1 /* 2 * Copyright (c) 2005, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/stringTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp" 30 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp" 31 #include "gc_implementation/parallelScavenge/pcTasks.hpp" 32 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp" 33 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp" 34 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp" 35 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp" 36 #include "gc_implementation/parallelScavenge/psOldGen.hpp" 37 #include "gc_implementation/parallelScavenge/psParallelCompact.inline.hpp" 38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 39 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 41 #include "gc_implementation/shared/gcHeapSummary.hpp" 42 #include "gc_implementation/shared/gcTimer.hpp" 43 #include "gc_implementation/shared/gcTrace.hpp" 44 #include "gc_implementation/shared/gcTraceTime.hpp" 45 #include "gc_implementation/shared/isGCActiveMark.hpp" 46 #include "gc_implementation/shared/spaceDecorator.hpp" 47 #include "gc_interface/gcCause.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/referencePolicy.hpp" 50 #include "memory/referenceProcessor.hpp" 51 #include "oops/instanceKlass.inline.hpp" 52 #include "oops/instanceMirrorKlass.inline.hpp" 53 #include "oops/methodData.hpp" 54 #include "oops/objArrayKlass.inline.hpp" 55 #include "oops/oop.inline.hpp" 56 #include "runtime/atomic.inline.hpp" 57 #include "runtime/fprofiler.hpp" 58 #include "runtime/safepoint.hpp" 59 #include "runtime/vmThread.hpp" 60 #include "services/management.hpp" 61 #include "services/memoryService.hpp" 62 #include "services/memTracker.hpp" 63 #include "utilities/events.hpp" 64 #include "utilities/stack.inline.hpp" 65 66 #include <math.h> 67 68 // All sizes are in HeapWords. 69 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 70 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 71 const size_t ParallelCompactData::RegionSizeBytes = 72 RegionSize << LogHeapWordSize; 73 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 74 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 75 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 76 77 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 78 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 79 const size_t ParallelCompactData::BlockSizeBytes = 80 BlockSize << LogHeapWordSize; 81 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 82 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 83 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 84 85 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 86 const size_t ParallelCompactData::Log2BlocksPerRegion = 87 Log2RegionSize - Log2BlockSize; 88 89 const ParallelCompactData::RegionData::region_sz_t 90 ParallelCompactData::RegionData::dc_shift = 27; 91 92 const ParallelCompactData::RegionData::region_sz_t 93 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 94 95 const ParallelCompactData::RegionData::region_sz_t 96 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 97 98 const ParallelCompactData::RegionData::region_sz_t 99 ParallelCompactData::RegionData::los_mask = ~dc_mask; 100 101 const ParallelCompactData::RegionData::region_sz_t 102 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 103 104 const ParallelCompactData::RegionData::region_sz_t 105 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 106 107 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 108 bool PSParallelCompact::_print_phases = false; 109 110 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 111 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL; 112 113 double PSParallelCompact::_dwl_mean; 114 double PSParallelCompact::_dwl_std_dev; 115 double PSParallelCompact::_dwl_first_term; 116 double PSParallelCompact::_dwl_adjustment; 117 #ifdef ASSERT 118 bool PSParallelCompact::_dwl_initialized = false; 119 #endif // #ifdef ASSERT 120 121 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 122 HeapWord* destination) 123 { 124 assert(src_region_idx != 0, "invalid src_region_idx"); 125 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 126 assert(destination != NULL, "invalid destination argument"); 127 128 _src_region_idx = src_region_idx; 129 _partial_obj_size = partial_obj_size; 130 _destination = destination; 131 132 // These fields may not be updated below, so make sure they're clear. 133 assert(_dest_region_addr == NULL, "should have been cleared"); 134 assert(_first_src_addr == NULL, "should have been cleared"); 135 136 // Determine the number of destination regions for the partial object. 137 HeapWord* const last_word = destination + partial_obj_size - 1; 138 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 139 HeapWord* const beg_region_addr = sd.region_align_down(destination); 140 HeapWord* const end_region_addr = sd.region_align_down(last_word); 141 142 if (beg_region_addr == end_region_addr) { 143 // One destination region. 144 _destination_count = 1; 145 if (end_region_addr == destination) { 146 // The destination falls on a region boundary, thus the first word of the 147 // partial object will be the first word copied to the destination region. 148 _dest_region_addr = end_region_addr; 149 _first_src_addr = sd.region_to_addr(src_region_idx); 150 } 151 } else { 152 // Two destination regions. When copied, the partial object will cross a 153 // destination region boundary, so a word somewhere within the partial 154 // object will be the first word copied to the second destination region. 155 _destination_count = 2; 156 _dest_region_addr = end_region_addr; 157 const size_t ofs = pointer_delta(end_region_addr, destination); 158 assert(ofs < _partial_obj_size, "sanity"); 159 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 160 } 161 } 162 163 void SplitInfo::clear() 164 { 165 _src_region_idx = 0; 166 _partial_obj_size = 0; 167 _destination = NULL; 168 _destination_count = 0; 169 _dest_region_addr = NULL; 170 _first_src_addr = NULL; 171 assert(!is_valid(), "sanity"); 172 } 173 174 #ifdef ASSERT 175 void SplitInfo::verify_clear() 176 { 177 assert(_src_region_idx == 0, "not clear"); 178 assert(_partial_obj_size == 0, "not clear"); 179 assert(_destination == NULL, "not clear"); 180 assert(_destination_count == 0, "not clear"); 181 assert(_dest_region_addr == NULL, "not clear"); 182 assert(_first_src_addr == NULL, "not clear"); 183 } 184 #endif // #ifdef ASSERT 185 186 187 void PSParallelCompact::print_on_error(outputStream* st) { 188 _mark_bitmap.print_on_error(st); 189 } 190 191 #ifndef PRODUCT 192 const char* PSParallelCompact::space_names[] = { 193 "old ", "eden", "from", "to " 194 }; 195 196 void PSParallelCompact::print_region_ranges() 197 { 198 tty->print_cr("space bottom top end new_top"); 199 tty->print_cr("------ ---------- ---------- ---------- ----------"); 200 201 for (unsigned int id = 0; id < last_space_id; ++id) { 202 const MutableSpace* space = _space_info[id].space(); 203 tty->print_cr("%u %s " 204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 205 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 206 id, space_names[id], 207 summary_data().addr_to_region_idx(space->bottom()), 208 summary_data().addr_to_region_idx(space->top()), 209 summary_data().addr_to_region_idx(space->end()), 210 summary_data().addr_to_region_idx(_space_info[id].new_top())); 211 } 212 } 213 214 void 215 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 216 { 217 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 218 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 219 220 ParallelCompactData& sd = PSParallelCompact::summary_data(); 221 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 222 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 223 REGION_IDX_FORMAT " " PTR_FORMAT " " 224 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 225 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 226 i, 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 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 824 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 825 826 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 827 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 828 829 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 830 831 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 832 klass->oops_do(_mark_and_push_closure); 833 } 834 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 835 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 836 } 837 838 void PSParallelCompact::post_initialize() { 839 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 840 MemRegion mr = heap->reserved_region(); 841 _ref_processor = 842 new ReferenceProcessor(mr, // span 843 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 844 (int) ParallelGCThreads, // mt processing degree 845 true, // mt discovery 846 (int) ParallelGCThreads, // mt discovery degree 847 true, // atomic_discovery 848 &_is_alive_closure); // non-header is alive closure 849 _counters = new CollectorCounters("PSParallelCompact", 1); 850 851 // Initialize static fields in ParCompactionManager. 852 ParCompactionManager::initialize(mark_bitmap()); 853 } 854 855 bool PSParallelCompact::initialize() { 856 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 857 MemRegion mr = heap->reserved_region(); 858 859 // Was the old gen get allocated successfully? 860 if (!heap->old_gen()->is_allocated()) { 861 return false; 862 } 863 864 initialize_space_info(); 865 initialize_dead_wood_limiter(); 866 867 if (!_mark_bitmap.initialize(mr)) { 868 vm_shutdown_during_initialization( 869 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 870 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 871 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 872 return false; 873 } 874 875 if (!_summary_data.initialize(mr)) { 876 vm_shutdown_during_initialization( 877 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 878 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 879 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 880 return false; 881 } 882 883 return true; 884 } 885 886 void PSParallelCompact::initialize_space_info() 887 { 888 memset(&_space_info, 0, sizeof(_space_info)); 889 890 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 891 PSYoungGen* young_gen = heap->young_gen(); 892 893 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 894 _space_info[eden_space_id].set_space(young_gen->eden_space()); 895 _space_info[from_space_id].set_space(young_gen->from_space()); 896 _space_info[to_space_id].set_space(young_gen->to_space()); 897 898 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 899 } 900 901 void PSParallelCompact::initialize_dead_wood_limiter() 902 { 903 const size_t max = 100; 904 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 905 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 906 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 907 DEBUG_ONLY(_dwl_initialized = true;) 908 _dwl_adjustment = normal_distribution(1.0); 909 } 910 911 // Simple class for storing info about the heap at the start of GC, to be used 912 // after GC for comparison/printing. 913 class PreGCValues { 914 public: 915 PreGCValues() { } 916 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 917 918 void fill(ParallelScavengeHeap* heap) { 919 _heap_used = heap->used(); 920 _young_gen_used = heap->young_gen()->used_in_bytes(); 921 _old_gen_used = heap->old_gen()->used_in_bytes(); 922 _metadata_used = MetaspaceAux::used_bytes(); 923 }; 924 925 size_t heap_used() const { return _heap_used; } 926 size_t young_gen_used() const { return _young_gen_used; } 927 size_t old_gen_used() const { return _old_gen_used; } 928 size_t metadata_used() const { return _metadata_used; } 929 930 private: 931 size_t _heap_used; 932 size_t _young_gen_used; 933 size_t _old_gen_used; 934 size_t _metadata_used; 935 }; 936 937 void 938 PSParallelCompact::clear_data_covering_space(SpaceId id) 939 { 940 // At this point, top is the value before GC, new_top() is the value that will 941 // be set at the end of GC. The marking bitmap is cleared to top; nothing 942 // should be marked above top. The summary data is cleared to the larger of 943 // top & new_top. 944 MutableSpace* const space = _space_info[id].space(); 945 HeapWord* const bot = space->bottom(); 946 HeapWord* const top = space->top(); 947 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 948 949 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 950 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 951 _mark_bitmap.clear_range(beg_bit, end_bit); 952 953 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 954 const size_t end_region = 955 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 956 _summary_data.clear_range(beg_region, end_region); 957 958 // Clear the data used to 'split' regions. 959 SplitInfo& split_info = _space_info[id].split_info(); 960 if (split_info.is_valid()) { 961 split_info.clear(); 962 } 963 DEBUG_ONLY(split_info.verify_clear();) 964 } 965 966 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 967 { 968 // Update the from & to space pointers in space_info, since they are swapped 969 // at each young gen gc. Do the update unconditionally (even though a 970 // promotion failure does not swap spaces) because an unknown number of minor 971 // collections will have swapped the spaces an unknown number of times. 