1 /* 2 * Copyright (c) 2005, 2019, 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/weakProcessor.hpp" 57 #include "gc/shared/workerPolicy.hpp" 58 #include "gc/shared/workgroup.hpp" 59 #include "logging/log.hpp" 60 #include "memory/iterator.inline.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "memory/universe.hpp" 63 #include "oops/access.inline.hpp" 64 #include "oops/instanceClassLoaderKlass.inline.hpp" 65 #include "oops/instanceKlass.inline.hpp" 66 #include "oops/instanceMirrorKlass.inline.hpp" 67 #include "oops/methodData.hpp" 68 #include "oops/objArrayKlass.inline.hpp" 69 #include "oops/oop.inline.hpp" 70 #include "runtime/atomic.hpp" 71 #include "runtime/handles.inline.hpp" 72 #include "runtime/safepoint.hpp" 73 #include "runtime/vmThread.hpp" 74 #include "services/management.hpp" 75 #include "services/memTracker.hpp" 76 #include "services/memoryService.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/debug.hpp" 79 #include "utilities/events.hpp" 80 #include "utilities/formatBuffer.hpp" 81 #include "utilities/macros.hpp" 82 #include "utilities/stack.inline.hpp" 83 #if INCLUDE_JVMCI 84 #include "jvmci/jvmci.hpp" 85 #endif 86 87 #include <math.h> 88 89 // All sizes are in HeapWords. 90 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 91 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 92 const size_t ParallelCompactData::RegionSizeBytes = 93 RegionSize << LogHeapWordSize; 94 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 95 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 96 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 97 98 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 99 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 100 const size_t ParallelCompactData::BlockSizeBytes = 101 BlockSize << LogHeapWordSize; 102 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 103 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 104 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 105 106 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 107 const size_t ParallelCompactData::Log2BlocksPerRegion = 108 Log2RegionSize - Log2BlockSize; 109 110 const ParallelCompactData::RegionData::region_sz_t 111 ParallelCompactData::RegionData::dc_shift = 27; 112 113 const ParallelCompactData::RegionData::region_sz_t 114 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 115 116 const ParallelCompactData::RegionData::region_sz_t 117 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 118 119 const ParallelCompactData::RegionData::region_sz_t 120 ParallelCompactData::RegionData::los_mask = ~dc_mask; 121 122 const ParallelCompactData::RegionData::region_sz_t 123 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 124 125 const ParallelCompactData::RegionData::region_sz_t 126 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 127 128 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 129 130 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 131 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 132 133 double PSParallelCompact::_dwl_mean; 134 double PSParallelCompact::_dwl_std_dev; 135 double PSParallelCompact::_dwl_first_term; 136 double PSParallelCompact::_dwl_adjustment; 137 #ifdef ASSERT 138 bool PSParallelCompact::_dwl_initialized = false; 139 #endif // #ifdef ASSERT 140 141 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 142 HeapWord* destination) 143 { 144 assert(src_region_idx != 0, "invalid src_region_idx"); 145 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 146 assert(destination != NULL, "invalid destination argument"); 147 148 _src_region_idx = src_region_idx; 149 _partial_obj_size = partial_obj_size; 150 _destination = destination; 151 152 // These fields may not be updated below, so make sure they're clear. 153 assert(_dest_region_addr == NULL, "should have been cleared"); 154 assert(_first_src_addr == NULL, "should have been cleared"); 155 156 // Determine the number of destination regions for the partial object. 157 HeapWord* const last_word = destination + partial_obj_size - 1; 158 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 159 HeapWord* const beg_region_addr = sd.region_align_down(destination); 160 HeapWord* const end_region_addr = sd.region_align_down(last_word); 161 162 if (beg_region_addr == end_region_addr) { 163 // One destination region. 164 _destination_count = 1; 165 if (end_region_addr == destination) { 166 // The destination falls on a region boundary, thus the first word of the 167 // partial object will be the first word copied to the destination region. 168 _dest_region_addr = end_region_addr; 169 _first_src_addr = sd.region_to_addr(src_region_idx); 170 } 171 } else { 172 // Two destination regions. When copied, the partial object will cross a 173 // destination region boundary, so a word somewhere within the partial 174 // object will be the first word copied to the second destination region. 175 _destination_count = 2; 176 _dest_region_addr = end_region_addr; 177 const size_t ofs = pointer_delta(end_region_addr, destination); 178 assert(ofs < _partial_obj_size, "sanity"); 179 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 180 } 181 } 182 183 void SplitInfo::clear() 184 { 185 _src_region_idx = 0; 186 _partial_obj_size = 0; 187 _destination = NULL; 188 _destination_count = 0; 189 _dest_region_addr = NULL; 190 _first_src_addr = NULL; 191 assert(!is_valid(), "sanity"); 192 } 193 194 #ifdef ASSERT 195 void SplitInfo::verify_clear() 196 { 197 assert(_src_region_idx == 0, "not clear"); 198 assert(_partial_obj_size == 0, "not clear"); 199 assert(_destination == NULL, "not clear"); 200 assert(_destination_count == 0, "not clear"); 201 assert(_dest_region_addr == NULL, "not clear"); 202 assert(_first_src_addr == NULL, "not clear"); 203 } 204 #endif // #ifdef ASSERT 205 206 207 void PSParallelCompact::print_on_error(outputStream* st) { 208 _mark_bitmap.print_on_error(st); 209 } 210 211 #ifndef PRODUCT 212 const char* PSParallelCompact::space_names[] = { 213 "old ", "eden", "from", "to " 214 }; 215 216 void PSParallelCompact::print_region_ranges() { 217 if (!log_develop_is_enabled(Trace, gc, compaction)) { 218 return; 219 } 220 Log(gc, compaction) log; 221 ResourceMark rm; 222 LogStream ls(log.trace()); 223 Universe::print_on(&ls); 224 log.trace("space bottom top end new_top"); 225 log.trace("------ ---------- ---------- ---------- ----------"); 226 227 for (unsigned int id = 0; id < last_space_id; ++id) { 228 const MutableSpace* space = _space_info[id].space(); 229 log.trace("%u %s " 230 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 231 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 232 id, space_names[id], 233 summary_data().addr_to_region_idx(space->bottom()), 234 summary_data().addr_to_region_idx(space->top()), 235 summary_data().addr_to_region_idx(space->end()), 236 summary_data().addr_to_region_idx(_space_info[id].new_top())); 237 } 238 } 239 240 void 241 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 242 { 243 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 244 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 245 246 ParallelCompactData& sd = PSParallelCompact::summary_data(); 247 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 248 log_develop_trace(gc, compaction)( 249 REGION_IDX_FORMAT " " PTR_FORMAT " " 250 REGION_IDX_FORMAT " " PTR_FORMAT " " 251 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 252 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 253 i, p2i(c->data_location()), dci, p2i(c->destination()), 254 c->partial_obj_size(), c->live_obj_size(), 255 c->data_size(), c->source_region(), c->destination_count()); 256 257 #undef REGION_IDX_FORMAT 258 #undef REGION_DATA_FORMAT 259 } 260 261 void 262 print_generic_summary_data(ParallelCompactData& summary_data, 263 HeapWord* const beg_addr, 264 HeapWord* const end_addr) 265 { 266 size_t total_words = 0; 267 size_t i = summary_data.addr_to_region_idx(beg_addr); 268 const size_t last = summary_data.addr_to_region_idx(end_addr); 269 HeapWord* pdest = 0; 270 271 while (i < last) { 272 ParallelCompactData::RegionData* c = summary_data.region(i); 273 if (c->data_size() != 0 || c->destination() != pdest) { 274 print_generic_summary_region(i, c); 275 total_words += c->data_size(); 276 pdest = c->destination(); 277 } 278 ++i; 279 } 280 281 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 282 } 283 284 void 285 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 286 HeapWord* const beg_addr, 287 HeapWord* const end_addr) { 288 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 289 } 290 291 void 292 print_generic_summary_data(ParallelCompactData& summary_data, 293 SpaceInfo* space_info) 294 { 295 if (!log_develop_is_enabled(Trace, gc, compaction)) { 296 return; 297 } 298 299 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 300 const MutableSpace* space = space_info[id].space(); 301 print_generic_summary_data(summary_data, space->bottom(), 302 MAX2(space->top(), space_info[id].new_top())); 303 } 304 } 305 306 void 307 print_initial_summary_data(ParallelCompactData& summary_data, 308 const MutableSpace* space) { 309 if (space->top() == space->bottom()) { 310 return; 311 } 312 313 const size_t region_size = ParallelCompactData::RegionSize; 314 typedef ParallelCompactData::RegionData RegionData; 315 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 316 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 317 const RegionData* c = summary_data.region(end_region - 1); 318 HeapWord* end_addr = c->destination() + c->data_size(); 319 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 320 321 // Print (and count) the full regions at the beginning of the space. 322 size_t full_region_count = 0; 323 size_t i = summary_data.addr_to_region_idx(space->bottom()); 324 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 325 ParallelCompactData::RegionData* c = summary_data.region(i); 326 log_develop_trace(gc, compaction)( 327 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 328 i, p2i(c->destination()), 329 c->partial_obj_size(), c->live_obj_size(), 330 c->data_size(), c->source_region(), c->destination_count()); 331 ++full_region_count; 332 ++i; 333 } 334 335 size_t live_to_right = live_in_space - full_region_count * region_size; 336 337 double max_reclaimed_ratio = 0.0; 338 size_t max_reclaimed_ratio_region = 0; 339 size_t max_dead_to_right = 0; 340 size_t max_live_to_right = 0; 341 342 // Print the 'reclaimed ratio' for regions while there is something live in 343 // the region or to the right of it. The remaining regions are empty (and 344 // uninteresting), and computing the ratio will result in division by 0. 345 while (i < end_region && live_to_right > 0) { 346 c = summary_data.region(i); 347 HeapWord* const region_addr = summary_data.region_to_addr(i); 348 const size_t used_to_right = pointer_delta(space->top(), region_addr); 349 const size_t dead_to_right = used_to_right - live_to_right; 350 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 351 352 if (reclaimed_ratio > max_reclaimed_ratio) { 353 max_reclaimed_ratio = reclaimed_ratio; 354 max_reclaimed_ratio_region = i; 355 max_dead_to_right = dead_to_right; 356 max_live_to_right = live_to_right; 357 } 358 359 ParallelCompactData::RegionData* c = summary_data.region(i); 360 log_develop_trace(gc, compaction)( 361 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 362 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 363 i, p2i(c->destination()), 364 c->partial_obj_size(), c->live_obj_size(), 365 c->data_size(), c->source_region(), c->destination_count(), 366 reclaimed_ratio, dead_to_right, live_to_right); 367 368 369 live_to_right -= c->data_size(); 370 ++i; 371 } 372 373 // Any remaining regions are empty. Print one more if there is one. 374 if (i < end_region) { 375 ParallelCompactData::RegionData* c = summary_data.region(i); 376 log_develop_trace(gc, compaction)( 377 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 378 i, p2i(c->destination()), 379 c->partial_obj_size(), c->live_obj_size(), 380 c->data_size(), c->source_region(), c->destination_count()); 381 } 382 383 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 384 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 385 } 386 387 void 388 print_initial_summary_data(ParallelCompactData& summary_data, 389 SpaceInfo* space_info) { 390 if (!log_develop_is_enabled(Trace, gc, compaction)) { 391 return; 392 } 393 394 unsigned int id = PSParallelCompact::old_space_id; 395 const MutableSpace* space; 396 do { 397 space = space_info[id].space(); 398 print_initial_summary_data(summary_data, space); 399 } while (++id < PSParallelCompact::eden_space_id); 400 401 do { 402 space = space_info[id].space(); 403 print_generic_summary_data(summary_data, space->bottom(), space->top()); 404 } while (++id < PSParallelCompact::last_space_id); 405 } 406 #endif // #ifndef PRODUCT 407 408 #ifdef ASSERT 409 size_t add_obj_count; 410 size_t add_obj_size; 411 size_t mark_bitmap_count; 412 size_t mark_bitmap_size; 413 #endif // #ifdef ASSERT 414 415 ParallelCompactData::ParallelCompactData() : 416 _region_start(NULL), 417 DEBUG_ONLY(_region_end(NULL) COMMA) 418 _region_vspace(NULL), 419 _reserved_byte_size(0), 420 _region_data(NULL), 421 _region_count(0), 422 _block_vspace(NULL), 423 _block_data(NULL), 424 _block_count(0) {} 425 426 bool ParallelCompactData::initialize(MemRegion covered_region) 427 { 428 _region_start = covered_region.start(); 429 const size_t region_size = covered_region.word_size(); 430 DEBUG_ONLY(_region_end = _region_start + region_size;) 431 432 assert(region_align_down(_region_start) == _region_start, 433 "region start not aligned"); 434 assert((region_size & RegionSizeOffsetMask) == 0, 435 "region size not a multiple of RegionSize"); 436 437 bool result = initialize_region_data(region_size) && initialize_block_data(); 438 return result; 439 } 440 441 PSVirtualSpace* 442 ParallelCompactData::create_vspace(size_t count, size_t element_size) 443 { 444 const size_t raw_bytes = count * element_size; 445 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 446 const size_t granularity = os::vm_allocation_granularity(); 447 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 448 449 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 450 MAX2(page_sz, granularity); 451 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 452 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 453 rs.size()); 454 455 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 456 457 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 458 if (vspace != 0) { 459 if (vspace->expand_by(_reserved_byte_size)) { 460 return vspace; 461 } 462 delete vspace; 463 // Release memory reserved in the space. 