1 /* 2 * Copyright (c) 2005, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp" 30 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp" 31 #include "gc_implementation/parallelScavenge/pcTasks.hpp" 32 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp" 33 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp" 34 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp" 35 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp" 36 #include "gc_implementation/parallelScavenge/psOldGen.hpp" 37 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp" 38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 39 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 41 #include "gc_implementation/shared/gcHeapSummary.hpp" 42 #include "gc_implementation/shared/gcTimer.hpp" 43 #include "gc_implementation/shared/gcTrace.hpp" 44 #include "gc_implementation/shared/gcTraceTime.hpp" 45 #include "gc_implementation/shared/isGCActiveMark.hpp" 46 #include "gc_interface/gcCause.hpp" 47 #include "memory/gcLocker.inline.hpp" 48 #include "memory/referencePolicy.hpp" 49 #include "memory/referenceProcessor.hpp" 50 #include "oops/methodData.hpp" 51 #include "oops/oop.inline.hpp" 52 #include "oops/oop.pcgc.inline.hpp" 53 #include "runtime/atomic.inline.hpp" 54 #include "runtime/fprofiler.hpp" 55 #include "runtime/safepoint.hpp" 56 #include "runtime/vmThread.hpp" 57 #include "services/management.hpp" 58 #include "services/memoryService.hpp" 59 #include "services/memTracker.hpp" 60 #include "utilities/events.hpp" 61 #include "utilities/stack.inline.hpp" 62 63 #include <math.h> 64 65 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 66 67 // All sizes are in HeapWords. 68 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 69 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 70 const size_t ParallelCompactData::RegionSizeBytes = 71 RegionSize << LogHeapWordSize; 72 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 73 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 74 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 75 76 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 77 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 78 const size_t ParallelCompactData::BlockSizeBytes = 79 BlockSize << LogHeapWordSize; 80 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 81 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 82 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 83 84 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 85 const size_t ParallelCompactData::Log2BlocksPerRegion = 86 Log2RegionSize - Log2BlockSize; 87 88 const ParallelCompactData::RegionData::region_sz_t 89 ParallelCompactData::RegionData::dc_shift = 27; 90 91 const ParallelCompactData::RegionData::region_sz_t 92 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 93 94 const ParallelCompactData::RegionData::region_sz_t 95 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 96 97 const ParallelCompactData::RegionData::region_sz_t 98 ParallelCompactData::RegionData::los_mask = ~dc_mask; 99 100 const ParallelCompactData::RegionData::region_sz_t 101 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 102 103 const ParallelCompactData::RegionData::region_sz_t 104 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 105 106 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 107 bool PSParallelCompact::_print_phases = false; 108 109 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 110 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL; 111 112 double PSParallelCompact::_dwl_mean; 113 double PSParallelCompact::_dwl_std_dev; 114 double PSParallelCompact::_dwl_first_term; 115 double PSParallelCompact::_dwl_adjustment; 116 #ifdef ASSERT 117 bool PSParallelCompact::_dwl_initialized = false; 118 #endif // #ifdef ASSERT 119 120 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 121 HeapWord* destination) 122 { 123 assert(src_region_idx != 0, "invalid src_region_idx"); 124 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 125 assert(destination != NULL, "invalid destination argument"); 126 127 _src_region_idx = src_region_idx; 128 _partial_obj_size = partial_obj_size; 129 _destination = destination; 130 131 // These fields may not be updated below, so make sure they're clear. 132 assert(_dest_region_addr == NULL, "should have been cleared"); 133 assert(_first_src_addr == NULL, "should have been cleared"); 134 135 // Determine the number of destination regions for the partial object. 136 HeapWord* const last_word = destination + partial_obj_size - 1; 137 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 138 HeapWord* const beg_region_addr = sd.region_align_down(destination); 139 HeapWord* const end_region_addr = sd.region_align_down(last_word); 140 141 if (beg_region_addr == end_region_addr) { 142 // One destination region. 143 _destination_count = 1; 144 if (end_region_addr == destination) { 145 // The destination falls on a region boundary, thus the first word of the 146 // partial object will be the first word copied to the destination region. 147 _dest_region_addr = end_region_addr; 148 _first_src_addr = sd.region_to_addr(src_region_idx); 149 } 150 } else { 151 // Two destination regions. When copied, the partial object will cross a 152 // destination region boundary, so a word somewhere within the partial 153 // object will be the first word copied to the second destination region. 154 _destination_count = 2; 155 _dest_region_addr = end_region_addr; 156 const size_t ofs = pointer_delta(end_region_addr, destination); 157 assert(ofs < _partial_obj_size, "sanity"); 158 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 159 } 160 } 161 162 void SplitInfo::clear() 163 { 164 _src_region_idx = 0; 165 _partial_obj_size = 0; 166 _destination = NULL; 167 _destination_count = 0; 168 _dest_region_addr = NULL; 169 _first_src_addr = NULL; 170 assert(!is_valid(), "sanity"); 171 } 172 173 #ifdef ASSERT 174 void SplitInfo::verify_clear() 175 { 176 assert(_src_region_idx == 0, "not clear"); 177 assert(_partial_obj_size == 0, "not clear"); 178 assert(_destination == NULL, "not clear"); 179 assert(_destination_count == 0, "not clear"); 180 assert(_dest_region_addr == NULL, "not clear"); 181 assert(_first_src_addr == NULL, "not clear"); 182 } 183 #endif // #ifdef ASSERT 184 185 186 void PSParallelCompact::print_on_error(outputStream* st) { 187 _mark_bitmap.print_on_error(st); 188 } 189 190 #ifndef PRODUCT 191 const char* PSParallelCompact::space_names[] = { 192 "old ", "eden", "from", "to " 193 }; 194 195 void PSParallelCompact::print_region_ranges() 196 { 197 tty->print_cr("space bottom top end new_top"); 198 tty->print_cr("------ ---------- ---------- ---------- ----------"); 199 200 for (unsigned int id = 0; id < last_space_id; ++id) { 201 const MutableSpace* space = _space_info[id].space(); 202 tty->print_cr("%u %s " 203 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 205 id, space_names[id], 206 summary_data().addr_to_region_idx(space->bottom()), 207 summary_data().addr_to_region_idx(space->top()), 208 summary_data().addr_to_region_idx(space->end()), 209 summary_data().addr_to_region_idx(_space_info[id].new_top())); 210 } 211 } 212 213 void 214 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 215 { 216 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 217 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 218 219 ParallelCompactData& sd = PSParallelCompact::summary_data(); 220 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 221 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 222 REGION_IDX_FORMAT " " PTR_FORMAT " " 223 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 224 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 225 i, c->data_location(), dci, c->destination(), 226 c->partial_obj_size(), c->live_obj_size(), 227 c->data_size(), c->source_region(), c->destination_count()); 228 229 #undef REGION_IDX_FORMAT 230 #undef REGION_DATA_FORMAT 231 } 232 233 void 234 print_generic_summary_data(ParallelCompactData& summary_data, 235 HeapWord* const beg_addr, 236 HeapWord* const end_addr) 237 { 238 size_t total_words = 0; 239 size_t i = summary_data.addr_to_region_idx(beg_addr); 240 const size_t last = summary_data.addr_to_region_idx(end_addr); 241 HeapWord* pdest = 0; 242 243 while (i <= last) { 244 ParallelCompactData::RegionData* c = summary_data.region(i); 245 if (c->data_size() != 0 || c->destination() != pdest) { 246 print_generic_summary_region(i, c); 247 total_words += c->data_size(); 248 pdest = c->destination(); 249 } 250 ++i; 251 } 252 253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 254 } 255 256 void 257 print_generic_summary_data(ParallelCompactData& summary_data, 258 SpaceInfo* space_info) 259 { 260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 261 const MutableSpace* space = space_info[id].space(); 262 print_generic_summary_data(summary_data, space->bottom(), 263 MAX2(space->top(), space_info[id].new_top())); 264 } 265 } 266 267 void 268 print_initial_summary_region(size_t i, 269 const ParallelCompactData::RegionData* c, 270 bool newline = true) 271 { 272 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 273 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 275 i, c->destination(), 276 c->partial_obj_size(), c->live_obj_size(), 277 c->data_size(), c->source_region(), c->destination_count()); 278 if (newline) tty->cr(); 279 } 280 281 void 282 print_initial_summary_data(ParallelCompactData& summary_data, 283 const MutableSpace* space) { 284 if (space->top() == space->bottom()) { 285 return; 286 } 287 288 const size_t region_size = ParallelCompactData::RegionSize; 289 typedef ParallelCompactData::RegionData RegionData; 290 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 291 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 292 const RegionData* c = summary_data.region(end_region - 1); 293 HeapWord* end_addr = c->destination() + c->data_size(); 294 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 295 296 // Print (and count) the full regions at the beginning of the space. 297 size_t full_region_count = 0; 298 size_t i = summary_data.addr_to_region_idx(space->bottom()); 299 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 300 print_initial_summary_region(i, summary_data.region(i)); 301 ++full_region_count; 302 ++i; 303 } 304 305 size_t live_to_right = live_in_space - full_region_count * region_size; 306 307 double max_reclaimed_ratio = 0.0; 308 size_t max_reclaimed_ratio_region = 0; 309 size_t max_dead_to_right = 0; 310 size_t max_live_to_right = 0; 311 312 // Print the 'reclaimed ratio' for regions while there is something live in 313 // the region or to the right of it. The remaining regions are empty (and 314 // uninteresting), and computing the ratio will result in division by 0. 315 while (i < end_region && live_to_right > 0) { 316 c = summary_data.region(i); 317 HeapWord* const region_addr = summary_data.region_to_addr(i); 318 const size_t used_to_right = pointer_delta(space->top(), region_addr); 319 const size_t dead_to_right = used_to_right - live_to_right; 320 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 321 322 if (reclaimed_ratio > max_reclaimed_ratio) { 323 max_reclaimed_ratio = reclaimed_ratio; 324 max_reclaimed_ratio_region = i; 325 max_dead_to_right = dead_to_right; 326 max_live_to_right = live_to_right; 327 } 328 329 print_initial_summary_region(i, c, false); 330 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 331 reclaimed_ratio, dead_to_right, live_to_right); 332 333 live_to_right -= c->data_size(); 334 ++i; 335 } 336 337 // Any remaining regions are empty. Print one more if there is one. 338 if (i < end_region) { 339 print_initial_summary_region(i, summary_data.region(i)); 340 } 341 342 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 343 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 344 max_reclaimed_ratio_region, max_dead_to_right, 345 max_live_to_right, max_reclaimed_ratio); 346 } 347 348 void 349 print_initial_summary_data(ParallelCompactData& summary_data, 350 SpaceInfo* space_info) { 351 unsigned int id = PSParallelCompact::old_space_id; 352 const MutableSpace* space; 353 do { 354 space = space_info[id].space(); 355 print_initial_summary_data(summary_data, space); 356 } while (++id < PSParallelCompact::eden_space_id); 357 358 do { 359 space = space_info[id].space(); 360 print_generic_summary_data(summary_data, space->bottom(), space->top()); 361 } while (++id < PSParallelCompact::last_space_id); 362 } 363 #endif // #ifndef PRODUCT 364 365 #ifdef ASSERT 366 size_t add_obj_count; 367 size_t add_obj_size; 368 size_t mark_bitmap_count; 369 size_t mark_bitmap_size; 370 #endif // #ifdef ASSERT 371 372 ParallelCompactData::ParallelCompactData() 373 { 374 _region_start = 0; 375 376 _region_vspace = 0; 377 _reserved_byte_size = 0; 378 _region_data = 0; 379 _region_count = 0; 380 381 _block_vspace = 0; 382 _block_data = 0; 383 _block_count = 0; 384 } 385 386 bool ParallelCompactData::initialize(MemRegion covered_region) 387 { 388 _region_start = covered_region.start(); 389 const size_t region_size = covered_region.word_size(); 390 DEBUG_ONLY(_region_end = _region_start + region_size;) 391 392 assert(region_align_down(_region_start) == _region_start, 393 "region start not aligned"); 394 assert((region_size & RegionSizeOffsetMask) == 0, 395 "region size not a multiple of RegionSize"); 396 397 bool result = initialize_region_data(region_size) && initialize_block_data(); 398 return result; 399 } 400 401 PSVirtualSpace* 402 ParallelCompactData::create_vspace(size_t count, size_t element_size) 403 { 404 const size_t raw_bytes = count * element_size; 405 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); 406 const size_t granularity = os::vm_allocation_granularity(); 407 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 408 409 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 410 MAX2(page_sz, granularity); 411 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 412 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 413 rs.size()); 414 415 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 416 417 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 418 if (vspace != 0) { 419 if (vspace->expand_by(_reserved_byte_size)) { 420 return vspace; 421 } 422 delete vspace; 423 // Release memory reserved in the space. 424 rs.release(); 425 } 426 427 return 0; 428 } 429 430 bool ParallelCompactData::initialize_region_data(size_t region_size) 431 { 432 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 433 _region_vspace = create_vspace(count, sizeof(RegionData)); 434 if (_region_vspace != 0) { 435 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 436 _region_count = count; 437 return true; 438 } 439 return false; 440 } 441 442 bool ParallelCompactData::initialize_block_data() 443 { 444 assert(_region_count != 0, "region data must be initialized first"); 445 const size_t count = _region_count << Log2BlocksPerRegion; 446 _block_vspace = create_vspace(count, sizeof(BlockData)); 447 if (_block_vspace != 0) { 448 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 449 _block_count = count; 450 return true; 451 } 452 return false; 453 } 454 455 void ParallelCompactData::clear() 456 { 457 memset(_region_data, 0, _region_vspace->committed_size()); 458 memset(_block_data, 0, _block_vspace->committed_size()); 459 } 460 461 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 462 assert(beg_region <= _region_count, "beg_region out of range"); 463 assert(end_region <= _region_count, "end_region out of range"); 464 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 465 466 const size_t region_cnt = end_region - beg_region; 467 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 468 469 const size_t beg_block = beg_region * BlocksPerRegion; 470 const size_t block_cnt = region_cnt * BlocksPerRegion; 471 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 472 } 473 474 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 475 { 476 const RegionData* cur_cp = region(region_idx); 477 const RegionData* const end_cp = region(region_count() - 1); 478 479 HeapWord* result = region_to_addr(region_idx); 480 if (cur_cp < end_cp) { 481 do { 482 result += cur_cp->partial_obj_size(); 483 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 484 } 485 return result; 486 } 487 488 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 489 { 490 const size_t obj_ofs = pointer_delta(addr, _region_start); 491 const size_t beg_region = obj_ofs >> Log2RegionSize; 492 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 493 494 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 495 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 496 497 if (beg_region == end_region) { 498 // All in one region. 