1 /* 2 * Copyright (c) 2005, 2012, 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/isGCActiveMark.hpp" 43 #include "gc_interface/gcCause.hpp" 44 #include "memory/gcLocker.inline.hpp" 45 #include "memory/referencePolicy.hpp" 46 #include "memory/referenceProcessor.hpp" 47 #include "oops/methodData.hpp" 48 #include "oops/oop.inline.hpp" 49 #include "oops/oop.pcgc.inline.hpp" 50 #include "runtime/fprofiler.hpp" 51 #include "runtime/safepoint.hpp" 52 #include "runtime/vmThread.hpp" 53 #include "services/management.hpp" 54 #include "services/memoryService.hpp" 55 #include "services/memTracker.hpp" 56 #include "utilities/events.hpp" 57 #include "utilities/stack.inline.hpp" 58 59 #include <math.h> 60 61 // All sizes are in HeapWords. 62 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words 63 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 64 const size_t ParallelCompactData::RegionSizeBytes = 65 RegionSize << LogHeapWordSize; 66 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 67 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 68 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 69 70 const ParallelCompactData::RegionData::region_sz_t 71 ParallelCompactData::RegionData::dc_shift = 27; 72 73 const ParallelCompactData::RegionData::region_sz_t 74 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 75 76 const ParallelCompactData::RegionData::region_sz_t 77 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 78 79 const ParallelCompactData::RegionData::region_sz_t 80 ParallelCompactData::RegionData::los_mask = ~dc_mask; 81 82 const ParallelCompactData::RegionData::region_sz_t 83 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 84 85 const ParallelCompactData::RegionData::region_sz_t 86 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 87 88 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 89 bool PSParallelCompact::_print_phases = false; 90 91 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 92 Klass* PSParallelCompact::_updated_int_array_klass_obj = NULL; 93 94 double PSParallelCompact::_dwl_mean; 95 double PSParallelCompact::_dwl_std_dev; 96 double PSParallelCompact::_dwl_first_term; 97 double PSParallelCompact::_dwl_adjustment; 98 #ifdef ASSERT 99 bool PSParallelCompact::_dwl_initialized = false; 100 #endif // #ifdef ASSERT 101 102 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 103 HeapWord* destination) 104 { 105 assert(src_region_idx != 0, "invalid src_region_idx"); 106 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 107 assert(destination != NULL, "invalid destination argument"); 108 109 _src_region_idx = src_region_idx; 110 _partial_obj_size = partial_obj_size; 111 _destination = destination; 112 113 // These fields may not be updated below, so make sure they're clear. 114 assert(_dest_region_addr == NULL, "should have been cleared"); 115 assert(_first_src_addr == NULL, "should have been cleared"); 116 117 // Determine the number of destination regions for the partial object. 118 HeapWord* const last_word = destination + partial_obj_size - 1; 119 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 120 HeapWord* const beg_region_addr = sd.region_align_down(destination); 121 HeapWord* const end_region_addr = sd.region_align_down(last_word); 122 123 if (beg_region_addr == end_region_addr) { 124 // One destination region. 125 _destination_count = 1; 126 if (end_region_addr == destination) { 127 // The destination falls on a region boundary, thus the first word of the 128 // partial object will be the first word copied to the destination region. 129 _dest_region_addr = end_region_addr; 130 _first_src_addr = sd.region_to_addr(src_region_idx); 131 } 132 } else { 133 // Two destination regions. When copied, the partial object will cross a 134 // destination region boundary, so a word somewhere within the partial 135 // object will be the first word copied to the second destination region. 136 _destination_count = 2; 137 _dest_region_addr = end_region_addr; 138 const size_t ofs = pointer_delta(end_region_addr, destination); 139 assert(ofs < _partial_obj_size, "sanity"); 140 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 141 } 142 } 143 144 void SplitInfo::clear() 145 { 146 _src_region_idx = 0; 147 _partial_obj_size = 0; 148 _destination = NULL; 149 _destination_count = 0; 150 _dest_region_addr = NULL; 151 _first_src_addr = NULL; 152 assert(!is_valid(), "sanity"); 153 } 154 155 #ifdef ASSERT 156 void SplitInfo::verify_clear() 157 { 158 assert(_src_region_idx == 0, "not clear"); 159 assert(_partial_obj_size == 0, "not clear"); 160 assert(_destination == NULL, "not clear"); 161 assert(_destination_count == 0, "not clear"); 162 assert(_dest_region_addr == NULL, "not clear"); 163 assert(_first_src_addr == NULL, "not clear"); 164 } 165 #endif // #ifdef ASSERT 166 167 168 void PSParallelCompact::print_on_error(outputStream* st) { 169 _mark_bitmap.print_on_error(st); 170 } 171 172 #ifndef PRODUCT 173 const char* PSParallelCompact::space_names[] = { 174 "old ", "eden", "from", "to " 175 }; 176 177 void PSParallelCompact::print_region_ranges() 178 { 179 tty->print_cr("space bottom top end new_top"); 180 tty->print_cr("------ ---------- ---------- ---------- ----------"); 181 182 for (unsigned int id = 0; id < last_space_id; ++id) { 183 const MutableSpace* space = _space_info[id].space(); 184 tty->print_cr("%u %s " 185 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 186 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 187 id, space_names[id], 188 summary_data().addr_to_region_idx(space->bottom()), 189 summary_data().addr_to_region_idx(space->top()), 190 summary_data().addr_to_region_idx(space->end()), 191 summary_data().addr_to_region_idx(_space_info[id].new_top())); 192 } 193 } 194 195 void 196 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 197 { 198 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 199 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 200 201 ParallelCompactData& sd = PSParallelCompact::summary_data(); 202 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 203 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 204 REGION_IDX_FORMAT " " PTR_FORMAT " " 205 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 206 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 207 i, c->data_location(), dci, c->destination(), 208 c->partial_obj_size(), c->live_obj_size(), 209 c->data_size(), c->source_region(), c->destination_count()); 210 211 #undef REGION_IDX_FORMAT 212 #undef REGION_DATA_FORMAT 213 } 214 215 void 216 print_generic_summary_data(ParallelCompactData& summary_data, 217 HeapWord* const beg_addr, 218 HeapWord* const end_addr) 219 { 220 size_t total_words = 0; 221 size_t i = summary_data.addr_to_region_idx(beg_addr); 222 const size_t last = summary_data.addr_to_region_idx(end_addr); 223 HeapWord* pdest = 0; 224 225 while (i <= last) { 226 ParallelCompactData::RegionData* c = summary_data.region(i); 227 if (c->data_size() != 0 || c->destination() != pdest) { 228 print_generic_summary_region(i, c); 229 total_words += c->data_size(); 230 pdest = c->destination(); 231 } 232 ++i; 233 } 234 235 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 236 } 237 238 void 239 print_generic_summary_data(ParallelCompactData& summary_data, 240 SpaceInfo* space_info) 241 { 242 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 243 const MutableSpace* space = space_info[id].space(); 244 print_generic_summary_data(summary_data, space->bottom(), 245 MAX2(space->top(), space_info[id].new_top())); 246 } 247 } 248 249 void 250 print_initial_summary_region(size_t i, 251 const ParallelCompactData::RegionData* c, 252 bool newline = true) 253 { 254 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 255 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 256 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 257 i, c->destination(), 258 c->partial_obj_size(), c->live_obj_size(), 259 c->data_size(), c->source_region(), c->destination_count()); 260 if (newline) tty->cr(); 261 } 262 263 void 264 print_initial_summary_data(ParallelCompactData& summary_data, 265 const MutableSpace* space) { 266 if (space->top() == space->bottom()) { 267 return; 268 } 269 270 const size_t region_size = ParallelCompactData::RegionSize; 271 typedef ParallelCompactData::RegionData RegionData; 272 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 273 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 274 const RegionData* c = summary_data.region(end_region - 1); 275 HeapWord* end_addr = c->destination() + c->data_size(); 276 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 277 278 // Print (and count) the full regions at the beginning of the space. 279 size_t full_region_count = 0; 280 size_t i = summary_data.addr_to_region_idx(space->bottom()); 281 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 282 print_initial_summary_region(i, summary_data.region(i)); 283 ++full_region_count; 284 ++i; 285 } 286 287 size_t live_to_right = live_in_space - full_region_count * region_size; 288 289 double max_reclaimed_ratio = 0.0; 290 size_t max_reclaimed_ratio_region = 0; 291 size_t max_dead_to_right = 0; 292 size_t max_live_to_right = 0; 293 294 // Print the 'reclaimed ratio' for regions while there is something live in 295 // the region or to the right of it. The remaining regions are empty (and 296 // uninteresting), and computing the ratio will result in division by 0. 297 while (i < end_region && live_to_right > 0) { 298 c = summary_data.region(i); 299 HeapWord* const region_addr = summary_data.region_to_addr(i); 300 const size_t used_to_right = pointer_delta(space->top(), region_addr); 301 const size_t dead_to_right = used_to_right - live_to_right; 302 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 303 304 if (reclaimed_ratio > max_reclaimed_ratio) { 305 max_reclaimed_ratio = reclaimed_ratio; 306 max_reclaimed_ratio_region = i; 307 max_dead_to_right = dead_to_right; 308 max_live_to_right = live_to_right; 309 } 310 311 print_initial_summary_region(i, c, false); 312 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 313 reclaimed_ratio, dead_to_right, live_to_right); 314 315 live_to_right -= c->data_size(); 316 ++i; 317 } 318 319 // Any remaining regions are empty. Print one more if there is one. 320 if (i < end_region) { 321 print_initial_summary_region(i, summary_data.region(i)); 322 } 323 324 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 325 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 326 max_reclaimed_ratio_region, max_dead_to_right, 327 max_live_to_right, max_reclaimed_ratio); 328 } 329 330 void 331 print_initial_summary_data(ParallelCompactData& summary_data, 332 SpaceInfo* space_info) { 333 unsigned int id = PSParallelCompact::old_space_id; 334 const MutableSpace* space; 335 do { 336 space = space_info[id].space(); 337 print_initial_summary_data(summary_data, space); 338 } while (++id < PSParallelCompact::eden_space_id); 339 340 do { 341 space = space_info[id].space(); 342 print_generic_summary_data(summary_data, space->bottom(), space->top()); 343 } while (++id < PSParallelCompact::last_space_id); 344 } 345 #endif // #ifndef PRODUCT 346 347 #ifdef ASSERT 348 size_t add_obj_count; 349 size_t add_obj_size; 350 size_t mark_bitmap_count; 351 size_t mark_bitmap_size; 352 #endif // #ifdef ASSERT 353 354 ParallelCompactData::ParallelCompactData() 355 { 356 _region_start = 0; 357 358 _region_vspace = 0; 359 _region_data = 0; 360 _region_count = 0; 361 } 362 363 bool ParallelCompactData::initialize(MemRegion covered_region) 364 { 365 _region_start = covered_region.start(); 366 const size_t region_size = covered_region.word_size(); 367 DEBUG_ONLY(_region_end = _region_start + region_size;) 368 369 assert(region_align_down(_region_start) == _region_start, 370 "region start not aligned"); 371 assert((region_size & RegionSizeOffsetMask) == 0, 372 "region size not a multiple of RegionSize"); 373 374 bool result = initialize_region_data(region_size); 375 376 return result; 377 } 378 379 PSVirtualSpace* 380 ParallelCompactData::create_vspace(size_t count, size_t element_size) 381 { 382 const size_t raw_bytes = count * element_size; 383 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); 384 const size_t granularity = os::vm_allocation_granularity(); 385 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 386 387 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 388 MAX2(page_sz, granularity); 389 ReservedSpace rs(bytes, rs_align, rs_align > 0); 390 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 391 rs.size()); 392 393 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 394 395 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 396 if (vspace != 0) { 397 if (vspace->expand_by(bytes)) { 398 return vspace; 399 } 400 delete vspace; 401 // Release memory reserved in the space. 402 rs.release(); 403 } 404 405 return 0; 406 } 407 408 bool ParallelCompactData::initialize_region_data(size_t region_size) 409 { 410 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 411 _region_vspace = create_vspace(count, sizeof(RegionData)); 412 if (_region_vspace != 0) { 413 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 414 _region_count = count; 415 return true; 416 } 417 return false; 418 } 419 420 void ParallelCompactData::clear() 421 { 422 memset(_region_data, 0, _region_vspace->committed_size()); 423 } 424 425 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 426 assert(beg_region <= _region_count, "beg_region out of range"); 427 assert(end_region <= _region_count, "end_region out of range"); 428 429 const size_t region_cnt = end_region - beg_region; 430 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 431 } 432 433 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 434 { 435 const RegionData* cur_cp = region(region_idx); 436 const RegionData* const end_cp = region(region_count() - 1); 437 438 HeapWord* result = region_to_addr(region_idx); 439 if (cur_cp < end_cp) { 440 do { 441 result += cur_cp->partial_obj_size(); 442 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 443 } 444 return result; 445 } 446 447 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 448 { 449 const size_t obj_ofs = pointer_delta(addr, _region_start); 450 const size_t beg_region = obj_ofs >> Log2RegionSize; 451 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 452 453 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 454 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 455 456 if (beg_region == end_region) { 457 // All in one region. 458 _region_data[beg_region].add_live_obj(len); 459 return; 460 } 461 462 // First region. 463 const size_t beg_ofs = region_offset(addr); 464 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 465 466 Klass* klass = ((oop)addr)->klass(); 467 // Middle regions--completely spanned by this object. 468 for (size_t region = beg_region + 1; region < end_region; ++region) { 469 _region_data[region].set_partial_obj_size(RegionSize); 470 _region_data[region].set_partial_obj_addr(addr); 471 } 472 473 // Last region. 