1 /* 2 * Copyright (c) 2005, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "aot/aotLoader.hpp" 27 #include "classfile/classLoaderDataGraph.hpp" 28 #include "classfile/javaClasses.inline.hpp" 29 #include "classfile/stringTable.hpp" 30 #include "classfile/symbolTable.hpp" 31 #include "classfile/systemDictionary.hpp" 32 #include "code/codeCache.hpp" 33 #include "gc/parallel/gcTaskManager.hpp" 34 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 35 #include "gc/parallel/parMarkBitMap.inline.hpp" 36 #include "gc/parallel/pcTasks.hpp" 37 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 38 #include "gc/parallel/psCompactionManager.inline.hpp" 39 #include "gc/parallel/psOldGen.hpp" 40 #include "gc/parallel/psParallelCompact.inline.hpp" 41 #include "gc/parallel/psPromotionManager.inline.hpp" 42 #include "gc/parallel/psScavenge.hpp" 43 #include "gc/parallel/psYoungGen.hpp" 44 #include "gc/shared/gcCause.hpp" 45 #include "gc/shared/gcHeapSummary.hpp" 46 #include "gc/shared/gcId.hpp" 47 #include "gc/shared/gcLocker.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/isGCActiveMark.hpp" 52 #include "gc/shared/referencePolicy.hpp" 53 #include "gc/shared/referenceProcessor.hpp" 54 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 55 #include "gc/shared/spaceDecorator.hpp" 56 #include "gc/shared/weakProcessor.hpp" 57 #include "logging/log.hpp" 58 #include "memory/iterator.inline.hpp" 59 #include "memory/resourceArea.hpp" 60 #include "oops/access.inline.hpp" 61 #include "oops/instanceClassLoaderKlass.inline.hpp" 62 #include "oops/instanceKlass.inline.hpp" 63 #include "oops/instanceMirrorKlass.inline.hpp" 64 #include "oops/methodData.hpp" 65 #include "oops/objArrayKlass.inline.hpp" 66 #include "oops/oop.inline.hpp" 67 #include "runtime/atomic.hpp" 68 #include "runtime/handles.inline.hpp" 69 #include "runtime/safepoint.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/management.hpp" 72 #include "services/memTracker.hpp" 73 #include "services/memoryService.hpp" 74 #include "utilities/align.hpp" 75 #include "utilities/debug.hpp" 76 #include "utilities/events.hpp" 77 #include "utilities/formatBuffer.hpp" 78 #include "utilities/macros.hpp" 79 #include "utilities/stack.inline.hpp" 80 81 #include <math.h> 82 83 // All sizes are in HeapWords. 84 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 85 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 86 const size_t ParallelCompactData::RegionSizeBytes = 87 RegionSize << LogHeapWordSize; 88 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 89 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 90 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 91 92 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 93 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 94 const size_t ParallelCompactData::BlockSizeBytes = 95 BlockSize << LogHeapWordSize; 96 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 97 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 98 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 99 100 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 101 const size_t ParallelCompactData::Log2BlocksPerRegion = 102 Log2RegionSize - Log2BlockSize; 103 104 const ParallelCompactData::RegionData::region_sz_t 105 ParallelCompactData::RegionData::dc_shift = 27; 106 107 const ParallelCompactData::RegionData::region_sz_t 108 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 109 110 const ParallelCompactData::RegionData::region_sz_t 111 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 112 113 const ParallelCompactData::RegionData::region_sz_t 114 ParallelCompactData::RegionData::los_mask = ~dc_mask; 115 116 const ParallelCompactData::RegionData::region_sz_t 117 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 118 119 const ParallelCompactData::RegionData::region_sz_t 120 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 121 122 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 123 124 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer; 125 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 126 127 double PSParallelCompact::_dwl_mean; 128 double PSParallelCompact::_dwl_std_dev; 129 double PSParallelCompact::_dwl_first_term; 130 double PSParallelCompact::_dwl_adjustment; 131 #ifdef ASSERT 132 bool PSParallelCompact::_dwl_initialized = false; 133 #endif // #ifdef ASSERT 134 135 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 136 HeapWord* destination) 137 { 138 assert(src_region_idx != 0, "invalid src_region_idx"); 139 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 140 assert(destination != NULL, "invalid destination argument"); 141 142 _src_region_idx = src_region_idx; 143 _partial_obj_size = partial_obj_size; 144 _destination = destination; 145 146 // These fields may not be updated below, so make sure they're clear. 147 assert(_dest_region_addr == NULL, "should have been cleared"); 148 assert(_first_src_addr == NULL, "should have been cleared"); 149 150 // Determine the number of destination regions for the partial object. 151 HeapWord* const last_word = destination + partial_obj_size - 1; 152 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 153 HeapWord* const beg_region_addr = sd.region_align_down(destination); 154 HeapWord* const end_region_addr = sd.region_align_down(last_word); 155 156 if (beg_region_addr == end_region_addr) { 157 // One destination region. 158 _destination_count = 1; 159 if (end_region_addr == destination) { 160 // The destination falls on a region boundary, thus the first word of the 161 // partial object will be the first word copied to the destination region. 162 _dest_region_addr = end_region_addr; 163 _first_src_addr = sd.region_to_addr(src_region_idx); 164 } 165 } else { 166 // Two destination regions. When copied, the partial object will cross a 167 // destination region boundary, so a word somewhere within the partial 168 // object will be the first word copied to the second destination region. 169 _destination_count = 2; 170 _dest_region_addr = end_region_addr; 171 const size_t ofs = pointer_delta(end_region_addr, destination); 172 assert(ofs < _partial_obj_size, "sanity"); 173 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 174 } 175 } 176 177 void SplitInfo::clear() 178 { 179 _src_region_idx = 0; 180 _partial_obj_size = 0; 181 _destination = NULL; 182 _destination_count = 0; 183 _dest_region_addr = NULL; 184 _first_src_addr = NULL; 185 assert(!is_valid(), "sanity"); 186 } 187 188 #ifdef ASSERT 189 void SplitInfo::verify_clear() 190 { 191 assert(_src_region_idx == 0, "not clear"); 192 assert(_partial_obj_size == 0, "not clear"); 193 assert(_destination == NULL, "not clear"); 194 assert(_destination_count == 0, "not clear"); 195 assert(_dest_region_addr == NULL, "not clear"); 196 assert(_first_src_addr == NULL, "not clear"); 197 } 198 #endif // #ifdef ASSERT 199 200 201 void PSParallelCompact::print_on_error(outputStream* st) { 202 _mark_bitmap.print_on_error(st); 203 } 204 205 #ifndef PRODUCT 206 const char* PSParallelCompact::space_names[] = { 207 "old ", "eden", "from", "to " 208 }; 209 210 void PSParallelCompact::print_region_ranges() { 211 if (!log_develop_is_enabled(Trace, gc, compaction)) { 212 return; 213 } 214 Log(gc, compaction) log; 215 ResourceMark rm; 216 LogStream ls(log.trace()); 217 Universe::print_on(&ls); 218 log.trace("space bottom top end new_top"); 219 log.trace("------ ---------- ---------- ---------- ----------"); 220 221 for (unsigned int id = 0; id < last_space_id; ++id) { 222 const MutableSpace* space = _space_info[id].space(); 223 log.trace("%u %s " 224 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 225 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 226 id, space_names[id], 227 summary_data().addr_to_region_idx(space->bottom()), 228 summary_data().addr_to_region_idx(space->top()), 229 summary_data().addr_to_region_idx(space->end()), 230 summary_data().addr_to_region_idx(_space_info[id].new_top())); 231 } 232 } 233 234 void 235 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 236 { 237 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 238 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 239 240 ParallelCompactData& sd = PSParallelCompact::summary_data(); 241 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 242 log_develop_trace(gc, compaction)( 243 REGION_IDX_FORMAT " " PTR_FORMAT " " 244 REGION_IDX_FORMAT " " PTR_FORMAT " " 245 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 246 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 247 i, p2i(c->data_location()), dci, p2i(c->destination()), 248 c->partial_obj_size(), c->live_obj_size(), 249 c->data_size(), c->source_region(), c->destination_count()); 250 251 #undef REGION_IDX_FORMAT 252 #undef REGION_DATA_FORMAT 253 } 254 255 void 256 print_generic_summary_data(ParallelCompactData& summary_data, 257 HeapWord* const beg_addr, 258 HeapWord* const end_addr) 259 { 260 size_t total_words = 0; 261 size_t i = summary_data.addr_to_region_idx(beg_addr); 262 const size_t last = summary_data.addr_to_region_idx(end_addr); 263 HeapWord* pdest = 0; 264 265 while (i < last) { 266 ParallelCompactData::RegionData* c = summary_data.region(i); 267 if (c->data_size() != 0 || c->destination() != pdest) { 268 print_generic_summary_region(i, c); 269 total_words += c->data_size(); 270 pdest = c->destination(); 271 } 272 ++i; 273 } 274 275 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 276 } 277 278 void 279 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 280 HeapWord* const beg_addr, 281 HeapWord* const end_addr) { 282 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 283 } 284 285 void 286 print_generic_summary_data(ParallelCompactData& summary_data, 287 SpaceInfo* space_info) 288 { 289 if (!log_develop_is_enabled(Trace, gc, compaction)) { 290 return; 291 } 292 293 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 294 const MutableSpace* space = space_info[id].space(); 295 print_generic_summary_data(summary_data, space->bottom(), 296 MAX2(space->top(), space_info[id].new_top())); 297 } 298 } 299 300 void 301 print_initial_summary_data(ParallelCompactData& summary_data, 302 const MutableSpace* space) { 303 if (space->top() == space->bottom()) { 304 return; 305 } 306 307 const size_t region_size = ParallelCompactData::RegionSize; 308 typedef ParallelCompactData::RegionData RegionData; 309 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 310 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 311 const RegionData* c = summary_data.region(end_region - 1); 312 HeapWord* end_addr = c->destination() + c->data_size(); 313 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 314 315 // Print (and count) the full regions at the beginning of the space. 316 size_t full_region_count = 0; 317 size_t i = summary_data.addr_to_region_idx(space->bottom()); 318 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 319 ParallelCompactData::RegionData* c = summary_data.region(i); 320 log_develop_trace(gc, compaction)( 321 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 322 i, p2i(c->destination()), 323 c->partial_obj_size(), c->live_obj_size(), 324 c->data_size(), c->source_region(), c->destination_count()); 325 ++full_region_count; 326 ++i; 327 } 328 329 size_t live_to_right = live_in_space - full_region_count * region_size; 330 331 double max_reclaimed_ratio = 0.0; 332 size_t max_reclaimed_ratio_region = 0; 333 size_t max_dead_to_right = 0; 334 size_t max_live_to_right = 0; 335 336 // Print the 'reclaimed ratio' for regions while there is something live in 337 // the region or to the right of it. The remaining regions are empty (and 338 // uninteresting), and computing the ratio will result in division by 0. 339 while (i < end_region && live_to_right > 0) { 340 c = summary_data.region(i); 341 HeapWord* const region_addr = summary_data.region_to_addr(i); 342 const size_t used_to_right = pointer_delta(space->top(), region_addr); 343 const size_t dead_to_right = used_to_right - live_to_right; 344 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 345 346 if (reclaimed_ratio > max_reclaimed_ratio) { 347 max_reclaimed_ratio = reclaimed_ratio; 348 max_reclaimed_ratio_region = i; 349 max_dead_to_right = dead_to_right; 350 max_live_to_right = live_to_right; 351 } 352 353 ParallelCompactData::RegionData* c = summary_data.region(i); 354 log_develop_trace(gc, compaction)( 355 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 356 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 357 i, p2i(c->destination()), 358 c->partial_obj_size(), c->live_obj_size(), 359 c->data_size(), c->source_region(), c->destination_count(), 360 reclaimed_ratio, dead_to_right, live_to_right); 361 362 363 live_to_right -= c->data_size(); 364 ++i; 365 } 366 367 // Any remaining regions are empty. Print one more if there is one. 368 if (i < end_region) { 369 ParallelCompactData::RegionData* c = summary_data.region(i); 370 log_develop_trace(gc, compaction)( 371 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 372 i, p2i(c->destination()), 373 c->partial_obj_size(), c->live_obj_size(), 374 c->data_size(), c->source_region(), c->destination_count()); 375 } 376 377 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 378 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 379 } 380 381 void 382 print_initial_summary_data(ParallelCompactData& summary_data, 383 SpaceInfo* space_info) { 384 if (!log_develop_is_enabled(Trace, gc, compaction)) { 385 return; 386 } 387 388 unsigned int id = PSParallelCompact::old_space_id; 389 const MutableSpace* space; 390 do { 391 space = space_info[id].space(); 392 print_initial_summary_data(summary_data, space); 393 } while (++id < PSParallelCompact::eden_space_id); 394 395 do { 396 space = space_info[id].space(); 397 print_generic_summary_data(summary_data, space->bottom(), space->top()); 398 } while (++id < PSParallelCompact::last_space_id); 399 } 400 #endif // #ifndef PRODUCT 401 402 #ifdef ASSERT 403 size_t add_obj_count; 404 size_t add_obj_size; 405 size_t mark_bitmap_count; 406 size_t mark_bitmap_size; 407 #endif // #ifdef ASSERT 408 409 ParallelCompactData::ParallelCompactData() : 410 _region_start(NULL), 411 DEBUG_ONLY(_region_end(NULL) COMMA) 412 _region_vspace(NULL), 413 _reserved_byte_size(0), 414 _region_data(NULL), 415 _region_count(0), 416 _block_vspace(NULL), 417 _block_data(NULL), 418 _block_count(0) {} 419 420 bool ParallelCompactData::initialize(MemRegion covered_region) 421 { 422 _region_start = covered_region.start(); 423 const size_t region_size = covered_region.word_size(); 424 DEBUG_ONLY(_region_end = _region_start + region_size;) 425 426 assert(region_align_down(_region_start) == _region_start, 427 "region start not aligned"); 428 assert((region_size & RegionSizeOffsetMask) == 0, 429 "region size not a multiple of RegionSize"); 430 431 bool result = initialize_region_data(region_size) && initialize_block_data(); 432 return result; 433 } 434 435 PSVirtualSpace* 436 ParallelCompactData::create_vspace(size_t count, size_t element_size) 437 { 438 const size_t raw_bytes = count * element_size; 439 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 440 const size_t granularity = os::vm_allocation_granularity(); 441 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 442 443 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 444 MAX2(page_sz, granularity); 445 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 446 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 447 rs.size()); 448 449 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 450 451 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 452 if (vspace != 0) { 453 if (vspace->expand_by(_reserved_byte_size)) { 454 return vspace; 455 } 456 delete vspace; 457 // Release memory reserved in the space. 