1 /* 2 * Copyright (c) 2001, 2014, 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/space.inline.hpp" 37 #include "memory/universe.inline.hpp" 38 #include "oops/oop.inline.hpp" 39 #include "runtime/globals.hpp" 40 #include "runtime/handles.inline.hpp" 41 #include "runtime/init.hpp" 42 #include "runtime/java.hpp" 43 #include "runtime/orderAccess.inline.hpp" 44 #include "runtime/vmThread.hpp" 45 #include "utilities/copy.hpp" 46 47 ///////////////////////////////////////////////////////////////////////// 48 //// CompactibleFreeListSpace 49 ///////////////////////////////////////////////////////////////////////// 50 51 // highest ranked free list lock rank 52 int CompactibleFreeListSpace::_lockRank = Mutex::leaf + 3; 53 54 // Defaults are 0 so things will break badly if incorrectly initialized. 55 size_t CompactibleFreeListSpace::IndexSetStart = 0; 56 size_t CompactibleFreeListSpace::IndexSetStride = 0; 57 58 size_t MinChunkSize = 0; 59 60 void CompactibleFreeListSpace::set_cms_values() { 61 // Set CMS global values 62 assert(MinChunkSize == 0, "already set"); 63 64 // MinChunkSize should be a multiple of MinObjAlignment and be large enough 65 // for chunks to contain a FreeChunk. 66 size_t min_chunk_size_in_bytes = align_size_up(sizeof(FreeChunk), MinObjAlignmentInBytes); 67 MinChunkSize = min_chunk_size_in_bytes / BytesPerWord; 68 69 assert(IndexSetStart == 0 && IndexSetStride == 0, "already set"); 70 IndexSetStart = MinChunkSize; 71 IndexSetStride = MinObjAlignment; 72 } 73 74 // Constructor 75 CompactibleFreeListSpace::CompactibleFreeListSpace(BlockOffsetSharedArray* bs, 76 MemRegion mr, bool use_adaptive_freelists, 77 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 78 _dictionaryChoice(dictionaryChoice), 79 _adaptive_freelists(use_adaptive_freelists), 80 _bt(bs, mr), 81 // free list locks are in the range of values taken by _lockRank 82 // This range currently is [_leaf+2, _leaf+3] 83 // Note: this requires that CFLspace c'tors 84 // are called serially in the order in which the locks are 85 // are acquired in the program text. This is true today. 86 _freelistLock(_lockRank--, "CompactibleFreeListSpace._lock", true, 87 Monitor::_safepoint_check_sometimes), 88 _parDictionaryAllocLock(Mutex::leaf - 1, // == rank(ExpandHeap_lock) - 1 89 "CompactibleFreeListSpace._dict_par_lock", true, 90 Monitor::_safepoint_check_never), 91 _rescan_task_size(CardTableModRefBS::card_size_in_words * BitsPerWord * 92 CMSRescanMultiple), 93 _marking_task_size(CardTableModRefBS::card_size_in_words * BitsPerWord * 94 CMSConcMarkMultiple), 95 _collector(NULL) 96 { 97 assert(sizeof(FreeChunk) / BytesPerWord <= MinChunkSize, 98 "FreeChunk is larger than expected"); 99 _bt.set_space(this); 100 initialize(mr, SpaceDecorator::Clear, SpaceDecorator::Mangle); 101 // We have all of "mr", all of which we place in the dictionary 102 // as one big chunk. We'll need to decide here which of several 103 // possible alternative dictionary implementations to use. For 104 // now the choice is easy, since we have only one working 105 // implementation, namely, the simple binary tree (splaying 106 // temporarily disabled). 107 switch (dictionaryChoice) { 108 case FreeBlockDictionary<FreeChunk>::dictionaryBinaryTree: 109 _dictionary = new AFLBinaryTreeDictionary(mr); 110 break; 111 case FreeBlockDictionary<FreeChunk>::dictionarySplayTree: 112 case FreeBlockDictionary<FreeChunk>::dictionarySkipList: 113 default: 114 warning("dictionaryChoice: selected option not understood; using" 115 " default BinaryTreeDictionary implementation instead."); 116 } 117 assert(_dictionary != NULL, "CMS dictionary initialization"); 118 // The indexed free lists are initially all empty and are lazily 119 // filled in on demand. Initialize the array elements to NULL. 120 initializeIndexedFreeListArray(); 121 122 // Not using adaptive free lists assumes that allocation is first 123 // from the linAB's. Also a cms perm gen which can be compacted 124 // has to have the klass's klassKlass allocated at a lower 125 // address in the heap than the klass so that the klassKlass is 126 // moved to its new location before the klass is moved. 127 // Set the _refillSize for the linear allocation blocks 128 if (!use_adaptive_freelists) { 129 FreeChunk* fc = _dictionary->get_chunk(mr.word_size(), 130 FreeBlockDictionary<FreeChunk>::atLeast); 131 // The small linAB initially has all the space and will allocate 132 // a chunk of any size. 133 HeapWord* addr = (HeapWord*) fc; 134 _smallLinearAllocBlock.set(addr, fc->size() , 135 1024*SmallForLinearAlloc, fc->size()); 136 // Note that _unallocated_block is not updated here. 137 // Allocations from the linear allocation block should 138 // update it. 139 } else { 140 _smallLinearAllocBlock.set(0, 0, 1024*SmallForLinearAlloc, 141 SmallForLinearAlloc); 142 } 143 // CMSIndexedFreeListReplenish should be at least 1 144 CMSIndexedFreeListReplenish = MAX2((uintx)1, CMSIndexedFreeListReplenish); 145 _promoInfo.setSpace(this); 146 if (UseCMSBestFit) { 147 _fitStrategy = FreeBlockBestFitFirst; 148 } else { 149 _fitStrategy = FreeBlockStrategyNone; 150 } 151 check_free_list_consistency(); 152 153 // Initialize locks for parallel case. 154 155 if (CollectedHeap::use_parallel_gc_threads()) { 156 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 157 _indexedFreeListParLocks[i] = new Mutex(Mutex::leaf - 1, // == ExpandHeap_lock - 1 158 "freelist par lock", true, Mutex::_safepoint_check_sometimes); 159 DEBUG_ONLY( 160 _indexedFreeList[i].set_protecting_lock(_indexedFreeListParLocks[i]); 161 ) 162 } 163 _dictionary->set_par_lock(&_parDictionaryAllocLock); 164 } 165 } 166 167 // Like CompactibleSpace forward() but always calls cross_threshold() to 168 // update the block offset table. Removed initialize_threshold call because 169 // CFLS does not use a block offset array for contiguous spaces. 170 HeapWord* CompactibleFreeListSpace::forward(oop q, size_t size, 171 CompactPoint* cp, HeapWord* compact_top) { 172 // q is alive 173 // First check if we should switch compaction space 174 assert(this == cp->space, "'this' should be current compaction space."); 175 size_t compaction_max_size = pointer_delta(end(), compact_top); 176 assert(adjustObjectSize(size) == cp->space->adjust_object_size_v(size), 177 "virtual adjustObjectSize_v() method is not correct"); 178 size_t adjusted_size = adjustObjectSize(size); 179 assert(compaction_max_size >= MinChunkSize || compaction_max_size == 0, 180 "no small fragments allowed"); 181 assert(minimum_free_block_size() == MinChunkSize, 182 "for de-virtualized reference below"); 183 // Can't leave a nonzero size, residual fragment smaller than MinChunkSize 184 if (adjusted_size + MinChunkSize > compaction_max_size && 185 adjusted_size != compaction_max_size) { 186 do { 187 // switch to next compaction space 188 cp->space->set_compaction_top(compact_top); 189 cp->space = cp->space->next_compaction_space(); 190 if (cp->space == NULL) { 191 cp->gen = GenCollectedHeap::heap()->prev_gen(cp->gen); 192 assert(cp->gen != NULL, "compaction must succeed"); 193 cp->space = cp->gen->first_compaction_space(); 194 assert(cp->space != NULL, "generation must have a first compaction space"); 195 } 196 compact_top = cp->space->bottom(); 197 cp->space->set_compaction_top(compact_top); 198 // The correct adjusted_size may not be the same as that for this method 199 // (i.e., cp->space may no longer be "this" so adjust the size again. 200 // Use the virtual method which is not used above to save the virtual 201 // dispatch. 202 adjusted_size = cp->space->adjust_object_size_v(size); 203 compaction_max_size = pointer_delta(cp->space->end(), compact_top); 204 assert(cp->space->minimum_free_block_size() == 0, "just checking"); 205 } while (adjusted_size > compaction_max_size); 206 } 207 208 // store the forwarding pointer into the mark word 209 if ((HeapWord*)q != compact_top) { 210 q->forward_to(oop(compact_top)); 211 assert(q->is_gc_marked(), "encoding the pointer should preserve the mark"); 212 } else { 213 // if the object isn't moving we can just set the mark to the default 214 // mark and handle it specially later on. 215 q->init_mark(); 216 assert(q->forwardee() == NULL, "should be forwarded to NULL"); 217 } 218 219 compact_top += adjusted_size; 220 221 // we need to update the offset table so that the beginnings of objects can be 222 // found during scavenge. Note that we are updating the offset table based on 223 // where the object will be once the compaction phase finishes. 224 225 // Always call cross_threshold(). A contiguous space can only call it when 226 // the compaction_top exceeds the current threshold but not for an 227 // non-contiguous space. 228 cp->threshold = 229 cp->space->cross_threshold(compact_top - adjusted_size, compact_top); 230 return compact_top; 231 } 232 233 // A modified copy of OffsetTableContigSpace::cross_threshold() with _offsets -> _bt 234 // and use of single_block instead of alloc_block. The name here is not really 235 // appropriate - maybe a more general name could be invented for both the 236 // contiguous and noncontiguous spaces. 237 238 HeapWord* CompactibleFreeListSpace::cross_threshold(HeapWord* start, HeapWord* the_end) { 239 _bt.single_block(start, the_end); 240 return end(); 241 } 242 243 // Initialize them to NULL. 244 void CompactibleFreeListSpace::initializeIndexedFreeListArray() { 245 for (size_t i = 0; i < IndexSetSize; i++) { 246 // Note that on platforms where objects are double word aligned, 247 // the odd array elements are not used. It is convenient, however, 248 // to map directly from the object size to the array element. 249 _indexedFreeList[i].reset(IndexSetSize); 250 _indexedFreeList[i].set_size(i); 251 assert(_indexedFreeList[i].count() == 0, "reset check failed"); 252 assert(_indexedFreeList[i].head() == NULL, "reset check failed"); 253 assert(_indexedFreeList[i].tail() == NULL, "reset check failed"); 254 assert(_indexedFreeList[i].hint() == IndexSetSize, "reset check failed"); 255 } 256 } 257 258 void CompactibleFreeListSpace::resetIndexedFreeListArray() { 259 for (size_t i = 1; i < IndexSetSize; i++) { 260 assert(_indexedFreeList[i].size() == (size_t) i, 261 "Indexed free list sizes are incorrect"); 262 _indexedFreeList[i].reset(IndexSetSize); 263 assert(_indexedFreeList[i].count() == 0, "reset check failed"); 264 assert(_indexedFreeList[i].head() == NULL, "reset check failed"); 265 assert(_indexedFreeList[i].tail() == NULL, "reset check failed"); 266 assert(_indexedFreeList[i].hint() == IndexSetSize, "reset check failed"); 267 } 268 } 269 270 void CompactibleFreeListSpace::reset(MemRegion mr) { 271 resetIndexedFreeListArray(); 272 dictionary()->reset(); 273 if (BlockOffsetArrayUseUnallocatedBlock) { 274 assert(end() == mr.end(), "We are compacting to the bottom of CMS gen"); 275 // Everything's allocated until proven otherwise. 276 _bt.set_unallocated_block(end()); 277 } 278 if (!mr.is_empty()) { 279 assert(mr.word_size() >= MinChunkSize, "Chunk size is too small"); 280 _bt.single_block(mr.start(), mr.word_size()); 281 FreeChunk* fc = (FreeChunk*) mr.start(); 282 fc->set_size(mr.word_size()); 283 if (mr.word_size() >= IndexSetSize ) { 284 returnChunkToDictionary(fc); 285 } else { 286 _bt.verify_not_unallocated((HeapWord*)fc, fc->size()); 287 _indexedFreeList[mr.word_size()].return_chunk_at_head(fc); 288 } 289 coalBirth(mr.word_size()); 290 } 291 _promoInfo.reset(); 292 _smallLinearAllocBlock._ptr = NULL; 293 _smallLinearAllocBlock._word_size = 0; 294 } 295 296 void CompactibleFreeListSpace::reset_after_compaction() { 297 // Reset the space to the new reality - one free chunk. 298 MemRegion mr(compaction_top(), end()); 299 reset(mr); 300 // Now refill the linear allocation block(s) if possible. 301 if (_adaptive_freelists) { 302 refillLinearAllocBlocksIfNeeded(); 303 } else { 304 // Place as much of mr in the linAB as we can get, 305 // provided it was big enough to go into the dictionary. 306 FreeChunk* fc = dictionary()->find_largest_dict(); 307 if (fc != NULL) { 308 assert(fc->size() == mr.word_size(), 309 "Why was the chunk broken up?"); 310 removeChunkFromDictionary(fc); 311 HeapWord* addr = (HeapWord*) fc; 312 _smallLinearAllocBlock.set(addr, fc->size() , 313 1024*SmallForLinearAlloc, fc->size()); 314 // Note that _unallocated_block is not updated here. 315 } 316 } 317 } 318 319 // Walks the entire dictionary, returning a coterminal 320 // chunk, if it exists. Use with caution since it involves 321 // a potentially complete walk of a potentially large tree. 322 FreeChunk* CompactibleFreeListSpace::find_chunk_at_end() { 323 324 assert_lock_strong(&_freelistLock); 325 326 return dictionary()->find_chunk_ends_at(end()); 327 } 328 329 330 #ifndef PRODUCT 331 void CompactibleFreeListSpace::initializeIndexedFreeListArrayReturnedBytes() { 332 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 333 _indexedFreeList[i].allocation_stats()->set_returned_bytes(0); 334 } 335 } 336 337 size_t CompactibleFreeListSpace::sumIndexedFreeListArrayReturnedBytes() { 338 size_t sum = 0; 339 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 340 sum += _indexedFreeList[i].allocation_stats()->returned_bytes(); 341 } 342 return sum; 343 } 344 345 size_t CompactibleFreeListSpace::totalCountInIndexedFreeLists() const { 346 size_t count = 0; 347 for (size_t i = IndexSetStart; i < IndexSetSize; i++) { 348 debug_only( 349 ssize_t total_list_count = 0; 350 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 351 fc = fc->next()) { 352 total_list_count++; 353 } 354 assert(total_list_count == _indexedFreeList[i].count(), 355 "Count in list is incorrect"); 356 ) 357 count += _indexedFreeList[i].count(); 358 } 359 return count; 360 } 361 362 size_t CompactibleFreeListSpace::totalCount() { 363 size_t num = totalCountInIndexedFreeLists(); 364 num += dictionary()->total_count(); 365 if (_smallLinearAllocBlock._word_size != 0) { 366 num++; 367 } 368 return num; 369 } 370 #endif 371 372 bool CompactibleFreeListSpace::is_free_block(const HeapWord* p) const { 373 FreeChunk* fc = (FreeChunk*) p; 374 return fc->is_free(); 375 } 376 377 size_t CompactibleFreeListSpace::used() const { 378 return capacity() - free(); 379 } 380 381 size_t CompactibleFreeListSpace::free() const { 382 // "MT-safe, but not MT-precise"(TM), if you will: i.e. 383 // if you do this while the structures are in flux you 384 // may get an approximate answer only; for instance 385 // because there is concurrent allocation either 386 // directly by mutators or for promotion during a GC. 387 // It's "MT-safe", however, in the sense that you are guaranteed 388 // not to crash and burn, for instance, because of walking 389 // pointers that could disappear as you were walking them. 390 // The approximation is because the various components 391 // that are read below are not read atomically (and 392 // further the computation of totalSizeInIndexedFreeLists() 393 // is itself a non-atomic computation. The normal use of 394 // this is during a resize operation at the end of GC 395 // and at that time you are guaranteed to get the 396 // correct actual value. However, for instance, this is 397 // also read completely asynchronously by the "perf-sampler" 398 // that supports jvmstat, and you are apt to see the values 399 // flicker in such cases. 