1 /* 2 * Copyright (c) 2005, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/stringTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/parallel/gcTaskManager.hpp" 30 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 31 #include "gc/parallel/pcTasks.hpp" 32 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 33 #include "gc/parallel/psCompactionManager.inline.hpp" 34 #include "gc/parallel/psMarkSweep.hpp" 35 #include "gc/parallel/psMarkSweepDecorator.hpp" 36 #include "gc/parallel/psOldGen.hpp" 37 #include "gc/parallel/psParallelCompact.inline.hpp" 38 #include "gc/parallel/psPromotionManager.inline.hpp" 39 #include "gc/parallel/psScavenge.hpp" 40 #include "gc/parallel/psYoungGen.hpp" 41 #include "gc/shared/gcCause.hpp" 42 #include "gc/shared/gcHeapSummary.hpp" 43 #include "gc/shared/gcId.hpp" 44 #include "gc/shared/gcLocker.inline.hpp" 45 #include "gc/shared/gcTimer.hpp" 46 #include "gc/shared/gcTrace.hpp" 47 #include "gc/shared/gcTraceTime.hpp" 48 #include "gc/shared/isGCActiveMark.hpp" 49 #include "gc/shared/referencePolicy.hpp" 50 #include "gc/shared/referenceProcessor.hpp" 51 #include "gc/shared/spaceDecorator.hpp" 52 #include "oops/instanceKlass.inline.hpp" 53 #include "oops/instanceMirrorKlass.inline.hpp" 54 #include "oops/methodData.hpp" 55 #include "oops/objArrayKlass.inline.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "runtime/atomic.inline.hpp" 58 #include "runtime/fprofiler.hpp" 59 #include "runtime/safepoint.hpp" 60 #include "runtime/vmThread.hpp" 61 #include "services/management.hpp" 62 #include "services/memTracker.hpp" 63 #include "services/memoryService.hpp" 64 #include "utilities/events.hpp" 65 #include "utilities/stack.inline.hpp" 66 67 #include <math.h> 68 69 // All sizes are in HeapWords. 70 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 71 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 72 const size_t ParallelCompactData::RegionSizeBytes = 73 RegionSize << LogHeapWordSize; 74 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 75 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 76 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 77 78 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 79 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 80 const size_t ParallelCompactData::BlockSizeBytes = 81 BlockSize << LogHeapWordSize; 82 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 83 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 84 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 85 86 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 87 const size_t ParallelCompactData::Log2BlocksPerRegion = 88 Log2RegionSize - Log2BlockSize; 89 90 const ParallelCompactData::RegionData::region_sz_t 91 ParallelCompactData::RegionData::dc_shift = 27; 92 93 const ParallelCompactData::RegionData::region_sz_t 94 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 95 96 const ParallelCompactData::RegionData::region_sz_t 97 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 98 99 const ParallelCompactData::RegionData::region_sz_t 100 ParallelCompactData::RegionData::los_mask = ~dc_mask; 101 102 const ParallelCompactData::RegionData::region_sz_t 103 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 104 105 const ParallelCompactData::RegionData::region_sz_t 106 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 107 108 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 109 bool PSParallelCompact::_print_phases = false; 110 111 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 112 113 double PSParallelCompact::_dwl_mean; 114 double PSParallelCompact::_dwl_std_dev; 115 double PSParallelCompact::_dwl_first_term; 116 double PSParallelCompact::_dwl_adjustment; 117 #ifdef ASSERT 118 bool PSParallelCompact::_dwl_initialized = false; 119 #endif // #ifdef ASSERT 120 121 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 122 HeapWord* destination) 123 { 124 assert(src_region_idx != 0, "invalid src_region_idx"); 125 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 126 assert(destination != NULL, "invalid destination argument"); 127 128 _src_region_idx = src_region_idx; 129 _partial_obj_size = partial_obj_size; 130 _destination = destination; 131 132 // These fields may not be updated below, so make sure they're clear. 133 assert(_dest_region_addr == NULL, "should have been cleared"); 134 assert(_first_src_addr == NULL, "should have been cleared"); 135 136 // Determine the number of destination regions for the partial object. 137 HeapWord* const last_word = destination + partial_obj_size - 1; 138 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 139 HeapWord* const beg_region_addr = sd.region_align_down(destination); 140 HeapWord* const end_region_addr = sd.region_align_down(last_word); 141 142 if (beg_region_addr == end_region_addr) { 143 // One destination region. 144 _destination_count = 1; 145 if (end_region_addr == destination) { 146 // The destination falls on a region boundary, thus the first word of the 147 // partial object will be the first word copied to the destination region. 148 _dest_region_addr = end_region_addr; 149 _first_src_addr = sd.region_to_addr(src_region_idx); 150 } 151 } else { 152 // Two destination regions. When copied, the partial object will cross a 153 // destination region boundary, so a word somewhere within the partial 154 // object will be the first word copied to the second destination region. 155 _destination_count = 2; 156 _dest_region_addr = end_region_addr; 157 const size_t ofs = pointer_delta(end_region_addr, destination); 158 assert(ofs < _partial_obj_size, "sanity"); 159 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 160 } 161 } 162 163 void SplitInfo::clear() 164 { 165 _src_region_idx = 0; 166 _partial_obj_size = 0; 167 _destination = NULL; 168 _destination_count = 0; 169 _dest_region_addr = NULL; 170 _first_src_addr = NULL; 171 assert(!is_valid(), "sanity"); 172 } 173 174 #ifdef ASSERT 175 void SplitInfo::verify_clear() 176 { 177 assert(_src_region_idx == 0, "not clear"); 178 assert(_partial_obj_size == 0, "not clear"); 179 assert(_destination == NULL, "not clear"); 180 assert(_destination_count == 0, "not clear"); 181 assert(_dest_region_addr == NULL, "not clear"); 182 assert(_first_src_addr == NULL, "not clear"); 183 } 184 #endif // #ifdef ASSERT 185 186 187 void PSParallelCompact::print_on_error(outputStream* st) { 188 _mark_bitmap.print_on_error(st); 189 } 190 191 #ifndef PRODUCT 192 const char* PSParallelCompact::space_names[] = { 193 "old ", "eden", "from", "to " 194 }; 195 196 void PSParallelCompact::print_region_ranges() 197 { 198 tty->print_cr("space bottom top end new_top"); 199 tty->print_cr("------ ---------- ---------- ---------- ----------"); 200 201 for (unsigned int id = 0; id < last_space_id; ++id) { 202 const MutableSpace* space = _space_info[id].space(); 203 tty->print_cr("%u %s " 204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 205 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 206 id, space_names[id], 207 summary_data().addr_to_region_idx(space->bottom()), 208 summary_data().addr_to_region_idx(space->top()), 209 summary_data().addr_to_region_idx(space->end()), 210 summary_data().addr_to_region_idx(_space_info[id].new_top())); 211 } 212 } 213 214 void 215 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 216 { 217 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 218 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 219 220 ParallelCompactData& sd = PSParallelCompact::summary_data(); 221 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 222 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 223 REGION_IDX_FORMAT " " PTR_FORMAT " " 224 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 225 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 226 i, p2i(c->data_location()), dci, p2i(c->destination()), 227 c->partial_obj_size(), c->live_obj_size(), 228 c->data_size(), c->source_region(), c->destination_count()); 229 230 #undef REGION_IDX_FORMAT 231 #undef REGION_DATA_FORMAT 232 } 233 234 void 235 print_generic_summary_data(ParallelCompactData& summary_data, 236 HeapWord* const beg_addr, 237 HeapWord* const end_addr) 238 { 239 size_t total_words = 0; 240 size_t i = summary_data.addr_to_region_idx(beg_addr); 241 const size_t last = summary_data.addr_to_region_idx(end_addr); 242 HeapWord* pdest = 0; 243 244 while (i <= last) { 245 ParallelCompactData::RegionData* c = summary_data.region(i); 246 if (c->data_size() != 0 || c->destination() != pdest) { 247 print_generic_summary_region(i, c); 248 total_words += c->data_size(); 249 pdest = c->destination(); 250 } 251 ++i; 252 } 253 254 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 255 } 256 257 void 258 print_generic_summary_data(ParallelCompactData& summary_data, 259 SpaceInfo* space_info) 260 { 261 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 262 const MutableSpace* space = space_info[id].space(); 263 print_generic_summary_data(summary_data, space->bottom(), 264 MAX2(space->top(), space_info[id].new_top())); 265 } 266 } 267 268 void 269 print_initial_summary_region(size_t i, 270 const ParallelCompactData::RegionData* c, 271 bool newline = true) 272 { 273 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 275 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 276 i, p2i(c->destination()), 277 c->partial_obj_size(), c->live_obj_size(), 278 c->data_size(), c->source_region(), c->destination_count()); 279 if (newline) tty->cr(); 280 } 281 282 void 283 print_initial_summary_data(ParallelCompactData& summary_data, 284 const MutableSpace* space) { 285 if (space->top() == space->bottom()) { 286 return; 287 } 288 289 const size_t region_size = ParallelCompactData::RegionSize; 290 typedef ParallelCompactData::RegionData RegionData; 291 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 292 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 293 const RegionData* c = summary_data.region(end_region - 1); 294 HeapWord* end_addr = c->destination() + c->data_size(); 295 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 296 297 // Print (and count) the full regions at the beginning of the space. 298 size_t full_region_count = 0; 299 size_t i = summary_data.addr_to_region_idx(space->bottom()); 300 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 301 print_initial_summary_region(i, summary_data.region(i)); 302 ++full_region_count; 303 ++i; 304 } 305 306 size_t live_to_right = live_in_space - full_region_count * region_size; 307 308 double max_reclaimed_ratio = 0.0; 309 size_t max_reclaimed_ratio_region = 0; 310 size_t max_dead_to_right = 0; 311 size_t max_live_to_right = 0; 312 313 // Print the 'reclaimed ratio' for regions while there is something live in 314 // the region or to the right of it. The remaining regions are empty (and 315 // uninteresting), and computing the ratio will result in division by 0. 316 while (i < end_region && live_to_right > 0) { 317 c = summary_data.region(i); 318 HeapWord* const region_addr = summary_data.region_to_addr(i); 319 const size_t used_to_right = pointer_delta(space->top(), region_addr); 320 const size_t dead_to_right = used_to_right - live_to_right; 321 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 322 323 if (reclaimed_ratio > max_reclaimed_ratio) { 324 max_reclaimed_ratio = reclaimed_ratio; 325 max_reclaimed_ratio_region = i; 326 max_dead_to_right = dead_to_right; 327 max_live_to_right = live_to_right; 328 } 329 330 print_initial_summary_region(i, c, false); 331 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 332 reclaimed_ratio, dead_to_right, live_to_right); 333 334 live_to_right -= c->data_size(); 335 ++i; 336 } 337 338 // Any remaining regions are empty. Print one more if there is one. 339 if (i < end_region) { 340 print_initial_summary_region(i, summary_data.region(i)); 341 } 342 343 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 344 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 345 max_reclaimed_ratio_region, max_dead_to_right, 346 max_live_to_right, max_reclaimed_ratio); 347 } 348 349 void 350 print_initial_summary_data(ParallelCompactData& summary_data, 351 SpaceInfo* space_info) { 352 unsigned int id = PSParallelCompact::old_space_id; 353 const MutableSpace* space; 354 do { 355 space = space_info[id].space(); 356 print_initial_summary_data(summary_data, space); 357 } while (++id < PSParallelCompact::eden_space_id); 358 359 do { 360 space = space_info[id].space(); 361 print_generic_summary_data(summary_data, space->bottom(), space->top()); 362 } while (++id < PSParallelCompact::last_space_id); 363 } 364 #endif // #ifndef PRODUCT 365 366 #ifdef ASSERT 367 size_t add_obj_count; 368 size_t add_obj_size; 369 size_t mark_bitmap_count; 370 size_t mark_bitmap_size; 371 #endif // #ifdef ASSERT 372 373 ParallelCompactData::ParallelCompactData() 374 { 375 _region_start = 0; 376 377 _region_vspace = 0; 378 _reserved_byte_size = 0; 379 _region_data = 0; 380 _region_count = 0; 381 382 _block_vspace = 0; 383 _block_data = 0; 384 _block_count = 0; 385 } 386 387 bool ParallelCompactData::initialize(MemRegion covered_region) 388 { 389 _region_start = covered_region.start(); 390 const size_t region_size = covered_region.word_size(); 391 DEBUG_ONLY(_region_end = _region_start + region_size;) 392 393 assert(region_align_down(_region_start) == _region_start, 394 "region start not aligned"); 395 assert((region_size & RegionSizeOffsetMask) == 0, 396 "region size not a multiple of RegionSize"); 397 398 bool result = initialize_region_data(region_size) && initialize_block_data(); 399 return result; 400 } 401 402 PSVirtualSpace* 403 ParallelCompactData::create_vspace(size_t count, size_t element_size) 404 { 405 const size_t raw_bytes = count * element_size; 406 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 407 const size_t granularity = os::vm_allocation_granularity(); 408 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 409 410 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 411 MAX2(page_sz, granularity); 412 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 413 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 414 rs.size()); 415 416 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 417 418 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 419 if (vspace != 0) { 420 if (vspace->expand_by(_reserved_byte_size)) { 421 return vspace; 422 } 423 delete vspace; 424 // Release memory reserved in the space. 425 rs.release(); 426 } 427 428 return 0; 429 } 430 431 bool ParallelCompactData::initialize_region_data(size_t region_size) 432 { 433 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 434 _region_vspace = create_vspace(count, sizeof(RegionData)); 435 if (_region_vspace != 0) { 436 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 437 _region_count = count; 438 return true; 439 } 440 return false; 441 } 442 443 bool ParallelCompactData::initialize_block_data() 444 { 445 assert(_region_count != 0, "region data must be initialized first"); 446 const size_t count = _region_count << Log2BlocksPerRegion; 447 _block_vspace = create_vspace(count, sizeof(BlockData)); 448 if (_block_vspace != 0) { 449 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 450 _block_count = count; 451 return true; 452 } 453 return false; 454 } 455 456 void ParallelCompactData::clear() 457 { 458 memset(_region_data, 0, _region_vspace->committed_size()); 459 memset(_block_data, 0, _block_vspace->committed_size()); 460 } 461 462 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 463 assert(beg_region <= _region_count, "beg_region out of range"); 464 assert(end_region <= _region_count, "end_region out of range"); 465 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 466 467 const size_t region_cnt = end_region - beg_region; 468 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 469 470 const size_t beg_block = beg_region * BlocksPerRegion; 471 const size_t block_cnt = region_cnt * BlocksPerRegion; 472 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 473 } 474 475 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 476 { 477 const RegionData* cur_cp = region(region_idx); 478 const RegionData* const end_cp = region(region_count() - 1); 479 480 HeapWord* result = region_to_addr(region_idx); 481 if (cur_cp < end_cp) { 482 do { 483 result += cur_cp->partial_obj_size(); 484 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 485 } 486 return result; 487 } 488 489 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 490 { 491 const size_t obj_ofs = pointer_delta(addr, _region_start); 492 const size_t beg_region = obj_ofs >> Log2RegionSize; 493 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 494 495 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 496 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 497 498 if (beg_region == end_region) { 499 // All in one region. 