1 /* 2 * Copyright (c) 2005, 2010, 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/psPermGen.hpp" 40 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp" 41 #include "gc_implementation/parallelScavenge/psScavenge.hpp" 42 #include "gc_implementation/parallelScavenge/psYoungGen.hpp" 43 #include "gc_implementation/shared/isGCActiveMark.hpp" 44 #include "gc_interface/gcCause.hpp" 45 #include "memory/gcLocker.inline.hpp" 46 #include "memory/referencePolicy.hpp" 47 #include "memory/referenceProcessor.hpp" 48 #include "oops/methodDataOop.hpp" 49 #include "oops/oop.inline.hpp" 50 #include "oops/oop.pcgc.inline.hpp" 51 #include "runtime/fprofiler.hpp" 52 #include "runtime/safepoint.hpp" 53 #include "runtime/vmThread.hpp" 54 #include "services/management.hpp" 55 #include "services/memoryService.hpp" 56 #include "utilities/events.hpp" 57 58 #include <math.h> 59 60 // All sizes are in HeapWords. 61 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words 62 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 63 const size_t ParallelCompactData::RegionSizeBytes = 64 RegionSize << LogHeapWordSize; 65 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 66 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 67 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 68 69 const ParallelCompactData::RegionData::region_sz_t 70 ParallelCompactData::RegionData::dc_shift = 27; 71 72 const ParallelCompactData::RegionData::region_sz_t 73 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 74 75 const ParallelCompactData::RegionData::region_sz_t 76 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 77 78 const ParallelCompactData::RegionData::region_sz_t 79 ParallelCompactData::RegionData::los_mask = ~dc_mask; 80 81 const ParallelCompactData::RegionData::region_sz_t 82 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 83 84 const ParallelCompactData::RegionData::region_sz_t 85 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 86 87 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 88 bool PSParallelCompact::_print_phases = false; 89 90 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 91 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL; 92 93 double PSParallelCompact::_dwl_mean; 94 double PSParallelCompact::_dwl_std_dev; 95 double PSParallelCompact::_dwl_first_term; 96 double PSParallelCompact::_dwl_adjustment; 97 #ifdef ASSERT 98 bool PSParallelCompact::_dwl_initialized = false; 99 #endif // #ifdef ASSERT 100 101 #ifdef VALIDATE_MARK_SWEEP 102 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL; 103 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL; 104 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL; 105 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL; 106 size_t PSParallelCompact::_live_oops_index = 0; 107 size_t PSParallelCompact::_live_oops_index_at_perm = 0; 108 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL; 109 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL; 110 bool PSParallelCompact::_pointer_tracking = false; 111 bool PSParallelCompact::_root_tracking = true; 112 113 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL; 114 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL; 115 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL; 116 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL; 117 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL; 118 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL; 119 #endif 120 121 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 122 HeapWord* destination) 123 { 124 assert(src_region_idx != 0, "invalid src_region_idx"); 125 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 126 assert(destination != NULL, "invalid destination argument"); 127 128 _src_region_idx = src_region_idx; 129 _partial_obj_size = partial_obj_size; 130 _destination = destination; 131 132 // These fields may not be updated below, so make sure they're clear. 133 assert(_dest_region_addr == NULL, "should have been cleared"); 134 assert(_first_src_addr == NULL, "should have been cleared"); 135 136 // Determine the number of destination regions for the partial object. 137 HeapWord* const last_word = destination + partial_obj_size - 1; 138 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 139 HeapWord* const beg_region_addr = sd.region_align_down(destination); 140 HeapWord* const end_region_addr = sd.region_align_down(last_word); 141 142 if (beg_region_addr == end_region_addr) { 143 // One destination region. 144 _destination_count = 1; 145 if (end_region_addr == destination) { 146 // The destination falls on a region boundary, thus the first word of the 147 // partial object will be the first word copied to the destination region. 148 _dest_region_addr = end_region_addr; 149 _first_src_addr = sd.region_to_addr(src_region_idx); 150 } 151 } else { 152 // Two destination regions. When copied, the partial object will cross a 153 // destination region boundary, so a word somewhere within the partial 154 // object will be the first word copied to the second destination region. 155 _destination_count = 2; 156 _dest_region_addr = end_region_addr; 157 const size_t ofs = pointer_delta(end_region_addr, destination); 158 assert(ofs < _partial_obj_size, "sanity"); 159 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 160 } 161 } 162 163 void SplitInfo::clear() 164 { 165 _src_region_idx = 0; 166 _partial_obj_size = 0; 167 _destination = NULL; 168 _destination_count = 0; 169 _dest_region_addr = NULL; 170 _first_src_addr = NULL; 171 assert(!is_valid(), "sanity"); 172 } 173 174 #ifdef ASSERT 175 void SplitInfo::verify_clear() 176 { 177 assert(_src_region_idx == 0, "not clear"); 178 assert(_partial_obj_size == 0, "not clear"); 179 assert(_destination == NULL, "not clear"); 180 assert(_destination_count == 0, "not clear"); 181 assert(_dest_region_addr == NULL, "not clear"); 182 assert(_first_src_addr == NULL, "not clear"); 183 } 184 #endif // #ifdef ASSERT 185 186 187 #ifndef PRODUCT 188 const char* PSParallelCompact::space_names[] = { 189 "perm", "old ", "eden", "from", "to " 190 }; 191 192 void PSParallelCompact::print_region_ranges() 193 { 194 tty->print_cr("space bottom top end new_top"); 195 tty->print_cr("------ ---------- ---------- ---------- ----------"); 196 197 for (unsigned int id = 0; id < last_space_id; ++id) { 198 const MutableSpace* space = _space_info[id].space(); 199 tty->print_cr("%u %s " 200 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 201 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 202 id, space_names[id], 203 summary_data().addr_to_region_idx(space->bottom()), 204 summary_data().addr_to_region_idx(space->top()), 205 summary_data().addr_to_region_idx(space->end()), 206 summary_data().addr_to_region_idx(_space_info[id].new_top())); 207 } 208 } 209 210 void 211 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 212 { 213 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 214 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 215 216 ParallelCompactData& sd = PSParallelCompact::summary_data(); 217 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 218 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 219 REGION_IDX_FORMAT " " PTR_FORMAT " " 220 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 221 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 222 i, c->data_location(), dci, c->destination(), 223 c->partial_obj_size(), c->live_obj_size(), 224 c->data_size(), c->source_region(), c->destination_count()); 225 226 #undef REGION_IDX_FORMAT 227 #undef REGION_DATA_FORMAT 228 } 229 230 void 231 print_generic_summary_data(ParallelCompactData& summary_data, 232 HeapWord* const beg_addr, 233 HeapWord* const end_addr) 234 { 235 size_t total_words = 0; 236 size_t i = summary_data.addr_to_region_idx(beg_addr); 237 const size_t last = summary_data.addr_to_region_idx(end_addr); 238 HeapWord* pdest = 0; 239 240 while (i <= last) { 241 ParallelCompactData::RegionData* c = summary_data.region(i); 242 if (c->data_size() != 0 || c->destination() != pdest) { 243 print_generic_summary_region(i, c); 244 total_words += c->data_size(); 245 pdest = c->destination(); 246 } 247 ++i; 248 } 249 250 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 251 } 252 253 void 254 print_generic_summary_data(ParallelCompactData& summary_data, 255 SpaceInfo* space_info) 256 { 257 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 258 const MutableSpace* space = space_info[id].space(); 259 print_generic_summary_data(summary_data, space->bottom(), 260 MAX2(space->top(), space_info[id].new_top())); 261 } 262 } 263 264 void 265 print_initial_summary_region(size_t i, 266 const ParallelCompactData::RegionData* c, 267 bool newline = true) 268 { 269 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 270 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 271 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 272 i, c->destination(), 273 c->partial_obj_size(), c->live_obj_size(), 274 c->data_size(), c->source_region(), c->destination_count()); 275 if (newline) tty->cr(); 276 } 277 278 void 279 print_initial_summary_data(ParallelCompactData& summary_data, 280 const MutableSpace* space) { 281 if (space->top() == space->bottom()) { 282 return; 283 } 284 285 const size_t region_size = ParallelCompactData::RegionSize; 286 typedef ParallelCompactData::RegionData RegionData; 287 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 288 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 289 const RegionData* c = summary_data.region(end_region - 1); 290 HeapWord* end_addr = c->destination() + c->data_size(); 291 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 292 293 // Print (and count) the full regions at the beginning of the space. 294 size_t full_region_count = 0; 295 size_t i = summary_data.addr_to_region_idx(space->bottom()); 296 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 297 print_initial_summary_region(i, summary_data.region(i)); 298 ++full_region_count; 299 ++i; 300 } 301 302 size_t live_to_right = live_in_space - full_region_count * region_size; 303 304 double max_reclaimed_ratio = 0.0; 305 size_t max_reclaimed_ratio_region = 0; 306 size_t max_dead_to_right = 0; 307 size_t max_live_to_right = 0; 308 309 // Print the 'reclaimed ratio' for regions while there is something live in 310 // the region or to the right of it. The remaining regions are empty (and 311 // uninteresting), and computing the ratio will result in division by 0. 312 while (i < end_region && live_to_right > 0) { 313 c = summary_data.region(i); 314 HeapWord* const region_addr = summary_data.region_to_addr(i); 315 const size_t used_to_right = pointer_delta(space->top(), region_addr); 316 const size_t dead_to_right = used_to_right - live_to_right; 317 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 318 319 if (reclaimed_ratio > max_reclaimed_ratio) { 320 max_reclaimed_ratio = reclaimed_ratio; 321 max_reclaimed_ratio_region = i; 322 max_dead_to_right = dead_to_right; 323 max_live_to_right = live_to_right; 324 } 325 326 print_initial_summary_region(i, c, false); 327 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 328 reclaimed_ratio, dead_to_right, live_to_right); 329 330 live_to_right -= c->data_size(); 331 ++i; 332 } 333 334 // Any remaining regions are empty. Print one more if there is one. 335 if (i < end_region) { 336 print_initial_summary_region(i, summary_data.region(i)); 337 } 338 339 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 340 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 341 max_reclaimed_ratio_region, max_dead_to_right, 342 max_live_to_right, max_reclaimed_ratio); 343 } 344 345 void 346 print_initial_summary_data(ParallelCompactData& summary_data, 347 SpaceInfo* space_info) { 348 unsigned int id = PSParallelCompact::perm_space_id; 349 const MutableSpace* space; 350 do { 351 space = space_info[id].space(); 352 print_initial_summary_data(summary_data, space); 353 } while (++id < PSParallelCompact::eden_space_id); 354 355 do { 356 space = space_info[id].space(); 357 print_generic_summary_data(summary_data, space->bottom(), space->top()); 358 } while (++id < PSParallelCompact::last_space_id); 359 } 360 #endif // #ifndef PRODUCT 361 362 #ifdef ASSERT 363 size_t add_obj_count; 364 size_t add_obj_size; 365 size_t mark_bitmap_count; 366 size_t mark_bitmap_size; 367 #endif // #ifdef ASSERT 368 369 ParallelCompactData::ParallelCompactData() 370 { 371 _region_start = 0; 372 373 _region_vspace = 0; 374 _region_data = 0; 375 _region_count = 0; 376 } 377 378 bool ParallelCompactData::initialize(MemRegion covered_region) 379 { 380 _region_start = covered_region.start(); 381 const size_t region_size = covered_region.word_size(); 382 DEBUG_ONLY(_region_end = _region_start + region_size;) 383 384 assert(region_align_down(_region_start) == _region_start, 385 "region start not aligned"); 386 assert((region_size & RegionSizeOffsetMask) == 0, 387 "region size not a multiple of RegionSize"); 388 389 bool result = initialize_region_data(region_size); 390 391 return result; 392 } 393 394 PSVirtualSpace* 395 ParallelCompactData::create_vspace(size_t count, size_t element_size) 396 { 397 const size_t raw_bytes = count * element_size; 398 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); 399 const size_t granularity = os::vm_allocation_granularity(); 400 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 401 402 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 403 MAX2(page_sz, granularity); 404 ReservedSpace rs(bytes, rs_align, rs_align > 0); 405 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 406 rs.size()); 407 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 408 if (vspace != 0) { 409 if (vspace->expand_by(bytes)) { 410 return vspace; 411 } 412 delete vspace; 413 // Release memory reserved in the space. 414 rs.release(); 415 } 416 417 return 0; 418 } 419 420 bool ParallelCompactData::initialize_region_data(size_t region_size) 421 { 422 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 423 _region_vspace = create_vspace(count, sizeof(RegionData)); 424 if (_region_vspace != 0) { 425 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 426 _region_count = count; 427 return true; 428 } 429 return false; 430 } 431 432 void ParallelCompactData::clear() 433 { 434 memset(_region_data, 0, _region_vspace->committed_size()); 435 } 436 437 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 438 assert(beg_region <= _region_count, "beg_region out of range"); 439 assert(end_region <= _region_count, "end_region out of range"); 440 441 const size_t region_cnt = end_region - beg_region; 442 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 443 } 444 445 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 446 { 447 const RegionData* cur_cp = region(region_idx); 448 const RegionData* const end_cp = region(region_count() - 1); 449 450 HeapWord* result = region_to_addr(region_idx); 451 if (cur_cp < end_cp) { 452 do { 453 result += cur_cp->partial_obj_size(); 454 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 455 } 456 return result; 457 } 458 459 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 460 { 461 const size_t obj_ofs = pointer_delta(addr, _region_start); 462 const size_t beg_region = obj_ofs >> Log2RegionSize; 463 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 464 465 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 466 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 467 468 if (beg_region == end_region) { 469 // All in one region. 470 _region_data[beg_region].add_live_obj(len); 471 return; 472 } 473 474 // First region. 475 const size_t beg_ofs = region_offset(addr); 476 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 477 478 klassOop klass = ((oop)addr)->klass(); 479 // Middle regions--completely spanned by this object. 480 for (size_t region = beg_region + 1; region < end_region; ++region) { 481 _region_data[region].set_partial_obj_size(RegionSize); 482 _region_data[region].set_partial_obj_addr(addr); 483 } 484 485 // Last region. 