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 (!UseParallelGC) { 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 // Dummy counter 875 new CollectorCounters("dummy", 2); 876 877 // Initialize static fields in ParCompactionManager. 878 ParCompactionManager::initialize(mark_bitmap()); 879 } 880 881 bool PSParallelCompact::initialize() { 882 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 883 MemRegion mr = heap->reserved_region(); 884 885 // Was the old gen get allocated successfully? 886 if (!heap->old_gen()->is_allocated()) { 887 return false; 888 } 889 890 initialize_space_info(); 891 initialize_dead_wood_limiter(); 892 893 if (!_mark_bitmap.initialize(mr)) { 894 vm_shutdown_during_initialization( 895 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 896 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 897 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 898 return false; 899 } 900 901 if (!_summary_data.initialize(mr)) { 902 vm_shutdown_during_initialization( 903 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 904 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 905 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 906 return false; 907 } 908 909 return true; 910 } 911 912 void PSParallelCompact::initialize_space_info() 913 { 914 memset(&_space_info, 0, sizeof(_space_info)); 915 916 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 917 PSYoungGen* young_gen = heap->young_gen(); 918 919 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 920 _space_info[eden_space_id].set_space(young_gen->eden_space()); 921 _space_info[from_space_id].set_space(young_gen->from_space()); 922 _space_info[to_space_id].set_space(young_gen->to_space()); 923 924 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 925 } 926 927 void PSParallelCompact::initialize_dead_wood_limiter() 928 { 929 const size_t max = 100; 930 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 931 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 932 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 933 DEBUG_ONLY(_dwl_initialized = true;) 934 _dwl_adjustment = normal_distribution(1.0); 935 } 936 937 void 938 PSParallelCompact::clear_data_covering_space(SpaceId id) 939 { 940 // At this point, top is the value before GC, new_top() is the value that will 941 // be set at the end of GC. The marking bitmap is cleared to top; nothing 942 // should be marked above top. The summary data is cleared to the larger of 943 // top & new_top. 944 MutableSpace* const space = _space_info[id].space(); 945 HeapWord* const bot = space->bottom(); 946 HeapWord* const top = space->top(); 947 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 948 949 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 950 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 951 _mark_bitmap.clear_range(beg_bit, end_bit); 952 953 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 954 const size_t end_region = 955 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 956 _summary_data.clear_range(beg_region, end_region); 957 958 // Clear the data used to 'split' regions. 959 SplitInfo& split_info = _space_info[id].split_info(); 960 if (split_info.is_valid()) { 961 split_info.clear(); 962 } 963 DEBUG_ONLY(split_info.verify_clear();) 964 } 965 966 void PSParallelCompact::pre_compact() 967 { 968 // Update the from & to space pointers in space_info, since they are swapped 969 // at each young gen gc. Do the update unconditionally (even though a 970 // promotion failure does not swap spaces) because an unknown number of young 971 // collections will have swapped the spaces an unknown number of times. 972 GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer); 973 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 974 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 975 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 976 977 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 978 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 979 980 // Increment the invocation count 981 heap->increment_total_collections(true); 982 983 // We need to track unique mark sweep invocations as well. 984 _total_invocations++; 985 986 heap->print_heap_before_gc(); 987 heap->trace_heap_before_gc(&_gc_tracer); 988 989 // Fill in TLABs 990 heap->accumulate_statistics_all_tlabs(); 991 heap->ensure_parsability(true); // retire TLABs 992 993 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 994 HandleMark hm; // Discard invalid handles created during verification 995 Universe::verify("Before GC"); 996 } 997 998 // Verify object start arrays 999 if (VerifyObjectStartArray && 1000 VerifyBeforeGC) { 1001 heap->old_gen()->verify_object_start_array(); 1002 } 1003 1004 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1005 DEBUG_ONLY(summary_data().verify_clear();) 1006 1007 // Have worker threads release resources the next time they run a task. 1008 gc_task_manager()->release_all_resources(); 1009 1010 ParCompactionManager::reset_all_bitmap_query_caches(); 1011 } 1012 1013 void PSParallelCompact::post_compact() 1014 { 1015 GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer); 1016 1017 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1018 // Clear the marking bitmap, summary data and split info. 1019 clear_data_covering_space(SpaceId(id)); 1020 // Update top(). Must be done after clearing the bitmap and summary data. 1021 _space_info[id].publish_new_top(); 1022 } 1023 1024 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1025 MutableSpace* const from_space = _space_info[from_space_id].space(); 1026 MutableSpace* const to_space = _space_info[to_space_id].space(); 1027 1028 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1029 bool eden_empty = eden_space->is_empty(); 1030 if (!eden_empty) { 1031 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1032 heap->young_gen(), heap->old_gen()); 1033 } 1034 1035 // Update heap occupancy information which is used as input to the soft ref 1036 // clearing policy at the next gc. 1037 Universe::update_heap_info_at_gc(); 1038 1039 bool young_gen_empty = eden_empty && from_space->is_empty() && 1040 to_space->is_empty(); 1041 1042 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1043 MemRegion old_mr = heap->old_gen()->reserved(); 1044 if (young_gen_empty) { 1045 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1046 } else { 1047 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1048 } 1049 1050 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1051 ClassLoaderDataGraph::purge(); 1052 MetaspaceAux::verify_metrics(); 1053 1054 CodeCache::gc_epilogue(); 1055 JvmtiExport::gc_epilogue(); 1056 1057 #if defined(COMPILER2) || INCLUDE_JVMCI 1058 DerivedPointerTable::update_pointers(); 1059 #endif 1060 1061 ref_processor()->enqueue_discovered_references(NULL); 1062 1063 if (ZapUnusedHeapArea) { 1064 heap->gen_mangle_unused_area(); 1065 } 1066 1067 // Update time of last GC 1068 reset_millis_since_last_gc(); 1069 } 1070 1071 HeapWord* 1072 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1073 bool maximum_compaction) 1074 { 1075 const size_t region_size = ParallelCompactData::RegionSize; 1076 const ParallelCompactData& sd = summary_data(); 1077 1078 const MutableSpace* const space = _space_info[id].space(); 1079 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1080 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1081 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1082 1083 // Skip full regions at the beginning of the space--they are necessarily part 1084 // of the dense prefix. 1085 size_t full_count = 0; 1086 const RegionData* cp; 1087 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1088 ++full_count; 1089 } 1090 1091 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1092 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1093 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1094 if (maximum_compaction || cp == end_cp || interval_ended) { 1095 _maximum_compaction_gc_num = total_invocations(); 1096 return sd.region_to_addr(cp); 1097 } 1098 1099 HeapWord* const new_top = _space_info[id].new_top(); 1100 const size_t space_live = pointer_delta(new_top, space->bottom()); 1101 const size_t space_used = space->used_in_words(); 1102 const size_t space_capacity = space->capacity_in_words(); 1103 1104 const double cur_density = double(space_live) / space_capacity; 1105 const double deadwood_density = 1106 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1107 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1108 1109 if (TraceParallelOldGCDensePrefix) { 1110 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1111 cur_density, deadwood_density, deadwood_goal); 1112 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1113 "space_cap=" SIZE_FORMAT, 1114 space_live, space_used, 1115 space_capacity); 1116 } 1117 1118 // XXX - Use binary search? 1119 HeapWord* dense_prefix = sd.region_to_addr(cp); 1120 const RegionData* full_cp = cp; 1121 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1122 while (cp < end_cp) { 1123 HeapWord* region_destination = cp->destination(); 1124 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1125 if (TraceParallelOldGCDensePrefix && Verbose) { 1126 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1127 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1128 sd.region(cp), p2i(region_destination), 1129 p2i(dense_prefix), cur_deadwood); 1130 } 1131 1132 if (cur_deadwood >= deadwood_goal) { 1133 // Found the region that has the correct amount of deadwood to the left. 1134 // This typically occurs after crossing a fairly sparse set of regions, so 1135 // iterate backwards over those sparse regions, looking for the region 1136 // that has the lowest density of live objects 'to the right.' 