.start(), newRegion.end()); |
ysr@777 | 3408 | } |
ysr@777 | 3409 | // Now push the part of the region we didn't scan on the |
ysr@777 | 3410 | // region stack to make sure a task scans it later. |
ysr@777 | 3411 | _cm->region_stack_push(newRegion); |
ysr@777 | 3412 | } |
ysr@777 | 3413 | // break from while |
ysr@777 | 3414 | mr = MemRegion(); |
ysr@777 | 3415 | } |
ysr@777 | 3416 | _region_finger = NULL; |
ysr@777 | 3417 | } |
ysr@777 | 3418 | |
ysr@777 | 3419 | // We only push regions on the region stack during evacuation |
ysr@777 | 3420 | // pauses. So if we come out the above iteration because we region |
ysr@777 | 3421 | // stack is empty, it will remain empty until the next yield |
ysr@777 | 3422 | // point. So, the guarantee below is safe. |
ysr@777 | 3423 | guarantee( has_aborted() || _cm->region_stack_empty(), |
ysr@777 | 3424 | "only way to exit the loop" ); |
ysr@777 | 3425 | |
ysr@777 | 3426 | if (_cm->verbose_low()) |
ysr@777 | 3427 | gclog_or_tty->print_cr("[%d] drained region stack, size = %d", |
ysr@777 | 3428 | _task_id, _cm->region_stack_size()); |
ysr@777 | 3429 | } |
ysr@777 | 3430 | } |
ysr@777 | 3431 | |
ysr@777 | 3432 | void CMTask::print_stats() { |
ysr@777 | 3433 | gclog_or_tty->print_cr("Marking Stats, task = %d, calls = %d", |
ysr@777 | 3434 | _task_id, _calls); |
ysr@777 | 3435 | gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", |
ysr@777 | 3436 | _elapsed_time_ms, _termination_time_ms); |
ysr@777 | 3437 | gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", |
ysr@777 | 3438 | _step_times_ms.num(), _step_times_ms.avg(), |
ysr@777 | 3439 | _step_times_ms.sd()); |
ysr@777 | 3440 | gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", |
ysr@777 | 3441 | _step_times_ms.maximum(), _step_times_ms.sum()); |
ysr@777 | 3442 | |
ysr@777 | 3443 | #if _MARKING_STATS_ |
ysr@777 | 3444 | gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", |
ysr@777 | 3445 | _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), |
ysr@777 | 3446 | _all_clock_intervals_ms.sd()); |
ysr@777 | 3447 | gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", |
ysr@777 | 3448 | _all_clock_intervals_ms.maximum(), |
ysr@777 | 3449 | _all_clock_intervals_ms.sum()); |
ysr@777 | 3450 | gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", |
ysr@777 | 3451 | _clock_due_to_scanning, _clock_due_to_marking); |
ysr@777 | 3452 | gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", |
ysr@777 | 3453 | _objs_scanned, _objs_found_on_bitmap); |
ysr@777 | 3454 | gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", |
ysr@777 | 3455 | _local_pushes, _local_pops, _local_max_size); |
ysr@777 | 3456 | gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", |
ysr@777 | 3457 | _global_pushes, _global_pops, _global_max_size); |
ysr@777 | 3458 | gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", |
ysr@777 | 3459 | _global_transfers_to,_global_transfers_from); |
ysr@777 | 3460 | gclog_or_tty->print_cr(" Regions: claimed = %d, Region Stack: pops = %d", |
ysr@777 | 3461 | _regions_claimed, _region_stack_pops); |
ysr@777 | 3462 | gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); |
ysr@777 | 3463 | gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", |
ysr@777 | 3464 | _steal_attempts, _steals); |
ysr@777 | 3465 | gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); |
ysr@777 | 3466 | gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", |
ysr@777 | 3467 | _aborted_overflow, _aborted_cm_aborted, _aborted_yield); |
ysr@777 | 3468 | gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", |
ysr@777 | 3469 | _aborted_timed_out, _aborted_satb, _aborted_termination); |
ysr@777 | 3470 | #endif // _MARKING_STATS_ |
ysr@777 | 3471 | } |
ysr@777 | 3472 | |
ysr@777 | 3473 | /***************************************************************************** |
ysr@777 | 3474 | |
ysr@777 | 3475 | The do_marking_step(time_target_ms) method is the building block |
ysr@777 | 3476 | of the parallel marking framework. It can be called in parallel |
ysr@777 | 3477 | with other invocations of do_marking_step() on different tasks |
