1 /*
  2  * Copyright (c) 2014, 2023, 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/continuationGCSupport.inline.hpp"
 37 #include "gc/shared/partialArrayTaskStepper.inline.hpp"
 38 #include "gc/shared/preservedMarks.inline.hpp"
 39 #include "gc/shared/stringdedup/stringDedup.hpp"
 40 #include "gc/shared/taskqueue.inline.hpp"
 41 #include "memory/allocation.inline.hpp"
 42 #include "oops/access.inline.hpp"
 43 #include "oops/oop.inline.hpp"
 44 #include "runtime/atomic.hpp"
 45 #include "runtime/prefetch.inline.hpp"
 46 #include "utilities/globalDefinitions.hpp"
 47 #include "utilities/macros.hpp"
 48 
 49 // In fastdebug builds the code size can get out of hand, potentially
 50 // tripping over compiler limits (which may be bugs, but nevertheless
 51 // need to be taken into consideration).  A side benefit of limiting
 52 // inlining is that we get more call frames that might aid debugging.
 53 // And the fastdebug compile time for this file is much reduced.
 54 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
 55 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
 56 
 57 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
 58                                            G1RedirtyCardsQueueSet* rdcqs,
 59                                            PreservedMarks* preserved_marks,
 60                                            uint worker_id,
 61                                            uint num_workers,
 62                                            G1CollectionSet* collection_set,
 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(nullptr),
 69     _plab_allocator(nullptr),
 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(nullptr),
 79     _surviving_young_words(nullptr),
 80     _surviving_words_length(collection_set->young_region_length() + 1),
 81     _old_gen_is_full(false),
 82     _partial_objarray_chunk_size(ParGCArrayScanChunk),
 83     _partial_array_stepper(num_workers),
 84     _string_dedup_requests(),
 85     _max_num_optional_regions(collection_set->optional_region_length()),
 86     _numa(g1h->numa()),
 87     _obj_alloc_stat(nullptr),
 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     _evac_failure_enqueued_cards(0)
 93 {
 94   // We allocate number of young gen regions in the collection set plus one
 95   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
 96   // We also add a few elements at the beginning and at the end in
 97   // an attempt to eliminate cache contention
 98   const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
 99   size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
100 
101   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
102   _surviving_young_words = _surviving_young_words_base + padding_elem_num;
103   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
104 
105   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
106 
107   _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
108                                                              this,
109                                                              collection_set->only_contains_young_regions());
110 
111   _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
112 
113   initialize_numa_stats();
114 }
115 
116 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
117   _rdc_local_qset.flush();
118   flush_numa_stats();
119   // Update allocation statistics.
120   _plab_allocator->flush_and_retire_stats(num_workers);
121   _g1h->policy()->record_age_table(&_age_table);
122 
123   if (_evacuation_failed_info.has_failed()) {
124      _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
125   }
126 
127   size_t sum = 0;
128   for (uint i = 0; i < _surviving_words_length; i++) {
129     surviving_young_words[i] += _surviving_young_words[i];
130     sum += _surviving_young_words[i];
131   }
132   return sum;
133 }
134 
135 G1ParScanThreadState::~G1ParScanThreadState() {
136   delete _plab_allocator;
137   delete _closures;
138   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
139   delete[] _oops_into_optional_regions;
140   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
141 }
142 
143 size_t G1ParScanThreadState::lab_waste_words() const {
144   return _plab_allocator->waste();
145 }
146 
147 size_t G1ParScanThreadState::lab_undo_waste_words() const {
148   return _plab_allocator->undo_waste();
149 }
150 
151 size_t G1ParScanThreadState::evac_failure_enqueued_cards() const {
152   return _evac_failure_enqueued_cards;
153 }
154 
155 #ifdef ASSERT
156 void G1ParScanThreadState::verify_task(narrowOop* task) const {
157   assert(task != nullptr, "invariant");
158   assert(UseCompressedOops, "sanity");
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(oop* task) const {
165   assert(task != nullptr, "invariant");
166   oop p = RawAccess<>::oop_load(task);
167   assert(_g1h->is_in_reserved(p),
168          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
169 }
170 
171 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
172   // Must be in the collection set--it's already been copied.
