1 /*
   2  * Copyright (c) 2014, 2020, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc/g1/g1Allocator.inline.hpp"
  27 #include "gc/g1/g1CollectedHeap.inline.hpp"
  28 #include "gc/g1/g1CollectionSet.hpp"
  29 #include "gc/g1/g1OopClosures.inline.hpp"
  30 #include "gc/g1/g1ParScanThreadState.inline.hpp"
  31 #include "gc/g1/g1RootClosures.hpp"
  32 #include "gc/g1/g1StringDedup.hpp"
  33 #include "gc/g1/g1Trace.hpp"
  34 #include "gc/shared/taskqueue.inline.hpp"
  35 #include "memory/allocation.inline.hpp"
  36 #include "oops/access.inline.hpp"
  37 #include "oops/oop.inline.hpp"
  38 #include "runtime/prefetch.inline.hpp"
  39 
  40 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
  41                                            G1RedirtyCardsQueueSet* rdcqs,
  42                                            uint worker_id,
  43                                            size_t young_cset_length,
  44                                            size_t optional_cset_length)
  45   : _g1h(g1h),
  46     _task_queue(g1h->task_queue(worker_id)),
  47     _rdcq(rdcqs),
  48     _ct(g1h->card_table()),
  49     _closures(NULL),
  50     _plab_allocator(NULL),
  51     _age_table(false),
  52     _tenuring_threshold(g1h->policy()->tenuring_threshold()),
  53     _scanner(g1h, this),
  54     _worker_id(worker_id),
  55     _last_enqueued_card(SIZE_MAX),
  56     _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
  57     _stack_trim_lower_threshold(GCDrainStackTargetSize),
  58     _trim_ticks(),
  59     _surviving_young_words_base(NULL),
  60     _surviving_young_words(NULL),
  61     _surviving_words_length(young_cset_length + 1),
  62     _old_gen_is_full(false),
  63     _num_optional_regions(optional_cset_length),
  64     _numa(g1h->numa()),
  65     _obj_alloc_stat(NULL)
  66 {
  67   // We allocate number of young gen regions in the collection set plus one
  68   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
  69   // We also add a few elements at the beginning and at the end in
  70   // an attempt to eliminate cache contention
  71   const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
  72   size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
  73 
  74   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
  75   _surviving_young_words = _surviving_young_words_base + padding_elem_num;
  76   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
  77 
  78   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
  79 
  80   // The dest for Young is used when the objects are aged enough to
  81   // need to be moved to the next space.
  82   _dest[G1HeapRegionAttr::Young] = G1HeapRegionAttr::Old;
  83   _dest[G1HeapRegionAttr::Old]   = G1HeapRegionAttr::Old;
  84 
  85   _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
  86 
  87   _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
  88 
  89   initialize_numa_stats();
  90 }
  91 
  92 size_t G1ParScanThreadState::flush(size_t* surviving_young_words) {
  93   _rdcq.flush();
  94   flush_numa_stats();
  95   // Update allocation statistics.
  96   _plab_allocator->flush_and_retire_stats();
  97   _g1h->policy()->record_age_table(&_age_table);
  98 
  99   size_t sum = 0;
 100   for (uint i = 0; i < _surviving_words_length; i++) {
 101     surviving_young_words[i] += _surviving_young_words[i];
 102     sum += _surviving_young_words[i];
 103   }
 104   return sum;
 105 }
 106 
 107 G1ParScanThreadState::~G1ParScanThreadState() {
 108   delete _plab_allocator;
 109   delete _closures;
 110   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
 111   delete[] _oops_into_optional_regions;
 112   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
 113 }
 114 
 115 size_t G1ParScanThreadState::lab_waste_words() const {
 116   return _plab_allocator->waste();
 117 }
 118 
 119 size_t G1ParScanThreadState::lab_undo_waste_words() const {
 120   return _plab_allocator->undo_waste();
 121 }
 122 
 123 #ifdef ASSERT
 124 void G1ParScanThreadState::verify_task(narrowOop* task) const {
 125   assert(task != NULL, "invariant");
 126   assert(UseCompressedOops, "sanity");
 127   oop p = RawAccess<>::oop_load(task);
 128   assert(_g1h->is_in_g1_reserved(p),
 129          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
 130 }
 131 
 132 void G1ParScanThreadState::verify_task(oop* task) const {
 133   assert(task != NULL, "invariant");
 134   oop p = RawAccess<>::oop_load(task);
 135   assert(_g1h->is_in_g1_reserved(p),
 136          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
 137 }
 138 
 139 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
 140   // Must be in the collection set--it's already been copied.
