1 /* 2 * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp" 30 #include "gc_implementation/parallelScavenge/generationSizer.hpp" 31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp" 32 #include "gc_implementation/parallelScavenge/pcTasks.hpp" 33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp" 34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp" 35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp" 36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp" 37 #include "gc_implementation/parallelScavenge/psOldGen.hpp" 38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp" 39 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 40 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 41 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 42 #include "gc_implementation/shared/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 NMTTrackOp op(NMTTrackOp::TypeOp); 394 op.execute_op((address)rs.base(), 0, mtGC); 395 396 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 397 if (vspace != 0) { 398 if (vspace->expand_by(bytes)) { 399 return vspace; 400 } 401 delete vspace; 402 // Release memory reserved in the space. 403 rs.release(); 404 } 405 406 return 0; 407 } 408 409 bool ParallelCompactData::initialize_region_data(size_t region_size) 410 { 411 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 412 _region_vspace = create_vspace(count, sizeof(RegionData)); 413 if (_region_vspace != 0) { 414 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 415 _region_count = count; 416 return true; 417 } 418 return false; 419 } 420 421 void ParallelCompactData::clear() 422 { 423 memset(_region_data, 0, _region_vspace->committed_size()); 424 } 425 426 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 427 assert(beg_region <= _region_count, "beg_region out of range"); 428 assert(end_region <= _region_count, "end_region out of range"); 429 430 const size_t region_cnt = end_region - beg_region; 431 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 432 } 433 434 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 435 { 436 const RegionData* cur_cp = region(region_idx); 437 const RegionData* const end_cp = region(region_count() - 1); 438 439 HeapWord* result = region_to_addr(region_idx); 440 if (cur_cp < end_cp) { 441 do { 442 result += cur_cp->partial_obj_size(); 443 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 444 } 445 return result; 446 } 447 448 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 449 { 450 const size_t obj_ofs = pointer_delta(addr, _region_start); 451 const size_t beg_region = obj_ofs >> Log2RegionSize; 452 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 453 454 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 455 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 456 457 if (beg_region == end_region) { 458 // All in one region. 459 _region_data[beg_region].add_live_obj(len); 460 return; 461 } 462 463 // First region. 464 const size_t beg_ofs = region_offset(addr); 465 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 466 467 Klass* klass = ((oop)addr)->klass(); 468 // Middle regions--completely spanned by this object. 469 for (size_t region = beg_region + 1; region < end_region; ++region) { 470 _region_data[region].set_partial_obj_size(RegionSize); 471 _region_data[region].set_partial_obj_addr(addr); 472 } 473 474 // Last region. 475 const size_t end_ofs = region_offset(addr + len - 1); 476 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 477 _region_data[end_region].set_partial_obj_addr(addr); 478 } 479 480 void 481 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 482 { 483 assert(region_offset(beg) == 0, "not RegionSize aligned"); 484 assert(region_offset(end) == 0, "not RegionSize aligned"); 485 486 size_t cur_region = addr_to_region_idx(beg); 487 const size_t end_region = addr_to_region_idx(end); 488 HeapWord* addr = beg; 489 while (cur_region < end_region) { 490 _region_data[cur_region].set_destination(addr); 491 _region_data[cur_region].set_destination_count(0); 492 _region_data[cur_region].set_source_region(cur_region); 493 _region_data[cur_region].set_data_location(addr); 494 495 // Update live_obj_size so the region appears completely full. 496 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 497 _region_data[cur_region].set_live_obj_size(live_size); 498 499 ++cur_region; 500 addr += RegionSize; 501 } 502 } 503 504 // Find the point at which a space can be split and, if necessary, record the 505 // split point. 506 // 507 // If the current src region (which overflowed the destination space) doesn't 508 // have a partial object, the split point is at the beginning of the current src 509 // region (an "easy" split, no extra bookkeeping required). 510 // 511 // If the current src region has a partial object, the split point is in the 512 // region where that partial object starts (call it the split_region). If 513 // split_region has a partial object, then the split point is just after that 514 // partial object (a "hard" split where we have to record the split data and 515 // zero the partial_obj_size field). With a "hard" split, we know that the 516 // partial_obj ends within split_region because the partial object that caused 517 // the overflow starts in split_region. If split_region doesn't have a partial 518 // obj, then the split is at the beginning of split_region (another "easy" 519 // split). 520 HeapWord* 521 ParallelCompactData::summarize_split_space(size_t src_region, 522 SplitInfo& split_info, 523 HeapWord* destination, 524 HeapWord* target_end, 525 HeapWord** target_next) 526 { 527 assert(destination <= target_end, "sanity"); 528 assert(destination + _region_data[src_region].data_size() > target_end, 529 "region should not fit into target space"); 530 assert(is_region_aligned(target_end), "sanity"); 531 532 size_t split_region = src_region; 533 HeapWord* split_destination = destination; 534 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 535 536 if (destination + partial_obj_size > target_end) { 537 // The split point is just after the partial object (if any) in the 538 // src_region that contains the start of the object that overflowed the 539 // destination space. 540 // 541 // Find the start of the "overflow" object and set split_region to the 542 // region containing it. 543 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 544 split_region = addr_to_region_idx(overflow_obj); 545 546 // Clear the source_region field of all destination regions whose first word 547 // came from data after the split point (a non-null source_region field 548 // implies a region must be filled). 549 // 550 // An alternative to the simple loop below: clear during post_compact(), 551 // which uses memcpy instead of individual stores, and is easy to 552 // parallelize. (The downside is that it clears the entire RegionData 553 // object as opposed to just one field.) 554 // 555 // post_compact() would have to clear the summary data up to the highest 556 // address that was written during the summary phase, which would be 557 // 558 // max(top, max(new_top, clear_top)) 559 // 560 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 561 // to target_end. 562 const RegionData* const sr = region(split_region); 563 const size_t beg_idx = 564 addr_to_region_idx(region_align_up(sr->destination() + 565 sr->partial_obj_size())); 566 const size_t end_idx = addr_to_region_idx(target_end); 567 568 if (TraceParallelOldGCSummaryPhase) { 569 gclog_or_tty->print_cr("split: clearing source_region field in [" 570 SIZE_FORMAT ", " SIZE_FORMAT ")", 571 beg_idx, end_idx); 572 } 573 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 574 _region_data[idx].set_source_region(0); 575 } 576 577 // Set split_destination and partial_obj_size to reflect the split region. 578 split_destination = sr->destination(); 579 partial_obj_size = sr->partial_obj_size(); 580 } 581 582 // The split is recorded only if a partial object extends onto the region. 583 if (partial_obj_size != 0) { 584 _region_data[split_region].set_partial_obj_size(0); 585 split_info.record(split_region, partial_obj_size, split_destination); 586 } 587 588 // Setup the continuation addresses. 589 *target_next = split_destination + partial_obj_size; 590 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 591 592 if (TraceParallelOldGCSummaryPhase) { 593 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 594 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 595 " pos=" SIZE_FORMAT, 596 split_type, source_next, split_region, 597 partial_obj_size); 598 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 599 " tn=" PTR_FORMAT, 600 split_type, split_destination, 601 addr_to_region_idx(split_destination), 602 *target_next); 603 604 if (partial_obj_size != 0) { 605 HeapWord* const po_beg = split_info.destination(); 606 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 607 gclog_or_tty->print_cr("%s split: " 608 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 609 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 610 split_type, 611 po_beg, addr_to_region_idx(po_beg), 612 po_end, addr_to_region_idx(po_end)); 613 } 614 } 615 616 return source_next; 617 } 618 619 bool ParallelCompactData::summarize(SplitInfo& split_info, 620 HeapWord* source_beg, HeapWord* source_end, 621 HeapWord** source_next, 622 HeapWord* target_beg, HeapWord* target_end, 623 HeapWord** target_next) 624 { 625 if (TraceParallelOldGCSummaryPhase) { 626 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 627 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 628 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 629 source_beg, source_end, source_next_val, 630 target_beg, target_end, *target_next); 631 } 632 633 size_t cur_region = addr_to_region_idx(source_beg); 634 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 635 636 HeapWord *dest_addr = target_beg; 637 while (cur_region < end_region) { 638 // The destination must be set even if the region has no data. 639 _region_data[cur_region].set_destination(dest_addr); 640 641 size_t words = _region_data[cur_region].data_size(); 642 if (words > 0) { 643 // If cur_region does not fit entirely into the target space, find a point 644 // at which the source space can be 'split' so that part is copied to the 645 // target space and the rest is copied elsewhere. 646 if (dest_addr + words > target_end) { 647 assert(source_next != NULL, "source_next is NULL when splitting"); 648 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 649 target_end, target_next); 650 return false; 651 } 652 653 // Compute the destination_count for cur_region, and if necessary, update 654 // source_region for a destination region. The source_region field is 655 // updated if cur_region is the first (left-most) region to be copied to a 656 // destination region. 657 // 658 // The destination_count calculation is a bit subtle. A region that has 659 // data that compacts into itself does not count itself as a destination. 660 // This maintains the invariant that a zero count means the region is 661 // available and can be claimed and then filled. 662 uint destination_count = 0; 663 if (split_info.is_split(cur_region)) { 664 // The current region has been split: the partial object will be copied 665 // to one destination space and the remaining data will be copied to 666 // another destination space. Adjust the initial destination_count and, 667 // if necessary, set the source_region field if the partial object will 668 // cross a destination region boundary. 669 destination_count = split_info.destination_count(); 670 if (destination_count == 2) { 671 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 672 _region_data[dest_idx].set_source_region(cur_region); 673 } 674 } 675 676 HeapWord* const last_addr = dest_addr + words - 1; 677 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 678 const size_t dest_region_2 = addr_to_region_idx(last_addr); 679 680 // Initially assume that the destination regions will be the same and 681 // adjust the value below if necessary. Under this assumption, if 682 // cur_region == dest_region_2, then cur_region will be compacted 683 // completely into itself. 684 destination_count += cur_region == dest_region_2 ? 0 : 1; 685 if (dest_region_1 != dest_region_2) { 686 // Destination regions differ; adjust destination_count. 687 destination_count += 1; 688 // Data from cur_region will be copied to the start of dest_region_2. 689 _region_data[dest_region_2].set_source_region(cur_region); 690 } else if (region_offset(dest_addr) == 0) { 691 // Data from cur_region will be copied to the start of the destination 692 // region. 693 _region_data[dest_region_1].set_source_region(cur_region); 694 } 695 696 _region_data[cur_region].set_destination_count(destination_count); 697 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 698 dest_addr += words; 699 } 700 701 ++cur_region; 702 } 703 704 *target_next = dest_addr; 705 return true; 706 } 707 708 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 709 assert(addr != NULL, "Should detect NULL oop earlier"); 710 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); 711 #ifdef ASSERT 712 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { 713 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); 714 } 715 #endif 716 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); 717 718 // Region covering the object. 719 size_t region_index = addr_to_region_idx(addr); 720 const RegionData* const region_ptr = region(region_index); 721 HeapWord* const region_addr = region_align_down(addr); 722 723 assert(addr < region_addr + RegionSize, "Region does not cover object"); 724 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check"); 725 726 HeapWord* result = region_ptr->destination(); 727 728 // If all the data in the region is live, then the new location of the object 729 // can be calculated from the destination of the region plus the offset of the 730 // object in the region. 