1 /* 2 * Copyright (c) 2005, 2014, 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 "classfile/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp" 30 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp" 31 #include "gc_implementation/parallelScavenge/pcTasks.hpp" 32 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp" 33 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp" 34 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp" 35 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp" 36 #include "gc_implementation/parallelScavenge/psOldGen.hpp" 37 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp" 38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 39 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 41 #include "gc_implementation/shared/gcHeapSummary.hpp" 42 #include "gc_implementation/shared/gcTimer.hpp" 43 #include "gc_implementation/shared/gcTrace.hpp" 44 #include "gc_implementation/shared/gcTraceTime.hpp" 45 #include "gc_implementation/shared/isGCActiveMark.hpp" 46 #include "gc_interface/gcCause.hpp" 47 #include "memory/gcLocker.inline.hpp" 48 #include "memory/referencePolicy.hpp" 49 #include "memory/referenceProcessor.hpp" 50 #include "oops/methodData.hpp" 51 #include "oops/oop.inline.hpp" 52 #include "oops/oop.pcgc.inline.hpp" 53 #include "runtime/fprofiler.hpp" 54 #include "runtime/safepoint.hpp" 55 #include "runtime/vmThread.hpp" 56 #include "services/management.hpp" 57 #include "services/memoryService.hpp" 58 #include "services/memTracker.hpp" 59 #include "utilities/events.hpp" 60 #include "utilities/stack.inline.hpp" 61 62 #include <math.h> 63 64 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 65 66 // All sizes are in HeapWords. 67 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 68 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 69 const size_t ParallelCompactData::RegionSizeBytes = 70 RegionSize << LogHeapWordSize; 71 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 72 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 73 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 74 75 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 76 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 77 const size_t ParallelCompactData::BlockSizeBytes = 78 BlockSize << LogHeapWordSize; 79 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 80 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 81 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 82 83 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 84 const size_t ParallelCompactData::Log2BlocksPerRegion = 85 Log2RegionSize - Log2BlockSize; 86 87 const ParallelCompactData::RegionData::region_sz_t 88 ParallelCompactData::RegionData::dc_shift = 27; 89 90 const ParallelCompactData::RegionData::region_sz_t 91 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 92 93 const ParallelCompactData::RegionData::region_sz_t 94 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 95 96 const ParallelCompactData::RegionData::region_sz_t 97 ParallelCompactData::RegionData::los_mask = ~dc_mask; 98 99 const ParallelCompactData::RegionData::region_sz_t 100 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 101 102 const ParallelCompactData::RegionData::region_sz_t 103 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 104 105 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 106 bool PSParallelCompact::_print_phases = false; 107 108 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 109 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL; 110 111 double PSParallelCompact::_dwl_mean; 112 double PSParallelCompact::_dwl_std_dev; 113 double PSParallelCompact::_dwl_first_term; 114 double PSParallelCompact::_dwl_adjustment; 115 #ifdef ASSERT 116 bool PSParallelCompact::_dwl_initialized = false; 117 #endif // #ifdef ASSERT 118 119 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 120 HeapWord* destination) 121 { 122 assert(src_region_idx != 0, "invalid src_region_idx"); 123 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 124 assert(destination != NULL, "invalid destination argument"); 125 126 _src_region_idx = src_region_idx; 127 _partial_obj_size = partial_obj_size; 128 _destination = destination; 129 130 // These fields may not be updated below, so make sure they're clear. 131 assert(_dest_region_addr == NULL, "should have been cleared"); 132 assert(_first_src_addr == NULL, "should have been cleared"); 133 134 // Determine the number of destination regions for the partial object. 135 HeapWord* const last_word = destination + partial_obj_size - 1; 136 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 137 HeapWord* const beg_region_addr = sd.region_align_down(destination); 138 HeapWord* const end_region_addr = sd.region_align_down(last_word); 139 140 if (beg_region_addr == end_region_addr) { 141 // One destination region. 142 _destination_count = 1; 143 if (end_region_addr == destination) { 144 // The destination falls on a region boundary, thus the first word of the 145 // partial object will be the first word copied to the destination region. 146 _dest_region_addr = end_region_addr; 147 _first_src_addr = sd.region_to_addr(src_region_idx); 148 } 149 } else { 150 // Two destination regions. When copied, the partial object will cross a 151 // destination region boundary, so a word somewhere within the partial 152 // object will be the first word copied to the second destination region. 153 _destination_count = 2; 154 _dest_region_addr = end_region_addr; 155 const size_t ofs = pointer_delta(end_region_addr, destination); 156 assert(ofs < _partial_obj_size, "sanity"); 157 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 158 } 159 } 160 161 void SplitInfo::clear() 162 { 163 _src_region_idx = 0; 164 _partial_obj_size = 0; 165 _destination = NULL; 166 _destination_count = 0; 167 _dest_region_addr = NULL; 168 _first_src_addr = NULL; 169 assert(!is_valid(), "sanity"); 170 } 171 172 #ifdef ASSERT 173 void SplitInfo::verify_clear() 174 { 175 assert(_src_region_idx == 0, "not clear"); 176 assert(_partial_obj_size == 0, "not clear"); 177 assert(_destination == NULL, "not clear"); 178 assert(_destination_count == 0, "not clear"); 179 assert(_dest_region_addr == NULL, "not clear"); 180 assert(_first_src_addr == NULL, "not clear"); 181 } 182 #endif // #ifdef ASSERT 183 184 185 void PSParallelCompact::print_on_error(outputStream* st) { 186 _mark_bitmap.print_on_error(st); 187 } 188 189 #ifndef PRODUCT 190 const char* PSParallelCompact::space_names[] = { 191 "old ", "eden", "from", "to " 192 }; 193 194 void PSParallelCompact::print_region_ranges() 195 { 196 tty->print_cr("space bottom top end new_top"); 197 tty->print_cr("------ ---------- ---------- ---------- ----------"); 198 199 for (unsigned int id = 0; id < last_space_id; ++id) { 200 const MutableSpace* space = _space_info[id].space(); 201 tty->print_cr("%u %s " 202 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 203 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 204 id, space_names[id], 205 summary_data().addr_to_region_idx(space->bottom()), 206 summary_data().addr_to_region_idx(space->top()), 207 summary_data().addr_to_region_idx(space->end()), 208 summary_data().addr_to_region_idx(_space_info[id].new_top())); 209 } 210 } 211 212 void 213 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 214 { 215 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 216 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 217 218 ParallelCompactData& sd = PSParallelCompact::summary_data(); 219 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 220 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 221 REGION_IDX_FORMAT " " PTR_FORMAT " " 222 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 223 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 224 i, c->data_location(), dci, c->destination(), 225 c->partial_obj_size(), c->live_obj_size(), 226 c->data_size(), c->source_region(), c->destination_count()); 227 228 #undef REGION_IDX_FORMAT 229 #undef REGION_DATA_FORMAT 230 } 231 232 void 233 print_generic_summary_data(ParallelCompactData& summary_data, 234 HeapWord* const beg_addr, 235 HeapWord* const end_addr) 236 { 237 size_t total_words = 0; 238 size_t i = summary_data.addr_to_region_idx(beg_addr); 239 const size_t last = summary_data.addr_to_region_idx(end_addr); 240 HeapWord* pdest = 0; 241 242 while (i <= last) { 243 ParallelCompactData::RegionData* c = summary_data.region(i); 244 if (c->data_size() != 0 || c->destination() != pdest) { 245 print_generic_summary_region(i, c); 246 total_words += c->data_size(); 247 pdest = c->destination(); 248 } 249 ++i; 250 } 251 252 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 253 } 254 255 void 256 print_generic_summary_data(ParallelCompactData& summary_data, 257 SpaceInfo* space_info) 258 { 259 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 260 const MutableSpace* space = space_info[id].space(); 261 print_generic_summary_data(summary_data, space->bottom(), 262 MAX2(space->top(), space_info[id].new_top())); 263 } 264 } 265 266 void 267 print_initial_summary_region(size_t i, 268 const ParallelCompactData::RegionData* c, 269 bool newline = true) 270 { 271 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 272 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 273 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 274 i, c->destination(), 275 c->partial_obj_size(), c->live_obj_size(), 276 c->data_size(), c->source_region(), c->destination_count()); 277 if (newline) tty->cr(); 278 } 279 280 void 281 print_initial_summary_data(ParallelCompactData& summary_data, 282 const MutableSpace* space) { 283 if (space->top() == space->bottom()) { 284 return; 285 } 286 287 const size_t region_size = ParallelCompactData::RegionSize; 288 typedef ParallelCompactData::RegionData RegionData; 289 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 290 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 291 const RegionData* c = summary_data.region(end_region - 1); 292 HeapWord* end_addr = c->destination() + c->data_size(); 293 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 294 295 // Print (and count) the full regions at the beginning of the space. 296 size_t full_region_count = 0; 297 size_t i = summary_data.addr_to_region_idx(space->bottom()); 298 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 299 print_initial_summary_region(i, summary_data.region(i)); 300 ++full_region_count; 301 ++i; 302 } 303 304 size_t live_to_right = live_in_space - full_region_count * region_size; 305 306 double max_reclaimed_ratio = 0.0; 307 size_t max_reclaimed_ratio_region = 0; 308 size_t max_dead_to_right = 0; 309 size_t max_live_to_right = 0; 310 311 // Print the 'reclaimed ratio' for regions while there is something live in 312 // the region or to the right of it. The remaining regions are empty (and 313 // uninteresting), and computing the ratio will result in division by 0. 314 while (i < end_region && live_to_right > 0) { 315 c = summary_data.region(i); 316 HeapWord* const region_addr = summary_data.region_to_addr(i); 317 const size_t used_to_right = pointer_delta(space->top(), region_addr); 318 const size_t dead_to_right = used_to_right - live_to_right; 319 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 320 321 if (reclaimed_ratio > max_reclaimed_ratio) { 322 max_reclaimed_ratio = reclaimed_ratio; 323 max_reclaimed_ratio_region = i; 324 max_dead_to_right = dead_to_right; 325 max_live_to_right = live_to_right; 326 } 327 328 print_initial_summary_region(i, c, false); 329 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 330 reclaimed_ratio, dead_to_right, live_to_right); 331 332 live_to_right -= c->data_size(); 333 ++i; 334 } 335 336 // Any remaining regions are empty. Print one more if there is one. 337 if (i < end_region) { 338 print_initial_summary_region(i, summary_data.region(i)); 339 } 340 341 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 342 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 343 max_reclaimed_ratio_region, max_dead_to_right, 344 max_live_to_right, max_reclaimed_ratio); 345 } 346 347 void 348 print_initial_summary_data(ParallelCompactData& summary_data, 349 SpaceInfo* space_info) { 350 unsigned int id = PSParallelCompact::old_space_id; 351 const MutableSpace* space; 352 do { 353 space = space_info[id].space(); 354 print_initial_summary_data(summary_data, space); 355 } while (++id < PSParallelCompact::eden_space_id); 356 357 do { 358 space = space_info[id].space(); 359 print_generic_summary_data(summary_data, space->bottom(), space->top()); 360 } while (++id < PSParallelCompact::last_space_id); 361 } 362 #endif // #ifndef PRODUCT 363 364 #ifdef ASSERT 365 size_t add_obj_count; 366 size_t add_obj_size; 367 size_t mark_bitmap_count; 368 size_t mark_bitmap_size; 369 #endif // #ifdef ASSERT 370 371 ParallelCompactData::ParallelCompactData() 372 { 373 _region_start = 0; 374 375 _region_vspace = 0; 376 _reserved_byte_size = 0; 377 _region_data = 0; 378 _region_count = 0; 379 380 _block_vspace = 0; 381 _block_data = 0; 382 _block_count = 0; 383 } 384 385 bool ParallelCompactData::initialize(MemRegion covered_region) 386 { 387 _region_start = covered_region.start(); 388 const size_t region_size = covered_region.word_size(); 389 DEBUG_ONLY(_region_end = _region_start + region_size;) 390 391 assert(region_align_down(_region_start) == _region_start, 392 "region start not aligned"); 393 assert((region_size & RegionSizeOffsetMask) == 0, 394 "region size not a multiple of RegionSize"); 395 396 bool result = initialize_region_data(region_size) && initialize_block_data(); 397 return result; 398 } 399 400 PSVirtualSpace* 401 ParallelCompactData::create_vspace(size_t count, size_t element_size) 402 { 403 const size_t raw_bytes = count * element_size; 404 const size_t page_sz = os::page_size_for_region(raw_bytes, 10); 405 const size_t granularity = os::vm_allocation_granularity(); 406 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 407 408 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 409 MAX2(page_sz, granularity); 410 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 411 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 412 rs.size()); 413 414 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 415 416 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 417 if (vspace != 0) { 418 if (vspace->expand_by(_reserved_byte_size)) { 419 return vspace; 420 } 421 delete vspace; 422 // Release memory reserved in the space. 