1 /*
  2  * Copyright (c) 2014, 2021, 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/g1EvacFailureRegions.inline.hpp"
 30 #include "gc/g1/g1OopClosures.inline.hpp"
 31 #include "gc/g1/g1ParScanThreadState.inline.hpp"
 32 #include "gc/g1/g1RootClosures.hpp"
 33 #include "gc/g1/g1StringDedup.hpp"
 34 #include "gc/g1/g1Trace.hpp"
 35 #include "gc/g1/g1YoungGCEvacFailureInjector.inline.hpp"
 36 #include "gc/shared/partialArrayTaskStepper.inline.hpp"
 37 #include "gc/shared/preservedMarks.inline.hpp"
 38 #include "gc/shared/stringdedup/stringDedup.hpp"
 39 #include "gc/shared/taskqueue.inline.hpp"
 40 #include "memory/allocation.inline.hpp"
 41 #include "oops/access.inline.hpp"
 42 #include "oops/oop.inline.hpp"
 43 #include "runtime/atomic.hpp"
 44 #include "runtime/prefetch.inline.hpp"
 45 #include "utilities/globalDefinitions.hpp"
 46 #include "utilities/macros.hpp"
 47 
 48 // In fastdebug builds the code size can get out of hand, potentially
 49 // tripping over compiler limits (which may be bugs, but nevertheless
 50 // need to be taken into consideration).  A side benefit of limiting
 51 // inlining is that we get more call frames that might aid debugging.
 52 // And the fastdebug compile time for this file is much reduced.
 53 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
 54 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
 55 
 56 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
 57                                            G1RedirtyCardsQueueSet* rdcqs,
 58                                            PreservedMarks* preserved_marks,
 59                                            uint worker_id,
 60                                            uint n_workers,
 61                                            size_t young_cset_length,
 62                                            size_t optional_cset_length,
 63                                            G1EvacFailureRegions* evac_failure_regions)
 64   : _g1h(g1h),
 65     _task_queue(g1h->task_queue(worker_id)),
 66     _rdc_local_qset(rdcqs),
 67     _ct(g1h->card_table()),
 68     _closures(NULL),
 69     _plab_allocator(NULL),
 70     _age_table(false),
 71     _tenuring_threshold(g1h->policy()->tenuring_threshold()),
 72     _scanner(g1h, this),
 73     _worker_id(worker_id),
 74     _last_enqueued_card(SIZE_MAX),
 75     _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
 76     _stack_trim_lower_threshold(GCDrainStackTargetSize),
 77     _trim_ticks(),
 78     _surviving_young_words_base(NULL),
 79     _surviving_young_words(NULL),
 80     _surviving_words_length(young_cset_length + 1),
 81     _old_gen_is_full(false),
 82     _partial_objarray_chunk_size(ParGCArrayScanChunk),
 83     _partial_array_stepper(n_workers),
 84     _string_dedup_requests(),
 85     _num_optional_regions(optional_cset_length),
 86     _numa(g1h->numa()),
 87     _obj_alloc_stat(NULL),
 88     EVAC_FAILURE_INJECTOR_ONLY(_evac_failure_inject_counter(0) COMMA)
 89     _preserved_marks(preserved_marks),
 90     _evacuation_failed_info(),
 91     _evac_failure_regions(evac_failure_regions)
 92 {
 93   // We allocate number of young gen regions in the collection set plus one
 94   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
 95   // We also add a few elements at the beginning and at the end in
 96   // an attempt to eliminate cache contention
 97   const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
 98   size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
 99 
100   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
101   _surviving_young_words = _surviving_young_words_base + padding_elem_num;
102   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
103 
104   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
105 
106   _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
107 
108   _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
109 
110   initialize_numa_stats();
111 }
112 
113 size_t G1ParScanThreadState::flush(size_t* surviving_young_words) {
114   _rdc_local_qset.flush();
115   flush_numa_stats();
116   // Update allocation statistics.