ysr@777 | 3478 | (but only one per task, obviously) and concurrently with the |
ysr@777 | 3479 | mutator threads, or during remark, hence it eliminates the need |
ysr@777 | 3480 | for two versions of the code. When called during remark, it will |
ysr@777 | 3481 | pick up from where the task left off during the concurrent marking |
ysr@777 | 3482 | phase. Interestingly, tasks are also claimable during evacuation |
ysr@777 | 3483 | pauses too, since do_marking_step() ensures that it aborts before |
ysr@777 | 3484 | it needs to yield. |
ysr@777 | 3485 | |
ysr@777 | 3486 | The data structures that is uses to do marking work are the |
ysr@777 | 3487 | following: |
ysr@777 | 3488 | |
ysr@777 | 3489 | (1) Marking Bitmap. If there are gray objects that appear only |
ysr@777 | 3490 | on the bitmap (this happens either when dealing with an overflow |
ysr@777 | 3491 | or when the initial marking phase has simply marked the roots |
ysr@777 | 3492 | and didn't push them on the stack), then tasks claim heap |
ysr@777 | 3493 | regions whose bitmap they then scan to find gray objects. A |
ysr@777 | 3494 | global finger indicates where the end of the last claimed region |
ysr@777 | 3495 | is. A local finger indicates how far into the region a task has |
ysr@777 | 3496 | scanned. The two fingers are used to determine how to gray an |
ysr@777 | 3497 | object (i.e. whether simply marking it is OK, as it will be |
ysr@777 | 3498 | visited by a task in the future, or whether it needs to be also |
ysr@777 | 3499 | pushed on a stack). |
ysr@777 | 3500 | |
ysr@777 | 3501 | (2) Local Queue. The local queue of the task which is accessed |
ysr@777 | 3502 | reasonably efficiently by the task. Other tasks can steal from |
ysr@777 | 3503 | it when they run out of work. Throughout the marking phase, a |
ysr@777 | 3504 | task attempts to keep its local queue short but not totally |
ysr@777 | 3505 | empty, so that entries are available for stealing by other |
ysr@777 | 3506 | tasks. Only when there is no more work, a task will totally |
ysr@777 | 3507 | drain its local queue. |
ysr@777 | 3508 | |
ysr@777 | 3509 | (3) Global Mark Stack. This handles local queue overflow. During |
ysr@777 | 3510 | marking only sets of entries are moved between it and the local |
ysr@777 | 3511 | queues, as access to it requires a mutex and more fine-grain |
ysr@777 | 3512 | interaction with it which might cause contention. If it |
ysr@777 | 3513 | overflows, then the marking phase should restart and iterate |
ysr@777 | 3514 | over the bitmap to identify gray objects. Throughout the marking |
ysr@777 | 3515 | phase, tasks attempt to keep the global mark stack at a small |
ysr@777 | 3516 | length but not totally empty, so that entries are available for |
ysr@777 | 3517 | popping by other tasks. Only when there is no more work, tasks |
ysr@777 | 3518 | will totally drain the global mark stack. |
ysr@777 | 3519 | |
ysr@777 | 3520 | (4) Global Region Stack. Entries on it correspond to areas of |
ysr@777 | 3521 | the bitmap that need to be scanned since they contain gray |
ysr@777 | 3522 | objects. Pushes on the region stack only happen during |
ysr@777 | 3523 | evacuation pauses and typically correspond to areas covered by |
ysr@777 | 3524 | GC LABS. If it overflows, then the marking phase should restart |
ysr@777 | 3525 | and iterate over the bitmap to identify gray objects. Tasks will |
ysr@777 | 3526 | try to totally drain the region stack as soon as possible. |
ysr@777 | 3527 | |
ysr@777 | 3528 | (5) SATB Buffer Queue. This is where completed SATB buffers are |
ysr@777 | 3529 | made available. Buffers are regularly removed from this queue |
ysr@777 | 3530 | and scanned for roots, so that the queue doesn't get too |
ysr@777 | 3531 | long. During remark, all completed buffers are processed, as |
ysr@777 | 3532 | well as the filled in parts of any uncompleted buffers. |
ysr@777 | 3533 | |
ysr@777 | 3534 | The do_marking_step() method tries to abort when the time target |
ysr@777 | 3535 | has been reached. There are a few other cases when the |