423 rs.release(); 424 } 425 426 return 0; 427 } 428 429 bool ParallelCompactData::initialize_region_data(size_t region_size) 430 { 431 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 432 _region_vspace = create_vspace(count, sizeof(RegionData)); 433 if (_region_vspace != 0) { 434 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 435 _region_count = count; 436 return true; 437 } 438 return false; 439 } 440 441 bool ParallelCompactData::initialize_block_data() 442 { 443 assert(_region_count != 0, "region data must be initialized first"); 444 const size_t count = _region_count << Log2BlocksPerRegion; 445 _block_vspace = create_vspace(count, sizeof(BlockData)); 446 if (_block_vspace != 0) { 447 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 448 _block_count = count; 449 return true; 450 } 451 return false; 452 } 453 454 void ParallelCompactData::clear() 455 { 456 memset(_region_data, 0, _region_vspace->committed_size()); 457 memset(_block_data, 0, _block_vspace->committed_size()); 458 } 459 460 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 461 assert(beg_region <= _region_count, "beg_region out of range"); 462 assert(end_region <= _region_count, "end_region out of range"); 463 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 464 465 const size_t region_cnt = end_region - beg_region; 466 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 467 468 const size_t beg_block = beg_region * BlocksPerRegion; 469 const size_t block_cnt = region_cnt * BlocksPerRegion; 470 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 471 } 472 473 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 474 { 475 const RegionData* cur_cp = region(region_idx); 476 const RegionData* const end_cp = region(region_count() - 1); 477 478 HeapWord* result = region_to_addr(region_idx); 479 if (cur_cp < end_cp) { 480 do { 481 result += cur_cp->partial_obj_size(); 482 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 483 } 484 return result; 485 } 486 487 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 488 { 489 const size_t obj_ofs = pointer_delta(addr, _region_start); 490 const size_t beg_region = obj_ofs >> Log2RegionSize; 491 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 492 493 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 494 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 495 496 if (beg_region == end_region) { 497 // All in one region. 498 _region_data[beg_region].add_live_obj(len); 499 return; 500 } 501 502 // First region. 503 const size_t beg_ofs = region_offset(addr); 504 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 505 506 Klass* klass = ((oop)addr)->klass(); 507 // Middle regions--completely spanned by this object. 508 for (size_t region = beg_region + 1; region < end_region; ++region) { 509 _region_data[region].set_partial_obj_size(RegionSize); 510 _region_data[region].set_partial_obj_addr(addr); 511 } 512 513 // Last region. 514 const size_t end_ofs = region_offset(addr + len - 1); 515 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 516 _region_data[end_region].set_partial_obj_addr(addr); 517 } 518 519 void 520 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 521 { 522 assert(region_offset(beg) == 0, "not RegionSize aligned"); 523 assert(region_offset(end) == 0, "not RegionSize aligned"); 524 525 size_t cur_region = addr_to_region_idx(beg); 526 const size_t end_region = addr_to_region_idx(end); 527 HeapWord* addr = beg; 528 while (cur_region < end_region) { 529 _region_data[cur_region].set_destination(addr); 530 _region_data[cur_region].set_destination_count(0); 531 _region_data[cur_region].set_source_region(cur_region); 532 _region_data[cur_region].set_data_location(addr); 533 534 // Update live_obj_size so the region appears completely full. 535 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 536 _region_data[cur_region].set_live_obj_size(live_size); 537 538 ++cur_region; 539 addr += RegionSize; 540 } 541 } 542 543 // Find the point at which a space can be split and, if necessary, record the 544 // split point. 545 // 546 // If the current src region (which overflowed the destination space) doesn't 547 // have a partial object, the split point is at the beginning of the current src 548 // region (an "easy" split, no extra bookkeeping required). 549 // 550 // If the current src region has a partial object, the split point is in the 551 // region where that partial object starts (call it the split_region). If 552 // split_region has a partial object, then the split point is just after that 553 // partial object (a "hard" split where we have to record the split data and 554 // zero the partial_obj_size field). With a "hard" split, we know that the 555 // partial_obj ends within split_region because the partial object that caused 556 // the overflow starts in split_region. If split_region doesn't have a partial 557 // obj, then the split is at the beginning of split_region (another "easy" 558 // split). 559 HeapWord* 560 ParallelCompactData::summarize_split_space(size_t src_region, 561 SplitInfo& split_info, 562 HeapWord* destination, 563 HeapWord* target_end, 564 HeapWord** target_next) 565 { 566 assert(destination <= target_end, "sanity"); 567 assert(destination + _region_data[src_region].data_size() > target_end, 568 "region should not fit into target space"); 569 assert(is_region_aligned(target_end), "sanity"); 570 571 size_t split_region = src_region; 572 HeapWord* split_destination = destination; 573 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 574 575 if (destination + partial_obj_size > target_end) { 576 // The split point is just after the partial object (if any) in the 577 // src_region that contains the start of the object that overflowed the 578 // destination space. 579 // 580 // Find the start of the "overflow" object and set split_region to the 581 // region containing it. 582 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 583 split_region = addr_to_region_idx(overflow_obj); 584 585 // Clear the source_region field of all destination regions whose first word 586 // came from data after the split point (a non-null source_region field 587 // implies a region must be filled). 588 // 589 // An alternative to the simple loop below: clear during post_compact(), 590 // which uses memcpy instead of individual stores, and is easy to 591 // parallelize. (The downside is that it clears the entire RegionData 592 // object as opposed to just one field.) 593 // 594 // post_compact() would have to clear the summary data up to the highest 595 // address that was written during the summary phase, which would be 596 // 597 // max(top, max(new_top, clear_top)) 598 // 599 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 600 // to target_end. 601 const RegionData* const sr = region(split_region); 602 const size_t beg_idx = 603 addr_to_region_idx(region_align_up(sr->destination() + 604 sr->partial_obj_size())); 605 const size_t end_idx = addr_to_region_idx(target_end); 606 607 if (TraceParallelOldGCSummaryPhase) { 608 gclog_or_tty->print_cr("split: clearing source_region field in [" 609 SIZE_FORMAT ", " SIZE_FORMAT ")", 610 beg_idx, end_idx); 611 } 612 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 613 _region_data[idx].set_source_region(0); 614 } 615 616 // Set split_destination and partial_obj_size to reflect the split region. 617 split_destination = sr->destination(); 618 partial_obj_size = sr->partial_obj_size(); 619 } 620 621 // The split is recorded only if a partial object extends onto the region. 622 if (partial_obj_size != 0) { 623 _region_data[split_region].set_partial_obj_size(0); 624 split_info.record(split_region, partial_obj_size, split_destination); 625 } 626 627 // Setup the continuation addresses. 628 *target_next = split_destination + partial_obj_size; 629 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 630 631 if (TraceParallelOldGCSummaryPhase) { 632 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 633 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 634 " pos=" SIZE_FORMAT, 635 split_type, source_next, split_region, 636 partial_obj_size); 637 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 638 " tn=" PTR_FORMAT, 639 split_type, split_destination, 640 addr_to_region_idx(split_destination), 641 *target_next); 642 643 if (partial_obj_size != 0) { 644 HeapWord* const po_beg = split_info.destination(); 645 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 646 gclog_or_tty->print_cr("%s split: " 647 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 648 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 649 split_type, 650 po_beg, addr_to_region_idx(po_beg), 651 po_end, addr_to_region_idx(po_end)); 652 } 653 } 654 655 return source_next; 656 } 657 658 bool ParallelCompactData::summarize(SplitInfo& split_info, 659 HeapWord* source_beg, HeapWord* source_end, 660 HeapWord** source_next, 661 HeapWord* target_beg, HeapWord* target_end, 662 HeapWord** target_next) 663 { 664 if (TraceParallelOldGCSummaryPhase) { 665 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 666 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 667 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 668 source_beg, source_end, source_next_val, 669 target_beg, target_end, *target_next); 670 } 671 672 size_t cur_region = addr_to_region_idx(source_beg); 673 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 674 675 HeapWord *dest_addr = target_beg; 676 while (cur_region < end_region) { 677 // The destination must be set even if the region has no data. 678 _region_data[cur_region].set_destination(dest_addr); 679 680 size_t words = _region_data[cur_region].data_size(); 681 if (words > 0) { 682 // If cur_region does not fit entirely into the target space, find a point 683 // at which the source space can be 'split' so that part is copied to the 684 // target space and the rest is copied elsewhere. 685 if (dest_addr + words > target_end) { 686 assert(source_next != NULL, "source_next is NULL when splitting"); 687 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 688 target_end, target_next); 689 return false; 690 } 691 692 // Compute the destination_count for cur_region, and if necessary, update 693 // source_region for a destination region. The source_region field is 694 // updated if cur_region is the first (left-most) region to be copied to a 695 // destination region. 696 // 697 // The destination_count calculation is a bit subtle. A region that has 698 // data that compacts into itself does not count itself as a destination. 699 // This maintains the invariant that a zero count means the region is 700 // available and can be claimed and then filled. 701 uint destination_count = 0; 702 if (split_info.is_split(cur_region)) { 703 // The current region has been split: the partial object will be copied 704 // to one destination space and the remaining data will be copied to 705 // another destination space. Adjust the initial destination_count and, 706 // if necessary, set the source_region field if the partial object will 707 // cross a destination region boundary. 708 destination_count = split_info.destination_count(); 709 if (destination_count == 2) { 710 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 711 _region_data[dest_idx].set_source_region(cur_region); 712 } 713 } 714 715 HeapWord* const last_addr = dest_addr + words - 1; 716 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 717 const size_t dest_region_2 = addr_to_region_idx(last_addr); 718 719 // Initially assume that the destination regions will be the same and 720 // adjust the value below if necessary. Under this assumption, if 721 // cur_region == dest_region_2, then cur_region will be compacted 722 // completely into itself. 723 destination_count += cur_region == dest_region_2 ? 0 : 1; 724 if (dest_region_1 != dest_region_2) { 725 // Destination regions differ; adjust destination_count. 726 destination_count += 1; 727 // Data from cur_region will be copied to the start of dest_region_2. 728 _region_data[dest_region_2].set_source_region(cur_region); 729 } else if (region_offset(dest_addr) == 0) { 730 // Data from cur_region will be copied to the start of the destination 731 // region. 732 _region_data[dest_region_1].set_source_region(cur_region); 733 } 734 735 _region_data[cur_region].set_destination_count(destination_count); 736 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 737 dest_addr += words; 738 } 739 740 ++cur_region; 741 } 742 743 *target_next = dest_addr; 744 return true; 745 } 746 747 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 748 assert(addr != NULL, "Should detect NULL oop earlier"); 749 assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap"); 750 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 751 752 // Region covering the object. 753 RegionData* const region_ptr = addr_to_region_ptr(addr); 754 HeapWord* result = region_ptr->destination(); 755 756 // If the entire Region is live, the new location is region->destination + the 757 // offset of the object within in the Region. 758 759 // Run some performance tests to determine if this special case pays off. It 760 // is worth it for pointers into the dense prefix. If the optimization to 761 // avoid pointer updates in regions that only point to the dense prefix is 762 // ever implemented, this should be revisited. 763 if (region_ptr->data_size() == RegionSize) { 764 result += region_offset(addr); 765 return result; 766 } 767 768 // Otherwise, the new location is region->destination + block offset + the 769 // number of live words in the Block that are (a) to the left of addr and (b) 770 // due to objects that start in the Block. 771 772 // Fill in the block table if necessary. This is unsynchronized, so multiple 773 // threads may fill the block table for a region (harmless, since it is 774 // idempotent). 775 if (!region_ptr->blocks_filled()) { 776 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 777 region_ptr->set_blocks_filled(); 778 } 779 780 HeapWord* const search_start = block_align_down(addr); 781 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 782 783 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 784 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 785 result += block_offset + live; 786 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 787 return result; 788 } 789 790 #ifdef ASSERT 791 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 792 { 793 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 794 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 795 for (const size_t* p = beg; p < end; ++p) { 796 assert(*p == 0, "not zero"); 797 } 798 } 799 800 void ParallelCompactData::verify_clear() 801 { 802 verify_clear(_region_vspace); 803 verify_clear(_block_vspace); 804 } 805 #endif // #ifdef ASSERT 806 807 STWGCTimer PSParallelCompact::_gc_timer; 808 ParallelOldTracer PSParallelCompact::_gc_tracer; 809 elapsedTimer PSParallelCompact::_accumulated_time; 810 unsigned int PSParallelCompact::_total_invocations = 0; 811 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 812 jlong PSParallelCompact::_time_of_last_gc = 0; 813 CollectorCounters* PSParallelCompact::_counters = NULL; 814 ParMarkBitMap PSParallelCompact::_mark_bitmap; 815 ParallelCompactData PSParallelCompact::_summary_data; 816 817 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 818 819 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 820 821 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 822 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 823 824 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 825 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 826 827 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); } 828 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); } 829 830 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 831 832 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { 833 mark_and_push(_compaction_manager, p); 834 } 835 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 836 837 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 838 klass->oops_do(_mark_and_push_closure); 839 } 840 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 841 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 842 } 843 844 void PSParallelCompact::post_initialize() { 845 ParallelScavengeHeap* heap = gc_heap(); 846 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 847 848 MemRegion mr = heap->reserved_region(); 849 _ref_processor = 850 new ReferenceProcessor(mr, // span 851 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 852 (int) ParallelGCThreads, // mt processing degree 853 true, // mt discovery 854 (int) ParallelGCThreads, // mt discovery degree 855 true, // atomic_discovery 856 &_is_alive_closure); // non-header is alive closure 857 _counters = new CollectorCounters("PSParallelCompact", 1); 858 859 // Initialize static fields in ParCompactionManager. 