400 assert(_dictionary != NULL, "No _dictionary?"); 401 return (_dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())) + 402 totalSizeInIndexedFreeLists() + 403 _smallLinearAllocBlock._word_size) * HeapWordSize; 404 } 405 406 size_t CompactibleFreeListSpace::max_alloc_in_words() const { 407 assert(_dictionary != NULL, "No _dictionary?"); 408 assert_locked(); 409 size_t res = _dictionary->max_chunk_size(); 410 res = MAX2(res, MIN2(_smallLinearAllocBlock._word_size, 411 (size_t) SmallForLinearAlloc - 1)); 412 // XXX the following could potentially be pretty slow; 413 // should one, pessimistically for the rare cases when res 414 // calculated above is less than IndexSetSize, 415 // just return res calculated above? My reasoning was that 416 // those cases will be so rare that the extra time spent doesn't 417 // really matter.... 418 // Note: do not change the loop test i >= res + IndexSetStride 419 // to i > res below, because i is unsigned and res may be zero. 420 for (size_t i = IndexSetSize - 1; i >= res + IndexSetStride; 421 i -= IndexSetStride) { 422 if (_indexedFreeList[i].head() != NULL) { 423 assert(_indexedFreeList[i].count() != 0, "Inconsistent FreeList"); 424 return i; 425 } 426 } 427 return res; 428 } 429 430 void LinearAllocBlock::print_on(outputStream* st) const { 431 st->print_cr(" LinearAllocBlock: ptr = " PTR_FORMAT ", word_size = " SIZE_FORMAT 432 ", refillsize = " SIZE_FORMAT ", allocation_size_limit = " SIZE_FORMAT, 433 p2i(_ptr), _word_size, _refillSize, _allocation_size_limit); 434 } 435 436 void CompactibleFreeListSpace::print_on(outputStream* st) const { 437 st->print_cr("COMPACTIBLE FREELIST SPACE"); 438 st->print_cr(" Space:"); 439 Space::print_on(st); 440 441 st->print_cr("promoInfo:"); 442 _promoInfo.print_on(st); 443 444 st->print_cr("_smallLinearAllocBlock"); 445 _smallLinearAllocBlock.print_on(st); 446 447 // dump_memory_block(_smallLinearAllocBlock->_ptr, 128); 448 449 st->print_cr(" _fitStrategy = %s, _adaptive_freelists = %s", 450 _fitStrategy?"true":"false", _adaptive_freelists?"true":"false"); 451 } 452 453 void CompactibleFreeListSpace::print_indexed_free_lists(outputStream* st) 454 const { 455 reportIndexedFreeListStatistics(); 456 gclog_or_tty->print_cr("Layout of Indexed Freelists"); 457 gclog_or_tty->print_cr("---------------------------"); 458 AdaptiveFreeList<FreeChunk>::print_labels_on(st, "size"); 459 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 460 _indexedFreeList[i].print_on(gclog_or_tty); 461 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 462 fc = fc->next()) { 463 gclog_or_tty->print_cr("\t[" PTR_FORMAT "," PTR_FORMAT ") %s", 464 p2i(fc), p2i((HeapWord*)fc + i), 465 fc->cantCoalesce() ? "\t CC" : ""); 466 } 467 } 468 } 469 470 void CompactibleFreeListSpace::print_promo_info_blocks(outputStream* st) 471 const { 472 _promoInfo.print_on(st); 473 } 474 475 void CompactibleFreeListSpace::print_dictionary_free_lists(outputStream* st) 476 const { 477 _dictionary->report_statistics(); 478 st->print_cr("Layout of Freelists in Tree"); 479 st->print_cr("---------------------------"); 480 _dictionary->print_free_lists(st); 481 } 482 483 class BlkPrintingClosure: public BlkClosure { 484 const CMSCollector* _collector; 485 const CompactibleFreeListSpace* _sp; 486 const CMSBitMap* _live_bit_map; 487 const bool _post_remark; 488 outputStream* _st; 489 public: 490 BlkPrintingClosure(const CMSCollector* collector, 491 const CompactibleFreeListSpace* sp, 492 const CMSBitMap* live_bit_map, 493 outputStream* st): 494 _collector(collector), 495 _sp(sp), 496 _live_bit_map(live_bit_map), 497 _post_remark(collector->abstract_state() > CMSCollector::FinalMarking), 498 _st(st) { } 499 size_t do_blk(HeapWord* addr); 500 }; 501 502 size_t BlkPrintingClosure::do_blk(HeapWord* addr) { 503 size_t sz = _sp->block_size_no_stall(addr, _collector); 504 assert(sz != 0, "Should always be able to compute a size"); 505 if (_sp->block_is_obj(addr)) { 506 const bool dead = _post_remark && !_live_bit_map->isMarked(addr); 507 _st->print_cr(PTR_FORMAT ": %s object of size " SIZE_FORMAT "%s", 508 p2i(addr), 509 dead ? "dead" : "live", 510 sz, 511 (!dead && CMSPrintObjectsInDump) ? ":" : "."); 512 if (CMSPrintObjectsInDump && !dead) { 513 oop(addr)->print_on(_st); 514 _st->print_cr("--------------------------------------"); 515 } 516 } else { // free block 517 _st->print_cr(PTR_FORMAT ": free block of size " SIZE_FORMAT "%s", 518 p2i(addr), sz, CMSPrintChunksInDump ? ":" : "."); 519 if (CMSPrintChunksInDump) { 520 ((FreeChunk*)addr)->print_on(_st); 521 _st->print_cr("--------------------------------------"); 522 } 523 } 524 return sz; 525 } 526 527 void CompactibleFreeListSpace::dump_at_safepoint_with_locks(CMSCollector* c, 528 outputStream* st) { 529 st->print_cr("\n========================="); 530 st->print_cr("Block layout in CMS Heap:"); 531 st->print_cr("========================="); 532 BlkPrintingClosure bpcl(c, this, c->markBitMap(), st); 533 blk_iterate(&bpcl); 534 535 st->print_cr("\n======================================="); 536 st->print_cr("Order & Layout of Promotion Info Blocks"); 537 st->print_cr("======================================="); 538 print_promo_info_blocks(st); 539 540 st->print_cr("\n==========================="); 541 st->print_cr("Order of Indexed Free Lists"); 542 st->print_cr("========================="); 543 print_indexed_free_lists(st); 544 545 st->print_cr("\n================================="); 546 st->print_cr("Order of Free Lists in Dictionary"); 547 st->print_cr("================================="); 548 print_dictionary_free_lists(st); 549 } 550 551 552 void CompactibleFreeListSpace::reportFreeListStatistics() const { 553 assert_lock_strong(&_freelistLock); 554 assert(PrintFLSStatistics != 0, "Reporting error"); 555 _dictionary->report_statistics(); 556 if (PrintFLSStatistics > 1) { 557 reportIndexedFreeListStatistics(); 558 size_t total_size = totalSizeInIndexedFreeLists() + 559 _dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())); 560 gclog_or_tty->print(" free=" SIZE_FORMAT " frag=%1.4f\n", total_size, flsFrag()); 561 } 562 } 563 564 void CompactibleFreeListSpace::reportIndexedFreeListStatistics() const { 565 assert_lock_strong(&_freelistLock); 566 gclog_or_tty->print("Statistics for IndexedFreeLists:\n" 567 "--------------------------------\n"); 568 size_t total_size = totalSizeInIndexedFreeLists(); 569 size_t free_blocks = numFreeBlocksInIndexedFreeLists(); 570 gclog_or_tty->print("Total Free Space: " SIZE_FORMAT "\n", total_size); 571 gclog_or_tty->print("Max Chunk Size: " SIZE_FORMAT "\n", maxChunkSizeInIndexedFreeLists()); 572 gclog_or_tty->print("Number of Blocks: " SIZE_FORMAT "\n", free_blocks); 573 if (free_blocks != 0) { 574 gclog_or_tty->print("Av. Block Size: " SIZE_FORMAT "\n", total_size/free_blocks); 575 } 576 } 577 578 size_t CompactibleFreeListSpace::numFreeBlocksInIndexedFreeLists() const { 579 size_t res = 0; 580 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 581 debug_only( 582 ssize_t recount = 0; 583 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 584 fc = fc->next()) { 585 recount += 1; 586 } 587 assert(recount == _indexedFreeList[i].count(), 588 "Incorrect count in list"); 589 ) 590 res += _indexedFreeList[i].count(); 591 } 592 return res; 593 } 594 595 size_t CompactibleFreeListSpace::maxChunkSizeInIndexedFreeLists() const { 596 for (size_t i = IndexSetSize - 1; i != 0; i -= IndexSetStride) { 597 if (_indexedFreeList[i].head() != NULL) { 598 assert(_indexedFreeList[i].count() != 0, "Inconsistent FreeList"); 599 return (size_t)i; 600 } 601 } 602 return 0; 603 } 604 605 void CompactibleFreeListSpace::set_end(HeapWord* value) { 606 HeapWord* prevEnd = end(); 607 assert(prevEnd != value, "unnecessary set_end call"); 608 assert(prevEnd == NULL || !BlockOffsetArrayUseUnallocatedBlock || value >= unallocated_block(), 609 "New end is below unallocated block"); 610 _end = value; 611 if (prevEnd != NULL) { 612 // Resize the underlying block offset table. 613 _bt.resize(pointer_delta(value, bottom())); 614 if (value <= prevEnd) { 615 assert(!BlockOffsetArrayUseUnallocatedBlock || value >= unallocated_block(), 616 "New end is below unallocated block"); 617 } else { 618 // Now, take this new chunk and add it to the free blocks. 619 // Note that the BOT has not yet been updated for this block. 620 size_t newFcSize = pointer_delta(value, prevEnd); 621 // XXX This is REALLY UGLY and should be fixed up. XXX 622 if (!_adaptive_freelists && _smallLinearAllocBlock._ptr == NULL) { 623 // Mark the boundary of the new block in BOT 624 _bt.mark_block(prevEnd, value); 625 // put it all in the linAB 626 if (ParallelGCThreads == 0) { 627 _smallLinearAllocBlock._ptr = prevEnd; 628 _smallLinearAllocBlock._word_size = newFcSize; 629 repairLinearAllocBlock(&_smallLinearAllocBlock); 630 } else { // ParallelGCThreads > 0 631 MutexLockerEx x(parDictionaryAllocLock(), 632 Mutex::_no_safepoint_check_flag); 633 _smallLinearAllocBlock._ptr = prevEnd; 634 _smallLinearAllocBlock._word_size = newFcSize; 635 repairLinearAllocBlock(&_smallLinearAllocBlock); 636 } 637 // Births of chunks put into a LinAB are not recorded. Births 638 // of chunks as they are allocated out of a LinAB are. 639 } else { 640 // Add the block to the free lists, if possible coalescing it 641 // with the last free block, and update the BOT and census data. 642 addChunkToFreeListsAtEndRecordingStats(prevEnd, newFcSize); 643 } 644 } 645 } 646 } 647 648 class FreeListSpace_DCTOC : public Filtering_DCTOC { 649 CompactibleFreeListSpace* _cfls; 650 CMSCollector* _collector; 651 protected: 652 // Override. 653 #define walk_mem_region_with_cl_DECL(ClosureType) \ 654 virtual void walk_mem_region_with_cl(MemRegion mr, \ 655 HeapWord* bottom, HeapWord* top, \ 656 ClosureType* cl); \ 657 void walk_mem_region_with_cl_par(MemRegion mr, \ 658 HeapWord* bottom, HeapWord* top, \ 659 ClosureType* cl); \ 660 void walk_mem_region_with_cl_nopar(MemRegion mr, \ 661 HeapWord* bottom, HeapWord* top, \ 662 ClosureType* cl) 663 walk_mem_region_with_cl_DECL(ExtendedOopClosure); 664 walk_mem_region_with_cl_DECL(FilteringClosure); 665 666 public: 667 FreeListSpace_DCTOC(CompactibleFreeListSpace* sp, 668 CMSCollector* collector, 669 ExtendedOopClosure* cl, 670 CardTableModRefBS::PrecisionStyle precision, 671 HeapWord* boundary) : 672 Filtering_DCTOC(sp, cl, precision, boundary), 673 _cfls(sp), _collector(collector) {} 674 }; 675 676 // We de-virtualize the block-related calls below, since we know that our 677 // space is a CompactibleFreeListSpace. 678 679 #define FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(ClosureType) \ 680 void FreeListSpace_DCTOC::walk_mem_region_with_cl(MemRegion mr, \ 681 HeapWord* bottom, \ 682 HeapWord* top, \ 683 ClosureType* cl) { \ 684 bool is_par = SharedHeap::heap()->n_par_threads() > 0; \ 685 if (is_par) { \ 686 assert(SharedHeap::heap()->n_par_threads() == \ 687 SharedHeap::heap()->workers()->active_workers(), "Mismatch"); \ 688 walk_mem_region_with_cl_par(mr, bottom, top, cl); \ 689 } else { \ 690 walk_mem_region_with_cl_nopar(mr, bottom, top, cl); \ 691 } \ 692 } \ 693 void FreeListSpace_DCTOC::walk_mem_region_with_cl_par(MemRegion mr, \ 694 HeapWord* bottom, \ 695 HeapWord* top, \ 696 ClosureType* cl) { \ 697 /* Skip parts that are before "mr", in case "block_start" sent us \ 698 back too far. */ \ 699 HeapWord* mr_start = mr.start(); \ 700 size_t bot_size = _cfls->CompactibleFreeListSpace::block_size(bottom); \ 701 HeapWord* next = bottom + bot_size; \ 702 while (next < mr_start) { \ 703 bottom = next; \ 704 bot_size = _cfls->CompactibleFreeListSpace::block_size(bottom); \ 705 next = bottom + bot_size; \ 706 } \ 707 \ 708 while (bottom < top) { \ 709 if (_cfls->CompactibleFreeListSpace::block_is_obj(bottom) && \ 710 !_cfls->CompactibleFreeListSpace::obj_allocated_since_save_marks( \ 711 oop(bottom)) && \ 712 !_collector->CMSCollector::is_dead_obj(oop(bottom))) { \ 713 size_t word_sz = oop(bottom)->oop_iterate(cl, mr); \ 714 bottom += _cfls->adjustObjectSize(word_sz); \ 715 } else { \ 716 bottom += _cfls->CompactibleFreeListSpace::block_size(bottom); \ 717 } \ 718 } \ 719 } \ 720 void FreeListSpace_DCTOC::walk_mem_region_with_cl_nopar(MemRegion mr, \ 721 HeapWord* bottom, \ 722 HeapWord* top, \ 723 ClosureType* cl) { \ 724 /* Skip parts that are before "mr", in case "block_start" sent us \ 725 back too far. */ \ 726 HeapWord* mr_start = mr.start(); \ 727 size_t bot_size = _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \ 728 HeapWord* next = bottom + bot_size; \ 729 while (next < mr_start) { \ 730 bottom = next; \ 731 bot_size = _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \ 732 next = bottom + bot_size; \ 733 } \ 734 \ 735 while (bottom < top) { \ 736 if (_cfls->CompactibleFreeListSpace::block_is_obj_nopar(bottom) && \ 737 !_cfls->CompactibleFreeListSpace::obj_allocated_since_save_marks( \ 738 oop(bottom)) && \ 739 !_collector->CMSCollector::is_dead_obj(oop(bottom))) { \ 740 size_t word_sz = oop(bottom)->oop_iterate(cl, mr); \ 741 bottom += _cfls->adjustObjectSize(word_sz); \ 742 } else { \ 743 bottom += _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \ 744 } \ 745 } \ 746 } 747 748 // (There are only two of these, rather than N, because the split is due 749 // only to the introduction of the FilteringClosure, a local part of the 750 // impl of this abstraction.) 751 FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(ExtendedOopClosure) 752 FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(FilteringClosure) 753 754 DirtyCardToOopClosure* 755 CompactibleFreeListSpace::new_dcto_cl(ExtendedOopClosure* cl, 756 CardTableModRefBS::PrecisionStyle precision, 757 HeapWord* boundary) { 758 return new FreeListSpace_DCTOC(this, _collector, cl, precision, boundary); 759 } 760 761 762 // Note on locking for the space iteration functions: 763 // since the collector's iteration activities are concurrent with 764 // allocation activities by mutators, absent a suitable mutual exclusion 765 // mechanism the iterators may go awry. For instance a block being iterated 766 // may suddenly be allocated or divided up and part of it allocated and 767 // so on. 768 769 // Apply the given closure to each block in the space. 