1 /* 2 * Copyright (c) 2005, 2020, 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 "aot/aotLoader.hpp" 27 #include "classfile/classLoaderDataGraph.hpp" 28 #include "classfile/javaClasses.inline.hpp" 29 #include "classfile/stringTable.hpp" 30 #include "classfile/symbolTable.hpp" 31 #include "classfile/systemDictionary.hpp" 32 #include "code/codeCache.hpp" 33 #include "gc/parallel/parallelArguments.hpp" 34 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 35 #include "gc/parallel/parMarkBitMap.inline.hpp" 36 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 37 #include "gc/parallel/psCompactionManager.inline.hpp" 38 #include "gc/parallel/psOldGen.hpp" 39 #include "gc/parallel/psParallelCompact.inline.hpp" 40 #include "gc/parallel/psPromotionManager.inline.hpp" 41 #include "gc/parallel/psRootType.hpp" 42 #include "gc/parallel/psScavenge.hpp" 43 #include "gc/parallel/psYoungGen.hpp" 44 #include "gc/shared/gcCause.hpp" 45 #include "gc/shared/gcHeapSummary.hpp" 46 #include "gc/shared/gcId.hpp" 47 #include "gc/shared/gcLocker.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/isGCActiveMark.hpp" 52 #include "gc/shared/oopStorage.inline.hpp" 53 #include "gc/shared/oopStorageSet.inline.hpp" 54 #include "gc/shared/oopStorageSetParState.inline.hpp" 55 #include "gc/shared/referencePolicy.hpp" 56 #include "gc/shared/referenceProcessor.hpp" 57 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 58 #include "gc/shared/spaceDecorator.inline.hpp" 59 #include "gc/shared/taskTerminator.hpp" 60 #include "gc/shared/weakProcessor.hpp" 61 #include "gc/shared/workerPolicy.hpp" 62 #include "gc/shared/workgroup.hpp" 63 #include "logging/log.hpp" 64 #include "memory/iterator.inline.hpp" 65 #include "memory/resourceArea.hpp" 66 #include "memory/universe.hpp" 67 #include "oops/access.inline.hpp" 68 #include "oops/instanceClassLoaderKlass.inline.hpp" 69 #include "oops/instanceKlass.inline.hpp" 70 #include "oops/instanceMirrorKlass.inline.hpp" 71 #include "oops/methodData.hpp" 72 #include "oops/objArrayKlass.inline.hpp" 73 #include "oops/oop.inline.hpp" 74 #include "runtime/atomic.hpp" 75 #include "runtime/handles.inline.hpp" 76 #include "runtime/safepoint.hpp" 77 #include "runtime/vmThread.hpp" 78 #include "services/management.hpp" 79 #include "services/memTracker.hpp" 80 #include "services/memoryService.hpp" 81 #include "utilities/align.hpp" 82 #include "utilities/debug.hpp" 83 #include "utilities/events.hpp" 84 #include "utilities/formatBuffer.hpp" 85 #include "utilities/macros.hpp" 86 #include "utilities/stack.inline.hpp" 87 #if INCLUDE_JVMCI 88 #include "jvmci/jvmci.hpp" 89 #endif 90 91 #include <math.h> 92 93 // All sizes are in HeapWords. 94 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 95 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 96 const size_t ParallelCompactData::RegionSizeBytes = 97 RegionSize << LogHeapWordSize; 98 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 99 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 100 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 101 102 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 103 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 104 const size_t ParallelCompactData::BlockSizeBytes = 105 BlockSize << LogHeapWordSize; 106 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 107 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 108 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 109 110 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 111 const size_t ParallelCompactData::Log2BlocksPerRegion = 112 Log2RegionSize - Log2BlockSize; 113 114 const ParallelCompactData::RegionData::region_sz_t 115 ParallelCompactData::RegionData::dc_shift = 27; 116 117 const ParallelCompactData::RegionData::region_sz_t 118 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 119 120 const ParallelCompactData::RegionData::region_sz_t 121 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 122 123 const ParallelCompactData::RegionData::region_sz_t 124 ParallelCompactData::RegionData::los_mask = ~dc_mask; 125 126 const ParallelCompactData::RegionData::region_sz_t 127 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 128 129 const ParallelCompactData::RegionData::region_sz_t 130 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 131 132 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 133 134 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 135 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 136 137 double PSParallelCompact::_dwl_mean; 138 double PSParallelCompact::_dwl_std_dev; 139 double PSParallelCompact::_dwl_first_term; 140 double PSParallelCompact::_dwl_adjustment; 141 #ifdef ASSERT 142 bool PSParallelCompact::_dwl_initialized = false; 143 #endif // #ifdef ASSERT 144 145 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 146 HeapWord* destination) 147 { 148 assert(src_region_idx != 0, "invalid src_region_idx"); 149 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 150 assert(destination != NULL, "invalid destination argument"); 151 152 _src_region_idx = src_region_idx; 153 _partial_obj_size = partial_obj_size; 154 _destination = destination; 155 156 // These fields may not be updated below, so make sure they're clear. 157 assert(_dest_region_addr == NULL, "should have been cleared"); 158 assert(_first_src_addr == NULL, "should have been cleared"); 159 160 // Determine the number of destination regions for the partial object. 161 HeapWord* const last_word = destination + partial_obj_size - 1; 162 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 163 HeapWord* const beg_region_addr = sd.region_align_down(destination); 164 HeapWord* const end_region_addr = sd.region_align_down(last_word); 165 166 if (beg_region_addr == end_region_addr) { 167 // One destination region. 168 _destination_count = 1; 169 if (end_region_addr == destination) { 170 // The destination falls on a region boundary, thus the first word of the 171 // partial object will be the first word copied to the destination region. 172 _dest_region_addr = end_region_addr; 173 _first_src_addr = sd.region_to_addr(src_region_idx); 174 } 175 } else { 176 // Two destination regions. When copied, the partial object will cross a 177 // destination region boundary, so a word somewhere within the partial 178 // object will be the first word copied to the second destination region. 179 _destination_count = 2; 180 _dest_region_addr = end_region_addr; 181 const size_t ofs = pointer_delta(end_region_addr, destination); 182 assert(ofs < _partial_obj_size, "sanity"); 183 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 184 } 185 } 186 187 void SplitInfo::clear() 188 { 189 _src_region_idx = 0; 190 _partial_obj_size = 0; 191 _destination = NULL; 192 _destination_count = 0; 193 _dest_region_addr = NULL; 194 _first_src_addr = NULL; 195 assert(!is_valid(), "sanity"); 196 } 197 198 #ifdef ASSERT 199 void SplitInfo::verify_clear() 200 { 201 assert(_src_region_idx == 0, "not clear"); 202 assert(_partial_obj_size == 0, "not clear"); 203 assert(_destination == NULL, "not clear"); 204 assert(_destination_count == 0, "not clear"); 205 assert(_dest_region_addr == NULL, "not clear"); 206 assert(_first_src_addr == NULL, "not clear"); 207 } 208 #endif // #ifdef ASSERT 209 210 211 void PSParallelCompact::print_on_error(outputStream* st) { 212 _mark_bitmap.print_on_error(st); 213 } 214 215 #ifndef PRODUCT 216 const char* PSParallelCompact::space_names[] = { 217 "old ", "eden", "from", "to " 218 }; 219 220 void PSParallelCompact::print_region_ranges() { 221 if (!log_develop_is_enabled(Trace, gc, compaction)) { 222 return; 223 } 224 Log(gc, compaction) log; 225 ResourceMark rm; 226 LogStream ls(log.trace()); 227 Universe::print_on(&ls); 228 log.trace("space bottom top end new_top"); 229 log.trace("------ ---------- ---------- ---------- ----------"); 230 231 for (unsigned int id = 0; id < last_space_id; ++id) { 232 const MutableSpace* space = _space_info[id].space(); 233 log.trace("%u %s " 234 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 235 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 236 id, space_names[id], 237 summary_data().addr_to_region_idx(space->bottom()), 238 summary_data().addr_to_region_idx(space->top()), 239 summary_data().addr_to_region_idx(space->end()), 240 summary_data().addr_to_region_idx(_space_info[id].new_top())); 241 } 242 } 243 244 void 245 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 246 { 247 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 248 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 249 250 ParallelCompactData& sd = PSParallelCompact::summary_data(); 251 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 252 log_develop_trace(gc, compaction)( 253 REGION_IDX_FORMAT " " PTR_FORMAT " " 254 REGION_IDX_FORMAT " " PTR_FORMAT " " 255 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 256 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 257 i, p2i(c->data_location()), dci, p2i(c->destination()), 258 c->partial_obj_size(), c->live_obj_size(), 259 c->data_size(), c->source_region(), c->destination_count()); 260 261 #undef REGION_IDX_FORMAT 262 #undef REGION_DATA_FORMAT 263 } 264 265 void 266 print_generic_summary_data(ParallelCompactData& summary_data, 267 HeapWord* const beg_addr, 268 HeapWord* const end_addr) 269 { 270 size_t total_words = 0; 271 size_t i = summary_data.addr_to_region_idx(beg_addr); 272 const size_t last = summary_data.addr_to_region_idx(end_addr); 273 HeapWord* pdest = 0; 274 275 while (i < last) { 276 ParallelCompactData::RegionData* c = summary_data.region(i); 277 if (c->data_size() != 0 || c->destination() != pdest) { 278 print_generic_summary_region(i, c); 279 total_words += c->data_size(); 280 pdest = c->destination(); 281 } 282 ++i; 283 } 284 285 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 286 } 287 288 void 289 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 290 HeapWord* const beg_addr, 291 HeapWord* const end_addr) { 292 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 293 } 294 295 void 296 print_generic_summary_data(ParallelCompactData& summary_data, 297 SpaceInfo* space_info) 298 { 299 if (!log_develop_is_enabled(Trace, gc, compaction)) { 300 return; 301 } 302 303 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 304 const MutableSpace* space = space_info[id].space(); 305 print_generic_summary_data(summary_data, space->bottom(), 306 MAX2(space->top(), space_info[id].new_top())); 307 } 308 } 309 310 void 311 print_initial_summary_data(ParallelCompactData& summary_data, 312 const MutableSpace* space) { 313 if (space->top() == space->bottom()) { 314 return; 315 } 316 317 const size_t region_size = ParallelCompactData::RegionSize; 318 typedef ParallelCompactData::RegionData RegionData; 319 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 320 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 321 const RegionData* c = summary_data.region(end_region - 1); 322 HeapWord* end_addr = c->destination() + c->data_size(); 323 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 324 325 // Print (and count) the full regions at the beginning of the space. 326 size_t full_region_count = 0; 327 size_t i = summary_data.addr_to_region_idx(space->bottom()); 328 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 329 ParallelCompactData::RegionData* c = summary_data.region(i); 330 log_develop_trace(gc, compaction)( 331 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 332 i, p2i(c->destination()), 333 c->partial_obj_size(), c->live_obj_size(), 334 c->data_size(), c->source_region(), c->destination_count()); 335 ++full_region_count; 336 ++i; 337 } 338 339 size_t live_to_right = live_in_space - full_region_count * region_size; 340 341 double max_reclaimed_ratio = 0.0; 342 size_t max_reclaimed_ratio_region = 0; 343 size_t max_dead_to_right = 0; 344 size_t max_live_to_right = 0; 345 346 // Print the 'reclaimed ratio' for regions while there is something live in 347 // the region or to the right of it. The remaining regions are empty (and 348 // uninteresting), and computing the ratio will result in division by 0. 349 while (i < end_region && live_to_right > 0) { 350 c = summary_data.region(i); 351 HeapWord* const region_addr = summary_data.region_to_addr(i); 352 const size_t used_to_right = pointer_delta(space->top(), region_addr); 353 const size_t dead_to_right = used_to_right - live_to_right; 354 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 355 356 if (reclaimed_ratio > max_reclaimed_ratio) { 357 max_reclaimed_ratio = reclaimed_ratio; 358 max_reclaimed_ratio_region = i; 359 max_dead_to_right = dead_to_right; 360 max_live_to_right = live_to_right; 361 } 362 363 ParallelCompactData::RegionData* c = summary_data.region(i); 364 log_develop_trace(gc, compaction)( 365 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 366 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 367 i, p2i(c->destination()), 368 c->partial_obj_size(), c->live_obj_size(), 369 c->data_size(), c->source_region(), c->destination_count(), 370 reclaimed_ratio, dead_to_right, live_to_right); 371 372 373 live_to_right -= c->data_size(); 374 ++i; 375 } 376 377 // Any remaining regions are empty. Print one more if there is one. 378 if (i < end_region) { 379 ParallelCompactData::RegionData* c = summary_data.region(i); 380 log_develop_trace(gc, compaction)( 381 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 382 i, p2i(c->destination()), 383 c->partial_obj_size(), c->live_obj_size(), 384 c->data_size(), c->source_region(), c->destination_count()); 385 } 386 387 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 388 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 389 } 390 391 void 392 print_initial_summary_data(ParallelCompactData& summary_data, 393 SpaceInfo* space_info) { 394 if (!log_develop_is_enabled(Trace, gc, compaction)) { 395 return; 396 } 397 398 unsigned int id = PSParallelCompact::old_space_id; 399 const MutableSpace* space; 400 do { 401 space = space_info[id].space(); 402 print_initial_summary_data(summary_data, space); 403 } while (++id < PSParallelCompact::eden_space_id); 404 405 do { 406 space = space_info[id].space(); 407 print_generic_summary_data(summary_data, space->bottom(), space->top()); 408 } while (++id < PSParallelCompact::last_space_id); 409 } 410 #endif // #ifndef PRODUCT 411 412 #ifdef ASSERT 413 size_t add_obj_count; 414 size_t add_obj_size; 415 size_t mark_bitmap_count; 416 size_t mark_bitmap_size; 417 #endif // #ifdef ASSERT 418 419 ParallelCompactData::ParallelCompactData() : 420 _region_start(NULL), 421 DEBUG_ONLY(_region_end(NULL) COMMA) 422 _region_vspace(NULL), 423 _reserved_byte_size(0), 424 _region_data(NULL), 425 _region_count(0), 426 _block_vspace(NULL), 427 _block_data(NULL), 428 _block_count(0) {} 429 430 bool ParallelCompactData::initialize(MemRegion covered_region) 431 { 432 _region_start = covered_region.start(); 433 const size_t region_size = covered_region.word_size(); 434 DEBUG_ONLY(_region_end = _region_start + region_size;) 435 436 assert(region_align_down(_region_start) == _region_start, 437 "region start not aligned"); 438 assert((region_size & RegionSizeOffsetMask) == 0, 439 "region size not a multiple of RegionSize"); 440 441 bool result = initialize_region_data(region_size) && initialize_block_data(); 442 return result; 443 } 444 445 PSVirtualSpace* 446 ParallelCompactData::create_vspace(size_t count, size_t element_size) 447 { 448 const size_t raw_bytes = count * element_size; 449 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 450 const size_t granularity = os::vm_allocation_granularity(); 451 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 452 453 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 454 MAX2(page_sz, granularity); 455 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 456 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 457 rs.size()); 458 459 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 460 461 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 462 if (vspace != 0) { 463 if (vspace->expand_by(_reserved_byte_size)) { 464 return vspace; 465 } 466 delete vspace; 467 // Release memory reserved in the space. 