500 _region_data[beg_region].add_live_obj(len); 501 return; 502 } 503 504 // First region. 505 const size_t beg_ofs = region_offset(addr); 506 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 507 508 Klass* klass = ((oop)addr)->klass(); 509 // Middle regions--completely spanned by this object. 510 for (size_t region = beg_region + 1; region < end_region; ++region) { 511 _region_data[region].set_partial_obj_size(RegionSize); 512 _region_data[region].set_partial_obj_addr(addr); 513 } 514 515 // Last region. 516 const size_t end_ofs = region_offset(addr + len - 1); 517 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 518 _region_data[end_region].set_partial_obj_addr(addr); 519 } 520 521 void 522 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 523 { 524 assert(region_offset(beg) == 0, "not RegionSize aligned"); 525 assert(region_offset(end) == 0, "not RegionSize aligned"); 526 527 size_t cur_region = addr_to_region_idx(beg); 528 const size_t end_region = addr_to_region_idx(end); 529 HeapWord* addr = beg; 530 while (cur_region < end_region) { 531 _region_data[cur_region].set_destination(addr); 532 _region_data[cur_region].set_destination_count(0); 533 _region_data[cur_region].set_source_region(cur_region); 534 _region_data[cur_region].set_data_location(addr); 535 536 // Update live_obj_size so the region appears completely full. 537 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 538 _region_data[cur_region].set_live_obj_size(live_size); 539 540 ++cur_region; 541 addr += RegionSize; 542 } 543 } 544 545 // Find the point at which a space can be split and, if necessary, record the 546 // split point. 547 // 548 // If the current src region (which overflowed the destination space) doesn't 549 // have a partial object, the split point is at the beginning of the current src 550 // region (an "easy" split, no extra bookkeeping required). 551 // 552 // If the current src region has a partial object, the split point is in the 553 // region where that partial object starts (call it the split_region). If 554 // split_region has a partial object, then the split point is just after that 555 // partial object (a "hard" split where we have to record the split data and 556 // zero the partial_obj_size field). With a "hard" split, we know that the 557 // partial_obj ends within split_region because the partial object that caused 558 // the overflow starts in split_region. If split_region doesn't have a partial 559 // obj, then the split is at the beginning of split_region (another "easy" 560 // split). 561 HeapWord* 562 ParallelCompactData::summarize_split_space(size_t src_region, 563 SplitInfo& split_info, 564 HeapWord* destination, 565 HeapWord* target_end, 566 HeapWord** target_next) 567 { 568 assert(destination <= target_end, "sanity"); 569 assert(destination + _region_data[src_region].data_size() > target_end, 570 "region should not fit into target space"); 571 assert(is_region_aligned(target_end), "sanity"); 572 573 size_t split_region = src_region; 574 HeapWord* split_destination = destination; 575 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 576 577 if (destination + partial_obj_size > target_end) { 578 // The split point is just after the partial object (if any) in the 579 // src_region that contains the start of the object that overflowed the 580 // destination space. 581 // 582 // Find the start of the "overflow" object and set split_region to the 583 // region containing it. 584 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 585 split_region = addr_to_region_idx(overflow_obj); 586 587 // Clear the source_region field of all destination regions whose first word 588 // came from data after the split point (a non-null source_region field 589 // implies a region must be filled). 590 // 591 // An alternative to the simple loop below: clear during post_compact(), 592 // which uses memcpy instead of individual stores, and is easy to 593 // parallelize. (The downside is that it clears the entire RegionData 594 // object as opposed to just one field.) 595 // 596 // post_compact() would have to clear the summary data up to the highest 597 // address that was written during the summary phase, which would be 598 // 599 // max(top, max(new_top, clear_top)) 600 // 601 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 602 // to target_end. 603 const RegionData* const sr = region(split_region); 604 const size_t beg_idx = 605 addr_to_region_idx(region_align_up(sr->destination() + 606 sr->partial_obj_size())); 607 const size_t end_idx = addr_to_region_idx(target_end); 608 609 if (TraceParallelOldGCSummaryPhase) { 610 gclog_or_tty->print_cr("split: clearing source_region field in [" 611 SIZE_FORMAT ", " SIZE_FORMAT ")", 612 beg_idx, end_idx); 613 } 614 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 615 _region_data[idx].set_source_region(0); 616 } 617 618 // Set split_destination and partial_obj_size to reflect the split region. 619 split_destination = sr->destination(); 620 partial_obj_size = sr->partial_obj_size(); 621 } 622 623 // The split is recorded only if a partial object extends onto the region. 624 if (partial_obj_size != 0) { 625 _region_data[split_region].set_partial_obj_size(0); 626 split_info.record(split_region, partial_obj_size, split_destination); 627 } 628 629 // Setup the continuation addresses. 630 *target_next = split_destination + partial_obj_size; 631 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 632 633 if (TraceParallelOldGCSummaryPhase) { 634 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 635 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 636 " pos=" SIZE_FORMAT, 637 split_type, p2i(source_next), split_region, 638 partial_obj_size); 639 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 640 " tn=" PTR_FORMAT, 641 split_type, p2i(split_destination), 642 addr_to_region_idx(split_destination), 643 p2i(*target_next)); 644 645 if (partial_obj_size != 0) { 646 HeapWord* const po_beg = split_info.destination(); 647 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 648 gclog_or_tty->print_cr("%s split: " 649 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 650 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 651 split_type, 652 p2i(po_beg), addr_to_region_idx(po_beg), 653 p2i(po_end), addr_to_region_idx(po_end)); 654 } 655 } 656 657 return source_next; 658 } 659 660 bool ParallelCompactData::summarize(SplitInfo& split_info, 661 HeapWord* source_beg, HeapWord* source_end, 662 HeapWord** source_next, 663 HeapWord* target_beg, HeapWord* target_end, 664 HeapWord** target_next) 665 { 666 if (TraceParallelOldGCSummaryPhase) { 667 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 668 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 669 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 670 p2i(source_beg), p2i(source_end), p2i(source_next_val), 671 p2i(target_beg), p2i(target_end), p2i(*target_next)); 672 } 673 674 size_t cur_region = addr_to_region_idx(source_beg); 675 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 676 677 HeapWord *dest_addr = target_beg; 678 while (cur_region < end_region) { 679 // The destination must be set even if the region has no data. 680 _region_data[cur_region].set_destination(dest_addr); 681 682 size_t words = _region_data[cur_region].data_size(); 683 if (words > 0) { 684 // If cur_region does not fit entirely into the target space, find a point 685 // at which the source space can be 'split' so that part is copied to the 686 // target space and the rest is copied elsewhere. 687 if (dest_addr + words > target_end) { 688 assert(source_next != NULL, "source_next is NULL when splitting"); 689 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 690 target_end, target_next); 691 return false; 692 } 693 694 // Compute the destination_count for cur_region, and if necessary, update 695 // source_region for a destination region. The source_region field is 696 // updated if cur_region is the first (left-most) region to be copied to a 697 // destination region. 698 // 699 // The destination_count calculation is a bit subtle. A region that has 700 // data that compacts into itself does not count itself as a destination. 701 // This maintains the invariant that a zero count means the region is 702 // available and can be claimed and then filled. 703 uint destination_count = 0; 704 if (split_info.is_split(cur_region)) { 705 // The current region has been split: the partial object will be copied 706 // to one destination space and the remaining data will be copied to 707 // another destination space. Adjust the initial destination_count and, 708 // if necessary, set the source_region field if the partial object will 709 // cross a destination region boundary. 710 destination_count = split_info.destination_count(); 711 if (destination_count == 2) { 712 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 713 _region_data[dest_idx].set_source_region(cur_region); 714 } 715 } 716 717 HeapWord* const last_addr = dest_addr + words - 1; 718 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 719 const size_t dest_region_2 = addr_to_region_idx(last_addr); 720 721 // Initially assume that the destination regions will be the same and 722 // adjust the value below if necessary. Under this assumption, if 723 // cur_region == dest_region_2, then cur_region will be compacted 724 // completely into itself. 725 destination_count += cur_region == dest_region_2 ? 0 : 1; 726 if (dest_region_1 != dest_region_2) { 727 // Destination regions differ; adjust destination_count. 728 destination_count += 1; 729 // Data from cur_region will be copied to the start of dest_region_2. 730 _region_data[dest_region_2].set_source_region(cur_region); 731 } else if (region_offset(dest_addr) == 0) { 732 // Data from cur_region will be copied to the start of the destination 733 // region. 734 _region_data[dest_region_1].set_source_region(cur_region); 735 } 736 737 _region_data[cur_region].set_destination_count(destination_count); 738 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 739 dest_addr += words; 740 } 741 742 ++cur_region; 743 } 744 745 *target_next = dest_addr; 746 return true; 747 } 748 749 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 750 assert(addr != NULL, "Should detect NULL oop earlier"); 751 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 752 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 753 754 // Region covering the object. 755 RegionData* const region_ptr = addr_to_region_ptr(addr); 756 HeapWord* result = region_ptr->destination(); 757 758 // If the entire Region is live, the new location is region->destination + the 759 // offset of the object within in the Region. 760 761 // Run some performance tests to determine if this special case pays off. It 762 // is worth it for pointers into the dense prefix. If the optimization to 763 // avoid pointer updates in regions that only point to the dense prefix is 764 // ever implemented, this should be revisited. 765 if (region_ptr->data_size() == RegionSize) { 766 result += region_offset(addr); 767 return result; 768 } 769 770 // Otherwise, the new location is region->destination + block offset + the 771 // number of live words in the Block that are (a) to the left of addr and (b) 772 // due to objects that start in the Block. 773 774 // Fill in the block table if necessary. This is unsynchronized, so multiple 775 // threads may fill the block table for a region (harmless, since it is 776 // idempotent). 777 if (!region_ptr->blocks_filled()) { 778 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 779 region_ptr->set_blocks_filled(); 780 } 781 782 HeapWord* const search_start = block_align_down(addr); 783 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 784 785 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 786 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 787 result += block_offset + live; 788 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 789 return result; 790 } 791 792 #ifdef ASSERT 793 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 794 { 795 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 796 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 797 for (const size_t* p = beg; p < end; ++p) { 798 assert(*p == 0, "not zero"); 799 } 800 } 801 802 void ParallelCompactData::verify_clear() 803 { 804 verify_clear(_region_vspace); 805 verify_clear(_block_vspace); 806 } 807 #endif // #ifdef ASSERT 808 809 STWGCTimer PSParallelCompact::_gc_timer; 810 ParallelOldTracer PSParallelCompact::_gc_tracer; 811 elapsedTimer PSParallelCompact::_accumulated_time; 812 unsigned int PSParallelCompact::_total_invocations = 0; 813 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 814 jlong PSParallelCompact::_time_of_last_gc = 0; 815 CollectorCounters* PSParallelCompact::_counters = NULL; 816 ParMarkBitMap PSParallelCompact::_mark_bitmap; 817 ParallelCompactData PSParallelCompact::_summary_data; 818 819 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 820 821 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 822 823 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 824 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 825 826 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 827 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 828 } 829 830 void PSParallelCompact::post_initialize() { 831 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 832 MemRegion mr = heap->reserved_region(); 833 _ref_processor = 834 new ReferenceProcessor(mr, // span 835 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 836 ParallelGCThreads, // mt processing degree 837 true, // mt discovery 838 ParallelGCThreads, // mt discovery degree 839 true, // atomic_discovery 840 &_is_alive_closure); // non-header is alive closure 841 _counters = new CollectorCounters("PSParallelCompact", 1); 842 843 // Initialize static fields in ParCompactionManager. 844 ParCompactionManager::initialize(mark_bitmap()); 845 } 846 847 bool PSParallelCompact::initialize() { 848 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 849 MemRegion mr = heap->reserved_region(); 850 851 // Was the old gen get allocated successfully? 852 if (!heap->old_gen()->is_allocated()) { 853 return false; 854 } 855 856 initialize_space_info(); 857 initialize_dead_wood_limiter(); 858 859 if (!_mark_bitmap.initialize(mr)) { 860 vm_shutdown_during_initialization( 861 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 862 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 863 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 864 return false; 865 } 866 867 if (!