117   _plab_allocator->flush_and_retire_stats();
118   _g1h->policy()->record_age_table(&_age_table);
119 
120   if (_evacuation_failed_info.has_failed()) {
121      _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
122   }
123 
124   size_t sum = 0;
125   for (uint i = 0; i < _surviving_words_length; i++) {
126     surviving_young_words[i] += _surviving_young_words[i];
127     sum += _surviving_young_words[i];
128   }
129   return sum;
130 }
131 
132 G1ParScanThreadState::~G1ParScanThreadState() {
133   delete _plab_allocator;
134   delete _closures;
135   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
136   delete[] _oops_into_optional_regions;
137   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
138 }
139 
140 size_t G1ParScanThreadState::lab_waste_words() const {
141   return _plab_allocator->waste();
142 }
143 
144 size_t G1ParScanThreadState::lab_undo_waste_words() const {
145   return _plab_allocator->undo_waste();
146 }
147 
148 #ifdef ASSERT
149 void G1ParScanThreadState::verify_task(narrowOop* task) const {
150   assert(task != NULL, "invariant");
151   assert(UseCompressedOops, "sanity");
152   oop p = RawAccess<>::oop_load(task);
153   assert(_g1h->is_in_reserved(p),
154          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
155 }
156 
157 void G1ParScanThreadState::verify_task(oop* task) const {
158   assert(task != NULL, "invariant");
159   oop p = RawAccess<>::oop_load(task);
160   assert(_g1h->is_in_reserved(p),
161          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
162 }
163 
164 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
165   // Must be in the collection set--it's already been copied.
166   oop p = task.to_source_array();
167   assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
168 }
169 
170 void G1ParScanThreadState::verify_task(ScannerTask task) const {
171   if (task.is_narrow_oop_ptr()) {
172     verify_task(task.to_narrow_oop_ptr());
173   } else if (task.is_oop_ptr()) {
174     verify_task(task.to_oop_ptr());
175   } else if (task.is_partial_array_task()) {
176     verify_task(task.to_partial_array_task());
177   } else {
178     ShouldNotReachHere();
179   }
180 }
181 #endif // ASSERT
182 
183 template <class T>
184 MAYBE_INLINE_EVACUATION
185 void G1ParScanThreadState::do_oop_evac(T* p) {
186   // Reference should not be NULL here as such are never pushed to the task queue.
187   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
188 
189   // Although we never intentionally push references outside of the collection
190   // set, due to (benign) races in the claim mechanism during RSet scanning more
191   // than one thread might claim the same card. So the same card may be
192   // processed multiple times, and so we might get references into old gen here.
193   // So we need to redo this check.
194   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
195   // References pushed onto the work stack should never point to a humongous region
196   // as they are not added to the collection set due to above precondition.
197   assert(!region_attr.is_humongous(),
198          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
199          p2i(obj), _g1h->addr_to_region(cast_from_oop<HeapWord*>(obj)), p2i(p));
200 
201   if (!region_attr.is_in_cset()) {
202     // In this case somebody else already did all the work.
203     return;
204   }
205 
206   markWord m = obj->mark();
207   if (m.is_marked()) {
208     obj = cast_to_oop(m.decode_pointer());
209   } else {
210     obj = do_copy_to_survivor_space(region_attr, obj, m);
211   }
212   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
213 
214   write_ref_field_post(p, obj);
215 }
216 
217 MAYBE_INLINE_EVACUATION
218 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
219   oop from_obj = task.to_source_array();
220 
221   assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
222   assert(from_obj->is_objArray(), "must be obj array");
223   assert(from_obj->is_forwarded(), "must be forwarded");
224 
225   oop to_obj = from_obj->forwardee();
226   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
227   assert(to_obj->is_objArray(), "must be obj array");
228   objArrayOop to_array = objArrayOop(to_obj);
229 
230   PartialArrayTaskStepper::Step step
231     = _partial_array_stepper.next(objArrayOop(from_obj),
232                                   to_array,
233                                   _partial_objarray_chunk_size);
234   for (uint i = 0; i < step._ncreate; ++i) {
235     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
236   }
237 
238   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
239   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
240   // Process claimed task.  The length of to_array is not correct, but
241   // fortunately the iteration ignores the length field and just relies
242   // on start/end.