972 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 973 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 974 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 975 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 976 977 pre_gc_values->fill(heap); 978 979 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 980 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 981 982 // Increment the invocation count 983 heap->increment_total_collections(true); 984 985 // We need to track unique mark sweep invocations as well. 986 _total_invocations++; 987 988 heap->print_heap_before_gc(); 989 heap->trace_heap_before_gc(&_gc_tracer); 990 991 // Fill in TLABs 992 heap->accumulate_statistics_all_tlabs(); 993 heap->ensure_parsability(true); // retire TLABs 994 995 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 996 HandleMark hm; // Discard invalid handles created during verification 997 Universe::verify(" VerifyBeforeGC:"); 998 } 999 1000 // Verify object start arrays 1001 if (VerifyObjectStartArray && 1002 VerifyBeforeGC) { 1003 heap->old_gen()->verify_object_start_array(); 1004 } 1005 1006 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1007 DEBUG_ONLY(summary_data().verify_clear();) 1008 1009 // Have worker threads release resources the next time they run a task. 1010 gc_task_manager()->release_all_resources(); 1011 } 1012 1013 void PSParallelCompact::post_compact() 1014 { 1015 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1016 1017 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1018 // Clear the marking bitmap, summary data and split info. 1019 clear_data_covering_space(SpaceId(id)); 1020 // Update top(). Must be done after clearing the bitmap and summary data. 1021 _space_info[id].publish_new_top(); 1022 } 1023 1024 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1025 MutableSpace* const from_space = _space_info[from_space_id].space(); 1026 MutableSpace* const to_space = _space_info[to_space_id].space(); 1027 1028 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1029 bool eden_empty = eden_space->is_empty(); 1030 if (!eden_empty) { 1031 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1032 heap->young_gen(), heap->old_gen()); 1033 } 1034 1035 // Update heap occupancy information which is used as input to the soft ref 1036 // clearing policy at the next gc. 1037 Universe::update_heap_info_at_gc(); 1038 1039 bool young_gen_empty = eden_empty && from_space->is_empty() && 1040 to_space->is_empty(); 1041 1042 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1043 MemRegion old_mr = heap->old_gen()->reserved(); 1044 if (young_gen_empty) { 1045 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1046 } else { 1047 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1048 } 1049 1050 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1051 ClassLoaderDataGraph::purge(); 1052 MetaspaceAux::verify_metrics(); 1053 1054 CodeCache::gc_epilogue(); 1055 JvmtiExport::gc_epilogue(); 1056 1057 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1058 1059 ref_processor()->enqueue_discovered_references(NULL); 1060 1061 if (ZapUnusedHeapArea) { 1062 heap->gen_mangle_unused_area(); 1063 } 1064 1065 // Update time of last GC 1066 reset_millis_since_last_gc(); 1067 } 1068 1069 HeapWord* 1070 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1071 bool maximum_compaction) 1072 { 1073 const size_t region_size = ParallelCompactData::RegionSize; 1074 const ParallelCompactData& sd = summary_data(); 1075 1076 const MutableSpace* const space = _space_info[id].space(); 1077 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1078 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1079 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1080 1081 // Skip full regions at the beginning of the space--they are necessarily part 1082 // of the dense prefix. 1083 size_t full_count = 0; 1084 const RegionData* cp; 1085 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1086 ++full_count; 1087 } 1088 1089 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1090 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1091 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1092 if (maximum_compaction || cp == end_cp || interval_ended) { 1093 _maximum_compaction_gc_num = total_invocations(); 1094 return sd.region_to_addr(cp); 1095 } 1096 1097 HeapWord* const new_top = _space_info[id].new_top(); 1098 const size_t space_live = pointer_delta(new_top, space->bottom()); 1099 const size_t space_used = space->used_in_words(); 1100 const size_t space_capacity = space->capacity_in_words(); 1101 1102 const double cur_density = double(space_live) / space_capacity; 1103 const double deadwood_density = 1104 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1105 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1106 1107 if (TraceParallelOldGCDensePrefix) { 1108 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1109 cur_density, deadwood_density, deadwood_goal); 1110 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1111 "space_cap=" SIZE_FORMAT, 1112 space_live, space_used, 1113 space_capacity); 1114 } 1115 1116 // XXX - Use binary search? 1117 HeapWord* dense_prefix = sd.region_to_addr(cp); 1118 const RegionData* full_cp = cp; 1119 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1120 while (cp < end_cp) { 1121 HeapWord* region_destination = cp->destination(); 1122 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1123 if (TraceParallelOldGCDensePrefix && Verbose) { 1124 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1125 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1126 sd.region(cp), p2i(region_destination), 1127 p2i(dense_prefix), cur_deadwood); 1128 } 1129 1130 if (cur_deadwood >= deadwood_goal) { 1131 // Found the region that has the correct amount of deadwood to the left. 1132 // This typically occurs after crossing a fairly sparse set of regions, so 1133 // iterate backwards over those sparse regions, looking for the region 1134 // that has the lowest density of live objects 'to the right.' 1135 size_t space_to_left = sd.region(cp) * region_size; 1136 size_t live_to_left = space_to_left - cur_deadwood; 1137 size_t space_to_right = space_capacity - space_to_left; 1138 size_t live_to_right = space_live - live_to_left; 1139 double density_to_right = double(live_to_right) / space_to_right; 1140 while (cp > full_cp) { 1141 --cp; 1142 const size_t prev_region_live_to_right = live_to_right - 1143 cp->data_size(); 1144 const size_t prev_region_space_to_right = space_to_right + region_size; 1145 double prev_region_density_to_right = 1146 double(prev_region_live_to_right) / prev_region_space_to_right; 1147 if (density_to_right <= prev_region_density_to_right) { 1148 return dense_prefix; 1149 } 1150 if (TraceParallelOldGCDensePrefix && Verbose) { 1151 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1152 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1153 prev_region_density_to_right); 1154 } 1155 dense_prefix -= region_size; 1156 live_to_right = prev_region_live_to_right; 1157 space_to_right = prev_region_space_to_right; 1158 density_to_right = prev_region_density_to_right; 1159 } 1160 return dense_prefix; 1161 } 1162 1163 dense_prefix += region_size; 1164 ++cp; 1165 } 1166 1167 return dense_prefix; 1168 } 1169 1170 #ifndef PRODUCT 1171 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1172 const SpaceId id, 1173 const bool maximum_compaction, 1174 HeapWord* const addr) 1175 { 1176 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1177 RegionData* const cp = summary_data().region(region_idx); 1178 const MutableSpace* const space = _space_info[id].space(); 1179 HeapWord* const new_top = _space_info[id].new_top(); 1180 1181 const size_t space_live = pointer_delta(new_top, space->bottom()); 1182 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1183 const size_t space_cap = space->capacity_in_words(); 1184 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1185 const size_t live_to_right = new_top - cp->destination(); 1186 const size_t dead_to_right = space->top() - addr - live_to_right; 1187 1188 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1189 "spl=" SIZE_FORMAT " " 1190 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1191 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1192 " ratio=%10.8f", 1193 algorithm, p2i(addr), region_idx, 1194 space_live, 1195 dead_to_left, dead_to_left_pct, 1196 dead_to_right, live_to_right, 1197 double(dead_to_right) / live_to_right); 1198 } 1199 #endif // #ifndef PRODUCT 1200 1201 // Return a fraction indicating how much of the generation can be treated as 1202 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1203 // based on the density of live objects in the generation to determine a limit, 1204 // which is then adjusted so the return value is min_percent when the density is 1205 // 1. 1206 // 1207 // The following table shows some return values for a different values of the 1208 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1209 // min_percent is 1. 1210 // 1211 // fraction allowed as dead wood 1212 // ----------------------------------------------------------------- 1213 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1214 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1215 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1216 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1217 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1218 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1219 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1220 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1221 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1222 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1223 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1224 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1225 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1226 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1227 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1228 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1229 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1230 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1231 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1232 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1233 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1234 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1235 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1236 1237 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1238 { 1239 assert(_dwl_initialized, "uninitialized"); 1240 1241 // The raw limit is the value of the normal distribution at x = density. 1242 const double raw_limit = normal_distribution(density); 1243 1244 // Adjust the raw limit so it becomes the minimum when the density is 1. 1245 // 1246 // First subtract the adjustment value (which is simply the precomputed value 1247 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1248 // Then add the minimum value, so the minimum is returned when the density is 1249 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1250 const double min = double(min_percent) / 100.0; 1251 const double limit = raw_limit - _dwl_adjustment + min; 1252 return MAX2(limit, 0.0); 1253 } 1254 1255 ParallelCompactData::RegionData* 1256 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1257 const RegionData* end) 1258 { 1259 const size_t region_size = ParallelCompactData::RegionSize; 1260 ParallelCompactData& sd = summary_data(); 1261 size_t left = sd.region(beg); 1262 size_t right = end > beg ? sd.region(end) - 1 : left; 1263 1264 // Binary search. 1265 while (left < right) { 1266 // Equivalent to (left + right) / 2, but does not overflow. 1267 const size_t middle = left + (right - left) / 2; 1268 RegionData* const middle_ptr = sd.region(middle); 1269 HeapWord* const dest = middle_ptr->destination(); 1270 HeapWord* const addr = sd.region_to_addr(middle); 1271 assert(dest != NULL, "sanity"); 1272 assert(dest <= addr, "must move left"); 1273 1274 if (middle > left && dest < addr) { 1275 right = middle - 1; 1276 } else if (middle < right && middle_ptr->data_size() == region_size) { 1277 left = middle + 1; 1278 } else { 1279 return middle_ptr; 1280 } 1281 } 1282 return sd.region(left); 1283 } 1284 1285 ParallelCompactData::RegionData* 1286 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1287 const RegionData* end, 1288 size_t dead_words) 1289 { 1290 ParallelCompactData& sd = summary_data(); 1291 size_t left = sd.region(beg); 1292 size_t right = end > beg ? sd.region(end) - 1 : left; 1293 1294 // Binary search. 1295 while (left < right) { 1296 // Equivalent to (left + right) / 2, but does not overflow. 1297 const size_t middle = left + (right - left) / 2; 1298 RegionData* const middle_ptr = sd.region(middle); 1299 HeapWord* const dest = middle_ptr->destination(); 1300 HeapWord* const addr = sd.region_to_addr(middle); 1301 assert(dest != NULL, "sanity"); 1302 assert(dest <= addr, "must move left"); 1303 1304 const size_t dead_to_left = pointer_delta(addr, dest); 1305 if (middle > left && dead_to_left > dead_words) { 1306 right = middle - 1; 1307 } else if (middle < right && dead_to_left < dead_words) { 1308 left = middle + 1; 1309 } else { 1310 return middle_ptr; 1311 } 1312 } 1313 return sd.region(left); 1314 } 1315 1316 // The result is valid during the summary phase, after the initial summarization 1317 // of each space into itself, and before final summarization. 