1 /* 2 * Copyright (c) 2005, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/stringTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp" 30 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp" 31 #include "gc_implementation/parallelScavenge/pcTasks.hpp" 32 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp" 33 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp" 34 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp" 35 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp" 36 #include "gc_implementation/parallelScavenge/psOldGen.hpp" 37 #include "gc_implementation/parallelScavenge/psParallelCompact.inline.hpp" 38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 39 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 41 #include "gc_implementation/shared/gcHeapSummary.hpp" 42 #include "gc_implementation/shared/gcTimer.hpp" 43 #include "gc_implementation/shared/gcTrace.hpp" 44 #include "gc_implementation/shared/gcTraceTime.hpp" 45 #include "gc_implementation/shared/isGCActiveMark.hpp" 46 #include "gc_implementation/shared/spaceDecorator.hpp" 47 #include "gc_interface/gcCause.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/referencePolicy.hpp" 50 #include "memory/referenceProcessor.hpp" 51 #include "oops/instanceKlass.inline.hpp" 52 #include "oops/instanceMirrorKlass.inline.hpp" 53 #include "oops/methodData.hpp" 54 #include "oops/objArrayKlass.inline.hpp" 55 #include "oops/oop.inline.hpp" 56 #include "runtime/atomic.inline.hpp" 57 #include "runtime/fprofiler.hpp" 58 #include "runtime/safepoint.hpp" 59 #include "runtime/vmThread.hpp" 60 #include "services/management.hpp" 61 #include "services/memoryService.hpp" 62 #include "services/memTracker.hpp" 63 #include "utilities/events.hpp" 64 #include "utilities/stack.inline.hpp" 65 66 #include <math.h> 67 68 // All sizes are in HeapWords. 69 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 70 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 71 const size_t ParallelCompactData::RegionSizeBytes = 72 RegionSize << LogHeapWordSize; 73 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 74 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 75 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 76 77 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 78 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 79 const size_t ParallelCompactData::BlockSizeBytes = 80 BlockSize << LogHeapWordSize; 81 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 82 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 83 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 84 85 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 86 const size_t ParallelCompactData::Log2BlocksPerRegion = 87 Log2RegionSize - Log2BlockSize; 88 89 const ParallelCompactData::RegionData::region_sz_t 90 ParallelCompactData::RegionData::dc_shift = 27; 91 92 const ParallelCompactData::RegionData::region_sz_t 93 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 94 95 const ParallelCompactData::RegionData::region_sz_t 96 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 97 98 const ParallelCompactData::RegionData::region_sz_t 99 ParallelCompactData::RegionData::los_mask = ~dc_mask; 100 101 const ParallelCompactData::RegionData::region_sz_t 102 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 103 104 const ParallelCompactData::RegionData::region_sz_t 105 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 106 107 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 108 bool PSParallelCompact::_print_phases = false; 109 110 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 111 112 double PSParallelCompact::_dwl_mean; 113 double PSParallelCompact::_dwl_std_dev; 114 double PSParallelCompact::_dwl_first_term; 115 double PSParallelCompact::_dwl_adjustment; 116 #ifdef ASSERT 117 bool PSParallelCompact::_dwl_initialized = false; 118 #endif // #ifdef ASSERT 119 120 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 121 HeapWord* destination) 122 { 123 assert(src_region_idx != 0, "invalid src_region_idx"); 124 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 125 assert(destination != NULL, "invalid destination argument"); 126 127 _src_region_idx = src_region_idx; 128 _partial_obj_size = partial_obj_size; 129 _destination = destination; 130 131 // These fields may not be updated below, so make sure they're clear. 132 assert(_dest_region_addr == NULL, "should have been cleared"); 133 assert(_first_src_addr == NULL, "should have been cleared"); 134 135 // Determine the number of destination regions for the partial object. 136 HeapWord* const last_word = destination + partial_obj_size - 1; 137 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 138 HeapWord* const beg_region_addr = sd.region_align_down(destination); 139 HeapWord* const end_region_addr = sd.region_align_down(last_word); 140 141 if (beg_region_addr == end_region_addr) { 142 // One destination region. 143 _destination_count = 1; 144 if (end_region_addr == destination) { 145 // The destination falls on a region boundary, thus the first word of the 146 // partial object will be the first word copied to the destination region. 147 _dest_region_addr = end_region_addr; 148 _first_src_addr = sd.region_to_addr(src_region_idx); 149 } 150 } else { 151 // Two destination regions. When copied, the partial object will cross a 152 // destination region boundary, so a word somewhere within the partial 153 // object will be the first word copied to the second destination region. 154 _destination_count = 2; 155 _dest_region_addr = end_region_addr; 156 const size_t ofs = pointer_delta(end_region_addr, destination); 157 assert(ofs < _partial_obj_size, "sanity"); 158 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 159 } 160 } 161 162 void SplitInfo::clear() 163 { 164 _src_region_idx = 0; 165 _partial_obj_size = 0; 166 _destination = NULL; 167 _destination_count = 0; 168 _dest_region_addr = NULL; 169 _first_src_addr = NULL; 170 assert(!is_valid(), "sanity"); 171 } 172 173 #ifdef ASSERT 174 void SplitInfo::verify_clear() 175 { 176 assert(_src_region_idx == 0, "not clear"); 177 assert(_partial_obj_size == 0, "not clear"); 178 assert(_destination == NULL, "not clear"); 179 assert(_destination_count == 0, "not clear"); 180 assert(_dest_region_addr == NULL, "not clear"); 181 assert(_first_src_addr == NULL, "not clear"); 182 } 183 #endif // #ifdef ASSERT 184 185 186 void PSParallelCompact::print_on_error(outputStream* st) { 187 _mark_bitmap.print_on_error(st); 188 } 189 190 #ifndef PRODUCT 191 const char* PSParallelCompact::space_names[] = { 192 "old ", "eden", "from", "to " 193 }; 194 195 void PSParallelCompact::print_region_ranges() 196 { 197 tty->print_cr("space bottom top end new_top"); 198 tty->print_cr("------ ---------- ---------- ---------- ----------"); 199 200 for (unsigned int id = 0; id < last_space_id; ++id) { 201 const MutableSpace* space = _space_info[id].space(); 202 tty->print_cr("%u %s " 203 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 205 id, space_names[id], 206 summary_data().addr_to_region_idx(space->bottom()), 207 summary_data().addr_to_region_idx(space->top()), 208 summary_data().addr_to_region_idx(space->end()), 209 summary_data().addr_to_region_idx(_space_info[id].new_top())); 210 } 211 } 212 213 void 214 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 215 { 216 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 217 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 218 219 ParallelCompactData& sd = PSParallelCompact::summary_data(); 220 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 221 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 222 REGION_IDX_FORMAT " " PTR_FORMAT " " 223 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 224 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 225 i, p2i(c->data_location()), dci, p2i(c->destination()), 226 c->partial_obj_size(), c->live_obj_size(), 227 c->data_size(), c->source_region(), c->destination_count()); 228 229 #undef REGION_IDX_FORMAT 230 #undef REGION_DATA_FORMAT 231 } 232 233 void 234 print_generic_summary_data(ParallelCompactData& summary_data, 235 HeapWord* const beg_addr, 236 HeapWord* const end_addr) 237 { 238 size_t total_words = 0; 239 size_t i = summary_data.addr_to_region_idx(beg_addr); 240 const size_t last = summary_data.addr_to_region_idx(end_addr); 241 HeapWord* pdest = 0; 242 243 while (i <= last) { 244 ParallelCompactData::RegionData* c = summary_data.region(i); 245 if (c->data_size() != 0 || c->destination() != pdest) { 246 print_generic_summary_region(i, c); 247 total_words += c->data_size(); 248 pdest = c->destination(); 249 } 250 ++i; 251 } 252 253 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 254 } 255 256 void 257 print_generic_summary_data(ParallelCompactData& summary_data, 258 SpaceInfo* space_info) 259 { 260 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 261 const MutableSpace* space = space_info[id].space(); 262 print_generic_summary_data(summary_data, space->bottom(), 263 MAX2(space->top(), space_info[id].new_top())); 264 } 265 } 266 267 void 268 print_initial_summary_region(size_t i, 269 const ParallelCompactData::RegionData* c, 270 bool newline = true) 271 { 272 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 273 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 275 i, p2i(c->destination()), 276 c->partial_obj_size(), c->live_obj_size(), 277 c->data_size(), c->source_region(), c->destination_count()); 278 if (newline) tty->cr(); 279 } 280 281 void 282 print_initial_summary_data(ParallelCompactData& summary_data, 283 const MutableSpace* space) { 284 if (space->top() == space->bottom()) { 285 return; 286 } 287 288 const size_t region_size = ParallelCompactData::RegionSize; 289 typedef ParallelCompactData::RegionData RegionData; 290 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 291 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 292 const RegionData* c = summary_data.region(end_region - 1); 293 HeapWord* end_addr = c->destination() + c->data_size(); 294 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 295 296 // Print (and count) the full regions at the beginning of the space. 297 size_t full_region_count = 0; 298 size_t i = summary_data.addr_to_region_idx(space->bottom()); 299 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 300 print_initial_summary_region(i, summary_data.region(i)); 301 ++full_region_count; 302 ++i; 303 } 304 305 size_t live_to_right = live_in_space - full_region_count * region_size; 306 307 double max_reclaimed_ratio = 0.0; 308 size_t max_reclaimed_ratio_region = 0; 309 size_t max_dead_to_right = 0; 310 size_t max_live_to_right = 0; 311 312 // Print the 'reclaimed ratio' for regions while there is something live in 313 // the region or to the right of it. The remaining regions are empty (and 314 // uninteresting), and computing the ratio will result in division by 0. 315 while (i < end_region && live_to_right > 0) { 316 c = summary_data.region(i); 317 HeapWord* const region_addr = summary_data.region_to_addr(i); 318 const size_t used_to_right = pointer_delta(space->top(), region_addr); 319 const size_t dead_to_right = used_to_right - live_to_right; 320 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 321 322 if (reclaimed_ratio > max_reclaimed_ratio) { 323 max_reclaimed_ratio = reclaimed_ratio; 324 max_reclaimed_ratio_region = i; 325 max_dead_to_right = dead_to_right; 326 max_live_to_right = live_to_right; 327 } 328 329 print_initial_summary_region(i, c, false); 330 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 331 reclaimed_ratio, dead_to_right, live_to_right); 332 333 live_to_right -= c->data_size(); 334 ++i; 335 } 336 337 // Any remaining regions are empty. Print one more if there is one. 338 if (i < end_region) { 339 print_initial_summary_region(i, summary_data.region(i)); 340 } 341 342 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 343 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 344 max_reclaimed_ratio_region, max_dead_to_right, 345 max_live_to_right, max_reclaimed_ratio); 346 } 347 348 void 349 print_initial_summary_data(ParallelCompactData& summary_data, 350 SpaceInfo* space_info) { 351 unsigned int id = PSParallelCompact::old_space_id; 352 const MutableSpace* space; 353 do { 354 space = space_info[id].space(); 355 print_initial_summary_data(summary_data, space); 356 } while (++id < PSParallelCompact::eden_space_id); 357 358 do { 359 space = space_info[id].space(); 360 print_generic_summary_data(summary_data, space->bottom(), space->top()); 361 } while (++id < PSParallelCompact::last_space_id); 362 } 363 #endif // #ifndef PRODUCT 364 365 #ifdef ASSERT 366 size_t add_obj_count; 367 size_t add_obj_size; 368 size_t mark_bitmap_count; 369 size_t mark_bitmap_size; 370 #endif // #ifdef ASSERT 371 372 ParallelCompactData::ParallelCompactData() 373 { 374 _region_start = 0; 375 376 _region_vspace = 0; 377 _reserved_byte_size = 0; 378 _region_data = 0; 379 _region_count = 0; 380 381 _block_vspace = 0; 382 _block_data = 0; 383 _block_count = 0; 384 } 385 386 bool ParallelCompactData::initialize(MemRegion covered_region) 387 { 388 _region_start = covered_region.start(); 389 const size_t region_size = covered_region.word_size(); 390 DEBUG_ONLY(_region_end = _region_start + region_size;) 391 392 assert(region_align_down(_region_start) == _region_start, 393 "region start not aligned"); 394 assert((region_size & RegionSizeOffsetMask) == 0, 395 "region size not a multiple of RegionSize"); 396 397 bool result = initialize_region_data(region_size) && initialize_block_data(); 398 return result; 399 } 400 401 PSVirtualSpace* 402 ParallelCompactData::create_vspace(size_t count, size_t element_size) 403 { 404 const size_t raw_bytes = count * element_size; 405 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 406 const size_t granularity = os::vm_allocation_granularity(); 407 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 408 409 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 410 MAX2(page_sz, granularity); 411 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 412 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 413 rs.size()); 414 415 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 416 417 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 418 if (vspace != 0) { 419 if (vspace->expand_by(_reserved_byte_size)) { 420 return vspace; 421 } 422 delete vspace; 423 // Release memory reserved in the space. 424 rs.release(); 425 } 426 427 return 0; 428 } 429 430 bool ParallelCompactData::initialize_region_data(size_t region_size) 431 { 432 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 433 _region_vspace = create_vspace(count, sizeof(RegionData)); 434 if (_region_vspace != 0) { 435 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 436 _region_count = count; 437 return true; 438 } 439 return false; 440 } 441 442 bool ParallelCompactData::initialize_block_data() 443 { 444 assert(_region_count != 0, "region data must be initialized first"); 445 const size_t count = _region_count << Log2BlocksPerRegion; 446 _block_vspace = create_vspace(count, sizeof(BlockData)); 447 if (_block_vspace != 0) { 448 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 449 _block_count = count; 450 return true; 451 } 452 return false; 453 } 454 455 void ParallelCompactData::clear() 456 { 457 memset(_region_data, 0, _region_vspace->committed_size()); 458 memset(_block_data, 0, _block_vspace->committed_size()); 459 } 460 461 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 462 assert(beg_region <= _region_count, "beg_region out of range"); 463 assert(end_region <= _region_count, "end_region out of range"); 464 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 465 466 const size_t region_cnt = end_region - beg_region; 467 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 468 469 const size_t beg_block = beg_region * BlocksPerRegion; 470 const size_t block_cnt = region_cnt * BlocksPerRegion; 471 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 472 } 473 474 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 475 { 476 const RegionData* cur_cp = region(region_idx); 477 const RegionData* const end_cp = region(region_count() - 1); 478 479 HeapWord* result = region_to_addr(region_idx); 480 if (cur_cp < end_cp) { 481 do { 482 result += cur_cp->partial_obj_size(); 483 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 484 } 485 return result; 486 } 487 488 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 489 { 490 const size_t obj_ofs = pointer_delta(addr, _region_start); 491 const size_t beg_region = obj_ofs >> Log2RegionSize; 492 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 493 494 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 495 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 496 497 if (beg_region == end_region) { 498 // All in one region. 