1 /* 2 * Copyright (c) 2005, 2013, 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/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp" 30 #include "gc_implementation/parallelScavenge/generationSizer.hpp" 31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp" 32 #include "gc_implementation/parallelScavenge/pcTasks.hpp" 33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp" 34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp" 35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp" 36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp" 37 #include "gc_implementation/parallelScavenge/psOldGen.hpp" 38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp" 39 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 40 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 41 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 42 #include "gc_implementation/shared/gcHeapSummary.hpp" 43 #include "gc_implementation/shared/gcTimer.hpp" 44 #include "gc_implementation/shared/gcTrace.hpp" 45 #include "gc_implementation/shared/gcTraceTime.hpp" 46 #include "gc_implementation/shared/isGCActiveMark.hpp" 47 #include "gc_interface/gcCause.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/referencePolicy.hpp" 50 #include "memory/referenceProcessor.hpp" 51 #include "oops/methodData.hpp" 52 #include "oops/oop.inline.hpp" 53 #include "oops/oop.pcgc.inline.hpp" 54 #include "runtime/fprofiler.hpp" 55 #include "runtime/safepoint.hpp" 56 #include "runtime/vmThread.hpp" 57 #include "services/management.hpp" 58 #include "services/memoryService.hpp" 59 #include "services/memTracker.hpp" 60 #include "utilities/events.hpp" 61 #include "utilities/stack.inline.hpp" 62 63 #include <math.h> 64 65 // All sizes are in HeapWords. 66 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words 67 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 68 const size_t ParallelCompactData::RegionSizeBytes = 69 RegionSize << LogHeapWordSize; 70 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 71 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 72 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 73 74 const ParallelCompactData::RegionData::region_sz_t 75 ParallelCompactData::RegionData::dc_shift = 27; 76 77 const ParallelCompactData::RegionData::region_sz_t 78 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 79 80 const ParallelCompactData::RegionData::region_sz_t 81 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 82 83 const ParallelCompactData::RegionData::region_sz_t 84 ParallelCompactData::RegionData::los_mask = ~dc_mask; 85 86 const ParallelCompactData::RegionData::region_sz_t 87 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 88 89 const ParallelCompactData::RegionData::region_sz_t 90 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 91 92 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 93 bool PSParallelCompact::_print_phases = false; 94 95 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 96 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL; 97 98 double PSParallelCompact::_dwl_mean; 99 double PSParallelCompact::_dwl_std_dev; 100 double PSParallelCompact::_dwl_first_term; 101 double PSParallelCompact::_dwl_adjustment; 102 #ifdef ASSERT 103 bool PSParallelCompact::_dwl_initialized = false; 104 #endif // #ifdef ASSERT 105 106 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 107 HeapWord* destination) 108 { 109 assert(src_region_idx != 0, "invalid src_region_idx"); 110 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 111 assert(destination != NULL, "invalid destination argument"); 112 113 _src_region_idx = src_region_idx; 114 _partial_obj_size = partial_obj_size; 115 _destination = destination; 116 117 // These fields may not be updated below, so make sure they're clear. 118 assert(_dest_region_addr == NULL, "should have been cleared"); 119 assert(_first_src_addr == NULL, "should have been cleared"); 120 121 // Determine the number of destination regions for the partial object. 122 HeapWord* const last_word = destination + partial_obj_size - 1; 123 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 124 HeapWord* const beg_region_addr = sd.region_align_down(destination); 125 HeapWord* const end_region_addr = sd.region_align_down(last_word); 126 127 if (beg_region_addr == end_region_addr) { 128 // One destination region. 129 _destination_count = 1; 130 if (end_region_addr == destination) { 131 // The destination falls on a region boundary, thus the first word of the 132 // partial object will be the first word copied to the destination region. 133 _dest_region_addr = end_region_addr; 134 _first_src_addr = sd.region_to_addr(src_region_idx); 135 } 136 } else { 137 // Two destination regions. When copied, the partial object will cross a 138 // destination region boundary, so a word somewhere within the partial 139 // object will be the first word copied to the second destination region. 140 _destination_count = 2; 141 _dest_region_addr = end_region_addr; 142 const size_t ofs = pointer_delta(end_region_addr, destination); 143 assert(ofs < _partial_obj_size, "sanity"); 144 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 145 } 146 } 147 148 void SplitInfo::clear() 149 { 150 _src_region_idx = 0; 151 _partial_obj_size = 0; 152 _destination = NULL; 153 _destination_count = 0; 154 _dest_region_addr = NULL; 155 _first_src_addr = NULL; 156 assert(!is_valid(), "sanity"); 157 } 158 159 #ifdef ASSERT 160 void SplitInfo::verify_clear() 161 { 162 assert(_src_region_idx == 0, "not clear"); 163 assert(_partial_obj_size == 0, "not clear"); 164 assert(_destination == NULL, "not clear"); 165 assert(_destination_count == 0, "not clear"); 166 assert(_dest_region_addr == NULL, "not clear"); 167 assert(_first_src_addr == NULL, "not clear"); 168 } 169 #endif // #ifdef ASSERT 170 171 172 void PSParallelCompact::print_on_error(outputStream* st) { 173 _mark_bitmap.print_on_error(st); 174 } 175 176 #ifndef PRODUCT 177 const char* PSParallelCompact::space_names[] = { 178 "old ", "eden", "from", "to " 179 }; 180 181 void PSParallelCompact::print_region_ranges() 182 { 183 tty->print_cr("space bottom top end new_top"); 184 tty->print_cr("------ ---------- ---------- ---------- ----------"); 185 186 for (unsigned int id = 0; id < last_space_id; ++id) { 187 const MutableSpace* space = _space_info[id].space(); 188 tty->print_cr("%u %s " 189 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 190 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 191 id, space_names[id], 192 summary_data().addr_to_region_idx(space->bottom()), 193 summary_data().addr_to_region_idx(space->top()), 194 summary_data().addr_to_region_idx(space->end()), 195 summary_data().addr_to_region_idx(_space_info[id].new_top())); 196 } 197 } 198 199 void 200 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 201 { 202 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 203 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 204 205 ParallelCompactData& sd = PSParallelCompact::summary_data(); 206 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 207 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 208 REGION_IDX_FORMAT " " PTR_FORMAT " " 209 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 210 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 211 i, c->data_location(), dci, c->destination(), 212 c->partial_obj_size(), c->live_obj_size(), 213 c->data_size(), c->source_region(), c->destination_count()); 214 215 #undef REGION_IDX_FORMAT 216 #undef REGION_DATA_FORMAT 217 } 218 219 void 220 print_generic_summary_data(ParallelCompactData& summary_data, 221 HeapWord* const beg_addr, 222 HeapWord* const end_addr) 223 { 224 size_t total_words = 0; 225 size_t i = summary_data.addr_to_region_idx(beg_addr); 226 const size_t last = summary_data.addr_to_region_idx(end_addr); 227 HeapWord* pdest = 0; 228 229 while (i <= last) { 230 ParallelCompactData::RegionData* c = summary_data.region(i); 231 if (c->data_size() != 0 || c->destination() != pdest) { 232 print_generic_summary_region(i, c); 233 total_words += c->data_size(); 234 pdest = c->destination(); 235 } 236 ++i; 237 } 238 239 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 240 } 241 242 void 243 print_generic_summary_data(ParallelCompactData& summary_data, 244 SpaceInfo* space_info) 245 { 246 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 247 const MutableSpace* space = space_info[id].space(); 248 print_generic_summary_data(summary_data, space->bottom(), 249 MAX2(space->top(), space_info[id].new_top())); 250 } 251 } 252 253 void 254 print_initial_summary_region(size_t i, 255 const ParallelCompactData::RegionData* c, 256 bool newline = true) 257 { 258 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 259 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 260 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 261 i, c->destination(), 262 c->partial_obj_size(), c->live_obj_size(), 263 c->data_size(), c->source_region(), c->destination_count()); 264 if (newline) tty->cr(); 265 } 266 267 void 268 print_initial_summary_data(ParallelCompactData& summary_data, 269 const MutableSpace* space) { 270 if (space->top() == space->bottom()) { 271 return; 272 } 273 274 const size_t region_size = ParallelCompactData::RegionSize; 275 typedef ParallelCompactData::RegionData RegionData; 276 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 277 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 278 const RegionData* c = summary_data.region(end_region - 1); 279 HeapWord* end_addr = c->destination() + c->data_size(); 280 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 281 282 // Print (and count) the full regions at the beginning of the space. 283 size_t full_region_count = 0; 284 size_t i = summary_data.addr_to_region_idx(space->bottom()); 285 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 286 print_initial_summary_region(i, summary_data.region(i)); 287 ++full_region_count; 288 ++i; 289 } 290 291 size_t live_to_right = live_in_space - full_region_count * region_size; 292 293 double max_reclaimed_ratio = 0.0; 294 size_t max_reclaimed_ratio_region = 0; 295 size_t max_dead_to_right = 0; 296 size_t max_live_to_right = 0; 297 298 // Print the 'reclaimed ratio' for regions while there is something live in 299 // the region or to the right of it. The remaining regions are empty (and 300 // uninteresting), and computing the ratio will result in division by 0. 301 while (i < end_region && live_to_right > 0) { 302 c = summary_data.region(i); 303 HeapWord* const region_addr = summary_data.region_to_addr(i); 304 const size_t used_to_right = pointer_delta(space->top(), region_addr); 305 const size_t dead_to_right = used_to_right - live_to_right; 306 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 307 308 if (reclaimed_ratio > max_reclaimed_ratio) { 309 max_reclaimed_ratio = reclaimed_ratio; 310 max_reclaimed_ratio_region = i; 311 max_dead_to_right = dead_to_right; 312 max_live_to_right = live_to_right; 313 } 314 315 print_initial_summary_region(i, c, false); 316 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 317 reclaimed_ratio, dead_to_right, live_to_right); 318 319 live_to_right -= c->data_size(); 320 ++i; 321 } 322 323 // Any remaining regions are empty. Print one more if there is one. 324 if (i < end_region) { 325 print_initial_summary_region(i, summary_data.region(i)); 326 } 327 328 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 329 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 330 max_reclaimed_ratio_region, max_dead_to_right, 331 max_live_to_right, max_reclaimed_ratio); 332 } 333 334 void 335 print_initial_summary_data(ParallelCompactData& summary_data, 336 SpaceInfo* space_info) { 337 unsigned int id = PSParallelCompact::old_space_id; 338 const MutableSpace* space; 339 do { 340 space = space_info[id].space(); 341 print_initial_summary_data(summary_data, space); 342 } while (++id < PSParallelCompact::eden_space_id); 343 344 do { 345 space = space_info[id].space(); 346 print_generic_summary_data(summary_data, space->bottom(), space->top()); 347 } while (++id < PSParallelCompact::last_space_id); 348 } 349 #endif // #ifndef PRODUCT 350 351 #ifdef ASSERT 352 size_t add_obj_count; 353 size_t add_obj_size; 354 size_t mark_bitmap_count; 355 size_t mark_bitmap_size; 356 #endif // #ifdef ASSERT 357 358 ParallelCompactData::ParallelCompactData() 359 { 360 _region_start = 0; 361 362 _region_vspace = 0; 363 _reserved_byte_size = 0; 364 _region_data = 0; 365 _region_count = 0; 366 } 367 368 bool ParallelCompactData::initialize(MemRegion covered_region) 369 { 370 _region_start = covered_region.start(); 371 const size_t region_size = covered_region.word_size(); 372 DEBUG_ONLY(_region_end = _region_start + region_size;) 373 374 assert(region_align_down(_region_start) == _region_start, 375 "region start not aligned"); 376 assert((region_size & RegionSizeOffsetMask) == 0, 377 "region size not a multiple of RegionSize"); 378 379 bool result = initialize_region_data(region_size); 380 381 return result; 382 } 383 384 PSVirtualSpace* 385 ParallelCompactData::create_vspace(size_t count, size_t element_size) 386 { 387 const size_t raw_bytes = count * element_size; 388 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); 389 const size_t granularity = os::vm_allocation_granularity(); 390 _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 391 392 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 393 MAX2(page_sz, granularity); 394 ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0); 395 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 396 rs.size()); 397 398 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 399 400 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 401 if (vspace != 0) { 402 if (vspace->expand_by(_reserved_byte_size)) { 403 return vspace; 404 } 405 delete vspace; 406 // Release memory reserved in the space. 407 rs.release(); 408 } 409 410 return 0; 411 } 412 413 bool ParallelCompactData::initialize_region_data(size_t region_size) 414 { 415 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 416 _region_vspace = create_vspace(count, sizeof(RegionData)); 417 if (_region_vspace != 0) { 418 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 419 _region_count = count; 420 return true; 421 } 422 return false; 423 } 424 425 void ParallelCompactData::clear() 426 { 427 memset(_region_data, 0, _region_vspace->committed_size()); 428 } 429 430 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 431 assert(beg_region <= _region_count, "beg_region out of range"); 432 assert(end_region <= _region_count, "end_region out of range"); 433 434 const size_t region_cnt = end_region - beg_region; 435 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 436 } 437 438 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 439 { 440 const RegionData* cur_cp = region(region_idx); 441 const RegionData* const end_cp = region(region_count() - 1); 442 443 HeapWord* result = region_to_addr(region_idx); 444 if (cur_cp < end_cp) { 445 do { 446 result += cur_cp->partial_obj_size(); 447 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 448 } 449 return result; 450 } 451 452 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 453 { 454 const size_t obj_ofs = pointer_delta(addr, _region_start); 455 const size_t beg_region = obj_ofs >> Log2RegionSize; 456 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 457 458 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 459 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 460 461 if (beg_region == end_region) { 462 // All in one region. 