1 /* 2 * Copyright (c) 2005, 2015, 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/stringTable.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_implementation/shared/spaceDecorator.hpp" 47 #include "gc_interface/gcCause.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/referencePolicy.hpp" 50 #include "memory/referenceProcessor.hpp" 51 #include "oops/methodData.hpp" 52 #include "oops/oop.inline.hpp" 53 #include "oops/oop.pcgc.inline.hpp" 54 #include "runtime/atomic.inline.hpp" 55 #include "runtime/fprofiler.hpp" 56 #include "runtime/safepoint.hpp" 57 #include "runtime/vmThread.hpp" 58 #include "services/management.hpp" 59 #include "services/memoryService.hpp" 60 #include "services/memTracker.hpp" 61 #include "utilities/events.hpp" 62 #include "utilities/stack.inline.hpp" 63 64 #include <math.h> 65 66 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 67 68 // All sizes are in HeapWords. 69 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 70 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 71 const size_t ParallelCompactData::RegionSizeBytes = 72 RegionSize << LogHeapWordSize; 73 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 74 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 75 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 76 77 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 78 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 79 const size_t ParallelCompactData::BlockSizeBytes = 80 BlockSize << LogHeapWordSize; 81 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 82 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 83 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 84 85 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 86 const size_t ParallelCompactData::Log2BlocksPerRegion = 87 Log2RegionSize - Log2BlockSize; 88 89 const ParallelCompactData::RegionData::region_sz_t 90 ParallelCompactData::RegionData::dc_shift = 27; 91 92 const ParallelCompactData::RegionData::region_sz_t 93 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 94 95 const ParallelCompactData::RegionData::region_sz_t 96 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 97 98 const ParallelCompactData::RegionData::region_sz_t 99 ParallelCompactData::RegionData::los_mask = ~dc_mask; 100 101 const ParallelCompactData::RegionData::region_sz_t 102 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 103 104 const ParallelCompactData::RegionData::region_sz_t 105 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 106 107 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 108 bool PSParallelCompact::_print_phases = false; 109 110 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 111 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL; 112 113 double PSParallelCompact::_dwl_mean; 114 double PSParallelCompact::_dwl_std_dev; 115 double PSParallelCompact::_dwl_first_term; 116 double PSParallelCompact::_dwl_adjustment; 117 #ifdef ASSERT 118 bool PSParallelCompact::_dwl_initialized = false; 119 #endif // #ifdef ASSERT 120 121 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 122 HeapWord* destination) 123 { 124 assert(src_region_idx != 0, "invalid src_region_idx"); 125 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 126 assert(destination != NULL, "invalid destination argument"); 127 128 _src_region_idx = src_region_idx; 129 _partial_obj_size = partial_obj_size; 130 _destination = destination; 131 132 // These fields may not be updated below, so make sure they're clear. 133 assert(_dest_region_addr == NULL, "should have been cleared"); 134 assert(_first_src_addr == NULL, "should have been cleared"); 135 136 // Determine the number of destination regions for the partial object. 137 HeapWord* const last_word = destination + partial_obj_size - 1; 138 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 139 HeapWord* const beg_region_addr = sd.region_align_down(destination); 140 HeapWord* const end_region_addr = sd.region_align_down(last_word); 141 142 if (beg_region_addr == end_region_addr) { 143 // One destination region. 144 _destination_count = 1; 145 if (end_region_addr == destination) { 146 // The destination falls on a region boundary, thus the first word of the 147 // partial object will be the first word copied to the destination region. 148 _dest_region_addr = end_region_addr; 149 _first_src_addr = sd.region_to_addr(src_region_idx); 150 } 151 } else { 152 // Two destination regions. When copied, the partial object will cross a 153 // destination region boundary, so a word somewhere within the partial 154 // object will be the first word copied to the second destination region. 155 _destination_count = 2; 156 _dest_region_addr = end_region_addr; 157 const size_t ofs = pointer_delta(end_region_addr, destination); 158 assert(ofs < _partial_obj_size, "sanity"); 159 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 160 } 161 } 162 163 void SplitInfo::clear() 164 { 165 _src_region_idx = 0; 166 _partial_obj_size = 0; 167 _destination = NULL; 168 _destination_count = 0; 169 _dest_region_addr = NULL; 170 _first_src_addr = NULL; 171 assert(!is_valid(), "sanity"); 172 } 173 174 #ifdef ASSERT 175 void SplitInfo::verify_clear() 176 { 177 assert(_src_region_idx == 0, "not clear"); 178 assert(_partial_obj_size == 0, "not clear"); 179 assert(_destination == NULL, "not clear"); 180 assert(_destination_count == 0, "not clear"); 181 assert(_dest_region_addr == NULL, "not clear"); 182 assert(_first_src_addr == NULL, "not clear"); 183 } 184 #endif // #ifdef ASSERT 185 186 187 void PSParallelCompact::print_on_error(outputStream* st) { 188 _mark_bitmap.print_on_error(st); 189 } 190 191 #ifndef PRODUCT 192 const char* PSParallelCompact::space_names[] = { 193 "old ", "eden", "from", "to " 194 }; 195 196 void PSParallelCompact::print_region_ranges() 197 { 198 tty->print_cr("space bottom top end new_top"); 199 tty->print_cr("------ ---------- ---------- ---------- ----------"); 200 201 for (unsigned int id = 0; id < last_space_id; ++id) { 202 const MutableSpace* space = _space_info[id].space(); 203 tty->print_cr("%u %s " 204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 205 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 206 id, space_names[id], 207 summary_data().addr_to_region_idx(space->bottom()), 208 summary_data().addr_to_region_idx(space->top()), 209 summary_data().addr_to_region_idx(space->end()), 210 summary_data().addr_to_region_idx(_space_info[id].new_top())); 211 } 212 } 213 214 void 215 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 216 { 217 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 218 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 219 220 ParallelCompactData& sd = PSParallelCompact::summary_data(); 221 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 222 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 223 REGION_IDX_FORMAT " " PTR_FORMAT " " 224 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 225 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 226 i, c->data_location(), dci, c->destination(), 227 c->partial_obj_size(), c->live_obj_size(), 228 c->data_size(), c->source_region(), c->destination_count()); 229 230 #undef REGION_IDX_FORMAT 231 #undef REGION_DATA_FORMAT 232 } 233 234 void 235 print_generic_summary_data(ParallelCompactData& summary_data, 236 HeapWord* const beg_addr, 237 HeapWord* const end_addr) 238 { 239 size_t total_words = 0; 240 size_t i = summary_data.addr_to_region_idx(beg_addr); 241 const size_t last = summary_data.addr_to_region_idx(end_addr); 242 HeapWord* pdest = 0; 243 244 while (i <= last) { 245 ParallelCompactData::RegionData* c = summary_data.region(i); 246 if (c->data_size() != 0 || c->destination() != pdest) { 247 print_generic_summary_region(i, c); 248 total_words += c->data_size(); 249 pdest = c->destination(); 250 } 251 ++i; 252 } 253 254 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 255 } 256 257 void 258 print_generic_summary_data(ParallelCompactData& summary_data, 259 SpaceInfo* space_info) 260 { 261 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 262 const MutableSpace* space = space_info[id].space(); 263 print_generic_summary_data(summary_data, space->bottom(), 264 MAX2(space->top(), space_info[id].new_top())); 265 } 266 } 267 268 void 269 print_initial_summary_region(size_t i, 270 const ParallelCompactData::RegionData* c, 271 bool newline = true) 272 { 273 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 275 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 276 i, c->destination(), 277 c->partial_obj_size(), c->live_obj_size(), 278 c->data_size(), c->source_region(), c->destination_count()); 279 if (newline) tty->cr(); 280 } 281 282 void 283 print_initial_summary_data(ParallelCompactData& summary_data, 284 const MutableSpace* space) { 285 if (space->top() == space->bottom()) { 286 return; 287 } 288 289 const size_t region_size = ParallelCompactData::RegionSize; 290 typedef ParallelCompactData::RegionData RegionData; 291 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 292 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 293 const RegionData* c = summary_data.region(end_region - 1); 294 HeapWord* end_addr = c->destination() + c->data_size(); 295 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 296 297 // Print (and count) the full regions at the beginning of the space. 298 size_t full_region_count = 0; 299 size_t i = summary_data.addr_to_region_idx(space->bottom()); 300 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 301 print_initial_summary_region(i, summary_data.region(i)); 302 ++full_region_count; 303 ++i; 304 } 305 306 size_t live_to_right = live_in_space - full_region_count * region_size; 307 308 double max_reclaimed_ratio = 0.0; 309 size_t max_reclaimed_ratio_region = 0; 310 size_t max_dead_to_right = 0; 311 size_t max_live_to_right = 0; 312 313 // Print the 'reclaimed ratio' for regions while there is something live in 314 // the region or to the right of it. The remaining regions are empty (and 315 // uninteresting), and computing the ratio will result in division by 0. 316 while (i < end_region && live_to_right > 0) { 317 c = summary_data.region(i); 318 HeapWord* const region_addr = summary_data.region_to_addr(i); 319 const size_t used_to_right = pointer_delta(space->top(), region_addr); 320 const size_t dead_to_right = used_to_right - live_to_right; 321 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 322 323 if (reclaimed_ratio > max_reclaimed_ratio) { 324 max_reclaimed_ratio = reclaimed_ratio; 325 max_reclaimed_ratio_region = i; 326 max_dead_to_right = dead_to_right; 327 max_live_to_right = live_to_right; 328 } 329 330 print_initial_summary_region(i, c, false); 331 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 332 reclaimed_ratio, dead_to_right, live_to_right); 333 334 live_to_right -= c->data_size(); 335 ++i; 336 } 337 338 // Any remaining regions are empty. Print one more if there is one. 339 if (i < end_region) { 340 print_initial_summary_region(i, summary_data.region(i)); 341 } 342 343 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 344 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 345 max_reclaimed_ratio_region, max_dead_to_right, 346 max_live_to_right, max_reclaimed_ratio); 347 } 348 349 void 350 print_initial_summary_data(ParallelCompactData& summary_data, 351 SpaceInfo* space_info) { 352 unsigned int id = PSParallelCompact::old_space_id; 353 const MutableSpace* space; 354 do { 355 space = space_info[id].space(); 356 print_initial_summary_data(summary_data, space); 357 } while (++id < PSParallelCompact::eden_space_id); 358 359 do { 360 space = space_info[id].space(); 361 print_generic_summary_data(summary_data, space->bottom(), space->top()); 362 } while (++id < PSParallelCompact::last_space_id); 363 } 364 #endif // #ifndef PRODUCT 365 366 #ifdef ASSERT 367 size_t add_obj_count; 368 size_t add_obj_size; 369 size_t mark_bitmap_count; 370 size_t mark_bitmap_size; 371 #endif // #ifdef ASSERT 372 373 ParallelCompactData::ParallelCompactData() 374 { 375 _region_start = 0; 376 377 _region_vspace = 0; 378 _reserved_byte_size = 0; 379 _region_data = 0; 380 _region_count = 0; 381 382 _block_vspace = 0; 383 _block_data = 0; 384 _block_count = 0; 385 } 386 387 bool ParallelCompactData::initialize(MemRegion covered_region) 388 { 389 _region_start = covered_region.start(); 390 const size_t region_size = covered_region.word_size(); 391 DEBUG_ONLY(_region_end = _region_start + region_size;) 392 393 assert(region_align_down(_region_start) == _region_start, 394 "region start not aligned"); 395 assert((region_size & RegionSizeOffsetMask) == 0, 396 "region size not a multiple of RegionSize"); 397 398 bool result = initialize_region_data(region_size) && initialize_block_data(); 399 return result; 400 } 401 402 PSVirtualSpace* 403 ParallelCompactData::create_vspace(size_t count, size_t element_size) 404 { 405 const size_t raw_bytes = count * element_size; 406 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 407 const size_t granularity = os::vm_allocation_granularity(); 408 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 409 410 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 411 MAX2(page_sz, granularity); 412 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 413 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 414 rs.size()); 415 416 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 417 418 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 419 if (vspace != 0) { 420 if (vspace->expand_by(_reserved_byte_size)) { 421 return vspace; 422 } 423 delete vspace; 424 // Release memory reserved in the space. 425 rs.release(); 426 } 427 428 return 0; 429 } 430 431 bool ParallelCompactData::initialize_region_data(size_t region_size) 432 { 433 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 434 _region_vspace = create_vspace(count, sizeof(RegionData)); 435 if (_region_vspace != 0) { 436 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 437 _region_count = count; 438 return true; 439 } 440 return false; 441 } 442 443 bool ParallelCompactData::initialize_block_data() 444 { 445 assert(_region_count != 0, "region data must be initialized first"); 446 const size_t count = _region_count << Log2BlocksPerRegion; 447 _block_vspace = create_vspace(count, sizeof(BlockData)); 448 if (_block_vspace != 0) { 449 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 450 _block_count = count; 451 return true; 452 } 453 return false; 454 } 455 456 void ParallelCompactData::clear() 457 { 458 memset(_region_data, 0, _region_vspace->committed_size()); 459 memset(_block_data, 0, _block_vspace->committed_size()); 460 } 461 462 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 463 assert(beg_region <= _region_count, "beg_region out of range"); 464 assert(end_region <= _region_count, "end_region out of range"); 465 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 466 467 const size_t region_cnt = end_region - beg_region; 468 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 469 470 const size_t beg_block = beg_region * BlocksPerRegion; 471 const size_t block_cnt = region_cnt * BlocksPerRegion; 472 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 473 } 474 475 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 476 { 477 const RegionData* cur_cp = region(region_idx); 478 const RegionData* const end_cp = region(region_count() - 1); 479 480 HeapWord* result = region_to_addr(region_idx); 481 if (cur_cp < end_cp) { 482 do { 483 result += cur_cp->partial_obj_size(); 484 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 485 } 486 return result; 487 } 488 489 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 490 { 491 const size_t obj_ofs = pointer_delta(addr, _region_start); 492 const size_t beg_region = obj_ofs >> Log2RegionSize; 493 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 494 495 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 496 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 497 498 if (beg_region == end_region) { 499 // All in one region. 