860 ParCompactionManager::initialize(mark_bitmap()); 861 } 862 863 bool PSParallelCompact::initialize() { 864 ParallelScavengeHeap* heap = gc_heap(); 865 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 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 = gc_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 // Simple class for storing info about the heap at the start of GC, to be used 921 // after GC for comparison/printing. 922 class PreGCValues { 923 public: 924 PreGCValues() { } 925 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 926 927 void fill(ParallelScavengeHeap* heap) { 928 _heap_used = heap->used(); 929 _young_gen_used = heap->young_gen()->used_in_bytes(); 930 _old_gen_used = heap->old_gen()->used_in_bytes(); 931 _metadata_used = MetaspaceAux::used_bytes(); 932 }; 933 934 size_t heap_used() const { return _heap_used; } 935 size_t young_gen_used() const { return _young_gen_used; } 936 size_t old_gen_used() const { return _old_gen_used; } 937 size_t metadata_used() const { return _metadata_used; } 938 939 private: 940 size_t _heap_used; 941 size_t _young_gen_used; 942 size_t _old_gen_used; 943 size_t _metadata_used; 944 }; 945 946 void 947 PSParallelCompact::clear_data_covering_space(SpaceId id) 948 { 949 // At this point, top is the value before GC, new_top() is the value that will 950 // be set at the end of GC. The marking bitmap is cleared to top; nothing 951 // should be marked above top. The summary data is cleared to the larger of 952 // top & new_top. 953 MutableSpace* const space = _space_info[id].space(); 954 HeapWord* const bot = space->bottom(); 955 HeapWord* const top = space->top(); 956 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 957 958 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 959 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 960 _mark_bitmap.clear_range(beg_bit, end_bit); 961 962 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 963 const size_t end_region = 964 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 965 _summary_data.clear_range(beg_region, end_region); 966 967 // Clear the data used to 'split' regions. 968 SplitInfo& split_info = _space_info[id].split_info(); 969 if (split_info.is_valid()) { 970 split_info.clear(); 971 } 972 DEBUG_ONLY(split_info.verify_clear();) 973 } 974 975 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 976 { 977 // Update the from & to space pointers in space_info, since they are swapped 978 // at each young gen gc. Do the update unconditionally (even though a 979 // promotion failure does not swap spaces) because an unknown number of minor 980 // collections will have swapped the spaces an unknown number of times. 981 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 982 ParallelScavengeHeap* heap = gc_heap(); 983 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 984 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 985 986 pre_gc_values->fill(heap); 987 988 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 989 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 990 991 // Increment the invocation count 992 heap->increment_total_collections(true); 993 994 // We need to track unique mark sweep invocations as well. 995 _total_invocations++; 996 997 heap->print_heap_before_gc(); 998 heap->trace_heap_before_gc(&_gc_tracer); 999 1000 // Fill in TLABs 1001 heap->accumulate_statistics_all_tlabs(); 1002 heap->ensure_parsability(true); // retire TLABs 1003 1004 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1005 HandleMark hm; // Discard invalid handles created during verification 1006 Universe::verify(" VerifyBeforeGC:"); 1007 } 1008 1009 // Verify object start arrays 1010 if (VerifyObjectStartArray && 1011 VerifyBeforeGC) { 1012 heap->old_gen()->verify_object_start_array(); 1013 } 1014 1015 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1016 DEBUG_ONLY(summary_data().verify_clear();) 1017 1018 // Have worker threads release resources the next time they run a task. 1019 gc_task_manager()->release_all_resources(); 1020 } 1021 1022 void PSParallelCompact::post_compact() 1023 { 1024 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1025 1026 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1027 // Clear the marking bitmap, summary data and split info. 1028 clear_data_covering_space(SpaceId(id)); 1029 // Update top(). Must be done after clearing the bitmap and summary data. 1030 _space_info[id].publish_new_top(); 1031 } 1032 1033 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1034 MutableSpace* const from_space = _space_info[from_space_id].space(); 1035 MutableSpace* const to_space = _space_info[to_space_id].space(); 1036 1037 ParallelScavengeHeap* heap = gc_heap(); 1038 bool eden_empty = eden_space->is_empty(); 1039 if (!eden_empty) { 1040 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1041 heap->young_gen(), heap->old_gen()); 1042 } 1043 1044 // Update heap occupancy information which is used as input to the soft ref 1045 // clearing policy at the next gc. 1046 Universe::update_heap_info_at_gc(); 1047 1048 bool young_gen_empty = eden_empty && from_space->is_empty() && 1049 to_space->is_empty(); 1050 1051 BarrierSet* bs = heap->barrier_set(); 1052 if (bs->is_a(BarrierSet::ModRef)) { 1053 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1054 MemRegion old_mr = heap->old_gen()->reserved(); 1055 1056 if (young_gen_empty) { 1057 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1058 } else { 1059 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1060 } 1061 } 1062 1063 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1064 ClassLoaderDataGraph::purge(); 1065 MetaspaceAux::verify_metrics(); 1066 1067 Threads::gc_epilogue(); 1068 CodeCache::gc_epilogue(); 1069 JvmtiExport::gc_epilogue(); 1070 1071 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1072 1073 ref_processor()->enqueue_discovered_references(NULL); 1074 1075 if (ZapUnusedHeapArea) { 1076 heap->gen_mangle_unused_area(); 1077 } 1078 1079 // Update time of last GC 1080 reset_millis_since_last_gc(); 1081 } 1082 1083 HeapWord* 1084 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1085 bool maximum_compaction) 1086 { 1087 const size_t region_size = ParallelCompactData::RegionSize; 1088 const ParallelCompactData& sd = summary_data(); 1089 1090 const MutableSpace* const space = _space_info[id].space(); 1091 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1092 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1093 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1094 1095 // Skip full regions at the beginning of the space--they are necessarily part 1096 // of the dense prefix. 1097 size_t full_count = 0; 1098 const RegionData* cp; 1099 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1100 ++full_count; 1101 } 1102 1103 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1104 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1105 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1106 if (maximum_compaction || cp == end_cp || interval_ended) { 1107 _maximum_compaction_gc_num = total_invocations(); 1108 return sd.region_to_addr(cp); 1109 } 1110 1111 HeapWord* const new_top = _space_info[id].new_top(); 1112 const size_t space_live = pointer_delta(new_top, space->bottom()); 1113 const size_t space_used = space->used_in_words(); 1114 const size_t space_capacity = space->capacity_in_words(); 1115 1116 const double cur_density = double(space_live) / space_capacity; 1117 const double deadwood_density = 1118 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1119 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1120 1121 if (TraceParallelOldGCDensePrefix) { 1122 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1123 cur_density, deadwood_density, deadwood_goal); 1124 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1125 "space_cap=" SIZE_FORMAT, 1126 space_live, space_used, 1127 space_capacity); 1128 } 1129 1130 // XXX - Use binary search? 1131 HeapWord* dense_prefix = sd.region_to_addr(cp); 1132 const RegionData* full_cp = cp; 1133 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1134 while (cp < end_cp) { 1135 HeapWord* region_destination = cp->destination(); 1136 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1137 if (TraceParallelOldGCDensePrefix && Verbose) { 1138 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1139 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1140 sd.region(cp), region_destination, 1141 dense_prefix, cur_deadwood); 1142 } 1143 1144 if (cur_deadwood >= deadwood_goal) { 1145 // Found the region that has the correct amount of deadwood to the left. 1146 // This typically occurs after crossing a fairly sparse set of regions, so 1147 // iterate backwards over those sparse regions, looking for the region 1148 // that has the lowest density of live objects 'to the right.' 1149 size_t space_to_left = sd.region(cp) * region_size; 1150 size_t live_to_left = space_to_left - cur_deadwood; 1151 size_t space_to_right = space_capacity - space_to_left; 1152 size_t live_to_right = space_live - live_to_left; 1153 double density_to_right = double(live_to_right) / space_to_right; 1154 while (cp > full_cp) { 1155 --cp; 1156 const size_t prev_region_live_to_right = live_to_right - 1157 cp->data_size(); 1158 const size_t prev_region_space_to_right = space_to_right + region_size; 1159 double prev_region_density_to_right = 1160 double(prev_region_live_to_right) / prev_region_space_to_right; 1161 if (density_to_right <= prev_region_density_to_right) { 1162 return dense_prefix; 1163 } 1164 if (TraceParallelOldGCDensePrefix && Verbose) { 1165 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1166 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1167 prev_region_density_to_right); 1168 } 1169 dense_prefix -= region_size; 1170 live_to_right = prev_region_live_to_right; 1171 space_to_right = prev_region_space_to_right; 1172 density_to_right = prev_region_density_to_right; 1173 } 1174 return dense_prefix; 1175 } 1176 1177 dense_prefix += region_size; 1178 ++cp; 1179 } 1180 1181 return dense_prefix; 1182 } 1183 1184 #ifndef PRODUCT 1185 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1186 const SpaceId id, 1187 const bool maximum_compaction, 1188 HeapWord* const addr) 1189 { 1190 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1191 RegionData* const cp = summary_data().region(region_idx); 1192 const MutableSpace* const space = _space_info[id].space(); 1193 HeapWord* const new_top = _space_info[id].new_top(); 1194 1195 const size_t space_live = pointer_delta(new_top, space->bottom()); 1196 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1197 const size_t space_cap = space->capacity_in_words(); 1198 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1199 const size_t live_to_right = new_top - cp->destination(); 1200 const size_t dead_to_right = space->top() - addr - live_to_right; 1201 1202 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1203 "spl=" SIZE_FORMAT " " 1204 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1205 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1206 " ratio=%10.8f", 1207 algorithm, addr, region_idx, 1208 space_live, 1209 dead_to_left, dead_to_left_pct, 1210 dead_to_right, live_to_right, 1211 double(dead_to_right) / live_to_right); 1212 } 1213 #endif // #ifndef PRODUCT 1214 1215 // Return a fraction indicating how much of the generation can be treated as 1216 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1217 // based on the density of live objects in the generation to determine a limit, 1218 // which is then adjusted so the return value is min_percent when the density is 1219 // 1. 1220 // 1221 // The following table shows some return values for a different values of the 1222 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1223 // min_percent is 1. 1224 // 1225 // fraction allowed as dead wood 1226 // ----------------------------------------------------------------- 1227 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1228 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1229 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1230 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1231 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1232 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1233 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1234 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1235 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1236 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1237 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1238 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1239 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1240 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1241 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1242 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1243 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1244 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1245 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1246 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1247 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1248 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1249 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1250 1251 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1252 { 1253 assert(_dwl_initialized, "uninitialized"); 1254 1255 // The raw limit is the value of the normal distribution at x = density. 1256 const double raw_limit = normal_distribution(density); 1257 1258 // Adjust the raw limit so it becomes the minimum when the density is 1. 1259 // 1260 // First subtract the adjustment value (which is simply the precomputed value 1261 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1262 // Then add the minimum value, so the minimum is returned when the density is 1263 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1264 const double min = double(min_percent) / 100.0; 1265 const double limit = raw_limit - _dwl_adjustment + min; 1266 return MAX2(limit, 0.0); 1267 } 1268 1269 ParallelCompactData::RegionData* 1270 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1271 const RegionData* end) 1272 { 1273 const size_t region_size = ParallelCompactData::RegionSize; 1274 ParallelCompactData& sd = summary_data(); 1275 size_t left = sd.region(beg); 1276 size_t right = end > beg ? sd.region(end) - 1 : left; 1277 1278 // Binary search. 1279 while (left < right) { 1280 // Equivalent to (left + right) / 2, but does not overflow. 1281 const size_t middle = left + (right - left) / 2; 1282 RegionData* const middle_ptr = sd.region(middle); 1283 HeapWord* const dest = middle_ptr->destination(); 1284 HeapWord* const addr = sd.region_to_addr(middle); 1285 assert(dest != NULL, "sanity"); 1286 assert(dest <= addr, "must move left"); 1287 1288 if (middle > left && dest < addr) { 1289 right = middle - 1; 1290 } else if (middle < right && middle_ptr->data_size() == region_size) { 1291 left = middle + 1; 1292 } else { 1293 return middle_ptr; 1294 } 1295 } 1296 return sd.region(left); 1297 } 1298 1299 ParallelCompactData::RegionData* 1300 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1301 const RegionData* end, 1302 size_t dead_words) 1303 { 1304 ParallelCompactData& sd = summary_data(); 1305 size_t left = sd.region(beg); 1306 size_t right = end > beg ? sd.region(end) - 1 : left; 1307 1308 // Binary search. 1309 while (left < right) { 1310 // Equivalent to (left + right) / 2, but does not overflow. 