1 /* 2 * Copyright (c) 2005, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "aot/aotLoader.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/parallel/gcTaskManager.hpp" 32 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 33 #include "gc/parallel/parMarkBitMap.inline.hpp" 34 #include "gc/parallel/pcTasks.hpp" 35 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 36 #include "gc/parallel/psCompactionManager.inline.hpp" 37 #include "gc/parallel/psMarkSweep.hpp" 38 #include "gc/parallel/psMarkSweepDecorator.hpp" 39 #include "gc/parallel/psOldGen.hpp" 40 #include "gc/parallel/psParallelCompact.inline.hpp" 41 #include "gc/parallel/psPromotionManager.inline.hpp" 42 #include "gc/parallel/psScavenge.hpp" 43 #include "gc/parallel/psYoungGen.hpp" 44 #include "gc/shared/gcCause.hpp" 45 #include "gc/shared/gcHeapSummary.hpp" 46 #include "gc/shared/gcId.hpp" 47 #include "gc/shared/gcLocker.inline.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/isGCActiveMark.hpp" 52 #include "gc/shared/referencePolicy.hpp" 53 #include "gc/shared/referenceProcessor.hpp" 54 #include "gc/shared/spaceDecorator.hpp" 55 #include "logging/log.hpp" 56 #include "memory/resourceArea.hpp" 57 #include "oops/instanceKlass.inline.hpp" 58 #include "oops/instanceMirrorKlass.inline.hpp" 59 #include "oops/methodData.hpp" 60 #include "oops/objArrayKlass.inline.hpp" 61 #include "oops/oop.inline.hpp" 62 #include "runtime/atomic.hpp" 63 #include "runtime/fprofiler.hpp" 64 #include "runtime/safepoint.hpp" 65 #include "runtime/vmThread.hpp" 66 #include "services/management.hpp" 67 #include "services/memTracker.hpp" 68 #include "services/memoryService.hpp" 69 #include "utilities/align.hpp" 70 #include "utilities/debug.hpp" 71 #include "utilities/events.hpp" 72 #include "utilities/formatBuffer.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 #include <math.h> 76 77 // All sizes are in HeapWords. 78 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 79 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 80 const size_t ParallelCompactData::RegionSizeBytes = 81 RegionSize << LogHeapWordSize; 82 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 83 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 84 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 85 86 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 87 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 88 const size_t ParallelCompactData::BlockSizeBytes = 89 BlockSize << LogHeapWordSize; 90 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 91 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 92 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 93 94 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 95 const size_t ParallelCompactData::Log2BlocksPerRegion = 96 Log2RegionSize - Log2BlockSize; 97 98 const ParallelCompactData::RegionData::region_sz_t 99 ParallelCompactData::RegionData::dc_shift = 27; 100 101 const ParallelCompactData::RegionData::region_sz_t 102 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 103 104 const ParallelCompactData::RegionData::region_sz_t 105 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 106 107 const ParallelCompactData::RegionData::region_sz_t 108 ParallelCompactData::RegionData::los_mask = ~dc_mask; 109 110 const ParallelCompactData::RegionData::region_sz_t 111 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 112 113 const ParallelCompactData::RegionData::region_sz_t 114 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 115 116 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 117 118 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 119 120 double PSParallelCompact::_dwl_mean; 121 double PSParallelCompact::_dwl_std_dev; 122 double PSParallelCompact::_dwl_first_term; 123 double PSParallelCompact::_dwl_adjustment; 124 #ifdef ASSERT 125 bool PSParallelCompact::_dwl_initialized = false; 126 #endif // #ifdef ASSERT 127 128 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 129 HeapWord* destination) 130 { 131 assert(src_region_idx != 0, "invalid src_region_idx"); 132 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 133 assert(destination != NULL, "invalid destination argument"); 134 135 _src_region_idx = src_region_idx; 136 _partial_obj_size = partial_obj_size; 137 _destination = destination; 138 139 // These fields may not be updated below, so make sure they're clear. 140 assert(_dest_region_addr == NULL, "should have been cleared"); 141 assert(_first_src_addr == NULL, "should have been cleared"); 142 143 // Determine the number of destination regions for the partial object. 144 HeapWord* const last_word = destination + partial_obj_size - 1; 145 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 146 HeapWord* const beg_region_addr = sd.region_align_down(destination); 147 HeapWord* const end_region_addr = sd.region_align_down(last_word); 148 149 if (beg_region_addr == end_region_addr) { 150 // One destination region. 151 _destination_count = 1; 152 if (end_region_addr == destination) { 153 // The destination falls on a region boundary, thus the first word of the 154 // partial object will be the first word copied to the destination region. 155 _dest_region_addr = end_region_addr; 156 _first_src_addr = sd.region_to_addr(src_region_idx); 157 } 158 } else { 159 // Two destination regions. When copied, the partial object will cross a 160 // destination region boundary, so a word somewhere within the partial 161 // object will be the first word copied to the second destination region. 162 _destination_count = 2; 163 _dest_region_addr = end_region_addr; 164 const size_t ofs = pointer_delta(end_region_addr, destination); 165 assert(ofs < _partial_obj_size, "sanity"); 166 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 167 } 168 } 169 170 void SplitInfo::clear() 171 { 172 _src_region_idx = 0; 173 _partial_obj_size = 0; 174 _destination = NULL; 175 _destination_count = 0; 176 _dest_region_addr = NULL; 177 _first_src_addr = NULL; 178 assert(!is_valid(), "sanity"); 179 } 180 181 #ifdef ASSERT 182 void SplitInfo::verify_clear() 183 { 184 assert(_src_region_idx == 0, "not clear"); 185 assert(_partial_obj_size == 0, "not clear"); 186 assert(_destination == NULL, "not clear"); 187 assert(_destination_count == 0, "not clear"); 188 assert(_dest_region_addr == NULL, "not clear"); 189 assert(_first_src_addr == NULL, "not clear"); 190 } 191 #endif // #ifdef ASSERT 192 193 194 void PSParallelCompact::print_on_error(outputStream* st) { 195 _mark_bitmap.print_on_error(st); 196 } 197 198 #ifndef PRODUCT 199 const char* PSParallelCompact::space_names[] = { 200 "old ", "eden", "from", "to " 201 }; 202 203 void PSParallelCompact::print_region_ranges() { 204 if (!log_develop_is_enabled(Trace, gc, compaction)) { 205 return; 206 } 207 Log(gc, compaction) log; 208 ResourceMark rm; 209 LogStream ls(log.trace()); 210 Universe::print_on(&ls); 211 log.trace("space bottom top end new_top"); 212 log.trace("------ ---------- ---------- ---------- ----------"); 213 214 for (unsigned int id = 0; id < last_space_id; ++id) { 215 const MutableSpace* space = _space_info[id].space(); 216 log.trace("%u %s " 217 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 218 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 219 id, space_names[id], 220 summary_data().addr_to_region_idx(space->bottom()), 221 summary_data().addr_to_region_idx(space->top()), 222 summary_data().addr_to_region_idx(space->end()), 223 summary_data().addr_to_region_idx(_space_info[id].new_top())); 224 } 225 } 226 227 void 228 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 229 { 230 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 231 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 232 233 ParallelCompactData& sd = PSParallelCompact::summary_data(); 234 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 235 log_develop_trace(gc, compaction)( 236 REGION_IDX_FORMAT " " PTR_FORMAT " " 237 REGION_IDX_FORMAT " " PTR_FORMAT " " 238 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 239 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 240 i, p2i(c->data_location()), dci, p2i(c->destination()), 241 c->partial_obj_size(), c->live_obj_size(), 242 c->data_size(), c->source_region(), c->destination_count()); 243 244 #undef REGION_IDX_FORMAT 245 #undef REGION_DATA_FORMAT 246 } 247 248 void 249 print_generic_summary_data(ParallelCompactData& summary_data, 250 HeapWord* const beg_addr, 251 HeapWord* const end_addr) 252 { 253 size_t total_words = 0; 254 size_t i = summary_data.addr_to_region_idx(beg_addr); 255 const size_t last = summary_data.addr_to_region_idx(end_addr); 256 HeapWord* pdest = 0; 257 258 while (i < last) { 259 ParallelCompactData::RegionData* c = summary_data.region(i); 260 if (c->data_size() != 0 || c->destination() != pdest) { 261 print_generic_summary_region(i, c); 262 total_words += c->data_size(); 263 pdest = c->destination(); 264 } 265 ++i; 266 } 267 268 log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 269 } 270 271 void 272 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data, 273 HeapWord* const beg_addr, 274 HeapWord* const end_addr) { 275 ::print_generic_summary_data(summary_data,beg_addr, end_addr); 276 } 277 278 void 279 print_generic_summary_data(ParallelCompactData& summary_data, 280 SpaceInfo* space_info) 281 { 282 if (!log_develop_is_enabled(Trace, gc, compaction)) { 283 return; 284 } 285 286 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 287 const MutableSpace* space = space_info[id].space(); 288 print_generic_summary_data(summary_data, space->bottom(), 289 MAX2(space->top(), space_info[id].new_top())); 290 } 291 } 292 293 void 294 print_initial_summary_data(ParallelCompactData& summary_data, 295 const MutableSpace* space) { 296 if (space->top() == space->bottom()) { 297 return; 298 } 299 300 const size_t region_size = ParallelCompactData::RegionSize; 301 typedef ParallelCompactData::RegionData RegionData; 302 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 303 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 304 const RegionData* c = summary_data.region(end_region - 1); 305 HeapWord* end_addr = c->destination() + c->data_size(); 306 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 307 308 // Print (and count) the full regions at the beginning of the space. 309 size_t full_region_count = 0; 310 size_t i = summary_data.addr_to_region_idx(space->bottom()); 311 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 312 ParallelCompactData::RegionData* c = summary_data.region(i); 313 log_develop_trace(gc, compaction)( 314 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 315 i, p2i(c->destination()), 316 c->partial_obj_size(), c->live_obj_size(), 317 c->data_size(), c->source_region(), c->destination_count()); 318 ++full_region_count; 319 ++i; 320 } 321 322 size_t live_to_right = live_in_space - full_region_count * region_size; 323 324 double max_reclaimed_ratio = 0.0; 325 size_t max_reclaimed_ratio_region = 0; 326 size_t max_dead_to_right = 0; 327 size_t max_live_to_right = 0; 328 329 // Print the 'reclaimed ratio' for regions while there is something live in 330 // the region or to the right of it. The remaining regions are empty (and 331 // uninteresting), and computing the ratio will result in division by 0. 332 while (i < end_region && live_to_right > 0) { 333 c = summary_data.region(i); 334 HeapWord* const region_addr = summary_data.region_to_addr(i); 335 const size_t used_to_right = pointer_delta(space->top(), region_addr); 336 const size_t dead_to_right = used_to_right - live_to_right; 337 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 338 339 if (reclaimed_ratio > max_reclaimed_ratio) { 340 max_reclaimed_ratio = reclaimed_ratio; 341 max_reclaimed_ratio_region = i; 342 max_dead_to_right = dead_to_right; 343 max_live_to_right = live_to_right; 344 } 345 346 ParallelCompactData::RegionData* c = summary_data.region(i); 347 log_develop_trace(gc, compaction)( 348 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d" 349 "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 350 i, p2i(c->destination()), 351 c->partial_obj_size(), c->live_obj_size(), 352 c->data_size(), c->source_region(), c->destination_count(), 353 reclaimed_ratio, dead_to_right, live_to_right); 354 355 356 live_to_right -= c->data_size(); 357 ++i; 358 } 359 360 // Any remaining regions are empty. Print one more if there is one. 361 if (i < end_region) { 362 ParallelCompactData::RegionData* c = summary_data.region(i); 363 log_develop_trace(gc, compaction)( 364 SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 365 i, p2i(c->destination()), 366 c->partial_obj_size(), c->live_obj_size(), 367 c->data_size(), c->source_region(), c->destination_count()); 368 } 369 370 log_develop_trace(gc, compaction)("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 371 max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio); 372 } 373 374 void 375 print_initial_summary_data(ParallelCompactData& summary_data, 376 SpaceInfo* space_info) { 377 if (!log_develop_is_enabled(Trace, gc, compaction)) { 378 return; 379 } 380 381 unsigned int id = PSParallelCompact::old_space_id; 382 const MutableSpace* space; 383 do { 384 space = space_info[id].space(); 385 print_initial_summary_data(summary_data, space); 386 } while (++id < PSParallelCompact::eden_space_id); 387 388 do { 389 space = space_info[id].space(); 390 print_generic_summary_data(summary_data, space->bottom(), space->top()); 391 } while (++id < PSParallelCompact::last_space_id); 392 } 393 #endif // #ifndef PRODUCT 394 395 #ifdef ASSERT 396 size_t add_obj_count; 397 size_t add_obj_size; 398 size_t mark_bitmap_count; 399 size_t mark_bitmap_size; 400 #endif // #ifdef ASSERT 401 402 ParallelCompactData::ParallelCompactData() 403 { 404 _region_start = 0; 405 406 _region_vspace = 0; 407 _reserved_byte_size = 0; 408 _region_data = 0; 409 _region_count = 0; 410 411 _block_vspace = 0; 412 _block_data = 0; 413 _block_count = 0; 414 } 415 416 bool ParallelCompactData::initialize(MemRegion covered_region) 417 { 418 _region_start = covered_region.start(); 419 const size_t region_size = covered_region.word_size(); 420 DEBUG_ONLY(_region_end = _region_start + region_size;) 421 422 assert(region_align_down(_region_start) == _region_start, 423 "region start not aligned"); 424 assert((region_size & RegionSizeOffsetMask) == 0, 425 "region size not a multiple of RegionSize"); 426 427 bool result = initialize_region_data(region_size) && initialize_block_data(); 428 return result; 429 } 430 431 PSVirtualSpace* 432 ParallelCompactData::create_vspace(size_t count, size_t element_size) 433 { 434 const size_t raw_bytes = count * element_size; 435 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 436 const size_t granularity = os::vm_allocation_granularity(); 437 _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity)); 438 439 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 440 MAX2(page_sz, granularity); 441 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 442 os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(), 443 rs.size()); 444 445 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 446 447 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 448 if (vspace != 0) { 449 if (vspace->expand_by(_reserved_byte_size)) { 450 return vspace; 451 } 452 delete vspace; 453 // Release memory reserved in the space. 