1 /* 2 * Copyright 2005-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_psParallelCompact.cpp.incl" 27 28 #include <math.h> 29 30 // All sizes are in HeapWords. 31 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words 32 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize; 33 const size_t ParallelCompactData::RegionSizeBytes = 34 RegionSize << LogHeapWordSize; 35 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1; 36 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1; 37 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask; 38 39 const ParallelCompactData::RegionData::region_sz_t 40 ParallelCompactData::RegionData::dc_shift = 27; 41 42 const ParallelCompactData::RegionData::region_sz_t 43 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift; 44 45 const ParallelCompactData::RegionData::region_sz_t 46 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift; 47 48 const ParallelCompactData::RegionData::region_sz_t 49 ParallelCompactData::RegionData::los_mask = ~dc_mask; 50 51 const ParallelCompactData::RegionData::region_sz_t 52 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift; 53 54 const ParallelCompactData::RegionData::region_sz_t 55 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift; 56 57 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id]; 58 bool PSParallelCompact::_print_phases = false; 59 60 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL; 61 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL; 62 63 double PSParallelCompact::_dwl_mean; 64 double PSParallelCompact::_dwl_std_dev; 65 double PSParallelCompact::_dwl_first_term; 66 double PSParallelCompact::_dwl_adjustment; 67 #ifdef ASSERT 68 bool PSParallelCompact::_dwl_initialized = false; 69 #endif // #ifdef ASSERT 70 71 #ifdef VALIDATE_MARK_SWEEP 72 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL; 73 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL; 74 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL; 75 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL; 76 size_t PSParallelCompact::_live_oops_index = 0; 77 size_t PSParallelCompact::_live_oops_index_at_perm = 0; 78 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL; 79 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL; 80 bool PSParallelCompact::_pointer_tracking = false; 81 bool PSParallelCompact::_root_tracking = true; 82 83 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL; 84 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL; 85 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL; 86 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL; 87 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL; 88 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL; 89 #endif 90 91 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size, 92 HeapWord* destination) 93 { 94 assert(src_region_idx != 0, "invalid src_region_idx"); 95 assert(partial_obj_size != 0, "invalid partial_obj_size argument"); 96 assert(destination != NULL, "invalid destination argument"); 97 98 _src_region_idx = src_region_idx; 99 _partial_obj_size = partial_obj_size; 100 _destination = destination; 101 102 // These fields may not be updated below, so make sure they're clear. 103 assert(_dest_region_addr == NULL, "should have been cleared"); 104 assert(_first_src_addr == NULL, "should have been cleared"); 105 106 // Determine the number of destination regions for the partial object. 107 HeapWord* const last_word = destination + partial_obj_size - 1; 108 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 109 HeapWord* const beg_region_addr = sd.region_align_down(destination); 110 HeapWord* const end_region_addr = sd.region_align_down(last_word); 111 112 if (beg_region_addr == end_region_addr) { 113 // One destination region. 114 _destination_count = 1; 115 if (end_region_addr == destination) { 116 // The destination falls on a region boundary, thus the first word of the 117 // partial object will be the first word copied to the destination region. 118 _dest_region_addr = end_region_addr; 119 _first_src_addr = sd.region_to_addr(src_region_idx); 120 } 121 } else { 122 // Two destination regions. When copied, the partial object will cross a 123 // destination region boundary, so a word somewhere within the partial 124 // object will be the first word copied to the second destination region. 125 _destination_count = 2; 126 _dest_region_addr = end_region_addr; 127 const size_t ofs = pointer_delta(end_region_addr, destination); 128 assert(ofs < _partial_obj_size, "sanity"); 129 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs; 130 } 131 } 132 133 void SplitInfo::clear() 134 { 135 _src_region_idx = 0; 136 _partial_obj_size = 0; 137 _destination = NULL; 138 _destination_count = 0; 139 _dest_region_addr = NULL; 140 _first_src_addr = NULL; 141 assert(!is_valid(), "sanity"); 142 } 143 144 #ifdef ASSERT 145 void SplitInfo::verify_clear() 146 { 147 assert(_src_region_idx == 0, "not clear"); 148 assert(_partial_obj_size == 0, "not clear"); 149 assert(_destination == NULL, "not clear"); 150 assert(_destination_count == 0, "not clear"); 151 assert(_dest_region_addr == NULL, "not clear"); 152 assert(_first_src_addr == NULL, "not clear"); 153 } 154 #endif // #ifdef ASSERT 155 156 157 #ifndef PRODUCT 158 const char* PSParallelCompact::space_names[] = { 159 "perm", "old ", "eden", "from", "to " 160 }; 161 162 void PSParallelCompact::print_region_ranges() 163 { 164 tty->print_cr("space bottom top end new_top"); 165 tty->print_cr("------ ---------- ---------- ---------- ----------"); 166 167 for (unsigned int id = 0; id < last_space_id; ++id) { 168 const MutableSpace* space = _space_info[id].space(); 169 tty->print_cr("%u %s " 170 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " " 171 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ", 172 id, space_names[id], 173 summary_data().addr_to_region_idx(space->bottom()), 174 summary_data().addr_to_region_idx(space->top()), 175 summary_data().addr_to_region_idx(space->end()), 176 summary_data().addr_to_region_idx(_space_info[id].new_top())); 177 } 178 } 179 180 void 181 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c) 182 { 183 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7) 184 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5) 185 186 ParallelCompactData& sd = PSParallelCompact::summary_data(); 187 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0; 188 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " " 189 REGION_IDX_FORMAT " " PTR_FORMAT " " 190 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " " 191 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d", 192 i, c->data_location(), dci, c->destination(), 193 c->partial_obj_size(), c->live_obj_size(), 194 c->data_size(), c->source_region(), c->destination_count()); 195 196 #undef REGION_IDX_FORMAT 197 #undef REGION_DATA_FORMAT 198 } 199 200 void 201 print_generic_summary_data(ParallelCompactData& summary_data, 202 HeapWord* const beg_addr, 203 HeapWord* const end_addr) 204 { 205 size_t total_words = 0; 206 size_t i = summary_data.addr_to_region_idx(beg_addr); 207 const size_t last = summary_data.addr_to_region_idx(end_addr); 208 HeapWord* pdest = 0; 209 210 while (i <= last) { 211 ParallelCompactData::RegionData* c = summary_data.region(i); 212 if (c->data_size() != 0 || c->destination() != pdest) { 213 print_generic_summary_region(i, c); 214 total_words += c->data_size(); 215 pdest = c->destination(); 216 } 217 ++i; 218 } 219 220 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize); 221 } 222 223 void 224 print_generic_summary_data(ParallelCompactData& summary_data, 225 SpaceInfo* space_info) 226 { 227 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) { 228 const MutableSpace* space = space_info[id].space(); 229 print_generic_summary_data(summary_data, space->bottom(), 230 MAX2(space->top(), space_info[id].new_top())); 231 } 232 } 233 234 void 235 print_initial_summary_region(size_t i, 236 const ParallelCompactData::RegionData* c, 237 bool newline = true) 238 { 239 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " " 240 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " 241 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d", 242 i, c->destination(), 243 c->partial_obj_size(), c->live_obj_size(), 244 c->data_size(), c->source_region(), c->destination_count()); 245 if (newline) tty->cr(); 246 } 247 248 void 249 print_initial_summary_data(ParallelCompactData& summary_data, 250 const MutableSpace* space) { 251 if (space->top() == space->bottom()) { 252 return; 253 } 254 255 const size_t region_size = ParallelCompactData::RegionSize; 256 typedef ParallelCompactData::RegionData RegionData; 257 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top()); 258 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up); 259 const RegionData* c = summary_data.region(end_region - 1); 260 HeapWord* end_addr = c->destination() + c->data_size(); 261 const size_t live_in_space = pointer_delta(end_addr, space->bottom()); 262 263 // Print (and count) the full regions at the beginning of the space. 264 size_t full_region_count = 0; 265 size_t i = summary_data.addr_to_region_idx(space->bottom()); 266 while (i < end_region && summary_data.region(i)->data_size() == region_size) { 267 print_initial_summary_region(i, summary_data.region(i)); 268 ++full_region_count; 269 ++i; 270 } 271 272 size_t live_to_right = live_in_space - full_region_count * region_size; 273 274 double max_reclaimed_ratio = 0.0; 275 size_t max_reclaimed_ratio_region = 0; 276 size_t max_dead_to_right = 0; 277 size_t max_live_to_right = 0; 278 279 // Print the 'reclaimed ratio' for regions while there is something live in 280 // the region or to the right of it. The remaining regions are empty (and 281 // uninteresting), and computing the ratio will result in division by 0. 282 while (i < end_region && live_to_right > 0) { 283 c = summary_data.region(i); 284 HeapWord* const region_addr = summary_data.region_to_addr(i); 285 const size_t used_to_right = pointer_delta(space->top(), region_addr); 286 const size_t dead_to_right = used_to_right - live_to_right; 287 const double reclaimed_ratio = double(dead_to_right) / live_to_right; 288 289 if (reclaimed_ratio > max_reclaimed_ratio) { 290 max_reclaimed_ratio = reclaimed_ratio; 291 max_reclaimed_ratio_region = i; 292 max_dead_to_right = dead_to_right; 293 max_live_to_right = live_to_right; 294 } 295 296 print_initial_summary_region(i, c, false); 297 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10), 298 reclaimed_ratio, dead_to_right, live_to_right); 299 300 live_to_right -= c->data_size(); 301 ++i; 302 } 303 304 // Any remaining regions are empty. Print one more if there is one. 305 if (i < end_region) { 306 print_initial_summary_region(i, summary_data.region(i)); 307 } 308 309 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " " 310 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f", 311 max_reclaimed_ratio_region, max_dead_to_right, 312 max_live_to_right, max_reclaimed_ratio); 313 } 314 315 void 316 print_initial_summary_data(ParallelCompactData& summary_data, 317 SpaceInfo* space_info) { 318 unsigned int id = PSParallelCompact::perm_space_id; 319 const MutableSpace* space; 320 do { 321 space = space_info[id].space(); 322 print_initial_summary_data(summary_data, space); 323 } while (++id < PSParallelCompact::eden_space_id); 324 325 do { 326 space = space_info[id].space(); 327 print_generic_summary_data(summary_data, space->bottom(), space->top()); 328 } while (++id < PSParallelCompact::last_space_id); 329 } 330 #endif // #ifndef PRODUCT 331 332 #ifdef ASSERT 333 size_t add_obj_count; 334 size_t add_obj_size; 335 size_t mark_bitmap_count; 336 size_t mark_bitmap_size; 337 #endif // #ifdef ASSERT 338 339 ParallelCompactData::ParallelCompactData() 340 { 341 _region_start = 0; 342 343 _region_vspace = 0; 344 _region_data = 0; 345 _region_count = 0; 346 } 347 348 bool ParallelCompactData::initialize(MemRegion covered_region) 349 { 350 _region_start = covered_region.start(); 351 const size_t region_size = covered_region.word_size(); 352 DEBUG_ONLY(_region_end = _region_start + region_size;) 353 354 assert(region_align_down(_region_start) == _region_start, 355 "region start not aligned"); 356 assert((region_size & RegionSizeOffsetMask) == 0, 357 "region size not a multiple of RegionSize"); 358 359 bool result = initialize_region_data(region_size); 360 361 return result; 362 } 363 364 PSVirtualSpace* 365 ParallelCompactData::create_vspace(size_t count, size_t element_size) 366 { 367 const size_t raw_bytes = count * element_size; 368 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10); 369 const size_t granularity = os::vm_allocation_granularity(); 370 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity)); 371 372 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 : 373 MAX2(page_sz, granularity); 374 ReservedSpace rs(bytes, rs_align, rs_align > 0); 375 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(), 376 rs.size()); 377 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz); 378 if (vspace != 0) { 379 if (vspace->expand_by(bytes)) { 380 return vspace; 381 } 382 delete vspace; 383 // Release memory reserved in the space. 384 rs.release(); 385 } 386 387 return 0; 388 } 389 390 bool ParallelCompactData::initialize_region_data(size_t region_size) 391 { 392 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize; 393 _region_vspace = create_vspace(count, sizeof(RegionData)); 394 if (_region_vspace != 0) { 395 _region_data = (RegionData*)_region_vspace->reserved_low_addr(); 396 _region_count = count; 397 return true; 398 } 399 return false; 400 } 401 402 void ParallelCompactData::clear() 403 { 404 memset(_region_data, 0, _region_vspace->committed_size()); 405 } 406 407 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) { 408 assert(beg_region <= _region_count, "beg_region out of range"); 409 assert(end_region <= _region_count, "end_region out of range"); 410 411 const size_t region_cnt = end_region - beg_region; 412 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData)); 413 } 414 415 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const 416 { 417 const RegionData* cur_cp = region(region_idx); 418 const RegionData* const end_cp = region(region_count() - 1); 419 420 HeapWord* result = region_to_addr(region_idx); 421 if (cur_cp < end_cp) { 422 do { 423 result += cur_cp->partial_obj_size(); 424 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp); 425 } 426 return result; 427 } 428 429 void ParallelCompactData::add_obj(HeapWord* addr, size_t len) 430 { 431 const size_t obj_ofs = pointer_delta(addr, _region_start); 432 const size_t beg_region = obj_ofs >> Log2RegionSize; 433 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize; 434 435 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);) 436 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);) 437 438 if (beg_region == end_region) { 439 // All in one region. 440 _region_data[beg_region].add_live_obj(len); 441 return; 442 } 443 444 // First region. 445 const size_t beg_ofs = region_offset(addr); 446 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs); 447 448 klassOop klass = ((oop)addr)->klass(); 449 // Middle regions--completely spanned by this object. 450 for (size_t region = beg_region + 1; region < end_region; ++region) { 451 _region_data[region].set_partial_obj_size(RegionSize); 452 _region_data[region].set_partial_obj_addr(addr); 453 } 454 455 // Last region. 456 const size_t end_ofs = region_offset(addr + len - 1); 457 _region_data[end_region].set_partial_obj_size(end_ofs + 1); 458 _region_data[end_region].set_partial_obj_addr(addr); 459 } 460 461 void 462 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end) 463 { 464 assert(region_offset(beg) == 0, "not RegionSize aligned"); 465 assert(region_offset(end) == 0, "not RegionSize aligned"); 466 467 size_t cur_region = addr_to_region_idx(beg); 468 const size_t end_region = addr_to_region_idx(end); 469 HeapWord* addr = beg; 470 while (cur_region < end_region) { 471 _region_data[cur_region].set_destination(addr); 472 _region_data[cur_region].set_destination_count(0); 473 _region_data[cur_region].set_source_region(cur_region); 474 _region_data[cur_region].set_data_location(addr); 475 476 // Update live_obj_size so the region appears completely full. 477 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size(); 478 _region_data[cur_region].set_live_obj_size(live_size); 479 480 ++cur_region; 481 addr += RegionSize; 482 } 483 } 484 485 // Find the point at which a space can be split and, if necessary, record the 486 // split point. 