458 rs.release(); 459 } 460 461 return 0; 462 } 463 464 bool ParallelCompactData::initialize_region_data(size_t region_size) 465 { 466 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 467 _region_vspace = create_vspace(count, sizeof(RegionData)); 468 if (_region_vspace != 0) { 469 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 470 _region_count = count; 471 return true; 472 } 473 return false; 474 } 475 476 bool ParallelCompactData::initialize_block_data() 477 { 478 assert(_region_count != 0, "region data must be initialized first"); 479 const size_t count = _region_count << Log2BlocksPerRegion; 480 _block_vspace = create_vspace(count, sizeof(BlockData)); 481 if (_block_vspace != 0) { 482 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 483 _block_count = count; 484 return true; 485 } 486 return false; 487 } 488 489 void ParallelCompactData::clear() 490 { 491 memset(_region_data, 0, _region_vspace->committed_size()); 492 memset(_block_data, 0, _block_vspace->committed_size()); 493 } 494 495 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 496 assert(beg_region <= _region_count, "beg_region out of range"); 497 assert(end_region <= _region_count, "end_region out of range"); 498 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 499 500 const size_t region_cnt = end_region - beg_region; 501 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 502 503 const size_t beg_block = beg_region * BlocksPerRegion; 504 const size_t block_cnt = region_cnt * BlocksPerRegion; 505 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 506 } 507 508 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 509 { 510 const RegionData* cur_cp = region(region_idx); 511 const RegionData* const end_cp = region(region_count() - 1); 512 513 HeapWord* result = region_to_addr(region_idx); 514 if (cur_cp < end_cp) { 515 do { 516 result += cur_cp->partial_obj_size(); 517 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 518 } 519 return result; 520 } 521 522 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 523 { 524 const size_t obj_ofs = pointer_delta(addr, _region_start); 525 const size_t beg_region = obj_ofs >> Log2RegionSize; 526 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 527 528 DEBUG_ONLY(Atomic::inc(&add_obj_count);) 529 DEBUG_ONLY(Atomic::add(len, &add_obj_size);) 530 531 if (beg_region == end_region) { 532 // All in one region. 533 _region_data[beg_region].add_live_obj(len); 534 return; 535 } 536 537 // First region. 538 const size_t beg_ofs = region_offset(addr); 539 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 540 541 Klass* klass = ((oop)addr)->klass(); 542 // Middle regions--completely spanned by this object. 543 for (size_t region = beg_region + 1; region < end_region; ++region) { 544 _region_data[region].set_partial_obj_size(RegionSize); 545 _region_data[region].set_partial_obj_addr(addr); 546 } 547 548 // Last region. 549 const size_t end_ofs = region_offset(addr + len - 1); 550 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 551 _region_data[end_region].set_partial_obj_addr(addr); 552 } 553 554 void 555 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 556 { 557 assert(region_offset(beg) == 0, "not RegionSize aligned"); 558 assert(region_offset(end) == 0, "not RegionSize aligned"); 559 560 size_t cur_region = addr_to_region_idx(beg); 561 const size_t end_region = addr_to_region_idx(end); 562 HeapWord* addr = beg; 563 while (cur_region < end_region) { 564 _region_data[cur_region].set_destination(addr); 565 _region_data[cur_region].set_destination_count(0); 566 _region_data[cur_region].set_source_region(cur_region); 567 _region_data[cur_region].set_data_location(addr); 568 569 // Update live_obj_size so the region appears completely full. 570 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 571 _region_data[cur_region].set_live_obj_size(live_size); 572 573 ++cur_region; 574 addr += RegionSize; 575 } 576 } 577 578 // Find the point at which a space can be split and, if necessary, record the 579 // split point. 580 // 581 // If the current src region (which overflowed the destination space) doesn't 582 // have a partial object, the split point is at the beginning of the current src 583 // region (an "easy" split, no extra bookkeeping required). 584 // 585 // If the current src region has a partial object, the split point is in the 586 // region where that partial object starts (call it the split_region). If 587 // split_region has a partial object, then the split point is just after that 588 // partial object (a "hard" split where we have to record the split data and 589 // zero the partial_obj_size field). With a "hard" split, we know that the 590 // partial_obj ends within split_region because the partial object that caused 591 // the overflow starts in split_region. If split_region doesn't have a partial 592 // obj, then the split is at the beginning of split_region (another "easy" 593 // split). 594 HeapWord* 595 ParallelCompactData::summarize_split_space(size_t src_region, 596 SplitInfo& split_info, 597 HeapWord* destination, 598 HeapWord* target_end, 599 HeapWord** target_next) 600 { 601 assert(destination <= target_end, "sanity"); 602 assert(destination + _region_data[src_region].data_size() > target_end, 603 "region should not fit into target space"); 604 assert(is_region_aligned(target_end), "sanity"); 605 606 size_t split_region = src_region; 607 HeapWord* split_destination = destination; 608 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 609 610 if (destination + partial_obj_size > target_end) { 611 // The split point is just after the partial object (if any) in the 612 // src_region that contains the start of the object that overflowed the 613 // destination space. 614 // 615 // Find the start of the "overflow" object and set split_region to the 616 // region containing it. 617 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 618 split_region = addr_to_region_idx(overflow_obj); 619 620 // Clear the source_region field of all destination regions whose first word 621 // came from data after the split point (a non-null source_region field 622 // implies a region must be filled). 623 // 624 // An alternative to the simple loop below: clear during post_compact(), 625 // which uses memcpy instead of individual stores, and is easy to 626 // parallelize. (The downside is that it clears the entire RegionData 627 // object as opposed to just one field.) 628 // 629 // post_compact() would have to clear the summary data up to the highest 630 // address that was written during the summary phase, which would be 631 // 632 // max(top, max(new_top, clear_top)) 633 // 634 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 635 // to target_end. 636 const RegionData* const sr = region(split_region); 637 const size_t beg_idx = 638 addr_to_region_idx(region_align_up(sr->destination() + 639 sr->partial_obj_size())); 640 const size_t end_idx = addr_to_region_idx(target_end); 641 642 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 643 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 644 _region_data[idx].set_source_region(0); 645 } 646 647 // Set split_destination and partial_obj_size to reflect the split region. 648 split_destination = sr->destination(); 649 partial_obj_size = sr->partial_obj_size(); 650 } 651 652 // The split is recorded only if a partial object extends onto the region. 653 if (partial_obj_size != 0) { 654 _region_data[split_region].set_partial_obj_size(0); 655 split_info.record(split_region, partial_obj_size, split_destination); 656 } 657 658 // Setup the continuation addresses. 659 *target_next = split_destination + partial_obj_size; 660 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 661 662 if (log_develop_is_enabled(Trace, gc, compaction)) { 663 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 664 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 665 split_type, p2i(source_next), split_region, partial_obj_size); 666 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 667 split_type, p2i(split_destination), 668 addr_to_region_idx(split_destination), 669 p2i(*target_next)); 670 671 if (partial_obj_size != 0) { 672 HeapWord* const po_beg = split_info.destination(); 673 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 674 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 675 split_type, 676 p2i(po_beg), addr_to_region_idx(po_beg), 677 p2i(po_end), addr_to_region_idx(po_end)); 678 } 679 } 680 681 return source_next; 682 } 683 684 bool ParallelCompactData::summarize(SplitInfo& split_info, 685 HeapWord* source_beg, HeapWord* source_end, 686 HeapWord** source_next, 687 HeapWord* target_beg, HeapWord* target_end, 688 HeapWord** target_next) 689 { 690 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 691 log_develop_trace(gc, compaction)( 692 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 693 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 694 p2i(source_beg), p2i(source_end), p2i(source_next_val), 695 p2i(target_beg), p2i(target_end), p2i(*target_next)); 696 697 size_t cur_region = addr_to_region_idx(source_beg); 698 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 699 700 HeapWord *dest_addr = target_beg; 701 while (cur_region < end_region) { 702 // The destination must be set even if the region has no data. 703 _region_data[cur_region].set_destination(dest_addr); 704 705 size_t words = _region_data[cur_region].data_size(); 706 if (words > 0) { 707 // If cur_region does not fit entirely into the target space, find a point 708 // at which the source space can be 'split' so that part is copied to the 709 // target space and the rest is copied elsewhere. 710 if (dest_addr + words > target_end) { 711 assert(source_next != NULL, "source_next is NULL when splitting"); 712 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 713 target_end, target_next); 714 return false; 715 } 716 717 // Compute the destination_count for cur_region, and if necessary, update 718 // source_region for a destination region. The source_region field is 719 // updated if cur_region is the first (left-most) region to be copied to a 720 // destination region. 721 // 722 // The destination_count calculation is a bit subtle. A region that has 723 // data that compacts into itself does not count itself as a destination. 724 // This maintains the invariant that a zero count means the region is 725 // available and can be claimed and then filled. 726 uint destination_count = 0; 727 if (split_info.is_split(cur_region)) { 728 // The current region has been split: the partial object will be copied 729 // to one destination space and the remaining data will be copied to 730 // another destination space. Adjust the initial destination_count and, 731 // if necessary, set the source_region field if the partial object will 732 // cross a destination region boundary. 733 destination_count = split_info.destination_count(); 734 if (destination_count == 2) { 735 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 736 _region_data[dest_idx].set_source_region(cur_region); 737 } 738 } 739 740 HeapWord* const last_addr = dest_addr + words - 1; 741 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 742 const size_t dest_region_2 = addr_to_region_idx(last_addr); 743 744 // Initially assume that the destination regions will be the same and 745 // adjust the value below if necessary. Under this assumption, if 746 // cur_region == dest_region_2, then cur_region will be compacted 747 // completely into itself. 748 destination_count += cur_region == dest_region_2 ? 0 : 1; 749 if (dest_region_1 != dest_region_2) { 750 // Destination regions differ; adjust destination_count. 751 destination_count += 1; 752 // Data from cur_region will be copied to the start of dest_region_2. 753 _region_data[dest_region_2].set_source_region(cur_region); 754 } else if (region_offset(dest_addr) == 0) { 755 // Data from cur_region will be copied to the start of the destination 756 // region. 757 _region_data[dest_region_1].set_source_region(cur_region); 758 } 759 760 _region_data[cur_region].set_destination_count(destination_count); 761 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 762 dest_addr += words; 763 } 764 765 ++cur_region; 766 } 767 768 *target_next = dest_addr; 769 return true; 770 } 771 772 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 773 assert(addr != NULL, "Should detect NULL oop earlier"); 774 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 775 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 776 777 // Region covering the object. 778 RegionData* const region_ptr = addr_to_region_ptr(addr); 779 HeapWord* result = region_ptr->destination(); 780 781 // If the entire Region is live, the new location is region->destination + the 782 // offset of the object within in the Region. 783 784 // Run some performance tests to determine if this special case pays off. It 785 // is worth it for pointers into the dense prefix. If the optimization to 786 // avoid pointer updates in regions that only point to the dense prefix is 787 // ever implemented, this should be revisited. 788 if (region_ptr->data_size() == RegionSize) { 789 result += region_offset(addr); 790 return result; 791 } 792 793 // Otherwise, the new location is region->destination + block offset + the 794 // number of live words in the Block that are (a) to the left of addr and (b) 795 // due to objects that start in the Block. 796 797 // Fill in the block table if necessary. This is unsynchronized, so multiple 798 // threads may fill the block table for a region (harmless, since it is 799 // idempotent). 