1 /*
   2  * Copyright (c) 2005, 2026, 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 "classfile/classLoaderDataGraph.hpp"
  26 #include "classfile/javaClasses.inline.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/symbolTable.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/codeCache.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/oopMap.hpp"
  33 #include "cppstdlib/new.hpp"
  34 #include "gc/parallel/objectStartArray.inline.hpp"
  35 #include "gc/parallel/parallelArguments.hpp"
  36 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
  37 #include "gc/parallel/parMarkBitMap.inline.hpp"
  38 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
  39 #include "gc/parallel/psCompactionManager.inline.hpp"
  40 #include "gc/parallel/psOldGen.hpp"
  41 #include "gc/parallel/psParallelCompact.inline.hpp"
  42 #include "gc/parallel/psPromotionManager.inline.hpp"
  43 #include "gc/parallel/psRootType.hpp"
  44 #include "gc/parallel/psScavenge.hpp"
  45 #include "gc/parallel/psStringDedup.hpp"
  46 #include "gc/parallel/psYoungGen.hpp"
  47 #include "gc/shared/classUnloadingContext.hpp"
  48 #include "gc/shared/collectedHeap.inline.hpp"
  49 #include "gc/shared/fullGCForwarding.inline.hpp"
  50 #include "gc/shared/gcCause.hpp"
  51 #include "gc/shared/gcHeapSummary.hpp"
  52 #include "gc/shared/gcId.hpp"
  53 #include "gc/shared/gcLocker.hpp"
  54 #include "gc/shared/gcTimer.hpp"
  55 #include "gc/shared/gcTrace.hpp"
  56 #include "gc/shared/gcTraceTime.inline.hpp"
  57 #include "gc/shared/gcVMOperations.hpp"
  58 #include "gc/shared/isGCActiveMark.hpp"
  59 #include "gc/shared/oopStorage.inline.hpp"
  60 #include "gc/shared/oopStorageSet.inline.hpp"
  61 #include "gc/shared/oopStorageSetParState.inline.hpp"
  62 #include "gc/shared/parallelCleaning.hpp"
  63 #include "gc/shared/preservedMarks.inline.hpp"
  64 #include "gc/shared/referencePolicy.hpp"
  65 #include "gc/shared/referenceProcessor.hpp"
  66 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
  67 #include "gc/shared/spaceDecorator.hpp"
  68 #include "gc/shared/taskTerminator.hpp"
  69 #include "gc/shared/weakProcessor.inline.hpp"
  70 #include "gc/shared/workerPolicy.hpp"
  71 #include "gc/shared/workerThread.hpp"
  72 #include "gc/shared/workerUtils.hpp"
  73 #include "logging/log.hpp"
  74 #include "memory/iterator.inline.hpp"
  75 #include "memory/memoryReserver.hpp"
  76 #include "memory/metaspaceUtils.hpp"
  77 #include "memory/resourceArea.hpp"
  78 #include "memory/universe.hpp"
  79 #include "nmt/memTracker.hpp"
  80 #include "oops/access.inline.hpp"
  81 #include "oops/instanceClassLoaderKlass.inline.hpp"
  82 #include "oops/instanceKlass.inline.hpp"
  83 #include "oops/instanceMirrorKlass.inline.hpp"
  84 #include "oops/methodData.hpp"
  85 #include "oops/objArrayKlass.inline.hpp"
  86 #include "oops/oop.inline.hpp"
  87 #include "runtime/atomicAccess.hpp"
  88 #include "runtime/handles.inline.hpp"
  89 #include "runtime/java.hpp"
  90 #include "runtime/safepoint.hpp"
  91 #include "runtime/threads.hpp"
  92 #include "runtime/vmThread.hpp"
  93 #include "services/memoryService.hpp"
  94 #include "utilities/align.hpp"
  95 #include "utilities/debug.hpp"
  96 #include "utilities/events.hpp"
  97 #include "utilities/formatBuffer.hpp"
  98 #include "utilities/macros.hpp"
  99 #include "utilities/stack.inline.hpp"
 100 #if INCLUDE_JVMCI
 101 #include "jvmci/jvmci.hpp"
 102 #endif
 103 
 104 #include <math.h>
 105 
 106 // All sizes are in HeapWords.
 107 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
 108 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
 109 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
 110 const size_t ParallelCompactData::RegionSizeBytes =
 111   RegionSize << LogHeapWordSize;
 112 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
 113 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
 114 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 115 
 116 const ParallelCompactData::RegionData::region_sz_t
 117 ParallelCompactData::RegionData::dc_shift = 27;
 118 
 119 const ParallelCompactData::RegionData::region_sz_t
 120 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 121 
 122 const ParallelCompactData::RegionData::region_sz_t
 123 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 124 
 125 const ParallelCompactData::RegionData::region_sz_t
 126 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 127 
 128 const ParallelCompactData::RegionData::region_sz_t
 129 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 130 
 131 const ParallelCompactData::RegionData::region_sz_t
 132 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 133 
 134 bool ParallelCompactData::RegionData::is_clear() {
 135   return (_destination == nullptr) &&
 136          (_source_region == 0) &&
 137          (_partial_obj_addr == nullptr) &&
 138          (_partial_obj_size == 0) &&
 139          (dc_and_los() == 0) &&
 140          (shadow_state() == 0);
 141 }
 142 
 143 #ifdef ASSERT
 144 void ParallelCompactData::RegionData::verify_clear() {
 145   assert(_destination == nullptr, "inv");
 146   assert(_source_region == 0, "inv");
 147   assert(_partial_obj_addr == nullptr, "inv");
 148   assert(_partial_obj_size == 0, "inv");
 149   assert(dc_and_los() == 0, "inv");
 150   assert(shadow_state() == 0, "inv");
 151 }
 152 #endif
 153 
 154 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 155 
 156 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 157 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
 158 
 159 void SplitInfo::record(size_t split_region_idx, HeapWord* split_point, size_t preceding_live_words) {
 160   assert(split_region_idx != 0, "precondition");
 161 
 162   // Obj denoted by split_point will be deferred to the next space.
 163   assert(split_point != nullptr, "precondition");
 164 
 165   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 166 
 167   PSParallelCompact::RegionData* split_region_ptr = sd.region(split_region_idx);
 168   assert(preceding_live_words < split_region_ptr->data_size(), "inv");
 169 
 170   HeapWord* preceding_destination = split_region_ptr->destination();
 171   assert(preceding_destination != nullptr, "inv");
 172 
 173   // How many regions does the preceding part occupy
 174   uint preceding_destination_count;
 175   if (preceding_live_words == 0) {
 176     preceding_destination_count = 0;
 177   } else {
 178     // -1 so that the ending address doesn't fall on the region-boundary
 179     if (sd.region_align_down(preceding_destination) ==
 180         sd.region_align_down(preceding_destination + preceding_live_words - 1)) {
 181       preceding_destination_count = 1;
 182     } else {
 183       preceding_destination_count = 2;
 184     }
 185   }
 186 
 187   _split_region_idx = split_region_idx;
 188   _split_point = split_point;
 189   _preceding_live_words = preceding_live_words;
 190   _preceding_destination = preceding_destination;
 191   _preceding_destination_count = preceding_destination_count;
 192 }
 193 
 194 void SplitInfo::clear()
 195 {
 196   _split_region_idx = 0;
 197   _split_point = nullptr;
 198   _preceding_live_words = 0;
 199   _preceding_destination = nullptr;
 200   _preceding_destination_count = 0;
 201   assert(!is_valid(), "sanity");
 202 }
 203 
 204 #ifdef  ASSERT
 205 void SplitInfo::verify_clear()
 206 {
 207   assert(_split_region_idx == 0, "not clear");
 208   assert(_split_point == nullptr, "not clear");
 209   assert(_preceding_live_words == 0, "not clear");
 210   assert(_preceding_destination == nullptr, "not clear");
 211   assert(_preceding_destination_count == 0, "not clear");
 212 }
 213 #endif  // #ifdef ASSERT
 214 
 215 
 216 void PSParallelCompact::print_on(outputStream* st) {
 217   _mark_bitmap.print_on(st);
 218 }
 219 
 220 ParallelCompactData::ParallelCompactData() :
 221   _heap_start(nullptr),
 222   DEBUG_ONLY(_heap_end(nullptr) COMMA)
 223   _region_vspace(nullptr),
 224   _reserved_byte_size(0),
 225   _region_data(nullptr),
 226   _region_count(0) {}
 227 
 228 bool ParallelCompactData::initialize(MemRegion reserved_heap)
 229 {
 230   _heap_start = reserved_heap.start();
 231   const size_t heap_size = reserved_heap.word_size();
 232   DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
 233 
 234   assert(region_align_down(_heap_start) == _heap_start,
 235          "region start not aligned");
 236 
 237   return initialize_region_data(heap_size);
 238 }
 239 
 240 PSVirtualSpace*
 241 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 242 {
 243   const size_t raw_bytes = count * element_size;
 244   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 245   const size_t granularity = os::vm_allocation_granularity();
 246   const size_t rs_align = MAX2(page_sz, granularity);
 247 
 248   _reserved_byte_size = align_up(raw_bytes, rs_align);
 249 
 250   ReservedSpace rs = MemoryReserver::reserve(_reserved_byte_size,
 251                                              rs_align,
 252                                              page_sz,
 253                                              mtGC);
 254 
 255   if (!rs.is_reserved()) {
 256     // Failed to reserve memory.
 257     return nullptr;
 258   }
 259 
 260   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
 261                        rs.size(), page_sz);
 262 
 263   MemTracker::record_virtual_memory_tag(rs, mtGC);
 264 
 265   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 266 
 267   if (!vspace->expand_by(_reserved_byte_size)) {
 268     // Failed to commit memory.
 269 
 270     delete vspace;
 271 
 272     // Release memory reserved in the space.
 273     MemoryReserver::release(rs);
 274 
 275     return nullptr;
 276   }
 277 
 278   return vspace;
 279 }
 280 
 281 bool ParallelCompactData::initialize_region_data(size_t heap_size)
 282 {
 283   assert(is_aligned(heap_size, RegionSize), "precondition");
 284 
 285   const size_t count = heap_size >> Log2RegionSize;
 286   _region_vspace = create_vspace(count, sizeof(RegionData));
 287   if (_region_vspace != nullptr) {
 288     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 289     _region_count = count;
 290     return true;
 291   }
 292   return false;
 293 }
 294 
 295 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 296   assert(beg_region <= _region_count, "beg_region out of range");
 297   assert(end_region <= _region_count, "end_region out of range");
 298 
 299   const size_t region_cnt = end_region - beg_region;
 300   for (size_t i = beg_region; i < end_region; i++) {
 301     ::new (&_region_data[i]) RegionData{};
 302   }
 303 }
 304 
 305 // The total live words on src_region would overflow the target space, so find
 306 // the overflowing object and record the split point. The invariant is that an
 307 // obj should not cross space boundary.