463 _region_data[beg_region].add_live_obj(len); 464 return; 465 } 466 467 // First region. 468 const size_t beg_ofs = region_offset(addr); 469 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 470 471 Klass* klass = ((oop)addr)->klass(); 472 // Middle regions--completely spanned by this object. 473 for (size_t region = beg_region + 1; region < end_region; ++region) { 474 _region_data[region].set_partial_obj_size(RegionSize); 475 _region_data[region].set_partial_obj_addr(addr); 476 } 477 478 // Last region. 479 const size_t end_ofs = region_offset(addr + len - 1); 480 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 481 _region_data[end_region].set_partial_obj_addr(addr); 482 } 483 484 void 485 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 486 { 487 assert(region_offset(beg) == 0, "not RegionSize aligned"); 488 assert(region_offset(end) == 0, "not RegionSize aligned"); 489 490 size_t cur_region = addr_to_region_idx(beg); 491 const size_t end_region = addr_to_region_idx(end); 492 HeapWord* addr = beg; 493 while (cur_region < end_region) { 494 _region_data[cur_region].set_destination(addr); 495 _region_data[cur_region].set_destination_count(0); 496 _region_data[cur_region].set_source_region(cur_region); 497 _region_data[cur_region].set_data_location(addr); 498 499 // Update live_obj_size so the region appears completely full. 500 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 501 _region_data[cur_region].set_live_obj_size(live_size); 502 503 ++cur_region; 504 addr += RegionSize; 505 } 506 } 507 508 // Find the point at which a space can be split and, if necessary, record the 509 // split point. 510 // 511 // If the current src region (which overflowed the destination space) doesn't 512 // have a partial object, the split point is at the beginning of the current src 513 // region (an "easy" split, no extra bookkeeping required). 514 // 515 // If the current src region has a partial object, the split point is in the 516 // region where that partial object starts (call it the split_region). If 517 // split_region has a partial object, then the split point is just after that 518 // partial object (a "hard" split where we have to record the split data and 519 // zero the partial_obj_size field). With a "hard" split, we know that the 520 // partial_obj ends within split_region because the partial object that caused 521 // the overflow starts in split_region. If split_region doesn't have a partial 522 // obj, then the split is at the beginning of split_region (another "easy" 523 // split). 524 HeapWord* 525 ParallelCompactData::summarize_split_space(size_t src_region, 526 SplitInfo& split_info, 527 HeapWord* destination, 528 HeapWord* target_end, 529 HeapWord** target_next) 530 { 531 assert(destination <= target_end, "sanity"); 532 assert(destination + _region_data[src_region].data_size() > target_end, 533 "region should not fit into target space"); 534 assert(is_region_aligned(target_end), "sanity"); 535 536 size_t split_region = src_region; 537 HeapWord* split_destination = destination; 538 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 539 540 if (destination + partial_obj_size > target_end) { 541 // The split point is just after the partial object (if any) in the 542 // src_region that contains the start of the object that overflowed the 543 // destination space. 544 // 545 // Find the start of the "overflow" object and set split_region to the 546 // region containing it. 547 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 548 split_region = addr_to_region_idx(overflow_obj); 549 550 // Clear the source_region field of all destination regions whose first word 551 // came from data after the split point (a non-null source_region field 552 // implies a region must be filled). 553 // 554 // An alternative to the simple loop below: clear during post_compact(), 555 // which uses memcpy instead of individual stores, and is easy to 556 // parallelize. (The downside is that it clears the entire RegionData 557 // object as opposed to just one field.) 558 // 559 // post_compact() would have to clear the summary data up to the highest 560 // address that was written during the summary phase, which would be 561 // 562 // max(top, max(new_top, clear_top)) 563 // 564 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 565 // to target_end. 566 const RegionData* const sr = region(split_region); 567 const size_t beg_idx = 568 addr_to_region_idx(region_align_up(sr->destination() + 569 sr->partial_obj_size())); 570 const size_t end_idx = addr_to_region_idx(target_end); 571 572 if (TraceParallelOldGCSummaryPhase) { 573 gclog_or_tty->print_cr("split: clearing source_region field in [" 574 SIZE_FORMAT ", " SIZE_FORMAT ")", 575 beg_idx, end_idx); 576 } 577 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 578 _region_data[idx].set_source_region(0); 579 } 580 581 // Set split_destination and partial_obj_size to reflect the split region. 582 split_destination = sr->destination(); 583 partial_obj_size = sr->partial_obj_size(); 584 } 585 586 // The split is recorded only if a partial object extends onto the region. 587 if (partial_obj_size != 0) { 588 _region_data[split_region].set_partial_obj_size(0); 589 split_info.record(split_region, partial_obj_size, split_destination); 590 } 591 592 // Setup the continuation addresses. 593 *target_next = split_destination + partial_obj_size; 594 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 595 596 if (TraceParallelOldGCSummaryPhase) { 597 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 598 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 599 " pos=" SIZE_FORMAT, 600 split_type, source_next, split_region, 601 partial_obj_size); 602 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 603 " tn=" PTR_FORMAT, 604 split_type, split_destination, 605 addr_to_region_idx(split_destination), 606 *target_next); 607 608 if (partial_obj_size != 0) { 609 HeapWord* const po_beg = split_info.destination(); 610 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 611 gclog_or_tty->print_cr("%s split: " 612 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 613 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 614 split_type, 615 po_beg, addr_to_region_idx(po_beg), 616 po_end, addr_to_region_idx(po_end)); 617 } 618 } 619 620 return source_next; 621 } 622 623 bool ParallelCompactData::summarize(SplitInfo& split_info, 624 HeapWord* source_beg, HeapWord* source_end, 625 HeapWord** source_next, 626 HeapWord* target_beg, HeapWord* target_end, 627 HeapWord** target_next) 628 { 629 if (TraceParallelOldGCSummaryPhase) { 630 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 631 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 632 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 633 source_beg, source_end, source_next_val, 634 target_beg, target_end, *target_next); 635 } 636 637 size_t cur_region = addr_to_region_idx(source_beg); 638 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 639 640 HeapWord *dest_addr = target_beg; 641 while (cur_region < end_region) { 642 // The destination must be set even if the region has no data. 643 _region_data[cur_region].set_destination(dest_addr); 644 645 size_t words = _region_data[cur_region].data_size(); 646 if (words > 0) { 647 // If cur_region does not fit entirely into the target space, find a point 648 // at which the source space can be 'split' so that part is copied to the 649 // target space and the rest is copied elsewhere. 650 if (dest_addr + words > target_end) { 651 assert(source_next != NULL, "source_next is NULL when splitting"); 652 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 653 target_end, target_next); 654 return false; 655 } 656 657 // Compute the destination_count for cur_region, and if necessary, update 658 // source_region for a destination region. The source_region field is 659 // updated if cur_region is the first (left-most) region to be copied to a 660 // destination region. 661 // 662 // The destination_count calculation is a bit subtle. A region that has 663 // data that compacts into itself does not count itself as a destination. 664 // This maintains the invariant that a zero count means the region is 665 // available and can be claimed and then filled. 666 uint destination_count = 0; 667 if (split_info.is_split(cur_region)) { 668 // The current region has been split: the partial object will be copied 669 // to one destination space and the remaining data will be copied to 670 // another destination space. Adjust the initial destination_count and, 671 // if necessary, set the source_region field if the partial object will 672 // cross a destination region boundary. 673 destination_count = split_info.destination_count(); 674 if (destination_count == 2) { 675 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 676 _region_data[dest_idx].set_source_region(cur_region); 677 } 678 } 679 680 HeapWord* const last_addr = dest_addr + words - 1; 681 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 682 const size_t dest_region_2 = addr_to_region_idx(last_addr); 683 684 // Initially assume that the destination regions will be the same and 685 // adjust the value below if necessary. Under this assumption, if 686 // cur_region == dest_region_2, then cur_region will be compacted 687 // completely into itself. 688 destination_count += cur_region == dest_region_2 ? 0 : 1; 689 if (dest_region_1 != dest_region_2) { 690 // Destination regions differ; adjust destination_count. 691 destination_count += 1; 692 // Data from cur_region will be copied to the start of dest_region_2. 693 _region_data[dest_region_2].set_source_region(cur_region); 694 } else if (region_offset(dest_addr) == 0) { 695 // Data from cur_region will be copied to the start of the destination 696 // region. 697 _region_data[dest_region_1].set_source_region(cur_region); 698 } 699 700 _region_data[cur_region].set_destination_count(destination_count); 701 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 702 dest_addr += words; 703 } 704 705 ++cur_region; 706 } 707 708 *target_next = dest_addr; 709 return true; 710 } 711 712 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 713 assert(addr != NULL, "Should detect NULL oop earlier"); 714 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); 715 #ifdef ASSERT 716 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { 717 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); 718 } 719 #endif 720 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); 721 722 // Region covering the object. 723 size_t region_index = addr_to_region_idx(addr); 724 const RegionData* const region_ptr = region(region_index); 725 HeapWord* const region_addr = region_align_down(addr); 726 727 assert(addr < region_addr + RegionSize, "Region does not cover object"); 728 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check"); 729 730 HeapWord* result = region_ptr->destination(); 731 732 // If all the data in the region is live, then the new location of the object 733 // can be calculated from the destination of the region plus the offset of the 734 // object in the region. 735 if (region_ptr->data_size() == RegionSize) { 736 result += pointer_delta(addr, region_addr); 737 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 738 return result; 739 } 740 741 // The new location of the object is 742 // region destination + 743 // size of the partial object extending onto the region + 744 // sizes of the live objects in the Region that are to the left of addr 745 const size_t partial_obj_size = region_ptr->partial_obj_size(); 746 HeapWord* const search_start = region_addr + partial_obj_size; 747 748 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 749 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 750 751 result += partial_obj_size + live_to_left; 752 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 753 return result; 754 } 755 756 #ifdef ASSERT 757 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 758 { 759 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 760 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 761 for (const size_t* p = beg; p < end; ++p) { 762 assert(*p == 0, "not zero"); 763 } 764 } 765 766 void ParallelCompactData::verify_clear() 767 { 768 verify_clear(_region_vspace); 769 } 770 #endif // #ifdef ASSERT 771 772 #ifdef NOT_PRODUCT 773 ParallelCompactData::RegionData* debug_region(size_t region_index) { 774 ParallelCompactData& sd = PSParallelCompact::summary_data(); 775 return sd.region(region_index); 776 } 777 #endif 778 779 STWGCTimer PSParallelCompact::_gc_timer; 780 ParallelOldTracer PSParallelCompact::_gc_tracer; 781 elapsedTimer PSParallelCompact::_accumulated_time; 782 unsigned int PSParallelCompact::_total_invocations = 0; 783 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 784 jlong PSParallelCompact::_time_of_last_gc = 0; 785 CollectorCounters* PSParallelCompact::_counters = NULL; 786 ParMarkBitMap PSParallelCompact::_mark_bitmap; 787 ParallelCompactData PSParallelCompact::_summary_data; 788 789 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 790 791 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 792 793 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 794 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 795 796 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 797 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 798 799 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); } 800 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); } 801 802 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 803 804 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { 805 mark_and_push(_compaction_manager, p); 806 } 807 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 808 809 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 810 klass->oops_do(_mark_and_push_closure); 811 } 812 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 813 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 814 } 815 816 void PSParallelCompact::post_initialize() { 817 ParallelScavengeHeap* heap = gc_heap(); 818 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 819 820 MemRegion mr = heap->reserved_region(); 821 _ref_processor = 822 new ReferenceProcessor(mr, // span 823 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 824 (int) ParallelGCThreads, // mt processing degree 825 true, // mt discovery 826 (int) ParallelGCThreads, // mt discovery degree 827 true, // atomic_discovery 828 &_is_alive_closure, // non-header is alive closure 829 false); // write barrier for next field updates 830 _counters = new CollectorCounters("PSParallelCompact", 1); 831 832 // Initialize static fields in ParCompactionManager. 