487 // 488 // If the current src region (which overflowed the destination space) doesn't 489 // have a partial object, the split point is at the beginning of the current src 490 // region (an "easy" split, no extra bookkeeping required). 491 // 492 // If the current src region has a partial object, the split point is in the 493 // region where that partial object starts (call it the split_region). If 494 // split_region has a partial object, then the split point is just after that 495 // partial object (a "hard" split where we have to record the split data and 496 // zero the partial_obj_size field). With a "hard" split, we know that the 497 // partial_obj ends within split_region because the partial object that caused 498 // the overflow starts in split_region. If split_region doesn't have a partial 499 // obj, then the split is at the beginning of split_region (another "easy" 500 // split). 501 HeapWord* 502 ParallelCompactData::summarize_split_space(size_t src_region, 503 SplitInfo& split_info, 504 HeapWord* destination, 505 HeapWord* target_end, 506 HeapWord** target_next) 507 { 508 assert(destination <= target_end, "sanity"); 509 assert(destination + _region_data[src_region].data_size() > target_end, 510 "region should not fit into target space"); 511 assert(is_region_aligned(target_end), "sanity"); 512 513 size_t split_region = src_region; 514 HeapWord* split_destination = destination; 515 size_t partial_obj_size = _region_data[src_region].partial_obj_size(); 516 517 if (destination + partial_obj_size > target_end) { 518 // The split point is just after the partial object (if any) in the 519 // src_region that contains the start of the object that overflowed the 520 // destination space. 521 // 522 // Find the start of the "overflow" object and set split_region to the 523 // region containing it. 524 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr(); 525 split_region = addr_to_region_idx(overflow_obj); 526 527 // Clear the source_region field of all destination regions whose first word 528 // came from data after the split point (a non-null source_region field 529 // implies a region must be filled). 530 // 531 // An alternative to the simple loop below: clear during post_compact(), 532 // which uses memcpy instead of individual stores, and is easy to 533 // parallelize. (The downside is that it clears the entire RegionData 534 // object as opposed to just one field.) 535 // 536 // post_compact() would have to clear the summary data up to the highest 537 // address that was written during the summary phase, which would be 538 // 539 // max(top, max(new_top, clear_top)) 540 // 541 // where clear_top is a new field in SpaceInfo. Would have to set clear_top 542 // to target_end. 543 const RegionData* const sr = region(split_region); 544 const size_t beg_idx = 545 addr_to_region_idx(region_align_up(sr->destination() + 546 sr->partial_obj_size())); 547 const size_t end_idx = addr_to_region_idx(target_end); 548 549 if (TraceParallelOldGCSummaryPhase) { 550 gclog_or_tty->print_cr("split: clearing source_region field in [" 551 SIZE_FORMAT ", " SIZE_FORMAT ")", 552 beg_idx, end_idx); 553 } 554 for (size_t idx = beg_idx; idx < end_idx; ++idx) { 555 _region_data[idx].set_source_region(0); 556 } 557 558 // Set split_destination and partial_obj_size to reflect the split region. 559 split_destination = sr->destination(); 560 partial_obj_size = sr->partial_obj_size(); 561 } 562 563 // The split is recorded only if a partial object extends onto the region. 564 if (partial_obj_size != 0) { 565 _region_data[split_region].set_partial_obj_size(0); 566 split_info.record(split_region, partial_obj_size, split_destination); 567 } 568 569 // Setup the continuation addresses. 570 *target_next = split_destination + partial_obj_size; 571 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size; 572 573 if (TraceParallelOldGCSummaryPhase) { 574 const char * split_type = partial_obj_size == 0 ? "easy" : "hard"; 575 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT 576 " pos=" SIZE_FORMAT, 577 split_type, source_next, split_region, 578 partial_obj_size); 579 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT 580 " tn=" PTR_FORMAT, 581 split_type, split_destination, 582 addr_to_region_idx(split_destination), 583 *target_next); 584 585 if (partial_obj_size != 0) { 586 HeapWord* const po_beg = split_info.destination(); 587 HeapWord* const po_end = po_beg + split_info.partial_obj_size(); 588 gclog_or_tty->print_cr("%s split: " 589 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " " 590 "po_end=" PTR_FORMAT " " SIZE_FORMAT, 591 split_type, 592 po_beg, addr_to_region_idx(po_beg), 593 po_end, addr_to_region_idx(po_end)); 594 } 595 } 596 597 return source_next; 598 } 599 600 bool ParallelCompactData::summarize(SplitInfo& split_info, 601 HeapWord* source_beg, HeapWord* source_end, 602 HeapWord** source_next, 603 HeapWord* target_beg, HeapWord* target_end, 604 HeapWord** target_next) 605 { 606 if (TraceParallelOldGCSummaryPhase) { 607 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next; 608 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT 609 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT, 610 source_beg, source_end, source_next_val, 611 target_beg, target_end, *target_next); 612 } 613 614 size_t cur_region = addr_to_region_idx(source_beg); 615 const size_t end_region = addr_to_region_idx(region_align_up(source_end)); 616 617 HeapWord *dest_addr = target_beg; 618 while (cur_region < end_region) { 619 // The destination must be set even if the region has no data. 620 _region_data[cur_region].set_destination(dest_addr); 621 622 size_t words = _region_data[cur_region].data_size(); 623 if (words > 0) { 624 // If cur_region does not fit entirely into the target space, find a point 625 // at which the source space can be 'split' so that part is copied to the 626 // target space and the rest is copied elsewhere. 627 if (dest_addr + words > target_end) { 628 assert(source_next != NULL, "source_next is NULL when splitting"); 629 *source_next = summarize_split_space(cur_region, split_info, dest_addr, 630 target_end, target_next); 631 return false; 632 } 633 634 // Compute the destination_count for cur_region, and if necessary, update 635 // source_region for a destination region. The source_region field is 636 // updated if cur_region is the first (left-most) region to be copied to a 637 // destination region. 638 // 639 // The destination_count calculation is a bit subtle. A region that has 640 // data that compacts into itself does not count itself as a destination. 641 // This maintains the invariant that a zero count means the region is 642 // available and can be claimed and then filled. 643 uint destination_count = 0; 644 if (split_info.is_split(cur_region)) { 645 // The current region has been split: the partial object will be copied 646 // to one destination space and the remaining data will be copied to 647 // another destination space. Adjust the initial destination_count and, 648 // if necessary, set the source_region field if the partial object will 649 // cross a destination region boundary. 650 destination_count = split_info.destination_count(); 651 if (destination_count == 2) { 652 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr()); 653 _region_data[dest_idx].set_source_region(cur_region); 654 } 655 } 656 657 HeapWord* const last_addr = dest_addr + words - 1; 658 const size_t dest_region_1 = addr_to_region_idx(dest_addr); 659 const size_t dest_region_2 = addr_to_region_idx(last_addr); 660 661 // Initially assume that the destination regions will be the same and 662 // adjust the value below if necessary. Under this assumption, if 663 // cur_region == dest_region_2, then cur_region will be compacted 664 // completely into itself. 665 destination_count += cur_region == dest_region_2 ? 0 : 1; 666 if (dest_region_1 != dest_region_2) { 667 // Destination regions differ; adjust destination_count. 668 destination_count += 1; 669 // Data from cur_region will be copied to the start of dest_region_2. 670 _region_data[dest_region_2].set_source_region(cur_region); 671 } else if (region_offset(dest_addr) == 0) { 672 // Data from cur_region will be copied to the start of the destination 673 // region. 674 _region_data[dest_region_1].set_source_region(cur_region); 675 } 676 677 _region_data[cur_region].set_destination_count(destination_count); 678 _region_data[cur_region].set_data_location(region_to_addr(cur_region)); 679 dest_addr += words; 680 } 681 682 ++cur_region; 683 } 684 685 *target_next = dest_addr; 686 return true; 687 } 688 689 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) { 690 assert(addr != NULL, "Should detect NULL oop earlier"); 691 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap"); 692 #ifdef ASSERT 693 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) { 694 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr); 695 } 696 #endif 697 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked"); 698 699 // Region covering the object. 700 size_t region_index = addr_to_region_idx(addr); 701 const RegionData* const region_ptr = region(region_index); 702 HeapWord* const region_addr = region_align_down(addr); 703 704 assert(addr < region_addr + RegionSize, "Region does not cover object"); 705 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check"); 706 707 HeapWord* result = region_ptr->destination(); 708 709 // If all the data in the region is live, then the new location of the object 710 // can be calculated from the destination of the region plus the offset of the 711 // object in the region. 712 if (region_ptr->data_size() == RegionSize) { 713 result += pointer_delta(addr, region_addr); 714 return result; 715 } 716 717 // The new location of the object is 718 // region destination + 719 // size of the partial object extending onto the region + 720 // sizes of the live objects in the Region that are to the left of addr 721 const size_t partial_obj_size = region_ptr->partial_obj_size(); 722 HeapWord* const search_start = region_addr + partial_obj_size; 723 724 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 725 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 726 727 result += partial_obj_size + live_to_left; 728 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 729 return result; 730 } 731 732 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) { 733 klassOop updated_klass; 734 if (PSParallelCompact::should_update_klass(old_klass)) { 735 updated_klass = (klassOop) calc_new_pointer(old_klass); 736 } else { 737 updated_klass = old_klass; 738 } 739 740 return updated_klass; 741 } 742 743 #ifdef ASSERT 744 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 745 { 746 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 747 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 748 for (const size_t* p = beg; p < end; ++p) { 749 assert(*p == 0, "not zero"); 750 } 751 } 752 753 void ParallelCompactData::verify_clear() 754 { 755 verify_clear(_region_vspace); 756 } 757 #endif // #ifdef ASSERT 758 759 #ifdef NOT_PRODUCT 760 ParallelCompactData::RegionData* debug_region(size_t region_index) { 761 ParallelCompactData& sd = PSParallelCompact::summary_data(); 762 return sd.region(region_index); 763 } 764 #endif 765 766 elapsedTimer PSParallelCompact::_accumulated_time; 767 unsigned int PSParallelCompact::_total_invocations = 0; 768 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 769 jlong PSParallelCompact::_time_of_last_gc = 0; 770 CollectorCounters* PSParallelCompact::_counters = NULL; 771 ParMarkBitMap PSParallelCompact::_mark_bitmap; 772 ParallelCompactData PSParallelCompact::_summary_data; 773 774 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 775 776 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } 777 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 778 779 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 780 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 781 782 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true); 783 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false); 784 785 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); } 786 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); } 787 788 void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); } 789 790 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); } 791 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 792 793 void PSParallelCompact::post_initialize() { 794 ParallelScavengeHeap* heap = gc_heap(); 795 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 796 797 MemRegion mr = heap->reserved_region(); 798 _ref_processor = ReferenceProcessor::create_ref_processor( 799 mr, // span 800 true, // atomic_discovery 801 true, // mt_discovery 802 &_is_alive_closure, 803 ParallelGCThreads, 804 ParallelRefProcEnabled); 805 _counters = new CollectorCounters("PSParallelCompact", 1); 806 807 // Initialize static fields in ParCompactionManager. 808 ParCompactionManager::initialize(mark_bitmap()); 809 } 810 811 bool PSParallelCompact::initialize() { 812 ParallelScavengeHeap* heap = gc_heap(); 813 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 814 MemRegion mr = heap->reserved_region(); 815 816 // Was the old gen get allocated successfully? 817 if (!heap->old_gen()->is_allocated()) { 818 return false; 819 } 820 821 initialize_space_info(); 822 initialize_dead_wood_limiter(); 823 824 if (!_mark_bitmap.initialize(mr)) { 825 vm_shutdown_during_initialization("Unable to allocate bit map for " 826 "parallel garbage collection for the requested heap size."); 827 return false; 828 } 829 830 if (!_summary_data.initialize(mr)) { 831 vm_shutdown_during_initialization("Unable to allocate tables for " 832 "parallel garbage collection for the requested heap size."); 833 return false; 834 } 835 836 return true; 837 } 838 839 void PSParallelCompact::initialize_space_info() 840 { 841 memset(&_space_info, 0, sizeof(_space_info)); 842 843 ParallelScavengeHeap* heap = gc_heap(); 844 PSYoungGen* young_gen = heap->young_gen(); 845 MutableSpace* perm_space = heap->perm_gen()->object_space(); 846 847 _space_info[perm_space_id].set_space(perm_space); 848 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 849 _space_info[eden_space_id].set_space(young_gen->eden_space()); 850 _space_info[from_space_id].set_space(young_gen->from_space()); 851 _space_info[to_space_id].set_space(young_gen->to_space()); 852 853 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array()); 854 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 855 856 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top()); 857 if (TraceParallelOldGCDensePrefix) { 858 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT, 859 _space_info[perm_space_id].min_dense_prefix()); 860 } 861 } 862 863 void PSParallelCompact::initialize_dead_wood_limiter() 864 { 865 const size_t max = 100; 866 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 867 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 868 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 869 DEBUG_ONLY(_dwl_initialized = true;) 870 _dwl_adjustment = normal_distribution(1.0); 871 } 872 873 // Simple class for storing info about the heap at the start of GC, to be used 874 // after GC for comparison/printing. 875 class PreGCValues { 876 public: 877 PreGCValues() { } 878 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 879 880 void fill(ParallelScavengeHeap* heap) { 881 _heap_used = heap->used(); 882 _young_gen_used = heap->young_gen()->used_in_bytes(); 883 _old_gen_used = heap->old_gen()->used_in_bytes(); 884 _perm_gen_used = heap->perm_gen()->used_in_bytes(); 885 }; 886 887 size_t heap_used() const { return _heap_used; } 888 size_t young_gen_used() const { return _young_gen_used; } 889 size_t old_gen_used() const { return _old_gen_used; } 890 size_t perm_gen_used() const { return _perm_gen_used; } 891 892 private: 893 size_t _heap_used; 894 size_t _young_gen_used; 895 size_t _old_gen_used; 896 size_t _perm_gen_used; 897 }; 898 899 void 900 PSParallelCompact::clear_data_covering_space(SpaceId id) 901 { 902 // At this point, top is the value before GC, new_top() is the value that will 903 // be set at the end of GC. The marking bitmap is cleared to top; nothing 904 // should be marked above top. The summary data is cleared to the larger of 905 // top & new_top. 906 MutableSpace* const space = _space_info[id].space(); 907 HeapWord* const bot = space->bottom(); 908 HeapWord* const top = space->top(); 909 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 910 911 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 912 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 913 _mark_bitmap.clear_range(beg_bit, end_bit); 914 915 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 916 const size_t end_region = 917 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 918 _summary_data.clear_range(beg_region, end_region); 919 920 // Clear the data used to 'split' regions. 