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