_summary_data.initialize(mr)) { 868 vm_shutdown_during_initialization( 869 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 870 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 871 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 872 return false; 873 } 874 875 return true; 876 } 877 878 void PSParallelCompact::initialize_space_info() 879 { 880 memset(&_space_info, 0, sizeof(_space_info)); 881 882 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 883 PSYoungGen* young_gen = heap->young_gen(); 884 885 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 886 _space_info[eden_space_id].set_space(young_gen->eden_space()); 887 _space_info[from_space_id].set_space(young_gen->from_space()); 888 _space_info[to_space_id].set_space(young_gen->to_space()); 889 890 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 891 } 892 893 void PSParallelCompact::initialize_dead_wood_limiter() 894 { 895 const size_t max = 100; 896 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 897 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 898 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 899 DEBUG_ONLY(_dwl_initialized = true;) 900 _dwl_adjustment = normal_distribution(1.0); 901 } 902 903 // Simple class for storing info about the heap at the start of GC, to be used 904 // after GC for comparison/printing. 905 class PreGCValues { 906 public: 907 PreGCValues() { } 908 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 909 910 void fill(ParallelScavengeHeap* heap) { 911 _heap_used = heap->used(); 912 _young_gen_used = heap->young_gen()->used_in_bytes(); 913 _old_gen_used = heap->old_gen()->used_in_bytes(); 914 _metadata_used = MetaspaceAux::used_bytes(); 915 }; 916 917 size_t heap_used() const { return _heap_used; } 918 size_t young_gen_used() const { return _young_gen_used; } 919 size_t old_gen_used() const { return _old_gen_used; } 920 size_t metadata_used() const { return _metadata_used; } 921 922 private: 923 size_t _heap_used; 924 size_t _young_gen_used; 925 size_t _old_gen_used; 926 size_t _metadata_used; 927 }; 928 929 void 930 PSParallelCompact::clear_data_covering_space(SpaceId id) 931 { 932 // At this point, top is the value before GC, new_top() is the value that will 933 // be set at the end of GC. The marking bitmap is cleared to top; nothing 934 // should be marked above top. The summary data is cleared to the larger of 935 // top & new_top. 936 MutableSpace* const space = _space_info[id].space(); 937 HeapWord* const bot = space->bottom(); 938 HeapWord* const top = space->top(); 939 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 940 941 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 942 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 943 _mark_bitmap.clear_range(beg_bit, end_bit); 944 945 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 946 const size_t end_region = 947 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 948 _summary_data.clear_range(beg_region, end_region); 949 950 // Clear the data used to 'split' regions. 951 SplitInfo& split_info = _space_info[id].split_info(); 952 if (split_info.is_valid()) { 953 split_info.clear(); 954 } 955 DEBUG_ONLY(split_info.verify_clear();) 956 } 957 958 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 959 { 960 // Update the from & to space pointers in space_info, since they are swapped 961 // at each young gen gc. Do the update unconditionally (even though a 962 // promotion failure does not swap spaces) because an unknown number of young 963 // collections will have swapped the spaces an unknown number of times. 964 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer); 965 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 966 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 967 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 968 969 pre_gc_values->fill(heap); 970 971 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 972 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 973 974 // Increment the invocation count 975 heap->increment_total_collections(true); 976 977 // We need to track unique mark sweep invocations as well. 978 _total_invocations++; 979 980 heap->print_heap_before_gc(); 981 heap->trace_heap_before_gc(&_gc_tracer); 982 983 // Fill in TLABs 984 heap->accumulate_statistics_all_tlabs(); 985 heap->ensure_parsability(true); // retire TLABs 986 987 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 988 HandleMark hm; // Discard invalid handles created during verification 989 Universe::verify(" VerifyBeforeGC:"); 990 } 991 992 // Verify object start arrays 993 if (VerifyObjectStartArray && 994 VerifyBeforeGC) { 995 heap->old_gen()->verify_object_start_array(); 996 } 997 998 DEBUG_ONLY(mark_bitmap()->verify_clear();) 999 DEBUG_ONLY(summary_data().verify_clear();) 1000 1001 // Have worker threads release resources the next time they run a task. 1002 gc_task_manager()->release_all_resources(); 1003 } 1004 1005 void PSParallelCompact::post_compact() 1006 { 1007 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer); 1008 1009 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1010 // Clear the marking bitmap, summary data and split info. 1011 clear_data_covering_space(SpaceId(id)); 1012 // Update top(). Must be done after clearing the bitmap and summary data. 1013 _space_info[id].publish_new_top(); 1014 } 1015 1016 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1017 MutableSpace* const from_space = _space_info[from_space_id].space(); 1018 MutableSpace* const to_space = _space_info[to_space_id].space(); 1019 1020 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1021 bool eden_empty = eden_space->is_empty(); 1022 if (!eden_empty) { 1023 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1024 heap->young_gen(), heap->old_gen()); 1025 } 1026 1027 // Update heap occupancy information which is used as input to the soft ref 1028 // clearing policy at the next gc. 1029 Universe::update_heap_info_at_gc(); 1030 1031 bool young_gen_empty = eden_empty && from_space->is_empty() && 1032 to_space->is_empty(); 1033 1034 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1035 MemRegion old_mr = heap->old_gen()->reserved(); 1036 if (young_gen_empty) { 1037 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1038 } else { 1039 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1040 } 1041 1042 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1043 ClassLoaderDataGraph::purge(); 1044 MetaspaceAux::verify_metrics(); 1045 1046 CodeCache::gc_epilogue(); 1047 JvmtiExport::gc_epilogue(); 1048 1049 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1050 1051 ref_processor()->enqueue_discovered_references(NULL); 1052 1053 if (ZapUnusedHeapArea) { 1054 heap->gen_mangle_unused_area(); 1055 } 1056 1057 // Update time of last GC 1058 reset_millis_since_last_gc(); 1059 } 1060 1061 HeapWord* 1062 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1063 bool maximum_compaction) 1064 { 1065 const size_t region_size = ParallelCompactData::RegionSize; 1066 const ParallelCompactData& sd = summary_data(); 1067 1068 const MutableSpace* const space = _space_info[id].space(); 1069 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1070 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1071 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1072 1073 // Skip full regions at the beginning of the space--they are necessarily part 1074 // of the dense prefix. 1075 size_t full_count = 0; 1076 const RegionData* cp; 1077 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1078 ++full_count; 1079 } 1080 1081 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1082 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1083 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1084 if (maximum_compaction || cp == end_cp || interval_ended) { 1085 _maximum_compaction_gc_num = total_invocations(); 1086 return sd.region_to_addr(cp); 1087 } 1088 1089 HeapWord* const new_top = _space_info[id].new_top(); 1090 const size_t space_live = pointer_delta(new_top, space->bottom()); 1091 const size_t space_used = space->used_in_words(); 1092 const size_t space_capacity = space->capacity_in_words(); 1093 1094 const double cur_density = double(space_live) / space_capacity; 1095 const double deadwood_density = 1096 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1097 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1098 1099 if (TraceParallelOldGCDensePrefix) { 1100 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1101 cur_density, deadwood_density, deadwood_goal); 1102 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1103 "space_cap=" SIZE_FORMAT, 1104 space_live, space_used, 1105 space_capacity); 1106 } 1107 1108 // XXX - Use binary search? 1109 HeapWord* dense_prefix = sd.region_to_addr(cp); 1110 const RegionData* full_cp = cp; 1111 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1112 while (cp < end_cp) { 1113 HeapWord* region_destination = cp->destination(); 1114 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1115 if (TraceParallelOldGCDensePrefix && Verbose) { 1116 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1117 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1118 sd.region(cp), p2i(region_destination), 1119 p2i(dense_prefix), cur_deadwood); 1120 } 1121 1122 if (cur_deadwood >= deadwood_goal) { 1123 // Found the region that has the correct amount of deadwood to the left. 1124 // This typically occurs after crossing a fairly sparse set of regions, so 1125 // iterate backwards over those sparse regions, looking for the region 1126 // that has the lowest density of live objects 'to the right.' 1127 size_t space_to_left = sd.region(cp) * region_size; 1128 size_t live_to_left = space_to_left - cur_deadwood; 1129 size_t space_to_right = space_capacity - space_to_left; 1130 size_t live_to_right = space_live - live_to_left; 1131 double density_to_right = double(live_to_right) / space_to_right; 1132 while (cp > full_cp) { 1133 --cp; 1134 const size_t prev_region_live_to_right = live_to_right - 1135 cp->data_size(); 1136 const size_t prev_region_space_to_right = space_to_right + region_size; 1137 double prev_region_density_to_right = 1138 double(prev_region_live_to_right) / prev_region_space_to_right; 1139 if (density_to_right <= prev_region_density_to_right) { 1140 return dense_prefix; 1141 } 1142 if (TraceParallelOldGCDensePrefix && Verbose) { 1143 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1144 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1145 prev_region_density_to_right); 1146 } 1147 dense_prefix -= region_size; 1148 live_to_right = prev_region_live_to_right; 1149 space_to_right = prev_region_space_to_right; 1150 density_to_right = prev_region_density_to_right; 1151 } 1152 return dense_prefix; 1153 } 1154 1155 dense_prefix += region_size; 1156 ++cp; 1157 } 1158 1159 return dense_prefix; 1160 } 1161 1162 #ifndef PRODUCT 1163 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1164 const SpaceId id, 1165 const bool maximum_compaction, 1166 HeapWord* const addr) 1167 { 1168 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1169 RegionData* const cp = summary_data().region(region_idx); 1170 const MutableSpace* const space = _space_info[id].space(); 1171 HeapWord* const new_top = _space_info[id].new_top(); 1172 1173 const size_t space_live = pointer_delta(new_top, space->bottom()); 1174 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1175 const size_t space_cap = space->capacity_in_words(); 1176 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1177 const size_t live_to_right = new_top - cp->destination(); 1178 const size_t dead_to_right = space->top() - addr - live_to_right; 1179 1180 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1181 "spl=" SIZE_FORMAT " " 1182 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1183 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1184 " ratio=%10.8f", 1185 algorithm, p2i(addr), region_idx, 1186 space_live, 1187 dead_to_left, dead_to_left_pct, 1188 dead_to_right, live_to_right, 1189 double(dead_to_right) / live_to_right); 1190 } 1191 #endif // #ifndef PRODUCT 1192 1193 // Return a fraction indicating how much of the generation can be treated as 1194 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1195 // based on the density of live objects in the generation to determine a limit, 1196 // which is then adjusted so the return value is min_percent when the density is 1197 // 1. 1198 // 1199 // The following table shows some return values for a different values of the 1200 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1201 // min_percent is 1. 1202 // 1203 // fraction allowed as dead wood 1204 // ----------------------------------------------------------------- 1205 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1206 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1207 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1208 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1209 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1210 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1211 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1212 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1213 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1214 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1215 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1216 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1217 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1218 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1219 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1220 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1221 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1222 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1223 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1224 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1225 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1226 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1227 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1228 1229 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1230 { 1231 assert(_dwl_initialized, "uninitialized"); 1232 1233 // The raw limit is the value of the normal distribution at x = density. 1234 const double raw_limit = normal_distribution(density); 1235 1236 // Adjust the raw limit so it becomes the minimum when the density is 1. 1237 // 1238 // First subtract the adjustment value (which is simply the precomputed value 1239 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1240 // Then add the minimum value, so the minimum is returned when the density is 1241 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1242 const double min = double(min_percent) / 100.0; 1243 const double limit = raw_limit - _dwl_adjustment + min; 1244 return MAX2(limit, 0.0); 1245 } 1246 1247 ParallelCompactData::RegionData* 1248 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1249 const RegionData* end) 1250 { 1251 const size_t region_size = ParallelCompactData::RegionSize; 1252 ParallelCompactData& sd = summary_data(); 1253 size_t left = sd.region(beg); 1254 size_t right = end > beg ? sd.region(end) - 1 : left; 1255 1256 // Binary search. 1257 while (left < right) { 1258 // Equivalent to (left + right) / 2, but does not overflow. 1259 const size_t middle = left + (right - left) / 2; 1260 RegionData* const middle_ptr = sd.region(middle); 1261 HeapWord* const dest = middle_ptr->destination(); 1262 HeapWord* const addr = sd.region_to_addr(middle); 1263 assert(dest != NULL, "sanity"); 1264 assert(dest <= addr, "must move left"); 1265 1266 if (middle > left && dest < addr) { 1267 right = middle - 1; 1268 } else if (middle < right && middle_ptr->data_size() == region_size) { 1269 left = middle + 1; 1270 } else { 1271 return middle_ptr; 1272 } 1273 } 1274 return sd.region(left); 1275 } 1276 1277 ParallelCompactData::RegionData* 1278 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1279 const RegionData* end, 1280 size_t dead_words) 1281 { 1282 ParallelCompactData& sd = summary_data(); 1283 size_t left = sd.region(beg); 1284 size_t right = end > beg ? sd.region(end) - 1 : left; 1285 1286 // Binary search. 1287 while (left < right) { 1288 // Equivalent to (left + right) / 2, but does not overflow. 