1 /* 2 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "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 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 715 return result; 716 } 717 718 // The new location of the object is 719 // region destination + 720 // size of the partial object extending onto the region + 721 // sizes of the live objects in the Region that are to the left of addr 722 const size_t partial_obj_size = region_ptr->partial_obj_size(); 723 HeapWord* const search_start = region_addr + partial_obj_size; 724 725 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap(); 726 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr)); 727 728 result += partial_obj_size + live_to_left; 729 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);) 730 return result; 731 } 732 733 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) { 734 klassOop updated_klass; 735 if (PSParallelCompact::should_update_klass(old_klass)) { 736 updated_klass = (klassOop) calc_new_pointer(old_klass); 737 } else { 738 updated_klass = old_klass; 739 } 740 741 return updated_klass; 742 } 743 744 #ifdef ASSERT 745 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace) 746 { 747 const size_t* const beg = (const size_t*)vspace->committed_low_addr(); 748 const size_t* const end = (const size_t*)vspace->committed_high_addr(); 749 for (const size_t* p = beg; p < end; ++p) { 750 assert(*p == 0, "not zero"); 751 } 752 } 753 754 void ParallelCompactData::verify_clear() 755 { 756 verify_clear(_region_vspace); 757 } 758 #endif // #ifdef ASSERT 759 760 #ifdef NOT_PRODUCT 761 ParallelCompactData::RegionData* debug_region(size_t region_index) { 762 ParallelCompactData& sd = PSParallelCompact::summary_data(); 763 return sd.region(region_index); 764 } 765 #endif 766 767 elapsedTimer PSParallelCompact::_accumulated_time; 768 unsigned int PSParallelCompact::_total_invocations = 0; 769 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0; 770 jlong PSParallelCompact::_time_of_last_gc = 0; 771 CollectorCounters* PSParallelCompact::_counters = NULL; 772 ParMarkBitMap PSParallelCompact::_mark_bitmap; 773 ParallelCompactData PSParallelCompact::_summary_data; 774 775 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure; 776 777 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); } 778 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); } 779 780 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 781 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); } 782 783 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true); 784 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false); 785 786 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); } 787 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); } 788 789 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); } 790 791 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); } 792 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); } 793 794 void PSParallelCompact::post_initialize() { 795 ParallelScavengeHeap* heap = gc_heap(); 796 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 797 798 MemRegion mr = heap->reserved_region(); 799 _ref_processor = ReferenceProcessor::create_ref_processor( 800 mr, // span 801 true, // atomic_discovery 802 true, // mt_discovery 803 &_is_alive_closure, 804 ParallelGCThreads, 805 ParallelRefProcEnabled); 806 _counters = new CollectorCounters("PSParallelCompact", 1); 807 808 // Initialize static fields in ParCompactionManager. 809 ParCompactionManager::initialize(mark_bitmap()); 810 } 811 812 bool PSParallelCompact::initialize() { 813 ParallelScavengeHeap* heap = gc_heap(); 814 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 815 MemRegion mr = heap->reserved_region(); 816 817 // Was the old gen get allocated successfully? 818 if (!heap->old_gen()->is_allocated()) { 819 return false; 820 } 821 822 initialize_space_info(); 823 initialize_dead_wood_limiter(); 824 825 if (!_mark_bitmap.initialize(mr)) { 826 vm_shutdown_during_initialization("Unable to allocate bit map for " 827 "parallel garbage collection for the requested heap size."); 828 return false; 829 } 830 831 if (!_summary_data.initialize(mr)) { 832 vm_shutdown_during_initialization("Unable to allocate tables for " 833 "parallel garbage collection for the requested heap size."); 834 return false; 835 } 836 837 return true; 838 } 839 840 void PSParallelCompact::initialize_space_info() 841 { 842 memset(&_space_info, 0, sizeof(_space_info)); 843 844 ParallelScavengeHeap* heap = gc_heap(); 845 PSYoungGen* young_gen = heap->young_gen(); 846 MutableSpace* perm_space = heap->perm_gen()->object_space(); 847 848 _space_info[perm_space_id].set_space(perm_space); 849 _space_info[old_space_id].set_space(heap->old_gen()->object_space()); 850 _space_info[eden_space_id].set_space(young_gen->eden_space()); 851 _space_info[from_space_id].set_space(young_gen->from_space()); 852 _space_info[to_space_id].set_space(young_gen->to_space()); 853 854 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array()); 855 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array()); 856 857 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top()); 858 if (TraceParallelOldGCDensePrefix) { 859 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT, 860 _space_info[perm_space_id].min_dense_prefix()); 861 } 862 } 863 864 void PSParallelCompact::initialize_dead_wood_limiter() 865 { 866 const size_t max = 100; 867 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0; 868 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0; 869 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev); 870 DEBUG_ONLY(_dwl_initialized = true;) 871 _dwl_adjustment = normal_distribution(1.0); 872 } 873 874 // Simple class for storing info about the heap at the start of GC, to be used 875 // after GC for comparison/printing. 876 class PreGCValues { 877 public: 878 PreGCValues() { } 879 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); } 880 881 void fill(ParallelScavengeHeap* heap) { 882 _heap_used = heap->used(); 883 _young_gen_used = heap->young_gen()->used_in_bytes(); 884 _old_gen_used = heap->old_gen()->used_in_bytes(); 885 _perm_gen_used = heap->perm_gen()->used_in_bytes(); 886 }; 887 888 size_t heap_used() const { return _heap_used; } 889 size_t young_gen_used() const { return _young_gen_used; } 890 size_t old_gen_used() const { return _old_gen_used; } 891 size_t perm_gen_used() const { return _perm_gen_used; } 892 893 private: 894 size_t _heap_used; 895 size_t _young_gen_used; 896 size_t _old_gen_used; 897 size_t _perm_gen_used; 898 }; 899 900 void 901 PSParallelCompact::clear_data_covering_space(SpaceId id) 902 { 903 // At this point, top is the value before GC, new_top() is the value that will 904 // be set at the end of GC. The marking bitmap is cleared to top; nothing 905 // should be marked above top. The summary data is cleared to the larger of 906 // top & new_top. 907 MutableSpace* const space = _space_info[id].space(); 908 HeapWord* const bot = space->bottom(); 909 HeapWord* const top = space->top(); 910 HeapWord* const max_top = MAX2(top, _space_info[id].new_top()); 911 912 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot); 913 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top)); 914 _mark_bitmap.clear_range(beg_bit, end_bit); 915 916 const size_t beg_region = _summary_data.addr_to_region_idx(bot); 917 const size_t end_region = 918 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top)); 919 _summary_data.clear_range(beg_region, end_region); 920 921 // Clear the data used to 'split' regions. 922 SplitInfo& split_info = _space_info[id].split_info(); 923 if (split_info.is_valid()) { 924 split_info.clear(); 925 } 926 DEBUG_ONLY(split_info.verify_clear();) 927 } 928 929 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values) 930 { 931 // Update the from & to space pointers in space_info, since they are swapped 932 // at each young gen gc. Do the update unconditionally (even though a 933 // promotion failure does not swap spaces) because an unknown number of minor 934 // collections will have swapped the spaces an unknown number of times. 935 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty); 936 ParallelScavengeHeap* heap = gc_heap(); 937 _space_info[from_space_id].set_space(heap->young_gen()->from_space()); 938 _space_info[to_space_id].set_space(heap->young_gen()->to_space()); 939 940 pre_gc_values->fill(heap); 941 942 ParCompactionManager::reset(); 943 NOT_PRODUCT(_mark_bitmap.reset_counters()); 944 DEBUG_ONLY(add_obj_count = add_obj_size = 0;) 945 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;) 946 947 // Increment the invocation count 948 heap->increment_total_collections(true); 949 950 // We need to track unique mark sweep invocations as well. 951 _total_invocations++; 952 953 if (PrintHeapAtGC) { 954 Universe::print_heap_before_gc(); 955 } 956 957 // Fill in TLABs 958 heap->accumulate_statistics_all_tlabs(); 959 heap->ensure_parsability(true); // retire TLABs 960 961 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) { 962 HandleMark hm; // Discard invalid handles created during verification 963 gclog_or_tty->print(" VerifyBeforeGC:"); 964 Universe::verify(true); 965 } 966 967 // Verify object start arrays 968 if (VerifyObjectStartArray && 969 VerifyBeforeGC) { 970 heap->old_gen()->verify_object_start_array(); 971 heap->perm_gen()->verify_object_start_array(); 972 } 973 974 DEBUG_ONLY(mark_bitmap()->verify_clear();) 975 DEBUG_ONLY(summary_data().verify_clear();) 976 977 // Have worker threads release resources the next time they run a task. 978 gc_task_manager()->release_all_resources(); 979 } 980 981 void PSParallelCompact::post_compact() 982 { 983 TraceTime tm("post compact", print_phases(), true, gclog_or_tty); 984 985 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 986 // Clear the marking bitmap, summary data and split info. 987 clear_data_covering_space(SpaceId(id)); 988 // Update top(). Must be done after clearing the bitmap and summary data. 989 _space_info[id].publish_new_top(); 990 } 991 992 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 993 MutableSpace* const from_space = _space_info[from_space_id].space(); 994 MutableSpace* const to_space = _space_info[to_space_id].space(); 995 996 ParallelScavengeHeap* heap = gc_heap(); 997 bool eden_empty = eden_space->is_empty(); 998 if (!eden_empty) { 999 eden_empty = absorb_live_data_from_eden(heap->size_policy(), 1000 heap->young_gen(), heap->old_gen()); 1001 } 1002 1003 // Update heap occupancy information which is used as input to the soft ref 1004 // clearing policy at the next gc. 1005 Universe::update_heap_info_at_gc(); 1006 1007 bool young_gen_empty = eden_empty && from_space->is_empty() && 1008 to_space->is_empty(); 1009 1010 BarrierSet* bs = heap->barrier_set(); 1011 if (bs->is_a(BarrierSet::ModRef)) { 1012 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs; 1013 MemRegion old_mr = heap->old_gen()->reserved(); 1014 MemRegion perm_mr = heap->perm_gen()->reserved(); 1015 assert(perm_mr.end() <= old_mr.start(), "Generations out of order"); 1016 1017 if (young_gen_empty) { 1018 modBS->clear(MemRegion(perm_mr.start(), old_mr.end())); 1019 } else { 1020 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end())); 1021 } 1022 } 1023 1024 Threads::gc_epilogue(); 1025 CodeCache::gc_epilogue(); 1026 1027 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1028 1029 ref_processor()->enqueue_discovered_references(NULL); 1030 1031 if (ZapUnusedHeapArea) { 1032 heap->gen_mangle_unused_area(); 1033 } 1034 1035 // Update time of last GC 1036 reset_millis_since_last_gc(); 1037 } 1038 1039 HeapWord* 1040 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id, 1041 bool maximum_compaction) 1042 { 1043 const size_t region_size = ParallelCompactData::RegionSize; 1044 const ParallelCompactData& sd = summary_data(); 1045 1046 const MutableSpace* const space = _space_info[id].space(); 1047 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 1048 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom()); 1049 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up); 1050 1051 // Skip full regions at the beginning of the space--they are necessarily part 1052 // of the dense prefix. 1053 size_t full_count = 0; 1054 const RegionData* cp; 1055 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) { 1056 ++full_count; 1057 } 1058 1059 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1060 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1061 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval; 1062 if (maximum_compaction || cp == end_cp || interval_ended) { 1063 _maximum_compaction_gc_num = total_invocations(); 1064 return sd.region_to_addr(cp); 1065 } 1066 1067 HeapWord* const new_top = _space_info[id].new_top(); 1068 const size_t space_live = pointer_delta(new_top, space->bottom()); 1069 const size_t space_used = space->used_in_words(); 1070 const size_t space_capacity = space->capacity_in_words(); 1071 1072 const double cur_density = double(space_live) / space_capacity; 1073 const double deadwood_density = 1074 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density; 1075 const size_t deadwood_goal = size_t(space_capacity * deadwood_density); 1076 1077 if (TraceParallelOldGCDensePrefix) { 1078 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT, 1079 cur_density, deadwood_density, deadwood_goal); 1080 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1081 "space_cap=" SIZE_FORMAT, 1082 space_live, space_used, 1083 space_capacity); 1084 } 1085 1086 // XXX - Use binary search? 