464 rs.release(); 465 } 466 467 return 0; 468 } 469 470 bool ParallelCompactData::initialize_region_data(size_t region_size) 471 { 472 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 473 _region_vspace = create_vspace(count, sizeof(RegionData)); 474 if (_region_vspace != 0) { 475 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 476 _region_count = count; 477 return true; 478 } 479 return false; 480 } 481 482 bool ParallelCompactData::initialize_block_data() 483 { 484 assert(_region_count != 0, "region data must be initialized first"); 485 const size_t count = _region_count << Log2BlocksPerRegion; 486 _block_vspace = create_vspace(count, sizeof(BlockData)); 487 if (_block_vspace != 0) { 488 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 489 _block_count = count; 490 return true; 491 } 492 return false; 493 } 494 495 void ParallelCompactData::clear() 496 { 497 memset(_region_data, 0, _region_vspace->committed_size()); 498 memset(_block_data, 0, _block_vspace->committed_size()); 499 } 500 501 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 502 assert(beg_region <= _region_count, "beg_region out of range"); 503 assert(end_region <= _region_count, "end_region out of range"); 504 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 505 506 const size_t region_cnt = end_region - beg_region; 507 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 508 509 const size_t beg_block = beg_region * BlocksPerRegion; 510 const size_t block_cnt = region_cnt * BlocksPerRegion; 511 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 512 } 513 514 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 515 { 516 const RegionData* cur_cp = region(region_idx); 517 const RegionData* const end_cp = region(region_count() - 1); 518 519 HeapWord* result = region_to_addr(region_idx); 520 if (cur_cp < end_cp) { 521 do { 522 result += cur_cp->partial_obj_size(); 523 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 524 } 525 return result; 526 } 527 528 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 529 { 530 const size_t obj_ofs = pointer_delta(addr, _region_start); 531 const size_t beg_region = obj_ofs >> Log2RegionSize; 532 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 533 534 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 535 DEBUG_ONLY(Atomic::add(len, &add_obj_size);) 536 537 if (beg_region == end_region) { 538 // All in one region. 539 _region_data[beg_region].add_live_obj(len); 540 return; 541 } 542 543 // First region. 544 const size_t beg_ofs = region_offset(addr); 545 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 546 547 Klass* klass = ((oop)addr)->klass(); 548 // Middle regions--completely spanned by this object. 549 for (size_t region = beg_region + 1; region < end_region; ++region) { 550 _region_data[region].set_partial_obj_size(RegionSize); 551 _region_data[region].set_partial_obj_addr(addr); 552 } 553 554 // Last region. 555 const size_t end_ofs = region_offset(addr + len - 1); 556 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 557 _region_data[end_region].set_partial_obj_addr(addr); 558 } 559 560 void 561 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 562 { 563 assert(region_offset(beg) == 0, "not RegionSize aligned"); 564 assert(region_offset(end) == 0, "not RegionSize aligned"); 565 566 size_t cur_region = addr_to_region_idx(beg); 567 const size_t end_region = addr_to_region_idx(end); 568 HeapWord* addr = beg; 569 while (cur_region < end_region) { 570 _region_data[cur_region].set_destination(addr); 571 _region_data[cur_region].set_destination_count(0); 572 _region_data[cur_region].set_source_region(cur_region); 573 _region_data[cur_region].set_data_location(addr); 574 575 // Update live_obj_size so the region appears completely full. 576 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 577 _region_data[cur_region].set_live_obj_size(live_size); 578 579 ++cur_region; 580 addr += RegionSize; 581 } 582 } 583 584 // Find the point at which a space can be split and, if necessary, record the 585 // split point. 586 // 587 // If the current src region (which overflowed the destination space) doesn't 588 // have a partial object, the split point is at the beginning of the current src 589 // region (an "easy" split, no extra bookkeeping required). 590 // 591 // If the current src region has a partial object, the split point is in the 592 // region where that partial object starts (call it the split_region). If 593 // split_region has a partial object, then the split point is just after that 594 // partial object (a "hard" split where we have to record the split data and 595 // zero the partial_obj_size field). With a "hard" split, we know that the 596 // partial_obj ends within split_region because the partial object that caused 597 // the overflow starts in split_region. If split_region doesn't have a partial 598 // obj, then the split is at the beginning of split_region (another "easy" 599 // split). 600 HeapWord* 601 ParallelCompactData::summarize_split_space(size_t src_region, 602 SplitInfo& split_info, 603 HeapWord* destination, 604 HeapWord* target_end, 605 HeapWord** target_next) 606 { 607 assert(destination <= target_end, "sanity"); 608 assert(destination + _region_data[src_region].data_size() > target_end, 609 "region should not fit into target space"); 610 assert(is_region_aligned(target_end), "sanity"); 611 612 size_t split_region = src_region; 613 HeapWord* split_destination = destination; 614 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 615 616 if (destination + partial_obj_size > target_end) { 617 // The split point is just after the partial object (if any) in the 618 // src_region that contains the start of the object that overflowed the 619 // destination space. 620 // 621 // Find the start of the "overflow" object and set split_region to the 622 // region containing it. 623 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 624 split_region = addr_to_region_idx(overflow_obj); 625 626 // Clear the source_region field of all destination regions whose first word 627 // came from data after the split point (a non-null source_region field 628 // implies a region must be filled). 629 // 630 // An alternative to the simple loop below: clear during post_compact(), 631 // which uses memcpy instead of individual stores, and is easy to 632 // parallelize. (The downside is that it clears the entire RegionData 633 // object as opposed to just one field.) 634 // 635 // post_compact() would have to clear the summary data up to the highest 636 // address that was written during the summary phase, which would be 637 // 638 // max(top, max(new_top, clear_top)) 639 // 640 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 641 // to target_end. 642 const RegionData* const sr = region(split_region); 643 const size_t beg_idx = 644 addr_to_region_idx(region_align_up(sr->destination() + 645 sr->partial_obj_size())); 646 const size_t end_idx = addr_to_region_idx(target_end); 647 648 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 649 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 650 _region_data[idx].set_source_region(0); 651 } 652 653 // Set split_destination and partial_obj_size to reflect the split region. 654 split_destination = sr->destination(); 655 partial_obj_size = sr->partial_obj_size(); 656 } 657 658 // The split is recorded only if a partial object extends onto the region. 659 if (partial_obj_size != 0) { 660 _region_data[split_region].set_partial_obj_size(0); 661 split_info.record(split_region, partial_obj_size, split_destination); 662 } 663 664 // Setup the continuation addresses. 665 *target_next = split_destination + partial_obj_size; 666 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 667 668 if (log_develop_is_enabled(Trace, gc, compaction)) { 669 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 670 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 671 split_type, p2i(source_next), split_region, partial_obj_size); 672 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 673 split_type, p2i(split_destination), 674 addr_to_region_idx(split_destination), 675 p2i(*target_next)); 676 677 if (partial_obj_size != 0) { 678 HeapWord* const po_beg = split_info.destination(); 679 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 680 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 681 split_type, 682 p2i(po_beg), addr_to_region_idx(po_beg), 683 p2i(po_end), addr_to_region_idx(po_end)); 684 } 685 } 686 687 return source_next; 688 } 689 690 bool ParallelCompactData::summarize(SplitInfo& split_info, 691 HeapWord* source_beg, HeapWord* source_end, 692 HeapWord** source_next, 693 HeapWord* target_beg, HeapWord* target_end, 694 HeapWord** target_next) 695 { 696 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 697 log_develop_trace(gc, compaction)( 698 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 699 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 700 p2i(source_beg), p2i(source_end), p2i(source_next_val), 701 p2i(target_beg), p2i(target_end), p2i(*target_next)); 702 703 size_t cur_region = addr_to_region_idx(source_beg); 704 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 705 706 HeapWord *dest_addr = target_beg; 707 while (cur_region < end_region) { 708 // The destination must be set even if the region has no data. 709 _region_data[cur_region].set_destination(dest_addr); 710 711 size_t words = _region_data[cur_region].data_size(); 712 if (words > 0) { 713 // If cur_region does not fit entirely into the target space, find a point 714 // at which the source space can be 'split' so that part is copied to the 715 // target space and the rest is copied elsewhere. 716 if (dest_addr + words > target_end) { 717 assert(source_next != NULL, "source_next is NULL when splitting"); 718 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 719 target_end, target_next); 720 return false; 721 } 722 723 // Compute the destination_count for cur_region, and if necessary, update 724 // source_region for a destination region. The source_region field is 725 // updated if cur_region is the first (left-most) region to be copied to a 726 // destination region. 727 // 728 // The destination_count calculation is a bit subtle. A region that has 729 // data that compacts into itself does not count itself as a destination. 730 // This maintains the invariant that a zero count means the region is 731 // available and can be claimed and then filled. 732 uint destination_count = 0; 733 if (split_info.is_split(cur_region)) { 734 // The current region has been split: the partial object will be copied 735 // to one destination space and the remaining data will be copied to 736 // another destination space. Adjust the initial destination_count and, 737 // if necessary, set the source_region field if the partial object will 738 // cross a destination region boundary. 739 destination_count = split_info.destination_count(); 740 if (destination_count == 2) { 741 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 742 _region_data[dest_idx].set_source_region(cur_region); 743 } 744 } 745 746 HeapWord* const last_addr = dest_addr + words - 1; 747 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 748 const size_t dest_region_2 = addr_to_region_idx(last_addr); 749 750 // Initially assume that the destination regions will be the same and 751 // adjust the value below if necessary. Under this assumption, if 752 // cur_region == dest_region_2, then cur_region will be compacted 753 // completely into itself. 754 destination_count += cur_region == dest_region_2 ? 0 : 1; 755 if (dest_region_1 != dest_region_2) { 756 // Destination regions differ; adjust destination_count. 757 destination_count += 1; 758 // Data from cur_region will be copied to the start of dest_region_2. 759 _region_data[dest_region_2].set_source_region(cur_region); 760 } else if (region_offset(dest_addr) == 0) { 761 // Data from cur_region will be copied to the start of the destination 762 // region. 763 _region_data[dest_region_1].set_source_region(cur_region); 764 } 765 766 _region_data[cur_region].set_destination_count(destination_count); 767 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 768 dest_addr += words; 769 } 770 771 ++cur_region; 772 } 773 774 *target_next = dest_addr; 775 return true; 776 } 777 778 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 779 assert(addr != NULL, "Should detect NULL oop earlier"); 780 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 781 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 782 783 // Region covering the object. 784 RegionData* const region_ptr = addr_to_region_ptr(addr); 785 HeapWord* result = region_ptr->destination(); 786 787 // If the entire Region is live, the new location is region->destination + the 788 // offset of the object within in the Region. 789 790 // Run some performance tests to determine if this special case pays off. It 791 // is worth it for pointers into the dense prefix. If the optimization to 792 // avoid pointer updates in regions that only point to the dense prefix is 793 // ever implemented, this should be revisited. 794 if (region_ptr->data_size() == RegionSize) { 795 result += region_offset(addr); 796 return result; 797 } 798 799 // Otherwise, the new location is region->destination + block offset + the 800 // number of live words in the Block that are (a) to the left of addr and (b) 801 // due to objects that start in the Block. 802 803 // Fill in the block table if necessary. This is unsynchronized, so multiple 804 // threads may fill the block table for a region (harmless, since it is 805 // idempotent). 