474 const size_t end_ofs = region_offset(addr + len - 1); 475 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 476 _region_data[end_region].set_partial_obj_addr(addr); 477 } 478 479 void 480 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 481 { 482 assert(region_offset(beg) == 0, "not RegionSize aligned"); 483 assert(region_offset(end) == 0, "not RegionSize aligned"); 484 485 size_t cur_region = addr_to_region_idx(beg); 486 const size_t end_region = addr_to_region_idx(end); 487 HeapWord* addr = beg; 488 while (cur_region < end_region) { 489 _region_data[cur_region].set_destination(addr); 490 _region_data[cur_region].set_destination_count(0); 491 _region_data[cur_region].set_source_region(cur_region); 492 _region_data[cur_region].set_data_location(addr); 493 494 // Update live_obj_size so the region appears completely full. 495 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 496 _region_data[cur_region].set_live_obj_size(live_size); 497 498 ++cur_region; 499 addr += RegionSize; 500 } 501 } 502 503 // Find the point at which a space can be split and, if necessary, record the 504 // split point. 505 // 506 // If the current src region (which overflowed the destination space) doesn't 507 // have a partial object, the split point is at the beginning of the current src 508 // region (an "easy" split, no extra bookkeeping required). 509 // 510 // If the current src region has a partial object, the split point is in the 511 // region where that partial object starts (call it the split_region). If 512 // split_region has a partial object, then the split point is just after that 513 // partial object (a "hard" split where we have to record the split data and 514 // zero the partial_obj_size field). With a "hard" split, we know that the 515 // partial_obj ends within split_region because the partial object that caused 516 // the overflow starts in split_region. If split_region doesn't have a partial 517 // obj, then the split is at the beginning of split_region (another "easy" 518 // split). 519 HeapWord* 520 ParallelCompactData::summarize_split_space(size_t src_region, 521 SplitInfo& split_info, 522 HeapWord* destination, 523 HeapWord* target_end, 524 HeapWord** target_next) 525 { 526 assert(destination <= target_end, "sanity"); 527 assert(destination + _region_data[src_region].data_size() > target_end, 528 "region should not fit into target space"); 529 assert(is_region_aligned(target_end), "sanity"); 530 531 size_t split_region = src_region; 532 HeapWord* split_destination = destination; 533 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 534 535 if (destination + partial_obj_size > target_end) { 536 // The split point is just after the partial object (if any) in the 537 // src_region that contains the start of the object that overflowed the 538 // destination space. 539 // 540 // Find the start of the "overflow" object and set split_region to the 541 // region containing it. 542 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 543 split_region = addr_to_region_idx(overflow_obj); 544 545 // Clear the source_region field of all destination regions whose first word 546 // came from data after the split point (a non-null source_region field 547 // implies a region must be filled). 548 // 549 // An alternative to the simple loop below: clear during post_compact(), 550 // which uses memcpy instead of individual stores, and is easy to 551 // parallelize. (The downside is that it clears the entire RegionData 552 // object as opposed to just one field.) 553 // 554 // post_compact() would have to clear the summary data up to the highest 555 // address that was written during the summary phase, which would be 556 // 557 // max(top, max(new_top, clear_top)) 558 // 559 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 560 // to target_end. 561 const RegionData* const sr = region(split_region); 562 const size_t beg_idx = 563 addr_to_region_idx(region_align_up(sr->destination() + 564 sr->partial_obj_size())); 565 const size_t end_idx = addr_to_region_idx(target_end); 566 567 if (TraceParallelOldGCSummaryPhase) { 568 gclog_or_tty->print_cr("split: clearing source_region field in [" 569 SIZE_FORMAT ", " SIZE_FORMAT ")", 570 beg_idx, end_idx); 571 } 572 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 573 _region_data[idx].set_source_region(0); 574 } 575 576 // Set split_destination and partial_obj_size to reflect the split region. 577 split_destination = sr->destination(); 578 partial_obj_size = sr->partial_obj_size(); 579 } 580 581 // The split is recorded only if a partial object extends onto the region. 582 if (partial_obj_size != 0) { 583 _region_data[split_region].set_partial_obj_size(0); 584 split_info.record(split_region, partial_obj_size, split_destination); 585 } 586 587 // Setup the continuation addresses. 588 *target_next = split_destination + partial_obj_size; 589 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 590 591 if (TraceParallelOldGCSummaryPhase) { 592 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 593 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 594 " pos=" SIZE_FORMAT, 595 split_type, source_next, split_region, 596 partial_obj_size); 597 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 598 " tn=" PTR_FORMAT, 599 split_type, split_destination, 600 addr_to_region_idx(split_destination), 601 *target_next); 602 603 if (partial_obj_size != 0) { 604 HeapWord* const po_beg = split_info.destination(); 605 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 606 gclog_or_tty->print_cr("%s split: " 607 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 608 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 609 split_type, 610 po_beg, addr_to_region_idx(po_beg), 611 po_end, addr_to_region_idx(po_end)); 612 } 613 } 614 615 return source_next; 616 } 617 618 bool ParallelCompactData::summarize(SplitInfo& split_info, 619 HeapWord* source_beg, HeapWord* source_end, 620 HeapWord** source_next, 621 HeapWord* target_beg, HeapWord* target_end, 622 HeapWord** target_next) 623 { 624 if (TraceParallelOldGCSummaryPhase) { 625 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 626 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 627 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 628 source_beg, source_end, source_next_val, 629 target_beg, target_end, *target_next); 630 } 631 632 size_t cur_region = addr_to_region_idx(source_beg); 633 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 634 635 HeapWord *dest_addr = target_beg; 636 while (cur_region < end_region) { 637 // The destination must be set even if the region has no data. 638 _region_data[cur_region].set_destination(dest_addr); 639 640 size_t words = _region_data[cur_region].data_size(); 641 if (words > 0) { 642 // If cur_region does not fit entirely into the target space, find a point 643 // at which the source space can be 'split' so that part is copied to the 644 // target space and the rest is copied elsewhere. 645 if (dest_addr + words > target_end) { 646 assert(source_next != NULL, "source_next is NULL when splitting"); 647 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 648 target_end, target_next); 649 return false; 650 } 651 652 // Compute the destination_count for cur_region, and if necessary, update 653 // source_region for a destination region. The source_region field is 654 // updated if cur_region is the first (left-most) region to be copied to a 655 // destination region. 656 // 657 // The destination_count calculation is a bit subtle. A region that has 658 // data that compacts into itself does not count itself as a destination. 659 // This maintains the invariant that a zero count means the region is 660 // available and can be claimed and then filled. 661 uint destination_count = 0; 662 if (split_info.is_split(cur_region)) { 663 // The current region has been split: the partial object will be copied 664 // to one destination space and the remaining data will be copied to 665 // another destination space. Adjust the initial destination_count and, 666 // if necessary, set the source_region field if the partial object will 667 // cross a destination region boundary. 668 destination_count = split_info.destination_count(); 669 if (destination_count == 2) { 670 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 671 _region_data[dest_idx].set_source_region(cur_region); 672 } 673 } 674 675 HeapWord* const last_addr = dest_addr + words - 1; 676 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 677 const size_t dest_region_2 = addr_to_region_idx(last_addr); 678 679 // Initially assume that the destination regions will be the same and 680 // adjust the value below if necessary. Under this assumption, if 681 // cur_region == dest_region_2, then cur_region will be compacted 682 // completely into itself. 683 destination_count += cur_region == dest_region_2 ? 0 : 1; 684 if (dest_region_1 != dest_region_2) { 685 // Destination regions differ; adjust destination_count. 686 destination_count += 1; 687 // Data from cur_region will be copied to the start of dest_region_2. 688 _region_data[dest_region_2].set_source_region(cur_region); 689 } else if (region_offset(dest_addr) == 0) { 690 // Data from cur_region will be copied to the start of the destination 691 // region. 692 _region_data[dest_region_1].set_source_region(cur_region); 693 } 694 695 _region_data[cur_region].set_destination_count(destination_count); 696 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 697 dest_addr += words; 698 } 699 700 ++cur_region; 701 } 702 703 *target_next = dest_addr; 704 return true; 705 } 706 707 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 708 assert(addr != NULL, "Should detect NULL oop earlier"); 709 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); 710 #ifdef ASSERT 711 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { 712 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); 713 } 714 #endif 715 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); 716 717 // Region covering the object. 718 size_t region_index = addr_to_region_idx(addr); 719 const RegionData* const region_ptr = region(region_index); 720 HeapWord* const region_addr = region_align_down(addr); 721 722 assert(addr < region_addr + RegionSize, "Region does not cover object"); 723 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check"); 724 725 HeapWord* result = region_ptr->destination(); 726 727 // If all the data in the region is live, then the new location of the object 728 // can be calculated from the destination of the region plus the offset of the 729 // object in the region. 730 if (region_ptr->data_size() == RegionSize) { 731 result += pointer_delta(addr, region_addr); 732 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 733 return result; 734 } 735 736 // The new location of the object is 737 // region destination + 738 // size of the partial object extending onto the region + 739 // sizes of the live objects in the Region that are to the left of addr 740 const size_t partial_obj_size = region_ptr->partial_obj_size(); 741 HeapWord* const search_start = region_addr + partial_obj_size; 742 743 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 744 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 745 746 result += partial_obj_size + live_to_left; 747 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 748 return result; 749 } 750 751 #ifdef ASSERT 752 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 753 { 754 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 755 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 756 for (const size_t* p = beg; p < end; ++p) { 757 assert(*p == 0, "not zero"); 758 } 759 } 760 761 void ParallelCompactData::verify_clear() 762 { 763 verify_clear(_region_vspace); 764 } 765 #endif // #ifdef ASSERT 766 767 #ifdef NOT_PRODUCT 768 ParallelCompactData::RegionData* debug_region(size_t region_index) { 769 ParallelCompactData& sd = PSParallelCompact::summary_data(); 770 return sd.region(region_index); 771 } 772 #endif 773 774 elapsedTimer PSParallelCompact::_accumulated_time; 775 unsigned int PSParallelCompact::_total_invocations = 0; 776 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 777 jlong PSParallelCompact::_time_of_last_gc = 0; 778 CollectorCounters* PSParallelCompact::_counters = NULL; 779 ParMarkBitMap PSParallelCompact::_mark_bitmap; 780 ParallelCompactData PSParallelCompact::_summary_data; 781 782 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 783 784 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } 785 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 786 787 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 788 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 789 790 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 791 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 792 793 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); } 794 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); } 795 796 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 797 798 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { 799 mark_and_push(_compaction_manager, p); 800 } 801 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 802 803 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 804 klass->oops_do(_mark_and_push_closure); 805 } 806 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 807 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 808 } 809 810 void PSParallelCompact::post_initialize() { 811 ParallelScavengeHeap* heap = gc_heap(); 812 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 813 814 MemRegion mr = heap->reserved_region(); 815 _ref_processor = 816 new ReferenceProcessor(mr, // span 817 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 818 (int) ParallelGCThreads, // mt processing degree 819 true, // mt discovery 820 (int) ParallelGCThreads, // mt discovery degree 821 true, // atomic_discovery 822 &_is_alive_closure, // non-header is alive closure 823 false); // write barrier for next field updates 824 _counters = new CollectorCounters("PSParallelCompact", 1); 825 826 // Initialize static fields in ParCompactionManager. 827 ParCompactionManager::initialize(mark_bitmap()); 828 } 829 830 bool PSParallelCompact::initialize() { 831 ParallelScavengeHeap* heap = gc_heap(); 832 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 833 MemRegion mr = heap->reserved_region(); 834 835 // Was the old gen get allocated successfully? 836 if (!heap->old_gen()->is_allocated()) { 837 return false; 838 } 839 840 initialize_space_info(); 841 initialize_dead_wood_limiter(); 842 843 if (!_mark_bitmap.initialize(mr)) { 844 vm_shutdown_during_initialization("Unable to allocate bit map for " 845 "parallel garbage collection for the requested heap size."); 846 return false; 847 } 848 849 if (!