833 ParCompactionManager::initialize(mark_bitmap()); 834 } 835 836 bool PSParallelCompact::initialize() { 837 ParallelScavengeHeap* heap = gc_heap(); 838 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 839 MemRegion mr = heap->reserved_region(); 840 841 // Was the old gen get allocated successfully? 842 if (!heap->old_gen()->is_allocated()) { 843 return false; 844 } 845 846 initialize_space_info(); 847 initialize_dead_wood_limiter(); 848 849 if (!_mark_bitmap.initialize(mr)) { 850 vm_shutdown_during_initialization( 851 err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel " 852 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 853 _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K)); 854 return false; 855 } 856 857 if (!_summary_data.initialize(mr)) { 858 vm_shutdown_during_initialization( 859 err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel " 860 "garbage collection for the requested " SIZE_FORMAT "KB heap.", 861 _summary_data.reserved_byte_size()/K, mr.byte_size()/K)); 862 return false; 863 } 864 865 return true; 866 } 867 868 void PSParallelCompact::initialize_space_info() 869 { 870 memset(&_space_info, 0, sizeof(_space_info)); 871 872 ParallelScavengeHeap* heap = gc_heap(); 873 PSYoungGen* young_gen = heap->young_gen(); 874 875 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 876 _space_info[eden_space_id].set_space(young_gen->eden_space()); 877 _space_info[from_space_id].set_space(young_gen->from_space()); 878 _space_info[to_space_id].set_space(young_gen->to_space()); 879 880 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 881 } 882 883 void PSParallelCompact::initialize_dead_wood_limiter() 884 { 885 const size_t max = 100; 886 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 887 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 888 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 889 DEBUG_ONLY(_dwl_initialized = true;) 890 _dwl_adjustment = normal_distribution(1.0); 891 } 892 893 // Simple class for storing info about the heap at the start of GC, to be used 894 // after GC for comparison/printing. 895 class PreGCValues { 896 public: 897 PreGCValues() { } 898 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 899 900 void fill(ParallelScavengeHeap* heap) { 901 _heap_used = heap->used(); 902 _young_gen_used = heap->young_gen()->used_in_bytes(); 903 _old_gen_used = heap->old_gen()->used_in_bytes(); 904 _metadata_used = MetaspaceAux::allocated_used_bytes(); 905 }; 906 907 size_t heap_used() const { return _heap_used; } 908 size_t young_gen_used() const { return _young_gen_used; } 909 size_t old_gen_used() const { return _old_gen_used; } 910 size_t metadata_used() const { return _metadata_used; } 911 912 private: 913 size_t _heap_used; 914 size_t _young_gen_used; 915 size_t _old_gen_used; 916 size_t _metadata_used; 917 }; 918 919 void 920 PSParallelCompact::clear_data_covering_space(SpaceId id) 921 { 922 // At this point, top is the value before GC, new_top() is the value that will 923 // be set at the end of GC. The marking bitmap is cleared to top; nothing 924 // should be marked above top. The summary data is cleared to the larger of 925 // top & new_top. 926 MutableSpace* const space = _space_info[id].space(); 927 HeapWord* const bot = space->bottom(); 928 HeapWord* const top = space->top(); 929 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 930 931 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 932 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 933 _mark_bitmap.clear_range(beg_bit, end_bit); 934 935 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 936 const size_t end_region = 937 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 938 _summary_data.clear_range(beg_region, end_region); 939 940 // Clear the data used to 'split' regions. 941 SplitInfo& split_info = _space_info[id].split_info(); 942 if (split_info.is_valid()) { 943 split_info.clear(); 944 } 945 DEBUG_ONLY(split_info.verify_clear();) 946 } 947 948 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 949 { 950 // Update the from & to space pointers in space_info, since they are swapped 951 // at each young gen gc. Do the update unconditionally (even though a 952 // promotion failure does not swap spaces) because an unknown number of minor 953 // collections will have swapped the spaces an unknown number of times. 954 GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer); 955 ParallelScavengeHeap* heap = gc_heap(); 956 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 957 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 958 959 pre_gc_values->fill(heap); 960 961 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 962 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 963 964 // Increment the invocation count 965 heap->increment_total_collections(true); 966 967 // We need to track unique mark sweep invocations as well. 968 _total_invocations++; 969 970 heap->print_heap_before_gc(); 971 heap->trace_heap_before_gc(&_gc_tracer); 972 973 // Fill in TLABs 974 heap->accumulate_statistics_all_tlabs(); 975 heap->ensure_parsability(true); // retire TLABs 976 977 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 978 HandleMark hm; // Discard invalid handles created during verification 979 Universe::verify(" VerifyBeforeGC:"); 980 } 981 982 // Verify object start arrays 983 if (VerifyObjectStartArray && 984 VerifyBeforeGC) { 985 heap->old_gen()->verify_object_start_array(); 986 } 987 988 DEBUG_ONLY(mark_bitmap()->verify_clear();) 989 DEBUG_ONLY(summary_data().verify_clear();) 990 991 // Have worker threads release resources the next time they run a task. 992 gc_task_manager()->release_all_resources(); 993 } 994 995 void PSParallelCompact::post_compact() 996 { 997 GCTraceTime tm("post compact", print_phases(), true, &_gc_timer); 998 999 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1000 // Clear the marking bitmap, summary data and split info. 1001 clear_data_covering_space(SpaceId(id)); 1002 // Update top(). Must be done after clearing the bitmap and summary data. 1003 _space_info[id].publish_new_top(); 1004 } 1005 1006 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1007 MutableSpace* const from_space = _space_info[from_space_id].space(); 1008 MutableSpace* const to_space = _space_info[to_space_id].space(); 1009 1010 ParallelScavengeHeap* heap = gc_heap(); 1011 bool eden_empty = eden_space->is_empty(); 1012 if (!eden_empty) { 1013 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1014 heap->young_gen(), heap->old_gen()); 1015 } 1016 1017 // Update heap occupancy information which is used as input to the soft ref 1018 // clearing policy at the next gc. 1019 Universe::update_heap_info_at_gc(); 1020 1021 bool young_gen_empty = eden_empty && from_space->is_empty() && 1022 to_space->is_empty(); 1023 1024 BarrierSet* bs = heap->barrier_set(); 1025 if (bs->is_a(BarrierSet::ModRef)) { 1026 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1027 MemRegion old_mr = heap->old_gen()->reserved(); 1028 1029 if (young_gen_empty) { 1030 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1031 } else { 1032 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1033 } 1034 } 1035 1036 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1037 ClassLoaderDataGraph::purge(); 1038 MetaspaceAux::verify_metrics(); 1039 1040 Threads::gc_epilogue(); 1041 CodeCache::gc_epilogue(); 1042 JvmtiExport::gc_epilogue(); 1043 1044 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1045 1046 ref_processor()->enqueue_discovered_references(NULL); 1047 1048 if (ZapUnusedHeapArea) { 1049 heap->gen_mangle_unused_area(); 1050 } 1051 1052 // Update time of last GC 1053 reset_millis_since_last_gc(); 1054 } 1055 1056 HeapWord* 1057 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1058 bool maximum_compaction) 1059 { 1060 const size_t region_size = ParallelCompactData::RegionSize; 1061 const ParallelCompactData& sd = summary_data(); 1062 1063 const MutableSpace* const space = _space_info[id].space(); 1064 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1065 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1066 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1067 1068 // Skip full regions at the beginning of the space--they are necessarily part 1069 // of the dense prefix. 1070 size_t full_count = 0; 1071 const RegionData* cp; 1072 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1073 ++full_count; 1074 } 1075 1076 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1077 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1078 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1079 if (maximum_compaction || cp == end_cp || interval_ended) { 1080 _maximum_compaction_gc_num = total_invocations(); 1081 return sd.region_to_addr(cp); 1082 } 1083 1084 HeapWord* const new_top = _space_info[id].new_top(); 1085 const size_t space_live = pointer_delta(new_top, space->bottom()); 1086 const size_t space_used = space->used_in_words(); 1087 const size_t space_capacity = space->capacity_in_words(); 1088 1089 const double cur_density = double(space_live) / space_capacity; 1090 const double deadwood_density = 1091 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1092 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1093 1094 if (TraceParallelOldGCDensePrefix) { 1095 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1096 cur_density, deadwood_density, deadwood_goal); 1097 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1098 "space_cap=" SIZE_FORMAT, 1099 space_live, space_used, 1100 space_capacity); 1101 } 1102 1103 // XXX - Use binary search? 1104 HeapWord* dense_prefix = sd.region_to_addr(cp); 1105 const RegionData* full_cp = cp; 1106 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1107 while (cp < end_cp) { 1108 HeapWord* region_destination = cp->destination(); 1109 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1110 if (TraceParallelOldGCDensePrefix && Verbose) { 1111 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1112 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1113 sd.region(cp), region_destination, 1114 dense_prefix, cur_deadwood); 1115 } 1116 1117 if (cur_deadwood >= deadwood_goal) { 1118 // Found the region that has the correct amount of deadwood to the left. 1119 // This typically occurs after crossing a fairly sparse set of regions, so 1120 // iterate backwards over those sparse regions, looking for the region 1121 // that has the lowest density of live objects 'to the right.' 1122 size_t space_to_left = sd.region(cp) * region_size; 1123 size_t live_to_left = space_to_left - cur_deadwood; 1124 size_t space_to_right = space_capacity - space_to_left; 1125 size_t live_to_right = space_live - live_to_left; 1126 double density_to_right = double(live_to_right) / space_to_right; 1127 while (cp > full_cp) { 1128 --cp; 1129 const size_t prev_region_live_to_right = live_to_right - 1130 cp->data_size(); 1131 const size_t prev_region_space_to_right = space_to_right + region_size; 1132 double prev_region_density_to_right = 1133 double(prev_region_live_to_right) / prev_region_space_to_right; 1134 if (density_to_right <= prev_region_density_to_right) { 1135 return dense_prefix; 1136 } 1137 if (TraceParallelOldGCDensePrefix && Verbose) { 1138 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1139 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1140 prev_region_density_to_right); 1141 } 1142 dense_prefix -= region_size; 1143 live_to_right = prev_region_live_to_right; 1144 space_to_right = prev_region_space_to_right; 1145 density_to_right = prev_region_density_to_right; 1146 } 1147 return dense_prefix; 1148 } 1149 1150 dense_prefix += region_size; 1151 ++cp; 1152 } 1153 1154 return dense_prefix; 1155 } 1156 1157 #ifndef PRODUCT 1158 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1159 const SpaceId id, 1160 const bool maximum_compaction, 1161 HeapWord* const addr) 1162 { 1163 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1164 RegionData* const cp = summary_data().region(region_idx); 1165 const MutableSpace* const space = _space_info[id].space(); 1166 HeapWord* const new_top = _space_info[id].new_top(); 1167 1168 const size_t space_live = pointer_delta(new_top, space->bottom()); 1169 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1170 const size_t space_cap = space->capacity_in_words(); 1171 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1172 const size_t live_to_right = new_top - cp->destination(); 1173 const size_t dead_to_right = space->top() - addr - live_to_right; 1174 1175 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1176 "spl=" SIZE_FORMAT " " 1177 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1178 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1179 " ratio=%10.8f", 1180 algorithm, addr, region_idx, 1181 space_live, 1182 dead_to_left, dead_to_left_pct, 1183 dead_to_right, live_to_right, 1184 double(dead_to_right) / live_to_right); 1185 } 1186 #endif // #ifndef PRODUCT 1187 1188 // Return a fraction indicating how much of the generation can be treated as 1189 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1190 // based on the density of live objects in the generation to determine a limit, 1191 // which is then adjusted so the return value is min_percent when the density is 1192 // 1. 1193 // 1194 // The following table shows some return values for a different values of the 1195 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1196 // min_percent is 1. 1197 // 1198 // fraction allowed as dead wood 1199 // ----------------------------------------------------------------- 1200 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1201 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1202 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1203 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1204 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1205 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1206 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1207 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1208 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1209 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1210 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1211 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1212 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1213 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1214 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1215 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1216 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1217 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1218 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1219 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1220 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1221 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1222 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1223 1224 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1225 { 1226 assert(_dwl_initialized, "uninitialized"); 1227 1228 // The raw limit is the value of the normal distribution at x = density. 