1087 HeapWord* dense_prefix = sd.region_to_addr(cp); 1088 const RegionData* full_cp = cp; 1089 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1); 1090 while (cp < end_cp) { 1091 HeapWord* region_destination = cp->destination(); 1092 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination); 1093 if (TraceParallelOldGCDensePrefix && Verbose) { 1094 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " " 1095 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8), 1096 sd.region(cp), region_destination, 1097 dense_prefix, cur_deadwood); 1098 } 1099 1100 if (cur_deadwood >= deadwood_goal) { 1101 // Found the region that has the correct amount of deadwood to the left. 1102 // This typically occurs after crossing a fairly sparse set of regions, so 1103 // iterate backwards over those sparse regions, looking for the region 1104 // that has the lowest density of live objects 'to the right.' 1105 size_t space_to_left = sd.region(cp) * region_size; 1106 size_t live_to_left = space_to_left - cur_deadwood; 1107 size_t space_to_right = space_capacity - space_to_left; 1108 size_t live_to_right = space_live - live_to_left; 1109 double density_to_right = double(live_to_right) / space_to_right; 1110 while (cp > full_cp) { 1111 --cp; 1112 const size_t prev_region_live_to_right = live_to_right - 1113 cp->data_size(); 1114 const size_t prev_region_space_to_right = space_to_right + region_size; 1115 double prev_region_density_to_right = 1116 double(prev_region_live_to_right) / prev_region_space_to_right; 1117 if (density_to_right <= prev_region_density_to_right) { 1118 return dense_prefix; 1119 } 1120 if (TraceParallelOldGCDensePrefix && Verbose) { 1121 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f " 1122 "pc_d2r=%10.8f", sd.region(cp), density_to_right, 1123 prev_region_density_to_right); 1124 } 1125 dense_prefix -= region_size; 1126 live_to_right = prev_region_live_to_right; 1127 space_to_right = prev_region_space_to_right; 1128 density_to_right = prev_region_density_to_right; 1129 } 1130 return dense_prefix; 1131 } 1132 1133 dense_prefix += region_size; 1134 ++cp; 1135 } 1136 1137 return dense_prefix; 1138 } 1139 1140 #ifndef PRODUCT 1141 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm, 1142 const SpaceId id, 1143 const bool maximum_compaction, 1144 HeapWord* const addr) 1145 { 1146 const size_t region_idx = summary_data().addr_to_region_idx(addr); 1147 RegionData* const cp = summary_data().region(region_idx); 1148 const MutableSpace* const space = _space_info[id].space(); 1149 HeapWord* const new_top = _space_info[id].new_top(); 1150 1151 const size_t space_live = pointer_delta(new_top, space->bottom()); 1152 const size_t dead_to_left = pointer_delta(addr, cp->destination()); 1153 const size_t space_cap = space->capacity_in_words(); 1154 const double dead_to_left_pct = double(dead_to_left) / space_cap; 1155 const size_t live_to_right = new_top - cp->destination(); 1156 const size_t dead_to_right = space->top() - addr - live_to_right; 1157 1158 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " " 1159 "spl=" SIZE_FORMAT " " 1160 "d2l=" SIZE_FORMAT " d2l%%=%6.4f " 1161 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT 1162 " ratio=%10.8f", 1163 algorithm, addr, region_idx, 1164 space_live, 1165 dead_to_left, dead_to_left_pct, 1166 dead_to_right, live_to_right, 1167 double(dead_to_right) / live_to_right); 1168 } 1169 #endif // #ifndef PRODUCT 1170 1171 // Return a fraction indicating how much of the generation can be treated as 1172 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution 1173 // based on the density of live objects in the generation to determine a limit, 1174 // which is then adjusted so the return value is min_percent when the density is 1175 // 1. 1176 // 1177 // The following table shows some return values for a different values of the 1178 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and 1179 // min_percent is 1. 1180 // 1181 // fraction allowed as dead wood 1182 // ----------------------------------------------------------------- 1183 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95 1184 // ------- ---------- ---------- ---------- ---------- ---------- ---------- 1185 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1186 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1187 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1188 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1189 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1190 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1191 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1192 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1193 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1194 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1195 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510 1196 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386 1197 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500 1198 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289 1199 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132 1200 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313 1201 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975 1202 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066 1203 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272 1204 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941 1205 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 1206 1207 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent) 1208 { 1209 assert(_dwl_initialized, "uninitialized"); 1210 1211 // The raw limit is the value of the normal distribution at x = density. 1212 const double raw_limit = normal_distribution(density); 1213 1214 // Adjust the raw limit so it becomes the minimum when the density is 1. 1215 // 1216 // First subtract the adjustment value (which is simply the precomputed value 1217 // normal_distribution(1.0)); this yields a value of 0 when the density is 1. 1218 // Then add the minimum value, so the minimum is returned when the density is 1219 // 1. Finally, prevent negative values, which occur when the mean is not 0.5. 1220 const double min = double(min_percent) / 100.0; 1221 const double limit = raw_limit - _dwl_adjustment + min; 1222 return MAX2(limit, 0.0); 1223 } 1224 1225 ParallelCompactData::RegionData* 1226 PSParallelCompact::first_dead_space_region(const RegionData* beg, 1227 const RegionData* end) 1228 { 1229 const size_t region_size = ParallelCompactData::RegionSize; 1230 ParallelCompactData& sd = summary_data(); 1231 size_t left = sd.region(beg); 1232 size_t right = end > beg ? sd.region(end) - 1 : left; 1233 1234 // Binary search. 1235 while (left < right) { 1236 // Equivalent to (left + right) / 2, but does not overflow. 1237 const size_t middle = left + (right - left) / 2; 1238 RegionData* const middle_ptr = sd.region(middle); 1239 HeapWord* const dest = middle_ptr->destination(); 1240 HeapWord* const addr = sd.region_to_addr(middle); 1241 assert(dest != NULL, "sanity"); 1242 assert(dest <= addr, "must move left"); 1243 1244 if (middle > left && dest < addr) { 1245 right = middle - 1; 1246 } else if (middle < right && middle_ptr->data_size() == region_size) { 1247 left = middle + 1; 1248 } else { 1249 return middle_ptr; 1250 } 1251 } 1252 return sd.region(left); 1253 } 1254 1255 ParallelCompactData::RegionData* 1256 PSParallelCompact::dead_wood_limit_region(const RegionData* beg, 1257 const RegionData* end, 1258 size_t dead_words) 1259 { 1260 ParallelCompactData& sd = summary_data(); 1261 size_t left = sd.region(beg); 1262 size_t right = end > beg ? sd.region(end) - 1 : left; 1263 1264 // Binary search. 1265 while (left < right) { 1266 // Equivalent to (left + right) / 2, but does not overflow. 1267 const size_t middle = left + (right - left) / 2; 1268 RegionData* const middle_ptr = sd.region(middle); 1269 HeapWord* const dest = middle_ptr->destination(); 1270 HeapWord* const addr = sd.region_to_addr(middle); 1271 assert(dest != NULL, "sanity"); 1272 assert(dest <= addr, "must move left"); 1273 1274 const size_t dead_to_left = pointer_delta(addr, dest); 1275 if (middle > left && dead_to_left > dead_words) { 1276 right = middle - 1; 1277 } else if (middle < right && dead_to_left < dead_words) { 1278 left = middle + 1; 1279 } else { 1280 return middle_ptr; 1281 } 1282 } 1283 return sd.region(left); 1284 } 1285 1286 // The result is valid during the summary phase, after the initial summarization 1287 // of each space into itself, and before final summarization. 1288 inline double 1289 PSParallelCompact::reclaimed_ratio(const RegionData* const cp, 1290 HeapWord* const bottom, 1291 HeapWord* const top, 1292 HeapWord* const new_top) 1293 { 1294 ParallelCompactData& sd = summary_data(); 1295 1296 assert(cp != NULL, "sanity"); 1297 assert(bottom != NULL, "sanity"); 1298 assert(top != NULL, "sanity"); 1299 assert(new_top != NULL, "sanity"); 1300 assert(top >= new_top, "summary data problem?"); 1301 assert(new_top > bottom, "space is empty; should not be here"); 1302 assert(new_top >= cp->destination(), "sanity"); 1303 assert(top >= sd.region_to_addr(cp), "sanity"); 1304 1305 HeapWord* const destination = cp->destination(); 1306 const size_t dense_prefix_live = pointer_delta(destination, bottom); 1307 const size_t compacted_region_live = pointer_delta(new_top, destination); 1308 const size_t compacted_region_used = pointer_delta(top, 1309 sd.region_to_addr(cp)); 1310 const size_t reclaimable = compacted_region_used - compacted_region_live; 1311 1312 const double divisor = dense_prefix_live + 1.25 * compacted_region_live; 1313 return double(reclaimable) / divisor; 1314 } 1315 1316 // Return the address of the end of the dense prefix, a.k.a. the start of the 1317 // compacted region. The address is always on a region boundary. 1318 // 1319 // Completely full regions at the left are skipped, since no compaction can 1320 // occur in those regions. Then the maximum amount of dead wood to allow is 1321 // computed, based on the density (amount live / capacity) of the generation; 1322 // the region with approximately that amount of dead space to the left is 1323 // identified as the limit region. Regions between the last completely full 1324 // region and the limit region are scanned and the one that has the best 1325 // (maximum) reclaimed_ratio() is selected. 1326 HeapWord* 1327 PSParallelCompact::compute_dense_prefix(const SpaceId id, 1328 bool maximum_compaction) 1329 { 1330 if (ParallelOldGCSplitALot) { 1331 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) { 1332 // The value was chosen to provoke splitting a young gen space; use it. 1333 return _space_info[id].dense_prefix(); 1334 } 1335 } 1336 1337 const size_t region_size = ParallelCompactData::RegionSize; 1338 const ParallelCompactData& sd = summary_data(); 1339 1340 const MutableSpace* const space = _space_info[id].space(); 1341 HeapWord* const top = space->top(); 1342 HeapWord* const top_aligned_up = sd.region_align_up(top); 1343 HeapWord* const new_top = _space_info[id].new_top(); 1344 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top); 1345 HeapWord* const bottom = space->bottom(); 1346 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom); 1347 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 1348 const RegionData* const new_top_cp = 1349 sd.addr_to_region_ptr(new_top_aligned_up); 1350 1351 // Skip full regions at the beginning of the space--they are necessarily part 1352 // of the dense prefix. 1353 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp); 1354 assert(full_cp->destination() == sd.region_to_addr(full_cp) || 1355 space->is_empty(), "no dead space allowed to the left"); 1356 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1, 1357 "region must have dead space"); 1358 1359 // The gc number is saved whenever a maximum compaction is done, and used to 1360 // determine when the maximum compaction interval has expired. This avoids 1361 // successive max compactions for different reasons. 1362 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity"); 1363 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num; 1364 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval || 1365 total_invocations() == HeapFirstMaximumCompactionCount; 1366 if (maximum_compaction || full_cp == top_cp || interval_ended) { 1367 _maximum_compaction_gc_num = total_invocations(); 1368 return sd.region_to_addr(full_cp); 1369 } 1370 1371 const size_t space_live = pointer_delta(new_top, bottom); 1372 const size_t space_used = space->used_in_words(); 1373 const size_t space_capacity = space->capacity_in_words(); 1374 1375 const double density = double(space_live) / double(space_capacity); 1376 const size_t min_percent_free = 1377 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio; 1378 const double limiter = dead_wood_limiter(density, min_percent_free); 1379 const size_t dead_wood_max = space_used - space_live; 1380 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter), 1381 dead_wood_max); 1382 1383 if (TraceParallelOldGCDensePrefix) { 1384 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " " 1385 "space_cap=" SIZE_FORMAT, 1386 space_live, space_used, 1387 space_capacity); 1388 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f " 1389 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT, 1390 density, min_percent_free, limiter, 1391 dead_wood_max, dead_wood_limit); 1392 } 1393 1394 // Locate the region with the desired amount of dead space to the left. 1395 const RegionData* const limit_cp = 1396 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit); 1397 1398 // Scan from the first region with dead space to the limit region and find the 1399 // one with the best (largest) reclaimed ratio. 