1137 size_t space_to_left = sd.region(cp) * region_size; 1138 size_t live_to_left = space_to_left - cur_deadwood; 1139 size_t space_to_right = space_capacity - space_to_left; 1140 size_t live_to_right = space_live - live_to_left; 1141 double density_to_right = double(live_to_right) / space_to_right; 1142 while (cp > full_cp) { 1143 --cp; 1144 const size_t prev_region_live_to_right = live_to_right - 1145 cp->data_size(); 1146 const size_t prev_region_space_to_right = space_to_right + region_size; 1147 double prev_region_density_to_right = 1148 double(prev_region_live_to_right) / prev_region_space_to_right; 1149 if (density_to_right <= prev_region_density_to_right) { 1150 return dense_prefix; 1151 } 1152 if (TraceParallelOldGCDensePrefix && Verbose) { 1153 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1154 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1155 prev_region_density_to_right); 1156 } 1157 dense_prefix -= region_size; 1158 live_to_right = prev_region_live_to_right; 1159 space_to_right = prev_region_space_to_right; 1160 density_to_right = prev_region_density_to_right; 1161 } 1162 return dense_prefix; 1163 } 1164 1165 dense_prefix += region_size; 1166 ++cp; 1167 } 1168 1169 return dense_prefix; 1170 } 1171 1172 #ifndef PRODUCT 1173 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1174 const SpaceId id, 1175 const bool maximum_compaction, 1176 HeapWord* const addr) 1177 { 1178 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1179 RegionData* const cp = summary_data().region(region_idx); 1180 const MutableSpace* const space = _space_info[id].space(); 1181 HeapWord* const new_top = _space_info[id].new_top(); 1182 1183 const size_t space_live = pointer_delta(new_top, space->bottom()); 1184 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1185 const size_t space_cap = space->capacity_in_words(); 1186 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1187 const size_t live_to_right = new_top - cp->destination(); 1188 const size_t dead_to_right = space->top() - addr - live_to_right; 1189 1190 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1191 "spl=" SIZE_FORMAT " " 1192 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1193 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1194 " ratio=%10.8f", 1195 algorithm, p2i(addr), region_idx, 1196 space_live, 1197 dead_to_left, dead_to_left_pct, 1198 dead_to_right, live_to_right, 1199 double(dead_to_right) / live_to_right); 1200 } 1201 #endif // #ifndef PRODUCT 1202 1203 // Return a fraction indicating how much of the generation can be treated as 1204 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1205 // based on the density of live objects in the generation to determine a limit, 1206 // which is then adjusted so the return value is min_percent when the density is 1207 // 1. 1208 // 1209 // The following table shows some return values for a different values of the 1210 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1211 // min_percent is 1. 1212 // 1213 // fraction allowed as dead wood 1214 // ----------------------------------------------------------------- 1215 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1216 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1217 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1218 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1219 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1220 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1221 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1222 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1223 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1224 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1225 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1226 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1227 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1228 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1229 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1230 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1231 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1232 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1233 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1234 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1235 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1236 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1237 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1238 1239 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1240 { 1241 assert(_dwl_initialized, "uninitialized"); 1242 1243 // The raw limit is the value of the normal distribution at x = density. 1244 const double raw_limit = normal_distribution(density); 1245 1246 // Adjust the raw limit so it becomes the minimum when the density is 1. 1247 // 1248 // First subtract the adjustment value (which is simply the precomputed value 1249 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1250 // Then add the minimum value, so the minimum is returned when the density is 1251 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1252 const double min = double(min_percent) / 100.0; 1253 const double limit = raw_limit - _dwl_adjustment + min; 1254 return MAX2(limit, 0.0); 1255 } 1256 1257 ParallelCompactData::RegionData* 1258 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1259 const RegionData* end) 1260 { 1261 const size_t region_size = ParallelCompactData::RegionSize; 1262 ParallelCompactData& sd = summary_data(); 1263 size_t left = sd.region(beg); 1264 size_t right = end > beg ? sd.region(end) - 1 : left; 1265 1266 // Binary search. 1267 while (left < right) { 1268 // Equivalent to (left + right) / 2, but does not overflow. 1269 const size_t middle = left + (right - left) / 2; 1270 RegionData* const middle_ptr = sd.region(middle); 1271 HeapWord* const dest = middle_ptr->destination(); 1272 HeapWord* const addr = sd.region_to_addr(middle); 1273 assert(dest != NULL, "sanity"); 1274 assert(dest <= addr, "must move left"); 1275 1276 if (middle > left && dest < addr) { 1277 right = middle - 1; 1278 } else if (middle < right && middle_ptr->data_size() == region_size) { 1279 left = middle + 1; 1280 } else { 1281 return middle_ptr; 1282 } 1283 } 1284 return sd.region(left); 1285 } 1286 1287 ParallelCompactData::RegionData* 1288 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1289 const RegionData* end, 1290 size_t dead_words) 1291 { 1292 ParallelCompactData& sd = summary_data(); 1293 size_t left = sd.region(beg); 1294 size_t right = end > beg ? sd.region(end) - 1 : left; 1295 1296 // Binary search. 1297 while (left < right) { 1298 // Equivalent to (left + right) / 2, but does not overflow. 1299 const size_t middle = left + (right - left) / 2; 1300 RegionData* const middle_ptr = sd.region(middle); 1301 HeapWord* const dest = middle_ptr->destination(); 1302 HeapWord* const addr = sd.region_to_addr(middle); 1303 assert(dest != NULL, "sanity"); 1304 assert(dest <= addr, "must move left"); 1305 1306 const size_t dead_to_left = pointer_delta(addr, dest); 1307 if (middle > left && dead_to_left > dead_words) { 1308 right = middle - 1; 1309 } else if (middle < right && dead_to_left < dead_words) { 1310 left = middle + 1; 1311 } else { 1312 return middle_ptr; 1313 } 1314 } 1315 return sd.region(left); 1316 } 1317 1318 // The result is valid during the summary phase, after the initial summarization 1319 // of each space into itself, and before final summarization. 1320 inline double 1321 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1322 HeapWord* const bottom, 1323 HeapWord* const top, 1324 HeapWord* const new_top) 1325 { 1326 ParallelCompactData& sd = summary_data(); 1327 1328 assert(cp != NULL, "sanity"); 1329 assert(bottom != NULL, "sanity"); 1330 assert(top != NULL, "sanity"); 1331 assert(new_top != NULL, "sanity"); 1332 assert(top >= new_top, "summary data problem?"); 1333 assert(new_top > bottom, "space is empty; should not be here"); 1334 assert(new_top >= cp->destination(), "sanity"); 1335 assert(top >= sd.region_to_addr(cp), "sanity"); 1336 1337 HeapWord* const destination = cp->destination(); 1338 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1339 const size_t compacted_region_live = pointer_delta(new_top, destination); 1340 const size_t compacted_region_used = pointer_delta(top, 1341 sd.region_to_addr(cp)); 1342 const size_t reclaimable = compacted_region_used - compacted_region_live; 1343 1344 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1345 return double(reclaimable) / divisor; 1346 } 1347 1348 // Return the address of the end of the dense prefix, a.k.a. the start of the 1349 // compacted region. The address is always on a region boundary. 1350 // 1351 // Completely full regions at the left are skipped, since no compaction can 1352 // occur in those regions. Then the maximum amount of dead wood to allow is 1353 // computed, based on the density (amount live / capacity) of the generation; 1354 // the region with approximately that amount of dead space to the left is 1355 // identified as the limit region. Regions between the last completely full 1356 // region and the limit region are scanned and the one that has the best 1357 // (maximum) reclaimed_ratio() is selected. 1358 HeapWord* 1359 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1360 bool maximum_compaction) 1361 { 1362 const size_t region_size = ParallelCompactData::RegionSize; 1363 const ParallelCompactData& sd = summary_data(); 1364 1365 const MutableSpace* const space = _space_info[id].space(); 1366 HeapWord* const top = space->top(); 1367 HeapWord* const top_aligned_up = sd.region_align_up(top); 1368 HeapWord* const new_top = _space_info[id].new_top(); 1369 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1370 HeapWord* const bottom = space->bottom(); 1371 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1372 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1373 const RegionData* const new_top_cp = 1374 sd.addr_to_region_ptr(new_top_aligned_up); 1375 1376 // Skip full regions at the beginning of the space--they are necessarily part 1377 // of the dense prefix. 1378 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1379 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1380 space->is_empty(), "no dead space allowed to the left"); 1381 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1382 "region must have dead space"); 1383 1384 // The gc number is saved whenever a maximum compaction is done, and used to 1385 // determine when the maximum compaction interval has expired. This avoids 1386 // successive max compactions for different reasons. 1387 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1388 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1389 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1390 total_invocations() == HeapFirstMaximumCompactionCount; 1391 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1392 _maximum_compaction_gc_num = total_invocations(); 1393 return sd.region_to_addr(full_cp); 1394 } 1395 1396 const size_t space_live = pointer_delta(new_top, bottom); 1397 const size_t space_used = space->used_in_words(); 1398 const size_t space_capacity = space->capacity_in_words(); 1399 1400 const double density = double(space_live) / double(space_capacity); 1401 const size_t min_percent_free = MarkSweepDeadRatio; 1402 const double limiter = dead_wood_limiter(density, min_percent_free); 1403 const size_t dead_wood_max = space_used - space_live; 1404 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1405 dead_wood_max); 1406 1407 if (TraceParallelOldGCDensePrefix) { 1408 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1409 "space_cap=" SIZE_FORMAT, 1410 space_live, space_used, 1411 space_capacity); 1412 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1413 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1414 density, min_percent_free, limiter, 1415 dead_wood_max, dead_wood_limit); 1416 } 1417 1418 // Locate the region with the desired amount of dead space to the left. 1419 const RegionData* const limit_cp = 1420 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1421 1422 // Scan from the first region with dead space to the limit region and find the 1423 // one with the best (largest) reclaimed ratio. 