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 const int ParallelCompactData::RegionData::UNUSED = 0; 129 const int ParallelCompactData::RegionData::SHADOW = 1; 130 const int ParallelCompactData::RegionData::FILLED = 2; 131 const int ParallelCompactData::RegionData::FINISH = 3; 132 133 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 134 135 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 136 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 137 138 double PSParallelCompact::_dwl_mean; 139 double PSParallelCompact::_dwl_std_dev; 140 double PSParallelCompact::_dwl_first_term; 141 double PSParallelCompact::_dwl_adjustment; 142 #ifdef ASSERT 143 bool PSParallelCompact::_dwl_initialized = false; 144 #endif // #ifdef ASSERT 145 146 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 147 HeapWord* destination) 148 { 149 assert(src_region_idx != 0, "invalid src_region_idx"); 150 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 151 assert(destination != NULL, "invalid destination argument"); 152 153 _src_region_idx = src_region_idx; 154 _partial_obj_size = partial_obj_size; 155 _destination = destination; 156 157 // These fields may not be updated below, so make sure they're clear. 158 assert(_dest_region_addr == NULL, "should have been cleared"); 159 assert(_first_src_addr == NULL, "should have been cleared"); 160 161 // Determine the number of destination regions for the partial object. 162 HeapWord* const last_word = destination + partial_obj_size - 1; 163 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 164 HeapWord* const beg_region_addr = sd.region_align_down(destination); 165 HeapWord* const end_region_addr = sd.region_align_down(last_word); 166 167 if (beg_region_addr == end_region_addr) { 168 // One destination region. 169 _destination_count = 1; 170 if (end_region_addr == destination) { 171 // The destination falls on a region boundary, thus the first word of the 172 // partial object will be the first word copied to the destination region. 173 _dest_region_addr = end_region_addr; 174 _first_src_addr = sd.region_to_addr(src_region_idx); 175 } 176 } else { 177 // Two destination regions. When copied, the partial object will cross a 178 // destination region boundary, so a word somewhere within the partial 179 // object will be the first word copied to the second destination region. 180 _destination_count = 2; 181 _dest_region_addr = end_region_addr; 182 const size_t ofs = pointer_delta(end_region_addr, destination); 183 assert(ofs < _partial_obj_size, "sanity"); 184 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 185 } 186 } 187 188 void SplitInfo::clear() 189 { 190 _src_region_idx = 0; 191 _partial_obj_size = 0; 192 _destination = NULL; 193 _destination_count = 0; 194 _dest_region_addr = NULL; 195 _first_src_addr = NULL; 196 assert(!is_valid(), "sanity"); 197 } 198 199 #ifdef ASSERT 200 void SplitInfo::verify_clear() 201 { 202 assert(_src_region_idx == 0, "not clear"); 203 assert(_partial_obj_size == 0, "not clear"); 204 assert(_destination == NULL, "not clear"); 205 assert(_destination_count == 0, "not clear"); 206 assert(_dest_region_addr == NULL, "not clear"); 207 assert(_first_src_addr == NULL, "not clear"); 208 } 209 #endif // #ifdef ASSERT 210 211 212 void PSParallelCompact::print_on_error(outputStream* st) { 213 _mark_bitmap.print_on_error(st); 214 } 215 216 #ifndef PRODUCT 217 const char* PSParallelCompact::space_names[] = { 218 "old ", "eden", "from", "to " 219 }; 220 221 void PSParallelCompact::print_region_ranges() { 222 if (!log_develop_is_enabled(Trace, gc, compaction)) { 223 return; 224 } 225 Log(gc, compaction) log; 226 ResourceMark rm; 227 LogStream ls(log.trace()); 228 Universe::print_on(&ls); 229 log.trace("space bottom top end new_top"); 230 log.trace("------ ---------- ---------- ---------- ----------"); 231 232 for (unsigned int id = 0; id < last_space_id; ++id) { 233 const MutableSpace* space = _space_info[id].space(); 234 log.trace("%u %s " 235 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 236 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 237 id, space_names[id], 238 summary_data().addr_to_region_idx(space->bottom()), 239 summary_data().addr_to_region_idx(space->top()), 240 summary_data().addr_to_region_idx(space->end()), 241 summary_data().addr_to_region_idx(_space_info[id].new_top())); 242 } 243 } 244 245 void 246 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 247 { 248 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 249 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 250 251 ParallelCompactData& sd = PSParallelCompact::summary_data(); 252 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 253 log_develop_trace(gc, compaction)( 254 REGION_IDX_FORMAT " " PTR_FORMAT " " 255 REGION_IDX_FORMAT " " PTR_FORMAT " " 256 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 257 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 258 i, p2i(c->data_location()), dci, p2i(c->destination()), 259 c->partial_obj_size(), c->live_obj_size(), 260 c->data_size(), c->source_region(), c->destination_count()); 261 262 #undef REGION_IDX_FORMAT 263 #undef REGION_DATA_FORMAT 264 } 265 266 void 267 print_generic_summary_data(ParallelCompactData& summary_data, 268 HeapWord* const beg_addr, 269 HeapWord* const end_addr) 270 { 271 size_t total_words = 0; 272 size_t i = summary_data.addr_to_region_idx(beg_addr); 273 const size_t last = summary_data.addr_to_region_idx(end_addr); 274 HeapWord* pdest = 0; 275 276 while (i < last) { 277 ParallelCompactData::RegionData* c = summary_data.region(i); 278 if (c->data_size() != 0 || c->destination() != pdest) { 279 print_generic_summary_region(i, c); 280 total_words += c->data_size(); 281 pdest = c->destination(); 282 } 283 ++i; 284 } 285 286 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 287 } 288 289 void 290 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 291 HeapWord* const beg_addr, 292 HeapWord* const end_addr) { 293 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 294 } 295 296 void 297 print_generic_summary_data(ParallelCompactData& summary_data, 298 SpaceInfo* space_info) 299 { 300 if (!log_develop_is_enabled(Trace, gc, compaction)) { 301 return; 302 } 303 304 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 305 const MutableSpace* space = space_info[id].space(); 306 print_generic_summary_data(summary_data, space->bottom(), 307 MAX2(space->top(), space_info[id].new_top())); 308 } 309 } 310 311 void 312 print_initial_summary_data(ParallelCompactData& summary_data, 313 const MutableSpace* space) { 314 if (space->top() == space->bottom()) { 315 return; 316 } 317 318 const size_t region_size = ParallelCompactData::RegionSize; 319 typedef ParallelCompactData::RegionData RegionData; 320 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 321 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 322 const RegionData* c = summary_data.region(end_region - 1); 323 HeapWord* end_addr = c->destination() + c->data_size(); 324 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 325 326 // Print (and count) the full regions at the beginning of the space. 327 size_t full_region_count = 0; 328 size_t i = summary_data.addr_to_region_idx(space->bottom()); 329 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 330 ParallelCompactData::RegionData* c = summary_data.region(i); 331 log_develop_trace(gc, compaction)( 332 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 333 i, p2i(c->destination()), 334 c->partial_obj_size(), c->live_obj_size(), 335 c->data_size(), c->source_region(), c->destination_count()); 336 ++full_region_count; 337 ++i; 338 } 339 340 size_t live_to_right = live_in_space - full_region_count * region_size; 341 342 double max_reclaimed_ratio = 0.0; 343 size_t max_reclaimed_ratio_region = 0; 344 size_t max_dead_to_right = 0; 345 size_t max_live_to_right = 0; 346 347 // Print the 'reclaimed ratio' for regions while there is something live in 348 // the region or to the right of it. The remaining regions are empty (and 349 // uninteresting), and computing the ratio will result in division by 0. 350 while (i < end_region && live_to_right > 0) { 351 c = summary_data.region(i); 352 HeapWord* const region_addr = summary_data.region_to_addr(i); 353 const size_t used_to_right = pointer_delta(space->top(), region_addr); 354 const size_t dead_to_right = used_to_right - live_to_right; 355 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 356 357 if (reclaimed_ratio > max_reclaimed_ratio) { 358 max_reclaimed_ratio = reclaimed_ratio; 359 max_reclaimed_ratio_region = i; 360 max_dead_to_right = dead_to_right; 361 max_live_to_right = live_to_right; 362 } 363 364 ParallelCompactData::RegionData* c = summary_data.region(i); 365 log_develop_trace(gc, compaction)( 366 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 367 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 368 i, p2i(c->destination()), 369 c->partial_obj_size(), c->live_obj_size(), 370 c->data_size(), c->source_region(), c->destination_count(), 371 reclaimed_ratio, dead_to_right, live_to_right); 372 373 374 live_to_right -= c->data_size(); 375 ++i; 376 } 377 378 // Any remaining regions are empty. Print one more if there is one. 379 if (i < end_region) { 380 ParallelCompactData::RegionData* c = summary_data.region(i); 381 log_develop_trace(gc, compaction)( 382 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 383 i, p2i(c->destination()), 384 c->partial_obj_size(), c->live_obj_size(), 385 c->data_size(), c->source_region(), c->destination_count()); 386 } 387 388 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 389 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 390 } 391 392 void 393 print_initial_summary_data(ParallelCompactData& summary_data, 394 SpaceInfo* space_info) { 395 if (!log_develop_is_enabled(Trace, gc, compaction)) { 396 return; 397 } 398 399 unsigned int id = PSParallelCompact::old_space_id; 400 const MutableSpace* space; 401 do { 402 space = space_info[id].space(); 403 print_initial_summary_data(summary_data, space); 404 } while (++id < PSParallelCompact::eden_space_id); 405 406 do { 407 space = space_info[id].space(); 408 print_generic_summary_data(summary_data, space->bottom(), space->top()); 409 } while (++id < PSParallelCompact::last_space_id); 410 } 411 #endif // #ifndef PRODUCT 412 413 #ifdef ASSERT 414 size_t add_obj_count; 415 size_t add_obj_size; 416 size_t mark_bitmap_count; 417 size_t mark_bitmap_size; 418 #endif // #ifdef ASSERT 419 420 ParallelCompactData::ParallelCompactData() : 421 _region_start(NULL), 422 DEBUG_ONLY(_region_end(NULL) COMMA) 423 _region_vspace(NULL), 424 _reserved_byte_size(0), 425 _region_data(NULL), 426 _region_count(0), 427 _block_vspace(NULL), 428 _block_data(NULL), 429 _block_count(0) {} 430 431 bool ParallelCompactData::initialize(MemRegion covered_region) 432 { 433 _region_start = covered_region.start(); 434 const size_t region_size = covered_region.word_size(); 435 DEBUG_ONLY(_region_end = _region_start + region_size;) 436 437 assert(region_align_down(_region_start) == _region_start, 438 "region start not aligned"); 439 assert((region_size & RegionSizeOffsetMask) == 0, 440 "region size not a multiple of RegionSize"); 441 442 bool result = initialize_region_data(region_size) && initialize_block_data(); 443 return result; 444 } 445 446 PSVirtualSpace* 447 ParallelCompactData::create_vspace(size_t count, size_t element_size) 448 { 449 const size_t raw_bytes = count * element_size; 450 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 451 const size_t granularity = os::vm_allocation_granularity(); 452 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 453 454 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 455 MAX2(page_sz, granularity); 456 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 457 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 458 rs.size()); 459 460 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 461 462 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 463 if (vspace != 0) { 464 if (vspace->expand_by(_reserved_byte_size)) { 465 return vspace; 466 } 467 delete vspace; 468 // Release memory reserved in the space. 469 rs.release(); 470 } 471 472 return 0; 473 } 474 475 bool ParallelCompactData::initialize_region_data(size_t region_size) 476 { 477 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 478 _region_vspace = create_vspace(count, sizeof(RegionData)); 479 if (_region_vspace != 0) { 480 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 481 _region_count = count; 482 return true; 483 } 484 return false; 485 } 486 487 bool ParallelCompactData::initialize_block_data() 488 { 489 assert(_region_count != 0, "region data must be initialized first"); 490 const size_t count = _region_count << Log2BlocksPerRegion; 491 _block_vspace = create_vspace(count, sizeof(BlockData)); 492 if (_block_vspace != 0) { 493 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 494 _block_count = count; 495 return true; 496 } 497 return false; 498 } 499 500 void ParallelCompactData::clear() 501 { 502 memset(_region_data, 0, _region_vspace->committed_size()); 503 memset(_block_data, 0, _block_vspace->committed_size()); 504 } 505 506 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 507 assert(beg_region <= _region_count, "beg_region out of range"); 508 assert(end_region <= _region_count, "end_region out of range"); 509 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 510 511 const size_t region_cnt = end_region - beg_region; 512 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 513 514 const size_t beg_block = beg_region * BlocksPerRegion; 515 const size_t block_cnt = region_cnt * BlocksPerRegion; 516 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 517 } 518 519 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 520 { 521 const RegionData* cur_cp = region(region_idx); 522 const RegionData* const end_cp = region(region_count() - 1); 523 524 HeapWord* result = region_to_addr(region_idx); 525 if (cur_cp < end_cp) { 526 do { 527 result += cur_cp->partial_obj_size(); 528 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 529 } 530 return result; 531 } 532 533 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 534 { 535 const size_t obj_ofs = pointer_delta(addr, _region_start); 536 const size_t beg_region = obj_ofs >> Log2RegionSize; 537 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 538 539 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 540 DEBUG_ONLY(Atomic::add(len, &add_obj_size);) 541 542 if (beg_region == end_region) { 543 // All in one region. 544 _region_data[beg_region].add_live_obj(len); 545 return; 546 } 547 548 // First region. 549 const size_t beg_ofs = region_offset(addr); 550 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 551 552 Klass* klass = ((oop)addr)->klass(); 553 // Middle regions--completely spanned by this object. 554 for (size_t region = beg_region + 1; region < end_region; ++region) { 555 _region_data[region].set_partial_obj_size(RegionSize); 556 _region_data[region].set_partial_obj_addr(addr); 557 } 558 559 // Last region. 