 141   oop p = task.to_source_array();
 142   assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
 143 }
 144 
 145 void G1ParScanThreadState::verify_task(ScannerTask task) const {
 146   if (task.is_narrow_oop_ptr()) {
 147     verify_task(task.to_narrow_oop_ptr());
 148   } else if (task.is_oop_ptr()) {
 149     verify_task(task.to_oop_ptr());
 150   } else if (task.is_partial_array_task()) {
 151     verify_task(task.to_partial_array_task());
 152   } else {
 153     ShouldNotReachHere();
 154   }
 155 }
 156 #endif // ASSERT
 157 
 158 template <class T> void G1ParScanThreadState::do_oop_evac(T* p) {
 159   // Reference should not be NULL here as such are never pushed to the task queue.
 160   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
 161 
 162   // Although we never intentionally push references outside of the collection
 163   // set, due to (benign) races in the claim mechanism during RSet scanning more
 164   // than one thread might claim the same card. So the same card may be
 165   // processed multiple times, and so we might get references into old gen here.
 166   // So we need to redo this check.
 167   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
 168   // References pushed onto the work stack should never point to a humongous region
 169   // as they are not added to the collection set due to above precondition.
 170   assert(!region_attr.is_humongous(),
 171          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
 172          p2i(obj), _g1h->addr_to_region(cast_from_oop<HeapWord*>(obj)), p2i(p));
 173 
 174   if (!region_attr.is_in_cset()) {
 175     // In this case somebody else already did all the work.
 176     return;
 177   }
 178 
 179   markWord m = obj->mark_raw();
 180   if (m.is_marked()) {
 181     obj = (oop) m.decode_pointer();
 182   } else {
 183     obj = do_copy_to_survivor_space(region_attr, obj, m);
 184   }
 185   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
 186 
 187   assert(obj != NULL, "Must be");
 188   if (HeapRegion::is_in_same_region(p, obj)) {
 189     return;
 190   }
 191   HeapRegion* from = _g1h->heap_region_containing(p);
 192   if (!from->is_young()) {
 193     enqueue_card_if_tracked(_g1h->region_attr(obj), p, obj);
 194   }
 195 }
 196 
 197 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
 198   oop from_obj = task.to_source_array();
 199 
 200   assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
 201   assert(from_obj->is_objArray(), "must be obj array");
 202   objArrayOop from_obj_array = objArrayOop(from_obj);
 203   // The from-space object contains the real length.
 204   int length                 = from_obj_array->length();
 205 
 206   assert(from_obj->is_forwarded(), "must be forwarded");
 207   oop to_obj                 = from_obj->forwardee();
 208   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
 209   objArrayOop to_obj_array   = objArrayOop(to_obj);
 210   // We keep track of the next start index in the length field of the
 211   // to-space object.
 212   int next_index             = to_obj_array->length();
 213   assert(0 <= next_index && next_index < length,
 214          "invariant, next index: %d, length: %d", next_index, length);
 215 
 216   int start                  = next_index;
 217   int end                    = length;
 218   int remainder              = end - start;
 219   // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
 220   if (remainder > 2 * ParGCArrayScanChunk) {
 221     end = start + ParGCArrayScanChunk;
 222     to_obj_array->set_length(end);
 223     // Push the remainder before we process the range in case another
 224     // worker has run out of things to do and can steal it.
 225     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
 226   } else {
 227     assert(length == end, "sanity");
 228     // We'll process the final range for this object. Restore the length
 229     // so that the heap remains parsable in case of evacuation failure.
 230     to_obj_array->set_length(end);
 231   }
 232 
 233   HeapRegion* hr = _g1h->heap_region_containing(to_obj);
 234   G1ScanInYoungSetter x(&_scanner, hr->is_young());
 235   // Process indexes [start,end). It will also process the header
 236   // along with the first chunk (i.e., the chunk with start == 0).
 237   // Note that at this point the length field of to_obj_array is not
 238   // correct given that we are using it to keep track of the next
 239   // start index. oop_iterate_range() (thankfully!) ignores the length
 240   // field and only relies on the start / end parameters.  It does
 241   // however return the size of the object which will be incorrect. So
 242   // we have to ignore it even if we wanted to use it.