770 void CompactibleFreeListSpace::blk_iterate_careful(BlkClosureCareful* cl) { 771 assert_lock_strong(freelistLock()); 772 HeapWord *cur, *limit; 773 for (cur = bottom(), limit = end(); cur < limit; 774 cur += cl->do_blk_careful(cur)); 775 } 776 777 // Apply the given closure to each block in the space. 778 void CompactibleFreeListSpace::blk_iterate(BlkClosure* cl) { 779 assert_lock_strong(freelistLock()); 780 HeapWord *cur, *limit; 781 for (cur = bottom(), limit = end(); cur < limit; 782 cur += cl->do_blk(cur)); 783 } 784 785 // Apply the given closure to each oop in the space. 786 void CompactibleFreeListSpace::oop_iterate(ExtendedOopClosure* cl) { 787 assert_lock_strong(freelistLock()); 788 HeapWord *cur, *limit; 789 size_t curSize; 790 for (cur = bottom(), limit = end(); cur < limit; 791 cur += curSize) { 792 curSize = block_size(cur); 793 if (block_is_obj(cur)) { 794 oop(cur)->oop_iterate(cl); 795 } 796 } 797 } 798 799 // NOTE: In the following methods, in order to safely be able to 800 // apply the closure to an object, we need to be sure that the 801 // object has been initialized. We are guaranteed that an object 802 // is initialized if we are holding the Heap_lock with the 803 // world stopped. 804 void CompactibleFreeListSpace::verify_objects_initialized() const { 805 if (is_init_completed()) { 806 assert_locked_or_safepoint(Heap_lock); 807 if (Universe::is_fully_initialized()) { 808 guarantee(SafepointSynchronize::is_at_safepoint(), 809 "Required for objects to be initialized"); 810 } 811 } // else make a concession at vm start-up 812 } 813 814 // Apply the given closure to each object in the space 815 void CompactibleFreeListSpace::object_iterate(ObjectClosure* blk) { 816 assert_lock_strong(freelistLock()); 817 NOT_PRODUCT(verify_objects_initialized()); 818 HeapWord *cur, *limit; 819 size_t curSize; 820 for (cur = bottom(), limit = end(); cur < limit; 821 cur += curSize) { 822 curSize = block_size(cur); 823 if (block_is_obj(cur)) { 824 blk->do_object(oop(cur)); 825 } 826 } 827 } 828 829 // Apply the given closure to each live object in the space 830 // The usage of CompactibleFreeListSpace 831 // by the ConcurrentMarkSweepGeneration for concurrent GC's allows 832 // objects in the space with references to objects that are no longer 833 // valid. For example, an object may reference another object 834 // that has already been sweep up (collected). This method uses 835 // obj_is_alive() to determine whether it is safe to apply the closure to 836 // an object. See obj_is_alive() for details on how liveness of an 837 // object is decided. 838 839 void CompactibleFreeListSpace::safe_object_iterate(ObjectClosure* blk) { 840 assert_lock_strong(freelistLock()); 841 NOT_PRODUCT(verify_objects_initialized()); 842 HeapWord *cur, *limit; 843 size_t curSize; 844 for (cur = bottom(), limit = end(); cur < limit; 845 cur += curSize) { 846 curSize = block_size(cur); 847 if (block_is_obj(cur) && obj_is_alive(cur)) { 848 blk->do_object(oop(cur)); 849 } 850 } 851 } 852 853 void CompactibleFreeListSpace::object_iterate_mem(MemRegion mr, 854 UpwardsObjectClosure* cl) { 855 assert_locked(freelistLock()); 856 NOT_PRODUCT(verify_objects_initialized()); 857 assert(!mr.is_empty(), "Should be non-empty"); 858 // We use MemRegion(bottom(), end()) rather than used_region() below 859 // because the two are not necessarily equal for some kinds of 860 // spaces, in particular, certain kinds of free list spaces. 861 // We could use the more complicated but more precise: 862 // MemRegion(used_region().start(), round_to(used_region().end(), CardSize)) 863 // but the slight imprecision seems acceptable in the assertion check. 864 assert(MemRegion(bottom(), end()).contains(mr), 865 "Should be within used space"); 866 HeapWord* prev = cl->previous(); // max address from last time 867 if (prev >= mr.end()) { // nothing to do 868 return; 869 } 870 // This assert will not work when we go from cms space to perm 871 // space, and use same closure. Easy fix deferred for later. XXX YSR 872 // assert(prev == NULL || contains(prev), "Should be within space"); 873 874 bool last_was_obj_array = false; 875 HeapWord *blk_start_addr, *region_start_addr; 876 if (prev > mr.start()) { 877 region_start_addr = prev; 878 blk_start_addr = prev; 879 // The previous invocation may have pushed "prev" beyond the 880 // last allocated block yet there may be still be blocks 881 // in this region due to a particular coalescing policy. 882 // Relax the assertion so that the case where the unallocated 883 // block is maintained and "prev" is beyond the unallocated 884 // block does not cause the assertion to fire. 885 assert((BlockOffsetArrayUseUnallocatedBlock && 886 (!is_in(prev))) || 887 (blk_start_addr == block_start(region_start_addr)), "invariant"); 888 } else { 889 region_start_addr = mr.start(); 890 blk_start_addr = block_start(region_start_addr); 891 } 892 HeapWord* region_end_addr = mr.end(); 893 MemRegion derived_mr(region_start_addr, region_end_addr); 894 while (blk_start_addr < region_end_addr) { 895 const size_t size = block_size(blk_start_addr); 896 if (block_is_obj(blk_start_addr)) { 897 last_was_obj_array = cl->do_object_bm(oop(blk_start_addr), derived_mr); 898 } else { 899 last_was_obj_array = false; 900 } 901 blk_start_addr += size; 902 } 903 if (!last_was_obj_array) { 904 assert((bottom() <= blk_start_addr) && (blk_start_addr <= end()), 905 "Should be within (closed) used space"); 906 assert(blk_start_addr > prev, "Invariant"); 907 cl->set_previous(blk_start_addr); // min address for next time 908 } 909 } 910 911 912 // Callers of this iterator beware: The closure application should 913 // be robust in the face of uninitialized objects and should (always) 914 // return a correct size so that the next addr + size below gives us a 915 // valid block boundary. [See for instance, 916 // ScanMarkedObjectsAgainCarefullyClosure::do_object_careful() 917 // in ConcurrentMarkSweepGeneration.cpp.] 918 HeapWord* 919 CompactibleFreeListSpace::object_iterate_careful_m(MemRegion mr, 920 ObjectClosureCareful* cl) { 921 assert_lock_strong(freelistLock()); 922 // Can't use used_region() below because it may not necessarily 923 // be the same as [bottom(),end()); although we could 924 // use [used_region().start(),round_to(used_region().end(),CardSize)), 925 // that appears too cumbersome, so we just do the simpler check 926 // in the assertion below. 927 assert(!mr.is_empty() && MemRegion(bottom(),end()).contains(mr), 928 "mr should be non-empty and within used space"); 929 HeapWord *addr, *end; 930 size_t size; 931 for (addr = block_start_careful(mr.start()), end = mr.end(); 932 addr < end; addr += size) { 933 FreeChunk* fc = (FreeChunk*)addr; 934 if (fc->is_free()) { 935 // Since we hold the free list lock, which protects direct 936 // allocation in this generation by mutators, a free object 937 // will remain free throughout this iteration code. 938 size = fc->size(); 939 } else { 940 // Note that the object need not necessarily be initialized, 941 // because (for instance) the free list lock does NOT protect 942 // object initialization. The closure application below must 943 // therefore be correct in the face of uninitialized objects. 944 size = cl->do_object_careful_m(oop(addr), mr); 945 if (size == 0) { 946 // An unparsable object found. Signal early termination. 947 return addr; 948 } 949 } 950 } 951 return NULL; 952 } 953 954 955 HeapWord* CompactibleFreeListSpace::block_start_const(const void* p) const { 956 NOT_PRODUCT(verify_objects_initialized()); 957 return _bt.block_start(p); 958 } 959 960 HeapWord* CompactibleFreeListSpace::block_start_careful(const void* p) const { 961 return _bt.block_start_careful(p); 962 } 963 964 size_t CompactibleFreeListSpace::block_size(const HeapWord* p) const { 965 NOT_PRODUCT(verify_objects_initialized()); 966 // This must be volatile, or else there is a danger that the compiler 967 // will compile the code below into a sometimes-infinite loop, by keeping 968 // the value read the first time in a register. 969 while (true) { 970 // We must do this until we get a consistent view of the object. 971 if (FreeChunk::indicatesFreeChunk(p)) { 972 volatile FreeChunk* fc = (volatile FreeChunk*)p; 973 size_t res = fc->size(); 974 975 // Bugfix for systems with weak memory model (PPC64/IA64). The 976 // block's free bit was set and we have read the size of the 977 // block. Acquire and check the free bit again. If the block is 978 // still free, the read size is correct. 979 OrderAccess::acquire(); 980 981 // If the object is still a free chunk, return the size, else it 982 // has been allocated so try again. 983 if (FreeChunk::indicatesFreeChunk(p)) { 984 assert(res != 0, "Block size should not be 0"); 985 return res; 986 } 987 } else { 988 // must read from what 'p' points to in each loop. 989 Klass* k = ((volatile oopDesc*)p)->klass_or_null(); 990 if (k != NULL) { 991 assert(k->is_klass(), "Should really be klass oop."); 992 oop o = (oop)p; 993 assert(o->is_oop(true /* ignore mark word */), "Should be an oop."); 994 995 // Bugfix for systems with weak memory model (PPC64/IA64). 996 // The object o may be an array. Acquire to make sure that the array 997 // size (third word) is consistent. 998 OrderAccess::acquire(); 999 1000 size_t res = o->size_given_klass(k); 1001 res = adjustObjectSize(res); 1002 assert(res != 0, "Block size should not be 0"); 1003 return res; 1004 } 1005 } 1006 } 1007 } 1008 1009 // TODO: Now that is_parsable is gone, we should combine these two functions. 1010 // A variant of the above that uses the Printezis bits for 1011 // unparsable but allocated objects. This avoids any possible 1012 // stalls waiting for mutators to initialize objects, and is 1013 // thus potentially faster than the variant above. However, 1014 // this variant may return a zero size for a block that is 1015 // under mutation and for which a consistent size cannot be 1016 // inferred without stalling; see CMSCollector::block_size_if_printezis_bits(). 1017 size_t CompactibleFreeListSpace::block_size_no_stall(HeapWord* p, 1018 const CMSCollector* c) 1019 const { 1020 assert(MemRegion(bottom(), end()).contains(p), "p not in space"); 1021 // This must be volatile, or else there is a danger that the compiler 1022 // will compile the code below into a sometimes-infinite loop, by keeping 1023 // the value read the first time in a register. 1024 DEBUG_ONLY(uint loops = 0;) 1025 while (true) { 1026 // We must do this until we get a consistent view of the object. 1027 if (FreeChunk::indicatesFreeChunk(p)) { 1028 volatile FreeChunk* fc = (volatile FreeChunk*)p; 1029 size_t res = fc->size(); 1030 1031 // Bugfix for systems with weak memory model (PPC64/IA64). The 1032 // free bit of the block was set and we have read the size of 1033 // the block. Acquire and check the free bit again. If the 1034 // block is still free, the read size is correct. 1035 OrderAccess::acquire(); 1036 1037 if (FreeChunk::indicatesFreeChunk(p)) { 1038 assert(res != 0, "Block size should not be 0"); 1039 assert(loops == 0, "Should be 0"); 1040 return res; 1041 } 1042 } else { 1043 // must read from what 'p' points to in each loop. 1044 Klass* k = ((volatile oopDesc*)p)->klass_or_null(); 1045 // We trust the size of any object that has a non-NULL 1046 // klass and (for those in the perm gen) is parsable 1047 // -- irrespective of its conc_safe-ty. 1048 if (k != NULL) { 1049 assert(k->is_klass(), "Should really be klass oop."); 1050 oop o = (oop)p; 1051 assert(o->is_oop(), "Should be an oop"); 1052 1053 // Bugfix for systems with weak memory model (PPC64/IA64). 1054 // The object o may be an array. Acquire to make sure that the array 1055 // size (third word) is consistent. 1056 OrderAccess::acquire(); 1057 1058 size_t res = o->size_given_klass(k); 1059 res = adjustObjectSize(res); 1060 assert(res != 0, "Block size should not be 0"); 1061 return res; 1062 } else { 1063 // May return 0 if P-bits not present. 1064 return c->block_size_if_printezis_bits(p); 1065 } 1066 } 1067 assert(loops == 0, "Can loop at most once"); 1068 DEBUG_ONLY(loops++;) 1069 } 1070 } 1071 1072 size_t CompactibleFreeListSpace::block_size_nopar(const HeapWord* p) const { 1073 NOT_PRODUCT(verify_objects_initialized()); 1074 assert(MemRegion(bottom(), end()).contains(p), "p not in space"); 1075 FreeChunk* fc = (FreeChunk*)p; 1076 if (fc->is_free()) { 1077 return fc->size(); 1078 } else { 1079 // Ignore mark word because this may be a recently promoted 1080 // object whose mark word is used to chain together grey 1081 // objects (the last one would have a null value). 1082 assert(oop(p)->is_oop(true), "Should be an oop"); 1083 return adjustObjectSize(oop(p)->size()); 1084 } 1085 } 1086 1087 // This implementation assumes that the property of "being an object" is 1088 // stable. But being a free chunk may not be (because of parallel 1089 // promotion.) 1090 bool CompactibleFreeListSpace::block_is_obj(const HeapWord* p) const { 1091 FreeChunk* fc = (FreeChunk*)p; 1092 assert(is_in_reserved(p), "Should be in space"); 1093 // When doing a mark-sweep-compact of the CMS generation, this 1094 // assertion may fail because prepare_for_compaction() uses 1095 // space that is garbage to maintain information on ranges of 1096 // live objects so that these live ranges can be moved as a whole. 1097 // Comment out this assertion until that problem can be solved 1098 // (i.e., that the block start calculation may look at objects 1099 // at address below "p" in finding the object that contains "p" 1100 // and those objects (if garbage) may have been modified to hold 1101 // live range information. 1102 // assert(CollectedHeap::use_parallel_gc_threads() || _bt.block_start(p) == p, 1103 // "Should be a block boundary"); 1104 if (FreeChunk::indicatesFreeChunk(p)) return false; 1105 Klass* k = oop(p)->klass_or_null(); 1106 if (k != NULL) { 1107 // Ignore mark word because it may have been used to 1108 // chain together promoted objects (the last one 1109 // would have a null value). 1110 assert(oop(p)->is_oop(true), "Should be an oop"); 1111 return true; 1112 } else { 1113 return false; // Was not an object at the start of collection. 1114 } 1115 } 1116 1117 // Check if the object is alive. This fact is checked either by consulting 1118 // the main marking bitmap in the sweeping phase or, if it's a permanent 1119 // generation and we're not in the sweeping phase, by checking the 1120 // perm_gen_verify_bit_map where we store the "deadness" information if 1121 // we did not sweep the perm gen in the most recent previous GC cycle. 1122 bool CompactibleFreeListSpace::obj_is_alive(const HeapWord* p) const { 1123 assert(SafepointSynchronize::is_at_safepoint() || !is_init_completed(), 1124 "Else races are possible"); 1125 assert(block_is_obj(p), "The address should point to an object"); 1126 1127 // If we're sweeping, we use object liveness information from the main bit map 1128 // for both perm gen and old gen. 1129 // We don't need to lock the bitmap (live_map or dead_map below), because 1130 // EITHER we are in the middle of the sweeping phase, and the 1131 // main marking bit map (live_map below) is locked, 1132 // OR we're in other phases and perm_gen_verify_bit_map (dead_map below) 1133 // is stable, because it's mutated only in the sweeping phase. 