500 _region_data[beg_region].add_live_obj(len); 501 return; 502 } 503 504 // First region. 505 const size_t beg_ofs = region_offset(addr); 506 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 507 508 Klass* klass = ((oop)addr)->klass(); 509 // Middle regions--completely spanned by this object. 510 for (size_t region = beg_region + 1; region < end_region; ++region) { 511 _region_data[region].set_partial_obj_size(RegionSize); 512 _region_data[region].set_partial_obj_addr(addr); 513 } 514 515 // Last region. 516 const size_t end_ofs = region_offset(addr + len - 1); 517 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 518 _region_data[end_region].set_partial_obj_addr(addr); 519 } 520 521 void 522 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 523 { 524 assert(region_offset(beg) == 0, "not RegionSize aligned"); 525 assert(region_offset(end) == 0, "not RegionSize aligned"); 526 527 size_t cur_region = addr_to_region_idx(beg); 528 const size_t end_region = addr_to_region_idx(end); 529 HeapWord* addr = beg; 530 while (cur_region < end_region) { 531 _region_data[cur_region].set_destination(addr); 532 _region_data[cur_region].set_destination_count(0); 533 _region_data[cur_region].set_source_region(cur_region); 534 _region_data[cur_region].set_data_location(addr); 535 536 // Update live_obj_size so the region appears completely full. 537 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 538 _region_data[cur_region].set_live_obj_size(live_size); 539 540 ++cur_region; 541 addr += RegionSize; 542 } 543 } 544 545 // Find the point at which a space can be split and, if necessary, record the 546 // split point. 547 // 548 // If the current src region (which overflowed the destination space) doesn't 549 // have a partial object, the split point is at the beginning of the current src 550 // region (an "easy" split, no extra bookkeeping required). 551 // 552 // If the current src region has a partial object, the split point is in the 553 // region where that partial object starts (call it the split_region). If 554 // split_region has a partial object, then the split point is just after that 555 // partial object (a "hard" split where we have to record the split data and 556 // zero the partial_obj_size field). With a "hard" split, we know that the 557 // partial_obj ends within split_region because the partial object that caused 558 // the overflow starts in split_region. If split_region doesn't have a partial 559 // obj, then the split is at the beginning of split_region (another "easy" 560 // split). 561 HeapWord* 562 ParallelCompactData::summarize_split_space(size_t src_region, 563 SplitInfo& split_info, 564 HeapWord* destination, 565 HeapWord* target_end, 566 HeapWord** target_next) 567 { 568 assert(destination <= target_end, "sanity"); 569 assert(destination + _region_data[src_region].data_size() > target_end, 570 "region should not fit into target space"); 571 assert(is_region_aligned(target_end), "sanity"); 572 573 size_t split_region = src_region; 574 HeapWord* split_destination = destination; 575 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 576 577 if (destination + partial_obj_size > target_end) { 578 // The split point is just after the partial object (if any) in the 579 // src_region that contains the start of the object that overflowed the 580 // destination space. 581 // 582 // Find the start of the "overflow" object and set split_region to the 583 // region containing it. 584 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 585 split_region = addr_to_region_idx(overflow_obj); 586 587 // Clear the source_region field of all destination regions whose first word 588 // came from data after the split point (a non-null source_region field 589 // implies a region must be filled). 590 // 591 // An alternative to the simple loop below: clear during post_compact(), 592 // which uses memcpy instead of individual stores, and is easy to 593 // parallelize. (The downside is that it clears the entire RegionData 594 // object as opposed to just one field.) 595 // 596 // post_compact() would have to clear the summary data up to the highest 597 // address that was written during the summary phase, which would be 598 // 599 // max(top, max(new_top, clear_top)) 600 // 601 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 602 // to target_end. 603 const RegionData* const sr = region(split_region); 604 const size_t beg_idx = 605 addr_to_region_idx(region_align_up(sr->destination() + 606 sr->partial_obj_size())); 607 const size_t end_idx = addr_to_region_idx(target_end); 608 609 if (TraceParallelOldGCSummaryPhase) { 610 gclog_or_tty->print_cr("split: clearing source_region field in [" 611 SIZE_FORMAT ", " SIZE_FORMAT ")", 612 beg_idx, end_idx); 613 } 614 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 615 _region_data[idx].set_source_region(0); 616 } 617 618 // Set split_destination and partial_obj_size to reflect the split region. 619 split_destination = sr->destination(); 620 partial_obj_size = sr->partial_obj_size(); 621 } 622 623 // The split is recorded only if a partial object extends onto the region. 624 if (partial_obj_size != 0) { 625 _region_data[split_region].set_partial_obj_size(0); 626 split_info.record(split_region, partial_obj_size, split_destination); 627 } 628 629 // Setup the continuation addresses. 630 *target_next = split_destination + partial_obj_size; 631 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 632 633 if (TraceParallelOldGCSummaryPhase) { 634 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 635 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 636 " pos=" SIZE_FORMAT, 637 split_type, source_next, split_region, 638 partial_obj_size); 639 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 640 " tn=" PTR_FORMAT, 641 split_type, split_destination, 642 addr_to_region_idx(split_destination), 643 *target_next); 644 645 if (partial_obj_size != 0) { 646 HeapWord* const po_beg = split_info.destination(); 647 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 648 gclog_or_tty->print_cr("%s split: " 649 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 650 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 651 split_type, 652 po_beg, addr_to_region_idx(po_beg), 653 po_end, addr_to_region_idx(po_end)); 654 } 655 } 656 657 return source_next; 658 } 659 660 bool ParallelCompactData::summarize(SplitInfo& split_info, 661 HeapWord* source_beg, HeapWord* source_end, 662 HeapWord** source_next, 663 HeapWord* target_beg, HeapWord* target_end, 664 HeapWord** target_next) 665 { 666 if (TraceParallelOldGCSummaryPhase) { 667 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 668 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 669 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 670 source_beg, source_end, source_next_val, 671 target_beg, target_end, *target_next); 672 } 673 674 size_t cur_region = addr_to_region_idx(source_beg); 675 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 676 677 HeapWord *dest_addr = target_beg; 678 while (cur_region < end_region) { 679 // The destination must be set even if the region has no data. 680 _region_data[cur_region].set_destination(dest_addr); 681 682 size_t words = _region_data[cur_region].data_size(); 683 if (words > 0) { 684 // If cur_region does not fit entirely into the target space, find a point 685 // at which the source space can be 'split' so that part is copied to the 686 // target space and the rest is copied elsewhere. 687 if (dest_addr + words > target_end) { 688 assert(source_next != NULL, "source_next is NULL when splitting"); 689 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 690 target_end, target_next); 691 return false; 692 } 693 694 // Compute the destination_count for cur_region, and if necessary, update 695 // source_region for a destination region. The source_region field is 696 // updated if cur_region is the first (left-most) region to be copied to a 697 // destination region. 698 // 699 // The destination_count calculation is a bit subtle. A region that has 700 // data that compacts into itself does not count itself as a destination. 701 // This maintains the invariant that a zero count means the region is 702 // available and can be claimed and then filled. 703 uint destination_count = 0; 704 if (split_info.is_split(cur_region)) { 705 // The current region has been split: the partial object will be copied 706 // to one destination space and the remaining data will be copied to 707 // another destination space. Adjust the initial destination_count and, 708 // if necessary, set the source_region field if the partial object will 709 // cross a destination region boundary. 710 destination_count = split_info.destination_count(); 711 if (destination_count == 2) { 712 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 713 _region_data[dest_idx].set_source_region(cur_region); 714 } 715 } 716 717 HeapWord* const last_addr = dest_addr + words - 1; 718 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 719 const size_t dest_region_2 = addr_to_region_idx(last_addr); 720 721 // Initially assume that the destination regions will be the same and 722 // adjust the value below if necessary. Under this assumption, if 723 // cur_region == dest_region_2, then cur_region will be compacted 724 // completely into itself. 725 destination_count += cur_region == dest_region_2 ? 0 : 1; 726 if (dest_region_1 != dest_region_2) { 727 // Destination regions differ; adjust destination_count. 728 destination_count += 1; 729 // Data from cur_region will be copied to the start of dest_region_2. 730 _region_data[dest_region_2].set_source_region(cur_region); 731 } else if (region_offset(dest_addr) == 0) { 732 // Data from cur_region will be copied to the start of the destination 733 // region. 734 _region_data[dest_region_1].set_source_region(cur_region); 735 } 736 737 _region_data[cur_region].set_destination_count(destination_count); 738 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 739 dest_addr += words; 740 } 741 742 ++cur_region; 743 } 744 745 *target_next = dest_addr; 746 return true; 747 } 748 749 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 750 assert(addr != NULL, "Should detect NULL oop earlier"); 751 assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap"); 752 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 753 754 // Region covering the object. 755 RegionData* const region_ptr = addr_to_region_ptr(addr); 756 HeapWord* result = region_ptr->destination(); 757 758 // If the entire Region is live, the new location is region->destination + the 759 // offset of the object within in the Region. 760 761 // Run some performance tests to determine if this special case pays off. It 762 // is worth it for pointers into the dense prefix. If the optimization to 763 // avoid pointer updates in regions that only point to the dense prefix is 764 // ever implemented, this should be revisited. 765 if (region_ptr->data_size() == RegionSize) { 766 result += region_offset(addr); 767 return result; 768 } 769 770 // Otherwise, the new location is region->destination + block offset + the 771 // number of live words in the Block that are (a) to the left of addr and (b) 772 // due to objects that start in the Block. 773 774 // Fill in the block table if necessary. This is unsynchronized, so multiple 775 // threads may fill the block table for a region (harmless, since it is 776 // idempotent). 777 if (!region_ptr->blocks_filled()) { 778 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 779 region_ptr->set_blocks_filled(); 780 } 781 782 HeapWord* const search_start = block_align_down(addr); 783 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 784 785 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 786 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 787 result += block_offset + live; 788 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 789 return result; 790 } 791 792 #ifdef ASSERT 793 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 794 { 795 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 796 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 797 for (const size_t* p = beg; p < end; ++p) { 798 assert(*p == 0, "not zero"); 799 } 800 } 801 802 void ParallelCompactData::verify_clear() 803 { 804 verify_clear(_region_vspace); 805 verify_clear(_block_vspace); 806 } 807 #endif // #ifdef ASSERT 808 809 STWGCTimer PSParallelCompact::_gc_timer; 810 ParallelOldTracer PSParallelCompact::_gc_tracer; 811 elapsedTimer PSParallelCompact::_accumulated_time; 812 unsigned int PSParallelCompact::_total_invocations = 0; 813 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 814 jlong PSParallelCompact::_time_of_last_gc = 0; 815 CollectorCounters* PSParallelCompact::_counters = NULL; 816 ParMarkBitMap PSParallelCompact::_mark_bitmap; 817 ParallelCompactData PSParallelCompact::_summary_data; 818 819 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 820 821 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 822 823 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 824 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 825 826 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 827 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 828 829 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); } 830 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); } 831 832 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 833 834 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { 835 mark_and_push(_compaction_manager, p); 836 } 837 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 838 839 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 840 klass->oops_do(_mark_and_push_closure); 841 } 842 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 843 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 844 } 845 846 void PSParallelCompact::post_initialize() { 847 ParallelScavengeHeap* heap = gc_heap(); 848 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 849 850 MemRegion mr = heap->reserved_region(); 851 _ref_processor = 852 new ReferenceProcessor(mr, // span 853 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 854 (int) ParallelGCThreads, // mt processing degree 855 true, // mt discovery 856 (int) ParallelGCThreads, // mt discovery degree 857 true, // atomic_discovery 858 &_is_alive_closure); // non-header is alive closure 859 _counters = new CollectorCounters("PSParallelCompact", 1); 860 861 // Initialize static fields in ParCompactionManager. 862 ParCompactionManager::initialize(mark_bitmap()); 863 } 864 865 bool PSParallelCompact::initialize() { 866 ParallelScavengeHeap* heap = gc_heap(); 867 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 868 MemRegion mr = heap->reserved_region(); 869 870 // Was the old gen get allocated successfully? 871 if (!heap->old_gen()->is_allocated()) { 872 return false; 873 } 874 875 initialize_space_info(); 876 initialize_dead_wood_limiter(); 877 878 if (!_mark_bitmap.initialize(mr)) { 879 vm_shutdown_during_initialization( 880 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 881 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 882 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 883 return false; 884 } 885 886 if (!