486 const size_t end_ofs = region_offset(addr + len - 1); 487 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 488 _region_data[end_region].set_partial_obj_addr(addr); 489 } 490 491 void 492 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 493 { 494 assert(region_offset(beg) == 0, "not RegionSize aligned"); 495 assert(region_offset(end) == 0, "not RegionSize aligned"); 496 497 size_t cur_region = addr_to_region_idx(beg); 498 const size_t end_region = addr_to_region_idx(end); 499 HeapWord* addr = beg; 500 while (cur_region < end_region) { 501 _region_data[cur_region].set_destination(addr); 502 _region_data[cur_region].set_destination_count(0); 503 _region_data[cur_region].set_source_region(cur_region); 504 _region_data[cur_region].set_data_location(addr); 505 506 // Update live_obj_size so the region appears completely full. 507 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 508 _region_data[cur_region].set_live_obj_size(live_size); 509 510 ++cur_region; 511 addr += RegionSize; 512 } 513 } 514 515 // Find the point at which a space can be split and, if necessary, record the 516 // split point. 517 // 518 // If the current src region (which overflowed the destination space) doesn't 519 // have a partial object, the split point is at the beginning of the current src 520 // region (an "easy" split, no extra bookkeeping required). 521 // 522 // If the current src region has a partial object, the split point is in the 523 // region where that partial object starts (call it the split_region). If 524 // split_region has a partial object, then the split point is just after that 525 // partial object (a "hard" split where we have to record the split data and 526 // zero the partial_obj_size field). With a "hard" split, we know that the 527 // partial_obj ends within split_region because the partial object that caused 528 // the overflow starts in split_region. If split_region doesn't have a partial 529 // obj, then the split is at the beginning of split_region (another "easy" 530 // split). 531 HeapWord* 532 ParallelCompactData::summarize_split_space(size_t src_region, 533 SplitInfo& split_info, 534 HeapWord* destination, 535 HeapWord* target_end, 536 HeapWord** target_next) 537 { 538 assert(destination <= target_end, "sanity"); 539 assert(destination + _region_data[src_region].data_size() > target_end, 540 "region should not fit into target space"); 541 assert(is_region_aligned(target_end), "sanity"); 542 543 size_t split_region = src_region; 544 HeapWord* split_destination = destination; 545 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 546 547 if (destination + partial_obj_size > target_end) { 548 // The split point is just after the partial object (if any) in the 549 // src_region that contains the start of the object that overflowed the 550 // destination space. 551 // 552 // Find the start of the "overflow" object and set split_region to the 553 // region containing it. 554 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 555 split_region = addr_to_region_idx(overflow_obj); 556 557 // Clear the source_region field of all destination regions whose first word 558 // came from data after the split point (a non-null source_region field 559 // implies a region must be filled). 560 // 561 // An alternative to the simple loop below: clear during post_compact(), 562 // which uses memcpy instead of individual stores, and is easy to 563 // parallelize. (The downside is that it clears the entire RegionData 564 // object as opposed to just one field.) 565 // 566 // post_compact() would have to clear the summary data up to the highest 567 // address that was written during the summary phase, which would be 568 // 569 // max(top, max(new_top, clear_top)) 570 // 571 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 572 // to target_end. 573 const RegionData* const sr = region(split_region); 574 const size_t beg_idx = 575 addr_to_region_idx(region_align_up(sr->destination() + 576 sr->partial_obj_size())); 577 const size_t end_idx = addr_to_region_idx(target_end); 578 579 if (TraceParallelOldGCSummaryPhase) { 580 gclog_or_tty->print_cr("split: clearing source_region field in [" 581 SIZE_FORMAT ", " SIZE_FORMAT ")", 582 beg_idx, end_idx); 583 } 584 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 585 _region_data[idx].set_source_region(0); 586 } 587 588 // Set split_destination and partial_obj_size to reflect the split region. 589 split_destination = sr->destination(); 590 partial_obj_size = sr->partial_obj_size(); 591 } 592 593 // The split is recorded only if a partial object extends onto the region. 594 if (partial_obj_size != 0) { 595 _region_data[split_region].set_partial_obj_size(0); 596 split_info.record(split_region, partial_obj_size, split_destination); 597 } 598 599 // Setup the continuation addresses. 600 *target_next = split_destination + partial_obj_size; 601 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 602 603 if (TraceParallelOldGCSummaryPhase) { 604 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 605 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 606 " pos=" SIZE_FORMAT, 607 split_type, source_next, split_region, 608 partial_obj_size); 609 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 610 " tn=" PTR_FORMAT, 611 split_type, split_destination, 612 addr_to_region_idx(split_destination), 613 *target_next); 614 615 if (partial_obj_size != 0) { 616 HeapWord* const po_beg = split_info.destination(); 617 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 618 gclog_or_tty->print_cr("%s split: " 619 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 620 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 621 split_type, 622 po_beg, addr_to_region_idx(po_beg), 623 po_end, addr_to_region_idx(po_end)); 624 } 625 } 626 627 return source_next; 628 } 629 630 bool ParallelCompactData::summarize(SplitInfo& split_info, 631 HeapWord* source_beg, HeapWord* source_end, 632 HeapWord** source_next, 633 HeapWord* target_beg, HeapWord* target_end, 634 HeapWord** target_next) 635 { 636 if (TraceParallelOldGCSummaryPhase) { 637 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 638 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 639 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 640 source_beg, source_end, source_next_val, 641 target_beg, target_end, *target_next); 642 } 643 644 size_t cur_region = addr_to_region_idx(source_beg); 645 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 646 647 HeapWord *dest_addr = target_beg; 648 while (cur_region < end_region) { 649 // The destination must be set even if the region has no data. 650 _region_data[cur_region].set_destination(dest_addr); 651 652 size_t words = _region_data[cur_region].data_size(); 653 if (words > 0) { 654 // If cur_region does not fit entirely into the target space, find a point 655 // at which the source space can be 'split' so that part is copied to the 656 // target space and the rest is copied elsewhere. 657 if (dest_addr + words > target_end) { 658 assert(source_next != NULL, "source_next is NULL when splitting"); 659 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 660 target_end, target_next); 661 return false; 662 } 663 664 // Compute the destination_count for cur_region, and if necessary, update 665 // source_region for a destination region. The source_region field is 666 // updated if cur_region is the first (left-most) region to be copied to a 667 // destination region. 668 // 669 // The destination_count calculation is a bit subtle. A region that has 670 // data that compacts into itself does not count itself as a destination. 671 // This maintains the invariant that a zero count means the region is 672 // available and can be claimed and then filled. 673 uint destination_count = 0; 674 if (split_info.is_split(cur_region)) { 675 // The current region has been split: the partial object will be copied 676 // to one destination space and the remaining data will be copied to 677 // another destination space. Adjust the initial destination_count and, 678 // if necessary, set the source_region field if the partial object will 679 // cross a destination region boundary. 680 destination_count = split_info.destination_count(); 681 if (destination_count == 2) { 682 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 683 _region_data[dest_idx].set_source_region(cur_region); 684 } 685 } 686 687 HeapWord* const last_addr = dest_addr + words - 1; 688 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 689 const size_t dest_region_2 = addr_to_region_idx(last_addr); 690 691 // Initially assume that the destination regions will be the same and 692 // adjust the value below if necessary. Under this assumption, if 693 // cur_region == dest_region_2, then cur_region will be compacted 694 // completely into itself. 695 destination_count += cur_region == dest_region_2 ? 0 : 1; 696 if (dest_region_1 != dest_region_2) { 697 // Destination regions differ; adjust destination_count. 698 destination_count += 1; 699 // Data from cur_region will be copied to the start of dest_region_2. 700 _region_data[dest_region_2].set_source_region(cur_region); 701 } else if (region_offset(dest_addr) == 0) { 702 // Data from cur_region will be copied to the start of the destination 703 // region. 704 _region_data[dest_region_1].set_source_region(cur_region); 705 } 706 707 _region_data[cur_region].set_destination_count(destination_count); 708 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 709 dest_addr += words; 710 } 711 712 ++cur_region; 713 } 714 715 *target_next = dest_addr; 716 return true; 717 } 718 719 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 720 assert(addr != NULL, "Should detect NULL oop earlier"); 721 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); 722 #ifdef ASSERT 723 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { 724 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); 725 } 726 #endif 727 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); 728 729 // Region covering the object. 730 size_t region_index = addr_to_region_idx(addr); 731 const RegionData* const region_ptr = region(region_index); 732 HeapWord* const region_addr = region_align_down(addr); 733 734 assert(addr < region_addr + RegionSize, "Region does not cover object"); 735 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check"); 736 737 HeapWord* result = region_ptr->destination(); 738 739 // If all the data in the region is live, then the new location of the object 740 // can be calculated from the destination of the region plus the offset of the 741 // object in the region. 742 if (region_ptr->data_size() == RegionSize) { 743 result += pointer_delta(addr, region_addr); 744 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 745 return result; 746 } 747 748 // The new location of the object is 749 // region destination + 750 // size of the partial object extending onto the region + 751 // sizes of the live objects in the Region that are to the left of addr 752 const size_t partial_obj_size = region_ptr->partial_obj_size(); 753 HeapWord* const search_start = region_addr + partial_obj_size; 754 755 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 756 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 757 758 result += partial_obj_size + live_to_left; 759 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 760 return result; 761 } 762 763 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) { 764 klassOop updated_klass; 765 if (PSParallelCompact::should_update_klass(old_klass)) { 766 updated_klass = (klassOop) calc_new_pointer(old_klass); 767 } else { 768 updated_klass = old_klass; 769 } 770 771 return updated_klass; 772 } 773 774 #ifdef ASSERT 775 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 776 { 777 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 778 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 779 for (const size_t* p = beg; p < end; ++p) { 780 assert(*p == 0, "not zero"); 781 } 782 } 783 784 void ParallelCompactData::verify_clear() 785 { 786 verify_clear(_region_vspace); 787 } 788 #endif // #ifdef ASSERT 789 790 #ifdef NOT_PRODUCT 791 ParallelCompactData::RegionData* debug_region(size_t region_index) { 792 ParallelCompactData& sd = PSParallelCompact::summary_data(); 793 return sd.region(region_index); 794 } 795 #endif 796 797 elapsedTimer PSParallelCompact::_accumulated_time; 798 unsigned int PSParallelCompact::_total_invocations = 0; 799 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 800 jlong PSParallelCompact::_time_of_last_gc = 0; 801 CollectorCounters* PSParallelCompact::_counters = NULL; 802 ParMarkBitMap PSParallelCompact::_mark_bitmap; 803 ParallelCompactData PSParallelCompact::_summary_data; 804 805 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 806 807 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } 808 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 809 810 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 811 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 812 813 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true); 814 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false); 815 816 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); } 817 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); } 818 819 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 820 821 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); } 822 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 823 824 void PSParallelCompact::post_initialize() { 825 ParallelScavengeHeap* heap = gc_heap(); 826 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 827 828 MemRegion mr = heap->reserved_region(); 829 _ref_processor = ReferenceProcessor::create_ref_processor( 830 mr, // span 831 true, // atomic_discovery 832 true, // mt_discovery 833 &_is_alive_closure, 834 ParallelGCThreads, 835 ParallelRefProcEnabled); 836 _counters = new CollectorCounters("PSParallelCompact", 1); 837 838 // Initialize static fields in ParCompactionManager. 839 ParCompactionManager::initialize(mark_bitmap()); 840 } 841 842 bool PSParallelCompact::initialize() { 843 ParallelScavengeHeap* heap = gc_heap(); 844 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 845 MemRegion mr = heap->reserved_region(); 846 847 // Was the old gen get allocated successfully? 848 if (!heap->old_gen()->is_allocated()) { 849 return false; 850 } 851 852 initialize_space_info(); 853 initialize_dead_wood_limiter(); 854 855 if (!_mark_bitmap.initialize(mr)) { 856 vm_shutdown_during_initialization("Unable to allocate bit map for " 857 "parallel garbage collection for the requested heap size."); 858 return false; 859 } 860 861 if (!_summary_data.initialize(mr)) { 862 vm_shutdown_during_initialization("Unable to allocate tables for " 863 "parallel garbage collection for the requested heap size."); 864 return false; 865 } 866 867 return true; 868 } 869 870 void PSParallelCompact::initialize_space_info() 871 { 872 memset(&_space_info, 0, sizeof(_space_info)); 873 874 ParallelScavengeHeap* heap = gc_heap(); 875 PSYoungGen* young_gen = heap->young_gen(); 876 MutableSpace* perm_space = heap->perm_gen()->object_space(); 877 878 _space_info[perm_space_id].set_space(perm_space); 879 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 880 _space_info[eden_space_id].set_space(young_gen->eden_space()); 881 _space_info[from_space_id].set_space(young_gen->from_space()); 882 _space_info[to_space_id].set_space(young_gen->to_space()); 883 884 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array()); 885 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 886 887 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top()); 888 if (TraceParallelOldGCDensePrefix) { 889 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT, 890 _space_info[perm_space_id].min_dense_prefix()); 891 } 892 } 893 894 void PSParallelCompact::initialize_dead_wood_limiter() 895 { 896 const size_t max = 100; 897 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 898 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 899 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 900 DEBUG_ONLY(_dwl_initialized = true;) 901 _dwl_adjustment = normal_distribution(1.0); 902 } 903 904 // Simple class for storing info about the heap at the start of GC, to be used 905 // after GC for comparison/printing. 