1134 // NOTE: This method is also used by jmap where, if class unloading is 1135 // off, the results can return "false" for legitimate perm objects, 1136 // when we are not in the midst of a sweeping phase, which can result 1137 // in jmap not reporting certain perm gen objects. This will be moot 1138 // if/when the perm gen goes away in the future. 1139 if (_collector->abstract_state() == CMSCollector::Sweeping) { 1140 CMSBitMap* live_map = _collector->markBitMap(); 1141 return live_map->par_isMarked((HeapWord*) p); 1142 } 1143 return true; 1144 } 1145 1146 bool CompactibleFreeListSpace::block_is_obj_nopar(const HeapWord* p) const { 1147 FreeChunk* fc = (FreeChunk*)p; 1148 assert(is_in_reserved(p), "Should be in space"); 1149 assert(_bt.block_start(p) == p, "Should be a block boundary"); 1150 if (!fc->is_free()) { 1151 // Ignore mark word because it may have been used to 1152 // chain together promoted objects (the last one 1153 // would have a null value). 1154 assert(oop(p)->is_oop(true), "Should be an oop"); 1155 return true; 1156 } 1157 return false; 1158 } 1159 1160 // "MT-safe but not guaranteed MT-precise" (TM); you may get an 1161 // approximate answer if you don't hold the freelistlock when you call this. 1162 size_t CompactibleFreeListSpace::totalSizeInIndexedFreeLists() const { 1163 size_t size = 0; 1164 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 1165 debug_only( 1166 // We may be calling here without the lock in which case we 1167 // won't do this modest sanity check. 1168 if (freelistLock()->owned_by_self()) { 1169 size_t total_list_size = 0; 1170 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL; 1171 fc = fc->next()) { 1172 total_list_size += i; 1173 } 1174 assert(total_list_size == i * _indexedFreeList[i].count(), 1175 "Count in list is incorrect"); 1176 } 1177 ) 1178 size += i * _indexedFreeList[i].count(); 1179 } 1180 return size; 1181 } 1182 1183 HeapWord* CompactibleFreeListSpace::par_allocate(size_t size) { 1184 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 1185 return allocate(size); 1186 } 1187 1188 HeapWord* 1189 CompactibleFreeListSpace::getChunkFromSmallLinearAllocBlockRemainder(size_t size) { 1190 return getChunkFromLinearAllocBlockRemainder(&_smallLinearAllocBlock, size); 1191 } 1192 1193 HeapWord* CompactibleFreeListSpace::allocate(size_t size) { 1194 assert_lock_strong(freelistLock()); 1195 HeapWord* res = NULL; 1196 assert(size == adjustObjectSize(size), 1197 "use adjustObjectSize() before calling into allocate()"); 1198 1199 if (_adaptive_freelists) { 1200 res = allocate_adaptive_freelists(size); 1201 } else { // non-adaptive free lists 1202 res = allocate_non_adaptive_freelists(size); 1203 } 1204 1205 if (res != NULL) { 1206 // check that res does lie in this space! 1207 assert(is_in_reserved(res), "Not in this space!"); 1208 assert(is_aligned((void*)res), "alignment check"); 1209 1210 FreeChunk* fc = (FreeChunk*)res; 1211 fc->markNotFree(); 1212 assert(!fc->is_free(), "shouldn't be marked free"); 1213 assert(oop(fc)->klass_or_null() == NULL, "should look uninitialized"); 1214 // Verify that the block offset table shows this to 1215 // be a single block, but not one which is unallocated. 1216 _bt.verify_single_block(res, size); 1217 _bt.verify_not_unallocated(res, size); 1218 // mangle a just allocated object with a distinct pattern. 1219 debug_only(fc->mangleAllocated(size)); 1220 } 1221 1222 return res; 1223 } 1224 1225 HeapWord* CompactibleFreeListSpace::allocate_non_adaptive_freelists(size_t size) { 1226 HeapWord* res = NULL; 1227 // try and use linear allocation for smaller blocks 1228 if (size < _smallLinearAllocBlock._allocation_size_limit) { 1229 // if successful, the following also adjusts block offset table 1230 res = getChunkFromSmallLinearAllocBlock(size); 1231 } 1232 // Else triage to indexed lists for smaller sizes 1233 if (res == NULL) { 1234 if (size < SmallForDictionary) { 1235 res = (HeapWord*) getChunkFromIndexedFreeList(size); 1236 } else { 1237 // else get it from the big dictionary; if even this doesn't 1238 // work we are out of luck. 1239 res = (HeapWord*)getChunkFromDictionaryExact(size); 1240 } 1241 } 1242 1243 return res; 1244 } 1245 1246 HeapWord* CompactibleFreeListSpace::allocate_adaptive_freelists(size_t size) { 1247 assert_lock_strong(freelistLock()); 1248 HeapWord* res = NULL; 1249 assert(size == adjustObjectSize(size), 1250 "use adjustObjectSize() before calling into allocate()"); 1251 1252 // Strategy 1253 // if small 1254 // exact size from small object indexed list if small 1255 // small or large linear allocation block (linAB) as appropriate 1256 // take from lists of greater sized chunks 1257 // else 1258 // dictionary 1259 // small or large linear allocation block if it has the space 1260 // Try allocating exact size from indexTable first 1261 if (size < IndexSetSize) { 1262 res = (HeapWord*) getChunkFromIndexedFreeList(size); 1263 if(res != NULL) { 1264 assert(res != (HeapWord*)_indexedFreeList[size].head(), 1265 "Not removed from free list"); 1266 // no block offset table adjustment is necessary on blocks in 1267 // the indexed lists. 1268 1269 // Try allocating from the small LinAB 1270 } else if (size < _smallLinearAllocBlock._allocation_size_limit && 1271 (res = getChunkFromSmallLinearAllocBlock(size)) != NULL) { 1272 // if successful, the above also adjusts block offset table 1273 // Note that this call will refill the LinAB to 1274 // satisfy the request. This is different that 1275 // evm. 1276 // Don't record chunk off a LinAB? smallSplitBirth(size); 1277 } else { 1278 // Raid the exact free lists larger than size, even if they are not 1279 // overpopulated. 1280 res = (HeapWord*) getChunkFromGreater(size); 1281 } 1282 } else { 1283 // Big objects get allocated directly from the dictionary. 1284 res = (HeapWord*) getChunkFromDictionaryExact(size); 1285 if (res == NULL) { 1286 // Try hard not to fail since an allocation failure will likely 1287 // trigger a synchronous GC. Try to get the space from the 1288 // allocation blocks. 1289 res = getChunkFromSmallLinearAllocBlockRemainder(size); 1290 } 1291 } 1292 1293 return res; 1294 } 1295 1296 // A worst-case estimate of the space required (in HeapWords) to expand the heap 1297 // when promoting obj. 1298 size_t CompactibleFreeListSpace::expansionSpaceRequired(size_t obj_size) const { 1299 // Depending on the object size, expansion may require refilling either a 1300 // bigLAB or a smallLAB plus refilling a PromotionInfo object. MinChunkSize 1301 // is added because the dictionary may over-allocate to avoid fragmentation. 1302 size_t space = obj_size; 1303 if (!_adaptive_freelists) { 1304 space = MAX2(space, _smallLinearAllocBlock._refillSize); 1305 } 1306 space += _promoInfo.refillSize() + 2 * MinChunkSize; 1307 return space; 1308 } 1309 1310 FreeChunk* CompactibleFreeListSpace::getChunkFromGreater(size_t numWords) { 1311 FreeChunk* ret; 1312 1313 assert(numWords >= MinChunkSize, "Size is less than minimum"); 1314 assert(linearAllocationWouldFail() || bestFitFirst(), 1315 "Should not be here"); 1316 1317 size_t i; 1318 size_t currSize = numWords + MinChunkSize; 1319 assert(currSize % MinObjAlignment == 0, "currSize should be aligned"); 1320 for (i = currSize; i < IndexSetSize; i += IndexSetStride) { 1321 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[i]; 1322 if (fl->head()) { 1323 ret = getFromListGreater(fl, numWords); 1324 assert(ret == NULL || ret->is_free(), "Should be returning a free chunk"); 1325 return ret; 1326 } 1327 } 1328 1329 currSize = MAX2((size_t)SmallForDictionary, 1330 (size_t)(numWords + MinChunkSize)); 1331 1332 /* Try to get a chunk that satisfies request, while avoiding 1333 fragmentation that can't be handled. */ 1334 { 1335 ret = dictionary()->get_chunk(currSize); 1336 if (ret != NULL) { 1337 assert(ret->size() - numWords >= MinChunkSize, 1338 "Chunk is too small"); 1339 _bt.allocated((HeapWord*)ret, ret->size()); 1340 /* Carve returned chunk. */ 1341 (void) splitChunkAndReturnRemainder(ret, numWords); 1342 /* Label this as no longer a free chunk. */ 1343 assert(ret->is_free(), "This chunk should be free"); 1344 ret->link_prev(NULL); 1345 } 1346 assert(ret == NULL || ret->is_free(), "Should be returning a free chunk"); 1347 return ret; 1348 } 1349 ShouldNotReachHere(); 1350 } 1351 1352 bool CompactibleFreeListSpace::verifyChunkInIndexedFreeLists(FreeChunk* fc) const { 1353 assert(fc->size() < IndexSetSize, "Size of chunk is too large"); 1354 return _indexedFreeList[fc->size()].verify_chunk_in_free_list(fc); 1355 } 1356 1357 bool CompactibleFreeListSpace::verify_chunk_is_linear_alloc_block(FreeChunk* fc) const { 1358 assert((_smallLinearAllocBlock._ptr != (HeapWord*)fc) || 1359 (_smallLinearAllocBlock._word_size == fc->size()), 1360 "Linear allocation block shows incorrect size"); 1361 return ((_smallLinearAllocBlock._ptr == (HeapWord*)fc) && 1362 (_smallLinearAllocBlock._word_size == fc->size())); 1363 } 1364 1365 // Check if the purported free chunk is present either as a linear 1366 // allocation block, the size-indexed table of (smaller) free blocks, 1367 // or the larger free blocks kept in the binary tree dictionary. 1368 bool CompactibleFreeListSpace::verify_chunk_in_free_list(FreeChunk* fc) const { 1369 if (verify_chunk_is_linear_alloc_block(fc)) { 1370 return true; 1371 } else if (fc->size() < IndexSetSize) { 1372 return verifyChunkInIndexedFreeLists(fc); 1373 } else { 1374 return dictionary()->verify_chunk_in_free_list(fc); 1375 } 1376 } 1377 1378 #ifndef PRODUCT 1379 void CompactibleFreeListSpace::assert_locked() const { 1380 CMSLockVerifier::assert_locked(freelistLock(), parDictionaryAllocLock()); 1381 } 1382 1383 void CompactibleFreeListSpace::assert_locked(const Mutex* lock) const { 1384 CMSLockVerifier::assert_locked(lock); 1385 } 1386 #endif 1387 1388 FreeChunk* CompactibleFreeListSpace::allocateScratch(size_t size) { 1389 // In the parallel case, the main thread holds the free list lock 1390 // on behalf the parallel threads. 1391 FreeChunk* fc; 1392 { 1393 // If GC is parallel, this might be called by several threads. 1394 // This should be rare enough that the locking overhead won't affect 1395 // the sequential code. 1396 MutexLockerEx x(parDictionaryAllocLock(), 1397 Mutex::_no_safepoint_check_flag); 1398 fc = getChunkFromDictionary(size); 1399 } 1400 if (fc != NULL) { 1401 fc->dontCoalesce(); 1402 assert(fc->is_free(), "Should be free, but not coalescable"); 1403 // Verify that the block offset table shows this to 1404 // be a single block, but not one which is unallocated. 1405 _bt.verify_single_block((HeapWord*)fc, fc->size()); 1406 _bt.verify_not_unallocated((HeapWord*)fc, fc->size()); 1407 } 1408 return fc; 1409 } 1410 1411 oop CompactibleFreeListSpace::promote(oop obj, size_t obj_size) { 1412 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1413 assert_locked(); 1414 1415 // if we are tracking promotions, then first ensure space for 1416 // promotion (including spooling space for saving header if necessary). 1417 // then allocate and copy, then track promoted info if needed. 1418 // When tracking (see PromotionInfo::track()), the mark word may 1419 // be displaced and in this case restoration of the mark word 1420 // occurs in the (oop_since_save_marks_)iterate phase. 1421 if (_promoInfo.tracking() && !_promoInfo.ensure_spooling_space()) { 1422 return NULL; 1423 } 1424 // Call the allocate(size_t, bool) form directly to avoid the 1425 // additional call through the allocate(size_t) form. Having 1426 // the compile inline the call is problematic because allocate(size_t) 1427 // is a virtual method. 1428 HeapWord* res = allocate(adjustObjectSize(obj_size)); 1429 if (res != NULL) { 1430 Copy::aligned_disjoint_words((HeapWord*)obj, res, obj_size); 1431 // if we should be tracking promotions, do so. 1432 if (_promoInfo.tracking()) { 1433 _promoInfo.track((PromotedObject*)res); 1434 } 1435 } 1436 return oop(res); 1437 } 1438 1439 HeapWord* 1440 CompactibleFreeListSpace::getChunkFromSmallLinearAllocBlock(size_t size) { 1441 assert_locked(); 1442 assert(size >= MinChunkSize, "minimum chunk size"); 1443 assert(size < _smallLinearAllocBlock._allocation_size_limit, 1444 "maximum from smallLinearAllocBlock"); 1445 return getChunkFromLinearAllocBlock(&_smallLinearAllocBlock, size); 1446 } 1447 1448 HeapWord* 1449 CompactibleFreeListSpace::getChunkFromLinearAllocBlock(LinearAllocBlock *blk, 1450 size_t size) { 1451 assert_locked(); 1452 assert(size >= MinChunkSize, "too small"); 1453 HeapWord* res = NULL; 1454 // Try to do linear allocation from blk, making sure that 1455 if (blk->_word_size == 0) { 1456 // We have probably been unable to fill this either in the prologue or 1457 // when it was exhausted at the last linear allocation. Bail out until 1458 // next time. 1459 assert(blk->_ptr == NULL, "consistency check"); 1460 return NULL; 1461 } 1462 assert(blk->_word_size != 0 && blk->_ptr != NULL, "consistency check"); 1463 res = getChunkFromLinearAllocBlockRemainder(blk, size); 1464 if (res != NULL) return res; 1465 1466 // about to exhaust this linear allocation block 1467 if (blk->_word_size == size) { // exactly satisfied 1468 res = blk->_ptr; 1469 _bt.allocated(res, blk->_word_size); 1470 } else if (size + MinChunkSize <= blk->_refillSize) { 1471 size_t sz = blk->_word_size; 1472 // Update _unallocated_block if the size is such that chunk would be 1473 // returned to the indexed free list. All other chunks in the indexed 1474 // free lists are allocated from the dictionary so that _unallocated_block 1475 // has already been adjusted for them. Do it here so that the cost 1476 // for all chunks added back to the indexed free lists. 1477 if (sz < SmallForDictionary) { 1478 _bt.allocated(blk->_ptr, sz); 1479 } 1480 // Return the chunk that isn't big enough, and then refill below. 1481 addChunkToFreeLists(blk->_ptr, sz); 1482 split_birth(sz); 1483 // Don't keep statistics on adding back chunk from a LinAB. 1484 } else { 1485 // A refilled block would not satisfy the request. 1486 return NULL; 1487 } 1488 1489 blk->_ptr = NULL; blk->_word_size = 0; 1490 refillLinearAllocBlock(blk); 1491 assert(blk->_ptr == NULL || blk->_word_size >= size + MinChunkSize, 1492 "block was replenished"); 1493 if (res != NULL) { 1494 split_birth(size); 1495 repairLinearAllocBlock(blk); 1496 } else if (blk->_ptr != NULL) { 1497 res = blk->_ptr; 1498 size_t blk_size = blk->_word_size; 1499 blk->_word_size -= size; 1500 blk->_ptr += size; 1501 split_birth(size); 1502 repairLinearAllocBlock(blk); 1503 // Update BOT last so that other (parallel) GC threads see a consistent 1504 // view of the BOT and free blocks. 