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