906 class PreGCValues { 907 public: 908 PreGCValues() { } 909 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 910 911 void fill(ParallelScavengeHeap* heap) { 912 _heap_used = heap->used(); 913 _young_gen_used = heap->young_gen()->used_in_bytes(); 914 _old_gen_used = heap->old_gen()->used_in_bytes(); 915 _perm_gen_used = heap->perm_gen()->used_in_bytes(); 916 }; 917 918 size_t heap_used() const { return _heap_used; } 919 size_t young_gen_used() const { return _young_gen_used; } 920 size_t old_gen_used() const { return _old_gen_used; } 921 size_t perm_gen_used() const { return _perm_gen_used; } 922 923 private: 924 size_t _heap_used; 925 size_t _young_gen_used; 926 size_t _old_gen_used; 927 size_t _perm_gen_used; 928 }; 929 930 void 931 PSParallelCompact::clear_data_covering_space(SpaceId id) 932 { 933 // At this point, top is the value before GC, new_top() is the value that will 934 // be set at the end of GC. The marking bitmap is cleared to top; nothing 935 // should be marked above top. The summary data is cleared to the larger of 936 // top & new_top. 937 MutableSpace* const space = _space_info[id].space(); 938 HeapWord* const bot = space->bottom(); 939 HeapWord* const top = space->top(); 940 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 941 942 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 943 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 944 _mark_bitmap.clear_range(beg_bit, end_bit); 945 946 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 947 const size_t end_region = 948 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 949 _summary_data.clear_range(beg_region, end_region); 950 951 // Clear the data used to 'split' regions. 952 SplitInfo& split_info = _space_info[id].split_info(); 953 if (split_info.is_valid()) { 954 split_info.clear(); 955 } 956 DEBUG_ONLY(split_info.verify_clear();) 957 } 958 959 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 960 { 961 // Update the from & to space pointers in space_info, since they are swapped 962 // at each young gen gc. Do the update unconditionally (even though a 963 // promotion failure does not swap spaces) because an unknown number of minor 964 // collections will have swapped the spaces an unknown number of times. 965 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); 966 ParallelScavengeHeap* heap = gc_heap(); 967 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 968 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 969 970 pre_gc_values->fill(heap); 971 972 ParCompactionManager::reset(); 973 NOT_PRODUCT(_mark_bitmap.reset_counters()); 974 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 975 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 976 977 // Increment the invocation count 978 heap->increment_total_collections(true); 979 980 // We need to track unique mark sweep invocations as well. 981 _total_invocations++; 982 983 if (PrintHeapAtGC) { 984 Universe::print_heap_before_gc(); 985 } 986 987 // Fill in TLABs 988 heap->accumulate_statistics_all_tlabs(); 989 heap->ensure_parsability(true); // retire TLABs 990 991 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 992 HandleMark hm; // Discard invalid handles created during verification 993 gclog_or_tty->print(" VerifyBeforeGC:"); 994 Universe::verify(true); 995 } 996 997 // Verify object start arrays 998 if (VerifyObjectStartArray && 999 VerifyBeforeGC) { 1000 heap->old_gen()->verify_object_start_array(); 1001 heap->perm_gen()->verify_object_start_array(); 1002 } 1003 1004 DEBUG_ONLY(mark_bitmap()->verify_clear();) 1005 DEBUG_ONLY(summary_data().verify_clear();) 1006 1007 // Have worker threads release resources the next time they run a task. 1008 gc_task_manager()->release_all_resources(); 1009 } 1010 1011 void PSParallelCompact::post_compact() 1012 { 1013 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); 1014 1015 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 1016 // Clear the marking bitmap, summary data and split info. 1017 clear_data_covering_space(SpaceId(id)); 1018 // Update top(). Must be done after clearing the bitmap and summary data. 1019 _space_info[id].publish_new_top(); 1020 } 1021 1022 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1023 MutableSpace* const from_space = _space_info[from_space_id].space(); 1024 MutableSpace* const to_space = _space_info[to_space_id].space(); 1025 1026 ParallelScavengeHeap* heap = gc_heap(); 1027 bool eden_empty = eden_space->is_empty(); 1028 if (!eden_empty) { 1029 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1030 heap->young_gen(), heap->old_gen()); 1031 } 1032 1033 // Update heap occupancy information which is used as input to the soft ref 1034 // clearing policy at the next gc. 1035 Universe::update_heap_info_at_gc(); 1036 1037 bool young_gen_empty = eden_empty && from_space->is_empty() && 1038 to_space->is_empty(); 1039 1040 BarrierSet* bs = heap->barrier_set(); 1041 if (bs->is_a(BarrierSet::ModRef)) { 1042 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1043 MemRegion old_mr = heap->old_gen()->reserved(); 1044 MemRegion perm_mr = heap->perm_gen()->reserved(); 1045 assert(perm_mr.end() <= old_mr.start(), "Generations out of order"); 1046 1047 if (young_gen_empty) { 1048 modBS->clear(MemRegion(perm_mr.start(), old_mr.end())); 1049 } else { 1050 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end())); 1051 } 1052 } 1053 1054 Threads::gc_epilogue(); 1055 CodeCache::gc_epilogue(); 1056 1057 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1058 1059 ref_processor()->enqueue_discovered_references(NULL); 1060 1061 if (ZapUnusedHeapArea) { 1062 heap->gen_mangle_unused_area(); 1063 } 1064 1065 // Update time of last GC 1066 reset_millis_since_last_gc(); 1067 } 1068 1069 HeapWord* 1070 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1071 bool maximum_compaction) 1072 { 1073 const size_t region_size = ParallelCompactData::RegionSize; 1074 const ParallelCompactData& sd = summary_data(); 1075 1076 const MutableSpace* const space = _space_info[id].space(); 1077 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1078 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1079 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1080 1081 // Skip full regions at the beginning of the space--they are necessarily part 1082 // of the dense prefix. 1083 size_t full_count = 0; 1084 const RegionData* cp; 1085 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1086 ++full_count; 1087 } 1088 1089 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1090 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1091 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1092 if (maximum_compaction || cp == end_cp || interval_ended) { 1093 _maximum_compaction_gc_num = total_invocations(); 1094 return sd.region_to_addr(cp); 1095 } 1096 1097 HeapWord* const new_top = _space_info[id].new_top(); 1098 const size_t space_live = pointer_delta(new_top, space->bottom()); 1099 const size_t space_used = space->used_in_words(); 1100 const size_t space_capacity = space->capacity_in_words(); 1101 1102 const double cur_density = double(space_live) / space_capacity; 1103 const double deadwood_density = 1104 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1105 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1106 1107 if (TraceParallelOldGCDensePrefix) { 1108 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1109 cur_density, deadwood_density, deadwood_goal); 1110 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1111 "space_cap=" SIZE_FORMAT, 1112 space_live, space_used, 1113 space_capacity); 1114 } 1115 1116 // XXX - Use binary search? 1117 HeapWord* dense_prefix = sd.region_to_addr(cp); 1118 const RegionData* full_cp = cp; 1119 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1120 while (cp < end_cp) { 1121 HeapWord* region_destination = cp->destination(); 1122 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1123 if (TraceParallelOldGCDensePrefix && Verbose) { 1124 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1125 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1126 sd.region(cp), region_destination, 1127 dense_prefix, cur_deadwood); 1128 } 1129 1130 if (cur_deadwood >= deadwood_goal) { 1131 // Found the region that has the correct amount of deadwood to the left. 1132 // This typically occurs after crossing a fairly sparse set of regions, so 1133 // iterate backwards over those sparse regions, looking for the region 1134 // that has the lowest density of live objects 'to the right.' 1135 size_t space_to_left = sd.region(cp) * region_size; 1136 size_t live_to_left = space_to_left - cur_deadwood; 1137 size_t space_to_right = space_capacity - space_to_left; 1138 size_t live_to_right = space_live - live_to_left; 1139 double density_to_right = double(live_to_right) / space_to_right; 1140 while (cp > full_cp) { 1141 --cp; 1142 const size_t prev_region_live_to_right = live_to_right - 1143 cp->data_size(); 1144 const size_t prev_region_space_to_right = space_to_right + region_size; 1145 double prev_region_density_to_right = 1146 double(prev_region_live_to_right) / prev_region_space_to_right; 1147 if (density_to_right <= prev_region_density_to_right) { 1148 return dense_prefix; 1149 } 1150 if (TraceParallelOldGCDensePrefix && Verbose) { 1151 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1152 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1153 prev_region_density_to_right); 1154 } 1155 dense_prefix -= region_size; 1156 live_to_right = prev_region_live_to_right; 1157 space_to_right = prev_region_space_to_right; 1158 density_to_right = prev_region_density_to_right; 1159 } 1160 return dense_prefix; 1161 } 1162 1163 dense_prefix += region_size; 1164 ++cp; 1165 } 1166 1167 return dense_prefix; 1168 } 1169 1170 #ifndef PRODUCT 1171 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1172 const SpaceId id, 1173 const bool maximum_compaction, 1174 HeapWord* const addr) 1175 { 1176 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1177 RegionData* const cp = summary_data().region(region_idx); 1178 const MutableSpace* const space = _space_info[id].space(); 1179 HeapWord* const new_top = _space_info[id].new_top(); 1180 1181 const size_t space_live = pointer_delta(new_top, space->bottom()); 1182 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1183 const size_t space_cap = space->capacity_in_words(); 1184 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1185 const size_t live_to_right = new_top - cp->destination(); 1186 const size_t dead_to_right = space->top() - addr - live_to_right; 1187 1188 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1189 "spl=" SIZE_FORMAT " " 1190 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1191 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1192 " ratio=%10.8f", 1193 algorithm, addr, region_idx, 1194 space_live, 1195 dead_to_left, dead_to_left_pct, 1196 dead_to_right, live_to_right, 1197 double(dead_to_right) / live_to_right); 1198 } 1199 #endif // #ifndef PRODUCT 1200 1201 // Return a fraction indicating how much of the generation can be treated as 1202 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1203 // based on the density of live objects in the generation to determine a limit, 1204 // which is then adjusted so the return value is min_percent when the density is 1205 // 1. 1206 // 1207 // The following table shows some return values for a different values of the 1208 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1209 // min_percent is 1. 1210 // 1211 // fraction allowed as dead wood 1212 // ----------------------------------------------------------------- 1213 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1214 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1215 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1216 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1217 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1218 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1219 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1220 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1221 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1222 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1223 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1224 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1225 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1226 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1227 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1228 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1229 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1230 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1231 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1232 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1233 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1234 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1235 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1236 1237 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1238 { 1239 assert(_dwl_initialized, "uninitialized"); 1240 1241 // The raw limit is the value of the normal distribution at x = density. 1242 const double raw_limit = normal_distribution(density); 1243 1244 // Adjust the raw limit so it becomes the minimum when the density is 1. 1245 // 1246 // First subtract the adjustment value (which is simply the precomputed value 1247 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1248 // Then add the minimum value, so the minimum is returned when the density is 1249 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1250 const double min = double(min_percent) / 100.0; 1251 const double limit = raw_limit - _dwl_adjustment + min; 1252 return MAX2(limit, 0.0); 1253 } 1254 1255 ParallelCompactData::RegionData* 1256 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1257 const RegionData* end) 1258 { 1259 const size_t region_size = ParallelCompactData::RegionSize; 1260 ParallelCompactData& sd = summary_data(); 1261 size_t left = sd.region(beg); 1262 size_t right = end > beg ? sd.region(end) - 1 : left; 1263 1264 // Binary search. 1265 while (left < right) { 1266 // Equivalent to (left + right) / 2, but does not overflow. 1267 const size_t middle = left + (right - left) / 2; 1268 RegionData* const middle_ptr = sd.region(middle); 1269 HeapWord* const dest = middle_ptr->destination(); 1270 HeapWord* const addr = sd.region_to_addr(middle); 1271 assert(dest != NULL, "sanity"); 1272 assert(dest <= addr, "must move left"); 1273 1274 if (middle > left && dest < addr) { 1275 right = middle - 1; 1276 } else if (middle < right && middle_ptr->data_size() == region_size) { 1277 left = middle + 1; 1278 } else { 1279 return middle_ptr; 1280 } 1281 } 1282 return sd.region(left); 1283 } 1284 1285 ParallelCompactData::RegionData* 1286 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1287 const RegionData* end, 1288 size_t dead_words) 1289 { 1290 ParallelCompactData& sd = summary_data(); 1291 size_t left = sd.region(beg); 1292 size_t right = end > beg ? sd.region(end) - 1 : left; 1293 1294 // Binary search. 1295 while (left < right) { 1296 // Equivalent to (left + right) / 2, but does not overflow. 1297 const size_t middle = left + (right - left) / 2; 1298 RegionData* const middle_ptr = sd.region(middle); 1299 HeapWord* const dest = middle_ptr->destination(); 1300 HeapWord* const addr = sd.region_to_addr(middle); 1301 assert(dest != NULL, "sanity"); 1302 assert(dest <= addr, "must move left"); 1303 1304 const size_t dead_to_left = pointer_delta(addr, dest); 1305 if (middle > left && dead_to_left > dead_words) { 1306 right = middle - 1; 1307 } else if (middle < right && dead_to_left < dead_words) { 1308 left = middle + 1; 1309 } else { 1310 return middle_ptr; 1311 } 1312 } 1313 return sd.region(left); 1314 } 1315 1316 // The result is valid during the summary phase, after the initial summarization 1317 // of each space into itself, and before final summarization. 1318 inline double 1319 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1320 HeapWord* const bottom, 1321 HeapWord* const top, 1322 HeapWord* const new_top) 1323 { 1324 ParallelCompactData& sd = summary_data(); 1325 1326 assert(cp != NULL, "sanity"); 1327 assert(bottom != NULL, "sanity"); 1328 assert(top != NULL, "sanity"); 1329 assert(new_top != NULL, "sanity"); 1330 assert(top >= new_top, "summary data problem?"); 1331 assert(new_top > bottom, "space is empty; should not be here"); 1332 assert(new_top >= cp->destination(), "sanity"); 1333 assert(top >= sd.region_to_addr(cp), "sanity"); 1334 1335 HeapWord* const destination = cp->destination(); 1336 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1337 const size_t compacted_region_live = pointer_delta(new_top, destination); 1338 const size_t compacted_region_used = pointer_delta(top, 1339 sd.region_to_addr(cp)); 1340 const size_t reclaimable = compacted_region_used - compacted_region_live; 1341 1342 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1343 return double(reclaimable) / divisor; 1344 } 1345 1346 // Return the address of the end of the dense prefix, a.k.a. the start of the 1347 // compacted region. The address is always on a region boundary. 