560 const size_t end_ofs = region_offset(addr + len - 1); 561 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 562 _region_data[end_region].set_partial_obj_addr(addr); 563 } 564 565 void 566 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 567 { 568 assert(region_offset(beg) == 0, "not RegionSize aligned"); 569 assert(region_offset(end) == 0, "not RegionSize aligned"); 570 571 size_t cur_region = addr_to_region_idx(beg); 572 const size_t end_region = addr_to_region_idx(end); 573 HeapWord* addr = beg; 574 while (cur_region < end_region) { 575 _region_data[cur_region].set_destination(addr); 576 _region_data[cur_region].set_destination_count(0); 577 _region_data[cur_region].set_source_region(cur_region); 578 _region_data[cur_region].set_data_location(addr); 579 580 // Update live_obj_size so the region appears completely full. 581 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 582 _region_data[cur_region].set_live_obj_size(live_size); 583 584 ++cur_region; 585 addr += RegionSize; 586 } 587 } 588 589 // Find the point at which a space can be split and, if necessary, record the 590 // split point. 591 // 592 // If the current src region (which overflowed the destination space) doesn't 593 // have a partial object, the split point is at the beginning of the current src 594 // region (an "easy" split, no extra bookkeeping required). 595 // 596 // If the current src region has a partial object, the split point is in the 597 // region where that partial object starts (call it the split_region). If 598 // split_region has a partial object, then the split point is just after that 599 // partial object (a "hard" split where we have to record the split data and 600 // zero the partial_obj_size field). With a "hard" split, we know that the 601 // partial_obj ends within split_region because the partial object that caused 602 // the overflow starts in split_region. If split_region doesn't have a partial 603 // obj, then the split is at the beginning of split_region (another "easy" 604 // split). 605 HeapWord* 606 ParallelCompactData::summarize_split_space(size_t src_region, 607 SplitInfo& split_info, 608 HeapWord* destination, 609 HeapWord* target_end, 610 HeapWord** target_next) 611 { 612 assert(destination <= target_end, "sanity"); 613 assert(destination + _region_data[src_region].data_size() > target_end, 614 "region should not fit into target space"); 615 assert(is_region_aligned(target_end), "sanity"); 616 617 size_t split_region = src_region; 618 HeapWord* split_destination = destination; 619 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 620 621 if (destination + partial_obj_size > target_end) { 622 // The split point is just after the partial object (if any) in the 623 // src_region that contains the start of the object that overflowed the 624 // destination space. 625 // 626 // Find the start of the "overflow" object and set split_region to the 627 // region containing it. 628 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 629 split_region = addr_to_region_idx(overflow_obj); 630 631 // Clear the source_region field of all destination regions whose first word 632 // came from data after the split point (a non-null source_region field 633 // implies a region must be filled). 634 // 635 // An alternative to the simple loop below: clear during post_compact(), 636 // which uses memcpy instead of individual stores, and is easy to 637 // parallelize. (The downside is that it clears the entire RegionData 638 // object as opposed to just one field.) 639 // 640 // post_compact() would have to clear the summary data up to the highest 641 // address that was written during the summary phase, which would be 642 // 643 // max(top, max(new_top, clear_top)) 644 // 645 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 646 // to target_end. 647 const RegionData* const sr = region(split_region); 648 const size_t beg_idx = 649 addr_to_region_idx(region_align_up(sr->destination() + 650 sr->partial_obj_size())); 651 const size_t end_idx = addr_to_region_idx(target_end); 652 653 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 654 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 655 _region_data[idx].set_source_region(0); 656 } 657 658 // Set split_destination and partial_obj_size to reflect the split region. 659 split_destination = sr->destination(); 660 partial_obj_size = sr->partial_obj_size(); 661 } 662 663 // The split is recorded only if a partial object extends onto the region. 664 if (partial_obj_size != 0) { 665 _region_data[split_region].set_partial_obj_size(0); 666 split_info.record(split_region, partial_obj_size, split_destination); 667 } 668 669 // Setup the continuation addresses. 670 *target_next = split_destination + partial_obj_size; 671 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 672 673 if (log_develop_is_enabled(Trace, gc, compaction)) { 674 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 675 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 676 split_type, p2i(source_next), split_region, partial_obj_size); 677 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 678 split_type, p2i(split_destination), 679 addr_to_region_idx(split_destination), 680 p2i(*target_next)); 681 682 if (partial_obj_size != 0) { 683 HeapWord* const po_beg = split_info.destination(); 684 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 685 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 686 split_type, 687 p2i(po_beg), addr_to_region_idx(po_beg), 688 p2i(po_end), addr_to_region_idx(po_end)); 689 } 690 } 691 692 return source_next; 693 } 694 695 bool ParallelCompactData::summarize(SplitInfo& split_info, 696 HeapWord* source_beg, HeapWord* source_end, 697 HeapWord** source_next, 698 HeapWord* target_beg, HeapWord* target_end, 699 HeapWord** target_next) 700 { 701 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 702 log_develop_trace(gc, compaction)( 703 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 704 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 705 p2i(source_beg), p2i(source_end), p2i(source_next_val), 706 p2i(target_beg), p2i(target_end), p2i(*target_next)); 707 708 size_t cur_region = addr_to_region_idx(source_beg); 709 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 710 711 HeapWord *dest_addr = target_beg; 712 while (cur_region < end_region) { 713 // The destination must be set even if the region has no data. 714 _region_data[cur_region].set_destination(dest_addr); 715 716 size_t words = _region_data[cur_region].data_size(); 717 if (words > 0) { 718 // If cur_region does not fit entirely into the target space, find a point 719 // at which the source space can be 'split' so that part is copied to the 720 // target space and the rest is copied elsewhere. 721 if (dest_addr + words > target_end) { 722 assert(source_next != NULL, "source_next is NULL when splitting"); 723 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 724 target_end, target_next); 725 return false; 726 } 727 728 // Compute the destination_count for cur_region, and if necessary, update 729 // source_region for a destination region. The source_region field is 730 // updated if cur_region is the first (left-most) region to be copied to a 731 // destination region. 732 // 733 // The destination_count calculation is a bit subtle. A region that has 734 // data that compacts into itself does not count itself as a destination. 735 // This maintains the invariant that a zero count means the region is 736 // available and can be claimed and then filled. 737 uint destination_count = 0; 738 if (split_info.is_split(cur_region)) { 739 // The current region has been split: the partial object will be copied 740 // to one destination space and the remaining data will be copied to 741 // another destination space. Adjust the initial destination_count and, 742 // if necessary, set the source_region field if the partial object will 743 // cross a destination region boundary. 744 destination_count = split_info.destination_count(); 745 if (destination_count == 2) { 746 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 747 _region_data[dest_idx].set_source_region(cur_region); 748 } 749 } 750 751 HeapWord* const last_addr = dest_addr + words - 1; 752 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 753 const size_t dest_region_2 = addr_to_region_idx(last_addr); 754 755 // Initially assume that the destination regions will be the same and 756 // adjust the value below if necessary. Under this assumption, if 757 // cur_region == dest_region_2, then cur_region will be compacted 758 // completely into itself. 759 destination_count += cur_region == dest_region_2 ? 0 : 1; 760 if (dest_region_1 != dest_region_2) { 761 // Destination regions differ; adjust destination_count. 762 destination_count += 1; 763 // Data from cur_region will be copied to the start of dest_region_2. 764 _region_data[dest_region_2].set_source_region(cur_region); 765 } else if (region_offset(dest_addr) == 0) { 766 // Data from cur_region will be copied to the start of the destination 767 // region. 768 _region_data[dest_region_1].set_source_region(cur_region); 769 } 770 771 _region_data[cur_region].set_destination_count(destination_count); 772 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 773 dest_addr += words; 774 } 775 776 ++cur_region; 777 } 778 779 *target_next = dest_addr; 780 return true; 781 } 782 783 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 784 assert(addr != NULL, "Should detect NULL oop earlier"); 785 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 786 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 787 788 // Region covering the object. 789 RegionData* const region_ptr = addr_to_region_ptr(addr); 790 HeapWord* result = region_ptr->destination(); 791 792 // If the entire Region is live, the new location is region->destination + the 793 // offset of the object within in the Region. 794 795 // Run some performance tests to determine if this special case pays off. It 796 // is worth it for pointers into the dense prefix. If the optimization to 797 // avoid pointer updates in regions that only point to the dense prefix is 798 // ever implemented, this should be revisited. 799 if (region_ptr->data_size() == RegionSize) { 800 result += region_offset(addr); 801 return result; 802 } 803 804 // Otherwise, the new location is region->destination + block offset + the 805 // number of live words in the Block that are (a) to the left of addr and (b) 806 // due to objects that start in the Block. 807 808 // Fill in the block table if necessary. This is unsynchronized, so multiple 809 // threads may fill the block table for a region (harmless, since it is 810 // idempotent). 811 if (!region_ptr->blocks_filled()) { 812 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 813 region_ptr->set_blocks_filled(); 814 } 815 816 HeapWord* const search_start = block_align_down(addr); 817 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 818 819 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 820 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 821 result += block_offset + live; 822 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 823 return result; 824 } 825 826 #ifdef ASSERT 827 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 828 { 829 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 830 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 831 for (const size_t* p = beg; p < end; ++p) { 832 assert(*p == 0, "not zero"); 833 } 834 } 835 836 void ParallelCompactData::verify_clear() 837 { 838 verify_clear(_region_vspace); 839 verify_clear(_block_vspace); 840 } 841 #endif // #ifdef ASSERT 842 843 STWGCTimer PSParallelCompact::_gc_timer; 844 ParallelOldTracer PSParallelCompact::_gc_tracer; 845 elapsedTimer PSParallelCompact::_accumulated_time; 846 unsigned int PSParallelCompact::_total_invocations = 0; 847 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 848 jlong PSParallelCompact::_time_of_last_gc = 0; 849 CollectorCounters* PSParallelCompact::_counters = NULL; 850 ParMarkBitMap PSParallelCompact::_mark_bitmap; 851 ParallelCompactData PSParallelCompact::_summary_data; 852 853 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 854 855 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 856 857 class PCReferenceProcessor: public ReferenceProcessor { 858 public: 859 PCReferenceProcessor( 860 BoolObjectClosure* is_subject_to_discovery, 861 BoolObjectClosure* is_alive_non_header) : 862 ReferenceProcessor(is_subject_to_discovery, 863 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 864 ParallelGCThreads, // mt processing degree 865 true, // mt discovery 866 ParallelGCThreads, // mt discovery degree 867 true, // atomic_discovery 868 is_alive_non_header) { 869 } 870 871 template<typename T> bool discover(oop obj, ReferenceType type) { 872 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj); 873 T heap_oop = RawAccess<>::oop_load(referent_addr); 874 oop referent = CompressedOops::decode_not_null(heap_oop); 875 return PSParallelCompact::mark_bitmap()->is_unmarked(referent) 876 && ReferenceProcessor::discover_reference(obj, type); 877 } 878 virtual bool discover_reference(oop obj, ReferenceType type) { 879 if (UseCompressedOops) { 880 return discover<narrowOop>(obj, type); 881 } else { 882 return discover<oop>(obj, type); 883 } 884 } 885 }; 886 887 void PSParallelCompact::post_initialize() { 888 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 889 _span_based_discoverer.set_span(heap->reserved_region()); 890 _ref_processor = 891 new PCReferenceProcessor(&_span_based_discoverer, 892 &_is_alive_closure); // non-header is alive closure 893 894 _counters = new CollectorCounters("Parallel full collection pauses", 1); 895 896 // Initialize static fields in ParCompactionManager. 897 ParCompactionManager::initialize(mark_bitmap()); 898 } 899 900 bool PSParallelCompact::initialize() { 901 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 902 MemRegion mr = heap->reserved_region(); 903 904 // Was the old gen get allocated successfully? 905 if (!heap->old_gen()->is_allocated()) { 906 return false; 907 } 908 909 initialize_space_info(); 910 initialize_dead_wood_limiter(); 911 912 if (!_mark_bitmap.initialize(mr)) { 913 vm_shutdown_during_initialization( 914 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 915 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 916 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 917 return false; 918 } 919 920 if (!_summary_data.initialize(mr)) { 921 vm_shutdown_during_initialization( 922 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 923 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 924 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 925 return false; 926 } 927 928 return true; 929 } 930 931 void PSParallelCompact::initialize_space_info() 932 { 933 memset(&_space_info, 0, sizeof(_space_info)); 934 935 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 936 PSYoungGen* young_gen = heap->young_gen(); 937 938 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 939 _space_info[eden_space_id].set_space(young_gen->eden_space()); 940 _space_info[from_space_id].set_space(young_gen->from_space()); 941 _space_info[to_space_id].set_space(young_gen->to_space()); 942 943 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 944 } 945 946 void PSParallelCompact::initialize_dead_wood_limiter() 947 { 948 const size_t max = 100; 949 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 950 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 951 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 952 DEBUG_ONLY(_dwl_initialized = true;) 953 _dwl_adjustment = normal_distribution(1.0); 954 } 955 956 void 957 PSParallelCompact::clear_data_covering_space(SpaceId id) 958 { 959 // At this point, top is the value before GC, new_top() is the value that will 960 // be set at the end of GC. The marking bitmap is cleared to top; nothing 961 // should be marked above top. The summary data is cleared to the larger of 962 // top & new_top. 963 MutableSpace* const space = _space_info[id].space(); 964 HeapWord* const bot = space->bottom(); 965 HeapWord* const top = space->top(); 966 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 967 968 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 969 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 970 _mark_bitmap.clear_range(beg_bit, end_bit); 971 972 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 973 const size_t end_region = 974 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 975 _summary_data.clear_range(beg_region, end_region); 976 977 // Clear the data used to 'split' regions. 