1229 const double raw_limit = normal_distribution(density); 1230 1231 // Adjust the raw limit so it becomes the minimum when the density is 1. 1232 // 1233 // First subtract the adjustment value (which is simply the precomputed value 1234 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1235 // Then add the minimum value, so the minimum is returned when the density is 1236 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1237 const double min = double(min_percent) / 100.0; 1238 const double limit = raw_limit - _dwl_adjustment + min; 1239 return MAX2(limit, 0.0); 1240 } 1241 1242 ParallelCompactData::RegionData* 1243 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1244 const RegionData* end) 1245 { 1246 const size_t region_size = ParallelCompactData::RegionSize; 1247 ParallelCompactData& sd = summary_data(); 1248 size_t left = sd.region(beg); 1249 size_t right = end > beg ? sd.region(end) - 1 : left; 1250 1251 // Binary search. 1252 while (left < right) { 1253 // Equivalent to (left + right) / 2, but does not overflow. 1254 const size_t middle = left + (right - left) / 2; 1255 RegionData* const middle_ptr = sd.region(middle); 1256 HeapWord* const dest = middle_ptr->destination(); 1257 HeapWord* const addr = sd.region_to_addr(middle); 1258 assert(dest != NULL, "sanity"); 1259 assert(dest <= addr, "must move left"); 1260 1261 if (middle > left && dest < addr) { 1262 right = middle - 1; 1263 } else if (middle < right && middle_ptr->data_size() == region_size) { 1264 left = middle + 1; 1265 } else { 1266 return middle_ptr; 1267 } 1268 } 1269 return sd.region(left); 1270 } 1271 1272 ParallelCompactData::RegionData* 1273 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1274 const RegionData* end, 1275 size_t dead_words) 1276 { 1277 ParallelCompactData& sd = summary_data(); 1278 size_t left = sd.region(beg); 1279 size_t right = end > beg ? sd.region(end) - 1 : left; 1280 1281 // Binary search. 1282 while (left < right) { 1283 // Equivalent to (left + right) / 2, but does not overflow. 1284 const size_t middle = left + (right - left) / 2; 1285 RegionData* const middle_ptr = sd.region(middle); 1286 HeapWord* const dest = middle_ptr->destination(); 1287 HeapWord* const addr = sd.region_to_addr(middle); 1288 assert(dest != NULL, "sanity"); 1289 assert(dest <= addr, "must move left"); 1290 1291 const size_t dead_to_left = pointer_delta(addr, dest); 1292 if (middle > left && dead_to_left > dead_words) { 1293 right = middle - 1; 1294 } else if (middle < right && dead_to_left < dead_words) { 1295 left = middle + 1; 1296 } else { 1297 return middle_ptr; 1298 } 1299 } 1300 return sd.region(left); 1301 } 1302 1303 // The result is valid during the summary phase, after the initial summarization 1304 // of each space into itself, and before final summarization. 1305 inline double 1306 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1307 HeapWord* const bottom, 1308 HeapWord* const top, 1309 HeapWord* const new_top) 1310 { 1311 ParallelCompactData& sd = summary_data(); 1312 1313 assert(cp != NULL, "sanity"); 1314 assert(bottom != NULL, "sanity"); 1315 assert(top != NULL, "sanity"); 1316 assert(new_top != NULL, "sanity"); 1317 assert(top >= new_top, "summary data problem?"); 1318 assert(new_top > bottom, "space is empty; should not be here"); 1319 assert(new_top >= cp->destination(), "sanity"); 1320 assert(top >= sd.region_to_addr(cp), "sanity"); 1321 1322 HeapWord* const destination = cp->destination(); 1323 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1324 const size_t compacted_region_live = pointer_delta(new_top, destination); 1325 const size_t compacted_region_used = pointer_delta(top, 1326 sd.region_to_addr(cp)); 1327 const size_t reclaimable = compacted_region_used - compacted_region_live; 1328 1329 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1330 return double(reclaimable) / divisor; 1331 } 1332 1333 // Return the address of the end of the dense prefix, a.k.a. the start of the 1334 // compacted region. The address is always on a region boundary. 1335 // 1336 // Completely full regions at the left are skipped, since no compaction can 1337 // occur in those regions. Then the maximum amount of dead wood to allow is 1338 // computed, based on the density (amount live / capacity) of the generation; 1339 // the region with approximately that amount of dead space to the left is 1340 // identified as the limit region. Regions between the last completely full 1341 // region and the limit region are scanned and the one that has the best 1342 // (maximum) reclaimed_ratio() is selected. 1343 HeapWord* 1344 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1345 bool maximum_compaction) 1346 { 1347 if (ParallelOldGCSplitALot) { 1348 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1349 // The value was chosen to provoke splitting a young gen space; use it. 1350 return _space_info[id].dense_prefix(); 1351 } 1352 } 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, %d)=%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 #if 0 1427 // Something to consider: if the region with the best ratio is 'close to' the 1428 // first region w/free space, choose the first region with free space 1429 // ("first-free"). The first-free region is usually near the start of the 1430 // heap, which means we are copying most of the heap already, so copy a bit 1431 // more to get complete compaction. 1432 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1433 _maximum_compaction_gc_num = total_invocations(); 1434 best_cp = full_cp; 1435 } 1436 #endif // #if 0 1437 1438 return sd.region_to_addr(best_cp); 1439 } 1440 1441 #ifndef PRODUCT 1442 void 1443 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1444 size_t words) 1445 { 1446 if (TraceParallelOldGCSummaryPhase) { 1447 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1448 SIZE_FORMAT, start, start + words, words); 1449 } 1450 1451 ObjectStartArray* const start_array = _space_info[id].start_array(); 1452 CollectedHeap::fill_with_objects(start, words); 1453 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1454 _mark_bitmap.mark_obj(p, words); 1455 _summary_data.add_obj(p, words); 1456 start_array->allocate_block(p); 1457 } 1458 } 1459 1460 void 1461 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1462 { 1463 ParallelCompactData& sd = summary_data(); 1464 MutableSpace* space = _space_info[id].space(); 1465 1466 // Find the source and destination start addresses. 1467 HeapWord* const src_addr = sd.region_align_down(start); 1468 HeapWord* dst_addr; 1469 if (src_addr < start) { 1470 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1471 } else if (src_addr > space->bottom()) { 1472 // The start (the original top() value) is aligned to a region boundary so 1473 // the associated region does not have a destination. Compute the 1474 // destination from the previous region. 1475 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1476 dst_addr = cp->destination() + cp->data_size(); 1477 } else { 1478 // Filling the entire space. 1479 dst_addr = space->bottom(); 1480 } 1481 assert(dst_addr != NULL, "sanity"); 1482 1483 // Update the summary data. 1484 bool result = _summary_data.summarize(_space_info[id].split_info(), 1485 src_addr, space->top(), NULL, 1486 dst_addr, space->end(), 1487 _space_info[id].new_top_addr()); 1488 assert(result, "should not fail: bad filler object size"); 1489 } 1490 1491 void 1492 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1493 { 1494 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1495 return; 1496 } 1497 1498 MutableSpace* const space = _space_info[id].space(); 1499 if (space->is_empty()) { 1500 HeapWord* b = space->bottom(); 1501 HeapWord* t = b + space->capacity_in_words() / 2; 1502 space->set_top(t); 1503 if (ZapUnusedHeapArea) { 1504 space->set_top_for_allocations(); 1505 } 1506 1507 size_t min_size = CollectedHeap::min_fill_size(); 1508 size_t obj_len = min_size; 1509 while (b + obj_len <= t) { 1510 CollectedHeap::fill_with_object(b, obj_len); 1511 mark_bitmap()->mark_obj(b, obj_len); 1512 summary_data().add_obj(b, obj_len); 1513 b += obj_len; 1514 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1515 } 1516 if (b < t) { 1517 // The loop didn't completely fill to t (top); adjust top downward. 1518 space->set_top(b); 1519 if (ZapUnusedHeapArea) { 1520 space->set_top_for_allocations(); 1521 } 1522 } 1523 1524 HeapWord** nta = _space_info[id].new_top_addr(); 1525 bool result = summary_data().summarize(_space_info[id].split_info(), 1526 space->bottom(), space->top(), NULL, 1527 space->bottom(), space->end(), nta); 1528 assert(result, "space must fit into itself"); 1529 } 1530 } 1531 1532 void 1533 PSParallelCompact::provoke_split(bool & max_compaction) 1534 { 1535 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1536 return; 1537 } 1538 1539 const size_t region_size = ParallelCompactData::RegionSize; 1540 ParallelCompactData& sd = summary_data(); 1541 1542 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1543 MutableSpace* const from_space = _space_info[from_space_id].space(); 1544 const size_t eden_live = pointer_delta(eden_space->top(), 1545 _space_info[eden_space_id].new_top()); 1546 const size_t from_live = pointer_delta(from_space->top(), 1547 _space_info[from_space_id].new_top()); 1548 1549 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1550 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1551 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1552 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1553 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1554 1555 // Choose the space to split; need at least 2 regions live (or fillable). 1556 SpaceId id; 1557 MutableSpace* space; 1558 size_t live_words; 1559 size_t fill_words; 1560 if (eden_live + eden_fillable >= region_size * 2) { 1561 id = eden_space_id; 1562 space = eden_space; 1563 live_words = eden_live; 1564 fill_words = eden_fillable; 1565 } else if (from_live + from_fillable >= region_size * 2) { 1566 id = from_space_id; 1567 space = from_space; 1568 live_words = from_live; 1569 fill_words = from_fillable; 1570 } else { 1571 return; // Give up. 1572 } 1573 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1574 1575 if (live_words < region_size * 2) { 1576 // Fill from top() to end() w/live objects of mixed sizes. 1577 HeapWord* const fill_start = space->top(); 1578 live_words += fill_words; 1579 1580 space->set_top(fill_start + fill_words); 1581 if (ZapUnusedHeapArea) { 1582 space->set_top_for_allocations(); 1583 } 1584 1585 HeapWord* cur_addr = fill_start; 1586 while (fill_words > 0) { 1587 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1588 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1589 if (fill_words - cur_size < min_fill_size) { 1590 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1591 } 1592 1593 CollectedHeap::fill_with_object(cur_addr, cur_size); 1594 mark_bitmap()->mark_obj(cur_addr, cur_size); 1595 sd.add_obj(cur_addr, cur_size); 1596 1597 cur_addr += cur_size; 1598 fill_words -= cur_size; 1599 } 1600 1601 summarize_new_objects(id, fill_start); 1602 } 1603 1604 max_compaction = false; 1605 1606 // Manipulate the old gen so that it has room for about half of the live data 1607 // in the target young gen space (live_words / 2). 1608 id = old_space_id; 1609 space = _space_info[id].space(); 1610 const size_t free_at_end = space->free_in_words(); 1611 const size_t free_target = align_object_size(live_words / 2); 1612 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1613 1614 if (free_at_end >= free_target + min_fill_size) { 1615 // Fill space above top() and set the dense prefix so everything survives. 1616 HeapWord* const fill_start = space->top(); 1617 const size_t fill_size = free_at_end - free_target; 1618 space->set_top(space->top() + fill_size); 1619 if (ZapUnusedHeapArea) { 1620 space->set_top_for_allocations(); 1621 } 1622 fill_with_live_objects(id, fill_start, fill_size); 1623 summarize_new_objects(id, fill_start); 1624 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1625 } else if (dead + free_at_end > free_target) { 1626 // Find a dense prefix that makes the right amount of space available. 1627 HeapWord* cur = sd.region_align_down(space->top()); 1628 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1629 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1630 while (dead_to_right < free_target) { 1631 cur -= region_size; 1632 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1633 dead_to_right = pointer_delta(space->end(), cur_destination); 1634 } 1635 _space_info[id].set_dense_prefix(cur); 1636 } 1637 } 1638 #endif // #ifndef PRODUCT 1639 1640 void PSParallelCompact::summarize_spaces_quick() 1641 { 1642 for (unsigned int i = 0; i < last_space_id; ++i) { 1643 const MutableSpace* space = _space_info[i].space(); 1644 HeapWord** nta = _space_info[i].new_top_addr(); 1645 bool result = _summary_data.summarize(_space_info[i].split_info(), 1646 space->bottom(), space->top(), NULL, 1647 space->bottom(), space->end(), nta); 1648 assert(result, "space must fit into itself"); 1649 _space_info[i].set_dense_prefix(space->bottom()); 1650 } 1651 1652 #ifndef PRODUCT 1653 if (ParallelOldGCSplitALot) { 1654 provoke_split_fill_survivor(to_space_id); 1655 } 1656 #endif // #ifndef PRODUCT 1657 } 1658 1659 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1660 { 1661 HeapWord* const dense_prefix_end = dense_prefix(id); 1662 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1663 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1664 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1665 // Only enough dead space is filled so that any remaining dead space to the 1666 // left is larger than the minimum filler object. (The remainder is filled 1667 // during the copy/update phase.) 1668 // 1669 // The size of the dead space to the right of the boundary is not a 1670 // concern, since compaction will be able to use whatever space is 1671 // available. 1672 // 1673 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1674 // surrounds the space to be filled with an object. 1675 // 1676 // In the 32-bit VM, each bit represents two 32-bit words: 1677 // +---+ 1678 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1679 // end_bits: ... x x x | 0 | || 0 x x ... 1680 // +---+ 1681 // 1682 // In the 64-bit VM, each bit represents one 64-bit word: 1683 // +------------+ 1684 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1685 // end_bits: ... x x 1 | 0 || 0 | x x ... 