1 /* 2 * Copyright (c) 2005, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "aot/aotLoader.hpp" 27 #include "classfile/classLoaderDataGraph.hpp" 28 #include "classfile/javaClasses.inline.hpp" 29 #include "classfile/stringTable.hpp" 30 #include "classfile/symbolTable.hpp" 31 #include "classfile/systemDictionary.hpp" 32 #include "code/codeCache.hpp" 33 #include "gc/parallel/parallelArguments.hpp" 34 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 35 #include "gc/parallel/parMarkBitMap.inline.hpp" 36 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 37 #include "gc/parallel/psCompactionManager.inline.hpp" 38 #include "gc/parallel/psOldGen.hpp" 39 #include "gc/parallel/psParallelCompact.inline.hpp" 40 #include "gc/parallel/psPromotionManager.inline.hpp" 41 #include "gc/parallel/psRootType.hpp" 42 #include "gc/parallel/psScavenge.hpp" 43 #include "gc/parallel/psYoungGen.hpp" 44 #include "gc/shared/gcCause.hpp" 45 #include "gc/shared/gcHeapSummary.hpp" 46 #include "gc/shared/gcId.hpp" 47 #include "gc/shared/gcLocker.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/isGCActiveMark.hpp" 52 #include "gc/shared/referencePolicy.hpp" 53 #include "gc/shared/referenceProcessor.hpp" 54 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 55 #include "gc/shared/spaceDecorator.inline.hpp" 56 #include "gc/shared/taskTerminator.hpp" 57 #include "gc/shared/weakProcessor.hpp" 58 #include "gc/shared/workerPolicy.hpp" 59 #include "gc/shared/workgroup.hpp" 60 #include "logging/log.hpp" 61 #include "memory/iterator.inline.hpp" 62 #include "memory/resourceArea.hpp" 63 #include "memory/universe.hpp" 64 #include "oops/access.inline.hpp" 65 #include "oops/instanceClassLoaderKlass.inline.hpp" 66 #include "oops/instanceKlass.inline.hpp" 67 #include "oops/instanceMirrorKlass.inline.hpp" 68 #include "oops/methodData.hpp" 69 #include "oops/objArrayKlass.inline.hpp" 70 #include "oops/oop.inline.hpp" 71 #include "runtime/atomic.hpp" 72 #include "runtime/handles.inline.hpp" 73 #include "runtime/safepoint.hpp" 74 #include "runtime/vmThread.hpp" 75 #include "services/management.hpp" 76 #include "services/memTracker.hpp" 77 #include "services/memoryService.hpp" 78 #include "utilities/align.hpp" 79 #include "utilities/debug.hpp" 80 #include "utilities/events.hpp" 81 #include "utilities/formatBuffer.hpp" 82 #include "utilities/macros.hpp" 83 #include "utilities/stack.inline.hpp" 84 #if INCLUDE_JVMCI 85 #include "jvmci/jvmci.hpp" 86 #endif 87 88 #include <math.h> 89 90 // All sizes are in HeapWords. 91 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 92 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 93 const size_t ParallelCompactData::RegionSizeBytes = 94 RegionSize << LogHeapWordSize; 95 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 96 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 97 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 98 99 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 100 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 101 const size_t ParallelCompactData::BlockSizeBytes = 102 BlockSize << LogHeapWordSize; 103 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 104 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 105 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 106 107 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 108 const size_t ParallelCompactData::Log2BlocksPerRegion = 109 Log2RegionSize - Log2BlockSize; 110 111 const ParallelCompactData::RegionData::region_sz_t 112 ParallelCompactData::RegionData::dc_shift = 27; 113 114 const ParallelCompactData::RegionData::region_sz_t 115 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 116 117 const ParallelCompactData::RegionData::region_sz_t 118 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 119 120 const ParallelCompactData::RegionData::region_sz_t 121 ParallelCompactData::RegionData::los_mask = ~dc_mask; 122 123 const ParallelCompactData::RegionData::region_sz_t 124 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 125 126 const ParallelCompactData::RegionData::region_sz_t 127 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 128 129 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 130 131 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 132 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 133 134 double PSParallelCompact::_dwl_mean; 135 double PSParallelCompact::_dwl_std_dev; 136 double PSParallelCompact::_dwl_first_term; 137 double PSParallelCompact::_dwl_adjustment; 138 #ifdef ASSERT 139 bool PSParallelCompact::_dwl_initialized = false; 140 #endif // #ifdef ASSERT 141 142 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 143 HeapWord* destination) 144 { 145 assert(src_region_idx != 0, "invalid src_region_idx"); 146 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 147 assert(destination != NULL, "invalid destination argument"); 148 149 _src_region_idx = src_region_idx; 150 _partial_obj_size = partial_obj_size; 151 _destination = destination; 152 153 // These fields may not be updated below, so make sure they're clear. 154 assert(_dest_region_addr == NULL, "should have been cleared"); 155 assert(_first_src_addr == NULL, "should have been cleared"); 156 157 // Determine the number of destination regions for the partial object. 158 HeapWord* const last_word = destination + partial_obj_size - 1; 159 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 160 HeapWord* const beg_region_addr = sd.region_align_down(destination); 161 HeapWord* const end_region_addr = sd.region_align_down(last_word); 162 163 if (beg_region_addr == end_region_addr) { 164 // One destination region. 165 _destination_count = 1; 166 if (end_region_addr == destination) { 167 // The destination falls on a region boundary, thus the first word of the 168 // partial object will be the first word copied to the destination region. 169 _dest_region_addr = end_region_addr; 170 _first_src_addr = sd.region_to_addr(src_region_idx); 171 } 172 } else { 173 // Two destination regions. When copied, the partial object will cross a 174 // destination region boundary, so a word somewhere within the partial 175 // object will be the first word copied to the second destination region. 176 _destination_count = 2; 177 _dest_region_addr = end_region_addr; 178 const size_t ofs = pointer_delta(end_region_addr, destination); 179 assert(ofs < _partial_obj_size, "sanity"); 180 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 181 } 182 } 183 184 void SplitInfo::clear() 185 { 186 _src_region_idx = 0; 187 _partial_obj_size = 0; 188 _destination = NULL; 189 _destination_count = 0; 190 _dest_region_addr = NULL; 191 _first_src_addr = NULL; 192 assert(!is_valid(), "sanity"); 193 } 194 195 #ifdef ASSERT 196 void SplitInfo::verify_clear() 197 { 198 assert(_src_region_idx == 0, "not clear"); 199 assert(_partial_obj_size == 0, "not clear"); 200 assert(_destination == NULL, "not clear"); 201 assert(_destination_count == 0, "not clear"); 202 assert(_dest_region_addr == NULL, "not clear"); 203 assert(_first_src_addr == NULL, "not clear"); 204 } 205 #endif // #ifdef ASSERT 206 207 208 void PSParallelCompact::print_on_error(outputStream* st) { 209 _mark_bitmap.print_on_error(st); 210 } 211 212 #ifndef PRODUCT 213 const char* PSParallelCompact::space_names[] = { 214 "old ", "eden", "from", "to " 215 }; 216 217 void PSParallelCompact::print_region_ranges() { 218 if (!log_develop_is_enabled(Trace, gc, compaction)) { 219 return; 220 } 221 Log(gc, compaction) log; 222 ResourceMark rm; 223 LogStream ls(log.trace()); 224 Universe::print_on(&ls); 225 log.trace("space bottom top end new_top"); 226 log.trace("------ ---------- ---------- ---------- ----------"); 227 228 for (unsigned int id = 0; id < last_space_id; ++id) { 229 const MutableSpace* space = _space_info[id].space(); 230 log.trace("%u %s " 231 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 232 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 233 id, space_names[id], 234 summary_data().addr_to_region_idx(space->bottom()), 235 summary_data().addr_to_region_idx(space->top()), 236 summary_data().addr_to_region_idx(space->end()), 237 summary_data().addr_to_region_idx(_space_info[id].new_top())); 238 } 239 } 240 241 void 242 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 243 { 244 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 245 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 246 247 ParallelCompactData& sd = PSParallelCompact::summary_data(); 248 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 249 log_develop_trace(gc, compaction)( 250 REGION_IDX_FORMAT " " PTR_FORMAT " " 251 REGION_IDX_FORMAT " " PTR_FORMAT " " 252 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 253 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 254 i, p2i(c->data_location()), dci, p2i(c->destination()), 255 c->partial_obj_size(), c->live_obj_size(), 256 c->data_size(), c->source_region(), c->destination_count()); 257 258 #undef REGION_IDX_FORMAT 259 #undef REGION_DATA_FORMAT 260 } 261 262 void 263 print_generic_summary_data(ParallelCompactData& summary_data, 264 HeapWord* const beg_addr, 265 HeapWord* const end_addr) 266 { 267 size_t total_words = 0; 268 size_t i = summary_data.addr_to_region_idx(beg_addr); 269 const size_t last = summary_data.addr_to_region_idx(end_addr); 270 HeapWord* pdest = 0; 271 272 while (i < last) { 273 ParallelCompactData::RegionData* c = summary_data.region(i); 274 if (c->data_size() != 0 || c->destination() != pdest) { 275 print_generic_summary_region(i, c); 276 total_words += c->data_size(); 277 pdest = c->destination(); 278 } 279 ++i; 280 } 281 282 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 283 } 284 285 void 286 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 287 HeapWord* const beg_addr, 288 HeapWord* const end_addr) { 289 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 290 } 291 292 void 293 print_generic_summary_data(ParallelCompactData& summary_data, 294 SpaceInfo* space_info) 295 { 296 if (!log_develop_is_enabled(Trace, gc, compaction)) { 297 return; 298 } 299 300 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 301 const MutableSpace* space = space_info[id].space(); 302 print_generic_summary_data(summary_data, space->bottom(), 303 MAX2(space->top(), space_info[id].new_top())); 304 } 305 } 306 307 void 308 print_initial_summary_data(ParallelCompactData& summary_data, 309 const MutableSpace* space) { 310 if (space->top() == space->bottom()) { 311 return; 312 } 313 314 const size_t region_size = ParallelCompactData::RegionSize; 315 typedef ParallelCompactData::RegionData RegionData; 316 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 317 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 318 const RegionData* c = summary_data.region(end_region - 1); 319 HeapWord* end_addr = c->destination() + c->data_size(); 320 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 321 322 // Print (and count) the full regions at the beginning of the space. 323 size_t full_region_count = 0; 324 size_t i = summary_data.addr_to_region_idx(space->bottom()); 325 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 326 ParallelCompactData::RegionData* c = summary_data.region(i); 327 log_develop_trace(gc, compaction)( 328 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 329 i, p2i(c->destination()), 330 c->partial_obj_size(), c->live_obj_size(), 331 c->data_size(), c->source_region(), c->destination_count()); 332 ++full_region_count; 333 ++i; 334 } 335 336 size_t live_to_right = live_in_space - full_region_count * region_size; 337 338 double max_reclaimed_ratio = 0.0; 339 size_t max_reclaimed_ratio_region = 0; 340 size_t max_dead_to_right = 0; 341 size_t max_live_to_right = 0; 342 343 // Print the 'reclaimed ratio' for regions while there is something live in 344 // the region or to the right of it. The remaining regions are empty (and 345 // uninteresting), and computing the ratio will result in division by 0. 346 while (i < end_region && live_to_right > 0) { 347 c = summary_data.region(i); 348 HeapWord* const region_addr = summary_data.region_to_addr(i); 349 const size_t used_to_right = pointer_delta(space->top(), region_addr); 350 const size_t dead_to_right = used_to_right - live_to_right; 351 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 352 353 if (reclaimed_ratio > max_reclaimed_ratio) { 354 max_reclaimed_ratio = reclaimed_ratio; 355 max_reclaimed_ratio_region = i; 356 max_dead_to_right = dead_to_right; 357 max_live_to_right = live_to_right; 358 } 359 360 ParallelCompactData::RegionData* c = summary_data.region(i); 361 log_develop_trace(gc, compaction)( 362 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 363 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 364 i, p2i(c->destination()), 365 c->partial_obj_size(), c->live_obj_size(), 366 c->data_size(), c->source_region(), c->destination_count(), 367 reclaimed_ratio, dead_to_right, live_to_right); 368 369 370 live_to_right -= c->data_size(); 371 ++i; 372 } 373 374 // Any remaining regions are empty. Print one more if there is one. 375 if (i < end_region) { 376 ParallelCompactData::RegionData* c = summary_data.region(i); 377 log_develop_trace(gc, compaction)( 378 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 379 i, p2i(c->destination()), 380 c->partial_obj_size(), c->live_obj_size(), 381 c->data_size(), c->source_region(), c->destination_count()); 382 } 383 384 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 385 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 386 } 387 388 void 389 print_initial_summary_data(ParallelCompactData& summary_data, 390 SpaceInfo* space_info) { 391 if (!log_develop_is_enabled(Trace, gc, compaction)) { 392 return; 393 } 394 395 unsigned int id = PSParallelCompact::old_space_id; 396 const MutableSpace* space; 397 do { 398 space = space_info[id].space(); 399 print_initial_summary_data(summary_data, space); 400 } while (++id < PSParallelCompact::eden_space_id); 401 402 do { 403 space = space_info[id].space(); 404 print_generic_summary_data(summary_data, space->bottom(), space->top()); 405 } while (++id < PSParallelCompact::last_space_id); 406 } 407 #endif // #ifndef PRODUCT 408 409 #ifdef ASSERT 410 size_t add_obj_count; 411 size_t add_obj_size; 412 size_t mark_bitmap_count; 413 size_t mark_bitmap_size; 414 #endif // #ifdef ASSERT 415 416 ParallelCompactData::ParallelCompactData() : 417 _region_start(NULL), 418 DEBUG_ONLY(_region_end(NULL) COMMA) 419 _region_vspace(NULL), 420 _reserved_byte_size(0), 421 _region_data(NULL), 422 _region_count(0), 423 _block_vspace(NULL), 424 _block_data(NULL), 425 _block_count(0) {} 426 427 bool ParallelCompactData::initialize(MemRegion covered_region) 428 { 429 _region_start = covered_region.start(); 430 const size_t region_size = covered_region.word_size(); 431 DEBUG_ONLY(_region_end = _region_start + region_size;) 432 433 assert(region_align_down(_region_start) == _region_start, 434 "region start not aligned"); 435 assert((region_size & RegionSizeOffsetMask) == 0, 436 "region size not a multiple of RegionSize"); 437 438 bool result = initialize_region_data(region_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 == (size_t) os::vm_page_size() ? 