978 SplitInfo& split_info = _space_info[id].split_info(); 979 if (split_info.is_valid()) { 980 split_info.clear(); 981 } 982 DEBUG_ONLY(split_info.verify_clear();) 983 } 984 985 void PSParallelCompact::pre_compact() 986 { 987 // Update the from & to space pointers in space_info, since they are swapped 988 // at each young gen gc. Do the update unconditionally (even though a 989 // promotion failure does not swap spaces) because an unknown number of young 990 // collections will have swapped the spaces an unknown number of times. 991 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 992 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 993 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 994 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 995 996 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 997 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 998 999 // Increment the invocation count 1000 heap->increment_total_collections(true); 1001 1002 // We need to track unique mark sweep invocations as well. 1003 _total_invocations++; 1004 1005 heap->print_heap_before_gc(); 1006 heap->trace_heap_before_gc(&_gc_tracer); 1007 1008 // Fill in TLABs 1009 heap->ensure_parsability(true); // retire TLABs 1010 1011 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1012 HandleMark hm; // Discard invalid handles created during verification 1013 Universe::verify("Before GC"); 1014 } 1015 1016 // Verify object start arrays 1017 if (VerifyObjectStartArray && 1018 VerifyBeforeGC) { 1019 heap->old_gen()->verify_object_start_array(); 1020 } 1021 1022 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1023 DEBUG_ONLY(summary_data().verify_clear();) 1024 1025 ParCompactionManager::reset_all_bitmap_query_caches(); 1026 } 1027 1028 void PSParallelCompact::post_compact() 1029 { 1030 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1031 ParCompactionManager::dequeue_shadow_region(); 1032 1033 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1034 // Clear the marking bitmap, summary data and split info. 1035 clear_data_covering_space(SpaceId(id)); 1036 // Update top(). Must be done after clearing the bitmap and summary data. 1037 _space_info[id].publish_new_top(); 1038 } 1039 1040 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1041 MutableSpace* const from_space = _space_info[from_space_id].space(); 1042 MutableSpace* const to_space = _space_info[to_space_id].space(); 1043 1044 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1045 bool eden_empty = eden_space->is_empty(); 1046 if (!eden_empty) { 1047 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1048 heap->young_gen(), heap->old_gen()); 1049 } 1050 1051 // Update heap occupancy information which is used as input to the soft ref 1052 // clearing policy at the next gc. 1053 Universe::update_heap_info_at_gc(); 1054 1055 bool young_gen_empty = eden_empty && from_space->is_empty() && 1056 to_space->is_empty(); 1057 1058 PSCardTable* ct = heap->card_table(); 1059 MemRegion old_mr = heap->old_gen()->reserved(); 1060 if (young_gen_empty) { 1061 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1062 } else { 1063 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1064 } 1065 1066 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1067 ClassLoaderDataGraph::purge(); 1068 MetaspaceUtils::verify_metrics(); 1069 1070 heap->prune_scavengable_nmethods(); 1071 1072 #if COMPILER2_OR_JVMCI 1073 DerivedPointerTable::update_pointers(); 1074 #endif 1075 1076 if (ZapUnusedHeapArea) { 1077 heap->gen_mangle_unused_area(); 1078 } 1079 1080 // Update time of last GC 1081 reset_millis_since_last_gc(); 1082 } 1083 1084 HeapWord* 1085 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1086 bool maximum_compaction) 1087 { 1088 const size_t region_size = ParallelCompactData::RegionSize; 1089 const ParallelCompactData& sd = summary_data(); 1090 1091 const MutableSpace* const space = _space_info[id].space(); 1092 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1093 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1094 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1095 1096 // Skip full regions at the beginning of the space--they are necessarily part 1097 // of the dense prefix. 1098 size_t full_count = 0; 1099 const RegionData* cp; 1100 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1101 ++full_count; 1102 } 1103 1104 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1105 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1106 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1107 if (maximum_compaction || cp == end_cp || interval_ended) { 1108 _maximum_compaction_gc_num = total_invocations(); 1109 return sd.region_to_addr(cp); 1110 } 1111 1112 HeapWord* const new_top = _space_info[id].new_top(); 1113 const size_t space_live = pointer_delta(new_top, space->bottom()); 1114 const size_t space_used = space->used_in_words(); 1115 const size_t space_capacity = space->capacity_in_words(); 1116 1117 const double cur_density = double(space_live) / space_capacity; 1118 const double deadwood_density = 1119 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1120 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1121 1122 if (TraceParallelOldGCDensePrefix) { 1123 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1124 cur_density, deadwood_density, deadwood_goal); 1125 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1126 "space_cap=" SIZE_FORMAT, 1127 space_live, space_used, 1128 space_capacity); 1129 } 1130 1131 // XXX - Use binary search? 1132 HeapWord* dense_prefix = sd.region_to_addr(cp); 1133 const RegionData* full_cp = cp; 1134 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1135 while (cp < end_cp) { 1136 HeapWord* region_destination = cp->destination(); 1137 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1138 if (TraceParallelOldGCDensePrefix && Verbose) { 1139 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1140 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1141 sd.region(cp), p2i(region_destination), 1142 p2i(dense_prefix), cur_deadwood); 1143 } 1144 1145 if (cur_deadwood >= deadwood_goal) { 1146 // Found the region that has the correct amount of deadwood to the left. 1147 // This typically occurs after crossing a fairly sparse set of regions, so 1148 // iterate backwards over those sparse regions, looking for the region 1149 // that has the lowest density of live objects 'to the right.' 1150 size_t space_to_left = sd.region(cp) * region_size; 1151 size_t live_to_left = space_to_left - cur_deadwood; 1152 size_t space_to_right = space_capacity - space_to_left; 1153 size_t live_to_right = space_live - live_to_left; 1154 double density_to_right = double(live_to_right) / space_to_right; 1155 while (cp > full_cp) { 1156 --cp; 1157 const size_t prev_region_live_to_right = live_to_right - 1158 cp->data_size(); 1159 const size_t prev_region_space_to_right = space_to_right + region_size; 1160 double prev_region_density_to_right = 1161 double(prev_region_live_to_right) / prev_region_space_to_right; 1162 if (density_to_right <= prev_region_density_to_right) { 1163 return dense_prefix; 1164 } 1165 if (TraceParallelOldGCDensePrefix && Verbose) { 1166 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1167 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1168 prev_region_density_to_right); 1169 } 1170 dense_prefix -= region_size; 1171 live_to_right = prev_region_live_to_right; 1172 space_to_right = prev_region_space_to_right; 1173 density_to_right = prev_region_density_to_right; 1174 } 1175 return dense_prefix; 1176 } 1177 1178 dense_prefix += region_size; 1179 ++cp; 1180 } 1181 1182 return dense_prefix; 1183 } 1184 1185 #ifndef PRODUCT 1186 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1187 const SpaceId id, 1188 const bool maximum_compaction, 1189 HeapWord* const addr) 1190 { 1191 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1192 RegionData* const cp = summary_data().region(region_idx); 1193 const MutableSpace* const space = _space_info[id].space(); 1194 HeapWord* const new_top = _space_info[id].new_top(); 1195 1196 const size_t space_live = pointer_delta(new_top, space->bottom()); 1197 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1198 const size_t space_cap = space->capacity_in_words(); 1199 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1200 const size_t live_to_right = new_top - cp->destination(); 1201 const size_t dead_to_right = space->top() - addr - live_to_right; 1202 1203 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1204 "spl=" SIZE_FORMAT " " 1205 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1206 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1207 " ratio=%10.8f", 1208 algorithm, p2i(addr), region_idx, 1209 space_live, 1210 dead_to_left, dead_to_left_pct, 1211 dead_to_right, live_to_right, 1212 double(dead_to_right) / live_to_right); 1213 } 1214 #endif // #ifndef PRODUCT 1215 1216 // Return a fraction indicating how much of the generation can be treated as 1217 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1218 // based on the density of live objects in the generation to determine a limit, 1219 // which is then adjusted so the return value is min_percent when the density is 1220 // 1. 1221 // 1222 // The following table shows some return values for a different values of the 1223 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1224 // min_percent is 1. 1225 // 1226 // fraction allowed as dead wood 1227 // ----------------------------------------------------------------- 1228 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1229 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1230 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1231 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1232 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1233 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1234 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1235 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1236 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1237 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1238 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1239 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1240 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1241 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1242 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1243 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1244 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1245 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1246 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1247 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1248 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1249 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1250 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1251 1252 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1253 { 1254 assert(_dwl_initialized, "uninitialized"); 1255 1256 // The raw limit is the value of the normal distribution at x = density. 1257 const double raw_limit = normal_distribution(density); 1258 1259 // Adjust the raw limit so it becomes the minimum when the density is 1. 1260 // 1261 // First subtract the adjustment value (which is simply the precomputed value 1262 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1263 // Then add the minimum value, so the minimum is returned when the density is 1264 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1265 const double min = double(min_percent) / 100.0; 1266 const double limit = raw_limit - _dwl_adjustment + min; 1267 return MAX2(limit, 0.0); 1268 } 1269 1270 ParallelCompactData::RegionData* 1271 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1272 const RegionData* end) 1273 { 1274 const size_t region_size = ParallelCompactData::RegionSize; 1275 ParallelCompactData& sd = summary_data(); 1276 size_t left = sd.region(beg); 1277 size_t right = end > beg ? sd.region(end) - 1 : left; 1278 1279 // Binary search. 1280 while (left < right) { 1281 // Equivalent to (left + right) / 2, but does not overflow. 1282 const size_t middle = left + (right - left) / 2; 1283 RegionData* const middle_ptr = sd.region(middle); 1284 HeapWord* const dest = middle_ptr->destination(); 1285 HeapWord* const addr = sd.region_to_addr(middle); 1286 assert(dest != NULL, "sanity"); 1287 assert(dest <= addr, "must move left"); 1288 1289 if (middle > left && dest < addr) { 1290 right = middle - 1; 1291 } else if (middle < right && middle_ptr->data_size() == region_size) { 1292 left = middle + 1; 1293 } else { 1294 return middle_ptr; 1295 } 1296 } 1297 return sd.region(left); 1298 } 1299 1300 ParallelCompactData::RegionData* 1301 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1302 const RegionData* end, 1303 size_t dead_words) 1304 { 1305 ParallelCompactData& sd = summary_data(); 1306 size_t left = sd.region(beg); 1307 size_t right = end > beg ? sd.region(end) - 1 : left; 1308 1309 // Binary search. 1310 while (left < right) { 1311 // Equivalent to (left + right) / 2, but does not overflow. 1312 const size_t middle = left + (right - left) / 2; 1313 RegionData* const middle_ptr = sd.region(middle); 1314 HeapWord* const dest = middle_ptr->destination(); 1315 HeapWord* const addr = sd.region_to_addr(middle); 1316 assert(dest != NULL, "sanity"); 1317 assert(dest <= addr, "must move left"); 1318 1319 const size_t dead_to_left = pointer_delta(addr, dest); 1320 if (middle > left && dead_to_left > dead_words) { 1321 right = middle - 1; 1322 } else if (middle < right && dead_to_left < dead_words) { 1323 left = middle + 1; 1324 } else { 1325 return middle_ptr; 1326 } 1327 } 1328 return sd.region(left); 1329 } 1330 1331 // The result is valid during the summary phase, after the initial summarization 1332 // of each space into itself, and before final summarization. 