ysr@777 | 3536 | do_marking_step() method also aborts: |
ysr@777 | 3537 | |
ysr@777 | 3538 | (1) When the marking phase has been aborted (after a Full GC). |
ysr@777 | 3539 | |
ysr@777 | 3540 | (2) When a global overflow (either on the global stack or the |
ysr@777 | 3541 | region stack) has been triggered. Before the task aborts, it |
ysr@777 | 3542 | will actually sync up with the other tasks to ensure that all |
ysr@777 | 3543 | the marking data structures (local queues, stacks, fingers etc.) |
ysr@777 | 3544 | are re-initialised so that when do_marking_step() completes, |
ysr@777 | 3545 | the marking phase can immediately restart. |
ysr@777 | 3546 | |
ysr@777 | 3547 | (3) When enough completed SATB buffers are available. The |
ysr@777 | 3548 | do_marking_step() method only tries to drain SATB buffers right |
ysr@777 | 3549 | at the beginning. So, if enough buffers are available, the |
ysr@777 | 3550 | marking step aborts and the SATB buffers are processed at |
ysr@777 | 3551 | the beginning of the next invocation. |
ysr@777 | 3552 | |
ysr@777 | 3553 | (4) To yield. when we have to yield then we abort and yield |
ysr@777 | 3554 | right at the end of do_marking_step(). This saves us from a lot |
ysr@777 | 3555 | of hassle as, by yielding we might allow a Full GC. If this |
ysr@777 | 3556 | happens then objects will be compacted underneath our feet, the |
ysr@777 | 3557 | heap might shrink, etc. We save checking for this by just |
ysr@777 | 3558 | aborting and doing the yield right at the end. |
ysr@777 | 3559 | |
ysr@777 | 3560 | From the above it follows that the do_marking_step() method should |
ysr@777 | 3561 | be called in a loop (or, otherwise, regularly) until it completes. |
ysr@777 | 3562 | |
ysr@777 | 3563 | If a marking step completes without its has_aborted() flag being |
ysr@777 | 3564 | true, it means it has completed the current marking phase (and |
ysr@777 | 3565 | also all other marking tasks have done so and have all synced up). |
ysr@777 | 3566 | |
ysr@777 | 3567 | A method called regular_clock_call() is invoked "regularly" (in |
ysr@777 | 3568 | sub ms intervals) throughout marking. It is this clock method that |
ysr@777 | 3569 | checks all the abort conditions which were mentioned above and |
ysr@777 | 3570 | decides when the task should abort. A work-based scheme is used to |
ysr@777 | 3571 | trigger this clock method: when the number of object words the |
ysr@777 | 3572 | marking phase has scanned or the number of references the marking |
ysr@777 | 3573 | phase has visited reach a given limit. Additional invocations to |
ysr@777 | 3574 | the method clock have been planted in a few other strategic places |
ysr@777 | 3575 | too. The initial reason for the clock method was to avoid calling |
ysr@777 | 3576 | vtime too regularly, as it is quite expensive. So, once it was in |
ysr@777 | 3577 | place, it was natural to piggy-back all the other conditions on it |
ysr@777 | 3578 | too and not constantly check them throughout the code. |
ysr@777 | 3579 | |
ysr@777 | 3580 | *****************************************************************************/ |
ysr@777 | 3581 | |
ysr@777 | 3582 | void CMTask::do_marking_step(double time_target_ms) { |
ysr@777 | 3583 | guarantee( time_target_ms >= 1.0, "minimum granularity is 1ms" ); |
ysr@777 | 3584 | guarantee( concurrent() == _cm->concurrent(), "they should be the same" ); |
ysr@777 | 3585 | |
ysr@777 | 3586 | guarantee( concurrent() || _cm->region_stack_empty(), |
ysr@777 | 3587 | "the region stack should have been cleared before remark" ); |
ysr@777 | 3588 | guarantee( _region_finger == NULL, |
ysr@777 | 3589 | "this should be non-null only when a region is being scanned" ); |
ysr@777 | 3590 | |
ysr@777 | 3591 | G1CollectorPolicy* g1_policy = _g1h->g1_policy(); |
ysr@777 | 3592 | guarantee( _task_queues != NULL, "invariant" ); |
ysr@777 | 3593 | guarantee( _task_queue != NULL, "invariant" ); |
ysr@777 | 3594 | guarantee( _task_queues->queue(_task_id) == _task_queue, "invariant" ); |
ysr@777 | 3595 | |
ysr@777 | 3596 | guarantee( !_claimed, |