1400 double best_ratio = 0.0; 1401 const RegionData* best_cp = full_cp; 1402 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) { 1403 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top); 1404 if (tmp_ratio > best_ratio) { 1405 best_cp = cp; 1406 best_ratio = tmp_ratio; 1407 } 1408 } 1409 1410 #if 0 1411 // Something to consider: if the region with the best ratio is 'close to' the 1412 // first region w/free space, choose the first region with free space 1413 // ("first-free"). The first-free region is usually near the start of the 1414 // heap, which means we are copying most of the heap already, so copy a bit 1415 // more to get complete compaction. 1416 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) { 1417 _maximum_compaction_gc_num = total_invocations(); 1418 best_cp = full_cp; 1419 } 1420 #endif // #if 0 1421 1422 return sd.region_to_addr(best_cp); 1423 } 1424 1425 #ifndef PRODUCT 1426 void 1427 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start, 1428 size_t words) 1429 { 1430 if (TraceParallelOldGCSummaryPhase) { 1431 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") " 1432 SIZE_FORMAT, start, start + words, words); 1433 } 1434 1435 ObjectStartArray* const start_array = _space_info[id].start_array(); 1436 CollectedHeap::fill_with_objects(start, words); 1437 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) { 1438 _mark_bitmap.mark_obj(p, words); 1439 _summary_data.add_obj(p, words); 1440 start_array->allocate_block(p); 1441 } 1442 } 1443 1444 void 1445 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start) 1446 { 1447 ParallelCompactData& sd = summary_data(); 1448 MutableSpace* space = _space_info[id].space(); 1449 1450 // Find the source and destination start addresses. 1451 HeapWord* const src_addr = sd.region_align_down(start); 1452 HeapWord* dst_addr; 1453 if (src_addr < start) { 1454 dst_addr = sd.addr_to_region_ptr(src_addr)->destination(); 1455 } else if (src_addr > space->bottom()) { 1456 // The start (the original top() value) is aligned to a region boundary so 1457 // the associated region does not have a destination. Compute the 1458 // destination from the previous region. 1459 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1; 1460 dst_addr = cp->destination() + cp->data_size(); 1461 } else { 1462 // Filling the entire space. 1463 dst_addr = space->bottom(); 1464 } 1465 assert(dst_addr != NULL, "sanity"); 1466 1467 // Update the summary data. 1468 bool result = _summary_data.summarize(_space_info[id].split_info(), 1469 src_addr, space->top(), NULL, 1470 dst_addr, space->end(), 1471 _space_info[id].new_top_addr()); 1472 assert(result, "should not fail: bad filler object size"); 1473 } 1474 1475 void 1476 PSParallelCompact::provoke_split_fill_survivor(SpaceId id) 1477 { 1478 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) { 1479 return; 1480 } 1481 1482 MutableSpace* const space = _space_info[id].space(); 1483 if (space->is_empty()) { 1484 HeapWord* b = space->bottom(); 1485 HeapWord* t = b + space->capacity_in_words() / 2; 1486 space->set_top(t); 1487 if (ZapUnusedHeapArea) { 1488 space->set_top_for_allocations(); 1489 } 1490 1491 size_t min_size = CollectedHeap::min_fill_size(); 1492 size_t obj_len = min_size; 1493 while (b + obj_len <= t) { 1494 CollectedHeap::fill_with_object(b, obj_len); 1495 mark_bitmap()->mark_obj(b, obj_len); 1496 summary_data().add_obj(b, obj_len); 1497 b += obj_len; 1498 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ... 1499 } 1500 if (b < t) { 1501 // The loop didn't completely fill to t (top); adjust top downward. 1502 space->set_top(b); 1503 if (ZapUnusedHeapArea) { 1504 space->set_top_for_allocations(); 1505 } 1506 } 1507 1508 HeapWord** nta = _space_info[id].new_top_addr(); 1509 bool result = summary_data().summarize(_space_info[id].split_info(), 1510 space->bottom(), space->top(), NULL, 1511 space->bottom(), space->end(), nta); 1512 assert(result, "space must fit into itself"); 1513 } 1514 } 1515 1516 void 1517 PSParallelCompact::provoke_split(bool & max_compaction) 1518 { 1519 if (total_invocations() % ParallelOldGCSplitInterval != 0) { 1520 return; 1521 } 1522 1523 const size_t region_size = ParallelCompactData::RegionSize; 1524 ParallelCompactData& sd = summary_data(); 1525 1526 MutableSpace* const eden_space = _space_info[eden_space_id].space(); 1527 MutableSpace* const from_space = _space_info[from_space_id].space(); 1528 const size_t eden_live = pointer_delta(eden_space->top(), 1529 _space_info[eden_space_id].new_top()); 1530 const size_t from_live = pointer_delta(from_space->top(), 1531 _space_info[from_space_id].new_top()); 1532 1533 const size_t min_fill_size = CollectedHeap::min_fill_size(); 1534 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top()); 1535 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0; 1536 const size_t from_free = pointer_delta(from_space->end(), from_space->top()); 1537 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0; 1538 1539 // Choose the space to split; need at least 2 regions live (or fillable). 1540 SpaceId id; 1541 MutableSpace* space; 1542 size_t live_words; 1543 size_t fill_words; 1544 if (eden_live + eden_fillable >= region_size * 2) { 1545 id = eden_space_id; 1546 space = eden_space; 1547 live_words = eden_live; 1548 fill_words = eden_fillable; 1549 } else if (from_live + from_fillable >= region_size * 2) { 1550 id = from_space_id; 1551 space = from_space; 1552 live_words = from_live; 1553 fill_words = from_fillable; 1554 } else { 1555 return; // Give up. 1556 } 1557 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity"); 1558 1559 if (live_words < region_size * 2) { 1560 // Fill from top() to end() w/live objects of mixed sizes. 1561 HeapWord* const fill_start = space->top(); 1562 live_words += fill_words; 1563 1564 space->set_top(fill_start + fill_words); 1565 if (ZapUnusedHeapArea) { 1566 space->set_top_for_allocations(); 1567 } 1568 1569 HeapWord* cur_addr = fill_start; 1570 while (fill_words > 0) { 1571 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size; 1572 size_t cur_size = MIN2(align_object_size_(r), fill_words); 1573 if (fill_words - cur_size < min_fill_size) { 1574 cur_size = fill_words; // Avoid leaving a fragment too small to fill. 1575 } 1576 1577 CollectedHeap::fill_with_object(cur_addr, cur_size); 1578 mark_bitmap()->mark_obj(cur_addr, cur_size); 1579 sd.add_obj(cur_addr, cur_size); 1580 1581 cur_addr += cur_size; 1582 fill_words -= cur_size; 1583 } 1584 1585 summarize_new_objects(id, fill_start); 1586 } 1587 1588 max_compaction = false; 1589 1590 // Manipulate the old gen so that it has room for about half of the live data 1591 // in the target young gen space (live_words / 2). 1592 id = old_space_id; 1593 space = _space_info[id].space(); 1594 const size_t free_at_end = space->free_in_words(); 1595 const size_t free_target = align_object_size(live_words / 2); 1596 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top()); 1597 1598 if (free_at_end >= free_target + min_fill_size) { 1599 // Fill space above top() and set the dense prefix so everything survives. 1600 HeapWord* const fill_start = space->top(); 1601 const size_t fill_size = free_at_end - free_target; 1602 space->set_top(space->top() + fill_size); 1603 if (ZapUnusedHeapArea) { 1604 space->set_top_for_allocations(); 1605 } 1606 fill_with_live_objects(id, fill_start, fill_size); 1607 summarize_new_objects(id, fill_start); 1608 _space_info[id].set_dense_prefix(sd.region_align_down(space->top())); 1609 } else if (dead + free_at_end > free_target) { 1610 // Find a dense prefix that makes the right amount of space available. 1611 HeapWord* cur = sd.region_align_down(space->top()); 1612 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1613 size_t dead_to_right = pointer_delta(space->end(), cur_destination); 1614 while (dead_to_right < free_target) { 1615 cur -= region_size; 1616 cur_destination = sd.addr_to_region_ptr(cur)->destination(); 1617 dead_to_right = pointer_delta(space->end(), cur_destination); 1618 } 1619 _space_info[id].set_dense_prefix(cur); 1620 } 1621 } 1622 #endif // #ifndef PRODUCT 1623 1624 void PSParallelCompact::summarize_spaces_quick() 1625 { 1626 for (unsigned int i = 0; i < last_space_id; ++i) { 1627 const MutableSpace* space = _space_info[i].space(); 1628 HeapWord** nta = _space_info[i].new_top_addr(); 1629 bool result = _summary_data.summarize(_space_info[i].split_info(), 1630 space->bottom(), space->top(), NULL, 1631 space->bottom(), space->end(), nta); 1632 assert(result, "space must fit into itself"); 1633 _space_info[i].set_dense_prefix(space->bottom()); 1634 } 1635 1636 #ifndef PRODUCT 1637 if (ParallelOldGCSplitALot) { 1638 provoke_split_fill_survivor(to_space_id); 1639 } 1640 #endif // #ifndef PRODUCT 1641 } 1642 1643 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) 1644 { 1645 HeapWord* const dense_prefix_end = dense_prefix(id); 1646 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end); 1647 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end); 1648 if (dead_space_crosses_boundary(region, dense_prefix_bit)) { 1649 // Only enough dead space is filled so that any remaining dead space to the 1650 // left is larger than the minimum filler object. (The remainder is filled 1651 // during the copy/update phase.) 1652 // 1653 // The size of the dead space to the right of the boundary is not a 1654 // concern, since compaction will be able to use whatever space is 1655 // available. 1656 // 1657 // Here '||' is the boundary, 'x' represents a don't care bit and a box 1658 // surrounds the space to be filled with an object. 1659 // 1660 // In the 32-bit VM, each bit represents two 32-bit words: 1661 // +---+ 1662 // a) beg_bits: ... x x x | 0 | || 0 x x ... 1663 // end_bits: ... x x x | 0 | || 0 x x ... 1664 // +---+ 1665 // 1666 // In the 64-bit VM, each bit represents one 64-bit word: 1667 // +------------+ 1668 // b) beg_bits: ... x x x | 0 || 0 | x x ... 1669 // end_bits: ... x x 1 | 0 || 0 | x x ... 1670 // +------------+ 1671 // +-------+ 1672 // c) beg_bits: ... x x | 0 0 | || 0 x x ... 1673 // end_bits: ... x 1 | 0 0 | || 0 x x ... 1674 // +-------+ 1675 // +-----------+ 1676 // d) beg_bits: ... x | 0 0 0 | || 0 x x ... 1677 // end_bits: ... 1 | 0 0 0 | || 0 x x ... 1678 // +-----------+ 1679 // +-------+ 1680 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ... 1681 // end_bits: ... 0 0 | 0 0 | || 0 x x ... 1682 // +-------+ 1683 1684 // Initially assume case a, c or e will apply. 1685 size_t obj_len = CollectedHeap::min_fill_size(); 1686 HeapWord* obj_beg = dense_prefix_end - obj_len; 1687 1688 #ifdef _LP64 1689 if (MinObjAlignment > 1) { // object alignment > heap word size 1690 // Cases a, c or e. 1691 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) { 1692 // Case b above. 1693 obj_beg = dense_prefix_end - 1; 1694 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) && 1695 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) { 1696 // Case d above. 1697 obj_beg = dense_prefix_end - 3; 1698 obj_len = 3; 1699 } 1700 #endif // #ifdef _LP64 1701 1702 CollectedHeap::fill_with_object(obj_beg, obj_len); 1703 _mark_bitmap.mark_obj(obj_beg, obj_len); 1704 _summary_data.add_obj(obj_beg, obj_len); 1705 assert(start_array(id) != NULL, "sanity"); 1706 start_array(id)->allocate_block(obj_beg); 1707 } 1708 } 1709 1710 void 1711 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr) 1712 { 1713 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr); 1714 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr); 1715 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up); 1716 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) { 1717 cur->set_source_region(0); 1718 } 1719 } 1720 1721 void 1722 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction) 1723 { 1724 assert(id < last_space_id, "id out of range"); 1725 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() || 1726 ParallelOldGCSplitALot && id == old_space_id, 1727 "should have been reset in summarize_spaces_quick()"); 1728 1729 const MutableSpace* space = _space_info[id].space(); 1730 if (_space_info[id].new_top() != space->bottom()) { 1731 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction); 1732 _space_info[id].set_dense_prefix(dense_prefix_end); 1733 1734 #ifndef PRODUCT 1735 if (TraceParallelOldGCDensePrefix) { 1736 print_dense_prefix_stats("ratio", id, maximum_compaction, 1737 dense_prefix_end); 1738 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction); 1739 print_dense_prefix_stats("density", id, maximum_compaction, addr); 1740 } 1741 #endif // #ifndef PRODUCT 1742 1743 // Recompute the summary data, taking into account the dense prefix. If 1744 // every last byte will be reclaimed, then the existing summary data which 1745 // compacts everything can be left in place. 