173   oop p = task.to_source_array();
174   assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
175 }
176 
177 void G1ParScanThreadState::verify_task(ScannerTask task) const {
178   if (task.is_narrow_oop_ptr()) {
179     verify_task(task.to_narrow_oop_ptr());
180   } else if (task.is_oop_ptr()) {
181     verify_task(task.to_oop_ptr());
182   } else if (task.is_partial_array_task()) {
183     verify_task(task.to_partial_array_task());
184   } else {
185     ShouldNotReachHere();
186   }
187 }
188 #endif // ASSERT
189 
190 template <class T>
191 MAYBE_INLINE_EVACUATION
192 void G1ParScanThreadState::do_oop_evac(T* p) {
193   // Reference should not be null here as such are never pushed to the task queue.
194   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
195 
196   // Although we never intentionally push references outside of the collection
197   // set, due to (benign) races in the claim mechanism during RSet scanning more
198   // than one thread might claim the same card. So the same card may be
199   // processed multiple times, and so we might get references into old gen here.
200   // So we need to redo this check.
201   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
202   // References pushed onto the work stack should never point to a humongous region
203   // as they are not added to the collection set due to above precondition.
204   assert(!region_attr.is_humongous_candidate(),
205          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
206          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
207 
208   if (!region_attr.is_in_cset()) {
209     // In this case somebody else already did all the work.
210     return;
211   }
212 
213   markWord m = obj->mark();
214   if (m.is_marked()) {
215     obj = cast_to_oop(m.decode_pointer());
216   } else {
217     obj = do_copy_to_survivor_space(region_attr, obj, m);
218   }
219   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
220 
221   write_ref_field_post(p, obj);
222 }
223 
224 MAYBE_INLINE_EVACUATION
225 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
226   oop from_obj = task.to_source_array();
227 
228   assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
229   assert(from_obj->is_objArray(), "must be obj array");
230   assert(from_obj->is_forwarded(), "must be forwarded");
231 
232   oop to_obj = from_obj->forwardee();
233   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
234   assert(to_obj->is_objArray(), "must be obj array");
235   objArrayOop to_array = objArrayOop(to_obj);
236 
237   PartialArrayTaskStepper::Step step
238     = _partial_array_stepper.next(objArrayOop(from_obj),
239                                   to_array,
240                                   _partial_objarray_chunk_size);
241   for (uint i = 0; i < step._ncreate; ++i) {
242     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
243   }
244 
245   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
246   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
247   // Process claimed task.  The length of to_array is not correct, but
248   // fortunately the iteration ignores the length field and just relies
249   // on start/end.
250   to_array->oop_iterate_range(&_scanner,
251                               step._index,
252                               step._index + _partial_objarray_chunk_size);
253 }
254 
255 MAYBE_INLINE_EVACUATION
256 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
257                                                   oop from_obj,
258                                                   oop to_obj) {
259   assert(from_obj->is_objArray(), "precondition");
260   assert(from_obj->is_forwarded(), "precondition");
261   assert(from_obj->forwardee() == to_obj, "precondition");
262   assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
263   assert(to_obj->is_objArray(), "precondition");
264 
265   objArrayOop to_array = objArrayOop(to_obj);
266 
267   PartialArrayTaskStepper::Step step
268     = _partial_array_stepper.start(objArrayOop(from_obj),
269                                    to_array,
270                                    _partial_objarray_chunk_size);
271 
272   // Push any needed partial scan tasks.  Pushed before processing the
273   // initial chunk to allow other workers to steal while we're processing.
274   for (uint i = 0; i < step._ncreate; ++i) {
275     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
276   }
277 
278   // Skip the card enqueue iff the object (to_array) is in survivor region.
279   // However, HeapRegion::is_survivor() is too expensive here.
280   // Instead, we use dest_attr.is_young() because the two values are always
281   // equal: successfully allocated young regions must be survivor regions.
282   assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
283   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
284   // Process the initial chunk.  No need to process the type in the
285   // klass, as it will already be handled by processing the built-in
286   // module. The length of to_array is not correct, but fortunately
287   // the iteration ignores that length field and relies on start/end.