1348 // 1349 // Completely full regions at the left are skipped, since no compaction can 1350 // occur in those regions. Then the maximum amount of dead wood to allow is 1351 // computed, based on the density (amount live / capacity) of the generation; 1352 // the region with approximately that amount of dead space to the left is 1353 // identified as the limit region. Regions between the last completely full 1354 // region and the limit region are scanned and the one that has the best 1355 // (maximum) reclaimed_ratio() is selected. 1356 HeapWord* 1357 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1358 bool maximum_compaction) 1359 { 1360 if (ParallelOldGCSplitALot) { 1361 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1362 // The value was chosen to provoke splitting a young gen space; use it. 1363 return _space_info[id].dense_prefix(); 1364 } 1365 } 1366 1367 const size_t region_size = ParallelCompactData::RegionSize; 1368 const ParallelCompactData& sd = summary_data(); 1369 1370 const MutableSpace* const space = _space_info[id].space(); 1371 HeapWord* const top = space->top(); 1372 HeapWord* const top_aligned_up = sd.region_align_up(top); 1373 HeapWord* const new_top = _space_info[id].new_top(); 1374 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1375 HeapWord* const bottom = space->bottom(); 1376 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1377 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1378 const RegionData* const new_top_cp = 1379 sd.addr_to_region_ptr(new_top_aligned_up); 1380 1381 // Skip full regions at the beginning of the space--they are necessarily part 1382 // of the dense prefix. 1383 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1384 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1385 space->is_empty(), "no dead space allowed to the left"); 1386 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1387 "region must have dead space"); 1388 1389 // The gc number is saved whenever a maximum compaction is done, and used to 1390 // determine when the maximum compaction interval has expired. This avoids 1391 // successive max compactions for different reasons. 1392 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1393 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1394 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1395 total_invocations() == HeapFirstMaximumCompactionCount; 1396 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1397 _maximum_compaction_gc_num = total_invocations(); 1398 return sd.region_to_addr(full_cp); 1399 } 1400 1401 const size_t space_live = pointer_delta(new_top, bottom); 1402 const size_t space_used = space->used_in_words(); 1403 const size_t space_capacity = space->capacity_in_words(); 1404 1405 const double density = double(space_live) / double(space_capacity); 1406 const size_t min_percent_free = 1407 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio; 1408 const double limiter = dead_wood_limiter(density, min_percent_free); 1409 const size_t dead_wood_max = space_used - space_live; 1410 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1411 dead_wood_max); 1412 1413 if (TraceParallelOldGCDensePrefix) { 1414 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1415 "space_cap=" SIZE_FORMAT, 1416 space_live, space_used, 1417 space_capacity); 1418 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1419 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1420 density, min_percent_free, limiter, 1421 dead_wood_max, dead_wood_limit); 1422 } 1423 1424 // Locate the region with the desired amount of dead space to the left. 1425 const RegionData* const limit_cp = 1426 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1427 1428 // Scan from the first region with dead space to the limit region and find the 1429 // one with the best (largest) reclaimed ratio. 1430 double best_ratio = 0.0; 1431 const RegionData* best_cp = full_cp; 1432 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1433 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1434 if (tmp_ratio > best_ratio) { 1435 best_cp = cp; 1436 best_ratio = tmp_ratio; 1437 } 1438 } 1439 1440 #if 0 1441 // Something to consider: if the region with the best ratio is 'close to' the 1442 // first region w/free space, choose the first region with free space 1443 // ("first-free"). The first-free region is usually near the start of the 1444 // heap, which means we are copying most of the heap already, so copy a bit 1445 // more to get complete compaction. 1446 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1447 _maximum_compaction_gc_num = total_invocations(); 1448 best_cp = full_cp; 1449 } 1450 #endif // #if 0 1451 1452 return sd.region_to_addr(best_cp); 1453 } 1454 1455 #ifndef PRODUCT 1456 void 1457 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1458 size_t words) 1459 { 1460 if (TraceParallelOldGCSummaryPhase) { 1461 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1462 SIZE_FORMAT, start, start + words, words); 1463 } 1464 1465 ObjectStartArray* const start_array = _space_info[id].start_array(); 1466 CollectedHeap::fill_with_objects(start, words); 1467 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1468 _mark_bitmap.mark_obj(p, words); 1469 _summary_data.add_obj(p, words); 1470 start_array->allocate_block(p); 1471 } 1472 } 1473 1474 void 1475 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1476 { 1477 ParallelCompactData& sd = summary_data(); 1478 MutableSpace* space = _space_info[id].space(); 1479 1480 // Find the source and destination start addresses. 1481 HeapWord* const src_addr = sd.region_align_down(start); 1482 HeapWord* dst_addr; 1483 if (src_addr < start) { 1484 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1485 } else if (src_addr > space->bottom()) { 1486 // The start (the original top() value) is aligned to a region boundary so 1487 // the associated region does not have a destination. Compute the 1488 // destination from the previous region. 1489 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1490 dst_addr = cp->destination() + cp->data_size(); 1491 } else { 1492 // Filling the entire space. 1493 dst_addr = space->bottom(); 1494 } 1495 assert(dst_addr != NULL, "sanity"); 1496 1497 // Update the summary data. 1498 bool result = _summary_data.summarize(_space_info[id].split_info(), 1499 src_addr, space->top(), NULL, 1500 dst_addr, space->end(), 1501 _space_info[id].new_top_addr()); 1502 assert(result, "should not fail: bad filler object size"); 1503 } 1504 1505 void 1506 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1507 { 1508 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1509 return; 1510 } 1511 1512 MutableSpace* const space = _space_info[id].space(); 1513 if (space->is_empty()) { 1514 HeapWord* b = space->bottom(); 1515 HeapWord* t = b + space->capacity_in_words() / 2; 1516 space->set_top(t); 1517 if (ZapUnusedHeapArea) { 1518 space->set_top_for_allocations(); 1519 } 1520 1521 size_t min_size = CollectedHeap::min_fill_size(); 1522 size_t obj_len = min_size; 1523 while (b + obj_len <= t) { 1524 CollectedHeap::fill_with_object(b, obj_len); 1525 mark_bitmap()->mark_obj(b, obj_len); 1526 summary_data().add_obj(b, obj_len); 1527 b += obj_len; 1528 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1529 } 1530 if (b < t) { 1531 // The loop didn't completely fill to t (top); adjust top downward. 1532 space->set_top(b); 1533 if (ZapUnusedHeapArea) { 1534 space->set_top_for_allocations(); 1535 } 1536 } 1537 1538 HeapWord** nta = _space_info[id].new_top_addr(); 1539 bool result = summary_data().summarize(_space_info[id].split_info(), 1540 space->bottom(), space->top(), NULL, 1541 space->bottom(), space->end(), nta); 1542 assert(result, "space must fit into itself"); 1543 } 1544 } 1545 1546 void 1547 PSParallelCompact::provoke_split(bool & max_compaction) 1548 { 1549 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1550 return; 1551 } 1552 1553 const size_t region_size = ParallelCompactData::RegionSize; 1554 ParallelCompactData& sd = summary_data(); 1555 1556 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1557 MutableSpace* const from_space = _space_info[from_space_id].space(); 1558 const size_t eden_live = pointer_delta(eden_space->top(), 1559 _space_info[eden_space_id].new_top()); 1560 const size_t from_live = pointer_delta(from_space->top(), 1561 _space_info[from_space_id].new_top()); 1562 1563 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1564 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1565 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1566 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1567 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1568 1569 // Choose the space to split; need at least 2 regions live (or fillable). 1570 SpaceId id; 1571 MutableSpace* space; 1572 size_t live_words; 1573 size_t fill_words; 1574 if (eden_live + eden_fillable >= region_size * 2) { 1575 id = eden_space_id; 1576 space = eden_space; 1577 live_words = eden_live; 1578 fill_words = eden_fillable; 1579 } else if (from_live + from_fillable >= region_size * 2) { 1580 id = from_space_id; 1581 space = from_space; 1582 live_words = from_live; 1583 fill_words = from_fillable; 1584 } else { 1585 return; // Give up. 1586 } 1587 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1588 1589 if (live_words < region_size * 2) { 1590 // Fill from top() to end() w/live objects of mixed sizes. 1591 HeapWord* const fill_start = space->top(); 1592 live_words += fill_words; 1593 1594 space->set_top(fill_start + fill_words); 1595 if (ZapUnusedHeapArea) { 1596 space->set_top_for_allocations(); 1597 } 1598 1599 HeapWord* cur_addr = fill_start; 1600 while (fill_words > 0) { 1601 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1602 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1603 if (fill_words - cur_size < min_fill_size) { 1604 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1605 } 1606 1607 CollectedHeap::fill_with_object(cur_addr, cur_size); 1608 mark_bitmap()->mark_obj(cur_addr, cur_size); 1609 sd.add_obj(cur_addr, cur_size); 1610 1611 cur_addr += cur_size; 1612 fill_words -= cur_size; 1613 } 1614 1615 summarize_new_objects(id, fill_start); 1616 } 1617 1618 max_compaction = false; 1619 1620 // Manipulate the old gen so that it has room for about half of the live data 1621 // in the target young gen space (live_words / 2). 1622 id = old_space_id; 1623 space = _space_info[id].space(); 1624 const size_t free_at_end = space->free_in_words(); 1625 const size_t free_target = align_object_size(live_words / 2); 1626 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1627 1628 if (free_at_end >= free_target + min_fill_size) { 1629 // Fill space above top() and set the dense prefix so everything survives. 1630 HeapWord* const fill_start = space->top(); 1631 const size_t fill_size = free_at_end - free_target; 1632 space->set_top(space->top() + fill_size); 1633 if (ZapUnusedHeapArea) { 1634 space->set_top_for_allocations(); 1635 } 1636 fill_with_live_objects(id, fill_start, fill_size); 1637 summarize_new_objects(id, fill_start); 1638 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1639 } else if (dead + free_at_end > free_target) { 1640 // Find a dense prefix that makes the right amount of space available. 1641 HeapWord* cur = sd.region_align_down(space->top()); 1642 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1643 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1644 while (dead_to_right < free_target) { 1645 cur -= region_size; 1646 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1647 dead_to_right = pointer_delta(space->end(), cur_destination); 1648 } 1649 _space_info[id].set_dense_prefix(cur); 1650 } 1651 } 1652 #endif // #ifndef PRODUCT 1653 1654 void PSParallelCompact::summarize_spaces_quick() 1655 { 1656 for (unsigned int i = 0; i < last_space_id; ++i) { 1657 const MutableSpace* space = _space_info[i].space(); 1658 HeapWord** nta = _space_info[i].new_top_addr(); 1659 bool result = _summary_data.summarize(_space_info[i].split_info(), 1660 space->bottom(), space->top(), NULL, 1661 space->bottom(), space->end(), nta); 1662 assert(result, "space must fit into itself"); 1663 _space_info[i].set_dense_prefix(space->bottom()); 1664 } 1665 1666 #ifndef PRODUCT 1667 if (ParallelOldGCSplitALot) { 1668 provoke_split_fill_survivor(to_space_id); 1669 } 1670 #endif // #ifndef PRODUCT 1671 } 1672 1673 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1674 { 1675 HeapWord* const dense_prefix_end = dense_prefix(id); 1676 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1677 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1678 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1679 // Only enough dead space is filled so that any remaining dead space to the 1680 // left is larger than the minimum filler object. (The remainder is filled 1681 // during the copy/update phase.) 1682 // 1683 // The size of the dead space to the right of the boundary is not a 1684 // concern, since compaction will be able to use whatever space is 1685 // available. 1686 // 1687 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1688 // surrounds the space to be filled with an object. 1689 // 1690 // In the 32-bit VM, each bit represents two 32-bit words: 1691 // +---+ 1692 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1693 // end_bits: ... x x x | 0 | || 0 x x ... 1694 // +---+ 1695 // 1696 // In the 64-bit VM, each bit represents one 64-bit word: 1697 // +------------+ 1698 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1699 // end_bits: ... x x 1 | 0 || 0 | x x ... 1700 // +------------+ 1701 // +-------+ 1702 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1703 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1704 // +-------+ 1705 // +-----------+ 1706 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1707 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1708 // +-----------+ 1709 // +-------+ 1710 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1711 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1712 // +-------+ 1713 1714 // Initially assume case a, c or e will apply. 1715 size_t obj_len = CollectedHeap::min_fill_size(); 1716 HeapWord* obj_beg = dense_prefix_end - obj_len; 1717 1718 #ifdef _LP64 1719 if (MinObjAlignment > 1) { // object alignment > heap word size 1720 // Cases a, c or e. 1721 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1722 // Case b above. 1723 obj_beg = dense_prefix_end - 1; 1724 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1725 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1726 // Case d above. 1727 obj_beg = dense_prefix_end - 3; 1728 obj_len = 3; 1729 } 1730 #endif // #ifdef _LP64 1731 1732 CollectedHeap::fill_with_object(obj_beg, obj_len); 1733 _mark_bitmap.mark_obj(obj_beg, obj_len); 1734 _summary_data.add_obj(obj_beg, obj_len); 1735 assert(start_array(id) != NULL, "sanity"); 1736 start_array(id)->allocate_block(obj_beg); 1737 } 1738 } 1739 1740 void 1741 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1742 { 1743 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1744 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1745 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1746 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1747 cur->set_source_region(0); 1748 } 1749 } 1750 1751 void 1752 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1753 { 1754 assert(id < last_space_id, "id out of range"); 1755 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1756 ParallelOldGCSplitALot && id == old_space_id, 1757 "should have been reset in summarize_spaces_quick()"); 1758 1759 const MutableSpace* space = _space_info[id].space(); 1760 if (_space_info[id].new_top() != space->bottom()) { 1761 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1762 _space_info[id].set_dense_prefix(dense_prefix_end); 1763 1764 #ifndef PRODUCT 1765 if (TraceParallelOldGCDensePrefix) { 1766 print_dense_prefix_stats("ratio", id, maximum_compaction, 1767 dense_prefix_end); 1768 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1769 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1770 } 1771 #endif // #ifndef PRODUCT 1772 1773 // Recompute the summary data, taking into account the dense prefix. If 1774 // every last byte will be reclaimed, then the existing summary data which 1775 // compacts everything can be left in place. 