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/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc/parallel/gcTaskManager.hpp" 31 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 32 #include "gc/parallel/pcTasks.hpp" 33 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 34 #include "gc/parallel/psCompactionManager.inline.hpp" 35 #include "gc/parallel/psMarkSweep.hpp" 36 #include "gc/parallel/psMarkSweepDecorator.hpp" 37 #include "gc/parallel/psOldGen.hpp" 38 #include "gc/parallel/psParallelCompact.inline.hpp" 39 #include "gc/parallel/psPromotionManager.inline.hpp" 40 #include "gc/parallel/psScavenge.hpp" 41 #include "gc/parallel/psYoungGen.hpp" 42 #include "gc/shared/gcCause.hpp" 43 #include "gc/shared/gcHeapSummary.hpp" 44 #include "gc/shared/gcId.hpp" 45 #include "gc/shared/gcLocker.inline.hpp" 46 #include "gc/shared/gcTimer.hpp" 47 #include "gc/shared/gcTrace.hpp" 48 #include "gc/shared/gcTraceTime.hpp" 49 #include "gc/shared/isGCActiveMark.hpp" 50 #include "gc/shared/referencePolicy.hpp" 51 #include "gc/shared/referenceProcessor.hpp" 52 #include "gc/shared/spaceDecorator.hpp" 53 #include "oops/instanceKlass.inline.hpp" 54 #include "oops/instanceMirrorKlass.inline.hpp" 55 #include "oops/methodData.hpp" 56 #include "oops/objArrayKlass.inline.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "runtime/atomic.inline.hpp" 59 #include "runtime/fprofiler.hpp" 60 #include "runtime/safepoint.hpp" 61 #include "runtime/vmThread.hpp" 62 #include "services/management.hpp" 63 #include "services/memTracker.hpp" 64 #include "services/memoryService.hpp" 65 #include "utilities/events.hpp" 66 #include "utilities/stack.inline.hpp" 67 68 #include <math.h> 69 70 // All sizes are in HeapWords. 71 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 72 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 73 const size_t ParallelCompactData::RegionSizeBytes = 74 RegionSize << LogHeapWordSize; 75 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 76 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 77 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 78 79 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 80 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 81 const size_t ParallelCompactData::BlockSizeBytes = 82 BlockSize << LogHeapWordSize; 83 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 84 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 85 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 86 87 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 88 const size_t ParallelCompactData::Log2BlocksPerRegion = 89 Log2RegionSize - Log2BlockSize; 90 91 const ParallelCompactData::RegionData::region_sz_t 92 ParallelCompactData::RegionData::dc_shift = 27; 93 94 const ParallelCompactData::RegionData::region_sz_t 95 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 96 97 const ParallelCompactData::RegionData::region_sz_t 98 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 99 100 const ParallelCompactData::RegionData::region_sz_t 101 ParallelCompactData::RegionData::los_mask = ~dc_mask; 102 103 const ParallelCompactData::RegionData::region_sz_t 104 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 105 106 const ParallelCompactData::RegionData::region_sz_t 107 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 108 109 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 110 bool PSParallelCompact::_print_phases = false; 111 112 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 113 114 double PSParallelCompact::_dwl_mean; 115 double PSParallelCompact::_dwl_std_dev; 116 double PSParallelCompact::_dwl_first_term; 117 double PSParallelCompact::_dwl_adjustment; 118 #ifdef ASSERT 119 bool PSParallelCompact::_dwl_initialized = false; 120 #endif // #ifdef ASSERT 121 122 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 123 HeapWord* destination) 124 { 125 assert(src_region_idx != 0, "invalid src_region_idx"); 126 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 127 assert(destination != NULL, "invalid destination argument"); 128 129 _src_region_idx = src_region_idx; 130 _partial_obj_size = partial_obj_size; 131 _destination = destination; 132 133 // These fields may not be updated below, so make sure they're clear. 134 assert(_dest_region_addr == NULL, "should have been cleared"); 135 assert(_first_src_addr == NULL, "should have been cleared"); 136 137 // Determine the number of destination regions for the partial object. 138 HeapWord* const last_word = destination + partial_obj_size - 1; 139 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 140 HeapWord* const beg_region_addr = sd.region_align_down(destination); 141 HeapWord* const end_region_addr = sd.region_align_down(last_word); 142 143 if (beg_region_addr == end_region_addr) { 144 // One destination region. 145 _destination_count = 1; 146 if (end_region_addr == destination) { 147 // The destination falls on a region boundary, thus the first word of the 148 // partial object will be the first word copied to the destination region. 149 _dest_region_addr = end_region_addr; 150 _first_src_addr = sd.region_to_addr(src_region_idx); 151 } 152 } else { 153 // Two destination regions. When copied, the partial object will cross a 154 // destination region boundary, so a word somewhere within the partial 155 // object will be the first word copied to the second destination region. 156 _destination_count = 2; 157 _dest_region_addr = end_region_addr; 158 const size_t ofs = pointer_delta(end_region_addr, destination); 159 assert(ofs < _partial_obj_size, "sanity"); 160 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 161 } 162 } 163 164 void SplitInfo::clear() 165 { 166 _src_region_idx = 0; 167 _partial_obj_size = 0; 168 _destination = NULL; 169 _destination_count = 0; 170 _dest_region_addr = NULL; 171 _first_src_addr = NULL; 172 assert(!is_valid(), "sanity"); 173 } 174 175 #ifdef ASSERT 176 void SplitInfo::verify_clear() 177 { 178 assert(_src_region_idx == 0, "not clear"); 179 assert(_partial_obj_size == 0, "not clear"); 180 assert(_destination == NULL, "not clear"); 181 assert(_destination_count == 0, "not clear"); 182 assert(_dest_region_addr == NULL, "not clear"); 183 assert(_first_src_addr == NULL, "not clear"); 184 } 185 #endif // #ifdef ASSERT 186 187 188 void PSParallelCompact::print_on_error(outputStream* st) { 189 _mark_bitmap.print_on_error(st); 190 } 191 192 #ifndef PRODUCT 193 const char* PSParallelCompact::space_names[] = { 194 "old ", "eden", "from", "to " 195 }; 196 197 void PSParallelCompact::print_region_ranges() 198 { 199 tty->print_cr("space bottom top end new_top"); 200 tty->print_cr("------ ---------- ---------- ---------- ----------"); 201 202 for (unsigned int id = 0; id < last_space_id; ++id) { 203 const MutableSpace* space = _space_info[id].space(); 204 tty->print_cr("%u %s " 205 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 206 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 207 id, space_names[id], 208 summary_data().addr_to_region_idx(space->bottom()), 209 summary_data().addr_to_region_idx(space->top()), 210 summary_data().addr_to_region_idx(space->end()), 211 summary_data().addr_to_region_idx(_space_info[id].new_top())); 212 } 213 } 214 215 void 216 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 217 { 218 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 219 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 220 221 ParallelCompactData& sd = PSParallelCompact::summary_data(); 222 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 223 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 224 REGION_IDX_FORMAT " " PTR_FORMAT " " 225 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 226 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 227 i, p2i(c->data_location()), dci, p2i(c->destination()), 228 c->partial_obj_size(), c->live_obj_size(), 229 c->data_size(), c->source_region(), c->destination_count()); 230 231 #undef REGION_IDX_FORMAT 232 #undef REGION_DATA_FORMAT 233 } 234 235 void 236 print_generic_summary_data(ParallelCompactData& summary_data, 237 HeapWord* const beg_addr, 238 HeapWord* const end_addr) 239 { 240 size_t total_words = 0; 241 size_t i = summary_data.addr_to_region_idx(beg_addr); 242 const size_t last = summary_data.addr_to_region_idx(end_addr); 243 HeapWord* pdest = 0; 244 245 while (i <= last) { 246 ParallelCompactData::RegionData* c = summary_data.region(i); 247 if (c->data_size() != 0 || c->destination() != pdest) { 248 print_generic_summary_region(i, c); 249 total_words += c->data_size(); 250 pdest = c->destination(); 251 } 252 ++i; 253 } 254 255 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 256 } 257 258 void 259 print_generic_summary_data(ParallelCompactData& summary_data, 260 SpaceInfo* space_info) 261 { 262 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 263 const MutableSpace* space = space_info[id].space(); 264 print_generic_summary_data(summary_data, space->bottom(), 265 MAX2(space->top(), space_info[id].new_top())); 266 } 267 } 268 269 void 270 print_initial_summary_region(size_t i, 271 const ParallelCompactData::RegionData* c, 272 bool newline = true) 273 { 274 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 275 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 276 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 277 i, p2i(c->destination()), 278 c->partial_obj_size(), c->live_obj_size(), 279 c->data_size(), c->source_region(), c->destination_count()); 280 if (newline) tty->cr(); 281 } 282 283 void 284 print_initial_summary_data(ParallelCompactData& summary_data, 285 const MutableSpace* space) { 286 if (space->top() == space->bottom()) { 287 return; 288 } 289 290 const size_t region_size = ParallelCompactData::RegionSize; 291 typedef ParallelCompactData::RegionData RegionData; 292 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 293 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 294 const RegionData* c = summary_data.region(end_region - 1); 295 HeapWord* end_addr = c->destination() + c->data_size(); 296 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 297 298 // Print (and count) the full regions at the beginning of the space. 299 size_t full_region_count = 0; 300 size_t i = summary_data.addr_to_region_idx(space->bottom()); 301 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 302 print_initial_summary_region(i, summary_data.region(i)); 303 ++full_region_count; 304 ++i; 305 } 306 307 size_t live_to_right = live_in_space - full_region_count * region_size; 308 309 double max_reclaimed_ratio = 0.0; 310 size_t max_reclaimed_ratio_region = 0; 311 size_t max_dead_to_right = 0; 312 size_t max_live_to_right = 0; 313 314 // Print the 'reclaimed ratio' for regions while there is something live in 315 // the region or to the right of it. The remaining regions are empty (and 316 // uninteresting), and computing the ratio will result in division by 0. 317 while (i < end_region && live_to_right > 0) { 318 c = summary_data.region(i); 319 HeapWord* const region_addr = summary_data.region_to_addr(i); 320 const size_t used_to_right = pointer_delta(space->top(), region_addr); 321 const size_t dead_to_right = used_to_right - live_to_right; 322 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 323 324 if (reclaimed_ratio > max_reclaimed_ratio) { 325 max_reclaimed_ratio = reclaimed_ratio; 326 max_reclaimed_ratio_region = i; 327 max_dead_to_right = dead_to_right; 328 max_live_to_right = live_to_right; 329 } 330 331 print_initial_summary_region(i, c, false); 332 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 333 reclaimed_ratio, dead_to_right, live_to_right); 334 335 live_to_right -= c->data_size(); 336 ++i; 337 } 338 339 // Any remaining regions are empty. Print one more if there is one. 340 if (i < end_region) { 341 print_initial_summary_region(i, summary_data.region(i)); 342 } 343 344 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 345 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 346 max_reclaimed_ratio_region, max_dead_to_right, 347 max_live_to_right, max_reclaimed_ratio); 348 } 349 350 void 351 print_initial_summary_data(ParallelCompactData& summary_data, 352 SpaceInfo* space_info) { 353 unsigned int id = PSParallelCompact::old_space_id; 354 const MutableSpace* space; 355 do { 356 space = space_info[id].space(); 357 print_initial_summary_data(summary_data, space); 358 } while (++id < PSParallelCompact::eden_space_id); 359 360 do { 361 space = space_info[id].space(); 362 print_generic_summary_data(summary_data, space->bottom(), space->top()); 363 } while (++id < PSParallelCompact::last_space_id); 364 } 365 #endif // #ifndef PRODUCT 366 367 #ifdef ASSERT 368 size_t add_obj_count; 369 size_t add_obj_size; 370 size_t mark_bitmap_count; 371 size_t mark_bitmap_size; 372 #endif // #ifdef ASSERT 373 374 ParallelCompactData::ParallelCompactData() 375 { 376 _region_start = 0; 377 378 _region_vspace = 0; 379 _reserved_byte_size = 0; 380 _region_data = 0; 381 _region_count = 0; 382 383 _block_vspace = 0; 384 _block_data = 0; 385 _block_count = 0; 386 } 387 388 bool ParallelCompactData::initialize(MemRegion covered_region) 389 { 390 _region_start = covered_region.start(); 391 const size_t region_size = covered_region.word_size(); 392 DEBUG_ONLY(_region_end = _region_start + region_size;) 393 394 assert(region_align_down(_region_start) == _region_start, 395 "region start not aligned"); 396 assert((region_size & RegionSizeOffsetMask) == 0, 397 "region size not a multiple of RegionSize"); 398 399 bool result = initialize_region_data(region_size) && initialize_block_data(); 400 return result; 401 } 402 403 PSVirtualSpace* 404 ParallelCompactData::create_vspace(size_t count, size_t element_size) 405 { 406 const size_t raw_bytes = count * element_size; 407 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 408 const size_t granularity = os::vm_allocation_granularity(); 409 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 410 411 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 412 MAX2(page_sz, granularity); 413 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 414 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 415 rs.size()); 416 417 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 418 419 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 420 if (vspace != 0) { 421 if (vspace->expand_by(_reserved_byte_size)) { 422 return vspace; 423 } 424 delete vspace; 425 // Release memory reserved in the space. 