806 if (!region_ptr->blocks_filled()) { 807 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 808 region_ptr->set_blocks_filled(); 809 } 810 811 HeapWord* const search_start = block_align_down(addr); 812 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 813 814 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 815 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 816 result += block_offset + live; 817 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 818 return result; 819 } 820 821 #ifdef ASSERT 822 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 823 { 824 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 825 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 826 for (const size_t* p = beg; p < end; ++p) { 827 assert(*p == 0, "not zero"); 828 } 829 } 830 831 void ParallelCompactData::verify_clear() 832 { 833 verify_clear(_region_vspace); 834 verify_clear(_block_vspace); 835 } 836 #endif // #ifdef ASSERT 837 838 STWGCTimer PSParallelCompact::_gc_timer; 839 ParallelOldTracer PSParallelCompact::_gc_tracer; 840 elapsedTimer PSParallelCompact::_accumulated_time; 841 unsigned int PSParallelCompact::_total_invocations = 0; 842 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 843 jlong PSParallelCompact::_time_of_last_gc = 0; 844 CollectorCounters* PSParallelCompact::_counters = NULL; 845 ParMarkBitMap PSParallelCompact::_mark_bitmap; 846 ParallelCompactData PSParallelCompact::_summary_data; 847 848 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 849 850 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 851 852 class PCReferenceProcessor: public ReferenceProcessor { 853 public: 854 PCReferenceProcessor( 855 BoolObjectClosure* is_subject_to_discovery, 856 BoolObjectClosure* is_alive_non_header) : 857 ReferenceProcessor(is_subject_to_discovery, 858 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 859 ParallelGCThreads, // mt processing degree 860 true, // mt discovery 861 ParallelGCThreads, // mt discovery degree 862 true, // atomic_discovery 863 is_alive_non_header) { 864 } 865 866 template<typename T> bool discover(oop obj, ReferenceType type) { 867 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj); 868 T heap_oop = RawAccess<>::oop_load(referent_addr); 869 oop referent = CompressedOops::decode_not_null(heap_oop); 870 return PSParallelCompact::mark_bitmap()->is_unmarked(referent) 871 && ReferenceProcessor::discover_reference(obj, type); 872 } 873 virtual bool discover_reference(oop obj, ReferenceType type) { 874 if (UseCompressedOops) { 875 return discover<narrowOop>(obj, type); 876 } else { 877 return discover<oop>(obj, type); 878 } 879 } 880 }; 881 882 void PSParallelCompact::post_initialize() { 883 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 884 _span_based_discoverer.set_span(heap->reserved_region()); 885 _ref_processor = 886 new PCReferenceProcessor(&_span_based_discoverer, 887 &_is_alive_closure); // non-header is alive closure 888 889 _counters = new CollectorCounters("Parallel full collection pauses", 1); 890 891 // Initialize static fields in ParCompactionManager. 892 ParCompactionManager::initialize(mark_bitmap()); 893 } 894 895 bool PSParallelCompact::initialize() { 896 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 897 MemRegion mr = heap->reserved_region(); 898 899 // Was the old gen get allocated successfully? 900 if (!heap->old_gen()->is_allocated()) { 901 return false; 902 } 903 904 initialize_space_info(); 905 initialize_dead_wood_limiter(); 906 907 if (!_mark_bitmap.initialize(mr)) { 908 vm_shutdown_during_initialization( 909 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 910 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 911 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 912 return false; 913 } 914 915 if (!_summary_data.initialize(mr)) { 916 vm_shutdown_during_initialization( 917 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 918 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 919 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 920 return false; 921 } 922 923 return true; 924 } 925 926 void PSParallelCompact::initialize_space_info() 927 { 928 memset(&_space_info, 0, sizeof(_space_info)); 929 930 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 931 PSYoungGen* young_gen = heap->young_gen(); 932 933 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 934 _space_info[eden_space_id].set_space(young_gen->eden_space()); 935 _space_info[from_space_id].set_space(young_gen->from_space()); 936 _space_info[to_space_id].set_space(young_gen->to_space()); 937 938 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 939 } 940 941 void PSParallelCompact::initialize_dead_wood_limiter() 942 { 943 const size_t max = 100; 944 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 945 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 946 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 947 DEBUG_ONLY(_dwl_initialized = true;) 948 _dwl_adjustment = normal_distribution(1.0); 949 } 950 951 void 952 PSParallelCompact::clear_data_covering_space(SpaceId id) 953 { 954 // At this point, top is the value before GC, new_top() is the value that will 955 // be set at the end of GC. The marking bitmap is cleared to top; nothing 956 // should be marked above top. The summary data is cleared to the larger of 957 // top & new_top. 958 MutableSpace* const space = _space_info[id].space(); 959 HeapWord* const bot = space->bottom(); 960 HeapWord* const top = space->top(); 961 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 962 963 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 964 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 965 _mark_bitmap.clear_range(beg_bit, end_bit); 966 967 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 968 const size_t end_region = 969 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 970 _summary_data.clear_range(beg_region, end_region); 971 972 // Clear the data used to 'split' regions. 973 SplitInfo& split_info = _space_info[id].split_info(); 974 if (split_info.is_valid()) { 975 split_info.clear(); 976 } 977 DEBUG_ONLY(split_info.verify_clear();) 978 } 979 980 void PSParallelCompact::pre_compact() 981 { 982 // Update the from & to space pointers in space_info, since they are swapped 983 // at each young gen gc. Do the update unconditionally (even though a 984 // promotion failure does not swap spaces) because an unknown number of young 985 // collections will have swapped the spaces an unknown number of times. 986 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 987 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 988 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 989 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 990 991 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 992 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 993 994 // Increment the invocation count 995 heap->increment_total_collections(true); 996 997 // We need to track unique mark sweep invocations as well. 998 _total_invocations++; 999 1000 heap->print_heap_before_gc(); 1001 heap->trace_heap_before_gc(&_gc_tracer); 1002 1003 // Fill in TLABs 1004 heap->ensure_parsability(true); // retire TLABs 1005 1006 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1007 HandleMark hm; // Discard invalid handles created during verification 1008 Universe::verify("Before GC"); 1009 } 1010 1011 // Verify object start arrays 1012 if (VerifyObjectStartArray && 1013 VerifyBeforeGC) { 1014 heap->old_gen()->verify_object_start_array(); 1015 } 1016 1017 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1018 DEBUG_ONLY(summary_data().verify_clear();) 1019 1020 ParCompactionManager::reset_all_bitmap_query_caches(); 1021 } 1022 1023 void PSParallelCompact::post_compact() 1024 { 1025 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1026 1027 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1028 // Clear the marking bitmap, summary data and split info. 1029 clear_data_covering_space(SpaceId(id)); 1030 // Update top(). Must be done after clearing the bitmap and summary data. 1031 _space_info[id].publish_new_top(); 1032 } 1033 1034 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1035 MutableSpace* const from_space = _space_info[from_space_id].space(); 1036 MutableSpace* const to_space = _space_info[to_space_id].space(); 1037 1038 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1039 bool eden_empty = eden_space->is_empty(); 1040 if (!eden_empty) { 1041 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1042 heap->young_gen(), heap->old_gen()); 1043 } 1044 1045 // Update heap occupancy information which is used as input to the soft ref 1046 // clearing policy at the next gc. 1047 Universe::update_heap_info_at_gc(); 1048 1049 bool young_gen_empty = eden_empty && from_space->is_empty() && 1050 to_space->is_empty(); 1051 1052 PSCardTable* ct = heap->card_table(); 1053 MemRegion old_mr = heap->old_gen()->reserved(); 1054 if (young_gen_empty) { 1055 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1056 } else { 1057 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1058 } 1059 1060 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1061 ClassLoaderDataGraph::purge(); 1062 MetaspaceUtils::verify_metrics(); 1063 1064 heap->prune_scavengable_nmethods(); 1065 1066 #if COMPILER2_OR_JVMCI 1067 DerivedPointerTable::update_pointers(); 1068 #endif 1069 1070 if (ZapUnusedHeapArea) { 1071 heap->gen_mangle_unused_area(); 1072 } 1073 1074 // Update time of last GC 1075 reset_millis_since_last_gc(); 1076 } 1077 1078 HeapWord* 1079 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1080 bool maximum_compaction) 1081 { 1082 const size_t region_size = ParallelCompactData::RegionSize; 1083 const ParallelCompactData& sd = summary_data(); 1084 1085 const MutableSpace* const space = _space_info[id].space(); 1086 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1087 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1088 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1089 1090 // Skip full regions at the beginning of the space--they are necessarily part 1091 // of the dense prefix. 1092 size_t full_count = 0; 1093 const RegionData* cp; 1094 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1095 ++full_count; 1096 } 1097 1098 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1099 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1100 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1101 if (maximum_compaction || cp == end_cp || interval_ended) { 1102 _maximum_compaction_gc_num = total_invocations(); 1103 return sd.region_to_addr(cp); 1104 } 1105 1106 HeapWord* const new_top = _space_info[id].new_top(); 1107 const size_t space_live = pointer_delta(new_top, space->bottom()); 1108 const size_t space_used = space->used_in_words(); 1109 const size_t space_capacity = space->capacity_in_words(); 1110 1111 const double cur_density = double(space_live) / space_capacity; 1112 const double deadwood_density = 1113 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1114 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1115 1116 if (TraceParallelOldGCDensePrefix) { 1117 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1118 cur_density, deadwood_density, deadwood_goal); 1119 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1120 "space_cap=" SIZE_FORMAT, 1121 space_live, space_used, 1122 space_capacity); 1123 } 1124 1125 // XXX - Use binary search? 1126 HeapWord* dense_prefix = sd.region_to_addr(cp); 1127 const RegionData* full_cp = cp; 1128 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1129 while (cp < end_cp) { 1130 HeapWord* region_destination = cp->destination(); 1131 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1132 if (TraceParallelOldGCDensePrefix && Verbose) { 1133 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1134 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1135 sd.region(cp), p2i(region_destination), 1136 p2i(dense_prefix), cur_deadwood); 1137 } 1138 1139 if (cur_deadwood >= deadwood_goal) { 1140 // Found the region that has the correct amount of deadwood to the left. 1141 // This typically occurs after crossing a fairly sparse set of regions, so 1142 // iterate backwards over those sparse regions, looking for the region 1143 // that has the lowest density of live objects 'to the right.' 1144 size_t space_to_left = sd.region(cp) * region_size; 1145 size_t live_to_left = space_to_left - cur_deadwood; 1146 size_t space_to_right = space_capacity - space_to_left; 1147 size_t live_to_right = space_live - live_to_left; 1148 double density_to_right = double(live_to_right) / space_to_right; 1149 while (cp > full_cp) { 1150 --cp; 1151 const size_t prev_region_live_to_right = live_to_right - 1152 cp->data_size(); 1153 const size_t prev_region_space_to_right = space_to_right + region_size; 1154 double prev_region_density_to_right = 1155 double(prev_region_live_to_right) / prev_region_space_to_right; 1156 if (density_to_right <= prev_region_density_to_right) { 1157 return dense_prefix; 1158 } 1159 if (TraceParallelOldGCDensePrefix && Verbose) { 1160 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1161 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1162 prev_region_density_to_right); 1163 } 1164 dense_prefix -= region_size; 1165 live_to_right = prev_region_live_to_right; 1166 space_to_right = prev_region_space_to_right; 1167 density_to_right = prev_region_density_to_right; 1168 } 1169 return dense_prefix; 1170 } 1171 1172 dense_prefix += region_size; 1173 ++cp; 1174 } 1175 1176 return dense_prefix; 1177 } 1178 1179 #ifndef PRODUCT 1180 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1181 const SpaceId id, 1182 const bool maximum_compaction, 1183 HeapWord* const addr) 1184 { 1185 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1186 RegionData* const cp = summary_data().region(region_idx); 1187 const MutableSpace* const space = _space_info[id].space(); 1188 HeapWord* const new_top = _space_info[id].new_top(); 1189 1190 const size_t space_live = pointer_delta(new_top, space->bottom()); 1191 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1192 const size_t space_cap = space->capacity_in_words(); 1193 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1194 const size_t live_to_right = new_top - cp->destination(); 1195 const size_t dead_to_right = space->top() - addr - live_to_right; 1196 1197 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1198 "spl=" SIZE_FORMAT " " 1199 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1200 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1201 " ratio=%10.8f", 1202 algorithm, p2i(addr), region_idx, 1203 space_live, 1204 dead_to_left, dead_to_left_pct, 1205 dead_to_right, live_to_right, 1206 double(dead_to_right) / live_to_right); 1207 } 1208 #endif // #ifndef PRODUCT 1209 1210 // Return a fraction indicating how much of the generation can be treated as 1211 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1212 // based on the density of live objects in the generation to determine a limit, 1213 // which is then adjusted so the return value is min_percent when the density is 1214 // 1. 1215 // 1216 // The following table shows some return values for a different values of the 1217 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1218 // min_percent is 1. 