454 rs.release(); 455 } 456 457 return 0; 458 } 459 460 bool ParallelCompactData::initialize_region_data(size_t region_size) 461 { 462 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 463 _region_vspace = create_vspace(count, sizeof(RegionData)); 464 if (_region_vspace != 0) { 465 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 466 _region_count = count; 467 return true; 468 } 469 return false; 470 } 471 472 bool ParallelCompactData::initialize_block_data() 473 { 474 assert(_region_count != 0, "region data must be initialized first"); 475 const size_t count = _region_count << Log2BlocksPerRegion; 476 _block_vspace = create_vspace(count, sizeof(BlockData)); 477 if (_block_vspace != 0) { 478 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 479 _block_count = count; 480 return true; 481 } 482 return false; 483 } 484 485 void ParallelCompactData::clear() 486 { 487 memset(_region_data, 0, _region_vspace->committed_size()); 488 memset(_block_data, 0, _block_vspace->committed_size()); 489 } 490 491 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 492 assert(beg_region <= _region_count, "beg_region out of range"); 493 assert(end_region <= _region_count, "end_region out of range"); 494 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 495 496 const size_t region_cnt = end_region - beg_region; 497 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 498 499 const size_t beg_block = beg_region * BlocksPerRegion; 500 const size_t block_cnt = region_cnt * BlocksPerRegion; 501 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 502 } 503 504 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 505 { 506 const RegionData* cur_cp = region(region_idx); 507 const RegionData* const end_cp = region(region_count() - 1); 508 509 HeapWord* result = region_to_addr(region_idx); 510 if (cur_cp < end_cp) { 511 do { 512 result += cur_cp->partial_obj_size(); 513 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 514 } 515 return result; 516 } 517 518 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 519 { 520 const size_t obj_ofs = pointer_delta(addr, _region_start); 521 const size_t beg_region = obj_ofs >> Log2RegionSize; 522 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 523 524 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 525 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 526 527 if (beg_region == end_region) { 528 // All in one region. 529 _region_data[beg_region].add_live_obj(len); 530 return; 531 } 532 533 // First region. 534 const size_t beg_ofs = region_offset(addr); 535 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 536 537 Klass* klass = ((oop)addr)->klass(); 538 // Middle regions--completely spanned by this object. 539 for (size_t region = beg_region + 1; region < end_region; ++region) { 540 _region_data[region].set_partial_obj_size(RegionSize); 541 _region_data[region].set_partial_obj_addr(addr); 542 } 543 544 // Last region. 545 const size_t end_ofs = region_offset(addr + len - 1); 546 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 547 _region_data[end_region].set_partial_obj_addr(addr); 548 } 549 550 void 551 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 552 { 553 assert(region_offset(beg) == 0, "not RegionSize aligned"); 554 assert(region_offset(end) == 0, "not RegionSize aligned"); 555 556 size_t cur_region = addr_to_region_idx(beg); 557 const size_t end_region = addr_to_region_idx(end); 558 HeapWord* addr = beg; 559 while (cur_region < end_region) { 560 _region_data[cur_region].set_destination(addr); 561 _region_data[cur_region].set_destination_count(0); 562 _region_data[cur_region].set_source_region(cur_region); 563 _region_data[cur_region].set_data_location(addr); 564 565 // Update live_obj_size so the region appears completely full. 566 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 567 _region_data[cur_region].set_live_obj_size(live_size); 568 569 ++cur_region; 570 addr += RegionSize; 571 } 572 } 573 574 // Find the point at which a space can be split and, if necessary, record the 575 // split point. 576 // 577 // If the current src region (which overflowed the destination space) doesn't 578 // have a partial object, the split point is at the beginning of the current src 579 // region (an "easy" split, no extra bookkeeping required). 580 // 581 // If the current src region has a partial object, the split point is in the 582 // region where that partial object starts (call it the split_region). If 583 // split_region has a partial object, then the split point is just after that 584 // partial object (a "hard" split where we have to record the split data and 585 // zero the partial_obj_size field). With a "hard" split, we know that the 586 // partial_obj ends within split_region because the partial object that caused 587 // the overflow starts in split_region. If split_region doesn't have a partial 588 // obj, then the split is at the beginning of split_region (another "easy" 589 // split). 590 HeapWord* 591 ParallelCompactData::summarize_split_space(size_t src_region, 592 SplitInfo& split_info, 593 HeapWord* destination, 594 HeapWord* target_end, 595 HeapWord** target_next) 596 { 597 assert(destination <= target_end, "sanity"); 598 assert(destination + _region_data[src_region].data_size() > target_end, 599 "region should not fit into target space"); 600 assert(is_region_aligned(target_end), "sanity"); 601 602 size_t split_region = src_region; 603 HeapWord* split_destination = destination; 604 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 605 606 if (destination + partial_obj_size > target_end) { 607 // The split point is just after the partial object (if any) in the 608 // src_region that contains the start of the object that overflowed the 609 // destination space. 610 // 611 // Find the start of the "overflow" object and set split_region to the 612 // region containing it. 613 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 614 split_region = addr_to_region_idx(overflow_obj); 615 616 // Clear the source_region field of all destination regions whose first word 617 // came from data after the split point (a non-null source_region field 618 // implies a region must be filled). 619 // 620 // An alternative to the simple loop below: clear during post_compact(), 621 // which uses memcpy instead of individual stores, and is easy to 622 // parallelize. (The downside is that it clears the entire RegionData 623 // object as opposed to just one field.) 624 // 625 // post_compact() would have to clear the summary data up to the highest 626 // address that was written during the summary phase, which would be 627 // 628 // max(top, max(new_top, clear_top)) 629 // 630 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 631 // to target_end. 632 const RegionData* const sr = region(split_region); 633 const size_t beg_idx = 634 addr_to_region_idx(region_align_up(sr->destination() + 635 sr->partial_obj_size())); 636 const size_t end_idx = addr_to_region_idx(target_end); 637 638 log_develop_trace(gc, compaction)("split: clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx); 639 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 640 _region_data[idx].set_source_region(0); 641 } 642 643 // Set split_destination and partial_obj_size to reflect the split region. 644 split_destination = sr->destination(); 645 partial_obj_size = sr->partial_obj_size(); 646 } 647 648 // The split is recorded only if a partial object extends onto the region. 649 if (partial_obj_size != 0) { 650 _region_data[split_region].set_partial_obj_size(0); 651 split_info.record(split_region, partial_obj_size, split_destination); 652 } 653 654 // Setup the continuation addresses. 655 *target_next = split_destination + partial_obj_size; 656 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 657 658 if (log_develop_is_enabled(Trace, gc, compaction)) { 659 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 660 log_develop_trace(gc, compaction)("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT, 661 split_type, p2i(source_next), split_region, partial_obj_size); 662 log_develop_trace(gc, compaction)("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT, 663 split_type, p2i(split_destination), 664 addr_to_region_idx(split_destination), 665 p2i(*target_next)); 666 667 if (partial_obj_size != 0) { 668 HeapWord* const po_beg = split_info.destination(); 669 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 670 log_develop_trace(gc, compaction)("%s split: po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT, 671 split_type, 672 p2i(po_beg), addr_to_region_idx(po_beg), 673 p2i(po_end), addr_to_region_idx(po_end)); 674 } 675 } 676 677 return source_next; 678 } 679 680 bool ParallelCompactData::summarize(SplitInfo& split_info, 681 HeapWord* source_beg, HeapWord* source_end, 682 HeapWord** source_next, 683 HeapWord* target_beg, HeapWord* target_end, 684 HeapWord** target_next) 685 { 686 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 687 log_develop_trace(gc, compaction)( 688 "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 689 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 690 p2i(source_beg), p2i(source_end), p2i(source_next_val), 691 p2i(target_beg), p2i(target_end), p2i(*target_next)); 692 693 size_t cur_region = addr_to_region_idx(source_beg); 694 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 695 696 HeapWord *dest_addr = target_beg; 697 while (cur_region < end_region) { 698 // The destination must be set even if the region has no data. 699 _region_data[cur_region].set_destination(dest_addr); 700 701 size_t words = _region_data[cur_region].data_size(); 702 if (words > 0) { 703 // If cur_region does not fit entirely into the target space, find a point 704 // at which the source space can be 'split' so that part is copied to the 705 // target space and the rest is copied elsewhere. 706 if (dest_addr + words > target_end) { 707 assert(source_next != NULL, "source_next is NULL when splitting"); 708 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 709 target_end, target_next); 710 return false; 711 } 712 713 // Compute the destination_count for cur_region, and if necessary, update 714 // source_region for a destination region. The source_region field is 715 // updated if cur_region is the first (left-most) region to be copied to a 716 // destination region. 717 // 718 // The destination_count calculation is a bit subtle. A region that has 719 // data that compacts into itself does not count itself as a destination. 720 // This maintains the invariant that a zero count means the region is 721 // available and can be claimed and then filled. 722 uint destination_count = 0; 723 if (split_info.is_split(cur_region)) { 724 // The current region has been split: the partial object will be copied 725 // to one destination space and the remaining data will be copied to 726 // another destination space. Adjust the initial destination_count and, 727 // if necessary, set the source_region field if the partial object will 728 // cross a destination region boundary. 729 destination_count = split_info.destination_count(); 730 if (destination_count == 2) { 731 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 732 _region_data[dest_idx].set_source_region(cur_region); 733 } 734 } 735 736 HeapWord* const last_addr = dest_addr + words - 1; 737 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 738 const size_t dest_region_2 = addr_to_region_idx(last_addr); 739 740 // Initially assume that the destination regions will be the same and 741 // adjust the value below if necessary. Under this assumption, if 742 // cur_region == dest_region_2, then cur_region will be compacted 743 // completely into itself. 744 destination_count += cur_region == dest_region_2 ? 0 : 1; 745 if (dest_region_1 != dest_region_2) { 746 // Destination regions differ; adjust destination_count. 747 destination_count += 1; 748 // Data from cur_region will be copied to the start of dest_region_2. 749 _region_data[dest_region_2].set_source_region(cur_region); 750 } else if (region_offset(dest_addr) == 0) { 751 // Data from cur_region will be copied to the start of the destination 752 // region. 753 _region_data[dest_region_1].set_source_region(cur_region); 754 } 755 756 _region_data[cur_region].set_destination_count(destination_count); 757 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 758 dest_addr += words; 759 } 760 761 ++cur_region; 762 } 763 764 *target_next = dest_addr; 765 return true; 766 } 767 768 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) { 769 assert(addr != NULL, "Should detect NULL oop earlier"); 770 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 771 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 772 773 // Region covering the object. 774 RegionData* const region_ptr = addr_to_region_ptr(addr); 775 HeapWord* result = region_ptr->destination(); 776 777 // If the entire Region is live, the new location is region->destination + the 778 // offset of the object within in the Region. 779 780 // Run some performance tests to determine if this special case pays off. It 781 // is worth it for pointers into the dense prefix. If the optimization to 782 // avoid pointer updates in regions that only point to the dense prefix is 783 // ever implemented, this should be revisited. 784 if (region_ptr->data_size() == RegionSize) { 785 result += region_offset(addr); 786 return result; 787 } 788 789 // Otherwise, the new location is region->destination + block offset + the 790 // number of live words in the Block that are (a) to the left of addr and (b) 791 // due to objects that start in the Block. 792 793 // Fill in the block table if necessary. This is unsynchronized, so multiple 794 // threads may fill the block table for a region (harmless, since it is 795 // idempotent). 