426 rs.release(); 427 } 428 429 return 0; 430 } 431 432 bool ParallelCompactData::initialize_region_data(size_t region_size) 433 { 434 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 435 _region_vspace = create_vspace(count, sizeof(RegionData)); 436 if (_region_vspace != 0) { 437 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 438 _region_count = count; 439 return true; 440 } 441 return false; 442 } 443 444 bool ParallelCompactData::initialize_block_data() 445 { 446 assert(_region_count != 0, "region data must be initialized first"); 447 const size_t count = _region_count << Log2BlocksPerRegion; 448 _block_vspace = create_vspace(count, sizeof(BlockData)); 449 if (_block_vspace != 0) { 450 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 451 _block_count = count; 452 return true; 453 } 454 return false; 455 } 456 457 void ParallelCompactData::clear() 458 { 459 memset(_region_data, 0, _region_vspace->committed_size()); 460 memset(_block_data, 0, _block_vspace->committed_size()); 461 } 462 463 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 464 assert(beg_region <= _region_count, "beg_region out of range"); 465 assert(end_region <= _region_count, "end_region out of range"); 466 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 467 468 const size_t region_cnt = end_region - beg_region; 469 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 470 471 const size_t beg_block = beg_region * BlocksPerRegion; 472 const size_t block_cnt = region_cnt * BlocksPerRegion; 473 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 474 } 475 476 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 477 { 478 const RegionData* cur_cp = region(region_idx); 479 const RegionData* const end_cp = region(region_count() - 1); 480 481 HeapWord* result = region_to_addr(region_idx); 482 if (cur_cp < end_cp) { 483 do { 484 result += cur_cp->partial_obj_size(); 485 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 486 } 487 return result; 488 } 489 490 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 491 { 492 const size_t obj_ofs = pointer_delta(addr, _region_start); 493 const size_t beg_region = obj_ofs >> Log2RegionSize; 494 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 495 496 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 497 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 498 499 if (beg_region == end_region) { 500 // All in one region. 501 _region_data[beg_region].add_live_obj(len); 502 return; 503 } 504 505 // First region. 506 const size_t beg_ofs = region_offset(addr); 507 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 508 509 Klass* klass = ((oop)addr)->klass(); 510 // Middle regions--completely spanned by this object. 511 for (size_t region = beg_region + 1; region < end_region; ++region) { 512 _region_data[region].set_partial_obj_size(RegionSize); 513 _region_data[region].set_partial_obj_addr(addr); 514 } 515 516 // Last region. 517 const size_t end_ofs = region_offset(addr + len - 1); 518 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 519 _region_data[end_region].set_partial_obj_addr(addr); 520 } 521 522 void 523 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 524 { 525 assert(region_offset(beg) == 0, "not RegionSize aligned"); 526 assert(region_offset(end) == 0, "not RegionSize aligned"); 527 528 size_t cur_region = addr_to_region_idx(beg); 529 const size_t end_region = addr_to_region_idx(end); 530 HeapWord* addr = beg; 531 while (cur_region < end_region) { 532 _region_data[cur_region].set_destination(addr); 533 _region_data[cur_region].set_destination_count(0); 534 _region_data[cur_region].set_source_region(cur_region); 535 _region_data[cur_region].set_data_location(addr); 536 537 // Update live_obj_size so the region appears completely full. 538 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 539 _region_data[cur_region].set_live_obj_size(live_size); 540 541 ++cur_region; 542 addr += RegionSize; 543 } 544 } 545 546 // Find the point at which a space can be split and, if necessary, record the 547 // split point. 548 // 549 // If the current src region (which overflowed the destination space) doesn't 550 // have a partial object, the split point is at the beginning of the current src 551 // region (an "easy" split, no extra bookkeeping required). 552 // 553 // If the current src region has a partial object, the split point is in the 554 // region where that partial object starts (call it the split_region). If 555 // split_region has a partial object, then the split point is just after that 556 // partial object (a "hard" split where we have to record the split data and 557 // zero the partial_obj_size field). With a "hard" split, we know that the 558 // partial_obj ends within split_region because the partial object that caused 559 // the overflow starts in split_region. If split_region doesn't have a partial 560 // obj, then the split is at the beginning of split_region (another "easy" 561 // split). 562 HeapWord* 563 ParallelCompactData::summarize_split_space(size_t src_region, 564 SplitInfo& split_info, 565 HeapWord* destination, 566 HeapWord* target_end, 567 HeapWord** target_next) 568 { 569 assert(destination <= target_end, "sanity"); 570 assert(destination + _region_data[src_region].data_size() > target_end, 571 "region should not fit into target space"); 572 assert(is_region_aligned(target_end), "sanity"); 573 574 size_t split_region = src_region; 575 HeapWord* split_destination = destination; 576 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 577 578 if (destination + partial_obj_size > target_end) { 579 // The split point is just after the partial object (if any) in the 580 // src_region that contains the start of the object that overflowed the 581 // destination space. 582 // 583 // Find the start of the "overflow" object and set split_region to the 584 // region containing it. 585 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 586 split_region = addr_to_region_idx(overflow_obj); 587 588 // Clear the source_region field of all destination regions whose first word 589 // came from data after the split point (a non-null source_region field 590 // implies a region must be filled). 591 // 592 // An alternative to the simple loop below: clear during post_compact(), 593 // which uses memcpy instead of individual stores, and is easy to 594 // parallelize. (The downside is that it clears the entire RegionData 595 // object as opposed to just one field.) 596 // 597 // post_compact() would have to clear the summary data up to the highest 598 // address that was written during the summary phase, which would be 599 // 600 // max(top, max(new_top, clear_top)) 601 // 602 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 603 // to target_end. 604 const RegionData* const sr = region(split_region); 605 const size_t beg_idx = 606 addr_to_region_idx(region_align_up(sr->destination() + 607 sr->partial_obj_size())); 608 const size_t end_idx = addr_to_region_idx(target_end); 609 610 if (TraceParallelOldGCSummaryPhase) { 611 gclog_or_tty->print_cr("split: clearing source_region field in [" 612 SIZE_FORMAT ", " SIZE_FORMAT ")", 613 beg_idx, end_idx); 614 } 615 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 616 _region_data[idx].set_source_region(0); 617 } 618 619 // Set split_destination and partial_obj_size to reflect the split region. 620 split_destination = sr->destination(); 621 partial_obj_size = sr->partial_obj_size(); 622 } 623 624 // The split is recorded only if a partial object extends onto the region. 625 if (partial_obj_size != 0) { 626 _region_data[split_region].set_partial_obj_size(0); 627 split_info.record(split_region, partial_obj_size, split_destination); 628 } 629 630 // Setup the continuation addresses. 631 *target_next = split_destination + partial_obj_size; 632 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 633 634 if (TraceParallelOldGCSummaryPhase) { 635 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 636 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 637 " pos=" SIZE_FORMAT, 638 split_type, p2i(source_next), split_region, 639 partial_obj_size); 640 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 641 " tn=" PTR_FORMAT, 642 split_type, p2i(split_destination), 643 addr_to_region_idx(split_destination), 644 p2i(*target_next)); 645 646 if (partial_obj_size != 0) { 647 HeapWord* const po_beg = split_info.destination(); 648 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 649 gclog_or_tty->print_cr("%s split: " 650 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 651 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 652 split_type, 653 p2i(po_beg), addr_to_region_idx(po_beg), 654 p2i(po_end), addr_to_region_idx(po_end)); 655 } 656 } 657 658 return source_next; 659 } 660 661 bool ParallelCompactData::summarize(SplitInfo& split_info, 662 HeapWord* source_beg, HeapWord* source_end, 663 HeapWord** source_next, 664 HeapWord* target_beg, HeapWord* target_end, 665 HeapWord** target_next) 666 { 667 if (TraceParallelOldGCSummaryPhase) { 668 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 669 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 670 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 671 p2i(source_beg), p2i(source_end), p2i(source_next_val), 672 p2i(target_beg), p2i(target_end), p2i(*target_next)); 673 } 674 675 size_t cur_region = addr_to_region_idx(source_beg); 676 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 677 678 HeapWord *dest_addr = target_beg; 679 while (cur_region < end_region) { 680 // The destination must be set even if the region has no data. 681 _region_data[cur_region].set_destination(dest_addr); 682 683 size_t words = _region_data[cur_region].data_size(); 684 if (words > 0) { 685 // If cur_region does not fit entirely into the target space, find a point 686 // at which the source space can be 'split' so that part is copied to the 687 // target space and the rest is copied elsewhere. 688 if (dest_addr + words > target_end) { 689 assert(source_next != NULL, "source_next is NULL when splitting"); 690 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 691 target_end, target_next); 692 return false; 693 } 694 695 // Compute the destination_count for cur_region, and if necessary, update 696 // source_region for a destination region. The source_region field is 697 // updated if cur_region is the first (left-most) region to be copied to a 698 // destination region. 699 // 700 // The destination_count calculation is a bit subtle. A region that has 701 // data that compacts into itself does not count itself as a destination. 702 // This maintains the invariant that a zero count means the region is 703 // available and can be claimed and then filled. 704 uint destination_count = 0; 705 if (split_info.is_split(cur_region)) { 706 // The current region has been split: the partial object will be copied 707 // to one destination space and the remaining data will be copied to 708 // another destination space. Adjust the initial destination_count and, 709 // if necessary, set the source_region field if the partial object will 710 // cross a destination region boundary. 711 destination_count = split_info.destination_count(); 712 if (destination_count == 2) { 713 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 714 _region_data[dest_idx].set_source_region(cur_region); 715 } 716 } 717 718 HeapWord* const last_addr = dest_addr + words - 1; 719 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 720 const size_t dest_region_2 = addr_to_region_idx(last_addr); 721 722 // Initially assume that the destination regions will be the same and 723 // adjust the value below if necessary. Under this assumption, if 724 // cur_region == dest_region_2, then cur_region will be compacted 725 // completely into itself. 726 destination_count += cur_region == dest_region_2 ? 0 : 1; 727 if (dest_region_1 != dest_region_2) { 728 // Destination regions differ; adjust destination_count. 729 destination_count += 1; 730 // Data from cur_region will be copied to the start of dest_region_2. 731 _region_data[dest_region_2].set_source_region(cur_region); 732 } else if (region_offset(dest_addr) == 0) { 733 // Data from cur_region will be copied to the start of the destination 734 // region. 735 _region_data[dest_region_1].set_source_region(cur_region); 736 } 737 738 _region_data[cur_region].set_destination_count(destination_count); 739 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 740 dest_addr += words; 741 } 742 743 ++cur_region; 744 } 745 746 *target_next = dest_addr; 747 return true; 748 } 749 750 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 751 assert(addr != NULL, "Should detect NULL oop earlier"); 752 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 753 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 754 755 // Region covering the object. 756 RegionData* const region_ptr = addr_to_region_ptr(addr); 757 HeapWord* result = region_ptr->destination(); 758 759 // If the entire Region is live, the new location is region->destination + the 760 // offset of the object within in the Region. 761 762 // Run some performance tests to determine if this special case pays off. It 763 // is worth it for pointers into the dense prefix. If the optimization to 764 // avoid pointer updates in regions that only point to the dense prefix is 765 // ever implemented, this should be revisited. 766 if (region_ptr->data_size() == RegionSize) { 767 result += region_offset(addr); 768 return result; 769 } 770 771 // Otherwise, the new location is region->destination + block offset + the 772 // number of live words in the Block that are (a) to the left of addr and (b) 773 // due to objects that start in the Block. 774 775 // Fill in the block table if necessary. This is unsynchronized, so multiple 776 // threads may fill the block table for a region (harmless, since it is 777 // idempotent). 778 if (!region_ptr->blocks_filled()) { 779 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 780 region_ptr->set_blocks_filled(); 781 } 782 783 HeapWord* const search_start = block_align_down(addr); 784 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 785 786 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 787 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 788 result += block_offset + live; 789 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 790 return result; 791 } 792 793 #ifdef ASSERT 794 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 795 { 796 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 797 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 798 for (const size_t* p = beg; p < end; ++p) { 799 assert(*p == 0, "not zero"); 800 } 801 } 802 803 void ParallelCompactData::verify_clear() 804 { 805 verify_clear(_region_vspace); 806 verify_clear(_block_vspace); 807 } 808 #endif // #ifdef ASSERT 809 810 STWGCTimer PSParallelCompact::_gc_timer; 811 ParallelOldTracer PSParallelCompact::_gc_tracer; 812 elapsedTimer PSParallelCompact::_accumulated_time; 813 unsigned int PSParallelCompact::_total_invocations = 0; 814 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 815 jlong PSParallelCompact::_time_of_last_gc = 0; 816 CollectorCounters* PSParallelCompact::_counters = NULL; 817 ParMarkBitMap PSParallelCompact::_mark_bitmap; 818 ParallelCompactData PSParallelCompact::_summary_data; 819 820 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 821 822 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 823 824 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 825 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 826 827 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 828 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 829 } 830 831 void PSParallelCompact::post_initialize() { 832 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 833 MemRegion mr = heap->reserved_region(); 834 _ref_processor = 835 new ReferenceProcessor(mr, // span 836 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 837 ParallelGCThreads, // mt processing degree 838 true, // mt discovery 839 ParallelGCThreads, // mt discovery degree 840 true, // atomic_discovery 841 &_is_alive_closure); // non-header is alive closure 842 _counters = new CollectorCounters("PSParallelCompact", 1); 843 844 // Initialize static fields in ParCompactionManager. 