499 _region_data[beg_region].add_live_obj(len); 500 return; 501 } 502 503 // First region. 504 const size_t beg_ofs = region_offset(addr); 505 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 506 507 Klass* klass = ((oop)addr)->klass(); 508 // Middle regions--completely spanned by this object. 509 for (size_t region = beg_region + 1; region < end_region; ++region) { 510 _region_data[region].set_partial_obj_size(RegionSize); 511 _region_data[region].set_partial_obj_addr(addr); 512 } 513 514 // Last region. 515 const size_t end_ofs = region_offset(addr + len - 1); 516 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 517 _region_data[end_region].set_partial_obj_addr(addr); 518 } 519 520 void 521 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 522 { 523 assert(region_offset(beg) == 0, "not RegionSize aligned"); 524 assert(region_offset(end) == 0, "not RegionSize aligned"); 525 526 size_t cur_region = addr_to_region_idx(beg); 527 const size_t end_region = addr_to_region_idx(end); 528 HeapWord* addr = beg; 529 while (cur_region < end_region) { 530 _region_data[cur_region].set_destination(addr); 531 _region_data[cur_region].set_destination_count(0); 532 _region_data[cur_region].set_source_region(cur_region); 533 _region_data[cur_region].set_data_location(addr); 534 535 // Update live_obj_size so the region appears completely full. 536 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 537 _region_data[cur_region].set_live_obj_size(live_size); 538 539 ++cur_region; 540 addr += RegionSize; 541 } 542 } 543 544 // Find the point at which a space can be split and, if necessary, record the 545 // split point. 546 // 547 // If the current src region (which overflowed the destination space) doesn't 548 // have a partial object, the split point is at the beginning of the current src 549 // region (an "easy" split, no extra bookkeeping required). 550 // 551 // If the current src region has a partial object, the split point is in the 552 // region where that partial object starts (call it the split_region). If 553 // split_region has a partial object, then the split point is just after that 554 // partial object (a "hard" split where we have to record the split data and 555 // zero the partial_obj_size field). With a "hard" split, we know that the 556 // partial_obj ends within split_region because the partial object that caused 557 // the overflow starts in split_region. If split_region doesn't have a partial 558 // obj, then the split is at the beginning of split_region (another "easy" 559 // split). 560 HeapWord* 561 ParallelCompactData::summarize_split_space(size_t src_region, 562 SplitInfo& split_info, 563 HeapWord* destination, 564 HeapWord* target_end, 565 HeapWord** target_next) 566 { 567 assert(destination <= target_end, "sanity"); 568 assert(destination + _region_data[src_region].data_size() > target_end, 569 "region should not fit into target space"); 570 assert(is_region_aligned(target_end), "sanity"); 571 572 size_t split_region = src_region; 573 HeapWord* split_destination = destination; 574 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 575 576 if (destination + partial_obj_size > target_end) { 577 // The split point is just after the partial object (if any) in the 578 // src_region that contains the start of the object that overflowed the 579 // destination space. 580 // 581 // Find the start of the "overflow" object and set split_region to the 582 // region containing it. 583 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 584 split_region = addr_to_region_idx(overflow_obj); 585 586 // Clear the source_region field of all destination regions whose first word 587 // came from data after the split point (a non-null source_region field 588 // implies a region must be filled). 589 // 590 // An alternative to the simple loop below: clear during post_compact(), 591 // which uses memcpy instead of individual stores, and is easy to 592 // parallelize. (The downside is that it clears the entire RegionData 593 // object as opposed to just one field.) 594 // 595 // post_compact() would have to clear the summary data up to the highest 596 // address that was written during the summary phase, which would be 597 // 598 // max(top, max(new_top, clear_top)) 599 // 600 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 601 // to target_end. 602 const RegionData* const sr = region(split_region); 603 const size_t beg_idx = 604 addr_to_region_idx(region_align_up(sr->destination() + 605 sr->partial_obj_size())); 606 const size_t end_idx = addr_to_region_idx(target_end); 607 608 if (TraceParallelOldGCSummaryPhase) { 609 gclog_or_tty->print_cr("split: clearing source_region field in [" 610 SIZE_FORMAT ", " SIZE_FORMAT ")", 611 beg_idx, end_idx); 612 } 613 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 614 _region_data[idx].set_source_region(0); 615 } 616 617 // Set split_destination and partial_obj_size to reflect the split region. 618 split_destination = sr->destination(); 619 partial_obj_size = sr->partial_obj_size(); 620 } 621 622 // The split is recorded only if a partial object extends onto the region. 623 if (partial_obj_size != 0) { 624 _region_data[split_region].set_partial_obj_size(0); 625 split_info.record(split_region, partial_obj_size, split_destination); 626 } 627 628 // Setup the continuation addresses. 629 *target_next = split_destination + partial_obj_size; 630 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 631 632 if (TraceParallelOldGCSummaryPhase) { 633 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 634 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 635 " pos=" SIZE_FORMAT, 636 split_type, p2i(source_next), split_region, 637 partial_obj_size); 638 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 639 " tn=" PTR_FORMAT, 640 split_type, p2i(split_destination), 641 addr_to_region_idx(split_destination), 642 p2i(*target_next)); 643 644 if (partial_obj_size != 0) { 645 HeapWord* const po_beg = split_info.destination(); 646 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 647 gclog_or_tty->print_cr("%s split: " 648 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 649 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 650 split_type, 651 p2i(po_beg), addr_to_region_idx(po_beg), 652 p2i(po_end), addr_to_region_idx(po_end)); 653 } 654 } 655 656 return source_next; 657 } 658 659 bool ParallelCompactData::summarize(SplitInfo& split_info, 660 HeapWord* source_beg, HeapWord* source_end, 661 HeapWord** source_next, 662 HeapWord* target_beg, HeapWord* target_end, 663 HeapWord** target_next) 664 { 665 if (TraceParallelOldGCSummaryPhase) { 666 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 667 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 668 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 669 p2i(source_beg), p2i(source_end), p2i(source_next_val), 670 p2i(target_beg), p2i(target_end), p2i(*target_next)); 671 } 672 673 size_t cur_region = addr_to_region_idx(source_beg); 674 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 675 676 HeapWord *dest_addr = target_beg; 677 while (cur_region < end_region) { 678 // The destination must be set even if the region has no data. 679 _region_data[cur_region].set_destination(dest_addr); 680 681 size_t words = _region_data[cur_region].data_size(); 682 if (words > 0) { 683 // If cur_region does not fit entirely into the target space, find a point 684 // at which the source space can be 'split' so that part is copied to the 685 // target space and the rest is copied elsewhere. 686 if (dest_addr + words > target_end) { 687 assert(source_next != NULL, "source_next is NULL when splitting"); 688 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 689 target_end, target_next); 690 return false; 691 } 692 693 // Compute the destination_count for cur_region, and if necessary, update 694 // source_region for a destination region. The source_region field is 695 // updated if cur_region is the first (left-most) region to be copied to a 696 // destination region. 697 // 698 // The destination_count calculation is a bit subtle. A region that has 699 // data that compacts into itself does not count itself as a destination. 700 // This maintains the invariant that a zero count means the region is 701 // available and can be claimed and then filled. 702 uint destination_count = 0; 703 if (split_info.is_split(cur_region)) { 704 // The current region has been split: the partial object will be copied 705 // to one destination space and the remaining data will be copied to 706 // another destination space. Adjust the initial destination_count and, 707 // if necessary, set the source_region field if the partial object will 708 // cross a destination region boundary. 709 destination_count = split_info.destination_count(); 710 if (destination_count == 2) { 711 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 712 _region_data[dest_idx].set_source_region(cur_region); 713 } 714 } 715 716 HeapWord* const last_addr = dest_addr + words - 1; 717 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 718 const size_t dest_region_2 = addr_to_region_idx(last_addr); 719 720 // Initially assume that the destination regions will be the same and 721 // adjust the value below if necessary. Under this assumption, if 722 // cur_region == dest_region_2, then cur_region will be compacted 723 // completely into itself. 724 destination_count += cur_region == dest_region_2 ? 0 : 1; 725 if (dest_region_1 != dest_region_2) { 726 // Destination regions differ; adjust destination_count. 727 destination_count += 1; 728 // Data from cur_region will be copied to the start of dest_region_2. 729 _region_data[dest_region_2].set_source_region(cur_region); 730 } else if (region_offset(dest_addr) == 0) { 731 // Data from cur_region will be copied to the start of the destination 732 // region. 733 _region_data[dest_region_1].set_source_region(cur_region); 734 } 735 736 _region_data[cur_region].set_destination_count(destination_count); 737 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 738 dest_addr += words; 739 } 740 741 ++cur_region; 742 } 743 744 *target_next = dest_addr; 745 return true; 746 } 747 748 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 749 assert(addr != NULL, "Should detect NULL oop earlier"); 750 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 751 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 752 753 // Region covering the object. 754 RegionData* const region_ptr = addr_to_region_ptr(addr); 755 HeapWord* result = region_ptr->destination(); 756 757 // If the entire Region is live, the new location is region->destination + the 758 // offset of the object within in the Region. 759 760 // Run some performance tests to determine if this special case pays off. It 761 // is worth it for pointers into the dense prefix. If the optimization to 762 // avoid pointer updates in regions that only point to the dense prefix is 763 // ever implemented, this should be revisited. 764 if (region_ptr->data_size() == RegionSize) { 765 result += region_offset(addr); 766 return result; 767 } 768 769 // Otherwise, the new location is region->destination + block offset + the 770 // number of live words in the Block that are (a) to the left of addr and (b) 771 // due to objects that start in the Block. 772 773 // Fill in the block table if necessary. This is unsynchronized, so multiple 774 // threads may fill the block table for a region (harmless, since it is 775 // idempotent). 776 if (!region_ptr->blocks_filled()) { 777 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 778 region_ptr->set_blocks_filled(); 779 } 780 781 HeapWord* const search_start = block_align_down(addr); 782 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 783 784 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 785 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 786 result += block_offset + live; 787 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 788 return result; 789 } 790 791 #ifdef ASSERT 792 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 793 { 794 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 795 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 796 for (const size_t* p = beg; p < end; ++p) { 797 assert(*p == 0, "not zero"); 798 } 799 } 800 801 void ParallelCompactData::verify_clear() 802 { 803 verify_clear(_region_vspace); 804 verify_clear(_block_vspace); 805 } 806 #endif // #ifdef ASSERT 807 808 STWGCTimer PSParallelCompact::_gc_timer; 809 ParallelOldTracer PSParallelCompact::_gc_tracer; 810 elapsedTimer PSParallelCompact::_accumulated_time; 811 unsigned int PSParallelCompact::_total_invocations = 0; 812 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 813 jlong PSParallelCompact::_time_of_last_gc = 0; 814 CollectorCounters* PSParallelCompact::_counters = NULL; 815 ParMarkBitMap PSParallelCompact::_mark_bitmap; 816 ParallelCompactData PSParallelCompact::_summary_data; 817 818 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 819 820 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 821 822 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 823 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 824 825 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 826 827 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 828 klass->oops_do(_mark_and_push_closure); 829 } 830 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 831 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 832 } 833 834 void PSParallelCompact::post_initialize() { 835 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 836 MemRegion mr = heap->reserved_region(); 837 _ref_processor = 838 new ReferenceProcessor(mr, // span 839 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 840 (int) ParallelGCThreads, // mt processing degree 841 true, // mt discovery 842 (int) ParallelGCThreads, // mt discovery degree 843 true, // atomic_discovery 844 &_is_alive_closure); // non-header is alive closure 845 _counters = new CollectorCounters("PSParallelCompact", 1); 846 847 // Initialize static fields in ParCompactionManager. 848 ParCompactionManager::initialize(mark_bitmap()); 849 } 850 851 bool PSParallelCompact::initialize() { 852 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 853 MemRegion mr = heap->reserved_region(); 854 855 // Was the old gen get allocated successfully? 856 if (!heap->old_gen()->is_allocated()) { 857 return false; 858 } 859 860 initialize_space_info(); 861 initialize_dead_wood_limiter(); 862 863 if (!_mark_bitmap.initialize(mr)) { 864 vm_shutdown_during_initialization( 865 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 866 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 867 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 868 return false; 869 } 870 871 if (!