1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc_implementation/concurrentMarkSweep/cmsLockVerifier.hpp" 27 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 28 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 29 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 30 #include "gc_implementation/shared/liveRange.hpp" 31 #include "gc_implementation/shared/spaceDecorator.hpp" 32 #include "gc_interface/collectedHeap.inline.hpp" 33 #include "memory/allocation.inline.hpp" 34 #include "memory/blockOffsetTable.inline.hpp" 35 #include "memory/resourceArea.hpp" 36 #include "memory/universe.inline.hpp" 37 #include "oops/oop.inline.hpp" 38 #include "runtime/globals.hpp" 39 #include "runtime/handles.inline.hpp" 40 #include "runtime/init.hpp" 41 #include "runtime/java.hpp" 42 #include "runtime/vmThread.hpp" 43 #include "utilities/copy.hpp" 44 45 ///////////////////////////////////////////////////////////////////////// 46 //// CompactibleFreeListSpace 47 ///////////////////////////////////////////////////////////////////////// 48 49 // highest ranked free list lock rank 50 int CompactibleFreeListSpace::_lockRank = Mutex::leaf + 3; 51 52 // Defaults are 0 so things will break badly if incorrectly initialized. 53 size_t CompactibleFreeListSpace::IndexSetStart = 0; 54 size_t CompactibleFreeListSpace::IndexSetStride = 0; 55 56 size_t MinChunkSize = 0; 57 58 void CompactibleFreeListSpace::set_cms_values() { 59 // Set CMS global values 60 assert(MinChunkSize == 0, "already set"); 61 62 // MinChunkSize should be a multiple of MinObjAlignment and be large enough 63 // for chunks to contain a FreeChunk. 64 size_t min_chunk_size_in_bytes = align_size_up(sizeof(FreeChunk), MinObjAlignmentInBytes); 65 MinChunkSize = min_chunk_size_in_bytes / BytesPerWord; 66 67 assert(IndexSetStart == 0 && IndexSetStride == 0, "already set"); 68 IndexSetStart = MinChunkSize; 69 IndexSetStride = MinObjAlignment; 70 } 71 72 // Constructor 73 CompactibleFreeListSpace::CompactibleFreeListSpace(BlockOffsetSharedArray* bs, 74 MemRegion mr, bool use_adaptive_freelists, 75 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 76 _dictionaryChoice(dictionaryChoice), 77 _adaptive_freelists(use_adaptive_freelists), 78 _bt(bs, mr), 79 // free list locks are in the range of values taken by _lockRank 80 // This range currently is [_leaf+2, _leaf+3] 81 // Note: this requires that CFLspace c'tors 82 // are called serially in the order in which the locks are 83 // are acquired in the program text. This is true today. 84 _freelistLock(_lockRank--, "CompactibleFreeListSpace._lock", true), 85 _parDictionaryAllocLock(Mutex::leaf - 1, // == rank(ExpandHeap_lock) - 1 86 "CompactibleFreeListSpace._dict_par_lock", true), 87 _rescan_task_size(CardTableModRefBS::card_size_in_words * BitsPerWord * 88 CMSRescanMultiple), 89 _marking_task_size(CardTableModRefBS::card_size_in_words * BitsPerWord * 90 CMSConcMarkMultiple), 91 _collector(NULL) 92 { 93 assert(sizeof(FreeChunk) / BytesPerWord <= MinChunkSize, 94 "FreeChunk is larger than expected"); 95 _bt.set_space(this); 96 initialize(mr, SpaceDecorator::Clear, SpaceDecorator::Mangle); 97 // We have all of "mr", all of which we place in the dictionary 98 // as one big chunk. We'll need to decide here which of several 99 // possible alternative dictionary implementations to use. For 100 // now the choice is easy, since we have only one working 101 // implementation, namely, the simple binary tree (splaying 102 // temporarily disabled). 103 switch (dictionaryChoice) { 104 case FreeBlockDictionary<FreeChunk>::dictionaryBinaryTree: 105 _dictionary = new AFLBinaryTreeDictionary(mr); 106 break; 107 case FreeBlockDictionary<FreeChunk>::dictionarySplayTree: 108 case FreeBlockDictionary<FreeChunk>::dictionarySkipList: 109 default: 110 warning("dictionaryChoice: selected option not understood; using" 111 " default BinaryTreeDictionary implementation instead."); 112 } 113 assert(_dictionary != NULL, "CMS dictionary initialization"); 114 // The indexed free lists are initially all empty and are lazily 115 // filled in on demand. Initialize the array elements to NULL. 116 initializeIndexedFreeListArray(); 117 118 // Not using adaptive free lists assumes that allocation is first 119 // from the linAB's. Also a cms perm gen which can be compacted 120 // has to have the klass's klassKlass allocated at a lower 121 // address in the heap than the klass so that the klassKlass is 122 // moved to its new location before the klass is moved. 123 // Set the _refillSize for the linear allocation blocks 124 if (!use_adaptive_freelists) { 125 FreeChunk* fc = _dictionary->get_chunk(mr.word_size(), 126 FreeBlockDictionary<FreeChunk>::atLeast); 127 // The small linAB initially has all the space and will allocate 128 // a chunk of any size. 129 HeapWord* addr = (HeapWord*) fc; 130 _smallLinearAllocBlock.set(addr, fc->size() , 131 1024*SmallForLinearAlloc, fc->size()); 132 // Note that _unallocated_block is not updated here. 133 // Allocations from the linear allocation block should 134 // update it. 135 } else { 136 _smallLinearAllocBlock.set(0, 0, 1024*SmallForLinearAlloc, 137 SmallForLinearAlloc); 138 } 139 // CMSIndexedFreeListReplenish should be at least 1 140 CMSIndexedFreeListReplenish = MAX2((uintx)1, CMSIndexedFreeListReplenish); 141 _promoInfo.setSpace(this); 142 if (UseCMSBestFit) { 143 _fitStrategy = FreeBlockBestFitFirst; 144 } else { 145 _fitStrategy = FreeBlockStrategyNone; 146 } 147 check_free_list_consistency(); 148 149 // Initialize locks for parallel case. 150 151 if (CollectedHeap::use_parallel_gc_threads()) { 152 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 153 _indexedFreeListParLocks[i] = new Mutex(Mutex::leaf - 1, // == ExpandHeap_lock - 1 154 "a freelist par lock", 155 true); 156 DEBUG_ONLY( 157 _indexedFreeList[i].set_protecting_lock(_indexedFreeListParLocks[i]); 158 ) 159 } 160 _dictionary->set_par_lock(&_parDictionaryAllocLock); 161 } 162 } 163 164 // Like CompactibleSpace forward() but always calls cross_threshold() to 165 // update the block offset table. Removed initialize_threshold call because 166 // CFLS does not use a block offset array for contiguous spaces. 167 HeapWord* CompactibleFreeListSpace::forward(oop q, size_t size, 168 CompactPoint* cp, HeapWord* compact_top) { 169 // q is alive 170 // First check if we should switch compaction space 171 assert(this == cp->space, "'this' should be current compaction space."); 172 size_t compaction_max_size = pointer_delta(end(), compact_top); 173 assert(adjustObjectSize(size) == cp->space->adjust_object_size_v(size), 174 "virtual adjustObjectSize_v() method is not correct"); 175 size_t adjusted_size = adjustObjectSize(size); 176 assert(compaction_max_size >= MinChunkSize || compaction_max_size == 0, 177 "no small fragments allowed"); 178 assert(minimum_free_block_size() == MinChunkSize, 179 "for de-virtualized reference below"); 180 // Can't leave a nonzero size, residual fragment smaller than MinChunkSize 181 if (adjusted_size + MinChunkSize > compaction_max_size && 182 adjusted_size != compaction_max_size) { 183 do { 184 // switch to next compaction space 185 cp->space->set_compaction_top(compact_top); 186 cp->space = cp->space->next_compaction_space(); 187 if (cp->space == NULL) { 188 cp->gen = GenCollectedHeap::heap()->prev_gen(cp->gen); 189 assert(cp->gen != NULL, "compaction must succeed"); 190 cp->space = cp->gen->first_compaction_space(); 191 assert(cp->space != NULL, "generation must have a first compaction space"); 192 } 193 compact_top = cp->space->bottom(); 194 cp->space->set_compaction_top(compact_top); 195 // The correct adjusted_size may not be the same as that for this method 196 // (i.e., cp->space may no longer be "this" so adjust the size again. 197 // Use the virtual method which is not used above to save the virtual 198 // dispatch. 199 adjusted_size = cp->space->adjust_object_size_v(size); 200 compaction_max_size = pointer_delta(cp->space->end(), compact_top); 201 assert(cp->space->minimum_free_block_size() == 0, "just checking"); 202 } while (adjusted_size > compaction_max_size); 203 } 204 205 // store the forwarding pointer into the mark word 206 if ((HeapWord*)q != compact_top) { 207 q->forward_to(oop(compact_top)); 208 assert(q->is_gc_marked(), "encoding the pointer should preserve the mark"); 209 } else { 210 // if the object isn't moving we can just set the mark to the default 211 // mark and handle it specially later on. 212 q->init_mark(); 213 assert(q->forwardee() == NULL, "should be forwarded to NULL"); 214 } 215 216 compact_top += adjusted_size; 217 218 // we need to update the offset table so that the beginnings of objects can be 219 // found during scavenge. Note that we are updating the offset table based on 220 // where the object will be once the compaction phase finishes. 221 222 // Always call cross_threshold(). A contiguous space can only call it when 223 // the compaction_top exceeds the current threshold but not for an 224 // non-contiguous space. 225 cp->threshold = 226 cp->space->cross_threshold(compact_top - adjusted_size, compact_top); 227 return compact_top; 228 } 229 230 // A modified copy of OffsetTableContigSpace::cross_threshold() with _offsets -> _bt 231 // and use of single_block instead of alloc_block. The name here is not really 232 // appropriate - maybe a more general name could be invented for both the 233 // contiguous and noncontiguous spaces. 234 235 HeapWord* CompactibleFreeListSpace::cross_threshold(HeapWord* start, HeapWord* the_end) { 236 _bt.single_block(start, the_end); 237 return end(); 238 } 239 240 // Initialize them to NULL. 241 void CompactibleFreeListSpace::initializeIndexedFreeListArray() { 242 for (size_t i = 0; i < IndexSetSize; i++) { 243 // Note that on platforms where objects are double word aligned, 244 // the odd array elements are not used. It is convenient, however, 245 // to map directly from the object size to the array element. 246 _indexedFreeList[i].reset(IndexSetSize); 247 _indexedFreeList[i].set_size(i); 248 assert(_indexedFreeList[i].count() == 0, "reset check failed"); 249 assert(_indexedFreeList[i].head() == NULL, "reset check failed"); 250 assert(_indexedFreeList[i].tail() == NULL, "reset check failed"); 251 assert(_indexedFreeList[i].hint() == IndexSetSize, "reset check failed"); 252 } 253 } 254 255 void CompactibleFreeListSpace::resetIndexedFreeListArray() { 256 for (size_t i = 1; i < IndexSetSize; i++) { 257 assert(_indexedFreeList[i].size() == (size_t) i, 258 "Indexed free list sizes are incorrect"); 259 _indexedFreeList[i].reset(IndexSetSize); 260 assert(_indexedFreeList[i].count() == 0, "reset check failed"); 261 assert(_indexedFreeList[i].head() == NULL, "reset check failed"); 262 assert(_indexedFreeList[i].tail() == NULL, "reset check failed"); 263 assert(_indexedFreeList[i].hint() == IndexSetSize, "reset check failed"); 264 } 265 } 266 267 void CompactibleFreeListSpace::reset(MemRegion mr) { 268 resetIndexedFreeListArray(); 269 dictionary()->reset(); 270 if (BlockOffsetArrayUseUnallocatedBlock) { 271 assert(end() == mr.end(), "We are compacting to the bottom of CMS gen"); 272 // Everything's allocated until proven otherwise. 273 _bt.set_unallocated_block(end()); 274 } 275 if (!mr.is_empty()) { 276 assert(mr.word_size() >= MinChunkSize, "Chunk size is too small"); 277 _bt.single_block(mr.start(), mr.word_size()); 278 FreeChunk* fc = (FreeChunk*) mr.start(); 279 fc->set_size(mr.word_size()); 280 if (mr.word_size() >= IndexSetSize ) { 281 returnChunkToDictionary(fc); 282 } else { 283 _bt.verify_not_unallocated((HeapWord*)fc, fc->size()); 284 _indexedFreeList[mr.word_size()].return_chunk_at_head(fc); 285 } 286 coalBirth(mr.word_size()); 287 } 288 _promoInfo.reset(); 289 _smallLinearAllocBlock._ptr = NULL; 290 _smallLinearAllocBlock._word_size = 0; 291 } 292 293 void CompactibleFreeListSpace::reset_after_compaction() { 294 // Reset the space to the new reality - one free chunk. 295 MemRegion mr(compaction_top(), end()); 296 reset(mr); 297 // Now refill the linear allocation block(s) if possible. 298 if (_adaptive_freelists) { 299 refillLinearAllocBlocksIfNeeded(); 300 } else { 301 // Place as much of mr in the linAB as we can get, 302 // provided it was big enough to go into the dictionary. 303 FreeChunk* fc = dictionary()->find_largest_dict(); 304 if (fc != NULL) { 305 assert(fc->size() == mr.word_size(), 306 "Why was the chunk broken up?"); 307 removeChunkFromDictionary(fc); 308 HeapWord* addr = (HeapWord*) fc; 309 _smallLinearAllocBlock.set(addr, fc->size() , 310 1024*SmallForLinearAlloc, fc->size()); 311 // Note that _unallocated_block is not updated here. 312 } 313 } 314 } 315 316 // Walks the entire dictionary, returning a coterminal 317 // chunk, if it exists. Use with caution since it involves 318 // a potentially complete walk of a potentially large tree. 319 FreeChunk* CompactibleFreeListSpace::find_chunk_at_end() { 320 321 assert_lock_strong(&_freelistLock); 322 323 return dictionary()->find_chunk_ends_at(end()); 324 } 325 326 327 #ifndef PRODUCT 328 void CompactibleFreeListSpace::initializeIndexedFreeListArrayReturnedBytes() { 329 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 330 _indexedFreeList[i].allocation_stats()->set_returned_bytes(0); 331 } 332 } 333 334 size_t CompactibleFreeListSpace::sumIndexedFreeListArrayReturnedBytes() { 335 size_t sum = 0; 336 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 337 sum += _indexedFreeList[i].allocation_stats()->returned_bytes(); 338 } 339 return sum; 340 } 341 342 size_t CompactibleFreeListSpace::totalCountInIndexedFreeLists() const { 343 size_t count = 0; 344 for (size_t i = IndexSetStart; i < IndexSetSize; i++) { 345 debug_only( 346 ssize_t total_list_count = 0; 347 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 348 fc = fc->next()) { 349 total_list_count++; 350 } 351 assert(total_list_count == _indexedFreeList[i].count(), 352 "Count in list is incorrect"); 353 ) 354 count += _indexedFreeList[i].count(); 355 } 356 return count; 357 } 358 359 size_t CompactibleFreeListSpace::totalCount() { 360 size_t num = totalCountInIndexedFreeLists(); 361 num += dictionary()->total_count(); 362 if (_smallLinearAllocBlock._word_size != 0) { 363 num++; 364 } 365 return num; 366 } 367 #endif 368 369 bool CompactibleFreeListSpace::is_free_block(const HeapWord* p) const { 370 FreeChunk* fc = (FreeChunk*) p; 371 return fc->is_free(); 372 } 373 374 size_t CompactibleFreeListSpace::used() const { 375 return capacity() - free(); 376 } 377 378 size_t CompactibleFreeListSpace::free() const { 379 // "MT-safe, but not MT-precise"(TM), if you will: i.e. 380 // if you do this while the structures are in flux you 381 // may get an approximate answer only; for instance 382 // because there is concurrent allocation either 383 // directly by mutators or for promotion during a GC. 384 // It's "MT-safe", however, in the sense that you are guaranteed 385 // not to crash and burn, for instance, because of walking 386 // pointers that could disappear as you were walking them. 387 // The approximation is because the various components 388 // that are read below are not read atomically (and 389 // further the computation of totalSizeInIndexedFreeLists() 390 // is itself a non-atomic computation. The normal use of 391 // this is during a resize operation at the end of GC 392 // and at that time you are guaranteed to get the 393 // correct actual value. However, for instance, this is 394 // also read completely asynchronously by the "perf-sampler" 395 // that supports jvmstat, and you are apt to see the values 396 // flicker in such cases. 