796 if (!region_ptr->blocks_filled()) { 797 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 798 region_ptr->set_blocks_filled(); 799 } 800 801 HeapWord* const search_start = block_align_down(addr); 802 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 803 804 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 805 const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr)); 806 result += block_offset + live; 807 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 808 return result; 809 } 810 811 #ifdef ASSERT 812 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 813 { 814 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 815 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 816 for (const size_t* p = beg; p < end; ++p) { 817 assert(*p == 0, "not zero"); 818 } 819 } 820 821 void ParallelCompactData::verify_clear() 822 { 823 verify_clear(_region_vspace); 824 verify_clear(_block_vspace); 825 } 826 #endif // #ifdef ASSERT 827 828 STWGCTimer PSParallelCompact::_gc_timer; 829 ParallelOldTracer PSParallelCompact::_gc_tracer; 830 elapsedTimer PSParallelCompact::_accumulated_time; 831 unsigned int PSParallelCompact::_total_invocations = 0; 832 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 833 jlong PSParallelCompact::_time_of_last_gc = 0; 834 CollectorCounters* PSParallelCompact::_counters = NULL; 835 ParMarkBitMap PSParallelCompact::_mark_bitmap; 836 ParallelCompactData PSParallelCompact::_summary_data; 837 838 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 839 840 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 841 842 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 843 PSParallelCompact::AdjustPointerClosure closure(_cm); 844 klass->oops_do(&closure); 845 } 846 847 void PSParallelCompact::post_initialize() { 848 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 849 MemRegion mr = heap->reserved_region(); 850 _ref_processor = 851 new ReferenceProcessor(mr, // span 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_closure); // non-header is alive closure 858 _counters = new CollectorCounters("PSParallelCompact", 1); 859 860 // Initialize static fields in ParCompactionManager. 861 ParCompactionManager::initialize(mark_bitmap()); 862 } 863 864 bool PSParallelCompact::initialize() { 865 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 866 MemRegion mr = heap->reserved_region(); 867 868 // Was the old gen get allocated successfully? 869 if (!heap->old_gen()->is_allocated()) { 870 return false; 871 } 872 873 initialize_space_info(); 874 initialize_dead_wood_limiter(); 875 876 if (!_mark_bitmap.initialize(mr)) { 877 vm_shutdown_during_initialization( 878 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 879 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 880 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 881 return false; 882 } 883 884 if (!_summary_data.initialize(mr)) { 885 vm_shutdown_during_initialization( 886 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 887 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 888 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 889 return false; 890 } 891 892 return true; 893 } 894 895 void PSParallelCompact::initialize_space_info() 896 { 897 memset(&_space_info, 0, sizeof(_space_info)); 898 899 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 900 PSYoungGen* young_gen = heap->young_gen(); 901 902 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 903 _space_info[eden_space_id].set_space(young_gen->eden_space()); 904 _space_info[from_space_id].set_space(young_gen->from_space()); 905 _space_info[to_space_id].set_space(young_gen->to_space()); 906 907 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 908 } 909 910 void PSParallelCompact::initialize_dead_wood_limiter() 911 { 912 const size_t max = 100; 913 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 914 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 915 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 916 DEBUG_ONLY(_dwl_initialized = true;) 917 _dwl_adjustment = normal_distribution(1.0); 918 } 919 920 void 921 PSParallelCompact::clear_data_covering_space(SpaceId id) 922 { 923 // At this point, top is the value before GC, new_top() is the value that will 924 // be set at the end of GC. The marking bitmap is cleared to top; nothing 925 // should be marked above top. The summary data is cleared to the larger of 926 // top & new_top. 927 MutableSpace* const space = _space_info[id].space(); 928 HeapWord* const bot = space->bottom(); 929 HeapWord* const top = space->top(); 930 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 931 932 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 933 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 934 _mark_bitmap.clear_range(beg_bit, end_bit); 935 936 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 937 const size_t end_region = 938 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 939 _summary_data.clear_range(beg_region, end_region); 940 941 // Clear the data used to 'split' regions. 942 SplitInfo& split_info = _space_info[id].split_info(); 943 if (split_info.is_valid()) { 944 split_info.clear(); 945 } 946 DEBUG_ONLY(split_info.verify_clear();) 947 } 948 949 void PSParallelCompact::pre_compact() 950 { 951 // Update the from & to space pointers in space_info, since they are swapped 952 // at each young gen gc. Do the update unconditionally (even though a 953 // promotion failure does not swap spaces) because an unknown number of young 954 // collections will have swapped the spaces an unknown number of times. 955 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 956 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 957 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 958 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 959 960 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 961 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 962 963 // Increment the invocation count 964 heap->increment_total_collections(true); 965 966 // We need to track unique mark sweep invocations as well. 967 _total_invocations++; 968 969 heap->print_heap_before_gc(); 970 heap->trace_heap_before_gc(&_gc_tracer); 971 972 // Fill in TLABs 973 heap->accumulate_statistics_all_tlabs(); 974 heap->ensure_parsability(true); // retire TLABs 975 976 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 977 HandleMark hm; // Discard invalid handles created during verification 978 Universe::verify("Before GC"); 979 } 980 981 // Verify object start arrays 982 if (VerifyObjectStartArray && 983 VerifyBeforeGC) { 984 heap->old_gen()->verify_object_start_array(); 985 } 986 987 DEBUG_ONLY(mark_bitmap()->verify_clear();) 988 DEBUG_ONLY(summary_data().verify_clear();) 989 990 // Have worker threads release resources the next time they run a task. 991 gc_task_manager()->release_all_resources(); 992 993 ParCompactionManager::reset_all_bitmap_query_caches(); 994 } 995 996 void PSParallelCompact::post_compact() 997 { 998 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 999 1000 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1001 // Clear the marking bitmap, summary data and split info. 1002 clear_data_covering_space(SpaceId(id)); 1003 // Update top(). Must be done after clearing the bitmap and summary data. 1004 _space_info[id].publish_new_top(); 1005 } 1006 1007 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1008 MutableSpace* const from_space = _space_info[from_space_id].space(); 1009 MutableSpace* const to_space = _space_info[to_space_id].space(); 1010 1011 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1012 bool eden_empty = eden_space->is_empty(); 1013 if (!eden_empty) { 1014 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1015 heap->young_gen(), heap->old_gen()); 1016 } 1017 1018 // Update heap occupancy information which is used as input to the soft ref 1019 // clearing policy at the next gc. 1020 Universe::update_heap_info_at_gc(); 1021 1022 bool young_gen_empty = eden_empty && from_space->is_empty() && 1023 to_space->is_empty(); 1024 1025 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1026 MemRegion old_mr = heap->old_gen()->reserved(); 1027 if (young_gen_empty) { 1028 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1029 } else { 1030 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1031 } 1032 1033 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1034 ClassLoaderDataGraph::purge(); 1035 MetaspaceAux::verify_metrics(); 1036 1037 CodeCache::gc_epilogue(); 1038 JvmtiExport::gc_epilogue(); 1039 1040 #if defined(COMPILER2) || INCLUDE_JVMCI 1041 DerivedPointerTable::update_pointers(); 1042 #endif 1043 1044 ref_processor()->enqueue_discovered_references(NULL); 1045 1046 if (ZapUnusedHeapArea) { 1047 heap->gen_mangle_unused_area(); 1048 } 1049 1050 // Update time of last GC 1051 reset_millis_since_last_gc(); 1052 } 1053 1054 HeapWord* 1055 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1056 bool maximum_compaction) 1057 { 1058 const size_t region_size = ParallelCompactData::RegionSize; 1059 const ParallelCompactData& sd = summary_data(); 1060 1061 const MutableSpace* const space = _space_info[id].space(); 1062 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1063 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1064 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1065 1066 // Skip full regions at the beginning of the space--they are necessarily part 1067 // of the dense prefix. 1068 size_t full_count = 0; 1069 const RegionData* cp; 1070 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1071 ++full_count; 1072 } 1073 1074 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1075 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1076 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1077 if (maximum_compaction || cp == end_cp || interval_ended) { 1078 _maximum_compaction_gc_num = total_invocations(); 1079 return sd.region_to_addr(cp); 1080 } 1081 1082 HeapWord* const new_top = _space_info[id].new_top(); 1083 const size_t space_live = pointer_delta(new_top, space->bottom()); 1084 const size_t space_used = space->used_in_words(); 1085 const size_t space_capacity = space->capacity_in_words(); 1086 1087 const double cur_density = double(space_live) / space_capacity; 1088 const double deadwood_density = 1089 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1090 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1091 1092 if (TraceParallelOldGCDensePrefix) { 1093 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1094 cur_density, deadwood_density, deadwood_goal); 1095 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1096 "space_cap=" SIZE_FORMAT, 1097 space_live, space_used, 1098 space_capacity); 1099 } 1100 1101 // XXX - Use binary search? 1102 HeapWord* dense_prefix = sd.region_to_addr(cp); 1103 const RegionData* full_cp = cp; 1104 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1105 while (cp < end_cp) { 1106 HeapWord* region_destination = cp->destination(); 1107 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1108 if (TraceParallelOldGCDensePrefix && Verbose) { 1109 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1110 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1111 sd.region(cp), p2i(region_destination), 1112 p2i(dense_prefix), cur_deadwood); 1113 } 1114 1115 if (cur_deadwood >= deadwood_goal) { 1116 // Found the region that has the correct amount of deadwood to the left. 1117 // This typically occurs after crossing a fairly sparse set of regions, so 1118 // iterate backwards over those sparse regions, looking for the region 1119 // that has the lowest density of live objects 'to the right.' 1120 size_t space_to_left = sd.region(cp) * region_size; 1121 size_t live_to_left = space_to_left - cur_deadwood; 1122 size_t space_to_right = space_capacity - space_to_left; 1123 size_t live_to_right = space_live - live_to_left; 1124 double density_to_right = double(live_to_right) / space_to_right; 1125 while (cp > full_cp) { 1126 --cp; 1127 const size_t prev_region_live_to_right = live_to_right - 1128 cp->data_size(); 1129 const size_t prev_region_space_to_right = space_to_right + region_size; 1130 double prev_region_density_to_right = 1131 double(prev_region_live_to_right) / prev_region_space_to_right; 1132 if (density_to_right <= prev_region_density_to_right) { 1133 return dense_prefix; 1134 } 1135 if (TraceParallelOldGCDensePrefix && Verbose) { 1136 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1137 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1138 prev_region_density_to_right); 1139 } 1140 dense_prefix -= region_size; 1141 live_to_right = prev_region_live_to_right; 1142 space_to_right = prev_region_space_to_right; 1143 density_to_right = prev_region_density_to_right; 1144 } 1145 return dense_prefix; 1146 } 1147 1148 dense_prefix += region_size; 1149 ++cp; 1150 } 1151 1152 return dense_prefix; 1153 } 1154 1155 #ifndef PRODUCT 1156 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1157 const SpaceId id, 1158 const bool maximum_compaction, 1159 HeapWord* const addr) 1160 { 1161 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1162 RegionData* const cp = summary_data().region(region_idx); 1163 const MutableSpace* const space = _space_info[id].space(); 1164 HeapWord* const new_top = _space_info[id].new_top(); 1165 1166 const size_t space_live = pointer_delta(new_top, space->bottom()); 1167 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1168 const size_t space_cap = space->capacity_in_words(); 1169 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1170 const size_t live_to_right = new_top - cp->destination(); 1171 const size_t dead_to_right = space->top() - addr - live_to_right; 1172 1173 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1174 "spl=" SIZE_FORMAT " " 1175 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1176 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1177 " ratio=%10.8f", 1178 algorithm, p2i(addr), region_idx, 1179 space_live, 1180 dead_to_left, dead_to_left_pct, 1181 dead_to_right, live_to_right, 1182 double(dead_to_right) / live_to_right); 1183 } 1184 #endif // #ifndef PRODUCT 1185 1186 // Return a fraction indicating how much of the generation can be treated as 1187 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1188 // based on the density of live objects in the generation to determine a limit, 1189 // which is then adjusted so the return value is min_percent when the density is 1190 // 1. 1191 // 1192 // The following table shows some return values for a different values of the 1193 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1194 // min_percent is 1. 1195 // 1196 // fraction allowed as dead wood 1197 // ----------------------------------------------------------------- 1198 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1199 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1200 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1201 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1202 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1203 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1204 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1205 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1206 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1207 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1208 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1209 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1210 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1211 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1212 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1213 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1214 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1215 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1216 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1217 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1218 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1219 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1220 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1221 1222 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1223 { 1224 assert(_dwl_initialized, "uninitialized"); 1225 1226 // The raw limit is the value of the normal distribution at x = density. 