_summary_data.initialize(mr)) { 872 vm_shutdown_during_initialization( 873 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 874 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 875 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 876 return false; 877 } 878 879 return true; 880 } 881 882 void PSParallelCompact::initialize_space_info() 883 { 884 memset(&_space_info, 0, sizeof(_space_info)); 885 886 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 887 PSYoungGen* young_gen = heap->young_gen(); 888 889 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 890 _space_info[eden_space_id].set_space(young_gen->eden_space()); 891 _space_info[from_space_id].set_space(young_gen->from_space()); 892 _space_info[to_space_id].set_space(young_gen->to_space()); 893 894 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 895 } 896 897 void PSParallelCompact::initialize_dead_wood_limiter() 898 { 899 const size_t max = 100; 900 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 901 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 902 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 903 DEBUG_ONLY(_dwl_initialized = true;) 904 _dwl_adjustment = normal_distribution(1.0); 905 } 906 907 // Simple class for storing info about the heap at the start of GC, to be used 908 // after GC for comparison/printing. 909 class PreGCValues { 910 public: 911 PreGCValues() { } 912 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 913 914 void fill(ParallelScavengeHeap* heap) { 915 _heap_used = heap->used(); 916 _young_gen_used = heap->young_gen()->used_in_bytes(); 917 _old_gen_used = heap->old_gen()->used_in_bytes(); 918 _metadata_used = MetaspaceAux::used_bytes(); 919 }; 920 921 size_t heap_used() const { return _heap_used; } 922 size_t young_gen_used() const { return _young_gen_used; } 923 size_t old_gen_used() const { return _old_gen_used; } 924 size_t metadata_used() const { return _metadata_used; } 925 926 private: 927 size_t _heap_used; 928 size_t _young_gen_used; 929 size_t _old_gen_used; 930 size_t _metadata_used; 931 }; 932 933 void 934 PSParallelCompact::clear_data_covering_space(SpaceId id) 935 { 936 // At this point, top is the value before GC, new_top() is the value that will 937 // be set at the end of GC. The marking bitmap is cleared to top; nothing 938 // should be marked above top. The summary data is cleared to the larger of 939 // top & new_top. 940 MutableSpace* const space = _space_info[id].space(); 941 HeapWord* const bot = space->bottom(); 942 HeapWord* const top = space->top(); 943 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 944 945 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 946 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 947 _mark_bitmap.clear_range(beg_bit, end_bit); 948 949 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 950 const size_t end_region = 951 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 952 _summary_data.clear_range(beg_region, end_region); 953 954 // Clear the data used to 'split' regions. 955 SplitInfo& split_info = _space_info[id].split_info(); 956 if (split_info.is_valid()) { 957 split_info.clear(); 958 } 959 DEBUG_ONLY(split_info.verify_clear();) 960 } 961 962 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 963 { 964 // Update the from & to space pointers in space_info, since they are swapped 965 // at each young gen gc. Do the update unconditionally (even though a 966 // promotion failure does not swap spaces) because an unknown number of minor 967 // collections will have swapped the spaces an unknown number of times. 968 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 969 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 970 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 971 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 972 973 pre_gc_values->fill(heap); 974 975 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 976 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 977 978 // Increment the invocation count 979 heap->increment_total_collections(true); 980 981 // We need to track unique mark sweep invocations as well. 982 _total_invocations++; 983 984 heap->print_heap_before_gc(); 985 heap->trace_heap_before_gc(&_gc_tracer); 986 987 // Fill in TLABs 988 heap->accumulate_statistics_all_tlabs(); 989 heap->ensure_parsability(true); // retire TLABs 990 991 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 992 HandleMark hm; // Discard invalid handles created during verification 993 Universe::verify(" VerifyBeforeGC:"); 994 } 995 996 // Verify object start arrays 997 if (VerifyObjectStartArray && 998 VerifyBeforeGC) { 999 heap->old_gen()->verify_object_start_array(); 1000 } 1001 1002 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1003 DEBUG_ONLY(summary_data().verify_clear();) 1004 1005 // Have worker threads release resources the next time they run a task. 1006 gc_task_manager()->release_all_resources(); 1007 } 1008 1009 void PSParallelCompact::post_compact() 1010 { 1011 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1012 1013 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1014 // Clear the marking bitmap, summary data and split info. 1015 clear_data_covering_space(SpaceId(id)); 1016 // Update top(). Must be done after clearing the bitmap and summary data. 1017 _space_info[id].publish_new_top(); 1018 } 1019 1020 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1021 MutableSpace* const from_space = _space_info[from_space_id].space(); 1022 MutableSpace* const to_space = _space_info[to_space_id].space(); 1023 1024 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1025 bool eden_empty = eden_space->is_empty(); 1026 if (!eden_empty) { 1027 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1028 heap->young_gen(), heap->old_gen()); 1029 } 1030 1031 // Update heap occupancy information which is used as input to the soft ref 1032 // clearing policy at the next gc. 1033 Universe::update_heap_info_at_gc(); 1034 1035 bool young_gen_empty = eden_empty && from_space->is_empty() && 1036 to_space->is_empty(); 1037 1038 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1039 MemRegion old_mr = heap->old_gen()->reserved(); 1040 if (young_gen_empty) { 1041 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1042 } else { 1043 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1044 } 1045 1046 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1047 ClassLoaderDataGraph::purge(); 1048 MetaspaceAux::verify_metrics(); 1049 1050 CodeCache::gc_epilogue(); 1051 JvmtiExport::gc_epilogue(); 1052 1053 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1054 1055 ref_processor()->enqueue_discovered_references(NULL); 1056 1057 if (ZapUnusedHeapArea) { 1058 heap->gen_mangle_unused_area(); 1059 } 1060 1061 // Update time of last GC 1062 reset_millis_since_last_gc(); 1063 } 1064 1065 HeapWord* 1066 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1067 bool maximum_compaction) 1068 { 1069 const size_t region_size = ParallelCompactData::RegionSize; 1070 const ParallelCompactData& sd = summary_data(); 1071 1072 const MutableSpace* const space = _space_info[id].space(); 1073 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1074 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1075 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1076 1077 // Skip full regions at the beginning of the space--they are necessarily part 1078 // of the dense prefix. 1079 size_t full_count = 0; 1080 const RegionData* cp; 1081 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1082 ++full_count; 1083 } 1084 1085 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1086 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1087 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1088 if (maximum_compaction || cp == end_cp || interval_ended) { 1089 _maximum_compaction_gc_num = total_invocations(); 1090 return sd.region_to_addr(cp); 1091 } 1092 1093 HeapWord* const new_top = _space_info[id].new_top(); 1094 const size_t space_live = pointer_delta(new_top, space->bottom()); 1095 const size_t space_used = space->used_in_words(); 1096 const size_t space_capacity = space->capacity_in_words(); 1097 1098 const double cur_density = double(space_live) / space_capacity; 1099 const double deadwood_density = 1100 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1101 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1102 1103 if (TraceParallelOldGCDensePrefix) { 1104 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1105 cur_density, deadwood_density, deadwood_goal); 1106 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1107 "space_cap=" SIZE_FORMAT, 1108 space_live, space_used, 1109 space_capacity); 1110 } 1111 1112 // XXX - Use binary search? 1113 HeapWord* dense_prefix = sd.region_to_addr(cp); 1114 const RegionData* full_cp = cp; 1115 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1116 while (cp < end_cp) { 1117 HeapWord* region_destination = cp->destination(); 1118 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1119 if (TraceParallelOldGCDensePrefix && Verbose) { 1120 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1121 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1122 sd.region(cp), p2i(region_destination), 1123 p2i(dense_prefix), cur_deadwood); 1124 } 1125 1126 if (cur_deadwood >= deadwood_goal) { 1127 // Found the region that has the correct amount of deadwood to the left. 1128 // This typically occurs after crossing a fairly sparse set of regions, so 1129 // iterate backwards over those sparse regions, looking for the region 1130 // that has the lowest density of live objects 'to the right.' 1131 size_t space_to_left = sd.region(cp) * region_size; 1132 size_t live_to_left = space_to_left - cur_deadwood; 1133 size_t space_to_right = space_capacity - space_to_left; 1134 size_t live_to_right = space_live - live_to_left; 1135 double density_to_right = double(live_to_right) / space_to_right; 1136 while (cp > full_cp) { 1137 --cp; 1138 const size_t prev_region_live_to_right = live_to_right - 1139 cp->data_size(); 1140 const size_t prev_region_space_to_right = space_to_right + region_size; 1141 double prev_region_density_to_right = 1142 double(prev_region_live_to_right) / prev_region_space_to_right; 1143 if (density_to_right <= prev_region_density_to_right) { 1144 return dense_prefix; 1145 } 1146 if (TraceParallelOldGCDensePrefix && Verbose) { 1147 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1148 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1149 prev_region_density_to_right); 1150 } 1151 dense_prefix -= region_size; 1152 live_to_right = prev_region_live_to_right; 1153 space_to_right = prev_region_space_to_right; 1154 density_to_right = prev_region_density_to_right; 1155 } 1156 return dense_prefix; 1157 } 1158 1159 dense_prefix += region_size; 1160 ++cp; 1161 } 1162 1163 return dense_prefix; 1164 } 1165 1166 #ifndef PRODUCT 1167 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1168 const SpaceId id, 1169 const bool maximum_compaction, 1170 HeapWord* const addr) 1171 { 1172 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1173 RegionData* const cp = summary_data().region(region_idx); 1174 const MutableSpace* const space = _space_info[id].space(); 1175 HeapWord* const new_top = _space_info[id].new_top(); 1176 1177 const size_t space_live = pointer_delta(new_top, space->bottom()); 1178 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1179 const size_t space_cap = space->capacity_in_words(); 1180 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1181 const size_t live_to_right = new_top - cp->destination(); 1182 const size_t dead_to_right = space->top() - addr - live_to_right; 1183 1184 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1185 "spl=" SIZE_FORMAT " " 1186 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1187 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1188 " ratio=%10.8f", 1189 algorithm, p2i(addr), region_idx, 1190 space_live, 1191 dead_to_left, dead_to_left_pct, 1192 dead_to_right, live_to_right, 1193 double(dead_to_right) / live_to_right); 1194 } 1195 #endif // #ifndef PRODUCT 1196 1197 // Return a fraction indicating how much of the generation can be treated as 1198 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1199 // based on the density of live objects in the generation to determine a limit, 1200 // which is then adjusted so the return value is min_percent when the density is 1201 // 1. 1202 // 1203 // The following table shows some return values for a different values of the 1204 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1205 // min_percent is 1. 1206 // 1207 // fraction allowed as dead wood 1208 // ----------------------------------------------------------------- 1209 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1210 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1211 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1212 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1213 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1214 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1215 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1216 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1217 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1218 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1219 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1220 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1221 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1222 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1223 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1224 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1225 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1226 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1227 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1228 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1229 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1230 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1231 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1232 1233 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1234 { 1235 assert(_dwl_initialized, "uninitialized"); 1236 1237 // The raw limit is the value of the normal distribution at x = density. 1238 const double raw_limit = normal_distribution(density); 1239 1240 // Adjust the raw limit so it becomes the minimum when the density is 1. 1241 // 1242 // First subtract the adjustment value (which is simply the precomputed value 1243 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1244 // Then add the minimum value, so the minimum is returned when the density is 1245 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1246 const double min = double(min_percent) / 100.0; 1247 const double limit = raw_limit - _dwl_adjustment + min; 1248 return MAX2(limit, 0.0); 1249 } 1250 1251 ParallelCompactData::RegionData* 1252 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1253 const RegionData* end) 1254 { 1255 const size_t region_size = ParallelCompactData::RegionSize; 1256 ParallelCompactData& sd = summary_data(); 1257 size_t left = sd.region(beg); 1258 size_t right = end > beg ? sd.region(end) - 1 : left; 1259 1260 // Binary search. 1261 while (left < right) { 1262 // Equivalent to (left + right) / 2, but does not overflow. 1263 const size_t middle = left + (right - left) / 2; 1264 RegionData* const middle_ptr = sd.region(middle); 1265 HeapWord* const dest = middle_ptr->destination(); 1266 HeapWord* const addr = sd.region_to_addr(middle); 1267 assert(dest != NULL, "sanity"); 1268 assert(dest <= addr, "must move left"); 1269 1270 if (middle > left && dest < addr) { 1271 right = middle - 1; 1272 } else if (middle < right && middle_ptr->data_size() == region_size) { 1273 left = middle + 1; 1274 } else { 1275 return middle_ptr; 1276 } 1277 } 1278 return sd.region(left); 1279 } 1280 1281 ParallelCompactData::RegionData* 1282 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1283 const RegionData* end, 1284 size_t dead_words) 1285 { 1286 ParallelCompactData& sd = summary_data(); 1287 size_t left = sd.region(beg); 1288 size_t right = end > beg ? sd.region(end) - 1 : left; 1289 1290 // Binary search. 1291 while (left < right) { 1292 // Equivalent to (left + right) / 2, but does not overflow. 