 308 HeapWord* ParallelCompactData::summarize_split_space(size_t src_region,
 309                                                      SplitInfo& split_info,
 310                                                      HeapWord* const destination,
 311                                                      HeapWord* const target_end,
 312                                                      HeapWord** target_next) {
 313   assert(destination <= target_end, "sanity");
 314   assert(destination + _region_data[src_region].data_size() > target_end,
 315     "region should not fit into target space");
 316   assert(is_region_aligned(target_end), "sanity");
 317 
 318   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 319 
 320   if (destination + partial_obj_size > target_end) {
 321     assert(partial_obj_size > 0, "inv");
 322     // The overflowing obj is from a previous region.
 323     //
 324     // source-regions:
 325     //
 326     // ***************
 327     // |     A|AA    |
 328     // ***************
 329     //       ^
 330     //       | split-point
 331     //
 332     // dest-region:
 333     //
 334     // ********
 335     // |~~~~A |
 336     // ********
 337     //       ^^
 338     //       || target-space-end
 339     //       |
 340     //       | destination
 341     //
 342     // AAA would overflow target-space.
 343     //
 344     HeapWord* overflowing_obj = _region_data[src_region].partial_obj_addr();
 345     size_t split_region = addr_to_region_idx(overflowing_obj);
 346 
 347     // The number of live words before the overflowing object on this split region
 348     size_t preceding_live_words;
 349     if (is_region_aligned(overflowing_obj)) {
 350       preceding_live_words = 0;
 351     } else {
 352       // Words accounted by the overflowing object on the split region
 353       size_t overflowing_size = pointer_delta(region_align_up(overflowing_obj), overflowing_obj);
 354       preceding_live_words = region(split_region)->data_size() - overflowing_size;
 355     }
 356 
 357     split_info.record(split_region, overflowing_obj, preceding_live_words);
 358 
 359     // The [overflowing_obj, src_region_start) part has been accounted for, so
 360     // must move back the new_top, now that this overflowing obj is deferred.
 361     HeapWord* new_top = destination - pointer_delta(region_to_addr(src_region), overflowing_obj);
 362 
 363     // If the overflowing obj was relocated to its original destination,
 364     // those destination regions would have their source_region set. Now that
 365     // this overflowing obj is relocated somewhere else, reset the
 366     // source_region.
 367     {
 368       size_t range_start = addr_to_region_idx(region_align_up(new_top));
 369       size_t range_end = addr_to_region_idx(region_align_up(destination));
 370       for (size_t i = range_start; i < range_end; ++i) {
 371         region(i)->set_source_region(0);
 372       }
 373     }
 374 
 375     // Update new top of target space
 376     *target_next = new_top;
 377 
 378     return overflowing_obj;
 379   }
 380 
 381   // Obj-iteration to locate the overflowing obj
 382   HeapWord* region_start = region_to_addr(src_region);
 383   HeapWord* region_end = region_start + RegionSize;
 384   HeapWord* cur_addr = region_start + partial_obj_size;
 385   size_t live_words = partial_obj_size;
 386 
 387   while (true) {
 388     assert(cur_addr < region_end, "inv");
 389     cur_addr = PSParallelCompact::mark_bitmap()->find_obj_beg(cur_addr, region_end);
 390     // There must be an overflowing obj in this region
 391     assert(cur_addr < region_end, "inv");
 392 
 393     oop obj = cast_to_oop(cur_addr);
 394     size_t obj_size = obj->size();
 395     if (destination + live_words + obj_size > target_end) {
 396       // Found the overflowing obj
 397       split_info.record(src_region, cur_addr, live_words);
 398       *target_next = destination + live_words;
 399       return cur_addr;
 400     }
 401 
 402     live_words += obj_size;
 403     cur_addr += obj_size;
 404   }
 405 }
 406 
 407 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
 408                                                 HeapWord** full_region_prefix_end) {
 409   size_t cur_region = addr_to_region_idx(space->bottom());
 410   const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
 411   size_t live_words = 0;
 412   if (full_region_prefix_end == nullptr) {
 413     for (/* empty */; cur_region < end_region; ++cur_region) {
 414       live_words += _region_data[cur_region].data_size();
 415     }
 416   } else {
 417     bool first_set = false;
 418     for (/* empty */; cur_region < end_region; ++cur_region) {
 419       size_t live_words_in_region = _region_data[cur_region].data_size();
 420       if (!first_set && live_words_in_region < RegionSize) {
 421         *full_region_prefix_end = region_to_addr(cur_region);
 422         first_set = true;
 423       }
 424       live_words += live_words_in_region;
 425     }
 426     if (!first_set) {
 427       // All regions are full of live objs.
 428       assert(is_region_aligned(space->top()), "inv");
 429       *full_region_prefix_end = space->top();
 430     }
 431     assert(*full_region_prefix_end != nullptr, "postcondition");
 432     assert(is_region_aligned(*full_region_prefix_end), "inv");
 433     assert(*full_region_prefix_end >= space->bottom(), "in-range");
 434     assert(*full_region_prefix_end <= space->top(), "in-range");
 435   }
 436   return live_words;
 437 }
 438 
 439 bool ParallelCompactData::summarize(SplitInfo& split_info,
 440                                     HeapWord* source_beg, HeapWord* source_end,
 441                                     HeapWord** source_next,
 442                                     HeapWord* target_beg, HeapWord* target_end,
 443                                     HeapWord** target_next)
 444 {
 445   HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
 446   log_develop_trace(gc, compaction)(
 447       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 448       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 449       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 450       p2i(target_beg), p2i(target_end), p2i(*target_next));
 451 
 452   size_t cur_region = addr_to_region_idx(source_beg);
 453   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 454 
 455   HeapWord *dest_addr = target_beg;
 456   for (/* empty */; cur_region < end_region; cur_region++) {
 457     size_t words = _region_data[cur_region].data_size();
 458 
 459     // Skip empty ones
 460     if (words == 0) {
 461       continue;
 462     }
 463 
 464     if (split_info.is_split(cur_region)) {
 465       assert(words > split_info.preceding_live_words(), "inv");
 466       words -= split_info.preceding_live_words();
 467     }
 468 
 469     _region_data[cur_region].set_destination(dest_addr);
 470 
 471     // If cur_region does not fit entirely into the target space, find a point
 472     // at which the source space can be 'split' so that part is copied to the
 473     // target space and the rest is copied elsewhere.
 474     if (dest_addr + words > target_end) {
 475       assert(source_next != nullptr, "source_next is null when splitting");
 476       *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 477                                            target_end, target_next);
 478       return false;
 479     }
 480 
 481     uint destination_count = split_info.is_split(cur_region)
 482                              ? split_info.preceding_destination_count()
 483                              : 0;
 484 
 485     HeapWord* const last_addr = dest_addr + words - 1;
 486     const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 487     const size_t dest_region_2 = addr_to_region_idx(last_addr);
 488 
 489     // Initially assume that the destination regions will be the same and
 490     // adjust the value below if necessary.  Under this assumption, if
 491     // cur_region == dest_region_2, then cur_region will be compacted
 492     // completely into itself.
 493     destination_count += cur_region == dest_region_2 ? 0 : 1;
 494     if (dest_region_1 != dest_region_2) {
 495       // Destination regions differ; adjust destination_count.
 496       destination_count += 1;
 497       // Data from cur_region will be copied to the start of dest_region_2.
 498       _region_data[dest_region_2].set_source_region(cur_region);
 499     } else if (is_region_aligned(dest_addr)) {
 500       // Data from cur_region will be copied to the start of the destination
 501       // region.
 502       _region_data[dest_region_1].set_source_region(cur_region);
 503     }
 504 
 505     _region_data[cur_region].set_destination_count(destination_count);
 506     dest_addr += words;
 507   }
 508 
 509   *target_next = dest_addr;
 510   return true;
 511 }
 512 
 513 #ifdef ASSERT
 514 void ParallelCompactData::verify_clear() {
 515   for (uint cur_idx = 0; cur_idx < region_count(); ++cur_idx) {
 516     if (!region(cur_idx)->is_clear()) {
 517       log_warning(gc)("Uncleared Region: %u", cur_idx);
 518       region(cur_idx)->verify_clear();
 519     }
 520   }
 521 }
 522 #endif  // #ifdef ASSERT
 523 
 524 STWGCTimer          PSParallelCompact::_gc_timer;
 525 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 526 elapsedTimer        PSParallelCompact::_accumulated_time;
 527 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 528 CollectorCounters*  PSParallelCompact::_counters = nullptr;
 529 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 530 ParallelCompactData PSParallelCompact::_summary_data;
 531 
 532 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 533 
 534 class PCAdjustPointerClosure: public BasicOopIterateClosure {
 535   template <typename T>
 536   void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
 537 
 538 public:
 539   virtual void do_oop(oop* p)                { do_oop_work(p); }
 540   virtual void do_oop(narrowOop* p)          { do_oop_work(p); }
 541 
 542   virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
 543 };
 544 
 545 static PCAdjustPointerClosure pc_adjust_pointer_closure;
 546 
 547 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 548 
 549 void PSParallelCompact::post_initialize() {
 550   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 551   _span_based_discoverer.set_span(heap->reserved_region());
 552   _ref_processor =
 553     new ReferenceProcessor(&_span_based_discoverer,
 554                            ParallelGCThreads,   // mt processing degree
 555                            ParallelGCThreads,   // mt discovery degree
 556                            false,               // concurrent_discovery
 557                            &_is_alive_closure); // non-header is alive closure
 558 
 559   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 560 
 561   // Initialize static fields in ParCompactionManager.
