1 #ifdef USE_PRAGMA_IDENT_SRC 2 #pragma ident "@(#)psParallelCompact.cpp 1.64 08/06/19 15:33:02 JVM" 3 #endif 4 /* 5 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved. 6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 7 * 8 * This code is free software; you can redistribute it and/or modify it 9 * under the terms of the GNU General Public License version 2 only, as 10 * published by the Free Software Foundation. 11 * 12 * This code is distributed in the hope that it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 * version 2 for more details (a copy is included in the LICENSE file that 16 * accompanied this code). 17 * 18 * You should have received a copy of the GNU General Public License version 19 * 2 along with this work; if not, write to the Free Software Foundation, 20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 21 * 22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 23 * CA 95054 USA or visit www.sun.com if you need additional information or 24 * have any questions. 25 * 26 */ 27 28 #include "incls/_precompiled.incl" 29 #include "incls/_psParallelCompact.cpp.incl" 30 31 #include <math.h> 32 33 // All sizes are in HeapWords. 34 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words 35 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 36 const size_t ParallelCompactData::RegionSizeBytes = 37 RegionSize << LogHeapWordSize; 38 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 39 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 40 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 41 42 const ParallelCompactData::RegionData::region_sz_t 43 ParallelCompactData::RegionData::dc_shift = 27; 44 45 const ParallelCompactData::RegionData::region_sz_t 46 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 47 48 const ParallelCompactData::RegionData::region_sz_t 49 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 50 51 const ParallelCompactData::RegionData::region_sz_t 52 ParallelCompactData::RegionData::los_mask = ~dc_mask; 53 54 const ParallelCompactData::RegionData::region_sz_t 55 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 56 57 const ParallelCompactData::RegionData::region_sz_t 58 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 59 60 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 61 bool PSParallelCompact::_print_phases = false; 62 63 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 64 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL; 65 66 double PSParallelCompact::_dwl_mean; 67 double PSParallelCompact::_dwl_std_dev; 68 double PSParallelCompact::_dwl_first_term; 69 double PSParallelCompact::_dwl_adjustment; 70 #ifdef ASSERT 71 bool PSParallelCompact::_dwl_initialized = false; 72 #endif // #ifdef ASSERT 73 74 #ifdef VALIDATE_MARK_SWEEP 75 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL; 76 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL; 77 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL; 78 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL; 79 size_t PSParallelCompact::_live_oops_index = 0; 80 size_t PSParallelCompact::_live_oops_index_at_perm = 0; 81 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL; 82 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL; 83 bool PSParallelCompact::_pointer_tracking = false; 84 bool PSParallelCompact::_root_tracking = true; 85 86 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL; 87 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL; 88 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL; 89 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL; 90 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL; 91 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL; 92 #endif 93 94 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 95 HeapWord* destination) 96 { 97 assert(src_region_idx != 0, "invalid src_region_idx"); 98 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 99 assert(destination != NULL, "invalid destination argument"); 100 101 _src_region_idx = src_region_idx; 102 _partial_obj_size = partial_obj_size; 103 _destination = destination; 104 105 // These fields may not be updated below, so make sure they're clear. 106 assert(_dest_region_addr == NULL, "should have been cleared"); 107 assert(_first_src_addr == NULL, "should have been cleared"); 108 109 // Determine the number of destination regions for the partial object. 110 HeapWord* const last_word = destination + partial_obj_size - 1; 111 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 112 HeapWord* const beg_region_addr = sd.region_align_down(destination); 113 HeapWord* const end_region_addr = sd.region_align_down(last_word); 114 115 if (beg_region_addr == end_region_addr) { 116 // One destination region. 117 _destination_count = 1; 118 if (end_region_addr == destination) { 119 // The destination falls on a region boundary, thus the first word of the 120 // partial object will be the first word copied to the destination region. 121 _dest_region_addr = end_region_addr; 122 _first_src_addr = sd.region_to_addr(src_region_idx); 123 } 124 } else { 125 // Two destination regions. When copied, the partial object will cross a 126 // destination region boundary, so a word somewhere within the partial 127 // object will be the first word copied to the second destination region. 128 _destination_count = 2; 129 _dest_region_addr = end_region_addr; 130 const size_t ofs = pointer_delta(end_region_addr, destination); 131 assert(ofs < _partial_obj_size, "sanity"); 132 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 133 } 134 } 135 136 void SplitInfo::clear() 137 { 138 _src_region_idx = 0; 139 _partial_obj_size = 0; 140 _destination = NULL; 141 _destination_count = 0; 142 _dest_region_addr = NULL; 143 _first_src_addr = NULL; 144 assert(!is_valid(), "sanity"); 145 } 146 147 #ifdef ASSERT 148 void SplitInfo::verify_clear() 149 { 150 assert(_src_region_idx == 0, "not clear"); 151 assert(_partial_obj_size == 0, "not clear"); 152 assert(_destination == NULL, "not clear"); 153 assert(_destination_count == 0, "not clear"); 154 assert(_dest_region_addr == NULL, "not clear"); 155 assert(_first_src_addr == NULL, "not clear"); 156 } 157 #endif // #ifdef ASSERT 158 159 160 #ifndef PRODUCT 161 const char* PSParallelCompact::space_names[] = { 162 "perm", "old ", "eden", "from", "to " 163 }; 164 165 void PSParallelCompact::print_region_ranges() 166 { 167 tty->print_cr("space bottom top end new_top"); 168 tty->print_cr("------ ---------- ---------- ---------- ----------"); 169 170 for (unsigned int id = 0; id < last_space_id; ++id) { 171 const MutableSpace* space = _space_info[id].space(); 172 tty->print_cr("%u %s " 173 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 174 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 175 id, space_names[id], 176 summary_data().addr_to_region_idx(space->bottom()), 177 summary_data().addr_to_region_idx(space->top()), 178 summary_data().addr_to_region_idx(space->end()), 179 summary_data().addr_to_region_idx(_space_info[id].new_top())); 180 } 181 } 182 183 void 184 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 185 { 186 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 187 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 188 189 ParallelCompactData& sd = PSParallelCompact::summary_data(); 190 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 191 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 192 REGION_IDX_FORMAT " " PTR_FORMAT " " 193 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 194 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 195 i, c->data_location(), dci, c->destination(), 196 c->partial_obj_size(), c->live_obj_size(), 197 c->data_size(), c->source_region(), c->destination_count()); 198 199 #undef REGION_IDX_FORMAT 200 #undef REGION_DATA_FORMAT 201 } 202 203 void 204 print_generic_summary_data(ParallelCompactData& summary_data, 205 HeapWord* const beg_addr, 206 HeapWord* const end_addr) 207 { 208 size_t total_words = 0; 209 size_t i = summary_data.addr_to_region_idx(beg_addr); 210 const size_t last = summary_data.addr_to_region_idx(end_addr); 211 HeapWord* pdest = 0; 212 213 while (i <= last) { 214 ParallelCompactData::RegionData* c = summary_data.region(i); 215 if (c->data_size() != 0 || c->destination() != pdest) { 216 print_generic_summary_region(i, c); 217 total_words += c->data_size(); 218 pdest = c->destination(); 219 } 220 ++i; 221 } 222 223 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 224 } 225 226 void 227 print_generic_summary_data(ParallelCompactData& summary_data, 228 SpaceInfo* space_info) 229 { 230 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 231 const MutableSpace* space = space_info[id].space(); 232 print_generic_summary_data(summary_data, space->bottom(), 233 MAX2(space->top(), space_info[id].new_top())); 234 } 235 } 236 237 void 238 print_initial_summary_region(size_t i, 239 const ParallelCompactData::RegionData* c, 240 bool newline = true) 241 { 242 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 243 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 244 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 245 i, c->destination(), 246 c->partial_obj_size(), c->live_obj_size(), 247 c->data_size(), c->source_region(), c->destination_count()); 248 if (newline) tty->cr(); 249 } 250 251 void 252 print_initial_summary_data(ParallelCompactData& summary_data, 253 const MutableSpace* space) { 254 if (space->top() == space->bottom()) { 255 return; 256 } 257 258 const size_t region_size = ParallelCompactData::RegionSize; 259 typedef ParallelCompactData::RegionData RegionData; 260 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 261 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 262 const RegionData* c = summary_data.region(end_region - 1); 263 HeapWord* end_addr = c->destination() + c->data_size(); 264 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 265 266 // Print (and count) the full regions at the beginning of the space. 267 size_t full_region_count = 0; 268 size_t i = summary_data.addr_to_region_idx(space->bottom()); 269 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 270 print_initial_summary_region(i, summary_data.region(i)); 271 ++full_region_count; 272 ++i; 273 } 274 275 size_t live_to_right = live_in_space - full_region_count * region_size; 276 277 double max_reclaimed_ratio = 0.0; 278 size_t max_reclaimed_ratio_region = 0; 279 size_t max_dead_to_right = 0; 280 size_t max_live_to_right = 0; 281 282 // Print the 'reclaimed ratio' for regions while there is something live in 283 // the region or to the right of it. The remaining regions are empty (and 284 // uninteresting), and computing the ratio will result in division by 0. 285 while (i < end_region && live_to_right > 0) { 286 c = summary_data.region(i); 287 HeapWord* const region_addr = summary_data.region_to_addr(i); 288 const size_t used_to_right = pointer_delta(space->top(), region_addr); 289 const size_t dead_to_right = used_to_right - live_to_right; 290 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 291 292 if (reclaimed_ratio > max_reclaimed_ratio) { 293 max_reclaimed_ratio = reclaimed_ratio; 294 max_reclaimed_ratio_region = i; 295 max_dead_to_right = dead_to_right; 296 max_live_to_right = live_to_right; 297 } 298 299 print_initial_summary_region(i, c, false); 300 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 301 reclaimed_ratio, dead_to_right, live_to_right); 302 303 live_to_right -= c->data_size(); 304 ++i; 305 } 306 307 // Any remaining regions are empty. Print one more if there is one. 308 if (i < end_region) { 309 print_initial_summary_region(i, summary_data.region(i)); 310 } 311 312 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 313 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 314 max_reclaimed_ratio_region, max_dead_to_right, 315 max_live_to_right, max_reclaimed_ratio); 316 } 317 318 void 319 print_initial_summary_data(ParallelCompactData& summary_data, 320 SpaceInfo* space_info) { 321 unsigned int id = PSParallelCompact::perm_space_id; 322 const MutableSpace* space; 323 do { 324 space = space_info[id].space(); 325 print_initial_summary_data(summary_data, space); 326 } while (++id < PSParallelCompact::eden_space_id); 327 328 do { 329 space = space_info[id].space(); 330 print_generic_summary_data(summary_data, space->bottom(), space->top()); 331 } while (++id < PSParallelCompact::last_space_id); 332 } 333 #endif // #ifndef PRODUCT 334 335 #ifdef ASSERT 336 size_t add_obj_count; 337 size_t add_obj_size; 338 size_t mark_bitmap_count; 339 size_t mark_bitmap_size; 340 #endif // #ifdef ASSERT 341 342 ParallelCompactData::ParallelCompactData() 343 { 344 _region_start = 0; 345 346 _region_vspace = 0; 347 _region_data = 0; 348 _region_count = 0; 349 } 350 351 bool ParallelCompactData::initialize(MemRegion covered_region) 352 { 353 _region_start = covered_region.start(); 354 const size_t region_size = covered_region.word_size(); 355 DEBUG_ONLY(_region_end = _region_start + region_size;) 356 357 assert(region_align_down(_region_start) == _region_start, 358 "region start not aligned"); 359 assert((region_size & RegionSizeOffsetMask) == 0, 360 "region size not a multiple of RegionSize"); 361 362 bool result = initialize_region_data(region_size); 363 364 return result; 365 } 366 367 PSVirtualSpace* 368 ParallelCompactData::create_vspace(size_t count, size_t element_size) 369 { 370 const size_t raw_bytes = count * element_size; 371 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); 372 const size_t granularity = os::vm_allocation_granularity(); 373 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 374 375 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 376 MAX2(page_sz, granularity); 377 ReservedSpace rs(bytes, rs_align, rs_align > 0); 378 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 379 rs.size()); 380 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 381 if (vspace != 0) { 382 if (vspace->expand_by(bytes)) { 383 return vspace; 384 } 385 delete vspace; 386 // Release memory reserved in the space. 387 rs.release(); 388 } 389 390 return 0; 391 } 392 393 bool ParallelCompactData::initialize_region_data(size_t region_size) 394 { 395 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 396 _region_vspace = create_vspace(count, sizeof(RegionData)); 397 if (_region_vspace != 0) { 398 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 399 _region_count = count; 400 return true; 401 } 402 return false; 403 } 404 405 void ParallelCompactData::clear() 406 { 407 memset(_region_data, 0, _region_vspace->committed_size()); 408 } 409 410 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 411 assert(beg_region <= _region_count, "beg_region out of range"); 412 assert(end_region <= _region_count, "end_region out of range"); 413 414 const size_t region_cnt = end_region - beg_region; 415 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 416 } 417 418 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 419 { 420 const RegionData* cur_cp = region(region_idx); 421 const RegionData* const end_cp = region(region_count() - 1); 422 423 HeapWord* result = region_to_addr(region_idx); 424 if (cur_cp < end_cp) { 425 do { 426 result += cur_cp->partial_obj_size(); 427 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 428 } 429 return result; 430 } 431 432 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 433 { 434 const size_t obj_ofs = pointer_delta(addr, _region_start); 435 const size_t beg_region = obj_ofs >> Log2RegionSize; 436 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 437 438 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 439 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 440 441 if (beg_region == end_region) { 442 // All in one region. 443 _region_data[beg_region].add_live_obj(len); 444 return; 445 } 446 447 // First region. 448 const size_t beg_ofs = region_offset(addr); 449 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 450 451 klassOop klass = ((oop)addr)->klass(); 452 // Middle regions--completely spanned by this object. 453 for (size_t region = beg_region + 1; region < end_region; ++region) { 454 _region_data[region].set_partial_obj_size(RegionSize); 455 _region_data[region].set_partial_obj_addr(addr); 456 } 457 458 // Last region. 