_summary_data.initialize(mr)) { 850 vm_shutdown_during_initialization("Unable to allocate tables for " 851 "parallel garbage collection for the requested heap size."); 852 return false; 853 } 854 855 return true; 856 } 857 858 void PSParallelCompact::initialize_space_info() 859 { 860 memset(&_space_info, 0, sizeof(_space_info)); 861 862 ParallelScavengeHeap* heap = gc_heap(); 863 PSYoungGen* young_gen = heap->young_gen(); 864 865 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 866 _space_info[eden_space_id].set_space(young_gen->eden_space()); 867 _space_info[from_space_id].set_space(young_gen->from_space()); 868 _space_info[to_space_id].set_space(young_gen->to_space()); 869 870 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 871 } 872 873 void PSParallelCompact::initialize_dead_wood_limiter() 874 { 875 const size_t max = 100; 876 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 877 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 878 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 879 DEBUG_ONLY(_dwl_initialized = true;) 880 _dwl_adjustment = normal_distribution(1.0); 881 } 882 883 // Simple class for storing info about the heap at the start of GC, to be used 884 // after GC for comparison/printing. 885 class PreGCValues { 886 public: 887 PreGCValues() { } 888 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 889 890 void fill(ParallelScavengeHeap* heap) { 891 _heap_used = heap->used(); 892 _young_gen_used = heap->young_gen()->used_in_bytes(); 893 _old_gen_used = heap->old_gen()->used_in_bytes(); 894 _metadata_used = MetaspaceAux::allocated_used_bytes(); 895 }; 896 897 size_t heap_used() const { return _heap_used; } 898 size_t young_gen_used() const { return _young_gen_used; } 899 size_t old_gen_used() const { return _old_gen_used; } 900 size_t metadata_used() const { return _metadata_used; } 901 902 private: 903 size_t _heap_used; 904 size_t _young_gen_used; 905 size_t _old_gen_used; 906 size_t _metadata_used; 907 }; 908 909 void 910 PSParallelCompact::clear_data_covering_space(SpaceId id) 911 { 912 // At this point, top is the value before GC, new_top() is the value that will 913 // be set at the end of GC. The marking bitmap is cleared to top; nothing 914 // should be marked above top. The summary data is cleared to the larger of 915 // top & new_top. 916 MutableSpace* const space = _space_info[id].space(); 917 HeapWord* const bot = space->bottom(); 918 HeapWord* const top = space->top(); 919 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 920 921 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 922 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 923 _mark_bitmap.clear_range(beg_bit, end_bit); 924 925 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 926 const size_t end_region = 927 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 928 _summary_data.clear_range(beg_region, end_region); 929 930 // Clear the data used to 'split' regions. 931 SplitInfo& split_info = _space_info[id].split_info(); 932 if (split_info.is_valid()) { 933 split_info.clear(); 934 } 935 DEBUG_ONLY(split_info.verify_clear();) 936 } 937 938 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 939 { 940 // Update the from & to space pointers in space_info, since they are swapped 941 // at each young gen gc. Do the update unconditionally (even though a 942 // promotion failure does not swap spaces) because an unknown number of minor 943 // collections will have swapped the spaces an unknown number of times. 944 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); 945 ParallelScavengeHeap* heap = gc_heap(); 946 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 947 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 948 949 pre_gc_values->fill(heap); 950 951 NOT_PRODUCT(_mark_bitmap.reset_counters()); 952 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 953 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 954 955 // Increment the invocation count 956 heap->increment_total_collections(true); 957 958 // We need to track unique mark sweep invocations as well. 959 _total_invocations++; 960 961 heap->print_heap_before_gc(); 962 963 // Fill in TLABs 964 heap->accumulate_statistics_all_tlabs(); 965 heap->ensure_parsability(true); // retire TLABs 966 967 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 968 HandleMark hm; // Discard invalid handles created during verification 969 Universe::verify(" VerifyBeforeGC:"); 970 } 971 972 // Verify object start arrays 973 if (VerifyObjectStartArray && 974 VerifyBeforeGC) { 975 heap->old_gen()->verify_object_start_array(); 976 } 977 978 DEBUG_ONLY(mark_bitmap()->verify_clear();) 979 DEBUG_ONLY(summary_data().verify_clear();) 980 981 // Have worker threads release resources the next time they run a task. 982 gc_task_manager()->release_all_resources(); 983 } 984 985 void PSParallelCompact::post_compact() 986 { 987 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); 988 989 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 990 // Clear the marking bitmap, summary data and split info. 991 clear_data_covering_space(SpaceId(id)); 992 // Update top(). Must be done after clearing the bitmap and summary data. 993 _space_info[id].publish_new_top(); 994 } 995 996 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 997 MutableSpace* const from_space = _space_info[from_space_id].space(); 998 MutableSpace* const to_space = _space_info[to_space_id].space(); 999 1000 ParallelScavengeHeap* heap = gc_heap(); 1001 bool eden_empty = eden_space->is_empty(); 1002 if (!eden_empty) { 1003 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1004 heap->young_gen(), heap->old_gen()); 1005 } 1006 1007 // Update heap occupancy information which is used as input to the soft ref 1008 // clearing policy at the next gc. 1009 Universe::update_heap_info_at_gc(); 1010 1011 bool young_gen_empty = eden_empty && from_space->is_empty() && 1012 to_space->is_empty(); 1013 1014 BarrierSet* bs = heap->barrier_set(); 1015 if (bs->is_a(BarrierSet::ModRef)) { 1016 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1017 MemRegion old_mr = heap->old_gen()->reserved(); 1018 1019 if (young_gen_empty) { 1020 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1021 } else { 1022 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1023 } 1024 } 1025 1026 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1027 ClassLoaderDataGraph::purge(); 1028 MetaspaceAux::verify_metrics(); 1029 1030 Threads::gc_epilogue(); 1031 CodeCache::gc_epilogue(); 1032 JvmtiExport::gc_epilogue(); 1033 1034 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1035 1036 ref_processor()->enqueue_discovered_references(NULL); 1037 1038 if (ZapUnusedHeapArea) { 1039 heap->gen_mangle_unused_area(); 1040 } 1041 1042 // Update time of last GC 1043 reset_millis_since_last_gc(); 1044 } 1045 1046 HeapWord* 1047 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1048 bool maximum_compaction) 1049 { 1050 const size_t region_size = ParallelCompactData::RegionSize; 1051 const ParallelCompactData& sd = summary_data(); 1052 1053 const MutableSpace* const space = _space_info[id].space(); 1054 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1055 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1056 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1057 1058 // Skip full regions at the beginning of the space--they are necessarily part 1059 // of the dense prefix. 1060 size_t full_count = 0; 1061 const RegionData* cp; 1062 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1063 ++full_count; 1064 } 1065 1066 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1067 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1068 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1069 if (maximum_compaction || cp == end_cp || interval_ended) { 1070 _maximum_compaction_gc_num = total_invocations(); 1071 return sd.region_to_addr(cp); 1072 } 1073 1074 HeapWord* const new_top = _space_info[id].new_top(); 1075 const size_t space_live = pointer_delta(new_top, space->bottom()); 1076 const size_t space_used = space->used_in_words(); 1077 const size_t space_capacity = space->capacity_in_words(); 1078 1079 const double cur_density = double(space_live) / space_capacity; 1080 const double deadwood_density = 1081 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1082 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1083 1084 if (TraceParallelOldGCDensePrefix) { 1085 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1086 cur_density, deadwood_density, deadwood_goal); 1087 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1088 "space_cap=" SIZE_FORMAT, 1089 space_live, space_used, 1090 space_capacity); 1091 } 1092 1093 // XXX - Use binary search? 1094 HeapWord* dense_prefix = sd.region_to_addr(cp); 1095 const RegionData* full_cp = cp; 1096 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1097 while (cp < end_cp) { 1098 HeapWord* region_destination = cp->destination(); 1099 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1100 if (TraceParallelOldGCDensePrefix && Verbose) { 1101 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1102 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1103 sd.region(cp), region_destination, 1104 dense_prefix, cur_deadwood); 1105 } 1106 1107 if (cur_deadwood >= deadwood_goal) { 1108 // Found the region that has the correct amount of deadwood to the left. 1109 // This typically occurs after crossing a fairly sparse set of regions, so 1110 // iterate backwards over those sparse regions, looking for the region 1111 // that has the lowest density of live objects 'to the right.' 1112 size_t space_to_left = sd.region(cp) * region_size; 1113 size_t live_to_left = space_to_left - cur_deadwood; 1114 size_t space_to_right = space_capacity - space_to_left; 1115 size_t live_to_right = space_live - live_to_left; 1116 double density_to_right = double(live_to_right) / space_to_right; 1117 while (cp > full_cp) { 1118 --cp; 1119 const size_t prev_region_live_to_right = live_to_right - 1120 cp->data_size(); 1121 const size_t prev_region_space_to_right = space_to_right + region_size; 1122 double prev_region_density_to_right = 1123 double(prev_region_live_to_right) / prev_region_space_to_right; 1124 if (density_to_right <= prev_region_density_to_right) { 1125 return dense_prefix; 1126 } 1127 if (TraceParallelOldGCDensePrefix && Verbose) { 1128 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1129 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1130 prev_region_density_to_right); 1131 } 1132 dense_prefix -= region_size; 1133 live_to_right = prev_region_live_to_right; 1134 space_to_right = prev_region_space_to_right; 1135 density_to_right = prev_region_density_to_right; 1136 } 1137 return dense_prefix; 1138 } 1139 1140 dense_prefix += region_size; 1141 ++cp; 1142 } 1143 1144 return dense_prefix; 1145 } 1146 1147 #ifndef PRODUCT 1148 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1149 const SpaceId id, 1150 const bool maximum_compaction, 1151 HeapWord* const addr) 1152 { 1153 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1154 RegionData* const cp = summary_data().region(region_idx); 1155 const MutableSpace* const space = _space_info[id].space(); 1156 HeapWord* const new_top = _space_info[id].new_top(); 1157 1158 const size_t space_live = pointer_delta(new_top, space->bottom()); 1159 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1160 const size_t space_cap = space->capacity_in_words(); 1161 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1162 const size_t live_to_right = new_top - cp->destination(); 1163 const size_t dead_to_right = space->top() - addr - live_to_right; 1164 1165 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1166 "spl=" SIZE_FORMAT " " 1167 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1168 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1169 " ratio=%10.8f", 1170 algorithm, addr, region_idx, 1171 space_live, 1172 dead_to_left, dead_to_left_pct, 1173 dead_to_right, live_to_right, 1174 double(dead_to_right) / live_to_right); 1175 } 1176 #endif // #ifndef PRODUCT 1177 1178 // Return a fraction indicating how much of the generation can be treated as 1179 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1180 // based on the density of live objects in the generation to determine a limit, 1181 // which is then adjusted so the return value is min_percent when the density is 1182 // 1. 1183 // 1184 // The following table shows some return values for a different values of the 1185 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1186 // min_percent is 1. 1187 // 1188 // fraction allowed as dead wood 1189 // ----------------------------------------------------------------- 1190 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1191 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1192 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1193 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1194 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1195 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1196 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1197 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1198 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1199 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1200 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1201 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1202 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1203 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1204 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1205 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1206 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1207 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1208 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1209 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1210 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1211 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1212 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1213 1214 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1215 { 1216 assert(_dwl_initialized, "uninitialized"); 1217 1218 // The raw limit is the value of the normal distribution at x = density. 1219 const double raw_limit = normal_distribution(density); 1220 1221 // Adjust the raw limit so it becomes the minimum when the density is 1. 1222 // 1223 // First subtract the adjustment value (which is simply the precomputed value 1224 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1225 // Then add the minimum value, so the minimum is returned when the density is 1226 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1227 const double min = double(min_percent) / 100.0; 1228 const double limit = raw_limit - _dwl_adjustment + min; 1229 return MAX2(limit, 0.0); 1230 } 1231 1232 ParallelCompactData::RegionData* 1233 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1234 const RegionData* end) 1235 { 1236 const size_t region_size = ParallelCompactData::RegionSize; 1237 ParallelCompactData& sd = summary_data(); 1238 size_t left = sd.region(beg); 1239 size_t right = end > beg ? sd.region(end) - 1 : left; 1240 1241 // Binary search. 1242 while (left < right) { 1243 // Equivalent to (left + right) / 2, but does not overflow. 