397 assert(_dictionary != NULL, "No _dictionary?"); 398 return (_dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())) + 399 totalSizeInIndexedFreeLists() + 400 _smallLinearAllocBlock._word_size) * HeapWordSize; 401 } 402 403 size_t CompactibleFreeListSpace::max_alloc_in_words() const { 404 assert(_dictionary != NULL, "No _dictionary?"); 405 assert_locked(); 406 size_t res = _dictionary->max_chunk_size(); 407 res = MAX2(res, MIN2(_smallLinearAllocBlock._word_size, 408 (size_t) SmallForLinearAlloc - 1)); 409 // XXX the following could potentially be pretty slow; 410 // should one, pesimally for the rare cases when res 411 // caclulated above is less than IndexSetSize, 412 // just return res calculated above? My reasoning was that 413 // those cases will be so rare that the extra time spent doesn't 414 // really matter.... 415 // Note: do not change the loop test i >= res + IndexSetStride 416 // to i > res below, because i is unsigned and res may be zero. 417 for (size_t i = IndexSetSize - 1; i >= res + IndexSetStride; 418 i -= IndexSetStride) { 419 if (_indexedFreeList[i].head() != NULL) { 420 assert(_indexedFreeList[i].count() != 0, "Inconsistent FreeList"); 421 return i; 422 } 423 } 424 return res; 425 } 426 427 void LinearAllocBlock::print_on(outputStream* st) const { 428 st->print_cr(" LinearAllocBlock: ptr = " PTR_FORMAT ", word_size = " SIZE_FORMAT 429 ", refillsize = " SIZE_FORMAT ", allocation_size_limit = " SIZE_FORMAT, 430 _ptr, _word_size, _refillSize, _allocation_size_limit); 431 } 432 433 void CompactibleFreeListSpace::print_on(outputStream* st) const { 434 st->print_cr("COMPACTIBLE FREELIST SPACE"); 435 st->print_cr(" Space:"); 436 Space::print_on(st); 437 438 st->print_cr("promoInfo:"); 439 _promoInfo.print_on(st); 440 441 st->print_cr("_smallLinearAllocBlock"); 442 _smallLinearAllocBlock.print_on(st); 443 444 // dump_memory_block(_smallLinearAllocBlock->_ptr, 128); 445 446 st->print_cr(" _fitStrategy = %s, _adaptive_freelists = %s", 447 _fitStrategy?"true":"false", _adaptive_freelists?"true":"false"); 448 } 449 450 void CompactibleFreeListSpace::print_indexed_free_lists(outputStream* st) 451 const { 452 reportIndexedFreeListStatistics(); 453 gclog_or_tty->print_cr("Layout of Indexed Freelists"); 454 gclog_or_tty->print_cr("---------------------------"); 455 AdaptiveFreeList<FreeChunk>::print_labels_on(st, "size"); 456 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 457 _indexedFreeList[i].print_on(gclog_or_tty); 458 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 459 fc = fc->next()) { 460 gclog_or_tty->print_cr("\t[" PTR_FORMAT "," PTR_FORMAT ") %s", 461 fc, (HeapWord*)fc + i, 462 fc->cantCoalesce() ? "\t CC" : ""); 463 } 464 } 465 } 466 467 void CompactibleFreeListSpace::print_promo_info_blocks(outputStream* st) 468 const { 469 _promoInfo.print_on(st); 470 } 471 472 void CompactibleFreeListSpace::print_dictionary_free_lists(outputStream* st) 473 const { 474 _dictionary->report_statistics(); 475 st->print_cr("Layout of Freelists in Tree"); 476 st->print_cr("---------------------------"); 477 _dictionary->print_free_lists(st); 478 } 479 480 class BlkPrintingClosure: public BlkClosure { 481 const CMSCollector* _collector; 482 const CompactibleFreeListSpace* _sp; 483 const CMSBitMap* _live_bit_map; 484 const bool _post_remark; 485 outputStream* _st; 486 public: 487 BlkPrintingClosure(const CMSCollector* collector, 488 const CompactibleFreeListSpace* sp, 489 const CMSBitMap* live_bit_map, 490 outputStream* st): 491 _collector(collector), 492 _sp(sp), 493 _live_bit_map(live_bit_map), 494 _post_remark(collector->abstract_state() > CMSCollector::FinalMarking), 495 _st(st) { } 496 size_t do_blk(HeapWord* addr); 497 }; 498 499 size_t BlkPrintingClosure::do_blk(HeapWord* addr) { 500 size_t sz = _sp->block_size_no_stall(addr, _collector); 501 assert(sz != 0, "Should always be able to compute a size"); 502 if (_sp->block_is_obj(addr)) { 503 const bool dead = _post_remark && !_live_bit_map->isMarked(addr); 504 _st->print_cr(PTR_FORMAT ": %s object of size " SIZE_FORMAT "%s", 505 addr, 506 dead ? "dead" : "live", 507 sz, 508 (!dead && CMSPrintObjectsInDump) ? ":" : "."); 509 if (CMSPrintObjectsInDump && !dead) { 510 oop(addr)->print_on(_st); 511 _st->print_cr("--------------------------------------"); 512 } 513 } else { // free block 514 _st->print_cr(PTR_FORMAT ": free block of size " SIZE_FORMAT "%s", 515 addr, sz, CMSPrintChunksInDump ? ":" : "."); 516 if (CMSPrintChunksInDump) { 517 ((FreeChunk*)addr)->print_on(_st); 518 _st->print_cr("--------------------------------------"); 519 } 520 } 521 return sz; 522 } 523 524 void CompactibleFreeListSpace::dump_at_safepoint_with_locks(CMSCollector* c, 525 outputStream* st) { 526 st->print_cr("\n========================="); 527 st->print_cr("Block layout in CMS Heap:"); 528 st->print_cr("========================="); 529 BlkPrintingClosure bpcl(c, this, c->markBitMap(), st); 530 blk_iterate(&bpcl); 531 532 st->print_cr("\n======================================="); 533 st->print_cr("Order & Layout of Promotion Info Blocks"); 534 st->print_cr("======================================="); 535 print_promo_info_blocks(st); 536 537 st->print_cr("\n==========================="); 538 st->print_cr("Order of Indexed Free Lists"); 539 st->print_cr("========================="); 540 print_indexed_free_lists(st); 541 542 st->print_cr("\n================================="); 543 st->print_cr("Order of Free Lists in Dictionary"); 544 st->print_cr("================================="); 545 print_dictionary_free_lists(st); 546 } 547 548 549 void CompactibleFreeListSpace::reportFreeListStatistics() const { 550 assert_lock_strong(&_freelistLock); 551 assert(PrintFLSStatistics != 0, "Reporting error"); 552 _dictionary->report_statistics(); 553 if (PrintFLSStatistics > 1) { 554 reportIndexedFreeListStatistics(); 555 size_t total_size = totalSizeInIndexedFreeLists() + 556 _dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())); 557 gclog_or_tty->print(" free=" SIZE_FORMAT " frag=%1.4f\n", total_size, flsFrag()); 558 } 559 } 560 561 void CompactibleFreeListSpace::reportIndexedFreeListStatistics() const { 562 assert_lock_strong(&_freelistLock); 563 gclog_or_tty->print("Statistics for IndexedFreeLists:\n" 564 "--------------------------------\n"); 565 size_t total_size = totalSizeInIndexedFreeLists(); 566 size_t free_blocks = numFreeBlocksInIndexedFreeLists(); 567 gclog_or_tty->print("Total Free Space: %d\n", total_size); 568 gclog_or_tty->print("Max Chunk Size: %d\n", maxChunkSizeInIndexedFreeLists()); 569 gclog_or_tty->print("Number of Blocks: %d\n", free_blocks); 570 if (free_blocks != 0) { 571 gclog_or_tty->print("Av. Block Size: %d\n", total_size/free_blocks); 572 } 573 } 574 575 size_t CompactibleFreeListSpace::numFreeBlocksInIndexedFreeLists() const { 576 size_t res = 0; 577 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 578 debug_only( 579 ssize_t recount = 0; 580 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 581 fc = fc->next()) { 582 recount += 1; 583 } 584 assert(recount == _indexedFreeList[i].count(), 585 "Incorrect count in list"); 586 ) 587 res += _indexedFreeList[i].count(); 588 } 589 return res; 590 } 591 592 size_t CompactibleFreeListSpace::maxChunkSizeInIndexedFreeLists() const { 593 for (size_t i = IndexSetSize - 1; i != 0; i -= IndexSetStride) { 594 if (_indexedFreeList[i].head() != NULL) { 595 assert(_indexedFreeList[i].count() != 0, "Inconsistent FreeList"); 596 return (size_t)i; 597 } 598 } 599 return 0; 600 } 601 602 void CompactibleFreeListSpace::set_end(HeapWord* value) { 603 HeapWord* prevEnd = end(); 604 assert(prevEnd != value, "unnecessary set_end call"); 605 assert(prevEnd == NULL || !BlockOffsetArrayUseUnallocatedBlock || value >= unallocated_block(), 606 "New end is below unallocated block"); 607 _end = value; 608 if (prevEnd != NULL) { 609 // Resize the underlying block offset table. 610 _bt.resize(pointer_delta(value, bottom())); 611 if (value <= prevEnd) { 612 assert(!BlockOffsetArrayUseUnallocatedBlock || value >= unallocated_block(), 613 "New end is below unallocated block"); 614 } else { 615 // Now, take this new chunk and add it to the free blocks. 616 // Note that the BOT has not yet been updated for this block. 617 size_t newFcSize = pointer_delta(value, prevEnd); 618 // XXX This is REALLY UGLY and should be fixed up. XXX 619 if (!_adaptive_freelists && _smallLinearAllocBlock._ptr == NULL) { 620 // Mark the boundary of the new block in BOT 621 _bt.mark_block(prevEnd, value); 622 // put it all in the linAB 623 if (ParallelGCThreads == 0) { 624 _smallLinearAllocBlock._ptr = prevEnd; 625 _smallLinearAllocBlock._word_size = newFcSize; 626 repairLinearAllocBlock(&_smallLinearAllocBlock); 627 } else { // ParallelGCThreads > 0 628 MutexLockerEx x(parDictionaryAllocLock(), 629 Mutex::_no_safepoint_check_flag); 630 _smallLinearAllocBlock._ptr = prevEnd; 631 _smallLinearAllocBlock._word_size = newFcSize; 632 repairLinearAllocBlock(&_smallLinearAllocBlock); 633 } 634 // Births of chunks put into a LinAB are not recorded. Births 635 // of chunks as they are allocated out of a LinAB are. 636 } else { 637 // Add the block to the free lists, if possible coalescing it 638 // with the last free block, and update the BOT and census data. 639 addChunkToFreeListsAtEndRecordingStats(prevEnd, newFcSize); 640 } 641 } 642 } 643 } 644 645 class FreeListSpace_DCTOC : public Filtering_DCTOC { 646 CompactibleFreeListSpace* _cfls; 647 CMSCollector* _collector; 648 protected: 649 // Override. 650 #define walk_mem_region_with_cl_DECL(ClosureType) \ 651 virtual void walk_mem_region_with_cl(MemRegion mr, \ 652 HeapWord* bottom, HeapWord* top, \ 653 ClosureType* cl); \ 654 void walk_mem_region_with_cl_par(MemRegion mr, \ 655 HeapWord* bottom, HeapWord* top, \ 656 ClosureType* cl); \ 657 void walk_mem_region_with_cl_nopar(MemRegion mr, \ 658 HeapWord* bottom, HeapWord* top, \ 659 ClosureType* cl) 660 walk_mem_region_with_cl_DECL(ExtendedOopClosure); 661 walk_mem_region_with_cl_DECL(FilteringClosure); 662 663 public: 664 FreeListSpace_DCTOC(CompactibleFreeListSpace* sp, 665 CMSCollector* collector, 666 ExtendedOopClosure* cl, 667 CardTableModRefBS::PrecisionStyle precision, 668 HeapWord* boundary) : 669 Filtering_DCTOC(sp, cl, precision, boundary), 670 _cfls(sp), _collector(collector) {} 671 }; 672 673 // We de-virtualize the block-related calls below, since we know that our 674 // space is a CompactibleFreeListSpace. 675 676 #define FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(ClosureType) \ 677 void FreeListSpace_DCTOC::walk_mem_region_with_cl(MemRegion mr, \ 678 HeapWord* bottom, \ 679 HeapWord* top, \ 680 ClosureType* cl) { \ 681 bool is_par = SharedHeap::heap()->n_par_threads() > 0; \ 682 if (is_par) { \ 683 assert(SharedHeap::heap()->n_par_threads() == \ 684 SharedHeap::heap()->workers()->active_workers(), "Mismatch"); \ 685 walk_mem_region_with_cl_par(mr, bottom, top, cl); \ 686 } else { \ 687 walk_mem_region_with_cl_nopar(mr, bottom, top, cl); \ 688 } \ 689 } \ 690 void FreeListSpace_DCTOC::walk_mem_region_with_cl_par(MemRegion mr, \ 691 HeapWord* bottom, \ 692 HeapWord* top, \ 693 ClosureType* cl) { \ 694 /* Skip parts that are before "mr", in case "block_start" sent us \ 695 back too far. */ \ 696 HeapWord* mr_start = mr.start(); \ 697 size_t bot_size = _cfls->CompactibleFreeListSpace::block_size(bottom); \ 698 HeapWord* next = bottom + bot_size; \ 699 while (next < mr_start) { \ 700 bottom = next; \ 701 bot_size = _cfls->CompactibleFreeListSpace::block_size(bottom); \ 702 next = bottom + bot_size; \ 703 } \ 704 \ 705 while (bottom < top) { \ 706 if (_cfls->CompactibleFreeListSpace::block_is_obj(bottom) && \ 707 !_cfls->CompactibleFreeListSpace::obj_allocated_since_save_marks( \ 708 oop(bottom)) && \ 709 !_collector->CMSCollector::is_dead_obj(oop(bottom))) { \ 710 size_t word_sz = oop(bottom)->oop_iterate(cl, mr); \ 711 bottom += _cfls->adjustObjectSize(word_sz); \ 712 } else { \ 713 bottom += _cfls->CompactibleFreeListSpace::block_size(bottom); \ 714 } \ 715 } \ 716 } \ 717 void FreeListSpace_DCTOC::walk_mem_region_with_cl_nopar(MemRegion mr, \ 718 HeapWord* bottom, \ 719 HeapWord* top, \ 720 ClosureType* cl) { \ 721 /* Skip parts that are before "mr", in case "block_start" sent us \ 722 back too far. */ \ 723 HeapWord* mr_start = mr.start(); \ 724 size_t bot_size = _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \ 725 HeapWord* next = bottom + bot_size; \ 726 while (next < mr_start) { \ 727 bottom = next; \ 728 bot_size = _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \ 729 next = bottom + bot_size; \ 730 } \ 731 \ 732 while (bottom < top) { \ 733 if (_cfls->CompactibleFreeListSpace::block_is_obj_nopar(bottom) && \ 734 !_cfls->CompactibleFreeListSpace::obj_allocated_since_save_marks( \ 735 oop(bottom)) && \ 736 !_collector->CMSCollector::is_dead_obj(oop(bottom))) { \ 737 size_t word_sz = oop(bottom)->oop_iterate(cl, mr); \ 738 bottom += _cfls->adjustObjectSize(word_sz); \ 739 } else { \ 740 bottom += _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \ 741 } \ 742 } \ 743 } 744 745 // (There are only two of these, rather than N, because the split is due 746 // only to the introduction of the FilteringClosure, a local part of the 747 // impl of this abstraction.) 748 FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(ExtendedOopClosure) 749 FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(FilteringClosure) 750 751 DirtyCardToOopClosure* 752 CompactibleFreeListSpace::new_dcto_cl(ExtendedOopClosure* cl, 753 CardTableModRefBS::PrecisionStyle precision, 754 HeapWord* boundary) { 755 return new FreeListSpace_DCTOC(this, _collector, cl, precision, boundary); 756 } 757 758 759 // Note on locking for the space iteration functions: 760 // since the collector's iteration activities are concurrent with 761 // allocation activities by mutators, absent a suitable mutual exclusion 762 // mechanism the iterators may go awry. For instace a block being iterated 763 // may suddenly be allocated or divided up and part of it allocated and 764 // so on. 765 766 // Apply the given closure to each block in the space. 