1 /* 2 * Copyright (c) 2005, 2015, 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/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/parallel/gcTaskManager.hpp" 30 #include "gc/parallel/parallelScavengeHeap.inline.hpp" 31 #include "gc/parallel/pcTasks.hpp" 32 #include "gc/parallel/psAdaptiveSizePolicy.hpp" 33 #include "gc/parallel/psCompactionManager.inline.hpp" 34 #include "gc/parallel/psMarkSweep.hpp" 35 #include "gc/parallel/psMarkSweepDecorator.hpp" 36 #include "gc/parallel/psOldGen.hpp" 37 #include "gc/parallel/psParallelCompact.inline.hpp" 38 #include "gc/parallel/psPromotionManager.inline.hpp" 39 #include "gc/parallel/psScavenge.hpp" 40 #include "gc/parallel/psYoungGen.hpp" 41 #include "gc/shared/gcCause.hpp" 42 #include "gc/shared/gcHeapSummary.hpp" 43 #include "gc/shared/gcId.hpp" 44 #include "gc/shared/gcLocker.inline.hpp" 45 #include "gc/shared/gcTimer.hpp" 46 #include "gc/shared/gcTrace.hpp" 47 #include "gc/shared/gcTraceTime.hpp" 48 #include "gc/shared/isGCActiveMark.hpp" 49 #include "gc/shared/referencePolicy.hpp" 50 #include "gc/shared/referenceProcessor.hpp" 51 #include "gc/shared/spaceDecorator.hpp" 52 #include "oops/instanceKlass.inline.hpp" 53 #include "oops/instanceMirrorKlass.inline.hpp" 54 #include "oops/methodData.hpp" 55 #include "oops/objArrayKlass.inline.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "runtime/atomic.inline.hpp" 58 #include "runtime/fprofiler.hpp" 59 #include "runtime/safepoint.hpp" 60 #include "runtime/vmThread.hpp" 61 #include "services/management.hpp" 62 #include "services/memTracker.hpp" 63 #include "services/memoryService.hpp" 64 #include "utilities/events.hpp" 65 #include "utilities/stack.inline.hpp" 66 67 #include <math.h> 68 69 // All sizes are in HeapWords. 70 const size_t ParallelCompactData::Log2RegionSize = 16; // 64K words 71 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 72 const size_t ParallelCompactData::RegionSizeBytes = 73 RegionSize << LogHeapWordSize; 74 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 75 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 76 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 77 78 const size_t ParallelCompactData::Log2BlockSize = 7; // 128 words 79 const size_t ParallelCompactData::BlockSize = (size_t)1 << Log2BlockSize; 80 const size_t ParallelCompactData::BlockSizeBytes = 81 BlockSize << LogHeapWordSize; 82 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1; 83 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1; 84 const size_t ParallelCompactData::BlockAddrMask = ~BlockAddrOffsetMask; 85 86 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize; 87 const size_t ParallelCompactData::Log2BlocksPerRegion = 88 Log2RegionSize - Log2BlockSize; 89 90 const ParallelCompactData::RegionData::region_sz_t 91 ParallelCompactData::RegionData::dc_shift = 27; 92 93 const ParallelCompactData::RegionData::region_sz_t 94 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 95 96 const ParallelCompactData::RegionData::region_sz_t 97 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 98 99 const ParallelCompactData::RegionData::region_sz_t 100 ParallelCompactData::RegionData::los_mask = ~dc_mask; 101 102 const ParallelCompactData::RegionData::region_sz_t 103 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 104 105 const ParallelCompactData::RegionData::region_sz_t 106 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 107 108 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 109 bool PSParallelCompact::_print_phases = false; 110 111 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 112 113 double PSParallelCompact::_dwl_mean; 114 double PSParallelCompact::_dwl_std_dev; 115 double PSParallelCompact::_dwl_first_term; 116 double PSParallelCompact::_dwl_adjustment; 117 #ifdef ASSERT 118 bool PSParallelCompact::_dwl_initialized = false; 119 #endif // #ifdef ASSERT 120 121 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 122 HeapWord* destination) 123 { 124 assert(src_region_idx != 0, "invalid src_region_idx"); 125 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 126 assert(destination != NULL, "invalid destination argument"); 127 128 _src_region_idx = src_region_idx; 129 _partial_obj_size = partial_obj_size; 130 _destination = destination; 131 132 // These fields may not be updated below, so make sure they're clear. 133 assert(_dest_region_addr == NULL, "should have been cleared"); 134 assert(_first_src_addr == NULL, "should have been cleared"); 135 136 // Determine the number of destination regions for the partial object. 137 HeapWord* const last_word = destination + partial_obj_size - 1; 138 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 139 HeapWord* const beg_region_addr = sd.region_align_down(destination); 140 HeapWord* const end_region_addr = sd.region_align_down(last_word); 141 142 if (beg_region_addr == end_region_addr) { 143 // One destination region. 144 _destination_count = 1; 145 if (end_region_addr == destination) { 146 // The destination falls on a region boundary, thus the first word of the 147 // partial object will be the first word copied to the destination region. 148 _dest_region_addr = end_region_addr; 149 _first_src_addr = sd.region_to_addr(src_region_idx); 150 } 151 } else { 152 // Two destination regions. When copied, the partial object will cross a 153 // destination region boundary, so a word somewhere within the partial 154 // object will be the first word copied to the second destination region. 155 _destination_count = 2; 156 _dest_region_addr = end_region_addr; 157 const size_t ofs = pointer_delta(end_region_addr, destination); 158 assert(ofs < _partial_obj_size, "sanity"); 159 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 160 } 161 } 162 163 void SplitInfo::clear() 164 { 165 _src_region_idx = 0; 166 _partial_obj_size = 0; 167 _destination = NULL; 168 _destination_count = 0; 169 _dest_region_addr = NULL; 170 _first_src_addr = NULL; 171 assert(!is_valid(), "sanity"); 172 } 173 174 #ifdef ASSERT 175 void SplitInfo::verify_clear() 176 { 177 assert(_src_region_idx == 0, "not clear"); 178 assert(_partial_obj_size == 0, "not clear"); 179 assert(_destination == NULL, "not clear"); 180 assert(_destination_count == 0, "not clear"); 181 assert(_dest_region_addr == NULL, "not clear"); 182 assert(_first_src_addr == NULL, "not clear"); 183 } 184 #endif // #ifdef ASSERT 185 186 187 void PSParallelCompact::print_on_error(outputStream* st) { 188 _mark_bitmap.print_on_error(st); 189 } 190 191 #ifndef PRODUCT 192 const char* PSParallelCompact::space_names[] = { 193 "old ", "eden", "from", "to " 194 }; 195 196 void PSParallelCompact::print_region_ranges() 197 { 198 tty->print_cr("space bottom top end new_top"); 199 tty->print_cr("------ ---------- ---------- ---------- ----------"); 200 201 for (unsigned int id = 0; id < last_space_id; ++id) { 202 const MutableSpace* space = _space_info[id].space(); 203 tty->print_cr("%u %s " 204 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 205 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 206 id, space_names[id], 207 summary_data().addr_to_region_idx(space->bottom()), 208 summary_data().addr_to_region_idx(space->top()), 209 summary_data().addr_to_region_idx(space->end()), 210 summary_data().addr_to_region_idx(_space_info[id].new_top())); 211 } 212 } 213 214 void 215 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 216 { 217 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 218 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 219 220 ParallelCompactData& sd = PSParallelCompact::summary_data(); 221 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 222 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 223 REGION_IDX_FORMAT " " PTR_FORMAT " " 224 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 225 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 226 i, p2i(c->data_location()), dci, p2i(c->destination()), 227 c->partial_obj_size(), c->live_obj_size(), 228 c->data_size(), c->source_region(), c->destination_count()); 229 230 #undef REGION_IDX_FORMAT 231 #undef REGION_DATA_FORMAT 232 } 233 234 void 235 print_generic_summary_data(ParallelCompactData& summary_data, 236 HeapWord* const beg_addr, 237 HeapWord* const end_addr) 238 { 239 size_t total_words = 0; 240 size_t i = summary_data.addr_to_region_idx(beg_addr); 241 const size_t last = summary_data.addr_to_region_idx(end_addr); 242 HeapWord* pdest = 0; 243 244 while (i <= last) { 245 ParallelCompactData::RegionData* c = summary_data.region(i); 246 if (c->data_size() != 0 || c->destination() != pdest) { 247 print_generic_summary_region(i, c); 248 total_words += c->data_size(); 249 pdest = c->destination(); 250 } 251 ++i; 252 } 253 254 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 255 } 256 257 void 258 print_generic_summary_data(ParallelCompactData& summary_data, 259 SpaceInfo* space_info) 260 { 261 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 262 const MutableSpace* space = space_info[id].space(); 263 print_generic_summary_data(summary_data, space->bottom(), 264 MAX2(space->top(), space_info[id].new_top())); 265 } 266 } 267 268 void 269 print_initial_summary_region(size_t i, 270 const ParallelCompactData::RegionData* c, 271 bool newline = true) 272 { 273 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 274 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 275 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 276 i, p2i(c->destination()), 277 c->partial_obj_size(), c->live_obj_size(), 278 c->data_size(), c->source_region(), c->destination_count()); 279 if (newline) tty->cr(); 280 } 281 282 void 283 print_initial_summary_data(ParallelCompactData& summary_data, 284 const MutableSpace* space) { 285 if (space->top() == space->bottom()) { 286 return; 287 } 288 289 const size_t region_size = ParallelCompactData::RegionSize; 290 typedef ParallelCompactData::RegionData RegionData; 291 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 292 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 293 const RegionData* c = summary_data.region(end_region - 1); 294 HeapWord* end_addr = c->destination() + c->data_size(); 295 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 296 297 // Print (and count) the full regions at the beginning of the space. 298 size_t full_region_count = 0; 299 size_t i = summary_data.addr_to_region_idx(space->bottom()); 300 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 301 print_initial_summary_region(i, summary_data.region(i)); 302 ++full_region_count; 303 ++i; 304 } 305 306 size_t live_to_right = live_in_space - full_region_count * region_size; 307 308 double max_reclaimed_ratio = 0.0; 309 size_t max_reclaimed_ratio_region = 0; 310 size_t max_dead_to_right = 0; 311 size_t max_live_to_right = 0; 312 313 // Print the 'reclaimed ratio' for regions while there is something live in 314 // the region or to the right of it. The remaining regions are empty (and 315 // uninteresting), and computing the ratio will result in division by 0. 316 while (i < end_region && live_to_right > 0) { 317 c = summary_data.region(i); 318 HeapWord* const region_addr = summary_data.region_to_addr(i); 319 const size_t used_to_right = pointer_delta(space->top(), region_addr); 320 const size_t dead_to_right = used_to_right - live_to_right; 321 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 322 323 if (reclaimed_ratio > max_reclaimed_ratio) { 324 max_reclaimed_ratio = reclaimed_ratio; 325 max_reclaimed_ratio_region = i; 326 max_dead_to_right = dead_to_right; 327 max_live_to_right = live_to_right; 328 } 329 330 print_initial_summary_region(i, c, false); 331 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 332 reclaimed_ratio, dead_to_right, live_to_right); 333 334 live_to_right -= c->data_size(); 335 ++i; 336 } 337 338 // Any remaining regions are empty. Print one more if there is one. 339 if (i < end_region) { 340 print_initial_summary_region(i, summary_data.region(i)); 341 } 342 343 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 344 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 345 max_reclaimed_ratio_region, max_dead_to_right, 346 max_live_to_right, max_reclaimed_ratio); 347 } 348 349 void 350 print_initial_summary_data(ParallelCompactData& summary_data, 351 SpaceInfo* space_info) { 352 unsigned int id = PSParallelCompact::old_space_id; 353 const MutableSpace* space; 354 do { 355 space = space_info[id].space(); 356 print_initial_summary_data(summary_data, space); 357 } while (++id < PSParallelCompact::eden_space_id); 358 359 do { 360 space = space_info[id].space(); 361 print_generic_summary_data(summary_data, space->bottom(), space->top()); 362 } while (++id < PSParallelCompact::last_space_id); 363 } 364 #endif // #ifndef PRODUCT 365 366 #ifdef ASSERT 367 size_t add_obj_count; 368 size_t add_obj_size; 369 size_t mark_bitmap_count; 370 size_t mark_bitmap_size; 371 #endif // #ifdef ASSERT 372 373 ParallelCompactData::ParallelCompactData() 374 { 375 _region_start = 0; 376 377 _region_vspace = 0; 378 _reserved_byte_size = 0; 379 _region_data = 0; 380 _region_count = 0; 381 382 _block_vspace = 0; 383 _block_data = 0; 384 _block_count = 0; 385 } 386 387 bool ParallelCompactData::initialize(MemRegion covered_region) 388 { 389 _region_start = covered_region.start(); 390 const size_t region_size = covered_region.word_size(); 391 DEBUG_ONLY(_region_end = _region_start + region_size;) 392 393 assert(region_align_down(_region_start) == _region_start, 394 "region start not aligned"); 395 assert((region_size & RegionSizeOffsetMask) == 0, 396 "region size not a multiple of RegionSize"); 397 398 bool result = initialize_region_data(region_size) && initialize_block_data(); 399 return result; 400 } 401 402 PSVirtualSpace* 403 ParallelCompactData::create_vspace(size_t count, size_t element_size) 404 { 405 const size_t raw_bytes = count * element_size; 406 const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10); 407 const size_t granularity = os::vm_allocation_granularity(); 408 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 409 410 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 411 MAX2(page_sz, granularity); 412 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 413 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 414 rs.size()); 415 416 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 417 418 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 419 if (vspace != 0) { 420 if (vspace->expand_by(_reserved_byte_size)) { 421 return vspace; 422 } 423 delete vspace; 424 // Release memory reserved in the space. 