1746 if (!maximum_compaction && dense_prefix_end != space->bottom()) { 1747 // If dead space crosses the dense prefix boundary, it is (at least 1748 // partially) filled with a dummy object, marked live and added to the 1749 // summary data. This simplifies the copy/update phase and must be done 1750 // before the final locations of objects are determined, to prevent 1751 // leaving a fragment of dead space that is too small to fill. 1752 fill_dense_prefix_end(id); 1753 1754 // Compute the destination of each Region, and thus each object. 1755 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end); 1756 _summary_data.summarize(_space_info[id].split_info(), 1757 dense_prefix_end, space->top(), NULL, 1758 dense_prefix_end, space->end(), 1759 _space_info[id].new_top_addr()); 1760 } 1761 } 1762 1763 if (TraceParallelOldGCSummaryPhase) { 1764 const size_t region_size = ParallelCompactData::RegionSize; 1765 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix(); 1766 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end); 1767 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom()); 1768 HeapWord* const new_top = _space_info[id].new_top(); 1769 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top); 1770 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end); 1771 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " " 1772 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " " 1773 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT, 1774 id, space->capacity_in_words(), dense_prefix_end, 1775 dp_region, dp_words / region_size, 1776 cr_words / region_size, new_top); 1777 } 1778 } 1779 1780 #ifndef PRODUCT 1781 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id, 1782 HeapWord* dst_beg, HeapWord* dst_end, 1783 SpaceId src_space_id, 1784 HeapWord* src_beg, HeapWord* src_end) 1785 { 1786 if (TraceParallelOldGCSummaryPhase) { 1787 tty->print_cr("summarizing %d [%s] into %d [%s]: " 1788 "src=" PTR_FORMAT "-" PTR_FORMAT " " 1789 SIZE_FORMAT "-" SIZE_FORMAT " " 1790 "dst=" PTR_FORMAT "-" PTR_FORMAT " " 1791 SIZE_FORMAT "-" SIZE_FORMAT, 1792 src_space_id, space_names[src_space_id], 1793 dst_space_id, space_names[dst_space_id], 1794 src_beg, src_end, 1795 _summary_data.addr_to_region_idx(src_beg), 1796 _summary_data.addr_to_region_idx(src_end), 1797 dst_beg, dst_end, 1798 _summary_data.addr_to_region_idx(dst_beg), 1799 _summary_data.addr_to_region_idx(dst_end)); 1800 } 1801 } 1802 #endif // #ifndef PRODUCT 1803 1804 void PSParallelCompact::summary_phase(ParCompactionManager* cm, 1805 bool maximum_compaction) 1806 { 1807 EventMark m("2 summarize"); 1808 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty); 1809 // trace("2"); 1810 1811 #ifdef ASSERT 1812 if (TraceParallelOldGCMarkingPhase) { 1813 tty->print_cr("add_obj_count=" SIZE_FORMAT " " 1814 "add_obj_bytes=" SIZE_FORMAT, 1815 add_obj_count, add_obj_size * HeapWordSize); 1816 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " " 1817 "mark_bitmap_bytes=" SIZE_FORMAT, 1818 mark_bitmap_count, mark_bitmap_size * HeapWordSize); 1819 } 1820 #endif // #ifdef ASSERT 1821 1822 // Quick summarization of each space into itself, to see how much is live. 1823 summarize_spaces_quick(); 1824 1825 if (TraceParallelOldGCSummaryPhase) { 1826 tty->print_cr("summary_phase: after summarizing each space to self"); 1827 Universe::print(); 1828 NOT_PRODUCT(print_region_ranges()); 1829 if (Verbose) { 1830 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info)); 1831 } 1832 } 1833 1834 // The amount of live data that will end up in old space (assuming it fits). 1835 size_t old_space_total_live = 0; 1836 assert(perm_space_id < old_space_id, "should not count perm data here"); 1837 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 1838 old_space_total_live += pointer_delta(_space_info[id].new_top(), 1839 _space_info[id].space()->bottom()); 1840 } 1841 1842 MutableSpace* const old_space = _space_info[old_space_id].space(); 1843 const size_t old_capacity = old_space->capacity_in_words(); 1844 if (old_space_total_live > old_capacity) { 1845 // XXX - should also try to expand 1846 maximum_compaction = true; 1847 } 1848 #ifndef PRODUCT 1849 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) { 1850 provoke_split(maximum_compaction); 1851 } 1852 #endif // #ifndef PRODUCT 1853 1854 // Permanent and Old generations. 1855 summarize_space(perm_space_id, maximum_compaction); 1856 summarize_space(old_space_id, maximum_compaction); 1857 1858 // Summarize the remaining spaces in the young gen. The initial target space 1859 // is the old gen. If a space does not fit entirely into the target, then the 1860 // remainder is compacted into the space itself and that space becomes the new 1861 // target. 1862 SpaceId dst_space_id = old_space_id; 1863 HeapWord* dst_space_end = old_space->end(); 1864 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr(); 1865 for (unsigned int id = eden_space_id; id < last_space_id; ++id) { 1866 const MutableSpace* space = _space_info[id].space(); 1867 const size_t live = pointer_delta(_space_info[id].new_top(), 1868 space->bottom()); 1869 const size_t available = pointer_delta(dst_space_end, *new_top_addr); 1870 1871 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end, 1872 SpaceId(id), space->bottom(), space->top());) 1873 if (live > 0 && live <= available) { 1874 // All the live data will fit. 1875 bool done = _summary_data.summarize(_space_info[id].split_info(), 1876 space->bottom(), space->top(), 1877 NULL, 1878 *new_top_addr, dst_space_end, 1879 new_top_addr); 1880 assert(done, "space must fit into old gen"); 1881 1882 // Reset the new_top value for the space. 1883 _space_info[id].set_new_top(space->bottom()); 1884 } else if (live > 0) { 1885 // Attempt to fit part of the source space into the target space. 1886 HeapWord* next_src_addr = NULL; 1887 bool done = _summary_data.summarize(_space_info[id].split_info(), 1888 space->bottom(), space->top(), 1889 &next_src_addr, 1890 *new_top_addr, dst_space_end, 1891 new_top_addr); 1892 assert(!done, "space should not fit into old gen"); 1893 assert(next_src_addr != NULL, "sanity"); 1894 1895 // The source space becomes the new target, so the remainder is compacted 1896 // within the space itself. 1897 dst_space_id = SpaceId(id); 1898 dst_space_end = space->end(); 1899 new_top_addr = _space_info[id].new_top_addr(); 1900 NOT_PRODUCT(summary_phase_msg(dst_space_id, 1901 space->bottom(), dst_space_end, 1902 SpaceId(id), next_src_addr, space->top());) 1903 done = _summary_data.summarize(_space_info[id].split_info(), 1904 next_src_addr, space->top(), 1905 NULL, 1906 space->bottom(), dst_space_end, 1907 new_top_addr); 1908 assert(done, "space must fit when compacted into itself"); 1909 assert(*new_top_addr <= space->top(), "usage should not grow"); 1910 } 1911 } 1912 1913 if (TraceParallelOldGCSummaryPhase) { 1914 tty->print_cr("summary_phase: after final summarization"); 1915 Universe::print(); 1916 NOT_PRODUCT(print_region_ranges()); 1917 if (Verbose) { 1918 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info)); 1919 } 1920 } 1921 } 1922 1923 // This method should contain all heap-specific policy for invoking a full 1924 // collection. invoke_no_policy() will only attempt to compact the heap; it 1925 // will do nothing further. If we need to bail out for policy reasons, scavenge 1926 // before full gc, or any other specialized behavior, it needs to be added here. 1927 // 1928 // Note that this method should only be called from the vm_thread while at a 1929 // safepoint. 1930 // 1931 // Note that the all_soft_refs_clear flag in the collector policy 1932 // may be true because this method can be called without intervening 1933 // activity. For example when the heap space is tight and full measure 1934 // are being taken to free space. 1935 void PSParallelCompact::invoke(bool maximum_heap_compaction) { 1936 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1937 assert(Thread::current() == (Thread*)VMThread::vm_thread(), 1938 "should be in vm thread"); 1939 1940 ParallelScavengeHeap* heap = gc_heap(); 1941 GCCause::Cause gc_cause = heap->gc_cause(); 1942 assert(!heap->is_gc_active(), "not reentrant"); 1943 1944 PSAdaptiveSizePolicy* policy = heap->size_policy(); 1945 IsGCActiveMark mark; 1946 1947 if (ScavengeBeforeFullGC) { 1948 PSScavenge::invoke_no_policy(); 1949 } 1950 1951 const bool clear_all_soft_refs = 1952 heap->collector_policy()->should_clear_all_soft_refs(); 1953 1954 PSParallelCompact::invoke_no_policy(clear_all_soft_refs || 1955 maximum_heap_compaction); 1956 } 1957 1958 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) { 1959 size_t addr_region_index = addr_to_region_idx(addr); 1960 return region_index == addr_region_index; 1961 } 1962 1963 // This method contains no policy. You should probably 1964 // be calling invoke() instead. 1965 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) { 1966 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint"); 1967 assert(ref_processor() != NULL, "Sanity"); 1968 1969 if (GC_locker::check_active_before_gc()) { 1970 return; 1971 } 1972 1973 TimeStamp marking_start; 1974 TimeStamp compaction_start; 1975 TimeStamp collection_exit; 1976 1977 ParallelScavengeHeap* heap = gc_heap(); 1978 GCCause::Cause gc_cause = heap->gc_cause(); 1979 PSYoungGen* young_gen = heap->young_gen(); 1980 PSOldGen* old_gen = heap->old_gen(); 1981 PSPermGen* perm_gen = heap->perm_gen(); 1982 PSAdaptiveSizePolicy* size_policy = heap->size_policy(); 1983 1984 // The scope of casr should end after code that can change 1985 // CollectorPolicy::_should_clear_all_soft_refs. 1986 ClearedAllSoftRefs casr(maximum_heap_compaction, 1987 heap->collector_policy()); 1988 1989 if (ZapUnusedHeapArea) { 1990 // Save information needed to minimize mangling 1991 heap->record_gen_tops_before_GC(); 1992 } 1993 1994 heap->pre_full_gc_dump(); 1995 1996 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes; 1997 1998 // Make sure data structures are sane, make the heap parsable, and do other 1999 // miscellaneous bookkeeping. 2000 PreGCValues pre_gc_values; 2001 pre_compact(&pre_gc_values); 2002 2003 // Get the compaction manager reserved for the VM thread. 2004 ParCompactionManager* const vmthread_cm = 2005 ParCompactionManager::manager_array(gc_task_manager()->workers()); 2006 2007 // Place after pre_compact() where the number of invocations is incremented. 2008 AdaptiveSizePolicyOutput(size_policy, heap->total_collections()); 2009 2010 { 2011 ResourceMark rm; 2012 HandleMark hm; 2013 2014 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc; 2015 2016 // This is useful for debugging but don't change the output the 2017 // the customer sees. 2018 const char* gc_cause_str = "Full GC"; 2019 if (is_system_gc && PrintGCDetails) { 2020 gc_cause_str = "Full GC (System)"; 2021 } 2022 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 2023 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 2024 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty); 2025 TraceCollectorStats tcs(counters()); 2026 TraceMemoryManagerStats tms(true /* Full GC */); 2027 2028 if (TraceGen1Time) accumulated_time()->start(); 2029 2030 // Let the size policy know we're starting 2031 size_policy->major_collection_begin(); 2032 2033 // When collecting the permanent generation methodOops may be moving, 2034 // so we either have to flush all bcp data or convert it into bci. 2035 CodeCache::gc_prologue(); 2036 Threads::gc_prologue(); 2037 2038 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2039 COMPILER2_PRESENT(DerivedPointerTable::clear()); 2040 2041 ref_processor()->enable_discovery(); 2042 ref_processor()->setup_policy(maximum_heap_compaction); 2043 2044 bool marked_for_unloading = false; 2045 2046 marking_start.update(); 2047 marking_phase(vmthread_cm, maximum_heap_compaction); 2048 2049 #ifndef PRODUCT 2050 if (TraceParallelOldGCMarkingPhase) { 2051 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d " 2052 "cas_by_another %d", 2053 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(), 2054 mark_bitmap()->cas_by_another()); 2055 } 2056 #endif // #ifndef PRODUCT 2057 2058 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc; 2059 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc); 2060 2061 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity")); 2062 COMPILER2_PRESENT(DerivedPointerTable::set_active(false)); 2063 2064 // adjust_roots() updates Universe::_intArrayKlassObj which is 2065 // needed by the compaction for filling holes in the dense prefix. 2066 adjust_roots(); 2067 2068 compaction_start.update(); 2069 // Does the perm gen always have to be done serially because 2070 // klasses are used in the update of an object? 2071 compact_perm(vmthread_cm); 2072 2073 if (UseParallelOldGCCompacting) { 2074 compact(); 2075 } else { 2076 compact_serial(vmthread_cm); 2077 } 2078 2079 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be 2080 // done before resizing. 2081 post_compact(); 2082 2083 // Let the size policy know we're done 2084 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause); 2085 2086 if (UseAdaptiveSizePolicy) { 2087 if (PrintAdaptiveSizePolicy) { 2088 gclog_or_tty->print("AdaptiveSizeStart: "); 2089 gclog_or_tty->stamp(); 2090 gclog_or_tty->print_cr(" collection: %d ", 2091 heap->total_collections()); 2092 if (Verbose) { 2093 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d" 2094 " perm_gen_capacity: %d ", 2095 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(), 2096 perm_gen->capacity_in_bytes()); 2097 } 2098 } 2099 2100 // Don't check if the size_policy is ready here. Let 2101 // the size_policy check that internally. 