1 /* 2 * Copyright (c) 2005, 2024, 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 "classfile/classLoaderDataGraph.hpp" 27 #include "classfile/javaClasses.inline.hpp" 28 #include "classfile/stringTable.hpp" 29 #include "classfile/symbolTable.hpp" 30 #include "classfile/systemDictionary.hpp" 31 #include "code/codeCache.hpp" 32 #include "compiler/oopMap.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/psStringDedup.hpp" 44 #include "gc/parallel/psYoungGen.hpp" 45 #include "gc/shared/classUnloadingContext.hpp" 46 #include "gc/shared/gcCause.hpp" 47 #include "gc/shared/gcHeapSummary.hpp" 48 #include "gc/shared/gcId.hpp" 49 #include "gc/shared/gcLocker.hpp" 50 #include "gc/shared/gcTimer.hpp" 51 #include "gc/shared/gcTrace.hpp" 52 #include "gc/shared/gcTraceTime.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/oopStorage.inline.hpp" 55 #include "gc/shared/oopStorageSet.inline.hpp" 56 #include "gc/shared/oopStorageSetParState.inline.hpp" 57 #include "gc/shared/referencePolicy.hpp" 58 #include "gc/shared/referenceProcessor.hpp" 59 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 60 #include "gc/shared/spaceDecorator.inline.hpp" 61 #include "gc/shared/taskTerminator.hpp" 62 #include "gc/shared/weakProcessor.inline.hpp" 63 #include "gc/shared/workerPolicy.hpp" 64 #include "gc/shared/workerThread.hpp" 65 #include "gc/shared/workerUtils.hpp" 66 #include "logging/log.hpp" 67 #include "memory/iterator.inline.hpp" 68 #include "memory/metaspaceUtils.hpp" 69 #include "memory/resourceArea.hpp" 70 #include "memory/universe.hpp" 71 #include "nmt/memTracker.hpp" 72 #include "oops/access.inline.hpp" 73 #include "oops/instanceClassLoaderKlass.inline.hpp" 74 #include "oops/instanceKlass.inline.hpp" 75 #include "oops/instanceMirrorKlass.inline.hpp" 76 #include "oops/methodData.hpp" 77 #include "oops/objArrayKlass.inline.hpp" 78 #include "oops/oop.inline.hpp" 79 #include "runtime/atomic.hpp" 80 #include "runtime/handles.inline.hpp" 81 #include "runtime/java.hpp" 82 #include "runtime/safepoint.hpp" 83 #include "runtime/threads.hpp" 84 #include "runtime/vmThread.hpp" 85 #include "services/memoryService.hpp" 86 #include "utilities/align.hpp" 87 #include "utilities/debug.hpp" 88 #include "utilities/events.hpp" 89 #include "utilities/formatBuffer.hpp" 90 #include "utilities/macros.hpp" 91 #include "utilities/stack.inline.hpp" 92 #if INCLUDE_JVMCI 93 #include "jvmci/jvmci.hpp" 94 #endif 95 96 #include <math.h> 97 98 // All sizes are in HeapWords. 99 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 100 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 101 const size_t ParallelCompactData::RegionSizeBytes = 102 RegionSize << LogHeapWordSize; 103 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 104 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 105 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 106 107 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 108 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 109 const size_t ParallelCompactData::BlockSizeBytes = 110 BlockSize << LogHeapWordSize; 111 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 112 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 113 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 114 115 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 116 const size_t ParallelCompactData::Log2BlocksPerRegion = 117 Log2RegionSize - Log2BlockSize; 118 119 const ParallelCompactData::RegionData::region_sz_t 120 ParallelCompactData::RegionData::dc_shift = 27; 121 122 const ParallelCompactData::RegionData::region_sz_t 123 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 124 125 const ParallelCompactData::RegionData::region_sz_t 126 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 127 128 const ParallelCompactData::RegionData::region_sz_t 129 ParallelCompactData::RegionData::los_mask = ~dc_mask; 130 131 const ParallelCompactData::RegionData::region_sz_t 132 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 133 134 const ParallelCompactData::RegionData::region_sz_t 135 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 136 137 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 138 139 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 140 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr; 141 142 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 143 HeapWord* destination) 144 { 145 assert(src_region_idx != 0, "invalid src_region_idx"); 146 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 147 assert(destination != nullptr, "invalid destination argument"); 148 149 _src_region_idx = src_region_idx; 150 _partial_obj_size = partial_obj_size; 151 _destination = destination; 152 153 // These fields may not be updated below, so make sure they're clear. 154 assert(_dest_region_addr == nullptr, "should have been cleared"); 155 assert(_first_src_addr == nullptr, "should have been cleared"); 156 157 // Determine the number of destination regions for the partial object. 158 HeapWord* const last_word = destination + partial_obj_size - 1; 159 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 160 HeapWord* const beg_region_addr = sd.region_align_down(destination); 161 HeapWord* const end_region_addr = sd.region_align_down(last_word); 162 163 if (beg_region_addr == end_region_addr) { 164 // One destination region. 165 _destination_count = 1; 166 if (end_region_addr == destination) { 167 // The destination falls on a region boundary, thus the first word of the 168 // partial object will be the first word copied to the destination region. 169 _dest_region_addr = end_region_addr; 170 _first_src_addr = sd.region_to_addr(src_region_idx); 171 } 172 } else { 173 // Two destination regions. When copied, the partial object will cross a 174 // destination region boundary, so a word somewhere within the partial 175 // object will be the first word copied to the second destination region. 176 _destination_count = 2; 177 _dest_region_addr = end_region_addr; 178 const size_t ofs = pointer_delta(end_region_addr, destination); 179 assert(ofs < _partial_obj_size, "sanity"); 180 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 181 } 182 } 183 184 void SplitInfo::clear() 185 { 186 _src_region_idx = 0; 187 _partial_obj_size = 0; 188 _destination = nullptr; 189 _destination_count = 0; 190 _dest_region_addr = nullptr; 191 _first_src_addr = nullptr; 192 assert(!is_valid(), "sanity"); 193 } 194 195 #ifdef ASSERT 196 void SplitInfo::verify_clear() 197 { 198 assert(_src_region_idx == 0, "not clear"); 199 assert(_partial_obj_size == 0, "not clear"); 200 assert(_destination == nullptr, "not clear"); 201 assert(_destination_count == 0, "not clear"); 202 assert(_dest_region_addr == nullptr, "not clear"); 203 assert(_first_src_addr == nullptr, "not clear"); 204 } 205 #endif // #ifdef ASSERT 206 207 208 void PSParallelCompact::print_on_error(outputStream* st) { 209 _mark_bitmap.print_on_error(st); 210 } 211 212 #ifndef PRODUCT 213 const char* PSParallelCompact::space_names[] = { 214 "old ", "eden", "from", "to " 215 }; 216 217 void PSParallelCompact::print_region_ranges() { 218 if (!log_develop_is_enabled(Trace, gc, compaction)) { 219 return; 220 } 221 Log(gc, compaction) log; 222 ResourceMark rm; 223 LogStream ls(log.trace()); 224 Universe::print_on(&ls); 225 log.trace("space bottom top end new_top"); 226 log.trace("------ ---------- ---------- ---------- ----------"); 227 228 for (unsigned int id = 0; id < last_space_id; ++id) { 229 const MutableSpace* space = _space_info[id].space(); 230 log.trace("%u %s " 231 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 232 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 233 id, space_names[id], 234 summary_data().addr_to_region_idx(space->bottom()), 235 summary_data().addr_to_region_idx(space->top()), 236 summary_data().addr_to_region_idx(space->end()), 237 summary_data().addr_to_region_idx(_space_info[id].new_top())); 238 } 239 } 240 241 static void 242 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 243 { 244 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 245 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 246 247 ParallelCompactData& sd = PSParallelCompact::summary_data(); 248 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 249 log_develop_trace(gc, compaction)( 250 REGION_IDX_FORMAT " " 251 REGION_IDX_FORMAT " " PTR_FORMAT " " 252 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 253 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 254 i, dci, p2i(c->destination()), 255 c->partial_obj_size(), c->live_obj_size(), 256 c->data_size(), c->source_region(), c->destination_count()); 257 258 #undef REGION_IDX_FORMAT 259 #undef REGION_DATA_FORMAT 260 } 261 262 void 263 print_generic_summary_data(ParallelCompactData& summary_data, 264 HeapWord* const beg_addr, 265 HeapWord* const end_addr) 266 { 267 size_t total_words = 0; 268 size_t i = summary_data.addr_to_region_idx(beg_addr); 269 const size_t last = summary_data.addr_to_region_idx(end_addr); 270 HeapWord* pdest = 0; 271 272 while (i < last) { 273 ParallelCompactData::RegionData* c = summary_data.region(i); 274 if (c->data_size() != 0 || c->destination() != pdest) { 275 print_generic_summary_region(i, c); 276 total_words += c->data_size(); 277 pdest = c->destination(); 278 } 279 ++i; 280 } 281 282 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 283 } 284 285 void 286 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 287 HeapWord* const beg_addr, 288 HeapWord* const end_addr) { 289 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 290 } 291 292 void 293 print_generic_summary_data(ParallelCompactData& summary_data, 294 SpaceInfo* space_info) 295 { 296 if (!log_develop_is_enabled(Trace, gc, compaction)) { 297 return; 298 } 299 300 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 301 const MutableSpace* space = space_info[id].space(); 302 print_generic_summary_data(summary_data, space->bottom(), 303 MAX2(space->top(), space_info[id].new_top())); 304 } 305 } 306 307 static void 308 print_initial_summary_data(ParallelCompactData& summary_data, 309 const MutableSpace* space) { 310 if (space->top() == space->bottom()) { 311 return; 312 } 313 314 const size_t region_size = ParallelCompactData::RegionSize; 315 typedef ParallelCompactData::RegionData RegionData; 316 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 317 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 318 const RegionData* c = summary_data.region(end_region - 1); 319 HeapWord* end_addr = c->destination() + c->data_size(); 320 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 321 322 // Print (and count) the full regions at the beginning of the space. 323 size_t full_region_count = 0; 324 size_t i = summary_data.addr_to_region_idx(space->bottom()); 325 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 326 ParallelCompactData::RegionData* c = summary_data.region(i); 327 log_develop_trace(gc, compaction)( 328 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 329 i, p2i(c->destination()), 330 c->partial_obj_size(), c->live_obj_size(), 331 c->data_size(), c->source_region(), c->destination_count()); 332 ++full_region_count; 333 ++i; 334 } 335 336 size_t live_to_right = live_in_space - full_region_count * region_size; 337 338 double max_reclaimed_ratio = 0.0; 339 size_t max_reclaimed_ratio_region = 0; 340 size_t max_dead_to_right = 0; 341 size_t max_live_to_right = 0; 342 343 // Print the 'reclaimed ratio' for regions while there is something live in 344 // the region or to the right of it. The remaining regions are empty (and 345 // uninteresting), and computing the ratio will result in division by 0. 