ysr@777 | 3597 | "only one thread should claim this task at any one time" ); |
ysr@777 | 3598 | |
ysr@777 | 3599 | // OK, this doesn't safeguard again all possible scenarios, as it is |
ysr@777 | 3600 | // possible for two threads to set the _claimed flag at the same |
ysr@777 | 3601 | // time. But it is only for debugging purposes anyway and it will |
ysr@777 | 3602 | // catch most problems. |
ysr@777 | 3603 | _claimed = true; |
ysr@777 | 3604 | |
ysr@777 | 3605 | _start_time_ms = os::elapsedVTime() * 1000.0; |
ysr@777 | 3606 | statsOnly( _interval_start_time_ms = _start_time_ms ); |
ysr@777 | 3607 | |
ysr@777 | 3608 | double diff_prediction_ms = |
ysr@777 | 3609 | g1_policy->get_new_prediction(&_marking_step_diffs_ms); |
ysr@777 | 3610 | _time_target_ms = time_target_ms - diff_prediction_ms; |
ysr@777 | 3611 | |
ysr@777 | 3612 | // set up the variables that are used in the work-based scheme to |
ysr@777 | 3613 | // call the regular clock method |
ysr@777 | 3614 | _words_scanned = 0; |
ysr@777 | 3615 | _refs_reached = 0; |
ysr@777 | 3616 | recalculate_limits(); |
ysr@777 | 3617 | |
ysr@777 | 3618 | // clear all flags |
ysr@777 | 3619 | clear_has_aborted(); |
ysr@777 | 3620 | _has_aborted_timed_out = false; |
ysr@777 | 3621 | _draining_satb_buffers = false; |
ysr@777 | 3622 | |
ysr@777 | 3623 | ++_calls; |
ysr@777 | 3624 | |
ysr@777 | 3625 | if (_cm->verbose_low()) |
ysr@777 | 3626 | gclog_or_tty->print_cr("[%d] >>>>>>>>>> START, call = %d, " |
ysr@777 | 3627 | "target = %1.2lfms >>>>>>>>>>", |
ysr@777 | 3628 | _task_id, _calls, _time_target_ms); |
ysr@777 | 3629 | |
ysr@777 | 3630 | // Set up the bitmap and oop closures. Anything that uses them is |
ysr@777 | 3631 | // eventually called from this method, so it is OK to allocate these |
ysr@777 | 3632 | // statically. |
ysr@777 | 3633 | CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); |
ysr@777 | 3634 | CMOopClosure oop_closure(_g1h, _cm, this); |
ysr@777 | 3635 | set_oop_closure(&oop_closure); |
ysr@777 | 3636 | |
ysr@777 | 3637 | if (_cm->has_overflown()) { |
ysr@777 | 3638 | // This can happen if the region stack or the mark stack overflows |
ysr@777 | 3639 | // during a GC pause and this task, after a yield point, |
ysr@777 | 3640 | // restarts. We have to abort as we need to get into the overflow |
ysr@777 | 3641 | // protocol which happens right at the end of this task. |
ysr@777 | 3642 | set_has_aborted(); |
ysr@777 | 3643 | } |
ysr@777 | 3644 | |
ysr@777 | 3645 | // First drain any available SATB buffers. After this, we will not |
ysr@777 | 3646 | // look at SATB buffers before the next invocation of this method. |
ysr@777 | 3647 | // If enough completed SATB buffers are queued up, the regular clock |
ysr@777 | 3648 | // will abort this task so that it restarts. |
ysr@777 | 3649 | drain_satb_buffers(); |
ysr@777 | 3650 | // ...then partially drain the local queue and the global stack |
ysr@777 | 3651 | drain_local_queue(true); |
ysr@777 | 3652 | drain_global_stack(true); |
ysr@777 | 3653 | |
ysr@777 | 3654 | // Then totally drain the region stack. We will not look at |
ysr@777 | 3655 | // it again before the next invocation of this method. Entries on |
ysr@777 | 3656 | // the region stack are only added during evacuation pauses, for |
ysr@777 | 3657 | // which we have to yield. When we do, we abort the task anyway so |
ysr@777 | 3658 | // it will look at the region stack again when it restarts. |
ysr@777 | 3659 | bitmap_closure.set_scanning_heap_region(false); |
ysr@777 | 3660 | drain_region_stack(&bitmap_closure); |
ysr@777 | 3661 | // ...then partially drain the local queue and the global stack |
ysr@777 | 3662 | drain_local_queue(true); |
ysr@777 | 3663 | drain_global_stack(true); |
ysr@777 | 3664 | |
ysr@777 | 3665 | do { |
ysr@777 | 3666 | if (!has_aborted() && _curr_region != NULL) { |
ysr@777 | 3667 | // This means that we're already holding on to a region. |
ysr@777 | 3668 | tmp_guarantee_CM( _finger != NULL, |
ysr@777 | 3669 | "if region is not NULL, then the finger " |
ysr@777 | 3670 | "should not be NULL either" ); |