 562   ParCompactionManager::initialize(mark_bitmap());
 563 }
 564 
 565 bool PSParallelCompact::initialize_aux_data() {
 566   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 567   MemRegion mr = heap->reserved_region();
 568   assert(mr.byte_size() != 0, "heap should be reserved");
 569 
 570   initialize_space_info();
 571 
 572   if (!_mark_bitmap.initialize(mr)) {
 573     vm_shutdown_during_initialization(
 574       err_msg("Unable to allocate %zuKB bitmaps for parallel "
 575       "garbage collection for the requested %zuKB heap.",
 576       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 577     return false;
 578   }
 579 
 580   if (!_summary_data.initialize(mr)) {
 581     vm_shutdown_during_initialization(
 582       err_msg("Unable to allocate %zuKB card tables for parallel "
 583       "garbage collection for the requested %zuKB heap.",
 584       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 585     return false;
 586   }
 587 
 588   return true;
 589 }
 590 
 591 void PSParallelCompact::initialize_space_info()
 592 {
 593   memset(&_space_info, 0, sizeof(_space_info));
 594 
 595   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 596   PSYoungGen* young_gen = heap->young_gen();
 597 
 598   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 599   _space_info[eden_space_id].set_space(young_gen->eden_space());
 600   _space_info[from_space_id].set_space(young_gen->from_space());
 601   _space_info[to_space_id].set_space(young_gen->to_space());
 602 
 603   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 604 }
 605 
 606 void
 607 PSParallelCompact::clear_data_covering_space(SpaceId id)
 608 {
 609   // At this point, top is the value before GC, new_top() is the value that will
 610   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 611   // should be marked above top.  The summary data is cleared to the larger of
 612   // top & new_top.
 613   MutableSpace* const space = _space_info[id].space();
 614   HeapWord* const bot = space->bottom();
 615   HeapWord* const top = space->top();
 616   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 617 
 618   _mark_bitmap.clear_range(bot, top);
 619 
 620   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 621   const size_t end_region =
 622     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 623   _summary_data.clear_range(beg_region, end_region);
 624 
 625   // Clear the data used to 'split' regions.
 626   SplitInfo& split_info = _space_info[id].split_info();
 627   if (split_info.is_valid()) {
 628     split_info.clear();
 629   }
 630   DEBUG_ONLY(split_info.verify_clear();)
 631 }
 632 
 633 void PSParallelCompact::pre_compact()
 634 {
 635   // Update the from & to space pointers in space_info, since they are swapped
 636   // at each young gen gc.  Do the update unconditionally (even though a
 637   // promotion failure does not swap spaces) because an unknown number of young
 638   // collections will have swapped the spaces an unknown number of times.
 639   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 640   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 641   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 642   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 643 
 644   heap->increment_total_collections(true);
 645 
 646   CodeCache::on_gc_marking_cycle_start();
 647 
 648   heap->print_before_gc();
 649   heap->trace_heap_before_gc(&_gc_tracer);
 650 
 651   // Fill in TLABs
 652   heap->ensure_parsability(true);  // retire TLABs
 653 
 654   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 655     Universe::verify("Before GC");
 656   }
 657 
 658   DEBUG_ONLY(mark_bitmap()->verify_clear();)
 659   DEBUG_ONLY(summary_data().verify_clear();)
 660 }
 661 
 662 void PSParallelCompact::post_compact()
 663 {
 664   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
 665   ParCompactionManager::remove_all_shadow_regions();
 666 
 667   CodeCache::on_gc_marking_cycle_finish();
 668   CodeCache::arm_all_nmethods();
 669 
 670   // Need to clear claim bits for the next full-gc (marking and adjust-pointers).
 671   ClassLoaderDataGraph::clear_claimed_marks();
 672 
 673   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
 674     // Clear the marking bitmap, summary data and split info.
 675     clear_data_covering_space(SpaceId(id));
 676     {
 677       MutableSpace* space = _space_info[id].space();
 678       HeapWord* top = space->top();
 679       HeapWord* new_top = _space_info[id].new_top();
 680       if (ZapUnusedHeapArea && new_top < top) {
 681         space->mangle_region(MemRegion(new_top, top));
 682       }
 683       // Update top().  Must be done after clearing the bitmap and summary data.
 684       space->set_top(new_top);
 685     }
 686   }
 687 
 688 #ifdef ASSERT
 689   {
 690     mark_bitmap()->verify_clear();
 691     summary_data().verify_clear();
 692   }
 693 #endif
 694 
 695   ParCompactionManager::flush_all_string_dedup_requests();
 696 
 697   MutableSpace* const eden_space = _space_info[eden_space_id].space();
 698   MutableSpace* const from_space = _space_info[from_space_id].space();
 699   MutableSpace* const to_space   = _space_info[to_space_id].space();
 700 
 701   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 702   bool eden_empty = eden_space->is_empty();
 703 
 704   // Update heap occupancy information which is used as input to the soft ref
 705   // clearing policy at the next gc.
 706   Universe::heap()->update_capacity_and_used_at_gc();
 707 
 708   bool young_gen_empty = eden_empty && from_space->is_empty() &&
 709     to_space->is_empty();
 710 
 711   PSCardTable* ct = heap->card_table();
 712   MemRegion old_mr = heap->old_gen()->committed();
 713   if (young_gen_empty) {
 714     ct->clear_MemRegion(old_mr);
 715   } else {
 716     ct->dirty_MemRegion(old_mr);
 717   }
 718 
 719   heap->prune_scavengable_nmethods();
 720 
 721 #if COMPILER2_OR_JVMCI
 722   DerivedPointerTable::update_pointers();
 723 #endif
 724 
 725   // Signal that we have completed a visit to all live objects.
 726   Universe::heap()->record_whole_heap_examined_timestamp();
 727 }
 728 
 729 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
 730                                                                 HeapWord* full_region_prefix_end) {
 731   const size_t region_size = ParallelCompactData::RegionSize;
 732   const ParallelCompactData& sd = summary_data();
 733 
 734   // Iteration starts with the region *after* the full-region-prefix-end.
 735   const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
 736   // If final region is not full, iteration stops before that region,
 737   // because fill_dense_prefix_end assumes that prefix_end <= top.
 738   const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
 739   assert(start_region <= end_region, "inv");
 740 
 741   size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
 742   const RegionData* cur_region = start_region;
 743   for (/* empty */; cur_region < end_region; ++cur_region) {
 744     assert(region_size >= cur_region->data_size(), "inv");
 745     size_t dead_size = region_size - cur_region->data_size();
 746     if (max_waste < dead_size) {
 747       break;
 748     }
 749     max_waste -= dead_size;
 750   }
 751 
 752   HeapWord* const prefix_end = sd.region_to_addr(cur_region);
 753   assert(sd.is_region_aligned(prefix_end), "postcondition");
 754   assert(prefix_end >= full_region_prefix_end, "in-range");
 755   assert(prefix_end <= old_space->top(), "in-range");
 756   return prefix_end;
 757 }
 758 
 759 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
 760   // Comparing two sizes to decide if filling is required:
 761   //
 762   // The size of the filler (min-obj-size) is 2 heap words with the default
 763   // MinObjAlignment, since both markword and klass take 1 heap word.
 764   // With +UseCompactObjectHeaders, the minimum filler size is only one word,
 765   // because the Klass* gets encoded in the mark-word.
 766   //
 767   // The size of the gap (if any) right before dense-prefix-end is
 768   // MinObjAlignment.
 769   //
 770   // Need to fill in the gap only if it's smaller than min-obj-size, and the
 771   // filler obj will extend to next region.
 772 
 773   if (MinObjAlignment >= checked_cast<int>(CollectedHeap::min_fill_size())) {
 774     return;
 775   }
 776 
 777   assert(!UseCompactObjectHeaders, "Compact headers can allocate small objects");
 778   assert(CollectedHeap::min_fill_size() == 2, "inv");
 779   HeapWord* const dense_prefix_end = dense_prefix(id);
 780   assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
 781   assert(dense_prefix_end <= space(id)->top(), "precondition");
 782   if (dense_prefix_end == space(id)->top()) {
 783     // Must not have single-word gap right before prefix-end/top.
 784     return;
 785   }
 786   RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
 787 
 788   if (region_after_dense_prefix->partial_obj_size() != 0 ||
 789       _mark_bitmap.is_marked(dense_prefix_end)) {
 790     // The region after the dense prefix starts with live bytes.
 791     return;
 792   }
 793 
 794   HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
 795   if (block_start == dense_prefix_end - 1) {
 796     assert(!_mark_bitmap.is_marked(block_start), "inv");
 797     // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
 798     // The filler object will extend into region_after_dense_prefix.
 799     const size_t obj_len = 2; // min-fill-size
 800     HeapWord* const obj_beg = dense_prefix_end - 1;
 801     CollectedHeap::fill_with_object(obj_beg, obj_len);
 802     _mark_bitmap.mark_obj(obj_beg);
 803     _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
 804     region_after_dense_prefix->set_partial_obj_size(1);
 805     region_after_dense_prefix->set_partial_obj_addr(obj_beg);
 806     assert(start_array(id) != nullptr, "sanity");
 807     start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
 808   }
 809 }
 810 
 811 bool PSParallelCompact::check_maximum_compaction(bool should_do_max_compaction,
 812                                                  size_t total_live_words,
 813                                                  MutableSpace* const old_space,
 814                                                  HeapWord* full_region_prefix_end) {
 815 
 816   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 817 
 818   // Check System.GC
 819   bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
 820                           && GCCause::is_user_requested_gc(heap->gc_cause());
 821 
 822   // Check if all live objs are too much for old-gen.
 823   const bool is_old_gen_too_full = (total_live_words >= old_space->capacity_in_words());
 824 
 825   // If all regions in old-gen are full
 826   const bool is_region_full =
 827     full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
 828 
 829   return should_do_max_compaction
 830       || is_max_on_system_gc
 831       || is_old_gen_too_full
 832       || is_region_full;
 833 }
 834 
 835 void PSParallelCompact::summary_phase(bool should_do_max_compaction)
 836 {
 837   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
 838 
 839   MutableSpace* const old_space = _space_info[old_space_id].space();
 840   {
 841     size_t total_live_words = 0;
 842     HeapWord* full_region_prefix_end = nullptr;
 843     {
 844       // old-gen
 845       size_t live_words = _summary_data.live_words_in_space(old_space,
 846                                                             &full_region_prefix_end);
 847       total_live_words += live_words;
 848     }
 849     // young-gen
 850     for (uint i = eden_space_id; i < last_space_id; ++i) {
 851       const MutableSpace* space = _space_info[i].space();
 852       size_t live_words = _summary_data.live_words_in_space(space);
 853       total_live_words += live_words;
 854       _space_info[i].set_new_top(space->bottom() + live_words);
 855       _space_info[i].set_dense_prefix(space->bottom());
 856     }
 857 
 858     should_do_max_compaction = check_maximum_compaction(should_do_max_compaction,
 859                                                         total_live_words,
 860                                                         old_space,
 861                                                         full_region_prefix_end);
 862     {
 863       GCTraceTime(Info, gc, phases) tm("Summary Phase: expand", &_gc_timer);
 864       // Try to expand old-gen in order to fit all live objs and waste.