1219 // 1220 // fraction allowed as dead wood 1221 // ----------------------------------------------------------------- 1222 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1223 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1224 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1225 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1226 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1227 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1228 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1229 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1230 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1231 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1232 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1233 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1234 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1235 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1236 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1237 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1238 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1239 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1240 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1241 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1242 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1243 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1244 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1245 1246 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1247 { 1248 assert(_dwl_initialized, "uninitialized"); 1249 1250 // The raw limit is the value of the normal distribution at x = density. 1251 const double raw_limit = normal_distribution(density); 1252 1253 // Adjust the raw limit so it becomes the minimum when the density is 1. 1254 // 1255 // First subtract the adjustment value (which is simply the precomputed value 1256 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1257 // Then add the minimum value, so the minimum is returned when the density is 1258 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1259 const double min = double(min_percent) / 100.0; 1260 const double limit = raw_limit - _dwl_adjustment + min; 1261 return MAX2(limit, 0.0); 1262 } 1263 1264 ParallelCompactData::RegionData* 1265 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1266 const RegionData* end) 1267 { 1268 const size_t region_size = ParallelCompactData::RegionSize; 1269 ParallelCompactData& sd = summary_data(); 1270 size_t left = sd.region(beg); 1271 size_t right = end > beg ? sd.region(end) - 1 : left; 1272 1273 // Binary search. 1274 while (left < right) { 1275 // Equivalent to (left + right) / 2, but does not overflow. 1276 const size_t middle = left + (right - left) / 2; 1277 RegionData* const middle_ptr = sd.region(middle); 1278 HeapWord* const dest = middle_ptr->destination(); 1279 HeapWord* const addr = sd.region_to_addr(middle); 1280 assert(dest != NULL, "sanity"); 1281 assert(dest <= addr, "must move left"); 1282 1283 if (middle > left && dest < addr) { 1284 right = middle - 1; 1285 } else if (middle < right && middle_ptr->data_size() == region_size) { 1286 left = middle + 1; 1287 } else { 1288 return middle_ptr; 1289 } 1290 } 1291 return sd.region(left); 1292 } 1293 1294 ParallelCompactData::RegionData* 1295 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1296 const RegionData* end, 1297 size_t dead_words) 1298 { 1299 ParallelCompactData& sd = summary_data(); 1300 size_t left = sd.region(beg); 1301 size_t right = end > beg ? sd.region(end) - 1 : left; 1302 1303 // Binary search. 1304 while (left < right) { 1305 // Equivalent to (left + right) / 2, but does not overflow. 1306 const size_t middle = left + (right - left) / 2; 1307 RegionData* const middle_ptr = sd.region(middle); 1308 HeapWord* const dest = middle_ptr->destination(); 1309 HeapWord* const addr = sd.region_to_addr(middle); 1310 assert(dest != NULL, "sanity"); 1311 assert(dest <= addr, "must move left"); 1312 1313 const size_t dead_to_left = pointer_delta(addr, dest); 1314 if (middle > left && dead_to_left > dead_words) { 1315 right = middle - 1; 1316 } else if (middle < right && dead_to_left < dead_words) { 1317 left = middle + 1; 1318 } else { 1319 return middle_ptr; 1320 } 1321 } 1322 return sd.region(left); 1323 } 1324 1325 // The result is valid during the summary phase, after the initial summarization 1326 // of each space into itself, and before final summarization. 1327 inline double 1328 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1329 HeapWord* const bottom, 1330 HeapWord* const top, 1331 HeapWord* const new_top) 1332 { 1333 ParallelCompactData& sd = summary_data(); 1334 1335 assert(cp != NULL, "sanity"); 1336 assert(bottom != NULL, "sanity"); 1337 assert(top != NULL, "sanity"); 1338 assert(new_top != NULL, "sanity"); 1339 assert(top >= new_top, "summary data problem?"); 1340 assert(new_top > bottom, "space is empty; should not be here"); 1341 assert(new_top >= cp->destination(), "sanity"); 1342 assert(top >= sd.region_to_addr(cp), "sanity"); 1343 1344 HeapWord* const destination = cp->destination(); 1345 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1346 const size_t compacted_region_live = pointer_delta(new_top, destination); 1347 const size_t compacted_region_used = pointer_delta(top, 1348 sd.region_to_addr(cp)); 1349 const size_t reclaimable = compacted_region_used - compacted_region_live; 1350 1351 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1352 return double(reclaimable) / divisor; 1353 } 1354 1355 // Return the address of the end of the dense prefix, a.k.a. the start of the 1356 // compacted region. The address is always on a region boundary. 1357 // 1358 // Completely full regions at the left are skipped, since no compaction can 1359 // occur in those regions. Then the maximum amount of dead wood to allow is 1360 // computed, based on the density (amount live / capacity) of the generation; 1361 // the region with approximately that amount of dead space to the left is 1362 // identified as the limit region. Regions between the last completely full 1363 // region and the limit region are scanned and the one that has the best 1364 // (maximum) reclaimed_ratio() is selected. 1365 HeapWord* 1366 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1367 bool maximum_compaction) 1368 { 1369 const size_t region_size = ParallelCompactData::RegionSize; 1370 const ParallelCompactData& sd = summary_data(); 1371 1372 const MutableSpace* const space = _space_info[id].space(); 1373 HeapWord* const top = space->top(); 1374 HeapWord* const top_aligned_up = sd.region_align_up(top); 1375 HeapWord* const new_top = _space_info[id].new_top(); 1376 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1377 HeapWord* const bottom = space->bottom(); 1378 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1379 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1380 const RegionData* const new_top_cp = 1381 sd.addr_to_region_ptr(new_top_aligned_up); 1382 1383 // Skip full regions at the beginning of the space--they are necessarily part 1384 // of the dense prefix. 1385 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1386 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1387 space->is_empty(), "no dead space allowed to the left"); 1388 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1389 "region must have dead space"); 1390 1391 // The gc number is saved whenever a maximum compaction is done, and used to 1392 // determine when the maximum compaction interval has expired. This avoids 1393 // successive max compactions for different reasons. 1394 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1395 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1396 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1397 total_invocations() == HeapFirstMaximumCompactionCount; 1398 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1399 _maximum_compaction_gc_num = total_invocations(); 1400 return sd.region_to_addr(full_cp); 1401 } 1402 1403 const size_t space_live = pointer_delta(new_top, bottom); 1404 const size_t space_used = space->used_in_words(); 1405 const size_t space_capacity = space->capacity_in_words(); 1406 1407 const double density = double(space_live) / double(space_capacity); 1408 const size_t min_percent_free = MarkSweepDeadRatio; 1409 const double limiter = dead_wood_limiter(density, min_percent_free); 1410 const size_t dead_wood_max = space_used - space_live; 1411 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1412 dead_wood_max); 1413 1414 if (TraceParallelOldGCDensePrefix) { 1415 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1416 "space_cap=" SIZE_FORMAT, 1417 space_live, space_used, 1418 space_capacity); 1419 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1420 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1421 density, min_percent_free, limiter, 1422 dead_wood_max, dead_wood_limit); 1423 } 1424 1425 // Locate the region with the desired amount of dead space to the left. 1426 const RegionData* const limit_cp = 1427 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1428 1429 // Scan from the first region with dead space to the limit region and find the 1430 // one with the best (largest) reclaimed ratio. 1431 double best_ratio = 0.0; 1432 const RegionData* best_cp = full_cp; 1433 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1434 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1435 if (tmp_ratio > best_ratio) { 1436 best_cp = cp; 1437 best_ratio = tmp_ratio; 1438 } 1439 } 1440 1441 return sd.region_to_addr(best_cp); 1442 } 1443 1444 void PSParallelCompact::summarize_spaces_quick() 1445 { 1446 for (unsigned int i = 0; i < last_space_id; ++i) { 1447 const MutableSpace* space = _space_info[i].space(); 1448 HeapWord** nta = _space_info[i].new_top_addr(); 1449 bool result = _summary_data.summarize(_space_info[i].split_info(), 1450 space->bottom(), space->top(), NULL, 1451 space->bottom(), space->end(), nta); 1452 assert(result, "space must fit into itself"); 1453 _space_info[i].set_dense_prefix(space->bottom()); 1454 } 1455 } 1456 1457 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1458 { 1459 HeapWord* const dense_prefix_end = dense_prefix(id); 1460 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1461 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1462 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1463 // Only enough dead space is filled so that any remaining dead space to the 1464 // left is larger than the minimum filler object. (The remainder is filled 1465 // during the copy/update phase.) 1466 // 1467 // The size of the dead space to the right of the boundary is not a 1468 // concern, since compaction will be able to use whatever space is 1469 // available. 1470 // 1471 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1472 // surrounds the space to be filled with an object. 1473 // 1474 // In the 32-bit VM, each bit represents two 32-bit words: 1475 // +---+ 1476 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1477 // end_bits: ... x x x | 0 | || 0 x x ... 1478 // +---+ 1479 // 1480 // In the 64-bit VM, each bit represents one 64-bit word: 1481 // +------------+ 1482 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1483 // end_bits: ... x x 1 | 0 || 0 | x x ... 1484 // +------------+ 1485 // +-------+ 1486 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1487 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1488 // +-------+ 1489 // +-----------+ 1490 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1491 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1492 // +-----------+ 1493 // +-------+ 1494 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1495 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1496 // +-------+ 1497 1498 // Initially assume case a, c or e will apply. 1499 size_t obj_len = CollectedHeap::min_fill_size(); 1500 HeapWord* obj_beg = dense_prefix_end - obj_len; 1501 1502 #ifdef _LP64 1503 if (MinObjAlignment > 1) { // object alignment > heap word size 1504 // Cases a, c or e. 1505 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1506 // Case b above. 1507 obj_beg = dense_prefix_end - 1; 1508 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1509 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1510 // Case d above. 1511 obj_beg = dense_prefix_end - 3; 1512 obj_len = 3; 1513 } 1514 #endif // #ifdef _LP64 1515 1516 CollectedHeap::fill_with_object(obj_beg, obj_len); 1517 _mark_bitmap.mark_obj(obj_beg, obj_len); 1518 _summary_data.add_obj(obj_beg, obj_len); 1519 assert(start_array(id) != NULL, "sanity"); 1520 start_array(id)->allocate_block(obj_beg); 1521 } 1522 } 1523 1524 void 1525 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1526 { 1527 assert(id < last_space_id, "id out of range"); 1528 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1529 "should have been reset in summarize_spaces_quick()"); 1530 1531 const MutableSpace* space = _space_info[id].space(); 1532 if (_space_info[id].new_top() != space->bottom()) { 1533 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1534 _space_info[id].set_dense_prefix(dense_prefix_end); 1535 1536 #ifndef PRODUCT 1537 if (TraceParallelOldGCDensePrefix) { 1538 print_dense_prefix_stats("ratio", id, maximum_compaction, 1539 dense_prefix_end); 1540 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1541 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1542 } 1543 #endif // #ifndef PRODUCT 1544 1545 // Recompute the summary data, taking into account the dense prefix. If 1546 // every last byte will be reclaimed, then the existing summary data which 1547 // compacts everything can be left in place. 1548 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1549 // If dead space crosses the dense prefix boundary, it is (at least 1550 // partially) filled with a dummy object, marked live and added to the 1551 // summary data. This simplifies the copy/update phase and must be done 1552 // before the final locations of objects are determined, to prevent 1553 // leaving a fragment of dead space that is too small to fill. 1554 fill_dense_prefix_end(id); 1555 1556 // Compute the destination of each Region, and thus each object. 