243   to_array->oop_iterate_range(&_scanner,
244                               step._index,
245                               step._index + _partial_objarray_chunk_size);
246 }
247 
248 MAYBE_INLINE_EVACUATION
249 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
250                                                   oop from_obj,
251                                                   oop to_obj) {
252   assert(from_obj->is_objArray(), "precondition");
253   assert(from_obj->is_forwarded(), "precondition");
254   assert(from_obj->forwardee() == to_obj, "precondition");
255   assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
256   assert(to_obj->is_objArray(), "precondition");
257 
258   objArrayOop to_array = objArrayOop(to_obj);
259 
260   PartialArrayTaskStepper::Step step
261     = _partial_array_stepper.start(objArrayOop(from_obj),
262                                    to_array,
263                                    _partial_objarray_chunk_size);
264 
265   // Push any needed partial scan tasks.  Pushed before processing the
266   // intitial chunk to allow other workers to steal while we're processing.
267   for (uint i = 0; i < step._ncreate; ++i) {
268     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
269   }
270 
271   // Skip the card enqueue iff the object (to_array) is in survivor region.
272   // However, HeapRegion::is_survivor() is too expensive here.
273   // Instead, we use dest_attr.is_young() because the two values are always
274   // equal: successfully allocated young regions must be survivor regions.
275   assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
276   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
277   // Process the initial chunk.  No need to process the type in the
278   // klass, as it will already be handled by processing the built-in
279   // module. The length of to_array is not correct, but fortunately
280   // the iteration ignores that length field and relies on start/end.
281   to_array->oop_iterate_range(&_scanner, 0, step._index);
282 }
283 
284 MAYBE_INLINE_EVACUATION
285 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
286   verify_task(task);
287   if (task.is_narrow_oop_ptr()) {
288     do_oop_evac(task.to_narrow_oop_ptr());
289   } else if (task.is_oop_ptr()) {
290     do_oop_evac(task.to_oop_ptr());
291   } else {
292     do_partial_array(task.to_partial_array_task());
293   }
294 }
295 
296 // Process tasks until overflow queue is empty and local queue
297 // contains no more than threshold entries.  NOINLINE to prevent
298 // inlining into steal_and_trim_queue.
299 ATTRIBUTE_FLATTEN NOINLINE
300 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
301   ScannerTask task;
302   do {
303     while (_task_queue->pop_overflow(task)) {
304       if (!_task_queue->try_push_to_taskqueue(task)) {
305         dispatch_task(task);
306       }
307     }
308     while (_task_queue->pop_local(task, threshold)) {
309       dispatch_task(task);
310     }
311   } while (!_task_queue->overflow_empty());
312 }
313 
314 ATTRIBUTE_FLATTEN
315 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
316   ScannerTask stolen_task;
317   while (task_queues->steal(_worker_id, stolen_task)) {
318     dispatch_task(stolen_task);
319     // Processing stolen task may have added tasks to our queue.
320     trim_queue();
321   }
322 }
323 
324 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
325                                                       size_t word_sz,
326                                                       bool previous_plab_refill_failed,
327                                                       uint node_index) {
328 
329   assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
330 
331   // Right now we only have two types of regions (young / old) so
332   // let's keep the logic here simple. We can generalize it when necessary.
333   if (dest->is_young()) {
334     bool plab_refill_in_old_failed = false;
335     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
336                                                         word_sz,
337                                                         &plab_refill_in_old_failed,
338                                                         node_index);
339     // Make sure that we won't attempt to copy any other objects out
340     // of a survivor region (given that apparently we cannot allocate
341     // any new ones) to avoid coming into this slow path again and again.