 865       size_t target_capacity_bytes = total_live_words * HeapWordSize
 866                                    + old_space->capacity_in_bytes() * (MarkSweepDeadRatio / 100);
 867       ParallelScavengeHeap::heap()->old_gen()->try_expand_till_size(target_capacity_bytes);
 868     }
 869 
 870     HeapWord* dense_prefix_end = should_do_max_compaction
 871                                  ? full_region_prefix_end
 872                                  : compute_dense_prefix_for_old_space(old_space,
 873                                                                       full_region_prefix_end);
 874     SpaceId id = old_space_id;
 875     _space_info[id].set_dense_prefix(dense_prefix_end);
 876 
 877     if (dense_prefix_end != old_space->bottom()) {
 878       fill_dense_prefix_end(id);
 879     }
 880 
 881     // Compacting objs in [dense_prefix_end, old_space->top())
 882     _summary_data.summarize(_space_info[id].split_info(),
 883                             dense_prefix_end, old_space->top(), nullptr,
 884                             dense_prefix_end, old_space->end(),
 885                             _space_info[id].new_top_addr());
 886   }
 887 
 888   // Summarize the remaining spaces in the young gen.  The initial target space
 889   // is the old gen.  If a space does not fit entirely into the target, then the
 890   // remainder is compacted into the space itself and that space becomes the new
 891   // target.
 892   SpaceId dst_space_id = old_space_id;
 893   HeapWord* dst_space_end = old_space->end();
 894   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
 895   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
 896     const MutableSpace* space = _space_info[id].space();
 897     const size_t live = pointer_delta(_space_info[id].new_top(),
 898                                       space->bottom());
 899     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
 900 
 901     if (live > 0 && live <= available) {
 902       // All the live data will fit.
 903       bool done = _summary_data.summarize(_space_info[id].split_info(),
 904                                           space->bottom(), space->top(),
 905                                           nullptr,
 906                                           *new_top_addr, dst_space_end,
 907                                           new_top_addr);
 908       assert(done, "space must fit into old gen");
 909 
 910       // Reset the new_top value for the space.
 911       _space_info[id].set_new_top(space->bottom());
 912     } else if (live > 0) {
 913       // Attempt to fit part of the source space into the target space.
 914       HeapWord* next_src_addr = nullptr;
 915       bool done = _summary_data.summarize(_space_info[id].split_info(),
 916                                           space->bottom(), space->top(),
 917                                           &next_src_addr,
 918                                           *new_top_addr, dst_space_end,
 919                                           new_top_addr);
 920       assert(!done, "space should not fit into old gen");
 921       assert(next_src_addr != nullptr, "sanity");
 922 
 923       // The source space becomes the new target, so the remainder is compacted
 924       // within the space itself.
 925       dst_space_id = SpaceId(id);
 926       dst_space_end = space->end();
 927       new_top_addr = _space_info[id].new_top_addr();
 928       done = _summary_data.summarize(_space_info[id].split_info(),
 929                                      next_src_addr, space->top(),
 930                                      nullptr,
 931                                      space->bottom(), dst_space_end,
 932                                      new_top_addr);
 933       assert(done, "space must fit when compacted into itself");
 934       assert(*new_top_addr <= space->top(), "usage should not grow");
 935     }
 936   }
 937 }
 938 
 939 void PSParallelCompact::report_object_count_after_gc() {
 940   GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
 941   // The heap is compacted, all objects are iterable. However there may be
 942   // filler objects in the heap which we should ignore.
 943   class SkipFillerObjectClosure : public BoolObjectClosure {
 944   public:
 945     bool do_object_b(oop obj) override { return !CollectedHeap::is_filler_object(obj); }
 946   } cl;
 947   _gc_tracer.report_object_count_after_gc(&cl, &ParallelScavengeHeap::heap()->workers());
 948 }
 949 
 950 bool PSParallelCompact::invoke(bool clear_all_soft_refs, bool should_do_max_compaction) {
 951   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
 952   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
 953          "should be in vm thread");
 954   assert(ref_processor() != nullptr, "Sanity");
 955 
 956   SvcGCMarker sgcm(SvcGCMarker::FULL);
 957   IsSTWGCActiveMark mark;
 958 
 959   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 960 
 961   GCIdMark gc_id_mark;
 962   _gc_timer.register_gc_start();
 963   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
 964 
 965   GCCause::Cause gc_cause = heap->gc_cause();
 966   PSOldGen* old_gen = heap->old_gen();
 967   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
 968 
 969   // Make sure data structures are sane, make the heap parsable, and do other
 970   // miscellaneous bookkeeping.
 971   pre_compact();
 972 
 973   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
 974 
 975   {
 976     const uint active_workers =
 977       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
 978                                         ParallelScavengeHeap::heap()->workers().active_workers(),
 979                                         Threads::number_of_non_daemon_threads());
 980     ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
 981 
 982     GCTraceCPUTime tcpu(&_gc_tracer);
 983     GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
 984 
 985     heap->pre_full_gc_dump(&_gc_timer);
 986 
 987     TraceCollectorStats tcs(counters());
 988     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
 989 
 990     if (log_is_enabled(Debug, gc, heap, exit)) {
 991       accumulated_time()->start();
 992     }
 993 
 994     // Let the size policy know we're starting
 995     size_policy->major_collection_begin();
 996 
 997 #if COMPILER2_OR_JVMCI
 998     DerivedPointerTable::clear();
 999 #endif
1000 
1001     ref_processor()->start_discovery(clear_all_soft_refs);
1002 
1003     marking_phase(&_gc_tracer);
1004 
1005     summary_phase(should_do_max_compaction);
1006 
1007 #if COMPILER2_OR_JVMCI
1008     assert(DerivedPointerTable::is_active(), "Sanity");
1009     DerivedPointerTable::set_active(false);
1010 #endif
1011 
1012     forward_to_new_addr();
1013 
1014     adjust_pointers();
1015 
1016     compact();
1017 
1018     ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1019 
1020     ParCompactionManager::verify_all_region_stack_empty();
1021 
1022     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1023     // done before resizing.
1024     post_compact();
1025 
1026     size_policy->major_collection_end();
1027 
1028     size_policy->sample_old_gen_used_bytes(MAX2(pre_gc_values.old_gen_used(), old_gen->used_in_bytes()));
1029 
1030     if (UseAdaptiveSizePolicy) {
1031       heap->resize_after_full_gc();
1032     }
1033 
1034     heap->resize_all_tlabs();
1035 
1036     // Resize the metaspace capacity after a collection
1037     MetaspaceGC::compute_new_size();
1038 
1039     if (log_is_enabled(Debug, gc, heap, exit)) {
1040       accumulated_time()->stop();
1041     }
1042 
1043     heap->print_heap_change(pre_gc_values);
1044 
1045     report_object_count_after_gc();
1046 
1047     // Track memory usage and detect low memory
1048     MemoryService::track_memory_usage();
1049     heap->update_counters();
1050 
1051     heap->post_full_gc_dump(&_gc_timer);
1052 
1053     size_policy->record_gc_pause_end_instant();
1054   }
1055 
1056   heap->gc_epilogue(true);
1057 
1058   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1059     Universe::verify("After GC");
1060   }
1061 
1062   heap->print_after_gc();
1063   heap->trace_heap_after_gc(&_gc_tracer);
1064 
1065   _gc_timer.register_gc_end();
1066 
1067   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1068   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1069 
1070   return true;
1071 }
1072 
1073 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1074   ParCompactionManager* _cm;
1075 
1076 public:
1077   PCAddThreadRootsMarkingTaskClosure(ParCompactionManager* cm) : _cm(cm) { }
1078   void do_thread(Thread* thread) {
1079     ResourceMark rm;
1080 
1081     MarkingNMethodClosure mark_and_push_in_blobs(&_cm->_mark_and_push_closure);
1082 
1083     thread->oops_do(&_cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1084 
1085     // Do the real work
1086     _cm->follow_marking_stacks();
1087   }
1088 };
1089 
1090 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1091   assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1092 
1093   ParCompactionManager* cm =
1094     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1095 
1096   do {
1097     ScannerTask task;
1098     if (ParCompactionManager::steal(worker_id, task)) {
1099       cm->follow_contents(task, true);
1100     }
1101     cm->follow_marking_stacks();
1102   } while (!terminator.offer_termination());
1103 }
1104 
1105 class MarkFromRootsTask : public WorkerTask {
1106   NMethodMarkingScope _nmethod_marking_scope;
1107   ThreadsClaimTokenScope _threads_claim_token_scope;
1108   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1109   TaskTerminator _terminator;
1110   uint _active_workers;
1111 
1112 public:
1113   MarkFromRootsTask(uint active_workers) :
1114       WorkerTask("MarkFromRootsTask"),
1115       _nmethod_marking_scope(),
1116       _threads_claim_token_scope(),
1117       _terminator(active_workers, ParCompactionManager::marking_stacks()),
1118       _active_workers(active_workers) {}
1119 
1120   virtual void work(uint worker_id) {
1121     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1122     cm->create_marking_stats_cache();
1123     {
1124       CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1125       ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1126 
1127       // Do the real work
1128       cm->follow_marking_stacks();
1129     }
1130 
1131     {
1132       PCAddThreadRootsMarkingTaskClosure closure(cm);
1133       Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1134     }
1135 
1136     // Mark from OopStorages
1137     {
1138       _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1139       // Do the real work
1140       cm->follow_marking_stacks();
1141     }
1142 
1143     if (_active_workers > 1) {
1144       steal_marking_work(_terminator, worker_id);
1145     }
1146   }
1147 };
1148 
1149 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1150   TaskTerminator _terminator;
1151 
1152 public:
1153   ParallelCompactRefProcProxyTask(uint max_workers)
1154     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1155       _terminator(_max_workers, ParCompactionManager::marking_stacks()) {}
1156 
1157   void work(uint worker_id) override {
1158     assert(worker_id < _max_workers, "sanity");
1159     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1160     BarrierEnqueueDiscoveredFieldClosure enqueue;
1161     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1162     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1163   }
1164 
1165   void prepare_run_task_hook() override {
1166     _terminator.reset_for_reuse(_queue_count);
1167   }
1168 };
1169 
1170 static void flush_marking_stats_cache(const uint num_workers) {
1171   for (uint i = 0; i < num_workers; ++i) {
1172     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1173     cm->flush_and_destroy_marking_stats_cache();
1174   }
1175 }
1176 
1177 class PSParallelCleaningTask : public WorkerTask {
1178   bool                    _unloading_occurred;
1179   CodeCacheUnloadingTask  _code_cache_task;
1180   // Prune dead klasses from subklass/sibling/implementor lists.