459 const size_t end_ofs = region_offset(addr + len - 1); 460 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 461 _region_data[end_region].set_partial_obj_addr(addr); 462 } 463 464 void 465 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 466 { 467 assert(region_offset(beg) == 0, "not RegionSize aligned"); 468 assert(region_offset(end) == 0, "not RegionSize aligned"); 469 470 size_t cur_region = addr_to_region_idx(beg); 471 const size_t end_region = addr_to_region_idx(end); 472 HeapWord* addr = beg; 473 while (cur_region < end_region) { 474 _region_data[cur_region].set_destination(addr); 475 _region_data[cur_region].set_destination_count(0); 476 _region_data[cur_region].set_source_region(cur_region); 477 _region_data[cur_region].set_data_location(addr); 478 479 // Update live_obj_size so the region appears completely full. 480 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 481 _region_data[cur_region].set_live_obj_size(live_size); 482 483 ++cur_region; 484 addr += RegionSize; 485 } 486 } 487 488 // Find the point at which a space can be split and, if necessary, record the 489 // split point. 490 // 491 // If the current src region (which overflowed the destination space) doesn't 492 // have a partial object, the split point is at the beginning of the current src 493 // region (an "easy" split, no extra bookkeeping required). 494 // 495 // If the current src region has a partial object, the split point is in the 496 // region where that partial object starts (call it the split_region). If 497 // split_region has a partial object, then the split point is just after that 498 // partial object (a "hard" split where we have to record the split data and 499 // zero the partial_obj_size field). With a "hard" split, we know that the 500 // partial_obj ends within split_region because the partial object that caused 501 // the overflow starts in split_region. If split_region doesn't have a partial 502 // obj, then the split is at the beginning of split_region (another "easy" 503 // split). 504 HeapWord* 505 ParallelCompactData::summarize_split_space(size_t src_region, 506 SplitInfo& split_info, 507 HeapWord* destination, 508 HeapWord* target_end, 509 HeapWord** target_next) 510 { 511 assert(destination <= target_end, "sanity"); 512 assert(destination + _region_data[src_region].data_size() > target_end, 513 "region should not fit into target space"); 514 515 size_t split_region = src_region; 516 HeapWord* split_destination = destination; 517 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 518 519 if (destination + partial_obj_size > target_end) { 520 // The split point is just after the partial object (if any) in the 521 // src_region that contains the start of the object that overflowed the 522 // destination space. 523 // 524 // Find the start of the "overflow" object and set split_region to the 525 // region containing it. 526 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 527 split_region = addr_to_region_idx(overflow_obj); 528 529 // Clear the source_region field of all destination regions whose first word 530 // came from data after the split point (a non-null source_region field 531 // implies a region must be filled). 532 // 533 // An alternative to the simple loop below: clear during post_compact(), 534 // which uses memcpy instead of individual stores, and is easy to 535 // parallelize. (The downside is that it clears the entire RegionData 536 // object as opposed to just one field.) 537 // 538 // post_compact() would have to clear the summary data up to the highest 539 // address that was written during the summary phase, which would be 540 // 541 // max(top, max(new_top, clear_top)) 542 // 543 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 544 // to destination + partial_obj_size, where both have the values passed to 545 // this routine. 546 const RegionData* const sr = region(split_region); 547 const size_t beg_idx = 548 addr_to_region_idx(region_align_up(sr->destination() + 549 sr->partial_obj_size())); 550 const size_t end_idx = 551 addr_to_region_idx(region_align_up(destination + partial_obj_size)); 552 553 if (TraceParallelOldGCSummaryPhase) { 554 gclog_or_tty->print_cr("split: clearing source_region field in [" 555 SIZE_FORMAT ", " SIZE_FORMAT ")", 556 beg_idx, end_idx); 557 } 558 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 559 _region_data[idx].set_source_region(0); 560 } 561 562 // Set split_destination and partial_obj_size to reflect the split region. 563 split_destination = sr->destination(); 564 partial_obj_size = sr->partial_obj_size(); 565 } 566 567 // The split is recorded only if a partial object extends onto the region. 568 if (partial_obj_size != 0) { 569 _region_data[split_region].set_partial_obj_size(0); 570 split_info.record(split_region, partial_obj_size, split_destination); 571 } 572 573 // Setup the continuation addresses. 574 *target_next = split_destination + partial_obj_size; 575 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 576 577 if (TraceParallelOldGCSummaryPhase) { 578 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 579 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 580 " pos=" SIZE_FORMAT, 581 split_type, source_next, split_region, 582 partial_obj_size); 583 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 584 " tn=" PTR_FORMAT, 585 split_type, split_destination, 586 addr_to_region_idx(split_destination), 587 *target_next); 588 589 if (partial_obj_size != 0) { 590 HeapWord* const po_beg = split_info.destination(); 591 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 592 gclog_or_tty->print_cr("%s split: " 593 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 594 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 595 split_type, 596 po_beg, addr_to_region_idx(po_beg), 597 po_end, addr_to_region_idx(po_end)); 598 } 599 } 600 601 return source_next; 602 } 603 604 bool ParallelCompactData::summarize(SplitInfo& split_info, 605 HeapWord* source_beg, HeapWord* source_end, 606 HeapWord** source_next, 607 HeapWord* target_beg, HeapWord* target_end, 608 HeapWord** target_next) 609 { 610 if (TraceParallelOldGCSummaryPhase) { 611 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 612 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 613 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 614 source_beg, source_end, source_next_val, 615 target_beg, target_end, *target_next); 616 } 617 618 size_t cur_region = addr_to_region_idx(source_beg); 619 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 620 621 HeapWord *dest_addr = target_beg; 622 while (cur_region < end_region) { 623 // The destination must be set even if the region has no data. 624 _region_data[cur_region].set_destination(dest_addr); 625 626 size_t words = _region_data[cur_region].data_size(); 627 if (words > 0) { 628 // If cur_region does not fit entirely into the target space, find a point 629 // at which the source space can be 'split' so that part is copied to the 630 // target space and the rest is copied elsewhere. 631 if (dest_addr + words > target_end) { 632 assert(source_next != NULL, "source_next is NULL when splitting"); 633 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 634 target_end, target_next); 635 return false; 636 } 637 638 // Compute the destination_count for cur_region, and if necessary, update 639 // source_region for a destination region. The source_region field is 640 // updated if cur_region is the first (left-most) region to be copied to a 641 // destination region. 642 // 643 // The destination_count calculation is a bit subtle. A region that has 644 // data that compacts into itself does not count itself as a destination. 645 // This maintains the invariant that a zero count means the region is 646 // available and can be claimed and then filled. 647 uint destination_count = 0; 648 if (split_info.is_split(cur_region)) { 649 // The current region has been split: the partial object will be copied 650 // to one destination space and the remaining data will be copied to 651 // another destination space. Adjust the initial destination_count and, 652 // if necessary, set the source_region field if the partial object will 653 // cross a destination region boundary. 654 destination_count = split_info.destination_count(); 655 if (destination_count == 2) { 656 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 657 _region_data[dest_idx].set_source_region(cur_region); 658 } 659 } 660 661 HeapWord* const last_addr = dest_addr + words - 1; 662 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 663 const size_t dest_region_2 = addr_to_region_idx(last_addr); 664 665 // Initially assume that the destination regions will be the same and 666 // adjust the value below if necessary. Under this assumption, if 667 // cur_region == dest_region_2, then cur_region will be compacted 668 // completely into itself. 669 destination_count += cur_region == dest_region_2 ? 0 : 1; 670 if (dest_region_1 != dest_region_2) { 671 // Destination regions differ; adjust destination_count. 672 destination_count += 1; 673 // Data from cur_region will be copied to the start of dest_region_2. 674 _region_data[dest_region_2].set_source_region(cur_region); 675 } else if (region_offset(dest_addr) == 0) { 676 // Data from cur_region will be copied to the start of the destination 677 // region. 678 _region_data[dest_region_1].set_source_region(cur_region); 679 } 680 681 _region_data[cur_region].set_destination_count(destination_count); 682 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 683 dest_addr += words; 684 } 685 686 ++cur_region; 687 } 688 689 *target_next = dest_addr; 690 return true; 691 } 692 693 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 694 assert(addr != NULL, "Should detect NULL oop earlier"); 695 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); 696 #ifdef ASSERT 697 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { 698 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); 699 } 700 #endif 701 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); 702 703 // Region covering the object. 704 size_t region_index = addr_to_region_idx(addr); 705 const RegionData* const region_ptr = region(region_index); 706 HeapWord* const region_addr = region_align_down(addr); 707 708 assert(addr < region_addr + RegionSize, "Region does not cover object"); 709 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check"); 710 711 HeapWord* result = region_ptr->destination(); 712 713 // If all the data in the region is live, then the new location of the object 714 // can be calculated from the destination of the region plus the offset of the 715 // object in the region. 716 if (region_ptr->data_size() == RegionSize) { 717 result += pointer_delta(addr, region_addr); 718 return result; 719 } 720 721 // The new location of the object is 722 // region destination + 723 // size of the partial object extending onto the region + 724 // sizes of the live objects in the Region that are to the left of addr 725 const size_t partial_obj_size = region_ptr->partial_obj_size(); 726 HeapWord* const search_start = region_addr + partial_obj_size; 727 728 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 729 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 730 731 result += partial_obj_size + live_to_left; 732 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 733 return result; 734 } 735 736 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) { 737 klassOop updated_klass; 738 if (PSParallelCompact::should_update_klass(old_klass)) { 739 updated_klass = (klassOop) calc_new_pointer(old_klass); 740 } else { 741 updated_klass = old_klass; 742 } 743 744 return updated_klass; 745 } 746 747 #ifdef ASSERT 748 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 749 { 750 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 751 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 752 for (const size_t* p = beg; p < end; ++p) { 753 assert(*p == 0, "not zero"); 754 } 755 } 756 757 void ParallelCompactData::verify_clear() 758 { 759 verify_clear(_region_vspace); 760 } 761 #endif // #ifdef ASSERT 762 763 #ifdef NOT_PRODUCT 764 ParallelCompactData::RegionData* debug_region(size_t region_index) { 765 ParallelCompactData& sd = PSParallelCompact::summary_data(); 766 return sd.region(region_index); 767 } 768 #endif 769 770 elapsedTimer PSParallelCompact::_accumulated_time; 771 unsigned int PSParallelCompact::_total_invocations = 0; 772 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 773 jlong PSParallelCompact::_time_of_last_gc = 0; 774 CollectorCounters* PSParallelCompact::_counters = NULL; 775 ParMarkBitMap PSParallelCompact::_mark_bitmap; 776 ParallelCompactData PSParallelCompact::_summary_data; 777 778 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 779 780 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } 781 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 782 783 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 784 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 785 786 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true); 787 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false); 788 789 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); } 790 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); } 791 792 void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); } 793 794 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); } 795 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 796 797 void PSParallelCompact::post_initialize() { 798 ParallelScavengeHeap* heap = gc_heap(); 799 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 800 801 MemRegion mr = heap->reserved_region(); 802 _ref_processor = ReferenceProcessor::create_ref_processor( 803 mr, // span 804 true, // atomic_discovery 805 true, // mt_discovery 806 &_is_alive_closure, 807 ParallelGCThreads, 808 ParallelRefProcEnabled); 809 _counters = new CollectorCounters("PSParallelCompact", 1); 810 811 // Initialize static fields in ParCompactionManager. 812 ParCompactionManager::initialize(mark_bitmap()); 813 } 814 815 bool PSParallelCompact::initialize() { 816 ParallelScavengeHeap* heap = gc_heap(); 817 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 818 MemRegion mr = heap->reserved_region(); 819 820 // Was the old gen get allocated successfully? 821 if (!heap->old_gen()->is_allocated()) { 822 return false; 823 } 824 825 initialize_space_info(); 826 initialize_dead_wood_limiter(); 827 828 if (!_mark_bitmap.initialize(mr)) { 829 vm_shutdown_during_initialization("Unable to allocate bit map for " 830 "parallel garbage collection for the requested heap size."); 831 return false; 832 } 833 834 if (!_summary_data.initialize(mr)) { 835 vm_shutdown_during_initialization("Unable to allocate tables for " 836 "parallel garbage collection for the requested heap size."); 837 return false; 838 } 839 840 return true; 841 } 842 843 void PSParallelCompact::initialize_space_info() 844 { 845 memset(&_space_info, 0, sizeof(_space_info)); 846 847 ParallelScavengeHeap* heap = gc_heap(); 848 PSYoungGen* young_gen = heap->young_gen(); 849 MutableSpace* perm_space = heap->perm_gen()->object_space(); 850 851 _space_info[perm_space_id].set_space(perm_space); 852 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 853 _space_info[eden_space_id].set_space(young_gen->eden_space()); 854 _space_info[from_space_id].set_space(young_gen->from_space()); 855 _space_info[to_space_id].set_space(young_gen->to_space()); 856 857 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array()); 858 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 859 860 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top()); 861 if (TraceParallelOldGCDensePrefix) { 862 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT, 863 _space_info[perm_space_id].min_dense_prefix()); 864 } 865 } 866 867 void PSParallelCompact::initialize_dead_wood_limiter() 868 { 869 const size_t max = 100; 870 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 871 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 872 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 873 DEBUG_ONLY(_dwl_initialized = true;) 874 _dwl_adjustment = normal_distribution(1.0); 875 } 876 877 // Simple class for storing info about the heap at the start of GC, to be used 878 // after GC for comparison/printing. 