346 while (i < end_region && live_to_right > 0) { 347 c = summary_data.region(i); 348 HeapWord* const region_addr = summary_data.region_to_addr(i); 349 const size_t used_to_right = pointer_delta(space->top(), region_addr); 350 const size_t dead_to_right = used_to_right - live_to_right; 351 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 352 353 if (reclaimed_ratio > max_reclaimed_ratio) { 354 max_reclaimed_ratio = reclaimed_ratio; 355 max_reclaimed_ratio_region = i; 356 max_dead_to_right = dead_to_right; 357 max_live_to_right = live_to_right; 358 } 359 360 ParallelCompactData::RegionData* c = summary_data.region(i); 361 log_develop_trace(gc, compaction)( 362 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 363 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 364 i, p2i(c->destination()), 365 c->partial_obj_size(), c->live_obj_size(), 366 c->data_size(), c->source_region(), c->destination_count(), 367 reclaimed_ratio, dead_to_right, live_to_right); 368 369 370 live_to_right -= c->data_size(); 371 ++i; 372 } 373 374 // Any remaining regions are empty. Print one more if there is one. 375 if (i < end_region) { 376 ParallelCompactData::RegionData* c = summary_data.region(i); 377 log_develop_trace(gc, compaction)( 378 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 379 i, p2i(c->destination()), 380 c->partial_obj_size(), c->live_obj_size(), 381 c->data_size(), c->source_region(), c->destination_count()); 382 } 383 384 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 385 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 386 } 387 388 static void 389 print_initial_summary_data(ParallelCompactData& summary_data, 390 SpaceInfo* space_info) { 391 if (!log_develop_is_enabled(Trace, gc, compaction)) { 392 return; 393 } 394 395 unsigned int id = PSParallelCompact::old_space_id; 396 const MutableSpace* space; 397 do { 398 space = space_info[id].space(); 399 print_initial_summary_data(summary_data, space); 400 } while (++id < PSParallelCompact::eden_space_id); 401 402 do { 403 space = space_info[id].space(); 404 print_generic_summary_data(summary_data, space->bottom(), space->top()); 405 } while (++id < PSParallelCompact::last_space_id); 406 } 407 #endif // #ifndef PRODUCT 408 409 ParallelCompactData::ParallelCompactData() : 410 _heap_start(nullptr), 411 DEBUG_ONLY(_heap_end(nullptr) COMMA) 412 _region_vspace(nullptr), 413 _reserved_byte_size(0), 414 _region_data(nullptr), 415 _region_count(0), 416 _block_vspace(nullptr), 417 _block_data(nullptr), 418 _block_count(0) {} 419 420 bool ParallelCompactData::initialize(MemRegion reserved_heap) 421 { 422 _heap_start = reserved_heap.start(); 423 const size_t heap_size = reserved_heap.word_size(); 424 DEBUG_ONLY(_heap_end = _heap_start + heap_size;) 425 426 assert(region_align_down(_heap_start) == _heap_start, 427 "region start not aligned"); 428 429 bool result = initialize_region_data(heap_size) && initialize_block_data(); 430 return result; 431 } 432 433 PSVirtualSpace* 434 ParallelCompactData::create_vspace(size_t count, size_t element_size) 435 { 436 const size_t raw_bytes = count * element_size; 437 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 438 const size_t granularity = os::vm_allocation_granularity(); 439 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 440 441 const size_t rs_align = page_sz == os::vm_page_size() ? 0 : 442 MAX2(page_sz, granularity); 443 ReservedSpace rs(_reserved_byte_size, rs_align, page_sz); 444 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(), 445 rs.size(), page_sz); 446 447 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 448 449 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 450 if (vspace != 0) { 451 if (vspace->expand_by(_reserved_byte_size)) { 452 return vspace; 453 } 454 delete vspace; 455 // Release memory reserved in the space. 456 rs.release(); 457 } 458 459 return 0; 460 } 461 462 bool ParallelCompactData::initialize_region_data(size_t heap_size) 463 { 464 assert(is_aligned(heap_size, RegionSize), "precondition"); 465 466 const size_t count = heap_size >> Log2RegionSize; 467 _region_vspace = create_vspace(count, sizeof(RegionData)); 468 if (_region_vspace != 0) { 469 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 470 _region_count = count; 471 return true; 472 } 473 return false; 474 } 475 476 bool ParallelCompactData::initialize_block_data() 477 { 478 assert(_region_count != 0, "region data must be initialized first"); 479 const size_t count = _region_count << Log2BlocksPerRegion; 480 _block_vspace = create_vspace(count, sizeof(BlockData)); 481 if (_block_vspace != 0) { 482 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 483 _block_count = count; 484 return true; 485 } 486 return false; 487 } 488 489 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 490 assert(beg_region <= _region_count, "beg_region out of range"); 491 assert(end_region <= _region_count, "end_region out of range"); 492 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 493 494 const size_t region_cnt = end_region - beg_region; 495 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 496 497 const size_t beg_block = beg_region * BlocksPerRegion; 498 const size_t block_cnt = region_cnt * BlocksPerRegion; 499 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 500 } 501 502 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 503 { 504 const RegionData* cur_cp = region(region_idx); 505 const RegionData* const end_cp = region(region_count() - 1); 506 507 HeapWord* result = region_to_addr(region_idx); 508 if (cur_cp < end_cp) { 509 do { 510 result += cur_cp->partial_obj_size(); 511 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 512 } 513 return result; 514 } 515 516 void 517 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 518 { 519 assert(is_region_aligned(beg), "not RegionSize aligned"); 520 assert(is_region_aligned(end), "not RegionSize aligned"); 521 522 size_t cur_region = addr_to_region_idx(beg); 523 const size_t end_region = addr_to_region_idx(end); 524 HeapWord* addr = beg; 525 while (cur_region < end_region) { 526 _region_data[cur_region].set_destination(addr); 527 _region_data[cur_region].set_destination_count(0); 528 _region_data[cur_region].set_source_region(cur_region); 529 530 // Update live_obj_size so the region appears completely full. 531 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 532 _region_data[cur_region].set_live_obj_size(live_size); 533 534 ++cur_region; 535 addr += RegionSize; 536 } 537 } 538 539 // Find the point at which a space can be split and, if necessary, record the 540 // split point. 541 // 542 // If the current src region (which overflowed the destination space) doesn't 543 // have a partial object, the split point is at the beginning of the current src 544 // region (an "easy" split, no extra bookkeeping required). 545 // 546 // If the current src region has a partial object, the split point is in the 547 // region where that partial object starts (call it the split_region). If 548 // split_region has a partial object, then the split point is just after that 549 // partial object (a "hard" split where we have to record the split data and 550 // zero the partial_obj_size field). With a "hard" split, we know that the 551 // partial_obj ends within split_region because the partial object that caused 552 // the overflow starts in split_region. If split_region doesn't have a partial 553 // obj, then the split is at the beginning of split_region (another "easy" 554 // split). 555 HeapWord* 556 ParallelCompactData::summarize_split_space(size_t src_region, 557 SplitInfo& split_info, 558 HeapWord* destination, 559 HeapWord* target_end, 560 HeapWord** target_next) 561 { 562 assert(destination <= target_end, "sanity"); 563 assert(destination + _region_data[src_region].data_size() > target_end, 564 "region should not fit into target space"); 565 assert(is_region_aligned(target_end), "sanity"); 566 567 size_t split_region = src_region; 568 HeapWord* split_destination = destination; 569 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 570 571 if (destination + partial_obj_size > target_end) { 572 // The split point is just after the partial object (if any) in the 573 // src_region that contains the start of the object that overflowed the 574 // destination space. 575 // 576 // Find the start of the "overflow" object and set split_region to the 577 // region containing it. 578 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 579 split_region = addr_to_region_idx(overflow_obj); 580 581 // Clear the source_region field of all destination regions whose first word 582 // came from data after the split point (a non-null source_region field 583 // implies a region must be filled). 584 // 585 // An alternative to the simple loop below: clear during post_compact(), 586 // which uses memcpy instead of individual stores, and is easy to 587 // parallelize. (The downside is that it clears the entire RegionData 588 // object as opposed to just one field.) 589 // 590 // post_compact() would have to clear the summary data up to the highest 591 // address that was written during the summary phase, which would be 592 // 593 // max(top, max(new_top, clear_top)) 594 // 595 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 596 // to target_end. 597 const RegionData* const sr = region(split_region); 598 const size_t beg_idx = 599 addr_to_region_idx(region_align_up(sr->destination() + 600 sr->partial_obj_size())); 601 const size_t end_idx = addr_to_region_idx(target_end); 602 603 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 604 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 605 _region_data[idx].set_source_region(0); 606 } 607 608 // Set split_destination and partial_obj_size to reflect the split region. 609 split_destination = sr->destination(); 610 partial_obj_size = sr->partial_obj_size(); 611 } 612 613 // The split is recorded only if a partial object extends onto the region. 614 if (partial_obj_size != 0) { 615 _region_data[split_region].set_partial_obj_size(0); 616 split_info.record(split_region, partial_obj_size, split_destination); 617 } 618 619 // Setup the continuation addresses. 620 *target_next = split_destination + partial_obj_size; 621 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 622 623 if (log_develop_is_enabled(Trace, gc, compaction)) { 624 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 625 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 626 split_type, p2i(source_next), split_region, partial_obj_size); 627 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 628 split_type, p2i(split_destination), 629 addr_to_region_idx(split_destination), 630 p2i(*target_next)); 631 632 if (partial_obj_size != 0) { 633 HeapWord* const po_beg = split_info.destination(); 634 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 635 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 636 split_type, 637 p2i(po_beg), addr_to_region_idx(po_beg), 638 p2i(po_end), addr_to_region_idx(po_end)); 639 } 640 } 641 642 return source_next; 643 } 644 645 bool ParallelCompactData::summarize(SplitInfo& split_info, 646 HeapWord* source_beg, HeapWord* source_end, 647 HeapWord** source_next, 648 HeapWord* target_beg, HeapWord* target_end, 649 HeapWord** target_next) 650 { 651 HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next; 652 log_develop_trace(gc, compaction)( 653 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 654 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 655 p2i(source_beg), p2i(source_end), p2i(source_next_val), 656 p2i(target_beg), p2i(target_end), p2i(*target_next)); 657 658 size_t cur_region = addr_to_region_idx(source_beg); 659 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 660 661 HeapWord *dest_addr = target_beg; 662 while (cur_region < end_region) { 663 // The destination must be set even if the region has no data. 664 _region_data[cur_region].set_destination(dest_addr); 665 666 size_t words = _region_data[cur_region].data_size(); 667 if (words > 0) { 668 // If cur_region does not fit entirely into the target space, find a point 669 // at which the source space can be 'split' so that part is copied to the 670 // target space and the rest is copied elsewhere. 671 if (dest_addr + words > target_end) { 672 assert(source_next != nullptr, "source_next is null when splitting"); 673 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 674 target_end, target_next); 675 return false; 676 } 677 678 // Compute the destination_count for cur_region, and if necessary, update 679 // source_region for a destination region. The source_region field is 680 // updated if cur_region is the first (left-most) region to be copied to a 681 // destination region. 682 // 683 // The destination_count calculation is a bit subtle. A region that has 684 // data that compacts into itself does not count itself as a destination. 685 // This maintains the invariant that a zero count means the region is 686 // available and can be claimed and then filled. 687 uint destination_count = 0; 688 if (split_info.is_split(cur_region)) { 689 // The current region has been split: the partial object will be copied 690 // to one destination space and the remaining data will be copied to 691 // another destination space. Adjust the initial destination_count and, 692 // if necessary, set the source_region field if the partial object will 693 // cross a destination region boundary. 694 destination_count = split_info.destination_count(); 695 if (destination_count == 2) { 696 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 697 _region_data[dest_idx].set_source_region(cur_region); 698 } 699 } 700 701 HeapWord* const last_addr = dest_addr + words - 1; 702 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 703 const size_t dest_region_2 = addr_to_region_idx(last_addr); 704 705 // Initially assume that the destination regions will be the same and 706 // adjust the value below if necessary. Under this assumption, if 707 // cur_region == dest_region_2, then cur_region will be compacted 708 // completely into itself. 709 destination_count += cur_region == dest_region_2 ? 0 : 1; 710 if (dest_region_1 != dest_region_2) { 711 // Destination regions differ; adjust destination_count. 712 destination_count += 1; 713 // Data from cur_region will be copied to the start of dest_region_2. 714 _region_data[dest_region_2].set_source_region(cur_region); 715 } else if (is_region_aligned(dest_addr)) { 716 // Data from cur_region will be copied to the start of the destination 717 // region. 