1 /* 2 * Copyright (c) 2005, 2018, 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/javaClasses.inline.hpp" 28 #include "classfile/stringTable.hpp" 29 #include "classfile/symbolTable.hpp" 30 #include "classfile/systemDictionary.hpp" 31 #include "code/codeCache.hpp" 32 #include "gc/parallel/gcTaskManager.hpp" 33 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 34 #include "gc/parallel/parMarkBitMap.inline.hpp" 35 #include "gc/parallel/pcTasks.hpp" 36 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 37 #include "gc/parallel/psCompactionManager.inline.hpp" 38 #include "gc/parallel/psOldGen.hpp" 39 #include "gc/parallel/psParallelCompact.inline.hpp" 40 #include "gc/parallel/psPromotionManager.inline.hpp" 41 #include "gc/parallel/psScavenge.hpp" 42 #include "gc/parallel/psYoungGen.hpp" 43 #include "gc/shared/gcCause.hpp" 44 #include "gc/shared/gcHeapSummary.hpp" 45 #include "gc/shared/gcId.hpp" 46 #include "gc/shared/gcLocker.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.inline.hpp" 50 #include "gc/shared/isGCActiveMark.hpp" 51 #include "gc/shared/referencePolicy.hpp" 52 #include "gc/shared/referenceProcessor.hpp" 53 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 54 #include "gc/shared/spaceDecorator.hpp" 55 #include "gc/shared/weakProcessor.hpp" 56 #include "logging/log.hpp" 57 #include "memory/iterator.inline.hpp" 58 #include "memory/resourceArea.hpp" 59 #include "oops/access.inline.hpp" 60 #include "oops/instanceClassLoaderKlass.inline.hpp" 61 #include "oops/instanceKlass.inline.hpp" 62 #include "oops/instanceMirrorKlass.inline.hpp" 63 #include "oops/methodData.hpp" 64 #include "oops/objArrayKlass.inline.hpp" 65 #include "oops/oop.inline.hpp" 66 #include "runtime/atomic.hpp" 67 #include "runtime/handles.inline.hpp" 68 #include "runtime/safepoint.hpp" 69 #include "runtime/vmThread.hpp" 70 #include "services/management.hpp" 71 #include "services/memTracker.hpp" 72 #include "services/memoryService.hpp" 73 #include "utilities/align.hpp" 74 #include "utilities/debug.hpp" 75 #include "utilities/events.hpp" 76 #include "utilities/formatBuffer.hpp" 77 #include "utilities/macros.hpp" 78 #include "utilities/stack.inline.hpp" 79 80 #include <math.h> 81 82 // All sizes are in HeapWords. 83 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 84 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 85 const size_t ParallelCompactData::RegionSizeBytes = 86 RegionSize << LogHeapWordSize; 87 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 88 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 89 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 90 91 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 92 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 93 const size_t ParallelCompactData::BlockSizeBytes = 94 BlockSize << LogHeapWordSize; 95 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 96 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 97 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 98 99 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 100 const size_t ParallelCompactData::Log2BlocksPerRegion = 101 Log2RegionSize - Log2BlockSize; 102 103 const ParallelCompactData::RegionData::region_sz_t 104 ParallelCompactData::RegionData::dc_shift = 27; 105 106 const ParallelCompactData::RegionData::region_sz_t 107 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 108 109 const ParallelCompactData::RegionData::region_sz_t 110 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 111 112 const ParallelCompactData::RegionData::region_sz_t 113 ParallelCompactData::RegionData::los_mask = ~dc_mask; 114 115 const ParallelCompactData::RegionData::region_sz_t 116 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 117 118 const ParallelCompactData::RegionData::region_sz_t 119 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 120 121 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 122 123 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 124 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 125 126 double PSParallelCompact::_dwl_mean; 127 double PSParallelCompact::_dwl_std_dev; 128 double PSParallelCompact::_dwl_first_term; 129 double PSParallelCompact::_dwl_adjustment; 130 #ifdef ASSERT 131 bool PSParallelCompact::_dwl_initialized = false; 132 #endif // #ifdef ASSERT 133 134 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 135 HeapWord* destination) 136 { 137 assert(src_region_idx != 0, "invalid src_region_idx"); 138 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 139 assert(destination != NULL, "invalid destination argument"); 140 141 _src_region_idx = src_region_idx; 142 _partial_obj_size = partial_obj_size; 143 _destination = destination; 144 145 // These fields may not be updated below, so make sure they're clear. 146 assert(_dest_region_addr == NULL, "should have been cleared"); 147 assert(_first_src_addr == NULL, "should have been cleared"); 148 149 // Determine the number of destination regions for the partial object. 150 HeapWord* const last_word = destination + partial_obj_size - 1; 151 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 152 HeapWord* const beg_region_addr = sd.region_align_down(destination); 153 HeapWord* const end_region_addr = sd.region_align_down(last_word); 154 155 if (beg_region_addr == end_region_addr) { 156 // One destination region. 157 _destination_count = 1; 158 if (end_region_addr == destination) { 159 // The destination falls on a region boundary, thus the first word of the 160 // partial object will be the first word copied to the destination region. 161 _dest_region_addr = end_region_addr; 162 _first_src_addr = sd.region_to_addr(src_region_idx); 163 } 164 } else { 165 // Two destination regions. When copied, the partial object will cross a 166 // destination region boundary, so a word somewhere within the partial 167 // object will be the first word copied to the second destination region. 168 _destination_count = 2; 169 _dest_region_addr = end_region_addr; 170 const size_t ofs = pointer_delta(end_region_addr, destination); 171 assert(ofs < _partial_obj_size, "sanity"); 172 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 173 } 174 } 175 176 void SplitInfo::clear() 177 { 178 _src_region_idx = 0; 179 _partial_obj_size = 0; 180 _destination = NULL; 181 _destination_count = 0; 182 _dest_region_addr = NULL; 183 _first_src_addr = NULL; 184 assert(!is_valid(), "sanity"); 185 } 186 187 #ifdef ASSERT 188 void SplitInfo::verify_clear() 189 { 190 assert(_src_region_idx == 0, "not clear"); 191 assert(_partial_obj_size == 0, "not clear"); 192 assert(_destination == NULL, "not clear"); 193 assert(_destination_count == 0, "not clear"); 194 assert(_dest_region_addr == NULL, "not clear"); 195 assert(_first_src_addr == NULL, "not clear"); 196 } 197 #endif // #ifdef ASSERT 198 199 200 void PSParallelCompact::print_on_error(outputStream* st) { 201 _mark_bitmap.print_on_error(st); 202 } 203 204 #ifndef PRODUCT 205 const char* PSParallelCompact::space_names[] = { 206 "old ", "eden", "from", "to " 207 }; 208 209 void PSParallelCompact::print_region_ranges() { 210 if (!log_develop_is_enabled(Trace, gc, compaction)) { 211 return; 212 } 213 Log(gc, compaction) log; 214 ResourceMark rm; 215 LogStream ls(log.trace()); 216 Universe::print_on(&ls); 217 log.trace("space bottom top end new_top"); 218 log.trace("------ ---------- ---------- ---------- ----------"); 219 220 for (unsigned int id = 0; id < last_space_id; ++id) { 221 const MutableSpace* space = _space_info[id].space(); 222 log.trace("%u %s " 223 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 224 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 225 id, space_names[id], 226 summary_data().addr_to_region_idx(space->bottom()), 227 summary_data().addr_to_region_idx(space->top()), 228 summary_data().addr_to_region_idx(space->end()), 229 summary_data().addr_to_region_idx(_space_info[id].new_top())); 230 } 231 } 232 233 void 234 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 235 { 236 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 237 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 238 239 ParallelCompactData& sd = PSParallelCompact::summary_data(); 240 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 241 log_develop_trace(gc, compaction)( 242 REGION_IDX_FORMAT " " PTR_FORMAT " " 243 REGION_IDX_FORMAT " " PTR_FORMAT " " 244 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 245 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 246 i, p2i(c->data_location()), dci, p2i(c->destination()), 247 c->partial_obj_size(), c->live_obj_size(), 248 c->data_size(), c->source_region(), c->destination_count()); 249 250 #undef REGION_IDX_FORMAT 251 #undef REGION_DATA_FORMAT 252 } 253 254 void 255 print_generic_summary_data(ParallelCompactData& summary_data, 256 HeapWord* const beg_addr, 257 HeapWord* const end_addr) 258 { 259 size_t total_words = 0; 260 size_t i = summary_data.addr_to_region_idx(beg_addr); 261 const size_t last = summary_data.addr_to_region_idx(end_addr); 262 HeapWord* pdest = 0; 263 264 while (i < last) { 265 ParallelCompactData::RegionData* c = summary_data.region(i); 266 if (c->data_size() != 0 || c->destination() != pdest) { 267 print_generic_summary_region(i, c); 268 total_words += c->data_size(); 269 pdest = c->destination(); 270 } 271 ++i; 272 } 273 274 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 275 } 276 277 void 278 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 279 HeapWord* const beg_addr, 280 HeapWord* const end_addr) { 281 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 282 } 283 284 void 285 print_generic_summary_data(ParallelCompactData& summary_data, 286 SpaceInfo* space_info) 287 { 288 if (!log_develop_is_enabled(Trace, gc, compaction)) { 289 return; 290 } 291 292 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 293 const MutableSpace* space = space_info[id].space(); 294 print_generic_summary_data(summary_data, space->bottom(), 295 MAX2(space->top(), space_info[id].new_top())); 296 } 297 } 298 299 void 300 print_initial_summary_data(ParallelCompactData& summary_data, 301 const MutableSpace* space) { 302 if (space->top() == space->bottom()) { 303 return; 304 } 305 306 const size_t region_size = ParallelCompactData::RegionSize; 307 typedef ParallelCompactData::RegionData RegionData; 308 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 309 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 310 const RegionData* c = summary_data.region(end_region - 1); 311 HeapWord* end_addr = c->destination() + c->data_size(); 312 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 313 314 // Print (and count) the full regions at the beginning of the space. 315 size_t full_region_count = 0; 316 size_t i = summary_data.addr_to_region_idx(space->bottom()); 317 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 318 ParallelCompactData::RegionData* c = summary_data.region(i); 319 log_develop_trace(gc, compaction)( 320 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 321 i, p2i(c->destination()), 322 c->partial_obj_size(), c->live_obj_size(), 323 c->data_size(), c->source_region(), c->destination_count()); 324 ++full_region_count; 325 ++i; 326 } 327 328 size_t live_to_right = live_in_space - full_region_count * region_size; 329 330 double max_reclaimed_ratio = 0.0; 331 size_t max_reclaimed_ratio_region = 0; 332 size_t max_dead_to_right = 0; 333 size_t max_live_to_right = 0; 334 335 // Print the 'reclaimed ratio' for regions while there is something live in 336 // the region or to the right of it. The remaining regions are empty (and 337 // uninteresting), and computing the ratio will result in division by 0. 338 while (i < end_region && live_to_right > 0) { 339 c = summary_data.region(i); 340 HeapWord* const region_addr = summary_data.region_to_addr(i); 341 const size_t used_to_right = pointer_delta(space->top(), region_addr); 342 const size_t dead_to_right = used_to_right - live_to_right; 343 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 344 345 if (reclaimed_ratio > max_reclaimed_ratio) { 346 max_reclaimed_ratio = reclaimed_ratio; 347 max_reclaimed_ratio_region = i; 348 max_dead_to_right = dead_to_right; 349 max_live_to_right = live_to_right; 350 } 351 352 ParallelCompactData::RegionData* c = summary_data.region(i); 353 log_develop_trace(gc, compaction)( 354 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 355 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 356 i, p2i(c->destination()), 357 c->partial_obj_size(), c->live_obj_size(), 358 c->data_size(), c->source_region(), c->destination_count(), 359 reclaimed_ratio, dead_to_right, live_to_right); 360 361 362 live_to_right -= c->data_size(); 363 ++i; 364 } 365 366 // Any remaining regions are empty. Print one more if there is one. 367 if (i < end_region) { 368 ParallelCompactData::RegionData* c = summary_data.region(i); 369 log_develop_trace(gc, compaction)( 370 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 371 i, p2i(c->destination()), 372 c->partial_obj_size(), c->live_obj_size(), 373 c->data_size(), c->source_region(), c->destination_count()); 374 } 375 376 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 377 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 378 } 379 380 void 381 print_initial_summary_data(ParallelCompactData& summary_data, 382 SpaceInfo* space_info) { 383 if (!log_develop_is_enabled(Trace, gc, compaction)) { 384 return; 385 } 386 387 unsigned int id = PSParallelCompact::old_space_id; 388 const MutableSpace* space; 389 do { 390 space = space_info[id].space(); 391 print_initial_summary_data(summary_data, space); 392 } while (++id < PSParallelCompact::eden_space_id); 393 394 do { 395 space = space_info[id].space(); 396 print_generic_summary_data(summary_data, space->bottom(), space->top()); 397 } while (++id < PSParallelCompact::last_space_id); 398 } 399 #endif // #ifndef PRODUCT 400 401 #ifdef ASSERT 402 size_t add_obj_count; 403 size_t add_obj_size; 404 size_t mark_bitmap_count; 405 size_t mark_bitmap_size; 406 #endif // #ifdef ASSERT 407 408 ParallelCompactData::ParallelCompactData() 409 { 410 _region_start = 0; 411 412 _region_vspace = 0; 413 _reserved_byte_size = 0; 414 _region_data = 0; 415 _region_count = 0; 416 417 _block_vspace = 0; 418 _block_data = 0; 419 _block_count = 0; 420 } 421 422 bool ParallelCompactData::initialize(MemRegion covered_region) 423 { 424 _region_start = covered_region.start(); 425 const size_t region_size = covered_region.word_size(); 426 DEBUG_ONLY(_region_end = _region_start + region_size;) 427 428 assert(region_align_down(_region_start) == _region_start, 429 "region start not aligned"); 430 assert((region_size & RegionSizeOffsetMask) == 0, 431 "region size not a multiple of RegionSize"); 432 433 bool result = initialize_region_data(region_size) && initialize_block_data(); 434 return result; 435 } 436 437 PSVirtualSpace* 438 ParallelCompactData::create_vspace(size_t count, size_t element_size) 439 { 440 const size_t raw_bytes = count * element_size; 441 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 442 const size_t granularity = os::vm_allocation_granularity(); 443 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 444 445 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 446 MAX2(page_sz, granularity); 447 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 448 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 449 rs.size()); 450 451 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 452 453 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 454 if (vspace != 0) { 455 if (vspace->expand_by(_reserved_byte_size)) { 456 return vspace; 457 } 458 delete vspace; 459 // Release memory reserved in the space. 