845 ParCompactionManager::initialize(mark_bitmap()); 846 } 847 848 bool PSParallelCompact::initialize() { 849 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 850 MemRegion mr = heap->reserved_region(); 851 852 // Was the old gen get allocated successfully? 853 if (!heap->old_gen()->is_allocated()) { 854 return false; 855 } 856 857 initialize_space_info(); 858 initialize_dead_wood_limiter(); 859 860 if (!_mark_bitmap.initialize(mr)) { 861 vm_shutdown_during_initialization( 862 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 863 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 864 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 865 return false; 866 } 867 868 if (!_summary_data.initialize(mr)) { 869 vm_shutdown_during_initialization( 870 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 871 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 872 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 873 return false; 874 } 875 876 return true; 877 } 878 879 void PSParallelCompact::initialize_space_info() 880 { 881 memset(&_space_info, 0, sizeof(_space_info)); 882 883 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 884 PSYoungGen* young_gen = heap->young_gen(); 885 886 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 887 _space_info[eden_space_id].set_space(young_gen->eden_space()); 888 _space_info[from_space_id].set_space(young_gen->from_space()); 889 _space_info[to_space_id].set_space(young_gen->to_space()); 890 891 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 892 } 893 894 void PSParallelCompact::initialize_dead_wood_limiter() 895 { 896 const size_t max = 100; 897 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 898 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 899 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 900 DEBUG_ONLY(_dwl_initialized = true;) 901 _dwl_adjustment = normal_distribution(1.0); 902 } 903 904 // Simple class for storing info about the heap at the start of GC, to be used 905 // after GC for comparison/printing. 906 class PreGCValues { 907 public: 908 PreGCValues() { } 909 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 910 911 void fill(ParallelScavengeHeap* heap) { 912 _heap_used = heap->used(); 913 _young_gen_used = heap->young_gen()->used_in_bytes(); 914 _old_gen_used = heap->old_gen()->used_in_bytes(); 915 _metadata_used = MetaspaceAux::used_bytes(); 916 }; 917 918 size_t heap_used() const { return _heap_used; } 919 size_t young_gen_used() const { return _young_gen_used; } 920 size_t old_gen_used() const { return _old_gen_used; } 921 size_t metadata_used() const { return _metadata_used; } 922 923 private: 924 size_t _heap_used; 925 size_t _young_gen_used; 926 size_t _old_gen_used; 927 size_t _metadata_used; 928 }; 929 930 void 931 PSParallelCompact::clear_data_covering_space(SpaceId id) 932 { 933 // At this point, top is the value before GC, new_top() is the value that will 934 // be set at the end of GC. The marking bitmap is cleared to top; nothing 935 // should be marked above top. The summary data is cleared to the larger of 936 // top & new_top. 937 MutableSpace* const space = _space_info[id].space(); 938 HeapWord* const bot = space->bottom(); 939 HeapWord* const top = space->top(); 940 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 941 942 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 943 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 944 _mark_bitmap.clear_range(beg_bit, end_bit); 945 946 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 947 const size_t end_region = 948 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 949 _summary_data.clear_range(beg_region, end_region); 950 951 // Clear the data used to 'split' regions. 952 SplitInfo& split_info = _space_info[id].split_info(); 953 if (split_info.is_valid()) { 954 split_info.clear(); 955 } 956 DEBUG_ONLY(split_info.verify_clear();) 957 } 958 959 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 960 { 961 // Update the from & to space pointers in space_info, since they are swapped 962 // at each young gen gc. Do the update unconditionally (even though a 963 // promotion failure does not swap spaces) because an unknown number of young 964 // collections will have swapped the spaces an unknown number of times. 965 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer); 966 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 967 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 968 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 969 970 pre_gc_values->fill(heap); 971 972 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 973 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 974 975 // Increment the invocation count 976 heap->increment_total_collections(true); 977 978 // We need to track unique mark sweep invocations as well. 979 _total_invocations++; 980 981 heap->print_heap_before_gc(); 982 heap->trace_heap_before_gc(&_gc_tracer); 983 984 // Fill in TLABs 985 heap->accumulate_statistics_all_tlabs(); 986 heap->ensure_parsability(true); // retire TLABs 987 988 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 989 HandleMark hm; // Discard invalid handles created during verification 990 Universe::verify(" VerifyBeforeGC:"); 991 } 992 993 // Verify object start arrays 994 if (VerifyObjectStartArray && 995 VerifyBeforeGC) { 996 heap->old_gen()->verify_object_start_array(); 997 } 998 999 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1000 DEBUG_ONLY(summary_data().verify_clear();) 1001 1002 // Have worker threads release resources the next time they run a task. 1003 gc_task_manager()->release_all_resources(); 1004 } 1005 1006 void PSParallelCompact::post_compact() 1007 { 1008 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer); 1009 1010 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1011 // Clear the marking bitmap, summary data and split info. 1012 clear_data_covering_space(SpaceId(id)); 1013 // Update top(). Must be done after clearing the bitmap and summary data. 1014 _space_info[id].publish_new_top(); 1015 } 1016 1017 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1018 MutableSpace* const from_space = _space_info[from_space_id].space(); 1019 MutableSpace* const to_space = _space_info[to_space_id].space(); 1020 1021 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1022 bool eden_empty = eden_space->is_empty(); 1023 if (!eden_empty) { 1024 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1025 heap->young_gen(), heap->old_gen()); 1026 } 1027 1028 // Update heap occupancy information which is used as input to the soft ref 1029 // clearing policy at the next gc. 1030 Universe::update_heap_info_at_gc(); 1031 1032 bool young_gen_empty = eden_empty && from_space->is_empty() && 1033 to_space->is_empty(); 1034 1035 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1036 MemRegion old_mr = heap->old_gen()->reserved(); 1037 if (young_gen_empty) { 1038 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1039 } else { 1040 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1041 } 1042 1043 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1044 ClassLoaderDataGraph::purge(); 1045 MetaspaceAux::verify_metrics(); 1046 1047 CodeCache::gc_epilogue(); 1048 JvmtiExport::gc_epilogue(); 1049 1050 #if defined(COMPILER2) || INCLUDE_JVMCI 1051 DerivedPointerTable::update_pointers(); 1052 #endif 1053 1054 ref_processor()->enqueue_discovered_references(NULL); 1055 1056 if (ZapUnusedHeapArea) { 1057 heap->gen_mangle_unused_area(); 1058 } 1059 1060 // Update time of last GC 1061 reset_millis_since_last_gc(); 1062 } 1063 1064 HeapWord* 1065 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1066 bool maximum_compaction) 1067 { 1068 const size_t region_size = ParallelCompactData::RegionSize; 1069 const ParallelCompactData& sd = summary_data(); 1070 1071 const MutableSpace* const space = _space_info[id].space(); 1072 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1073 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1074 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1075 1076 // Skip full regions at the beginning of the space--they are necessarily part 1077 // of the dense prefix. 1078 size_t full_count = 0; 1079 const RegionData* cp; 1080 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1081 ++full_count; 1082 } 1083 1084 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1085 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1086 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1087 if (maximum_compaction || cp == end_cp || interval_ended) { 1088 _maximum_compaction_gc_num = total_invocations(); 1089 return sd.region_to_addr(cp); 1090 } 1091 1092 HeapWord* const new_top = _space_info[id].new_top(); 1093 const size_t space_live = pointer_delta(new_top, space->bottom()); 1094 const size_t space_used = space->used_in_words(); 1095 const size_t space_capacity = space->capacity_in_words(); 1096 1097 const double cur_density = double(space_live) / space_capacity; 1098 const double deadwood_density = 1099 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1100 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1101 1102 if (TraceParallelOldGCDensePrefix) { 1103 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1104 cur_density, deadwood_density, deadwood_goal); 1105 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1106 "space_cap=" SIZE_FORMAT, 1107 space_live, space_used, 1108 space_capacity); 1109 } 1110 1111 // XXX - Use binary search? 1112 HeapWord* dense_prefix = sd.region_to_addr(cp); 1113 const RegionData* full_cp = cp; 1114 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1115 while (cp < end_cp) { 1116 HeapWord* region_destination = cp->destination(); 1117 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1118 if (TraceParallelOldGCDensePrefix && Verbose) { 1119 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1120 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1121 sd.region(cp), p2i(region_destination), 1122 p2i(dense_prefix), cur_deadwood); 1123 } 1124 1125 if (cur_deadwood >= deadwood_goal) { 1126 // Found the region that has the correct amount of deadwood to the left. 1127 // This typically occurs after crossing a fairly sparse set of regions, so 1128 // iterate backwards over those sparse regions, looking for the region 1129 // that has the lowest density of live objects 'to the right.' 1130 size_t space_to_left = sd.region(cp) * region_size; 1131 size_t live_to_left = space_to_left - cur_deadwood; 1132 size_t space_to_right = space_capacity - space_to_left; 1133 size_t live_to_right = space_live - live_to_left; 1134 double density_to_right = double(live_to_right) / space_to_right; 1135 while (cp > full_cp) { 1136 --cp; 1137 const size_t prev_region_live_to_right = live_to_right - 1138 cp->data_size(); 1139 const size_t prev_region_space_to_right = space_to_right + region_size; 1140 double prev_region_density_to_right = 1141 double(prev_region_live_to_right) / prev_region_space_to_right; 1142 if (density_to_right <= prev_region_density_to_right) { 1143 return dense_prefix; 1144 } 1145 if (TraceParallelOldGCDensePrefix && Verbose) { 1146 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1147 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1148 prev_region_density_to_right); 1149 } 1150 dense_prefix -= region_size; 1151 live_to_right = prev_region_live_to_right; 1152 space_to_right = prev_region_space_to_right; 1153 density_to_right = prev_region_density_to_right; 1154 } 1155 return dense_prefix; 1156 } 1157 1158 dense_prefix += region_size; 1159 ++cp; 1160 } 1161 1162 return dense_prefix; 1163 } 1164 1165 #ifndef PRODUCT 1166 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1167 const SpaceId id, 1168 const bool maximum_compaction, 1169 HeapWord* const addr) 1170 { 1171 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1172 RegionData* const cp = summary_data().region(region_idx); 1173 const MutableSpace* const space = _space_info[id].space(); 1174 HeapWord* const new_top = _space_info[id].new_top(); 1175 1176 const size_t space_live = pointer_delta(new_top, space->bottom()); 1177 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1178 const size_t space_cap = space->capacity_in_words(); 1179 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1180 const size_t live_to_right = new_top - cp->destination(); 1181 const size_t dead_to_right = space->top() - addr - live_to_right; 1182 1183 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1184 "spl=" SIZE_FORMAT " " 1185 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1186 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1187 " ratio=%10.8f", 1188 algorithm, p2i(addr), region_idx, 1189 space_live, 1190 dead_to_left, dead_to_left_pct, 1191 dead_to_right, live_to_right, 1192 double(dead_to_right) / live_to_right); 1193 } 1194 #endif // #ifndef PRODUCT 1195 1196 // Return a fraction indicating how much of the generation can be treated as 1197 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1198 // based on the density of live objects in the generation to determine a limit, 1199 // which is then adjusted so the return value is min_percent when the density is 1200 // 1. 1201 // 1202 // The following table shows some return values for a different values of the 1203 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1204 // min_percent is 1. 1205 // 1206 // fraction allowed as dead wood 1207 // ----------------------------------------------------------------- 1208 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1209 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1210 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1211 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1212 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1213 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1214 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1215 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1216 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1217 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1218 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1219 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1220 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1221 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1222 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1223 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1224 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1225 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1226 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1227 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1228 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1229 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1230 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1231 1232 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1233 { 1234 assert(_dwl_initialized, "uninitialized"); 1235 1236 // The raw limit is the value of the normal distribution at x = density. 