767 void CompactibleFreeListSpace::blk_iterate_careful(BlkClosureCareful* cl) { 768 assert_lock_strong(freelistLock()); 769 HeapWord *cur, *limit; 770 for (cur = bottom(), limit = end(); cur < limit; 771 cur += cl->do_blk_careful(cur)); 772 } 773 774 // Apply the given closure to each block in the space. 775 void CompactibleFreeListSpace::blk_iterate(BlkClosure* cl) { 776 assert_lock_strong(freelistLock()); 777 HeapWord *cur, *limit; 778 for (cur = bottom(), limit = end(); cur < limit; 779 cur += cl->do_blk(cur)); 780 } 781 782 // Apply the given closure to each oop in the space. 783 void CompactibleFreeListSpace::oop_iterate(ExtendedOopClosure* cl) { 784 assert_lock_strong(freelistLock()); 785 HeapWord *cur, *limit; 786 size_t curSize; 787 for (cur = bottom(), limit = end(); cur < limit; 788 cur += curSize) { 789 curSize = block_size(cur); 790 if (block_is_obj(cur)) { 791 oop(cur)->oop_iterate(cl); 792 } 793 } 794 } 795 796 // Apply the given closure to each oop in the space \intersect memory region. 797 void CompactibleFreeListSpace::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 798 assert_lock_strong(freelistLock()); 799 if (is_empty()) { 800 return; 801 } 802 MemRegion cur = MemRegion(bottom(), end()); 803 mr = mr.intersection(cur); 804 if (mr.is_empty()) { 805 return; 806 } 807 if (mr.equals(cur)) { 808 oop_iterate(cl); 809 return; 810 } 811 assert(mr.end() <= end(), "just took an intersection above"); 812 HeapWord* obj_addr = block_start(mr.start()); 813 HeapWord* t = mr.end(); 814 815 SpaceMemRegionOopsIterClosure smr_blk(cl, mr); 816 if (block_is_obj(obj_addr)) { 817 // Handle first object specially. 818 oop obj = oop(obj_addr); 819 obj_addr += adjustObjectSize(obj->oop_iterate(&smr_blk)); 820 } else { 821 FreeChunk* fc = (FreeChunk*)obj_addr; 822 obj_addr += fc->size(); 823 } 824 while (obj_addr < t) { 825 HeapWord* obj = obj_addr; 826 obj_addr += block_size(obj_addr); 827 // If "obj_addr" is not greater than top, then the 828 // entire object "obj" is within the region. 829 if (obj_addr <= t) { 830 if (block_is_obj(obj)) { 831 oop(obj)->oop_iterate(cl); 832 } 833 } else { 834 // "obj" extends beyond end of region 835 if (block_is_obj(obj)) { 836 oop(obj)->oop_iterate(&smr_blk); 837 } 838 break; 839 } 840 } 841 } 842 843 // NOTE: In the following methods, in order to safely be able to 844 // apply the closure to an object, we need to be sure that the 845 // object has been initialized. We are guaranteed that an object 846 // is initialized if we are holding the Heap_lock with the 847 // world stopped. 848 void CompactibleFreeListSpace::verify_objects_initialized() const { 849 if (is_init_completed()) { 850 assert_locked_or_safepoint(Heap_lock); 851 if (Universe::is_fully_initialized()) { 852 guarantee(SafepointSynchronize::is_at_safepoint(), 853 "Required for objects to be initialized"); 854 } 855 } // else make a concession at vm start-up 856 } 857 858 // Apply the given closure to each object in the space 859 void CompactibleFreeListSpace::object_iterate(ObjectClosure* blk) { 860 assert_lock_strong(freelistLock()); 861 NOT_PRODUCT(verify_objects_initialized()); 862 HeapWord *cur, *limit; 863 size_t curSize; 864 for (cur = bottom(), limit = end(); cur < limit; 865 cur += curSize) { 866 curSize = block_size(cur); 867 if (block_is_obj(cur)) { 868 blk->do_object(oop(cur)); 869 } 870 } 871 } 872 873 // Apply the given closure to each live object in the space 874 // The usage of CompactibleFreeListSpace 875 // by the ConcurrentMarkSweepGeneration for concurrent GC's allows 876 // objects in the space with references to objects that are no longer 877 // valid. For example, an object may reference another object 878 // that has already been sweep up (collected). This method uses 879 // obj_is_alive() to determine whether it is safe to apply the closure to 880 // an object. See obj_is_alive() for details on how liveness of an 881 // object is decided. 882 883 void CompactibleFreeListSpace::safe_object_iterate(ObjectClosure* blk) { 884 assert_lock_strong(freelistLock()); 885 NOT_PRODUCT(verify_objects_initialized()); 886 HeapWord *cur, *limit; 887 size_t curSize; 888 for (cur = bottom(), limit = end(); cur < limit; 889 cur += curSize) { 890 curSize = block_size(cur); 891 if (block_is_obj(cur) && obj_is_alive(cur)) { 892 blk->do_object(oop(cur)); 893 } 894 } 895 } 896 897 void CompactibleFreeListSpace::object_iterate_mem(MemRegion mr, 898 UpwardsObjectClosure* cl) { 899 assert_locked(freelistLock()); 900 NOT_PRODUCT(verify_objects_initialized()); 901 Space::object_iterate_mem(mr, cl); 902 } 903 904 // Callers of this iterator beware: The closure application should 905 // be robust in the face of uninitialized objects and should (always) 906 // return a correct size so that the next addr + size below gives us a 907 // valid block boundary. [See for instance, 908 // ScanMarkedObjectsAgainCarefullyClosure::do_object_careful() 909 // in ConcurrentMarkSweepGeneration.cpp.] 910 HeapWord* 911 CompactibleFreeListSpace::object_iterate_careful(ObjectClosureCareful* cl) { 912 assert_lock_strong(freelistLock()); 913 HeapWord *addr, *last; 914 size_t size; 915 for (addr = bottom(), last = end(); 916 addr < last; addr += size) { 917 FreeChunk* fc = (FreeChunk*)addr; 918 if (fc->is_free()) { 919 // Since we hold the free list lock, which protects direct 920 // allocation in this generation by mutators, a free object 921 // will remain free throughout this iteration code. 922 size = fc->size(); 923 } else { 924 // Note that the object need not necessarily be initialized, 925 // because (for instance) the free list lock does NOT protect 926 // object initialization. The closure application below must 927 // therefore be correct in the face of uninitialized objects. 928 size = cl->do_object_careful(oop(addr)); 929 if (size == 0) { 930 // An unparsable object found. Signal early termination. 931 return addr; 932 } 933 } 934 } 935 return NULL; 936 } 937 938 // Callers of this iterator beware: The closure application should 939 // be robust in the face of uninitialized objects and should (always) 940 // return a correct size so that the next addr + size below gives us a 941 // valid block boundary. [See for instance, 942 // ScanMarkedObjectsAgainCarefullyClosure::do_object_careful() 943 // in ConcurrentMarkSweepGeneration.cpp.] 944 HeapWord* 945 CompactibleFreeListSpace::object_iterate_careful_m(MemRegion mr, 946 ObjectClosureCareful* cl) { 947 assert_lock_strong(freelistLock()); 948 // Can't use used_region() below because it may not necessarily 949 // be the same as [bottom(),end()); although we could 950 // use [used_region().start(),round_to(used_region().end(),CardSize)), 951 // that appears too cumbersome, so we just do the simpler check 952 // in the assertion below. 953 assert(!mr.is_empty() && MemRegion(bottom(),end()).contains(mr), 954 "mr should be non-empty and within used space"); 955 HeapWord *addr, *end; 956 size_t size; 957 for (addr = block_start_careful(mr.start()), end = mr.end(); 958 addr < end; addr += size) { 959 FreeChunk* fc = (FreeChunk*)addr; 960 if (fc->is_free()) { 961 // Since we hold the free list lock, which protects direct 962 // allocation in this generation by mutators, a free object 963 // will remain free throughout this iteration code. 964 size = fc->size(); 965 } else { 966 // Note that the object need not necessarily be initialized, 967 // because (for instance) the free list lock does NOT protect 968 // object initialization. The closure application below must 969 // therefore be correct in the face of uninitialized objects. 970 size = cl->do_object_careful_m(oop(addr), mr); 971 if (size == 0) { 972 // An unparsable object found. Signal early termination. 973 return addr; 974 } 975 } 976 } 977 return NULL; 978 } 979 980 981 HeapWord* CompactibleFreeListSpace::block_start_const(const void* p) const { 982 NOT_PRODUCT(verify_objects_initialized()); 983 return _bt.block_start(p); 984 } 985 986 HeapWord* CompactibleFreeListSpace::block_start_careful(const void* p) const { 987 return _bt.block_start_careful(p); 988 } 989 990 size_t CompactibleFreeListSpace::block_size(const HeapWord* p) const { 991 NOT_PRODUCT(verify_objects_initialized()); 992 // This must be volatile, or else there is a danger that the compiler 993 // will compile the code below into a sometimes-infinite loop, by keeping 994 // the value read the first time in a register. 995 while (true) { 996 // We must do this until we get a consistent view of the object. 997 if (FreeChunk::indicatesFreeChunk(p)) { 998 volatile FreeChunk* fc = (volatile FreeChunk*)p; 999 size_t res = fc->size(); 1000 1001 // Bugfix for systems with weak memory model (PPC64/IA64). The 1002 // block's free bit was set and we have read the size of the 1003 // block. Acquire and check the free bit again. If the block is 1004 // still free, the read size is correct. 1005 OrderAccess::acquire(); 1006 1007 // If the object is still a free chunk, return the size, else it 1008 // has been allocated so try again. 1009 if (FreeChunk::indicatesFreeChunk(p)) { 1010 assert(res != 0, "Block size should not be 0"); 1011 return res; 1012 } 1013 } else { 1014 // must read from what 'p' points to in each loop. 1015 Klass* k = ((volatile oopDesc*)p)->klass_or_null(); 1016 if (k != NULL) { 1017 assert(k->is_klass(), "Should really be klass oop."); 1018 oop o = (oop)p; 1019 assert(o->is_oop(true /* ignore mark word */), "Should be an oop."); 1020 1021 // Bugfix for systems with weak memory model (PPC64/IA64). 1022 // The object o may be an array. Acquire to make sure that the array 1023 // size (third word) is consistent. 1024 OrderAccess::acquire(); 1025 1026 size_t res = o->size_given_klass(k); 1027 res = adjustObjectSize(res); 1028 assert(res != 0, "Block size should not be 0"); 1029 return res; 1030 } 1031 } 1032 } 1033 } 1034 1035 // TODO: Now that is_parsable is gone, we should combine these two functions. 1036 // A variant of the above that uses the Printezis bits for 1037 // unparsable but allocated objects. This avoids any possible 1038 // stalls waiting for mutators to initialize objects, and is 1039 // thus potentially faster than the variant above. However, 1040 // this variant may return a zero size for a block that is 1041 // under mutation and for which a consistent size cannot be 1042 // inferred without stalling; see CMSCollector::block_size_if_printezis_bits(). 1043 size_t CompactibleFreeListSpace::block_size_no_stall(HeapWord* p, 1044 const CMSCollector* c) 1045 const { 1046 assert(MemRegion(bottom(), end()).contains(p), "p not in space"); 1047 // This must be volatile, or else there is a danger that the compiler 1048 // will compile the code below into a sometimes-infinite loop, by keeping 1049 // the value read the first time in a register. 1050 DEBUG_ONLY(uint loops = 0;) 1051 while (true) { 1052 // We must do this until we get a consistent view of the object. 1053 if (FreeChunk::indicatesFreeChunk(p)) { 1054 volatile FreeChunk* fc = (volatile FreeChunk*)p; 1055 size_t res = fc->size(); 1056 1057 // Bugfix for systems with weak memory model (PPC64/IA64). The 1058 // free bit of the block was set and we have read the size of 1059 // the block. Acquire and check the free bit again. If the 1060 // block is still free, the read size is correct. 1061 OrderAccess::acquire(); 1062 1063 if (FreeChunk::indicatesFreeChunk(p)) { 1064 assert(res != 0, "Block size should not be 0"); 1065 assert(loops == 0, "Should be 0"); 1066 return res; 1067 } 1068 } else { 1069 // must read from what 'p' points to in each loop. 1070 Klass* k = ((volatile oopDesc*)p)->klass_or_null(); 1071 // We trust the size of any object that has a non-NULL 1072 // klass and (for those in the perm gen) is parsable 1073 // -- irrespective of its conc_safe-ty. 1074 if (k != NULL) { 1075 assert(k->is_klass(), "Should really be klass oop."); 1076 oop o = (oop)p; 1077 assert(o->is_oop(), "Should be an oop"); 1078 1079 // Bugfix for systems with weak memory model (PPC64/IA64). 1080 // The object o may be an array. Acquire to make sure that the array 1081 // size (third word) is consistent. 1082 OrderAccess::acquire(); 1083 1084 size_t res = o->size_given_klass(k); 1085 res = adjustObjectSize(res); 1086 assert(res != 0, "Block size should not be 0"); 1087 return res; 1088 } else { 1089 // May return 0 if P-bits not present. 1090 return c->block_size_if_printezis_bits(p); 1091 } 1092 } 1093 assert(loops == 0, "Can loop at most once"); 1094 DEBUG_ONLY(loops++;) 1095 } 1096 } 1097 1098 size_t CompactibleFreeListSpace::block_size_nopar(const HeapWord* p) const { 1099 NOT_PRODUCT(verify_objects_initialized()); 1100 assert(MemRegion(bottom(), end()).contains(p), "p not in space"); 1101 FreeChunk* fc = (FreeChunk*)p; 1102 if (fc->is_free()) { 1103 return fc->size(); 1104 } else { 1105 // Ignore mark word because this may be a recently promoted 1106 // object whose mark word is used to chain together grey 1107 // objects (the last one would have a null value). 1108 assert(oop(p)->is_oop(true), "Should be an oop"); 1109 return adjustObjectSize(oop(p)->size()); 1110 } 1111 } 1112 1113 // This implementation assumes that the property of "being an object" is 1114 // stable. But being a free chunk may not be (because of parallel 1115 // promotion.) 1116 bool CompactibleFreeListSpace::block_is_obj(const HeapWord* p) const { 1117 FreeChunk* fc = (FreeChunk*)p; 1118 assert(is_in_reserved(p), "Should be in space"); 1119 // When doing a mark-sweep-compact of the CMS generation, this 1120 // assertion may fail because prepare_for_compaction() uses 1121 // space that is garbage to maintain information on ranges of 1122 // live objects so that these live ranges can be moved as a whole. 1123 // Comment out this assertion until that problem can be solved 1124 // (i.e., that the block start calculation may look at objects 1125 // at address below "p" in finding the object that contains "p" 1126 // and those objects (if garbage) may have been modified to hold 1127 // live range information. 1128 // assert(CollectedHeap::use_parallel_gc_threads() || _bt.block_start(p) == p, 1129 // "Should be a block boundary"); 1130 if (FreeChunk::indicatesFreeChunk(p)) return false; 1131 Klass* k = oop(p)->klass_or_null(); 1132 if (k != NULL) { 1133 // Ignore mark word because it may have been used to 1134 // chain together promoted objects (the last one 1135 // would have a null value). 