718 _region_data[dest_region_1].set_source_region(cur_region); 719 } 720 721 _region_data[cur_region].set_destination_count(destination_count); 722 dest_addr += words; 723 } 724 725 ++cur_region; 726 } 727 728 *target_next = dest_addr; 729 return true; 730 } 731 732 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const { 733 assert(addr != nullptr, "Should detect null oop earlier"); 734 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 735 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 736 737 // Region covering the object. 738 RegionData* const region_ptr = addr_to_region_ptr(addr); 739 HeapWord* result = region_ptr->destination(); 740 741 // If the entire Region is live, the new location is region->destination + the 742 // offset of the object within in the Region. 743 744 // Run some performance tests to determine if this special case pays off. It 745 // is worth it for pointers into the dense prefix. If the optimization to 746 // avoid pointer updates in regions that only point to the dense prefix is 747 // ever implemented, this should be revisited. 748 if (region_ptr->data_size() == RegionSize) { 749 result += region_offset(addr); 750 return result; 751 } 752 753 // Otherwise, the new location is region->destination + block offset + the 754 // number of live words in the Block that are (a) to the left of addr and (b) 755 // due to objects that start in the Block. 756 757 // Fill in the block table if necessary. This is unsynchronized, so multiple 758 // threads may fill the block table for a region (harmless, since it is 759 // idempotent). 760 if (!region_ptr->blocks_filled()) { 761 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 762 region_ptr->set_blocks_filled(); 763 } 764 765 HeapWord* const search_start = block_align_down(addr); 766 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 767 768 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 769 const size_t live = bitmap->live_words_in_range(cm, search_start, cast_to_oop(addr)); 770 result += block_offset + live; 771 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 772 return result; 773 } 774 775 #ifdef ASSERT 776 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 777 { 778 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 779 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 780 for (const size_t* p = beg; p < end; ++p) { 781 assert(*p == 0, "not zero"); 782 } 783 } 784 785 void ParallelCompactData::verify_clear() 786 { 787 verify_clear(_region_vspace); 788 verify_clear(_block_vspace); 789 } 790 #endif // #ifdef ASSERT 791 792 STWGCTimer PSParallelCompact::_gc_timer; 793 ParallelOldTracer PSParallelCompact::_gc_tracer; 794 elapsedTimer PSParallelCompact::_accumulated_time; 795 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 796 CollectorCounters* PSParallelCompact::_counters = nullptr; 797 ParMarkBitMap PSParallelCompact::_mark_bitmap; 798 ParallelCompactData PSParallelCompact::_summary_data; 799 800 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 801 802 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 803 804 void PSParallelCompact::post_initialize() { 805 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 806 _span_based_discoverer.set_span(heap->reserved_region()); 807 _ref_processor = 808 new ReferenceProcessor(&_span_based_discoverer, 809 ParallelGCThreads, // mt processing degree 810 ParallelGCThreads, // mt discovery degree 811 false, // concurrent_discovery 812 &_is_alive_closure); // non-header is alive closure 813 814 _counters = new CollectorCounters("Parallel full collection pauses", 1); 815 816 // Initialize static fields in ParCompactionManager. 817 ParCompactionManager::initialize(mark_bitmap()); 818 } 819 820 bool PSParallelCompact::initialize_aux_data() { 821 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 822 MemRegion mr = heap->reserved_region(); 823 assert(mr.byte_size() != 0, "heap should be reserved"); 824 825 initialize_space_info(); 826 827 if (!_mark_bitmap.initialize(mr)) { 828 vm_shutdown_during_initialization( 829 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 830 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 831 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 832 return false; 833 } 834 835 if (!_summary_data.initialize(mr)) { 836 vm_shutdown_during_initialization( 837 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 838 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 839 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 840 return false; 841 } 842 843 return true; 844 } 845 846 void PSParallelCompact::initialize_space_info() 847 { 848 memset(&_space_info, 0, sizeof(_space_info)); 849 850 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 851 PSYoungGen* young_gen = heap->young_gen(); 852 853 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 854 _space_info[eden_space_id].set_space(young_gen->eden_space()); 855 _space_info[from_space_id].set_space(young_gen->from_space()); 856 _space_info[to_space_id].set_space(young_gen->to_space()); 857 858 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 859 } 860 861 void 862 PSParallelCompact::clear_data_covering_space(SpaceId id) 863 { 864 // At this point, top is the value before GC, new_top() is the value that will 865 // be set at the end of GC. The marking bitmap is cleared to top; nothing 866 // should be marked above top. The summary data is cleared to the larger of 867 // top & new_top. 868 MutableSpace* const space = _space_info[id].space(); 869 HeapWord* const bot = space->bottom(); 870 HeapWord* const top = space->top(); 871 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 872 873 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 874 const idx_t end_bit = _mark_bitmap.addr_to_bit(top); 875 _mark_bitmap.clear_range(beg_bit, end_bit); 876 877 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 878 const size_t end_region = 879 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 880 _summary_data.clear_range(beg_region, end_region); 881 882 // Clear the data used to 'split' regions. 883 SplitInfo& split_info = _space_info[id].split_info(); 884 if (split_info.is_valid()) { 885 split_info.clear(); 886 } 887 DEBUG_ONLY(split_info.verify_clear();) 888 } 889 890 void PSParallelCompact::pre_compact() 891 { 892 // Update the from & to space pointers in space_info, since they are swapped 893 // at each young gen gc. Do the update unconditionally (even though a 894 // promotion failure does not swap spaces) because an unknown number of young 895 // collections will have swapped the spaces an unknown number of times. 896 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 897 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 898 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 899 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 900 901 // Increment the invocation count 902 heap->increment_total_collections(true); 903 904 CodeCache::on_gc_marking_cycle_start(); 905 906 heap->print_heap_before_gc(); 907 heap->trace_heap_before_gc(&_gc_tracer); 908 909 // Fill in TLABs 910 heap->ensure_parsability(true); // retire TLABs 911 912 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 913 Universe::verify("Before GC"); 914 } 915 916 DEBUG_ONLY(mark_bitmap()->verify_clear();) 917 DEBUG_ONLY(summary_data().verify_clear();) 918 919 ParCompactionManager::reset_all_bitmap_query_caches(); 920 } 921 922 void PSParallelCompact::post_compact() 923 { 924 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 925 ParCompactionManager::remove_all_shadow_regions(); 926 927 CodeCache::on_gc_marking_cycle_finish(); 928 CodeCache::arm_all_nmethods(); 929 930 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 931 // Clear the marking bitmap, summary data and split info. 932 clear_data_covering_space(SpaceId(id)); 933 // Update top(). Must be done after clearing the bitmap and summary data. 934 _space_info[id].publish_new_top(); 935 } 936 937 ParCompactionManager::flush_all_string_dedup_requests(); 938 939 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 940 MutableSpace* const from_space = _space_info[from_space_id].space(); 941 MutableSpace* const to_space = _space_info[to_space_id].space(); 942 943 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 944 bool eden_empty = eden_space->is_empty(); 945 946 // Update heap occupancy information which is used as input to the soft ref 947 // clearing policy at the next gc. 948 Universe::heap()->update_capacity_and_used_at_gc(); 949 950 bool young_gen_empty = eden_empty && from_space->is_empty() && 951 to_space->is_empty(); 952 953 PSCardTable* ct = heap->card_table(); 954 MemRegion old_mr = heap->old_gen()->committed(); 955 if (young_gen_empty) { 956 ct->clear_MemRegion(old_mr); 957 } else { 958 ct->dirty_MemRegion(old_mr); 959 } 960 961 { 962 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 963 GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer()); 964 ClassLoaderDataGraph::purge(true /* at_safepoint */); 965 DEBUG_ONLY(MetaspaceUtils::verify();) 966 } 967 968 // Need to clear claim bits for the next mark. 969 ClassLoaderDataGraph::clear_claimed_marks(); 970 971 heap->prune_scavengable_nmethods(); 972 973 #if COMPILER2_OR_JVMCI 974 DerivedPointerTable::update_pointers(); 975 #endif 976 977 if (ZapUnusedHeapArea) { 978 heap->gen_mangle_unused_area(); 979 } 980 981 // Signal that we have completed a visit to all live objects. 982 Universe::heap()->record_whole_heap_examined_timestamp(); 983 } 984 985 ParallelCompactData::RegionData* 986 PSParallelCompact::first_dead_space_region(const RegionData* beg, 987 const RegionData* end) 988 { 989 const size_t region_size = ParallelCompactData::RegionSize; 990 ParallelCompactData& sd = summary_data(); 991 size_t left = sd.region(beg); 992 size_t right = end > beg ? sd.region(end) - 1 : left; 993 994 // Binary search. 995 while (left < right) { 996 // Equivalent to (left + right) / 2, but does not overflow. 997 const size_t middle = left + (right - left) / 2; 998 RegionData* const middle_ptr = sd.region(middle); 999 HeapWord* const dest = middle_ptr->destination(); 1000 HeapWord* const addr = sd.region_to_addr(middle); 1001 assert(dest != nullptr, "sanity"); 1002 assert(dest <= addr, "must move left"); 1003 1004 if (middle > left && dest < addr) { 1005 right = middle - 1; 1006 } else if (middle < right && middle_ptr->data_size() == region_size) { 1007 left = middle + 1; 1008 } else { 1009 return middle_ptr; 1010 } 1011 } 1012 return sd.region(left); 1013 } 1014 1015 // Return the address of the end of the dense prefix, a.k.a. the start of the 1016 // compacted region. The address is always on a region boundary. 1017 // 1018 // Completely full regions at the left are skipped, since no compaction can 1019 // occur in those regions. Then the maximum amount of dead wood to allow is 1020 // computed, based on the density (amount live / capacity) of the generation; 1021 // the region with approximately that amount of dead space to the left is 1022 // identified as the limit region. Regions between the last completely full 1023 // region and the limit region are scanned and the one that has the best 1024 // (maximum) reclaimed_ratio() is selected. 1025 HeapWord* 1026 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1027 bool maximum_compaction) 1028 { 1029 const size_t region_size = ParallelCompactData::RegionSize; 1030 const ParallelCompactData& sd = summary_data(); 1031 1032 const MutableSpace* const space = _space_info[id].space(); 1033 HeapWord* const top = space->top(); 1034 HeapWord* const top_aligned_up = sd.region_align_up(top); 1035 HeapWord* const new_top = _space_info[id].new_top(); 1036 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1037 HeapWord* const bottom = space->bottom(); 1038 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1039 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1040 const RegionData* const new_top_cp = 1041 sd.addr_to_region_ptr(new_top_aligned_up); 1042 1043 // Skip full regions at the beginning of the space--they are necessarily part 1044 // of the dense prefix. 1045 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1046 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1047 space->is_empty(), "no dead space allowed to the left"); 1048 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1049 "region must have dead space"); 1050 1051 // The gc number is saved whenever a maximum compaction is done, and used to 1052 // determine when the maximum compaction interval has expired. This avoids 1053 // successive max compactions for different reasons. 