800 if (!region_ptr->blocks_filled()) { 801 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 802 region_ptr->set_blocks_filled(); 803 } 804 805 HeapWord* const search_start = block_align_down(addr); 806 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 807 808 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 809 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 810 result += block_offset + live; 811 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 812 return result; 813 } 814 815 #ifdef ASSERT 816 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 817 { 818 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 819 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 820 for (const size_t* p = beg; p < end; ++p) { 821 assert(*p == 0, "not zero"); 822 } 823 } 824 825 void ParallelCompactData::verify_clear() 826 { 827 verify_clear(_region_vspace); 828 verify_clear(_block_vspace); 829 } 830 #endif // #ifdef ASSERT 831 832 STWGCTimer PSParallelCompact::_gc_timer; 833 ParallelOldTracer PSParallelCompact::_gc_tracer; 834 elapsedTimer PSParallelCompact::_accumulated_time; 835 unsigned int PSParallelCompact::_total_invocations = 0; 836 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 837 jlong PSParallelCompact::_time_of_last_gc = 0; 838 CollectorCounters* PSParallelCompact::_counters = NULL; 839 ParMarkBitMap PSParallelCompact::_mark_bitmap; 840 ParallelCompactData PSParallelCompact::_summary_data; 841 842 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 843 844 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 845 846 class PCReferenceProcessor: public ReferenceProcessor { 847 public: 848 PCReferenceProcessor( 849 BoolObjectClosure* is_subject_to_discovery, 850 BoolObjectClosure* is_alive_non_header) : 851 ReferenceProcessor(is_subject_to_discovery, 852 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 853 ParallelGCThreads, // mt processing degree 854 true, // mt discovery 855 ParallelGCThreads, // mt discovery degree 856 true, // atomic_discovery 857 is_alive_non_header) { 858 } 859 860 template<typename T> bool discover(oop obj, ReferenceType type) { 861 T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj); 862 T heap_oop = RawAccess<>::oop_load(referent_addr); 863 oop referent = CompressedOops::decode_not_null(heap_oop); 864 return PSParallelCompact::mark_bitmap()->is_unmarked(referent) 865 && ReferenceProcessor::discover_reference(obj, type); 866 } 867 virtual bool discover_reference(oop obj, ReferenceType type) { 868 if (UseCompressedOops) { 869 return discover<narrowOop>(obj, type); 870 } else { 871 return discover<oop>(obj, type); 872 } 873 } 874 }; 875 876 void PSParallelCompact::post_initialize() { 877 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 878 _span_based_discoverer.set_span(heap->reserved_region()); 879 _ref_processor = 880 new PCReferenceProcessor(&_span_based_discoverer, 881 &_is_alive_closure); // non-header is alive closure 882 883 _counters = new CollectorCounters("Parallel full collection pauses", 1); 884 885 // Initialize static fields in ParCompactionManager. 886 ParCompactionManager::initialize(mark_bitmap()); 887 } 888 889 bool PSParallelCompact::initialize() { 890 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 891 MemRegion mr = heap->reserved_region(); 892 893 // Was the old gen get allocated successfully? 894 if (!heap->old_gen()->is_allocated()) { 895 return false; 896 } 897 898 initialize_space_info(); 899 initialize_dead_wood_limiter(); 900 901 if (!_mark_bitmap.initialize(mr)) { 902 vm_shutdown_during_initialization( 903 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 904 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 905 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 906 return false; 907 } 908 909 if (!_summary_data.initialize(mr)) { 910 vm_shutdown_during_initialization( 911 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 912 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 913 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 914 return false; 915 } 916 917 return true; 918 } 919 920 void PSParallelCompact::initialize_space_info() 921 { 922 memset(&_space_info, 0, sizeof(_space_info)); 923 924 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 925 PSYoungGen* young_gen = heap->young_gen(); 926 927 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 928 _space_info[eden_space_id].set_space(young_gen->eden_space()); 929 _space_info[from_space_id].set_space(young_gen->from_space()); 930 _space_info[to_space_id].set_space(young_gen->to_space()); 931 932 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 933 } 934 935 void PSParallelCompact::initialize_dead_wood_limiter() 936 { 937 const size_t max = 100; 938 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 939 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 940 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 941 DEBUG_ONLY(_dwl_initialized = true;) 942 _dwl_adjustment = normal_distribution(1.0); 943 } 944 945 void 946 PSParallelCompact::clear_data_covering_space(SpaceId id) 947 { 948 // At this point, top is the value before GC, new_top() is the value that will 949 // be set at the end of GC. The marking bitmap is cleared to top; nothing 950 // should be marked above top. The summary data is cleared to the larger of 951 // top & new_top. 952 MutableSpace* const space = _space_info[id].space(); 953 HeapWord* const bot = space->bottom(); 954 HeapWord* const top = space->top(); 955 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 956 957 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 958 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 959 _mark_bitmap.clear_range(beg_bit, end_bit); 960 961 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 962 const size_t end_region = 963 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 964 _summary_data.clear_range(beg_region, end_region); 965 966 // Clear the data used to 'split' regions. 967 SplitInfo& split_info = _space_info[id].split_info(); 968 if (split_info.is_valid()) { 969 split_info.clear(); 970 } 971 DEBUG_ONLY(split_info.verify_clear();) 972 } 973 974 void PSParallelCompact::pre_compact() 975 { 976 // Update the from & to space pointers in space_info, since they are swapped 977 // at each young gen gc. Do the update unconditionally (even though a 978 // promotion failure does not swap spaces) because an unknown number of young 979 // collections will have swapped the spaces an unknown number of times. 980 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 981 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 982 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 983 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 984 985 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 986 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 987 988 // Increment the invocation count 989 heap->increment_total_collections(true); 990 991 // We need to track unique mark sweep invocations as well. 992 _total_invocations++; 993 994 heap->print_heap_before_gc(); 995 heap->trace_heap_before_gc(&_gc_tracer); 996 997 // Fill in TLABs 998 heap->ensure_parsability(true); // retire TLABs 999 1000 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1001 HandleMark hm; // Discard invalid handles created during verification 1002 Universe::verify("Before GC"); 1003 } 1004 1005 // Verify object start arrays 1006 if (VerifyObjectStartArray && 1007 VerifyBeforeGC) { 1008 heap->old_gen()->verify_object_start_array(); 1009 } 1010 1011 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1012 DEBUG_ONLY(summary_data().verify_clear();) 1013 1014 // Have worker threads release resources the next time they run a task. 1015 gc_task_manager()->release_all_resources(); 1016 1017 ParCompactionManager::reset_all_bitmap_query_caches(); 1018 } 1019 1020 void PSParallelCompact::post_compact() 1021 { 1022 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1023 1024 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1025 // Clear the marking bitmap, summary data and split info. 1026 clear_data_covering_space(SpaceId(id)); 1027 // Update top(). Must be done after clearing the bitmap and summary data. 1028 _space_info[id].publish_new_top(); 1029 } 1030 1031 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1032 MutableSpace* const from_space = _space_info[from_space_id].space(); 1033 MutableSpace* const to_space = _space_info[to_space_id].space(); 1034 1035 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1036 bool eden_empty = eden_space->is_empty(); 1037 if (!eden_empty) { 1038 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1039 heap->young_gen(), heap->old_gen()); 1040 } 1041 1042 // Update heap occupancy information which is used as input to the soft ref 1043 // clearing policy at the next gc. 1044 Universe::update_heap_info_at_gc(); 1045 1046 bool young_gen_empty = eden_empty && from_space->is_empty() && 1047 to_space->is_empty(); 1048 1049 PSCardTable* ct = heap->card_table(); 1050 MemRegion old_mr = heap->old_gen()->reserved(); 1051 if (young_gen_empty) { 1052 ct->clear(MemRegion(old_mr.start(), old_mr.end())); 1053 } else { 1054 ct->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1055 } 1056 1057 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1058 ClassLoaderDataGraph::purge(); 1059 MetaspaceUtils::verify_metrics(); 1060 1061 heap->prune_scavengable_nmethods(); 1062 JvmtiExport::gc_epilogue(); 1063 1064 #if COMPILER2_OR_JVMCI 1065 DerivedPointerTable::update_pointers(); 1066 #endif 1067 1068 if (ZapUnusedHeapArea) { 1069 heap->gen_mangle_unused_area(); 1070 } 1071 1072 // Update time of last GC 1073 reset_millis_since_last_gc(); 1074 } 1075 1076 HeapWord* 1077 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1078 bool maximum_compaction) 1079 { 1080 const size_t region_size = ParallelCompactData::RegionSize; 1081 const ParallelCompactData& sd = summary_data(); 1082 1083 const MutableSpace* const space = _space_info[id].space(); 1084 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1085 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1086 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1087 1088 // Skip full regions at the beginning of the space--they are necessarily part 1089 // of the dense prefix. 1090 size_t full_count = 0; 1091 const RegionData* cp; 1092 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1093 ++full_count; 1094 } 1095 1096 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1097 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1098 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1099 if (maximum_compaction || cp == end_cp || interval_ended) { 1100 _maximum_compaction_gc_num = total_invocations(); 1101 return sd.region_to_addr(cp); 1102 } 1103 1104 HeapWord* const new_top = _space_info[id].new_top(); 1105 const size_t space_live = pointer_delta(new_top, space->bottom()); 1106 const size_t space_used = space->used_in_words(); 1107 const size_t space_capacity = space->capacity_in_words(); 1108 1109 const double cur_density = double(space_live) / space_capacity; 1110 const double deadwood_density = 1111 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1112 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1113 1114 if (TraceParallelOldGCDensePrefix) { 1115 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1116 cur_density, deadwood_density, deadwood_goal); 1117 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1118 "space_cap=" SIZE_FORMAT, 1119 space_live, space_used, 1120 space_capacity); 1121 } 1122 1123 // XXX - Use binary search? 1124 HeapWord* dense_prefix = sd.region_to_addr(cp); 1125 const RegionData* full_cp = cp; 1126 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1127 while (cp < end_cp) { 1128 HeapWord* region_destination = cp->destination(); 1129 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1130 if (TraceParallelOldGCDensePrefix && Verbose) { 1131 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1132 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1133 sd.region(cp), p2i(region_destination), 1134 p2i(dense_prefix), cur_deadwood); 1135 } 1136 1137 if (cur_deadwood >= deadwood_goal) { 1138 // Found the region that has the correct amount of deadwood to the left. 1139 // This typically occurs after crossing a fairly sparse set of regions, so 1140 // iterate backwards over those sparse regions, looking for the region 1141 // that has the lowest density of live objects 'to the right.' 1142 size_t space_to_left = sd.region(cp) * region_size; 1143 size_t live_to_left = space_to_left - cur_deadwood; 1144 size_t space_to_right = space_capacity - space_to_left; 1145 size_t live_to_right = space_live - live_to_left; 1146 double density_to_right = double(live_to_right) / space_to_right; 1147 while (cp > full_cp) { 1148 --cp; 1149 const size_t prev_region_live_to_right = live_to_right - 1150 cp->data_size(); 1151 const size_t prev_region_space_to_right = space_to_right + region_size; 1152 double prev_region_density_to_right = 1153 double(prev_region_live_to_right) / prev_region_space_to_right; 1154 if (density_to_right <= prev_region_density_to_right) { 1155 return dense_prefix; 1156 } 1157 if (TraceParallelOldGCDensePrefix && Verbose) { 1158 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1159 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1160 prev_region_density_to_right); 1161 } 1162 dense_prefix -= region_size; 1163 live_to_right = prev_region_live_to_right; 1164 space_to_right = prev_region_space_to_right; 1165 density_to_right = prev_region_density_to_right; 1166 } 1167 return dense_prefix; 1168 } 1169 1170 dense_prefix += region_size; 1171 ++cp; 1172 } 1173 1174 return dense_prefix; 1175 } 1176 1177 #ifndef PRODUCT 1178 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1179 const SpaceId id, 1180 const bool maximum_compaction, 1181 HeapWord* const addr) 1182 { 1183 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1184 RegionData* const cp = summary_data().region(region_idx); 1185 const MutableSpace* const space = _space_info[id].space(); 1186 HeapWord* const new_top = _space_info[id].new_top(); 1187 1188 const size_t space_live = pointer_delta(new_top, space->bottom()); 1189 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1190 const size_t space_cap = space->capacity_in_words(); 1191 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1192 const size_t live_to_right = new_top - cp->destination(); 1193 const size_t dead_to_right = space->top() - addr - live_to_right; 1194 1195 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1196 "spl=" SIZE_FORMAT " " 1197 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1198 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1199 " ratio=%10.8f", 1200 algorithm, p2i(addr), region_idx, 1201 space_live, 1202 dead_to_left, dead_to_left_pct, 1203 dead_to_right, live_to_right, 1204 double(dead_to_right) / live_to_right); 1205 } 1206 #endif // #ifndef PRODUCT 1207 1208 // Return a fraction indicating how much of the generation can be treated as 1209 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1210 // based on the density of live objects in the generation to determine a limit, 1211 // which is then adjusted so the return value is min_percent when the density is 1212 // 1. 