1244 const size_t middle = left + (right - left) / 2; 1245 RegionData* const middle_ptr = sd.region(middle); 1246 HeapWord* const dest = middle_ptr->destination(); 1247 HeapWord* const addr = sd.region_to_addr(middle); 1248 assert(dest != NULL, "sanity"); 1249 assert(dest <= addr, "must move left"); 1250 1251 if (middle > left && dest < addr) { 1252 right = middle - 1; 1253 } else if (middle < right && middle_ptr->data_size() == region_size) { 1254 left = middle + 1; 1255 } else { 1256 return middle_ptr; 1257 } 1258 } 1259 return sd.region(left); 1260 } 1261 1262 ParallelCompactData::RegionData* 1263 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1264 const RegionData* end, 1265 size_t dead_words) 1266 { 1267 ParallelCompactData& sd = summary_data(); 1268 size_t left = sd.region(beg); 1269 size_t right = end > beg ? sd.region(end) - 1 : left; 1270 1271 // Binary search. 1272 while (left < right) { 1273 // Equivalent to (left + right) / 2, but does not overflow. 1274 const size_t middle = left + (right - left) / 2; 1275 RegionData* const middle_ptr = sd.region(middle); 1276 HeapWord* const dest = middle_ptr->destination(); 1277 HeapWord* const addr = sd.region_to_addr(middle); 1278 assert(dest != NULL, "sanity"); 1279 assert(dest <= addr, "must move left"); 1280 1281 const size_t dead_to_left = pointer_delta(addr, dest); 1282 if (middle > left && dead_to_left > dead_words) { 1283 right = middle - 1; 1284 } else if (middle < right && dead_to_left < dead_words) { 1285 left = middle + 1; 1286 } else { 1287 return middle_ptr; 1288 } 1289 } 1290 return sd.region(left); 1291 } 1292 1293 // The result is valid during the summary phase, after the initial summarization 1294 // of each space into itself, and before final summarization. 1295 inline double 1296 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1297 HeapWord* const bottom, 1298 HeapWord* const top, 1299 HeapWord* const new_top) 1300 { 1301 ParallelCompactData& sd = summary_data(); 1302 1303 assert(cp != NULL, "sanity"); 1304 assert(bottom != NULL, "sanity"); 1305 assert(top != NULL, "sanity"); 1306 assert(new_top != NULL, "sanity"); 1307 assert(top >= new_top, "summary data problem?"); 1308 assert(new_top > bottom, "space is empty; should not be here"); 1309 assert(new_top >= cp->destination(), "sanity"); 1310 assert(top >= sd.region_to_addr(cp), "sanity"); 1311 1312 HeapWord* const destination = cp->destination(); 1313 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1314 const size_t compacted_region_live = pointer_delta(new_top, destination); 1315 const size_t compacted_region_used = pointer_delta(top, 1316 sd.region_to_addr(cp)); 1317 const size_t reclaimable = compacted_region_used - compacted_region_live; 1318 1319 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1320 return double(reclaimable) / divisor; 1321 } 1322 1323 // Return the address of the end of the dense prefix, a.k.a. the start of the 1324 // compacted region. The address is always on a region boundary. 1325 // 1326 // Completely full regions at the left are skipped, since no compaction can 1327 // occur in those regions. Then the maximum amount of dead wood to allow is 1328 // computed, based on the density (amount live / capacity) of the generation; 1329 // the region with approximately that amount of dead space to the left is 1330 // identified as the limit region. Regions between the last completely full 1331 // region and the limit region are scanned and the one that has the best 1332 // (maximum) reclaimed_ratio() is selected. 1333 HeapWord* 1334 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1335 bool maximum_compaction) 1336 { 1337 if (ParallelOldGCSplitALot) { 1338 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1339 // The value was chosen to provoke splitting a young gen space; use it. 1340 return _space_info[id].dense_prefix(); 1341 } 1342 } 1343 1344 const size_t region_size = ParallelCompactData::RegionSize; 1345 const ParallelCompactData& sd = summary_data(); 1346 1347 const MutableSpace* const space = _space_info[id].space(); 1348 HeapWord* const top = space->top(); 1349 HeapWord* const top_aligned_up = sd.region_align_up(top); 1350 HeapWord* const new_top = _space_info[id].new_top(); 1351 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1352 HeapWord* const bottom = space->bottom(); 1353 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1354 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1355 const RegionData* const new_top_cp = 1356 sd.addr_to_region_ptr(new_top_aligned_up); 1357 1358 // Skip full regions at the beginning of the space--they are necessarily part 1359 // of the dense prefix. 1360 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1361 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1362 space->is_empty(), "no dead space allowed to the left"); 1363 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1364 "region must have dead space"); 1365 1366 // The gc number is saved whenever a maximum compaction is done, and used to 1367 // determine when the maximum compaction interval has expired. This avoids 1368 // successive max compactions for different reasons. 1369 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1370 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1371 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1372 total_invocations() == HeapFirstMaximumCompactionCount; 1373 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1374 _maximum_compaction_gc_num = total_invocations(); 1375 return sd.region_to_addr(full_cp); 1376 } 1377 1378 const size_t space_live = pointer_delta(new_top, bottom); 1379 const size_t space_used = space->used_in_words(); 1380 const size_t space_capacity = space->capacity_in_words(); 1381 1382 const double density = double(space_live) / double(space_capacity); 1383 const size_t min_percent_free = MarkSweepDeadRatio; 1384 const double limiter = dead_wood_limiter(density, min_percent_free); 1385 const size_t dead_wood_max = space_used - space_live; 1386 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1387 dead_wood_max); 1388 1389 if (TraceParallelOldGCDensePrefix) { 1390 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1391 "space_cap=" SIZE_FORMAT, 1392 space_live, space_used, 1393 space_capacity); 1394 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1395 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1396 density, min_percent_free, limiter, 1397 dead_wood_max, dead_wood_limit); 1398 } 1399 1400 // Locate the region with the desired amount of dead space to the left. 1401 const RegionData* const limit_cp = 1402 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1403 1404 // Scan from the first region with dead space to the limit region and find the 1405 // one with the best (largest) reclaimed ratio. 1406 double best_ratio = 0.0; 1407 const RegionData* best_cp = full_cp; 1408 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1409 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1410 if (tmp_ratio > best_ratio) { 1411 best_cp = cp; 1412 best_ratio = tmp_ratio; 1413 } 1414 } 1415 1416 #if 0 1417 // Something to consider: if the region with the best ratio is 'close to' the 1418 // first region w/free space, choose the first region with free space 1419 // ("first-free"). The first-free region is usually near the start of the 1420 // heap, which means we are copying most of the heap already, so copy a bit 1421 // more to get complete compaction. 1422 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1423 _maximum_compaction_gc_num = total_invocations(); 1424 best_cp = full_cp; 1425 } 1426 #endif // #if 0 1427 1428 return sd.region_to_addr(best_cp); 1429 } 1430 1431 #ifndef PRODUCT 1432 void 1433 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1434 size_t words) 1435 { 1436 if (TraceParallelOldGCSummaryPhase) { 1437 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1438 SIZE_FORMAT, start, start + words, words); 1439 } 1440 1441 ObjectStartArray* const start_array = _space_info[id].start_array(); 1442 CollectedHeap::fill_with_objects(start, words); 1443 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1444 _mark_bitmap.mark_obj(p, words); 1445 _summary_data.add_obj(p, words); 1446 start_array->allocate_block(p); 1447 } 1448 } 1449 1450 void 1451 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1452 { 1453 ParallelCompactData& sd = summary_data(); 1454 MutableSpace* space = _space_info[id].space(); 1455 1456 // Find the source and destination start addresses. 1457 HeapWord* const src_addr = sd.region_align_down(start); 1458 HeapWord* dst_addr; 1459 if (src_addr < start) { 1460 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1461 } else if (src_addr > space->bottom()) { 1462 // The start (the original top() value) is aligned to a region boundary so 1463 // the associated region does not have a destination. Compute the 1464 // destination from the previous region. 1465 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1466 dst_addr = cp->destination() + cp->data_size(); 1467 } else { 1468 // Filling the entire space. 1469 dst_addr = space->bottom(); 1470 } 1471 assert(dst_addr != NULL, "sanity"); 1472 1473 // Update the summary data. 1474 bool result = _summary_data.summarize(_space_info[id].split_info(), 1475 src_addr, space->top(), NULL, 1476 dst_addr, space->end(), 1477 _space_info[id].new_top_addr()); 1478 assert(result, "should not fail: bad filler object size"); 1479 } 1480 1481 void 1482 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1483 { 1484 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1485 return; 1486 } 1487 1488 MutableSpace* const space = _space_info[id].space(); 1489 if (space->is_empty()) { 1490 HeapWord* b = space->bottom(); 1491 HeapWord* t = b + space->capacity_in_words() / 2; 1492 space->set_top(t); 1493 if (ZapUnusedHeapArea) { 1494 space->set_top_for_allocations(); 1495 } 1496 1497 size_t min_size = CollectedHeap::min_fill_size(); 1498 size_t obj_len = min_size; 1499 while (b + obj_len <= t) { 1500 CollectedHeap::fill_with_object(b, obj_len); 1501 mark_bitmap()->mark_obj(b, obj_len); 1502 summary_data().add_obj(b, obj_len); 1503 b += obj_len; 1504 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1505 } 1506 if (b < t) { 1507 // The loop didn't completely fill to t (top); adjust top downward. 1508 space->set_top(b); 1509 if (ZapUnusedHeapArea) { 1510 space->set_top_for_allocations(); 1511 } 1512 } 1513 1514 HeapWord** nta = _space_info[id].new_top_addr(); 1515 bool result = summary_data().summarize(_space_info[id].split_info(), 1516 space->bottom(), space->top(), NULL, 1517 space->bottom(), space->end(), nta); 1518 assert(result, "space must fit into itself"); 1519 } 1520 } 1521 1522 void 1523 PSParallelCompact::provoke_split(bool & max_compaction) 1524 { 1525 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1526 return; 1527 } 1528 1529 const size_t region_size = ParallelCompactData::RegionSize; 1530 ParallelCompactData& sd = summary_data(); 1531 1532 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1533 MutableSpace* const from_space = _space_info[from_space_id].space(); 1534 const size_t eden_live = pointer_delta(eden_space->top(), 1535 _space_info[eden_space_id].new_top()); 1536 const size_t from_live = pointer_delta(from_space->top(), 1537 _space_info[from_space_id].new_top()); 1538 1539 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1540 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1541 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1542 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1543 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1544 1545 // Choose the space to split; need at least 2 regions live (or fillable). 1546 SpaceId id; 1547 MutableSpace* space; 1548 size_t live_words; 1549 size_t fill_words; 1550 if (eden_live + eden_fillable >= region_size * 2) { 1551 id = eden_space_id; 1552 space = eden_space; 1553 live_words = eden_live; 1554 fill_words = eden_fillable; 1555 } else if (from_live + from_fillable >= region_size * 2) { 1556 id = from_space_id; 1557 space = from_space; 1558 live_words = from_live; 1559 fill_words = from_fillable; 1560 } else { 1561 return; // Give up. 1562 } 1563 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1564 1565 if (live_words < region_size * 2) { 1566 // Fill from top() to end() w/live objects of mixed sizes. 1567 HeapWord* const fill_start = space->top(); 1568 live_words += fill_words; 1569 1570 space->set_top(fill_start + fill_words); 1571 if (ZapUnusedHeapArea) { 1572 space->set_top_for_allocations(); 1573 } 1574 1575 HeapWord* cur_addr = fill_start; 1576 while (fill_words > 0) { 1577 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1578 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1579 if (fill_words - cur_size < min_fill_size) { 1580 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1581 } 1582 1583 CollectedHeap::fill_with_object(cur_addr, cur_size); 1584 mark_bitmap()->mark_obj(cur_addr, cur_size); 1585 sd.add_obj(cur_addr, cur_size); 1586 1587 cur_addr += cur_size; 1588 fill_words -= cur_size; 1589 } 1590 1591 summarize_new_objects(id, fill_start); 1592 } 1593 1594 max_compaction = false; 1595 1596 // Manipulate the old gen so that it has room for about half of the live data 1597 // in the target young gen space (live_words / 2). 1598 id = old_space_id; 1599 space = _space_info[id].space(); 1600 const size_t free_at_end = space->free_in_words(); 1601 const size_t free_target = align_object_size(live_words / 2); 1602 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1603 1604 if (free_at_end >= free_target + min_fill_size) { 1605 // Fill space above top() and set the dense prefix so everything survives. 1606 HeapWord* const fill_start = space->top(); 1607 const size_t fill_size = free_at_end - free_target; 1608 space->set_top(space->top() + fill_size); 1609 if (ZapUnusedHeapArea) { 1610 space->set_top_for_allocations(); 1611 } 1612 fill_with_live_objects(id, fill_start, fill_size); 1613 summarize_new_objects(id, fill_start); 1614 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1615 } else if (dead + free_at_end > free_target) { 1616 // Find a dense prefix that makes the right amount of space available. 1617 HeapWord* cur = sd.region_align_down(space->top()); 1618 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1619 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1620 while (dead_to_right < free_target) { 1621 cur -= region_size; 1622 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1623 dead_to_right = pointer_delta(space->end(), cur_destination); 1624 } 1625 _space_info[id].set_dense_prefix(cur); 1626 } 1627 } 1628 #endif // #ifndef PRODUCT 1629 1630 void PSParallelCompact::summarize_spaces_quick() 1631 { 1632 for (unsigned int i = 0; i < last_space_id; ++i) { 1633 const MutableSpace* space = _space_info[i].space(); 1634 HeapWord** nta = _space_info[i].new_top_addr(); 1635 bool result = _summary_data.summarize(_space_info[i].split_info(), 1636 space->bottom(), space->top(), NULL, 1637 space->bottom(), space->end(), nta); 1638 assert(result, "space must fit into itself"); 1639 _space_info[i].set_dense_prefix(space->bottom()); 1640 } 1641 1642 #ifndef PRODUCT 1643 if (ParallelOldGCSplitALot) { 1644 provoke_split_fill_survivor(to_space_id); 1645 } 1646 #endif // #ifndef PRODUCT 1647 } 1648 1649 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1650 { 1651 HeapWord* const dense_prefix_end = dense_prefix(id); 1652 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1653 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1654 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1655 // Only enough dead space is filled so that any remaining dead space to the 1656 // left is larger than the minimum filler object. (The remainder is filled 1657 // during the copy/update phase.) 