1237 const double raw_limit = normal_distribution(density); 1238 1239 // Adjust the raw limit so it becomes the minimum when the density is 1. 1240 // 1241 // First subtract the adjustment value (which is simply the precomputed value 1242 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1243 // Then add the minimum value, so the minimum is returned when the density is 1244 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1245 const double min = double(min_percent) / 100.0; 1246 const double limit = raw_limit - _dwl_adjustment + min; 1247 return MAX2(limit, 0.0); 1248 } 1249 1250 ParallelCompactData::RegionData* 1251 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1252 const RegionData* end) 1253 { 1254 const size_t region_size = ParallelCompactData::RegionSize; 1255 ParallelCompactData& sd = summary_data(); 1256 size_t left = sd.region(beg); 1257 size_t right = end > beg ? sd.region(end) - 1 : left; 1258 1259 // Binary search. 1260 while (left < right) { 1261 // Equivalent to (left + right) / 2, but does not overflow. 1262 const size_t middle = left + (right - left) / 2; 1263 RegionData* const middle_ptr = sd.region(middle); 1264 HeapWord* const dest = middle_ptr->destination(); 1265 HeapWord* const addr = sd.region_to_addr(middle); 1266 assert(dest != NULL, "sanity"); 1267 assert(dest <= addr, "must move left"); 1268 1269 if (middle > left && dest < addr) { 1270 right = middle - 1; 1271 } else if (middle < right && middle_ptr->data_size() == region_size) { 1272 left = middle + 1; 1273 } else { 1274 return middle_ptr; 1275 } 1276 } 1277 return sd.region(left); 1278 } 1279 1280 ParallelCompactData::RegionData* 1281 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1282 const RegionData* end, 1283 size_t dead_words) 1284 { 1285 ParallelCompactData& sd = summary_data(); 1286 size_t left = sd.region(beg); 1287 size_t right = end > beg ? sd.region(end) - 1 : left; 1288 1289 // Binary search. 1290 while (left < right) { 1291 // Equivalent to (left + right) / 2, but does not overflow. 1292 const size_t middle = left + (right - left) / 2; 1293 RegionData* const middle_ptr = sd.region(middle); 1294 HeapWord* const dest = middle_ptr->destination(); 1295 HeapWord* const addr = sd.region_to_addr(middle); 1296 assert(dest != NULL, "sanity"); 1297 assert(dest <= addr, "must move left"); 1298 1299 const size_t dead_to_left = pointer_delta(addr, dest); 1300 if (middle > left && dead_to_left > dead_words) { 1301 right = middle - 1; 1302 } else if (middle < right && dead_to_left < dead_words) { 1303 left = middle + 1; 1304 } else { 1305 return middle_ptr; 1306 } 1307 } 1308 return sd.region(left); 1309 } 1310 1311 // The result is valid during the summary phase, after the initial summarization 1312 // of each space into itself, and before final summarization. 1313 inline double 1314 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1315 HeapWord* const bottom, 1316 HeapWord* const top, 1317 HeapWord* const new_top) 1318 { 1319 ParallelCompactData& sd = summary_data(); 1320 1321 assert(cp != NULL, "sanity"); 1322 assert(bottom != NULL, "sanity"); 1323 assert(top != NULL, "sanity"); 1324 assert(new_top != NULL, "sanity"); 1325 assert(top >= new_top, "summary data problem?"); 1326 assert(new_top > bottom, "space is empty; should not be here"); 1327 assert(new_top >= cp->destination(), "sanity"); 1328 assert(top >= sd.region_to_addr(cp), "sanity"); 1329 1330 HeapWord* const destination = cp->destination(); 1331 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1332 const size_t compacted_region_live = pointer_delta(new_top, destination); 1333 const size_t compacted_region_used = pointer_delta(top, 1334 sd.region_to_addr(cp)); 1335 const size_t reclaimable = compacted_region_used - compacted_region_live; 1336 1337 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1338 return double(reclaimable) / divisor; 1339 } 1340 1341 // Return the address of the end of the dense prefix, a.k.a. the start of the 1342 // compacted region. The address is always on a region boundary. 1343 // 1344 // Completely full regions at the left are skipped, since no compaction can 1345 // occur in those regions. Then the maximum amount of dead wood to allow is 1346 // computed, based on the density (amount live / capacity) of the generation; 1347 // the region with approximately that amount of dead space to the left is 1348 // identified as the limit region. Regions between the last completely full 1349 // region and the limit region are scanned and the one that has the best 1350 // (maximum) reclaimed_ratio() is selected. 1351 HeapWord* 1352 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1353 bool maximum_compaction) 1354 { 1355 const size_t region_size = ParallelCompactData::RegionSize; 1356 const ParallelCompactData& sd = summary_data(); 1357 1358 const MutableSpace* const space = _space_info[id].space(); 1359 HeapWord* const top = space->top(); 1360 HeapWord* const top_aligned_up = sd.region_align_up(top); 1361 HeapWord* const new_top = _space_info[id].new_top(); 1362 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1363 HeapWord* const bottom = space->bottom(); 1364 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1365 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1366 const RegionData* const new_top_cp = 1367 sd.addr_to_region_ptr(new_top_aligned_up); 1368 1369 // Skip full regions at the beginning of the space--they are necessarily part 1370 // of the dense prefix. 1371 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1372 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1373 space->is_empty(), "no dead space allowed to the left"); 1374 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1375 "region must have dead space"); 1376 1377 // The gc number is saved whenever a maximum compaction is done, and used to 1378 // determine when the maximum compaction interval has expired. This avoids 1379 // successive max compactions for different reasons. 1380 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1381 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1382 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1383 total_invocations() == HeapFirstMaximumCompactionCount; 1384 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1385 _maximum_compaction_gc_num = total_invocations(); 1386 return sd.region_to_addr(full_cp); 1387 } 1388 1389 const size_t space_live = pointer_delta(new_top, bottom); 1390 const size_t space_used = space->used_in_words(); 1391 const size_t space_capacity = space->capacity_in_words(); 1392 1393 const double density = double(space_live) / double(space_capacity); 1394 const size_t min_percent_free = MarkSweepDeadRatio; 1395 const double limiter = dead_wood_limiter(density, min_percent_free); 1396 const size_t dead_wood_max = space_used - space_live; 1397 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1398 dead_wood_max); 1399 1400 if (TraceParallelOldGCDensePrefix) { 1401 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1402 "space_cap=" SIZE_FORMAT, 1403 space_live, space_used, 1404 space_capacity); 1405 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1406 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1407 density, min_percent_free, limiter, 1408 dead_wood_max, dead_wood_limit); 1409 } 1410 1411 // Locate the region with the desired amount of dead space to the left. 1412 const RegionData* const limit_cp = 1413 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1414 1415 // Scan from the first region with dead space to the limit region and find the 1416 // one with the best (largest) reclaimed ratio. 1417 double best_ratio = 0.0; 1418 const RegionData* best_cp = full_cp; 1419 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1420 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1421 if (tmp_ratio > best_ratio) { 1422 best_cp = cp; 1423 best_ratio = tmp_ratio; 1424 } 1425 } 1426 1427 return sd.region_to_addr(best_cp); 1428 } 1429 1430 void PSParallelCompact::summarize_spaces_quick() 1431 { 1432 for (unsigned int i = 0; i < last_space_id; ++i) { 1433 const MutableSpace* space = _space_info[i].space(); 1434 HeapWord** nta = _space_info[i].new_top_addr(); 1435 bool result = _summary_data.summarize(_space_info[i].split_info(), 1436 space->bottom(), space->top(), NULL, 1437 space->bottom(), space->end(), nta); 1438 assert(result, "space must fit into itself"); 1439 _space_info[i].set_dense_prefix(space->bottom()); 1440 } 1441 } 1442 1443 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1444 { 1445 HeapWord* const dense_prefix_end = dense_prefix(id); 1446 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1447 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1448 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1449 // Only enough dead space is filled so that any remaining dead space to the 1450 // left is larger than the minimum filler object. (The remainder is filled 1451 // during the copy/update phase.) 1452 // 1453 // The size of the dead space to the right of the boundary is not a 1454 // concern, since compaction will be able to use whatever space is 1455 // available. 1456 // 1457 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1458 // surrounds the space to be filled with an object. 1459 // 1460 // In the 32-bit VM, each bit represents two 32-bit words: 1461 // +---+ 1462 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1463 // end_bits: ... x x x | 0 | || 0 x x ... 1464 // +---+ 1465 // 1466 // In the 64-bit VM, each bit represents one 64-bit word: 1467 // +------------+ 1468 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1469 // end_bits: ... x x 1 | 0 || 0 | x x ... 1470 // +------------+ 1471 // +-------+ 1472 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1473 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1474 // +-------+ 1475 // +-----------+ 1476 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1477 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1478 // +-----------+ 1479 // +-------+ 1480 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1481 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1482 // +-------+ 1483 1484 // Initially assume case a, c or e will apply. 1485 size_t obj_len = CollectedHeap::min_fill_size(); 1486 HeapWord* obj_beg = dense_prefix_end - obj_len; 1487 1488 #ifdef _LP64 1489 if (MinObjAlignment > 1) { // object alignment > heap word size 1490 // Cases a, c or e. 1491 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1492 // Case b above. 1493 obj_beg = dense_prefix_end - 1; 1494 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1495 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1496 // Case d above. 1497 obj_beg = dense_prefix_end - 3; 1498 obj_len = 3; 1499 } 1500 #endif // #ifdef _LP64 1501 1502 CollectedHeap::fill_with_object(obj_beg, obj_len); 1503 _mark_bitmap.mark_obj(obj_beg, obj_len); 1504 _summary_data.add_obj(obj_beg, obj_len); 1505 assert(start_array(id) != NULL, "sanity"); 1506 start_array(id)->allocate_block(obj_beg); 1507 } 1508 } 1509 1510 void 1511 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1512 { 1513 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1514 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1515 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1516 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1517 cur->set_source_region(0); 1518 } 1519 } 1520 1521 void 1522 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1523 { 1524 assert(id < last_space_id, "id out of range"); 1525 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1526 "should have been reset in summarize_spaces_quick()"); 1527 1528 const MutableSpace* space = _space_info[id].space(); 1529 if (_space_info[id].new_top() != space->bottom()) { 1530 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1531 _space_info[id].set_dense_prefix(dense_prefix_end); 1532 1533 #ifndef PRODUCT 1534 if (TraceParallelOldGCDensePrefix) { 1535 print_dense_prefix_stats("ratio", id, maximum_compaction, 1536 dense_prefix_end); 1537 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1538 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1539 } 1540 #endif // #ifndef PRODUCT 1541 1542 // Recompute the summary data, taking into account the dense prefix. If 1543 // every last byte will be reclaimed, then the existing summary data which 1544 // compacts everything can be left in place. 1545 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1546 // If dead space crosses the dense prefix boundary, it is (at least 1547 // partially) filled with a dummy object, marked live and added to the 1548 // summary data. This simplifies the copy/update phase and must be done 1549 // before the final locations of objects are determined, to prevent 1550 // leaving a fragment of dead space that is too small to fill. 1551 fill_dense_prefix_end(id); 1552 1553 // Compute the destination of each Region, and thus each object. 1554 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1555 _summary_data.summarize(_space_info[id].split_info(), 1556 dense_prefix_end, space->top(), NULL, 1557 dense_prefix_end, space->end(), 1558 _space_info[id].new_top_addr()); 1559 } 1560 } 1561 1562 if (TraceParallelOldGCSummaryPhase) { 1563 const size_t region_size = ParallelCompactData::RegionSize; 1564 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1565 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1566 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1567 HeapWord* const new_top = _space_info[id].new_top(); 1568 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1569 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1570 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1571 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1572 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1573 id, space->capacity_in_words(), p2i(dense_prefix_end), 1574 dp_region, dp_words / region_size, 1575 cr_words / region_size, p2i(new_top)); 1576 } 1577 } 1578 1579 #ifndef PRODUCT 1580 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1581 HeapWord* dst_beg, HeapWord* dst_end, 1582 SpaceId src_space_id, 1583 HeapWord* src_beg, HeapWord* src_end) 1584 { 1585 if (TraceParallelOldGCSummaryPhase) { 1586 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1587 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1588 SIZE_FORMAT "-" SIZE_FORMAT " " 1589 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1590 SIZE_FORMAT "-" SIZE_FORMAT, 1591 src_space_id, space_names[src_space_id], 1592 dst_space_id, space_names[dst_space_id], 1593 p2i(src_beg), p2i(src_end), 1594 _summary_data.addr_to_region_idx(src_beg), 1595 _summary_data.addr_to_region_idx(src_end), 1596 p2i(dst_beg), p2i(dst_end), 1597 _summary_data.addr_to_region_idx(dst_beg), 1598 _summary_data.addr_to_region_idx(dst_end)); 1599 } 1600 } 1601 #endif // #ifndef PRODUCT 1602 1603 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1604 bool maximum_compaction) 1605 { 1606 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer); 1607 // trace("2"); 1608 1609 #ifdef ASSERT 1610 if (TraceParallelOldGCMarkingPhase) { 1611 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1612 "add_obj_bytes=" SIZE_FORMAT, 1613 add_obj_count, add_obj_size * HeapWordSize); 1614 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1615 "mark_bitmap_bytes=" SIZE_FORMAT, 1616 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1617 } 1618 #endif // #ifdef ASSERT 1619 1620 // Quick summarization of each space into itself, to see how much is live. 1621 summarize_spaces_quick(); 1622 1623 if (TraceParallelOldGCSummaryPhase) { 1624 tty->print_cr("summary_phase: after summarizing each space to self"); 1625 Universe::print(); 1626 NOT_PRODUCT(print_region_ranges()); 1627 if (Verbose) { 1628 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1629 } 1630 } 1631 1632 // The amount of live data that will end up in old space (assuming it fits). 