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