1424 double best_ratio = 0.0; 1425 const RegionData* best_cp = full_cp; 1426 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1427 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1428 if (tmp_ratio > best_ratio) { 1429 best_cp = cp; 1430 best_ratio = tmp_ratio; 1431 } 1432 } 1433 1434 return sd.region_to_addr(best_cp); 1435 } 1436 1437 void PSParallelCompact::summarize_spaces_quick() 1438 { 1439 for (unsigned int i = 0; i < last_space_id; ++i) { 1440 const MutableSpace* space = _space_info[i].space(); 1441 HeapWord** nta = _space_info[i].new_top_addr(); 1442 bool result = _summary_data.summarize(_space_info[i].split_info(), 1443 space->bottom(), space->top(), NULL, 1444 space->bottom(), space->end(), nta); 1445 assert(result, "space must fit into itself"); 1446 _space_info[i].set_dense_prefix(space->bottom()); 1447 } 1448 } 1449 1450 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1451 { 1452 HeapWord* const dense_prefix_end = dense_prefix(id); 1453 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1454 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1455 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1456 // Only enough dead space is filled so that any remaining dead space to the 1457 // left is larger than the minimum filler object. (The remainder is filled 1458 // during the copy/update phase.) 1459 // 1460 // The size of the dead space to the right of the boundary is not a 1461 // concern, since compaction will be able to use whatever space is 1462 // available. 1463 // 1464 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1465 // surrounds the space to be filled with an object. 1466 // 1467 // In the 32-bit VM, each bit represents two 32-bit words: 1468 // +---+ 1469 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1470 // end_bits: ... x x x | 0 | || 0 x x ... 1471 // +---+ 1472 // 1473 // In the 64-bit VM, each bit represents one 64-bit word: 1474 // +------------+ 1475 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1476 // end_bits: ... x x 1 | 0 || 0 | x x ... 1477 // +------------+ 1478 // +-------+ 1479 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1480 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1481 // +-------+ 1482 // +-----------+ 1483 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1484 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1485 // +-----------+ 1486 // +-------+ 1487 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1488 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1489 // +-------+ 1490 1491 // Initially assume case a, c or e will apply. 1492 size_t obj_len = CollectedHeap::min_fill_size(); 1493 HeapWord* obj_beg = dense_prefix_end - obj_len; 1494 1495 #ifdef _LP64 1496 if (MinObjAlignment > 1) { // object alignment > heap word size 1497 // Cases a, c or e. 1498 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1499 // Case b above. 1500 obj_beg = dense_prefix_end - 1; 1501 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1502 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1503 // Case d above. 1504 obj_beg = dense_prefix_end - 3; 1505 obj_len = 3; 1506 } 1507 #endif // #ifdef _LP64 1508 1509 CollectedHeap::fill_with_object(obj_beg, obj_len); 1510 _mark_bitmap.mark_obj(obj_beg, obj_len); 1511 _summary_data.add_obj(obj_beg, obj_len); 1512 assert(start_array(id) != NULL, "sanity"); 1513 start_array(id)->allocate_block(obj_beg); 1514 } 1515 } 1516 1517 void 1518 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1519 { 1520 assert(id < last_space_id, "id out of range"); 1521 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1522 "should have been reset in summarize_spaces_quick()"); 1523 1524 const MutableSpace* space = _space_info[id].space(); 1525 if (_space_info[id].new_top() != space->bottom()) { 1526 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1527 _space_info[id].set_dense_prefix(dense_prefix_end); 1528 1529 #ifndef PRODUCT 1530 if (TraceParallelOldGCDensePrefix) { 1531 print_dense_prefix_stats("ratio", id, maximum_compaction, 1532 dense_prefix_end); 1533 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1534 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1535 } 1536 #endif // #ifndef PRODUCT 1537 1538 // Recompute the summary data, taking into account the dense prefix. If 1539 // every last byte will be reclaimed, then the existing summary data which 1540 // compacts everything can be left in place. 1541 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1542 // If dead space crosses the dense prefix boundary, it is (at least 1543 // partially) filled with a dummy object, marked live and added to the 1544 // summary data. This simplifies the copy/update phase and must be done 1545 // before the final locations of objects are determined, to prevent 1546 // leaving a fragment of dead space that is too small to fill. 1547 fill_dense_prefix_end(id); 1548 1549 // Compute the destination of each Region, and thus each object. 1550 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1551 _summary_data.summarize(_space_info[id].split_info(), 1552 dense_prefix_end, space->top(), NULL, 1553 dense_prefix_end, space->end(), 1554 _space_info[id].new_top_addr()); 1555 } 1556 } 1557 1558 if (log_develop_is_enabled(Trace, gc, compaction)) { 1559 const size_t region_size = ParallelCompactData::RegionSize; 1560 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1561 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1562 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1563 HeapWord* const new_top = _space_info[id].new_top(); 1564 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1565 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1566 log_develop_trace(gc, compaction)( 1567 "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1568 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1569 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1570 id, space->capacity_in_words(), p2i(dense_prefix_end), 1571 dp_region, dp_words / region_size, 1572 cr_words / region_size, p2i(new_top)); 1573 } 1574 } 1575 1576 #ifndef PRODUCT 1577 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1578 HeapWord* dst_beg, HeapWord* dst_end, 1579 SpaceId src_space_id, 1580 HeapWord* src_beg, HeapWord* src_end) 1581 { 1582 log_develop_trace(gc, compaction)( 1583 "Summarizing %d [%s] into %d [%s]: " 1584 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1585 SIZE_FORMAT "-" SIZE_FORMAT " " 1586 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1587 SIZE_FORMAT "-" SIZE_FORMAT, 1588 src_space_id, space_names[src_space_id], 1589 dst_space_id, space_names[dst_space_id], 1590 p2i(src_beg), p2i(src_end), 1591 _summary_data.addr_to_region_idx(src_beg), 1592 _summary_data.addr_to_region_idx(src_end), 1593 p2i(dst_beg), p2i(dst_end), 1594 _summary_data.addr_to_region_idx(dst_beg), 1595 _summary_data.addr_to_region_idx(dst_end)); 1596 } 1597 #endif // #ifndef PRODUCT 1598 1599 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1600 bool maximum_compaction) 1601 { 1602 GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer); 1603 1604 #ifdef ASSERT 1605 if (TraceParallelOldGCMarkingPhase) { 1606 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1607 "add_obj_bytes=" SIZE_FORMAT, 1608 add_obj_count, add_obj_size * HeapWordSize); 1609 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1610 "mark_bitmap_bytes=" SIZE_FORMAT, 1611 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1612 } 1613 #endif // #ifdef ASSERT 1614 1615 // Quick summarization of each space into itself, to see how much is live. 1616 summarize_spaces_quick(); 1617 1618 log_develop_trace(gc, compaction)("summary phase: after summarizing each space to self"); 1619 NOT_PRODUCT(print_region_ranges()); 1620 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1621 1622 // The amount of live data that will end up in old space (assuming it fits). 1623 size_t old_space_total_live = 0; 1624 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1625 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1626 _space_info[id].space()->bottom()); 1627 } 1628 1629 MutableSpace* const old_space = _space_info[old_space_id].space(); 1630 const size_t old_capacity = old_space->capacity_in_words(); 1631 if (old_space_total_live > old_capacity) { 1632 // XXX - should also try to expand 1633 maximum_compaction = true; 1634 } 1635 1636 // Old generations. 1637 summarize_space(old_space_id, maximum_compaction); 1638 1639 // Summarize the remaining spaces in the young gen. The initial target space 1640 // is the old gen. If a space does not fit entirely into the target, then the 1641 // remainder is compacted into the space itself and that space becomes the new 1642 // target. 1643 SpaceId dst_space_id = old_space_id; 1644 HeapWord* dst_space_end = old_space->end(); 1645 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1646 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1647 const MutableSpace* space = _space_info[id].space(); 1648 const size_t live = pointer_delta(_space_info[id].new_top(), 1649 space->bottom()); 1650 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1651 1652 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1653 SpaceId(id), space->bottom(), space->top());) 1654 if (live > 0 && live <= available) { 1655 // All the live data will fit. 1656 bool done = _summary_data.summarize(_space_info[id].split_info(), 1657 space->bottom(), space->top(), 1658 NULL, 1659 *new_top_addr, dst_space_end, 1660 new_top_addr); 1661 assert(done, "space must fit into old gen"); 1662 1663 // Reset the new_top value for the space. 1664 _space_info[id].set_new_top(space->bottom()); 1665 } else if (live > 0) { 1666 // Attempt to fit part of the source space into the target space. 1667 HeapWord* next_src_addr = NULL; 1668 bool done = _summary_data.summarize(_space_info[id].split_info(), 1669 space->bottom(), space->top(), 1670 &next_src_addr, 1671 *new_top_addr, dst_space_end, 1672 new_top_addr); 1673 assert(!done, "space should not fit into old gen"); 1674 assert(next_src_addr != NULL, "sanity"); 1675 1676 // The source space becomes the new target, so the remainder is compacted 1677 // within the space itself. 1678 dst_space_id = SpaceId(id); 1679 dst_space_end = space->end(); 1680 new_top_addr = _space_info[id].new_top_addr(); 1681 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1682 space->bottom(), dst_space_end, 1683 SpaceId(id), next_src_addr, space->top());) 1684 done = _summary_data.summarize(_space_info[id].split_info(), 1685 next_src_addr, space->top(), 1686 NULL, 1687 space->bottom(), dst_space_end, 1688 new_top_addr); 1689 assert(done, "space must fit when compacted into itself"); 1690 assert(*new_top_addr <= space->top(), "usage should not grow"); 1691 } 1692 } 1693 1694 log_develop_trace(gc, compaction)("Summary_phase: after final summarization"); 1695 NOT_PRODUCT(print_region_ranges()); 1696 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1697 } 1698 1699 // This method should contain all heap-specific policy for invoking a full 1700 // collection. invoke_no_policy() will only attempt to compact the heap; it 1701 // will do nothing further. If we need to bail out for policy reasons, scavenge 1702 // before full gc, or any other specialized behavior, it needs to be added here. 1703 // 1704 // Note that this method should only be called from the vm_thread while at a 1705 // safepoint. 