1293 const size_t middle = left + (right - left) / 2; 1294 RegionData* const middle_ptr = sd.region(middle); 1295 HeapWord* const dest = middle_ptr->destination(); 1296 HeapWord* const addr = sd.region_to_addr(middle); 1297 assert(dest != NULL, "sanity"); 1298 assert(dest <= addr, "must move left"); 1299 1300 const size_t dead_to_left = pointer_delta(addr, dest); 1301 if (middle > left && dead_to_left > dead_words) { 1302 right = middle - 1; 1303 } else if (middle < right && dead_to_left < dead_words) { 1304 left = middle + 1; 1305 } else { 1306 return middle_ptr; 1307 } 1308 } 1309 return sd.region(left); 1310 } 1311 1312 // The result is valid during the summary phase, after the initial summarization 1313 // of each space into itself, and before final summarization. 1314 inline double 1315 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1316 HeapWord* const bottom, 1317 HeapWord* const top, 1318 HeapWord* const new_top) 1319 { 1320 ParallelCompactData& sd = summary_data(); 1321 1322 assert(cp != NULL, "sanity"); 1323 assert(bottom != NULL, "sanity"); 1324 assert(top != NULL, "sanity"); 1325 assert(new_top != NULL, "sanity"); 1326 assert(top >= new_top, "summary data problem?"); 1327 assert(new_top > bottom, "space is empty; should not be here"); 1328 assert(new_top >= cp->destination(), "sanity"); 1329 assert(top >= sd.region_to_addr(cp), "sanity"); 1330 1331 HeapWord* const destination = cp->destination(); 1332 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1333 const size_t compacted_region_live = pointer_delta(new_top, destination); 1334 const size_t compacted_region_used = pointer_delta(top, 1335 sd.region_to_addr(cp)); 1336 const size_t reclaimable = compacted_region_used - compacted_region_live; 1337 1338 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1339 return double(reclaimable) / divisor; 1340 } 1341 1342 // Return the address of the end of the dense prefix, a.k.a. the start of the 1343 // compacted region. The address is always on a region boundary. 1344 // 1345 // Completely full regions at the left are skipped, since no compaction can 1346 // occur in those regions. Then the maximum amount of dead wood to allow is 1347 // computed, based on the density (amount live / capacity) of the generation; 1348 // the region with approximately that amount of dead space to the left is 1349 // identified as the limit region. Regions between the last completely full 1350 // region and the limit region are scanned and the one that has the best 1351 // (maximum) reclaimed_ratio() is selected. 1352 HeapWord* 1353 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1354 bool maximum_compaction) 1355 { 1356 if (ParallelOldGCSplitALot) { 1357 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1358 // The value was chosen to provoke splitting a young gen space; use it. 1359 return _space_info[id].dense_prefix(); 1360 } 1361 } 1362 1363 const size_t region_size = ParallelCompactData::RegionSize; 1364 const ParallelCompactData& sd = summary_data(); 1365 1366 const MutableSpace* const space = _space_info[id].space(); 1367 HeapWord* const top = space->top(); 1368 HeapWord* const top_aligned_up = sd.region_align_up(top); 1369 HeapWord* const new_top = _space_info[id].new_top(); 1370 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1371 HeapWord* const bottom = space->bottom(); 1372 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1373 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1374 const RegionData* const new_top_cp = 1375 sd.addr_to_region_ptr(new_top_aligned_up); 1376 1377 // Skip full regions at the beginning of the space--they are necessarily part 1378 // of the dense prefix. 1379 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1380 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1381 space->is_empty(), "no dead space allowed to the left"); 1382 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1383 "region must have dead space"); 1384 1385 // The gc number is saved whenever a maximum compaction is done, and used to 1386 // determine when the maximum compaction interval has expired. This avoids 1387 // successive max compactions for different reasons. 1388 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1389 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1390 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1391 total_invocations() == HeapFirstMaximumCompactionCount; 1392 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1393 _maximum_compaction_gc_num = total_invocations(); 1394 return sd.region_to_addr(full_cp); 1395 } 1396 1397 const size_t space_live = pointer_delta(new_top, bottom); 1398 const size_t space_used = space->used_in_words(); 1399 const size_t space_capacity = space->capacity_in_words(); 1400 1401 const double density = double(space_live) / double(space_capacity); 1402 const size_t min_percent_free = MarkSweepDeadRatio; 1403 const double limiter = dead_wood_limiter(density, min_percent_free); 1404 const size_t dead_wood_max = space_used - space_live; 1405 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1406 dead_wood_max); 1407 1408 if (TraceParallelOldGCDensePrefix) { 1409 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1410 "space_cap=" SIZE_FORMAT, 1411 space_live, space_used, 1412 space_capacity); 1413 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1414 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1415 density, min_percent_free, limiter, 1416 dead_wood_max, dead_wood_limit); 1417 } 1418 1419 // Locate the region with the desired amount of dead space to the left. 1420 const RegionData* const limit_cp = 1421 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1422 1423 // Scan from the first region with dead space to the limit region and find the 1424 // one with the best (largest) reclaimed ratio. 1425 double best_ratio = 0.0; 1426 const RegionData* best_cp = full_cp; 1427 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1428 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1429 if (tmp_ratio > best_ratio) { 1430 best_cp = cp; 1431 best_ratio = tmp_ratio; 1432 } 1433 } 1434 1435 #if 0 1436 // Something to consider: if the region with the best ratio is 'close to' the 1437 // first region w/free space, choose the first region with free space 1438 // ("first-free"). The first-free region is usually near the start of the 1439 // heap, which means we are copying most of the heap already, so copy a bit 1440 // more to get complete compaction. 1441 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1442 _maximum_compaction_gc_num = total_invocations(); 1443 best_cp = full_cp; 1444 } 1445 #endif // #if 0 1446 1447 return sd.region_to_addr(best_cp); 1448 } 1449 1450 #ifndef PRODUCT 1451 void 1452 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1453 size_t words) 1454 { 1455 if (TraceParallelOldGCSummaryPhase) { 1456 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1457 SIZE_FORMAT, p2i(start), p2i(start + words), words); 1458 } 1459 1460 ObjectStartArray* const start_array = _space_info[id].start_array(); 1461 CollectedHeap::fill_with_objects(start, words); 1462 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1463 _mark_bitmap.mark_obj(p, words); 1464 _summary_data.add_obj(p, words); 1465 start_array->allocate_block(p); 1466 } 1467 } 1468 1469 void 1470 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1471 { 1472 ParallelCompactData& sd = summary_data(); 1473 MutableSpace* space = _space_info[id].space(); 1474 1475 // Find the source and destination start addresses. 1476 HeapWord* const src_addr = sd.region_align_down(start); 1477 HeapWord* dst_addr; 1478 if (src_addr < start) { 1479 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1480 } else if (src_addr > space->bottom()) { 1481 // The start (the original top() value) is aligned to a region boundary so 1482 // the associated region does not have a destination. Compute the 1483 // destination from the previous region. 1484 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1485 dst_addr = cp->destination() + cp->data_size(); 1486 } else { 1487 // Filling the entire space. 1488 dst_addr = space->bottom(); 1489 } 1490 assert(dst_addr != NULL, "sanity"); 1491 1492 // Update the summary data. 1493 bool result = _summary_data.summarize(_space_info[id].split_info(), 1494 src_addr, space->top(), NULL, 1495 dst_addr, space->end(), 1496 _space_info[id].new_top_addr()); 1497 assert(result, "should not fail: bad filler object size"); 1498 } 1499 1500 void 1501 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1502 { 1503 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1504 return; 1505 } 1506 1507 MutableSpace* const space = _space_info[id].space(); 1508 if (space->is_empty()) { 1509 HeapWord* b = space->bottom(); 1510 HeapWord* t = b + space->capacity_in_words() / 2; 1511 space->set_top(t); 1512 if (ZapUnusedHeapArea) { 1513 space->set_top_for_allocations(); 1514 } 1515 1516 size_t min_size = CollectedHeap::min_fill_size(); 1517 size_t obj_len = min_size; 1518 while (b + obj_len <= t) { 1519 CollectedHeap::fill_with_object(b, obj_len); 1520 mark_bitmap()->mark_obj(b, obj_len); 1521 summary_data().add_obj(b, obj_len); 1522 b += obj_len; 1523 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1524 } 1525 if (b < t) { 1526 // The loop didn't completely fill to t (top); adjust top downward. 1527 space->set_top(b); 1528 if (ZapUnusedHeapArea) { 1529 space->set_top_for_allocations(); 1530 } 1531 } 1532 1533 HeapWord** nta = _space_info[id].new_top_addr(); 1534 bool result = summary_data().summarize(_space_info[id].split_info(), 1535 space->bottom(), space->top(), NULL, 1536 space->bottom(), space->end(), nta); 1537 assert(result, "space must fit into itself"); 1538 } 1539 } 1540 1541 void 1542 PSParallelCompact::provoke_split(bool & max_compaction) 1543 { 1544 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1545 return; 1546 } 1547 1548 const size_t region_size = ParallelCompactData::RegionSize; 1549 ParallelCompactData& sd = summary_data(); 1550 1551 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1552 MutableSpace* const from_space = _space_info[from_space_id].space(); 1553 const size_t eden_live = pointer_delta(eden_space->top(), 1554 _space_info[eden_space_id].new_top()); 1555 const size_t from_live = pointer_delta(from_space->top(), 1556 _space_info[from_space_id].new_top()); 1557 1558 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1559 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1560 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1561 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1562 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1563 1564 // Choose the space to split; need at least 2 regions live (or fillable). 1565 SpaceId id; 1566 MutableSpace* space; 1567 size_t live_words; 1568 size_t fill_words; 1569 if (eden_live + eden_fillable >= region_size * 2) { 1570 id = eden_space_id; 1571 space = eden_space; 1572 live_words = eden_live; 1573 fill_words = eden_fillable; 1574 } else if (from_live + from_fillable >= region_size * 2) { 1575 id = from_space_id; 1576 space = from_space; 1577 live_words = from_live; 1578 fill_words = from_fillable; 1579 } else { 1580 return; // Give up. 1581 } 1582 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1583 1584 if (live_words < region_size * 2) { 1585 // Fill from top() to end() w/live objects of mixed sizes. 1586 HeapWord* const fill_start = space->top(); 1587 live_words += fill_words; 1588 1589 space->set_top(fill_start + fill_words); 1590 if (ZapUnusedHeapArea) { 1591 space->set_top_for_allocations(); 1592 } 1593 1594 HeapWord* cur_addr = fill_start; 1595 while (fill_words > 0) { 1596 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1597 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1598 if (fill_words - cur_size < min_fill_size) { 1599 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1600 } 1601 1602 CollectedHeap::fill_with_object(cur_addr, cur_size); 1603 mark_bitmap()->mark_obj(cur_addr, cur_size); 1604 sd.add_obj(cur_addr, cur_size); 1605 1606 cur_addr += cur_size; 1607 fill_words -= cur_size; 1608 } 1609 1610 summarize_new_objects(id, fill_start); 1611 } 1612 1613 max_compaction = false; 1614 1615 // Manipulate the old gen so that it has room for about half of the live data 1616 // in the target young gen space (live_words / 2). 1617 id = old_space_id; 1618 space = _space_info[id].space(); 1619 const size_t free_at_end = space->free_in_words(); 1620 const size_t free_target = align_object_size(live_words / 2); 1621 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1622 1623 if (free_at_end >= free_target + min_fill_size) { 1624 // Fill space above top() and set the dense prefix so everything survives. 1625 HeapWord* const fill_start = space->top(); 1626 const size_t fill_size = free_at_end - free_target; 1627 space->set_top(space->top() + fill_size); 1628 if (ZapUnusedHeapArea) { 1629 space->set_top_for_allocations(); 1630 } 1631 fill_with_live_objects(id, fill_start, fill_size); 1632 summarize_new_objects(id, fill_start); 1633 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1634 } else if (dead + free_at_end > free_target) { 1635 // Find a dense prefix that makes the right amount of space available. 1636 HeapWord* cur = sd.region_align_down(space->top()); 1637 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1638 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1639 while (dead_to_right < free_target) { 1640 cur -= region_size; 1641 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1642 dead_to_right = pointer_delta(space->end(), cur_destination); 1643 } 1644 _space_info[id].set_dense_prefix(cur); 1645 } 1646 } 1647 #endif // #ifndef PRODUCT 1648 1649 void PSParallelCompact::summarize_spaces_quick() 1650 { 1651 for (unsigned int i = 0; i < last_space_id; ++i) { 1652 const MutableSpace* space = _space_info[i].space(); 1653 HeapWord** nta = _space_info[i].new_top_addr(); 1654 bool result = _summary_data.summarize(_space_info[i].split_info(), 1655 space->bottom(), space->top(), NULL, 1656 space->bottom(), space->end(), nta); 1657 assert(result, "space must fit into itself"); 1658 _space_info[i].set_dense_prefix(space->bottom()); 1659 } 1660 1661 #ifndef PRODUCT 1662 if (ParallelOldGCSplitALot) { 1663 provoke_split_fill_survivor(to_space_id); 1664 } 1665 #endif // #ifndef PRODUCT 1666 } 1667 1668 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1669 { 1670 HeapWord* const dense_prefix_end = dense_prefix(id); 1671 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1672 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1673 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1674 // Only enough dead space is filled so that any remaining dead space to the 1675 // left is larger than the minimum filler object. (The remainder is filled 1676 // during the copy/update phase.) 1677 // 1678 // The size of the dead space to the right of the boundary is not a 1679 // concern, since compaction will be able to use whatever space is 1680 // available. 1681 // 1682 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1683 // surrounds the space to be filled with an object. 1684 // 1685 // In the 32-bit VM, each bit represents two 32-bit words: 1686 // +---+ 1687 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1688 // end_bits: ... x x x | 0 | || 0 x x ... 1689 // +---+ 1690 // 1691 // In the 64-bit VM, each bit represents one 64-bit word: 1692 // +------------+ 1693 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1694 // end_bits: ... x x 1 | 0 || 0 | x x ... 1695 // +------------+ 1696 // +-------+ 1697 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1698 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1699 // +-------+ 1700 // +-----------+ 1701 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1702 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1703 // +-----------+ 1704 // +-------+ 1705 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1706 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1707 // +-------+ 1708 1709 // Initially assume case a, c or e will apply. 1710 size_t obj_len = CollectedHeap::min_fill_size(); 1711 HeapWord* obj_beg = dense_prefix_end - obj_len; 1712 1713 #ifdef _LP64 1714 if (MinObjAlignment > 1) { // object alignment > heap word size 1715 // Cases a, c or e. 1716 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1717 // Case b above. 