468 rs.release(); 469 } 470 471 return 0; 472 } 473 474 bool ParallelCompactData::initialize_region_data(size_t region_size) 475 { 476 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 477 _region_vspace = create_vspace(count, sizeof(RegionData)); 478 if (_region_vspace != 0) { 479 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 480 _region_count = count; 481 return true; 482 } 483 return false; 484 } 485 486 bool ParallelCompactData::initialize_block_data() 487 { 488 assert(_region_count != 0, "region data must be initialized first"); 489 const size_t count = _region_count << Log2BlocksPerRegion; 490 _block_vspace = create_vspace(count, sizeof(BlockData)); 491 if (_block_vspace != 0) { 492 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 493 _block_count = count; 494 return true; 495 } 496 return false; 497 } 498 499 void ParallelCompactData::clear() 500 { 501 memset(_region_data, 0, _region_vspace->committed_size()); 502 memset(_block_data, 0, _block_vspace->committed_size()); 503 } 504 505 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 506 assert(beg_region <= _region_count, "beg_region out of range"); 507 assert(end_region <= _region_count, "end_region out of range"); 508 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 509 510 const size_t region_cnt = end_region - beg_region; 511 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 512 513 const size_t beg_block = beg_region * BlocksPerRegion; 514 const size_t block_cnt = region_cnt * BlocksPerRegion; 515 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 516 } 517 518 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 519 { 520 const RegionData* cur_cp = region(region_idx); 521 const RegionData* const end_cp = region(region_count() - 1); 522 523 HeapWord* result = region_to_addr(region_idx); 524 if (cur_cp < end_cp) { 525 do { 526 result += cur_cp->partial_obj_size(); 527 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 528 } 529 return result; 530 } 531 532 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 533 { 534 const size_t obj_ofs = pointer_delta(addr, _region_start); 535 const size_t beg_region = obj_ofs >> Log2RegionSize; 536 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 537 538 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 539 DEBUG_ONLY(Atomic::add(&add_obj_size, len);) 540 541 if (beg_region == end_region) { 542 // All in one region. 543 _region_data[beg_region].add_live_obj(len); 544 return; 545 } 546 547 // First region. 548 const size_t beg_ofs = region_offset(addr); 549 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 550 551 Klass* klass = ((oop)addr)->klass(); 552 // Middle regions--completely spanned by this object. 553 for (size_t region = beg_region + 1; region < end_region; ++region) { 554 _region_data[region].set_partial_obj_size(RegionSize); 555 _region_data[region].set_partial_obj_addr(addr); 556 } 557 558 // Last region. 559 const size_t end_ofs = region_offset(addr + len - 1); 560 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 561 _region_data[end_region].set_partial_obj_addr(addr); 562 } 563 564 void 565 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 566 { 567 assert(region_offset(beg) == 0, "not RegionSize aligned"); 568 assert(region_offset(end) == 0, "not RegionSize aligned"); 569 570 size_t cur_region = addr_to_region_idx(beg); 571 const size_t end_region = addr_to_region_idx(end); 572 HeapWord* addr = beg; 573 while (cur_region < end_region) { 574 _region_data[cur_region].set_destination(addr); 575 _region_data[cur_region].set_destination_count(0); 576 _region_data[cur_region].set_source_region(cur_region); 577 _region_data[cur_region].set_data_location(addr); 578 579 // Update live_obj_size so the region appears completely full. 580 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 581 _region_data[cur_region].set_live_obj_size(live_size); 582 583 ++cur_region; 584 addr += RegionSize; 585 } 586 } 587 588 // Find the point at which a space can be split and, if necessary, record the 589 // split point. 590 // 591 // If the current src region (which overflowed the destination space) doesn't 592 // have a partial object, the split point is at the beginning of the current src 593 // region (an "easy" split, no extra bookkeeping required). 594 // 595 // If the current src region has a partial object, the split point is in the 596 // region where that partial object starts (call it the split_region). If 597 // split_region has a partial object, then the split point is just after that 598 // partial object (a "hard" split where we have to record the split data and 599 // zero the partial_obj_size field). With a "hard" split, we know that the 600 // partial_obj ends within split_region because the partial object that caused 601 // the overflow starts in split_region. If split_region doesn't have a partial 602 // obj, then the split is at the beginning of split_region (another "easy" 603 // split). 604 HeapWord* 605 ParallelCompactData::summarize_split_space(size_t src_region, 606 SplitInfo& split_info, 607 HeapWord* destination, 608 HeapWord* target_end, 609 HeapWord** target_next) 610 { 611 assert(destination <= target_end, "sanity"); 612 assert(destination + _region_data[src_region].data_size() > target_end, 613 "region should not fit into target space"); 614 assert(is_region_aligned(target_end), "sanity"); 615 616 size_t split_region = src_region; 617 HeapWord* split_destination = destination; 618 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 619 620 if (destination + partial_obj_size > target_end) { 621 // The split point is just after the partial object (if any) in the 622 // src_region that contains the start of the object that overflowed the 623 // destination space. 624 // 625 // Find the start of the "overflow" object and set split_region to the 626 // region containing it. 627 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 628 split_region = addr_to_region_idx(overflow_obj); 629 630 // Clear the source_region field of all destination regions whose first word 631 // came from data after the split point (a non-null source_region field 632 // implies a region must be filled). 633 // 634 // An alternative to the simple loop below: clear during post_compact(), 635 // which uses memcpy instead of individual stores, and is easy to 636 // parallelize. (The downside is that it clears the entire RegionData 637 // object as opposed to just one field.) 638 // 639 // post_compact() would have to clear the summary data up to the highest 640 // address that was written during the summary phase, which would be 641 // 642 // max(top, max(new_top, clear_top)) 643 // 644 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 645 // to target_end. 646 const RegionData* const sr = region(split_region); 647 const size_t beg_idx = 648 addr_to_region_idx(region_align_up(sr->destination() + 649 sr->partial_obj_size())); 650 const size_t end_idx = addr_to_region_idx(target_end); 651 652 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 653 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 654 _region_data[idx].set_source_region(0); 655 } 656 657 // Set split_destination and partial_obj_size to reflect the split region. 658 split_destination = sr->destination(); 659 partial_obj_size = sr->partial_obj_size(); 660 } 661 662 // The split is recorded only if a partial object extends onto the region. 663 if (partial_obj_size != 0) { 664 _region_data[split_region].set_partial_obj_size(0); 665 split_info.record(split_region, partial_obj_size, split_destination); 666 } 667 668 // Setup the continuation addresses. 669 *target_next = split_destination + partial_obj_size; 670 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 671 672 if (log_develop_is_enabled(Trace, gc, compaction)) { 673 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 674 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 675 split_type, p2i(source_next), split_region, partial_obj_size); 676 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 677 split_type, p2i(split_destination), 678 addr_to_region_idx(split_destination), 679 p2i(*target_next)); 680 681 if (partial_obj_size != 0) { 682 HeapWord* const po_beg = split_info.destination(); 683 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 684 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 685 split_type, 686 p2i(po_beg), addr_to_region_idx(po_beg), 687 p2i(po_end), addr_to_region_idx(po_end)); 688 } 689 } 690 691 return source_next; 692 } 693 694 bool ParallelCompactData::summarize(SplitInfo& split_info, 695 HeapWord* source_beg, HeapWord* source_end, 696 HeapWord** source_next, 697 HeapWord* target_beg, HeapWord* target_end, 698 HeapWord** target_next) 699 { 700 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 701 log_develop_trace(gc, compaction)( 702 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 703 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 704 p2i(source_beg), p2i(source_end), p2i(source_next_val), 705 p2i(target_beg), p2i(target_end), p2i(*target_next)); 706 707 size_t cur_region = addr_to_region_idx(source_beg); 708 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 709 710 HeapWord *dest_addr = target_beg; 711 while (cur_region < end_region) { 712 // The destination must be set even if the region has no data. 713 _region_data[cur_region].set_destination(dest_addr); 714 715 size_t words = _region_data[cur_region].data_size(); 716 if (words > 0) { 717 // If cur_region does not fit entirely into the target space, find a point 718 // at which the source space can be 'split' so that part is copied to the 719 // target space and the rest is copied elsewhere. 720 if (dest_addr + words > target_end) { 721 assert(source_next != NULL, "source_next is NULL when splitting"); 722 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 723 target_end, target_next); 724 return false; 725 } 726 727 // Compute the destination_count for cur_region, and if necessary, update 728 // source_region for a destination region. The source_region field is 729 // updated if cur_region is the first (left-most) region to be copied to a 730 // destination region. 731 // 732 // The destination_count calculation is a bit subtle. A region that has 733 // data that compacts into itself does not count itself as a destination. 734 // This maintains the invariant that a zero count means the region is 735 // available and can be claimed and then filled. 736 uint destination_count = 0; 737 if (split_info.is_split(cur_region)) { 738 // The current region has been split: the partial object will be copied 739 // to one destination space and the remaining data will be copied to 740 // another destination space. Adjust the initial destination_count and, 741 // if necessary, set the source_region field if the partial object will 742 // cross a destination region boundary. 743 destination_count = split_info.destination_count(); 744 if (destination_count == 2) { 745 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 746 _region_data[dest_idx].set_source_region(cur_region); 747 } 748 } 749 750 HeapWord* const last_addr = dest_addr + words - 1; 751 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 752 const size_t dest_region_2 = addr_to_region_idx(last_addr); 753 754 // Initially assume that the destination regions will be the same and 755 // adjust the value below if necessary. Under this assumption, if 756 // cur_region == dest_region_2, then cur_region will be compacted 757 // completely into itself. 758 destination_count += cur_region == dest_region_2 ? 0 : 1; 759 if (dest_region_1 != dest_region_2) { 760 // Destination regions differ; adjust destination_count. 761 destination_count += 1; 762 // Data from cur_region will be copied to the start of dest_region_2. 763 _region_data[dest_region_2].set_source_region(cur_region); 764 } else if (region_offset(dest_addr) == 0) { 765 // Data from cur_region will be copied to the start of the destination 766 // region. 767 _region_data[dest_region_1].set_source_region(cur_region); 768 } 769 770 _region_data[cur_region].set_destination_count(destination_count); 771 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 772 dest_addr += words; 773 } 774 775 ++cur_region; 776 } 777 778 *target_next = dest_addr; 779 return true; 780 } 781 782 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 783 assert(addr != NULL, "Should detect NULL oop earlier"); 784 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 785 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 786 787 // Region covering the object. 788 RegionData* const region_ptr = addr_to_region_ptr(addr); 789 HeapWord* result = region_ptr->destination(); 790 791 // If the entire Region is live, the new location is region->destination + the 792 // offset of the object within in the Region. 793 794 // Run some performance tests to determine if this special case pays off. It 795 // is worth it for pointers into the dense prefix. If the optimization to 796 // avoid pointer updates in regions that only point to the dense prefix is 797 // ever implemented, this should be revisited. 798 if (region_ptr->data_size() == RegionSize) { 799 result += region_offset(addr); 800 return result; 801 } 802 803 // Otherwise, the new location is region->destination + block offset + the 804 // number of live words in the Block that are (a) to the left of addr and (b) 805 // due to objects that start in the Block. 806 807 // Fill in the block table if necessary. This is unsynchronized, so multiple 808 // threads may fill the block table for a region (harmless, since it is 809 // idempotent). 