879 class PreGCValues { 880 public: 881 PreGCValues() { } 882 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 883 884 void fill(ParallelScavengeHeap* heap) { 885 _heap_used = heap->used(); 886 _young_gen_used = heap->young_gen()->used_in_bytes(); 887 _old_gen_used = heap->old_gen()->used_in_bytes(); 888 _perm_gen_used = heap->perm_gen()->used_in_bytes(); 889 }; 890 891 size_t heap_used() const { return _heap_used; } 892 size_t young_gen_used() const { return _young_gen_used; } 893 size_t old_gen_used() const { return _old_gen_used; } 894 size_t perm_gen_used() const { return _perm_gen_used; } 895 896 private: 897 size_t _heap_used; 898 size_t _young_gen_used; 899 size_t _old_gen_used; 900 size_t _perm_gen_used; 901 }; 902 903 void 904 PSParallelCompact::clear_data_covering_space(SpaceId id) 905 { 906 // At this point, top is the value before GC, new_top() is the value that will 907 // be set at the end of GC. The marking bitmap is cleared to top; nothing 908 // should be marked above top. The summary data is cleared to the larger of 909 // top & new_top. 910 MutableSpace* const space = _space_info[id].space(); 911 HeapWord* const bot = space->bottom(); 912 HeapWord* const top = space->top(); 913 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 914 915 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 916 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 917 _mark_bitmap.clear_range(beg_bit, end_bit); 918 919 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 920 const size_t end_region = 921 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 922 _summary_data.clear_range(beg_region, end_region); 923 924 // Clear the data used to 'split' regions. 925 SplitInfo& split_info = _space_info[id].split_info(); 926 if (split_info.is_valid()) { 927 split_info.clear(); 928 } 929 DEBUG_ONLY(split_info.verify_clear();) 930 } 931 932 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 933 { 934 // Update the from & to space pointers in space_info, since they are swapped 935 // at each young gen gc. Do the update unconditionally (even though a 936 // promotion failure does not swap spaces) because an unknown number of minor 937 // collections will have swapped the spaces an unknown number of times. 938 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); 939 ParallelScavengeHeap* heap = gc_heap(); 940 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 941 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 942 943 pre_gc_values->fill(heap); 944 945 ParCompactionManager::reset(); 946 NOT_PRODUCT(_mark_bitmap.reset_counters()); 947 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 948 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 949 950 // Increment the invocation count 951 heap->increment_total_collections(true); 952 953 // We need to track unique mark sweep invocations as well. 954 _total_invocations++; 955 956 if (PrintHeapAtGC) { 957 Universe::print_heap_before_gc(); 958 } 959 960 // Fill in TLABs 961 heap->accumulate_statistics_all_tlabs(); 962 heap->ensure_parsability(true); // retire TLABs 963 964 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 965 HandleMark hm; // Discard invalid handles created during verification 966 gclog_or_tty->print(" VerifyBeforeGC:"); 967 Universe::verify(true); 968 } 969 970 // Verify object start arrays 971 if (VerifyObjectStartArray && 972 VerifyBeforeGC) { 973 heap->old_gen()->verify_object_start_array(); 974 heap->perm_gen()->verify_object_start_array(); 975 } 976 977 DEBUG_ONLY(mark_bitmap()->verify_clear();) 978 DEBUG_ONLY(summary_data().verify_clear();) 979 980 // Have worker threads release resources the next time they run a task. 981 gc_task_manager()->release_all_resources(); 982 } 983 984 void PSParallelCompact::post_compact() 985 { 986 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); 987 988 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 989 // Clear the marking bitmap, summary data and split info. 990 clear_data_covering_space(SpaceId(id)); 991 // Update top(). Must be done after clearing the bitmap and summary data. 992 _space_info[id].publish_new_top(); 993 } 994 995 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 996 MutableSpace* const from_space = _space_info[from_space_id].space(); 997 MutableSpace* const to_space = _space_info[to_space_id].space(); 998 999 ParallelScavengeHeap* heap = gc_heap(); 1000 bool eden_empty = eden_space->is_empty(); 1001 if (!eden_empty) { 1002 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1003 heap->young_gen(), heap->old_gen()); 1004 } 1005 1006 // Update heap occupancy information which is used as input to the soft ref 1007 // clearing policy at the next gc. 1008 Universe::update_heap_info_at_gc(); 1009 1010 bool young_gen_empty = eden_empty && from_space->is_empty() && 1011 to_space->is_empty(); 1012 1013 BarrierSet* bs = heap->barrier_set(); 1014 if (bs->is_a(BarrierSet::ModRef)) { 1015 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1016 MemRegion old_mr = heap->old_gen()->reserved(); 1017 MemRegion perm_mr = heap->perm_gen()->reserved(); 1018 assert(perm_mr.end() <= old_mr.start(), "Generations out of order"); 1019 1020 if (young_gen_empty) { 1021 modBS->clear(MemRegion(perm_mr.start(), old_mr.end())); 1022 } else { 1023 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end())); 1024 } 1025 } 1026 1027 Threads::gc_epilogue(); 1028 CodeCache::gc_epilogue(); 1029 1030 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1031 1032 ref_processor()->enqueue_discovered_references(NULL); 1033 1034 if (ZapUnusedHeapArea) { 1035 heap->gen_mangle_unused_area(); 1036 } 1037 1038 // Update time of last GC 1039 reset_millis_since_last_gc(); 1040 } 1041 1042 HeapWord* 1043 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1044 bool maximum_compaction) 1045 { 1046 const size_t region_size = ParallelCompactData::RegionSize; 1047 const ParallelCompactData& sd = summary_data(); 1048 1049 const MutableSpace* const space = _space_info[id].space(); 1050 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1051 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1052 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1053 1054 // Skip full regions at the beginning of the space--they are necessarily part 1055 // of the dense prefix. 1056 size_t full_count = 0; 1057 const RegionData* cp; 1058 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1059 ++full_count; 1060 } 1061 1062 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1063 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1064 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1065 if (maximum_compaction || cp == end_cp || interval_ended) { 1066 _maximum_compaction_gc_num = total_invocations(); 1067 return sd.region_to_addr(cp); 1068 } 1069 1070 HeapWord* const new_top = _space_info[id].new_top(); 1071 const size_t space_live = pointer_delta(new_top, space->bottom()); 1072 const size_t space_used = space->used_in_words(); 1073 const size_t space_capacity = space->capacity_in_words(); 1074 1075 const double cur_density = double(space_live) / space_capacity; 1076 const double deadwood_density = 1077 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1078 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1079 1080 if (TraceParallelOldGCDensePrefix) { 1081 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1082 cur_density, deadwood_density, deadwood_goal); 1083 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1084 "space_cap=" SIZE_FORMAT, 1085 space_live, space_used, 1086 space_capacity); 1087 } 1088 1089 // XXX - Use binary search? 1090 HeapWord* dense_prefix = sd.region_to_addr(cp); 1091 const RegionData* full_cp = cp; 1092 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1093 while (cp < end_cp) { 1094 HeapWord* region_destination = cp->destination(); 1095 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1096 if (TraceParallelOldGCDensePrefix && Verbose) { 1097 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1098 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1099 sd.region(cp), region_destination, 1100 dense_prefix, cur_deadwood); 1101 } 1102 1103 if (cur_deadwood >= deadwood_goal) { 1104 // Found the region that has the correct amount of deadwood to the left. 1105 // This typically occurs after crossing a fairly sparse set of regions, so 1106 // iterate backwards over those sparse regions, looking for the region 1107 // that has the lowest density of live objects 'to the right.' 1108 size_t space_to_left = sd.region(cp) * region_size; 1109 size_t live_to_left = space_to_left - cur_deadwood; 1110 size_t space_to_right = space_capacity - space_to_left; 1111 size_t live_to_right = space_live - live_to_left; 1112 double density_to_right = double(live_to_right) / space_to_right; 1113 while (cp > full_cp) { 1114 --cp; 1115 const size_t prev_region_live_to_right = live_to_right - 1116 cp->data_size(); 1117 const size_t prev_region_space_to_right = space_to_right + region_size; 1118 double prev_region_density_to_right = 1119 double(prev_region_live_to_right) / prev_region_space_to_right; 1120 if (density_to_right <= prev_region_density_to_right) { 1121 return dense_prefix; 1122 } 1123 if (TraceParallelOldGCDensePrefix && Verbose) { 1124 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1125 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1126 prev_region_density_to_right); 1127 } 1128 dense_prefix -= region_size; 1129 live_to_right = prev_region_live_to_right; 1130 space_to_right = prev_region_space_to_right; 1131 density_to_right = prev_region_density_to_right; 1132 } 1133 return dense_prefix; 1134 } 1135 1136 dense_prefix += region_size; 1137 ++cp; 1138 } 1139 1140 return dense_prefix; 1141 } 1142 1143 #ifndef PRODUCT 1144 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1145 const SpaceId id, 1146 const bool maximum_compaction, 1147 HeapWord* const addr) 1148 { 1149 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1150 RegionData* const cp = summary_data().region(region_idx); 1151 const MutableSpace* const space = _space_info[id].space(); 1152 HeapWord* const new_top = _space_info[id].new_top(); 1153 1154 const size_t space_live = pointer_delta(new_top, space->bottom()); 1155 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1156 const size_t space_cap = space->capacity_in_words(); 1157 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1158 const size_t live_to_right = new_top - cp->destination(); 1159 const size_t dead_to_right = space->top() - addr - live_to_right; 1160 1161 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1162 "spl=" SIZE_FORMAT " " 1163 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1164 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1165 " ratio=%10.8f", 1166 algorithm, addr, region_idx, 1167 space_live, 1168 dead_to_left, dead_to_left_pct, 1169 dead_to_right, live_to_right, 1170 double(dead_to_right) / live_to_right); 1171 } 1172 #endif // #ifndef PRODUCT 1173 1174 // Return a fraction indicating how much of the generation can be treated as 1175 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1176 // based on the density of live objects in the generation to determine a limit, 1177 // which is then adjusted so the return value is min_percent when the density is 1178 // 1. 1179 // 1180 // The following table shows some return values for a different values of the 1181 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1182 // min_percent is 1. 1183 // 1184 // fraction allowed as dead wood 1185 // ----------------------------------------------------------------- 1186 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1187 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1188 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1189 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1190 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1191 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1192 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1193 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1194 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1195 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1196 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1197 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1198 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1199 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1200 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1201 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1202 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1203 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1204 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1205 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1206 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1207 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1208 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1209 1210 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1211 { 1212 assert(_dwl_initialized, "uninitialized"); 1213 1214 // The raw limit is the value of the normal distribution at x = density. 1215 const double raw_limit = normal_distribution(density); 1216 1217 // Adjust the raw limit so it becomes the minimum when the density is 1. 1218 // 1219 // First subtract the adjustment value (which is simply the precomputed value 1220 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1221 // Then add the minimum value, so the minimum is returned when the density is 1222 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1223 const double min = double(min_percent) / 100.0; 1224 const double limit = raw_limit - _dwl_adjustment + min; 1225 return MAX2(limit, 0.0); 1226 } 1227 1228 ParallelCompactData::RegionData* 1229 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1230 const RegionData* end) 1231 { 1232 const size_t region_size = ParallelCompactData::RegionSize; 1233 ParallelCompactData& sd = summary_data(); 1234 size_t left = sd.region(beg); 1235 size_t right = end > beg ? sd.region(end) - 1 : left; 1236 1237 // Binary search. 1238 while (left < right) { 1239 // Equivalent to (left + right) / 2, but does not overflow. 1240 const size_t middle = left + (right - left) / 2; 1241 RegionData* const middle_ptr = sd.region(middle); 1242 HeapWord* const dest = middle_ptr->destination(); 1243 HeapWord* const addr = sd.region_to_addr(middle); 1244 assert(dest != NULL, "sanity"); 1245 assert(dest <= addr, "must move left"); 1246 1247 if (middle > left && dest < addr) { 1248 right = middle - 1; 1249 } else if (middle < right && middle_ptr->data_size() == region_size) { 1250 left = middle + 1; 1251 } else { 1252 return middle_ptr; 1253 } 1254 } 1255 return sd.region(left); 1256 } 1257 1258 ParallelCompactData::RegionData* 1259 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1260 const RegionData* end, 1261 size_t dead_words) 1262 { 1263 ParallelCompactData& sd = summary_data(); 1264 size_t left = sd.region(beg); 1265 size_t right = end > beg ? sd.region(end) - 1 : left; 1266 1267 // Binary search. 1268 while (left < right) { 1269 // Equivalent to (left + right) / 2, but does not overflow. 1270 const size_t middle = left + (right - left) / 2; 1271 RegionData* const middle_ptr = sd.region(middle); 1272 HeapWord* const dest = middle_ptr->destination(); 1273 HeapWord* const addr = sd.region_to_addr(middle); 1274 assert(dest != NULL, "sanity"); 1275 assert(dest <= addr, "must move left"); 1276 1277 const size_t dead_to_left = pointer_delta(addr, dest); 1278 if (middle > left && dead_to_left > dead_words) { 1279 right = middle - 1; 1280 } else if (middle < right && dead_to_left < dead_words) { 1281 left = middle + 1; 1282 } else { 1283 return middle_ptr; 1284 } 1285 } 1286 return sd.region(left); 1287 } 1288 1289 // The result is valid during the summary phase, after the initial summarization 1290 // of each space into itself, and before final summarization. 