1718 obj_beg = dense_prefix_end - 1; 1719 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1720 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1721 // Case d above. 1722 obj_beg = dense_prefix_end - 3; 1723 obj_len = 3; 1724 } 1725 #endif // #ifdef _LP64 1726 1727 CollectedHeap::fill_with_object(obj_beg, obj_len); 1728 _mark_bitmap.mark_obj(obj_beg, obj_len); 1729 _summary_data.add_obj(obj_beg, obj_len); 1730 assert(start_array(id) != NULL, "sanity"); 1731 start_array(id)->allocate_block(obj_beg); 1732 } 1733 } 1734 1735 void 1736 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1737 { 1738 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1739 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1740 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1741 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1742 cur->set_source_region(0); 1743 } 1744 } 1745 1746 void 1747 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1748 { 1749 assert(id < last_space_id, "id out of range"); 1750 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1751 ParallelOldGCSplitALot && id == old_space_id, 1752 "should have been reset in summarize_spaces_quick()"); 1753 1754 const MutableSpace* space = _space_info[id].space(); 1755 if (_space_info[id].new_top() != space->bottom()) { 1756 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1757 _space_info[id].set_dense_prefix(dense_prefix_end); 1758 1759 #ifndef PRODUCT 1760 if (TraceParallelOldGCDensePrefix) { 1761 print_dense_prefix_stats("ratio", id, maximum_compaction, 1762 dense_prefix_end); 1763 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1764 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1765 } 1766 #endif // #ifndef PRODUCT 1767 1768 // Recompute the summary data, taking into account the dense prefix. If 1769 // every last byte will be reclaimed, then the existing summary data which 1770 // compacts everything can be left in place. 1771 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1772 // If dead space crosses the dense prefix boundary, it is (at least 1773 // partially) filled with a dummy object, marked live and added to the 1774 // summary data. This simplifies the copy/update phase and must be done 1775 // before the final locations of objects are determined, to prevent 1776 // leaving a fragment of dead space that is too small to fill. 1777 fill_dense_prefix_end(id); 1778 1779 // Compute the destination of each Region, and thus each object. 1780 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1781 _summary_data.summarize(_space_info[id].split_info(), 1782 dense_prefix_end, space->top(), NULL, 1783 dense_prefix_end, space->end(), 1784 _space_info[id].new_top_addr()); 1785 } 1786 } 1787 1788 if (TraceParallelOldGCSummaryPhase) { 1789 const size_t region_size = ParallelCompactData::RegionSize; 1790 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1791 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1792 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1793 HeapWord* const new_top = _space_info[id].new_top(); 1794 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1795 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1796 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1797 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1798 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1799 id, space->capacity_in_words(), p2i(dense_prefix_end), 1800 dp_region, dp_words / region_size, 1801 cr_words / region_size, p2i(new_top)); 1802 } 1803 } 1804 1805 #ifndef PRODUCT 1806 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1807 HeapWord* dst_beg, HeapWord* dst_end, 1808 SpaceId src_space_id, 1809 HeapWord* src_beg, HeapWord* src_end) 1810 { 1811 if (TraceParallelOldGCSummaryPhase) { 1812 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1813 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1814 SIZE_FORMAT "-" SIZE_FORMAT " " 1815 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1816 SIZE_FORMAT "-" SIZE_FORMAT, 1817 src_space_id, space_names[src_space_id], 1818 dst_space_id, space_names[dst_space_id], 1819 p2i(src_beg), p2i(src_end), 1820 _summary_data.addr_to_region_idx(src_beg), 1821 _summary_data.addr_to_region_idx(src_end), 1822 p2i(dst_beg), p2i(dst_end), 1823 _summary_data.addr_to_region_idx(dst_beg), 1824 _summary_data.addr_to_region_idx(dst_end)); 1825 } 1826 } 1827 #endif // #ifndef PRODUCT 1828 1829 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1830 bool maximum_compaction) 1831 { 1832 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 1833 // trace("2"); 1834 1835 #ifdef ASSERT 1836 if (TraceParallelOldGCMarkingPhase) { 1837 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1838 "add_obj_bytes=" SIZE_FORMAT, 1839 add_obj_count, add_obj_size * HeapWordSize); 1840 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1841 "mark_bitmap_bytes=" SIZE_FORMAT, 1842 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1843 } 1844 #endif // #ifdef ASSERT 1845 1846 // Quick summarization of each space into itself, to see how much is live. 1847 summarize_spaces_quick(); 1848 1849 if (TraceParallelOldGCSummaryPhase) { 1850 tty->print_cr("summary_phase: after summarizing each space to self"); 1851 Universe::print(); 1852 NOT_PRODUCT(print_region_ranges()); 1853 if (Verbose) { 1854 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1855 } 1856 } 1857 1858 // The amount of live data that will end up in old space (assuming it fits). 1859 size_t old_space_total_live = 0; 1860 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1861 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1862 _space_info[id].space()->bottom()); 1863 } 1864 1865 MutableSpace* const old_space = _space_info[old_space_id].space(); 1866 const size_t old_capacity = old_space->capacity_in_words(); 1867 if (old_space_total_live > old_capacity) { 1868 // XXX - should also try to expand 1869 maximum_compaction = true; 1870 } 1871 #ifndef PRODUCT 1872 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1873 provoke_split(maximum_compaction); 1874 } 1875 #endif // #ifndef PRODUCT 1876 1877 // Old generations. 1878 summarize_space(old_space_id, maximum_compaction); 1879 1880 // Summarize the remaining spaces in the young gen. The initial target space 1881 // is the old gen. If a space does not fit entirely into the target, then the 1882 // remainder is compacted into the space itself and that space becomes the new 1883 // target. 1884 SpaceId dst_space_id = old_space_id; 1885 HeapWord* dst_space_end = old_space->end(); 1886 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1887 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1888 const MutableSpace* space = _space_info[id].space(); 1889 const size_t live = pointer_delta(_space_info[id].new_top(), 1890 space->bottom()); 1891 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1892 1893 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1894 SpaceId(id), space->bottom(), space->top());) 1895 if (live > 0 && live <= available) { 1896 // All the live data will fit. 1897 bool done = _summary_data.summarize(_space_info[id].split_info(), 1898 space->bottom(), space->top(), 1899 NULL, 1900 *new_top_addr, dst_space_end, 1901 new_top_addr); 1902 assert(done, "space must fit into old gen"); 1903 1904 // Reset the new_top value for the space. 1905 _space_info[id].set_new_top(space->bottom()); 1906 } else if (live > 0) { 1907 // Attempt to fit part of the source space into the target space. 1908 HeapWord* next_src_addr = NULL; 1909 bool done = _summary_data.summarize(_space_info[id].split_info(), 1910 space->bottom(), space->top(), 1911 &next_src_addr, 1912 *new_top_addr, dst_space_end, 1913 new_top_addr); 1914 assert(!done, "space should not fit into old gen"); 1915 assert(next_src_addr != NULL, "sanity"); 1916 1917 // The source space becomes the new target, so the remainder is compacted 1918 // within the space itself. 1919 dst_space_id = SpaceId(id); 1920 dst_space_end = space->end(); 1921 new_top_addr = _space_info[id].new_top_addr(); 1922 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1923 space->bottom(), dst_space_end, 1924 SpaceId(id), next_src_addr, space->top());) 1925 done = _summary_data.summarize(_space_info[id].split_info(), 1926 next_src_addr, space->top(), 1927 NULL, 1928 space->bottom(), dst_space_end, 1929 new_top_addr); 1930 assert(done, "space must fit when compacted into itself"); 1931 assert(*new_top_addr <= space->top(), "usage should not grow"); 1932 } 1933 } 1934 1935 if (TraceParallelOldGCSummaryPhase) { 1936 tty->print_cr("summary_phase: after final summarization"); 1937 Universe::print(); 1938 NOT_PRODUCT(print_region_ranges()); 1939 if (Verbose) { 1940 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1941 } 1942 } 1943 } 1944 1945 // This method should contain all heap-specific policy for invoking a full 1946 // collection. invoke_no_policy() will only attempt to compact the heap; it 1947 // will do nothing further. If we need to bail out for policy reasons, scavenge 1948 // before full gc, or any other specialized behavior, it needs to be added here. 1949 // 1950 // Note that this method should only be called from the vm_thread while at a 1951 // safepoint. 1952 // 1953 // Note that the all_soft_refs_clear flag in the collector policy 1954 // may be true because this method can be called without intervening 1955 // activity. For example when the heap space is tight and full measure 1956 // are being taken to free space. 1957 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1958 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1959 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1960 "should be in vm thread"); 1961 1962 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1963 GCCause::Cause gc_cause = heap->gc_cause(); 1964 assert(!heap->is_gc_active(), "not reentrant"); 1965 1966 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1967 IsGCActiveMark mark; 1968 1969 if (ScavengeBeforeFullGC) { 1970 PSScavenge::invoke_no_policy(); 1971 } 1972 1973 const bool clear_all_soft_refs = 1974 heap->collector_policy()->should_clear_all_soft_refs(); 1975 1976 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1977 maximum_heap_compaction); 1978 } 1979 1980 // This method contains no policy. You should probably 1981 // be calling invoke() instead. 1982 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1983 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1984 assert(ref_processor() != NULL, "Sanity"); 1985 1986 if (GC_locker::check_active_before_gc()) { 1987 return false; 1988 } 1989 1990 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1991 1992 _gc_timer.register_gc_start(); 1993 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1994 1995 TimeStamp marking_start; 1996 TimeStamp compaction_start; 1997 TimeStamp collection_exit; 1998 1999 GCCause::Cause gc_cause = heap->gc_cause(); 2000 PSYoungGen* young_gen = heap->young_gen(); 2001 PSOldGen* old_gen = heap->old_gen(); 2002 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2003 2004 // The scope of casr should end after code that can change 2005 // CollectorPolicy::_should_clear_all_soft_refs. 2006 ClearedAllSoftRefs casr(maximum_heap_compaction, 2007 heap->collector_policy()); 2008 2009 if (ZapUnusedHeapArea) { 2010 // Save information needed to minimize mangling 2011 heap->record_gen_tops_before_GC(); 2012 } 2013 2014 heap->pre_full_gc_dump(&_gc_timer); 2015 2016 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2017 2018 // Make sure data structures are sane, make the heap parsable, and do other 2019 // miscellaneous bookkeeping. 2020 PreGCValues pre_gc_values; 2021 pre_compact(&pre_gc_values); 2022 2023 // Get the compaction manager reserved for the VM thread. 2024 ParCompactionManager* const vmthread_cm = 2025 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2026 2027 // Place after pre_compact() where the number of invocations is incremented. 2028 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2029 2030 { 2031 ResourceMark rm; 2032 HandleMark hm; 2033 2034 // Set the number of GC threads to be used in this collection 2035 gc_task_manager()->set_active_gang(); 2036 gc_task_manager()->task_idle_workers(); 2037 heap->set_par_threads(gc_task_manager()->active_workers()); 2038 2039 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2040 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id()); 2041 TraceCollectorStats tcs(counters()); 2042 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2043 2044 if (TraceOldGenTime) accumulated_time()->start(); 2045 2046 // Let the size policy know we're starting 2047 size_policy->major_collection_begin(); 2048 2049 CodeCache::gc_prologue(); 2050 2051 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2052 2053 ref_processor()->enable_discovery(); 2054 ref_processor()->setup_policy(maximum_heap_compaction); 2055 2056 bool marked_for_unloading = false; 2057 2058 marking_start.update(); 2059 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2060 2061 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2062 && gc_cause == GCCause::_java_lang_system_gc; 2063 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2064 2065 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2066 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2067 2068 // adjust_roots() updates Universe::_intArrayKlassObj which is 2069 // needed by the compaction for filling holes in the dense prefix. 2070 adjust_roots(); 2071 2072 compaction_start.update(); 2073 compact(); 2074 2075 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2076 // done before resizing. 2077 post_compact(); 2078 2079 // Let the size policy know we're done 2080 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2081 2082 if (UseAdaptiveSizePolicy) { 2083 if (PrintAdaptiveSizePolicy) { 2084 gclog_or_tty->print("AdaptiveSizeStart: "); 2085 gclog_or_tty->stamp(); 2086 gclog_or_tty->print_cr(" collection: %d ", 2087 heap->total_collections()); 2088 if (Verbose) { 2089 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT 2090 " young_gen_capacity: " SIZE_FORMAT, 2091 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2092 } 2093 } 2094 2095 // Don't check if the size_policy is ready here. Let 2096 // the size_policy check that internally. 2097 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2098 ((gc_cause != GCCause::_java_lang_system_gc) || 2099 UseAdaptiveSizePolicyWithSystemGC)) { 2100 // Swap the survivor spaces if from_space is empty. The 2101 // resize_young_gen() called below is normally used after 2102 // a successful young GC and swapping of survivor spaces; 2103 // otherwise, it will fail to resize the young gen with 2104 // the current implementation. 2105 if (young_gen->from_space()->is_empty()) { 2106 young_gen->from_space()->clear(SpaceDecorator::Mangle); 2107 young_gen->swap_spaces(); 2108 } 2109 2110 // Calculate optimal free space amounts 2111 assert(young_gen->max_size() > 2112 young_gen->from_space()->capacity_in_bytes() + 2113 young_gen->to_space()->capacity_in_bytes(), 2114 "Sizes of space in young gen are out-of-bounds"); 2115 2116 size_t young_live = young_gen->used_in_bytes(); 2117 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2118 size_t old_live = old_gen->used_in_bytes(); 2119 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2120 size_t max_old_gen_size = old_gen->max_gen_size(); 2121 size_t max_eden_size = young_gen->max_size() - 2122 young_gen->from_space()->capacity_in_bytes() - 2123 young_gen->to_space()->capacity_in_bytes(); 2124 2125 // Used for diagnostics 2126 size_policy->clear_generation_free_space_flags(); 2127 2128 size_policy->compute_generations_free_space(young_live, 2129 eden_live, 2130 old_live, 2131 cur_eden, 2132 max_old_gen_size, 2133 max_eden_size, 2134 true /* full gc*/); 2135 2136 size_policy->check_gc_overhead_limit(young_live, 2137 eden_live, 2138 max_old_gen_size, 2139 max_eden_size, 2140 true /* full gc*/, 2141 gc_cause, 2142 heap->collector_policy()); 2143 2144 size_policy->decay_supplemental_growth(true /* full gc*/); 2145 2146 heap->resize_old_gen( 2147 size_policy->calculated_old_free_size_in_bytes()); 2148 2149 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 2150 size_policy->calculated_survivor_size_in_bytes()); 2151 } 2152 if (PrintAdaptiveSizePolicy) { 2153 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2154 heap->total_collections()); 2155 } 2156 } 2157 2158 if (UsePerfData) { 2159 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2160 counters->update_counters(); 2161 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2162 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2163 } 2164 2165 heap->resize_all_tlabs(); 2166 2167 // Resize the metaspace capacity after a collection 2168 MetaspaceGC::compute_new_size(); 2169 2170 if (TraceOldGenTime) accumulated_time()->stop(); 2171 2172 if (PrintGC) { 2173 if (PrintGCDetails) { 2174 // No GC timestamp here. This is after GC so it would be confusing. 