810 if (!region_ptr->blocks_filled()) { 811 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 812 region_ptr->set_blocks_filled(); 813 } 814 815 HeapWord* const search_start = block_align_down(addr); 816 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 817 818 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 819 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 820 result += block_offset + live; 821 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 822 return result; 823 } 824 825 #ifdef ASSERT 826 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 827 { 828 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 829 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 830 for (const size_t* p = beg; p < end; ++p) { 831 assert(*p == 0, "not zero"); 832 } 833 } 834 835 void ParallelCompactData::verify_clear() 836 { 837 verify_clear(_region_vspace); 838 verify_clear(_block_vspace); 839 } 840 #endif // #ifdef ASSERT 841 842 STWGCTimer PSParallelCompact::_gc_timer; 843 ParallelOldTracer PSParallelCompact::_gc_tracer; 844 elapsedTimer PSParallelCompact::_accumulated_time; 845 unsigned int PSParallelCompact::_total_invocations = 0; 846 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 847 jlong PSParallelCompact::_time_of_last_gc = 0; 848 CollectorCounters* PSParallelCompact::_counters = NULL; 849 ParMarkBitMap PSParallelCompact::_mark_bitmap; 850 ParallelCompactData PSParallelCompact::_summary_data; 851 852 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 853 854 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 855 856 class PCReferenceProcessor: public ReferenceProcessor { 857 public: 858 PCReferenceProcessor( 859 BoolObjectClosure* is_subject_to_discovery, 860 BoolObjectClosure* is_alive_non_header) : 861 ReferenceProcessor(is_subject_to_discovery, 862 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 863 ParallelGCThreads, // mt processing degree 864 true, // mt discovery 865 ParallelGCThreads, // mt discovery degree 866 true, // atomic_discovery 867 is_alive_non_header) { 868 } 869 870 template<typename T> bool discover(oop obj, ReferenceType type) { 871 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj); 872 T heap_oop = RawAccess<>::oop_load(referent_addr); 873 oop referent = CompressedOops::decode_not_null(heap_oop); 874 return PSParallelCompact::mark_bitmap()->is_unmarked(referent) 875 && ReferenceProcessor::discover_reference(obj, type); 876 } 877 virtual bool discover_reference(oop obj, ReferenceType type) { 878 if (UseCompressedOops) { 879 return discover<narrowOop>(obj, type); 880 } else { 881 return discover<oop>(obj, type); 882 } 883 } 884 }; 885 886 void PSParallelCompact::post_initialize() { 887 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 888 _span_based_discoverer.set_span(heap->reserved_region()); 889 _ref_processor = 890 new PCReferenceProcessor(&_span_based_discoverer, 891 &_is_alive_closure); // non-header is alive closure 892 893 _counters = new CollectorCounters("Parallel full collection pauses", 1); 894 895 // Initialize static fields in ParCompactionManager. 896 ParCompactionManager::initialize(mark_bitmap()); 897 } 898 899 bool PSParallelCompact::initialize() { 900 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 901 MemRegion mr = heap->reserved_region(); 902 903 // Was the old gen get allocated successfully? 904 if (!heap->old_gen()->is_allocated()) { 905 return false; 906 } 907 908 initialize_space_info(); 909 initialize_dead_wood_limiter(); 910 911 if (!_mark_bitmap.initialize(mr)) { 912 vm_shutdown_during_initialization( 913 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 914 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 915 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 916 return false; 917 } 918 919 if (!_summary_data.initialize(mr)) { 920 vm_shutdown_during_initialization( 921 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 922 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 923 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 924 return false; 925 } 926 927 return true; 928 } 929 930 void PSParallelCompact::initialize_space_info() 931 { 932 memset(&_space_info, 0, sizeof(_space_info)); 933 934 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 935 PSYoungGen* young_gen = heap->young_gen(); 936 937 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 938 _space_info[eden_space_id].set_space(young_gen->eden_space()); 939 _space_info[from_space_id].set_space(young_gen->from_space()); 940 _space_info[to_space_id].set_space(young_gen->to_space()); 941 942 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 943 } 944 945 void PSParallelCompact::initialize_dead_wood_limiter() 946 { 947 const size_t max = 100; 948 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 949 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 950 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 951 DEBUG_ONLY(_dwl_initialized = true;) 952 _dwl_adjustment = normal_distribution(1.0); 953 } 954 955 void 956 PSParallelCompact::clear_data_covering_space(SpaceId id) 957 { 958 // At this point, top is the value before GC, new_top() is the value that will 959 // be set at the end of GC. The marking bitmap is cleared to top; nothing 960 // should be marked above top. The summary data is cleared to the larger of 961 // top & new_top. 962 MutableSpace* const space = _space_info[id].space(); 963 HeapWord* const bot = space->bottom(); 964 HeapWord* const top = space->top(); 965 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 966 967 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 968 const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top)); 969 _mark_bitmap.clear_range(beg_bit, end_bit); 970 971 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 972 const size_t end_region = 973 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 974 _summary_data.clear_range(beg_region, end_region); 975 976 // Clear the data used to 'split' regions. 977 SplitInfo& split_info = _space_info[id].split_info(); 978 if (split_info.is_valid()) { 979 split_info.clear(); 980 } 981 DEBUG_ONLY(split_info.verify_clear();) 982 } 983 984 void PSParallelCompact::pre_compact() 985 { 986 // Update the from & to space pointers in space_info, since they are swapped 987 // at each young gen gc. Do the update unconditionally (even though a 988 // promotion failure does not swap spaces) because an unknown number of young 989 // collections will have swapped the spaces an unknown number of times. 990 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 991 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 992 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 993 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 994 995 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 996 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 997 998 // Increment the invocation count 999 heap->increment_total_collections(true); 1000 1001 // We need to track unique mark sweep invocations as well. 1002 _total_invocations++; 1003 1004 heap->print_heap_before_gc(); 1005 heap->trace_heap_before_gc(&_gc_tracer); 1006 1007 // Fill in TLABs 1008 heap->ensure_parsability(true); // retire TLABs 1009 1010 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1011 HandleMark hm; // Discard invalid handles created during verification 1012 Universe::verify("Before GC"); 1013 } 1014 1015 // Verify object start arrays 1016 if (VerifyObjectStartArray && 1017 VerifyBeforeGC) { 1018 heap->old_gen()->verify_object_start_array(); 1019 } 1020 1021 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1022 DEBUG_ONLY(summary_data().verify_clear();) 1023 1024 ParCompactionManager::reset_all_bitmap_query_caches(); 1025 } 1026 1027 void PSParallelCompact::post_compact() 1028 { 1029 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1030 ParCompactionManager::remove_all_shadow_regions(); 1031 1032 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1033 // Clear the marking bitmap, summary data and split info. 1034 clear_data_covering_space(SpaceId(id)); 1035 // Update top(). Must be done after clearing the bitmap and summary data. 1036 _space_info[id].publish_new_top(); 1037 } 1038 1039 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1040 MutableSpace* const from_space = _space_info[from_space_id].space(); 1041 MutableSpace* const to_space = _space_info[to_space_id].space(); 1042 1043 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1044 bool eden_empty = eden_space->is_empty(); 1045 1046 // Update heap occupancy information which is used as input to the soft ref 1047 // clearing policy at the next gc. 1048 Universe::update_heap_info_at_gc(); 1049 1050 bool young_gen_empty = eden_empty && from_space->is_empty() && 1051 to_space->is_empty(); 1052 1053 PSCardTable* ct = heap->card_table(); 1054 MemRegion old_mr = heap->old_gen()->reserved(); 1055 if (young_gen_empty) { 1056 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1057 } else { 1058 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1059 } 1060 1061 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1062 ClassLoaderDataGraph::purge(); 1063 MetaspaceUtils::verify_metrics(); 1064 1065 heap->prune_scavengable_nmethods(); 1066 1067 #if COMPILER2_OR_JVMCI 1068 DerivedPointerTable::update_pointers(); 1069 #endif 1070 1071 if (ZapUnusedHeapArea) { 1072 heap->gen_mangle_unused_area(); 1073 } 1074 1075 // Update time of last GC 1076 reset_millis_since_last_gc(); 1077 } 1078 1079 HeapWord* 1080 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1081 bool maximum_compaction) 1082 { 1083 const size_t region_size = ParallelCompactData::RegionSize; 1084 const ParallelCompactData& sd = summary_data(); 1085 1086 const MutableSpace* const space = _space_info[id].space(); 1087 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1088 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1089 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1090 1091 // Skip full regions at the beginning of the space--they are necessarily part 1092 // of the dense prefix. 1093 size_t full_count = 0; 1094 const RegionData* cp; 1095 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1096 ++full_count; 1097 } 1098 1099 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1100 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1101 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1102 if (maximum_compaction || cp == end_cp || interval_ended) { 1103 _maximum_compaction_gc_num = total_invocations(); 1104 return sd.region_to_addr(cp); 1105 } 1106 1107 HeapWord* const new_top = _space_info[id].new_top(); 1108 const size_t space_live = pointer_delta(new_top, space->bottom()); 1109 const size_t space_used = space->used_in_words(); 1110 const size_t space_capacity = space->capacity_in_words(); 1111 1112 const double cur_density = double(space_live) / space_capacity; 1113 const double deadwood_density = 1114 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1115 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1116 1117 log_develop_debug(gc, compaction)( 1118 "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1119 cur_density, deadwood_density, deadwood_goal); 1120 log_develop_debug(gc, compaction)( 1121 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " " 1122 "space_cap=" SIZE_FORMAT, 1123 space_live, space_used, 1124 space_capacity); 1125 1126 // XXX - Use binary search? 1127 HeapWord* dense_prefix = sd.region_to_addr(cp); 1128 const RegionData* full_cp = cp; 1129 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1130 while (cp < end_cp) { 1131 HeapWord* region_destination = cp->destination(); 1132 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1133 1134 log_develop_trace(gc, compaction)( 1135 "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1136 "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8), 1137 sd.region(cp), p2i(region_destination), 1138 p2i(dense_prefix), cur_deadwood); 1139 1140 if (cur_deadwood >= deadwood_goal) { 1141 // Found the region that has the correct amount of deadwood to the left. 1142 // This typically occurs after crossing a fairly sparse set of regions, so 1143 // iterate backwards over those sparse regions, looking for the region 1144 // that has the lowest density of live objects 'to the right.' 1145 size_t space_to_left = sd.region(cp) * region_size; 1146 size_t live_to_left = space_to_left - cur_deadwood; 1147 size_t space_to_right = space_capacity - space_to_left; 1148 size_t live_to_right = space_live - live_to_left; 1149 double density_to_right = double(live_to_right) / space_to_right; 1150 while (cp > full_cp) { 1151 --cp; 1152 const size_t prev_region_live_to_right = live_to_right - 1153 cp->data_size(); 1154 const size_t prev_region_space_to_right = space_to_right + region_size; 1155 double prev_region_density_to_right = 1156 double(prev_region_live_to_right) / prev_region_space_to_right; 1157 if (density_to_right <= prev_region_density_to_right) { 1158 return dense_prefix; 1159 } 1160 1161 log_develop_trace(gc, compaction)( 1162 "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1163 "pc_d2r=%10.8f", 1164 sd.region(cp), density_to_right, 1165 prev_region_density_to_right); 1166 1167 dense_prefix -= region_size; 1168 live_to_right = prev_region_live_to_right; 1169 space_to_right = prev_region_space_to_right; 1170 density_to_right = prev_region_density_to_right; 1171 } 1172 return dense_prefix; 1173 } 1174 1175 dense_prefix += region_size; 1176 ++cp; 1177 } 1178 1179 return dense_prefix; 1180 } 1181 1182 #ifndef PRODUCT 1183 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1184 const SpaceId id, 1185 const bool maximum_compaction, 1186 HeapWord* const addr) 1187 { 1188 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1189 RegionData* const cp = summary_data().region(region_idx); 1190 const MutableSpace* const space = _space_info[id].space(); 1191 HeapWord* const new_top = _space_info[id].new_top(); 1192 1193 const size_t space_live = pointer_delta(new_top, space->bottom()); 1194 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1195 const size_t space_cap = space->capacity_in_words(); 1196 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1197 const size_t live_to_right = new_top - cp->destination(); 1198 const size_t dead_to_right = space->top() - addr - live_to_right; 1199 1200 log_develop_debug(gc, compaction)( 1201 "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1202 "spl=" SIZE_FORMAT " " 1203 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1204 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " " 1205 "ratio=%10.8f", 1206 algorithm, p2i(addr), region_idx, 1207 space_live, 1208 dead_to_left, dead_to_left_pct, 1209 dead_to_right, live_to_right, 1210 double(dead_to_right) / live_to_right); 1211 } 1212 #endif // #ifndef PRODUCT 1213 1214 // Return a fraction indicating how much of the generation can be treated as 1215 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1216 // based on the density of live objects in the generation to determine a limit, 1217 // which is then adjusted so the return value is min_percent when the density is 1218 // 1. 1219 // 1220 // The following table shows some return values for a different values of the 1221 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1222 // min_percent is 1. 1223 // 1224 // fraction allowed as dead wood 1225 // ----------------------------------------------------------------- 1226 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1227 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1228 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1229 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1230 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1231 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1232 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1233 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1234 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1235 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1236 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1237 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1238 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1239 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1240 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1241 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1242 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1243 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1244 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1245 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1246 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1247 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1248 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1249 1250 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1251 { 1252 assert(_dwl_initialized, "uninitialized"); 1253 1254 // The raw limit is the value of the normal distribution at x = density. 