1505 // Above must occur before BOT is updated below. 1506 OrderAccess::storestore(); 1507 _bt.split_block(res, blk_size, size); // adjust block offset table 1508 } 1509 return res; 1510 } 1511 1512 HeapWord* CompactibleFreeListSpace::getChunkFromLinearAllocBlockRemainder( 1513 LinearAllocBlock* blk, 1514 size_t size) { 1515 assert_locked(); 1516 assert(size >= MinChunkSize, "too small"); 1517 1518 HeapWord* res = NULL; 1519 // This is the common case. Keep it simple. 1520 if (blk->_word_size >= size + MinChunkSize) { 1521 assert(blk->_ptr != NULL, "consistency check"); 1522 res = blk->_ptr; 1523 // Note that the BOT is up-to-date for the linAB before allocation. It 1524 // indicates the start of the linAB. The split_block() updates the 1525 // BOT for the linAB after the allocation (indicates the start of the 1526 // next chunk to be allocated). 1527 size_t blk_size = blk->_word_size; 1528 blk->_word_size -= size; 1529 blk->_ptr += size; 1530 split_birth(size); 1531 repairLinearAllocBlock(blk); 1532 // Update BOT last so that other (parallel) GC threads see a consistent 1533 // view of the BOT and free blocks. 1534 // Above must occur before BOT is updated below. 1535 OrderAccess::storestore(); 1536 _bt.split_block(res, blk_size, size); // adjust block offset table 1537 _bt.allocated(res, size); 1538 } 1539 return res; 1540 } 1541 1542 FreeChunk* 1543 CompactibleFreeListSpace::getChunkFromIndexedFreeList(size_t size) { 1544 assert_locked(); 1545 assert(size < SmallForDictionary, "just checking"); 1546 FreeChunk* res; 1547 res = _indexedFreeList[size].get_chunk_at_head(); 1548 if (res == NULL) { 1549 res = getChunkFromIndexedFreeListHelper(size); 1550 } 1551 _bt.verify_not_unallocated((HeapWord*) res, size); 1552 assert(res == NULL || res->size() == size, "Incorrect block size"); 1553 return res; 1554 } 1555 1556 FreeChunk* 1557 CompactibleFreeListSpace::getChunkFromIndexedFreeListHelper(size_t size, 1558 bool replenish) { 1559 assert_locked(); 1560 FreeChunk* fc = NULL; 1561 if (size < SmallForDictionary) { 1562 assert(_indexedFreeList[size].head() == NULL || 1563 _indexedFreeList[size].surplus() <= 0, 1564 "List for this size should be empty or under populated"); 1565 // Try best fit in exact lists before replenishing the list 1566 if (!bestFitFirst() || (fc = bestFitSmall(size)) == NULL) { 1567 // Replenish list. 1568 // 1569 // Things tried that failed. 1570 // Tried allocating out of the two LinAB's first before 1571 // replenishing lists. 1572 // Tried small linAB of size 256 (size in indexed list) 1573 // and replenishing indexed lists from the small linAB. 1574 // 1575 FreeChunk* newFc = NULL; 1576 const size_t replenish_size = CMSIndexedFreeListReplenish * size; 1577 if (replenish_size < SmallForDictionary) { 1578 // Do not replenish from an underpopulated size. 1579 if (_indexedFreeList[replenish_size].surplus() > 0 && 1580 _indexedFreeList[replenish_size].head() != NULL) { 1581 newFc = _indexedFreeList[replenish_size].get_chunk_at_head(); 1582 } else if (bestFitFirst()) { 1583 newFc = bestFitSmall(replenish_size); 1584 } 1585 } 1586 if (newFc == NULL && replenish_size > size) { 1587 assert(CMSIndexedFreeListReplenish > 1, "ctl pt invariant"); 1588 newFc = getChunkFromIndexedFreeListHelper(replenish_size, false); 1589 } 1590 // Note: The stats update re split-death of block obtained above 1591 // will be recorded below precisely when we know we are going to 1592 // be actually splitting it into more than one pieces below. 1593 if (newFc != NULL) { 1594 if (replenish || CMSReplenishIntermediate) { 1595 // Replenish this list and return one block to caller. 1596 size_t i; 1597 FreeChunk *curFc, *nextFc; 1598 size_t num_blk = newFc->size() / size; 1599 assert(num_blk >= 1, "Smaller than requested?"); 1600 assert(newFc->size() % size == 0, "Should be integral multiple of request"); 1601 if (num_blk > 1) { 1602 // we are sure we will be splitting the block just obtained 1603 // into multiple pieces; record the split-death of the original 1604 splitDeath(replenish_size); 1605 } 1606 // carve up and link blocks 0, ..., num_blk - 2 1607 // The last chunk is not added to the lists but is returned as the 1608 // free chunk. 1609 for (curFc = newFc, nextFc = (FreeChunk*)((HeapWord*)curFc + size), 1610 i = 0; 1611 i < (num_blk - 1); 1612 curFc = nextFc, nextFc = (FreeChunk*)((HeapWord*)nextFc + size), 1613 i++) { 1614 curFc->set_size(size); 1615 // Don't record this as a return in order to try and 1616 // determine the "returns" from a GC. 1617 _bt.verify_not_unallocated((HeapWord*) fc, size); 1618 _indexedFreeList[size].return_chunk_at_tail(curFc, false); 1619 _bt.mark_block((HeapWord*)curFc, size); 1620 split_birth(size); 1621 // Don't record the initial population of the indexed list 1622 // as a split birth. 1623 } 1624 1625 // check that the arithmetic was OK above 1626 assert((HeapWord*)nextFc == (HeapWord*)newFc + num_blk*size, 1627 "inconsistency in carving newFc"); 1628 curFc->set_size(size); 1629 _bt.mark_block((HeapWord*)curFc, size); 1630 split_birth(size); 1631 fc = curFc; 1632 } else { 1633 // Return entire block to caller 1634 fc = newFc; 1635 } 1636 } 1637 } 1638 } else { 1639 // Get a free chunk from the free chunk dictionary to be returned to 1640 // replenish the indexed free list. 1641 fc = getChunkFromDictionaryExact(size); 1642 } 1643 // assert(fc == NULL || fc->is_free(), "Should be returning a free chunk"); 1644 return fc; 1645 } 1646 1647 FreeChunk* 1648 CompactibleFreeListSpace::getChunkFromDictionary(size_t size) { 1649 assert_locked(); 1650 FreeChunk* fc = _dictionary->get_chunk(size, 1651 FreeBlockDictionary<FreeChunk>::atLeast); 1652 if (fc == NULL) { 1653 return NULL; 1654 } 1655 _bt.allocated((HeapWord*)fc, fc->size()); 1656 if (fc->size() >= size + MinChunkSize) { 1657 fc = splitChunkAndReturnRemainder(fc, size); 1658 } 1659 assert(fc->size() >= size, "chunk too small"); 1660 assert(fc->size() < size + MinChunkSize, "chunk too big"); 1661 _bt.verify_single_block((HeapWord*)fc, fc->size()); 1662 return fc; 1663 } 1664 1665 FreeChunk* 1666 CompactibleFreeListSpace::getChunkFromDictionaryExact(size_t size) { 1667 assert_locked(); 1668 FreeChunk* fc = _dictionary->get_chunk(size, 1669 FreeBlockDictionary<FreeChunk>::atLeast); 1670 if (fc == NULL) { 1671 return fc; 1672 } 1673 _bt.allocated((HeapWord*)fc, fc->size()); 1674 if (fc->size() == size) { 1675 _bt.verify_single_block((HeapWord*)fc, size); 1676 return fc; 1677 } 1678 assert(fc->size() > size, "get_chunk() guarantee"); 1679 if (fc->size() < size + MinChunkSize) { 1680 // Return the chunk to the dictionary and go get a bigger one. 1681 returnChunkToDictionary(fc); 1682 fc = _dictionary->get_chunk(size + MinChunkSize, 1683 FreeBlockDictionary<FreeChunk>::atLeast); 1684 if (fc == NULL) { 1685 return NULL; 1686 } 1687 _bt.allocated((HeapWord*)fc, fc->size()); 1688 } 1689 assert(fc->size() >= size + MinChunkSize, "tautology"); 1690 fc = splitChunkAndReturnRemainder(fc, size); 1691 assert(fc->size() == size, "chunk is wrong size"); 1692 _bt.verify_single_block((HeapWord*)fc, size); 1693 return fc; 1694 } 1695 1696 void 1697 CompactibleFreeListSpace::returnChunkToDictionary(FreeChunk* chunk) { 1698 assert_locked(); 1699 1700 size_t size = chunk->size(); 1701 _bt.verify_single_block((HeapWord*)chunk, size); 1702 // adjust _unallocated_block downward, as necessary 1703 _bt.freed((HeapWord*)chunk, size); 1704 _dictionary->return_chunk(chunk); 1705 #ifndef PRODUCT 1706 if (CMSCollector::abstract_state() != CMSCollector::Sweeping) { 1707 TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >* tc = TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >::as_TreeChunk(chunk); 1708 TreeList<FreeChunk, AdaptiveFreeList<FreeChunk> >* tl = tc->list(); 1709 tl->verify_stats(); 1710 } 1711 #endif // PRODUCT 1712 } 1713 1714 void 1715 CompactibleFreeListSpace::returnChunkToFreeList(FreeChunk* fc) { 1716 assert_locked(); 1717 size_t size = fc->size(); 1718 _bt.verify_single_block((HeapWord*) fc, size); 1719 _bt.verify_not_unallocated((HeapWord*) fc, size); 1720 if (_adaptive_freelists) { 1721 _indexedFreeList[size].return_chunk_at_tail(fc); 1722 } else { 1723 _indexedFreeList[size].return_chunk_at_head(fc); 1724 } 1725 #ifndef PRODUCT 1726 if (CMSCollector::abstract_state() != CMSCollector::Sweeping) { 1727 _indexedFreeList[size].verify_stats(); 1728 } 1729 #endif // PRODUCT 1730 } 1731 1732 // Add chunk to end of last block -- if it's the largest 1733 // block -- and update BOT and census data. We would 1734 // of course have preferred to coalesce it with the 1735 // last block, but it's currently less expensive to find the 1736 // largest block than it is to find the last. 1737 void 1738 CompactibleFreeListSpace::addChunkToFreeListsAtEndRecordingStats( 1739 HeapWord* chunk, size_t size) { 1740 // check that the chunk does lie in this space! 1741 assert(chunk != NULL && is_in_reserved(chunk), "Not in this space!"); 1742 // One of the parallel gc task threads may be here 1743 // whilst others are allocating. 1744 Mutex* lock = NULL; 1745 if (ParallelGCThreads != 0) { 1746 lock = &_parDictionaryAllocLock; 1747 } 1748 FreeChunk* ec; 1749 { 1750 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 1751 ec = dictionary()->find_largest_dict(); // get largest block 1752 if (ec != NULL && ec->end() == (uintptr_t*) chunk) { 1753 // It's a coterminal block - we can coalesce. 1754 size_t old_size = ec->size(); 1755 coalDeath(old_size); 1756 removeChunkFromDictionary(ec); 1757 size += old_size; 1758 } else { 1759 ec = (FreeChunk*)chunk; 1760 } 1761 } 1762 ec->set_size(size); 1763 debug_only(ec->mangleFreed(size)); 1764 if (size < SmallForDictionary && ParallelGCThreads != 0) { 1765 lock = _indexedFreeListParLocks[size]; 1766 } 1767 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 1768 addChunkAndRepairOffsetTable((HeapWord*)ec, size, true); 1769 // record the birth under the lock since the recording involves 1770 // manipulation of the list on which the chunk lives and 1771 // if the chunk is allocated and is the last on the list, 1772 // the list can go away. 1773 coalBirth(size); 1774 } 1775 1776 void 1777 CompactibleFreeListSpace::addChunkToFreeLists(HeapWord* chunk, 1778 size_t size) { 1779 // check that the chunk does lie in this space! 1780 assert(chunk != NULL && is_in_reserved(chunk), "Not in this space!"); 1781 assert_locked(); 1782 _bt.verify_single_block(chunk, size); 1783 1784 FreeChunk* fc = (FreeChunk*) chunk; 1785 fc->set_size(size); 1786 debug_only(fc->mangleFreed(size)); 1787 if (size < SmallForDictionary) { 1788 returnChunkToFreeList(fc); 1789 } else { 1790 returnChunkToDictionary(fc); 1791 } 1792 } 1793 1794 void 1795 CompactibleFreeListSpace::addChunkAndRepairOffsetTable(HeapWord* chunk, 1796 size_t size, bool coalesced) { 1797 assert_locked(); 1798 assert(chunk != NULL, "null chunk"); 1799 if (coalesced) { 1800 // repair BOT 1801 _bt.single_block(chunk, size); 1802 } 1803 addChunkToFreeLists(chunk, size); 1804 } 1805 1806 // We _must_ find the purported chunk on our free lists; 1807 // we assert if we don't. 1808 void 1809 CompactibleFreeListSpace::removeFreeChunkFromFreeLists(FreeChunk* fc) { 1810 size_t size = fc->size(); 1811 assert_locked(); 1812 debug_only(verifyFreeLists()); 1813 if (size < SmallForDictionary) { 1814 removeChunkFromIndexedFreeList(fc); 1815 } else { 1816 removeChunkFromDictionary(fc); 1817 } 1818 _bt.verify_single_block((HeapWord*)fc, size); 1819 debug_only(verifyFreeLists()); 1820 } 1821 1822 void 1823 CompactibleFreeListSpace::removeChunkFromDictionary(FreeChunk* fc) { 1824 size_t size = fc->size(); 1825 assert_locked(); 1826 assert(fc != NULL, "null chunk"); 1827 _bt.verify_single_block((HeapWord*)fc, size); 1828 _dictionary->remove_chunk(fc); 1829 // adjust _unallocated_block upward, as necessary 1830 _bt.allocated((HeapWord*)fc, size); 1831 } 1832 1833 void 1834 CompactibleFreeListSpace::removeChunkFromIndexedFreeList(FreeChunk* fc) { 1835 assert_locked(); 1836 size_t size = fc->size(); 1837 _bt.verify_single_block((HeapWord*)fc, size); 1838 NOT_PRODUCT( 1839 if (FLSVerifyIndexTable) { 1840 verifyIndexedFreeList(size); 1841 } 1842 ) 1843 _indexedFreeList[size].remove_chunk(fc); 1844 NOT_PRODUCT( 1845 if (FLSVerifyIndexTable) { 1846 verifyIndexedFreeList(size); 1847 } 1848 ) 1849 } 1850 1851 FreeChunk* CompactibleFreeListSpace::bestFitSmall(size_t numWords) { 1852 /* A hint is the next larger size that has a surplus. 1853 Start search at a size large enough to guarantee that 1854 the excess is >= MIN_CHUNK. */ 1855 size_t start = align_object_size(numWords + MinChunkSize); 1856 if (start < IndexSetSize) { 1857 AdaptiveFreeList<FreeChunk>* it = _indexedFreeList; 1858 size_t hint = _indexedFreeList[start].hint(); 1859 while (hint < IndexSetSize) { 1860 assert(hint % MinObjAlignment == 0, "hint should be aligned"); 1861 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[hint]; 1862 if (fl->surplus() > 0 && fl->head() != NULL) { 1863 // Found a list with surplus, reset original hint 1864 // and split out a free chunk which is returned. 1865 _indexedFreeList[start].set_hint(hint); 1866 FreeChunk* res = getFromListGreater(fl, numWords); 1867 assert(res == NULL || res->is_free(), 1868 "Should be returning a free chunk"); 1869 return res; 1870 } 1871 hint = fl->hint(); /* keep looking */ 1872 } 1873 /* None found. */ 1874 it[start].set_hint(IndexSetSize); 1875 } 1876 return NULL; 1877 } 1878 1879 /* Requires fl->size >= numWords + MinChunkSize */ 1880 FreeChunk* CompactibleFreeListSpace::getFromListGreater(AdaptiveFreeList<FreeChunk>* fl, 1881 size_t numWords) { 1882 FreeChunk *curr = fl->head(); 1883 size_t oldNumWords = curr->size(); 1884 assert(numWords >= MinChunkSize, "Word size is too small"); 1885 assert(curr != NULL, "List is empty"); 1886 assert(oldNumWords >= numWords + MinChunkSize, 1887 "Size of chunks in the list is too small"); 1888 1889 fl->remove_chunk(curr); 1890 // recorded indirectly by splitChunkAndReturnRemainder - 1891 // smallSplit(oldNumWords, numWords); 1892 FreeChunk* new_chunk = splitChunkAndReturnRemainder(curr, numWords); 1893 // Does anything have to be done for the remainder in terms of 1894 // fixing the card table? 