1289 const size_t middle = left + (right - left) / 2; 1290 RegionData* const middle_ptr = sd.region(middle); 1291 HeapWord* const dest = middle_ptr->destination(); 1292 HeapWord* const addr = sd.region_to_addr(middle); 1293 assert(dest != NULL, "sanity"); 1294 assert(dest <= addr, "must move left"); 1295 1296 const size_t dead_to_left = pointer_delta(addr, dest); 1297 if (middle > left && dead_to_left > dead_words) { 1298 right = middle - 1; 1299 } else if (middle < right && dead_to_left < dead_words) { 1300 left = middle + 1; 1301 } else { 1302 return middle_ptr; 1303 } 1304 } 1305 return sd.region(left); 1306 } 1307 1308 // The result is valid during the summary phase, after the initial summarization 1309 // of each space into itself, and before final summarization. 1310 inline double 1311 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1312 HeapWord* const bottom, 1313 HeapWord* const top, 1314 HeapWord* const new_top) 1315 { 1316 ParallelCompactData& sd = summary_data(); 1317 1318 assert(cp != NULL, "sanity"); 1319 assert(bottom != NULL, "sanity"); 1320 assert(top != NULL, "sanity"); 1321 assert(new_top != NULL, "sanity"); 1322 assert(top >= new_top, "summary data problem?"); 1323 assert(new_top > bottom, "space is empty; should not be here"); 1324 assert(new_top >= cp->destination(), "sanity"); 1325 assert(top >= sd.region_to_addr(cp), "sanity"); 1326 1327 HeapWord* const destination = cp->destination(); 1328 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1329 const size_t compacted_region_live = pointer_delta(new_top, destination); 1330 const size_t compacted_region_used = pointer_delta(top, 1331 sd.region_to_addr(cp)); 1332 const size_t reclaimable = compacted_region_used - compacted_region_live; 1333 1334 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1335 return double(reclaimable) / divisor; 1336 } 1337 1338 // Return the address of the end of the dense prefix, a.k.a. the start of the 1339 // compacted region. The address is always on a region boundary. 1340 // 1341 // Completely full regions at the left are skipped, since no compaction can 1342 // occur in those regions. Then the maximum amount of dead wood to allow is 1343 // computed, based on the density (amount live / capacity) of the generation; 1344 // the region with approximately that amount of dead space to the left is 1345 // identified as the limit region. Regions between the last completely full 1346 // region and the limit region are scanned and the one that has the best 1347 // (maximum) reclaimed_ratio() is selected. 1348 HeapWord* 1349 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1350 bool maximum_compaction) 1351 { 1352 if (ParallelOldGCSplitALot) { 1353 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1354 // The value was chosen to provoke splitting a young gen space; use it. 1355 return _space_info[id].dense_prefix(); 1356 } 1357 } 1358 1359 const size_t region_size = ParallelCompactData::RegionSize; 1360 const ParallelCompactData& sd = summary_data(); 1361 1362 const MutableSpace* const space = _space_info[id].space(); 1363 HeapWord* const top = space->top(); 1364 HeapWord* const top_aligned_up = sd.region_align_up(top); 1365 HeapWord* const new_top = _space_info[id].new_top(); 1366 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1367 HeapWord* const bottom = space->bottom(); 1368 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1369 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1370 const RegionData* const new_top_cp = 1371 sd.addr_to_region_ptr(new_top_aligned_up); 1372 1373 // Skip full regions at the beginning of the space--they are necessarily part 1374 // of the dense prefix. 1375 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1376 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1377 space->is_empty(), "no dead space allowed to the left"); 1378 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1379 "region must have dead space"); 1380 1381 // The gc number is saved whenever a maximum compaction is done, and used to 1382 // determine when the maximum compaction interval has expired. This avoids 1383 // successive max compactions for different reasons. 1384 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1385 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1386 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1387 total_invocations() == HeapFirstMaximumCompactionCount; 1388 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1389 _maximum_compaction_gc_num = total_invocations(); 1390 return sd.region_to_addr(full_cp); 1391 } 1392 1393 const size_t space_live = pointer_delta(new_top, bottom); 1394 const size_t space_used = space->used_in_words(); 1395 const size_t space_capacity = space->capacity_in_words(); 1396 1397 const double density = double(space_live) / double(space_capacity); 1398 const size_t min_percent_free = MarkSweepDeadRatio; 1399 const double limiter = dead_wood_limiter(density, min_percent_free); 1400 const size_t dead_wood_max = space_used - space_live; 1401 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1402 dead_wood_max); 1403 1404 if (TraceParallelOldGCDensePrefix) { 1405 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1406 "space_cap=" SIZE_FORMAT, 1407 space_live, space_used, 1408 space_capacity); 1409 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1410 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1411 density, min_percent_free, limiter, 1412 dead_wood_max, dead_wood_limit); 1413 } 1414 1415 // Locate the region with the desired amount of dead space to the left. 1416 const RegionData* const limit_cp = 1417 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1418 1419 // Scan from the first region with dead space to the limit region and find the 1420 // one with the best (largest) reclaimed ratio. 1421 double best_ratio = 0.0; 1422 const RegionData* best_cp = full_cp; 1423 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1424 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1425 if (tmp_ratio > best_ratio) { 1426 best_cp = cp; 1427 best_ratio = tmp_ratio; 1428 } 1429 } 1430 1431 #if 0 1432 // Something to consider: if the region with the best ratio is 'close to' the 1433 // first region w/free space, choose the first region with free space 1434 // ("first-free"). The first-free region is usually near the start of the 1435 // heap, which means we are copying most of the heap already, so copy a bit 1436 // more to get complete compaction. 1437 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1438 _maximum_compaction_gc_num = total_invocations(); 1439 best_cp = full_cp; 1440 } 1441 #endif // #if 0 1442 1443 return sd.region_to_addr(best_cp); 1444 } 1445 1446 #ifndef PRODUCT 1447 void 1448 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1449 size_t words) 1450 { 1451 if (TraceParallelOldGCSummaryPhase) { 1452 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1453 SIZE_FORMAT, p2i(start), p2i(start + words), words); 1454 } 1455 1456 ObjectStartArray* const start_array = _space_info[id].start_array(); 1457 CollectedHeap::fill_with_objects(start, words); 1458 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1459 _mark_bitmap.mark_obj(p, words); 1460 _summary_data.add_obj(p, words); 1461 start_array->allocate_block(p); 1462 } 1463 } 1464 1465 void 1466 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1467 { 1468 ParallelCompactData& sd = summary_data(); 1469 MutableSpace* space = _space_info[id].space(); 1470 1471 // Find the source and destination start addresses. 1472 HeapWord* const src_addr = sd.region_align_down(start); 1473 HeapWord* dst_addr; 1474 if (src_addr < start) { 1475 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1476 } else if (src_addr > space->bottom()) { 1477 // The start (the original top() value) is aligned to a region boundary so 1478 // the associated region does not have a destination. Compute the 1479 // destination from the previous region. 1480 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1481 dst_addr = cp->destination() + cp->data_size(); 1482 } else { 1483 // Filling the entire space. 1484 dst_addr = space->bottom(); 1485 } 1486 assert(dst_addr != NULL, "sanity"); 1487 1488 // Update the summary data. 1489 bool result = _summary_data.summarize(_space_info[id].split_info(), 1490 src_addr, space->top(), NULL, 1491 dst_addr, space->end(), 1492 _space_info[id].new_top_addr()); 1493 assert(result, "should not fail: bad filler object size"); 1494 } 1495 1496 void 1497 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1498 { 1499 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1500 return; 1501 } 1502 1503 MutableSpace* const space = _space_info[id].space(); 1504 if (space->is_empty()) { 1505 HeapWord* b = space->bottom(); 1506 HeapWord* t = b + space->capacity_in_words() / 2; 1507 space->set_top(t); 1508 if (ZapUnusedHeapArea) { 1509 space->set_top_for_allocations(); 1510 } 1511 1512 size_t min_size = CollectedHeap::min_fill_size(); 1513 size_t obj_len = min_size; 1514 while (b + obj_len <= t) { 1515 CollectedHeap::fill_with_object(b, obj_len); 1516 mark_bitmap()->mark_obj(b, obj_len); 1517 summary_data().add_obj(b, obj_len); 1518 b += obj_len; 1519 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1520 } 1521 if (b < t) { 1522 // The loop didn't completely fill to t (top); adjust top downward. 1523 space->set_top(b); 1524 if (ZapUnusedHeapArea) { 1525 space->set_top_for_allocations(); 1526 } 1527 } 1528 1529 HeapWord** nta = _space_info[id].new_top_addr(); 1530 bool result = summary_data().summarize(_space_info[id].split_info(), 1531 space->bottom(), space->top(), NULL, 1532 space->bottom(), space->end(), nta); 1533 assert(result, "space must fit into itself"); 1534 } 1535 } 1536 1537 void 1538 PSParallelCompact::provoke_split(bool & max_compaction) 1539 { 1540 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1541 return; 1542 } 1543 1544 const size_t region_size = ParallelCompactData::RegionSize; 1545 ParallelCompactData& sd = summary_data(); 1546 1547 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1548 MutableSpace* const from_space = _space_info[from_space_id].space(); 1549 const size_t eden_live = pointer_delta(eden_space->top(), 1550 _space_info[eden_space_id].new_top()); 1551 const size_t from_live = pointer_delta(from_space->top(), 1552 _space_info[from_space_id].new_top()); 1553 1554 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1555 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1556 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1557 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1558 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1559 1560 // Choose the space to split; need at least 2 regions live (or fillable). 1561 SpaceId id; 1562 MutableSpace* space; 1563 size_t live_words; 1564 size_t fill_words; 1565 if (eden_live + eden_fillable >= region_size * 2) { 1566 id = eden_space_id; 1567 space = eden_space; 1568 live_words = eden_live; 1569 fill_words = eden_fillable; 1570 } else if (from_live + from_fillable >= region_size * 2) { 1571 id = from_space_id; 1572 space = from_space; 1573 live_words = from_live; 1574 fill_words = from_fillable; 1575 } else { 1576 return; // Give up. 1577 } 1578 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1579 1580 if (live_words < region_size * 2) { 1581 // Fill from top() to end() w/live objects of mixed sizes. 1582 HeapWord* const fill_start = space->top(); 1583 live_words += fill_words; 1584 1585 space->set_top(fill_start + fill_words); 1586 if (ZapUnusedHeapArea) { 1587 space->set_top_for_allocations(); 1588 } 1589 1590 HeapWord* cur_addr = fill_start; 1591 while (fill_words > 0) { 1592 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1593 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1594 if (fill_words - cur_size < min_fill_size) { 1595 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1596 } 1597 1598 CollectedHeap::fill_with_object(cur_addr, cur_size); 1599 mark_bitmap()->mark_obj(cur_addr, cur_size); 1600 sd.add_obj(cur_addr, cur_size); 1601 1602 cur_addr += cur_size; 1603 fill_words -= cur_size; 1604 } 1605 1606 summarize_new_objects(id, fill_start); 1607 } 1608 1609 max_compaction = false; 1610 1611 // Manipulate the old gen so that it has room for about half of the live data 1612 // in the target young gen space (live_words / 2). 1613 id = old_space_id; 1614 space = _space_info[id].space(); 1615 const size_t free_at_end = space->free_in_words(); 1616 const size_t free_target = align_object_size(live_words / 2); 1617 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1618 1619 if (free_at_end >= free_target + min_fill_size) { 1620 // Fill space above top() and set the dense prefix so everything survives. 1621 HeapWord* const fill_start = space->top(); 1622 const size_t fill_size = free_at_end - free_target; 1623 space->set_top(space->top() + fill_size); 1624 if (ZapUnusedHeapArea) { 1625 space->set_top_for_allocations(); 1626 } 1627 fill_with_live_objects(id, fill_start, fill_size); 1628 summarize_new_objects(id, fill_start); 1629 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1630 } else if (dead + free_at_end > free_target) { 1631 // Find a dense prefix that makes the right amount of space available. 1632 HeapWord* cur = sd.region_align_down(space->top()); 1633 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1634 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1635 while (dead_to_right < free_target) { 1636 cur -= region_size; 1637 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1638 dead_to_right = pointer_delta(space->end(), cur_destination); 1639 } 1640 _space_info[id].set_dense_prefix(cur); 1641 } 1642 } 1643 #endif // #ifndef PRODUCT 1644 1645 void PSParallelCompact::summarize_spaces_quick() 1646 { 1647 for (unsigned int i = 0; i < last_space_id; ++i) { 1648 const MutableSpace* space = _space_info[i].space(); 1649 HeapWord** nta = _space_info[i].new_top_addr(); 1650 bool result = _summary_data.summarize(_space_info[i].split_info(), 1651 space->bottom(), space->top(), NULL, 1652 space->bottom(), space->end(), nta); 1653 assert(result, "space must fit into itself"); 1654 _space_info[i].set_dense_prefix(space->bottom()); 1655 } 1656 1657 #ifndef PRODUCT 1658 if (ParallelOldGCSplitALot) { 1659 provoke_split_fill_survivor(to_space_id); 1660 } 1661 #endif // #ifndef PRODUCT 1662 } 1663 1664 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1665 { 1666 HeapWord* const dense_prefix_end = dense_prefix(id); 1667 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1668 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1669 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1670 // Only enough dead space is filled so that any remaining dead space to the 1671 // left is larger than the minimum filler object. (The remainder is filled 1672 // during the copy/update phase.) 1673 // 1674 // The size of the dead space to the right of the boundary is not a 1675 // concern, since compaction will be able to use whatever space is 1676 // available. 1677 // 1678 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1679 // surrounds the space to be filled with an object. 1680 // 1681 // In the 32-bit VM, each bit represents two 32-bit words: 1682 // +---+ 1683 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1684 // end_bits: ... x x x | 0 | || 0 x x ... 1685 // +---+ 1686 // 1687 // In the 64-bit VM, each bit represents one 64-bit word: 1688 // +------------+ 1689 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1690 // end_bits: ... x x 1 | 0 || 0 | x x ... 1691 // +------------+ 1692 // +-------+ 1693 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1694 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1695 // +-------+ 1696 // +-----------+ 1697 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1698 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1699 // +-----------+ 1700 // +-------+ 1701 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1702 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1703 // +-------+ 1704 1705 // Initially assume case a, c or e will apply. 1706 size_t obj_len = CollectedHeap::min_fill_size(); 1707 HeapWord* obj_beg = dense_prefix_end - obj_len; 1708 1709 #ifdef _LP64 1710 if (MinObjAlignment > 1) { // object alignment > heap word size 1711 // Cases a, c or e. 1712 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1713 // Case b above. 