425 rs.release(); 426 } 427 428 return 0; 429 } 430 431 bool ParallelCompactData::initialize_region_data(size_t region_size) 432 { 433 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 434 _region_vspace = create_vspace(count, sizeof(RegionData)); 435 if (_region_vspace != 0) { 436 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 437 _region_count = count; 438 return true; 439 } 440 return false; 441 } 442 443 bool ParallelCompactData::initialize_block_data() 444 { 445 assert(_region_count != 0, "region data must be initialized first"); 446 const size_t count = _region_count << Log2BlocksPerRegion; 447 _block_vspace = create_vspace(count, sizeof(BlockData)); 448 if (_block_vspace != 0) { 449 _block_data = (BlockData*)_block_vspace->reserved_low_addr(); 450 _block_count = count; 451 return true; 452 } 453 return false; 454 } 455 456 void ParallelCompactData::clear() 457 { 458 memset(_region_data, 0, _region_vspace->committed_size()); 459 memset(_block_data, 0, _block_vspace->committed_size()); 460 } 461 462 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 463 assert(beg_region <= _region_count, "beg_region out of range"); 464 assert(end_region <= _region_count, "end_region out of range"); 465 assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize"); 466 467 const size_t region_cnt = end_region - beg_region; 468 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 469 470 const size_t beg_block = beg_region * BlocksPerRegion; 471 const size_t block_cnt = region_cnt * BlocksPerRegion; 472 memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData)); 473 } 474 475 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 476 { 477 const RegionData* cur_cp = region(region_idx); 478 const RegionData* const end_cp = region(region_count() - 1); 479 480 HeapWord* result = region_to_addr(region_idx); 481 if (cur_cp < end_cp) { 482 do { 483 result += cur_cp->partial_obj_size(); 484 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 485 } 486 return result; 487 } 488 489 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 490 { 491 const size_t obj_ofs = pointer_delta(addr, _region_start); 492 const size_t beg_region = obj_ofs >> Log2RegionSize; 493 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 494 495 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 496 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 497 498 if (beg_region == end_region) { 499 // All in one region. 500 _region_data[beg_region].add_live_obj(len); 501 return; 502 } 503 504 // First region. 505 const size_t beg_ofs = region_offset(addr); 506 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 507 508 Klass* klass = ((oop)addr)->klass(); 509 // Middle regions--completely spanned by this object. 510 for (size_t region = beg_region + 1; region < end_region; ++region) { 511 _region_data[region].set_partial_obj_size(RegionSize); 512 _region_data[region].set_partial_obj_addr(addr); 513 } 514 515 // Last region. 516 const size_t end_ofs = region_offset(addr + len - 1); 517 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 518 _region_data[end_region].set_partial_obj_addr(addr); 519 } 520 521 void 522 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 523 { 524 assert(region_offset(beg) == 0, "not RegionSize aligned"); 525 assert(region_offset(end) == 0, "not RegionSize aligned"); 526 527 size_t cur_region = addr_to_region_idx(beg); 528 const size_t end_region = addr_to_region_idx(end); 529 HeapWord* addr = beg; 530 while (cur_region < end_region) { 531 _region_data[cur_region].set_destination(addr); 532 _region_data[cur_region].set_destination_count(0); 533 _region_data[cur_region].set_source_region(cur_region); 534 _region_data[cur_region].set_data_location(addr); 535 536 // Update live_obj_size so the region appears completely full. 537 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 538 _region_data[cur_region].set_live_obj_size(live_size); 539 540 ++cur_region; 541 addr += RegionSize; 542 } 543 } 544 545 // Find the point at which a space can be split and, if necessary, record the 546 // split point. 547 // 548 // If the current src region (which overflowed the destination space) doesn't 549 // have a partial object, the split point is at the beginning of the current src 550 // region (an "easy" split, no extra bookkeeping required). 551 // 552 // If the current src region has a partial object, the split point is in the 553 // region where that partial object starts (call it the split_region). If 554 // split_region has a partial object, then the split point is just after that 555 // partial object (a "hard" split where we have to record the split data and 556 // zero the partial_obj_size field). With a "hard" split, we know that the 557 // partial_obj ends within split_region because the partial object that caused 558 // the overflow starts in split_region. If split_region doesn't have a partial 559 // obj, then the split is at the beginning of split_region (another "easy" 560 // split). 561 HeapWord* 562 ParallelCompactData::summarize_split_space(size_t src_region, 563 SplitInfo& split_info, 564 HeapWord* destination, 565 HeapWord* target_end, 566 HeapWord** target_next) 567 { 568 assert(destination <= target_end, "sanity"); 569 assert(destination + _region_data[src_region].data_size() > target_end, 570 "region should not fit into target space"); 571 assert(is_region_aligned(target_end), "sanity"); 572 573 size_t split_region = src_region; 574 HeapWord* split_destination = destination; 575 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 576 577 if (destination + partial_obj_size > target_end) { 578 // The split point is just after the partial object (if any) in the 579 // src_region that contains the start of the object that overflowed the 580 // destination space. 581 // 582 // Find the start of the "overflow" object and set split_region to the 583 // region containing it. 584 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 585 split_region = addr_to_region_idx(overflow_obj); 586 587 // Clear the source_region field of all destination regions whose first word 588 // came from data after the split point (a non-null source_region field 589 // implies a region must be filled). 590 // 591 // An alternative to the simple loop below: clear during post_compact(), 592 // which uses memcpy instead of individual stores, and is easy to 593 // parallelize. (The downside is that it clears the entire RegionData 594 // object as opposed to just one field.) 595 // 596 // post_compact() would have to clear the summary data up to the highest 597 // address that was written during the summary phase, which would be 598 // 599 // max(top, max(new_top, clear_top)) 600 // 601 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 602 // to target_end. 603 const RegionData* const sr = region(split_region); 604 const size_t beg_idx = 605 addr_to_region_idx(region_align_up(sr->destination() + 606 sr->partial_obj_size())); 607 const size_t end_idx = addr_to_region_idx(target_end); 608 609 if (TraceParallelOldGCSummaryPhase) { 610 gclog_or_tty->print_cr("split: clearing source_region field in [" 611 SIZE_FORMAT ", " SIZE_FORMAT ")", 612 beg_idx, end_idx); 613 } 614 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 615 _region_data[idx].set_source_region(0); 616 } 617 618 // Set split_destination and partial_obj_size to reflect the split region. 619 split_destination = sr->destination(); 620 partial_obj_size = sr->partial_obj_size(); 621 } 622 623 // The split is recorded only if a partial object extends onto the region. 624 if (partial_obj_size != 0) { 625 _region_data[split_region].set_partial_obj_size(0); 626 split_info.record(split_region, partial_obj_size, split_destination); 627 } 628 629 // Setup the continuation addresses. 630 *target_next = split_destination + partial_obj_size; 631 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 632 633 if (TraceParallelOldGCSummaryPhase) { 634 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 635 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 636 " pos=" SIZE_FORMAT, 637 split_type, p2i(source_next), split_region, 638 partial_obj_size); 639 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 640 " tn=" PTR_FORMAT, 641 split_type, p2i(split_destination), 642 addr_to_region_idx(split_destination), 643 p2i(*target_next)); 644 645 if (partial_obj_size != 0) { 646 HeapWord* const po_beg = split_info.destination(); 647 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 648 gclog_or_tty->print_cr("%s split: " 649 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 650 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 651 split_type, 652 p2i(po_beg), addr_to_region_idx(po_beg), 653 p2i(po_end), addr_to_region_idx(po_end)); 654 } 655 } 656 657 return source_next; 658 } 659 660 bool ParallelCompactData::summarize(SplitInfo& split_info, 661 HeapWord* source_beg, HeapWord* source_end, 662 HeapWord** source_next, 663 HeapWord* target_beg, HeapWord* target_end, 664 HeapWord** target_next) 665 { 666 if (TraceParallelOldGCSummaryPhase) { 667 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 668 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 669 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 670 p2i(source_beg), p2i(source_end), p2i(source_next_val), 671 p2i(target_beg), p2i(target_end), p2i(*target_next)); 672 } 673 674 size_t cur_region = addr_to_region_idx(source_beg); 675 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 676 677 HeapWord *dest_addr = target_beg; 678 while (cur_region < end_region) { 679 // The destination must be set even if the region has no data. 680 _region_data[cur_region].set_destination(dest_addr); 681 682 size_t words = _region_data[cur_region].data_size(); 683 if (words > 0) { 684 // If cur_region does not fit entirely into the target space, find a point 685 // at which the source space can be 'split' so that part is copied to the 686 // target space and the rest is copied elsewhere. 687 if (dest_addr + words > target_end) { 688 assert(source_next != NULL, "source_next is NULL when splitting"); 689 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 690 target_end, target_next); 691 return false; 692 } 693 694 // Compute the destination_count for cur_region, and if necessary, update 695 // source_region for a destination region. The source_region field is 696 // updated if cur_region is the first (left-most) region to be copied to a 697 // destination region. 698 // 699 // The destination_count calculation is a bit subtle. A region that has 700 // data that compacts into itself does not count itself as a destination. 701 // This maintains the invariant that a zero count means the region is 702 // available and can be claimed and then filled. 703 uint destination_count = 0; 704 if (split_info.is_split(cur_region)) { 705 // The current region has been split: the partial object will be copied 706 // to one destination space and the remaining data will be copied to 707 // another destination space. Adjust the initial destination_count and, 708 // if necessary, set the source_region field if the partial object will 709 // cross a destination region boundary. 710 destination_count = split_info.destination_count(); 711 if (destination_count == 2) { 712 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 713 _region_data[dest_idx].set_source_region(cur_region); 714 } 715 } 716 717 HeapWord* const last_addr = dest_addr + words - 1; 718 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 719 const size_t dest_region_2 = addr_to_region_idx(last_addr); 720 721 // Initially assume that the destination regions will be the same and 722 // adjust the value below if necessary. Under this assumption, if 723 // cur_region == dest_region_2, then cur_region will be compacted 724 // completely into itself. 725 destination_count += cur_region == dest_region_2 ? 0 : 1; 726 if (dest_region_1 != dest_region_2) { 727 // Destination regions differ; adjust destination_count. 728 destination_count += 1; 729 // Data from cur_region will be copied to the start of dest_region_2. 730 _region_data[dest_region_2].set_source_region(cur_region); 731 } else if (region_offset(dest_addr) == 0) { 732 // Data from cur_region will be copied to the start of the destination 733 // region. 734 _region_data[dest_region_1].set_source_region(cur_region); 735 } 736 737 _region_data[cur_region].set_destination_count(destination_count); 738 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 739 dest_addr += words; 740 } 741 742 ++cur_region; 743 } 744 745 *target_next = dest_addr; 746 return true; 747 } 748 749 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 750 assert(addr != NULL, "Should detect NULL oop earlier"); 751 assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap"); 752 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked"); 753 754 // Region covering the object. 755 RegionData* const region_ptr = addr_to_region_ptr(addr); 756 HeapWord* result = region_ptr->destination(); 757 758 // If the entire Region is live, the new location is region->destination + the 759 // offset of the object within in the Region. 760 761 // Run some performance tests to determine if this special case pays off. It 762 // is worth it for pointers into the dense prefix. If the optimization to 763 // avoid pointer updates in regions that only point to the dense prefix is 764 // ever implemented, this should be revisited. 765 if (region_ptr->data_size() == RegionSize) { 766 result += region_offset(addr); 767 return result; 768 } 769 770 // Otherwise, the new location is region->destination + block offset + the 771 // number of live words in the Block that are (a) to the left of addr and (b) 772 // due to objects that start in the Block. 773 774 // Fill in the block table if necessary. This is unsynchronized, so multiple 775 // threads may fill the block table for a region (harmless, since it is 776 // idempotent). 777 if (!region_ptr->blocks_filled()) { 778 PSParallelCompact::fill_blocks(addr_to_region_idx(addr)); 779 region_ptr->set_blocks_filled(); 780 } 781 782 HeapWord* const search_start = block_align_down(addr); 783 const size_t block_offset = addr_to_block_ptr(addr)->offset(); 784 785 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 786 const size_t live = bitmap->live_words_in_range(search_start, oop(addr)); 787 result += block_offset + live; 788 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result)); 789 return result; 790 } 791 792 #ifdef ASSERT 793 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 794 { 795 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 796 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 797 for (const size_t* p = beg; p < end; ++p) { 798 assert(*p == 0, "not zero"); 799 } 800 } 801 802 void ParallelCompactData::verify_clear() 803 { 804 verify_clear(_region_vspace); 805 verify_clear(_block_vspace); 806 } 807 #endif // #ifdef ASSERT 808 809 STWGCTimer PSParallelCompact::_gc_timer; 810 ParallelOldTracer PSParallelCompact::_gc_tracer; 811 elapsedTimer PSParallelCompact::_accumulated_time; 812 unsigned int PSParallelCompact::_total_invocations = 0; 813 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 814 jlong PSParallelCompact::_time_of_last_gc = 0; 815 CollectorCounters* PSParallelCompact::_counters = NULL; 816 ParMarkBitMap PSParallelCompact::_mark_bitmap; 817 ParallelCompactData PSParallelCompact::_summary_data; 818 819 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 820 821 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 822 823 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 824 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 825 826 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 827 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 828 } 829 830 void PSParallelCompact::post_initialize() { 831 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 832 MemRegion mr = heap->reserved_region(); 833 _ref_processor = 834 new ReferenceProcessor(mr, // span 835 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 836 ParallelGCThreads, // mt processing degree 837 true, // mt discovery 838 ParallelGCThreads, // mt discovery degree 839 true, // atomic_discovery 840 &_is_alive_closure); // non-header is alive closure 841 _counters = new CollectorCounters("PSParallelCompact", 1); 842 843 // Initialize static fields in ParCompactionManager. 