1054 const uint total_invocations = ParallelScavengeHeap::heap()->total_full_collections(); 1055 assert(total_invocations >= _maximum_compaction_gc_num, "sanity"); 1056 const size_t gcs_since_max = total_invocations - _maximum_compaction_gc_num; 1057 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1058 total_invocations == HeapFirstMaximumCompactionCount; 1059 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1060 _maximum_compaction_gc_num = total_invocations; 1061 return sd.region_to_addr(full_cp); 1062 } 1063 1064 // Iteration starts with the region *after* the full-region-prefix-end. 1065 const RegionData* const start_region = full_cp; 1066 // If final region is not full, iteration stops before that region, 1067 // because fill_dense_prefix_end assumes that prefix_end <= top. 1068 const RegionData* const end_region = sd.addr_to_region_ptr(space->top()); 1069 assert(start_region <= end_region, "inv"); 1070 1071 size_t max_waste = space->capacity_in_words() * (MarkSweepDeadRatio / 100.0); 1072 const RegionData* cur_region = start_region; 1073 for (/* empty */; cur_region < end_region; ++cur_region) { 1074 assert(region_size >= cur_region->data_size(), "inv"); 1075 size_t dead_size = region_size - cur_region->data_size(); 1076 if (max_waste < dead_size) { 1077 break; 1078 } 1079 max_waste -= dead_size; 1080 } 1081 1082 HeapWord* const prefix_end = sd.region_to_addr(cur_region); 1083 assert(sd.is_region_aligned(prefix_end), "postcondition"); 1084 assert(prefix_end >= sd.region_to_addr(full_cp), "in-range"); 1085 assert(prefix_end <= space->top(), "in-range"); 1086 return prefix_end; 1087 } 1088 1089 void PSParallelCompact::summarize_spaces_quick() 1090 { 1091 for (unsigned int i = 0; i < last_space_id; ++i) { 1092 const MutableSpace* space = _space_info[i].space(); 1093 HeapWord** nta = _space_info[i].new_top_addr(); 1094 bool result = _summary_data.summarize(_space_info[i].split_info(), 1095 space->bottom(), space->top(), nullptr, 1096 space->bottom(), space->end(), nta); 1097 assert(result, "space must fit into itself"); 1098 _space_info[i].set_dense_prefix(space->bottom()); 1099 } 1100 } 1101 1102 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) { 1103 // Comparing two sizes to decide if filling is required: 1104 // 1105 // The size of the filler (min-obj-size) is 2 heap words with the default 1106 // MinObjAlignment, since both markword and klass take 1 heap word. 1107 // 1108 // The size of the gap (if any) right before dense-prefix-end is 1109 // MinObjAlignment. 1110 // 1111 // Need to fill in the gap only if it's smaller than min-obj-size, and the 1112 // filler obj will extend to next region. 1113 1114 // Note: If min-fill-size decreases to 1, this whole method becomes redundant. 1115 if (UseCompactObjectHeaders) { 1116 // The gap is always equal to min-fill-size, so nothing to do. 1117 return; 1118 } 1119 assert(CollectedHeap::min_fill_size() >= 2, "inv"); 1120 #ifndef _LP64 1121 // In 32-bit system, each heap word is 4 bytes, so MinObjAlignment == 2. 1122 // The gap is always equal to min-fill-size, so nothing to do. 1123 return; 1124 #endif 1125 if (MinObjAlignment > 1) { 1126 return; 1127 } 1128 assert(CollectedHeap::min_fill_size() == 2, "inv"); 1129 HeapWord* const dense_prefix_end = dense_prefix(id); 1130 RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end); 1131 idx_t const dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1132 1133 if (region_after_dense_prefix->partial_obj_size() != 0 || 1134 _mark_bitmap.is_obj_beg(dense_prefix_bit)) { 1135 // The region after the dense prefix starts with live bytes. 1136 return; 1137 } 1138 1139 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1140 // There is exactly one heap word gap right before the dense prefix end, so we need a filler object. 1141 // The filler object will extend into the region after the last dense prefix region. 1142 const size_t obj_len = 2; // min-fill-size 1143 HeapWord* const obj_beg = dense_prefix_end - 1; 1144 CollectedHeap::fill_with_object(obj_beg, obj_len); 1145 _mark_bitmap.mark_obj(obj_beg, obj_len); 1146 _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1); 1147 region_after_dense_prefix->set_partial_obj_size(1); 1148 region_after_dense_prefix->set_partial_obj_addr(obj_beg); 1149 assert(start_array(id) != nullptr, "sanity"); 1150 start_array(id)->update_for_block(obj_beg, obj_beg + obj_len); 1151 } 1152 } 1153 1154 void 1155 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1156 { 1157 assert(id < last_space_id, "id out of range"); 1158 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1159 "should have been reset in summarize_spaces_quick()"); 1160 1161 const MutableSpace* space = _space_info[id].space(); 1162 if (_space_info[id].new_top() != space->bottom()) { 1163 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1164 _space_info[id].set_dense_prefix(dense_prefix_end); 1165 1166 // Recompute the summary data, taking into account the dense prefix. If 1167 // every last byte will be reclaimed, then the existing summary data which 1168 // compacts everything can be left in place. 1169 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1170 // If dead space crosses the dense prefix boundary, it is (at least 1171 // partially) filled with a dummy object, marked live and added to the 1172 // summary data. This simplifies the copy/update phase and must be done 1173 // before the final locations of objects are determined, to prevent 1174 // leaving a fragment of dead space that is too small to fill. 1175 fill_dense_prefix_end(id); 1176 1177 // Compute the destination of each Region, and thus each object. 1178 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1179 _summary_data.summarize(_space_info[id].split_info(), 1180 dense_prefix_end, space->top(), nullptr, 1181 dense_prefix_end, space->end(), 1182 _space_info[id].new_top_addr()); 1183 } 1184 } 1185 1186 if (log_develop_is_enabled(Trace, gc, compaction)) { 1187 const size_t region_size = ParallelCompactData::RegionSize; 1188 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1189 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1190 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1191 HeapWord* const new_top = _space_info[id].new_top(); 1192 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1193 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1194 log_develop_trace(gc, compaction)( 1195 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1196 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1197 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1198 id, space->capacity_in_words(), p2i(dense_prefix_end), 1199 dp_region, dp_words / region_size, 1200 cr_words / region_size, p2i(new_top)); 1201 } 1202 } 1203 1204 #ifndef PRODUCT 1205 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1206 HeapWord* dst_beg, HeapWord* dst_end, 1207 SpaceId src_space_id, 1208 HeapWord* src_beg, HeapWord* src_end) 1209 { 1210 log_develop_trace(gc, compaction)( 1211 "Summarizing %d [%s] into %d [%s]: " 1212 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1213 SIZE_FORMAT "-" SIZE_FORMAT " " 1214 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1215 SIZE_FORMAT "-" SIZE_FORMAT, 1216 src_space_id, space_names[src_space_id], 1217 dst_space_id, space_names[dst_space_id], 1218 p2i(src_beg), p2i(src_end), 1219 _summary_data.addr_to_region_idx(src_beg), 1220 _summary_data.addr_to_region_idx(src_end), 1221 p2i(dst_beg), p2i(dst_end), 1222 _summary_data.addr_to_region_idx(dst_beg), 1223 _summary_data.addr_to_region_idx(dst_end)); 1224 } 1225 #endif // #ifndef PRODUCT 1226 1227 void PSParallelCompact::summary_phase(bool maximum_compaction) 1228 { 1229 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1230 1231 // Quick summarization of each space into itself, to see how much is live. 1232 summarize_spaces_quick(); 1233 1234 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1235 NOT_PRODUCT(print_region_ranges()); 1236 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1237 1238 // The amount of live data that will end up in old space (assuming it fits). 1239 size_t old_space_total_live = 0; 1240 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1241 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1242 _space_info[id].space()->bottom()); 1243 } 1244 1245 MutableSpace* const old_space = _space_info[old_space_id].space(); 1246 const size_t old_capacity = old_space->capacity_in_words(); 1247 if (old_space_total_live > old_capacity) { 1248 // XXX - should also try to expand 1249 maximum_compaction = true; 1250 } 1251 1252 // Old generations. 1253 summarize_space(old_space_id, maximum_compaction); 1254 1255 // Summarize the remaining spaces in the young gen. The initial target space 1256 // is the old gen. If a space does not fit entirely into the target, then the 1257 // remainder is compacted into the space itself and that space becomes the new 1258 // target. 1259 SpaceId dst_space_id = old_space_id; 1260 HeapWord* dst_space_end = old_space->end(); 1261 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1262 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1263 const MutableSpace* space = _space_info[id].space(); 1264 const size_t live = pointer_delta(_space_info[id].new_top(), 1265 space->bottom()); 1266 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1267 1268 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1269 SpaceId(id), space->bottom(), space->top());) 1270 if (live > 0 && live <= available) { 1271 // All the live data will fit. 1272 bool done = _summary_data.summarize(_space_info[id].split_info(), 1273 space->bottom(), space->top(), 1274 nullptr, 1275 *new_top_addr, dst_space_end, 1276 new_top_addr); 1277 assert(done, "space must fit into old gen"); 1278 1279 // Reset the new_top value for the space. 1280 _space_info[id].set_new_top(space->bottom()); 1281 } else if (live > 0) { 1282 // Attempt to fit part of the source space into the target space. 1283 HeapWord* next_src_addr = nullptr; 1284 bool done = _summary_data.summarize(_space_info[id].split_info(), 1285 space->bottom(), space->top(), 1286 &next_src_addr, 1287 *new_top_addr, dst_space_end, 1288 new_top_addr); 1289 assert(!done, "space should not fit into old gen"); 1290 assert(next_src_addr != nullptr, "sanity"); 1291 1292 // The source space becomes the new target, so the remainder is compacted 1293 // within the space itself. 1294 dst_space_id = SpaceId(id); 1295 dst_space_end = space->end(); 1296 new_top_addr = _space_info[id].new_top_addr(); 1297 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1298 space->bottom(), dst_space_end, 1299 SpaceId(id), next_src_addr, space->top());) 1300 done = _summary_data.summarize(_space_info[id].split_info(), 1301 next_src_addr, space->top(), 1302 nullptr, 1303 space->bottom(), dst_space_end, 1304 new_top_addr); 1305 assert(done, "space must fit when compacted into itself"); 1306 assert(*new_top_addr <= space->top(), "usage should not grow"); 1307 } 1308 } 1309 1310 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1311 NOT_PRODUCT(print_region_ranges()); 1312 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1313 } 1314 1315 // This method should contain all heap-specific policy for invoking a full 1316 // collection. invoke_no_policy() will only attempt to compact the heap; it 1317 // will do nothing further. If we need to bail out for policy reasons, scavenge 1318 // before full gc, or any other specialized behavior, it needs to be added here. 1319 // 1320 // Note that this method should only be called from the vm_thread while at a 1321 // safepoint. 1322 // 1323 // Note that the all_soft_refs_clear flag in the soft ref policy 1324 // may be true because this method can be called without intervening 1325 // activity. For example when the heap space is tight and full measure 1326 // are being taken to free space. 1327 bool PSParallelCompact::invoke(bool maximum_heap_compaction) { 1328 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1329 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1330 "should be in vm thread"); 1331 1332 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1333 assert(!heap->is_gc_active(), "not reentrant"); 1334 1335 IsGCActiveMark mark; 1336 1337 if (ScavengeBeforeFullGC) { 1338 PSScavenge::invoke_no_policy(); 1339 } 1340 1341 const bool clear_all_soft_refs = 1342 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1343 1344 return PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1345 maximum_heap_compaction); 1346 } 1347 1348 // This method contains no policy. You should probably 1349 // be calling invoke() instead. 1350 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1351 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1352 assert(ref_processor() != nullptr, "Sanity"); 1353 1354 if (GCLocker::check_active_before_gc()) { 1355 return false; 1356 } 1357 1358 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1359 1360 GCIdMark gc_id_mark; 1361 _gc_timer.register_gc_start(); 1362 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1363 1364 GCCause::Cause gc_cause = heap->gc_cause(); 1365 PSYoungGen* young_gen = heap->young_gen(); 1366 PSOldGen* old_gen = heap->old_gen(); 1367 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1368 1369 // The scope of casr should end after code that can change 1370 // SoftRefPolicy::_should_clear_all_soft_refs. 1371 ClearedAllSoftRefs casr(maximum_heap_compaction, 1372 heap->soft_ref_policy()); 1373 1374 if (ZapUnusedHeapArea) { 1375 // Save information needed to minimize mangling 1376 heap->record_gen_tops_before_GC(); 1377 } 1378 1379 // Make sure data structures are sane, make the heap parsable, and do other 1380 // miscellaneous bookkeeping. 