0 : 451 MAX2(page_sz, granularity); 452 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 453 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 454 rs.size()); 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 region_size) 472 { 473 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 474 _region_vspace = create_vspace(count, sizeof(RegionData)); 475 if (_region_vspace != 0) { 476 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 477 _region_count = count; 478 return true; 479 } 480 return false; 481 } 482 483 bool ParallelCompactData::initialize_block_data() 484 { 485 assert(_region_count != 0, "region data must be initialized first"); 486 const size_t count = _region_count << Log2BlocksPerRegion; 487 _block_vspace = create_vspace(count, sizeof(BlockData)); 488 if (_block_vspace != 0) { 489 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 490 _block_count = count; 491 return true; 492 } 493 return false; 494 } 495 496 void ParallelCompactData::clear() 497 { 498 memset(_region_data, 0, _region_vspace->committed_size()); 499 memset(_block_data, 0, _block_vspace->committed_size()); 500 } 501 502 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 503 assert(beg_region <= _region_count, "beg_region out of range"); 504 assert(end_region <= _region_count, "end_region out of range"); 505 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 506 507 const size_t region_cnt = end_region - beg_region; 508 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 509 510 const size_t beg_block = beg_region * BlocksPerRegion; 511 const size_t block_cnt = region_cnt * BlocksPerRegion; 512 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 513 } 514 515 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 516 { 517 const RegionData* cur_cp = region(region_idx); 518 const RegionData* const end_cp = region(region_count() - 1); 519 520 HeapWord* result = region_to_addr(region_idx); 521 if (cur_cp < end_cp) { 522 do { 523 result += cur_cp->partial_obj_size(); 524 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 525 } 526 return result; 527 } 528 529 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 530 { 531 const size_t obj_ofs = pointer_delta(addr, _region_start); 532 const size_t beg_region = obj_ofs >> Log2RegionSize; 533 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 534 535 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 536 DEBUG_ONLY(Atomic::add(&add_obj_size, len);) 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 Klass* klass = ((oop)addr)->klass(); 549 // Middle regions--completely spanned by this object. 550 for (size_t region = beg_region + 1; region < end_region; ++region) { 551 _region_data[region].set_partial_obj_size(RegionSize); 552 _region_data[region].set_partial_obj_addr(addr); 553 } 554 555 // Last region. 556 const size_t end_ofs = region_offset(addr + len - 1); 557 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 558 _region_data[end_region].set_partial_obj_addr(addr); 559 } 560 561 void 562 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 563 { 564 assert(region_offset(beg) == 0, "not RegionSize aligned"); 565 assert(region_offset(end) == 0, "not RegionSize aligned"); 566 567 size_t cur_region = addr_to_region_idx(beg); 568 const size_t end_region = addr_to_region_idx(end); 569 HeapWord* addr = beg; 570 while (cur_region < end_region) { 571 _region_data[cur_region].set_destination(addr); 572 _region_data[cur_region].set_destination_count(0); 573 _region_data[cur_region].set_source_region(cur_region); 574 _region_data[cur_region].set_data_location(addr); 575 576 // Update live_obj_size so the region appears completely full. 577 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 578 _region_data[cur_region].set_live_obj_size(live_size); 579 580 ++cur_region; 581 addr += RegionSize; 582 } 583 } 584 585 // Find the point at which a space can be split and, if necessary, record the 586 // split point. 587 // 588 // If the current src region (which overflowed the destination space) doesn't 589 // have a partial object, the split point is at the beginning of the current src 590 // region (an "easy" split, no extra bookkeeping required). 591 // 592 // If the current src region has a partial object, the split point is in the 593 // region where that partial object starts (call it the split_region). If 594 // split_region has a partial object, then the split point is just after that 595 // partial object (a "hard" split where we have to record the split data and 596 // zero the partial_obj_size field). With a "hard" split, we know that the 597 // partial_obj ends within split_region because the partial object that caused 598 // the overflow starts in split_region. If split_region doesn't have a partial 599 // obj, then the split is at the beginning of split_region (another "easy" 600 // split). 601 HeapWord* 602 ParallelCompactData::summarize_split_space(size_t src_region, 603 SplitInfo& split_info, 604 HeapWord* destination, 605 HeapWord* target_end, 606 HeapWord** target_next) 607 { 608 assert(destination <= target_end, "sanity"); 609 assert(destination + _region_data[src_region].data_size() > target_end, 610 "region should not fit into target space"); 611 assert(is_region_aligned(target_end), "sanity"); 612 613 size_t split_region = src_region; 614 HeapWord* split_destination = destination; 615 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 616 617 if (destination + partial_obj_size > target_end) { 618 // The split point is just after the partial object (if any) in the 619 // src_region that contains the start of the object that overflowed the 620 // destination space. 621 // 622 // Find the start of the "overflow" object and set split_region to the 623 // region containing it. 624 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 625 split_region = addr_to_region_idx(overflow_obj); 626 627 // Clear the source_region field of all destination regions whose first word 628 // came from data after the split point (a non-null source_region field 629 // implies a region must be filled). 630 // 631 // An alternative to the simple loop below: clear during post_compact(), 632 // which uses memcpy instead of individual stores, and is easy to 633 // parallelize. (The downside is that it clears the entire RegionData 634 // object as opposed to just one field.) 635 // 636 // post_compact() would have to clear the summary data up to the highest 637 // address that was written during the summary phase, which would be 638 // 639 // max(top, max(new_top, clear_top)) 640 // 641 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 642 // to target_end. 643 const RegionData* const sr = region(split_region); 644 const size_t beg_idx = 645 addr_to_region_idx(region_align_up(sr->destination() + 646 sr->partial_obj_size())); 647 const size_t end_idx = addr_to_region_idx(target_end); 648 649 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 650 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 651 _region_data[idx].set_source_region(0); 652 } 653 654 // Set split_destination and partial_obj_size to reflect the split region. 655 split_destination = sr->destination(); 656 partial_obj_size = sr->partial_obj_size(); 657 } 658 659 // The split is recorded only if a partial object extends onto the region. 660 if (partial_obj_size != 0) { 661 _region_data[split_region].set_partial_obj_size(0); 662 split_info.record(split_region, partial_obj_size, split_destination); 663 } 664 665 // Setup the continuation addresses. 666 *target_next = split_destination + partial_obj_size; 667 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 668 669 if (log_develop_is_enabled(Trace, gc, compaction)) { 670 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 671 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 672 split_type, p2i(source_next), split_region, partial_obj_size); 673 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 674 split_type, p2i(split_destination), 675 addr_to_region_idx(split_destination), 676 p2i(*target_next)); 677 678 if (partial_obj_size != 0) { 679 HeapWord* const po_beg = split_info.destination(); 680 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 681 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 682 split_type, 683 p2i(po_beg), addr_to_region_idx(po_beg), 684 p2i(po_end), addr_to_region_idx(po_end)); 685 } 686 } 687 688 return source_next; 689 } 690 691 bool ParallelCompactData::summarize(SplitInfo& split_info, 692 HeapWord* source_beg, HeapWord* source_end, 693 HeapWord** source_next, 694 HeapWord* target_beg, HeapWord* target_end, 695 HeapWord** target_next) 696 { 697 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 698 log_develop_trace(gc, compaction)( 699 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 700 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 701 p2i(source_beg), p2i(source_end), p2i(source_next_val), 702 p2i(target_beg), p2i(target_end), p2i(*target_next)); 703 704 size_t cur_region = addr_to_region_idx(source_beg); 705 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 706 707 HeapWord *dest_addr = target_beg; 708 while (cur_region < end_region) { 709 // The destination must be set even if the region has no data. 710 _region_data[cur_region].set_destination(dest_addr); 711 712 size_t words = _region_data[cur_region].data_size(); 713 if (words > 0) { 714 // If cur_region does not fit entirely into the target space, find a point 715 // at which the source space can be 'split' so that part is copied to the 716 // target space and the rest is copied elsewhere. 717 if (dest_addr + words > target_end) { 718 assert(source_next != NULL, "source_next is NULL when splitting"); 719 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 720 target_end, target_next); 721 return false; 722 } 723 724 // Compute the destination_count for cur_region, and if necessary, update 725 // source_region for a destination region. The source_region field is 726 // updated if cur_region is the first (left-most) region to be copied to a 727 // destination region. 728 // 729 // The destination_count calculation is a bit subtle. A region that has 730 // data that compacts into itself does not count itself as a destination. 731 // This maintains the invariant that a zero count means the region is 732 // available and can be claimed and then filled. 733 uint destination_count = 0; 734 if (split_info.is_split(cur_region)) { 735 // The current region has been split: the partial object will be copied 736 // to one destination space and the remaining data will be copied to 737 // another destination space. Adjust the initial destination_count and, 738 // if necessary, set the source_region field if the partial object will 739 // cross a destination region boundary. 740 destination_count = split_info.destination_count(); 741 if (destination_count == 2) { 742 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 743 _region_data[dest_idx].set_source_region(cur_region); 744 } 745 } 746 747 HeapWord* const last_addr = dest_addr + words - 1; 748 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 749 const size_t dest_region_2 = addr_to_region_idx(last_addr); 750 751 // Initially assume that the destination regions will be the same and 752 // adjust the value below if necessary. Under this assumption, if 753 // cur_region == dest_region_2, then cur_region will be compacted 754 // completely into itself. 755 destination_count += cur_region == dest_region_2 ? 0 : 1; 756 if (dest_region_1 != dest_region_2) { 757 // Destination regions differ; adjust destination_count. 758 destination_count += 1; 759 // Data from cur_region will be copied to the start of dest_region_2. 760 _region_data[dest_region_2].set_source_region(cur_region); 761 } else if (region_offset(dest_addr) == 0) { 762 // Data from cur_region will be copied to the start of the destination 763 // region. 764 _region_data[dest_region_1].set_source_region(cur_region); 765 } 766 767 _region_data[cur_region].set_destination_count(destination_count); 768 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 769 dest_addr += words; 770 } 771 772 ++cur_region; 773 } 774 775 *target_next = dest_addr; 776 return true; 777 } 778 779 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 780 assert(addr != NULL, "Should detect NULL oop earlier"); 781 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 782 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 783 784 // Region covering the object. 785 RegionData* const region_ptr = addr_to_region_ptr(addr); 786 HeapWord* result = region_ptr->destination(); 787 788 // If the entire Region is live, the new location is region->destination + the 789 // offset of the object within in the Region. 790 791 // Run some performance tests to determine if this special case pays off. It 792 // is worth it for pointers into the dense prefix. If the optimization to 793 // avoid pointer updates in regions that only point to the dense prefix is 794 // ever implemented, this should be revisited. 795 if (region_ptr->data_size() == RegionSize) { 796 result += region_offset(addr); 797 return result; 798 } 799 800 // Otherwise, the new location is region->destination + block offset + the 801 // number of live words in the Block that are (a) to the left of addr and (b) 802 // due to objects that start in the Block. 803 804 // Fill in the block table if necessary. This is unsynchronized, so multiple 805 // threads may fill the block table for a region (harmless, since it is 806 // idempotent). 807 if (!region_ptr->blocks_filled()) { 808 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 809 region_ptr->set_blocks_filled(); 810 } 811 812 HeapWord* const search_start = block_align_down(addr); 813 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 814 815 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 816 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 817 result += block_offset + live; 818 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 819 return result; 820 } 821 822 #ifdef ASSERT 823 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 824 { 825 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 826 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 827 for (const size_t* p = beg; p < end; ++p) { 828 assert(*p == 0, "not zero"); 829 } 830 } 831 832 void ParallelCompactData::verify_clear() 833 { 834 verify_clear(_region_vspace); 835 verify_clear(_block_vspace); 836 } 837 #endif // #ifdef ASSERT 838 839 STWGCTimer PSParallelCompact::_gc_timer; 840 ParallelOldTracer PSParallelCompact::_gc_tracer; 841 elapsedTimer PSParallelCompact::_accumulated_time; 842 unsigned int PSParallelCompact::_total_invocations = 0; 843 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 844 jlong PSParallelCompact::_time_of_last_gc = 0; 845 CollectorCounters* PSParallelCompact::_counters = NULL; 846 ParMarkBitMap PSParallelCompact::_mark_bitmap; 847 ParallelCompactData PSParallelCompact::_summary_data; 848 849 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 850 851 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 852 853 class PCReferenceProcessor: public ReferenceProcessor { 854 public: 855 PCReferenceProcessor( 856 BoolObjectClosure* is_subject_to_discovery, 857 BoolObjectClosure* is_alive_non_header) : 858 ReferenceProcessor(is_subject_to_discovery, 859 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 860 ParallelGCThreads, // mt processing degree 861 true, // mt discovery 862 ParallelGCThreads, // mt discovery degree 863 true, // atomic_discovery 864 is_alive_non_header) { 865 } 866 867 template<typename T> bool discover(oop obj, ReferenceType type) { 868 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj); 869 T heap_oop = RawAccess<>::oop_load(referent_addr); 870 oop referent = CompressedOops::decode_not_null(heap_oop); 871 return PSParallelCompact::mark_bitmap()->is_unmarked(referent) 872 && ReferenceProcessor::discover_reference(obj, type); 873 } 874 virtual bool discover_reference(oop obj, ReferenceType type) { 875 if (UseCompressedOops) { 876 return discover<narrowOop>(obj, type); 877 } else { 878 return discover<oop>(obj, type); 879 } 880 } 881 }; 882 883 void PSParallelCompact::post_initialize() { 884 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 885 _span_based_discoverer.set_span(heap->reserved_region()); 886 _ref_processor = 887 new PCReferenceProcessor(&_span_based_discoverer, 888 &_is_alive_closure); // non-header is alive closure 889 890 _counters = new CollectorCounters("Parallel full collection pauses", 1); 891 892 // Initialize static fields in ParCompactionManager. 