1181   KlassCleaningTask       _klass_cleaning_task;
1182 
1183 public:
1184   PSParallelCleaningTask(bool unloading_occurred) :
1185     WorkerTask("PS Parallel Cleaning"),
1186     _unloading_occurred(unloading_occurred),
1187     _code_cache_task(unloading_occurred),
1188     _klass_cleaning_task() {}
1189 
1190   void work(uint worker_id) {
1191 #if INCLUDE_JVMCI
1192     if (EnableJVMCI && worker_id == 0) {
1193       // Serial work; only first worker.
1194       // Clean JVMCI metadata handles.
1195       JVMCI::do_unloading(_unloading_occurred);
1196     }
1197 #endif
1198 
1199     // Do first pass of code cache cleaning.
1200     _code_cache_task.work(worker_id);
1201 
1202     // Clean all klasses that were not unloaded.
1203     // The weak metadata in klass doesn't need to be
1204     // processed if there was no unloading.
1205     if (_unloading_occurred) {
1206       _klass_cleaning_task.work();
1207     }
1208   }
1209 };
1210 
1211 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1212   // Recursively traverse all live objects and mark them
1213   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1214 
1215   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1216 
1217   ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1218   {
1219     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1220 
1221     MarkFromRootsTask task(active_gc_threads);
1222     ParallelScavengeHeap::heap()->workers().run_task(&task);
1223   }
1224 
1225   // Process reference objects found during marking
1226   {
1227     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1228 
1229     ReferenceProcessorStats stats;
1230     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1231 
1232     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1233     stats = ref_processor()->process_discovered_references(task, &ParallelScavengeHeap::heap()->workers(), pt);
1234 
1235     gc_tracer->report_gc_reference_stats(stats);
1236     pt.print_all_references();
1237   }
1238 
1239   {
1240     GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1241 
1242     flush_marking_stats_cache(active_gc_threads);
1243   }
1244 
1245   // This is the point where the entire marking should have completed.
1246   ParCompactionManager::verify_all_marking_stack_empty();
1247 
1248   {
1249     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1250     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1251                                 is_alive_closure(),
1252                                 &do_nothing_cl,
1253                                 1);
1254   }
1255 
1256   {
1257     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1258 
1259     ClassUnloadingContext ctx(active_gc_threads /* num_nmethod_unlink_workers */,
1260                               false /* unregister_nmethods_during_purge */,
1261                               false /* lock_nmethod_free_separately */);
1262 
1263     {
1264       CodeCache::UnlinkingScope scope(is_alive_closure());
1265 
1266       // Follow system dictionary roots and unload classes.
1267       bool unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1268 
1269       PSParallelCleaningTask task{unloading_occurred};
1270       ParallelScavengeHeap::heap()->workers().run_task(&task);
1271     }
1272 
1273     {
1274       GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1275       // Release unloaded nmethod's memory.
1276       ctx.purge_nmethods();
1277     }
1278     {
1279       GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1280       ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1281     }
1282     {
1283       GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1284       ctx.free_nmethods();
1285     }
1286     {
1287       // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1288       GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
1289       ClassLoaderDataGraph::purge(true /* at_safepoint */);
1290       DEBUG_ONLY(MetaspaceUtils::verify();)
1291     }
1292   }
1293 
1294 #if TASKQUEUE_STATS
1295   ParCompactionManager::print_and_reset_taskqueue_stats();
1296 #endif
1297 }
1298 
1299 template<typename Func>
1300 void PSParallelCompact::adjust_in_space_helper(SpaceId id, Atomic<uint>* claim_counter, Func&& on_stripe) {
1301   MutableSpace* sp = PSParallelCompact::space(id);
1302   HeapWord* const bottom = sp->bottom();
1303   HeapWord* const top = sp->top();
1304   if (bottom == top) {
1305     return;
1306   }
1307 
1308   const uint num_regions_per_stripe = 2;
1309   const size_t region_size = ParallelCompactData::RegionSize;
1310   const size_t stripe_size = num_regions_per_stripe * region_size;
1311 
1312   while (true) {
1313     uint counter = claim_counter->fetch_then_add(num_regions_per_stripe);
1314     HeapWord* cur_stripe = bottom + counter * region_size;
1315     if (cur_stripe >= top) {
1316       break;
1317     }
1318     HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1319     on_stripe(cur_stripe, stripe_end);
1320   }
1321 }
1322 
1323 void PSParallelCompact::adjust_in_old_space(Atomic<uint>* claim_counter) {
1324   // Regions in old-space shouldn't be split.
1325   assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1326 
1327   auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1328     assert(mark_bitmap()->is_marked(obj_start), "inv");
1329     oop obj = cast_to_oop(obj_start);
1330     return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1331   };
1332 
1333   adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1334     assert(_summary_data.is_region_aligned(stripe_start), "inv");
1335     RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1336     HeapWord* obj_start;
1337     if (cur_region->partial_obj_size() != 0) {
1338       obj_start = cur_region->partial_obj_addr();
1339       obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1340     } else {
1341       obj_start = stripe_start;
1342     }
1343 
1344     while (obj_start < stripe_end) {
1345       obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1346       if (obj_start >= stripe_end) {
1347         break;
1348       }
1349       obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1350     }
1351   });
1352 }
1353 
1354 void PSParallelCompact::adjust_in_young_space(SpaceId id, Atomic<uint>* claim_counter) {
1355   adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1356     HeapWord* obj_start = stripe_start;
1357     while (obj_start < stripe_end) {
1358       obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1359       if (obj_start >= stripe_end) {
1360         break;
1361       }
1362       oop obj = cast_to_oop(obj_start);
1363       obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1364     }
1365   });
1366 }
1367 
1368 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, Atomic<uint>* claim_counters) {
1369   auto start_time = Ticks::now();
1370   adjust_in_old_space(&claim_counters[0]);
1371   for (uint id = eden_space_id; id < last_space_id; ++id) {
1372     adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1373   }
1374   log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1375 }
1376 
1377 class PSAdjustTask final : public WorkerTask {
1378   ThreadsClaimTokenScope                     _threads_claim_token_scope;
1379   WeakProcessor::Task                        _weak_proc_task;
1380   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
1381   uint                                       _nworkers;
1382   Atomic<bool>                               _code_cache_claimed;
1383   Atomic<uint> _claim_counters[PSParallelCompact::last_space_id];
1384 
1385   bool try_claim_code_cache_task() {
1386     return _code_cache_claimed.load_relaxed() == false
1387         && _code_cache_claimed.compare_set(false, true);
1388   }
1389 
1390 public:
1391   PSAdjustTask(uint nworkers) :
1392     WorkerTask("PSAdjust task"),
1393     _threads_claim_token_scope(),
1394     _weak_proc_task(nworkers),
1395     _oop_storage_iter(),
1396     _nworkers(nworkers),
1397     _code_cache_claimed(false) {
1398 
1399     for (unsigned int i = PSParallelCompact::old_space_id; i < PSParallelCompact::last_space_id; ++i) {
1400       ::new (&_claim_counters[i]) Atomic<uint>{};
1401     }
1402     ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1403   }
1404 
1405   void work(uint worker_id) {
1406     {
1407       // Pointers in heap.
1408       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1409       cm->preserved_marks()->adjust_during_full_gc();
1410 
1411       PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1412     }
1413 
1414     {
1415       // All (strong and weak) CLDs.
1416       CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1417       ClassLoaderDataGraph::cld_do(&cld_closure);
1418     }
1419 
1420     {
1421       // Threads stack frames. No need to visit on-stack nmethods, because all
1422       // nmethods are visited in one go via CodeCache::nmethods_do.
1423       ResourceMark rm;
1424       Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1425       if (try_claim_code_cache_task()) {
1426         NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1427         CodeCache::nmethods_do(&adjust_code);
1428       }
1429     }
1430 
1431     {
1432       // VM internal strong and weak roots.
1433       _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1434       AlwaysTrueClosure always_alive;
1435       _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1436     }
1437   }
1438 };
1439 
1440 void PSParallelCompact::adjust_pointers() {
1441   // Adjust the pointers to reflect the new locations
1442   GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1443   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1444   PSAdjustTask task(nworkers);
1445   ParallelScavengeHeap::heap()->workers().run_task(&task);
1446 }
1447 
1448 // Split [start, end) evenly for a number of workers and return the
1449 // range for worker_id.
1450 static void split_regions_for_worker(size_t start, size_t end,
1451                                      uint worker_id, uint num_workers,
1452                                      size_t* worker_start, size_t* worker_end) {
1453   assert(start < end, "precondition");
1454   assert(num_workers > 0, "precondition");
1455   assert(worker_id < num_workers, "precondition");
1456 
1457   size_t num_regions = end - start;
1458   size_t num_regions_per_worker = num_regions / num_workers;
1459   size_t remainder = num_regions % num_workers;
1460   // The first few workers will get one extra.