1291 inline double 1292 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1293 HeapWord* const bottom, 1294 HeapWord* const top, 1295 HeapWord* const new_top) 1296 { 1297 ParallelCompactData& sd = summary_data(); 1298 1299 assert(cp != NULL, "sanity"); 1300 assert(bottom != NULL, "sanity"); 1301 assert(top != NULL, "sanity"); 1302 assert(new_top != NULL, "sanity"); 1303 assert(top >= new_top, "summary data problem?"); 1304 assert(new_top > bottom, "space is empty; should not be here"); 1305 assert(new_top >= cp->destination(), "sanity"); 1306 assert(top >= sd.region_to_addr(cp), "sanity"); 1307 1308 HeapWord* const destination = cp->destination(); 1309 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1310 const size_t compacted_region_live = pointer_delta(new_top, destination); 1311 const size_t compacted_region_used = pointer_delta(top, 1312 sd.region_to_addr(cp)); 1313 const size_t reclaimable = compacted_region_used - compacted_region_live; 1314 1315 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1316 return double(reclaimable) / divisor; 1317 } 1318 1319 // Return the address of the end of the dense prefix, a.k.a. the start of the 1320 // compacted region. The address is always on a region boundary. 1321 // 1322 // Completely full regions at the left are skipped, since no compaction can 1323 // occur in those regions. Then the maximum amount of dead wood to allow is 1324 // computed, based on the density (amount live / capacity) of the generation; 1325 // the region with approximately that amount of dead space to the left is 1326 // identified as the limit region. Regions between the last completely full 1327 // region and the limit region are scanned and the one that has the best 1328 // (maximum) reclaimed_ratio() is selected. 1329 HeapWord* 1330 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1331 bool maximum_compaction) 1332 { 1333 if (ParallelOldGCSplitALot) { 1334 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1335 // The value was chosen to provoke splitting a young gen space; use it. 1336 return _space_info[id].dense_prefix(); 1337 } 1338 } 1339 1340 const size_t region_size = ParallelCompactData::RegionSize; 1341 const ParallelCompactData& sd = summary_data(); 1342 1343 const MutableSpace* const space = _space_info[id].space(); 1344 HeapWord* const top = space->top(); 1345 HeapWord* const top_aligned_up = sd.region_align_up(top); 1346 HeapWord* const new_top = _space_info[id].new_top(); 1347 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1348 HeapWord* const bottom = space->bottom(); 1349 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1350 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1351 const RegionData* const new_top_cp = 1352 sd.addr_to_region_ptr(new_top_aligned_up); 1353 1354 // Skip full regions at the beginning of the space--they are necessarily part 1355 // of the dense prefix. 1356 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1357 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1358 space->is_empty(), "no dead space allowed to the left"); 1359 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1360 "region must have dead space"); 1361 1362 // The gc number is saved whenever a maximum compaction is done, and used to 1363 // determine when the maximum compaction interval has expired. This avoids 1364 // successive max compactions for different reasons. 1365 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1366 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1367 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1368 total_invocations() == HeapFirstMaximumCompactionCount; 1369 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1370 _maximum_compaction_gc_num = total_invocations(); 1371 return sd.region_to_addr(full_cp); 1372 } 1373 1374 const size_t space_live = pointer_delta(new_top, bottom); 1375 const size_t space_used = space->used_in_words(); 1376 const size_t space_capacity = space->capacity_in_words(); 1377 1378 const double density = double(space_live) / double(space_capacity); 1379 const size_t min_percent_free = 1380 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio; 1381 const double limiter = dead_wood_limiter(density, min_percent_free); 1382 const size_t dead_wood_max = space_used - space_live; 1383 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1384 dead_wood_max); 1385 1386 if (TraceParallelOldGCDensePrefix) { 1387 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1388 "space_cap=" SIZE_FORMAT, 1389 space_live, space_used, 1390 space_capacity); 1391 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1392 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1393 density, min_percent_free, limiter, 1394 dead_wood_max, dead_wood_limit); 1395 } 1396 1397 // Locate the region with the desired amount of dead space to the left. 1398 const RegionData* const limit_cp = 1399 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1400 1401 // Scan from the first region with dead space to the limit region and find the 1402 // one with the best (largest) reclaimed ratio. 1403 double best_ratio = 0.0; 1404 const RegionData* best_cp = full_cp; 1405 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1406 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1407 if (tmp_ratio > best_ratio) { 1408 best_cp = cp; 1409 best_ratio = tmp_ratio; 1410 } 1411 } 1412 1413 #if 0 1414 // Something to consider: if the region with the best ratio is 'close to' the 1415 // first region w/free space, choose the first region with free space 1416 // ("first-free"). The first-free region is usually near the start of the 1417 // heap, which means we are copying most of the heap already, so copy a bit 1418 // more to get complete compaction. 1419 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1420 _maximum_compaction_gc_num = total_invocations(); 1421 best_cp = full_cp; 1422 } 1423 #endif // #if 0 1424 1425 return sd.region_to_addr(best_cp); 1426 } 1427 1428 #ifndef PRODUCT 1429 void 1430 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1431 size_t words) 1432 { 1433 if (TraceParallelOldGCSummaryPhase) { 1434 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1435 SIZE_FORMAT, start, start + words, words); 1436 } 1437 1438 ObjectStartArray* const start_array = _space_info[id].start_array(); 1439 CollectedHeap::fill_with_objects(start, words); 1440 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1441 _mark_bitmap.mark_obj(p, words); 1442 _summary_data.add_obj(p, words); 1443 start_array->allocate_block(p); 1444 } 1445 } 1446 1447 void 1448 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1449 { 1450 ParallelCompactData& sd = summary_data(); 1451 MutableSpace* space = _space_info[id].space(); 1452 1453 // Find the source and destination start addresses. 1454 HeapWord* const src_addr = sd.region_align_down(start); 1455 HeapWord* dst_addr; 1456 if (src_addr < start) { 1457 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1458 } else if (src_addr > space->bottom()) { 1459 // The start (the original top() value) is aligned to a region boundary so 1460 // the associated region does not have a destination. Compute the 1461 // destination from the previous region. 1462 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1463 dst_addr = cp->destination() + cp->data_size(); 1464 } else { 1465 // Filling the entire space. 1466 dst_addr = space->bottom(); 1467 } 1468 assert(dst_addr != NULL, "sanity"); 1469 1470 // Update the summary data. 1471 bool result = _summary_data.summarize(_space_info[id].split_info(), 1472 src_addr, space->top(), NULL, 1473 dst_addr, space->end(), 1474 _space_info[id].new_top_addr()); 1475 assert(result, "should not fail: bad filler object size"); 1476 } 1477 1478 void 1479 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1480 { 1481 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1482 return; 1483 } 1484 1485 MutableSpace* const space = _space_info[id].space(); 1486 if (space->is_empty()) { 1487 HeapWord* b = space->bottom(); 1488 HeapWord* t = b + space->capacity_in_words() / 2; 1489 space->set_top(t); 1490 if (ZapUnusedHeapArea) { 1491 space->set_top_for_allocations(); 1492 } 1493 1494 size_t obj_len = 8; 1495 while (b + obj_len <= t) { 1496 CollectedHeap::fill_with_object(b, obj_len); 1497 mark_bitmap()->mark_obj(b, obj_len); 1498 summary_data().add_obj(b, obj_len); 1499 b += obj_len; 1500 obj_len = (obj_len & 0x18) + 8; // 8 16 24 32 8 16 24 32 ... 1501 } 1502 if (b < t) { 1503 // The loop didn't completely fill to t (top); adjust top downward. 1504 space->set_top(b); 1505 if (ZapUnusedHeapArea) { 1506 space->set_top_for_allocations(); 1507 } 1508 } 1509 1510 HeapWord** nta = _space_info[id].new_top_addr(); 1511 bool result = summary_data().summarize(_space_info[id].split_info(), 1512 space->bottom(), space->top(), NULL, 1513 space->bottom(), space->end(), nta); 1514 assert(result, "space must fit into itself"); 1515 } 1516 } 1517 1518 void 1519 PSParallelCompact::provoke_split(bool & max_compaction) 1520 { 1521 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1522 return; 1523 } 1524 1525 const size_t region_size = ParallelCompactData::RegionSize; 1526 ParallelCompactData& sd = summary_data(); 1527 1528 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1529 MutableSpace* const from_space = _space_info[from_space_id].space(); 1530 const size_t eden_live = pointer_delta(eden_space->top(), 1531 _space_info[eden_space_id].new_top()); 1532 const size_t from_live = pointer_delta(from_space->top(), 1533 _space_info[from_space_id].new_top()); 1534 1535 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1536 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1537 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1538 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1539 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1540 1541 // Choose the space to split; need at least 2 regions live (or fillable). 1542 SpaceId id; 1543 MutableSpace* space; 1544 size_t live_words; 1545 size_t fill_words; 1546 if (eden_live + eden_fillable >= region_size * 2) { 1547 id = eden_space_id; 1548 space = eden_space; 1549 live_words = eden_live; 1550 fill_words = eden_fillable; 1551 } else if (from_live + from_fillable >= region_size * 2) { 1552 id = from_space_id; 1553 space = from_space; 1554 live_words = from_live; 1555 fill_words = from_fillable; 1556 } else { 1557 return; // Give up. 1558 } 1559 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1560 1561 if (live_words < region_size * 2) { 1562 // Fill from top() to end() w/live objects of mixed sizes. 1563 HeapWord* const fill_start = space->top(); 1564 live_words += fill_words; 1565 1566 space->set_top(fill_start + fill_words); 1567 if (ZapUnusedHeapArea) { 1568 space->set_top_for_allocations(); 1569 } 1570 1571 HeapWord* cur_addr = fill_start; 1572 while (fill_words > 0) { 1573 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1574 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1575 if (fill_words - cur_size < min_fill_size) { 1576 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1577 } 1578 1579 CollectedHeap::fill_with_object(cur_addr, cur_size); 1580 mark_bitmap()->mark_obj(cur_addr, cur_size); 1581 sd.add_obj(cur_addr, cur_size); 1582 1583 cur_addr += cur_size; 1584 fill_words -= cur_size; 1585 } 1586 1587 summarize_new_objects(id, fill_start); 1588 } 1589 1590 max_compaction = false; 1591 1592 // Manipulate the old gen so that it has room for about half of the live data 1593 // in the target young gen space (live_words / 2). 1594 id = old_space_id; 1595 space = _space_info[id].space(); 1596 const size_t free_at_end = space->free_in_words(); 1597 const size_t free_target = align_object_size(live_words / 2); 1598 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1599 1600 if (free_at_end >= free_target + min_fill_size) { 1601 // Fill space above top() and set the dense prefix so everything survives. 1602 HeapWord* const fill_start = space->top(); 1603 const size_t fill_size = free_at_end - free_target; 1604 space->set_top(space->top() + fill_size); 1605 if (ZapUnusedHeapArea) { 1606 space->set_top_for_allocations(); 1607 } 1608 fill_with_live_objects(id, fill_start, fill_size); 1609 summarize_new_objects(id, fill_start); 1610 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1611 } else if (dead + free_at_end > free_target) { 1612 // Find a dense prefix that makes the right amount of space available. 1613 HeapWord* cur = sd.region_align_down(space->top()); 1614 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1615 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1616 while (dead_to_right < free_target) { 1617 cur -= region_size; 1618 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1619 dead_to_right = pointer_delta(space->end(), cur_destination); 1620 } 1621 _space_info[id].set_dense_prefix(cur); 1622 } 1623 } 1624 #endif // #ifndef PRODUCT 1625 1626 void PSParallelCompact::summarize_spaces_quick() 1627 { 1628 for (unsigned int i = 0; i < last_space_id; ++i) { 1629 const MutableSpace* space = _space_info[i].space(); 1630 HeapWord** nta = _space_info[i].new_top_addr(); 1631 bool result = _summary_data.summarize(_space_info[i].split_info(), 1632 space->bottom(), space->top(), NULL, 1633 space->bottom(), space->end(), nta); 1634 assert(result, "space must fit into itself"); 1635 _space_info[i].set_dense_prefix(space->bottom()); 1636 } 1637 1638 #ifndef PRODUCT 1639 if (ParallelOldGCSplitALot) { 1640 provoke_split_fill_survivor(to_space_id); 1641 } 1642 #endif // #ifndef PRODUCT 1643 } 1644 1645 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1646 { 1647 HeapWord* const dense_prefix_end = dense_prefix(id); 1648 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1649 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1650 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1651 // Only enough dead space is filled so that any remaining dead space to the 1652 // left is larger than the minimum filler object. (The remainder is filled 1653 // during the copy/update phase.) 1654 // 1655 // The size of the dead space to the right of the boundary is not a 1656 // concern, since compaction will be able to use whatever space is 1657 // available. 1658 // 1659 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1660 // surrounds the space to be filled with an object. 1661 // 1662 // In the 32-bit VM, each bit represents two 32-bit words: 1663 // +---+ 1664 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1665 // end_bits: ... x x x | 0 | || 0 x x ... 1666 // +---+ 1667 // 1668 // In the 64-bit VM, each bit represents one 64-bit word: 1669 // +------------+ 1670 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1671 // end_bits: ... x x 1 | 0 || 0 | x x ... 1672 // +------------+ 1673 // +-------+ 1674 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1675 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1676 // +-------+ 1677 // +-----------+ 1678 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1679 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1680 // +-----------+ 1681 // +-------+ 1682 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1683 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1684 // +-------+ 1685 1686 // Initially assume case a, c or e will apply. 1687 size_t obj_len = (size_t)oopDesc::header_size(); 1688 HeapWord* obj_beg = dense_prefix_end - obj_len; 1689 1690 #ifdef _LP64 1691 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1692 // Case b above. 1693 obj_beg = dense_prefix_end - 1; 1694 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1695 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1696 // Case d above. 1697 obj_beg = dense_prefix_end - 3; 1698 obj_len = 3; 1699 } 1700 #endif // #ifdef _LP64 1701 1702 CollectedHeap::fill_with_object(obj_beg, obj_len); 1703 _mark_bitmap.mark_obj(obj_beg, obj_len); 1704 _summary_data.add_obj(obj_beg, obj_len); 1705 assert(start_array(id) != NULL, "sanity"); 1706 start_array(id)->allocate_block(obj_beg); 1707 } 1708 } 1709 1710 void 1711 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1712 { 1713 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1714 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1715 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1716 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1717 cur->set_source_region(0); 1718 } 1719 } 1720 1721 void 1722 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1723 { 1724 assert(id < last_space_id, "id out of range"); 1725 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1726 ParallelOldGCSplitALot && id == old_space_id, 1727 "should have been reset in summarize_spaces_quick()"); 1728 1729 const MutableSpace* space = _space_info[id].space(); 1730 if (_space_info[id].new_top() != space->bottom()) { 1731 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1732 _space_info[id].set_dense_prefix(dense_prefix_end); 1733 1734 #ifndef PRODUCT 1735 if (TraceParallelOldGCDensePrefix) { 1736 print_dense_prefix_stats("ratio", id, maximum_compaction, 1737 dense_prefix_end); 1738 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1739 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1740 } 1741 #endif // #ifndef PRODUCT 1742 1743 // Recompute the summary data, taking into account the dense prefix. If 1744 // every last byte will be reclaimed, then the existing summary data which 1745 // compacts everything can be left in place. 