_summary_data.initialize(mr)) { 887 vm_shutdown_during_initialization( 888 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 889 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 890 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 891 return false; 892 } 893 894 return true; 895 } 896 897 void PSParallelCompact::initialize_space_info() 898 { 899 memset(&_space_info, 0, sizeof(_space_info)); 900 901 ParallelScavengeHeap* heap = gc_heap(); 902 PSYoungGen* young_gen = heap->young_gen(); 903 904 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 905 _space_info[eden_space_id].set_space(young_gen->eden_space()); 906 _space_info[from_space_id].set_space(young_gen->from_space()); 907 _space_info[to_space_id].set_space(young_gen->to_space()); 908 909 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 910 } 911 912 void PSParallelCompact::initialize_dead_wood_limiter() 913 { 914 const size_t max = 100; 915 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 916 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 917 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 918 DEBUG_ONLY(_dwl_initialized = true;) 919 _dwl_adjustment = normal_distribution(1.0); 920 } 921 922 // Simple class for storing info about the heap at the start of GC, to be used 923 // after GC for comparison/printing. 924 class PreGCValues { 925 public: 926 PreGCValues() { } 927 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 928 929 void fill(ParallelScavengeHeap* heap) { 930 _heap_used = heap->used(); 931 _young_gen_used = heap->young_gen()->used_in_bytes(); 932 _old_gen_used = heap->old_gen()->used_in_bytes(); 933 _metadata_used = MetaspaceAux::used_bytes(); 934 }; 935 936 size_t heap_used() const { return _heap_used; } 937 size_t young_gen_used() const { return _young_gen_used; } 938 size_t old_gen_used() const { return _old_gen_used; } 939 size_t metadata_used() const { return _metadata_used; } 940 941 private: 942 size_t _heap_used; 943 size_t _young_gen_used; 944 size_t _old_gen_used; 945 size_t _metadata_used; 946 }; 947 948 void 949 PSParallelCompact::clear_data_covering_space(SpaceId id) 950 { 951 // At this point, top is the value before GC, new_top() is the value that will 952 // be set at the end of GC. The marking bitmap is cleared to top; nothing 953 // should be marked above top. The summary data is cleared to the larger of 954 // top & new_top. 955 MutableSpace* const space = _space_info[id].space(); 956 HeapWord* const bot = space->bottom(); 957 HeapWord* const top = space->top(); 958 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 959 960 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 961 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 962 _mark_bitmap.clear_range(beg_bit, end_bit); 963 964 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 965 const size_t end_region = 966 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 967 _summary_data.clear_range(beg_region, end_region); 968 969 // Clear the data used to 'split' regions. 970 SplitInfo& split_info = _space_info[id].split_info(); 971 if (split_info.is_valid()) { 972 split_info.clear(); 973 } 974 DEBUG_ONLY(split_info.verify_clear();) 975 } 976 977 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 978 { 979 // Update the from & to space pointers in space_info, since they are swapped 980 // at each young gen gc. Do the update unconditionally (even though a 981 // promotion failure does not swap spaces) because an unknown number of minor 982 // collections will have swapped the spaces an unknown number of times. 983 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 984 ParallelScavengeHeap* heap = gc_heap(); 985 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 986 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 987 988 pre_gc_values->fill(heap); 989 990 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 991 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 992 993 // Increment the invocation count 994 heap->increment_total_collections(true); 995 996 // We need to track unique mark sweep invocations as well. 997 _total_invocations++; 998 999 heap->print_heap_before_gc(); 1000 heap->trace_heap_before_gc(&_gc_tracer); 1001 1002 // Fill in TLABs 1003 heap->accumulate_statistics_all_tlabs(); 1004 heap->ensure_parsability(true); // retire TLABs 1005 1006 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1007 HandleMark hm; // Discard invalid handles created during verification 1008 Universe::verify(" VerifyBeforeGC:"); 1009 } 1010 1011 // Verify object start arrays 1012 if (VerifyObjectStartArray && 1013 VerifyBeforeGC) { 1014 heap->old_gen()->verify_object_start_array(); 1015 } 1016 1017 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1018 DEBUG_ONLY(summary_data().verify_clear();) 1019 1020 // Have worker threads release resources the next time they run a task. 1021 gc_task_manager()->release_all_resources(); 1022 } 1023 1024 void PSParallelCompact::post_compact() 1025 { 1026 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1027 1028 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1029 // Clear the marking bitmap, summary data and split info. 1030 clear_data_covering_space(SpaceId(id)); 1031 // Update top(). Must be done after clearing the bitmap and summary data. 1032 _space_info[id].publish_new_top(); 1033 } 1034 1035 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1036 MutableSpace* const from_space = _space_info[from_space_id].space(); 1037 MutableSpace* const to_space = _space_info[to_space_id].space(); 1038 1039 ParallelScavengeHeap* heap = gc_heap(); 1040 bool eden_empty = eden_space->is_empty(); 1041 if (!eden_empty) { 1042 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1043 heap->young_gen(), heap->old_gen()); 1044 } 1045 1046 // Update heap occupancy information which is used as input to the soft ref 1047 // clearing policy at the next gc. 1048 Universe::update_heap_info_at_gc(); 1049 1050 bool young_gen_empty = eden_empty && from_space->is_empty() && 1051 to_space->is_empty(); 1052 1053 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1054 MemRegion old_mr = heap->old_gen()->reserved(); 1055 if (young_gen_empty) { 1056 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1057 } else { 1058 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1059 } 1060 1061 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1062 ClassLoaderDataGraph::purge(); 1063 MetaspaceAux::verify_metrics(); 1064 1065 CodeCache::gc_epilogue(); 1066 JvmtiExport::gc_epilogue(); 1067 1068 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1069 1070 ref_processor()->enqueue_discovered_references(NULL); 1071 1072 if (ZapUnusedHeapArea) { 1073 heap->gen_mangle_unused_area(); 1074 } 1075 1076 // Update time of last GC 1077 reset_millis_since_last_gc(); 1078 } 1079 1080 HeapWord* 1081 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1082 bool maximum_compaction) 1083 { 1084 const size_t region_size = ParallelCompactData::RegionSize; 1085 const ParallelCompactData& sd = summary_data(); 1086 1087 const MutableSpace* const space = _space_info[id].space(); 1088 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1089 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1090 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1091 1092 // Skip full regions at the beginning of the space--they are necessarily part 1093 // of the dense prefix. 1094 size_t full_count = 0; 1095 const RegionData* cp; 1096 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1097 ++full_count; 1098 } 1099 1100 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1101 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1102 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1103 if (maximum_compaction || cp == end_cp || interval_ended) { 1104 _maximum_compaction_gc_num = total_invocations(); 1105 return sd.region_to_addr(cp); 1106 } 1107 1108 HeapWord* const new_top = _space_info[id].new_top(); 1109 const size_t space_live = pointer_delta(new_top, space->bottom()); 1110 const size_t space_used = space->used_in_words(); 1111 const size_t space_capacity = space->capacity_in_words(); 1112 1113 const double cur_density = double(space_live) / space_capacity; 1114 const double deadwood_density = 1115 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1116 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1117 1118 if (TraceParallelOldGCDensePrefix) { 1119 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1120 cur_density, deadwood_density, deadwood_goal); 1121 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1122 "space_cap=" SIZE_FORMAT, 1123 space_live, space_used, 1124 space_capacity); 1125 } 1126 1127 // XXX - Use binary search? 1128 HeapWord* dense_prefix = sd.region_to_addr(cp); 1129 const RegionData* full_cp = cp; 1130 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1131 while (cp < end_cp) { 1132 HeapWord* region_destination = cp->destination(); 1133 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1134 if (TraceParallelOldGCDensePrefix && Verbose) { 1135 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1136 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1137 sd.region(cp), region_destination, 1138 dense_prefix, cur_deadwood); 1139 } 1140 1141 if (cur_deadwood >= deadwood_goal) { 1142 // Found the region that has the correct amount of deadwood to the left. 1143 // This typically occurs after crossing a fairly sparse set of regions, so 1144 // iterate backwards over those sparse regions, looking for the region 1145 // that has the lowest density of live objects 'to the right.' 1146 size_t space_to_left = sd.region(cp) * region_size; 1147 size_t live_to_left = space_to_left - cur_deadwood; 1148 size_t space_to_right = space_capacity - space_to_left; 1149 size_t live_to_right = space_live - live_to_left; 1150 double density_to_right = double(live_to_right) / space_to_right; 1151 while (cp > full_cp) { 1152 --cp; 1153 const size_t prev_region_live_to_right = live_to_right - 1154 cp->data_size(); 1155 const size_t prev_region_space_to_right = space_to_right + region_size; 1156 double prev_region_density_to_right = 1157 double(prev_region_live_to_right) / prev_region_space_to_right; 1158 if (density_to_right <= prev_region_density_to_right) { 1159 return dense_prefix; 1160 } 1161 if (TraceParallelOldGCDensePrefix && Verbose) { 1162 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1163 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1164 prev_region_density_to_right); 1165 } 1166 dense_prefix -= region_size; 1167 live_to_right = prev_region_live_to_right; 1168 space_to_right = prev_region_space_to_right; 1169 density_to_right = prev_region_density_to_right; 1170 } 1171 return dense_prefix; 1172 } 1173 1174 dense_prefix += region_size; 1175 ++cp; 1176 } 1177 1178 return dense_prefix; 1179 } 1180 1181 #ifndef PRODUCT 1182 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1183 const SpaceId id, 1184 const bool maximum_compaction, 1185 HeapWord* const addr) 1186 { 1187 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1188 RegionData* const cp = summary_data().region(region_idx); 1189 const MutableSpace* const space = _space_info[id].space(); 1190 HeapWord* const new_top = _space_info[id].new_top(); 1191 1192 const size_t space_live = pointer_delta(new_top, space->bottom()); 1193 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1194 const size_t space_cap = space->capacity_in_words(); 1195 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1196 const size_t live_to_right = new_top - cp->destination(); 1197 const size_t dead_to_right = space->top() - addr - live_to_right; 1198 1199 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1200 "spl=" SIZE_FORMAT " " 1201 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1202 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1203 " ratio=%10.8f", 1204 algorithm, addr, region_idx, 1205 space_live, 1206 dead_to_left, dead_to_left_pct, 1207 dead_to_right, live_to_right, 1208 double(dead_to_right) / live_to_right); 1209 } 1210 #endif // #ifndef PRODUCT 1211 1212 // Return a fraction indicating how much of the generation can be treated as 1213 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1214 // based on the density of live objects in the generation to determine a limit, 1215 // which is then adjusted so the return value is min_percent when the density is 1216 // 1. 1217 // 1218 // The following table shows some return values for a different values of the 1219 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1220 // min_percent is 1. 1221 // 1222 // fraction allowed as dead wood 1223 // ----------------------------------------------------------------- 1224 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1225 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1226 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1227 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1228 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1229 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1230 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1231 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1232 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1233 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1234 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1235 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1236 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1237 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1238 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1239 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1240 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1241 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1242 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1243 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1244 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1245 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1246 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1247 1248 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1249 { 1250 assert(_dwl_initialized, "uninitialized"); 1251 1252 // The raw limit is the value of the normal distribution at x = density. 1253 const double raw_limit = normal_distribution(density); 1254 1255 // Adjust the raw limit so it becomes the minimum when the density is 1. 1256 // 1257 // First subtract the adjustment value (which is simply the precomputed value 1258 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1259 // Then add the minimum value, so the minimum is returned when the density is 1260 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1261 const double min = double(min_percent) / 100.0; 1262 const double limit = raw_limit - _dwl_adjustment + min; 1263 return MAX2(limit, 0.0); 1264 } 1265 1266 ParallelCompactData::RegionData* 1267 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1268 const RegionData* end) 1269 { 1270 const size_t region_size = ParallelCompactData::RegionSize; 1271 ParallelCompactData& sd = summary_data(); 1272 size_t left = sd.region(beg); 1273 size_t right = end > beg ? sd.region(end) - 1 : left; 1274 1275 // Binary search. 1276 while (left < right) { 1277 // Equivalent to (left + right) / 2, but does not overflow. 1278 const size_t middle = left + (right - left) / 2; 1279 RegionData* const middle_ptr = sd.region(middle); 1280 HeapWord* const dest = middle_ptr->destination(); 1281 HeapWord* const addr = sd.region_to_addr(middle); 1282 assert(dest != NULL, "sanity"); 1283 assert(dest <= addr, "must move left"); 1284 1285 if (middle > left && dest < addr) { 1286 right = middle - 1; 1287 } else if (middle < right && middle_ptr->data_size() == region_size) { 1288 left = middle + 1; 1289 } else { 1290 return middle_ptr; 1291 } 1292 } 1293 return sd.region(left); 1294 } 1295 1296 ParallelCompactData::RegionData* 1297 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1298 const RegionData* end, 1299 size_t dead_words) 1300 { 1301 ParallelCompactData& sd = summary_data(); 1302 size_t left = sd.region(beg); 1303 size_t right = end > beg ? sd.region(end) - 1 : left; 1304 1305 // Binary search. 