1213 // 1214 // The following table shows some return values for a different values of the 1215 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1216 // min_percent is 1. 1217 // 1218 // fraction allowed as dead wood 1219 // ----------------------------------------------------------------- 1220 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1221 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1222 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1223 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1224 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1225 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1226 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1227 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1228 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1229 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1230 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1231 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1232 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1233 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1234 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1235 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1236 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1237 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1238 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1239 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1240 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1241 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1242 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1243 1244 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1245 { 1246 assert(_dwl_initialized, "uninitialized"); 1247 1248 // The raw limit is the value of the normal distribution at x = density. 1249 const double raw_limit = normal_distribution(density); 1250 1251 // Adjust the raw limit so it becomes the minimum when the density is 1. 1252 // 1253 // First subtract the adjustment value (which is simply the precomputed value 1254 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1255 // Then add the minimum value, so the minimum is returned when the density is 1256 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1257 const double min = double(min_percent) / 100.0; 1258 const double limit = raw_limit - _dwl_adjustment + min; 1259 return MAX2(limit, 0.0); 1260 } 1261 1262 ParallelCompactData::RegionData* 1263 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1264 const RegionData* end) 1265 { 1266 const size_t region_size = ParallelCompactData::RegionSize; 1267 ParallelCompactData& sd = summary_data(); 1268 size_t left = sd.region(beg); 1269 size_t right = end > beg ? sd.region(end) - 1 : left; 1270 1271 // Binary search. 1272 while (left < right) { 1273 // Equivalent to (left + right) / 2, but does not overflow. 1274 const size_t middle = left + (right - left) / 2; 1275 RegionData* const middle_ptr = sd.region(middle); 1276 HeapWord* const dest = middle_ptr->destination(); 1277 HeapWord* const addr = sd.region_to_addr(middle); 1278 assert(dest != NULL, "sanity"); 1279 assert(dest <= addr, "must move left"); 1280 1281 if (middle > left && dest < addr) { 1282 right = middle - 1; 1283 } else if (middle < right && middle_ptr->data_size() == region_size) { 1284 left = middle + 1; 1285 } else { 1286 return middle_ptr; 1287 } 1288 } 1289 return sd.region(left); 1290 } 1291 1292 ParallelCompactData::RegionData* 1293 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1294 const RegionData* end, 1295 size_t dead_words) 1296 { 1297 ParallelCompactData& sd = summary_data(); 1298 size_t left = sd.region(beg); 1299 size_t right = end > beg ? sd.region(end) - 1 : left; 1300 1301 // Binary search. 1302 while (left < right) { 1303 // Equivalent to (left + right) / 2, but does not overflow. 1304 const size_t middle = left + (right - left) / 2; 1305 RegionData* const middle_ptr = sd.region(middle); 1306 HeapWord* const dest = middle_ptr->destination(); 1307 HeapWord* const addr = sd.region_to_addr(middle); 1308 assert(dest != NULL, "sanity"); 1309 assert(dest <= addr, "must move left"); 1310 1311 const size_t dead_to_left = pointer_delta(addr, dest); 1312 if (middle > left && dead_to_left > dead_words) { 1313 right = middle - 1; 1314 } else if (middle < right && dead_to_left < dead_words) { 1315 left = middle + 1; 1316 } else { 1317 return middle_ptr; 1318 } 1319 } 1320 return sd.region(left); 1321 } 1322 1323 // The result is valid during the summary phase, after the initial summarization 1324 // of each space into itself, and before final summarization. 1325 inline double 1326 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1327 HeapWord* const bottom, 1328 HeapWord* const top, 1329 HeapWord* const new_top) 1330 { 1331 ParallelCompactData& sd = summary_data(); 1332 1333 assert(cp != NULL, "sanity"); 1334 assert(bottom != NULL, "sanity"); 1335 assert(top != NULL, "sanity"); 1336 assert(new_top != NULL, "sanity"); 1337 assert(top >= new_top, "summary data problem?"); 1338 assert(new_top > bottom, "space is empty; should not be here"); 1339 assert(new_top >= cp->destination(), "sanity"); 1340 assert(top >= sd.region_to_addr(cp), "sanity"); 1341 1342 HeapWord* const destination = cp->destination(); 1343 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1344 const size_t compacted_region_live = pointer_delta(new_top, destination); 1345 const size_t compacted_region_used = pointer_delta(top, 1346 sd.region_to_addr(cp)); 1347 const size_t reclaimable = compacted_region_used - compacted_region_live; 1348 1349 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1350 return double(reclaimable) / divisor; 1351 } 1352 1353 // Return the address of the end of the dense prefix, a.k.a. the start of the 1354 // compacted region. The address is always on a region boundary. 1355 // 1356 // Completely full regions at the left are skipped, since no compaction can 1357 // occur in those regions. Then the maximum amount of dead wood to allow is 1358 // computed, based on the density (amount live / capacity) of the generation; 1359 // the region with approximately that amount of dead space to the left is 1360 // identified as the limit region. Regions between the last completely full 1361 // region and the limit region are scanned and the one that has the best 1362 // (maximum) reclaimed_ratio() is selected. 1363 HeapWord* 1364 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1365 bool maximum_compaction) 1366 { 1367 const size_t region_size = ParallelCompactData::RegionSize; 1368 const ParallelCompactData& sd = summary_data(); 1369 1370 const MutableSpace* const space = _space_info[id].space(); 1371 HeapWord* const top = space->top(); 1372 HeapWord* const top_aligned_up = sd.region_align_up(top); 1373 HeapWord* const new_top = _space_info[id].new_top(); 1374 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1375 HeapWord* const bottom = space->bottom(); 1376 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1377 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1378 const RegionData* const new_top_cp = 1379 sd.addr_to_region_ptr(new_top_aligned_up); 1380 1381 // Skip full regions at the beginning of the space--they are necessarily part 1382 // of the dense prefix. 1383 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1384 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1385 space->is_empty(), "no dead space allowed to the left"); 1386 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1387 "region must have dead space"); 1388 1389 // The gc number is saved whenever a maximum compaction is done, and used to 1390 // determine when the maximum compaction interval has expired. This avoids 1391 // successive max compactions for different reasons. 1392 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1393 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1394 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1395 total_invocations() == HeapFirstMaximumCompactionCount; 1396 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1397 _maximum_compaction_gc_num = total_invocations(); 1398 return sd.region_to_addr(full_cp); 1399 } 1400 1401 const size_t space_live = pointer_delta(new_top, bottom); 1402 const size_t space_used = space->used_in_words(); 1403 const size_t space_capacity = space->capacity_in_words(); 1404 1405 const double density = double(space_live) / double(space_capacity); 1406 const size_t min_percent_free = MarkSweepDeadRatio; 1407 const double limiter = dead_wood_limiter(density, min_percent_free); 1408 const size_t dead_wood_max = space_used - space_live; 1409 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1410 dead_wood_max); 1411 1412 if (TraceParallelOldGCDensePrefix) { 1413 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1414 "space_cap=" SIZE_FORMAT, 1415 space_live, space_used, 1416 space_capacity); 1417 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1418 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1419 density, min_percent_free, limiter, 1420 dead_wood_max, dead_wood_limit); 1421 } 1422 1423 // Locate the region with the desired amount of dead space to the left. 1424 const RegionData* const limit_cp = 1425 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1426 1427 // Scan from the first region with dead space to the limit region and find the 1428 // one with the best (largest) reclaimed ratio. 1429 double best_ratio = 0.0; 1430 const RegionData* best_cp = full_cp; 1431 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1432 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1433 if (tmp_ratio > best_ratio) { 1434 best_cp = cp; 1435 best_ratio = tmp_ratio; 1436 } 1437 } 1438 1439 return sd.region_to_addr(best_cp); 1440 } 1441 1442 void PSParallelCompact::summarize_spaces_quick() 1443 { 1444 for (unsigned int i = 0; i < last_space_id; ++i) { 1445 const MutableSpace* space = _space_info[i].space(); 1446 HeapWord** nta = _space_info[i].new_top_addr(); 1447 bool result = _summary_data.summarize(_space_info[i].split_info(), 1448 space->bottom(), space->top(), NULL, 1449 space->bottom(), space->end(), nta); 1450 assert(result, "space must fit into itself"); 1451 _space_info[i].set_dense_prefix(space->bottom()); 1452 } 1453 } 1454 1455 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1456 { 1457 HeapWord* const dense_prefix_end = dense_prefix(id); 1458 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1459 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1460 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1461 // Only enough dead space is filled so that any remaining dead space to the 1462 // left is larger than the minimum filler object. (The remainder is filled 1463 // during the copy/update phase.) 1464 // 1465 // The size of the dead space to the right of the boundary is not a 1466 // concern, since compaction will be able to use whatever space is 1467 // available. 1468 // 1469 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1470 // surrounds the space to be filled with an object. 1471 // 1472 // In the 32-bit VM, each bit represents two 32-bit words: 1473 // +---+ 1474 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1475 // end_bits: ... x x x | 0 | || 0 x x ... 1476 // +---+ 1477 // 1478 // In the 64-bit VM, each bit represents one 64-bit word: 1479 // +------------+ 1480 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1481 // end_bits: ... x x 1 | 0 || 0 | x x ... 1482 // +------------+ 1483 // +-------+ 1484 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1485 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1486 // +-------+ 1487 // +-----------+ 1488 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1489 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1490 // +-----------+ 1491 // +-------+ 1492 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1493 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1494 // +-------+ 1495 1496 // Initially assume case a, c or e will apply. 1497 size_t obj_len = CollectedHeap::min_fill_size(); 1498 HeapWord* obj_beg = dense_prefix_end - obj_len; 1499 1500 #ifdef _LP64 1501 if (MinObjAlignment > 1) { // object alignment > heap word size 1502 // Cases a, c or e. 1503 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1504 // Case b above. 1505 obj_beg = dense_prefix_end - 1; 1506 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1507 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1508 // Case d above. 1509 obj_beg = dense_prefix_end - 3; 1510 obj_len = 3; 1511 } 1512 #endif // #ifdef _LP64 1513 1514 CollectedHeap::fill_with_object(obj_beg, obj_len); 1515 _mark_bitmap.mark_obj(obj_beg, obj_len); 1516 _summary_data.add_obj(obj_beg, obj_len); 1517 assert(start_array(id) != NULL, "sanity"); 1518 start_array(id)->allocate_block(obj_beg); 1519 } 1520 } 1521 1522 void 1523 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1524 { 1525 assert(id < last_space_id, "id out of range"); 1526 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1527 "should have been reset in summarize_spaces_quick()"); 1528 1529 const MutableSpace* space = _space_info[id].space(); 1530 if (_space_info[id].new_top() != space->bottom()) { 1531 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1532 _space_info[id].set_dense_prefix(dense_prefix_end); 1533 1534 #ifndef PRODUCT 1535 if (TraceParallelOldGCDensePrefix) { 1536 print_dense_prefix_stats("ratio", id, maximum_compaction, 1537 dense_prefix_end); 1538 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1539 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1540 } 1541 #endif // #ifndef PRODUCT 1542 1543 // Recompute the summary data, taking into account the dense prefix. If 1544 // every last byte will be reclaimed, then the existing summary data which 1545 // compacts everything can be left in place. 1546 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1547 // If dead space crosses the dense prefix boundary, it is (at least 1548 // partially) filled with a dummy object, marked live and added to the 1549 // summary data. This simplifies the copy/update phase and must be done 1550 // before the final locations of objects are determined, to prevent 1551 // leaving a fragment of dead space that is too small to fill. 1552 fill_dense_prefix_end(id); 1553 1554 // Compute the destination of each Region, and thus each object. 