1311 const size_t middle = left + (right - left) / 2; 1312 RegionData* const middle_ptr = sd.region(middle); 1313 HeapWord* const dest = middle_ptr->destination(); 1314 HeapWord* const addr = sd.region_to_addr(middle); 1315 assert(dest != NULL, "sanity"); 1316 assert(dest <= addr, "must move left"); 1317 1318 const size_t dead_to_left = pointer_delta(addr, dest); 1319 if (middle > left && dead_to_left > dead_words) { 1320 right = middle - 1; 1321 } else if (middle < right && dead_to_left < dead_words) { 1322 left = middle + 1; 1323 } else { 1324 return middle_ptr; 1325 } 1326 } 1327 return sd.region(left); 1328 } 1329 1330 // The result is valid during the summary phase, after the initial summarization 1331 // of each space into itself, and before final summarization. 1332 inline double 1333 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1334 HeapWord* const bottom, 1335 HeapWord* const top, 1336 HeapWord* const new_top) 1337 { 1338 ParallelCompactData& sd = summary_data(); 1339 1340 assert(cp != NULL, "sanity"); 1341 assert(bottom != NULL, "sanity"); 1342 assert(top != NULL, "sanity"); 1343 assert(new_top != NULL, "sanity"); 1344 assert(top >= new_top, "summary data problem?"); 1345 assert(new_top > bottom, "space is empty; should not be here"); 1346 assert(new_top >= cp->destination(), "sanity"); 1347 assert(top >= sd.region_to_addr(cp), "sanity"); 1348 1349 HeapWord* const destination = cp->destination(); 1350 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1351 const size_t compacted_region_live = pointer_delta(new_top, destination); 1352 const size_t compacted_region_used = pointer_delta(top, 1353 sd.region_to_addr(cp)); 1354 const size_t reclaimable = compacted_region_used - compacted_region_live; 1355 1356 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1357 return double(reclaimable) / divisor; 1358 } 1359 1360 // Return the address of the end of the dense prefix, a.k.a. the start of the 1361 // compacted region. The address is always on a region boundary. 1362 // 1363 // Completely full regions at the left are skipped, since no compaction can 1364 // occur in those regions. Then the maximum amount of dead wood to allow is 1365 // computed, based on the density (amount live / capacity) of the generation; 1366 // the region with approximately that amount of dead space to the left is 1367 // identified as the limit region. Regions between the last completely full 1368 // region and the limit region are scanned and the one that has the best 1369 // (maximum) reclaimed_ratio() is selected. 1370 HeapWord* 1371 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1372 bool maximum_compaction) 1373 { 1374 if (ParallelOldGCSplitALot) { 1375 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1376 // The value was chosen to provoke splitting a young gen space; use it. 1377 return _space_info[id].dense_prefix(); 1378 } 1379 } 1380 1381 const size_t region_size = ParallelCompactData::RegionSize; 1382 const ParallelCompactData& sd = summary_data(); 1383 1384 const MutableSpace* const space = _space_info[id].space(); 1385 HeapWord* const top = space->top(); 1386 HeapWord* const top_aligned_up = sd.region_align_up(top); 1387 HeapWord* const new_top = _space_info[id].new_top(); 1388 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1389 HeapWord* const bottom = space->bottom(); 1390 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1391 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1392 const RegionData* const new_top_cp = 1393 sd.addr_to_region_ptr(new_top_aligned_up); 1394 1395 // Skip full regions at the beginning of the space--they are necessarily part 1396 // of the dense prefix. 1397 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1398 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1399 space->is_empty(), "no dead space allowed to the left"); 1400 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1401 "region must have dead space"); 1402 1403 // The gc number is saved whenever a maximum compaction is done, and used to 1404 // determine when the maximum compaction interval has expired. This avoids 1405 // successive max compactions for different reasons. 1406 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1407 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1408 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1409 total_invocations() == HeapFirstMaximumCompactionCount; 1410 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1411 _maximum_compaction_gc_num = total_invocations(); 1412 return sd.region_to_addr(full_cp); 1413 } 1414 1415 const size_t space_live = pointer_delta(new_top, bottom); 1416 const size_t space_used = space->used_in_words(); 1417 const size_t space_capacity = space->capacity_in_words(); 1418 1419 const double density = double(space_live) / double(space_capacity); 1420 const size_t min_percent_free = MarkSweepDeadRatio; 1421 const double limiter = dead_wood_limiter(density, min_percent_free); 1422 const size_t dead_wood_max = space_used - space_live; 1423 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1424 dead_wood_max); 1425 1426 if (TraceParallelOldGCDensePrefix) { 1427 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1428 "space_cap=" SIZE_FORMAT, 1429 space_live, space_used, 1430 space_capacity); 1431 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1432 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1433 density, min_percent_free, limiter, 1434 dead_wood_max, dead_wood_limit); 1435 } 1436 1437 // Locate the region with the desired amount of dead space to the left. 1438 const RegionData* const limit_cp = 1439 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1440 1441 // Scan from the first region with dead space to the limit region and find the 1442 // one with the best (largest) reclaimed ratio. 1443 double best_ratio = 0.0; 1444 const RegionData* best_cp = full_cp; 1445 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1446 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1447 if (tmp_ratio > best_ratio) { 1448 best_cp = cp; 1449 best_ratio = tmp_ratio; 1450 } 1451 } 1452 1453 #if 0 1454 // Something to consider: if the region with the best ratio is 'close to' the 1455 // first region w/free space, choose the first region with free space 1456 // ("first-free"). The first-free region is usually near the start of the 1457 // heap, which means we are copying most of the heap already, so copy a bit 1458 // more to get complete compaction. 1459 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1460 _maximum_compaction_gc_num = total_invocations(); 1461 best_cp = full_cp; 1462 } 1463 #endif // #if 0 1464 1465 return sd.region_to_addr(best_cp); 1466 } 1467 1468 #ifndef PRODUCT 1469 void 1470 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1471 size_t words) 1472 { 1473 if (TraceParallelOldGCSummaryPhase) { 1474 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1475 SIZE_FORMAT, start, start + words, words); 1476 } 1477 1478 ObjectStartArray* const start_array = _space_info[id].start_array(); 1479 CollectedHeap::fill_with_objects(start, words); 1480 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1481 _mark_bitmap.mark_obj(p, words); 1482 _summary_data.add_obj(p, words); 1483 start_array->allocate_block(p); 1484 } 1485 } 1486 1487 void 1488 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1489 { 1490 ParallelCompactData& sd = summary_data(); 1491 MutableSpace* space = _space_info[id].space(); 1492 1493 // Find the source and destination start addresses. 1494 HeapWord* const src_addr = sd.region_align_down(start); 1495 HeapWord* dst_addr; 1496 if (src_addr < start) { 1497 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1498 } else if (src_addr > space->bottom()) { 1499 // The start (the original top() value) is aligned to a region boundary so 1500 // the associated region does not have a destination. Compute the 1501 // destination from the previous region. 1502 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1503 dst_addr = cp->destination() + cp->data_size(); 1504 } else { 1505 // Filling the entire space. 1506 dst_addr = space->bottom(); 1507 } 1508 assert(dst_addr != NULL, "sanity"); 1509 1510 // Update the summary data. 1511 bool result = _summary_data.summarize(_space_info[id].split_info(), 1512 src_addr, space->top(), NULL, 1513 dst_addr, space->end(), 1514 _space_info[id].new_top_addr()); 1515 assert(result, "should not fail: bad filler object size"); 1516 } 1517 1518 void 1519 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1520 { 1521 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1522 return; 1523 } 1524 1525 MutableSpace* const space = _space_info[id].space(); 1526 if (space->is_empty()) { 1527 HeapWord* b = space->bottom(); 1528 HeapWord* t = b + space->capacity_in_words() / 2; 1529 space->set_top(t); 1530 if (ZapUnusedHeapArea) { 1531 space->set_top_for_allocations(); 1532 } 1533 1534 size_t min_size = CollectedHeap::min_fill_size(); 1535 size_t obj_len = min_size; 1536 while (b + obj_len <= t) { 1537 CollectedHeap::fill_with_object(b, obj_len); 1538 mark_bitmap()->mark_obj(b, obj_len); 1539 summary_data().add_obj(b, obj_len); 1540 b += obj_len; 1541 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1542 } 1543 if (b < t) { 1544 // The loop didn't completely fill to t (top); adjust top downward. 1545 space->set_top(b); 1546 if (ZapUnusedHeapArea) { 1547 space->set_top_for_allocations(); 1548 } 1549 } 1550 1551 HeapWord** nta = _space_info[id].new_top_addr(); 1552 bool result = summary_data().summarize(_space_info[id].split_info(), 1553 space->bottom(), space->top(), NULL, 1554 space->bottom(), space->end(), nta); 1555 assert(result, "space must fit into itself"); 1556 } 1557 } 1558 1559 void 1560 PSParallelCompact::provoke_split(bool & max_compaction) 1561 { 1562 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1563 return; 1564 } 1565 1566 const size_t region_size = ParallelCompactData::RegionSize; 1567 ParallelCompactData& sd = summary_data(); 1568 1569 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1570 MutableSpace* const from_space = _space_info[from_space_id].space(); 1571 const size_t eden_live = pointer_delta(eden_space->top(), 1572 _space_info[eden_space_id].new_top()); 1573 const size_t from_live = pointer_delta(from_space->top(), 1574 _space_info[from_space_id].new_top()); 1575 1576 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1577 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1578 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1579 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1580 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1581 1582 // Choose the space to split; need at least 2 regions live (or fillable). 1583 SpaceId id; 1584 MutableSpace* space; 1585 size_t live_words; 1586 size_t fill_words; 1587 if (eden_live + eden_fillable >= region_size * 2) { 1588 id = eden_space_id; 1589 space = eden_space; 1590 live_words = eden_live; 1591 fill_words = eden_fillable; 1592 } else if (from_live + from_fillable >= region_size * 2) { 1593 id = from_space_id; 1594 space = from_space; 1595 live_words = from_live; 1596 fill_words = from_fillable; 1597 } else { 1598 return; // Give up. 1599 } 1600 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1601 1602 if (live_words < region_size * 2) { 1603 // Fill from top() to end() w/live objects of mixed sizes. 1604 HeapWord* const fill_start = space->top(); 1605 live_words += fill_words; 1606 1607 space->set_top(fill_start + fill_words); 1608 if (ZapUnusedHeapArea) { 1609 space->set_top_for_allocations(); 1610 } 1611 1612 HeapWord* cur_addr = fill_start; 1613 while (fill_words > 0) { 1614 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1615 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1616 if (fill_words - cur_size < min_fill_size) { 1617 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1618 } 1619 1620 CollectedHeap::fill_with_object(cur_addr, cur_size); 1621 mark_bitmap()->mark_obj(cur_addr, cur_size); 1622 sd.add_obj(cur_addr, cur_size); 1623 1624 cur_addr += cur_size; 1625 fill_words -= cur_size; 1626 } 1627 1628 summarize_new_objects(id, fill_start); 1629 } 1630 1631 max_compaction = false; 1632 1633 // Manipulate the old gen so that it has room for about half of the live data 1634 // in the target young gen space (live_words / 2). 1635 id = old_space_id; 1636 space = _space_info[id].space(); 1637 const size_t free_at_end = space->free_in_words(); 1638 const size_t free_target = align_object_size(live_words / 2); 1639 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1640 1641 if (free_at_end >= free_target + min_fill_size) { 1642 // Fill space above top() and set the dense prefix so everything survives. 1643 HeapWord* const fill_start = space->top(); 1644 const size_t fill_size = free_at_end - free_target; 1645 space->set_top(space->top() + fill_size); 1646 if (ZapUnusedHeapArea) { 1647 space->set_top_for_allocations(); 1648 } 1649 fill_with_live_objects(id, fill_start, fill_size); 1650 summarize_new_objects(id, fill_start); 1651 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1652 } else if (dead + free_at_end > free_target) { 1653 // Find a dense prefix that makes the right amount of space available. 1654 HeapWord* cur = sd.region_align_down(space->top()); 1655 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1656 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1657 while (dead_to_right < free_target) { 1658 cur -= region_size; 1659 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1660 dead_to_right = pointer_delta(space->end(), cur_destination); 1661 } 1662 _space_info[id].set_dense_prefix(cur); 1663 } 1664 } 1665 #endif // #ifndef PRODUCT 1666 1667 void PSParallelCompact::summarize_spaces_quick() 1668 { 1669 for (unsigned int i = 0; i < last_space_id; ++i) { 1670 const MutableSpace* space = _space_info[i].space(); 1671 HeapWord** nta = _space_info[i].new_top_addr(); 1672 bool result = _summary_data.summarize(_space_info[i].split_info(), 1673 space->bottom(), space->top(), NULL, 1674 space->bottom(), space->end(), nta); 1675 assert(result, "space must fit into itself"); 1676 _space_info[i].set_dense_prefix(space->bottom()); 1677 } 1678 1679 #ifndef PRODUCT 1680 if (ParallelOldGCSplitALot) { 1681 provoke_split_fill_survivor(to_space_id); 1682 } 1683 #endif // #ifndef PRODUCT 1684 } 1685 1686 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1687 { 1688 HeapWord* const dense_prefix_end = dense_prefix(id); 1689 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1690 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1691 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1692 // Only enough dead space is filled so that any remaining dead space to the 1693 // left is larger than the minimum filler object. (The remainder is filled 1694 // during the copy/update phase.) 1695 // 1696 // The size of the dead space to the right of the boundary is not a 1697 // concern, since compaction will be able to use whatever space is 1698 // available. 1699 // 1700 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1701 // surrounds the space to be filled with an object. 1702 // 1703 // In the 32-bit VM, each bit represents two 32-bit words: 1704 // +---+ 1705 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1706 // end_bits: ... x x x | 0 | || 0 x x ... 1707 // +---+ 1708 // 1709 // In the 64-bit VM, each bit represents one 64-bit word: 1710 // +------------+ 1711 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1712 // end_bits: ... x x 1 | 0 || 0 | x x ... 