2102 if (UseAdaptiveGenerationSizePolicyAtMajorCollection && 2103 ((gc_cause != GCCause::_java_lang_system_gc) || 2104 UseAdaptiveSizePolicyWithSystemGC)) { 2105 // Calculate optimal free space amounts 2106 assert(young_gen->max_size() > 2107 young_gen->from_space()->capacity_in_bytes() + 2108 young_gen->to_space()->capacity_in_bytes(), 2109 "Sizes of space in young gen are out-of-bounds"); 2110 size_t max_eden_size = young_gen->max_size() - 2111 young_gen->from_space()->capacity_in_bytes() - 2112 young_gen->to_space()->capacity_in_bytes(); 2113 size_policy->compute_generation_free_space( 2114 young_gen->used_in_bytes(), 2115 young_gen->eden_space()->used_in_bytes(), 2116 old_gen->used_in_bytes(), 2117 perm_gen->used_in_bytes(), 2118 young_gen->eden_space()->capacity_in_bytes(), 2119 old_gen->max_gen_size(), 2120 max_eden_size, 2121 true /* full gc*/, 2122 gc_cause, 2123 heap->collector_policy()); 2124 2125 heap->resize_old_gen( 2126 size_policy->calculated_old_free_size_in_bytes()); 2127 2128 // Don't resize the young generation at an major collection. A 2129 // desired young generation size may have been calculated but 2130 // resizing the young generation complicates the code because the 2131 // resizing of the old generation may have moved the boundary 2132 // between the young generation and the old generation. Let the 2133 // young generation resizing happen at the minor collections. 2134 } 2135 if (PrintAdaptiveSizePolicy) { 2136 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ", 2137 heap->total_collections()); 2138 } 2139 } 2140 2141 if (UsePerfData) { 2142 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters(); 2143 counters->update_counters(); 2144 counters->update_old_capacity(old_gen->capacity_in_bytes()); 2145 counters->update_young_capacity(young_gen->capacity_in_bytes()); 2146 } 2147 2148 heap->resize_all_tlabs(); 2149 2150 // We collected the perm gen, so we'll resize it here. 2151 perm_gen->compute_new_size(pre_gc_values.perm_gen_used()); 2152 2153 if (TraceGen1Time) accumulated_time()->stop(); 2154 2155 if (PrintGC) { 2156 if (PrintGCDetails) { 2157 // No GC timestamp here. This is after GC so it would be confusing. 2158 young_gen->print_used_change(pre_gc_values.young_gen_used()); 2159 old_gen->print_used_change(pre_gc_values.old_gen_used()); 2160 heap->print_heap_change(pre_gc_values.heap_used()); 2161 // Print perm gen last (print_heap_change() excludes the perm gen). 2162 perm_gen->print_used_change(pre_gc_values.perm_gen_used()); 2163 } else { 2164 heap->print_heap_change(pre_gc_values.heap_used()); 2165 } 2166 } 2167 2168 // Track memory usage and detect low memory 2169 MemoryService::track_memory_usage(); 2170 heap->update_counters(); 2171 } 2172 2173 #ifdef ASSERT 2174 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) { 2175 ParCompactionManager* const cm = 2176 ParCompactionManager::manager_array(int(i)); 2177 assert(cm->marking_stack()->is_empty(), "should be empty"); 2178 assert(cm->region_stack()->is_empty(), "should be empty"); 2179 assert(cm->revisit_klass_stack()->is_empty(), "should be empty"); 2180 } 2181 #endif // ASSERT 2182 2183 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) { 2184 HandleMark hm; // Discard invalid handles created during verification 2185 gclog_or_tty->print(" VerifyAfterGC:"); 2186 Universe::verify(false); 2187 } 2188 2189 // Re-verify object start arrays 2190 if (VerifyObjectStartArray && 2191 VerifyAfterGC) { 2192 old_gen->verify_object_start_array(); 2193 perm_gen->verify_object_start_array(); 2194 } 2195 2196 if (ZapUnusedHeapArea) { 2197 old_gen->object_space()->check_mangled_unused_area_complete(); 2198 perm_gen->object_space()->check_mangled_unused_area_complete(); 2199 } 2200 2201 NOT_PRODUCT(ref_processor()->verify_no_references_recorded()); 2202 2203 collection_exit.update(); 2204 2205 if (PrintHeapAtGC) { 2206 Universe::print_heap_after_gc(); 2207 } 2208 if (PrintGCTaskTimeStamps) { 2209 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " " 2210 INT64_FORMAT, 2211 marking_start.ticks(), compaction_start.ticks(), 2212 collection_exit.ticks()); 2213 gc_task_manager()->print_task_time_stamps(); 2214 } 2215 2216 heap->post_full_gc_dump(); 2217 2218 #ifdef TRACESPINNING 2219 ParallelTaskTerminator::print_termination_counts(); 2220 #endif 2221 } 2222 2223 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy, 2224 PSYoungGen* young_gen, 2225 PSOldGen* old_gen) { 2226 MutableSpace* const eden_space = young_gen->eden_space(); 2227 assert(!eden_space->is_empty(), "eden must be non-empty"); 2228 assert(young_gen->virtual_space()->alignment() == 2229 old_gen->virtual_space()->alignment(), "alignments do not match"); 2230 2231 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) { 2232 return false; 2233 } 2234 2235 // Both generations must be completely committed. 2236 if (young_gen->virtual_space()->uncommitted_size() != 0) { 2237 return false; 2238 } 2239 if (old_gen->virtual_space()->uncommitted_size() != 0) { 2240 return false; 2241 } 2242 2243 // Figure out how much to take from eden. Include the average amount promoted 2244 // in the total; otherwise the next young gen GC will simply bail out to a 2245 // full GC. 2246 const size_t alignment = old_gen->virtual_space()->alignment(); 2247 const size_t eden_used = eden_space->used_in_bytes(); 2248 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average(); 2249 const size_t absorb_size = align_size_up(eden_used + promoted, alignment); 2250 const size_t eden_capacity = eden_space->capacity_in_bytes(); 2251 2252 if (absorb_size >= eden_capacity) { 2253 return false; // Must leave some space in eden. 2254 } 2255 2256 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size; 2257 if (new_young_size < young_gen->min_gen_size()) { 2258 return false; // Respect young gen minimum size. 2259 } 2260 2261 if (TraceAdaptiveGCBoundary && Verbose) { 2262 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: " 2263 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K " 2264 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K " 2265 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ", 2266 absorb_size / K, 2267 eden_capacity / K, (eden_capacity - absorb_size) / K, 2268 young_gen->from_space()->used_in_bytes() / K, 2269 young_gen->to_space()->used_in_bytes() / K, 2270 young_gen->capacity_in_bytes() / K, new_young_size / K); 2271 } 2272 2273 // Fill the unused part of the old gen. 2274 MutableSpace* const old_space = old_gen->object_space(); 2275 HeapWord* const unused_start = old_space->top(); 2276 size_t const unused_words = pointer_delta(old_space->end(), unused_start); 2277 2278 if (unused_words > 0) { 2279 if (unused_words < CollectedHeap::min_fill_size()) { 2280 return false; // If the old gen cannot be filled, must give up. 2281 } 2282 CollectedHeap::fill_with_objects(unused_start, unused_words); 2283 } 2284 2285 // Take the live data from eden and set both top and end in the old gen to 2286 // eden top. (Need to set end because reset_after_change() mangles the region 2287 // from end to virtual_space->high() in debug builds). 2288 HeapWord* const new_top = eden_space->top(); 2289 old_gen->virtual_space()->expand_into(young_gen->virtual_space(), 2290 absorb_size); 2291 young_gen->reset_after_change(); 2292 old_space->set_top(new_top); 2293 old_space->set_end(new_top); 2294 old_gen->reset_after_change(); 2295 2296 // Update the object start array for the filler object and the data from eden. 2297 ObjectStartArray* const start_array = old_gen->start_array(); 2298 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) { 2299 start_array->allocate_block(p); 2300 } 2301 2302 // Could update the promoted average here, but it is not typically updated at 2303 // full GCs and the value to use is unclear. Something like 2304 // 2305 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc. 2306 2307 size_policy->set_bytes_absorbed_from_eden(absorb_size); 2308 return true; 2309 } 2310 2311 GCTaskManager* const PSParallelCompact::gc_task_manager() { 2312 assert(ParallelScavengeHeap::gc_task_manager() != NULL, 2313 "shouldn't return NULL"); 2314 return ParallelScavengeHeap::gc_task_manager(); 2315 } 2316 2317 void PSParallelCompact::marking_phase(ParCompactionManager* cm, 2318 bool maximum_heap_compaction) { 2319 // Recursively traverse all live objects and mark them 2320 EventMark m("1 mark object"); 2321 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty); 2322 2323 ParallelScavengeHeap* heap = gc_heap(); 2324 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2325 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2326 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2327 2328 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm); 2329 PSParallelCompact::FollowStackClosure follow_stack_closure(cm); 2330 2331 { 2332 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty); 2333 ParallelScavengeHeap::ParStrongRootsScope psrs; 2334 2335 GCTaskQueue* q = GCTaskQueue::create(); 2336 2337 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe)); 2338 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles)); 2339 // We scan the thread roots in parallel 2340 Threads::create_thread_roots_marking_tasks(q); 2341 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer)); 2342 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler)); 2343 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management)); 2344 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary)); 2345 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti)); 2346 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols)); 2347 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache)); 2348 2349 if (parallel_gc_threads > 1) { 2350 for (uint j = 0; j < parallel_gc_threads; j++) { 2351 q->enqueue(new StealMarkingTask(&terminator)); 2352 } 2353 } 2354 2355 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2356 q->enqueue(fin); 2357 2358 gc_task_manager()->add_list(q); 2359 2360 fin->wait_for(); 2361 2362 // We have to release the barrier tasks! 2363 WaitForBarrierGCTask::destroy(fin); 2364 } 2365 2366 // Process reference objects found during marking 2367 { 2368 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty); 2369 if (ref_processor()->processing_is_mt()) { 2370 RefProcTaskExecutor task_executor; 2371 ref_processor()->process_discovered_references( 2372 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, 2373 &task_executor); 2374 } else { 2375 ref_processor()->process_discovered_references( 2376 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL); 2377 } 2378 } 2379 2380 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty); 2381 // Follow system dictionary roots and unload classes. 2382 bool purged_class = SystemDictionary::do_unloading(is_alive_closure()); 2383 2384 // Follow code cache roots. 2385 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure, 2386 purged_class); 2387 cm->follow_marking_stacks(); // Flush marking stack. 2388 2389 // Update subklass/sibling/implementor links of live klasses 2390 // revisit_klass_stack is used in follow_weak_klass_links(). 2391 follow_weak_klass_links(); 2392 2393 // Revisit memoized MDO's and clear any unmarked weak refs 2394 follow_mdo_weak_refs(); 2395 2396 // Visit symbol and interned string tables and delete unmarked oops 2397 SymbolTable::unlink(is_alive_closure()); 2398 StringTable::unlink(is_alive_closure()); 2399 2400 assert(cm->marking_stacks_empty(), "marking stacks should be empty"); 2401 } 2402 2403 // This should be moved to the shared markSweep code! 2404 class PSAlwaysTrueClosure: public BoolObjectClosure { 2405 public: 2406 void do_object(oop p) { ShouldNotReachHere(); } 2407 bool do_object_b(oop p) { return true; } 2408 }; 2409 static PSAlwaysTrueClosure always_true; 2410 2411 void PSParallelCompact::adjust_roots() { 2412 // Adjust the pointers to reflect the new locations 2413 EventMark m("3 adjust roots"); 2414 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty); 2415 2416 // General strong roots. 2417 Universe::oops_do(adjust_root_pointer_closure()); 2418 ReferenceProcessor::oops_do(adjust_root_pointer_closure()); 2419 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles 2420 Threads::oops_do(adjust_root_pointer_closure(), NULL); 2421 ObjectSynchronizer::oops_do(adjust_root_pointer_closure()); 2422 FlatProfiler::oops_do(adjust_root_pointer_closure()); 2423 Management::oops_do(adjust_root_pointer_closure()); 2424 JvmtiExport::oops_do(adjust_root_pointer_closure()); 2425 // SO_AllClasses 2426 SystemDictionary::oops_do(adjust_root_pointer_closure()); 2427 vmSymbols::oops_do(adjust_root_pointer_closure()); 2428 2429 // Now adjust pointers in remaining weak roots. (All of which should 2430 // have been cleared if they pointed to non-surviving objects.) 2431 // Global (weak) JNI handles 2432 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure()); 2433 2434 CodeCache::oops_do(adjust_pointer_closure()); 2435 SymbolTable::oops_do(adjust_root_pointer_closure()); 2436 StringTable::oops_do(adjust_root_pointer_closure()); 2437 ref_processor()->weak_oops_do(adjust_root_pointer_closure()); 2438 // Roots were visited so references into the young gen in roots 2439 // may have been scanned. Process them also. 2440 // Should the reference processor have a span that excludes 2441 // young gen objects? 