921 SplitInfo& split_info = _space_info[id].split_info(); 922 if (split_info.is_valid()) { 923 split_info.clear(); 924 } 925 DEBUG_ONLY(split_info.verify_clear();) 926 } 927 928 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 929 { 930 // Update the from & to space pointers in space_info, since they are swapped 931 // at each young gen gc. Do the update unconditionally (even though a 932 // promotion failure does not swap spaces) because an unknown number of minor 933 // collections will have swapped the spaces an unknown number of times. 934 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); 935 ParallelScavengeHeap* heap = gc_heap(); 936 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 937 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 938 939 pre_gc_values->fill(heap); 940 941 ParCompactionManager::reset(); 942 NOT_PRODUCT(_mark_bitmap.reset_counters()); 943 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 944 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 945 946 // Increment the invocation count 947 heap->increment_total_collections(true); 948 949 // We need to track unique mark sweep invocations as well. 950 _total_invocations++; 951 952 if (PrintHeapAtGC) { 953 Universe::print_heap_before_gc(); 954 } 955 956 // Fill in TLABs 957 heap->accumulate_statistics_all_tlabs(); 958 heap->ensure_parsability(true); // retire TLABs 959 960 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 961 HandleMark hm; // Discard invalid handles created during verification 962 gclog_or_tty->print(" VerifyBeforeGC:"); 963 Universe::verify(true); 964 } 965 966 // Verify object start arrays 967 if (VerifyObjectStartArray && 968 VerifyBeforeGC) { 969 heap->old_gen()->verify_object_start_array(); 970 heap->perm_gen()->verify_object_start_array(); 971 } 972 973 DEBUG_ONLY(mark_bitmap()->verify_clear();) 974 DEBUG_ONLY(summary_data().verify_clear();) 975 976 // Have worker threads release resources the next time they run a task. 977 gc_task_manager()->release_all_resources(); 978 } 979 980 void PSParallelCompact::post_compact() 981 { 982 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); 983 984 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 985 // Clear the marking bitmap, summary data and split info. 986 clear_data_covering_space(SpaceId(id)); 987 // Update top(). Must be done after clearing the bitmap and summary data. 988 _space_info[id].publish_new_top(); 989 } 990 991 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 992 MutableSpace* const from_space = _space_info[from_space_id].space(); 993 MutableSpace* const to_space = _space_info[to_space_id].space(); 994 995 ParallelScavengeHeap* heap = gc_heap(); 996 bool eden_empty = eden_space->is_empty(); 997 if (!eden_empty) { 998 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 999 heap->young_gen(), heap->old_gen()); 1000 } 1001 1002 // Update heap occupancy information which is used as input to the soft ref 1003 // clearing policy at the next gc. 1004 Universe::update_heap_info_at_gc(); 1005 1006 bool young_gen_empty = eden_empty && from_space->is_empty() && 1007 to_space->is_empty(); 1008 1009 BarrierSet* bs = heap->barrier_set(); 1010 if (bs->is_a(BarrierSet::ModRef)) { 1011 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1012 MemRegion old_mr = heap->old_gen()->reserved(); 1013 MemRegion perm_mr = heap->perm_gen()->reserved(); 1014 assert(perm_mr.end() <= old_mr.start(), "Generations out of order"); 1015 1016 if (young_gen_empty) { 1017 modBS->clear(MemRegion(perm_mr.start(), old_mr.end())); 1018 } else { 1019 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end())); 1020 } 1021 } 1022 1023 Threads::gc_epilogue(); 1024 CodeCache::gc_epilogue(); 1025 1026 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1027 1028 ref_processor()->enqueue_discovered_references(NULL); 1029 1030 if (ZapUnusedHeapArea) { 1031 heap->gen_mangle_unused_area(); 1032 } 1033 1034 // Update time of last GC 1035 reset_millis_since_last_gc(); 1036 } 1037 1038 HeapWord* 1039 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1040 bool maximum_compaction) 1041 { 1042 const size_t region_size = ParallelCompactData::RegionSize; 1043 const ParallelCompactData& sd = summary_data(); 1044 1045 const MutableSpace* const space = _space_info[id].space(); 1046 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1047 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1048 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1049 1050 // Skip full regions at the beginning of the space--they are necessarily part 1051 // of the dense prefix. 1052 size_t full_count = 0; 1053 const RegionData* cp; 1054 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1055 ++full_count; 1056 } 1057 1058 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1059 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1060 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1061 if (maximum_compaction || cp == end_cp || interval_ended) { 1062 _maximum_compaction_gc_num = total_invocations(); 1063 return sd.region_to_addr(cp); 1064 } 1065 1066 HeapWord* const new_top = _space_info[id].new_top(); 1067 const size_t space_live = pointer_delta(new_top, space->bottom()); 1068 const size_t space_used = space->used_in_words(); 1069 const size_t space_capacity = space->capacity_in_words(); 1070 1071 const double cur_density = double(space_live) / space_capacity; 1072 const double deadwood_density = 1073 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1074 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1075 1076 if (TraceParallelOldGCDensePrefix) { 1077 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1078 cur_density, deadwood_density, deadwood_goal); 1079 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1080 "space_cap=" SIZE_FORMAT, 1081 space_live, space_used, 1082 space_capacity); 1083 } 1084 1085 // XXX - Use binary search? 1086 HeapWord* dense_prefix = sd.region_to_addr(cp); 1087 const RegionData* full_cp = cp; 1088 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1089 while (cp < end_cp) { 1090 HeapWord* region_destination = cp->destination(); 1091 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1092 if (TraceParallelOldGCDensePrefix && Verbose) { 1093 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1094 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1095 sd.region(cp), region_destination, 1096 dense_prefix, cur_deadwood); 1097 } 1098 1099 if (cur_deadwood >= deadwood_goal) { 1100 // Found the region that has the correct amount of deadwood to the left. 1101 // This typically occurs after crossing a fairly sparse set of regions, so 1102 // iterate backwards over those sparse regions, looking for the region 1103 // that has the lowest density of live objects 'to the right.' 1104 size_t space_to_left = sd.region(cp) * region_size; 1105 size_t live_to_left = space_to_left - cur_deadwood; 1106 size_t space_to_right = space_capacity - space_to_left; 1107 size_t live_to_right = space_live - live_to_left; 1108 double density_to_right = double(live_to_right) / space_to_right; 1109 while (cp > full_cp) { 1110 --cp; 1111 const size_t prev_region_live_to_right = live_to_right - 1112 cp->data_size(); 1113 const size_t prev_region_space_to_right = space_to_right + region_size; 1114 double prev_region_density_to_right = 1115 double(prev_region_live_to_right) / prev_region_space_to_right; 1116 if (density_to_right <= prev_region_density_to_right) { 1117 return dense_prefix; 1118 } 1119 if (TraceParallelOldGCDensePrefix && Verbose) { 1120 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1121 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1122 prev_region_density_to_right); 1123 } 1124 dense_prefix -= region_size; 1125 live_to_right = prev_region_live_to_right; 1126 space_to_right = prev_region_space_to_right; 1127 density_to_right = prev_region_density_to_right; 1128 } 1129 return dense_prefix; 1130 } 1131 1132 dense_prefix += region_size; 1133 ++cp; 1134 } 1135 1136 return dense_prefix; 1137 } 1138 1139 #ifndef PRODUCT 1140 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1141 const SpaceId id, 1142 const bool maximum_compaction, 1143 HeapWord* const addr) 1144 { 1145 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1146 RegionData* const cp = summary_data().region(region_idx); 1147 const MutableSpace* const space = _space_info[id].space(); 1148 HeapWord* const new_top = _space_info[id].new_top(); 1149 1150 const size_t space_live = pointer_delta(new_top, space->bottom()); 1151 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1152 const size_t space_cap = space->capacity_in_words(); 1153 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1154 const size_t live_to_right = new_top - cp->destination(); 1155 const size_t dead_to_right = space->top() - addr - live_to_right; 1156 1157 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1158 "spl=" SIZE_FORMAT " " 1159 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1160 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1161 " ratio=%10.8f", 1162 algorithm, addr, region_idx, 1163 space_live, 1164 dead_to_left, dead_to_left_pct, 1165 dead_to_right, live_to_right, 1166 double(dead_to_right) / live_to_right); 1167 } 1168 #endif // #ifndef PRODUCT 1169 1170 // Return a fraction indicating how much of the generation can be treated as 1171 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1172 // based on the density of live objects in the generation to determine a limit, 1173 // which is then adjusted so the return value is min_percent when the density is 1174 // 1. 1175 // 1176 // The following table shows some return values for a different values of the 1177 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1178 // min_percent is 1. 1179 // 1180 // fraction allowed as dead wood 1181 // ----------------------------------------------------------------- 1182 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1183 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1184 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1185 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1186 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1187 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1188 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1189 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1190 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1191 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1192 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1193 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1194 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1195 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1196 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1197 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1198 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1199 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1200 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1201 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1202 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1203 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1204 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1205 1206 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1207 { 1208 assert(_dwl_initialized, "uninitialized"); 1209 1210 // The raw limit is the value of the normal distribution at x = density. 1211 const double raw_limit = normal_distribution(density); 1212 1213 // Adjust the raw limit so it becomes the minimum when the density is 1. 1214 // 1215 // First subtract the adjustment value (which is simply the precomputed value 1216 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1217 // Then add the minimum value, so the minimum is returned when the density is 1218 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1219 const double min = double(min_percent) / 100.0; 1220 const double limit = raw_limit - _dwl_adjustment + min; 1221 return MAX2(limit, 0.0); 1222 } 1223 1224 ParallelCompactData::RegionData* 1225 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1226 const RegionData* end) 1227 { 1228 const size_t region_size = ParallelCompactData::RegionSize; 1229 ParallelCompactData& sd = summary_data(); 1230 size_t left = sd.region(beg); 1231 size_t right = end > beg ? sd.region(end) - 1 : left; 1232 1233 // Binary search. 1234 while (left < right) { 1235 // Equivalent to (left + right) / 2, but does not overflow. 1236 const size_t middle = left + (right - left) / 2; 1237 RegionData* const middle_ptr = sd.region(middle); 1238 HeapWord* const dest = middle_ptr->destination(); 1239 HeapWord* const addr = sd.region_to_addr(middle); 1240 assert(dest != NULL, "sanity"); 1241 assert(dest <= addr, "must move left"); 1242 1243 if (middle > left && dest < addr) { 1244 right = middle - 1; 1245 } else if (middle < right && middle_ptr->data_size() == region_size) { 1246 left = middle + 1; 1247 } else { 1248 return middle_ptr; 1249 } 1250 } 1251 return sd.region(left); 1252 } 1253 1254 ParallelCompactData::RegionData* 1255 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1256 const RegionData* end, 1257 size_t dead_words) 1258 { 1259 ParallelCompactData& sd = summary_data(); 1260 size_t left = sd.region(beg); 1261 size_t right = end > beg ? sd.region(end) - 1 : left; 1262 1263 // Binary search. 1264 while (left < right) { 1265 // Equivalent to (left + right) / 2, but does not overflow. 1266 const size_t middle = left + (right - left) / 2; 1267 RegionData* const middle_ptr = sd.region(middle); 1268 HeapWord* const dest = middle_ptr->destination(); 1269 HeapWord* const addr = sd.region_to_addr(middle); 1270 assert(dest != NULL, "sanity"); 1271 assert(dest <= addr, "must move left"); 1272 1273 const size_t dead_to_left = pointer_delta(addr, dest); 1274 if (middle > left && dead_to_left > dead_words) { 1275 right = middle - 1; 1276 } else if (middle < right && dead_to_left < dead_words) { 1277 left = middle + 1; 1278 } else { 1279 return middle_ptr; 1280 } 1281 } 1282 return sd.region(left); 1283 } 1284 1285 // The result is valid during the summary phase, after the initial summarization 1286 // of each space into itself, and before final summarization. 1287 inline double 1288 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1289 HeapWord* const bottom, 1290 HeapWord* const top, 1291 HeapWord* const new_top) 1292 { 1293 ParallelCompactData& sd = summary_data(); 1294 1295 assert(cp != NULL, "sanity"); 1296 assert(bottom != NULL, "sanity"); 1297 assert(top != NULL, "sanity"); 1298 assert(new_top != NULL, "sanity"); 1299 assert(top >= new_top, "summary data problem?"); 1300 assert(new_top > bottom, "space is empty; should not be here"); 1301 assert(new_top >= cp->destination(), "sanity"); 1302 assert(top >= sd.region_to_addr(cp), "sanity"); 1303 1304 HeapWord* const destination = cp->destination(); 1305 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1306 const size_t compacted_region_live = pointer_delta(new_top, destination); 1307 const size_t compacted_region_used = pointer_delta(top, 1308 sd.region_to_addr(cp)); 1309 const size_t reclaimable = compacted_region_used - compacted_region_live; 1310 1311 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1312 return double(reclaimable) / divisor; 1313 } 1314 1315 // Return the address of the end of the dense prefix, a.k.a. the start of the 1316 // compacted region. The address is always on a region boundary. 