1318 inline double 1319 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1320 HeapWord* const bottom, 1321 HeapWord* const top, 1322 HeapWord* const new_top) 1323 { 1324 ParallelCompactData& sd = summary_data(); 1325 1326 assert(cp != NULL, "sanity"); 1327 assert(bottom != NULL, "sanity"); 1328 assert(top != NULL, "sanity"); 1329 assert(new_top != NULL, "sanity"); 1330 assert(top >= new_top, "summary data problem?"); 1331 assert(new_top > bottom, "space is empty; should not be here"); 1332 assert(new_top >= cp->destination(), "sanity"); 1333 assert(top >= sd.region_to_addr(cp), "sanity"); 1334 1335 HeapWord* const destination = cp->destination(); 1336 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1337 const size_t compacted_region_live = pointer_delta(new_top, destination); 1338 const size_t compacted_region_used = pointer_delta(top, 1339 sd.region_to_addr(cp)); 1340 const size_t reclaimable = compacted_region_used - compacted_region_live; 1341 1342 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1343 return double(reclaimable) / divisor; 1344 } 1345 1346 // Return the address of the end of the dense prefix, a.k.a. the start of the 1347 // compacted region. The address is always on a region boundary. 1348 // 1349 // Completely full regions at the left are skipped, since no compaction can 1350 // occur in those regions. Then the maximum amount of dead wood to allow is 1351 // computed, based on the density (amount live / capacity) of the generation; 1352 // the region with approximately that amount of dead space to the left is 1353 // identified as the limit region. Regions between the last completely full 1354 // region and the limit region are scanned and the one that has the best 1355 // (maximum) reclaimed_ratio() is selected. 1356 HeapWord* 1357 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1358 bool maximum_compaction) 1359 { 1360 if (ParallelOldGCSplitALot) { 1361 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1362 // The value was chosen to provoke splitting a young gen space; use it. 1363 return _space_info[id].dense_prefix(); 1364 } 1365 } 1366 1367 const size_t region_size = ParallelCompactData::RegionSize; 1368 const ParallelCompactData& sd = summary_data(); 1369 1370 const MutableSpace* const space = _space_info[id].space(); 1371 HeapWord* const top = space->top(); 1372 HeapWord* const top_aligned_up = sd.region_align_up(top); 1373 HeapWord* const new_top = _space_info[id].new_top(); 1374 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1375 HeapWord* const bottom = space->bottom(); 1376 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1377 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1378 const RegionData* const new_top_cp = 1379 sd.addr_to_region_ptr(new_top_aligned_up); 1380 1381 // Skip full regions at the beginning of the space--they are necessarily part 1382 // of the dense prefix. 1383 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1384 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1385 space->is_empty(), "no dead space allowed to the left"); 1386 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1387 "region must have dead space"); 1388 1389 // The gc number is saved whenever a maximum compaction is done, and used to 1390 // determine when the maximum compaction interval has expired. This avoids 1391 // successive max compactions for different reasons. 1392 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1393 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1394 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1395 total_invocations() == HeapFirstMaximumCompactionCount; 1396 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1397 _maximum_compaction_gc_num = total_invocations(); 1398 return sd.region_to_addr(full_cp); 1399 } 1400 1401 const size_t space_live = pointer_delta(new_top, bottom); 1402 const size_t space_used = space->used_in_words(); 1403 const size_t space_capacity = space->capacity_in_words(); 1404 1405 const double density = double(space_live) / double(space_capacity); 1406 const size_t min_percent_free = MarkSweepDeadRatio; 1407 const double limiter = dead_wood_limiter(density, min_percent_free); 1408 const size_t dead_wood_max = space_used - space_live; 1409 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1410 dead_wood_max); 1411 1412 if (TraceParallelOldGCDensePrefix) { 1413 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1414 "space_cap=" SIZE_FORMAT, 1415 space_live, space_used, 1416 space_capacity); 1417 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1418 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1419 density, min_percent_free, limiter, 1420 dead_wood_max, dead_wood_limit); 1421 } 1422 1423 // Locate the region with the desired amount of dead space to the left. 1424 const RegionData* const limit_cp = 1425 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1426 1427 // Scan from the first region with dead space to the limit region and find the 1428 // one with the best (largest) reclaimed ratio. 1429 double best_ratio = 0.0; 1430 const RegionData* best_cp = full_cp; 1431 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1432 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1433 if (tmp_ratio > best_ratio) { 1434 best_cp = cp; 1435 best_ratio = tmp_ratio; 1436 } 1437 } 1438 1439 #if 0 1440 // Something to consider: if the region with the best ratio is 'close to' the 1441 // first region w/free space, choose the first region with free space 1442 // ("first-free"). The first-free region is usually near the start of the 1443 // heap, which means we are copying most of the heap already, so copy a bit 1444 // more to get complete compaction. 1445 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1446 _maximum_compaction_gc_num = total_invocations(); 1447 best_cp = full_cp; 1448 } 1449 #endif // #if 0 1450 1451 return sd.region_to_addr(best_cp); 1452 } 1453 1454 #ifndef PRODUCT 1455 void 1456 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1457 size_t words) 1458 { 1459 if (TraceParallelOldGCSummaryPhase) { 1460 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1461 SIZE_FORMAT, p2i(start), p2i(start + words), words); 1462 } 1463 1464 ObjectStartArray* const start_array = _space_info[id].start_array(); 1465 CollectedHeap::fill_with_objects(start, words); 1466 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1467 _mark_bitmap.mark_obj(p, words); 1468 _summary_data.add_obj(p, words); 1469 start_array->allocate_block(p); 1470 } 1471 } 1472 1473 void 1474 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1475 { 1476 ParallelCompactData& sd = summary_data(); 1477 MutableSpace* space = _space_info[id].space(); 1478 1479 // Find the source and destination start addresses. 1480 HeapWord* const src_addr = sd.region_align_down(start); 1481 HeapWord* dst_addr; 1482 if (src_addr < start) { 1483 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1484 } else if (src_addr > space->bottom()) { 1485 // The start (the original top() value) is aligned to a region boundary so 1486 // the associated region does not have a destination. Compute the 1487 // destination from the previous region. 1488 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1489 dst_addr = cp->destination() + cp->data_size(); 1490 } else { 1491 // Filling the entire space. 1492 dst_addr = space->bottom(); 1493 } 1494 assert(dst_addr != NULL, "sanity"); 1495 1496 // Update the summary data. 1497 bool result = _summary_data.summarize(_space_info[id].split_info(), 1498 src_addr, space->top(), NULL, 1499 dst_addr, space->end(), 1500 _space_info[id].new_top_addr()); 1501 assert(result, "should not fail: bad filler object size"); 1502 } 1503 1504 void 1505 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1506 { 1507 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1508 return; 1509 } 1510 1511 MutableSpace* const space = _space_info[id].space(); 1512 if (space->is_empty()) { 1513 HeapWord* b = space->bottom(); 1514 HeapWord* t = b + space->capacity_in_words() / 2; 1515 space->set_top(t); 1516 if (ZapUnusedHeapArea) { 1517 space->set_top_for_allocations(); 1518 } 1519 1520 size_t min_size = CollectedHeap::min_fill_size(); 1521 size_t obj_len = min_size; 1522 while (b + obj_len <= t) { 1523 CollectedHeap::fill_with_object(b, obj_len); 1524 mark_bitmap()->mark_obj(b, obj_len); 1525 summary_data().add_obj(b, obj_len); 1526 b += obj_len; 1527 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1528 } 1529 if (b < t) { 1530 // The loop didn't completely fill to t (top); adjust top downward. 1531 space->set_top(b); 1532 if (ZapUnusedHeapArea) { 1533 space->set_top_for_allocations(); 1534 } 1535 } 1536 1537 HeapWord** nta = _space_info[id].new_top_addr(); 1538 bool result = summary_data().summarize(_space_info[id].split_info(), 1539 space->bottom(), space->top(), NULL, 1540 space->bottom(), space->end(), nta); 1541 assert(result, "space must fit into itself"); 1542 } 1543 } 1544 1545 void 1546 PSParallelCompact::provoke_split(bool & max_compaction) 1547 { 1548 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1549 return; 1550 } 1551 1552 const size_t region_size = ParallelCompactData::RegionSize; 1553 ParallelCompactData& sd = summary_data(); 1554 1555 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1556 MutableSpace* const from_space = _space_info[from_space_id].space(); 1557 const size_t eden_live = pointer_delta(eden_space->top(), 1558 _space_info[eden_space_id].new_top()); 1559 const size_t from_live = pointer_delta(from_space->top(), 1560 _space_info[from_space_id].new_top()); 1561 1562 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1563 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1564 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1565 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1566 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1567 1568 // Choose the space to split; need at least 2 regions live (or fillable). 1569 SpaceId id; 1570 MutableSpace* space; 1571 size_t live_words; 1572 size_t fill_words; 1573 if (eden_live + eden_fillable >= region_size * 2) { 1574 id = eden_space_id; 1575 space = eden_space; 1576 live_words = eden_live; 1577 fill_words = eden_fillable; 1578 } else if (from_live + from_fillable >= region_size * 2) { 1579 id = from_space_id; 1580 space = from_space; 1581 live_words = from_live; 1582 fill_words = from_fillable; 1583 } else { 1584 return; // Give up. 1585 } 1586 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1587 1588 if (live_words < region_size * 2) { 1589 // Fill from top() to end() w/live objects of mixed sizes. 1590 HeapWord* const fill_start = space->top(); 1591 live_words += fill_words; 1592 1593 space->set_top(fill_start + fill_words); 1594 if (ZapUnusedHeapArea) { 1595 space->set_top_for_allocations(); 1596 } 1597 1598 HeapWord* cur_addr = fill_start; 1599 while (fill_words > 0) { 1600 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1601 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1602 if (fill_words - cur_size < min_fill_size) { 1603 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1604 } 1605 1606 CollectedHeap::fill_with_object(cur_addr, cur_size); 1607 mark_bitmap()->mark_obj(cur_addr, cur_size); 1608 sd.add_obj(cur_addr, cur_size); 1609 1610 cur_addr += cur_size; 1611 fill_words -= cur_size; 1612 } 1613 1614 summarize_new_objects(id, fill_start); 1615 } 1616 1617 max_compaction = false; 1618 1619 // Manipulate the old gen so that it has room for about half of the live data 1620 // in the target young gen space (live_words / 2). 1621 id = old_space_id; 1622 space = _space_info[id].space(); 1623 const size_t free_at_end = space->free_in_words(); 1624 const size_t free_target = align_object_size(live_words / 2); 1625 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1626 1627 if (free_at_end >= free_target + min_fill_size) { 1628 // Fill space above top() and set the dense prefix so everything survives. 1629 HeapWord* const fill_start = space->top(); 1630 const size_t fill_size = free_at_end - free_target; 1631 space->set_top(space->top() + fill_size); 1632 if (ZapUnusedHeapArea) { 1633 space->set_top_for_allocations(); 1634 } 1635 fill_with_live_objects(id, fill_start, fill_size); 1636 summarize_new_objects(id, fill_start); 1637 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1638 } else if (dead + free_at_end > free_target) { 1639 // Find a dense prefix that makes the right amount of space available. 