1557 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1558 _summary_data.summarize(_space_info[id].split_info(), 1559 dense_prefix_end, space->top(), NULL, 1560 dense_prefix_end, space->end(), 1561 _space_info[id].new_top_addr()); 1562 } 1563 } 1564 1565 if (log_develop_is_enabled(Trace, gc, compaction)) { 1566 const size_t region_size = ParallelCompactData::RegionSize; 1567 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1568 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1569 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1570 HeapWord* const new_top = _space_info[id].new_top(); 1571 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1572 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1573 log_develop_trace(gc, compaction)( 1574 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1575 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1576 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1577 id, space->capacity_in_words(), p2i(dense_prefix_end), 1578 dp_region, dp_words / region_size, 1579 cr_words / region_size, p2i(new_top)); 1580 } 1581 } 1582 1583 #ifndef PRODUCT 1584 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1585 HeapWord* dst_beg, HeapWord* dst_end, 1586 SpaceId src_space_id, 1587 HeapWord* src_beg, HeapWord* src_end) 1588 { 1589 log_develop_trace(gc, compaction)( 1590 "Summarizing %d [%s] into %d [%s]: " 1591 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1592 SIZE_FORMAT "-" SIZE_FORMAT " " 1593 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1594 SIZE_FORMAT "-" SIZE_FORMAT, 1595 src_space_id, space_names[src_space_id], 1596 dst_space_id, space_names[dst_space_id], 1597 p2i(src_beg), p2i(src_end), 1598 _summary_data.addr_to_region_idx(src_beg), 1599 _summary_data.addr_to_region_idx(src_end), 1600 p2i(dst_beg), p2i(dst_end), 1601 _summary_data.addr_to_region_idx(dst_beg), 1602 _summary_data.addr_to_region_idx(dst_end)); 1603 } 1604 #endif // #ifndef PRODUCT 1605 1606 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1607 bool maximum_compaction) 1608 { 1609 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1610 1611 #ifdef ASSERT 1612 if (TraceParallelOldGCMarkingPhase) { 1613 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1614 "add_obj_bytes=" SIZE_FORMAT, 1615 add_obj_count, add_obj_size * HeapWordSize); 1616 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1617 "mark_bitmap_bytes=" SIZE_FORMAT, 1618 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1619 } 1620 #endif // #ifdef ASSERT 1621 1622 // Quick summarization of each space into itself, to see how much is live. 1623 summarize_spaces_quick(); 1624 1625 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1626 NOT_PRODUCT(print_region_ranges()); 1627 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1628 1629 // The amount of live data that will end up in old space (assuming it fits). 1630 size_t old_space_total_live = 0; 1631 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1632 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1633 _space_info[id].space()->bottom()); 1634 } 1635 1636 MutableSpace* const old_space = _space_info[old_space_id].space(); 1637 const size_t old_capacity = old_space->capacity_in_words(); 1638 if (old_space_total_live > old_capacity) { 1639 // XXX - should also try to expand 1640 maximum_compaction = true; 1641 } 1642 1643 // Old generations. 1644 summarize_space(old_space_id, maximum_compaction); 1645 1646 // Summarize the remaining spaces in the young gen. The initial target space 1647 // is the old gen. If a space does not fit entirely into the target, then the 1648 // remainder is compacted into the space itself and that space becomes the new 1649 // target. 1650 SpaceId dst_space_id = old_space_id; 1651 HeapWord* dst_space_end = old_space->end(); 1652 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1653 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1654 const MutableSpace* space = _space_info[id].space(); 1655 const size_t live = pointer_delta(_space_info[id].new_top(), 1656 space->bottom()); 1657 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1658 1659 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1660 SpaceId(id), space->bottom(), space->top());) 1661 if (live > 0 && live <= available) { 1662 // All the live data will fit. 1663 bool done = _summary_data.summarize(_space_info[id].split_info(), 1664 space->bottom(), space->top(), 1665 NULL, 1666 *new_top_addr, dst_space_end, 1667 new_top_addr); 1668 assert(done, "space must fit into old gen"); 1669 1670 // Reset the new_top value for the space. 1671 _space_info[id].set_new_top(space->bottom()); 1672 } else if (live > 0) { 1673 // Attempt to fit part of the source space into the target space. 1674 HeapWord* next_src_addr = NULL; 1675 bool done = _summary_data.summarize(_space_info[id].split_info(), 1676 space->bottom(), space->top(), 1677 &next_src_addr, 1678 *new_top_addr, dst_space_end, 1679 new_top_addr); 1680 assert(!done, "space should not fit into old gen"); 1681 assert(next_src_addr != NULL, "sanity"); 1682 1683 // The source space becomes the new target, so the remainder is compacted 1684 // within the space itself. 1685 dst_space_id = SpaceId(id); 1686 dst_space_end = space->end(); 1687 new_top_addr = _space_info[id].new_top_addr(); 1688 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1689 space->bottom(), dst_space_end, 1690 SpaceId(id), next_src_addr, space->top());) 1691 done = _summary_data.summarize(_space_info[id].split_info(), 1692 next_src_addr, space->top(), 1693 NULL, 1694 space->bottom(), dst_space_end, 1695 new_top_addr); 1696 assert(done, "space must fit when compacted into itself"); 1697 assert(*new_top_addr <= space->top(), "usage should not grow"); 1698 } 1699 } 1700 1701 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1702 NOT_PRODUCT(print_region_ranges()); 1703 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1704 } 1705 1706 // This method should contain all heap-specific policy for invoking a full 1707 // collection. invoke_no_policy() will only attempt to compact the heap; it 1708 // will do nothing further. If we need to bail out for policy reasons, scavenge 1709 // before full gc, or any other specialized behavior, it needs to be added here. 1710 // 1711 // Note that this method should only be called from the vm_thread while at a 1712 // safepoint. 1713 // 1714 // Note that the all_soft_refs_clear flag in the soft ref policy 1715 // may be true because this method can be called without intervening 1716 // activity. For example when the heap space is tight and full measure 1717 // are being taken to free space. 1718 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1719 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1720 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1721 "should be in vm thread"); 1722 1723 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1724 GCCause::Cause gc_cause = heap->gc_cause(); 1725 assert(!heap->is_gc_active(), "not reentrant"); 1726 1727 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1728 IsGCActiveMark mark; 1729 1730 if (ScavengeBeforeFullGC) { 1731 PSScavenge::invoke_no_policy(); 1732 } 1733 1734 const bool clear_all_soft_refs = 1735 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1736 1737 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1738 maximum_heap_compaction); 1739 } 1740 1741 // This method contains no policy. You should probably 1742 // be calling invoke() instead. 1743 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1744 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1745 assert(ref_processor() != NULL, "Sanity"); 1746 1747 if (GCLocker::check_active_before_gc()) { 1748 return false; 1749 } 1750 1751 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1752 1753 GCIdMark gc_id_mark; 1754 _gc_timer.register_gc_start(); 1755 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1756 1757 TimeStamp marking_start; 1758 TimeStamp compaction_start; 1759 TimeStamp collection_exit; 1760 1761 GCCause::Cause gc_cause = heap->gc_cause(); 1762 PSYoungGen* young_gen = heap->young_gen(); 1763 PSOldGen* old_gen = heap->old_gen(); 1764 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1765 1766 // The scope of casr should end after code that can change 1767 // SoftRefPolicy::_should_clear_all_soft_refs. 1768 ClearedAllSoftRefs casr(maximum_heap_compaction, 1769 heap->soft_ref_policy()); 1770 1771 if (ZapUnusedHeapArea) { 1772 // Save information needed to minimize mangling 1773 heap->record_gen_tops_before_GC(); 1774 } 1775 1776 // Make sure data structures are sane, make the heap parsable, and do other 1777 // miscellaneous bookkeeping. 1778 pre_compact(); 1779 1780 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values(); 1781 1782 // Get the compaction manager reserved for the VM thread. 1783 ParCompactionManager* const vmthread_cm = 1784 ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers()); 1785 1786 { 1787 ResourceMark rm; 1788 HandleMark hm; 1789 1790 const uint active_workers = 1791 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(), 1792 ParallelScavengeHeap::heap()->workers().active_workers(), 1793 Threads::number_of_non_daemon_threads()); 1794 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers); 1795 1796 GCTraceCPUTime tcpu; 1797 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1798 1799 heap->pre_full_gc_dump(&_gc_timer); 1800 1801 TraceCollectorStats tcs(counters()); 1802 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1803 1804 if (log_is_enabled(Debug, gc, heap, exit)) { 1805 accumulated_time()->start(); 1806 } 1807 1808 // Let the size policy know we're starting 1809 size_policy->major_collection_begin(); 1810 1811 #if COMPILER2_OR_JVMCI 1812 DerivedPointerTable::clear(); 1813 #endif 1814 1815 ref_processor()->enable_discovery(); 1816 ref_processor()->setup_policy(maximum_heap_compaction); 1817 1818 bool marked_for_unloading = false; 1819 1820 marking_start.update(); 1821 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1822 1823 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1824 && GCCause::is_user_requested_gc(gc_cause); 1825 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1826 1827 #if COMPILER2_OR_JVMCI 1828 assert(DerivedPointerTable::is_active(), "Sanity"); 1829 DerivedPointerTable::set_active(false); 1830 #endif 1831 1832 // adjust_roots() updates Universe::_intArrayKlassObj which is 1833 // needed by the compaction for filling holes in the dense prefix. 1834 adjust_roots(vmthread_cm); 1835 1836 compaction_start.update(); 1837 compact(); 1838 1839 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1840 // done before resizing. 1841 post_compact(); 1842 1843 // Let the size policy know we're done 1844 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1845 1846 if (UseAdaptiveSizePolicy) { 1847 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1848 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1849 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1850 1851 // Don't check if the size_policy is ready here. Let 1852 // the size_policy check that internally. 1853 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1854 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1855 // Swap the survivor spaces if from_space is empty. The 1856 // resize_young_gen() called below is normally used after 1857 // a successful young GC and swapping of survivor spaces; 1858 // otherwise, it will fail to resize the young gen with 1859 // the current implementation. 1860 if (young_gen->from_space()->is_empty()) { 1861 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1862 young_gen->swap_spaces(); 1863 } 1864 1865 // Calculate optimal free space amounts 1866 assert(young_gen->max_size() > 1867 young_gen->from_space()->capacity_in_bytes() + 1868 young_gen->to_space()->capacity_in_bytes(), 1869 "Sizes of space in young gen are out-of-bounds"); 1870 1871 size_t young_live = young_gen->used_in_bytes(); 1872 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1873 size_t old_live = old_gen->used_in_bytes(); 1874 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1875 size_t max_old_gen_size = old_gen->max_gen_size(); 1876 size_t max_eden_size = young_gen->max_size() - 1877 young_gen->from_space()->capacity_in_bytes() - 1878 young_gen->to_space()->capacity_in_bytes(); 1879 1880 // Used for diagnostics 1881 size_policy->clear_generation_free_space_flags(); 1882 1883 size_policy->compute_generations_free_space(young_live, 1884 eden_live, 1885 old_live, 1886 cur_eden, 1887 max_old_gen_size, 1888 max_eden_size, 1889 true /* full gc*/); 1890 1891 size_policy->check_gc_overhead_limit(eden_live, 1892 max_old_gen_size, 1893 max_eden_size, 1894 true /* full gc*/, 1895 gc_cause, 1896 heap->soft_ref_policy()); 1897 1898 size_policy->decay_supplemental_growth(true /* full gc*/); 1899 1900 heap->resize_old_gen( 1901 size_policy->calculated_old_free_size_in_bytes()); 1902 1903 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1904 size_policy->calculated_survivor_size_in_bytes()); 1905 } 1906 1907 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1908 } 1909 1910 if (UsePerfData) { 1911 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1912 counters->update_counters(); 1913 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1914 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1915 } 1916 1917 heap->resize_all_tlabs(); 1918 1919 // Resize the metaspace capacity after a collection 1920 MetaspaceGC::compute_new_size(); 1921 1922 if (log_is_enabled(Debug, gc, heap, exit)) { 1923 accumulated_time()->stop(); 1924 } 1925 1926 heap->print_heap_change(pre_gc_values); 1927 1928 // Track memory usage and detect low memory 1929 MemoryService::track_memory_usage(); 1930 heap->update_counters(); 1931 1932 heap->post_full_gc_dump(&_gc_timer); 1933 } 1934 1935 #ifdef ASSERT 1936 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1937 ParCompactionManager* const cm = 1938 ParCompactionManager::manager_array(int(i)); 1939 assert(cm->marking_stack()->is_empty(), "should be empty"); 1940 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1941 } 1942 #endif // ASSERT 1943 1944 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1945 HandleMark hm; // Discard invalid handles created during verification 1946 Universe::verify("After GC"); 1947 } 1948 1949 // Re-verify object start arrays 1950 if (VerifyObjectStartArray && 1951 VerifyAfterGC) { 1952 old_gen->verify_object_start_array(); 1953 } 1954 1955 if (ZapUnusedHeapArea) { 1956 old_gen->object_space()->check_mangled_unused_area_complete(); 1957 } 1958 1959 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1960 1961 collection_exit.update(); 1962 1963 heap->print_heap_after_gc(); 1964 heap->trace_heap_after_gc(&_gc_tracer); 1965 1966 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1967 marking_start.ticks(), compaction_start.ticks(), 1968 collection_exit.ticks()); 1969 1970 #ifdef TRACESPINNING 1971 ParallelTaskTerminator::print_termination_counts(); 1972 #endif 1973 1974 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1975 1976 _gc_timer.register_gc_end(); 1977 1978 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1979 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1980 1981 return true; 1982 } 1983 1984 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1985 PSYoungGen* young_gen, 1986 PSOldGen* old_gen) { 1987 MutableSpace* const eden_space = young_gen->eden_space(); 1988 assert(!eden_space->is_empty(), "eden must be non-empty"); 1989 assert(young_gen->virtual_space()->alignment() == 1990 old_gen->virtual_space()->alignment(), "alignments do not match"); 1991 1992 // We also return false when it's a heterogeneous heap because old generation cannot absorb data from eden 1993 // when it is allocated on different memory (example, nv-dimm) than young. 