1 /* 2 * Copyright (c) 2005, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "aot/aotLoader.hpp" 27 #include "classfile/classLoaderDataGraph.hpp" 28 #include "classfile/javaClasses.inline.hpp" 29 #include "classfile/stringTable.hpp" 30 #include "classfile/symbolTable.hpp" 31 #include "classfile/systemDictionary.hpp" 32 #include "code/codeCache.hpp" 33 #include "gc/parallel/parallelArguments.hpp" 34 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 35 #include "gc/parallel/parMarkBitMap.inline.hpp" 36 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 37 #include "gc/parallel/psCompactionManager.inline.hpp" 38 #include "gc/parallel/psOldGen.hpp" 39 #include "gc/parallel/psParallelCompact.inline.hpp" 40 #include "gc/parallel/psPromotionManager.inline.hpp" 41 #include "gc/parallel/psRootType.hpp" 42 #include "gc/parallel/psScavenge.hpp" 43 #include "gc/parallel/psYoungGen.hpp" 44 #include "gc/shared/gcCause.hpp" 45 #include "gc/shared/gcHeapSummary.hpp" 46 #include "gc/shared/gcId.hpp" 47 #include "gc/shared/gcLocker.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/isGCActiveMark.hpp" 52 #include "gc/shared/oopStorage.inline.hpp" 53 #include "gc/shared/oopStorageSet.inline.hpp" 54 #include "gc/shared/oopStorageSetParState.inline.hpp" 55 #include "gc/shared/referencePolicy.hpp" 56 #include "gc/shared/referenceProcessor.hpp" 57 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 58 #include "gc/shared/spaceDecorator.inline.hpp" 59 #include "gc/shared/taskTerminator.hpp" 60 #include "gc/shared/weakProcessor.hpp" 61 #include "gc/shared/workerPolicy.hpp" 62 #include "gc/shared/workgroup.hpp" 63 #include "logging/log.hpp" 64 #include "memory/iterator.inline.hpp" 65 #include "memory/resourceArea.hpp" 66 #include "memory/universe.hpp" 67 #include "oops/access.inline.hpp" 68 #include "oops/instanceClassLoaderKlass.inline.hpp" 69 #include "oops/instanceKlass.inline.hpp" 70 #include "oops/instanceMirrorKlass.inline.hpp" 71 #include "oops/methodData.hpp" 72 #include "oops/objArrayKlass.inline.hpp" 73 #include "oops/oop.inline.hpp" 74 #include "runtime/atomic.hpp" 75 #include "runtime/handles.inline.hpp" 76 #include "runtime/safepoint.hpp" 77 #include "runtime/vmThread.hpp" 78 #include "services/memTracker.hpp" 79 #include "services/memoryService.hpp" 80 #include "utilities/align.hpp" 81 #include "utilities/debug.hpp" 82 #include "utilities/events.hpp" 83 #include "utilities/formatBuffer.hpp" 84 #include "utilities/macros.hpp" 85 #include "utilities/stack.inline.hpp" 86 #if INCLUDE_JVMCI 87 #include "jvmci/jvmci.hpp" 88 #endif 89 90 #include <math.h> 91 92 // All sizes are in HeapWords. 93 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 94 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 95 const size_t ParallelCompactData::RegionSizeBytes = 96 RegionSize << LogHeapWordSize; 97 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 98 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 99 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 100 101 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 102 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 103 const size_t ParallelCompactData::BlockSizeBytes = 104 BlockSize << LogHeapWordSize; 105 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 106 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 107 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 108 109 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 110 const size_t ParallelCompactData::Log2BlocksPerRegion = 111 Log2RegionSize - Log2BlockSize; 112 113 const ParallelCompactData::RegionData::region_sz_t 114 ParallelCompactData::RegionData::dc_shift = 27; 115 116 const ParallelCompactData::RegionData::region_sz_t 117 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 118 119 const ParallelCompactData::RegionData::region_sz_t 120 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 121 122 const ParallelCompactData::RegionData::region_sz_t 123 ParallelCompactData::RegionData::los_mask = ~dc_mask; 124 125 const ParallelCompactData::RegionData::region_sz_t 126 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 127 128 const ParallelCompactData::RegionData::region_sz_t 129 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 130 131 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 132 133 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 134 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 135 136 double PSParallelCompact::_dwl_mean; 137 double PSParallelCompact::_dwl_std_dev; 138 double PSParallelCompact::_dwl_first_term; 139 double PSParallelCompact::_dwl_adjustment; 140 #ifdef ASSERT 141 bool PSParallelCompact::_dwl_initialized = false; 142 #endif // #ifdef ASSERT 143 144 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 145 HeapWord* destination) 146 { 147 assert(src_region_idx != 0, "invalid src_region_idx"); 148 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 149 assert(destination != NULL, "invalid destination argument"); 150 151 _src_region_idx = src_region_idx; 152 _partial_obj_size = partial_obj_size; 153 _destination = destination; 154 155 // These fields may not be updated below, so make sure they're clear. 156 assert(_dest_region_addr == NULL, "should have been cleared"); 157 assert(_first_src_addr == NULL, "should have been cleared"); 158 159 // Determine the number of destination regions for the partial object. 160 HeapWord* const last_word = destination + partial_obj_size - 1; 161 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 162 HeapWord* const beg_region_addr = sd.region_align_down(destination); 163 HeapWord* const end_region_addr = sd.region_align_down(last_word); 164 165 if (beg_region_addr == end_region_addr) { 166 // One destination region. 167 _destination_count = 1; 168 if (end_region_addr == destination) { 169 // The destination falls on a region boundary, thus the first word of the 170 // partial object will be the first word copied to the destination region. 171 _dest_region_addr = end_region_addr; 172 _first_src_addr = sd.region_to_addr(src_region_idx); 173 } 174 } else { 175 // Two destination regions. When copied, the partial object will cross a 176 // destination region boundary, so a word somewhere within the partial 177 // object will be the first word copied to the second destination region. 178 _destination_count = 2; 179 _dest_region_addr = end_region_addr; 180 const size_t ofs = pointer_delta(end_region_addr, destination); 181 assert(ofs < _partial_obj_size, "sanity"); 182 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 183 } 184 } 185 186 void SplitInfo::clear() 187 { 188 _src_region_idx = 0; 189 _partial_obj_size = 0; 190 _destination = NULL; 191 _destination_count = 0; 192 _dest_region_addr = NULL; 193 _first_src_addr = NULL; 194 assert(!is_valid(), "sanity"); 195 } 196 197 #ifdef ASSERT 198 void SplitInfo::verify_clear() 199 { 200 assert(_src_region_idx == 0, "not clear"); 201 assert(_partial_obj_size == 0, "not clear"); 202 assert(_destination == NULL, "not clear"); 203 assert(_destination_count == 0, "not clear"); 204 assert(_dest_region_addr == NULL, "not clear"); 205 assert(_first_src_addr == NULL, "not clear"); 206 } 207 #endif // #ifdef ASSERT 208 209 210 void PSParallelCompact::print_on_error(outputStream* st) { 211 _mark_bitmap.print_on_error(st); 212 } 213 214 #ifndef PRODUCT 215 const char* PSParallelCompact::space_names[] = { 216 "old ", "eden", "from", "to " 217 }; 218 219 void PSParallelCompact::print_region_ranges() { 220 if (!log_develop_is_enabled(Trace, gc, compaction)) { 221 return; 222 } 223 Log(gc, compaction) log; 224 ResourceMark rm; 225 LogStream ls(log.trace()); 226 Universe::print_on(&ls); 227 log.trace("space bottom top end new_top"); 228 log.trace("------ ---------- ---------- ---------- ----------"); 229 230 for (unsigned int id = 0; id < last_space_id; ++id) { 231 const MutableSpace* space = _space_info[id].space(); 232 log.trace("%u %s " 233 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 234 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 235 id, space_names[id], 236 summary_data().addr_to_region_idx(space->bottom()), 237 summary_data().addr_to_region_idx(space->top()), 238 summary_data().addr_to_region_idx(space->end()), 239 summary_data().addr_to_region_idx(_space_info[id].new_top())); 240 } 241 } 242 243 void 244 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 245 { 246 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 247 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 248 249 ParallelCompactData& sd = PSParallelCompact::summary_data(); 250 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 251 log_develop_trace(gc, compaction)( 252 REGION_IDX_FORMAT " " PTR_FORMAT " " 253 REGION_IDX_FORMAT " " PTR_FORMAT " " 254 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 255 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 256 i, p2i(c->data_location()), dci, p2i(c->destination()), 257 c->partial_obj_size(), c->live_obj_size(), 258 c->data_size(), c->source_region(), c->destination_count()); 259 260 #undef REGION_IDX_FORMAT 261 #undef REGION_DATA_FORMAT 262 } 263 264 void 265 print_generic_summary_data(ParallelCompactData& summary_data, 266 HeapWord* const beg_addr, 267 HeapWord* const end_addr) 268 { 269 size_t total_words = 0; 270 size_t i = summary_data.addr_to_region_idx(beg_addr); 271 const size_t last = summary_data.addr_to_region_idx(end_addr); 272 HeapWord* pdest = 0; 273 274 while (i < last) { 275 ParallelCompactData::RegionData* c = summary_data.region(i); 276 if (c->data_size() != 0 || c->destination() != pdest) { 277 print_generic_summary_region(i, c); 278 total_words += c->data_size(); 279 pdest = c->destination(); 280 } 281 ++i; 282 } 283 284 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 285 } 286 287 void 288 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 289 HeapWord* const beg_addr, 290 HeapWord* const end_addr) { 291 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 292 } 293 294 void 295 print_generic_summary_data(ParallelCompactData& summary_data, 296 SpaceInfo* space_info) 297 { 298 if (!log_develop_is_enabled(Trace, gc, compaction)) { 299 return; 300 } 301 302 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 303 const MutableSpace* space = space_info[id].space(); 304 print_generic_summary_data(summary_data, space->bottom(), 305 MAX2(space->top(), space_info[id].new_top())); 306 } 307 } 308 309 void 310 print_initial_summary_data(ParallelCompactData& summary_data, 311 const MutableSpace* space) { 312 if (space->top() == space->bottom()) { 313 return; 314 } 315 316 const size_t region_size = ParallelCompactData::RegionSize; 317 typedef ParallelCompactData::RegionData RegionData; 318 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 319 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 320 const RegionData* c = summary_data.region(end_region - 1); 321 HeapWord* end_addr = c->destination() + c->data_size(); 322 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 323 324 // Print (and count) the full regions at the beginning of the space. 325 size_t full_region_count = 0; 326 size_t i = summary_data.addr_to_region_idx(space->bottom()); 327 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 328 ParallelCompactData::RegionData* c = summary_data.region(i); 329 log_develop_trace(gc, compaction)( 330 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 331 i, p2i(c->destination()), 332 c->partial_obj_size(), c->live_obj_size(), 333 c->data_size(), c->source_region(), c->destination_count()); 334 ++full_region_count; 335 ++i; 336 } 337 338 size_t live_to_right = live_in_space - full_region_count * region_size; 339 340 double max_reclaimed_ratio = 0.0; 341 size_t max_reclaimed_ratio_region = 0; 342 size_t max_dead_to_right = 0; 343 size_t max_live_to_right = 0; 344 345 // Print the 'reclaimed ratio' for regions while there is something live in 346 // the region or to the right of it. The remaining regions are empty (and 347 // uninteresting), and computing the ratio will result in division by 0. 348 while (i < end_region && live_to_right > 0) { 349 c = summary_data.region(i); 350 HeapWord* const region_addr = summary_data.region_to_addr(i); 351 const size_t used_to_right = pointer_delta(space->top(), region_addr); 352 const size_t dead_to_right = used_to_right - live_to_right; 353 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 354 355 if (reclaimed_ratio > max_reclaimed_ratio) { 356 max_reclaimed_ratio = reclaimed_ratio; 357 max_reclaimed_ratio_region = i; 358 max_dead_to_right = dead_to_right; 359 max_live_to_right = live_to_right; 360 } 361 362 ParallelCompactData::RegionData* c = summary_data.region(i); 363 log_develop_trace(gc, compaction)( 364 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 365 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 366 i, p2i(c->destination()), 367 c->partial_obj_size(), c->live_obj_size(), 368 c->data_size(), c->source_region(), c->destination_count(), 369 reclaimed_ratio, dead_to_right, live_to_right); 370 371 372 live_to_right -= c->data_size(); 373 ++i; 374 } 375 376 // Any remaining regions are empty. Print one more if there is one. 377 if (i < end_region) { 378 ParallelCompactData::RegionData* c = summary_data.region(i); 379 log_develop_trace(gc, compaction)( 380 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 381 i, p2i(c->destination()), 382 c->partial_obj_size(), c->live_obj_size(), 383 c->data_size(), c->source_region(), c->destination_count()); 384 } 385 386 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 387 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 388 } 389 390 void 391 print_initial_summary_data(ParallelCompactData& summary_data, 392 SpaceInfo* space_info) { 393 if (!log_develop_is_enabled(Trace, gc, compaction)) { 394 return; 395 } 396 397 unsigned int id = PSParallelCompact::old_space_id; 398 const MutableSpace* space; 399 do { 400 space = space_info[id].space(); 401 print_initial_summary_data(summary_data, space); 402 } while (++id < PSParallelCompact::eden_space_id); 403 404 do { 405 space = space_info[id].space(); 406 print_generic_summary_data(summary_data, space->bottom(), space->top()); 407 } while (++id < PSParallelCompact::last_space_id); 408 } 409 #endif // #ifndef PRODUCT 410 411 #ifdef ASSERT 412 size_t add_obj_count; 413 size_t add_obj_size; 414 size_t mark_bitmap_count; 415 size_t mark_bitmap_size; 416 #endif // #ifdef ASSERT 417 418 ParallelCompactData::ParallelCompactData() : 419 _region_start(NULL), 420 DEBUG_ONLY(_region_end(NULL) COMMA) 421 _region_vspace(NULL), 422 _reserved_byte_size(0), 423 _region_data(NULL), 424 _region_count(0), 425 _block_vspace(NULL), 426 _block_data(NULL), 427 _block_count(0) {} 428 429 bool ParallelCompactData::initialize(MemRegion covered_region) 430 { 431 _region_start = covered_region.start(); 432 const size_t region_size = covered_region.word_size(); 433 DEBUG_ONLY(_region_end = _region_start + region_size;) 434 435 assert(region_align_down(_region_start) == _region_start, 436 "region start not aligned"); 437 assert((region_size & RegionSizeOffsetMask) == 0, 438 "region size not a multiple of RegionSize"); 439 440 bool result = initialize_region_data(region_size) && initialize_block_data(); 441 return result; 442 } 443 444 PSVirtualSpace* 445 ParallelCompactData::create_vspace(size_t count, size_t element_size) 446 { 447 const size_t raw_bytes = count * element_size; 448 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 449 const size_t granularity = os::vm_allocation_granularity(); 450 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 451 452 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 453 MAX2(page_sz, granularity); 454 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 455 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 456 rs.size()); 457 458 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 459 460 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 461 if (vspace != 0) { 462 if (vspace->expand_by(_reserved_byte_size)) { 463 return vspace; 464 } 465 delete vspace; 466 // Release memory reserved in the space. 