893 ParCompactionManager::initialize(mark_bitmap()); 894 } 895 896 bool PSParallelCompact::initialize() { 897 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 898 MemRegion mr = heap->reserved_region(); 899 900 // Was the old gen get allocated successfully? 901 if (!heap->old_gen()->is_allocated()) { 902 return false; 903 } 904 905 initialize_space_info(); 906 initialize_dead_wood_limiter(); 907 908 if (!_mark_bitmap.initialize(mr)) { 909 vm_shutdown_during_initialization( 910 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 911 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 912 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 913 return false; 914 } 915 916 if (!_summary_data.initialize(mr)) { 917 vm_shutdown_during_initialization( 918 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 919 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 920 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 921 return false; 922 } 923 924 return true; 925 } 926 927 void PSParallelCompact::initialize_space_info() 928 { 929 memset(&_space_info, 0, sizeof(_space_info)); 930 931 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 932 PSYoungGen* young_gen = heap->young_gen(); 933 934 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 935 _space_info[eden_space_id].set_space(young_gen->eden_space()); 936 _space_info[from_space_id].set_space(young_gen->from_space()); 937 _space_info[to_space_id].set_space(young_gen->to_space()); 938 939 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 940 } 941 942 void PSParallelCompact::initialize_dead_wood_limiter() 943 { 944 const size_t max = 100; 945 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 946 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 947 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 948 DEBUG_ONLY(_dwl_initialized = true;) 949 _dwl_adjustment = normal_distribution(1.0); 950 } 951 952 void 953 PSParallelCompact::clear_data_covering_space(SpaceId id) 954 { 955 // At this point, top is the value before GC, new_top() is the value that will 956 // be set at the end of GC. The marking bitmap is cleared to top; nothing 957 // should be marked above top. The summary data is cleared to the larger of 958 // top & new_top. 959 MutableSpace* const space = _space_info[id].space(); 960 HeapWord* const bot = space->bottom(); 961 HeapWord* const top = space->top(); 962 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 963 964 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 965 const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top)); 966 _mark_bitmap.clear_range(beg_bit, end_bit); 967 968 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 969 const size_t end_region = 970 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 971 _summary_data.clear_range(beg_region, end_region); 972 973 // Clear the data used to 'split' regions. 974 SplitInfo& split_info = _space_info[id].split_info(); 975 if (split_info.is_valid()) { 976 split_info.clear(); 977 } 978 DEBUG_ONLY(split_info.verify_clear();) 979 } 980 981 void PSParallelCompact::pre_compact() 982 { 983 // Update the from & to space pointers in space_info, since they are swapped 984 // at each young gen gc. Do the update unconditionally (even though a 985 // promotion failure does not swap spaces) because an unknown number of young 986 // collections will have swapped the spaces an unknown number of times. 987 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 988 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 989 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 990 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 991 992 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 993 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 994 995 // Increment the invocation count 996 heap->increment_total_collections(true); 997 998 // We need to track unique mark sweep invocations as well. 999 _total_invocations++; 1000 1001 heap->print_heap_before_gc(); 1002 heap->trace_heap_before_gc(&_gc_tracer); 1003 1004 // Fill in TLABs 1005 heap->ensure_parsability(true); // retire TLABs 1006 1007 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1008 HandleMark hm; // Discard invalid handles created during verification 1009 Universe::verify("Before GC"); 1010 } 1011 1012 // Verify object start arrays 1013 if (VerifyObjectStartArray && 1014 VerifyBeforeGC) { 1015 heap->old_gen()->verify_object_start_array(); 1016 } 1017 1018 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1019 DEBUG_ONLY(summary_data().verify_clear();) 1020 1021 ParCompactionManager::reset_all_bitmap_query_caches(); 1022 } 1023 1024 void PSParallelCompact::post_compact() 1025 { 1026 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1027 ParCompactionManager::remove_all_shadow_regions(); 1028 1029 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1030 // Clear the marking bitmap, summary data and split info. 1031 clear_data_covering_space(SpaceId(id)); 1032 // Update top(). Must be done after clearing the bitmap and summary data. 1033 _space_info[id].publish_new_top(); 1034 } 1035 1036 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1037 MutableSpace* const from_space = _space_info[from_space_id].space(); 1038 MutableSpace* const to_space = _space_info[to_space_id].space(); 1039 1040 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1041 bool eden_empty = eden_space->is_empty(); 1042 if (!eden_empty) { 1043 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1044 heap->young_gen(), heap->old_gen()); 1045 } 1046 1047 // Update heap occupancy information which is used as input to the soft ref 1048 // clearing policy at the next gc. 1049 Universe::update_heap_info_at_gc(); 1050 1051 bool young_gen_empty = eden_empty && from_space->is_empty() && 1052 to_space->is_empty(); 1053 1054 PSCardTable* ct = heap->card_table(); 1055 MemRegion old_mr = heap->old_gen()->reserved(); 1056 if (young_gen_empty) { 1057 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1058 } else { 1059 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1060 } 1061 1062 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1063 ClassLoaderDataGraph::purge(); 1064 MetaspaceUtils::verify_metrics(); 1065 1066 heap->prune_scavengable_nmethods(); 1067 1068 #if COMPILER2_OR_JVMCI 1069 DerivedPointerTable::update_pointers(); 1070 #endif 1071 1072 if (ZapUnusedHeapArea) { 1073 heap->gen_mangle_unused_area(); 1074 } 1075 1076 // Update time of last GC 1077 reset_millis_since_last_gc(); 1078 } 1079 1080 HeapWord* 1081 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1082 bool maximum_compaction) 1083 { 1084 const size_t region_size = ParallelCompactData::RegionSize; 1085 const ParallelCompactData& sd = summary_data(); 1086 1087 const MutableSpace* const space = _space_info[id].space(); 1088 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1089 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1090 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1091 1092 // Skip full regions at the beginning of the space--they are necessarily part 1093 // of the dense prefix. 1094 size_t full_count = 0; 1095 const RegionData* cp; 1096 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1097 ++full_count; 1098 } 1099 1100 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1101 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1102 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1103 if (maximum_compaction || cp == end_cp || interval_ended) { 1104 _maximum_compaction_gc_num = total_invocations(); 1105 return sd.region_to_addr(cp); 1106 } 1107 1108 HeapWord* const new_top = _space_info[id].new_top(); 1109 const size_t space_live = pointer_delta(new_top, space->bottom()); 1110 const size_t space_used = space->used_in_words(); 1111 const size_t space_capacity = space->capacity_in_words(); 1112 1113 const double cur_density = double(space_live) / space_capacity; 1114 const double deadwood_density = 1115 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1116 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1117 1118 if (TraceParallelOldGCDensePrefix) { 1119 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1120 cur_density, deadwood_density, deadwood_goal); 1121 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1122 "space_cap=" SIZE_FORMAT, 1123 space_live, space_used, 1124 space_capacity); 1125 } 1126 1127 // XXX - Use binary search? 1128 HeapWord* dense_prefix = sd.region_to_addr(cp); 1129 const RegionData* full_cp = cp; 1130 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1131 while (cp < end_cp) { 1132 HeapWord* region_destination = cp->destination(); 1133 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1134 if (TraceParallelOldGCDensePrefix && Verbose) { 1135 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1136 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1137 sd.region(cp), p2i(region_destination), 1138 p2i(dense_prefix), cur_deadwood); 1139 } 1140 1141 if (cur_deadwood >= deadwood_goal) { 1142 // Found the region that has the correct amount of deadwood to the left. 1143 // This typically occurs after crossing a fairly sparse set of regions, so 1144 // iterate backwards over those sparse regions, looking for the region 1145 // that has the lowest density of live objects 'to the right.' 1146 size_t space_to_left = sd.region(cp) * region_size; 1147 size_t live_to_left = space_to_left - cur_deadwood; 1148 size_t space_to_right = space_capacity - space_to_left; 1149 size_t live_to_right = space_live - live_to_left; 1150 double density_to_right = double(live_to_right) / space_to_right; 1151 while (cp > full_cp) { 1152 --cp; 1153 const size_t prev_region_live_to_right = live_to_right - 1154 cp->data_size(); 1155 const size_t prev_region_space_to_right = space_to_right + region_size; 1156 double prev_region_density_to_right = 1157 double(prev_region_live_to_right) / prev_region_space_to_right; 1158 if (density_to_right <= prev_region_density_to_right) { 1159 return dense_prefix; 1160 } 1161 if (TraceParallelOldGCDensePrefix && Verbose) { 1162 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1163 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1164 prev_region_density_to_right); 1165 } 1166 dense_prefix -= region_size; 1167 live_to_right = prev_region_live_to_right; 1168 space_to_right = prev_region_space_to_right; 1169 density_to_right = prev_region_density_to_right; 1170 } 1171 return dense_prefix; 1172 } 1173 1174 dense_prefix += region_size; 1175 ++cp; 1176 } 1177 1178 return dense_prefix; 1179 } 1180 1181 #ifndef PRODUCT 1182 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1183 const SpaceId id, 1184 const bool maximum_compaction, 1185 HeapWord* const addr) 1186 { 1187 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1188 RegionData* const cp = summary_data().region(region_idx); 1189 const MutableSpace* const space = _space_info[id].space(); 1190 HeapWord* const new_top = _space_info[id].new_top(); 1191 1192 const size_t space_live = pointer_delta(new_top, space->bottom()); 1193 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1194 const size_t space_cap = space->capacity_in_words(); 1195 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1196 const size_t live_to_right = new_top - cp->destination(); 1197 const size_t dead_to_right = space->top() - addr - live_to_right; 1198 1199 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1200 "spl=" SIZE_FORMAT " " 1201 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1202 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1203 " ratio=%10.8f", 1204 algorithm, p2i(addr), region_idx, 1205 space_live, 1206 dead_to_left, dead_to_left_pct, 1207 dead_to_right, live_to_right, 1208 double(dead_to_right) / live_to_right); 1209 } 1210 #endif // #ifndef PRODUCT 1211 1212 // Return a fraction indicating how much of the generation can be treated as 1213 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1214 // based on the density of live objects in the generation to determine a limit, 1215 // which is then adjusted so the return value is min_percent when the density is 1216 // 1. 1217 // 1218 // The following table shows some return values for a different values of the 1219 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1220 // min_percent is 1. 1221 // 1222 // fraction allowed as dead wood 1223 // ----------------------------------------------------------------- 1224 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1225 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1226 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1227 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1228 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1229 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1230 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1231 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1232 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1233 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1234 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1235 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1236 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1237 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1238 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1239 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1240 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1241 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1242 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1243 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1244 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1245 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1246 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1247 1248 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1249 { 1250 assert(_dwl_initialized, "uninitialized"); 1251 1252 // The raw limit is the value of the normal distribution at x = density. 1253 const double raw_limit = normal_distribution(density); 1254 1255 // Adjust the raw limit so it becomes the minimum when the density is 1. 1256 // 1257 // First subtract the adjustment value (which is simply the precomputed value 1258 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1259 // Then add the minimum value, so the minimum is returned when the density is 1260 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1261 const double min = double(min_percent) / 100.0; 1262 const double limit = raw_limit - _dwl_adjustment + min; 1263 return MAX2(limit, 0.0); 1264 } 1265 1266 ParallelCompactData::RegionData* 1267 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1268 const RegionData* end) 1269 { 1270 const size_t region_size = ParallelCompactData::RegionSize; 1271 ParallelCompactData& sd = summary_data(); 1272 size_t left = sd.region(beg); 1273 size_t right = end > beg ? sd.region(end) - 1 : left; 1274 1275 // Binary search. 1276 while (left < right) { 1277 // Equivalent to (left + right) / 2, but does not overflow. 1278 const size_t middle = left + (right - left) / 2; 1279 RegionData* const middle_ptr = sd.region(middle); 1280 HeapWord* const dest = middle_ptr->destination(); 1281 HeapWord* const addr = sd.region_to_addr(middle); 1282 assert(dest != NULL, "sanity"); 1283 assert(dest <= addr, "must move left"); 1284 1285 if (middle > left && dest < addr) { 1286 right = middle - 1; 1287 } else if (middle < right && middle_ptr->data_size() == region_size) { 1288 left = middle + 1; 1289 } else { 1290 return middle_ptr; 1291 } 1292 } 1293 return sd.region(left); 1294 } 1295 1296 ParallelCompactData::RegionData* 1297 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1298 const RegionData* end, 1299 size_t dead_words) 1300 { 1301 ParallelCompactData& sd = summary_data(); 1302 size_t left = sd.region(beg); 1303 size_t right = end > beg ? sd.region(end) - 1 : left; 1304 1305 // Binary search. 1306 while (left < right) { 1307 // Equivalent to (left + right) / 2, but does not overflow. 