288   to_array->oop_iterate_range(&_scanner, 0, step._index);
289 }
290 
291 MAYBE_INLINE_EVACUATION
292 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
293   verify_task(task);
294   if (task.is_narrow_oop_ptr()) {
295     do_oop_evac(task.to_narrow_oop_ptr());
296   } else if (task.is_oop_ptr()) {
297     do_oop_evac(task.to_oop_ptr());
298   } else {
299     do_partial_array(task.to_partial_array_task());
300   }
301 }
302 
303 // Process tasks until overflow queue is empty and local queue
304 // contains no more than threshold entries.  NOINLINE to prevent
305 // inlining into steal_and_trim_queue.
306 ATTRIBUTE_FLATTEN NOINLINE
307 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
308   ScannerTask task;
309   do {
310     while (_task_queue->pop_overflow(task)) {
311       if (!_task_queue->try_push_to_taskqueue(task)) {
312         dispatch_task(task);
313       }
314     }
315     while (_task_queue->pop_local(task, threshold)) {
316       dispatch_task(task);
317     }
318   } while (!_task_queue->overflow_empty());
319 }
320 
321 ATTRIBUTE_FLATTEN
322 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
323   ScannerTask stolen_task;
324   while (task_queues->steal(_worker_id, stolen_task)) {
325     dispatch_task(stolen_task);
326     // Processing stolen task may have added tasks to our queue.
327     trim_queue();
328   }
329 }
330 
331 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
332                                                       size_t word_sz,
333                                                       bool previous_plab_refill_failed,
334                                                       uint node_index) {
335 
336   assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
337 
338   // Right now we only have two types of regions (young / old) so
339   // let's keep the logic here simple. We can generalize it when necessary.
340   if (dest->is_young()) {
341     bool plab_refill_in_old_failed = false;
342     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
343                                                         word_sz,
344                                                         &plab_refill_in_old_failed,
345                                                         node_index);
346     // Make sure that we won't attempt to copy any other objects out
347     // of a survivor region (given that apparently we cannot allocate
348     // any new ones) to avoid coming into this slow path again and again.
349     // Only consider failed PLAB refill here: failed inline allocations are
350     // typically large, so not indicative of remaining space.
351     if (previous_plab_refill_failed) {
352       _tenuring_threshold = 0;
353     }
354 
355     if (obj_ptr != nullptr) {
356       dest->set_old();
357     } else {
358       // We just failed to allocate in old gen. The same idea as explained above
359       // for making survivor gen unavailable for allocation applies for old gen.
360       _old_gen_is_full = plab_refill_in_old_failed;
361     }
362     return obj_ptr;
363   } else {
364     _old_gen_is_full = previous_plab_refill_failed;
365     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
366     // no other space to try.
367     return nullptr;
368   }
369 }
370 
371 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
372   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
373 
374   if (region_attr.is_young()) {
375     age = !m.has_displaced_mark_helper() ? m.age()
376                                          : m.displaced_mark_helper().age();
377     if (age < _tenuring_threshold) {
378       return region_attr;
379     }
380   }
381   // young-to-old (promotion) or old-to-old; destination is old in both cases.
382   return G1HeapRegionAttr::Old;
383 }
384 
385 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
386                                                   oop const old, size_t word_sz, uint age,
387                                                   HeapWord * const obj_ptr, uint node_index) const {
388   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
389   if (alloc_buf->contains(obj_ptr)) {
390     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
391                                                               dest_attr.type() == G1HeapRegionAttr::Old,
392                                                               alloc_buf->word_sz() * HeapWordSize);
393   } else {
394     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
395                                                                dest_attr.type() == G1HeapRegionAttr::Old);
396   }
397 }
398 
399 NOINLINE
400 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
401                                                    oop old,
402                                                    size_t word_sz,
403                                                    uint age,
404                                                    uint node_index) {
405   HeapWord* obj_ptr = nullptr;
406   // Try slow-path allocation unless we're allocating old and old is already full.