1640 HeapWord* cur = sd.region_align_down(space->top()); 1641 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1642 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1643 while (dead_to_right < free_target) { 1644 cur -= region_size; 1645 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1646 dead_to_right = pointer_delta(space->end(), cur_destination); 1647 } 1648 _space_info[id].set_dense_prefix(cur); 1649 } 1650 } 1651 #endif // #ifndef PRODUCT 1652 1653 void PSParallelCompact::summarize_spaces_quick() 1654 { 1655 for (unsigned int i = 0; i < last_space_id; ++i) { 1656 const MutableSpace* space = _space_info[i].space(); 1657 HeapWord** nta = _space_info[i].new_top_addr(); 1658 bool result = _summary_data.summarize(_space_info[i].split_info(), 1659 space->bottom(), space->top(), NULL, 1660 space->bottom(), space->end(), nta); 1661 assert(result, "space must fit into itself"); 1662 _space_info[i].set_dense_prefix(space->bottom()); 1663 } 1664 1665 #ifndef PRODUCT 1666 if (ParallelOldGCSplitALot) { 1667 provoke_split_fill_survivor(to_space_id); 1668 } 1669 #endif // #ifndef PRODUCT 1670 } 1671 1672 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1673 { 1674 HeapWord* const dense_prefix_end = dense_prefix(id); 1675 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1676 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1677 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1678 // Only enough dead space is filled so that any remaining dead space to the 1679 // left is larger than the minimum filler object. (The remainder is filled 1680 // during the copy/update phase.) 1681 // 1682 // The size of the dead space to the right of the boundary is not a 1683 // concern, since compaction will be able to use whatever space is 1684 // available. 1685 // 1686 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1687 // surrounds the space to be filled with an object. 1688 // 1689 // In the 32-bit VM, each bit represents two 32-bit words: 1690 // +---+ 1691 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1692 // end_bits: ... x x x | 0 | || 0 x x ... 1693 // +---+ 1694 // 1695 // In the 64-bit VM, each bit represents one 64-bit word: 1696 // +------------+ 1697 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1698 // end_bits: ... x x 1 | 0 || 0 | x x ... 1699 // +------------+ 1700 // +-------+ 1701 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1702 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1703 // +-------+ 1704 // +-----------+ 1705 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1706 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1707 // +-----------+ 1708 // +-------+ 1709 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1710 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1711 // +-------+ 1712 1713 // Initially assume case a, c or e will apply. 1714 size_t obj_len = CollectedHeap::min_fill_size(); 1715 HeapWord* obj_beg = dense_prefix_end - obj_len; 1716 1717 #ifdef _LP64 1718 if (MinObjAlignment > 1) { // object alignment > heap word size 1719 // Cases a, c or e. 1720 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1721 // Case b above. 1722 obj_beg = dense_prefix_end - 1; 1723 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1724 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1725 // Case d above. 1726 obj_beg = dense_prefix_end - 3; 1727 obj_len = 3; 1728 } 1729 #endif // #ifdef _LP64 1730 1731 CollectedHeap::fill_with_object(obj_beg, obj_len); 1732 _mark_bitmap.mark_obj(obj_beg, obj_len); 1733 _summary_data.add_obj(obj_beg, obj_len); 1734 assert(start_array(id) != NULL, "sanity"); 1735 start_array(id)->allocate_block(obj_beg); 1736 } 1737 } 1738 1739 void 1740 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1741 { 1742 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1743 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1744 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1745 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1746 cur->set_source_region(0); 1747 } 1748 } 1749 1750 void 1751 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1752 { 1753 assert(id < last_space_id, "id out of range"); 1754 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1755 ParallelOldGCSplitALot && id == old_space_id, 1756 "should have been reset in summarize_spaces_quick()"); 1757 1758 const MutableSpace* space = _space_info[id].space(); 1759 if (_space_info[id].new_top() != space->bottom()) { 1760 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1761 _space_info[id].set_dense_prefix(dense_prefix_end); 1762 1763 #ifndef PRODUCT 1764 if (TraceParallelOldGCDensePrefix) { 1765 print_dense_prefix_stats("ratio", id, maximum_compaction, 1766 dense_prefix_end); 1767 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1768 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1769 } 1770 #endif // #ifndef PRODUCT 1771 1772 // Recompute the summary data, taking into account the dense prefix. If 1773 // every last byte will be reclaimed, then the existing summary data which 1774 // compacts everything can be left in place. 1775 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1776 // If dead space crosses the dense prefix boundary, it is (at least 1777 // partially) filled with a dummy object, marked live and added to the 1778 // summary data. This simplifies the copy/update phase and must be done 1779 // before the final locations of objects are determined, to prevent 1780 // leaving a fragment of dead space that is too small to fill. 1781 fill_dense_prefix_end(id); 1782 1783 // Compute the destination of each Region, and thus each object. 1784 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1785 _summary_data.summarize(_space_info[id].split_info(), 1786 dense_prefix_end, space->top(), NULL, 1787 dense_prefix_end, space->end(), 1788 _space_info[id].new_top_addr()); 1789 } 1790 } 1791 1792 if (TraceParallelOldGCSummaryPhase) { 1793 const size_t region_size = ParallelCompactData::RegionSize; 1794 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1795 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1796 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1797 HeapWord* const new_top = _space_info[id].new_top(); 1798 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1799 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1800 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1801 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1802 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1803 id, space->capacity_in_words(), p2i(dense_prefix_end), 1804 dp_region, dp_words / region_size, 1805 cr_words / region_size, p2i(new_top)); 1806 } 1807 } 1808 1809 #ifndef PRODUCT 1810 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1811 HeapWord* dst_beg, HeapWord* dst_end, 1812 SpaceId src_space_id, 1813 HeapWord* src_beg, HeapWord* src_end) 1814 { 1815 if (TraceParallelOldGCSummaryPhase) { 1816 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1817 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1818 SIZE_FORMAT "-" SIZE_FORMAT " " 1819 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1820 SIZE_FORMAT "-" SIZE_FORMAT, 1821 src_space_id, space_names[src_space_id], 1822 dst_space_id, space_names[dst_space_id], 1823 p2i(src_beg), p2i(src_end), 1824 _summary_data.addr_to_region_idx(src_beg), 1825 _summary_data.addr_to_region_idx(src_end), 1826 p2i(dst_beg), p2i(dst_end), 1827 _summary_data.addr_to_region_idx(dst_beg), 1828 _summary_data.addr_to_region_idx(dst_end)); 1829 } 1830 } 1831 #endif // #ifndef PRODUCT 1832 1833 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1834 bool maximum_compaction) 1835 { 1836 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1837 // trace("2"); 1838 1839 #ifdef ASSERT 1840 if (TraceParallelOldGCMarkingPhase) { 1841 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1842 "add_obj_bytes=" SIZE_FORMAT, 1843 add_obj_count, add_obj_size * HeapWordSize); 1844 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1845 "mark_bitmap_bytes=" SIZE_FORMAT, 1846 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1847 } 1848 #endif // #ifdef ASSERT 1849 1850 // Quick summarization of each space into itself, to see how much is live. 1851 summarize_spaces_quick(); 1852 1853 if (TraceParallelOldGCSummaryPhase) { 1854 tty->print_cr("summary_phase: after summarizing each space to self"); 1855 Universe::print(); 1856 NOT_PRODUCT(print_region_ranges()); 1857 if (Verbose) { 1858 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1859 } 1860 } 1861 1862 // The amount of live data that will end up in old space (assuming it fits). 1863 size_t old_space_total_live = 0; 1864 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1865 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1866 _space_info[id].space()->bottom()); 1867 } 1868 1869 MutableSpace* const old_space = _space_info[old_space_id].space(); 1870 const size_t old_capacity = old_space->capacity_in_words(); 1871 if (old_space_total_live > old_capacity) { 1872 // XXX - should also try to expand 1873 maximum_compaction = true; 1874 } 1875 #ifndef PRODUCT 1876 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1877 provoke_split(maximum_compaction); 1878 } 1879 #endif // #ifndef PRODUCT 1880 1881 // Old generations. 1882 summarize_space(old_space_id, maximum_compaction); 1883 1884 // Summarize the remaining spaces in the young gen. The initial target space 1885 // is the old gen. If a space does not fit entirely into the target, then the 1886 // remainder is compacted into the space itself and that space becomes the new 1887 // target. 1888 SpaceId dst_space_id = old_space_id; 1889 HeapWord* dst_space_end = old_space->end(); 1890 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1891 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1892 const MutableSpace* space = _space_info[id].space(); 1893 const size_t live = pointer_delta(_space_info[id].new_top(), 1894 space->bottom()); 1895 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1896 1897 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1898 SpaceId(id), space->bottom(), space->top());) 1899 if (live > 0 && live <= available) { 1900 // All the live data will fit. 1901 bool done = _summary_data.summarize(_space_info[id].split_info(), 1902 space->bottom(), space->top(), 1903 NULL, 1904 *new_top_addr, dst_space_end, 1905 new_top_addr); 1906 assert(done, "space must fit into old gen"); 1907 1908 // Reset the new_top value for the space. 1909 _space_info[id].set_new_top(space->bottom()); 1910 } else if (live > 0) { 1911 // Attempt to fit part of the source space into the target space. 1912 HeapWord* next_src_addr = NULL; 1913 bool done = _summary_data.summarize(_space_info[id].split_info(), 1914 space->bottom(), space->top(), 1915 &next_src_addr, 1916 *new_top_addr, dst_space_end, 1917 new_top_addr); 1918 assert(!done, "space should not fit into old gen"); 1919 assert(next_src_addr != NULL, "sanity"); 1920 1921 // The source space becomes the new target, so the remainder is compacted 1922 // within the space itself. 1923 dst_space_id = SpaceId(id); 1924 dst_space_end = space->end(); 1925 new_top_addr = _space_info[id].new_top_addr(); 1926 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1927 space->bottom(), dst_space_end, 1928 SpaceId(id), next_src_addr, space->top());) 1929 done = _summary_data.summarize(_space_info[id].split_info(), 1930 next_src_addr, space->top(), 1931 NULL, 1932 space->bottom(), dst_space_end, 1933 new_top_addr); 1934 assert(done, "space must fit when compacted into itself"); 1935 assert(*new_top_addr <= space->top(), "usage should not grow"); 1936 } 1937 } 1938 1939 if (TraceParallelOldGCSummaryPhase) { 1940 tty->print_cr("summary_phase: after final summarization"); 1941 Universe::print(); 1942 NOT_PRODUCT(print_region_ranges()); 1943 if (Verbose) { 1944 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1945 } 1946 } 1947 } 1948 1949 // This method should contain all heap-specific policy for invoking a full 1950 // collection. invoke_no_policy() will only attempt to compact the heap; it 1951 // will do nothing further. If we need to bail out for policy reasons, scavenge 1952 // before full gc, or any other specialized behavior, it needs to be added here. 1953 // 1954 // Note that this method should only be called from the vm_thread while at a 1955 // safepoint. 1956 // 1957 // Note that the all_soft_refs_clear flag in the collector policy 1958 // may be true because this method can be called without intervening 1959 // activity. For example when the heap space is tight and full measure 1960 // are being taken to free space. 1961 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1962 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1963 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1964 "should be in vm thread"); 1965 1966 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1967 GCCause::Cause gc_cause = heap->gc_cause(); 1968 assert(!heap->is_gc_active(), "not reentrant"); 1969 1970 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1971 IsGCActiveMark mark; 1972 1973 if (ScavengeBeforeFullGC) { 1974 PSScavenge::invoke_no_policy(); 1975 } 1976 1977 const bool clear_all_soft_refs = 1978 heap->collector_policy()->should_clear_all_soft_refs(); 1979 1980 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1981 maximum_heap_compaction); 1982 } 1983 1984 // This method contains no policy. You should probably 1985 // be calling invoke() instead. 