1895 assert(new_chunk == NULL || new_chunk->is_free(), 1896 "Should be returning a free chunk"); 1897 return new_chunk; 1898 } 1899 1900 FreeChunk* 1901 CompactibleFreeListSpace::splitChunkAndReturnRemainder(FreeChunk* chunk, 1902 size_t new_size) { 1903 assert_locked(); 1904 size_t size = chunk->size(); 1905 assert(size > new_size, "Split from a smaller block?"); 1906 assert(is_aligned(chunk), "alignment problem"); 1907 assert(size == adjustObjectSize(size), "alignment problem"); 1908 size_t rem_sz = size - new_size; 1909 assert(rem_sz == adjustObjectSize(rem_sz), "alignment problem"); 1910 assert(rem_sz >= MinChunkSize, "Free chunk smaller than minimum"); 1911 FreeChunk* ffc = (FreeChunk*)((HeapWord*)chunk + new_size); 1912 assert(is_aligned(ffc), "alignment problem"); 1913 ffc->set_size(rem_sz); 1914 ffc->link_next(NULL); 1915 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 1916 // Above must occur before BOT is updated below. 1917 // adjust block offset table 1918 OrderAccess::storestore(); 1919 assert(chunk->is_free() && ffc->is_free(), "Error"); 1920 _bt.split_block((HeapWord*)chunk, chunk->size(), new_size); 1921 if (rem_sz < SmallForDictionary) { 1922 bool is_par = (SharedHeap::heap()->n_par_threads() > 0); 1923 if (is_par) _indexedFreeListParLocks[rem_sz]->lock(); 1924 assert(!is_par || 1925 (SharedHeap::heap()->n_par_threads() == 1926 SharedHeap::heap()->workers()->active_workers()), "Mismatch"); 1927 returnChunkToFreeList(ffc); 1928 split(size, rem_sz); 1929 if (is_par) _indexedFreeListParLocks[rem_sz]->unlock(); 1930 } else { 1931 returnChunkToDictionary(ffc); 1932 split(size, rem_sz); 1933 } 1934 chunk->set_size(new_size); 1935 return chunk; 1936 } 1937 1938 void 1939 CompactibleFreeListSpace::sweep_completed() { 1940 // Now that space is probably plentiful, refill linear 1941 // allocation blocks as needed. 1942 refillLinearAllocBlocksIfNeeded(); 1943 } 1944 1945 void 1946 CompactibleFreeListSpace::gc_prologue() { 1947 assert_locked(); 1948 if (PrintFLSStatistics != 0) { 1949 gclog_or_tty->print("Before GC:\n"); 1950 reportFreeListStatistics(); 1951 } 1952 refillLinearAllocBlocksIfNeeded(); 1953 } 1954 1955 void 1956 CompactibleFreeListSpace::gc_epilogue() { 1957 assert_locked(); 1958 if (PrintGCDetails && Verbose && !_adaptive_freelists) { 1959 if (_smallLinearAllocBlock._word_size == 0) 1960 warning("CompactibleFreeListSpace(epilogue):: Linear allocation failure"); 1961 } 1962 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); 1963 _promoInfo.stopTrackingPromotions(); 1964 repairLinearAllocationBlocks(); 1965 // Print Space's stats 1966 if (PrintFLSStatistics != 0) { 1967 gclog_or_tty->print("After GC:\n"); 1968 reportFreeListStatistics(); 1969 } 1970 } 1971 1972 // Iteration support, mostly delegated from a CMS generation 1973 1974 void CompactibleFreeListSpace::save_marks() { 1975 assert(Thread::current()->is_VM_thread(), 1976 "Global variable should only be set when single-threaded"); 1977 // Mark the "end" of the used space at the time of this call; 1978 // note, however, that promoted objects from this point 1979 // on are tracked in the _promoInfo below. 1980 set_saved_mark_word(unallocated_block()); 1981 #ifdef ASSERT 1982 // Check the sanity of save_marks() etc. 1983 MemRegion ur = used_region(); 1984 MemRegion urasm = used_region_at_save_marks(); 1985 assert(ur.contains(urasm), 1986 err_msg(" Error at save_marks(): [" PTR_FORMAT "," PTR_FORMAT ")" 1987 " should contain [" PTR_FORMAT "," PTR_FORMAT ")", 1988 p2i(ur.start()), p2i(ur.end()), p2i(urasm.start()), p2i(urasm.end()))); 1989 #endif 1990 // inform allocator that promotions should be tracked. 1991 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); 1992 _promoInfo.startTrackingPromotions(); 1993 } 1994 1995 bool CompactibleFreeListSpace::no_allocs_since_save_marks() { 1996 assert(_promoInfo.tracking(), "No preceding save_marks?"); 1997 assert(SharedHeap::heap()->n_par_threads() == 0, 1998 "Shouldn't be called if using parallel gc."); 1999 return _promoInfo.noPromotions(); 2000 } 2001 2002 #define CFLS_OOP_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2003 \ 2004 void CompactibleFreeListSpace:: \ 2005 oop_since_save_marks_iterate##nv_suffix(OopClosureType* blk) { \ 2006 assert(SharedHeap::heap()->n_par_threads() == 0, \ 2007 "Shouldn't be called (yet) during parallel part of gc."); \ 2008 _promoInfo.promoted_oops_iterate##nv_suffix(blk); \ 2009 /* \ 2010 * This also restores any displaced headers and removes the elements from \ 2011 * the iteration set as they are processed, so that we have a clean slate \ 2012 * at the end of the iteration. Note, thus, that if new objects are \ 2013 * promoted as a result of the iteration they are iterated over as well. \ 2014 */ \ 2015 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); \ 2016 } 2017 2018 ALL_SINCE_SAVE_MARKS_CLOSURES(CFLS_OOP_SINCE_SAVE_MARKS_DEFN) 2019 2020 bool CompactibleFreeListSpace::linearAllocationWouldFail() const { 2021 return _smallLinearAllocBlock._word_size == 0; 2022 } 2023 2024 void CompactibleFreeListSpace::repairLinearAllocationBlocks() { 2025 // Fix up linear allocation blocks to look like free blocks 2026 repairLinearAllocBlock(&_smallLinearAllocBlock); 2027 } 2028 2029 void CompactibleFreeListSpace::repairLinearAllocBlock(LinearAllocBlock* blk) { 2030 assert_locked(); 2031 if (blk->_ptr != NULL) { 2032 assert(blk->_word_size != 0 && blk->_word_size >= MinChunkSize, 2033 "Minimum block size requirement"); 2034 FreeChunk* fc = (FreeChunk*)(blk->_ptr); 2035 fc->set_size(blk->_word_size); 2036 fc->link_prev(NULL); // mark as free 2037 fc->dontCoalesce(); 2038 assert(fc->is_free(), "just marked it free"); 2039 assert(fc->cantCoalesce(), "just marked it uncoalescable"); 2040 } 2041 } 2042 2043 void CompactibleFreeListSpace::refillLinearAllocBlocksIfNeeded() { 2044 assert_locked(); 2045 if (_smallLinearAllocBlock._ptr == NULL) { 2046 assert(_smallLinearAllocBlock._word_size == 0, 2047 "Size of linAB should be zero if the ptr is NULL"); 2048 // Reset the linAB refill and allocation size limit. 2049 _smallLinearAllocBlock.set(0, 0, 1024*SmallForLinearAlloc, SmallForLinearAlloc); 2050 } 2051 refillLinearAllocBlockIfNeeded(&_smallLinearAllocBlock); 2052 } 2053 2054 void 2055 CompactibleFreeListSpace::refillLinearAllocBlockIfNeeded(LinearAllocBlock* blk) { 2056 assert_locked(); 2057 assert((blk->_ptr == NULL && blk->_word_size == 0) || 2058 (blk->_ptr != NULL && blk->_word_size >= MinChunkSize), 2059 "blk invariant"); 2060 if (blk->_ptr == NULL) { 2061 refillLinearAllocBlock(blk); 2062 } 2063 if (PrintMiscellaneous && Verbose) { 2064 if (blk->_word_size == 0) { 2065 warning("CompactibleFreeListSpace(prologue):: Linear allocation failure"); 2066 } 2067 } 2068 } 2069 2070 void 2071 CompactibleFreeListSpace::refillLinearAllocBlock(LinearAllocBlock* blk) { 2072 assert_locked(); 2073 assert(blk->_word_size == 0 && blk->_ptr == NULL, 2074 "linear allocation block should be empty"); 2075 FreeChunk* fc; 2076 if (blk->_refillSize < SmallForDictionary && 2077 (fc = getChunkFromIndexedFreeList(blk->_refillSize)) != NULL) { 2078 // A linAB's strategy might be to use small sizes to reduce 2079 // fragmentation but still get the benefits of allocation from a 2080 // linAB. 2081 } else { 2082 fc = getChunkFromDictionary(blk->_refillSize); 2083 } 2084 if (fc != NULL) { 2085 blk->_ptr = (HeapWord*)fc; 2086 blk->_word_size = fc->size(); 2087 fc->dontCoalesce(); // to prevent sweeper from sweeping us up 2088 } 2089 } 2090 2091 // Support for concurrent collection policy decisions. 2092 bool CompactibleFreeListSpace::should_concurrent_collect() const { 2093 // In the future we might want to add in fragmentation stats -- 2094 // including erosion of the "mountain" into this decision as well. 2095 return !adaptive_freelists() && linearAllocationWouldFail(); 2096 } 2097 2098 // Support for compaction 2099 2100 void CompactibleFreeListSpace::prepare_for_compaction(CompactPoint* cp) { 2101 SCAN_AND_FORWARD(cp,end,block_is_obj,block_size); 2102 // Prepare_for_compaction() uses the space between live objects 2103 // so that later phase can skip dead space quickly. So verification 2104 // of the free lists doesn't work after. 2105 } 2106 2107 #define obj_size(q) adjustObjectSize(oop(q)->size()) 2108 #define adjust_obj_size(s) adjustObjectSize(s) 2109 2110 void CompactibleFreeListSpace::adjust_pointers() { 2111 // In other versions of adjust_pointers(), a bail out 2112 // based on the amount of live data in the generation 2113 // (i.e., if 0, bail out) may be used. 2114 // Cannot test used() == 0 here because the free lists have already 2115 // been mangled by the compaction. 2116 2117 SCAN_AND_ADJUST_POINTERS(adjust_obj_size); 2118 // See note about verification in prepare_for_compaction(). 2119 } 2120 2121 void CompactibleFreeListSpace::compact() { 2122 SCAN_AND_COMPACT(obj_size); 2123 } 2124 2125 // Fragmentation metric = 1 - [sum of (fbs**2) / (sum of fbs)**2] 2126 // where fbs is free block sizes 2127 double CompactibleFreeListSpace::flsFrag() const { 2128 size_t itabFree = totalSizeInIndexedFreeLists(); 2129 double frag = 0.0; 2130 size_t i; 2131 2132 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2133 double sz = i; 2134 frag += _indexedFreeList[i].count() * (sz * sz); 2135 } 2136 2137 double totFree = itabFree + 2138 _dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())); 2139 if (totFree > 0) { 2140 frag = ((frag + _dictionary->sum_of_squared_block_sizes()) / 2141 (totFree * totFree)); 2142 frag = (double)1.0 - frag; 2143 } else { 2144 assert(frag == 0.0, "Follows from totFree == 0"); 2145 } 2146 return frag; 2147 } 2148 2149 void CompactibleFreeListSpace::beginSweepFLCensus( 2150 float inter_sweep_current, 2151 float inter_sweep_estimate, 2152 float intra_sweep_estimate) { 2153 assert_locked(); 2154 size_t i; 2155 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2156 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[i]; 2157 if (PrintFLSStatistics > 1) { 2158 gclog_or_tty->print("size[" SIZE_FORMAT "] : ", i); 2159 } 2160 fl->compute_desired(inter_sweep_current, inter_sweep_estimate, intra_sweep_estimate); 2161 fl->set_coal_desired((ssize_t)((double)fl->desired() * CMSSmallCoalSurplusPercent)); 2162 fl->set_before_sweep(fl->count()); 2163 fl->set_bfr_surp(fl->surplus()); 2164 } 2165 _dictionary->begin_sweep_dict_census(CMSLargeCoalSurplusPercent, 2166 inter_sweep_current, 2167 inter_sweep_estimate, 2168 intra_sweep_estimate); 2169 } 2170 2171 void CompactibleFreeListSpace::setFLSurplus() { 2172 assert_locked(); 2173 size_t i; 2174 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2175 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2176 fl->set_surplus(fl->count() - 2177 (ssize_t)((double)fl->desired() * CMSSmallSplitSurplusPercent)); 2178 } 2179 } 2180 2181 void CompactibleFreeListSpace::setFLHints() { 2182 assert_locked(); 2183 size_t i; 2184 size_t h = IndexSetSize; 2185 for (i = IndexSetSize - 1; i != 0; i -= IndexSetStride) { 2186 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2187 fl->set_hint(h); 2188 if (fl->surplus() > 0) { 2189 h = i; 2190 } 2191 } 2192 } 2193 2194 void CompactibleFreeListSpace::clearFLCensus() { 2195 assert_locked(); 2196 size_t i; 2197 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2198 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2199 fl->set_prev_sweep(fl->count()); 2200 fl->set_coal_births(0); 2201 fl->set_coal_deaths(0); 2202 fl->set_split_births(0); 2203 fl->set_split_deaths(0); 2204 } 2205 } 2206 2207 void CompactibleFreeListSpace::endSweepFLCensus(size_t sweep_count) { 2208 if (PrintFLSStatistics > 0) { 2209 HeapWord* largestAddr = (HeapWord*) dictionary()->find_largest_dict(); 2210 gclog_or_tty->print_cr("CMS: Large block " PTR_FORMAT, 2211 p2i(largestAddr)); 2212 } 2213 setFLSurplus(); 2214 setFLHints(); 2215 if (PrintGC && PrintFLSCensus > 0) { 2216 printFLCensus(sweep_count); 2217 } 2218 clearFLCensus(); 2219 assert_locked(); 2220 _dictionary->end_sweep_dict_census(CMSLargeSplitSurplusPercent); 2221 } 2222 2223 bool CompactibleFreeListSpace::coalOverPopulated(size_t size) { 2224 if (size < SmallForDictionary) { 2225 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2226 return (fl->coal_desired() < 0) || 2227 ((int)fl->count() > fl->coal_desired()); 2228 } else { 2229 return dictionary()->coal_dict_over_populated(size); 2230 } 2231 } 2232 2233 void CompactibleFreeListSpace::smallCoalBirth(size_t size) { 2234 assert(size < SmallForDictionary, "Size too large for indexed list"); 2235 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2236 fl->increment_coal_births(); 2237 fl->increment_surplus(); 2238 } 2239 2240 void CompactibleFreeListSpace::smallCoalDeath(size_t size) { 2241 assert(size < SmallForDictionary, "Size too large for indexed list"); 2242 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2243 fl->increment_coal_deaths(); 2244 fl->decrement_surplus(); 2245 } 2246 2247 void CompactibleFreeListSpace::coalBirth(size_t size) { 2248 if (size < SmallForDictionary) { 2249 smallCoalBirth(size); 2250 } else { 2251 dictionary()->dict_census_update(size, 2252 false /* split */, 2253 true /* birth */); 2254 } 2255 } 2256 2257 void CompactibleFreeListSpace::coalDeath(size_t size) { 2258 if(size < SmallForDictionary) { 2259 smallCoalDeath(size); 2260 } else { 2261 dictionary()->dict_census_update(size, 2262 false /* split */, 2263 false /* birth */); 2264 } 2265 } 2266 2267 void CompactibleFreeListSpace::smallSplitBirth(size_t size) { 2268 assert(size < SmallForDictionary, "Size too large for indexed list"); 2269 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2270 fl->increment_split_births(); 2271 fl->increment_surplus(); 2272 } 2273 2274 void CompactibleFreeListSpace::smallSplitDeath(size_t size) { 2275 assert(size < SmallForDictionary, "Size too large for indexed list"); 2276 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size]; 2277 fl->increment_split_deaths(); 2278 fl->decrement_surplus(); 2279 } 2280 2281 void CompactibleFreeListSpace::split_birth(size_t size) { 2282 if (size < SmallForDictionary) { 2283 smallSplitBirth(size); 2284 } else { 2285 dictionary()->dict_census_update(size, 2286 true /* split */, 2287 true /* birth */); 2288 } 2289 } 2290 2291 void CompactibleFreeListSpace::splitDeath(size_t size) { 2292 if (size < SmallForDictionary) { 2293 smallSplitDeath(size); 2294 } else { 2295 dictionary()->dict_census_update(size, 2296 true /* split */, 2297 false /* birth */); 2298 } 2299 } 2300 2301 void CompactibleFreeListSpace::split(size_t from, size_t to1) { 2302 size_t to2 = from - to1; 2303 splitDeath(from); 2304 split_birth(to1); 2305 split_birth(to2); 2306 } 2307 2308 void CompactibleFreeListSpace::print() const { 2309 print_on(tty); 2310 } 2311 2312 void CompactibleFreeListSpace::prepare_for_verify() { 2313 assert_locked(); 2314 repairLinearAllocationBlocks(); 2315 // Verify that the SpoolBlocks look like free blocks of 2316 // appropriate sizes... To be done ... 