731 if (region_ptr->data_size() == RegionSize) { 732 result += pointer_delta(addr, region_addr); 733 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 734 return result; 735 } 736 737 // The new location of the object is 738 // region destination + 739 // size of the partial object extending onto the region + 740 // sizes of the live objects in the Region that are to the left of addr 741 const size_t partial_obj_size = region_ptr->partial_obj_size(); 742 HeapWord* const search_start = region_addr + partial_obj_size; 743 744 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 745 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 746 747 result += partial_obj_size + live_to_left; 748 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 749 return result; 750 } 751 752 #ifdef ASSERT 753 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 754 { 755 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 756 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 757 for (const size_t* p = beg; p < end; ++p) { 758 assert(*p == 0, "not zero"); 759 } 760 } 761 762 void ParallelCompactData::verify_clear() 763 { 764 verify_clear(_region_vspace); 765 } 766 #endif // #ifdef ASSERT 767 768 #ifdef NOT_PRODUCT 769 ParallelCompactData::RegionData* debug_region(size_t region_index) { 770 ParallelCompactData& sd = PSParallelCompact::summary_data(); 771 return sd.region(region_index); 772 } 773 #endif 774 775 elapsedTimer PSParallelCompact::_accumulated_time; 776 unsigned int PSParallelCompact::_total_invocations = 0; 777 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 778 jlong PSParallelCompact::_time_of_last_gc = 0; 779 CollectorCounters* PSParallelCompact::_counters = NULL; 780 ParMarkBitMap PSParallelCompact::_mark_bitmap; 781 ParallelCompactData PSParallelCompact::_summary_data; 782 783 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 784 785 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } 786 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 787 788 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 789 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 790 791 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure; 792 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure; 793 794 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p); } 795 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); } 796 797 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 798 799 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { 800 mark_and_push(_compaction_manager, p); 801 } 802 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 803 804 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) { 805 klass->oops_do(_mark_and_push_closure); 806 } 807 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) { 808 klass->oops_do(&PSParallelCompact::_adjust_pointer_closure); 809 } 810 811 void PSParallelCompact::post_initialize() { 812 ParallelScavengeHeap* heap = gc_heap(); 813 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 814 815 MemRegion mr = heap->reserved_region(); 816 _ref_processor = 817 new ReferenceProcessor(mr, // span 818 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing 819 (int) ParallelGCThreads, // mt processing degree 820 true, // mt discovery 821 (int) ParallelGCThreads, // mt discovery degree 822 true, // atomic_discovery 823 &_is_alive_closure, // non-header is alive closure 824 false); // write barrier for next field updates 825 _counters = new CollectorCounters("PSParallelCompact", 1); 826 827 // Initialize static fields in ParCompactionManager. 828 ParCompactionManager::initialize(mark_bitmap()); 829 } 830 831 bool PSParallelCompact::initialize() { 832 ParallelScavengeHeap* heap = gc_heap(); 833 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 834 MemRegion mr = heap->reserved_region(); 835 836 // Was the old gen get allocated successfully? 837 if (!heap->old_gen()->is_allocated()) { 838 return false; 839 } 840 841 initialize_space_info(); 842 initialize_dead_wood_limiter(); 843 844 if (!_mark_bitmap.initialize(mr)) { 845 vm_shutdown_during_initialization("Unable to allocate bit map for " 846 "parallel garbage collection for the requested heap size."); 847 return false; 848 } 849 850 if (!_summary_data.initialize(mr)) { 851 vm_shutdown_during_initialization("Unable to allocate tables for " 852 "parallel garbage collection for the requested heap size."); 853 return false; 854 } 855 856 return true; 857 } 858 859 void PSParallelCompact::initialize_space_info() 860 { 861 memset(&_space_info, 0, sizeof(_space_info)); 862 863 ParallelScavengeHeap* heap = gc_heap(); 864 PSYoungGen* young_gen = heap->young_gen(); 865 866 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 867 _space_info[eden_space_id].set_space(young_gen->eden_space()); 868 _space_info[from_space_id].set_space(young_gen->from_space()); 869 _space_info[to_space_id].set_space(young_gen->to_space()); 870 871 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 872 } 873 874 void PSParallelCompact::initialize_dead_wood_limiter() 875 { 876 const size_t max = 100; 877 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 878 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 879 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 880 DEBUG_ONLY(_dwl_initialized = true;) 881 _dwl_adjustment = normal_distribution(1.0); 882 } 883 884 // Simple class for storing info about the heap at the start of GC, to be used 885 // after GC for comparison/printing. 886 class PreGCValues { 887 public: 888 PreGCValues() { } 889 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 890 891 void fill(ParallelScavengeHeap* heap) { 892 _heap_used = heap->used(); 893 _young_gen_used = heap->young_gen()->used_in_bytes(); 894 _old_gen_used = heap->old_gen()->used_in_bytes(); 895 _metadata_used = MetaspaceAux::allocated_used_bytes(); 896 }; 897 898 size_t heap_used() const { return _heap_used; } 899 size_t young_gen_used() const { return _young_gen_used; } 900 size_t old_gen_used() const { return _old_gen_used; } 901 size_t metadata_used() const { return _metadata_used; } 902 903 private: 904 size_t _heap_used; 905 size_t _young_gen_used; 906 size_t _old_gen_used; 907 size_t _metadata_used; 908 }; 909 910 void 911 PSParallelCompact::clear_data_covering_space(SpaceId id) 912 { 913 // At this point, top is the value before GC, new_top() is the value that will 914 // be set at the end of GC. The marking bitmap is cleared to top; nothing 915 // should be marked above top. The summary data is cleared to the larger of 916 // top & new_top. 917 MutableSpace* const space = _space_info[id].space(); 918 HeapWord* const bot = space->bottom(); 919 HeapWord* const top = space->top(); 920 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 921 922 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 923 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 924 _mark_bitmap.clear_range(beg_bit, end_bit); 925 926 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 927 const size_t end_region = 928 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 929 _summary_data.clear_range(beg_region, end_region); 930 931 // Clear the data used to 'split' regions. 932 SplitInfo& split_info = _space_info[id].split_info(); 933 if (split_info.is_valid()) { 934 split_info.clear(); 935 } 936 DEBUG_ONLY(split_info.verify_clear();) 937 } 938 939 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 940 { 941 // Update the from & to space pointers in space_info, since they are swapped 942 // at each young gen gc. Do the update unconditionally (even though a 943 // promotion failure does not swap spaces) because an unknown number of minor 944 // collections will have swapped the spaces an unknown number of times. 945 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); 946 ParallelScavengeHeap* heap = gc_heap(); 947 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 948 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 949 950 pre_gc_values->fill(heap); 951 952 NOT_PRODUCT(_mark_bitmap.reset_counters()); 953 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 954 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 955 956 // Increment the invocation count 957 heap->increment_total_collections(true); 958 959 // We need to track unique mark sweep invocations as well. 960 _total_invocations++; 961 962 heap->print_heap_before_gc(); 963 964 // Fill in TLABs 965 heap->accumulate_statistics_all_tlabs(); 966 heap->ensure_parsability(true); // retire TLABs 967 968 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 969 HandleMark hm; // Discard invalid handles created during verification 970 Universe::verify(" VerifyBeforeGC:"); 971 } 972 973 // Verify object start arrays 974 if (VerifyObjectStartArray && 975 VerifyBeforeGC) { 976 heap->old_gen()->verify_object_start_array(); 977 } 978 979 DEBUG_ONLY(mark_bitmap()->verify_clear();) 980 DEBUG_ONLY(summary_data().verify_clear();) 981 982 // Have worker threads release resources the next time they run a task. 983 gc_task_manager()->release_all_resources(); 984 } 985 986 void PSParallelCompact::post_compact() 987 { 988 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); 989 990 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 991 // Clear the marking bitmap, summary data and split info. 992 clear_data_covering_space(SpaceId(id)); 993 // Update top(). Must be done after clearing the bitmap and summary data. 994 _space_info[id].publish_new_top(); 995 } 996 997 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 998 MutableSpace* const from_space = _space_info[from_space_id].space(); 999 MutableSpace* const to_space = _space_info[to_space_id].space(); 1000 1001 ParallelScavengeHeap* heap = gc_heap(); 1002 bool eden_empty = eden_space->is_empty(); 1003 if (!eden_empty) { 1004 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1005 heap->young_gen(), heap->old_gen()); 1006 } 1007 1008 // Update heap occupancy information which is used as input to the soft ref 1009 // clearing policy at the next gc. 1010 Universe::update_heap_info_at_gc(); 1011 1012 bool young_gen_empty = eden_empty && from_space->is_empty() && 1013 to_space->is_empty(); 1014 1015 BarrierSet* bs = heap->barrier_set(); 1016 if (bs->is_a(BarrierSet::ModRef)) { 1017 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1018 MemRegion old_mr = heap->old_gen()->reserved(); 1019 1020 if (young_gen_empty) { 1021 modBS->clear(MemRegion(old_mr.start(), old_mr.end())); 1022 } else { 1023 modBS->invalidate(MemRegion(old_mr.start(), old_mr.end())); 1024 } 1025 } 1026 1027 // Delete metaspaces for unloaded class loaders and clean up loader_data graph 1028 ClassLoaderDataGraph::purge(); 1029 MetaspaceAux::verify_metrics(); 1030 1031 Threads::gc_epilogue(); 1032 CodeCache::gc_epilogue(); 1033 JvmtiExport::gc_epilogue(); 1034 1035 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1036 1037 ref_processor()->enqueue_discovered_references(NULL); 1038 1039 if (ZapUnusedHeapArea) { 1040 heap->gen_mangle_unused_area(); 1041 } 1042 1043 // Update time of last GC 1044 reset_millis_since_last_gc(); 1045 } 1046 1047 HeapWord* 1048 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1049 bool maximum_compaction) 1050 { 1051 const size_t region_size = ParallelCompactData::RegionSize; 1052 const ParallelCompactData& sd = summary_data(); 1053 1054 const MutableSpace* const space = _space_info[id].space(); 1055 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1056 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1057 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1058 1059 // Skip full regions at the beginning of the space--they are necessarily part 1060 // of the dense prefix. 1061 size_t full_count = 0; 1062 const RegionData* cp; 1063 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1064 ++full_count; 1065 } 1066 1067 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1068 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1069 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1070 if (maximum_compaction || cp == end_cp || interval_ended) { 1071 _maximum_compaction_gc_num = total_invocations(); 1072 return sd.region_to_addr(cp); 1073 } 1074 1075 HeapWord* const new_top = _space_info[id].new_top(); 1076 const size_t space_live = pointer_delta(new_top, space->bottom()); 1077 const size_t space_used = space->used_in_words(); 1078 const size_t space_capacity = space->capacity_in_words(); 1079 1080 const double cur_density = double(space_live) / space_capacity; 1081 const double deadwood_density = 1082 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1083 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1084 1085 if (TraceParallelOldGCDensePrefix) { 1086 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1087 cur_density, deadwood_density, deadwood_goal); 1088 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1089 "space_cap=" SIZE_FORMAT, 1090 space_live, space_used, 1091 space_capacity); 1092 } 1093 1094 // XXX - Use binary search? 1095 HeapWord* dense_prefix = sd.region_to_addr(cp); 1096 const RegionData* full_cp = cp; 1097 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1098 while (cp < end_cp) { 1099 HeapWord* region_destination = cp->destination(); 1100 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1101 if (TraceParallelOldGCDensePrefix && Verbose) { 1102 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1103 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1104 sd.region(cp), region_destination, 1105 dense_prefix, cur_deadwood); 1106 } 1107 1108 if (cur_deadwood >= deadwood_goal) { 1109 // Found the region that has the correct amount of deadwood to the left. 1110 // This typically occurs after crossing a fairly sparse set of regions, so 1111 // iterate backwards over those sparse regions, looking for the region 1112 // that has the lowest density of live objects 'to the right.' 