1317 // 1318 // Completely full regions at the left are skipped, since no compaction can 1319 // occur in those regions. Then the maximum amount of dead wood to allow is 1320 // computed, based on the density (amount live / capacity) of the generation; 1321 // the region with approximately that amount of dead space to the left is 1322 // identified as the limit region. Regions between the last completely full 1323 // region and the limit region are scanned and the one that has the best 1324 // (maximum) reclaimed_ratio() is selected. 1325 HeapWord* 1326 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1327 bool maximum_compaction) 1328 { 1329 if (ParallelOldGCSplitALot) { 1330 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1331 // The value was chosen to provoke splitting a young gen space; use it. 1332 return _space_info[id].dense_prefix(); 1333 } 1334 } 1335 1336 const size_t region_size = ParallelCompactData::RegionSize; 1337 const ParallelCompactData& sd = summary_data(); 1338 1339 const MutableSpace* const space = _space_info[id].space(); 1340 HeapWord* const top = space->top(); 1341 HeapWord* const top_aligned_up = sd.region_align_up(top); 1342 HeapWord* const new_top = _space_info[id].new_top(); 1343 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1344 HeapWord* const bottom = space->bottom(); 1345 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1346 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1347 const RegionData* const new_top_cp = 1348 sd.addr_to_region_ptr(new_top_aligned_up); 1349 1350 // Skip full regions at the beginning of the space--they are necessarily part 1351 // of the dense prefix. 1352 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1353 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1354 space->is_empty(), "no dead space allowed to the left"); 1355 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1356 "region must have dead space"); 1357 1358 // The gc number is saved whenever a maximum compaction is done, and used to 1359 // determine when the maximum compaction interval has expired. This avoids 1360 // successive max compactions for different reasons. 1361 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1362 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1363 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1364 total_invocations() == HeapFirstMaximumCompactionCount; 1365 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1366 _maximum_compaction_gc_num = total_invocations(); 1367 return sd.region_to_addr(full_cp); 1368 } 1369 1370 const size_t space_live = pointer_delta(new_top, bottom); 1371 const size_t space_used = space->used_in_words(); 1372 const size_t space_capacity = space->capacity_in_words(); 1373 1374 const double density = double(space_live) / double(space_capacity); 1375 const size_t min_percent_free = 1376 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio; 1377 const double limiter = dead_wood_limiter(density, min_percent_free); 1378 const size_t dead_wood_max = space_used - space_live; 1379 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1380 dead_wood_max); 1381 1382 if (TraceParallelOldGCDensePrefix) { 1383 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1384 "space_cap=" SIZE_FORMAT, 1385 space_live, space_used, 1386 space_capacity); 1387 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1388 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1389 density, min_percent_free, limiter, 1390 dead_wood_max, dead_wood_limit); 1391 } 1392 1393 // Locate the region with the desired amount of dead space to the left. 1394 const RegionData* const limit_cp = 1395 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1396 1397 // Scan from the first region with dead space to the limit region and find the 1398 // one with the best (largest) reclaimed ratio. 1399 double best_ratio = 0.0; 1400 const RegionData* best_cp = full_cp; 1401 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1402 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1403 if (tmp_ratio > best_ratio) { 1404 best_cp = cp; 1405 best_ratio = tmp_ratio; 1406 } 1407 } 1408 1409 #if 0 1410 // Something to consider: if the region with the best ratio is 'close to' the 1411 // first region w/free space, choose the first region with free space 1412 // ("first-free"). The first-free region is usually near the start of the 1413 // heap, which means we are copying most of the heap already, so copy a bit 1414 // more to get complete compaction. 1415 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1416 _maximum_compaction_gc_num = total_invocations(); 1417 best_cp = full_cp; 1418 } 1419 #endif // #if 0 1420 1421 return sd.region_to_addr(best_cp); 1422 } 1423 1424 #ifndef PRODUCT 1425 void 1426 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1427 size_t words) 1428 { 1429 if (TraceParallelOldGCSummaryPhase) { 1430 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1431 SIZE_FORMAT, start, start + words, words); 1432 } 1433 1434 ObjectStartArray* const start_array = _space_info[id].start_array(); 1435 CollectedHeap::fill_with_objects(start, words); 1436 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1437 _mark_bitmap.mark_obj(p, words); 1438 _summary_data.add_obj(p, words); 1439 start_array->allocate_block(p); 1440 } 1441 } 1442 1443 void 1444 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1445 { 1446 ParallelCompactData& sd = summary_data(); 1447 MutableSpace* space = _space_info[id].space(); 1448 1449 // Find the source and destination start addresses. 1450 HeapWord* const src_addr = sd.region_align_down(start); 1451 HeapWord* dst_addr; 1452 if (src_addr < start) { 1453 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1454 } else if (src_addr > space->bottom()) { 1455 // The start (the original top() value) is aligned to a region boundary so 1456 // the associated region does not have a destination. Compute the 1457 // destination from the previous region. 1458 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1459 dst_addr = cp->destination() + cp->data_size(); 1460 } else { 1461 // Filling the entire space. 1462 dst_addr = space->bottom(); 1463 } 1464 assert(dst_addr != NULL, "sanity"); 1465 1466 // Update the summary data. 1467 bool result = _summary_data.summarize(_space_info[id].split_info(), 1468 src_addr, space->top(), NULL, 1469 dst_addr, space->end(), 1470 _space_info[id].new_top_addr()); 1471 assert(result, "should not fail: bad filler object size"); 1472 } 1473 1474 void 1475 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1476 { 1477 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1478 return; 1479 } 1480 1481 MutableSpace* const space = _space_info[id].space(); 1482 if (space->is_empty()) { 1483 HeapWord* b = space->bottom(); 1484 HeapWord* t = b + space->capacity_in_words() / 2; 1485 space->set_top(t); 1486 if (ZapUnusedHeapArea) { 1487 space->set_top_for_allocations(); 1488 } 1489 1490 size_t obj_len = 8; 1491 while (b + obj_len <= t) { 1492 CollectedHeap::fill_with_object(b, obj_len); 1493 mark_bitmap()->mark_obj(b, obj_len); 1494 summary_data().add_obj(b, obj_len); 1495 b += obj_len; 1496 obj_len = (obj_len & 0x18) + 8; // 8 16 24 32 8 16 24 32 ... 1497 } 1498 if (b < t) { 1499 // The loop didn't completely fill to t (top); adjust top downward. 1500 space->set_top(b); 1501 if (ZapUnusedHeapArea) { 1502 space->set_top_for_allocations(); 1503 } 1504 } 1505 1506 HeapWord** nta = _space_info[id].new_top_addr(); 1507 bool result = summary_data().summarize(_space_info[id].split_info(), 1508 space->bottom(), space->top(), NULL, 1509 space->bottom(), space->end(), nta); 1510 assert(result, "space must fit into itself"); 1511 } 1512 } 1513 1514 void 1515 PSParallelCompact::provoke_split(bool & max_compaction) 1516 { 1517 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1518 return; 1519 } 1520 1521 const size_t region_size = ParallelCompactData::RegionSize; 1522 ParallelCompactData& sd = summary_data(); 1523 1524 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1525 MutableSpace* const from_space = _space_info[from_space_id].space(); 1526 const size_t eden_live = pointer_delta(eden_space->top(), 1527 _space_info[eden_space_id].new_top()); 1528 const size_t from_live = pointer_delta(from_space->top(), 1529 _space_info[from_space_id].new_top()); 1530 1531 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1532 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1533 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1534 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1535 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1536 1537 // Choose the space to split; need at least 2 regions live (or fillable). 1538 SpaceId id; 1539 MutableSpace* space; 1540 size_t live_words; 1541 size_t fill_words; 1542 if (eden_live + eden_fillable >= region_size * 2) { 1543 id = eden_space_id; 1544 space = eden_space; 1545 live_words = eden_live; 1546 fill_words = eden_fillable; 1547 } else if (from_live + from_fillable >= region_size * 2) { 1548 id = from_space_id; 1549 space = from_space; 1550 live_words = from_live; 1551 fill_words = from_fillable; 1552 } else { 1553 return; // Give up. 1554 } 1555 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1556 1557 if (live_words < region_size * 2) { 1558 // Fill from top() to end() w/live objects of mixed sizes. 1559 HeapWord* const fill_start = space->top(); 1560 live_words += fill_words; 1561 1562 space->set_top(fill_start + fill_words); 1563 if (ZapUnusedHeapArea) { 1564 space->set_top_for_allocations(); 1565 } 1566 1567 HeapWord* cur_addr = fill_start; 1568 while (fill_words > 0) { 1569 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1570 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1571 if (fill_words - cur_size < min_fill_size) { 1572 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1573 } 1574 1575 CollectedHeap::fill_with_object(cur_addr, cur_size); 1576 mark_bitmap()->mark_obj(cur_addr, cur_size); 1577 sd.add_obj(cur_addr, cur_size); 1578 1579 cur_addr += cur_size; 1580 fill_words -= cur_size; 1581 } 1582 1583 summarize_new_objects(id, fill_start); 1584 } 1585 1586 max_compaction = false; 1587 1588 // Manipulate the old gen so that it has room for about half of the live data 1589 // in the target young gen space (live_words / 2). 1590 id = old_space_id; 1591 space = _space_info[id].space(); 1592 const size_t free_at_end = space->free_in_words(); 1593 const size_t free_target = align_object_size(live_words / 2); 1594 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1595 1596 if (free_at_end >= free_target + min_fill_size) { 1597 // Fill space above top() and set the dense prefix so everything survives. 1598 HeapWord* const fill_start = space->top(); 1599 const size_t fill_size = free_at_end - free_target; 1600 space->set_top(space->top() + fill_size); 1601 if (ZapUnusedHeapArea) { 1602 space->set_top_for_allocations(); 1603 } 1604 fill_with_live_objects(id, fill_start, fill_size); 1605 summarize_new_objects(id, fill_start); 1606 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1607 } else if (dead + free_at_end > free_target) { 1608 // Find a dense prefix that makes the right amount of space available. 1609 HeapWord* cur = sd.region_align_down(space->top()); 1610 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1611 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1612 while (dead_to_right < free_target) { 1613 cur -= region_size; 1614 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1615 dead_to_right = pointer_delta(space->end(), cur_destination); 1616 } 1617 _space_info[id].set_dense_prefix(cur); 1618 } 1619 } 1620 #endif // #ifndef PRODUCT 1621 1622 void PSParallelCompact::summarize_spaces_quick() 1623 { 1624 for (unsigned int i = 0; i < last_space_id; ++i) { 1625 const MutableSpace* space = _space_info[i].space(); 1626 HeapWord** nta = _space_info[i].new_top_addr(); 1627 bool result = _summary_data.summarize(_space_info[i].split_info(), 1628 space->bottom(), space->top(), NULL, 1629 space->bottom(), space->end(), nta); 1630 assert(result, "space must fit into itself"); 1631 _space_info[i].set_dense_prefix(space->bottom()); 1632 } 1633 1634 #ifndef PRODUCT 1635 if (ParallelOldGCSplitALot) { 1636 provoke_split_fill_survivor(to_space_id); 1637 } 1638 #endif // #ifndef PRODUCT 1639 } 1640 1641 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1642 { 1643 HeapWord* const dense_prefix_end = dense_prefix(id); 1644 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1645 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1646 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1647 // Only enough dead space is filled so that any remaining dead space to the 1648 // left is larger than the minimum filler object. (The remainder is filled 1649 // during the copy/update phase.) 1650 // 1651 // The size of the dead space to the right of the boundary is not a 1652 // concern, since compaction will be able to use whatever space is 1653 // available. 1654 // 1655 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1656 // surrounds the space to be filled with an object. 1657 // 1658 // In the 32-bit VM, each bit represents two 32-bit words: 1659 // +---+ 1660 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1661 // end_bits: ... x x x | 0 | || 0 x x ... 1662 // +---+ 1663 // 1664 // In the 64-bit VM, each bit represents one 64-bit word: 1665 // +------------+ 1666 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1667 // end_bits: ... x x 1 | 0 || 0 | x x ... 1668 // +------------+ 1669 // +-------+ 1670 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1671 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1672 // +-------+ 1673 // +-----------+ 1674 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1675 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1676 // +-----------+ 1677 // +-------+ 1678 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1679 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1680 // +-------+ 1681 1682 // Initially assume case a, c or e will apply. 1683 size_t obj_len = (size_t)oopDesc::header_size(); 1684 HeapWord* obj_beg = dense_prefix_end - obj_len; 1685 1686 #ifdef _LP64 1687 if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1688 // Case b above. 1689 obj_beg = dense_prefix_end - 1; 1690 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1691 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1692 // Case d above. 1693 obj_beg = dense_prefix_end - 3; 1694 obj_len = 3; 1695 } 1696 #endif // #ifdef _LP64 1697 1698 CollectedHeap::fill_with_object(obj_beg, obj_len); 1699 _mark_bitmap.mark_obj(obj_beg, obj_len); 1700 _summary_data.add_obj(obj_beg, obj_len); 1701 assert(start_array(id) != NULL, "sanity"); 1702 start_array(id)->allocate_block(obj_beg); 1703 } 1704 } 1705 1706 void 1707 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1708 { 1709 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1710 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1711 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1712 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1713 cur->set_source_region(0); 1714 } 1715 } 1716 1717 void 1718 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1719 { 1720 assert(id < last_space_id, "id out of range"); 1721 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1722 ParallelOldGCSplitALot && id == old_space_id, 1723 "should have been reset in summarize_spaces_quick()"); 1724 1725 const MutableSpace* space = _space_info[id].space(); 1726 if (_space_info[id].new_top() != space->bottom()) { 1727 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1728 _space_info[id].set_dense_prefix(dense_prefix_end); 1729 1730 #ifndef PRODUCT 1731 if (TraceParallelOldGCDensePrefix) { 1732 print_dense_prefix_stats("ratio", id, maximum_compaction, 1733 dense_prefix_end); 1734 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1735 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1736 } 1737 #endif // #ifndef PRODUCT 1738 1739 // Recompute the summary data, taking into account the dense prefix. If 1740 // every last byte will be reclaimed, then the existing summary data which 1741 // compacts everything can be left in place. 1742 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1743 // If dead space crosses the dense prefix boundary, it is (at least 1744 // partially) filled with a dummy object, marked live and added to the 1745 // summary data. This simplifies the copy/update phase and must be done 1746 // before the final locations of objects are determined, to prevent 1747 // leaving a fragment of dead space that is too small to fill. 1748 fill_dense_prefix_end(id); 1749 1750 // Compute the destination of each Region, and thus each object. 