844 ParCompactionManager::initialize(mark_bitmap()); 845 } 846 847 bool PSParallelCompact::initialize() { 848 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 849 MemRegion mr = heap->reserved_region(); 850 851 // Was the old gen get allocated successfully? 852 if (!heap->old_gen()->is_allocated()) { 853 return false; 854 } 855 856 initialize_space_info(); 857 initialize_dead_wood_limiter(); 858 859 if (!_mark_bitmap.initialize(mr)) { 860 vm_shutdown_during_initialization( 861 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 862 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 863 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 864 return false; 865 } 866 867 if (!_summary_data.initialize(mr)) { 868 vm_shutdown_during_initialization( 869 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 870 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 871 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 872 return false; 873 } 874 875 return true; 876 } 877 878 void PSParallelCompact::initialize_space_info() 879 { 880 memset(&_space_info, 0, sizeof(_space_info)); 881 882 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 883 PSYoungGen* young_gen = heap->young_gen(); 884 885 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 886 _space_info[eden_space_id].set_space(young_gen->eden_space()); 887 _space_info[from_space_id].set_space(young_gen->from_space()); 888 _space_info[to_space_id].set_space(young_gen->to_space()); 889 890 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 891 } 892 893 void PSParallelCompact::initialize_dead_wood_limiter() 894 { 895 const size_t max = 100; 896 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 897 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 898 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 899 DEBUG_ONLY(_dwl_initialized = true;) 900 _dwl_adjustment = normal_distribution(1.0); 901 } 902 903 // Simple class for storing info about the heap at the start of GC, to be used 904 // after GC for comparison/printing. 905 class PreGCValues { 906 public: 907 PreGCValues() { } 908 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 909 910 void fill(ParallelScavengeHeap* heap) { 911 _heap_used = heap->used(); 912 _young_gen_used = heap->young_gen()->used_in_bytes(); 913 _old_gen_used = heap->old_gen()->used_in_bytes(); 914 _metadata_used = MetaspaceAux::used_bytes(); 915 }; 916 917 size_t heap_used() const { return _heap_used; } 918 size_t young_gen_used() const { return _young_gen_used; } 919 size_t old_gen_used() const { return _old_gen_used; } 920 size_t metadata_used() const { return _metadata_used; } 921 922 private: 923 size_t _heap_used; 924 size_t _young_gen_used; 925 size_t _old_gen_used; 926 size_t _metadata_used; 927 }; 928 929 void 930 PSParallelCompact::clear_data_covering_space(SpaceId id) 931 { 932 // At this point, top is the value before GC, new_top() is the value that will 933 // be set at the end of GC. The marking bitmap is cleared to top; nothing 934 // should be marked above top. The summary data is cleared to the larger of 935 // top & new_top. 936 MutableSpace* const space = _space_info[id].space(); 937 HeapWord* const bot = space->bottom(); 938 HeapWord* const top = space->top(); 939 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 940 941 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 942 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 943 _mark_bitmap.clear_range(beg_bit, end_bit); 944 945 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 946 const size_t end_region = 947 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 948 _summary_data.clear_range(beg_region, end_region); 949 950 // Clear the data used to 'split' regions. 951 SplitInfo& split_info = _space_info[id].split_info(); 952 if (split_info.is_valid()) { 953 split_info.clear(); 954 } 955 DEBUG_ONLY(split_info.verify_clear();) 956 } 957 958 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 959 { 960 // Update the from & to space pointers in space_info, since they are swapped 961 // at each young gen gc. Do the update unconditionally (even though a 962 // promotion failure does not swap spaces) because an unknown number of young 963 // collections will have swapped the spaces an unknown number of times. 964 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer); 965 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 966 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 967 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 968 969 pre_gc_values->fill(heap); 970 971 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 972 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 973 974 // Increment the invocation count 975 heap->increment_total_collections(true); 976 977 // We need to track unique mark sweep invocations as well. 978 _total_invocations++; 979 980 heap->print_heap_before_gc(); 981 heap->trace_heap_before_gc(&_gc_tracer); 982 983 // Fill in TLABs 984 heap->accumulate_statistics_all_tlabs(); 985 heap->ensure_parsability(true); // retire TLABs 986 987 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 988 HandleMark hm; // Discard invalid handles created during verification 989 Universe::verify(" VerifyBeforeGC:"); 990 } 991 992 // Verify object start arrays 993 if (VerifyObjectStartArray && 994 VerifyBeforeGC) { 995 heap->old_gen()->verify_object_start_array(); 996 } 997 998 DEBUG_ONLY(mark_bitmap()->verify_clear();) 999 DEBUG_ONLY(summary_data().verify_clear();) 1000 1001 // Have worker threads release resources the next time they run a task. 1002 gc_task_manager()->release_all_resources(); 1003 } 1004 1005 void PSParallelCompact::post_compact() 1006 { 1007 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer); 1008 1009 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1010 // Clear the marking bitmap, summary data and split info. 1011 clear_data_covering_space(SpaceId(id)); 1012 // Update top(). Must be done after clearing the bitmap and summary data. 1013 _space_info[id].publish_new_top(); 1014 } 1015 1016 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1017 MutableSpace* const from_space = _space_info[from_space_id].space(); 1018 MutableSpace* const to_space = _space_info[to_space_id].space(); 1019 1020 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1021 bool eden_empty = eden_space->is_empty(); 1022 if (!eden_empty) { 1023 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1024 heap->young_gen(), heap->old_gen()); 1025 } 1026 1027 // Update heap occupancy information which is used as input to the soft ref 1028 // clearing policy at the next gc. 1029 Universe::update_heap_info_at_gc(); 1030 1031 bool young_gen_empty = eden_empty && from_space->is_empty() && 1032 to_space->is_empty(); 1033 1034 ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set()); 1035 MemRegion old_mr = heap->old_gen()->reserved(); 1036 if (young_gen_empty) { 1037 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1038 } else { 1039 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1040 } 1041 1042 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1043 ClassLoaderDataGraph::purge(); 1044 MetaspaceAux::verify_metrics(); 1045 1046 CodeCache::gc_epilogue(); 1047 JvmtiExport::gc_epilogue(); 1048 1049 #if defined(COMPILER2) || INCLUDE_JVMCI 1050 DerivedPointerTable::update_pointers(); 1051 #endif 1052 1053 ref_processor()->enqueue_discovered_references(NULL); 1054 1055 if (ZapUnusedHeapArea) { 1056 heap->gen_mangle_unused_area(); 1057 } 1058 1059 // Update time of last GC 1060 reset_millis_since_last_gc(); 1061 } 1062 1063 HeapWord* 1064 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1065 bool maximum_compaction) 1066 { 1067 const size_t region_size = ParallelCompactData::RegionSize; 1068 const ParallelCompactData& sd = summary_data(); 1069 1070 const MutableSpace* const space = _space_info[id].space(); 1071 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1072 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1073 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1074 1075 // Skip full regions at the beginning of the space--they are necessarily part 1076 // of the dense prefix. 1077 size_t full_count = 0; 1078 const RegionData* cp; 1079 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1080 ++full_count; 1081 } 1082 1083 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1084 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1085 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1086 if (maximum_compaction || cp == end_cp || interval_ended) { 1087 _maximum_compaction_gc_num = total_invocations(); 1088 return sd.region_to_addr(cp); 1089 } 1090 1091 HeapWord* const new_top = _space_info[id].new_top(); 1092 const size_t space_live = pointer_delta(new_top, space->bottom()); 1093 const size_t space_used = space->used_in_words(); 1094 const size_t space_capacity = space->capacity_in_words(); 1095 1096 const double cur_density = double(space_live) / space_capacity; 1097 const double deadwood_density = 1098 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1099 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1100 1101 if (TraceParallelOldGCDensePrefix) { 1102 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1103 cur_density, deadwood_density, deadwood_goal); 1104 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1105 "space_cap=" SIZE_FORMAT, 1106 space_live, space_used, 1107 space_capacity); 1108 } 1109 1110 // XXX - Use binary search? 1111 HeapWord* dense_prefix = sd.region_to_addr(cp); 1112 const RegionData* full_cp = cp; 1113 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1114 while (cp < end_cp) { 1115 HeapWord* region_destination = cp->destination(); 1116 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1117 if (TraceParallelOldGCDensePrefix && Verbose) { 1118 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1119 "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8), 1120 sd.region(cp), p2i(region_destination), 1121 p2i(dense_prefix), cur_deadwood); 1122 } 1123 1124 if (cur_deadwood >= deadwood_goal) { 1125 // Found the region that has the correct amount of deadwood to the left. 1126 // This typically occurs after crossing a fairly sparse set of regions, so 1127 // iterate backwards over those sparse regions, looking for the region 1128 // that has the lowest density of live objects 'to the right.' 1129 size_t space_to_left = sd.region(cp) * region_size; 1130 size_t live_to_left = space_to_left - cur_deadwood; 1131 size_t space_to_right = space_capacity - space_to_left; 1132 size_t live_to_right = space_live - live_to_left; 1133 double density_to_right = double(live_to_right) / space_to_right; 1134 while (cp > full_cp) { 1135 --cp; 1136 const size_t prev_region_live_to_right = live_to_right - 1137 cp->data_size(); 1138 const size_t prev_region_space_to_right = space_to_right + region_size; 1139 double prev_region_density_to_right = 1140 double(prev_region_live_to_right) / prev_region_space_to_right; 1141 if (density_to_right <= prev_region_density_to_right) { 1142 return dense_prefix; 1143 } 1144 if (TraceParallelOldGCDensePrefix && Verbose) { 1145 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1146 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1147 prev_region_density_to_right); 1148 } 1149 dense_prefix -= region_size; 1150 live_to_right = prev_region_live_to_right; 1151 space_to_right = prev_region_space_to_right; 1152 density_to_right = prev_region_density_to_right; 1153 } 1154 return dense_prefix; 1155 } 1156 1157 dense_prefix += region_size; 1158 ++cp; 1159 } 1160 1161 return dense_prefix; 1162 } 1163 1164 #ifndef PRODUCT 1165 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1166 const SpaceId id, 1167 const bool maximum_compaction, 1168 HeapWord* const addr) 1169 { 1170 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1171 RegionData* const cp = summary_data().region(region_idx); 1172 const MutableSpace* const space = _space_info[id].space(); 1173 HeapWord* const new_top = _space_info[id].new_top(); 1174 1175 const size_t space_live = pointer_delta(new_top, space->bottom()); 1176 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1177 const size_t space_cap = space->capacity_in_words(); 1178 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1179 const size_t live_to_right = new_top - cp->destination(); 1180 const size_t dead_to_right = space->top() - addr - live_to_right; 1181 1182 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1183 "spl=" SIZE_FORMAT " " 1184 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1185 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1186 " ratio=%10.8f", 1187 algorithm, p2i(addr), region_idx, 1188 space_live, 1189 dead_to_left, dead_to_left_pct, 1190 dead_to_right, live_to_right, 1191 double(dead_to_right) / live_to_right); 1192 } 1193 #endif // #ifndef PRODUCT 1194 1195 // Return a fraction indicating how much of the generation can be treated as 1196 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1197 // based on the density of live objects in the generation to determine a limit, 1198 // which is then adjusted so the return value is min_percent when the density is 1199 // 1. 1200 // 1201 // The following table shows some return values for a different values of the 1202 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1203 // min_percent is 1. 1204 // 1205 // fraction allowed as dead wood 1206 // ----------------------------------------------------------------- 1207 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1208 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1209 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1210 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1211 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1212 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1213 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1214 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1215 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1216 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1217 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1218 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1219 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1220 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1221 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1222 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1223 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1224 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1225 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1226 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1227 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1228 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1229 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1230 1231 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1232 { 1233 assert(_dwl_initialized, "uninitialized"); 1234 1235 // The raw limit is the value of the normal distribution at x = density. 