ysr@777 | 3671 | |
ysr@777 | 3672 | // We might have restarted this task after an evacuation pause |
ysr@777 | 3673 | // which might have evacuated the region we're holding on to |
ysr@777 | 3674 | // underneath our feet. Let's read its limit again to make sure |
ysr@777 | 3675 | // that we do not iterate over a region of the heap that |
ysr@777 | 3676 | // contains garbage (update_region_limit() will also move |
ysr@777 | 3677 | // _finger to the start of the region if it is found empty). |
ysr@777 | 3678 | update_region_limit(); |
ysr@777 | 3679 | // We will start from _finger not from the start of the region, |
ysr@777 | 3680 | // as we might be restarting this task after aborting half-way |
ysr@777 | 3681 | // through scanning this region. In this case, _finger points to |
ysr@777 | 3682 | // the address where we last found a marked object. If this is a |
ysr@777 | 3683 | // fresh region, _finger points to start(). |
ysr@777 | 3684 | MemRegion mr = MemRegion(_finger, _region_limit); |
ysr@777 | 3685 | |
ysr@777 | 3686 | if (_cm->verbose_low()) |
ysr@777 | 3687 | gclog_or_tty->print_cr("[%d] we're scanning part " |
ysr@777 | 3688 | "["PTR_FORMAT", "PTR_FORMAT") " |
ysr@777 | 3689 | "of region "PTR_FORMAT, |
ysr@777 | 3690 | _task_id, _finger, _region_limit, _curr_region); |
ysr@777 | 3691 | |
ysr@777 | 3692 | // Let's iterate over the bitmap of the part of the |
ysr@777 | 3693 | // region that is left. |
ysr@777 | 3694 | bitmap_closure.set_scanning_heap_region(true); |
ysr@777 | 3695 | if (mr.is_empty() || |
ysr@777 | 3696 | _nextMarkBitMap->iterate(&bitmap_closure, mr)) { |
ysr@777 | 3697 | // We successfully completed iterating over the region. Now, |
ysr@777 | 3698 | // let's give up the region. |
ysr@777 | 3699 | giveup_current_region(); |
ysr@777 | 3700 | regular_clock_call(); |
ysr@777 | 3701 | } else { |
ysr@777 | 3702 | guarantee( has_aborted(), "currently the only way to do so" ); |
ysr@777 | 3703 | // The only way to abort the bitmap iteration is to return |
ysr@777 | 3704 | // false from the do_bit() method. However, inside the |
ysr@777 | 3705 | // do_bit() method we move the _finger to point to the |
ysr@777 | 3706 | // object currently being looked at. So, if we bail out, we |
ysr@777 | 3707 | // have definitely set _finger to something non-null. |
ysr@777 | 3708 | guarantee( _finger != NULL, "invariant" ); |
ysr@777 | 3709 | |
ysr@777 | 3710 | // Region iteration was actually aborted. So now _finger |
ysr@777 | 3711 | // points to the address of the object we last scanned. If we |
ysr@777 | 3712 | // leave it there, when we restart this task, we will rescan |
ysr@777 | 3713 | // the object. It is easy to avoid this. We move the finger by |
ysr@777 | 3714 | // enough to point to the next possible object header (the |
ysr@777 | 3715 | // bitmap knows by how much we need to move it as it knows its |
ysr@777 | 3716 | // granularity). |
ysr@777 | 3717 | move_finger_to(_nextMarkBitMap->nextWord(_finger)); |
ysr@777 | 3718 | } |
ysr@777 | 3719 | } |
ysr@777 | 3720 | // At this point we have either completed iterating over the |
ysr@777 | 3721 | // region we were holding on to, or we have aborted. |
ysr@777 | 3722 | |
ysr@777 | 3723 | // We then partially drain the local queue and the global stack. |
ysr@777 | 3724 | // (Do we really need this?) |
ysr@777 | 3725 | drain_local_queue(true); |
ysr@777 | 3726 | drain_global_stack(true); |
ysr@777 | 3727 | |
ysr@777 | 3728 | // Read the note on the claim_region() method on why it might |
ysr@777 | 3729 | // return NULL with potentially more regions available for |
ysr@777 | 3730 | // claiming and why we have to check out_of_regions() to determine |
ysr@777 | 3731 | // whether we're done or not. |
ysr@777 | 3732 | while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { |
ysr@777 | 3733 | // We are going to try to claim a new region. We should have |
ysr@777 | 3734 | // given up on the previous one. |
ysr@777 | 3735 | tmp_guarantee_CM( _curr_region == NULL && |
ysr@777 | 3736 | _finger == NULL && |