1227 const double raw_limit = normal_distribution(density); 1228 1229 // Adjust the raw limit so it becomes the minimum when the density is 1. 1230 // 1231 // First subtract the adjustment value (which is simply the precomputed value 1232 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1233 // Then add the minimum value, so the minimum is returned when the density is 1234 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1235 const double min = double(min_percent) / 100.0; 1236 const double limit = raw_limit - _dwl_adjustment + min; 1237 return MAX2(limit, 0.0); 1238 } 1239 1240 ParallelCompactData::RegionData* 1241 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1242 const RegionData* end) 1243 { 1244 const size_t region_size = ParallelCompactData::RegionSize; 1245 ParallelCompactData& sd = summary_data(); 1246 size_t left = sd.region(beg); 1247 size_t right = end > beg ? sd.region(end) - 1 : left; 1248 1249 // Binary search. 1250 while (left < right) { 1251 // Equivalent to (left + right) / 2, but does not overflow. 1252 const size_t middle = left + (right - left) / 2; 1253 RegionData* const middle_ptr = sd.region(middle); 1254 HeapWord* const dest = middle_ptr->destination(); 1255 HeapWord* const addr = sd.region_to_addr(middle); 1256 assert(dest != NULL, "sanity"); 1257 assert(dest <= addr, "must move left"); 1258 1259 if (middle > left && dest < addr) { 1260 right = middle - 1; 1261 } else if (middle < right && middle_ptr->data_size() == region_size) { 1262 left = middle + 1; 1263 } else { 1264 return middle_ptr; 1265 } 1266 } 1267 return sd.region(left); 1268 } 1269 1270 ParallelCompactData::RegionData* 1271 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1272 const RegionData* end, 1273 size_t dead_words) 1274 { 1275 ParallelCompactData& sd = summary_data(); 1276 size_t left = sd.region(beg); 1277 size_t right = end > beg ? sd.region(end) - 1 : left; 1278 1279 // Binary search. 1280 while (left < right) { 1281 // Equivalent to (left + right) / 2, but does not overflow. 1282 const size_t middle = left + (right - left) / 2; 1283 RegionData* const middle_ptr = sd.region(middle); 1284 HeapWord* const dest = middle_ptr->destination(); 1285 HeapWord* const addr = sd.region_to_addr(middle); 1286 assert(dest != NULL, "sanity"); 1287 assert(dest <= addr, "must move left"); 1288 1289 const size_t dead_to_left = pointer_delta(addr, dest); 1290 if (middle > left && dead_to_left > dead_words) { 1291 right = middle - 1; 1292 } else if (middle < right && dead_to_left < dead_words) { 1293 left = middle + 1; 1294 } else { 1295 return middle_ptr; 1296 } 1297 } 1298 return sd.region(left); 1299 } 1300 1301 // The result is valid during the summary phase, after the initial summarization 1302 // of each space into itself, and before final summarization. 1303 inline double 1304 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1305 HeapWord* const bottom, 1306 HeapWord* const top, 1307 HeapWord* const new_top) 1308 { 1309 ParallelCompactData& sd = summary_data(); 1310 1311 assert(cp != NULL, "sanity"); 1312 assert(bottom != NULL, "sanity"); 1313 assert(top != NULL, "sanity"); 1314 assert(new_top != NULL, "sanity"); 1315 assert(top >= new_top, "summary data problem?"); 1316 assert(new_top > bottom, "space is empty; should not be here"); 1317 assert(new_top >= cp->destination(), "sanity"); 1318 assert(top >= sd.region_to_addr(cp), "sanity"); 1319 1320 HeapWord* const destination = cp->destination(); 1321 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1322 const size_t compacted_region_live = pointer_delta(new_top, destination); 1323 const size_t compacted_region_used = pointer_delta(top, 1324 sd.region_to_addr(cp)); 1325 const size_t reclaimable = compacted_region_used - compacted_region_live; 1326 1327 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1328 return double(reclaimable) / divisor; 1329 } 1330 1331 // Return the address of the end of the dense prefix, a.k.a. the start of the 1332 // compacted region. The address is always on a region boundary. 1333 // 1334 // Completely full regions at the left are skipped, since no compaction can 1335 // occur in those regions. Then the maximum amount of dead wood to allow is 1336 // computed, based on the density (amount live / capacity) of the generation; 1337 // the region with approximately that amount of dead space to the left is 1338 // identified as the limit region. Regions between the last completely full 1339 // region and the limit region are scanned and the one that has the best 1340 // (maximum) reclaimed_ratio() is selected. 1341 HeapWord* 1342 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1343 bool maximum_compaction) 1344 { 1345 const size_t region_size = ParallelCompactData::RegionSize; 1346 const ParallelCompactData& sd = summary_data(); 1347 1348 const MutableSpace* const space = _space_info[id].space(); 1349 HeapWord* const top = space->top(); 1350 HeapWord* const top_aligned_up = sd.region_align_up(top); 1351 HeapWord* const new_top = _space_info[id].new_top(); 1352 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1353 HeapWord* const bottom = space->bottom(); 1354 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1355 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1356 const RegionData* const new_top_cp = 1357 sd.addr_to_region_ptr(new_top_aligned_up); 1358 1359 // Skip full regions at the beginning of the space--they are necessarily part 1360 // of the dense prefix. 1361 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1362 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1363 space->is_empty(), "no dead space allowed to the left"); 1364 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1365 "region must have dead space"); 1366 1367 // The gc number is saved whenever a maximum compaction is done, and used to 1368 // determine when the maximum compaction interval has expired. This avoids 1369 // successive max compactions for different reasons. 1370 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1371 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1372 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1373 total_invocations() == HeapFirstMaximumCompactionCount; 1374 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1375 _maximum_compaction_gc_num = total_invocations(); 1376 return sd.region_to_addr(full_cp); 1377 } 1378 1379 const size_t space_live = pointer_delta(new_top, bottom); 1380 const size_t space_used = space->used_in_words(); 1381 const size_t space_capacity = space->capacity_in_words(); 1382 1383 const double density = double(space_live) / double(space_capacity); 1384 const size_t min_percent_free = MarkSweepDeadRatio; 1385 const double limiter = dead_wood_limiter(density, min_percent_free); 1386 const size_t dead_wood_max = space_used - space_live; 1387 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1388 dead_wood_max); 1389 1390 if (TraceParallelOldGCDensePrefix) { 1391 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1392 "space_cap=" SIZE_FORMAT, 1393 space_live, space_used, 1394 space_capacity); 1395 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1396 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1397 density, min_percent_free, limiter, 1398 dead_wood_max, dead_wood_limit); 1399 } 1400 1401 // Locate the region with the desired amount of dead space to the left. 1402 const RegionData* const limit_cp = 1403 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1404 1405 // Scan from the first region with dead space to the limit region and find the 1406 // one with the best (largest) reclaimed ratio. 1407 double best_ratio = 0.0; 1408 const RegionData* best_cp = full_cp; 1409 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1410 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1411 if (tmp_ratio > best_ratio) { 1412 best_cp = cp; 1413 best_ratio = tmp_ratio; 1414 } 1415 } 1416 1417 return sd.region_to_addr(best_cp); 1418 } 1419 1420 void PSParallelCompact::summarize_spaces_quick() 1421 { 1422 for (unsigned int i = 0; i < last_space_id; ++i) { 1423 const MutableSpace* space = _space_info[i].space(); 1424 HeapWord** nta = _space_info[i].new_top_addr(); 1425 bool result = _summary_data.summarize(_space_info[i].split_info(), 1426 space->bottom(), space->top(), NULL, 1427 space->bottom(), space->end(), nta); 1428 assert(result, "space must fit into itself"); 1429 _space_info[i].set_dense_prefix(space->bottom()); 1430 } 1431 } 1432 1433 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1434 { 1435 HeapWord* const dense_prefix_end = dense_prefix(id); 1436 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1437 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1438 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1439 // Only enough dead space is filled so that any remaining dead space to the 1440 // left is larger than the minimum filler object. (The remainder is filled 1441 // during the copy/update phase.) 1442 // 1443 // The size of the dead space to the right of the boundary is not a 1444 // concern, since compaction will be able to use whatever space is 1445 // available. 1446 // 1447 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1448 // surrounds the space to be filled with an object. 1449 // 1450 // In the 32-bit VM, each bit represents two 32-bit words: 1451 // +---+ 1452 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1453 // end_bits: ... x x x | 0 | || 0 x x ... 1454 // +---+ 1455 // 1456 // In the 64-bit VM, each bit represents one 64-bit word: 1457 // +------------+ 1458 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1459 // end_bits: ... x x 1 | 0 || 0 | x x ... 1460 // +------------+ 1461 // +-------+ 1462 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1463 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1464 // +-------+ 1465 // +-----------+ 1466 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1467 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1468 // +-----------+ 1469 // +-------+ 1470 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1471 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1472 // +-------+ 1473 1474 // Initially assume case a, c or e will apply. 1475 size_t obj_len = CollectedHeap::min_fill_size(); 1476 HeapWord* obj_beg = dense_prefix_end - obj_len; 1477 1478 #ifdef _LP64 1479 if (MinObjAlignment > 1) { // object alignment > heap word size 1480 // Cases a, c or e. 1481 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1482 // Case b above. 1483 obj_beg = dense_prefix_end - 1; 1484 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1485 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1486 // Case d above. 1487 obj_beg = dense_prefix_end - 3; 1488 obj_len = 3; 1489 } 1490 #endif // #ifdef _LP64 1491 1492 CollectedHeap::fill_with_object(obj_beg, obj_len); 1493 _mark_bitmap.mark_obj(obj_beg, obj_len); 1494 _summary_data.add_obj(obj_beg, obj_len); 1495 assert(start_array(id) != NULL, "sanity"); 1496 start_array(id)->allocate_block(obj_beg); 1497 } 1498 } 1499 1500 void 1501 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1502 { 1503 assert(id < last_space_id, "id out of range"); 1504 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1505 "should have been reset in summarize_spaces_quick()"); 1506 1507 const MutableSpace* space = _space_info[id].space(); 1508 if (_space_info[id].new_top() != space->bottom()) { 1509 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1510 _space_info[id].set_dense_prefix(dense_prefix_end); 1511 1512 #ifndef PRODUCT 1513 if (TraceParallelOldGCDensePrefix) { 1514 print_dense_prefix_stats("ratio", id, maximum_compaction, 1515 dense_prefix_end); 1516 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1517 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1518 } 1519 #endif // #ifndef PRODUCT 1520 1521 // Recompute the summary data, taking into account the dense prefix. If 1522 // every last byte will be reclaimed, then the existing summary data which 1523 // compacts everything can be left in place. 1524 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1525 // If dead space crosses the dense prefix boundary, it is (at least 1526 // partially) filled with a dummy object, marked live and added to the 1527 // summary data. This simplifies the copy/update phase and must be done 1528 // before the final locations of objects are determined, to prevent 1529 // leaving a fragment of dead space that is too small to fill. 1530 fill_dense_prefix_end(id); 1531 1532 // Compute the destination of each Region, and thus each object. 1533 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1534 _summary_data.summarize(_space_info[id].split_info(), 1535 dense_prefix_end, space->top(), NULL, 1536 dense_prefix_end, space->end(), 1537 _space_info[id].new_top_addr()); 1538 } 1539 } 1540 1541 if (log_develop_is_enabled(Trace, gc, compaction)) { 1542 const size_t region_size = ParallelCompactData::RegionSize; 1543 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1544 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1545 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1546 HeapWord* const new_top = _space_info[id].new_top(); 1547 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1548 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1549 log_develop_trace(gc, compaction)( 1550 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1551 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1552 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1553 id, space->capacity_in_words(), p2i(dense_prefix_end), 1554 dp_region, dp_words / region_size, 1555 cr_words / region_size, p2i(new_top)); 1556 } 1557 } 1558 1559 #ifndef PRODUCT 1560 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1561 HeapWord* dst_beg, HeapWord* dst_end, 1562 SpaceId src_space_id, 1563 HeapWord* src_beg, HeapWord* src_end) 1564 { 1565 log_develop_trace(gc, compaction)( 1566 "Summarizing %d [%s] into %d [%s]: " 1567 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1568 SIZE_FORMAT "-" SIZE_FORMAT " " 1569 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1570 SIZE_FORMAT "-" SIZE_FORMAT, 1571 src_space_id, space_names[src_space_id], 1572 dst_space_id, space_names[dst_space_id], 1573 p2i(src_beg), p2i(src_end), 1574 _summary_data.addr_to_region_idx(src_beg), 1575 _summary_data.addr_to_region_idx(src_end), 1576 p2i(dst_beg), p2i(dst_end), 1577 _summary_data.addr_to_region_idx(dst_beg), 1578 _summary_data.addr_to_region_idx(dst_end)); 1579 } 1580 #endif // #ifndef PRODUCT 1581 1582 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1583 bool maximum_compaction) 1584 { 1585 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1586 1587 #ifdef ASSERT 1588 if (TraceParallelOldGCMarkingPhase) { 1589 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1590 "add_obj_bytes=" SIZE_FORMAT, 1591 add_obj_count, add_obj_size * HeapWordSize); 1592 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1593 "mark_bitmap_bytes=" SIZE_FORMAT, 1594 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1595 } 1596 #endif // #ifdef ASSERT 1597 1598 // Quick summarization of each space into itself, to see how much is live. 1599 summarize_spaces_quick(); 1600 1601 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1602 NOT_PRODUCT(print_region_ranges()); 1603 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1604 1605 // The amount of live data that will end up in old space (assuming it fits). 