460 rs.release(); 461 } 462 463 return 0; 464 } 465 466 bool ParallelCompactData::initialize_region_data(size_t region_size) 467 { 468 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 469 _region_vspace = create_vspace(count, sizeof(RegionData)); 470 if (_region_vspace != 0) { 471 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 472 _region_count = count; 473 return true; 474 } 475 return false; 476 } 477 478 bool ParallelCompactData::initialize_block_data() 479 { 480 assert(_region_count != 0, "region data must be initialized first"); 481 const size_t count = _region_count << Log2BlocksPerRegion; 482 _block_vspace = create_vspace(count, sizeof(BlockData)); 483 if (_block_vspace != 0) { 484 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 485 _block_count = count; 486 return true; 487 } 488 return false; 489 } 490 491 void ParallelCompactData::clear() 492 { 493 memset(_region_data, 0, _region_vspace->committed_size()); 494 memset(_block_data, 0, _block_vspace->committed_size()); 495 } 496 497 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 498 assert(beg_region <= _region_count, "beg_region out of range"); 499 assert(end_region <= _region_count, "end_region out of range"); 500 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 501 502 const size_t region_cnt = end_region - beg_region; 503 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 504 505 const size_t beg_block = beg_region * BlocksPerRegion; 506 const size_t block_cnt = region_cnt * BlocksPerRegion; 507 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 508 } 509 510 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 511 { 512 const RegionData* cur_cp = region(region_idx); 513 const RegionData* const end_cp = region(region_count() - 1); 514 515 HeapWord* result = region_to_addr(region_idx); 516 if (cur_cp < end_cp) { 517 do { 518 result += cur_cp->partial_obj_size(); 519 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 520 } 521 return result; 522 } 523 524 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 525 { 526 const size_t obj_ofs = pointer_delta(addr, _region_start); 527 const size_t beg_region = obj_ofs >> Log2RegionSize; 528 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 529 530 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 531 DEBUG_ONLY(Atomic::add(len, &add_obj_size);) 532 533 if (beg_region == end_region) { 534 // All in one region. 535 _region_data[beg_region].add_live_obj(len); 536 return; 537 } 538 539 // First region. 540 const size_t beg_ofs = region_offset(addr); 541 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 542 543 Klass* klass = ((oop)addr)->klass(); 544 // Middle regions--completely spanned by this object. 545 for (size_t region = beg_region + 1; region < end_region; ++region) { 546 _region_data[region].set_partial_obj_size(RegionSize); 547 _region_data[region].set_partial_obj_addr(addr); 548 } 549 550 // Last region. 551 const size_t end_ofs = region_offset(addr + len - 1); 552 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 553 _region_data[end_region].set_partial_obj_addr(addr); 554 } 555 556 void 557 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 558 { 559 assert(region_offset(beg) == 0, "not RegionSize aligned"); 560 assert(region_offset(end) == 0, "not RegionSize aligned"); 561 562 size_t cur_region = addr_to_region_idx(beg); 563 const size_t end_region = addr_to_region_idx(end); 564 HeapWord* addr = beg; 565 while (cur_region < end_region) { 566 _region_data[cur_region].set_destination(addr); 567 _region_data[cur_region].set_destination_count(0); 568 _region_data[cur_region].set_source_region(cur_region); 569 _region_data[cur_region].set_data_location(addr); 570 571 // Update live_obj_size so the region appears completely full. 572 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 573 _region_data[cur_region].set_live_obj_size(live_size); 574 575 ++cur_region; 576 addr += RegionSize; 577 } 578 } 579 580 // Find the point at which a space can be split and, if necessary, record the 581 // split point. 582 // 583 // If the current src region (which overflowed the destination space) doesn't 584 // have a partial object, the split point is at the beginning of the current src 585 // region (an "easy" split, no extra bookkeeping required). 586 // 587 // If the current src region has a partial object, the split point is in the 588 // region where that partial object starts (call it the split_region). If 589 // split_region has a partial object, then the split point is just after that 590 // partial object (a "hard" split where we have to record the split data and 591 // zero the partial_obj_size field). With a "hard" split, we know that the 592 // partial_obj ends within split_region because the partial object that caused 593 // the overflow starts in split_region. If split_region doesn't have a partial 594 // obj, then the split is at the beginning of split_region (another "easy" 595 // split). 596 HeapWord* 597 ParallelCompactData::summarize_split_space(size_t src_region, 598 SplitInfo& split_info, 599 HeapWord* destination, 600 HeapWord* target_end, 601 HeapWord** target_next) 602 { 603 assert(destination <= target_end, "sanity"); 604 assert(destination + _region_data[src_region].data_size() > target_end, 605 "region should not fit into target space"); 606 assert(is_region_aligned(target_end), "sanity"); 607 608 size_t split_region = src_region; 609 HeapWord* split_destination = destination; 610 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 611 612 if (destination + partial_obj_size > target_end) { 613 // The split point is just after the partial object (if any) in the 614 // src_region that contains the start of the object that overflowed the 615 // destination space. 616 // 617 // Find the start of the "overflow" object and set split_region to the 618 // region containing it. 619 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 620 split_region = addr_to_region_idx(overflow_obj); 621 622 // Clear the source_region field of all destination regions whose first word 623 // came from data after the split point (a non-null source_region field 624 // implies a region must be filled). 625 // 626 // An alternative to the simple loop below: clear during post_compact(), 627 // which uses memcpy instead of individual stores, and is easy to 628 // parallelize. (The downside is that it clears the entire RegionData 629 // object as opposed to just one field.) 630 // 631 // post_compact() would have to clear the summary data up to the highest 632 // address that was written during the summary phase, which would be 633 // 634 // max(top, max(new_top, clear_top)) 635 // 636 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 637 // to target_end. 638 const RegionData* const sr = region(split_region); 639 const size_t beg_idx = 640 addr_to_region_idx(region_align_up(sr->destination() + 641 sr->partial_obj_size())); 642 const size_t end_idx = addr_to_region_idx(target_end); 643 644 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 645 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 646 _region_data[idx].set_source_region(0); 647 } 648 649 // Set split_destination and partial_obj_size to reflect the split region. 650 split_destination = sr->destination(); 651 partial_obj_size = sr->partial_obj_size(); 652 } 653 654 // The split is recorded only if a partial object extends onto the region. 655 if (partial_obj_size != 0) { 656 _region_data[split_region].set_partial_obj_size(0); 657 split_info.record(split_region, partial_obj_size, split_destination); 658 } 659 660 // Setup the continuation addresses. 661 *target_next = split_destination + partial_obj_size; 662 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 663 664 if (log_develop_is_enabled(Trace, gc, compaction)) { 665 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 666 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 667 split_type, p2i(source_next), split_region, partial_obj_size); 668 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 669 split_type, p2i(split_destination), 670 addr_to_region_idx(split_destination), 671 p2i(*target_next)); 672 673 if (partial_obj_size != 0) { 674 HeapWord* const po_beg = split_info.destination(); 675 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 676 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 677 split_type, 678 p2i(po_beg), addr_to_region_idx(po_beg), 679 p2i(po_end), addr_to_region_idx(po_end)); 680 } 681 } 682 683 return source_next; 684 } 685 686 bool ParallelCompactData::summarize(SplitInfo& split_info, 687 HeapWord* source_beg, HeapWord* source_end, 688 HeapWord** source_next, 689 HeapWord* target_beg, HeapWord* target_end, 690 HeapWord** target_next) 691 { 692 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 693 log_develop_trace(gc, compaction)( 694 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 695 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 696 p2i(source_beg), p2i(source_end), p2i(source_next_val), 697 p2i(target_beg), p2i(target_end), p2i(*target_next)); 698 699 size_t cur_region = addr_to_region_idx(source_beg); 700 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 701 702 HeapWord *dest_addr = target_beg; 703 while (cur_region < end_region) { 704 // The destination must be set even if the region has no data. 705 _region_data[cur_region].set_destination(dest_addr); 706 707 size_t words = _region_data[cur_region].data_size(); 708 if (words > 0) { 709 // If cur_region does not fit entirely into the target space, find a point 710 // at which the source space can be 'split' so that part is copied to the 711 // target space and the rest is copied elsewhere. 712 if (dest_addr + words > target_end) { 713 assert(source_next != NULL, "source_next is NULL when splitting"); 714 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 715 target_end, target_next); 716 return false; 717 } 718 719 // Compute the destination_count for cur_region, and if necessary, update 720 // source_region for a destination region. The source_region field is 721 // updated if cur_region is the first (left-most) region to be copied to a 722 // destination region. 723 // 724 // The destination_count calculation is a bit subtle. A region that has 725 // data that compacts into itself does not count itself as a destination. 726 // This maintains the invariant that a zero count means the region is 727 // available and can be claimed and then filled. 728 uint destination_count = 0; 729 if (split_info.is_split(cur_region)) { 730 // The current region has been split: the partial object will be copied 731 // to one destination space and the remaining data will be copied to 732 // another destination space. Adjust the initial destination_count and, 733 // if necessary, set the source_region field if the partial object will 734 // cross a destination region boundary. 735 destination_count = split_info.destination_count(); 736 if (destination_count == 2) { 737 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 738 _region_data[dest_idx].set_source_region(cur_region); 739 } 740 } 741 742 HeapWord* const last_addr = dest_addr + words - 1; 743 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 744 const size_t dest_region_2 = addr_to_region_idx(last_addr); 745 746 // Initially assume that the destination regions will be the same and 747 // adjust the value below if necessary. Under this assumption, if 748 // cur_region == dest_region_2, then cur_region will be compacted 749 // completely into itself. 750 destination_count += cur_region == dest_region_2 ? 0 : 1; 751 if (dest_region_1 != dest_region_2) { 752 // Destination regions differ; adjust destination_count. 753 destination_count += 1; 754 // Data from cur_region will be copied to the start of dest_region_2. 755 _region_data[dest_region_2].set_source_region(cur_region); 756 } else if (region_offset(dest_addr) == 0) { 757 // Data from cur_region will be copied to the start of the destination 758 // region. 759 _region_data[dest_region_1].set_source_region(cur_region); 760 } 761 762 _region_data[cur_region].set_destination_count(destination_count); 763 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 764 dest_addr += words; 765 } 766 767 ++cur_region; 768 } 769 770 *target_next = dest_addr; 771 return true; 772 } 773 774 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 775 assert(addr != NULL, "Should detect NULL oop earlier"); 776 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 777 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 778 779 // Region covering the object. 780 RegionData* const region_ptr = addr_to_region_ptr(addr); 781 HeapWord* result = region_ptr->destination(); 782 783 // If the entire Region is live, the new location is region->destination + the 784 // offset of the object within in the Region. 785 786 // Run some performance tests to determine if this special case pays off. It 787 // is worth it for pointers into the dense prefix. If the optimization to 788 // avoid pointer updates in regions that only point to the dense prefix is 789 // ever implemented, this should be revisited. 790 if (region_ptr->data_size() == RegionSize) { 791 result += region_offset(addr); 792 return result; 793 } 794 795 // Otherwise, the new location is region->destination + block offset + the 796 // number of live words in the Block that are (a) to the left of addr and (b) 797 // due to objects that start in the Block. 798 799 // Fill in the block table if necessary. This is unsynchronized, so multiple 800 // threads may fill the block table for a region (harmless, since it is 801 // idempotent). 