1461   *worker_start = start + worker_id * num_regions_per_worker
1462                   + MIN2(checked_cast<size_t>(worker_id), remainder);
1463   *worker_end = *worker_start + num_regions_per_worker
1464                 + (worker_id < remainder ? 1 : 0);
1465 }
1466 
1467 void PSParallelCompact::forward_to_new_addr() {
1468   GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1469   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1470 
1471   struct ForwardTask final : public WorkerTask {
1472     uint _num_workers;
1473 
1474     explicit ForwardTask(uint num_workers) :
1475       WorkerTask("PSForward task"),
1476       _num_workers(num_workers) {}
1477 
1478     static void forward_objs_in_range(ParCompactionManager* cm,
1479                                       HeapWord* start,
1480                                       HeapWord* end,
1481                                       HeapWord* destination) {
1482       HeapWord* cur_addr = start;
1483       HeapWord* new_addr = destination;
1484 
1485       while (cur_addr < end) {
1486         cur_addr = mark_bitmap()->find_obj_beg(cur_addr, end);
1487         if (cur_addr >= end) {
1488           return;
1489         }
1490         assert(mark_bitmap()->is_marked(cur_addr), "inv");
1491         oop obj = cast_to_oop(cur_addr);
1492         if (new_addr != cur_addr) {
1493           cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1494           FullGCForwarding::forward_to(obj, cast_to_oop(new_addr));
1495         }
1496         size_t obj_size = obj->size();
1497         new_addr += obj_size;
1498         cur_addr += obj_size;
1499       }
1500     }
1501 
1502     void work(uint worker_id) override {
1503       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1504       for (uint id = old_space_id; id < last_space_id; ++id) {
1505         MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1506         HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1507         HeapWord* top = sp->top();
1508 
1509         if (dense_prefix_addr == top) {
1510           // Empty space
1511           continue;
1512         }
1513 
1514         const SplitInfo& split_info = _space_info[SpaceId(id)].split_info();
1515         size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1516         size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1517         size_t start_region;
1518         size_t end_region;
1519         split_regions_for_worker(dense_prefix_region, top_region,
1520                                  worker_id, _num_workers,
1521                                  &start_region, &end_region);
1522         for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1523           RegionData* region_ptr = _summary_data.region(cur_region);
1524           size_t partial_obj_size = region_ptr->partial_obj_size();
1525 
1526           if (partial_obj_size == ParallelCompactData::RegionSize) {
1527             // No obj-start
1528             continue;
1529           }
1530 
1531           HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1532           HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1533 
1534           if (split_info.is_split(cur_region)) {
1535             // Part 1: will be relocated to space-1
1536             HeapWord* preceding_destination = split_info.preceding_destination();
1537             HeapWord* split_point = split_info.split_point();
1538             forward_objs_in_range(cm, region_start + partial_obj_size, split_point, preceding_destination + partial_obj_size);
1539 
1540             // Part 2: will be relocated to space-2
1541             HeapWord* destination = region_ptr->destination();
1542             forward_objs_in_range(cm, split_point, region_end, destination);
1543           } else {
1544             HeapWord* destination = region_ptr->destination();
1545             forward_objs_in_range(cm, region_start + partial_obj_size, region_end, destination + partial_obj_size);
1546           }
1547         }
1548       }
1549     }
1550   } task(nworkers);
1551 
1552   ParallelScavengeHeap::heap()->workers().run_task(&task);
1553   DEBUG_ONLY(verify_forward();)
1554 }
1555 
1556 #ifdef ASSERT
1557 void PSParallelCompact::verify_forward() {
1558   HeapWord* const old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1559   // The destination addr for the first live obj after dense-prefix.
1560   HeapWord* bump_ptr = old_dense_prefix_addr
1561                      + _summary_data.addr_to_region_ptr(old_dense_prefix_addr)->partial_obj_size();
1562   SpaceId bump_ptr_space = old_space_id;
1563 
1564   for (uint id = old_space_id; id < last_space_id; ++id) {
1565     MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1566     // Only verify objs after dense-prefix, because those before dense-prefix are not moved (forwarded).
1567     HeapWord* cur_addr = dense_prefix(SpaceId(id));
1568     HeapWord* top = sp->top();
1569 
1570     while (cur_addr < top) {
1571       cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1572       if (cur_addr >= top) {
1573         break;
1574       }
1575       assert(mark_bitmap()->is_marked(cur_addr), "inv");
1576       assert(bump_ptr <= _space_info[bump_ptr_space].new_top(), "inv");
1577       // Move to the space containing cur_addr
1578       if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1579         bump_ptr = space(space_id(cur_addr))->bottom();
1580         bump_ptr_space = space_id(bump_ptr);
1581       }
1582       oop obj = cast_to_oop(cur_addr);
1583       if (cur_addr == bump_ptr) {
1584         assert(!FullGCForwarding::is_forwarded(obj), "inv");
1585       } else {
1586         assert(FullGCForwarding::forwardee(obj) == cast_to_oop(bump_ptr), "inv");
1587       }
1588       bump_ptr += obj->size();
1589       cur_addr += obj->size();
1590     }
1591   }
1592 }
1593 #endif
1594 
1595 // Helper class to print 8 region numbers per line and then print the total at the end.
1596 class FillableRegionLogger : public StackObj {
1597 private:
1598   Log(gc, compaction) log;
1599   static const int LineLength = 8;
1600   size_t _regions[LineLength];
1601   int _next_index;
1602   bool _enabled;
1603   size_t _total_regions;
1604 public:
1605   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1606   ~FillableRegionLogger() {
1607     log.trace("%zu initially fillable regions", _total_regions);
1608   }
1609 
1610   void print_line() {
1611     if (!_enabled || _next_index == 0) {
1612       return;
1613     }
1614     FormatBuffer<> line("Fillable: ");
1615     for (int i = 0; i < _next_index; i++) {
1616       line.append(" %7zu", _regions[i]);
1617     }
1618     log.trace("%s", line.buffer());
1619     _next_index = 0;
1620   }
1621 
1622   void handle(size_t region) {
1623     if (!_enabled) {
1624       return;
1625     }
1626     _regions[_next_index++] = region;
1627     if (_next_index == LineLength) {
1628       print_line();
1629     }
1630     _total_regions++;
1631   }
1632 };
1633 
1634 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1635 {
1636   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1637 
1638   // Find the threads that are active
1639   uint worker_id = 0;
1640 
1641   // Find all regions that are available (can be filled immediately) and
1642   // distribute them to the thread stacks.  The iteration is done in reverse
1643   // order (high to low) so the regions will be removed in ascending order.
1644 
1645   const ParallelCompactData& sd = PSParallelCompact::summary_data();
1646 
1647   // id + 1 is used to test termination so unsigned  can
1648   // be used with an old_space_id == 0.
1649   FillableRegionLogger region_logger;
1650   for (unsigned int id = last_space_id - 1; id + 1 > old_space_id; --id) {
1651     SpaceInfo* const space_info = _space_info + id;
1652     HeapWord* const new_top = space_info->new_top();
1653 
1654     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1655     const size_t end_region =
1656       sd.addr_to_region_idx(sd.region_align_up(new_top));
1657 
1658     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1659       if (sd.region(cur)->claim_unsafe()) {
1660         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1661         bool result = sd.region(cur)->mark_normal();
1662         assert(result, "Must succeed at this point.");
1663         cm->region_stack()->push(cur);
1664         region_logger.handle(cur);
1665         // Assign regions to tasks in round-robin fashion.
1666         if (++worker_id == parallel_gc_threads) {
1667           worker_id = 0;
1668         }
1669       }
1670     }
1671     region_logger.print_line();
1672   }
1673 }
1674 
1675 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1676   assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1677 
1678   ParCompactionManager* cm =
1679     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1680 
1681   // Drain the stacks that have been preloaded with regions
1682   // that are ready to fill.
1683 
1684   cm->drain_region_stacks();
1685 
1686   guarantee(cm->region_stack()->is_empty(), "Not empty");
1687 
1688   size_t region_index = 0;
1689 
1690   while (true) {
1691     if (ParCompactionManager::steal(worker_id, region_index)) {
1692       PSParallelCompact::fill_and_update_region(cm, region_index);
1693       cm->drain_region_stacks();
1694     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1695       // Fill and update an unavailable region with the help of a shadow region
1696       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1697       cm->drain_region_stacks();
1698     } else {
1699       if (terminator->offer_termination()) {
1700         break;
1701       }
1702       // Go around again.
1703     }
1704   }
1705 }
1706 
1707 class FillDensePrefixAndCompactionTask: public WorkerTask {
1708   TaskTerminator _terminator;
1709 
1710 public:
1711   FillDensePrefixAndCompactionTask(uint active_workers) :
1712       WorkerTask("FillDensePrefixAndCompactionTask"),
1713       _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1714   }
1715 
1716   virtual void work(uint worker_id) {
1717     if (worker_id == 0) {
1718       auto start = Ticks::now();
1719       PSParallelCompact::fill_dead_objs_in_dense_prefix();
1720       log_trace(gc, phases)("Fill dense prefix by worker 0: %.3f ms", (Ticks::now() - start).seconds() * 1000);
1721     }
1722     compaction_with_stealing_work(&_terminator, worker_id);
1723   }
1724 };
1725 
1726 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1727 #ifdef ASSERT
1728   {
1729     assert(start < end, "precondition");
1730     assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1731     HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1732     if (start != bottom) {
1733       // The preceding live obj.
1734       HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1735       HeapWord* obj_end = obj_start + cast_to_oop(obj_start)->size();
1736       assert(obj_end == start, "precondition");
1737     }
1738   }
1739 #endif
1740 
1741   CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1742   HeapWord* addr = start;
1743   do {
1744     size_t size = cast_to_oop(addr)->size();
1745     start_array(old_space_id)->update_for_block(addr, addr + size);
1746     addr += size;
1747   } while (addr < end);
1748 }
1749 
1750 void PSParallelCompact::fill_dead_objs_in_dense_prefix() {
1751   ParMarkBitMap* bitmap = mark_bitmap();
1752 
1753   HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1754   HeapWord* const prefix_end = dense_prefix(old_space_id);
1755 
1756   const size_t region_size = ParallelCompactData::RegionSize;
1757 
1758   // Fill dead space in [start_addr, end_addr)
1759   HeapWord* const start_addr = bottom;
1760   HeapWord* const end_addr   = prefix_end;
1761 
1762   for (HeapWord* cur_addr = start_addr; cur_addr < end_addr; /* empty */) {
1763     RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1764     if (cur_region_ptr->data_size() == region_size) {
1765       // Full; no dead space. Next region.
1766       if (_summary_data.is_region_aligned(cur_addr)) {
1767         cur_addr += region_size;
1768       } else {
1769         cur_addr = _summary_data.region_align_up(cur_addr);
1770       }
1771       continue;
1772     }
1773 
1774     // Fill dead space inside cur_region.
1775     if (_summary_data.is_region_aligned(cur_addr)) {
1776       cur_addr += cur_region_ptr->partial_obj_size();
1777     }
1778 
1779     HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1780     assert(region_end_addr <= end_addr, "inv");
1781     while (cur_addr < region_end_addr) {
1782       // Use end_addr to allow filler-obj to cross region boundary.
1783       HeapWord* live_start = bitmap->find_obj_beg(cur_addr, end_addr);
1784       if (cur_addr != live_start) {
1785         // Found dead space [cur_addr, live_start).
