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