1555 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1556 _summary_data.summarize(_space_info[id].split_info(), 1557 dense_prefix_end, space->top(), NULL, 1558 dense_prefix_end, space->end(), 1559 _space_info[id].new_top_addr()); 1560 } 1561 } 1562 1563 if (log_develop_is_enabled(Trace, gc, compaction)) { 1564 const size_t region_size = ParallelCompactData::RegionSize; 1565 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1566 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1567 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1568 HeapWord* const new_top = _space_info[id].new_top(); 1569 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1570 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1571 log_develop_trace(gc, compaction)( 1572 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1573 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1574 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1575 id, space->capacity_in_words(), p2i(dense_prefix_end), 1576 dp_region, dp_words / region_size, 1577 cr_words / region_size, p2i(new_top)); 1578 } 1579 } 1580 1581 #ifndef PRODUCT 1582 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1583 HeapWord* dst_beg, HeapWord* dst_end, 1584 SpaceId src_space_id, 1585 HeapWord* src_beg, HeapWord* src_end) 1586 { 1587 log_develop_trace(gc, compaction)( 1588 "Summarizing %d [%s] into %d [%s]: " 1589 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1590 SIZE_FORMAT "-" SIZE_FORMAT " " 1591 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1592 SIZE_FORMAT "-" SIZE_FORMAT, 1593 src_space_id, space_names[src_space_id], 1594 dst_space_id, space_names[dst_space_id], 1595 p2i(src_beg), p2i(src_end), 1596 _summary_data.addr_to_region_idx(src_beg), 1597 _summary_data.addr_to_region_idx(src_end), 1598 p2i(dst_beg), p2i(dst_end), 1599 _summary_data.addr_to_region_idx(dst_beg), 1600 _summary_data.addr_to_region_idx(dst_end)); 1601 } 1602 #endif // #ifndef PRODUCT 1603 1604 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1605 bool maximum_compaction) 1606 { 1607 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1608 1609 #ifdef ASSERT 1610 if (TraceParallelOldGCMarkingPhase) { 1611 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1612 "add_obj_bytes=" SIZE_FORMAT, 1613 add_obj_count, add_obj_size * HeapWordSize); 1614 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1615 "mark_bitmap_bytes=" SIZE_FORMAT, 1616 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1617 } 1618 #endif // #ifdef ASSERT 1619 1620 // Quick summarization of each space into itself, to see how much is live. 1621 summarize_spaces_quick(); 1622 1623 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1624 NOT_PRODUCT(print_region_ranges()); 1625 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1626 1627 // The amount of live data that will end up in old space (assuming it fits). 1628 size_t old_space_total_live = 0; 1629 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1630 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1631 _space_info[id].space()->bottom()); 1632 } 1633 1634 MutableSpace* const old_space = _space_info[old_space_id].space(); 1635 const size_t old_capacity = old_space->capacity_in_words(); 1636 if (old_space_total_live > old_capacity) { 1637 // XXX - should also try to expand 1638 maximum_compaction = true; 1639 } 1640 1641 // Old generations. 1642 summarize_space(old_space_id, maximum_compaction); 1643 1644 // Summarize the remaining spaces in the young gen. The initial target space 1645 // is the old gen. If a space does not fit entirely into the target, then the 1646 // remainder is compacted into the space itself and that space becomes the new 1647 // target. 1648 SpaceId dst_space_id = old_space_id; 1649 HeapWord* dst_space_end = old_space->end(); 1650 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1651 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1652 const MutableSpace* space = _space_info[id].space(); 1653 const size_t live = pointer_delta(_space_info[id].new_top(), 1654 space->bottom()); 1655 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1656 1657 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1658 SpaceId(id), space->bottom(), space->top());) 1659 if (live > 0 && live <= available) { 1660 // All the live data will fit. 1661 bool done = _summary_data.summarize(_space_info[id].split_info(), 1662 space->bottom(), space->top(), 1663 NULL, 1664 *new_top_addr, dst_space_end, 1665 new_top_addr); 1666 assert(done, "space must fit into old gen"); 1667 1668 // Reset the new_top value for the space. 1669 _space_info[id].set_new_top(space->bottom()); 1670 } else if (live > 0) { 1671 // Attempt to fit part of the source space into the target space. 1672 HeapWord* next_src_addr = NULL; 1673 bool done = _summary_data.summarize(_space_info[id].split_info(), 1674 space->bottom(), space->top(), 1675 &next_src_addr, 1676 *new_top_addr, dst_space_end, 1677 new_top_addr); 1678 assert(!done, "space should not fit into old gen"); 1679 assert(next_src_addr != NULL, "sanity"); 1680 1681 // The source space becomes the new target, so the remainder is compacted 1682 // within the space itself. 1683 dst_space_id = SpaceId(id); 1684 dst_space_end = space->end(); 1685 new_top_addr = _space_info[id].new_top_addr(); 1686 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1687 space->bottom(), dst_space_end, 1688 SpaceId(id), next_src_addr, space->top());) 1689 done = _summary_data.summarize(_space_info[id].split_info(), 1690 next_src_addr, space->top(), 1691 NULL, 1692 space->bottom(), dst_space_end, 1693 new_top_addr); 1694 assert(done, "space must fit when compacted into itself"); 1695 assert(*new_top_addr <= space->top(), "usage should not grow"); 1696 } 1697 } 1698 1699 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1700 NOT_PRODUCT(print_region_ranges()); 1701 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1702 } 1703 1704 // This method should contain all heap-specific policy for invoking a full 1705 // collection. invoke_no_policy() will only attempt to compact the heap; it 1706 // will do nothing further. If we need to bail out for policy reasons, scavenge 1707 // before full gc, or any other specialized behavior, it needs to be added here. 1708 // 1709 // Note that this method should only be called from the vm_thread while at a 1710 // safepoint. 1711 // 1712 // Note that the all_soft_refs_clear flag in the collector policy 1713 // may be true because this method can be called without intervening 1714 // activity. For example when the heap space is tight and full measure 1715 // are being taken to free space. 1716 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1717 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1718 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1719 "should be in vm thread"); 1720 1721 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1722 GCCause::Cause gc_cause = heap->gc_cause(); 1723 assert(!heap->is_gc_active(), "not reentrant"); 1724 1725 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1726 IsGCActiveMark mark; 1727 1728 if (ScavengeBeforeFullGC) { 1729 PSScavenge::invoke_no_policy(); 1730 } 1731 1732 const bool clear_all_soft_refs = 1733 heap->soft_ref_policy()->should_clear_all_soft_refs(); 1734 1735 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1736 maximum_heap_compaction); 1737 } 1738 1739 // This method contains no policy. You should probably 1740 // be calling invoke() instead. 1741 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1742 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1743 assert(ref_processor() != NULL, "Sanity"); 1744 1745 if (GCLocker::check_active_before_gc()) { 1746 return false; 1747 } 1748 1749 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1750 1751 GCIdMark gc_id_mark; 1752 _gc_timer.register_gc_start(); 1753 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1754 1755 TimeStamp marking_start; 1756 TimeStamp compaction_start; 1757 TimeStamp collection_exit; 1758 1759 GCCause::Cause gc_cause = heap->gc_cause(); 1760 PSYoungGen* young_gen = heap->young_gen(); 1761 PSOldGen* old_gen = heap->old_gen(); 1762 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1763 1764 // The scope of casr should end after code that can change 1765 // CollectorPolicy::_should_clear_all_soft_refs. 1766 ClearedAllSoftRefs casr(maximum_heap_compaction, 1767 heap->soft_ref_policy()); 1768 1769 if (ZapUnusedHeapArea) { 1770 // Save information needed to minimize mangling 1771 heap->record_gen_tops_before_GC(); 1772 } 1773 1774 // Make sure data structures are sane, make the heap parsable, and do other 1775 // miscellaneous bookkeeping. 1776 pre_compact(); 1777 1778 PreGCValues pre_gc_values(heap); 1779 1780 // Get the compaction manager reserved for the VM thread. 1781 ParCompactionManager* const vmthread_cm = 1782 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1783 1784 { 1785 ResourceMark rm; 1786 HandleMark hm; 1787 1788 // Set the number of GC threads to be used in this collection 1789 gc_task_manager()->set_active_gang(); 1790 gc_task_manager()->task_idle_workers(); 1791 1792 GCTraceCPUTime tcpu; 1793 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1794 1795 heap->pre_full_gc_dump(&_gc_timer); 1796 1797 TraceCollectorStats tcs(counters()); 1798 TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause); 1799 1800 if (log_is_enabled(Debug, gc, heap, exit)) { 1801 accumulated_time()->start(); 1802 } 1803 1804 // Let the size policy know we're starting 1805 size_policy->major_collection_begin(); 1806 1807 #if COMPILER2_OR_JVMCI 1808 DerivedPointerTable::clear(); 1809 #endif 1810 1811 ref_processor()->enable_discovery(); 1812 ref_processor()->setup_policy(maximum_heap_compaction); 1813 1814 bool marked_for_unloading = false; 1815 1816 marking_start.update(); 1817 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1818 1819 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1820 && GCCause::is_user_requested_gc(gc_cause); 1821 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1822 1823 #if COMPILER2_OR_JVMCI 1824 assert(DerivedPointerTable::is_active(), "Sanity"); 1825 DerivedPointerTable::set_active(false); 1826 #endif 1827 1828 // adjust_roots() updates Universe::_intArrayKlassObj which is 1829 // needed by the compaction for filling holes in the dense prefix. 1830 adjust_roots(vmthread_cm); 1831 1832 compaction_start.update(); 1833 compact(); 1834 1835 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1836 // done before resizing. 1837 post_compact(); 1838 1839 // Let the size policy know we're done 1840 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1841 1842 if (UseAdaptiveSizePolicy) { 1843 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1844 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1845 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1846 1847 // Don't check if the size_policy is ready here. Let 1848 // the size_policy check that internally. 1849 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1850 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1851 // Swap the survivor spaces if from_space is empty. The 1852 // resize_young_gen() called below is normally used after 1853 // a successful young GC and swapping of survivor spaces; 1854 // otherwise, it will fail to resize the young gen with 1855 // the current implementation. 1856 if (young_gen->from_space()->is_empty()) { 1857 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1858 young_gen->swap_spaces(); 1859 } 1860 1861 // Calculate optimal free space amounts 1862 assert(young_gen->max_size() > 1863 young_gen->from_space()->capacity_in_bytes() + 1864 young_gen->to_space()->capacity_in_bytes(), 1865 "Sizes of space in young gen are out-of-bounds"); 1866 1867 size_t young_live = young_gen->used_in_bytes(); 1868 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1869 size_t old_live = old_gen->used_in_bytes(); 1870 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1871 size_t max_old_gen_size = old_gen->max_gen_size(); 1872 size_t max_eden_size = young_gen->max_size() - 1873 young_gen->from_space()->capacity_in_bytes() - 1874 young_gen->to_space()->capacity_in_bytes(); 1875 1876 // Used for diagnostics 1877 size_policy->clear_generation_free_space_flags(); 1878 1879 size_policy->compute_generations_free_space(young_live, 1880 eden_live, 1881 old_live, 1882 cur_eden, 1883 max_old_gen_size, 1884 max_eden_size, 1885 true /* full gc*/); 1886 1887 size_policy->check_gc_overhead_limit(eden_live, 1888 max_old_gen_size, 1889 max_eden_size, 1890 true /* full gc*/, 1891 gc_cause, 1892 heap->soft_ref_policy()); 1893 1894 size_policy->decay_supplemental_growth(true /* full gc*/); 1895 1896 heap->resize_old_gen( 1897 size_policy->calculated_old_free_size_in_bytes()); 1898 1899 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1900 size_policy->calculated_survivor_size_in_bytes()); 1901 } 1902 1903 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1904 } 1905 1906 if (UsePerfData) { 1907 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1908 counters->update_counters(); 1909 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1910 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1911 } 1912 1913 heap->resize_all_tlabs(); 1914 1915 // Resize the metaspace capacity after a collection 1916 MetaspaceGC::compute_new_size(); 1917 1918 if (log_is_enabled(Debug, gc, heap, exit)) { 1919 accumulated_time()->stop(); 1920 } 1921 1922 young_gen->print_used_change(pre_gc_values.young_gen_used()); 1923 old_gen->print_used_change(pre_gc_values.old_gen_used()); 1924 MetaspaceUtils::print_metaspace_change(pre_gc_values.metadata_used()); 1925 1926 // Track memory usage and detect low memory 1927 MemoryService::track_memory_usage(); 1928 heap->update_counters(); 1929 gc_task_manager()->release_idle_workers(); 1930 1931 heap->post_full_gc_dump(&_gc_timer); 1932 } 1933 1934 #ifdef ASSERT 1935 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1936 ParCompactionManager* const cm = 1937 ParCompactionManager::manager_array(int(i)); 1938 assert(cm->marking_stack()->is_empty(), "should be empty"); 1939 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1940 } 1941 #endif // ASSERT 1942 1943 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1944 HandleMark hm; // Discard invalid handles created during verification 1945 Universe::verify("After GC"); 1946 } 1947 1948 // Re-verify object start arrays 1949 if (VerifyObjectStartArray && 1950 VerifyAfterGC) { 1951 old_gen->verify_object_start_array(); 1952 } 1953 1954 if (ZapUnusedHeapArea) { 1955 old_gen->object_space()->check_mangled_unused_area_complete(); 1956 } 1957 1958 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1959 1960 collection_exit.update(); 1961 1962 heap->print_heap_after_gc(); 1963 heap->trace_heap_after_gc(&_gc_tracer); 1964 1965 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1966 marking_start.ticks(), compaction_start.ticks(), 1967 collection_exit.ticks()); 1968 gc_task_manager()->print_task_time_stamps(); 1969 1970 #ifdef TRACESPINNING 1971 ParallelTaskTerminator::print_termination_counts(); 1972 #endif 1973 1974 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1975 1976 _gc_timer.register_gc_end(); 1977 1978 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1979 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1980 1981 return true; 1982 } 1983 1984 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1985 PSYoungGen* young_gen, 1986 PSOldGen* old_gen) { 1987 MutableSpace* const eden_space = young_gen->eden_space(); 1988 assert(!eden_space->is_empty(), "eden must be non-empty"); 1989 assert(young_gen->virtual_space()->alignment() == 1990 old_gen->virtual_space()->alignment(), "alignments do not match"); 1991 1992 // We also return false when it's a heterogenous heap because old generation cannot absorb data from eden 1993 // when it is allocated on different memory (example, nv-dimm) than young. 