1306 while (left < right) { 1307 // Equivalent to (left + right) / 2, but does not overflow. 1308 const size_t middle = left + (right - left) / 2; 1309 RegionData* const middle_ptr = sd.region(middle); 1310 HeapWord* const dest = middle_ptr->destination(); 1311 HeapWord* const addr = sd.region_to_addr(middle); 1312 assert(dest != NULL, "sanity"); 1313 assert(dest <= addr, "must move left"); 1314 1315 const size_t dead_to_left = pointer_delta(addr, dest); 1316 if (middle > left && dead_to_left > dead_words) { 1317 right = middle - 1; 1318 } else if (middle < right && dead_to_left < dead_words) { 1319 left = middle + 1; 1320 } else { 1321 return middle_ptr; 1322 } 1323 } 1324 return sd.region(left); 1325 } 1326 1327 // The result is valid during the summary phase, after the initial summarization 1328 // of each space into itself, and before final summarization. 1329 inline double 1330 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1331 HeapWord* const bottom, 1332 HeapWord* const top, 1333 HeapWord* const new_top) 1334 { 1335 ParallelCompactData& sd = summary_data(); 1336 1337 assert(cp != NULL, "sanity"); 1338 assert(bottom != NULL, "sanity"); 1339 assert(top != NULL, "sanity"); 1340 assert(new_top != NULL, "sanity"); 1341 assert(top >= new_top, "summary data problem?"); 1342 assert(new_top > bottom, "space is empty; should not be here"); 1343 assert(new_top >= cp->destination(), "sanity"); 1344 assert(top >= sd.region_to_addr(cp), "sanity"); 1345 1346 HeapWord* const destination = cp->destination(); 1347 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1348 const size_t compacted_region_live = pointer_delta(new_top, destination); 1349 const size_t compacted_region_used = pointer_delta(top, 1350 sd.region_to_addr(cp)); 1351 const size_t reclaimable = compacted_region_used - compacted_region_live; 1352 1353 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1354 return double(reclaimable) / divisor; 1355 } 1356 1357 // Return the address of the end of the dense prefix, a.k.a. the start of the 1358 // compacted region. The address is always on a region boundary. 1359 // 1360 // Completely full regions at the left are skipped, since no compaction can 1361 // occur in those regions. Then the maximum amount of dead wood to allow is 1362 // computed, based on the density (amount live / capacity) of the generation; 1363 // the region with approximately that amount of dead space to the left is 1364 // identified as the limit region. Regions between the last completely full 1365 // region and the limit region are scanned and the one that has the best 1366 // (maximum) reclaimed_ratio() is selected. 1367 HeapWord* 1368 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1369 bool maximum_compaction) 1370 { 1371 if (ParallelOldGCSplitALot) { 1372 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1373 // The value was chosen to provoke splitting a young gen space; use it. 1374 return _space_info[id].dense_prefix(); 1375 } 1376 } 1377 1378 const size_t region_size = ParallelCompactData::RegionSize; 1379 const ParallelCompactData& sd = summary_data(); 1380 1381 const MutableSpace* const space = _space_info[id].space(); 1382 HeapWord* const top = space->top(); 1383 HeapWord* const top_aligned_up = sd.region_align_up(top); 1384 HeapWord* const new_top = _space_info[id].new_top(); 1385 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1386 HeapWord* const bottom = space->bottom(); 1387 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1388 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1389 const RegionData* const new_top_cp = 1390 sd.addr_to_region_ptr(new_top_aligned_up); 1391 1392 // Skip full regions at the beginning of the space--they are necessarily part 1393 // of the dense prefix. 1394 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1395 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1396 space->is_empty(), "no dead space allowed to the left"); 1397 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1398 "region must have dead space"); 1399 1400 // The gc number is saved whenever a maximum compaction is done, and used to 1401 // determine when the maximum compaction interval has expired. This avoids 1402 // successive max compactions for different reasons. 1403 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1404 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1405 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1406 total_invocations() == HeapFirstMaximumCompactionCount; 1407 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1408 _maximum_compaction_gc_num = total_invocations(); 1409 return sd.region_to_addr(full_cp); 1410 } 1411 1412 const size_t space_live = pointer_delta(new_top, bottom); 1413 const size_t space_used = space->used_in_words(); 1414 const size_t space_capacity = space->capacity_in_words(); 1415 1416 const double density = double(space_live) / double(space_capacity); 1417 const size_t min_percent_free = MarkSweepDeadRatio; 1418 const double limiter = dead_wood_limiter(density, min_percent_free); 1419 const size_t dead_wood_max = space_used - space_live; 1420 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1421 dead_wood_max); 1422 1423 if (TraceParallelOldGCDensePrefix) { 1424 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1425 "space_cap=" SIZE_FORMAT, 1426 space_live, space_used, 1427 space_capacity); 1428 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1429 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1430 density, min_percent_free, limiter, 1431 dead_wood_max, dead_wood_limit); 1432 } 1433 1434 // Locate the region with the desired amount of dead space to the left. 1435 const RegionData* const limit_cp = 1436 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1437 1438 // Scan from the first region with dead space to the limit region and find the 1439 // one with the best (largest) reclaimed ratio. 1440 double best_ratio = 0.0; 1441 const RegionData* best_cp = full_cp; 1442 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1443 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1444 if (tmp_ratio > best_ratio) { 1445 best_cp = cp; 1446 best_ratio = tmp_ratio; 1447 } 1448 } 1449 1450 #if 0 1451 // Something to consider: if the region with the best ratio is 'close to' the 1452 // first region w/free space, choose the first region with free space 1453 // ("first-free"). The first-free region is usually near the start of the 1454 // heap, which means we are copying most of the heap already, so copy a bit 1455 // more to get complete compaction. 1456 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1457 _maximum_compaction_gc_num = total_invocations(); 1458 best_cp = full_cp; 1459 } 1460 #endif // #if 0 1461 1462 return sd.region_to_addr(best_cp); 1463 } 1464 1465 #ifndef PRODUCT 1466 void 1467 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1468 size_t words) 1469 { 1470 if (TraceParallelOldGCSummaryPhase) { 1471 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1472 SIZE_FORMAT, start, start + words, words); 1473 } 1474 1475 ObjectStartArray* const start_array = _space_info[id].start_array(); 1476 CollectedHeap::fill_with_objects(start, words); 1477 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1478 _mark_bitmap.mark_obj(p, words); 1479 _summary_data.add_obj(p, words); 1480 start_array->allocate_block(p); 1481 } 1482 } 1483 1484 void 1485 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1486 { 1487 ParallelCompactData& sd = summary_data(); 1488 MutableSpace* space = _space_info[id].space(); 1489 1490 // Find the source and destination start addresses. 1491 HeapWord* const src_addr = sd.region_align_down(start); 1492 HeapWord* dst_addr; 1493 if (src_addr < start) { 1494 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1495 } else if (src_addr > space->bottom()) { 1496 // The start (the original top() value) is aligned to a region boundary so 1497 // the associated region does not have a destination. Compute the 1498 // destination from the previous region. 1499 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1500 dst_addr = cp->destination() + cp->data_size(); 1501 } else { 1502 // Filling the entire space. 1503 dst_addr = space->bottom(); 1504 } 1505 assert(dst_addr != NULL, "sanity"); 1506 1507 // Update the summary data. 1508 bool result = _summary_data.summarize(_space_info[id].split_info(), 1509 src_addr, space->top(), NULL, 1510 dst_addr, space->end(), 1511 _space_info[id].new_top_addr()); 1512 assert(result, "should not fail: bad filler object size"); 1513 } 1514 1515 void 1516 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1517 { 1518 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1519 return; 1520 } 1521 1522 MutableSpace* const space = _space_info[id].space(); 1523 if (space->is_empty()) { 1524 HeapWord* b = space->bottom(); 1525 HeapWord* t = b + space->capacity_in_words() / 2; 1526 space->set_top(t); 1527 if (ZapUnusedHeapArea) { 1528 space->set_top_for_allocations(); 1529 } 1530 1531 size_t min_size = CollectedHeap::min_fill_size(); 1532 size_t obj_len = min_size; 1533 while (b + obj_len <= t) { 1534 CollectedHeap::fill_with_object(b, obj_len); 1535 mark_bitmap()->mark_obj(b, obj_len); 1536 summary_data().add_obj(b, obj_len); 1537 b += obj_len; 1538 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1539 } 1540 if (b < t) { 1541 // The loop didn't completely fill to t (top); adjust top downward. 1542 space->set_top(b); 1543 if (ZapUnusedHeapArea) { 1544 space->set_top_for_allocations(); 1545 } 1546 } 1547 1548 HeapWord** nta = _space_info[id].new_top_addr(); 1549 bool result = summary_data().summarize(_space_info[id].split_info(), 1550 space->bottom(), space->top(), NULL, 1551 space->bottom(), space->end(), nta); 1552 assert(result, "space must fit into itself"); 1553 } 1554 } 1555 1556 void 1557 PSParallelCompact::provoke_split(bool & max_compaction) 1558 { 1559 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1560 return; 1561 } 1562 1563 const size_t region_size = ParallelCompactData::RegionSize; 1564 ParallelCompactData& sd = summary_data(); 1565 1566 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1567 MutableSpace* const from_space = _space_info[from_space_id].space(); 1568 const size_t eden_live = pointer_delta(eden_space->top(), 1569 _space_info[eden_space_id].new_top()); 1570 const size_t from_live = pointer_delta(from_space->top(), 1571 _space_info[from_space_id].new_top()); 1572 1573 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1574 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1575 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1576 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1577 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1578 1579 // Choose the space to split; need at least 2 regions live (or fillable). 1580 SpaceId id; 1581 MutableSpace* space; 1582 size_t live_words; 1583 size_t fill_words; 1584 if (eden_live + eden_fillable >= region_size * 2) { 1585 id = eden_space_id; 1586 space = eden_space; 1587 live_words = eden_live; 1588 fill_words = eden_fillable; 1589 } else if (from_live + from_fillable >= region_size * 2) { 1590 id = from_space_id; 1591 space = from_space; 1592 live_words = from_live; 1593 fill_words = from_fillable; 1594 } else { 1595 return; // Give up. 1596 } 1597 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1598 1599 if (live_words < region_size * 2) { 1600 // Fill from top() to end() w/live objects of mixed sizes. 1601 HeapWord* const fill_start = space->top(); 1602 live_words += fill_words; 1603 1604 space->set_top(fill_start + fill_words); 1605 if (ZapUnusedHeapArea) { 1606 space->set_top_for_allocations(); 1607 } 1608 1609 HeapWord* cur_addr = fill_start; 1610 while (fill_words > 0) { 1611 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1612 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1613 if (fill_words - cur_size < min_fill_size) { 1614 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1615 } 1616 1617 CollectedHeap::fill_with_object(cur_addr, cur_size); 1618 mark_bitmap()->mark_obj(cur_addr, cur_size); 1619 sd.add_obj(cur_addr, cur_size); 1620 1621 cur_addr += cur_size; 1622 fill_words -= cur_size; 1623 } 1624 1625 summarize_new_objects(id, fill_start); 1626 } 1627 1628 max_compaction = false; 1629 1630 // Manipulate the old gen so that it has room for about half of the live data 1631 // in the target young gen space (live_words / 2). 1632 id = old_space_id; 1633 space = _space_info[id].space(); 1634 const size_t free_at_end = space->free_in_words(); 1635 const size_t free_target = align_object_size(live_words / 2); 1636 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1637 1638 if (free_at_end >= free_target + min_fill_size) { 1639 // Fill space above top() and set the dense prefix so everything survives. 1640 HeapWord* const fill_start = space->top(); 1641 const size_t fill_size = free_at_end - free_target; 1642 space->set_top(space->top() + fill_size); 1643 if (ZapUnusedHeapArea) { 1644 space->set_top_for_allocations(); 1645 } 1646 fill_with_live_objects(id, fill_start, fill_size); 1647 summarize_new_objects(id, fill_start); 1648 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1649 } else if (dead + free_at_end > free_target) { 1650 // Find a dense prefix that makes the right amount of space available. 1651 HeapWord* cur = sd.region_align_down(space->top()); 1652 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1653 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1654 while (dead_to_right < free_target) { 1655 cur -= region_size; 1656 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1657 dead_to_right = pointer_delta(space->end(), cur_destination); 1658 } 1659 _space_info[id].set_dense_prefix(cur); 1660 } 1661 } 1662 #endif // #ifndef PRODUCT 1663 1664 void PSParallelCompact::summarize_spaces_quick() 1665 { 1666 for (unsigned int i = 0; i < last_space_id; ++i) { 1667 const MutableSpace* space = _space_info[i].space(); 1668 HeapWord** nta = _space_info[i].new_top_addr(); 1669 bool result = _summary_data.summarize(_space_info[i].split_info(), 1670 space->bottom(), space->top(), NULL, 1671 space->bottom(), space->end(), nta); 1672 assert(result, "space must fit into itself"); 1673 _space_info[i].set_dense_prefix(space->bottom()); 1674 } 1675 1676 #ifndef PRODUCT 1677 if (ParallelOldGCSplitALot) { 1678 provoke_split_fill_survivor(to_space_id); 1679 } 1680 #endif // #ifndef PRODUCT 1681 } 1682 1683 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1684 { 1685 HeapWord* const dense_prefix_end = dense_prefix(id); 1686 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1687 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1688 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1689 // Only enough dead space is filled so that any remaining dead space to the 1690 // left is larger than the minimum filler object. (The remainder is filled 1691 // during the copy/update phase.) 1692 // 1693 // The size of the dead space to the right of the boundary is not a 1694 // concern, since compaction will be able to use whatever space is 1695 // available. 1696 // 1697 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1698 // surrounds the space to be filled with an object. 1699 // 1700 // In the 32-bit VM, each bit represents two 32-bit words: 1701 // +---+ 1702 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1703 // end_bits: ... x x x | 0 | || 0 x x ... 