1333 inline double 1334 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1335 HeapWord* const bottom, 1336 HeapWord* const top, 1337 HeapWord* const new_top) 1338 { 1339 ParallelCompactData& sd = summary_data(); 1340 1341 assert(cp != NULL, "sanity"); 1342 assert(bottom != NULL, "sanity"); 1343 assert(top != NULL, "sanity"); 1344 assert(new_top != NULL, "sanity"); 1345 assert(top >= new_top, "summary data problem?"); 1346 assert(new_top > bottom, "space is empty; should not be here"); 1347 assert(new_top >= cp->destination(), "sanity"); 1348 assert(top >= sd.region_to_addr(cp), "sanity"); 1349 1350 HeapWord* const destination = cp->destination(); 1351 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1352 const size_t compacted_region_live = pointer_delta(new_top, destination); 1353 const size_t compacted_region_used = pointer_delta(top, 1354 sd.region_to_addr(cp)); 1355 const size_t reclaimable = compacted_region_used - compacted_region_live; 1356 1357 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1358 return double(reclaimable) / divisor; 1359 } 1360 1361 // Return the address of the end of the dense prefix, a.k.a. the start of the 1362 // compacted region. The address is always on a region boundary. 1363 // 1364 // Completely full regions at the left are skipped, since no compaction can 1365 // occur in those regions. Then the maximum amount of dead wood to allow is 1366 // computed, based on the density (amount live / capacity) of the generation; 1367 // the region with approximately that amount of dead space to the left is 1368 // identified as the limit region. Regions between the last completely full 1369 // region and the limit region are scanned and the one that has the best 1370 // (maximum) reclaimed_ratio() is selected. 1371 HeapWord* 1372 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1373 bool maximum_compaction) 1374 { 1375 const size_t region_size = ParallelCompactData::RegionSize; 1376 const ParallelCompactData& sd = summary_data(); 1377 1378 const MutableSpace* const space = _space_info[id].space(); 1379 HeapWord* const top = space->top(); 1380 HeapWord* const top_aligned_up = sd.region_align_up(top); 1381 HeapWord* const new_top = _space_info[id].new_top(); 1382 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1383 HeapWord* const bottom = space->bottom(); 1384 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1385 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1386 const RegionData* const new_top_cp = 1387 sd.addr_to_region_ptr(new_top_aligned_up); 1388 1389 // Skip full regions at the beginning of the space--they are necessarily part 1390 // of the dense prefix. 1391 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1392 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1393 space->is_empty(), "no dead space allowed to the left"); 1394 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1395 "region must have dead space"); 1396 1397 // The gc number is saved whenever a maximum compaction is done, and used to 1398 // determine when the maximum compaction interval has expired. This avoids 1399 // successive max compactions for different reasons. 1400 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1401 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1402 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1403 total_invocations() == HeapFirstMaximumCompactionCount; 1404 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1405 _maximum_compaction_gc_num = total_invocations(); 1406 return sd.region_to_addr(full_cp); 1407 } 1408 1409 const size_t space_live = pointer_delta(new_top, bottom); 1410 const size_t space_used = space->used_in_words(); 1411 const size_t space_capacity = space->capacity_in_words(); 1412 1413 const double density = double(space_live) / double(space_capacity); 1414 const size_t min_percent_free = MarkSweepDeadRatio; 1415 const double limiter = dead_wood_limiter(density, min_percent_free); 1416 const size_t dead_wood_max = space_used - space_live; 1417 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1418 dead_wood_max); 1419 1420 if (TraceParallelOldGCDensePrefix) { 1421 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1422 "space_cap=" SIZE_FORMAT, 1423 space_live, space_used, 1424 space_capacity); 1425 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1426 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1427 density, min_percent_free, limiter, 1428 dead_wood_max, dead_wood_limit); 1429 } 1430 1431 // Locate the region with the desired amount of dead space to the left. 1432 const RegionData* const limit_cp = 1433 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1434 1435 // Scan from the first region with dead space to the limit region and find the 1436 // one with the best (largest) reclaimed ratio. 1437 double best_ratio = 0.0; 1438 const RegionData* best_cp = full_cp; 1439 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1440 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1441 if (tmp_ratio > best_ratio) { 1442 best_cp = cp; 1443 best_ratio = tmp_ratio; 1444 } 1445 } 1446 1447 return sd.region_to_addr(best_cp); 1448 } 1449 1450 void PSParallelCompact::summarize_spaces_quick() 1451 { 1452 for (unsigned int i = 0; i < last_space_id; ++i) { 1453 const MutableSpace* space = _space_info[i].space(); 1454 HeapWord** nta = _space_info[i].new_top_addr(); 1455 bool result = _summary_data.summarize(_space_info[i].split_info(), 1456 space->bottom(), space->top(), NULL, 1457 space->bottom(), space->end(), nta); 1458 assert(result, "space must fit into itself"); 1459 _space_info[i].set_dense_prefix(space->bottom()); 1460 } 1461 } 1462 1463 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1464 { 1465 HeapWord* const dense_prefix_end = dense_prefix(id); 1466 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1467 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1468 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1469 // Only enough dead space is filled so that any remaining dead space to the 1470 // left is larger than the minimum filler object. (The remainder is filled 1471 // during the copy/update phase.) 1472 // 1473 // The size of the dead space to the right of the boundary is not a 1474 // concern, since compaction will be able to use whatever space is 1475 // available. 1476 // 1477 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1478 // surrounds the space to be filled with an object. 1479 // 1480 // In the 32-bit VM, each bit represents two 32-bit words: 1481 // +---+ 1482 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1483 // end_bits: ... x x x | 0 | || 0 x x ... 1484 // +---+ 1485 // 1486 // In the 64-bit VM, each bit represents one 64-bit word: 1487 // +------------+ 1488 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1489 // end_bits: ... x x 1 | 0 || 0 | x x ... 1490 // +------------+ 1491 // +-------+ 1492 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1493 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1494 // +-------+ 1495 // +-----------+ 1496 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1497 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1498 // +-----------+ 1499 // +-------+ 1500 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1501 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1502 // +-------+ 1503 1504 // Initially assume case a, c or e will apply. 1505 size_t obj_len = CollectedHeap::min_fill_size(); 1506 HeapWord* obj_beg = dense_prefix_end - obj_len; 1507 1508 #ifdef _LP64 1509 if (MinObjAlignment > 1) { // object alignment > heap word size 1510 // Cases a, c or e. 1511 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1512 // Case b above. 1513 obj_beg = dense_prefix_end - 1; 1514 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1515 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1516 // Case d above. 1517 obj_beg = dense_prefix_end - 3; 1518 obj_len = 3; 1519 } 1520 #endif // #ifdef _LP64 1521 1522 CollectedHeap::fill_with_object(obj_beg, obj_len); 1523 _mark_bitmap.mark_obj(obj_beg, obj_len); 1524 _summary_data.add_obj(obj_beg, obj_len); 1525 assert(start_array(id) != NULL, "sanity"); 1526 start_array(id)->allocate_block(obj_beg); 1527 } 1528 } 1529 1530 void 1531 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1532 { 1533 assert(id < last_space_id, "id out of range"); 1534 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1535 "should have been reset in summarize_spaces_quick()"); 1536 1537 const MutableSpace* space = _space_info[id].space(); 1538 if (_space_info[id].new_top() != space->bottom()) { 1539 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1540 _space_info[id].set_dense_prefix(dense_prefix_end); 1541 1542 #ifndef PRODUCT 1543 if (TraceParallelOldGCDensePrefix) { 1544 print_dense_prefix_stats("ratio", id, maximum_compaction, 1545 dense_prefix_end); 1546 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1547 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1548 } 1549 #endif // #ifndef PRODUCT 1550 1551 // Recompute the summary data, taking into account the dense prefix. If 1552 // every last byte will be reclaimed, then the existing summary data which 1553 // compacts everything can be left in place. 1554 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1555 // If dead space crosses the dense prefix boundary, it is (at least 1556 // partially) filled with a dummy object, marked live and added to the 1557 // summary data. This simplifies the copy/update phase and must be done 1558 // before the final locations of objects are determined, to prevent 1559 // leaving a fragment of dead space that is too small to fill. 1560 fill_dense_prefix_end(id); 1561 1562 // Compute the destination of each Region, and thus each object. 1563 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1564 _summary_data.summarize(_space_info[id].split_info(), 1565 dense_prefix_end, space->top(), NULL, 1566 dense_prefix_end, space->end(), 1567 _space_info[id].new_top_addr()); 1568 } 1569 } 1570 1571 if (log_develop_is_enabled(Trace, gc, compaction)) { 1572 const size_t region_size = ParallelCompactData::RegionSize; 1573 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1574 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1575 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1576 HeapWord* const new_top = _space_info[id].new_top(); 1577 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1578 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1579 log_develop_trace(gc, compaction)( 1580 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1581 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1582 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1583 id, space->capacity_in_words(), p2i(dense_prefix_end), 1584 dp_region, dp_words / region_size, 1585 cr_words / region_size, p2i(new_top)); 1586 } 1587 } 1588 1589 #ifndef PRODUCT 1590 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1591 HeapWord* dst_beg, HeapWord* dst_end, 1592 SpaceId src_space_id, 1593 HeapWord* src_beg, HeapWord* src_end) 1594 { 1595 log_develop_trace(gc, compaction)( 1596 "Summarizing %d [%s] into %d [%s]: " 1597 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1598 SIZE_FORMAT "-" SIZE_FORMAT " " 1599 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1600 SIZE_FORMAT "-" SIZE_FORMAT, 1601 src_space_id, space_names[src_space_id], 1602 dst_space_id, space_names[dst_space_id], 1603 p2i(src_beg), p2i(src_end), 1604 _summary_data.addr_to_region_idx(src_beg), 1605 _summary_data.addr_to_region_idx(src_end), 1606 p2i(dst_beg), p2i(dst_end), 1607 _summary_data.addr_to_region_idx(dst_beg), 1608 _summary_data.addr_to_region_idx(dst_end)); 1609 } 1610 #endif // #ifndef PRODUCT 1611 1612 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1613 bool maximum_compaction) 1614 { 1615 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1616 1617 #ifdef ASSERT 1618 if (TraceParallelOldGCMarkingPhase) { 1619 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1620 "add_obj_bytes=" SIZE_FORMAT, 1621 add_obj_count, add_obj_size * HeapWordSize); 1622 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1623 "mark_bitmap_bytes=" SIZE_FORMAT, 1624 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1625 } 1626 #endif // #ifdef ASSERT 1627 1628 // Quick summarization of each space into itself, to see how much is live. 1629 summarize_spaces_quick(); 1630 1631 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1632 NOT_PRODUCT(print_region_ranges()); 1633 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1634 1635 // The amount of live data that will end up in old space (assuming it fits). 1636 size_t old_space_total_live = 0; 1637 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1638 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1639 _space_info[id].space()->bottom()); 1640 } 1641 1642 MutableSpace* const old_space = _space_info[old_space_id].space(); 1643 const size_t old_capacity = old_space->capacity_in_words(); 1644 if (old_space_total_live > old_capacity) { 1645 // XXX - should also try to expand 1646 maximum_compaction = true; 1647 } 1648 1649 // Old generations. 1650 summarize_space(old_space_id, maximum_compaction); 1651 1652 // Summarize the remaining spaces in the young gen. The initial target space 1653 // is the old gen. If a space does not fit entirely into the target, then the 1654 // remainder is compacted into the space itself and that space becomes the new 1655 // target. 1656 SpaceId dst_space_id = old_space_id; 1657 HeapWord* dst_space_end = old_space->end(); 1658 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1659 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1660 const MutableSpace* space = _space_info[id].space(); 1661 const size_t live = pointer_delta(_space_info[id].new_top(), 1662 space->bottom()); 1663 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1664 1665 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1666 SpaceId(id), space->bottom(), space->top());) 1667 if (live > 0 && live <= available) { 1668 // All the live data will fit. 1669 bool done = _summary_data.summarize(_space_info[id].split_info(), 1670 space->bottom(), space->top(), 1671 NULL, 1672 *new_top_addr, dst_space_end, 1673 new_top_addr); 1674 assert(done, "space must fit into old gen"); 1675 1676 // Reset the new_top value for the space. 1677 _space_info[id].set_new_top(space->bottom()); 1678 } else if (live > 0) { 1679 // Attempt to fit part of the source space into the target space. 1680 HeapWord* next_src_addr = NULL; 1681 bool done = _summary_data.summarize(_space_info[id].split_info(), 1682 space->bottom(), space->top(), 1683 &next_src_addr, 1684 *new_top_addr, dst_space_end, 1685 new_top_addr); 1686 assert(!done, "space should not fit into old gen"); 1687 assert(next_src_addr != NULL, "sanity"); 1688 1689 // The source space becomes the new target, so the remainder is compacted 1690 // within the space itself. 1691 dst_space_id = SpaceId(id); 1692 dst_space_end = space->end(); 1693 new_top_addr = _space_info[id].new_top_addr(); 1694 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1695 space->bottom(), dst_space_end, 1696 SpaceId(id), next_src_addr, space->top());) 1697 done = _summary_data.summarize(_space_info[id].split_info(), 1698 next_src_addr, space->top(), 1699 NULL, 1700 space->bottom(), dst_space_end, 1701 new_top_addr); 1702 assert(done, "space must fit when compacted into itself"); 1703 assert(*new_top_addr <= space->top(), "usage should not grow"); 1704 } 1705 } 1706 1707 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1708 NOT_PRODUCT(print_region_ranges()); 1709 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1710 } 1711 1712 // This method should contain all heap-specific policy for invoking a full 1713 // collection. invoke_no_policy() will only attempt to compact the heap; it 1714 // will do nothing further. If we need to bail out for policy reasons, scavenge 1715 // before full gc, or any other specialized behavior, it needs to be added here. 1716 // 1717 // Note that this method should only be called from the vm_thread while at a 1718 // safepoint. 