1255 const double raw_limit = normal_distribution(density); 1256 1257 // Adjust the raw limit so it becomes the minimum when the density is 1. 1258 // 1259 // First subtract the adjustment value (which is simply the precomputed value 1260 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1261 // Then add the minimum value, so the minimum is returned when the density is 1262 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1263 const double min = double(min_percent) / 100.0; 1264 const double limit = raw_limit - _dwl_adjustment + min; 1265 return MAX2(limit, 0.0); 1266 } 1267 1268 ParallelCompactData::RegionData* 1269 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1270 const RegionData* end) 1271 { 1272 const size_t region_size = ParallelCompactData::RegionSize; 1273 ParallelCompactData& sd = summary_data(); 1274 size_t left = sd.region(beg); 1275 size_t right = end > beg ? sd.region(end) - 1 : left; 1276 1277 // Binary search. 1278 while (left < right) { 1279 // Equivalent to (left + right) / 2, but does not overflow. 1280 const size_t middle = left + (right - left) / 2; 1281 RegionData* const middle_ptr = sd.region(middle); 1282 HeapWord* const dest = middle_ptr->destination(); 1283 HeapWord* const addr = sd.region_to_addr(middle); 1284 assert(dest != NULL, "sanity"); 1285 assert(dest <= addr, "must move left"); 1286 1287 if (middle > left && dest < addr) { 1288 right = middle - 1; 1289 } else if (middle < right && middle_ptr->data_size() == region_size) { 1290 left = middle + 1; 1291 } else { 1292 return middle_ptr; 1293 } 1294 } 1295 return sd.region(left); 1296 } 1297 1298 ParallelCompactData::RegionData* 1299 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1300 const RegionData* end, 1301 size_t dead_words) 1302 { 1303 ParallelCompactData& sd = summary_data(); 1304 size_t left = sd.region(beg); 1305 size_t right = end > beg ? sd.region(end) - 1 : left; 1306 1307 // Binary search. 1308 while (left < right) { 1309 // Equivalent to (left + right) / 2, but does not overflow. 1310 const size_t middle = left + (right - left) / 2; 1311 RegionData* const middle_ptr = sd.region(middle); 1312 HeapWord* const dest = middle_ptr->destination(); 1313 HeapWord* const addr = sd.region_to_addr(middle); 1314 assert(dest != NULL, "sanity"); 1315 assert(dest <= addr, "must move left"); 1316 1317 const size_t dead_to_left = pointer_delta(addr, dest); 1318 if (middle > left && dead_to_left > dead_words) { 1319 right = middle - 1; 1320 } else if (middle < right && dead_to_left < dead_words) { 1321 left = middle + 1; 1322 } else { 1323 return middle_ptr; 1324 } 1325 } 1326 return sd.region(left); 1327 } 1328 1329 // The result is valid during the summary phase, after the initial summarization 1330 // of each space into itself, and before final summarization. 1331 inline double 1332 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1333 HeapWord* const bottom, 1334 HeapWord* const top, 1335 HeapWord* const new_top) 1336 { 1337 ParallelCompactData& sd = summary_data(); 1338 1339 assert(cp != NULL, "sanity"); 1340 assert(bottom != NULL, "sanity"); 1341 assert(top != NULL, "sanity"); 1342 assert(new_top != NULL, "sanity"); 1343 assert(top >= new_top, "summary data problem?"); 1344 assert(new_top > bottom, "space is empty; should not be here"); 1345 assert(new_top >= cp->destination(), "sanity"); 1346 assert(top >= sd.region_to_addr(cp), "sanity"); 1347 1348 HeapWord* const destination = cp->destination(); 1349 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1350 const size_t compacted_region_live = pointer_delta(new_top, destination); 1351 const size_t compacted_region_used = pointer_delta(top, 1352 sd.region_to_addr(cp)); 1353 const size_t reclaimable = compacted_region_used - compacted_region_live; 1354 1355 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1356 return double(reclaimable) / divisor; 1357 } 1358 1359 // Return the address of the end of the dense prefix, a.k.a. the start of the 1360 // compacted region. The address is always on a region boundary. 1361 // 1362 // Completely full regions at the left are skipped, since no compaction can 1363 // occur in those regions. Then the maximum amount of dead wood to allow is 1364 // computed, based on the density (amount live / capacity) of the generation; 1365 // the region with approximately that amount of dead space to the left is 1366 // identified as the limit region. Regions between the last completely full 1367 // region and the limit region are scanned and the one that has the best 1368 // (maximum) reclaimed_ratio() is selected. 1369 HeapWord* 1370 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1371 bool maximum_compaction) 1372 { 1373 const size_t region_size = ParallelCompactData::RegionSize; 1374 const ParallelCompactData& sd = summary_data(); 1375 1376 const MutableSpace* const space = _space_info[id].space(); 1377 HeapWord* const top = space->top(); 1378 HeapWord* const top_aligned_up = sd.region_align_up(top); 1379 HeapWord* const new_top = _space_info[id].new_top(); 1380 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1381 HeapWord* const bottom = space->bottom(); 1382 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1383 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1384 const RegionData* const new_top_cp = 1385 sd.addr_to_region_ptr(new_top_aligned_up); 1386 1387 // Skip full regions at the beginning of the space--they are necessarily part 1388 // of the dense prefix. 1389 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1390 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1391 space->is_empty(), "no dead space allowed to the left"); 1392 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1393 "region must have dead space"); 1394 1395 // The gc number is saved whenever a maximum compaction is done, and used to 1396 // determine when the maximum compaction interval has expired. This avoids 1397 // successive max compactions for different reasons. 1398 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1399 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1400 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1401 total_invocations() == HeapFirstMaximumCompactionCount; 1402 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1403 _maximum_compaction_gc_num = total_invocations(); 1404 return sd.region_to_addr(full_cp); 1405 } 1406 1407 const size_t space_live = pointer_delta(new_top, bottom); 1408 const size_t space_used = space->used_in_words(); 1409 const size_t space_capacity = space->capacity_in_words(); 1410 1411 const double density = double(space_live) / double(space_capacity); 1412 const size_t min_percent_free = MarkSweepDeadRatio; 1413 const double limiter = dead_wood_limiter(density, min_percent_free); 1414 const size_t dead_wood_max = space_used - space_live; 1415 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1416 dead_wood_max); 1417 1418 log_develop_debug(gc, compaction)( 1419 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " " 1420 "space_cap=" SIZE_FORMAT, 1421 space_live, space_used, 1422 space_capacity); 1423 log_develop_debug(gc, compaction)( 1424 "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1425 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1426 density, min_percent_free, limiter, 1427 dead_wood_max, dead_wood_limit); 1428 1429 // Locate the region with the desired amount of dead space to the left. 1430 const RegionData* const limit_cp = 1431 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1432 1433 // Scan from the first region with dead space to the limit region and find the 1434 // one with the best (largest) reclaimed ratio. 1435 double best_ratio = 0.0; 1436 const RegionData* best_cp = full_cp; 1437 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1438 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1439 if (tmp_ratio > best_ratio) { 1440 best_cp = cp; 1441 best_ratio = tmp_ratio; 1442 } 1443 } 1444 1445 return sd.region_to_addr(best_cp); 1446 } 1447 1448 void PSParallelCompact::summarize_spaces_quick() 1449 { 1450 for (unsigned int i = 0; i < last_space_id; ++i) { 1451 const MutableSpace* space = _space_info[i].space(); 1452 HeapWord** nta = _space_info[i].new_top_addr(); 1453 bool result = _summary_data.summarize(_space_info[i].split_info(), 1454 space->bottom(), space->top(), NULL, 1455 space->bottom(), space->end(), nta); 1456 assert(result, "space must fit into itself"); 1457 _space_info[i].set_dense_prefix(space->bottom()); 1458 } 1459 } 1460 1461 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1462 { 1463 HeapWord* const dense_prefix_end = dense_prefix(id); 1464 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1465 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1466 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1467 // Only enough dead space is filled so that any remaining dead space to the 1468 // left is larger than the minimum filler object. (The remainder is filled 1469 // during the copy/update phase.) 1470 // 1471 // The size of the dead space to the right of the boundary is not a 1472 // concern, since compaction will be able to use whatever space is 1473 // available. 1474 // 1475 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1476 // surrounds the space to be filled with an object. 1477 // 1478 // In the 32-bit VM, each bit represents two 32-bit words: 1479 // +---+ 1480 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1481 // end_bits: ... x x x | 0 | || 0 x x ... 1482 // +---+ 1483 // 1484 // In the 64-bit VM, each bit represents one 64-bit word: 1485 // +------------+ 1486 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1487 // end_bits: ... x x 1 | 0 || 0 | x x ... 1488 // +------------+ 1489 // +-------+ 1490 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1491 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1492 // +-------+ 1493 // +-----------+ 1494 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1495 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1496 // +-----------+ 1497 // +-------+ 1498 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1499 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1500 // +-------+ 1501 1502 // Initially assume case a, c or e will apply. 1503 size_t obj_len = CollectedHeap::min_fill_size(); 1504 HeapWord* obj_beg = dense_prefix_end - obj_len; 1505 1506 #ifdef _LP64 1507 if (MinObjAlignment > 1) { // object alignment > heap word size 1508 // Cases a, c or e. 1509 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1510 // Case b above. 1511 obj_beg = dense_prefix_end - 1; 1512 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1513 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1514 // Case d above. 1515 obj_beg = dense_prefix_end - 3; 1516 obj_len = 3; 1517 } 1518 #endif // #ifdef _LP64 1519 1520 CollectedHeap::fill_with_object(obj_beg, obj_len); 1521 _mark_bitmap.mark_obj(obj_beg, obj_len); 1522 _summary_data.add_obj(obj_beg, obj_len); 1523 assert(start_array(id) != NULL, "sanity"); 1524 start_array(id)->allocate_block(obj_beg); 1525 } 1526 } 1527 1528 void 1529 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1530 { 1531 assert(id < last_space_id, "id out of range"); 1532 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1533 "should have been reset in summarize_spaces_quick()"); 1534 1535 const MutableSpace* space = _space_info[id].space(); 1536 if (_space_info[id].new_top() != space->bottom()) { 1537 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1538 _space_info[id].set_dense_prefix(dense_prefix_end); 1539 1540 #ifndef PRODUCT 1541 if (log_is_enabled(Debug, gc, compaction)) { 1542 print_dense_prefix_stats("ratio", id, maximum_compaction, 1543 dense_prefix_end); 1544 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1545 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1546 } 1547 #endif // #ifndef PRODUCT 1548 1549 // Recompute the summary data, taking into account the dense prefix. If 1550 // every last byte will be reclaimed, then the existing summary data which 1551 // compacts everything can be left in place. 1552 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1553 // If dead space crosses the dense prefix boundary, it is (at least 1554 // partially) filled with a dummy object, marked live and added to the 1555 // summary data. This simplifies the copy/update phase and must be done 1556 // before the final locations of objects are determined, to prevent 1557 // leaving a fragment of dead space that is too small to fill. 1558 fill_dense_prefix_end(id); 1559 1560 // Compute the destination of each Region, and thus each object. 1561 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1562 _summary_data.summarize(_space_info[id].split_info(), 1563 dense_prefix_end, space->top(), NULL, 1564 dense_prefix_end, space->end(), 1565 _space_info[id].new_top_addr()); 1566 } 1567 } 1568 1569 if (log_develop_is_enabled(Trace, gc, compaction)) { 1570 const size_t region_size = ParallelCompactData::RegionSize; 1571 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1572 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1573 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1574 HeapWord* const new_top = _space_info[id].new_top(); 1575 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1576 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1577 log_develop_trace(gc, compaction)( 1578 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1579 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1580 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1581 id, space->capacity_in_words(), p2i(dense_prefix_end), 1582 dp_region, dp_words / region_size, 1583 cr_words / region_size, p2i(new_top)); 1584 } 1585 } 1586 1587 #ifndef PRODUCT 1588 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1589 HeapWord* dst_beg, HeapWord* dst_end, 1590 SpaceId src_space_id, 1591 HeapWord* src_beg, HeapWord* src_end) 1592 { 1593 log_develop_trace(gc, compaction)( 1594 "Summarizing %d [%s] into %d [%s]: " 1595 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1596 SIZE_FORMAT "-" SIZE_FORMAT " " 1597 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1598 SIZE_FORMAT "-" SIZE_FORMAT, 1599 src_space_id, space_names[src_space_id], 1600 dst_space_id, space_names[dst_space_id], 1601 p2i(src_beg), p2i(src_end), 1602 _summary_data.addr_to_region_idx(src_beg), 1603 _summary_data.addr_to_region_idx(src_end), 1604 p2i(dst_beg), p2i(dst_end), 1605 _summary_data.addr_to_region_idx(dst_beg), 1606 _summary_data.addr_to_region_idx(dst_end)); 1607 } 1608 #endif // #ifndef PRODUCT 1609 1610 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1611 bool maximum_compaction) 1612 { 1613 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1614 1615 log_develop_debug(gc, marking)( 1616 "add_obj_count=" SIZE_FORMAT " " 1617 "add_obj_bytes=" SIZE_FORMAT, 1618 add_obj_count, 1619 add_obj_size * HeapWordSize); 1620 log_develop_debug(gc, marking)( 1621 "mark_bitmap_count=" SIZE_FORMAT " " 1622 "mark_bitmap_bytes=" SIZE_FORMAT, 1623 mark_bitmap_count, 1624 mark_bitmap_size * HeapWordSize); 1625 1626 // Quick summarization of each space into itself, to see how much is live. 1627 summarize_spaces_quick(); 1628 1629 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1630 NOT_PRODUCT(print_region_ranges()); 1631 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1632 1633 // The amount of live data that will end up in old space (assuming it fits). 