1704 // +---+ 1705 // 1706 // In the 64-bit VM, each bit represents one 64-bit word: 1707 // +------------+ 1708 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1709 // end_bits: ... x x 1 | 0 || 0 | x x ... 1710 // +------------+ 1711 // +-------+ 1712 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1713 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1714 // +-------+ 1715 // +-----------+ 1716 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1717 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1718 // +-----------+ 1719 // +-------+ 1720 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1721 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1722 // +-------+ 1723 1724 // Initially assume case a, c or e will apply. 1725 size_t obj_len = CollectedHeap::min_fill_size(); 1726 HeapWord* obj_beg = dense_prefix_end - obj_len; 1727 1728 #ifdef _LP64 1729 if (MinObjAlignment > 1) { // object alignment > heap word size 1730 // Cases a, c or e. 1731 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1732 // Case b above. 1733 obj_beg = dense_prefix_end - 1; 1734 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1735 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1736 // Case d above. 1737 obj_beg = dense_prefix_end - 3; 1738 obj_len = 3; 1739 } 1740 #endif // #ifdef _LP64 1741 1742 CollectedHeap::fill_with_object(obj_beg, obj_len); 1743 _mark_bitmap.mark_obj(obj_beg, obj_len); 1744 _summary_data.add_obj(obj_beg, obj_len); 1745 assert(start_array(id) != NULL, "sanity"); 1746 start_array(id)->allocate_block(obj_beg); 1747 } 1748 } 1749 1750 void 1751 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1752 { 1753 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1754 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1755 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1756 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1757 cur->set_source_region(0); 1758 } 1759 } 1760 1761 void 1762 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1763 { 1764 assert(id < last_space_id, "id out of range"); 1765 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1766 ParallelOldGCSplitALot && id == old_space_id, 1767 "should have been reset in summarize_spaces_quick()"); 1768 1769 const MutableSpace* space = _space_info[id].space(); 1770 if (_space_info[id].new_top() != space->bottom()) { 1771 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1772 _space_info[id].set_dense_prefix(dense_prefix_end); 1773 1774 #ifndef PRODUCT 1775 if (TraceParallelOldGCDensePrefix) { 1776 print_dense_prefix_stats("ratio", id, maximum_compaction, 1777 dense_prefix_end); 1778 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1779 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1780 } 1781 #endif // #ifndef PRODUCT 1782 1783 // Recompute the summary data, taking into account the dense prefix. If 1784 // every last byte will be reclaimed, then the existing summary data which 1785 // compacts everything can be left in place. 1786 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1787 // If dead space crosses the dense prefix boundary, it is (at least 1788 // partially) filled with a dummy object, marked live and added to the 1789 // summary data. This simplifies the copy/update phase and must be done 1790 // before the final locations of objects are determined, to prevent 1791 // leaving a fragment of dead space that is too small to fill. 1792 fill_dense_prefix_end(id); 1793 1794 // Compute the destination of each Region, and thus each object. 1795 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1796 _summary_data.summarize(_space_info[id].split_info(), 1797 dense_prefix_end, space->top(), NULL, 1798 dense_prefix_end, space->end(), 1799 _space_info[id].new_top_addr()); 1800 } 1801 } 1802 1803 if (TraceParallelOldGCSummaryPhase) { 1804 const size_t region_size = ParallelCompactData::RegionSize; 1805 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1806 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1807 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1808 HeapWord* const new_top = _space_info[id].new_top(); 1809 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1810 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1811 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1812 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1813 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1814 id, space->capacity_in_words(), dense_prefix_end, 1815 dp_region, dp_words / region_size, 1816 cr_words / region_size, new_top); 1817 } 1818 } 1819 1820 #ifndef PRODUCT 1821 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1822 HeapWord* dst_beg, HeapWord* dst_end, 1823 SpaceId src_space_id, 1824 HeapWord* src_beg, HeapWord* src_end) 1825 { 1826 if (TraceParallelOldGCSummaryPhase) { 1827 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1828 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1829 SIZE_FORMAT "-" SIZE_FORMAT " " 1830 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1831 SIZE_FORMAT "-" SIZE_FORMAT, 1832 src_space_id, space_names[src_space_id], 1833 dst_space_id, space_names[dst_space_id], 1834 src_beg, src_end, 1835 _summary_data.addr_to_region_idx(src_beg), 1836 _summary_data.addr_to_region_idx(src_end), 1837 dst_beg, dst_end, 1838 _summary_data.addr_to_region_idx(dst_beg), 1839 _summary_data.addr_to_region_idx(dst_end)); 1840 } 1841 } 1842 #endif // #ifndef PRODUCT 1843 1844 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1845 bool maximum_compaction) 1846 { 1847 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1848 // trace("2"); 1849 1850 #ifdef ASSERT 1851 if (TraceParallelOldGCMarkingPhase) { 1852 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1853 "add_obj_bytes=" SIZE_FORMAT, 1854 add_obj_count, add_obj_size * HeapWordSize); 1855 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1856 "mark_bitmap_bytes=" SIZE_FORMAT, 1857 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1858 } 1859 #endif // #ifdef ASSERT 1860 1861 // Quick summarization of each space into itself, to see how much is live. 1862 summarize_spaces_quick(); 1863 1864 if (TraceParallelOldGCSummaryPhase) { 1865 tty->print_cr("summary_phase: after summarizing each space to self"); 1866 Universe::print(); 1867 NOT_PRODUCT(print_region_ranges()); 1868 if (Verbose) { 1869 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1870 } 1871 } 1872 1873 // The amount of live data that will end up in old space (assuming it fits). 1874 size_t old_space_total_live = 0; 1875 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1876 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1877 _space_info[id].space()->bottom()); 1878 } 1879 1880 MutableSpace* const old_space = _space_info[old_space_id].space(); 1881 const size_t old_capacity = old_space->capacity_in_words(); 1882 if (old_space_total_live > old_capacity) { 1883 // XXX - should also try to expand 1884 maximum_compaction = true; 1885 } 1886 #ifndef PRODUCT 1887 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1888 provoke_split(maximum_compaction); 1889 } 1890 #endif // #ifndef PRODUCT 1891 1892 // Old generations. 1893 summarize_space(old_space_id, maximum_compaction); 1894 1895 // Summarize the remaining spaces in the young gen. The initial target space 1896 // is the old gen. If a space does not fit entirely into the target, then the 1897 // remainder is compacted into the space itself and that space becomes the new 1898 // target. 1899 SpaceId dst_space_id = old_space_id; 1900 HeapWord* dst_space_end = old_space->end(); 1901 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1902 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1903 const MutableSpace* space = _space_info[id].space(); 1904 const size_t live = pointer_delta(_space_info[id].new_top(), 1905 space->bottom()); 1906 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1907 1908 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1909 SpaceId(id), space->bottom(), space->top());) 1910 if (live > 0 && live <= available) { 1911 // All the live data will fit. 1912 bool done = _summary_data.summarize(_space_info[id].split_info(), 1913 space->bottom(), space->top(), 1914 NULL, 1915 *new_top_addr, dst_space_end, 1916 new_top_addr); 1917 assert(done, "space must fit into old gen"); 1918 1919 // Reset the new_top value for the space. 1920 _space_info[id].set_new_top(space->bottom()); 1921 } else if (live > 0) { 1922 // Attempt to fit part of the source space into the target space. 1923 HeapWord* next_src_addr = NULL; 1924 bool done = _summary_data.summarize(_space_info[id].split_info(), 1925 space->bottom(), space->top(), 1926 &next_src_addr, 1927 *new_top_addr, dst_space_end, 1928 new_top_addr); 1929 assert(!done, "space should not fit into old gen"); 1930 assert(next_src_addr != NULL, "sanity"); 1931 1932 // The source space becomes the new target, so the remainder is compacted 1933 // within the space itself. 1934 dst_space_id = SpaceId(id); 1935 dst_space_end = space->end(); 1936 new_top_addr = _space_info[id].new_top_addr(); 1937 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1938 space->bottom(), dst_space_end, 1939 SpaceId(id), next_src_addr, space->top());) 1940 done = _summary_data.summarize(_space_info[id].split_info(), 1941 next_src_addr, space->top(), 1942 NULL, 1943 space->bottom(), dst_space_end, 1944 new_top_addr); 1945 assert(done, "space must fit when compacted into itself"); 1946 assert(*new_top_addr <= space->top(), "usage should not grow"); 1947 } 1948 } 1949 1950 if (TraceParallelOldGCSummaryPhase) { 1951 tty->print_cr("summary_phase: after final summarization"); 1952 Universe::print(); 1953 NOT_PRODUCT(print_region_ranges()); 1954 if (Verbose) { 1955 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1956 } 1957 } 1958 } 1959 1960 // This method should contain all heap-specific policy for invoking a full 1961 // collection. invoke_no_policy() will only attempt to compact the heap; it 1962 // will do nothing further. If we need to bail out for policy reasons, scavenge 1963 // before full gc, or any other specialized behavior, it needs to be added here. 1964 // 1965 // Note that this method should only be called from the vm_thread while at a 1966 // safepoint. 1967 // 1968 // Note that the all_soft_refs_clear flag in the collector policy 1969 // may be true because this method can be called without intervening 1970 // activity. For example when the heap space is tight and full measure 1971 // are being taken to free space. 1972 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1973 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1974 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1975 "should be in vm thread"); 1976 1977 ParallelScavengeHeap* heap = gc_heap(); 1978 GCCause::Cause gc_cause = heap->gc_cause(); 1979 assert(!heap->is_gc_active(), "not reentrant"); 1980 1981 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1982 IsGCActiveMark mark; 1983 1984 if (ScavengeBeforeFullGC) { 1985 PSScavenge::invoke_no_policy(); 1986 } 1987 1988 const bool clear_all_soft_refs = 1989 heap->collector_policy()->should_clear_all_soft_refs(); 1990 1991 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1992 maximum_heap_compaction); 1993 } 1994 1995 // This method contains no policy. You should probably 1996 // be calling invoke() instead. 1997 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1998 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1999 assert(ref_processor() != NULL, "Sanity"); 2000 2001 if (GC_locker::check_active_before_gc()) { 2002 return false; 2003 } 2004 2005 ParallelScavengeHeap* heap = gc_heap(); 2006 2007 _gc_timer.register_gc_start(); 2008 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 2009 2010 TimeStamp marking_start; 2011 TimeStamp compaction_start; 2012 TimeStamp collection_exit; 2013 2014 GCCause::Cause gc_cause = heap->gc_cause(); 2015 PSYoungGen* young_gen = heap->young_gen(); 2016 PSOldGen* old_gen = heap->old_gen(); 2017 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2018 2019 // The scope of casr should end after code that can change 2020 // CollectorPolicy::_should_clear_all_soft_refs. 2021 ClearedAllSoftRefs casr(maximum_heap_compaction, 2022 heap->collector_policy()); 2023 2024 if (ZapUnusedHeapArea) { 2025 // Save information needed to minimize mangling 2026 heap->record_gen_tops_before_GC(); 2027 } 2028 2029 heap->pre_full_gc_dump(&_gc_timer); 2030 2031 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2032 2033 // Make sure data structures are sane, make the heap parsable, and do other 2034 // miscellaneous bookkeeping. 2035 PreGCValues pre_gc_values; 2036 pre_compact(&pre_gc_values); 2037 2038 // Get the compaction manager reserved for the VM thread. 2039 ParCompactionManager* const vmthread_cm = 2040 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2041 2042 // Place after pre_compact() where the number of invocations is incremented. 2043 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2044 2045 { 2046 ResourceMark rm; 2047 HandleMark hm; 2048 2049 // Set the number of GC threads to be used in this collection 2050 gc_task_manager()->set_active_gang(); 2051 gc_task_manager()->task_idle_workers(); 2052 heap->set_par_threads(gc_task_manager()->active_workers()); 2053 2054 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2055 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id()); 2056 TraceCollectorStats tcs(counters()); 2057 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2058 2059 if (TraceOldGenTime) accumulated_time()->start(); 2060 2061 // Let the size policy know we're starting 2062 size_policy->major_collection_begin(); 2063 2064 CodeCache::gc_prologue(); 2065 2066 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2067 2068 ref_processor()->enable_discovery(); 2069 ref_processor()->setup_policy(maximum_heap_compaction); 2070 2071 bool marked_for_unloading = false; 2072 2073 marking_start.update(); 2074 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2075 2076 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2077 && GCCause::is_user_requested_gc(gc_cause); 2078 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2079 2080 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2081 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2082 2083 // adjust_roots() updates Universe::_intArrayKlassObj which is 2084 // needed by the compaction for filling holes in the dense prefix. 2085 adjust_roots(); 2086 2087 compaction_start.update(); 2088 compact(); 2089 2090 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2091 // done before resizing. 2092 post_compact(); 2093 2094 // Let the size policy know we're done 2095 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2096 2097 if (UseAdaptiveSizePolicy) { 2098 if (PrintAdaptiveSizePolicy) { 2099 gclog_or_tty->print("AdaptiveSizeStart: "); 2100 gclog_or_tty->stamp(); 2101 gclog_or_tty->print_cr(" collection: %d ", 2102 heap->total_collections()); 2103 if (Verbose) { 2104 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT 2105 " young_gen_capacity: " SIZE_FORMAT, 2106 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2107 } 2108 } 2109 2110 // Don't check if the size_policy is ready here. Let 2111 // the size_policy check that internally. 2112 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2113 (!GCCause::is_user_requested_gc(gc_cause) || 2114 UseAdaptiveSizePolicyWithSystemGC)) { 2115 // Swap the survivor spaces if from_space is empty. The 2116 // resize_young_gen() called below is normally used after 2117 // a successful young GC and swapping of survivor spaces; 2118 // otherwise, it will fail to resize the young gen with 2119 // the current implementation. 