499 _region_data[beg_region].add_live_obj(len); 500 return; 501 } 502 503 // First region. 504 const size_t beg_ofs = region_offset(addr); 505 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 506 507 Klass* klass = ((oop)addr)->klass(); 508 // Middle regions--completely spanned by this object. 509 for (size_t region = beg_region + 1; region < end_region; ++region) { 510 _region_data[region].set_partial_obj_size(RegionSize); 511 _region_data[region].set_partial_obj_addr(addr); 512 } 513 514 // Last region. 515 const size_t end_ofs = region_offset(addr + len - 1); 516 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 517 _region_data[end_region].set_partial_obj_addr(addr); 518 } 519 520 void 521 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 522 { 523 assert(region_offset(beg) == 0, "not RegionSize aligned"); 524 assert(region_offset(end) == 0, "not RegionSize aligned"); 525 526 size_t cur_region = addr_to_region_idx(beg); 527 const size_t end_region = addr_to_region_idx(end); 528 HeapWord* addr = beg; 529 while (cur_region < end_region) { 530 _region_data[cur_region].set_destination(addr); 531 _region_data[cur_region].set_destination_count(0); 532 _region_data[cur_region].set_source_region(cur_region); 533 _region_data[cur_region].set_data_location(addr); 534 535 // Update live_obj_size so the region appears completely full. 536 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 537 _region_data[cur_region].set_live_obj_size(live_size); 538 539 ++cur_region; 540 addr += RegionSize; 541 } 542 } 543 544 // Find the point at which a space can be split and, if necessary, record the 545 // split point. 546 // 547 // If the current src region (which overflowed the destination space) doesn't 548 // have a partial object, the split point is at the beginning of the current src 549 // region (an "easy" split, no extra bookkeeping required). 550 // 551 // If the current src region has a partial object, the split point is in the 552 // region where that partial object starts (call it the split_region). If 553 // split_region has a partial object, then the split point is just after that 554 // partial object (a "hard" split where we have to record the split data and 555 // zero the partial_obj_size field). With a "hard" split, we know that the 556 // partial_obj ends within split_region because the partial object that caused 557 // the overflow starts in split_region. If split_region doesn't have a partial 558 // obj, then the split is at the beginning of split_region (another "easy" 559 // split). 560 HeapWord* 561 ParallelCompactData::summarize_split_space(size_t src_region, 562 SplitInfo& split_info, 563 HeapWord* destination, 564 HeapWord* target_end, 565 HeapWord** target_next) 566 { 567 assert(destination <= target_end, "sanity"); 568 assert(destination + _region_data[src_region].data_size() > target_end, 569 "region should not fit into target space"); 570 assert(is_region_aligned(target_end), "sanity"); 571 572 size_t split_region = src_region; 573 HeapWord* split_destination = destination; 574 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 575 576 if (destination + partial_obj_size > target_end) { 577 // The split point is just after the partial object (if any) in the 578 // src_region that contains the start of the object that overflowed the 579 // destination space. 580 // 581 // Find the start of the "overflow" object and set split_region to the 582 // region containing it. 583 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 584 split_region = addr_to_region_idx(overflow_obj); 585 586 // Clear the source_region field of all destination regions whose first word 587 // came from data after the split point (a non-null source_region field 588 // implies a region must be filled). 589 // 590 // An alternative to the simple loop below: clear during post_compact(), 591 // which uses memcpy instead of individual stores, and is easy to 592 // parallelize. (The downside is that it clears the entire RegionData 593 // object as opposed to just one field.) 594 // 595 // post_compact() would have to clear the summary data up to the highest 596 // address that was written during the summary phase, which would be 597 // 598 // max(top, max(new_top, clear_top)) 599 // 600 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 601 // to target_end. 602 const RegionData* const sr = region(split_region); 603 const size_t beg_idx = 604 addr_to_region_idx(region_align_up(sr->destination() + 605 sr->partial_obj_size())); 606 const size_t end_idx = addr_to_region_idx(target_end); 607 608 if (TraceParallelOldGCSummaryPhase) { 609 gclog_or_tty->print_cr("split: clearing source_region field in [" 610 SIZE_FORMAT ", " SIZE_FORMAT ")", 611 beg_idx, end_idx); 612 } 613 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 614 _region_data[idx].set_source_region(0); 615 } 616 617 // Set split_destination and partial_obj_size to reflect the split region. 618 split_destination = sr->destination(); 619 partial_obj_size = sr->partial_obj_size(); 620 } 621 622 // The split is recorded only if a partial object extends onto the region. 623 if (partial_obj_size != 0) { 624 _region_data[split_region].set_partial_obj_size(0); 625 split_info.record(split_region, partial_obj_size, split_destination); 626 } 627 628 // Setup the continuation addresses. 629 *target_next = split_destination + partial_obj_size; 630 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 631 632 if (TraceParallelOldGCSummaryPhase) { 633 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 634 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 635 " pos=" SIZE_FORMAT, 636 split_type, source_next, split_region, 637 partial_obj_size); 638 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 639 " tn=" PTR_FORMAT, 640 split_type, split_destination, 641 addr_to_region_idx(split_destination), 642 *target_next); 643 644 if (partial_obj_size != 0) { 645 HeapWord* const po_beg = split_info.destination(); 646 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 647 gclog_or_tty->print_cr("%s split: " 648 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 649 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 650 split_type, 651 po_beg, addr_to_region_idx(po_beg), 652 po_end, addr_to_region_idx(po_end)); 653 } 654 } 655 656 return source_next; 657 } 658 659 bool ParallelCompactData::summarize(SplitInfo& split_info, 660 HeapWord* source_beg, HeapWord* source_end, 661 HeapWord** source_next, 662 HeapWord* target_beg, HeapWord* target_end, 663 HeapWord** target_next) 664 { 665 if (TraceParallelOldGCSummaryPhase) { 666 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 667 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 668 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 669 source_beg, source_end, source_next_val, 670 target_beg, target_end, *target_next); 671 } 672 673 size_t cur_region = addr_to_region_idx(source_beg); 674 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 675 676 HeapWord *dest_addr = target_beg; 677 while (cur_region < end_region) { 678 // The destination must be set even if the region has no data. 679 _region_data[cur_region].set_destination(dest_addr); 680 681 size_t words = _region_data[cur_region].data_size(); 682 if (words > 0) { 683 // If cur_region does not fit entirely into the target space, find a point 684 // at which the source space can be 'split' so that part is copied to the 685 // target space and the rest is copied elsewhere. 686 if (dest_addr + words > target_end) { 687 assert(source_next != NULL, "source_next is NULL when splitting"); 688 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 689 target_end, target_next); 690 return false; 691 } 692 693 // Compute the destination_count for cur_region, and if necessary, update 694 // source_region for a destination region. The source_region field is 695 // updated if cur_region is the first (left-most) region to be copied to a 696 // destination region. 697 // 698 // The destination_count calculation is a bit subtle. A region that has 699 // data that compacts into itself does not count itself as a destination. 700 // This maintains the invariant that a zero count means the region is 701 // available and can be claimed and then filled. 702 uint destination_count = 0; 703 if (split_info.is_split(cur_region)) { 704 // The current region has been split: the partial object will be copied 705 // to one destination space and the remaining data will be copied to 706 // another destination space. Adjust the initial destination_count and, 707 // if necessary, set the source_region field if the partial object will 708 // cross a destination region boundary. 709 destination_count = split_info.destination_count(); 710 if (destination_count == 2) { 711 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 712 _region_data[dest_idx].set_source_region(cur_region); 713 } 714 } 715 716 HeapWord* const last_addr = dest_addr + words - 1; 717 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 718 const size_t dest_region_2 = addr_to_region_idx(last_addr); 719 720 // Initially assume that the destination regions will be the same and 721 // adjust the value below if necessary. Under this assumption, if 722 // cur_region == dest_region_2, then cur_region will be compacted 723 // completely into itself. 724 destination_count += cur_region == dest_region_2 ? 0 : 1; 725 if (dest_region_1 != dest_region_2) { 726 // Destination regions differ; adjust destination_count. 727 destination_count += 1; 728 // Data from cur_region will be copied to the start of dest_region_2. 729 _region_data[dest_region_2].set_source_region(cur_region); 730 } else if (region_offset(dest_addr) == 0) { 731 // Data from cur_region will be copied to the start of the destination 732 // region. 733 _region_data[dest_region_1].set_source_region(cur_region); 734 } 735 736 _region_data[cur_region].set_destination_count(destination_count); 737 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 738 dest_addr += words; 739 } 740 741 ++cur_region; 742 } 743 744 *target_next = dest_addr; 745 return true; 746 } 747 748 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 749 assert(addr != NULL, "Should detect NULL oop earlier"); 750 assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap"); 751 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 752 753 // Region covering the object. 754 RegionData* const region_ptr = addr_to_region_ptr(addr); 755 HeapWord* result = region_ptr->destination(); 756 757 // If the entire Region is live, the new location is region->destination + the 758 // offset of the object within in the Region. 759 760 // Run some performance tests to determine if this special case pays off. It 761 // is worth it for pointers into the dense prefix. If the optimization to 762 // avoid pointer updates in regions that only point to the dense prefix is 763 // ever implemented, this should be revisited. 764 if (region_ptr->data_size() == RegionSize) { 765 result += region_offset(addr); 766 return result; 767 } 768 769 // Otherwise, the new location is region->destination + block offset + the 770 // number of live words in the Block that are (a) to the left of addr and (b) 771 // due to objects that start in the Block. 772 773 // Fill in the block table if necessary. This is unsynchronized, so multiple 774 // threads may fill the block table for a region (harmless, since it is 775 // idempotent). 776 if (!region_ptr->blocks_filled()) { 777 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 778 region_ptr->set_blocks_filled(); 779 } 780 781 HeapWord* const search_start = block_align_down(addr); 782 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 783 784 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 785 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 786 result += block_offset + live; 787 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 788 return result; 789 } 790 791 #ifdef ASSERT 792 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 793 { 794 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 795 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 796 for (const size_t* p = beg; p < end; ++p) { 797 assert(*p == 0, "not zero"); 798 } 799 } 800 801 void ParallelCompactData::verify_clear() 802 { 803 verify_clear(_region_vspace); 804 verify_clear(_block_vspace); 805 } 806 #endif // #ifdef ASSERT 807 808 STWGCTimer PSParallelCompact::_gc_timer; 809 ParallelOldTracer PSParallelCompact::_gc_tracer; 810 elapsedTimer PSParallelCompact::_accumulated_time; 811 unsigned int PSParallelCompact::_total_invocations = 0; 812 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 813 jlong PSParallelCompact::_time_of_last_gc = 0; 814 CollectorCounters* PSParallelCompact::_counters = NULL; 815 ParMarkBitMap PSParallelCompact::_mark_bitmap; 816 ParallelCompactData PSParallelCompact::_summary_data; 817 818 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 819 820 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 821 822 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 823 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 824 825 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 826 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 827 828 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); } 829 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); } 830 831 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 832 833 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { 834 mark_and_push(_compaction_manager, p); 835 } 836 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 837 838 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 839 klass->oops_do(_mark_and_push_closure); 840 } 841 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 842 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 843 } 844 845 void PSParallelCompact::post_initialize() { 846 ParallelScavengeHeap* heap = gc_heap(); 847 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 848 849 MemRegion mr = heap->reserved_region(); 850 _ref_processor = 851 new ReferenceProcessor(mr, // span 852 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 853 (int) ParallelGCThreads, // mt processing degree 854 true, // mt discovery 855 (int) ParallelGCThreads, // mt discovery degree 856 true, // atomic_discovery 857 &_is_alive_closure); // non-header is alive closure 858 _counters = new CollectorCounters("PSParallelCompact", 1); 859 860 // Initialize static fields in ParCompactionManager. 861 ParCompactionManager::initialize(mark_bitmap()); 862 } 863 864 bool PSParallelCompact::initialize() { 865 ParallelScavengeHeap* heap = gc_heap(); 866 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 867 MemRegion mr = heap->reserved_region(); 868 869 // Was the old gen get allocated successfully? 870 if (!heap->old_gen()->is_allocated()) { 871 return false; 872 } 873 874 initialize_space_info(); 875 initialize_dead_wood_limiter(); 876 877 if (!