1713 // +------------+ 1714 // +-------+ 1715 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1716 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1717 // +-------+ 1718 // +-----------+ 1719 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1720 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1721 // +-----------+ 1722 // +-------+ 1723 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1724 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1725 // +-------+ 1726 1727 // Initially assume case a, c or e will apply. 1728 size_t obj_len = CollectedHeap::min_fill_size(); 1729 HeapWord* obj_beg = dense_prefix_end - obj_len; 1730 1731 #ifdef _LP64 1732 if (MinObjAlignment > 1) { // object alignment > heap word size 1733 // Cases a, c or e. 1734 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1735 // Case b above. 1736 obj_beg = dense_prefix_end - 1; 1737 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1738 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1739 // Case d above. 1740 obj_beg = dense_prefix_end - 3; 1741 obj_len = 3; 1742 } 1743 #endif // #ifdef _LP64 1744 1745 CollectedHeap::fill_with_object(obj_beg, obj_len); 1746 _mark_bitmap.mark_obj(obj_beg, obj_len); 1747 _summary_data.add_obj(obj_beg, obj_len); 1748 assert(start_array(id) != NULL, "sanity"); 1749 start_array(id)->allocate_block(obj_beg); 1750 } 1751 } 1752 1753 void 1754 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1755 { 1756 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1757 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1758 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1759 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1760 cur->set_source_region(0); 1761 } 1762 } 1763 1764 void 1765 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1766 { 1767 assert(id < last_space_id, "id out of range"); 1768 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1769 ParallelOldGCSplitALot && id == old_space_id, 1770 "should have been reset in summarize_spaces_quick()"); 1771 1772 const MutableSpace* space = _space_info[id].space(); 1773 if (_space_info[id].new_top() != space->bottom()) { 1774 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1775 _space_info[id].set_dense_prefix(dense_prefix_end); 1776 1777 #ifndef PRODUCT 1778 if (TraceParallelOldGCDensePrefix) { 1779 print_dense_prefix_stats("ratio", id, maximum_compaction, 1780 dense_prefix_end); 1781 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1782 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1783 } 1784 #endif // #ifndef PRODUCT 1785 1786 // Recompute the summary data, taking into account the dense prefix. If 1787 // every last byte will be reclaimed, then the existing summary data which 1788 // compacts everything can be left in place. 1789 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1790 // If dead space crosses the dense prefix boundary, it is (at least 1791 // partially) filled with a dummy object, marked live and added to the 1792 // summary data. This simplifies the copy/update phase and must be done 1793 // before the final locations of objects are determined, to prevent 1794 // leaving a fragment of dead space that is too small to fill. 1795 fill_dense_prefix_end(id); 1796 1797 // Compute the destination of each Region, and thus each object. 1798 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1799 _summary_data.summarize(_space_info[id].split_info(), 1800 dense_prefix_end, space->top(), NULL, 1801 dense_prefix_end, space->end(), 1802 _space_info[id].new_top_addr()); 1803 } 1804 } 1805 1806 if (TraceParallelOldGCSummaryPhase) { 1807 const size_t region_size = ParallelCompactData::RegionSize; 1808 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1809 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1810 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1811 HeapWord* const new_top = _space_info[id].new_top(); 1812 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1813 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1814 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1815 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1816 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1817 id, space->capacity_in_words(), dense_prefix_end, 1818 dp_region, dp_words / region_size, 1819 cr_words / region_size, new_top); 1820 } 1821 } 1822 1823 #ifndef PRODUCT 1824 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1825 HeapWord* dst_beg, HeapWord* dst_end, 1826 SpaceId src_space_id, 1827 HeapWord* src_beg, HeapWord* src_end) 1828 { 1829 if (TraceParallelOldGCSummaryPhase) { 1830 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1831 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1832 SIZE_FORMAT "-" SIZE_FORMAT " " 1833 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1834 SIZE_FORMAT "-" SIZE_FORMAT, 1835 src_space_id, space_names[src_space_id], 1836 dst_space_id, space_names[dst_space_id], 1837 src_beg, src_end, 1838 _summary_data.addr_to_region_idx(src_beg), 1839 _summary_data.addr_to_region_idx(src_end), 1840 dst_beg, dst_end, 1841 _summary_data.addr_to_region_idx(dst_beg), 1842 _summary_data.addr_to_region_idx(dst_end)); 1843 } 1844 } 1845 #endif // #ifndef PRODUCT 1846 1847 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1848 bool maximum_compaction) 1849 { 1850 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1851 // trace("2"); 1852 1853 #ifdef ASSERT 1854 if (TraceParallelOldGCMarkingPhase) { 1855 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1856 "add_obj_bytes=" SIZE_FORMAT, 1857 add_obj_count, add_obj_size * HeapWordSize); 1858 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1859 "mark_bitmap_bytes=" SIZE_FORMAT, 1860 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1861 } 1862 #endif // #ifdef ASSERT 1863 1864 // Quick summarization of each space into itself, to see how much is live. 1865 summarize_spaces_quick(); 1866 1867 if (TraceParallelOldGCSummaryPhase) { 1868 tty->print_cr("summary_phase: after summarizing each space to self"); 1869 Universe::print(); 1870 NOT_PRODUCT(print_region_ranges()); 1871 if (Verbose) { 1872 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1873 } 1874 } 1875 1876 // The amount of live data that will end up in old space (assuming it fits). 1877 size_t old_space_total_live = 0; 1878 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1879 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1880 _space_info[id].space()->bottom()); 1881 } 1882 1883 MutableSpace* const old_space = _space_info[old_space_id].space(); 1884 const size_t old_capacity = old_space->capacity_in_words(); 1885 if (old_space_total_live > old_capacity) { 1886 // XXX - should also try to expand 1887 maximum_compaction = true; 1888 } 1889 #ifndef PRODUCT 1890 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1891 provoke_split(maximum_compaction); 1892 } 1893 #endif // #ifndef PRODUCT 1894 1895 // Old generations. 1896 summarize_space(old_space_id, maximum_compaction); 1897 1898 // Summarize the remaining spaces in the young gen. The initial target space 1899 // is the old gen. If a space does not fit entirely into the target, then the 1900 // remainder is compacted into the space itself and that space becomes the new 1901 // target. 1902 SpaceId dst_space_id = old_space_id; 1903 HeapWord* dst_space_end = old_space->end(); 1904 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1905 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1906 const MutableSpace* space = _space_info[id].space(); 1907 const size_t live = pointer_delta(_space_info[id].new_top(), 1908 space->bottom()); 1909 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1910 1911 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1912 SpaceId(id), space->bottom(), space->top());) 1913 if (live > 0 && live <= available) { 1914 // All the live data will fit. 1915 bool done = _summary_data.summarize(_space_info[id].split_info(), 1916 space->bottom(), space->top(), 1917 NULL, 1918 *new_top_addr, dst_space_end, 1919 new_top_addr); 1920 assert(done, "space must fit into old gen"); 1921 1922 // Reset the new_top value for the space. 1923 _space_info[id].set_new_top(space->bottom()); 1924 } else if (live > 0) { 1925 // Attempt to fit part of the source space into the target space. 1926 HeapWord* next_src_addr = NULL; 1927 bool done = _summary_data.summarize(_space_info[id].split_info(), 1928 space->bottom(), space->top(), 1929 &next_src_addr, 1930 *new_top_addr, dst_space_end, 1931 new_top_addr); 1932 assert(!done, "space should not fit into old gen"); 1933 assert(next_src_addr != NULL, "sanity"); 1934 1935 // The source space becomes the new target, so the remainder is compacted 1936 // within the space itself. 1937 dst_space_id = SpaceId(id); 1938 dst_space_end = space->end(); 1939 new_top_addr = _space_info[id].new_top_addr(); 1940 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1941 space->bottom(), dst_space_end, 1942 SpaceId(id), next_src_addr, space->top());) 1943 done = _summary_data.summarize(_space_info[id].split_info(), 1944 next_src_addr, space->top(), 1945 NULL, 1946 space->bottom(), dst_space_end, 1947 new_top_addr); 1948 assert(done, "space must fit when compacted into itself"); 1949 assert(*new_top_addr <= space->top(), "usage should not grow"); 1950 } 1951 } 1952 1953 if (TraceParallelOldGCSummaryPhase) { 1954 tty->print_cr("summary_phase: after final summarization"); 1955 Universe::print(); 1956 NOT_PRODUCT(print_region_ranges()); 1957 if (Verbose) { 1958 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1959 } 1960 } 1961 } 1962 1963 // This method should contain all heap-specific policy for invoking a full 1964 // collection. invoke_no_policy() will only attempt to compact the heap; it 1965 // will do nothing further. If we need to bail out for policy reasons, scavenge 1966 // before full gc, or any other specialized behavior, it needs to be added here. 1967 // 1968 // Note that this method should only be called from the vm_thread while at a 1969 // safepoint. 1970 // 1971 // Note that the all_soft_refs_clear flag in the collector policy 1972 // may be true because this method can be called without intervening 1973 // activity. For example when the heap space is tight and full measure 1974 // are being taken to free space. 1975 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1976 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1977 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1978 "should be in vm thread"); 1979 1980 ParallelScavengeHeap* heap = gc_heap(); 1981 GCCause::Cause gc_cause = heap->gc_cause(); 1982 assert(!heap->is_gc_active(), "not reentrant"); 1983 1984 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1985 IsGCActiveMark mark; 1986 1987 if (ScavengeBeforeFullGC) { 1988 PSScavenge::invoke_no_policy(); 1989 } 1990 1991 const bool clear_all_soft_refs = 1992 heap->collector_policy()->should_clear_all_soft_refs(); 1993 1994 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1995 maximum_heap_compaction); 1996 } 1997 1998 // This method contains no policy. You should probably 1999 // be calling invoke() instead. 2000 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 2001 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 2002 assert(ref_processor() != NULL, "Sanity"); 2003 2004 if (GC_locker::check_active_before_gc()) { 2005 return false; 2006 } 2007 2008 ParallelScavengeHeap* heap = gc_heap(); 2009 2010 _gc_timer.register_gc_start(); 2011 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 2012 2013 TimeStamp marking_start; 2014 TimeStamp compaction_start; 2015 TimeStamp collection_exit; 2016 2017 GCCause::Cause gc_cause = heap->gc_cause(); 2018 PSYoungGen* young_gen = heap->young_gen(); 2019 PSOldGen* old_gen = heap->old_gen(); 2020 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2021 2022 // The scope of casr should end after code that can change 2023 // CollectorPolicy::_should_clear_all_soft_refs. 2024 ClearedAllSoftRefs casr(maximum_heap_compaction, 2025 heap->collector_policy()); 2026 2027 if (ZapUnusedHeapArea) { 2028 // Save information needed to minimize mangling 2029 heap->record_gen_tops_before_GC(); 2030 } 2031 2032 heap->pre_full_gc_dump(&_gc_timer); 2033 2034 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2035 2036 // Make sure data structures are sane, make the heap parsable, and do other 2037 // miscellaneous bookkeeping. 2038 PreGCValues pre_gc_values; 2039 pre_compact(&pre_gc_values); 2040 2041 // Get the compaction manager reserved for the VM thread. 2042 ParCompactionManager* const vmthread_cm = 2043 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2044 2045 // Place after pre_compact() where the number of invocations is incremented. 2046 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2047 2048 { 2049 ResourceMark rm; 2050 HandleMark hm; 2051 2052 // Set the number of GC threads to be used in this collection 2053 gc_task_manager()->set_active_gang(); 2054 gc_task_manager()->task_idle_workers(); 2055 heap->set_par_threads(gc_task_manager()->active_workers()); 2056 2057 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2058 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id()); 2059 TraceCollectorStats tcs(counters()); 2060 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2061 2062 if (TraceGen1Time) accumulated_time()->start(); 2063 2064 // Let the size policy know we're starting 2065 size_policy->major_collection_begin(); 2066 2067 CodeCache::gc_prologue(); 2068 Threads::gc_prologue(); 2069 2070 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2071 2072 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 2073 ref_processor()->setup_policy(maximum_heap_compaction); 2074 2075 bool marked_for_unloading = false; 2076 2077 marking_start.update(); 2078 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2079 2080 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2081 && gc_cause == GCCause::_java_lang_system_gc; 2082 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2083 2084 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2085 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2086 2087 // adjust_roots() updates Universe::_intArrayKlassObj which is 2088 // needed by the compaction for filling holes in the dense prefix. 2089 adjust_roots(); 2090 2091 compaction_start.update(); 2092 compact(); 2093 2094 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2095 // done before resizing. 2096 post_compact(); 2097 2098 // Let the size policy know we're done 2099 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2100 2101 if (UseAdaptiveSizePolicy) { 2102 if (PrintAdaptiveSizePolicy) { 2103 gclog_or_tty->print("AdaptiveSizeStart: "); 2104 gclog_or_tty->stamp(); 2105 gclog_or_tty->print_cr(" collection: %d ", 2106 heap->total_collections()); 2107 if (Verbose) { 2108 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d", 2109 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2110 } 2111 } 2112 2113 // Don't check if the size_policy is ready here. Let 2114 // the size_policy check that internally. 