1994 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) || 1995 ParallelArguments::is_heterogeneous_heap()) { 1996 return false; 1997 } 1998 1999 // Both generations must be completely committed. 2000 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2001 return false; 2002 } 2003 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2004 return false; 2005 } 2006 2007 // Figure out how much to take from eden. Include the average amount promoted 2008 // in the total; otherwise the next young gen GC will simply bail out to a 2009 // full GC. 2010 const size_t alignment = old_gen->virtual_space()->alignment(); 2011 const size_t eden_used = eden_space->used_in_bytes(); 2012 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2013 const size_t absorb_size = align_up(eden_used + promoted, alignment); 2014 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2015 2016 if (absorb_size >= eden_capacity) { 2017 return false; // Must leave some space in eden. 2018 } 2019 2020 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2021 if (new_young_size < young_gen->min_gen_size()) { 2022 return false; // Respect young gen minimum size. 2023 } 2024 2025 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: " 2026 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2027 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2028 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2029 absorb_size / K, 2030 eden_capacity / K, (eden_capacity - absorb_size) / K, 2031 young_gen->from_space()->used_in_bytes() / K, 2032 young_gen->to_space()->used_in_bytes() / K, 2033 young_gen->capacity_in_bytes() / K, new_young_size / K); 2034 2035 // Fill the unused part of the old gen. 2036 MutableSpace* const old_space = old_gen->object_space(); 2037 HeapWord* const unused_start = old_space->top(); 2038 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2039 2040 if (unused_words > 0) { 2041 if (unused_words < CollectedHeap::min_fill_size()) { 2042 return false; // If the old gen cannot be filled, must give up. 2043 } 2044 CollectedHeap::fill_with_objects(unused_start, unused_words); 2045 } 2046 2047 // Take the live data from eden and set both top and end in the old gen to 2048 // eden top. (Need to set end because reset_after_change() mangles the region 2049 // from end to virtual_space->high() in debug builds). 2050 HeapWord* const new_top = eden_space->top(); 2051 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2052 absorb_size); 2053 young_gen->reset_after_change(); 2054 old_space->set_top(new_top); 2055 old_space->set_end(new_top); 2056 old_gen->reset_after_change(); 2057 2058 // Update the object start array for the filler object and the data from eden. 2059 ObjectStartArray* const start_array = old_gen->start_array(); 2060 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2061 start_array->allocate_block(p); 2062 } 2063 2064 // Could update the promoted average here, but it is not typically updated at 2065 // full GCs and the value to use is unclear. Something like 2066 // 2067 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2068 2069 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2070 return true; 2071 } 2072 2073 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 2074 private: 2075 uint _worker_id; 2076 2077 public: 2078 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { } 2079 void do_thread(Thread* thread) { 2080 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2081 2082 ResourceMark rm; 2083 2084 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id); 2085 2086 PCMarkAndPushClosure mark_and_push_closure(cm); 2087 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations); 2088 2089 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs); 2090 2091 // Do the real work 2092 cm->follow_marking_stacks(); 2093 } 2094 }; 2095 2096 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) { 2097 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2098 2099 ParCompactionManager* cm = 2100 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2101 PCMarkAndPushClosure mark_and_push_closure(cm); 2102 2103 switch (root_type) { 2104 case ParallelRootType::universe: 2105 Universe::oops_do(&mark_and_push_closure); 2106 break; 2107 2108 case ParallelRootType::jni_handles: 2109 JNIHandles::oops_do(&mark_and_push_closure); 2110 break; 2111 2112 case ParallelRootType::object_synchronizer: 2113 ObjectSynchronizer::oops_do(&mark_and_push_closure); 2114 break; 2115 2116 case ParallelRootType::management: 2117 Management::oops_do(&mark_and_push_closure); 2118 break; 2119 2120 case ParallelRootType::jvmti: 2121 JvmtiExport::oops_do(&mark_and_push_closure); 2122 break; 2123 2124 case ParallelRootType::system_dictionary: 2125 SystemDictionary::oops_do(&mark_and_push_closure); 2126 break; 2127 2128 case ParallelRootType::class_loader_data: 2129 { 2130 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong); 2131 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure); 2132 } 2133 break; 2134 2135 case ParallelRootType::code_cache: 2136 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them. 2137 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure)); 2138 AOTLoader::oops_do(&mark_and_push_closure); 2139 break; 2140 2141 case ParallelRootType::sentinel: 2142 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds 2143 fatal("Bad enumeration value: %u", root_type); 2144 break; 2145 } 2146 2147 // Do the real work 2148 cm->follow_marking_stacks(); 2149 } 2150 2151 static void steal_marking_work(ParallelTaskTerminator& terminator, uint worker_id) { 2152 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2153 2154 ParCompactionManager* cm = 2155 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2156 2157 oop obj = NULL; 2158 ObjArrayTask task; 2159 do { 2160 while (ParCompactionManager::steal_objarray(worker_id, task)) { 2161 cm->follow_array((objArrayOop)task.obj(), task.index()); 2162 cm->follow_marking_stacks(); 2163 } 2164 while (ParCompactionManager::steal(worker_id, obj)) { 2165 cm->follow_contents(obj); 2166 cm->follow_marking_stacks(); 2167 } 2168 } while (!terminator.offer_termination()); 2169 } 2170 2171 class MarkFromRootsTask : public AbstractGangTask { 2172 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2173 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do 2174 SequentialSubTasksDone _subtasks; 2175 TaskTerminator _terminator; 2176 uint _active_workers; 2177 2178 public: 2179 MarkFromRootsTask(uint active_workers) : 2180 AbstractGangTask("MarkFromRootsTask"), 2181 _strong_roots_scope(active_workers), 2182 _subtasks(), 2183 _terminator(active_workers, ParCompactionManager::stack_array()), 2184 _active_workers(active_workers) { 2185 _subtasks.set_n_threads(active_workers); 2186 _subtasks.set_n_tasks(ParallelRootType::sentinel); 2187 } 2188 2189 virtual void work(uint worker_id) { 2190 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) { 2191 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id); 2192 } 2193 _subtasks.all_tasks_completed(); 2194 2195 PCAddThreadRootsMarkingTaskClosure closure(worker_id); 2196 Threads::possibly_parallel_threads_do(true /*parallel */, &closure); 2197 2198 if (_active_workers > 1) { 2199 steal_marking_work(*_terminator.terminator(), worker_id); 2200 } 2201 } 2202 }; 2203 2204 class PCRefProcTask : public AbstractGangTask { 2205 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2206 ProcessTask& _task; 2207 uint _ergo_workers; 2208 TaskTerminator _terminator; 2209 2210 public: 2211 PCRefProcTask(ProcessTask& task, uint ergo_workers) : 2212 AbstractGangTask("PCRefProcTask"), 2213 _task(task), 2214 _ergo_workers(ergo_workers), 2215 _terminator(_ergo_workers, ParCompactionManager::stack_array()) { 2216 } 2217 2218 virtual void work(uint worker_id) { 2219 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2220 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2221 2222 ParCompactionManager* cm = 2223 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2224 PCMarkAndPushClosure mark_and_push_closure(cm); 2225 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2226 _task.work(worker_id, *PSParallelCompact::is_alive_closure(), 2227 mark_and_push_closure, follow_stack_closure); 2228 2229 steal_marking_work(*_terminator.terminator(), worker_id); 2230 } 2231 }; 2232 2233 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2234 void execute(ProcessTask& process_task, uint ergo_workers) { 2235 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers, 2236 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 2237 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers()); 2238 2239 PCRefProcTask task(process_task, ergo_workers); 2240 ParallelScavengeHeap::heap()->workers().run_task(&task); 2241 } 2242 }; 2243 2244 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2245 bool maximum_heap_compaction, 2246 ParallelOldTracer *gc_tracer) { 2247 // Recursively traverse all live objects and mark them 2248 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2249 2250 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2251 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2252 2253 PCMarkAndPushClosure mark_and_push_closure(cm); 2254 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2255 2256 // Need new claim bits before marking starts. 2257 ClassLoaderDataGraph::clear_claimed_marks(); 2258 2259 { 2260 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2261 2262 MarkFromRootsTask task(active_gc_threads); 2263 ParallelScavengeHeap::heap()->workers().run_task(&task); 2264 } 2265 2266 // Process reference objects found during marking 2267 { 2268 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2269 2270 ReferenceProcessorStats stats; 2271 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2272 2273 if (ref_processor()->processing_is_mt()) { 2274 ref_processor()->set_active_mt_degree(active_gc_threads); 2275 2276 RefProcTaskExecutor task_executor; 2277 stats = ref_processor()->process_discovered_references( 2278 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2279 &task_executor, &pt); 2280 } else { 2281 stats = ref_processor()->process_discovered_references( 2282 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2283 &pt); 2284 } 2285 2286 gc_tracer->report_gc_reference_stats(stats); 2287 pt.print_all_references(); 2288 } 2289 2290 // This is the point where the entire marking should have completed. 2291 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2292 2293 { 2294 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2295 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2296 } 2297 2298 { 2299 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2300 2301 // Follow system dictionary roots and unload classes. 2302 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2303 2304 // Unload nmethods. 2305 CodeCache::do_unloading(is_alive_closure(), purged_class); 2306 2307 // Prune dead klasses from subklass/sibling/implementor lists. 2308 Klass::clean_weak_klass_links(purged_class); 2309 2310 // Clean JVMCI metadata handles. 2311 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 2312 } 2313 2314 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2315 } 2316 2317 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2318 // Adjust the pointers to reflect the new locations 2319 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2320 2321 // Need new claim bits when tracing through and adjusting pointers. 2322 ClassLoaderDataGraph::clear_claimed_marks(); 2323 2324 PCAdjustPointerClosure oop_closure(cm); 2325 2326 // General strong roots. 2327 Universe::oops_do(&oop_closure); 2328 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2329 Threads::oops_do(&oop_closure, NULL); 2330 ObjectSynchronizer::oops_do(&oop_closure); 2331 Management::oops_do(&oop_closure); 2332 JvmtiExport::oops_do(&oop_closure); 2333 SystemDictionary::oops_do(&oop_closure); 2334 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong); 2335 ClassLoaderDataGraph::cld_do(&cld_closure); 2336 2337 // Now adjust pointers in remaining weak roots. (All of which should 2338 // have been cleared if they pointed to non-surviving objects.) 2339 WeakProcessor::oops_do(&oop_closure); 2340 2341 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2342 CodeCache::blobs_do(&adjust_from_blobs); 2343 AOT_ONLY(AOTLoader::oops_do(&oop_closure);) 2344 2345 ref_processor()->weak_oops_do(&oop_closure); 2346 // Roots were visited so references into the young gen in roots 2347 // may have been scanned. Process them also. 2348 // Should the reference processor have a span that excludes 2349 // young gen objects? 2350 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2351 } 2352 2353 // Helper class to print 8 region numbers per line and then print the total at the end. 2354 class FillableRegionLogger : public StackObj { 2355 private: 2356 Log(gc, compaction) log; 2357 static const int LineLength = 8; 2358 size_t _regions[LineLength]; 2359 int _next_index; 2360 bool _enabled; 2361 size_t _total_regions; 2362 public: 2363 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2364 ~FillableRegionLogger() { 2365 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2366 } 2367 2368 void print_line() { 2369 if (!_enabled || _next_index == 0) { 2370 return; 2371 } 2372 FormatBuffer<> line("Fillable: "); 2373 for (int i = 0; i < _next_index; i++) { 2374 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2375 } 2376 log.trace("%s", line.buffer()); 2377 _next_index = 0; 2378 } 2379 2380 void handle(size_t region) { 2381 if (!_enabled) { 2382 return; 2383 } 2384 _regions[_next_index++] = region; 2385 if (_next_index == LineLength) { 2386 print_line(); 2387 } 2388 _total_regions++; 2389 } 2390 }; 2391 2392 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads) 2393 { 2394 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2395 2396 // Find the threads that are active 2397 uint worker_id = 0; 2398 2399 // Find all regions that are available (can be filled immediately) and 2400 // distribute them to the thread stacks. The iteration is done in reverse 2401 // order (high to low) so the regions will be removed in ascending order. 