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