1113 size_t space_to_left = sd.region(cp) * region_size; 1114 size_t live_to_left = space_to_left - cur_deadwood; 1115 size_t space_to_right = space_capacity - space_to_left; 1116 size_t live_to_right = space_live - live_to_left; 1117 double density_to_right = double(live_to_right) / space_to_right; 1118 while (cp > full_cp) { 1119 --cp; 1120 const size_t prev_region_live_to_right = live_to_right - 1121 cp->data_size(); 1122 const size_t prev_region_space_to_right = space_to_right + region_size; 1123 double prev_region_density_to_right = 1124 double(prev_region_live_to_right) / prev_region_space_to_right; 1125 if (density_to_right <= prev_region_density_to_right) { 1126 return dense_prefix; 1127 } 1128 if (TraceParallelOldGCDensePrefix && Verbose) { 1129 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1130 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1131 prev_region_density_to_right); 1132 } 1133 dense_prefix -= region_size; 1134 live_to_right = prev_region_live_to_right; 1135 space_to_right = prev_region_space_to_right; 1136 density_to_right = prev_region_density_to_right; 1137 } 1138 return dense_prefix; 1139 } 1140 1141 dense_prefix += region_size; 1142 ++cp; 1143 } 1144 1145 return dense_prefix; 1146 } 1147 1148 #ifndef PRODUCT 1149 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1150 const SpaceId id, 1151 const bool maximum_compaction, 1152 HeapWord* const addr) 1153 { 1154 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1155 RegionData* const cp = summary_data().region(region_idx); 1156 const MutableSpace* const space = _space_info[id].space(); 1157 HeapWord* const new_top = _space_info[id].new_top(); 1158 1159 const size_t space_live = pointer_delta(new_top, space->bottom()); 1160 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1161 const size_t space_cap = space->capacity_in_words(); 1162 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1163 const size_t live_to_right = new_top - cp->destination(); 1164 const size_t dead_to_right = space->top() - addr - live_to_right; 1165 1166 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1167 "spl=" SIZE_FORMAT " " 1168 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1169 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1170 " ratio=%10.8f", 1171 algorithm, addr, region_idx, 1172 space_live, 1173 dead_to_left, dead_to_left_pct, 1174 dead_to_right, live_to_right, 1175 double(dead_to_right) / live_to_right); 1176 } 1177 #endif // #ifndef PRODUCT 1178 1179 // Return a fraction indicating how much of the generation can be treated as 1180 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1181 // based on the density of live objects in the generation to determine a limit, 1182 // which is then adjusted so the return value is min_percent when the density is 1183 // 1. 1184 // 1185 // The following table shows some return values for a different values of the 1186 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1187 // min_percent is 1. 1188 // 1189 // fraction allowed as dead wood 1190 // ----------------------------------------------------------------- 1191 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1192 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1193 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1194 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1195 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1196 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1197 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1198 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1199 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1200 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1201 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1202 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1203 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1204 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1205 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1206 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1207 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1208 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1209 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1210 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1211 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1212 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1213 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1214 1215 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1216 { 1217 assert(_dwl_initialized, "uninitialized"); 1218 1219 // The raw limit is the value of the normal distribution at x = density. 1220 const double raw_limit = normal_distribution(density); 1221 1222 // Adjust the raw limit so it becomes the minimum when the density is 1. 1223 // 1224 // First subtract the adjustment value (which is simply the precomputed value 1225 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1226 // Then add the minimum value, so the minimum is returned when the density is 1227 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1228 const double min = double(min_percent) / 100.0; 1229 const double limit = raw_limit - _dwl_adjustment + min; 1230 return MAX2(limit, 0.0); 1231 } 1232 1233 ParallelCompactData::RegionData* 1234 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1235 const RegionData* end) 1236 { 1237 const size_t region_size = ParallelCompactData::RegionSize; 1238 ParallelCompactData& sd = summary_data(); 1239 size_t left = sd.region(beg); 1240 size_t right = end > beg ? sd.region(end) - 1 : left; 1241 1242 // Binary search. 1243 while (left < right) { 1244 // Equivalent to (left + right) / 2, but does not overflow. 1245 const size_t middle = left + (right - left) / 2; 1246 RegionData* const middle_ptr = sd.region(middle); 1247 HeapWord* const dest = middle_ptr->destination(); 1248 HeapWord* const addr = sd.region_to_addr(middle); 1249 assert(dest != NULL, "sanity"); 1250 assert(dest <= addr, "must move left"); 1251 1252 if (middle > left && dest < addr) { 1253 right = middle - 1; 1254 } else if (middle < right && middle_ptr->data_size() == region_size) { 1255 left = middle + 1; 1256 } else { 1257 return middle_ptr; 1258 } 1259 } 1260 return sd.region(left); 1261 } 1262 1263 ParallelCompactData::RegionData* 1264 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1265 const RegionData* end, 1266 size_t dead_words) 1267 { 1268 ParallelCompactData& sd = summary_data(); 1269 size_t left = sd.region(beg); 1270 size_t right = end > beg ? sd.region(end) - 1 : left; 1271 1272 // Binary search. 1273 while (left < right) { 1274 // Equivalent to (left + right) / 2, but does not overflow. 1275 const size_t middle = left + (right - left) / 2; 1276 RegionData* const middle_ptr = sd.region(middle); 1277 HeapWord* const dest = middle_ptr->destination(); 1278 HeapWord* const addr = sd.region_to_addr(middle); 1279 assert(dest != NULL, "sanity"); 1280 assert(dest <= addr, "must move left"); 1281 1282 const size_t dead_to_left = pointer_delta(addr, dest); 1283 if (middle > left && dead_to_left > dead_words) { 1284 right = middle - 1; 1285 } else if (middle < right && dead_to_left < dead_words) { 1286 left = middle + 1; 1287 } else { 1288 return middle_ptr; 1289 } 1290 } 1291 return sd.region(left); 1292 } 1293 1294 // The result is valid during the summary phase, after the initial summarization 1295 // of each space into itself, and before final summarization. 1296 inline double 1297 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1298 HeapWord* const bottom, 1299 HeapWord* const top, 1300 HeapWord* const new_top) 1301 { 1302 ParallelCompactData& sd = summary_data(); 1303 1304 assert(cp != NULL, "sanity"); 1305 assert(bottom != NULL, "sanity"); 1306 assert(top != NULL, "sanity"); 1307 assert(new_top != NULL, "sanity"); 1308 assert(top >= new_top, "summary data problem?"); 1309 assert(new_top > bottom, "space is empty; should not be here"); 1310 assert(new_top >= cp->destination(), "sanity"); 1311 assert(top >= sd.region_to_addr(cp), "sanity"); 1312 1313 HeapWord* const destination = cp->destination(); 1314 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1315 const size_t compacted_region_live = pointer_delta(new_top, destination); 1316 const size_t compacted_region_used = pointer_delta(top, 1317 sd.region_to_addr(cp)); 1318 const size_t reclaimable = compacted_region_used - compacted_region_live; 1319 1320 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1321 return double(reclaimable) / divisor; 1322 } 1323 1324 // Return the address of the end of the dense prefix, a.k.a. the start of the 1325 // compacted region. The address is always on a region boundary. 1326 // 1327 // Completely full regions at the left are skipped, since no compaction can 1328 // occur in those regions. Then the maximum amount of dead wood to allow is 1329 // computed, based on the density (amount live / capacity) of the generation; 1330 // the region with approximately that amount of dead space to the left is 1331 // identified as the limit region. Regions between the last completely full 1332 // region and the limit region are scanned and the one that has the best 1333 // (maximum) reclaimed_ratio() is selected. 1334 HeapWord* 1335 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1336 bool maximum_compaction) 1337 { 1338 if (ParallelOldGCSplitALot) { 1339 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1340 // The value was chosen to provoke splitting a young gen space; use it. 1341 return _space_info[id].dense_prefix(); 1342 } 1343 } 1344 1345 const size_t region_size = ParallelCompactData::RegionSize; 1346 const ParallelCompactData& sd = summary_data(); 1347 1348 const MutableSpace* const space = _space_info[id].space(); 1349 HeapWord* const top = space->top(); 1350 HeapWord* const top_aligned_up = sd.region_align_up(top); 1351 HeapWord* const new_top = _space_info[id].new_top(); 1352 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1353 HeapWord* const bottom = space->bottom(); 1354 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1355 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1356 const RegionData* const new_top_cp = 1357 sd.addr_to_region_ptr(new_top_aligned_up); 1358 1359 // Skip full regions at the beginning of the space--they are necessarily part 1360 // of the dense prefix. 1361 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1362 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1363 space->is_empty(), "no dead space allowed to the left"); 1364 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1365 "region must have dead space"); 1366 1367 // The gc number is saved whenever a maximum compaction is done, and used to 1368 // determine when the maximum compaction interval has expired. This avoids 1369 // successive max compactions for different reasons. 1370 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1371 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1372 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1373 total_invocations() == HeapFirstMaximumCompactionCount; 1374 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1375 _maximum_compaction_gc_num = total_invocations(); 1376 return sd.region_to_addr(full_cp); 1377 } 1378 1379 const size_t space_live = pointer_delta(new_top, bottom); 1380 const size_t space_used = space->used_in_words(); 1381 const size_t space_capacity = space->capacity_in_words(); 1382 1383 const double density = double(space_live) / double(space_capacity); 1384 const size_t min_percent_free = MarkSweepDeadRatio; 1385 const double limiter = dead_wood_limiter(density, min_percent_free); 1386 const size_t dead_wood_max = space_used - space_live; 1387 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1388 dead_wood_max); 1389 1390 if (TraceParallelOldGCDensePrefix) { 1391 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1392 "space_cap=" SIZE_FORMAT, 1393 space_live, space_used, 1394 space_capacity); 1395 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1396 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1397 density, min_percent_free, limiter, 1398 dead_wood_max, dead_wood_limit); 1399 } 1400 1401 // Locate the region with the desired amount of dead space to the left. 1402 const RegionData* const limit_cp = 1403 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1404 1405 // Scan from the first region with dead space to the limit region and find the 1406 // one with the best (largest) reclaimed ratio. 1407 double best_ratio = 0.0; 1408 const RegionData* best_cp = full_cp; 1409 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1410 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1411 if (tmp_ratio > best_ratio) { 1412 best_cp = cp; 1413 best_ratio = tmp_ratio; 1414 } 1415 } 1416 1417 #if 0 1418 // Something to consider: if the region with the best ratio is 'close to' the 1419 // first region w/free space, choose the first region with free space 1420 // ("first-free"). The first-free region is usually near the start of the 1421 // heap, which means we are copying most of the heap already, so copy a bit 1422 // more to get complete compaction. 