2442 PSScavenge::reference_processor()->weak_oops_do( 2443 adjust_root_pointer_closure()); 2444 } 2445 2446 void PSParallelCompact::compact_perm(ParCompactionManager* cm) { 2447 EventMark m("4 compact perm"); 2448 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty); 2449 // trace("4"); 2450 2451 gc_heap()->perm_gen()->start_array()->reset(); 2452 move_and_update(cm, perm_space_id); 2453 } 2454 2455 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q, 2456 uint parallel_gc_threads) 2457 { 2458 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty); 2459 2460 const unsigned int task_count = MAX2(parallel_gc_threads, 1U); 2461 for (unsigned int j = 0; j < task_count; j++) { 2462 q->enqueue(new DrainStacksCompactionTask(j)); 2463 } 2464 2465 // Find all regions that are available (can be filled immediately) and 2466 // distribute them to the thread stacks. The iteration is done in reverse 2467 // order (high to low) so the regions will be removed in ascending order. 2468 2469 const ParallelCompactData& sd = PSParallelCompact::summary_data(); 2470 2471 size_t fillable_regions = 0; // A count for diagnostic purposes. 2472 unsigned int which = 0; // The worker thread number. 2473 2474 for (unsigned int id = to_space_id; id > perm_space_id; --id) { 2475 SpaceInfo* const space_info = _space_info + id; 2476 MutableSpace* const space = space_info->space(); 2477 HeapWord* const new_top = space_info->new_top(); 2478 2479 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix()); 2480 const size_t end_region = 2481 sd.addr_to_region_idx(sd.region_align_up(new_top)); 2482 assert(end_region > 0, "perm gen cannot be empty"); 2483 2484 for (size_t cur = end_region - 1; cur >= beg_region; --cur) { 2485 if (sd.region(cur)->claim_unsafe()) { 2486 ParCompactionManager* cm = ParCompactionManager::manager_array(which); 2487 cm->push_region(cur); 2488 2489 if (TraceParallelOldGCCompactionPhase && Verbose) { 2490 const size_t count_mod_8 = fillable_regions & 7; 2491 if (count_mod_8 == 0) gclog_or_tty->print("fillable: "); 2492 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur); 2493 if (count_mod_8 == 7) gclog_or_tty->cr(); 2494 } 2495 2496 NOT_PRODUCT(++fillable_regions;) 2497 2498 // Assign regions to threads in round-robin fashion. 2499 if (++which == task_count) { 2500 which = 0; 2501 } 2502 } 2503 } 2504 } 2505 2506 if (TraceParallelOldGCCompactionPhase) { 2507 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr(); 2508 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions); 2509 } 2510 } 2511 2512 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4 2513 2514 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q, 2515 uint parallel_gc_threads) { 2516 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty); 2517 2518 ParallelCompactData& sd = PSParallelCompact::summary_data(); 2519 2520 // Iterate over all the spaces adding tasks for updating 2521 // regions in the dense prefix. Assume that 1 gc thread 2522 // will work on opening the gaps and the remaining gc threads 2523 // will work on the dense prefix. 2524 unsigned int space_id; 2525 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) { 2526 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix(); 2527 const MutableSpace* const space = _space_info[space_id].space(); 2528 2529 if (dense_prefix_end == space->bottom()) { 2530 // There is no dense prefix for this space. 2531 continue; 2532 } 2533 2534 // The dense prefix is before this region. 2535 size_t region_index_end_dense_prefix = 2536 sd.addr_to_region_idx(dense_prefix_end); 2537 RegionData* const dense_prefix_cp = 2538 sd.region(region_index_end_dense_prefix); 2539 assert(dense_prefix_end == space->end() || 2540 dense_prefix_cp->available() || 2541 dense_prefix_cp->claimed(), 2542 "The region after the dense prefix should always be ready to fill"); 2543 2544 size_t region_index_start = sd.addr_to_region_idx(space->bottom()); 2545 2546 // Is there dense prefix work? 2547 size_t total_dense_prefix_regions = 2548 region_index_end_dense_prefix - region_index_start; 2549 // How many regions of the dense prefix should be given to 2550 // each thread? 2551 if (total_dense_prefix_regions > 0) { 2552 uint tasks_for_dense_prefix = 1; 2553 if (UseParallelDensePrefixUpdate) { 2554 if (total_dense_prefix_regions <= 2555 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) { 2556 // Don't over partition. This assumes that 2557 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value 2558 // so there are not many regions to process. 2559 tasks_for_dense_prefix = parallel_gc_threads; 2560 } else { 2561 // Over partition 2562 tasks_for_dense_prefix = parallel_gc_threads * 2563 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING; 2564 } 2565 } 2566 size_t regions_per_thread = total_dense_prefix_regions / 2567 tasks_for_dense_prefix; 2568 // Give each thread at least 1 region. 2569 if (regions_per_thread == 0) { 2570 regions_per_thread = 1; 2571 } 2572 2573 for (uint k = 0; k < tasks_for_dense_prefix; k++) { 2574 if (region_index_start >= region_index_end_dense_prefix) { 2575 break; 2576 } 2577 // region_index_end is not processed 2578 size_t region_index_end = MIN2(region_index_start + regions_per_thread, 2579 region_index_end_dense_prefix); 2580 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2581 region_index_start, 2582 region_index_end)); 2583 region_index_start = region_index_end; 2584 } 2585 } 2586 // This gets any part of the dense prefix that did not 2587 // fit evenly. 2588 if (region_index_start < region_index_end_dense_prefix) { 2589 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id), 2590 region_index_start, 2591 region_index_end_dense_prefix)); 2592 } 2593 } 2594 } 2595 2596 void PSParallelCompact::enqueue_region_stealing_tasks( 2597 GCTaskQueue* q, 2598 ParallelTaskTerminator* terminator_ptr, 2599 uint parallel_gc_threads) { 2600 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty); 2601 2602 // Once a thread has drained it's stack, it should try to steal regions from 2603 // other threads. 2604 if (parallel_gc_threads > 1) { 2605 for (uint j = 0; j < parallel_gc_threads; j++) { 2606 q->enqueue(new StealRegionCompactionTask(terminator_ptr)); 2607 } 2608 } 2609 } 2610 2611 void PSParallelCompact::compact() { 2612 EventMark m("5 compact"); 2613 // trace("5"); 2614 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty); 2615 2616 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2617 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2618 PSOldGen* old_gen = heap->old_gen(); 2619 old_gen->start_array()->reset(); 2620 uint parallel_gc_threads = heap->gc_task_manager()->workers(); 2621 TaskQueueSetSuper* qset = ParCompactionManager::region_array(); 2622 ParallelTaskTerminator terminator(parallel_gc_threads, qset); 2623 2624 GCTaskQueue* q = GCTaskQueue::create(); 2625 enqueue_region_draining_tasks(q, parallel_gc_threads); 2626 enqueue_dense_prefix_tasks(q, parallel_gc_threads); 2627 enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads); 2628 2629 { 2630 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty); 2631 2632 WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create(); 2633 q->enqueue(fin); 2634 2635 gc_task_manager()->add_list(q); 2636 2637 fin->wait_for(); 2638 2639 // We have to release the barrier tasks! 2640 WaitForBarrierGCTask::destroy(fin); 2641 2642 #ifdef ASSERT 2643 // Verify that all regions have been processed before the deferred updates. 2644 // Note that perm_space_id is skipped; this type of verification is not 2645 // valid until the perm gen is compacted by regions. 2646 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2647 verify_complete(SpaceId(id)); 2648 } 2649 #endif 2650 } 2651 2652 { 2653 // Update the deferred objects, if any. Any compaction manager can be used. 2654 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty); 2655 ParCompactionManager* cm = ParCompactionManager::manager_array(0); 2656 for (unsigned int id = old_space_id; id < last_space_id; ++id) { 2657 update_deferred_objects(cm, SpaceId(id)); 2658 } 2659 } 2660 } 2661 2662 #ifdef ASSERT 2663 void PSParallelCompact::verify_complete(SpaceId space_id) { 2664 // All Regions between space bottom() to new_top() should be marked as filled 2665 // and all Regions between new_top() and top() should be available (i.e., 2666 // should have been emptied). 2667 ParallelCompactData& sd = summary_data(); 2668 SpaceInfo si = _space_info[space_id]; 2669 HeapWord* new_top_addr = sd.region_align_up(si.new_top()); 2670 HeapWord* old_top_addr = sd.region_align_up(si.space()->top()); 2671 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom()); 2672 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr); 2673 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr); 2674 2675 bool issued_a_warning = false; 2676 2677 size_t cur_region; 2678 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) { 2679 const RegionData* const c = sd.region(cur_region); 2680 if (!c->completed()) { 2681 warning("region " SIZE_FORMAT " not filled: " 2682 "destination_count=" SIZE_FORMAT, 2683 cur_region, c->destination_count()); 2684 issued_a_warning = true; 2685 } 2686 } 2687 2688 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) { 2689 const RegionData* const c = sd.region(cur_region); 2690 if (!c->available()) { 2691 warning("region " SIZE_FORMAT " not empty: " 2692 "destination_count=" SIZE_FORMAT, 2693 cur_region, c->destination_count()); 2694 issued_a_warning = true; 2695 } 2696 } 2697 2698 if (issued_a_warning) { 2699 print_region_ranges(); 2700 } 2701 } 2702 #endif // #ifdef ASSERT 2703 2704 void PSParallelCompact::compact_serial(ParCompactionManager* cm) { 2705 EventMark m("5 compact serial"); 2706 TraceTime tm("compact serial", print_phases(), true, gclog_or_tty); 2707 2708 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap(); 2709 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity"); 2710 2711 PSYoungGen* young_gen = heap->young_gen(); 2712 PSOldGen* old_gen = heap->old_gen(); 2713 2714 old_gen->start_array()->reset(); 2715 old_gen->move_and_update(cm); 2716 young_gen->move_and_update(cm); 2717 } 2718 2719 void 2720 PSParallelCompact::follow_weak_klass_links() { 2721 // All klasses on the revisit stack are marked at this point. 2722 // Update and follow all subklass, sibling and implementor links. 2723 if (PrintRevisitStats) { 2724 gclog_or_tty->print_cr("#classes in system dictionary = %d", 2725 SystemDictionary::number_of_classes()); 2726 } 2727 for (uint i = 0; i < ParallelGCThreads + 1; i++) { 2728 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2729 KeepAliveClosure keep_alive_closure(cm); 2730 Stack<Klass*>* const rks = cm->revisit_klass_stack(); 2731 if (PrintRevisitStats) { 2732 gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT, 2733 i, rks->size()); 2734 } 2735 while (!rks->is_empty()) { 2736 Klass* const k = rks->pop(); 2737 k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure); 2738 } 2739 2740 cm->follow_marking_stacks(); 2741 } 2742 } 2743 2744 void 2745 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) { 2746 cm->revisit_klass_stack()->push(k); 2747 } 2748 2749 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) { 2750 cm->revisit_mdo_stack()->push(p); 2751 } 2752 2753 void PSParallelCompact::follow_mdo_weak_refs() { 2754 // All strongly reachable oops have been marked at this point; 2755 // we can visit and clear any weak references from MDO's which 2756 // we memoized during the strong marking phase. 