2115 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2116 ((gc_cause != GCCause::_java_lang_system_gc) || 2117 UseAdaptiveSizePolicyWithSystemGC)) { 2118 // Calculate optimal free space amounts 2119 assert(young_gen->max_size() > 2120 young_gen->from_space()->capacity_in_bytes() + 2121 young_gen->to_space()->capacity_in_bytes(), 2122 "Sizes of space in young gen are out-of-bounds"); 2123 2124 size_t young_live = young_gen->used_in_bytes(); 2125 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2126 size_t old_live = old_gen->used_in_bytes(); 2127 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2128 size_t max_old_gen_size = old_gen->max_gen_size(); 2129 size_t max_eden_size = young_gen->max_size() - 2130 young_gen->from_space()->capacity_in_bytes() - 2131 young_gen->to_space()->capacity_in_bytes(); 2132 2133 // Used for diagnostics 2134 size_policy->clear_generation_free_space_flags(); 2135 2136 size_policy->compute_generations_free_space(young_live, 2137 eden_live, 2138 old_live, 2139 cur_eden, 2140 max_old_gen_size, 2141 max_eden_size, 2142 true /* full gc*/); 2143 2144 size_policy->check_gc_overhead_limit(young_live, 2145 eden_live, 2146 max_old_gen_size, 2147 max_eden_size, 2148 true /* full gc*/, 2149 gc_cause, 2150 heap->collector_policy()); 2151 2152 size_policy->decay_supplemental_growth(true /* full gc*/); 2153 2154 heap->resize_old_gen( 2155 size_policy->calculated_old_free_size_in_bytes()); 2156 2157 // Don't resize the young generation at an major collection. A 2158 // desired young generation size may have been calculated but 2159 // resizing the young generation complicates the code because the 2160 // resizing of the old generation may have moved the boundary 2161 // between the young generation and the old generation. Let the 2162 // young generation resizing happen at the minor collections. 2163 } 2164 if (PrintAdaptiveSizePolicy) { 2165 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2166 heap->total_collections()); 2167 } 2168 } 2169 2170 if (UsePerfData) { 2171 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2172 counters->update_counters(); 2173 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2174 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2175 } 2176 2177 heap->resize_all_tlabs(); 2178 2179 // Resize the metaspace capactiy after a collection 2180 MetaspaceGC::compute_new_size(); 2181 2182 if (TraceGen1Time) accumulated_time()->stop(); 2183 2184 if (PrintGC) { 2185 if (PrintGCDetails) { 2186 // No GC timestamp here. This is after GC so it would be confusing. 2187 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2188 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2189 heap->print_heap_change(pre_gc_values.heap_used()); 2190 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2191 } else { 2192 heap->print_heap_change(pre_gc_values.heap_used()); 2193 } 2194 } 2195 2196 // Track memory usage and detect low memory 2197 MemoryService::track_memory_usage(); 2198 heap->update_counters(); 2199 gc_task_manager()->release_idle_workers(); 2200 } 2201 2202 #ifdef ASSERT 2203 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2204 ParCompactionManager* const cm = 2205 ParCompactionManager::manager_array(int(i)); 2206 assert(cm->marking_stack()->is_empty(), "should be empty"); 2207 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2208 } 2209 #endif // ASSERT 2210 2211 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2212 HandleMark hm; // Discard invalid handles created during verification 2213 Universe::verify(" VerifyAfterGC:"); 2214 } 2215 2216 // Re-verify object start arrays 2217 if (VerifyObjectStartArray && 2218 VerifyAfterGC) { 2219 old_gen->verify_object_start_array(); 2220 } 2221 2222 if (ZapUnusedHeapArea) { 2223 old_gen->object_space()->check_mangled_unused_area_complete(); 2224 } 2225 2226 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2227 2228 collection_exit.update(); 2229 2230 heap->print_heap_after_gc(); 2231 heap->trace_heap_after_gc(&_gc_tracer); 2232 2233 if (PrintGCTaskTimeStamps) { 2234 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2235 INT64_FORMAT, 2236 marking_start.ticks(), compaction_start.ticks(), 2237 collection_exit.ticks()); 2238 gc_task_manager()->print_task_time_stamps(); 2239 } 2240 2241 heap->post_full_gc_dump(&_gc_timer); 2242 2243 #ifdef TRACESPINNING 2244 ParallelTaskTerminator::print_termination_counts(); 2245 #endif 2246 2247 _gc_timer.register_gc_end(); 2248 2249 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2250 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2251 2252 return true; 2253 } 2254 2255 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2256 PSYoungGen* young_gen, 2257 PSOldGen* old_gen) { 2258 MutableSpace* const eden_space = young_gen->eden_space(); 2259 assert(!eden_space->is_empty(), "eden must be non-empty"); 2260 assert(young_gen->virtual_space()->alignment() == 2261 old_gen->virtual_space()->alignment(), "alignments do not match"); 2262 2263 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2264 return false; 2265 } 2266 2267 // Both generations must be completely committed. 2268 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2269 return false; 2270 } 2271 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2272 return false; 2273 } 2274 2275 // Figure out how much to take from eden. Include the average amount promoted 2276 // in the total; otherwise the next young gen GC will simply bail out to a 2277 // full GC. 2278 const size_t alignment = old_gen->virtual_space()->alignment(); 2279 const size_t eden_used = eden_space->used_in_bytes(); 2280 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2281 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2282 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2283 2284 if (absorb_size >= eden_capacity) { 2285 return false; // Must leave some space in eden. 2286 } 2287 2288 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2289 if (new_young_size < young_gen->min_gen_size()) { 2290 return false; // Respect young gen minimum size. 2291 } 2292 2293 if (TraceAdaptiveGCBoundary && Verbose) { 2294 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2295 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2296 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2297 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2298 absorb_size / K, 2299 eden_capacity / K, (eden_capacity - absorb_size) / K, 2300 young_gen->from_space()->used_in_bytes() / K, 2301 young_gen->to_space()->used_in_bytes() / K, 2302 young_gen->capacity_in_bytes() / K, new_young_size / K); 2303 } 2304 2305 // Fill the unused part of the old gen. 2306 MutableSpace* const old_space = old_gen->object_space(); 2307 HeapWord* const unused_start = old_space->top(); 2308 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2309 2310 if (unused_words > 0) { 2311 if (unused_words < CollectedHeap::min_fill_size()) { 2312 return false; // If the old gen cannot be filled, must give up. 2313 } 2314 CollectedHeap::fill_with_objects(unused_start, unused_words); 2315 } 2316 2317 // Take the live data from eden and set both top and end in the old gen to 2318 // eden top. (Need to set end because reset_after_change() mangles the region 2319 // from end to virtual_space->high() in debug builds). 2320 HeapWord* const new_top = eden_space->top(); 2321 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2322 absorb_size); 2323 young_gen->reset_after_change(); 2324 old_space->set_top(new_top); 2325 old_space->set_end(new_top); 2326 old_gen->reset_after_change(); 2327 2328 // Update the object start array for the filler object and the data from eden. 2329 ObjectStartArray* const start_array = old_gen->start_array(); 2330 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2331 start_array->allocate_block(p); 2332 } 2333 2334 // Could update the promoted average here, but it is not typically updated at 2335 // full GCs and the value to use is unclear. Something like 2336 // 2337 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2338 2339 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2340 return true; 2341 } 2342 2343 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2344 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2345 "shouldn't return NULL"); 2346 return ParallelScavengeHeap::gc_task_manager(); 2347 } 2348 2349 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2350 bool maximum_heap_compaction, 2351 ParallelOldTracer *gc_tracer) { 2352 // Recursively traverse all live objects and mark them 2353 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2354 2355 ParallelScavengeHeap* heap = gc_heap(); 2356 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2357 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2358 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2359 ParallelTaskTerminator terminator(active_gc_threads, qset); 2360 2361 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2362 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2363 2364 // Need new claim bits before marking starts. 2365 ClassLoaderDataGraph::clear_claimed_marks(); 2366 2367 { 2368 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2369 2370 ParallelScavengeHeap::ParStrongRootsScope psrs; 2371 2372 GCTaskQueue* q = GCTaskQueue::create(); 2373 2374 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2375 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2376 // We scan the thread roots in parallel 2377 Threads::create_thread_roots_marking_tasks(q); 2378 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2379 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2380 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2381 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2382 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2383 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2384 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2385 2386 if (active_gc_threads > 1) { 2387 for (uint j = 0; j < active_gc_threads; j++) { 2388 q->enqueue(new StealMarkingTask(&terminator)); 2389 } 2390 } 2391 2392 gc_task_manager()->execute_and_wait(q); 2393 } 2394 2395 // Process reference objects found during marking 2396 { 2397 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2398 2399 ReferenceProcessorStats stats; 2400 if (ref_processor()->processing_is_mt()) { 2401 RefProcTaskExecutor task_executor; 2402 stats = ref_processor()->process_discovered_references( 2403 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2404 &task_executor, &_gc_timer, _gc_tracer.gc_id()); 2405 } else { 2406 stats = ref_processor()->process_discovered_references( 2407 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2408 &_gc_timer, _gc_tracer.gc_id()); 2409 } 2410 2411 gc_tracer->report_gc_reference_stats(stats); 2412 } 2413 2414 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2415 2416 // This is the point where the entire marking should have completed. 2417 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2418 2419 // Follow system dictionary roots and unload classes. 2420 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2421 2422 // Unload nmethods. 2423 CodeCache::do_unloading(is_alive_closure(), purged_class); 2424 2425 // Prune dead klasses from subklass/sibling/implementor lists. 2426 Klass::clean_weak_klass_links(is_alive_closure()); 2427 2428 // Delete entries for dead interned strings. 2429 StringTable::unlink(is_alive_closure()); 2430 2431 // Clean up unreferenced symbols in symbol table. 2432 SymbolTable::unlink(); 2433 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2434 } 2435 2436 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2437 ClassLoaderData* cld) { 2438 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2439 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2440 2441 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2442 } 2443 2444 // This should be moved to the shared markSweep code! 2445 class PSAlwaysTrueClosure: public BoolObjectClosure { 2446 public: 2447 bool do_object_b(oop p) { return true; } 2448 }; 2449 static PSAlwaysTrueClosure always_true; 2450 2451 void PSParallelCompact::adjust_roots() { 2452 // Adjust the pointers to reflect the new locations 2453 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2454 2455 // Need new claim bits when tracing through and adjusting pointers. 2456 ClassLoaderDataGraph::clear_claimed_marks(); 2457 2458 // General strong roots. 2459 Universe::oops_do(adjust_pointer_closure()); 2460 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2461 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2462 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2463 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2464 FlatProfiler::oops_do(adjust_pointer_closure()); 2465 Management::oops_do(adjust_pointer_closure()); 2466 JvmtiExport::oops_do(adjust_pointer_closure()); 2467 SystemDictionary::oops_do(adjust_pointer_closure()); 2468 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2469 2470 // Now adjust pointers in remaining weak roots. (All of which should 2471 // have been cleared if they pointed to non-surviving objects.) 2472 // Global (weak) JNI handles 2473 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2474 2475 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2476 CodeCache::blobs_do(&adjust_from_blobs); 2477 StringTable::oops_do(adjust_pointer_closure()); 2478 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2479 // Roots were visited so references into the young gen in roots 2480 // may have been scanned. Process them also. 2481 // Should the reference processor have a span that excludes 2482 // young gen objects? 2483 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2484 } 2485 2486 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2487 uint parallel_gc_threads) 2488 { 2489 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2490 2491 // Find the threads that are active 2492 unsigned int which = 0; 2493 2494 const uint task_count = MAX2(parallel_gc_threads, 1U); 2495 for (uint j = 0; j < task_count; j++) { 2496 q->enqueue(new DrainStacksCompactionTask(j)); 2497 ParCompactionManager::verify_region_list_empty(j); 2498 // Set the region stacks variables to "no" region stack values 2499 // so that they will be recognized and needing a region stack 2500 // in the stealing tasks if they do not get one by executing 2501 // a draining stack. 2502 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2503 cm->set_region_stack(NULL); 2504 cm->set_region_stack_index((uint)max_uintx); 2505 } 2506 ParCompactionManager::reset_recycled_stack_index(); 2507 2508 // Find all regions that are available (can be filled immediately) and 2509 // distribute them to the thread stacks. The iteration is done in reverse 2510 // order (high to low) so the regions will be removed in ascending order. 2511 2512 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2513 2514 size_t fillable_regions = 0; // A count for diagnostic purposes. 2515 // A region index which corresponds to the tasks created above. 2516 // "which" must be 0 <= which < task_count 2517 2518 which = 0; 2519 // id + 1 is used to test termination so unsigned can 2520 // be used with an old_space_id == 0. 2521 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2522 SpaceInfo* const space_info = _space_info + id; 2523 MutableSpace* const space = space_info->space(); 2524 HeapWord* const new_top = space_info->new_top(); 2525 2526 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2527 const size_t end_region = 2528 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2529 2530 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2531 if (sd.region(cur)->claim_unsafe()) { 2532 ParCompactionManager::region_list_push(which, cur); 2533 2534 if (TraceParallelOldGCCompactionPhase && Verbose) { 2535 const size_t count_mod_8 = fillable_regions & 7; 2536 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2537 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2538 if (count_mod_8 == 7) gclog_or_tty->cr(); 2539 } 2540 2541 NOT_PRODUCT(++fillable_regions;) 2542 2543 // Assign regions to tasks in round-robin fashion. 