1751 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1752 _summary_data.summarize(_space_info[id].split_info(), 1753 dense_prefix_end, space->top(), NULL, 1754 dense_prefix_end, space->end(), 1755 _space_info[id].new_top_addr()); 1756 } 1757 } 1758 1759 if (TraceParallelOldGCSummaryPhase) { 1760 const size_t region_size = ParallelCompactData::RegionSize; 1761 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1762 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1763 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1764 HeapWord* const new_top = _space_info[id].new_top(); 1765 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1766 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1767 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1768 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1769 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1770 id, space->capacity_in_words(), dense_prefix_end, 1771 dp_region, dp_words / region_size, 1772 cr_words / region_size, new_top); 1773 } 1774 } 1775 1776 #ifndef PRODUCT 1777 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1778 HeapWord* dst_beg, HeapWord* dst_end, 1779 SpaceId src_space_id, 1780 HeapWord* src_beg, HeapWord* src_end) 1781 { 1782 if (TraceParallelOldGCSummaryPhase) { 1783 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1784 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1785 SIZE_FORMAT "-" SIZE_FORMAT " " 1786 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1787 SIZE_FORMAT "-" SIZE_FORMAT, 1788 src_space_id, space_names[src_space_id], 1789 dst_space_id, space_names[dst_space_id], 1790 src_beg, src_end, 1791 _summary_data.addr_to_region_idx(src_beg), 1792 _summary_data.addr_to_region_idx(src_end), 1793 dst_beg, dst_end, 1794 _summary_data.addr_to_region_idx(dst_beg), 1795 _summary_data.addr_to_region_idx(dst_end)); 1796 } 1797 } 1798 #endif // #ifndef PRODUCT 1799 1800 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1801 bool maximum_compaction) 1802 { 1803 EventMark m("2 summarize"); 1804 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); 1805 // trace("2"); 1806 1807 #ifdef ASSERT 1808 if (TraceParallelOldGCMarkingPhase) { 1809 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1810 "add_obj_bytes=" SIZE_FORMAT, 1811 add_obj_count, add_obj_size * HeapWordSize); 1812 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1813 "mark_bitmap_bytes=" SIZE_FORMAT, 1814 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1815 } 1816 #endif // #ifdef ASSERT 1817 1818 // Quick summarization of each space into itself, to see how much is live. 1819 summarize_spaces_quick(); 1820 1821 if (TraceParallelOldGCSummaryPhase) { 1822 tty->print_cr("summary_phase: after summarizing each space to self"); 1823 Universe::print(); 1824 NOT_PRODUCT(print_region_ranges()); 1825 if (Verbose) { 1826 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1827 } 1828 } 1829 1830 // The amount of live data that will end up in old space (assuming it fits). 1831 size_t old_space_total_live = 0; 1832 assert(perm_space_id < old_space_id, "should not count perm data here"); 1833 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1834 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1835 _space_info[id].space()->bottom()); 1836 } 1837 1838 MutableSpace* const old_space = _space_info[old_space_id].space(); 1839 const size_t old_capacity = old_space->capacity_in_words(); 1840 if (old_space_total_live > old_capacity) { 1841 // XXX - should also try to expand 1842 maximum_compaction = true; 1843 } 1844 #ifndef PRODUCT 1845 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1846 provoke_split(maximum_compaction); 1847 } 1848 #endif // #ifndef PRODUCT 1849 1850 // Permanent and Old generations. 1851 summarize_space(perm_space_id, maximum_compaction); 1852 summarize_space(old_space_id, maximum_compaction); 1853 1854 // Summarize the remaining spaces in the young gen. The initial target space 1855 // is the old gen. If a space does not fit entirely into the target, then the 1856 // remainder is compacted into the space itself and that space becomes the new 1857 // target. 1858 SpaceId dst_space_id = old_space_id; 1859 HeapWord* dst_space_end = old_space->end(); 1860 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1861 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1862 const MutableSpace* space = _space_info[id].space(); 1863 const size_t live = pointer_delta(_space_info[id].new_top(), 1864 space->bottom()); 1865 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1866 1867 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1868 SpaceId(id), space->bottom(), space->top());) 1869 if (live > 0 && live <= available) { 1870 // All the live data will fit. 1871 bool done = _summary_data.summarize(_space_info[id].split_info(), 1872 space->bottom(), space->top(), 1873 NULL, 1874 *new_top_addr, dst_space_end, 1875 new_top_addr); 1876 assert(done, "space must fit into old gen"); 1877 1878 // Reset the new_top value for the space. 1879 _space_info[id].set_new_top(space->bottom()); 1880 } else if (live > 0) { 1881 // Attempt to fit part of the source space into the target space. 1882 HeapWord* next_src_addr = NULL; 1883 bool done = _summary_data.summarize(_space_info[id].split_info(), 1884 space->bottom(), space->top(), 1885 &next_src_addr, 1886 *new_top_addr, dst_space_end, 1887 new_top_addr); 1888 assert(!done, "space should not fit into old gen"); 1889 assert(next_src_addr != NULL, "sanity"); 1890 1891 // The source space becomes the new target, so the remainder is compacted 1892 // within the space itself. 1893 dst_space_id = SpaceId(id); 1894 dst_space_end = space->end(); 1895 new_top_addr = _space_info[id].new_top_addr(); 1896 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1897 space->bottom(), dst_space_end, 1898 SpaceId(id), next_src_addr, space->top());) 1899 done = _summary_data.summarize(_space_info[id].split_info(), 1900 next_src_addr, space->top(), 1901 NULL, 1902 space->bottom(), dst_space_end, 1903 new_top_addr); 1904 assert(done, "space must fit when compacted into itself"); 1905 assert(*new_top_addr <= space->top(), "usage should not grow"); 1906 } 1907 } 1908 1909 if (TraceParallelOldGCSummaryPhase) { 1910 tty->print_cr("summary_phase: after final summarization"); 1911 Universe::print(); 1912 NOT_PRODUCT(print_region_ranges()); 1913 if (Verbose) { 1914 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1915 } 1916 } 1917 } 1918 1919 // This method should contain all heap-specific policy for invoking a full 1920 // collection. invoke_no_policy() will only attempt to compact the heap; it 1921 // will do nothing further. If we need to bail out for policy reasons, scavenge 1922 // before full gc, or any other specialized behavior, it needs to be added here. 1923 // 1924 // Note that this method should only be called from the vm_thread while at a 1925 // safepoint. 1926 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1927 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1928 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1929 "should be in vm thread"); 1930 ParallelScavengeHeap* heap = gc_heap(); 1931 GCCause::Cause gc_cause = heap->gc_cause(); 1932 assert(!heap->is_gc_active(), "not reentrant"); 1933 1934 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1935 1936 // Before each allocation/collection attempt, find out from the 1937 // policy object if GCs are, on the whole, taking too long. If so, 1938 // bail out without attempting a collection. The exceptions are 1939 // for explicitly requested GC's. 1940 if (!policy->gc_time_limit_exceeded() || 1941 GCCause::is_user_requested_gc(gc_cause) || 1942 GCCause::is_serviceability_requested_gc(gc_cause)) { 1943 IsGCActiveMark mark; 1944 1945 if (ScavengeBeforeFullGC) { 1946 PSScavenge::invoke_no_policy(); 1947 } 1948 1949 PSParallelCompact::invoke_no_policy(maximum_heap_compaction); 1950 } 1951 } 1952 1953 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1954 size_t addr_region_index = addr_to_region_idx(addr); 1955 return region_index == addr_region_index; 1956 } 1957 1958 // This method contains no policy. You should probably 1959 // be calling invoke() instead. 1960 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1961 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1962 assert(ref_processor() != NULL, "Sanity"); 1963 1964 if (GC_locker::check_active_before_gc()) { 1965 return; 1966 } 1967 1968 TimeStamp marking_start; 1969 TimeStamp compaction_start; 1970 TimeStamp collection_exit; 1971 1972 ParallelScavengeHeap* heap = gc_heap(); 1973 GCCause::Cause gc_cause = heap->gc_cause(); 1974 PSYoungGen* young_gen = heap->young_gen(); 1975 PSOldGen* old_gen = heap->old_gen(); 1976 PSPermGen* perm_gen = heap->perm_gen(); 1977 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1978 1979 if (ZapUnusedHeapArea) { 1980 // Save information needed to minimize mangling 1981 heap->record_gen_tops_before_GC(); 1982 } 1983 1984 heap->pre_full_gc_dump(); 1985 1986 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 1987 1988 // Make sure data structures are sane, make the heap parsable, and do other 1989 // miscellaneous bookkeeping. 1990 PreGCValues pre_gc_values; 1991 pre_compact(&pre_gc_values); 1992 1993 // Get the compaction manager reserved for the VM thread. 1994 ParCompactionManager* const vmthread_cm = 1995 ParCompactionManager::manager_array(gc_task_manager()->workers()); 1996 1997 // Place after pre_compact() where the number of invocations is incremented. 1998 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 1999 2000 { 2001 ResourceMark rm; 2002 HandleMark hm; 2003 2004 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc; 2005 2006 // This is useful for debugging but don't change the output the 2007 // the customer sees. 2008 const char* gc_cause_str = "Full GC"; 2009 if (is_system_gc && PrintGCDetails) { 2010 gc_cause_str = "Full GC (System)"; 2011 } 2012 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2013 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2014 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty); 2015 TraceCollectorStats tcs(counters()); 2016 TraceMemoryManagerStats tms(true /* Full GC */); 2017 2018 if (TraceGen1Time) accumulated_time()->start(); 2019 2020 // Let the size policy know we're starting 2021 size_policy->major_collection_begin(); 2022 2023 // When collecting the permanent generation methodOops may be moving, 2024 // so we either have to flush all bcp data or convert it into bci. 2025 CodeCache::gc_prologue(); 2026 Threads::gc_prologue(); 2027 2028 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2029 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2030 2031 ref_processor()->enable_discovery(); 2032 ref_processor()->setup_policy(maximum_heap_compaction); 2033 2034 bool marked_for_unloading = false; 2035 2036 marking_start.update(); 2037 marking_phase(vmthread_cm, maximum_heap_compaction); 2038 2039 #ifndef PRODUCT 2040 if (TraceParallelOldGCMarkingPhase) { 2041 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " 2042 "cas_by_another %d", 2043 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), 2044 mark_bitmap()->cas_by_another()); 2045 } 2046 #endif // #ifndef PRODUCT 2047 2048 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc; 2049 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2050 2051 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2052 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2053 2054 // adjust_roots() updates Universe::_intArrayKlassObj which is 2055 // needed by the compaction for filling holes in the dense prefix. 2056 adjust_roots(); 2057 2058 compaction_start.update(); 2059 // Does the perm gen always have to be done serially because 2060 // klasses are used in the update of an object? 2061 compact_perm(vmthread_cm); 2062 2063 if (UseParallelOldGCCompacting) { 2064 compact(); 2065 } else { 2066 compact_serial(vmthread_cm); 2067 } 2068 2069 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2070 // done before resizing. 2071 post_compact(); 2072 2073 // Let the size policy know we're done 2074 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2075 2076 if (UseAdaptiveSizePolicy) { 2077 if (PrintAdaptiveSizePolicy) { 2078 gclog_or_tty->print("AdaptiveSizeStart: "); 2079 gclog_or_tty->stamp(); 2080 gclog_or_tty->print_cr(" collection: %d ", 2081 heap->total_collections()); 2082 if (Verbose) { 2083 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d" 2084 " perm_gen_capacity: %d ", 2085 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(), 2086 perm_gen->capacity_in_bytes()); 2087 } 2088 } 2089 2090 // Don't check if the size_policy is ready here. Let 2091 // the size_policy check that internally. 2092 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2093 ((gc_cause != GCCause::_java_lang_system_gc) || 2094 UseAdaptiveSizePolicyWithSystemGC)) { 2095 // Calculate optimal free space amounts 2096 assert(young_gen->max_size() > 2097 young_gen->from_space()->capacity_in_bytes() + 2098 young_gen->to_space()->capacity_in_bytes(), 2099 "Sizes of space in young gen are out-of-bounds"); 2100 size_t max_eden_size = young_gen->max_size() - 2101 young_gen->from_space()->capacity_in_bytes() - 2102 young_gen->to_space()->capacity_in_bytes(); 2103 size_policy->compute_generation_free_space( 2104 young_gen->used_in_bytes(), 2105 young_gen->eden_space()->used_in_bytes(), 2106 old_gen->used_in_bytes(), 2107 perm_gen->used_in_bytes(), 2108 young_gen->eden_space()->capacity_in_bytes(), 2109 old_gen->max_gen_size(), 2110 max_eden_size, 2111 true /* full gc*/, 2112 gc_cause); 2113 2114 heap->resize_old_gen( 2115 size_policy->calculated_old_free_size_in_bytes()); 2116 2117 // Don't resize the young generation at an major collection. A 2118 // desired young generation size may have been calculated but 2119 // resizing the young generation complicates the code because the 2120 // resizing of the old generation may have moved the boundary 2121 // between the young generation and the old generation. Let the 2122 // young generation resizing happen at the minor collections. 2123 } 2124 if (PrintAdaptiveSizePolicy) { 2125 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2126 heap->total_collections()); 2127 } 2128 } 2129 2130 if (UsePerfData) { 2131 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2132 counters->update_counters(); 2133 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2134 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2135 } 2136 2137 heap->resize_all_tlabs(); 2138 2139 // We collected the perm gen, so we'll resize it here. 2140 perm_gen->compute_new_size(pre_gc_values.perm_gen_used()); 2141 2142 if (TraceGen1Time) accumulated_time()->stop(); 2143 2144 if (PrintGC) { 2145 if (PrintGCDetails) { 2146 // No GC timestamp here. This is after GC so it would be confusing. 2147 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2148 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2149 heap->print_heap_change(pre_gc_values.heap_used()); 2150 // Print perm gen last (print_heap_change() excludes the perm gen). 