1786         fill_range_in_dense_prefix(cur_addr, live_start);
1787       }
1788       if (live_start >= region_end_addr) {
1789         cur_addr = live_start;
1790         break;
1791       }
1792       assert(bitmap->is_marked(live_start), "inv");
1793       cur_addr = live_start + cast_to_oop(live_start)->size();
1794     }
1795   }
1796 }
1797 
1798 void PSParallelCompact::compact() {
1799   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1800 
1801   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1802 
1803   initialize_shadow_regions(active_gc_threads);
1804   prepare_region_draining_tasks(active_gc_threads);
1805 
1806   {
1807     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1808 
1809     FillDensePrefixAndCompactionTask task(active_gc_threads);
1810     ParallelScavengeHeap::heap()->workers().run_task(&task);
1811 
1812 #ifdef  ASSERT
1813     verify_filler_in_dense_prefix();
1814 
1815     // Verify that all regions have been processed.
1816     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1817       verify_complete(SpaceId(id));
1818     }
1819 #endif
1820   }
1821 }
1822 
1823 #ifdef  ASSERT
1824 void PSParallelCompact::verify_filler_in_dense_prefix() {
1825   HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1826   HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1827 
1828   const size_t region_size = ParallelCompactData::RegionSize;
1829 
1830   for (HeapWord* cur_addr = bottom; cur_addr < dense_prefix_end; /* empty */) {
1831     RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1832     if (cur_region_ptr->data_size() == region_size) {
1833       // Full; no dead space. Next region.
1834       if (_summary_data.is_region_aligned(cur_addr)) {
1835         cur_addr += region_size;
1836       } else {
1837         cur_addr = _summary_data.region_align_up(cur_addr);
1838       }
1839       continue;
1840     }
1841 
1842     // This region contains filler objs.
1843     if (_summary_data.is_region_aligned(cur_addr)) {
1844       cur_addr += cur_region_ptr->partial_obj_size();
1845     }
1846 
1847     HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1848     assert(region_end_addr <= dense_prefix_end, "inv");
1849 
1850     while (cur_addr < region_end_addr) {
1851       oop obj = cast_to_oop(cur_addr);
1852       oopDesc::verify(obj);
1853       if (!mark_bitmap()->is_marked(cur_addr)) {
1854         assert(CollectedHeap::is_filler_object(cast_to_oop(cur_addr)), "inv");
1855       }
1856       cur_addr += obj->size();
1857     }
1858   }
1859 }
1860 
1861 void PSParallelCompact::verify_complete(SpaceId space_id) {
1862   // All Regions served as compaction targets, from dense_prefix() to
1863   // new_top(), should be marked as filled and all Regions between new_top()
1864   // and top() should be available (i.e., should have been emptied).
1865   ParallelCompactData& sd = summary_data();
1866   SpaceInfo si = _space_info[space_id];
1867   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1868   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1869   const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1870   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1871   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1872 
1873   size_t cur_region;
1874   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1875     const RegionData* const c = sd.region(cur_region);
1876     assert(c->completed(), "region %zu not filled: destination_count=%u",
1877            cur_region, c->destination_count());
1878   }
1879 
1880   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1881     const RegionData* const c = sd.region(cur_region);
1882     assert(c->available(), "region %zu not empty: destination_count=%u",
1883            cur_region, c->destination_count());
1884   }
1885 }
1886 #endif  // #ifdef ASSERT
1887 
1888 // Return the SpaceId for the space containing addr.  If addr is not in the
1889 // heap, last_space_id is returned.  In debug mode it expects the address to be
1890 // in the heap and asserts such.
1891 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1892   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1893 
1894   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1895     if (_space_info[id].space()->contains(addr)) {
1896       return SpaceId(id);
1897     }
1898   }
1899 
1900   assert(false, "no space contains the addr");
1901   return last_space_id;
1902 }
1903 
1904 // Skip over count live words starting from beg, and return the address of the
1905 // next live word. Callers must also ensure that there are enough live words in
1906 // the range [beg, end) to skip.
1907 HeapWord* PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1908 {
1909   ParMarkBitMap* m = mark_bitmap();
1910   HeapWord* cur_addr = beg;
1911   while (true) {
1912     cur_addr = m->find_obj_beg(cur_addr, end);
1913     assert(cur_addr < end, "inv");
1914     size_t obj_size = cast_to_oop(cur_addr)->size();
1915     // Strictly greater-than
1916     if (obj_size > count) {
1917       return cur_addr + count;
1918     }
1919     count -= obj_size;
1920     cur_addr += obj_size;
1921   }
1922 }
1923 
1924 // On starting to fill a destination region (dest-region), we need to know the
1925 // location of the word that will be at the start of the dest-region after
1926 // compaction. A dest-region can have one or more source regions, but only the
1927 // first source-region contains this location. This location is retrieved by
1928 // calling `first_src_addr` on a dest-region.
1929 // Conversely, a source-region has a dest-region which holds the destination of
1930 // the first live word on this source-region, based on which the destination
1931 // for the rest of live words can be derived.
1932 //
1933 // Note:
1934 // There is some complication due to space-boundary-fragmentation (an obj can't
1935 // cross space-boundary) -- a source-region may be split and behave like two
1936 // distinct regions with their own dest-region, as depicted below.
1937 //
1938 // source-region: region-n
1939 //
1940 // **********************
1941 // |     A|A~~~~B|B     |
1942 // **********************
1943 //    n-1     n     n+1
1944 //
1945 // AA, BB denote two live objs. ~~~~ denotes unknown number of live objs.
1946 //
1947 // Assuming the dest-region for region-n is the final region before
1948 // old-space-end and its first-live-word is the middle of AA, the heap content
1949 // will look like the following after compaction:
1950 //
1951 // **************                  *************
1952 //      A|A~~~~ |                  |BB    |
1953 // **************                  *************
1954 //              ^                  ^
1955 //              | old-space-end    | eden-space-start
1956 //
1957 // Therefore, in this example, region-n will have two dest-regions:
1958 // 1. the final region in old-space
1959 // 2. the first region in eden-space.
1960 // To handle this special case, we introduce the concept of split-region, whose
1961 // contents are relocated to two spaces. `SplitInfo` captures all necessary
1962 // info about the split, the first part, spliting-point, and the second part.
1963 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1964                                             SpaceId src_space_id,
1965                                             size_t src_region_idx)
1966 {
1967   const size_t RegionSize = ParallelCompactData::RegionSize;
1968   const ParallelCompactData& sd = summary_data();
1969   assert(sd.is_region_aligned(dest_addr), "precondition");
1970 
1971   const RegionData* const src_region_ptr = sd.region(src_region_idx);
1972   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1973 
1974   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1975   HeapWord* const src_region_destination = src_region_ptr->destination();
1976 
1977   HeapWord* const region_start = sd.region_to_addr(src_region_idx);
1978   HeapWord* const region_end = sd.region_to_addr(src_region_idx) + RegionSize;
1979 
1980   // Identify the actual destination for the first live words on this region,
1981   // taking split-region into account.
1982   HeapWord* region_start_destination;
1983   const SplitInfo& split_info = _space_info[src_space_id].split_info();
1984   if (split_info.is_split(src_region_idx)) {
1985     // The second part of this split region; use the recorded split point.
1986     if (dest_addr == src_region_destination) {
1987       return split_info.split_point();
1988     }
1989     region_start_destination = split_info.preceding_destination();
1990   } else {
1991     region_start_destination = src_region_destination;
1992   }
1993 
1994   // Calculate the offset to be skipped
1995   size_t words_to_skip = pointer_delta(dest_addr, region_start_destination);
1996 
1997   HeapWord* result;
1998   if (partial_obj_size > words_to_skip) {
1999     result = region_start + words_to_skip;
2000   } else {
2001     words_to_skip -= partial_obj_size;
2002     result = skip_live_words(region_start + partial_obj_size, region_end, words_to_skip);
2003   }
2004 
2005   if (split_info.is_split(src_region_idx)) {
2006     assert(result < split_info.split_point(), "postcondition");
2007   } else {
2008     assert(result < region_end, "postcondition");
2009   }
2010 
2011   return result;
2012 }
2013 
2014 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2015                                                      SpaceId src_space_id,
2016                                                      size_t beg_region,
2017                                                      HeapWord* end_addr)
2018 {
2019   ParallelCompactData& sd = summary_data();
2020 
2021 #ifdef ASSERT
2022   MutableSpace* const src_space = _space_info[src_space_id].space();
2023   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2024   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2025          "src_space_id does not match beg_addr");
2026   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2027          "src_space_id does not match end_addr");
2028 #endif // #ifdef ASSERT
2029 
2030   RegionData* const beg = sd.region(beg_region);
2031   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2032 
2033   // Regions up to new_top() are enqueued if they become available.
2034   HeapWord* const new_top = _space_info[src_space_id].new_top();
2035   RegionData* const enqueue_end =
2036     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2037 
2038   for (RegionData* cur = beg; cur < end; ++cur) {
2039     assert(cur->data_size() > 0, "region must have live data");
2040     cur->decrement_destination_count();
2041     if (cur < enqueue_end && cur->available() && cur->claim()) {
2042       if (cur->mark_normal()) {
2043         cm->push_region(sd.region(cur));
2044       } else if (cur->mark_copied()) {
2045         // Try to copy the content of the shadow region back to its corresponding
2046         // heap region if the shadow region is filled. Otherwise, the GC thread
2047         // fills the shadow region will copy the data back (see
2048         // MoveAndUpdateShadowClosure::complete_region).
2049         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2050         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2051         cur->set_completed();
2052       }
2053     }
2054   }
2055 }
2056 
2057 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2058                                           SpaceId& src_space_id,
2059                                           HeapWord*& src_space_top,
2060                                           HeapWord* end_addr)
2061 {
2062   ParallelCompactData& sd = PSParallelCompact::summary_data();
2063 
2064   size_t src_region_idx = 0;
2065 
2066   // Skip empty regions (if any) up to the top of the space.
2067   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2068   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2069   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2070   const RegionData* const top_region_ptr = sd.addr_to_region_ptr(top_aligned_up);
2071 
2072   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2073     ++src_region_ptr;
2074   }
2075 
2076   if (src_region_ptr < top_region_ptr) {
2077     // Found the first non-empty region in the same space.