467 rs.release(); 468 } 469 470 return 0; 471 } 472 473 bool ParallelCompactData::initialize_region_data(size_t region_size) 474 { 475 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 476 _region_vspace = create_vspace(count, sizeof(RegionData)); 477 if (_region_vspace != 0) { 478 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 479 _region_count = count; 480 return true; 481 } 482 return false; 483 } 484 485 bool ParallelCompactData::initialize_block_data() 486 { 487 assert(_region_count != 0, "region data must be initialized first"); 488 const size_t count = _region_count << Log2BlocksPerRegion; 489 _block_vspace = create_vspace(count, sizeof(BlockData)); 490 if (_block_vspace != 0) { 491 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 492 _block_count = count; 493 return true; 494 } 495 return false; 496 } 497 498 void ParallelCompactData::clear() 499 { 500 memset(_region_data, 0, _region_vspace->committed_size()); 501 memset(_block_data, 0, _block_vspace->committed_size()); 502 } 503 504 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 505 assert(beg_region <= _region_count, "beg_region out of range"); 506 assert(end_region <= _region_count, "end_region out of range"); 507 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 508 509 const size_t region_cnt = end_region - beg_region; 510 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 511 512 const size_t beg_block = beg_region * BlocksPerRegion; 513 const size_t block_cnt = region_cnt * BlocksPerRegion; 514 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 515 } 516 517 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 518 { 519 const RegionData* cur_cp = region(region_idx); 520 const RegionData* const end_cp = region(region_count() - 1); 521 522 HeapWord* result = region_to_addr(region_idx); 523 if (cur_cp < end_cp) { 524 do { 525 result += cur_cp->partial_obj_size(); 526 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 527 } 528 return result; 529 } 530 531 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 532 { 533 const size_t obj_ofs = pointer_delta(addr, _region_start); 534 const size_t beg_region = obj_ofs >> Log2RegionSize; 535 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 536 537 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 538 DEBUG_ONLY(Atomic::add(&add_obj_size, len);) 539 540 if (beg_region == end_region) { 541 // All in one region. 542 _region_data[beg_region].add_live_obj(len); 543 return; 544 } 545 546 // First region. 547 const size_t beg_ofs = region_offset(addr); 548 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 549 550 Klass* klass = ((oop)addr)->klass(); 551 // Middle regions--completely spanned by this object. 552 for (size_t region = beg_region + 1; region < end_region; ++region) { 553 _region_data[region].set_partial_obj_size(RegionSize); 554 _region_data[region].set_partial_obj_addr(addr); 555 } 556 557 // Last region. 558 const size_t end_ofs = region_offset(addr + len - 1); 559 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 560 _region_data[end_region].set_partial_obj_addr(addr); 561 } 562 563 void 564 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 565 { 566 assert(region_offset(beg) == 0, "not RegionSize aligned"); 567 assert(region_offset(end) == 0, "not RegionSize aligned"); 568 569 size_t cur_region = addr_to_region_idx(beg); 570 const size_t end_region = addr_to_region_idx(end); 571 HeapWord* addr = beg; 572 while (cur_region < end_region) { 573 _region_data[cur_region].set_destination(addr); 574 _region_data[cur_region].set_destination_count(0); 575 _region_data[cur_region].set_source_region(cur_region); 576 _region_data[cur_region].set_data_location(addr); 577 578 // Update live_obj_size so the region appears completely full. 579 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 580 _region_data[cur_region].set_live_obj_size(live_size); 581 582 ++cur_region; 583 addr += RegionSize; 584 } 585 } 586 587 // Find the point at which a space can be split and, if necessary, record the 588 // split point. 589 // 590 // If the current src region (which overflowed the destination space) doesn't 591 // have a partial object, the split point is at the beginning of the current src 592 // region (an "easy" split, no extra bookkeeping required). 593 // 594 // If the current src region has a partial object, the split point is in the 595 // region where that partial object starts (call it the split_region). If 596 // split_region has a partial object, then the split point is just after that 597 // partial object (a "hard" split where we have to record the split data and 598 // zero the partial_obj_size field). With a "hard" split, we know that the 599 // partial_obj ends within split_region because the partial object that caused 600 // the overflow starts in split_region. If split_region doesn't have a partial 601 // obj, then the split is at the beginning of split_region (another "easy" 602 // split). 603 HeapWord* 604 ParallelCompactData::summarize_split_space(size_t src_region, 605 SplitInfo& split_info, 606 HeapWord* destination, 607 HeapWord* target_end, 608 HeapWord** target_next) 609 { 610 assert(destination <= target_end, "sanity"); 611 assert(destination + _region_data[src_region].data_size() > target_end, 612 "region should not fit into target space"); 613 assert(is_region_aligned(target_end), "sanity"); 614 615 size_t split_region = src_region; 616 HeapWord* split_destination = destination; 617 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 618 619 if (destination + partial_obj_size > target_end) { 620 // The split point is just after the partial object (if any) in the 621 // src_region that contains the start of the object that overflowed the 622 // destination space. 623 // 624 // Find the start of the "overflow" object and set split_region to the 625 // region containing it. 626 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 627 split_region = addr_to_region_idx(overflow_obj); 628 629 // Clear the source_region field of all destination regions whose first word 630 // came from data after the split point (a non-null source_region field 631 // implies a region must be filled). 632 // 633 // An alternative to the simple loop below: clear during post_compact(), 634 // which uses memcpy instead of individual stores, and is easy to 635 // parallelize. (The downside is that it clears the entire RegionData 636 // object as opposed to just one field.) 637 // 638 // post_compact() would have to clear the summary data up to the highest 639 // address that was written during the summary phase, which would be 640 // 641 // max(top, max(new_top, clear_top)) 642 // 643 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 644 // to target_end. 645 const RegionData* const sr = region(split_region); 646 const size_t beg_idx = 647 addr_to_region_idx(region_align_up(sr->destination() + 648 sr->partial_obj_size())); 649 const size_t end_idx = addr_to_region_idx(target_end); 650 651 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 652 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 653 _region_data[idx].set_source_region(0); 654 } 655 656 // Set split_destination and partial_obj_size to reflect the split region. 657 split_destination = sr->destination(); 658 partial_obj_size = sr->partial_obj_size(); 659 } 660 661 // The split is recorded only if a partial object extends onto the region. 662 if (partial_obj_size != 0) { 663 _region_data[split_region].set_partial_obj_size(0); 664 split_info.record(split_region, partial_obj_size, split_destination); 665 } 666 667 // Setup the continuation addresses. 668 *target_next = split_destination + partial_obj_size; 669 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 670 671 if (log_develop_is_enabled(Trace, gc, compaction)) { 672 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 673 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 674 split_type, p2i(source_next), split_region, partial_obj_size); 675 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 676 split_type, p2i(split_destination), 677 addr_to_region_idx(split_destination), 678 p2i(*target_next)); 679 680 if (partial_obj_size != 0) { 681 HeapWord* const po_beg = split_info.destination(); 682 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 683 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 684 split_type, 685 p2i(po_beg), addr_to_region_idx(po_beg), 686 p2i(po_end), addr_to_region_idx(po_end)); 687 } 688 } 689 690 return source_next; 691 } 692 693 bool ParallelCompactData::summarize(SplitInfo& split_info, 694 HeapWord* source_beg, HeapWord* source_end, 695 HeapWord** source_next, 696 HeapWord* target_beg, HeapWord* target_end, 697 HeapWord** target_next) 698 { 699 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 700 log_develop_trace(gc, compaction)( 701 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 702 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 703 p2i(source_beg), p2i(source_end), p2i(source_next_val), 704 p2i(target_beg), p2i(target_end), p2i(*target_next)); 705 706 size_t cur_region = addr_to_region_idx(source_beg); 707 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 708 709 HeapWord *dest_addr = target_beg; 710 while (cur_region < end_region) { 711 // The destination must be set even if the region has no data. 712 _region_data[cur_region].set_destination(dest_addr); 713 714 size_t words = _region_data[cur_region].data_size(); 715 if (words > 0) { 716 // If cur_region does not fit entirely into the target space, find a point 717 // at which the source space can be 'split' so that part is copied to the 718 // target space and the rest is copied elsewhere. 719 if (dest_addr + words > target_end) { 720 assert(source_next != NULL, "source_next is NULL when splitting"); 721 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 722 target_end, target_next); 723 return false; 724 } 725 726 // Compute the destination_count for cur_region, and if necessary, update 727 // source_region for a destination region. The source_region field is 728 // updated if cur_region is the first (left-most) region to be copied to a 729 // destination region. 730 // 731 // The destination_count calculation is a bit subtle. A region that has 732 // data that compacts into itself does not count itself as a destination. 733 // This maintains the invariant that a zero count means the region is 734 // available and can be claimed and then filled. 735 uint destination_count = 0; 736 if (split_info.is_split(cur_region)) { 737 // The current region has been split: the partial object will be copied 738 // to one destination space and the remaining data will be copied to 739 // another destination space. Adjust the initial destination_count and, 740 // if necessary, set the source_region field if the partial object will 741 // cross a destination region boundary. 742 destination_count = split_info.destination_count(); 743 if (destination_count == 2) { 744 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 745 _region_data[dest_idx].set_source_region(cur_region); 746 } 747 } 748 749 HeapWord* const last_addr = dest_addr + words - 1; 750 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 751 const size_t dest_region_2 = addr_to_region_idx(last_addr); 752 753 // Initially assume that the destination regions will be the same and 754 // adjust the value below if necessary. Under this assumption, if 755 // cur_region == dest_region_2, then cur_region will be compacted 756 // completely into itself. 757 destination_count += cur_region == dest_region_2 ? 0 : 1; 758 if (dest_region_1 != dest_region_2) { 759 // Destination regions differ; adjust destination_count. 760 destination_count += 1; 761 // Data from cur_region will be copied to the start of dest_region_2. 762 _region_data[dest_region_2].set_source_region(cur_region); 763 } else if (region_offset(dest_addr) == 0) { 764 // Data from cur_region will be copied to the start of the destination 765 // region. 766 _region_data[dest_region_1].set_source_region(cur_region); 767 } 768 769 _region_data[cur_region].set_destination_count(destination_count); 770 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 771 dest_addr += words; 772 } 773 774 ++cur_region; 775 } 776 777 *target_next = dest_addr; 778 return true; 779 } 780 781 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 782 assert(addr != NULL, "Should detect NULL oop earlier"); 783 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 784 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 785 786 // Region covering the object. 787 RegionData* const region_ptr = addr_to_region_ptr(addr); 788 HeapWord* result = region_ptr->destination(); 789 790 // If the entire Region is live, the new location is region->destination + the 791 // offset of the object within in the Region. 792 793 // Run some performance tests to determine if this special case pays off. It 794 // is worth it for pointers into the dense prefix. If the optimization to 795 // avoid pointer updates in regions that only point to the dense prefix is 796 // ever implemented, this should be revisited. 797 if (region_ptr->data_size() == RegionSize) { 798 result += region_offset(addr); 799 return result; 800 } 801 802 // Otherwise, the new location is region->destination + block offset + the 803 // number of live words in the Block that are (a) to the left of addr and (b) 804 // due to objects that start in the Block. 805 806 // Fill in the block table if necessary. This is unsynchronized, so multiple 807 // threads may fill the block table for a region (harmless, since it is 808 // idempotent). 