2544 if (++which == task_count) { 2545 assert(which <= parallel_gc_threads, 2546 "Inconsistent number of workers"); 2547 which = 0; 2548 } 2549 } 2550 } 2551 } 2552 2553 if (TraceParallelOldGCCompactionPhase) { 2554 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2555 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2556 } 2557 } 2558 2559 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2560 2561 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2562 uint parallel_gc_threads) { 2563 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2564 2565 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2566 2567 // Iterate over all the spaces adding tasks for updating 2568 // regions in the dense prefix. Assume that 1 gc thread 2569 // will work on opening the gaps and the remaining gc threads 2570 // will work on the dense prefix. 2571 unsigned int space_id; 2572 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2573 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2574 const MutableSpace* const space = _space_info[space_id].space(); 2575 2576 if (dense_prefix_end == space->bottom()) { 2577 // There is no dense prefix for this space. 2578 continue; 2579 } 2580 2581 // The dense prefix is before this region. 2582 size_t region_index_end_dense_prefix = 2583 sd.addr_to_region_idx(dense_prefix_end); 2584 RegionData* const dense_prefix_cp = 2585 sd.region(region_index_end_dense_prefix); 2586 assert(dense_prefix_end == space->end() || 2587 dense_prefix_cp->available() || 2588 dense_prefix_cp->claimed(), 2589 "The region after the dense prefix should always be ready to fill"); 2590 2591 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2592 2593 // Is there dense prefix work? 2594 size_t total_dense_prefix_regions = 2595 region_index_end_dense_prefix - region_index_start; 2596 // How many regions of the dense prefix should be given to 2597 // each thread? 2598 if (total_dense_prefix_regions > 0) { 2599 uint tasks_for_dense_prefix = 1; 2600 if (total_dense_prefix_regions <= 2601 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2602 // Don't over partition. This assumes that 2603 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2604 // so there are not many regions to process. 2605 tasks_for_dense_prefix = parallel_gc_threads; 2606 } else { 2607 // Over partition 2608 tasks_for_dense_prefix = parallel_gc_threads * 2609 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2610 } 2611 size_t regions_per_thread = total_dense_prefix_regions / 2612 tasks_for_dense_prefix; 2613 // Give each thread at least 1 region. 2614 if (regions_per_thread == 0) { 2615 regions_per_thread = 1; 2616 } 2617 2618 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2619 if (region_index_start >= region_index_end_dense_prefix) { 2620 break; 2621 } 2622 // region_index_end is not processed 2623 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2624 region_index_end_dense_prefix); 2625 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2626 region_index_start, 2627 region_index_end)); 2628 region_index_start = region_index_end; 2629 } 2630 } 2631 // This gets any part of the dense prefix that did not 2632 // fit evenly. 2633 if (region_index_start < region_index_end_dense_prefix) { 2634 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2635 region_index_start, 2636 region_index_end_dense_prefix)); 2637 } 2638 } 2639 } 2640 2641 void PSParallelCompact::enqueue_region_stealing_tasks( 2642 GCTaskQueue* q, 2643 ParallelTaskTerminator* terminator_ptr, 2644 uint parallel_gc_threads) { 2645 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2646 2647 // Once a thread has drained it's stack, it should try to steal regions from 2648 // other threads. 2649 if (parallel_gc_threads > 1) { 2650 for (uint j = 0; j < parallel_gc_threads; j++) { 2651 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2652 } 2653 } 2654 } 2655 2656 #ifdef ASSERT 2657 // Write a histogram of the number of times the block table was filled for a 2658 // region. 2659 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2660 { 2661 if (!TraceParallelOldGCCompactionPhase) return; 2662 2663 typedef ParallelCompactData::RegionData rd_t; 2664 ParallelCompactData& sd = summary_data(); 2665 2666 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2667 MutableSpace* const spc = _space_info[id].space(); 2668 if (spc->bottom() != spc->top()) { 2669 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2670 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2671 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2672 2673 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2674 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2675 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2676 2677 for (const rd_t* cur = beg; cur < end; ++cur) { 2678 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2679 } 2680 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2681 for (size_t i = 0; i < histo_len; ++i) { 2682 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2683 histo[i], 100.0 * histo[i] / region_cnt); 2684 } 2685 out->cr(); 2686 } 2687 } 2688 } 2689 #endif // #ifdef ASSERT 2690 2691 void PSParallelCompact::compact() { 2692 // trace("5"); 2693 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2694 2695 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2696 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2697 PSOldGen* old_gen = heap->old_gen(); 2698 old_gen->start_array()->reset(); 2699 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2700 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2701 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2702 ParallelTaskTerminator terminator(active_gc_threads, qset); 2703 2704 GCTaskQueue* q = GCTaskQueue::create(); 2705 enqueue_region_draining_tasks(q, active_gc_threads); 2706 enqueue_dense_prefix_tasks(q, active_gc_threads); 2707 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2708 2709 { 2710 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2711 2712 gc_task_manager()->execute_and_wait(q); 2713 2714 #ifdef ASSERT 2715 // Verify that all regions have been processed before the deferred updates. 2716 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2717 verify_complete(SpaceId(id)); 2718 } 2719 #endif 2720 } 2721 2722 { 2723 // Update the deferred objects, if any. Any compaction manager can be used. 2724 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2725 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2726 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2727 update_deferred_objects(cm, SpaceId(id)); 2728 } 2729 } 2730 2731 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2732 } 2733 2734 #ifdef ASSERT 2735 void PSParallelCompact::verify_complete(SpaceId space_id) { 2736 // All Regions between space bottom() to new_top() should be marked as filled 2737 // and all Regions between new_top() and top() should be available (i.e., 2738 // should have been emptied). 2739 ParallelCompactData& sd = summary_data(); 2740 SpaceInfo si = _space_info[space_id]; 2741 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2742 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2743 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2744 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2745 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2746 2747 bool issued_a_warning = false; 2748 2749 size_t cur_region; 2750 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2751 const RegionData* const c = sd.region(cur_region); 2752 if (!c->completed()) { 2753 warning("region " SIZE_FORMAT " not filled: " 2754 "destination_count=" SIZE_FORMAT, 2755 cur_region, c->destination_count()); 2756 issued_a_warning = true; 2757 } 2758 } 2759 2760 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2761 const RegionData* const c = sd.region(cur_region); 2762 if (!c->available()) { 2763 warning("region " SIZE_FORMAT " not empty: " 2764 "destination_count=" SIZE_FORMAT, 2765 cur_region, c->destination_count()); 2766 issued_a_warning = true; 2767 } 2768 } 2769 2770 if (issued_a_warning) { 2771 print_region_ranges(); 2772 } 2773 } 2774 #endif // #ifdef ASSERT 2775 2776 // Update interior oops in the ranges of regions [beg_region, end_region). 2777 void 2778 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2779 SpaceId space_id, 2780 size_t beg_region, 2781 size_t end_region) { 2782 ParallelCompactData& sd = summary_data(); 2783 ParMarkBitMap* const mbm = mark_bitmap(); 2784 2785 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2786 HeapWord* const end_addr = sd.region_to_addr(end_region); 2787 assert(beg_region <= end_region, "bad region range"); 2788 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2789 2790 #ifdef ASSERT 2791 // Claim the regions to avoid triggering an assert when they are marked as 2792 // filled. 2793 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2794 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2795 } 2796 #endif // #ifdef ASSERT 2797 2798 if (beg_addr != space(space_id)->bottom()) { 2799 // Find the first live object or block of dead space that *starts* in this 2800 // range of regions. If a partial object crosses onto the region, skip it; 2801 // it will be marked for 'deferred update' when the object head is 2802 // processed. If dead space crosses onto the region, it is also skipped; it 2803 // will be filled when the prior region is processed. If neither of those 2804 // apply, the first word in the region is the start of a live object or dead 2805 // space. 2806 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2807 const RegionData* const cp = sd.region(beg_region); 2808 if (cp->partial_obj_size() != 0) { 2809 beg_addr = sd.partial_obj_end(beg_region); 2810 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2811 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2812 } 2813 } 2814 2815 if (beg_addr < end_addr) { 2816 // A live object or block of dead space starts in this range of Regions. 2817 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2818 2819 // Create closures and iterate. 2820 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2821 FillClosure fill_closure(cm, space_id); 2822 ParMarkBitMap::IterationStatus status; 2823 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2824 dense_prefix_end); 2825 if (status == ParMarkBitMap::incomplete) { 2826 update_closure.do_addr(update_closure.source()); 2827 } 2828 } 2829 2830 // Mark the regions as filled. 2831 RegionData* const beg_cp = sd.region(beg_region); 2832 RegionData* const end_cp = sd.region(end_region); 2833 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2834 cp->set_completed(); 2835 } 2836 } 2837 2838 // Return the SpaceId for the space containing addr. If addr is not in the 2839 // heap, last_space_id is returned. In debug mode it expects the address to be 2840 // in the heap and asserts such. 2841 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2842 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2843 2844 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2845 if (_space_info[id].space()->contains(addr)) { 2846 return SpaceId(id); 2847 } 2848 } 2849 2850 assert(false, "no space contains the addr"); 2851 return last_space_id; 2852 } 2853 2854 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2855 SpaceId id) { 2856 assert(id < last_space_id, "bad space id"); 2857 2858 ParallelCompactData& sd = summary_data(); 2859 const SpaceInfo* const space_info = _space_info + id; 2860 ObjectStartArray* const start_array = space_info->start_array(); 2861 2862 const MutableSpace* const space = space_info->space(); 2863 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2864 HeapWord* const beg_addr = space_info->dense_prefix(); 2865 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2866 2867 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2868 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2869 const RegionData* cur_region; 2870 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2871 HeapWord* const addr = cur_region->deferred_obj_addr(); 2872 if (addr != NULL) { 2873 if (start_array != NULL) { 2874 start_array->allocate_block(addr); 2875 } 2876 oop(addr)->update_contents(cm); 2877 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 2878 } 2879 } 2880 } 2881 2882 // Skip over count live words starting from beg, and return the address of the 2883 // next live word. Unless marked, the word corresponding to beg is assumed to 2884 // be dead. Callers must either ensure beg does not correspond to the middle of 2885 // an object, or account for those live words in some other way. Callers must 2886 // also ensure that there are enough live words in the range [beg, end) to skip. 2887 HeapWord* 2888 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2889 { 2890 assert(count > 0, "sanity"); 2891 2892 ParMarkBitMap* m = mark_bitmap(); 2893 idx_t bits_to_skip = m->words_to_bits(count); 2894 idx_t cur_beg = m->addr_to_bit(beg); 2895 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2896 2897 do { 2898 cur_beg = m->find_obj_beg(cur_beg, search_end); 2899 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2900 const size_t obj_bits = cur_end - cur_beg + 1; 2901 if (obj_bits > bits_to_skip) { 2902 return m->bit_to_addr(cur_beg + bits_to_skip); 2903 } 2904 bits_to_skip -= obj_bits; 2905 cur_beg = cur_end + 1; 2906 } while (bits_to_skip > 0); 2907 2908 // Skipping the desired number of words landed just past the end of an object. 2909 // Find the start of the next object. 2910 cur_beg = m->find_obj_beg(cur_beg, search_end); 2911 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2912 return m->bit_to_addr(cur_beg); 2913 } 2914 2915 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2916 SpaceId src_space_id, 2917 size_t src_region_idx) 2918 { 2919 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2920 2921 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2922 if (split_info.dest_region_addr() == dest_addr) { 2923 // The partial object ending at the split point contains the first word to 2924 // be copied to dest_addr. 2925 return split_info.first_src_addr(); 2926 } 2927 2928 const ParallelCompactData& sd = summary_data(); 2929 ParMarkBitMap* const bitmap = mark_bitmap(); 2930 const size_t RegionSize = ParallelCompactData::RegionSize; 2931 2932 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2933 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2934 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2935 HeapWord* const src_region_destination = src_region_ptr->destination(); 2936 2937 assert(dest_addr >= src_region_destination, "wrong src region"); 2938 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2939 2940 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2941 HeapWord* const src_region_end = src_region_beg + RegionSize; 2942 2943 HeapWord* addr = src_region_beg; 2944 if (dest_addr == src_region_destination) { 2945 // Return the first live word in the source region. 2946 if (partial_obj_size == 0) { 2947 addr = bitmap->find_obj_beg(addr, src_region_end); 2948 assert(addr < src_region_end, "no objects start in src region"); 2949 } 2950 return addr; 2951 } 2952 2953 // Must skip some live data. 2954 size_t words_to_skip = dest_addr - src_region_destination; 2955 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2956 2957 if (partial_obj_size >= words_to_skip) { 2958 // All the live words to skip are part of the partial object. 2959 addr += words_to_skip; 2960 if (partial_obj_size == words_to_skip) { 2961 // Find the first live word past the partial object. 2962 addr = bitmap->find_obj_beg(addr, src_region_end); 2963 assert(addr < src_region_end, "wrong src region"); 2964 } 2965 return addr; 2966 } 2967 2968 // Skip over the partial object (if any). 