1986 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1987 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1988 assert(ref_processor() != NULL, "Sanity"); 1989 1990 if (GC_locker::check_active_before_gc()) { 1991 return false; 1992 } 1993 1994 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1995 1996 _gc_timer.register_gc_start(); 1997 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1998 1999 TimeStamp marking_start; 2000 TimeStamp compaction_start; 2001 TimeStamp collection_exit; 2002 2003 GCCause::Cause gc_cause = heap->gc_cause(); 2004 PSYoungGen* young_gen = heap->young_gen(); 2005 PSOldGen* old_gen = heap->old_gen(); 2006 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2007 2008 // The scope of casr should end after code that can change 2009 // CollectorPolicy::_should_clear_all_soft_refs. 2010 ClearedAllSoftRefs casr(maximum_heap_compaction, 2011 heap->collector_policy()); 2012 2013 if (ZapUnusedHeapArea) { 2014 // Save information needed to minimize mangling 2015 heap->record_gen_tops_before_GC(); 2016 } 2017 2018 heap->pre_full_gc_dump(&_gc_timer); 2019 2020 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2021 2022 // Make sure data structures are sane, make the heap parsable, and do other 2023 // miscellaneous bookkeeping. 2024 PreGCValues pre_gc_values; 2025 pre_compact(&pre_gc_values); 2026 2027 // Get the compaction manager reserved for the VM thread. 2028 ParCompactionManager* const vmthread_cm = 2029 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2030 2031 // Place after pre_compact() where the number of invocations is incremented. 2032 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2033 2034 { 2035 ResourceMark rm; 2036 HandleMark hm; 2037 2038 // Set the number of GC threads to be used in this collection 2039 gc_task_manager()->set_active_gang(); 2040 gc_task_manager()->task_idle_workers(); 2041 heap->set_par_threads(gc_task_manager()->active_workers()); 2042 2043 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2044 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id()); 2045 TraceCollectorStats tcs(counters()); 2046 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2047 2048 if (TraceOldGenTime) accumulated_time()->start(); 2049 2050 // Let the size policy know we're starting 2051 size_policy->major_collection_begin(); 2052 2053 CodeCache::gc_prologue(); 2054 2055 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2056 2057 ref_processor()->enable_discovery(); 2058 ref_processor()->setup_policy(maximum_heap_compaction); 2059 2060 bool marked_for_unloading = false; 2061 2062 marking_start.update(); 2063 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2064 2065 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2066 && gc_cause == GCCause::_java_lang_system_gc; 2067 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2068 2069 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2070 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2071 2072 // adjust_roots() updates Universe::_intArrayKlassObj which is 2073 // needed by the compaction for filling holes in the dense prefix. 2074 adjust_roots(); 2075 2076 compaction_start.update(); 2077 compact(); 2078 2079 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2080 // done before resizing. 2081 post_compact(); 2082 2083 // Let the size policy know we're done 2084 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2085 2086 if (UseAdaptiveSizePolicy) { 2087 if (PrintAdaptiveSizePolicy) { 2088 gclog_or_tty->print("AdaptiveSizeStart: "); 2089 gclog_or_tty->stamp(); 2090 gclog_or_tty->print_cr(" collection: %d ", 2091 heap->total_collections()); 2092 if (Verbose) { 2093 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT 2094 " young_gen_capacity: " SIZE_FORMAT, 2095 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2096 } 2097 } 2098 2099 // Don't check if the size_policy is ready here. Let 2100 // the size_policy check that internally. 2101 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2102 ((gc_cause != GCCause::_java_lang_system_gc) || 2103 UseAdaptiveSizePolicyWithSystemGC)) { 2104 // Swap the survivor spaces if from_space is empty. The 2105 // resize_young_gen() called below is normally used after 2106 // a successful young GC and swapping of survivor spaces; 2107 // otherwise, it will fail to resize the young gen with 2108 // the current implementation. 2109 if (young_gen->from_space()->is_empty()) { 2110 young_gen->from_space()->clear(SpaceDecorator::Mangle); 2111 young_gen->swap_spaces(); 2112 } 2113 2114 // Calculate optimal free space amounts 2115 assert(young_gen->max_size() > 2116 young_gen->from_space()->capacity_in_bytes() + 2117 young_gen->to_space()->capacity_in_bytes(), 2118 "Sizes of space in young gen are out-of-bounds"); 2119 2120 size_t young_live = young_gen->used_in_bytes(); 2121 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2122 size_t old_live = old_gen->used_in_bytes(); 2123 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2124 size_t max_old_gen_size = old_gen->max_gen_size(); 2125 size_t max_eden_size = young_gen->max_size() - 2126 young_gen->from_space()->capacity_in_bytes() - 2127 young_gen->to_space()->capacity_in_bytes(); 2128 2129 // Used for diagnostics 2130 size_policy->clear_generation_free_space_flags(); 2131 2132 size_policy->compute_generations_free_space(young_live, 2133 eden_live, 2134 old_live, 2135 cur_eden, 2136 max_old_gen_size, 2137 max_eden_size, 2138 true /* full gc*/); 2139 2140 size_policy->check_gc_overhead_limit(young_live, 2141 eden_live, 2142 max_old_gen_size, 2143 max_eden_size, 2144 true /* full gc*/, 2145 gc_cause, 2146 heap->collector_policy()); 2147 2148 size_policy->decay_supplemental_growth(true /* full gc*/); 2149 2150 heap->resize_old_gen( 2151 size_policy->calculated_old_free_size_in_bytes()); 2152 2153 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 2154 size_policy->calculated_survivor_size_in_bytes()); 2155 } 2156 if (PrintAdaptiveSizePolicy) { 2157 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2158 heap->total_collections()); 2159 } 2160 } 2161 2162 if (UsePerfData) { 2163 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2164 counters->update_counters(); 2165 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2166 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2167 } 2168 2169 heap->resize_all_tlabs(); 2170 2171 // Resize the metaspace capacity after a collection 2172 MetaspaceGC::compute_new_size(); 2173 2174 if (TraceOldGenTime) accumulated_time()->stop(); 2175 2176 if (PrintGC) { 2177 if (PrintGCDetails) { 2178 // No GC timestamp here. This is after GC so it would be confusing. 2179 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2180 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2181 heap->print_heap_change(pre_gc_values.heap_used()); 2182 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2183 } else { 2184 heap->print_heap_change(pre_gc_values.heap_used()); 2185 } 2186 } 2187 2188 // Track memory usage and detect low memory 2189 MemoryService::track_memory_usage(); 2190 heap->update_counters(); 2191 gc_task_manager()->release_idle_workers(); 2192 } 2193 2194 #ifdef ASSERT 2195 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2196 ParCompactionManager* const cm = 2197 ParCompactionManager::manager_array(int(i)); 2198 assert(cm->marking_stack()->is_empty(), "should be empty"); 2199 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2200 } 2201 #endif // ASSERT 2202 2203 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2204 HandleMark hm; // Discard invalid handles created during verification 2205 Universe::verify(" VerifyAfterGC:"); 2206 } 2207 2208 // Re-verify object start arrays 2209 if (VerifyObjectStartArray && 2210 VerifyAfterGC) { 2211 old_gen->verify_object_start_array(); 2212 } 2213 2214 if (ZapUnusedHeapArea) { 2215 old_gen->object_space()->check_mangled_unused_area_complete(); 2216 } 2217 2218 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2219 2220 collection_exit.update(); 2221 2222 heap->print_heap_after_gc(); 2223 heap->trace_heap_after_gc(&_gc_tracer); 2224 2225 if (PrintGCTaskTimeStamps) { 2226 gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " 2227 JLONG_FORMAT, 2228 marking_start.ticks(), compaction_start.ticks(), 2229 collection_exit.ticks()); 2230 gc_task_manager()->print_task_time_stamps(); 2231 } 2232 2233 heap->post_full_gc_dump(&_gc_timer); 2234 2235 #ifdef TRACESPINNING 2236 ParallelTaskTerminator::print_termination_counts(); 2237 #endif 2238 2239 _gc_timer.register_gc_end(); 2240 2241 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2242 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2243 2244 return true; 2245 } 2246 2247 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2248 PSYoungGen* young_gen, 2249 PSOldGen* old_gen) { 2250 MutableSpace* const eden_space = young_gen->eden_space(); 2251 assert(!eden_space->is_empty(), "eden must be non-empty"); 2252 assert(young_gen->virtual_space()->alignment() == 2253 old_gen->virtual_space()->alignment(), "alignments do not match"); 2254 2255 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2256 return false; 2257 } 2258 2259 // Both generations must be completely committed. 2260 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2261 return false; 2262 } 2263 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2264 return false; 2265 } 2266 2267 // Figure out how much to take from eden. Include the average amount promoted 2268 // in the total; otherwise the next young gen GC will simply bail out to a 2269 // full GC. 2270 const size_t alignment = old_gen->virtual_space()->alignment(); 2271 const size_t eden_used = eden_space->used_in_bytes(); 2272 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2273 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2274 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2275 2276 if (absorb_size >= eden_capacity) { 2277 return false; // Must leave some space in eden. 2278 } 2279 2280 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2281 if (new_young_size < young_gen->min_gen_size()) { 2282 return false; // Respect young gen minimum size. 2283 } 2284 2285 if (TraceAdaptiveGCBoundary && Verbose) { 2286 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2287 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2288 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2289 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2290 absorb_size / K, 2291 eden_capacity / K, (eden_capacity - absorb_size) / K, 2292 young_gen->from_space()->used_in_bytes() / K, 2293 young_gen->to_space()->used_in_bytes() / K, 2294 young_gen->capacity_in_bytes() / K, new_young_size / K); 2295 } 2296 2297 // Fill the unused part of the old gen. 2298 MutableSpace* const old_space = old_gen->object_space(); 2299 HeapWord* const unused_start = old_space->top(); 2300 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2301 2302 if (unused_words > 0) { 2303 if (unused_words < CollectedHeap::min_fill_size()) { 2304 return false; // If the old gen cannot be filled, must give up. 2305 } 2306 CollectedHeap::fill_with_objects(unused_start, unused_words); 2307 } 2308 2309 // Take the live data from eden and set both top and end in the old gen to 2310 // eden top. (Need to set end because reset_after_change() mangles the region 2311 // from end to virtual_space->high() in debug builds). 2312 HeapWord* const new_top = eden_space->top(); 2313 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2314 absorb_size); 2315 young_gen->reset_after_change(); 2316 old_space->set_top(new_top); 2317 old_space->set_end(new_top); 2318 old_gen->reset_after_change(); 2319 2320 // Update the object start array for the filler object and the data from eden. 2321 ObjectStartArray* const start_array = old_gen->start_array(); 2322 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2323 start_array->allocate_block(p); 2324 } 2325 2326 // Could update the promoted average here, but it is not typically updated at 2327 // full GCs and the value to use is unclear. Something like 2328 // 2329 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2330 2331 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2332 return true; 2333 } 2334 2335 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2336 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2337 "shouldn't return NULL"); 2338 return ParallelScavengeHeap::gc_task_manager(); 2339 } 2340 2341 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2342 bool maximum_heap_compaction, 2343 ParallelOldTracer *gc_tracer) { 2344 // Recursively traverse all live objects and mark them 2345 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2346 2347 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2348 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2349 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2350 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2351 ParallelTaskTerminator terminator(active_gc_threads, qset); 2352 2353 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2354 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2355 2356 // Need new claim bits before marking starts. 