2120 if (young_gen->from_space()->is_empty()) { 2121 young_gen->from_space()->clear(SpaceDecorator::Mangle); 2122 young_gen->swap_spaces(); 2123 } 2124 2125 // Calculate optimal free space amounts 2126 assert(young_gen->max_size() > 2127 young_gen->from_space()->capacity_in_bytes() + 2128 young_gen->to_space()->capacity_in_bytes(), 2129 "Sizes of space in young gen are out-of-bounds"); 2130 2131 size_t young_live = young_gen->used_in_bytes(); 2132 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2133 size_t old_live = old_gen->used_in_bytes(); 2134 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2135 size_t max_old_gen_size = old_gen->max_gen_size(); 2136 size_t max_eden_size = young_gen->max_size() - 2137 young_gen->from_space()->capacity_in_bytes() - 2138 young_gen->to_space()->capacity_in_bytes(); 2139 2140 // Used for diagnostics 2141 size_policy->clear_generation_free_space_flags(); 2142 2143 size_policy->compute_generations_free_space(young_live, 2144 eden_live, 2145 old_live, 2146 cur_eden, 2147 max_old_gen_size, 2148 max_eden_size, 2149 true /* full gc*/); 2150 2151 size_policy->check_gc_overhead_limit(young_live, 2152 eden_live, 2153 max_old_gen_size, 2154 max_eden_size, 2155 true /* full gc*/, 2156 gc_cause, 2157 heap->collector_policy()); 2158 2159 size_policy->decay_supplemental_growth(true /* full gc*/); 2160 2161 heap->resize_old_gen( 2162 size_policy->calculated_old_free_size_in_bytes()); 2163 2164 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 2165 size_policy->calculated_survivor_size_in_bytes()); 2166 } 2167 if (PrintAdaptiveSizePolicy) { 2168 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2169 heap->total_collections()); 2170 } 2171 } 2172 2173 if (UsePerfData) { 2174 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2175 counters->update_counters(); 2176 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2177 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2178 } 2179 2180 heap->resize_all_tlabs(); 2181 2182 // Resize the metaspace capacity after a collection 2183 MetaspaceGC::compute_new_size(); 2184 2185 if (TraceOldGenTime) accumulated_time()->stop(); 2186 2187 if (PrintGC) { 2188 if (PrintGCDetails) { 2189 // No GC timestamp here. This is after GC so it would be confusing. 2190 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2191 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2192 heap->print_heap_change(pre_gc_values.heap_used()); 2193 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2194 } else { 2195 heap->print_heap_change(pre_gc_values.heap_used()); 2196 } 2197 } 2198 2199 // Track memory usage and detect low memory 2200 MemoryService::track_memory_usage(); 2201 heap->update_counters(); 2202 gc_task_manager()->release_idle_workers(); 2203 } 2204 2205 #ifdef ASSERT 2206 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2207 ParCompactionManager* const cm = 2208 ParCompactionManager::manager_array(int(i)); 2209 assert(cm->marking_stack()->is_empty(), "should be empty"); 2210 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2211 } 2212 #endif // ASSERT 2213 2214 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2215 HandleMark hm; // Discard invalid handles created during verification 2216 Universe::verify(" VerifyAfterGC:"); 2217 } 2218 2219 // Re-verify object start arrays 2220 if (VerifyObjectStartArray && 2221 VerifyAfterGC) { 2222 old_gen->verify_object_start_array(); 2223 } 2224 2225 if (ZapUnusedHeapArea) { 2226 old_gen->object_space()->check_mangled_unused_area_complete(); 2227 } 2228 2229 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2230 2231 collection_exit.update(); 2232 2233 heap->print_heap_after_gc(); 2234 heap->trace_heap_after_gc(&_gc_tracer); 2235 2236 if (PrintGCTaskTimeStamps) { 2237 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2238 INT64_FORMAT, 2239 marking_start.ticks(), compaction_start.ticks(), 2240 collection_exit.ticks()); 2241 gc_task_manager()->print_task_time_stamps(); 2242 } 2243 2244 heap->post_full_gc_dump(&_gc_timer); 2245 2246 #ifdef TRACESPINNING 2247 ParallelTaskTerminator::print_termination_counts(); 2248 #endif 2249 2250 _gc_timer.register_gc_end(); 2251 2252 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2253 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2254 2255 return true; 2256 } 2257 2258 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2259 PSYoungGen* young_gen, 2260 PSOldGen* old_gen) { 2261 MutableSpace* const eden_space = young_gen->eden_space(); 2262 assert(!eden_space->is_empty(), "eden must be non-empty"); 2263 assert(young_gen->virtual_space()->alignment() == 2264 old_gen->virtual_space()->alignment(), "alignments do not match"); 2265 2266 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2267 return false; 2268 } 2269 2270 // Both generations must be completely committed. 2271 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2272 return false; 2273 } 2274 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2275 return false; 2276 } 2277 2278 // Figure out how much to take from eden. Include the average amount promoted 2279 // in the total; otherwise the next young gen GC will simply bail out to a 2280 // full GC. 2281 const size_t alignment = old_gen->virtual_space()->alignment(); 2282 const size_t eden_used = eden_space->used_in_bytes(); 2283 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2284 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2285 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2286 2287 if (absorb_size >= eden_capacity) { 2288 return false; // Must leave some space in eden. 2289 } 2290 2291 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2292 if (new_young_size < young_gen->min_gen_size()) { 2293 return false; // Respect young gen minimum size. 2294 } 2295 2296 if (TraceAdaptiveGCBoundary && Verbose) { 2297 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2298 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2299 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2300 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2301 absorb_size / K, 2302 eden_capacity / K, (eden_capacity - absorb_size) / K, 2303 young_gen->from_space()->used_in_bytes() / K, 2304 young_gen->to_space()->used_in_bytes() / K, 2305 young_gen->capacity_in_bytes() / K, new_young_size / K); 2306 } 2307 2308 // Fill the unused part of the old gen. 2309 MutableSpace* const old_space = old_gen->object_space(); 2310 HeapWord* const unused_start = old_space->top(); 2311 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2312 2313 if (unused_words > 0) { 2314 if (unused_words < CollectedHeap::min_fill_size()) { 2315 return false; // If the old gen cannot be filled, must give up. 2316 } 2317 CollectedHeap::fill_with_objects(unused_start, unused_words); 2318 } 2319 2320 // Take the live data from eden and set both top and end in the old gen to 2321 // eden top. (Need to set end because reset_after_change() mangles the region 2322 // from end to virtual_space->high() in debug builds). 2323 HeapWord* const new_top = eden_space->top(); 2324 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2325 absorb_size); 2326 young_gen->reset_after_change(); 2327 old_space->set_top(new_top); 2328 old_space->set_end(new_top); 2329 old_gen->reset_after_change(); 2330 2331 // Update the object start array for the filler object and the data from eden. 2332 ObjectStartArray* const start_array = old_gen->start_array(); 2333 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2334 start_array->allocate_block(p); 2335 } 2336 2337 // Could update the promoted average here, but it is not typically updated at 2338 // full GCs and the value to use is unclear. Something like 2339 // 2340 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2341 2342 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2343 return true; 2344 } 2345 2346 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2347 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2348 "shouldn't return NULL"); 2349 return ParallelScavengeHeap::gc_task_manager(); 2350 } 2351 2352 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2353 bool maximum_heap_compaction, 2354 ParallelOldTracer *gc_tracer) { 2355 // Recursively traverse all live objects and mark them 2356 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2357 2358 ParallelScavengeHeap* heap = gc_heap(); 2359 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2360 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2361 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2362 ParallelTaskTerminator terminator(active_gc_threads, qset); 2363 2364 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2365 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2366 2367 // Need new claim bits before marking starts. 2368 ClassLoaderDataGraph::clear_claimed_marks(); 2369 2370 { 2371 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2372 2373 ParallelScavengeHeap::ParStrongRootsScope psrs; 2374 2375 GCTaskQueue* q = GCTaskQueue::create(); 2376 2377 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2378 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2379 // We scan the thread roots in parallel 2380 Threads::create_thread_roots_marking_tasks(q); 2381 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2382 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2383 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2384 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2385 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2386 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2387 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2388 2389 if (active_gc_threads > 1) { 2390 for (uint j = 0; j < active_gc_threads; j++) { 2391 q->enqueue(new StealMarkingTask(&terminator)); 2392 } 2393 } 2394 2395 gc_task_manager()->execute_and_wait(q); 2396 } 2397 2398 // Process reference objects found during marking 2399 { 2400 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2401 2402 ReferenceProcessorStats stats; 2403 if (ref_processor()->processing_is_mt()) { 2404 RefProcTaskExecutor task_executor; 2405 stats = ref_processor()->process_discovered_references( 2406 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2407 &task_executor, &_gc_timer, _gc_tracer.gc_id()); 2408 } else { 2409 stats = ref_processor()->process_discovered_references( 2410 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2411 &_gc_timer, _gc_tracer.gc_id()); 2412 } 2413 2414 gc_tracer->report_gc_reference_stats(stats); 2415 } 2416 2417 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2418 2419 // This is the point where the entire marking should have completed. 2420 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2421 2422 // Follow system dictionary roots and unload classes. 2423 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2424 2425 // Unload nmethods. 2426 CodeCache::do_unloading(is_alive_closure(), purged_class); 2427 2428 // Prune dead klasses from subklass/sibling/implementor lists. 2429 Klass::clean_weak_klass_links(is_alive_closure()); 2430 2431 // Delete entries for dead interned strings. 2432 StringTable::unlink(is_alive_closure()); 2433 2434 // Clean up unreferenced symbols in symbol table. 2435 SymbolTable::unlink(); 2436 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2437 } 2438 2439 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2440 ClassLoaderData* cld) { 2441 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2442 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2443 2444 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2445 } 2446 2447 // This should be moved to the shared markSweep code! 2448 class PSAlwaysTrueClosure: public BoolObjectClosure { 2449 public: 2450 bool do_object_b(oop p) { return true; } 2451 }; 2452 static PSAlwaysTrueClosure always_true; 2453 2454 void PSParallelCompact::adjust_roots() { 2455 // Adjust the pointers to reflect the new locations 2456 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2457 2458 // Need new claim bits when tracing through and adjusting pointers. 2459 ClassLoaderDataGraph::clear_claimed_marks(); 2460 2461 // General strong roots. 2462 Universe::oops_do(adjust_pointer_closure()); 2463 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2464 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2465 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2466 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2467 FlatProfiler::oops_do(adjust_pointer_closure()); 2468 Management::oops_do(adjust_pointer_closure()); 2469 JvmtiExport::oops_do(adjust_pointer_closure()); 2470 SystemDictionary::oops_do(adjust_pointer_closure()); 2471 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2472 2473 // Now adjust pointers in remaining weak roots. (All of which should 2474 // have been cleared if they pointed to non-surviving objects.) 2475 // Global (weak) JNI handles 2476 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2477 2478 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2479 CodeCache::blobs_do(&adjust_from_blobs); 2480 StringTable::oops_do(adjust_pointer_closure()); 2481 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2482 // Roots were visited so references into the young gen in roots 2483 // may have been scanned. Process them also. 2484 // Should the reference processor have a span that excludes 2485 // young gen objects? 2486 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2487 } 2488 2489 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2490 uint parallel_gc_threads) 2491 { 2492 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2493 2494 // Find the threads that are active 2495 unsigned int which = 0; 2496 2497 const uint task_count = MAX2(parallel_gc_threads, 1U); 2498 for (uint j = 0; j < task_count; j++) { 2499 q->enqueue(new DrainStacksCompactionTask(j)); 2500 ParCompactionManager::verify_region_list_empty(j); 2501 // Set the region stacks variables to "no" region stack values 2502 // so that they will be recognized and needing a region stack 2503 // in the stealing tasks if they do not get one by executing 2504 // a draining stack. 2505 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2506 cm->set_region_stack(NULL); 2507 cm->set_region_stack_index((uint)max_uintx); 2508 } 2509 ParCompactionManager::reset_recycled_stack_index(); 2510 2511 // Find all regions that are available (can be filled immediately) and 2512 // distribute them to the thread stacks. The iteration is done in reverse 2513 // order (high to low) so the regions will be removed in ascending order. 2514 2515 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2516 2517 size_t fillable_regions = 0; // A count for diagnostic purposes. 2518 // A region index which corresponds to the tasks created above. 2519 // "which" must be 0 <= which < task_count 2520 2521 which = 0; 2522 // id + 1 is used to test termination so unsigned can 2523 // be used with an old_space_id == 0. 2524 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2525 SpaceInfo* const space_info = _space_info + id; 2526 MutableSpace* const space = space_info->space(); 2527 HeapWord* const new_top = space_info->new_top(); 2528 2529 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2530 const size_t end_region = 2531 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2532 2533 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2534 if (sd.region(cur)->claim_unsafe()) { 2535 ParCompactionManager::region_list_push(which, cur); 2536 2537 if (TraceParallelOldGCCompactionPhase && Verbose) { 2538 const size_t count_mod_8 = fillable_regions & 7; 2539 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2540 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2541 if (count_mod_8 == 7) gclog_or_tty->cr(); 2542 } 2543 2544 NOT_PRODUCT(++fillable_regions;) 2545 2546 // Assign regions to tasks in round-robin fashion. 