1994 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary) || 1995 ParallelScavengeHeap::heap()->ps_collector_policy()->is_hetero_heap()) { 1996 return false; 1997 } 1998 1999 // Both generations must be completely committed. 2000 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2001 return false; 2002 } 2003 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2004 return false; 2005 } 2006 2007 // Figure out how much to take from eden. Include the average amount promoted 2008 // in the total; otherwise the next young gen GC will simply bail out to a 2009 // full GC. 2010 const size_t alignment = old_gen->virtual_space()->alignment(); 2011 const size_t eden_used = eden_space->used_in_bytes(); 2012 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2013 const size_t absorb_size = align_up(eden_used + promoted, alignment); 2014 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2015 2016 if (absorb_size >= eden_capacity) { 2017 return false; // Must leave some space in eden. 2018 } 2019 2020 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2021 if (new_young_size < young_gen->min_gen_size()) { 2022 return false; // Respect young gen minimum size. 2023 } 2024 2025 log_trace(gc, ergo, heap)(" absorbing " SIZE_FORMAT "K: " 2026 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2027 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2028 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2029 absorb_size / K, 2030 eden_capacity / K, (eden_capacity - absorb_size) / K, 2031 young_gen->from_space()->used_in_bytes() / K, 2032 young_gen->to_space()->used_in_bytes() / K, 2033 young_gen->capacity_in_bytes() / K, new_young_size / K); 2034 2035 // Fill the unused part of the old gen. 2036 MutableSpace* const old_space = old_gen->object_space(); 2037 HeapWord* const unused_start = old_space->top(); 2038 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2039 2040 if (unused_words > 0) { 2041 if (unused_words < CollectedHeap::min_fill_size()) { 2042 return false; // If the old gen cannot be filled, must give up. 2043 } 2044 CollectedHeap::fill_with_objects(unused_start, unused_words); 2045 } 2046 2047 // Take the live data from eden and set both top and end in the old gen to 2048 // eden top. (Need to set end because reset_after_change() mangles the region 2049 // from end to virtual_space->high() in debug builds). 2050 HeapWord* const new_top = eden_space->top(); 2051 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2052 absorb_size); 2053 young_gen->reset_after_change(); 2054 old_space->set_top(new_top); 2055 old_space->set_end(new_top); 2056 old_gen->reset_after_change(); 2057 2058 // Update the object start array for the filler object and the data from eden. 2059 ObjectStartArray* const start_array = old_gen->start_array(); 2060 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2061 start_array->allocate_block(p); 2062 } 2063 2064 // Could update the promoted average here, but it is not typically updated at 2065 // full GCs and the value to use is unclear. Something like 2066 // 2067 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2068 2069 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2070 return true; 2071 } 2072 2073 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2074 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2075 "shouldn't return NULL"); 2076 return ParallelScavengeHeap::gc_task_manager(); 2077 } 2078 2079 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure { 2080 private: 2081 GCTaskQueue* _q; 2082 2083 public: 2084 PCAddThreadRootsMarkingTaskClosure(GCTaskQueue* q) : _q(q) { } 2085 void do_thread(Thread* t) { 2086 _q->enqueue(new ThreadRootsMarkingTask(t)); 2087 } 2088 }; 2089 2090 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2091 bool maximum_heap_compaction, 2092 ParallelOldTracer *gc_tracer) { 2093 // Recursively traverse all live objects and mark them 2094 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2095 2096 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2097 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2098 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2099 TaskQueueSetSuper* qset = ParCompactionManager::stack_array(); 2100 TaskTerminator terminator(active_gc_threads, qset); 2101 2102 PCMarkAndPushClosure mark_and_push_closure(cm); 2103 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2104 2105 // Need new claim bits before marking starts. 2106 ClassLoaderDataGraph::clear_claimed_marks(); 2107 2108 { 2109 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2110 2111 ParallelScavengeHeap::ParStrongRootsScope psrs; 2112 2113 GCTaskQueue* q = GCTaskQueue::create(); 2114 2115 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2116 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2117 // We scan the thread roots in parallel 2118 PCAddThreadRootsMarkingTaskClosure cl(q); 2119 Threads::java_threads_and_vm_thread_do(&cl); 2120 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2121 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2122 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2123 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2124 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2125 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2126 2127 if (active_gc_threads > 1) { 2128 for (uint j = 0; j < active_gc_threads; j++) { 2129 q->enqueue(new StealMarkingTask(terminator.terminator())); 2130 } 2131 } 2132 2133 gc_task_manager()->execute_and_wait(q); 2134 } 2135 2136 // Process reference objects found during marking 2137 { 2138 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2139 2140 ReferenceProcessorStats stats; 2141 ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues()); 2142 2143 if (ref_processor()->processing_is_mt()) { 2144 ref_processor()->set_active_mt_degree(active_gc_threads); 2145 2146 RefProcTaskExecutor task_executor; 2147 stats = ref_processor()->process_discovered_references( 2148 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2149 &task_executor, &pt); 2150 } else { 2151 stats = ref_processor()->process_discovered_references( 2152 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2153 &pt); 2154 } 2155 2156 gc_tracer->report_gc_reference_stats(stats); 2157 pt.print_all_references(); 2158 } 2159 2160 // This is the point where the entire marking should have completed. 2161 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2162 2163 { 2164 GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer); 2165 WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl); 2166 } 2167 2168 { 2169 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2170 2171 // Follow system dictionary roots and unload classes. 2172 bool purged_class = SystemDictionary::do_unloading(&_gc_timer); 2173 2174 // Unload nmethods. 2175 CodeCache::do_unloading(is_alive_closure(), purged_class); 2176 2177 // Prune dead klasses from subklass/sibling/implementor lists. 2178 Klass::clean_weak_klass_links(purged_class); 2179 } 2180 2181 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2182 } 2183 2184 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2185 // Adjust the pointers to reflect the new locations 2186 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2187 2188 // Need new claim bits when tracing through and adjusting pointers. 2189 ClassLoaderDataGraph::clear_claimed_marks(); 2190 2191 PCAdjustPointerClosure oop_closure(cm); 2192 2193 // General strong roots. 2194 Universe::oops_do(&oop_closure); 2195 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2196 Threads::oops_do(&oop_closure, NULL); 2197 ObjectSynchronizer::oops_do(&oop_closure); 2198 Management::oops_do(&oop_closure); 2199 JvmtiExport::oops_do(&oop_closure); 2200 SystemDictionary::oops_do(&oop_closure); 2201 CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong); 2202 ClassLoaderDataGraph::cld_do(&cld_closure); 2203 2204 // Now adjust pointers in remaining weak roots. (All of which should 2205 // have been cleared if they pointed to non-surviving objects.) 2206 WeakProcessor::oops_do(&oop_closure); 2207 2208 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2209 CodeCache::blobs_do(&adjust_from_blobs); 2210 AOTLoader::oops_do(&oop_closure); 2211 ref_processor()->weak_oops_do(&oop_closure); 2212 // Roots were visited so references into the young gen in roots 2213 // may have been scanned. Process them also. 2214 // Should the reference processor have a span that excludes 2215 // young gen objects? 2216 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2217 } 2218 2219 // Helper class to print 8 region numbers per line and then print the total at the end. 2220 class FillableRegionLogger : public StackObj { 2221 private: 2222 Log(gc, compaction) log; 2223 static const int LineLength = 8; 2224 size_t _regions[LineLength]; 2225 int _next_index; 2226 bool _enabled; 2227 size_t _total_regions; 2228 public: 2229 FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { } 2230 ~FillableRegionLogger() { 2231 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2232 } 2233 2234 void print_line() { 2235 if (!_enabled || _next_index == 0) { 2236 return; 2237 } 2238 FormatBuffer<> line("Fillable: "); 2239 for (int i = 0; i < _next_index; i++) { 2240 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2241 } 2242 log.trace("%s", line.buffer()); 2243 _next_index = 0; 2244 } 2245 2246 void handle(size_t region) { 2247 if (!_enabled) { 2248 return; 2249 } 2250 _regions[_next_index++] = region; 2251 if (_next_index == LineLength) { 2252 print_line(); 2253 } 2254 _total_regions++; 2255 } 2256 }; 2257 2258 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q, 2259 uint parallel_gc_threads) 2260 { 2261 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2262 2263 // Find the threads that are active 2264 unsigned int which = 0; 2265 2266 // Find all regions that are available (can be filled immediately) and 2267 // distribute them to the thread stacks. The iteration is done in reverse 2268 // order (high to low) so the regions will be removed in ascending order. 2269 2270 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2271 2272 which = 0; 2273 // id + 1 is used to test termination so unsigned can 2274 // be used with an old_space_id == 0. 2275 FillableRegionLogger region_logger; 2276 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2277 SpaceInfo* const space_info = _space_info + id; 2278 MutableSpace* const space = space_info->space(); 2279 HeapWord* const new_top = space_info->new_top(); 2280 2281 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2282 const size_t end_region = 2283 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2284 2285 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2286 if (sd.region(cur)->claim_unsafe()) { 2287 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2288 cm->region_stack()->push(cur); 2289 region_logger.handle(cur); 2290 // Assign regions to tasks in round-robin fashion. 2291 if (++which == parallel_gc_threads) { 2292 which = 0; 2293 } 2294 } 2295 } 2296 region_logger.print_line(); 2297 } 2298 } 2299 2300 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2301 2302 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2303 uint parallel_gc_threads) { 2304 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2305 2306 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2307 2308 // Iterate over all the spaces adding tasks for updating 2309 // regions in the dense prefix. Assume that 1 gc thread 2310 // will work on opening the gaps and the remaining gc threads 2311 // will work on the dense prefix. 2312 unsigned int space_id; 2313 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2314 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2315 const MutableSpace* const space = _space_info[space_id].space(); 2316 2317 if (dense_prefix_end == space->bottom()) { 2318 // There is no dense prefix for this space. 2319 continue; 2320 } 2321 2322 // The dense prefix is before this region. 2323 size_t region_index_end_dense_prefix = 2324 sd.addr_to_region_idx(dense_prefix_end); 2325 RegionData* const dense_prefix_cp = 2326 sd.region(region_index_end_dense_prefix); 2327 assert(dense_prefix_end == space->end() || 2328 dense_prefix_cp->available() || 2329 dense_prefix_cp->claimed(), 2330 "The region after the dense prefix should always be ready to fill"); 2331 2332 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2333 2334 // Is there dense prefix work? 2335 size_t total_dense_prefix_regions = 2336 region_index_end_dense_prefix - region_index_start; 2337 // How many regions of the dense prefix should be given to 2338 // each thread? 2339 if (total_dense_prefix_regions > 0) { 2340 uint tasks_for_dense_prefix = 1; 2341 if (total_dense_prefix_regions <= 2342 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2343 // Don't over partition. This assumes that 2344 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2345 // so there are not many regions to process. 2346 tasks_for_dense_prefix = parallel_gc_threads; 2347 } else { 2348 // Over partition 2349 tasks_for_dense_prefix = parallel_gc_threads * 2350 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2351 } 2352 size_t regions_per_thread = total_dense_prefix_regions / 2353 tasks_for_dense_prefix; 2354 // Give each thread at least 1 region. 