1686 // +------------+ 1687 // +-------+ 1688 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1689 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1690 // +-------+ 1691 // +-----------+ 1692 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1693 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1694 // +-----------+ 1695 // +-------+ 1696 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1697 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1698 // +-------+ 1699 1700 // Initially assume case a, c or e will apply. 1701 size_t obj_len = CollectedHeap::min_fill_size(); 1702 HeapWord* obj_beg = dense_prefix_end - obj_len; 1703 1704 #ifdef _LP64 1705 if (MinObjAlignment > 1) { // object alignment > heap word size 1706 // Cases a, c or e. 1707 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1708 // Case b above. 1709 obj_beg = dense_prefix_end - 1; 1710 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1711 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1712 // Case d above. 1713 obj_beg = dense_prefix_end - 3; 1714 obj_len = 3; 1715 } 1716 #endif // #ifdef _LP64 1717 1718 CollectedHeap::fill_with_object(obj_beg, obj_len); 1719 _mark_bitmap.mark_obj(obj_beg, obj_len); 1720 _summary_data.add_obj(obj_beg, obj_len); 1721 assert(start_array(id) != NULL, "sanity"); 1722 start_array(id)->allocate_block(obj_beg); 1723 } 1724 } 1725 1726 void 1727 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1728 { 1729 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1730 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1731 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1732 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1733 cur->set_source_region(0); 1734 } 1735 } 1736 1737 void 1738 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1739 { 1740 assert(id < last_space_id, "id out of range"); 1741 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1742 ParallelOldGCSplitALot && id == old_space_id, 1743 "should have been reset in summarize_spaces_quick()"); 1744 1745 const MutableSpace* space = _space_info[id].space(); 1746 if (_space_info[id].new_top() != space->bottom()) { 1747 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1748 _space_info[id].set_dense_prefix(dense_prefix_end); 1749 1750 #ifndef PRODUCT 1751 if (TraceParallelOldGCDensePrefix) { 1752 print_dense_prefix_stats("ratio", id, maximum_compaction, 1753 dense_prefix_end); 1754 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1755 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1756 } 1757 #endif // #ifndef PRODUCT 1758 1759 // Recompute the summary data, taking into account the dense prefix. If 1760 // every last byte will be reclaimed, then the existing summary data which 1761 // compacts everything can be left in place. 1762 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1763 // If dead space crosses the dense prefix boundary, it is (at least 1764 // partially) filled with a dummy object, marked live and added to the 1765 // summary data. This simplifies the copy/update phase and must be done 1766 // before the final locations of objects are determined, to prevent 1767 // leaving a fragment of dead space that is too small to fill. 1768 fill_dense_prefix_end(id); 1769 1770 // Compute the destination of each Region, and thus each object. 1771 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1772 _summary_data.summarize(_space_info[id].split_info(), 1773 dense_prefix_end, space->top(), NULL, 1774 dense_prefix_end, space->end(), 1775 _space_info[id].new_top_addr()); 1776 } 1777 } 1778 1779 if (TraceParallelOldGCSummaryPhase) { 1780 const size_t region_size = ParallelCompactData::RegionSize; 1781 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1782 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1783 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1784 HeapWord* const new_top = _space_info[id].new_top(); 1785 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1786 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1787 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1788 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1789 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1790 id, space->capacity_in_words(), dense_prefix_end, 1791 dp_region, dp_words / region_size, 1792 cr_words / region_size, new_top); 1793 } 1794 } 1795 1796 #ifndef PRODUCT 1797 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1798 HeapWord* dst_beg, HeapWord* dst_end, 1799 SpaceId src_space_id, 1800 HeapWord* src_beg, HeapWord* src_end) 1801 { 1802 if (TraceParallelOldGCSummaryPhase) { 1803 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1804 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1805 SIZE_FORMAT "-" SIZE_FORMAT " " 1806 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1807 SIZE_FORMAT "-" SIZE_FORMAT, 1808 src_space_id, space_names[src_space_id], 1809 dst_space_id, space_names[dst_space_id], 1810 src_beg, src_end, 1811 _summary_data.addr_to_region_idx(src_beg), 1812 _summary_data.addr_to_region_idx(src_end), 1813 dst_beg, dst_end, 1814 _summary_data.addr_to_region_idx(dst_beg), 1815 _summary_data.addr_to_region_idx(dst_end)); 1816 } 1817 } 1818 #endif // #ifndef PRODUCT 1819 1820 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1821 bool maximum_compaction) 1822 { 1823 GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer); 1824 // trace("2"); 1825 1826 #ifdef ASSERT 1827 if (TraceParallelOldGCMarkingPhase) { 1828 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1829 "add_obj_bytes=" SIZE_FORMAT, 1830 add_obj_count, add_obj_size * HeapWordSize); 1831 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1832 "mark_bitmap_bytes=" SIZE_FORMAT, 1833 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1834 } 1835 #endif // #ifdef ASSERT 1836 1837 // Quick summarization of each space into itself, to see how much is live. 1838 summarize_spaces_quick(); 1839 1840 if (TraceParallelOldGCSummaryPhase) { 1841 tty->print_cr("summary_phase: after summarizing each space to self"); 1842 Universe::print(); 1843 NOT_PRODUCT(print_region_ranges()); 1844 if (Verbose) { 1845 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1846 } 1847 } 1848 1849 // The amount of live data that will end up in old space (assuming it fits). 1850 size_t old_space_total_live = 0; 1851 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1852 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1853 _space_info[id].space()->bottom()); 1854 } 1855 1856 MutableSpace* const old_space = _space_info[old_space_id].space(); 1857 const size_t old_capacity = old_space->capacity_in_words(); 1858 if (old_space_total_live > old_capacity) { 1859 // XXX - should also try to expand 1860 maximum_compaction = true; 1861 } 1862 #ifndef PRODUCT 1863 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1864 provoke_split(maximum_compaction); 1865 } 1866 #endif // #ifndef PRODUCT 1867 1868 // Old generations. 1869 summarize_space(old_space_id, maximum_compaction); 1870 1871 // Summarize the remaining spaces in the young gen. The initial target space 1872 // is the old gen. If a space does not fit entirely into the target, then the 1873 // remainder is compacted into the space itself and that space becomes the new 1874 // target. 1875 SpaceId dst_space_id = old_space_id; 1876 HeapWord* dst_space_end = old_space->end(); 1877 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1878 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1879 const MutableSpace* space = _space_info[id].space(); 1880 const size_t live = pointer_delta(_space_info[id].new_top(), 1881 space->bottom()); 1882 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1883 1884 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1885 SpaceId(id), space->bottom(), space->top());) 1886 if (live > 0 && live <= available) { 1887 // All the live data will fit. 1888 bool done = _summary_data.summarize(_space_info[id].split_info(), 1889 space->bottom(), space->top(), 1890 NULL, 1891 *new_top_addr, dst_space_end, 1892 new_top_addr); 1893 assert(done, "space must fit into old gen"); 1894 1895 // Reset the new_top value for the space. 1896 _space_info[id].set_new_top(space->bottom()); 1897 } else if (live > 0) { 1898 // Attempt to fit part of the source space into the target space. 1899 HeapWord* next_src_addr = NULL; 1900 bool done = _summary_data.summarize(_space_info[id].split_info(), 1901 space->bottom(), space->top(), 1902 &next_src_addr, 1903 *new_top_addr, dst_space_end, 1904 new_top_addr); 1905 assert(!done, "space should not fit into old gen"); 1906 assert(next_src_addr != NULL, "sanity"); 1907 1908 // The source space becomes the new target, so the remainder is compacted 1909 // within the space itself. 1910 dst_space_id = SpaceId(id); 1911 dst_space_end = space->end(); 1912 new_top_addr = _space_info[id].new_top_addr(); 1913 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1914 space->bottom(), dst_space_end, 1915 SpaceId(id), next_src_addr, space->top());) 1916 done = _summary_data.summarize(_space_info[id].split_info(), 1917 next_src_addr, space->top(), 1918 NULL, 1919 space->bottom(), dst_space_end, 1920 new_top_addr); 1921 assert(done, "space must fit when compacted into itself"); 1922 assert(*new_top_addr <= space->top(), "usage should not grow"); 1923 } 1924 } 1925 1926 if (TraceParallelOldGCSummaryPhase) { 1927 tty->print_cr("summary_phase: after final summarization"); 1928 Universe::print(); 1929 NOT_PRODUCT(print_region_ranges()); 1930 if (Verbose) { 1931 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1932 } 1933 } 1934 } 1935 1936 // This method should contain all heap-specific policy for invoking a full 1937 // collection. invoke_no_policy() will only attempt to compact the heap; it 1938 // will do nothing further. If we need to bail out for policy reasons, scavenge 1939 // before full gc, or any other specialized behavior, it needs to be added here. 1940 // 1941 // Note that this method should only be called from the vm_thread while at a 1942 // safepoint. 1943 // 1944 // Note that the all_soft_refs_clear flag in the collector policy 1945 // may be true because this method can be called without intervening 1946 // activity. For example when the heap space is tight and full measure 1947 // are being taken to free space. 1948 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1949 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1950 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1951 "should be in vm thread"); 1952 1953 ParallelScavengeHeap* heap = gc_heap(); 1954 GCCause::Cause gc_cause = heap->gc_cause(); 1955 assert(!heap->is_gc_active(), "not reentrant"); 1956 1957 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1958 IsGCActiveMark mark; 1959 1960 if (ScavengeBeforeFullGC) { 1961 PSScavenge::invoke_no_policy(); 1962 } 1963 1964 const bool clear_all_soft_refs = 1965 heap->collector_policy()->should_clear_all_soft_refs(); 1966 1967 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1968 maximum_heap_compaction); 1969 } 1970 1971 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1972 size_t addr_region_index = addr_to_region_idx(addr); 1973 return region_index == addr_region_index; 1974 } 1975 1976 // This method contains no policy. You should probably 1977 // be calling invoke() instead. 1978 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1979 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1980 assert(ref_processor() != NULL, "Sanity"); 1981 1982 if (GC_locker::check_active_before_gc()) { 1983 return false; 1984 } 1985 1986 ParallelScavengeHeap* heap = gc_heap(); 1987 1988 _gc_timer.register_gc_start(os::elapsed_counter()); 1989 _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start()); 1990 1991 TimeStamp marking_start; 1992 TimeStamp compaction_start; 1993 TimeStamp collection_exit; 1994 1995 GCCause::Cause gc_cause = heap->gc_cause(); 1996 PSYoungGen* young_gen = heap->young_gen(); 1997 PSOldGen* old_gen = heap->old_gen(); 1998 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1999 2000 // The scope of casr should end after code that can change 2001 // CollectorPolicy::_should_clear_all_soft_refs. 2002 ClearedAllSoftRefs casr(maximum_heap_compaction, 2003 heap->collector_policy()); 2004 2005 if (ZapUnusedHeapArea) { 2006 // Save information needed to minimize mangling 2007 heap->record_gen_tops_before_GC(); 2008 } 2009 2010 heap->pre_full_gc_dump(&_gc_timer); 2011 2012 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2013 2014 // Make sure data structures are sane, make the heap parsable, and do other 2015 // miscellaneous bookkeeping. 2016 PreGCValues pre_gc_values; 2017 pre_compact(&pre_gc_values); 2018 2019 // Get the compaction manager reserved for the VM thread. 2020 ParCompactionManager* const vmthread_cm = 2021 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2022 2023 // Place after pre_compact() where the number of invocations is incremented. 2024 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2025 2026 { 2027 ResourceMark rm; 2028 HandleMark hm; 2029 2030 // Set the number of GC threads to be used in this collection 2031 gc_task_manager()->set_active_gang(); 2032 gc_task_manager()->task_idle_workers(); 2033 heap->set_par_threads(gc_task_manager()->active_workers()); 2034 2035 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2036 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2037 GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 2038 TraceCollectorStats tcs(counters()); 2039 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2040 2041 if (TraceGen1Time) accumulated_time()->start(); 2042 2043 // Let the size policy know we're starting 2044 size_policy->major_collection_begin(); 2045 2046 CodeCache::gc_prologue(); 2047 Threads::gc_prologue(); 2048 2049 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2050 2051 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 2052 ref_processor()->setup_policy(maximum_heap_compaction); 2053 2054 bool marked_for_unloading = false; 2055 2056 marking_start.update(); 2057 marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer); 2058 2059 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2060 && gc_cause == GCCause::_java_lang_system_gc; 2061 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2062 2063 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2064 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2065 2066 // adjust_roots() updates Universe::_intArrayKlassObj which is 2067 // needed by the compaction for filling holes in the dense prefix. 2068 adjust_roots(); 2069 2070 compaction_start.update(); 2071 compact(); 2072 2073 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2074 // done before resizing. 2075 post_compact(); 2076 2077 // Let the size policy know we're done 2078 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2079 2080 if (UseAdaptiveSizePolicy) { 2081 if (PrintAdaptiveSizePolicy) { 2082 gclog_or_tty->print("AdaptiveSizeStart: "); 2083 gclog_or_tty->stamp(); 2084 gclog_or_tty->print_cr(" collection: %d ", 2085 heap->total_collections()); 2086 if (Verbose) { 2087 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d", 2088 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2089 } 2090 } 2091 2092 // Don't check if the size_policy is ready here. Let 2093 // the size_policy check that internally. 