802 if (!region_ptr->blocks_filled()) { 803 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 804 region_ptr->set_blocks_filled(); 805 } 806 807 HeapWord* const search_start = block_align_down(addr); 808 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 809 810 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 811 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 812 result += block_offset + live; 813 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 814 return result; 815 } 816 817 #ifdef ASSERT 818 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 819 { 820 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 821 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 822 for (const size_t* p = beg; p < end; ++p) { 823 assert(*p == 0, "not zero"); 824 } 825 } 826 827 void ParallelCompactData::verify_clear() 828 { 829 verify_clear(_region_vspace); 830 verify_clear(_block_vspace); 831 } 832 #endif // #ifdef ASSERT 833 834 STWGCTimer PSParallelCompact::_gc_timer; 835 ParallelOldTracer PSParallelCompact::_gc_tracer; 836 elapsedTimer PSParallelCompact::_accumulated_time; 837 unsigned int PSParallelCompact::_total_invocations = 0; 838 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 839 jlong PSParallelCompact::_time_of_last_gc = 0; 840 CollectorCounters* PSParallelCompact::_counters = NULL; 841 ParMarkBitMap PSParallelCompact::_mark_bitmap; 842 ParallelCompactData PSParallelCompact::_summary_data; 843 844 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 845 846 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 847 848 void PSParallelCompact::post_initialize() { 849 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 850 _span_based_discoverer.set_span(heap->reserved_region()); 851 _ref_processor = 852 new ReferenceProcessor(&_span_based_discoverer, 853 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 854 ParallelGCThreads, // mt processing degree 855 true, // mt discovery 856 ParallelGCThreads, // mt discovery degree 857 true, // atomic_discovery 858 &_is_alive_closure, // non-header is alive closure 859 false); // disable adjusting number of processing threads 860 _counters = new CollectorCounters("PSParallelCompact", 1); 861 862 // Initialize static fields in ParCompactionManager. 863 ParCompactionManager::initialize(mark_bitmap()); 864 } 865 866 bool PSParallelCompact::initialize() { 867 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 868 MemRegion mr = heap->reserved_region(); 869 870 // Was the old gen get allocated successfully? 871 if (!heap->old_gen()->is_allocated()) { 872 return false; 873 } 874 875 initialize_space_info(); 876 initialize_dead_wood_limiter(); 877 878 if (!_mark_bitmap.initialize(mr)) { 879 vm_shutdown_during_initialization( 880 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 881 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 882 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 883 return false; 884 } 885 886 if (!_summary_data.initialize(mr)) { 887 vm_shutdown_during_initialization( 888 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 889 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 890 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 891 return false; 892 } 893 894 return true; 895 } 896 897 void PSParallelCompact::initialize_space_info() 898 { 899 memset(&_space_info, 0, sizeof(_space_info)); 900 901 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 902 PSYoungGen* young_gen = heap->young_gen(); 903 904 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 905 _space_info[eden_space_id].set_space(young_gen->eden_space()); 906 _space_info[from_space_id].set_space(young_gen->from_space()); 907 _space_info[to_space_id].set_space(young_gen->to_space()); 908 909 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 910 } 911 912 void PSParallelCompact::initialize_dead_wood_limiter() 913 { 914 const size_t max = 100; 915 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 916 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 917 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 918 DEBUG_ONLY(_dwl_initialized = true;) 919 _dwl_adjustment = normal_distribution(1.0); 920 } 921 922 void 923 PSParallelCompact::clear_data_covering_space(SpaceId id) 924 { 925 // At this point, top is the value before GC, new_top() is the value that will 926 // be set at the end of GC. The marking bitmap is cleared to top; nothing 927 // should be marked above top. The summary data is cleared to the larger of 928 // top & new_top. 929 MutableSpace* const space = _space_info[id].space(); 930 HeapWord* const bot = space->bottom(); 931 HeapWord* const top = space->top(); 932 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 933 934 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 935 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 936 _mark_bitmap.clear_range(beg_bit, end_bit); 937 938 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 939 const size_t end_region = 940 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 941 _summary_data.clear_range(beg_region, end_region); 942 943 // Clear the data used to 'split' regions. 944 SplitInfo& split_info = _space_info[id].split_info(); 945 if (split_info.is_valid()) { 946 split_info.clear(); 947 } 948 DEBUG_ONLY(split_info.verify_clear();) 949 } 950 951 void PSParallelCompact::pre_compact() 952 { 953 // Update the from & to space pointers in space_info, since they are swapped 954 // at each young gen gc. Do the update unconditionally (even though a 955 // promotion failure does not swap spaces) because an unknown number of young 956 // collections will have swapped the spaces an unknown number of times. 957 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 958 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 959 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 960 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 961 962 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 963 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 964 965 // Increment the invocation count 966 heap->increment_total_collections(true); 967 968 // We need to track unique mark sweep invocations as well. 969 _total_invocations++; 970 971 heap->print_heap_before_gc(); 972 heap->trace_heap_before_gc(&_gc_tracer); 973 974 // Fill in TLABs 975 heap->ensure_parsability(true); // retire TLABs 976 977 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 978 HandleMark hm; // Discard invalid handles created during verification 979 Universe::verify("Before GC"); 980 } 981 982 // Verify object start arrays 983 if (VerifyObjectStartArray && 984 VerifyBeforeGC) { 985 heap->old_gen()->verify_object_start_array(); 986 } 987 988 DEBUG_ONLY(mark_bitmap()->verify_clear();) 989 DEBUG_ONLY(summary_data().verify_clear();) 990 991 // Have worker threads release resources the next time they run a task. 992 gc_task_manager()->release_all_resources(); 993 994 ParCompactionManager::reset_all_bitmap_query_caches(); 995 } 996 997 void PSParallelCompact::post_compact() 998 { 999 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1000 1001 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1002 // Clear the marking bitmap, summary data and split info. 1003 clear_data_covering_space(SpaceId(id)); 1004 // Update top(). Must be done after clearing the bitmap and summary data. 1005 _space_info[id].publish_new_top(); 1006 } 1007 1008 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1009 MutableSpace* const from_space = _space_info[from_space_id].space(); 1010 MutableSpace* const to_space = _space_info[to_space_id].space(); 1011 1012 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1013 bool eden_empty = eden_space->is_empty(); 1014 if (!eden_empty) { 1015 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1016 heap->young_gen(), heap->old_gen()); 1017 } 1018 1019 // Update heap occupancy information which is used as input to the soft ref 1020 // clearing policy at the next gc. 1021 Universe::update_heap_info_at_gc(); 1022 1023 bool young_gen_empty = eden_empty && from_space->is_empty() && 1024 to_space->is_empty(); 1025 1026 PSCardTable* ct = heap->card_table(); 1027 MemRegion old_mr = heap->old_gen()->reserved(); 1028 if (young_gen_empty) { 1029 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1030 } else { 1031 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1032 } 1033 1034 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1035 ClassLoaderDataGraph::purge(); 1036 MetaspaceUtils::verify_metrics(); 1037 1038 CodeCache::gc_epilogue(); 1039 JvmtiExport::gc_epilogue(); 1040 1041 #if COMPILER2_OR_JVMCI 1042 DerivedPointerTable::update_pointers(); 1043 #endif 1044 1045 if (ZapUnusedHeapArea) { 1046 heap->gen_mangle_unused_area(); 1047 } 1048 1049 // Update time of last GC 1050 reset_millis_since_last_gc(); 1051 } 1052 1053 HeapWord* 1054 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1055 bool maximum_compaction) 1056 { 1057 const size_t region_size = ParallelCompactData::RegionSize; 1058 const ParallelCompactData& sd = summary_data(); 1059 1060 const MutableSpace* const space = _space_info[id].space(); 1061 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1062 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1063 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1064 1065 // Skip full regions at the beginning of the space--they are necessarily part 1066 // of the dense prefix. 1067 size_t full_count = 0; 1068 const RegionData* cp; 1069 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1070 ++full_count; 1071 } 1072 1073 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1074 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1075 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1076 if (maximum_compaction || cp == end_cp || interval_ended) { 1077 _maximum_compaction_gc_num = total_invocations(); 1078 return sd.region_to_addr(cp); 1079 } 1080 1081 HeapWord* const new_top = _space_info[id].new_top(); 1082 const size_t space_live = pointer_delta(new_top, space->bottom()); 1083 const size_t space_used = space->used_in_words(); 1084 const size_t space_capacity = space->capacity_in_words(); 1085 1086 const double cur_density = double(space_live) / space_capacity; 1087 const double deadwood_density = 1088 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1089 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1090 1091 if (TraceParallelOldGCDensePrefix) { 1092 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1093 cur_density, deadwood_density, deadwood_goal); 1094 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1095 "space_cap=" SIZE_FORMAT, 1096 space_live, space_used, 1097 space_capacity); 1098 } 1099 1100 // XXX - Use binary search? 1101 HeapWord* dense_prefix = sd.region_to_addr(cp); 1102 const RegionData* full_cp = cp; 1103 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1104 while (cp < end_cp) { 1105 HeapWord* region_destination = cp->destination(); 1106 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1107 if (TraceParallelOldGCDensePrefix && Verbose) { 1108 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1109 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1110 sd.region(cp), p2i(region_destination), 1111 p2i(dense_prefix), cur_deadwood); 1112 } 1113 1114 if (cur_deadwood >= deadwood_goal) { 1115 // Found the region that has the correct amount of deadwood to the left. 1116 // This typically occurs after crossing a fairly sparse set of regions, so 1117 // iterate backwards over those sparse regions, looking for the region 1118 // that has the lowest density of live objects 'to the right.' 1119 size_t space_to_left = sd.region(cp) * region_size; 1120 size_t live_to_left = space_to_left - cur_deadwood; 1121 size_t space_to_right = space_capacity - space_to_left; 1122 size_t live_to_right = space_live - live_to_left; 1123 double density_to_right = double(live_to_right) / space_to_right; 1124 while (cp > full_cp) { 1125 --cp; 1126 const size_t prev_region_live_to_right = live_to_right - 1127 cp->data_size(); 1128 const size_t prev_region_space_to_right = space_to_right + region_size; 1129 double prev_region_density_to_right = 1130 double(prev_region_live_to_right) / prev_region_space_to_right; 1131 if (density_to_right <= prev_region_density_to_right) { 1132 return dense_prefix; 1133 } 1134 if (TraceParallelOldGCDensePrefix && Verbose) { 1135 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1136 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1137 prev_region_density_to_right); 1138 } 1139 dense_prefix -= region_size; 1140 live_to_right = prev_region_live_to_right; 1141 space_to_right = prev_region_space_to_right; 1142 density_to_right = prev_region_density_to_right; 1143 } 1144 return dense_prefix; 1145 } 1146 1147 dense_prefix += region_size; 1148 ++cp; 1149 } 1150 1151 return dense_prefix; 1152 } 1153 1154 #ifndef PRODUCT 1155 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1156 const SpaceId id, 1157 const bool maximum_compaction, 1158 HeapWord* const addr) 1159 { 1160 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1161 RegionData* const cp = summary_data().region(region_idx); 1162 const MutableSpace* const space = _space_info[id].space(); 1163 HeapWord* const new_top = _space_info[id].new_top(); 1164 1165 const size_t space_live = pointer_delta(new_top, space->bottom()); 1166 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1167 const size_t space_cap = space->capacity_in_words(); 1168 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1169 const size_t live_to_right = new_top - cp->destination(); 1170 const size_t dead_to_right = space->top() - addr - live_to_right; 1171 1172 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1173 "spl=" SIZE_FORMAT " " 1174 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1175 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1176 " ratio=%10.8f", 1177 algorithm, p2i(addr), region_idx, 1178 space_live, 1179 dead_to_left, dead_to_left_pct, 1180 dead_to_right, live_to_right, 1181 double(dead_to_right) / live_to_right); 1182 } 1183 #endif // #ifndef PRODUCT 1184 1185 // Return a fraction indicating how much of the generation can be treated as 1186 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1187 // based on the density of live objects in the generation to determine a limit, 1188 // which is then adjusted so the return value is min_percent when the density is 1189 // 1. 1190 // 1191 // The following table shows some return values for a different values of the 1192 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1193 // min_percent is 1. 1194 // 1195 // fraction allowed as dead wood 1196 // ----------------------------------------------------------------- 1197 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1198 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1199 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1200 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1201 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1202 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1203 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1204 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1205 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1206 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1207 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1208 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1209 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1210 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1211 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1212 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1213 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1214 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1215 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1216 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1217 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1218 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1219 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1220 1221 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1222 { 1223 assert(_dwl_initialized, "uninitialized"); 1224 1225 // The raw limit is the value of the normal distribution at x = density. 