1308 const size_t middle = left + (right - left) / 2; 1309 RegionData* const middle_ptr = sd.region(middle); 1310 HeapWord* const dest = middle_ptr->destination(); 1311 HeapWord* const addr = sd.region_to_addr(middle); 1312 assert(dest != NULL, "sanity"); 1313 assert(dest <= addr, "must move left"); 1314 1315 const size_t dead_to_left = pointer_delta(addr, dest); 1316 if (middle > left && dead_to_left > dead_words) { 1317 right = middle - 1; 1318 } else if (middle < right && dead_to_left < dead_words) { 1319 left = middle + 1; 1320 } else { 1321 return middle_ptr; 1322 } 1323 } 1324 return sd.region(left); 1325 } 1326 1327 // The result is valid during the summary phase, after the initial summarization 1328 // of each space into itself, and before final summarization. 1329 inline double 1330 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1331 HeapWord* const bottom, 1332 HeapWord* const top, 1333 HeapWord* const new_top) 1334 { 1335 ParallelCompactData& sd = summary_data(); 1336 1337 assert(cp != NULL, "sanity"); 1338 assert(bottom != NULL, "sanity"); 1339 assert(top != NULL, "sanity"); 1340 assert(new_top != NULL, "sanity"); 1341 assert(top >= new_top, "summary data problem?"); 1342 assert(new_top > bottom, "space is empty; should not be here"); 1343 assert(new_top >= cp->destination(), "sanity"); 1344 assert(top >= sd.region_to_addr(cp), "sanity"); 1345 1346 HeapWord* const destination = cp->destination(); 1347 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1348 const size_t compacted_region_live = pointer_delta(new_top, destination); 1349 const size_t compacted_region_used = pointer_delta(top, 1350 sd.region_to_addr(cp)); 1351 const size_t reclaimable = compacted_region_used - compacted_region_live; 1352 1353 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1354 return double(reclaimable) / divisor; 1355 } 1356 1357 // Return the address of the end of the dense prefix, a.k.a. the start of the 1358 // compacted region. The address is always on a region boundary. 1359 // 1360 // Completely full regions at the left are skipped, since no compaction can 1361 // occur in those regions. Then the maximum amount of dead wood to allow is 1362 // computed, based on the density (amount live / capacity) of the generation; 1363 // the region with approximately that amount of dead space to the left is 1364 // identified as the limit region. Regions between the last completely full 1365 // region and the limit region are scanned and the one that has the best 1366 // (maximum) reclaimed_ratio() is selected. 1367 HeapWord* 1368 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1369 bool maximum_compaction) 1370 { 1371 const size_t region_size = ParallelCompactData::RegionSize; 1372 const ParallelCompactData& sd = summary_data(); 1373 1374 const MutableSpace* const space = _space_info[id].space(); 1375 HeapWord* const top = space->top(); 1376 HeapWord* const top_aligned_up = sd.region_align_up(top); 1377 HeapWord* const new_top = _space_info[id].new_top(); 1378 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1379 HeapWord* const bottom = space->bottom(); 1380 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1381 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1382 const RegionData* const new_top_cp = 1383 sd.addr_to_region_ptr(new_top_aligned_up); 1384 1385 // Skip full regions at the beginning of the space--they are necessarily part 1386 // of the dense prefix. 1387 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1388 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1389 space->is_empty(), "no dead space allowed to the left"); 1390 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1391 "region must have dead space"); 1392 1393 // The gc number is saved whenever a maximum compaction is done, and used to 1394 // determine when the maximum compaction interval has expired. This avoids 1395 // successive max compactions for different reasons. 1396 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1397 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1398 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1399 total_invocations() == HeapFirstMaximumCompactionCount; 1400 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1401 _maximum_compaction_gc_num = total_invocations(); 1402 return sd.region_to_addr(full_cp); 1403 } 1404 1405 const size_t space_live = pointer_delta(new_top, bottom); 1406 const size_t space_used = space->used_in_words(); 1407 const size_t space_capacity = space->capacity_in_words(); 1408 1409 const double density = double(space_live) / double(space_capacity); 1410 const size_t min_percent_free = MarkSweepDeadRatio; 1411 const double limiter = dead_wood_limiter(density, min_percent_free); 1412 const size_t dead_wood_max = space_used - space_live; 1413 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1414 dead_wood_max); 1415 1416 if (TraceParallelOldGCDensePrefix) { 1417 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1418 "space_cap=" SIZE_FORMAT, 1419 space_live, space_used, 1420 space_capacity); 1421 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1422 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1423 density, min_percent_free, limiter, 1424 dead_wood_max, dead_wood_limit); 1425 } 1426 1427 // Locate the region with the desired amount of dead space to the left. 1428 const RegionData* const limit_cp = 1429 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1430 1431 // Scan from the first region with dead space to the limit region and find the 1432 // one with the best (largest) reclaimed ratio. 1433 double best_ratio = 0.0; 1434 const RegionData* best_cp = full_cp; 1435 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1436 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1437 if (tmp_ratio > best_ratio) { 1438 best_cp = cp; 1439 best_ratio = tmp_ratio; 1440 } 1441 } 1442 1443 return sd.region_to_addr(best_cp); 1444 } 1445 1446 void PSParallelCompact::summarize_spaces_quick() 1447 { 1448 for (unsigned int i = 0; i < last_space_id; ++i) { 1449 const MutableSpace* space = _space_info[i].space(); 1450 HeapWord** nta = _space_info[i].new_top_addr(); 1451 bool result = _summary_data.summarize(_space_info[i].split_info(), 1452 space->bottom(), space->top(), NULL, 1453 space->bottom(), space->end(), nta); 1454 assert(result, "space must fit into itself"); 1455 _space_info[i].set_dense_prefix(space->bottom()); 1456 } 1457 } 1458 1459 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1460 { 1461 HeapWord* const dense_prefix_end = dense_prefix(id); 1462 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1463 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1464 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1465 // Only enough dead space is filled so that any remaining dead space to the 1466 // left is larger than the minimum filler object. (The remainder is filled 1467 // during the copy/update phase.) 1468 // 1469 // The size of the dead space to the right of the boundary is not a 1470 // concern, since compaction will be able to use whatever space is 1471 // available. 1472 // 1473 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1474 // surrounds the space to be filled with an object. 1475 // 1476 // In the 32-bit VM, each bit represents two 32-bit words: 1477 // +---+ 1478 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1479 // end_bits: ... x x x | 0 | || 0 x x ... 1480 // +---+ 1481 // 1482 // In the 64-bit VM, each bit represents one 64-bit word: 1483 // +------------+ 1484 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1485 // end_bits: ... x x 1 | 0 || 0 | x x ... 1486 // +------------+ 1487 // +-------+ 1488 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1489 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1490 // +-------+ 1491 // +-----------+ 1492 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1493 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1494 // +-----------+ 1495 // +-------+ 1496 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1497 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1498 // +-------+ 1499 1500 // Initially assume case a, c or e will apply. 1501 size_t obj_len = CollectedHeap::min_fill_size(); 1502 HeapWord* obj_beg = dense_prefix_end - obj_len; 1503 1504 #ifdef _LP64 1505 if (MinObjAlignment > 1) { // object alignment > heap word size 1506 // Cases a, c or e. 1507 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1508 // Case b above. 1509 obj_beg = dense_prefix_end - 1; 1510 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1511 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1512 // Case d above. 1513 obj_beg = dense_prefix_end - 3; 1514 obj_len = 3; 1515 } 1516 #endif // #ifdef _LP64 1517 1518 CollectedHeap::fill_with_object(obj_beg, obj_len); 1519 _mark_bitmap.mark_obj(obj_beg, obj_len); 1520 _summary_data.add_obj(obj_beg, obj_len); 1521 assert(start_array(id) != NULL, "sanity"); 1522 start_array(id)->allocate_block(obj_beg); 1523 } 1524 } 1525 1526 void 1527 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1528 { 1529 assert(id < last_space_id, "id out of range"); 1530 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1531 "should have been reset in summarize_spaces_quick()"); 1532 1533 const MutableSpace* space = _space_info[id].space(); 1534 if (_space_info[id].new_top() != space->bottom()) { 1535 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1536 _space_info[id].set_dense_prefix(dense_prefix_end); 1537 1538 #ifndef PRODUCT 1539 if (TraceParallelOldGCDensePrefix) { 1540 print_dense_prefix_stats("ratio", id, maximum_compaction, 1541 dense_prefix_end); 1542 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1543 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1544 } 1545 #endif // #ifndef PRODUCT 1546 1547 // Recompute the summary data, taking into account the dense prefix. If 1548 // every last byte will be reclaimed, then the existing summary data which 1549 // compacts everything can be left in place. 1550 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1551 // If dead space crosses the dense prefix boundary, it is (at least 1552 // partially) filled with a dummy object, marked live and added to the 1553 // summary data. This simplifies the copy/update phase and must be done 1554 // before the final locations of objects are determined, to prevent 1555 // leaving a fragment of dead space that is too small to fill. 1556 fill_dense_prefix_end(id); 1557 1558 // Compute the destination of each Region, and thus each object. 1559 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1560 _summary_data.summarize(_space_info[id].split_info(), 1561 dense_prefix_end, space->top(), NULL, 1562 dense_prefix_end, space->end(), 1563 _space_info[id].new_top_addr()); 1564 } 1565 } 1566 1567 if (log_develop_is_enabled(Trace, gc, compaction)) { 1568 const size_t region_size = ParallelCompactData::RegionSize; 1569 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1570 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1571 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1572 HeapWord* const new_top = _space_info[id].new_top(); 1573 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1574 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1575 log_develop_trace(gc, compaction)( 1576 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1577 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1578 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1579 id, space->capacity_in_words(), p2i(dense_prefix_end), 1580 dp_region, dp_words / region_size, 1581 cr_words / region_size, p2i(new_top)); 1582 } 1583 } 1584 1585 #ifndef PRODUCT 1586 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1587 HeapWord* dst_beg, HeapWord* dst_end, 1588 SpaceId src_space_id, 1589 HeapWord* src_beg, HeapWord* src_end) 1590 { 1591 log_develop_trace(gc, compaction)( 1592 "Summarizing %d [%s] into %d [%s]: " 1593 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1594 SIZE_FORMAT "-" SIZE_FORMAT " " 1595 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1596 SIZE_FORMAT "-" SIZE_FORMAT, 1597 src_space_id, space_names[src_space_id], 1598 dst_space_id, space_names[dst_space_id], 1599 p2i(src_beg), p2i(src_end), 1600 _summary_data.addr_to_region_idx(src_beg), 1601 _summary_data.addr_to_region_idx(src_end), 1602 p2i(dst_beg), p2i(dst_end), 1603 _summary_data.addr_to_region_idx(dst_beg), 1604 _summary_data.addr_to_region_idx(dst_end)); 1605 } 1606 #endif // #ifndef PRODUCT 1607 1608 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1609 bool maximum_compaction) 1610 { 1611 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1612 1613 #ifdef ASSERT 1614 if (TraceParallelOldGCMarkingPhase) { 1615 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1616 "add_obj_bytes=" SIZE_FORMAT, 1617 add_obj_count, add_obj_size * HeapWordSize); 1618 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1619 "mark_bitmap_bytes=" SIZE_FORMAT, 1620 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1621 } 1622 #endif // #ifdef ASSERT 1623 1624 // Quick summarization of each space into itself, to see how much is live. 1625 summarize_spaces_quick(); 1626 1627 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1628 NOT_PRODUCT(print_region_ranges()); 1629 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1630 1631 // The amount of live data that will end up in old space (assuming it fits). 1632 size_t old_space_total_live = 0; 1633 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1634 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1635 _space_info[id].space()->bottom()); 1636 } 1637 1638 MutableSpace* const old_space = _space_info[old_space_id].space(); 1639 const size_t old_capacity = old_space->capacity_in_words(); 1640 if (old_space_total_live > old_capacity) { 1641 // XXX - should also try to expand 1642 maximum_compaction = true; 1643 } 1644 1645 // Old generations. 1646 summarize_space(old_space_id, maximum_compaction); 1647 1648 // Summarize the remaining spaces in the young gen. The initial target space 1649 // is the old gen. If a space does not fit entirely into the target, then the 1650 // remainder is compacted into the space itself and that space becomes the new 1651 // target. 1652 SpaceId dst_space_id = old_space_id; 1653 HeapWord* dst_space_end = old_space->end(); 1654 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1655 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1656 const MutableSpace* space = _space_info[id].space(); 1657 const size_t live = pointer_delta(_space_info[id].new_top(), 1658 space->bottom()); 1659 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1660 1661 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1662 SpaceId(id), space->bottom(), space->top());) 1663 if (live > 0 && live <= available) { 1664 // All the live data will fit. 1665 bool done = _summary_data.summarize(_space_info[id].split_info(), 1666 space->bottom(), space->top(), 1667 NULL, 1668 *new_top_addr, dst_space_end, 1669 new_top_addr); 1670 assert(done, "space must fit into old gen"); 1671 1672 // Reset the new_top value for the space. 1673 _space_info[id].set_new_top(space->bottom()); 1674 } else if (live > 0) { 1675 // Attempt to fit part of the source space into the target space. 1676 HeapWord* next_src_addr = NULL; 1677 bool done = _summary_data.summarize(_space_info[id].split_info(), 1678 space->bottom(), space->top(), 1679 &next_src_addr, 1680 *new_top_addr, dst_space_end, 1681 new_top_addr); 1682 assert(!done, "space should not fit into old gen"); 1683 assert(next_src_addr != NULL, "sanity"); 1684 1685 // The source space becomes the new target, so the remainder is compacted 1686 // within the space itself. 