_mark_bitmap.initialize(mr)) { 878 vm_shutdown_during_initialization( 879 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 880 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 881 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 882 return false; 883 } 884 885 if (!_summary_data.initialize(mr)) { 886 vm_shutdown_during_initialization( 887 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 888 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 889 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 890 return false; 891 } 892 893 return true; 894 } 895 896 void PSParallelCompact::initialize_space_info() 897 { 898 memset(&_space_info, 0, sizeof(_space_info)); 899 900 ParallelScavengeHeap* heap = gc_heap(); 901 PSYoungGen* young_gen = heap->young_gen(); 902 903 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 904 _space_info[eden_space_id].set_space(young_gen->eden_space()); 905 _space_info[from_space_id].set_space(young_gen->from_space()); 906 _space_info[to_space_id].set_space(young_gen->to_space()); 907 908 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 909 } 910 911 void PSParallelCompact::initialize_dead_wood_limiter() 912 { 913 const size_t max = 100; 914 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 915 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 916 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 917 DEBUG_ONLY(_dwl_initialized = true;) 918 _dwl_adjustment = normal_distribution(1.0); 919 } 920 921 // Simple class for storing info about the heap at the start of GC, to be used 922 // after GC for comparison/printing. 923 class PreGCValues { 924 public: 925 PreGCValues() { } 926 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 927 928 void fill(ParallelScavengeHeap* heap) { 929 _heap_used = heap->used(); 930 _young_gen_used = heap->young_gen()->used_in_bytes(); 931 _old_gen_used = heap->old_gen()->used_in_bytes(); 932 _metadata_used = MetaspaceAux::used_bytes(); 933 }; 934 935 size_t heap_used() const { return _heap_used; } 936 size_t young_gen_used() const { return _young_gen_used; } 937 size_t old_gen_used() const { return _old_gen_used; } 938 size_t metadata_used() const { return _metadata_used; } 939 940 private: 941 size_t _heap_used; 942 size_t _young_gen_used; 943 size_t _old_gen_used; 944 size_t _metadata_used; 945 }; 946 947 void 948 PSParallelCompact::clear_data_covering_space(SpaceId id) 949 { 950 // At this point, top is the value before GC, new_top() is the value that will 951 // be set at the end of GC. The marking bitmap is cleared to top; nothing 952 // should be marked above top. The summary data is cleared to the larger of 953 // top & new_top. 954 MutableSpace* const space = _space_info[id].space(); 955 HeapWord* const bot = space->bottom(); 956 HeapWord* const top = space->top(); 957 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 958 959 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 960 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 961 _mark_bitmap.clear_range(beg_bit, end_bit); 962 963 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 964 const size_t end_region = 965 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 966 _summary_data.clear_range(beg_region, end_region); 967 968 // Clear the data used to 'split' regions. 969 SplitInfo& split_info = _space_info[id].split_info(); 970 if (split_info.is_valid()) { 971 split_info.clear(); 972 } 973 DEBUG_ONLY(split_info.verify_clear();) 974 } 975 976 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 977 { 978 // Update the from & to space pointers in space_info, since they are swapped 979 // at each young gen gc. Do the update unconditionally (even though a 980 // promotion failure does not swap spaces) because an unknown number of minor 981 // collections will have swapped the spaces an unknown number of times. 982 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 983 ParallelScavengeHeap* heap = gc_heap(); 984 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 985 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 986 987 pre_gc_values->fill(heap); 988 989 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 990 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 991 992 // Increment the invocation count 993 heap->increment_total_collections(true); 994 995 // We need to track unique mark sweep invocations as well. 996 _total_invocations++; 997 998 heap->print_heap_before_gc(); 999 heap->trace_heap_before_gc(&_gc_tracer); 1000 1001 // Fill in TLABs 1002 heap->accumulate_statistics_all_tlabs(); 1003 heap->ensure_parsability(true); // retire TLABs 1004 1005 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 1006 HandleMark hm; // Discard invalid handles created during verification 1007 Universe::verify(" VerifyBeforeGC:"); 1008 } 1009 1010 // Verify object start arrays 1011 if (VerifyObjectStartArray && 1012 VerifyBeforeGC) { 1013 heap->old_gen()->verify_object_start_array(); 1014 } 1015 1016 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1017 DEBUG_ONLY(summary_data().verify_clear();) 1018 1019 // Have worker threads release resources the next time they run a task. 1020 gc_task_manager()->release_all_resources(); 1021 } 1022 1023 void PSParallelCompact::post_compact() 1024 { 1025 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1026 1027 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1028 // Clear the marking bitmap, summary data and split info. 1029 clear_data_covering_space(SpaceId(id)); 1030 // Update top(). Must be done after clearing the bitmap and summary data. 1031 _space_info[id].publish_new_top(); 1032 } 1033 1034 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1035 MutableSpace* const from_space = _space_info[from_space_id].space(); 1036 MutableSpace* const to_space = _space_info[to_space_id].space(); 1037 1038 ParallelScavengeHeap* heap = gc_heap(); 1039 bool eden_empty = eden_space->is_empty(); 1040 if (!eden_empty) { 1041 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1042 heap->young_gen(), heap->old_gen()); 1043 } 1044 1045 // Update heap occupancy information which is used as input to the soft ref 1046 // clearing policy at the next gc. 1047 Universe::update_heap_info_at_gc(); 1048 1049 bool young_gen_empty = eden_empty && from_space->is_empty() && 1050 to_space->is_empty(); 1051 1052 BarrierSet* bs = heap->barrier_set(); 1053 if (bs->is_a(BarrierSet::ModRef)) { 1054 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1055 MemRegion old_mr = heap->old_gen()->reserved(); 1056 1057 if (young_gen_empty) { 1058 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1059 } else { 1060 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1061 } 1062 } 1063 1064 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1065 ClassLoaderDataGraph::purge(); 1066 MetaspaceAux::verify_metrics(); 1067 1068 Threads::gc_epilogue(); 1069 CodeCache::gc_epilogue(); 1070 JvmtiExport::gc_epilogue(); 1071 1072 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1073 1074 ref_processor()->enqueue_discovered_references(NULL); 1075 1076 if (ZapUnusedHeapArea) { 1077 heap->gen_mangle_unused_area(); 1078 } 1079 1080 // Update time of last GC 1081 reset_millis_since_last_gc(); 1082 } 1083 1084 HeapWord* 1085 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1086 bool maximum_compaction) 1087 { 1088 const size_t region_size = ParallelCompactData::RegionSize; 1089 const ParallelCompactData& sd = summary_data(); 1090 1091 const MutableSpace* const space = _space_info[id].space(); 1092 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1093 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1094 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1095 1096 // Skip full regions at the beginning of the space--they are necessarily part 1097 // of the dense prefix. 1098 size_t full_count = 0; 1099 const RegionData* cp; 1100 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1101 ++full_count; 1102 } 1103 1104 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1105 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1106 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1107 if (maximum_compaction || cp == end_cp || interval_ended) { 1108 _maximum_compaction_gc_num = total_invocations(); 1109 return sd.region_to_addr(cp); 1110 } 1111 1112 HeapWord* const new_top = _space_info[id].new_top(); 1113 const size_t space_live = pointer_delta(new_top, space->bottom()); 1114 const size_t space_used = space->used_in_words(); 1115 const size_t space_capacity = space->capacity_in_words(); 1116 1117 const double cur_density = double(space_live) / space_capacity; 1118 const double deadwood_density = 1119 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1120 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1121 1122 if (TraceParallelOldGCDensePrefix) { 1123 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1124 cur_density, deadwood_density, deadwood_goal); 1125 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1126 "space_cap=" SIZE_FORMAT, 1127 space_live, space_used, 1128 space_capacity); 1129 } 1130 1131 // XXX - Use binary search? 1132 HeapWord* dense_prefix = sd.region_to_addr(cp); 1133 const RegionData* full_cp = cp; 1134 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1135 while (cp < end_cp) { 1136 HeapWord* region_destination = cp->destination(); 1137 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1138 if (TraceParallelOldGCDensePrefix && Verbose) { 1139 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1140 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1141 sd.region(cp), region_destination, 1142 dense_prefix, cur_deadwood); 1143 } 1144 1145 if (cur_deadwood >= deadwood_goal) { 1146 // Found the region that has the correct amount of deadwood to the left. 1147 // This typically occurs after crossing a fairly sparse set of regions, so 1148 // iterate backwards over those sparse regions, looking for the region 1149 // that has the lowest density of live objects 'to the right.' 1150 size_t space_to_left = sd.region(cp) * region_size; 1151 size_t live_to_left = space_to_left - cur_deadwood; 1152 size_t space_to_right = space_capacity - space_to_left; 1153 size_t live_to_right = space_live - live_to_left; 1154 double density_to_right = double(live_to_right) / space_to_right; 1155 while (cp > full_cp) { 1156 --cp; 1157 const size_t prev_region_live_to_right = live_to_right - 1158 cp->data_size(); 1159 const size_t prev_region_space_to_right = space_to_right + region_size; 1160 double prev_region_density_to_right = 1161 double(prev_region_live_to_right) / prev_region_space_to_right; 1162 if (density_to_right <= prev_region_density_to_right) { 1163 return dense_prefix; 1164 } 1165 if (TraceParallelOldGCDensePrefix && Verbose) { 1166 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1167 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1168 prev_region_density_to_right); 1169 } 1170 dense_prefix -= region_size; 1171 live_to_right = prev_region_live_to_right; 1172 space_to_right = prev_region_space_to_right; 1173 density_to_right = prev_region_density_to_right; 1174 } 1175 return dense_prefix; 1176 } 1177 1178 dense_prefix += region_size; 1179 ++cp; 1180 } 1181 1182 return dense_prefix; 1183 } 1184 1185 #ifndef PRODUCT 1186 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1187 const SpaceId id, 1188 const bool maximum_compaction, 1189 HeapWord* const addr) 1190 { 1191 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1192 RegionData* const cp = summary_data().region(region_idx); 1193 const MutableSpace* const space = _space_info[id].space(); 1194 HeapWord* const new_top = _space_info[id].new_top(); 1195 1196 const size_t space_live = pointer_delta(new_top, space->bottom()); 1197 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1198 const size_t space_cap = space->capacity_in_words(); 1199 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1200 const size_t live_to_right = new_top - cp->destination(); 1201 const size_t dead_to_right = space->top() - addr - live_to_right; 1202 1203 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1204 "spl=" SIZE_FORMAT " " 1205 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1206 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1207 " ratio=%10.8f", 1208 algorithm, addr, region_idx, 1209 space_live, 1210 dead_to_left, dead_to_left_pct, 1211 dead_to_right, live_to_right, 1212 double(dead_to_right) / live_to_right); 1213 } 1214 #endif // #ifndef PRODUCT 1215 1216 // Return a fraction indicating how much of the generation can be treated as 1217 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1218 // based on the density of live objects in the generation to determine a limit, 1219 // which is then adjusted so the return value is min_percent when the density is 1220 // 1. 1221 // 1222 // The following table shows some return values for a different values of the 1223 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1224 // min_percent is 1. 1225 // 1226 // fraction allowed as dead wood 1227 // ----------------------------------------------------------------- 1228 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1229 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1230 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1231 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1232 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1233 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1234 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1235 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1236 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1237 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1238 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1239 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1240 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1241 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1242 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1243 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1244 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1245 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1246 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1247 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1248 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1249 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1250 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1251 1252 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1253 { 1254 assert(_dwl_initialized, "uninitialized"); 1255 1256 // The raw limit is the value of the normal distribution at x = density. 1257 const double raw_limit = normal_distribution(density); 1258 1259 // Adjust the raw limit so it becomes the minimum when the density is 1. 1260 // 1261 // First subtract the adjustment value (which is simply the precomputed value 1262 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1263 // Then add the minimum value, so the minimum is returned when the density is 1264 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1265 const double min = double(min_percent) / 100.0; 1266 const double limit = raw_limit - _dwl_adjustment + min; 1267 return MAX2(limit, 0.0); 1268 } 1269 1270 ParallelCompactData::RegionData* 1271 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1272 const RegionData* end) 1273 { 1274 const size_t region_size = ParallelCompactData::RegionSize; 1275 ParallelCompactData& sd = summary_data(); 1276 size_t left = sd.region(beg); 1277 size_t right = end > beg ? sd.region(end) - 1 : left; 1278 1279 // Binary search. 1280 while (left < right) { 1281 // Equivalent to (left + right) / 2, but does not overflow. 