 243   to_obj_array->oop_iterate_range(&_scanner, start, end);
 244 }
 245 
 246 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
 247   verify_task(task);
 248   if (task.is_narrow_oop_ptr()) {
 249     do_oop_evac(task.to_narrow_oop_ptr());
 250   } else if (task.is_oop_ptr()) {
 251     do_oop_evac(task.to_oop_ptr());
 252   } else {
 253     do_partial_array(task.to_partial_array_task());
 254   }
 255 }
 256 
 257 // Process tasks until overflow queue is empty and local queue
 258 // contains no more than threshold entries.  NOINLINE to prevent
 259 // inlining into steal_and_trim_queue.
 260 ATTRIBUTE_FLATTEN NOINLINE
 261 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
 262   ScannerTask task;
 263   do {
 264     while (_task_queue->pop_overflow(task)) {
 265       if (!_task_queue->try_push_to_taskqueue(task)) {
 266         dispatch_task(task);
 267       }
 268     }
 269     while (_task_queue->pop_local(task, threshold)) {
 270       dispatch_task(task);
 271     }
 272   } while (!_task_queue->overflow_empty());
 273 }
 274 
 275 ATTRIBUTE_FLATTEN
 276 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
 277   ScannerTask stolen_task;
 278   while (task_queues->steal(_worker_id, stolen_task)) {
 279     dispatch_task(stolen_task);
 280     // Processing stolen task may have added tasks to our queue.
 281     trim_queue();
 282   }
 283 }
 284 
 285 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
 286                                                       size_t word_sz,
 287                                                       bool previous_plab_refill_failed,
 288                                                       uint node_index) {
 289 
 290   assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
 291 
 292   // Right now we only have two types of regions (young / old) so
 293   // let's keep the logic here simple. We can generalize it when necessary.
 294   if (dest->is_young()) {
 295     bool plab_refill_in_old_failed = false;
 296     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
 297                                                         word_sz,
 298                                                         &plab_refill_in_old_failed,
 299                                                         node_index);
 300     // Make sure that we won't attempt to copy any other objects out
 301     // of a survivor region (given that apparently we cannot allocate
 302     // any new ones) to avoid coming into this slow path again and again.
 303     // Only consider failed PLAB refill here: failed inline allocations are
 304     // typically large, so not indicative of remaining space.
 305     if (previous_plab_refill_failed) {
 306       _tenuring_threshold = 0;
 307     }
 308 
 309     if (obj_ptr != NULL) {
 310       dest->set_old();
 311     } else {
 312       // We just failed to allocate in old gen. The same idea as explained above
 313       // for making survivor gen unavailable for allocation applies for old gen.
 314       _old_gen_is_full = plab_refill_in_old_failed;
 315     }
 316     return obj_ptr;
 317   } else {
 318     _old_gen_is_full = previous_plab_refill_failed;
 319     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
 320     // no other space to try.
 321     return NULL;
 322   }
 323 }
 324 
 325 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
 326   if (region_attr.is_young()) {
 327     age = !m.has_displaced_mark_helper() ? m.age()
 328                                          : m.displaced_mark_helper().age();
 329     if (age < _tenuring_threshold) {
 330       return region_attr;
 331     }
 332   }
 333   return dest(region_attr);
 334 }
 335 
 336 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
 337                                                   oop const old, size_t word_sz, uint age,
 338                                                   HeapWord * const obj_ptr, uint node_index) const {
 339   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
 340   if (alloc_buf->contains(obj_ptr)) {
 341     _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
 342                                                              dest_attr.type() == G1HeapRegionAttr::Old,
 343                                                              alloc_buf->word_sz() * HeapWordSize);
 344   } else {
 345     _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
 346                                                               dest_attr.type() == G1HeapRegionAttr::Old);
 347   }
 348 }
 349 
 350 NOINLINE
 351 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
 352                                                    oop old,
 353                                                    size_t word_sz,
 354                                                    uint age,
 355                                                    uint node_index) {
 356   HeapWord* obj_ptr = NULL;
 357   // Try slow-path allocation unless we're allocating old and old is already full.