1687 dst_space_id = SpaceId(id); 1688 dst_space_end = space->end(); 1689 new_top_addr = _space_info[id].new_top_addr(); 1690 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1691 space->bottom(), dst_space_end, 1692 SpaceId(id), next_src_addr, space->top());) 1693 done = _summary_data.summarize(_space_info[id].split_info(), 1694 next_src_addr, space->top(), 1695 NULL, 1696 space->bottom(), dst_space_end, 1697 new_top_addr); 1698 assert(done, "space must fit when compacted into itself"); 1699 assert(*new_top_addr <= space->top(), "usage should not grow"); 1700 } 1701 } 1702 1703 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1704 NOT_PRODUCT(print_region_ranges()); 1705 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1706 } 1707 1708 // This method should contain all heap-specific policy for invoking a full 1709 // collection. invoke_no_policy() will only attempt to compact the heap; it 1710 // will do nothing further. If we need to bail out for policy reasons, scavenge 1711 // before full gc, or any other specialized behavior, it needs to be added here. 1712 // 1713 // Note that this method should only be called from the vm_thread while at a 1714 // safepoint. 1715 // 1716 // Note that the all_soft_refs_clear flag in the soft ref policy 1717 // may be true because this method can be called without intervening 1718 // activity. For example when the heap space is tight and full measure 1719 // are being taken to free space. 1720 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1721 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1722 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1723 "should be in vm thread"); 1724 1725 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1726 GCCause::Cause gc_cause = heap->gc_cause(); 1727 assert(!heap->is_gc_active(), "not reentrant"); 1728 1729 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1730 IsGCActiveMark mark; 1731 1732 if (ScavengeBeforeFullGC) { 1733 PSScavenge::invoke_no_policy(); 1734 } 1735 1736 const bool clear_all_soft_refs = 1737 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1738 1739 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1740 maximum_heap_compaction); 1741 } 1742 1743 // This method contains no policy. You should probably 1744 // be calling invoke() instead. 1745 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1746 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1747 assert(ref_processor() != NULL, "Sanity"); 1748 1749 if (GCLocker::check_active_before_gc()) { 1750 return false; 1751 } 1752 1753 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1754 1755 GCIdMark gc_id_mark; 1756 _gc_timer.register_gc_start(); 1757 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1758 1759 TimeStamp marking_start; 1760 TimeStamp compaction_start; 1761 TimeStamp collection_exit; 1762 1763 GCCause::Cause gc_cause = heap->gc_cause(); 1764 PSYoungGen* young_gen = heap->young_gen(); 1765 PSOldGen* old_gen = heap->old_gen(); 1766 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1767 1768 // The scope of casr should end after code that can change 1769 // SoftRefPolicy::_should_clear_all_soft_refs. 1770 ClearedAllSoftRefs casr(maximum_heap_compaction, 1771 heap->soft_ref_policy()); 1772 1773 if (ZapUnusedHeapArea) { 1774 // Save information needed to minimize mangling 1775 heap->record_gen_tops_before_GC(); 1776 } 1777 1778 // Make sure data structures are sane, make the heap parsable, and do other 1779 // miscellaneous bookkeeping. 1780 pre_compact(); 1781 1782 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values(); 1783 1784 // Get the compaction manager reserved for the VM thread. 1785 ParCompactionManager* const vmthread_cm = 1786 ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers()); 1787 1788 { 1789 ResourceMark rm; 1790 HandleMark hm; 1791 1792 const uint active_workers = 1793 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(), 1794 ParallelScavengeHeap::heap()->workers().active_workers(), 1795 Threads::number_of_non_daemon_threads()); 1796 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers); 1797 1798 GCTraceCPUTime tcpu; 1799 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1800 1801 heap->pre_full_gc_dump(&_gc_timer); 1802 1803 TraceCollectorStats tcs(counters()); 1804 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1805 1806 if (log_is_enabled(Debug, gc, heap, exit)) { 1807 accumulated_time()->start(); 1808 } 1809 1810 // Let the size policy know we're starting 1811 size_policy->major_collection_begin(); 1812 1813 #if COMPILER2_OR_JVMCI 1814 DerivedPointerTable::clear(); 1815 #endif 1816 1817 ref_processor()->enable_discovery(); 1818 ref_processor()->setup_policy(maximum_heap_compaction); 1819 1820 bool marked_for_unloading = false; 1821 1822 marking_start.update(); 1823 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1824 1825 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1826 && GCCause::is_user_requested_gc(gc_cause); 1827 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1828 1829 #if COMPILER2_OR_JVMCI 1830 assert(DerivedPointerTable::is_active(), "Sanity"); 1831 DerivedPointerTable::set_active(false); 1832 #endif 1833 1834 // adjust_roots() updates Universe::_intArrayKlassObj which is 1835 // needed by the compaction for filling holes in the dense prefix. 1836 adjust_roots(vmthread_cm); 1837 1838 compaction_start.update(); 1839 compact(); 1840 1841 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1842 // done before resizing. 1843 post_compact(); 1844 1845 // Let the size policy know we're done 1846 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1847 1848 if (UseAdaptiveSizePolicy) { 1849 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1850 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1851 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1852 1853 // Don't check if the size_policy is ready here. Let 1854 // the size_policy check that internally. 1855 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1856 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1857 // Swap the survivor spaces if from_space is empty. The 1858 // resize_young_gen() called below is normally used after 1859 // a successful young GC and swapping of survivor spaces; 1860 // otherwise, it will fail to resize the young gen with 1861 // the current implementation. 1862 if (young_gen->from_space()->is_empty()) { 1863 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1864 young_gen->swap_spaces(); 1865 } 1866 1867 // Calculate optimal free space amounts 1868 assert(young_gen->max_size() > 1869 young_gen->from_space()->capacity_in_bytes() + 1870 young_gen->to_space()->capacity_in_bytes(), 1871 "Sizes of space in young gen are out-of-bounds"); 1872 1873 size_t young_live = young_gen->used_in_bytes(); 1874 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1875 size_t old_live = old_gen->used_in_bytes(); 1876 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1877 size_t max_old_gen_size = old_gen->max_gen_size(); 1878 size_t max_eden_size = young_gen->max_size() - 1879 young_gen->from_space()->capacity_in_bytes() - 1880 young_gen->to_space()->capacity_in_bytes(); 1881 1882 // Used for diagnostics 1883 size_policy->clear_generation_free_space_flags(); 1884 1885 size_policy->compute_generations_free_space(young_live, 1886 eden_live, 1887 old_live, 1888 cur_eden, 1889 max_old_gen_size, 1890 max_eden_size, 1891 true /* full gc*/); 1892 1893 size_policy->check_gc_overhead_limit(eden_live, 1894 max_old_gen_size, 1895 max_eden_size, 1896 true /* full gc*/, 1897 gc_cause, 1898 heap->soft_ref_policy()); 1899 1900 size_policy->decay_supplemental_growth(true /* full gc*/); 1901 1902 heap->resize_old_gen( 1903 size_policy->calculated_old_free_size_in_bytes()); 1904 1905 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1906 size_policy->calculated_survivor_size_in_bytes()); 1907 } 1908 1909 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1910 } 1911 1912 if (UsePerfData) { 1913 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1914 counters->update_counters(); 1915 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1916 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1917 } 1918 1919 heap->resize_all_tlabs(); 1920 1921 // Resize the metaspace capacity after a collection 1922 MetaspaceGC::compute_new_size(); 1923 1924 if (log_is_enabled(Debug, gc, heap, exit)) { 1925 accumulated_time()->stop(); 1926 } 1927 1928 heap->print_heap_change(pre_gc_values); 1929 1930 // Track memory usage and detect low memory 1931 MemoryService::track_memory_usage(); 1932 heap->update_counters(); 1933 1934 heap->post_full_gc_dump(&_gc_timer); 1935 } 1936 1937 #ifdef ASSERT 1938 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1939 ParCompactionManager* const cm = 1940 ParCompactionManager::manager_array(int(i)); 1941 assert(cm->marking_stack()->is_empty(), "should be empty"); 1942 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1943 } 1944 #endif // ASSERT 1945 1946 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1947 HandleMark hm; // Discard invalid handles created during verification 1948 Universe::verify("After GC"); 1949 } 1950 1951 // Re-verify object start arrays 1952 if (VerifyObjectStartArray && 1953 VerifyAfterGC) { 1954 old_gen->verify_object_start_array(); 1955 } 1956 1957 if (ZapUnusedHeapArea) { 1958 old_gen->object_space()->check_mangled_unused_area_complete(); 1959 } 1960 1961 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1962 1963 collection_exit.update(); 1964 1965 heap->print_heap_after_gc(); 1966 heap->trace_heap_after_gc(&_gc_tracer); 1967 1968 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1969 marking_start.ticks(), compaction_start.ticks(), 1970 collection_exit.ticks()); 1971 1972 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1973 1974 _gc_timer.register_gc_end(); 1975 1976 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1977 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1978 1979 return true; 1980 } 1981 1982 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1983 PSYoungGen* young_gen, 1984 PSOldGen* old_gen) { 1985 MutableSpace* const eden_space = young_gen->eden_space(); 1986 assert(!eden_space->is_empty(), "eden must be non-empty"); 1987 assert(young_gen->virtual_space()->alignment() == 1988 old_gen->virtual_space()->alignment(), "alignments do not match"); 1989 1990 // We also return false when it's a heterogeneous heap because old generation cannot absorb data from eden 1991 // when it is allocated on different memory (example, nv-dimm) than young. 1992 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) || 1993 ParallelArguments::is_heterogeneous_heap()) { 1994 return false; 1995 } 1996 1997 // Both generations must be completely committed. 1998 if (young_gen->virtual_space()->uncommitted_size() != 0) { 1999 return false; 2000 } 2001 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2002 return false; 2003 } 2004 2005 // Figure out how much to take from eden. Include the average amount promoted 2006 // in the total; otherwise the next young gen GC will simply bail out to a 2007 // full GC. 2008 const size_t alignment = old_gen->virtual_space()->alignment(); 2009 const size_t eden_used = eden_space->used_in_bytes(); 2010 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2011 const size_t absorb_size = align_up(eden_used + promoted, alignment); 2012 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2013 2014 if (absorb_size >= eden_capacity) { 2015 return false; // Must leave some space in eden. 2016 } 2017 2018 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2019 if (new_young_size < young_gen->min_gen_size()) { 2020 return false; // Respect young gen minimum size. 2021 } 2022 2023 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: " 2024 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2025 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2026 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2027 absorb_size / K, 2028 eden_capacity / K, (eden_capacity - absorb_size) / K, 2029 young_gen->from_space()->used_in_bytes() / K, 2030 young_gen->to_space()->used_in_bytes() / K, 2031 young_gen->capacity_in_bytes() / K, new_young_size / K); 2032 2033 // Fill the unused part of the old gen. 2034 MutableSpace* const old_space = old_gen->object_space(); 2035 HeapWord* const unused_start = old_space->top(); 2036 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2037 2038 if (unused_words > 0) { 2039 if (unused_words < CollectedHeap::min_fill_size()) { 2040 return false; // If the old gen cannot be filled, must give up. 2041 } 2042 CollectedHeap::fill_with_objects(unused_start, unused_words); 2043 } 2044 2045 // Take the live data from eden and set both top and end in the old gen to 2046 // eden top. (Need to set end because reset_after_change() mangles the region 2047 // from end to virtual_space->high() in debug builds). 2048 HeapWord* const new_top = eden_space->top(); 2049 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2050 absorb_size); 2051 young_gen->reset_after_change(); 2052 old_space->set_top(new_top); 2053 old_space->set_end(new_top); 2054 old_gen->reset_after_change(); 2055 2056 // Update the object start array for the filler object and the data from eden. 2057 ObjectStartArray* const start_array = old_gen->start_array(); 2058 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2059 start_array->allocate_block(p); 2060 } 2061 2062 // Could update the promoted average here, but it is not typically updated at 2063 // full GCs and the value to use is unclear. Something like 2064 // 2065 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2066 2067 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2068 return true; 2069 } 2070 2071 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 2072 private: 2073 uint _worker_id; 2074 2075 public: 2076 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { } 2077 void do_thread(Thread* thread) { 2078 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2079 2080 ResourceMark rm; 2081 2082 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id); 2083 2084 PCMarkAndPushClosure mark_and_push_closure(cm); 2085 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations); 2086 2087 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs); 2088 2089 // Do the real work 2090 cm->follow_marking_stacks(); 2091 } 2092 }; 2093 2094 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) { 2095 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2096 2097 ParCompactionManager* cm = 2098 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2099 PCMarkAndPushClosure mark_and_push_closure(cm); 2100 2101 switch (root_type) { 2102 case ParallelRootType::universe: 2103 Universe::oops_do(&mark_and_push_closure); 2104 break; 2105 2106 case ParallelRootType::jni_handles: 2107 JNIHandles::oops_do(&mark_and_push_closure); 2108 break; 2109 2110 case ParallelRootType::object_synchronizer: 2111 ObjectSynchronizer::oops_do(&mark_and_push_closure); 2112 break; 2113 2114 case ParallelRootType::management: 2115 Management::oops_do(&mark_and_push_closure); 2116 break; 2117 2118 case ParallelRootType::jvmti: 2119 JvmtiExport::oops_do(&mark_and_push_closure); 2120 break; 2121 2122 case ParallelRootType::system_dictionary: 2123 SystemDictionary::oops_do(&mark_and_push_closure); 2124 break; 2125 2126 case ParallelRootType::class_loader_data: 2127 { 2128 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong); 2129 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure); 2130 } 2131 break; 2132 2133 case ParallelRootType::code_cache: 2134 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them. 2135 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure)); 2136 AOTLoader::oops_do(&mark_and_push_closure); 2137 break; 2138 2139 case ParallelRootType::sentinel: 2140 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds 2141 fatal("Bad enumeration value: %u", root_type); 2142 break; 2143 } 2144 2145 // Do the real work 2146 cm->follow_marking_stacks(); 2147 } 2148 2149 static void