1746 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1747 // If dead space crosses the dense prefix boundary, it is (at least 1748 // partially) filled with a dummy object, marked live and added to the 1749 // summary data. This simplifies the copy/update phase and must be done 1750 // before the final locations of objects are determined, to prevent 1751 // leaving a fragment of dead space that is too small to fill. 1752 fill_dense_prefix_end(id); 1753 1754 // Compute the destination of each Region, and thus each object. 1755 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1756 _summary_data.summarize(_space_info[id].split_info(), 1757 dense_prefix_end, space->top(), NULL, 1758 dense_prefix_end, space->end(), 1759 _space_info[id].new_top_addr()); 1760 } 1761 } 1762 1763 if (TraceParallelOldGCSummaryPhase) { 1764 const size_t region_size = ParallelCompactData::RegionSize; 1765 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1766 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1767 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1768 HeapWord* const new_top = _space_info[id].new_top(); 1769 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1770 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1771 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1772 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1773 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1774 id, space->capacity_in_words(), dense_prefix_end, 1775 dp_region, dp_words / region_size, 1776 cr_words / region_size, new_top); 1777 } 1778 } 1779 1780 #ifndef PRODUCT 1781 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1782 HeapWord* dst_beg, HeapWord* dst_end, 1783 SpaceId src_space_id, 1784 HeapWord* src_beg, HeapWord* src_end) 1785 { 1786 if (TraceParallelOldGCSummaryPhase) { 1787 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1788 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1789 SIZE_FORMAT "-" SIZE_FORMAT " " 1790 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1791 SIZE_FORMAT "-" SIZE_FORMAT, 1792 src_space_id, space_names[src_space_id], 1793 dst_space_id, space_names[dst_space_id], 1794 src_beg, src_end, 1795 _summary_data.addr_to_region_idx(src_beg), 1796 _summary_data.addr_to_region_idx(src_end), 1797 dst_beg, dst_end, 1798 _summary_data.addr_to_region_idx(dst_beg), 1799 _summary_data.addr_to_region_idx(dst_end)); 1800 } 1801 } 1802 #endif // #ifndef PRODUCT 1803 1804 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1805 bool maximum_compaction) 1806 { 1807 EventMark m("2 summarize"); 1808 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); 1809 // trace("2"); 1810 1811 #ifdef ASSERT 1812 if (TraceParallelOldGCMarkingPhase) { 1813 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1814 "add_obj_bytes=" SIZE_FORMAT, 1815 add_obj_count, add_obj_size * HeapWordSize); 1816 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1817 "mark_bitmap_bytes=" SIZE_FORMAT, 1818 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1819 } 1820 #endif // #ifdef ASSERT 1821 1822 // Quick summarization of each space into itself, to see how much is live. 1823 summarize_spaces_quick(); 1824 1825 if (TraceParallelOldGCSummaryPhase) { 1826 tty->print_cr("summary_phase: after summarizing each space to self"); 1827 Universe::print(); 1828 NOT_PRODUCT(print_region_ranges()); 1829 if (Verbose) { 1830 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1831 } 1832 } 1833 1834 // The amount of live data that will end up in old space (assuming it fits). 1835 size_t old_space_total_live = 0; 1836 assert(perm_space_id < old_space_id, "should not count perm data here"); 1837 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1838 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1839 _space_info[id].space()->bottom()); 1840 } 1841 1842 MutableSpace* const old_space = _space_info[old_space_id].space(); 1843 const size_t old_capacity = old_space->capacity_in_words(); 1844 if (old_space_total_live > old_capacity) { 1845 // XXX - should also try to expand 1846 maximum_compaction = true; 1847 } 1848 #ifndef PRODUCT 1849 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1850 provoke_split(maximum_compaction); 1851 } 1852 #endif // #ifndef PRODUCT 1853 1854 // Permanent and Old generations. 1855 summarize_space(perm_space_id, maximum_compaction); 1856 summarize_space(old_space_id, maximum_compaction); 1857 1858 // Summarize the remaining spaces in the young gen. The initial target space 1859 // is the old gen. If a space does not fit entirely into the target, then the 1860 // remainder is compacted into the space itself and that space becomes the new 1861 // target. 1862 SpaceId dst_space_id = old_space_id; 1863 HeapWord* dst_space_end = old_space->end(); 1864 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1865 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1866 const MutableSpace* space = _space_info[id].space(); 1867 const size_t live = pointer_delta(_space_info[id].new_top(), 1868 space->bottom()); 1869 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1870 1871 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1872 SpaceId(id), space->bottom(), space->top());) 1873 if (live > 0 && live <= available) { 1874 // All the live data will fit. 1875 bool done = _summary_data.summarize(_space_info[id].split_info(), 1876 space->bottom(), space->top(), 1877 NULL, 1878 *new_top_addr, dst_space_end, 1879 new_top_addr); 1880 assert(done, "space must fit into old gen"); 1881 1882 // Reset the new_top value for the space. 1883 _space_info[id].set_new_top(space->bottom()); 1884 } else if (live > 0) { 1885 // Attempt to fit part of the source space into the target space. 1886 HeapWord* next_src_addr = NULL; 1887 bool done = _summary_data.summarize(_space_info[id].split_info(), 1888 space->bottom(), space->top(), 1889 &next_src_addr, 1890 *new_top_addr, dst_space_end, 1891 new_top_addr); 1892 assert(!done, "space should not fit into old gen"); 1893 assert(next_src_addr != NULL, "sanity"); 1894 1895 // The source space becomes the new target, so the remainder is compacted 1896 // within the space itself. 1897 dst_space_id = SpaceId(id); 1898 dst_space_end = space->end(); 1899 new_top_addr = _space_info[id].new_top_addr(); 1900 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1901 space->bottom(), dst_space_end, 1902 SpaceId(id), next_src_addr, space->top());) 1903 done = _summary_data.summarize(_space_info[id].split_info(), 1904 next_src_addr, space->top(), 1905 NULL, 1906 space->bottom(), dst_space_end, 1907 new_top_addr); 1908 assert(done, "space must fit when compacted into itself"); 1909 assert(*new_top_addr <= space->top(), "usage should not grow"); 1910 } 1911 } 1912 1913 if (TraceParallelOldGCSummaryPhase) { 1914 tty->print_cr("summary_phase: after final summarization"); 1915 Universe::print(); 1916 NOT_PRODUCT(print_region_ranges()); 1917 if (Verbose) { 1918 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1919 } 1920 } 1921 } 1922 1923 // This method should contain all heap-specific policy for invoking a full 1924 // collection. invoke_no_policy() will only attempt to compact the heap; it 1925 // will do nothing further. If we need to bail out for policy reasons, scavenge 1926 // before full gc, or any other specialized behavior, it needs to be added here. 1927 // 1928 // Note that this method should only be called from the vm_thread while at a 1929 // safepoint. 1930 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1931 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1932 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1933 "should be in vm thread"); 1934 ParallelScavengeHeap* heap = gc_heap(); 1935 GCCause::Cause gc_cause = heap->gc_cause(); 1936 assert(!heap->is_gc_active(), "not reentrant"); 1937 1938 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1939 1940 // Before each allocation/collection attempt, find out from the 1941 // policy object if GCs are, on the whole, taking too long. If so, 1942 // bail out without attempting a collection. The exceptions are 1943 // for explicitly requested GC's. 1944 if (!policy->gc_time_limit_exceeded() || 1945 GCCause::is_user_requested_gc(gc_cause) || 1946 GCCause::is_serviceability_requested_gc(gc_cause)) { 1947 IsGCActiveMark mark; 1948 1949 if (ScavengeBeforeFullGC) { 1950 PSScavenge::invoke_no_policy(); 1951 } 1952 1953 PSParallelCompact::invoke_no_policy(maximum_heap_compaction); 1954 } 1955 } 1956 1957 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1958 size_t addr_region_index = addr_to_region_idx(addr); 1959 return region_index == addr_region_index; 1960 } 1961 1962 // This method contains no policy. You should probably 1963 // be calling invoke() instead. 1964 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1965 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1966 assert(ref_processor() != NULL, "Sanity"); 1967 1968 if (GC_locker::check_active_before_gc()) { 1969 return; 1970 } 1971 1972 TimeStamp marking_start; 1973 TimeStamp compaction_start; 1974 TimeStamp collection_exit; 1975 1976 ParallelScavengeHeap* heap = gc_heap(); 1977 GCCause::Cause gc_cause = heap->gc_cause(); 1978 PSYoungGen* young_gen = heap->young_gen(); 1979 PSOldGen* old_gen = heap->old_gen(); 1980 PSPermGen* perm_gen = heap->perm_gen(); 1981 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1982 1983 if (ZapUnusedHeapArea) { 1984 // Save information needed to minimize mangling 1985 heap->record_gen_tops_before_GC(); 1986 } 1987 1988 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 1989 1990 // Make sure data structures are sane, make the heap parsable, and do other 1991 // miscellaneous bookkeeping. 1992 PreGCValues pre_gc_values; 1993 pre_compact(&pre_gc_values); 1994 1995 // Get the compaction manager reserved for the VM thread. 1996 ParCompactionManager* const vmthread_cm = 1997 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1998 1999 // Place after pre_compact() where the number of invocations is incremented. 2000 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2001 2002 { 2003 ResourceMark rm; 2004 HandleMark hm; 2005 2006 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc; 2007 2008 // This is useful for debugging but don't change the output the 2009 // the customer sees. 2010 const char* gc_cause_str = "Full GC"; 2011 if (is_system_gc && PrintGCDetails) { 2012 gc_cause_str = "Full GC (System)"; 2013 } 2014 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2015 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2016 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty); 2017 TraceCollectorStats tcs(counters()); 2018 TraceMemoryManagerStats tms(true /* Full GC */); 2019 2020 if (TraceGen1Time) accumulated_time()->start(); 2021 2022 // Let the size policy know we're starting 2023 size_policy->major_collection_begin(); 2024 2025 // When collecting the permanent generation methodOops may be moving, 2026 // so we either have to flush all bcp data or convert it into bci. 2027 CodeCache::gc_prologue(); 2028 Threads::gc_prologue(); 2029 2030 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2031 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2032 2033 ref_processor()->enable_discovery(); 2034 ref_processor()->setup_policy(maximum_heap_compaction); 2035 2036 bool marked_for_unloading = false; 2037 2038 marking_start.update(); 2039 marking_phase(vmthread_cm, maximum_heap_compaction); 2040 2041 #ifndef PRODUCT 2042 if (TraceParallelOldGCMarkingPhase) { 2043 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " 2044 "cas_by_another %d", 2045 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), 2046 mark_bitmap()->cas_by_another()); 2047 } 2048 #endif // #ifndef PRODUCT 2049 2050 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc; 2051 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2052 2053 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2054 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2055 2056 // adjust_roots() updates Universe::_intArrayKlassObj which is 2057 // needed by the compaction for filling holes in the dense prefix. 2058 adjust_roots(); 2059 2060 compaction_start.update(); 2061 // Does the perm gen always have to be done serially because 2062 // klasses are used in the update of an object? 2063 compact_perm(vmthread_cm); 2064 2065 if (UseParallelOldGCCompacting) { 2066 compact(); 2067 } else { 2068 compact_serial(vmthread_cm); 2069 } 2070 2071 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2072 // done before resizing. 2073 post_compact(); 2074 2075 // Let the size policy know we're done 2076 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2077 2078 if (UseAdaptiveSizePolicy) { 2079 if (PrintAdaptiveSizePolicy) { 2080 gclog_or_tty->print("AdaptiveSizeStart: "); 2081 gclog_or_tty->stamp(); 2082 gclog_or_tty->print_cr(" collection: %d ", 2083 heap->total_collections()); 2084 if (Verbose) { 2085 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d" 2086 " perm_gen_capacity: %d ", 2087 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(), 2088 perm_gen->capacity_in_bytes()); 2089 } 2090 } 2091 2092 // Don't check if the size_policy is ready here. Let 2093 // the size_policy check that internally. 2094 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2095 ((gc_cause != GCCause::_java_lang_system_gc) || 2096 UseAdaptiveSizePolicyWithSystemGC)) { 2097 // Calculate optimal free space amounts 2098 assert(young_gen->max_size() > 2099 young_gen->from_space()->capacity_in_bytes() + 2100 young_gen->to_space()->capacity_in_bytes(), 2101 "Sizes of space in young gen are out-of-bounds"); 2102 size_t max_eden_size = young_gen->max_size() - 2103 young_gen->from_space()->capacity_in_bytes() - 2104 young_gen->to_space()->capacity_in_bytes(); 2105 size_policy->compute_generation_free_space( 2106 young_gen->used_in_bytes(), 2107 young_gen->eden_space()->used_in_bytes(), 2108 old_gen->used_in_bytes(), 2109 perm_gen->used_in_bytes(), 2110 young_gen->eden_space()->capacity_in_bytes(), 2111 old_gen->max_gen_size(), 2112 max_eden_size, 2113 true /* full gc*/, 2114 gc_cause); 2115 2116 heap->resize_old_gen( 2117 size_policy->calculated_old_free_size_in_bytes()); 2118 2119 // Don't resize the young generation at an major collection. A 2120 // desired young generation size may have been calculated but 2121 // resizing the young generation complicates the code because the 2122 // resizing of the old generation may have moved the boundary 2123 // between the young generation and the old generation. Let the 2124 // young generation resizing happen at the minor collections. 2125 } 2126 if (PrintAdaptiveSizePolicy) { 2127 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2128 heap->total_collections()); 2129 } 2130 } 2131 2132 if (UsePerfData) { 2133 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2134 counters->update_counters(); 2135 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2136 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2137 } 2138 2139 heap->resize_all_tlabs(); 2140 2141 // We collected the perm gen, so we'll resize it here. 2142 perm_gen->compute_new_size(pre_gc_values.perm_gen_used()); 2143 2144 if (TraceGen1Time) accumulated_time()->stop(); 2145 2146 if (PrintGC) { 2147 if (PrintGCDetails) { 2148 // No GC timestamp here. This is after GC so it would be confusing. 