1226 const double raw_limit = normal_distribution(density); 1227 1228 // Adjust the raw limit so it becomes the minimum when the density is 1. 1229 // 1230 // First subtract the adjustment value (which is simply the precomputed value 1231 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1232 // Then add the minimum value, so the minimum is returned when the density is 1233 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1234 const double min = double(min_percent) / 100.0; 1235 const double limit = raw_limit - _dwl_adjustment + min; 1236 return MAX2(limit, 0.0); 1237 } 1238 1239 ParallelCompactData::RegionData* 1240 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1241 const RegionData* end) 1242 { 1243 const size_t region_size = ParallelCompactData::RegionSize; 1244 ParallelCompactData& sd = summary_data(); 1245 size_t left = sd.region(beg); 1246 size_t right = end > beg ? sd.region(end) - 1 : left; 1247 1248 // Binary search. 1249 while (left < right) { 1250 // Equivalent to (left + right) / 2, but does not overflow. 1251 const size_t middle = left + (right - left) / 2; 1252 RegionData* const middle_ptr = sd.region(middle); 1253 HeapWord* const dest = middle_ptr->destination(); 1254 HeapWord* const addr = sd.region_to_addr(middle); 1255 assert(dest != NULL, "sanity"); 1256 assert(dest <= addr, "must move left"); 1257 1258 if (middle > left && dest < addr) { 1259 right = middle - 1; 1260 } else if (middle < right && middle_ptr->data_size() == region_size) { 1261 left = middle + 1; 1262 } else { 1263 return middle_ptr; 1264 } 1265 } 1266 return sd.region(left); 1267 } 1268 1269 ParallelCompactData::RegionData* 1270 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1271 const RegionData* end, 1272 size_t dead_words) 1273 { 1274 ParallelCompactData& sd = summary_data(); 1275 size_t left = sd.region(beg); 1276 size_t right = end > beg ? sd.region(end) - 1 : left; 1277 1278 // Binary search. 1279 while (left < right) { 1280 // Equivalent to (left + right) / 2, but does not overflow. 1281 const size_t middle = left + (right - left) / 2; 1282 RegionData* const middle_ptr = sd.region(middle); 1283 HeapWord* const dest = middle_ptr->destination(); 1284 HeapWord* const addr = sd.region_to_addr(middle); 1285 assert(dest != NULL, "sanity"); 1286 assert(dest <= addr, "must move left"); 1287 1288 const size_t dead_to_left = pointer_delta(addr, dest); 1289 if (middle > left && dead_to_left > dead_words) { 1290 right = middle - 1; 1291 } else if (middle < right && dead_to_left < dead_words) { 1292 left = middle + 1; 1293 } else { 1294 return middle_ptr; 1295 } 1296 } 1297 return sd.region(left); 1298 } 1299 1300 // The result is valid during the summary phase, after the initial summarization 1301 // of each space into itself, and before final summarization. 1302 inline double 1303 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1304 HeapWord* const bottom, 1305 HeapWord* const top, 1306 HeapWord* const new_top) 1307 { 1308 ParallelCompactData& sd = summary_data(); 1309 1310 assert(cp != NULL, "sanity"); 1311 assert(bottom != NULL, "sanity"); 1312 assert(top != NULL, "sanity"); 1313 assert(new_top != NULL, "sanity"); 1314 assert(top >= new_top, "summary data problem?"); 1315 assert(new_top > bottom, "space is empty; should not be here"); 1316 assert(new_top >= cp->destination(), "sanity"); 1317 assert(top >= sd.region_to_addr(cp), "sanity"); 1318 1319 HeapWord* const destination = cp->destination(); 1320 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1321 const size_t compacted_region_live = pointer_delta(new_top, destination); 1322 const size_t compacted_region_used = pointer_delta(top, 1323 sd.region_to_addr(cp)); 1324 const size_t reclaimable = compacted_region_used - compacted_region_live; 1325 1326 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1327 return double(reclaimable) / divisor; 1328 } 1329 1330 // Return the address of the end of the dense prefix, a.k.a. the start of the 1331 // compacted region. The address is always on a region boundary. 1332 // 1333 // Completely full regions at the left are skipped, since no compaction can 1334 // occur in those regions. Then the maximum amount of dead wood to allow is 1335 // computed, based on the density (amount live / capacity) of the generation; 1336 // the region with approximately that amount of dead space to the left is 1337 // identified as the limit region. Regions between the last completely full 1338 // region and the limit region are scanned and the one that has the best 1339 // (maximum) reclaimed_ratio() is selected. 1340 HeapWord* 1341 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1342 bool maximum_compaction) 1343 { 1344 const size_t region_size = ParallelCompactData::RegionSize; 1345 const ParallelCompactData& sd = summary_data(); 1346 1347 const MutableSpace* const space = _space_info[id].space(); 1348 HeapWord* const top = space->top(); 1349 HeapWord* const top_aligned_up = sd.region_align_up(top); 1350 HeapWord* const new_top = _space_info[id].new_top(); 1351 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1352 HeapWord* const bottom = space->bottom(); 1353 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1354 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1355 const RegionData* const new_top_cp = 1356 sd.addr_to_region_ptr(new_top_aligned_up); 1357 1358 // Skip full regions at the beginning of the space--they are necessarily part 1359 // of the dense prefix. 1360 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1361 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1362 space->is_empty(), "no dead space allowed to the left"); 1363 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1364 "region must have dead space"); 1365 1366 // The gc number is saved whenever a maximum compaction is done, and used to 1367 // determine when the maximum compaction interval has expired. This avoids 1368 // successive max compactions for different reasons. 1369 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1370 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1371 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1372 total_invocations() == HeapFirstMaximumCompactionCount; 1373 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1374 _maximum_compaction_gc_num = total_invocations(); 1375 return sd.region_to_addr(full_cp); 1376 } 1377 1378 const size_t space_live = pointer_delta(new_top, bottom); 1379 const size_t space_used = space->used_in_words(); 1380 const size_t space_capacity = space->capacity_in_words(); 1381 1382 const double density = double(space_live) / double(space_capacity); 1383 const size_t min_percent_free = MarkSweepDeadRatio; 1384 const double limiter = dead_wood_limiter(density, min_percent_free); 1385 const size_t dead_wood_max = space_used - space_live; 1386 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1387 dead_wood_max); 1388 1389 if (TraceParallelOldGCDensePrefix) { 1390 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1391 "space_cap=" SIZE_FORMAT, 1392 space_live, space_used, 1393 space_capacity); 1394 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1395 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1396 density, min_percent_free, limiter, 1397 dead_wood_max, dead_wood_limit); 1398 } 1399 1400 // Locate the region with the desired amount of dead space to the left. 1401 const RegionData* const limit_cp = 1402 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1403 1404 // Scan from the first region with dead space to the limit region and find the 1405 // one with the best (largest) reclaimed ratio. 1406 double best_ratio = 0.0; 1407 const RegionData* best_cp = full_cp; 1408 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1409 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1410 if (tmp_ratio > best_ratio) { 1411 best_cp = cp; 1412 best_ratio = tmp_ratio; 1413 } 1414 } 1415 1416 return sd.region_to_addr(best_cp); 1417 } 1418 1419 void PSParallelCompact::summarize_spaces_quick() 1420 { 1421 for (unsigned int i = 0; i < last_space_id; ++i) { 1422 const MutableSpace* space = _space_info[i].space(); 1423 HeapWord** nta = _space_info[i].new_top_addr(); 1424 bool result = _summary_data.summarize(_space_info[i].split_info(), 1425 space->bottom(), space->top(), NULL, 1426 space->bottom(), space->end(), nta); 1427 assert(result, "space must fit into itself"); 1428 _space_info[i].set_dense_prefix(space->bottom()); 1429 } 1430 } 1431 1432 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1433 { 1434 HeapWord* const dense_prefix_end = dense_prefix(id); 1435 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1436 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1437 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1438 // Only enough dead space is filled so that any remaining dead space to the 1439 // left is larger than the minimum filler object. (The remainder is filled 1440 // during the copy/update phase.) 1441 // 1442 // The size of the dead space to the right of the boundary is not a 1443 // concern, since compaction will be able to use whatever space is 1444 // available. 1445 // 1446 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1447 // surrounds the space to be filled with an object. 1448 // 1449 // In the 32-bit VM, each bit represents two 32-bit words: 1450 // +---+ 1451 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1452 // end_bits: ... x x x | 0 | || 0 x x ... 1453 // +---+ 1454 // 1455 // In the 64-bit VM, each bit represents one 64-bit word: 1456 // +------------+ 1457 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1458 // end_bits: ... x x 1 | 0 || 0 | x x ... 1459 // +------------+ 1460 // +-------+ 1461 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1462 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1463 // +-------+ 1464 // +-----------+ 1465 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1466 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1467 // +-----------+ 1468 // +-------+ 1469 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1470 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1471 // +-------+ 1472 1473 // Initially assume case a, c or e will apply. 1474 size_t obj_len = CollectedHeap::min_fill_size(); 1475 HeapWord* obj_beg = dense_prefix_end - obj_len; 1476 1477 #ifdef _LP64 1478 if (MinObjAlignment > 1) { // object alignment > heap word size 1479 // Cases a, c or e. 1480 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1481 // Case b above. 1482 obj_beg = dense_prefix_end - 1; 1483 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1484 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1485 // Case d above. 1486 obj_beg = dense_prefix_end - 3; 1487 obj_len = 3; 1488 } 1489 #endif // #ifdef _LP64 1490 1491 CollectedHeap::fill_with_object(obj_beg, obj_len); 1492 _mark_bitmap.mark_obj(obj_beg, obj_len); 1493 _summary_data.add_obj(obj_beg, obj_len); 1494 assert(start_array(id) != NULL, "sanity"); 1495 start_array(id)->allocate_block(obj_beg); 1496 } 1497 } 1498 1499 void 1500 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1501 { 1502 assert(id < last_space_id, "id out of range"); 1503 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1504 "should have been reset in summarize_spaces_quick()"); 1505 1506 const MutableSpace* space = _space_info[id].space(); 1507 if (_space_info[id].new_top() != space->bottom()) { 1508 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1509 _space_info[id].set_dense_prefix(dense_prefix_end); 1510 1511 #ifndef PRODUCT 1512 if (TraceParallelOldGCDensePrefix) { 1513 print_dense_prefix_stats("ratio", id, maximum_compaction, 1514 dense_prefix_end); 1515 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1516 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1517 } 1518 #endif // #ifndef PRODUCT 1519 1520 // Recompute the summary data, taking into account the dense prefix. If 1521 // every last byte will be reclaimed, then the existing summary data which 1522 // compacts everything can be left in place. 1523 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1524 // If dead space crosses the dense prefix boundary, it is (at least 1525 // partially) filled with a dummy object, marked live and added to the 1526 // summary data. This simplifies the copy/update phase and must be done 1527 // before the final locations of objects are determined, to prevent 1528 // leaving a fragment of dead space that is too small to fill. 1529 fill_dense_prefix_end(id); 1530 1531 // Compute the destination of each Region, and thus each object. 1532 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1533 _summary_data.summarize(_space_info[id].split_info(), 1534 dense_prefix_end, space->top(), NULL, 1535 dense_prefix_end, space->end(), 1536 _space_info[id].new_top_addr()); 1537 } 1538 } 1539 1540 if (log_develop_is_enabled(Trace, gc, compaction)) { 1541 const size_t region_size = ParallelCompactData::RegionSize; 1542 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1543 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1544 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1545 HeapWord* const new_top = _space_info[id].new_top(); 1546 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1547 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1548 log_develop_trace(gc, compaction)( 1549 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1550 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1551 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1552 id, space->capacity_in_words(), p2i(dense_prefix_end), 1553 dp_region, dp_words / region_size, 1554 cr_words / region_size, p2i(new_top)); 1555 } 1556 } 1557 1558 #ifndef PRODUCT 1559 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1560 HeapWord* dst_beg, HeapWord* dst_end, 1561 SpaceId src_space_id, 1562 HeapWord* src_beg, HeapWord* src_end) 1563 { 1564 log_develop_trace(gc, compaction)( 1565 "Summarizing %d [%s] into %d [%s]: " 1566 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1567 SIZE_FORMAT "-" SIZE_FORMAT " " 1568 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1569 SIZE_FORMAT "-" SIZE_FORMAT, 1570 src_space_id, space_names[src_space_id], 1571 dst_space_id, space_names[dst_space_id], 1572 p2i(src_beg), p2i(src_end), 1573 _summary_data.addr_to_region_idx(src_beg), 1574 _summary_data.addr_to_region_idx(src_end), 1575 p2i(dst_beg), p2i(dst_end), 1576 _summary_data.addr_to_region_idx(dst_beg), 1577 _summary_data.addr_to_region_idx(dst_end)); 1578 } 1579 #endif // #ifndef PRODUCT 1580 1581 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1582 bool maximum_compaction) 1583 { 1584 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1585 1586 #ifdef ASSERT 1587 if (TraceParallelOldGCMarkingPhase) { 1588 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1589 "add_obj_bytes=" SIZE_FORMAT, 1590 add_obj_count, add_obj_size * HeapWordSize); 1591 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1592 "mark_bitmap_bytes=" SIZE_FORMAT, 1593 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1594 } 1595 #endif // #ifdef ASSERT 1596 1597 // Quick summarization of each space into itself, to see how much is live. 1598 summarize_spaces_quick(); 1599 1600 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1601 NOT_PRODUCT(print_region_ranges()); 1602 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1603 1604 // The amount of live data that will end up in old space (assuming it fits). 