steal_marking_work(TaskTerminator& terminator, uint worker_id) { 2150 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2151 2152 ParCompactionManager* cm = 2153 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2154 2155 oop obj = NULL; 2156 ObjArrayTask task; 2157 do { 2158 while (ParCompactionManager::steal_objarray(worker_id, task)) { 2159 cm->follow_array((objArrayOop)task.obj(), task.index()); 2160 cm->follow_marking_stacks(); 2161 } 2162 while (ParCompactionManager::steal(worker_id, obj)) { 2163 cm->follow_contents(obj); 2164 cm->follow_marking_stacks(); 2165 } 2166 } while (!terminator.offer_termination()); 2167 } 2168 2169 class MarkFromRootsTask : public AbstractGangTask { 2170 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2171 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do 2172 SequentialSubTasksDone _subtasks; 2173 TaskTerminator _terminator; 2174 uint _active_workers; 2175 2176 public: 2177 MarkFromRootsTask(uint active_workers) : 2178 AbstractGangTask("MarkFromRootsTask"), 2179 _strong_roots_scope(active_workers), 2180 _subtasks(), 2181 _terminator(active_workers, ParCompactionManager::stack_array()), 2182 _active_workers(active_workers) { 2183 _subtasks.set_n_threads(active_workers); 2184 _subtasks.set_n_tasks(ParallelRootType::sentinel); 2185 } 2186 2187 virtual void work(uint worker_id) { 2188 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) { 2189 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id); 2190 } 2191 _subtasks.all_tasks_completed(); 2192 2193 PCAddThreadRootsMarkingTaskClosure closure(worker_id); 2194 Threads::possibly_parallel_threads_do(true /*parallel */, &closure); 2195 2196 if (_active_workers > 1) { 2197 steal_marking_work(_terminator, worker_id); 2198 } 2199 } 2200 }; 2201 2202 class PCRefProcTask : public AbstractGangTask { 2203 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2204 ProcessTask& _task; 2205 uint _ergo_workers; 2206 TaskTerminator _terminator; 2207 2208 public: 2209 PCRefProcTask(ProcessTask& task, uint ergo_workers) : 2210 AbstractGangTask("PCRefProcTask"), 2211 _task(task), 2212 _ergo_workers(ergo_workers), 2213 _terminator(_ergo_workers, ParCompactionManager::stack_array()) { 2214 } 2215 2216 virtual void work(uint worker_id) { 2217 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2218 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2219 2220 ParCompactionManager* cm = 2221 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2222 PCMarkAndPushClosure mark_and_push_closure(cm); 2223 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2224 _task.work(worker_id, *PSParallelCompact::is_alive_closure(), 2225 mark_and_push_closure, follow_stack_closure); 2226 2227 steal_marking_work(_terminator, worker_id); 2228 } 2229 }; 2230 2231 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2232 void execute(ProcessTask& process_task, uint ergo_workers) { 2233 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers, 2234 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 2235 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers()); 2236 2237 PCRefProcTask task(process_task, ergo_workers); 2238 ParallelScavengeHeap::heap()->workers().run_task(&task); 2239 } 2240 }; 2241 2242 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2243 bool maximum_heap_compaction, 2244 ParallelOldTracer *gc_tracer) { 2245 // Recursively traverse all live objects and mark them 2246 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2247 2248 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2249 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2250 2251 PCMarkAndPushClosure mark_and_push_closure(cm); 2252 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2253 2254 // Need new claim bits before marking starts. 2255 ClassLoaderDataGraph::clear_claimed_marks(); 2256 2257 { 2258 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2259 2260 MarkFromRootsTask task(active_gc_threads); 2261 ParallelScavengeHeap::heap()->workers().run_task(&task); 2262 } 2263 2264 // Process reference objects found during marking 2265 { 2266 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2267 2268 ReferenceProcessorStats stats; 2269 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2270 2271 if (ref_processor()->processing_is_mt()) { 2272 ref_processor()->set_active_mt_degree(active_gc_threads); 2273 2274 RefProcTaskExecutor task_executor; 2275 stats = ref_processor()->process_discovered_references( 2276 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2277 &task_executor, &pt); 2278 } else { 2279 stats = ref_processor()->process_discovered_references( 2280 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2281 &pt); 2282 } 2283 2284 gc_tracer->report_gc_reference_stats(stats); 2285 pt.print_all_references(); 2286 } 2287 2288 // This is the point where the entire marking should have completed. 2289 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2290 2291 { 2292 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2293 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2294 } 2295 2296 { 2297 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2298 2299 // Follow system dictionary roots and unload classes. 2300 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2301 2302 // Unload nmethods. 2303 CodeCache::do_unloading(is_alive_closure(), purged_class); 2304 2305 // Prune dead klasses from subklass/sibling/implementor lists. 2306 Klass::clean_weak_klass_links(purged_class); 2307 2308 // Clean JVMCI metadata handles. 2309 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 2310 } 2311 2312 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2313 } 2314 2315 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2316 // Adjust the pointers to reflect the new locations 2317 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2318 2319 // Need new claim bits when tracing through and adjusting pointers. 2320 ClassLoaderDataGraph::clear_claimed_marks(); 2321 2322 PCAdjustPointerClosure oop_closure(cm); 2323 2324 // General strong roots. 2325 Universe::oops_do(&oop_closure); 2326 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2327 Threads::oops_do(&oop_closure, NULL); 2328 ObjectSynchronizer::oops_do(&oop_closure); 2329 Management::oops_do(&oop_closure); 2330 JvmtiExport::oops_do(&oop_closure); 2331 SystemDictionary::oops_do(&oop_closure); 2332 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong); 2333 ClassLoaderDataGraph::cld_do(&cld_closure); 2334 2335 // Now adjust pointers in remaining weak roots. (All of which should 2336 // have been cleared if they pointed to non-surviving objects.) 2337 WeakProcessor::oops_do(&oop_closure); 2338 2339 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2340 CodeCache::blobs_do(&adjust_from_blobs); 2341 AOT_ONLY(AOTLoader::oops_do(&oop_closure);) 2342 2343 ref_processor()->weak_oops_do(&oop_closure); 2344 // Roots were visited so references into the young gen in roots 2345 // may have been scanned. Process them also. 2346 // Should the reference processor have a span that excludes 2347 // young gen objects? 2348 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2349 } 2350 2351 // Helper class to print 8 region numbers per line and then print the total at the end. 2352 class FillableRegionLogger : public StackObj { 2353 private: 2354 Log(gc, compaction) log; 2355 static const int LineLength = 8; 2356 size_t _regions[LineLength]; 2357 int _next_index; 2358 bool _enabled; 2359 size_t _total_regions; 2360 public: 2361 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2362 ~FillableRegionLogger() { 2363 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2364 } 2365 2366 void print_line() { 2367 if (!_enabled || _next_index == 0) { 2368 return; 2369 } 2370 FormatBuffer<> line("Fillable: "); 2371 for (int i = 0; i < _next_index; i++) { 2372 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2373 } 2374 log.trace("%s", line.buffer()); 2375 _next_index = 0; 2376 } 2377 2378 void handle(size_t region) { 2379 if (!_enabled) { 2380 return; 2381 } 2382 _regions[_next_index++] = region; 2383 if (_next_index == LineLength) { 2384 print_line(); 2385 } 2386 _total_regions++; 2387 } 2388 }; 2389 2390 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads) 2391 { 2392 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2393 2394 // Find the threads that are active 2395 uint worker_id = 0; 2396 2397 // Find all regions that are available (can be filled immediately) and 2398 // distribute them to the thread stacks. The iteration is done in reverse 2399 // order (high to low) so the regions will be removed in ascending order. 2400 2401 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2402 2403 // id + 1 is used to test termination so unsigned can 2404 // be used with an old_space_id == 0. 2405 FillableRegionLogger region_logger; 2406 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2407 SpaceInfo* const space_info = _space_info + id; 2408 MutableSpace* const space = space_info->space(); 2409 HeapWord* const new_top = space_info->new_top(); 2410 2411 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2412 const size_t end_region = 2413 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2414 2415 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2416 if (sd.region(cur)->claim_unsafe()) { 2417 ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id); 2418 bool result = sd.region(cur)->mark_normal(); 2419 assert(result, "Must succeed at this point."); 2420 cm->region_stack()->push(cur); 2421 region_logger.handle(cur); 2422 // Assign regions to tasks in round-robin fashion. 2423 if (++worker_id == parallel_gc_threads) { 2424 worker_id = 0; 2425 } 2426 } 2427 } 2428 region_logger.print_line(); 2429 } 2430 } 2431 2432 class TaskQueue : StackObj { 2433 volatile uint _counter; 2434 uint _size; 2435 uint _insert_index; 2436 PSParallelCompact::UpdateDensePrefixTask* _backing_array; 2437 public: 2438 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) { 2439 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC); 2440 } 2441 ~TaskQueue() { 2442 assert(_counter >= _insert_index, "not all queue elements were claimed"); 2443 FREE_C_HEAP_ARRAY(T, _backing_array); 2444 } 2445 2446 void push(const PSParallelCompact::UpdateDensePrefixTask& value) { 2447 assert(_insert_index < _size, "too small backing array"); 2448 _backing_array[_insert_index++] = value; 2449 } 2450 2451 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) { 2452 uint claimed = Atomic::fetch_and_add(&_counter, 1u); 2453 if (claimed < _insert_index) { 2454 reference = _backing_array[claimed]; 2455 return true; 2456 } else { 2457 return false; 2458 } 2459 } 2460 }; 2461 2462 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2463 2464 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue, 2465 uint parallel_gc_threads) { 2466 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2467 2468 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2469 2470 // Iterate over all the spaces adding tasks for updating 2471 // regions in the dense prefix. Assume that 1 gc thread 2472 // will work on opening the gaps and the remaining gc threads 2473 // will work on the dense prefix. 2474 unsigned int space_id; 2475 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2476 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2477 const MutableSpace* const space = _space_info[space_id].space(); 2478 2479 if (dense_prefix_end == space->bottom()) { 2480 // There is no dense prefix for this space. 2481 continue; 2482 } 2483 2484 // The dense prefix is before this region. 2485 size_t region_index_end_dense_prefix = 2486 sd.addr_to_region_idx(dense_prefix_end); 2487 RegionData* const dense_prefix_cp = 2488 sd.region(region_index_end_dense_prefix); 2489 assert(dense_prefix_end == space->end() || 2490 dense_prefix_cp->available() || 2491 dense_prefix_cp->claimed(), 2492 "The region after the dense prefix should always be ready to fill"); 2493 2494 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2495 2496 // Is there dense prefix work? 2497 size_t total_dense_prefix_regions = 2498 region_index_end_dense_prefix - region_index_start; 2499 // How many regions of the dense prefix should be given to 2500 // each thread? 2501 if (total_dense_prefix_regions > 0) { 2502 uint tasks_for_dense_prefix = 1; 2503 if (total_dense_prefix_regions <= 2504 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2505 // Don't over partition. This assumes that 2506 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2507 // so there are not many regions to process. 2508 tasks_for_dense_prefix = parallel_gc_threads; 2509 } else { 2510 // Over partition 2511 tasks_for_dense_prefix = parallel_gc_threads * 2512 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2513 } 2514 size_t regions_per_thread = total_dense_prefix_regions / 2515 tasks_for_dense_prefix; 2516 // Give each thread at least 1 region. 2517 if (regions_per_thread == 0) { 2518 regions_per_thread = 1; 2519 } 2520 2521 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2522 if (region_index_start >= region_index_end_dense_prefix) { 2523 break; 2524 } 2525 // region_index_end is not processed 2526 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2527 region_index_end_dense_prefix); 2528 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2529 region_index_start, 2530 region_index_end)); 2531 region_index_start = region_index_end; 2532 } 2533 } 2534 // This gets any part of the dense prefix that did not 2535 // fit evenly. 2536 if (region_index_start < region_index_end_dense_prefix) { 2537 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2538 region_index_start, 2539 region_index_end_dense_prefix)); 2540 } 2541 } 2542 } 2543 2544 #ifdef ASSERT 2545 // Write a histogram of the number of times the block table was filled for a 2546 // region. 2547 void PSParallelCompact::write_block_fill_histogram() 2548 { 2549 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2550 return; 2551 } 2552 2553 Log(gc, compaction) log; 2554 ResourceMark rm; 2555 LogStream ls(log.trace()); 2556 outputStream* out = &ls; 2557 2558 typedef ParallelCompactData::RegionData rd_t; 2559 ParallelCompactData& sd = summary_data(); 2560 2561 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2562 MutableSpace* const spc = _space_info[id].space(); 2563 if (spc->bottom() != spc->top()) { 2564 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2565 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2566 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2567 2568 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2569 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2570 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2571 2572 for (const rd_t* cur = beg; cur < end; ++cur) { 2573 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2574 } 2575 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2576 for (size_t i = 0; i < histo_len; ++i) { 2577 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2578 histo[i], 100.0 * histo[i] / region_cnt); 2579 } 2580 out->cr(); 2581 } 2582 } 2583 } 2584 #endif // #ifdef ASSERT 2585 2586 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) { 2587 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2588 2589 ParCompactionManager* cm = 2590 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2591 2592 // Drain the stacks that have been preloaded with regions 2593 // that are ready to fill. 2594 2595 cm->drain_region_stacks(); 2596 2597 guarantee(cm->region_stack()->is_empty(), "Not empty"); 2598 2599 size_t region_index = 0; 2600 2601 while (true) { 2602 if (ParCompactionManager::steal(worker_id, region_index)) { 2603 PSParallelCompact::fill_and_update_region(cm, region_index); 2604 cm->drain_region_stacks(); 2605 } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) { 2606 // Fill and update an unavailable region with the help of a shadow region 2607 PSParallelCompact::fill_and_update_shadow_region(cm, region_index); 2608 cm->drain_region_stacks(); 2609 } else { 2610 if (terminator->offer_termination()) { 2611 break; 2612 } 2613 // Go around again. 