1236 const double raw_limit = normal_distribution(density); 1237 1238 // Adjust the raw limit so it becomes the minimum when the density is 1. 1239 // 1240 // First subtract the adjustment value (which is simply the precomputed value 1241 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1242 // Then add the minimum value, so the minimum is returned when the density is 1243 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1244 const double min = double(min_percent) / 100.0; 1245 const double limit = raw_limit - _dwl_adjustment + min; 1246 return MAX2(limit, 0.0); 1247 } 1248 1249 ParallelCompactData::RegionData* 1250 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1251 const RegionData* end) 1252 { 1253 const size_t region_size = ParallelCompactData::RegionSize; 1254 ParallelCompactData& sd = summary_data(); 1255 size_t left = sd.region(beg); 1256 size_t right = end > beg ? sd.region(end) - 1 : left; 1257 1258 // Binary search. 1259 while (left < right) { 1260 // Equivalent to (left + right) / 2, but does not overflow. 1261 const size_t middle = left + (right - left) / 2; 1262 RegionData* const middle_ptr = sd.region(middle); 1263 HeapWord* const dest = middle_ptr->destination(); 1264 HeapWord* const addr = sd.region_to_addr(middle); 1265 assert(dest != NULL, "sanity"); 1266 assert(dest <= addr, "must move left"); 1267 1268 if (middle > left && dest < addr) { 1269 right = middle - 1; 1270 } else if (middle < right && middle_ptr->data_size() == region_size) { 1271 left = middle + 1; 1272 } else { 1273 return middle_ptr; 1274 } 1275 } 1276 return sd.region(left); 1277 } 1278 1279 ParallelCompactData::RegionData* 1280 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1281 const RegionData* end, 1282 size_t dead_words) 1283 { 1284 ParallelCompactData& sd = summary_data(); 1285 size_t left = sd.region(beg); 1286 size_t right = end > beg ? sd.region(end) - 1 : left; 1287 1288 // Binary search. 1289 while (left < right) { 1290 // Equivalent to (left + right) / 2, but does not overflow. 1291 const size_t middle = left + (right - left) / 2; 1292 RegionData* const middle_ptr = sd.region(middle); 1293 HeapWord* const dest = middle_ptr->destination(); 1294 HeapWord* const addr = sd.region_to_addr(middle); 1295 assert(dest != NULL, "sanity"); 1296 assert(dest <= addr, "must move left"); 1297 1298 const size_t dead_to_left = pointer_delta(addr, dest); 1299 if (middle > left && dead_to_left > dead_words) { 1300 right = middle - 1; 1301 } else if (middle < right && dead_to_left < dead_words) { 1302 left = middle + 1; 1303 } else { 1304 return middle_ptr; 1305 } 1306 } 1307 return sd.region(left); 1308 } 1309 1310 // The result is valid during the summary phase, after the initial summarization 1311 // of each space into itself, and before final summarization. 1312 inline double 1313 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1314 HeapWord* const bottom, 1315 HeapWord* const top, 1316 HeapWord* const new_top) 1317 { 1318 ParallelCompactData& sd = summary_data(); 1319 1320 assert(cp != NULL, "sanity"); 1321 assert(bottom != NULL, "sanity"); 1322 assert(top != NULL, "sanity"); 1323 assert(new_top != NULL, "sanity"); 1324 assert(top >= new_top, "summary data problem?"); 1325 assert(new_top > bottom, "space is empty; should not be here"); 1326 assert(new_top >= cp->destination(), "sanity"); 1327 assert(top >= sd.region_to_addr(cp), "sanity"); 1328 1329 HeapWord* const destination = cp->destination(); 1330 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1331 const size_t compacted_region_live = pointer_delta(new_top, destination); 1332 const size_t compacted_region_used = pointer_delta(top, 1333 sd.region_to_addr(cp)); 1334 const size_t reclaimable = compacted_region_used - compacted_region_live; 1335 1336 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1337 return double(reclaimable) / divisor; 1338 } 1339 1340 // Return the address of the end of the dense prefix, a.k.a. the start of the 1341 // compacted region. The address is always on a region boundary. 1342 // 1343 // Completely full regions at the left are skipped, since no compaction can 1344 // occur in those regions. Then the maximum amount of dead wood to allow is 1345 // computed, based on the density (amount live / capacity) of the generation; 1346 // the region with approximately that amount of dead space to the left is 1347 // identified as the limit region. Regions between the last completely full 1348 // region and the limit region are scanned and the one that has the best 1349 // (maximum) reclaimed_ratio() is selected. 1350 HeapWord* 1351 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1352 bool maximum_compaction) 1353 { 1354 const size_t region_size = ParallelCompactData::RegionSize; 1355 const ParallelCompactData& sd = summary_data(); 1356 1357 const MutableSpace* const space = _space_info[id].space(); 1358 HeapWord* const top = space->top(); 1359 HeapWord* const top_aligned_up = sd.region_align_up(top); 1360 HeapWord* const new_top = _space_info[id].new_top(); 1361 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1362 HeapWord* const bottom = space->bottom(); 1363 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1364 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1365 const RegionData* const new_top_cp = 1366 sd.addr_to_region_ptr(new_top_aligned_up); 1367 1368 // Skip full regions at the beginning of the space--they are necessarily part 1369 // of the dense prefix. 1370 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1371 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1372 space->is_empty(), "no dead space allowed to the left"); 1373 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1374 "region must have dead space"); 1375 1376 // The gc number is saved whenever a maximum compaction is done, and used to 1377 // determine when the maximum compaction interval has expired. This avoids 1378 // successive max compactions for different reasons. 1379 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1380 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1381 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1382 total_invocations() == HeapFirstMaximumCompactionCount; 1383 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1384 _maximum_compaction_gc_num = total_invocations(); 1385 return sd.region_to_addr(full_cp); 1386 } 1387 1388 const size_t space_live = pointer_delta(new_top, bottom); 1389 const size_t space_used = space->used_in_words(); 1390 const size_t space_capacity = space->capacity_in_words(); 1391 1392 const double density = double(space_live) / double(space_capacity); 1393 const size_t min_percent_free = MarkSweepDeadRatio; 1394 const double limiter = dead_wood_limiter(density, min_percent_free); 1395 const size_t dead_wood_max = space_used - space_live; 1396 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1397 dead_wood_max); 1398 1399 if (TraceParallelOldGCDensePrefix) { 1400 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1401 "space_cap=" SIZE_FORMAT, 1402 space_live, space_used, 1403 space_capacity); 1404 tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f " 1405 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1406 density, min_percent_free, limiter, 1407 dead_wood_max, dead_wood_limit); 1408 } 1409 1410 // Locate the region with the desired amount of dead space to the left. 1411 const RegionData* const limit_cp = 1412 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1413 1414 // Scan from the first region with dead space to the limit region and find the 1415 // one with the best (largest) reclaimed ratio. 1416 double best_ratio = 0.0; 1417 const RegionData* best_cp = full_cp; 1418 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1419 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1420 if (tmp_ratio > best_ratio) { 1421 best_cp = cp; 1422 best_ratio = tmp_ratio; 1423 } 1424 } 1425 1426 return sd.region_to_addr(best_cp); 1427 } 1428 1429 void PSParallelCompact::summarize_spaces_quick() 1430 { 1431 for (unsigned int i = 0; i < last_space_id; ++i) { 1432 const MutableSpace* space = _space_info[i].space(); 1433 HeapWord** nta = _space_info[i].new_top_addr(); 1434 bool result = _summary_data.summarize(_space_info[i].split_info(), 1435 space->bottom(), space->top(), NULL, 1436 space->bottom(), space->end(), nta); 1437 assert(result, "space must fit into itself"); 1438 _space_info[i].set_dense_prefix(space->bottom()); 1439 } 1440 } 1441 1442 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1443 { 1444 HeapWord* const dense_prefix_end = dense_prefix(id); 1445 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1446 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1447 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1448 // Only enough dead space is filled so that any remaining dead space to the 1449 // left is larger than the minimum filler object. (The remainder is filled 1450 // during the copy/update phase.) 1451 // 1452 // The size of the dead space to the right of the boundary is not a 1453 // concern, since compaction will be able to use whatever space is 1454 // available. 1455 // 1456 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1457 // surrounds the space to be filled with an object. 1458 // 1459 // In the 32-bit VM, each bit represents two 32-bit words: 1460 // +---+ 1461 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1462 // end_bits: ... x x x | 0 | || 0 x x ... 1463 // +---+ 1464 // 1465 // In the 64-bit VM, each bit represents one 64-bit word: 1466 // +------------+ 1467 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1468 // end_bits: ... x x 1 | 0 || 0 | x x ... 1469 // +------------+ 1470 // +-------+ 1471 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1472 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1473 // +-------+ 1474 // +-----------+ 1475 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1476 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1477 // +-----------+ 1478 // +-------+ 1479 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1480 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1481 // +-------+ 1482 1483 // Initially assume case a, c or e will apply. 1484 size_t obj_len = CollectedHeap::min_fill_size(); 1485 HeapWord* obj_beg = dense_prefix_end - obj_len; 1486 1487 #ifdef _LP64 1488 if (MinObjAlignment > 1) { // object alignment > heap word size 1489 // Cases a, c or e. 1490 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1491 // Case b above. 1492 obj_beg = dense_prefix_end - 1; 1493 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1494 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1495 // Case d above. 1496 obj_beg = dense_prefix_end - 3; 1497 obj_len = 3; 1498 } 1499 #endif // #ifdef _LP64 1500 1501 CollectedHeap::fill_with_object(obj_beg, obj_len); 1502 _mark_bitmap.mark_obj(obj_beg, obj_len); 1503 _summary_data.add_obj(obj_beg, obj_len); 1504 assert(start_array(id) != NULL, "sanity"); 1505 start_array(id)->allocate_block(obj_beg); 1506 } 1507 } 1508 1509 void 1510 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1511 { 1512 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1513 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1514 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1515 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1516 cur->set_source_region(0); 1517 } 1518 } 1519 1520 void 1521 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1522 { 1523 assert(id < last_space_id, "id out of range"); 1524 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(), 1525 "should have been reset in summarize_spaces_quick()"); 1526 1527 const MutableSpace* space = _space_info[id].space(); 1528 if (_space_info[id].new_top() != space->bottom()) { 1529 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1530 _space_info[id].set_dense_prefix(dense_prefix_end); 1531 1532 #ifndef PRODUCT 1533 if (TraceParallelOldGCDensePrefix) { 1534 print_dense_prefix_stats("ratio", id, maximum_compaction, 1535 dense_prefix_end); 1536 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1537 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1538 } 1539 #endif // #ifndef PRODUCT 1540 1541 // Recompute the summary data, taking into account the dense prefix. If 1542 // every last byte will be reclaimed, then the existing summary data which 1543 // compacts everything can be left in place. 1544 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1545 // If dead space crosses the dense prefix boundary, it is (at least 1546 // partially) filled with a dummy object, marked live and added to the 1547 // summary data. This simplifies the copy/update phase and must be done 1548 // before the final locations of objects are determined, to prevent 1549 // leaving a fragment of dead space that is too small to fill. 1550 fill_dense_prefix_end(id); 1551 1552 // Compute the destination of each Region, and thus each object. 1553 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1554 _summary_data.summarize(_space_info[id].split_info(), 1555 dense_prefix_end, space->top(), NULL, 1556 dense_prefix_end, space->end(), 1557 _space_info[id].new_top_addr()); 1558 } 1559 } 1560 1561 if (TraceParallelOldGCSummaryPhase) { 1562 const size_t region_size = ParallelCompactData::RegionSize; 1563 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1564 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1565 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1566 HeapWord* const new_top = _space_info[id].new_top(); 1567 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1568 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1569 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1570 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1571 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1572 id, space->capacity_in_words(), p2i(dense_prefix_end), 1573 dp_region, dp_words / region_size, 1574 cr_words / region_size, p2i(new_top)); 1575 } 1576 } 1577 1578 #ifndef PRODUCT 1579 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1580 HeapWord* dst_beg, HeapWord* dst_end, 1581 SpaceId src_space_id, 1582 HeapWord* src_beg, HeapWord* src_end) 1583 { 1584 if (TraceParallelOldGCSummaryPhase) { 1585 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1586 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1587 SIZE_FORMAT "-" SIZE_FORMAT " " 1588 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1589 SIZE_FORMAT "-" SIZE_FORMAT, 1590 src_space_id, space_names[src_space_id], 1591 dst_space_id, space_names[dst_space_id], 1592 p2i(src_beg), p2i(src_end), 1593 _summary_data.addr_to_region_idx(src_beg), 1594 _summary_data.addr_to_region_idx(src_end), 1595 p2i(dst_beg), p2i(dst_end), 1596 _summary_data.addr_to_region_idx(dst_beg), 1597 _summary_data.addr_to_region_idx(dst_end)); 1598 } 1599 } 1600 #endif // #ifndef PRODUCT 1601 1602 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1603 bool maximum_compaction) 1604 { 1605 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer); 1606 // trace("2"); 1607 1608 #ifdef ASSERT 1609 if (TraceParallelOldGCMarkingPhase) { 1610 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1611 "add_obj_bytes=" SIZE_FORMAT, 1612 add_obj_count, add_obj_size * HeapWordSize); 1613 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1614 "mark_bitmap_bytes=" SIZE_FORMAT, 1615 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1616 } 1617 #endif // #ifdef ASSERT 1618 1619 // Quick summarization of each space into itself, to see how much is live. 1620 summarize_spaces_quick(); 1621 1622 if (TraceParallelOldGCSummaryPhase) { 1623 tty->print_cr("summary_phase: after summarizing each space to self"); 1624 Universe::print(); 1625 NOT_PRODUCT(print_region_ranges()); 1626 if (Verbose) { 1627 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1628 } 1629 } 1630 1631 // The amount of live data that will end up in old space (assuming it fits). 