1381 pre_compact(); 1382 1383 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values(); 1384 1385 { 1386 const uint active_workers = 1387 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(), 1388 ParallelScavengeHeap::heap()->workers().active_workers(), 1389 Threads::number_of_non_daemon_threads()); 1390 ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers); 1391 1392 GCTraceCPUTime tcpu(&_gc_tracer); 1393 GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true); 1394 1395 heap->pre_full_gc_dump(&_gc_timer); 1396 1397 TraceCollectorStats tcs(counters()); 1398 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC"); 1399 1400 if (log_is_enabled(Debug, gc, heap, exit)) { 1401 accumulated_time()->start(); 1402 } 1403 1404 // Let the size policy know we're starting 1405 size_policy->major_collection_begin(); 1406 1407 #if COMPILER2_OR_JVMCI 1408 DerivedPointerTable::clear(); 1409 #endif 1410 1411 ref_processor()->start_discovery(maximum_heap_compaction); 1412 1413 ClassUnloadingContext ctx(1 /* num_nmethod_unlink_workers */, 1414 false /* unregister_nmethods_during_purge */, 1415 false /* lock_nmethod_free_separately */); 1416 1417 marking_phase(&_gc_tracer); 1418 1419 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1420 && GCCause::is_user_requested_gc(gc_cause); 1421 summary_phase(maximum_heap_compaction || max_on_system_gc); 1422 1423 #if COMPILER2_OR_JVMCI 1424 assert(DerivedPointerTable::is_active(), "Sanity"); 1425 DerivedPointerTable::set_active(false); 1426 #endif 1427 1428 // adjust_roots() updates Universe::_intArrayKlass which is 1429 // needed by the compaction for filling holes in the dense prefix. 1430 adjust_roots(); 1431 1432 compact(); 1433 1434 ParCompactionManager::verify_all_region_stack_empty(); 1435 1436 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1437 // done before resizing. 1438 post_compact(); 1439 1440 // Let the size policy know we're done 1441 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1442 1443 if (UseAdaptiveSizePolicy) { 1444 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1445 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1446 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1447 1448 // Don't check if the size_policy is ready here. Let 1449 // the size_policy check that internally. 1450 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1451 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1452 // Swap the survivor spaces if from_space is empty. The 1453 // resize_young_gen() called below is normally used after 1454 // a successful young GC and swapping of survivor spaces; 1455 // otherwise, it will fail to resize the young gen with 1456 // the current implementation. 1457 if (young_gen->from_space()->is_empty()) { 1458 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1459 young_gen->swap_spaces(); 1460 } 1461 1462 // Calculate optimal free space amounts 1463 assert(young_gen->max_gen_size() > 1464 young_gen->from_space()->capacity_in_bytes() + 1465 young_gen->to_space()->capacity_in_bytes(), 1466 "Sizes of space in young gen are out-of-bounds"); 1467 1468 size_t young_live = young_gen->used_in_bytes(); 1469 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1470 size_t old_live = old_gen->used_in_bytes(); 1471 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1472 size_t max_old_gen_size = old_gen->max_gen_size(); 1473 size_t max_eden_size = young_gen->max_gen_size() - 1474 young_gen->from_space()->capacity_in_bytes() - 1475 young_gen->to_space()->capacity_in_bytes(); 1476 1477 // Used for diagnostics 1478 size_policy->clear_generation_free_space_flags(); 1479 1480 size_policy->compute_generations_free_space(young_live, 1481 eden_live, 1482 old_live, 1483 cur_eden, 1484 max_old_gen_size, 1485 max_eden_size, 1486 true /* full gc*/); 1487 1488 size_policy->check_gc_overhead_limit(eden_live, 1489 max_old_gen_size, 1490 max_eden_size, 1491 true /* full gc*/, 1492 gc_cause, 1493 heap->soft_ref_policy()); 1494 1495 size_policy->decay_supplemental_growth(true /* full gc*/); 1496 1497 heap->resize_old_gen( 1498 size_policy->calculated_old_free_size_in_bytes()); 1499 1500 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1501 size_policy->calculated_survivor_size_in_bytes()); 1502 } 1503 1504 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1505 } 1506 1507 if (UsePerfData) { 1508 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1509 counters->update_counters(); 1510 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1511 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1512 } 1513 1514 heap->resize_all_tlabs(); 1515 1516 // Resize the metaspace capacity after a collection 1517 MetaspaceGC::compute_new_size(); 1518 1519 if (log_is_enabled(Debug, gc, heap, exit)) { 1520 accumulated_time()->stop(); 1521 } 1522 1523 heap->print_heap_change(pre_gc_values); 1524 1525 // Track memory usage and detect low memory 1526 MemoryService::track_memory_usage(); 1527 heap->update_counters(); 1528 1529 heap->post_full_gc_dump(&_gc_timer); 1530 } 1531 1532 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1533 Universe::verify("After GC"); 1534 } 1535 1536 if (ZapUnusedHeapArea) { 1537 old_gen->object_space()->check_mangled_unused_area_complete(); 1538 } 1539 1540 heap->print_heap_after_gc(); 1541 heap->trace_heap_after_gc(&_gc_tracer); 1542 1543 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1544 1545 _gc_timer.register_gc_end(); 1546 1547 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1548 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1549 1550 return true; 1551 } 1552 1553 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 1554 private: 1555 uint _worker_id; 1556 1557 public: 1558 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { } 1559 void do_thread(Thread* thread) { 1560 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 1561 1562 ResourceMark rm; 1563 1564 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id); 1565 1566 PCMarkAndPushClosure mark_and_push_closure(cm); 1567 MarkingNMethodClosure mark_and_push_in_blobs(&mark_and_push_closure, !NMethodToOopClosure::FixRelocations, true /* keepalive nmethods */); 1568 1569 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs); 1570 1571 // Do the real work 1572 cm->follow_marking_stacks(); 1573 } 1574 }; 1575 1576 void steal_marking_work(TaskTerminator& terminator, uint worker_id) { 1577 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 1578 1579 ParCompactionManager* cm = 1580 ParCompactionManager::gc_thread_compaction_manager(worker_id); 1581 1582 do { 1583 oop obj = nullptr; 1584 ObjArrayTask task; 1585 if (ParCompactionManager::steal_objarray(worker_id, task)) { 1586 cm->follow_array((objArrayOop)task.obj(), task.index()); 1587 } else if (ParCompactionManager::steal(worker_id, obj)) { 1588 cm->follow_contents(obj); 1589 } 1590 cm->follow_marking_stacks(); 1591 } while (!terminator.offer_termination()); 1592 } 1593 1594 class MarkFromRootsTask : public WorkerTask { 1595 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do 1596 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state; 1597 TaskTerminator _terminator; 1598 uint _active_workers; 1599 1600 public: 1601 MarkFromRootsTask(uint active_workers) : 1602 WorkerTask("MarkFromRootsTask"), 1603 _strong_roots_scope(active_workers), 1604 _terminator(active_workers, ParCompactionManager::oop_task_queues()), 1605 _active_workers(active_workers) {} 1606 1607 virtual void work(uint worker_id) { 1608 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 1609 cm->create_marking_stats_cache(); 1610 PCMarkAndPushClosure mark_and_push_closure(cm); 1611 1612 { 1613 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark); 1614 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure); 1615 1616 // Do the real work 1617 cm->follow_marking_stacks(); 1618 } 1619 1620 PCAddThreadRootsMarkingTaskClosure closure(worker_id); 1621 Threads::possibly_parallel_threads_do(true /* is_par */, &closure); 1622 1623 // Mark from OopStorages 1624 { 1625 _oop_storage_set_par_state.oops_do(&mark_and_push_closure); 1626 // Do the real work 1627 cm->follow_marking_stacks(); 1628 } 1629 1630 if (_active_workers > 1) { 1631 steal_marking_work(_terminator, worker_id); 1632 } 1633 } 1634 }; 1635 1636 class ParallelCompactRefProcProxyTask : public RefProcProxyTask { 1637 TaskTerminator _terminator; 1638 1639 public: 1640 ParallelCompactRefProcProxyTask(uint max_workers) 1641 : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers), 1642 _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {} 1643 1644 void work(uint worker_id) override { 1645 assert(worker_id < _max_workers, "sanity"); 1646 ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id); 1647 PCMarkAndPushClosure keep_alive(cm); 1648 BarrierEnqueueDiscoveredFieldClosure enqueue; 1649 ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id); 1650 _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc); 1651 } 1652 1653 void prepare_run_task_hook() override { 1654 _terminator.reset_for_reuse(_queue_count); 1655 } 1656 }; 1657 1658 static void flush_marking_stats_cache(const uint num_workers) { 1659 for (uint i = 0; i < num_workers; ++i) { 1660 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i); 1661 cm->flush_and_destroy_marking_stats_cache(); 1662 } 1663 } 1664 1665 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) { 1666 // Recursively traverse all live objects and mark them 1667 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 1668 1669 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 1670 1671 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark); 1672 { 1673 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 1674 1675 MarkFromRootsTask task(active_gc_threads); 1676 ParallelScavengeHeap::heap()->workers().run_task(&task); 1677 } 1678 1679 // Process reference objects found during marking 1680 { 1681 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 1682 1683 ReferenceProcessorStats stats; 1684 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 1685 1686 ref_processor()->set_active_mt_degree(active_gc_threads); 1687 ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues()); 1688 stats = ref_processor()->process_discovered_references(task, pt); 1689 1690 gc_tracer->report_gc_reference_stats(stats); 1691 pt.print_all_references(); 1692 } 1693 1694 { 1695 GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer); 1696 1697 flush_marking_stats_cache(active_gc_threads); 1698 } 1699 1700 // This is the point where the entire marking should have completed. 1701 ParCompactionManager::verify_all_marking_stack_empty(); 1702 1703 { 1704 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 1705 WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(), 1706 is_alive_closure(), 1707 &do_nothing_cl, 1708 1); 1709 } 1710 1711 { 1712 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 1713 1714 ClassUnloadingContext* ctx = ClassUnloadingContext::context(); 1715 1716 bool unloading_occurred; 1717 { 1718 CodeCache::UnlinkingScope scope(is_alive_closure()); 1719 1720 // Follow system dictionary roots and unload classes. 1721 unloading_occurred = SystemDictionary::do_unloading(&_gc_timer); 1722 1723 // Unload nmethods. 1724 CodeCache::do_unloading(unloading_occurred); 1725 } 1726 1727 { 1728 GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer()); 1729 // Release unloaded nmethod's memory. 1730 ctx->purge_nmethods(); 1731 } 1732 { 1733 GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer); 1734 ParallelScavengeHeap::heap()->prune_unlinked_nmethods(); 1735 } 1736 { 1737 GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer()); 1738 ctx->free_nmethods(); 1739 } 1740 1741 // Prune dead klasses from subklass/sibling/implementor lists. 1742 Klass::clean_weak_klass_links(unloading_occurred); 1743 1744 // Clean JVMCI metadata handles. 