1282 const size_t middle = left + (right - left) / 2; 1283 RegionData* const middle_ptr = sd.region(middle); 1284 HeapWord* const dest = middle_ptr->destination(); 1285 HeapWord* const addr = sd.region_to_addr(middle); 1286 assert(dest != NULL, "sanity"); 1287 assert(dest <= addr, "must move left"); 1288 1289 if (middle > left && dest < addr) { 1290 right = middle - 1; 1291 } else if (middle < right && middle_ptr->data_size() == region_size) { 1292 left = middle + 1; 1293 } else { 1294 return middle_ptr; 1295 } 1296 } 1297 return sd.region(left); 1298 } 1299 1300 ParallelCompactData::RegionData* 1301 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1302 const RegionData* end, 1303 size_t dead_words) 1304 { 1305 ParallelCompactData& sd = summary_data(); 1306 size_t left = sd.region(beg); 1307 size_t right = end > beg ? sd.region(end) - 1 : left; 1308 1309 // Binary search. 1310 while (left < right) { 1311 // Equivalent to (left + right) / 2, but does not overflow. 1312 const size_t middle = left + (right - left) / 2; 1313 RegionData* const middle_ptr = sd.region(middle); 1314 HeapWord* const dest = middle_ptr->destination(); 1315 HeapWord* const addr = sd.region_to_addr(middle); 1316 assert(dest != NULL, "sanity"); 1317 assert(dest <= addr, "must move left"); 1318 1319 const size_t dead_to_left = pointer_delta(addr, dest); 1320 if (middle > left && dead_to_left > dead_words) { 1321 right = middle - 1; 1322 } else if (middle < right && dead_to_left < dead_words) { 1323 left = middle + 1; 1324 } else { 1325 return middle_ptr; 1326 } 1327 } 1328 return sd.region(left); 1329 } 1330 1331 // The result is valid during the summary phase, after the initial summarization 1332 // of each space into itself, and before final summarization. 1333 inline double 1334 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1335 HeapWord* const bottom, 1336 HeapWord* const top, 1337 HeapWord* const new_top) 1338 { 1339 ParallelCompactData& sd = summary_data(); 1340 1341 assert(cp != NULL, "sanity"); 1342 assert(bottom != NULL, "sanity"); 1343 assert(top != NULL, "sanity"); 1344 assert(new_top != NULL, "sanity"); 1345 assert(top >= new_top, "summary data problem?"); 1346 assert(new_top > bottom, "space is empty; should not be here"); 1347 assert(new_top >= cp->destination(), "sanity"); 1348 assert(top >= sd.region_to_addr(cp), "sanity"); 1349 1350 HeapWord* const destination = cp->destination(); 1351 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1352 const size_t compacted_region_live = pointer_delta(new_top, destination); 1353 const size_t compacted_region_used = pointer_delta(top, 1354 sd.region_to_addr(cp)); 1355 const size_t reclaimable = compacted_region_used - compacted_region_live; 1356 1357 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1358 return double(reclaimable) / divisor; 1359 } 1360 1361 // Return the address of the end of the dense prefix, a.k.a. the start of the 1362 // compacted region. The address is always on a region boundary. 1363 // 1364 // Completely full regions at the left are skipped, since no compaction can 1365 // occur in those regions. Then the maximum amount of dead wood to allow is 1366 // computed, based on the density (amount live / capacity) of the generation; 1367 // the region with approximately that amount of dead space to the left is 1368 // identified as the limit region. Regions between the last completely full 1369 // region and the limit region are scanned and the one that has the best 1370 // (maximum) reclaimed_ratio() is selected. 1371 HeapWord* 1372 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1373 bool maximum_compaction) 1374 { 1375 if (ParallelOldGCSplitALot) { 1376 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1377 // The value was chosen to provoke splitting a young gen space; use it. 1378 return _space_info[id].dense_prefix(); 1379 } 1380 } 1381 1382 const size_t region_size = ParallelCompactData::RegionSize; 1383 const ParallelCompactData& sd = summary_data(); 1384 1385 const MutableSpace* const space = _space_info[id].space(); 1386 HeapWord* const top = space->top(); 1387 HeapWord* const top_aligned_up = sd.region_align_up(top); 1388 HeapWord* const new_top = _space_info[id].new_top(); 1389 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1390 HeapWord* const bottom = space->bottom(); 1391 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1392 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1393 const RegionData* const new_top_cp = 1394 sd.addr_to_region_ptr(new_top_aligned_up); 1395 1396 // Skip full regions at the beginning of the space--they are necessarily part 1397 // of the dense prefix. 1398 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1399 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1400 space->is_empty(), "no dead space allowed to the left"); 1401 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1402 "region must have dead space"); 1403 1404 // The gc number is saved whenever a maximum compaction is done, and used to 1405 // determine when the maximum compaction interval has expired. This avoids 1406 // successive max compactions for different reasons. 1407 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1408 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1409 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1410 total_invocations() == HeapFirstMaximumCompactionCount; 1411 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1412 _maximum_compaction_gc_num = total_invocations(); 1413 return sd.region_to_addr(full_cp); 1414 } 1415 1416 const size_t space_live = pointer_delta(new_top, bottom); 1417 const size_t space_used = space->used_in_words(); 1418 const size_t space_capacity = space->capacity_in_words(); 1419 1420 const double density = double(space_live) / double(space_capacity); 1421 const size_t min_percent_free = MarkSweepDeadRatio; 1422 const double limiter = dead_wood_limiter(density, min_percent_free); 1423 const size_t dead_wood_max = space_used - space_live; 1424 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1425 dead_wood_max); 1426 1427 if (TraceParallelOldGCDensePrefix) { 1428 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1429 "space_cap=" SIZE_FORMAT, 1430 space_live, space_used, 1431 space_capacity); 1432 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1433 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1434 density, min_percent_free, limiter, 1435 dead_wood_max, dead_wood_limit); 1436 } 1437 1438 // Locate the region with the desired amount of dead space to the left. 1439 const RegionData* const limit_cp = 1440 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1441 1442 // Scan from the first region with dead space to the limit region and find the 1443 // one with the best (largest) reclaimed ratio. 1444 double best_ratio = 0.0; 1445 const RegionData* best_cp = full_cp; 1446 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1447 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1448 if (tmp_ratio > best_ratio) { 1449 best_cp = cp; 1450 best_ratio = tmp_ratio; 1451 } 1452 } 1453 1454 #if 0 1455 // Something to consider: if the region with the best ratio is 'close to' the 1456 // first region w/free space, choose the first region with free space 1457 // ("first-free"). The first-free region is usually near the start of the 1458 // heap, which means we are copying most of the heap already, so copy a bit 1459 // more to get complete compaction. 1460 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1461 _maximum_compaction_gc_num = total_invocations(); 1462 best_cp = full_cp; 1463 } 1464 #endif // #if 0 1465 1466 return sd.region_to_addr(best_cp); 1467 } 1468 1469 #ifndef PRODUCT 1470 void 1471 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1472 size_t words) 1473 { 1474 if (TraceParallelOldGCSummaryPhase) { 1475 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1476 SIZE_FORMAT, start, start + words, words); 1477 } 1478 1479 ObjectStartArray* const start_array = _space_info[id].start_array(); 1480 CollectedHeap::fill_with_objects(start, words); 1481 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1482 _mark_bitmap.mark_obj(p, words); 1483 _summary_data.add_obj(p, words); 1484 start_array->allocate_block(p); 1485 } 1486 } 1487 1488 void 1489 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1490 { 1491 ParallelCompactData& sd = summary_data(); 1492 MutableSpace* space = _space_info[id].space(); 1493 1494 // Find the source and destination start addresses. 1495 HeapWord* const src_addr = sd.region_align_down(start); 1496 HeapWord* dst_addr; 1497 if (src_addr < start) { 1498 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1499 } else if (src_addr > space->bottom()) { 1500 // The start (the original top() value) is aligned to a region boundary so 1501 // the associated region does not have a destination. Compute the 1502 // destination from the previous region. 1503 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1504 dst_addr = cp->destination() + cp->data_size(); 1505 } else { 1506 // Filling the entire space. 1507 dst_addr = space->bottom(); 1508 } 1509 assert(dst_addr != NULL, "sanity"); 1510 1511 // Update the summary data. 1512 bool result = _summary_data.summarize(_space_info[id].split_info(), 1513 src_addr, space->top(), NULL, 1514 dst_addr, space->end(), 1515 _space_info[id].new_top_addr()); 1516 assert(result, "should not fail: bad filler object size"); 1517 } 1518 1519 void 1520 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1521 { 1522 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1523 return; 1524 } 1525 1526 MutableSpace* const space = _space_info[id].space(); 1527 if (space->is_empty()) { 1528 HeapWord* b = space->bottom(); 1529 HeapWord* t = b + space->capacity_in_words() / 2; 1530 space->set_top(t); 1531 if (ZapUnusedHeapArea) { 1532 space->set_top_for_allocations(); 1533 } 1534 1535 size_t min_size = CollectedHeap::min_fill_size(); 1536 size_t obj_len = min_size; 1537 while (b + obj_len <= t) { 1538 CollectedHeap::fill_with_object(b, obj_len); 1539 mark_bitmap()->mark_obj(b, obj_len); 1540 summary_data().add_obj(b, obj_len); 1541 b += obj_len; 1542 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1543 } 1544 if (b < t) { 1545 // The loop didn't completely fill to t (top); adjust top downward. 1546 space->set_top(b); 1547 if (ZapUnusedHeapArea) { 1548 space->set_top_for_allocations(); 1549 } 1550 } 1551 1552 HeapWord** nta = _space_info[id].new_top_addr(); 1553 bool result = summary_data().summarize(_space_info[id].split_info(), 1554 space->bottom(), space->top(), NULL, 1555 space->bottom(), space->end(), nta); 1556 assert(result, "space must fit into itself"); 1557 } 1558 } 1559 1560 void 1561 PSParallelCompact::provoke_split(bool & max_compaction) 1562 { 1563 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1564 return; 1565 } 1566 1567 const size_t region_size = ParallelCompactData::RegionSize; 1568 ParallelCompactData& sd = summary_data(); 1569 1570 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1571 MutableSpace* const from_space = _space_info[from_space_id].space(); 1572 const size_t eden_live = pointer_delta(eden_space->top(), 1573 _space_info[eden_space_id].new_top()); 1574 const size_t from_live = pointer_delta(from_space->top(), 1575 _space_info[from_space_id].new_top()); 1576 1577 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1578 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1579 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1580 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1581 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1582 1583 // Choose the space to split; need at least 2 regions live (or fillable). 1584 SpaceId id; 1585 MutableSpace* space; 1586 size_t live_words; 1587 size_t fill_words; 1588 if (eden_live + eden_fillable >= region_size * 2) { 1589 id = eden_space_id; 1590 space = eden_space; 1591 live_words = eden_live; 1592 fill_words = eden_fillable; 1593 } else if (from_live + from_fillable >= region_size * 2) { 1594 id = from_space_id; 1595 space = from_space; 1596 live_words = from_live; 1597 fill_words = from_fillable; 1598 } else { 1599 return; // Give up. 1600 } 1601 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1602 1603 if (live_words < region_size * 2) { 1604 // Fill from top() to end() w/live objects of mixed sizes. 1605 HeapWord* const fill_start = space->top(); 1606 live_words += fill_words; 1607 1608 space->set_top(fill_start + fill_words); 1609 if (ZapUnusedHeapArea) { 1610 space->set_top_for_allocations(); 1611 } 1612 1613 HeapWord* cur_addr = fill_start; 1614 while (fill_words > 0) { 1615 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1616 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1617 if (fill_words - cur_size < min_fill_size) { 1618 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1619 } 1620 1621 CollectedHeap::fill_with_object(cur_addr, cur_size); 1622 mark_bitmap()->mark_obj(cur_addr, cur_size); 1623 sd.add_obj(cur_addr, cur_size); 1624 1625 cur_addr += cur_size; 1626 fill_words -= cur_size; 1627 } 1628 1629 summarize_new_objects(id, fill_start); 1630 } 1631 1632 max_compaction = false; 1633 1634 // Manipulate the old gen so that it has room for about half of the live data 1635 // in the target young gen space (live_words / 2). 1636 id = old_space_id; 1637 space = _space_info[id].space(); 1638 const size_t free_at_end = space->free_in_words(); 1639 const size_t free_target = align_object_size(live_words / 2); 1640 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1641 1642 if (free_at_end >= free_target + min_fill_size) { 1643 // Fill space above top() and set the dense prefix so everything survives. 1644 HeapWord* const fill_start = space->top(); 1645 const size_t fill_size = free_at_end - free_target; 1646 space->set_top(space->top() + fill_size); 1647 if (ZapUnusedHeapArea) { 1648 space->set_top_for_allocations(); 1649 } 1650 fill_with_live_objects(id, fill_start, fill_size); 1651 summarize_new_objects(id, fill_start); 1652 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1653 } else if (dead + free_at_end > free_target) { 1654 // Find a dense prefix that makes the right amount of space available. 1655 HeapWord* cur = sd.region_align_down(space->top()); 1656 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1657 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1658 while (dead_to_right < free_target) { 1659 cur -= region_size; 1660 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1661 dead_to_right = pointer_delta(space->end(), cur_destination); 1662 } 1663 _space_info[id].set_dense_prefix(cur); 1664 } 1665 } 1666 #endif // #ifndef PRODUCT 1667 1668 void PSParallelCompact::summarize_spaces_quick() 1669 { 1670 for (unsigned int i = 0; i < last_space_id; ++i) { 1671 const MutableSpace* space = _space_info[i].space(); 1672 HeapWord** nta = _space_info[i].new_top_addr(); 1673 bool result = _summary_data.summarize(_space_info[i].split_info(), 1674 space->bottom(), space->top(), NULL, 1675 space->bottom(), space->end(), nta); 1676 assert(result, "space must fit into itself"); 1677 _space_info[i].set_dense_prefix(space->bottom()); 1678 } 1679 1680 #ifndef PRODUCT 1681 if (ParallelOldGCSplitALot) { 1682 provoke_split_fill_survivor(to_space_id); 1683 } 1684 #endif // #ifndef PRODUCT 1685 } 1686 1687 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1688 { 1689 HeapWord* const dense_prefix_end = dense_prefix(id); 1690 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1691 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1692 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1693 // Only enough dead space is filled so that any remaining dead space to the 1694 // left is larger than the minimum filler object. (The remainder is filled 1695 // during the copy/update phase.) 