1714 obj_beg = dense_prefix_end - 1; 1715 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1716 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1717 // Case d above. 1718 obj_beg = dense_prefix_end - 3; 1719 obj_len = 3; 1720 } 1721 #endif // #ifdef _LP64 1722 1723 CollectedHeap::fill_with_object(obj_beg, obj_len); 1724 _mark_bitmap.mark_obj(obj_beg, obj_len); 1725 _summary_data.add_obj(obj_beg, obj_len); 1726 assert(start_array(id) != NULL, "sanity"); 1727 start_array(id)->allocate_block(obj_beg); 1728 } 1729 } 1730 1731 void 1732 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1733 { 1734 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1735 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1736 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1737 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1738 cur->set_source_region(0); 1739 } 1740 } 1741 1742 void 1743 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1744 { 1745 assert(id < last_space_id, "id out of range"); 1746 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1747 ParallelOldGCSplitALot && id == old_space_id, 1748 "should have been reset in summarize_spaces_quick()"); 1749 1750 const MutableSpace* space = _space_info[id].space(); 1751 if (_space_info[id].new_top() != space->bottom()) { 1752 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1753 _space_info[id].set_dense_prefix(dense_prefix_end); 1754 1755 #ifndef PRODUCT 1756 if (TraceParallelOldGCDensePrefix) { 1757 print_dense_prefix_stats("ratio", id, maximum_compaction, 1758 dense_prefix_end); 1759 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1760 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1761 } 1762 #endif // #ifndef PRODUCT 1763 1764 // Recompute the summary data, taking into account the dense prefix. If 1765 // every last byte will be reclaimed, then the existing summary data which 1766 // compacts everything can be left in place. 1767 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1768 // If dead space crosses the dense prefix boundary, it is (at least 1769 // partially) filled with a dummy object, marked live and added to the 1770 // summary data. This simplifies the copy/update phase and must be done 1771 // before the final locations of objects are determined, to prevent 1772 // leaving a fragment of dead space that is too small to fill. 1773 fill_dense_prefix_end(id); 1774 1775 // Compute the destination of each Region, and thus each object. 1776 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1777 _summary_data.summarize(_space_info[id].split_info(), 1778 dense_prefix_end, space->top(), NULL, 1779 dense_prefix_end, space->end(), 1780 _space_info[id].new_top_addr()); 1781 } 1782 } 1783 1784 if (TraceParallelOldGCSummaryPhase) { 1785 const size_t region_size = ParallelCompactData::RegionSize; 1786 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1787 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1788 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1789 HeapWord* const new_top = _space_info[id].new_top(); 1790 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1791 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1792 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1793 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1794 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1795 id, space->capacity_in_words(), p2i(dense_prefix_end), 1796 dp_region, dp_words / region_size, 1797 cr_words / region_size, p2i(new_top)); 1798 } 1799 } 1800 1801 #ifndef PRODUCT 1802 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1803 HeapWord* dst_beg, HeapWord* dst_end, 1804 SpaceId src_space_id, 1805 HeapWord* src_beg, HeapWord* src_end) 1806 { 1807 if (TraceParallelOldGCSummaryPhase) { 1808 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1809 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1810 SIZE_FORMAT "-" SIZE_FORMAT " " 1811 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1812 SIZE_FORMAT "-" SIZE_FORMAT, 1813 src_space_id, space_names[src_space_id], 1814 dst_space_id, space_names[dst_space_id], 1815 p2i(src_beg), p2i(src_end), 1816 _summary_data.addr_to_region_idx(src_beg), 1817 _summary_data.addr_to_region_idx(src_end), 1818 p2i(dst_beg), p2i(dst_end), 1819 _summary_data.addr_to_region_idx(dst_beg), 1820 _summary_data.addr_to_region_idx(dst_end)); 1821 } 1822 } 1823 #endif // #ifndef PRODUCT 1824 1825 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1826 bool maximum_compaction) 1827 { 1828 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer); 1829 // trace("2"); 1830 1831 #ifdef ASSERT 1832 if (TraceParallelOldGCMarkingPhase) { 1833 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1834 "add_obj_bytes=" SIZE_FORMAT, 1835 add_obj_count, add_obj_size * HeapWordSize); 1836 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1837 "mark_bitmap_bytes=" SIZE_FORMAT, 1838 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1839 } 1840 #endif // #ifdef ASSERT 1841 1842 // Quick summarization of each space into itself, to see how much is live. 1843 summarize_spaces_quick(); 1844 1845 if (TraceParallelOldGCSummaryPhase) { 1846 tty->print_cr("summary_phase: after summarizing each space to self"); 1847 Universe::print(); 1848 NOT_PRODUCT(print_region_ranges()); 1849 if (Verbose) { 1850 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1851 } 1852 } 1853 1854 // The amount of live data that will end up in old space (assuming it fits). 1855 size_t old_space_total_live = 0; 1856 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1857 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1858 _space_info[id].space()->bottom()); 1859 } 1860 1861 MutableSpace* const old_space = _space_info[old_space_id].space(); 1862 const size_t old_capacity = old_space->capacity_in_words(); 1863 if (old_space_total_live > old_capacity) { 1864 // XXX - should also try to expand 1865 maximum_compaction = true; 1866 } 1867 #ifndef PRODUCT 1868 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1869 provoke_split(maximum_compaction); 1870 } 1871 #endif // #ifndef PRODUCT 1872 1873 // Old generations. 1874 summarize_space(old_space_id, maximum_compaction); 1875 1876 // Summarize the remaining spaces in the young gen. The initial target space 1877 // is the old gen. If a space does not fit entirely into the target, then the 1878 // remainder is compacted into the space itself and that space becomes the new 1879 // target. 1880 SpaceId dst_space_id = old_space_id; 1881 HeapWord* dst_space_end = old_space->end(); 1882 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1883 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1884 const MutableSpace* space = _space_info[id].space(); 1885 const size_t live = pointer_delta(_space_info[id].new_top(), 1886 space->bottom()); 1887 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1888 1889 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1890 SpaceId(id), space->bottom(), space->top());) 1891 if (live > 0 && live <= available) { 1892 // All the live data will fit. 1893 bool done = _summary_data.summarize(_space_info[id].split_info(), 1894 space->bottom(), space->top(), 1895 NULL, 1896 *new_top_addr, dst_space_end, 1897 new_top_addr); 1898 assert(done, "space must fit into old gen"); 1899 1900 // Reset the new_top value for the space. 1901 _space_info[id].set_new_top(space->bottom()); 1902 } else if (live > 0) { 1903 // Attempt to fit part of the source space into the target space. 1904 HeapWord* next_src_addr = NULL; 1905 bool done = _summary_data.summarize(_space_info[id].split_info(), 1906 space->bottom(), space->top(), 1907 &next_src_addr, 1908 *new_top_addr, dst_space_end, 1909 new_top_addr); 1910 assert(!done, "space should not fit into old gen"); 1911 assert(next_src_addr != NULL, "sanity"); 1912 1913 // The source space becomes the new target, so the remainder is compacted 1914 // within the space itself. 1915 dst_space_id = SpaceId(id); 1916 dst_space_end = space->end(); 1917 new_top_addr = _space_info[id].new_top_addr(); 1918 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1919 space->bottom(), dst_space_end, 1920 SpaceId(id), next_src_addr, space->top());) 1921 done = _summary_data.summarize(_space_info[id].split_info(), 1922 next_src_addr, space->top(), 1923 NULL, 1924 space->bottom(), dst_space_end, 1925 new_top_addr); 1926 assert(done, "space must fit when compacted into itself"); 1927 assert(*new_top_addr <= space->top(), "usage should not grow"); 1928 } 1929 } 1930 1931 if (TraceParallelOldGCSummaryPhase) { 1932 tty->print_cr("summary_phase: after final summarization"); 1933 Universe::print(); 1934 NOT_PRODUCT(print_region_ranges()); 1935 if (Verbose) { 1936 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1937 } 1938 } 1939 } 1940 1941 // This method should contain all heap-specific policy for invoking a full 1942 // collection. invoke_no_policy() will only attempt to compact the heap; it 1943 // will do nothing further. If we need to bail out for policy reasons, scavenge 1944 // before full gc, or any other specialized behavior, it needs to be added here. 1945 // 1946 // Note that this method should only be called from the vm_thread while at a 1947 // safepoint. 1948 // 1949 // Note that the all_soft_refs_clear flag in the collector policy 1950 // may be true because this method can be called without intervening 1951 // activity. For example when the heap space is tight and full measure 1952 // are being taken to free space. 1953 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1954 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1955 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1956 "should be in vm thread"); 1957 1958 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1959 GCCause::Cause gc_cause = heap->gc_cause(); 1960 assert(!heap->is_gc_active(), "not reentrant"); 1961 1962 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1963 IsGCActiveMark mark; 1964 1965 if (ScavengeBeforeFullGC) { 1966 PSScavenge::invoke_no_policy(); 1967 } 1968 1969 const bool clear_all_soft_refs = 1970 heap->collector_policy()->should_clear_all_soft_refs(); 1971 1972 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1973 maximum_heap_compaction); 1974 } 1975 1976 // This method contains no policy. You should probably 1977 // be calling invoke() instead. 1978 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1979 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1980 assert(ref_processor() != NULL, "Sanity"); 1981 1982 if (GC_locker::check_active_before_gc()) { 1983 return false; 1984 } 1985 1986 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1987 1988 GCIdMark gc_id_mark; 1989 _gc_timer.register_gc_start(); 1990 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1991 1992 TimeStamp marking_start; 1993 TimeStamp compaction_start; 1994 TimeStamp collection_exit; 1995 1996 GCCause::Cause gc_cause = heap->gc_cause(); 1997 PSYoungGen* young_gen = heap->young_gen(); 1998 PSOldGen* old_gen = heap->old_gen(); 1999 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2000 2001 // The scope of casr should end after code that can change 2002 // CollectorPolicy::_should_clear_all_soft_refs. 2003 ClearedAllSoftRefs casr(maximum_heap_compaction, 2004 heap->collector_policy()); 2005 2006 if (ZapUnusedHeapArea) { 2007 // Save information needed to minimize mangling 2008 heap->record_gen_tops_before_GC(); 2009 } 2010 2011 heap->pre_full_gc_dump(&_gc_timer); 2012 2013 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2014 2015 // Make sure data structures are sane, make the heap parsable, and do other 2016 // miscellaneous bookkeeping. 2017 PreGCValues pre_gc_values; 2018 pre_compact(&pre_gc_values); 2019 2020 // Get the compaction manager reserved for the VM thread. 2021 ParCompactionManager* const vmthread_cm = 2022 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2023 2024 // Place after pre_compact() where the number of invocations is incremented. 2025 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2026 2027 { 2028 ResourceMark rm; 2029 HandleMark hm; 2030 2031 // Set the number of GC threads to be used in this collection 2032 gc_task_manager()->set_active_gang(); 2033 gc_task_manager()->task_idle_workers(); 2034 2035 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2036 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 2037 TraceCollectorStats tcs(counters()); 2038 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2039 2040 if (TraceOldGenTime) accumulated_time()->start(); 2041 2042 // Let the size policy know we're starting 2043 size_policy->major_collection_begin(); 2044 2045 CodeCache::gc_prologue(); 2046 2047 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2048 2049 ref_processor()->enable_discovery(); 2050 ref_processor()->setup_policy(maximum_heap_compaction); 2051 2052 bool marked_for_unloading = false; 2053 2054 marking_start.update(); 2055 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2056 2057 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2058 && GCCause::is_user_requested_gc(gc_cause); 2059 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2060 2061 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2062 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2063 2064 // adjust_roots() updates Universe::_intArrayKlassObj which is 2065 // needed by the compaction for filling holes in the dense prefix. 2066 adjust_roots(); 2067 2068 compaction_start.update(); 2069 compact(); 2070 2071 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2072 // done before resizing. 2073 post_compact(); 2074 2075 // Let the size policy know we're done 2076 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2077 2078 if (UseAdaptiveSizePolicy) { 2079 if (PrintAdaptiveSizePolicy) { 2080 gclog_or_tty->print("AdaptiveSizeStart: "); 2081 gclog_or_tty->stamp(); 2082 gclog_or_tty->print_cr(" collection: %d ", 2083 heap->total_collections()); 2084 if (Verbose) { 2085 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT 2086 " young_gen_capacity: " SIZE_FORMAT, 2087 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2088 } 2089 } 2090 2091 // Don't check if the size_policy is ready here. Let 2092 // the size_policy check that internally. 2093 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2094 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 2095 // Swap the survivor spaces if from_space is empty. The 2096 // resize_young_gen() called below is normally used after 2097 // a successful young GC and swapping of survivor spaces; 2098 // otherwise, it will fail to resize the young gen with 2099 // the current implementation. 2100 if (young_gen->from_space()->is_empty()) { 2101 young_gen->from_space()->clear(SpaceDecorator::Mangle); 2102 young_gen->swap_spaces(); 2103 } 2104 2105 // Calculate optimal free space amounts 2106 assert(young_gen->max_size() > 2107 young_gen->from_space()->capacity_in_bytes() + 2108 young_gen->to_space()->capacity_in_bytes(), 2109 "Sizes of space in young gen are out-of-bounds"); 2110 2111 size_t young_live = young_gen->used_in_bytes(); 2112 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2113 size_t old_live = old_gen->used_in_bytes(); 2114 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2115 size_t max_old_gen_size = old_gen->max_gen_size(); 2116 size_t max_eden_size = young_gen->max_size() - 2117 young_gen->from_space()->capacity_in_bytes() - 2118 young_gen->to_space()->capacity_in_bytes(); 2119 2120 // Used for diagnostics 2121 size_policy->clear_generation_free_space_flags(); 2122 2123 size_policy->compute_generations_free_space(young_live, 2124 eden_live, 2125 old_live, 2126 cur_eden, 2127 max_old_gen_size, 2128 max_eden_size, 2129 true /* full gc*/); 2130 2131 size_policy->check_gc_overhead_limit(young_live, 2132 eden_live, 2133 max_old_gen_size, 2134 max_eden_size, 2135 true /* full gc*/, 2136 gc_cause, 2137 heap->collector_policy()); 2138 2139 size_policy->decay_supplemental_growth(true /* full gc*/); 2140 2141 heap->resize_old_gen( 2142 size_policy->calculated_old_free_size_in_bytes()); 2143 2144 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 2145 size_policy->calculated_survivor_size_in_bytes()); 2146 } 2147 if (PrintAdaptiveSizePolicy) { 2148 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2149 heap->total_collections()); 2150 } 2151 } 2152 2153 if (UsePerfData) { 2154 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2155 counters->update_counters(); 2156 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2157 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2158 } 2159 2160 heap->resize_all_tlabs(); 2161 2162 // Resize the metaspace capacity after a collection 2163 MetaspaceGC::compute_new_size(); 2164 2165 if (TraceOldGenTime) accumulated_time()->stop(); 2166 2167 if (PrintGC) { 2168 if (PrintGCDetails) { 2169 // No GC timestamp here. This is after GC so it would be confusing. 