1745 JVMCI_ONLY(JVMCI::do_unloading(unloading_occurred)); 1746 } 1747 1748 { 1749 GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer); 1750 _gc_tracer.report_object_count_after_gc(is_alive_closure(), &ParallelScavengeHeap::heap()->workers()); 1751 } 1752 #if TASKQUEUE_STATS 1753 ParCompactionManager::oop_task_queues()->print_and_reset_taskqueue_stats("Oop Queue"); 1754 ParCompactionManager::_objarray_task_queues->print_and_reset_taskqueue_stats("ObjArrayOop Queue"); 1755 #endif 1756 } 1757 1758 class PSAdjustTask final : public WorkerTask { 1759 SubTasksDone _sub_tasks; 1760 WeakProcessor::Task _weak_proc_task; 1761 OopStorageSetStrongParState<false, false> _oop_storage_iter; 1762 uint _nworkers; 1763 1764 enum PSAdjustSubTask { 1765 PSAdjustSubTask_code_cache, 1766 1767 PSAdjustSubTask_num_elements 1768 }; 1769 1770 public: 1771 PSAdjustTask(uint nworkers) : 1772 WorkerTask("PSAdjust task"), 1773 _sub_tasks(PSAdjustSubTask_num_elements), 1774 _weak_proc_task(nworkers), 1775 _nworkers(nworkers) { 1776 1777 ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust); 1778 if (nworkers > 1) { 1779 Threads::change_thread_claim_token(); 1780 } 1781 } 1782 1783 ~PSAdjustTask() { 1784 Threads::assert_all_threads_claimed(); 1785 } 1786 1787 void work(uint worker_id) { 1788 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 1789 PCAdjustPointerClosure adjust(cm); 1790 { 1791 ResourceMark rm; 1792 Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr); 1793 } 1794 _oop_storage_iter.oops_do(&adjust); 1795 { 1796 CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_stw_fullgc_adjust); 1797 ClassLoaderDataGraph::cld_do(&cld_closure); 1798 } 1799 { 1800 AlwaysTrueClosure always_alive; 1801 _weak_proc_task.work(worker_id, &always_alive, &adjust); 1802 } 1803 if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) { 1804 NMethodToOopClosure adjust_code(&adjust, NMethodToOopClosure::FixRelocations); 1805 CodeCache::nmethods_do(&adjust_code); 1806 } 1807 _sub_tasks.all_tasks_claimed(); 1808 } 1809 }; 1810 1811 void PSParallelCompact::adjust_roots() { 1812 // Adjust the pointers to reflect the new locations 1813 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 1814 uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers(); 1815 PSAdjustTask task(nworkers); 1816 ParallelScavengeHeap::heap()->workers().run_task(&task); 1817 } 1818 1819 // Helper class to print 8 region numbers per line and then print the total at the end. 1820 class FillableRegionLogger : public StackObj { 1821 private: 1822 Log(gc, compaction) log; 1823 static const int LineLength = 8; 1824 size_t _regions[LineLength]; 1825 int _next_index; 1826 bool _enabled; 1827 size_t _total_regions; 1828 public: 1829 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 1830 ~FillableRegionLogger() { 1831 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 1832 } 1833 1834 void print_line() { 1835 if (!_enabled || _next_index == 0) { 1836 return; 1837 } 1838 FormatBuffer<> line("Fillable: "); 1839 for (int i = 0; i < _next_index; i++) { 1840 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 1841 } 1842 log.trace("%s", line.buffer()); 1843 _next_index = 0; 1844 } 1845 1846 void handle(size_t region) { 1847 if (!_enabled) { 1848 return; 1849 } 1850 _regions[_next_index++] = region; 1851 if (_next_index == LineLength) { 1852 print_line(); 1853 } 1854 _total_regions++; 1855 } 1856 }; 1857 1858 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads) 1859 { 1860 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 1861 1862 // Find the threads that are active 1863 uint worker_id = 0; 1864 1865 // Find all regions that are available (can be filled immediately) and 1866 // distribute them to the thread stacks. The iteration is done in reverse 1867 // order (high to low) so the regions will be removed in ascending order. 1868 1869 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 1870 1871 // id + 1 is used to test termination so unsigned can 1872 // be used with an old_space_id == 0. 1873 FillableRegionLogger region_logger; 1874 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 1875 SpaceInfo* const space_info = _space_info + id; 1876 HeapWord* const new_top = space_info->new_top(); 1877 1878 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 1879 const size_t end_region = 1880 sd.addr_to_region_idx(sd.region_align_up(new_top)); 1881 1882 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 1883 if (sd.region(cur)->claim_unsafe()) { 1884 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 1885 bool result = sd.region(cur)->mark_normal(); 1886 assert(result, "Must succeed at this point."); 1887 cm->region_stack()->push(cur); 1888 region_logger.handle(cur); 1889 // Assign regions to tasks in round-robin fashion. 1890 if (++worker_id == parallel_gc_threads) { 1891 worker_id = 0; 1892 } 1893 } 1894 } 1895 region_logger.print_line(); 1896 } 1897 } 1898 1899 class TaskQueue : StackObj { 1900 volatile uint _counter; 1901 uint _size; 1902 uint _insert_index; 1903 PSParallelCompact::UpdateDensePrefixTask* _backing_array; 1904 public: 1905 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(nullptr) { 1906 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC); 1907 } 1908 ~TaskQueue() { 1909 assert(_counter >= _insert_index, "not all queue elements were claimed"); 1910 FREE_C_HEAP_ARRAY(T, _backing_array); 1911 } 1912 1913 void push(const PSParallelCompact::UpdateDensePrefixTask& value) { 1914 assert(_insert_index < _size, "too small backing array"); 1915 _backing_array[_insert_index++] = value; 1916 } 1917 1918 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) { 1919 uint claimed = Atomic::fetch_then_add(&_counter, 1u); 1920 if (claimed < _insert_index) { 1921 reference = _backing_array[claimed]; 1922 return true; 1923 } else { 1924 return false; 1925 } 1926 } 1927 }; 1928 1929 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 1930 1931 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue, 1932 uint parallel_gc_threads) { 1933 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 1934 1935 ParallelCompactData& sd = PSParallelCompact::summary_data(); 1936 1937 // Iterate over all the spaces adding tasks for updating 1938 // regions in the dense prefix. Assume that 1 gc thread 1939 // will work on opening the gaps and the remaining gc threads 1940 // will work on the dense prefix. 1941 unsigned int space_id; 1942 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 1943 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 1944 const MutableSpace* const space = _space_info[space_id].space(); 1945 1946 if (dense_prefix_end == space->bottom()) { 1947 // There is no dense prefix for this space. 1948 continue; 1949 } 1950 1951 // The dense prefix is before this region. 1952 size_t region_index_end_dense_prefix = 1953 sd.addr_to_region_idx(dense_prefix_end); 1954 RegionData* const dense_prefix_cp = 1955 sd.region(region_index_end_dense_prefix); 1956 assert(dense_prefix_end == space->end() || 1957 dense_prefix_cp->available() || 1958 dense_prefix_cp->claimed(), 1959 "The region after the dense prefix should always be ready to fill"); 1960 1961 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 1962 1963 // Is there dense prefix work? 1964 size_t total_dense_prefix_regions = 1965 region_index_end_dense_prefix - region_index_start; 1966 // How many regions of the dense prefix should be given to 1967 // each thread? 1968 if (total_dense_prefix_regions > 0) { 1969 uint tasks_for_dense_prefix = 1; 1970 if (total_dense_prefix_regions <= 1971 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 1972 // Don't over partition. This assumes that 1973 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 1974 // so there are not many regions to process. 1975 tasks_for_dense_prefix = parallel_gc_threads; 1976 } else { 1977 // Over partition 1978 tasks_for_dense_prefix = parallel_gc_threads * 1979 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 1980 } 1981 size_t regions_per_thread = total_dense_prefix_regions / 1982 tasks_for_dense_prefix; 1983 // Give each thread at least 1 region. 1984 if (regions_per_thread == 0) { 1985 regions_per_thread = 1; 1986 } 1987 1988 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 1989 if (region_index_start >= region_index_end_dense_prefix) { 1990 break; 1991 } 1992 // region_index_end is not processed 1993 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 1994 region_index_end_dense_prefix); 1995 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 1996 region_index_start, 1997 region_index_end)); 1998 region_index_start = region_index_end; 1999 } 2000 } 2001 // This gets any part of the dense prefix that did not 2002 // fit evenly. 2003 if (region_index_start < region_index_end_dense_prefix) { 2004 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2005 region_index_start, 2006 region_index_end_dense_prefix)); 2007 } 2008 } 2009 } 2010 2011 #ifdef ASSERT 2012 // Write a histogram of the number of times the block table was filled for a 2013 // region. 2014 void PSParallelCompact::write_block_fill_histogram() 2015 { 2016 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2017 return; 2018 } 2019 2020 Log(gc, compaction) log; 2021 ResourceMark rm; 2022 LogStream ls(log.trace()); 2023 outputStream* out = &ls; 2024 2025 typedef ParallelCompactData::RegionData rd_t; 2026 ParallelCompactData& sd = summary_data(); 2027 2028 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2029 MutableSpace* const spc = _space_info[id].space(); 2030 if (spc->bottom() != spc->top()) { 2031 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2032 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2033 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2034 2035 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2036 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2037 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2038 2039 for (const rd_t* cur = beg; cur < end; ++cur) { 2040 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2041 } 2042 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2043 for (size_t i = 0; i < histo_len; ++i) { 2044 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2045 histo[i], 100.0 * histo[i] / region_cnt); 2046 } 2047 out->cr(); 2048 } 2049 } 2050 } 2051 #endif // #ifdef ASSERT 2052 2053 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) { 2054 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2055 2056 ParCompactionManager* cm = 2057 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2058 2059 // Drain the stacks that have been preloaded with regions 2060 // that are ready to fill. 2061 2062 cm->drain_region_stacks(); 2063 2064 guarantee(cm->region_stack()->is_empty(), "Not empty"); 2065 2066 size_t region_index = 0; 2067 2068 while (true) { 2069 if (ParCompactionManager::steal(worker_id, region_index)) { 2070 PSParallelCompact::fill_and_update_region(cm, region_index); 2071 cm->drain_region_stacks(); 2072 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) { 2073 // Fill and update an unavailable region with the help of a shadow region 2074 PSParallelCompact::fill_and_update_shadow_region(cm, region_index); 2075 cm->drain_region_stacks(); 2076 } else { 2077 if (terminator->offer_termination()) { 2078 break; 2079 } 2080 // Go around again. 2081 } 2082 } 2083 } 2084 2085 class UpdateDensePrefixAndCompactionTask: public WorkerTask { 2086 TaskQueue& _tq; 2087 TaskTerminator _terminator; 2088 2089 public: 2090 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) : 2091 WorkerTask("UpdateDensePrefixAndCompactionTask"), 2092 _tq(tq), 2093 _terminator(active_workers, ParCompactionManager::region_task_queues()) { 2094 } 2095 virtual void work(uint worker_id) { 2096 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2097 2098 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) { 2099 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm, 2100 task._space_id, 2101 task._region_index_start, 2102 task._region_index_end); 2103 } 2104 2105 // Once a thread has drained it's stack, it should try to steal regions from 2106 // other threads. 2107 compaction_with_stealing_work(&_terminator, worker_id); 2108 2109 // At this point all regions have been compacted, so it's now safe 2110 // to update the deferred objects that cross region boundaries. 2111 cm->drain_deferred_objects(); 2112 } 2113 }; 2114 2115 void PSParallelCompact::compact() { 2116 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2117 2118 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2119 PSOldGen* old_gen = heap->old_gen(); 2120 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2121 2122 // for [0..last_space_id) 2123 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING) 2124 // push 2125 // push 2126 // 2127 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1) 2128 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)); 2129 initialize_shadow_regions(active_gc_threads); 2130 prepare_region_draining_tasks(active_gc_threads); 2131 enqueue_dense_prefix_tasks(task_queue, active_gc_threads); 2132 2133 { 2134 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2135 2136 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads); 2137 ParallelScavengeHeap::heap()->workers().run_task(&task); 2138 2139 #ifdef ASSERT 2140 // Verify that all regions have been processed. 2141 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2142 verify_complete(SpaceId(id)); 2143 } 2144 #endif 2145 } 2146 2147 DEBUG_ONLY(write_block_fill_histogram()); 2148 } 2149 2150 #ifdef ASSERT 2151 void PSParallelCompact::verify_complete(SpaceId space_id) { 2152 // All Regions between space bottom() to new_top() should be marked as filled 2153 // and all Regions between new_top() and top() should be available (i.e., 2154 // should have been emptied). 