1633 size_t old_space_total_live = 0; 1634 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1635 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1636 _space_info[id].space()->bottom()); 1637 } 1638 1639 MutableSpace* const old_space = _space_info[old_space_id].space(); 1640 const size_t old_capacity = old_space->capacity_in_words(); 1641 if (old_space_total_live > old_capacity) { 1642 // XXX - should also try to expand 1643 maximum_compaction = true; 1644 } 1645 1646 // Old generations. 1647 summarize_space(old_space_id, maximum_compaction); 1648 1649 // Summarize the remaining spaces in the young gen. The initial target space 1650 // is the old gen. If a space does not fit entirely into the target, then the 1651 // remainder is compacted into the space itself and that space becomes the new 1652 // target. 1653 SpaceId dst_space_id = old_space_id; 1654 HeapWord* dst_space_end = old_space->end(); 1655 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1656 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1657 const MutableSpace* space = _space_info[id].space(); 1658 const size_t live = pointer_delta(_space_info[id].new_top(), 1659 space->bottom()); 1660 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1661 1662 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1663 SpaceId(id), space->bottom(), space->top());) 1664 if (live > 0 && live <= available) { 1665 // All the live data will fit. 1666 bool done = _summary_data.summarize(_space_info[id].split_info(), 1667 space->bottom(), space->top(), 1668 NULL, 1669 *new_top_addr, dst_space_end, 1670 new_top_addr); 1671 assert(done, "space must fit into old gen"); 1672 1673 // Reset the new_top value for the space. 1674 _space_info[id].set_new_top(space->bottom()); 1675 } else if (live > 0) { 1676 // Attempt to fit part of the source space into the target space. 1677 HeapWord* next_src_addr = NULL; 1678 bool done = _summary_data.summarize(_space_info[id].split_info(), 1679 space->bottom(), space->top(), 1680 &next_src_addr, 1681 *new_top_addr, dst_space_end, 1682 new_top_addr); 1683 assert(!done, "space should not fit into old gen"); 1684 assert(next_src_addr != NULL, "sanity"); 1685 1686 // The source space becomes the new target, so the remainder is compacted 1687 // within the space itself. 1688 dst_space_id = SpaceId(id); 1689 dst_space_end = space->end(); 1690 new_top_addr = _space_info[id].new_top_addr(); 1691 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1692 space->bottom(), dst_space_end, 1693 SpaceId(id), next_src_addr, space->top());) 1694 done = _summary_data.summarize(_space_info[id].split_info(), 1695 next_src_addr, space->top(), 1696 NULL, 1697 space->bottom(), dst_space_end, 1698 new_top_addr); 1699 assert(done, "space must fit when compacted into itself"); 1700 assert(*new_top_addr <= space->top(), "usage should not grow"); 1701 } 1702 } 1703 1704 if (TraceParallelOldGCSummaryPhase) { 1705 tty->print_cr("summary_phase: after final summarization"); 1706 Universe::print(); 1707 NOT_PRODUCT(print_region_ranges()); 1708 if (Verbose) { 1709 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1710 } 1711 } 1712 } 1713 1714 // This method should contain all heap-specific policy for invoking a full 1715 // collection. invoke_no_policy() will only attempt to compact the heap; it 1716 // will do nothing further. If we need to bail out for policy reasons, scavenge 1717 // before full gc, or any other specialized behavior, it needs to be added here. 1718 // 1719 // Note that this method should only be called from the vm_thread while at a 1720 // safepoint. 1721 // 1722 // Note that the all_soft_refs_clear flag in the collector policy 1723 // may be true because this method can be called without intervening 1724 // activity. For example when the heap space is tight and full measure 1725 // are being taken to free space. 1726 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1727 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1728 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1729 "should be in vm thread"); 1730 1731 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1732 GCCause::Cause gc_cause = heap->gc_cause(); 1733 assert(!heap->is_gc_active(), "not reentrant"); 1734 1735 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1736 IsGCActiveMark mark; 1737 1738 if (ScavengeBeforeFullGC) { 1739 PSScavenge::invoke_no_policy(); 1740 } 1741 1742 const bool clear_all_soft_refs = 1743 heap->collector_policy()->should_clear_all_soft_refs(); 1744 1745 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1746 maximum_heap_compaction); 1747 } 1748 1749 // This method contains no policy. You should probably 1750 // be calling invoke() instead. 1751 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1752 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1753 assert(ref_processor() != NULL, "Sanity"); 1754 1755 if (GC_locker::check_active_before_gc()) { 1756 return false; 1757 } 1758 1759 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1760 1761 GCIdMark gc_id_mark; 1762 _gc_timer.register_gc_start(); 1763 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1764 1765 TimeStamp marking_start; 1766 TimeStamp compaction_start; 1767 TimeStamp collection_exit; 1768 1769 GCCause::Cause gc_cause = heap->gc_cause(); 1770 PSYoungGen* young_gen = heap->young_gen(); 1771 PSOldGen* old_gen = heap->old_gen(); 1772 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1773 1774 // The scope of casr should end after code that can change 1775 // CollectorPolicy::_should_clear_all_soft_refs. 1776 ClearedAllSoftRefs casr(maximum_heap_compaction, 1777 heap->collector_policy()); 1778 1779 if (ZapUnusedHeapArea) { 1780 // Save information needed to minimize mangling 1781 heap->record_gen_tops_before_GC(); 1782 } 1783 1784 heap->pre_full_gc_dump(&_gc_timer); 1785 1786 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 1787 1788 // Make sure data structures are sane, make the heap parsable, and do other 1789 // miscellaneous bookkeeping. 1790 PreGCValues pre_gc_values; 1791 pre_compact(&pre_gc_values); 1792 1793 // Get the compaction manager reserved for the VM thread. 1794 ParCompactionManager* const vmthread_cm = 1795 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1796 1797 // Place after pre_compact() where the number of invocations is incremented. 1798 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 1799 1800 { 1801 ResourceMark rm; 1802 HandleMark hm; 1803 1804 // Set the number of GC threads to be used in this collection 1805 gc_task_manager()->set_active_gang(); 1806 gc_task_manager()->task_idle_workers(); 1807 1808 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 1809 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 1810 TraceCollectorStats tcs(counters()); 1811 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 1812 1813 if (TraceOldGenTime) accumulated_time()->start(); 1814 1815 // Let the size policy know we're starting 1816 size_policy->major_collection_begin(); 1817 1818 CodeCache::gc_prologue(); 1819 1820 #if defined(COMPILER2) || INCLUDE_JVMCI 1821 DerivedPointerTable::clear(); 1822 #endif 1823 1824 ref_processor()->enable_discovery(); 1825 ref_processor()->setup_policy(maximum_heap_compaction); 1826 1827 bool marked_for_unloading = false; 1828 1829 marking_start.update(); 1830 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1831 1832 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1833 && GCCause::is_user_requested_gc(gc_cause); 1834 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1835 1836 #if defined(COMPILER2) || INCLUDE_JVMCI 1837 assert(DerivedPointerTable::is_active(), "Sanity"); 1838 DerivedPointerTable::set_active(false); 1839 #endif 1840 1841 // adjust_roots() updates Universe::_intArrayKlassObj which is 1842 // needed by the compaction for filling holes in the dense prefix. 1843 adjust_roots(); 1844 1845 compaction_start.update(); 1846 compact(); 1847 1848 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1849 // done before resizing. 1850 post_compact(); 1851 1852 // Let the size policy know we're done 1853 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1854 1855 if (UseAdaptiveSizePolicy) { 1856 if (PrintAdaptiveSizePolicy) { 1857 gclog_or_tty->print("AdaptiveSizeStart: "); 1858 gclog_or_tty->stamp(); 1859 gclog_or_tty->print_cr(" collection: %d ", 1860 heap->total_collections()); 1861 if (Verbose) { 1862 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT 1863 " young_gen_capacity: " SIZE_FORMAT, 1864 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1865 } 1866 } 1867 1868 // Don't check if the size_policy is ready here. Let 1869 // the size_policy check that internally. 1870 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1871 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1872 // Swap the survivor spaces if from_space is empty. The 1873 // resize_young_gen() called below is normally used after 1874 // a successful young GC and swapping of survivor spaces; 1875 // otherwise, it will fail to resize the young gen with 1876 // the current implementation. 1877 if (young_gen->from_space()->is_empty()) { 1878 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1879 young_gen->swap_spaces(); 1880 } 1881 1882 // Calculate optimal free space amounts 1883 assert(young_gen->max_size() > 1884 young_gen->from_space()->capacity_in_bytes() + 1885 young_gen->to_space()->capacity_in_bytes(), 1886 "Sizes of space in young gen are out-of-bounds"); 1887 1888 size_t young_live = young_gen->used_in_bytes(); 1889 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1890 size_t old_live = old_gen->used_in_bytes(); 1891 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1892 size_t max_old_gen_size = old_gen->max_gen_size(); 1893 size_t max_eden_size = young_gen->max_size() - 1894 young_gen->from_space()->capacity_in_bytes() - 1895 young_gen->to_space()->capacity_in_bytes(); 1896 1897 // Used for diagnostics 1898 size_policy->clear_generation_free_space_flags(); 1899 1900 size_policy->compute_generations_free_space(young_live, 1901 eden_live, 1902 old_live, 1903 cur_eden, 1904 max_old_gen_size, 1905 max_eden_size, 1906 true /* full gc*/); 1907 1908 size_policy->check_gc_overhead_limit(young_live, 1909 eden_live, 1910 max_old_gen_size, 1911 max_eden_size, 1912 true /* full gc*/, 1913 gc_cause, 1914 heap->collector_policy()); 1915 1916 size_policy->decay_supplemental_growth(true /* full gc*/); 1917 1918 heap->resize_old_gen( 1919 size_policy->calculated_old_free_size_in_bytes()); 1920 1921 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1922 size_policy->calculated_survivor_size_in_bytes()); 1923 } 1924 if (PrintAdaptiveSizePolicy) { 1925 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 1926 heap->total_collections()); 1927 } 1928 } 1929 1930 if (UsePerfData) { 1931 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1932 counters->update_counters(); 1933 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1934 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1935 } 1936 1937 heap->resize_all_tlabs(); 1938 1939 // Resize the metaspace capacity after a collection 1940 MetaspaceGC::compute_new_size(); 1941 1942 if (TraceOldGenTime) accumulated_time()->stop(); 1943 1944 if (PrintGC) { 1945 if (PrintGCDetails) { 1946 // No GC timestamp here. This is after GC so it would be confusing. 1947 young_gen->print_used_change(pre_gc_values.young_gen_used()); 1948 old_gen->print_used_change(pre_gc_values.old_gen_used()); 1949 heap->print_heap_change(pre_gc_values.heap_used()); 1950 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 1951 } else { 1952 heap->print_heap_change(pre_gc_values.heap_used()); 1953 } 1954 } 1955 1956 // Track memory usage and detect low memory 1957 MemoryService::track_memory_usage(); 1958 heap->update_counters(); 1959 gc_task_manager()->release_idle_workers(); 1960 } 1961 1962 #ifdef ASSERT 1963 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1964 ParCompactionManager* const cm = 1965 ParCompactionManager::manager_array(int(i)); 1966 assert(cm->marking_stack()->is_empty(), "should be empty"); 1967 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 1968 } 1969 #endif // ASSERT 1970 1971 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1972 HandleMark hm; // Discard invalid handles created during verification 1973 Universe::verify(" VerifyAfterGC:"); 1974 } 1975 1976 // Re-verify object start arrays 1977 if (VerifyObjectStartArray && 1978 VerifyAfterGC) { 1979 old_gen->verify_object_start_array(); 1980 } 1981 1982 if (ZapUnusedHeapArea) { 1983 old_gen->object_space()->check_mangled_unused_area_complete(); 1984 } 1985 1986 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1987 1988 collection_exit.update(); 1989 1990 heap->print_heap_after_gc(); 1991 heap->trace_heap_after_gc(&_gc_tracer); 1992 1993 if (PrintGCTaskTimeStamps) { 1994 gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " 1995 JLONG_FORMAT, 1996 marking_start.ticks(), compaction_start.ticks(), 1997 collection_exit.ticks()); 1998 gc_task_manager()->print_task_time_stamps(); 1999 } 2000 2001 heap->post_full_gc_dump(&_gc_timer); 2002 2003 #ifdef TRACESPINNING 2004 ParallelTaskTerminator::print_termination_counts(); 2005 #endif 2006 2007 _gc_timer.register_gc_end(); 2008 2009 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2010 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2011 2012 return true; 2013 } 2014 2015 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2016 PSYoungGen* young_gen, 2017 PSOldGen* old_gen) { 2018 MutableSpace* const eden_space = young_gen->eden_space(); 2019 assert(!eden_space->is_empty(), "eden must be non-empty"); 2020 assert(young_gen->virtual_space()->alignment() == 2021 old_gen->virtual_space()->alignment(), "alignments do not match"); 2022 2023 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2024 return false; 2025 } 2026 2027 // Both generations must be completely committed. 