1706 // 1707 // Note that the all_soft_refs_clear flag in the collector policy 1708 // may be true because this method can be called without intervening 1709 // activity. For example when the heap space is tight and full measure 1710 // are being taken to free space. 1711 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1712 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1713 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1714 "should be in vm thread"); 1715 1716 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1717 GCCause::Cause gc_cause = heap->gc_cause(); 1718 assert(!heap->is_gc_active(), "not reentrant"); 1719 1720 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1721 IsGCActiveMark mark; 1722 1723 if (ScavengeBeforeFullGC) { 1724 PSScavenge::invoke_no_policy(); 1725 } 1726 1727 const bool clear_all_soft_refs = 1728 heap->collector_policy()->should_clear_all_soft_refs(); 1729 1730 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1731 maximum_heap_compaction); 1732 } 1733 1734 // This method contains no policy. You should probably 1735 // be calling invoke() instead. 1736 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1737 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1738 assert(ref_processor() != NULL, "Sanity"); 1739 1740 if (GCLocker::check_active_before_gc()) { 1741 return false; 1742 } 1743 1744 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1745 1746 GCIdMark gc_id_mark; 1747 _gc_timer.register_gc_start(); 1748 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1749 1750 TimeStamp marking_start; 1751 TimeStamp compaction_start; 1752 TimeStamp collection_exit; 1753 1754 GCCause::Cause gc_cause = heap->gc_cause(); 1755 PSYoungGen* young_gen = heap->young_gen(); 1756 PSOldGen* old_gen = heap->old_gen(); 1757 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1758 1759 // The scope of casr should end after code that can change 1760 // CollectorPolicy::_should_clear_all_soft_refs. 1761 ClearedAllSoftRefs casr(maximum_heap_compaction, 1762 heap->collector_policy()); 1763 1764 if (ZapUnusedHeapArea) { 1765 // Save information needed to minimize mangling 1766 heap->record_gen_tops_before_GC(); 1767 } 1768 1769 // Make sure data structures are sane, make the heap parsable, and do other 1770 // miscellaneous bookkeeping. 1771 pre_compact(); 1772 1773 PreGCValues pre_gc_values(heap); 1774 1775 // Get the compaction manager reserved for the VM thread. 1776 ParCompactionManager* const vmthread_cm = 1777 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1778 1779 { 1780 ResourceMark rm; 1781 HandleMark hm; 1782 1783 // Set the number of GC threads to be used in this collection 1784 gc_task_manager()->set_active_gang(); 1785 gc_task_manager()->task_idle_workers(); 1786 1787 GCTraceCPUTime tcpu; 1788 GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true); 1789 1790 heap->pre_full_gc_dump(&_gc_timer); 1791 1792 TraceCollectorStats tcs(counters()); 1793 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 1794 1795 if (TraceOldGenTime) accumulated_time()->start(); 1796 1797 // Let the size policy know we're starting 1798 size_policy->major_collection_begin(); 1799 1800 CodeCache::gc_prologue(); 1801 1802 #if defined(COMPILER2) || INCLUDE_JVMCI 1803 DerivedPointerTable::clear(); 1804 #endif 1805 1806 ref_processor()->enable_discovery(); 1807 ref_processor()->setup_policy(maximum_heap_compaction); 1808 1809 bool marked_for_unloading = false; 1810 1811 marking_start.update(); 1812 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1813 1814 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1815 && GCCause::is_user_requested_gc(gc_cause); 1816 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1817 1818 #if defined(COMPILER2) || INCLUDE_JVMCI 1819 assert(DerivedPointerTable::is_active(), "Sanity"); 1820 DerivedPointerTable::set_active(false); 1821 #endif 1822 1823 // adjust_roots() updates Universe::_intArrayKlassObj which is 1824 // needed by the compaction for filling holes in the dense prefix. 1825 adjust_roots(vmthread_cm); 1826 1827 compaction_start.update(); 1828 compact(); 1829 1830 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1831 // done before resizing. 1832 post_compact(); 1833 1834 // Let the size policy know we're done 1835 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1836 1837 if (UseAdaptiveSizePolicy) { 1838 log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections()); 1839 log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT, 1840 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1841 1842 // Don't check if the size_policy is ready here. Let 1843 // the size_policy check that internally. 1844 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1845 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1846 // Swap the survivor spaces if from_space is empty. The 1847 // resize_young_gen() called below is normally used after 1848 // a successful young GC and swapping of survivor spaces; 1849 // otherwise, it will fail to resize the young gen with 1850 // the current implementation. 1851 if (young_gen->from_space()->is_empty()) { 1852 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1853 young_gen->swap_spaces(); 1854 } 1855 1856 // Calculate optimal free space amounts 1857 assert(young_gen->max_size() > 1858 young_gen->from_space()->capacity_in_bytes() + 1859 young_gen->to_space()->capacity_in_bytes(), 1860 "Sizes of space in young gen are out-of-bounds"); 1861 1862 size_t young_live = young_gen->used_in_bytes(); 1863 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1864 size_t old_live = old_gen->used_in_bytes(); 1865 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1866 size_t max_old_gen_size = old_gen->max_gen_size(); 1867 size_t max_eden_size = young_gen->max_size() - 1868 young_gen->from_space()->capacity_in_bytes() - 1869 young_gen->to_space()->capacity_in_bytes(); 1870 1871 // Used for diagnostics 1872 size_policy->clear_generation_free_space_flags(); 1873 1874 size_policy->compute_generations_free_space(young_live, 1875 eden_live, 1876 old_live, 1877 cur_eden, 1878 max_old_gen_size, 1879 max_eden_size, 1880 true /* full gc*/); 1881 1882 size_policy->check_gc_overhead_limit(young_live, 1883 eden_live, 1884 max_old_gen_size, 1885 max_eden_size, 1886 true /* full gc*/, 1887 gc_cause, 1888 heap->collector_policy()); 1889 1890 size_policy->decay_supplemental_growth(true /* full gc*/); 1891 1892 heap->resize_old_gen( 1893 size_policy->calculated_old_free_size_in_bytes()); 1894 1895 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1896 size_policy->calculated_survivor_size_in_bytes()); 1897 } 1898 1899 log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections()); 1900 } 1901 1902 if (UsePerfData) { 1903 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1904 counters->update_counters(); 1905 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1906 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1907 } 1908 1909 heap->resize_all_tlabs(); 1910 1911 // Resize the metaspace capacity after a collection 1912 MetaspaceGC::compute_new_size(); 1913 1914 if (TraceOldGenTime) { 1915 accumulated_time()->stop(); 1916 } 1917 1918 young_gen->print_used_change(pre_gc_values.young_gen_used()); 1919 old_gen->print_used_change(pre_gc_values.old_gen_used()); 1920 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 1921 1922 // Track memory usage and detect low memory 1923 MemoryService::track_memory_usage(); 1924 heap->update_counters(); 1925 gc_task_manager()->release_idle_workers(); 1926 1927 heap->post_full_gc_dump(&_gc_timer); 1928 } 1929 1930 #ifdef ASSERT 1931 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1932 ParCompactionManager* const cm = 1933 ParCompactionManager::manager_array(int(i)); 1934 assert(cm->marking_stack()->is_empty(), "should be empty"); 1935 assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i); 1936 } 1937 #endif // ASSERT 1938 1939 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1940 HandleMark hm; // Discard invalid handles created during verification 1941 Universe::verify("After GC"); 1942 } 1943 1944 // Re-verify object start arrays 1945 if (VerifyObjectStartArray && 1946 VerifyAfterGC) { 1947 old_gen->verify_object_start_array(); 1948 } 1949 1950 if (ZapUnusedHeapArea) { 1951 old_gen->object_space()->check_mangled_unused_area_complete(); 1952 } 1953 1954 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1955 1956 collection_exit.update(); 1957 1958 heap->print_heap_after_gc(); 1959 heap->trace_heap_after_gc(&_gc_tracer); 1960 1961 log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT, 1962 marking_start.ticks(), compaction_start.ticks(), 1963 collection_exit.ticks()); 1964 gc_task_manager()->print_task_time_stamps(); 1965 1966 #ifdef TRACESPINNING 1967 ParallelTaskTerminator::print_termination_counts(); 1968 #endif 1969 1970 AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections()); 1971 1972 _gc_timer.register_gc_end(); 1973 1974 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 1975 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 1976 1977 return true; 1978 } 1979 1980 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 1981 PSYoungGen* young_gen, 1982 PSOldGen* old_gen) { 1983 MutableSpace* const eden_space = young_gen->eden_space(); 1984 assert(!eden_space->is_empty(), "eden must be non-empty"); 1985 assert(young_gen->virtual_space()->alignment() == 1986 old_gen->virtual_space()->alignment(), "alignments do not match"); 1987 1988 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 1989 return false; 1990 } 1991 1992 // Both generations must be completely committed. 