2151 perm_gen->print_used_change(pre_gc_values.perm_gen_used()); 2152 } else { 2153 heap->print_heap_change(pre_gc_values.heap_used()); 2154 } 2155 } 2156 2157 // Track memory usage and detect low memory 2158 MemoryService::track_memory_usage(); 2159 heap->update_counters(); 2160 2161 if (PrintGCDetails) { 2162 if (size_policy->print_gc_time_limit_would_be_exceeded()) { 2163 if (size_policy->gc_time_limit_exceeded()) { 2164 gclog_or_tty->print_cr(" GC time is exceeding GCTimeLimit " 2165 "of %d%%", GCTimeLimit); 2166 } else { 2167 gclog_or_tty->print_cr(" GC time would exceed GCTimeLimit " 2168 "of %d%%", GCTimeLimit); 2169 } 2170 } 2171 size_policy->set_print_gc_time_limit_would_be_exceeded(false); 2172 } 2173 } 2174 2175 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2176 HandleMark hm; // Discard invalid handles created during verification 2177 gclog_or_tty->print(" VerifyAfterGC:"); 2178 Universe::verify(false); 2179 } 2180 2181 // Re-verify object start arrays 2182 if (VerifyObjectStartArray && 2183 VerifyAfterGC) { 2184 old_gen->verify_object_start_array(); 2185 perm_gen->verify_object_start_array(); 2186 } 2187 2188 if (ZapUnusedHeapArea) { 2189 old_gen->object_space()->check_mangled_unused_area_complete(); 2190 perm_gen->object_space()->check_mangled_unused_area_complete(); 2191 } 2192 2193 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2194 2195 collection_exit.update(); 2196 2197 if (PrintHeapAtGC) { 2198 Universe::print_heap_after_gc(); 2199 } 2200 if (PrintGCTaskTimeStamps) { 2201 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2202 INT64_FORMAT, 2203 marking_start.ticks(), compaction_start.ticks(), 2204 collection_exit.ticks()); 2205 gc_task_manager()->print_task_time_stamps(); 2206 } 2207 2208 heap->post_full_gc_dump(); 2209 2210 #ifdef TRACESPINNING 2211 ParallelTaskTerminator::print_termination_counts(); 2212 #endif 2213 } 2214 2215 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2216 PSYoungGen* young_gen, 2217 PSOldGen* old_gen) { 2218 MutableSpace* const eden_space = young_gen->eden_space(); 2219 assert(!eden_space->is_empty(), "eden must be non-empty"); 2220 assert(young_gen->virtual_space()->alignment() == 2221 old_gen->virtual_space()->alignment(), "alignments do not match"); 2222 2223 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2224 return false; 2225 } 2226 2227 // Both generations must be completely committed. 2228 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2229 return false; 2230 } 2231 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2232 return false; 2233 } 2234 2235 // Figure out how much to take from eden. Include the average amount promoted 2236 // in the total; otherwise the next young gen GC will simply bail out to a 2237 // full GC. 2238 const size_t alignment = old_gen->virtual_space()->alignment(); 2239 const size_t eden_used = eden_space->used_in_bytes(); 2240 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2241 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2242 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2243 2244 if (absorb_size >= eden_capacity) { 2245 return false; // Must leave some space in eden. 2246 } 2247 2248 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2249 if (new_young_size < young_gen->min_gen_size()) { 2250 return false; // Respect young gen minimum size. 2251 } 2252 2253 if (TraceAdaptiveGCBoundary && Verbose) { 2254 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2255 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2256 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2257 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2258 absorb_size / K, 2259 eden_capacity / K, (eden_capacity - absorb_size) / K, 2260 young_gen->from_space()->used_in_bytes() / K, 2261 young_gen->to_space()->used_in_bytes() / K, 2262 young_gen->capacity_in_bytes() / K, new_young_size / K); 2263 } 2264 2265 // Fill the unused part of the old gen. 2266 MutableSpace* const old_space = old_gen->object_space(); 2267 HeapWord* const unused_start = old_space->top(); 2268 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2269 2270 if (unused_words > 0) { 2271 if (unused_words < CollectedHeap::min_fill_size()) { 2272 return false; // If the old gen cannot be filled, must give up. 2273 } 2274 CollectedHeap::fill_with_objects(unused_start, unused_words); 2275 } 2276 2277 // Take the live data from eden and set both top and end in the old gen to 2278 // eden top. (Need to set end because reset_after_change() mangles the region 2279 // from end to virtual_space->high() in debug builds). 2280 HeapWord* const new_top = eden_space->top(); 2281 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2282 absorb_size); 2283 young_gen->reset_after_change(); 2284 old_space->set_top(new_top); 2285 old_space->set_end(new_top); 2286 old_gen->reset_after_change(); 2287 2288 // Update the object start array for the filler object and the data from eden. 2289 ObjectStartArray* const start_array = old_gen->start_array(); 2290 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2291 start_array->allocate_block(p); 2292 } 2293 2294 // Could update the promoted average here, but it is not typically updated at 2295 // full GCs and the value to use is unclear. Something like 2296 // 2297 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2298 2299 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2300 return true; 2301 } 2302 2303 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2304 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2305 "shouldn't return NULL"); 2306 return ParallelScavengeHeap::gc_task_manager(); 2307 } 2308 2309 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2310 bool maximum_heap_compaction) { 2311 // Recursively traverse all live objects and mark them 2312 EventMark m("1 mark object"); 2313 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); 2314 2315 ParallelScavengeHeap* heap = gc_heap(); 2316 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2317 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2318 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2319 2320 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2321 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2322 2323 { 2324 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); 2325 ParallelScavengeHeap::ParStrongRootsScope psrs; 2326 2327 GCTaskQueue* q = GCTaskQueue::create(); 2328 2329 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2330 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2331 // We scan the thread roots in parallel 2332 Threads::create_thread_roots_marking_tasks(q); 2333 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2334 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2335 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2336 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2337 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2338 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols)); 2339 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2340 2341 if (parallel_gc_threads > 1) { 2342 for (uint j = 0; j < parallel_gc_threads; j++) { 2343 q->enqueue(new StealMarkingTask(&terminator)); 2344 } 2345 } 2346 2347 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2348 q->enqueue(fin); 2349 2350 gc_task_manager()->add_list(q); 2351 2352 fin->wait_for(); 2353 2354 // We have to release the barrier tasks! 2355 WaitForBarrierGCTask::destroy(fin); 2356 } 2357 2358 // Process reference objects found during marking 2359 { 2360 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); 2361 if (ref_processor()->processing_is_mt()) { 2362 RefProcTaskExecutor task_executor; 2363 ref_processor()->process_discovered_references( 2364 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2365 &task_executor); 2366 } else { 2367 ref_processor()->process_discovered_references( 2368 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL); 2369 } 2370 } 2371 2372 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); 2373 // Follow system dictionary roots and unload classes. 2374 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2375 2376 // Follow code cache roots. 2377 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure, 2378 purged_class); 2379 follow_stack(cm); // Flush marking stack. 2380 2381 // Update subklass/sibling/implementor links of live klasses 2382 // revisit_klass_stack is used in follow_weak_klass_links(). 2383 follow_weak_klass_links(); 2384 2385 // Revisit memoized MDO's and clear any unmarked weak refs 2386 follow_mdo_weak_refs(); 2387 2388 // Visit symbol and interned string tables and delete unmarked oops 2389 SymbolTable::unlink(is_alive_closure()); 2390 StringTable::unlink(is_alive_closure()); 2391 2392 assert(cm->marking_stack()->size() == 0, "stack should be empty by now"); 2393 assert(cm->overflow_stack()->is_empty(), "stack should be empty by now"); 2394 } 2395 2396 // This should be moved to the shared markSweep code! 2397 class PSAlwaysTrueClosure: public BoolObjectClosure { 2398 public: 2399 void do_object(oop p) { ShouldNotReachHere(); } 2400 bool do_object_b(oop p) { return true; } 2401 }; 2402 static PSAlwaysTrueClosure always_true; 2403 2404 void PSParallelCompact::adjust_roots() { 2405 // Adjust the pointers to reflect the new locations 2406 EventMark m("3 adjust roots"); 2407 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); 2408 2409 // General strong roots. 2410 Universe::oops_do(adjust_root_pointer_closure()); 2411 ReferenceProcessor::oops_do(adjust_root_pointer_closure()); 2412 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles 2413 Threads::oops_do(adjust_root_pointer_closure(), NULL); 2414 ObjectSynchronizer::oops_do(adjust_root_pointer_closure()); 2415 FlatProfiler::oops_do(adjust_root_pointer_closure()); 2416 Management::oops_do(adjust_root_pointer_closure()); 2417 JvmtiExport::oops_do(adjust_root_pointer_closure()); 2418 // SO_AllClasses 2419 SystemDictionary::oops_do(adjust_root_pointer_closure()); 2420 vmSymbols::oops_do(adjust_root_pointer_closure()); 2421 2422 // Now adjust pointers in remaining weak roots. (All of which should 2423 // have been cleared if they pointed to non-surviving objects.) 2424 // Global (weak) JNI handles 2425 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure()); 2426 2427 CodeCache::oops_do(adjust_pointer_closure()); 2428 SymbolTable::oops_do(adjust_root_pointer_closure()); 2429 StringTable::oops_do(adjust_root_pointer_closure()); 2430 ref_processor()->weak_oops_do(adjust_root_pointer_closure()); 2431 // Roots were visited so references into the young gen in roots 2432 // may have been scanned. Process them also. 2433 // Should the reference processor have a span that excludes 2434 // young gen objects? 2435 PSScavenge::reference_processor()->weak_oops_do( 2436 adjust_root_pointer_closure()); 2437 } 2438 2439 void PSParallelCompact::compact_perm(ParCompactionManager* cm) { 2440 EventMark m("4 compact perm"); 2441 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty); 2442 // trace("4"); 2443 2444 gc_heap()->perm_gen()->start_array()->reset(); 2445 move_and_update(cm, perm_space_id); 2446 } 2447 2448 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2449 uint parallel_gc_threads) 2450 { 2451 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); 2452 2453 const unsigned int task_count = MAX2(parallel_gc_threads, 1U); 2454 for (unsigned int j = 0; j < task_count; j++) { 2455 q->enqueue(new DrainStacksCompactionTask()); 2456 } 2457 2458 // Find all regions that are available (can be filled immediately) and 2459 // distribute them to the thread stacks. The iteration is done in reverse 2460 // order (high to low) so the regions will be removed in ascending order. 2461 2462 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2463 2464 size_t fillable_regions = 0; // A count for diagnostic purposes. 2465 unsigned int which = 0; // The worker thread number. 2466 2467 for (unsigned int id = to_space_id; id > perm_space_id; --id) { 2468 SpaceInfo* const space_info = _space_info + id; 2469 MutableSpace* const space = space_info->space(); 2470 HeapWord* const new_top = space_info->new_top(); 2471 2472 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2473 const size_t end_region = 2474 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2475 assert(end_region > 0, "perm gen cannot be empty"); 2476 2477 for (size_t cur = end_region - 1; cur >= beg_region; --cur) { 2478 if (sd.region(cur)->claim_unsafe()) { 2479 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2480 cm->save_for_processing(cur); 2481 2482 if (TraceParallelOldGCCompactionPhase && Verbose) { 2483 const size_t count_mod_8 = fillable_regions & 7; 2484 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2485 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2486 if (count_mod_8 == 7) gclog_or_tty->cr(); 2487 } 2488 2489 NOT_PRODUCT(++fillable_regions;) 2490 2491 // Assign regions to threads in round-robin fashion. 2492 if (++which == task_count) { 2493 which = 0; 2494 } 2495 } 2496 } 2497 } 2498 2499 if (TraceParallelOldGCCompactionPhase) { 2500 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2501 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2502 } 2503 } 2504 2505 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2506 2507 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2508 uint parallel_gc_threads) { 2509 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); 2510 2511 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2512 2513 // Iterate over all the spaces adding tasks for updating 2514 // regions in the dense prefix. Assume that 1 gc thread 2515 // will work on opening the gaps and the remaining gc threads 2516 // will work on the dense prefix. 2517 unsigned int space_id; 2518 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2519 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2520 const MutableSpace* const space = _space_info[space_id].space(); 2521 2522 if (dense_prefix_end == space->bottom()) { 2523 // There is no dense prefix for this space. 2524 continue; 2525 } 2526 2527 // The dense prefix is before this region. 2528 size_t region_index_end_dense_prefix = 2529 sd.addr_to_region_idx(dense_prefix_end); 2530 RegionData* const dense_prefix_cp = 2531 sd.region(region_index_end_dense_prefix); 2532 assert(dense_prefix_end == space->end() || 2533 dense_prefix_cp->available() || 2534 dense_prefix_cp->claimed(), 2535 "The region after the dense prefix should always be ready to fill"); 2536 2537 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2538 2539 // Is there dense prefix work? 2540 size_t total_dense_prefix_regions = 2541 region_index_end_dense_prefix - region_index_start; 2542 // How many regions of the dense prefix should be given to 2543 // each thread? 2544 if (total_dense_prefix_regions > 0) { 2545 uint tasks_for_dense_prefix = 1; 2546 if (UseParallelDensePrefixUpdate) { 2547 if (total_dense_prefix_regions <= 2548 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2549 // Don't over partition. This assumes that 2550 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2551 // so there are not many regions to process. 2552 tasks_for_dense_prefix = parallel_gc_threads; 2553 } else { 2554 // Over partition 2555 tasks_for_dense_prefix = parallel_gc_threads * 2556 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2557 } 2558 } 2559 size_t regions_per_thread = total_dense_prefix_regions / 2560 tasks_for_dense_prefix; 2561 // Give each thread at least 1 region. 2562 if (regions_per_thread == 0) { 2563 regions_per_thread = 1; 2564 } 2565 2566 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2567 if (region_index_start >= region_index_end_dense_prefix) { 2568 break; 2569 } 2570 // region_index_end is not processed 2571 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2572 region_index_end_dense_prefix); 2573 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2574 region_index_start, 2575 region_index_end)); 2576 region_index_start = region_index_end; 2577 } 2578 } 2579 // This gets any part of the dense prefix that did not 2580 // fit evenly. 2581 if (region_index_start < region_index_end_dense_prefix) { 2582 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2583 region_index_start, 2584 region_index_end_dense_prefix)); 2585 } 2586 } 2587 } 2588 2589 void PSParallelCompact::enqueue_region_stealing_tasks( 2590 GCTaskQueue* q, 2591 ParallelTaskTerminator* terminator_ptr, 2592 uint parallel_gc_threads) { 2593 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); 2594 2595 // Once a thread has drained it's stack, it should try to steal regions from 2596 // other threads. 2597 if (parallel_gc_threads > 1) { 2598 for (uint j = 0; j < parallel_gc_threads; j++) { 2599 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2600 } 2601 } 2602 } 2603 2604 void PSParallelCompact::compact() { 2605 EventMark m("5 compact"); 2606 // trace("5"); 2607 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); 2608 2609 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2610 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2611 PSOldGen* old_gen = heap->old_gen(); 2612 old_gen->start_array()->reset(); 2613 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2614 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2615 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2616 2617 GCTaskQueue* q = GCTaskQueue::create(); 2618 enqueue_region_draining_tasks(q, parallel_gc_threads); 2619 enqueue_dense_prefix_tasks(q, parallel_gc_threads); 2620 enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads); 2621 2622 { 2623 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); 2624 2625 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2626 q->enqueue(fin); 2627 2628 gc_task_manager()->add_list(q); 2629 2630 fin->wait_for(); 2631 2632 // We have to release the barrier tasks! 