2155 ParallelCompactData& sd = summary_data(); 2156 SpaceInfo si = _space_info[space_id]; 2157 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2158 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2159 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2160 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2161 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2162 2163 bool issued_a_warning = false; 2164 2165 size_t cur_region; 2166 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2167 const RegionData* const c = sd.region(cur_region); 2168 if (!c->completed()) { 2169 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2170 cur_region, c->destination_count()); 2171 issued_a_warning = true; 2172 } 2173 } 2174 2175 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2176 const RegionData* const c = sd.region(cur_region); 2177 if (!c->available()) { 2178 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2179 cur_region, c->destination_count()); 2180 issued_a_warning = true; 2181 } 2182 } 2183 2184 if (issued_a_warning) { 2185 print_region_ranges(); 2186 } 2187 } 2188 #endif // #ifdef ASSERT 2189 2190 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2191 compaction_manager()->update_contents(cast_to_oop(addr)); 2192 } 2193 2194 // Update interior oops in the ranges of regions [beg_region, end_region). 2195 void 2196 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2197 SpaceId space_id, 2198 size_t beg_region, 2199 size_t end_region) { 2200 ParallelCompactData& sd = summary_data(); 2201 ParMarkBitMap* const mbm = mark_bitmap(); 2202 2203 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2204 HeapWord* const end_addr = sd.region_to_addr(end_region); 2205 assert(beg_region <= end_region, "bad region range"); 2206 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2207 2208 #ifdef ASSERT 2209 // Claim the regions to avoid triggering an assert when they are marked as 2210 // filled. 2211 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2212 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2213 } 2214 #endif // #ifdef ASSERT 2215 HeapWord* const space_bottom = space(space_id)->bottom(); 2216 2217 // Check if it's the first region in this space. 2218 if (beg_addr != space_bottom) { 2219 // Find the first live object or block of dead space that *starts* in this 2220 // range of regions. If a partial object crosses onto the region, skip it; 2221 // it will be marked for 'deferred update' when the object head is 2222 // processed. If dead space crosses onto the region, it is also skipped; it 2223 // will be filled when the prior region is processed. If neither of those 2224 // apply, the first word in the region is the start of a live object or dead 2225 // space. 2226 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2227 const RegionData* const cp = sd.region(beg_region); 2228 if (cp->partial_obj_size() != 0) { 2229 beg_addr = sd.partial_obj_end(beg_region); 2230 } else { 2231 idx_t beg_bit = mbm->addr_to_bit(beg_addr); 2232 if (!mbm->is_obj_beg(beg_bit) && !mbm->is_obj_end(beg_bit - 1)) { 2233 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2234 } 2235 } 2236 } 2237 2238 if (beg_addr < end_addr) { 2239 // A live object or block of dead space starts in this range of Regions. 2240 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2241 2242 // Create closures and iterate. 2243 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2244 FillClosure fill_closure(cm, space_id); 2245 mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, dense_prefix_end); 2246 } 2247 2248 // Mark the regions as filled. 2249 RegionData* const beg_cp = sd.region(beg_region); 2250 RegionData* const end_cp = sd.region(end_region); 2251 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2252 cp->set_completed(); 2253 } 2254 } 2255 2256 // Return the SpaceId for the space containing addr. If addr is not in the 2257 // heap, last_space_id is returned. In debug mode it expects the address to be 2258 // in the heap and asserts such. 2259 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2260 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2261 2262 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2263 if (_space_info[id].space()->contains(addr)) { 2264 return SpaceId(id); 2265 } 2266 } 2267 2268 assert(false, "no space contains the addr"); 2269 return last_space_id; 2270 } 2271 2272 void PSParallelCompact::update_deferred_object(ParCompactionManager* cm, HeapWord *addr) { 2273 #ifdef ASSERT 2274 ParallelCompactData& sd = summary_data(); 2275 size_t region_idx = sd.addr_to_region_idx(addr); 2276 assert(sd.region(region_idx)->completed(), "first region must be completed before deferred updates"); 2277 assert(sd.region(region_idx + 1)->completed(), "second region must be completed before deferred updates"); 2278 #endif 2279 2280 const SpaceInfo* const space_info = _space_info + space_id(addr); 2281 ObjectStartArray* const start_array = space_info->start_array(); 2282 if (start_array != nullptr) { 2283 start_array->update_for_block(addr, addr + cast_to_oop(addr)->size()); 2284 } 2285 2286 cm->update_contents(cast_to_oop(addr)); 2287 assert(oopDesc::is_oop(cast_to_oop(addr)), "Expected an oop at " PTR_FORMAT, p2i(cast_to_oop(addr))); 2288 } 2289 2290 // Skip over count live words starting from beg, and return the address of the 2291 // next live word. Unless marked, the word corresponding to beg is assumed to 2292 // be dead. Callers must either ensure beg does not correspond to the middle of 2293 // an object, or account for those live words in some other way. Callers must 2294 // also ensure that there are enough live words in the range [beg, end) to skip. 2295 HeapWord* 2296 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2297 { 2298 assert(count > 0, "sanity"); 2299 2300 ParMarkBitMap* m = mark_bitmap(); 2301 idx_t bits_to_skip = m->words_to_bits(count); 2302 idx_t cur_beg = m->addr_to_bit(beg); 2303 const idx_t search_end = m->align_range_end(m->addr_to_bit(end)); 2304 2305 do { 2306 cur_beg = m->find_obj_beg(cur_beg, search_end); 2307 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2308 const size_t obj_bits = cur_end - cur_beg + 1; 2309 if (obj_bits > bits_to_skip) { 2310 return m->bit_to_addr(cur_beg + bits_to_skip); 2311 } 2312 bits_to_skip -= obj_bits; 2313 cur_beg = cur_end + 1; 2314 } while (bits_to_skip > 0); 2315 2316 // Skipping the desired number of words landed just past the end of an object. 2317 // Find the start of the next object. 2318 cur_beg = m->find_obj_beg(cur_beg, search_end); 2319 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2320 return m->bit_to_addr(cur_beg); 2321 } 2322 2323 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2324 SpaceId src_space_id, 2325 size_t src_region_idx) 2326 { 2327 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2328 2329 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2330 if (split_info.dest_region_addr() == dest_addr) { 2331 // The partial object ending at the split point contains the first word to 2332 // be copied to dest_addr. 2333 return split_info.first_src_addr(); 2334 } 2335 2336 const ParallelCompactData& sd = summary_data(); 2337 ParMarkBitMap* const bitmap = mark_bitmap(); 2338 const size_t RegionSize = ParallelCompactData::RegionSize; 2339 2340 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2341 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2342 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2343 HeapWord* const src_region_destination = src_region_ptr->destination(); 2344 2345 assert(dest_addr >= src_region_destination, "wrong src region"); 2346 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2347 2348 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2349 HeapWord* const src_region_end = src_region_beg + RegionSize; 2350 2351 HeapWord* addr = src_region_beg; 2352 if (dest_addr == src_region_destination) { 2353 // Return the first live word in the source region. 2354 if (partial_obj_size == 0) { 2355 addr = bitmap->find_obj_beg(addr, src_region_end); 2356 assert(addr < src_region_end, "no objects start in src region"); 2357 } 2358 return addr; 2359 } 2360 2361 // Must skip some live data. 2362 size_t words_to_skip = dest_addr - src_region_destination; 2363 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2364 2365 if (partial_obj_size >= words_to_skip) { 2366 // All the live words to skip are part of the partial object. 2367 addr += words_to_skip; 2368 if (partial_obj_size == words_to_skip) { 2369 // Find the first live word past the partial object. 2370 addr = bitmap->find_obj_beg(addr, src_region_end); 2371 assert(addr < src_region_end, "wrong src region"); 2372 } 2373 return addr; 2374 } 2375 2376 // Skip over the partial object (if any). 2377 if (partial_obj_size != 0) { 2378 words_to_skip -= partial_obj_size; 2379 addr += partial_obj_size; 2380 } 2381 2382 // Skip over live words due to objects that start in the region. 2383 addr = skip_live_words(addr, src_region_end, words_to_skip); 2384 assert(addr < src_region_end, "wrong src region"); 2385 return addr; 2386 } 2387 2388 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2389 SpaceId src_space_id, 2390 size_t beg_region, 2391 HeapWord* end_addr) 2392 { 2393 ParallelCompactData& sd = summary_data(); 2394 2395 #ifdef ASSERT 2396 MutableSpace* const src_space = _space_info[src_space_id].space(); 2397 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2398 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2399 "src_space_id does not match beg_addr"); 2400 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2401 "src_space_id does not match end_addr"); 2402 #endif // #ifdef ASSERT 2403 2404 RegionData* const beg = sd.region(beg_region); 2405 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2406 2407 // Regions up to new_top() are enqueued if they become available. 2408 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2409 RegionData* const enqueue_end = 2410 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2411 2412 for (RegionData* cur = beg; cur < end; ++cur) { 2413 assert(cur->data_size() > 0, "region must have live data"); 2414 cur->decrement_destination_count(); 2415 if (cur < enqueue_end && cur->available() && cur->claim()) { 2416 if (cur->mark_normal()) { 2417 cm->push_region(sd.region(cur)); 2418 } else if (cur->mark_copied()) { 2419 // Try to copy the content of the shadow region back to its corresponding 2420 // heap region if the shadow region is filled. Otherwise, the GC thread 2421 // fills the shadow region will copy the data back (see 2422 // MoveAndUpdateShadowClosure::complete_region). 2423 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur)); 2424 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region()); 2425 cur->set_completed(); 2426 } 2427 } 2428 } 2429 } 2430 2431 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2432 SpaceId& src_space_id, 2433 HeapWord*& src_space_top, 2434 HeapWord* end_addr) 2435 { 2436 typedef ParallelCompactData::RegionData RegionData; 2437 2438 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2439 const size_t region_size = ParallelCompactData::RegionSize; 2440 2441 size_t src_region_idx = 0; 2442 2443 // Skip empty regions (if any) up to the top of the space. 2444 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2445 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2446 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2447 const RegionData* const top_region_ptr = 2448 sd.addr_to_region_ptr(top_aligned_up); 2449 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2450 ++src_region_ptr; 2451 } 2452 2453 if (src_region_ptr < top_region_ptr) { 2454 // The next source region is in the current space. Update src_region_idx 2455 // and the source address to match src_region_ptr. 2456 src_region_idx = sd.region(src_region_ptr); 2457 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2458 if (src_region_addr > closure.source()) { 2459 closure.set_source(src_region_addr); 2460 } 2461 return src_region_idx; 2462 } 2463 2464 // Switch to a new source space and find the first non-empty region. 2465 unsigned int space_id = src_space_id + 1; 2466 assert(space_id < last_space_id, "not enough spaces"); 2467 2468 HeapWord* const destination = closure.destination(); 2469 2470 do { 2471 MutableSpace* space = _space_info[space_id].space(); 2472 HeapWord* const bottom = space->bottom(); 2473 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2474 2475 // Iterate over the spaces that do not compact into themselves. 2476 if (bottom_cp->destination() != bottom) { 2477 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2478 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2479 2480 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2481 if (src_cp->live_obj_size() > 0) { 2482 // Found it. 2483 assert(src_cp->destination() == destination, 2484 "first live obj in the space must match the destination"); 2485 assert(src_cp->partial_obj_size() == 0, 2486 "a space cannot begin with a partial obj"); 2487 2488 src_space_id = SpaceId(space_id); 2489 src_space_top = space->top(); 2490 const size_t src_region_idx = sd.region(src_cp); 2491 closure.set_source(sd.region_to_addr(src_region_idx)); 2492 return src_region_idx; 2493 } else { 2494 assert(src_cp->data_size() == 0, "sanity"); 2495 } 2496 } 2497 } 2498 } while (++space_id < last_space_id); 2499 2500 assert(false, "no source region was found"); 2501 return 0; 2502 } 2503 2504 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx) 2505 { 2506 typedef ParMarkBitMap::IterationStatus IterationStatus; 2507 ParMarkBitMap* const bitmap = mark_bitmap(); 2508 ParallelCompactData& sd = summary_data(); 2509 RegionData* const region_ptr = sd.region(region_idx); 2510 2511 // Get the source region and related info. 2512 size_t src_region_idx = region_ptr->source_region(); 2513 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2514 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2515 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2516 2517 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2518 2519 // Adjust src_region_idx to prepare for decrementing destination counts (the 2520 // destination count is not decremented when a region is copied to itself). 