ysr@777 | 3737 | _region_limit == NULL, "invariant" ); |
ysr@777 | 3738 | if (_cm->verbose_low()) |
ysr@777 | 3739 | gclog_or_tty->print_cr("[%d] trying to claim a new region", _task_id); |
ysr@777 | 3740 | HeapRegion* claimed_region = _cm->claim_region(_task_id); |
ysr@777 | 3741 | if (claimed_region != NULL) { |
ysr@777 | 3742 | // Yes, we managed to claim one |
ysr@777 | 3743 | statsOnly( ++_regions_claimed ); |
ysr@777 | 3744 | |
ysr@777 | 3745 | if (_cm->verbose_low()) |
ysr@777 | 3746 | gclog_or_tty->print_cr("[%d] we successfully claimed " |
ysr@777 | 3747 | "region "PTR_FORMAT, |
ysr@777 | 3748 | _task_id, claimed_region); |
ysr@777 | 3749 | |
ysr@777 | 3750 | setup_for_region(claimed_region); |
ysr@777 | 3751 | tmp_guarantee_CM( _curr_region == claimed_region, "invariant" ); |
ysr@777 | 3752 | } |
ysr@777 | 3753 | // It is important to call the regular clock here. It might take |
ysr@777 | 3754 | // a while to claim a region if, for example, we hit a large |
ysr@777 | 3755 | // block of empty regions. So we need to call the regular clock |
ysr@777 | 3756 | // method once round the loop to make sure it's called |
ysr@777 | 3757 | // frequently enough. |
ysr@777 | 3758 | regular_clock_call(); |
ysr@777 | 3759 | } |
ysr@777 | 3760 | |
ysr@777 | 3761 | if (!has_aborted() && _curr_region == NULL) { |
ysr@777 | 3762 | tmp_guarantee_CM( _cm->out_of_regions(), |
ysr@777 | 3763 | "at this point we should be out of regions" ); |
ysr@777 | 3764 | } |
ysr@777 | 3765 | } while ( _curr_region != NULL && !has_aborted()); |
ysr@777 | 3766 | |
ysr@777 | 3767 | if (!has_aborted()) { |
ysr@777 | 3768 | // We cannot check whether the global stack is empty, since other |
iveresov@778 | 3769 | // tasks might be pushing objects to it concurrently. We also cannot |
iveresov@778 | 3770 | // check if the region stack is empty because if a thread is aborting |
iveresov@778 | 3771 | // it can push a partially done region back. |
iveresov@778 | 3772 | tmp_guarantee_CM( _cm->out_of_regions(), |
ysr@777 | 3773 | "at this point we should be out of regions" ); |
ysr@777 | 3774 | |
ysr@777 | 3775 | if (_cm->verbose_low()) |
ysr@777 | 3776 | gclog_or_tty->print_cr("[%d] all regions claimed", _task_id); |
ysr@777 | 3777 | |
ysr@777 | 3778 | // Try to reduce the number of available SATB buffers so that |
ysr@777 | 3779 | // remark has less work to do. |
ysr@777 | 3780 | drain_satb_buffers(); |
ysr@777 | 3781 | } |
ysr@777 | 3782 | |
ysr@777 | 3783 | // Since we've done everything else, we can now totally drain the |
ysr@777 | 3784 | // local queue and global stack. |
ysr@777 | 3785 | drain_local_queue(false); |
ysr@777 | 3786 | drain_global_stack(false); |
ysr@777 | 3787 | |
ysr@777 | 3788 | // Attempt at work stealing from other task's queues. |
ysr@777 | 3789 | if (!has_aborted()) { |
ysr@777 | 3790 | // We have not aborted. This means that we have finished all that |
ysr@777 | 3791 | // we could. Let's try to do some stealing... |
ysr@777 | 3792 | |
ysr@777 | 3793 | // We cannot check whether the global stack is empty, since other |
iveresov@778 | 3794 | // tasks might be pushing objects to it concurrently. We also cannot |
iveresov@778 | 3795 | // check if the region stack is empty because if a thread is aborting |
iveresov@778 | 3796 | // it can push a partially done region back. |
ysr@777 | 3797 | guarantee( _cm->out_of_regions() && |
ysr@777 | 3798 | _task_queue->size() == 0, "only way to reach here" ); |
ysr@777 | 3799 | |
ysr@777 | 3800 | if (_cm->verbose_low()) |
ysr@777 | 3801 | gclog_or_tty->print_cr("[%d] starting to steal", _task_id); |
ysr@777 | 3802 | |
ysr@777 | 3803 | while (!has_aborted()) { |
ysr@777 | 3804 | oop obj; |
ysr@777 | 3805 | statsOnly( ++_steal_attempts ); |
ysr@777 | 3806 | |
ysr@777 | 3807 | if (_cm->try_stealing(_task_id, &_hash_seed, obj)) { |
ysr@777 | 3808 | if (_cm->verbose_medium()) |
ysr@777 | 3809 | gclog_or_tty->print_cr("[%d] stolen "PTR_FORMAT" successfully", |
ysr@777 | 3810 | _task_id, (void*) obj); |
ysr@777 | 3811 | |
ysr@777 | 3812 | statsOnly( ++_steals ); |
ysr@777 | 3813 | |