342     // Only consider failed PLAB refill here: failed inline allocations are
343     // typically large, so not indicative of remaining space.
344     if (previous_plab_refill_failed) {
345       _tenuring_threshold = 0;
346     }
347 
348     if (obj_ptr != NULL) {
349       dest->set_old();
350     } else {
351       // We just failed to allocate in old gen. The same idea as explained above
352       // for making survivor gen unavailable for allocation applies for old gen.
353       _old_gen_is_full = plab_refill_in_old_failed;
354     }
355     return obj_ptr;
356   } else {
357     _old_gen_is_full = previous_plab_refill_failed;
358     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
359     // no other space to try.
360     return NULL;
361   }
362 }
363 
364 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
365   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
366 
367   if (region_attr.is_young()) {
368     age = !m.has_displaced_mark_helper() ? m.age()
369                                          : m.displaced_mark_helper().age();
370     if (age < _tenuring_threshold) {
371       return region_attr;
372     }
373   }
374   // young-to-old (promotion) or old-to-old; destination is old in both cases.
375   return G1HeapRegionAttr::Old;
376 }
377 
378 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
379                                                   oop const old, size_t word_sz, uint age,
380                                                   HeapWord * const obj_ptr, uint node_index) const {
381   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
382   if (alloc_buf->contains(obj_ptr)) {
383     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
384                                                               dest_attr.type() == G1HeapRegionAttr::Old,
385                                                               alloc_buf->word_sz() * HeapWordSize);
386   } else {
387     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
388                                                                dest_attr.type() == G1HeapRegionAttr::Old);
389   }
390 }
391 
392 NOINLINE
393 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
394                                                    oop old,
395                                                    size_t word_sz,
396                                                    uint age,
397                                                    uint node_index) {
398   HeapWord* obj_ptr = NULL;
399   // Try slow-path allocation unless we're allocating old and old is already full.
400   if (!(dest_attr->is_old() && _old_gen_is_full)) {
401     bool plab_refill_failed = false;
402     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
403                                                            word_sz,
404                                                            &plab_refill_failed,
405                                                            node_index);
406     if (obj_ptr == NULL) {
407       obj_ptr = allocate_in_next_plab(dest_attr,
408                                       word_sz,
409                                       plab_refill_failed,
410                                       node_index);
411     }
412   }
413   if (obj_ptr != NULL) {
414     update_numa_stats(node_index);
415     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
416       // The events are checked individually as part of the actual commit
417       report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
418     }
419   }
420   return obj_ptr;
421 }
422 
423 #if EVAC_FAILURE_INJECTOR
424 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
425   return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
426 }
427 #endif
428 
429 NOINLINE
430 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
431                                            HeapWord* obj_ptr,
432                                            size_t word_sz,
433                                            uint node_index) {
434   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
435 }
436 
437 // Private inline function, for direct internal use and providing the
438 // implementation of the public not-inline function.
439 MAYBE_INLINE_EVACUATION
440 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
441                                                     oop const old,
442                                                     markWord const old_mark) {
443   assert(region_attr.is_in_cset(),
444          "Unexpected region attr type: %s", region_attr.get_type_str());
445 
446   // Get the klass once.  We'll need it again later, and this avoids
447   // re-decoding when it's compressed.
448   Klass* klass = old->klass();
449   const size_t word_sz = old->size_given_klass(klass);
450 
451   uint age = 0;
452   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
453   HeapRegion* const from_region = _g1h->heap_region_containing(old);
454   uint node_index = from_region->node_index();
455 
456   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
457 
458   // PLAB allocations should succeed most of the time, so we'll
459   // normally check against NULL once and that's it.
460   if (obj_ptr == NULL) {
461     obj_ptr = allocate_copy_slow(&dest_attr, old, word_sz, age, node_index);
462     if (obj_ptr == NULL) {
463       // This will either forward-to-self, or detect that someone else has
464       // installed a forwarding pointer.