1632 size_t old_space_total_live = 0; 1633 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1634 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1635 _space_info[id].space()->bottom()); 1636 } 1637 1638 MutableSpace* const old_space = _space_info[old_space_id].space(); 1639 const size_t old_capacity = old_space->capacity_in_words(); 1640 if (old_space_total_live > old_capacity) { 1641 // XXX - should also try to expand 1642 maximum_compaction = true; 1643 } 1644 1645 // Old generations. 1646 summarize_space(old_space_id, maximum_compaction); 1647 1648 // Summarize the remaining spaces in the young gen. The initial target space 1649 // is the old gen. If a space does not fit entirely into the target, then the 1650 // remainder is compacted into the space itself and that space becomes the new 1651 // target. 1652 SpaceId dst_space_id = old_space_id; 1653 HeapWord* dst_space_end = old_space->end(); 1654 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1655 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1656 const MutableSpace* space = _space_info[id].space(); 1657 const size_t live = pointer_delta(_space_info[id].new_top(), 1658 space->bottom()); 1659 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1660 1661 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1662 SpaceId(id), space->bottom(), space->top());) 1663 if (live > 0 && live <= available) { 1664 // All the live data will fit. 1665 bool done = _summary_data.summarize(_space_info[id].split_info(), 1666 space->bottom(), space->top(), 1667 NULL, 1668 *new_top_addr, dst_space_end, 1669 new_top_addr); 1670 assert(done, "space must fit into old gen"); 1671 1672 // Reset the new_top value for the space. 1673 _space_info[id].set_new_top(space->bottom()); 1674 } else if (live > 0) { 1675 // Attempt to fit part of the source space into the target space. 1676 HeapWord* next_src_addr = NULL; 1677 bool done = _summary_data.summarize(_space_info[id].split_info(), 1678 space->bottom(), space->top(), 1679 &next_src_addr, 1680 *new_top_addr, dst_space_end, 1681 new_top_addr); 1682 assert(!done, "space should not fit into old gen"); 1683 assert(next_src_addr != NULL, "sanity"); 1684 1685 // The source space becomes the new target, so the remainder is compacted 1686 // within the space itself. 1687 dst_space_id = SpaceId(id); 1688 dst_space_end = space->end(); 1689 new_top_addr = _space_info[id].new_top_addr(); 1690 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1691 space->bottom(), dst_space_end, 1692 SpaceId(id), next_src_addr, space->top());) 1693 done = _summary_data.summarize(_space_info[id].split_info(), 1694 next_src_addr, space->top(), 1695 NULL, 1696 space->bottom(), dst_space_end, 1697 new_top_addr); 1698 assert(done, "space must fit when compacted into itself"); 1699 assert(*new_top_addr <= space->top(), "usage should not grow"); 1700 } 1701 } 1702 1703 if (TraceParallelOldGCSummaryPhase) { 1704 tty->print_cr("summary_phase: after final summarization"); 1705 Universe::print(); 1706 NOT_PRODUCT(print_region_ranges()); 1707 if (Verbose) { 1708 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1709 } 1710 } 1711 } 1712 1713 // This method should contain all heap-specific policy for invoking a full 1714 // collection. invoke_no_policy() will only attempt to compact the heap; it 1715 // will do nothing further. If we need to bail out for policy reasons, scavenge 1716 // before full gc, or any other specialized behavior, it needs to be added here. 1717 // 1718 // Note that this method should only be called from the vm_thread while at a 1719 // safepoint. 1720 // 1721 // Note that the all_soft_refs_clear flag in the collector policy 1722 // may be true because this method can be called without intervening 1723 // activity. For example when the heap space is tight and full measure 1724 // are being taken to free space. 1725 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1726 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1727 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1728 "should be in vm thread"); 1729 1730 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1731 GCCause::Cause gc_cause = heap->gc_cause(); 1732 assert(!heap->is_gc_active(), "not reentrant"); 1733 1734 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1735 IsGCActiveMark mark; 1736 1737 if (ScavengeBeforeFullGC) { 1738 PSScavenge::invoke_no_policy(); 1739 } 1740 1741 const bool clear_all_soft_refs = 1742 heap->collector_policy()->should_clear_all_soft_refs(); 1743 1744 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1745 maximum_heap_compaction); 1746 } 1747 1748 // This method contains no policy. You should probably 1749 // be calling invoke() instead. 1750 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1751 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1752 assert(ref_processor() != NULL, "Sanity"); 1753 1754 if (GC_locker::check_active_before_gc()) { 1755 return false; 1756 } 1757 1758 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 1759 1760 GCIdMark gc_id_mark; 1761 _gc_timer.register_gc_start(); 1762 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1763 1764 TimeStamp marking_start; 1765 TimeStamp compaction_start; 1766 TimeStamp collection_exit; 1767 1768 GCCause::Cause gc_cause = heap->gc_cause(); 1769 PSYoungGen* young_gen = heap->young_gen(); 1770 PSOldGen* old_gen = heap->old_gen(); 1771 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1772 1773 // The scope of casr should end after code that can change 1774 // CollectorPolicy::_should_clear_all_soft_refs. 1775 ClearedAllSoftRefs casr(maximum_heap_compaction, 1776 heap->collector_policy()); 1777 1778 if (ZapUnusedHeapArea) { 1779 // Save information needed to minimize mangling 1780 heap->record_gen_tops_before_GC(); 1781 } 1782 1783 heap->pre_full_gc_dump(&_gc_timer); 1784 1785 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 1786 1787 // Make sure data structures are sane, make the heap parsable, and do other 1788 // miscellaneous bookkeeping. 1789 PreGCValues pre_gc_values; 1790 pre_compact(&pre_gc_values); 1791 1792 // Get the compaction manager reserved for the VM thread. 1793 ParCompactionManager* const vmthread_cm = 1794 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1795 1796 // Place after pre_compact() where the number of invocations is incremented. 1797 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 1798 1799 { 1800 ResourceMark rm; 1801 HandleMark hm; 1802 1803 // Set the number of GC threads to be used in this collection 1804 gc_task_manager()->set_active_gang(); 1805 gc_task_manager()->task_idle_workers(); 1806 1807 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 1808 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 1809 TraceCollectorStats tcs(counters()); 1810 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 1811 1812 if (TraceOldGenTime) accumulated_time()->start(); 1813 1814 // Let the size policy know we're starting 1815 size_policy->major_collection_begin(); 1816 1817 CodeCache::gc_prologue(); 1818 1819 #if defined(COMPILER2) || INCLUDE_JVMCI 1820 DerivedPointerTable::clear(); 1821 #endif 1822 1823 ref_processor()->enable_discovery(); 1824 ref_processor()->setup_policy(maximum_heap_compaction); 1825 1826 bool marked_for_unloading = false; 1827 1828 marking_start.update(); 1829 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 1830 1831 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 1832 && GCCause::is_user_requested_gc(gc_cause); 1833 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 1834 1835 #if defined(COMPILER2) || INCLUDE_JVMCI 1836 assert(DerivedPointerTable::is_active(), "Sanity"); 1837 DerivedPointerTable::set_active(false); 1838 #endif 1839 1840 // adjust_roots() updates Universe::_intArrayKlassObj which is 1841 // needed by the compaction for filling holes in the dense prefix. 1842 adjust_roots(); 1843 1844 compaction_start.update(); 1845 compact(); 1846 1847 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 1848 // done before resizing. 1849 post_compact(); 1850 1851 // Let the size policy know we're done 1852 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 1853 1854 if (UseAdaptiveSizePolicy) { 1855 if (PrintAdaptiveSizePolicy) { 1856 gclog_or_tty->print("AdaptiveSizeStart: "); 1857 gclog_or_tty->stamp(); 1858 gclog_or_tty->print_cr(" collection: %d ", 1859 heap->total_collections()); 1860 if (Verbose) { 1861 gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT 1862 " young_gen_capacity: " SIZE_FORMAT, 1863 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 1864 } 1865 } 1866 1867 // Don't check if the size_policy is ready here. Let 1868 // the size_policy check that internally. 1869 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 1870 AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) { 1871 // Swap the survivor spaces if from_space is empty. The 1872 // resize_young_gen() called below is normally used after 1873 // a successful young GC and swapping of survivor spaces; 1874 // otherwise, it will fail to resize the young gen with 1875 // the current implementation. 1876 if (young_gen->from_space()->is_empty()) { 1877 young_gen->from_space()->clear(SpaceDecorator::Mangle); 1878 young_gen->swap_spaces(); 1879 } 1880 1881 // Calculate optimal free space amounts 1882 assert(young_gen->max_size() > 1883 young_gen->from_space()->capacity_in_bytes() + 1884 young_gen->to_space()->capacity_in_bytes(), 1885 "Sizes of space in young gen are out-of-bounds"); 1886 1887 size_t young_live = young_gen->used_in_bytes(); 1888 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 1889 size_t old_live = old_gen->used_in_bytes(); 1890 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 1891 size_t max_old_gen_size = old_gen->max_gen_size(); 1892 size_t max_eden_size = young_gen->max_size() - 1893 young_gen->from_space()->capacity_in_bytes() - 1894 young_gen->to_space()->capacity_in_bytes(); 1895 1896 // Used for diagnostics 1897 size_policy->clear_generation_free_space_flags(); 1898 1899 size_policy->compute_generations_free_space(young_live, 1900 eden_live, 1901 old_live, 1902 cur_eden, 1903 max_old_gen_size, 1904 max_eden_size, 1905 true /* full gc*/); 1906 1907 size_policy->check_gc_overhead_limit(young_live, 1908 eden_live, 1909 max_old_gen_size, 1910 max_eden_size, 1911 true /* full gc*/, 1912 gc_cause, 1913 heap->collector_policy()); 1914 1915 size_policy->decay_supplemental_growth(true /* full gc*/); 1916 1917 heap->resize_old_gen( 1918 size_policy->calculated_old_free_size_in_bytes()); 1919 1920 heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(), 1921 size_policy->calculated_survivor_size_in_bytes()); 1922 } 1923 if (PrintAdaptiveSizePolicy) { 1924 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 1925 heap->total_collections()); 1926 } 1927 } 1928 1929 if (UsePerfData) { 1930 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 1931 counters->update_counters(); 1932 counters->update_old_capacity(old_gen->capacity_in_bytes()); 1933 counters->update_young_capacity(young_gen->capacity_in_bytes()); 1934 } 1935 1936 heap->resize_all_tlabs(); 1937 1938 // Resize the metaspace capacity after a collection 1939 MetaspaceGC::compute_new_size(); 1940 1941 if (TraceOldGenTime) accumulated_time()->stop(); 1942 1943 if (PrintGC) { 1944 if (PrintGCDetails) { 1945 // No GC timestamp here. This is after GC so it would be confusing. 1946 young_gen->print_used_change(pre_gc_values.young_gen_used()); 1947 old_gen->print_used_change(pre_gc_values.old_gen_used()); 1948 heap->print_heap_change(pre_gc_values.heap_used()); 1949 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 1950 } else { 1951 heap->print_heap_change(pre_gc_values.heap_used()); 1952 } 1953 } 1954 1955 // Track memory usage and detect low memory 1956 MemoryService::track_memory_usage(); 1957 heap->update_counters(); 1958 gc_task_manager()->release_idle_workers(); 1959 } 1960 1961 #ifdef ASSERT 1962 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 1963 ParCompactionManager* const cm = 1964 ParCompactionManager::manager_array(int(i)); 1965 assert(cm->marking_stack()->is_empty(), "should be empty"); 1966 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 1967 } 1968 #endif // ASSERT 1969 1970 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 1971 HandleMark hm; // Discard invalid handles created during verification 1972 Universe::verify(" VerifyAfterGC:"); 1973 } 1974 1975 // Re-verify object start arrays 1976 if (VerifyObjectStartArray && 1977 VerifyAfterGC) { 1978 old_gen->verify_object_start_array(); 1979 } 1980 1981 if (ZapUnusedHeapArea) { 1982 old_gen->object_space()->check_mangled_unused_area_complete(); 1983 } 1984 1985 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 1986 1987 collection_exit.update(); 1988 1989 heap->print_heap_after_gc(); 1990 heap->trace_heap_after_gc(&_gc_tracer); 1991 1992 if (PrintGCTaskTimeStamps) { 1993 gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " 1994 JLONG_FORMAT, 1995 marking_start.ticks(), compaction_start.ticks(), 1996 collection_exit.ticks()); 1997 gc_task_manager()->print_task_time_stamps(); 1998 } 1999 2000 heap->post_full_gc_dump(&_gc_timer); 2001 2002 #ifdef TRACESPINNING 2003 ParallelTaskTerminator::print_termination_counts(); 2004 #endif 2005 2006 _gc_timer.register_gc_end(); 2007 2008 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2009 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2010 2011 return true; 2012 } 2013 2014 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2015 PSYoungGen* young_gen, 2016 PSOldGen* old_gen) { 2017 MutableSpace* const eden_space = young_gen->eden_space(); 2018 assert(!eden_space->is_empty(), "eden must be non-empty"); 2019 assert(young_gen->virtual_space()->alignment() == 2020 old_gen->virtual_space()->alignment(), "alignments do not match"); 2021 2022 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2023 return false; 2024 } 2025 2026 // Both generations must be completely committed. 