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