2402 2403 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2404 2405 // id + 1 is used to test termination so unsigned can 2406 // be used with an old_space_id == 0. 2407 FillableRegionLogger region_logger; 2408 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2409 SpaceInfo* const space_info = _space_info + id; 2410 MutableSpace* const space = space_info->space(); 2411 HeapWord* const new_top = space_info->new_top(); 2412 2413 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2414 const size_t end_region = 2415 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2416 2417 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2418 if (sd.region(cur)->claim_unsafe()) { 2419 ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id); 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::add(1u, &_counter) - 1; // -1 is so that we start with zero 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(ParallelTaskTerminator* 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 { 2606 if (terminator->offer_termination()) { 2607 break; 2608 } 2609 // Go around again. 2610 } 2611 } 2612 return; 2613 } 2614 2615 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask { 2616 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2617 TaskQueue& _tq; 2618 TaskTerminator _terminator; 2619 uint _active_workers; 2620 2621 public: 2622 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) : 2623 AbstractGangTask("UpdateDensePrefixAndCompactionTask"), 2624 _tq(tq), 2625 _terminator(active_workers, ParCompactionManager::region_array()), 2626 _active_workers(active_workers) { 2627 } 2628 virtual void work(uint worker_id) { 2629 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2630 2631 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) { 2632 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm, 2633 task._space_id, 2634 task._region_index_start, 2635 task._region_index_end); 2636 } 2637 2638 // Once a thread has drained it's stack, it should try to steal regions from 2639 // other threads. 2640 compaction_with_stealing_work(_terminator.terminator(), worker_id); 2641 } 2642 }; 2643 2644 void PSParallelCompact::compact() { 2645 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2646 2647 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2648 PSOldGen* old_gen = heap->old_gen(); 2649 old_gen->start_array()->reset(); 2650 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2651 2652 // for [0..last_space_id) 2653 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING) 2654 // push 2655 // push 2656 // 2657 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1) 2658 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)); 2659 prepare_region_draining_tasks(active_gc_threads); 2660 enqueue_dense_prefix_tasks(task_queue, active_gc_threads); 2661 2662 { 2663 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2664 2665 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads); 2666 ParallelScavengeHeap::heap()->workers().run_task(&task); 2667 2668 #ifdef ASSERT 2669 // Verify that all regions have been processed before the deferred updates. 2670 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2671 verify_complete(SpaceId(id)); 2672 } 2673 #endif 2674 } 2675 2676 { 2677 // Update the deferred objects, if any. Any compaction manager can be used. 2678 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2679 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2680 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2681 update_deferred_objects(cm, SpaceId(id)); 2682 } 2683 } 2684 2685 DEBUG_ONLY(write_block_fill_histogram()); 2686 } 2687 2688 #ifdef ASSERT 2689 void PSParallelCompact::verify_complete(SpaceId space_id) { 2690 // All Regions between space bottom() to new_top() should be marked as filled 2691 // and all Regions between new_top() and top() should be available (i.e., 2692 // should have been emptied). 2693 ParallelCompactData& sd = summary_data(); 2694 SpaceInfo si = _space_info[space_id]; 2695 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2696 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2697 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2698 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2699 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2700 2701 bool issued_a_warning = false; 2702 2703 size_t cur_region; 2704 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2705 const RegionData* const c = sd.region(cur_region); 2706 if (!c->completed()) { 2707 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2708 cur_region, c->destination_count()); 2709 issued_a_warning = true; 2710 } 2711 } 2712 2713 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2714 const RegionData* const c = sd.region(cur_region); 2715 if (!c->available()) { 2716 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2717 cur_region, c->destination_count()); 2718 issued_a_warning = true; 2719 } 2720 } 2721 2722 if (issued_a_warning) { 2723 print_region_ranges(); 2724 } 2725 } 2726 #endif // #ifdef ASSERT 2727 2728 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2729 _start_array->allocate_block(addr); 2730 compaction_manager()->update_contents(oop(addr)); 2731 } 2732 2733 // Update interior oops in the ranges of regions [beg_region, end_region). 2734 void 2735 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2736 SpaceId space_id, 2737 size_t beg_region, 2738 size_t end_region) { 2739 ParallelCompactData& sd = summary_data(); 2740 ParMarkBitMap* const mbm = mark_bitmap(); 2741 2742 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2743 HeapWord* const end_addr = sd.region_to_addr(end_region); 2744 assert(beg_region <= end_region, "bad region range"); 2745 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2746 2747 #ifdef ASSERT 2748 // Claim the regions to avoid triggering an assert when they are marked as 2749 // filled. 2750 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2751 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2752 } 2753 #endif // #ifdef ASSERT 2754 2755 if (beg_addr != space(space_id)->bottom()) { 2756 // Find the first live object or block of dead space that *starts* in this 2757 // range of regions. If a partial object crosses onto the region, skip it; 2758 // it will be marked for 'deferred update' when the object head is 2759 // processed. If dead space crosses onto the region, it is also skipped; it 2760 // will be filled when the prior region is processed. If neither of those 2761 // apply, the first word in the region is the start of a live object or dead 2762 // space. 2763 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2764 const RegionData* const cp = sd.region(beg_region); 2765 if (cp->partial_obj_size() != 0) { 2766 beg_addr = sd.partial_obj_end(beg_region); 2767 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2768 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2769 } 2770 } 2771 2772 if (beg_addr < end_addr) { 2773 // A live object or block of dead space starts in this range of Regions. 2774 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2775 2776 // Create closures and iterate. 2777 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2778 FillClosure fill_closure(cm, space_id); 2779 ParMarkBitMap::IterationStatus status; 2780 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2781 dense_prefix_end); 2782 if (status == ParMarkBitMap::incomplete) { 2783 update_closure.do_addr(update_closure.source()); 2784 } 2785 } 2786 2787 // Mark the regions as filled. 2788 RegionData* const beg_cp = sd.region(beg_region); 2789 RegionData* const end_cp = sd.region(end_region); 2790 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2791 cp->set_completed(); 2792 } 2793 } 2794 2795 // Return the SpaceId for the space containing addr. If addr is not in the 2796 // heap, last_space_id is returned. In debug mode it expects the address to be 2797 // in the heap and asserts such. 2798 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2799 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2800 2801 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2802 if (_space_info[id].space()->contains(addr)) { 2803 return SpaceId(id); 2804 } 2805 } 2806 2807 assert(false, "no space contains the addr"); 2808 return last_space_id; 2809 } 2810 2811 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2812 SpaceId id) { 2813 assert(id < last_space_id, "bad space id"); 2814 2815 ParallelCompactData& sd = summary_data(); 2816 const SpaceInfo* const space_info = _space_info + id; 2817 ObjectStartArray* const start_array = space_info->start_array(); 2818 2819 const MutableSpace* const space = space_info->space(); 2820 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2821 HeapWord* const beg_addr = space_info->dense_prefix(); 2822 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2823 2824 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2825 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2826 const RegionData* cur_region; 2827 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2828 HeapWord* const addr = cur_region->deferred_obj_addr(); 2829 if (addr != NULL) { 2830 if (start_array != NULL) { 2831 start_array->allocate_block(addr); 2832 } 2833 cm->update_contents(oop(addr)); 2834 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2835 } 2836 } 2837 } 2838 2839 // Skip over count live words starting from beg, and return the address of the 2840 // next live word. Unless marked, the word corresponding to beg is assumed to 2841 // be dead. Callers must either ensure beg does not correspond to the middle of 2842 // an object, or account for those live words in some other way. Callers must 2843 // also ensure that there are enough live words in the range [beg, end) to skip. 2844 HeapWord* 2845 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2846 { 2847 assert(count > 0, "sanity"); 2848 2849 ParMarkBitMap* m = mark_bitmap(); 2850 idx_t bits_to_skip = m->words_to_bits(count); 2851 idx_t cur_beg = m->addr_to_bit(beg); 2852 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2853 2854 do { 2855 cur_beg = m->find_obj_beg(cur_beg, search_end); 2856 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2857 const size_t obj_bits = cur_end - cur_beg + 1; 2858 if (obj_bits > bits_to_skip) { 2859 return m->bit_to_addr(cur_beg + bits_to_skip); 2860 } 2861 bits_to_skip -= obj_bits; 2862 cur_beg = cur_end + 1; 2863 } while (bits_to_skip > 0); 2864 2865 // Skipping the desired number of words landed just past the end of an object. 2866 // Find the start of the next object. 2867 cur_beg = m->find_obj_beg(cur_beg, search_end); 2868 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2869 return m->bit_to_addr(cur_beg); 2870 } 2871 2872 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2873 SpaceId src_space_id, 2874 size_t src_region_idx) 2875 { 2876 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2877 2878 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2879 if (split_info.dest_region_addr() == dest_addr) { 2880 // The partial object ending at the split point contains the first word to 2881 // be copied to dest_addr. 2882 return split_info.first_src_addr(); 2883 } 2884 2885 const ParallelCompactData& sd = summary_data(); 2886 ParMarkBitMap* const bitmap = mark_bitmap(); 2887 const size_t RegionSize = ParallelCompactData::RegionSize; 2888 2889 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2890 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2891 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2892 HeapWord* const src_region_destination = src_region_ptr->destination(); 2893 2894 assert(dest_addr >= src_region_destination, "wrong src region"); 2895 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2896 2897 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2898 HeapWord* const src_region_end = src_region_beg + RegionSize; 2899 2900 HeapWord* addr = src_region_beg; 2901 if (dest_addr == src_region_destination) { 2902 // Return the first live word in the source region. 2903 if (partial_obj_size == 0) { 2904 addr = bitmap->find_obj_beg(addr, src_region_end); 2905 assert(addr < src_region_end, "no objects start in src region"); 2906 } 2907 return addr; 2908 } 2909 2910 // Must skip some live data. 2911 size_t words_to_skip = dest_addr - src_region_destination; 2912 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2913 2914 if (partial_obj_size >= words_to_skip) { 2915 // All the live words to skip are part of the partial object. 2916 addr += words_to_skip; 2917 if (partial_obj_size == words_to_skip) { 2918 // Find the first live word past the partial object. 2919 addr = bitmap->find_obj_beg(addr, src_region_end); 2920 assert(addr < src_region_end, "wrong src region"); 2921 } 2922 return addr; 2923 } 2924 2925 // Skip over the partial object (if any). 2926 if (partial_obj_size != 0) { 2927 words_to_skip -= partial_obj_size; 2928 addr += partial_obj_size; 2929 } 2930 2931 // Skip over live words due to objects that start in the region. 2932 addr = skip_live_words(addr, src_region_end, words_to_skip); 2933 assert(addr < src_region_end, "wrong src region"); 2934 return addr; 2935 } 2936 2937 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2938 SpaceId src_space_id, 2939 size_t beg_region, 2940 HeapWord* end_addr) 2941 { 2942 ParallelCompactData& sd = summary_data(); 2943 2944 #ifdef ASSERT 2945 MutableSpace* const src_space = _space_info[src_space_id].space(); 2946 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2947 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2948 "src_space_id does not match beg_addr"); 2949 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2950 "src_space_id does not match end_addr"); 2951 #endif // #ifdef ASSERT 2952 2953 RegionData* const beg = sd.region(beg_region); 2954 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2955 2956 // Regions up to new_top() are enqueued if they become available. 2957 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2958 RegionData* const enqueue_end = 2959 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2960 2961 for (RegionData* cur = beg; cur < end; ++cur) { 2962 assert(cur->data_size() > 0, "region must have live data"); 2963 cur->decrement_destination_count(); 2964 if (cur < enqueue_end && cur->available() && cur->claim()) { 2965 cm->push_region(sd.region(cur)); 2966 } 2967 } 2968 } 2969 2970 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2971 SpaceId& src_space_id, 2972 HeapWord*& src_space_top, 2973 HeapWord* end_addr) 2974 { 2975 typedef ParallelCompactData::RegionData RegionData; 2976 2977 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2978 const size_t region_size = ParallelCompactData::RegionSize; 2979 2980 size_t src_region_idx = 0; 2981 2982 // Skip empty regions (if any) up to the top of the space. 