1658 // 1659 // The size of the dead space to the right of the boundary is not a 1660 // concern, since compaction will be able to use whatever space is 1661 // available. 1662 // 1663 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1664 // surrounds the space to be filled with an object. 1665 // 1666 // In the 32-bit VM, each bit represents two 32-bit words: 1667 // +---+ 1668 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1669 // end_bits: ... x x x | 0 | || 0 x x ... 1670 // +---+ 1671 // 1672 // In the 64-bit VM, each bit represents one 64-bit word: 1673 // +------------+ 1674 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1675 // end_bits: ... x x 1 | 0 || 0 | x x ... 1676 // +------------+ 1677 // +-------+ 1678 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1679 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1680 // +-------+ 1681 // +-----------+ 1682 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1683 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1684 // +-----------+ 1685 // +-------+ 1686 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1687 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1688 // +-------+ 1689 1690 // Initially assume case a, c or e will apply. 1691 size_t obj_len = CollectedHeap::min_fill_size(); 1692 HeapWord* obj_beg = dense_prefix_end - obj_len; 1693 1694 #ifdef _LP64 1695 if (MinObjAlignment > 1) { // object alignment > heap word size 1696 // Cases a, c or e. 1697 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1698 // Case b above. 1699 obj_beg = dense_prefix_end - 1; 1700 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1701 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1702 // Case d above. 1703 obj_beg = dense_prefix_end - 3; 1704 obj_len = 3; 1705 } 1706 #endif // #ifdef _LP64 1707 1708 CollectedHeap::fill_with_object(obj_beg, obj_len); 1709 _mark_bitmap.mark_obj(obj_beg, obj_len); 1710 _summary_data.add_obj(obj_beg, obj_len); 1711 assert(start_array(id) != NULL, "sanity"); 1712 start_array(id)->allocate_block(obj_beg); 1713 } 1714 } 1715 1716 void 1717 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1718 { 1719 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1720 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1721 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1722 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1723 cur->set_source_region(0); 1724 } 1725 } 1726 1727 void 1728 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1729 { 1730 assert(id < last_space_id, "id out of range"); 1731 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1732 ParallelOldGCSplitALot && id == old_space_id, 1733 "should have been reset in summarize_spaces_quick()"); 1734 1735 const MutableSpace* space = _space_info[id].space(); 1736 if (_space_info[id].new_top() != space->bottom()) { 1737 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1738 _space_info[id].set_dense_prefix(dense_prefix_end); 1739 1740 #ifndef PRODUCT 1741 if (TraceParallelOldGCDensePrefix) { 1742 print_dense_prefix_stats("ratio", id, maximum_compaction, 1743 dense_prefix_end); 1744 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1745 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1746 } 1747 #endif // #ifndef PRODUCT 1748 1749 // Recompute the summary data, taking into account the dense prefix. If 1750 // every last byte will be reclaimed, then the existing summary data which 1751 // compacts everything can be left in place. 1752 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1753 // If dead space crosses the dense prefix boundary, it is (at least 1754 // partially) filled with a dummy object, marked live and added to the 1755 // summary data. This simplifies the copy/update phase and must be done 1756 // before the final locations of objects are determined, to prevent 1757 // leaving a fragment of dead space that is too small to fill. 1758 fill_dense_prefix_end(id); 1759 1760 // Compute the destination of each Region, and thus each object. 1761 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1762 _summary_data.summarize(_space_info[id].split_info(), 1763 dense_prefix_end, space->top(), NULL, 1764 dense_prefix_end, space->end(), 1765 _space_info[id].new_top_addr()); 1766 } 1767 } 1768 1769 if (TraceParallelOldGCSummaryPhase) { 1770 const size_t region_size = ParallelCompactData::RegionSize; 1771 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1772 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1773 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1774 HeapWord* const new_top = _space_info[id].new_top(); 1775 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1776 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1777 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1778 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1779 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1780 id, space->capacity_in_words(), dense_prefix_end, 1781 dp_region, dp_words / region_size, 1782 cr_words / region_size, new_top); 1783 } 1784 } 1785 1786 #ifndef PRODUCT 1787 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1788 HeapWord* dst_beg, HeapWord* dst_end, 1789 SpaceId src_space_id, 1790 HeapWord* src_beg, HeapWord* src_end) 1791 { 1792 if (TraceParallelOldGCSummaryPhase) { 1793 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1794 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1795 SIZE_FORMAT "-" SIZE_FORMAT " " 1796 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1797 SIZE_FORMAT "-" SIZE_FORMAT, 1798 src_space_id, space_names[src_space_id], 1799 dst_space_id, space_names[dst_space_id], 1800 src_beg, src_end, 1801 _summary_data.addr_to_region_idx(src_beg), 1802 _summary_data.addr_to_region_idx(src_end), 1803 dst_beg, dst_end, 1804 _summary_data.addr_to_region_idx(dst_beg), 1805 _summary_data.addr_to_region_idx(dst_end)); 1806 } 1807 } 1808 #endif // #ifndef PRODUCT 1809 1810 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1811 bool maximum_compaction) 1812 { 1813 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); 1814 // trace("2"); 1815 1816 #ifdef ASSERT 1817 if (TraceParallelOldGCMarkingPhase) { 1818 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1819 "add_obj_bytes=" SIZE_FORMAT, 1820 add_obj_count, add_obj_size * HeapWordSize); 1821 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1822 "mark_bitmap_bytes=" SIZE_FORMAT, 1823 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1824 } 1825 #endif // #ifdef ASSERT 1826 1827 // Quick summarization of each space into itself, to see how much is live. 1828 summarize_spaces_quick(); 1829 1830 if (TraceParallelOldGCSummaryPhase) { 1831 tty->print_cr("summary_phase: after summarizing each space to self"); 1832 Universe::print(); 1833 NOT_PRODUCT(print_region_ranges()); 1834 if (Verbose) { 1835 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1836 } 1837 } 1838 1839 // The amount of live data that will end up in old space (assuming it fits). 1840 size_t old_space_total_live = 0; 1841 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1842 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1843 _space_info[id].space()->bottom()); 1844 } 1845 1846 MutableSpace* const old_space = _space_info[old_space_id].space(); 1847 const size_t old_capacity = old_space->capacity_in_words(); 1848 if (old_space_total_live > old_capacity) { 1849 // XXX - should also try to expand 1850 maximum_compaction = true; 1851 } 1852 #ifndef PRODUCT 1853 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1854 provoke_split(maximum_compaction); 1855 } 1856 #endif // #ifndef PRODUCT 1857 1858 // Old generations. 1859 summarize_space(old_space_id, maximum_compaction); 1860 1861 // Summarize the remaining spaces in the young gen. The initial target space 1862 // is the old gen. If a space does not fit entirely into the target, then the 1863 // remainder is compacted into the space itself and that space becomes the new 1864 // target. 1865 SpaceId dst_space_id = old_space_id; 1866 HeapWord* dst_space_end = old_space->end(); 1867 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1868 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1869 const MutableSpace* space = _space_info[id].space(); 1870 const size_t live = pointer_delta(_space_info[id].new_top(), 1871 space->bottom()); 1872 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1873 1874 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1875 SpaceId(id), space->bottom(), space->top());) 1876 if (live > 0 && live <= available) { 1877 // All the live data will fit. 1878 bool done = _summary_data.summarize(_space_info[id].split_info(), 1879 space->bottom(), space->top(), 1880 NULL, 1881 *new_top_addr, dst_space_end, 1882 new_top_addr); 1883 assert(done, "space must fit into old gen"); 1884 1885 // Reset the new_top value for the space. 1886 _space_info[id].set_new_top(space->bottom()); 1887 } else if (live > 0) { 1888 // Attempt to fit part of the source space into the target space. 1889 HeapWord* next_src_addr = NULL; 1890 bool done = _summary_data.summarize(_space_info[id].split_info(), 1891 space->bottom(), space->top(), 1892 &next_src_addr, 1893 *new_top_addr, dst_space_end, 1894 new_top_addr); 1895 assert(!done, "space should not fit into old gen"); 1896 assert(next_src_addr != NULL, "sanity"); 1897 1898 // The source space becomes the new target, so the remainder is compacted 1899 // within the space itself. 1900 dst_space_id = SpaceId(id); 1901 dst_space_end = space->end(); 1902 new_top_addr = _space_info[id].new_top_addr(); 1903 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1904 space->bottom(), dst_space_end, 1905 SpaceId(id), next_src_addr, space->top());) 1906 done = _summary_data.summarize(_space_info[id].split_info(), 1907 next_src_addr, space->top(), 1908 NULL, 1909 space->bottom(), dst_space_end, 1910 new_top_addr); 1911 assert(done, "space must fit when compacted into itself"); 1912 assert(*new_top_addr <= space->top(), "usage should not grow"); 1913 } 1914 } 1915 1916 if (TraceParallelOldGCSummaryPhase) { 1917 tty->print_cr("summary_phase: after final summarization"); 1918 Universe::print(); 1919 NOT_PRODUCT(print_region_ranges()); 1920 if (Verbose) { 1921 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1922 } 1923 } 1924 } 1925 1926 // This method should contain all heap-specific policy for invoking a full 1927 // collection. invoke_no_policy() will only attempt to compact the heap; it 1928 // will do nothing further. If we need to bail out for policy reasons, scavenge 1929 // before full gc, or any other specialized behavior, it needs to be added here. 1930 // 1931 // Note that this method should only be called from the vm_thread while at a 1932 // safepoint. 1933 // 1934 // Note that the all_soft_refs_clear flag in the collector policy 1935 // may be true because this method can be called without intervening 1936 // activity. For example when the heap space is tight and full measure 1937 // are being taken to free space. 1938 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1939 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1940 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1941 "should be in vm thread"); 1942 1943 ParallelScavengeHeap* heap = gc_heap(); 1944 GCCause::Cause gc_cause = heap->gc_cause(); 1945 assert(!heap->is_gc_active(), "not reentrant"); 1946 1947 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1948 IsGCActiveMark mark; 1949 1950 if (ScavengeBeforeFullGC) { 1951 PSScavenge::invoke_no_policy(); 1952 } 1953 1954 const bool clear_all_soft_refs = 1955 heap->collector_policy()->should_clear_all_soft_refs(); 1956 1957 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1958 maximum_heap_compaction); 1959 } 1960 1961 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1962 size_t addr_region_index = addr_to_region_idx(addr); 1963 return region_index == addr_region_index; 1964 } 1965 1966 // This method contains no policy. You should probably 1967 // be calling invoke() instead. 1968 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1969 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1970 assert(ref_processor() != NULL, "Sanity"); 1971 1972 if (GC_locker::check_active_before_gc()) { 1973 return false; 1974 } 1975 1976 TimeStamp marking_start; 1977 TimeStamp compaction_start; 1978 TimeStamp collection_exit; 1979 1980 ParallelScavengeHeap* heap = gc_heap(); 1981 GCCause::Cause gc_cause = heap->gc_cause(); 1982 PSYoungGen* young_gen = heap->young_gen(); 1983 PSOldGen* old_gen = heap->old_gen(); 1984 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1985 1986 // The scope of casr should end after code that can change 1987 // CollectorPolicy::_should_clear_all_soft_refs. 1988 ClearedAllSoftRefs casr(maximum_heap_compaction, 1989 heap->collector_policy()); 1990 1991 if (ZapUnusedHeapArea) { 1992 // Save information needed to minimize mangling 1993 heap->record_gen_tops_before_GC(); 1994 } 1995 1996 heap->pre_full_gc_dump(); 1997 1998 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 1999 2000 // Make sure data structures are sane, make the heap parsable, and do other 2001 // miscellaneous bookkeeping. 2002 PreGCValues pre_gc_values; 2003 pre_compact(&pre_gc_values); 2004 2005 // Get the compaction manager reserved for the VM thread. 2006 ParCompactionManager* const vmthread_cm = 2007 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2008 2009 // Place after pre_compact() where the number of invocations is incremented. 2010 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2011 2012 { 2013 ResourceMark rm; 2014 HandleMark hm; 2015 2016 // Set the number of GC threads to be used in this collection 2017 gc_task_manager()->set_active_gang(); 2018 gc_task_manager()->task_idle_workers(); 2019 heap->set_par_threads(gc_task_manager()->active_workers()); 2020 2021 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2022 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2023 TraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 2024 TraceCollectorStats tcs(counters()); 2025 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2026 2027 if (TraceGen1Time) accumulated_time()->start(); 2028 2029 // Let the size policy know we're starting 2030 size_policy->major_collection_begin(); 2031 2032 CodeCache::gc_prologue(); 2033 Threads::gc_prologue(); 2034 2035 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2036 2037 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 2038 ref_processor()->setup_policy(maximum_heap_compaction); 2039 2040 bool marked_for_unloading = false; 2041 2042 marking_start.update(); 2043 marking_phase(vmthread_cm, maximum_heap_compaction); 2044 2045 #ifndef PRODUCT 2046 if (TraceParallelOldGCMarkingPhase) { 2047 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " 2048 "cas_by_another %d", 2049 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), 2050 mark_bitmap()->cas_by_another()); 2051 } 2052 #endif // #ifndef PRODUCT 2053 2054 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2055 && gc_cause == GCCause::_java_lang_system_gc; 2056 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2057 2058 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2059 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2060 2061 // adjust_roots() updates Universe::_intArrayKlassObj which is 2062 // needed by the compaction for filling holes in the dense prefix. 