1776 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1777 // If dead space crosses the dense prefix boundary, it is (at least 1778 // partially) filled with a dummy object, marked live and added to the 1779 // summary data. This simplifies the copy/update phase and must be done 1780 // before the final locations of objects are determined, to prevent 1781 // leaving a fragment of dead space that is too small to fill. 1782 fill_dense_prefix_end(id); 1783 1784 // Compute the destination of each Region, and thus each object. 1785 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1786 _summary_data.summarize(_space_info[id].split_info(), 1787 dense_prefix_end, space->top(), NULL, 1788 dense_prefix_end, space->end(), 1789 _space_info[id].new_top_addr()); 1790 } 1791 } 1792 1793 if (TraceParallelOldGCSummaryPhase) { 1794 const size_t region_size = ParallelCompactData::RegionSize; 1795 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1796 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1797 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1798 HeapWord* const new_top = _space_info[id].new_top(); 1799 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1800 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1801 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1802 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1803 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1804 id, space->capacity_in_words(), dense_prefix_end, 1805 dp_region, dp_words / region_size, 1806 cr_words / region_size, new_top); 1807 } 1808 } 1809 1810 #ifndef PRODUCT 1811 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1812 HeapWord* dst_beg, HeapWord* dst_end, 1813 SpaceId src_space_id, 1814 HeapWord* src_beg, HeapWord* src_end) 1815 { 1816 if (TraceParallelOldGCSummaryPhase) { 1817 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1818 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1819 SIZE_FORMAT "-" SIZE_FORMAT " " 1820 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1821 SIZE_FORMAT "-" SIZE_FORMAT, 1822 src_space_id, space_names[src_space_id], 1823 dst_space_id, space_names[dst_space_id], 1824 src_beg, src_end, 1825 _summary_data.addr_to_region_idx(src_beg), 1826 _summary_data.addr_to_region_idx(src_end), 1827 dst_beg, dst_end, 1828 _summary_data.addr_to_region_idx(dst_beg), 1829 _summary_data.addr_to_region_idx(dst_end)); 1830 } 1831 } 1832 #endif // #ifndef PRODUCT 1833 1834 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1835 bool maximum_compaction) 1836 { 1837 EventMark m("2 summarize"); 1838 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); 1839 // trace("2"); 1840 1841 #ifdef ASSERT 1842 if (TraceParallelOldGCMarkingPhase) { 1843 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1844 "add_obj_bytes=" SIZE_FORMAT, 1845 add_obj_count, add_obj_size * HeapWordSize); 1846 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1847 "mark_bitmap_bytes=" SIZE_FORMAT, 1848 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1849 } 1850 #endif // #ifdef ASSERT 1851 1852 // Quick summarization of each space into itself, to see how much is live. 1853 summarize_spaces_quick(); 1854 1855 if (TraceParallelOldGCSummaryPhase) { 1856 tty->print_cr("summary_phase: after summarizing each space to self"); 1857 Universe::print(); 1858 NOT_PRODUCT(print_region_ranges()); 1859 if (Verbose) { 1860 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1861 } 1862 } 1863 1864 // The amount of live data that will end up in old space (assuming it fits). 1865 size_t old_space_total_live = 0; 1866 assert(perm_space_id < old_space_id, "should not count perm data here"); 1867 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1868 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1869 _space_info[id].space()->bottom()); 1870 } 1871 1872 MutableSpace* const old_space = _space_info[old_space_id].space(); 1873 const size_t old_capacity = old_space->capacity_in_words(); 1874 if (old_space_total_live > old_capacity) { 1875 // XXX - should also try to expand 1876 maximum_compaction = true; 1877 } 1878 #ifndef PRODUCT 1879 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1880 provoke_split(maximum_compaction); 1881 } 1882 #endif // #ifndef PRODUCT 1883 1884 // Permanent and Old generations. 1885 summarize_space(perm_space_id, maximum_compaction); 1886 summarize_space(old_space_id, maximum_compaction); 1887 1888 // Summarize the remaining spaces in the young gen. The initial target space 1889 // is the old gen. If a space does not fit entirely into the target, then the 1890 // remainder is compacted into the space itself and that space becomes the new 1891 // target. 1892 SpaceId dst_space_id = old_space_id; 1893 HeapWord* dst_space_end = old_space->end(); 1894 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1895 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1896 const MutableSpace* space = _space_info[id].space(); 1897 const size_t live = pointer_delta(_space_info[id].new_top(), 1898 space->bottom()); 1899 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1900 1901 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1902 SpaceId(id), space->bottom(), space->top());) 1903 if (live > 0 && live <= available) { 1904 // All the live data will fit. 1905 bool done = _summary_data.summarize(_space_info[id].split_info(), 1906 space->bottom(), space->top(), 1907 NULL, 1908 *new_top_addr, dst_space_end, 1909 new_top_addr); 1910 assert(done, "space must fit into old gen"); 1911 1912 // Reset the new_top value for the space. 1913 _space_info[id].set_new_top(space->bottom()); 1914 } else if (live > 0) { 1915 // Attempt to fit part of the source space into the target space. 1916 HeapWord* next_src_addr = NULL; 1917 bool done = _summary_data.summarize(_space_info[id].split_info(), 1918 space->bottom(), space->top(), 1919 &next_src_addr, 1920 *new_top_addr, dst_space_end, 1921 new_top_addr); 1922 assert(!done, "space should not fit into old gen"); 1923 assert(next_src_addr != NULL, "sanity"); 1924 1925 // The source space becomes the new target, so the remainder is compacted 1926 // within the space itself. 1927 dst_space_id = SpaceId(id); 1928 dst_space_end = space->end(); 1929 new_top_addr = _space_info[id].new_top_addr(); 1930 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1931 space->bottom(), dst_space_end, 1932 SpaceId(id), next_src_addr, space->top());) 1933 done = _summary_data.summarize(_space_info[id].split_info(), 1934 next_src_addr, space->top(), 1935 NULL, 1936 space->bottom(), dst_space_end, 1937 new_top_addr); 1938 assert(done, "space must fit when compacted into itself"); 1939 assert(*new_top_addr <= space->top(), "usage should not grow"); 1940 } 1941 } 1942 1943 if (TraceParallelOldGCSummaryPhase) { 1944 tty->print_cr("summary_phase: after final summarization"); 1945 Universe::print(); 1946 NOT_PRODUCT(print_region_ranges()); 1947 if (Verbose) { 1948 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1949 } 1950 } 1951 } 1952 1953 // This method should contain all heap-specific policy for invoking a full 1954 // collection. invoke_no_policy() will only attempt to compact the heap; it 1955 // will do nothing further. If we need to bail out for policy reasons, scavenge 1956 // before full gc, or any other specialized behavior, it needs to be added here. 1957 // 1958 // Note that this method should only be called from the vm_thread while at a 1959 // safepoint. 1960 // 1961 // Note that the all_soft_refs_clear flag in the collector policy 1962 // may be true because this method can be called without intervening 1963 // activity. For example when the heap space is tight and full measure 1964 // are being taken to free space. 1965 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1966 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1967 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1968 "should be in vm thread"); 1969 1970 ParallelScavengeHeap* heap = gc_heap(); 1971 GCCause::Cause gc_cause = heap->gc_cause(); 1972 assert(!heap->is_gc_active(), "not reentrant"); 1973 1974 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1975 IsGCActiveMark mark; 1976 1977 if (ScavengeBeforeFullGC) { 1978 PSScavenge::invoke_no_policy(); 1979 } 1980 1981 const bool clear_all_soft_refs = 1982 heap->collector_policy()->should_clear_all_soft_refs(); 1983 1984 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1985 maximum_heap_compaction); 1986 } 1987 1988 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1989 size_t addr_region_index = addr_to_region_idx(addr); 1990 return region_index == addr_region_index; 1991 } 1992 1993 // This method contains no policy. You should probably 1994 // be calling invoke() instead. 1995 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1996 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1997 assert(ref_processor() != NULL, "Sanity"); 1998 1999 if (GC_locker::check_active_before_gc()) { 2000 return; 2001 } 2002 2003 TimeStamp marking_start; 2004 TimeStamp compaction_start; 2005 TimeStamp collection_exit; 2006 2007 ParallelScavengeHeap* heap = gc_heap(); 2008 GCCause::Cause gc_cause = heap->gc_cause(); 2009 PSYoungGen* young_gen = heap->young_gen(); 2010 PSOldGen* old_gen = heap->old_gen(); 2011 PSPermGen* perm_gen = heap->perm_gen(); 2012 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 2013 2014 // The scope of casr should end after code that can change 2015 // CollectorPolicy::_should_clear_all_soft_refs. 2016 ClearedAllSoftRefs casr(maximum_heap_compaction, 2017 heap->collector_policy()); 2018 2019 if (ZapUnusedHeapArea) { 2020 // Save information needed to minimize mangling 2021 heap->record_gen_tops_before_GC(); 2022 } 2023 2024 heap->pre_full_gc_dump(); 2025 2026 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 2027 2028 // Make sure data structures are sane, make the heap parsable, and do other 2029 // miscellaneous bookkeeping. 2030 PreGCValues pre_gc_values; 2031 pre_compact(&pre_gc_values); 2032 2033 // Get the compaction manager reserved for the VM thread. 2034 ParCompactionManager* const vmthread_cm = 2035 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2036 2037 // Place after pre_compact() where the number of invocations is incremented. 2038 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2039 2040 { 2041 ResourceMark rm; 2042 HandleMark hm; 2043 2044 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc; 2045 2046 // This is useful for debugging but don't change the output the 2047 // the customer sees. 2048 const char* gc_cause_str = "Full GC"; 2049 if (is_system_gc && PrintGCDetails) { 2050 gc_cause_str = "Full GC (System)"; 2051 } 2052 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2053 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2054 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty); 2055 TraceCollectorStats tcs(counters()); 2056 TraceMemoryManagerStats tms(true /* Full GC */); 2057 2058 if (TraceGen1Time) accumulated_time()->start(); 2059 2060 // Let the size policy know we're starting 2061 size_policy->major_collection_begin(); 2062 2063 // When collecting the permanent generation methodOops may be moving, 2064 // so we either have to flush all bcp data or convert it into bci. 2065 CodeCache::gc_prologue(); 2066 Threads::gc_prologue(); 2067 2068 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2069 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2070 2071 ref_processor()->enable_discovery(); 2072 ref_processor()->setup_policy(maximum_heap_compaction); 2073 2074 bool marked_for_unloading = false; 2075 2076 marking_start.update(); 2077 marking_phase(vmthread_cm, maximum_heap_compaction); 2078 2079 #ifndef PRODUCT 2080 if (TraceParallelOldGCMarkingPhase) { 2081 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " 2082 "cas_by_another %d", 2083 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), 2084 mark_bitmap()->cas_by_another()); 2085 } 2086 #endif // #ifndef PRODUCT 2087 2088 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc; 2089 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2090 2091 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2092 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2093 2094 // adjust_roots() updates Universe::_intArrayKlassObj which is 2095 // needed by the compaction for filling holes in the dense prefix. 2096 adjust_roots(); 2097 2098 compaction_start.update(); 2099 // Does the perm gen always have to be done serially because 2100 // klasses are used in the update of an object? 2101 compact_perm(vmthread_cm); 2102 2103 if (UseParallelOldGCCompacting) { 2104 compact(); 2105 } else { 2106 compact_serial(vmthread_cm); 2107 } 2108 2109 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2110 // done before resizing. 2111 post_compact(); 2112 2113 // Let the size policy know we're done 2114 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2115 2116 if (UseAdaptiveSizePolicy) { 2117 if (PrintAdaptiveSizePolicy) { 2118 gclog_or_tty->print("AdaptiveSizeStart: "); 2119 gclog_or_tty->stamp(); 2120 gclog_or_tty->print_cr(" collection: %d ", 2121 heap->total_collections()); 2122 if (Verbose) { 2123 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d" 2124 " perm_gen_capacity: %d ", 2125 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(), 2126 perm_gen->capacity_in_bytes()); 2127 } 2128 } 2129 2130 // Don't check if the size_policy is ready here. Let 2131 // the size_policy check that internally. 2132 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2133 ((gc_cause != GCCause::_java_lang_system_gc) || 2134 UseAdaptiveSizePolicyWithSystemGC)) { 2135 // Calculate optimal free space amounts 2136 assert(young_gen->max_size() > 2137 young_gen->from_space()->capacity_in_bytes() + 2138 young_gen->to_space()->capacity_in_bytes(), 2139 "Sizes of space in young gen are out-of-bounds"); 2140 size_t max_eden_size = young_gen->max_size() - 2141 young_gen->from_space()->capacity_in_bytes() - 2142 young_gen->to_space()->capacity_in_bytes(); 2143 size_policy->compute_generation_free_space( 2144 young_gen->used_in_bytes(), 2145 young_gen->eden_space()->used_in_bytes(), 2146 old_gen->used_in_bytes(), 2147 perm_gen->used_in_bytes(), 2148 young_gen->eden_space()->capacity_in_bytes(), 2149 old_gen->max_gen_size(), 2150 max_eden_size, 2151 true /* full gc*/, 2152 gc_cause, 2153 heap->collector_policy()); 2154 2155 heap->resize_old_gen( 2156 size_policy->calculated_old_free_size_in_bytes()); 2157 2158 // Don't resize the young generation at an major collection. A 2159 // desired young generation size may have been calculated but 2160 // resizing the young generation complicates the code because the 2161 // resizing of the old generation may have moved the boundary 2162 // between the young generation and the old generation. Let the 2163 // young generation resizing happen at the minor collections. 2164 } 2165 if (PrintAdaptiveSizePolicy) { 2166 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2167 heap->total_collections()); 2168 } 2169 } 2170 2171 if (UsePerfData) { 2172 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2173 counters->update_counters(); 2174 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2175 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2176 } 2177 2178 heap->resize_all_tlabs(); 2179 2180 // We collected the perm gen, so we'll resize it here. 2181 perm_gen->compute_new_size(pre_gc_values.perm_gen_used()); 2182 2183 if (TraceGen1Time) accumulated_time()->stop(); 2184 2185 if (PrintGC) { 2186 if (PrintGCDetails) { 2187 // No GC timestamp here. This is after GC so it would be confusing. 2188 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2189 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2190 heap->print_heap_change(pre_gc_values.heap_used()); 2191 // Print perm gen last (print_heap_change() excludes the perm gen). 