809 if (!region_ptr->blocks_filled()) { 810 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 811 region_ptr->set_blocks_filled(); 812 } 813 814 HeapWord* const search_start = block_align_down(addr); 815 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 816 817 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 818 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 819 result += block_offset + live; 820 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 821 return result; 822 } 823 824 #ifdef ASSERT 825 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 826 { 827 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 828 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 829 for (const size_t* p = beg; p < end; ++p) { 830 assert(*p == 0, "not zero"); 831 } 832 } 833 834 void ParallelCompactData::verify_clear() 835 { 836 verify_clear(_region_vspace); 837 verify_clear(_block_vspace); 838 } 839 #endif // #ifdef ASSERT 840 841 STWGCTimer PSParallelCompact::_gc_timer; 842 ParallelOldTracer PSParallelCompact::_gc_tracer; 843 elapsedTimer PSParallelCompact::_accumulated_time; 844 unsigned int PSParallelCompact::_total_invocations = 0; 845 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 846 jlong PSParallelCompact::_time_of_last_gc = 0; 847 CollectorCounters* PSParallelCompact::_counters = NULL; 848 ParMarkBitMap PSParallelCompact::_mark_bitmap; 849 ParallelCompactData PSParallelCompact::_summary_data; 850 851 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 852 853 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 854 855 class PCReferenceProcessor: public ReferenceProcessor { 856 public: 857 PCReferenceProcessor( 858 BoolObjectClosure* is_subject_to_discovery, 859 BoolObjectClosure* is_alive_non_header) : 860 ReferenceProcessor(is_subject_to_discovery, 861 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 862 ParallelGCThreads, // mt processing degree 863 true, // mt discovery 864 ParallelGCThreads, // mt discovery degree 865 true, // atomic_discovery 866 is_alive_non_header) { 867 } 868 869 template<typename T> bool discover(oop obj, ReferenceType type) { 870 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj); 871 T heap_oop = RawAccess<>::oop_load(referent_addr); 872 oop referent = CompressedOops::decode_not_null(heap_oop); 873 return PSParallelCompact::mark_bitmap()->is_unmarked(referent) 874 && ReferenceProcessor::discover_reference(obj, type); 875 } 876 virtual bool discover_reference(oop obj, ReferenceType type) { 877 if (UseCompressedOops) { 878 return discover<narrowOop>(obj, type); 879 } else { 880 return discover<oop>(obj, type); 881 } 882 } 883 }; 884 885 void PSParallelCompact::post_initialize() { 886 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 887 _span_based_discoverer.set_span(heap->reserved_region()); 888 _ref_processor = 889 new PCReferenceProcessor(&_span_based_discoverer, 890 &_is_alive_closure); // non-header is alive closure 891 892 _counters = new CollectorCounters("Parallel full collection pauses", 1); 893 894 // Initialize static fields in ParCompactionManager. 895 ParCompactionManager::initialize(mark_bitmap()); 896 } 897 898 bool PSParallelCompact::initialize() { 899 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 900 MemRegion mr = heap->reserved_region(); 901 902 // Was the old gen get allocated successfully? 903 if (!heap->old_gen()->is_allocated()) { 904 return false; 905 } 906 907 initialize_space_info(); 908 initialize_dead_wood_limiter(); 909 910 if (!_mark_bitmap.initialize(mr)) { 911 vm_shutdown_during_initialization( 912 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 913 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 914 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 915 return false; 916 } 917 918 if (!_summary_data.initialize(mr)) { 919 vm_shutdown_during_initialization( 920 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 921 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 922 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 923 return false; 924 } 925 926 return true; 927 } 928 929 void PSParallelCompact::initialize_space_info() 930 { 931 memset(&_space_info, 0, sizeof(_space_info)); 932 933 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 934 PSYoungGen* young_gen = heap->young_gen(); 935 936 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 937 _space_info[eden_space_id].set_space(young_gen->eden_space()); 938 _space_info[from_space_id].set_space(young_gen->from_space()); 939 _space_info[to_space_id].set_space(young_gen->to_space()); 940 941 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 942 } 943 944 void PSParallelCompact::initialize_dead_wood_limiter() 945 { 946 const size_t max = 100; 947 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 948 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 949 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 950 DEBUG_ONLY(_dwl_initialized = true;) 951 _dwl_adjustment = normal_distribution(1.0); 952 } 953 954 void 955 PSParallelCompact::clear_data_covering_space(SpaceId id) 956 { 957 // At this point, top is the value before GC, new_top() is the value that will 958 // be set at the end of GC. The marking bitmap is cleared to top; nothing 959 // should be marked above top. The summary data is cleared to the larger of 960 // top & new_top. 961 MutableSpace* const space = _space_info[id].space(); 962 HeapWord* const bot = space->bottom(); 963 HeapWord* const top = space->top(); 964 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 965 966 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 967 const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top)); 968 _mark_bitmap.clear_range(beg_bit, end_bit); 969 970 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 971 const size_t end_region = 972 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 973 _summary_data.clear_range(beg_region, end_region); 974 975 // Clear the data used to 'split' regions. 976 SplitInfo& split_info = _space_info[id].split_info(); 977 if (split_info.is_valid()) { 978 split_info.clear(); 979 } 980 DEBUG_ONLY(split_info.verify_clear();) 981 } 982 983 void PSParallelCompact::pre_compact() 984 { 985 // Update the from & to space pointers in space_info, since they are swapped 986 // at each young gen gc. Do the update unconditionally (even though a 987 // promotion failure does not swap spaces) because an unknown number of young 988 // collections will have swapped the spaces an unknown number of times. 989 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 990 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 991 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 992 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 993 994 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 995 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 996 997 // Increment the invocation count 998 heap->increment_total_collections(true); 999 1000 // We need to track unique mark sweep invocations as well. 1001 _total_invocations++; 1002 1003 heap->print_heap_before_gc(); 1004 heap->trace_heap_before_gc(&_gc_tracer); 1005 1006 // Fill in TLABs 1007 heap->ensure_parsability(true); // retire TLABs 1008 1009 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1010 HandleMark hm; // Discard invalid handles created during verification 1011 Universe::verify("Before GC"); 1012 } 1013 1014 // Verify object start arrays 1015 if (VerifyObjectStartArray && 1016 VerifyBeforeGC) { 1017 heap->old_gen()->verify_object_start_array(); 1018 } 1019 1020 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1021 DEBUG_ONLY(summary_data().verify_clear();) 1022 1023 ParCompactionManager::reset_all_bitmap_query_caches(); 1024 } 1025 1026 void PSParallelCompact::post_compact() 1027 { 1028 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1029 ParCompactionManager::remove_all_shadow_regions(); 1030 1031 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1032 // Clear the marking bitmap, summary data and split info. 1033 clear_data_covering_space(SpaceId(id)); 1034 // Update top(). Must be done after clearing the bitmap and summary data. 1035 _space_info[id].publish_new_top(); 1036 } 1037 1038 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1039 MutableSpace* const from_space = _space_info[from_space_id].space(); 1040 MutableSpace* const to_space = _space_info[to_space_id].space(); 1041 1042 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1043 bool eden_empty = eden_space->is_empty(); 1044 1045 // Update heap occupancy information which is used as input to the soft ref 1046 // clearing policy at the next gc. 1047 Universe::update_heap_info_at_gc(); 1048 1049 bool young_gen_empty = eden_empty && from_space->is_empty() && 1050 to_space->is_empty(); 1051 1052 PSCardTable* ct = heap->card_table(); 1053 MemRegion old_mr = heap->old_gen()->reserved(); 1054 if (young_gen_empty) { 1055 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1056 } else { 1057 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1058 } 1059 1060 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1061 ClassLoaderDataGraph::purge(); 1062 MetaspaceUtils::verify_metrics(); 1063 1064 heap->prune_scavengable_nmethods(); 1065 1066 #if COMPILER2_OR_JVMCI 1067 DerivedPointerTable::update_pointers(); 1068 #endif 1069 1070 if (ZapUnusedHeapArea) { 1071 heap->gen_mangle_unused_area(); 1072 } 1073 1074 // Update time of last GC 1075 reset_millis_since_last_gc(); 1076 } 1077 1078 HeapWord* 1079 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1080 bool maximum_compaction) 1081 { 1082 const size_t region_size = ParallelCompactData::RegionSize; 1083 const ParallelCompactData& sd = summary_data(); 1084 1085 const MutableSpace* const space = _space_info[id].space(); 1086 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1087 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1088 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1089 1090 // Skip full regions at the beginning of the space--they are necessarily part 1091 // of the dense prefix. 1092 size_t full_count = 0; 1093 const RegionData* cp; 1094 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1095 ++full_count; 1096 } 1097 1098 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1099 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1100 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1101 if (maximum_compaction || cp == end_cp || interval_ended) { 1102 _maximum_compaction_gc_num = total_invocations(); 1103 return sd.region_to_addr(cp); 1104 } 1105 1106 HeapWord* const new_top = _space_info[id].new_top(); 1107 const size_t space_live = pointer_delta(new_top, space->bottom()); 1108 const size_t space_used = space->used_in_words(); 1109 const size_t space_capacity = space->capacity_in_words(); 1110 1111 const double cur_density = double(space_live) / space_capacity; 1112 const double deadwood_density = 1113 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1114 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1115 1116 log_develop_debug(gc, compaction)( 1117 "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1118 cur_density, deadwood_density, deadwood_goal); 1119 log_develop_debug(gc, compaction)( 1120 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " " 1121 "space_cap=" SIZE_FORMAT, 1122 space_live, space_used, 1123 space_capacity); 1124 1125 // XXX - Use binary search? 1126 HeapWord* dense_prefix = sd.region_to_addr(cp); 1127 const RegionData* full_cp = cp; 1128 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1129 while (cp < end_cp) { 1130 HeapWord* region_destination = cp->destination(); 1131 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1132 1133 log_develop_trace(gc, compaction)( 1134 "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1135 "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8), 1136 sd.region(cp), p2i(region_destination), 1137 p2i(dense_prefix), cur_deadwood); 1138 1139 if (cur_deadwood >= deadwood_goal) { 1140 // Found the region that has the correct amount of deadwood to the left. 1141 // This typically occurs after crossing a fairly sparse set of regions, so 1142 // iterate backwards over those sparse regions, looking for the region 1143 // that has the lowest density of live objects 'to the right.' 1144 size_t space_to_left = sd.region(cp) * region_size; 1145 size_t live_to_left = space_to_left - cur_deadwood; 1146 size_t space_to_right = space_capacity - space_to_left; 1147 size_t live_to_right = space_live - live_to_left; 1148 double density_to_right = double(live_to_right) / space_to_right; 1149 while (cp > full_cp) { 1150 --cp; 1151 const size_t prev_region_live_to_right = live_to_right - 1152 cp->data_size(); 1153 const size_t prev_region_space_to_right = space_to_right + region_size; 1154 double prev_region_density_to_right = 1155 double(prev_region_live_to_right) / prev_region_space_to_right; 1156 if (density_to_right <= prev_region_density_to_right) { 1157 return dense_prefix; 1158 } 1159 1160 log_develop_trace(gc, compaction)( 1161 "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1162 "pc_d2r=%10.8f", 1163 sd.region(cp), density_to_right, 1164 prev_region_density_to_right); 1165 1166 dense_prefix -= region_size; 1167 live_to_right = prev_region_live_to_right; 1168 space_to_right = prev_region_space_to_right; 1169 density_to_right = prev_region_density_to_right; 1170 } 1171 return dense_prefix; 1172 } 1173 1174 dense_prefix += region_size; 1175 ++cp; 1176 } 1177 1178 return dense_prefix; 1179 } 1180 1181 #ifndef PRODUCT 1182 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1183 const SpaceId id, 1184 const bool maximum_compaction, 1185 HeapWord* const addr) 1186 { 1187 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1188 RegionData* const cp = summary_data().region(region_idx); 1189 const MutableSpace* const space = _space_info[id].space(); 1190 HeapWord* const new_top = _space_info[id].new_top(); 1191 1192 const size_t space_live = pointer_delta(new_top, space->bottom()); 1193 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1194 const size_t space_cap = space->capacity_in_words(); 1195 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1196 const size_t live_to_right = new_top - cp->destination(); 1197 const size_t dead_to_right = space->top() - addr - live_to_right; 1198 1199 log_develop_debug(gc, compaction)( 1200 "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1201 "spl=" SIZE_FORMAT " " 1202 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1203 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " " 1204 "ratio=%10.8f", 1205 algorithm, p2i(addr), region_idx, 1206 space_live, 1207 dead_to_left, dead_to_left_pct, 1208 dead_to_right, live_to_right, 1209 double(dead_to_right) / live_to_right); 1210 } 1211 #endif // #ifndef PRODUCT 1212 1213 // Return a fraction indicating how much of the generation can be treated as 1214 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1215 // based on the density of live objects in the generation to determine a limit, 1216 // which is then adjusted so the return value is min_percent when the density is 1217 // 1. 1218 // 1219 // The following table shows some return values for a different values of the 1220 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1221 // min_percent is 1. 1222 // 1223 // fraction allowed as dead wood 1224 // ----------------------------------------------------------------- 1225 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1226 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1227 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1228 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1229 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1230 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1231 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1232 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1233 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1234 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1235 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1236 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1237 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1238 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1239 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1240 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1241 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1242 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1243 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1244 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1245 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1246 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1247 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1248 1249 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1250 { 1251 assert(_dwl_initialized, "uninitialized"); 1252 1253 // The raw limit is the value of the normal distribution at x = density. 