2357 ClassLoaderDataGraph::clear_claimed_marks(); 2358 2359 { 2360 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2361 2362 ParallelScavengeHeap::ParStrongRootsScope psrs; 2363 2364 GCTaskQueue* q = GCTaskQueue::create(); 2365 2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2368 // We scan the thread roots in parallel 2369 Threads::create_thread_roots_marking_tasks(q); 2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2371 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2372 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2373 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2374 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2375 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2376 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2377 2378 if (active_gc_threads > 1) { 2379 for (uint j = 0; j < active_gc_threads; j++) { 2380 q->enqueue(new StealMarkingTask(&terminator)); 2381 } 2382 } 2383 2384 gc_task_manager()->execute_and_wait(q); 2385 } 2386 2387 // Process reference objects found during marking 2388 { 2389 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2390 2391 ReferenceProcessorStats stats; 2392 if (ref_processor()->processing_is_mt()) { 2393 RefProcTaskExecutor task_executor; 2394 stats = ref_processor()->process_discovered_references( 2395 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2396 &task_executor, &_gc_timer, _gc_tracer.gc_id()); 2397 } else { 2398 stats = ref_processor()->process_discovered_references( 2399 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2400 &_gc_timer, _gc_tracer.gc_id()); 2401 } 2402 2403 gc_tracer->report_gc_reference_stats(stats); 2404 } 2405 2406 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2407 2408 // This is the point where the entire marking should have completed. 2409 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2410 2411 // Follow system dictionary roots and unload classes. 2412 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2413 2414 // Unload nmethods. 2415 CodeCache::do_unloading(is_alive_closure(), purged_class); 2416 2417 // Prune dead klasses from subklass/sibling/implementor lists. 2418 Klass::clean_weak_klass_links(is_alive_closure()); 2419 2420 // Delete entries for dead interned strings. 2421 StringTable::unlink(is_alive_closure()); 2422 2423 // Clean up unreferenced symbols in symbol table. 2424 SymbolTable::unlink(); 2425 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2426 } 2427 2428 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2429 ClassLoaderData* cld) { 2430 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2431 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2432 2433 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2434 } 2435 2436 // This should be moved to the shared markSweep code! 2437 class PSAlwaysTrueClosure: public BoolObjectClosure { 2438 public: 2439 bool do_object_b(oop p) { return true; } 2440 }; 2441 static PSAlwaysTrueClosure always_true; 2442 2443 void PSParallelCompact::adjust_roots() { 2444 // Adjust the pointers to reflect the new locations 2445 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2446 2447 // Need new claim bits when tracing through and adjusting pointers. 2448 ClassLoaderDataGraph::clear_claimed_marks(); 2449 2450 // General strong roots. 2451 Universe::oops_do(adjust_pointer_closure()); 2452 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2453 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2454 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2455 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2456 FlatProfiler::oops_do(adjust_pointer_closure()); 2457 Management::oops_do(adjust_pointer_closure()); 2458 JvmtiExport::oops_do(adjust_pointer_closure()); 2459 SystemDictionary::oops_do(adjust_pointer_closure()); 2460 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2461 2462 // Now adjust pointers in remaining weak roots. (All of which should 2463 // have been cleared if they pointed to non-surviving objects.) 2464 // Global (weak) JNI handles 2465 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2466 2467 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2468 CodeCache::blobs_do(&adjust_from_blobs); 2469 StringTable::oops_do(adjust_pointer_closure()); 2470 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2471 // Roots were visited so references into the young gen in roots 2472 // may have been scanned. Process them also. 2473 // Should the reference processor have a span that excludes 2474 // young gen objects? 2475 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2476 } 2477 2478 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2479 uint parallel_gc_threads) 2480 { 2481 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2482 2483 // Find the threads that are active 2484 unsigned int which = 0; 2485 2486 const uint task_count = MAX2(parallel_gc_threads, 1U); 2487 for (uint j = 0; j < task_count; j++) { 2488 q->enqueue(new DrainStacksCompactionTask(j)); 2489 ParCompactionManager::verify_region_list_empty(j); 2490 // Set the region stacks variables to "no" region stack values 2491 // so that they will be recognized and needing a region stack 2492 // in the stealing tasks if they do not get one by executing 2493 // a draining stack. 2494 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2495 cm->set_region_stack(NULL); 2496 cm->set_region_stack_index((uint)max_uintx); 2497 } 2498 ParCompactionManager::reset_recycled_stack_index(); 2499 2500 // Find all regions that are available (can be filled immediately) and 2501 // distribute them to the thread stacks. The iteration is done in reverse 2502 // order (high to low) so the regions will be removed in ascending order. 2503 2504 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2505 2506 size_t fillable_regions = 0; // A count for diagnostic purposes. 2507 // A region index which corresponds to the tasks created above. 2508 // "which" must be 0 <= which < task_count 2509 2510 which = 0; 2511 // id + 1 is used to test termination so unsigned can 2512 // be used with an old_space_id == 0. 2513 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2514 SpaceInfo* const space_info = _space_info + id; 2515 MutableSpace* const space = space_info->space(); 2516 HeapWord* const new_top = space_info->new_top(); 2517 2518 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2519 const size_t end_region = 2520 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2521 2522 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2523 if (sd.region(cur)->claim_unsafe()) { 2524 ParCompactionManager::region_list_push(which, cur); 2525 2526 if (TraceParallelOldGCCompactionPhase && Verbose) { 2527 const size_t count_mod_8 = fillable_regions & 7; 2528 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2529 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2530 if (count_mod_8 == 7) gclog_or_tty->cr(); 2531 } 2532 2533 NOT_PRODUCT(++fillable_regions;) 2534 2535 // Assign regions to tasks in round-robin fashion. 2536 if (++which == task_count) { 2537 assert(which <= parallel_gc_threads, 2538 "Inconsistent number of workers"); 2539 which = 0; 2540 } 2541 } 2542 } 2543 } 2544 2545 if (TraceParallelOldGCCompactionPhase) { 2546 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2547 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions); 2548 } 2549 } 2550 2551 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2552 2553 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2554 uint parallel_gc_threads) { 2555 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2556 2557 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2558 2559 // Iterate over all the spaces adding tasks for updating 2560 // regions in the dense prefix. Assume that 1 gc thread 2561 // will work on opening the gaps and the remaining gc threads 2562 // will work on the dense prefix. 2563 unsigned int space_id; 2564 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2565 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2566 const MutableSpace* const space = _space_info[space_id].space(); 2567 2568 if (dense_prefix_end == space->bottom()) { 2569 // There is no dense prefix for this space. 2570 continue; 2571 } 2572 2573 // The dense prefix is before this region. 2574 size_t region_index_end_dense_prefix = 2575 sd.addr_to_region_idx(dense_prefix_end); 2576 RegionData* const dense_prefix_cp = 2577 sd.region(region_index_end_dense_prefix); 2578 assert(dense_prefix_end == space->end() || 2579 dense_prefix_cp->available() || 2580 dense_prefix_cp->claimed(), 2581 "The region after the dense prefix should always be ready to fill"); 2582 2583 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2584 2585 // Is there dense prefix work? 2586 size_t total_dense_prefix_regions = 2587 region_index_end_dense_prefix - region_index_start; 2588 // How many regions of the dense prefix should be given to 2589 // each thread? 2590 if (total_dense_prefix_regions > 0) { 2591 uint tasks_for_dense_prefix = 1; 2592 if (total_dense_prefix_regions <= 2593 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2594 // Don't over partition. This assumes that 2595 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2596 // so there are not many regions to process. 2597 tasks_for_dense_prefix = parallel_gc_threads; 2598 } else { 2599 // Over partition 2600 tasks_for_dense_prefix = parallel_gc_threads * 2601 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2602 } 2603 size_t regions_per_thread = total_dense_prefix_regions / 2604 tasks_for_dense_prefix; 2605 // Give each thread at least 1 region. 2606 if (regions_per_thread == 0) { 2607 regions_per_thread = 1; 2608 } 2609 2610 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2611 if (region_index_start >= region_index_end_dense_prefix) { 2612 break; 2613 } 2614 // region_index_end is not processed 2615 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2616 region_index_end_dense_prefix); 2617 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2618 region_index_start, 2619 region_index_end)); 2620 region_index_start = region_index_end; 2621 } 2622 } 2623 // This gets any part of the dense prefix that did not 2624 // fit evenly. 2625 if (region_index_start < region_index_end_dense_prefix) { 2626 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2627 region_index_start, 2628 region_index_end_dense_prefix)); 2629 } 2630 } 2631 } 2632 2633 void PSParallelCompact::enqueue_region_stealing_tasks( 2634 GCTaskQueue* q, 2635 ParallelTaskTerminator* terminator_ptr, 2636 uint parallel_gc_threads) { 2637 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2638 2639 // Once a thread has drained it's stack, it should try to steal regions from 2640 // other threads. 2641 if (parallel_gc_threads > 1) { 2642 for (uint j = 0; j < parallel_gc_threads; j++) { 2643 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2644 } 2645 } 2646 } 2647 2648 #ifdef ASSERT 2649 // Write a histogram of the number of times the block table was filled for a 2650 // region. 2651 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2652 { 2653 if (!TraceParallelOldGCCompactionPhase) return; 2654 2655 typedef ParallelCompactData::RegionData rd_t; 2656 ParallelCompactData& sd = summary_data(); 2657 2658 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2659 MutableSpace* const spc = _space_info[id].space(); 2660 if (spc->bottom() != spc->top()) { 2661 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2662 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2663 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2664 2665 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2666 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2667 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2668 2669 for (const rd_t* cur = beg; cur < end; ++cur) { 2670 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2671 } 2672 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2673 for (size_t i = 0; i < histo_len; ++i) { 2674 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2675 histo[i], 100.0 * histo[i] / region_cnt); 2676 } 2677 out->cr(); 2678 } 2679 } 2680 } 2681 #endif // #ifdef ASSERT 2682 2683 void PSParallelCompact::compact() { 2684 // trace("5"); 2685 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2686 2687 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2688 PSOldGen* old_gen = heap->old_gen(); 2689 old_gen->start_array()->reset(); 2690 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2691 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2692 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2693 ParallelTaskTerminator terminator(active_gc_threads, qset); 2694 2695 GCTaskQueue* q = GCTaskQueue::create(); 2696 enqueue_region_draining_tasks(q, active_gc_threads); 2697 enqueue_dense_prefix_tasks(q, active_gc_threads); 2698 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2699 2700 { 2701 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2702 2703 gc_task_manager()->execute_and_wait(q); 2704 2705 #ifdef ASSERT 2706 // Verify that all regions have been processed before the deferred updates. 2707 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2708 verify_complete(SpaceId(id)); 2709 } 2710 #endif 2711 } 2712 2713 { 2714 // Update the deferred objects, if any. Any compaction manager can be used. 2715 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2716 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2717 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2718 update_deferred_objects(cm, SpaceId(id)); 2719 } 2720 } 2721 2722 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2723 } 2724 2725 #ifdef ASSERT 2726 void PSParallelCompact::verify_complete(SpaceId space_id) { 2727 // All Regions between space bottom() to new_top() should be marked as filled 2728 // and all Regions between new_top() and top() should be available (i.e., 2729 // should have been emptied). 