1136 assert(oop(p)->is_oop(true), "Should be an oop"); 1137 return true; 1138 } else { 1139 return false; // Was not an object at the start of collection. 1140 } 1141 } 1142 1143 // Check if the object is alive. This fact is checked either by consulting 1144 // the main marking bitmap in the sweeping phase or, if it's a permanent 1145 // generation and we're not in the sweeping phase, by checking the 1146 // perm_gen_verify_bit_map where we store the "deadness" information if 1147 // we did not sweep the perm gen in the most recent previous GC cycle. 1148 bool CompactibleFreeListSpace::obj_is_alive(const HeapWord* p) const { 1149 assert(SafepointSynchronize::is_at_safepoint() || !is_init_completed(), 1150 "Else races are possible"); 1151 assert(block_is_obj(p), "The address should point to an object"); 1152 1153 // If we're sweeping, we use object liveness information from the main bit map 1154 // for both perm gen and old gen. 1155 // We don't need to lock the bitmap (live_map or dead_map below), because 1156 // EITHER we are in the middle of the sweeping phase, and the 1157 // main marking bit map (live_map below) is locked, 1158 // OR we're in other phases and perm_gen_verify_bit_map (dead_map below) 1159 // is stable, because it's mutated only in the sweeping phase. 1160 // NOTE: This method is also used by jmap where, if class unloading is 1161 // off, the results can return "false" for legitimate perm objects, 1162 // when we are not in the midst of a sweeping phase, which can result 1163 // in jmap not reporting certain perm gen objects. This will be moot 1164 // if/when the perm gen goes away in the future. 1165 if (_collector->abstract_state() == CMSCollector::Sweeping) { 1166 CMSBitMap* live_map = _collector->markBitMap(); 1167 return live_map->par_isMarked((HeapWord*) p); 1168 } 1169 return true; 1170 } 1171 1172 bool CompactibleFreeListSpace::block_is_obj_nopar(const HeapWord* p) const { 1173 FreeChunk* fc = (FreeChunk*)p; 1174 assert(is_in_reserved(p), "Should be in space"); 1175 assert(_bt.block_start(p) == p, "Should be a block boundary"); 1176 if (!fc->is_free()) { 1177 // Ignore mark word because it may have been used to 1178 // chain together promoted objects (the last one 1179 // would have a null value). 1180 assert(oop(p)->is_oop(true), "Should be an oop"); 1181 return true; 1182 } 1183 return false; 1184 } 1185 1186 // "MT-safe but not guaranteed MT-precise" (TM); you may get an 1187 // approximate answer if you don't hold the freelistlock when you call this. 1188 size_t CompactibleFreeListSpace::totalSizeInIndexedFreeLists() const { 1189 size_t size = 0; 1190 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 1191 debug_only( 1192 // We may be calling here without the lock in which case we 1193 // won't do this modest sanity check. 1194 if (freelistLock()->owned_by_self()) { 1195 size_t total_list_size = 0; 1196 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 1197 fc = fc->next()) { 1198 total_list_size += i; 1199 } 1200 assert(total_list_size == i * _indexedFreeList[i].count(), 1201 "Count in list is incorrect"); 1202 } 1203 ) 1204 size += i * _indexedFreeList[i].count(); 1205 } 1206 return size; 1207 } 1208 1209 HeapWord* CompactibleFreeListSpace::par_allocate(size_t size) { 1210 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 1211 return allocate(size); 1212 } 1213 1214 HeapWord* 1215 CompactibleFreeListSpace::getChunkFromSmallLinearAllocBlockRemainder(size_t size) { 1216 return getChunkFromLinearAllocBlockRemainder(&_smallLinearAllocBlock, size); 1217 } 1218 1219 HeapWord* CompactibleFreeListSpace::allocate(size_t size) { 1220 assert_lock_strong(freelistLock()); 1221 HeapWord* res = NULL; 1222 assert(size == adjustObjectSize(size), 1223 "use adjustObjectSize() before calling into allocate()"); 1224 1225 if (_adaptive_freelists) { 1226 res = allocate_adaptive_freelists(size); 1227 } else { // non-adaptive free lists 1228 res = allocate_non_adaptive_freelists(size); 1229 } 1230 1231 if (res != NULL) { 1232 // check that res does lie in this space! 1233 assert(is_in_reserved(res), "Not in this space!"); 1234 assert(is_aligned((void*)res), "alignment check"); 1235 1236 FreeChunk* fc = (FreeChunk*)res; 1237 fc->markNotFree(); 1238 assert(!fc->is_free(), "shouldn't be marked free"); 1239 assert(oop(fc)->klass_or_null() == NULL, "should look uninitialized"); 1240 // Verify that the block offset table shows this to 1241 // be a single block, but not one which is unallocated. 1242 _bt.verify_single_block(res, size); 1243 _bt.verify_not_unallocated(res, size); 1244 // mangle a just allocated object with a distinct pattern. 1245 debug_only(fc->mangleAllocated(size)); 1246 } 1247 1248 return res; 1249 } 1250 1251 HeapWord* CompactibleFreeListSpace::allocate_non_adaptive_freelists(size_t size) { 1252 HeapWord* res = NULL; 1253 // try and use linear allocation for smaller blocks 1254 if (size < _smallLinearAllocBlock._allocation_size_limit) { 1255 // if successful, the following also adjusts block offset table 1256 res = getChunkFromSmallLinearAllocBlock(size); 1257 } 1258 // Else triage to indexed lists for smaller sizes 1259 if (res == NULL) { 1260 if (size < SmallForDictionary) { 1261 res = (HeapWord*) getChunkFromIndexedFreeList(size); 1262 } else { 1263 // else get it from the big dictionary; if even this doesn't 1264 // work we are out of luck. 1265 res = (HeapWord*)getChunkFromDictionaryExact(size); 1266 } 1267 } 1268 1269 return res; 1270 } 1271 1272 HeapWord* CompactibleFreeListSpace::allocate_adaptive_freelists(size_t size) { 1273 assert_lock_strong(freelistLock()); 1274 HeapWord* res = NULL; 1275 assert(size == adjustObjectSize(size), 1276 "use adjustObjectSize() before calling into allocate()"); 1277 1278 // Strategy 1279 // if small 1280 // exact size from small object indexed list if small 1281 // small or large linear allocation block (linAB) as appropriate 1282 // take from lists of greater sized chunks 1283 // else 1284 // dictionary 1285 // small or large linear allocation block if it has the space 1286 // Try allocating exact size from indexTable first 1287 if (size < IndexSetSize) { 1288 res = (HeapWord*) getChunkFromIndexedFreeList(size); 1289 if(res != NULL) { 1290 assert(res != (HeapWord*)_indexedFreeList[size].head(), 1291 "Not removed from free list"); 1292 // no block offset table adjustment is necessary on blocks in 1293 // the indexed lists. 1294 1295 // Try allocating from the small LinAB 1296 } else if (size < _smallLinearAllocBlock._allocation_size_limit && 1297 (res = getChunkFromSmallLinearAllocBlock(size)) != NULL) { 1298 // if successful, the above also adjusts block offset table 1299 // Note that this call will refill the LinAB to 1300 // satisfy the request. This is different that 1301 // evm. 1302 // Don't record chunk off a LinAB? smallSplitBirth(size); 1303 } else { 1304 // Raid the exact free lists larger than size, even if they are not 1305 // overpopulated. 1306 res = (HeapWord*) getChunkFromGreater(size); 1307 } 1308 } else { 1309 // Big objects get allocated directly from the dictionary. 1310 res = (HeapWord*) getChunkFromDictionaryExact(size); 1311 if (res == NULL) { 1312 // Try hard not to fail since an allocation failure will likely 1313 // trigger a synchronous GC. Try to get the space from the 1314 // allocation blocks. 1315 res = getChunkFromSmallLinearAllocBlockRemainder(size); 1316 } 1317 } 1318 1319 return res; 1320 } 1321 1322 // A worst-case estimate of the space required (in HeapWords) to expand the heap 1323 // when promoting obj. 1324 size_t CompactibleFreeListSpace::expansionSpaceRequired(size_t obj_size) const { 1325 // Depending on the object size, expansion may require refilling either a 1326 // bigLAB or a smallLAB plus refilling a PromotionInfo object. MinChunkSize 1327 // is added because the dictionary may over-allocate to avoid fragmentation. 1328 size_t space = obj_size; 1329 if (!_adaptive_freelists) { 1330 space = MAX2(space, _smallLinearAllocBlock._refillSize); 1331 } 1332 space += _promoInfo.refillSize() + 2 * MinChunkSize; 1333 return space; 1334 } 1335 1336 FreeChunk* CompactibleFreeListSpace::getChunkFromGreater(size_t numWords) { 1337 FreeChunk* ret; 1338 1339 assert(numWords >= MinChunkSize, "Size is less than minimum"); 1340 assert(linearAllocationWouldFail() || bestFitFirst(), 1341 "Should not be here"); 1342 1343 size_t i; 1344 size_t currSize = numWords + MinChunkSize; 1345 assert(currSize % MinObjAlignment == 0, "currSize should be aligned"); 1346 for (i = currSize; i < IndexSetSize; i += IndexSetStride) { 1347 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[i]; 1348 if (fl->head()) { 1349 ret = getFromListGreater(fl, numWords); 1350 assert(ret == NULL || ret->is_free(), "Should be returning a free chunk"); 1351 return ret; 1352 } 1353 } 1354 1355 currSize = MAX2((size_t)SmallForDictionary, 1356 (size_t)(numWords + MinChunkSize)); 1357 1358 /* Try to get a chunk that satisfies request, while avoiding 1359 fragmentation that can't be handled. */ 1360 { 1361 ret = dictionary()->get_chunk(currSize); 1362 if (ret != NULL) { 1363 assert(ret->size() - numWords >= MinChunkSize, 1364 "Chunk is too small"); 1365 _bt.allocated((HeapWord*)ret, ret->size()); 1366 /* Carve returned chunk. */ 1367 (void) splitChunkAndReturnRemainder(ret, numWords); 1368 /* Label this as no longer a free chunk. */ 1369 assert(ret->is_free(), "This chunk should be free"); 1370 ret->link_prev(NULL); 1371 } 1372 assert(ret == NULL || ret->is_free(), "Should be returning a free chunk"); 1373 return ret; 1374 } 1375 ShouldNotReachHere(); 1376 } 1377 1378 bool CompactibleFreeListSpace::verifyChunkInIndexedFreeLists(FreeChunk* fc) const { 1379 assert(fc->size() < IndexSetSize, "Size of chunk is too large"); 1380 return _indexedFreeList[fc->size()].verify_chunk_in_free_list(fc); 1381 } 1382 1383 bool CompactibleFreeListSpace::verify_chunk_is_linear_alloc_block(FreeChunk* fc) const { 1384 assert((_smallLinearAllocBlock._ptr != (HeapWord*)fc) || 1385 (_smallLinearAllocBlock._word_size == fc->size()), 1386 "Linear allocation block shows incorrect size"); 1387 return ((_smallLinearAllocBlock._ptr == (HeapWord*)fc) && 1388 (_smallLinearAllocBlock._word_size == fc->size())); 1389 } 1390 1391 // Check if the purported free chunk is present either as a linear 1392 // allocation block, the size-indexed table of (smaller) free blocks, 1393 // or the larger free blocks kept in the binary tree dictionary. 1394 bool CompactibleFreeListSpace::verify_chunk_in_free_list(FreeChunk* fc) const { 1395 if (verify_chunk_is_linear_alloc_block(fc)) { 1396 return true; 1397 } else if (fc->size() < IndexSetSize) { 1398 return verifyChunkInIndexedFreeLists(fc); 1399 } else { 1400 return dictionary()->verify_chunk_in_free_list(fc); 1401 } 1402 } 1403 1404 #ifndef PRODUCT 1405 void CompactibleFreeListSpace::assert_locked() const { 1406 CMSLockVerifier::assert_locked(freelistLock(), parDictionaryAllocLock()); 1407 } 1408 1409 void CompactibleFreeListSpace::assert_locked(const Mutex* lock) const { 1410 CMSLockVerifier::assert_locked(lock); 1411 } 1412 #endif 1413 1414 FreeChunk* CompactibleFreeListSpace::allocateScratch(size_t size) { 1415 // In the parallel case, the main thread holds the free list lock 1416 // on behalf the parallel threads. 1417 FreeChunk* fc; 1418 { 1419 // If GC is parallel, this might be called by several threads. 1420 // This should be rare enough that the locking overhead won't affect 1421 // the sequential code. 1422 MutexLockerEx x(parDictionaryAllocLock(), 1423 Mutex::_no_safepoint_check_flag); 1424 fc = getChunkFromDictionary(size); 1425 } 1426 if (fc != NULL) { 1427 fc->dontCoalesce(); 1428 assert(fc->is_free(), "Should be free, but not coalescable"); 1429 // Verify that the block offset table shows this to 1430 // be a single block, but not one which is unallocated. 1431 _bt.verify_single_block((HeapWord*)fc, fc->size()); 1432 _bt.verify_not_unallocated((HeapWord*)fc, fc->size()); 1433 } 1434 return fc; 1435 } 1436 1437 oop CompactibleFreeListSpace::promote(oop obj, size_t obj_size) { 1438 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1439 assert_locked(); 1440 1441 // if we are tracking promotions, then first ensure space for 1442 // promotion (including spooling space for saving header if necessary). 1443 // then allocate and copy, then track promoted info if needed. 1444 // When tracking (see PromotionInfo::track()), the mark word may 1445 // be displaced and in this case restoration of the mark word 1446 // occurs in the (oop_since_save_marks_)iterate phase. 1447 if (_promoInfo.tracking() && !_promoInfo.ensure_spooling_space()) { 1448 return NULL; 1449 } 1450 // Call the allocate(size_t, bool) form directly to avoid the 1451 // additional call through the allocate(size_t) form. Having 1452 // the compile inline the call is problematic because allocate(size_t) 1453 // is a virtual method. 1454 HeapWord* res = allocate(adjustObjectSize(obj_size)); 1455 if (res != NULL) { 1456 Copy::aligned_disjoint_words((HeapWord*)obj, res, obj_size); 1457 // if we should be tracking promotions, do so. 1458 if (_promoInfo.tracking()) { 1459 _promoInfo.track((PromotedObject*)res); 1460 } 1461 } 1462 return oop(res); 1463 } 1464 1465 HeapWord* 1466 CompactibleFreeListSpace::getChunkFromSmallLinearAllocBlock(size_t size) { 1467 assert_locked(); 1468 assert(size >= MinChunkSize, "minimum chunk size"); 1469 assert(size < _smallLinearAllocBlock._allocation_size_limit, 1470 "maximum from smallLinearAllocBlock"); 1471 return getChunkFromLinearAllocBlock(&_smallLinearAllocBlock, size); 1472 } 1473 1474 HeapWord* 1475 CompactibleFreeListSpace::getChunkFromLinearAllocBlock(LinearAllocBlock *blk, 1476 size_t size) { 1477 assert_locked(); 1478 assert(size >= MinChunkSize, "too small"); 1479 HeapWord* res = NULL; 1480 // Try to do linear allocation from blk, making sure that 1481 if (blk->_word_size == 0) { 1482 // We have probably been unable to fill this either in the prologue or 1483 // when it was exhausted at the last linear allocation. Bail out until 1484 // next time. 1485 assert(blk->_ptr == NULL, "consistency check"); 1486 return NULL; 1487 } 1488 assert(blk->_word_size != 0 && blk->_ptr != NULL, "consistency check"); 1489 res = getChunkFromLinearAllocBlockRemainder(blk, size); 1490 if (res != NULL) return res; 1491 1492 // about to exhaust this linear allocation block 1493 if (blk->_word_size == size) { // exactly satisfied 1494 res = blk->_ptr; 1495 _bt.allocated(res, blk->_word_size); 1496 } else if (size + MinChunkSize <= blk->_refillSize) { 1497 size_t sz = blk->_word_size; 1498 // Update _unallocated_block if the size is such that chunk would be 1499 // returned to the indexed free list. All other chunks in the indexed 1500 // free lists are allocated from the dictionary so that _unallocated_block 1501 // has already been adjusted for them. Do it here so that the cost 1502 // for all chunks added back to the indexed free lists. 