2317 } 2318 2319 class VerifyAllBlksClosure: public BlkClosure { 2320 private: 2321 const CompactibleFreeListSpace* _sp; 2322 const MemRegion _span; 2323 HeapWord* _last_addr; 2324 size_t _last_size; 2325 bool _last_was_obj; 2326 bool _last_was_live; 2327 2328 public: 2329 VerifyAllBlksClosure(const CompactibleFreeListSpace* sp, 2330 MemRegion span) : _sp(sp), _span(span), 2331 _last_addr(NULL), _last_size(0), 2332 _last_was_obj(false), _last_was_live(false) { } 2333 2334 virtual size_t do_blk(HeapWord* addr) { 2335 size_t res; 2336 bool was_obj = false; 2337 bool was_live = false; 2338 if (_sp->block_is_obj(addr)) { 2339 was_obj = true; 2340 oop p = oop(addr); 2341 guarantee(p->is_oop(), "Should be an oop"); 2342 res = _sp->adjustObjectSize(p->size()); 2343 if (_sp->obj_is_alive(addr)) { 2344 was_live = true; 2345 p->verify(); 2346 } 2347 } else { 2348 FreeChunk* fc = (FreeChunk*)addr; 2349 res = fc->size(); 2350 if (FLSVerifyLists && !fc->cantCoalesce()) { 2351 guarantee(_sp->verify_chunk_in_free_list(fc), 2352 "Chunk should be on a free list"); 2353 } 2354 } 2355 if (res == 0) { 2356 gclog_or_tty->print_cr("Livelock: no rank reduction!"); 2357 gclog_or_tty->print_cr( 2358 " Current: addr = " PTR_FORMAT ", size = " SIZE_FORMAT ", obj = %s, live = %s \n" 2359 " Previous: addr = " PTR_FORMAT ", size = " SIZE_FORMAT ", obj = %s, live = %s \n", 2360 p2i(addr), res, was_obj ?"true":"false", was_live ?"true":"false", 2361 p2i(_last_addr), _last_size, _last_was_obj?"true":"false", _last_was_live?"true":"false"); 2362 _sp->print_on(gclog_or_tty); 2363 guarantee(false, "Seppuku!"); 2364 } 2365 _last_addr = addr; 2366 _last_size = res; 2367 _last_was_obj = was_obj; 2368 _last_was_live = was_live; 2369 return res; 2370 } 2371 }; 2372 2373 class VerifyAllOopsClosure: public OopClosure { 2374 private: 2375 const CMSCollector* _collector; 2376 const CompactibleFreeListSpace* _sp; 2377 const MemRegion _span; 2378 const bool _past_remark; 2379 const CMSBitMap* _bit_map; 2380 2381 protected: 2382 void do_oop(void* p, oop obj) { 2383 if (_span.contains(obj)) { // the interior oop points into CMS heap 2384 if (!_span.contains(p)) { // reference from outside CMS heap 2385 // Should be a valid object; the first disjunct below allows 2386 // us to sidestep an assertion in block_is_obj() that insists 2387 // that p be in _sp. Note that several generations (and spaces) 2388 // are spanned by _span (CMS heap) above. 2389 guarantee(!_sp->is_in_reserved(obj) || 2390 _sp->block_is_obj((HeapWord*)obj), 2391 "Should be an object"); 2392 guarantee(obj->is_oop(), "Should be an oop"); 2393 obj->verify(); 2394 if (_past_remark) { 2395 // Remark has been completed, the object should be marked 2396 _bit_map->isMarked((HeapWord*)obj); 2397 } 2398 } else { // reference within CMS heap 2399 if (_past_remark) { 2400 // Remark has been completed -- so the referent should have 2401 // been marked, if referring object is. 2402 if (_bit_map->isMarked(_collector->block_start(p))) { 2403 guarantee(_bit_map->isMarked((HeapWord*)obj), "Marking error?"); 2404 } 2405 } 2406 } 2407 } else if (_sp->is_in_reserved(p)) { 2408 // the reference is from FLS, and points out of FLS 2409 guarantee(obj->is_oop(), "Should be an oop"); 2410 obj->verify(); 2411 } 2412 } 2413 2414 template <class T> void do_oop_work(T* p) { 2415 T heap_oop = oopDesc::load_heap_oop(p); 2416 if (!oopDesc::is_null(heap_oop)) { 2417 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 2418 do_oop(p, obj); 2419 } 2420 } 2421 2422 public: 2423 VerifyAllOopsClosure(const CMSCollector* collector, 2424 const CompactibleFreeListSpace* sp, MemRegion span, 2425 bool past_remark, CMSBitMap* bit_map) : 2426 _collector(collector), _sp(sp), _span(span), 2427 _past_remark(past_remark), _bit_map(bit_map) { } 2428 2429 virtual void do_oop(oop* p) { VerifyAllOopsClosure::do_oop_work(p); } 2430 virtual void do_oop(narrowOop* p) { VerifyAllOopsClosure::do_oop_work(p); } 2431 }; 2432 2433 void CompactibleFreeListSpace::verify() const { 2434 assert_lock_strong(&_freelistLock); 2435 verify_objects_initialized(); 2436 MemRegion span = _collector->_span; 2437 bool past_remark = (_collector->abstract_state() == 2438 CMSCollector::Sweeping); 2439 2440 ResourceMark rm; 2441 HandleMark hm; 2442 2443 // Check integrity of CFL data structures 2444 _promoInfo.verify(); 2445 _dictionary->verify(); 2446 if (FLSVerifyIndexTable) { 2447 verifyIndexedFreeLists(); 2448 } 2449 // Check integrity of all objects and free blocks in space 2450 { 2451 VerifyAllBlksClosure cl(this, span); 2452 ((CompactibleFreeListSpace*)this)->blk_iterate(&cl); // cast off const 2453 } 2454 // Check that all references in the heap to FLS 2455 // are to valid objects in FLS or that references in 2456 // FLS are to valid objects elsewhere in the heap 2457 if (FLSVerifyAllHeapReferences) 2458 { 2459 VerifyAllOopsClosure cl(_collector, this, span, past_remark, 2460 _collector->markBitMap()); 2461 CollectedHeap* ch = Universe::heap(); 2462 2463 // Iterate over all oops in the heap. Uses the _no_header version 2464 // since we are not interested in following the klass pointers. 2465 ch->oop_iterate_no_header(&cl); 2466 } 2467 2468 if (VerifyObjectStartArray) { 2469 // Verify the block offset table 2470 _bt.verify(); 2471 } 2472 } 2473 2474 #ifndef PRODUCT 2475 void CompactibleFreeListSpace::verifyFreeLists() const { 2476 if (FLSVerifyLists) { 2477 _dictionary->verify(); 2478 verifyIndexedFreeLists(); 2479 } else { 2480 if (FLSVerifyDictionary) { 2481 _dictionary->verify(); 2482 } 2483 if (FLSVerifyIndexTable) { 2484 verifyIndexedFreeLists(); 2485 } 2486 } 2487 } 2488 #endif 2489 2490 void CompactibleFreeListSpace::verifyIndexedFreeLists() const { 2491 size_t i = 0; 2492 for (; i < IndexSetStart; i++) { 2493 guarantee(_indexedFreeList[i].head() == NULL, "should be NULL"); 2494 } 2495 for (; i < IndexSetSize; i++) { 2496 verifyIndexedFreeList(i); 2497 } 2498 } 2499 2500 void CompactibleFreeListSpace::verifyIndexedFreeList(size_t size) const { 2501 FreeChunk* fc = _indexedFreeList[size].head(); 2502 FreeChunk* tail = _indexedFreeList[size].tail(); 2503 size_t num = _indexedFreeList[size].count(); 2504 size_t n = 0; 2505 guarantee(((size >= IndexSetStart) && (size % IndexSetStride == 0)) || fc == NULL, 2506 "Slot should have been empty"); 2507 for (; fc != NULL; fc = fc->next(), n++) { 2508 guarantee(fc->size() == size, "Size inconsistency"); 2509 guarantee(fc->is_free(), "!free?"); 2510 guarantee(fc->next() == NULL || fc->next()->prev() == fc, "Broken list"); 2511 guarantee((fc->next() == NULL) == (fc == tail), "Incorrect tail"); 2512 } 2513 guarantee(n == num, "Incorrect count"); 2514 } 2515 2516 #ifndef PRODUCT 2517 void CompactibleFreeListSpace::check_free_list_consistency() const { 2518 assert((TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >::min_size() <= IndexSetSize), 2519 "Some sizes can't be allocated without recourse to" 2520 " linear allocation buffers"); 2521 assert((TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >::min_size()*HeapWordSize == sizeof(TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >)), 2522 "else MIN_TREE_CHUNK_SIZE is wrong"); 2523 assert(IndexSetStart != 0, "IndexSetStart not initialized"); 2524 assert(IndexSetStride != 0, "IndexSetStride not initialized"); 2525 } 2526 #endif 2527 2528 void CompactibleFreeListSpace::printFLCensus(size_t sweep_count) const { 2529 assert_lock_strong(&_freelistLock); 2530 AdaptiveFreeList<FreeChunk> total; 2531 gclog_or_tty->print("end sweep# " SIZE_FORMAT "\n", sweep_count); 2532 AdaptiveFreeList<FreeChunk>::print_labels_on(gclog_or_tty, "size"); 2533 size_t total_free = 0; 2534 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) { 2535 const AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i]; 2536 total_free += fl->count() * fl->size(); 2537 if (i % (40*IndexSetStride) == 0) { 2538 AdaptiveFreeList<FreeChunk>::print_labels_on(gclog_or_tty, "size"); 2539 } 2540 fl->print_on(gclog_or_tty); 2541 total.set_bfr_surp( total.bfr_surp() + fl->bfr_surp() ); 2542 total.set_surplus( total.surplus() + fl->surplus() ); 2543 total.set_desired( total.desired() + fl->desired() ); 2544 total.set_prev_sweep( total.prev_sweep() + fl->prev_sweep() ); 2545 total.set_before_sweep(total.before_sweep() + fl->before_sweep()); 2546 total.set_count( total.count() + fl->count() ); 2547 total.set_coal_births( total.coal_births() + fl->coal_births() ); 2548 total.set_coal_deaths( total.coal_deaths() + fl->coal_deaths() ); 2549 total.set_split_births(total.split_births() + fl->split_births()); 2550 total.set_split_deaths(total.split_deaths() + fl->split_deaths()); 2551 } 2552 total.print_on(gclog_or_tty, "TOTAL"); 2553 gclog_or_tty->print_cr("Total free in indexed lists " 2554 SIZE_FORMAT " words", total_free); 2555 gclog_or_tty->print("growth: %8.5f deficit: %8.5f\n", 2556 (double)(total.split_births()+total.coal_births()-total.split_deaths()-total.coal_deaths())/ 2557 (total.prev_sweep() != 0 ? (double)total.prev_sweep() : 1.0), 2558 (double)(total.desired() - total.count())/(total.desired() != 0 ? (double)total.desired() : 1.0)); 2559 _dictionary->print_dict_census(); 2560 } 2561 2562 /////////////////////////////////////////////////////////////////////////// 2563 // CFLS_LAB 2564 /////////////////////////////////////////////////////////////////////////// 2565 2566 #define VECTOR_257(x) \ 2567 /* 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 */ \ 2568 { 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, \ 2569 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, \ 2570 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, \ 2571 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, \ 2572 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, \ 2573 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, \ 2574 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, \ 2575 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, \ 2576 x } 2577 2578 // Initialize with default setting of CMSParPromoteBlocksToClaim, _not_ 2579 // OldPLABSize, whose static default is different; if overridden at the 2580 // command-line, this will get reinitialized via a call to 2581 // modify_initialization() below. 2582 AdaptiveWeightedAverage CFLS_LAB::_blocks_to_claim[] = 2583 VECTOR_257(AdaptiveWeightedAverage(OldPLABWeight, (float)CMSParPromoteBlocksToClaim)); 2584 size_t CFLS_LAB::_global_num_blocks[] = VECTOR_257(0); 2585 uint CFLS_LAB::_global_num_workers[] = VECTOR_257(0); 2586 2587 CFLS_LAB::CFLS_LAB(CompactibleFreeListSpace* cfls) : 2588 _cfls(cfls) 2589 { 2590 assert(CompactibleFreeListSpace::IndexSetSize == 257, "Modify VECTOR_257() macro above"); 2591 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2592 i < CompactibleFreeListSpace::IndexSetSize; 2593 i += CompactibleFreeListSpace::IndexSetStride) { 2594 _indexedFreeList[i].set_size(i); 2595 _num_blocks[i] = 0; 2596 } 2597 } 2598 2599 static bool _CFLS_LAB_modified = false; 2600 2601 void CFLS_LAB::modify_initialization(size_t n, unsigned wt) { 2602 assert(!_CFLS_LAB_modified, "Call only once"); 2603 _CFLS_LAB_modified = true; 2604 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2605 i < CompactibleFreeListSpace::IndexSetSize; 2606 i += CompactibleFreeListSpace::IndexSetStride) { 2607 _blocks_to_claim[i].modify(n, wt, true /* force */); 2608 } 2609 } 2610 2611 HeapWord* CFLS_LAB::alloc(size_t word_sz) { 2612 FreeChunk* res; 2613 assert(word_sz == _cfls->adjustObjectSize(word_sz), "Error"); 2614 if (word_sz >= CompactibleFreeListSpace::IndexSetSize) { 2615 // This locking manages sync with other large object allocations. 2616 MutexLockerEx x(_cfls->parDictionaryAllocLock(), 2617 Mutex::_no_safepoint_check_flag); 2618 res = _cfls->getChunkFromDictionaryExact(word_sz); 2619 if (res == NULL) return NULL; 2620 } else { 2621 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[word_sz]; 2622 if (fl->count() == 0) { 2623 // Attempt to refill this local free list. 2624 get_from_global_pool(word_sz, fl); 2625 // If it didn't work, give up. 2626 if (fl->count() == 0) return NULL; 2627 } 2628 res = fl->get_chunk_at_head(); 2629 assert(res != NULL, "Why was count non-zero?"); 2630 } 2631 res->markNotFree(); 2632 assert(!res->is_free(), "shouldn't be marked free"); 2633 assert(oop(res)->klass_or_null() == NULL, "should look uninitialized"); 2634 // mangle a just allocated object with a distinct pattern. 2635 debug_only(res->mangleAllocated(word_sz)); 2636 return (HeapWord*)res; 2637 } 2638 2639 // Get a chunk of blocks of the right size and update related 2640 // book-keeping stats 2641 void CFLS_LAB::get_from_global_pool(size_t word_sz, AdaptiveFreeList<FreeChunk>* fl) { 2642 // Get the #blocks we want to claim 2643 size_t n_blks = (size_t)_blocks_to_claim[word_sz].average(); 2644 assert(n_blks > 0, "Error"); 2645 assert(ResizePLAB || n_blks == OldPLABSize, "Error"); 2646 // In some cases, when the application has a phase change, 2647 // there may be a sudden and sharp shift in the object survival 2648 // profile, and updating the counts at the end of a scavenge 2649 // may not be quick enough, giving rise to large scavenge pauses 2650 // during these phase changes. It is beneficial to detect such 2651 // changes on-the-fly during a scavenge and avoid such a phase-change 2652 // pothole. The following code is a heuristic attempt to do that. 2653 // It is protected by a product flag until we have gained 2654 // enough experience with this heuristic and fine-tuned its behavior. 2655 // WARNING: This might increase fragmentation if we overreact to 2656 // small spikes, so some kind of historical smoothing based on 2657 // previous experience with the greater reactivity might be useful. 2658 // Lacking sufficient experience, CMSOldPLABResizeQuicker is disabled by 2659 // default. 2660 if (ResizeOldPLAB && CMSOldPLABResizeQuicker) { 2661 size_t multiple = _num_blocks[word_sz]/(CMSOldPLABToleranceFactor*CMSOldPLABNumRefills*n_blks); 2662 n_blks += CMSOldPLABReactivityFactor*multiple*n_blks; 2663 n_blks = MIN2(n_blks, CMSOldPLABMax); 2664 } 2665 assert(n_blks > 0, "Error"); 2666 _cfls->par_get_chunk_of_blocks(word_sz, n_blks, fl); 2667 // Update stats table entry for this block size 2668 _num_blocks[word_sz] += fl->count(); 2669 } 2670 2671 void CFLS_LAB::compute_desired_plab_size() { 2672 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2673 i < CompactibleFreeListSpace::IndexSetSize; 2674 i += CompactibleFreeListSpace::IndexSetStride) { 2675 assert((_global_num_workers[i] == 0) == (_global_num_blocks[i] == 0), 2676 "Counter inconsistency"); 2677 if (_global_num_workers[i] > 0) { 2678 // Need to smooth wrt historical average 2679 if (ResizeOldPLAB) { 2680 _blocks_to_claim[i].sample( 2681 MAX2((size_t)CMSOldPLABMin, 2682 MIN2((size_t)CMSOldPLABMax, 2683 _global_num_blocks[i]/(_global_num_workers[i]*CMSOldPLABNumRefills)))); 2684 } 2685 // Reset counters for next round 2686 _global_num_workers[i] = 0; 2687 _global_num_blocks[i] = 0; 2688 if (PrintOldPLAB) { 2689 gclog_or_tty->print_cr("[" SIZE_FORMAT "]: " SIZE_FORMAT, 2690 i, (size_t)_blocks_to_claim[i].average()); 2691 } 2692 } 2693 } 2694 } 2695 2696 // If this is changed in the future to allow parallel 2697 // access, one would need to take the FL locks and, 2698 // depending on how it is used, stagger access from 2699 // parallel threads to reduce contention. 