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