2028 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2029 return false; 2030 } 2031 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2032 return false; 2033 } 2034 2035 // Figure out how much to take from eden. Include the average amount promoted 2036 // in the total; otherwise the next young gen GC will simply bail out to a 2037 // full GC. 2038 const size_t alignment = old_gen->virtual_space()->alignment(); 2039 const size_t eden_used = eden_space->used_in_bytes(); 2040 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2041 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2042 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2043 2044 if (absorb_size >= eden_capacity) { 2045 return false; // Must leave some space in eden. 2046 } 2047 2048 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2049 if (new_young_size < young_gen->min_gen_size()) { 2050 return false; // Respect young gen minimum size. 2051 } 2052 2053 if (TraceAdaptiveGCBoundary && Verbose) { 2054 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2055 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2056 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2057 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2058 absorb_size / K, 2059 eden_capacity / K, (eden_capacity - absorb_size) / K, 2060 young_gen->from_space()->used_in_bytes() / K, 2061 young_gen->to_space()->used_in_bytes() / K, 2062 young_gen->capacity_in_bytes() / K, new_young_size / K); 2063 } 2064 2065 // Fill the unused part of the old gen. 2066 MutableSpace* const old_space = old_gen->object_space(); 2067 HeapWord* const unused_start = old_space->top(); 2068 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2069 2070 if (unused_words > 0) { 2071 if (unused_words < CollectedHeap::min_fill_size()) { 2072 return false; // If the old gen cannot be filled, must give up. 2073 } 2074 CollectedHeap::fill_with_objects(unused_start, unused_words); 2075 } 2076 2077 // Take the live data from eden and set both top and end in the old gen to 2078 // eden top. (Need to set end because reset_after_change() mangles the region 2079 // from end to virtual_space->high() in debug builds). 2080 HeapWord* const new_top = eden_space->top(); 2081 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2082 absorb_size); 2083 young_gen->reset_after_change(); 2084 old_space->set_top(new_top); 2085 old_space->set_end(new_top); 2086 old_gen->reset_after_change(); 2087 2088 // Update the object start array for the filler object and the data from eden. 2089 ObjectStartArray* const start_array = old_gen->start_array(); 2090 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2091 start_array->allocate_block(p); 2092 } 2093 2094 // Could update the promoted average here, but it is not typically updated at 2095 // full GCs and the value to use is unclear. Something like 2096 // 2097 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2098 2099 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2100 return true; 2101 } 2102 2103 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2104 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2105 "shouldn't return NULL"); 2106 return ParallelScavengeHeap::gc_task_manager(); 2107 } 2108 2109 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2110 bool maximum_heap_compaction, 2111 ParallelOldTracer *gc_tracer) { 2112 // Recursively traverse all live objects and mark them 2113 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer); 2114 2115 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2116 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2117 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2118 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2119 ParallelTaskTerminator terminator(active_gc_threads, qset); 2120 2121 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm); 2122 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2123 2124 // Need new claim bits before marking starts. 2125 ClassLoaderDataGraph::clear_claimed_marks(); 2126 2127 { 2128 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer); 2129 2130 ParallelScavengeHeap::ParStrongRootsScope psrs; 2131 2132 GCTaskQueue* q = GCTaskQueue::create(); 2133 2134 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2135 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2136 // We scan the thread roots in parallel 2137 Threads::create_thread_roots_marking_tasks(q); 2138 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2139 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2140 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2141 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2142 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2143 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2144 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2145 2146 if (active_gc_threads > 1) { 2147 for (uint j = 0; j < active_gc_threads; j++) { 2148 q->enqueue(new StealMarkingTask(&terminator)); 2149 } 2150 } 2151 2152 gc_task_manager()->execute_and_wait(q); 2153 } 2154 2155 // Process reference objects found during marking 2156 { 2157 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer); 2158 2159 ReferenceProcessorStats stats; 2160 if (ref_processor()->processing_is_mt()) { 2161 RefProcTaskExecutor task_executor; 2162 stats = ref_processor()->process_discovered_references( 2163 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2164 &task_executor, &_gc_timer); 2165 } else { 2166 stats = ref_processor()->process_discovered_references( 2167 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2168 &_gc_timer); 2169 } 2170 2171 gc_tracer->report_gc_reference_stats(stats); 2172 } 2173 2174 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer); 2175 2176 // This is the point where the entire marking should have completed. 2177 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2178 2179 // Follow system dictionary roots and unload classes. 2180 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2181 2182 // Unload nmethods. 2183 CodeCache::do_unloading(is_alive_closure(), purged_class); 2184 2185 // Prune dead klasses from subklass/sibling/implementor lists. 2186 Klass::clean_weak_klass_links(is_alive_closure()); 2187 2188 // Delete entries for dead interned strings. 2189 StringTable::unlink(is_alive_closure()); 2190 2191 // Clean up unreferenced symbols in symbol table. 2192 SymbolTable::unlink(); 2193 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2194 } 2195 2196 // This should be moved to the shared markSweep code! 2197 class PSAlwaysTrueClosure: public BoolObjectClosure { 2198 public: 2199 bool do_object_b(oop p) { return true; } 2200 }; 2201 static PSAlwaysTrueClosure always_true; 2202 2203 void PSParallelCompact::adjust_roots() { 2204 // Adjust the pointers to reflect the new locations 2205 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer); 2206 2207 // Need new claim bits when tracing through and adjusting pointers. 2208 ClassLoaderDataGraph::clear_claimed_marks(); 2209 2210 // General strong roots. 2211 Universe::oops_do(adjust_pointer_closure()); 2212 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2213 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2214 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2215 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2216 FlatProfiler::oops_do(adjust_pointer_closure()); 2217 Management::oops_do(adjust_pointer_closure()); 2218 JvmtiExport::oops_do(adjust_pointer_closure()); 2219 SystemDictionary::oops_do(adjust_pointer_closure()); 2220 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2221 2222 // Now adjust pointers in remaining weak roots. (All of which should 2223 // have been cleared if they pointed to non-surviving objects.) 2224 // Global (weak) JNI handles 2225 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2226 2227 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2228 CodeCache::blobs_do(&adjust_from_blobs); 2229 StringTable::oops_do(adjust_pointer_closure()); 2230 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2231 // Roots were visited so references into the young gen in roots 2232 // may have been scanned. Process them also. 2233 // Should the reference processor have a span that excludes 2234 // young gen objects? 2235 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2236 } 2237 2238 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2239 uint parallel_gc_threads) 2240 { 2241 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer); 2242 2243 // Find the threads that are active 2244 unsigned int which = 0; 2245 2246 const uint task_count = MAX2(parallel_gc_threads, 1U); 2247 for (uint j = 0; j < task_count; j++) { 2248 q->enqueue(new DrainStacksCompactionTask(j)); 2249 ParCompactionManager::verify_region_list_empty(j); 2250 // Set the region stacks variables to "no" region stack values 2251 // so that they will be recognized and needing a region stack 2252 // in the stealing tasks if they do not get one by executing 2253 // a draining stack. 2254 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2255 cm->set_region_stack(NULL); 2256 cm->set_region_stack_index((uint)max_uintx); 2257 } 2258 ParCompactionManager::reset_recycled_stack_index(); 2259 2260 // Find all regions that are available (can be filled immediately) and 2261 // distribute them to the thread stacks. The iteration is done in reverse 2262 // order (high to low) so the regions will be removed in ascending order. 2263 2264 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2265 2266 size_t fillable_regions = 0; // A count for diagnostic purposes. 2267 // A region index which corresponds to the tasks created above. 2268 // "which" must be 0 <= which < task_count 2269 2270 which = 0; 2271 // id + 1 is used to test termination so unsigned can 2272 // be used with an old_space_id == 0. 2273 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2274 SpaceInfo* const space_info = _space_info + id; 2275 MutableSpace* const space = space_info->space(); 2276 HeapWord* const new_top = space_info->new_top(); 2277 2278 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2279 const size_t end_region = 2280 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2281 2282 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2283 if (sd.region(cur)->claim_unsafe()) { 2284 ParCompactionManager::region_list_push(which, cur); 2285 2286 if (TraceParallelOldGCCompactionPhase && Verbose) { 2287 const size_t count_mod_8 = fillable_regions & 7; 2288 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2289 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2290 if (count_mod_8 == 7) gclog_or_tty->cr(); 2291 } 2292 2293 NOT_PRODUCT(++fillable_regions;) 2294 2295 // Assign regions to tasks in round-robin fashion. 2296 if (++which == task_count) { 2297 assert(which <= parallel_gc_threads, 2298 "Inconsistent number of workers"); 2299 which = 0; 2300 } 2301 } 2302 } 2303 } 2304 2305 if (TraceParallelOldGCCompactionPhase) { 2306 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2307 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions); 2308 } 2309 } 2310 2311 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2312 2313 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2314 uint parallel_gc_threads) { 2315 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer); 2316 2317 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2318 2319 // Iterate over all the spaces adding tasks for updating 2320 // regions in the dense prefix. Assume that 1 gc thread 2321 // will work on opening the gaps and the remaining gc threads 2322 // will work on the dense prefix. 2323 unsigned int space_id; 2324 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2325 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2326 const MutableSpace* const space = _space_info[space_id].space(); 2327 2328 if (dense_prefix_end == space->bottom()) { 2329 // There is no dense prefix for this space. 2330 continue; 2331 } 2332 2333 // The dense prefix is before this region. 2334 size_t region_index_end_dense_prefix = 2335 sd.addr_to_region_idx(dense_prefix_end); 2336 RegionData* const dense_prefix_cp = 2337 sd.region(region_index_end_dense_prefix); 2338 assert(dense_prefix_end == space->end() || 2339 dense_prefix_cp->available() || 2340 dense_prefix_cp->claimed(), 2341 "The region after the dense prefix should always be ready to fill"); 2342 2343 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2344 2345 // Is there dense prefix work? 2346 size_t total_dense_prefix_regions = 2347 region_index_end_dense_prefix - region_index_start; 2348 // How many regions of the dense prefix should be given to 2349 // each thread? 2350 if (total_dense_prefix_regions > 0) { 2351 uint tasks_for_dense_prefix = 1; 2352 if (total_dense_prefix_regions <= 2353 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2354 // Don't over partition. This assumes that 2355 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2356 // so there are not many regions to process. 2357 tasks_for_dense_prefix = parallel_gc_threads; 2358 } else { 2359 // Over partition 2360 tasks_for_dense_prefix = parallel_gc_threads * 2361 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2362 } 2363 size_t regions_per_thread = total_dense_prefix_regions / 2364 tasks_for_dense_prefix; 2365 // Give each thread at least 1 region. 2366 if (regions_per_thread == 0) { 2367 regions_per_thread = 1; 2368 } 2369 2370 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2371 if (region_index_start >= region_index_end_dense_prefix) { 2372 break; 2373 } 2374 // region_index_end is not processed 2375 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2376 region_index_end_dense_prefix); 2377 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2378 region_index_start, 2379 region_index_end)); 2380 region_index_start = region_index_end; 2381 } 2382 } 2383 // This gets any part of the dense prefix that did not 2384 // fit evenly. 2385 if (region_index_start < region_index_end_dense_prefix) { 2386 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2387 region_index_start, 2388 region_index_end_dense_prefix)); 2389 } 2390 } 2391 } 2392 2393 void PSParallelCompact::enqueue_region_stealing_tasks( 2394 GCTaskQueue* q, 2395 ParallelTaskTerminator* terminator_ptr, 2396 uint parallel_gc_threads) { 2397 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer); 2398 2399 // Once a thread has drained it's stack, it should try to steal regions from 2400 // other threads. 2401 if (parallel_gc_threads > 1) { 2402 for (uint j = 0; j < parallel_gc_threads; j++) { 2403 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2404 } 2405 } 2406 } 2407 2408 #ifdef ASSERT 2409 // Write a histogram of the number of times the block table was filled for a 2410 // region. 