2170 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2171 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2172 heap->print_heap_change(pre_gc_values.heap_used()); 2173 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2174 } else { 2175 heap->print_heap_change(pre_gc_values.heap_used()); 2176 } 2177 } 2178 2179 // Track memory usage and detect low memory 2180 MemoryService::track_memory_usage(); 2181 heap->update_counters(); 2182 gc_task_manager()->release_idle_workers(); 2183 } 2184 2185 #ifdef ASSERT 2186 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2187 ParCompactionManager* const cm = 2188 ParCompactionManager::manager_array(int(i)); 2189 assert(cm->marking_stack()->is_empty(), "should be empty"); 2190 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2191 } 2192 #endif // ASSERT 2193 2194 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2195 HandleMark hm; // Discard invalid handles created during verification 2196 Universe::verify(" VerifyAfterGC:"); 2197 } 2198 2199 // Re-verify object start arrays 2200 if (VerifyObjectStartArray && 2201 VerifyAfterGC) { 2202 old_gen->verify_object_start_array(); 2203 } 2204 2205 if (ZapUnusedHeapArea) { 2206 old_gen->object_space()->check_mangled_unused_area_complete(); 2207 } 2208 2209 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2210 2211 collection_exit.update(); 2212 2213 heap->print_heap_after_gc(); 2214 heap->trace_heap_after_gc(&_gc_tracer); 2215 2216 if (PrintGCTaskTimeStamps) { 2217 gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " 2218 JLONG_FORMAT, 2219 marking_start.ticks(), compaction_start.ticks(), 2220 collection_exit.ticks()); 2221 gc_task_manager()->print_task_time_stamps(); 2222 } 2223 2224 heap->post_full_gc_dump(&_gc_timer); 2225 2226 #ifdef TRACESPINNING 2227 ParallelTaskTerminator::print_termination_counts(); 2228 #endif 2229 2230 _gc_timer.register_gc_end(); 2231 2232 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2233 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2234 2235 return true; 2236 } 2237 2238 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2239 PSYoungGen* young_gen, 2240 PSOldGen* old_gen) { 2241 MutableSpace* const eden_space = young_gen->eden_space(); 2242 assert(!eden_space->is_empty(), "eden must be non-empty"); 2243 assert(young_gen->virtual_space()->alignment() == 2244 old_gen->virtual_space()->alignment(), "alignments do not match"); 2245 2246 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2247 return false; 2248 } 2249 2250 // Both generations must be completely committed. 2251 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2252 return false; 2253 } 2254 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2255 return false; 2256 } 2257 2258 // Figure out how much to take from eden. Include the average amount promoted 2259 // in the total; otherwise the next young gen GC will simply bail out to a 2260 // full GC. 2261 const size_t alignment = old_gen->virtual_space()->alignment(); 2262 const size_t eden_used = eden_space->used_in_bytes(); 2263 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2264 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2265 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2266 2267 if (absorb_size >= eden_capacity) { 2268 return false; // Must leave some space in eden. 2269 } 2270 2271 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2272 if (new_young_size < young_gen->min_gen_size()) { 2273 return false; // Respect young gen minimum size. 2274 } 2275 2276 if (TraceAdaptiveGCBoundary && Verbose) { 2277 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2278 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2279 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2280 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2281 absorb_size / K, 2282 eden_capacity / K, (eden_capacity - absorb_size) / K, 2283 young_gen->from_space()->used_in_bytes() / K, 2284 young_gen->to_space()->used_in_bytes() / K, 2285 young_gen->capacity_in_bytes() / K, new_young_size / K); 2286 } 2287 2288 // Fill the unused part of the old gen. 2289 MutableSpace* const old_space = old_gen->object_space(); 2290 HeapWord* const unused_start = old_space->top(); 2291 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2292 2293 if (unused_words > 0) { 2294 if (unused_words < CollectedHeap::min_fill_size()) { 2295 return false; // If the old gen cannot be filled, must give up. 2296 } 2297 CollectedHeap::fill_with_objects(unused_start, unused_words); 2298 } 2299 2300 // Take the live data from eden and set both top and end in the old gen to 2301 // eden top. (Need to set end because reset_after_change() mangles the region 2302 // from end to virtual_space->high() in debug builds). 2303 HeapWord* const new_top = eden_space->top(); 2304 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2305 absorb_size); 2306 young_gen->reset_after_change(); 2307 old_space->set_top(new_top); 2308 old_space->set_end(new_top); 2309 old_gen->reset_after_change(); 2310 2311 // Update the object start array for the filler object and the data from eden. 2312 ObjectStartArray* const start_array = old_gen->start_array(); 2313 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2314 start_array->allocate_block(p); 2315 } 2316 2317 // Could update the promoted average here, but it is not typically updated at 2318 // full GCs and the value to use is unclear. Something like 2319 // 2320 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2321 2322 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2323 return true; 2324 } 2325 2326 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2327 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2328 "shouldn't return NULL"); 2329 return ParallelScavengeHeap::gc_task_manager(); 2330 } 2331 2332 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2333 bool maximum_heap_compaction, 2334 ParallelOldTracer *gc_tracer) { 2335 // Recursively traverse all live objects and mark them 2336 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer); 2337 2338 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2339 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2340 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2341 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2342 ParallelTaskTerminator terminator(active_gc_threads, qset); 2343 2344 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm); 2345 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2346 2347 // Need new claim bits before marking starts. 2348 ClassLoaderDataGraph::clear_claimed_marks(); 2349 2350 { 2351 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer); 2352 2353 ParallelScavengeHeap::ParStrongRootsScope psrs; 2354 2355 GCTaskQueue* q = GCTaskQueue::create(); 2356 2357 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2358 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2359 // We scan the thread roots in parallel 2360 Threads::create_thread_roots_marking_tasks(q); 2361 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2362 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2363 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2364 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2365 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2368 2369 if (active_gc_threads > 1) { 2370 for (uint j = 0; j < active_gc_threads; j++) { 2371 q->enqueue(new StealMarkingTask(&terminator)); 2372 } 2373 } 2374 2375 gc_task_manager()->execute_and_wait(q); 2376 } 2377 2378 // Process reference objects found during marking 2379 { 2380 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer); 2381 2382 ReferenceProcessorStats stats; 2383 if (ref_processor()->processing_is_mt()) { 2384 RefProcTaskExecutor task_executor; 2385 stats = ref_processor()->process_discovered_references( 2386 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2387 &task_executor, &_gc_timer); 2388 } else { 2389 stats = ref_processor()->process_discovered_references( 2390 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2391 &_gc_timer); 2392 } 2393 2394 gc_tracer->report_gc_reference_stats(stats); 2395 } 2396 2397 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer); 2398 2399 // This is the point where the entire marking should have completed. 2400 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2401 2402 // Follow system dictionary roots and unload classes. 2403 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2404 2405 // Unload nmethods. 2406 CodeCache::do_unloading(is_alive_closure(), purged_class); 2407 2408 // Prune dead klasses from subklass/sibling/implementor lists. 2409 Klass::clean_weak_klass_links(is_alive_closure()); 2410 2411 // Delete entries for dead interned strings. 2412 StringTable::unlink(is_alive_closure()); 2413 2414 // Clean up unreferenced symbols in symbol table. 2415 SymbolTable::unlink(); 2416 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2417 } 2418 2419 // This should be moved to the shared markSweep code! 2420 class PSAlwaysTrueClosure: public BoolObjectClosure { 2421 public: 2422 bool do_object_b(oop p) { return true; } 2423 }; 2424 static PSAlwaysTrueClosure always_true; 2425 2426 void PSParallelCompact::adjust_roots() { 2427 // Adjust the pointers to reflect the new locations 2428 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer); 2429 2430 // Need new claim bits when tracing through and adjusting pointers. 2431 ClassLoaderDataGraph::clear_claimed_marks(); 2432 2433 // General strong roots. 2434 Universe::oops_do(adjust_pointer_closure()); 2435 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2436 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2437 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2438 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2439 FlatProfiler::oops_do(adjust_pointer_closure()); 2440 Management::oops_do(adjust_pointer_closure()); 2441 JvmtiExport::oops_do(adjust_pointer_closure()); 2442 SystemDictionary::oops_do(adjust_pointer_closure()); 2443 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2444 2445 // Now adjust pointers in remaining weak roots. (All of which should 2446 // have been cleared if they pointed to non-surviving objects.) 2447 // Global (weak) JNI handles 2448 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2449 2450 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2451 CodeCache::blobs_do(&adjust_from_blobs); 2452 StringTable::oops_do(adjust_pointer_closure()); 2453 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2454 // Roots were visited so references into the young gen in roots 2455 // may have been scanned. Process them also. 2456 // Should the reference processor have a span that excludes 2457 // young gen objects? 2458 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2459 } 2460 2461 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2462 uint parallel_gc_threads) 2463 { 2464 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer); 2465 2466 // Find the threads that are active 2467 unsigned int which = 0; 2468 2469 const uint task_count = MAX2(parallel_gc_threads, 1U); 2470 for (uint j = 0; j < task_count; j++) { 2471 q->enqueue(new DrainStacksCompactionTask(j)); 2472 ParCompactionManager::verify_region_list_empty(j); 2473 // Set the region stacks variables to "no" region stack values 2474 // so that they will be recognized and needing a region stack 2475 // in the stealing tasks if they do not get one by executing 2476 // a draining stack. 2477 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2478 cm->set_region_stack(NULL); 2479 cm->set_region_stack_index((uint)max_uintx); 2480 } 2481 ParCompactionManager::reset_recycled_stack_index(); 2482 2483 // Find all regions that are available (can be filled immediately) and 2484 // distribute them to the thread stacks. The iteration is done in reverse 2485 // order (high to low) so the regions will be removed in ascending order. 2486 2487 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2488 2489 size_t fillable_regions = 0; // A count for diagnostic purposes. 2490 // A region index which corresponds to the tasks created above. 2491 // "which" must be 0 <= which < task_count 2492 2493 which = 0; 2494 // id + 1 is used to test termination so unsigned can 2495 // be used with an old_space_id == 0. 2496 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2497 SpaceInfo* const space_info = _space_info + id; 2498 MutableSpace* const space = space_info->space(); 2499 HeapWord* const new_top = space_info->new_top(); 2500 2501 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2502 const size_t end_region = 2503 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2504 2505 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2506 if (sd.region(cur)->claim_unsafe()) { 2507 ParCompactionManager::region_list_push(which, cur); 2508 2509 if (TraceParallelOldGCCompactionPhase && Verbose) { 2510 const size_t count_mod_8 = fillable_regions & 7; 2511 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2512 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2513 if (count_mod_8 == 7) gclog_or_tty->cr(); 2514 } 2515 2516 NOT_PRODUCT(++fillable_regions;) 2517 2518 // Assign regions to tasks in round-robin fashion. 2519 if (++which == task_count) { 2520 assert(which <= parallel_gc_threads, 2521 "Inconsistent number of workers"); 2522 which = 0; 2523 } 2524 } 2525 } 2526 } 2527 2528 if (TraceParallelOldGCCompactionPhase) { 2529 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2530 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions); 2531 } 2532 } 2533 2534 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2535 2536 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2537 uint parallel_gc_threads) { 2538 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer); 2539 2540 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2541 2542 // Iterate over all the spaces adding tasks for updating 2543 // regions in the dense prefix. Assume that 1 gc thread 2544 // will work on opening the gaps and the remaining gc threads 2545 // will work on the dense prefix. 2546 unsigned int space_id; 2547 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2548 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2549 const MutableSpace* const space = _space_info[space_id].space(); 2550 2551 if (dense_prefix_end == space->bottom()) { 2552 // There is no dense prefix for this space. 2553 continue; 2554 } 2555 2556 // The dense prefix is before this region. 2557 size_t region_index_end_dense_prefix = 2558 sd.addr_to_region_idx(dense_prefix_end); 2559 RegionData* const dense_prefix_cp = 2560 sd.region(region_index_end_dense_prefix); 2561 assert(dense_prefix_end == space->end() || 2562 dense_prefix_cp->available() || 2563 dense_prefix_cp->claimed(), 2564 "The region after the dense prefix should always be ready to fill"); 2565 2566 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2567 2568 // Is there dense prefix work? 2569 size_t total_dense_prefix_regions = 2570 region_index_end_dense_prefix - region_index_start; 2571 // How many regions of the dense prefix should be given to 2572 // each thread? 