1503 if (sz < SmallForDictionary) { 1504 _bt.allocated(blk->_ptr, sz); 1505 } 1506 // Return the chunk that isn't big enough, and then refill below. 1507 addChunkToFreeLists(blk->_ptr, sz); 1508 split_birth(sz); 1509 // Don't keep statistics on adding back chunk from a LinAB. 1510 } else { 1511 // A refilled block would not satisfy the request. 1512 return NULL; 1513 } 1514 1515 blk->_ptr = NULL; blk->_word_size = 0; 1516 refillLinearAllocBlock(blk); 1517 assert(blk->_ptr == NULL || blk->_word_size >= size + MinChunkSize, 1518 "block was replenished"); 1519 if (res != NULL) { 1520 split_birth(size); 1521 repairLinearAllocBlock(blk); 1522 } else if (blk->_ptr != NULL) { 1523 res = blk->_ptr; 1524 size_t blk_size = blk->_word_size; 1525 blk->_word_size -= size; 1526 blk->_ptr += size; 1527 split_birth(size); 1528 repairLinearAllocBlock(blk); 1529 // Update BOT last so that other (parallel) GC threads see a consistent 1530 // view of the BOT and free blocks. 1531 // Above must occur before BOT is updated below. 1532 OrderAccess::storestore(); 1533 _bt.split_block(res, blk_size, size); // adjust block offset table 1534 } 1535 return res; 1536 } 1537 1538 HeapWord* CompactibleFreeListSpace::getChunkFromLinearAllocBlockRemainder( 1539 LinearAllocBlock* blk, 1540 size_t size) { 1541 assert_locked(); 1542 assert(size >= MinChunkSize, "too small"); 1543 1544 HeapWord* res = NULL; 1545 // This is the common case. Keep it simple. 1546 if (blk->_word_size >= size + MinChunkSize) { 1547 assert(blk->_ptr != NULL, "consistency check"); 1548 res = blk->_ptr; 1549 // Note that the BOT is up-to-date for the linAB before allocation. It 1550 // indicates the start of the linAB. The split_block() updates the 1551 // BOT for the linAB after the allocation (indicates the start of the 1552 // next chunk to be allocated). 1553 size_t blk_size = blk->_word_size; 1554 blk->_word_size -= size; 1555 blk->_ptr += size; 1556 split_birth(size); 1557 repairLinearAllocBlock(blk); 1558 // Update BOT last so that other (parallel) GC threads see a consistent 1559 // view of the BOT and free blocks. 1560 // Above must occur before BOT is updated below. 1561 OrderAccess::storestore(); 1562 _bt.split_block(res, blk_size, size); // adjust block offset table 1563 _bt.allocated(res, size); 1564 } 1565 return res; 1566 } 1567 1568 FreeChunk* 1569 CompactibleFreeListSpace::getChunkFromIndexedFreeList(size_t size) { 1570 assert_locked(); 1571 assert(size < SmallForDictionary, "just checking"); 1572 FreeChunk* res; 1573 res = _indexedFreeList[size].get_chunk_at_head(); 1574 if (res == NULL) { 1575 res = getChunkFromIndexedFreeListHelper(size); 1576 } 1577 _bt.verify_not_unallocated((HeapWord*) res, size); 1578 assert(res == NULL || res->size() == size, "Incorrect block size"); 1579 return res; 1580 } 1581 1582 FreeChunk* 1583 CompactibleFreeListSpace::getChunkFromIndexedFreeListHelper(size_t size, 1584 bool replenish) { 1585 assert_locked(); 1586 FreeChunk* fc = NULL; 1587 if (size < SmallForDictionary) { 1588 assert(_indexedFreeList[size].head() == NULL || 1589 _indexedFreeList[size].surplus() <= 0, 1590 "List for this size should be empty or under populated"); 1591 // Try best fit in exact lists before replenishing the list 1592 if (!bestFitFirst() || (fc = bestFitSmall(size)) == NULL) { 1593 // Replenish list. 1594 // 1595 // Things tried that failed. 1596 // Tried allocating out of the two LinAB's first before 1597 // replenishing lists. 1598 // Tried small linAB of size 256 (size in indexed list) 1599 // and replenishing indexed lists from the small linAB. 1600 // 1601 FreeChunk* newFc = NULL; 1602 const size_t replenish_size = CMSIndexedFreeListReplenish * size; 1603 if (replenish_size < SmallForDictionary) { 1604 // Do not replenish from an underpopulated size. 1605 if (_indexedFreeList[replenish_size].surplus() > 0 && 1606 _indexedFreeList[replenish_size].head() != NULL) { 1607 newFc = _indexedFreeList[replenish_size].get_chunk_at_head(); 1608 } else if (bestFitFirst()) { 1609 newFc = bestFitSmall(replenish_size); 1610 } 1611 } 1612 if (newFc == NULL && replenish_size > size) { 1613 assert(CMSIndexedFreeListReplenish > 1, "ctl pt invariant"); 1614 newFc = getChunkFromIndexedFreeListHelper(replenish_size, false); 1615 } 1616 // Note: The stats update re split-death of block obtained above 1617 // will be recorded below precisely when we know we are going to 1618 // be actually splitting it into more than one pieces below. 1619 if (newFc != NULL) { 1620 if (replenish || CMSReplenishIntermediate) { 1621 // Replenish this list and return one block to caller. 1622 size_t i; 1623 FreeChunk *curFc, *nextFc; 1624 size_t num_blk = newFc->size() / size; 1625 assert(num_blk >= 1, "Smaller than requested?"); 1626 assert(newFc->size() % size == 0, "Should be integral multiple of request"); 1627 if (num_blk > 1) { 1628 // we are sure we will be splitting the block just obtained 1629 // into multiple pieces; record the split-death of the original 1630 splitDeath(replenish_size); 1631 } 1632 // carve up and link blocks 0, ..., num_blk - 2 1633 // The last chunk is not added to the lists but is returned as the 1634 // free chunk. 1635 for (curFc = newFc, nextFc = (FreeChunk*)((HeapWord*)curFc + size), 1636 i = 0; 1637 i < (num_blk - 1); 1638 curFc = nextFc, nextFc = (FreeChunk*)((HeapWord*)nextFc + size), 1639 i++) { 1640 curFc->set_size(size); 1641 // Don't record this as a return in order to try and 1642 // determine the "returns" from a GC. 1643 _bt.verify_not_unallocated((HeapWord*) fc, size); 1644 _indexedFreeList[size].return_chunk_at_tail(curFc, false); 1645 _bt.mark_block((HeapWord*)curFc, size); 1646 split_birth(size); 1647 // Don't record the initial population of the indexed list 1648 // as a split birth. 1649 } 1650 1651 // check that the arithmetic was OK above 1652 assert((HeapWord*)nextFc == (HeapWord*)newFc + num_blk*size, 1653 "inconsistency in carving newFc"); 1654 curFc->set_size(size); 1655 _bt.mark_block((HeapWord*)curFc, size); 1656 split_birth(size); 1657 fc = curFc; 1658 } else { 1659 // Return entire block to caller 1660 fc = newFc; 1661 } 1662 } 1663 } 1664 } else { 1665 // Get a free chunk from the free chunk dictionary to be returned to 1666 // replenish the indexed free list. 1667 fc = getChunkFromDictionaryExact(size); 1668 } 1669 // assert(fc == NULL || fc->is_free(), "Should be returning a free chunk"); 1670 return fc; 1671 } 1672 1673 FreeChunk* 1674 CompactibleFreeListSpace::getChunkFromDictionary(size_t size) { 1675 assert_locked(); 1676 FreeChunk* fc = _dictionary->get_chunk(size, 1677 FreeBlockDictionary<FreeChunk>::atLeast); 1678 if (fc == NULL) { 1679 return NULL; 1680 } 1681 _bt.allocated((HeapWord*)fc, fc->size()); 1682 if (fc->size() >= size + MinChunkSize) { 1683 fc = splitChunkAndReturnRemainder(fc, size); 1684 } 1685 assert(fc->size() >= size, "chunk too small"); 1686 assert(fc->size() < size + MinChunkSize, "chunk too big"); 1687 _bt.verify_single_block((HeapWord*)fc, fc->size()); 1688 return fc; 1689 } 1690 1691 FreeChunk* 1692 CompactibleFreeListSpace::getChunkFromDictionaryExact(size_t size) { 1693 assert_locked(); 1694 FreeChunk* fc = _dictionary->get_chunk(size, 1695 FreeBlockDictionary<FreeChunk>::atLeast); 1696 if (fc == NULL) { 1697 return fc; 1698 } 1699 _bt.allocated((HeapWord*)fc, fc->size()); 1700 if (fc->size() == size) { 1701 _bt.verify_single_block((HeapWord*)fc, size); 1702 return fc; 1703 } 1704 assert(fc->size() > size, "get_chunk() guarantee"); 1705 if (fc->size() < size + MinChunkSize) { 1706 // Return the chunk to the dictionary and go get a bigger one. 1707 returnChunkToDictionary(fc); 1708 fc = _dictionary->get_chunk(size + MinChunkSize, 1709 FreeBlockDictionary<FreeChunk>::atLeast); 1710 if (fc == NULL) { 1711 return NULL; 1712 } 1713 _bt.allocated((HeapWord*)fc, fc->size()); 1714 } 1715 assert(fc->size() >= size + MinChunkSize, "tautology"); 1716 fc = splitChunkAndReturnRemainder(fc, size); 1717 assert(fc->size() == size, "chunk is wrong size"); 1718 _bt.verify_single_block((HeapWord*)fc, size); 1719 return fc; 1720 } 1721 1722 void 1723 CompactibleFreeListSpace::returnChunkToDictionary(FreeChunk* chunk) { 1724 assert_locked(); 1725 1726 size_t size = chunk->size(); 1727 _bt.verify_single_block((HeapWord*)chunk, size); 1728 // adjust _unallocated_block downward, as necessary 1729 _bt.freed((HeapWord*)chunk, size); 1730 _dictionary->return_chunk(chunk); 1731 #ifndef PRODUCT 1732 if (CMSCollector::abstract_state() != CMSCollector::Sweeping) { 1733 TreeChunk<FreeChunk, AdaptiveFreeList>* tc = TreeChunk<FreeChunk, AdaptiveFreeList>::as_TreeChunk(chunk); 1734 TreeList<FreeChunk, AdaptiveFreeList>* tl = tc->list(); 1735 tl->verify_stats(); 1736 } 1737 #endif // PRODUCT 1738 } 1739 1740 void 1741 CompactibleFreeListSpace::returnChunkToFreeList(FreeChunk* fc) { 1742 assert_locked(); 1743 size_t size = fc->size(); 1744 _bt.verify_single_block((HeapWord*) fc, size); 1745 _bt.verify_not_unallocated((HeapWord*) fc, size); 1746 if (_adaptive_freelists) { 1747 _indexedFreeList[size].return_chunk_at_tail(fc); 1748 } else { 1749 _indexedFreeList[size].return_chunk_at_head(fc); 1750 } 1751 #ifndef PRODUCT 1752 if (CMSCollector::abstract_state() != CMSCollector::Sweeping) { 1753 _indexedFreeList[size].verify_stats(); 1754 } 1755 #endif // PRODUCT 1756 } 1757 1758 // Add chunk to end of last block -- if it's the largest 1759 // block -- and update BOT and census data. We would 1760 // of course have preferred to coalesce it with the 1761 // last block, but it's currently less expensive to find the 1762 // largest block than it is to find the last. 1763 void 1764 CompactibleFreeListSpace::addChunkToFreeListsAtEndRecordingStats( 1765 HeapWord* chunk, size_t size) { 1766 // check that the chunk does lie in this space! 1767 assert(chunk != NULL && is_in_reserved(chunk), "Not in this space!"); 1768 // One of the parallel gc task threads may be here 1769 // whilst others are allocating. 1770 Mutex* lock = NULL; 1771 if (ParallelGCThreads != 0) { 1772 lock = &_parDictionaryAllocLock; 1773 } 1774 FreeChunk* ec; 1775 { 1776 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 1777 ec = dictionary()->find_largest_dict(); // get largest block 1778 if (ec != NULL && ec->end() == (uintptr_t*) chunk) { 1779 // It's a coterminal block - we can coalesce. 1780 size_t old_size = ec->size(); 1781 coalDeath(old_size); 1782 removeChunkFromDictionary(ec); 1783 size += old_size; 1784 } else { 1785 ec = (FreeChunk*)chunk; 1786 } 1787 } 1788 ec->set_size(size); 1789 debug_only(ec->mangleFreed(size)); 1790 if (size < SmallForDictionary && ParallelGCThreads != 0) { 1791 lock = _indexedFreeListParLocks[size]; 1792 } 1793 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 1794 addChunkAndRepairOffsetTable((HeapWord*)ec, size, true); 1795 // record the birth under the lock since the recording involves 1796 // manipulation of the list on which the chunk lives and 1797 // if the chunk is allocated and is the last on the list, 1798 // the list can go away. 1799 coalBirth(size); 1800 } 1801 1802 void 1803 CompactibleFreeListSpace::addChunkToFreeLists(HeapWord* chunk, 1804 size_t size) { 1805 // check that the chunk does lie in this space! 1806 assert(chunk != NULL && is_in_reserved(chunk), "Not in this space!"); 1807 assert_locked(); 1808 _bt.verify_single_block(chunk, size); 1809 1810 FreeChunk* fc = (FreeChunk*) chunk; 1811 fc->set_size(size); 1812 debug_only(fc->mangleFreed(size)); 1813 if (size < SmallForDictionary) { 1814 returnChunkToFreeList(fc); 1815 } else { 1816 returnChunkToDictionary(fc); 1817 } 1818 } 1819 1820 void 1821 CompactibleFreeListSpace::addChunkAndRepairOffsetTable(HeapWord* chunk, 1822 size_t size, bool coalesced) { 1823 assert_locked(); 1824 assert(chunk != NULL, "null chunk"); 1825 if (coalesced) { 1826 // repair BOT 1827 _bt.single_block(chunk, size); 1828 } 1829 addChunkToFreeLists(chunk, size); 1830 } 1831 1832 // We _must_ find the purported chunk on our free lists; 1833 // we assert if we don't. 1834 void 1835 CompactibleFreeListSpace::removeFreeChunkFromFreeLists(FreeChunk* fc) { 1836 size_t size = fc->size(); 1837 assert_locked(); 1838 debug_only(verifyFreeLists()); 1839 if (size < SmallForDictionary) { 1840 removeChunkFromIndexedFreeList(fc); 1841 } else { 1842 removeChunkFromDictionary(fc); 1843 } 1844 _bt.verify_single_block((HeapWord*)fc, size); 1845 debug_only(verifyFreeLists()); 1846 } 1847 1848 void 1849 CompactibleFreeListSpace::removeChunkFromDictionary(FreeChunk* fc) { 1850 size_t size = fc->size(); 1851 assert_locked(); 1852 assert(fc != NULL, "null chunk"); 1853 _bt.verify_single_block((HeapWord*)fc, size); 1854 _dictionary->remove_chunk(fc); 1855 // adjust _unallocated_block upward, as necessary 1856 _bt.allocated((HeapWord*)fc, size); 1857 } 1858 1859 void 1860 CompactibleFreeListSpace::removeChunkFromIndexedFreeList(FreeChunk* fc) { 1861 assert_locked(); 1862 size_t size = fc->size(); 1863 _bt.verify_single_block((HeapWord*)fc, size); 1864 NOT_PRODUCT( 1865 if (FLSVerifyIndexTable) { 1866 verifyIndexedFreeList(size); 1867 } 1868 ) 1869 _indexedFreeList[size].remove_chunk(fc); 1870 NOT_PRODUCT( 1871 if (FLSVerifyIndexTable) { 1872 verifyIndexedFreeList(size); 1873 } 1874 ) 1875 } 1876 1877 FreeChunk* CompactibleFreeListSpace::bestFitSmall(size_t numWords) { 1878 /* A hint is the next larger size that has a surplus. 1879 Start search at a size large enough to guarantee that 1880 the excess is >= MIN_CHUNK. */ 1881 size_t start = align_object_size(numWords + MinChunkSize); 1882 if (start < IndexSetSize) { 1883 AdaptiveFreeList<FreeChunk>* it = _indexedFreeList; 1884 size_t hint = _indexedFreeList[start].hint(); 1885 while (hint < IndexSetSize) { 1886 assert(hint % MinObjAlignment == 0, "hint should be aligned"); 1887 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[hint]; 1888 if (fl->surplus() > 0 && fl->head() != NULL) { 1889 // Found a list with surplus, reset original hint 1890 // and split out a free chunk which is returned. 