2175 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2176 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2177 heap->print_heap_change(pre_gc_values.heap_used()); 2178 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2179 } else { 2180 heap->print_heap_change(pre_gc_values.heap_used()); 2181 } 2182 } 2183 2184 // Track memory usage and detect low memory 2185 MemoryService::track_memory_usage(); 2186 heap->update_counters(); 2187 gc_task_manager()->release_idle_workers(); 2188 } 2189 2190 #ifdef ASSERT 2191 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2192 ParCompactionManager* const cm = 2193 ParCompactionManager::manager_array(int(i)); 2194 assert(cm->marking_stack()->is_empty(), "should be empty"); 2195 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2196 } 2197 #endif // ASSERT 2198 2199 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2200 HandleMark hm; // Discard invalid handles created during verification 2201 Universe::verify(" VerifyAfterGC:"); 2202 } 2203 2204 // Re-verify object start arrays 2205 if (VerifyObjectStartArray && 2206 VerifyAfterGC) { 2207 old_gen->verify_object_start_array(); 2208 } 2209 2210 if (ZapUnusedHeapArea) { 2211 old_gen->object_space()->check_mangled_unused_area_complete(); 2212 } 2213 2214 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2215 2216 collection_exit.update(); 2217 2218 heap->print_heap_after_gc(); 2219 heap->trace_heap_after_gc(&_gc_tracer); 2220 2221 if (PrintGCTaskTimeStamps) { 2222 gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " 2223 JLONG_FORMAT, 2224 marking_start.ticks(), compaction_start.ticks(), 2225 collection_exit.ticks()); 2226 gc_task_manager()->print_task_time_stamps(); 2227 } 2228 2229 heap->post_full_gc_dump(&_gc_timer); 2230 2231 #ifdef TRACESPINNING 2232 ParallelTaskTerminator::print_termination_counts(); 2233 #endif 2234 2235 _gc_timer.register_gc_end(); 2236 2237 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2238 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2239 2240 return true; 2241 } 2242 2243 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2244 PSYoungGen* young_gen, 2245 PSOldGen* old_gen) { 2246 MutableSpace* const eden_space = young_gen->eden_space(); 2247 assert(!eden_space->is_empty(), "eden must be non-empty"); 2248 assert(young_gen->virtual_space()->alignment() == 2249 old_gen->virtual_space()->alignment(), "alignments do not match"); 2250 2251 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2252 return false; 2253 } 2254 2255 // Both generations must be completely committed. 2256 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2257 return false; 2258 } 2259 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2260 return false; 2261 } 2262 2263 // Figure out how much to take from eden. Include the average amount promoted 2264 // in the total; otherwise the next young gen GC will simply bail out to a 2265 // full GC. 2266 const size_t alignment = old_gen->virtual_space()->alignment(); 2267 const size_t eden_used = eden_space->used_in_bytes(); 2268 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2269 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2270 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2271 2272 if (absorb_size >= eden_capacity) { 2273 return false; // Must leave some space in eden. 2274 } 2275 2276 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2277 if (new_young_size < young_gen->min_gen_size()) { 2278 return false; // Respect young gen minimum size. 2279 } 2280 2281 if (TraceAdaptiveGCBoundary && Verbose) { 2282 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2283 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2284 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2285 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2286 absorb_size / K, 2287 eden_capacity / K, (eden_capacity - absorb_size) / K, 2288 young_gen->from_space()->used_in_bytes() / K, 2289 young_gen->to_space()->used_in_bytes() / K, 2290 young_gen->capacity_in_bytes() / K, new_young_size / K); 2291 } 2292 2293 // Fill the unused part of the old gen. 2294 MutableSpace* const old_space = old_gen->object_space(); 2295 HeapWord* const unused_start = old_space->top(); 2296 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2297 2298 if (unused_words > 0) { 2299 if (unused_words < CollectedHeap::min_fill_size()) { 2300 return false; // If the old gen cannot be filled, must give up. 2301 } 2302 CollectedHeap::fill_with_objects(unused_start, unused_words); 2303 } 2304 2305 // Take the live data from eden and set both top and end in the old gen to 2306 // eden top. (Need to set end because reset_after_change() mangles the region 2307 // from end to virtual_space->high() in debug builds). 2308 HeapWord* const new_top = eden_space->top(); 2309 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2310 absorb_size); 2311 young_gen->reset_after_change(); 2312 old_space->set_top(new_top); 2313 old_space->set_end(new_top); 2314 old_gen->reset_after_change(); 2315 2316 // Update the object start array for the filler object and the data from eden. 2317 ObjectStartArray* const start_array = old_gen->start_array(); 2318 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2319 start_array->allocate_block(p); 2320 } 2321 2322 // Could update the promoted average here, but it is not typically updated at 2323 // full GCs and the value to use is unclear. Something like 2324 // 2325 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2326 2327 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2328 return true; 2329 } 2330 2331 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2332 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2333 "shouldn't return NULL"); 2334 return ParallelScavengeHeap::gc_task_manager(); 2335 } 2336 2337 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2338 bool maximum_heap_compaction, 2339 ParallelOldTracer *gc_tracer) { 2340 // Recursively traverse all live objects and mark them 2341 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2342 2343 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2344 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2345 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2346 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2347 ParallelTaskTerminator terminator(active_gc_threads, qset); 2348 2349 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2350 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2351 2352 // Need new claim bits before marking starts. 2353 ClassLoaderDataGraph::clear_claimed_marks(); 2354 2355 { 2356 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2357 2358 ParallelScavengeHeap::ParStrongRootsScope psrs; 2359 2360 GCTaskQueue* q = GCTaskQueue::create(); 2361 2362 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2363 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2364 // We scan the thread roots in parallel 2365 Threads::create_thread_roots_marking_tasks(q); 2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2368 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2369 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2371 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2372 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2373 2374 if (active_gc_threads > 1) { 2375 for (uint j = 0; j < active_gc_threads; j++) { 2376 q->enqueue(new StealMarkingTask(&terminator)); 2377 } 2378 } 2379 2380 gc_task_manager()->execute_and_wait(q); 2381 } 2382 2383 // Process reference objects found during marking 2384 { 2385 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2386 2387 ReferenceProcessorStats stats; 2388 if (ref_processor()->processing_is_mt()) { 2389 RefProcTaskExecutor task_executor; 2390 stats = ref_processor()->process_discovered_references( 2391 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2392 &task_executor, &_gc_timer, _gc_tracer.gc_id()); 2393 } else { 2394 stats = ref_processor()->process_discovered_references( 2395 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2396 &_gc_timer, _gc_tracer.gc_id()); 2397 } 2398 2399 gc_tracer->report_gc_reference_stats(stats); 2400 } 2401 2402 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2403 2404 // This is the point where the entire marking should have completed. 2405 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2406 2407 // Follow system dictionary roots and unload classes. 2408 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2409 2410 // Unload nmethods. 2411 CodeCache::do_unloading(is_alive_closure(), purged_class); 2412 2413 // Prune dead klasses from subklass/sibling/implementor lists. 2414 Klass::clean_weak_klass_links(is_alive_closure()); 2415 2416 // Delete entries for dead interned strings. 2417 StringTable::unlink(is_alive_closure()); 2418 2419 // Clean up unreferenced symbols in symbol table. 2420 SymbolTable::unlink(); 2421 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2422 } 2423 2424 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2425 ClassLoaderData* cld) { 2426 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2427 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2428 2429 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2430 } 2431 2432 // This should be moved to the shared markSweep code! 2433 class PSAlwaysTrueClosure: public BoolObjectClosure { 2434 public: 2435 bool do_object_b(oop p) { return true; } 2436 }; 2437 static PSAlwaysTrueClosure always_true; 2438 2439 void PSParallelCompact::adjust_roots() { 2440 // Adjust the pointers to reflect the new locations 2441 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2442 2443 // Need new claim bits when tracing through and adjusting pointers. 2444 ClassLoaderDataGraph::clear_claimed_marks(); 2445 2446 // General strong roots. 2447 Universe::oops_do(adjust_pointer_closure()); 2448 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2449 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2450 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2451 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2452 FlatProfiler::oops_do(adjust_pointer_closure()); 2453 Management::oops_do(adjust_pointer_closure()); 2454 JvmtiExport::oops_do(adjust_pointer_closure()); 2455 SystemDictionary::oops_do(adjust_pointer_closure()); 2456 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2457 2458 // Now adjust pointers in remaining weak roots. (All of which should 2459 // have been cleared if they pointed to non-surviving objects.) 2460 // Global (weak) JNI handles 2461 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2462 2463 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2464 CodeCache::blobs_do(&adjust_from_blobs); 2465 StringTable::oops_do(adjust_pointer_closure()); 2466 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2467 // Roots were visited so references into the young gen in roots 2468 // may have been scanned. Process them also. 2469 // Should the reference processor have a span that excludes 2470 // young gen objects? 2471 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2472 } 2473 2474 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2475 uint parallel_gc_threads) 2476 { 2477 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2478 2479 // Find the threads that are active 2480 unsigned int which = 0; 2481 2482 const uint task_count = MAX2(parallel_gc_threads, 1U); 2483 for (uint j = 0; j < task_count; j++) { 2484 q->enqueue(new DrainStacksCompactionTask(j)); 2485 ParCompactionManager::verify_region_list_empty(j); 2486 // Set the region stacks variables to "no" region stack values 2487 // so that they will be recognized and needing a region stack 2488 // in the stealing tasks if they do not get one by executing 2489 // a draining stack. 2490 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2491 cm->set_region_stack(NULL); 2492 cm->set_region_stack_index((uint)max_uintx); 2493 } 2494 ParCompactionManager::reset_recycled_stack_index(); 2495 2496 // Find all regions that are available (can be filled immediately) and 2497 // distribute them to the thread stacks. The iteration is done in reverse 2498 // order (high to low) so the regions will be removed in ascending order. 2499 2500 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2501 2502 size_t fillable_regions = 0; // A count for diagnostic purposes. 2503 // A region index which corresponds to the tasks created above. 2504 // "which" must be 0 <= which < task_count 2505 2506 which = 0; 2507 // id + 1 is used to test termination so unsigned can 2508 // be used with an old_space_id == 0. 2509 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2510 SpaceInfo* const space_info = _space_info + id; 2511 MutableSpace* const space = space_info->space(); 2512 HeapWord* const new_top = space_info->new_top(); 2513 2514 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2515 const size_t end_region = 2516 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2517 2518 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2519 if (sd.region(cur)->claim_unsafe()) { 2520 ParCompactionManager::region_list_push(which, cur); 2521 2522 if (TraceParallelOldGCCompactionPhase && Verbose) { 2523 const size_t count_mod_8 = fillable_regions & 7; 2524 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2525 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2526 if (count_mod_8 == 7) gclog_or_tty->cr(); 2527 } 2528 2529 NOT_PRODUCT(++fillable_regions;) 2530 2531 // Assign regions to tasks in round-robin fashion. 2532 if (++which == task_count) { 2533 assert(which <= parallel_gc_threads, 2534 "Inconsistent number of workers"); 2535 which = 0; 2536 } 2537 } 2538 } 2539 } 2540 2541 if (TraceParallelOldGCCompactionPhase) { 2542 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2543 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions); 2544 } 2545 } 2546 2547 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2548 2549 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2550 uint parallel_gc_threads) { 2551 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2552 2553 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2554 2555 // Iterate over all the spaces adding tasks for updating 2556 // regions in the dense prefix. Assume that 1 gc thread 2557 // will work on opening the gaps and the remaining gc threads 2558 // will work on the dense prefix. 2559 unsigned int space_id; 2560 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2561 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2562 const MutableSpace* const space = _space_info[space_id].space(); 2563 2564 if (dense_prefix_end == space->bottom()) { 2565 // There is no dense prefix for this space. 2566 continue; 2567 } 2568 2569 // The dense prefix is before this region. 2570 size_t region_index_end_dense_prefix = 2571 sd.addr_to_region_idx(dense_prefix_end); 2572 RegionData* const dense_prefix_cp = 2573 sd.region(region_index_end_dense_prefix); 2574 assert(dense_prefix_end == space->end() || 2575 dense_prefix_cp->available() || 2576 dense_prefix_cp->claimed(), 2577 "The region after the dense prefix should always be ready to fill"); 2578 2579 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2580 2581 // Is there dense prefix work? 2582 size_t total_dense_prefix_regions = 2583 region_index_end_dense_prefix - region_index_start; 2584 // How many regions of the dense prefix should be given to 2585 // each thread? 