2411 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2412 { 2413 if (!TraceParallelOldGCCompactionPhase) return; 2414 2415 typedef ParallelCompactData::RegionData rd_t; 2416 ParallelCompactData& sd = summary_data(); 2417 2418 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2419 MutableSpace* const spc = _space_info[id].space(); 2420 if (spc->bottom() != spc->top()) { 2421 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2422 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2423 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2424 2425 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2426 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2427 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2428 2429 for (const rd_t* cur = beg; cur < end; ++cur) { 2430 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2431 } 2432 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2433 for (size_t i = 0; i < histo_len; ++i) { 2434 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2435 histo[i], 100.0 * histo[i] / region_cnt); 2436 } 2437 out->cr(); 2438 } 2439 } 2440 } 2441 #endif // #ifdef ASSERT 2442 2443 void PSParallelCompact::compact() { 2444 // trace("5"); 2445 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer); 2446 2447 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2448 PSOldGen* old_gen = heap->old_gen(); 2449 old_gen->start_array()->reset(); 2450 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2451 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2452 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2453 ParallelTaskTerminator terminator(active_gc_threads, qset); 2454 2455 GCTaskQueue* q = GCTaskQueue::create(); 2456 enqueue_region_draining_tasks(q, active_gc_threads); 2457 enqueue_dense_prefix_tasks(q, active_gc_threads); 2458 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2459 2460 { 2461 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer); 2462 2463 gc_task_manager()->execute_and_wait(q); 2464 2465 #ifdef ASSERT 2466 // Verify that all regions have been processed before the deferred updates. 2467 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2468 verify_complete(SpaceId(id)); 2469 } 2470 #endif 2471 } 2472 2473 { 2474 // Update the deferred objects, if any. Any compaction manager can be used. 2475 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer); 2476 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2477 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2478 update_deferred_objects(cm, SpaceId(id)); 2479 } 2480 } 2481 2482 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2483 } 2484 2485 #ifdef ASSERT 2486 void PSParallelCompact::verify_complete(SpaceId space_id) { 2487 // All Regions between space bottom() to new_top() should be marked as filled 2488 // and all Regions between new_top() and top() should be available (i.e., 2489 // should have been emptied). 2490 ParallelCompactData& sd = summary_data(); 2491 SpaceInfo si = _space_info[space_id]; 2492 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2493 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2494 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2495 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2496 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2497 2498 bool issued_a_warning = false; 2499 2500 size_t cur_region; 2501 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2502 const RegionData* const c = sd.region(cur_region); 2503 if (!c->completed()) { 2504 warning("region " SIZE_FORMAT " not filled: " 2505 "destination_count=%u", 2506 cur_region, c->destination_count()); 2507 issued_a_warning = true; 2508 } 2509 } 2510 2511 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2512 const RegionData* const c = sd.region(cur_region); 2513 if (!c->available()) { 2514 warning("region " SIZE_FORMAT " not empty: " 2515 "destination_count=%u", 2516 cur_region, c->destination_count()); 2517 issued_a_warning = true; 2518 } 2519 } 2520 2521 if (issued_a_warning) { 2522 print_region_ranges(); 2523 } 2524 } 2525 #endif // #ifdef ASSERT 2526 2527 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2528 _start_array->allocate_block(addr); 2529 compaction_manager()->update_contents(oop(addr)); 2530 } 2531 2532 // Update interior oops in the ranges of regions [beg_region, end_region). 2533 void 2534 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2535 SpaceId space_id, 2536 size_t beg_region, 2537 size_t end_region) { 2538 ParallelCompactData& sd = summary_data(); 2539 ParMarkBitMap* const mbm = mark_bitmap(); 2540 2541 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2542 HeapWord* const end_addr = sd.region_to_addr(end_region); 2543 assert(beg_region <= end_region, "bad region range"); 2544 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2545 2546 #ifdef ASSERT 2547 // Claim the regions to avoid triggering an assert when they are marked as 2548 // filled. 2549 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2550 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2551 } 2552 #endif // #ifdef ASSERT 2553 2554 if (beg_addr != space(space_id)->bottom()) { 2555 // Find the first live object or block of dead space that *starts* in this 2556 // range of regions. If a partial object crosses onto the region, skip it; 2557 // it will be marked for 'deferred update' when the object head is 2558 // processed. If dead space crosses onto the region, it is also skipped; it 2559 // will be filled when the prior region is processed. If neither of those 2560 // apply, the first word in the region is the start of a live object or dead 2561 // space. 2562 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2563 const RegionData* const cp = sd.region(beg_region); 2564 if (cp->partial_obj_size() != 0) { 2565 beg_addr = sd.partial_obj_end(beg_region); 2566 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2567 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2568 } 2569 } 2570 2571 if (beg_addr < end_addr) { 2572 // A live object or block of dead space starts in this range of Regions. 2573 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2574 2575 // Create closures and iterate. 2576 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2577 FillClosure fill_closure(cm, space_id); 2578 ParMarkBitMap::IterationStatus status; 2579 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2580 dense_prefix_end); 2581 if (status == ParMarkBitMap::incomplete) { 2582 update_closure.do_addr(update_closure.source()); 2583 } 2584 } 2585 2586 // Mark the regions as filled. 2587 RegionData* const beg_cp = sd.region(beg_region); 2588 RegionData* const end_cp = sd.region(end_region); 2589 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2590 cp->set_completed(); 2591 } 2592 } 2593 2594 // Return the SpaceId for the space containing addr. If addr is not in the 2595 // heap, last_space_id is returned. In debug mode it expects the address to be 2596 // in the heap and asserts such. 2597 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2598 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2599 2600 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2601 if (_space_info[id].space()->contains(addr)) { 2602 return SpaceId(id); 2603 } 2604 } 2605 2606 assert(false, "no space contains the addr"); 2607 return last_space_id; 2608 } 2609 2610 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2611 SpaceId id) { 2612 assert(id < last_space_id, "bad space id"); 2613 2614 ParallelCompactData& sd = summary_data(); 2615 const SpaceInfo* const space_info = _space_info + id; 2616 ObjectStartArray* const start_array = space_info->start_array(); 2617 2618 const MutableSpace* const space = space_info->space(); 2619 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2620 HeapWord* const beg_addr = space_info->dense_prefix(); 2621 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2622 2623 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2624 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2625 const RegionData* cur_region; 2626 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2627 HeapWord* const addr = cur_region->deferred_obj_addr(); 2628 if (addr != NULL) { 2629 if (start_array != NULL) { 2630 start_array->allocate_block(addr); 2631 } 2632 cm->update_contents(oop(addr)); 2633 assert(oop(addr)->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2634 } 2635 } 2636 } 2637 2638 // Skip over count live words starting from beg, and return the address of the 2639 // next live word. Unless marked, the word corresponding to beg is assumed to 2640 // be dead. Callers must either ensure beg does not correspond to the middle of 2641 // an object, or account for those live words in some other way. Callers must 2642 // also ensure that there are enough live words in the range [beg, end) to skip. 2643 HeapWord* 2644 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2645 { 2646 assert(count > 0, "sanity"); 2647 2648 ParMarkBitMap* m = mark_bitmap(); 2649 idx_t bits_to_skip = m->words_to_bits(count); 2650 idx_t cur_beg = m->addr_to_bit(beg); 2651 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2652 2653 do { 2654 cur_beg = m->find_obj_beg(cur_beg, search_end); 2655 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2656 const size_t obj_bits = cur_end - cur_beg + 1; 2657 if (obj_bits > bits_to_skip) { 2658 return m->bit_to_addr(cur_beg + bits_to_skip); 2659 } 2660 bits_to_skip -= obj_bits; 2661 cur_beg = cur_end + 1; 2662 } while (bits_to_skip > 0); 2663 2664 // Skipping the desired number of words landed just past the end of an object. 2665 // Find the start of the next object. 2666 cur_beg = m->find_obj_beg(cur_beg, search_end); 2667 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2668 return m->bit_to_addr(cur_beg); 2669 } 2670 2671 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2672 SpaceId src_space_id, 2673 size_t src_region_idx) 2674 { 2675 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2676 2677 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2678 if (split_info.dest_region_addr() == dest_addr) { 2679 // The partial object ending at the split point contains the first word to 2680 // be copied to dest_addr. 2681 return split_info.first_src_addr(); 2682 } 2683 2684 const ParallelCompactData& sd = summary_data(); 2685 ParMarkBitMap* const bitmap = mark_bitmap(); 2686 const size_t RegionSize = ParallelCompactData::RegionSize; 2687 2688 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2689 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2690 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2691 HeapWord* const src_region_destination = src_region_ptr->destination(); 2692 2693 assert(dest_addr >= src_region_destination, "wrong src region"); 2694 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2695 2696 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2697 HeapWord* const src_region_end = src_region_beg + RegionSize; 2698 2699 HeapWord* addr = src_region_beg; 2700 if (dest_addr == src_region_destination) { 2701 // Return the first live word in the source region. 2702 if (partial_obj_size == 0) { 2703 addr = bitmap->find_obj_beg(addr, src_region_end); 2704 assert(addr < src_region_end, "no objects start in src region"); 2705 } 2706 return addr; 2707 } 2708 2709 // Must skip some live data. 2710 size_t words_to_skip = dest_addr - src_region_destination; 2711 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2712 2713 if (partial_obj_size >= words_to_skip) { 2714 // All the live words to skip are part of the partial object. 2715 addr += words_to_skip; 2716 if (partial_obj_size == words_to_skip) { 2717 // Find the first live word past the partial object. 2718 addr = bitmap->find_obj_beg(addr, src_region_end); 2719 assert(addr < src_region_end, "wrong src region"); 2720 } 2721 return addr; 2722 } 2723 2724 // Skip over the partial object (if any). 2725 if (partial_obj_size != 0) { 2726 words_to_skip -= partial_obj_size; 2727 addr += partial_obj_size; 2728 } 2729 2730 // Skip over live words due to objects that start in the region. 2731 addr = skip_live_words(addr, src_region_end, words_to_skip); 2732 assert(addr < src_region_end, "wrong src region"); 2733 return addr; 2734 } 2735 2736 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2737 SpaceId src_space_id, 2738 size_t beg_region, 2739 HeapWord* end_addr) 2740 { 2741 ParallelCompactData& sd = summary_data(); 2742 2743 #ifdef ASSERT 2744 MutableSpace* const src_space = _space_info[src_space_id].space(); 2745 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2746 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2747 "src_space_id does not match beg_addr"); 2748 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2749 "src_space_id does not match end_addr"); 2750 #endif // #ifdef ASSERT 2751 2752 RegionData* const beg = sd.region(beg_region); 2753 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2754 2755 // Regions up to new_top() are enqueued if they become available. 2756 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2757 RegionData* const enqueue_end = 2758 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2759 2760 for (RegionData* cur = beg; cur < end; ++cur) { 2761 assert(cur->data_size() > 0, "region must have live data"); 2762 cur->decrement_destination_count(); 2763 if (cur < enqueue_end && cur->available() && cur->claim()) { 2764 cm->push_region(sd.region(cur)); 2765 } 2766 } 2767 } 2768 2769 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2770 SpaceId& src_space_id, 2771 HeapWord*& src_space_top, 2772 HeapWord* end_addr) 2773 { 2774 typedef ParallelCompactData::RegionData RegionData; 2775 2776 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2777 const size_t region_size = ParallelCompactData::RegionSize; 2778 2779 size_t src_region_idx = 0; 2780 2781 // Skip empty regions (if any) up to the top of the space. 2782 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2783 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2784 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2785 const RegionData* const top_region_ptr = 2786 sd.addr_to_region_ptr(top_aligned_up); 2787 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2788 ++src_region_ptr; 2789 } 2790 2791 if (src_region_ptr < top_region_ptr) { 2792 // The next source region is in the current space. Update src_region_idx 2793 // and the source address to match src_region_ptr. 2794 src_region_idx = sd.region(src_region_ptr); 2795 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2796 if (src_region_addr > closure.source()) { 2797 closure.set_source(src_region_addr); 2798 } 2799 return src_region_idx; 2800 } 2801 2802 // Switch to a new source space and find the first non-empty region. 2803 unsigned int space_id = src_space_id + 1; 2804 assert(space_id < last_space_id, "not enough spaces"); 2805 2806 HeapWord* const destination = closure.destination(); 2807 2808 do { 2809 MutableSpace* space = _space_info[space_id].space(); 2810 HeapWord* const bottom = space->bottom(); 2811 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2812 2813 // Iterate over the spaces that do not compact into themselves. 