1605 size_t old_space_total_live = 0; 1606 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1607 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1608 _space_info[id].space()->bottom()); 1609 } 1610 1611 MutableSpace* const old_space = _space_info[old_space_id].space(); 1612 const size_t old_capacity = old_space->capacity_in_words(); 1613 if (old_space_total_live > old_capacity) { 1614 // XXX - should also try to expand 1615 maximum_compaction = true; 1616 } 1617 1618 // Old generations. 1619 summarize_space(old_space_id, maximum_compaction); 1620 1621 // Summarize the remaining spaces in the young gen. The initial target space 1622 // is the old gen. If a space does not fit entirely into the target, then the 1623 // remainder is compacted into the space itself and that space becomes the new 1624 // target. 1625 SpaceId dst_space_id = old_space_id; 1626 HeapWord* dst_space_end = old_space->end(); 1627 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1628 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1629 const MutableSpace* space = _space_info[id].space(); 1630 const size_t live = pointer_delta(_space_info[id].new_top(), 1631 space->bottom()); 1632 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1633 1634 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1635 SpaceId(id), space->bottom(), space->top());) 1636 if (live > 0 && live <= available) { 1637 // All the live data will fit. 1638 bool done = _summary_data.summarize(_space_info[id].split_info(), 1639 space->bottom(), space->top(), 1640 NULL, 1641 *new_top_addr, dst_space_end, 1642 new_top_addr); 1643 assert(done, "space must fit into old gen"); 1644 1645 // Reset the new_top value for the space. 1646 _space_info[id].set_new_top(space->bottom()); 1647 } else if (live > 0) { 1648 // Attempt to fit part of the source space into the target space. 1649 HeapWord* next_src_addr = NULL; 1650 bool done = _summary_data.summarize(_space_info[id].split_info(), 1651 space->bottom(), space->top(), 1652 &next_src_addr, 1653 *new_top_addr, dst_space_end, 1654 new_top_addr); 1655 assert(!done, "space should not fit into old gen"); 1656 assert(next_src_addr != NULL, "sanity"); 1657 1658 // The source space becomes the new target, so the remainder is compacted 1659 // within the space itself. 1660 dst_space_id = SpaceId(id); 1661 dst_space_end = space->end(); 1662 new_top_addr = _space_info[id].new_top_addr(); 1663 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1664 space->bottom(), dst_space_end, 1665 SpaceId(id), next_src_addr, space->top());) 1666 done = _summary_data.summarize(_space_info[id].split_info(), 1667 next_src_addr, space->top(), 1668 NULL, 1669 space->bottom(), dst_space_end, 1670 new_top_addr); 1671 assert(done, "space must fit when compacted into itself"); 1672 assert(*new_top_addr <= space->top(), "usage should not grow"); 1673 } 1674 } 1675 1676 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1677 NOT_PRODUCT(print_region_ranges()); 1678 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1679 } 1680 1681 // This method should contain all heap-specific policy for invoking a full 1682 // collection. invoke_no_policy() will only attempt to compact the heap; it 1683 // will do nothing further. If we need to bail out for policy reasons, scavenge 1684 // before full gc, or any other specialized behavior, it needs to be added here. 1685 // 1686 // Note that this method should only be called from the vm_thread while at a 1687 // safepoint. 1688 // 1689 // Note that the all_soft_refs_clear flag in the collector policy 1690 // may be true because this method can be called without intervening 1691 // activity. For example when the heap space is tight and full measure 1692 // are being taken to free space. 1693 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1694 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1695 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1696 "should be in vm thread"); 1697 1698 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1699 GCCause::Cause gc_cause = heap->gc_cause(); 1700 assert(!heap->is_gc_active(), "not reentrant"); 1701 1702 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1703 IsGCActiveMark mark; 1704 1705 if (ScavengeBeforeFullGC) { 1706 PSScavenge::invoke_no_policy(); 1707 } 1708 1709 const bool clear_all_soft_refs = 1710 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1711 1712 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1713 maximum_heap_compaction); 1714 } 1715 1716 // This method contains no policy. You should probably 1717 // be calling invoke() instead. 1718 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1719 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1720 assert(ref_processor() != NULL, "Sanity"); 1721 1722 if (GCLocker::check_active_before_gc()) { 1723 return false; 1724 } 1725 1726 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1727 1728 GCIdMark gc_id_mark; 1729 _gc_timer.register_gc_start(); 1730 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1731 1732 TimeStamp marking_start; 1733 TimeStamp compaction_start; 1734 TimeStamp collection_exit; 1735 1736 GCCause::Cause gc_cause = heap->gc_cause(); 1737 PSYoungGen* young_gen = heap->young_gen(); 1738 PSOldGen* old_gen = heap->old_gen(); 1739 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1740 1741 // The scope of casr should end after code that can change 1742 // CollectorPolicy::_should_clear_all_soft_refs. 1743 ClearedAllSoftRefs casr(maximum_heap_compaction, 1744 heap->soft_ref_policy()); 1745 1746 if (ZapUnusedHeapArea) { 1747 // Save information needed to minimize mangling 1748 heap->record_gen_tops_before_GC(); 1749 } 1750 1751 // Make sure data structures are sane, make the heap parsable, and do other 1752 // miscellaneous bookkeeping. 1753 pre_compact(); 1754 1755 PreGCValues pre_gc_values(heap); 1756 1757 // Get the compaction manager reserved for the VM thread. 1758 ParCompactionManager* const vmthread_cm = 1759 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1760 1761 { 1762 ResourceMark rm; 1763 HandleMark hm; 1764 1765 // Set the number of GC threads to be used in this collection 1766 gc_task_manager()->set_active_gang(); 1767 gc_task_manager()->task_idle_workers(); 1768 1769 GCTraceCPUTime tcpu; 1770 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1771 1772 heap->pre_full_gc_dump(&_gc_timer); 1773 1774 TraceCollectorStats tcs(counters()); 1775 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1776 1777 if (log_is_enabled(Debug, gc, heap, exit)) { 1778 accumulated_time()->start(); 1779 } 1780 1781 // Let the size policy know we're starting 1782 size_policy->major_collection_begin(); 1783 1784 CodeCache::gc_prologue(); 1785 1786 #if COMPILER2_OR_JVMCI 1787 DerivedPointerTable::clear(); 1788 #endif 1789 1790 ref_processor()->enable_discovery(); 1791 ref_processor()->setup_policy(maximum_heap_compaction); 1792 1793 bool marked_for_unloading = false; 1794 1795 marking_start.update(); 1796 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1797 1798 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1799 && GCCause::is_user_requested_gc(gc_cause); 1800 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1801 1802 #if COMPILER2_OR_JVMCI 1803 assert(DerivedPointerTable::is_active(), "Sanity"); 1804 DerivedPointerTable::set_active(false); 1805 #endif 1806 1807 // adjust_roots() updates Universe::_intArrayKlassObj which is 1808 // needed by the compaction for filling holes in the dense prefix. 1809 adjust_roots(vmthread_cm); 1810 1811 compaction_start.update(); 1812 compact(); 1813 1814 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1815 // done before resizing. 1816 post_compact(); 1817 1818 // Let the size policy know we're done 1819 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1820 1821 if (UseAdaptiveSizePolicy) { 1822 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1823 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1824 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1825 1826 // Don't check if the size_policy is ready here. Let 1827 // the size_policy check that internally. 1828 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1829 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1830 // Swap the survivor spaces if from_space is empty. The 1831 // resize_young_gen() called below is normally used after 1832 // a successful young GC and swapping of survivor spaces; 1833 // otherwise, it will fail to resize the young gen with 1834 // the current implementation. 1835 if (young_gen->from_space()->is_empty()) { 1836 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1837 young_gen->swap_spaces(); 1838 } 1839 1840 // Calculate optimal free space amounts 1841 assert(young_gen->max_size() > 1842 young_gen->from_space()->capacity_in_bytes() + 1843 young_gen->to_space()->capacity_in_bytes(), 1844 "Sizes of space in young gen are out-of-bounds"); 1845 1846 size_t young_live = young_gen->used_in_bytes(); 1847 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1848 size_t old_live = old_gen->used_in_bytes(); 1849 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1850 size_t max_old_gen_size = old_gen->max_gen_size(); 1851 size_t max_eden_size = young_gen->max_size() - 1852 young_gen->from_space()->capacity_in_bytes() - 1853 young_gen->to_space()->capacity_in_bytes(); 1854 1855 // Used for diagnostics 1856 size_policy->clear_generation_free_space_flags(); 1857 1858 size_policy->compute_generations_free_space(young_live, 1859 eden_live, 1860 old_live, 1861 cur_eden, 1862 max_old_gen_size, 1863 max_eden_size, 1864 true /* full gc*/); 1865 1866 size_policy->check_gc_overhead_limit(young_live, 1867 eden_live, 1868 max_old_gen_size, 1869 max_eden_size, 1870 true /* full gc*/, 1871 gc_cause, 1872 heap->soft_ref_policy()); 1873 1874 size_policy->decay_supplemental_growth(true /* full gc*/); 1875 1876 heap->resize_old_gen( 1877 size_policy->calculated_old_free_size_in_bytes()); 1878 1879 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1880 size_policy->calculated_survivor_size_in_bytes()); 1881 } 1882 1883 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1884 } 1885 1886 if (UsePerfData) { 1887 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1888 counters->update_counters(); 1889 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1890 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1891 } 1892 1893 heap->resize_all_tlabs(); 1894 1895 // Resize the metaspace capacity after a collection 1896 MetaspaceGC::compute_new_size(); 1897 1898 if (log_is_enabled(Debug, gc, heap, exit)) { 1899 accumulated_time()->stop(); 1900 } 1901 1902 young_gen->print_used_change(pre_gc_values.young_gen_used()); 1903 old_gen->print_used_change(pre_gc_values.old_gen_used()); 1904 MetaspaceUtils::print_metaspace_change(pre_gc_values.metadata_used()); 1905 1906 // Track memory usage and detect low memory 1907 MemoryService::track_memory_usage(); 1908 heap->update_counters(); 1909 gc_task_manager()->release_idle_workers(); 1910 1911 heap->post_full_gc_dump(&_gc_timer); 1912 } 1913 1914 #ifdef ASSERT 1915 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1916 ParCompactionManager* const cm = 1917 ParCompactionManager::manager_array(int(i)); 1918 assert(cm->marking_stack()->is_empty(), "should be empty"); 1919 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1920 } 1921 #endif // ASSERT 1922 1923 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1924 HandleMark hm; // Discard invalid handles created during verification 1925 Universe::verify("After GC"); 1926 } 1927 1928 // Re-verify object start arrays 1929 if (VerifyObjectStartArray && 1930 VerifyAfterGC) { 1931 old_gen->verify_object_start_array(); 1932 } 1933 1934 if (ZapUnusedHeapArea) { 1935 old_gen->object_space()->check_mangled_unused_area_complete(); 1936 } 1937 1938 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1939 1940 collection_exit.update(); 1941 1942 heap->print_heap_after_gc(); 1943 heap->trace_heap_after_gc(&_gc_tracer); 1944 1945 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1946 marking_start.ticks(), compaction_start.ticks(), 1947 collection_exit.ticks()); 1948 gc_task_manager()->print_task_time_stamps(); 1949 1950 #ifdef TRACESPINNING 1951 ParallelTaskTerminator::print_termination_counts(); 1952 #endif 1953 1954 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1955 1956 _gc_timer.register_gc_end(); 1957 1958 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1959 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1960 1961 return true; 1962 } 1963 1964 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1965 PSYoungGen* young_gen, 1966 PSOldGen* old_gen) { 1967 MutableSpace* const eden_space = young_gen->eden_space(); 1968 assert(!eden_space->is_empty(), "eden must be non-empty"); 1969 assert(young_gen->virtual_space()->alignment() == 1970 old_gen->virtual_space()->alignment(), "alignments do not match"); 1971 1972 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 1973 return false; 1974 } 1975 1976 // Both generations must be completely committed. 