2633 WaitForBarrierGCTask::destroy(fin); 2634 2635 #ifdef ASSERT 2636 // Verify that all regions have been processed before the deferred updates. 2637 // Note that perm_space_id is skipped; this type of verification is not 2638 // valid until the perm gen is compacted by regions. 2639 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2640 verify_complete(SpaceId(id)); 2641 } 2642 #endif 2643 } 2644 2645 { 2646 // Update the deferred objects, if any. Any compaction manager can be used. 2647 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); 2648 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2649 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2650 update_deferred_objects(cm, SpaceId(id)); 2651 } 2652 } 2653 } 2654 2655 #ifdef ASSERT 2656 void PSParallelCompact::verify_complete(SpaceId space_id) { 2657 // All Regions between space bottom() to new_top() should be marked as filled 2658 // and all Regions between new_top() and top() should be available (i.e., 2659 // should have been emptied). 2660 ParallelCompactData& sd = summary_data(); 2661 SpaceInfo si = _space_info[space_id]; 2662 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2663 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2664 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2665 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2666 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2667 2668 bool issued_a_warning = false; 2669 2670 size_t cur_region; 2671 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2672 const RegionData* const c = sd.region(cur_region); 2673 if (!c->completed()) { 2674 warning("region " SIZE_FORMAT " not filled: " 2675 "destination_count=" SIZE_FORMAT, 2676 cur_region, c->destination_count()); 2677 issued_a_warning = true; 2678 } 2679 } 2680 2681 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2682 const RegionData* const c = sd.region(cur_region); 2683 if (!c->available()) { 2684 warning("region " SIZE_FORMAT " not empty: " 2685 "destination_count=" SIZE_FORMAT, 2686 cur_region, c->destination_count()); 2687 issued_a_warning = true; 2688 } 2689 } 2690 2691 if (issued_a_warning) { 2692 print_region_ranges(); 2693 } 2694 } 2695 #endif // #ifdef ASSERT 2696 2697 void PSParallelCompact::compact_serial(ParCompactionManager* cm) { 2698 EventMark m("5 compact serial"); 2699 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty); 2700 2701 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2702 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2703 2704 PSYoungGen* young_gen = heap->young_gen(); 2705 PSOldGen* old_gen = heap->old_gen(); 2706 2707 old_gen->start_array()->reset(); 2708 old_gen->move_and_update(cm); 2709 young_gen->move_and_update(cm); 2710 } 2711 2712 2713 void PSParallelCompact::follow_stack(ParCompactionManager* cm) { 2714 while(!cm->overflow_stack()->is_empty()) { 2715 oop obj = cm->overflow_stack()->pop(); 2716 obj->follow_contents(cm); 2717 } 2718 2719 oop obj; 2720 // obj is a reference!!! 2721 while (cm->marking_stack()->pop_local(obj)) { 2722 // It would be nice to assert about the type of objects we might 2723 // pop, but they can come from anywhere, unfortunately. 2724 obj->follow_contents(cm); 2725 } 2726 } 2727 2728 void 2729 PSParallelCompact::follow_weak_klass_links() { 2730 // All klasses on the revisit stack are marked at this point. 2731 // Update and follow all subklass, sibling and implementor links. 2732 if (PrintRevisitStats) { 2733 gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes()); 2734 } 2735 for (uint i = 0; i < ParallelGCThreads + 1; i++) { 2736 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2737 KeepAliveClosure keep_alive_closure(cm); 2738 int length = cm->revisit_klass_stack()->length(); 2739 if (PrintRevisitStats) { 2740 gclog_or_tty->print_cr("Revisit klass stack[%d] length = %d", i, length); 2741 } 2742 for (int j = 0; j < length; j++) { 2743 cm->revisit_klass_stack()->at(j)->follow_weak_klass_links( 2744 is_alive_closure(), 2745 &keep_alive_closure); 2746 } 2747 // revisit_klass_stack is cleared in reset() 2748 follow_stack(cm); 2749 } 2750 } 2751 2752 void 2753 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) { 2754 cm->revisit_klass_stack()->push(k); 2755 } 2756 2757 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) { 2758 cm->revisit_mdo_stack()->push(p); 2759 } 2760 2761 void PSParallelCompact::follow_mdo_weak_refs() { 2762 // All strongly reachable oops have been marked at this point; 2763 // we can visit and clear any weak references from MDO's which 2764 // we memoized during the strong marking phase. 2765 if (PrintRevisitStats) { 2766 gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes()); 2767 } 2768 for (uint i = 0; i < ParallelGCThreads + 1; i++) { 2769 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2770 GrowableArray<DataLayout*>* rms = cm->revisit_mdo_stack(); 2771 int length = rms->length(); 2772 if (PrintRevisitStats) { 2773 gclog_or_tty->print_cr("Revisit MDO stack[%d] length = %d", i, length); 2774 } 2775 for (int j = 0; j < length; j++) { 2776 rms->at(j)->follow_weak_refs(is_alive_closure()); 2777 } 2778 // revisit_mdo_stack is cleared in reset() 2779 follow_stack(cm); 2780 } 2781 } 2782 2783 2784 #ifdef VALIDATE_MARK_SWEEP 2785 2786 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) { 2787 if (!ValidateMarkSweep) 2788 return; 2789 2790 if (!isroot) { 2791 if (_pointer_tracking) { 2792 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer"); 2793 _adjusted_pointers->remove(p); 2794 } 2795 } else { 2796 ptrdiff_t index = _root_refs_stack->find(p); 2797 if (index != -1) { 2798 int l = _root_refs_stack->length(); 2799 if (l > 0 && l - 1 != index) { 2800 void* last = _root_refs_stack->pop(); 2801 assert(last != p, "should be different"); 2802 _root_refs_stack->at_put(index, last); 2803 } else { 2804 _root_refs_stack->remove(p); 2805 } 2806 } 2807 } 2808 } 2809 2810 2811 void PSParallelCompact::check_adjust_pointer(void* p) { 2812 _adjusted_pointers->push(p); 2813 } 2814 2815 2816 class AdjusterTracker: public OopClosure { 2817 public: 2818 AdjusterTracker() {}; 2819 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); } 2820 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); } 2821 }; 2822 2823 2824 void PSParallelCompact::track_interior_pointers(oop obj) { 2825 if (ValidateMarkSweep) { 2826 _adjusted_pointers->clear(); 2827 _pointer_tracking = true; 2828 2829 AdjusterTracker checker; 2830 obj->oop_iterate(&checker); 2831 } 2832 } 2833 2834 2835 void PSParallelCompact::check_interior_pointers() { 2836 if (ValidateMarkSweep) { 2837 _pointer_tracking = false; 2838 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers"); 2839 } 2840 } 2841 2842 2843 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) { 2844 if (ValidateMarkSweep) { 2845 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops"); 2846 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0; 2847 } 2848 } 2849 2850 2851 void PSParallelCompact::register_live_oop(oop p, size_t size) { 2852 if (ValidateMarkSweep) { 2853 _live_oops->push(p); 2854 _live_oops_size->push(size); 2855 _live_oops_index++; 2856 } 2857 } 2858 2859 void PSParallelCompact::validate_live_oop(oop p, size_t size) { 2860 if (ValidateMarkSweep) { 2861 oop obj = _live_oops->at((int)_live_oops_index); 2862 guarantee(obj == p, "should be the same object"); 2863 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size"); 2864 _live_oops_index++; 2865 } 2866 } 2867 2868 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size, 2869 HeapWord* compaction_top) { 2870 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top), 2871 "should be moved to forwarded location"); 2872 if (ValidateMarkSweep) { 2873 PSParallelCompact::validate_live_oop(oop(q), size); 2874 _live_oops_moved_to->push(oop(compaction_top)); 2875 } 2876 if (RecordMarkSweepCompaction) { 2877 _cur_gc_live_oops->push(q); 2878 _cur_gc_live_oops_moved_to->push(compaction_top); 2879 _cur_gc_live_oops_size->push(size); 2880 } 2881 } 2882 2883 2884 void PSParallelCompact::compaction_complete() { 2885 if (RecordMarkSweepCompaction) { 2886 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops; 2887 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to; 2888 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size; 2889 2890 _cur_gc_live_oops = _last_gc_live_oops; 2891 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to; 2892 _cur_gc_live_oops_size = _last_gc_live_oops_size; 2893 _last_gc_live_oops = _tmp_live_oops; 2894 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to; 2895 _last_gc_live_oops_size = _tmp_live_oops_size; 2896 } 2897 } 2898 2899 2900 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) { 2901 if (!RecordMarkSweepCompaction) { 2902 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled"); 2903 return; 2904 } 2905 2906 if (_last_gc_live_oops == NULL) { 2907 tty->print_cr("No compaction information gathered yet"); 2908 return; 2909 } 2910 2911 for (int i = 0; i < _last_gc_live_oops->length(); i++) { 2912 HeapWord* old_oop = _last_gc_live_oops->at(i); 2913 size_t sz = _last_gc_live_oops_size->at(i); 2914 if (old_oop <= q && q < (old_oop + sz)) { 2915 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i); 2916 size_t offset = (q - old_oop); 2917 tty->print_cr("Address " PTR_FORMAT, q); 2918 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset); 2919 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset); 2920 return; 2921 } 2922 } 2923 2924 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q); 2925 } 2926 #endif //VALIDATE_MARK_SWEEP 2927 2928 // Update interior oops in the ranges of regions [beg_region, end_region). 2929 void 2930 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2931 SpaceId space_id, 2932 size_t beg_region, 2933 size_t end_region) { 2934 ParallelCompactData& sd = summary_data(); 2935 ParMarkBitMap* const mbm = mark_bitmap(); 2936 2937 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2938 HeapWord* const end_addr = sd.region_to_addr(end_region); 2939 assert(beg_region <= end_region, "bad region range"); 2940 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2941 2942 #ifdef ASSERT 2943 // Claim the regions to avoid triggering an assert when they are marked as 2944 // filled. 2945 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2946 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2947 } 2948 #endif // #ifdef ASSERT 2949 2950 if (beg_addr != space(space_id)->bottom()) { 2951 // Find the first live object or block of dead space that *starts* in this 2952 // range of regions. If a partial object crosses onto the region, skip it; 2953 // it will be marked for 'deferred update' when the object head is 2954 // processed. If dead space crosses onto the region, it is also skipped; it 2955 // will be filled when the prior region is processed. If neither of those 2956 // apply, the first word in the region is the start of a live object or dead 2957 // space. 2958 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2959 const RegionData* const cp = sd.region(beg_region); 2960 if (cp->partial_obj_size() != 0) { 2961 beg_addr = sd.partial_obj_end(beg_region); 2962 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2963 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2964 } 2965 } 2966 2967 if (beg_addr < end_addr) { 2968 // A live object or block of dead space starts in this range of Regions. 2969 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2970 2971 // Create closures and iterate. 2972 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2973 FillClosure fill_closure(cm, space_id); 2974 ParMarkBitMap::IterationStatus status; 2975 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2976 dense_prefix_end); 2977 if (status == ParMarkBitMap::incomplete) { 2978 update_closure.do_addr(update_closure.source()); 2979 } 2980 } 2981 2982 // Mark the regions as filled. 2983 RegionData* const beg_cp = sd.region(beg_region); 2984 RegionData* const end_cp = sd.region(end_region); 2985 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2986 cp->set_completed(); 2987 } 2988 } 2989 2990 // Return the SpaceId for the space containing addr. If addr is not in the 2991 // heap, last_space_id is returned. In debug mode it expects the address to be 2992 // in the heap and asserts such. 2993 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2994 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2995 2996 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 2997 if (_space_info[id].space()->contains(addr)) { 2998 return SpaceId(id); 2999 } 3000 } 3001 3002 assert(false, "no space contains the addr"); 3003 return last_space_id; 3004 } 3005 3006 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 3007 SpaceId id) { 3008 assert(id < last_space_id, "bad space id"); 3009 3010 ParallelCompactData& sd = summary_data(); 3011 const SpaceInfo* const space_info = _space_info + id; 3012 ObjectStartArray* const start_array = space_info->start_array(); 3013 3014 const MutableSpace* const space = space_info->space(); 3015 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 3016 HeapWord* const beg_addr = space_info->dense_prefix(); 3017 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 3018 3019 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 3020 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 3021 const RegionData* cur_region; 3022 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 3023 HeapWord* const addr = cur_region->deferred_obj_addr(); 3024 if (addr != NULL) { 3025 if (start_array != NULL) { 3026 start_array->allocate_block(addr); 3027 } 3028 oop(addr)->update_contents(cm); 3029 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 3030 } 3031 } 3032 } 3033 3034 // Skip over count live words starting from beg, and return the address of the 3035 // next live word. Unless marked, the word corresponding to beg is assumed to 3036 // be dead. Callers must either ensure beg does not correspond to the middle of 3037 // an object, or account for those live words in some other way. Callers must 3038 // also ensure that there are enough live words in the range [beg, end) to skip. 3039 HeapWord* 3040 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 3041 { 3042 assert(count > 0, "sanity"); 3043 3044 ParMarkBitMap* m = mark_bitmap(); 3045 idx_t bits_to_skip = m->words_to_bits(count); 3046 idx_t cur_beg = m->addr_to_bit(beg); 3047 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 3048 3049 do { 3050 cur_beg = m->find_obj_beg(cur_beg, search_end); 3051 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 3052 const size_t obj_bits = cur_end - cur_beg + 1; 3053 if (obj_bits > bits_to_skip) { 3054 return m->bit_to_addr(cur_beg + bits_to_skip); 3055 } 3056 bits_to_skip -= obj_bits; 3057 cur_beg = cur_end + 1; 3058 } while (bits_to_skip > 0); 3059 3060 // Skipping the desired number of words landed just past the end of an object. 3061 // Find the start of the next object. 3062 cur_beg = m->find_obj_beg(cur_beg, search_end); 3063 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 3064 return m->bit_to_addr(cur_beg); 3065 } 3066 3067 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 3068 SpaceId src_space_id, 3069 size_t src_region_idx) 3070 { 3071 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 3072 3073 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 3074 if (split_info.dest_region_addr() == dest_addr) { 3075 // The partial object ending at the split point contains the first word to 3076 // be copied to dest_addr. 3077 return split_info.first_src_addr(); 3078 } 3079 3080 const ParallelCompactData& sd = summary_data(); 3081 ParMarkBitMap* const bitmap = mark_bitmap(); 3082 const size_t RegionSize = ParallelCompactData::RegionSize; 3083 3084 assert(sd.is_region_aligned(dest_addr), "not aligned"); 3085 const RegionData* const src_region_ptr = sd.region(src_region_idx); 3086 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 3087 HeapWord* const src_region_destination = src_region_ptr->destination(); 3088 3089 assert(dest_addr >= src_region_destination, "wrong src region"); 3090 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 3091 3092 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 3093 HeapWord* const src_region_end = src_region_beg + RegionSize; 3094 3095 HeapWord* addr = src_region_beg; 3096 if (dest_addr == src_region_destination) { 3097 // Return the first live word in the source region. 