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