2547 if (++which == task_count) { 2548 assert(which <= parallel_gc_threads, 2549 "Inconsistent number of workers"); 2550 which = 0; 2551 } 2552 } 2553 } 2554 } 2555 2556 if (TraceParallelOldGCCompactionPhase) { 2557 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2558 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions); 2559 } 2560 } 2561 2562 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2563 2564 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2565 uint parallel_gc_threads) { 2566 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2567 2568 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2569 2570 // Iterate over all the spaces adding tasks for updating 2571 // regions in the dense prefix. Assume that 1 gc thread 2572 // will work on opening the gaps and the remaining gc threads 2573 // will work on the dense prefix. 2574 unsigned int space_id; 2575 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2576 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2577 const MutableSpace* const space = _space_info[space_id].space(); 2578 2579 if (dense_prefix_end == space->bottom()) { 2580 // There is no dense prefix for this space. 2581 continue; 2582 } 2583 2584 // The dense prefix is before this region. 2585 size_t region_index_end_dense_prefix = 2586 sd.addr_to_region_idx(dense_prefix_end); 2587 RegionData* const dense_prefix_cp = 2588 sd.region(region_index_end_dense_prefix); 2589 assert(dense_prefix_end == space->end() || 2590 dense_prefix_cp->available() || 2591 dense_prefix_cp->claimed(), 2592 "The region after the dense prefix should always be ready to fill"); 2593 2594 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2595 2596 // Is there dense prefix work? 2597 size_t total_dense_prefix_regions = 2598 region_index_end_dense_prefix - region_index_start; 2599 // How many regions of the dense prefix should be given to 2600 // each thread? 2601 if (total_dense_prefix_regions > 0) { 2602 uint tasks_for_dense_prefix = 1; 2603 if (total_dense_prefix_regions <= 2604 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2605 // Don't over partition. This assumes that 2606 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2607 // so there are not many regions to process. 2608 tasks_for_dense_prefix = parallel_gc_threads; 2609 } else { 2610 // Over partition 2611 tasks_for_dense_prefix = parallel_gc_threads * 2612 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2613 } 2614 size_t regions_per_thread = total_dense_prefix_regions / 2615 tasks_for_dense_prefix; 2616 // Give each thread at least 1 region. 2617 if (regions_per_thread == 0) { 2618 regions_per_thread = 1; 2619 } 2620 2621 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2622 if (region_index_start >= region_index_end_dense_prefix) { 2623 break; 2624 } 2625 // region_index_end is not processed 2626 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2627 region_index_end_dense_prefix); 2628 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2629 region_index_start, 2630 region_index_end)); 2631 region_index_start = region_index_end; 2632 } 2633 } 2634 // This gets any part of the dense prefix that did not 2635 // fit evenly. 2636 if (region_index_start < region_index_end_dense_prefix) { 2637 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2638 region_index_start, 2639 region_index_end_dense_prefix)); 2640 } 2641 } 2642 } 2643 2644 void PSParallelCompact::enqueue_region_stealing_tasks( 2645 GCTaskQueue* q, 2646 ParallelTaskTerminator* terminator_ptr, 2647 uint parallel_gc_threads) { 2648 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2649 2650 // Once a thread has drained it's stack, it should try to steal regions from 2651 // other threads. 2652 if (parallel_gc_threads > 1) { 2653 for (uint j = 0; j < parallel_gc_threads; j++) { 2654 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2655 } 2656 } 2657 } 2658 2659 #ifdef ASSERT 2660 // Write a histogram of the number of times the block table was filled for a 2661 // region. 2662 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2663 { 2664 if (!TraceParallelOldGCCompactionPhase) return; 2665 2666 typedef ParallelCompactData::RegionData rd_t; 2667 ParallelCompactData& sd = summary_data(); 2668 2669 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2670 MutableSpace* const spc = _space_info[id].space(); 2671 if (spc->bottom() != spc->top()) { 2672 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2673 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2674 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2675 2676 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2677 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2678 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2679 2680 for (const rd_t* cur = beg; cur < end; ++cur) { 2681 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2682 } 2683 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2684 for (size_t i = 0; i < histo_len; ++i) { 2685 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2686 histo[i], 100.0 * histo[i] / region_cnt); 2687 } 2688 out->cr(); 2689 } 2690 } 2691 } 2692 #endif // #ifdef ASSERT 2693 2694 void PSParallelCompact::compact() { 2695 // trace("5"); 2696 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2697 2698 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2699 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2700 PSOldGen* old_gen = heap->old_gen(); 2701 old_gen->start_array()->reset(); 2702 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2703 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2704 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2705 ParallelTaskTerminator terminator(active_gc_threads, qset); 2706 2707 GCTaskQueue* q = GCTaskQueue::create(); 2708 enqueue_region_draining_tasks(q, active_gc_threads); 2709 enqueue_dense_prefix_tasks(q, active_gc_threads); 2710 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2711 2712 { 2713 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2714 2715 gc_task_manager()->execute_and_wait(q); 2716 2717 #ifdef ASSERT 2718 // Verify that all regions have been processed before the deferred updates. 2719 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2720 verify_complete(SpaceId(id)); 2721 } 2722 #endif 2723 } 2724 2725 { 2726 // Update the deferred objects, if any. Any compaction manager can be used. 2727 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2728 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2729 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2730 update_deferred_objects(cm, SpaceId(id)); 2731 } 2732 } 2733 2734 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2735 } 2736 2737 #ifdef ASSERT 2738 void PSParallelCompact::verify_complete(SpaceId space_id) { 2739 // All Regions between space bottom() to new_top() should be marked as filled 2740 // and all Regions between new_top() and top() should be available (i.e., 2741 // should have been emptied). 2742 ParallelCompactData& sd = summary_data(); 2743 SpaceInfo si = _space_info[space_id]; 2744 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2745 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2746 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2747 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2748 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2749 2750 bool issued_a_warning = false; 2751 2752 size_t cur_region; 2753 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2754 const RegionData* const c = sd.region(cur_region); 2755 if (!c->completed()) { 2756 warning("region " SIZE_FORMAT " not filled: " 2757 "destination_count=" SIZE_FORMAT, 2758 cur_region, c->destination_count()); 2759 issued_a_warning = true; 2760 } 2761 } 2762 2763 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2764 const RegionData* const c = sd.region(cur_region); 2765 if (!c->available()) { 2766 warning("region " SIZE_FORMAT " not empty: " 2767 "destination_count=" SIZE_FORMAT, 2768 cur_region, c->destination_count()); 2769 issued_a_warning = true; 2770 } 2771 } 2772 2773 if (issued_a_warning) { 2774 print_region_ranges(); 2775 } 2776 } 2777 #endif // #ifdef ASSERT 2778 2779 // Update interior oops in the ranges of regions [beg_region, end_region). 2780 void 2781 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2782 SpaceId space_id, 2783 size_t beg_region, 2784 size_t end_region) { 2785 ParallelCompactData& sd = summary_data(); 2786 ParMarkBitMap* const mbm = mark_bitmap(); 2787 2788 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2789 HeapWord* const end_addr = sd.region_to_addr(end_region); 2790 assert(beg_region <= end_region, "bad region range"); 2791 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2792 2793 #ifdef ASSERT 2794 // Claim the regions to avoid triggering an assert when they are marked as 2795 // filled. 2796 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2797 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2798 } 2799 #endif // #ifdef ASSERT 2800 2801 if (beg_addr != space(space_id)->bottom()) { 2802 // Find the first live object or block of dead space that *starts* in this 2803 // range of regions. If a partial object crosses onto the region, skip it; 2804 // it will be marked for 'deferred update' when the object head is 2805 // processed. If dead space crosses onto the region, it is also skipped; it 2806 // will be filled when the prior region is processed. If neither of those 2807 // apply, the first word in the region is the start of a live object or dead 2808 // space. 2809 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2810 const RegionData* const cp = sd.region(beg_region); 2811 if (cp->partial_obj_size() != 0) { 2812 beg_addr = sd.partial_obj_end(beg_region); 2813 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2814 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2815 } 2816 } 2817 2818 if (beg_addr < end_addr) { 2819 // A live object or block of dead space starts in this range of Regions. 2820 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2821 2822 // Create closures and iterate. 2823 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2824 FillClosure fill_closure(cm, space_id); 2825 ParMarkBitMap::IterationStatus status; 2826 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2827 dense_prefix_end); 2828 if (status == ParMarkBitMap::incomplete) { 2829 update_closure.do_addr(update_closure.source()); 2830 } 2831 } 2832 2833 // Mark the regions as filled. 2834 RegionData* const beg_cp = sd.region(beg_region); 2835 RegionData* const end_cp = sd.region(end_region); 2836 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2837 cp->set_completed(); 2838 } 2839 } 2840 2841 // Return the SpaceId for the space containing addr. If addr is not in the 2842 // heap, last_space_id is returned. In debug mode it expects the address to be 2843 // in the heap and asserts such. 2844 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2845 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2846 2847 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2848 if (_space_info[id].space()->contains(addr)) { 2849 return SpaceId(id); 2850 } 2851 } 2852 2853 assert(false, "no space contains the addr"); 2854 return last_space_id; 2855 } 2856 2857 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2858 SpaceId id) { 2859 assert(id < last_space_id, "bad space id"); 2860 2861 ParallelCompactData& sd = summary_data(); 2862 const SpaceInfo* const space_info = _space_info + id; 2863 ObjectStartArray* const start_array = space_info->start_array(); 2864 2865 const MutableSpace* const space = space_info->space(); 2866 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2867 HeapWord* const beg_addr = space_info->dense_prefix(); 2868 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2869 2870 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2871 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2872 const RegionData* cur_region; 2873 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2874 HeapWord* const addr = cur_region->deferred_obj_addr(); 2875 if (addr != NULL) { 2876 if (start_array != NULL) { 2877 start_array->allocate_block(addr); 2878 } 2879 oop(addr)->update_contents(cm); 2880 assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)))); 2881 } 2882 } 2883 } 2884 2885 // Skip over count live words starting from beg, and return the address of the 2886 // next live word. Unless marked, the word corresponding to beg is assumed to 2887 // be dead. Callers must either ensure beg does not correspond to the middle of 2888 // an object, or account for those live words in some other way. Callers must 2889 // also ensure that there are enough live words in the range [beg, end) to skip. 2890 HeapWord* 2891 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2892 { 2893 assert(count > 0, "sanity"); 2894 2895 ParMarkBitMap* m = mark_bitmap(); 2896 idx_t bits_to_skip = m->words_to_bits(count); 2897 idx_t cur_beg = m->addr_to_bit(beg); 2898 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2899 2900 do { 2901 cur_beg = m->find_obj_beg(cur_beg, search_end); 2902 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2903 const size_t obj_bits = cur_end - cur_beg + 1; 2904 if (obj_bits > bits_to_skip) { 2905 return m->bit_to_addr(cur_beg + bits_to_skip); 2906 } 2907 bits_to_skip -= obj_bits; 2908 cur_beg = cur_end + 1; 2909 } while (bits_to_skip > 0); 2910 2911 // Skipping the desired number of words landed just past the end of an object. 2912 // Find the start of the next object. 2913 cur_beg = m->find_obj_beg(cur_beg, search_end); 2914 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2915 return m->bit_to_addr(cur_beg); 2916 } 2917 2918 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2919 SpaceId src_space_id, 2920 size_t src_region_idx) 2921 { 2922 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2923 2924 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2925 if (split_info.dest_region_addr() == dest_addr) { 2926 // The partial object ending at the split point contains the first word to 2927 // be copied to dest_addr. 2928 return split_info.first_src_addr(); 2929 } 2930 2931 const ParallelCompactData& sd = summary_data(); 2932 ParMarkBitMap* const bitmap = mark_bitmap(); 2933 const size_t RegionSize = ParallelCompactData::RegionSize; 2934 2935 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2936 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2937 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2938 HeapWord* const src_region_destination = src_region_ptr->destination(); 2939 2940 assert(dest_addr >= src_region_destination, "wrong src region"); 2941 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2942 2943 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2944 HeapWord* const src_region_end = src_region_beg + RegionSize; 2945 2946 HeapWord* addr = src_region_beg; 2947 if (dest_addr == src_region_destination) { 2948 // Return the first live word in the source region. 2949 if (partial_obj_size == 0) { 2950 addr = bitmap->find_obj_beg(addr, src_region_end); 2951 assert(addr < src_region_end, "no objects start in src region"); 2952 } 2953 return addr; 2954 } 2955 2956 // Must skip some live data. 2957 size_t words_to_skip = dest_addr - src_region_destination; 2958 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2959 2960 if (partial_obj_size >= words_to_skip) { 2961 // All the live words to skip are part of the partial object. 2962 addr += words_to_skip; 2963 if (partial_obj_size == words_to_skip) { 2964 // Find the first live word past the partial object. 