1634 size_t old_space_total_live = 0; 1635 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1636 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1637 _space_info[id].space()->bottom()); 1638 } 1639 1640 MutableSpace* const old_space = _space_info[old_space_id].space(); 1641 const size_t old_capacity = old_space->capacity_in_words(); 1642 if (old_space_total_live > old_capacity) { 1643 // XXX - should also try to expand 1644 maximum_compaction = true; 1645 } 1646 1647 // Old generations. 1648 summarize_space(old_space_id, maximum_compaction); 1649 1650 // Summarize the remaining spaces in the young gen. The initial target space 1651 // is the old gen. If a space does not fit entirely into the target, then the 1652 // remainder is compacted into the space itself and that space becomes the new 1653 // target. 1654 SpaceId dst_space_id = old_space_id; 1655 HeapWord* dst_space_end = old_space->end(); 1656 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1657 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1658 const MutableSpace* space = _space_info[id].space(); 1659 const size_t live = pointer_delta(_space_info[id].new_top(), 1660 space->bottom()); 1661 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1662 1663 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1664 SpaceId(id), space->bottom(), space->top());) 1665 if (live > 0 && live <= available) { 1666 // All the live data will fit. 1667 bool done = _summary_data.summarize(_space_info[id].split_info(), 1668 space->bottom(), space->top(), 1669 NULL, 1670 *new_top_addr, dst_space_end, 1671 new_top_addr); 1672 assert(done, "space must fit into old gen"); 1673 1674 // Reset the new_top value for the space. 1675 _space_info[id].set_new_top(space->bottom()); 1676 } else if (live > 0) { 1677 // Attempt to fit part of the source space into the target space. 1678 HeapWord* next_src_addr = NULL; 1679 bool done = _summary_data.summarize(_space_info[id].split_info(), 1680 space->bottom(), space->top(), 1681 &next_src_addr, 1682 *new_top_addr, dst_space_end, 1683 new_top_addr); 1684 assert(!done, "space should not fit into old gen"); 1685 assert(next_src_addr != NULL, "sanity"); 1686 1687 // The source space becomes the new target, so the remainder is compacted 1688 // within the space itself. 1689 dst_space_id = SpaceId(id); 1690 dst_space_end = space->end(); 1691 new_top_addr = _space_info[id].new_top_addr(); 1692 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1693 space->bottom(), dst_space_end, 1694 SpaceId(id), next_src_addr, space->top());) 1695 done = _summary_data.summarize(_space_info[id].split_info(), 1696 next_src_addr, space->top(), 1697 NULL, 1698 space->bottom(), dst_space_end, 1699 new_top_addr); 1700 assert(done, "space must fit when compacted into itself"); 1701 assert(*new_top_addr <= space->top(), "usage should not grow"); 1702 } 1703 } 1704 1705 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1706 NOT_PRODUCT(print_region_ranges()); 1707 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1708 } 1709 1710 // This method should contain all heap-specific policy for invoking a full 1711 // collection. invoke_no_policy() will only attempt to compact the heap; it 1712 // will do nothing further. If we need to bail out for policy reasons, scavenge 1713 // before full gc, or any other specialized behavior, it needs to be added here. 1714 // 1715 // Note that this method should only be called from the vm_thread while at a 1716 // safepoint. 1717 // 1718 // Note that the all_soft_refs_clear flag in the soft ref policy 1719 // may be true because this method can be called without intervening 1720 // activity. For example when the heap space is tight and full measure 1721 // are being taken to free space. 1722 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1723 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1724 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1725 "should be in vm thread"); 1726 1727 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1728 GCCause::Cause gc_cause = heap->gc_cause(); 1729 assert(!heap->is_gc_active(), "not reentrant"); 1730 1731 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1732 IsGCActiveMark mark; 1733 1734 if (ScavengeBeforeFullGC) { 1735 PSScavenge::invoke_no_policy(); 1736 } 1737 1738 const bool clear_all_soft_refs = 1739 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1740 1741 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1742 maximum_heap_compaction); 1743 } 1744 1745 // This method contains no policy. You should probably 1746 // be calling invoke() instead. 1747 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1748 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1749 assert(ref_processor() != NULL, "Sanity"); 1750 1751 if (GCLocker::check_active_before_gc()) { 1752 return false; 1753 } 1754 1755 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1756 1757 GCIdMark gc_id_mark; 1758 _gc_timer.register_gc_start(); 1759 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1760 1761 TimeStamp marking_start; 1762 TimeStamp compaction_start; 1763 TimeStamp collection_exit; 1764 1765 GCCause::Cause gc_cause = heap->gc_cause(); 1766 PSYoungGen* young_gen = heap->young_gen(); 1767 PSOldGen* old_gen = heap->old_gen(); 1768 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1769 1770 // The scope of casr should end after code that can change 1771 // SoftRefPolicy::_should_clear_all_soft_refs. 1772 ClearedAllSoftRefs casr(maximum_heap_compaction, 1773 heap->soft_ref_policy()); 1774 1775 if (ZapUnusedHeapArea) { 1776 // Save information needed to minimize mangling 1777 heap->record_gen_tops_before_GC(); 1778 } 1779 1780 // Make sure data structures are sane, make the heap parsable, and do other 1781 // miscellaneous bookkeeping. 1782 pre_compact(); 1783 1784 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values(); 1785 1786 // Get the compaction manager reserved for the VM thread. 1787 ParCompactionManager* const vmthread_cm = 1788 ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers()); 1789 1790 { 1791 ResourceMark rm; 1792 HandleMark hm; 1793 1794 const uint active_workers = 1795 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(), 1796 ParallelScavengeHeap::heap()->workers().active_workers(), 1797 Threads::number_of_non_daemon_threads()); 1798 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers); 1799 1800 GCTraceCPUTime tcpu; 1801 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1802 1803 heap->pre_full_gc_dump(&_gc_timer); 1804 1805 TraceCollectorStats tcs(counters()); 1806 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1807 1808 if (log_is_enabled(Debug, gc, heap, exit)) { 1809 accumulated_time()->start(); 1810 } 1811 1812 // Let the size policy know we're starting 1813 size_policy->major_collection_begin(); 1814 1815 #if COMPILER2_OR_JVMCI 1816 DerivedPointerTable::clear(); 1817 #endif 1818 1819 ref_processor()->enable_discovery(); 1820 ref_processor()->setup_policy(maximum_heap_compaction); 1821 1822 bool marked_for_unloading = false; 1823 1824 marking_start.update(); 1825 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1826 1827 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1828 && GCCause::is_user_requested_gc(gc_cause); 1829 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1830 1831 #if COMPILER2_OR_JVMCI 1832 assert(DerivedPointerTable::is_active(), "Sanity"); 1833 DerivedPointerTable::set_active(false); 1834 #endif 1835 1836 // adjust_roots() updates Universe::_intArrayKlassObj which is 1837 // needed by the compaction for filling holes in the dense prefix. 1838 adjust_roots(vmthread_cm); 1839 1840 compaction_start.update(); 1841 compact(); 1842 1843 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1844 // done before resizing. 1845 post_compact(); 1846 1847 // Let the size policy know we're done 1848 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1849 1850 if (UseAdaptiveSizePolicy) { 1851 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1852 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1853 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1854 1855 // Don't check if the size_policy is ready here. Let 1856 // the size_policy check that internally. 1857 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1858 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1859 // Swap the survivor spaces if from_space is empty. The 1860 // resize_young_gen() called below is normally used after 1861 // a successful young GC and swapping of survivor spaces; 1862 // otherwise, it will fail to resize the young gen with 1863 // the current implementation. 1864 if (young_gen->from_space()->is_empty()) { 1865 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1866 young_gen->swap_spaces(); 1867 } 1868 1869 // Calculate optimal free space amounts 1870 assert(young_gen->max_gen_size() > 1871 young_gen->from_space()->capacity_in_bytes() + 1872 young_gen->to_space()->capacity_in_bytes(), 1873 "Sizes of space in young gen are out-of-bounds"); 1874 1875 size_t young_live = young_gen->used_in_bytes(); 1876 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1877 size_t old_live = old_gen->used_in_bytes(); 1878 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1879 size_t max_old_gen_size = old_gen->max_gen_size(); 1880 size_t max_eden_size = young_gen->max_gen_size() - 1881 young_gen->from_space()->capacity_in_bytes() - 1882 young_gen->to_space()->capacity_in_bytes(); 1883 1884 // Used for diagnostics 1885 size_policy->clear_generation_free_space_flags(); 1886 1887 size_policy->compute_generations_free_space(young_live, 1888 eden_live, 1889 old_live, 1890 cur_eden, 1891 max_old_gen_size, 1892 max_eden_size, 1893 true /* full gc*/); 1894 1895 size_policy->check_gc_overhead_limit(eden_live, 1896 max_old_gen_size, 1897 max_eden_size, 1898 true /* full gc*/, 1899 gc_cause, 1900 heap->soft_ref_policy()); 1901 1902 size_policy->decay_supplemental_growth(true /* full gc*/); 1903 1904 heap->resize_old_gen( 1905 size_policy->calculated_old_free_size_in_bytes()); 1906 1907 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1908 size_policy->calculated_survivor_size_in_bytes()); 1909 } 1910 1911 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1912 } 1913 1914 if (UsePerfData) { 1915 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1916 counters->update_counters(); 1917 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1918 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1919 } 1920 1921 heap->resize_all_tlabs(); 1922 1923 // Resize the metaspace capacity after a collection 1924 MetaspaceGC::compute_new_size(); 1925 1926 if (log_is_enabled(Debug, gc, heap, exit)) { 1927 accumulated_time()->stop(); 1928 } 1929 1930 heap->print_heap_change(pre_gc_values); 1931 1932 // Track memory usage and detect low memory 1933 MemoryService::track_memory_usage(); 1934 heap->update_counters(); 1935 1936 heap->post_full_gc_dump(&_gc_timer); 1937 } 1938 1939 #ifdef ASSERT 1940 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1941 ParCompactionManager* const cm = 1942 ParCompactionManager::manager_array(int(i)); 1943 assert(cm->marking_stack()->is_empty(), "should be empty"); 1944 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1945 } 1946 #endif // ASSERT 1947 1948 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1949 HandleMark hm; // Discard invalid handles created during verification 1950 Universe::verify("After GC"); 1951 } 1952 1953 // Re-verify object start arrays 1954 if (VerifyObjectStartArray && 1955 VerifyAfterGC) { 1956 old_gen->verify_object_start_array(); 1957 } 1958 1959 if (ZapUnusedHeapArea) { 1960 old_gen->object_space()->check_mangled_unused_area_complete(); 1961 } 1962 1963 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1964 1965 collection_exit.update(); 1966 1967 heap->print_heap_after_gc(); 1968 heap->trace_heap_after_gc(&_gc_tracer); 1969 1970 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1971 marking_start.ticks(), compaction_start.ticks(), 1972 collection_exit.ticks()); 1973 1974 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1975 1976 _gc_timer.register_gc_end(); 1977 1978 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1979 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1980 1981 return true; 1982 } 1983 1984 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 1985 private: 1986 uint _worker_id; 1987 1988 public: 1989 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { } 1990 void do_thread(Thread* thread) { 1991 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 1992 1993 ResourceMark rm; 1994 1995 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id); 1996 1997 PCMarkAndPushClosure mark_and_push_closure(cm); 1998 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations); 1999 2000 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs); 2001 2002 // Do the real work 2003 cm->follow_marking_stacks(); 2004 } 2005 }; 2006 2007 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) { 2008 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2009 2010 ParCompactionManager* cm = 2011 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2012 PCMarkAndPushClosure mark_and_push_closure(cm); 2013 2014 switch (root_type) { 2015 case ParallelRootType::universe: 2016 Universe::oops_do(&mark_and_push_closure); 2017 break; 2018 2019 case ParallelRootType::object_synchronizer: 2020 ObjectSynchronizer::oops_do(&mark_and_push_closure); 2021 break; 2022 2023 case ParallelRootType::management: 2024 Management::oops_do(&mark_and_push_closure); 2025 break; 2026 2027 case ParallelRootType::jvmti: 2028 JvmtiExport::oops_do(&mark_and_push_closure); 2029 break; 2030 2031 case ParallelRootType::class_loader_data: 2032 { 2033 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong); 2034 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure); 2035 } 2036 break; 2037 2038 case ParallelRootType::code_cache: 2039 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them. 2040 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure)); 2041 AOTLoader::oops_do(&mark_and_push_closure); 2042 break; 2043 2044 case ParallelRootType::sentinel: 2045 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds 2046 fatal("Bad enumeration value: %u", root_type); 2047 break; 2048 } 2049 2050 // Do the real work 2051 cm->follow_marking_stacks(); 2052 } 2053 2054 static void steal_marking_work(TaskTerminator& terminator, uint worker_id) { 2055 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2056 2057 ParCompactionManager* cm = 2058 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2059 2060 oop obj = NULL; 2061 ObjArrayTask task; 2062 do { 2063 while (ParCompactionManager::steal_objarray(worker_id, task)) { 2064 cm->follow_array((objArrayOop)task.obj(), task.index()); 2065 cm->follow_marking_stacks(); 2066 } 2067 while (ParCompactionManager::steal(worker_id, obj)) { 2068 cm->follow_contents(obj); 2069 cm->follow_marking_stacks(); 2070 } 2071 } while (!terminator.offer_termination()); 2072 } 2073 2074 class MarkFromRootsTask : public AbstractGangTask { 2075 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2076 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do 2077 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state; 2078 SequentialSubTasksDone _subtasks; 2079 TaskTerminator _terminator; 2080 uint _active_workers; 2081 2082 public: 2083 MarkFromRootsTask(uint active_workers) : 2084 AbstractGangTask("MarkFromRootsTask"), 2085 _strong_roots_scope(active_workers), 2086 _subtasks(), 2087 _terminator(active_workers, ParCompactionManager::oop_task_queues()), 2088 _active_workers(active_workers) { 2089 _subtasks.set_n_threads(active_workers); 2090 _subtasks.set_n_tasks(ParallelRootType::sentinel); 2091 } 2092 2093 virtual void work(uint worker_id) { 2094 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) { 2095 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id); 2096 } 2097 _subtasks.all_tasks_completed(); 2098 2099 PCAddThreadRootsMarkingTaskClosure closure(worker_id); 2100 Threads::possibly_parallel_threads_do(true /*parallel */, &closure); 2101 2102 // Mark from OopStorages 2103 { 2104 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2105 PCMarkAndPushClosure closure(cm); 2106 _oop_storage_set_par_state.oops_do(&closure); 2107 // Do the real work 2108 cm->follow_marking_stacks(); 2109 } 2110 2111 if (_active_workers > 1) { 2112 steal_marking_work(_terminator, worker_id); 2113 } 2114 } 2115 }; 2116 2117 class PCRefProcTask : public AbstractGangTask { 2118 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2119 ProcessTask& _task; 2120 uint _ergo_workers; 2121 TaskTerminator _terminator; 2122 2123 public: 2124 PCRefProcTask(ProcessTask& task, uint ergo_workers) : 2125 AbstractGangTask("PCRefProcTask"), 2126 _task(task), 2127 _ergo_workers(ergo_workers), 2128 _terminator(_ergo_workers, ParCompactionManager::oop_task_queues()) { 2129 } 2130 2131 virtual void work(uint worker_id) { 2132 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2133 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2134 2135 ParCompactionManager* cm = 2136 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2137 PCMarkAndPushClosure mark_and_push_closure(cm); 2138 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2139 _task.work(worker_id, *PSParallelCompact::is_alive_closure(), 2140 mark_and_push_closure, follow_stack_closure); 2141 2142 steal_marking_work(_terminator, worker_id); 2143 } 2144 }; 2145 2146 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2147 void execute(ProcessTask& process_task, uint ergo_workers) { 2148 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers, 2149 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 2150 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers()); 2151 2152 PCRefProcTask task(process_task, ergo_workers); 2153 ParallelScavengeHeap::heap()->workers().run_task(&task); 2154 } 2155 }; 2156 2157 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2158 bool maximum_heap_compaction, 2159 ParallelOldTracer *gc_tracer) { 2160 // Recursively traverse all live objects and mark them 2161 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2162 2163 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2164 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2165 2166 PCMarkAndPushClosure mark_and_push_closure(cm); 2167 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2168 2169 // Need new claim bits before marking starts. 2170 ClassLoaderDataGraph::clear_claimed_marks(); 2171 2172 { 2173 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2174 2175 MarkFromRootsTask task(active_gc_threads); 2176 ParallelScavengeHeap::heap()->workers().run_task(&task); 2177 } 2178 2179 // Process reference objects found during marking 2180 { 2181 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2182 2183 ReferenceProcessorStats stats; 2184 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2185 2186 if (ref_processor()->processing_is_mt()) { 2187 ref_processor()->set_active_mt_degree(active_gc_threads); 2188 2189 RefProcTaskExecutor task_executor; 2190 stats = ref_processor()->process_discovered_references( 2191 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2192 &task_executor, &pt); 2193 } else { 2194 stats = ref_processor()->process_discovered_references( 2195 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2196 &pt); 2197 } 2198 2199 gc_tracer->report_gc_reference_stats(stats); 2200 pt.print_all_references(); 2201 } 2202 2203 // This is the point where the entire marking should have completed. 2204 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2205 2206 { 2207 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2208 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2209 } 2210 2211 { 2212 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2213 2214 // Follow system dictionary roots and unload classes. 2215 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2216 2217 // Unload nmethods. 2218 CodeCache::do_unloading(is_alive_closure(), purged_class); 2219 2220 // Prune dead klasses from subklass/sibling/implementor lists. 2221 Klass::clean_weak_klass_links(purged_class); 2222 2223 // Clean JVMCI metadata handles. 2224 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 2225 } 2226 2227 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2228 } 2229 2230 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2231 // Adjust the pointers to reflect the new locations 2232 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2233 2234 // Need new claim bits when tracing through and adjusting pointers. 2235 ClassLoaderDataGraph::clear_claimed_marks(); 2236 2237 PCAdjustPointerClosure oop_closure(cm); 2238 2239 // General strong roots. 2240 Universe::oops_do(&oop_closure); 2241 Threads::oops_do(&oop_closure, NULL); 2242 ObjectSynchronizer::oops_do(&oop_closure); 2243 Management::oops_do(&oop_closure); 2244 JvmtiExport::oops_do(&oop_closure); 2245 OopStorageSet::strong_oops_do(&oop_closure); 2246 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong); 2247 ClassLoaderDataGraph::cld_do(&cld_closure); 2248 2249 // Now adjust pointers in remaining weak roots. (All of which should 2250 // have been cleared if they pointed to non-surviving objects.) 2251 WeakProcessor::oops_do(&oop_closure); 2252 2253 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2254 CodeCache::blobs_do(&adjust_from_blobs); 2255 AOT_ONLY(AOTLoader::oops_do(&oop_closure);) 2256 2257 ref_processor()->weak_oops_do(&oop_closure); 2258 // Roots were visited so references into the young gen in roots 2259 // may have been scanned. Process them also. 2260 // Should the reference processor have a span that excludes 2261 // young gen objects? 2262 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2263 } 2264 2265 // Helper class to print 8 region numbers per line and then print the total at the end. 2266 class FillableRegionLogger : public StackObj { 2267 private: 2268 Log(gc, compaction) log; 2269 static const int LineLength = 8; 2270 size_t _regions[LineLength]; 2271 int _next_index; 2272 bool _enabled; 2273 size_t _total_regions; 2274 public: 2275 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2276 ~FillableRegionLogger() { 2277 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2278 } 2279 2280 void print_line() { 2281 if (!_enabled || _next_index == 0) { 2282 return; 2283 } 2284 FormatBuffer<> line("Fillable: "); 2285 for (int i = 0; i < _next_index; i++) { 2286 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2287 } 2288 log.trace("%s", line.buffer()); 2289 _next_index = 0; 2290 } 2291 2292 void handle(size_t region) { 2293 if (!_enabled) { 2294 return; 2295 } 2296 _regions[_next_index++] = region; 2297 if (_next_index == LineLength) { 2298 print_line(); 2299 } 2300 _total_regions++; 2301 } 2302 }; 2303 2304 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads) 2305 { 2306 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2307 2308 // Find the threads that are active 2309 uint worker_id = 0; 2310 2311 // Find all regions that are available (can be filled immediately) and 2312 // distribute them to the thread stacks. The iteration is done in reverse 2313 // order (high to low) so the regions will be removed in ascending order. 2314 2315 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2316 2317 // id + 1 is used to test termination so unsigned can 2318 // be used with an old_space_id == 0. 2319 FillableRegionLogger region_logger; 2320 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2321 SpaceInfo* const space_info = _space_info + id; 2322 MutableSpace* const space = space_info->space(); 2323 HeapWord* const new_top = space_info->new_top(); 2324 2325 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2326 const size_t end_region = 2327 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2328 2329 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2330 if (sd.region(cur)->claim_unsafe()) { 2331 ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id); 2332 bool result = sd.region(cur)->mark_normal(); 2333 assert(result, "Must succeed at this point."); 2334 cm->region_stack()->push(cur); 2335 region_logger.handle(cur); 2336 // Assign regions to tasks in round-robin fashion. 2337 if (++worker_id == parallel_gc_threads) { 2338 worker_id = 0; 2339 } 2340 } 2341 } 2342 region_logger.print_line(); 2343 } 2344 } 2345 2346 class TaskQueue : StackObj { 2347 volatile uint _counter; 2348 uint _size; 2349 uint _insert_index; 2350 PSParallelCompact::UpdateDensePrefixTask* _backing_array; 2351 public: 2352 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) { 2353 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC); 2354 } 2355 ~TaskQueue() { 2356 assert(_counter >= _insert_index, "not all queue elements were claimed"); 2357 FREE_C_HEAP_ARRAY(T, _backing_array); 2358 } 2359 2360 void push(const PSParallelCompact::UpdateDensePrefixTask& value) { 2361 assert(_insert_index < _size, "too small backing array"); 2362 _backing_array[_insert_index++] = value; 2363 } 2364 2365 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) { 2366 uint claimed = Atomic::fetch_and_add(&_counter, 1u); 2367 if (claimed < _insert_index) { 2368 reference = _backing_array[claimed]; 2369 return true; 2370 } else { 2371 return false; 2372 } 2373 } 2374 }; 2375 2376 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2377 2378 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue, 2379 uint parallel_gc_threads) { 2380 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2381 2382 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2383 2384 // Iterate over all the spaces adding tasks for updating 2385 // regions in the dense prefix. Assume that 1 gc thread 2386 // will work on opening the gaps and the remaining gc threads 2387 // will work on the dense prefix. 2388 unsigned int space_id; 2389 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2390 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2391 const MutableSpace* const space = _space_info[space_id].space(); 2392 2393 if (dense_prefix_end == space->bottom()) { 2394 // There is no dense prefix for this space. 2395 continue; 2396 } 2397 2398 // The dense prefix is before this region. 2399 size_t region_index_end_dense_prefix = 2400 sd.addr_to_region_idx(dense_prefix_end); 2401 RegionData* const dense_prefix_cp = 2402 sd.region(region_index_end_dense_prefix); 2403 assert(dense_prefix_end == space->end() || 2404 dense_prefix_cp->available() || 2405 dense_prefix_cp->claimed(), 2406 "The region after the dense prefix should always be ready to fill"); 2407 2408 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2409 2410 // Is there dense prefix work? 2411 size_t total_dense_prefix_regions = 2412 region_index_end_dense_prefix - region_index_start; 2413 // How many regions of the dense prefix should be given to 2414 // each thread? 2415 if (total_dense_prefix_regions > 0) { 2416 uint tasks_for_dense_prefix = 1; 2417 if (total_dense_prefix_regions <= 2418 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2419 // Don't over partition. This assumes that 2420 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2421 // so there are not many regions to process. 2422 tasks_for_dense_prefix = parallel_gc_threads; 2423 } else { 2424 // Over partition 2425 tasks_for_dense_prefix = parallel_gc_threads * 2426 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2427 } 2428 size_t regions_per_thread = total_dense_prefix_regions / 2429 tasks_for_dense_prefix; 2430 // Give each thread at least 1 region. 2431 if (regions_per_thread == 0) { 2432 regions_per_thread = 1; 2433 } 2434 2435 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2436 if (region_index_start >= region_index_end_dense_prefix) { 2437 break; 2438 } 2439 // region_index_end is not processed 2440 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2441 region_index_end_dense_prefix); 2442 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2443 region_index_start, 2444 region_index_end)); 2445 region_index_start = region_index_end; 2446 } 2447 } 2448 // This gets any part of the dense prefix that did not 2449 // fit evenly. 2450 if (region_index_start < region_index_end_dense_prefix) { 2451 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2452 region_index_start, 2453 region_index_end_dense_prefix)); 2454 } 2455 } 2456 } 2457 2458 #ifdef ASSERT 2459 // Write a histogram of the number of times the block table was filled for a 2460 // region. 2461 void PSParallelCompact::write_block_fill_histogram() 2462 { 2463 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2464 return; 2465 } 2466 2467 Log(gc, compaction) log; 2468 ResourceMark rm; 2469 LogStream ls(log.trace()); 2470 outputStream* out = &ls; 2471 2472 typedef ParallelCompactData::RegionData rd_t; 2473 ParallelCompactData& sd = summary_data(); 2474 2475 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2476 MutableSpace* const spc = _space_info[id].space(); 2477 if (spc->bottom() != spc->top()) { 2478 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2479 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2480 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2481 2482 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2483 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2484 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2485 2486 for (const rd_t* cur = beg; cur < end; ++cur) { 2487 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2488 } 2489 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2490 for (size_t i = 0; i < histo_len; ++i) { 2491 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2492 histo[i], 100.0 * histo[i] / region_cnt); 2493 } 2494 out->cr(); 2495 } 2496 } 2497 } 2498 #endif // #ifdef ASSERT 2499 2500 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) { 2501 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2502 2503 ParCompactionManager* cm = 2504 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2505 2506 // Drain the stacks that have been preloaded with regions 2507 // that are ready to fill. 2508 2509 cm->drain_region_stacks(); 2510 2511 guarantee(cm->region_stack()->is_empty(), "Not empty"); 2512 2513 size_t region_index = 0; 2514 2515 while (true) { 2516 if (ParCompactionManager::steal(worker_id, region_index)) { 2517 PSParallelCompact::fill_and_update_region(cm, region_index); 2518 cm->drain_region_stacks(); 2519 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) { 2520 // Fill and update an unavailable region with the help of a shadow region 2521 PSParallelCompact::fill_and_update_shadow_region(cm, region_index); 2522 cm->drain_region_stacks(); 2523 } else { 2524 if (terminator->offer_termination()) { 2525 break; 2526 } 2527 // Go around again. 