1606 size_t old_space_total_live = 0; 1607 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1608 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1609 _space_info[id].space()->bottom()); 1610 } 1611 1612 MutableSpace* const old_space = _space_info[old_space_id].space(); 1613 const size_t old_capacity = old_space->capacity_in_words(); 1614 if (old_space_total_live > old_capacity) { 1615 // XXX - should also try to expand 1616 maximum_compaction = true; 1617 } 1618 1619 // Old generations. 1620 summarize_space(old_space_id, maximum_compaction); 1621 1622 // Summarize the remaining spaces in the young gen. The initial target space 1623 // is the old gen. If a space does not fit entirely into the target, then the 1624 // remainder is compacted into the space itself and that space becomes the new 1625 // target. 1626 SpaceId dst_space_id = old_space_id; 1627 HeapWord* dst_space_end = old_space->end(); 1628 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1629 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1630 const MutableSpace* space = _space_info[id].space(); 1631 const size_t live = pointer_delta(_space_info[id].new_top(), 1632 space->bottom()); 1633 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1634 1635 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1636 SpaceId(id), space->bottom(), space->top());) 1637 if (live > 0 && live <= available) { 1638 // All the live data will fit. 1639 bool done = _summary_data.summarize(_space_info[id].split_info(), 1640 space->bottom(), space->top(), 1641 NULL, 1642 *new_top_addr, dst_space_end, 1643 new_top_addr); 1644 assert(done, "space must fit into old gen"); 1645 1646 // Reset the new_top value for the space. 1647 _space_info[id].set_new_top(space->bottom()); 1648 } else if (live > 0) { 1649 // Attempt to fit part of the source space into the target space. 1650 HeapWord* next_src_addr = NULL; 1651 bool done = _summary_data.summarize(_space_info[id].split_info(), 1652 space->bottom(), space->top(), 1653 &next_src_addr, 1654 *new_top_addr, dst_space_end, 1655 new_top_addr); 1656 assert(!done, "space should not fit into old gen"); 1657 assert(next_src_addr != NULL, "sanity"); 1658 1659 // The source space becomes the new target, so the remainder is compacted 1660 // within the space itself. 1661 dst_space_id = SpaceId(id); 1662 dst_space_end = space->end(); 1663 new_top_addr = _space_info[id].new_top_addr(); 1664 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1665 space->bottom(), dst_space_end, 1666 SpaceId(id), next_src_addr, space->top());) 1667 done = _summary_data.summarize(_space_info[id].split_info(), 1668 next_src_addr, space->top(), 1669 NULL, 1670 space->bottom(), dst_space_end, 1671 new_top_addr); 1672 assert(done, "space must fit when compacted into itself"); 1673 assert(*new_top_addr <= space->top(), "usage should not grow"); 1674 } 1675 } 1676 1677 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1678 NOT_PRODUCT(print_region_ranges()); 1679 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1680 } 1681 1682 // This method should contain all heap-specific policy for invoking a full 1683 // collection. invoke_no_policy() will only attempt to compact the heap; it 1684 // will do nothing further. If we need to bail out for policy reasons, scavenge 1685 // before full gc, or any other specialized behavior, it needs to be added here. 1686 // 1687 // Note that this method should only be called from the vm_thread while at a 1688 // safepoint. 1689 // 1690 // Note that the all_soft_refs_clear flag in the collector policy 1691 // may be true because this method can be called without intervening 1692 // activity. For example when the heap space is tight and full measure 1693 // are being taken to free space. 1694 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1695 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1696 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1697 "should be in vm thread"); 1698 1699 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1700 GCCause::Cause gc_cause = heap->gc_cause(); 1701 assert(!heap->is_gc_active(), "not reentrant"); 1702 1703 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1704 IsGCActiveMark mark; 1705 1706 if (ScavengeBeforeFullGC) { 1707 PSScavenge::invoke_no_policy(); 1708 } 1709 1710 const bool clear_all_soft_refs = 1711 heap->collector_policy()->should_clear_all_soft_refs(); 1712 1713 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1714 maximum_heap_compaction); 1715 } 1716 1717 // This method contains no policy. You should probably 1718 // be calling invoke() instead. 1719 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1720 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1721 assert(ref_processor() != NULL, "Sanity"); 1722 1723 if (GCLocker::check_active_before_gc()) { 1724 return false; 1725 } 1726 1727 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1728 1729 GCIdMark gc_id_mark; 1730 _gc_timer.register_gc_start(); 1731 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1732 1733 TimeStamp marking_start; 1734 TimeStamp compaction_start; 1735 TimeStamp collection_exit; 1736 1737 GCCause::Cause gc_cause = heap->gc_cause(); 1738 PSYoungGen* young_gen = heap->young_gen(); 1739 PSOldGen* old_gen = heap->old_gen(); 1740 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1741 1742 // The scope of casr should end after code that can change 1743 // CollectorPolicy::_should_clear_all_soft_refs. 1744 ClearedAllSoftRefs casr(maximum_heap_compaction, 1745 heap->collector_policy()); 1746 1747 if (ZapUnusedHeapArea) { 1748 // Save information needed to minimize mangling 1749 heap->record_gen_tops_before_GC(); 1750 } 1751 1752 // Make sure data structures are sane, make the heap parsable, and do other 1753 // miscellaneous bookkeeping. 1754 pre_compact(); 1755 1756 PreGCValues pre_gc_values(heap); 1757 1758 // Get the compaction manager reserved for the VM thread. 1759 ParCompactionManager* const vmthread_cm = 1760 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1761 1762 { 1763 ResourceMark rm; 1764 HandleMark hm; 1765 1766 // Set the number of GC threads to be used in this collection 1767 gc_task_manager()->set_active_gang(); 1768 gc_task_manager()->task_idle_workers(); 1769 1770 GCTraceCPUTime tcpu; 1771 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1772 1773 heap->pre_full_gc_dump(&_gc_timer); 1774 1775 TraceCollectorStats tcs(counters()); 1776 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 1777 1778 if (TraceOldGenTime) accumulated_time()->start(); 1779 1780 // Let the size policy know we're starting 1781 size_policy->major_collection_begin(); 1782 1783 CodeCache::gc_prologue(); 1784 1785 #if defined(COMPILER2) || INCLUDE_JVMCI 1786 DerivedPointerTable::clear(); 1787 #endif 1788 1789 ref_processor()->enable_discovery(); 1790 ref_processor()->setup_policy(maximum_heap_compaction); 1791 1792 bool marked_for_unloading = false; 1793 1794 marking_start.update(); 1795 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1796 1797 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1798 && GCCause::is_user_requested_gc(gc_cause); 1799 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1800 1801 #if defined(COMPILER2) || INCLUDE_JVMCI 1802 assert(DerivedPointerTable::is_active(), "Sanity"); 1803 DerivedPointerTable::set_active(false); 1804 #endif 1805 1806 // adjust_roots() updates Universe::_intArrayKlassObj which is 1807 // needed by the compaction for filling holes in the dense prefix. 1808 adjust_roots(vmthread_cm); 1809 1810 compaction_start.update(); 1811 compact(); 1812 1813 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1814 // done before resizing. 1815 post_compact(); 1816 1817 // Let the size policy know we're done 1818 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1819 1820 if (UseAdaptiveSizePolicy) { 1821 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1822 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1823 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1824 1825 // Don't check if the size_policy is ready here. Let 1826 // the size_policy check that internally. 1827 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1828 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1829 // Swap the survivor spaces if from_space is empty. The 1830 // resize_young_gen() called below is normally used after 1831 // a successful young GC and swapping of survivor spaces; 1832 // otherwise, it will fail to resize the young gen with 1833 // the current implementation. 1834 if (young_gen->from_space()->is_empty()) { 1835 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1836 young_gen->swap_spaces(); 1837 } 1838 1839 // Calculate optimal free space amounts 1840 assert(young_gen->max_size() > 1841 young_gen->from_space()->capacity_in_bytes() + 1842 young_gen->to_space()->capacity_in_bytes(), 1843 "Sizes of space in young gen are out-of-bounds"); 1844 1845 size_t young_live = young_gen->used_in_bytes(); 1846 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1847 size_t old_live = old_gen->used_in_bytes(); 1848 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1849 size_t max_old_gen_size = old_gen->max_gen_size(); 1850 size_t max_eden_size = young_gen->max_size() - 1851 young_gen->from_space()->capacity_in_bytes() - 1852 young_gen->to_space()->capacity_in_bytes(); 1853 1854 // Used for diagnostics 1855 size_policy->clear_generation_free_space_flags(); 1856 1857 size_policy->compute_generations_free_space(young_live, 1858 eden_live, 1859 old_live, 1860 cur_eden, 1861 max_old_gen_size, 1862 max_eden_size, 1863 true /* full gc*/); 1864 1865 size_policy->check_gc_overhead_limit(young_live, 1866 eden_live, 1867 max_old_gen_size, 1868 max_eden_size, 1869 true /* full gc*/, 1870 gc_cause, 1871 heap->collector_policy()); 1872 1873 size_policy->decay_supplemental_growth(true /* full gc*/); 1874 1875 heap->resize_old_gen( 1876 size_policy->calculated_old_free_size_in_bytes()); 1877 1878 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1879 size_policy->calculated_survivor_size_in_bytes()); 1880 } 1881 1882 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1883 } 1884 1885 if (UsePerfData) { 1886 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1887 counters->update_counters(); 1888 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1889 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1890 } 1891 1892 heap->resize_all_tlabs(); 1893 1894 // Resize the metaspace capacity after a collection 1895 MetaspaceGC::compute_new_size(); 1896 1897 if (TraceOldGenTime) { 1898 accumulated_time()->stop(); 1899 } 1900 1901 young_gen->print_used_change(pre_gc_values.young_gen_used()); 1902 old_gen->print_used_change(pre_gc_values.old_gen_used()); 1903 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 1904 1905 // Track memory usage and detect low memory 1906 MemoryService::track_memory_usage(); 1907 heap->update_counters(); 1908 gc_task_manager()->release_idle_workers(); 1909 1910 heap->post_full_gc_dump(&_gc_timer); 1911 } 1912 1913 #ifdef ASSERT 1914 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1915 ParCompactionManager* const cm = 1916 ParCompactionManager::manager_array(int(i)); 1917 assert(cm->marking_stack()->is_empty(), "should be empty"); 1918 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1919 } 1920 #endif // ASSERT 1921 1922 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1923 HandleMark hm; // Discard invalid handles created during verification 1924 Universe::verify("After GC"); 1925 } 1926 1927 // Re-verify object start arrays 1928 if (VerifyObjectStartArray && 1929 VerifyAfterGC) { 1930 old_gen->verify_object_start_array(); 1931 } 1932 1933 if (ZapUnusedHeapArea) { 1934 old_gen->object_space()->check_mangled_unused_area_complete(); 1935 } 1936 1937 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1938 1939 collection_exit.update(); 1940 1941 heap->print_heap_after_gc(); 1942 heap->trace_heap_after_gc(&_gc_tracer); 1943 1944 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1945 marking_start.ticks(), compaction_start.ticks(), 1946 collection_exit.ticks()); 1947 gc_task_manager()->print_task_time_stamps(); 1948 1949 #ifdef TRACESPINNING 1950 ParallelTaskTerminator::print_termination_counts(); 1951 #endif 1952 1953 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1954 1955 _gc_timer.register_gc_end(); 1956 1957 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1958 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1959 1960 return true; 1961 } 1962 1963 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1964 PSYoungGen* young_gen, 1965 PSOldGen* old_gen) { 1966 MutableSpace* const eden_space = young_gen->eden_space(); 1967 assert(!eden_space->is_empty(), "eden must be non-empty"); 1968 assert(young_gen->virtual_space()->alignment() == 1969 old_gen->virtual_space()->alignment(), "alignments do not match"); 1970 1971 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 1972 return false; 1973 } 1974 1975 // Both generations must be completely committed. 