2757 if (PrintRevisitStats) { 2758 gclog_or_tty->print_cr("#classes in system dictionary = %d", 2759 SystemDictionary::number_of_classes()); 2760 } 2761 for (uint i = 0; i < ParallelGCThreads + 1; i++) { 2762 ParCompactionManager* cm = ParCompactionManager::manager_array(i); 2763 Stack<DataLayout*>* rms = cm->revisit_mdo_stack(); 2764 if (PrintRevisitStats) { 2765 gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT, 2766 i, rms->size()); 2767 } 2768 while (!rms->is_empty()) { 2769 rms->pop()->follow_weak_refs(is_alive_closure()); 2770 } 2771 2772 cm->follow_marking_stacks(); 2773 } 2774 } 2775 2776 2777 #ifdef VALIDATE_MARK_SWEEP 2778 2779 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) { 2780 if (!ValidateMarkSweep) 2781 return; 2782 2783 if (!isroot) { 2784 if (_pointer_tracking) { 2785 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer"); 2786 _adjusted_pointers->remove(p); 2787 } 2788 } else { 2789 ptrdiff_t index = _root_refs_stack->find(p); 2790 if (index != -1) { 2791 int l = _root_refs_stack->length(); 2792 if (l > 0 && l - 1 != index) { 2793 void* last = _root_refs_stack->pop(); 2794 assert(last != p, "should be different"); 2795 _root_refs_stack->at_put(index, last); 2796 } else { 2797 _root_refs_stack->remove(p); 2798 } 2799 } 2800 } 2801 } 2802 2803 2804 void PSParallelCompact::check_adjust_pointer(void* p) { 2805 _adjusted_pointers->push(p); 2806 } 2807 2808 2809 class AdjusterTracker: public OopClosure { 2810 public: 2811 AdjusterTracker() {}; 2812 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); } 2813 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); } 2814 }; 2815 2816 2817 void PSParallelCompact::track_interior_pointers(oop obj) { 2818 if (ValidateMarkSweep) { 2819 _adjusted_pointers->clear(); 2820 _pointer_tracking = true; 2821 2822 AdjusterTracker checker; 2823 obj->oop_iterate(&checker); 2824 } 2825 } 2826 2827 2828 void PSParallelCompact::check_interior_pointers() { 2829 if (ValidateMarkSweep) { 2830 _pointer_tracking = false; 2831 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers"); 2832 } 2833 } 2834 2835 2836 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) { 2837 if (ValidateMarkSweep) { 2838 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops"); 2839 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0; 2840 } 2841 } 2842 2843 2844 void PSParallelCompact::register_live_oop(oop p, size_t size) { 2845 if (ValidateMarkSweep) { 2846 _live_oops->push(p); 2847 _live_oops_size->push(size); 2848 _live_oops_index++; 2849 } 2850 } 2851 2852 void PSParallelCompact::validate_live_oop(oop p, size_t size) { 2853 if (ValidateMarkSweep) { 2854 oop obj = _live_oops->at((int)_live_oops_index); 2855 guarantee(obj == p, "should be the same object"); 2856 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size"); 2857 _live_oops_index++; 2858 } 2859 } 2860 2861 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size, 2862 HeapWord* compaction_top) { 2863 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top), 2864 "should be moved to forwarded location"); 2865 if (ValidateMarkSweep) { 2866 PSParallelCompact::validate_live_oop(oop(q), size); 2867 _live_oops_moved_to->push(oop(compaction_top)); 2868 } 2869 if (RecordMarkSweepCompaction) { 2870 _cur_gc_live_oops->push(q); 2871 _cur_gc_live_oops_moved_to->push(compaction_top); 2872 _cur_gc_live_oops_size->push(size); 2873 } 2874 } 2875 2876 2877 void PSParallelCompact::compaction_complete() { 2878 if (RecordMarkSweepCompaction) { 2879 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops; 2880 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to; 2881 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size; 2882 2883 _cur_gc_live_oops = _last_gc_live_oops; 2884 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to; 2885 _cur_gc_live_oops_size = _last_gc_live_oops_size; 2886 _last_gc_live_oops = _tmp_live_oops; 2887 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to; 2888 _last_gc_live_oops_size = _tmp_live_oops_size; 2889 } 2890 } 2891 2892 2893 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) { 2894 if (!RecordMarkSweepCompaction) { 2895 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled"); 2896 return; 2897 } 2898 2899 if (_last_gc_live_oops == NULL) { 2900 tty->print_cr("No compaction information gathered yet"); 2901 return; 2902 } 2903 2904 for (int i = 0; i < _last_gc_live_oops->length(); i++) { 2905 HeapWord* old_oop = _last_gc_live_oops->at(i); 2906 size_t sz = _last_gc_live_oops_size->at(i); 2907 if (old_oop <= q && q < (old_oop + sz)) { 2908 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i); 2909 size_t offset = (q - old_oop); 2910 tty->print_cr("Address " PTR_FORMAT, q); 2911 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset); 2912 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset); 2913 return; 2914 } 2915 } 2916 2917 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q); 2918 } 2919 #endif //VALIDATE_MARK_SWEEP 2920 2921 // Update interior oops in the ranges of regions [beg_region, end_region). 2922 void 2923 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm, 2924 SpaceId space_id, 2925 size_t beg_region, 2926 size_t end_region) { 2927 ParallelCompactData& sd = summary_data(); 2928 ParMarkBitMap* const mbm = mark_bitmap(); 2929 2930 HeapWord* beg_addr = sd.region_to_addr(beg_region); 2931 HeapWord* const end_addr = sd.region_to_addr(end_region); 2932 assert(beg_region <= end_region, "bad region range"); 2933 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix"); 2934 2935 #ifdef ASSERT 2936 // Claim the regions to avoid triggering an assert when they are marked as 2937 // filled. 2938 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) { 2939 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed"); 2940 } 2941 #endif // #ifdef ASSERT 2942 2943 if (beg_addr != space(space_id)->bottom()) { 2944 // Find the first live object or block of dead space that *starts* in this 2945 // range of regions. If a partial object crosses onto the region, skip it; 2946 // it will be marked for 'deferred update' when the object head is 2947 // processed. If dead space crosses onto the region, it is also skipped; it 2948 // will be filled when the prior region is processed. If neither of those 2949 // apply, the first word in the region is the start of a live object or dead 2950 // space. 2951 assert(beg_addr > space(space_id)->bottom(), "sanity"); 2952 const RegionData* const cp = sd.region(beg_region); 2953 if (cp->partial_obj_size() != 0) { 2954 beg_addr = sd.partial_obj_end(beg_region); 2955 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) { 2956 beg_addr = mbm->find_obj_beg(beg_addr, end_addr); 2957 } 2958 } 2959 2960 if (beg_addr < end_addr) { 2961 // A live object or block of dead space starts in this range of Regions. 2962 HeapWord* const dense_prefix_end = dense_prefix(space_id); 2963 2964 // Create closures and iterate. 2965 UpdateOnlyClosure update_closure(mbm, cm, space_id); 2966 FillClosure fill_closure(cm, space_id); 2967 ParMarkBitMap::IterationStatus status; 2968 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr, 2969 dense_prefix_end); 2970 if (status == ParMarkBitMap::incomplete) { 2971 update_closure.do_addr(update_closure.source()); 2972 } 2973 } 2974 2975 // Mark the regions as filled. 2976 RegionData* const beg_cp = sd.region(beg_region); 2977 RegionData* const end_cp = sd.region(end_region); 2978 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) { 2979 cp->set_completed(); 2980 } 2981 } 2982 2983 // Return the SpaceId for the space containing addr. If addr is not in the 2984 // heap, last_space_id is returned. In debug mode it expects the address to be 2985 // in the heap and asserts such. 2986 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) { 2987 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap"); 2988 2989 for (unsigned int id = perm_space_id; id < last_space_id; ++id) { 2990 if (_space_info[id].space()->contains(addr)) { 2991 return SpaceId(id); 2992 } 2993 } 2994 2995 assert(false, "no space contains the addr"); 2996 return last_space_id; 2997 } 2998 2999 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm, 3000 SpaceId id) { 3001 assert(id < last_space_id, "bad space id"); 3002 3003 ParallelCompactData& sd = summary_data(); 3004 const SpaceInfo* const space_info = _space_info + id; 3005 ObjectStartArray* const start_array = space_info->start_array(); 3006 3007 const MutableSpace* const space = space_info->space(); 3008 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set"); 3009 HeapWord* const beg_addr = space_info->dense_prefix(); 3010 HeapWord* const end_addr = sd.region_align_up(space_info->new_top()); 3011 3012 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr); 3013 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr); 3014 const RegionData* cur_region; 3015 for (cur_region = beg_region; cur_region < end_region; ++cur_region) { 3016 HeapWord* const addr = cur_region->deferred_obj_addr(); 3017 if (addr != NULL) { 3018 if (start_array != NULL) { 3019 start_array->allocate_block(addr); 3020 } 3021 oop(addr)->update_contents(cm); 3022 assert(oop(addr)->is_oop_or_null(), "should be an oop now"); 3023 } 3024 } 3025 } 3026 3027 // Skip over count live words starting from beg, and return the address of the 3028 // next live word. Unless marked, the word corresponding to beg is assumed to 3029 // be dead. Callers must either ensure beg does not correspond to the middle of 3030 // an object, or account for those live words in some other way. Callers must 3031 // also ensure that there are enough live words in the range [beg, end) to skip. 3032 HeapWord* 3033 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count) 3034 { 3035 assert(count > 0, "sanity"); 3036 3037 ParMarkBitMap* m = mark_bitmap(); 3038 idx_t bits_to_skip = m->words_to_bits(count); 3039 idx_t cur_beg = m->addr_to_bit(beg); 3040 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end)); 3041 3042 do { 3043 cur_beg = m->find_obj_beg(cur_beg, search_end); 3044 idx_t cur_end = m->find_obj_end(cur_beg, search_end); 3045 const size_t obj_bits = cur_end - cur_beg + 1; 3046 if (obj_bits > bits_to_skip) { 3047 return m->bit_to_addr(cur_beg + bits_to_skip); 3048 } 3049 bits_to_skip -= obj_bits; 3050 cur_beg = cur_end + 1; 3051 } while (bits_to_skip > 0); 3052 3053 // Skipping the desired number of words landed just past the end of an object. 3054 // Find the start of the next object. 3055 cur_beg = m->find_obj_beg(cur_beg, search_end); 3056 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip"); 3057 return m->bit_to_addr(cur_beg); 3058 } 3059 3060 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr, 3061 SpaceId src_space_id, 3062 size_t src_region_idx) 3063 { 3064 assert(summary_data().is_region_aligned(dest_addr), "not aligned"); 3065 3066 const SplitInfo& split_info = _space_info[src_space_id].split_info(); 3067 if (split_info.dest_region_addr() == dest_addr) { 3068 // The partial object ending at the split point contains the first word to 3069 // be copied to dest_addr. 3070 return split_info.first_src_addr(); 3071 } 3072 3073 const ParallelCompactData& sd = summary_data(); 3074 ParMarkBitMap* const bitmap = mark_bitmap(); 3075 const size_t RegionSize = ParallelCompactData::RegionSize; 3076 3077 assert(sd.is_region_aligned(dest_addr), "not aligned"); 3078 const RegionData* const src_region_ptr = sd.region(src_region_idx); 3079 const size_t partial_obj_size = src_region_ptr->partial_obj_size(); 3080 HeapWord* const src_region_destination = src_region_ptr->destination(); 3081 3082 assert(dest_addr >= src_region_destination, "wrong src region"); 3083 assert(src_region_ptr->data_size() > 0, "src region cannot be empty"); 3084 3085 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx); 3086 HeapWord* const src_region_end = src_region_beg + RegionSize; 3087 3088 HeapWord* addr = src_region_beg; 3089 if (dest_addr == src_region_destination) { 3090 // Return the first live word in the source region. 3091 if (partial_obj_size == 0) { 3092 addr = bitmap->find_obj_beg(addr, src_region_end); 3093 assert(addr < src_region_end, "no objects start in src region"); 3094 } 3095 return addr; 3096 } 3097 3098 // Must skip some live data. 3099 size_t words_to_skip = dest_addr - src_region_destination; 3100 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region"); 3101 3102 if (partial_obj_size >= words_to_skip) { 3103 // All the live words to skip are part of the partial object. 3104 addr += words_to_skip; 3105 if (partial_obj_size == words_to_skip) { 3106 // Find the first live word past the partial object. 3107 addr = bitmap->find_obj_beg(addr, src_region_end); 3108 assert(addr < src_region_end, "wrong src region"); 3109 } 3110 return addr; 3111 } 3112 3113 // Skip over the partial object (if any). 3114 if (partial_obj_size != 0) { 3115 words_to_skip -= partial_obj_size; 3116 addr += partial_obj_size; 3117 } 3118 3119 // Skip over live words due to objects that start in the region. 3120 addr = skip_live_words(addr, src_region_end, words_to_skip); 3121 assert(addr < src_region_end, "wrong src region"); 3122 return addr; 3123 } 3124 3125 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm, 3126 SpaceId src_space_id, 3127 size_t beg_region, 3128 HeapWord* end_addr) 3129 { 3130 ParallelCompactData& sd = summary_data(); 3131 3132 #ifdef ASSERT 3133 MutableSpace* const src_space = _space_info[src_space_id].space(); 3134 HeapWord* const beg_addr = sd.region_to_addr(beg_region); 3135 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(), 3136 "src_space_id does not match beg_addr"); 3137 assert(src_space->contains(end_addr) || end_addr == src_space->end(), 3138 "src_space_id does not match end_addr"); 3139 #endif // #ifdef ASSERT 3140 3141 RegionData* const beg = sd.region(beg_region); 3142 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr)); 3143 3144 // Regions up to new_top() are enqueued if they become available. 3145 HeapWord* const new_top = _space_info[src_space_id].new_top(); 3146 RegionData* const enqueue_end = 3147 sd.addr_to_region_ptr(sd.region_align_up(new_top)); 3148 3149 for (RegionData* cur = beg; cur < end; ++cur) { 3150 assert(cur->data_size() > 0, "region must have live data"); 3151 cur->decrement_destination_count(); 3152 if (cur < enqueue_end && cur->available() && cur->claim()) { 3153 cm->push_region(sd.region(cur)); 3154 } 3155 } 3156 } 3157 3158 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure, 3159 SpaceId& src_space_id, 3160 HeapWord*& src_space_top, 3161 HeapWord* end_addr) 3162 { 3163 typedef ParallelCompactData::RegionData RegionData; 3164 3165 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3166 const size_t region_size = ParallelCompactData::RegionSize; 3167 3168 size_t src_region_idx = 0; 3169 3170 // Skip empty regions (if any) up to the top of the space. 