2573 if (total_dense_prefix_regions > 0) { 2574 uint tasks_for_dense_prefix = 1; 2575 if (total_dense_prefix_regions <= 2576 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2577 // Don't over partition. This assumes that 2578 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2579 // so there are not many regions to process. 2580 tasks_for_dense_prefix = parallel_gc_threads; 2581 } else { 2582 // Over partition 2583 tasks_for_dense_prefix = parallel_gc_threads * 2584 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2585 } 2586 size_t regions_per_thread = total_dense_prefix_regions / 2587 tasks_for_dense_prefix; 2588 // Give each thread at least 1 region. 2589 if (regions_per_thread == 0) { 2590 regions_per_thread = 1; 2591 } 2592 2593 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2594 if (region_index_start >= region_index_end_dense_prefix) { 2595 break; 2596 } 2597 // region_index_end is not processed 2598 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2599 region_index_end_dense_prefix); 2600 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2601 region_index_start, 2602 region_index_end)); 2603 region_index_start = region_index_end; 2604 } 2605 } 2606 // This gets any part of the dense prefix that did not 2607 // fit evenly. 2608 if (region_index_start < region_index_end_dense_prefix) { 2609 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2610 region_index_start, 2611 region_index_end_dense_prefix)); 2612 } 2613 } 2614 } 2615 2616 void PSParallelCompact::enqueue_region_stealing_tasks( 2617 GCTaskQueue* q, 2618 ParallelTaskTerminator* terminator_ptr, 2619 uint parallel_gc_threads) { 2620 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer); 2621 2622 // Once a thread has drained it's stack, it should try to steal regions from 2623 // other threads. 2624 if (parallel_gc_threads > 1) { 2625 for (uint j = 0; j < parallel_gc_threads; j++) { 2626 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2627 } 2628 } 2629 } 2630 2631 #ifdef ASSERT 2632 // Write a histogram of the number of times the block table was filled for a 2633 // region. 2634 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2635 { 2636 if (!TraceParallelOldGCCompactionPhase) return; 2637 2638 typedef ParallelCompactData::RegionData rd_t; 2639 ParallelCompactData& sd = summary_data(); 2640 2641 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2642 MutableSpace* const spc = _space_info[id].space(); 2643 if (spc->bottom() != spc->top()) { 2644 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2645 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2646 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2647 2648 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2649 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2650 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2651 2652 for (const rd_t* cur = beg; cur < end; ++cur) { 2653 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2654 } 2655 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2656 for (size_t i = 0; i < histo_len; ++i) { 2657 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2658 histo[i], 100.0 * histo[i] / region_cnt); 2659 } 2660 out->cr(); 2661 } 2662 } 2663 } 2664 #endif // #ifdef ASSERT 2665 2666 void PSParallelCompact::compact() { 2667 // trace("5"); 2668 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer); 2669 2670 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2671 PSOldGen* old_gen = heap->old_gen(); 2672 old_gen->start_array()->reset(); 2673 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2674 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2675 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2676 ParallelTaskTerminator terminator(active_gc_threads, qset); 2677 2678 GCTaskQueue* q = GCTaskQueue::create(); 2679 enqueue_region_draining_tasks(q, active_gc_threads); 2680 enqueue_dense_prefix_tasks(q, active_gc_threads); 2681 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2682 2683 { 2684 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer); 2685 2686 gc_task_manager()->execute_and_wait(q); 2687 2688 #ifdef ASSERT 2689 // Verify that all regions have been processed before the deferred updates. 2690 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2691 verify_complete(SpaceId(id)); 2692 } 2693 #endif 2694 } 2695 2696 { 2697 // Update the deferred objects, if any. Any compaction manager can be used. 2698 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer); 2699 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2700 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2701 update_deferred_objects(cm, SpaceId(id)); 2702 } 2703 } 2704 2705 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2706 } 2707 2708 #ifdef ASSERT 2709 void PSParallelCompact::verify_complete(SpaceId space_id) { 2710 // All Regions between space bottom() to new_top() should be marked as filled 2711 // and all Regions between new_top() and top() should be available (i.e., 2712 // should have been emptied). 2713 ParallelCompactData& sd = summary_data(); 2714 SpaceInfo si = _space_info[space_id]; 2715 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2716 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2717 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2718 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2719 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2720 2721 bool issued_a_warning = false; 2722 2723 size_t cur_region; 2724 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2725 const RegionData* const c = sd.region(cur_region); 2726 if (!c->completed()) { 2727 warning("region " SIZE_FORMAT " not filled: " 2728 "destination_count=%u", 2729 cur_region, c->destination_count()); 2730 issued_a_warning = true; 2731 } 2732 } 2733 2734 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2735 const RegionData* const c = sd.region(cur_region); 2736 if (!c->available()) { 2737 warning("region " SIZE_FORMAT " not empty: " 2738 "destination_count=%u", 2739 cur_region, c->destination_count()); 2740 issued_a_warning = true; 2741 } 2742 } 2743 2744 if (issued_a_warning) { 2745 print_region_ranges(); 2746 } 2747 } 2748 #endif // #ifdef ASSERT 2749 2750 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2751 _start_array->allocate_block(addr); 2752 compaction_manager()->update_contents(oop(addr)); 2753 } 2754 2755 // Update interior oops in the ranges of regions [beg_region, end_region). 2756 void 2757 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2758 SpaceId space_id, 2759 size_t beg_region, 2760 size_t end_region) { 2761 ParallelCompactData& sd = summary_data(); 2762 ParMarkBitMap* const mbm = mark_bitmap(); 2763 2764 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2765 HeapWord* const end_addr = sd.region_to_addr(end_region); 2766 assert(beg_region <= end_region, "bad region range"); 2767 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2768 2769 #ifdef ASSERT 2770 // Claim the regions to avoid triggering an assert when they are marked as 2771 // filled. 2772 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2773 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2774 } 2775 #endif // #ifdef ASSERT 2776 2777 if (beg_addr != space(space_id)->bottom()) { 2778 // Find the first live object or block of dead space that *starts* in this 2779 // range of regions. If a partial object crosses onto the region, skip it; 2780 // it will be marked for 'deferred update' when the object head is 2781 // processed. If dead space crosses onto the region, it is also skipped; it 2782 // will be filled when the prior region is processed. If neither of those 2783 // apply, the first word in the region is the start of a live object or dead 2784 // space. 2785 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2786 const RegionData* const cp = sd.region(beg_region); 2787 if (cp->partial_obj_size() != 0) { 2788 beg_addr = sd.partial_obj_end(beg_region); 2789 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2790 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2791 } 2792 } 2793 2794 if (beg_addr < end_addr) { 2795 // A live object or block of dead space starts in this range of Regions. 2796 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2797 2798 // Create closures and iterate. 2799 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2800 FillClosure fill_closure(cm, space_id); 2801 ParMarkBitMap::IterationStatus status; 2802 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2803 dense_prefix_end); 2804 if (status == ParMarkBitMap::incomplete) { 2805 update_closure.do_addr(update_closure.source()); 2806 } 2807 } 2808 2809 // Mark the regions as filled. 2810 RegionData* const beg_cp = sd.region(beg_region); 2811 RegionData* const end_cp = sd.region(end_region); 2812 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2813 cp->set_completed(); 2814 } 2815 } 2816 2817 // Return the SpaceId for the space containing addr. If addr is not in the 2818 // heap, last_space_id is returned. In debug mode it expects the address to be 2819 // in the heap and asserts such. 2820 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2821 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2822 2823 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2824 if (_space_info[id].space()->contains(addr)) { 2825 return SpaceId(id); 2826 } 2827 } 2828 2829 assert(false, "no space contains the addr"); 2830 return last_space_id; 2831 } 2832 2833 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2834 SpaceId id) { 2835 assert(id < last_space_id, "bad space id"); 2836 2837 ParallelCompactData& sd = summary_data(); 2838 const SpaceInfo* const space_info = _space_info + id; 2839 ObjectStartArray* const start_array = space_info->start_array(); 2840 2841 const MutableSpace* const space = space_info->space(); 2842 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2843 HeapWord* const beg_addr = space_info->dense_prefix(); 2844 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2845 2846 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2847 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2848 const RegionData* cur_region; 2849 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2850 HeapWord* const addr = cur_region->deferred_obj_addr(); 2851 if (addr != NULL) { 2852 if (start_array != NULL) { 2853 start_array->allocate_block(addr); 2854 } 2855 cm->update_contents(oop(addr)); 2856 assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)))); 2857 } 2858 } 2859 } 2860 2861 // Skip over count live words starting from beg, and return the address of the 2862 // next live word. Unless marked, the word corresponding to beg is assumed to 2863 // be dead. Callers must either ensure beg does not correspond to the middle of 2864 // an object, or account for those live words in some other way. Callers must 2865 // also ensure that there are enough live words in the range [beg, end) to skip. 2866 HeapWord* 2867 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2868 { 2869 assert(count > 0, "sanity"); 2870 2871 ParMarkBitMap* m = mark_bitmap(); 2872 idx_t bits_to_skip = m->words_to_bits(count); 2873 idx_t cur_beg = m->addr_to_bit(beg); 2874 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2875 2876 do { 2877 cur_beg = m->find_obj_beg(cur_beg, search_end); 2878 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2879 const size_t obj_bits = cur_end - cur_beg + 1; 2880 if (obj_bits > bits_to_skip) { 2881 return m->bit_to_addr(cur_beg + bits_to_skip); 2882 } 2883 bits_to_skip -= obj_bits; 2884 cur_beg = cur_end + 1; 2885 } while (bits_to_skip > 0); 2886 2887 // Skipping the desired number of words landed just past the end of an object. 2888 // Find the start of the next object. 2889 cur_beg = m->find_obj_beg(cur_beg, search_end); 2890 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2891 return m->bit_to_addr(cur_beg); 2892 } 2893 2894 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2895 SpaceId src_space_id, 2896 size_t src_region_idx) 2897 { 2898 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2899 2900 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2901 if (split_info.dest_region_addr() == dest_addr) { 2902 // The partial object ending at the split point contains the first word to 2903 // be copied to dest_addr. 2904 return split_info.first_src_addr(); 2905 } 2906 2907 const ParallelCompactData& sd = summary_data(); 2908 ParMarkBitMap* const bitmap = mark_bitmap(); 2909 const size_t RegionSize = ParallelCompactData::RegionSize; 2910 2911 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2912 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2913 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2914 HeapWord* const src_region_destination = src_region_ptr->destination(); 2915 2916 assert(dest_addr >= src_region_destination, "wrong src region"); 2917 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2918 2919 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2920 HeapWord* const src_region_end = src_region_beg + RegionSize; 2921 2922 HeapWord* addr = src_region_beg; 2923 if (dest_addr == src_region_destination) { 2924 // Return the first live word in the source region. 2925 if (partial_obj_size == 0) { 2926 addr = bitmap->find_obj_beg(addr, src_region_end); 2927 assert(addr < src_region_end, "no objects start in src region"); 2928 } 2929 return addr; 2930 } 2931 2932 // Must skip some live data. 2933 size_t words_to_skip = dest_addr - src_region_destination; 2934 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2935 2936 if (partial_obj_size >= words_to_skip) { 2937 // All the live words to skip are part of the partial object. 2938 addr += words_to_skip; 2939 if (partial_obj_size == words_to_skip) { 2940 // Find the first live word past the partial object. 2941 addr = bitmap->find_obj_beg(addr, src_region_end); 2942 assert(addr < src_region_end, "wrong src region"); 2943 } 2944 return addr; 2945 } 2946 2947 // Skip over the partial object (if any). 2948 if (partial_obj_size != 0) { 2949 words_to_skip -= partial_obj_size; 2950 addr += partial_obj_size; 2951 } 2952 2953 // Skip over live words due to objects that start in the region. 2954 addr = skip_live_words(addr, src_region_end, words_to_skip); 2955 assert(addr < src_region_end, "wrong src region"); 2956 return addr; 2957 } 2958 2959 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2960 SpaceId src_space_id, 2961 size_t beg_region, 2962 HeapWord* end_addr) 2963 { 2964 ParallelCompactData& sd = summary_data(); 2965 2966 #ifdef ASSERT 2967 MutableSpace* const src_space = _space_info[src_space_id].space(); 2968 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2969 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2970 "src_space_id does not match beg_addr"); 2971 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2972 "src_space_id does not match end_addr"); 2973 #endif // #ifdef ASSERT 2974 2975 RegionData* const beg = sd.region(beg_region); 2976 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2977 2978 // Regions up to new_top() are enqueued if they become available. 2979 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2980 RegionData* const enqueue_end = 2981 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2982 2983 for (RegionData* cur = beg; cur < end; ++cur) { 2984 assert(cur->data_size() > 0, "region must have live data"); 2985 cur->decrement_destination_count(); 2986 if (cur < enqueue_end && cur->available() && cur->claim()) { 2987 cm->push_region(sd.region(cur)); 2988 } 2989 } 2990 } 2991 2992 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2993 SpaceId& src_space_id, 2994 HeapWord*& src_space_top, 2995 HeapWord* end_addr) 2996 { 2997 typedef ParallelCompactData::RegionData RegionData; 2998 2999 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3000 const size_t region_size = ParallelCompactData::RegionSize; 3001 3002 size_t src_region_idx = 0; 3003 3004 // Skip empty regions (if any) up to the top of the space. 