2355 if (regions_per_thread == 0) { 2356 regions_per_thread = 1; 2357 } 2358 2359 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2360 if (region_index_start >= region_index_end_dense_prefix) { 2361 break; 2362 } 2363 // region_index_end is not processed 2364 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2365 region_index_end_dense_prefix); 2366 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2367 region_index_start, 2368 region_index_end)); 2369 region_index_start = region_index_end; 2370 } 2371 } 2372 // This gets any part of the dense prefix that did not 2373 // fit evenly. 2374 if (region_index_start < region_index_end_dense_prefix) { 2375 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2376 region_index_start, 2377 region_index_end_dense_prefix)); 2378 } 2379 } 2380 } 2381 2382 void PSParallelCompact::enqueue_region_stealing_tasks( 2383 GCTaskQueue* q, 2384 ParallelTaskTerminator* terminator_ptr, 2385 uint parallel_gc_threads) { 2386 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer); 2387 2388 // Once a thread has drained it's stack, it should try to steal regions from 2389 // other threads. 2390 for (uint j = 0; j < parallel_gc_threads; j++) { 2391 q->enqueue(new CompactionWithStealingTask(terminator_ptr)); 2392 } 2393 } 2394 2395 #ifdef ASSERT 2396 // Write a histogram of the number of times the block table was filled for a 2397 // region. 2398 void PSParallelCompact::write_block_fill_histogram() 2399 { 2400 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2401 return; 2402 } 2403 2404 Log(gc, compaction) log; 2405 ResourceMark rm; 2406 LogStream ls(log.trace()); 2407 outputStream* out = &ls; 2408 2409 typedef ParallelCompactData::RegionData rd_t; 2410 ParallelCompactData& sd = summary_data(); 2411 2412 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2413 MutableSpace* const spc = _space_info[id].space(); 2414 if (spc->bottom() != spc->top()) { 2415 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2416 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2417 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2418 2419 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2420 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2421 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2422 2423 for (const rd_t* cur = beg; cur < end; ++cur) { 2424 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2425 } 2426 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2427 for (size_t i = 0; i < histo_len; ++i) { 2428 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2429 histo[i], 100.0 * histo[i] / region_cnt); 2430 } 2431 out->cr(); 2432 } 2433 } 2434 } 2435 #endif // #ifdef ASSERT 2436 2437 void PSParallelCompact::compact() { 2438 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2439 2440 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2441 PSOldGen* old_gen = heap->old_gen(); 2442 old_gen->start_array()->reset(); 2443 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2444 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2445 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2446 TaskTerminator terminator(active_gc_threads, qset); 2447 2448 GCTaskQueue* q = GCTaskQueue::create(); 2449 prepare_region_draining_tasks(q, active_gc_threads); 2450 enqueue_dense_prefix_tasks(q, active_gc_threads); 2451 enqueue_region_stealing_tasks(q, terminator.terminator(), active_gc_threads); 2452 2453 { 2454 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2455 2456 gc_task_manager()->execute_and_wait(q); 2457 2458 #ifdef ASSERT 2459 // Verify that all regions have been processed before the deferred updates. 2460 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2461 verify_complete(SpaceId(id)); 2462 } 2463 #endif 2464 } 2465 2466 { 2467 // Update the deferred objects, if any. Any compaction manager can be used. 2468 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2469 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2470 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2471 update_deferred_objects(cm, SpaceId(id)); 2472 } 2473 } 2474 2475 DEBUG_ONLY(write_block_fill_histogram()); 2476 } 2477 2478 #ifdef ASSERT 2479 void PSParallelCompact::verify_complete(SpaceId space_id) { 2480 // All Regions between space bottom() to new_top() should be marked as filled 2481 // and all Regions between new_top() and top() should be available (i.e., 2482 // should have been emptied). 2483 ParallelCompactData& sd = summary_data(); 2484 SpaceInfo si = _space_info[space_id]; 2485 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2486 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2487 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2488 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2489 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2490 2491 bool issued_a_warning = false; 2492 2493 size_t cur_region; 2494 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2495 const RegionData* const c = sd.region(cur_region); 2496 if (!c->completed()) { 2497 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2498 cur_region, c->destination_count()); 2499 issued_a_warning = true; 2500 } 2501 } 2502 2503 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2504 const RegionData* const c = sd.region(cur_region); 2505 if (!c->available()) { 2506 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2507 cur_region, c->destination_count()); 2508 issued_a_warning = true; 2509 } 2510 } 2511 2512 if (issued_a_warning) { 2513 print_region_ranges(); 2514 } 2515 } 2516 #endif // #ifdef ASSERT 2517 2518 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2519 _start_array->allocate_block(addr); 2520 compaction_manager()->update_contents(oop(addr)); 2521 } 2522 2523 // Update interior oops in the ranges of regions [beg_region, end_region). 2524 void 2525 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2526 SpaceId space_id, 2527 size_t beg_region, 2528 size_t end_region) { 2529 ParallelCompactData& sd = summary_data(); 2530 ParMarkBitMap* const mbm = mark_bitmap(); 2531 2532 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2533 HeapWord* const end_addr = sd.region_to_addr(end_region); 2534 assert(beg_region <= end_region, "bad region range"); 2535 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2536 2537 #ifdef ASSERT 2538 // Claim the regions to avoid triggering an assert when they are marked as 2539 // filled. 2540 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2541 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2542 } 2543 #endif // #ifdef ASSERT 2544 2545 if (beg_addr != space(space_id)->bottom()) { 2546 // Find the first live object or block of dead space that *starts* in this 2547 // range of regions. If a partial object crosses onto the region, skip it; 2548 // it will be marked for 'deferred update' when the object head is 2549 // processed. If dead space crosses onto the region, it is also skipped; it 2550 // will be filled when the prior region is processed. If neither of those 2551 // apply, the first word in the region is the start of a live object or dead 2552 // space. 2553 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2554 const RegionData* const cp = sd.region(beg_region); 2555 if (cp->partial_obj_size() != 0) { 2556 beg_addr = sd.partial_obj_end(beg_region); 2557 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2558 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2559 } 2560 } 2561 2562 if (beg_addr < end_addr) { 2563 // A live object or block of dead space starts in this range of Regions. 2564 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2565 2566 // Create closures and iterate. 2567 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2568 FillClosure fill_closure(cm, space_id); 2569 ParMarkBitMap::IterationStatus status; 2570 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2571 dense_prefix_end); 2572 if (status == ParMarkBitMap::incomplete) { 2573 update_closure.do_addr(update_closure.source()); 2574 } 2575 } 2576 2577 // Mark the regions as filled. 2578 RegionData* const beg_cp = sd.region(beg_region); 2579 RegionData* const end_cp = sd.region(end_region); 2580 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2581 cp->set_completed(); 2582 } 2583 } 2584 2585 // Return the SpaceId for the space containing addr. If addr is not in the 2586 // heap, last_space_id is returned. In debug mode it expects the address to be 2587 // in the heap and asserts such. 2588 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2589 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2590 2591 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2592 if (_space_info[id].space()->contains(addr)) { 2593 return SpaceId(id); 2594 } 2595 } 2596 2597 assert(false, "no space contains the addr"); 2598 return last_space_id; 2599 } 2600 2601 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2602 SpaceId id) { 2603 assert(id < last_space_id, "bad space id"); 2604 2605 ParallelCompactData& sd = summary_data(); 2606 const SpaceInfo* const space_info = _space_info + id; 2607 ObjectStartArray* const start_array = space_info->start_array(); 2608 2609 const MutableSpace* const space = space_info->space(); 2610 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2611 HeapWord* const beg_addr = space_info->dense_prefix(); 2612 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2613 2614 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2615 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2616 const RegionData* cur_region; 2617 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2618 HeapWord* const addr = cur_region->deferred_obj_addr(); 2619 if (addr != NULL) { 2620 if (start_array != NULL) { 2621 start_array->allocate_block(addr); 2622 } 2623 cm->update_contents(oop(addr)); 2624 assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2625 } 2626 } 2627 } 2628 2629 // Skip over count live words starting from beg, and return the address of the 2630 // next live word. Unless marked, the word corresponding to beg is assumed to 2631 // be dead. Callers must either ensure beg does not correspond to the middle of 2632 // an object, or account for those live words in some other way. Callers must 2633 // also ensure that there are enough live words in the range [beg, end) to skip. 2634 HeapWord* 2635 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2636 { 2637 assert(count > 0, "sanity"); 2638 2639 ParMarkBitMap* m = mark_bitmap(); 2640 idx_t bits_to_skip = m->words_to_bits(count); 2641 idx_t cur_beg = m->addr_to_bit(beg); 2642 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2643 2644 do { 2645 cur_beg = m->find_obj_beg(cur_beg, search_end); 2646 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2647 const size_t obj_bits = cur_end - cur_beg + 1; 2648 if (obj_bits > bits_to_skip) { 2649 return m->bit_to_addr(cur_beg + bits_to_skip); 2650 } 2651 bits_to_skip -= obj_bits; 2652 cur_beg = cur_end + 1; 2653 } while (bits_to_skip > 0); 2654 2655 // Skipping the desired number of words landed just past the end of an object. 2656 // Find the start of the next object. 2657 cur_beg = m->find_obj_beg(cur_beg, search_end); 2658 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2659 return m->bit_to_addr(cur_beg); 2660 } 2661 2662 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2663 SpaceId src_space_id, 2664 size_t src_region_idx) 2665 { 2666 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2667 2668 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2669 if (split_info.dest_region_addr() == dest_addr) { 2670 // The partial object ending at the split point contains the first word to 2671 // be copied to dest_addr. 2672 return split_info.first_src_addr(); 2673 } 2674 2675 const ParallelCompactData& sd = summary_data(); 2676 ParMarkBitMap* const bitmap = mark_bitmap(); 2677 const size_t RegionSize = ParallelCompactData::RegionSize; 2678 2679 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2680 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2681 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2682 HeapWord* const src_region_destination = src_region_ptr->destination(); 2683 2684 assert(dest_addr >= src_region_destination, "wrong src region"); 2685 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2686 2687 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2688 HeapWord* const src_region_end = src_region_beg + RegionSize; 2689 2690 HeapWord* addr = src_region_beg; 2691 if (dest_addr == src_region_destination) { 2692 // Return the first live word in the source region. 2693 if (partial_obj_size == 0) { 2694 addr = bitmap->find_obj_beg(addr, src_region_end); 2695 assert(addr < src_region_end, "no objects start in src region"); 2696 } 2697 return addr; 2698 } 2699 2700 // Must skip some live data. 2701 size_t words_to_skip = dest_addr - src_region_destination; 2702 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2703 2704 if (partial_obj_size >= words_to_skip) { 2705 // All the live words to skip are part of the partial object. 2706 addr += words_to_skip; 2707 if (partial_obj_size == words_to_skip) { 2708 // Find the first live word past the partial object. 2709 addr = bitmap->find_obj_beg(addr, src_region_end); 2710 assert(addr < src_region_end, "wrong src region"); 2711 } 2712 return addr; 2713 } 2714 2715 // Skip over the partial object (if any). 2716 if (partial_obj_size != 0) { 2717 words_to_skip -= partial_obj_size; 2718 addr += partial_obj_size; 2719 } 2720 2721 // Skip over live words due to objects that start in the region. 2722 addr = skip_live_words(addr, src_region_end, words_to_skip); 2723 assert(addr < src_region_end, "wrong src region"); 2724 return addr; 2725 } 2726 2727 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2728 SpaceId src_space_id, 2729 size_t beg_region, 2730 HeapWord* end_addr) 2731 { 2732 ParallelCompactData& sd = summary_data(); 2733 2734 #ifdef ASSERT 2735 MutableSpace* const src_space = _space_info[src_space_id].space(); 2736 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2737 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2738 "src_space_id does not match beg_addr"); 2739 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2740 "src_space_id does not match end_addr"); 2741 #endif // #ifdef ASSERT 2742 2743 RegionData* const beg = sd.region(beg_region); 2744 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2745 2746 // Regions up to new_top() are enqueued if they become available. 