1993 if (young_gen->virtual_space()->uncommitted_size() != 0) { 1994 return false; 1995 } 1996 if (old_gen->virtual_space()->uncommitted_size() != 0) { 1997 return false; 1998 } 1999 2000 // Figure out how much to take from eden. Include the average amount promoted 2001 // in the total; otherwise the next young gen GC will simply bail out to a 2002 // full GC. 2003 const size_t alignment = old_gen->virtual_space()->alignment(); 2004 const size_t eden_used = eden_space->used_in_bytes(); 2005 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2006 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2007 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2008 2009 if (absorb_size >= eden_capacity) { 2010 return false; // Must leave some space in eden. 2011 } 2012 2013 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2014 if (new_young_size < young_gen->min_gen_size()) { 2015 return false; // Respect young gen minimum size. 2016 } 2017 2018 log_trace(heap, ergo)(" absorbing " SIZE_FORMAT "K: " 2019 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2020 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2021 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2022 absorb_size / K, 2023 eden_capacity / K, (eden_capacity - absorb_size) / K, 2024 young_gen->from_space()->used_in_bytes() / K, 2025 young_gen->to_space()->used_in_bytes() / K, 2026 young_gen->capacity_in_bytes() / K, new_young_size / K); 2027 2028 // Fill the unused part of the old gen. 2029 MutableSpace* const old_space = old_gen->object_space(); 2030 HeapWord* const unused_start = old_space->top(); 2031 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2032 2033 if (unused_words > 0) { 2034 if (unused_words < CollectedHeap::min_fill_size()) { 2035 return false; // If the old gen cannot be filled, must give up. 2036 } 2037 CollectedHeap::fill_with_objects(unused_start, unused_words); 2038 } 2039 2040 // Take the live data from eden and set both top and end in the old gen to 2041 // eden top. (Need to set end because reset_after_change() mangles the region 2042 // from end to virtual_space->high() in debug builds). 2043 HeapWord* const new_top = eden_space->top(); 2044 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2045 absorb_size); 2046 young_gen->reset_after_change(); 2047 old_space->set_top(new_top); 2048 old_space->set_end(new_top); 2049 old_gen->reset_after_change(); 2050 2051 // Update the object start array for the filler object and the data from eden. 2052 ObjectStartArray* const start_array = old_gen->start_array(); 2053 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2054 start_array->allocate_block(p); 2055 } 2056 2057 // Could update the promoted average here, but it is not typically updated at 2058 // full GCs and the value to use is unclear. Something like 2059 // 2060 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2061 2062 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2063 return true; 2064 } 2065 2066 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2067 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2068 "shouldn't return NULL"); 2069 return ParallelScavengeHeap::gc_task_manager(); 2070 } 2071 2072 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2073 bool maximum_heap_compaction, 2074 ParallelOldTracer *gc_tracer) { 2075 // Recursively traverse all live objects and mark them 2076 GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer); 2077 2078 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2079 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2080 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2081 TaskQueueSetSuper* qset = ParCompactionManager::stack_array(); 2082 ParallelTaskTerminator terminator(active_gc_threads, qset); 2083 2084 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm); 2085 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2086 2087 // Need new claim bits before marking starts. 2088 ClassLoaderDataGraph::clear_claimed_marks(); 2089 2090 { 2091 GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer); 2092 2093 ParallelScavengeHeap::ParStrongRootsScope psrs; 2094 2095 GCTaskQueue* q = GCTaskQueue::create(); 2096 2097 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2098 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2099 // We scan the thread roots in parallel 2100 Threads::create_thread_roots_marking_tasks(q); 2101 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2102 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2103 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2104 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2105 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2106 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2107 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2108 2109 if (active_gc_threads > 1) { 2110 for (uint j = 0; j < active_gc_threads; j++) { 2111 q->enqueue(new StealMarkingTask(&terminator)); 2112 } 2113 } 2114 2115 gc_task_manager()->execute_and_wait(q); 2116 } 2117 2118 // Process reference objects found during marking 2119 { 2120 GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer); 2121 2122 ReferenceProcessorStats stats; 2123 if (ref_processor()->processing_is_mt()) { 2124 RefProcTaskExecutor task_executor; 2125 stats = ref_processor()->process_discovered_references( 2126 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2127 &task_executor, &_gc_timer); 2128 } else { 2129 stats = ref_processor()->process_discovered_references( 2130 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2131 &_gc_timer); 2132 } 2133 2134 gc_tracer->report_gc_reference_stats(stats); 2135 } 2136 2137 // This is the point where the entire marking should have completed. 2138 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2139 2140 { 2141 GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer); 2142 2143 // Follow system dictionary roots and unload classes. 2144 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2145 2146 // Unload nmethods. 2147 CodeCache::do_unloading(is_alive_closure(), purged_class); 2148 2149 // Prune dead klasses from subklass/sibling/implementor lists. 2150 Klass::clean_weak_klass_links(is_alive_closure()); 2151 } 2152 2153 { 2154 GCTraceTime(Debug, gc, phases) t("Scrub String Table", &_gc_timer); 2155 // Delete entries for dead interned strings. 2156 StringTable::unlink(is_alive_closure()); 2157 } 2158 2159 { 2160 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", &_gc_timer); 2161 // Clean up unreferenced symbols in symbol table. 2162 SymbolTable::unlink(); 2163 } 2164 2165 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2166 } 2167 2168 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) { 2169 // Adjust the pointers to reflect the new locations 2170 GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer); 2171 2172 // Need new claim bits when tracing through and adjusting pointers. 2173 ClassLoaderDataGraph::clear_claimed_marks(); 2174 2175 PSParallelCompact::AdjustPointerClosure oop_closure(cm); 2176 PSParallelCompact::AdjustKlassClosure klass_closure(cm); 2177 2178 // General strong roots. 2179 Universe::oops_do(&oop_closure); 2180 JNIHandles::oops_do(&oop_closure); // Global (strong) JNI handles 2181 Threads::oops_do(&oop_closure, NULL); 2182 ObjectSynchronizer::oops_do(&oop_closure); 2183 FlatProfiler::oops_do(&oop_closure); 2184 Management::oops_do(&oop_closure); 2185 JvmtiExport::oops_do(&oop_closure); 2186 SystemDictionary::oops_do(&oop_closure); 2187 ClassLoaderDataGraph::oops_do(&oop_closure, &klass_closure, true); 2188 2189 // Now adjust pointers in remaining weak roots. (All of which should 2190 // have been cleared if they pointed to non-surviving objects.) 2191 // Global (weak) JNI handles 2192 JNIHandles::weak_oops_do(&oop_closure); 2193 2194 CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations); 2195 CodeCache::blobs_do(&adjust_from_blobs); 2196 StringTable::oops_do(&oop_closure); 2197 ref_processor()->weak_oops_do(&oop_closure); 2198 // Roots were visited so references into the young gen in roots 2199 // may have been scanned. Process them also. 2200 // Should the reference processor have a span that excludes 2201 // young gen objects? 2202 PSScavenge::reference_processor()->weak_oops_do(&oop_closure); 2203 } 2204 2205 // Helper class to print 8 region numbers per line and then print the total at the end. 2206 class FillableRegionLogger : public StackObj { 2207 private: 2208 Log(gc, compaction) log; 2209 static const int LineLength = 8; 2210 size_t _regions[LineLength]; 2211 int _next_index; 2212 bool _enabled; 2213 size_t _total_regions; 2214 public: 2215 FillableRegionLogger() : _next_index(0), _total_regions(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)) { } 2216 ~FillableRegionLogger() { 2217 log.trace(SIZE_FORMAT " initially fillable regions", _total_regions); 2218 } 2219 2220 void print_line() { 2221 if (!_enabled || _next_index == 0) { 2222 return; 2223 } 2224 FormatBuffer<> line("Fillable: "); 2225 for (int i = 0; i < _next_index; i++) { 2226 line.append(" " SIZE_FORMAT_W(7), _regions[i]); 2227 } 2228 log.trace("%s", line.buffer()); 2229 _next_index = 0; 2230 } 2231 2232 void handle(size_t region) { 2233 if (!_enabled) { 2234 return; 2235 } 2236 _regions[_next_index++] = region; 2237 if (_next_index == LineLength) { 2238 print_line(); 2239 } 2240 _total_regions++; 2241 } 2242 }; 2243 2244 void PSParallelCompact::prepare_region_draining_tasks(GCTaskQueue* q, 2245 uint parallel_gc_threads) 2246 { 2247 GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer); 2248 2249 // Find the threads that are active 2250 unsigned int which = 0; 2251 2252 // Find all regions that are available (can be filled immediately) and 2253 // distribute them to the thread stacks. The iteration is done in reverse 2254 // order (high to low) so the regions will be removed in ascending order. 2255 2256 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2257 2258 which = 0; 2259 // id + 1 is used to test termination so unsigned can 2260 // be used with an old_space_id == 0. 2261 FillableRegionLogger region_logger; 2262 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2263 SpaceInfo* const space_info = _space_info + id; 2264 MutableSpace* const space = space_info->space(); 2265 HeapWord* const new_top = space_info->new_top(); 2266 2267 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2268 const size_t end_region = 2269 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2270 2271 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2272 if (sd.region(cur)->claim_unsafe()) { 2273 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2274 cm->region_stack()->push(cur); 2275 region_logger.handle(cur); 2276 // Assign regions to tasks in round-robin fashion. 2277 if (++which == parallel_gc_threads) { 2278 which = 0; 2279 } 2280 } 2281 } 2282 region_logger.print_line(); 2283 } 2284 } 2285 2286 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2287 2288 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2289 uint parallel_gc_threads) { 2290 GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer); 2291 2292 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2293 2294 // Iterate over all the spaces adding tasks for updating 2295 // regions in the dense prefix. Assume that 1 gc thread 2296 // will work on opening the gaps and the remaining gc threads 2297 // will work on the dense prefix. 2298 unsigned int space_id; 2299 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2300 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2301 const MutableSpace* const space = _space_info[space_id].space(); 2302 2303 if (dense_prefix_end == space->bottom()) { 2304 // There is no dense prefix for this space. 2305 continue; 2306 } 2307 2308 // The dense prefix is before this region. 