1891 _indexedFreeList[start].set_hint(hint); 1892 FreeChunk* res = getFromListGreater(fl, numWords); 1893 assert(res == NULL || res->is_free(), 1894 "Should be returning a free chunk"); 1895 return res; 1896 } 1897 hint = fl->hint(); /* keep looking */ 1898 } 1899 /* None found. */ 1900 it[start].set_hint(IndexSetSize); 1901 } 1902 return NULL; 1903 } 1904 1905 /* Requires fl->size >= numWords + MinChunkSize */ 1906 FreeChunk* CompactibleFreeListSpace::getFromListGreater(AdaptiveFreeList<FreeChunk>* fl, 1907 size_t numWords) { 1908 FreeChunk *curr = fl->head(); 1909 size_t oldNumWords = curr->size(); 1910 assert(numWords >= MinChunkSize, "Word size is too small"); 1911 assert(curr != NULL, "List is empty"); 1912 assert(oldNumWords >= numWords + MinChunkSize, 1913 "Size of chunks in the list is too small"); 1914 1915 fl->remove_chunk(curr); 1916 // recorded indirectly by splitChunkAndReturnRemainder - 1917 // smallSplit(oldNumWords, numWords); 1918 FreeChunk* new_chunk = splitChunkAndReturnRemainder(curr, numWords); 1919 // Does anything have to be done for the remainder in terms of 1920 // fixing the card table? 1921 assert(new_chunk == NULL || new_chunk->is_free(), 1922 "Should be returning a free chunk"); 1923 return new_chunk; 1924 } 1925 1926 FreeChunk* 1927 CompactibleFreeListSpace::splitChunkAndReturnRemainder(FreeChunk* chunk, 1928 size_t new_size) { 1929 assert_locked(); 1930 size_t size = chunk->size(); 1931 assert(size > new_size, "Split from a smaller block?"); 1932 assert(is_aligned(chunk), "alignment problem"); 1933 assert(size == adjustObjectSize(size), "alignment problem"); 1934 size_t rem_size = size - new_size; 1935 assert(rem_size == adjustObjectSize(rem_size), "alignment problem"); 1936 assert(rem_size >= MinChunkSize, "Free chunk smaller than minimum"); 1937 FreeChunk* ffc = (FreeChunk*)((HeapWord*)chunk + new_size); 1938 assert(is_aligned(ffc), "alignment problem"); 1939 ffc->set_size(rem_size); 1940 ffc->link_next(NULL); 1941 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 1942 // Above must occur before BOT is updated below. 1943 // adjust block offset table 1944 OrderAccess::storestore(); 1945 assert(chunk->is_free() && ffc->is_free(), "Error"); 1946 _bt.split_block((HeapWord*)chunk, chunk->size(), new_size); 1947 if (rem_size < SmallForDictionary) { 1948 bool is_par = (SharedHeap::heap()->n_par_threads() > 0); 1949 if (is_par) _indexedFreeListParLocks[rem_size]->lock(); 1950 assert(!is_par || 1951 (SharedHeap::heap()->n_par_threads() == 1952 SharedHeap::heap()->workers()->active_workers()), "Mismatch"); 1953 returnChunkToFreeList(ffc); 1954 split(size, rem_size); 1955 if (is_par) _indexedFreeListParLocks[rem_size]->unlock(); 1956 } else { 1957 returnChunkToDictionary(ffc); 1958 split(size ,rem_size); 1959 } 1960 chunk->set_size(new_size); 1961 return chunk; 1962 } 1963 1964 void 1965 CompactibleFreeListSpace::sweep_completed() { 1966 // Now that space is probably plentiful, refill linear 1967 // allocation blocks as needed. 1968 refillLinearAllocBlocksIfNeeded(); 1969 } 1970 1971 void 1972 CompactibleFreeListSpace::gc_prologue() { 1973 assert_locked(); 1974 if (PrintFLSStatistics != 0) { 1975 gclog_or_tty->print("Before GC:\n"); 1976 reportFreeListStatistics(); 1977 } 1978 refillLinearAllocBlocksIfNeeded(); 1979 } 1980 1981 void 1982 CompactibleFreeListSpace::gc_epilogue() { 1983 assert_locked(); 1984 if (PrintGCDetails && Verbose && !_adaptive_freelists) { 1985 if (_smallLinearAllocBlock._word_size == 0) 1986 warning("CompactibleFreeListSpace(epilogue):: Linear allocation failure"); 1987 } 1988 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); 1989 _promoInfo.stopTrackingPromotions(); 1990 repairLinearAllocationBlocks(); 1991 // Print Space's stats 1992 if (PrintFLSStatistics != 0) { 1993 gclog_or_tty->print("After GC:\n"); 1994 reportFreeListStatistics(); 1995 } 1996 } 1997 1998 // Iteration support, mostly delegated from a CMS generation 1999 2000 void CompactibleFreeListSpace::save_marks() { 2001 assert(Thread::current()->is_VM_thread(), 2002 "Global variable should only be set when single-threaded"); 2003 // Mark the "end" of the used space at the time of this call; 2004 // note, however, that promoted objects from this point 2005 // on are tracked in the _promoInfo below. 2006 set_saved_mark_word(unallocated_block()); 2007 #ifdef ASSERT 2008 // Check the sanity of save_marks() etc. 2009 MemRegion ur = used_region(); 2010 MemRegion urasm = used_region_at_save_marks(); 2011 assert(ur.contains(urasm), 2012 err_msg(" Error at save_marks(): [" PTR_FORMAT "," PTR_FORMAT ")" 2013 " should contain [" PTR_FORMAT "," PTR_FORMAT ")", 2014 ur.start(), ur.end(), urasm.start(), urasm.end())); 2015 #endif 2016 // inform allocator that promotions should be tracked. 2017 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); 2018 _promoInfo.startTrackingPromotions(); 2019 } 2020 2021 bool CompactibleFreeListSpace::no_allocs_since_save_marks() { 2022 assert(_promoInfo.tracking(), "No preceding save_marks?"); 2023 assert(SharedHeap::heap()->n_par_threads() == 0, 2024 "Shouldn't be called if using parallel gc."); 2025 return _promoInfo.noPromotions(); 2026 } 2027 2028 #define CFLS_OOP_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2029 \ 2030 void CompactibleFreeListSpace:: \ 2031 oop_since_save_marks_iterate##nv_suffix(OopClosureType* blk) { \ 2032 assert(SharedHeap::heap()->n_par_threads() == 0, \ 2033 "Shouldn't be called (yet) during parallel part of gc."); \ 2034 _promoInfo.promoted_oops_iterate##nv_suffix(blk); \ 2035 /* \ 2036 * This also restores any displaced headers and removes the elements from \ 2037 * the iteration set as they are processed, so that we have a clean slate \ 2038 * at the end of the iteration. Note, thus, that if new objects are \ 2039 * promoted as a result of the iteration they are iterated over as well. \ 2040 */ \ 2041 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); \ 2042 } 2043 2044 ALL_SINCE_SAVE_MARKS_CLOSURES(CFLS_OOP_SINCE_SAVE_MARKS_DEFN) 2045 2046 bool CompactibleFreeListSpace::linearAllocationWouldFail() const { 2047 return _smallLinearAllocBlock._word_size == 0; 2048 } 2049 2050 void CompactibleFreeListSpace::repairLinearAllocationBlocks() { 2051 // Fix up linear allocation blocks to look like free blocks 2052 repairLinearAllocBlock(&_smallLinearAllocBlock); 2053 } 2054 2055 void CompactibleFreeListSpace::repairLinearAllocBlock(LinearAllocBlock* blk) { 2056 assert_locked(); 2057 if (blk->_ptr != NULL) { 2058 assert(blk->_word_size != 0 && blk->_word_size >= MinChunkSize, 2059 "Minimum block size requirement"); 2060 FreeChunk* fc = (FreeChunk*)(blk->_ptr); 2061 fc->set_size(blk->_word_size); 2062 fc->link_prev(NULL); // mark as free 2063 fc->dontCoalesce(); 2064 assert(fc->is_free(), "just marked it free"); 2065 assert(fc->cantCoalesce(), "just marked it uncoalescable"); 2066 } 2067 } 2068 2069 void CompactibleFreeListSpace::refillLinearAllocBlocksIfNeeded() { 2070 assert_locked(); 2071 if (_smallLinearAllocBlock._ptr == NULL) { 2072 assert(_smallLinearAllocBlock._word_size == 0, 2073 "Size of linAB should be zero if the ptr is NULL"); 2074 // Reset the linAB refill and allocation size limit. 2075 _smallLinearAllocBlock.set(0, 0, 1024*SmallForLinearAlloc, SmallForLinearAlloc); 2076 } 2077 refillLinearAllocBlockIfNeeded(&_smallLinearAllocBlock); 2078 } 2079 2080 void 2081 CompactibleFreeListSpace::refillLinearAllocBlockIfNeeded(LinearAllocBlock* blk) { 2082 assert_locked(); 2083 assert((blk->_ptr == NULL && blk->_word_size == 0) || 2084 (blk->_ptr != NULL && blk->_word_size >= MinChunkSize), 2085 "blk invariant"); 2086 if (blk->_ptr == NULL) { 2087 refillLinearAllocBlock(blk); 2088 } 2089 if (PrintMiscellaneous && Verbose) { 2090 if (blk->_word_size == 0) { 2091 warning("CompactibleFreeListSpace(prologue):: Linear allocation failure"); 2092 } 2093 } 2094 } 2095 2096 void 2097 CompactibleFreeListSpace::refillLinearAllocBlock(LinearAllocBlock* blk) { 2098 assert_locked(); 2099 assert(blk->_word_size == 0 && blk->_ptr == NULL, 2100 "linear allocation block should be empty"); 2101 FreeChunk* fc; 2102 if (blk->_refillSize < SmallForDictionary && 2103 (fc = getChunkFromIndexedFreeList(blk->_refillSize)) != NULL) { 2104 // A linAB's strategy might be to use small sizes to reduce 2105 // fragmentation but still get the benefits of allocation from a 2106 // linAB. 2107 } else { 2108 fc = getChunkFromDictionary(blk->_refillSize); 2109 } 2110 if (fc != NULL) { 2111 blk->_ptr = (HeapWord*)fc; 2112 blk->_word_size = fc->size(); 2113 fc->dontCoalesce(); // to prevent sweeper from sweeping us up 2114 } 2115 } 2116 2117 // Support for concurrent collection policy decisions. 2118 bool CompactibleFreeListSpace::should_concurrent_collect() const { 2119 // In the future we might want to add in frgamentation stats -- 2120 // including erosion of the "mountain" into this decision as well. 2121 return !adaptive_freelists() && linearAllocationWouldFail(); 2122 } 2123 2124 // Support for compaction 2125 2126 void CompactibleFreeListSpace::prepare_for_compaction(CompactPoint* cp) { 2127 SCAN_AND_FORWARD(cp,end,block_is_obj,block_size); 2128 // prepare_for_compaction() uses the space between live objects 2129 // so that later phase can skip dead space quickly. So verification 2130 // of the free lists doesn't work after. 2131 } 2132 2133 #define obj_size(q) adjustObjectSize(oop(q)->size()) 2134 #define adjust_obj_size(s) adjustObjectSize(s) 2135 2136 void CompactibleFreeListSpace::adjust_pointers() { 2137 // In other versions of adjust_pointers(), a bail out 2138 // based on the amount of live data in the generation 2139 // (i.e., if 0, bail out) may be used. 2140 // Cannot test used() == 0 here because the free lists have already 2141 // been mangled by the compaction. 2142 2143 SCAN_AND_ADJUST_POINTERS(adjust_obj_size); 2144 // See note about verification in prepare_for_compaction(). 2145 } 2146 2147 void CompactibleFreeListSpace::compact() { 2148 SCAN_AND_COMPACT(obj_size); 2149 } 2150 2151 // fragmentation_metric = 1 - [sum of (fbs**2) / (sum of fbs)**2] 2152 // where fbs is free block sizes 2153 double CompactibleFreeListSpace::flsFrag() const { 2154 size_t itabFree = totalSizeInIndexedFreeLists(); 2155 double frag = 0.0; 2156 size_t i; 2157 2158 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2159 double sz = i; 2160 frag += _indexedFreeList[i].count() * (sz * sz); 2161 } 2162 2163 double totFree = itabFree + 2164 _dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())); 2165 if (totFree > 0) { 2166 frag = ((frag + _dictionary->sum_of_squared_block_sizes()) / 2167 (totFree * totFree)); 2168 frag = (double)1.0 - frag; 2169 } else { 2170 assert(frag == 0.0, "Follows from totFree == 0"); 2171 } 2172 return frag; 2173 } 2174 2175 void CompactibleFreeListSpace::beginSweepFLCensus( 2176 float inter_sweep_current, 2177 float inter_sweep_estimate, 2178 float intra_sweep_estimate) { 2179 assert_locked(); 2180 size_t i; 2181 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2182 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[i]; 2183 if (PrintFLSStatistics > 1) { 2184 gclog_or_tty->print("size[%d] : ", i); 2185 } 2186 fl->compute_desired(inter_sweep_current, inter_sweep_estimate, intra_sweep_estimate); 2187 fl->set_coal_desired((ssize_t)((double)fl->desired() * CMSSmallCoalSurplusPercent)); 2188 fl->set_before_sweep(fl->count()); 2189 fl->set_bfr_surp(fl->surplus()); 2190 } 2191 _dictionary->begin_sweep_dict_census(CMSLargeCoalSurplusPercent, 2192 inter_sweep_current, 2193 inter_sweep_estimate, 2194 intra_sweep_estimate); 2195 } 2196 2197 void CompactibleFreeListSpace::setFLSurplus() { 2198 assert_locked(); 2199 size_t i; 2200 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2201 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2202 fl->set_surplus(fl->count() - 2203 (ssize_t)((double)fl->desired() * CMSSmallSplitSurplusPercent)); 2204 } 2205 } 2206 2207 void CompactibleFreeListSpace::setFLHints() { 2208 assert_locked(); 2209 size_t i; 2210 size_t h = IndexSetSize; 2211 for (i = IndexSetSize - 1; i != 0; i -= IndexSetStride) { 2212 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2213 fl->set_hint(h); 2214 if (fl->surplus() > 0) { 2215 h = i; 2216 } 2217 } 2218 } 2219 2220 void CompactibleFreeListSpace::clearFLCensus() { 2221 assert_locked(); 2222 size_t i; 2223 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2224 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2225 fl->set_prev_sweep(fl->count()); 2226 fl->set_coal_births(0); 2227 fl->set_coal_deaths(0); 2228 fl->set_split_births(0); 2229 fl->set_split_deaths(0); 2230 } 2231 } 2232 2233 void CompactibleFreeListSpace::endSweepFLCensus(size_t sweep_count) { 2234 if (PrintFLSStatistics > 0) { 2235 HeapWord* largestAddr = (HeapWord*) dictionary()->find_largest_dict(); 2236 gclog_or_tty->print_cr("CMS: Large block " PTR_FORMAT, 2237 largestAddr); 2238 } 2239 setFLSurplus(); 2240 setFLHints(); 2241 if (PrintGC && PrintFLSCensus > 0) { 2242 printFLCensus(sweep_count); 2243 } 2244 clearFLCensus(); 2245 assert_locked(); 2246 _dictionary->end_sweep_dict_census(CMSLargeSplitSurplusPercent); 2247 } 2248 2249 bool CompactibleFreeListSpace::coalOverPopulated(size_t size) { 2250 if (size < SmallForDictionary) { 2251 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2252 return (fl->coal_desired() < 0) || 2253 ((int)fl->count() > fl->coal_desired()); 2254 } else { 2255 return dictionary()->coal_dict_over_populated(size); 2256 } 2257 } 2258 2259 void CompactibleFreeListSpace::smallCoalBirth(size_t size) { 2260 assert(size < SmallForDictionary, "Size too large for indexed list"); 2261 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2262 fl->increment_coal_births(); 2263 fl->increment_surplus(); 2264 } 2265 2266 void CompactibleFreeListSpace::smallCoalDeath(size_t size) { 2267 assert(size < SmallForDictionary, "Size too large for indexed list"); 2268 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2269 fl->increment_coal_deaths(); 2270 fl->decrement_surplus(); 2271 } 2272 2273 void CompactibleFreeListSpace::coalBirth(size_t size) { 2274 if (size < SmallForDictionary) { 2275 smallCoalBirth(size); 2276 } else { 2277 dictionary()->dict_census_update(size, 2278 false /* split */, 2279 true /* birth */); 2280 } 2281 } 2282 2283 void CompactibleFreeListSpace::coalDeath(size_t size) { 2284 if(size < SmallForDictionary) { 2285 smallCoalDeath(size); 2286 } else { 2287 dictionary()->dict_census_update(size, 2288 false /* split */, 2289 false /* birth */); 2290 } 2291 } 2292 2293 void CompactibleFreeListSpace::smallSplitBirth(size_t size) { 2294 assert(size < SmallForDictionary, "Size too large for indexed list"); 2295 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2296 fl->increment_split_births(); 2297 fl->increment_surplus(); 2298 } 2299 2300 void CompactibleFreeListSpace::smallSplitDeath(size_t size) { 2301 assert(size < SmallForDictionary, "Size too large for indexed list"); 2302 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2303 fl->increment_split_deaths(); 2304 fl->decrement_surplus(); 2305 } 2306 2307 void CompactibleFreeListSpace::split_birth(size_t size) { 2308 if (size < SmallForDictionary) { 2309 smallSplitBirth(size); 2310 } else { 2311 dictionary()->dict_census_update(size, 2312 true /* split */, 2313 true /* birth */); 2314 } 2315 } 2316 2317 void CompactibleFreeListSpace::splitDeath(size_t size) { 2318 if (size < SmallForDictionary) { 2319 smallSplitDeath(size); 2320 } else { 2321 dictionary()->dict_census_update(size, 2322 true /* split */, 2323 false /* birth */); 2324 } 2325 } 2326 2327 void CompactibleFreeListSpace::split(size_t from, size_t to1) { 2328 size_t to2 = from - to1; 2329 splitDeath(from); 2330 split_birth(to1); 2331 split_birth(to2); 2332 } 2333 2334 void CompactibleFreeListSpace::print() const { 2335 print_on(tty); 2336 } 2337 2338 void CompactibleFreeListSpace::prepare_for_verify() { 2339 assert_locked(); 2340 repairLinearAllocationBlocks(); 2341 // Verify that the SpoolBlocks look like free blocks of 2342 // appropriate sizes... To be done ... 