2747 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2748 RegionData* const enqueue_end = 2749 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2750 2751 for (RegionData* cur = beg; cur < end; ++cur) { 2752 assert(cur->data_size() > 0, "region must have live data"); 2753 cur->decrement_destination_count(); 2754 if (cur < enqueue_end && cur->available() && cur->claim()) { 2755 cm->push_region(sd.region(cur)); 2756 } 2757 } 2758 } 2759 2760 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2761 SpaceId& src_space_id, 2762 HeapWord*& src_space_top, 2763 HeapWord* end_addr) 2764 { 2765 typedef ParallelCompactData::RegionData RegionData; 2766 2767 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2768 const size_t region_size = ParallelCompactData::RegionSize; 2769 2770 size_t src_region_idx = 0; 2771 2772 // Skip empty regions (if any) up to the top of the space. 2773 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2774 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2775 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2776 const RegionData* const top_region_ptr = 2777 sd.addr_to_region_ptr(top_aligned_up); 2778 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2779 ++src_region_ptr; 2780 } 2781 2782 if (src_region_ptr < top_region_ptr) { 2783 // The next source region is in the current space. Update src_region_idx 2784 // and the source address to match src_region_ptr. 2785 src_region_idx = sd.region(src_region_ptr); 2786 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2787 if (src_region_addr > closure.source()) { 2788 closure.set_source(src_region_addr); 2789 } 2790 return src_region_idx; 2791 } 2792 2793 // Switch to a new source space and find the first non-empty region. 2794 unsigned int space_id = src_space_id + 1; 2795 assert(space_id < last_space_id, "not enough spaces"); 2796 2797 HeapWord* const destination = closure.destination(); 2798 2799 do { 2800 MutableSpace* space = _space_info[space_id].space(); 2801 HeapWord* const bottom = space->bottom(); 2802 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2803 2804 // Iterate over the spaces that do not compact into themselves. 2805 if (bottom_cp->destination() != bottom) { 2806 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2807 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2808 2809 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2810 if (src_cp->live_obj_size() > 0) { 2811 // Found it. 2812 assert(src_cp->destination() == destination, 2813 "first live obj in the space must match the destination"); 2814 assert(src_cp->partial_obj_size() == 0, 2815 "a space cannot begin with a partial obj"); 2816 2817 src_space_id = SpaceId(space_id); 2818 src_space_top = space->top(); 2819 const size_t src_region_idx = sd.region(src_cp); 2820 closure.set_source(sd.region_to_addr(src_region_idx)); 2821 return src_region_idx; 2822 } else { 2823 assert(src_cp->data_size() == 0, "sanity"); 2824 } 2825 } 2826 } 2827 } while (++space_id < last_space_id); 2828 2829 assert(false, "no source region was found"); 2830 return 0; 2831 } 2832 2833 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 2834 { 2835 typedef ParMarkBitMap::IterationStatus IterationStatus; 2836 const size_t RegionSize = ParallelCompactData::RegionSize; 2837 ParMarkBitMap* const bitmap = mark_bitmap(); 2838 ParallelCompactData& sd = summary_data(); 2839 RegionData* const region_ptr = sd.region(region_idx); 2840 2841 // Get the items needed to construct the closure. 2842 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2843 SpaceId dest_space_id = space_id(dest_addr); 2844 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 2845 HeapWord* new_top = _space_info[dest_space_id].new_top(); 2846 assert(dest_addr < new_top, "sanity"); 2847 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 2848 2849 // Get the source region and related info. 2850 size_t src_region_idx = region_ptr->source_region(); 2851 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2852 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2853 2854 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2855 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2856 2857 // Adjust src_region_idx to prepare for decrementing destination counts (the 2858 // destination count is not decremented when a region is copied to itself). 2859 if (src_region_idx == region_idx) { 2860 src_region_idx += 1; 2861 } 2862 2863 if (bitmap->is_unmarked(closure.source())) { 2864 // The first source word is in the middle of an object; copy the remainder 2865 // of the object or as much as will fit. The fact that pointer updates were 2866 // deferred will be noted when the object header is processed. 2867 HeapWord* const old_src_addr = closure.source(); 2868 closure.copy_partial_obj(); 2869 if (closure.is_full()) { 2870 decrement_destination_counts(cm, src_space_id, src_region_idx, 2871 closure.source()); 2872 region_ptr->set_deferred_obj_addr(NULL); 2873 region_ptr->set_completed(); 2874 return; 2875 } 2876 2877 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2878 if (sd.region_align_down(old_src_addr) != end_addr) { 2879 // The partial object was copied from more than one source region. 2880 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2881 2882 // Move to the next source region, possibly switching spaces as well. All 2883 // args except end_addr may be modified. 2884 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2885 end_addr); 2886 } 2887 } 2888 2889 do { 2890 HeapWord* const cur_addr = closure.source(); 2891 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2892 src_space_top); 2893 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2894 2895 if (status == ParMarkBitMap::incomplete) { 2896 // The last obj that starts in the source region does not end in the 2897 // region. 2898 assert(closure.source() < end_addr, "sanity"); 2899 HeapWord* const obj_beg = closure.source(); 2900 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 2901 src_space_top); 2902 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 2903 if (obj_end < range_end) { 2904 // The end was found; the entire object will fit. 2905 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 2906 assert(status != ParMarkBitMap::would_overflow, "sanity"); 2907 } else { 2908 // The end was not found; the object will not fit. 2909 assert(range_end < src_space_top, "obj cannot cross space boundary"); 2910 status = ParMarkBitMap::would_overflow; 2911 } 2912 } 2913 2914 if (status == ParMarkBitMap::would_overflow) { 2915 // The last object did not fit. Note that interior oop updates were 2916 // deferred, then copy enough of the object to fill the region. 2917 region_ptr->set_deferred_obj_addr(closure.destination()); 2918 status = closure.copy_until_full(); // copies from closure.source() 2919 2920 decrement_destination_counts(cm, src_space_id, src_region_idx, 2921 closure.source()); 2922 region_ptr->set_completed(); 2923 return; 2924 } 2925 2926 if (status == ParMarkBitMap::full) { 2927 decrement_destination_counts(cm, src_space_id, src_region_idx, 2928 closure.source()); 2929 region_ptr->set_deferred_obj_addr(NULL); 2930 region_ptr->set_completed(); 2931 return; 2932 } 2933 2934 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2935 2936 // Move to the next source region, possibly switching spaces as well. All 2937 // args except end_addr may be modified. 2938 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2939 end_addr); 2940 } while (true); 2941 } 2942 2943 void PSParallelCompact::fill_blocks(size_t region_idx) 2944 { 2945 // Fill in the block table elements for the specified region. Each block 2946 // table element holds the number of live words in the region that are to the 2947 // left of the first object that starts in the block. Thus only blocks in 2948 // which an object starts need to be filled. 2949 // 2950 // The algorithm scans the section of the bitmap that corresponds to the 2951 // region, keeping a running total of the live words. When an object start is 2952 // found, if it's the first to start in the block that contains it, the 2953 // current total is written to the block table element. 2954 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2955 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2956 const size_t RegionSize = ParallelCompactData::RegionSize; 2957 2958 ParallelCompactData& sd = summary_data(); 2959 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2960 if (partial_obj_size >= RegionSize) { 2961 return; // No objects start in this region. 2962 } 2963 2964 // Ensure the first loop iteration decides that the block has changed. 2965 size_t cur_block = sd.block_count(); 2966 2967 const ParMarkBitMap* const bitmap = mark_bitmap(); 2968 2969 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2970 assert((size_t)1 << Log2BitsPerBlock == 2971 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2972 2973 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2974 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2975 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2976 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2977 while (beg_bit < range_end) { 2978 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2979 if (new_block != cur_block) { 2980 cur_block = new_block; 2981 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2982 } 2983 2984 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2985 if (end_bit < range_end - 1) { 2986 live_bits += end_bit - beg_bit + 1; 2987 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2988 } else { 2989 return; 2990 } 2991 } 2992 } 2993 2994 void 2995 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 2996 const MutableSpace* sp = space(space_id); 2997 if (sp->is_empty()) { 2998 return; 2999 } 3000 3001 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3002 ParMarkBitMap* const bitmap = mark_bitmap(); 3003 HeapWord* const dp_addr = dense_prefix(space_id); 3004 HeapWord* beg_addr = sp->bottom(); 3005 HeapWord* end_addr = sp->top(); 3006 3007 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3008 3009 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3010 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3011 if (beg_region < dp_region) { 3012 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3013 } 3014 3015 // The destination of the first live object that starts in the region is one 3016 // past the end of the partial object entering the region (if any). 3017 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3018 HeapWord* const new_top = _space_info[space_id].new_top(); 3019 assert(new_top >= dest_addr, "bad new_top value"); 3020 const size_t words = pointer_delta(new_top, dest_addr); 3021 3022 if (words > 0) { 3023 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3024 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3025 3026 ParMarkBitMap::IterationStatus status; 3027 status = bitmap->iterate(&closure, dest_addr, end_addr); 3028 assert(status == ParMarkBitMap::full, "iteration not complete"); 3029 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3030 "live objects skipped because closure is full"); 3031 } 3032 } 3033 3034 jlong PSParallelCompact::millis_since_last_gc() { 3035 // We need a monotonically non-decreasing time in ms but 3036 // os::javaTimeMillis() does not guarantee monotonicity. 3037 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3038 jlong ret_val = now - _time_of_last_gc; 3039 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3040 if (ret_val < 0) { 3041 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3042 return 0; 3043 } 3044 return ret_val; 3045 } 3046 3047 void PSParallelCompact::reset_millis_since_last_gc() { 3048 // We need a monotonically non-decreasing time in ms but 3049 // os::javaTimeMillis() does not guarantee monotonicity. 3050 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3051 } 3052 3053 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3054 { 3055 if (source() != destination()) { 3056 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3057 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3058 } 3059 update_state(words_remaining()); 3060 assert(is_full(), "sanity"); 3061 return ParMarkBitMap::full; 3062 } 3063 3064 void MoveAndUpdateClosure::copy_partial_obj() 3065 { 3066 size_t words = words_remaining(); 3067 3068 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3069 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3070 if (end_addr < range_end) { 3071 words = bitmap()->obj_size(source(), end_addr); 3072 } 3073 3074 // This test is necessary; if omitted, the pointer updates to a partial object 3075 // that crosses the dense prefix boundary could be overwritten. 3076 if (source() != destination()) { 3077 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3078 Copy::aligned_conjoint_words(source(), destination(), words); 3079 } 3080 update_state(words); 3081 } 3082 3083 ParMarkBitMapClosure::IterationStatus 3084 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3085 assert(destination() != NULL, "sanity"); 3086 assert(bitmap()->obj_size(addr) == words, "bad size"); 3087 3088 _source = addr; 3089 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3090 destination(), "wrong destination"); 3091 3092 if (words > words_remaining()) { 3093 return ParMarkBitMap::would_overflow; 3094 } 3095 3096 // The start_array must be updated even if the object is not moving. 3097 if (_start_array != NULL) { 3098 _start_array->allocate_block(destination()); 3099 } 3100 3101 if (destination() != source()) { 3102 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3103 Copy::aligned_conjoint_words(source(), destination(), words); 3104 } 3105 3106 oop moved_oop = (oop) destination(); 3107 compaction_manager()->update_contents(moved_oop); 3108 assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3109 3110 update_state(words); 3111 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3112 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3113 } 3114 3115 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3116 ParCompactionManager* cm, 3117 PSParallelCompact::SpaceId space_id) : 3118 ParMarkBitMapClosure(mbm, cm), 3119 _space_id(space_id), 3120 _start_array(PSParallelCompact::start_array(space_id)) 3121 { 3122 } 3123 3124 // Updates the references in the object to their new values. 3125 ParMarkBitMapClosure::IterationStatus 3126 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3127 do_addr(addr); 3128 return ParMarkBitMap::incomplete; 3129 } 3130 3131 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3132 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3133 _start_array(PSParallelCompact::start_array(space_id)) 3134 { 3135 assert(space_id == PSParallelCompact::old_space_id, 3136 "cannot use FillClosure in the young gen"); 3137 } 3138 3139 ParMarkBitMapClosure::IterationStatus 3140 FillClosure::do_addr(HeapWord* addr, size_t size) { 3141 CollectedHeap::fill_with_objects(addr, size); 3142 HeapWord* const end = addr + size; 3143 do { 3144 _start_array->allocate_block(addr); 3145 addr += oop(addr)->size(); 3146 } while (addr < end); 3147 return ParMarkBitMap::incomplete; 3148 }