1696 // 1697 // The size of the dead space to the right of the boundary is not a 1698 // concern, since compaction will be able to use whatever space is 1699 // available. 1700 // 1701 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1702 // surrounds the space to be filled with an object. 1703 // 1704 // In the 32-bit VM, each bit represents two 32-bit words: 1705 // +---+ 1706 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1707 // end_bits: ... x x x | 0 | || 0 x x ... 1708 // +---+ 1709 // 1710 // In the 64-bit VM, each bit represents one 64-bit word: 1711 // +------------+ 1712 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1713 // end_bits: ... x x 1 | 0 || 0 | x x ... 1714 // +------------+ 1715 // +-------+ 1716 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1717 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1718 // +-------+ 1719 // +-----------+ 1720 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1721 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1722 // +-----------+ 1723 // +-------+ 1724 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1725 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1726 // +-------+ 1727 1728 // Initially assume case a, c or e will apply. 1729 size_t obj_len = CollectedHeap::min_fill_size(); 1730 HeapWord* obj_beg = dense_prefix_end - obj_len; 1731 1732 #ifdef _LP64 1733 if (MinObjAlignment > 1) { // object alignment > heap word size 1734 // Cases a, c or e. 1735 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1736 // Case b above. 1737 obj_beg = dense_prefix_end - 1; 1738 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1739 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1740 // Case d above. 1741 obj_beg = dense_prefix_end - 3; 1742 obj_len = 3; 1743 } 1744 #endif // #ifdef _LP64 1745 1746 CollectedHeap::fill_with_object(obj_beg, obj_len); 1747 _mark_bitmap.mark_obj(obj_beg, obj_len); 1748 _summary_data.add_obj(obj_beg, obj_len); 1749 assert(start_array(id) != NULL, "sanity"); 1750 start_array(id)->allocate_block(obj_beg); 1751 } 1752 } 1753 1754 void 1755 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1756 { 1757 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1758 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1759 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1760 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1761 cur->set_source_region(0); 1762 } 1763 } 1764 1765 void 1766 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1767 { 1768 assert(id < last_space_id, "id out of range"); 1769 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1770 ParallelOldGCSplitALot && id == old_space_id, 1771 "should have been reset in summarize_spaces_quick()"); 1772 1773 const MutableSpace* space = _space_info[id].space(); 1774 if (_space_info[id].new_top() != space->bottom()) { 1775 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1776 _space_info[id].set_dense_prefix(dense_prefix_end); 1777 1778 #ifndef PRODUCT 1779 if (TraceParallelOldGCDensePrefix) { 1780 print_dense_prefix_stats("ratio", id, maximum_compaction, 1781 dense_prefix_end); 1782 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1783 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1784 } 1785 #endif // #ifndef PRODUCT 1786 1787 // Recompute the summary data, taking into account the dense prefix. If 1788 // every last byte will be reclaimed, then the existing summary data which 1789 // compacts everything can be left in place. 1790 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1791 // If dead space crosses the dense prefix boundary, it is (at least 1792 // partially) filled with a dummy object, marked live and added to the 1793 // summary data. This simplifies the copy/update phase and must be done 1794 // before the final locations of objects are determined, to prevent 1795 // leaving a fragment of dead space that is too small to fill. 1796 fill_dense_prefix_end(id); 1797 1798 // Compute the destination of each Region, and thus each object. 1799 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1800 _summary_data.summarize(_space_info[id].split_info(), 1801 dense_prefix_end, space->top(), NULL, 1802 dense_prefix_end, space->end(), 1803 _space_info[id].new_top_addr()); 1804 } 1805 } 1806 1807 if (TraceParallelOldGCSummaryPhase) { 1808 const size_t region_size = ParallelCompactData::RegionSize; 1809 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1810 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1811 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1812 HeapWord* const new_top = _space_info[id].new_top(); 1813 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1814 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1815 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1816 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1817 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1818 id, space->capacity_in_words(), dense_prefix_end, 1819 dp_region, dp_words / region_size, 1820 cr_words / region_size, new_top); 1821 } 1822 } 1823 1824 #ifndef PRODUCT 1825 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1826 HeapWord* dst_beg, HeapWord* dst_end, 1827 SpaceId src_space_id, 1828 HeapWord* src_beg, HeapWord* src_end) 1829 { 1830 if (TraceParallelOldGCSummaryPhase) { 1831 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1832 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1833 SIZE_FORMAT "-" SIZE_FORMAT " " 1834 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1835 SIZE_FORMAT "-" SIZE_FORMAT, 1836 src_space_id, space_names[src_space_id], 1837 dst_space_id, space_names[dst_space_id], 1838 src_beg, src_end, 1839 _summary_data.addr_to_region_idx(src_beg), 1840 _summary_data.addr_to_region_idx(src_end), 1841 dst_beg, dst_end, 1842 _summary_data.addr_to_region_idx(dst_beg), 1843 _summary_data.addr_to_region_idx(dst_end)); 1844 } 1845 } 1846 #endif // #ifndef PRODUCT 1847 1848 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1849 bool maximum_compaction) 1850 { 1851 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1852 // trace("2"); 1853 1854 #ifdef ASSERT 1855 if (TraceParallelOldGCMarkingPhase) { 1856 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1857 "add_obj_bytes=" SIZE_FORMAT, 1858 add_obj_count, add_obj_size * HeapWordSize); 1859 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1860 "mark_bitmap_bytes=" SIZE_FORMAT, 1861 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1862 } 1863 #endif // #ifdef ASSERT 1864 1865 // Quick summarization of each space into itself, to see how much is live. 1866 summarize_spaces_quick(); 1867 1868 if (TraceParallelOldGCSummaryPhase) { 1869 tty->print_cr("summary_phase: after summarizing each space to self"); 1870 Universe::print(); 1871 NOT_PRODUCT(print_region_ranges()); 1872 if (Verbose) { 1873 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1874 } 1875 } 1876 1877 // The amount of live data that will end up in old space (assuming it fits). 1878 size_t old_space_total_live = 0; 1879 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1880 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1881 _space_info[id].space()->bottom()); 1882 } 1883 1884 MutableSpace* const old_space = _space_info[old_space_id].space(); 1885 const size_t old_capacity = old_space->capacity_in_words(); 1886 if (old_space_total_live > old_capacity) { 1887 // XXX - should also try to expand 1888 maximum_compaction = true; 1889 } 1890 #ifndef PRODUCT 1891 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1892 provoke_split(maximum_compaction); 1893 } 1894 #endif // #ifndef PRODUCT 1895 1896 // Old generations. 1897 summarize_space(old_space_id, maximum_compaction); 1898 1899 // Summarize the remaining spaces in the young gen. The initial target space 1900 // is the old gen. If a space does not fit entirely into the target, then the 1901 // remainder is compacted into the space itself and that space becomes the new 1902 // target. 1903 SpaceId dst_space_id = old_space_id; 1904 HeapWord* dst_space_end = old_space->end(); 1905 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1906 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1907 const MutableSpace* space = _space_info[id].space(); 1908 const size_t live = pointer_delta(_space_info[id].new_top(), 1909 space->bottom()); 1910 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1911 1912 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1913 SpaceId(id), space->bottom(), space->top());) 1914 if (live > 0 && live <= available) { 1915 // All the live data will fit. 1916 bool done = _summary_data.summarize(_space_info[id].split_info(), 1917 space->bottom(), space->top(), 1918 NULL, 1919 *new_top_addr, dst_space_end, 1920 new_top_addr); 1921 assert(done, "space must fit into old gen"); 1922 1923 // Reset the new_top value for the space. 1924 _space_info[id].set_new_top(space->bottom()); 1925 } else if (live > 0) { 1926 // Attempt to fit part of the source space into the target space. 1927 HeapWord* next_src_addr = NULL; 1928 bool done = _summary_data.summarize(_space_info[id].split_info(), 1929 space->bottom(), space->top(), 1930 &next_src_addr, 1931 *new_top_addr, dst_space_end, 1932 new_top_addr); 1933 assert(!done, "space should not fit into old gen"); 1934 assert(next_src_addr != NULL, "sanity"); 1935 1936 // The source space becomes the new target, so the remainder is compacted 1937 // within the space itself. 1938 dst_space_id = SpaceId(id); 1939 dst_space_end = space->end(); 1940 new_top_addr = _space_info[id].new_top_addr(); 1941 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1942 space->bottom(), dst_space_end, 1943 SpaceId(id), next_src_addr, space->top());) 1944 done = _summary_data.summarize(_space_info[id].split_info(), 1945 next_src_addr, space->top(), 1946 NULL, 1947 space->bottom(), dst_space_end, 1948 new_top_addr); 1949 assert(done, "space must fit when compacted into itself"); 1950 assert(*new_top_addr <= space->top(), "usage should not grow"); 1951 } 1952 } 1953 1954 if (TraceParallelOldGCSummaryPhase) { 1955 tty->print_cr("summary_phase: after final summarization"); 1956 Universe::print(); 1957 NOT_PRODUCT(print_region_ranges()); 1958 if (Verbose) { 1959 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1960 } 1961 } 1962 } 1963 1964 // This method should contain all heap-specific policy for invoking a full 1965 // collection. invoke_no_policy() will only attempt to compact the heap; it 1966 // will do nothing further. If we need to bail out for policy reasons, scavenge 1967 // before full gc, or any other specialized behavior, it needs to be added here. 1968 // 1969 // Note that this method should only be called from the vm_thread while at a 1970 // safepoint. 1971 // 1972 // Note that the all_soft_refs_clear flag in the collector policy 1973 // may be true because this method can be called without intervening 1974 // activity. For example when the heap space is tight and full measure 1975 // are being taken to free space. 1976 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1977 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1978 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1979 "should be in vm thread"); 1980 1981 ParallelScavengeHeap* heap = gc_heap(); 1982 GCCause::Cause gc_cause = heap->gc_cause(); 1983 assert(!heap->is_gc_active(), "not reentrant"); 1984 1985 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1986 IsGCActiveMark mark; 1987 1988 if (ScavengeBeforeFullGC) { 1989 PSScavenge::invoke_no_policy(); 1990 } 1991 1992 const bool clear_all_soft_refs = 1993 heap->collector_policy()->should_clear_all_soft_refs(); 1994 1995 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1996 maximum_heap_compaction); 1997 } 1998 1999 // This method contains no policy. You should probably 2000 // be calling invoke() instead. 2001 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 2002 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 2003 assert(ref_processor() != NULL, "Sanity"); 2004 2005 if (GC_locker::check_active_before_gc()) { 2006 return false; 2007 } 2008 2009 ParallelScavengeHeap* heap = gc_heap(); 2010 2011 _gc_timer.register_gc_start(); 2012 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 2013 2014 TimeStamp marking_start; 2015 TimeStamp compaction_start; 2016 TimeStamp collection_exit; 2017 2018 GCCause::Cause gc_cause = heap->gc_cause(); 2019 PSYoungGen* young_gen = heap->young_gen(); 2020 PSOldGen* old_gen = heap->old_gen(); 2021 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2022 2023 // The scope of casr should end after code that can change 2024 // CollectorPolicy::_should_clear_all_soft_refs. 2025 ClearedAllSoftRefs casr(maximum_heap_compaction, 2026 heap->collector_policy()); 2027 2028 if (ZapUnusedHeapArea) { 2029 // Save information needed to minimize mangling 2030 heap->record_gen_tops_before_GC(); 2031 } 2032 2033 heap->pre_full_gc_dump(&_gc_timer); 2034 2035 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2036 2037 // Make sure data structures are sane, make the heap parsable, and do other 2038 // miscellaneous bookkeeping. 2039 PreGCValues pre_gc_values; 2040 pre_compact(&pre_gc_values); 2041 2042 // Get the compaction manager reserved for the VM thread. 2043 ParCompactionManager* const vmthread_cm = 2044 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2045 2046 // Place after pre_compact() where the number of invocations is incremented. 2047 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2048 2049 { 2050 ResourceMark rm; 2051 HandleMark hm; 2052 2053 // Set the number of GC threads to be used in this collection 2054 gc_task_manager()->set_active_gang(); 2055 gc_task_manager()->task_idle_workers(); 2056 heap->set_par_threads(gc_task_manager()->active_workers()); 2057 2058 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2059 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2060 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id()); 2061 TraceCollectorStats tcs(counters()); 2062 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2063 2064 if (TraceGen1Time) accumulated_time()->start(); 2065 2066 // Let the size policy know we're starting 2067 size_policy->major_collection_begin(); 2068 2069 CodeCache::gc_prologue(); 2070 Threads::gc_prologue(); 2071 2072 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2073 2074 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 2075 ref_processor()->setup_policy(maximum_heap_compaction); 2076 2077 bool marked_for_unloading = false; 2078 2079 marking_start.update(); 2080 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2081 2082 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2083 && gc_cause == GCCause::_java_lang_system_gc; 2084 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2085 2086 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2087 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2088 2089 // adjust_roots() updates Universe::_intArrayKlassObj which is 2090 // needed by the compaction for filling holes in the dense prefix. 2091 adjust_roots(); 2092 2093 compaction_start.update(); 2094 compact(); 2095 2096 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2097 // done before resizing. 2098 post_compact(); 2099 2100 // Let the size policy know we're done 2101 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2102 2103 if (UseAdaptiveSizePolicy) { 2104 if (PrintAdaptiveSizePolicy) { 2105 gclog_or_tty->print("AdaptiveSizeStart: "); 2106 gclog_or_tty->stamp(); 2107 gclog_or_tty->print_cr(" collection: %d ", 2108 heap->total_collections()); 2109 if (Verbose) { 2110 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d", 2111 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2112 } 2113 } 2114 2115 // Don't check if the size_policy is ready here. Let 2116 // the size_policy check that internally. 