2149 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2150 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2151 heap->print_heap_change(pre_gc_values.heap_used()); 2152 // Print perm gen last (print_heap_change() excludes the perm gen). 2153 perm_gen->print_used_change(pre_gc_values.perm_gen_used()); 2154 } else { 2155 heap->print_heap_change(pre_gc_values.heap_used()); 2156 } 2157 } 2158 2159 // Track memory usage and detect low memory 2160 MemoryService::track_memory_usage(); 2161 heap->update_counters(); 2162 2163 if (PrintGCDetails) { 2164 if (size_policy->print_gc_time_limit_would_be_exceeded()) { 2165 if (size_policy->gc_time_limit_exceeded()) { 2166 gclog_or_tty->print_cr(" GC time is exceeding GCTimeLimit " 2167 "of %d%%", GCTimeLimit); 2168 } else { 2169 gclog_or_tty->print_cr(" GC time would exceed GCTimeLimit " 2170 "of %d%%", GCTimeLimit); 2171 } 2172 } 2173 size_policy->set_print_gc_time_limit_would_be_exceeded(false); 2174 } 2175 } 2176 2177 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2178 HandleMark hm; // Discard invalid handles created during verification 2179 gclog_or_tty->print(" VerifyAfterGC:"); 2180 Universe::verify(false); 2181 } 2182 2183 // Re-verify object start arrays 2184 if (VerifyObjectStartArray && 2185 VerifyAfterGC) { 2186 old_gen->verify_object_start_array(); 2187 perm_gen->verify_object_start_array(); 2188 } 2189 2190 if (ZapUnusedHeapArea) { 2191 old_gen->object_space()->check_mangled_unused_area_complete(); 2192 perm_gen->object_space()->check_mangled_unused_area_complete(); 2193 } 2194 2195 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2196 2197 collection_exit.update(); 2198 2199 if (PrintHeapAtGC) { 2200 Universe::print_heap_after_gc(); 2201 } 2202 if (PrintGCTaskTimeStamps) { 2203 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2204 INT64_FORMAT, 2205 marking_start.ticks(), compaction_start.ticks(), 2206 collection_exit.ticks()); 2207 gc_task_manager()->print_task_time_stamps(); 2208 } 2209 } 2210 2211 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2212 PSYoungGen* young_gen, 2213 PSOldGen* old_gen) { 2214 MutableSpace* const eden_space = young_gen->eden_space(); 2215 assert(!eden_space->is_empty(), "eden must be non-empty"); 2216 assert(young_gen->virtual_space()->alignment() == 2217 old_gen->virtual_space()->alignment(), "alignments do not match"); 2218 2219 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2220 return false; 2221 } 2222 2223 // Both generations must be completely committed. 2224 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2225 return false; 2226 } 2227 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2228 return false; 2229 } 2230 2231 // Figure out how much to take from eden. Include the average amount promoted 2232 // in the total; otherwise the next young gen GC will simply bail out to a 2233 // full GC. 2234 const size_t alignment = old_gen->virtual_space()->alignment(); 2235 const size_t eden_used = eden_space->used_in_bytes(); 2236 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2237 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2238 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2239 2240 if (absorb_size >= eden_capacity) { 2241 return false; // Must leave some space in eden. 2242 } 2243 2244 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2245 if (new_young_size < young_gen->min_gen_size()) { 2246 return false; // Respect young gen minimum size. 2247 } 2248 2249 if (TraceAdaptiveGCBoundary && Verbose) { 2250 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2251 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2252 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2253 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2254 absorb_size / K, 2255 eden_capacity / K, (eden_capacity - absorb_size) / K, 2256 young_gen->from_space()->used_in_bytes() / K, 2257 young_gen->to_space()->used_in_bytes() / K, 2258 young_gen->capacity_in_bytes() / K, new_young_size / K); 2259 } 2260 2261 // Fill the unused part of the old gen. 2262 MutableSpace* const old_space = old_gen->object_space(); 2263 HeapWord* const unused_start = old_space->top(); 2264 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2265 2266 if (unused_words > 0) { 2267 if (unused_words < CollectedHeap::min_fill_size()) { 2268 return false; // If the old gen cannot be filled, must give up. 2269 } 2270 CollectedHeap::fill_with_objects(unused_start, unused_words); 2271 } 2272 2273 // Take the live data from eden and set both top and end in the old gen to 2274 // eden top. (Need to set end because reset_after_change() mangles the region 2275 // from end to virtual_space->high() in debug builds). 2276 HeapWord* const new_top = eden_space->top(); 2277 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2278 absorb_size); 2279 young_gen->reset_after_change(); 2280 old_space->set_top(new_top); 2281 old_space->set_end(new_top); 2282 old_gen->reset_after_change(); 2283 2284 // Update the object start array for the filler object and the data from eden. 2285 ObjectStartArray* const start_array = old_gen->start_array(); 2286 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2287 start_array->allocate_block(p); 2288 } 2289 2290 // Could update the promoted average here, but it is not typically updated at 2291 // full GCs and the value to use is unclear. Something like 2292 // 2293 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2294 2295 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2296 return true; 2297 } 2298 2299 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2300 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2301 "shouldn't return NULL"); 2302 return ParallelScavengeHeap::gc_task_manager(); 2303 } 2304 2305 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2306 bool maximum_heap_compaction) { 2307 // Recursively traverse all live objects and mark them 2308 EventMark m("1 mark object"); 2309 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); 2310 2311 ParallelScavengeHeap* heap = gc_heap(); 2312 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2313 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2314 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2315 2316 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2317 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2318 2319 { 2320 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); 2321 2322 GCTaskQueue* q = GCTaskQueue::create(); 2323 2324 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2325 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2326 // We scan the thread roots in parallel 2327 Threads::create_thread_roots_marking_tasks(q); 2328 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2329 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2330 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2331 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2332 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2333 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols)); 2334 2335 if (parallel_gc_threads > 1) { 2336 for (uint j = 0; j < parallel_gc_threads; j++) { 2337 q->enqueue(new StealMarkingTask(&terminator)); 2338 } 2339 } 2340 2341 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2342 q->enqueue(fin); 2343 2344 gc_task_manager()->add_list(q); 2345 2346 fin->wait_for(); 2347 2348 // We have to release the barrier tasks! 2349 WaitForBarrierGCTask::destroy(fin); 2350 } 2351 2352 // Process reference objects found during marking 2353 { 2354 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); 2355 if (ref_processor()->processing_is_mt()) { 2356 RefProcTaskExecutor task_executor; 2357 ref_processor()->process_discovered_references( 2358 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2359 &task_executor); 2360 } else { 2361 ref_processor()->process_discovered_references( 2362 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL); 2363 } 2364 } 2365 2366 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); 2367 // Follow system dictionary roots and unload classes. 2368 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2369 2370 // Follow code cache roots. 2371 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure, 2372 purged_class); 2373 follow_stack(cm); // Flush marking stack. 2374 2375 // Update subklass/sibling/implementor links of live klasses 2376 // revisit_klass_stack is used in follow_weak_klass_links(). 2377 follow_weak_klass_links(cm); 2378 2379 // Visit symbol and interned string tables and delete unmarked oops 2380 SymbolTable::unlink(is_alive_closure()); 2381 StringTable::unlink(is_alive_closure()); 2382 2383 assert(cm->marking_stack()->size() == 0, "stack should be empty by now"); 2384 assert(cm->overflow_stack()->is_empty(), "stack should be empty by now"); 2385 } 2386 2387 // This should be moved to the shared markSweep code! 2388 class PSAlwaysTrueClosure: public BoolObjectClosure { 2389 public: 2390 void do_object(oop p) { ShouldNotReachHere(); } 2391 bool do_object_b(oop p) { return true; } 2392 }; 2393 static PSAlwaysTrueClosure always_true; 2394 2395 void PSParallelCompact::adjust_roots() { 2396 // Adjust the pointers to reflect the new locations 2397 EventMark m("3 adjust roots"); 2398 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); 2399 2400 // General strong roots. 2401 Universe::oops_do(adjust_root_pointer_closure()); 2402 ReferenceProcessor::oops_do(adjust_root_pointer_closure()); 2403 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles 2404 Threads::oops_do(adjust_root_pointer_closure()); 2405 ObjectSynchronizer::oops_do(adjust_root_pointer_closure()); 2406 FlatProfiler::oops_do(adjust_root_pointer_closure()); 2407 Management::oops_do(adjust_root_pointer_closure()); 2408 JvmtiExport::oops_do(adjust_root_pointer_closure()); 2409 // SO_AllClasses 2410 SystemDictionary::oops_do(adjust_root_pointer_closure()); 2411 vmSymbols::oops_do(adjust_root_pointer_closure()); 2412 2413 // Now adjust pointers in remaining weak roots. (All of which should 2414 // have been cleared if they pointed to non-surviving objects.) 2415 // Global (weak) JNI handles 2416 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure()); 2417 2418 CodeCache::oops_do(adjust_pointer_closure()); 2419 SymbolTable::oops_do(adjust_root_pointer_closure()); 2420 StringTable::oops_do(adjust_root_pointer_closure()); 2421 ref_processor()->weak_oops_do(adjust_root_pointer_closure()); 2422 // Roots were visited so references into the young gen in roots 2423 // may have been scanned. Process them also. 2424 // Should the reference processor have a span that excludes 2425 // young gen objects? 2426 PSScavenge::reference_processor()->weak_oops_do( 2427 adjust_root_pointer_closure()); 2428 } 2429 2430 void PSParallelCompact::compact_perm(ParCompactionManager* cm) { 2431 EventMark m("4 compact perm"); 2432 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty); 2433 // trace("4"); 2434 2435 gc_heap()->perm_gen()->start_array()->reset(); 2436 move_and_update(cm, perm_space_id); 2437 } 2438 2439 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2440 uint parallel_gc_threads) 2441 { 2442 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); 2443 2444 const unsigned int task_count = MAX2(parallel_gc_threads, 1U); 2445 for (unsigned int j = 0; j < task_count; j++) { 2446 q->enqueue(new DrainStacksCompactionTask()); 2447 } 2448 2449 // Find all regions that are available (can be filled immediately) and 2450 // distribute them to the thread stacks. The iteration is done in reverse 2451 // order (high to low) so the regions will be removed in ascending order. 2452 2453 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2454 2455 size_t fillable_regions = 0; // A count for diagnostic purposes. 2456 unsigned int which = 0; // The worker thread number. 2457 2458 for (unsigned int id = to_space_id; id > perm_space_id; --id) { 2459 SpaceInfo* const space_info = _space_info + id; 2460 MutableSpace* const space = space_info->space(); 2461 HeapWord* const new_top = space_info->new_top(); 2462 2463 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2464 const size_t end_region = 2465 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2466 assert(end_region > 0, "perm gen cannot be empty"); 2467 2468 for (size_t cur = end_region - 1; cur >= beg_region; --cur) { 2469 if (sd.region(cur)->claim_unsafe()) { 2470 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2471 cm->save_for_processing(cur); 2472 2473 if (TraceParallelOldGCCompactionPhase && Verbose) { 2474 const size_t count_mod_8 = fillable_regions & 7; 2475 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2476 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2477 if (count_mod_8 == 7) gclog_or_tty->cr(); 2478 } 2479 2480 NOT_PRODUCT(++fillable_regions;) 2481 2482 // Assign regions to threads in round-robin fashion. 2483 if (++which == task_count) { 2484 which = 0; 2485 } 2486 } 2487 } 2488 } 2489 2490 if (TraceParallelOldGCCompactionPhase) { 2491 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2492 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2493 } 2494 } 2495 2496 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2497 2498 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2499 uint parallel_gc_threads) { 2500 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); 2501 2502 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2503 2504 // Iterate over all the spaces adding tasks for updating 2505 // regions in the dense prefix. Assume that 1 gc thread 2506 // will work on opening the gaps and the remaining gc threads 2507 // will work on the dense prefix. 2508 unsigned int space_id; 2509 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2510 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2511 const MutableSpace* const space = _space_info[space_id].space(); 2512 2513 if (dense_prefix_end == space->bottom()) { 2514 // There is no dense prefix for this space. 2515 continue; 2516 } 2517 2518 // The dense prefix is before this region. 2519 size_t region_index_end_dense_prefix = 2520 sd.addr_to_region_idx(dense_prefix_end); 2521 RegionData* const dense_prefix_cp = 2522 sd.region(region_index_end_dense_prefix); 2523 assert(dense_prefix_end == space->end() || 2524 dense_prefix_cp->available() || 2525 dense_prefix_cp->claimed(), 2526 "The region after the dense prefix should always be ready to fill"); 2527 2528 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2529 2530 // Is there dense prefix work? 2531 size_t total_dense_prefix_regions = 2532 region_index_end_dense_prefix - region_index_start; 2533 // How many regions of the dense prefix should be given to 2534 // each thread? 2535 if (total_dense_prefix_regions > 0) { 2536 uint tasks_for_dense_prefix = 1; 2537 if (UseParallelDensePrefixUpdate) { 2538 if (total_dense_prefix_regions <= 2539 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2540 // Don't over partition. This assumes that 2541 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2542 // so there are not many regions to process. 2543 tasks_for_dense_prefix = parallel_gc_threads; 2544 } else { 2545 // Over partition 2546 tasks_for_dense_prefix = parallel_gc_threads * 2547 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2548 } 2549 } 2550 size_t regions_per_thread = total_dense_prefix_regions / 2551 tasks_for_dense_prefix; 2552 // Give each thread at least 1 region. 2553 if (regions_per_thread == 0) { 2554 regions_per_thread = 1; 2555 } 2556 2557 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2558 if (region_index_start >= region_index_end_dense_prefix) { 2559 break; 2560 } 2561 // region_index_end is not processed 2562 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2563 region_index_end_dense_prefix); 2564 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2565 region_index_start, 2566 region_index_end)); 2567 region_index_start = region_index_end; 2568 } 2569 } 2570 // This gets any part of the dense prefix that did not 2571 // fit evenly. 2572 if (region_index_start < region_index_end_dense_prefix) { 2573 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2574 region_index_start, 2575 region_index_end_dense_prefix)); 2576 } 2577 } 2578 } 2579 2580 void PSParallelCompact::enqueue_region_stealing_tasks( 2581 GCTaskQueue* q, 2582 ParallelTaskTerminator* terminator_ptr, 2583 uint parallel_gc_threads) { 2584 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); 2585 2586 // Once a thread has drained it's stack, it should try to steal regions from 2587 // other threads. 