2965 addr = bitmap->find_obj_beg(addr, src_region_end); 2966 assert(addr < src_region_end, "wrong src region"); 2967 } 2968 return addr; 2969 } 2970 2971 // Skip over the partial object (if any). 2972 if (partial_obj_size != 0) { 2973 words_to_skip -= partial_obj_size; 2974 addr += partial_obj_size; 2975 } 2976 2977 // Skip over live words due to objects that start in the region. 2978 addr = skip_live_words(addr, src_region_end, words_to_skip); 2979 assert(addr < src_region_end, "wrong src region"); 2980 return addr; 2981 } 2982 2983 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2984 SpaceId src_space_id, 2985 size_t beg_region, 2986 HeapWord* end_addr) 2987 { 2988 ParallelCompactData& sd = summary_data(); 2989 2990 #ifdef ASSERT 2991 MutableSpace* const src_space = _space_info[src_space_id].space(); 2992 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2993 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2994 "src_space_id does not match beg_addr"); 2995 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2996 "src_space_id does not match end_addr"); 2997 #endif // #ifdef ASSERT 2998 2999 RegionData* const beg = sd.region(beg_region); 3000 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 3001 3002 // Regions up to new_top() are enqueued if they become available. 3003 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3004 RegionData* const enqueue_end = 3005 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3006 3007 for (RegionData* cur = beg; cur < end; ++cur) { 3008 assert(cur->data_size() > 0, "region must have live data"); 3009 cur->decrement_destination_count(); 3010 if (cur < enqueue_end && cur->available() && cur->claim()) { 3011 cm->push_region(sd.region(cur)); 3012 } 3013 } 3014 } 3015 3016 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3017 SpaceId& src_space_id, 3018 HeapWord*& src_space_top, 3019 HeapWord* end_addr) 3020 { 3021 typedef ParallelCompactData::RegionData RegionData; 3022 3023 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3024 const size_t region_size = ParallelCompactData::RegionSize; 3025 3026 size_t src_region_idx = 0; 3027 3028 // Skip empty regions (if any) up to the top of the space. 3029 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3030 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3031 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3032 const RegionData* const top_region_ptr = 3033 sd.addr_to_region_ptr(top_aligned_up); 3034 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3035 ++src_region_ptr; 3036 } 3037 3038 if (src_region_ptr < top_region_ptr) { 3039 // The next source region is in the current space. Update src_region_idx 3040 // and the source address to match src_region_ptr. 3041 src_region_idx = sd.region(src_region_ptr); 3042 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3043 if (src_region_addr > closure.source()) { 3044 closure.set_source(src_region_addr); 3045 } 3046 return src_region_idx; 3047 } 3048 3049 // Switch to a new source space and find the first non-empty region. 3050 unsigned int space_id = src_space_id + 1; 3051 assert(space_id < last_space_id, "not enough spaces"); 3052 3053 HeapWord* const destination = closure.destination(); 3054 3055 do { 3056 MutableSpace* space = _space_info[space_id].space(); 3057 HeapWord* const bottom = space->bottom(); 3058 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3059 3060 // Iterate over the spaces that do not compact into themselves. 3061 if (bottom_cp->destination() != bottom) { 3062 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3063 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3064 3065 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3066 if (src_cp->live_obj_size() > 0) { 3067 // Found it. 3068 assert(src_cp->destination() == destination, 3069 "first live obj in the space must match the destination"); 3070 assert(src_cp->partial_obj_size() == 0, 3071 "a space cannot begin with a partial obj"); 3072 3073 src_space_id = SpaceId(space_id); 3074 src_space_top = space->top(); 3075 const size_t src_region_idx = sd.region(src_cp); 3076 closure.set_source(sd.region_to_addr(src_region_idx)); 3077 return src_region_idx; 3078 } else { 3079 assert(src_cp->data_size() == 0, "sanity"); 3080 } 3081 } 3082 } 3083 } while (++space_id < last_space_id); 3084 3085 assert(false, "no source region was found"); 3086 return 0; 3087 } 3088 3089 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3090 { 3091 typedef ParMarkBitMap::IterationStatus IterationStatus; 3092 const size_t RegionSize = ParallelCompactData::RegionSize; 3093 ParMarkBitMap* const bitmap = mark_bitmap(); 3094 ParallelCompactData& sd = summary_data(); 3095 RegionData* const region_ptr = sd.region(region_idx); 3096 3097 // Get the items needed to construct the closure. 3098 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3099 SpaceId dest_space_id = space_id(dest_addr); 3100 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3101 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3102 assert(dest_addr < new_top, "sanity"); 3103 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3104 3105 // Get the source region and related info. 3106 size_t src_region_idx = region_ptr->source_region(); 3107 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3108 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3109 3110 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3111 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3112 3113 // Adjust src_region_idx to prepare for decrementing destination counts (the 3114 // destination count is not decremented when a region is copied to itself). 3115 if (src_region_idx == region_idx) { 3116 src_region_idx += 1; 3117 } 3118 3119 if (bitmap->is_unmarked(closure.source())) { 3120 // The first source word is in the middle of an object; copy the remainder 3121 // of the object or as much as will fit. The fact that pointer updates were 3122 // deferred will be noted when the object header is processed. 3123 HeapWord* const old_src_addr = closure.source(); 3124 closure.copy_partial_obj(); 3125 if (closure.is_full()) { 3126 decrement_destination_counts(cm, src_space_id, src_region_idx, 3127 closure.source()); 3128 region_ptr->set_deferred_obj_addr(NULL); 3129 region_ptr->set_completed(); 3130 return; 3131 } 3132 3133 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3134 if (sd.region_align_down(old_src_addr) != end_addr) { 3135 // The partial object was copied from more than one source region. 3136 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3137 3138 // Move to the next source region, possibly switching spaces as well. All 3139 // args except end_addr may be modified. 3140 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3141 end_addr); 3142 } 3143 } 3144 3145 do { 3146 HeapWord* const cur_addr = closure.source(); 3147 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3148 src_space_top); 3149 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3150 3151 if (status == ParMarkBitMap::incomplete) { 3152 // The last obj that starts in the source region does not end in the 3153 // region. 3154 assert(closure.source() < end_addr, "sanity"); 3155 HeapWord* const obj_beg = closure.source(); 3156 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3157 src_space_top); 3158 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3159 if (obj_end < range_end) { 3160 // The end was found; the entire object will fit. 3161 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3162 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3163 } else { 3164 // The end was not found; the object will not fit. 3165 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3166 status = ParMarkBitMap::would_overflow; 3167 } 3168 } 3169 3170 if (status == ParMarkBitMap::would_overflow) { 3171 // The last object did not fit. Note that interior oop updates were 3172 // deferred, then copy enough of the object to fill the region. 3173 region_ptr->set_deferred_obj_addr(closure.destination()); 3174 status = closure.copy_until_full(); // copies from closure.source() 3175 3176 decrement_destination_counts(cm, src_space_id, src_region_idx, 3177 closure.source()); 3178 region_ptr->set_completed(); 3179 return; 3180 } 3181 3182 if (status == ParMarkBitMap::full) { 3183 decrement_destination_counts(cm, src_space_id, src_region_idx, 3184 closure.source()); 3185 region_ptr->set_deferred_obj_addr(NULL); 3186 region_ptr->set_completed(); 3187 return; 3188 } 3189 3190 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3191 3192 // Move to the next source region, possibly switching spaces as well. All 3193 // args except end_addr may be modified. 3194 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3195 end_addr); 3196 } while (true); 3197 } 3198 3199 void PSParallelCompact::fill_blocks(size_t region_idx) 3200 { 3201 // Fill in the block table elements for the specified region. Each block 3202 // table element holds the number of live words in the region that are to the 3203 // left of the first object that starts in the block. Thus only blocks in 3204 // which an object starts need to be filled. 3205 // 3206 // The algorithm scans the section of the bitmap that corresponds to the 3207 // region, keeping a running total of the live words. When an object start is 3208 // found, if it's the first to start in the block that contains it, the 3209 // current total is written to the block table element. 3210 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3211 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3212 const size_t RegionSize = ParallelCompactData::RegionSize; 3213 3214 ParallelCompactData& sd = summary_data(); 3215 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3216 if (partial_obj_size >= RegionSize) { 3217 return; // No objects start in this region. 3218 } 3219 3220 // Ensure the first loop iteration decides that the block has changed. 3221 size_t cur_block = sd.block_count(); 3222 3223 const ParMarkBitMap* const bitmap = mark_bitmap(); 3224 3225 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3226 assert((size_t)1 << Log2BitsPerBlock == 3227 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3228 3229 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3230 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3231 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3232 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3233 while (beg_bit < range_end) { 3234 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3235 if (new_block != cur_block) { 3236 cur_block = new_block; 3237 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3238 } 3239 3240 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3241 if (end_bit < range_end - 1) { 3242 live_bits += end_bit - beg_bit + 1; 3243 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3244 } else { 3245 return; 3246 } 3247 } 3248 } 3249 3250 void 3251 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3252 const MutableSpace* sp = space(space_id); 3253 if (sp->is_empty()) { 3254 return; 3255 } 3256 3257 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3258 ParMarkBitMap* const bitmap = mark_bitmap(); 3259 HeapWord* const dp_addr = dense_prefix(space_id); 3260 HeapWord* beg_addr = sp->bottom(); 3261 HeapWord* end_addr = sp->top(); 3262 3263 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3264 3265 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3266 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3267 if (beg_region < dp_region) { 3268 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3269 } 3270 3271 // The destination of the first live object that starts in the region is one 3272 // past the end of the partial object entering the region (if any). 3273 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3274 HeapWord* const new_top = _space_info[space_id].new_top(); 3275 assert(new_top >= dest_addr, "bad new_top value"); 3276 const size_t words = pointer_delta(new_top, dest_addr); 3277 3278 if (words > 0) { 3279 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3280 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3281 3282 ParMarkBitMap::IterationStatus status; 3283 status = bitmap->iterate(&closure, dest_addr, end_addr); 3284 assert(status == ParMarkBitMap::full, "iteration not complete"); 3285 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3286 "live objects skipped because closure is full"); 3287 } 3288 } 3289 3290 jlong PSParallelCompact::millis_since_last_gc() { 3291 // We need a monotonically non-decreasing time in ms but 3292 // os::javaTimeMillis() does not guarantee monotonicity. 3293 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3294 jlong ret_val = now - _time_of_last_gc; 3295 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3296 if (ret_val < 0) { 3297 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);) 3298 return 0; 3299 } 3300 return ret_val; 3301 } 3302 3303 void PSParallelCompact::reset_millis_since_last_gc() { 3304 // We need a monotonically non-decreasing time in ms but 3305 // os::javaTimeMillis() does not guarantee monotonicity. 3306 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3307 } 3308 3309 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3310 { 3311 if (source() != destination()) { 3312 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3313 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3314 } 3315 update_state(words_remaining()); 3316 assert(is_full(), "sanity"); 3317 return ParMarkBitMap::full; 3318 } 3319 3320 void MoveAndUpdateClosure::copy_partial_obj() 3321 { 3322 size_t words = words_remaining(); 3323 3324 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3325 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3326 if (end_addr < range_end) { 3327 words = bitmap()->obj_size(source(), end_addr); 3328 } 3329 3330 // This test is necessary; if omitted, the pointer updates to a partial object 3331 // that crosses the dense prefix boundary could be overwritten. 3332 if (source() != destination()) { 3333 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3334 Copy::aligned_conjoint_words(source(), destination(), words); 3335 } 3336 update_state(words); 3337 } 3338 3339 ParMarkBitMapClosure::IterationStatus 3340 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3341 assert(destination() != NULL, "sanity"); 3342 assert(bitmap()->obj_size(addr) == words, "bad size"); 3343 3344 _source = addr; 3345 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3346 destination(), "wrong destination"); 3347 3348 if (words > words_remaining()) { 3349 return ParMarkBitMap::would_overflow; 3350 } 3351 3352 // The start_array must be updated even if the object is not moving. 3353 if (_start_array != NULL) { 3354 _start_array->allocate_block(destination()); 3355 } 3356 3357 if (destination() != source()) { 3358 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3359 Copy::aligned_conjoint_words(source(), destination(), words); 3360 } 3361 3362 oop moved_oop = (oop) destination(); 3363 moved_oop->update_contents(compaction_manager()); 3364 assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop))); 3365 3366 update_state(words); 3367 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3368 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3369 } 3370 3371 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3372 ParCompactionManager* cm, 3373 PSParallelCompact::SpaceId space_id) : 3374 ParMarkBitMapClosure(mbm, cm), 3375 _space_id(space_id), 3376 _start_array(PSParallelCompact::start_array(space_id)) 3377 { 3378 } 3379 3380 // Updates the references in the object to their new values. 3381 ParMarkBitMapClosure::IterationStatus 3382 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3383 do_addr(addr); 3384 return ParMarkBitMap::incomplete; 3385 }