1254 const double raw_limit = normal_distribution(density); 1255 1256 // Adjust the raw limit so it becomes the minimum when the density is 1. 1257 // 1258 // First subtract the adjustment value (which is simply the precomputed value 1259 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1260 // Then add the minimum value, so the minimum is returned when the density is 1261 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1262 const double min = double(min_percent) / 100.0; 1263 const double limit = raw_limit - _dwl_adjustment + min; 1264 return MAX2(limit, 0.0); 1265 } 1266 1267 ParallelCompactData::RegionData* 1268 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1269 const RegionData* end) 1270 { 1271 const size_t region_size = ParallelCompactData::RegionSize; 1272 ParallelCompactData& sd = summary_data(); 1273 size_t left = sd.region(beg); 1274 size_t right = end > beg ? sd.region(end) - 1 : left; 1275 1276 // Binary search. 1277 while (left < right) { 1278 // Equivalent to (left + right) / 2, but does not overflow. 1279 const size_t middle = left + (right - left) / 2; 1280 RegionData* const middle_ptr = sd.region(middle); 1281 HeapWord* const dest = middle_ptr->destination(); 1282 HeapWord* const addr = sd.region_to_addr(middle); 1283 assert(dest != NULL, "sanity"); 1284 assert(dest <= addr, "must move left"); 1285 1286 if (middle > left && dest < addr) { 1287 right = middle - 1; 1288 } else if (middle < right && middle_ptr->data_size() == region_size) { 1289 left = middle + 1; 1290 } else { 1291 return middle_ptr; 1292 } 1293 } 1294 return sd.region(left); 1295 } 1296 1297 ParallelCompactData::RegionData* 1298 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1299 const RegionData* end, 1300 size_t dead_words) 1301 { 1302 ParallelCompactData& sd = summary_data(); 1303 size_t left = sd.region(beg); 1304 size_t right = end > beg ? sd.region(end) - 1 : left; 1305 1306 // Binary search. 1307 while (left < right) { 1308 // Equivalent to (left + right) / 2, but does not overflow. 1309 const size_t middle = left + (right - left) / 2; 1310 RegionData* const middle_ptr = sd.region(middle); 1311 HeapWord* const dest = middle_ptr->destination(); 1312 HeapWord* const addr = sd.region_to_addr(middle); 1313 assert(dest != NULL, "sanity"); 1314 assert(dest <= addr, "must move left"); 1315 1316 const size_t dead_to_left = pointer_delta(addr, dest); 1317 if (middle > left && dead_to_left > dead_words) { 1318 right = middle - 1; 1319 } else if (middle < right && dead_to_left < dead_words) { 1320 left = middle + 1; 1321 } else { 1322 return middle_ptr; 1323 } 1324 } 1325 return sd.region(left); 1326 } 1327 1328 // The result is valid during the summary phase, after the initial summarization 1329 // of each space into itself, and before final summarization. 1330 inline double 1331 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1332 HeapWord* const bottom, 1333 HeapWord* const top, 1334 HeapWord* const new_top) 1335 { 1336 ParallelCompactData& sd = summary_data(); 1337 1338 assert(cp != NULL, "sanity"); 1339 assert(bottom != NULL, "sanity"); 1340 assert(top != NULL, "sanity"); 1341 assert(new_top != NULL, "sanity"); 1342 assert(top >= new_top, "summary data problem?"); 1343 assert(new_top > bottom, "space is empty; should not be here"); 1344 assert(new_top >= cp->destination(), "sanity"); 1345 assert(top >= sd.region_to_addr(cp), "sanity"); 1346 1347 HeapWord* const destination = cp->destination(); 1348 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1349 const size_t compacted_region_live = pointer_delta(new_top, destination); 1350 const size_t compacted_region_used = pointer_delta(top, 1351 sd.region_to_addr(cp)); 1352 const size_t reclaimable = compacted_region_used - compacted_region_live; 1353 1354 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1355 return double(reclaimable) / divisor; 1356 } 1357 1358 // Return the address of the end of the dense prefix, a.k.a. the start of the 1359 // compacted region. The address is always on a region boundary. 1360 // 1361 // Completely full regions at the left are skipped, since no compaction can 1362 // occur in those regions. Then the maximum amount of dead wood to allow is 1363 // computed, based on the density (amount live / capacity) of the generation; 1364 // the region with approximately that amount of dead space to the left is 1365 // identified as the limit region. Regions between the last completely full 1366 // region and the limit region are scanned and the one that has the best 1367 // (maximum) reclaimed_ratio() is selected. 1368 HeapWord* 1369 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1370 bool maximum_compaction) 1371 { 1372 const size_t region_size = ParallelCompactData::RegionSize; 1373 const ParallelCompactData& sd = summary_data(); 1374 1375 const MutableSpace* const space = _space_info[id].space(); 1376 HeapWord* const top = space->top(); 1377 HeapWord* const top_aligned_up = sd.region_align_up(top); 1378 HeapWord* const new_top = _space_info[id].new_top(); 1379 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1380 HeapWord* const bottom = space->bottom(); 1381 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1382 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1383 const RegionData* const new_top_cp = 1384 sd.addr_to_region_ptr(new_top_aligned_up); 1385 1386 // Skip full regions at the beginning of the space--they are necessarily part 1387 // of the dense prefix. 1388 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1389 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1390 space->is_empty(), "no dead space allowed to the left"); 1391 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1392 "region must have dead space"); 1393 1394 // The gc number is saved whenever a maximum compaction is done, and used to 1395 // determine when the maximum compaction interval has expired. This avoids 1396 // successive max compactions for different reasons. 1397 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1398 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1399 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1400 total_invocations() == HeapFirstMaximumCompactionCount; 1401 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1402 _maximum_compaction_gc_num = total_invocations(); 1403 return sd.region_to_addr(full_cp); 1404 } 1405 1406 const size_t space_live = pointer_delta(new_top, bottom); 1407 const size_t space_used = space->used_in_words(); 1408 const size_t space_capacity = space->capacity_in_words(); 1409 1410 const double density = double(space_live) / double(space_capacity); 1411 const size_t min_percent_free = MarkSweepDeadRatio; 1412 const double limiter = dead_wood_limiter(density, min_percent_free); 1413 const size_t dead_wood_max = space_used - space_live; 1414 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1415 dead_wood_max); 1416 1417 log_develop_debug(gc, compaction)( 1418 "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " " 1419 "space_cap=" SIZE_FORMAT, 1420 space_live, space_used, 1421 space_capacity); 1422 log_develop_debug(gc, compaction)( 1423 "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1424 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1425 density, min_percent_free, limiter, 1426 dead_wood_max, dead_wood_limit); 1427 1428 // Locate the region with the desired amount of dead space to the left. 1429 const RegionData* const limit_cp = 1430 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1431 1432 // Scan from the first region with dead space to the limit region and find the 1433 // one with the best (largest) reclaimed ratio. 1434 double best_ratio = 0.0; 1435 const RegionData* best_cp = full_cp; 1436 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1437 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1438 if (tmp_ratio > best_ratio) { 1439 best_cp = cp; 1440 best_ratio = tmp_ratio; 1441 } 1442 } 1443 1444 return sd.region_to_addr(best_cp); 1445 } 1446 1447 void PSParallelCompact::summarize_spaces_quick() 1448 { 1449 for (unsigned int i = 0; i < last_space_id; ++i) { 1450 const MutableSpace* space = _space_info[i].space(); 1451 HeapWord** nta = _space_info[i].new_top_addr(); 1452 bool result = _summary_data.summarize(_space_info[i].split_info(), 1453 space->bottom(), space->top(), NULL, 1454 space->bottom(), space->end(), nta); 1455 assert(result, "space must fit into itself"); 1456 _space_info[i].set_dense_prefix(space->bottom()); 1457 } 1458 } 1459 1460 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1461 { 1462 HeapWord* const dense_prefix_end = dense_prefix(id); 1463 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1464 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1465 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1466 // Only enough dead space is filled so that any remaining dead space to the 1467 // left is larger than the minimum filler object. (The remainder is filled 1468 // during the copy/update phase.) 1469 // 1470 // The size of the dead space to the right of the boundary is not a 1471 // concern, since compaction will be able to use whatever space is 1472 // available. 1473 // 1474 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1475 // surrounds the space to be filled with an object. 1476 // 1477 // In the 32-bit VM, each bit represents two 32-bit words: 1478 // +---+ 1479 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1480 // end_bits: ... x x x | 0 | || 0 x x ... 1481 // +---+ 1482 // 1483 // In the 64-bit VM, each bit represents one 64-bit word: 1484 // +------------+ 1485 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1486 // end_bits: ... x x 1 | 0 || 0 | x x ... 1487 // +------------+ 1488 // +-------+ 1489 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1490 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1491 // +-------+ 1492 // +-----------+ 1493 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1494 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1495 // +-----------+ 1496 // +-------+ 1497 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1498 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1499 // +-------+ 1500 1501 // Initially assume case a, c or e will apply. 1502 size_t obj_len = CollectedHeap::min_fill_size(); 1503 HeapWord* obj_beg = dense_prefix_end - obj_len; 1504 1505 #ifdef _LP64 1506 if (MinObjAlignment > 1) { // object alignment > heap word size 1507 // Cases a, c or e. 1508 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1509 // Case b above. 1510 obj_beg = dense_prefix_end - 1; 1511 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1512 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1513 // Case d above. 1514 obj_beg = dense_prefix_end - 3; 1515 obj_len = 3; 1516 } 1517 #endif // #ifdef _LP64 1518 1519 CollectedHeap::fill_with_object(obj_beg, obj_len); 1520 _mark_bitmap.mark_obj(obj_beg, obj_len); 1521 _summary_data.add_obj(obj_beg, obj_len); 1522 assert(start_array(id) != NULL, "sanity"); 1523 start_array(id)->allocate_block(obj_beg); 1524 } 1525 } 1526 1527 void 1528 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1529 { 1530 assert(id < last_space_id, "id out of range"); 1531 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1532 "should have been reset in summarize_spaces_quick()"); 1533 1534 const MutableSpace* space = _space_info[id].space(); 1535 if (_space_info[id].new_top() != space->bottom()) { 1536 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1537 _space_info[id].set_dense_prefix(dense_prefix_end); 1538 1539 #ifndef PRODUCT 1540 if (log_is_enabled(Debug, gc, compaction)) { 1541 print_dense_prefix_stats("ratio", id, maximum_compaction, 1542 dense_prefix_end); 1543 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1544 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1545 } 1546 #endif // #ifndef PRODUCT 1547 1548 // Recompute the summary data, taking into account the dense prefix. If 1549 // every last byte will be reclaimed, then the existing summary data which 1550 // compacts everything can be left in place. 1551 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1552 // If dead space crosses the dense prefix boundary, it is (at least 1553 // partially) filled with a dummy object, marked live and added to the 1554 // summary data. This simplifies the copy/update phase and must be done 1555 // before the final locations of objects are determined, to prevent 1556 // leaving a fragment of dead space that is too small to fill. 1557 fill_dense_prefix_end(id); 1558 1559 // Compute the destination of each Region, and thus each object. 1560 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1561 _summary_data.summarize(_space_info[id].split_info(), 1562 dense_prefix_end, space->top(), NULL, 1563 dense_prefix_end, space->end(), 1564 _space_info[id].new_top_addr()); 1565 } 1566 } 1567 1568 if (log_develop_is_enabled(Trace, gc, compaction)) { 1569 const size_t region_size = ParallelCompactData::RegionSize; 1570 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1571 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1572 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1573 HeapWord* const new_top = _space_info[id].new_top(); 1574 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1575 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1576 log_develop_trace(gc, compaction)( 1577 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1578 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1579 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1580 id, space->capacity_in_words(), p2i(dense_prefix_end), 1581 dp_region, dp_words / region_size, 1582 cr_words / region_size, p2i(new_top)); 1583 } 1584 } 1585 1586 #ifndef PRODUCT 1587 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1588 HeapWord* dst_beg, HeapWord* dst_end, 1589 SpaceId src_space_id, 1590 HeapWord* src_beg, HeapWord* src_end) 1591 { 1592 log_develop_trace(gc, compaction)( 1593 "Summarizing %d [%s] into %d [%s]: " 1594 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1595 SIZE_FORMAT "-" SIZE_FORMAT " " 1596 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1597 SIZE_FORMAT "-" SIZE_FORMAT, 1598 src_space_id, space_names[src_space_id], 1599 dst_space_id, space_names[dst_space_id], 1600 p2i(src_beg), p2i(src_end), 1601 _summary_data.addr_to_region_idx(src_beg), 1602 _summary_data.addr_to_region_idx(src_end), 1603 p2i(dst_beg), p2i(dst_end), 1604 _summary_data.addr_to_region_idx(dst_beg), 1605 _summary_data.addr_to_region_idx(dst_end)); 1606 } 1607 #endif // #ifndef PRODUCT 1608 1609 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1610 bool maximum_compaction) 1611 { 1612 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1613 1614 log_develop_debug(gc, marking)( 1615 "add_obj_count=" SIZE_FORMAT " " 1616 "add_obj_bytes=" SIZE_FORMAT, 1617 add_obj_count, 1618 add_obj_size * HeapWordSize); 1619 log_develop_debug(gc, marking)( 1620 "mark_bitmap_count=" SIZE_FORMAT " " 1621 "mark_bitmap_bytes=" SIZE_FORMAT, 1622 mark_bitmap_count, 1623 mark_bitmap_size * HeapWordSize); 1624 1625 // Quick summarization of each space into itself, to see how much is live. 1626 summarize_spaces_quick(); 1627 1628 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1629 NOT_PRODUCT(print_region_ranges()); 1630 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1631 1632 // The amount of live data that will end up in old space (assuming it fits). 