2192 perm_gen->print_used_change(pre_gc_values.perm_gen_used()); 2193 } else { 2194 heap->print_heap_change(pre_gc_values.heap_used()); 2195 } 2196 } 2197 2198 // Track memory usage and detect low memory 2199 MemoryService::track_memory_usage(); 2200 heap->update_counters(); 2201 } 2202 2203 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2204 HandleMark hm; // Discard invalid handles created during verification 2205 gclog_or_tty->print(" VerifyAfterGC:"); 2206 Universe::verify(false); 2207 } 2208 2209 // Re-verify object start arrays 2210 if (VerifyObjectStartArray && 2211 VerifyAfterGC) { 2212 old_gen->verify_object_start_array(); 2213 perm_gen->verify_object_start_array(); 2214 } 2215 2216 if (ZapUnusedHeapArea) { 2217 old_gen->object_space()->check_mangled_unused_area_complete(); 2218 perm_gen->object_space()->check_mangled_unused_area_complete(); 2219 } 2220 2221 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2222 2223 collection_exit.update(); 2224 2225 if (PrintHeapAtGC) { 2226 Universe::print_heap_after_gc(); 2227 } 2228 if (PrintGCTaskTimeStamps) { 2229 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2230 INT64_FORMAT, 2231 marking_start.ticks(), compaction_start.ticks(), 2232 collection_exit.ticks()); 2233 gc_task_manager()->print_task_time_stamps(); 2234 } 2235 2236 heap->post_full_gc_dump(); 2237 2238 #ifdef TRACESPINNING 2239 ParallelTaskTerminator::print_termination_counts(); 2240 #endif 2241 } 2242 2243 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2244 PSYoungGen* young_gen, 2245 PSOldGen* old_gen) { 2246 MutableSpace* const eden_space = young_gen->eden_space(); 2247 assert(!eden_space->is_empty(), "eden must be non-empty"); 2248 assert(young_gen->virtual_space()->alignment() == 2249 old_gen->virtual_space()->alignment(), "alignments do not match"); 2250 2251 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2252 return false; 2253 } 2254 2255 // Both generations must be completely committed. 2256 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2257 return false; 2258 } 2259 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2260 return false; 2261 } 2262 2263 // Figure out how much to take from eden. Include the average amount promoted 2264 // in the total; otherwise the next young gen GC will simply bail out to a 2265 // full GC. 2266 const size_t alignment = old_gen->virtual_space()->alignment(); 2267 const size_t eden_used = eden_space->used_in_bytes(); 2268 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2269 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2270 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2271 2272 if (absorb_size >= eden_capacity) { 2273 return false; // Must leave some space in eden. 2274 } 2275 2276 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2277 if (new_young_size < young_gen->min_gen_size()) { 2278 return false; // Respect young gen minimum size. 2279 } 2280 2281 if (TraceAdaptiveGCBoundary && Verbose) { 2282 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2283 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2284 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2285 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2286 absorb_size / K, 2287 eden_capacity / K, (eden_capacity - absorb_size) / K, 2288 young_gen->from_space()->used_in_bytes() / K, 2289 young_gen->to_space()->used_in_bytes() / K, 2290 young_gen->capacity_in_bytes() / K, new_young_size / K); 2291 } 2292 2293 // Fill the unused part of the old gen. 2294 MutableSpace* const old_space = old_gen->object_space(); 2295 HeapWord* const unused_start = old_space->top(); 2296 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2297 2298 if (unused_words > 0) { 2299 if (unused_words < CollectedHeap::min_fill_size()) { 2300 return false; // If the old gen cannot be filled, must give up. 2301 } 2302 CollectedHeap::fill_with_objects(unused_start, unused_words); 2303 } 2304 2305 // Take the live data from eden and set both top and end in the old gen to 2306 // eden top. (Need to set end because reset_after_change() mangles the region 2307 // from end to virtual_space->high() in debug builds). 2308 HeapWord* const new_top = eden_space->top(); 2309 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2310 absorb_size); 2311 young_gen->reset_after_change(); 2312 old_space->set_top(new_top); 2313 old_space->set_end(new_top); 2314 old_gen->reset_after_change(); 2315 2316 // Update the object start array for the filler object and the data from eden. 2317 ObjectStartArray* const start_array = old_gen->start_array(); 2318 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2319 start_array->allocate_block(p); 2320 } 2321 2322 // Could update the promoted average here, but it is not typically updated at 2323 // full GCs and the value to use is unclear. Something like 2324 // 2325 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2326 2327 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2328 return true; 2329 } 2330 2331 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2332 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2333 "shouldn't return NULL"); 2334 return ParallelScavengeHeap::gc_task_manager(); 2335 } 2336 2337 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2338 bool maximum_heap_compaction) { 2339 // Recursively traverse all live objects and mark them 2340 EventMark m("1 mark object"); 2341 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); 2342 2343 ParallelScavengeHeap* heap = gc_heap(); 2344 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2345 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2346 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2347 2348 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2349 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2350 2351 { 2352 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); 2353 ParallelScavengeHeap::ParStrongRootsScope psrs; 2354 2355 GCTaskQueue* q = GCTaskQueue::create(); 2356 2357 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2358 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2359 // We scan the thread roots in parallel 2360 Threads::create_thread_roots_marking_tasks(q); 2361 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2362 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2363 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2364 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2365 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2366 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols)); 2367 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2368 2369 if (parallel_gc_threads > 1) { 2370 for (uint j = 0; j < parallel_gc_threads; j++) { 2371 q->enqueue(new StealMarkingTask(&terminator)); 2372 } 2373 } 2374 2375 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2376 q->enqueue(fin); 2377 2378 gc_task_manager()->add_list(q); 2379 2380 fin->wait_for(); 2381 2382 // We have to release the barrier tasks! 2383 WaitForBarrierGCTask::destroy(fin); 2384 } 2385 2386 // Process reference objects found during marking 2387 { 2388 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); 2389 if (ref_processor()->processing_is_mt()) { 2390 RefProcTaskExecutor task_executor; 2391 ref_processor()->process_discovered_references( 2392 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2393 &task_executor); 2394 } else { 2395 ref_processor()->process_discovered_references( 2396 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL); 2397 } 2398 } 2399 2400 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); 2401 // Follow system dictionary roots and unload classes. 2402 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2403 2404 // Follow code cache roots. 2405 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure, 2406 purged_class); 2407 cm->follow_marking_stacks(); // Flush marking stack. 2408 2409 // Update subklass/sibling/implementor links of live klasses 2410 // revisit_klass_stack is used in follow_weak_klass_links(). 2411 follow_weak_klass_links(); 2412 2413 // Revisit memoized MDO's and clear any unmarked weak refs 2414 follow_mdo_weak_refs(); 2415 2416 // Visit symbol and interned string tables and delete unmarked oops 2417 SymbolTable::unlink(is_alive_closure()); 2418 StringTable::unlink(is_alive_closure()); 2419 2420 assert(cm->marking_stacks_empty(), "marking stacks should be empty"); 2421 } 2422 2423 // This should be moved to the shared markSweep code! 2424 class PSAlwaysTrueClosure: public BoolObjectClosure { 2425 public: 2426 void do_object(oop p) { ShouldNotReachHere(); } 2427 bool do_object_b(oop p) { return true; } 2428 }; 2429 static PSAlwaysTrueClosure always_true; 2430 2431 void PSParallelCompact::adjust_roots() { 2432 // Adjust the pointers to reflect the new locations 2433 EventMark m("3 adjust roots"); 2434 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); 2435 2436 // General strong roots. 2437 Universe::oops_do(adjust_root_pointer_closure()); 2438 ReferenceProcessor::oops_do(adjust_root_pointer_closure()); 2439 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles 2440 Threads::oops_do(adjust_root_pointer_closure(), NULL); 2441 ObjectSynchronizer::oops_do(adjust_root_pointer_closure()); 2442 FlatProfiler::oops_do(adjust_root_pointer_closure()); 2443 Management::oops_do(adjust_root_pointer_closure()); 2444 JvmtiExport::oops_do(adjust_root_pointer_closure()); 2445 // SO_AllClasses 2446 SystemDictionary::oops_do(adjust_root_pointer_closure()); 2447 vmSymbols::oops_do(adjust_root_pointer_closure()); 2448 2449 // Now adjust pointers in remaining weak roots. (All of which should 2450 // have been cleared if they pointed to non-surviving objects.) 2451 // Global (weak) JNI handles 2452 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure()); 2453 2454 CodeCache::oops_do(adjust_pointer_closure()); 2455 SymbolTable::oops_do(adjust_root_pointer_closure()); 2456 StringTable::oops_do(adjust_root_pointer_closure()); 2457 ref_processor()->weak_oops_do(adjust_root_pointer_closure()); 2458 // Roots were visited so references into the young gen in roots 2459 // may have been scanned. Process them also. 2460 // Should the reference processor have a span that excludes 2461 // young gen objects? 2462 PSScavenge::reference_processor()->weak_oops_do( 2463 adjust_root_pointer_closure()); 2464 } 2465 2466 void PSParallelCompact::compact_perm(ParCompactionManager* cm) { 2467 EventMark m("4 compact perm"); 2468 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty); 2469 // trace("4"); 2470 2471 gc_heap()->perm_gen()->start_array()->reset(); 2472 move_and_update(cm, perm_space_id); 2473 } 2474 2475 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2476 uint parallel_gc_threads) 2477 { 2478 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); 2479 2480 const unsigned int task_count = MAX2(parallel_gc_threads, 1U); 2481 for (unsigned int j = 0; j < task_count; j++) { 2482 q->enqueue(new DrainStacksCompactionTask()); 2483 } 2484 2485 // Find all regions that are available (can be filled immediately) and 2486 // distribute them to the thread stacks. The iteration is done in reverse 2487 // order (high to low) so the regions will be removed in ascending order. 2488 2489 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2490 2491 size_t fillable_regions = 0; // A count for diagnostic purposes. 2492 unsigned int which = 0; // The worker thread number. 2493 2494 for (unsigned int id = to_space_id; id > perm_space_id; --id) { 2495 SpaceInfo* const space_info = _space_info + id; 2496 MutableSpace* const space = space_info->space(); 2497 HeapWord* const new_top = space_info->new_top(); 2498 2499 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2500 const size_t end_region = 2501 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2502 assert(end_region > 0, "perm gen cannot be empty"); 2503 2504 for (size_t cur = end_region - 1; cur >= beg_region; --cur) { 2505 if (sd.region(cur)->claim_unsafe()) { 2506 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2507 cm->push_region(cur); 2508 2509 if (TraceParallelOldGCCompactionPhase && Verbose) { 2510 const size_t count_mod_8 = fillable_regions & 7; 2511 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2512 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2513 if (count_mod_8 == 7) gclog_or_tty->cr(); 2514 } 2515 2516 NOT_PRODUCT(++fillable_regions;) 2517 2518 // Assign regions to threads in round-robin fashion. 2519 if (++which == task_count) { 2520 which = 0; 2521 } 2522 } 2523 } 2524 } 2525 2526 if (TraceParallelOldGCCompactionPhase) { 2527 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2528 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2529 } 2530 } 2531 2532 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2533 2534 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2535 uint parallel_gc_threads) { 2536 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); 2537 2538 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2539 2540 // Iterate over all the spaces adding tasks for updating 2541 // regions in the dense prefix. Assume that 1 gc thread 2542 // will work on opening the gaps and the remaining gc threads 2543 // will work on the dense prefix. 2544 unsigned int space_id; 2545 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2546 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2547 const MutableSpace* const space = _space_info[space_id].space(); 2548 2549 if (dense_prefix_end == space->bottom()) { 2550 // There is no dense prefix for this space. 2551 continue; 2552 } 2553 2554 // The dense prefix is before this region. 2555 size_t region_index_end_dense_prefix = 2556 sd.addr_to_region_idx(dense_prefix_end); 2557 RegionData* const dense_prefix_cp = 2558 sd.region(region_index_end_dense_prefix); 2559 assert(dense_prefix_end == space->end() || 2560 dense_prefix_cp->available() || 2561 dense_prefix_cp->claimed(), 2562 "The region after the dense prefix should always be ready to fill"); 2563 2564 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2565 2566 // Is there dense prefix work? 2567 size_t total_dense_prefix_regions = 2568 region_index_end_dense_prefix - region_index_start; 2569 // How many regions of the dense prefix should be given to 2570 // each thread? 2571 if (total_dense_prefix_regions > 0) { 2572 uint tasks_for_dense_prefix = 1; 2573 if (UseParallelDensePrefixUpdate) { 2574 if (total_dense_prefix_regions <= 2575 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2576 // Don't over partition. This assumes that 2577 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2578 // so there are not many regions to process. 2579 tasks_for_dense_prefix = parallel_gc_threads; 2580 } else { 2581 // Over partition 2582 tasks_for_dense_prefix = parallel_gc_threads * 2583 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2584 } 2585 } 2586 size_t regions_per_thread = total_dense_prefix_regions / 2587 tasks_for_dense_prefix; 2588 // Give each thread at least 1 region. 2589 if (regions_per_thread == 0) { 2590 regions_per_thread = 1; 2591 } 2592 2593 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2594 if (region_index_start >= region_index_end_dense_prefix) { 2595 break; 2596 } 2597 // region_index_end is not processed 2598 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2599 region_index_end_dense_prefix); 2600 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2601 region_index_start, 2602 region_index_end)); 2603 region_index_start = region_index_end; 2604 } 2605 } 2606 // This gets any part of the dense prefix that did not 2607 // fit evenly. 2608 if (region_index_start < region_index_end_dense_prefix) { 2609 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2610 region_index_start, 2611 region_index_end_dense_prefix)); 2612 } 2613 } 2614 } 2615 2616 void PSParallelCompact::enqueue_region_stealing_tasks( 2617 GCTaskQueue* q, 2618 ParallelTaskTerminator* terminator_ptr, 2619 uint parallel_gc_threads) { 2620 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); 2621 2622 // Once a thread has drained it's stack, it should try to steal regions from 2623 // other threads. 