2063 adjust_roots(); 2064 2065 compaction_start.update(); 2066 compact(); 2067 2068 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2069 // done before resizing. 2070 post_compact(); 2071 2072 // Let the size policy know we're done 2073 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2074 2075 if (UseAdaptiveSizePolicy) { 2076 if (PrintAdaptiveSizePolicy) { 2077 gclog_or_tty->print("AdaptiveSizeStart: "); 2078 gclog_or_tty->stamp(); 2079 gclog_or_tty->print_cr(" collection: %d ", 2080 heap->total_collections()); 2081 if (Verbose) { 2082 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d", 2083 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2084 } 2085 } 2086 2087 // Don't check if the size_policy is ready here. Let 2088 // the size_policy check that internally. 2089 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2090 ((gc_cause != GCCause::_java_lang_system_gc) || 2091 UseAdaptiveSizePolicyWithSystemGC)) { 2092 // Calculate optimal free space amounts 2093 assert(young_gen->max_size() > 2094 young_gen->from_space()->capacity_in_bytes() + 2095 young_gen->to_space()->capacity_in_bytes(), 2096 "Sizes of space in young gen are out-of-bounds"); 2097 size_t max_eden_size = young_gen->max_size() - 2098 young_gen->from_space()->capacity_in_bytes() - 2099 young_gen->to_space()->capacity_in_bytes(); 2100 size_policy->compute_generation_free_space( 2101 young_gen->used_in_bytes(), 2102 young_gen->eden_space()->used_in_bytes(), 2103 old_gen->used_in_bytes(), 2104 young_gen->eden_space()->capacity_in_bytes(), 2105 old_gen->max_gen_size(), 2106 max_eden_size, 2107 true /* full gc*/, 2108 gc_cause, 2109 heap->collector_policy()); 2110 2111 heap->resize_old_gen( 2112 size_policy->calculated_old_free_size_in_bytes()); 2113 2114 // Don't resize the young generation at an major collection. A 2115 // desired young generation size may have been calculated but 2116 // resizing the young generation complicates the code because the 2117 // resizing of the old generation may have moved the boundary 2118 // between the young generation and the old generation. Let the 2119 // young generation resizing happen at the minor collections. 2120 } 2121 if (PrintAdaptiveSizePolicy) { 2122 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2123 heap->total_collections()); 2124 } 2125 } 2126 2127 if (UsePerfData) { 2128 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2129 counters->update_counters(); 2130 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2131 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2132 } 2133 2134 heap->resize_all_tlabs(); 2135 2136 // Resize the metaspace capactiy after a collection 2137 MetaspaceGC::compute_new_size(); 2138 2139 if (TraceGen1Time) accumulated_time()->stop(); 2140 2141 if (PrintGC) { 2142 if (PrintGCDetails) { 2143 // No GC timestamp here. This is after GC so it would be confusing. 2144 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2145 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2146 heap->print_heap_change(pre_gc_values.heap_used()); 2147 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2148 } else { 2149 heap->print_heap_change(pre_gc_values.heap_used()); 2150 } 2151 } 2152 2153 // Track memory usage and detect low memory 2154 MemoryService::track_memory_usage(); 2155 heap->update_counters(); 2156 gc_task_manager()->release_idle_workers(); 2157 } 2158 2159 #ifdef ASSERT 2160 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2161 ParCompactionManager* const cm = 2162 ParCompactionManager::manager_array(int(i)); 2163 assert(cm->marking_stack()->is_empty(), "should be empty"); 2164 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2165 } 2166 #endif // ASSERT 2167 2168 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2169 HandleMark hm; // Discard invalid handles created during verification 2170 Universe::verify(" VerifyAfterGC:"); 2171 } 2172 2173 // Re-verify object start arrays 2174 if (VerifyObjectStartArray && 2175 VerifyAfterGC) { 2176 old_gen->verify_object_start_array(); 2177 } 2178 2179 if (ZapUnusedHeapArea) { 2180 old_gen->object_space()->check_mangled_unused_area_complete(); 2181 } 2182 2183 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2184 2185 collection_exit.update(); 2186 2187 heap->print_heap_after_gc(); 2188 if (PrintGCTaskTimeStamps) { 2189 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2190 INT64_FORMAT, 2191 marking_start.ticks(), compaction_start.ticks(), 2192 collection_exit.ticks()); 2193 gc_task_manager()->print_task_time_stamps(); 2194 } 2195 2196 heap->post_full_gc_dump(); 2197 2198 #ifdef TRACESPINNING 2199 ParallelTaskTerminator::print_termination_counts(); 2200 #endif 2201 2202 return true; 2203 } 2204 2205 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2206 PSYoungGen* young_gen, 2207 PSOldGen* old_gen) { 2208 MutableSpace* const eden_space = young_gen->eden_space(); 2209 assert(!eden_space->is_empty(), "eden must be non-empty"); 2210 assert(young_gen->virtual_space()->alignment() == 2211 old_gen->virtual_space()->alignment(), "alignments do not match"); 2212 2213 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2214 return false; 2215 } 2216 2217 // Both generations must be completely committed. 2218 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2219 return false; 2220 } 2221 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2222 return false; 2223 } 2224 2225 // Figure out how much to take from eden. Include the average amount promoted 2226 // in the total; otherwise the next young gen GC will simply bail out to a 2227 // full GC. 2228 const size_t alignment = old_gen->virtual_space()->alignment(); 2229 const size_t eden_used = eden_space->used_in_bytes(); 2230 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2231 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2232 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2233 2234 if (absorb_size >= eden_capacity) { 2235 return false; // Must leave some space in eden. 2236 } 2237 2238 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2239 if (new_young_size < young_gen->min_gen_size()) { 2240 return false; // Respect young gen minimum size. 2241 } 2242 2243 if (TraceAdaptiveGCBoundary && Verbose) { 2244 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2245 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2246 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2247 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2248 absorb_size / K, 2249 eden_capacity / K, (eden_capacity - absorb_size) / K, 2250 young_gen->from_space()->used_in_bytes() / K, 2251 young_gen->to_space()->used_in_bytes() / K, 2252 young_gen->capacity_in_bytes() / K, new_young_size / K); 2253 } 2254 2255 // Fill the unused part of the old gen. 2256 MutableSpace* const old_space = old_gen->object_space(); 2257 HeapWord* const unused_start = old_space->top(); 2258 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2259 2260 if (unused_words > 0) { 2261 if (unused_words < CollectedHeap::min_fill_size()) { 2262 return false; // If the old gen cannot be filled, must give up. 2263 } 2264 CollectedHeap::fill_with_objects(unused_start, unused_words); 2265 } 2266 2267 // Take the live data from eden and set both top and end in the old gen to 2268 // eden top. (Need to set end because reset_after_change() mangles the region 2269 // from end to virtual_space->high() in debug builds). 2270 HeapWord* const new_top = eden_space->top(); 2271 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2272 absorb_size); 2273 young_gen->reset_after_change(); 2274 old_space->set_top(new_top); 2275 old_space->set_end(new_top); 2276 old_gen->reset_after_change(); 2277 2278 // Update the object start array for the filler object and the data from eden. 2279 ObjectStartArray* const start_array = old_gen->start_array(); 2280 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2281 start_array->allocate_block(p); 2282 } 2283 2284 // Could update the promoted average here, but it is not typically updated at 2285 // full GCs and the value to use is unclear. Something like 2286 // 2287 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2288 2289 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2290 return true; 2291 } 2292 2293 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2294 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2295 "shouldn't return NULL"); 2296 return ParallelScavengeHeap::gc_task_manager(); 2297 } 2298 2299 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2300 bool maximum_heap_compaction) { 2301 // Recursively traverse all live objects and mark them 2302 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); 2303 2304 ParallelScavengeHeap* heap = gc_heap(); 2305 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2306 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2307 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2308 ParallelTaskTerminator terminator(active_gc_threads, qset); 2309 2310 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2311 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2312 2313 // Need new claim bits before marking starts. 2314 ClassLoaderDataGraph::clear_claimed_marks(); 2315 2316 { 2317 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); 2318 ParallelScavengeHeap::ParStrongRootsScope psrs; 2319 2320 GCTaskQueue* q = GCTaskQueue::create(); 2321 2322 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2323 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2324 // We scan the thread roots in parallel 2325 Threads::create_thread_roots_marking_tasks(q); 2326 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2327 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2328 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2329 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2330 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2331 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2332 2333 if (active_gc_threads > 1) { 2334 for (uint j = 0; j < active_gc_threads; j++) { 2335 q->enqueue(new StealMarkingTask(&terminator)); 2336 } 2337 } 2338 2339 gc_task_manager()->execute_and_wait(q); 2340 } 2341 2342 // Process reference objects found during marking 2343 { 2344 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); 2345 if (ref_processor()->processing_is_mt()) { 2346 RefProcTaskExecutor task_executor; 2347 ref_processor()->process_discovered_references( 2348 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2349 &task_executor); 2350 } else { 2351 ref_processor()->process_discovered_references( 2352 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL); 2353 } 2354 } 2355 2356 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); 2357 2358 // This is the point where the entire marking should have completed. 2359 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2360 2361 // Follow system dictionary roots and unload classes. 2362 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2363 2364 // Unload nmethods. 2365 CodeCache::do_unloading(is_alive_closure(), purged_class); 2366 2367 // Prune dead klasses from subklass/sibling/implementor lists. 2368 Klass::clean_weak_klass_links(is_alive_closure()); 2369 2370 // Delete entries for dead interned strings. 2371 StringTable::unlink(is_alive_closure()); 2372 2373 // Clean up unreferenced symbols in symbol table. 2374 SymbolTable::unlink(); 2375 } 2376 2377 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) { 2378 ClassLoaderData* cld = klass->class_loader_data(); 2379 // The actual processing of the klass is done when we 2380 // traverse the list of Klasses in the class loader data. 2381 PSParallelCompact::follow_class_loader(cm, cld); 2382 } 2383 2384 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) { 2385 ClassLoaderData* cld = klass->class_loader_data(); 2386 // The actual processing of the klass is done when we 2387 // traverse the list of Klasses in the class loader data. 2388 PSParallelCompact::adjust_class_loader(cm, cld); 2389 } 2390 2391 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2392 ClassLoaderData* cld) { 2393 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2394 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2395 2396 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2397 } 2398 2399 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm, 2400 ClassLoaderData* cld) { 2401 cld->oops_do(PSParallelCompact::adjust_pointer_closure(), 2402 PSParallelCompact::adjust_klass_closure(), 2403 true); 2404 } 2405 2406 // This should be moved to the shared markSweep code! 2407 class PSAlwaysTrueClosure: public BoolObjectClosure { 2408 public: 2409 void do_object(oop p) { ShouldNotReachHere(); } 2410 bool do_object_b(oop p) { return true; } 2411 }; 2412 static PSAlwaysTrueClosure always_true; 2413 2414 void PSParallelCompact::adjust_roots() { 2415 // Adjust the pointers to reflect the new locations 2416 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); 2417 2418 // Need new claim bits when tracing through and adjusting pointers. 2419 ClassLoaderDataGraph::clear_claimed_marks(); 2420 2421 // General strong roots. 2422 Universe::oops_do(adjust_pointer_closure()); 2423 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2424 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2425 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2426 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2427 FlatProfiler::oops_do(adjust_pointer_closure()); 2428 Management::oops_do(adjust_pointer_closure()); 2429 JvmtiExport::oops_do(adjust_pointer_closure()); 2430 // SO_AllClasses 2431 SystemDictionary::oops_do(adjust_pointer_closure()); 2432 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2433 2434 // Now adjust pointers in remaining weak roots. (All of which should 2435 // have been cleared if they pointed to non-surviving objects.) 2436 // Global (weak) JNI handles 2437 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2438 2439 CodeCache::oops_do(adjust_pointer_closure()); 2440 StringTable::oops_do(adjust_pointer_closure()); 2441 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2442 // Roots were visited so references into the young gen in roots 2443 // may have been scanned. Process them also. 2444 // Should the reference processor have a span that excludes 2445 // young gen objects? 