2094 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2095 ((gc_cause != GCCause::_java_lang_system_gc) || 2096 UseAdaptiveSizePolicyWithSystemGC)) { 2097 // Calculate optimal free space amounts 2098 assert(young_gen->max_size() > 2099 young_gen->from_space()->capacity_in_bytes() + 2100 young_gen->to_space()->capacity_in_bytes(), 2101 "Sizes of space in young gen are out-of-bounds"); 2102 2103 size_t young_live = young_gen->used_in_bytes(); 2104 size_t eden_live = young_gen->eden_space()->used_in_bytes(); 2105 size_t old_live = old_gen->used_in_bytes(); 2106 size_t cur_eden = young_gen->eden_space()->capacity_in_bytes(); 2107 size_t max_old_gen_size = old_gen->max_gen_size(); 2108 size_t max_eden_size = young_gen->max_size() - 2109 young_gen->from_space()->capacity_in_bytes() - 2110 young_gen->to_space()->capacity_in_bytes(); 2111 2112 // Used for diagnostics 2113 size_policy->clear_generation_free_space_flags(); 2114 2115 size_policy->compute_generations_free_space(young_live, 2116 eden_live, 2117 old_live, 2118 cur_eden, 2119 max_old_gen_size, 2120 max_eden_size, 2121 true /* full gc*/); 2122 2123 size_policy->check_gc_overhead_limit(young_live, 2124 eden_live, 2125 max_old_gen_size, 2126 max_eden_size, 2127 true /* full gc*/, 2128 gc_cause, 2129 heap->collector_policy()); 2130 2131 size_policy->decay_supplemental_growth(true /* full gc*/); 2132 2133 heap->resize_old_gen( 2134 size_policy->calculated_old_free_size_in_bytes()); 2135 2136 // Don't resize the young generation at an major collection. A 2137 // desired young generation size may have been calculated but 2138 // resizing the young generation complicates the code because the 2139 // resizing of the old generation may have moved the boundary 2140 // between the young generation and the old generation. Let the 2141 // young generation resizing happen at the minor collections. 2142 } 2143 if (PrintAdaptiveSizePolicy) { 2144 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2145 heap->total_collections()); 2146 } 2147 } 2148 2149 if (UsePerfData) { 2150 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2151 counters->update_counters(); 2152 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2153 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2154 } 2155 2156 heap->resize_all_tlabs(); 2157 2158 // Resize the metaspace capactiy after a collection 2159 MetaspaceGC::compute_new_size(); 2160 2161 if (TraceGen1Time) accumulated_time()->stop(); 2162 2163 if (PrintGC) { 2164 if (PrintGCDetails) { 2165 // No GC timestamp here. This is after GC so it would be confusing. 2166 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2167 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2168 heap->print_heap_change(pre_gc_values.heap_used()); 2169 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2170 } else { 2171 heap->print_heap_change(pre_gc_values.heap_used()); 2172 } 2173 } 2174 2175 // Track memory usage and detect low memory 2176 MemoryService::track_memory_usage(); 2177 heap->update_counters(); 2178 gc_task_manager()->release_idle_workers(); 2179 } 2180 2181 #ifdef ASSERT 2182 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2183 ParCompactionManager* const cm = 2184 ParCompactionManager::manager_array(int(i)); 2185 assert(cm->marking_stack()->is_empty(), "should be empty"); 2186 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2187 } 2188 #endif // ASSERT 2189 2190 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2191 HandleMark hm; // Discard invalid handles created during verification 2192 Universe::verify(" VerifyAfterGC:"); 2193 } 2194 2195 // Re-verify object start arrays 2196 if (VerifyObjectStartArray && 2197 VerifyAfterGC) { 2198 old_gen->verify_object_start_array(); 2199 } 2200 2201 if (ZapUnusedHeapArea) { 2202 old_gen->object_space()->check_mangled_unused_area_complete(); 2203 } 2204 2205 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2206 2207 collection_exit.update(); 2208 2209 heap->print_heap_after_gc(); 2210 heap->trace_heap_after_gc(&_gc_tracer); 2211 2212 if (PrintGCTaskTimeStamps) { 2213 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2214 INT64_FORMAT, 2215 marking_start.ticks(), compaction_start.ticks(), 2216 collection_exit.ticks()); 2217 gc_task_manager()->print_task_time_stamps(); 2218 } 2219 2220 heap->post_full_gc_dump(&_gc_timer); 2221 2222 #ifdef TRACESPINNING 2223 ParallelTaskTerminator::print_termination_counts(); 2224 #endif 2225 2226 _gc_timer.register_gc_end(os::elapsed_counter()); 2227 2228 _gc_tracer.report_dense_prefix(dense_prefix(old_space_id)); 2229 _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions()); 2230 2231 return true; 2232 } 2233 2234 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2235 PSYoungGen* young_gen, 2236 PSOldGen* old_gen) { 2237 MutableSpace* const eden_space = young_gen->eden_space(); 2238 assert(!eden_space->is_empty(), "eden must be non-empty"); 2239 assert(young_gen->virtual_space()->alignment() == 2240 old_gen->virtual_space()->alignment(), "alignments do not match"); 2241 2242 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2243 return false; 2244 } 2245 2246 // Both generations must be completely committed. 2247 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2248 return false; 2249 } 2250 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2251 return false; 2252 } 2253 2254 // Figure out how much to take from eden. Include the average amount promoted 2255 // in the total; otherwise the next young gen GC will simply bail out to a 2256 // full GC. 2257 const size_t alignment = old_gen->virtual_space()->alignment(); 2258 const size_t eden_used = eden_space->used_in_bytes(); 2259 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2260 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2261 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2262 2263 if (absorb_size >= eden_capacity) { 2264 return false; // Must leave some space in eden. 2265 } 2266 2267 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2268 if (new_young_size < young_gen->min_gen_size()) { 2269 return false; // Respect young gen minimum size. 2270 } 2271 2272 if (TraceAdaptiveGCBoundary && Verbose) { 2273 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2274 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2275 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2276 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2277 absorb_size / K, 2278 eden_capacity / K, (eden_capacity - absorb_size) / K, 2279 young_gen->from_space()->used_in_bytes() / K, 2280 young_gen->to_space()->used_in_bytes() / K, 2281 young_gen->capacity_in_bytes() / K, new_young_size / K); 2282 } 2283 2284 // Fill the unused part of the old gen. 2285 MutableSpace* const old_space = old_gen->object_space(); 2286 HeapWord* const unused_start = old_space->top(); 2287 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2288 2289 if (unused_words > 0) { 2290 if (unused_words < CollectedHeap::min_fill_size()) { 2291 return false; // If the old gen cannot be filled, must give up. 2292 } 2293 CollectedHeap::fill_with_objects(unused_start, unused_words); 2294 } 2295 2296 // Take the live data from eden and set both top and end in the old gen to 2297 // eden top. (Need to set end because reset_after_change() mangles the region 2298 // from end to virtual_space->high() in debug builds). 2299 HeapWord* const new_top = eden_space->top(); 2300 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2301 absorb_size); 2302 young_gen->reset_after_change(); 2303 old_space->set_top(new_top); 2304 old_space->set_end(new_top); 2305 old_gen->reset_after_change(); 2306 2307 // Update the object start array for the filler object and the data from eden. 2308 ObjectStartArray* const start_array = old_gen->start_array(); 2309 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2310 start_array->allocate_block(p); 2311 } 2312 2313 // Could update the promoted average here, but it is not typically updated at 2314 // full GCs and the value to use is unclear. Something like 2315 // 2316 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2317 2318 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2319 return true; 2320 } 2321 2322 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2323 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2324 "shouldn't return NULL"); 2325 return ParallelScavengeHeap::gc_task_manager(); 2326 } 2327 2328 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2329 bool maximum_heap_compaction, 2330 ParallelOldTracer *gc_tracer) { 2331 // Recursively traverse all live objects and mark them 2332 GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer); 2333 2334 ParallelScavengeHeap* heap = gc_heap(); 2335 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2336 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2337 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2338 ParallelTaskTerminator terminator(active_gc_threads, qset); 2339 2340 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2341 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2342 2343 // Need new claim bits before marking starts. 2344 ClassLoaderDataGraph::clear_claimed_marks(); 2345 2346 { 2347 GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer); 2348 2349 ParallelScavengeHeap::ParStrongRootsScope psrs; 2350 2351 GCTaskQueue* q = GCTaskQueue::create(); 2352 2353 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2354 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2355 // We scan the thread roots in parallel 2356 Threads::create_thread_roots_marking_tasks(q); 2357 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2358 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2359 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2360 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2361 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data)); 2362 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2363 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2364 2365 if (active_gc_threads > 1) { 2366 for (uint j = 0; j < active_gc_threads; j++) { 2367 q->enqueue(new StealMarkingTask(&terminator)); 2368 } 2369 } 2370 2371 gc_task_manager()->execute_and_wait(q); 2372 } 2373 2374 // Process reference objects found during marking 2375 { 2376 GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer); 2377 2378 ReferenceProcessorStats stats; 2379 if (ref_processor()->processing_is_mt()) { 2380 RefProcTaskExecutor task_executor; 2381 stats = ref_processor()->process_discovered_references( 2382 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2383 &task_executor, &_gc_timer); 2384 } else { 2385 stats = ref_processor()->process_discovered_references( 2386 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL, 2387 &_gc_timer); 2388 } 2389 2390 gc_tracer->report_gc_reference_stats(stats); 2391 } 2392 2393 GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer); 2394 2395 // This is the point where the entire marking should have completed. 2396 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2397 2398 // Follow system dictionary roots and unload classes. 2399 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2400 2401 // Unload nmethods. 2402 CodeCache::do_unloading(is_alive_closure(), purged_class); 2403 2404 // Prune dead klasses from subklass/sibling/implementor lists. 2405 Klass::clean_weak_klass_links(is_alive_closure()); 2406 2407 // Delete entries for dead interned strings. 2408 StringTable::unlink(is_alive_closure()); 2409 2410 // Clean up unreferenced symbols in symbol table. 2411 SymbolTable::unlink(); 2412 _gc_tracer.report_object_count_after_gc(is_alive_closure()); 2413 } 2414 2415 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) { 2416 ClassLoaderData* cld = klass->class_loader_data(); 2417 // The actual processing of the klass is done when we 2418 // traverse the list of Klasses in the class loader data. 2419 PSParallelCompact::follow_class_loader(cm, cld); 2420 } 2421 2422 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) { 2423 ClassLoaderData* cld = klass->class_loader_data(); 2424 // The actual processing of the klass is done when we 2425 // traverse the list of Klasses in the class loader data. 2426 PSParallelCompact::adjust_class_loader(cm, cld); 2427 } 2428 2429 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2430 ClassLoaderData* cld) { 2431 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2432 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2433 2434 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2435 } 2436 2437 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm, 2438 ClassLoaderData* cld) { 2439 cld->oops_do(PSParallelCompact::adjust_pointer_closure(), 2440 PSParallelCompact::adjust_klass_closure(), 2441 true); 2442 } 2443 2444 // This should be moved to the shared markSweep code! 2445 class PSAlwaysTrueClosure: public BoolObjectClosure { 2446 public: 2447 bool do_object_b(oop p) { return true; } 2448 }; 2449 static PSAlwaysTrueClosure always_true; 2450 2451 void PSParallelCompact::adjust_roots() { 2452 // Adjust the pointers to reflect the new locations 2453 GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer); 2454 2455 // Need new claim bits when tracing through and adjusting pointers. 2456 ClassLoaderDataGraph::clear_claimed_marks(); 2457 2458 // General strong roots. 2459 Universe::oops_do(adjust_pointer_closure()); 2460 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2461 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2462 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2463 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2464 FlatProfiler::oops_do(adjust_pointer_closure()); 2465 Management::oops_do(adjust_pointer_closure()); 2466 JvmtiExport::oops_do(adjust_pointer_closure()); 2467 // SO_AllClasses 2468 SystemDictionary::oops_do(adjust_pointer_closure()); 2469 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2470 2471 // Now adjust pointers in remaining weak roots. (All of which should 2472 // have been cleared if they pointed to non-surviving objects.) 2473 // Global (weak) JNI handles 2474 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2475 2476 CodeCache::oops_do(adjust_pointer_closure()); 2477 StringTable::oops_do(adjust_pointer_closure()); 2478 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2479 // Roots were visited so references into the young gen in roots 2480 // may have been scanned. Process them also. 2481 // Should the reference processor have a span that excludes 2482 // young gen objects? 2483 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2484 } 2485 2486 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2487 uint parallel_gc_threads) 2488 { 2489 GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer); 2490 2491 // Find the threads that are active 2492 unsigned int which = 0; 2493 2494 const uint task_count = MAX2(parallel_gc_threads, 1U); 2495 for (uint j = 0; j < task_count; j++) { 2496 q->enqueue(new DrainStacksCompactionTask(j)); 2497 ParCompactionManager::verify_region_list_empty(j); 2498 // Set the region stacks variables to "no" region stack values 2499 // so that they will be recognized and needing a region stack 2500 // in the stealing tasks if they do not get one by executing 2501 // a draining stack. 