1719 // 1720 // Note that the all_soft_refs_clear flag in the soft ref policy 1721 // may be true because this method can be called without intervening 1722 // activity. For example when the heap space is tight and full measure 1723 // are being taken to free space. 1724 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1725 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1726 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1727 "should be in vm thread"); 1728 1729 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1730 GCCause::Cause gc_cause = heap->gc_cause(); 1731 assert(!heap->is_gc_active(), "not reentrant"); 1732 1733 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1734 IsGCActiveMark mark; 1735 1736 if (ScavengeBeforeFullGC) { 1737 PSScavenge::invoke_no_policy(); 1738 } 1739 1740 const bool clear_all_soft_refs = 1741 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1742 1743 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1744 maximum_heap_compaction); 1745 } 1746 1747 // This method contains no policy. You should probably 1748 // be calling invoke() instead. 1749 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1750 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1751 assert(ref_processor() != NULL, "Sanity"); 1752 1753 if (GCLocker::check_active_before_gc()) { 1754 return false; 1755 } 1756 1757 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1758 1759 GCIdMark gc_id_mark; 1760 _gc_timer.register_gc_start(); 1761 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1762 1763 TimeStamp marking_start; 1764 TimeStamp compaction_start; 1765 TimeStamp collection_exit; 1766 1767 GCCause::Cause gc_cause = heap->gc_cause(); 1768 PSYoungGen* young_gen = heap->young_gen(); 1769 PSOldGen* old_gen = heap->old_gen(); 1770 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1771 1772 // The scope of casr should end after code that can change 1773 // SoftRefPolicy::_should_clear_all_soft_refs. 1774 ClearedAllSoftRefs casr(maximum_heap_compaction, 1775 heap->soft_ref_policy()); 1776 1777 if (ZapUnusedHeapArea) { 1778 // Save information needed to minimize mangling 1779 heap->record_gen_tops_before_GC(); 1780 } 1781 1782 // Make sure data structures are sane, make the heap parsable, and do other 1783 // miscellaneous bookkeeping. 1784 pre_compact(); 1785 1786 const PreGenGCValues pre_gc_values = heap->get_pre_gc_values(); 1787 1788 // Get the compaction manager reserved for the VM thread. 1789 ParCompactionManager* const vmthread_cm = 1790 ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers()); 1791 1792 { 1793 ResourceMark rm; 1794 HandleMark hm; 1795 1796 const uint active_workers = 1797 WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(), 1798 ParallelScavengeHeap::heap()->workers().active_workers(), 1799 Threads::number_of_non_daemon_threads()); 1800 ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers); 1801 1802 GCTraceCPUTime tcpu; 1803 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1804 1805 heap->pre_full_gc_dump(&_gc_timer); 1806 1807 TraceCollectorStats tcs(counters()); 1808 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1809 1810 if (log_is_enabled(Debug, gc, heap, exit)) { 1811 accumulated_time()->start(); 1812 } 1813 1814 // Let the size policy know we're starting 1815 size_policy->major_collection_begin(); 1816 1817 #if COMPILER2_OR_JVMCI 1818 DerivedPointerTable::clear(); 1819 #endif 1820 1821 ref_processor()->enable_discovery(); 1822 ref_processor()->setup_policy(maximum_heap_compaction); 1823 1824 bool marked_for_unloading = false; 1825 1826 marking_start.update(); 1827 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1828 1829 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1830 && GCCause::is_user_requested_gc(gc_cause); 1831 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1832 1833 #if COMPILER2_OR_JVMCI 1834 assert(DerivedPointerTable::is_active(), "Sanity"); 1835 DerivedPointerTable::set_active(false); 1836 #endif 1837 1838 // adjust_roots() updates Universe::_intArrayKlassObj which is 1839 // needed by the compaction for filling holes in the dense prefix. 1840 adjust_roots(vmthread_cm); 1841 1842 compaction_start.update(); 1843 compact(); 1844 1845 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1846 // done before resizing. 1847 post_compact(); 1848 1849 // Let the size policy know we're done 1850 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1851 1852 if (UseAdaptiveSizePolicy) { 1853 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1854 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1855 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1856 1857 // Don't check if the size_policy is ready here. Let 1858 // the size_policy check that internally. 1859 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1860 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1861 // Swap the survivor spaces if from_space is empty. The 1862 // resize_young_gen() called below is normally used after 1863 // a successful young GC and swapping of survivor spaces; 1864 // otherwise, it will fail to resize the young gen with 1865 // the current implementation. 1866 if (young_gen->from_space()->is_empty()) { 1867 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1868 young_gen->swap_spaces(); 1869 } 1870 1871 // Calculate optimal free space amounts 1872 assert(young_gen->max_size() > 1873 young_gen->from_space()->capacity_in_bytes() + 1874 young_gen->to_space()->capacity_in_bytes(), 1875 "Sizes of space in young gen are out-of-bounds"); 1876 1877 size_t young_live = young_gen->used_in_bytes(); 1878 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1879 size_t old_live = old_gen->used_in_bytes(); 1880 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1881 size_t max_old_gen_size = old_gen->max_gen_size(); 1882 size_t max_eden_size = young_gen->max_size() - 1883 young_gen->from_space()->capacity_in_bytes() - 1884 young_gen->to_space()->capacity_in_bytes(); 1885 1886 // Used for diagnostics 1887 size_policy->clear_generation_free_space_flags(); 1888 1889 size_policy->compute_generations_free_space(young_live, 1890 eden_live, 1891 old_live, 1892 cur_eden, 1893 max_old_gen_size, 1894 max_eden_size, 1895 true /* full gc*/); 1896 1897 size_policy->check_gc_overhead_limit(eden_live, 1898 max_old_gen_size, 1899 max_eden_size, 1900 true /* full gc*/, 1901 gc_cause, 1902 heap->soft_ref_policy()); 1903 1904 size_policy->decay_supplemental_growth(true /* full gc*/); 1905 1906 heap->resize_old_gen( 1907 size_policy->calculated_old_free_size_in_bytes()); 1908 1909 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1910 size_policy->calculated_survivor_size_in_bytes()); 1911 } 1912 1913 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1914 } 1915 1916 if (UsePerfData) { 1917 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1918 counters->update_counters(); 1919 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1920 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1921 } 1922 1923 heap->resize_all_tlabs(); 1924 1925 // Resize the metaspace capacity after a collection 1926 MetaspaceGC::compute_new_size(); 1927 1928 if (log_is_enabled(Debug, gc, heap, exit)) { 1929 accumulated_time()->stop(); 1930 } 1931 1932 heap->print_heap_change(pre_gc_values); 1933 1934 // Track memory usage and detect low memory 1935 MemoryService::track_memory_usage(); 1936 heap->update_counters(); 1937 1938 heap->post_full_gc_dump(&_gc_timer); 1939 } 1940 1941 #ifdef ASSERT 1942 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1943 ParCompactionManager* const cm = 1944 ParCompactionManager::manager_array(int(i)); 1945 assert(cm->marking_stack()->is_empty(), "should be empty"); 1946 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1947 } 1948 #endif // ASSERT 1949 1950 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1951 HandleMark hm; // Discard invalid handles created during verification 1952 Universe::verify("After GC"); 1953 } 1954 1955 // Re-verify object start arrays 1956 if (VerifyObjectStartArray && 1957 VerifyAfterGC) { 1958 old_gen->verify_object_start_array(); 1959 } 1960 1961 if (ZapUnusedHeapArea) { 1962 old_gen->object_space()->check_mangled_unused_area_complete(); 1963 } 1964 1965 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1966 1967 collection_exit.update(); 1968 1969 heap->print_heap_after_gc(); 1970 heap->trace_heap_after_gc(&_gc_tracer); 1971 1972 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1973 marking_start.ticks(), compaction_start.ticks(), 1974 collection_exit.ticks()); 1975 1976 #ifdef TRACESPINNING 1977 ParallelTaskTerminator::print_termination_counts(); 1978 #endif 1979 1980 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1981 1982 _gc_timer.register_gc_end(); 1983 1984 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1985 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1986 1987 return true; 1988 } 1989 1990 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1991 PSYoungGen* young_gen, 1992 PSOldGen* old_gen) { 1993 MutableSpace* const eden_space = young_gen->eden_space(); 1994 assert(!eden_space->is_empty(), "eden must be non-empty"); 1995 assert(young_gen->virtual_space()->alignment() == 1996 old_gen->virtual_space()->alignment(), "alignments do not match"); 1997 1998 // We also return false when it's a heterogeneous heap because old generation cannot absorb data from eden 1999 // when it is allocated on different memory (example, nv-dimm) than young. 2000 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) || 2001 ParallelArguments::is_heterogeneous_heap()) { 2002 return false; 2003 } 2004 2005 // Both generations must be completely committed. 2006 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2007 return false; 2008 } 2009 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2010 return false; 2011 } 2012 2013 // Figure out how much to take from eden. Include the average amount promoted 2014 // in the total; otherwise the next young gen GC will simply bail out to a 2015 // full GC. 2016 const size_t alignment = old_gen->virtual_space()->alignment(); 2017 const size_t eden_used = eden_space->used_in_bytes(); 2018 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2019 const size_t absorb_size = align_up(eden_used + promoted, alignment); 2020 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2021 2022 if (absorb_size >= eden_capacity) { 2023 return false; // Must leave some space in eden. 2024 } 2025 2026 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2027 if (new_young_size < young_gen->min_gen_size()) { 2028 return false; // Respect young gen minimum size. 2029 } 2030 2031 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: " 2032 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2033 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2034 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2035 absorb_size / K, 2036 eden_capacity / K, (eden_capacity - absorb_size) / K, 2037 young_gen->from_space()->used_in_bytes() / K, 2038 young_gen->to_space()->used_in_bytes() / K, 2039 young_gen->capacity_in_bytes() / K, new_young_size / K); 2040 2041 // Fill the unused part of the old gen. 2042 MutableSpace* const old_space = old_gen->object_space(); 2043 HeapWord* const unused_start = old_space->top(); 2044 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2045 2046 if (unused_words > 0) { 2047 if (unused_words < CollectedHeap::min_fill_size()) { 2048 return false; // If the old gen cannot be filled, must give up. 2049 } 2050 CollectedHeap::fill_with_objects(unused_start, unused_words); 2051 } 2052 2053 // Take the live data from eden and set both top and end in the old gen to 2054 // eden top. (Need to set end because reset_after_change() mangles the region 2055 // from end to virtual_space->high() in debug builds). 2056 HeapWord* const new_top = eden_space->top(); 2057 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2058 absorb_size); 2059 young_gen->reset_after_change(); 2060 old_space->set_top(new_top); 2061 old_space->set_end(new_top); 2062 old_gen->reset_after_change(); 2063 2064 // Update the object start array for the filler object and the data from eden. 2065 ObjectStartArray* const start_array = old_gen->start_array(); 2066 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2067 start_array->allocate_block(p); 2068 } 2069 2070 // Could update the promoted average here, but it is not typically updated at 2071 // full GCs and the value to use is unclear. Something like 2072 // 2073 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2074 2075 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2076 return true; 2077 } 2078 2079 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 2080 private: 2081 uint _worker_id; 2082 2083 public: 2084 PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { } 2085 void do_thread(Thread* thread) { 2086 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2087 2088 ResourceMark rm; 2089 2090 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id); 2091 2092 PCMarkAndPushClosure mark_and_push_closure(cm); 2093 MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations); 2094 2095 thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs); 2096 2097 // Do the real work 2098 cm->follow_marking_stacks(); 2099 } 2100 }; 2101 2102 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) { 2103 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2104 2105 ParCompactionManager* cm = 2106 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2107 PCMarkAndPushClosure mark_and_push_closure(cm); 2108 2109 switch (root_type) { 2110 case ParallelRootType::universe: 2111 Universe::oops_do(&mark_and_push_closure); 2112 break; 2113 2114 case ParallelRootType::jni_handles: 2115 JNIHandles::oops_do(&mark_and_push_closure); 2116 break; 2117 2118 case ParallelRootType::object_synchronizer: 2119 ObjectSynchronizer::oops_do(&mark_and_push_closure); 2120 break; 2121 2122 case ParallelRootType::management: 2123 Management::oops_do(&mark_and_push_closure); 2124 break; 2125 2126 case ParallelRootType::jvmti: 2127 JvmtiExport::oops_do(&mark_and_push_closure); 2128 break; 2129 2130 case ParallelRootType::system_dictionary: 2131 SystemDictionary::oops_do(&mark_and_push_closure); 2132 break; 2133 2134 case ParallelRootType::class_loader_data: 2135 { 2136 CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong); 2137 ClassLoaderDataGraph::always_strong_cld_do(&cld_closure); 2138 } 2139 break; 2140 2141 case ParallelRootType::code_cache: 2142 // Do not treat nmethods as strong roots for mark/sweep, since we can unload them. 2143 //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure)); 2144 AOTLoader::oops_do(&mark_and_push_closure); 2145 break; 2146 2147 case ParallelRootType::sentinel: 2148 DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds 2149 fatal("Bad enumeration value: %u", root_type); 2150 break; 2151 } 2152 2153 // Do the real work 2154 cm->follow_marking_stacks(); 2155 } 2156 2157 static void steal_marking_work(ParallelTaskTerminator& terminator, uint worker_id) { 2158 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2159 2160 ParCompactionManager* cm = 2161 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2162 