2614 } 2615 } 2616 return; 2617 } 2618 2619 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask { 2620 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2621 TaskQueue& _tq; 2622 TaskTerminator _terminator; 2623 uint _active_workers; 2624 2625 public: 2626 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) : 2627 AbstractGangTask("UpdateDensePrefixAndCompactionTask"), 2628 _tq(tq), 2629 _terminator(active_workers, ParCompactionManager::region_array()), 2630 _active_workers(active_workers) { 2631 } 2632 virtual void work(uint worker_id) { 2633 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2634 2635 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) { 2636 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm, 2637 task._space_id, 2638 task._region_index_start, 2639 task._region_index_end); 2640 } 2641 2642 // Once a thread has drained it's stack, it should try to steal regions from 2643 // other threads. 2644 compaction_with_stealing_work(&_terminator, worker_id); 2645 } 2646 }; 2647 2648 void PSParallelCompact::compact() { 2649 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2650 2651 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2652 PSOldGen* old_gen = heap->old_gen(); 2653 old_gen->start_array()->reset(); 2654 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2655 2656 // for [0..last_space_id) 2657 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING) 2658 // push 2659 // push 2660 // 2661 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1) 2662 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)); 2663 initialize_shadow_regions(active_gc_threads); 2664 prepare_region_draining_tasks(active_gc_threads); 2665 enqueue_dense_prefix_tasks(task_queue, active_gc_threads); 2666 2667 { 2668 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2669 2670 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads); 2671 ParallelScavengeHeap::heap()->workers().run_task(&task); 2672 2673 #ifdef ASSERT 2674 // Verify that all regions have been processed before the deferred updates. 2675 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2676 verify_complete(SpaceId(id)); 2677 } 2678 #endif 2679 } 2680 2681 { 2682 // Update the deferred objects, if any. Any compaction manager can be used. 2683 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2684 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2685 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2686 update_deferred_objects(cm, SpaceId(id)); 2687 } 2688 } 2689 2690 DEBUG_ONLY(write_block_fill_histogram()); 2691 } 2692 2693 #ifdef ASSERT 2694 void PSParallelCompact::verify_complete(SpaceId space_id) { 2695 // All Regions between space bottom() to new_top() should be marked as filled 2696 // and all Regions between new_top() and top() should be available (i.e., 2697 // should have been emptied). 2698 ParallelCompactData& sd = summary_data(); 2699 SpaceInfo si = _space_info[space_id]; 2700 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2701 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2702 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2703 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2704 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2705 2706 bool issued_a_warning = false; 2707 2708 size_t cur_region; 2709 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2710 const RegionData* const c = sd.region(cur_region); 2711 if (!c->completed()) { 2712 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2713 cur_region, c->destination_count()); 2714 issued_a_warning = true; 2715 } 2716 } 2717 2718 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2719 const RegionData* const c = sd.region(cur_region); 2720 if (!c->available()) { 2721 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2722 cur_region, c->destination_count()); 2723 issued_a_warning = true; 2724 } 2725 } 2726 2727 if (issued_a_warning) { 2728 print_region_ranges(); 2729 } 2730 } 2731 #endif // #ifdef ASSERT 2732 2733 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2734 _start_array->allocate_block(addr); 2735 compaction_manager()->update_contents(oop(addr)); 2736 } 2737 2738 // Update interior oops in the ranges of regions [beg_region, end_region). 2739 void 2740 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2741 SpaceId space_id, 2742 size_t beg_region, 2743 size_t end_region) { 2744 ParallelCompactData& sd = summary_data(); 2745 ParMarkBitMap* const mbm = mark_bitmap(); 2746 2747 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2748 HeapWord* const end_addr = sd.region_to_addr(end_region); 2749 assert(beg_region <= end_region, "bad region range"); 2750 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2751 2752 #ifdef ASSERT 2753 // Claim the regions to avoid triggering an assert when they are marked as 2754 // filled. 2755 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2756 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2757 } 2758 #endif // #ifdef ASSERT 2759 2760 if (beg_addr != space(space_id)->bottom()) { 2761 // Find the first live object or block of dead space that *starts* in this 2762 // range of regions. If a partial object crosses onto the region, skip it; 2763 // it will be marked for 'deferred update' when the object head is 2764 // processed. If dead space crosses onto the region, it is also skipped; it 2765 // will be filled when the prior region is processed. If neither of those 2766 // apply, the first word in the region is the start of a live object or dead 2767 // space. 2768 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2769 const RegionData* const cp = sd.region(beg_region); 2770 if (cp->partial_obj_size() != 0) { 2771 beg_addr = sd.partial_obj_end(beg_region); 2772 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2773 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2774 } 2775 } 2776 2777 if (beg_addr < end_addr) { 2778 // A live object or block of dead space starts in this range of Regions. 2779 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2780 2781 // Create closures and iterate. 2782 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2783 FillClosure fill_closure(cm, space_id); 2784 ParMarkBitMap::IterationStatus status; 2785 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2786 dense_prefix_end); 2787 if (status == ParMarkBitMap::incomplete) { 2788 update_closure.do_addr(update_closure.source()); 2789 } 2790 } 2791 2792 // Mark the regions as filled. 2793 RegionData* const beg_cp = sd.region(beg_region); 2794 RegionData* const end_cp = sd.region(end_region); 2795 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2796 cp->set_completed(); 2797 } 2798 } 2799 2800 // Return the SpaceId for the space containing addr. If addr is not in the 2801 // heap, last_space_id is returned. In debug mode it expects the address to be 2802 // in the heap and asserts such. 2803 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2804 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2805 2806 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2807 if (_space_info[id].space()->contains(addr)) { 2808 return SpaceId(id); 2809 } 2810 } 2811 2812 assert(false, "no space contains the addr"); 2813 return last_space_id; 2814 } 2815 2816 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2817 SpaceId id) { 2818 assert(id < last_space_id, "bad space id"); 2819 2820 ParallelCompactData& sd = summary_data(); 2821 const SpaceInfo* const space_info = _space_info + id; 2822 ObjectStartArray* const start_array = space_info->start_array(); 2823 2824 const MutableSpace* const space = space_info->space(); 2825 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2826 HeapWord* const beg_addr = space_info->dense_prefix(); 2827 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2828 2829 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2830 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2831 const RegionData* cur_region; 2832 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2833 HeapWord* const addr = cur_region->deferred_obj_addr(); 2834 if (addr != NULL) { 2835 if (start_array != NULL) { 2836 start_array->allocate_block(addr); 2837 } 2838 cm->update_contents(oop(addr)); 2839 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2840 } 2841 } 2842 } 2843 2844 // Skip over count live words starting from beg, and return the address of the 2845 // next live word. Unless marked, the word corresponding to beg is assumed to 2846 // be dead. Callers must either ensure beg does not correspond to the middle of 2847 // an object, or account for those live words in some other way. Callers must 2848 // also ensure that there are enough live words in the range [beg, end) to skip. 2849 HeapWord* 2850 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2851 { 2852 assert(count > 0, "sanity"); 2853 2854 ParMarkBitMap* m = mark_bitmap(); 2855 idx_t bits_to_skip = m->words_to_bits(count); 2856 idx_t cur_beg = m->addr_to_bit(beg); 2857 const idx_t search_end = m->align_range_end(m->addr_to_bit(end)); 2858 2859 do { 2860 cur_beg = m->find_obj_beg(cur_beg, search_end); 2861 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2862 const size_t obj_bits = cur_end - cur_beg + 1; 2863 if (obj_bits > bits_to_skip) { 2864 return m->bit_to_addr(cur_beg + bits_to_skip); 2865 } 2866 bits_to_skip -= obj_bits; 2867 cur_beg = cur_end + 1; 2868 } while (bits_to_skip > 0); 2869 2870 // Skipping the desired number of words landed just past the end of an object. 2871 // Find the start of the next object. 2872 cur_beg = m->find_obj_beg(cur_beg, search_end); 2873 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2874 return m->bit_to_addr(cur_beg); 2875 } 2876 2877 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2878 SpaceId src_space_id, 2879 size_t src_region_idx) 2880 { 2881 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2882 2883 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2884 if (split_info.dest_region_addr() == dest_addr) { 2885 // The partial object ending at the split point contains the first word to 2886 // be copied to dest_addr. 2887 return split_info.first_src_addr(); 2888 } 2889 2890 const ParallelCompactData& sd = summary_data(); 2891 ParMarkBitMap* const bitmap = mark_bitmap(); 2892 const size_t RegionSize = ParallelCompactData::RegionSize; 2893 2894 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2895 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2896 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2897 HeapWord* const src_region_destination = src_region_ptr->destination(); 2898 2899 assert(dest_addr >= src_region_destination, "wrong src region"); 2900 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2901 2902 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2903 HeapWord* const src_region_end = src_region_beg + RegionSize; 2904 2905 HeapWord* addr = src_region_beg; 2906 if (dest_addr == src_region_destination) { 2907 // Return the first live word in the source region. 2908 if (partial_obj_size == 0) { 2909 addr = bitmap->find_obj_beg(addr, src_region_end); 2910 assert(addr < src_region_end, "no objects start in src region"); 2911 } 2912 return addr; 2913 } 2914 2915 // Must skip some live data. 2916 size_t words_to_skip = dest_addr - src_region_destination; 2917 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2918 2919 if (partial_obj_size >= words_to_skip) { 2920 // All the live words to skip are part of the partial object. 2921 addr += words_to_skip; 2922 if (partial_obj_size == words_to_skip) { 2923 // Find the first live word past the partial object. 2924 addr = bitmap->find_obj_beg(addr, src_region_end); 2925 assert(addr < src_region_end, "wrong src region"); 2926 } 2927 return addr; 2928 } 2929 2930 // Skip over the partial object (if any). 2931 if (partial_obj_size != 0) { 2932 words_to_skip -= partial_obj_size; 2933 addr += partial_obj_size; 2934 } 2935 2936 // Skip over live words due to objects that start in the region. 2937 addr = skip_live_words(addr, src_region_end, words_to_skip); 2938 assert(addr < src_region_end, "wrong src region"); 2939 return addr; 2940 } 2941 2942 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2943 SpaceId src_space_id, 2944 size_t beg_region, 2945 HeapWord* end_addr) 2946 { 2947 ParallelCompactData& sd = summary_data(); 2948 2949 #ifdef ASSERT 2950 MutableSpace* const src_space = _space_info[src_space_id].space(); 2951 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2952 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2953 "src_space_id does not match beg_addr"); 2954 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2955 "src_space_id does not match end_addr"); 2956 #endif // #ifdef ASSERT 2957 2958 RegionData* const beg = sd.region(beg_region); 2959 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2960 2961 // Regions up to new_top() are enqueued if they become available. 2962 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2963 RegionData* const enqueue_end = 2964 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2965 2966 for (RegionData* cur = beg; cur < end; ++cur) { 2967 assert(cur->data_size() > 0, "region must have live data"); 2968 cur->decrement_destination_count(); 2969 if (cur < enqueue_end && cur->available() && cur->claim()) { 2970 if (cur->mark_normal()) { 2971 cm->push_region(sd.region(cur)); 2972 } else if (cur->mark_copied()) { 2973 // Try to copy the content of the shadow region back to its corresponding 2974 // heap region if the shadow region is filled. Otherwise, the GC thread 2975 // fills the shadow region will copy the data back (see 2976 // MoveAndUpdateShadowClosure::complete_region). 2977 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur)); 2978 ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region()); 2979 cur->set_completed(); 2980 } 2981 } 2982 } 2983 } 2984 2985 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2986 SpaceId& src_space_id, 2987 HeapWord*& src_space_top, 2988 HeapWord* end_addr) 2989 { 2990 typedef ParallelCompactData::RegionData RegionData; 2991 2992 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2993 const size_t region_size = ParallelCompactData::RegionSize; 2994 2995 size_t src_region_idx = 0; 2996 2997 // Skip empty regions (if any) up to the top of the space. 2998 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2999 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3000 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3001 const RegionData* const top_region_ptr = 3002 sd.addr_to_region_ptr(top_aligned_up); 3003 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3004 ++src_region_ptr; 3005 } 3006 3007 if (src_region_ptr < top_region_ptr) { 3008 // The next source region is in the current space. Update src_region_idx 3009 // and the source address to match src_region_ptr. 3010 src_region_idx = sd.region(src_region_ptr); 3011 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3012 if (src_region_addr > closure.source()) { 3013 closure.set_source(src_region_addr); 3014 } 3015 return src_region_idx; 3016 } 3017 3018 // Switch to a new source space and find the first non-empty region. 3019 unsigned int space_id = src_space_id + 1; 3020 assert(space_id < last_space_id, "not enough spaces"); 3021 3022 HeapWord* const destination = closure.destination(); 3023 3024 do { 3025 MutableSpace* space = _space_info[space_id].space(); 3026 HeapWord* const bottom = space->bottom(); 3027 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3028 3029 // Iterate over the spaces that do not compact into themselves. 