2502 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2503 cm->set_region_stack(NULL); 2504 cm->set_region_stack_index((uint)max_uintx); 2505 } 2506 ParCompactionManager::reset_recycled_stack_index(); 2507 2508 // Find all regions that are available (can be filled immediately) and 2509 // distribute them to the thread stacks. The iteration is done in reverse 2510 // order (high to low) so the regions will be removed in ascending order. 2511 2512 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2513 2514 size_t fillable_regions = 0; // A count for diagnostic purposes. 2515 // A region index which corresponds to the tasks created above. 2516 // "which" must be 0 <= which < task_count 2517 2518 which = 0; 2519 // id + 1 is used to test termination so unsigned can 2520 // be used with an old_space_id == 0. 2521 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2522 SpaceInfo* const space_info = _space_info + id; 2523 MutableSpace* const space = space_info->space(); 2524 HeapWord* const new_top = space_info->new_top(); 2525 2526 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2527 const size_t end_region = 2528 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2529 2530 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2531 if (sd.region(cur)->claim_unsafe()) { 2532 ParCompactionManager::region_list_push(which, cur); 2533 2534 if (TraceParallelOldGCCompactionPhase && Verbose) { 2535 const size_t count_mod_8 = fillable_regions & 7; 2536 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2537 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2538 if (count_mod_8 == 7) gclog_or_tty->cr(); 2539 } 2540 2541 NOT_PRODUCT(++fillable_regions;) 2542 2543 // Assign regions to tasks in round-robin fashion. 2544 if (++which == task_count) { 2545 assert(which <= parallel_gc_threads, 2546 "Inconsistent number of workers"); 2547 which = 0; 2548 } 2549 } 2550 } 2551 } 2552 2553 if (TraceParallelOldGCCompactionPhase) { 2554 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2555 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2556 } 2557 } 2558 2559 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2560 2561 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2562 uint parallel_gc_threads) { 2563 GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer); 2564 2565 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2566 2567 // Iterate over all the spaces adding tasks for updating 2568 // regions in the dense prefix. Assume that 1 gc thread 2569 // will work on opening the gaps and the remaining gc threads 2570 // will work on the dense prefix. 2571 unsigned int space_id; 2572 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2573 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2574 const MutableSpace* const space = _space_info[space_id].space(); 2575 2576 if (dense_prefix_end == space->bottom()) { 2577 // There is no dense prefix for this space. 2578 continue; 2579 } 2580 2581 // The dense prefix is before this region. 2582 size_t region_index_end_dense_prefix = 2583 sd.addr_to_region_idx(dense_prefix_end); 2584 RegionData* const dense_prefix_cp = 2585 sd.region(region_index_end_dense_prefix); 2586 assert(dense_prefix_end == space->end() || 2587 dense_prefix_cp->available() || 2588 dense_prefix_cp->claimed(), 2589 "The region after the dense prefix should always be ready to fill"); 2590 2591 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2592 2593 // Is there dense prefix work? 2594 size_t total_dense_prefix_regions = 2595 region_index_end_dense_prefix - region_index_start; 2596 // How many regions of the dense prefix should be given to 2597 // each thread? 2598 if (total_dense_prefix_regions > 0) { 2599 uint tasks_for_dense_prefix = 1; 2600 if (total_dense_prefix_regions <= 2601 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2602 // Don't over partition. This assumes that 2603 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2604 // so there are not many regions to process. 2605 tasks_for_dense_prefix = parallel_gc_threads; 2606 } else { 2607 // Over partition 2608 tasks_for_dense_prefix = parallel_gc_threads * 2609 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2610 } 2611 size_t regions_per_thread = total_dense_prefix_regions / 2612 tasks_for_dense_prefix; 2613 // Give each thread at least 1 region. 2614 if (regions_per_thread == 0) { 2615 regions_per_thread = 1; 2616 } 2617 2618 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2619 if (region_index_start >= region_index_end_dense_prefix) { 2620 break; 2621 } 2622 // region_index_end is not processed 2623 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2624 region_index_end_dense_prefix); 2625 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2626 region_index_start, 2627 region_index_end)); 2628 region_index_start = region_index_end; 2629 } 2630 } 2631 // This gets any part of the dense prefix that did not 2632 // fit evenly. 2633 if (region_index_start < region_index_end_dense_prefix) { 2634 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2635 region_index_start, 2636 region_index_end_dense_prefix)); 2637 } 2638 } 2639 } 2640 2641 void PSParallelCompact::enqueue_region_stealing_tasks( 2642 GCTaskQueue* q, 2643 ParallelTaskTerminator* terminator_ptr, 2644 uint parallel_gc_threads) { 2645 GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer); 2646 2647 // Once a thread has drained it's stack, it should try to steal regions from 2648 // other threads. 2649 if (parallel_gc_threads > 1) { 2650 for (uint j = 0; j < parallel_gc_threads; j++) { 2651 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2652 } 2653 } 2654 } 2655 2656 void PSParallelCompact::compact() { 2657 // trace("5"); 2658 GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer); 2659 2660 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2661 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2662 PSOldGen* old_gen = heap->old_gen(); 2663 old_gen->start_array()->reset(); 2664 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2665 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2666 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2667 ParallelTaskTerminator terminator(active_gc_threads, qset); 2668 2669 GCTaskQueue* q = GCTaskQueue::create(); 2670 enqueue_region_draining_tasks(q, active_gc_threads); 2671 enqueue_dense_prefix_tasks(q, active_gc_threads); 2672 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2673 2674 { 2675 GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer); 2676 2677 gc_task_manager()->execute_and_wait(q); 2678 2679 #ifdef ASSERT 2680 // Verify that all regions have been processed before the deferred updates. 2681 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2682 verify_complete(SpaceId(id)); 2683 } 2684 #endif 2685 } 2686 2687 { 2688 // Update the deferred objects, if any. Any compaction manager can be used. 2689 GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer); 2690 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2691 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2692 update_deferred_objects(cm, SpaceId(id)); 2693 } 2694 } 2695 } 2696 2697 #ifdef ASSERT 2698 void PSParallelCompact::verify_complete(SpaceId space_id) { 2699 // All Regions between space bottom() to new_top() should be marked as filled 2700 // and all Regions between new_top() and top() should be available (i.e., 2701 // should have been emptied). 2702 ParallelCompactData& sd = summary_data(); 2703 SpaceInfo si = _space_info[space_id]; 2704 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2705 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2706 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2707 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2708 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2709 2710 bool issued_a_warning = false; 2711 2712 size_t cur_region; 2713 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2714 const RegionData* const c = sd.region(cur_region); 2715 if (!c->completed()) { 2716 warning("region " SIZE_FORMAT " not filled: " 2717 "destination_count=" SIZE_FORMAT, 2718 cur_region, c->destination_count()); 2719 issued_a_warning = true; 2720 } 2721 } 2722 2723 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2724 const RegionData* const c = sd.region(cur_region); 2725 if (!c->available()) { 2726 warning("region " SIZE_FORMAT " not empty: " 2727 "destination_count=" SIZE_FORMAT, 2728 cur_region, c->destination_count()); 2729 issued_a_warning = true; 2730 } 2731 } 2732 2733 if (issued_a_warning) { 2734 print_region_ranges(); 2735 } 2736 } 2737 #endif // #ifdef ASSERT 2738 2739 // Update interior oops in the ranges of regions [beg_region, end_region). 2740 void 2741 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2742 SpaceId space_id, 2743 size_t beg_region, 2744 size_t end_region) { 2745 ParallelCompactData& sd = summary_data(); 2746 ParMarkBitMap* const mbm = mark_bitmap(); 2747 2748 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2749 HeapWord* const end_addr = sd.region_to_addr(end_region); 2750 assert(beg_region <= end_region, "bad region range"); 2751 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2752 2753 #ifdef ASSERT 2754 // Claim the regions to avoid triggering an assert when they are marked as 2755 // filled. 2756 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2757 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2758 } 2759 #endif // #ifdef ASSERT 2760 2761 if (beg_addr != space(space_id)->bottom()) { 2762 // Find the first live object or block of dead space that *starts* in this 2763 // range of regions. If a partial object crosses onto the region, skip it; 2764 // it will be marked for 'deferred update' when the object head is 2765 // processed. If dead space crosses onto the region, it is also skipped; it 2766 // will be filled when the prior region is processed. If neither of those 2767 // apply, the first word in the region is the start of a live object or dead 2768 // space. 2769 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2770 const RegionData* const cp = sd.region(beg_region); 2771 if (cp->partial_obj_size() != 0) { 2772 beg_addr = sd.partial_obj_end(beg_region); 2773 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2774 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2775 } 2776 } 2777 2778 if (beg_addr < end_addr) { 2779 // A live object or block of dead space starts in this range of Regions. 2780 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2781 2782 // Create closures and iterate. 2783 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2784 FillClosure fill_closure(cm, space_id); 2785 ParMarkBitMap::IterationStatus status; 2786 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2787 dense_prefix_end); 2788 if (status == ParMarkBitMap::incomplete) { 2789 update_closure.do_addr(update_closure.source()); 2790 } 2791 } 2792 2793 // Mark the regions as filled. 2794 RegionData* const beg_cp = sd.region(beg_region); 2795 RegionData* const end_cp = sd.region(end_region); 2796 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2797 cp->set_completed(); 2798 } 2799 } 2800 2801 // Return the SpaceId for the space containing addr. If addr is not in the 2802 // heap, last_space_id is returned. In debug mode it expects the address to be 2803 // in the heap and asserts such. 2804 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2805 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2806 2807 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2808 if (_space_info[id].space()->contains(addr)) { 2809 return SpaceId(id); 2810 } 2811 } 2812 2813 assert(false, "no space contains the addr"); 2814 return last_space_id; 2815 } 2816 2817 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2818 SpaceId id) { 2819 assert(id < last_space_id, "bad space id"); 2820 2821 ParallelCompactData& sd = summary_data(); 2822 const SpaceInfo* const space_info = _space_info + id; 2823 ObjectStartArray* const start_array = space_info->start_array(); 2824 2825 const MutableSpace* const space = space_info->space(); 2826 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2827 HeapWord* const beg_addr = space_info->dense_prefix(); 2828 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2829 2830 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2831 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2832 const RegionData* cur_region; 2833 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2834 HeapWord* const addr = cur_region->deferred_obj_addr(); 2835 if (addr != NULL) { 2836 if (start_array != NULL) { 2837 start_array->allocate_block(addr); 2838 } 2839 oop(addr)->update_contents(cm); 2840 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 2841 } 2842 } 2843 } 2844 2845 // Skip over count live words starting from beg, and return the address of the 2846 // next live word. Unless marked, the word corresponding to beg is assumed to 2847 // be dead. Callers must either ensure beg does not correspond to the middle of 2848 // an object, or account for those live words in some other way. Callers must 2849 // also ensure that there are enough live words in the range [beg, end) to skip. 2850 HeapWord* 2851 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2852 { 2853 assert(count > 0, "sanity"); 2854 2855 ParMarkBitMap* m = mark_bitmap(); 2856 idx_t bits_to_skip = m->words_to_bits(count); 2857 idx_t cur_beg = m->addr_to_bit(beg); 2858 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2859 2860 do { 2861 cur_beg = m->find_obj_beg(cur_beg, search_end); 2862 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2863 const size_t obj_bits = cur_end - cur_beg + 1; 2864 if (obj_bits > bits_to_skip) { 2865 return m->bit_to_addr(cur_beg + bits_to_skip); 2866 } 2867 bits_to_skip -= obj_bits; 2868 cur_beg = cur_end + 1; 2869 } while (bits_to_skip > 0); 2870 2871 // Skipping the desired number of words landed just past the end of an object. 2872 // Find the start of the next object. 