2343 } 2344 2345 class VerifyAllBlksClosure: public BlkClosure { 2346 private: 2347 const CompactibleFreeListSpace* _sp; 2348 const MemRegion _span; 2349 HeapWord* _last_addr; 2350 size_t _last_size; 2351 bool _last_was_obj; 2352 bool _last_was_live; 2353 2354 public: 2355 VerifyAllBlksClosure(const CompactibleFreeListSpace* sp, 2356 MemRegion span) : _sp(sp), _span(span), 2357 _last_addr(NULL), _last_size(0), 2358 _last_was_obj(false), _last_was_live(false) { } 2359 2360 virtual size_t do_blk(HeapWord* addr) { 2361 size_t res; 2362 bool was_obj = false; 2363 bool was_live = false; 2364 if (_sp->block_is_obj(addr)) { 2365 was_obj = true; 2366 oop p = oop(addr); 2367 guarantee(p->is_oop(), "Should be an oop"); 2368 res = _sp->adjustObjectSize(p->size()); 2369 if (_sp->obj_is_alive(addr)) { 2370 was_live = true; 2371 p->verify(); 2372 } 2373 } else { 2374 FreeChunk* fc = (FreeChunk*)addr; 2375 res = fc->size(); 2376 if (FLSVerifyLists && !fc->cantCoalesce()) { 2377 guarantee(_sp->verify_chunk_in_free_list(fc), 2378 "Chunk should be on a free list"); 2379 } 2380 } 2381 if (res == 0) { 2382 gclog_or_tty->print_cr("Livelock: no rank reduction!"); 2383 gclog_or_tty->print_cr( 2384 " Current: addr = " PTR_FORMAT ", size = " SIZE_FORMAT ", obj = %s, live = %s \n" 2385 " Previous: addr = " PTR_FORMAT ", size = " SIZE_FORMAT ", obj = %s, live = %s \n", 2386 addr, res, was_obj ?"true":"false", was_live ?"true":"false", 2387 _last_addr, _last_size, _last_was_obj?"true":"false", _last_was_live?"true":"false"); 2388 _sp->print_on(gclog_or_tty); 2389 guarantee(false, "Seppuku!"); 2390 } 2391 _last_addr = addr; 2392 _last_size = res; 2393 _last_was_obj = was_obj; 2394 _last_was_live = was_live; 2395 return res; 2396 } 2397 }; 2398 2399 class VerifyAllOopsClosure: public OopClosure { 2400 private: 2401 const CMSCollector* _collector; 2402 const CompactibleFreeListSpace* _sp; 2403 const MemRegion _span; 2404 const bool _past_remark; 2405 const CMSBitMap* _bit_map; 2406 2407 protected: 2408 void do_oop(void* p, oop obj) { 2409 if (_span.contains(obj)) { // the interior oop points into CMS heap 2410 if (!_span.contains(p)) { // reference from outside CMS heap 2411 // Should be a valid object; the first disjunct below allows 2412 // us to sidestep an assertion in block_is_obj() that insists 2413 // that p be in _sp. Note that several generations (and spaces) 2414 // are spanned by _span (CMS heap) above. 2415 guarantee(!_sp->is_in_reserved(obj) || 2416 _sp->block_is_obj((HeapWord*)obj), 2417 "Should be an object"); 2418 guarantee(obj->is_oop(), "Should be an oop"); 2419 obj->verify(); 2420 if (_past_remark) { 2421 // Remark has been completed, the object should be marked 2422 _bit_map->isMarked((HeapWord*)obj); 2423 } 2424 } else { // reference within CMS heap 2425 if (_past_remark) { 2426 // Remark has been completed -- so the referent should have 2427 // been marked, if referring object is. 2428 if (_bit_map->isMarked(_collector->block_start(p))) { 2429 guarantee(_bit_map->isMarked((HeapWord*)obj), "Marking error?"); 2430 } 2431 } 2432 } 2433 } else if (_sp->is_in_reserved(p)) { 2434 // the reference is from FLS, and points out of FLS 2435 guarantee(obj->is_oop(), "Should be an oop"); 2436 obj->verify(); 2437 } 2438 } 2439 2440 template <class T> void do_oop_work(T* p) { 2441 T heap_oop = oopDesc::load_heap_oop(p); 2442 if (!oopDesc::is_null(heap_oop)) { 2443 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 2444 do_oop(p, obj); 2445 } 2446 } 2447 2448 public: 2449 VerifyAllOopsClosure(const CMSCollector* collector, 2450 const CompactibleFreeListSpace* sp, MemRegion span, 2451 bool past_remark, CMSBitMap* bit_map) : 2452 _collector(collector), _sp(sp), _span(span), 2453 _past_remark(past_remark), _bit_map(bit_map) { } 2454 2455 virtual void do_oop(oop* p) { VerifyAllOopsClosure::do_oop_work(p); } 2456 virtual void do_oop(narrowOop* p) { VerifyAllOopsClosure::do_oop_work(p); } 2457 }; 2458 2459 void CompactibleFreeListSpace::verify() const { 2460 assert_lock_strong(&_freelistLock); 2461 verify_objects_initialized(); 2462 MemRegion span = _collector->_span; 2463 bool past_remark = (_collector->abstract_state() == 2464 CMSCollector::Sweeping); 2465 2466 ResourceMark rm; 2467 HandleMark hm; 2468 2469 // Check integrity of CFL data structures 2470 _promoInfo.verify(); 2471 _dictionary->verify(); 2472 if (FLSVerifyIndexTable) { 2473 verifyIndexedFreeLists(); 2474 } 2475 // Check integrity of all objects and free blocks in space 2476 { 2477 VerifyAllBlksClosure cl(this, span); 2478 ((CompactibleFreeListSpace*)this)->blk_iterate(&cl); // cast off const 2479 } 2480 // Check that all references in the heap to FLS 2481 // are to valid objects in FLS or that references in 2482 // FLS are to valid objects elsewhere in the heap 2483 if (FLSVerifyAllHeapReferences) 2484 { 2485 VerifyAllOopsClosure cl(_collector, this, span, past_remark, 2486 _collector->markBitMap()); 2487 CollectedHeap* ch = Universe::heap(); 2488 2489 // Iterate over all oops in the heap. Uses the _no_header version 2490 // since we are not interested in following the klass pointers. 2491 ch->oop_iterate_no_header(&cl); 2492 } 2493 2494 if (VerifyObjectStartArray) { 2495 // Verify the block offset table 2496 _bt.verify(); 2497 } 2498 } 2499 2500 #ifndef PRODUCT 2501 void CompactibleFreeListSpace::verifyFreeLists() const { 2502 if (FLSVerifyLists) { 2503 _dictionary->verify(); 2504 verifyIndexedFreeLists(); 2505 } else { 2506 if (FLSVerifyDictionary) { 2507 _dictionary->verify(); 2508 } 2509 if (FLSVerifyIndexTable) { 2510 verifyIndexedFreeLists(); 2511 } 2512 } 2513 } 2514 #endif 2515 2516 void CompactibleFreeListSpace::verifyIndexedFreeLists() const { 2517 size_t i = 0; 2518 for (; i < IndexSetStart; i++) { 2519 guarantee(_indexedFreeList[i].head() == NULL, "should be NULL"); 2520 } 2521 for (; i < IndexSetSize; i++) { 2522 verifyIndexedFreeList(i); 2523 } 2524 } 2525 2526 void CompactibleFreeListSpace::verifyIndexedFreeList(size_t size) const { 2527 FreeChunk* fc = _indexedFreeList[size].head(); 2528 FreeChunk* tail = _indexedFreeList[size].tail(); 2529 size_t num = _indexedFreeList[size].count(); 2530 size_t n = 0; 2531 guarantee(((size >= IndexSetStart) && (size % IndexSetStride == 0)) || fc == NULL, 2532 "Slot should have been empty"); 2533 for (; fc != NULL; fc = fc->next(), n++) { 2534 guarantee(fc->size() == size, "Size inconsistency"); 2535 guarantee(fc->is_free(), "!free?"); 2536 guarantee(fc->next() == NULL || fc->next()->prev() == fc, "Broken list"); 2537 guarantee((fc->next() == NULL) == (fc == tail), "Incorrect tail"); 2538 } 2539 guarantee(n == num, "Incorrect count"); 2540 } 2541 2542 #ifndef PRODUCT 2543 void CompactibleFreeListSpace::check_free_list_consistency() const { 2544 assert((TreeChunk<FreeChunk, AdaptiveFreeList>::min_size() <= IndexSetSize), 2545 "Some sizes can't be allocated without recourse to" 2546 " linear allocation buffers"); 2547 assert((TreeChunk<FreeChunk, AdaptiveFreeList>::min_size()*HeapWordSize == sizeof(TreeChunk<FreeChunk, AdaptiveFreeList>)), 2548 "else MIN_TREE_CHUNK_SIZE is wrong"); 2549 assert(IndexSetStart != 0, "IndexSetStart not initialized"); 2550 assert(IndexSetStride != 0, "IndexSetStride not initialized"); 2551 } 2552 #endif 2553 2554 void CompactibleFreeListSpace::printFLCensus(size_t sweep_count) const { 2555 assert_lock_strong(&_freelistLock); 2556 AdaptiveFreeList<FreeChunk> total; 2557 gclog_or_tty->print("end sweep# " SIZE_FORMAT "\n", sweep_count); 2558 AdaptiveFreeList<FreeChunk>::print_labels_on(gclog_or_tty, "size"); 2559 size_t total_free = 0; 2560 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2561 const AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2562 total_free += fl->count() * fl->size(); 2563 if (i % (40*IndexSetStride) == 0) { 2564 AdaptiveFreeList<FreeChunk>::print_labels_on(gclog_or_tty, "size"); 2565 } 2566 fl->print_on(gclog_or_tty); 2567 total.set_bfr_surp( total.bfr_surp() + fl->bfr_surp() ); 2568 total.set_surplus( total.surplus() + fl->surplus() ); 2569 total.set_desired( total.desired() + fl->desired() ); 2570 total.set_prev_sweep( total.prev_sweep() + fl->prev_sweep() ); 2571 total.set_before_sweep(total.before_sweep() + fl->before_sweep()); 2572 total.set_count( total.count() + fl->count() ); 2573 total.set_coal_births( total.coal_births() + fl->coal_births() ); 2574 total.set_coal_deaths( total.coal_deaths() + fl->coal_deaths() ); 2575 total.set_split_births(total.split_births() + fl->split_births()); 2576 total.set_split_deaths(total.split_deaths() + fl->split_deaths()); 2577 } 2578 total.print_on(gclog_or_tty, "TOTAL"); 2579 gclog_or_tty->print_cr("Total free in indexed lists " 2580 SIZE_FORMAT " words", total_free); 2581 gclog_or_tty->print("growth: %8.5f deficit: %8.5f\n", 2582 (double)(total.split_births()+total.coal_births()-total.split_deaths()-total.coal_deaths())/ 2583 (total.prev_sweep() != 0 ? (double)total.prev_sweep() : 1.0), 2584 (double)(total.desired() - total.count())/(total.desired() != 0 ? (double)total.desired() : 1.0)); 2585 _dictionary->print_dict_census(); 2586 } 2587 2588 /////////////////////////////////////////////////////////////////////////// 2589 // CFLS_LAB 2590 /////////////////////////////////////////////////////////////////////////// 2591 2592 #define VECTOR_257(x) \ 2593 /* 1 2 3 4 5 6 7 8 9 1x 11 12 13 14 15 16 17 18 19 2x 21 22 23 24 25 26 27 28 29 3x 31 32 */ \ 2594 { x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2595 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2596 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2597 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2598 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2599 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2600 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2601 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \ 2602 x } 2603 2604 // Initialize with default setting of CMSParPromoteBlocksToClaim, _not_ 2605 // OldPLABSize, whose static default is different; if overridden at the 2606 // command-line, this will get reinitialized via a call to 2607 // modify_initialization() below. 2608 AdaptiveWeightedAverage CFLS_LAB::_blocks_to_claim[] = 2609 VECTOR_257(AdaptiveWeightedAverage(OldPLABWeight, (float)CMSParPromoteBlocksToClaim)); 2610 size_t CFLS_LAB::_global_num_blocks[] = VECTOR_257(0); 2611 uint CFLS_LAB::_global_num_workers[] = VECTOR_257(0); 2612 2613 CFLS_LAB::CFLS_LAB(CompactibleFreeListSpace* cfls) : 2614 _cfls(cfls) 2615 { 2616 assert(CompactibleFreeListSpace::IndexSetSize == 257, "Modify VECTOR_257() macro above"); 2617 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2618 i < CompactibleFreeListSpace::IndexSetSize; 2619 i += CompactibleFreeListSpace::IndexSetStride) { 2620 _indexedFreeList[i].set_size(i); 2621 _num_blocks[i] = 0; 2622 } 2623 } 2624 2625 static bool _CFLS_LAB_modified = false; 2626 2627 void CFLS_LAB::modify_initialization(size_t n, unsigned wt) { 2628 assert(!_CFLS_LAB_modified, "Call only once"); 2629 _CFLS_LAB_modified = true; 2630 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2631 i < CompactibleFreeListSpace::IndexSetSize; 2632 i += CompactibleFreeListSpace::IndexSetStride) { 2633 _blocks_to_claim[i].modify(n, wt, true /* force */); 2634 } 2635 } 2636 2637 HeapWord* CFLS_LAB::alloc(size_t word_sz) { 2638 FreeChunk* res; 2639 assert(word_sz == _cfls->adjustObjectSize(word_sz), "Error"); 2640 if (word_sz >= CompactibleFreeListSpace::IndexSetSize) { 2641 // This locking manages sync with other large object allocations. 2642 MutexLockerEx x(_cfls->parDictionaryAllocLock(), 2643 Mutex::_no_safepoint_check_flag); 2644 res = _cfls->getChunkFromDictionaryExact(word_sz); 2645 if (res == NULL) return NULL; 2646 } else { 2647 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[word_sz]; 2648 if (fl->count() == 0) { 2649 // Attempt to refill this local free list. 2650 get_from_global_pool(word_sz, fl); 2651 // If it didn't work, give up. 2652 if (fl->count() == 0) return NULL; 2653 } 2654 res = fl->get_chunk_at_head(); 2655 assert(res != NULL, "Why was count non-zero?"); 2656 } 2657 res->markNotFree(); 2658 assert(!res->is_free(), "shouldn't be marked free"); 2659 assert(oop(res)->klass_or_null() == NULL, "should look uninitialized"); 2660 // mangle a just allocated object with a distinct pattern. 2661 debug_only(res->mangleAllocated(word_sz)); 2662 return (HeapWord*)res; 2663 } 2664 2665 // Get a chunk of blocks of the right size and update related 2666 // book-keeping stats 2667 void CFLS_LAB::get_from_global_pool(size_t word_sz, AdaptiveFreeList<FreeChunk>* fl) { 2668 // Get the #blocks we want to claim 2669 size_t n_blks = (size_t)_blocks_to_claim[word_sz].average(); 2670 assert(n_blks > 0, "Error"); 2671 assert(ResizePLAB || n_blks == OldPLABSize, "Error"); 2672 // In some cases, when the application has a phase change, 2673 // there may be a sudden and sharp shift in the object survival 2674 // profile, and updating the counts at the end of a scavenge 2675 // may not be quick enough, giving rise to large scavenge pauses 2676 // during these phase changes. It is beneficial to detect such 2677 // changes on-the-fly during a scavenge and avoid such a phase-change 2678 // pothole. The following code is a heuristic attempt to do that. 2679 // It is protected by a product flag until we have gained 2680 // enough experience with this heuristic and fine-tuned its behaviour. 2681 // WARNING: This might increase fragmentation if we overreact to 2682 // small spikes, so some kind of historical smoothing based on 2683 // previous experience with the greater reactivity might be useful. 2684 // Lacking sufficient experience, CMSOldPLABResizeQuicker is disabled by 2685 // default. 