465       return handle_evacuation_failure_par(old, old_mark, word_sz);
466     }
467   }
468 
469   assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
470   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
471 
472   // Should this evacuation fail?
473   if (inject_evacuation_failure(from_region->hrm_index())) {
474     // Doing this after all the allocation attempts also tests the
475     // undo_allocation() method too.
476     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
477     return handle_evacuation_failure_par(old, old_mark, word_sz);
478   }
479 
480   // We're going to allocate linearly, so might as well prefetch ahead.
481   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
482   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
483 
484   const oop obj = cast_to_oop(obj_ptr);
485   // Because the forwarding is done with memory_order_relaxed there is no
486   // ordering with the above copy.  Clients that get the forwardee must not
487   // examine its contents without other synchronization, since the contents
488   // may not be up to date for them.
489   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
490   if (forward_ptr == NULL) {
491 
492     {
493       const uint young_index = from_region->young_index_in_cset();
494       assert((from_region->is_young() && young_index >  0) ||
495              (!from_region->is_young() && young_index == 0), "invariant" );
496       _surviving_young_words[young_index] += word_sz;
497     }
498 
499     if (dest_attr.is_young()) {
500       if (age < markWord::max_age) {
501         age++;
502         obj->incr_age();
503       }
504       _age_table.add(age, word_sz);
505     } else {
506       // Currently we only have two destinations and we only need BOT updates for
507       // old. If the current allocation was done outside the PLAB this call will
508       // have no effect since the _top of the PLAB has not changed.
509       _plab_allocator->update_bot_for_plab_allocation(dest_attr, word_sz, node_index);
510     }
511 
512     // Most objects are not arrays, so do one array check rather than
513     // checking for each array category for each object.
514     if (klass->is_array_klass()) {
515       if (klass->is_objArray_klass()) {
516         start_partial_objarray(dest_attr, old, obj);
517       } else {
518         // Nothing needs to be done for typeArrays.  Body doesn't contain
519         // any oops to scan, and the type in the klass will already be handled
520         // by processing the built-in module.
521         assert(klass->is_typeArray_klass(), "invariant");
522       }
523       return obj;
524     }
525 
526     // Check for deduplicating young Strings.
527     if (G1StringDedup::is_candidate_from_evacuation(klass,
528                                                     region_attr,
529                                                     dest_attr,
530                                                     age)) {
531       // Record old; request adds a new weak reference, which reference
532       // processing expects to refer to a from-space object.
533       _string_dedup_requests.add(old);
534     }
535 
536     // Skip the card enqueue iff the object (obj) is in survivor region.
537     // However, HeapRegion::is_survivor() is too expensive here.
538     // Instead, we use dest_attr.is_young() because the two values are always
539     // equal: successfully allocated young regions must be survivor regions.
540     assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
541     G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
542     obj->oop_iterate_backwards(&_scanner, klass);
543     return obj;
544   } else {
545     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
546     return forward_ptr;
547   }
548 }
549 
550 // Public not-inline entry point.
551 ATTRIBUTE_FLATTEN
552 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
553                                                  oop old,
554                                                  markWord old_mark) {
555   return do_copy_to_survivor_space(region_attr, old, old_mark);
556 }
557 
558 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
559   assert(worker_id < _n_workers, "out of bounds access");
560   if (_states[worker_id] == NULL) {
561     _states[worker_id] =
562       new G1ParScanThreadState(_g1h, rdcqs(),
563                                _preserved_marks_set->get(worker_id),
564                                worker_id, _n_workers,
565                                _young_cset_length, _optional_cset_length,
566                                _evac_failure_regions);
567   }
568   return _states[worker_id];
569 }
570 
571 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
572   assert(_flushed, "thread local state from the per thread states should have been flushed");
573   return _surviving_young_words_total;
574 }
575 
576 void G1ParScanThreadStateSet::flush() {
577   assert(!_flushed, "thread local state from the per thread states should be flushed once");
578 
579   for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
580     G1ParScanThreadState* pss = _states[worker_id];
581     assert(pss != nullptr, "must be initialized");
582 
583     G1GCPhaseTimes* p = _g1h->phase_times();
584 
585     // Need to get the following two before the call to G1ParThreadScanState::flush()
586     // because it resets the PLAB allocator where we get this info from.