2117 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2118 ((gc_cause != GCCause::_java_lang_system_gc) || 2119 UseAdaptiveSizePolicyWithSystemGC)) { 2120 // Calculate optimal free space amounts 2121 assert(young_gen->max_size() > 2122 young_gen->from_space()->capacity_in_bytes() + 2123 young_gen->to_space()->capacity_in_bytes(), 2124 "Sizes of space in young gen are out-of-bounds"); 2125 2126 size_t young_live = young_gen->used_in_bytes(); 2127 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2128 size_t old_live = old_gen->used_in_bytes(); 2129 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2130 size_t max_old_gen_size = old_gen->max_gen_size(); 2131 size_t max_eden_size = young_gen->max_size() - 2132 young_gen->from_space()->capacity_in_bytes() - 2133 young_gen->to_space()->capacity_in_bytes(); 2134 2135 // Used for diagnostics 2136 size_policy->clear_generation_free_space_flags(); 2137 2138 size_policy->compute_generations_free_space(young_live, 2139 eden_live, 2140 old_live, 2141 cur_eden, 2142 max_old_gen_size, 2143 max_eden_size, 2144 true /* full gc*/); 2145 2146 size_policy->check_gc_overhead_limit(young_live, 2147 eden_live, 2148 max_old_gen_size, 2149 max_eden_size, 2150 true /* full gc*/, 2151 gc_cause, 2152 heap->collector_policy()); 2153 2154 size_policy->decay_supplemental_growth(true /* full gc*/); 2155 2156 heap->resize_old_gen( 2157 size_policy->calculated_old_free_size_in_bytes()); 2158 2159 // Don't resize the young generation at an major collection. A 2160 // desired young generation size may have been calculated but 2161 // resizing the young generation complicates the code because the 2162 // resizing of the old generation may have moved the boundary 2163 // between the young generation and the old generation. Let the 2164 // young generation resizing happen at the minor collections. 2165 } 2166 if (PrintAdaptiveSizePolicy) { 2167 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2168 heap->total_collections()); 2169 } 2170 } 2171 2172 if (UsePerfData) { 2173 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2174 counters->update_counters(); 2175 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2176 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2177 } 2178 2179 heap->resize_all_tlabs(); 2180 2181 // Resize the metaspace capactiy after a collection 2182 MetaspaceGC::compute_new_size(); 2183 2184 if (TraceGen1Time) accumulated_time()->stop(); 2185 2186 if (PrintGC) { 2187 if (PrintGCDetails) { 2188 // No GC timestamp here. This is after GC so it would be confusing. 2189 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2190 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2191 heap->print_heap_change(pre_gc_values.heap_used()); 2192 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2193 } else { 2194 heap->print_heap_change(pre_gc_values.heap_used()); 2195 } 2196 } 2197 2198 // Track memory usage and detect low memory 2199 MemoryService::track_memory_usage(); 2200 heap->update_counters(); 2201 gc_task_manager()->release_idle_workers(); 2202 } 2203 2204 #ifdef ASSERT 2205 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2206 ParCompactionManager* const cm = 2207 ParCompactionManager::manager_array(int(i)); 2208 assert(cm->marking_stack()->is_empty(), "should be empty"); 2209 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2210 } 2211 #endif // ASSERT 2212 2213 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2214 HandleMark hm; // Discard invalid handles created during verification 2215 Universe::verify(" VerifyAfterGC:"); 2216 } 2217 2218 // Re-verify object start arrays 2219 if (VerifyObjectStartArray && 2220 VerifyAfterGC) { 2221 old_gen->verify_object_start_array(); 2222 } 2223 2224 if (ZapUnusedHeapArea) { 2225 old_gen->object_space()->check_mangled_unused_area_complete(); 2226 } 2227 2228 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2229 2230 collection_exit.update(); 2231 2232 heap->print_heap_after_gc(); 2233 heap->trace_heap_after_gc(&_gc_tracer); 2234 2235 if (PrintGCTaskTimeStamps) { 2236 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2237 INT64_FORMAT, 2238 marking_start.ticks(), compaction_start.ticks(), 2239 collection_exit.ticks()); 2240 gc_task_manager()->print_task_time_stamps(); 2241 } 2242 2243 heap->post_full_gc_dump(&_gc_timer); 2244 2245 #ifdef TRACESPINNING 2246 ParallelTaskTerminator::print_termination_counts(); 2247 #endif 2248 2249 _gc_timer.register_gc_end(); 2250 2251 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2252 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2253 2254 return true; 2255 } 2256 2257 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2258 PSYoungGen* young_gen, 2259 PSOldGen* old_gen) { 2260 MutableSpace* const eden_space = young_gen->eden_space(); 2261 assert(!eden_space->is_empty(), "eden must be non-empty"); 2262 assert(young_gen->virtual_space()->alignment() == 2263 old_gen->virtual_space()->alignment(), "alignments do not match"); 2264 2265 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2266 return false; 2267 } 2268 2269 // Both generations must be completely committed. 2270 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2271 return false; 2272 } 2273 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2274 return false; 2275 } 2276 2277 // Figure out how much to take from eden. Include the average amount promoted 2278 // in the total; otherwise the next young gen GC will simply bail out to a 2279 // full GC. 2280 const size_t alignment = old_gen->virtual_space()->alignment(); 2281 const size_t eden_used = eden_space->used_in_bytes(); 2282 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2283 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2284 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2285 2286 if (absorb_size >= eden_capacity) { 2287 return false; // Must leave some space in eden. 2288 } 2289 2290 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2291 if (new_young_size < young_gen->min_gen_size()) { 2292 return false; // Respect young gen minimum size. 2293 } 2294 2295 if (TraceAdaptiveGCBoundary && Verbose) { 2296 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2297 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2298 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2299 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2300 absorb_size / K, 2301 eden_capacity / K, (eden_capacity - absorb_size) / K, 2302 young_gen->from_space()->used_in_bytes() / K, 2303 young_gen->to_space()->used_in_bytes() / K, 2304 young_gen->capacity_in_bytes() / K, new_young_size / K); 2305 } 2306 2307 // Fill the unused part of the old gen. 2308 MutableSpace* const old_space = old_gen->object_space(); 2309 HeapWord* const unused_start = old_space->top(); 2310 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2311 2312 if (unused_words > 0) { 2313 if (unused_words < CollectedHeap::min_fill_size()) { 2314 return false; // If the old gen cannot be filled, must give up. 2315 } 2316 CollectedHeap::fill_with_objects(unused_start, unused_words); 2317 } 2318 2319 // Take the live data from eden and set both top and end in the old gen to 2320 // eden top. (Need to set end because reset_after_change() mangles the region 2321 // from end to virtual_space->high() in debug builds). 2322 HeapWord* const new_top = eden_space->top(); 2323 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2324 absorb_size); 2325 young_gen->reset_after_change(); 2326 old_space->set_top(new_top); 2327 old_space->set_end(new_top); 2328 old_gen->reset_after_change(); 2329 2330 // Update the object start array for the filler object and the data from eden. 2331 ObjectStartArray* const start_array = old_gen->start_array(); 2332 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2333 start_array->allocate_block(p); 2334 } 2335 2336 // Could update the promoted average here, but it is not typically updated at 2337 // full GCs and the value to use is unclear. Something like 2338 // 2339 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2340 2341 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2342 return true; 2343 } 2344 2345 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2346 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2347 "shouldn't return NULL"); 2348 return ParallelScavengeHeap::gc_task_manager(); 2349 } 2350 2351 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2352 bool maximum_heap_compaction, 2353 ParallelOldTracer *gc_tracer) { 2354 // Recursively traverse all live objects and mark them 2355 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2356 2357 ParallelScavengeHeap* heap = gc_heap(); 2358 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2359 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2360 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2361 ParallelTaskTerminator terminator(active_gc_threads, qset); 2362 2363 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2364 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2365 2366 // Need new claim bits before marking starts. 2367 ClassLoaderDataGraph::clear_claimed_marks(); 2368 2369 { 2370 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2371 2372 ParallelScavengeHeap::ParStrongRootsScope psrs; 2373 2374 GCTaskQueue* q = GCTaskQueue::create(); 2375 2376 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2377 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2378 // We scan the thread roots in parallel 2379 Threads::create_thread_roots_marking_tasks(q); 2380 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2381 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2382 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2383 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2384 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2385 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2386 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2387 2388 if (active_gc_threads > 1) { 2389 for (uint j = 0; j < active_gc_threads; j++) { 2390 q->enqueue(new StealMarkingTask(&terminator)); 2391 } 2392 } 2393 2394 gc_task_manager()->execute_and_wait(q); 2395 } 2396 2397 // Process reference objects found during marking 2398 { 2399 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2400 2401 ReferenceProcessorStats stats; 2402 if (ref_processor()->processing_is_mt()) { 2403 RefProcTaskExecutor task_executor; 2404 stats = ref_processor()->process_discovered_references( 2405 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2406 &task_executor, &_gc_timer, _gc_tracer.gc_id()); 2407 } else { 2408 stats = ref_processor()->process_discovered_references( 2409 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2410 &_gc_timer, _gc_tracer.gc_id()); 2411 } 2412 2413 gc_tracer->report_gc_reference_stats(stats); 2414 } 2415 2416 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2417 2418 // This is the point where the entire marking should have completed. 2419 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2420 2421 // Follow system dictionary roots and unload classes. 2422 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2423 2424 // Unload nmethods. 2425 CodeCache::do_unloading(is_alive_closure(), purged_class); 2426 2427 // Prune dead klasses from subklass/sibling/implementor lists. 2428 Klass::clean_weak_klass_links(is_alive_closure()); 2429 2430 // Delete entries for dead interned strings. 2431 StringTable::unlink(is_alive_closure()); 2432 2433 // Clean up unreferenced symbols in symbol table. 2434 SymbolTable::unlink(); 2435 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2436 } 2437 2438 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2439 ClassLoaderData* cld) { 2440 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2441 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2442 2443 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2444 } 2445 2446 // This should be moved to the shared markSweep code! 2447 class PSAlwaysTrueClosure: public BoolObjectClosure { 2448 public: 2449 bool do_object_b(oop p) { return true; } 2450 }; 2451 static PSAlwaysTrueClosure always_true; 2452 2453 void PSParallelCompact::adjust_roots() { 2454 // Adjust the pointers to reflect the new locations 2455 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2456 2457 // Need new claim bits when tracing through and adjusting pointers. 2458 ClassLoaderDataGraph::clear_claimed_marks(); 2459 2460 // General strong roots. 2461 Universe::oops_do(adjust_pointer_closure()); 2462 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2463 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2464 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2465 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2466 FlatProfiler::oops_do(adjust_pointer_closure()); 2467 Management::oops_do(adjust_pointer_closure()); 2468 JvmtiExport::oops_do(adjust_pointer_closure()); 2469 SystemDictionary::oops_do(adjust_pointer_closure()); 2470 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2471 2472 // Now adjust pointers in remaining weak roots. (All of which should 2473 // have been cleared if they pointed to non-surviving objects.) 2474 // Global (weak) JNI handles 2475 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2476 2477 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2478 CodeCache::blobs_do(&adjust_from_blobs); 2479 StringTable::oops_do(adjust_pointer_closure()); 2480 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2481 // Roots were visited so references into the young gen in roots 2482 // may have been scanned. Process them also. 2483 // Should the reference processor have a span that excludes 2484 // young gen objects? 2485 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2486 } 2487 2488 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2489 uint parallel_gc_threads) 2490 { 2491 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2492 2493 // Find the threads that are active 2494 unsigned int which = 0; 2495 2496 const uint task_count = MAX2(parallel_gc_threads, 1U); 2497 for (uint j = 0; j < task_count; j++) { 2498 q->enqueue(new DrainStacksCompactionTask(j)); 2499 ParCompactionManager::verify_region_list_empty(j); 2500 // Set the region stacks variables to "no" region stack values 2501 // so that they will be recognized and needing a region stack 2502 // in the stealing tasks if they do not get one by executing 2503 // a draining stack. 2504 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2505 cm->set_region_stack(NULL); 2506 cm->set_region_stack_index((uint)max_uintx); 2507 } 2508 ParCompactionManager::reset_recycled_stack_index(); 2509 2510 // Find all regions that are available (can be filled immediately) and 2511 // distribute them to the thread stacks. The iteration is done in reverse 2512 // order (high to low) so the regions will be removed in ascending order. 2513 2514 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2515 2516 size_t fillable_regions = 0; // A count for diagnostic purposes. 2517 // A region index which corresponds to the tasks created above. 2518 // "which" must be 0 <= which < task_count 2519 2520 which = 0; 2521 // id + 1 is used to test termination so unsigned can 2522 // be used with an old_space_id == 0. 2523 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2524 SpaceInfo* const space_info = _space_info + id; 2525 MutableSpace* const space = space_info->space(); 2526 HeapWord* const new_top = space_info->new_top(); 2527 2528 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2529 const size_t end_region = 2530 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2531 2532 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2533 if (sd.region(cur)->claim_unsafe()) { 2534 ParCompactionManager::region_list_push(which, cur); 2535 2536 if (TraceParallelOldGCCompactionPhase && Verbose) { 2537 const size_t count_mod_8 = fillable_regions & 7; 2538 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2539 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2540 if (count_mod_8 == 7) gclog_or_tty->cr(); 2541 } 2542 2543 NOT_PRODUCT(++fillable_regions;) 2544 2545 // Assign regions to tasks in round-robin fashion. 