2686 if (ResizeOldPLAB && CMSOldPLABResizeQuicker) { 2687 size_t multiple = _num_blocks[word_sz]/(CMSOldPLABToleranceFactor*CMSOldPLABNumRefills*n_blks); 2688 n_blks += CMSOldPLABReactivityFactor*multiple*n_blks; 2689 n_blks = MIN2(n_blks, CMSOldPLABMax); 2690 } 2691 assert(n_blks > 0, "Error"); 2692 _cfls->par_get_chunk_of_blocks(word_sz, n_blks, fl); 2693 // Update stats table entry for this block size 2694 _num_blocks[word_sz] += fl->count(); 2695 } 2696 2697 void CFLS_LAB::compute_desired_plab_size() { 2698 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2699 i < CompactibleFreeListSpace::IndexSetSize; 2700 i += CompactibleFreeListSpace::IndexSetStride) { 2701 assert((_global_num_workers[i] == 0) == (_global_num_blocks[i] == 0), 2702 "Counter inconsistency"); 2703 if (_global_num_workers[i] > 0) { 2704 // Need to smooth wrt historical average 2705 if (ResizeOldPLAB) { 2706 _blocks_to_claim[i].sample( 2707 MAX2((size_t)CMSOldPLABMin, 2708 MIN2((size_t)CMSOldPLABMax, 2709 _global_num_blocks[i]/(_global_num_workers[i]*CMSOldPLABNumRefills)))); 2710 } 2711 // Reset counters for next round 2712 _global_num_workers[i] = 0; 2713 _global_num_blocks[i] = 0; 2714 if (PrintOldPLAB) { 2715 gclog_or_tty->print_cr("[%d]: %d", i, (size_t)_blocks_to_claim[i].average()); 2716 } 2717 } 2718 } 2719 } 2720 2721 // If this is changed in the future to allow parallel 2722 // access, one would need to take the FL locks and, 2723 // depending on how it is used, stagger access from 2724 // parallel threads to reduce contention. 2725 void CFLS_LAB::retire(int tid) { 2726 // We run this single threaded with the world stopped; 2727 // so no need for locks and such. 2728 NOT_PRODUCT(Thread* t = Thread::current();) 2729 assert(Thread::current()->is_VM_thread(), "Error"); 2730 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2731 i < CompactibleFreeListSpace::IndexSetSize; 2732 i += CompactibleFreeListSpace::IndexSetStride) { 2733 assert(_num_blocks[i] >= (size_t)_indexedFreeList[i].count(), 2734 "Can't retire more than what we obtained"); 2735 if (_num_blocks[i] > 0) { 2736 size_t num_retire = _indexedFreeList[i].count(); 2737 assert(_num_blocks[i] > num_retire, "Should have used at least one"); 2738 { 2739 // MutexLockerEx x(_cfls->_indexedFreeListParLocks[i], 2740 // Mutex::_no_safepoint_check_flag); 2741 2742 // Update globals stats for num_blocks used 2743 _global_num_blocks[i] += (_num_blocks[i] - num_retire); 2744 _global_num_workers[i]++; 2745 assert(_global_num_workers[i] <= ParallelGCThreads, "Too big"); 2746 if (num_retire > 0) { 2747 _cfls->_indexedFreeList[i].prepend(&_indexedFreeList[i]); 2748 // Reset this list. 2749 _indexedFreeList[i] = AdaptiveFreeList<FreeChunk>(); 2750 _indexedFreeList[i].set_size(i); 2751 } 2752 } 2753 if (PrintOldPLAB) { 2754 gclog_or_tty->print_cr("%d[%d]: %d/%d/%d", 2755 tid, i, num_retire, _num_blocks[i], (size_t)_blocks_to_claim[i].average()); 2756 } 2757 // Reset stats for next round 2758 _num_blocks[i] = 0; 2759 } 2760 } 2761 } 2762 2763 void CompactibleFreeListSpace:: par_get_chunk_of_blocks(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) { 2764 assert(fl->count() == 0, "Precondition."); 2765 assert(word_sz < CompactibleFreeListSpace::IndexSetSize, 2766 "Precondition"); 2767 2768 // We'll try all multiples of word_sz in the indexed set, starting with 2769 // word_sz itself and, if CMSSplitIndexedFreeListBlocks, try larger multiples, 2770 // then try getting a big chunk and splitting it. 2771 { 2772 bool found; 2773 int k; 2774 size_t cur_sz; 2775 for (k = 1, cur_sz = k * word_sz, found = false; 2776 (cur_sz < CompactibleFreeListSpace::IndexSetSize) && 2777 (CMSSplitIndexedFreeListBlocks || k <= 1); 2778 k++, cur_sz = k * word_sz) { 2779 AdaptiveFreeList<FreeChunk> fl_for_cur_sz; // Empty. 2780 fl_for_cur_sz.set_size(cur_sz); 2781 { 2782 MutexLockerEx x(_indexedFreeListParLocks[cur_sz], 2783 Mutex::_no_safepoint_check_flag); 2784 AdaptiveFreeList<FreeChunk>* gfl = &_indexedFreeList[cur_sz]; 2785 if (gfl->count() != 0) { 2786 // nn is the number of chunks of size cur_sz that 2787 // we'd need to split k-ways each, in order to create 2788 // "n" chunks of size word_sz each. 2789 const size_t nn = MAX2(n/k, (size_t)1); 2790 gfl->getFirstNChunksFromList(nn, &fl_for_cur_sz); 2791 found = true; 2792 if (k > 1) { 2793 // Update split death stats for the cur_sz-size blocks list: 2794 // we increment the split death count by the number of blocks 2795 // we just took from the cur_sz-size blocks list and which 2796 // we will be splitting below. 2797 ssize_t deaths = gfl->split_deaths() + 2798 fl_for_cur_sz.count(); 2799 gfl->set_split_deaths(deaths); 2800 } 2801 } 2802 } 2803 // Now transfer fl_for_cur_sz to fl. Common case, we hope, is k = 1. 2804 if (found) { 2805 if (k == 1) { 2806 fl->prepend(&fl_for_cur_sz); 2807 } else { 2808 // Divide each block on fl_for_cur_sz up k ways. 2809 FreeChunk* fc; 2810 while ((fc = fl_for_cur_sz.get_chunk_at_head()) != NULL) { 2811 // Must do this in reverse order, so that anybody attempting to 2812 // access the main chunk sees it as a single free block until we 2813 // change it. 2814 size_t fc_size = fc->size(); 2815 assert(fc->is_free(), "Error"); 2816 for (int i = k-1; i >= 0; i--) { 2817 FreeChunk* ffc = (FreeChunk*)((HeapWord*)fc + i * word_sz); 2818 assert((i != 0) || 2819 ((fc == ffc) && ffc->is_free() && 2820 (ffc->size() == k*word_sz) && (fc_size == word_sz)), 2821 "Counting error"); 2822 ffc->set_size(word_sz); 2823 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 2824 ffc->link_next(NULL); 2825 // Above must occur before BOT is updated below. 2826 OrderAccess::storestore(); 2827 // splitting from the right, fc_size == i * word_sz 2828 _bt.mark_block((HeapWord*)ffc, word_sz, true /* reducing */); 2829 fc_size -= word_sz; 2830 assert(fc_size == i*word_sz, "Error"); 2831 _bt.verify_not_unallocated((HeapWord*)ffc, word_sz); 2832 _bt.verify_single_block((HeapWord*)fc, fc_size); 2833 _bt.verify_single_block((HeapWord*)ffc, word_sz); 2834 // Push this on "fl". 2835 fl->return_chunk_at_head(ffc); 2836 } 2837 // TRAP 2838 assert(fl->tail()->next() == NULL, "List invariant."); 2839 } 2840 } 2841 // Update birth stats for this block size. 2842 size_t num = fl->count(); 2843 MutexLockerEx x(_indexedFreeListParLocks[word_sz], 2844 Mutex::_no_safepoint_check_flag); 2845 ssize_t births = _indexedFreeList[word_sz].split_births() + num; 2846 _indexedFreeList[word_sz].set_split_births(births); 2847 return; 2848 } 2849 } 2850 } 2851 // Otherwise, we'll split a block from the dictionary. 2852 FreeChunk* fc = NULL; 2853 FreeChunk* rem_fc = NULL; 2854 size_t rem; 2855 { 2856 MutexLockerEx x(parDictionaryAllocLock(), 2857 Mutex::_no_safepoint_check_flag); 2858 while (n > 0) { 2859 fc = dictionary()->get_chunk(MAX2(n * word_sz, _dictionary->min_size()), 2860 FreeBlockDictionary<FreeChunk>::atLeast); 2861 if (fc != NULL) { 2862 _bt.allocated((HeapWord*)fc, fc->size(), true /* reducing */); // update _unallocated_blk 2863 dictionary()->dict_census_update(fc->size(), 2864 true /*split*/, 2865 false /*birth*/); 2866 break; 2867 } else { 2868 n--; 2869 } 2870 } 2871 if (fc == NULL) return; 2872 // Otherwise, split up that block. 2873 assert((ssize_t)n >= 1, "Control point invariant"); 2874 assert(fc->is_free(), "Error: should be a free block"); 2875 _bt.verify_single_block((HeapWord*)fc, fc->size()); 2876 const size_t nn = fc->size() / word_sz; 2877 n = MIN2(nn, n); 2878 assert((ssize_t)n >= 1, "Control point invariant"); 2879 rem = fc->size() - n * word_sz; 2880 // If there is a remainder, and it's too small, allocate one fewer. 2881 if (rem > 0 && rem < MinChunkSize) { 2882 n--; rem += word_sz; 2883 } 2884 // Note that at this point we may have n == 0. 2885 assert((ssize_t)n >= 0, "Control point invariant"); 2886 2887 // If n is 0, the chunk fc that was found is not large 2888 // enough to leave a viable remainder. We are unable to 2889 // allocate even one block. Return fc to the 2890 // dictionary and return, leaving "fl" empty. 2891 if (n == 0) { 2892 returnChunkToDictionary(fc); 2893 assert(fl->count() == 0, "We never allocated any blocks"); 2894 return; 2895 } 2896 2897 // First return the remainder, if any. 2898 // Note that we hold the lock until we decide if we're going to give 2899 // back the remainder to the dictionary, since a concurrent allocation 2900 // may otherwise see the heap as empty. (We're willing to take that 2901 // hit if the block is a small block.) 2902 if (rem > 0) { 2903 size_t prefix_size = n * word_sz; 2904 rem_fc = (FreeChunk*)((HeapWord*)fc + prefix_size); 2905 rem_fc->set_size(rem); 2906 rem_fc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 2907 rem_fc->link_next(NULL); 2908 // Above must occur before BOT is updated below. 2909 assert((ssize_t)n > 0 && prefix_size > 0 && rem_fc > fc, "Error"); 2910 OrderAccess::storestore(); 2911 _bt.split_block((HeapWord*)fc, fc->size(), prefix_size); 2912 assert(fc->is_free(), "Error"); 2913 fc->set_size(prefix_size); 2914 if (rem >= IndexSetSize) { 2915 returnChunkToDictionary(rem_fc); 2916 dictionary()->dict_census_update(rem, true /*split*/, true /*birth*/); 2917 rem_fc = NULL; 2918 } 2919 // Otherwise, return it to the small list below. 2920 } 2921 } 2922 if (rem_fc != NULL) { 2923 MutexLockerEx x(_indexedFreeListParLocks[rem], 2924 Mutex::_no_safepoint_check_flag); 2925 _bt.verify_not_unallocated((HeapWord*)rem_fc, rem_fc->size()); 2926 _indexedFreeList[rem].return_chunk_at_head(rem_fc); 2927 smallSplitBirth(rem); 2928 } 2929 assert((ssize_t)n > 0 && fc != NULL, "Consistency"); 2930 // Now do the splitting up. 2931 // Must do this in reverse order, so that anybody attempting to 2932 // access the main chunk sees it as a single free block until we 2933 // change it. 2934 size_t fc_size = n * word_sz; 2935 // All but first chunk in this loop 2936 for (ssize_t i = n-1; i > 0; i--) { 2937 FreeChunk* ffc = (FreeChunk*)((HeapWord*)fc + i * word_sz); 2938 ffc->set_size(word_sz); 2939 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 2940 ffc->link_next(NULL); 2941 // Above must occur before BOT is updated below. 2942 OrderAccess::storestore(); 2943 // splitting from the right, fc_size == (n - i + 1) * wordsize 2944 _bt.mark_block((HeapWord*)ffc, word_sz, true /* reducing */); 2945 fc_size -= word_sz; 2946 _bt.verify_not_unallocated((HeapWord*)ffc, ffc->size()); 2947 _bt.verify_single_block((HeapWord*)ffc, ffc->size()); 2948 _bt.verify_single_block((HeapWord*)fc, fc_size); 2949 // Push this on "fl". 2950 fl->return_chunk_at_head(ffc); 2951 } 2952 // First chunk 2953 assert(fc->is_free() && fc->size() == n*word_sz, "Error: should still be a free block"); 2954 // The blocks above should show their new sizes before the first block below 2955 fc->set_size(word_sz); 2956 fc->link_prev(NULL); // idempotent wrt free-ness, see assert above 2957 fc->link_next(NULL); 2958 _bt.verify_not_unallocated((HeapWord*)fc, fc->size()); 2959 _bt.verify_single_block((HeapWord*)fc, fc->size()); 2960 fl->return_chunk_at_head(fc); 2961 2962 assert((ssize_t)n > 0 && (ssize_t)n == fl->count(), "Incorrect number of blocks"); 2963 { 2964 // Update the stats for this block size. 2965 MutexLockerEx x(_indexedFreeListParLocks[word_sz], 2966 Mutex::_no_safepoint_check_flag); 2967 const ssize_t births = _indexedFreeList[word_sz].split_births() + n; 2968 _indexedFreeList[word_sz].set_split_births(births); 2969 // ssize_t new_surplus = _indexedFreeList[word_sz].surplus() + n; 2970 // _indexedFreeList[word_sz].set_surplus(new_surplus); 2971 } 2972 2973 // TRAP 2974 assert(fl->tail()->next() == NULL, "List invariant."); 2975 } 2976 2977 // Set up the space's par_seq_tasks structure for work claiming 2978 // for parallel rescan. See CMSParRemarkTask where this is currently used. 2979 // XXX Need to suitably abstract and generalize this and the next 2980 // method into one. 2981 void 2982 CompactibleFreeListSpace:: 2983 initialize_sequential_subtasks_for_rescan(int n_threads) { 2984 // The "size" of each task is fixed according to rescan_task_size. 2985 assert(n_threads > 0, "Unexpected n_threads argument"); 2986 const size_t task_size = rescan_task_size(); 2987 size_t n_tasks = (used_region().word_size() + task_size - 1)/task_size; 2988 assert((n_tasks == 0) == used_region().is_empty(), "n_tasks incorrect"); 2989 assert(n_tasks == 0 || 2990 ((used_region().start() + (n_tasks - 1)*task_size < used_region().end()) && 2991 (used_region().start() + n_tasks*task_size >= used_region().end())), 2992 "n_tasks calculation incorrect"); 2993 SequentialSubTasksDone* pst = conc_par_seq_tasks(); 2994 assert(!pst->valid(), "Clobbering existing data?"); 2995 // Sets the condition for completion of the subtask (how many threads 2996 // need to finish in order to be done). 2997 pst->set_n_threads(n_threads); 2998 pst->set_n_tasks((int)n_tasks); 2999 } 3000 3001 // Set up the space's par_seq_tasks structure for work claiming 3002 // for parallel concurrent marking. See CMSConcMarkTask where this is currently used. 3003 void 3004 CompactibleFreeListSpace:: 3005 initialize_sequential_subtasks_for_marking(int n_threads, 3006 HeapWord* low) { 3007 // The "size" of each task is fixed according to rescan_task_size. 3008 assert(n_threads > 0, "Unexpected n_threads argument"); 3009 const size_t task_size = marking_task_size(); 3010 assert(task_size > CardTableModRefBS::card_size_in_words && 3011 (task_size % CardTableModRefBS::card_size_in_words == 0), 3012 "Otherwise arithmetic below would be incorrect"); 3013 MemRegion span = _gen->reserved(); 3014 if (low != NULL) { 3015 if (span.contains(low)) { 3016 // Align low down to a card boundary so that 3017 // we can use block_offset_careful() on span boundaries. 3018 HeapWord* aligned_low = (HeapWord*)align_size_down((uintptr_t)low, 3019 CardTableModRefBS::card_size); 3020 // Clip span prefix at aligned_low 3021 span = span.intersection(MemRegion(aligned_low, span.end())); 3022 } else if (low > span.end()) { 3023 span = MemRegion(low, low); // Null region 3024 } // else use entire span 3025 } 3026 assert(span.is_empty() || 3027 ((uintptr_t)span.start() % CardTableModRefBS::card_size == 0), 3028 "span should start at a card boundary"); 3029 size_t n_tasks = (span.word_size() + task_size - 1)/task_size; 3030 assert((n_tasks == 0) == span.is_empty(), "Inconsistency"); 3031 assert(n_tasks == 0 || 3032 ((span.start() + (n_tasks - 1)*task_size < span.end()) && 3033 (span.start() + n_tasks*task_size >= span.end())), 3034 "n_tasks calculation incorrect"); 3035 SequentialSubTasksDone* pst = conc_par_seq_tasks(); 3036 assert(!pst->valid(), "Clobbering existing data?"); 3037 // Sets the condition for completion of the subtask (how many threads 3038 // need to finish in order to be done). 3039 pst->set_n_threads(n_threads); 3040 pst->set_n_tasks((int)n_tasks); 3041 }