2700 void CFLS_LAB::retire(int tid) { 2701 // We run this single threaded with the world stopped; 2702 // so no need for locks and such. 2703 NOT_PRODUCT(Thread* t = Thread::current();) 2704 assert(Thread::current()->is_VM_thread(), "Error"); 2705 for (size_t i = CompactibleFreeListSpace::IndexSetStart; 2706 i < CompactibleFreeListSpace::IndexSetSize; 2707 i += CompactibleFreeListSpace::IndexSetStride) { 2708 assert(_num_blocks[i] >= (size_t)_indexedFreeList[i].count(), 2709 "Can't retire more than what we obtained"); 2710 if (_num_blocks[i] > 0) { 2711 size_t num_retire = _indexedFreeList[i].count(); 2712 assert(_num_blocks[i] > num_retire, "Should have used at least one"); 2713 { 2714 // MutexLockerEx x(_cfls->_indexedFreeListParLocks[i], 2715 // Mutex::_no_safepoint_check_flag); 2716 2717 // Update globals stats for num_blocks used 2718 _global_num_blocks[i] += (_num_blocks[i] - num_retire); 2719 _global_num_workers[i]++; 2720 assert(_global_num_workers[i] <= ParallelGCThreads, "Too big"); 2721 if (num_retire > 0) { 2722 _cfls->_indexedFreeList[i].prepend(&_indexedFreeList[i]); 2723 // Reset this list. 2724 _indexedFreeList[i] = AdaptiveFreeList<FreeChunk>(); 2725 _indexedFreeList[i].set_size(i); 2726 } 2727 } 2728 if (PrintOldPLAB) { 2729 gclog_or_tty->print_cr("%d[" SIZE_FORMAT "]: " SIZE_FORMAT "/" SIZE_FORMAT "/" SIZE_FORMAT, 2730 tid, i, num_retire, _num_blocks[i], (size_t)_blocks_to_claim[i].average()); 2731 } 2732 // Reset stats for next round 2733 _num_blocks[i] = 0; 2734 } 2735 } 2736 } 2737 2738 // Used by par_get_chunk_of_blocks() for the chunks from the 2739 // indexed_free_lists. Looks for a chunk with size that is a multiple 2740 // of "word_sz" and if found, splits it into "word_sz" chunks and add 2741 // to the free list "fl". "n" is the maximum number of chunks to 2742 // be added to "fl". 2743 bool CompactibleFreeListSpace:: par_get_chunk_of_blocks_IFL(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) { 2744 2745 // We'll try all multiples of word_sz in the indexed set, starting with 2746 // word_sz itself and, if CMSSplitIndexedFreeListBlocks, try larger multiples, 2747 // then try getting a big chunk and splitting it. 2748 { 2749 bool found; 2750 int k; 2751 size_t cur_sz; 2752 for (k = 1, cur_sz = k * word_sz, found = false; 2753 (cur_sz < CompactibleFreeListSpace::IndexSetSize) && 2754 (CMSSplitIndexedFreeListBlocks || k <= 1); 2755 k++, cur_sz = k * word_sz) { 2756 AdaptiveFreeList<FreeChunk> fl_for_cur_sz; // Empty. 2757 fl_for_cur_sz.set_size(cur_sz); 2758 { 2759 MutexLockerEx x(_indexedFreeListParLocks[cur_sz], 2760 Mutex::_no_safepoint_check_flag); 2761 AdaptiveFreeList<FreeChunk>* gfl = &_indexedFreeList[cur_sz]; 2762 if (gfl->count() != 0) { 2763 // nn is the number of chunks of size cur_sz that 2764 // we'd need to split k-ways each, in order to create 2765 // "n" chunks of size word_sz each. 2766 const size_t nn = MAX2(n/k, (size_t)1); 2767 gfl->getFirstNChunksFromList(nn, &fl_for_cur_sz); 2768 found = true; 2769 if (k > 1) { 2770 // Update split death stats for the cur_sz-size blocks list: 2771 // we increment the split death count by the number of blocks 2772 // we just took from the cur_sz-size blocks list and which 2773 // we will be splitting below. 2774 ssize_t deaths = gfl->split_deaths() + 2775 fl_for_cur_sz.count(); 2776 gfl->set_split_deaths(deaths); 2777 } 2778 } 2779 } 2780 // Now transfer fl_for_cur_sz to fl. Common case, we hope, is k = 1. 2781 if (found) { 2782 if (k == 1) { 2783 fl->prepend(&fl_for_cur_sz); 2784 } else { 2785 // Divide each block on fl_for_cur_sz up k ways. 2786 FreeChunk* fc; 2787 while ((fc = fl_for_cur_sz.get_chunk_at_head()) != NULL) { 2788 // Must do this in reverse order, so that anybody attempting to 2789 // access the main chunk sees it as a single free block until we 2790 // change it. 2791 size_t fc_size = fc->size(); 2792 assert(fc->is_free(), "Error"); 2793 for (int i = k-1; i >= 0; i--) { 2794 FreeChunk* ffc = (FreeChunk*)((HeapWord*)fc + i * word_sz); 2795 assert((i != 0) || 2796 ((fc == ffc) && ffc->is_free() && 2797 (ffc->size() == k*word_sz) && (fc_size == word_sz)), 2798 "Counting error"); 2799 ffc->set_size(word_sz); 2800 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 2801 ffc->link_next(NULL); 2802 // Above must occur before BOT is updated below. 2803 OrderAccess::storestore(); 2804 // splitting from the right, fc_size == i * word_sz 2805 _bt.mark_block((HeapWord*)ffc, word_sz, true /* reducing */); 2806 fc_size -= word_sz; 2807 assert(fc_size == i*word_sz, "Error"); 2808 _bt.verify_not_unallocated((HeapWord*)ffc, word_sz); 2809 _bt.verify_single_block((HeapWord*)fc, fc_size); 2810 _bt.verify_single_block((HeapWord*)ffc, word_sz); 2811 // Push this on "fl". 2812 fl->return_chunk_at_head(ffc); 2813 } 2814 // TRAP 2815 assert(fl->tail()->next() == NULL, "List invariant."); 2816 } 2817 } 2818 // Update birth stats for this block size. 2819 size_t num = fl->count(); 2820 MutexLockerEx x(_indexedFreeListParLocks[word_sz], 2821 Mutex::_no_safepoint_check_flag); 2822 ssize_t births = _indexedFreeList[word_sz].split_births() + num; 2823 _indexedFreeList[word_sz].set_split_births(births); 2824 return true; 2825 } 2826 } 2827 return found; 2828 } 2829 } 2830 2831 FreeChunk* CompactibleFreeListSpace::get_n_way_chunk_to_split(size_t word_sz, size_t n) { 2832 2833 FreeChunk* fc = NULL; 2834 FreeChunk* rem_fc = NULL; 2835 size_t rem; 2836 { 2837 MutexLockerEx x(parDictionaryAllocLock(), 2838 Mutex::_no_safepoint_check_flag); 2839 while (n > 0) { 2840 fc = dictionary()->get_chunk(MAX2(n * word_sz, _dictionary->min_size()), 2841 FreeBlockDictionary<FreeChunk>::atLeast); 2842 if (fc != NULL) { 2843 break; 2844 } else { 2845 n--; 2846 } 2847 } 2848 if (fc == NULL) return NULL; 2849 // Otherwise, split up that block. 2850 assert((ssize_t)n >= 1, "Control point invariant"); 2851 assert(fc->is_free(), "Error: should be a free block"); 2852 _bt.verify_single_block((HeapWord*)fc, fc->size()); 2853 const size_t nn = fc->size() / word_sz; 2854 n = MIN2(nn, n); 2855 assert((ssize_t)n >= 1, "Control point invariant"); 2856 rem = fc->size() - n * word_sz; 2857 // If there is a remainder, and it's too small, allocate one fewer. 2858 if (rem > 0 && rem < MinChunkSize) { 2859 n--; rem += word_sz; 2860 } 2861 // Note that at this point we may have n == 0. 2862 assert((ssize_t)n >= 0, "Control point invariant"); 2863 2864 // If n is 0, the chunk fc that was found is not large 2865 // enough to leave a viable remainder. We are unable to 2866 // allocate even one block. Return fc to the 2867 // dictionary and return, leaving "fl" empty. 2868 if (n == 0) { 2869 returnChunkToDictionary(fc); 2870 return NULL; 2871 } 2872 2873 _bt.allocated((HeapWord*)fc, fc->size(), true /* reducing */); // update _unallocated_blk 2874 dictionary()->dict_census_update(fc->size(), 2875 true /*split*/, 2876 false /*birth*/); 2877 2878 // First return the remainder, if any. 2879 // Note that we hold the lock until we decide if we're going to give 2880 // back the remainder to the dictionary, since a concurrent allocation 2881 // may otherwise see the heap as empty. (We're willing to take that 2882 // hit if the block is a small block.) 2883 if (rem > 0) { 2884 size_t prefix_size = n * word_sz; 2885 rem_fc = (FreeChunk*)((HeapWord*)fc + prefix_size); 2886 rem_fc->set_size(rem); 2887 rem_fc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 2888 rem_fc->link_next(NULL); 2889 // Above must occur before BOT is updated below. 2890 assert((ssize_t)n > 0 && prefix_size > 0 && rem_fc > fc, "Error"); 2891 OrderAccess::storestore(); 2892 _bt.split_block((HeapWord*)fc, fc->size(), prefix_size); 2893 assert(fc->is_free(), "Error"); 2894 fc->set_size(prefix_size); 2895 if (rem >= IndexSetSize) { 2896 returnChunkToDictionary(rem_fc); 2897 dictionary()->dict_census_update(rem, true /*split*/, true /*birth*/); 2898 rem_fc = NULL; 2899 } 2900 // Otherwise, return it to the small list below. 2901 } 2902 } 2903 if (rem_fc != NULL) { 2904 MutexLockerEx x(_indexedFreeListParLocks[rem], 2905 Mutex::_no_safepoint_check_flag); 2906 _bt.verify_not_unallocated((HeapWord*)rem_fc, rem_fc->size()); 2907 _indexedFreeList[rem].return_chunk_at_head(rem_fc); 2908 smallSplitBirth(rem); 2909 } 2910 assert(n * word_sz == fc->size(), 2911 err_msg("Chunk size " SIZE_FORMAT " is not exactly splittable by " 2912 SIZE_FORMAT " sized chunks of size " SIZE_FORMAT, 2913 fc->size(), n, word_sz)); 2914 return fc; 2915 } 2916 2917 void CompactibleFreeListSpace:: par_get_chunk_of_blocks_dictionary(size_t word_sz, size_t targetted_number_of_chunks, AdaptiveFreeList<FreeChunk>* fl) { 2918 2919 FreeChunk* fc = get_n_way_chunk_to_split(word_sz, targetted_number_of_chunks); 2920 2921 if (fc == NULL) { 2922 return; 2923 } 2924 2925 size_t n = fc->size() / word_sz; 2926 2927 assert((ssize_t)n > 0, "Consistency"); 2928 // Now do the splitting up. 2929 // Must do this in reverse order, so that anybody attempting to 2930 // access the main chunk sees it as a single free block until we 2931 // change it. 2932 size_t fc_size = n * word_sz; 2933 // All but first chunk in this loop 2934 for (ssize_t i = n-1; i > 0; i--) { 2935 FreeChunk* ffc = (FreeChunk*)((HeapWord*)fc + i * word_sz); 2936 ffc->set_size(word_sz); 2937 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads. 2938 ffc->link_next(NULL); 2939 // Above must occur before BOT is updated below. 2940 OrderAccess::storestore(); 2941 // splitting from the right, fc_size == (n - i + 1) * wordsize 2942 _bt.mark_block((HeapWord*)ffc, word_sz, true /* reducing */); 2943 fc_size -= word_sz; 2944 _bt.verify_not_unallocated((HeapWord*)ffc, ffc->size()); 2945 _bt.verify_single_block((HeapWord*)ffc, ffc->size()); 2946 _bt.verify_single_block((HeapWord*)fc, fc_size); 2947 // Push this on "fl". 2948 fl->return_chunk_at_head(ffc); 2949 } 2950 // First chunk 2951 assert(fc->is_free() && fc->size() == n*word_sz, "Error: should still be a free block"); 2952 // The blocks above should show their new sizes before the first block below 2953 fc->set_size(word_sz); 2954 fc->link_prev(NULL); // idempotent wrt free-ness, see assert above 2955 fc->link_next(NULL); 2956 _bt.verify_not_unallocated((HeapWord*)fc, fc->size()); 2957 _bt.verify_single_block((HeapWord*)fc, fc->size()); 2958 fl->return_chunk_at_head(fc); 2959 2960 assert((ssize_t)n > 0 && (ssize_t)n == fl->count(), "Incorrect number of blocks"); 2961 { 2962 // Update the stats for this block size. 2963 MutexLockerEx x(_indexedFreeListParLocks[word_sz], 2964 Mutex::_no_safepoint_check_flag); 2965 const ssize_t births = _indexedFreeList[word_sz].split_births() + n; 2966 _indexedFreeList[word_sz].set_split_births(births); 2967 // ssize_t new_surplus = _indexedFreeList[word_sz].surplus() + n; 2968 // _indexedFreeList[word_sz].set_surplus(new_surplus); 2969 } 2970 2971 // TRAP 2972 assert(fl->tail()->next() == NULL, "List invariant."); 2973 } 2974 2975 void CompactibleFreeListSpace:: par_get_chunk_of_blocks(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) { 2976 assert(fl->count() == 0, "Precondition."); 2977 assert(word_sz < CompactibleFreeListSpace::IndexSetSize, 2978 "Precondition"); 2979 2980 if (par_get_chunk_of_blocks_IFL(word_sz, n, fl)) { 2981 // Got it 2982 return; 2983 } 2984 2985 // Otherwise, we'll split a block from the dictionary. 2986 par_get_chunk_of_blocks_dictionary(word_sz, n, fl); 2987 } 2988 2989 // Set up the space's par_seq_tasks structure for work claiming 2990 // for parallel rescan. See CMSParRemarkTask where this is currently used. 2991 // XXX Need to suitably abstract and generalize this and the next 2992 // method into one. 2993 void 2994 CompactibleFreeListSpace:: 2995 initialize_sequential_subtasks_for_rescan(int n_threads) { 2996 // The "size" of each task is fixed according to rescan_task_size. 2997 assert(n_threads > 0, "Unexpected n_threads argument"); 2998 const size_t task_size = rescan_task_size(); 2999 size_t n_tasks = (used_region().word_size() + task_size - 1)/task_size; 3000 assert((n_tasks == 0) == used_region().is_empty(), "n_tasks incorrect"); 3001 assert(n_tasks == 0 || 3002 ((used_region().start() + (n_tasks - 1)*task_size < used_region().end()) && 3003 (used_region().start() + n_tasks*task_size >= used_region().end())), 3004 "n_tasks calculation incorrect"); 3005 SequentialSubTasksDone* pst = conc_par_seq_tasks(); 3006 assert(!pst->valid(), "Clobbering existing data?"); 3007 // Sets the condition for completion of the subtask (how many threads 3008 // need to finish in order to be done). 3009 pst->set_n_threads(n_threads); 3010 pst->set_n_tasks((int)n_tasks); 3011 } 3012 3013 // Set up the space's par_seq_tasks structure for work claiming 3014 // for parallel concurrent marking. See CMSConcMarkTask where this is currently used. 3015 void 3016 CompactibleFreeListSpace:: 3017 initialize_sequential_subtasks_for_marking(int n_threads, 3018 HeapWord* low) { 3019 // The "size" of each task is fixed according to rescan_task_size. 3020 assert(n_threads > 0, "Unexpected n_threads argument"); 3021 const size_t task_size = marking_task_size(); 3022 assert(task_size > CardTableModRefBS::card_size_in_words && 3023 (task_size % CardTableModRefBS::card_size_in_words == 0), 3024 "Otherwise arithmetic below would be incorrect"); 3025 MemRegion span = _gen->reserved(); 3026 if (low != NULL) { 3027 if (span.contains(low)) { 3028 // Align low down to a card boundary so that 3029 // we can use block_offset_careful() on span boundaries. 3030 HeapWord* aligned_low = (HeapWord*)align_size_down((uintptr_t)low, 3031 CardTableModRefBS::card_size); 3032 // Clip span prefix at aligned_low 3033 span = span.intersection(MemRegion(aligned_low, span.end())); 3034 } else if (low > span.end()) { 3035 span = MemRegion(low, low); // Null region 3036 } // else use entire span 3037 } 3038 assert(span.is_empty() || 3039 ((uintptr_t)span.start() % CardTableModRefBS::card_size == 0), 3040 "span should start at a card boundary"); 3041 size_t n_tasks = (span.word_size() + task_size - 1)/task_size; 3042 assert((n_tasks == 0) == span.is_empty(), "Inconsistency"); 3043 assert(n_tasks == 0 || 3044 ((span.start() + (n_tasks - 1)*task_size < span.end()) && 3045 (span.start() + n_tasks*task_size >= span.end())), 3046 "n_tasks calculation incorrect"); 3047 SequentialSubTasksDone* pst = conc_par_seq_tasks(); 3048 assert(!pst->valid(), "Clobbering existing data?"); 3049 // Sets the condition for completion of the subtask (how many threads 3050 // need to finish in order to be done). 3051 pst->set_n_threads(n_threads); 3052 pst->set_n_tasks((int)n_tasks); 3053 }