2521 if (src_region_idx == region_idx) { 2522 src_region_idx += 1; 2523 } 2524 2525 if (bitmap->is_unmarked(closure.source())) { 2526 // The first source word is in the middle of an object; copy the remainder 2527 // of the object or as much as will fit. The fact that pointer updates were 2528 // deferred will be noted when the object header is processed. 2529 HeapWord* const old_src_addr = closure.source(); 2530 closure.copy_partial_obj(); 2531 if (closure.is_full()) { 2532 decrement_destination_counts(cm, src_space_id, src_region_idx, 2533 closure.source()); 2534 closure.complete_region(cm, dest_addr, region_ptr); 2535 return; 2536 } 2537 2538 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2539 if (sd.region_align_down(old_src_addr) != end_addr) { 2540 // The partial object was copied from more than one source region. 2541 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2542 2543 // Move to the next source region, possibly switching spaces as well. All 2544 // args except end_addr may be modified. 2545 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2546 end_addr); 2547 } 2548 } 2549 2550 do { 2551 HeapWord* const cur_addr = closure.source(); 2552 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2553 src_space_top); 2554 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2555 2556 if (status == ParMarkBitMap::would_overflow) { 2557 // The last object did not fit. Note that interior oop updates were 2558 // deferred, then copy enough of the object to fill the region. 2559 cm->push_deferred_object(closure.destination()); 2560 status = closure.copy_until_full(); // copies from closure.source() 2561 2562 decrement_destination_counts(cm, src_space_id, src_region_idx, 2563 closure.source()); 2564 closure.complete_region(cm, dest_addr, region_ptr); 2565 return; 2566 } 2567 2568 if (status == ParMarkBitMap::full) { 2569 decrement_destination_counts(cm, src_space_id, src_region_idx, 2570 closure.source()); 2571 closure.complete_region(cm, dest_addr, region_ptr); 2572 return; 2573 } 2574 2575 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2576 2577 // Move to the next source region, possibly switching spaces as well. All 2578 // args except end_addr may be modified. 2579 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2580 end_addr); 2581 } while (true); 2582 } 2583 2584 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx) 2585 { 2586 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 2587 fill_region(cm, cl, region_idx); 2588 } 2589 2590 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx) 2591 { 2592 // Get a shadow region first 2593 ParallelCompactData& sd = summary_data(); 2594 RegionData* const region_ptr = sd.region(region_idx); 2595 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr); 2596 // The InvalidShadow return value indicates the corresponding heap region is available, 2597 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use 2598 // MoveAndUpdateShadowClosure to fill the acquired shadow region. 2599 if (shadow_region == ParCompactionManager::InvalidShadow) { 2600 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 2601 region_ptr->shadow_to_normal(); 2602 return fill_region(cm, cl, region_idx); 2603 } else { 2604 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region); 2605 return fill_region(cm, cl, region_idx); 2606 } 2607 } 2608 2609 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr) 2610 { 2611 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize); 2612 } 2613 2614 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx) 2615 { 2616 size_t next = cm->next_shadow_region(); 2617 ParallelCompactData& sd = summary_data(); 2618 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top()); 2619 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2620 2621 while (next < old_new_top) { 2622 if (sd.region(next)->mark_shadow()) { 2623 region_idx = next; 2624 return true; 2625 } 2626 next = cm->move_next_shadow_region_by(active_gc_threads); 2627 } 2628 2629 return false; 2630 } 2631 2632 // The shadow region is an optimization to address region dependencies in full GC. The basic 2633 // idea is making more regions available by temporally storing their live objects in empty 2634 // shadow regions to resolve dependencies between them and the destination regions. Therefore, 2635 // GC threads need not wait destination regions to be available before processing sources. 2636 // 2637 // A typical workflow would be: 2638 // After draining its own stack and failing to steal from others, a GC worker would pick an 2639 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills 2640 // the shadow region by copying live objects from source regions of the unavailable one. Once 2641 // the unavailable region becomes available, the data in the shadow region will be copied back. 2642 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces. 2643 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads) 2644 { 2645 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2646 2647 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2648 SpaceInfo* const space_info = _space_info + id; 2649 MutableSpace* const space = space_info->space(); 2650 2651 const size_t beg_region = 2652 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top()))); 2653 const size_t end_region = 2654 sd.addr_to_region_idx(sd.region_align_down(space->end())); 2655 2656 for (size_t cur = beg_region; cur < end_region; ++cur) { 2657 ParCompactionManager::push_shadow_region(cur); 2658 } 2659 } 2660 2661 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix()); 2662 for (uint i = 0; i < parallel_gc_threads; i++) { 2663 ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i); 2664 cm->set_next_shadow_region(beg_region + i); 2665 } 2666 } 2667 2668 void PSParallelCompact::fill_blocks(size_t region_idx) 2669 { 2670 // Fill in the block table elements for the specified region. Each block 2671 // table element holds the number of live words in the region that are to the 2672 // left of the first object that starts in the block. Thus only blocks in 2673 // which an object starts need to be filled. 2674 // 2675 // The algorithm scans the section of the bitmap that corresponds to the 2676 // region, keeping a running total of the live words. When an object start is 2677 // found, if it's the first to start in the block that contains it, the 2678 // current total is written to the block table element. 2679 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2680 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2681 const size_t RegionSize = ParallelCompactData::RegionSize; 2682 2683 ParallelCompactData& sd = summary_data(); 2684 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2685 if (partial_obj_size >= RegionSize) { 2686 return; // No objects start in this region. 2687 } 2688 2689 // Ensure the first loop iteration decides that the block has changed. 2690 size_t cur_block = sd.block_count(); 2691 2692 const ParMarkBitMap* const bitmap = mark_bitmap(); 2693 2694 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2695 assert((size_t)1 << Log2BitsPerBlock == 2696 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2697 2698 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2699 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2700 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2701 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2702 while (beg_bit < range_end) { 2703 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2704 if (new_block != cur_block) { 2705 cur_block = new_block; 2706 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2707 } 2708 2709 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2710 if (end_bit < range_end - 1) { 2711 live_bits += end_bit - beg_bit + 1; 2712 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2713 } else { 2714 return; 2715 } 2716 } 2717 } 2718 2719 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 2720 { 2721 if (source() != copy_destination()) { 2722 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 2723 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining()); 2724 } 2725 update_state(words_remaining()); 2726 assert(is_full(), "sanity"); 2727 return ParMarkBitMap::full; 2728 } 2729 2730 void MoveAndUpdateClosure::copy_partial_obj() 2731 { 2732 size_t words = words_remaining(); 2733 2734 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 2735 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 2736 if (end_addr < range_end) { 2737 words = bitmap()->obj_size(source(), end_addr); 2738 } 2739 2740 // This test is necessary; if omitted, the pointer updates to a partial object 2741 // that crosses the dense prefix boundary could be overwritten. 2742 if (source() != copy_destination()) { 2743 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 2744 Copy::aligned_conjoint_words(source(), copy_destination(), words); 2745 } 2746 update_state(words); 2747 } 2748 2749 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 2750 PSParallelCompact::RegionData *region_ptr) { 2751 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished"); 2752 region_ptr->set_completed(); 2753 } 2754 2755 ParMarkBitMapClosure::IterationStatus 2756 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 2757 assert(destination() != nullptr, "sanity"); 2758 assert(bitmap()->obj_size(addr) == words, "bad size"); 2759 2760 _source = addr; 2761 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 2762 destination(), "wrong destination"); 2763 2764 if (words > words_remaining()) { 2765 return ParMarkBitMap::would_overflow; 2766 } 2767 2768 // The start_array must be updated even if the object is not moving. 2769 if (_start_array != nullptr) { 2770 _start_array->update_for_block(destination(), destination() + words); 2771 } 2772 2773 if (copy_destination() != source()) { 2774 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 2775 Copy::aligned_conjoint_words(source(), copy_destination(), words); 2776 } 2777 2778 oop moved_oop = cast_to_oop(copy_destination()); 2779 compaction_manager()->update_contents(moved_oop); 2780 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or null at " PTR_FORMAT, p2i(moved_oop)); 2781 2782 update_state(words); 2783 assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity"); 2784 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 2785 } 2786 2787 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 2788 PSParallelCompact::RegionData *region_ptr) { 2789 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow"); 2790 // Record the shadow region index 2791 region_ptr->set_shadow_region(_shadow); 2792 // Mark the shadow region as filled to indicate the data is ready to be 2793 // copied back 2794 region_ptr->mark_filled(); 2795 // Try to copy the content of the shadow region back to its corresponding 2796 // heap region if available; the GC thread that decreases the destination 2797 // count to zero will do the copying otherwise (see 2798 // PSParallelCompact::decrement_destination_counts). 2799 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) { 2800 region_ptr->set_completed(); 2801 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr); 2802 ParCompactionManager::push_shadow_region_mt_safe(_shadow); 2803 } 2804 } 2805 2806 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 2807 ParCompactionManager* cm, 2808 PSParallelCompact::SpaceId space_id) : 2809 ParMarkBitMapClosure(mbm, cm), 2810 _start_array(PSParallelCompact::start_array(space_id)) 2811 { 2812 } 2813 2814 // Updates the references in the object to their new values. 2815 ParMarkBitMapClosure::IterationStatus 2816 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 2817 do_addr(addr); 2818 return ParMarkBitMap::incomplete; 2819 } 2820 2821 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 2822 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 2823 _start_array(PSParallelCompact::start_array(space_id)) 2824 { 2825 assert(space_id == PSParallelCompact::old_space_id, 2826 "cannot use FillClosure in the young gen"); 2827 } 2828 2829 ParMarkBitMapClosure::IterationStatus 2830 FillClosure::do_addr(HeapWord* addr, size_t size) { 2831 CollectedHeap::fill_with_objects(addr, size); 2832 HeapWord* const end = addr + size; 2833 do { 2834 size_t size = cast_to_oop(addr)->size(); 2835 _start_array->update_for_block(addr, addr + size); 2836 addr += size; 2837 } while (addr < end); 2838 return ParMarkBitMap::incomplete; 2839 }