2624 if (parallel_gc_threads > 1) { 2625 for (uint j = 0; j < parallel_gc_threads; j++) { 2626 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2627 } 2628 } 2629 } 2630 2631 void PSParallelCompact::compact() { 2632 EventMark m("5 compact"); 2633 // trace("5"); 2634 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); 2635 2636 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2637 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2638 PSOldGen* old_gen = heap->old_gen(); 2639 old_gen->start_array()->reset(); 2640 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2641 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2642 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2643 2644 GCTaskQueue* q = GCTaskQueue::create(); 2645 enqueue_region_draining_tasks(q, parallel_gc_threads); 2646 enqueue_dense_prefix_tasks(q, parallel_gc_threads); 2647 enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads); 2648 2649 { 2650 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); 2651 2652 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2653 q->enqueue(fin); 2654 2655 gc_task_manager()->add_list(q); 2656 2657 fin->wait_for(); 2658 2659 // We have to release the barrier tasks! 2660 WaitForBarrierGCTask::destroy(fin); 2661 2662 #ifdef ASSERT 2663 // Verify that all regions have been processed before the deferred updates. 2664 // Note that perm_space_id is skipped; this type of verification is not 2665 // valid until the perm gen is compacted by regions. 2666 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2667 verify_complete(SpaceId(id)); 2668 } 2669 #endif 2670 } 2671 2672 { 2673 // Update the deferred objects, if any. Any compaction manager can be used. 2674 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); 2675 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2676 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2677 update_deferred_objects(cm, SpaceId(id)); 2678 } 2679 } 2680 } 2681 2682 #ifdef ASSERT 2683 void PSParallelCompact::verify_complete(SpaceId space_id) { 2684 // All Regions between space bottom() to new_top() should be marked as filled 2685 // and all Regions between new_top() and top() should be available (i.e., 2686 // should have been emptied). 2687 ParallelCompactData& sd = summary_data(); 2688 SpaceInfo si = _space_info[space_id]; 2689 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2690 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2691 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2692 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2693 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2694 2695 bool issued_a_warning = false; 2696 2697 size_t cur_region; 2698 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2699 const RegionData* const c = sd.region(cur_region); 2700 if (!c->completed()) { 2701 warning("region " SIZE_FORMAT " not filled: " 2702 "destination_count=" SIZE_FORMAT, 2703 cur_region, c->destination_count()); 2704 issued_a_warning = true; 2705 } 2706 } 2707 2708 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2709 const RegionData* const c = sd.region(cur_region); 2710 if (!c->available()) { 2711 warning("region " SIZE_FORMAT " not empty: " 2712 "destination_count=" SIZE_FORMAT, 2713 cur_region, c->destination_count()); 2714 issued_a_warning = true; 2715 } 2716 } 2717 2718 if (issued_a_warning) { 2719 print_region_ranges(); 2720 } 2721 } 2722 #endif // #ifdef ASSERT 2723 2724 void PSParallelCompact::compact_serial(ParCompactionManager* cm) { 2725 EventMark m("5 compact serial"); 2726 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty); 2727 2728 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2729 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2730 2731 PSYoungGen* young_gen = heap->young_gen(); 2732 PSOldGen* old_gen = heap->old_gen(); 2733 2734 old_gen->start_array()->reset(); 2735 old_gen->move_and_update(cm); 2736 young_gen->move_and_update(cm); 2737 } 2738 2739 void 2740 PSParallelCompact::follow_weak_klass_links() { 2741 // All klasses on the revisit stack are marked at this point. 2742 // Update and follow all subklass, sibling and implementor links. 2743 if (PrintRevisitStats) { 2744 gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes()); 2745 } 2746 for (uint i = 0; i < ParallelGCThreads + 1; i++) { 2747 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2748 KeepAliveClosure keep_alive_closure(cm); 2749 int length = cm->revisit_klass_stack()->length(); 2750 if (PrintRevisitStats) { 2751 gclog_or_tty->print_cr("Revisit klass stack[%d] length = %d", i, length); 2752 } 2753 for (int j = 0; j < length; j++) { 2754 cm->revisit_klass_stack()->at(j)->follow_weak_klass_links( 2755 is_alive_closure(), 2756 &keep_alive_closure); 2757 } 2758 // revisit_klass_stack is cleared in reset() 2759 cm->follow_marking_stacks(); 2760 } 2761 } 2762 2763 void 2764 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) { 2765 cm->revisit_klass_stack()->push(k); 2766 } 2767 2768 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) { 2769 cm->revisit_mdo_stack()->push(p); 2770 } 2771 2772 void PSParallelCompact::follow_mdo_weak_refs() { 2773 // All strongly reachable oops have been marked at this point; 2774 // we can visit and clear any weak references from MDO's which 2775 // we memoized during the strong marking phase. 2776 if (PrintRevisitStats) { 2777 gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes()); 2778 } 2779 for (uint i = 0; i < ParallelGCThreads + 1; i++) { 2780 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2781 GrowableArray<DataLayout*>* rms = cm->revisit_mdo_stack(); 2782 int length = rms->length(); 2783 if (PrintRevisitStats) { 2784 gclog_or_tty->print_cr("Revisit MDO stack[%d] length = %d", i, length); 2785 } 2786 for (int j = 0; j < length; j++) { 2787 rms->at(j)->follow_weak_refs(is_alive_closure()); 2788 } 2789 // revisit_mdo_stack is cleared in reset() 2790 cm->follow_marking_stacks(); 2791 } 2792 } 2793 2794 2795 #ifdef VALIDATE_MARK_SWEEP 2796 2797 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) { 2798 if (!ValidateMarkSweep) 2799 return; 2800 2801 if (!isroot) { 2802 if (_pointer_tracking) { 2803 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer"); 2804 _adjusted_pointers->remove(p); 2805 } 2806 } else { 2807 ptrdiff_t index = _root_refs_stack->find(p); 2808 if (index != -1) { 2809 int l = _root_refs_stack->length(); 2810 if (l > 0 && l - 1 != index) { 2811 void* last = _root_refs_stack->pop(); 2812 assert(last != p, "should be different"); 2813 _root_refs_stack->at_put(index, last); 2814 } else { 2815 _root_refs_stack->remove(p); 2816 } 2817 } 2818 } 2819 } 2820 2821 2822 void PSParallelCompact::check_adjust_pointer(void* p) { 2823 _adjusted_pointers->push(p); 2824 } 2825 2826 2827 class AdjusterTracker: public OopClosure { 2828 public: 2829 AdjusterTracker() {}; 2830 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); } 2831 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); } 2832 }; 2833 2834 2835 void PSParallelCompact::track_interior_pointers(oop obj) { 2836 if (ValidateMarkSweep) { 2837 _adjusted_pointers->clear(); 2838 _pointer_tracking = true; 2839 2840 AdjusterTracker checker; 2841 obj->oop_iterate(&checker); 2842 } 2843 } 2844 2845 2846 void PSParallelCompact::check_interior_pointers() { 2847 if (ValidateMarkSweep) { 2848 _pointer_tracking = false; 2849 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers"); 2850 } 2851 } 2852 2853 2854 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) { 2855 if (ValidateMarkSweep) { 2856 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops"); 2857 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0; 2858 } 2859 } 2860 2861 2862 void PSParallelCompact::register_live_oop(oop p, size_t size) { 2863 if (ValidateMarkSweep) { 2864 _live_oops->push(p); 2865 _live_oops_size->push(size); 2866 _live_oops_index++; 2867 } 2868 } 2869 2870 void PSParallelCompact::validate_live_oop(oop p, size_t size) { 2871 if (ValidateMarkSweep) { 2872 oop obj = _live_oops->at((int)_live_oops_index); 2873 guarantee(obj == p, "should be the same object"); 2874 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size"); 2875 _live_oops_index++; 2876 } 2877 } 2878 2879 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size, 2880 HeapWord* compaction_top) { 2881 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top), 2882 "should be moved to forwarded location"); 2883 if (ValidateMarkSweep) { 2884 PSParallelCompact::validate_live_oop(oop(q), size); 2885 _live_oops_moved_to->push(oop(compaction_top)); 2886 } 2887 if (RecordMarkSweepCompaction) { 2888 _cur_gc_live_oops->push(q); 2889 _cur_gc_live_oops_moved_to->push(compaction_top); 2890 _cur_gc_live_oops_size->push(size); 2891 } 2892 } 2893 2894 2895 void PSParallelCompact::compaction_complete() { 2896 if (RecordMarkSweepCompaction) { 2897 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops; 2898 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to; 2899 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size; 2900 2901 _cur_gc_live_oops = _last_gc_live_oops; 2902 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to; 2903 _cur_gc_live_oops_size = _last_gc_live_oops_size; 2904 _last_gc_live_oops = _tmp_live_oops; 2905 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to; 2906 _last_gc_live_oops_size = _tmp_live_oops_size; 2907 } 2908 } 2909 2910 2911 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) { 2912 if (!RecordMarkSweepCompaction) { 2913 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled"); 2914 return; 2915 } 2916 2917 if (_last_gc_live_oops == NULL) { 2918 tty->print_cr("No compaction information gathered yet"); 2919 return; 2920 } 2921 2922 for (int i = 0; i < _last_gc_live_oops->length(); i++) { 2923 HeapWord* old_oop = _last_gc_live_oops->at(i); 2924 size_t sz = _last_gc_live_oops_size->at(i); 2925 if (old_oop <= q && q < (old_oop + sz)) { 2926 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i); 2927 size_t offset = (q - old_oop); 2928 tty->print_cr("Address " PTR_FORMAT, q); 2929 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset); 2930 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset); 2931 return; 2932 } 2933 } 2934 2935 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q); 2936 } 2937 #endif //VALIDATE_MARK_SWEEP 2938 2939 // Update interior oops in the ranges of regions [beg_region, end_region). 2940 void 2941 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2942 SpaceId space_id, 2943 size_t beg_region, 2944 size_t end_region) { 2945 ParallelCompactData& sd = summary_data(); 2946 ParMarkBitMap* const mbm = mark_bitmap(); 2947 2948 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2949 HeapWord* const end_addr = sd.region_to_addr(end_region); 2950 assert(beg_region <= end_region, "bad region range"); 2951 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2952 2953 #ifdef ASSERT 2954 // Claim the regions to avoid triggering an assert when they are marked as 2955 // filled. 2956 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2957 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2958 } 2959 #endif // #ifdef ASSERT 2960 2961 if (beg_addr != space(space_id)->bottom()) { 2962 // Find the first live object or block of dead space that *starts* in this 2963 // range of regions. If a partial object crosses onto the region, skip it; 2964 // it will be marked for 'deferred update' when the object head is 2965 // processed. If dead space crosses onto the region, it is also skipped; it 2966 // will be filled when the prior region is processed. If neither of those 2967 // apply, the first word in the region is the start of a live object or dead 2968 // space. 2969 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2970 const RegionData* const cp = sd.region(beg_region); 2971 if (cp->partial_obj_size() != 0) { 2972 beg_addr = sd.partial_obj_end(beg_region); 2973 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2974 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2975 } 2976 } 2977 2978 if (beg_addr < end_addr) { 2979 // A live object or block of dead space starts in this range of Regions. 2980 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2981 2982 // Create closures and iterate. 2983 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2984 FillClosure fill_closure(cm, space_id); 2985 ParMarkBitMap::IterationStatus status; 2986 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2987 dense_prefix_end); 2988 if (status == ParMarkBitMap::incomplete) { 2989 update_closure.do_addr(update_closure.source()); 2990 } 2991 } 2992 2993 // Mark the regions as filled. 2994 RegionData* const beg_cp = sd.region(beg_region); 2995 RegionData* const end_cp = sd.region(end_region); 2996 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2997 cp->set_completed(); 2998 } 2999 } 3000 3001 // Return the SpaceId for the space containing addr. If addr is not in the 3002 // heap, last_space_id is returned. In debug mode it expects the address to be 3003 // in the heap and asserts such. 3004 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 3005 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 3006 3007 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 3008 if (_space_info[id].space()->contains(addr)) { 3009 return SpaceId(id); 3010 } 3011 } 3012 3013 assert(false, "no space contains the addr"); 3014 return last_space_id; 3015 } 3016 3017 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 3018 SpaceId id) { 3019 assert(id < last_space_id, "bad space id"); 3020 3021 ParallelCompactData& sd = summary_data(); 3022 const SpaceInfo* const space_info = _space_info + id; 3023 ObjectStartArray* const start_array = space_info->start_array(); 3024 3025 const MutableSpace* const space = space_info->space(); 3026 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 3027 HeapWord* const beg_addr = space_info->dense_prefix(); 3028 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 3029 3030 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 3031 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 3032 const RegionData* cur_region; 3033 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 3034 HeapWord* const addr = cur_region->deferred_obj_addr(); 3035 if (addr != NULL) { 3036 if (start_array != NULL) { 3037 start_array->allocate_block(addr); 3038 } 3039 oop(addr)->update_contents(cm); 3040 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 3041 } 3042 } 3043 } 3044 3045 // Skip over count live words starting from beg, and return the address of the 3046 // next live word. Unless marked, the word corresponding to beg is assumed to 3047 // be dead. Callers must either ensure beg does not correspond to the middle of 3048 // an object, or account for those live words in some other way. Callers must 3049 // also ensure that there are enough live words in the range [beg, end) to skip. 3050 HeapWord* 3051 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 3052 { 3053 assert(count > 0, "sanity"); 3054 3055 ParMarkBitMap* m = mark_bitmap(); 3056 idx_t bits_to_skip = m->words_to_bits(count); 3057 idx_t cur_beg = m->addr_to_bit(beg); 3058 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 3059 3060 do { 3061 cur_beg = m->find_obj_beg(cur_beg, search_end); 3062 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 3063 const size_t obj_bits = cur_end - cur_beg + 1; 3064 if (obj_bits > bits_to_skip) { 3065 return m->bit_to_addr(cur_beg + bits_to_skip); 3066 } 3067 bits_to_skip -= obj_bits; 3068 cur_beg = cur_end + 1; 3069 } while (bits_to_skip > 0); 3070 3071 // Skipping the desired number of words landed just past the end of an object. 3072 // Find the start of the next object. 3073 cur_beg = m->find_obj_beg(cur_beg, search_end); 3074 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 3075 return m->bit_to_addr(cur_beg); 3076 } 3077 3078 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 3079 SpaceId src_space_id, 3080 size_t src_region_idx) 3081 { 3082 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 3083 3084 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 3085 if (split_info.dest_region_addr() == dest_addr) { 3086 // The partial object ending at the split point contains the first word to 3087 // be copied to dest_addr. 3088 return split_info.first_src_addr(); 3089 } 3090 3091 const ParallelCompactData& sd = summary_data(); 3092 ParMarkBitMap* const bitmap = mark_bitmap(); 3093 const size_t RegionSize = ParallelCompactData::RegionSize; 3094 3095 assert(sd.is_region_aligned(dest_addr), "not aligned"); 3096 const RegionData* const src_region_ptr = sd.region(src_region_idx); 3097 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 3098 HeapWord* const src_region_destination = src_region_ptr->destination(); 3099 3100 assert(dest_addr >= src_region_destination, "wrong src region"); 3101 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 3102 3103 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 3104 HeapWord* const src_region_end = src_region_beg + RegionSize; 3105 3106 HeapWord* addr = src_region_beg; 3107 if (dest_addr == src_region_destination) { 3108 // Return the first live word in the source region. 