2983 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2984 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2985 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2986 const RegionData* const top_region_ptr = 2987 sd.addr_to_region_ptr(top_aligned_up); 2988 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2989 ++src_region_ptr; 2990 } 2991 2992 if (src_region_ptr < top_region_ptr) { 2993 // The next source region is in the current space. Update src_region_idx 2994 // and the source address to match src_region_ptr. 2995 src_region_idx = sd.region(src_region_ptr); 2996 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2997 if (src_region_addr > closure.source()) { 2998 closure.set_source(src_region_addr); 2999 } 3000 return src_region_idx; 3001 } 3002 3003 // Switch to a new source space and find the first non-empty region. 3004 unsigned int space_id = src_space_id + 1; 3005 assert(space_id < last_space_id, "not enough spaces"); 3006 3007 HeapWord* const destination = closure.destination(); 3008 3009 do { 3010 MutableSpace* space = _space_info[space_id].space(); 3011 HeapWord* const bottom = space->bottom(); 3012 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3013 3014 // Iterate over the spaces that do not compact into themselves. 3015 if (bottom_cp->destination() != bottom) { 3016 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3017 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3018 3019 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3020 if (src_cp->live_obj_size() > 0) { 3021 // Found it. 3022 assert(src_cp->destination() == destination, 3023 "first live obj in the space must match the destination"); 3024 assert(src_cp->partial_obj_size() == 0, 3025 "a space cannot begin with a partial obj"); 3026 3027 src_space_id = SpaceId(space_id); 3028 src_space_top = space->top(); 3029 const size_t src_region_idx = sd.region(src_cp); 3030 closure.set_source(sd.region_to_addr(src_region_idx)); 3031 return src_region_idx; 3032 } else { 3033 assert(src_cp->data_size() == 0, "sanity"); 3034 } 3035 } 3036 } 3037 } while (++space_id < last_space_id); 3038 3039 assert(false, "no source region was found"); 3040 return 0; 3041 } 3042 3043 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3044 { 3045 typedef ParMarkBitMap::IterationStatus IterationStatus; 3046 const size_t RegionSize = ParallelCompactData::RegionSize; 3047 ParMarkBitMap* const bitmap = mark_bitmap(); 3048 ParallelCompactData& sd = summary_data(); 3049 RegionData* const region_ptr = sd.region(region_idx); 3050 3051 // Get the items needed to construct the closure. 3052 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3053 SpaceId dest_space_id = space_id(dest_addr); 3054 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3055 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3056 assert(dest_addr < new_top, "sanity"); 3057 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3058 3059 // Get the source region and related info. 3060 size_t src_region_idx = region_ptr->source_region(); 3061 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3062 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3063 3064 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3065 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3066 3067 // Adjust src_region_idx to prepare for decrementing destination counts (the 3068 // destination count is not decremented when a region is copied to itself). 3069 if (src_region_idx == region_idx) { 3070 src_region_idx += 1; 3071 } 3072 3073 if (bitmap->is_unmarked(closure.source())) { 3074 // The first source word is in the middle of an object; copy the remainder 3075 // of the object or as much as will fit. The fact that pointer updates were 3076 // deferred will be noted when the object header is processed. 3077 HeapWord* const old_src_addr = closure.source(); 3078 closure.copy_partial_obj(); 3079 if (closure.is_full()) { 3080 decrement_destination_counts(cm, src_space_id, src_region_idx, 3081 closure.source()); 3082 region_ptr->set_deferred_obj_addr(NULL); 3083 region_ptr->set_completed(); 3084 return; 3085 } 3086 3087 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3088 if (sd.region_align_down(old_src_addr) != end_addr) { 3089 // The partial object was copied from more than one source region. 3090 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3091 3092 // Move to the next source region, possibly switching spaces as well. All 3093 // args except end_addr may be modified. 3094 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3095 end_addr); 3096 } 3097 } 3098 3099 do { 3100 HeapWord* const cur_addr = closure.source(); 3101 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3102 src_space_top); 3103 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3104 3105 if (status == ParMarkBitMap::incomplete) { 3106 // The last obj that starts in the source region does not end in the 3107 // region. 3108 assert(closure.source() < end_addr, "sanity"); 3109 HeapWord* const obj_beg = closure.source(); 3110 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3111 src_space_top); 3112 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3113 if (obj_end < range_end) { 3114 // The end was found; the entire object will fit. 3115 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3116 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3117 } else { 3118 // The end was not found; the object will not fit. 3119 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3120 status = ParMarkBitMap::would_overflow; 3121 } 3122 } 3123 3124 if (status == ParMarkBitMap::would_overflow) { 3125 // The last object did not fit. Note that interior oop updates were 3126 // deferred, then copy enough of the object to fill the region. 3127 region_ptr->set_deferred_obj_addr(closure.destination()); 3128 status = closure.copy_until_full(); // copies from closure.source() 3129 3130 decrement_destination_counts(cm, src_space_id, src_region_idx, 3131 closure.source()); 3132 region_ptr->set_completed(); 3133 return; 3134 } 3135 3136 if (status == ParMarkBitMap::full) { 3137 decrement_destination_counts(cm, src_space_id, src_region_idx, 3138 closure.source()); 3139 region_ptr->set_deferred_obj_addr(NULL); 3140 region_ptr->set_completed(); 3141 return; 3142 } 3143 3144 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3145 3146 // Move to the next source region, possibly switching spaces as well. All 3147 // args except end_addr may be modified. 3148 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3149 end_addr); 3150 } while (true); 3151 } 3152 3153 void PSParallelCompact::fill_blocks(size_t region_idx) 3154 { 3155 // Fill in the block table elements for the specified region. Each block 3156 // table element holds the number of live words in the region that are to the 3157 // left of the first object that starts in the block. Thus only blocks in 3158 // which an object starts need to be filled. 3159 // 3160 // The algorithm scans the section of the bitmap that corresponds to the 3161 // region, keeping a running total of the live words. When an object start is 3162 // found, if it's the first to start in the block that contains it, the 3163 // current total is written to the block table element. 3164 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3165 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3166 const size_t RegionSize = ParallelCompactData::RegionSize; 3167 3168 ParallelCompactData& sd = summary_data(); 3169 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3170 if (partial_obj_size >= RegionSize) { 3171 return; // No objects start in this region. 3172 } 3173 3174 // Ensure the first loop iteration decides that the block has changed. 3175 size_t cur_block = sd.block_count(); 3176 3177 const ParMarkBitMap* const bitmap = mark_bitmap(); 3178 3179 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3180 assert((size_t)1 << Log2BitsPerBlock == 3181 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3182 3183 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3184 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3185 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3186 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3187 while (beg_bit < range_end) { 3188 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3189 if (new_block != cur_block) { 3190 cur_block = new_block; 3191 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3192 } 3193 3194 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3195 if (end_bit < range_end - 1) { 3196 live_bits += end_bit - beg_bit + 1; 3197 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3198 } else { 3199 return; 3200 } 3201 } 3202 } 3203 3204 void 3205 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3206 const MutableSpace* sp = space(space_id); 3207 if (sp->is_empty()) { 3208 return; 3209 } 3210 3211 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3212 ParMarkBitMap* const bitmap = mark_bitmap(); 3213 HeapWord* const dp_addr = dense_prefix(space_id); 3214 HeapWord* beg_addr = sp->bottom(); 3215 HeapWord* end_addr = sp->top(); 3216 3217 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3218 3219 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3220 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3221 if (beg_region < dp_region) { 3222 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3223 } 3224 3225 // The destination of the first live object that starts in the region is one 3226 // past the end of the partial object entering the region (if any). 3227 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3228 HeapWord* const new_top = _space_info[space_id].new_top(); 3229 assert(new_top >= dest_addr, "bad new_top value"); 3230 const size_t words = pointer_delta(new_top, dest_addr); 3231 3232 if (words > 0) { 3233 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3234 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3235 3236 ParMarkBitMap::IterationStatus status; 3237 status = bitmap->iterate(&closure, dest_addr, end_addr); 3238 assert(status == ParMarkBitMap::full, "iteration not complete"); 3239 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3240 "live objects skipped because closure is full"); 3241 } 3242 } 3243 3244 jlong PSParallelCompact::millis_since_last_gc() { 3245 // We need a monotonically non-decreasing time in ms but 3246 // os::javaTimeMillis() does not guarantee monotonicity. 3247 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3248 jlong ret_val = now - _time_of_last_gc; 3249 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3250 if (ret_val < 0) { 3251 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3252 return 0; 3253 } 3254 return ret_val; 3255 } 3256 3257 void PSParallelCompact::reset_millis_since_last_gc() { 3258 // We need a monotonically non-decreasing time in ms but 3259 // os::javaTimeMillis() does not guarantee monotonicity. 3260 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3261 } 3262 3263 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3264 { 3265 if (source() != destination()) { 3266 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3267 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3268 } 3269 update_state(words_remaining()); 3270 assert(is_full(), "sanity"); 3271 return ParMarkBitMap::full; 3272 } 3273 3274 void MoveAndUpdateClosure::copy_partial_obj() 3275 { 3276 size_t words = words_remaining(); 3277 3278 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3279 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3280 if (end_addr < range_end) { 3281 words = bitmap()->obj_size(source(), end_addr); 3282 } 3283 3284 // This test is necessary; if omitted, the pointer updates to a partial object 3285 // that crosses the dense prefix boundary could be overwritten. 3286 if (source() != destination()) { 3287 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3288 Copy::aligned_conjoint_words(source(), destination(), words); 3289 } 3290 update_state(words); 3291 } 3292 3293 ParMarkBitMapClosure::IterationStatus 3294 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3295 assert(destination() != NULL, "sanity"); 3296 assert(bitmap()->obj_size(addr) == words, "bad size"); 3297 3298 _source = addr; 3299 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3300 destination(), "wrong destination"); 3301 3302 if (words > words_remaining()) { 3303 return ParMarkBitMap::would_overflow; 3304 } 3305 3306 // The start_array must be updated even if the object is not moving. 3307 if (_start_array != NULL) { 3308 _start_array->allocate_block(destination()); 3309 } 3310 3311 if (destination() != source()) { 3312 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3313 Copy::aligned_conjoint_words(source(), destination(), words); 3314 } 3315 3316 oop moved_oop = (oop) destination(); 3317 compaction_manager()->update_contents(moved_oop); 3318 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3319 3320 update_state(words); 3321 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3322 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3323 } 3324 3325 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3326 ParCompactionManager* cm, 3327 PSParallelCompact::SpaceId space_id) : 3328 ParMarkBitMapClosure(mbm, cm), 3329 _space_id(space_id), 3330 _start_array(PSParallelCompact::start_array(space_id)) 3331 { 3332 } 3333 3334 // Updates the references in the object to their new values. 3335 ParMarkBitMapClosure::IterationStatus 3336 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3337 do_addr(addr); 3338 return ParMarkBitMap::incomplete; 3339 } 3340 3341 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3342 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3343 _start_array(PSParallelCompact::start_array(space_id)) 3344 { 3345 assert(space_id == PSParallelCompact::old_space_id, 3346 "cannot use FillClosure in the young gen"); 3347 } 3348 3349 ParMarkBitMapClosure::IterationStatus 3350 FillClosure::do_addr(HeapWord* addr, size_t size) { 3351 CollectedHeap::fill_with_objects(addr, size); 3352 HeapWord* const end = addr + size; 3353 do { 3354 _start_array->allocate_block(addr); 3355 addr += oop(addr)->size(); 3356 } while (addr < end); 3357 return ParMarkBitMap::incomplete; 3358 }