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