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