2309 size_t region_index_end_dense_prefix = 2310 sd.addr_to_region_idx(dense_prefix_end); 2311 RegionData* const dense_prefix_cp = 2312 sd.region(region_index_end_dense_prefix); 2313 assert(dense_prefix_end == space->end() || 2314 dense_prefix_cp->available() || 2315 dense_prefix_cp->claimed(), 2316 "The region after the dense prefix should always be ready to fill"); 2317 2318 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2319 2320 // Is there dense prefix work? 2321 size_t total_dense_prefix_regions = 2322 region_index_end_dense_prefix - region_index_start; 2323 // How many regions of the dense prefix should be given to 2324 // each thread? 2325 if (total_dense_prefix_regions > 0) { 2326 uint tasks_for_dense_prefix = 1; 2327 if (total_dense_prefix_regions <= 2328 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2329 // Don't over partition. This assumes that 2330 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2331 // so there are not many regions to process. 2332 tasks_for_dense_prefix = parallel_gc_threads; 2333 } else { 2334 // Over partition 2335 tasks_for_dense_prefix = parallel_gc_threads * 2336 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2337 } 2338 size_t regions_per_thread = total_dense_prefix_regions / 2339 tasks_for_dense_prefix; 2340 // Give each thread at least 1 region. 2341 if (regions_per_thread == 0) { 2342 regions_per_thread = 1; 2343 } 2344 2345 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2346 if (region_index_start >= region_index_end_dense_prefix) { 2347 break; 2348 } 2349 // region_index_end is not processed 2350 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2351 region_index_end_dense_prefix); 2352 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2353 region_index_start, 2354 region_index_end)); 2355 region_index_start = region_index_end; 2356 } 2357 } 2358 // This gets any part of the dense prefix that did not 2359 // fit evenly. 2360 if (region_index_start < region_index_end_dense_prefix) { 2361 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2362 region_index_start, 2363 region_index_end_dense_prefix)); 2364 } 2365 } 2366 } 2367 2368 void PSParallelCompact::enqueue_region_stealing_tasks( 2369 GCTaskQueue* q, 2370 ParallelTaskTerminator* terminator_ptr, 2371 uint parallel_gc_threads) { 2372 GCTraceTime(Trace, gc, phases) tm("Steal Task Setup", &_gc_timer); 2373 2374 // Once a thread has drained it's stack, it should try to steal regions from 2375 // other threads. 2376 for (uint j = 0; j < parallel_gc_threads; j++) { 2377 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2378 } 2379 } 2380 2381 #ifdef ASSERT 2382 // Write a histogram of the number of times the block table was filled for a 2383 // region. 2384 void PSParallelCompact::write_block_fill_histogram() 2385 { 2386 if (!log_develop_is_enabled(Trace, gc, compaction)) { 2387 return; 2388 } 2389 2390 Log(gc, compaction) log; 2391 ResourceMark rm; 2392 outputStream* out = log.trace_stream(); 2393 2394 typedef ParallelCompactData::RegionData rd_t; 2395 ParallelCompactData& sd = summary_data(); 2396 2397 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2398 MutableSpace* const spc = _space_info[id].space(); 2399 if (spc->bottom() != spc->top()) { 2400 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2401 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2402 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2403 2404 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2405 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2406 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2407 2408 for (const rd_t* cur = beg; cur < end; ++cur) { 2409 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2410 } 2411 out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2412 for (size_t i = 0; i < histo_len; ++i) { 2413 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2414 histo[i], 100.0 * histo[i] / region_cnt); 2415 } 2416 out->cr(); 2417 } 2418 } 2419 } 2420 #endif // #ifdef ASSERT 2421 2422 void PSParallelCompact::compact() { 2423 GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer); 2424 2425 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2426 PSOldGen* old_gen = heap->old_gen(); 2427 old_gen->start_array()->reset(); 2428 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2429 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2430 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2431 ParallelTaskTerminator terminator(active_gc_threads, qset); 2432 2433 GCTaskQueue* q = GCTaskQueue::create(); 2434 prepare_region_draining_tasks(q, active_gc_threads); 2435 enqueue_dense_prefix_tasks(q, active_gc_threads); 2436 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2437 2438 { 2439 GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer); 2440 2441 gc_task_manager()->execute_and_wait(q); 2442 2443 #ifdef ASSERT 2444 // Verify that all regions have been processed before the deferred updates. 2445 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2446 verify_complete(SpaceId(id)); 2447 } 2448 #endif 2449 } 2450 2451 { 2452 // Update the deferred objects, if any. Any compaction manager can be used. 2453 GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer); 2454 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2455 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2456 update_deferred_objects(cm, SpaceId(id)); 2457 } 2458 } 2459 2460 DEBUG_ONLY(write_block_fill_histogram()); 2461 } 2462 2463 #ifdef ASSERT 2464 void PSParallelCompact::verify_complete(SpaceId space_id) { 2465 // All Regions between space bottom() to new_top() should be marked as filled 2466 // and all Regions between new_top() and top() should be available (i.e., 2467 // should have been emptied). 2468 ParallelCompactData& sd = summary_data(); 2469 SpaceInfo si = _space_info[space_id]; 2470 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2471 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2472 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2473 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2474 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2475 2476 bool issued_a_warning = false; 2477 2478 size_t cur_region; 2479 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2480 const RegionData* const c = sd.region(cur_region); 2481 if (!c->completed()) { 2482 log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u", 2483 cur_region, c->destination_count()); 2484 issued_a_warning = true; 2485 } 2486 } 2487 2488 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2489 const RegionData* const c = sd.region(cur_region); 2490 if (!c->available()) { 2491 log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u", 2492 cur_region, c->destination_count()); 2493 issued_a_warning = true; 2494 } 2495 } 2496 2497 if (issued_a_warning) { 2498 print_region_ranges(); 2499 } 2500 } 2501 #endif // #ifdef ASSERT 2502 2503 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2504 _start_array->allocate_block(addr); 2505 compaction_manager()->update_contents(oop(addr)); 2506 } 2507 2508 // Update interior oops in the ranges of regions [beg_region, end_region). 2509 void 2510 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2511 SpaceId space_id, 2512 size_t beg_region, 2513 size_t end_region) { 2514 ParallelCompactData& sd = summary_data(); 2515 ParMarkBitMap* const mbm = mark_bitmap(); 2516 2517 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2518 HeapWord* const end_addr = sd.region_to_addr(end_region); 2519 assert(beg_region <= end_region, "bad region range"); 2520 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2521 2522 #ifdef ASSERT 2523 // Claim the regions to avoid triggering an assert when they are marked as 2524 // filled. 2525 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2526 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2527 } 2528 #endif // #ifdef ASSERT 2529 2530 if (beg_addr != space(space_id)->bottom()) { 2531 // Find the first live object or block of dead space that *starts* in this 2532 // range of regions. If a partial object crosses onto the region, skip it; 2533 // it will be marked for 'deferred update' when the object head is 2534 // processed. If dead space crosses onto the region, it is also skipped; it 2535 // will be filled when the prior region is processed. If neither of those 2536 // apply, the first word in the region is the start of a live object or dead 2537 // space. 2538 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2539 const RegionData* const cp = sd.region(beg_region); 2540 if (cp->partial_obj_size() != 0) { 2541 beg_addr = sd.partial_obj_end(beg_region); 2542 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2543 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2544 } 2545 } 2546 2547 if (beg_addr < end_addr) { 2548 // A live object or block of dead space starts in this range of Regions. 2549 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2550 2551 // Create closures and iterate. 2552 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2553 FillClosure fill_closure(cm, space_id); 2554 ParMarkBitMap::IterationStatus status; 2555 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2556 dense_prefix_end); 2557 if (status == ParMarkBitMap::incomplete) { 2558 update_closure.do_addr(update_closure.source()); 2559 } 2560 } 2561 2562 // Mark the regions as filled. 2563 RegionData* const beg_cp = sd.region(beg_region); 2564 RegionData* const end_cp = sd.region(end_region); 2565 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2566 cp->set_completed(); 2567 } 2568 } 2569 2570 // Return the SpaceId for the space containing addr. If addr is not in the 2571 // heap, last_space_id is returned. In debug mode it expects the address to be 2572 // in the heap and asserts such. 2573 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2574 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2575 2576 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2577 if (_space_info[id].space()->contains(addr)) { 2578 return SpaceId(id); 2579 } 2580 } 2581 2582 assert(false, "no space contains the addr"); 2583 return last_space_id; 2584 } 2585 2586 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2587 SpaceId id) { 2588 assert(id < last_space_id, "bad space id"); 2589 2590 ParallelCompactData& sd = summary_data(); 2591 const SpaceInfo* const space_info = _space_info + id; 2592 ObjectStartArray* const start_array = space_info->start_array(); 2593 2594 const MutableSpace* const space = space_info->space(); 2595 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2596 HeapWord* const beg_addr = space_info->dense_prefix(); 2597 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2598 2599 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2600 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2601 const RegionData* cur_region; 2602 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2603 HeapWord* const addr = cur_region->deferred_obj_addr(); 2604 if (addr != NULL) { 2605 if (start_array != NULL) { 2606 start_array->allocate_block(addr); 2607 } 2608 cm->update_contents(oop(addr)); 2609 assert(oop(addr)->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2610 } 2611 } 2612 } 2613 2614 // Skip over count live words starting from beg, and return the address of the 2615 // next live word. Unless marked, the word corresponding to beg is assumed to 2616 // be dead. Callers must either ensure beg does not correspond to the middle of 2617 // an object, or account for those live words in some other way. Callers must 2618 // also ensure that there are enough live words in the range [beg, end) to skip. 2619 HeapWord* 2620 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2621 { 2622 assert(count > 0, "sanity"); 2623 2624 ParMarkBitMap* m = mark_bitmap(); 2625 idx_t bits_to_skip = m->words_to_bits(count); 2626 idx_t cur_beg = m->addr_to_bit(beg); 2627 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2628 2629 do { 2630 cur_beg = m->find_obj_beg(cur_beg, search_end); 2631 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2632 const size_t obj_bits = cur_end - cur_beg + 1; 2633 if (obj_bits > bits_to_skip) { 2634 return m->bit_to_addr(cur_beg + bits_to_skip); 2635 } 2636 bits_to_skip -= obj_bits; 2637 cur_beg = cur_end + 1; 2638 } while (bits_to_skip > 0); 2639 2640 // Skipping the desired number of words landed just past the end of an object. 