1633 size_t old_space_total_live = 0; 1634 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1635 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1636 _space_info[id].space()->bottom()); 1637 } 1638 1639 MutableSpace* const old_space = _space_info[old_space_id].space(); 1640 const size_t old_capacity = old_space->capacity_in_words(); 1641 if (old_space_total_live > old_capacity) { 1642 // XXX - should also try to expand 1643 maximum_compaction = true; 1644 } 1645 1646 // Old generations. 1647 summarize_space(old_space_id, maximum_compaction); 1648 1649 // Summarize the remaining spaces in the young gen. The initial target space 1650 // is the old gen. If a space does not fit entirely into the target, then the 1651 // remainder is compacted into the space itself and that space becomes the new 1652 // target. 1653 SpaceId dst_space_id = old_space_id; 1654 HeapWord* dst_space_end = old_space->end(); 1655 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1656 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1657 const MutableSpace* space = _space_info[id].space(); 1658 const size_t live = pointer_delta(_space_info[id].new_top(), 1659 space->bottom()); 1660 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1661 1662 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1663 SpaceId(id), space->bottom(), space->top());) 1664 if (live > 0 && live <= available) { 1665 // All the live data will fit. 1666 bool done = _summary_data.summarize(_space_info[id].split_info(), 1667 space->bottom(), space->top(), 1668 NULL, 1669 *new_top_addr, dst_space_end, 1670 new_top_addr); 1671 assert(done, "space must fit into old gen"); 1672 1673 // Reset the new_top value for the space. 1674 _space_info[id].set_new_top(space->bottom()); 1675 } else if (live > 0) { 1676 // Attempt to fit part of the source space into the target space. 1677 HeapWord* next_src_addr = NULL; 1678 bool done = _summary_data.summarize(_space_info[id].split_info(), 1679 space->bottom(), space->top(), 1680 &next_src_addr, 1681 *new_top_addr, dst_space_end, 1682 new_top_addr); 1683 assert(!done, "space should not fit into old gen"); 1684 assert(next_src_addr != NULL, "sanity"); 1685 1686 // The source space becomes the new target, so the remainder is compacted 1687 // within the space itself. 1688 dst_space_id = SpaceId(id); 1689 dst_space_end = space->end(); 1690 new_top_addr = _space_info[id].new_top_addr(); 1691 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1692 space->bottom(), dst_space_end, 1693 SpaceId(id), next_src_addr, space->top());) 1694 done = _summary_data.summarize(_space_info[id].split_info(), 1695 next_src_addr, space->top(), 1696 NULL, 1697 space->bottom(), dst_space_end, 1698 new_top_addr); 1699 assert(done, "space must fit when compacted into itself"); 1700 assert(*new_top_addr <= space->top(), "usage should not grow"); 1701 } 1702 } 1703 1704 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1705 NOT_PRODUCT(print_region_ranges()); 1706 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1707 } 1708 1709 // This method should contain all heap-specific policy for invoking a full 1710 // collection. invoke_no_policy() will only attempt to compact the heap; it 1711 // will do nothing further. If we need to bail out for policy reasons, scavenge 1712 // before full gc, or any other specialized behavior, it needs to be added here. 1713 // 1714 // Note that this method should only be called from the vm_thread while at a 1715 // safepoint. 1716 // 1717 // Note that the all_soft_refs_clear flag in the soft ref policy 1718 // may be true because this method can be called without intervening 1719 // activity. For example when the heap space is tight and full measure 1720 // are being taken to free space. 1721 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1722 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1723 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1724 "should be in vm thread"); 1725 1726 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1727 GCCause::Cause gc_cause = heap->gc_cause(); 1728 assert(!heap->is_gc_active(), "not reentrant"); 1729 1730 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1731 IsGCActiveMark mark; 1732 1733 if (ScavengeBeforeFullGC) { 1734 PSScavenge::invoke_no_policy(); 1735 } 1736 1737 const bool clear_all_soft_refs = 1738 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1739 1740 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1741 maximum_heap_compaction); 1742 } 1743 1744 // This method contains no policy. You should probably 1745 // be calling invoke() instead. 1746 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1747 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1748 assert(ref_processor() != NULL, "Sanity"); 1749 1750 if (GCLocker::check_active_before_gc()) { 1751 return false; 1752 } 1753 1754 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1755 1756 GCIdMark gc_id_mark; 1757 _gc_timer.register_gc_start(); 1758 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1759 1760 TimeStamp marking_start; 1761 TimeStamp compaction_start; 1762 TimeStamp collection_exit; 1763 1764 GCCause::Cause gc_cause = heap->gc_cause(); 1765 PSYoungGen* young_gen = heap->young_gen(); 1766 PSOldGen* old_gen = heap->old_gen(); 1767 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1768 1769 // The scope of casr should end after code that can change 1770 // SoftRefPolicy::_should_clear_all_soft_refs. 1771 ClearedAllSoftRefs casr(maximum_heap_compaction, 1772 heap->soft_ref_policy()); 1773 1774 if (ZapUnusedHeapArea) { 1775 // Save information needed to minimize mangling 1776 heap->record_gen_tops_before_GC(); 1777 } 1778 1779 // Make sure data structures are sane, make the heap parsable, and do other 1780 // miscellaneous bookkeeping. 1781 pre_compact(); 1782 1783 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values(); 1784 1785 // Get the compaction manager reserved for the VM thread. 1786 ParCompactionManager* const vmthread_cm = 1787 ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers()); 1788 1789 { 1790 ResourceMark rm; 1791 HandleMark hm; 1792 1793 const uint active_workers = 1794 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(), 1795 ParallelScavengeHeap::heap()->workers().active_workers(), 1796 Threads::number_of_non_daemon_threads()); 1797 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers); 1798 1799 GCTraceCPUTime tcpu; 1800 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1801 1802 heap->pre_full_gc_dump(&_gc_timer); 1803 1804 TraceCollectorStats tcs(counters()); 1805 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1806 1807 if (log_is_enabled(Debug, gc, heap, exit)) { 1808 accumulated_time()->start(); 1809 } 1810 1811 // Let the size policy know we're starting 1812 size_policy->major_collection_begin(); 1813 1814 #if COMPILER2_OR_JVMCI 1815 DerivedPointerTable::clear(); 1816 #endif 1817 1818 ref_processor()->enable_discovery(); 1819 ref_processor()->setup_policy(maximum_heap_compaction); 1820 1821 bool marked_for_unloading = false; 1822 1823 marking_start.update(); 1824 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1825 1826 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1827 && GCCause::is_user_requested_gc(gc_cause); 1828 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1829 1830 #if COMPILER2_OR_JVMCI 1831 assert(DerivedPointerTable::is_active(), "Sanity"); 1832 DerivedPointerTable::set_active(false); 1833 #endif 1834 1835 // adjust_roots() updates Universe::_intArrayKlassObj which is 1836 // needed by the compaction for filling holes in the dense prefix. 1837 adjust_roots(vmthread_cm); 1838 1839 compaction_start.update(); 1840 compact(); 1841 1842 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1843 // done before resizing. 1844 post_compact(); 1845 1846 // Let the size policy know we're done 1847 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1848 1849 if (UseAdaptiveSizePolicy) { 1850 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1851 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1852 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1853 1854 // Don't check if the size_policy is ready here. Let 1855 // the size_policy check that internally. 1856 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1857 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1858 // Swap the survivor spaces if from_space is empty. The 1859 // resize_young_gen() called below is normally used after 1860 // a successful young GC and swapping of survivor spaces; 1861 // otherwise, it will fail to resize the young gen with 1862 // the current implementation. 1863 if (young_gen->from_space()->is_empty()) { 1864 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1865 young_gen->swap_spaces(); 1866 } 1867 1868 // Calculate optimal free space amounts 1869 assert(young_gen->max_gen_size() > 1870 young_gen->from_space()->capacity_in_bytes() + 1871 young_gen->to_space()->capacity_in_bytes(), 1872 "Sizes of space in young gen are out-of-bounds"); 1873 1874 size_t young_live = young_gen->used_in_bytes(); 1875 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1876 size_t old_live = old_gen->used_in_bytes(); 1877 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1878 size_t max_old_gen_size = old_gen->max_gen_size(); 1879 size_t max_eden_size = young_gen->max_gen_size() - 1880 young_gen->from_space()->capacity_in_bytes() - 1881 young_gen->to_space()->capacity_in_bytes(); 1882 1883 // Used for diagnostics 1884 size_policy->clear_generation_free_space_flags(); 1885 1886 size_policy->compute_generations_free_space(young_live, 1887 eden_live, 1888 old_live, 1889 cur_eden, 1890 max_old_gen_size, 1891 max_eden_size, 1892 true /* full gc*/); 1893 1894 size_policy->check_gc_overhead_limit(eden_live, 1895 max_old_gen_size, 1896 max_eden_size, 1897 true /* full gc*/, 1898 gc_cause, 1899 heap->soft_ref_policy()); 1900 1901 size_policy->decay_supplemental_growth(true /* full gc*/); 1902 1903 heap->resize_old_gen( 1904 size_policy->calculated_old_free_size_in_bytes()); 1905 1906 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1907 size_policy->calculated_survivor_size_in_bytes()); 1908 } 1909 1910 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1911 } 1912 1913 if (UsePerfData) { 1914 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1915 counters->update_counters(); 1916 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1917 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1918 } 1919 1920 heap->resize_all_tlabs(); 1921 1922 // Resize the metaspace capacity after a collection 1923 MetaspaceGC::compute_new_size(); 1924 1925 if (log_is_enabled(Debug, gc, heap, exit)) { 1926 accumulated_time()->stop(); 1927 } 1928 1929 heap->print_heap_change(pre_gc_values); 1930 1931 // Track memory usage and detect low memory 1932 MemoryService::track_memory_usage(); 1933 heap->update_counters(); 1934 1935 heap->post_full_gc_dump(&_gc_timer); 1936 } 1937 1938 #ifdef ASSERT 1939 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1940 ParCompactionManager* const cm = 1941 ParCompactionManager::manager_array(int(i)); 1942 assert(cm->marking_stack()->is_empty(), "should be empty"); 1943 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1944 } 1945 #endif // ASSERT 1946 1947 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1948 HandleMark hm; // Discard invalid handles created during verification 1949 Universe::verify("After GC"); 1950 } 1951 1952 // Re-verify object start arrays 1953 if (VerifyObjectStartArray && 1954 VerifyAfterGC) { 1955 old_gen->verify_object_start_array(); 1956 } 1957 1958 if (ZapUnusedHeapArea) { 1959 old_gen->object_space()->check_mangled_unused_area_complete(); 1960 } 1961 1962 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1963 1964 collection_exit.update(); 1965 1966 heap->print_heap_after_gc(); 1967 heap->trace_heap_after_gc(&_gc_tracer); 1968 1969 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1970 marking_start.ticks(), compaction_start.ticks(), 1971 collection_exit.ticks()); 1972 1973 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1974 1975 _gc_timer.register_gc_end(); 1976 1977 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1978 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1979 1980 return true; 1981 } 1982 1983 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 1984 private: 1985 uint _worker_id; 1986 1987 public: 1988 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { } 1989 void do_thread(Thread* thread) { 1990 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 1991 1992 ResourceMark rm; 1993 1994 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id); 1995 1996 PCMarkAndPushClosure mark_and_push_closure(cm); 1997 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations); 1998 1999 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs); 2000 2001 // Do the real work 2002 cm->follow_marking_stacks(); 2003 } 2004 }; 2005 2006 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) { 2007 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2008 2009 ParCompactionManager* cm = 2010 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2011 PCMarkAndPushClosure mark_and_push_closure(cm); 2012 2013 switch (root_type) { 2014 case ParallelRootType::universe: 2015 Universe::oops_do(&mark_and_push_closure); 2016 break; 2017 2018 case ParallelRootType::object_synchronizer: 2019 ObjectSynchronizer::oops_do(&mark_and_push_closure); 2020 break; 2021 2022 case ParallelRootType::class_loader_data: 2023 { 2024 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong); 2025 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure); 2026 } 2027 break; 2028 2029 case ParallelRootType::code_cache: 2030 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them. 2031 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure)); 2032 AOTLoader::oops_do(&mark_and_push_closure); 2033 break; 2034 2035 case ParallelRootType::sentinel: 2036 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds 2037 fatal("Bad enumeration value: %u", root_type); 2038 break; 2039 } 2040 2041 // Do the real work 2042 cm->follow_marking_stacks(); 2043 } 2044 2045 static void steal_marking_work(TaskTerminator& terminator, uint worker_id) { 2046 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2047 2048 ParCompactionManager* cm = 2049 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2050 2051 oop obj = NULL; 2052 ObjArrayTask task; 2053 do { 2054 while (ParCompactionManager::steal_objarray(worker_id, task)) { 2055 cm->follow_array((objArrayOop)task.obj(), task.index()); 2056 cm->follow_marking_stacks(); 2057 } 2058 while (ParCompactionManager::steal(worker_id, obj)) { 2059 cm->follow_contents(obj); 2060 cm->follow_marking_stacks(); 2061 } 2062 } while (!terminator.offer_termination()); 2063 } 2064 2065 class MarkFromRootsTask : public AbstractGangTask { 2066 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2067 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do 2068 OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state; 2069 SequentialSubTasksDone _subtasks; 2070 TaskTerminator _terminator; 2071 uint _active_workers; 2072 2073 public: 2074 MarkFromRootsTask(uint active_workers) : 2075 AbstractGangTask("MarkFromRootsTask"), 2076 _strong_roots_scope(active_workers), 2077 _subtasks(), 2078 _terminator(active_workers, ParCompactionManager::oop_task_queues()), 2079 _active_workers(active_workers) { 2080 _subtasks.set_n_threads(active_workers); 2081 _subtasks.set_n_tasks(ParallelRootType::sentinel); 2082 } 2083 2084 virtual void work(uint worker_id) { 2085 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) { 2086 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id); 2087 } 2088 _subtasks.all_tasks_completed(); 2089 2090 PCAddThreadRootsMarkingTaskClosure closure(worker_id); 2091 Threads::possibly_parallel_threads_do(true /*parallel */, &closure); 2092 2093 // Mark from OopStorages 2094 { 2095 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2096 PCMarkAndPushClosure closure(cm); 2097 _oop_storage_set_par_state.oops_do(&closure); 2098 // Do the real work 2099 cm->follow_marking_stacks(); 2100 } 2101 2102 if (_active_workers > 1) { 2103 steal_marking_work(_terminator, worker_id); 2104 } 2105 } 2106 }; 2107 2108 class PCRefProcTask : public AbstractGangTask { 2109 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2110 ProcessTask& _task; 2111 uint _ergo_workers; 2112 TaskTerminator _terminator; 2113 2114 public: 2115 PCRefProcTask(ProcessTask& task, uint ergo_workers) : 2116 AbstractGangTask("PCRefProcTask"), 2117 _task(task), 2118 _ergo_workers(ergo_workers), 2119 _terminator(_ergo_workers, ParCompactionManager::oop_task_queues()) { 2120 } 2121 2122 virtual void work(uint worker_id) { 2123 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2124 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2125 2126 ParCompactionManager* cm = 2127 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2128 PCMarkAndPushClosure mark_and_push_closure(cm); 2129 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2130 _task.work(worker_id, *PSParallelCompact::is_alive_closure(), 2131 mark_and_push_closure, follow_stack_closure); 2132 2133 steal_marking_work(_terminator, worker_id); 2134 } 2135 }; 2136 2137 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2138 void execute(ProcessTask& process_task, uint ergo_workers) { 2139 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers, 2140 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 2141 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers()); 2142 2143 PCRefProcTask task(process_task, ergo_workers); 2144 ParallelScavengeHeap::heap()->workers().run_task(&task); 2145 } 2146 }; 2147 2148 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2149 bool maximum_heap_compaction, 2150 ParallelOldTracer *gc_tracer) { 2151 // Recursively traverse all live objects and mark them 2152 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2153 2154 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2155 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2156 2157 PCMarkAndPushClosure mark_and_push_closure(cm); 2158 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2159 2160 // Need new claim bits before marking starts. 2161 ClassLoaderDataGraph::clear_claimed_marks(); 2162 2163 { 2164 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2165 2166 MarkFromRootsTask task(active_gc_threads); 2167 ParallelScavengeHeap::heap()->workers().run_task(&task); 2168 } 2169 2170 // Process reference objects found during marking 2171 { 2172 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2173 2174 ReferenceProcessorStats stats; 2175 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2176 2177 if (ref_processor()->processing_is_mt()) { 2178 ref_processor()->set_active_mt_degree(active_gc_threads); 2179 2180 RefProcTaskExecutor task_executor; 2181 stats = ref_processor()->process_discovered_references( 2182 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2183 &task_executor, &pt); 2184 } else { 2185 stats = ref_processor()->process_discovered_references( 2186 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2187 &pt); 2188 } 2189 2190 gc_tracer->report_gc_reference_stats(stats); 2191 pt.print_all_references(); 2192 } 2193 2194 // This is the point where the entire marking should have completed. 2195 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2196 2197 { 2198 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2199 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2200 } 2201 2202 { 2203 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2204 2205 // Follow system dictionary roots and unload classes. 2206 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2207 2208 // Unload nmethods. 2209 CodeCache::do_unloading(is_alive_closure(), purged_class); 2210 2211 // Prune dead klasses from subklass/sibling/implementor lists. 2212 Klass::clean_weak_klass_links(purged_class); 2213 2214 // Clean JVMCI metadata handles. 2215 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 2216 } 2217 2218 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2219 } 2220 2221 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2222 // Adjust the pointers to reflect the new locations 2223 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2224 2225 // Need new claim bits when tracing through and adjusting pointers. 2226 ClassLoaderDataGraph::clear_claimed_marks(); 2227 2228 PCAdjustPointerClosure oop_closure(cm); 2229 2230 // General strong roots. 2231 Universe::oops_do(&oop_closure); 2232 Threads::oops_do(&oop_closure, NULL); 2233 ObjectSynchronizer::oops_do(&oop_closure); 2234 OopStorageSet::strong_oops_do(&oop_closure); 2235 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong); 2236 ClassLoaderDataGraph::cld_do(&cld_closure); 2237 2238 // Now adjust pointers in remaining weak roots. (All of which should 2239 // have been cleared if they pointed to non-surviving objects.) 2240 WeakProcessor::oops_do(&oop_closure); 2241 2242 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2243 CodeCache::blobs_do(&adjust_from_blobs); 2244 AOT_ONLY(AOTLoader::oops_do(&oop_closure);) 2245 2246 ref_processor()->weak_oops_do(&oop_closure); 2247 // Roots were visited so references into the young gen in roots 2248 // may have been scanned. Process them also. 2249 // Should the reference processor have a span that excludes 2250 // young gen objects? 2251 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2252 } 2253 2254 // Helper class to print 8 region numbers per line and then print the total at the end. 2255 class FillableRegionLogger : public StackObj { 2256 private: 2257 Log(gc, compaction) log; 2258 static const int LineLength = 8; 2259 size_t _regions[LineLength]; 2260 int _next_index; 2261 bool _enabled; 2262 size_t _total_regions; 2263 public: 2264 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2265 ~FillableRegionLogger() { 2266 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2267 } 2268 2269 void print_line() { 2270 if (!_enabled || _next_index == 0) { 2271 return; 2272 } 2273 FormatBuffer<> line("Fillable: "); 2274 for (int i = 0; i < _next_index; i++) { 2275 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2276 } 2277 log.trace("%s", line.buffer()); 2278 _next_index = 0; 2279 } 2280 2281 void handle(size_t region) { 2282 if (!_enabled) { 2283 return; 2284 } 2285 _regions[_next_index++] = region; 2286 if (_next_index == LineLength) { 2287 print_line(); 2288 } 2289 _total_regions++; 2290 } 2291 }; 2292 2293 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads) 2294 { 2295 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2296 2297 // Find the threads that are active 2298 uint worker_id = 0; 2299 2300 // Find all regions that are available (can be filled immediately) and 2301 // distribute them to the thread stacks. The iteration is done in reverse 2302 // order (high to low) so the regions will be removed in ascending order. 2303 2304 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2305 2306 // id + 1 is used to test termination so unsigned can 2307 // be used with an old_space_id == 0. 2308 FillableRegionLogger region_logger; 2309 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2310 SpaceInfo* const space_info = _space_info + id; 2311 MutableSpace* const space = space_info->space(); 2312 HeapWord* const new_top = space_info->new_top(); 2313 2314 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2315 const size_t end_region = 2316 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2317 2318 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2319 if (sd.region(cur)->claim_unsafe()) { 2320 ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id); 2321 bool result = sd.region(cur)->mark_normal(); 2322 assert(result, "Must succeed at this point."); 2323 cm->region_stack()->push(cur); 2324 region_logger.handle(cur); 2325 // Assign regions to tasks in round-robin fashion. 2326 if (++worker_id == parallel_gc_threads) { 2327 worker_id = 0; 2328 } 2329 } 2330 } 2331 region_logger.print_line(); 2332 } 2333 } 2334 2335 class TaskQueue : StackObj { 2336 volatile uint _counter; 2337 uint _size; 2338 uint _insert_index; 2339 PSParallelCompact::UpdateDensePrefixTask* _backing_array; 2340 public: 2341 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) { 2342 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC); 2343 } 2344 ~TaskQueue() { 2345 assert(_counter >= _insert_index, "not all queue elements were claimed"); 2346 FREE_C_HEAP_ARRAY(T, _backing_array); 2347 } 2348 2349 void push(const PSParallelCompact::UpdateDensePrefixTask& value) { 2350 assert(_insert_index < _size, "too small backing array"); 2351 _backing_array[_insert_index++] = value; 2352 } 2353 2354 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) { 2355 uint claimed = Atomic::fetch_and_add(&_counter, 1u); 2356 if (claimed < _insert_index) { 2357 reference = _backing_array[claimed]; 2358 return true; 2359 } else { 2360 return false; 2361 } 2362 } 2363 }; 2364 2365 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2366 2367 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue, 2368 uint parallel_gc_threads) { 2369 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2370 2371 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2372 2373 // Iterate over all the spaces adding tasks for updating 2374 // regions in the dense prefix. Assume that 1 gc thread 2375 // will work on opening the gaps and the remaining gc threads 2376 // will work on the dense prefix. 2377 unsigned int space_id; 2378 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2379 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2380 const MutableSpace* const space = _space_info[space_id].space(); 2381 2382 if (dense_prefix_end == space->bottom()) { 2383 // There is no dense prefix for this space. 2384 continue; 2385 } 2386 2387 // The dense prefix is before this region. 2388 size_t region_index_end_dense_prefix = 2389 sd.addr_to_region_idx(dense_prefix_end); 2390 RegionData* const dense_prefix_cp = 2391 sd.region(region_index_end_dense_prefix); 2392 assert(dense_prefix_end == space->end() || 2393 dense_prefix_cp->available() || 2394 dense_prefix_cp->claimed(), 2395 "The region after the dense prefix should always be ready to fill"); 2396 2397 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2398 2399 // Is there dense prefix work? 2400 size_t total_dense_prefix_regions = 2401 region_index_end_dense_prefix - region_index_start; 2402 // How many regions of the dense prefix should be given to 2403 // each thread? 2404 if (total_dense_prefix_regions > 0) { 2405 uint tasks_for_dense_prefix = 1; 2406 if (total_dense_prefix_regions <= 2407 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2408 // Don't over partition. This assumes that 2409 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2410 // so there are not many regions to process. 2411 tasks_for_dense_prefix = parallel_gc_threads; 2412 } else { 2413 // Over partition 2414 tasks_for_dense_prefix = parallel_gc_threads * 2415 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2416 } 2417 size_t regions_per_thread = total_dense_prefix_regions / 2418 tasks_for_dense_prefix; 2419 // Give each thread at least 1 region. 2420 if (regions_per_thread == 0) { 2421 regions_per_thread = 1; 2422 } 2423 2424 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2425 if (region_index_start >= region_index_end_dense_prefix) { 2426 break; 2427 } 2428 // region_index_end is not processed 2429 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2430 region_index_end_dense_prefix); 2431 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2432 region_index_start, 2433 region_index_end)); 2434 region_index_start = region_index_end; 2435 } 2436 } 2437 // This gets any part of the dense prefix that did not 2438 // fit evenly. 2439 if (region_index_start < region_index_end_dense_prefix) { 2440 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2441 region_index_start, 2442 region_index_end_dense_prefix)); 2443 } 2444 } 2445 } 2446 2447 #ifdef ASSERT 2448 // Write a histogram of the number of times the block table was filled for a 2449 // region. 2450 void PSParallelCompact::write_block_fill_histogram() 2451 { 2452 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2453 return; 2454 } 2455 2456 Log(gc, compaction) log; 2457 ResourceMark rm; 2458 LogStream ls(log.trace()); 2459 outputStream* out = &ls; 2460 2461 typedef ParallelCompactData::RegionData rd_t; 2462 ParallelCompactData& sd = summary_data(); 2463 2464 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2465 MutableSpace* const spc = _space_info[id].space(); 2466 if (spc->bottom() != spc->top()) { 2467 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2468 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2469 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2470 2471 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2472 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2473 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2474 2475 for (const rd_t* cur = beg; cur < end; ++cur) { 2476 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2477 } 2478 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2479 for (size_t i = 0; i < histo_len; ++i) { 2480 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2481 histo[i], 100.0 * histo[i] / region_cnt); 2482 } 2483 out->cr(); 2484 } 2485 } 2486 } 2487 #endif // #ifdef ASSERT 2488 2489 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) { 2490 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2491 2492 ParCompactionManager* cm = 2493 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2494 2495 // Drain the stacks that have been preloaded with regions 2496 // that are ready to fill. 2497 2498 cm->drain_region_stacks(); 2499 2500 guarantee(cm->region_stack()->is_empty(), "Not empty"); 2501 2502 size_t region_index = 0; 2503 2504 while (true) { 2505 if (ParCompactionManager::steal(worker_id, region_index)) { 2506 PSParallelCompact::fill_and_update_region(cm, region_index); 2507 cm->drain_region_stacks(); 2508 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) { 2509 // Fill and update an unavailable region with the help of a shadow region 2510 PSParallelCompact::fill_and_update_shadow_region(cm, region_index); 2511 cm->drain_region_stacks(); 2512 } else { 2513 if (terminator->offer_termination()) { 2514 break; 2515 } 2516 // Go around again. 2517 } 2518 } 2519 return; 2520 } 2521 2522 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask { 2523 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2524 TaskQueue& _tq; 2525 TaskTerminator _terminator; 2526 uint _active_workers; 2527 2528 public: 2529 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) : 2530 AbstractGangTask("UpdateDensePrefixAndCompactionTask"), 2531 _tq(tq), 2532 _terminator(active_workers, ParCompactionManager::region_task_queues()), 2533 _active_workers(active_workers) { 2534 } 2535 virtual void work(uint worker_id) { 2536 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2537 2538 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) { 2539 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm, 2540 task._space_id, 2541 task._region_index_start, 2542 task._region_index_end); 2543 } 2544 2545 // Once a thread has drained it's stack, it should try to steal regions from 2546 // other threads. 