2588 if (parallel_gc_threads > 1) { 2589 for (uint j = 0; j < parallel_gc_threads; j++) { 2590 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2591 } 2592 } 2593 } 2594 2595 void PSParallelCompact::compact() { 2596 EventMark m("5 compact"); 2597 // trace("5"); 2598 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); 2599 2600 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2601 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2602 PSOldGen* old_gen = heap->old_gen(); 2603 old_gen->start_array()->reset(); 2604 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2605 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2606 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2607 2608 GCTaskQueue* q = GCTaskQueue::create(); 2609 enqueue_region_draining_tasks(q, parallel_gc_threads); 2610 enqueue_dense_prefix_tasks(q, parallel_gc_threads); 2611 enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads); 2612 2613 { 2614 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); 2615 2616 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2617 q->enqueue(fin); 2618 2619 gc_task_manager()->add_list(q); 2620 2621 fin->wait_for(); 2622 2623 // We have to release the barrier tasks! 2624 WaitForBarrierGCTask::destroy(fin); 2625 2626 #ifdef ASSERT 2627 // Verify that all regions have been processed before the deferred updates. 2628 // Note that perm_space_id is skipped; this type of verification is not 2629 // valid until the perm gen is compacted by regions. 2630 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2631 verify_complete(SpaceId(id)); 2632 } 2633 #endif 2634 } 2635 2636 { 2637 // Update the deferred objects, if any. Any compaction manager can be used. 2638 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); 2639 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2640 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2641 update_deferred_objects(cm, SpaceId(id)); 2642 } 2643 } 2644 } 2645 2646 #ifdef ASSERT 2647 void PSParallelCompact::verify_complete(SpaceId space_id) { 2648 // All Regions between space bottom() to new_top() should be marked as filled 2649 // and all Regions between new_top() and top() should be available (i.e., 2650 // should have been emptied). 2651 ParallelCompactData& sd = summary_data(); 2652 SpaceInfo si = _space_info[space_id]; 2653 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2654 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2655 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2656 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2657 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2658 2659 bool issued_a_warning = false; 2660 2661 size_t cur_region; 2662 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2663 const RegionData* const c = sd.region(cur_region); 2664 if (!c->completed()) { 2665 warning("region " SIZE_FORMAT " not filled: " 2666 "destination_count=" SIZE_FORMAT, 2667 cur_region, c->destination_count()); 2668 issued_a_warning = true; 2669 } 2670 } 2671 2672 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2673 const RegionData* const c = sd.region(cur_region); 2674 if (!c->available()) { 2675 warning("region " SIZE_FORMAT " not empty: " 2676 "destination_count=" SIZE_FORMAT, 2677 cur_region, c->destination_count()); 2678 issued_a_warning = true; 2679 } 2680 } 2681 2682 if (issued_a_warning) { 2683 print_region_ranges(); 2684 } 2685 } 2686 #endif // #ifdef ASSERT 2687 2688 void PSParallelCompact::compact_serial(ParCompactionManager* cm) { 2689 EventMark m("5 compact serial"); 2690 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty); 2691 2692 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2693 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2694 2695 PSYoungGen* young_gen = heap->young_gen(); 2696 PSOldGen* old_gen = heap->old_gen(); 2697 2698 old_gen->start_array()->reset(); 2699 old_gen->move_and_update(cm); 2700 young_gen->move_and_update(cm); 2701 } 2702 2703 2704 void PSParallelCompact::follow_stack(ParCompactionManager* cm) { 2705 while(!cm->overflow_stack()->is_empty()) { 2706 oop obj = cm->overflow_stack()->pop(); 2707 obj->follow_contents(cm); 2708 } 2709 2710 oop obj; 2711 // obj is a reference!!! 2712 while (cm->marking_stack()->pop_local(obj)) { 2713 // It would be nice to assert about the type of objects we might 2714 // pop, but they can come from anywhere, unfortunately. 2715 obj->follow_contents(cm); 2716 } 2717 } 2718 2719 void 2720 PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) { 2721 // All klasses on the revisit stack are marked at this point. 2722 // Update and follow all subklass, sibling and implementor links. 2723 for (uint i = 0; i < ParallelGCThreads+1; i++) { 2724 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2725 KeepAliveClosure keep_alive_closure(cm); 2726 for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) { 2727 cm->revisit_klass_stack()->at(i)->follow_weak_klass_links( 2728 is_alive_closure(), 2729 &keep_alive_closure); 2730 } 2731 follow_stack(cm); 2732 } 2733 } 2734 2735 void 2736 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) { 2737 cm->revisit_klass_stack()->push(k); 2738 } 2739 2740 #ifdef VALIDATE_MARK_SWEEP 2741 2742 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) { 2743 if (!ValidateMarkSweep) 2744 return; 2745 2746 if (!isroot) { 2747 if (_pointer_tracking) { 2748 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer"); 2749 _adjusted_pointers->remove(p); 2750 } 2751 } else { 2752 ptrdiff_t index = _root_refs_stack->find(p); 2753 if (index != -1) { 2754 int l = _root_refs_stack->length(); 2755 if (l > 0 && l - 1 != index) { 2756 void* last = _root_refs_stack->pop(); 2757 assert(last != p, "should be different"); 2758 _root_refs_stack->at_put(index, last); 2759 } else { 2760 _root_refs_stack->remove(p); 2761 } 2762 } 2763 } 2764 } 2765 2766 2767 void PSParallelCompact::check_adjust_pointer(void* p) { 2768 _adjusted_pointers->push(p); 2769 } 2770 2771 2772 class AdjusterTracker: public OopClosure { 2773 public: 2774 AdjusterTracker() {}; 2775 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); } 2776 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); } 2777 }; 2778 2779 2780 void PSParallelCompact::track_interior_pointers(oop obj) { 2781 if (ValidateMarkSweep) { 2782 _adjusted_pointers->clear(); 2783 _pointer_tracking = true; 2784 2785 AdjusterTracker checker; 2786 obj->oop_iterate(&checker); 2787 } 2788 } 2789 2790 2791 void PSParallelCompact::check_interior_pointers() { 2792 if (ValidateMarkSweep) { 2793 _pointer_tracking = false; 2794 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers"); 2795 } 2796 } 2797 2798 2799 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) { 2800 if (ValidateMarkSweep) { 2801 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops"); 2802 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0; 2803 } 2804 } 2805 2806 2807 void PSParallelCompact::register_live_oop(oop p, size_t size) { 2808 if (ValidateMarkSweep) { 2809 _live_oops->push(p); 2810 _live_oops_size->push(size); 2811 _live_oops_index++; 2812 } 2813 } 2814 2815 void PSParallelCompact::validate_live_oop(oop p, size_t size) { 2816 if (ValidateMarkSweep) { 2817 oop obj = _live_oops->at((int)_live_oops_index); 2818 guarantee(obj == p, "should be the same object"); 2819 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size"); 2820 _live_oops_index++; 2821 } 2822 } 2823 2824 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size, 2825 HeapWord* compaction_top) { 2826 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top), 2827 "should be moved to forwarded location"); 2828 if (ValidateMarkSweep) { 2829 PSParallelCompact::validate_live_oop(oop(q), size); 2830 _live_oops_moved_to->push(oop(compaction_top)); 2831 } 2832 if (RecordMarkSweepCompaction) { 2833 _cur_gc_live_oops->push(q); 2834 _cur_gc_live_oops_moved_to->push(compaction_top); 2835 _cur_gc_live_oops_size->push(size); 2836 } 2837 } 2838 2839 2840 void PSParallelCompact::compaction_complete() { 2841 if (RecordMarkSweepCompaction) { 2842 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops; 2843 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to; 2844 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size; 2845 2846 _cur_gc_live_oops = _last_gc_live_oops; 2847 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to; 2848 _cur_gc_live_oops_size = _last_gc_live_oops_size; 2849 _last_gc_live_oops = _tmp_live_oops; 2850 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to; 2851 _last_gc_live_oops_size = _tmp_live_oops_size; 2852 } 2853 } 2854 2855 2856 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) { 2857 if (!RecordMarkSweepCompaction) { 2858 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled"); 2859 return; 2860 } 2861 2862 if (_last_gc_live_oops == NULL) { 2863 tty->print_cr("No compaction information gathered yet"); 2864 return; 2865 } 2866 2867 for (int i = 0; i < _last_gc_live_oops->length(); i++) { 2868 HeapWord* old_oop = _last_gc_live_oops->at(i); 2869 size_t sz = _last_gc_live_oops_size->at(i); 2870 if (old_oop <= q && q < (old_oop + sz)) { 2871 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i); 2872 size_t offset = (q - old_oop); 2873 tty->print_cr("Address " PTR_FORMAT, q); 2874 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset); 2875 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset); 2876 return; 2877 } 2878 } 2879 2880 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q); 2881 } 2882 #endif //VALIDATE_MARK_SWEEP 2883 2884 // Update interior oops in the ranges of regions [beg_region, end_region). 2885 void 2886 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2887 SpaceId space_id, 2888 size_t beg_region, 2889 size_t end_region) { 2890 ParallelCompactData& sd = summary_data(); 2891 ParMarkBitMap* const mbm = mark_bitmap(); 2892 2893 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2894 HeapWord* const end_addr = sd.region_to_addr(end_region); 2895 assert(beg_region <= end_region, "bad region range"); 2896 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2897 2898 #ifdef ASSERT 2899 // Claim the regions to avoid triggering an assert when they are marked as 2900 // filled. 2901 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2902 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2903 } 2904 #endif // #ifdef ASSERT 2905 2906 if (beg_addr != space(space_id)->bottom()) { 2907 // Find the first live object or block of dead space that *starts* in this 2908 // range of regions. If a partial object crosses onto the region, skip it; 2909 // it will be marked for 'deferred update' when the object head is 2910 // processed. If dead space crosses onto the region, it is also skipped; it 2911 // will be filled when the prior region is processed. If neither of those 2912 // apply, the first word in the region is the start of a live object or dead 2913 // space. 2914 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2915 const RegionData* const cp = sd.region(beg_region); 2916 if (cp->partial_obj_size() != 0) { 2917 beg_addr = sd.partial_obj_end(beg_region); 2918 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2919 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2920 } 2921 } 2922 2923 if (beg_addr < end_addr) { 2924 // A live object or block of dead space starts in this range of Regions. 2925 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2926 2927 // Create closures and iterate. 2928 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2929 FillClosure fill_closure(cm, space_id); 2930 ParMarkBitMap::IterationStatus status; 2931 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2932 dense_prefix_end); 2933 if (status == ParMarkBitMap::incomplete) { 2934 update_closure.do_addr(update_closure.source()); 2935 } 2936 } 2937 2938 // Mark the regions as filled. 2939 RegionData* const beg_cp = sd.region(beg_region); 2940 RegionData* const end_cp = sd.region(end_region); 2941 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2942 cp->set_completed(); 2943 } 2944 } 2945 2946 // Return the SpaceId for the space containing addr. If addr is not in the 2947 // heap, last_space_id is returned. In debug mode it expects the address to be 2948 // in the heap and asserts such. 2949 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2950 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2951 2952 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 2953 if (_space_info[id].space()->contains(addr)) { 2954 return SpaceId(id); 2955 } 2956 } 2957 2958 assert(false, "no space contains the addr"); 2959 return last_space_id; 2960 } 2961 2962 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 2963 SpaceId id) { 2964 assert(id < last_space_id, "bad space id"); 2965 2966 ParallelCompactData& sd = summary_data(); 2967 const SpaceInfo* const space_info = _space_info + id; 2968 ObjectStartArray* const start_array = space_info->start_array(); 2969 2970 const MutableSpace* const space = space_info->space(); 2971 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 2972 HeapWord* const beg_addr = space_info->dense_prefix(); 2973 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 2974 2975 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 2976 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 2977 const RegionData* cur_region; 2978 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 2979 HeapWord* const addr = cur_region->deferred_obj_addr(); 2980 if (addr != NULL) { 2981 if (start_array != NULL) { 2982 start_array->allocate_block(addr); 2983 } 2984 oop(addr)->update_contents(cm); 2985 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 2986 } 2987 } 2988 } 2989 2990 // Skip over count live words starting from beg, and return the address of the 2991 // next live word. Unless marked, the word corresponding to beg is assumed to 2992 // be dead. Callers must either ensure beg does not correspond to the middle of 2993 // an object, or account for those live words in some other way. Callers must 2994 // also ensure that there are enough live words in the range [beg, end) to skip. 2995 HeapWord* 2996 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 2997 { 2998 assert(count > 0, "sanity"); 2999 3000 ParMarkBitMap* m = mark_bitmap(); 3001 idx_t bits_to_skip = m->words_to_bits(count); 3002 idx_t cur_beg = m->addr_to_bit(beg); 3003 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 3004 3005 do { 3006 cur_beg = m->find_obj_beg(cur_beg, search_end); 3007 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 3008 const size_t obj_bits = cur_end - cur_beg + 1; 3009 if (obj_bits > bits_to_skip) { 3010 return m->bit_to_addr(cur_beg + bits_to_skip); 3011 } 3012 bits_to_skip -= obj_bits; 3013 cur_beg = cur_end + 1; 3014 } while (bits_to_skip > 0); 3015 3016 // Skipping the desired number of words landed just past the end of an object. 3017 // Find the start of the next object. 3018 cur_beg = m->find_obj_beg(cur_beg, search_end); 3019 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 3020 return m->bit_to_addr(cur_beg); 3021 } 3022 3023 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 3024 SpaceId src_space_id, 3025 size_t src_region_idx) 3026 { 3027 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 3028 3029 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 3030 if (split_info.dest_region_addr() == dest_addr) { 3031 // The partial object ending at the split point contains the first word to 3032 // be copied to dest_addr. 3033 return split_info.first_src_addr(); 3034 } 3035 3036 const ParallelCompactData& sd = summary_data(); 3037 ParMarkBitMap* const bitmap = mark_bitmap(); 3038 const size_t RegionSize = ParallelCompactData::RegionSize; 3039 3040 assert(sd.is_region_aligned(dest_addr), "not aligned"); 3041 const RegionData* const src_region_ptr = sd.region(src_region_idx); 3042 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 3043 HeapWord* const src_region_destination = src_region_ptr->destination(); 3044 3045 assert(dest_addr >= src_region_destination, "wrong src region"); 3046 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 3047 3048 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 3049 HeapWord* const src_region_end = src_region_beg + RegionSize; 3050 3051 HeapWord* addr = src_region_beg; 3052 if (dest_addr == src_region_destination) { 3053 // Return the first live word in the source region. 