 358   if (!(dest_attr->is_old() && _old_gen_is_full)) {
 359     bool plab_refill_failed = false;
 360     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
 361                                                            word_sz,
 362                                                            &plab_refill_failed,
 363                                                            node_index);
 364     if (obj_ptr == NULL) {
 365       obj_ptr = allocate_in_next_plab(dest_attr,
 366                                       word_sz,
 367                                       plab_refill_failed,
 368                                       node_index);
 369     }
 370   }
 371   if (obj_ptr != NULL) {
 372     update_numa_stats(node_index);
 373     if (_g1h->_gc_tracer_stw->should_report_promotion_events()) {
 374       // The events are checked individually as part of the actual commit
 375       report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
 376     }
 377   }
 378   return obj_ptr;
 379 }
 380 
 381 NOINLINE
 382 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
 383                                            HeapWord* obj_ptr,
 384                                            size_t word_sz,
 385                                            uint node_index) {
 386   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 387 }
 388 
 389 // Private inline function, for direct internal use and providing the
 390 // implementation of the public not-inline function.
 391 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
 392                                                     oop const old,
 393                                                     markWord const old_mark) {
 394    assert(region_attr.is_in_cset(),
 395           "Unexpected region attr type: %s", region_attr.get_type_str());
 396 
 397   const size_t word_sz = old->size();
 398 
 399   uint age = 0;
 400   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
 401   HeapRegion* const from_region = _g1h->heap_region_containing(old);
 402   uint node_index = from_region->node_index();
 403 
 404   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
 405 
 406   // PLAB allocations should succeed most of the time, so we'll
 407   // normally check against NULL once and that's it.
 408   if (obj_ptr == NULL) {
 409     obj_ptr = allocate_copy_slow(&dest_attr, old, word_sz, age, node_index);
 410     if (obj_ptr == NULL) {
 411       // This will either forward-to-self, or detect that someone else has
 412       // installed a forwarding pointer.
 413       return handle_evacuation_failure_par(old, old_mark);
 414     }
 415   }
 416 
 417   assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
 418   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
 419 
 420 #ifndef PRODUCT
 421   // Should this evacuation fail?
 422   if (_g1h->evacuation_should_fail()) {
 423     // Doing this after all the allocation attempts also tests the
 424     // undo_allocation() method too.
 425     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 426     return handle_evacuation_failure_par(old, old_mark);
 427   }
 428 #endif // !PRODUCT
 429 
 430   // We're going to allocate linearly, so might as well prefetch ahead.
 431   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
 432 
 433   const oop obj = oop(obj_ptr);
 434   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
 435   if (forward_ptr == NULL) {
 436     Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
 437 
 438     {
 439       const uint young_index = from_region->young_index_in_cset();
 440       assert((from_region->is_young() && young_index >  0) ||
 441              (!from_region->is_young() && young_index == 0), "invariant" );
 442       _surviving_young_words[young_index] += word_sz;
 443     }
 444 
 445     if (dest_attr.is_young()) {
 446       if (age < markWord::max_age) {
 447         age++;
 448       }
 449       if (old_mark.has_displaced_mark_helper()) {
 450         // In this case, we have to install the mark word first,
 451         // otherwise obj looks to be forwarded (the old mark word,
 452         // which contains the forward pointer, was copied)
 453         obj->set_mark_raw(old_mark);
 454         markWord new_mark = old_mark.displaced_mark_helper().set_age(age);
 455         old_mark.set_displaced_mark_helper(new_mark);
 456       } else {
 457         obj->set_mark_raw(old_mark.set_age(age));
 458       }
 459       _age_table.add(age, word_sz);
 460     } else {
 461       obj->set_mark_raw(old_mark);
 462     }
 463 
 464     if (G1StringDedup::is_enabled()) {
 465       const bool is_from_young = region_attr.is_young();
 466       const bool is_to_young = dest_attr.is_young();
 467       assert(is_from_young == from_region->is_young(),
 468              "sanity");
 469       assert(is_to_young == _g1h->heap_region_containing(obj)->is_young(),
 470              "sanity");
 471       G1StringDedup::enqueue_from_evacuation(is_from_young,
 472                                              is_to_young,
 473                                              _worker_id,
 474                                              obj);
 475     }
 476 
 477     if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
 478       // We keep track of the next start index in the length field of
 479       // the to-space object. The actual length can be found in the
 480       // length field of the from-space object.
 481       arrayOop(obj)->set_length(0);
 482       do_partial_array(PartialArrayScanTask(old));
 483     } else {
 484       G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
 485       obj->oop_iterate_backwards(&_scanner);
 486     }
 487     return obj;
 488   } else {
 489     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 490     return forward_ptr;
 491   }
 492 }
 493 
 494 // Public not-inline entry point.