407   if (!(dest_attr->is_old() && _old_gen_is_full)) {
408     bool plab_refill_failed = false;
409     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
410                                                            word_sz,
411                                                            &plab_refill_failed,
412                                                            node_index);
413     if (obj_ptr == nullptr) {
414       obj_ptr = allocate_in_next_plab(dest_attr,
415                                       word_sz,
416                                       plab_refill_failed,
417                                       node_index);
418     }
419   }
420   if (obj_ptr != nullptr) {
421     update_numa_stats(node_index);
422     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
423       // The events are checked individually as part of the actual commit
424       report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
425     }
426   }
427   return obj_ptr;
428 }
429 
430 #if EVAC_FAILURE_INJECTOR
431 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
432   return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
433 }
434 #endif
435 
436 NOINLINE
437 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
438                                            HeapWord* obj_ptr,
439                                            size_t word_sz,
440                                            uint node_index) {
441   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
442 }
443 
444 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
445   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
446   HeapRegion* region = _g1h->heap_region_containing(obj_start);
447   region->update_bot_for_obj(obj_start, word_sz);
448 }
449 
450 // Private inline function, for direct internal use and providing the
451 // implementation of the public not-inline function.
452 MAYBE_INLINE_EVACUATION
453 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
454                                                     oop const old,
455                                                     markWord const old_mark) {
456   assert(region_attr.is_in_cset(),
457          "Unexpected region attr type: %s", region_attr.get_type_str());
458 
459   // Get the klass once.  We'll need it again later, and this avoids
460   // re-decoding when it's compressed.
461   Klass* klass = old->klass();
462   const size_t word_sz = old->size_given_klass(klass);
463 
464   uint age = 0;
465   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
466   HeapRegion* const from_region = _g1h->heap_region_containing(old);
467   uint node_index = from_region->node_index();
468 
469   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
470 
471   // PLAB allocations should succeed most of the time, so we'll
472   // normally check against null once and that's it.
473   if (obj_ptr == nullptr) {
474     obj_ptr = allocate_copy_slow(&dest_attr, old, word_sz, age, node_index);
475     if (obj_ptr == nullptr) {
476       // This will either forward-to-self, or detect that someone else has
477       // installed a forwarding pointer.
478       return handle_evacuation_failure_par(old, old_mark, word_sz);
479     }
480   }
481 
482   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
483   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
484 
485   // Should this evacuation fail?
486   if (inject_evacuation_failure(from_region->hrm_index())) {
487     // Doing this after all the allocation attempts also tests the
488     // undo_allocation() method too.
489     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
490     return handle_evacuation_failure_par(old, old_mark, word_sz);
491   }
492 
493   // We're going to allocate linearly, so might as well prefetch ahead.
494   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
495   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
496 
497   const oop obj = cast_to_oop(obj_ptr);
498   // Because the forwarding is done with memory_order_relaxed there is no
499   // ordering with the above copy.  Clients that get the forwardee must not
500   // examine its contents without other synchronization, since the contents
501   // may not be up to date for them.
502   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
503   if (forward_ptr == nullptr) {
504 
505     {
506       const uint young_index = from_region->young_index_in_cset();
507       assert((from_region->is_young() && young_index >  0) ||
508              (!from_region->is_young() && young_index == 0), "invariant" );
509       _surviving_young_words[young_index] += word_sz;
510     }
511 
512     if (dest_attr.is_young()) {
513       if (age < markWord::max_age) {
514         age++;
515         obj->incr_age();
516       }
517       _age_table.add(age, word_sz);
518     } else {
519       update_bot_after_copying(obj, word_sz);
520     }
521 
522     // Most objects are not arrays, so do one array check rather than
523     // checking for each array category for each object.
524     // CMH: Valhalla flat arrays can split this work up, but for now, doesn't
525     if (klass->is_array_klass() && !klass->is_flatArray_klass()) {
526       if (klass->is_objArray_klass()) {
527         start_partial_objarray(dest_attr, old, obj);
528       } else {
529         // Nothing needs to be done for typeArrays.  Body doesn't contain
530         // any oops to scan, and the type in the klass will already be handled
531         // by processing the built-in module.
532         assert(klass->is_typeArray_klass(), "invariant");
533       }
534       return obj;
535     }
536 
537     ContinuationGCSupport::transform_stack_chunk(obj);
538 
539     // Check for deduplicating young Strings.