2969 if (partial_obj_size != 0) { 2970 words_to_skip -= partial_obj_size; 2971 addr += partial_obj_size; 2972 } 2973 2974 // Skip over live words due to objects that start in the region. 2975 addr = skip_live_words(addr, src_region_end, words_to_skip); 2976 assert(addr < src_region_end, "wrong src region"); 2977 return addr; 2978 } 2979 2980 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2981 SpaceId src_space_id, 2982 size_t beg_region, 2983 HeapWord* end_addr) 2984 { 2985 ParallelCompactData& sd = summary_data(); 2986 2987 #ifdef ASSERT 2988 MutableSpace* const src_space = _space_info[src_space_id].space(); 2989 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2990 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2991 "src_space_id does not match beg_addr"); 2992 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2993 "src_space_id does not match end_addr"); 2994 #endif // #ifdef ASSERT 2995 2996 RegionData* const beg = sd.region(beg_region); 2997 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2998 2999 // Regions up to new_top() are enqueued if they become available. 3000 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3001 RegionData* const enqueue_end = 3002 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3003 3004 for (RegionData* cur = beg; cur < end; ++cur) { 3005 assert(cur->data_size() > 0, "region must have live data"); 3006 cur->decrement_destination_count(); 3007 if (cur < enqueue_end && cur->available() && cur->claim()) { 3008 cm->push_region(sd.region(cur)); 3009 } 3010 } 3011 } 3012 3013 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3014 SpaceId& src_space_id, 3015 HeapWord*& src_space_top, 3016 HeapWord* end_addr) 3017 { 3018 typedef ParallelCompactData::RegionData RegionData; 3019 3020 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3021 const size_t region_size = ParallelCompactData::RegionSize; 3022 3023 size_t src_region_idx = 0; 3024 3025 // Skip empty regions (if any) up to the top of the space. 3026 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3027 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3028 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3029 const RegionData* const top_region_ptr = 3030 sd.addr_to_region_ptr(top_aligned_up); 3031 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3032 ++src_region_ptr; 3033 } 3034 3035 if (src_region_ptr < top_region_ptr) { 3036 // The next source region is in the current space. Update src_region_idx 3037 // and the source address to match src_region_ptr. 3038 src_region_idx = sd.region(src_region_ptr); 3039 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3040 if (src_region_addr > closure.source()) { 3041 closure.set_source(src_region_addr); 3042 } 3043 return src_region_idx; 3044 } 3045 3046 // Switch to a new source space and find the first non-empty region. 3047 unsigned int space_id = src_space_id + 1; 3048 assert(space_id < last_space_id, "not enough spaces"); 3049 3050 HeapWord* const destination = closure.destination(); 3051 3052 do { 3053 MutableSpace* space = _space_info[space_id].space(); 3054 HeapWord* const bottom = space->bottom(); 3055 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3056 3057 // Iterate over the spaces that do not compact into themselves. 3058 if (bottom_cp->destination() != bottom) { 3059 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3060 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3061 3062 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3063 if (src_cp->live_obj_size() > 0) { 3064 // Found it. 3065 assert(src_cp->destination() == destination, 3066 "first live obj in the space must match the destination"); 3067 assert(src_cp->partial_obj_size() == 0, 3068 "a space cannot begin with a partial obj"); 3069 3070 src_space_id = SpaceId(space_id); 3071 src_space_top = space->top(); 3072 const size_t src_region_idx = sd.region(src_cp); 3073 closure.set_source(sd.region_to_addr(src_region_idx)); 3074 return src_region_idx; 3075 } else { 3076 assert(src_cp->data_size() == 0, "sanity"); 3077 } 3078 } 3079 } 3080 } while (++space_id < last_space_id); 3081 3082 assert(false, "no source region was found"); 3083 return 0; 3084 } 3085 3086 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3087 { 3088 typedef ParMarkBitMap::IterationStatus IterationStatus; 3089 const size_t RegionSize = ParallelCompactData::RegionSize; 3090 ParMarkBitMap* const bitmap = mark_bitmap(); 3091 ParallelCompactData& sd = summary_data(); 3092 RegionData* const region_ptr = sd.region(region_idx); 3093 3094 // Get the items needed to construct the closure. 3095 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3096 SpaceId dest_space_id = space_id(dest_addr); 3097 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3098 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3099 assert(dest_addr < new_top, "sanity"); 3100 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3101 3102 // Get the source region and related info. 3103 size_t src_region_idx = region_ptr->source_region(); 3104 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3105 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3106 3107 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3108 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3109 3110 // Adjust src_region_idx to prepare for decrementing destination counts (the 3111 // destination count is not decremented when a region is copied to itself). 3112 if (src_region_idx == region_idx) { 3113 src_region_idx += 1; 3114 } 3115 3116 if (bitmap->is_unmarked(closure.source())) { 3117 // The first source word is in the middle of an object; copy the remainder 3118 // of the object or as much as will fit. The fact that pointer updates were 3119 // deferred will be noted when the object header is processed. 3120 HeapWord* const old_src_addr = closure.source(); 3121 closure.copy_partial_obj(); 3122 if (closure.is_full()) { 3123 decrement_destination_counts(cm, src_space_id, src_region_idx, 3124 closure.source()); 3125 region_ptr->set_deferred_obj_addr(NULL); 3126 region_ptr->set_completed(); 3127 return; 3128 } 3129 3130 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3131 if (sd.region_align_down(old_src_addr) != end_addr) { 3132 // The partial object was copied from more than one source region. 3133 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3134 3135 // Move to the next source region, possibly switching spaces as well. All 3136 // args except end_addr may be modified. 3137 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3138 end_addr); 3139 } 3140 } 3141 3142 do { 3143 HeapWord* const cur_addr = closure.source(); 3144 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3145 src_space_top); 3146 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3147 3148 if (status == ParMarkBitMap::incomplete) { 3149 // The last obj that starts in the source region does not end in the 3150 // region. 3151 assert(closure.source() < end_addr, "sanity"); 3152 HeapWord* const obj_beg = closure.source(); 3153 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3154 src_space_top); 3155 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3156 if (obj_end < range_end) { 3157 // The end was found; the entire object will fit. 3158 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3159 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3160 } else { 3161 // The end was not found; the object will not fit. 3162 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3163 status = ParMarkBitMap::would_overflow; 3164 } 3165 } 3166 3167 if (status == ParMarkBitMap::would_overflow) { 3168 // The last object did not fit. Note that interior oop updates were 3169 // deferred, then copy enough of the object to fill the region. 3170 region_ptr->set_deferred_obj_addr(closure.destination()); 3171 status = closure.copy_until_full(); // copies from closure.source() 3172 3173 decrement_destination_counts(cm, src_space_id, src_region_idx, 3174 closure.source()); 3175 region_ptr->set_completed(); 3176 return; 3177 } 3178 3179 if (status == ParMarkBitMap::full) { 3180 decrement_destination_counts(cm, src_space_id, src_region_idx, 3181 closure.source()); 3182 region_ptr->set_deferred_obj_addr(NULL); 3183 region_ptr->set_completed(); 3184 return; 3185 } 3186 3187 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3188 3189 // Move to the next source region, possibly switching spaces as well. All 3190 // args except end_addr may be modified. 3191 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3192 end_addr); 3193 } while (true); 3194 } 3195 3196 void PSParallelCompact::fill_blocks(size_t region_idx) 3197 { 3198 // Fill in the block table elements for the specified region. Each block 3199 // table element holds the number of live words in the region that are to the 3200 // left of the first object that starts in the block. Thus only blocks in 3201 // which an object starts need to be filled. 3202 // 3203 // The algorithm scans the section of the bitmap that corresponds to the 3204 // region, keeping a running total of the live words. When an object start is 3205 // found, if it's the first to start in the block that contains it, the 3206 // current total is written to the block table element. 3207 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3208 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3209 const size_t RegionSize = ParallelCompactData::RegionSize; 3210 3211 ParallelCompactData& sd = summary_data(); 3212 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3213 if (partial_obj_size >= RegionSize) { 3214 return; // No objects start in this region. 3215 } 3216 3217 // Ensure the first loop iteration decides that the block has changed. 3218 size_t cur_block = sd.block_count(); 3219 3220 const ParMarkBitMap* const bitmap = mark_bitmap(); 3221 3222 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3223 assert((size_t)1 << Log2BitsPerBlock == 3224 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3225 3226 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3227 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3228 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3229 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3230 while (beg_bit < range_end) { 3231 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3232 if (new_block != cur_block) { 3233 cur_block = new_block; 3234 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3235 } 3236 3237 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3238 if (end_bit < range_end - 1) { 3239 live_bits += end_bit - beg_bit + 1; 3240 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3241 } else { 3242 return; 3243 } 3244 } 3245 } 3246 3247 void 3248 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3249 const MutableSpace* sp = space(space_id); 3250 if (sp->is_empty()) { 3251 return; 3252 } 3253 3254 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3255 ParMarkBitMap* const bitmap = mark_bitmap(); 3256 HeapWord* const dp_addr = dense_prefix(space_id); 3257 HeapWord* beg_addr = sp->bottom(); 3258 HeapWord* end_addr = sp->top(); 3259 3260 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3261 3262 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3263 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3264 if (beg_region < dp_region) { 3265 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3266 } 3267 3268 // The destination of the first live object that starts in the region is one 3269 // past the end of the partial object entering the region (if any). 3270 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3271 HeapWord* const new_top = _space_info[space_id].new_top(); 3272 assert(new_top >= dest_addr, "bad new_top value"); 3273 const size_t words = pointer_delta(new_top, dest_addr); 3274 3275 if (words > 0) { 3276 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3277 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3278 3279 ParMarkBitMap::IterationStatus status; 3280 status = bitmap->iterate(&closure, dest_addr, end_addr); 3281 assert(status == ParMarkBitMap::full, "iteration not complete"); 3282 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3283 "live objects skipped because closure is full"); 3284 } 3285 } 3286 3287 jlong PSParallelCompact::millis_since_last_gc() { 3288 // We need a monotonically non-deccreasing time in ms but 3289 // os::javaTimeMillis() does not guarantee monotonicity. 3290 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3291 jlong ret_val = now - _time_of_last_gc; 3292 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3293 if (ret_val < 0) { 3294 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);) 3295 return 0; 3296 } 3297 return ret_val; 3298 } 3299 3300 void PSParallelCompact::reset_millis_since_last_gc() { 3301 // We need a monotonically non-deccreasing time in ms but 3302 // os::javaTimeMillis() does not guarantee monotonicity. 3303 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3304 } 3305 3306 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3307 { 3308 if (source() != destination()) { 3309 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3310 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3311 } 3312 update_state(words_remaining()); 3313 assert(is_full(), "sanity"); 3314 return ParMarkBitMap::full; 3315 } 3316 3317 void MoveAndUpdateClosure::copy_partial_obj() 3318 { 3319 size_t words = words_remaining(); 3320 3321 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3322 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3323 if (end_addr < range_end) { 3324 words = bitmap()->obj_size(source(), end_addr); 3325 } 3326 3327 // This test is necessary; if omitted, the pointer updates to a partial object 3328 // that crosses the dense prefix boundary could be overwritten. 3329 if (source() != destination()) { 3330 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3331 Copy::aligned_conjoint_words(source(), destination(), words); 3332 } 3333 update_state(words); 3334 } 3335 3336 ParMarkBitMapClosure::IterationStatus 3337 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3338 assert(destination() != NULL, "sanity"); 3339 assert(bitmap()->obj_size(addr) == words, "bad size"); 3340 3341 _source = addr; 3342 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3343 destination(), "wrong destination"); 3344 3345 if (words > words_remaining()) { 3346 return ParMarkBitMap::would_overflow; 3347 } 3348 3349 // The start_array must be updated even if the object is not moving. 3350 if (_start_array != NULL) { 3351 _start_array->allocate_block(destination()); 3352 } 3353 3354 if (destination() != source()) { 3355 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3356 Copy::aligned_conjoint_words(source(), destination(), words); 3357 } 3358 3359 oop moved_oop = (oop) destination(); 3360 moved_oop->update_contents(compaction_manager()); 3361 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3362 3363 update_state(words); 3364 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3365 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3366 } 3367 3368 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3369 ParCompactionManager* cm, 3370 PSParallelCompact::SpaceId space_id) : 3371 ParMarkBitMapClosure(mbm, cm), 3372 _space_id(space_id), 3373 _start_array(PSParallelCompact::start_array(space_id)) 3374 { 3375 } 3376 3377 // Updates the references in the object to their new values. 3378 ParMarkBitMapClosure::IterationStatus 3379 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3380 do_addr(addr); 3381 return ParMarkBitMap::incomplete; 3382 }