2546 if (++which == task_count) { 2547 assert(which <= parallel_gc_threads, 2548 "Inconsistent number of workers"); 2549 which = 0; 2550 } 2551 } 2552 } 2553 } 2554 2555 if (TraceParallelOldGCCompactionPhase) { 2556 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2557 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2558 } 2559 } 2560 2561 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2562 2563 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2564 uint parallel_gc_threads) { 2565 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2566 2567 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2568 2569 // Iterate over all the spaces adding tasks for updating 2570 // regions in the dense prefix. Assume that 1 gc thread 2571 // will work on opening the gaps and the remaining gc threads 2572 // will work on the dense prefix. 2573 unsigned int space_id; 2574 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2575 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2576 const MutableSpace* const space = _space_info[space_id].space(); 2577 2578 if (dense_prefix_end == space->bottom()) { 2579 // There is no dense prefix for this space. 2580 continue; 2581 } 2582 2583 // The dense prefix is before this region. 2584 size_t region_index_end_dense_prefix = 2585 sd.addr_to_region_idx(dense_prefix_end); 2586 RegionData* const dense_prefix_cp = 2587 sd.region(region_index_end_dense_prefix); 2588 assert(dense_prefix_end == space->end() || 2589 dense_prefix_cp->available() || 2590 dense_prefix_cp->claimed(), 2591 "The region after the dense prefix should always be ready to fill"); 2592 2593 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2594 2595 // Is there dense prefix work? 2596 size_t total_dense_prefix_regions = 2597 region_index_end_dense_prefix - region_index_start; 2598 // How many regions of the dense prefix should be given to 2599 // each thread? 2600 if (total_dense_prefix_regions > 0) { 2601 uint tasks_for_dense_prefix = 1; 2602 if (total_dense_prefix_regions <= 2603 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2604 // Don't over partition. This assumes that 2605 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2606 // so there are not many regions to process. 2607 tasks_for_dense_prefix = parallel_gc_threads; 2608 } else { 2609 // Over partition 2610 tasks_for_dense_prefix = parallel_gc_threads * 2611 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2612 } 2613 size_t regions_per_thread = total_dense_prefix_regions / 2614 tasks_for_dense_prefix; 2615 // Give each thread at least 1 region. 2616 if (regions_per_thread == 0) { 2617 regions_per_thread = 1; 2618 } 2619 2620 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2621 if (region_index_start >= region_index_end_dense_prefix) { 2622 break; 2623 } 2624 // region_index_end is not processed 2625 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2626 region_index_end_dense_prefix); 2627 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2628 region_index_start, 2629 region_index_end)); 2630 region_index_start = region_index_end; 2631 } 2632 } 2633 // This gets any part of the dense prefix that did not 2634 // fit evenly. 2635 if (region_index_start < region_index_end_dense_prefix) { 2636 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2637 region_index_start, 2638 region_index_end_dense_prefix)); 2639 } 2640 } 2641 } 2642 2643 void PSParallelCompact::enqueue_region_stealing_tasks( 2644 GCTaskQueue* q, 2645 ParallelTaskTerminator* terminator_ptr, 2646 uint parallel_gc_threads) { 2647 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2648 2649 // Once a thread has drained it's stack, it should try to steal regions from 2650 // other threads. 2651 if (parallel_gc_threads > 1) { 2652 for (uint j = 0; j < parallel_gc_threads; j++) { 2653 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2654 } 2655 } 2656 } 2657 2658 #ifdef ASSERT 2659 // Write a histogram of the number of times the block table was filled for a 2660 // region. 2661 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2662 { 2663 if (!TraceParallelOldGCCompactionPhase) return; 2664 2665 typedef ParallelCompactData::RegionData rd_t; 2666 ParallelCompactData& sd = summary_data(); 2667 2668 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2669 MutableSpace* const spc = _space_info[id].space(); 2670 if (spc->bottom() != spc->top()) { 2671 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2672 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2673 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2674 2675 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2676 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2677 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2678 2679 for (const rd_t* cur = beg; cur < end; ++cur) { 2680 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2681 } 2682 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2683 for (size_t i = 0; i < histo_len; ++i) { 2684 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2685 histo[i], 100.0 * histo[i] / region_cnt); 2686 } 2687 out->cr(); 2688 } 2689 } 2690 } 2691 #endif // #ifdef ASSERT 2692 2693 void PSParallelCompact::compact() { 2694 // trace("5"); 2695 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2696 2697 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2698 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2699 PSOldGen* old_gen = heap->old_gen(); 2700 old_gen->start_array()->reset(); 2701 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2702 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2703 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2704 ParallelTaskTerminator terminator(active_gc_threads, qset); 2705 2706 GCTaskQueue* q = GCTaskQueue::create(); 2707 enqueue_region_draining_tasks(q, active_gc_threads); 2708 enqueue_dense_prefix_tasks(q, active_gc_threads); 2709 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2710 2711 { 2712 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2713 2714 gc_task_manager()->execute_and_wait(q); 2715 2716 #ifdef ASSERT 2717 // Verify that all regions have been processed before the deferred updates. 2718 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2719 verify_complete(SpaceId(id)); 2720 } 2721 #endif 2722 } 2723 2724 { 2725 // Update the deferred objects, if any. Any compaction manager can be used. 2726 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2727 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2728 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2729 update_deferred_objects(cm, SpaceId(id)); 2730 } 2731 } 2732 2733 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2734 } 2735 2736 #ifdef ASSERT 2737 void PSParallelCompact::verify_complete(SpaceId space_id) { 2738 // All Regions between space bottom() to new_top() should be marked as filled 2739 // and all Regions between new_top() and top() should be available (i.e., 2740 // should have been emptied). 2741 ParallelCompactData& sd = summary_data(); 2742 SpaceInfo si = _space_info[space_id]; 2743 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2744 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2745 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2746 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2747 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2748 2749 bool issued_a_warning = false; 2750 2751 size_t cur_region; 2752 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2753 const RegionData* const c = sd.region(cur_region); 2754 if (!c->completed()) { 2755 warning("region " SIZE_FORMAT " not filled: " 2756 "destination_count=" SIZE_FORMAT, 2757 cur_region, c->destination_count()); 2758 issued_a_warning = true; 2759 } 2760 } 2761 2762 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2763 const RegionData* const c = sd.region(cur_region); 2764 if (!c->available()) { 2765 warning("region " SIZE_FORMAT " not empty: " 2766 "destination_count=" SIZE_FORMAT, 2767 cur_region, c->destination_count()); 2768 issued_a_warning = true; 2769 } 2770 } 2771 2772 if (issued_a_warning) { 2773 print_region_ranges(); 2774 } 2775 } 2776 #endif // #ifdef ASSERT 2777 2778 // Update interior oops in the ranges of regions [beg_region, end_region). 2779 void 2780 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2781 SpaceId space_id, 2782 size_t beg_region, 2783 size_t end_region) { 2784 ParallelCompactData& sd = summary_data(); 2785 ParMarkBitMap* const mbm = mark_bitmap(); 2786 2787 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2788 HeapWord* const end_addr = sd.region_to_addr(end_region); 2789 assert(beg_region <= end_region, "bad region range"); 2790 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2791 2792 #ifdef ASSERT 2793 // Claim the regions to avoid triggering an assert when they are marked as 2794 // filled. 2795 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2796 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2797 } 2798 #endif // #ifdef ASSERT 2799 2800 if (beg_addr != space(space_id)->bottom()) { 2801 // Find the first live object or block of dead space that *starts* in this 2802 // range of regions. If a partial object crosses onto the region, skip it; 2803 // it will be marked for 'deferred update' when the object head is 2804 // processed. If dead space crosses onto the region, it is also skipped; it 2805 // will be filled when the prior region is processed. If neither of those 2806 // apply, the first word in the region is the start of a live object or dead 2807 // space. 2808 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2809 const RegionData* const cp = sd.region(beg_region); 2810 if (cp->partial_obj_size() != 0) { 2811 beg_addr = sd.partial_obj_end(beg_region); 2812 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2813 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2814 } 2815 } 2816 2817 if (beg_addr < end_addr) { 2818 // A live object or block of dead space starts in this range of Regions. 2819 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2820 2821 // Create closures and iterate. 2822 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2823 FillClosure fill_closure(cm, space_id); 2824 ParMarkBitMap::IterationStatus status; 2825 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2826 dense_prefix_end); 2827 if (status == ParMarkBitMap::incomplete) { 2828 update_closure.do_addr(update_closure.source()); 2829 } 2830 } 2831 2832 // Mark the regions as filled. 2833 RegionData* const beg_cp = sd.region(beg_region); 2834 RegionData* const end_cp = sd.region(end_region); 2835 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2836 cp->set_completed(); 2837 } 2838 } 2839 2840 // Return the SpaceId for the space containing addr. If addr is not in the 2841 // heap, last_space_id is returned. In debug mode it expects the address to be 2842 // in the heap and asserts such. 2843 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2844 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2845 2846 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2847 if (_space_info[id].space()->contains(addr)) { 2848 return SpaceId(id); 2849 } 2850 } 2851 2852 assert(false, "no space contains the addr"); 2853 return last_space_id; 2854 } 2855 2856 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2857 SpaceId id) { 2858 assert(id < last_space_id, "bad space id"); 2859 2860 ParallelCompactData& sd = summary_data(); 2861 const SpaceInfo* const space_info = _space_info + id; 2862 ObjectStartArray* const start_array = space_info->start_array(); 2863 2864 const MutableSpace* const space = space_info->space(); 2865 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2866 HeapWord* const beg_addr = space_info->dense_prefix(); 2867 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2868 2869 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2870 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2871 const RegionData* cur_region; 2872 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2873 HeapWord* const addr = cur_region->deferred_obj_addr(); 2874 if (addr != NULL) { 2875 if (start_array != NULL) { 2876 start_array->allocate_block(addr); 2877 } 2878 oop(addr)->update_contents(cm); 2879 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 2880 } 2881 } 2882 } 2883 2884 // Skip over count live words starting from beg, and return the address of the 2885 // next live word. Unless marked, the word corresponding to beg is assumed to 2886 // be dead. Callers must either ensure beg does not correspond to the middle of 2887 // an object, or account for those live words in some other way. Callers must 2888 // also ensure that there are enough live words in the range [beg, end) to skip. 2889 HeapWord* 2890 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2891 { 2892 assert(count > 0, "sanity"); 2893 2894 ParMarkBitMap* m = mark_bitmap(); 2895 idx_t bits_to_skip = m->words_to_bits(count); 2896 idx_t cur_beg = m->addr_to_bit(beg); 2897 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2898 2899 do { 2900 cur_beg = m->find_obj_beg(cur_beg, search_end); 2901 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2902 const size_t obj_bits = cur_end - cur_beg + 1; 2903 if (obj_bits > bits_to_skip) { 2904 return m->bit_to_addr(cur_beg + bits_to_skip); 2905 } 2906 bits_to_skip -= obj_bits; 2907 cur_beg = cur_end + 1; 2908 } while (bits_to_skip > 0); 2909 2910 // Skipping the desired number of words landed just past the end of an object. 2911 // Find the start of the next object. 2912 cur_beg = m->find_obj_beg(cur_beg, search_end); 2913 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2914 return m->bit_to_addr(cur_beg); 2915 } 2916 2917 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2918 SpaceId src_space_id, 2919 size_t src_region_idx) 2920 { 2921 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2922 2923 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2924 if (split_info.dest_region_addr() == dest_addr) { 2925 // The partial object ending at the split point contains the first word to 2926 // be copied to dest_addr. 2927 return split_info.first_src_addr(); 2928 } 2929 2930 const ParallelCompactData& sd = summary_data(); 2931 ParMarkBitMap* const bitmap = mark_bitmap(); 2932 const size_t RegionSize = ParallelCompactData::RegionSize; 2933 2934 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2935 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2936 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2937 HeapWord* const src_region_destination = src_region_ptr->destination(); 2938 2939 assert(dest_addr >= src_region_destination, "wrong src region"); 2940 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2941 2942 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2943 HeapWord* const src_region_end = src_region_beg + RegionSize; 2944 2945 HeapWord* addr = src_region_beg; 2946 if (dest_addr == src_region_destination) { 2947 // Return the first live word in the source region. 2948 if (partial_obj_size == 0) { 2949 addr = bitmap->find_obj_beg(addr, src_region_end); 2950 assert(addr < src_region_end, "no objects start in src region"); 2951 } 2952 return addr; 2953 } 2954 2955 // Must skip some live data. 2956 size_t words_to_skip = dest_addr - src_region_destination; 2957 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2958 2959 if (partial_obj_size >= words_to_skip) { 2960 // All the live words to skip are part of the partial object. 2961 addr += words_to_skip; 2962 if (partial_obj_size == words_to_skip) { 2963 // Find the first live word past the partial object. 