2027 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2028 return false; 2029 } 2030 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2031 return false; 2032 } 2033 2034 // Figure out how much to take from eden. Include the average amount promoted 2035 // in the total; otherwise the next young gen GC will simply bail out to a 2036 // full GC. 2037 const size_t alignment = old_gen->virtual_space()->alignment(); 2038 const size_t eden_used = eden_space->used_in_bytes(); 2039 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2040 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2041 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2042 2043 if (absorb_size >= eden_capacity) { 2044 return false; // Must leave some space in eden. 2045 } 2046 2047 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2048 if (new_young_size < young_gen->min_gen_size()) { 2049 return false; // Respect young gen minimum size. 2050 } 2051 2052 if (TraceAdaptiveGCBoundary && Verbose) { 2053 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2054 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2055 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2056 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2057 absorb_size / K, 2058 eden_capacity / K, (eden_capacity - absorb_size) / K, 2059 young_gen->from_space()->used_in_bytes() / K, 2060 young_gen->to_space()->used_in_bytes() / K, 2061 young_gen->capacity_in_bytes() / K, new_young_size / K); 2062 } 2063 2064 // Fill the unused part of the old gen. 2065 MutableSpace* const old_space = old_gen->object_space(); 2066 HeapWord* const unused_start = old_space->top(); 2067 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2068 2069 if (unused_words > 0) { 2070 if (unused_words < CollectedHeap::min_fill_size()) { 2071 return false; // If the old gen cannot be filled, must give up. 2072 } 2073 CollectedHeap::fill_with_objects(unused_start, unused_words); 2074 } 2075 2076 // Take the live data from eden and set both top and end in the old gen to 2077 // eden top. (Need to set end because reset_after_change() mangles the region 2078 // from end to virtual_space->high() in debug builds). 2079 HeapWord* const new_top = eden_space->top(); 2080 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2081 absorb_size); 2082 young_gen->reset_after_change(); 2083 old_space->set_top(new_top); 2084 old_space->set_end(new_top); 2085 old_gen->reset_after_change(); 2086 2087 // Update the object start array for the filler object and the data from eden. 2088 ObjectStartArray* const start_array = old_gen->start_array(); 2089 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2090 start_array->allocate_block(p); 2091 } 2092 2093 // Could update the promoted average here, but it is not typically updated at 2094 // full GCs and the value to use is unclear. Something like 2095 // 2096 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2097 2098 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2099 return true; 2100 } 2101 2102 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2103 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2104 "shouldn't return NULL"); 2105 return ParallelScavengeHeap::gc_task_manager(); 2106 } 2107 2108 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2109 bool maximum_heap_compaction, 2110 ParallelOldTracer *gc_tracer) { 2111 // Recursively traverse all live objects and mark them 2112 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer); 2113 2114 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2115 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2116 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2117 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2118 ParallelTaskTerminator terminator(active_gc_threads, qset); 2119 2120 ParCompactionManager::MarkAndPushClosure mark_and_push_closure(cm); 2121 ParCompactionManager::FollowStackClosure follow_stack_closure(cm); 2122 2123 // Need new claim bits before marking starts. 2124 ClassLoaderDataGraph::clear_claimed_marks(); 2125 2126 { 2127 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer); 2128 2129 ParallelScavengeHeap::ParStrongRootsScope psrs; 2130 2131 GCTaskQueue* q = GCTaskQueue::create(); 2132 2133 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2134 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2135 // We scan the thread roots in parallel 2136 Threads::create_thread_roots_marking_tasks(q); 2137 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2138 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2139 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2140 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2141 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2142 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2143 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2144 2145 if (active_gc_threads > 1) { 2146 for (uint j = 0; j < active_gc_threads; j++) { 2147 q->enqueue(new StealMarkingTask(&terminator)); 2148 } 2149 } 2150 2151 gc_task_manager()->execute_and_wait(q); 2152 } 2153 2154 // Process reference objects found during marking 2155 { 2156 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer); 2157 2158 ReferenceProcessorStats stats; 2159 if (ref_processor()->processing_is_mt()) { 2160 RefProcTaskExecutor task_executor; 2161 stats = ref_processor()->process_discovered_references( 2162 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2163 &task_executor, &_gc_timer); 2164 } else { 2165 stats = ref_processor()->process_discovered_references( 2166 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2167 &_gc_timer); 2168 } 2169 2170 gc_tracer->report_gc_reference_stats(stats); 2171 } 2172 2173 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer); 2174 2175 // This is the point where the entire marking should have completed. 2176 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2177 2178 // Follow system dictionary roots and unload classes. 2179 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2180 2181 // Unload nmethods. 2182 CodeCache::do_unloading(is_alive_closure(), purged_class); 2183 2184 // Prune dead klasses from subklass/sibling/implementor lists. 2185 Klass::clean_weak_klass_links(is_alive_closure()); 2186 2187 // Delete entries for dead interned strings. 2188 StringTable::unlink(is_alive_closure()); 2189 2190 // Clean up unreferenced symbols in symbol table. 2191 SymbolTable::unlink(); 2192 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2193 } 2194 2195 // This should be moved to the shared markSweep code! 2196 class PSAlwaysTrueClosure: public BoolObjectClosure { 2197 public: 2198 bool do_object_b(oop p) { return true; } 2199 }; 2200 static PSAlwaysTrueClosure always_true; 2201 2202 void PSParallelCompact::adjust_roots() { 2203 // Adjust the pointers to reflect the new locations 2204 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer); 2205 2206 // Need new claim bits when tracing through and adjusting pointers. 2207 ClassLoaderDataGraph::clear_claimed_marks(); 2208 2209 // General strong roots. 2210 Universe::oops_do(adjust_pointer_closure()); 2211 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2212 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2213 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2214 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2215 FlatProfiler::oops_do(adjust_pointer_closure()); 2216 Management::oops_do(adjust_pointer_closure()); 2217 JvmtiExport::oops_do(adjust_pointer_closure()); 2218 SystemDictionary::oops_do(adjust_pointer_closure()); 2219 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2220 2221 // Now adjust pointers in remaining weak roots. (All of which should 2222 // have been cleared if they pointed to non-surviving objects.) 2223 // Global (weak) JNI handles 2224 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2225 2226 CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations); 2227 CodeCache::blobs_do(&adjust_from_blobs); 2228 StringTable::oops_do(adjust_pointer_closure()); 2229 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2230 // Roots were visited so references into the young gen in roots 2231 // may have been scanned. Process them also. 2232 // Should the reference processor have a span that excludes 2233 // young gen objects? 2234 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2235 } 2236 2237 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2238 uint parallel_gc_threads) 2239 { 2240 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer); 2241 2242 // Find the threads that are active 2243 unsigned int which = 0; 2244 2245 const uint task_count = MAX2(parallel_gc_threads, 1U); 2246 for (uint j = 0; j < task_count; j++) { 2247 q->enqueue(new DrainStacksCompactionTask(j)); 2248 ParCompactionManager::verify_region_list_empty(j); 2249 // Set the region stacks variables to "no" region stack values 2250 // so that they will be recognized and needing a region stack 2251 // in the stealing tasks if they do not get one by executing 2252 // a draining stack. 2253 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2254 cm->set_region_stack(NULL); 2255 cm->set_region_stack_index((uint)max_uintx); 2256 } 2257 ParCompactionManager::reset_recycled_stack_index(); 2258 2259 // Find all regions that are available (can be filled immediately) and 2260 // distribute them to the thread stacks. The iteration is done in reverse 2261 // order (high to low) so the regions will be removed in ascending order. 2262 2263 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2264 2265 size_t fillable_regions = 0; // A count for diagnostic purposes. 2266 // A region index which corresponds to the tasks created above. 2267 // "which" must be 0 <= which < task_count 2268 2269 which = 0; 2270 // id + 1 is used to test termination so unsigned can 2271 // be used with an old_space_id == 0. 2272 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2273 SpaceInfo* const space_info = _space_info + id; 2274 MutableSpace* const space = space_info->space(); 2275 HeapWord* const new_top = space_info->new_top(); 2276 2277 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2278 const size_t end_region = 2279 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2280 2281 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2282 if (sd.region(cur)->claim_unsafe()) { 2283 ParCompactionManager::region_list_push(which, cur); 2284 2285 if (TraceParallelOldGCCompactionPhase && Verbose) { 2286 const size_t count_mod_8 = fillable_regions & 7; 2287 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2288 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2289 if (count_mod_8 == 7) gclog_or_tty->cr(); 2290 } 2291 2292 NOT_PRODUCT(++fillable_regions;) 2293 2294 // Assign regions to tasks in round-robin fashion. 2295 if (++which == task_count) { 2296 assert(which <= parallel_gc_threads, 2297 "Inconsistent number of workers"); 2298 which = 0; 2299 } 2300 } 2301 } 2302 } 2303 2304 if (TraceParallelOldGCCompactionPhase) { 2305 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2306 gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions); 2307 } 2308 } 2309 2310 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2311 2312 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2313 uint parallel_gc_threads) { 2314 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer); 2315 2316 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2317 2318 // Iterate over all the spaces adding tasks for updating 2319 // regions in the dense prefix. Assume that 1 gc thread 2320 // will work on opening the gaps and the remaining gc threads 2321 // will work on the dense prefix. 2322 unsigned int space_id; 2323 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2324 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2325 const MutableSpace* const space = _space_info[space_id].space(); 2326 2327 if (dense_prefix_end == space->bottom()) { 2328 // There is no dense prefix for this space. 2329 continue; 2330 } 2331 2332 // The dense prefix is before this region. 2333 size_t region_index_end_dense_prefix = 2334 sd.addr_to_region_idx(dense_prefix_end); 2335 RegionData* const dense_prefix_cp = 2336 sd.region(region_index_end_dense_prefix); 2337 assert(dense_prefix_end == space->end() || 2338 dense_prefix_cp->available() || 2339 dense_prefix_cp->claimed(), 2340 "The region after the dense prefix should always be ready to fill"); 2341 2342 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2343 2344 // Is there dense prefix work? 2345 size_t total_dense_prefix_regions = 2346 region_index_end_dense_prefix - region_index_start; 2347 // How many regions of the dense prefix should be given to 2348 // each thread? 2349 if (total_dense_prefix_regions > 0) { 2350 uint tasks_for_dense_prefix = 1; 2351 if (total_dense_prefix_regions <= 2352 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2353 // Don't over partition. This assumes that 2354 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2355 // so there are not many regions to process. 2356 tasks_for_dense_prefix = parallel_gc_threads; 2357 } else { 2358 // Over partition 2359 tasks_for_dense_prefix = parallel_gc_threads * 2360 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2361 } 2362 size_t regions_per_thread = total_dense_prefix_regions / 2363 tasks_for_dense_prefix; 2364 // Give each thread at least 1 region. 2365 if (regions_per_thread == 0) { 2366 regions_per_thread = 1; 2367 } 2368 2369 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2370 if (region_index_start >= region_index_end_dense_prefix) { 2371 break; 2372 } 2373 // region_index_end is not processed 2374 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2375 region_index_end_dense_prefix); 2376 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2377 region_index_start, 2378 region_index_end)); 2379 region_index_start = region_index_end; 2380 } 2381 } 2382 // This gets any part of the dense prefix that did not 2383 // fit evenly. 2384 if (region_index_start < region_index_end_dense_prefix) { 2385 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2386 region_index_start, 2387 region_index_end_dense_prefix)); 2388 } 2389 } 2390 } 2391 2392 void PSParallelCompact::enqueue_region_stealing_tasks( 2393 GCTaskQueue* q, 2394 ParallelTaskTerminator* terminator_ptr, 2395 uint parallel_gc_threads) { 2396 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer); 2397 2398 // Once a thread has drained it's stack, it should try to steal regions from 2399 // other threads. 2400 if (parallel_gc_threads > 1) { 2401 for (uint j = 0; j < parallel_gc_threads; j++) { 2402 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2403 } 2404 } 2405 } 2406 2407 #ifdef ASSERT 2408 // Write a histogram of the number of times the block table was filled for a 2409 // region. 