ysr@777 | 3814 | tmp_guarantee_CM( _nextMarkBitMap->isMarked((HeapWord*) obj), |
ysr@777 | 3815 | "any stolen object should be marked" ); |
ysr@777 | 3816 | scan_object(obj); |
ysr@777 | 3817 | |
ysr@777 | 3818 | // And since we're towards the end, let's totally drain the |
ysr@777 | 3819 | // local queue and global stack. |
ysr@777 | 3820 | drain_local_queue(false); |
ysr@777 | 3821 | drain_global_stack(false); |
ysr@777 | 3822 | } else { |
ysr@777 | 3823 | break; |
ysr@777 | 3824 | } |
ysr@777 | 3825 | } |
ysr@777 | 3826 | } |
ysr@777 | 3827 | |
ysr@777 | 3828 | // We still haven't aborted. Now, let's try to get into the |
ysr@777 | 3829 | // termination protocol. |
ysr@777 | 3830 | if (!has_aborted()) { |
ysr@777 | 3831 | // We cannot check whether the global stack is empty, since other |
iveresov@778 | 3832 | // tasks might be concurrently pushing objects on it. We also cannot |
iveresov@778 | 3833 | // check if the region stack is empty because if a thread is aborting |
iveresov@778 | 3834 | // it can push a partially done region back. |
ysr@777 | 3835 | guarantee( _cm->out_of_regions() && |
ysr@777 | 3836 | _task_queue->size() == 0, "only way to reach here" ); |
ysr@777 | 3837 | |
ysr@777 | 3838 | if (_cm->verbose_low()) |
ysr@777 | 3839 | gclog_or_tty->print_cr("[%d] starting termination protocol", _task_id); |
ysr@777 | 3840 | |
ysr@777 | 3841 | _termination_start_time_ms = os::elapsedVTime() * 1000.0; |
ysr@777 | 3842 | // The CMTask class also extends the TerminatorTerminator class, |
ysr@777 | 3843 | // hence its should_exit_termination() method will also decide |
ysr@777 | 3844 | // whether to exit the termination protocol or not. |
ysr@777 | 3845 | bool finished = _cm->terminator()->offer_termination(this); |
ysr@777 | 3846 | double termination_end_time_ms = os::elapsedVTime() * 1000.0; |
ysr@777 | 3847 | _termination_time_ms += |
ysr@777 | 3848 | termination_end_time_ms - _termination_start_time_ms; |
ysr@777 | 3849 | |
ysr@777 | 3850 | if (finished) { |
ysr@777 | 3851 | // We're all done. |
ysr@777 | 3852 | |
ysr@777 | 3853 | if (_task_id == 0) { |
ysr@777 | 3854 | // let's allow task 0 to do this |
ysr@777 | 3855 | if (concurrent()) { |
ysr@777 | 3856 | guarantee( _cm->concurrent_marking_in_progress(), "invariant" ); |
ysr@777 | 3857 | // we need to set this to false before the next |
ysr@777 | 3858 | // safepoint. This way we ensure that the marking phase |
ysr@777 | 3859 | // doesn't observe any more heap expansions. |
ysr@777 | 3860 | _cm->clear_concurrent_marking_in_progress(); |
ysr@777 | 3861 | } |
ysr@777 | 3862 | } |
ysr@777 | 3863 | |
ysr@777 | 3864 | // We can now guarantee that the global stack is empty, since |
ysr@777 | 3865 | // all other tasks have finished. |
ysr@777 | 3866 | guarantee( _cm->out_of_regions() && |
ysr@777 | 3867 | _cm->region_stack_empty() && |
ysr@777 | 3868 | _cm->mark_stack_empty() && |
ysr@777 | 3869 | _task_queue->size() == 0 && |
ysr@777 | 3870 | !_cm->has_overflown() && |
ysr@777 | 3871 | !_cm->mark_stack_overflow() && |
ysr@777 | 3872 | !_cm->region_stack_overflow(), |
ysr@777 | 3873 | "only way to reach here" ); |
ysr@777 | 3874 | |
ysr@777 | 3875 | if (_cm->verbose_low()) |
ysr@777 | 3876 | gclog_or_tty->print_cr("[%d] all tasks terminated", _task_id); |
ysr@777 | 3877 | } else { |
ysr@777 | 3878 | // Apparently there's more work to do. Let's abort this task. It |
ysr@777 | 3879 | // will restart it and we can hopefully find more things to do. |
ysr@777 | 3880 | |
ysr@777 | 3881 | if (_cm->verbose_low()) |
ysr@777 | 3882 | gclog_or_tty->print_cr("[%d] apparently there is more work to do", _task_id); |
ysr@777 | 3883 | |
ysr@777 | 3884 | set_has_aborted(); |
ysr@777 | 3885 | statsOnly( ++_aborted_termination ); |
ysr@777 | 3886 | } |
ysr@777 | 3887 | } |
ysr@777 | 3888 | |
ysr@777 | 3889 | // Mainly for debugging purposes to make sure that a pointer to the |
ysr@777 | 3890 | // closure which was statically allocated in this frame doesn't |
ysr@777 | 3891 | // escape it by accident. |
ysr@777 | 3892 | set_oop_closure(NULL); |
ysr@777 | 3893 | double end_time_ms = os::elapsedVTime() * 1000.0; |