2964 addr = bitmap->find_obj_beg(addr, src_region_end); 2965 assert(addr < src_region_end, "wrong src region"); 2966 } 2967 return addr; 2968 } 2969 2970 // Skip over the partial object (if any). 2971 if (partial_obj_size != 0) { 2972 words_to_skip -= partial_obj_size; 2973 addr += partial_obj_size; 2974 } 2975 2976 // Skip over live words due to objects that start in the region. 2977 addr = skip_live_words(addr, src_region_end, words_to_skip); 2978 assert(addr < src_region_end, "wrong src region"); 2979 return addr; 2980 } 2981 2982 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2983 SpaceId src_space_id, 2984 size_t beg_region, 2985 HeapWord* end_addr) 2986 { 2987 ParallelCompactData& sd = summary_data(); 2988 2989 #ifdef ASSERT 2990 MutableSpace* const src_space = _space_info[src_space_id].space(); 2991 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2992 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2993 "src_space_id does not match beg_addr"); 2994 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2995 "src_space_id does not match end_addr"); 2996 #endif // #ifdef ASSERT 2997 2998 RegionData* const beg = sd.region(beg_region); 2999 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 3000 3001 // Regions up to new_top() are enqueued if they become available. 3002 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3003 RegionData* const enqueue_end = 3004 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3005 3006 for (RegionData* cur = beg; cur < end; ++cur) { 3007 assert(cur->data_size() > 0, "region must have live data"); 3008 cur->decrement_destination_count(); 3009 if (cur < enqueue_end && cur->available() && cur->claim()) { 3010 cm->push_region(sd.region(cur)); 3011 } 3012 } 3013 } 3014 3015 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3016 SpaceId& src_space_id, 3017 HeapWord*& src_space_top, 3018 HeapWord* end_addr) 3019 { 3020 typedef ParallelCompactData::RegionData RegionData; 3021 3022 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3023 const size_t region_size = ParallelCompactData::RegionSize; 3024 3025 size_t src_region_idx = 0; 3026 3027 // Skip empty regions (if any) up to the top of the space. 3028 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3029 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3030 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3031 const RegionData* const top_region_ptr = 3032 sd.addr_to_region_ptr(top_aligned_up); 3033 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3034 ++src_region_ptr; 3035 } 3036 3037 if (src_region_ptr < top_region_ptr) { 3038 // The next source region is in the current space. Update src_region_idx 3039 // and the source address to match src_region_ptr. 3040 src_region_idx = sd.region(src_region_ptr); 3041 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3042 if (src_region_addr > closure.source()) { 3043 closure.set_source(src_region_addr); 3044 } 3045 return src_region_idx; 3046 } 3047 3048 // Switch to a new source space and find the first non-empty region. 3049 unsigned int space_id = src_space_id + 1; 3050 assert(space_id < last_space_id, "not enough spaces"); 3051 3052 HeapWord* const destination = closure.destination(); 3053 3054 do { 3055 MutableSpace* space = _space_info[space_id].space(); 3056 HeapWord* const bottom = space->bottom(); 3057 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3058 3059 // Iterate over the spaces that do not compact into themselves. 3060 if (bottom_cp->destination() != bottom) { 3061 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3062 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3063 3064 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3065 if (src_cp->live_obj_size() > 0) { 3066 // Found it. 3067 assert(src_cp->destination() == destination, 3068 "first live obj in the space must match the destination"); 3069 assert(src_cp->partial_obj_size() == 0, 3070 "a space cannot begin with a partial obj"); 3071 3072 src_space_id = SpaceId(space_id); 3073 src_space_top = space->top(); 3074 const size_t src_region_idx = sd.region(src_cp); 3075 closure.set_source(sd.region_to_addr(src_region_idx)); 3076 return src_region_idx; 3077 } else { 3078 assert(src_cp->data_size() == 0, "sanity"); 3079 } 3080 } 3081 } 3082 } while (++space_id < last_space_id); 3083 3084 assert(false, "no source region was found"); 3085 return 0; 3086 } 3087 3088 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3089 { 3090 typedef ParMarkBitMap::IterationStatus IterationStatus; 3091 const size_t RegionSize = ParallelCompactData::RegionSize; 3092 ParMarkBitMap* const bitmap = mark_bitmap(); 3093 ParallelCompactData& sd = summary_data(); 3094 RegionData* const region_ptr = sd.region(region_idx); 3095 3096 // Get the items needed to construct the closure. 3097 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3098 SpaceId dest_space_id = space_id(dest_addr); 3099 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3100 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3101 assert(dest_addr < new_top, "sanity"); 3102 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3103 3104 // Get the source region and related info. 3105 size_t src_region_idx = region_ptr->source_region(); 3106 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3107 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3108 3109 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3110 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3111 3112 // Adjust src_region_idx to prepare for decrementing destination counts (the 3113 // destination count is not decremented when a region is copied to itself). 3114 if (src_region_idx == region_idx) { 3115 src_region_idx += 1; 3116 } 3117 3118 if (bitmap->is_unmarked(closure.source())) { 3119 // The first source word is in the middle of an object; copy the remainder 3120 // of the object or as much as will fit. The fact that pointer updates were 3121 // deferred will be noted when the object header is processed. 3122 HeapWord* const old_src_addr = closure.source(); 3123 closure.copy_partial_obj(); 3124 if (closure.is_full()) { 3125 decrement_destination_counts(cm, src_space_id, src_region_idx, 3126 closure.source()); 3127 region_ptr->set_deferred_obj_addr(NULL); 3128 region_ptr->set_completed(); 3129 return; 3130 } 3131 3132 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3133 if (sd.region_align_down(old_src_addr) != end_addr) { 3134 // The partial object was copied from more than one source region. 3135 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3136 3137 // Move to the next source region, possibly switching spaces as well. All 3138 // args except end_addr may be modified. 3139 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3140 end_addr); 3141 } 3142 } 3143 3144 do { 3145 HeapWord* const cur_addr = closure.source(); 3146 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3147 src_space_top); 3148 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3149 3150 if (status == ParMarkBitMap::incomplete) { 3151 // The last obj that starts in the source region does not end in the 3152 // region. 3153 assert(closure.source() < end_addr, "sanity"); 3154 HeapWord* const obj_beg = closure.source(); 3155 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3156 src_space_top); 3157 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3158 if (obj_end < range_end) { 3159 // The end was found; the entire object will fit. 3160 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3161 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3162 } else { 3163 // The end was not found; the object will not fit. 3164 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3165 status = ParMarkBitMap::would_overflow; 3166 } 3167 } 3168 3169 if (status == ParMarkBitMap::would_overflow) { 3170 // The last object did not fit. Note that interior oop updates were 3171 // deferred, then copy enough of the object to fill the region. 3172 region_ptr->set_deferred_obj_addr(closure.destination()); 3173 status = closure.copy_until_full(); // copies from closure.source() 3174 3175 decrement_destination_counts(cm, src_space_id, src_region_idx, 3176 closure.source()); 3177 region_ptr->set_completed(); 3178 return; 3179 } 3180 3181 if (status == ParMarkBitMap::full) { 3182 decrement_destination_counts(cm, src_space_id, src_region_idx, 3183 closure.source()); 3184 region_ptr->set_deferred_obj_addr(NULL); 3185 region_ptr->set_completed(); 3186 return; 3187 } 3188 3189 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3190 3191 // Move to the next source region, possibly switching spaces as well. All 3192 // args except end_addr may be modified. 3193 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3194 end_addr); 3195 } while (true); 3196 } 3197 3198 void PSParallelCompact::fill_blocks(size_t region_idx) 3199 { 3200 // Fill in the block table elements for the specified region. Each block 3201 // table element holds the number of live words in the region that are to the 3202 // left of the first object that starts in the block. Thus only blocks in 3203 // which an object starts need to be filled. 3204 // 3205 // The algorithm scans the section of the bitmap that corresponds to the 3206 // region, keeping a running total of the live words. When an object start is 3207 // found, if it's the first to start in the block that contains it, the 3208 // current total is written to the block table element. 3209 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3210 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3211 const size_t RegionSize = ParallelCompactData::RegionSize; 3212 3213 ParallelCompactData& sd = summary_data(); 3214 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3215 if (partial_obj_size >= RegionSize) { 3216 return; // No objects start in this region. 3217 } 3218 3219 // Ensure the first loop iteration decides that the block has changed. 3220 size_t cur_block = sd.block_count(); 3221 3222 const ParMarkBitMap* const bitmap = mark_bitmap(); 3223 3224 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3225 assert((size_t)1 << Log2BitsPerBlock == 3226 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3227 3228 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3229 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3230 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3231 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3232 while (beg_bit < range_end) { 3233 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3234 if (new_block != cur_block) { 3235 cur_block = new_block; 3236 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3237 } 3238 3239 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3240 if (end_bit < range_end - 1) { 3241 live_bits += end_bit - beg_bit + 1; 3242 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3243 } else { 3244 return; 3245 } 3246 } 3247 } 3248 3249 void 3250 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3251 const MutableSpace* sp = space(space_id); 3252 if (sp->is_empty()) { 3253 return; 3254 } 3255 3256 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3257 ParMarkBitMap* const bitmap = mark_bitmap(); 3258 HeapWord* const dp_addr = dense_prefix(space_id); 3259 HeapWord* beg_addr = sp->bottom(); 3260 HeapWord* end_addr = sp->top(); 3261 3262 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3263 3264 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3265 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3266 if (beg_region < dp_region) { 3267 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3268 } 3269 3270 // The destination of the first live object that starts in the region is one 3271 // past the end of the partial object entering the region (if any). 3272 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3273 HeapWord* const new_top = _space_info[space_id].new_top(); 3274 assert(new_top >= dest_addr, "bad new_top value"); 3275 const size_t words = pointer_delta(new_top, dest_addr); 3276 3277 if (words > 0) { 3278 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3279 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3280 3281 ParMarkBitMap::IterationStatus status; 3282 status = bitmap->iterate(&closure, dest_addr, end_addr); 3283 assert(status == ParMarkBitMap::full, "iteration not complete"); 3284 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3285 "live objects skipped because closure is full"); 3286 } 3287 } 3288 3289 jlong PSParallelCompact::millis_since_last_gc() { 3290 // We need a monotonically non-deccreasing time in ms but 3291 // os::javaTimeMillis() does not guarantee monotonicity. 3292 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3293 jlong ret_val = now - _time_of_last_gc; 3294 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3295 if (ret_val < 0) { 3296 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);) 3297 return 0; 3298 } 3299 return ret_val; 3300 } 3301 3302 void PSParallelCompact::reset_millis_since_last_gc() { 3303 // We need a monotonically non-deccreasing time in ms but 3304 // os::javaTimeMillis() does not guarantee monotonicity. 3305 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3306 } 3307 3308 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3309 { 3310 if (source() != destination()) { 3311 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3312 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3313 } 3314 update_state(words_remaining()); 3315 assert(is_full(), "sanity"); 3316 return ParMarkBitMap::full; 3317 } 3318 3319 void MoveAndUpdateClosure::copy_partial_obj() 3320 { 3321 size_t words = words_remaining(); 3322 3323 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3324 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3325 if (end_addr < range_end) { 3326 words = bitmap()->obj_size(source(), end_addr); 3327 } 3328 3329 // This test is necessary; if omitted, the pointer updates to a partial object 3330 // that crosses the dense prefix boundary could be overwritten. 3331 if (source() != destination()) { 3332 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3333 Copy::aligned_conjoint_words(source(), destination(), words); 3334 } 3335 update_state(words); 3336 } 3337 3338 ParMarkBitMapClosure::IterationStatus 3339 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3340 assert(destination() != NULL, "sanity"); 3341 assert(bitmap()->obj_size(addr) == words, "bad size"); 3342 3343 _source = addr; 3344 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3345 destination(), "wrong destination"); 3346 3347 if (words > words_remaining()) { 3348 return ParMarkBitMap::would_overflow; 3349 } 3350 3351 // The start_array must be updated even if the object is not moving. 3352 if (_start_array != NULL) { 3353 _start_array->allocate_block(destination()); 3354 } 3355 3356 if (destination() != source()) { 3357 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3358 Copy::aligned_conjoint_words(source(), destination(), words); 3359 } 3360 3361 oop moved_oop = (oop) destination(); 3362 moved_oop->update_contents(compaction_manager()); 3363 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3364 3365 update_state(words); 3366 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3367 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3368 } 3369 3370 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3371 ParCompactionManager* cm, 3372 PSParallelCompact::SpaceId space_id) : 3373 ParMarkBitMapClosure(mbm, cm), 3374 _space_id(space_id), 3375 _start_array(PSParallelCompact::start_array(space_id)) 3376 { 3377 } 3378 3379 // Updates the references in the object to their new values. 3380 ParMarkBitMapClosure::IterationStatus 3381 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3382 do_addr(addr); 3383 return ParMarkBitMap::incomplete; 3384 }