2528 } 2529 } 2530 return; 2531 } 2532 2533 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask { 2534 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2535 TaskQueue& _tq; 2536 TaskTerminator _terminator; 2537 uint _active_workers; 2538 2539 public: 2540 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) : 2541 AbstractGangTask("UpdateDensePrefixAndCompactionTask"), 2542 _tq(tq), 2543 _terminator(active_workers, ParCompactionManager::region_task_queues()), 2544 _active_workers(active_workers) { 2545 } 2546 virtual void work(uint worker_id) { 2547 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2548 2549 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) { 2550 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm, 2551 task._space_id, 2552 task._region_index_start, 2553 task._region_index_end); 2554 } 2555 2556 // Once a thread has drained it's stack, it should try to steal regions from 2557 // other threads. 2558 compaction_with_stealing_work(&_terminator, worker_id); 2559 } 2560 }; 2561 2562 void PSParallelCompact::compact() { 2563 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2564 2565 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2566 PSOldGen* old_gen = heap->old_gen(); 2567 old_gen->start_array()->reset(); 2568 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2569 2570 // for [0..last_space_id) 2571 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING) 2572 // push 2573 // push 2574 // 2575 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1) 2576 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)); 2577 initialize_shadow_regions(active_gc_threads); 2578 prepare_region_draining_tasks(active_gc_threads); 2579 enqueue_dense_prefix_tasks(task_queue, active_gc_threads); 2580 2581 { 2582 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2583 2584 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads); 2585 ParallelScavengeHeap::heap()->workers().run_task(&task); 2586 2587 #ifdef ASSERT 2588 // Verify that all regions have been processed before the deferred updates. 2589 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2590 verify_complete(SpaceId(id)); 2591 } 2592 #endif 2593 } 2594 2595 { 2596 // Update the deferred objects, if any. Any compaction manager can be used. 2597 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2598 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2599 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2600 update_deferred_objects(cm, SpaceId(id)); 2601 } 2602 } 2603 2604 DEBUG_ONLY(write_block_fill_histogram()); 2605 } 2606 2607 #ifdef ASSERT 2608 void PSParallelCompact::verify_complete(SpaceId space_id) { 2609 // All Regions between space bottom() to new_top() should be marked as filled 2610 // and all Regions between new_top() and top() should be available (i.e., 2611 // should have been emptied). 2612 ParallelCompactData& sd = summary_data(); 2613 SpaceInfo si = _space_info[space_id]; 2614 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2615 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2616 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2617 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2618 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2619 2620 bool issued_a_warning = false; 2621 2622 size_t cur_region; 2623 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2624 const RegionData* const c = sd.region(cur_region); 2625 if (!c->completed()) { 2626 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2627 cur_region, c->destination_count()); 2628 issued_a_warning = true; 2629 } 2630 } 2631 2632 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2633 const RegionData* const c = sd.region(cur_region); 2634 if (!c->available()) { 2635 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2636 cur_region, c->destination_count()); 2637 issued_a_warning = true; 2638 } 2639 } 2640 2641 if (issued_a_warning) { 2642 print_region_ranges(); 2643 } 2644 } 2645 #endif // #ifdef ASSERT 2646 2647 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2648 _start_array->allocate_block(addr); 2649 compaction_manager()->update_contents(oop(addr)); 2650 } 2651 2652 // Update interior oops in the ranges of regions [beg_region, end_region). 2653 void 2654 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2655 SpaceId space_id, 2656 size_t beg_region, 2657 size_t end_region) { 2658 ParallelCompactData& sd = summary_data(); 2659 ParMarkBitMap* const mbm = mark_bitmap(); 2660 2661 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2662 HeapWord* const end_addr = sd.region_to_addr(end_region); 2663 assert(beg_region <= end_region, "bad region range"); 2664 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2665 2666 #ifdef ASSERT 2667 // Claim the regions to avoid triggering an assert when they are marked as 2668 // filled. 2669 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2670 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2671 } 2672 #endif // #ifdef ASSERT 2673 2674 if (beg_addr != space(space_id)->bottom()) { 2675 // Find the first live object or block of dead space that *starts* in this 2676 // range of regions. If a partial object crosses onto the region, skip it; 2677 // it will be marked for 'deferred update' when the object head is 2678 // processed. If dead space crosses onto the region, it is also skipped; it 2679 // will be filled when the prior region is processed. If neither of those 2680 // apply, the first word in the region is the start of a live object or dead 2681 // space. 2682 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2683 const RegionData* const cp = sd.region(beg_region); 2684 if (cp->partial_obj_size() != 0) { 2685 beg_addr = sd.partial_obj_end(beg_region); 2686 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2687 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2688 } 2689 } 2690 2691 if (beg_addr < end_addr) { 2692 // A live object or block of dead space starts in this range of Regions. 2693 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2694 2695 // Create closures and iterate. 2696 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2697 FillClosure fill_closure(cm, space_id); 2698 ParMarkBitMap::IterationStatus status; 2699 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2700 dense_prefix_end); 2701 if (status == ParMarkBitMap::incomplete) { 2702 update_closure.do_addr(update_closure.source()); 2703 } 2704 } 2705 2706 // Mark the regions as filled. 2707 RegionData* const beg_cp = sd.region(beg_region); 2708 RegionData* const end_cp = sd.region(end_region); 2709 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2710 cp->set_completed(); 2711 } 2712 } 2713 2714 // Return the SpaceId for the space containing addr. If addr is not in the 2715 // heap, last_space_id is returned. In debug mode it expects the address to be 2716 // in the heap and asserts such. 2717 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2718 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2719 2720 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2721 if (_space_info[id].space()->contains(addr)) { 2722 return SpaceId(id); 2723 } 2724 } 2725 2726 assert(false, "no space contains the addr"); 2727 return last_space_id; 2728 } 2729 2730 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2731 SpaceId id) { 2732 assert(id < last_space_id, "bad space id"); 2733 2734 ParallelCompactData& sd = summary_data(); 2735 const SpaceInfo* const space_info = _space_info + id; 2736 ObjectStartArray* const start_array = space_info->start_array(); 2737 2738 const MutableSpace* const space = space_info->space(); 2739 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2740 HeapWord* const beg_addr = space_info->dense_prefix(); 2741 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2742 2743 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2744 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2745 const RegionData* cur_region; 2746 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2747 HeapWord* const addr = cur_region->deferred_obj_addr(); 2748 if (addr != NULL) { 2749 if (start_array != NULL) { 2750 start_array->allocate_block(addr); 2751 } 2752 cm->update_contents(oop(addr)); 2753 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2754 } 2755 } 2756 } 2757 2758 // Skip over count live words starting from beg, and return the address of the 2759 // next live word. Unless marked, the word corresponding to beg is assumed to 2760 // be dead. Callers must either ensure beg does not correspond to the middle of 2761 // an object, or account for those live words in some other way. Callers must 2762 // also ensure that there are enough live words in the range [beg, end) to skip. 2763 HeapWord* 2764 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2765 { 2766 assert(count > 0, "sanity"); 2767 2768 ParMarkBitMap* m = mark_bitmap(); 2769 idx_t bits_to_skip = m->words_to_bits(count); 2770 idx_t cur_beg = m->addr_to_bit(beg); 2771 const idx_t search_end = m->align_range_end(m->addr_to_bit(end)); 2772 2773 do { 2774 cur_beg = m->find_obj_beg(cur_beg, search_end); 2775 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2776 const size_t obj_bits = cur_end - cur_beg + 1; 2777 if (obj_bits > bits_to_skip) { 2778 return m->bit_to_addr(cur_beg + bits_to_skip); 2779 } 2780 bits_to_skip -= obj_bits; 2781 cur_beg = cur_end + 1; 2782 } while (bits_to_skip > 0); 2783 2784 // Skipping the desired number of words landed just past the end of an object. 2785 // Find the start of the next object. 2786 cur_beg = m->find_obj_beg(cur_beg, search_end); 2787 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2788 return m->bit_to_addr(cur_beg); 2789 } 2790 2791 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2792 SpaceId src_space_id, 2793 size_t src_region_idx) 2794 { 2795 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2796 2797 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2798 if (split_info.dest_region_addr() == dest_addr) { 2799 // The partial object ending at the split point contains the first word to 2800 // be copied to dest_addr. 2801 return split_info.first_src_addr(); 2802 } 2803 2804 const ParallelCompactData& sd = summary_data(); 2805 ParMarkBitMap* const bitmap = mark_bitmap(); 2806 const size_t RegionSize = ParallelCompactData::RegionSize; 2807 2808 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2809 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2810 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2811 HeapWord* const src_region_destination = src_region_ptr->destination(); 2812 2813 assert(dest_addr >= src_region_destination, "wrong src region"); 2814 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2815 2816 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2817 HeapWord* const src_region_end = src_region_beg + RegionSize; 2818 2819 HeapWord* addr = src_region_beg; 2820 if (dest_addr == src_region_destination) { 2821 // Return the first live word in the source region. 2822 if (partial_obj_size == 0) { 2823 addr = bitmap->find_obj_beg(addr, src_region_end); 2824 assert(addr < src_region_end, "no objects start in src region"); 2825 } 2826 return addr; 2827 } 2828 2829 // Must skip some live data. 2830 size_t words_to_skip = dest_addr - src_region_destination; 2831 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2832 2833 if (partial_obj_size >= words_to_skip) { 2834 // All the live words to skip are part of the partial object. 2835 addr += words_to_skip; 2836 if (partial_obj_size == words_to_skip) { 2837 // Find the first live word past the partial object. 2838 addr = bitmap->find_obj_beg(addr, src_region_end); 2839 assert(addr < src_region_end, "wrong src region"); 2840 } 2841 return addr; 2842 } 2843 2844 // Skip over the partial object (if any). 2845 if (partial_obj_size != 0) { 2846 words_to_skip -= partial_obj_size; 2847 addr += partial_obj_size; 2848 } 2849 2850 // Skip over live words due to objects that start in the region. 2851 addr = skip_live_words(addr, src_region_end, words_to_skip); 2852 assert(addr < src_region_end, "wrong src region"); 2853 return addr; 2854 } 2855 2856 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2857 SpaceId src_space_id, 2858 size_t beg_region, 2859 HeapWord* end_addr) 2860 { 2861 ParallelCompactData& sd = summary_data(); 2862 2863 #ifdef ASSERT 2864 MutableSpace* const src_space = _space_info[src_space_id].space(); 2865 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2866 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2867 "src_space_id does not match beg_addr"); 2868 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2869 "src_space_id does not match end_addr"); 2870 #endif // #ifdef ASSERT 2871 2872 RegionData* const beg = sd.region(beg_region); 2873 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2874 2875 // Regions up to new_top() are enqueued if they become available. 2876 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2877 RegionData* const enqueue_end = 2878 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2879 2880 for (RegionData* cur = beg; cur < end; ++cur) { 2881 assert(cur->data_size() > 0, "region must have live data"); 2882 cur->decrement_destination_count(); 2883 if (cur < enqueue_end && cur->available() && cur->claim()) { 2884 if (cur->mark_normal()) { 2885 cm->push_region(sd.region(cur)); 2886 } else if (cur->mark_copied()) { 2887 // Try to copy the content of the shadow region back to its corresponding 2888 // heap region if the shadow region is filled. Otherwise, the GC thread 2889 // fills the shadow region will copy the data back (see 2890 // MoveAndUpdateShadowClosure::complete_region). 2891 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur)); 2892 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region()); 2893 cur->set_completed(); 2894 } 2895 } 2896 } 2897 } 2898 2899 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2900 SpaceId& src_space_id, 2901 HeapWord*& src_space_top, 2902 HeapWord* end_addr) 2903 { 2904 typedef ParallelCompactData::RegionData RegionData; 2905 2906 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2907 const size_t region_size = ParallelCompactData::RegionSize; 2908 2909 size_t src_region_idx = 0; 2910 2911 // Skip empty regions (if any) up to the top of the space. 2912 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2913 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2914 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2915 const RegionData* const top_region_ptr = 2916 sd.addr_to_region_ptr(top_aligned_up); 2917 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2918 ++src_region_ptr; 2919 } 2920 2921 if (src_region_ptr < top_region_ptr) { 2922 // The next source region is in the current space. Update src_region_idx 2923 // and the source address to match src_region_ptr. 2924 src_region_idx = sd.region(src_region_ptr); 2925 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2926 if (src_region_addr > closure.source()) { 2927 closure.set_source(src_region_addr); 2928 } 2929 return src_region_idx; 2930 } 2931 2932 // Switch to a new source space and find the first non-empty region. 2933 unsigned int space_id = src_space_id + 1; 2934 assert(space_id < last_space_id, "not enough spaces"); 2935 2936 HeapWord* const destination = closure.destination(); 2937 2938 do { 2939 MutableSpace* space = _space_info[space_id].space(); 2940 HeapWord* const bottom = space->bottom(); 2941 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2942 2943 // Iterate over the spaces that do not compact into themselves. 