2873 cur_beg = m->find_obj_beg(cur_beg, search_end); 2874 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2875 return m->bit_to_addr(cur_beg); 2876 } 2877 2878 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2879 SpaceId src_space_id, 2880 size_t src_region_idx) 2881 { 2882 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2883 2884 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2885 if (split_info.dest_region_addr() == dest_addr) { 2886 // The partial object ending at the split point contains the first word to 2887 // be copied to dest_addr. 2888 return split_info.first_src_addr(); 2889 } 2890 2891 const ParallelCompactData& sd = summary_data(); 2892 ParMarkBitMap* const bitmap = mark_bitmap(); 2893 const size_t RegionSize = ParallelCompactData::RegionSize; 2894 2895 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2896 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2897 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2898 HeapWord* const src_region_destination = src_region_ptr->destination(); 2899 2900 assert(dest_addr >= src_region_destination, "wrong src region"); 2901 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2902 2903 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2904 HeapWord* const src_region_end = src_region_beg + RegionSize; 2905 2906 HeapWord* addr = src_region_beg; 2907 if (dest_addr == src_region_destination) { 2908 // Return the first live word in the source region. 2909 if (partial_obj_size == 0) { 2910 addr = bitmap->find_obj_beg(addr, src_region_end); 2911 assert(addr < src_region_end, "no objects start in src region"); 2912 } 2913 return addr; 2914 } 2915 2916 // Must skip some live data. 2917 size_t words_to_skip = dest_addr - src_region_destination; 2918 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2919 2920 if (partial_obj_size >= words_to_skip) { 2921 // All the live words to skip are part of the partial object. 2922 addr += words_to_skip; 2923 if (partial_obj_size == words_to_skip) { 2924 // Find the first live word past the partial object. 2925 addr = bitmap->find_obj_beg(addr, src_region_end); 2926 assert(addr < src_region_end, "wrong src region"); 2927 } 2928 return addr; 2929 } 2930 2931 // Skip over the partial object (if any). 2932 if (partial_obj_size != 0) { 2933 words_to_skip -= partial_obj_size; 2934 addr += partial_obj_size; 2935 } 2936 2937 // Skip over live words due to objects that start in the region. 2938 addr = skip_live_words(addr, src_region_end, words_to_skip); 2939 assert(addr < src_region_end, "wrong src region"); 2940 return addr; 2941 } 2942 2943 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2944 SpaceId src_space_id, 2945 size_t beg_region, 2946 HeapWord* end_addr) 2947 { 2948 ParallelCompactData& sd = summary_data(); 2949 2950 #ifdef ASSERT 2951 MutableSpace* const src_space = _space_info[src_space_id].space(); 2952 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2953 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2954 "src_space_id does not match beg_addr"); 2955 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2956 "src_space_id does not match end_addr"); 2957 #endif // #ifdef ASSERT 2958 2959 RegionData* const beg = sd.region(beg_region); 2960 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2961 2962 // Regions up to new_top() are enqueued if they become available. 2963 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2964 RegionData* const enqueue_end = 2965 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2966 2967 for (RegionData* cur = beg; cur < end; ++cur) { 2968 assert(cur->data_size() > 0, "region must have live data"); 2969 cur->decrement_destination_count(); 2970 if (cur < enqueue_end && cur->available() && cur->claim()) { 2971 cm->push_region(sd.region(cur)); 2972 } 2973 } 2974 } 2975 2976 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2977 SpaceId& src_space_id, 2978 HeapWord*& src_space_top, 2979 HeapWord* end_addr) 2980 { 2981 typedef ParallelCompactData::RegionData RegionData; 2982 2983 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2984 const size_t region_size = ParallelCompactData::RegionSize; 2985 2986 size_t src_region_idx = 0; 2987 2988 // Skip empty regions (if any) up to the top of the space. 2989 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2990 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2991 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2992 const RegionData* const top_region_ptr = 2993 sd.addr_to_region_ptr(top_aligned_up); 2994 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2995 ++src_region_ptr; 2996 } 2997 2998 if (src_region_ptr < top_region_ptr) { 2999 // The next source region is in the current space. Update src_region_idx 3000 // and the source address to match src_region_ptr. 3001 src_region_idx = sd.region(src_region_ptr); 3002 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3003 if (src_region_addr > closure.source()) { 3004 closure.set_source(src_region_addr); 3005 } 3006 return src_region_idx; 3007 } 3008 3009 // Switch to a new source space and find the first non-empty region. 3010 unsigned int space_id = src_space_id + 1; 3011 assert(space_id < last_space_id, "not enough spaces"); 3012 3013 HeapWord* const destination = closure.destination(); 3014 3015 do { 3016 MutableSpace* space = _space_info[space_id].space(); 3017 HeapWord* const bottom = space->bottom(); 3018 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3019 3020 // Iterate over the spaces that do not compact into themselves. 3021 if (bottom_cp->destination() != bottom) { 3022 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3023 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3024 3025 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3026 if (src_cp->live_obj_size() > 0) { 3027 // Found it. 3028 assert(src_cp->destination() == destination, 3029 "first live obj in the space must match the destination"); 3030 assert(src_cp->partial_obj_size() == 0, 3031 "a space cannot begin with a partial obj"); 3032 3033 src_space_id = SpaceId(space_id); 3034 src_space_top = space->top(); 3035 const size_t src_region_idx = sd.region(src_cp); 3036 closure.set_source(sd.region_to_addr(src_region_idx)); 3037 return src_region_idx; 3038 } else { 3039 assert(src_cp->data_size() == 0, "sanity"); 3040 } 3041 } 3042 } 3043 } while (++space_id < last_space_id); 3044 3045 assert(false, "no source region was found"); 3046 return 0; 3047 } 3048 3049 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3050 { 3051 typedef ParMarkBitMap::IterationStatus IterationStatus; 3052 const size_t RegionSize = ParallelCompactData::RegionSize; 3053 ParMarkBitMap* const bitmap = mark_bitmap(); 3054 ParallelCompactData& sd = summary_data(); 3055 RegionData* const region_ptr = sd.region(region_idx); 3056 3057 // Get the items needed to construct the closure. 3058 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3059 SpaceId dest_space_id = space_id(dest_addr); 3060 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3061 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3062 assert(dest_addr < new_top, "sanity"); 3063 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3064 3065 // Get the source region and related info. 3066 size_t src_region_idx = region_ptr->source_region(); 3067 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3068 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3069 3070 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3071 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3072 3073 // Adjust src_region_idx to prepare for decrementing destination counts (the 3074 // destination count is not decremented when a region is copied to itself). 3075 if (src_region_idx == region_idx) { 3076 src_region_idx += 1; 3077 } 3078 3079 if (bitmap->is_unmarked(closure.source())) { 3080 // The first source word is in the middle of an object; copy the remainder 3081 // of the object or as much as will fit. The fact that pointer updates were 3082 // deferred will be noted when the object header is processed. 3083 HeapWord* const old_src_addr = closure.source(); 3084 closure.copy_partial_obj(); 3085 if (closure.is_full()) { 3086 decrement_destination_counts(cm, src_space_id, src_region_idx, 3087 closure.source()); 3088 region_ptr->set_deferred_obj_addr(NULL); 3089 region_ptr->set_completed(); 3090 return; 3091 } 3092 3093 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3094 if (sd.region_align_down(old_src_addr) != end_addr) { 3095 // The partial object was copied from more than one source region. 3096 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3097 3098 // Move to the next source region, possibly switching spaces as well. All 3099 // args except end_addr may be modified. 3100 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3101 end_addr); 3102 } 3103 } 3104 3105 do { 3106 HeapWord* const cur_addr = closure.source(); 3107 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3108 src_space_top); 3109 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3110 3111 if (status == ParMarkBitMap::incomplete) { 3112 // The last obj that starts in the source region does not end in the 3113 // region. 3114 assert(closure.source() < end_addr, "sanity"); 3115 HeapWord* const obj_beg = closure.source(); 3116 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3117 src_space_top); 3118 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3119 if (obj_end < range_end) { 3120 // The end was found; the entire object will fit. 3121 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3122 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3123 } else { 3124 // The end was not found; the object will not fit. 3125 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3126 status = ParMarkBitMap::would_overflow; 3127 } 3128 } 3129 3130 if (status == ParMarkBitMap::would_overflow) { 3131 // The last object did not fit. Note that interior oop updates were 3132 // deferred, then copy enough of the object to fill the region. 3133 region_ptr->set_deferred_obj_addr(closure.destination()); 3134 status = closure.copy_until_full(); // copies from closure.source() 3135 3136 decrement_destination_counts(cm, src_space_id, src_region_idx, 3137 closure.source()); 3138 region_ptr->set_completed(); 3139 return; 3140 } 3141 3142 if (status == ParMarkBitMap::full) { 3143 decrement_destination_counts(cm, src_space_id, src_region_idx, 3144 closure.source()); 3145 region_ptr->set_deferred_obj_addr(NULL); 3146 region_ptr->set_completed(); 3147 return; 3148 } 3149 3150 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3151 3152 // Move to the next source region, possibly switching spaces as well. All 3153 // args except end_addr may be modified. 3154 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3155 end_addr); 3156 } while (true); 3157 } 3158 3159 void 3160 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3161 const MutableSpace* sp = space(space_id); 3162 if (sp->is_empty()) { 3163 return; 3164 } 3165 3166 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3167 ParMarkBitMap* const bitmap = mark_bitmap(); 3168 HeapWord* const dp_addr = dense_prefix(space_id); 3169 HeapWord* beg_addr = sp->bottom(); 3170 HeapWord* end_addr = sp->top(); 3171 3172 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3173 3174 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3175 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3176 if (beg_region < dp_region) { 3177 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3178 } 3179 3180 // The destination of the first live object that starts in the region is one 3181 // past the end of the partial object entering the region (if any). 3182 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3183 HeapWord* const new_top = _space_info[space_id].new_top(); 3184 assert(new_top >= dest_addr, "bad new_top value"); 3185 const size_t words = pointer_delta(new_top, dest_addr); 3186 3187 if (words > 0) { 3188 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3189 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3190 3191 ParMarkBitMap::IterationStatus status; 3192 status = bitmap->iterate(&closure, dest_addr, end_addr); 3193 assert(status == ParMarkBitMap::full, "iteration not complete"); 3194 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3195 "live objects skipped because closure is full"); 3196 } 3197 } 3198 3199 jlong PSParallelCompact::millis_since_last_gc() { 3200 // We need a monotonically non-deccreasing time in ms but 3201 // os::javaTimeMillis() does not guarantee monotonicity. 3202 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3203 jlong ret_val = now - _time_of_last_gc; 3204 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3205 if (ret_val < 0) { 3206 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);) 3207 return 0; 3208 } 3209 return ret_val; 3210 } 3211 3212 void PSParallelCompact::reset_millis_since_last_gc() { 3213 // We need a monotonically non-deccreasing time in ms but 3214 // os::javaTimeMillis() does not guarantee monotonicity. 3215 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3216 } 3217 3218 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3219 { 3220 if (source() != destination()) { 3221 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3222 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3223 } 3224 update_state(words_remaining()); 3225 assert(is_full(), "sanity"); 3226 return ParMarkBitMap::full; 3227 } 3228 3229 void MoveAndUpdateClosure::copy_partial_obj() 3230 { 3231 size_t words = words_remaining(); 3232 3233 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3234 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3235 if (end_addr < range_end) { 3236 words = bitmap()->obj_size(source(), end_addr); 3237 } 3238 3239 // This test is necessary; if omitted, the pointer updates to a partial object 3240 // that crosses the dense prefix boundary could be overwritten. 3241 if (source() != destination()) { 3242 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3243 Copy::aligned_conjoint_words(source(), destination(), words); 3244 } 3245 update_state(words); 3246 } 3247 3248 ParMarkBitMapClosure::IterationStatus 3249 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3250 assert(destination() != NULL, "sanity"); 3251 assert(bitmap()->obj_size(addr) == words, "bad size"); 3252 3253 _source = addr; 3254 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3255 destination(), "wrong destination"); 3256 3257 if (words > words_remaining()) { 3258 return ParMarkBitMap::would_overflow; 3259 } 3260 3261 // The start_array must be updated even if the object is not moving. 3262 if (_start_array != NULL) { 3263 _start_array->allocate_block(destination()); 3264 } 3265 3266 if (destination() != source()) { 3267 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3268 Copy::aligned_conjoint_words(source(), destination(), words); 3269 } 3270 3271 oop moved_oop = (oop) destination(); 3272 moved_oop->update_contents(compaction_manager()); 3273 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3274 3275 update_state(words); 3276 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3277 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3278 } 3279 3280 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3281 ParCompactionManager* cm, 3282 PSParallelCompact::SpaceId space_id) : 3283 ParMarkBitMapClosure(mbm, cm), 3284 _space_id(space_id), 3285 _start_array(PSParallelCompact::start_array(space_id)) 3286 { 3287 } 3288 3289 // Updates the references in the object to their new values. 3290 ParMarkBitMapClosure::IterationStatus 3291 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3292 do_addr(addr); 3293 return ParMarkBitMap::incomplete; 3294 }