1423 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1424 _maximum_compaction_gc_num = total_invocations(); 1425 best_cp = full_cp; 1426 } 1427 #endif // #if 0 1428 1429 return sd.region_to_addr(best_cp); 1430 } 1431 1432 #ifndef PRODUCT 1433 void 1434 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1435 size_t words) 1436 { 1437 if (TraceParallelOldGCSummaryPhase) { 1438 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1439 SIZE_FORMAT, start, start + words, words); 1440 } 1441 1442 ObjectStartArray* const start_array = _space_info[id].start_array(); 1443 CollectedHeap::fill_with_objects(start, words); 1444 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1445 _mark_bitmap.mark_obj(p, words); 1446 _summary_data.add_obj(p, words); 1447 start_array->allocate_block(p); 1448 } 1449 } 1450 1451 void 1452 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1453 { 1454 ParallelCompactData& sd = summary_data(); 1455 MutableSpace* space = _space_info[id].space(); 1456 1457 // Find the source and destination start addresses. 1458 HeapWord* const src_addr = sd.region_align_down(start); 1459 HeapWord* dst_addr; 1460 if (src_addr < start) { 1461 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1462 } else if (src_addr > space->bottom()) { 1463 // The start (the original top() value) is aligned to a region boundary so 1464 // the associated region does not have a destination. Compute the 1465 // destination from the previous region. 1466 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1467 dst_addr = cp->destination() + cp->data_size(); 1468 } else { 1469 // Filling the entire space. 1470 dst_addr = space->bottom(); 1471 } 1472 assert(dst_addr != NULL, "sanity"); 1473 1474 // Update the summary data. 1475 bool result = _summary_data.summarize(_space_info[id].split_info(), 1476 src_addr, space->top(), NULL, 1477 dst_addr, space->end(), 1478 _space_info[id].new_top_addr()); 1479 assert(result, "should not fail: bad filler object size"); 1480 } 1481 1482 void 1483 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1484 { 1485 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1486 return; 1487 } 1488 1489 MutableSpace* const space = _space_info[id].space(); 1490 if (space->is_empty()) { 1491 HeapWord* b = space->bottom(); 1492 HeapWord* t = b + space->capacity_in_words() / 2; 1493 space->set_top(t); 1494 if (ZapUnusedHeapArea) { 1495 space->set_top_for_allocations(); 1496 } 1497 1498 size_t min_size = CollectedHeap::min_fill_size(); 1499 size_t obj_len = min_size; 1500 while (b + obj_len <= t) { 1501 CollectedHeap::fill_with_object(b, obj_len); 1502 mark_bitmap()->mark_obj(b, obj_len); 1503 summary_data().add_obj(b, obj_len); 1504 b += obj_len; 1505 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1506 } 1507 if (b < t) { 1508 // The loop didn't completely fill to t (top); adjust top downward. 1509 space->set_top(b); 1510 if (ZapUnusedHeapArea) { 1511 space->set_top_for_allocations(); 1512 } 1513 } 1514 1515 HeapWord** nta = _space_info[id].new_top_addr(); 1516 bool result = summary_data().summarize(_space_info[id].split_info(), 1517 space->bottom(), space->top(), NULL, 1518 space->bottom(), space->end(), nta); 1519 assert(result, "space must fit into itself"); 1520 } 1521 } 1522 1523 void 1524 PSParallelCompact::provoke_split(bool & max_compaction) 1525 { 1526 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1527 return; 1528 } 1529 1530 const size_t region_size = ParallelCompactData::RegionSize; 1531 ParallelCompactData& sd = summary_data(); 1532 1533 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1534 MutableSpace* const from_space = _space_info[from_space_id].space(); 1535 const size_t eden_live = pointer_delta(eden_space->top(), 1536 _space_info[eden_space_id].new_top()); 1537 const size_t from_live = pointer_delta(from_space->top(), 1538 _space_info[from_space_id].new_top()); 1539 1540 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1541 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1542 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1543 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1544 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1545 1546 // Choose the space to split; need at least 2 regions live (or fillable). 1547 SpaceId id; 1548 MutableSpace* space; 1549 size_t live_words; 1550 size_t fill_words; 1551 if (eden_live + eden_fillable >= region_size * 2) { 1552 id = eden_space_id; 1553 space = eden_space; 1554 live_words = eden_live; 1555 fill_words = eden_fillable; 1556 } else if (from_live + from_fillable >= region_size * 2) { 1557 id = from_space_id; 1558 space = from_space; 1559 live_words = from_live; 1560 fill_words = from_fillable; 1561 } else { 1562 return; // Give up. 1563 } 1564 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1565 1566 if (live_words < region_size * 2) { 1567 // Fill from top() to end() w/live objects of mixed sizes. 1568 HeapWord* const fill_start = space->top(); 1569 live_words += fill_words; 1570 1571 space->set_top(fill_start + fill_words); 1572 if (ZapUnusedHeapArea) { 1573 space->set_top_for_allocations(); 1574 } 1575 1576 HeapWord* cur_addr = fill_start; 1577 while (fill_words > 0) { 1578 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1579 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1580 if (fill_words - cur_size < min_fill_size) { 1581 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1582 } 1583 1584 CollectedHeap::fill_with_object(cur_addr, cur_size); 1585 mark_bitmap()->mark_obj(cur_addr, cur_size); 1586 sd.add_obj(cur_addr, cur_size); 1587 1588 cur_addr += cur_size; 1589 fill_words -= cur_size; 1590 } 1591 1592 summarize_new_objects(id, fill_start); 1593 } 1594 1595 max_compaction = false; 1596 1597 // Manipulate the old gen so that it has room for about half of the live data 1598 // in the target young gen space (live_words / 2). 1599 id = old_space_id; 1600 space = _space_info[id].space(); 1601 const size_t free_at_end = space->free_in_words(); 1602 const size_t free_target = align_object_size(live_words / 2); 1603 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1604 1605 if (free_at_end >= free_target + min_fill_size) { 1606 // Fill space above top() and set the dense prefix so everything survives. 1607 HeapWord* const fill_start = space->top(); 1608 const size_t fill_size = free_at_end - free_target; 1609 space->set_top(space->top() + fill_size); 1610 if (ZapUnusedHeapArea) { 1611 space->set_top_for_allocations(); 1612 } 1613 fill_with_live_objects(id, fill_start, fill_size); 1614 summarize_new_objects(id, fill_start); 1615 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1616 } else if (dead + free_at_end > free_target) { 1617 // Find a dense prefix that makes the right amount of space available. 1618 HeapWord* cur = sd.region_align_down(space->top()); 1619 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1620 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1621 while (dead_to_right < free_target) { 1622 cur -= region_size; 1623 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1624 dead_to_right = pointer_delta(space->end(), cur_destination); 1625 } 1626 _space_info[id].set_dense_prefix(cur); 1627 } 1628 } 1629 #endif // #ifndef PRODUCT 1630 1631 void PSParallelCompact::summarize_spaces_quick() 1632 { 1633 for (unsigned int i = 0; i < last_space_id; ++i) { 1634 const MutableSpace* space = _space_info[i].space(); 1635 HeapWord** nta = _space_info[i].new_top_addr(); 1636 bool result = _summary_data.summarize(_space_info[i].split_info(), 1637 space->bottom(), space->top(), NULL, 1638 space->bottom(), space->end(), nta); 1639 assert(result, "space must fit into itself"); 1640 _space_info[i].set_dense_prefix(space->bottom()); 1641 } 1642 1643 #ifndef PRODUCT 1644 if (ParallelOldGCSplitALot) { 1645 provoke_split_fill_survivor(to_space_id); 1646 } 1647 #endif // #ifndef PRODUCT 1648 } 1649 1650 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1651 { 1652 HeapWord* const dense_prefix_end = dense_prefix(id); 1653 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1654 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1655 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1656 // Only enough dead space is filled so that any remaining dead space to the 1657 // left is larger than the minimum filler object. (The remainder is filled 1658 // during the copy/update phase.) 1659 // 1660 // The size of the dead space to the right of the boundary is not a 1661 // concern, since compaction will be able to use whatever space is 1662 // available. 1663 // 1664 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1665 // surrounds the space to be filled with an object. 1666 // 1667 // In the 32-bit VM, each bit represents two 32-bit words: 1668 // +---+ 1669 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1670 // end_bits: ... x x x | 0 | || 0 x x ... 1671 // +---+ 1672 // 1673 // In the 64-bit VM, each bit represents one 64-bit word: 1674 // +------------+ 1675 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1676 // end_bits: ... x x 1 | 0 || 0 | x x ... 1677 // +------------+ 1678 // +-------+ 1679 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1680 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1681 // +-------+ 1682 // +-----------+ 1683 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1684 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1685 // +-----------+ 1686 // +-------+ 1687 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1688 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1689 // +-------+ 1690 1691 // Initially assume case a, c or e will apply. 1692 size_t obj_len = CollectedHeap::min_fill_size(); 1693 HeapWord* obj_beg = dense_prefix_end - obj_len; 1694 1695 #ifdef _LP64 1696 if (MinObjAlignment > 1) { // object alignment > heap word size 1697 // Cases a, c or e. 1698 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1699 // Case b above. 1700 obj_beg = dense_prefix_end - 1; 1701 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1702 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1703 // Case d above. 1704 obj_beg = dense_prefix_end - 3; 1705 obj_len = 3; 1706 } 1707 #endif // #ifdef _LP64 1708 1709 CollectedHeap::fill_with_object(obj_beg, obj_len); 1710 _mark_bitmap.mark_obj(obj_beg, obj_len); 1711 _summary_data.add_obj(obj_beg, obj_len); 1712 assert(start_array(id) != NULL, "sanity"); 1713 start_array(id)->allocate_block(obj_beg); 1714 } 1715 } 1716 1717 void 1718 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1719 { 1720 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1721 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1722 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1723 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1724 cur->set_source_region(0); 1725 } 1726 } 1727 1728 void 1729 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1730 { 1731 assert(id < last_space_id, "id out of range"); 1732 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1733 ParallelOldGCSplitALot && id == old_space_id, 1734 "should have been reset in summarize_spaces_quick()"); 1735 1736 const MutableSpace* space = _space_info[id].space(); 1737 if (_space_info[id].new_top() != space->bottom()) { 1738 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1739 _space_info[id].set_dense_prefix(dense_prefix_end); 1740 1741 #ifndef PRODUCT 1742 if (TraceParallelOldGCDensePrefix) { 1743 print_dense_prefix_stats("ratio", id, maximum_compaction, 1744 dense_prefix_end); 1745 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1746 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1747 } 1748 #endif // #ifndef PRODUCT 1749 1750 // Recompute the summary data, taking into account the dense prefix. If 1751 // every last byte will be reclaimed, then the existing summary data which 1752 // compacts everything can be left in place. 1753 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1754 // If dead space crosses the dense prefix boundary, it is (at least 1755 // partially) filled with a dummy object, marked live and added to the 1756 // summary data. This simplifies the copy/update phase and must be done 1757 // before the final locations of objects are determined, to prevent 1758 // leaving a fragment of dead space that is too small to fill. 1759 fill_dense_prefix_end(id); 1760 1761 // Compute the destination of each Region, and thus each object. 