1976 if (young_gen->virtual_space()->uncommitted_size() != 0) { 1977 return false; 1978 } 1979 if (old_gen->virtual_space()->uncommitted_size() != 0) { 1980 return false; 1981 } 1982 1983 // Figure out how much to take from eden. Include the average amount promoted 1984 // in the total; otherwise the next young gen GC will simply bail out to a 1985 // full GC. 1986 const size_t alignment = old_gen->virtual_space()->alignment(); 1987 const size_t eden_used = eden_space->used_in_bytes(); 1988 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 1989 const size_t absorb_size = align_up(eden_used + promoted, alignment); 1990 const size_t eden_capacity = eden_space->capacity_in_bytes(); 1991 1992 if (absorb_size >= eden_capacity) { 1993 return false; // Must leave some space in eden. 1994 } 1995 1996 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 1997 if (new_young_size < young_gen->min_gen_size()) { 1998 return false; // Respect young gen minimum size. 1999 } 2000 2001 log_trace(heap, ergo)(" absorbing " SIZE_FORMAT "K: " 2002 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2003 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2004 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2005 absorb_size / K, 2006 eden_capacity / K, (eden_capacity - absorb_size) / K, 2007 young_gen->from_space()->used_in_bytes() / K, 2008 young_gen->to_space()->used_in_bytes() / K, 2009 young_gen->capacity_in_bytes() / K, new_young_size / K); 2010 2011 // Fill the unused part of the old gen. 2012 MutableSpace* const old_space = old_gen->object_space(); 2013 HeapWord* const unused_start = old_space->top(); 2014 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2015 2016 if (unused_words > 0) { 2017 if (unused_words < CollectedHeap::min_fill_size()) { 2018 return false; // If the old gen cannot be filled, must give up. 2019 } 2020 CollectedHeap::fill_with_objects(unused_start, unused_words); 2021 } 2022 2023 // Take the live data from eden and set both top and end in the old gen to 2024 // eden top. (Need to set end because reset_after_change() mangles the region 2025 // from end to virtual_space->high() in debug builds). 2026 HeapWord* const new_top = eden_space->top(); 2027 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2028 absorb_size); 2029 young_gen->reset_after_change(); 2030 old_space->set_top(new_top); 2031 old_space->set_end(new_top); 2032 old_gen->reset_after_change(); 2033 2034 // Update the object start array for the filler object and the data from eden. 2035 ObjectStartArray* const start_array = old_gen->start_array(); 2036 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2037 start_array->allocate_block(p); 2038 } 2039 2040 // Could update the promoted average here, but it is not typically updated at 2041 // full GCs and the value to use is unclear. Something like 2042 // 2043 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2044 2045 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2046 return true; 2047 } 2048 2049 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2050 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2051 "shouldn't return NULL"); 2052 return ParallelScavengeHeap::gc_task_manager(); 2053 } 2054 2055 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2056 bool maximum_heap_compaction, 2057 ParallelOldTracer *gc_tracer) { 2058 // Recursively traverse all live objects and mark them 2059 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2060 2061 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2062 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2063 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2064 TaskQueueSetSuper* qset = ParCompactionManager::stack_array(); 2065 ParallelTaskTerminator terminator(active_gc_threads, qset); 2066 2067 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm); 2068 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2069 2070 // Need new claim bits before marking starts. 2071 ClassLoaderDataGraph::clear_claimed_marks(); 2072 2073 { 2074 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2075 2076 ParallelScavengeHeap::ParStrongRootsScope psrs; 2077 2078 GCTaskQueue* q = GCTaskQueue::create(); 2079 2080 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2081 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2082 // We scan the thread roots in parallel 2083 Threads::create_thread_roots_marking_tasks(q); 2084 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2085 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2086 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2087 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2088 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2089 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2090 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2091 2092 if (active_gc_threads > 1) { 2093 for (uint j = 0; j < active_gc_threads; j++) { 2094 q->enqueue(new StealMarkingTask(&terminator)); 2095 } 2096 } 2097 2098 gc_task_manager()->execute_and_wait(q); 2099 } 2100 2101 // Process reference objects found during marking 2102 { 2103 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2104 2105 ReferenceProcessorStats stats; 2106 if (ref_processor()->processing_is_mt()) { 2107 RefProcTaskExecutor task_executor; 2108 stats = ref_processor()->process_discovered_references( 2109 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2110 &task_executor, &_gc_timer); 2111 } else { 2112 stats = ref_processor()->process_discovered_references( 2113 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2114 &_gc_timer); 2115 } 2116 2117 gc_tracer->report_gc_reference_stats(stats); 2118 } 2119 2120 // This is the point where the entire marking should have completed. 2121 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2122 2123 { 2124 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2125 2126 // Follow system dictionary roots and unload classes. 2127 bool purged_class = SystemDictionary::do_unloading(is_alive_closure(), &_gc_timer); 2128 2129 // Unload nmethods. 2130 CodeCache::do_unloading(is_alive_closure(), purged_class); 2131 2132 // Prune dead klasses from subklass/sibling/implementor lists. 2133 Klass::clean_weak_klass_links(is_alive_closure()); 2134 } 2135 2136 { 2137 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer); 2138 // Delete entries for dead interned strings. 2139 StringTable::unlink(is_alive_closure()); 2140 } 2141 2142 { 2143 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer); 2144 // Clean up unreferenced symbols in symbol table. 2145 SymbolTable::unlink(); 2146 } 2147 2148 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2149 } 2150 2151 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2152 // Adjust the pointers to reflect the new locations 2153 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2154 2155 // Need new claim bits when tracing through and adjusting pointers. 2156 ClassLoaderDataGraph::clear_claimed_marks(); 2157 2158 PSParallelCompact::AdjustPointerClosure oop_closure(cm); 2159 PSParallelCompact::AdjustKlassClosure klass_closure(cm); 2160 2161 // General strong roots. 2162 Universe::oops_do(&oop_closure); 2163 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2164 Threads::oops_do(&oop_closure, NULL); 2165 ObjectSynchronizer::oops_do(&oop_closure); 2166 FlatProfiler::oops_do(&oop_closure); 2167 Management::oops_do(&oop_closure); 2168 JvmtiExport::oops_do(&oop_closure); 2169 SystemDictionary::oops_do(&oop_closure); 2170 ClassLoaderDataGraph::oops_do(&oop_closure, &klass_closure, true); 2171 2172 // Now adjust pointers in remaining weak roots. (All of which should 2173 // have been cleared if they pointed to non-surviving objects.) 2174 // Global (weak) JNI handles 2175 JNIHandles::weak_oops_do(&oop_closure); 2176 2177 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2178 CodeCache::blobs_do(&adjust_from_blobs); 2179 AOTLoader::oops_do(&oop_closure); 2180 StringTable::oops_do(&oop_closure); 2181 ref_processor()->weak_oops_do(&oop_closure); 2182 // Roots were visited so references into the young gen in roots 2183 // may have been scanned. Process them also. 2184 // Should the reference processor have a span that excludes 2185 // young gen objects? 2186 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2187 } 2188 2189 // Helper class to print 8 region numbers per line and then print the total at the end. 2190 class FillableRegionLogger : public StackObj { 2191 private: 2192 Log(gc, compaction) log; 2193 static const int LineLength = 8; 2194 size_t _regions[LineLength]; 2195 int _next_index; 2196 bool _enabled; 2197 size_t _total_regions; 2198 public: 2199 FillableRegionLogger() : _next_index(0), _total_regions(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)) { } 2200 ~FillableRegionLogger() { 2201 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2202 } 2203 2204 void print_line() { 2205 if (!_enabled || _next_index == 0) { 2206 return; 2207 } 2208 FormatBuffer<> line("Fillable: "); 2209 for (int i = 0; i < _next_index; i++) { 2210 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2211 } 2212 log.trace("%s", line.buffer()); 2213 _next_index = 0; 2214 } 2215 2216 void handle(size_t region) { 2217 if (!_enabled) { 2218 return; 2219 } 2220 _regions[_next_index++] = region; 2221 if (_next_index == LineLength) { 2222 print_line(); 2223 } 2224 _total_regions++; 2225 } 2226 }; 2227 2228 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q, 2229 uint parallel_gc_threads) 2230 { 2231 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2232 2233 // Find the threads that are active 2234 unsigned int which = 0; 2235 2236 // Find all regions that are available (can be filled immediately) and 2237 // distribute them to the thread stacks. The iteration is done in reverse 2238 // order (high to low) so the regions will be removed in ascending order. 2239 2240 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2241 2242 which = 0; 2243 // id + 1 is used to test termination so unsigned can 2244 // be used with an old_space_id == 0. 2245 FillableRegionLogger region_logger; 2246 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2247 SpaceInfo* const space_info = _space_info + id; 2248 MutableSpace* const space = space_info->space(); 2249 HeapWord* const new_top = space_info->new_top(); 2250 2251 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2252 const size_t end_region = 2253 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2254 2255 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2256 if (sd.region(cur)->claim_unsafe()) { 2257 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2258 cm->region_stack()->push(cur); 2259 region_logger.handle(cur); 2260 // Assign regions to tasks in round-robin fashion. 2261 if (++which == parallel_gc_threads) { 2262 which = 0; 2263 } 2264 } 2265 } 2266 region_logger.print_line(); 2267 } 2268 } 2269 2270 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2271 2272 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2273 uint parallel_gc_threads) { 2274 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2275 2276 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2277 2278 // Iterate over all the spaces adding tasks for updating 2279 // regions in the dense prefix. Assume that 1 gc thread 2280 // will work on opening the gaps and the remaining gc threads 2281 // will work on the dense prefix. 2282 unsigned int space_id; 2283 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2284 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2285 const MutableSpace* const space = _space_info[space_id].space(); 2286 2287 if (dense_prefix_end == space->bottom()) { 2288 // There is no dense prefix for this space. 2289 continue; 2290 } 2291 2292 // The dense prefix is before this region. 2293 size_t region_index_end_dense_prefix = 2294 sd.addr_to_region_idx(dense_prefix_end); 2295 RegionData* const dense_prefix_cp = 2296 sd.region(region_index_end_dense_prefix); 2297 assert(dense_prefix_end == space->end() || 2298 dense_prefix_cp->available() || 2299 dense_prefix_cp->claimed(), 2300 "The region after the dense prefix should always be ready to fill"); 2301 2302 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2303 2304 // Is there dense prefix work? 2305 size_t total_dense_prefix_regions = 2306 region_index_end_dense_prefix - region_index_start; 2307 // How many regions of the dense prefix should be given to 2308 // each thread? 2309 if (total_dense_prefix_regions > 0) { 2310 uint tasks_for_dense_prefix = 1; 2311 if (total_dense_prefix_regions <= 2312 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2313 // Don't over partition. This assumes that 2314 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2315 // so there are not many regions to process. 2316 tasks_for_dense_prefix = parallel_gc_threads; 2317 } else { 2318 // Over partition 2319 tasks_for_dense_prefix = parallel_gc_threads * 2320 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2321 } 2322 size_t regions_per_thread = total_dense_prefix_regions / 2323 tasks_for_dense_prefix; 2324 // Give each thread at least 1 region. 2325 if (regions_per_thread == 0) { 2326 regions_per_thread = 1; 2327 } 2328 2329 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2330 if (region_index_start >= region_index_end_dense_prefix) { 2331 break; 2332 } 2333 // region_index_end is not processed 2334 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2335 region_index_end_dense_prefix); 2336 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2337 region_index_start, 2338 region_index_end)); 2339 region_index_start = region_index_end; 2340 } 2341 } 2342 // This gets any part of the dense prefix that did not 2343 // fit evenly. 2344 if (region_index_start < region_index_end_dense_prefix) { 2345 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2346 region_index_start, 2347 region_index_end_dense_prefix)); 2348 } 2349 } 2350 } 2351 2352 void PSParallelCompact::enqueue_region_stealing_tasks( 2353 GCTaskQueue* q, 2354 ParallelTaskTerminator* terminator_ptr, 2355 uint parallel_gc_threads) { 2356 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer); 2357 2358 // Once a thread has drained it's stack, it should try to steal regions from 2359 // other threads. 2360 for (uint j = 0; j < parallel_gc_threads; j++) { 2361 q->enqueue(new CompactionWithStealingTask(terminator_ptr)); 2362 } 2363 } 2364 2365 #ifdef ASSERT 2366 // Write a histogram of the number of times the block table was filled for a 2367 // region. 