2163 oop obj = NULL; 2164 ObjArrayTask task; 2165 do { 2166 while (ParCompactionManager::steal_objarray(worker_id, task)) { 2167 cm->follow_array((objArrayOop)task.obj(), task.index()); 2168 cm->follow_marking_stacks(); 2169 } 2170 while (ParCompactionManager::steal(worker_id, obj)) { 2171 cm->follow_contents(obj); 2172 cm->follow_marking_stacks(); 2173 } 2174 } while (!terminator.offer_termination()); 2175 } 2176 2177 class MarkFromRootsTask : public AbstractGangTask { 2178 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2179 StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do 2180 SequentialSubTasksDone _subtasks; 2181 TaskTerminator _terminator; 2182 uint _active_workers; 2183 2184 public: 2185 MarkFromRootsTask(uint active_workers) : 2186 AbstractGangTask("MarkFromRootsTask"), 2187 _strong_roots_scope(active_workers), 2188 _subtasks(), 2189 _terminator(active_workers, ParCompactionManager::stack_array()), 2190 _active_workers(active_workers) { 2191 _subtasks.set_n_threads(active_workers); 2192 _subtasks.set_n_tasks(ParallelRootType::sentinel); 2193 } 2194 2195 virtual void work(uint worker_id) { 2196 for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) { 2197 mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id); 2198 } 2199 _subtasks.all_tasks_completed(); 2200 2201 PCAddThreadRootsMarkingTaskClosure closure(worker_id); 2202 Threads::possibly_parallel_threads_do(true /*parallel */, &closure); 2203 2204 if (_active_workers > 1) { 2205 steal_marking_work(*_terminator.terminator(), worker_id); 2206 } 2207 } 2208 }; 2209 2210 class PCRefProcTask : public AbstractGangTask { 2211 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2212 ProcessTask& _task; 2213 uint _ergo_workers; 2214 TaskTerminator _terminator; 2215 2216 public: 2217 PCRefProcTask(ProcessTask& task, uint ergo_workers) : 2218 AbstractGangTask("PCRefProcTask"), 2219 _task(task), 2220 _ergo_workers(ergo_workers), 2221 _terminator(_ergo_workers, ParCompactionManager::stack_array()) { 2222 } 2223 2224 virtual void work(uint worker_id) { 2225 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2226 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2227 2228 ParCompactionManager* cm = 2229 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2230 PCMarkAndPushClosure mark_and_push_closure(cm); 2231 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2232 _task.work(worker_id, *PSParallelCompact::is_alive_closure(), 2233 mark_and_push_closure, follow_stack_closure); 2234 2235 steal_marking_work(*_terminator.terminator(), worker_id); 2236 } 2237 }; 2238 2239 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2240 void execute(ProcessTask& process_task, uint ergo_workers) { 2241 assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers, 2242 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 2243 ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers()); 2244 2245 PCRefProcTask task(process_task, ergo_workers); 2246 ParallelScavengeHeap::heap()->workers().run_task(&task); 2247 } 2248 }; 2249 2250 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2251 bool maximum_heap_compaction, 2252 ParallelOldTracer *gc_tracer) { 2253 // Recursively traverse all live objects and mark them 2254 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2255 2256 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2257 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2258 2259 PCMarkAndPushClosure mark_and_push_closure(cm); 2260 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2261 2262 // Need new claim bits before marking starts. 2263 ClassLoaderDataGraph::clear_claimed_marks(); 2264 2265 { 2266 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2267 2268 MarkFromRootsTask task(active_gc_threads); 2269 ParallelScavengeHeap::heap()->workers().run_task(&task); 2270 } 2271 2272 // Process reference objects found during marking 2273 { 2274 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2275 2276 ReferenceProcessorStats stats; 2277 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2278 2279 if (ref_processor()->processing_is_mt()) { 2280 ref_processor()->set_active_mt_degree(active_gc_threads); 2281 2282 RefProcTaskExecutor task_executor; 2283 stats = ref_processor()->process_discovered_references( 2284 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2285 &task_executor, &pt); 2286 } else { 2287 stats = ref_processor()->process_discovered_references( 2288 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2289 &pt); 2290 } 2291 2292 gc_tracer->report_gc_reference_stats(stats); 2293 pt.print_all_references(); 2294 } 2295 2296 // This is the point where the entire marking should have completed. 2297 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2298 2299 { 2300 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2301 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2302 } 2303 2304 { 2305 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2306 2307 // Follow system dictionary roots and unload classes. 2308 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2309 2310 // Unload nmethods. 2311 CodeCache::do_unloading(is_alive_closure(), purged_class); 2312 2313 // Prune dead klasses from subklass/sibling/implementor lists. 2314 Klass::clean_weak_klass_links(purged_class); 2315 2316 // Clean JVMCI metadata handles. 2317 JVMCI_ONLY(JVMCI::do_unloading(purged_class)); 2318 } 2319 2320 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2321 } 2322 2323 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2324 // Adjust the pointers to reflect the new locations 2325 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2326 2327 // Need new claim bits when tracing through and adjusting pointers. 2328 ClassLoaderDataGraph::clear_claimed_marks(); 2329 2330 PCAdjustPointerClosure oop_closure(cm); 2331 2332 // General strong roots. 2333 Universe::oops_do(&oop_closure); 2334 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2335 Threads::oops_do(&oop_closure, NULL); 2336 ObjectSynchronizer::oops_do(&oop_closure); 2337 Management::oops_do(&oop_closure); 2338 JvmtiExport::oops_do(&oop_closure); 2339 SystemDictionary::oops_do(&oop_closure); 2340 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong); 2341 ClassLoaderDataGraph::cld_do(&cld_closure); 2342 2343 // Now adjust pointers in remaining weak roots. (All of which should 2344 // have been cleared if they pointed to non-surviving objects.) 2345 WeakProcessor::oops_do(&oop_closure); 2346 2347 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2348 CodeCache::blobs_do(&adjust_from_blobs); 2349 AOT_ONLY(AOTLoader::oops_do(&oop_closure);) 2350 2351 ref_processor()->weak_oops_do(&oop_closure); 2352 // Roots were visited so references into the young gen in roots 2353 // may have been scanned. Process them also. 2354 // Should the reference processor have a span that excludes 2355 // young gen objects? 2356 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2357 } 2358 2359 // Helper class to print 8 region numbers per line and then print the total at the end. 2360 class FillableRegionLogger : public StackObj { 2361 private: 2362 Log(gc, compaction) log; 2363 static const int LineLength = 8; 2364 size_t _regions[LineLength]; 2365 int _next_index; 2366 bool _enabled; 2367 size_t _total_regions; 2368 public: 2369 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2370 ~FillableRegionLogger() { 2371 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2372 } 2373 2374 void print_line() { 2375 if (!_enabled || _next_index == 0) { 2376 return; 2377 } 2378 FormatBuffer<> line("Fillable: "); 2379 for (int i = 0; i < _next_index; i++) { 2380 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2381 } 2382 log.trace("%s", line.buffer()); 2383 _next_index = 0; 2384 } 2385 2386 void handle(size_t region) { 2387 if (!_enabled) { 2388 return; 2389 } 2390 _regions[_next_index++] = region; 2391 if (_next_index == LineLength) { 2392 print_line(); 2393 } 2394 _total_regions++; 2395 } 2396 }; 2397 2398 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads) 2399 { 2400 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2401 2402 // Find the threads that are active 2403 uint worker_id = 0; 2404 2405 // Find all regions that are available (can be filled immediately) and 2406 // distribute them to the thread stacks. The iteration is done in reverse 2407 // order (high to low) so the regions will be removed in ascending order. 2408 2409 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2410 2411 // id + 1 is used to test termination so unsigned can 2412 // be used with an old_space_id == 0. 2413 FillableRegionLogger region_logger; 2414 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2415 SpaceInfo* const space_info = _space_info + id; 2416 MutableSpace* const space = space_info->space(); 2417 HeapWord* const new_top = space_info->new_top(); 2418 2419 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2420 const size_t end_region = 2421 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2422 2423 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2424 if (sd.region(cur)->claim_unsafe()) { 2425 ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id); 2426 if (sd.region(cur)->try_push()) { 2427 cm->region_stack()->push(cur); 2428 region_logger.handle(cur); 2429 } 2430 // Assign regions to tasks in round-robin fashion. 2431 if (++worker_id == parallel_gc_threads) { 2432 worker_id = 0; 2433 } 2434 } 2435 } 2436 region_logger.print_line(); 2437 } 2438 } 2439 2440 class TaskQueue : StackObj { 2441 volatile uint _counter; 2442 uint _size; 2443 uint _insert_index; 2444 PSParallelCompact::UpdateDensePrefixTask* _backing_array; 2445 public: 2446 explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) { 2447 _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC); 2448 } 2449 ~TaskQueue() { 2450 assert(_counter >= _insert_index, "not all queue elements were claimed"); 2451 FREE_C_HEAP_ARRAY(T, _backing_array); 2452 } 2453 2454 void push(const PSParallelCompact::UpdateDensePrefixTask& value) { 2455 assert(_insert_index < _size, "too small backing array"); 2456 _backing_array[_insert_index++] = value; 2457 } 2458 2459 bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) { 2460 uint claimed = Atomic::add(1u, &_counter) - 1; // -1 is so that we start with zero 2461 if (claimed < _insert_index) { 2462 reference = _backing_array[claimed]; 2463 return true; 2464 } else { 2465 return false; 2466 } 2467 } 2468 }; 2469 2470 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2471 2472 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue, 2473 uint parallel_gc_threads) { 2474 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2475 2476 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2477 2478 // Iterate over all the spaces adding tasks for updating 2479 // regions in the dense prefix. Assume that 1 gc thread 2480 // will work on opening the gaps and the remaining gc threads 2481 // will work on the dense prefix. 2482 unsigned int space_id; 2483 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2484 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2485 const MutableSpace* const space = _space_info[space_id].space(); 2486 2487 if (dense_prefix_end == space->bottom()) { 2488 // There is no dense prefix for this space. 2489 continue; 2490 } 2491 2492 // The dense prefix is before this region. 2493 size_t region_index_end_dense_prefix = 2494 sd.addr_to_region_idx(dense_prefix_end); 2495 RegionData* const dense_prefix_cp = 2496 sd.region(region_index_end_dense_prefix); 2497 assert(dense_prefix_end == space->end() || 2498 dense_prefix_cp->available() || 2499 dense_prefix_cp->claimed(), 2500 "The region after the dense prefix should always be ready to fill"); 2501 2502 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2503 2504 // Is there dense prefix work? 2505 size_t total_dense_prefix_regions = 2506 region_index_end_dense_prefix - region_index_start; 2507 // How many regions of the dense prefix should be given to 2508 // each thread? 2509 if (total_dense_prefix_regions > 0) { 2510 uint tasks_for_dense_prefix = 1; 2511 if (total_dense_prefix_regions <= 2512 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2513 // Don't over partition. This assumes that 2514 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2515 // so there are not many regions to process. 2516 tasks_for_dense_prefix = parallel_gc_threads; 2517 } else { 2518 // Over partition 2519 tasks_for_dense_prefix = parallel_gc_threads * 2520 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2521 } 2522 size_t regions_per_thread = total_dense_prefix_regions / 2523 tasks_for_dense_prefix; 2524 // Give each thread at least 1 region. 2525 if (regions_per_thread == 0) { 2526 regions_per_thread = 1; 2527 } 2528 2529 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2530 if (region_index_start >= region_index_end_dense_prefix) { 2531 break; 2532 } 2533 // region_index_end is not processed 2534 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2535 region_index_end_dense_prefix); 2536 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2537 region_index_start, 2538 region_index_end)); 2539 region_index_start = region_index_end; 2540 } 2541 } 2542 // This gets any part of the dense prefix that did not 2543 // fit evenly. 2544 if (region_index_start < region_index_end_dense_prefix) { 2545 task_queue.push(UpdateDensePrefixTask(SpaceId(space_id), 2546 region_index_start, 2547 region_index_end_dense_prefix)); 2548 } 2549 } 2550 } 2551 2552 #ifdef ASSERT 2553 // Write a histogram of the number of times the block table was filled for a 2554 // region. 2555 void PSParallelCompact::write_block_fill_histogram() 2556 { 2557 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2558 return; 2559 } 2560 2561 Log(gc, compaction) log; 2562 ResourceMark rm; 2563 LogStream ls(log.trace()); 2564 outputStream* out = &ls; 2565 2566 typedef ParallelCompactData::RegionData rd_t; 2567 ParallelCompactData& sd = summary_data(); 2568 2569 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2570 MutableSpace* const spc = _space_info[id].space(); 2571 if (spc->bottom() != spc->top()) { 2572 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2573 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2574 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2575 2576 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2577 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2578 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2579 2580 for (const rd_t* cur = beg; cur < end; ++cur) { 2581 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2582 } 2583 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2584 for (size_t i = 0; i < histo_len; ++i) { 2585 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2586 histo[i], 100.0 * histo[i] / region_cnt); 2587 } 2588 out->cr(); 2589 } 2590 } 2591 } 2592 #endif // #ifdef ASSERT 2593 2594 static void compaction_with_stealing_work(ParallelTaskTerminator* terminator, uint worker_id) { 2595 assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc"); 2596 2597 ParCompactionManager* cm = 2598 ParCompactionManager::gc_thread_compaction_manager(worker_id); 2599 2600 // Drain the stacks that have been preloaded with regions 2601 // that are ready to fill. 2602 2603 cm->drain_region_stacks(); 2604 2605 guarantee(cm->region_stack()->is_empty(), "Not empty"); 2606 2607 size_t region_index = 0; 2608 2609 PSParallelCompact::initialize_steal_record(worker_id); 2610 while (true) { 2611 if (ParCompactionManager::steal(worker_id, region_index)) { 2612 PSParallelCompact::fill_and_update_region(cm, region_index); 2613 cm->drain_region_stacks(); 2614 } else if (PSParallelCompact::steal_shadow_region(cm, region_index)) { 2615 // Keep working with the help of shadow regions 2616 PSParallelCompact::fill_and_update_shadow_region(cm, region_index); 2617 cm->drain_region_stacks(); 2618 } else { 2619 if (terminator->offer_termination()) { 2620 break; 2621 } 2622 // Go around again. 