2446 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2447 } 2448 2449 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2450 uint parallel_gc_threads) 2451 { 2452 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); 2453 2454 // Find the threads that are active 2455 unsigned int which = 0; 2456 2457 const uint task_count = MAX2(parallel_gc_threads, 1U); 2458 for (uint j = 0; j < task_count; j++) { 2459 q->enqueue(new DrainStacksCompactionTask(j)); 2460 ParCompactionManager::verify_region_list_empty(j); 2461 // Set the region stacks variables to "no" region stack values 2462 // so that they will be recognized and needing a region stack 2463 // in the stealing tasks if they do not get one by executing 2464 // a draining stack. 2465 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2466 cm->set_region_stack(NULL); 2467 cm->set_region_stack_index((uint)max_uintx); 2468 } 2469 ParCompactionManager::reset_recycled_stack_index(); 2470 2471 // Find all regions that are available (can be filled immediately) and 2472 // distribute them to the thread stacks. The iteration is done in reverse 2473 // order (high to low) so the regions will be removed in ascending order. 2474 2475 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2476 2477 size_t fillable_regions = 0; // A count for diagnostic purposes. 2478 // A region index which corresponds to the tasks created above. 2479 // "which" must be 0 <= which < task_count 2480 2481 which = 0; 2482 // id + 1 is used to test termination so unsigned can 2483 // be used with an old_space_id == 0. 2484 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2485 SpaceInfo* const space_info = _space_info + id; 2486 MutableSpace* const space = space_info->space(); 2487 HeapWord* const new_top = space_info->new_top(); 2488 2489 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2490 const size_t end_region = 2491 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2492 2493 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2494 if (sd.region(cur)->claim_unsafe()) { 2495 ParCompactionManager::region_list_push(which, cur); 2496 2497 if (TraceParallelOldGCCompactionPhase && Verbose) { 2498 const size_t count_mod_8 = fillable_regions & 7; 2499 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2500 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2501 if (count_mod_8 == 7) gclog_or_tty->cr(); 2502 } 2503 2504 NOT_PRODUCT(++fillable_regions;) 2505 2506 // Assign regions to tasks in round-robin fashion. 2507 if (++which == task_count) { 2508 assert(which <= parallel_gc_threads, 2509 "Inconsistent number of workers"); 2510 which = 0; 2511 } 2512 } 2513 } 2514 } 2515 2516 if (TraceParallelOldGCCompactionPhase) { 2517 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2518 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2519 } 2520 } 2521 2522 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2523 2524 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2525 uint parallel_gc_threads) { 2526 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); 2527 2528 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2529 2530 // Iterate over all the spaces adding tasks for updating 2531 // regions in the dense prefix. Assume that 1 gc thread 2532 // will work on opening the gaps and the remaining gc threads 2533 // will work on the dense prefix. 2534 unsigned int space_id; 2535 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2536 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2537 const MutableSpace* const space = _space_info[space_id].space(); 2538 2539 if (dense_prefix_end == space->bottom()) { 2540 // There is no dense prefix for this space. 2541 continue; 2542 } 2543 2544 // The dense prefix is before this region. 2545 size_t region_index_end_dense_prefix = 2546 sd.addr_to_region_idx(dense_prefix_end); 2547 RegionData* const dense_prefix_cp = 2548 sd.region(region_index_end_dense_prefix); 2549 assert(dense_prefix_end == space->end() || 2550 dense_prefix_cp->available() || 2551 dense_prefix_cp->claimed(), 2552 "The region after the dense prefix should always be ready to fill"); 2553 2554 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2555 2556 // Is there dense prefix work? 2557 size_t total_dense_prefix_regions = 2558 region_index_end_dense_prefix - region_index_start; 2559 // How many regions of the dense prefix should be given to 2560 // each thread? 2561 if (total_dense_prefix_regions > 0) { 2562 uint tasks_for_dense_prefix = 1; 2563 if (total_dense_prefix_regions <= 2564 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2565 // Don't over partition. This assumes that 2566 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2567 // so there are not many regions to process. 2568 tasks_for_dense_prefix = parallel_gc_threads; 2569 } else { 2570 // Over partition 2571 tasks_for_dense_prefix = parallel_gc_threads * 2572 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2573 } 2574 size_t regions_per_thread = total_dense_prefix_regions / 2575 tasks_for_dense_prefix; 2576 // Give each thread at least 1 region. 2577 if (regions_per_thread == 0) { 2578 regions_per_thread = 1; 2579 } 2580 2581 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2582 if (region_index_start >= region_index_end_dense_prefix) { 2583 break; 2584 } 2585 // region_index_end is not processed 2586 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2587 region_index_end_dense_prefix); 2588 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2589 region_index_start, 2590 region_index_end)); 2591 region_index_start = region_index_end; 2592 } 2593 } 2594 // This gets any part of the dense prefix that did not 2595 // fit evenly. 2596 if (region_index_start < region_index_end_dense_prefix) { 2597 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2598 region_index_start, 2599 region_index_end_dense_prefix)); 2600 } 2601 } 2602 } 2603 2604 void PSParallelCompact::enqueue_region_stealing_tasks( 2605 GCTaskQueue* q, 2606 ParallelTaskTerminator* terminator_ptr, 2607 uint parallel_gc_threads) { 2608 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); 2609 2610 // Once a thread has drained it's stack, it should try to steal regions from 2611 // other threads. 2612 if (parallel_gc_threads > 1) { 2613 for (uint j = 0; j < parallel_gc_threads; j++) { 2614 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2615 } 2616 } 2617 } 2618 2619 void PSParallelCompact::compact() { 2620 // trace("5"); 2621 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); 2622 2623 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2624 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2625 PSOldGen* old_gen = heap->old_gen(); 2626 old_gen->start_array()->reset(); 2627 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2628 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2629 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2630 ParallelTaskTerminator terminator(active_gc_threads, qset); 2631 2632 GCTaskQueue* q = GCTaskQueue::create(); 2633 enqueue_region_draining_tasks(q, active_gc_threads); 2634 enqueue_dense_prefix_tasks(q, active_gc_threads); 2635 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2636 2637 { 2638 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); 2639 2640 gc_task_manager()->execute_and_wait(q); 2641 2642 #ifdef ASSERT 2643 // Verify that all regions have been processed before the deferred updates. 2644 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2645 verify_complete(SpaceId(id)); 2646 } 2647 #endif 2648 } 2649 2650 { 2651 // Update the deferred objects, if any. Any compaction manager can be used. 2652 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); 2653 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2654 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2655 update_deferred_objects(cm, SpaceId(id)); 2656 } 2657 } 2658 } 2659 2660 #ifdef ASSERT 2661 void PSParallelCompact::verify_complete(SpaceId space_id) { 2662 // All Regions between space bottom() to new_top() should be marked as filled 2663 // and all Regions between new_top() and top() should be available (i.e., 2664 // should have been emptied). 2665 ParallelCompactData& sd = summary_data(); 2666 SpaceInfo si = _space_info[space_id]; 2667 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2668 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2669 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2670 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2671 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2672 2673 bool issued_a_warning = false; 2674 2675 size_t cur_region; 2676 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2677 const RegionData* const c = sd.region(cur_region); 2678 if (!c->completed()) { 2679 warning("region " SIZE_FORMAT " not filled: " 2680 "destination_count=" SIZE_FORMAT, 2681 cur_region, c->destination_count()); 2682 issued_a_warning = true; 2683 } 2684 } 2685 2686 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2687 const RegionData* const c = sd.region(cur_region); 2688 if (!c->available()) { 2689 warning("region " SIZE_FORMAT " not empty: " 2690 "destination_count=" SIZE_FORMAT, 2691 cur_region, c->destination_count()); 2692 issued_a_warning = true; 2693 } 2694 } 2695 2696 if (issued_a_warning) { 2697 print_region_ranges(); 2698 } 2699 } 2700 #endif // #ifdef ASSERT 2701 2702 // Update interior oops in the ranges of regions [beg_region, end_region). 2703 void 2704 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2705 SpaceId space_id, 2706 size_t beg_region, 2707 size_t end_region) { 2708 ParallelCompactData& sd = summary_data(); 2709 ParMarkBitMap* const mbm = mark_bitmap(); 2710 2711 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2712 HeapWord* const end_addr = sd.region_to_addr(end_region); 2713 assert(beg_region <= end_region, "bad region range"); 2714 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2715 2716 #ifdef ASSERT 2717 // Claim the regions to avoid triggering an assert when they are marked as 2718 // filled. 2719 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2720 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2721 } 2722 #endif // #ifdef ASSERT 2723 2724 if (beg_addr != space(space_id)->bottom()) { 2725 // Find the first live object or block of dead space that *starts* in this 2726 // range of regions. If a partial object crosses onto the region, skip it; 2727 // it will be marked for 'deferred update' when the object head is 2728 // processed. If dead space crosses onto the region, it is also skipped; it 2729 // will be filled when the prior region is processed. If neither of those 2730 // apply, the first word in the region is the start of a live object or dead 2731 // space. 2732 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2733 const RegionData* const cp = sd.region(beg_region); 2734 if (cp->partial_obj_size() != 0) { 2735 beg_addr = sd.partial_obj_end(beg_region); 2736 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2737 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2738 } 2739 } 2740 2741 if (beg_addr < end_addr) { 2742 // A live object or block of dead space starts in this range of Regions. 2743 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2744 2745 // Create closures and iterate. 2746 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2747 FillClosure fill_closure(cm, space_id); 2748 ParMarkBitMap::IterationStatus status; 2749 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2750 dense_prefix_end); 2751 if (status == ParMarkBitMap::incomplete) { 2752 update_closure.do_addr(update_closure.source()); 2753 } 2754 } 2755 2756 // Mark the regions as filled. 2757 RegionData* const beg_cp = sd.region(beg_region); 2758 RegionData* const end_cp = sd.region(end_region); 2759 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2760 cp->set_completed(); 2761 } 2762 } 2763 2764 // Return the SpaceId for the space containing addr. If addr is not in the 2765 // heap, last_space_id is returned. In debug mode it expects the address to be 2766 // in the heap and asserts such. 2767 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2768 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2769 2770 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2771 if (_space_info[id].space()->contains(addr)) { 2772 return SpaceId(id); 2773 } 2774 } 2775 2776 assert(false, "no space contains the addr"); 2777 return last_space_id; 2778 } 2779 2780 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2781 SpaceId id) { 2782 assert(id < last_space_id, "bad space id"); 2783 2784 ParallelCompactData& sd = summary_data(); 2785 const SpaceInfo* const space_info = _space_info + id; 2786 ObjectStartArray* const start_array = space_info->start_array(); 2787 2788 const MutableSpace* const space = space_info->space(); 2789 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2790 HeapWord* const beg_addr = space_info->dense_prefix(); 2791 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2792 2793 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2794 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2795 const RegionData* cur_region; 2796 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2797 HeapWord* const addr = cur_region->deferred_obj_addr(); 2798 if (addr != NULL) { 2799 if (start_array != NULL) { 2800 start_array->allocate_block(addr); 2801 } 2802 oop(addr)->update_contents(cm); 2803 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 2804 } 2805 } 2806 } 2807 2808 // Skip over count live words starting from beg, and return the address of the 2809 // next live word. Unless marked, the word corresponding to beg is assumed to 2810 // be dead. Callers must either ensure beg does not correspond to the middle of 2811 // an object, or account for those live words in some other way. Callers must 2812 // also ensure that there are enough live words in the range [beg, end) to skip. 2813 HeapWord* 2814 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2815 { 2816 assert(count > 0, "sanity"); 2817 2818 ParMarkBitMap* m = mark_bitmap(); 2819 idx_t bits_to_skip = m->words_to_bits(count); 2820 idx_t cur_beg = m->addr_to_bit(beg); 2821 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2822 2823 do { 2824 cur_beg = m->find_obj_beg(cur_beg, search_end); 2825 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2826 const size_t obj_bits = cur_end - cur_beg + 1; 2827 if (obj_bits > bits_to_skip) { 2828 return m->bit_to_addr(cur_beg + bits_to_skip); 2829 } 2830 bits_to_skip -= obj_bits; 2831 cur_beg = cur_end + 1; 2832 } while (bits_to_skip > 0); 2833 2834 // Skipping the desired number of words landed just past the end of an object. 2835 // Find the start of the next object. 