2547 compaction_with_stealing_work(&_terminator, worker_id); 2548 } 2549 }; 2550 2551 void PSParallelCompact::compact() { 2552 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2553 2554 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2555 PSOldGen* old_gen = heap->old_gen(); 2556 old_gen->start_array()->reset(); 2557 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2558 2559 // for [0..last_space_id) 2560 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING) 2561 // push 2562 // push 2563 // 2564 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1) 2565 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)); 2566 initialize_shadow_regions(active_gc_threads); 2567 prepare_region_draining_tasks(active_gc_threads); 2568 enqueue_dense_prefix_tasks(task_queue, active_gc_threads); 2569 2570 { 2571 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2572 2573 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads); 2574 ParallelScavengeHeap::heap()->workers().run_task(&task); 2575 2576 #ifdef ASSERT 2577 // Verify that all regions have been processed before the deferred updates. 2578 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2579 verify_complete(SpaceId(id)); 2580 } 2581 #endif 2582 } 2583 2584 { 2585 // Update the deferred objects, if any. Any compaction manager can be used. 2586 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2587 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2588 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2589 update_deferred_objects(cm, SpaceId(id)); 2590 } 2591 } 2592 2593 DEBUG_ONLY(write_block_fill_histogram()); 2594 } 2595 2596 #ifdef ASSERT 2597 void PSParallelCompact::verify_complete(SpaceId space_id) { 2598 // All Regions between space bottom() to new_top() should be marked as filled 2599 // and all Regions between new_top() and top() should be available (i.e., 2600 // should have been emptied). 2601 ParallelCompactData& sd = summary_data(); 2602 SpaceInfo si = _space_info[space_id]; 2603 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2604 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2605 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2606 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2607 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2608 2609 bool issued_a_warning = false; 2610 2611 size_t cur_region; 2612 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2613 const RegionData* const c = sd.region(cur_region); 2614 if (!c->completed()) { 2615 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2616 cur_region, c->destination_count()); 2617 issued_a_warning = true; 2618 } 2619 } 2620 2621 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2622 const RegionData* const c = sd.region(cur_region); 2623 if (!c->available()) { 2624 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2625 cur_region, c->destination_count()); 2626 issued_a_warning = true; 2627 } 2628 } 2629 2630 if (issued_a_warning) { 2631 print_region_ranges(); 2632 } 2633 } 2634 #endif // #ifdef ASSERT 2635 2636 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2637 _start_array->allocate_block(addr); 2638 compaction_manager()->update_contents(oop(addr)); 2639 } 2640 2641 // Update interior oops in the ranges of regions [beg_region, end_region). 2642 void 2643 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2644 SpaceId space_id, 2645 size_t beg_region, 2646 size_t end_region) { 2647 ParallelCompactData& sd = summary_data(); 2648 ParMarkBitMap* const mbm = mark_bitmap(); 2649 2650 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2651 HeapWord* const end_addr = sd.region_to_addr(end_region); 2652 assert(beg_region <= end_region, "bad region range"); 2653 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2654 2655 #ifdef ASSERT 2656 // Claim the regions to avoid triggering an assert when they are marked as 2657 // filled. 2658 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2659 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2660 } 2661 #endif // #ifdef ASSERT 2662 2663 if (beg_addr != space(space_id)->bottom()) { 2664 // Find the first live object or block of dead space that *starts* in this 2665 // range of regions. If a partial object crosses onto the region, skip it; 2666 // it will be marked for 'deferred update' when the object head is 2667 // processed. If dead space crosses onto the region, it is also skipped; it 2668 // will be filled when the prior region is processed. If neither of those 2669 // apply, the first word in the region is the start of a live object or dead 2670 // space. 2671 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2672 const RegionData* const cp = sd.region(beg_region); 2673 if (cp->partial_obj_size() != 0) { 2674 beg_addr = sd.partial_obj_end(beg_region); 2675 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2676 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2677 } 2678 } 2679 2680 if (beg_addr < end_addr) { 2681 // A live object or block of dead space starts in this range of Regions. 2682 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2683 2684 // Create closures and iterate. 2685 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2686 FillClosure fill_closure(cm, space_id); 2687 ParMarkBitMap::IterationStatus status; 2688 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2689 dense_prefix_end); 2690 if (status == ParMarkBitMap::incomplete) { 2691 update_closure.do_addr(update_closure.source()); 2692 } 2693 } 2694 2695 // Mark the regions as filled. 2696 RegionData* const beg_cp = sd.region(beg_region); 2697 RegionData* const end_cp = sd.region(end_region); 2698 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2699 cp->set_completed(); 2700 } 2701 } 2702 2703 // Return the SpaceId for the space containing addr. If addr is not in the 2704 // heap, last_space_id is returned. In debug mode it expects the address to be 2705 // in the heap and asserts such. 2706 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2707 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2708 2709 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2710 if (_space_info[id].space()->contains(addr)) { 2711 return SpaceId(id); 2712 } 2713 } 2714 2715 assert(false, "no space contains the addr"); 2716 return last_space_id; 2717 } 2718 2719 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2720 SpaceId id) { 2721 assert(id < last_space_id, "bad space id"); 2722 2723 ParallelCompactData& sd = summary_data(); 2724 const SpaceInfo* const space_info = _space_info + id; 2725 ObjectStartArray* const start_array = space_info->start_array(); 2726 2727 const MutableSpace* const space = space_info->space(); 2728 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2729 HeapWord* const beg_addr = space_info->dense_prefix(); 2730 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2731 2732 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2733 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2734 const RegionData* cur_region; 2735 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2736 HeapWord* const addr = cur_region->deferred_obj_addr(); 2737 if (addr != NULL) { 2738 if (start_array != NULL) { 2739 start_array->allocate_block(addr); 2740 } 2741 cm->update_contents(oop(addr)); 2742 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2743 } 2744 } 2745 } 2746 2747 // Skip over count live words starting from beg, and return the address of the 2748 // next live word. Unless marked, the word corresponding to beg is assumed to 2749 // be dead. Callers must either ensure beg does not correspond to the middle of 2750 // an object, or account for those live words in some other way. Callers must 2751 // also ensure that there are enough live words in the range [beg, end) to skip. 2752 HeapWord* 2753 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2754 { 2755 assert(count > 0, "sanity"); 2756 2757 ParMarkBitMap* m = mark_bitmap(); 2758 idx_t bits_to_skip = m->words_to_bits(count); 2759 idx_t cur_beg = m->addr_to_bit(beg); 2760 const idx_t search_end = m->align_range_end(m->addr_to_bit(end)); 2761 2762 do { 2763 cur_beg = m->find_obj_beg(cur_beg, search_end); 2764 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2765 const size_t obj_bits = cur_end - cur_beg + 1; 2766 if (obj_bits > bits_to_skip) { 2767 return m->bit_to_addr(cur_beg + bits_to_skip); 2768 } 2769 bits_to_skip -= obj_bits; 2770 cur_beg = cur_end + 1; 2771 } while (bits_to_skip > 0); 2772 2773 // Skipping the desired number of words landed just past the end of an object. 2774 // Find the start of the next object. 2775 cur_beg = m->find_obj_beg(cur_beg, search_end); 2776 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2777 return m->bit_to_addr(cur_beg); 2778 } 2779 2780 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2781 SpaceId src_space_id, 2782 size_t src_region_idx) 2783 { 2784 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2785 2786 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2787 if (split_info.dest_region_addr() == dest_addr) { 2788 // The partial object ending at the split point contains the first word to 2789 // be copied to dest_addr. 2790 return split_info.first_src_addr(); 2791 } 2792 2793 const ParallelCompactData& sd = summary_data(); 2794 ParMarkBitMap* const bitmap = mark_bitmap(); 2795 const size_t RegionSize = ParallelCompactData::RegionSize; 2796 2797 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2798 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2799 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2800 HeapWord* const src_region_destination = src_region_ptr->destination(); 2801 2802 assert(dest_addr >= src_region_destination, "wrong src region"); 2803 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2804 2805 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2806 HeapWord* const src_region_end = src_region_beg + RegionSize; 2807 2808 HeapWord* addr = src_region_beg; 2809 if (dest_addr == src_region_destination) { 2810 // Return the first live word in the source region. 2811 if (partial_obj_size == 0) { 2812 addr = bitmap->find_obj_beg(addr, src_region_end); 2813 assert(addr < src_region_end, "no objects start in src region"); 2814 } 2815 return addr; 2816 } 2817 2818 // Must skip some live data. 2819 size_t words_to_skip = dest_addr - src_region_destination; 2820 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2821 2822 if (partial_obj_size >= words_to_skip) { 2823 // All the live words to skip are part of the partial object. 2824 addr += words_to_skip; 2825 if (partial_obj_size == words_to_skip) { 2826 // Find the first live word past the partial object. 2827 addr = bitmap->find_obj_beg(addr, src_region_end); 2828 assert(addr < src_region_end, "wrong src region"); 2829 } 2830 return addr; 2831 } 2832 2833 // Skip over the partial object (if any). 2834 if (partial_obj_size != 0) { 2835 words_to_skip -= partial_obj_size; 2836 addr += partial_obj_size; 2837 } 2838 2839 // Skip over live words due to objects that start in the region. 2840 addr = skip_live_words(addr, src_region_end, words_to_skip); 2841 assert(addr < src_region_end, "wrong src region"); 2842 return addr; 2843 } 2844 2845 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2846 SpaceId src_space_id, 2847 size_t beg_region, 2848 HeapWord* end_addr) 2849 { 2850 ParallelCompactData& sd = summary_data(); 2851 2852 #ifdef ASSERT 2853 MutableSpace* const src_space = _space_info[src_space_id].space(); 2854 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2855 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2856 "src_space_id does not match beg_addr"); 2857 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2858 "src_space_id does not match end_addr"); 2859 #endif // #ifdef ASSERT 2860 2861 RegionData* const beg = sd.region(beg_region); 2862 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2863 2864 // Regions up to new_top() are enqueued if they become available. 2865 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2866 RegionData* const enqueue_end = 2867 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2868 2869 for (RegionData* cur = beg; cur < end; ++cur) { 2870 assert(cur->data_size() > 0, "region must have live data"); 2871 cur->decrement_destination_count(); 2872 if (cur < enqueue_end && cur->available() && cur->claim()) { 2873 if (cur->mark_normal()) { 2874 cm->push_region(sd.region(cur)); 2875 } else if (cur->mark_copied()) { 2876 // Try to copy the content of the shadow region back to its corresponding 2877 // heap region if the shadow region is filled. Otherwise, the GC thread 2878 // fills the shadow region will copy the data back (see 2879 // MoveAndUpdateShadowClosure::complete_region). 2880 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur)); 2881 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region()); 2882 cur->set_completed(); 2883 } 2884 } 2885 } 2886 } 2887 2888 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2889 SpaceId& src_space_id, 2890 HeapWord*& src_space_top, 2891 HeapWord* end_addr) 2892 { 2893 typedef ParallelCompactData::RegionData RegionData; 2894 2895 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2896 const size_t region_size = ParallelCompactData::RegionSize; 2897 2898 size_t src_region_idx = 0; 2899 2900 // Skip empty regions (if any) up to the top of the space. 2901 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2902 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2903 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2904 const RegionData* const top_region_ptr = 2905 sd.addr_to_region_ptr(top_aligned_up); 2906 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2907 ++src_region_ptr; 2908 } 2909 2910 if (src_region_ptr < top_region_ptr) { 2911 // The next source region is in the current space. Update src_region_idx 2912 // and the source address to match src_region_ptr. 2913 src_region_idx = sd.region(src_region_ptr); 2914 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2915 if (src_region_addr > closure.source()) { 2916 closure.set_source(src_region_addr); 2917 } 2918 return src_region_idx; 2919 } 2920 2921 // Switch to a new source space and find the first non-empty region. 2922 unsigned int space_id = src_space_id + 1; 2923 assert(space_id < last_space_id, "not enough spaces"); 2924 2925 HeapWord* const destination = closure.destination(); 2926 2927 do { 2928 MutableSpace* space = _space_info[space_id].space(); 2929 HeapWord* const bottom = space->bottom(); 2930 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2931 2932 // Iterate over the spaces that do not compact into themselves. 