2641 // Find the start of the next object. 2642 cur_beg = m->find_obj_beg(cur_beg, search_end); 2643 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2644 return m->bit_to_addr(cur_beg); 2645 } 2646 2647 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2648 SpaceId src_space_id, 2649 size_t src_region_idx) 2650 { 2651 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2652 2653 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2654 if (split_info.dest_region_addr() == dest_addr) { 2655 // The partial object ending at the split point contains the first word to 2656 // be copied to dest_addr. 2657 return split_info.first_src_addr(); 2658 } 2659 2660 const ParallelCompactData& sd = summary_data(); 2661 ParMarkBitMap* const bitmap = mark_bitmap(); 2662 const size_t RegionSize = ParallelCompactData::RegionSize; 2663 2664 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2665 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2666 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2667 HeapWord* const src_region_destination = src_region_ptr->destination(); 2668 2669 assert(dest_addr >= src_region_destination, "wrong src region"); 2670 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2671 2672 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2673 HeapWord* const src_region_end = src_region_beg + RegionSize; 2674 2675 HeapWord* addr = src_region_beg; 2676 if (dest_addr == src_region_destination) { 2677 // Return the first live word in the source region. 2678 if (partial_obj_size == 0) { 2679 addr = bitmap->find_obj_beg(addr, src_region_end); 2680 assert(addr < src_region_end, "no objects start in src region"); 2681 } 2682 return addr; 2683 } 2684 2685 // Must skip some live data. 2686 size_t words_to_skip = dest_addr - src_region_destination; 2687 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2688 2689 if (partial_obj_size >= words_to_skip) { 2690 // All the live words to skip are part of the partial object. 2691 addr += words_to_skip; 2692 if (partial_obj_size == words_to_skip) { 2693 // Find the first live word past the partial object. 2694 addr = bitmap->find_obj_beg(addr, src_region_end); 2695 assert(addr < src_region_end, "wrong src region"); 2696 } 2697 return addr; 2698 } 2699 2700 // Skip over the partial object (if any). 2701 if (partial_obj_size != 0) { 2702 words_to_skip -= partial_obj_size; 2703 addr += partial_obj_size; 2704 } 2705 2706 // Skip over live words due to objects that start in the region. 2707 addr = skip_live_words(addr, src_region_end, words_to_skip); 2708 assert(addr < src_region_end, "wrong src region"); 2709 return addr; 2710 } 2711 2712 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2713 SpaceId src_space_id, 2714 size_t beg_region, 2715 HeapWord* end_addr) 2716 { 2717 ParallelCompactData& sd = summary_data(); 2718 2719 #ifdef ASSERT 2720 MutableSpace* const src_space = _space_info[src_space_id].space(); 2721 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2722 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2723 "src_space_id does not match beg_addr"); 2724 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2725 "src_space_id does not match end_addr"); 2726 #endif // #ifdef ASSERT 2727 2728 RegionData* const beg = sd.region(beg_region); 2729 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2730 2731 // Regions up to new_top() are enqueued if they become available. 2732 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2733 RegionData* const enqueue_end = 2734 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2735 2736 for (RegionData* cur = beg; cur < end; ++cur) { 2737 assert(cur->data_size() > 0, "region must have live data"); 2738 cur->decrement_destination_count(); 2739 if (cur < enqueue_end && cur->available() && cur->claim()) { 2740 cm->push_region(sd.region(cur)); 2741 } 2742 } 2743 } 2744 2745 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2746 SpaceId& src_space_id, 2747 HeapWord*& src_space_top, 2748 HeapWord* end_addr) 2749 { 2750 typedef ParallelCompactData::RegionData RegionData; 2751 2752 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2753 const size_t region_size = ParallelCompactData::RegionSize; 2754 2755 size_t src_region_idx = 0; 2756 2757 // Skip empty regions (if any) up to the top of the space. 2758 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2759 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2760 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2761 const RegionData* const top_region_ptr = 2762 sd.addr_to_region_ptr(top_aligned_up); 2763 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2764 ++src_region_ptr; 2765 } 2766 2767 if (src_region_ptr < top_region_ptr) { 2768 // The next source region is in the current space. Update src_region_idx 2769 // and the source address to match src_region_ptr. 2770 src_region_idx = sd.region(src_region_ptr); 2771 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2772 if (src_region_addr > closure.source()) { 2773 closure.set_source(src_region_addr); 2774 } 2775 return src_region_idx; 2776 } 2777 2778 // Switch to a new source space and find the first non-empty region. 2779 unsigned int space_id = src_space_id + 1; 2780 assert(space_id < last_space_id, "not enough spaces"); 2781 2782 HeapWord* const destination = closure.destination(); 2783 2784 do { 2785 MutableSpace* space = _space_info[space_id].space(); 2786 HeapWord* const bottom = space->bottom(); 2787 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2788 2789 // Iterate over the spaces that do not compact into themselves. 2790 if (bottom_cp->destination() != bottom) { 2791 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2792 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2793 2794 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2795 if (src_cp->live_obj_size() > 0) { 2796 // Found it. 2797 assert(src_cp->destination() == destination, 2798 "first live obj in the space must match the destination"); 2799 assert(src_cp->partial_obj_size() == 0, 2800 "a space cannot begin with a partial obj"); 2801 2802 src_space_id = SpaceId(space_id); 2803 src_space_top = space->top(); 2804 const size_t src_region_idx = sd.region(src_cp); 2805 closure.set_source(sd.region_to_addr(src_region_idx)); 2806 return src_region_idx; 2807 } else { 2808 assert(src_cp->data_size() == 0, "sanity"); 2809 } 2810 } 2811 } 2812 } while (++space_id < last_space_id); 2813 2814 assert(false, "no source region was found"); 2815 return 0; 2816 } 2817 2818 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 2819 { 2820 typedef ParMarkBitMap::IterationStatus IterationStatus; 2821 const size_t RegionSize = ParallelCompactData::RegionSize; 2822 ParMarkBitMap* const bitmap = mark_bitmap(); 2823 ParallelCompactData& sd = summary_data(); 2824 RegionData* const region_ptr = sd.region(region_idx); 2825 2826 // Get the items needed to construct the closure. 2827 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2828 SpaceId dest_space_id = space_id(dest_addr); 2829 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 2830 HeapWord* new_top = _space_info[dest_space_id].new_top(); 2831 assert(dest_addr < new_top, "sanity"); 2832 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 2833 2834 // Get the source region and related info. 2835 size_t src_region_idx = region_ptr->source_region(); 2836 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2837 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2838 2839 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2840 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2841 2842 // Adjust src_region_idx to prepare for decrementing destination counts (the 2843 // destination count is not decremented when a region is copied to itself). 2844 if (src_region_idx == region_idx) { 2845 src_region_idx += 1; 2846 } 2847 2848 if (bitmap->is_unmarked(closure.source())) { 2849 // The first source word is in the middle of an object; copy the remainder 2850 // of the object or as much as will fit. The fact that pointer updates were 2851 // deferred will be noted when the object header is processed. 2852 HeapWord* const old_src_addr = closure.source(); 2853 closure.copy_partial_obj(); 2854 if (closure.is_full()) { 2855 decrement_destination_counts(cm, src_space_id, src_region_idx, 2856 closure.source()); 2857 region_ptr->set_deferred_obj_addr(NULL); 2858 region_ptr->set_completed(); 2859 return; 2860 } 2861 2862 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2863 if (sd.region_align_down(old_src_addr) != end_addr) { 2864 // The partial object was copied from more than one source region. 2865 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2866 2867 // Move to the next source region, possibly switching spaces as well. All 2868 // args except end_addr may be modified. 2869 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2870 end_addr); 2871 } 2872 } 2873 2874 do { 2875 HeapWord* const cur_addr = closure.source(); 2876 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2877 src_space_top); 2878 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2879 2880 if (status == ParMarkBitMap::incomplete) { 2881 // The last obj that starts in the source region does not end in the 2882 // region. 2883 assert(closure.source() < end_addr, "sanity"); 2884 HeapWord* const obj_beg = closure.source(); 2885 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 2886 src_space_top); 2887 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 2888 if (obj_end < range_end) { 2889 // The end was found; the entire object will fit. 2890 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 2891 assert(status != ParMarkBitMap::would_overflow, "sanity"); 2892 } else { 2893 // The end was not found; the object will not fit. 2894 assert(range_end < src_space_top, "obj cannot cross space boundary"); 2895 status = ParMarkBitMap::would_overflow; 2896 } 2897 } 2898 2899 if (status == ParMarkBitMap::would_overflow) { 2900 // The last object did not fit. Note that interior oop updates were 2901 // deferred, then copy enough of the object to fill the region. 