2730 ParallelCompactData& sd = summary_data(); 2731 SpaceInfo si = _space_info[space_id]; 2732 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2733 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2734 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2735 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2736 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2737 2738 bool issued_a_warning = false; 2739 2740 size_t cur_region; 2741 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2742 const RegionData* const c = sd.region(cur_region); 2743 if (!c->completed()) { 2744 warning("region " SIZE_FORMAT " not filled: " 2745 "destination_count=%u", 2746 cur_region, c->destination_count()); 2747 issued_a_warning = true; 2748 } 2749 } 2750 2751 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2752 const RegionData* const c = sd.region(cur_region); 2753 if (!c->available()) { 2754 warning("region " SIZE_FORMAT " not empty: " 2755 "destination_count=%u", 2756 cur_region, c->destination_count()); 2757 issued_a_warning = true; 2758 } 2759 } 2760 2761 if (issued_a_warning) { 2762 print_region_ranges(); 2763 } 2764 } 2765 #endif // #ifdef ASSERT 2766 2767 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2768 _start_array->allocate_block(addr); 2769 compaction_manager()->update_contents(oop(addr)); 2770 } 2771 2772 // Update interior oops in the ranges of regions [beg_region, end_region). 2773 void 2774 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2775 SpaceId space_id, 2776 size_t beg_region, 2777 size_t end_region) { 2778 ParallelCompactData& sd = summary_data(); 2779 ParMarkBitMap* const mbm = mark_bitmap(); 2780 2781 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2782 HeapWord* const end_addr = sd.region_to_addr(end_region); 2783 assert(beg_region <= end_region, "bad region range"); 2784 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2785 2786 #ifdef ASSERT 2787 // Claim the regions to avoid triggering an assert when they are marked as 2788 // filled. 2789 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2790 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2791 } 2792 #endif // #ifdef ASSERT 2793 2794 if (beg_addr != space(space_id)->bottom()) { 2795 // Find the first live object or block of dead space that *starts* in this 2796 // range of regions. If a partial object crosses onto the region, skip it; 2797 // it will be marked for 'deferred update' when the object head is 2798 // processed. If dead space crosses onto the region, it is also skipped; it 2799 // will be filled when the prior region is processed. If neither of those 2800 // apply, the first word in the region is the start of a live object or dead 2801 // space. 2802 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2803 const RegionData* const cp = sd.region(beg_region); 2804 if (cp->partial_obj_size() != 0) { 2805 beg_addr = sd.partial_obj_end(beg_region); 2806 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2807 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2808 } 2809 } 2810 2811 if (beg_addr < end_addr) { 2812 // A live object or block of dead space starts in this range of Regions. 2813 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2814 2815 // Create closures and iterate. 2816 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2817 FillClosure fill_closure(cm, space_id); 2818 ParMarkBitMap::IterationStatus status; 2819 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2820 dense_prefix_end); 2821 if (status == ParMarkBitMap::incomplete) { 2822 update_closure.do_addr(update_closure.source()); 2823 } 2824 } 2825 2826 // Mark the regions as filled. 2827 RegionData* const beg_cp = sd.region(beg_region); 2828 RegionData* const end_cp = sd.region(end_region); 2829 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2830 cp->set_completed(); 2831 } 2832 } 2833 2834 // Return the SpaceId for the space containing addr. If addr is not in the 2835 // heap, last_space_id is returned. In debug mode it expects the address to be 2836 // in the heap and asserts such. 2837 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2838 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2839 2840 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2841 if (_space_info[id].space()->contains(addr)) { 2842 return SpaceId(id); 2843 } 2844 } 2845 2846 assert(false, "no space contains the addr"); 2847 return last_space_id; 2848 } 2849 2850 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2851 SpaceId id) { 2852 assert(id < last_space_id, "bad space id"); 2853 2854 ParallelCompactData& sd = summary_data(); 2855 const SpaceInfo* const space_info = _space_info + id; 2856 ObjectStartArray* const start_array = space_info->start_array(); 2857 2858 const MutableSpace* const space = space_info->space(); 2859 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2860 HeapWord* const beg_addr = space_info->dense_prefix(); 2861 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2862 2863 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2864 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2865 const RegionData* cur_region; 2866 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2867 HeapWord* const addr = cur_region->deferred_obj_addr(); 2868 if (addr != NULL) { 2869 if (start_array != NULL) { 2870 start_array->allocate_block(addr); 2871 } 2872 cm->update_contents(oop(addr)); 2873 assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)))); 2874 } 2875 } 2876 } 2877 2878 // Skip over count live words starting from beg, and return the address of the 2879 // next live word. Unless marked, the word corresponding to beg is assumed to 2880 // be dead. Callers must either ensure beg does not correspond to the middle of 2881 // an object, or account for those live words in some other way. Callers must 2882 // also ensure that there are enough live words in the range [beg, end) to skip. 2883 HeapWord* 2884 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2885 { 2886 assert(count > 0, "sanity"); 2887 2888 ParMarkBitMap* m = mark_bitmap(); 2889 idx_t bits_to_skip = m->words_to_bits(count); 2890 idx_t cur_beg = m->addr_to_bit(beg); 2891 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2892 2893 do { 2894 cur_beg = m->find_obj_beg(cur_beg, search_end); 2895 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2896 const size_t obj_bits = cur_end - cur_beg + 1; 2897 if (obj_bits > bits_to_skip) { 2898 return m->bit_to_addr(cur_beg + bits_to_skip); 2899 } 2900 bits_to_skip -= obj_bits; 2901 cur_beg = cur_end + 1; 2902 } while (bits_to_skip > 0); 2903 2904 // Skipping the desired number of words landed just past the end of an object. 2905 // Find the start of the next object. 2906 cur_beg = m->find_obj_beg(cur_beg, search_end); 2907 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2908 return m->bit_to_addr(cur_beg); 2909 } 2910 2911 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2912 SpaceId src_space_id, 2913 size_t src_region_idx) 2914 { 2915 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2916 2917 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2918 if (split_info.dest_region_addr() == dest_addr) { 2919 // The partial object ending at the split point contains the first word to 2920 // be copied to dest_addr. 2921 return split_info.first_src_addr(); 2922 } 2923 2924 const ParallelCompactData& sd = summary_data(); 2925 ParMarkBitMap* const bitmap = mark_bitmap(); 2926 const size_t RegionSize = ParallelCompactData::RegionSize; 2927 2928 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2929 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2930 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2931 HeapWord* const src_region_destination = src_region_ptr->destination(); 2932 2933 assert(dest_addr >= src_region_destination, "wrong src region"); 2934 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2935 2936 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2937 HeapWord* const src_region_end = src_region_beg + RegionSize; 2938 2939 HeapWord* addr = src_region_beg; 2940 if (dest_addr == src_region_destination) { 2941 // Return the first live word in the source region. 2942 if (partial_obj_size == 0) { 2943 addr = bitmap->find_obj_beg(addr, src_region_end); 2944 assert(addr < src_region_end, "no objects start in src region"); 2945 } 2946 return addr; 2947 } 2948 2949 // Must skip some live data. 2950 size_t words_to_skip = dest_addr - src_region_destination; 2951 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2952 2953 if (partial_obj_size >= words_to_skip) { 2954 // All the live words to skip are part of the partial object. 2955 addr += words_to_skip; 2956 if (partial_obj_size == words_to_skip) { 2957 // Find the first live word past the partial object. 2958 addr = bitmap->find_obj_beg(addr, src_region_end); 2959 assert(addr < src_region_end, "wrong src region"); 2960 } 2961 return addr; 2962 } 2963 2964 // Skip over the partial object (if any). 2965 if (partial_obj_size != 0) { 2966 words_to_skip -= partial_obj_size; 2967 addr += partial_obj_size; 2968 } 2969 2970 // Skip over live words due to objects that start in the region. 2971 addr = skip_live_words(addr, src_region_end, words_to_skip); 2972 assert(addr < src_region_end, "wrong src region"); 2973 return addr; 2974 } 2975 2976 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2977 SpaceId src_space_id, 2978 size_t beg_region, 2979 HeapWord* end_addr) 2980 { 2981 ParallelCompactData& sd = summary_data(); 2982 2983 #ifdef ASSERT 2984 MutableSpace* const src_space = _space_info[src_space_id].space(); 2985 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2986 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2987 "src_space_id does not match beg_addr"); 2988 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2989 "src_space_id does not match end_addr"); 2990 #endif // #ifdef ASSERT 2991 2992 RegionData* const beg = sd.region(beg_region); 2993 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2994 2995 // Regions up to new_top() are enqueued if they become available. 2996 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2997 RegionData* const enqueue_end = 2998 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2999 3000 for (RegionData* cur = beg; cur < end; ++cur) { 3001 assert(cur->data_size() > 0, "region must have live data"); 3002 cur->decrement_destination_count(); 3003 if (cur < enqueue_end && cur->available() && cur->claim()) { 3004 cm->push_region(sd.region(cur)); 3005 } 3006 } 3007 } 3008 3009 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3010 SpaceId& src_space_id, 3011 HeapWord*& src_space_top, 3012 HeapWord* end_addr) 3013 { 3014 typedef ParallelCompactData::RegionData RegionData; 3015 3016 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3017 const size_t region_size = ParallelCompactData::RegionSize; 3018 3019 size_t src_region_idx = 0; 3020 3021 // Skip empty regions (if any) up to the top of the space. 3022 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3023 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3024 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3025 const RegionData* const top_region_ptr = 3026 sd.addr_to_region_ptr(top_aligned_up); 3027 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3028 ++src_region_ptr; 3029 } 3030 3031 if (src_region_ptr < top_region_ptr) { 3032 // The next source region is in the current space. Update src_region_idx 3033 // and the source address to match src_region_ptr. 3034 src_region_idx = sd.region(src_region_ptr); 3035 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3036 if (src_region_addr > closure.source()) { 3037 closure.set_source(src_region_addr); 3038 } 3039 return src_region_idx; 3040 } 3041 3042 // Switch to a new source space and find the first non-empty region. 3043 unsigned int space_id = src_space_id + 1; 3044 assert(space_id < last_space_id, "not enough spaces"); 3045 3046 HeapWord* const destination = closure.destination(); 3047 3048 do { 3049 MutableSpace* space = _space_info[space_id].space(); 3050 HeapWord* const bottom = space->bottom(); 3051 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3052 3053 // Iterate over the spaces that do not compact into themselves. 3054 if (bottom_cp->destination() != bottom) { 3055 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3056 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3057 3058 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3059 if (src_cp->live_obj_size() > 0) { 3060 // Found it. 3061 assert(src_cp->destination() == destination, 3062 "first live obj in the space must match the destination"); 3063 assert(src_cp->partial_obj_size() == 0, 3064 "a space cannot begin with a partial obj"); 3065 3066 src_space_id = SpaceId(space_id); 3067 src_space_top = space->top(); 3068 const size_t src_region_idx = sd.region(src_cp); 3069 closure.set_source(sd.region_to_addr(src_region_idx)); 3070 return src_region_idx; 3071 } else { 3072 assert(src_cp->data_size() == 0, "sanity"); 3073 } 3074 } 3075 } 3076 } while (++space_id < last_space_id); 3077 3078 assert(false, "no source region was found"); 3079 return 0; 3080 } 3081 3082 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3083 { 3084 typedef ParMarkBitMap::IterationStatus IterationStatus; 3085 const size_t RegionSize = ParallelCompactData::RegionSize; 3086 ParMarkBitMap* const bitmap = mark_bitmap(); 3087 ParallelCompactData& sd = summary_data(); 3088 RegionData* const region_ptr = sd.region(region_idx); 3089 3090 // Get the items needed to construct the closure. 