1977 if (young_gen->virtual_space()->uncommitted_size() != 0) { 1978 return false; 1979 } 1980 if (old_gen->virtual_space()->uncommitted_size() != 0) { 1981 return false; 1982 } 1983 1984 // Figure out how much to take from eden. Include the average amount promoted 1985 // in the total; otherwise the next young gen GC will simply bail out to a 1986 // full GC. 1987 const size_t alignment = old_gen->virtual_space()->alignment(); 1988 const size_t eden_used = eden_space->used_in_bytes(); 1989 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 1990 const size_t absorb_size = align_up(eden_used + promoted, alignment); 1991 const size_t eden_capacity = eden_space->capacity_in_bytes(); 1992 1993 if (absorb_size >= eden_capacity) { 1994 return false; // Must leave some space in eden. 1995 } 1996 1997 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 1998 if (new_young_size < young_gen->min_gen_size()) { 1999 return false; // Respect young gen minimum size. 2000 } 2001 2002 log_trace(heap, ergo)(" absorbing " SIZE_FORMAT "K: " 2003 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2004 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2005 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2006 absorb_size / K, 2007 eden_capacity / K, (eden_capacity - absorb_size) / K, 2008 young_gen->from_space()->used_in_bytes() / K, 2009 young_gen->to_space()->used_in_bytes() / K, 2010 young_gen->capacity_in_bytes() / K, new_young_size / K); 2011 2012 // Fill the unused part of the old gen. 2013 MutableSpace* const old_space = old_gen->object_space(); 2014 HeapWord* const unused_start = old_space->top(); 2015 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2016 2017 if (unused_words > 0) { 2018 if (unused_words < CollectedHeap::min_fill_size()) { 2019 return false; // If the old gen cannot be filled, must give up. 2020 } 2021 CollectedHeap::fill_with_objects(unused_start, unused_words); 2022 } 2023 2024 // Take the live data from eden and set both top and end in the old gen to 2025 // eden top. (Need to set end because reset_after_change() mangles the region 2026 // from end to virtual_space->high() in debug builds). 2027 HeapWord* const new_top = eden_space->top(); 2028 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2029 absorb_size); 2030 young_gen->reset_after_change(); 2031 old_space->set_top(new_top); 2032 old_space->set_end(new_top); 2033 old_gen->reset_after_change(); 2034 2035 // Update the object start array for the filler object and the data from eden. 2036 ObjectStartArray* const start_array = old_gen->start_array(); 2037 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2038 start_array->allocate_block(p); 2039 } 2040 2041 // Could update the promoted average here, but it is not typically updated at 2042 // full GCs and the value to use is unclear. Something like 2043 // 2044 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2045 2046 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2047 return true; 2048 } 2049 2050 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2051 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2052 "shouldn't return NULL"); 2053 return ParallelScavengeHeap::gc_task_manager(); 2054 } 2055 2056 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 2057 private: 2058 GCTaskQueue* _q; 2059 2060 public: 2061 PCAddThreadRootsMarkingTaskClosure(GCTaskQueue* q) : _q(q) { } 2062 void do_thread(Thread* t) { 2063 _q->enqueue(new ThreadRootsMarkingTask(t)); 2064 } 2065 }; 2066 2067 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2068 bool maximum_heap_compaction, 2069 ParallelOldTracer *gc_tracer) { 2070 // Recursively traverse all live objects and mark them 2071 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2072 2073 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2074 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2075 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2076 TaskQueueSetSuper* qset = ParCompactionManager::stack_array(); 2077 ParallelTaskTerminator terminator(active_gc_threads, qset); 2078 2079 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm); 2080 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2081 2082 // Need new claim bits before marking starts. 2083 ClassLoaderDataGraph::clear_claimed_marks(); 2084 2085 { 2086 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2087 2088 ParallelScavengeHeap::ParStrongRootsScope psrs; 2089 2090 GCTaskQueue* q = GCTaskQueue::create(); 2091 2092 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2093 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2094 // We scan the thread roots in parallel 2095 PCAddThreadRootsMarkingTaskClosure cl(q); 2096 Threads::java_threads_and_vm_thread_do(&cl); 2097 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2098 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2099 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2100 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2101 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2102 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2103 2104 if (active_gc_threads > 1) { 2105 for (uint j = 0; j < active_gc_threads; j++) { 2106 q->enqueue(new StealMarkingTask(&terminator)); 2107 } 2108 } 2109 2110 gc_task_manager()->execute_and_wait(q); 2111 } 2112 2113 // Process reference objects found during marking 2114 { 2115 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2116 2117 ReferenceProcessorStats stats; 2118 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2119 2120 if (ref_processor()->processing_is_mt()) { 2121 ref_processor()->set_active_mt_degree(active_gc_threads); 2122 2123 RefProcTaskExecutor task_executor; 2124 stats = ref_processor()->process_discovered_references( 2125 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2126 &task_executor, &pt); 2127 } else { 2128 stats = ref_processor()->process_discovered_references( 2129 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2130 &pt); 2131 } 2132 2133 gc_tracer->report_gc_reference_stats(stats); 2134 pt.print_all_references(); 2135 } 2136 2137 // This is the point where the entire marking should have completed. 2138 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2139 2140 { 2141 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2142 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2143 } 2144 2145 { 2146 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2147 2148 // Follow system dictionary roots and unload classes. 2149 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2150 2151 // Unload nmethods. 2152 CodeCache::do_unloading(is_alive_closure(), purged_class); 2153 2154 // Prune dead klasses from subklass/sibling/implementor lists. 2155 Klass::clean_weak_klass_links(purged_class); 2156 } 2157 2158 { 2159 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer); 2160 // Delete entries for dead interned strings. 2161 StringTable::unlink(is_alive_closure()); 2162 } 2163 2164 { 2165 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer); 2166 // Clean up unreferenced symbols in symbol table. 2167 SymbolTable::unlink(); 2168 } 2169 2170 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2171 } 2172 2173 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2174 // Adjust the pointers to reflect the new locations 2175 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2176 2177 // Need new claim bits when tracing through and adjusting pointers. 2178 ClassLoaderDataGraph::clear_claimed_marks(); 2179 2180 PSParallelCompact::AdjustPointerClosure oop_closure(cm); 2181 2182 // General strong roots. 2183 Universe::oops_do(&oop_closure); 2184 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2185 Threads::oops_do(&oop_closure, NULL); 2186 ObjectSynchronizer::oops_do(&oop_closure); 2187 Management::oops_do(&oop_closure); 2188 JvmtiExport::oops_do(&oop_closure); 2189 SystemDictionary::oops_do(&oop_closure); 2190 CLDToOopClosure cld_closure(&oop_closure); 2191 ClassLoaderDataGraph::cld_do(&cld_closure); 2192 2193 // Now adjust pointers in remaining weak roots. (All of which should 2194 // have been cleared if they pointed to non-surviving objects.) 2195 WeakProcessor::oops_do(&oop_closure); 2196 2197 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2198 CodeCache::blobs_do(&adjust_from_blobs); 2199 AOTLoader::oops_do(&oop_closure); 2200 StringTable::oops_do(&oop_closure); 2201 ref_processor()->weak_oops_do(&oop_closure); 2202 // Roots were visited so references into the young gen in roots 2203 // may have been scanned. Process them also. 2204 // Should the reference processor have a span that excludes 2205 // young gen objects? 2206 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2207 } 2208 2209 // Helper class to print 8 region numbers per line and then print the total at the end. 2210 class FillableRegionLogger : public StackObj { 2211 private: 2212 Log(gc, compaction) log; 2213 static const int LineLength = 8; 2214 size_t _regions[LineLength]; 2215 int _next_index; 2216 bool _enabled; 2217 size_t _total_regions; 2218 public: 2219 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2220 ~FillableRegionLogger() { 2221 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2222 } 2223 2224 void print_line() { 2225 if (!_enabled || _next_index == 0) { 2226 return; 2227 } 2228 FormatBuffer<> line("Fillable: "); 2229 for (int i = 0; i < _next_index; i++) { 2230 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2231 } 2232 log.trace("%s", line.buffer()); 2233 _next_index = 0; 2234 } 2235 2236 void handle(size_t region) { 2237 if (!_enabled) { 2238 return; 2239 } 2240 _regions[_next_index++] = region; 2241 if (_next_index == LineLength) { 2242 print_line(); 2243 } 2244 _total_regions++; 2245 } 2246 }; 2247 2248 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q, 2249 uint parallel_gc_threads) 2250 { 2251 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2252 2253 // Find the threads that are active 2254 unsigned int which = 0; 2255 2256 // Find all regions that are available (can be filled immediately) and 2257 // distribute them to the thread stacks. The iteration is done in reverse 2258 // order (high to low) so the regions will be removed in ascending order. 2259 2260 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2261 2262 which = 0; 2263 // id + 1 is used to test termination so unsigned can 2264 // be used with an old_space_id == 0. 2265 FillableRegionLogger region_logger; 2266 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2267 SpaceInfo* const space_info = _space_info + id; 2268 MutableSpace* const space = space_info->space(); 2269 HeapWord* const new_top = space_info->new_top(); 2270 2271 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2272 const size_t end_region = 2273 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2274 2275 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2276 if (sd.region(cur)->claim_unsafe()) { 2277 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2278 cm->region_stack()->push(cur); 2279 region_logger.handle(cur); 2280 // Assign regions to tasks in round-robin fashion. 2281 if (++which == parallel_gc_threads) { 2282 which = 0; 2283 } 2284 } 2285 } 2286 region_logger.print_line(); 2287 } 2288 } 2289 2290 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2291 2292 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2293 uint parallel_gc_threads) { 2294 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2295 2296 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2297 2298 // Iterate over all the spaces adding tasks for updating 2299 // regions in the dense prefix. Assume that 1 gc thread 2300 // will work on opening the gaps and the remaining gc threads 2301 // will work on the dense prefix. 2302 unsigned int space_id; 2303 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2304 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2305 const MutableSpace* const space = _space_info[space_id].space(); 2306 2307 if (dense_prefix_end == space->bottom()) { 2308 // There is no dense prefix for this space. 2309 continue; 2310 } 2311 2312 // The dense prefix is before this region. 2313 size_t region_index_end_dense_prefix = 2314 sd.addr_to_region_idx(dense_prefix_end); 2315 RegionData* const dense_prefix_cp = 2316 sd.region(region_index_end_dense_prefix); 2317 assert(dense_prefix_end == space->end() || 2318 dense_prefix_cp->available() || 2319 dense_prefix_cp->claimed(), 2320 "The region after the dense prefix should always be ready to fill"); 2321 2322 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2323 2324 // Is there dense prefix work? 2325 size_t total_dense_prefix_regions = 2326 region_index_end_dense_prefix - region_index_start; 2327 // How many regions of the dense prefix should be given to 2328 // each thread? 2329 if (total_dense_prefix_regions > 0) { 2330 uint tasks_for_dense_prefix = 1; 2331 if (total_dense_prefix_regions <= 2332 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2333 // Don't over partition. This assumes that 2334 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2335 // so there are not many regions to process. 2336 tasks_for_dense_prefix = parallel_gc_threads; 2337 } else { 2338 // Over partition 2339 tasks_for_dense_prefix = parallel_gc_threads * 2340 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2341 } 2342 size_t regions_per_thread = total_dense_prefix_regions / 2343 tasks_for_dense_prefix; 2344 // Give each thread at least 1 region. 2345 if (regions_per_thread == 0) { 2346 regions_per_thread = 1; 2347 } 2348 2349 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2350 if (region_index_start >= region_index_end_dense_prefix) { 2351 break; 2352 } 2353 // region_index_end is not processed 2354 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2355 region_index_end_dense_prefix); 2356 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2357 region_index_start, 2358 region_index_end)); 2359 region_index_start = region_index_end; 2360 } 2361 } 2362 // This gets any part of the dense prefix that did not 2363 // fit evenly. 2364 if (region_index_start < region_index_end_dense_prefix) { 2365 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2366 region_index_start, 2367 region_index_end_dense_prefix)); 2368 } 2369 } 2370 } 2371 2372 void PSParallelCompact::enqueue_region_stealing_tasks( 2373 GCTaskQueue* q, 2374 ParallelTaskTerminator* terminator_ptr, 2375 uint parallel_gc_threads) { 2376 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer); 2377 2378 // Once a thread has drained it's stack, it should try to steal regions from 2379 // other threads. 2380 for (uint j = 0; j < parallel_gc_threads; j++) { 2381 q->enqueue(new CompactionWithStealingTask(terminator_ptr)); 2382 } 2383 } 2384 2385 #ifdef ASSERT 2386 // Write a histogram of the number of times the block table was filled for a 2387 // region. 