1762 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1763 _summary_data.summarize(_space_info[id].split_info(), 1764 dense_prefix_end, space->top(), NULL, 1765 dense_prefix_end, space->end(), 1766 _space_info[id].new_top_addr()); 1767 } 1768 } 1769 1770 if (TraceParallelOldGCSummaryPhase) { 1771 const size_t region_size = ParallelCompactData::RegionSize; 1772 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1773 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1774 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1775 HeapWord* const new_top = _space_info[id].new_top(); 1776 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1777 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1778 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1779 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1780 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1781 id, space->capacity_in_words(), dense_prefix_end, 1782 dp_region, dp_words / region_size, 1783 cr_words / region_size, new_top); 1784 } 1785 } 1786 1787 #ifndef PRODUCT 1788 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1789 HeapWord* dst_beg, HeapWord* dst_end, 1790 SpaceId src_space_id, 1791 HeapWord* src_beg, HeapWord* src_end) 1792 { 1793 if (TraceParallelOldGCSummaryPhase) { 1794 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1795 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1796 SIZE_FORMAT "-" SIZE_FORMAT " " 1797 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1798 SIZE_FORMAT "-" SIZE_FORMAT, 1799 src_space_id, space_names[src_space_id], 1800 dst_space_id, space_names[dst_space_id], 1801 src_beg, src_end, 1802 _summary_data.addr_to_region_idx(src_beg), 1803 _summary_data.addr_to_region_idx(src_end), 1804 dst_beg, dst_end, 1805 _summary_data.addr_to_region_idx(dst_beg), 1806 _summary_data.addr_to_region_idx(dst_end)); 1807 } 1808 } 1809 #endif // #ifndef PRODUCT 1810 1811 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1812 bool maximum_compaction) 1813 { 1814 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); 1815 // trace("2"); 1816 1817 #ifdef ASSERT 1818 if (TraceParallelOldGCMarkingPhase) { 1819 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1820 "add_obj_bytes=" SIZE_FORMAT, 1821 add_obj_count, add_obj_size * HeapWordSize); 1822 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1823 "mark_bitmap_bytes=" SIZE_FORMAT, 1824 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1825 } 1826 #endif // #ifdef ASSERT 1827 1828 // Quick summarization of each space into itself, to see how much is live. 1829 summarize_spaces_quick(); 1830 1831 if (TraceParallelOldGCSummaryPhase) { 1832 tty->print_cr("summary_phase: after summarizing each space to self"); 1833 Universe::print(); 1834 NOT_PRODUCT(print_region_ranges()); 1835 if (Verbose) { 1836 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1837 } 1838 } 1839 1840 // The amount of live data that will end up in old space (assuming it fits). 1841 size_t old_space_total_live = 0; 1842 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1843 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1844 _space_info[id].space()->bottom()); 1845 } 1846 1847 MutableSpace* const old_space = _space_info[old_space_id].space(); 1848 const size_t old_capacity = old_space->capacity_in_words(); 1849 if (old_space_total_live > old_capacity) { 1850 // XXX - should also try to expand 1851 maximum_compaction = true; 1852 } 1853 #ifndef PRODUCT 1854 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1855 provoke_split(maximum_compaction); 1856 } 1857 #endif // #ifndef PRODUCT 1858 1859 // Old generations. 1860 summarize_space(old_space_id, maximum_compaction); 1861 1862 // Summarize the remaining spaces in the young gen. The initial target space 1863 // is the old gen. If a space does not fit entirely into the target, then the 1864 // remainder is compacted into the space itself and that space becomes the new 1865 // target. 1866 SpaceId dst_space_id = old_space_id; 1867 HeapWord* dst_space_end = old_space->end(); 1868 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1869 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1870 const MutableSpace* space = _space_info[id].space(); 1871 const size_t live = pointer_delta(_space_info[id].new_top(), 1872 space->bottom()); 1873 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1874 1875 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1876 SpaceId(id), space->bottom(), space->top());) 1877 if (live > 0 && live <= available) { 1878 // All the live data will fit. 1879 bool done = _summary_data.summarize(_space_info[id].split_info(), 1880 space->bottom(), space->top(), 1881 NULL, 1882 *new_top_addr, dst_space_end, 1883 new_top_addr); 1884 assert(done, "space must fit into old gen"); 1885 1886 // Reset the new_top value for the space. 1887 _space_info[id].set_new_top(space->bottom()); 1888 } else if (live > 0) { 1889 // Attempt to fit part of the source space into the target space. 1890 HeapWord* next_src_addr = NULL; 1891 bool done = _summary_data.summarize(_space_info[id].split_info(), 1892 space->bottom(), space->top(), 1893 &next_src_addr, 1894 *new_top_addr, dst_space_end, 1895 new_top_addr); 1896 assert(!done, "space should not fit into old gen"); 1897 assert(next_src_addr != NULL, "sanity"); 1898 1899 // The source space becomes the new target, so the remainder is compacted 1900 // within the space itself. 1901 dst_space_id = SpaceId(id); 1902 dst_space_end = space->end(); 1903 new_top_addr = _space_info[id].new_top_addr(); 1904 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1905 space->bottom(), dst_space_end, 1906 SpaceId(id), next_src_addr, space->top());) 1907 done = _summary_data.summarize(_space_info[id].split_info(), 1908 next_src_addr, space->top(), 1909 NULL, 1910 space->bottom(), dst_space_end, 1911 new_top_addr); 1912 assert(done, "space must fit when compacted into itself"); 1913 assert(*new_top_addr <= space->top(), "usage should not grow"); 1914 } 1915 } 1916 1917 if (TraceParallelOldGCSummaryPhase) { 1918 tty->print_cr("summary_phase: after final summarization"); 1919 Universe::print(); 1920 NOT_PRODUCT(print_region_ranges()); 1921 if (Verbose) { 1922 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1923 } 1924 } 1925 } 1926 1927 // This method should contain all heap-specific policy for invoking a full 1928 // collection. invoke_no_policy() will only attempt to compact the heap; it 1929 // will do nothing further. If we need to bail out for policy reasons, scavenge 1930 // before full gc, or any other specialized behavior, it needs to be added here. 1931 // 1932 // Note that this method should only be called from the vm_thread while at a 1933 // safepoint. 1934 // 1935 // Note that the all_soft_refs_clear flag in the collector policy 1936 // may be true because this method can be called without intervening 1937 // activity. For example when the heap space is tight and full measure 1938 // are being taken to free space. 1939 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1940 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1941 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1942 "should be in vm thread"); 1943 1944 ParallelScavengeHeap* heap = gc_heap(); 1945 GCCause::Cause gc_cause = heap->gc_cause(); 1946 assert(!heap->is_gc_active(), "not reentrant"); 1947 1948 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1949 IsGCActiveMark mark; 1950 1951 if (ScavengeBeforeFullGC) { 1952 PSScavenge::invoke_no_policy(); 1953 } 1954 1955 const bool clear_all_soft_refs = 1956 heap->collector_policy()->should_clear_all_soft_refs(); 1957 1958 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1959 maximum_heap_compaction); 1960 } 1961 1962 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1963 size_t addr_region_index = addr_to_region_idx(addr); 1964 return region_index == addr_region_index; 1965 } 1966 1967 // This method contains no policy. You should probably 1968 // be calling invoke() instead. 1969 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1970 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1971 assert(ref_processor() != NULL, "Sanity"); 1972 1973 if (GC_locker::check_active_before_gc()) { 1974 return false; 1975 } 1976 1977 TimeStamp marking_start; 1978 TimeStamp compaction_start; 1979 TimeStamp collection_exit; 1980 1981 ParallelScavengeHeap* heap = gc_heap(); 1982 GCCause::Cause gc_cause = heap->gc_cause(); 1983 PSYoungGen* young_gen = heap->young_gen(); 1984 PSOldGen* old_gen = heap->old_gen(); 1985 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1986 1987 // The scope of casr should end after code that can change 1988 // CollectorPolicy::_should_clear_all_soft_refs. 1989 ClearedAllSoftRefs casr(maximum_heap_compaction, 1990 heap->collector_policy()); 1991 1992 if (ZapUnusedHeapArea) { 1993 // Save information needed to minimize mangling 1994 heap->record_gen_tops_before_GC(); 1995 } 1996 1997 heap->pre_full_gc_dump(); 1998 1999 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2000 2001 // Make sure data structures are sane, make the heap parsable, and do other 2002 // miscellaneous bookkeeping. 2003 PreGCValues pre_gc_values; 2004 pre_compact(&pre_gc_values); 2005 2006 // Get the compaction manager reserved for the VM thread. 2007 ParCompactionManager* const vmthread_cm = 2008 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2009 2010 // Place after pre_compact() where the number of invocations is incremented. 2011 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2012 2013 { 2014 ResourceMark rm; 2015 HandleMark hm; 2016 2017 // Set the number of GC threads to be used in this collection 2018 gc_task_manager()->set_active_gang(); 2019 gc_task_manager()->task_idle_workers(); 2020 heap->set_par_threads(gc_task_manager()->active_workers()); 2021 2022 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2023 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2024 TraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 2025 TraceCollectorStats tcs(counters()); 2026 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause); 2027 2028 if (TraceGen1Time) accumulated_time()->start(); 2029 2030 // Let the size policy know we're starting 2031 size_policy->major_collection_begin(); 2032 2033 CodeCache::gc_prologue(); 2034 Threads::gc_prologue(); 2035 2036 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2037 2038 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 2039 ref_processor()->setup_policy(maximum_heap_compaction); 2040 2041 bool marked_for_unloading = false; 2042 2043 marking_start.update(); 2044 marking_phase(vmthread_cm, maximum_heap_compaction); 2045 2046 #ifndef PRODUCT 2047 if (TraceParallelOldGCMarkingPhase) { 2048 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " 2049 "cas_by_another %d", 2050 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), 2051 mark_bitmap()->cas_by_another()); 2052 } 2053 #endif // #ifndef PRODUCT 2054 2055 bool max_on_system_gc = UseMaximumCompactionOnSystemGC 2056 && gc_cause == GCCause::_java_lang_system_gc; 2057 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2058 2059 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2060 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2061 2062 // adjust_roots() updates Universe::_intArrayKlassObj which is 2063 // needed by the compaction for filling holes in the dense prefix. 2064 adjust_roots(); 2065 2066 compaction_start.update(); 2067 compact(); 2068 2069 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2070 // done before resizing. 2071 post_compact(); 2072 2073 // Let the size policy know we're done 2074 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2075 2076 if (UseAdaptiveSizePolicy) { 2077 if (PrintAdaptiveSizePolicy) { 2078 gclog_or_tty->print("AdaptiveSizeStart: "); 2079 gclog_or_tty->stamp(); 2080 gclog_or_tty->print_cr(" collection: %d ", 2081 heap->total_collections()); 2082 if (Verbose) { 2083 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d", 2084 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes()); 2085 } 2086 } 2087 2088 // Don't check if the size_policy is ready here. Let 2089 // the size_policy check that internally. 2090 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2091 ((gc_cause != GCCause::_java_lang_system_gc) || 2092 UseAdaptiveSizePolicyWithSystemGC)) { 2093 // Calculate optimal free space amounts 2094 assert(young_gen->max_size() > 2095 young_gen->from_space()->capacity_in_bytes() + 2096 young_gen->to_space()->capacity_in_bytes(), 2097 "Sizes of space in young gen are out-of-bounds"); 2098 size_t max_eden_size = young_gen->max_size() - 2099 young_gen->from_space()->capacity_in_bytes() - 2100 young_gen->to_space()->capacity_in_bytes(); 2101 size_policy->compute_generation_free_space( 2102 young_gen->used_in_bytes(), 2103 young_gen->eden_space()->used_in_bytes(), 2104 old_gen->used_in_bytes(), 2105 young_gen->eden_space()->capacity_in_bytes(), 2106 old_gen->max_gen_size(), 2107 max_eden_size, 2108 true /* full gc*/, 2109 gc_cause, 2110 heap->collector_policy()); 2111 2112 heap->resize_old_gen( 2113 size_policy->calculated_old_free_size_in_bytes()); 2114 2115 // Don't resize the young generation at an major collection. A 2116 // desired young generation size may have been calculated but 2117 // resizing the young generation complicates the code because the 2118 // resizing of the old generation may have moved the boundary 2119 // between the young generation and the old generation. Let the 2120 // young generation resizing happen at the minor collections. 2121 } 2122 if (PrintAdaptiveSizePolicy) { 2123 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2124 heap->total_collections()); 2125 } 2126 } 2127 2128 if (UsePerfData) { 2129 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2130 counters->update_counters(); 2131 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2132 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2133 } 2134 2135 heap->resize_all_tlabs(); 2136 2137 // Resize the metaspace capactiy after a collection 2138 MetaspaceGC::compute_new_size(); 2139 2140 if (TraceGen1Time) accumulated_time()->stop(); 2141 2142 if (PrintGC) { 2143 if (PrintGCDetails) { 2144 // No GC timestamp here. This is after GC so it would be confusing. 