2410 void PSParallelCompact::write_block_fill_histogram(outputStream* const out) 2411 { 2412 if (!TraceParallelOldGCCompactionPhase) return; 2413 2414 typedef ParallelCompactData::RegionData rd_t; 2415 ParallelCompactData& sd = summary_data(); 2416 2417 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2418 MutableSpace* const spc = _space_info[id].space(); 2419 if (spc->bottom() != spc->top()) { 2420 const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom()); 2421 HeapWord* const top_aligned_up = sd.region_align_up(spc->top()); 2422 const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up); 2423 2424 size_t histo[5] = { 0, 0, 0, 0, 0 }; 2425 const size_t histo_len = sizeof(histo) / sizeof(size_t); 2426 const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t)); 2427 2428 for (const rd_t* cur = beg; cur < end; ++cur) { 2429 ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)]; 2430 } 2431 out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt); 2432 for (size_t i = 0; i < histo_len; ++i) { 2433 out->print(" " SIZE_FORMAT_W(5) " %5.1f%%", 2434 histo[i], 100.0 * histo[i] / region_cnt); 2435 } 2436 out->cr(); 2437 } 2438 } 2439 } 2440 #endif // #ifdef ASSERT 2441 2442 void PSParallelCompact::compact() { 2443 // trace("5"); 2444 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer); 2445 2446 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap(); 2447 PSOldGen* old_gen = heap->old_gen(); 2448 old_gen->start_array()->reset(); 2449 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2450 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2451 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2452 ParallelTaskTerminator terminator(active_gc_threads, qset); 2453 2454 GCTaskQueue* q = GCTaskQueue::create(); 2455 enqueue_region_draining_tasks(q, active_gc_threads); 2456 enqueue_dense_prefix_tasks(q, active_gc_threads); 2457 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2458 2459 { 2460 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer); 2461 2462 gc_task_manager()->execute_and_wait(q); 2463 2464 #ifdef ASSERT 2465 // Verify that all regions have been processed before the deferred updates. 2466 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2467 verify_complete(SpaceId(id)); 2468 } 2469 #endif 2470 } 2471 2472 { 2473 // Update the deferred objects, if any. Any compaction manager can be used. 2474 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer); 2475 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2476 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2477 update_deferred_objects(cm, SpaceId(id)); 2478 } 2479 } 2480 2481 DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty)); 2482 } 2483 2484 #ifdef ASSERT 2485 void PSParallelCompact::verify_complete(SpaceId space_id) { 2486 // All Regions between space bottom() to new_top() should be marked as filled 2487 // and all Regions between new_top() and top() should be available (i.e., 2488 // should have been emptied). 2489 ParallelCompactData& sd = summary_data(); 2490 SpaceInfo si = _space_info[space_id]; 2491 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2492 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2493 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2494 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2495 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2496 2497 bool issued_a_warning = false; 2498 2499 size_t cur_region; 2500 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2501 const RegionData* const c = sd.region(cur_region); 2502 if (!c->completed()) { 2503 warning("region " SIZE_FORMAT " not filled: " 2504 "destination_count=%u", 2505 cur_region, c->destination_count()); 2506 issued_a_warning = true; 2507 } 2508 } 2509 2510 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2511 const RegionData* const c = sd.region(cur_region); 2512 if (!c->available()) { 2513 warning("region " SIZE_FORMAT " not empty: " 2514 "destination_count=%u", 2515 cur_region, c->destination_count()); 2516 issued_a_warning = true; 2517 } 2518 } 2519 2520 if (issued_a_warning) { 2521 print_region_ranges(); 2522 } 2523 } 2524 #endif // #ifdef ASSERT 2525 2526 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) { 2527 _start_array->allocate_block(addr); 2528 compaction_manager()->update_contents(oop(addr)); 2529 } 2530 2531 // Update interior oops in the ranges of regions [beg_region, end_region). 2532 void 2533 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2534 SpaceId space_id, 2535 size_t beg_region, 2536 size_t end_region) { 2537 ParallelCompactData& sd = summary_data(); 2538 ParMarkBitMap* const mbm = mark_bitmap(); 2539 2540 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2541 HeapWord* const end_addr = sd.region_to_addr(end_region); 2542 assert(beg_region <= end_region, "bad region range"); 2543 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2544 2545 #ifdef ASSERT 2546 // Claim the regions to avoid triggering an assert when they are marked as 2547 // filled. 2548 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2549 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2550 } 2551 #endif // #ifdef ASSERT 2552 2553 if (beg_addr != space(space_id)->bottom()) { 2554 // Find the first live object or block of dead space that *starts* in this 2555 // range of regions. If a partial object crosses onto the region, skip it; 2556 // it will be marked for 'deferred update' when the object head is 2557 // processed. If dead space crosses onto the region, it is also skipped; it 2558 // will be filled when the prior region is processed. If neither of those 2559 // apply, the first word in the region is the start of a live object or dead 2560 // space. 2561 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2562 const RegionData* const cp = sd.region(beg_region); 2563 if (cp->partial_obj_size() != 0) { 2564 beg_addr = sd.partial_obj_end(beg_region); 2565 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2566 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2567 } 2568 } 2569 2570 if (beg_addr < end_addr) { 2571 // A live object or block of dead space starts in this range of Regions. 2572 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2573 2574 // Create closures and iterate. 2575 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2576 FillClosure fill_closure(cm, space_id); 2577 ParMarkBitMap::IterationStatus status; 2578 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2579 dense_prefix_end); 2580 if (status == ParMarkBitMap::incomplete) { 2581 update_closure.do_addr(update_closure.source()); 2582 } 2583 } 2584 2585 // Mark the regions as filled. 2586 RegionData* const beg_cp = sd.region(beg_region); 2587 RegionData* const end_cp = sd.region(end_region); 2588 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2589 cp->set_completed(); 2590 } 2591 } 2592 2593 // Return the SpaceId for the space containing addr. If addr is not in the 2594 // heap, last_space_id is returned. In debug mode it expects the address to be 2595 // in the heap and asserts such. 2596 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2597 assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap"); 2598 2599 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2600 if (_space_info[id].space()->contains(addr)) { 2601 return SpaceId(id); 2602 } 2603 } 2604 2605 assert(false, "no space contains the addr"); 2606 return last_space_id; 2607 } 2608 2609 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2610 SpaceId id) { 2611 assert(id < last_space_id, "bad space id"); 2612 2613 ParallelCompactData& sd = summary_data(); 2614 const SpaceInfo* const space_info = _space_info + id; 2615 ObjectStartArray* const start_array = space_info->start_array(); 2616 2617 const MutableSpace* const space = space_info->space(); 2618 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2619 HeapWord* const beg_addr = space_info->dense_prefix(); 2620 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2621 2622 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2623 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2624 const RegionData* cur_region; 2625 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2626 HeapWord* const addr = cur_region->deferred_obj_addr(); 2627 if (addr != NULL) { 2628 if (start_array != NULL) { 2629 start_array->allocate_block(addr); 2630 } 2631 cm->update_contents(oop(addr)); 2632 assert(oop(addr)->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))); 2633 } 2634 } 2635 } 2636 2637 // Skip over count live words starting from beg, and return the address of the 2638 // next live word. Unless marked, the word corresponding to beg is assumed to 2639 // be dead. Callers must either ensure beg does not correspond to the middle of 2640 // an object, or account for those live words in some other way. Callers must 2641 // also ensure that there are enough live words in the range [beg, end) to skip. 2642 HeapWord* 2643 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2644 { 2645 assert(count > 0, "sanity"); 2646 2647 ParMarkBitMap* m = mark_bitmap(); 2648 idx_t bits_to_skip = m->words_to_bits(count); 2649 idx_t cur_beg = m->addr_to_bit(beg); 2650 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2651 2652 do { 2653 cur_beg = m->find_obj_beg(cur_beg, search_end); 2654 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2655 const size_t obj_bits = cur_end - cur_beg + 1; 2656 if (obj_bits > bits_to_skip) { 2657 return m->bit_to_addr(cur_beg + bits_to_skip); 2658 } 2659 bits_to_skip -= obj_bits; 2660 cur_beg = cur_end + 1; 2661 } while (bits_to_skip > 0); 2662 2663 // Skipping the desired number of words landed just past the end of an object. 2664 // Find the start of the next object. 2665 cur_beg = m->find_obj_beg(cur_beg, search_end); 2666 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2667 return m->bit_to_addr(cur_beg); 2668 } 2669 2670 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2671 SpaceId src_space_id, 2672 size_t src_region_idx) 2673 { 2674 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2675 2676 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2677 if (split_info.dest_region_addr() == dest_addr) { 2678 // The partial object ending at the split point contains the first word to 2679 // be copied to dest_addr. 2680 return split_info.first_src_addr(); 2681 } 2682 2683 const ParallelCompactData& sd = summary_data(); 2684 ParMarkBitMap* const bitmap = mark_bitmap(); 2685 const size_t RegionSize = ParallelCompactData::RegionSize; 2686 2687 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2688 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2689 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2690 HeapWord* const src_region_destination = src_region_ptr->destination(); 2691 2692 assert(dest_addr >= src_region_destination, "wrong src region"); 2693 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2694 2695 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2696 HeapWord* const src_region_end = src_region_beg + RegionSize; 2697 2698 HeapWord* addr = src_region_beg; 2699 if (dest_addr == src_region_destination) { 2700 // Return the first live word in the source region. 2701 if (partial_obj_size == 0) { 2702 addr = bitmap->find_obj_beg(addr, src_region_end); 2703 assert(addr < src_region_end, "no objects start in src region"); 2704 } 2705 return addr; 2706 } 2707 2708 // Must skip some live data. 2709 size_t words_to_skip = dest_addr - src_region_destination; 2710 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2711 2712 if (partial_obj_size >= words_to_skip) { 2713 // All the live words to skip are part of the partial object. 2714 addr += words_to_skip; 2715 if (partial_obj_size == words_to_skip) { 2716 // Find the first live word past the partial object. 2717 addr = bitmap->find_obj_beg(addr, src_region_end); 2718 assert(addr < src_region_end, "wrong src region"); 2719 } 2720 return addr; 2721 } 2722 2723 // Skip over the partial object (if any). 2724 if (partial_obj_size != 0) { 2725 words_to_skip -= partial_obj_size; 2726 addr += partial_obj_size; 2727 } 2728 2729 // Skip over live words due to objects that start in the region. 2730 addr = skip_live_words(addr, src_region_end, words_to_skip); 2731 assert(addr < src_region_end, "wrong src region"); 2732 return addr; 2733 } 2734 2735 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2736 SpaceId src_space_id, 2737 size_t beg_region, 2738 HeapWord* end_addr) 2739 { 2740 ParallelCompactData& sd = summary_data(); 2741 2742 #ifdef ASSERT 2743 MutableSpace* const src_space = _space_info[src_space_id].space(); 2744 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2745 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2746 "src_space_id does not match beg_addr"); 2747 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2748 "src_space_id does not match end_addr"); 2749 #endif // #ifdef ASSERT 2750 2751 RegionData* const beg = sd.region(beg_region); 2752 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2753 2754 // Regions up to new_top() are enqueued if they become available. 2755 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2756 RegionData* const enqueue_end = 2757 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2758 2759 for (RegionData* cur = beg; cur < end; ++cur) { 2760 assert(cur->data_size() > 0, "region must have live data"); 2761 cur->decrement_destination_count(); 2762 if (cur < enqueue_end && cur->available() && cur->claim()) { 2763 cm->push_region(sd.region(cur)); 2764 } 2765 } 2766 } 2767 2768 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2769 SpaceId& src_space_id, 2770 HeapWord*& src_space_top, 2771 HeapWord* end_addr) 2772 { 2773 typedef ParallelCompactData::RegionData RegionData; 2774 2775 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2776 const size_t region_size = ParallelCompactData::RegionSize; 2777 2778 size_t src_region_idx = 0; 2779 2780 // Skip empty regions (if any) up to the top of the space. 2781 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2782 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2783 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2784 const RegionData* const top_region_ptr = 2785 sd.addr_to_region_ptr(top_aligned_up); 2786 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2787 ++src_region_ptr; 2788 } 2789 2790 if (src_region_ptr < top_region_ptr) { 2791 // The next source region is in the current space. Update src_region_idx 2792 // and the source address to match src_region_ptr. 2793 src_region_idx = sd.region(src_region_ptr); 2794 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2795 if (src_region_addr > closure.source()) { 2796 closure.set_source(src_region_addr); 2797 } 2798 return src_region_idx; 2799 } 2800 2801 // Switch to a new source space and find the first non-empty region. 2802 unsigned int space_id = src_space_id + 1; 2803 assert(space_id < last_space_id, "not enough spaces"); 2804 2805 HeapWord* const destination = closure.destination(); 2806 2807 do { 2808 MutableSpace* space = _space_info[space_id].space(); 2809 HeapWord* const bottom = space->bottom(); 2810 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2811 2812 // Iterate over the spaces that do not compact into themselves. 