2623 } 2624 } 2625 return; 2626 } 2627 2628 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask { 2629 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2630 TaskQueue& _tq; 2631 TaskTerminator _terminator; 2632 uint _active_workers; 2633 2634 public: 2635 UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) : 2636 AbstractGangTask("UpdateDensePrefixAndCompactionTask"), 2637 _tq(tq), 2638 _terminator(active_workers, ParCompactionManager::region_array()), 2639 _active_workers(active_workers) { 2640 } 2641 virtual void work(uint worker_id) { 2642 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id); 2643 2644 for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) { 2645 PSParallelCompact::update_and_deadwood_in_dense_prefix(cm, 2646 task._space_id, 2647 task._region_index_start, 2648 task._region_index_end); 2649 } 2650 2651 // Once a thread has drained it's stack, it should try to steal regions from 2652 // other threads. 2653 compaction_with_stealing_work(_terminator.terminator(), worker_id); 2654 } 2655 }; 2656 2657 void PSParallelCompact::compact() { 2658 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2659 2660 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2661 PSOldGen* old_gen = heap->old_gen(); 2662 old_gen->start_array()->reset(); 2663 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 2664 2665 // for [0..last_space_id) 2666 // for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING) 2667 // push 2668 // push 2669 // 2670 // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1) 2671 TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)); 2672 enqueue_shadow_region(); 2673 prepare_region_draining_tasks(active_gc_threads); 2674 enqueue_dense_prefix_tasks(task_queue, active_gc_threads); 2675 2676 { 2677 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2678 2679 UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads); 2680 ParallelScavengeHeap::heap()->workers().run_task(&task); 2681 2682 #ifdef ASSERT 2683 // Verify that all regions have been processed before the deferred updates. 2684 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2685 verify_complete(SpaceId(id)); 2686 } 2687 #endif 2688 } 2689 2690 { 2691 // Update the deferred objects, if any. Any compaction manager can be used. 2692 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2693 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2694 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2695 update_deferred_objects(cm, SpaceId(id)); 2696 } 2697 } 2698 2699 DEBUG_ONLY(write_block_fill_histogram()); 2700 } 2701 2702 #ifdef ASSERT 2703 void PSParallelCompact::verify_complete(SpaceId space_id) { 2704 // All Regions between space bottom() to new_top() should be marked as filled 2705 // and all Regions between new_top() and top() should be available (i.e., 2706 // should have been emptied). 2707 ParallelCompactData& sd = summary_data(); 2708 SpaceInfo si = _space_info[space_id]; 2709 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2710 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2711 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2712 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2713 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2714 2715 bool issued_a_warning = false; 2716 2717 size_t cur_region; 2718 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2719 const RegionData* const c = sd.region(cur_region); 2720 if (!c->completed()) { 2721 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2722 cur_region, c->destination_count()); 2723 issued_a_warning = true; 2724 } 2725 } 2726 2727 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2728 const RegionData* const c = sd.region(cur_region); 2729 if (!c->available()) { 2730 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2731 cur_region, c->destination_count()); 2732 issued_a_warning = true; 2733 } 2734 } 2735 2736 if (issued_a_warning) { 2737 print_region_ranges(); 2738 } 2739 } 2740 #endif // #ifdef ASSERT 2741 2742 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2743 _start_array->allocate_block(addr); 2744 compaction_manager()->update_contents(oop(addr)); 2745 } 2746 2747 // Update interior oops in the ranges of regions [beg_region, end_region). 2748 void 2749 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2750 SpaceId space_id, 2751 size_t beg_region, 2752 size_t end_region) { 2753 ParallelCompactData& sd = summary_data(); 2754 ParMarkBitMap* const mbm = mark_bitmap(); 2755 2756 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2757 HeapWord* const end_addr = sd.region_to_addr(end_region); 2758 assert(beg_region <= end_region, "bad region range"); 2759 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2760 2761 #ifdef ASSERT 2762 // Claim the regions to avoid triggering an assert when they are marked as 2763 // filled. 2764 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2765 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2766 } 2767 #endif // #ifdef ASSERT 2768 2769 if (beg_addr != space(space_id)->bottom()) { 2770 // Find the first live object or block of dead space that *starts* in this 2771 // range of regions. If a partial object crosses onto the region, skip it; 2772 // it will be marked for 'deferred update' when the object head is 2773 // processed. If dead space crosses onto the region, it is also skipped; it 2774 // will be filled when the prior region is processed. If neither of those 2775 // apply, the first word in the region is the start of a live object or dead 2776 // space. 2777 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2778 const RegionData* const cp = sd.region(beg_region); 2779 if (cp->partial_obj_size() != 0) { 2780 beg_addr = sd.partial_obj_end(beg_region); 2781 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2782 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2783 } 2784 } 2785 2786 if (beg_addr < end_addr) { 2787 // A live object or block of dead space starts in this range of Regions. 2788 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2789 2790 // Create closures and iterate. 2791 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2792 FillClosure fill_closure(cm, space_id); 2793 ParMarkBitMap::IterationStatus status; 2794 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2795 dense_prefix_end); 2796 if (status == ParMarkBitMap::incomplete) { 2797 update_closure.do_addr(update_closure.source()); 2798 } 2799 } 2800 2801 // Mark the regions as filled. 2802 RegionData* const beg_cp = sd.region(beg_region); 2803 RegionData* const end_cp = sd.region(end_region); 2804 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2805 cp->set_completed(); 2806 } 2807 } 2808 2809 // Return the SpaceId for the space containing addr. If addr is not in the 2810 // heap, last_space_id is returned. In debug mode it expects the address to be 2811 // in the heap and asserts such. 2812 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2813 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2814 2815 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2816 if (_space_info[id].space()->contains(addr)) { 2817 return SpaceId(id); 2818 } 2819 } 2820 2821 assert(false, "no space contains the addr"); 2822 return last_space_id; 2823 } 2824 2825 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2826 SpaceId id) { 2827 assert(id < last_space_id, "bad space id"); 2828 2829 ParallelCompactData& sd = summary_data(); 2830 const SpaceInfo* const space_info = _space_info + id; 2831 ObjectStartArray* const start_array = space_info->start_array(); 2832 2833 const MutableSpace* const space = space_info->space(); 2834 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2835 HeapWord* const beg_addr = space_info->dense_prefix(); 2836 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2837 2838 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2839 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2840 const RegionData* cur_region; 2841 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2842 HeapWord* const addr = cur_region->deferred_obj_addr(); 2843 if (addr != NULL) { 2844 if (start_array != NULL) { 2845 start_array->allocate_block(addr); 2846 } 2847 cm->update_contents(oop(addr)); 2848 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2849 } 2850 } 2851 } 2852 2853 // Skip over count live words starting from beg, and return the address of the 2854 // next live word. Unless marked, the word corresponding to beg is assumed to 2855 // be dead. Callers must either ensure beg does not correspond to the middle of 2856 // an object, or account for those live words in some other way. Callers must 2857 // also ensure that there are enough live words in the range [beg, end) to skip. 2858 HeapWord* 2859 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2860 { 2861 assert(count > 0, "sanity"); 2862 2863 ParMarkBitMap* m = mark_bitmap(); 2864 idx_t bits_to_skip = m->words_to_bits(count); 2865 idx_t cur_beg = m->addr_to_bit(beg); 2866 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2867 2868 do { 2869 cur_beg = m->find_obj_beg(cur_beg, search_end); 2870 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2871 const size_t obj_bits = cur_end - cur_beg + 1; 2872 if (obj_bits > bits_to_skip) { 2873 return m->bit_to_addr(cur_beg + bits_to_skip); 2874 } 2875 bits_to_skip -= obj_bits; 2876 cur_beg = cur_end + 1; 2877 } while (bits_to_skip > 0); 2878 2879 // Skipping the desired number of words landed just past the end of an object. 2880 // Find the start of the next object. 2881 cur_beg = m->find_obj_beg(cur_beg, search_end); 2882 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2883 return m->bit_to_addr(cur_beg); 2884 } 2885 2886 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2887 SpaceId src_space_id, 2888 size_t src_region_idx) 2889 { 2890 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2891 2892 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2893 if (split_info.dest_region_addr() == dest_addr) { 2894 // The partial object ending at the split point contains the first word to 2895 // be copied to dest_addr. 2896 return split_info.first_src_addr(); 2897 } 2898 2899 const ParallelCompactData& sd = summary_data(); 2900 ParMarkBitMap* const bitmap = mark_bitmap(); 2901 const size_t RegionSize = ParallelCompactData::RegionSize; 2902 2903 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2904 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2905 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2906 HeapWord* const src_region_destination = src_region_ptr->destination(); 2907 2908 assert(dest_addr >= src_region_destination, "wrong src region"); 2909 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2910 2911 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2912 HeapWord* const src_region_end = src_region_beg + RegionSize; 2913 2914 HeapWord* addr = src_region_beg; 2915 if (dest_addr == src_region_destination) { 2916 // Return the first live word in the source region. 2917 if (partial_obj_size == 0) { 2918 addr = bitmap->find_obj_beg(addr, src_region_end); 2919 assert(addr < src_region_end, "no objects start in src region"); 2920 } 2921 return addr; 2922 } 2923 2924 // Must skip some live data. 2925 size_t words_to_skip = dest_addr - src_region_destination; 2926 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2927 2928 if (partial_obj_size >= words_to_skip) { 2929 // All the live words to skip are part of the partial object. 2930 addr += words_to_skip; 2931 if (partial_obj_size == words_to_skip) { 2932 // Find the first live word past the partial object. 2933 addr = bitmap->find_obj_beg(addr, src_region_end); 2934 assert(addr < src_region_end, "wrong src region"); 2935 } 2936 return addr; 2937 } 2938 2939 // Skip over the partial object (if any). 2940 if (partial_obj_size != 0) { 2941 words_to_skip -= partial_obj_size; 2942 addr += partial_obj_size; 2943 } 2944 2945 // Skip over live words due to objects that start in the region. 2946 addr = skip_live_words(addr, src_region_end, words_to_skip); 2947 assert(addr < src_region_end, "wrong src region"); 2948 return addr; 2949 } 2950 2951 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2952 SpaceId src_space_id, 2953 size_t beg_region, 2954 HeapWord* end_addr) 2955 { 2956 ParallelCompactData& sd = summary_data(); 2957 2958 #ifdef ASSERT 2959 MutableSpace* const src_space = _space_info[src_space_id].space(); 2960 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2961 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2962 "src_space_id does not match beg_addr"); 2963 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2964 "src_space_id does not match end_addr"); 2965 #endif // #ifdef ASSERT 2966 2967 RegionData* const beg = sd.region(beg_region); 2968 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2969 2970 // Regions up to new_top() are enqueued if they become available. 2971 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2972 RegionData* const enqueue_end = 2973 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2974 2975 for (RegionData* cur = beg; cur < end; ++cur) { 2976 assert(cur->data_size() > 0, "region must have live data"); 2977 cur->decrement_destination_count(); 2978 if (cur < enqueue_end && cur->available() && cur->claim()) { 2979 if (cur->try_push()) { 2980 cm->push_region(sd.region(cur)); 2981 } else if (cur->try_copy()) { 2982 // Try to copy the content of the shadow region back to its corresponding 2983 // heap region if the shadow region is filled 2984 copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur)); 2985 cm->release_shadow_region(cur->shadow_region()); 2986 cur->set_completed(); 2987 } 2988 } 2989 } 2990 } 2991 2992 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2993 SpaceId& src_space_id, 2994 HeapWord*& src_space_top, 2995 HeapWord* end_addr) 2996 { 2997 typedef ParallelCompactData::RegionData RegionData; 2998 2999 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3000 const size_t region_size = ParallelCompactData::RegionSize; 3001 3002 size_t src_region_idx = 0; 3003 3004 // Skip empty regions (if any) up to the top of the space. 3005 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3006 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3007 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3008 const RegionData* const top_region_ptr = 3009 sd.addr_to_region_ptr(top_aligned_up); 3010 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3011 ++src_region_ptr; 3012 } 3013 3014 if (src_region_ptr < top_region_ptr) { 3015 // The next source region is in the current space. Update src_region_idx 3016 // and the source address to match src_region_ptr. 3017 src_region_idx = sd.region(src_region_ptr); 3018 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3019 if (src_region_addr > closure.source()) { 3020 closure.set_source(src_region_addr); 3021 } 3022 return src_region_idx; 3023 } 3024 3025 // Switch to a new source space and find the first non-empty region. 