2933 if (bottom_cp->destination() != bottom) { 2934 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2935 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2936 2937 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2938 if (src_cp->live_obj_size() > 0) { 2939 // Found it. 2940 assert(src_cp->destination() == destination, 2941 "first live obj in the space must match the destination"); 2942 assert(src_cp->partial_obj_size() == 0, 2943 "a space cannot begin with a partial obj"); 2944 2945 src_space_id = SpaceId(space_id); 2946 src_space_top = space->top(); 2947 const size_t src_region_idx = sd.region(src_cp); 2948 closure.set_source(sd.region_to_addr(src_region_idx)); 2949 return src_region_idx; 2950 } else { 2951 assert(src_cp->data_size() == 0, "sanity"); 2952 } 2953 } 2954 } 2955 } while (++space_id < last_space_id); 2956 2957 assert(false, "no source region was found"); 2958 return 0; 2959 } 2960 2961 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx) 2962 { 2963 typedef ParMarkBitMap::IterationStatus IterationStatus; 2964 ParMarkBitMap* const bitmap = mark_bitmap(); 2965 ParallelCompactData& sd = summary_data(); 2966 RegionData* const region_ptr = sd.region(region_idx); 2967 2968 // Get the source region and related info. 2969 size_t src_region_idx = region_ptr->source_region(); 2970 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2971 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2972 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2973 2974 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2975 2976 // Adjust src_region_idx to prepare for decrementing destination counts (the 2977 // destination count is not decremented when a region is copied to itself). 2978 if (src_region_idx == region_idx) { 2979 src_region_idx += 1; 2980 } 2981 2982 if (bitmap->is_unmarked(closure.source())) { 2983 // The first source word is in the middle of an object; copy the remainder 2984 // of the object or as much as will fit. The fact that pointer updates were 2985 // deferred will be noted when the object header is processed. 2986 HeapWord* const old_src_addr = closure.source(); 2987 closure.copy_partial_obj(); 2988 if (closure.is_full()) { 2989 decrement_destination_counts(cm, src_space_id, src_region_idx, 2990 closure.source()); 2991 region_ptr->set_deferred_obj_addr(NULL); 2992 closure.complete_region(cm, dest_addr, region_ptr); 2993 return; 2994 } 2995 2996 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2997 if (sd.region_align_down(old_src_addr) != end_addr) { 2998 // The partial object was copied from more than one source region. 2999 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3000 3001 // Move to the next source region, possibly switching spaces as well. All 3002 // args except end_addr may be modified. 3003 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3004 end_addr); 3005 } 3006 } 3007 3008 do { 3009 HeapWord* const cur_addr = closure.source(); 3010 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3011 src_space_top); 3012 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3013 3014 if (status == ParMarkBitMap::incomplete) { 3015 // The last obj that starts in the source region does not end in the 3016 // region. 3017 assert(closure.source() < end_addr, "sanity"); 3018 HeapWord* const obj_beg = closure.source(); 3019 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3020 src_space_top); 3021 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3022 if (obj_end < range_end) { 3023 // The end was found; the entire object will fit. 3024 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3025 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3026 } else { 3027 // The end was not found; the object will not fit. 3028 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3029 status = ParMarkBitMap::would_overflow; 3030 } 3031 } 3032 3033 if (status == ParMarkBitMap::would_overflow) { 3034 // The last object did not fit. Note that interior oop updates were 3035 // deferred, then copy enough of the object to fill the region. 3036 region_ptr->set_deferred_obj_addr(closure.destination()); 3037 status = closure.copy_until_full(); // copies from closure.source() 3038 3039 decrement_destination_counts(cm, src_space_id, src_region_idx, 3040 closure.source()); 3041 closure.complete_region(cm, dest_addr, region_ptr); 3042 return; 3043 } 3044 3045 if (status == ParMarkBitMap::full) { 3046 decrement_destination_counts(cm, src_space_id, src_region_idx, 3047 closure.source()); 3048 region_ptr->set_deferred_obj_addr(NULL); 3049 closure.complete_region(cm, dest_addr, region_ptr); 3050 return; 3051 } 3052 3053 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3054 3055 // Move to the next source region, possibly switching spaces as well. All 3056 // args except end_addr may be modified. 3057 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3058 end_addr); 3059 } while (true); 3060 } 3061 3062 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx) 3063 { 3064 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3065 fill_region(cm, cl, region_idx); 3066 } 3067 3068 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx) 3069 { 3070 // Get a shadow region first 3071 ParallelCompactData& sd = summary_data(); 3072 RegionData* const region_ptr = sd.region(region_idx); 3073 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr); 3074 // The InvalidShadow return value indicates the corresponding heap region is available, 3075 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use 3076 // MoveAndUpdateShadowClosure to fill the acquired shadow region. 3077 if (shadow_region == ParCompactionManager::InvalidShadow) { 3078 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3079 region_ptr->shadow_to_normal(); 3080 return fill_region(cm, cl, region_idx); 3081 } else { 3082 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region); 3083 return fill_region(cm, cl, region_idx); 3084 } 3085 } 3086 3087 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr) 3088 { 3089 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize); 3090 } 3091 3092 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx) 3093 { 3094 size_t next = cm->next_shadow_region(); 3095 ParallelCompactData& sd = summary_data(); 3096 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top()); 3097 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 3098 3099 while (next < old_new_top) { 3100 if (sd.region(next)->mark_shadow()) { 3101 region_idx = next; 3102 return true; 3103 } 3104 next = cm->move_next_shadow_region_by(active_gc_threads); 3105 } 3106 3107 return false; 3108 } 3109 3110 // The shadow region is an optimization to address region dependencies in full GC. The basic 3111 // idea is making more regions available by temporally storing their live objects in empty 3112 // shadow regions to resolve dependencies between them and the destination regions. Therefore, 3113 // GC threads need not wait destination regions to be available before processing sources. 3114 // 3115 // A typical workflow would be: 3116 // After draining its own stack and failing to steal from others, a GC worker would pick an 3117 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills 3118 // the shadow region by copying live objects from source regions of the unavailable one. Once 3119 // the unavailable region becomes available, the data in the shadow region will be copied back. 3120 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces. 3121 // 3122 // For more details, please refer to ยง4.2 of the VEE'19 paper: 3123 // Haoyu Li, Mingyu Wu, Binyu Zang, and Haibo Chen. 2019. ScissorGC: scalable and efficient 3124 // compaction for Java full garbage collection. In Proceedings of the 15th ACM SIGPLAN/SIGOPS 3125 // International Conference on Virtual Execution Environments (VEE 2019). ACM, New York, NY, USA, 3126 // 108-121. DOI: https://doi.org/10.1145/3313808.3313820 3127 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads) 3128 { 3129 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 3130 3131 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 3132 SpaceInfo* const space_info = _space_info + id; 3133 MutableSpace* const space = space_info->space(); 3134 3135 const size_t beg_region = 3136 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top()))); 3137 const size_t end_region = 3138 sd.addr_to_region_idx(sd.region_align_down(space->end())); 3139 3140 for (size_t cur = beg_region; cur < end_region; ++cur) { 3141 ParCompactionManager::push_shadow_region(cur); 3142 } 3143 } 3144 3145 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix()); 3146 for (uint i = 0; i < parallel_gc_threads; i++) { 3147 ParCompactionManager *cm = ParCompactionManager::manager_array(i); 3148 cm->set_next_shadow_region(beg_region + i); 3149 } 3150 } 3151 3152 void PSParallelCompact::fill_blocks(size_t region_idx) 3153 { 3154 // Fill in the block table elements for the specified region. Each block 3155 // table element holds the number of live words in the region that are to the 3156 // left of the first object that starts in the block. Thus only blocks in 3157 // which an object starts need to be filled. 3158 // 3159 // The algorithm scans the section of the bitmap that corresponds to the 3160 // region, keeping a running total of the live words. When an object start is 3161 // found, if it's the first to start in the block that contains it, the 3162 // current total is written to the block table element. 3163 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3164 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3165 const size_t RegionSize = ParallelCompactData::RegionSize; 3166 3167 ParallelCompactData& sd = summary_data(); 3168 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3169 if (partial_obj_size >= RegionSize) { 3170 return; // No objects start in this region. 3171 } 3172 3173 // Ensure the first loop iteration decides that the block has changed. 3174 size_t cur_block = sd.block_count(); 3175 3176 const ParMarkBitMap* const bitmap = mark_bitmap(); 3177 3178 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3179 assert((size_t)1 << Log2BitsPerBlock == 3180 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3181 3182 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3183 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3184 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3185 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3186 while (beg_bit < range_end) { 3187 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3188 if (new_block != cur_block) { 3189 cur_block = new_block; 3190 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3191 } 3192 3193 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3194 if (end_bit < range_end - 1) { 3195 live_bits += end_bit - beg_bit + 1; 3196 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3197 } else { 3198 return; 3199 } 3200 } 3201 } 3202 3203 jlong PSParallelCompact::millis_since_last_gc() { 3204 // We need a monotonically non-decreasing time in ms but 3205 // os::javaTimeMillis() does not guarantee monotonicity. 3206 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3207 jlong ret_val = now - _time_of_last_gc; 3208 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3209 if (ret_val < 0) { 3210 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3211 return 0; 3212 } 3213 return ret_val; 3214 } 3215 3216 void PSParallelCompact::reset_millis_since_last_gc() { 3217 // We need a monotonically non-decreasing time in ms but 3218 // os::javaTimeMillis() does not guarantee monotonicity. 3219 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3220 } 3221 3222 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3223 { 3224 if (source() != copy_destination()) { 3225 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3226 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining()); 3227 } 3228 update_state(words_remaining()); 3229 assert(is_full(), "sanity"); 3230 return ParMarkBitMap::full; 3231 } 3232 3233 void MoveAndUpdateClosure::copy_partial_obj() 3234 { 3235 size_t words = words_remaining(); 3236 3237 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3238 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3239 if (end_addr < range_end) { 3240 words = bitmap()->obj_size(source(), end_addr); 3241 } 3242 3243 // This test is necessary; if omitted, the pointer updates to a partial object 3244 // that crosses the dense prefix boundary could be overwritten. 3245 if (source() != copy_destination()) { 3246 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3247 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3248 } 3249 update_state(words); 3250 } 3251 3252 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3253 PSParallelCompact::RegionData *region_ptr) { 3254 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished"); 3255 region_ptr->set_completed(); 3256 } 3257 3258 ParMarkBitMapClosure::IterationStatus 3259 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3260 assert(destination() != NULL, "sanity"); 3261 assert(bitmap()->obj_size(addr) == words, "bad size"); 3262 3263 _source = addr; 3264 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3265 destination(), "wrong destination"); 3266 3267 if (words > words_remaining()) { 3268 return ParMarkBitMap::would_overflow; 3269 } 3270 3271 // The start_array must be updated even if the object is not moving. 3272 if (_start_array != NULL) { 3273 _start_array->allocate_block(destination()); 3274 } 3275 3276 if (copy_destination() != source()) { 3277 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3278 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3279 } 3280 3281 oop moved_oop = (oop) copy_destination(); 3282 compaction_manager()->update_contents(moved_oop); 3283 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3284 3285 update_state(words); 3286 assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity"); 3287 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3288 } 3289 3290 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3291 PSParallelCompact::RegionData *region_ptr) { 3292 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow"); 3293 // Record the shadow region index 3294 region_ptr->set_shadow_region(_shadow); 3295 // Mark the shadow region as filled to indicate the data is ready to be 3296 // copied back 3297 region_ptr->mark_filled(); 3298 // Try to copy the content of the shadow region back to its corresponding 3299 // heap region if available; the GC thread that decreases the destination 3300 // count to zero will do the copying otherwise (see 3301 // PSParallelCompact::decrement_destination_counts). 3302 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) { 3303 region_ptr->set_completed(); 3304 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr); 3305 ParCompactionManager::push_shadow_region_mt_safe(_shadow); 3306 } 3307 } 3308 3309 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3310 ParCompactionManager* cm, 3311 PSParallelCompact::SpaceId space_id) : 3312 ParMarkBitMapClosure(mbm, cm), 3313 _space_id(space_id), 3314 _start_array(PSParallelCompact::start_array(space_id)) 3315 { 3316 } 3317 3318 // Updates the references in the object to their new values. 3319 ParMarkBitMapClosure::IterationStatus 3320 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3321 do_addr(addr); 3322 return ParMarkBitMap::incomplete; 3323 } 3324 3325 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3326 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3327 _start_array(PSParallelCompact::start_array(space_id)) 3328 { 3329 assert(space_id == PSParallelCompact::old_space_id, 3330 "cannot use FillClosure in the young gen"); 3331 } 3332 3333 ParMarkBitMapClosure::IterationStatus 3334 FillClosure::do_addr(HeapWord* addr, size_t size) { 3335 CollectedHeap::fill_with_objects(addr, size); 3336 HeapWord* const end = addr + size; 3337 do { 3338 _start_array->allocate_block(addr); 3339 addr += oop(addr)->size(); 3340 } while (addr < end); 3341 return ParMarkBitMap::incomplete; 3342 }