587     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
588     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
589     size_t copied_bytes = pss->flush(_surviving_young_words_total) * HeapWordSize;
590 
591     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
592     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
593     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
594 
595     delete pss;
596     _states[worker_id] = NULL;
597   }
598   _flushed = true;
599 }
600 
601 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
602   for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
603     G1ParScanThreadState* pss = _states[worker_index];
604     assert(pss != nullptr, "must be initialized");
605 
606     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
607     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
608   }
609 }
610 
611 NOINLINE
612 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
613   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
614 
615   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
616   if (forward_ptr == NULL) {
617     // Forward-to-self succeeded. We are the "owner" of the object.
618     HeapRegion* r = _g1h->heap_region_containing(old);
619     // Records evac failure objs, this will help speed up iteration
620     // of these objs later in *remove self forward* phase of post evacuation.
621     r->record_evac_failure_obj(old);
622 
623     if (_evac_failure_regions->record(r->hrm_index())) {
624       _g1h->hr_printer()->evac_failure(r);
625     }
626 
627     _preserved_marks->push_if_necessary(old, m);
628     _evacuation_failed_info.register_copy_failure(word_sz);
629 
630     // For iterating objects that failed evacuation currently we can reuse the
631     // existing closure to scan evacuated objects because:
632     // - for objects referring into the collection set we do not need to gather
633     // cards at this time. The regions they are in will be unconditionally turned
634     // to old regions without remembered sets.
635     // - since we are iterating from a collection set region (i.e. never a Survivor
636     // region), we always need to gather cards for this case.
637     G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
638     old->oop_iterate_backwards(&_scanner);
639 
640     return old;
641   } else {
642     // Forward-to-self failed. Either someone else managed to allocate
643     // space for this object (old != forward_ptr) or they beat us in
644     // self-forwarding it (old == forward_ptr).
645     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
646            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
647            "should not be in the CSet",
648            p2i(old), p2i(forward_ptr));
649     return forward_ptr;
650   }
651 }
652 
653 void G1ParScanThreadState::initialize_numa_stats() {
654   if (_numa->is_enabled()) {
655     LogTarget(Info, gc, heap, numa) lt;
656 
657     if (lt.is_enabled()) {
658       uint num_nodes = _numa->num_active_nodes();
659       // Record only if there are multiple active nodes.
660       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
661       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
662     }
663   }
664 }
665 
666 void G1ParScanThreadState::flush_numa_stats() {
667   if (_obj_alloc_stat != NULL) {
668     uint node_index = _numa->index_of_current_thread();
669     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
670   }
671 }
672 
673 void G1ParScanThreadState::update_numa_stats(uint node_index) {
674   if (_obj_alloc_stat != NULL) {
675     _obj_alloc_stat[node_index]++;
676   }
677 }
678 
679 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
680                                                  G1RedirtyCardsQueueSet* rdcqs,
681                                                  PreservedMarksSet* preserved_marks_set,
682                                                  uint n_workers,
683                                                  size_t young_cset_length,
684                                                  size_t optional_cset_length,
685                                                  G1EvacFailureRegions* evac_failure_regions) :
686     _g1h(g1h),
687     _rdcqs(rdcqs),
688     _preserved_marks_set(preserved_marks_set),
689     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
690     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
691     _young_cset_length(young_cset_length),
692     _optional_cset_length(optional_cset_length),
693     _n_workers(n_workers),
694     _flushed(false),
695     _evac_failure_regions(evac_failure_regions) {
696   for (uint i = 0; i < n_workers; ++i) {
697     _states[i] = NULL;
698   }
699   memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
700 }
701 
702 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
703   assert(_flushed, "thread local state from the per thread states should have been flushed");
704   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
705   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
706 }