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