2145 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2146 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2147 heap->print_heap_change(pre_gc_values.heap_used()); 2148 MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used()); 2149 } else { 2150 heap->print_heap_change(pre_gc_values.heap_used()); 2151 } 2152 } 2153 2154 // Track memory usage and detect low memory 2155 MemoryService::track_memory_usage(); 2156 heap->update_counters(); 2157 gc_task_manager()->release_idle_workers(); 2158 } 2159 2160 #ifdef ASSERT 2161 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2162 ParCompactionManager* const cm = 2163 ParCompactionManager::manager_array(int(i)); 2164 assert(cm->marking_stack()->is_empty(), "should be empty"); 2165 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty"); 2166 } 2167 #endif // ASSERT 2168 2169 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2170 HandleMark hm; // Discard invalid handles created during verification 2171 Universe::verify(" VerifyAfterGC:"); 2172 } 2173 2174 // Re-verify object start arrays 2175 if (VerifyObjectStartArray && 2176 VerifyAfterGC) { 2177 old_gen->verify_object_start_array(); 2178 } 2179 2180 if (ZapUnusedHeapArea) { 2181 old_gen->object_space()->check_mangled_unused_area_complete(); 2182 } 2183 2184 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2185 2186 collection_exit.update(); 2187 2188 heap->print_heap_after_gc(); 2189 if (PrintGCTaskTimeStamps) { 2190 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2191 INT64_FORMAT, 2192 marking_start.ticks(), compaction_start.ticks(), 2193 collection_exit.ticks()); 2194 gc_task_manager()->print_task_time_stamps(); 2195 } 2196 2197 heap->post_full_gc_dump(); 2198 2199 #ifdef TRACESPINNING 2200 ParallelTaskTerminator::print_termination_counts(); 2201 #endif 2202 2203 return true; 2204 } 2205 2206 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2207 PSYoungGen* young_gen, 2208 PSOldGen* old_gen) { 2209 MutableSpace* const eden_space = young_gen->eden_space(); 2210 assert(!eden_space->is_empty(), "eden must be non-empty"); 2211 assert(young_gen->virtual_space()->alignment() == 2212 old_gen->virtual_space()->alignment(), "alignments do not match"); 2213 2214 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2215 return false; 2216 } 2217 2218 // Both generations must be completely committed. 2219 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2220 return false; 2221 } 2222 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2223 return false; 2224 } 2225 2226 // Figure out how much to take from eden. Include the average amount promoted 2227 // in the total; otherwise the next young gen GC will simply bail out to a 2228 // full GC. 2229 const size_t alignment = old_gen->virtual_space()->alignment(); 2230 const size_t eden_used = eden_space->used_in_bytes(); 2231 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2232 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2233 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2234 2235 if (absorb_size >= eden_capacity) { 2236 return false; // Must leave some space in eden. 2237 } 2238 2239 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2240 if (new_young_size < young_gen->min_gen_size()) { 2241 return false; // Respect young gen minimum size. 2242 } 2243 2244 if (TraceAdaptiveGCBoundary && Verbose) { 2245 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2246 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2247 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2248 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2249 absorb_size / K, 2250 eden_capacity / K, (eden_capacity - absorb_size) / K, 2251 young_gen->from_space()->used_in_bytes() / K, 2252 young_gen->to_space()->used_in_bytes() / K, 2253 young_gen->capacity_in_bytes() / K, new_young_size / K); 2254 } 2255 2256 // Fill the unused part of the old gen. 2257 MutableSpace* const old_space = old_gen->object_space(); 2258 HeapWord* const unused_start = old_space->top(); 2259 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2260 2261 if (unused_words > 0) { 2262 if (unused_words < CollectedHeap::min_fill_size()) { 2263 return false; // If the old gen cannot be filled, must give up. 2264 } 2265 CollectedHeap::fill_with_objects(unused_start, unused_words); 2266 } 2267 2268 // Take the live data from eden and set both top and end in the old gen to 2269 // eden top. (Need to set end because reset_after_change() mangles the region 2270 // from end to virtual_space->high() in debug builds). 2271 HeapWord* const new_top = eden_space->top(); 2272 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2273 absorb_size); 2274 young_gen->reset_after_change(); 2275 old_space->set_top(new_top); 2276 old_space->set_end(new_top); 2277 old_gen->reset_after_change(); 2278 2279 // Update the object start array for the filler object and the data from eden. 2280 ObjectStartArray* const start_array = old_gen->start_array(); 2281 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2282 start_array->allocate_block(p); 2283 } 2284 2285 // Could update the promoted average here, but it is not typically updated at 2286 // full GCs and the value to use is unclear. Something like 2287 // 2288 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2289 2290 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2291 return true; 2292 } 2293 2294 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2295 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2296 "shouldn't return NULL"); 2297 return ParallelScavengeHeap::gc_task_manager(); 2298 } 2299 2300 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2301 bool maximum_heap_compaction) { 2302 // Recursively traverse all live objects and mark them 2303 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); 2304 2305 ParallelScavengeHeap* heap = gc_heap(); 2306 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2307 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2308 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2309 ParallelTaskTerminator terminator(active_gc_threads, qset); 2310 2311 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2312 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2313 2314 // Need new claim bits before marking starts. 2315 ClassLoaderDataGraph::clear_claimed_marks(); 2316 2317 { 2318 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); 2319 ParallelScavengeHeap::ParStrongRootsScope psrs; 2320 2321 GCTaskQueue* q = GCTaskQueue::create(); 2322 2323 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2324 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2325 // We scan the thread roots in parallel 2326 Threads::create_thread_roots_marking_tasks(q); 2327 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2328 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2329 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2330 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2331 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2332 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2333 2334 if (active_gc_threads > 1) { 2335 for (uint j = 0; j < active_gc_threads; j++) { 2336 q->enqueue(new StealMarkingTask(&terminator)); 2337 } 2338 } 2339 2340 gc_task_manager()->execute_and_wait(q); 2341 } 2342 2343 // Process reference objects found during marking 2344 { 2345 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); 2346 if (ref_processor()->processing_is_mt()) { 2347 RefProcTaskExecutor task_executor; 2348 ref_processor()->process_discovered_references( 2349 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2350 &task_executor); 2351 } else { 2352 ref_processor()->process_discovered_references( 2353 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL); 2354 } 2355 } 2356 2357 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); 2358 2359 // This is the point where the entire marking should have completed. 2360 assert(cm->marking_stacks_empty(), "Marking should have completed"); 2361 2362 // Follow system dictionary roots and unload classes. 2363 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2364 2365 // Unload nmethods. 2366 CodeCache::do_unloading(is_alive_closure(), purged_class); 2367 2368 // Prune dead klasses from subklass/sibling/implementor lists. 2369 Klass::clean_weak_klass_links(is_alive_closure()); 2370 2371 // Delete entries for dead interned strings. 2372 StringTable::unlink(is_alive_closure()); 2373 2374 // Clean up unreferenced symbols in symbol table. 2375 SymbolTable::unlink(); 2376 } 2377 2378 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) { 2379 ClassLoaderData* cld = klass->class_loader_data(); 2380 // The actual processing of the klass is done when we 2381 // traverse the list of Klasses in the class loader data. 2382 PSParallelCompact::follow_class_loader(cm, cld); 2383 } 2384 2385 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) { 2386 ClassLoaderData* cld = klass->class_loader_data(); 2387 // The actual processing of the klass is done when we 2388 // traverse the list of Klasses in the class loader data. 2389 PSParallelCompact::adjust_class_loader(cm, cld); 2390 } 2391 2392 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm, 2393 ClassLoaderData* cld) { 2394 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2395 PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure); 2396 2397 cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true); 2398 } 2399 2400 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm, 2401 ClassLoaderData* cld) { 2402 cld->oops_do(PSParallelCompact::adjust_pointer_closure(), 2403 PSParallelCompact::adjust_klass_closure(), 2404 true); 2405 } 2406 2407 // This should be moved to the shared markSweep code! 2408 class PSAlwaysTrueClosure: public BoolObjectClosure { 2409 public: 2410 void do_object(oop p) { ShouldNotReachHere(); } 2411 bool do_object_b(oop p) { return true; } 2412 }; 2413 static PSAlwaysTrueClosure always_true; 2414 2415 void PSParallelCompact::adjust_roots() { 2416 // Adjust the pointers to reflect the new locations 2417 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); 2418 2419 // Need new claim bits when tracing through and adjusting pointers. 2420 ClassLoaderDataGraph::clear_claimed_marks(); 2421 2422 // General strong roots. 2423 Universe::oops_do(adjust_pointer_closure()); 2424 JNIHandles::oops_do(adjust_pointer_closure()); // Global (strong) JNI handles 2425 CLDToOopClosure adjust_from_cld(adjust_pointer_closure()); 2426 Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL); 2427 ObjectSynchronizer::oops_do(adjust_pointer_closure()); 2428 FlatProfiler::oops_do(adjust_pointer_closure()); 2429 Management::oops_do(adjust_pointer_closure()); 2430 JvmtiExport::oops_do(adjust_pointer_closure()); 2431 // SO_AllClasses 2432 SystemDictionary::oops_do(adjust_pointer_closure()); 2433 ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true); 2434 2435 // Now adjust pointers in remaining weak roots. (All of which should 2436 // have been cleared if they pointed to non-surviving objects.) 2437 // Global (weak) JNI handles 2438 JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure()); 2439 2440 CodeCache::oops_do(adjust_pointer_closure()); 2441 StringTable::oops_do(adjust_pointer_closure()); 2442 ref_processor()->weak_oops_do(adjust_pointer_closure()); 2443 // Roots were visited so references into the young gen in roots 2444 // may have been scanned. Process them also. 2445 // Should the reference processor have a span that excludes 2446 // young gen objects? 2447 PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure()); 2448 } 2449 2450 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2451 uint parallel_gc_threads) 2452 { 2453 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); 2454 2455 // Find the threads that are active 2456 unsigned int which = 0; 2457 2458 const uint task_count = MAX2(parallel_gc_threads, 1U); 2459 for (uint j = 0; j < task_count; j++) { 2460 q->enqueue(new DrainStacksCompactionTask(j)); 2461 ParCompactionManager::verify_region_list_empty(j); 2462 // Set the region stacks variables to "no" region stack values 2463 // so that they will be recognized and needing a region stack 2464 // in the stealing tasks if they do not get one by executing 2465 // a draining stack. 2466 ParCompactionManager* cm = ParCompactionManager::manager_array(j); 2467 cm->set_region_stack(NULL); 2468 cm->set_region_stack_index((uint)max_uintx); 2469 } 2470 ParCompactionManager::reset_recycled_stack_index(); 2471 2472 // Find all regions that are available (can be filled immediately) and 2473 // distribute them to the thread stacks. The iteration is done in reverse 2474 // order (high to low) so the regions will be removed in ascending order. 2475 2476 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2477 2478 size_t fillable_regions = 0; // A count for diagnostic purposes. 2479 // A region index which corresponds to the tasks created above. 2480 // "which" must be 0 <= which < task_count 2481 2482 which = 0; 2483 // id + 1 is used to test termination so unsigned can 2484 // be used with an old_space_id == 0. 2485 for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) { 2486 SpaceInfo* const space_info = _space_info + id; 2487 MutableSpace* const space = space_info->space(); 2488 HeapWord* const new_top = space_info->new_top(); 2489 2490 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2491 const size_t end_region = 2492 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2493 2494 for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) { 2495 if (sd.region(cur)->claim_unsafe()) { 2496 ParCompactionManager::region_list_push(which, cur); 2497 2498 if (TraceParallelOldGCCompactionPhase && Verbose) { 2499 const size_t count_mod_8 = fillable_regions & 7; 2500 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2501 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2502 if (count_mod_8 == 7) gclog_or_tty->cr(); 2503 } 2504 2505 NOT_PRODUCT(++fillable_regions;) 2506 2507 // Assign regions to tasks in round-robin fashion. 2508 if (++which == task_count) { 2509 assert(which <= parallel_gc_threads, 2510 "Inconsistent number of workers"); 2511 which = 0; 2512 } 2513 } 2514 } 2515 } 2516 2517 if (TraceParallelOldGCCompactionPhase) { 2518 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2519 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2520 } 2521 } 2522 2523 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2524 2525 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2526 uint parallel_gc_threads) { 2527 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); 2528 2529 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2530 2531 // Iterate over all the spaces adding tasks for updating 2532 // regions in the dense prefix. Assume that 1 gc thread 2533 // will work on opening the gaps and the remaining gc threads 2534 // will work on the dense prefix. 