2586 if (total_dense_prefix_regions > 0) { 2587 uint tasks_for_dense_prefix = 1; 2588 if (total_dense_prefix_regions <= 2589 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2590 // Don't over partition. This assumes that 2591 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2592 // so there are not many regions to process. 2593 tasks_for_dense_prefix = parallel_gc_threads; 2594 } else { 2595 // Over partition 2596 tasks_for_dense_prefix = parallel_gc_threads * 2597 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2598 } 2599 size_t regions_per_thread = total_dense_prefix_regions / 2600 tasks_for_dense_prefix; 2601 // Give each thread at least 1 region. 2602 if (regions_per_thread == 0) { 2603 regions_per_thread = 1; 2604 } 2605 2606 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2607 if (region_index_start >= region_index_end_dense_prefix) { 2608 break; 2609 } 2610 // region_index_end is not processed 2611 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2612 region_index_end_dense_prefix); 2613 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2614 region_index_start, 2615 region_index_end)); 2616 region_index_start = region_index_end; 2617 } 2618 } 2619 // This gets any part of the dense prefix that did not 2620 // fit evenly. 2621 if (region_index_start < region_index_end_dense_prefix) { 2622 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2623 region_index_start, 2624 region_index_end_dense_prefix)); 2625 } 2626 } 2627 } 2628 2629 void PSParallelCompact::enqueue_region_stealing_tasks( 2630 GCTaskQueue* q, 2631 ParallelTaskTerminator* terminator_ptr, 2632 uint parallel_gc_threads) { 2633 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2634 2635 // Once a thread has drained it's stack, it should try to steal regions from 2636 // other threads. 2637 if (parallel_gc_threads > 1) { 2638 for (uint j = 0; j < parallel_gc_threads; j++) { 2639 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2640 } 2641 } 2642 } 2643 2644 #ifdef ASSERT 2645 // Write a histogram of the number of times the block table was filled for a 2646 // region. 2647 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2648 { 2649 if (!TraceParallelOldGCCompactionPhase) return; 2650 2651 typedef ParallelCompactData::RegionData rd_t; 2652 ParallelCompactData& sd = summary_data(); 2653 2654 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2655 MutableSpace* const spc = _space_info[id].space(); 2656 if (spc->bottom() != spc->top()) { 2657 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2658 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2659 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2660 2661 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2662 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2663 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2664 2665 for (const rd_t* cur = beg; cur < end; ++cur) { 2666 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2667 } 2668 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2669 for (size_t i = 0; i < histo_len; ++i) { 2670 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2671 histo[i], 100.0 * histo[i] / region_cnt); 2672 } 2673 out->cr(); 2674 } 2675 } 2676 } 2677 #endif // #ifdef ASSERT 2678 2679 void PSParallelCompact::compact() { 2680 // trace("5"); 2681 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2682 2683 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2684 PSOldGen* old_gen = heap->old_gen(); 2685 old_gen->start_array()->reset(); 2686 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2687 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2688 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2689 ParallelTaskTerminator terminator(active_gc_threads, qset); 2690 2691 GCTaskQueue* q = GCTaskQueue::create(); 2692 enqueue_region_draining_tasks(q, active_gc_threads); 2693 enqueue_dense_prefix_tasks(q, active_gc_threads); 2694 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2695 2696 { 2697 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2698 2699 gc_task_manager()->execute_and_wait(q); 2700 2701 #ifdef ASSERT 2702 // Verify that all regions have been processed before the deferred updates. 2703 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2704 verify_complete(SpaceId(id)); 2705 } 2706 #endif 2707 } 2708 2709 { 2710 // Update the deferred objects, if any. Any compaction manager can be used. 2711 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id()); 2712 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2713 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2714 update_deferred_objects(cm, SpaceId(id)); 2715 } 2716 } 2717 2718 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2719 } 2720 2721 #ifdef ASSERT 2722 void PSParallelCompact::verify_complete(SpaceId space_id) { 2723 // All Regions between space bottom() to new_top() should be marked as filled 2724 // and all Regions between new_top() and top() should be available (i.e., 2725 // should have been emptied). 2726 ParallelCompactData& sd = summary_data(); 2727 SpaceInfo si = _space_info[space_id]; 2728 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2729 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2730 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2731 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2732 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2733 2734 bool issued_a_warning = false; 2735 2736 size_t cur_region; 2737 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2738 const RegionData* const c = sd.region(cur_region); 2739 if (!c->completed()) { 2740 warning("region " SIZE_FORMAT " not filled: " 2741 "destination_count=%u", 2742 cur_region, c->destination_count()); 2743 issued_a_warning = true; 2744 } 2745 } 2746 2747 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2748 const RegionData* const c = sd.region(cur_region); 2749 if (!c->available()) { 2750 warning("region " SIZE_FORMAT " not empty: " 2751 "destination_count=%u", 2752 cur_region, c->destination_count()); 2753 issued_a_warning = true; 2754 } 2755 } 2756 2757 if (issued_a_warning) { 2758 print_region_ranges(); 2759 } 2760 } 2761 #endif // #ifdef ASSERT 2762 2763 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2764 _start_array->allocate_block(addr); 2765 compaction_manager()->update_contents(oop(addr)); 2766 } 2767 2768 // Update interior oops in the ranges of regions [beg_region, end_region). 2769 void 2770 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2771 SpaceId space_id, 2772 size_t beg_region, 2773 size_t end_region) { 2774 ParallelCompactData& sd = summary_data(); 2775 ParMarkBitMap* const mbm = mark_bitmap(); 2776 2777 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2778 HeapWord* const end_addr = sd.region_to_addr(end_region); 2779 assert(beg_region <= end_region, "bad region range"); 2780 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2781 2782 #ifdef ASSERT 2783 // Claim the regions to avoid triggering an assert when they are marked as 2784 // filled. 2785 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2786 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2787 } 2788 #endif // #ifdef ASSERT 2789 2790 if (beg_addr != space(space_id)->bottom()) { 2791 // Find the first live object or block of dead space that *starts* in this 2792 // range of regions. If a partial object crosses onto the region, skip it; 2793 // it will be marked for 'deferred update' when the object head is 2794 // processed. If dead space crosses onto the region, it is also skipped; it 2795 // will be filled when the prior region is processed. If neither of those 2796 // apply, the first word in the region is the start of a live object or dead 2797 // space. 2798 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2799 const RegionData* const cp = sd.region(beg_region); 2800 if (cp->partial_obj_size() != 0) { 2801 beg_addr = sd.partial_obj_end(beg_region); 2802 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2803 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2804 } 2805 } 2806 2807 if (beg_addr < end_addr) { 2808 // A live object or block of dead space starts in this range of Regions. 2809 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2810 2811 // Create closures and iterate. 2812 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2813 FillClosure fill_closure(cm, space_id); 2814 ParMarkBitMap::IterationStatus status; 2815 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2816 dense_prefix_end); 2817 if (status == ParMarkBitMap::incomplete) { 2818 update_closure.do_addr(update_closure.source()); 2819 } 2820 } 2821 2822 // Mark the regions as filled. 2823 RegionData* const beg_cp = sd.region(beg_region); 2824 RegionData* const end_cp = sd.region(end_region); 2825 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2826 cp->set_completed(); 2827 } 2828 } 2829 2830 // Return the SpaceId for the space containing addr. If addr is not in the 2831 // heap, last_space_id is returned. In debug mode it expects the address to be 2832 // in the heap and asserts such. 2833 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2834 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2835 2836 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2837 if (_space_info[id].space()->contains(addr)) { 2838 return SpaceId(id); 2839 } 2840 } 2841 2842 assert(false, "no space contains the addr"); 2843 return last_space_id; 2844 } 2845 2846 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2847 SpaceId id) { 2848 assert(id < last_space_id, "bad space id"); 2849 2850 ParallelCompactData& sd = summary_data(); 2851 const SpaceInfo* const space_info = _space_info + id; 2852 ObjectStartArray* const start_array = space_info->start_array(); 2853 2854 const MutableSpace* const space = space_info->space(); 2855 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2856 HeapWord* const beg_addr = space_info->dense_prefix(); 2857 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2858 2859 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2860 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2861 const RegionData* cur_region; 2862 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2863 HeapWord* const addr = cur_region->deferred_obj_addr(); 2864 if (addr != NULL) { 2865 if (start_array != NULL) { 2866 start_array->allocate_block(addr); 2867 } 2868 cm->update_contents(oop(addr)); 2869 assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)))); 2870 } 2871 } 2872 } 2873 2874 // Skip over count live words starting from beg, and return the address of the 2875 // next live word. Unless marked, the word corresponding to beg is assumed to 2876 // be dead. Callers must either ensure beg does not correspond to the middle of 2877 // an object, or account for those live words in some other way. Callers must 2878 // also ensure that there are enough live words in the range [beg, end) to skip. 2879 HeapWord* 2880 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2881 { 2882 assert(count > 0, "sanity"); 2883 2884 ParMarkBitMap* m = mark_bitmap(); 2885 idx_t bits_to_skip = m->words_to_bits(count); 2886 idx_t cur_beg = m->addr_to_bit(beg); 2887 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2888 2889 do { 2890 cur_beg = m->find_obj_beg(cur_beg, search_end); 2891 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2892 const size_t obj_bits = cur_end - cur_beg + 1; 2893 if (obj_bits > bits_to_skip) { 2894 return m->bit_to_addr(cur_beg + bits_to_skip); 2895 } 2896 bits_to_skip -= obj_bits; 2897 cur_beg = cur_end + 1; 2898 } while (bits_to_skip > 0); 2899 2900 // Skipping the desired number of words landed just past the end of an object. 2901 // Find the start of the next object. 2902 cur_beg = m->find_obj_beg(cur_beg, search_end); 2903 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2904 return m->bit_to_addr(cur_beg); 2905 } 2906 2907 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2908 SpaceId src_space_id, 2909 size_t src_region_idx) 2910 { 2911 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2912 2913 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2914 if (split_info.dest_region_addr() == dest_addr) { 2915 // The partial object ending at the split point contains the first word to 2916 // be copied to dest_addr. 2917 return split_info.first_src_addr(); 2918 } 2919 2920 const ParallelCompactData& sd = summary_data(); 2921 ParMarkBitMap* const bitmap = mark_bitmap(); 2922 const size_t RegionSize = ParallelCompactData::RegionSize; 2923 2924 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2925 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2926 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2927 HeapWord* const src_region_destination = src_region_ptr->destination(); 2928 2929 assert(dest_addr >= src_region_destination, "wrong src region"); 2930 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2931 2932 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2933 HeapWord* const src_region_end = src_region_beg + RegionSize; 2934 2935 HeapWord* addr = src_region_beg; 2936 if (dest_addr == src_region_destination) { 2937 // Return the first live word in the source region. 2938 if (partial_obj_size == 0) { 2939 addr = bitmap->find_obj_beg(addr, src_region_end); 2940 assert(addr < src_region_end, "no objects start in src region"); 2941 } 2942 return addr; 2943 } 2944 2945 // Must skip some live data. 2946 size_t words_to_skip = dest_addr - src_region_destination; 2947 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2948 2949 if (partial_obj_size >= words_to_skip) { 2950 // All the live words to skip are part of the partial object. 2951 addr += words_to_skip; 2952 if (partial_obj_size == words_to_skip) { 2953 // Find the first live word past the partial object. 2954 addr = bitmap->find_obj_beg(addr, src_region_end); 2955 assert(addr < src_region_end, "wrong src region"); 2956 } 2957 return addr; 2958 } 2959 2960 // Skip over the partial object (if any). 2961 if (partial_obj_size != 0) { 2962 words_to_skip -= partial_obj_size; 2963 addr += partial_obj_size; 2964 } 2965 2966 // Skip over live words due to objects that start in the region. 2967 addr = skip_live_words(addr, src_region_end, words_to_skip); 2968 assert(addr < src_region_end, "wrong src region"); 2969 return addr; 2970 } 2971 2972 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2973 SpaceId src_space_id, 2974 size_t beg_region, 2975 HeapWord* end_addr) 2976 { 2977 ParallelCompactData& sd = summary_data(); 2978 2979 #ifdef ASSERT 2980 MutableSpace* const src_space = _space_info[src_space_id].space(); 2981 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2982 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2983 "src_space_id does not match beg_addr"); 2984 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2985 "src_space_id does not match end_addr"); 2986 #endif // #ifdef ASSERT 2987 2988 RegionData* const beg = sd.region(beg_region); 2989 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2990 2991 // Regions up to new_top() are enqueued if they become available. 2992 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2993 RegionData* const enqueue_end = 2994 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2995 2996 for (RegionData* cur = beg; cur < end; ++cur) { 2997 assert(cur->data_size() > 0, "region must have live data"); 2998 cur->decrement_destination_count(); 2999 if (cur < enqueue_end && cur->available() && cur->claim()) { 3000 cm->push_region(sd.region(cur)); 3001 } 3002 } 3003 } 3004 3005 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3006 SpaceId& src_space_id, 3007 HeapWord*& src_space_top, 3008 HeapWord* end_addr) 3009 { 3010 typedef ParallelCompactData::RegionData RegionData; 3011 3012 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3013 const size_t region_size = ParallelCompactData::RegionSize; 3014 3015 size_t src_region_idx = 0; 3016 3017 // Skip empty regions (if any) up to the top of the space. 