ysr@777 | 3894 | double elapsed_time_ms = end_time_ms - _start_time_ms; |
ysr@777 | 3895 | // Update the step history. |
ysr@777 | 3896 | _step_times_ms.add(elapsed_time_ms); |
ysr@777 | 3897 | |
ysr@777 | 3898 | if (has_aborted()) { |
ysr@777 | 3899 | // The task was aborted for some reason. |
ysr@777 | 3900 | |
ysr@777 | 3901 | statsOnly( ++_aborted ); |
ysr@777 | 3902 | |
ysr@777 | 3903 | if (_has_aborted_timed_out) { |
ysr@777 | 3904 | double diff_ms = elapsed_time_ms - _time_target_ms; |
ysr@777 | 3905 | // Keep statistics of how well we did with respect to hitting |
ysr@777 | 3906 | // our target only if we actually timed out (if we aborted for |
ysr@777 | 3907 | // other reasons, then the results might get skewed). |
ysr@777 | 3908 | _marking_step_diffs_ms.add(diff_ms); |
ysr@777 | 3909 | } |
ysr@777 | 3910 | |
ysr@777 | 3911 | if (_cm->has_overflown()) { |
ysr@777 | 3912 | // This is the interesting one. We aborted because a global |
ysr@777 | 3913 | // overflow was raised. This means we have to restart the |
ysr@777 | 3914 | // marking phase and start iterating over regions. However, in |
ysr@777 | 3915 | // order to do this we have to make sure that all tasks stop |
ysr@777 | 3916 | // what they are doing and re-initialise in a safe manner. We |
ysr@777 | 3917 | // will achieve this with the use of two barrier sync points. |
ysr@777 | 3918 | |
ysr@777 | 3919 | if (_cm->verbose_low()) |
ysr@777 | 3920 | gclog_or_tty->print_cr("[%d] detected overflow", _task_id); |
ysr@777 | 3921 | |
ysr@777 | 3922 | _cm->enter_first_sync_barrier(_task_id); |
ysr@777 | 3923 | // When we exit this sync barrier we know that all tasks have |
ysr@777 | 3924 | // stopped doing marking work. So, it's now safe to |
ysr@777 | 3925 | // re-initialise our data structures. At the end of this method, |
ysr@777 | 3926 | // task 0 will clear the global data structures. |
ysr@777 | 3927 | |
ysr@777 | 3928 | statsOnly( ++_aborted_overflow ); |
ysr@777 | 3929 | |
ysr@777 | 3930 | // We clear the local state of this task... |
ysr@777 | 3931 | clear_region_fields(); |
ysr@777 | 3932 | |
ysr@777 | 3933 | // ...and enter the second barrier. |
ysr@777 | 3934 | _cm->enter_second_sync_barrier(_task_id); |
ysr@777 | 3935 | // At this point everything has bee re-initialised and we're |
ysr@777 | 3936 | // ready to restart. |
ysr@777 | 3937 | } |
ysr@777 | 3938 | |
ysr@777 | 3939 | if (_cm->verbose_low()) { |
ysr@777 | 3940 | gclog_or_tty->print_cr("[%d] <<<<<<<<<< ABORTING, target = %1.2lfms, " |
ysr@777 | 3941 | "elapsed = %1.2lfms <<<<<<<<<<", |
ysr@777 | 3942 | _task_id, _time_target_ms, elapsed_time_ms); |
ysr@777 | 3943 | if (_cm->has_aborted()) |
ysr@777 | 3944 | gclog_or_tty->print_cr("[%d] ========== MARKING ABORTED ==========", |
ysr@777 | 3945 | _task_id); |
ysr@777 | 3946 | } |
ysr@777 | 3947 | } else { |
ysr@777 | 3948 | if (_cm->verbose_low()) |
ysr@777 | 3949 | gclog_or_tty->print_cr("[%d] <<<<<<<<<< FINISHED, target = %1.2lfms, " |
ysr@777 | 3950 | "elapsed = %1.2lfms <<<<<<<<<<", |
ysr@777 | 3951 | _task_id, _time_target_ms, elapsed_time_ms); |
ysr@777 | 3952 | } |
ysr@777 | 3953 | |
ysr@777 | 3954 | _claimed = false; |
ysr@777 | 3955 | } |
ysr@777 | 3956 | |
ysr@777 | 3957 | CMTask::CMTask(int task_id, |
ysr@777 | 3958 | ConcurrentMark* cm, |
ysr@777 | 3959 | CMTaskQueue* task_queue, |
ysr@777 | 3960 | CMTaskQueueSet* task_queues) |
ysr@777 | 3961 | : _g1h(G1CollectedHeap::heap()), |
ysr@777 | 3962 | _task_id(task_id), _cm(cm), |
ysr@777 | 3963 | _claimed(false), |
ysr@777 | 3964 | _nextMarkBitMap(NULL), _hash_seed(17), |
ysr@777 | 3965 | _task_queue(task_queue), |
ysr@777 | 3966 | _task_queues(task_queues), |
ysr@777 | 3967 | _oop_closure(NULL) { |
ysr@777 | 3968 | guarantee( task_queue != NULL, "invariant" ); |
ysr@777 | 3969 | guarantee( task_queues != NULL, "invariant" ); |
ysr@777 | 3970 | |
ysr@777 | 3971 | statsOnly( _clock_due_to_scanning = 0; |
ysr@777 | 3972 | _clock_due_to_marking = 0 ); |
ysr@777 | 3973 | |
ysr@777 | 3974 | _marking_step_diffs_ms.add(0.5); |
ysr@777 | 3975 | } |