2535 unsigned int space_id; 2536 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2537 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2538 const MutableSpace* const space = _space_info[space_id].space(); 2539 2540 if (dense_prefix_end == space->bottom()) { 2541 // There is no dense prefix for this space. 2542 continue; 2543 } 2544 2545 // The dense prefix is before this region. 2546 size_t region_index_end_dense_prefix = 2547 sd.addr_to_region_idx(dense_prefix_end); 2548 RegionData* const dense_prefix_cp = 2549 sd.region(region_index_end_dense_prefix); 2550 assert(dense_prefix_end == space->end() || 2551 dense_prefix_cp->available() || 2552 dense_prefix_cp->claimed(), 2553 "The region after the dense prefix should always be ready to fill"); 2554 2555 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2556 2557 // Is there dense prefix work? 2558 size_t total_dense_prefix_regions = 2559 region_index_end_dense_prefix - region_index_start; 2560 // How many regions of the dense prefix should be given to 2561 // each thread? 2562 if (total_dense_prefix_regions > 0) { 2563 uint tasks_for_dense_prefix = 1; 2564 if (total_dense_prefix_regions <= 2565 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2566 // Don't over partition. This assumes that 2567 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2568 // so there are not many regions to process. 2569 tasks_for_dense_prefix = parallel_gc_threads; 2570 } else { 2571 // Over partition 2572 tasks_for_dense_prefix = parallel_gc_threads * 2573 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2574 } 2575 size_t regions_per_thread = total_dense_prefix_regions / 2576 tasks_for_dense_prefix; 2577 // Give each thread at least 1 region. 2578 if (regions_per_thread == 0) { 2579 regions_per_thread = 1; 2580 } 2581 2582 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2583 if (region_index_start >= region_index_end_dense_prefix) { 2584 break; 2585 } 2586 // region_index_end is not processed 2587 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2588 region_index_end_dense_prefix); 2589 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2590 region_index_start, 2591 region_index_end)); 2592 region_index_start = region_index_end; 2593 } 2594 } 2595 // This gets any part of the dense prefix that did not 2596 // fit evenly. 2597 if (region_index_start < region_index_end_dense_prefix) { 2598 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2599 region_index_start, 2600 region_index_end_dense_prefix)); 2601 } 2602 } 2603 } 2604 2605 void PSParallelCompact::enqueue_region_stealing_tasks( 2606 GCTaskQueue* q, 2607 ParallelTaskTerminator* terminator_ptr, 2608 uint parallel_gc_threads) { 2609 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); 2610 2611 // Once a thread has drained it's stack, it should try to steal regions from 2612 // other threads. 2613 if (parallel_gc_threads > 1) { 2614 for (uint j = 0; j < parallel_gc_threads; j++) { 2615 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2616 } 2617 } 2618 } 2619 2620 void PSParallelCompact::compact() { 2621 // trace("5"); 2622 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); 2623 2624 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2625 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2626 PSOldGen* old_gen = heap->old_gen(); 2627 old_gen->start_array()->reset(); 2628 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2629 uint active_gc_threads = heap->gc_task_manager()->active_workers(); 2630 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2631 ParallelTaskTerminator terminator(active_gc_threads, qset); 2632 2633 GCTaskQueue* q = GCTaskQueue::create(); 2634 enqueue_region_draining_tasks(q, active_gc_threads); 2635 enqueue_dense_prefix_tasks(q, active_gc_threads); 2636 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads); 2637 2638 { 2639 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); 2640 2641 gc_task_manager()->execute_and_wait(q); 2642 2643 #ifdef ASSERT 2644 // Verify that all regions have been processed before the deferred updates. 2645 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2646 verify_complete(SpaceId(id)); 2647 } 2648 #endif 2649 } 2650 2651 { 2652 // Update the deferred objects, if any. Any compaction manager can be used. 2653 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); 2654 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2655 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2656 update_deferred_objects(cm, SpaceId(id)); 2657 } 2658 } 2659 } 2660 2661 #ifdef ASSERT 2662 void PSParallelCompact::verify_complete(SpaceId space_id) { 2663 // All Regions between space bottom() to new_top() should be marked as filled 2664 // and all Regions between new_top() and top() should be available (i.e., 2665 // should have been emptied). 2666 ParallelCompactData& sd = summary_data(); 2667 SpaceInfo si = _space_info[space_id]; 2668 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2669 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2670 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2671 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2672 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2673 2674 bool issued_a_warning = false; 2675 2676 size_t cur_region; 2677 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2678 const RegionData* const c = sd.region(cur_region); 2679 if (!c->completed()) { 2680 warning("region " SIZE_FORMAT " not filled: " 2681 "destination_count=" SIZE_FORMAT, 2682 cur_region, c->destination_count()); 2683 issued_a_warning = true; 2684 } 2685 } 2686 2687 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2688 const RegionData* const c = sd.region(cur_region); 2689 if (!c->available()) { 2690 warning("region " SIZE_FORMAT " not empty: " 2691 "destination_count=" SIZE_FORMAT, 2692 cur_region, c->destination_count()); 2693 issued_a_warning = true; 2694 } 2695 } 2696 2697 if (issued_a_warning) { 2698 print_region_ranges(); 2699 } 2700 } 2701 #endif // #ifdef ASSERT 2702 2703 // Update interior oops in the ranges of regions [beg_region, end_region). 2704 void 2705 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2706 SpaceId space_id, 2707 size_t beg_region, 2708 size_t end_region) { 2709 ParallelCompactData& sd = summary_data(); 2710 ParMarkBitMap* const mbm = mark_bitmap(); 2711 2712 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2713 HeapWord* const end_addr = sd.region_to_addr(end_region); 2714 assert(beg_region <= end_region, "bad region range"); 2715 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2716 2717 #ifdef ASSERT 2718 // Claim the regions to avoid triggering an assert when they are marked as 2719 // filled. 2720 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2721 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2722 } 2723 #endif // #ifdef ASSERT 2724 2725 if (beg_addr != space(space_id)->bottom()) { 2726 // Find the first live object or block of dead space that *starts* in this 2727 // range of regions. If a partial object crosses onto the region, skip it; 2728 // it will be marked for 'deferred update' when the object head is 2729 // processed. If dead space crosses onto the region, it is also skipped; it 2730 // will be filled when the prior region is processed. If neither of those 2731 // apply, the first word in the region is the start of a live object or dead 2732 // space. 2733 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2734 const RegionData* const cp = sd.region(beg_region); 2735 if (cp->partial_obj_size() != 0) { 2736 beg_addr = sd.partial_obj_end(beg_region); 2737 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2738 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2739 } 2740 } 2741 2742 if (beg_addr < end_addr) { 2743 // A live object or block of dead space starts in this range of Regions. 2744 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2745 2746 // Create closures and iterate. 2747 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2748 FillClosure fill_closure(cm, space_id); 2749 ParMarkBitMap::IterationStatus status; 2750 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2751 dense_prefix_end); 2752 if (status == ParMarkBitMap::incomplete) { 2753 update_closure.do_addr(update_closure.source()); 2754 } 2755 } 2756 2757 // Mark the regions as filled. 2758 RegionData* const beg_cp = sd.region(beg_region); 2759 RegionData* const end_cp = sd.region(end_region); 2760 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2761 cp->set_completed(); 2762 } 2763 } 2764 2765 // Return the SpaceId for the space containing addr. If addr is not in the 2766 // heap, last_space_id is returned. In debug mode it expects the address to be 2767 // in the heap and asserts such. 2768 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2769 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2770 2771 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2772 if (_space_info[id].space()->contains(addr)) { 2773 return SpaceId(id); 2774 } 2775 } 2776 2777 assert(false, "no space contains the addr"); 2778 return last_space_id; 2779 } 2780 2781 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2782 SpaceId id) { 2783 assert(id < last_space_id, "bad space id"); 2784 2785 ParallelCompactData& sd = summary_data(); 2786 const SpaceInfo* const space_info = _space_info + id; 2787 ObjectStartArray* const start_array = space_info->start_array(); 2788 2789 const MutableSpace* const space = space_info->space(); 2790 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2791 HeapWord* const beg_addr = space_info->dense_prefix(); 2792 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2793 2794 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2795 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2796 const RegionData* cur_region; 2797 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2798 HeapWord* const addr = cur_region->deferred_obj_addr(); 2799 if (addr != NULL) { 2800 if (start_array != NULL) { 2801 start_array->allocate_block(addr); 2802 } 2803 oop(addr)->update_contents(cm); 2804 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 2805 } 2806 } 2807 } 2808 2809 // Skip over count live words starting from beg, and return the address of the 2810 // next live word. Unless marked, the word corresponding to beg is assumed to 2811 // be dead. Callers must either ensure beg does not correspond to the middle of 2812 // an object, or account for those live words in some other way. Callers must 2813 // also ensure that there are enough live words in the range [beg, end) to skip. 2814 HeapWord* 2815 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2816 { 2817 assert(count > 0, "sanity"); 2818 2819 ParMarkBitMap* m = mark_bitmap(); 2820 idx_t bits_to_skip = m->words_to_bits(count); 2821 idx_t cur_beg = m->addr_to_bit(beg); 2822 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 2823 2824 do { 2825 cur_beg = m->find_obj_beg(cur_beg, search_end); 2826 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 2827 const size_t obj_bits = cur_end - cur_beg + 1; 2828 if (obj_bits > bits_to_skip) { 2829 return m->bit_to_addr(cur_beg + bits_to_skip); 2830 } 2831 bits_to_skip -= obj_bits; 2832 cur_beg = cur_end + 1; 2833 } while (bits_to_skip > 0); 2834 2835 // Skipping the desired number of words landed just past the end of an object. 2836 // Find the start of the next object. 2837 cur_beg = m->find_obj_beg(cur_beg, search_end); 2838 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 2839 return m->bit_to_addr(cur_beg); 2840 } 2841 2842 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 2843 SpaceId src_space_id, 2844 size_t src_region_idx) 2845 { 2846 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 2847 2848 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 2849 if (split_info.dest_region_addr() == dest_addr) { 2850 // The partial object ending at the split point contains the first word to 2851 // be copied to dest_addr. 2852 return split_info.first_src_addr(); 2853 } 2854 2855 const ParallelCompactData& sd = summary_data(); 2856 ParMarkBitMap* const bitmap = mark_bitmap(); 2857 const size_t RegionSize = ParallelCompactData::RegionSize; 2858 2859 assert(sd.is_region_aligned(dest_addr), "not aligned"); 2860 const RegionData* const src_region_ptr = sd.region(src_region_idx); 2861 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 2862 HeapWord* const src_region_destination = src_region_ptr->destination(); 2863 2864 assert(dest_addr >= src_region_destination, "wrong src region"); 2865 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 2866 2867 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 2868 HeapWord* const src_region_end = src_region_beg + RegionSize; 2869 2870 HeapWord* addr = src_region_beg; 2871 if (dest_addr == src_region_destination) { 2872 // Return the first live word in the source region. 2873 if (partial_obj_size == 0) { 2874 addr = bitmap->find_obj_beg(addr, src_region_end); 2875 assert(addr < src_region_end, "no objects start in src region"); 2876 } 2877 return addr; 2878 } 2879 2880 // Must skip some live data. 2881 size_t words_to_skip = dest_addr - src_region_destination; 2882 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 2883 2884 if (partial_obj_size >= words_to_skip) { 2885 // All the live words to skip are part of the partial object. 2886 addr += words_to_skip; 2887 if (partial_obj_size == words_to_skip) { 2888 // Find the first live word past the partial object. 