2944 if (bottom_cp->destination() != bottom) { 2945 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2946 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2947 2948 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2949 if (src_cp->live_obj_size() > 0) { 2950 // Found it. 2951 assert(src_cp->destination() == destination, 2952 "first live obj in the space must match the destination"); 2953 assert(src_cp->partial_obj_size() == 0, 2954 "a space cannot begin with a partial obj"); 2955 2956 src_space_id = SpaceId(space_id); 2957 src_space_top = space->top(); 2958 const size_t src_region_idx = sd.region(src_cp); 2959 closure.set_source(sd.region_to_addr(src_region_idx)); 2960 return src_region_idx; 2961 } else { 2962 assert(src_cp->data_size() == 0, "sanity"); 2963 } 2964 } 2965 } 2966 } while (++space_id < last_space_id); 2967 2968 assert(false, "no source region was found"); 2969 return 0; 2970 } 2971 2972 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx) 2973 { 2974 typedef ParMarkBitMap::IterationStatus IterationStatus; 2975 ParMarkBitMap* const bitmap = mark_bitmap(); 2976 ParallelCompactData& sd = summary_data(); 2977 RegionData* const region_ptr = sd.region(region_idx); 2978 2979 // Get the source region and related info. 2980 size_t src_region_idx = region_ptr->source_region(); 2981 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2982 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2983 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2984 2985 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2986 2987 // Adjust src_region_idx to prepare for decrementing destination counts (the 2988 // destination count is not decremented when a region is copied to itself). 2989 if (src_region_idx == region_idx) { 2990 src_region_idx += 1; 2991 } 2992 2993 if (bitmap->is_unmarked(closure.source())) { 2994 // The first source word is in the middle of an object; copy the remainder 2995 // of the object or as much as will fit. The fact that pointer updates were 2996 // deferred will be noted when the object header is processed. 2997 HeapWord* const old_src_addr = closure.source(); 2998 closure.copy_partial_obj(); 2999 if (closure.is_full()) { 3000 decrement_destination_counts(cm, src_space_id, src_region_idx, 3001 closure.source()); 3002 region_ptr->set_deferred_obj_addr(NULL); 3003 closure.complete_region(cm, dest_addr, region_ptr); 3004 return; 3005 } 3006 3007 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3008 if (sd.region_align_down(old_src_addr) != end_addr) { 3009 // The partial object was copied from more than one source region. 3010 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3011 3012 // Move to the next source region, possibly switching spaces as well. All 3013 // args except end_addr may be modified. 3014 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3015 end_addr); 3016 } 3017 } 3018 3019 do { 3020 HeapWord* const cur_addr = closure.source(); 3021 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3022 src_space_top); 3023 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3024 3025 if (status == ParMarkBitMap::incomplete) { 3026 // The last obj that starts in the source region does not end in the 3027 // region. 3028 assert(closure.source() < end_addr, "sanity"); 3029 HeapWord* const obj_beg = closure.source(); 3030 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3031 src_space_top); 3032 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3033 if (obj_end < range_end) { 3034 // The end was found; the entire object will fit. 3035 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3036 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3037 } else { 3038 // The end was not found; the object will not fit. 3039 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3040 status = ParMarkBitMap::would_overflow; 3041 } 3042 } 3043 3044 if (status == ParMarkBitMap::would_overflow) { 3045 // The last object did not fit. Note that interior oop updates were 3046 // deferred, then copy enough of the object to fill the region. 3047 region_ptr->set_deferred_obj_addr(closure.destination()); 3048 status = closure.copy_until_full(); // copies from closure.source() 3049 3050 decrement_destination_counts(cm, src_space_id, src_region_idx, 3051 closure.source()); 3052 closure.complete_region(cm, dest_addr, region_ptr); 3053 return; 3054 } 3055 3056 if (status == ParMarkBitMap::full) { 3057 decrement_destination_counts(cm, src_space_id, src_region_idx, 3058 closure.source()); 3059 region_ptr->set_deferred_obj_addr(NULL); 3060 closure.complete_region(cm, dest_addr, region_ptr); 3061 return; 3062 } 3063 3064 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3065 3066 // Move to the next source region, possibly switching spaces as well. All 3067 // args except end_addr may be modified. 3068 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3069 end_addr); 3070 } while (true); 3071 } 3072 3073 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx) 3074 { 3075 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3076 fill_region(cm, cl, region_idx); 3077 } 3078 3079 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx) 3080 { 3081 // Get a shadow region first 3082 ParallelCompactData& sd = summary_data(); 3083 RegionData* const region_ptr = sd.region(region_idx); 3084 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr); 3085 // The InvalidShadow return value indicates the corresponding heap region is available, 3086 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use 3087 // MoveAndUpdateShadowClosure to fill the acquired shadow region. 3088 if (shadow_region == ParCompactionManager::InvalidShadow) { 3089 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3090 region_ptr->shadow_to_normal(); 3091 return fill_region(cm, cl, region_idx); 3092 } else { 3093 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region); 3094 return fill_region(cm, cl, region_idx); 3095 } 3096 } 3097 3098 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr) 3099 { 3100 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize); 3101 } 3102 3103 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx) 3104 { 3105 size_t next = cm->next_shadow_region(); 3106 ParallelCompactData& sd = summary_data(); 3107 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top()); 3108 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 3109 3110 while (next < old_new_top) { 3111 if (sd.region(next)->mark_shadow()) { 3112 region_idx = next; 3113 return true; 3114 } 3115 next = cm->move_next_shadow_region_by(active_gc_threads); 3116 } 3117 3118 return false; 3119 } 3120 3121 // The shadow region is an optimization to address region dependencies in full GC. The basic 3122 // idea is making more regions available by temporally storing their live objects in empty 3123 // shadow regions to resolve dependencies between them and the destination regions. Therefore, 3124 // GC threads need not wait destination regions to be available before processing sources. 3125 // 3126 // A typical workflow would be: 3127 // After draining its own stack and failing to steal from others, a GC worker would pick an 3128 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills 3129 // the shadow region by copying live objects from source regions of the unavailable one. Once 3130 // the unavailable region becomes available, the data in the shadow region will be copied back. 3131 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces. 3132 // 3133 // For more details, please refer to ยง4.2 of the VEE'19 paper: 3134 // Haoyu Li, Mingyu Wu, Binyu Zang, and Haibo Chen. 2019. ScissorGC: scalable and efficient 3135 // compaction for Java full garbage collection. In Proceedings of the 15th ACM SIGPLAN/SIGOPS 3136 // International Conference on Virtual Execution Environments (VEE 2019). ACM, New York, NY, USA, 3137 // 108-121. DOI: https://doi.org/10.1145/3313808.3313820 3138 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads) 3139 { 3140 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 3141 3142 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 3143 SpaceInfo* const space_info = _space_info + id; 3144 MutableSpace* const space = space_info->space(); 3145 3146 const size_t beg_region = 3147 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top()))); 3148 const size_t end_region = 3149 sd.addr_to_region_idx(sd.region_align_down(space->end())); 3150 3151 for (size_t cur = beg_region; cur < end_region; ++cur) { 3152 ParCompactionManager::push_shadow_region(cur); 3153 } 3154 } 3155 3156 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix()); 3157 for (uint i = 0; i < parallel_gc_threads; i++) { 3158 ParCompactionManager *cm = ParCompactionManager::manager_array(i); 3159 cm->set_next_shadow_region(beg_region + i); 3160 } 3161 } 3162 3163 void PSParallelCompact::fill_blocks(size_t region_idx) 3164 { 3165 // Fill in the block table elements for the specified region. Each block 3166 // table element holds the number of live words in the region that are to the 3167 // left of the first object that starts in the block. Thus only blocks in 3168 // which an object starts need to be filled. 3169 // 3170 // The algorithm scans the section of the bitmap that corresponds to the 3171 // region, keeping a running total of the live words. When an object start is 3172 // found, if it's the first to start in the block that contains it, the 3173 // current total is written to the block table element. 3174 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3175 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3176 const size_t RegionSize = ParallelCompactData::RegionSize; 3177 3178 ParallelCompactData& sd = summary_data(); 3179 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3180 if (partial_obj_size >= RegionSize) { 3181 return; // No objects start in this region. 3182 } 3183 3184 // Ensure the first loop iteration decides that the block has changed. 3185 size_t cur_block = sd.block_count(); 3186 3187 const ParMarkBitMap* const bitmap = mark_bitmap(); 3188 3189 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3190 assert((size_t)1 << Log2BitsPerBlock == 3191 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3192 3193 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3194 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3195 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3196 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3197 while (beg_bit < range_end) { 3198 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3199 if (new_block != cur_block) { 3200 cur_block = new_block; 3201 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3202 } 3203 3204 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3205 if (end_bit < range_end - 1) { 3206 live_bits += end_bit - beg_bit + 1; 3207 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3208 } else { 3209 return; 3210 } 3211 } 3212 } 3213 3214 jlong PSParallelCompact::millis_since_last_gc() { 3215 // We need a monotonically non-decreasing time in ms but 3216 // os::javaTimeMillis() does not guarantee monotonicity. 3217 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3218 jlong ret_val = now - _time_of_last_gc; 3219 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3220 if (ret_val < 0) { 3221 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3222 return 0; 3223 } 3224 return ret_val; 3225 } 3226 3227 void PSParallelCompact::reset_millis_since_last_gc() { 3228 // We need a monotonically non-decreasing time in ms but 3229 // os::javaTimeMillis() does not guarantee monotonicity. 3230 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3231 } 3232 3233 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3234 { 3235 if (source() != copy_destination()) { 3236 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3237 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining()); 3238 } 3239 update_state(words_remaining()); 3240 assert(is_full(), "sanity"); 3241 return ParMarkBitMap::full; 3242 } 3243 3244 void MoveAndUpdateClosure::copy_partial_obj() 3245 { 3246 size_t words = words_remaining(); 3247 3248 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3249 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3250 if (end_addr < range_end) { 3251 words = bitmap()->obj_size(source(), end_addr); 3252 } 3253 3254 // This test is necessary; if omitted, the pointer updates to a partial object 3255 // that crosses the dense prefix boundary could be overwritten. 3256 if (source() != copy_destination()) { 3257 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3258 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3259 } 3260 update_state(words); 3261 } 3262 3263 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3264 PSParallelCompact::RegionData *region_ptr) { 3265 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished"); 3266 region_ptr->set_completed(); 3267 } 3268 3269 ParMarkBitMapClosure::IterationStatus 3270 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3271 assert(destination() != NULL, "sanity"); 3272 assert(bitmap()->obj_size(addr) == words, "bad size"); 3273 3274 _source = addr; 3275 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3276 destination(), "wrong destination"); 3277 3278 if (words > words_remaining()) { 3279 return ParMarkBitMap::would_overflow; 3280 } 3281 3282 // The start_array must be updated even if the object is not moving. 3283 if (_start_array != NULL) { 3284 _start_array->allocate_block(destination()); 3285 } 3286 3287 if (copy_destination() != source()) { 3288 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3289 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3290 } 3291 3292 oop moved_oop = (oop) copy_destination(); 3293 compaction_manager()->update_contents(moved_oop); 3294 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3295 3296 update_state(words); 3297 assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity"); 3298 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3299 } 3300 3301 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3302 PSParallelCompact::RegionData *region_ptr) { 3303 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow"); 3304 // Record the shadow region index 3305 region_ptr->set_shadow_region(_shadow); 3306 // Mark the shadow region as filled to indicate the data is ready to be 3307 // copied back 3308 region_ptr->mark_filled(); 3309 // Try to copy the content of the shadow region back to its corresponding 3310 // heap region if available; the GC thread that decreases the destination 3311 // count to zero will do the copying otherwise (see 3312 // PSParallelCompact::decrement_destination_counts). 3313 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) { 3314 region_ptr->set_completed(); 3315 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr); 3316 ParCompactionManager::push_shadow_region_mt_safe(_shadow); 3317 } 3318 } 3319 3320 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3321 ParCompactionManager* cm, 3322 PSParallelCompact::SpaceId space_id) : 3323 ParMarkBitMapClosure(mbm, cm), 3324 _space_id(space_id), 3325 _start_array(PSParallelCompact::start_array(space_id)) 3326 { 3327 } 3328 3329 // Updates the references in the object to their new values. 3330 ParMarkBitMapClosure::IterationStatus 3331 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3332 do_addr(addr); 3333 return ParMarkBitMap::incomplete; 3334 } 3335 3336 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3337 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3338 _start_array(PSParallelCompact::start_array(space_id)) 3339 { 3340 assert(space_id == PSParallelCompact::old_space_id, 3341 "cannot use FillClosure in the young gen"); 3342 } 3343 3344 ParMarkBitMapClosure::IterationStatus 3345 FillClosure::do_addr(HeapWord* addr, size_t size) { 3346 CollectedHeap::fill_with_objects(addr, size); 3347 HeapWord* const end = addr + size; 3348 do { 3349 _start_array->allocate_block(addr); 3350 addr += oop(addr)->size(); 3351 } while (addr < end); 3352 return ParMarkBitMap::incomplete; 3353 }