3098 if (partial_obj_size == 0) { 3099 addr = bitmap->find_obj_beg(addr, src_region_end); 3100 assert(addr < src_region_end, "no objects start in src region"); 3101 } 3102 return addr; 3103 } 3104 3105 // Must skip some live data. 3106 size_t words_to_skip = dest_addr - src_region_destination; 3107 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 3108 3109 if (partial_obj_size >= words_to_skip) { 3110 // All the live words to skip are part of the partial object. 3111 addr += words_to_skip; 3112 if (partial_obj_size == words_to_skip) { 3113 // Find the first live word past the partial object. 3114 addr = bitmap->find_obj_beg(addr, src_region_end); 3115 assert(addr < src_region_end, "wrong src region"); 3116 } 3117 return addr; 3118 } 3119 3120 // Skip over the partial object (if any). 3121 if (partial_obj_size != 0) { 3122 words_to_skip -= partial_obj_size; 3123 addr += partial_obj_size; 3124 } 3125 3126 // Skip over live words due to objects that start in the region. 3127 addr = skip_live_words(addr, src_region_end, words_to_skip); 3128 assert(addr < src_region_end, "wrong src region"); 3129 return addr; 3130 } 3131 3132 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 3133 SpaceId src_space_id, 3134 size_t beg_region, 3135 HeapWord* end_addr) 3136 { 3137 ParallelCompactData& sd = summary_data(); 3138 3139 #ifdef ASSERT 3140 MutableSpace* const src_space = _space_info[src_space_id].space(); 3141 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 3142 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 3143 "src_space_id does not match beg_addr"); 3144 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 3145 "src_space_id does not match end_addr"); 3146 #endif // #ifdef ASSERT 3147 3148 RegionData* const beg = sd.region(beg_region); 3149 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 3150 3151 // Regions up to new_top() are enqueued if they become available. 3152 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3153 RegionData* const enqueue_end = 3154 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3155 3156 for (RegionData* cur = beg; cur < end; ++cur) { 3157 assert(cur->data_size() > 0, "region must have live data"); 3158 cur->decrement_destination_count(); 3159 if (cur < enqueue_end && cur->available() && cur->claim()) { 3160 cm->save_for_processing(sd.region(cur)); 3161 } 3162 } 3163 } 3164 3165 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3166 SpaceId& src_space_id, 3167 HeapWord*& src_space_top, 3168 HeapWord* end_addr) 3169 { 3170 typedef ParallelCompactData::RegionData RegionData; 3171 3172 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3173 const size_t region_size = ParallelCompactData::RegionSize; 3174 3175 size_t src_region_idx = 0; 3176 3177 // Skip empty regions (if any) up to the top of the space. 3178 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3179 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3180 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3181 const RegionData* const top_region_ptr = 3182 sd.addr_to_region_ptr(top_aligned_up); 3183 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3184 ++src_region_ptr; 3185 } 3186 3187 if (src_region_ptr < top_region_ptr) { 3188 // The next source region is in the current space. Update src_region_idx 3189 // and the source address to match src_region_ptr. 3190 src_region_idx = sd.region(src_region_ptr); 3191 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3192 if (src_region_addr > closure.source()) { 3193 closure.set_source(src_region_addr); 3194 } 3195 return src_region_idx; 3196 } 3197 3198 // Switch to a new source space and find the first non-empty region. 3199 unsigned int space_id = src_space_id + 1; 3200 assert(space_id < last_space_id, "not enough spaces"); 3201 3202 HeapWord* const destination = closure.destination(); 3203 3204 do { 3205 MutableSpace* space = _space_info[space_id].space(); 3206 HeapWord* const bottom = space->bottom(); 3207 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3208 3209 // Iterate over the spaces that do not compact into themselves. 3210 if (bottom_cp->destination() != bottom) { 3211 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3212 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3213 3214 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3215 if (src_cp->live_obj_size() > 0) { 3216 // Found it. 3217 assert(src_cp->destination() == destination, 3218 "first live obj in the space must match the destination"); 3219 assert(src_cp->partial_obj_size() == 0, 3220 "a space cannot begin with a partial obj"); 3221 3222 src_space_id = SpaceId(space_id); 3223 src_space_top = space->top(); 3224 const size_t src_region_idx = sd.region(src_cp); 3225 closure.set_source(sd.region_to_addr(src_region_idx)); 3226 return src_region_idx; 3227 } else { 3228 assert(src_cp->data_size() == 0, "sanity"); 3229 } 3230 } 3231 } 3232 } while (++space_id < last_space_id); 3233 3234 assert(false, "no source region was found"); 3235 return 0; 3236 } 3237 3238 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3239 { 3240 typedef ParMarkBitMap::IterationStatus IterationStatus; 3241 const size_t RegionSize = ParallelCompactData::RegionSize; 3242 ParMarkBitMap* const bitmap = mark_bitmap(); 3243 ParallelCompactData& sd = summary_data(); 3244 RegionData* const region_ptr = sd.region(region_idx); 3245 3246 // Get the items needed to construct the closure. 3247 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3248 SpaceId dest_space_id = space_id(dest_addr); 3249 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3250 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3251 assert(dest_addr < new_top, "sanity"); 3252 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3253 3254 // Get the source region and related info. 3255 size_t src_region_idx = region_ptr->source_region(); 3256 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3257 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3258 3259 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3260 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3261 3262 // Adjust src_region_idx to prepare for decrementing destination counts (the 3263 // destination count is not decremented when a region is copied to itself). 3264 if (src_region_idx == region_idx) { 3265 src_region_idx += 1; 3266 } 3267 3268 if (bitmap->is_unmarked(closure.source())) { 3269 // The first source word is in the middle of an object; copy the remainder 3270 // of the object or as much as will fit. The fact that pointer updates were 3271 // deferred will be noted when the object header is processed. 3272 HeapWord* const old_src_addr = closure.source(); 3273 closure.copy_partial_obj(); 3274 if (closure.is_full()) { 3275 decrement_destination_counts(cm, src_space_id, src_region_idx, 3276 closure.source()); 3277 region_ptr->set_deferred_obj_addr(NULL); 3278 region_ptr->set_completed(); 3279 return; 3280 } 3281 3282 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3283 if (sd.region_align_down(old_src_addr) != end_addr) { 3284 // The partial object was copied from more than one source region. 3285 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3286 3287 // Move to the next source region, possibly switching spaces as well. All 3288 // args except end_addr may be modified. 3289 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3290 end_addr); 3291 } 3292 } 3293 3294 do { 3295 HeapWord* const cur_addr = closure.source(); 3296 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3297 src_space_top); 3298 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3299 3300 if (status == ParMarkBitMap::incomplete) { 3301 // The last obj that starts in the source region does not end in the 3302 // region. 3303 assert(closure.source() < end_addr, "sanity") 3304 HeapWord* const obj_beg = closure.source(); 3305 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3306 src_space_top); 3307 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3308 if (obj_end < range_end) { 3309 // The end was found; the entire object will fit. 3310 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3311 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3312 } else { 3313 // The end was not found; the object will not fit. 3314 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3315 status = ParMarkBitMap::would_overflow; 3316 } 3317 } 3318 3319 if (status == ParMarkBitMap::would_overflow) { 3320 // The last object did not fit. Note that interior oop updates were 3321 // deferred, then copy enough of the object to fill the region. 3322 region_ptr->set_deferred_obj_addr(closure.destination()); 3323 status = closure.copy_until_full(); // copies from closure.source() 3324 3325 decrement_destination_counts(cm, src_space_id, src_region_idx, 3326 closure.source()); 3327 region_ptr->set_completed(); 3328 return; 3329 } 3330 3331 if (status == ParMarkBitMap::full) { 3332 decrement_destination_counts(cm, src_space_id, src_region_idx, 3333 closure.source()); 3334 region_ptr->set_deferred_obj_addr(NULL); 3335 region_ptr->set_completed(); 3336 return; 3337 } 3338 3339 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3340 3341 // Move to the next source region, possibly switching spaces as well. All 3342 // args except end_addr may be modified. 3343 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3344 end_addr); 3345 } while (true); 3346 } 3347 3348 void 3349 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3350 const MutableSpace* sp = space(space_id); 3351 if (sp->is_empty()) { 3352 return; 3353 } 3354 3355 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3356 ParMarkBitMap* const bitmap = mark_bitmap(); 3357 HeapWord* const dp_addr = dense_prefix(space_id); 3358 HeapWord* beg_addr = sp->bottom(); 3359 HeapWord* end_addr = sp->top(); 3360 3361 #ifdef ASSERT 3362 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3363 if (cm->should_verify_only()) { 3364 VerifyUpdateClosure verify_update(cm, sp); 3365 bitmap->iterate(&verify_update, beg_addr, end_addr); 3366 return; 3367 } 3368 3369 if (cm->should_reset_only()) { 3370 ResetObjectsClosure reset_objects(cm); 3371 bitmap->iterate(&reset_objects, beg_addr, end_addr); 3372 return; 3373 } 3374 #endif 3375 3376 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3377 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3378 if (beg_region < dp_region) { 3379 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3380 } 3381 3382 // The destination of the first live object that starts in the region is one 3383 // past the end of the partial object entering the region (if any). 3384 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3385 HeapWord* const new_top = _space_info[space_id].new_top(); 3386 assert(new_top >= dest_addr, "bad new_top value"); 3387 const size_t words = pointer_delta(new_top, dest_addr); 3388 3389 if (words > 0) { 3390 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3391 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3392 3393 ParMarkBitMap::IterationStatus status; 3394 status = bitmap->iterate(&closure, dest_addr, end_addr); 3395 assert(status == ParMarkBitMap::full, "iteration not complete"); 3396 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3397 "live objects skipped because closure is full"); 3398 } 3399 } 3400 3401 jlong PSParallelCompact::millis_since_last_gc() { 3402 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc; 3403 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3404 if (ret_val < 0) { 3405 NOT_PRODUCT(warning("time warp: %d", ret_val);) 3406 return 0; 3407 } 3408 return ret_val; 3409 } 3410 3411 void PSParallelCompact::reset_millis_since_last_gc() { 3412 _time_of_last_gc = os::javaTimeMillis(); 3413 } 3414 3415 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3416 { 3417 if (source() != destination()) { 3418 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3419 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3420 } 3421 update_state(words_remaining()); 3422 assert(is_full(), "sanity"); 3423 return ParMarkBitMap::full; 3424 } 3425 3426 void MoveAndUpdateClosure::copy_partial_obj() 3427 { 3428 size_t words = words_remaining(); 3429 3430 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3431 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3432 if (end_addr < range_end) { 3433 words = bitmap()->obj_size(source(), end_addr); 3434 } 3435 3436 // This test is necessary; if omitted, the pointer updates to a partial object 3437 // that crosses the dense prefix boundary could be overwritten. 3438 if (source() != destination()) { 3439 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3440 Copy::aligned_conjoint_words(source(), destination(), words); 3441 } 3442 update_state(words); 3443 } 3444 3445 ParMarkBitMapClosure::IterationStatus 3446 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3447 assert(destination() != NULL, "sanity"); 3448 assert(bitmap()->obj_size(addr) == words, "bad size"); 3449 3450 _source = addr; 3451 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3452 destination(), "wrong destination"); 3453 3454 if (words > words_remaining()) { 3455 return ParMarkBitMap::would_overflow; 3456 } 3457 3458 // The start_array must be updated even if the object is not moving. 3459 if (_start_array != NULL) { 3460 _start_array->allocate_block(destination()); 3461 } 3462 3463 if (destination() != source()) { 3464 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3465 Copy::aligned_conjoint_words(source(), destination(), words); 3466 } 3467 3468 oop moved_oop = (oop) destination(); 3469 moved_oop->update_contents(compaction_manager()); 3470 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3471 3472 update_state(words); 3473 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3474 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3475 } 3476 3477 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3478 ParCompactionManager* cm, 3479 PSParallelCompact::SpaceId space_id) : 3480 ParMarkBitMapClosure(mbm, cm), 3481 _space_id(space_id), 3482 _start_array(PSParallelCompact::start_array(space_id)) 3483 { 3484 } 3485 3486 // Updates the references in the object to their new values. 3487 ParMarkBitMapClosure::IterationStatus 3488 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3489 do_addr(addr); 3490 return ParMarkBitMap::incomplete; 3491 } 3492 3493 // Verify the new location using the forwarding pointer 3494 // from MarkSweep::mark_sweep_phase2(). Set the mark_word 3495 // to the initial value. 3496 ParMarkBitMapClosure::IterationStatus 3497 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3498 // The second arg (words) is not used. 3499 oop obj = (oop) addr; 3500 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer(); 3501 HeapWord* new_pointer = summary_data().calc_new_pointer(obj); 3502 if (forwarding_ptr == NULL) { 3503 // The object is dead or not moving. 3504 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj), 3505 "Object liveness is wrong."); 3506 return ParMarkBitMap::incomplete; 3507 } 3508 assert(UseParallelOldGCDensePrefix || 3509 (HeapMaximumCompactionInterval > 1) || 3510 (MarkSweepAlwaysCompactCount > 1) || 3511 (forwarding_ptr == new_pointer), 3512 "Calculation of new location is incorrect"); 3513 return ParMarkBitMap::incomplete; 3514 } 3515 3516 // Reset objects modified for debug checking. 3517 ParMarkBitMapClosure::IterationStatus 3518 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) { 3519 // The second arg (words) is not used. 3520 oop obj = (oop) addr; 3521 obj->init_mark(); 3522 return ParMarkBitMap::incomplete; 3523 } 3524 3525 // Prepare for compaction. This method is executed once 3526 // (i.e., by a single thread) before compaction. 3527 // Save the updated location of the intArrayKlassObj for 3528 // filling holes in the dense prefix. 3529 void PSParallelCompact::compact_prologue() { 3530 _updated_int_array_klass_obj = (klassOop) 3531 summary_data().calc_new_pointer(Universe::intArrayKlassObj()); 3532 } 3533