2388 void PSParallelCompact::write_block_fill_histogram() 2389 { 2390 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2391 return; 2392 } 2393 2394 Log(gc, compaction) log; 2395 ResourceMark rm; 2396 LogStream ls(log.trace()); 2397 outputStream* out = &ls; 2398 2399 typedef ParallelCompactData::RegionData rd_t; 2400 ParallelCompactData& sd = summary_data(); 2401 2402 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2403 MutableSpace* const spc = _space_info[id].space(); 2404 if (spc->bottom() != spc->top()) { 2405 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2406 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2407 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2408 2409 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2410 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2411 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2412 2413 for (const rd_t* cur = beg; cur < end; ++cur) { 2414 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2415 } 2416 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2417 for (size_t i = 0; i < histo_len; ++i) { 2418 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2419 histo[i], 100.0 * histo[i] / region_cnt); 2420 } 2421 out->cr(); 2422 } 2423 } 2424 } 2425 #endif // #ifdef ASSERT 2426 2427 void PSParallelCompact::compact() { 2428 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2429 2430 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2431 PSOldGen* old_gen = heap->old_gen(); 2432 old_gen->start_array()->reset(); 2433 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2434 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2435 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2436 ParallelTaskTerminator terminator(active_gc_threads, qset); 2437 2438 GCTaskQueue* q = GCTaskQueue::create(); 2439 prepare_region_draining_tasks(q, active_gc_threads); 2440 enqueue_dense_prefix_tasks(q, active_gc_threads); 2441 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2442 2443 { 2444 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2445 2446 gc_task_manager()->execute_and_wait(q); 2447 2448 #ifdef ASSERT 2449 // Verify that all regions have been processed before the deferred updates. 2450 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2451 verify_complete(SpaceId(id)); 2452 } 2453 #endif 2454 } 2455 2456 { 2457 // Update the deferred objects, if any. Any compaction manager can be used. 2458 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2459 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2460 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2461 update_deferred_objects(cm, SpaceId(id)); 2462 } 2463 } 2464 2465 DEBUG_ONLY(write_block_fill_histogram()); 2466 } 2467 2468 #ifdef ASSERT 2469 void PSParallelCompact::verify_complete(SpaceId space_id) { 2470 // All Regions between space bottom() to new_top() should be marked as filled 2471 // and all Regions between new_top() and top() should be available (i.e., 2472 // should have been emptied). 2473 ParallelCompactData& sd = summary_data(); 2474 SpaceInfo si = _space_info[space_id]; 2475 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2476 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2477 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2478 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2479 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2480 2481 bool issued_a_warning = false; 2482 2483 size_t cur_region; 2484 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2485 const RegionData* const c = sd.region(cur_region); 2486 if (!c->completed()) { 2487 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2488 cur_region, c->destination_count()); 2489 issued_a_warning = true; 2490 } 2491 } 2492 2493 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2494 const RegionData* const c = sd.region(cur_region); 2495 if (!c->available()) { 2496 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2497 cur_region, c->destination_count()); 2498 issued_a_warning = true; 2499 } 2500 } 2501 2502 if (issued_a_warning) { 2503 print_region_ranges(); 2504 } 2505 } 2506 #endif // #ifdef ASSERT 2507 2508 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2509 _start_array->allocate_block(addr); 2510 compaction_manager()->update_contents(oop(addr)); 2511 } 2512 2513 // Update interior oops in the ranges of regions [beg_region, end_region). 2514 void 2515 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2516 SpaceId space_id, 2517 size_t beg_region, 2518 size_t end_region) { 2519 ParallelCompactData& sd = summary_data(); 2520 ParMarkBitMap* const mbm = mark_bitmap(); 2521 2522 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2523 HeapWord* const end_addr = sd.region_to_addr(end_region); 2524 assert(beg_region <= end_region, "bad region range"); 2525 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2526 2527 #ifdef ASSERT 2528 // Claim the regions to avoid triggering an assert when they are marked as 2529 // filled. 2530 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2531 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2532 } 2533 #endif // #ifdef ASSERT 2534 2535 if (beg_addr != space(space_id)->bottom()) { 2536 // Find the first live object or block of dead space that *starts* in this 2537 // range of regions. If a partial object crosses onto the region, skip it; 2538 // it will be marked for 'deferred update' when the object head is 2539 // processed. If dead space crosses onto the region, it is also skipped; it 2540 // will be filled when the prior region is processed. If neither of those 2541 // apply, the first word in the region is the start of a live object or dead 2542 // space. 2543 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2544 const RegionData* const cp = sd.region(beg_region); 2545 if (cp->partial_obj_size() != 0) { 2546 beg_addr = sd.partial_obj_end(beg_region); 2547 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2548 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2549 } 2550 } 2551 2552 if (beg_addr < end_addr) { 2553 // A live object or block of dead space starts in this range of Regions. 2554 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2555 2556 // Create closures and iterate. 2557 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2558 FillClosure fill_closure(cm, space_id); 2559 ParMarkBitMap::IterationStatus status; 2560 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2561 dense_prefix_end); 2562 if (status == ParMarkBitMap::incomplete) { 2563 update_closure.do_addr(update_closure.source()); 2564 } 2565 } 2566 2567 // Mark the regions as filled. 2568 RegionData* const beg_cp = sd.region(beg_region); 2569 RegionData* const end_cp = sd.region(end_region); 2570 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2571 cp->set_completed(); 2572 } 2573 } 2574 2575 // Return the SpaceId for the space containing addr. If addr is not in the 2576 // heap, last_space_id is returned. In debug mode it expects the address to be 2577 // in the heap and asserts such. 2578 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2579 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2580 2581 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2582 if (_space_info[id].space()->contains(addr)) { 2583 return SpaceId(id); 2584 } 2585 } 2586 2587 assert(false, "no space contains the addr"); 2588 return last_space_id; 2589 } 2590 2591 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2592 SpaceId id) { 2593 assert(id < last_space_id, "bad space id"); 2594 2595 ParallelCompactData& sd = summary_data(); 2596 const SpaceInfo* const space_info = _space_info + id; 2597 ObjectStartArray* const start_array = space_info->start_array(); 2598 2599 const MutableSpace* const space = space_info->space(); 2600 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2601 HeapWord* const beg_addr = space_info->dense_prefix(); 2602 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2603 2604 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2605 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2606 const RegionData* cur_region; 2607 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2608 HeapWord* const addr = cur_region->deferred_obj_addr(); 2609 if (addr != NULL) { 2610 if (start_array != NULL) { 2611 start_array->allocate_block(addr); 2612 } 2613 cm->update_contents(oop(addr)); 2614 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2615 } 2616 } 2617 } 2618 2619 // Skip over count live words starting from beg, and return the address of the 2620 // next live word. Unless marked, the word corresponding to beg is assumed to 2621 // be dead. Callers must either ensure beg does not correspond to the middle of 2622 // an object, or account for those live words in some other way. Callers must 2623 // also ensure that there are enough live words in the range [beg, end) to skip. 2624 HeapWord* 2625 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2626 { 2627 assert(count > 0, "sanity"); 2628 2629 ParMarkBitMap* m = mark_bitmap(); 2630 idx_t bits_to_skip = m->words_to_bits(count); 2631 idx_t cur_beg = m->addr_to_bit(beg); 2632 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2633 2634 do { 2635 cur_beg = m->find_obj_beg(cur_beg, search_end); 2636 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2637 const size_t obj_bits = cur_end - cur_beg + 1; 2638 if (obj_bits > bits_to_skip) { 2639 return m->bit_to_addr(cur_beg + bits_to_skip); 2640 } 2641 bits_to_skip -= obj_bits; 2642 cur_beg = cur_end + 1; 2643 } while (bits_to_skip > 0); 2644 2645 // Skipping the desired number of words landed just past the end of an object. 2646 // Find the start of the next object. 2647 cur_beg = m->find_obj_beg(cur_beg, search_end); 2648 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2649 return m->bit_to_addr(cur_beg); 2650 } 2651 2652 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2653 SpaceId src_space_id, 2654 size_t src_region_idx) 2655 { 2656 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2657 2658 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2659 if (split_info.dest_region_addr() == dest_addr) { 2660 // The partial object ending at the split point contains the first word to 2661 // be copied to dest_addr. 2662 return split_info.first_src_addr(); 2663 } 2664 2665 const ParallelCompactData& sd = summary_data(); 2666 ParMarkBitMap* const bitmap = mark_bitmap(); 2667 const size_t RegionSize = ParallelCompactData::RegionSize; 2668 2669 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2670 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2671 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2672 HeapWord* const src_region_destination = src_region_ptr->destination(); 2673 2674 assert(dest_addr >= src_region_destination, "wrong src region"); 2675 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2676 2677 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2678 HeapWord* const src_region_end = src_region_beg + RegionSize; 2679 2680 HeapWord* addr = src_region_beg; 2681 if (dest_addr == src_region_destination) { 2682 // Return the first live word in the source region. 2683 if (partial_obj_size == 0) { 2684 addr = bitmap->find_obj_beg(addr, src_region_end); 2685 assert(addr < src_region_end, "no objects start in src region"); 2686 } 2687 return addr; 2688 } 2689 2690 // Must skip some live data. 2691 size_t words_to_skip = dest_addr - src_region_destination; 2692 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2693 2694 if (partial_obj_size >= words_to_skip) { 2695 // All the live words to skip are part of the partial object. 2696 addr += words_to_skip; 2697 if (partial_obj_size == words_to_skip) { 2698 // Find the first live word past the partial object. 2699 addr = bitmap->find_obj_beg(addr, src_region_end); 2700 assert(addr < src_region_end, "wrong src region"); 2701 } 2702 return addr; 2703 } 2704 2705 // Skip over the partial object (if any). 2706 if (partial_obj_size != 0) { 2707 words_to_skip -= partial_obj_size; 2708 addr += partial_obj_size; 2709 } 2710 2711 // Skip over live words due to objects that start in the region. 2712 addr = skip_live_words(addr, src_region_end, words_to_skip); 2713 assert(addr < src_region_end, "wrong src region"); 2714 return addr; 2715 } 2716 2717 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2718 SpaceId src_space_id, 2719 size_t beg_region, 2720 HeapWord* end_addr) 2721 { 2722 ParallelCompactData& sd = summary_data(); 2723 2724 #ifdef ASSERT 2725 MutableSpace* const src_space = _space_info[src_space_id].space(); 2726 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2727 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2728 "src_space_id does not match beg_addr"); 2729 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2730 "src_space_id does not match end_addr"); 2731 #endif // #ifdef ASSERT 2732 2733 RegionData* const beg = sd.region(beg_region); 2734 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2735 2736 // Regions up to new_top() are enqueued if they become available. 2737 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2738 RegionData* const enqueue_end = 2739 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2740 2741 for (RegionData* cur = beg; cur < end; ++cur) { 2742 assert(cur->data_size() > 0, "region must have live data"); 2743 cur->decrement_destination_count(); 2744 if (cur < enqueue_end && cur->available() && cur->claim()) { 2745 cm->push_region(sd.region(cur)); 2746 } 2747 } 2748 } 2749 2750 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2751 SpaceId& src_space_id, 2752 HeapWord*& src_space_top, 2753 HeapWord* end_addr) 2754 { 2755 typedef ParallelCompactData::RegionData RegionData; 2756 2757 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2758 const size_t region_size = ParallelCompactData::RegionSize; 2759 2760 size_t src_region_idx = 0; 2761 2762 // Skip empty regions (if any) up to the top of the space. 2763 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2764 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2765 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2766 const RegionData* const top_region_ptr = 2767 sd.addr_to_region_ptr(top_aligned_up); 2768 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2769 ++src_region_ptr; 2770 } 2771 2772 if (src_region_ptr < top_region_ptr) { 2773 // The next source region is in the current space. Update src_region_idx 2774 // and the source address to match src_region_ptr. 2775 src_region_idx = sd.region(src_region_ptr); 2776 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2777 if (src_region_addr > closure.source()) { 2778 closure.set_source(src_region_addr); 2779 } 2780 return src_region_idx; 2781 } 2782 2783 // Switch to a new source space and find the first non-empty region. 2784 unsigned int space_id = src_space_id + 1; 2785 assert(space_id < last_space_id, "not enough spaces"); 2786 2787 HeapWord* const destination = closure.destination(); 2788 2789 do { 2790 MutableSpace* space = _space_info[space_id].space(); 2791 HeapWord* const bottom = space->bottom(); 2792 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2793 2794 // Iterate over the spaces that do not compact into themselves. 