2902 region_ptr->set_deferred_obj_addr(closure.destination()); 2903 status = closure.copy_until_full(); // copies from closure.source() 2904 2905 decrement_destination_counts(cm, src_space_id, src_region_idx, 2906 closure.source()); 2907 region_ptr->set_completed(); 2908 return; 2909 } 2910 2911 if (status == ParMarkBitMap::full) { 2912 decrement_destination_counts(cm, src_space_id, src_region_idx, 2913 closure.source()); 2914 region_ptr->set_deferred_obj_addr(NULL); 2915 region_ptr->set_completed(); 2916 return; 2917 } 2918 2919 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2920 2921 // Move to the next source region, possibly switching spaces as well. All 2922 // args except end_addr may be modified. 2923 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2924 end_addr); 2925 } while (true); 2926 } 2927 2928 void PSParallelCompact::fill_blocks(size_t region_idx) 2929 { 2930 // Fill in the block table elements for the specified region. Each block 2931 // table element holds the number of live words in the region that are to the 2932 // left of the first object that starts in the block. Thus only blocks in 2933 // which an object starts need to be filled. 2934 // 2935 // The algorithm scans the section of the bitmap that corresponds to the 2936 // region, keeping a running total of the live words. When an object start is 2937 // found, if it's the first to start in the block that contains it, the 2938 // current total is written to the block table element. 2939 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2940 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2941 const size_t RegionSize = ParallelCompactData::RegionSize; 2942 2943 ParallelCompactData& sd = summary_data(); 2944 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2945 if (partial_obj_size >= RegionSize) { 2946 return; // No objects start in this region. 2947 } 2948 2949 // Ensure the first loop iteration decides that the block has changed. 2950 size_t cur_block = sd.block_count(); 2951 2952 const ParMarkBitMap* const bitmap = mark_bitmap(); 2953 2954 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2955 assert((size_t)1 << Log2BitsPerBlock == 2956 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2957 2958 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2959 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2960 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2961 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2962 while (beg_bit < range_end) { 2963 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2964 if (new_block != cur_block) { 2965 cur_block = new_block; 2966 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2967 } 2968 2969 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2970 if (end_bit < range_end - 1) { 2971 live_bits += end_bit - beg_bit + 1; 2972 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2973 } else { 2974 return; 2975 } 2976 } 2977 } 2978 2979 void 2980 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 2981 const MutableSpace* sp = space(space_id); 2982 if (sp->is_empty()) { 2983 return; 2984 } 2985 2986 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2987 ParMarkBitMap* const bitmap = mark_bitmap(); 2988 HeapWord* const dp_addr = dense_prefix(space_id); 2989 HeapWord* beg_addr = sp->bottom(); 2990 HeapWord* end_addr = sp->top(); 2991 2992 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 2993 2994 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 2995 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 2996 if (beg_region < dp_region) { 2997 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 2998 } 2999 3000 // The destination of the first live object that starts in the region is one 3001 // past the end of the partial object entering the region (if any). 3002 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3003 HeapWord* const new_top = _space_info[space_id].new_top(); 3004 assert(new_top >= dest_addr, "bad new_top value"); 3005 const size_t words = pointer_delta(new_top, dest_addr); 3006 3007 if (words > 0) { 3008 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3009 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3010 3011 ParMarkBitMap::IterationStatus status; 3012 status = bitmap->iterate(&closure, dest_addr, end_addr); 3013 assert(status == ParMarkBitMap::full, "iteration not complete"); 3014 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3015 "live objects skipped because closure is full"); 3016 } 3017 } 3018 3019 jlong PSParallelCompact::millis_since_last_gc() { 3020 // We need a monotonically non-decreasing time in ms but 3021 // os::javaTimeMillis() does not guarantee monotonicity. 3022 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3023 jlong ret_val = now - _time_of_last_gc; 3024 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3025 if (ret_val < 0) { 3026 NOT_PRODUCT(log_warning(gc)("time warp: " JLONG_FORMAT, ret_val);) 3027 return 0; 3028 } 3029 return ret_val; 3030 } 3031 3032 void PSParallelCompact::reset_millis_since_last_gc() { 3033 // We need a monotonically non-decreasing time in ms but 3034 // os::javaTimeMillis() does not guarantee monotonicity. 3035 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3036 } 3037 3038 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3039 { 3040 if (source() != destination()) { 3041 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3042 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3043 } 3044 update_state(words_remaining()); 3045 assert(is_full(), "sanity"); 3046 return ParMarkBitMap::full; 3047 } 3048 3049 void MoveAndUpdateClosure::copy_partial_obj() 3050 { 3051 size_t words = words_remaining(); 3052 3053 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3054 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3055 if (end_addr < range_end) { 3056 words = bitmap()->obj_size(source(), end_addr); 3057 } 3058 3059 // This test is necessary; if omitted, the pointer updates to a partial object 3060 // that crosses the dense prefix boundary could be overwritten. 3061 if (source() != destination()) { 3062 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3063 Copy::aligned_conjoint_words(source(), destination(), words); 3064 } 3065 update_state(words); 3066 } 3067 3068 void InstanceKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3069 PSParallelCompact::AdjustPointerClosure closure(cm); 3070 oop_oop_iterate_oop_maps<true>(obj, &closure); 3071 } 3072 3073 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3074 InstanceKlass::oop_pc_update_pointers(obj, cm); 3075 3076 PSParallelCompact::AdjustPointerClosure closure(cm); 3077 oop_oop_iterate_statics<true>(obj, &closure); 3078 } 3079 3080 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3081 InstanceKlass::oop_pc_update_pointers(obj, cm); 3082 } 3083 3084 #ifdef ASSERT 3085 template <class T> static void trace_reference_gc(const char *s, oop obj, 3086 T* referent_addr, 3087 T* next_addr, 3088 T* discovered_addr) { 3089 log_develop_trace(gc, ref)("%s obj " PTR_FORMAT, s, p2i(obj)); 3090 log_develop_trace(gc, ref)(" referent_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3091 p2i(referent_addr), referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3092 log_develop_trace(gc, ref)(" next_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3093 p2i(next_addr), next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3094 log_develop_trace(gc, ref)(" discovered_addr/* " PTR_FORMAT " / " PTR_FORMAT, 3095 p2i(discovered_addr), discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3096 } 3097 #endif 3098 3099 template <class T> 3100 static void oop_pc_update_pointers_specialized(oop obj, ParCompactionManager* cm) { 3101 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3102 PSParallelCompact::adjust_pointer(referent_addr, cm); 3103 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3104 PSParallelCompact::adjust_pointer(next_addr, cm); 3105 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3106 PSParallelCompact::adjust_pointer(discovered_addr, cm); 3107 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3108 referent_addr, next_addr, discovered_addr);) 3109 } 3110 3111 void InstanceRefKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3112 InstanceKlass::oop_pc_update_pointers(obj, cm); 3113 3114 if (UseCompressedOops) { 3115 oop_pc_update_pointers_specialized<narrowOop>(obj, cm); 3116 } else { 3117 oop_pc_update_pointers_specialized<oop>(obj, cm); 3118 } 3119 } 3120 3121 void ObjArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3122 assert(obj->is_objArray(), "obj must be obj array"); 3123 PSParallelCompact::AdjustPointerClosure closure(cm); 3124 oop_oop_iterate_elements<true>(objArrayOop(obj), &closure); 3125 } 3126 3127 void TypeArrayKlass::oop_pc_update_pointers(oop obj, ParCompactionManager* cm) { 3128 assert(obj->is_typeArray(),"must be a type array"); 3129 } 3130 3131 ParMarkBitMapClosure::IterationStatus 3132 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3133 assert(destination() != NULL, "sanity"); 3134 assert(bitmap()->obj_size(addr) == words, "bad size"); 3135 3136 _source = addr; 3137 assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) == 3138 destination(), "wrong destination"); 3139 3140 if (words > words_remaining()) { 3141 return ParMarkBitMap::would_overflow; 3142 } 3143 3144 // The start_array must be updated even if the object is not moving. 3145 if (_start_array != NULL) { 3146 _start_array->allocate_block(destination()); 3147 } 3148 3149 if (destination() != source()) { 3150 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3151 Copy::aligned_conjoint_words(source(), destination(), words); 3152 } 3153 3154 oop moved_oop = (oop) destination(); 3155 compaction_manager()->update_contents(moved_oop); 3156 assert(moved_oop->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3157 3158 update_state(words); 3159 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3160 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3161 } 3162 3163 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3164 ParCompactionManager* cm, 3165 PSParallelCompact::SpaceId space_id) : 3166 ParMarkBitMapClosure(mbm, cm), 3167 _space_id(space_id), 3168 _start_array(PSParallelCompact::start_array(space_id)) 3169 { 3170 } 3171 3172 // Updates the references in the object to their new values. 3173 ParMarkBitMapClosure::IterationStatus 3174 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3175 do_addr(addr); 3176 return ParMarkBitMap::incomplete; 3177 } 3178 3179 ParMarkBitMapClosure::IterationStatus 3180 FillClosure::do_addr(HeapWord* addr, size_t size) { 3181 CollectedHeap::fill_with_objects(addr, size); 3182 HeapWord* const end = addr + size; 3183 do { 3184 _start_array->allocate_block(addr); 3185 addr += oop(addr)->size(); 3186 } while (addr < end); 3187 return ParMarkBitMap::incomplete; 3188 }