2814 if (bottom_cp->destination() != bottom) { 2815 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2816 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2817 2818 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2819 if (src_cp->live_obj_size() > 0) { 2820 // Found it. 2821 assert(src_cp->destination() == destination, 2822 "first live obj in the space must match the destination"); 2823 assert(src_cp->partial_obj_size() == 0, 2824 "a space cannot begin with a partial obj"); 2825 2826 src_space_id = SpaceId(space_id); 2827 src_space_top = space->top(); 2828 const size_t src_region_idx = sd.region(src_cp); 2829 closure.set_source(sd.region_to_addr(src_region_idx)); 2830 return src_region_idx; 2831 } else { 2832 assert(src_cp->data_size() == 0, "sanity"); 2833 } 2834 } 2835 } 2836 } while (++space_id < last_space_id); 2837 2838 assert(false, "no source region was found"); 2839 return 0; 2840 } 2841 2842 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 2843 { 2844 typedef ParMarkBitMap::IterationStatus IterationStatus; 2845 const size_t RegionSize = ParallelCompactData::RegionSize; 2846 ParMarkBitMap* const bitmap = mark_bitmap(); 2847 ParallelCompactData& sd = summary_data(); 2848 RegionData* const region_ptr = sd.region(region_idx); 2849 2850 // Get the items needed to construct the closure. 2851 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2852 SpaceId dest_space_id = space_id(dest_addr); 2853 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 2854 HeapWord* new_top = _space_info[dest_space_id].new_top(); 2855 assert(dest_addr < new_top, "sanity"); 2856 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 2857 2858 // Get the source region and related info. 2859 size_t src_region_idx = region_ptr->source_region(); 2860 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2861 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2862 2863 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2864 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2865 2866 // Adjust src_region_idx to prepare for decrementing destination counts (the 2867 // destination count is not decremented when a region is copied to itself). 2868 if (src_region_idx == region_idx) { 2869 src_region_idx += 1; 2870 } 2871 2872 if (bitmap->is_unmarked(closure.source())) { 2873 // The first source word is in the middle of an object; copy the remainder 2874 // of the object or as much as will fit. The fact that pointer updates were 2875 // deferred will be noted when the object header is processed. 2876 HeapWord* const old_src_addr = closure.source(); 2877 closure.copy_partial_obj(); 2878 if (closure.is_full()) { 2879 decrement_destination_counts(cm, src_space_id, src_region_idx, 2880 closure.source()); 2881 region_ptr->set_deferred_obj_addr(NULL); 2882 region_ptr->set_completed(); 2883 return; 2884 } 2885 2886 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2887 if (sd.region_align_down(old_src_addr) != end_addr) { 2888 // The partial object was copied from more than one source region. 2889 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2890 2891 // Move to the next source region, possibly switching spaces as well. All 2892 // args except end_addr may be modified. 2893 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2894 end_addr); 2895 } 2896 } 2897 2898 do { 2899 HeapWord* const cur_addr = closure.source(); 2900 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2901 src_space_top); 2902 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2903 2904 if (status == ParMarkBitMap::incomplete) { 2905 // The last obj that starts in the source region does not end in the 2906 // region. 2907 assert(closure.source() < end_addr, "sanity"); 2908 HeapWord* const obj_beg = closure.source(); 2909 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 2910 src_space_top); 2911 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 2912 if (obj_end < range_end) { 2913 // The end was found; the entire object will fit. 2914 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 2915 assert(status != ParMarkBitMap::would_overflow, "sanity"); 2916 } else { 2917 // The end was not found; the object will not fit. 2918 assert(range_end < src_space_top, "obj cannot cross space boundary"); 2919 status = ParMarkBitMap::would_overflow; 2920 } 2921 } 2922 2923 if (status == ParMarkBitMap::would_overflow) { 2924 // The last object did not fit. Note that interior oop updates were 2925 // deferred, then copy enough of the object to fill the region. 2926 region_ptr->set_deferred_obj_addr(closure.destination()); 2927 status = closure.copy_until_full(); // copies from closure.source() 2928 2929 decrement_destination_counts(cm, src_space_id, src_region_idx, 2930 closure.source()); 2931 region_ptr->set_completed(); 2932 return; 2933 } 2934 2935 if (status == ParMarkBitMap::full) { 2936 decrement_destination_counts(cm, src_space_id, src_region_idx, 2937 closure.source()); 2938 region_ptr->set_deferred_obj_addr(NULL); 2939 region_ptr->set_completed(); 2940 return; 2941 } 2942 2943 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2944 2945 // Move to the next source region, possibly switching spaces as well. All 2946 // args except end_addr may be modified. 2947 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2948 end_addr); 2949 } while (true); 2950 } 2951 2952 void PSParallelCompact::fill_blocks(size_t region_idx) 2953 { 2954 // Fill in the block table elements for the specified region. Each block 2955 // table element holds the number of live words in the region that are to the 2956 // left of the first object that starts in the block. Thus only blocks in 2957 // which an object starts need to be filled. 2958 // 2959 // The algorithm scans the section of the bitmap that corresponds to the 2960 // region, keeping a running total of the live words. When an object start is 2961 // found, if it's the first to start in the block that contains it, the 2962 // current total is written to the block table element. 2963 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2964 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2965 const size_t RegionSize = ParallelCompactData::RegionSize; 2966 2967 ParallelCompactData& sd = summary_data(); 2968 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2969 if (partial_obj_size >= RegionSize) { 2970 return; // No objects start in this region. 2971 } 2972 2973 // Ensure the first loop iteration decides that the block has changed. 2974 size_t cur_block = sd.block_count(); 2975 2976 const ParMarkBitMap* const bitmap = mark_bitmap(); 2977 2978 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2979 assert((size_t)1 << Log2BitsPerBlock == 2980 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2981 2982 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2983 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2984 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2985 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2986 while (beg_bit < range_end) { 2987 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2988 if (new_block != cur_block) { 2989 cur_block = new_block; 2990 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2991 } 2992 2993 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2994 if (end_bit < range_end - 1) { 2995 live_bits += end_bit - beg_bit + 1; 2996 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2997 } else { 2998 return; 2999 } 3000 } 3001 } 3002 3003 void 3004 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3005 const MutableSpace* sp = space(space_id); 3006 if (sp->is_empty()) { 3007 return; 3008 } 3009 3010 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3011 ParMarkBitMap* const bitmap = mark_bitmap(); 3012 HeapWord* const dp_addr = dense_prefix(space_id); 3013 HeapWord* beg_addr = sp->bottom(); 3014 HeapWord* end_addr = sp->top(); 3015 3016 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3017 3018 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3019 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3020 if (beg_region < dp_region) { 3021 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3022 } 3023 3024 // The destination of the first live object that starts in the region is one 3025 // past the end of the partial object entering the region (if any). 3026 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3027 HeapWord* const new_top = _space_info[space_id].new_top(); 3028 assert(new_top >= dest_addr, "bad new_top value"); 3029 const size_t words = pointer_delta(new_top, dest_addr); 3030 3031 if (words > 0) { 3032 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3033 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3034 3035 ParMarkBitMap::IterationStatus status; 3036 status = bitmap->iterate(&closure, dest_addr, end_addr); 3037 assert(status == ParMarkBitMap::full, "iteration not complete"); 3038 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3039 "live objects skipped because closure is full"); 3040 } 3041 } 3042 3043 jlong PSParallelCompact::millis_since_last_gc() { 3044 // We need a monotonically non-decreasing time in ms but 3045 // os::javaTimeMillis() does not guarantee monotonicity. 3046 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3047 jlong ret_val = now - _time_of_last_gc; 3048 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3049 if (ret_val < 0) { 3050 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);) 3051 return 0; 3052 } 3053 return ret_val; 3054 } 3055 3056 void PSParallelCompact::reset_millis_since_last_gc() { 3057 // We need a monotonically non-decreasing time in ms but 3058 // os::javaTimeMillis() does not guarantee monotonicity. 3059 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3060 } 3061 3062 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3063 { 3064 if (source() != destination()) { 3065 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3066 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3067 } 3068 update_state(words_remaining()); 3069 assert(is_full(), "sanity"); 3070 return ParMarkBitMap::full; 3071 } 3072 3073 void MoveAndUpdateClosure::copy_partial_obj() 3074 { 3075 size_t words = words_remaining(); 3076 3077 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3078 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3079 if (end_addr < range_end) { 3080 words = bitmap()->obj_size(source(), end_addr); 3081 } 3082 3083 // This test is necessary; if omitted, the pointer updates to a partial object 3084 // that crosses the dense prefix boundary could be overwritten. 3085 if (source() != destination()) { 3086 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3087 Copy::aligned_conjoint_words(source(), destination(), words); 3088 } 3089 update_state(words); 3090 } 3091 3092 void InstanceKlass::oop_pc_update_pointers(oop obj) { 3093 oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3094 } 3095 3096 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) { 3097 InstanceKlass::oop_pc_update_pointers(obj); 3098 3099 oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3100 } 3101 3102 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) { 3103 InstanceKlass::oop_pc_update_pointers(obj); 3104 } 3105 3106 #ifdef ASSERT 3107 template <class T> static void trace_reference_gc(const char *s, oop obj, 3108 bool is_ephemeron, 3109 T* referent_addr, 3110 T* next_addr, 3111 T* discovered_addr, 3112 T* value_addr) { 3113 if(TraceReferenceGC && PrintGCDetails) { 3114 gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj)); 3115 gclog_or_tty->print_cr(" referent_addr/* " PTR_FORMAT " / " 3116 PTR_FORMAT, p2i(referent_addr), 3117 referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3118 gclog_or_tty->print_cr(" next_addr/* " PTR_FORMAT " / " 3119 PTR_FORMAT, p2i(next_addr), 3120 next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3121 gclog_or_tty->print_cr(" discovered_addr/* " PTR_FORMAT " / " 3122 PTR_FORMAT, p2i(discovered_addr), 3123 discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3124 if (is_ephemeron) { 3125 gclog_or_tty->print_cr(" value_addr/* " PTR_FORMAT " / " 3126 PTR_FORMAT, p2i(value_addr), 3127 value_addr ? p2i(oopDesc::load_decode_heap_oop(value_addr)) : NULL); 3128 } 3129 } 3130 } 3131 #endif 3132 3133 template <class T> 3134 static void oop_pc_update_pointers_specialized(oop obj, InstanceRefKlass* klass) { 3135 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3136 PSParallelCompact::adjust_pointer(referent_addr); 3137 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3138 PSParallelCompact::adjust_pointer(next_addr); 3139 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3140 PSParallelCompact::adjust_pointer(discovered_addr); 3141 if (klass->reference_type() == REF_EPHEMERON) { 3142 T* value_addr = (T*)java_lang_ref_Ephemeron::value_addr(obj); 3143 PSParallelCompact::adjust_pointer(value_addr); 3144 debug_only(trace_reference_gc("InstanceRefKlass<Ephemeron>::oop_update_ptrs", obj, true, 3145 referent_addr, next_addr, discovered_addr, value_addr);) 3146 } else { 3147 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, false, 3148 referent_addr, next_addr, discovered_addr, (T*)NULL);) 3149 } 3150 } 3151 3152 void InstanceRefKlass::oop_pc_update_pointers(oop obj) { 3153 InstanceKlass::oop_pc_update_pointers(obj); 3154 3155 if (UseCompressedOops) { 3156 oop_pc_update_pointers_specialized<narrowOop>(obj, this); 3157 } else { 3158 oop_pc_update_pointers_specialized<oop>(obj, this); 3159 } 3160 } 3161 3162 void ObjArrayKlass::oop_pc_update_pointers(oop obj) { 3163 assert(obj->is_objArray(), "obj must be obj array"); 3164 oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure()); 3165 } 3166 3167 void TypeArrayKlass::oop_pc_update_pointers(oop obj) { 3168 assert(obj->is_typeArray(),"must be a type array"); 3169 } 3170 3171 ParMarkBitMapClosure::IterationStatus 3172 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3173 assert(destination() != NULL, "sanity"); 3174 assert(bitmap()->obj_size(addr) == words, "bad size"); 3175 3176 _source = addr; 3177 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3178 destination(), "wrong destination"); 3179 3180 if (words > words_remaining()) { 3181 return ParMarkBitMap::would_overflow; 3182 } 3183 3184 // The start_array must be updated even if the object is not moving. 3185 if (_start_array != NULL) { 3186 _start_array->allocate_block(destination()); 3187 } 3188 3189 if (destination() != source()) { 3190 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3191 Copy::aligned_conjoint_words(source(), destination(), words); 3192 } 3193 3194 oop moved_oop = (oop) destination(); 3195 compaction_manager()->update_contents(moved_oop); 3196 assert(moved_oop->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3197 3198 update_state(words); 3199 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3200 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3201 } 3202 3203 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3204 ParCompactionManager* cm, 3205 PSParallelCompact::SpaceId space_id) : 3206 ParMarkBitMapClosure(mbm, cm), 3207 _space_id(space_id), 3208 _start_array(PSParallelCompact::start_array(space_id)) 3209 { 3210 } 3211 3212 // Updates the references in the object to their new values. 3213 ParMarkBitMapClosure::IterationStatus 3214 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3215 do_addr(addr); 3216 return ParMarkBitMap::incomplete; 3217 }