2889 addr = bitmap->find_obj_beg(addr, src_region_end); 2890 assert(addr < src_region_end, "wrong src region"); 2891 } 2892 return addr; 2893 } 2894 2895 // Skip over the partial object (if any). 2896 if (partial_obj_size != 0) { 2897 words_to_skip -= partial_obj_size; 2898 addr += partial_obj_size; 2899 } 2900 2901 // Skip over live words due to objects that start in the region. 2902 addr = skip_live_words(addr, src_region_end, words_to_skip); 2903 assert(addr < src_region_end, "wrong src region"); 2904 return addr; 2905 } 2906 2907 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 2908 SpaceId src_space_id, 2909 size_t beg_region, 2910 HeapWord* end_addr) 2911 { 2912 ParallelCompactData& sd = summary_data(); 2913 2914 #ifdef ASSERT 2915 MutableSpace* const src_space = _space_info[src_space_id].space(); 2916 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 2917 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 2918 "src_space_id does not match beg_addr"); 2919 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 2920 "src_space_id does not match end_addr"); 2921 #endif // #ifdef ASSERT 2922 2923 RegionData* const beg = sd.region(beg_region); 2924 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 2925 2926 // Regions up to new_top() are enqueued if they become available. 2927 HeapWord* const new_top = _space_info[src_space_id].new_top(); 2928 RegionData* const enqueue_end = 2929 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 2930 2931 for (RegionData* cur = beg; cur < end; ++cur) { 2932 assert(cur->data_size() > 0, "region must have live data"); 2933 cur->decrement_destination_count(); 2934 if (cur < enqueue_end && cur->available() && cur->claim()) { 2935 cm->push_region(sd.region(cur)); 2936 } 2937 } 2938 } 2939 2940 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 2941 SpaceId& src_space_id, 2942 HeapWord*& src_space_top, 2943 HeapWord* end_addr) 2944 { 2945 typedef ParallelCompactData::RegionData RegionData; 2946 2947 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2948 const size_t region_size = ParallelCompactData::RegionSize; 2949 2950 size_t src_region_idx = 0; 2951 2952 // Skip empty regions (if any) up to the top of the space. 2953 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 2954 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 2955 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 2956 const RegionData* const top_region_ptr = 2957 sd.addr_to_region_ptr(top_aligned_up); 2958 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 2959 ++src_region_ptr; 2960 } 2961 2962 if (src_region_ptr < top_region_ptr) { 2963 // The next source region is in the current space. Update src_region_idx 2964 // and the source address to match src_region_ptr. 2965 src_region_idx = sd.region(src_region_ptr); 2966 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 2967 if (src_region_addr > closure.source()) { 2968 closure.set_source(src_region_addr); 2969 } 2970 return src_region_idx; 2971 } 2972 2973 // Switch to a new source space and find the first non-empty region. 2974 unsigned int space_id = src_space_id + 1; 2975 assert(space_id < last_space_id, "not enough spaces"); 2976 2977 HeapWord* const destination = closure.destination(); 2978 2979 do { 2980 MutableSpace* space = _space_info[space_id].space(); 2981 HeapWord* const bottom = space->bottom(); 2982 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 2983 2984 // Iterate over the spaces that do not compact into themselves. 2985 if (bottom_cp->destination() != bottom) { 2986 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 2987 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 2988 2989 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 2990 if (src_cp->live_obj_size() > 0) { 2991 // Found it. 2992 assert(src_cp->destination() == destination, 2993 "first live obj in the space must match the destination"); 2994 assert(src_cp->partial_obj_size() == 0, 2995 "a space cannot begin with a partial obj"); 2996 2997 src_space_id = SpaceId(space_id); 2998 src_space_top = space->top(); 2999 const size_t src_region_idx = sd.region(src_cp); 3000 closure.set_source(sd.region_to_addr(src_region_idx)); 3001 return src_region_idx; 3002 } else { 3003 assert(src_cp->data_size() == 0, "sanity"); 3004 } 3005 } 3006 } 3007 } while (++space_id < last_space_id); 3008 3009 assert(false, "no source region was found"); 3010 return 0; 3011 } 3012 3013 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3014 { 3015 typedef ParMarkBitMap::IterationStatus IterationStatus; 3016 const size_t RegionSize = ParallelCompactData::RegionSize; 3017 ParMarkBitMap* const bitmap = mark_bitmap(); 3018 ParallelCompactData& sd = summary_data(); 3019 RegionData* const region_ptr = sd.region(region_idx); 3020 3021 // Get the items needed to construct the closure. 3022 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3023 SpaceId dest_space_id = space_id(dest_addr); 3024 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3025 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3026 assert(dest_addr < new_top, "sanity"); 3027 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3028 3029 // Get the source region and related info. 3030 size_t src_region_idx = region_ptr->source_region(); 3031 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3032 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3033 3034 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3035 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3036 3037 // Adjust src_region_idx to prepare for decrementing destination counts (the 3038 // destination count is not decremented when a region is copied to itself). 3039 if (src_region_idx == region_idx) { 3040 src_region_idx += 1; 3041 } 3042 3043 if (bitmap->is_unmarked(closure.source())) { 3044 // The first source word is in the middle of an object; copy the remainder 3045 // of the object or as much as will fit. The fact that pointer updates were 3046 // deferred will be noted when the object header is processed. 3047 HeapWord* const old_src_addr = closure.source(); 3048 closure.copy_partial_obj(); 3049 if (closure.is_full()) { 3050 decrement_destination_counts(cm, src_space_id, src_region_idx, 3051 closure.source()); 3052 region_ptr->set_deferred_obj_addr(NULL); 3053 region_ptr->set_completed(); 3054 return; 3055 } 3056 3057 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3058 if (sd.region_align_down(old_src_addr) != end_addr) { 3059 // The partial object was copied from more than one source region. 3060 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3061 3062 // Move to the next source region, possibly switching spaces as well. All 3063 // args except end_addr may be modified. 3064 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3065 end_addr); 3066 } 3067 } 3068 3069 do { 3070 HeapWord* const cur_addr = closure.source(); 3071 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3072 src_space_top); 3073 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3074 3075 if (status == ParMarkBitMap::incomplete) { 3076 // The last obj that starts in the source region does not end in the 3077 // region. 3078 assert(closure.source() < end_addr, "sanity"); 3079 HeapWord* const obj_beg = closure.source(); 3080 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3081 src_space_top); 3082 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3083 if (obj_end < range_end) { 3084 // The end was found; the entire object will fit. 3085 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3086 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3087 } else { 3088 // The end was not found; the object will not fit. 3089 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3090 status = ParMarkBitMap::would_overflow; 3091 } 3092 } 3093 3094 if (status == ParMarkBitMap::would_overflow) { 3095 // The last object did not fit. Note that interior oop updates were 3096 // deferred, then copy enough of the object to fill the region. 3097 region_ptr->set_deferred_obj_addr(closure.destination()); 3098 status = closure.copy_until_full(); // copies from closure.source() 3099 3100 decrement_destination_counts(cm, src_space_id, src_region_idx, 3101 closure.source()); 3102 region_ptr->set_completed(); 3103 return; 3104 } 3105 3106 if (status == ParMarkBitMap::full) { 3107 decrement_destination_counts(cm, src_space_id, src_region_idx, 3108 closure.source()); 3109 region_ptr->set_deferred_obj_addr(NULL); 3110 region_ptr->set_completed(); 3111 return; 3112 } 3113 3114 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3115 3116 // Move to the next source region, possibly switching spaces as well. All 3117 // args except end_addr may be modified. 3118 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3119 end_addr); 3120 } while (true); 3121 } 3122 3123 void 3124 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3125 const MutableSpace* sp = space(space_id); 3126 if (sp->is_empty()) { 3127 return; 3128 } 3129 3130 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3131 ParMarkBitMap* const bitmap = mark_bitmap(); 3132 HeapWord* const dp_addr = dense_prefix(space_id); 3133 HeapWord* beg_addr = sp->bottom(); 3134 HeapWord* end_addr = sp->top(); 3135 3136 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3137 3138 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3139 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3140 if (beg_region < dp_region) { 3141 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3142 } 3143 3144 // The destination of the first live object that starts in the region is one 3145 // past the end of the partial object entering the region (if any). 3146 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3147 HeapWord* const new_top = _space_info[space_id].new_top(); 3148 assert(new_top >= dest_addr, "bad new_top value"); 3149 const size_t words = pointer_delta(new_top, dest_addr); 3150 3151 if (words > 0) { 3152 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3153 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3154 3155 ParMarkBitMap::IterationStatus status; 3156 status = bitmap->iterate(&closure, dest_addr, end_addr); 3157 assert(status == ParMarkBitMap::full, "iteration not complete"); 3158 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3159 "live objects skipped because closure is full"); 3160 } 3161 } 3162 3163 jlong PSParallelCompact::millis_since_last_gc() { 3164 // We need a monotonically non-deccreasing time in ms but 3165 // os::javaTimeMillis() does not guarantee monotonicity. 3166 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3167 jlong ret_val = now - _time_of_last_gc; 3168 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3169 if (ret_val < 0) { 3170 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);) 3171 return 0; 3172 } 3173 return ret_val; 3174 } 3175 3176 void PSParallelCompact::reset_millis_since_last_gc() { 3177 // We need a monotonically non-deccreasing time in ms but 3178 // os::javaTimeMillis() does not guarantee monotonicity. 3179 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 3180 } 3181 3182 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3183 { 3184 if (source() != destination()) { 3185 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3186 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3187 } 3188 update_state(words_remaining()); 3189 assert(is_full(), "sanity"); 3190 return ParMarkBitMap::full; 3191 } 3192 3193 void MoveAndUpdateClosure::copy_partial_obj() 3194 { 3195 size_t words = words_remaining(); 3196 3197 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3198 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3199 if (end_addr < range_end) { 3200 words = bitmap()->obj_size(source(), end_addr); 3201 } 3202 3203 // This test is necessary; if omitted, the pointer updates to a partial object 3204 // that crosses the dense prefix boundary could be overwritten. 3205 if (source() != destination()) { 3206 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3207 Copy::aligned_conjoint_words(source(), destination(), words); 3208 } 3209 update_state(words); 3210 } 3211 3212 ParMarkBitMapClosure::IterationStatus 3213 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3214 assert(destination() != NULL, "sanity"); 3215 assert(bitmap()->obj_size(addr) == words, "bad size"); 3216 3217 _source = addr; 3218 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3219 destination(), "wrong destination"); 3220 3221 if (words > words_remaining()) { 3222 return ParMarkBitMap::would_overflow; 3223 } 3224 3225 // The start_array must be updated even if the object is not moving. 3226 if (_start_array != NULL) { 3227 _start_array->allocate_block(destination()); 3228 } 3229 3230 if (destination() != source()) { 3231 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3232 Copy::aligned_conjoint_words(source(), destination(), words); 3233 } 3234 3235 oop moved_oop = (oop) destination(); 3236 moved_oop->update_contents(compaction_manager()); 3237 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3238 3239 update_state(words); 3240 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3241 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3242 } 3243 3244 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3245 ParCompactionManager* cm, 3246 PSParallelCompact::SpaceId space_id) : 3247 ParMarkBitMapClosure(mbm, cm), 3248 _space_id(space_id), 3249 _start_array(PSParallelCompact::start_array(space_id)) 3250 { 3251 } 3252 3253 // Updates the references in the object to their new values. 3254 ParMarkBitMapClosure::IterationStatus 3255 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3256 do_addr(addr); 3257 return ParMarkBitMap::incomplete; 3258 }