3018 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3019 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3020 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3021 const RegionData* const top_region_ptr = 3022 sd.addr_to_region_ptr(top_aligned_up); 3023 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3024 ++src_region_ptr; 3025 } 3026 3027 if (src_region_ptr < top_region_ptr) { 3028 // The next source region is in the current space. Update src_region_idx 3029 // and the source address to match src_region_ptr. 3030 src_region_idx = sd.region(src_region_ptr); 3031 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3032 if (src_region_addr > closure.source()) { 3033 closure.set_source(src_region_addr); 3034 } 3035 return src_region_idx; 3036 } 3037 3038 // Switch to a new source space and find the first non-empty region. 3039 unsigned int space_id = src_space_id + 1; 3040 assert(space_id < last_space_id, "not enough spaces"); 3041 3042 HeapWord* const destination = closure.destination(); 3043 3044 do { 3045 MutableSpace* space = _space_info[space_id].space(); 3046 HeapWord* const bottom = space->bottom(); 3047 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3048 3049 // Iterate over the spaces that do not compact into themselves. 3050 if (bottom_cp->destination() != bottom) { 3051 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3052 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3053 3054 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3055 if (src_cp->live_obj_size() > 0) { 3056 // Found it. 3057 assert(src_cp->destination() == destination, 3058 "first live obj in the space must match the destination"); 3059 assert(src_cp->partial_obj_size() == 0, 3060 "a space cannot begin with a partial obj"); 3061 3062 src_space_id = SpaceId(space_id); 3063 src_space_top = space->top(); 3064 const size_t src_region_idx = sd.region(src_cp); 3065 closure.set_source(sd.region_to_addr(src_region_idx)); 3066 return src_region_idx; 3067 } else { 3068 assert(src_cp->data_size() == 0, "sanity"); 3069 } 3070 } 3071 } 3072 } while (++space_id < last_space_id); 3073 3074 assert(false, "no source region was found"); 3075 return 0; 3076 } 3077 3078 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3079 { 3080 typedef ParMarkBitMap::IterationStatus IterationStatus; 3081 const size_t RegionSize = ParallelCompactData::RegionSize; 3082 ParMarkBitMap* const bitmap = mark_bitmap(); 3083 ParallelCompactData& sd = summary_data(); 3084 RegionData* const region_ptr = sd.region(region_idx); 3085 3086 // Get the items needed to construct the closure. 3087 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3088 SpaceId dest_space_id = space_id(dest_addr); 3089 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3090 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3091 assert(dest_addr < new_top, "sanity"); 3092 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3093 3094 // Get the source region and related info. 3095 size_t src_region_idx = region_ptr->source_region(); 3096 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3097 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3098 3099 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3100 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3101 3102 // Adjust src_region_idx to prepare for decrementing destination counts (the 3103 // destination count is not decremented when a region is copied to itself). 3104 if (src_region_idx == region_idx) { 3105 src_region_idx += 1; 3106 } 3107 3108 if (bitmap->is_unmarked(closure.source())) { 3109 // The first source word is in the middle of an object; copy the remainder 3110 // of the object or as much as will fit. The fact that pointer updates were 3111 // deferred will be noted when the object header is processed. 3112 HeapWord* const old_src_addr = closure.source(); 3113 closure.copy_partial_obj(); 3114 if (closure.is_full()) { 3115 decrement_destination_counts(cm, src_space_id, src_region_idx, 3116 closure.source()); 3117 region_ptr->set_deferred_obj_addr(NULL); 3118 region_ptr->set_completed(); 3119 return; 3120 } 3121 3122 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3123 if (sd.region_align_down(old_src_addr) != end_addr) { 3124 // The partial object was copied from more than one source region. 3125 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3126 3127 // Move to the next source region, possibly switching spaces as well. All 3128 // args except end_addr may be modified. 3129 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3130 end_addr); 3131 } 3132 } 3133 3134 do { 3135 HeapWord* const cur_addr = closure.source(); 3136 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3137 src_space_top); 3138 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3139 3140 if (status == ParMarkBitMap::incomplete) { 3141 // The last obj that starts in the source region does not end in the 3142 // region. 3143 assert(closure.source() < end_addr, "sanity"); 3144 HeapWord* const obj_beg = closure.source(); 3145 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3146 src_space_top); 3147 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3148 if (obj_end < range_end) { 3149 // The end was found; the entire object will fit. 3150 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3151 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3152 } else { 3153 // The end was not found; the object will not fit. 3154 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3155 status = ParMarkBitMap::would_overflow; 3156 } 3157 } 3158 3159 if (status == ParMarkBitMap::would_overflow) { 3160 // The last object did not fit. Note that interior oop updates were 3161 // deferred, then copy enough of the object to fill the region. 3162 region_ptr->set_deferred_obj_addr(closure.destination()); 3163 status = closure.copy_until_full(); // copies from closure.source() 3164 3165 decrement_destination_counts(cm, src_space_id, src_region_idx, 3166 closure.source()); 3167 region_ptr->set_completed(); 3168 return; 3169 } 3170 3171 if (status == ParMarkBitMap::full) { 3172 decrement_destination_counts(cm, src_space_id, src_region_idx, 3173 closure.source()); 3174 region_ptr->set_deferred_obj_addr(NULL); 3175 region_ptr->set_completed(); 3176 return; 3177 } 3178 3179 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3180 3181 // Move to the next source region, possibly switching spaces as well. All 3182 // args except end_addr may be modified. 3183 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3184 end_addr); 3185 } while (true); 3186 } 3187 3188 void PSParallelCompact::fill_blocks(size_t region_idx) 3189 { 3190 // Fill in the block table elements for the specified region. Each block 3191 // table element holds the number of live words in the region that are to the 3192 // left of the first object that starts in the block. Thus only blocks in 3193 // which an object starts need to be filled. 3194 // 3195 // The algorithm scans the section of the bitmap that corresponds to the 3196 // region, keeping a running total of the live words. When an object start is 3197 // found, if it's the first to start in the block that contains it, the 3198 // current total is written to the block table element. 3199 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 3200 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 3201 const size_t RegionSize = ParallelCompactData::RegionSize; 3202 3203 ParallelCompactData& sd = summary_data(); 3204 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 3205 if (partial_obj_size >= RegionSize) { 3206 return; // No objects start in this region. 3207 } 3208 3209 // Ensure the first loop iteration decides that the block has changed. 3210 size_t cur_block = sd.block_count(); 3211 3212 const ParMarkBitMap* const bitmap = mark_bitmap(); 3213 3214 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 3215 assert((size_t)1 << Log2BitsPerBlock == 3216 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 3217 3218 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 3219 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 3220 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 3221 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 3222 while (beg_bit < range_end) { 3223 const size_t new_block = beg_bit >> Log2BitsPerBlock; 3224 if (new_block != cur_block) { 3225 cur_block = new_block; 3226 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 3227 } 3228 3229 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 3230 if (end_bit < range_end - 1) { 3231 live_bits += end_bit - beg_bit + 1; 3232 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 3233 } else { 3234 return; 3235 } 3236 } 3237 } 3238 3239 void 3240 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3241 const MutableSpace* sp = space(space_id); 3242 if (sp->is_empty()) { 3243 return; 3244 } 3245 3246 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3247 ParMarkBitMap* const bitmap = mark_bitmap(); 3248 HeapWord* const dp_addr = dense_prefix(space_id); 3249 HeapWord* beg_addr = sp->bottom(); 3250 HeapWord* end_addr = sp->top(); 3251 3252 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3253 3254 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3255 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3256 if (beg_region < dp_region) { 3257 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3258 } 3259 3260 // The destination of the first live object that starts in the region is one 3261 // past the end of the partial object entering the region (if any). 3262 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3263 HeapWord* const new_top = _space_info[space_id].new_top(); 3264 assert(new_top >= dest_addr, "bad new_top value"); 3265 const size_t words = pointer_delta(new_top, dest_addr); 3266 3267 if (words > 0) { 3268 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3269 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3270 3271 ParMarkBitMap::IterationStatus status; 3272 status = bitmap->iterate(&closure, dest_addr, end_addr); 3273 assert(status == ParMarkBitMap::full, "iteration not complete"); 3274 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3275 "live objects skipped because closure is full"); 3276 } 3277 } 3278 3279 jlong PSParallelCompact::millis_since_last_gc() { 3280 // We need a monotonically non-decreasing time in ms but 3281 // os::javaTimeMillis() does not guarantee monotonicity. 3282 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3283 jlong ret_val = now - _time_of_last_gc; 3284 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3285 if (ret_val < 0) { 3286 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);) 3287 return 0; 3288 } 3289 return ret_val; 3290 } 3291 3292 void PSParallelCompact::reset_millis_since_last_gc() { 3293 // We need a monotonically non-decreasing time in ms but 3294 // os::javaTimeMillis() does not guarantee monotonicity. 3295 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3296 } 3297 3298 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3299 { 3300 if (source() != destination()) { 3301 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3302 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3303 } 3304 update_state(words_remaining()); 3305 assert(is_full(), "sanity"); 3306 return ParMarkBitMap::full; 3307 } 3308 3309 void MoveAndUpdateClosure::copy_partial_obj() 3310 { 3311 size_t words = words_remaining(); 3312 3313 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3314 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3315 if (end_addr < range_end) { 3316 words = bitmap()->obj_size(source(), end_addr); 3317 } 3318 3319 // This test is necessary; if omitted, the pointer updates to a partial object 3320 // that crosses the dense prefix boundary could be overwritten. 3321 if (source() != destination()) { 3322 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3323 Copy::aligned_conjoint_words(source(), destination(), words); 3324 } 3325 update_state(words); 3326 } 3327 3328 void InstanceKlass::oop_pc_update_pointers(oop obj) { 3329 oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3330 } 3331 3332 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) { 3333 InstanceKlass::oop_pc_update_pointers(obj); 3334 3335 oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3336 } 3337 3338 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) { 3339 InstanceKlass::oop_pc_update_pointers(obj); 3340 } 3341 3342 #ifdef ASSERT 3343 template <class T> static void trace_reference_gc(const char *s, oop obj, 3344 T* referent_addr, 3345 T* next_addr, 3346 T* discovered_addr) { 3347 if(TraceReferenceGC && PrintGCDetails) { 3348 gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj)); 3349 gclog_or_tty->print_cr(" referent_addr/* " PTR_FORMAT " / " 3350 PTR_FORMAT, p2i(referent_addr), 3351 referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3352 gclog_or_tty->print_cr(" next_addr/* " PTR_FORMAT " / " 3353 PTR_FORMAT, p2i(next_addr), 3354 next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3355 gclog_or_tty->print_cr(" discovered_addr/* " PTR_FORMAT " / " 3356 PTR_FORMAT, p2i(discovered_addr), 3357 discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3358 } 3359 } 3360 #endif 3361 3362 template <class T> 3363 static void oop_pc_update_pointers_specialized(oop obj) { 3364 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3365 PSParallelCompact::adjust_pointer(referent_addr); 3366 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3367 PSParallelCompact::adjust_pointer(next_addr); 3368 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3369 PSParallelCompact::adjust_pointer(discovered_addr); 3370 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3371 referent_addr, next_addr, discovered_addr);) 3372 } 3373 3374 void InstanceRefKlass::oop_pc_update_pointers(oop obj) { 3375 InstanceKlass::oop_pc_update_pointers(obj); 3376 3377 if (UseCompressedOops) { 3378 oop_pc_update_pointers_specialized<narrowOop>(obj); 3379 } else { 3380 oop_pc_update_pointers_specialized<oop>(obj); 3381 } 3382 } 3383 3384 void ObjArrayKlass::oop_pc_update_pointers(oop obj) { 3385 assert(obj->is_objArray(), "obj must be obj array"); 3386 oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure()); 3387 } 3388 3389 void TypeArrayKlass::oop_pc_update_pointers(oop obj) { 3390 assert(obj->is_typeArray(),"must be a type array"); 3391 } 3392 3393 ParMarkBitMapClosure::IterationStatus 3394 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3395 assert(destination() != NULL, "sanity"); 3396 assert(bitmap()->obj_size(addr) == words, "bad size"); 3397 3398 _source = addr; 3399 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3400 destination(), "wrong destination"); 3401 3402 if (words > words_remaining()) { 3403 return ParMarkBitMap::would_overflow; 3404 } 3405 3406 // The start_array must be updated even if the object is not moving. 3407 if (_start_array != NULL) { 3408 _start_array->allocate_block(destination()); 3409 } 3410 3411 if (destination() != source()) { 3412 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3413 Copy::aligned_conjoint_words(source(), destination(), words); 3414 } 3415 3416 oop moved_oop = (oop) destination(); 3417 compaction_manager()->update_contents(moved_oop); 3418 assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop))); 3419 3420 update_state(words); 3421 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3422 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3423 } 3424 3425 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3426 ParCompactionManager* cm, 3427 PSParallelCompact::SpaceId space_id) : 3428 ParMarkBitMapClosure(mbm, cm), 3429 _space_id(space_id), 3430 _start_array(PSParallelCompact::start_array(space_id)) 3431 { 3432 } 3433 3434 // Updates the references in the object to their new values. 3435 ParMarkBitMapClosure::IterationStatus 3436 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3437 do_addr(addr); 3438 return ParMarkBitMap::incomplete; 3439 }