3054 if (partial_obj_size == 0) { 3055 addr = bitmap->find_obj_beg(addr, src_region_end); 3056 assert(addr < src_region_end, "no objects start in src region"); 3057 } 3058 return addr; 3059 } 3060 3061 // Must skip some live data. 3062 size_t words_to_skip = dest_addr - src_region_destination; 3063 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 3064 3065 if (partial_obj_size >= words_to_skip) { 3066 // All the live words to skip are part of the partial object. 3067 addr += words_to_skip; 3068 if (partial_obj_size == words_to_skip) { 3069 // Find the first live word past the partial object. 3070 addr = bitmap->find_obj_beg(addr, src_region_end); 3071 assert(addr < src_region_end, "wrong src region"); 3072 } 3073 return addr; 3074 } 3075 3076 // Skip over the partial object (if any). 3077 if (partial_obj_size != 0) { 3078 words_to_skip -= partial_obj_size; 3079 addr += partial_obj_size; 3080 } 3081 3082 // Skip over live words due to objects that start in the region. 3083 addr = skip_live_words(addr, src_region_end, words_to_skip); 3084 assert(addr < src_region_end, "wrong src region"); 3085 return addr; 3086 } 3087 3088 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 3089 SpaceId src_space_id, 3090 size_t beg_region, 3091 HeapWord* end_addr) 3092 { 3093 ParallelCompactData& sd = summary_data(); 3094 3095 #ifdef ASSERT 3096 MutableSpace* const src_space = _space_info[src_space_id].space(); 3097 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 3098 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 3099 "src_space_id does not match beg_addr"); 3100 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 3101 "src_space_id does not match end_addr"); 3102 #endif // #ifdef ASSERT 3103 3104 RegionData* const beg = sd.region(beg_region); 3105 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 3106 3107 // Regions up to new_top() are enqueued if they become available. 3108 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3109 RegionData* const enqueue_end = 3110 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3111 3112 for (RegionData* cur = beg; cur < end; ++cur) { 3113 assert(cur->data_size() > 0, "region must have live data"); 3114 cur->decrement_destination_count(); 3115 if (cur < enqueue_end && cur->available() && cur->claim()) { 3116 cm->save_for_processing(sd.region(cur)); 3117 } 3118 } 3119 } 3120 3121 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3122 SpaceId& src_space_id, 3123 HeapWord*& src_space_top, 3124 HeapWord* end_addr) 3125 { 3126 typedef ParallelCompactData::RegionData RegionData; 3127 3128 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3129 const size_t region_size = ParallelCompactData::RegionSize; 3130 3131 size_t src_region_idx = 0; 3132 3133 // Skip empty regions (if any) up to the top of the space. 3134 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3135 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3136 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3137 const RegionData* const top_region_ptr = 3138 sd.addr_to_region_ptr(top_aligned_up); 3139 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3140 ++src_region_ptr; 3141 } 3142 3143 if (src_region_ptr < top_region_ptr) { 3144 // The next source region is in the current space. Update src_region_idx 3145 // and the source address to match src_region_ptr. 3146 src_region_idx = sd.region(src_region_ptr); 3147 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3148 if (src_region_addr > closure.source()) { 3149 closure.set_source(src_region_addr); 3150 } 3151 return src_region_idx; 3152 } 3153 3154 // Switch to a new source space and find the first non-empty region. 3155 unsigned int space_id = src_space_id + 1; 3156 assert(space_id < last_space_id, "not enough spaces"); 3157 3158 HeapWord* const destination = closure.destination(); 3159 3160 do { 3161 MutableSpace* space = _space_info[space_id].space(); 3162 HeapWord* const bottom = space->bottom(); 3163 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3164 3165 // Iterate over the spaces that do not compact into themselves. 3166 if (bottom_cp->destination() != bottom) { 3167 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3168 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3169 3170 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3171 if (src_cp->live_obj_size() > 0) { 3172 // Found it. 3173 assert(src_cp->destination() == destination, 3174 "first live obj in the space must match the destination"); 3175 assert(src_cp->partial_obj_size() == 0, 3176 "a space cannot begin with a partial obj"); 3177 3178 src_space_id = SpaceId(space_id); 3179 src_space_top = space->top(); 3180 const size_t src_region_idx = sd.region(src_cp); 3181 closure.set_source(sd.region_to_addr(src_region_idx)); 3182 return src_region_idx; 3183 } else { 3184 assert(src_cp->data_size() == 0, "sanity"); 3185 } 3186 } 3187 } 3188 } while (++space_id < last_space_id); 3189 3190 assert(false, "no source region was found"); 3191 return 0; 3192 } 3193 3194 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3195 { 3196 typedef ParMarkBitMap::IterationStatus IterationStatus; 3197 const size_t RegionSize = ParallelCompactData::RegionSize; 3198 ParMarkBitMap* const bitmap = mark_bitmap(); 3199 ParallelCompactData& sd = summary_data(); 3200 RegionData* const region_ptr = sd.region(region_idx); 3201 3202 // Get the items needed to construct the closure. 3203 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3204 SpaceId dest_space_id = space_id(dest_addr); 3205 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3206 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3207 assert(dest_addr < new_top, "sanity"); 3208 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3209 3210 // Get the source region and related info. 3211 size_t src_region_idx = region_ptr->source_region(); 3212 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3213 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3214 3215 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3216 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3217 3218 // Adjust src_region_idx to prepare for decrementing destination counts (the 3219 // destination count is not decremented when a region is copied to itself). 3220 if (src_region_idx == region_idx) { 3221 src_region_idx += 1; 3222 } 3223 3224 if (bitmap->is_unmarked(closure.source())) { 3225 // The first source word is in the middle of an object; copy the remainder 3226 // of the object or as much as will fit. The fact that pointer updates were 3227 // deferred will be noted when the object header is processed. 3228 HeapWord* const old_src_addr = closure.source(); 3229 closure.copy_partial_obj(); 3230 if (closure.is_full()) { 3231 decrement_destination_counts(cm, src_space_id, src_region_idx, 3232 closure.source()); 3233 region_ptr->set_deferred_obj_addr(NULL); 3234 region_ptr->set_completed(); 3235 return; 3236 } 3237 3238 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3239 if (sd.region_align_down(old_src_addr) != end_addr) { 3240 // The partial object was copied from more than one source region. 3241 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3242 3243 // Move to the next source region, possibly switching spaces as well. All 3244 // args except end_addr may be modified. 3245 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3246 end_addr); 3247 } 3248 } 3249 3250 do { 3251 HeapWord* const cur_addr = closure.source(); 3252 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3253 src_space_top); 3254 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3255 3256 if (status == ParMarkBitMap::incomplete) { 3257 // The last obj that starts in the source region does not end in the 3258 // region. 3259 assert(closure.source() < end_addr, "sanity") 3260 HeapWord* const obj_beg = closure.source(); 3261 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3262 src_space_top); 3263 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3264 if (obj_end < range_end) { 3265 // The end was found; the entire object will fit. 3266 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3267 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3268 } else { 3269 // The end was not found; the object will not fit. 3270 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3271 status = ParMarkBitMap::would_overflow; 3272 } 3273 } 3274 3275 if (status == ParMarkBitMap::would_overflow) { 3276 // The last object did not fit. Note that interior oop updates were 3277 // deferred, then copy enough of the object to fill the region. 3278 region_ptr->set_deferred_obj_addr(closure.destination()); 3279 status = closure.copy_until_full(); // copies from closure.source() 3280 3281 decrement_destination_counts(cm, src_space_id, src_region_idx, 3282 closure.source()); 3283 region_ptr->set_completed(); 3284 return; 3285 } 3286 3287 if (status == ParMarkBitMap::full) { 3288 decrement_destination_counts(cm, src_space_id, src_region_idx, 3289 closure.source()); 3290 region_ptr->set_deferred_obj_addr(NULL); 3291 region_ptr->set_completed(); 3292 return; 3293 } 3294 3295 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3296 3297 // Move to the next source region, possibly switching spaces as well. All 3298 // args except end_addr may be modified. 3299 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3300 end_addr); 3301 } while (true); 3302 } 3303 3304 void 3305 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3306 const MutableSpace* sp = space(space_id); 3307 if (sp->is_empty()) { 3308 return; 3309 } 3310 3311 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3312 ParMarkBitMap* const bitmap = mark_bitmap(); 3313 HeapWord* const dp_addr = dense_prefix(space_id); 3314 HeapWord* beg_addr = sp->bottom(); 3315 HeapWord* end_addr = sp->top(); 3316 3317 #ifdef ASSERT 3318 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3319 if (cm->should_verify_only()) { 3320 VerifyUpdateClosure verify_update(cm, sp); 3321 bitmap->iterate(&verify_update, beg_addr, end_addr); 3322 return; 3323 } 3324 3325 if (cm->should_reset_only()) { 3326 ResetObjectsClosure reset_objects(cm); 3327 bitmap->iterate(&reset_objects, beg_addr, end_addr); 3328 return; 3329 } 3330 #endif 3331 3332 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3333 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3334 if (beg_region < dp_region) { 3335 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3336 } 3337 3338 // The destination of the first live object that starts in the region is one 3339 // past the end of the partial object entering the region (if any). 3340 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3341 HeapWord* const new_top = _space_info[space_id].new_top(); 3342 assert(new_top >= dest_addr, "bad new_top value"); 3343 const size_t words = pointer_delta(new_top, dest_addr); 3344 3345 if (words > 0) { 3346 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3347 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3348 3349 ParMarkBitMap::IterationStatus status; 3350 status = bitmap->iterate(&closure, dest_addr, end_addr); 3351 assert(status == ParMarkBitMap::full, "iteration not complete"); 3352 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3353 "live objects skipped because closure is full"); 3354 } 3355 } 3356 3357 jlong PSParallelCompact::millis_since_last_gc() { 3358 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc; 3359 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3360 if (ret_val < 0) { 3361 NOT_PRODUCT(warning("time warp: %d", ret_val);) 3362 return 0; 3363 } 3364 return ret_val; 3365 } 3366 3367 void PSParallelCompact::reset_millis_since_last_gc() { 3368 _time_of_last_gc = os::javaTimeMillis(); 3369 } 3370 3371 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3372 { 3373 if (source() != destination()) { 3374 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3375 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3376 } 3377 update_state(words_remaining()); 3378 assert(is_full(), "sanity"); 3379 return ParMarkBitMap::full; 3380 } 3381 3382 void MoveAndUpdateClosure::copy_partial_obj() 3383 { 3384 size_t words = words_remaining(); 3385 3386 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3387 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3388 if (end_addr < range_end) { 3389 words = bitmap()->obj_size(source(), end_addr); 3390 } 3391 3392 // This test is necessary; if omitted, the pointer updates to a partial object 3393 // that crosses the dense prefix boundary could be overwritten. 3394 if (source() != destination()) { 3395 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3396 Copy::aligned_conjoint_words(source(), destination(), words); 3397 } 3398 update_state(words); 3399 } 3400 3401 ParMarkBitMapClosure::IterationStatus 3402 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3403 assert(destination() != NULL, "sanity"); 3404 assert(bitmap()->obj_size(addr) == words, "bad size"); 3405 3406 _source = addr; 3407 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3408 destination(), "wrong destination"); 3409 3410 if (words > words_remaining()) { 3411 return ParMarkBitMap::would_overflow; 3412 } 3413 3414 // The start_array must be updated even if the object is not moving. 3415 if (_start_array != NULL) { 3416 _start_array->allocate_block(destination()); 3417 } 3418 3419 if (destination() != source()) { 3420 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3421 Copy::aligned_conjoint_words(source(), destination(), words); 3422 } 3423 3424 oop moved_oop = (oop) destination(); 3425 moved_oop->update_contents(compaction_manager()); 3426 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3427 3428 update_state(words); 3429 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3430 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3431 } 3432 3433 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3434 ParCompactionManager* cm, 3435 PSParallelCompact::SpaceId space_id) : 3436 ParMarkBitMapClosure(mbm, cm), 3437 _space_id(space_id), 3438 _start_array(PSParallelCompact::start_array(space_id)) 3439 { 3440 } 3441 3442 // Updates the references in the object to their new values. 3443 ParMarkBitMapClosure::IterationStatus 3444 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3445 do_addr(addr); 3446 return ParMarkBitMap::incomplete; 3447 } 3448 3449 // Verify the new location using the forwarding pointer 3450 // from MarkSweep::mark_sweep_phase2(). Set the mark_word 3451 // to the initial value. 3452 ParMarkBitMapClosure::IterationStatus 3453 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3454 // The second arg (words) is not used. 3455 oop obj = (oop) addr; 3456 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer(); 3457 HeapWord* new_pointer = summary_data().calc_new_pointer(obj); 3458 if (forwarding_ptr == NULL) { 3459 // The object is dead or not moving. 3460 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj), 3461 "Object liveness is wrong."); 3462 return ParMarkBitMap::incomplete; 3463 } 3464 assert(UseParallelOldGCDensePrefix || 3465 (HeapMaximumCompactionInterval > 1) || 3466 (MarkSweepAlwaysCompactCount > 1) || 3467 (forwarding_ptr == new_pointer), 3468 "Calculation of new location is incorrect"); 3469 return ParMarkBitMap::incomplete; 3470 } 3471 3472 // Reset objects modified for debug checking. 3473 ParMarkBitMapClosure::IterationStatus 3474 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) { 3475 // The second arg (words) is not used. 3476 oop obj = (oop) addr; 3477 obj->init_mark(); 3478 return ParMarkBitMap::incomplete; 3479 } 3480 3481 // Prepare for compaction. This method is executed once 3482 // (i.e., by a single thread) before compaction. 3483 // Save the updated location of the intArrayKlassObj for 3484 // filling holes in the dense prefix. 3485 void PSParallelCompact::compact_prologue() { 3486 _updated_int_array_klass_obj = (klassOop) 3487 summary_data().calc_new_pointer(Universe::intArrayKlassObj()); 3488 } 3489