3171 HeapWord* const src_aligned_up = sd.region_align_up(end_addr); 3172 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up); 3173 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top); 3174 const RegionData* const top_region_ptr = 3175 sd.addr_to_region_ptr(top_aligned_up); 3176 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) { 3177 ++src_region_ptr; 3178 } 3179 3180 if (src_region_ptr < top_region_ptr) { 3181 // The next source region is in the current space. Update src_region_idx 3182 // and the source address to match src_region_ptr. 3183 src_region_idx = sd.region(src_region_ptr); 3184 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx); 3185 if (src_region_addr > closure.source()) { 3186 closure.set_source(src_region_addr); 3187 } 3188 return src_region_idx; 3189 } 3190 3191 // Switch to a new source space and find the first non-empty region. 3192 unsigned int space_id = src_space_id + 1; 3193 assert(space_id < last_space_id, "not enough spaces"); 3194 3195 HeapWord* const destination = closure.destination(); 3196 3197 do { 3198 MutableSpace* space = _space_info[space_id].space(); 3199 HeapWord* const bottom = space->bottom(); 3200 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom); 3201 3202 // Iterate over the spaces that do not compact into themselves. 3203 if (bottom_cp->destination() != bottom) { 3204 HeapWord* const top_aligned_up = sd.region_align_up(space->top()); 3205 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up); 3206 3207 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) { 3208 if (src_cp->live_obj_size() > 0) { 3209 // Found it. 3210 assert(src_cp->destination() == destination, 3211 "first live obj in the space must match the destination"); 3212 assert(src_cp->partial_obj_size() == 0, 3213 "a space cannot begin with a partial obj"); 3214 3215 src_space_id = SpaceId(space_id); 3216 src_space_top = space->top(); 3217 const size_t src_region_idx = sd.region(src_cp); 3218 closure.set_source(sd.region_to_addr(src_region_idx)); 3219 return src_region_idx; 3220 } else { 3221 assert(src_cp->data_size() == 0, "sanity"); 3222 } 3223 } 3224 } 3225 } while (++space_id < last_space_id); 3226 3227 assert(false, "no source region was found"); 3228 return 0; 3229 } 3230 3231 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx) 3232 { 3233 typedef ParMarkBitMap::IterationStatus IterationStatus; 3234 const size_t RegionSize = ParallelCompactData::RegionSize; 3235 ParMarkBitMap* const bitmap = mark_bitmap(); 3236 ParallelCompactData& sd = summary_data(); 3237 RegionData* const region_ptr = sd.region(region_idx); 3238 3239 // Get the items needed to construct the closure. 3240 HeapWord* dest_addr = sd.region_to_addr(region_idx); 3241 SpaceId dest_space_id = space_id(dest_addr); 3242 ObjectStartArray* start_array = _space_info[dest_space_id].start_array(); 3243 HeapWord* new_top = _space_info[dest_space_id].new_top(); 3244 assert(dest_addr < new_top, "sanity"); 3245 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize); 3246 3247 // Get the source region and related info. 3248 size_t src_region_idx = region_ptr->source_region(); 3249 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx)); 3250 HeapWord* src_space_top = _space_info[src_space_id].space()->top(); 3251 3252 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3253 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx)); 3254 3255 // Adjust src_region_idx to prepare for decrementing destination counts (the 3256 // destination count is not decremented when a region is copied to itself). 3257 if (src_region_idx == region_idx) { 3258 src_region_idx += 1; 3259 } 3260 3261 if (bitmap->is_unmarked(closure.source())) { 3262 // The first source word is in the middle of an object; copy the remainder 3263 // of the object or as much as will fit. The fact that pointer updates were 3264 // deferred will be noted when the object header is processed. 3265 HeapWord* const old_src_addr = closure.source(); 3266 closure.copy_partial_obj(); 3267 if (closure.is_full()) { 3268 decrement_destination_counts(cm, src_space_id, src_region_idx, 3269 closure.source()); 3270 region_ptr->set_deferred_obj_addr(NULL); 3271 region_ptr->set_completed(); 3272 return; 3273 } 3274 3275 HeapWord* const end_addr = sd.region_align_down(closure.source()); 3276 if (sd.region_align_down(old_src_addr) != end_addr) { 3277 // The partial object was copied from more than one source region. 3278 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3279 3280 // Move to the next source region, possibly switching spaces as well. All 3281 // args except end_addr may be modified. 3282 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3283 end_addr); 3284 } 3285 } 3286 3287 do { 3288 HeapWord* const cur_addr = closure.source(); 3289 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1), 3290 src_space_top); 3291 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr); 3292 3293 if (status == ParMarkBitMap::incomplete) { 3294 // The last obj that starts in the source region does not end in the 3295 // region. 3296 assert(closure.source() < end_addr, "sanity"); 3297 HeapWord* const obj_beg = closure.source(); 3298 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(), 3299 src_space_top); 3300 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end); 3301 if (obj_end < range_end) { 3302 // The end was found; the entire object will fit. 3303 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end)); 3304 assert(status != ParMarkBitMap::would_overflow, "sanity"); 3305 } else { 3306 // The end was not found; the object will not fit. 3307 assert(range_end < src_space_top, "obj cannot cross space boundary"); 3308 status = ParMarkBitMap::would_overflow; 3309 } 3310 } 3311 3312 if (status == ParMarkBitMap::would_overflow) { 3313 // The last object did not fit. Note that interior oop updates were 3314 // deferred, then copy enough of the object to fill the region. 3315 region_ptr->set_deferred_obj_addr(closure.destination()); 3316 status = closure.copy_until_full(); // copies from closure.source() 3317 3318 decrement_destination_counts(cm, src_space_id, src_region_idx, 3319 closure.source()); 3320 region_ptr->set_completed(); 3321 return; 3322 } 3323 3324 if (status == ParMarkBitMap::full) { 3325 decrement_destination_counts(cm, src_space_id, src_region_idx, 3326 closure.source()); 3327 region_ptr->set_deferred_obj_addr(NULL); 3328 region_ptr->set_completed(); 3329 return; 3330 } 3331 3332 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr); 3333 3334 // Move to the next source region, possibly switching spaces as well. All 3335 // args except end_addr may be modified. 3336 src_region_idx = next_src_region(closure, src_space_id, src_space_top, 3337 end_addr); 3338 } while (true); 3339 } 3340 3341 void 3342 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) { 3343 const MutableSpace* sp = space(space_id); 3344 if (sp->is_empty()) { 3345 return; 3346 } 3347 3348 ParallelCompactData& sd = PSParallelCompact::summary_data(); 3349 ParMarkBitMap* const bitmap = mark_bitmap(); 3350 HeapWord* const dp_addr = dense_prefix(space_id); 3351 HeapWord* beg_addr = sp->bottom(); 3352 HeapWord* end_addr = sp->top(); 3353 3354 #ifdef ASSERT 3355 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix"); 3356 if (cm->should_verify_only()) { 3357 VerifyUpdateClosure verify_update(cm, sp); 3358 bitmap->iterate(&verify_update, beg_addr, end_addr); 3359 return; 3360 } 3361 3362 if (cm->should_reset_only()) { 3363 ResetObjectsClosure reset_objects(cm); 3364 bitmap->iterate(&reset_objects, beg_addr, end_addr); 3365 return; 3366 } 3367 #endif 3368 3369 const size_t beg_region = sd.addr_to_region_idx(beg_addr); 3370 const size_t dp_region = sd.addr_to_region_idx(dp_addr); 3371 if (beg_region < dp_region) { 3372 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region); 3373 } 3374 3375 // The destination of the first live object that starts in the region is one 3376 // past the end of the partial object entering the region (if any). 3377 HeapWord* const dest_addr = sd.partial_obj_end(dp_region); 3378 HeapWord* const new_top = _space_info[space_id].new_top(); 3379 assert(new_top >= dest_addr, "bad new_top value"); 3380 const size_t words = pointer_delta(new_top, dest_addr); 3381 3382 if (words > 0) { 3383 ObjectStartArray* start_array = _space_info[space_id].start_array(); 3384 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words); 3385 3386 ParMarkBitMap::IterationStatus status; 3387 status = bitmap->iterate(&closure, dest_addr, end_addr); 3388 assert(status == ParMarkBitMap::full, "iteration not complete"); 3389 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr, 3390 "live objects skipped because closure is full"); 3391 } 3392 } 3393 3394 jlong PSParallelCompact::millis_since_last_gc() { 3395 jlong ret_val = os::javaTimeMillis() - _time_of_last_gc; 3396 // XXX See note in genCollectedHeap::millis_since_last_gc(). 3397 if (ret_val < 0) { 3398 NOT_PRODUCT(warning("time warp: %d", ret_val);) 3399 return 0; 3400 } 3401 return ret_val; 3402 } 3403 3404 void PSParallelCompact::reset_millis_since_last_gc() { 3405 _time_of_last_gc = os::javaTimeMillis(); 3406 } 3407 3408 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full() 3409 { 3410 if (source() != destination()) { 3411 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3412 Copy::aligned_conjoint_words(source(), destination(), words_remaining()); 3413 } 3414 update_state(words_remaining()); 3415 assert(is_full(), "sanity"); 3416 return ParMarkBitMap::full; 3417 } 3418 3419 void MoveAndUpdateClosure::copy_partial_obj() 3420 { 3421 size_t words = words_remaining(); 3422 3423 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end()); 3424 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end); 3425 if (end_addr < range_end) { 3426 words = bitmap()->obj_size(source(), end_addr); 3427 } 3428 3429 // This test is necessary; if omitted, the pointer updates to a partial object 3430 // that crosses the dense prefix boundary could be overwritten. 3431 if (source() != destination()) { 3432 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3433 Copy::aligned_conjoint_words(source(), destination(), words); 3434 } 3435 update_state(words); 3436 } 3437 3438 ParMarkBitMapClosure::IterationStatus 3439 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3440 assert(destination() != NULL, "sanity"); 3441 assert(bitmap()->obj_size(addr) == words, "bad size"); 3442 3443 _source = addr; 3444 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) == 3445 destination(), "wrong destination"); 3446 3447 if (words > words_remaining()) { 3448 return ParMarkBitMap::would_overflow; 3449 } 3450 3451 // The start_array must be updated even if the object is not moving. 3452 if (_start_array != NULL) { 3453 _start_array->allocate_block(destination()); 3454 } 3455 3456 if (destination() != source()) { 3457 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());) 3458 Copy::aligned_conjoint_words(source(), destination(), words); 3459 } 3460 3461 oop moved_oop = (oop) destination(); 3462 moved_oop->update_contents(compaction_manager()); 3463 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point"); 3464 3465 update_state(words); 3466 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity"); 3467 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete; 3468 } 3469 3470 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm, 3471 ParCompactionManager* cm, 3472 PSParallelCompact::SpaceId space_id) : 3473 ParMarkBitMapClosure(mbm, cm), 3474 _space_id(space_id), 3475 _start_array(PSParallelCompact::start_array(space_id)) 3476 { 3477 } 3478 3479 // Updates the references in the object to their new values. 3480 ParMarkBitMapClosure::IterationStatus 3481 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) { 3482 do_addr(addr); 3483 return ParMarkBitMap::incomplete; 3484 } 3485 3486 // Verify the new location using the forwarding pointer 3487 // from MarkSweep::mark_sweep_phase2(). Set the mark_word 3488 // to the initial value. 3489 ParMarkBitMapClosure::IterationStatus 3490 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) { 3491 // The second arg (words) is not used. 3492 oop obj = (oop) addr; 3493 HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer(); 3494 HeapWord* new_pointer = summary_data().calc_new_pointer(obj); 3495 if (forwarding_ptr == NULL) { 3496 // The object is dead or not moving. 3497 assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj), 3498 "Object liveness is wrong."); 3499 return ParMarkBitMap::incomplete; 3500 } 3501 assert(UseParallelOldGCDensePrefix || 3502 (HeapMaximumCompactionInterval > 1) || 3503 (MarkSweepAlwaysCompactCount > 1) || 3504 (forwarding_ptr == new_pointer), 3505 "Calculation of new location is incorrect"); 3506 return ParMarkBitMap::incomplete; 3507 } 3508 3509 // Reset objects modified for debug checking. 3510 ParMarkBitMapClosure::IterationStatus 3511 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) { 3512 // The second arg (words) is not used. 3513 oop obj = (oop) addr; 3514 obj->init_mark(); 3515 return ParMarkBitMap::incomplete; 3516 } 3517 3518 // Prepare for compaction. This method is executed once 3519 // (i.e., by a single thread) before compaction. 3520 // Save the updated location of the intArrayKlassObj for 3521 // filling holes in the dense prefix. 3522 void PSParallelCompact::compact_prologue() { 3523 _updated_int_array_klass_obj = (klassOop) 3524 summary_data().calc_new_pointer(Universe::intArrayKlassObj()); 3525 } 3526