2078     src_region_idx = sd.region(src_region_ptr);
2079     closure.set_source(sd.region_to_addr(src_region_idx));
2080     return src_region_idx;
2081   }
2082 
2083   // Switch to a new source space and find the first non-empty region.
2084   uint space_id = src_space_id + 1;
2085   assert(space_id < last_space_id, "not enough spaces");
2086 
2087   for (/* empty */; space_id < last_space_id; ++space_id) {
2088     HeapWord* bottom = _space_info[space_id].space()->bottom();
2089     HeapWord* top = _space_info[space_id].space()->top();
2090     // Skip empty space
2091     if (bottom == top) {
2092       continue;
2093     }
2094 
2095     // Identify the first region that contains live words in this space
2096     size_t cur_region = sd.addr_to_region_idx(bottom);
2097     size_t end_region = sd.addr_to_region_idx(sd.region_align_up(top));
2098 
2099     for (/* empty */ ; cur_region < end_region; ++cur_region) {
2100       RegionData* cur = sd.region(cur_region);
2101       if (cur->live_obj_size() > 0) {
2102         HeapWord* region_start_addr = sd.region_to_addr(cur_region);
2103 
2104         src_space_id = SpaceId(space_id);
2105         src_space_top = top;
2106         closure.set_source(region_start_addr);
2107         return cur_region;
2108       }
2109     }
2110   }
2111 
2112   ShouldNotReachHere();
2113 }
2114 
2115 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2116   ParallelCompactData& sd = summary_data();
2117   assert(sd.is_region_aligned(region_start_addr), "precondition");
2118 
2119   // Use per-region partial_obj_size to locate the end of the obj, that extends
2120   // to region_start_addr.
2121   size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2122   size_t end_region_idx = sd.region_count();
2123   size_t accumulated_size = 0;
2124   for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2125     size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2126     accumulated_size += cur_partial_obj_size;
2127     if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2128       break;
2129     }
2130   }
2131   return region_start_addr + accumulated_size;
2132 }
2133 
2134 // Use region_idx as the destination region, and evacuate all live objs on its
2135 // source regions to this destination region.
2136 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2137 {
2138   ParMarkBitMap* const bitmap = mark_bitmap();
2139   ParallelCompactData& sd = summary_data();
2140   RegionData* const region_ptr = sd.region(region_idx);
2141 
2142   // Get the source region and related info.
2143   size_t src_region_idx = region_ptr->source_region();
2144   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2145   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2146   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2147 
2148   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2149 
2150   // Adjust src_region_idx to prepare for decrementing destination counts (the
2151   // destination count is not decremented when a region is copied to itself).
2152   if (src_region_idx == region_idx) {
2153     src_region_idx += 1;
2154   }
2155 
2156   // source-region:
2157   //
2158   // **********
2159   // |   ~~~  |
2160   // **********
2161   //      ^
2162   //      |-- closure.source() / first_src_addr
2163   //
2164   //
2165   // ~~~ : live words
2166   //
2167   // destination-region:
2168   //
2169   // **********
2170   // |        |
2171   // **********
2172   // ^
2173   // |-- region-start
2174   if (bitmap->is_unmarked(closure.source())) {
2175     // An object overflows the previous destination region, so this
2176     // destination region should copy the remainder of the object or as much as
2177     // will fit.
2178     HeapWord* const old_src_addr = closure.source();
2179     {
2180       HeapWord* region_start = sd.region_align_down(closure.source());
2181       HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2182       HeapWord* obj_end;
2183       if (obj_start != closure.source()) {
2184         assert(bitmap->is_marked(obj_start), "inv");
2185         // Found the actual obj-start, try to find the obj-end using either
2186         // size() if this obj is completely contained in the current region.
2187         HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2188         HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2189                                       ? nullptr
2190                                       : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2191         // This obj extends to next region iff partial_obj_addr of the *next*
2192         // region is the same as obj-start.
2193         if (partial_obj_start == obj_start) {
2194           // This obj extends to next region.
2195           obj_end = partial_obj_end(next_region_start);
2196         } else {
2197           // Completely contained in this region; safe to use size().
2198           obj_end = obj_start + cast_to_oop(obj_start)->size();
2199         }
2200       } else {
2201         // This obj extends to current region.
2202         obj_end = partial_obj_end(region_start);
2203       }
2204       size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2205       closure.copy_partial_obj(partial_obj_size);
2206     }
2207 
2208     if (closure.is_full()) {
2209       decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2210       closure.complete_region(dest_addr, region_ptr);
2211       return;
2212     }
2213 
2214     // Finished copying without using up the current destination-region
2215     HeapWord* const end_addr = sd.region_align_down(closure.source());
2216     if (sd.region_align_down(old_src_addr) != end_addr) {
2217       assert(sd.region_align_up(old_src_addr) == end_addr, "only one region");
2218       // The partial object was copied from more than one source region.
2219       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2220 
2221       // Move to the next source region, possibly switching spaces as well.  All
2222       // args except end_addr may be modified.
2223       src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2224     }
2225   }
2226 
2227   // Handle the rest obj-by-obj, where we know obj-start.
2228   do {
2229     HeapWord* cur_addr = closure.source();
2230     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2231                                     src_space_top);
2232     // To handle the case where the final obj in source region extends to next region.
2233     HeapWord* final_obj_start = (end_addr == src_space_top)
2234                                 ? nullptr
2235                                 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2236     // Apply closure on objs inside [cur_addr, end_addr)
2237     do {
2238       cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2239       if (cur_addr == end_addr) {
2240         break;
2241       }
2242       size_t obj_size;
2243       if (final_obj_start == cur_addr) {
2244         obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2245       } else {
2246         // This obj doesn't extend into next region; size() is safe to use.
2247         obj_size = cast_to_oop(cur_addr)->size();
2248       }
2249       closure.do_addr(cur_addr, obj_size);
2250       cur_addr += obj_size;
2251     } while (cur_addr < end_addr && !closure.is_full());
2252 
2253     if (closure.is_full()) {
2254       decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2255       closure.complete_region(dest_addr, region_ptr);
2256       return;
2257     }
2258 
2259     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2260 
2261     // Move to the next source region, possibly switching spaces as well.  All
2262     // args except end_addr may be modified.
2263     src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2264   } while (true);
2265 }
2266 
2267 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2268 {
2269   MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2270   fill_region(cm, cl, region_idx);
2271 }
2272 
2273 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2274 {
2275   // Get a shadow region first
2276   ParallelCompactData& sd = summary_data();
2277   RegionData* const region_ptr = sd.region(region_idx);
2278   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2279   // The InvalidShadow return value indicates the corresponding heap region is available,
2280   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2281   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2282   if (shadow_region == ParCompactionManager::InvalidShadow) {
2283     MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2284     region_ptr->shadow_to_normal();
2285     return fill_region(cm, cl, region_idx);
2286   } else {
2287     MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2288     return fill_region(cm, cl, region_idx);
2289   }
2290 }
2291 
2292 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2293 {
2294   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2295 }
2296 
2297 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
2298 {
2299   size_t next = cm->next_shadow_region();
2300   ParallelCompactData& sd = summary_data();
2301   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2302   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2303 
2304   while (next < old_new_top) {
2305     if (sd.region(next)->mark_shadow()) {
2306       region_idx = next;
2307       return true;
2308     }
2309     next = cm->move_next_shadow_region_by(active_gc_threads);
2310   }
2311 
2312   return false;
2313 }
2314 
2315 // The shadow region is an optimization to address region dependencies in full GC. The basic
2316 // idea is making more regions available by temporally storing their live objects in empty
2317 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2318 // GC threads need not wait destination regions to be available before processing sources.
2319 //
2320 // A typical workflow would be:
2321 // After draining its own stack and failing to steal from others, a GC worker would pick an
2322 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2323 // the shadow region by copying live objects from source regions of the unavailable one. Once
2324 // the unavailable region becomes available, the data in the shadow region will be copied back.
2325 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2326 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2327 {
2328   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2329 
2330   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2331     SpaceInfo* const space_info = _space_info + id;
2332     MutableSpace* const space = space_info->space();
2333 
2334     const size_t beg_region =
2335       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2336     const size_t end_region =
2337       sd.addr_to_region_idx(sd.region_align_down(space->end()));
2338 
2339     for (size_t cur = beg_region; cur < end_region; ++cur) {
2340       ParCompactionManager::push_shadow_region(cur);
2341     }
2342   }
2343 
2344   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2345   for (uint i = 0; i < parallel_gc_threads; i++) {
2346     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2347     cm->set_next_shadow_region(beg_region + i);
2348   }
2349 }
2350 
2351 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2352 {
2353   size_t words = MIN2(partial_obj_size, words_remaining());
2354 
2355   // This test is necessary; if omitted, the pointer updates to a partial object
2356   // that crosses the dense prefix boundary could be overwritten.
2357   if (source() != copy_destination()) {
2358     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2359     Copy::aligned_conjoint_words(source(), copy_destination(), words);
2360   }
2361   update_state(words);
2362 }
2363 
2364 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2365   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2366   region_ptr->set_completed();
2367 }
2368 
2369 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2370   assert(destination() != nullptr, "sanity");
2371   _source = addr;
2372 
2373   // The start_array must be updated even if the object is not moving.
2374   if (_start_array != nullptr) {
2375     _start_array->update_for_block(destination(), destination() + words);
2376   }
2377 
2378   // Avoid overflow
2379   words = MIN2(words, words_remaining());
2380   assert(words > 0, "inv");
2381 
2382   if (copy_destination() != source()) {
2383     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2384     assert(source() != destination(), "inv");
2385     assert(FullGCForwarding::is_forwarded(cast_to_oop(source())), "inv");
2386     assert(FullGCForwarding::forwardee(cast_to_oop(source())) == cast_to_oop(destination()), "inv");
2387     Copy::aligned_conjoint_words(source(), copy_destination(), words);
2388     cast_to_oop(copy_destination())->init_mark();
2389   }
2390 
2391   update_state(words);
2392 }
2393 
2394 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2395   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2396   // Record the shadow region index
2397   region_ptr->set_shadow_region(_shadow);
2398   // Mark the shadow region as filled to indicate the data is ready to be
2399   // copied back
2400   region_ptr->mark_filled();
2401   // Try to copy the content of the shadow region back to its corresponding
2402   // heap region if available; the GC thread that decreases the destination
2403   // count to zero will do the copying otherwise (see
2404   // PSParallelCompact::decrement_destination_counts).
2405   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2406     region_ptr->set_completed();
2407     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2408     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2409   }
2410 }