3005 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3006 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3007 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3008 const RegionData* const top_region_ptr = 3009 sd.addr_to_region_ptr(top_aligned_up); 3010 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3011 ++src_region_ptr; 3012 } 3013 3014 if (src_region_ptr < top_region_ptr) { 3015 // The next source region is in the current space. Update src_region_idx 3016 // and the source address to match src_region_ptr. 3017 src_region_idx = sd.region(src_region_ptr); 3018 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3019 if (src_region_addr > closure.source()) { 3020 closure.set_source(src_region_addr); 3021 } 3022 return src_region_idx; 3023 } 3024 3025 // Switch to a new source space and find the first non-empty region. 3026 unsigned int space_id = src_space_id + 1; 3027 assert(space_id < last_space_id, "not enough spaces"); 3028 3029 HeapWord* const destination = closure.destination(); 3030 3031 do { 3032 MutableSpace* space = _space_info[space_id].space(); 3033 HeapWord* const bottom = space->bottom(); 3034 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3035 3036 // Iterate over the spaces that do not compact into themselves. 3037 if (bottom_cp->destination() != bottom) { 3038 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3039 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3040 3041 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3042 if (src_cp->live_obj_size() > 0) { 3043 // Found it. 3044 assert(src_cp->destination() == destination, 3045 "first live obj in the space must match the destination"); 3046 assert(src_cp->partial_obj_size() == 0, 3047 "a space cannot begin with a partial obj"); 3048 3049 src_space_id = SpaceId(space_id); 3050 src_space_top = space->top(); 3051 const size_t src_region_idx = sd.region(src_cp); 3052 closure.set_source(sd.region_to_addr(src_region_idx)); 3053 return src_region_idx; 3054 } else { 3055 assert(src_cp->data_size() == 0, "sanity"); 3056 } 3057 } 3058 } 3059 } while (++space_id < last_space_id); 3060 3061 assert(false, "no source region was found"); 3062 return 0; 3063 } 3064 3065 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3066 { 3067 typedef ParMarkBitMap::IterationStatus IterationStatus; 3068 const size_t RegionSize = ParallelCompactData::RegionSize; 3069 ParMarkBitMap* const bitmap = mark_bitmap(); 3070 ParallelCompactData& sd = summary_data(); 3071 RegionData* const region_ptr = sd.region(region_idx); 3072 3073 // Get the items needed to construct the closure. 3074 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3075 SpaceId dest_space_id = space_id(dest_addr); 3076 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3077 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3078 assert(dest_addr < new_top, "sanity"); 3079 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3080 3081 // Get the source region and related info. 3082 size_t src_region_idx = region_ptr->source_region(); 3083 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3084 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3085 3086 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3087 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3088 3089 // Adjust src_region_idx to prepare for decrementing destination counts (the 3090 // destination count is not decremented when a region is copied to itself). 3091 if (src_region_idx == region_idx) { 3092 src_region_idx += 1; 3093 } 3094 3095 if (bitmap->is_unmarked(closure.source())) { 3096 // The first source word is in the middle of an object; copy the remainder 3097 // of the object or as much as will fit. The fact that pointer updates were 3098 // deferred will be noted when the object header is processed. 3099 HeapWord* const old_src_addr = closure.source(); 3100 closure.copy_partial_obj(); 3101 if (closure.is_full()) { 3102 decrement_destination_counts(cm, src_space_id, src_region_idx, 3103 closure.source()); 3104 region_ptr->set_deferred_obj_addr(NULL); 3105 region_ptr->set_completed(); 3106 return; 3107 } 3108 3109 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3110 if (sd.region_align_down(old_src_addr) != end_addr) { 3111 // The partial object was copied from more than one source region. 3112 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3113 3114 // Move to the next source region, possibly switching spaces as well. All 3115 // args except end_addr may be modified. 3116 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3117 end_addr); 3118 } 3119 } 3120 3121 do { 3122 HeapWord* const cur_addr = closure.source(); 3123 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3124 src_space_top); 3125 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3126 3127 if (status == ParMarkBitMap::incomplete) { 3128 // The last obj that starts in the source region does not end in the 3129 // region. 3130 assert(closure.source() < end_addr, "sanity"); 3131 HeapWord* const obj_beg = closure.source(); 3132 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3133 src_space_top); 3134 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3135 if (obj_end < range_end) { 3136 // The end was found; the entire object will fit. 3137 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3138 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3139 } else { 3140 // The end was not found; the object will not fit. 3141 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3142 status = ParMarkBitMap::would_overflow; 3143 } 3144 } 3145 3146 if (status == ParMarkBitMap::would_overflow) { 3147 // The last object did not fit. Note that interior oop updates were 3148 // deferred, then copy enough of the object to fill the region. 3149 region_ptr->set_deferred_obj_addr(closure.destination()); 3150 status = closure.copy_until_full(); // copies from closure.source() 3151 3152 decrement_destination_counts(cm, src_space_id, src_region_idx, 3153 closure.source()); 3154 region_ptr->set_completed(); 3155 return; 3156 } 3157 3158 if (status == ParMarkBitMap::full) { 3159 decrement_destination_counts(cm, src_space_id, src_region_idx, 3160 closure.source()); 3161 region_ptr->set_deferred_obj_addr(NULL); 3162 region_ptr->set_completed(); 3163 return; 3164 } 3165 3166 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3167 3168 // Move to the next source region, possibly switching spaces as well. All 3169 // args except end_addr may be modified. 3170 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3171 end_addr); 3172 } while (true); 3173 } 3174 3175 void PSParallelCompact::fill_blocks(size_t region_idx) 3176 { 3177 // Fill in the block table elements for the specified region. Each block 3178 // table element holds the number of live words in the region that are to the 3179 // left of the first object that starts in the block. Thus only blocks in 3180 // which an object starts need to be filled. 3181 // 3182 // The algorithm scans the section of the bitmap that corresponds to the 3183 // region, keeping a running total of the live words. When an object start is 3184 // found, if it's the first to start in the block that contains it, the 3185 // current total is written to the block table element. 3186 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3187 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3188 const size_t RegionSize = ParallelCompactData::RegionSize; 3189 3190 ParallelCompactData& sd = summary_data(); 3191 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3192 if (partial_obj_size >= RegionSize) { 3193 return; // No objects start in this region. 3194 } 3195 3196 // Ensure the first loop iteration decides that the block has changed. 3197 size_t cur_block = sd.block_count(); 3198 3199 const ParMarkBitMap* const bitmap = mark_bitmap(); 3200 3201 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3202 assert((size_t)1 << Log2BitsPerBlock == 3203 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3204 3205 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3206 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3207 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3208 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3209 while (beg_bit < range_end) { 3210 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3211 if (new_block != cur_block) { 3212 cur_block = new_block; 3213 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3214 } 3215 3216 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3217 if (end_bit < range_end - 1) { 3218 live_bits += end_bit - beg_bit + 1; 3219 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3220 } else { 3221 return; 3222 } 3223 } 3224 } 3225 3226 void 3227 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3228 const MutableSpace* sp = space(space_id); 3229 if (sp->is_empty()) { 3230 return; 3231 } 3232 3233 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3234 ParMarkBitMap* const bitmap = mark_bitmap(); 3235 HeapWord* const dp_addr = dense_prefix(space_id); 3236 HeapWord* beg_addr = sp->bottom(); 3237 HeapWord* end_addr = sp->top(); 3238 3239 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3240 3241 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3242 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3243 if (beg_region < dp_region) { 3244 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3245 } 3246 3247 // The destination of the first live object that starts in the region is one 3248 // past the end of the partial object entering the region (if any). 3249 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3250 HeapWord* const new_top = _space_info[space_id].new_top(); 3251 assert(new_top >= dest_addr, "bad new_top value"); 3252 const size_t words = pointer_delta(new_top, dest_addr); 3253 3254 if (words > 0) { 3255 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3256 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3257 3258 ParMarkBitMap::IterationStatus status; 3259 status = bitmap->iterate(&closure, dest_addr, end_addr); 3260 assert(status == ParMarkBitMap::full, "iteration not complete"); 3261 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3262 "live objects skipped because closure is full"); 3263 } 3264 } 3265 3266 jlong PSParallelCompact::millis_since_last_gc() { 3267 // We need a monotonically non-decreasing time in ms but 3268 // os::javaTimeMillis() does not guarantee monotonicity. 3269 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3270 jlong ret_val = now - _time_of_last_gc; 3271 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3272 if (ret_val < 0) { 3273 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);) 3274 return 0; 3275 } 3276 return ret_val; 3277 } 3278 3279 void PSParallelCompact::reset_millis_since_last_gc() { 3280 // We need a monotonically non-decreasing time in ms but 3281 // os::javaTimeMillis() does not guarantee monotonicity. 3282 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3283 } 3284 3285 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3286 { 3287 if (source() != destination()) { 3288 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3289 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3290 } 3291 update_state(words_remaining()); 3292 assert(is_full(), "sanity"); 3293 return ParMarkBitMap::full; 3294 } 3295 3296 void MoveAndUpdateClosure::copy_partial_obj() 3297 { 3298 size_t words = words_remaining(); 3299 3300 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3301 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3302 if (end_addr < range_end) { 3303 words = bitmap()->obj_size(source(), end_addr); 3304 } 3305 3306 // This test is necessary; if omitted, the pointer updates to a partial object 3307 // that crosses the dense prefix boundary could be overwritten. 3308 if (source() != destination()) { 3309 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3310 Copy::aligned_conjoint_words(source(), destination(), words); 3311 } 3312 update_state(words); 3313 } 3314 3315 void InstanceKlass::oop_pc_update_pointers(oop obj) { 3316 oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3317 } 3318 3319 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) { 3320 InstanceKlass::oop_pc_update_pointers(obj); 3321 3322 oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3323 } 3324 3325 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) { 3326 InstanceKlass::oop_pc_update_pointers(obj); 3327 } 3328 3329 #ifdef ASSERT 3330 template <class T> static void trace_reference_gc(const char *s, oop obj, 3331 T* referent_addr, 3332 T* next_addr, 3333 T* discovered_addr) { 3334 if(TraceReferenceGC && PrintGCDetails) { 3335 gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj)); 3336 gclog_or_tty->print_cr(" referent_addr/* " PTR_FORMAT " / " 3337 PTR_FORMAT, p2i(referent_addr), 3338 referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3339 gclog_or_tty->print_cr(" next_addr/* " PTR_FORMAT " / " 3340 PTR_FORMAT, p2i(next_addr), 3341 next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3342 gclog_or_tty->print_cr(" discovered_addr/* " PTR_FORMAT " / " 3343 PTR_FORMAT, p2i(discovered_addr), 3344 discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3345 } 3346 } 3347 #endif 3348 3349 template <class T> 3350 static void oop_pc_update_pointers_specialized(oop obj) { 3351 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3352 PSParallelCompact::adjust_pointer(referent_addr); 3353 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3354 PSParallelCompact::adjust_pointer(next_addr); 3355 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3356 PSParallelCompact::adjust_pointer(discovered_addr); 3357 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3358 referent_addr, next_addr, discovered_addr);) 3359 } 3360 3361 void InstanceRefKlass::oop_pc_update_pointers(oop obj) { 3362 InstanceKlass::oop_pc_update_pointers(obj); 3363 3364 if (UseCompressedOops) { 3365 oop_pc_update_pointers_specialized<narrowOop>(obj); 3366 } else { 3367 oop_pc_update_pointers_specialized<oop>(obj); 3368 } 3369 } 3370 3371 void ObjArrayKlass::oop_pc_update_pointers(oop obj) { 3372 assert(obj->is_objArray(), "obj must be obj array"); 3373 oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure()); 3374 } 3375 3376 void TypeArrayKlass::oop_pc_update_pointers(oop obj) { 3377 assert(obj->is_typeArray(),"must be a type array"); 3378 } 3379 3380 ParMarkBitMapClosure::IterationStatus 3381 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3382 assert(destination() != NULL, "sanity"); 3383 assert(bitmap()->obj_size(addr) == words, "bad size"); 3384 3385 _source = addr; 3386 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3387 destination(), "wrong destination"); 3388 3389 if (words > words_remaining()) { 3390 return ParMarkBitMap::would_overflow; 3391 } 3392 3393 // The start_array must be updated even if the object is not moving. 3394 if (_start_array != NULL) { 3395 _start_array->allocate_block(destination()); 3396 } 3397 3398 if (destination() != source()) { 3399 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3400 Copy::aligned_conjoint_words(source(), destination(), words); 3401 } 3402 3403 oop moved_oop = (oop) destination(); 3404 compaction_manager()->update_contents(moved_oop); 3405 assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop))); 3406 3407 update_state(words); 3408 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3409 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3410 } 3411 3412 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3413 ParCompactionManager* cm, 3414 PSParallelCompact::SpaceId space_id) : 3415 ParMarkBitMapClosure(mbm, cm), 3416 _space_id(space_id), 3417 _start_array(PSParallelCompact::start_array(space_id)) 3418 { 3419 } 3420 3421 // Updates the references in the object to their new values. 3422 ParMarkBitMapClosure::IterationStatus 3423 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3424 do_addr(addr); 3425 return ParMarkBitMap::incomplete; 3426 }