2795 if (bottom_cp->destination() != bottom) { 2796 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2797 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2798 2799 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2800 if (src_cp->live_obj_size() > 0) { 2801 // Found it. 2802 assert(src_cp->destination() == destination, 2803 "first live obj in the space must match the destination"); 2804 assert(src_cp->partial_obj_size() == 0, 2805 "a space cannot begin with a partial obj"); 2806 2807 src_space_id = SpaceId(space_id); 2808 src_space_top = space->top(); 2809 const size_t src_region_idx = sd.region(src_cp); 2810 closure.set_source(sd.region_to_addr(src_region_idx)); 2811 return src_region_idx; 2812 } else { 2813 assert(src_cp->data_size() == 0, "sanity"); 2814 } 2815 } 2816 } 2817 } while (++space_id < last_space_id); 2818 2819 assert(false, "no source region was found"); 2820 return 0; 2821 } 2822 2823 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 2824 { 2825 typedef ParMarkBitMap::IterationStatus IterationStatus; 2826 const size_t RegionSize = ParallelCompactData::RegionSize; 2827 ParMarkBitMap* const bitmap = mark_bitmap(); 2828 ParallelCompactData& sd = summary_data(); 2829 RegionData* const region_ptr = sd.region(region_idx); 2830 2831 // Get the items needed to construct the closure. 2832 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2833 SpaceId dest_space_id = space_id(dest_addr); 2834 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 2835 HeapWord* new_top = _space_info[dest_space_id].new_top(); 2836 assert(dest_addr < new_top, "sanity"); 2837 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 2838 2839 // Get the source region and related info. 2840 size_t src_region_idx = region_ptr->source_region(); 2841 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2842 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2843 2844 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2845 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2846 2847 // Adjust src_region_idx to prepare for decrementing destination counts (the 2848 // destination count is not decremented when a region is copied to itself). 2849 if (src_region_idx == region_idx) { 2850 src_region_idx += 1; 2851 } 2852 2853 if (bitmap->is_unmarked(closure.source())) { 2854 // The first source word is in the middle of an object; copy the remainder 2855 // of the object or as much as will fit. The fact that pointer updates were 2856 // deferred will be noted when the object header is processed. 2857 HeapWord* const old_src_addr = closure.source(); 2858 closure.copy_partial_obj(); 2859 if (closure.is_full()) { 2860 decrement_destination_counts(cm, src_space_id, src_region_idx, 2861 closure.source()); 2862 region_ptr->set_deferred_obj_addr(NULL); 2863 region_ptr->set_completed(); 2864 return; 2865 } 2866 2867 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2868 if (sd.region_align_down(old_src_addr) != end_addr) { 2869 // The partial object was copied from more than one source region. 2870 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2871 2872 // Move to the next source region, possibly switching spaces as well. All 2873 // args except end_addr may be modified. 2874 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2875 end_addr); 2876 } 2877 } 2878 2879 do { 2880 HeapWord* const cur_addr = closure.source(); 2881 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2882 src_space_top); 2883 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2884 2885 if (status == ParMarkBitMap::incomplete) { 2886 // The last obj that starts in the source region does not end in the 2887 // region. 2888 assert(closure.source() < end_addr, "sanity"); 2889 HeapWord* const obj_beg = closure.source(); 2890 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 2891 src_space_top); 2892 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 2893 if (obj_end < range_end) { 2894 // The end was found; the entire object will fit. 2895 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 2896 assert(status != ParMarkBitMap::would_overflow, "sanity"); 2897 } else { 2898 // The end was not found; the object will not fit. 2899 assert(range_end < src_space_top, "obj cannot cross space boundary"); 2900 status = ParMarkBitMap::would_overflow; 2901 } 2902 } 2903 2904 if (status == ParMarkBitMap::would_overflow) { 2905 // The last object did not fit. Note that interior oop updates were 2906 // deferred, then copy enough of the object to fill the region. 2907 region_ptr->set_deferred_obj_addr(closure.destination()); 2908 status = closure.copy_until_full(); // copies from closure.source() 2909 2910 decrement_destination_counts(cm, src_space_id, src_region_idx, 2911 closure.source()); 2912 region_ptr->set_completed(); 2913 return; 2914 } 2915 2916 if (status == ParMarkBitMap::full) { 2917 decrement_destination_counts(cm, src_space_id, src_region_idx, 2918 closure.source()); 2919 region_ptr->set_deferred_obj_addr(NULL); 2920 region_ptr->set_completed(); 2921 return; 2922 } 2923 2924 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2925 2926 // Move to the next source region, possibly switching spaces as well. All 2927 // args except end_addr may be modified. 2928 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2929 end_addr); 2930 } while (true); 2931 } 2932 2933 void PSParallelCompact::fill_blocks(size_t region_idx) 2934 { 2935 // Fill in the block table elements for the specified region. Each block 2936 // table element holds the number of live words in the region that are to the 2937 // left of the first object that starts in the block. Thus only blocks in 2938 // which an object starts need to be filled. 2939 // 2940 // The algorithm scans the section of the bitmap that corresponds to the 2941 // region, keeping a running total of the live words. When an object start is 2942 // found, if it's the first to start in the block that contains it, the 2943 // current total is written to the block table element. 2944 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2945 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2946 const size_t RegionSize = ParallelCompactData::RegionSize; 2947 2948 ParallelCompactData& sd = summary_data(); 2949 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2950 if (partial_obj_size >= RegionSize) { 2951 return; // No objects start in this region. 2952 } 2953 2954 // Ensure the first loop iteration decides that the block has changed. 2955 size_t cur_block = sd.block_count(); 2956 2957 const ParMarkBitMap* const bitmap = mark_bitmap(); 2958 2959 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2960 assert((size_t)1 << Log2BitsPerBlock == 2961 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2962 2963 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2964 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2965 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2966 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2967 while (beg_bit < range_end) { 2968 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2969 if (new_block != cur_block) { 2970 cur_block = new_block; 2971 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2972 } 2973 2974 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2975 if (end_bit < range_end - 1) { 2976 live_bits += end_bit - beg_bit + 1; 2977 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2978 } else { 2979 return; 2980 } 2981 } 2982 } 2983 2984 void 2985 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 2986 const MutableSpace* sp = space(space_id); 2987 if (sp->is_empty()) { 2988 return; 2989 } 2990 2991 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2992 ParMarkBitMap* const bitmap = mark_bitmap(); 2993 HeapWord* const dp_addr = dense_prefix(space_id); 2994 HeapWord* beg_addr = sp->bottom(); 2995 HeapWord* end_addr = sp->top(); 2996 2997 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 2998 2999 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3000 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3001 if (beg_region < dp_region) { 3002 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3003 } 3004 3005 // The destination of the first live object that starts in the region is one 3006 // past the end of the partial object entering the region (if any). 3007 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3008 HeapWord* const new_top = _space_info[space_id].new_top(); 3009 assert(new_top >= dest_addr, "bad new_top value"); 3010 const size_t words = pointer_delta(new_top, dest_addr); 3011 3012 if (words > 0) { 3013 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3014 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3015 3016 ParMarkBitMap::IterationStatus status; 3017 status = bitmap->iterate(&closure, dest_addr, end_addr); 3018 assert(status == ParMarkBitMap::full, "iteration not complete"); 3019 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3020 "live objects skipped because closure is full"); 3021 } 3022 } 3023 3024 jlong PSParallelCompact::millis_since_last_gc() { 3025 // We need a monotonically non-decreasing time in ms but 3026 // os::javaTimeMillis() does not guarantee monotonicity. 3027 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3028 jlong ret_val = now - _time_of_last_gc; 3029 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3030 if (ret_val < 0) { 3031 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3032 return 0; 3033 } 3034 return ret_val; 3035 } 3036 3037 void PSParallelCompact::reset_millis_since_last_gc() { 3038 // We need a monotonically non-decreasing time in ms but 3039 // os::javaTimeMillis() does not guarantee monotonicity. 3040 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3041 } 3042 3043 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3044 { 3045 if (source() != destination()) { 3046 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3047 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3048 } 3049 update_state(words_remaining()); 3050 assert(is_full(), "sanity"); 3051 return ParMarkBitMap::full; 3052 } 3053 3054 void MoveAndUpdateClosure::copy_partial_obj() 3055 { 3056 size_t words = words_remaining(); 3057 3058 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3059 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3060 if (end_addr < range_end) { 3061 words = bitmap()->obj_size(source(), end_addr); 3062 } 3063 3064 // This test is necessary; if omitted, the pointer updates to a partial object 3065 // that crosses the dense prefix boundary could be overwritten. 3066 if (source() != destination()) { 3067 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3068 Copy::aligned_conjoint_words(source(), destination(), words); 3069 } 3070 update_state(words); 3071 } 3072 3073 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3074 PSParallelCompact::AdjustPointerClosure closure(cm); 3075 if (UseCompressedOops) { 3076 oop_oop_iterate_oop_maps<narrowOop>(obj, &closure); 3077 } else { 3078 oop_oop_iterate_oop_maps<oop>(obj, &closure); 3079 } 3080 } 3081 3082 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3083 InstanceKlass::oop_pc_update_pointers(obj, cm); 3084 3085 PSParallelCompact::AdjustPointerClosure closure(cm); 3086 if (UseCompressedOops) { 3087 oop_oop_iterate_statics<narrowOop>(obj, &closure); 3088 } else { 3089 oop_oop_iterate_statics<oop>(obj, &closure); 3090 } 3091 } 3092 3093 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3094 InstanceKlass::oop_pc_update_pointers(obj, cm); 3095 } 3096 3097 #ifdef ASSERT 3098 template <class T> static void trace_reference_gc(const char *s, oop obj, 3099 T* referent_addr, 3100 T* discovered_addr) { 3101 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj)); 3102 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3103 p2i(referent_addr), referent_addr ? p2i((oop)RawAccess<>::oop_load(referent_addr)) : NULL); 3104 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3105 p2i(discovered_addr), discovered_addr ? p2i((oop)RawAccess<>::oop_load(discovered_addr)) : NULL); 3106 } 3107 #endif 3108 3109 template <class T> 3110 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) { 3111 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr_raw(obj); 3112 PSParallelCompact::adjust_pointer(referent_addr, cm); 3113 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr_raw(obj); 3114 PSParallelCompact::adjust_pointer(discovered_addr, cm); 3115 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3116 referent_addr, discovered_addr);) 3117 } 3118 3119 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3120 InstanceKlass::oop_pc_update_pointers(obj, cm); 3121 3122 if (UseCompressedOops) { 3123 oop_pc_update_pointers_specialized<narrowOop>(obj, cm); 3124 } else { 3125 oop_pc_update_pointers_specialized<oop>(obj, cm); 3126 } 3127 } 3128 3129 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3130 assert(obj->is_objArray(), "obj must be obj array"); 3131 PSParallelCompact::AdjustPointerClosure closure(cm); 3132 if (UseCompressedOops) { 3133 oop_oop_iterate_elements<narrowOop>(objArrayOop(obj), &closure); 3134 } else { 3135 oop_oop_iterate_elements<oop>(objArrayOop(obj), &closure); 3136 } 3137 } 3138 3139 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3140 assert(obj->is_typeArray(),"must be a type array"); 3141 } 3142 3143 ParMarkBitMapClosure::IterationStatus 3144 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3145 assert(destination() != NULL, "sanity"); 3146 assert(bitmap()->obj_size(addr) == words, "bad size"); 3147 3148 _source = addr; 3149 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3150 destination(), "wrong destination"); 3151 3152 if (words > words_remaining()) { 3153 return ParMarkBitMap::would_overflow; 3154 } 3155 3156 // The start_array must be updated even if the object is not moving. 3157 if (_start_array != NULL) { 3158 _start_array->allocate_block(destination()); 3159 } 3160 3161 if (destination() != source()) { 3162 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3163 Copy::aligned_conjoint_words(source(), destination(), words); 3164 } 3165 3166 oop moved_oop = (oop) destination(); 3167 compaction_manager()->update_contents(moved_oop); 3168 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3169 3170 update_state(words); 3171 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3172 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3173 } 3174 3175 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3176 ParCompactionManager* cm, 3177 PSParallelCompact::SpaceId space_id) : 3178 ParMarkBitMapClosure(mbm, cm), 3179 _space_id(space_id), 3180 _start_array(PSParallelCompact::start_array(space_id)) 3181 { 3182 } 3183 3184 // Updates the references in the object to their new values. 3185 ParMarkBitMapClosure::IterationStatus 3186 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3187 do_addr(addr); 3188 return ParMarkBitMap::incomplete; 3189 } 3190 3191 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3192 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3193 _start_array(PSParallelCompact::start_array(space_id)) 3194 { 3195 assert(space_id == PSParallelCompact::old_space_id, 3196 "cannot use FillClosure in the young gen"); 3197 } 3198 3199 ParMarkBitMapClosure::IterationStatus 3200 FillClosure::do_addr(HeapWord* addr, size_t size) { 3201 CollectedHeap::fill_with_objects(addr, size); 3202 HeapWord* const end = addr + size; 3203 do { 3204 _start_array->allocate_block(addr); 3205 addr += oop(addr)->size(); 3206 } while (addr < end); 3207 return ParMarkBitMap::incomplete; 3208 }