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