3109 if (partial_obj_size == 0) { 3110 addr = bitmap->find_obj_beg(addr, src_region_end); 3111 assert(addr < src_region_end, "no objects start in src region"); 3112 } 3113 return addr; 3114 } 3115 3116 // Must skip some live data. 3117 size_t words_to_skip = dest_addr - src_region_destination; 3118 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 3119 3120 if (partial_obj_size >= words_to_skip) { 3121 // All the live words to skip are part of the partial object. 3122 addr += words_to_skip; 3123 if (partial_obj_size == words_to_skip) { 3124 // Find the first live word past the partial object. 3125 addr = bitmap->find_obj_beg(addr, src_region_end); 3126 assert(addr < src_region_end, "wrong src region"); 3127 } 3128 return addr; 3129 } 3130 3131 // Skip over the partial object (if any). 3132 if (partial_obj_size != 0) { 3133 words_to_skip -= partial_obj_size; 3134 addr += partial_obj_size; 3135 } 3136 3137 // Skip over live words due to objects that start in the region. 3138 addr = skip_live_words(addr, src_region_end, words_to_skip); 3139 assert(addr < src_region_end, "wrong src region"); 3140 return addr; 3141 } 3142 3143 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 3144 SpaceId src_space_id, 3145 size_t beg_region, 3146 HeapWord* end_addr) 3147 { 3148 ParallelCompactData& sd = summary_data(); 3149 3150 #ifdef ASSERT 3151 MutableSpace* const src_space = _space_info[src_space_id].space(); 3152 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 3153 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 3154 "src_space_id does not match beg_addr"); 3155 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 3156 "src_space_id does not match end_addr"); 3157 #endif // #ifdef ASSERT 3158 3159 RegionData* const beg = sd.region(beg_region); 3160 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 3161 3162 // Regions up to new_top() are enqueued if they become available. 3163 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3164 RegionData* const enqueue_end = 3165 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3166 3167 for (RegionData* cur = beg; cur < end; ++cur) { 3168 assert(cur->data_size() > 0, "region must have live data"); 3169 cur->decrement_destination_count(); 3170 if (cur < enqueue_end && cur->available() && cur->claim()) { 3171 cm->push_region(sd.region(cur)); 3172 } 3173 } 3174 } 3175 3176 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3177 SpaceId& src_space_id, 3178 HeapWord*& src_space_top, 3179 HeapWord* end_addr) 3180 { 3181 typedef ParallelCompactData::RegionData RegionData; 3182 3183 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3184 const size_t region_size = ParallelCompactData::RegionSize; 3185 3186 size_t src_region_idx = 0; 3187 3188 // Skip empty regions (if any) up to the top of the space. 3189 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3190 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3191 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3192 const RegionData* const top_region_ptr = 3193 sd.addr_to_region_ptr(top_aligned_up); 3194 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3195 ++src_region_ptr; 3196 } 3197 3198 if (src_region_ptr < top_region_ptr) { 3199 // The next source region is in the current space. Update src_region_idx 3200 // and the source address to match src_region_ptr. 3201 src_region_idx = sd.region(src_region_ptr); 3202 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3203 if (src_region_addr > closure.source()) { 3204 closure.set_source(src_region_addr); 3205 } 3206 return src_region_idx; 3207 } 3208 3209 // Switch to a new source space and find the first non-empty region. 3210 unsigned int space_id = src_space_id + 1; 3211 assert(space_id < last_space_id, "not enough spaces"); 3212 3213 HeapWord* const destination = closure.destination(); 3214 3215 do { 3216 MutableSpace* space = _space_info[space_id].space(); 3217 HeapWord* const bottom = space->bottom(); 3218 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3219 3220 // Iterate over the spaces that do not compact into themselves. 3221 if (bottom_cp->destination() != bottom) { 3222 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3223 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3224 3225 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3226 if (src_cp->live_obj_size() > 0) { 3227 // Found it. 3228 assert(src_cp->destination() == destination, 3229 "first live obj in the space must match the destination"); 3230 assert(src_cp->partial_obj_size() == 0, 3231 "a space cannot begin with a partial obj"); 3232 3233 src_space_id = SpaceId(space_id); 3234 src_space_top = space->top(); 3235 const size_t src_region_idx = sd.region(src_cp); 3236 closure.set_source(sd.region_to_addr(src_region_idx)); 3237 return src_region_idx; 3238 } else { 3239 assert(src_cp->data_size() == 0, "sanity"); 3240 } 3241 } 3242 } 3243 } while (++space_id < last_space_id); 3244 3245 assert(false, "no source region was found"); 3246 return 0; 3247 } 3248 3249 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3250 { 3251 typedef ParMarkBitMap::IterationStatus IterationStatus; 3252 const size_t RegionSize = ParallelCompactData::RegionSize; 3253 ParMarkBitMap* const bitmap = mark_bitmap(); 3254 ParallelCompactData& sd = summary_data(); 3255 RegionData* const region_ptr = sd.region(region_idx); 3256 3257 // Get the items needed to construct the closure. 3258 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3259 SpaceId dest_space_id = space_id(dest_addr); 3260 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3261 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3262 assert(dest_addr < new_top, "sanity"); 3263 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3264 3265 // Get the source region and related info. 3266 size_t src_region_idx = region_ptr->source_region(); 3267 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3268 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3269 3270 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3271 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3272 3273 // Adjust src_region_idx to prepare for decrementing destination counts (the 3274 // destination count is not decremented when a region is copied to itself). 3275 if (src_region_idx == region_idx) { 3276 src_region_idx += 1; 3277 } 3278 3279 if (bitmap->is_unmarked(closure.source())) { 3280 // The first source word is in the middle of an object; copy the remainder 3281 // of the object or as much as will fit. The fact that pointer updates were 3282 // deferred will be noted when the object header is processed. 3283 HeapWord* const old_src_addr = closure.source(); 3284 closure.copy_partial_obj(); 3285 if (closure.is_full()) { 3286 decrement_destination_counts(cm, src_space_id, src_region_idx, 3287 closure.source()); 3288 region_ptr->set_deferred_obj_addr(NULL); 3289 region_ptr->set_completed(); 3290 return; 3291 } 3292 3293 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3294 if (sd.region_align_down(old_src_addr) != end_addr) { 3295 // The partial object was copied from more than one source region. 3296 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3297 3298 // Move to the next source region, possibly switching spaces as well. All 3299 // args except end_addr may be modified. 3300 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3301 end_addr); 3302 } 3303 } 3304 3305 do { 3306 HeapWord* const cur_addr = closure.source(); 3307 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3308 src_space_top); 3309 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3310 3311 if (status == ParMarkBitMap::incomplete) { 3312 // The last obj that starts in the source region does not end in the 3313 // region. 3314 assert(closure.source() < end_addr, "sanity"); 3315 HeapWord* const obj_beg = closure.source(); 3316 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3317 src_space_top); 3318 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3319 if (obj_end < range_end) { 3320 // The end was found; the entire object will fit. 3321 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3322 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3323 } else { 3324 // The end was not found; the object will not fit. 3325 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3326 status = ParMarkBitMap::would_overflow; 3327 } 3328 } 3329 3330 if (status == ParMarkBitMap::would_overflow) { 3331 // The last object did not fit. Note that interior oop updates were 3332 // deferred, then copy enough of the object to fill the region. 3333 region_ptr->set_deferred_obj_addr(closure.destination()); 3334 status = closure.copy_until_full(); // copies from closure.source() 3335 3336 decrement_destination_counts(cm, src_space_id, src_region_idx, 3337 closure.source()); 3338 region_ptr->set_completed(); 3339 return; 3340 } 3341 3342 if (status == ParMarkBitMap::full) { 3343 decrement_destination_counts(cm, src_space_id, src_region_idx, 3344 closure.source()); 3345 region_ptr->set_deferred_obj_addr(NULL); 3346 region_ptr->set_completed(); 3347 return; 3348 } 3349 3350 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3351 3352 // Move to the next source region, possibly switching spaces as well. All 3353 // args except end_addr may be modified. 3354 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3355 end_addr); 3356 } while (true); 3357 } 3358 3359 void 3360 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3361 const MutableSpace* sp = space(space_id); 3362 if (sp->is_empty()) { 3363 return; 3364 } 3365 3366 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3367 ParMarkBitMap* const bitmap = mark_bitmap(); 3368 HeapWord* const dp_addr = dense_prefix(space_id); 3369 HeapWord* beg_addr = sp->bottom(); 3370 HeapWord* end_addr = sp->top(); 3371 3372 #ifdef ASSERT 3373 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3374 if (cm->should_verify_only()) { 3375 VerifyUpdateClosure verify_update(cm, sp); 3376 bitmap->iterate(&verify_update, beg_addr, end_addr); 3377 return; 3378 } 3379 3380 if (cm->should_reset_only()) { 3381 ResetObjectsClosure reset_objects(cm); 3382 bitmap->iterate(&reset_objects, beg_addr, end_addr); 3383 return; 3384 } 3385 #endif 3386 3387 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3388 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3389 if (beg_region < dp_region) { 3390 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3391 } 3392 3393 // The destination of the first live object that starts in the region is one 3394 // past the end of the partial object entering the region (if any). 3395 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3396 HeapWord* const new_top = _space_info[space_id].new_top(); 3397 assert(new_top >= dest_addr, "bad new_top value"); 3398 const size_t words = pointer_delta(new_top, dest_addr); 3399 3400 if (words > 0) { 3401 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3402 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3403 3404 ParMarkBitMap::IterationStatus status; 3405 status = bitmap->iterate(&closure, dest_addr, end_addr); 3406 assert(status == ParMarkBitMap::full, "iteration not complete"); 3407 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3408 "live objects skipped because closure is full"); 3409 } 3410 } 3411 3412 jlong PSParallelCompact::millis_since_last_gc() { 3413 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc; 3414 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3415 if (ret_val < 0) { 3416 NOT_PRODUCT(warning("time warp: %d", ret_val);) 3417 return 0; 3418 } 3419 return ret_val; 3420 } 3421 3422 void PSParallelCompact::reset_millis_since_last_gc() { 3423 _time_of_last_gc = os::javaTimeMillis(); 3424 } 3425 3426 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3427 { 3428 if (source() != destination()) { 3429 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3430 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3431 } 3432 update_state(words_remaining()); 3433 assert(is_full(), "sanity"); 3434 return ParMarkBitMap::full; 3435 } 3436 3437 void MoveAndUpdateClosure::copy_partial_obj() 3438 { 3439 size_t words = words_remaining(); 3440 3441 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3442 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3443 if (end_addr < range_end) { 3444 words = bitmap()->obj_size(source(), end_addr); 3445 } 3446 3447 // This test is necessary; if omitted, the pointer updates to a partial object 3448 // that crosses the dense prefix boundary could be overwritten. 3449 if (source() != destination()) { 3450 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3451 Copy::aligned_conjoint_words(source(), destination(), words); 3452 } 3453 update_state(words); 3454 } 3455 3456 ParMarkBitMapClosure::IterationStatus 3457 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3458 assert(destination() != NULL, "sanity"); 3459 assert(bitmap()->obj_size(addr) == words, "bad size"); 3460 3461 _source = addr; 3462 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3463 destination(), "wrong destination"); 3464 3465 if (words > words_remaining()) { 3466 return ParMarkBitMap::would_overflow; 3467 } 3468 3469 // The start_array must be updated even if the object is not moving. 3470 if (_start_array != NULL) { 3471 _start_array->allocate_block(destination()); 3472 } 3473 3474 if (destination() != source()) { 3475 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3476 Copy::aligned_conjoint_words(source(), destination(), words); 3477 } 3478 3479 oop moved_oop = (oop) destination(); 3480 moved_oop->update_contents(compaction_manager()); 3481 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3482 3483 update_state(words); 3484 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3485 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3486 } 3487 3488 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3489 ParCompactionManager* cm, 3490 PSParallelCompact::SpaceId space_id) : 3491 ParMarkBitMapClosure(mbm, cm), 3492 _space_id(space_id), 3493 _start_array(PSParallelCompact::start_array(space_id)) 3494 { 3495 } 3496 3497 // Updates the references in the object to their new values. 3498 ParMarkBitMapClosure::IterationStatus 3499 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3500 do_addr(addr); 3501 return ParMarkBitMap::incomplete; 3502 } 3503 3504 // Verify the new location using the forwarding pointer 3505 // from MarkSweep::mark_sweep_phase2(). Set the mark_word 3506 // to the initial value. 3507 ParMarkBitMapClosure::IterationStatus 3508 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3509 // The second arg (words) is not used. 3510 oop obj = (oop) addr; 3511 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer(); 3512 HeapWord* new_pointer = summary_data().calc_new_pointer(obj); 3513 if (forwarding_ptr == NULL) { 3514 // The object is dead or not moving. 3515 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj), 3516 "Object liveness is wrong."); 3517 return ParMarkBitMap::incomplete; 3518 } 3519 assert(UseParallelOldGCDensePrefix || 3520 (HeapMaximumCompactionInterval > 1) || 3521 (MarkSweepAlwaysCompactCount > 1) || 3522 (forwarding_ptr == new_pointer), 3523 "Calculation of new location is incorrect"); 3524 return ParMarkBitMap::incomplete; 3525 } 3526 3527 // Reset objects modified for debug checking. 3528 ParMarkBitMapClosure::IterationStatus 3529 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) { 3530 // The second arg (words) is not used. 3531 oop obj = (oop) addr; 3532 obj->init_mark(); 3533 return ParMarkBitMap::incomplete; 3534 } 3535 3536 // Prepare for compaction. This method is executed once 3537 // (i.e., by a single thread) before compaction. 3538 // Save the updated location of the intArrayKlassObj for 3539 // filling holes in the dense prefix. 3540 void PSParallelCompact::compact_prologue() { 3541 _updated_int_array_klass_obj = (klassOop) 3542 summary_data().calc_new_pointer(Universe::intArrayKlassObj()); 3543 } 3544