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