2813 if (bottom_cp->destination() != bottom) { 2814 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2815 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2816 2817 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2818 if (src_cp->live_obj_size() > 0) { 2819 // Found it. 2820 assert(src_cp->destination() == destination, 2821 "first live obj in the space must match the destination"); 2822 assert(src_cp->partial_obj_size() == 0, 2823 "a space cannot begin with a partial obj"); 2824 2825 src_space_id = SpaceId(space_id); 2826 src_space_top = space->top(); 2827 const size_t src_region_idx = sd.region(src_cp); 2828 closure.set_source(sd.region_to_addr(src_region_idx)); 2829 return src_region_idx; 2830 } else { 2831 assert(src_cp->data_size() == 0, "sanity"); 2832 } 2833 } 2834 } 2835 } while (++space_id < last_space_id); 2836 2837 assert(false, "no source region was found"); 2838 return 0; 2839 } 2840 2841 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 2842 { 2843 typedef ParMarkBitMap::IterationStatus IterationStatus; 2844 const size_t RegionSize = ParallelCompactData::RegionSize; 2845 ParMarkBitMap* const bitmap = mark_bitmap(); 2846 ParallelCompactData& sd = summary_data(); 2847 RegionData* const region_ptr = sd.region(region_idx); 2848 2849 // Get the items needed to construct the closure. 2850 HeapWord* dest_addr = sd.region_to_addr(region_idx); 2851 SpaceId dest_space_id = space_id(dest_addr); 2852 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 2853 HeapWord* new_top = _space_info[dest_space_id].new_top(); 2854 assert(dest_addr < new_top, "sanity"); 2855 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 2856 2857 // Get the source region and related info. 2858 size_t src_region_idx = region_ptr->source_region(); 2859 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 2860 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 2861 2862 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 2863 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 2864 2865 // Adjust src_region_idx to prepare for decrementing destination counts (the 2866 // destination count is not decremented when a region is copied to itself). 2867 if (src_region_idx == region_idx) { 2868 src_region_idx += 1; 2869 } 2870 2871 if (bitmap->is_unmarked(closure.source())) { 2872 // The first source word is in the middle of an object; copy the remainder 2873 // of the object or as much as will fit. The fact that pointer updates were 2874 // deferred will be noted when the object header is processed. 2875 HeapWord* const old_src_addr = closure.source(); 2876 closure.copy_partial_obj(); 2877 if (closure.is_full()) { 2878 decrement_destination_counts(cm, src_space_id, src_region_idx, 2879 closure.source()); 2880 region_ptr->set_deferred_obj_addr(NULL); 2881 region_ptr->set_completed(); 2882 return; 2883 } 2884 2885 HeapWord* const end_addr = sd.region_align_down(closure.source()); 2886 if (sd.region_align_down(old_src_addr) != end_addr) { 2887 // The partial object was copied from more than one source region. 2888 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2889 2890 // Move to the next source region, possibly switching spaces as well. All 2891 // args except end_addr may be modified. 2892 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2893 end_addr); 2894 } 2895 } 2896 2897 do { 2898 HeapWord* const cur_addr = closure.source(); 2899 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 2900 src_space_top); 2901 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 2902 2903 if (status == ParMarkBitMap::incomplete) { 2904 // The last obj that starts in the source region does not end in the 2905 // region. 2906 assert(closure.source() < end_addr, "sanity"); 2907 HeapWord* const obj_beg = closure.source(); 2908 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 2909 src_space_top); 2910 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 2911 if (obj_end < range_end) { 2912 // The end was found; the entire object will fit. 2913 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 2914 assert(status != ParMarkBitMap::would_overflow, "sanity"); 2915 } else { 2916 // The end was not found; the object will not fit. 2917 assert(range_end < src_space_top, "obj cannot cross space boundary"); 2918 status = ParMarkBitMap::would_overflow; 2919 } 2920 } 2921 2922 if (status == ParMarkBitMap::would_overflow) { 2923 // The last object did not fit. Note that interior oop updates were 2924 // deferred, then copy enough of the object to fill the region. 2925 region_ptr->set_deferred_obj_addr(closure.destination()); 2926 status = closure.copy_until_full(); // copies from closure.source() 2927 2928 decrement_destination_counts(cm, src_space_id, src_region_idx, 2929 closure.source()); 2930 region_ptr->set_completed(); 2931 return; 2932 } 2933 2934 if (status == ParMarkBitMap::full) { 2935 decrement_destination_counts(cm, src_space_id, src_region_idx, 2936 closure.source()); 2937 region_ptr->set_deferred_obj_addr(NULL); 2938 region_ptr->set_completed(); 2939 return; 2940 } 2941 2942 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 2943 2944 // Move to the next source region, possibly switching spaces as well. All 2945 // args except end_addr may be modified. 2946 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 2947 end_addr); 2948 } while (true); 2949 } 2950 2951 void PSParallelCompact::fill_blocks(size_t region_idx) 2952 { 2953 // Fill in the block table elements for the specified region. Each block 2954 // table element holds the number of live words in the region that are to the 2955 // left of the first object that starts in the block. Thus only blocks in 2956 // which an object starts need to be filled. 2957 // 2958 // The algorithm scans the section of the bitmap that corresponds to the 2959 // region, keeping a running total of the live words. When an object start is 2960 // found, if it's the first to start in the block that contains it, the 2961 // current total is written to the block table element. 2962 const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize; 2963 const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize; 2964 const size_t RegionSize = ParallelCompactData::RegionSize; 2965 2966 ParallelCompactData& sd = summary_data(); 2967 const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size(); 2968 if (partial_obj_size >= RegionSize) { 2969 return; // No objects start in this region. 2970 } 2971 2972 // Ensure the first loop iteration decides that the block has changed. 2973 size_t cur_block = sd.block_count(); 2974 2975 const ParMarkBitMap* const bitmap = mark_bitmap(); 2976 2977 const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment; 2978 assert((size_t)1 << Log2BitsPerBlock == 2979 bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity"); 2980 2981 size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize); 2982 const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize); 2983 size_t live_bits = bitmap->words_to_bits(partial_obj_size); 2984 beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end); 2985 while (beg_bit < range_end) { 2986 const size_t new_block = beg_bit >> Log2BitsPerBlock; 2987 if (new_block != cur_block) { 2988 cur_block = new_block; 2989 sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits)); 2990 } 2991 2992 const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end); 2993 if (end_bit < range_end - 1) { 2994 live_bits += end_bit - beg_bit + 1; 2995 beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end); 2996 } else { 2997 return; 2998 } 2999 } 3000 } 3001 3002 void 3003 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3004 const MutableSpace* sp = space(space_id); 3005 if (sp->is_empty()) { 3006 return; 3007 } 3008 3009 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3010 ParMarkBitMap* const bitmap = mark_bitmap(); 3011 HeapWord* const dp_addr = dense_prefix(space_id); 3012 HeapWord* beg_addr = sp->bottom(); 3013 HeapWord* end_addr = sp->top(); 3014 3015 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3016 3017 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3018 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3019 if (beg_region < dp_region) { 3020 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3021 } 3022 3023 // The destination of the first live object that starts in the region is one 3024 // past the end of the partial object entering the region (if any). 3025 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3026 HeapWord* const new_top = _space_info[space_id].new_top(); 3027 assert(new_top >= dest_addr, "bad new_top value"); 3028 const size_t words = pointer_delta(new_top, dest_addr); 3029 3030 if (words > 0) { 3031 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3032 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3033 3034 ParMarkBitMap::IterationStatus status; 3035 status = bitmap->iterate(&closure, dest_addr, end_addr); 3036 assert(status == ParMarkBitMap::full, "iteration not complete"); 3037 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3038 "live objects skipped because closure is full"); 3039 } 3040 } 3041 3042 jlong PSParallelCompact::millis_since_last_gc() { 3043 // We need a monotonically non-decreasing time in ms but 3044 // os::javaTimeMillis() does not guarantee monotonicity. 3045 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3046 jlong ret_val = now - _time_of_last_gc; 3047 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3048 if (ret_val < 0) { 3049 NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);) 3050 return 0; 3051 } 3052 return ret_val; 3053 } 3054 3055 void PSParallelCompact::reset_millis_since_last_gc() { 3056 // We need a monotonically non-decreasing time in ms but 3057 // os::javaTimeMillis() does not guarantee monotonicity. 3058 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3059 } 3060 3061 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3062 { 3063 if (source() != destination()) { 3064 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3065 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3066 } 3067 update_state(words_remaining()); 3068 assert(is_full(), "sanity"); 3069 return ParMarkBitMap::full; 3070 } 3071 3072 void MoveAndUpdateClosure::copy_partial_obj() 3073 { 3074 size_t words = words_remaining(); 3075 3076 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3077 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3078 if (end_addr < range_end) { 3079 words = bitmap()->obj_size(source(), end_addr); 3080 } 3081 3082 // This test is necessary; if omitted, the pointer updates to a partial object 3083 // that crosses the dense prefix boundary could be overwritten. 3084 if (source() != destination()) { 3085 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3086 Copy::aligned_conjoint_words(source(), destination(), words); 3087 } 3088 update_state(words); 3089 } 3090 3091 void InstanceKlass::oop_pc_update_pointers(oop obj) { 3092 oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3093 } 3094 3095 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) { 3096 InstanceKlass::oop_pc_update_pointers(obj); 3097 3098 oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure()); 3099 } 3100 3101 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) { 3102 InstanceKlass::oop_pc_update_pointers(obj); 3103 } 3104 3105 #ifdef ASSERT 3106 template <class T> static void trace_reference_gc(const char *s, oop obj, 3107 T* referent_addr, 3108 T* next_addr, 3109 T* discovered_addr) { 3110 if(TraceReferenceGC && PrintGCDetails) { 3111 gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj)); 3112 gclog_or_tty->print_cr(" referent_addr/* " PTR_FORMAT " / " 3113 PTR_FORMAT, p2i(referent_addr), 3114 referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL); 3115 gclog_or_tty->print_cr(" next_addr/* " PTR_FORMAT " / " 3116 PTR_FORMAT, p2i(next_addr), 3117 next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL); 3118 gclog_or_tty->print_cr(" discovered_addr/* " PTR_FORMAT " / " 3119 PTR_FORMAT, p2i(discovered_addr), 3120 discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL); 3121 } 3122 } 3123 #endif 3124 3125 template <class T> 3126 static void oop_pc_update_pointers_specialized(oop obj) { 3127 T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj); 3128 PSParallelCompact::adjust_pointer(referent_addr); 3129 T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj); 3130 PSParallelCompact::adjust_pointer(next_addr); 3131 T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj); 3132 PSParallelCompact::adjust_pointer(discovered_addr); 3133 debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj, 3134 referent_addr, next_addr, discovered_addr);) 3135 } 3136 3137 void InstanceRefKlass::oop_pc_update_pointers(oop obj) { 3138 InstanceKlass::oop_pc_update_pointers(obj); 3139 3140 if (UseCompressedOops) { 3141 oop_pc_update_pointers_specialized<narrowOop>(obj); 3142 } else { 3143 oop_pc_update_pointers_specialized<oop>(obj); 3144 } 3145 } 3146 3147 void ObjArrayKlass::oop_pc_update_pointers(oop obj) { 3148 assert(obj->is_objArray(), "obj must be obj array"); 3149 oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure()); 3150 } 3151 3152 void TypeArrayKlass::oop_pc_update_pointers(oop obj) { 3153 assert(obj->is_typeArray(),"must be a type array"); 3154 } 3155 3156 ParMarkBitMapClosure::IterationStatus 3157 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3158 assert(destination() != NULL, "sanity"); 3159 assert(bitmap()->obj_size(addr) == words, "bad size"); 3160 3161 _source = addr; 3162 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3163 destination(), "wrong destination"); 3164 3165 if (words > words_remaining()) { 3166 return ParMarkBitMap::would_overflow; 3167 } 3168 3169 // The start_array must be updated even if the object is not moving. 3170 if (_start_array != NULL) { 3171 _start_array->allocate_block(destination()); 3172 } 3173 3174 if (destination() != source()) { 3175 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3176 Copy::aligned_conjoint_words(source(), destination(), words); 3177 } 3178 3179 oop moved_oop = (oop) destination(); 3180 compaction_manager()->update_contents(moved_oop); 3181 assert(moved_oop->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)); 3182 3183 update_state(words); 3184 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3185 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3186 } 3187 3188 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3189 ParCompactionManager* cm, 3190 PSParallelCompact::SpaceId space_id) : 3191 ParMarkBitMapClosure(mbm, cm), 3192 _space_id(space_id), 3193 _start_array(PSParallelCompact::start_array(space_id)) 3194 { 3195 } 3196 3197 // Updates the references in the object to their new values. 3198 ParMarkBitMapClosure::IterationStatus 3199 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3200 do_addr(addr); 3201 return ParMarkBitMap::incomplete; 3202 }