3030 if (bottom_cp->destination() != bottom) { 3031 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3032 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3033 3034 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3035 if (src_cp->live_obj_size() > 0) { 3036 // Found it. 3037 assert(src_cp->destination() == destination, 3038 "first live obj in the space must match the destination"); 3039 assert(src_cp->partial_obj_size() == 0, 3040 "a space cannot begin with a partial obj"); 3041 3042 src_space_id = SpaceId(space_id); 3043 src_space_top = space->top(); 3044 const size_t src_region_idx = sd.region(src_cp); 3045 closure.set_source(sd.region_to_addr(src_region_idx)); 3046 return src_region_idx; 3047 } else { 3048 assert(src_cp->data_size() == 0, "sanity"); 3049 } 3050 } 3051 } 3052 } while (++space_id < last_space_id); 3053 3054 assert(false, "no source region was found"); 3055 return 0; 3056 } 3057 3058 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx) 3059 { 3060 typedef ParMarkBitMap::IterationStatus IterationStatus; 3061 ParMarkBitMap* const bitmap = mark_bitmap(); 3062 ParallelCompactData& sd = summary_data(); 3063 RegionData* const region_ptr = sd.region(region_idx); 3064 3065 // Get the source region and related info. 3066 size_t src_region_idx = region_ptr->source_region(); 3067 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3068 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3069 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3070 3071 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3072 3073 // Adjust src_region_idx to prepare for decrementing destination counts (the 3074 // destination count is not decremented when a region is copied to itself). 3075 if (src_region_idx == region_idx) { 3076 src_region_idx += 1; 3077 } 3078 3079 if (bitmap->is_unmarked(closure.source())) { 3080 // The first source word is in the middle of an object; copy the remainder 3081 // of the object or as much as will fit. The fact that pointer updates were 3082 // deferred will be noted when the object header is processed. 3083 HeapWord* const old_src_addr = closure.source(); 3084 closure.copy_partial_obj(); 3085 if (closure.is_full()) { 3086 decrement_destination_counts(cm, src_space_id, src_region_idx, 3087 closure.source()); 3088 region_ptr->set_deferred_obj_addr(NULL); 3089 closure.complete_region(cm, dest_addr, region_ptr); 3090 return; 3091 } 3092 3093 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3094 if (sd.region_align_down(old_src_addr) != end_addr) { 3095 // The partial object was copied from more than one source region. 3096 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3097 3098 // Move to the next source region, possibly switching spaces as well. All 3099 // args except end_addr may be modified. 3100 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3101 end_addr); 3102 } 3103 } 3104 3105 do { 3106 HeapWord* const cur_addr = closure.source(); 3107 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3108 src_space_top); 3109 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3110 3111 if (status == ParMarkBitMap::incomplete) { 3112 // The last obj that starts in the source region does not end in the 3113 // region. 3114 assert(closure.source() < end_addr, "sanity"); 3115 HeapWord* const obj_beg = closure.source(); 3116 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3117 src_space_top); 3118 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3119 if (obj_end < range_end) { 3120 // The end was found; the entire object will fit. 3121 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3122 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3123 } else { 3124 // The end was not found; the object will not fit. 3125 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3126 status = ParMarkBitMap::would_overflow; 3127 } 3128 } 3129 3130 if (status == ParMarkBitMap::would_overflow) { 3131 // The last object did not fit. Note that interior oop updates were 3132 // deferred, then copy enough of the object to fill the region. 3133 region_ptr->set_deferred_obj_addr(closure.destination()); 3134 status = closure.copy_until_full(); // copies from closure.source() 3135 3136 decrement_destination_counts(cm, src_space_id, src_region_idx, 3137 closure.source()); 3138 closure.complete_region(cm, dest_addr, region_ptr); 3139 return; 3140 } 3141 3142 if (status == ParMarkBitMap::full) { 3143 decrement_destination_counts(cm, src_space_id, src_region_idx, 3144 closure.source()); 3145 region_ptr->set_deferred_obj_addr(NULL); 3146 closure.complete_region(cm, dest_addr, region_ptr); 3147 return; 3148 } 3149 3150 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3151 3152 // Move to the next source region, possibly switching spaces as well. All 3153 // args except end_addr may be modified. 3154 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3155 end_addr); 3156 } while (true); 3157 } 3158 3159 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx) 3160 { 3161 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3162 fill_region(cm, cl, region_idx); 3163 } 3164 3165 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx) 3166 { 3167 // Get a shadow region first 3168 ParallelCompactData& sd = summary_data(); 3169 RegionData* const region_ptr = sd.region(region_idx); 3170 size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr); 3171 // The InvalidShadow return value indicates the corresponding heap region is available, 3172 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use 3173 // MoveAndUpdateShadowClosure to fill the acquired shadow region. 3174 if (shadow_region == ParCompactionManager::InvalidShadow) { 3175 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3176 region_ptr->shadow_to_normal(); 3177 return fill_region(cm, cl, region_idx); 3178 } else { 3179 MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region); 3180 return fill_region(cm, cl, region_idx); 3181 } 3182 } 3183 3184 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr) 3185 { 3186 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize); 3187 } 3188 3189 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t ®ion_idx) 3190 { 3191 size_t next = cm->next_shadow_region(); 3192 ParallelCompactData& sd = summary_data(); 3193 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top()); 3194 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 3195 3196 while (next < old_new_top) { 3197 if (sd.region(next)->mark_shadow()) { 3198 region_idx = next; 3199 return true; 3200 } 3201 next = cm->move_next_shadow_region_by(active_gc_threads); 3202 } 3203 3204 return false; 3205 } 3206 3207 // The shadow region is an optimization to address region dependencies in full GC. The basic 3208 // idea is making more regions available by temporally storing their live objects in empty 3209 // shadow regions to resolve dependencies between them and the destination regions. Therefore, 3210 // GC threads need not wait destination regions to be available before processing sources. 3211 // 3212 // A typical workflow would be: 3213 // After draining its own stack and failing to steal from others, a GC worker would pick an 3214 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills 3215 // the shadow region by copying live objects from source regions of the unavailable one. Once 3216 // the unavailable region becomes available, the data in the shadow region will be copied back. 3217 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces. 3218 // 3219 // For more details, please refer to ยง4.2 of the VEE'19 paper: 3220 // Haoyu Li, Mingyu Wu, Binyu Zang, and Haibo Chen. 2019. ScissorGC: scalable and efficient 3221 // compaction for Java full garbage collection. In Proceedings of the 15th ACM SIGPLAN/SIGOPS 3222 // International Conference on Virtual Execution Environments (VEE 2019). ACM, New York, NY, USA, 3223 // 108-121. DOI: https://doi.org/10.1145/3313808.3313820 3224 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads) 3225 { 3226 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 3227 3228 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 3229 SpaceInfo* const space_info = _space_info + id; 3230 MutableSpace* const space = space_info->space(); 3231 3232 const size_t beg_region = 3233 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top()))); 3234 const size_t end_region = 3235 sd.addr_to_region_idx(sd.region_align_down(space->end())); 3236 3237 for (size_t cur = beg_region; cur < end_region; ++cur) { 3238 ParCompactionManager::push_shadow_region(cur); 3239 } 3240 } 3241 3242 size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix()); 3243 for (uint i = 0; i < parallel_gc_threads; i++) { 3244 ParCompactionManager *cm = ParCompactionManager::manager_array(i); 3245 cm->set_next_shadow_region(beg_region + i); 3246 } 3247 } 3248 3249 void PSParallelCompact::fill_blocks(size_t region_idx) 3250 { 3251 // Fill in the block table elements for the specified region. Each block 3252 // table element holds the number of live words in the region that are to the 3253 // left of the first object that starts in the block. Thus only blocks in 3254 // which an object starts need to be filled. 3255 // 3256 // The algorithm scans the section of the bitmap that corresponds to the 3257 // region, keeping a running total of the live words. When an object start is 3258 // found, if it's the first to start in the block that contains it, the 3259 // current total is written to the block table element. 3260 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3261 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3262 const size_t RegionSize = ParallelCompactData::RegionSize; 3263 3264 ParallelCompactData& sd = summary_data(); 3265 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3266 if (partial_obj_size >= RegionSize) { 3267 return; // No objects start in this region. 3268 } 3269 3270 // Ensure the first loop iteration decides that the block has changed. 3271 size_t cur_block = sd.block_count(); 3272 3273 const ParMarkBitMap* const bitmap = mark_bitmap(); 3274 3275 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3276 assert((size_t)1 << Log2BitsPerBlock == 3277 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3278 3279 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3280 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3281 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3282 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3283 while (beg_bit < range_end) { 3284 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3285 if (new_block != cur_block) { 3286 cur_block = new_block; 3287 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3288 } 3289 3290 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3291 if (end_bit < range_end - 1) { 3292 live_bits += end_bit - beg_bit + 1; 3293 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3294 } else { 3295 return; 3296 } 3297 } 3298 } 3299 3300 jlong PSParallelCompact::millis_since_last_gc() { 3301 // We need a monotonically non-decreasing time in ms but 3302 // os::javaTimeMillis() does not guarantee monotonicity. 3303 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3304 jlong ret_val = now - _time_of_last_gc; 3305 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3306 if (ret_val < 0) { 3307 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3308 return 0; 3309 } 3310 return ret_val; 3311 } 3312 3313 void PSParallelCompact::reset_millis_since_last_gc() { 3314 // We need a monotonically non-decreasing time in ms but 3315 // os::javaTimeMillis() does not guarantee monotonicity. 3316 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3317 } 3318 3319 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3320 { 3321 if (source() != copy_destination()) { 3322 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3323 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining()); 3324 } 3325 update_state(words_remaining()); 3326 assert(is_full(), "sanity"); 3327 return ParMarkBitMap::full; 3328 } 3329 3330 void MoveAndUpdateClosure::copy_partial_obj() 3331 { 3332 size_t words = words_remaining(); 3333 3334 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3335 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3336 if (end_addr < range_end) { 3337 words = bitmap()->obj_size(source(), end_addr); 3338 } 3339 3340 // This test is necessary; if omitted, the pointer updates to a partial object 3341 // that crosses the dense prefix boundary could be overwritten. 3342 if (source() != copy_destination()) { 3343 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3344 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3345 } 3346 update_state(words); 3347 } 3348 3349 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3350 PSParallelCompact::RegionData *region_ptr) { 3351 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished"); 3352 region_ptr->set_completed(); 3353 } 3354 3355 ParMarkBitMapClosure::IterationStatus 3356 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3357 assert(destination() != NULL, "sanity"); 3358 assert(bitmap()->obj_size(addr) == words, "bad size"); 3359 3360 _source = addr; 3361 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3362 destination(), "wrong destination"); 3363 3364 if (words > words_remaining()) { 3365 return ParMarkBitMap::would_overflow; 3366 } 3367 3368 // The start_array must be updated even if the object is not moving. 3369 if (_start_array != NULL) { 3370 _start_array->allocate_block(destination()); 3371 } 3372 3373 if (copy_destination() != source()) { 3374 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3375 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3376 } 3377 3378 oop moved_oop = (oop) copy_destination(); 3379 compaction_manager()->update_contents(moved_oop); 3380 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3381 3382 update_state(words); 3383 assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity"); 3384 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3385 } 3386 3387 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3388 PSParallelCompact::RegionData *region_ptr) { 3389 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow"); 3390 // Record the shadow region index 3391 region_ptr->set_shadow_region(_shadow); 3392 // Mark the shadow region as filled to indicate the data is ready to be 3393 // copied back 3394 region_ptr->mark_filled(); 3395 // Try to copy the content of the shadow region back to its corresponding 3396 // heap region if available; the GC thread that decreases the destination 3397 // count to zero will do the copying otherwise (see 3398 // PSParallelCompact::decrement_destination_counts). 3399 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) { 3400 region_ptr->set_completed(); 3401 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr); 3402 ParCompactionManager::push_shadow_region_mt_safe(_shadow); 3403 } 3404 } 3405 3406 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3407 ParCompactionManager* cm, 3408 PSParallelCompact::SpaceId space_id) : 3409 ParMarkBitMapClosure(mbm, cm), 3410 _space_id(space_id), 3411 _start_array(PSParallelCompact::start_array(space_id)) 3412 { 3413 } 3414 3415 // Updates the references in the object to their new values. 3416 ParMarkBitMapClosure::IterationStatus 3417 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3418 do_addr(addr); 3419 return ParMarkBitMap::incomplete; 3420 } 3421 3422 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3423 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3424 _start_array(PSParallelCompact::start_array(space_id)) 3425 { 3426 assert(space_id == PSParallelCompact::old_space_id, 3427 "cannot use FillClosure in the young gen"); 3428 } 3429 3430 ParMarkBitMapClosure::IterationStatus 3431 FillClosure::do_addr(HeapWord* addr, size_t size) { 3432 CollectedHeap::fill_with_objects(addr, size); 3433 HeapWord* const end = addr + size; 3434 do { 3435 _start_array->allocate_block(addr); 3436 addr += oop(addr)->size(); 3437 } while (addr < end); 3438 return ParMarkBitMap::incomplete; 3439 }