2368 void PSParallelCompact::write_block_fill_histogram() 2369 { 2370 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2371 return; 2372 } 2373 2374 Log(gc, compaction) log; 2375 ResourceMark rm; 2376 LogStream ls(log.trace()); 2377 outputStream* out = &ls; 2378 2379 typedef ParallelCompactData::RegionData rd_t; 2380 ParallelCompactData& sd = summary_data(); 2381 2382 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2383 MutableSpace* const spc = _space_info[id].space(); 2384 if (spc->bottom() != spc->top()) { 2385 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2386 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2387 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2388 2389 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2390 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2391 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2392 2393 for (const rd_t* cur = beg; cur < end; ++cur) { 2394 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2395 } 2396 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2397 for (size_t i = 0; i < histo_len; ++i) { 2398 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2399 histo[i], 100.0 * histo[i] / region_cnt); 2400 } 2401 out->cr(); 2402 } 2403 } 2404 } 2405 #endif // #ifdef ASSERT 2406 2407 void PSParallelCompact::compact() { 2408 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2409 2410 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2411 PSOldGen* old_gen = heap->old_gen(); 2412 old_gen->start_array()->reset(); 2413 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2414 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2415 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2416 ParallelTaskTerminator terminator(active_gc_threads, qset); 2417 2418 GCTaskQueue* q = GCTaskQueue::create(); 2419 prepare_region_draining_tasks(q, active_gc_threads); 2420 enqueue_dense_prefix_tasks(q, active_gc_threads); 2421 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2422 2423 { 2424 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2425 2426 gc_task_manager()->execute_and_wait(q); 2427 2428 #ifdef ASSERT 2429 // Verify that all regions have been processed before the deferred updates. 2430 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2431 verify_complete(SpaceId(id)); 2432 } 2433 #endif 2434 } 2435 2436 { 2437 // Update the deferred objects, if any. Any compaction manager can be used. 2438 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2439 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2440 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2441 update_deferred_objects(cm, SpaceId(id)); 2442 } 2443 } 2444 2445 DEBUG_ONLY(write_block_fill_histogram()); 2446 } 2447 2448 #ifdef ASSERT 2449 void PSParallelCompact::verify_complete(SpaceId space_id) { 2450 // All Regions between space bottom() to new_top() should be marked as filled 2451 // and all Regions between new_top() and top() should be available (i.e., 2452 // should have been emptied). 2453 ParallelCompactData& sd = summary_data(); 2454 SpaceInfo si = _space_info[space_id]; 2455 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2456 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2457 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2458 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2459 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2460 2461 bool issued_a_warning = false; 2462 2463 size_t cur_region; 2464 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2465 const RegionData* const c = sd.region(cur_region); 2466 if (!c->completed()) { 2467 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2468 cur_region, c->destination_count()); 2469 issued_a_warning = true; 2470 } 2471 } 2472 2473 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2474 const RegionData* const c = sd.region(cur_region); 2475 if (!c->available()) { 2476 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2477 cur_region, c->destination_count()); 2478 issued_a_warning = true; 2479 } 2480 } 2481 2482 if (issued_a_warning) { 2483 print_region_ranges(); 2484 } 2485 } 2486 #endif // #ifdef ASSERT 2487 2488 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2489 _start_array->allocate_block(addr); 2490 compaction_manager()->update_contents(oop(addr)); 2491 } 2492 2493 // Update interior oops in the ranges of regions [beg_region, end_region). 2494 void 2495 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2496 SpaceId space_id, 2497 size_t beg_region, 2498 size_t end_region) { 2499 ParallelCompactData& sd = summary_data(); 2500 ParMarkBitMap* const mbm = mark_bitmap(); 2501 2502 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2503 HeapWord* const end_addr = sd.region_to_addr(end_region); 2504 assert(beg_region <= end_region, "bad region range"); 2505 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2506 2507 #ifdef ASSERT 2508 // Claim the regions to avoid triggering an assert when they are marked as 2509 // filled. 2510 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2511 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2512 } 2513 #endif // #ifdef ASSERT 2514 2515 if (beg_addr != space(space_id)->bottom()) { 2516 // Find the first live object or block of dead space that *starts* in this 2517 // range of regions. If a partial object crosses onto the region, skip it; 2518 // it will be marked for 'deferred update' when the object head is 2519 // processed. If dead space crosses onto the region, it is also skipped; it 2520 // will be filled when the prior region is processed. If neither of those 2521 // apply, the first word in the region is the start of a live object or dead 2522 // space. 2523 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2524 const RegionData* const cp = sd.region(beg_region); 2525 if (cp->partial_obj_size() != 0) { 2526 beg_addr = sd.partial_obj_end(beg_region); 2527 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2528 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2529 } 2530 } 2531 2532 if (beg_addr < end_addr) { 2533 // A live object or block of dead space starts in this range of Regions. 2534 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2535 2536 // Create closures and iterate. 2537 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2538 FillClosure fill_closure(cm, space_id); 2539 ParMarkBitMap::IterationStatus status; 2540 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2541 dense_prefix_end); 2542 if (status == ParMarkBitMap::incomplete) { 2543 update_closure.do_addr(update_closure.source()); 2544 } 2545 } 2546 2547 // Mark the regions as filled. 2548 RegionData* const beg_cp = sd.region(beg_region); 2549 RegionData* const end_cp = sd.region(end_region); 2550 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2551 cp->set_completed(); 2552 } 2553 } 2554 2555 // Return the SpaceId for the space containing addr. If addr is not in the 2556 // heap, last_space_id is returned. In debug mode it expects the address to be 2557 // in the heap and asserts such. 2558 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2559 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2560 2561 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2562 if (_space_info[id].space()->contains(addr)) { 2563 return SpaceId(id); 2564 } 2565 } 2566 2567 assert(false, "no space contains the addr"); 2568 return last_space_id; 2569 } 2570 2571 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2572 SpaceId id) { 2573 assert(id < last_space_id, "bad space id"); 2574 2575 ParallelCompactData& sd = summary_data(); 2576 const SpaceInfo* const space_info = _space_info + id; 2577 ObjectStartArray* const start_array = space_info->start_array(); 2578 2579 const MutableSpace* const space = space_info->space(); 2580 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2581 HeapWord* const beg_addr = space_info->dense_prefix(); 2582 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2583 2584 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2585 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2586 const RegionData* cur_region; 2587 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2588 HeapWord* const addr = cur_region->deferred_obj_addr(); 2589 if (addr != NULL) { 2590 if (start_array != NULL) { 2591 start_array->allocate_block(addr); 2592 } 2593 cm->update_contents(oop(addr)); 2594 assert(oop(addr)->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2595 } 2596 } 2597 } 2598 2599 // Skip over count live words starting from beg, and return the address of the 2600 // next live word. Unless marked, the word corresponding to beg is assumed to 2601 // be dead. Callers must either ensure beg does not correspond to the middle of 2602 // an object, or account for those live words in some other way. Callers must 2603 // also ensure that there are enough live words in the range [beg, end) to skip. 2604 HeapWord* 2605 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2606 { 2607 assert(count > 0, "sanity"); 2608 2609 ParMarkBitMap* m = mark_bitmap(); 2610 idx_t bits_to_skip = m->words_to_bits(count); 2611 idx_t cur_beg = m->addr_to_bit(beg); 2612 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2613 2614 do { 2615 cur_beg = m->find_obj_beg(cur_beg, search_end); 2616 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2617 const size_t obj_bits = cur_end - cur_beg + 1; 2618 if (obj_bits > bits_to_skip) { 2619 return m->bit_to_addr(cur_beg + bits_to_skip); 2620 } 2621 bits_to_skip -= obj_bits; 2622 cur_beg = cur_end + 1; 2623 } while (bits_to_skip > 0); 2624 2625 // Skipping the desired number of words landed just past the end of an object. 2626 // Find the start of the next object. 2627 cur_beg = m->find_obj_beg(cur_beg, search_end); 2628 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2629 return m->bit_to_addr(cur_beg); 2630 } 2631 2632 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2633 SpaceId src_space_id, 2634 size_t src_region_idx) 2635 { 2636 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2637 2638 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2639 if (split_info.dest_region_addr() == dest_addr) { 2640 // The partial object ending at the split point contains the first word to 2641 // be copied to dest_addr. 2642 return split_info.first_src_addr(); 2643 } 2644 2645 const ParallelCompactData& sd = summary_data(); 2646 ParMarkBitMap* const bitmap = mark_bitmap(); 2647 const size_t RegionSize = ParallelCompactData::RegionSize; 2648 2649 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2650 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2651 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2652 HeapWord* const src_region_destination = src_region_ptr->destination(); 2653 2654 assert(dest_addr >= src_region_destination, "wrong src region"); 2655 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2656 2657 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2658 HeapWord* const src_region_end = src_region_beg + RegionSize; 2659 2660 HeapWord* addr = src_region_beg; 2661 if (dest_addr == src_region_destination) { 2662 // Return the first live word in the source region. 2663 if (partial_obj_size == 0) { 2664 addr = bitmap->find_obj_beg(addr, src_region_end); 2665 assert(addr < src_region_end, "no objects start in src region"); 2666 } 2667 return addr; 2668 } 2669 2670 // Must skip some live data. 2671 size_t words_to_skip = dest_addr - src_region_destination; 2672 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2673 2674 if (partial_obj_size >= words_to_skip) { 2675 // All the live words to skip are part of the partial object. 2676 addr += words_to_skip; 2677 if (partial_obj_size == words_to_skip) { 2678 // Find the first live word past the partial object. 2679 addr = bitmap->find_obj_beg(addr, src_region_end); 2680 assert(addr < src_region_end, "wrong src region"); 2681 } 2682 return addr; 2683 } 2684 2685 // Skip over the partial object (if any). 2686 if (partial_obj_size != 0) { 2687 words_to_skip -= partial_obj_size; 2688 addr += partial_obj_size; 2689 } 2690 2691 // Skip over live words due to objects that start in the region. 2692 addr = skip_live_words(addr, src_region_end, words_to_skip); 2693 assert(addr < src_region_end, "wrong src region"); 2694 return addr; 2695 } 2696 2697 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2698 SpaceId src_space_id, 2699 size_t beg_region, 2700 HeapWord* end_addr) 2701 { 2702 ParallelCompactData& sd = summary_data(); 2703 2704 #ifdef ASSERT 2705 MutableSpace* const src_space = _space_info[src_space_id].space(); 2706 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2707 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2708 "src_space_id does not match beg_addr"); 2709 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2710 "src_space_id does not match end_addr"); 2711 #endif // #ifdef ASSERT 2712 2713 RegionData* const beg = sd.region(beg_region); 2714 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2715 2716 // Regions up to new_top() are enqueued if they become available. 2717 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2718 RegionData* const enqueue_end = 2719 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2720 2721 for (RegionData* cur = beg; cur < end; ++cur) { 2722 assert(cur->data_size() > 0, "region must have live data"); 2723 cur->decrement_destination_count(); 2724 if (cur < enqueue_end && cur->available() && cur->claim()) { 2725 cm->push_region(sd.region(cur)); 2726 } 2727 } 2728 } 2729 2730 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2731 SpaceId& src_space_id, 2732 HeapWord*& src_space_top, 2733 HeapWord* end_addr) 2734 { 2735 typedef ParallelCompactData::RegionData RegionData; 2736 2737 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2738 const size_t region_size = ParallelCompactData::RegionSize; 2739 2740 size_t src_region_idx = 0; 2741 2742 // Skip empty regions (if any) up to the top of the space. 2743 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2744 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2745 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2746 const RegionData* const top_region_ptr = 2747 sd.addr_to_region_ptr(top_aligned_up); 2748 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2749 ++src_region_ptr; 2750 } 2751 2752 if (src_region_ptr < top_region_ptr) { 2753 // The next source region is in the current space. Update src_region_idx 2754 // and the source address to match src_region_ptr. 2755 src_region_idx = sd.region(src_region_ptr); 2756 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2757 if (src_region_addr > closure.source()) { 2758 closure.set_source(src_region_addr); 2759 } 2760 return src_region_idx; 2761 } 2762 2763 // Switch to a new source space and find the first non-empty region. 2764 unsigned int space_id = src_space_id + 1; 2765 assert(space_id < last_space_id, "not enough spaces"); 2766 2767 HeapWord* const destination = closure.destination(); 2768 2769 do { 2770 MutableSpace* space = _space_info[space_id].space(); 2771 HeapWord* const bottom = space->bottom(); 2772 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2773 2774 // Iterate over the spaces that do not compact into themselves. 