3026 unsigned int space_id = src_space_id + 1; 3027 assert(space_id < last_space_id, "not enough spaces"); 3028 3029 HeapWord* const destination = closure.destination(); 3030 3031 do { 3032 MutableSpace* space = _space_info[space_id].space(); 3033 HeapWord* const bottom = space->bottom(); 3034 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3035 3036 // Iterate over the spaces that do not compact into themselves. 3037 if (bottom_cp->destination() != bottom) { 3038 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3039 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3040 3041 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3042 if (src_cp->live_obj_size() > 0) { 3043 // Found it. 3044 assert(src_cp->destination() == destination, 3045 "first live obj in the space must match the destination"); 3046 assert(src_cp->partial_obj_size() == 0, 3047 "a space cannot begin with a partial obj"); 3048 3049 src_space_id = SpaceId(space_id); 3050 src_space_top = space->top(); 3051 const size_t src_region_idx = sd.region(src_cp); 3052 closure.set_source(sd.region_to_addr(src_region_idx)); 3053 return src_region_idx; 3054 } else { 3055 assert(src_cp->data_size() == 0, "sanity"); 3056 } 3057 } 3058 } 3059 } while (++space_id < last_space_id); 3060 3061 assert(false, "no source region was found"); 3062 return 0; 3063 } 3064 3065 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx) 3066 { 3067 typedef ParMarkBitMap::IterationStatus IterationStatus; 3068 const size_t RegionSize = ParallelCompactData::RegionSize; 3069 ParMarkBitMap* const bitmap = mark_bitmap(); 3070 ParallelCompactData& sd = summary_data(); 3071 RegionData* const region_ptr = sd.region(region_idx); 3072 3073 // Get the source region and related info. 3074 size_t src_region_idx = region_ptr->source_region(); 3075 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3076 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3077 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3078 3079 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3080 3081 // Adjust src_region_idx to prepare for decrementing destination counts (the 3082 // destination count is not decremented when a region is copied to itself). 3083 if (src_region_idx == region_idx) { 3084 src_region_idx += 1; 3085 } 3086 3087 if (bitmap->is_unmarked(closure.source())) { 3088 // The first source word is in the middle of an object; copy the remainder 3089 // of the object or as much as will fit. The fact that pointer updates were 3090 // deferred will be noted when the object header is processed. 3091 HeapWord* const old_src_addr = closure.source(); 3092 closure.copy_partial_obj(); 3093 if (closure.is_full()) { 3094 decrement_destination_counts(cm, src_space_id, src_region_idx, 3095 closure.source()); 3096 region_ptr->set_deferred_obj_addr(NULL); 3097 closure.complete_region(cm, dest_addr, region_ptr); 3098 return; 3099 } 3100 3101 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3102 if (sd.region_align_down(old_src_addr) != end_addr) { 3103 // The partial object was copied from more than one source region. 3104 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3105 3106 // Move to the next source region, possibly switching spaces as well. All 3107 // args except end_addr may be modified. 3108 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3109 end_addr); 3110 } 3111 } 3112 3113 do { 3114 HeapWord* const cur_addr = closure.source(); 3115 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3116 src_space_top); 3117 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3118 3119 if (status == ParMarkBitMap::incomplete) { 3120 // The last obj that starts in the source region does not end in the 3121 // region. 3122 assert(closure.source() < end_addr, "sanity"); 3123 HeapWord* const obj_beg = closure.source(); 3124 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3125 src_space_top); 3126 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3127 if (obj_end < range_end) { 3128 // The end was found; the entire object will fit. 3129 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3130 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3131 } else { 3132 // The end was not found; the object will not fit. 3133 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3134 status = ParMarkBitMap::would_overflow; 3135 } 3136 } 3137 3138 if (status == ParMarkBitMap::would_overflow) { 3139 // The last object did not fit. Note that interior oop updates were 3140 // deferred, then copy enough of the object to fill the region. 3141 region_ptr->set_deferred_obj_addr(closure.destination()); 3142 3143 status = closure.copy_until_full(); // copies from closure.source() 3144 3145 decrement_destination_counts(cm, src_space_id, src_region_idx, 3146 closure.source()); 3147 closure.complete_region(cm, dest_addr, region_ptr); 3148 return; 3149 } 3150 3151 if (status == ParMarkBitMap::full) { 3152 decrement_destination_counts(cm, src_space_id, src_region_idx, 3153 closure.source()); 3154 region_ptr->set_deferred_obj_addr(NULL); 3155 closure.complete_region(cm, dest_addr, region_ptr); 3156 return; 3157 } 3158 3159 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3160 3161 // Move to the next source region, possibly switching spaces as well. All 3162 // args except end_addr may be modified. 3163 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3164 end_addr); 3165 } while (true); 3166 } 3167 3168 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx) { 3169 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3170 fill_region(cm, cl, region_idx); 3171 } 3172 3173 void PSParallelCompact::fill_shadow_region(ParCompactionManager* cm, size_t region_idx) 3174 { 3175 // Acquire a shadow region at first 3176 ParallelCompactData& sd = summary_data(); 3177 RegionData* const region_ptr = sd.region(region_idx); 3178 size_t shadow_region = cm->acquire_shadow_region(region_ptr); 3179 // The zero return value indicates the corresponding heap region is available, 3180 // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use 3181 // ShadowClosure to fill the acquired shadow region. 3182 if (shadow_region == 0) { 3183 MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx); 3184 region_ptr->mark_normal(); 3185 return fill_region(cm, cl, region_idx); 3186 } else { 3187 ShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region); 3188 return fill_region(cm, cl, region_idx); 3189 } 3190 } 3191 3192 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr) { 3193 Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize); 3194 } 3195 3196 bool PSParallelCompact::steal_shadow_region(ParCompactionManager* cm, size_t ®ion_idx) { 3197 size_t record = cm->shadow_record(); 3198 ParallelCompactData& sd = _summary_data; 3199 size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top()); 3200 uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers(); 3201 3202 while (record < old_new_top) { 3203 if (sd.region(record)->try_steal()) { 3204 region_idx = record; 3205 return true; 3206 } 3207 record = cm->next_shadow_record(active_gc_threads); 3208 } 3209 3210 return false; 3211 } 3212 3213 void PSParallelCompact::enqueue_shadow_region() { 3214 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 3215 3216 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 3217 SpaceInfo* const space_info = _space_info + id; 3218 MutableSpace* const space = space_info->space(); 3219 3220 const size_t beg_region = 3221 sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top()))); 3222 const size_t end_region = 3223 sd.addr_to_region_idx(sd.region_align_down(space->end())); 3224 3225 for (size_t cur = beg_region + 1; cur < end_region; ++cur) { 3226 ParCompactionManager::enqueue_shadow_region(cur); 3227 } 3228 } 3229 } 3230 3231 void PSParallelCompact::initialize_steal_record(uint which) { 3232 ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(which); 3233 3234 size_t record = _summary_data.addr_to_region_idx(_space_info[old_space_id].dense_prefix()); 3235 cm->set_shadow_record(record + which); 3236 } 3237 3238 void PSParallelCompact::fill_blocks(size_t region_idx) 3239 { 3240 // Fill in the block table elements for the specified region. Each block 3241 // table element holds the number of live words in the region that are to the 3242 // left of the first object that starts in the block. Thus only blocks in 3243 // which an object starts need to be filled. 3244 // 3245 // The algorithm scans the section of the bitmap that corresponds to the 3246 // region, keeping a running total of the live words. When an object start is 3247 // found, if it's the first to start in the block that contains it, the 3248 // current total is written to the block table element. 3249 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3250 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3251 const size_t RegionSize = ParallelCompactData::RegionSize; 3252 3253 ParallelCompactData& sd = summary_data(); 3254 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3255 if (partial_obj_size >= RegionSize) { 3256 return; // No objects start in this region. 3257 } 3258 3259 // Ensure the first loop iteration decides that the block has changed. 3260 size_t cur_block = sd.block_count(); 3261 3262 const ParMarkBitMap* const bitmap = mark_bitmap(); 3263 3264 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3265 assert((size_t)1 << Log2BitsPerBlock == 3266 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3267 3268 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3269 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3270 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3271 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3272 while (beg_bit < range_end) { 3273 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3274 if (new_block != cur_block) { 3275 cur_block = new_block; 3276 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3277 } 3278 3279 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3280 if (end_bit < range_end - 1) { 3281 live_bits += end_bit - beg_bit + 1; 3282 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3283 } else { 3284 return; 3285 } 3286 } 3287 } 3288 3289 jlong PSParallelCompact::millis_since_last_gc() { 3290 // We need a monotonically non-decreasing time in ms but 3291 // os::javaTimeMillis() does not guarantee monotonicity. 3292 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3293 jlong ret_val = now - _time_of_last_gc; 3294 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3295 if (ret_val < 0) { 3296 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3297 return 0; 3298 } 3299 return ret_val; 3300 } 3301 3302 void PSParallelCompact::reset_millis_since_last_gc() { 3303 // We need a monotonically non-decreasing time in ms but 3304 // os::javaTimeMillis() does not guarantee monotonicity. 3305 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3306 } 3307 3308 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3309 { 3310 if (source() != copy_destination()) { 3311 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3312 Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining()); 3313 } 3314 update_state(words_remaining()); 3315 assert(is_full(), "sanity"); 3316 return ParMarkBitMap::full; 3317 } 3318 3319 void MoveAndUpdateClosure::copy_partial_obj() 3320 { 3321 size_t words = words_remaining(); 3322 3323 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3324 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3325 if (end_addr < range_end) { 3326 words = bitmap()->obj_size(source(), end_addr); 3327 } 3328 3329 // This test is necessary; if omitted, the pointer updates to a partial object 3330 // that crosses the dense prefix boundary could be overwritten. 3331 if (source() != copy_destination()) { 3332 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3333 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3334 } 3335 update_state(words); 3336 } 3337 3338 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3339 PSParallelCompact::RegionData *region_ptr) { 3340 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::FINISH, "Region should be finished"); 3341 region_ptr->set_completed(); 3342 } 3343 3344 ParMarkBitMapClosure::IterationStatus 3345 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3346 assert(destination() != NULL, "sanity"); 3347 assert(bitmap()->obj_size(addr) == words, "bad size"); 3348 3349 _source = addr; 3350 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3351 destination(), "wrong destination"); 3352 3353 if (words > words_remaining()) { 3354 return ParMarkBitMap::would_overflow; 3355 } 3356 3357 // The start_array must be updated even if the object is not moving. 3358 if (_start_array != NULL) { 3359 _start_array->allocate_block(destination()); 3360 } 3361 3362 if (copy_destination() != source()) { 3363 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3364 Copy::aligned_conjoint_words(source(), copy_destination(), words); 3365 } 3366 3367 oop moved_oop = (oop) copy_destination(); 3368 compaction_manager()->update_contents(moved_oop); 3369 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3370 3371 update_state(words); 3372 assert(copy_destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3373 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3374 } 3375 3376 void ShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr, 3377 PSParallelCompact::RegionData *region_ptr) { 3378 assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::SHADOW, "Region should be shadow"); 3379 // Record the shadow region index 3380 region_ptr->set_shadow_region(_shadow); 3381 // Mark the shadow region filled 3382 region_ptr->mark_filled(); 3383 // Try to copy the content of the shadow region back to its corresponding 3384 // heap region if available 3385 if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->try_copy()) { 3386 region_ptr->set_completed(); 3387 PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr); 3388 cm->release_shadow_region(_shadow); 3389 } 3390 } 3391 3392 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3393 ParCompactionManager* cm, 3394 PSParallelCompact::SpaceId space_id) : 3395 ParMarkBitMapClosure(mbm, cm), 3396 _space_id(space_id), 3397 _start_array(PSParallelCompact::start_array(space_id)) 3398 { 3399 } 3400 3401 // Updates the references in the object to their new values. 3402 ParMarkBitMapClosure::IterationStatus 3403 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3404 do_addr(addr); 3405 return ParMarkBitMap::incomplete; 3406 } 3407 3408 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3409 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3410 _start_array(PSParallelCompact::start_array(space_id)) 3411 { 3412 assert(space_id == PSParallelCompact::old_space_id, 3413 "cannot use FillClosure in the young gen"); 3414 } 3415 3416 ParMarkBitMapClosure::IterationStatus 3417 FillClosure::do_addr(HeapWord* addr, size_t size) { 3418 CollectedHeap::fill_with_objects(addr, size); 3419 HeapWord* const end = addr + size; 3420 do { 3421 _start_array->allocate_block(addr); 3422 addr += oop(addr)->size(); 3423 } while (addr < end); 3424 return ParMarkBitMap::incomplete; 3425 }