3091 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3092 SpaceId dest_space_id = space_id(dest_addr); 3093 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3094 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3095 assert(dest_addr < new_top, "sanity"); 3096 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3097 3098 // Get the source region and related info. 3099 size_t src_region_idx = region_ptr->source_region(); 3100 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3101 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3102 3103 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3104 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3105 3106 // Adjust src_region_idx to prepare for decrementing destination counts (the 3107 // destination count is not decremented when a region is copied to itself). 3108 if (src_region_idx == region_idx) { 3109 src_region_idx += 1; 3110 } 3111 3112 if (bitmap->is_unmarked(closure.source())) { 3113 // The first source word is in the middle of an object; copy the remainder 3114 // of the object or as much as will fit. The fact that pointer updates were 3115 // deferred will be noted when the object header is processed. 3116 HeapWord* const old_src_addr = closure.source(); 3117 closure.copy_partial_obj(); 3118 if (closure.is_full()) { 3119 decrement_destination_counts(cm, src_space_id, src_region_idx, 3120 closure.source()); 3121 region_ptr->set_deferred_obj_addr(NULL); 3122 region_ptr->set_completed(); 3123 return; 3124 } 3125 3126 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3127 if (sd.region_align_down(old_src_addr) != end_addr) { 3128 // The partial object was copied from more than one source region. 3129 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3130 3131 // Move to the next source region, possibly switching spaces as well. All 3132 // args except end_addr may be modified. 3133 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3134 end_addr); 3135 } 3136 } 3137 3138 do { 3139 HeapWord* const cur_addr = closure.source(); 3140 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3141 src_space_top); 3142 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3143 3144 if (status == ParMarkBitMap::incomplete) { 3145 // The last obj that starts in the source region does not end in the 3146 // region. 3147 assert(closure.source() < end_addr, "sanity"); 3148 HeapWord* const obj_beg = closure.source(); 3149 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3150 src_space_top); 3151 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3152 if (obj_end < range_end) { 3153 // The end was found; the entire object will fit. 3154 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3155 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3156 } else { 3157 // The end was not found; the object will not fit. 3158 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3159 status = ParMarkBitMap::would_overflow; 3160 } 3161 } 3162 3163 if (status == ParMarkBitMap::would_overflow) { 3164 // The last object did not fit. Note that interior oop updates were 3165 // deferred, then copy enough of the object to fill the region. 3166 region_ptr->set_deferred_obj_addr(closure.destination()); 3167 status = closure.copy_until_full(); // copies from closure.source() 3168 3169 decrement_destination_counts(cm, src_space_id, src_region_idx, 3170 closure.source()); 3171 region_ptr->set_completed(); 3172 return; 3173 } 3174 3175 if (status == ParMarkBitMap::full) { 3176 decrement_destination_counts(cm, src_space_id, src_region_idx, 3177 closure.source()); 3178 region_ptr->set_deferred_obj_addr(NULL); 3179 region_ptr->set_completed(); 3180 return; 3181 } 3182 3183 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3184 3185 // Move to the next source region, possibly switching spaces as well. All 3186 // args except end_addr may be modified. 3187 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3188 end_addr); 3189 } while (true); 3190 } 3191 3192 void PSParallelCompact::fill_blocks(size_t region_idx) 3193 { 3194 // Fill in the block table elements for the specified region. Each block 3195 // table element holds the number of live words in the region that are to the 3196 // left of the first object that starts in the block. Thus only blocks in 3197 // which an object starts need to be filled. 3198 // 3199 // The algorithm scans the section of the bitmap that corresponds to the 3200 // region, keeping a running total of the live words. When an object start is 3201 // found, if it's the first to start in the block that contains it, the 3202 // current total is written to the block table element. 3203 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3204 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3205 const size_t RegionSize = ParallelCompactData::RegionSize; 3206 3207 ParallelCompactData& sd = summary_data(); 3208 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3209 if (partial_obj_size >= RegionSize) { 3210 return; // No objects start in this region. 3211 } 3212 3213 // Ensure the first loop iteration decides that the block has changed. 3214 size_t cur_block = sd.block_count(); 3215 3216 const ParMarkBitMap* const bitmap = mark_bitmap(); 3217 3218 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3219 assert((size_t)1 << Log2BitsPerBlock == 3220 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3221 3222 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3223 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3224 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3225 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3226 while (beg_bit < range_end) { 3227 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3228 if (new_block != cur_block) { 3229 cur_block = new_block; 3230 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3231 } 3232 3233 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3234 if (end_bit < range_end - 1) { 3235 live_bits += end_bit - beg_bit + 1; 3236 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3237 } else { 3238 return; 3239 } 3240 } 3241 } 3242 3243 void 3244 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3245 const MutableSpace* sp = space(space_id); 3246 if (sp->is_empty()) { 3247 return; 3248 } 3249 3250 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3251 ParMarkBitMap* const bitmap = mark_bitmap(); 3252 HeapWord* const dp_addr = dense_prefix(space_id); 3253 HeapWord* beg_addr = sp->bottom(); 3254 HeapWord* end_addr = sp->top(); 3255 3256 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3257 3258 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3259 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3260 if (beg_region < dp_region) { 3261 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3262 } 3263 3264 // The destination of the first live object that starts in the region is one 3265 // past the end of the partial object entering the region (if any). 3266 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3267 HeapWord* const new_top = _space_info[space_id].new_top(); 3268 assert(new_top >= dest_addr, "bad new_top value"); 3269 const size_t words = pointer_delta(new_top, dest_addr); 3270 3271 if (words > 0) { 3272 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3273 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3274 3275 ParMarkBitMap::IterationStatus status; 3276 status = bitmap->iterate(&closure, dest_addr, end_addr); 3277 assert(status == ParMarkBitMap::full, "iteration not complete"); 3278 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3279 "live objects skipped because closure is full"); 3280 } 3281 } 3282 3283 jlong PSParallelCompact::millis_since_last_gc() { 3284 // We need a monotonically non-decreasing time in ms but 3285 // os::javaTimeMillis() does not guarantee monotonicity. 3286 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3287 jlong ret_val = now - _time_of_last_gc; 3288 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3289 if (ret_val < 0) { 3290 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);) 3291 return 0; 3292 } 3293 return ret_val; 3294 } 3295 3296 void PSParallelCompact::reset_millis_since_last_gc() { 3297 // We need a monotonically non-decreasing time in ms but 3298 // os::javaTimeMillis() does not guarantee monotonicity. 3299 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3300 } 3301 3302 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3303 { 3304 if (source() != destination()) { 3305 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3306 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3307 } 3308 update_state(words_remaining()); 3309 assert(is_full(), "sanity"); 3310 return ParMarkBitMap::full; 3311 } 3312 3313 void MoveAndUpdateClosure::copy_partial_obj() 3314 { 3315 size_t words = words_remaining(); 3316 3317 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3318 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3319 if (end_addr < range_end) { 3320 words = bitmap()->obj_size(source(), end_addr); 3321 } 3322 3323 // This test is necessary; if omitted, the pointer updates to a partial object 3324 // that crosses the dense prefix boundary could be overwritten. 3325 if (source() != destination()) { 3326 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3327 Copy::aligned_conjoint_words(source(), destination(), words); 3328 } 3329 update_state(words); 3330 } 3331 3332 void InstanceKlass::oop_pc_update_pointers(oop obj) { 3333 oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3334 } 3335 3336 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) { 3337 InstanceKlass::oop_pc_update_pointers(obj); 3338 3339 oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3340 } 3341 3342 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) { 3343 InstanceKlass::oop_pc_update_pointers(obj); 3344 } 3345 3346 #ifdef ASSERT 3347 template <class T> static void trace_reference_gc(const char *s, oop obj, 3348 T* referent_addr, 3349 T* next_addr, 3350 T* discovered_addr) { 3351 if(TraceReferenceGC && PrintGCDetails) { 3352 gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj)); 3353 gclog_or_tty->print_cr(" referent_addr/* " PTR_FORMAT " / " 3354 PTR_FORMAT, p2i(referent_addr), 3355 referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3356 gclog_or_tty->print_cr(" next_addr/* " PTR_FORMAT " / " 3357 PTR_FORMAT, p2i(next_addr), 3358 next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3359 gclog_or_tty->print_cr(" discovered_addr/* " PTR_FORMAT " / " 3360 PTR_FORMAT, p2i(discovered_addr), 3361 discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3362 } 3363 } 3364 #endif 3365 3366 template <class T> 3367 static void oop_pc_update_pointers_specialized(oop obj) { 3368 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3369 PSParallelCompact::adjust_pointer(referent_addr); 3370 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3371 PSParallelCompact::adjust_pointer(next_addr); 3372 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3373 PSParallelCompact::adjust_pointer(discovered_addr); 3374 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3375 referent_addr, next_addr, discovered_addr);) 3376 } 3377 3378 void InstanceRefKlass::oop_pc_update_pointers(oop obj) { 3379 InstanceKlass::oop_pc_update_pointers(obj); 3380 3381 if (UseCompressedOops) { 3382 oop_pc_update_pointers_specialized<narrowOop>(obj); 3383 } else { 3384 oop_pc_update_pointers_specialized<oop>(obj); 3385 } 3386 } 3387 3388 void ObjArrayKlass::oop_pc_update_pointers(oop obj) { 3389 assert(obj->is_objArray(), "obj must be obj array"); 3390 oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure()); 3391 } 3392 3393 void TypeArrayKlass::oop_pc_update_pointers(oop obj) { 3394 assert(obj->is_typeArray(),"must be a type array"); 3395 } 3396 3397 ParMarkBitMapClosure::IterationStatus 3398 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3399 assert(destination() != NULL, "sanity"); 3400 assert(bitmap()->obj_size(addr) == words, "bad size"); 3401 3402 _source = addr; 3403 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3404 destination(), "wrong destination"); 3405 3406 if (words > words_remaining()) { 3407 return ParMarkBitMap::would_overflow; 3408 } 3409 3410 // The start_array must be updated even if the object is not moving. 3411 if (_start_array != NULL) { 3412 _start_array->allocate_block(destination()); 3413 } 3414 3415 if (destination() != source()) { 3416 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3417 Copy::aligned_conjoint_words(source(), destination(), words); 3418 } 3419 3420 oop moved_oop = (oop) destination(); 3421 compaction_manager()->update_contents(moved_oop); 3422 assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop))); 3423 3424 update_state(words); 3425 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3426 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3427 } 3428 3429 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3430 ParCompactionManager* cm, 3431 PSParallelCompact::SpaceId space_id) : 3432 ParMarkBitMapClosure(mbm, cm), 3433 _space_id(space_id), 3434 _start_array(PSParallelCompact::start_array(space_id)) 3435 { 3436 } 3437 3438 // Updates the references in the object to their new values. 3439 ParMarkBitMapClosure::IterationStatus 3440 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3441 do_addr(addr); 3442 return ParMarkBitMap::incomplete; 3443 }