2775 if (bottom_cp->destination() != bottom) { 2776 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2777 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2778 2779 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2780 if (src_cp->live_obj_size() > 0) { 2781 // Found it. 2782 assert(src_cp->destination() == destination, 2783 "first live obj in the space must match the destination"); 2784 assert(src_cp->partial_obj_size() == 0, 2785 "a space cannot begin with a partial obj"); 2786 2787 src_space_id = SpaceId(space_id); 2788 src_space_top = space->top(); 2789 const size_t src_region_idx = sd.region(src_cp); 2790 closure.set_source(sd.region_to_addr(src_region_idx)); 2791 return src_region_idx; 2792 } else { 2793 assert(src_cp->data_size() == 0, "sanity"); 2794 } 2795 } 2796 } 2797 } while (++space_id < last_space_id); 2798 2799 assert(false, "no source region was found"); 2800 return 0; 2801 } 2802 2803 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 2804 { 2805 typedef ParMarkBitMap::IterationStatus IterationStatus; 2806 const size_t RegionSize = ParallelCompactData::RegionSize; 2807 ParMarkBitMap* const bitmap = mark_bitmap(); 2808 ParallelCompactData& sd = summary_data(); 2809 RegionData* const region_ptr = sd.region(region_idx); 2810 2811 // Get the items needed to construct the closure. 2812 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2813 SpaceId dest_space_id = space_id(dest_addr); 2814 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 2815 HeapWord* new_top = _space_info[dest_space_id].new_top(); 2816 assert(dest_addr < new_top, "sanity"); 2817 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 2818 2819 // Get the source region and related info. 2820 size_t src_region_idx = region_ptr->source_region(); 2821 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2822 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2823 2824 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2825 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2826 2827 // Adjust src_region_idx to prepare for decrementing destination counts (the 2828 // destination count is not decremented when a region is copied to itself). 2829 if (src_region_idx == region_idx) { 2830 src_region_idx += 1; 2831 } 2832 2833 if (bitmap->is_unmarked(closure.source())) { 2834 // The first source word is in the middle of an object; copy the remainder 2835 // of the object or as much as will fit. The fact that pointer updates were 2836 // deferred will be noted when the object header is processed. 2837 HeapWord* const old_src_addr = closure.source(); 2838 closure.copy_partial_obj(); 2839 if (closure.is_full()) { 2840 decrement_destination_counts(cm, src_space_id, src_region_idx, 2841 closure.source()); 2842 region_ptr->set_deferred_obj_addr(NULL); 2843 region_ptr->set_completed(); 2844 return; 2845 } 2846 2847 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2848 if (sd.region_align_down(old_src_addr) != end_addr) { 2849 // The partial object was copied from more than one source region. 2850 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2851 2852 // Move to the next source region, possibly switching spaces as well. All 2853 // args except end_addr may be modified. 2854 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2855 end_addr); 2856 } 2857 } 2858 2859 do { 2860 HeapWord* const cur_addr = closure.source(); 2861 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2862 src_space_top); 2863 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2864 2865 if (status == ParMarkBitMap::incomplete) { 2866 // The last obj that starts in the source region does not end in the 2867 // region. 2868 assert(closure.source() < end_addr, "sanity"); 2869 HeapWord* const obj_beg = closure.source(); 2870 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 2871 src_space_top); 2872 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 2873 if (obj_end < range_end) { 2874 // The end was found; the entire object will fit. 2875 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 2876 assert(status != ParMarkBitMap::would_overflow, "sanity"); 2877 } else { 2878 // The end was not found; the object will not fit. 2879 assert(range_end < src_space_top, "obj cannot cross space boundary"); 2880 status = ParMarkBitMap::would_overflow; 2881 } 2882 } 2883 2884 if (status == ParMarkBitMap::would_overflow) { 2885 // The last object did not fit. Note that interior oop updates were 2886 // deferred, then copy enough of the object to fill the region. 2887 region_ptr->set_deferred_obj_addr(closure.destination()); 2888 status = closure.copy_until_full(); // copies from closure.source() 2889 2890 decrement_destination_counts(cm, src_space_id, src_region_idx, 2891 closure.source()); 2892 region_ptr->set_completed(); 2893 return; 2894 } 2895 2896 if (status == ParMarkBitMap::full) { 2897 decrement_destination_counts(cm, src_space_id, src_region_idx, 2898 closure.source()); 2899 region_ptr->set_deferred_obj_addr(NULL); 2900 region_ptr->set_completed(); 2901 return; 2902 } 2903 2904 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2905 2906 // Move to the next source region, possibly switching spaces as well. All 2907 // args except end_addr may be modified. 2908 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2909 end_addr); 2910 } while (true); 2911 } 2912 2913 void PSParallelCompact::fill_blocks(size_t region_idx) 2914 { 2915 // Fill in the block table elements for the specified region. Each block 2916 // table element holds the number of live words in the region that are to the 2917 // left of the first object that starts in the block. Thus only blocks in 2918 // which an object starts need to be filled. 2919 // 2920 // The algorithm scans the section of the bitmap that corresponds to the 2921 // region, keeping a running total of the live words. When an object start is 2922 // found, if it's the first to start in the block that contains it, the 2923 // current total is written to the block table element. 2924 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2925 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2926 const size_t RegionSize = ParallelCompactData::RegionSize; 2927 2928 ParallelCompactData& sd = summary_data(); 2929 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2930 if (partial_obj_size >= RegionSize) { 2931 return; // No objects start in this region. 2932 } 2933 2934 // Ensure the first loop iteration decides that the block has changed. 2935 size_t cur_block = sd.block_count(); 2936 2937 const ParMarkBitMap* const bitmap = mark_bitmap(); 2938 2939 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2940 assert((size_t)1 << Log2BitsPerBlock == 2941 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2942 2943 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2944 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2945 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2946 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2947 while (beg_bit < range_end) { 2948 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2949 if (new_block != cur_block) { 2950 cur_block = new_block; 2951 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2952 } 2953 2954 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2955 if (end_bit < range_end - 1) { 2956 live_bits += end_bit - beg_bit + 1; 2957 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2958 } else { 2959 return; 2960 } 2961 } 2962 } 2963 2964 void 2965 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 2966 const MutableSpace* sp = space(space_id); 2967 if (sp->is_empty()) { 2968 return; 2969 } 2970 2971 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2972 ParMarkBitMap* const bitmap = mark_bitmap(); 2973 HeapWord* const dp_addr = dense_prefix(space_id); 2974 HeapWord* beg_addr = sp->bottom(); 2975 HeapWord* end_addr = sp->top(); 2976 2977 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 2978 2979 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 2980 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 2981 if (beg_region < dp_region) { 2982 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 2983 } 2984 2985 // The destination of the first live object that starts in the region is one 2986 // past the end of the partial object entering the region (if any). 2987 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 2988 HeapWord* const new_top = _space_info[space_id].new_top(); 2989 assert(new_top >= dest_addr, "bad new_top value"); 2990 const size_t words = pointer_delta(new_top, dest_addr); 2991 2992 if (words > 0) { 2993 ObjectStartArray* start_array = _space_info[space_id].start_array(); 2994 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2995 2996 ParMarkBitMap::IterationStatus status; 2997 status = bitmap->iterate(&closure, dest_addr, end_addr); 2998 assert(status == ParMarkBitMap::full, "iteration not complete"); 2999 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3000 "live objects skipped because closure is full"); 3001 } 3002 } 3003 3004 jlong PSParallelCompact::millis_since_last_gc() { 3005 // We need a monotonically non-decreasing time in ms but 3006 // os::javaTimeMillis() does not guarantee monotonicity. 3007 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3008 jlong ret_val = now - _time_of_last_gc; 3009 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3010 if (ret_val < 0) { 3011 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3012 return 0; 3013 } 3014 return ret_val; 3015 } 3016 3017 void PSParallelCompact::reset_millis_since_last_gc() { 3018 // We need a monotonically non-decreasing time in ms but 3019 // os::javaTimeMillis() does not guarantee monotonicity. 3020 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3021 } 3022 3023 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3024 { 3025 if (source() != destination()) { 3026 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3027 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3028 } 3029 update_state(words_remaining()); 3030 assert(is_full(), "sanity"); 3031 return ParMarkBitMap::full; 3032 } 3033 3034 void MoveAndUpdateClosure::copy_partial_obj() 3035 { 3036 size_t words = words_remaining(); 3037 3038 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3039 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3040 if (end_addr < range_end) { 3041 words = bitmap()->obj_size(source(), end_addr); 3042 } 3043 3044 // This test is necessary; if omitted, the pointer updates to a partial object 3045 // that crosses the dense prefix boundary could be overwritten. 3046 if (source() != destination()) { 3047 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3048 Copy::aligned_conjoint_words(source(), destination(), words); 3049 } 3050 update_state(words); 3051 } 3052 3053 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3054 PSParallelCompact::AdjustPointerClosure closure(cm); 3055 oop_oop_iterate_oop_maps<true>(obj, &closure); 3056 } 3057 3058 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3059 InstanceKlass::oop_pc_update_pointers(obj, cm); 3060 3061 PSParallelCompact::AdjustPointerClosure closure(cm); 3062 oop_oop_iterate_statics<true>(obj, &closure); 3063 } 3064 3065 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3066 InstanceKlass::oop_pc_update_pointers(obj, cm); 3067 } 3068 3069 #ifdef ASSERT 3070 template <class T> static void trace_reference_gc(const char *s, oop obj, 3071 T* referent_addr, 3072 T* next_addr, 3073 T* discovered_addr) { 3074 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj)); 3075 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3076 p2i(referent_addr), referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3077 log_develop_trace(gc, ref)(" next_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3078 p2i(next_addr), next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3079 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3080 p2i(discovered_addr), discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3081 } 3082 #endif 3083 3084 template <class T> 3085 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) { 3086 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3087 PSParallelCompact::adjust_pointer(referent_addr, cm); 3088 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3089 PSParallelCompact::adjust_pointer(next_addr, cm); 3090 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3091 PSParallelCompact::adjust_pointer(discovered_addr, cm); 3092 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3093 referent_addr, next_addr, discovered_addr);) 3094 } 3095 3096 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3097 InstanceKlass::oop_pc_update_pointers(obj, cm); 3098 3099 if (UseCompressedOops) { 3100 oop_pc_update_pointers_specialized<narrowOop>(obj, cm); 3101 } else { 3102 oop_pc_update_pointers_specialized<oop>(obj, cm); 3103 } 3104 } 3105 3106 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3107 assert(obj->is_objArray(), "obj must be obj array"); 3108 PSParallelCompact::AdjustPointerClosure closure(cm); 3109 oop_oop_iterate_elements<true>(objArrayOop(obj), &closure); 3110 } 3111 3112 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3113 assert(obj->is_typeArray(),"must be a type array"); 3114 } 3115 3116 ParMarkBitMapClosure::IterationStatus 3117 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3118 assert(destination() != NULL, "sanity"); 3119 assert(bitmap()->obj_size(addr) == words, "bad size"); 3120 3121 _source = addr; 3122 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3123 destination(), "wrong destination"); 3124 3125 if (words > words_remaining()) { 3126 return ParMarkBitMap::would_overflow; 3127 } 3128 3129 // The start_array must be updated even if the object is not moving. 3130 if (_start_array != NULL) { 3131 _start_array->allocate_block(destination()); 3132 } 3133 3134 if (destination() != source()) { 3135 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3136 Copy::aligned_conjoint_words(source(), destination(), words); 3137 } 3138 3139 oop moved_oop = (oop) destination(); 3140 compaction_manager()->update_contents(moved_oop); 3141 assert(moved_oop->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3142 3143 update_state(words); 3144 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3145 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3146 } 3147 3148 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3149 ParCompactionManager* cm, 3150 PSParallelCompact::SpaceId space_id) : 3151 ParMarkBitMapClosure(mbm, cm), 3152 _space_id(space_id), 3153 _start_array(PSParallelCompact::start_array(space_id)) 3154 { 3155 } 3156 3157 // Updates the references in the object to their new values. 3158 ParMarkBitMapClosure::IterationStatus 3159 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3160 do_addr(addr); 3161 return ParMarkBitMap::incomplete; 3162 } 3163 3164 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) : 3165 ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm), 3166 _start_array(PSParallelCompact::start_array(space_id)) 3167 { 3168 assert(space_id == PSParallelCompact::old_space_id, 3169 "cannot use FillClosure in the young gen"); 3170 } 3171 3172 ParMarkBitMapClosure::IterationStatus 3173 FillClosure::do_addr(HeapWord* addr, size_t size) { 3174 CollectedHeap::fill_with_objects(addr, size); 3175 HeapWord* const end = addr + size; 3176 do { 3177 _start_array->allocate_block(addr); 3178 addr += oop(addr)->size(); 3179 } while (addr < end); 3180 return ParMarkBitMap::incomplete; 3181 }