 495 ATTRIBUTE_FLATTEN
 496 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
 497                                                  oop old,
 498                                                  markWord old_mark) {
 499   return do_copy_to_survivor_space(region_attr, old, old_mark);
 500 }
 501 
 502 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
 503   assert(worker_id < _n_workers, "out of bounds access");
 504   if (_states[worker_id] == NULL) {
 505     _states[worker_id] =
 506       new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length);
 507   }
 508   return _states[worker_id];
 509 }
 510 
 511 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
 512   assert(_flushed, "thread local state from the per thread states should have been flushed");
 513   return _surviving_young_words_total;
 514 }
 515 
 516 void G1ParScanThreadStateSet::flush() {
 517   assert(!_flushed, "thread local state from the per thread states should be flushed once");
 518 
 519   for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
 520     G1ParScanThreadState* pss = _states[worker_id];
 521 
 522     if (pss == NULL) {
 523       continue;
 524     }
 525 
 526     G1GCPhaseTimes* p = _g1h->phase_times();
 527 
 528     // Need to get the following two before the call to G1ParThreadScanState::flush()
 529     // because it resets the PLAB allocator where we get this info from.
 530     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
 531     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
 532     size_t copied_bytes = pss->flush(_surviving_young_words_total) * HeapWordSize;
 533 
 534     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
 535     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
 536     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
 537 
 538     delete pss;
 539     _states[worker_id] = NULL;
 540   }
 541   _flushed = true;
 542 }
 543 
 544 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
 545   for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
 546     G1ParScanThreadState* pss = _states[worker_index];
 547 
 548     if (pss == NULL) {
 549       continue;
 550     }
 551 
 552     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
 553     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
 554   }
 555 }
 556 
 557 NOINLINE
 558 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m) {
 559   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
 560 
 561   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
 562   if (forward_ptr == NULL) {
 563     // Forward-to-self succeeded. We are the "owner" of the object.
 564     HeapRegion* r = _g1h->heap_region_containing(old);
 565 
 566     if (!r->evacuation_failed()) {
 567       r->set_evacuation_failed(true);
 568      _g1h->hr_printer()->evac_failure(r);
 569     }
 570 
 571     _g1h->preserve_mark_during_evac_failure(_worker_id, old, m);
 572 
 573     G1ScanInYoungSetter x(&_scanner, r->is_young());
 574     old->oop_iterate_backwards(&_scanner);
 575 
 576     return old;
 577   } else {
 578     // Forward-to-self failed. Either someone else managed to allocate
 579     // space for this object (old != forward_ptr) or they beat us in
 580     // self-forwarding it (old == forward_ptr).
 581     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
 582            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
 583            "should not be in the CSet",
 584            p2i(old), p2i(forward_ptr));
 585     return forward_ptr;
 586   }
 587 }
 588 
 589 void G1ParScanThreadState::initialize_numa_stats() {
 590   if (_numa->is_enabled()) {
 591     LogTarget(Info, gc, heap, numa) lt;
 592 
 593     if (lt.is_enabled()) {
 594       uint num_nodes = _numa->num_active_nodes();
 595       // Record only if there are multiple active nodes.
 596       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
 597       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
 598     }
 599   }
 600 }
 601 
 602 void G1ParScanThreadState::flush_numa_stats() {
 603   if (_obj_alloc_stat != NULL) {
 604     uint node_index = _numa->index_of_current_thread();
 605     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
 606   }
 607 }
 608 
 609 void G1ParScanThreadState::update_numa_stats(uint node_index) {
 610   if (_obj_alloc_stat != NULL) {
 611     _obj_alloc_stat[node_index]++;
 612   }
 613 }
 614 
 615 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
 616                                                  G1RedirtyCardsQueueSet* rdcqs,
 617                                                  uint n_workers,
 618                                                  size_t young_cset_length,
 619                                                  size_t optional_cset_length) :
 620     _g1h(g1h),
 621     _rdcqs(rdcqs),
 622     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
 623     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
 624     _young_cset_length(young_cset_length),
 625     _optional_cset_length(optional_cset_length),
 626     _n_workers(n_workers),
 627     _flushed(false) {
 628   for (uint i = 0; i < n_workers; ++i) {
 629     _states[i] = NULL;
 630   }
 631   memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
 632 }
 633 
 634 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
 635   assert(_flushed, "thread local state from the per thread states should have been flushed");
 636   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
 637   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
 638 }