2836 cur_beg = m->find_obj_beg(cur_beg, search_end); 2837 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2838 return m->bit_to_addr(cur_beg); 2839 } 2840 2841 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2842 SpaceId src_space_id, 2843 size_t src_region_idx) 2844 { 2845 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2846 2847 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2848 if (split_info.dest_region_addr() == dest_addr) { 2849 // The partial object ending at the split point contains the first word to 2850 // be copied to dest_addr. 2851 return split_info.first_src_addr(); 2852 } 2853 2854 const ParallelCompactData& sd = summary_data(); 2855 ParMarkBitMap* const bitmap = mark_bitmap(); 2856 const size_t RegionSize = ParallelCompactData::RegionSize; 2857 2858 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2859 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2860 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2861 HeapWord* const src_region_destination = src_region_ptr->destination(); 2862 2863 assert(dest_addr >= src_region_destination, "wrong src region"); 2864 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2865 2866 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2867 HeapWord* const src_region_end = src_region_beg + RegionSize; 2868 2869 HeapWord* addr = src_region_beg; 2870 if (dest_addr == src_region_destination) { 2871 // Return the first live word in the source region. 2872 if (partial_obj_size == 0) { 2873 addr = bitmap->find_obj_beg(addr, src_region_end); 2874 assert(addr < src_region_end, "no objects start in src region"); 2875 } 2876 return addr; 2877 } 2878 2879 // Must skip some live data. 2880 size_t words_to_skip = dest_addr - src_region_destination; 2881 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2882 2883 if (partial_obj_size >= words_to_skip) { 2884 // All the live words to skip are part of the partial object. 2885 addr += words_to_skip; 2886 if (partial_obj_size == words_to_skip) { 2887 // Find the first live word past the partial object. 2888 addr = bitmap->find_obj_beg(addr, src_region_end); 2889 assert(addr < src_region_end, "wrong src region"); 2890 } 2891 return addr; 2892 } 2893 2894 // Skip over the partial object (if any). 2895 if (partial_obj_size != 0) { 2896 words_to_skip -= partial_obj_size; 2897 addr += partial_obj_size; 2898 } 2899 2900 // Skip over live words due to objects that start in the region. 2901 addr = skip_live_words(addr, src_region_end, words_to_skip); 2902 assert(addr < src_region_end, "wrong src region"); 2903 return addr; 2904 } 2905 2906 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2907 SpaceId src_space_id, 2908 size_t beg_region, 2909 HeapWord* end_addr) 2910 { 2911 ParallelCompactData& sd = summary_data(); 2912 2913 #ifdef ASSERT 2914 MutableSpace* const src_space = _space_info[src_space_id].space(); 2915 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2916 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2917 "src_space_id does not match beg_addr"); 2918 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2919 "src_space_id does not match end_addr"); 2920 #endif // #ifdef ASSERT 2921 2922 RegionData* const beg = sd.region(beg_region); 2923 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2924 2925 // Regions up to new_top() are enqueued if they become available. 2926 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2927 RegionData* const enqueue_end = 2928 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2929 2930 for (RegionData* cur = beg; cur < end; ++cur) { 2931 assert(cur->data_size() > 0, "region must have live data"); 2932 cur->decrement_destination_count(); 2933 if (cur < enqueue_end && cur->available() && cur->claim()) { 2934 cm->push_region(sd.region(cur)); 2935 } 2936 } 2937 } 2938 2939 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2940 SpaceId& src_space_id, 2941 HeapWord*& src_space_top, 2942 HeapWord* end_addr) 2943 { 2944 typedef ParallelCompactData::RegionData RegionData; 2945 2946 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2947 const size_t region_size = ParallelCompactData::RegionSize; 2948 2949 size_t src_region_idx = 0; 2950 2951 // Skip empty regions (if any) up to the top of the space. 2952 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2953 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2954 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2955 const RegionData* const top_region_ptr = 2956 sd.addr_to_region_ptr(top_aligned_up); 2957 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2958 ++src_region_ptr; 2959 } 2960 2961 if (src_region_ptr < top_region_ptr) { 2962 // The next source region is in the current space. Update src_region_idx 2963 // and the source address to match src_region_ptr. 2964 src_region_idx = sd.region(src_region_ptr); 2965 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2966 if (src_region_addr > closure.source()) { 2967 closure.set_source(src_region_addr); 2968 } 2969 return src_region_idx; 2970 } 2971 2972 // Switch to a new source space and find the first non-empty region. 2973 unsigned int space_id = src_space_id + 1; 2974 assert(space_id < last_space_id, "not enough spaces"); 2975 2976 HeapWord* const destination = closure.destination(); 2977 2978 do { 2979 MutableSpace* space = _space_info[space_id].space(); 2980 HeapWord* const bottom = space->bottom(); 2981 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2982 2983 // Iterate over the spaces that do not compact into themselves. 2984 if (bottom_cp->destination() != bottom) { 2985 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2986 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2987 2988 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2989 if (src_cp->live_obj_size() > 0) { 2990 // Found it. 2991 assert(src_cp->destination() == destination, 2992 "first live obj in the space must match the destination"); 2993 assert(src_cp->partial_obj_size() == 0, 2994 "a space cannot begin with a partial obj"); 2995 2996 src_space_id = SpaceId(space_id); 2997 src_space_top = space->top(); 2998 const size_t src_region_idx = sd.region(src_cp); 2999 closure.set_source(sd.region_to_addr(src_region_idx)); 3000 return src_region_idx; 3001 } else { 3002 assert(src_cp->data_size() == 0, "sanity"); 3003 } 3004 } 3005 } 3006 } while (++space_id < last_space_id); 3007 3008 assert(false, "no source region was found"); 3009 return 0; 3010 } 3011 3012 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3013 { 3014 typedef ParMarkBitMap::IterationStatus IterationStatus; 3015 const size_t RegionSize = ParallelCompactData::RegionSize; 3016 ParMarkBitMap* const bitmap = mark_bitmap(); 3017 ParallelCompactData& sd = summary_data(); 3018 RegionData* const region_ptr = sd.region(region_idx); 3019 3020 // Get the items needed to construct the closure. 3021 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3022 SpaceId dest_space_id = space_id(dest_addr); 3023 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3024 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3025 assert(dest_addr < new_top, "sanity"); 3026 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3027 3028 // Get the source region and related info. 3029 size_t src_region_idx = region_ptr->source_region(); 3030 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3031 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3032 3033 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3034 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3035 3036 // Adjust src_region_idx to prepare for decrementing destination counts (the 3037 // destination count is not decremented when a region is copied to itself). 3038 if (src_region_idx == region_idx) { 3039 src_region_idx += 1; 3040 } 3041 3042 if (bitmap->is_unmarked(closure.source())) { 3043 // The first source word is in the middle of an object; copy the remainder 3044 // of the object or as much as will fit. The fact that pointer updates were 3045 // deferred will be noted when the object header is processed. 3046 HeapWord* const old_src_addr = closure.source(); 3047 closure.copy_partial_obj(); 3048 if (closure.is_full()) { 3049 decrement_destination_counts(cm, src_space_id, src_region_idx, 3050 closure.source()); 3051 region_ptr->set_deferred_obj_addr(NULL); 3052 region_ptr->set_completed(); 3053 return; 3054 } 3055 3056 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3057 if (sd.region_align_down(old_src_addr) != end_addr) { 3058 // The partial object was copied from more than one source region. 3059 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3060 3061 // Move to the next source region, possibly switching spaces as well. All 3062 // args except end_addr may be modified. 3063 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3064 end_addr); 3065 } 3066 } 3067 3068 do { 3069 HeapWord* const cur_addr = closure.source(); 3070 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3071 src_space_top); 3072 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3073 3074 if (status == ParMarkBitMap::incomplete) { 3075 // The last obj that starts in the source region does not end in the 3076 // region. 3077 assert(closure.source() < end_addr, "sanity"); 3078 HeapWord* const obj_beg = closure.source(); 3079 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3080 src_space_top); 3081 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3082 if (obj_end < range_end) { 3083 // The end was found; the entire object will fit. 3084 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3085 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3086 } else { 3087 // The end was not found; the object will not fit. 3088 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3089 status = ParMarkBitMap::would_overflow; 3090 } 3091 } 3092 3093 if (status == ParMarkBitMap::would_overflow) { 3094 // The last object did not fit. Note that interior oop updates were 3095 // deferred, then copy enough of the object to fill the region. 3096 region_ptr->set_deferred_obj_addr(closure.destination()); 3097 status = closure.copy_until_full(); // copies from closure.source() 3098 3099 decrement_destination_counts(cm, src_space_id, src_region_idx, 3100 closure.source()); 3101 region_ptr->set_completed(); 3102 return; 3103 } 3104 3105 if (status == ParMarkBitMap::full) { 3106 decrement_destination_counts(cm, src_space_id, src_region_idx, 3107 closure.source()); 3108 region_ptr->set_deferred_obj_addr(NULL); 3109 region_ptr->set_completed(); 3110 return; 3111 } 3112 3113 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3114 3115 // Move to the next source region, possibly switching spaces as well. All 3116 // args except end_addr may be modified. 3117 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3118 end_addr); 3119 } while (true); 3120 } 3121 3122 void 3123 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3124 const MutableSpace* sp = space(space_id); 3125 if (sp->is_empty()) { 3126 return; 3127 } 3128 3129 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3130 ParMarkBitMap* const bitmap = mark_bitmap(); 3131 HeapWord* const dp_addr = dense_prefix(space_id); 3132 HeapWord* beg_addr = sp->bottom(); 3133 HeapWord* end_addr = sp->top(); 3134 3135 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3136 3137 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3138 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3139 if (beg_region < dp_region) { 3140 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3141 } 3142 3143 // The destination of the first live object that starts in the region is one 3144 // past the end of the partial object entering the region (if any). 3145 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3146 HeapWord* const new_top = _space_info[space_id].new_top(); 3147 assert(new_top >= dest_addr, "bad new_top value"); 3148 const size_t words = pointer_delta(new_top, dest_addr); 3149 3150 if (words > 0) { 3151 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3152 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3153 3154 ParMarkBitMap::IterationStatus status; 3155 status = bitmap->iterate(&closure, dest_addr, end_addr); 3156 assert(status == ParMarkBitMap::full, "iteration not complete"); 3157 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3158 "live objects skipped because closure is full"); 3159 } 3160 } 3161 3162 jlong PSParallelCompact::millis_since_last_gc() { 3163 // We need a monotonically non-deccreasing time in ms but 3164 // os::javaTimeMillis() does not guarantee monotonicity. 3165 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3166 jlong ret_val = now - _time_of_last_gc; 3167 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3168 if (ret_val < 0) { 3169 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);) 3170 return 0; 3171 } 3172 return ret_val; 3173 } 3174 3175 void PSParallelCompact::reset_millis_since_last_gc() { 3176 // We need a monotonically non-deccreasing time in ms but 3177 // os::javaTimeMillis() does not guarantee monotonicity. 3178 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3179 } 3180 3181 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3182 { 3183 if (source() != destination()) { 3184 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3185 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3186 } 3187 update_state(words_remaining()); 3188 assert(is_full(), "sanity"); 3189 return ParMarkBitMap::full; 3190 } 3191 3192 void MoveAndUpdateClosure::copy_partial_obj() 3193 { 3194 size_t words = words_remaining(); 3195 3196 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3197 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3198 if (end_addr < range_end) { 3199 words = bitmap()->obj_size(source(), end_addr); 3200 } 3201 3202 // This test is necessary; if omitted, the pointer updates to a partial object 3203 // that crosses the dense prefix boundary could be overwritten. 3204 if (source() != destination()) { 3205 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3206 Copy::aligned_conjoint_words(source(), destination(), words); 3207 } 3208 update_state(words); 3209 } 3210 3211 ParMarkBitMapClosure::IterationStatus 3212 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3213 assert(destination() != NULL, "sanity"); 3214 assert(bitmap()->obj_size(addr) == words, "bad size"); 3215 3216 _source = addr; 3217 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3218 destination(), "wrong destination"); 3219 3220 if (words > words_remaining()) { 3221 return ParMarkBitMap::would_overflow; 3222 } 3223 3224 // The start_array must be updated even if the object is not moving. 3225 if (_start_array != NULL) { 3226 _start_array->allocate_block(destination()); 3227 } 3228 3229 if (destination() != source()) { 3230 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3231 Copy::aligned_conjoint_words(source(), destination(), words); 3232 } 3233 3234 oop moved_oop = (oop) destination(); 3235 moved_oop->update_contents(compaction_manager()); 3236 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3237 3238 update_state(words); 3239 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3240 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3241 } 3242 3243 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3244 ParCompactionManager* cm, 3245 PSParallelCompact::SpaceId space_id) : 3246 ParMarkBitMapClosure(mbm, cm), 3247 _space_id(space_id), 3248 _start_array(PSParallelCompact::start_array(space_id)) 3249 { 3250 } 3251 3252 // Updates the references in the object to their new values. 3253 ParMarkBitMapClosure::IterationStatus 3254 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3255 do_addr(addr); 3256 return ParMarkBitMap::incomplete; 3257 }