540     if (G1StringDedup::is_candidate_from_evacuation(klass,
541                                                     region_attr,
542                                                     dest_attr,
543                                                     age)) {
544       // Record old; request adds a new weak reference, which reference
545       // processing expects to refer to a from-space object.
546       _string_dedup_requests.add(old);
547     }
548 
549     // Skip the card enqueue iff the object (obj) is in survivor region.
550     // However, HeapRegion::is_survivor() is too expensive here.
551     // Instead, we use dest_attr.is_young() because the two values are always
552     // equal: successfully allocated young regions must be survivor regions.
553     assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
554     G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
555     obj->oop_iterate_backwards(&_scanner, klass);
556     return obj;
557   } else {
558     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
559     return forward_ptr;
560   }
561 }
562 
563 // Public not-inline entry point.
564 ATTRIBUTE_FLATTEN
565 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
566                                                  oop old,
567                                                  markWord old_mark) {
568   return do_copy_to_survivor_space(region_attr, old, old_mark);
569 }
570 
571 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
572   assert(worker_id < _num_workers, "out of bounds access");
573   if (_states[worker_id] == nullptr) {
574     _states[worker_id] =
575       new G1ParScanThreadState(_g1h, rdcqs(),
576                                _preserved_marks_set.get(worker_id),
577                                worker_id,
578                                _num_workers,
579                                _collection_set,
580                                _evac_failure_regions);
581   }
582   return _states[worker_id];
583 }
584 
585 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
586   assert(_flushed, "thread local state from the per thread states should have been flushed");
587   return _surviving_young_words_total;
588 }
589 
590 void G1ParScanThreadStateSet::flush_stats() {
591   assert(!_flushed, "thread local state from the per thread states should be flushed once");
592 
593   for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
594     G1ParScanThreadState* pss = _states[worker_id];
595     assert(pss != nullptr, "must be initialized");
596 
597     G1GCPhaseTimes* p = _g1h->phase_times();
598 
599     // Need to get the following two before the call to G1ParThreadScanState::flush()
600     // because it resets the PLAB allocator where we get this info from.
601     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
602     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
603     size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers) * HeapWordSize;
604     size_t evac_fail_enqueued_cards = pss->evac_failure_enqueued_cards();
605 
606     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
607     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
608     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
609     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_fail_enqueued_cards, G1GCPhaseTimes::MergePSSEvacFailExtra);
610 
611     delete pss;
612     _states[worker_id] = nullptr;
613   }
614   _flushed = true;
615 }
616 
617 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
618   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
619     G1ParScanThreadState* pss = _states[worker_index];
620     assert(pss != nullptr, "must be initialized");
621 
622     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
623     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
624   }
625 }
626 
627 NOINLINE
628 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
629   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
630 
631   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
632   if (forward_ptr == nullptr) {
633     // Forward-to-self succeeded. We are the "owner" of the object.
634     HeapRegion* r = _g1h->heap_region_containing(old);
635 
636     if (_evac_failure_regions->record(r->hrm_index())) {
637       _g1h->hr_printer()->evac_failure(r);
638     }
639 
640     // Mark the failing object in the marking bitmap and later use the bitmap to handle
641     // evacuation failure recovery.
642     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
643 
644     _preserved_marks->push_if_necessary(old, m);
645 
646     ContinuationGCSupport::transform_stack_chunk(old);
647 
648     _evacuation_failed_info.register_copy_failure(word_sz);
649 
650     // For iterating objects that failed evacuation currently we can reuse the
651     // existing closure to scan evacuated objects; since we are iterating from a
652     // collection set region (i.e. never a Survivor region), we always need to
653     // gather cards for this case.
654     G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
655     old->oop_iterate_backwards(&_scanner);
656 
657     return old;
658   } else {
659     // Forward-to-self failed. Either someone else managed to allocate
660     // space for this object (old != forward_ptr) or they beat us in
661     // self-forwarding it (old == forward_ptr).
662     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
663            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
664            "should not be in the CSet",
665            p2i(old), p2i(forward_ptr));
666     return forward_ptr;
667   }
668 }
669 
670 void G1ParScanThreadState::initialize_numa_stats() {
671   if (_numa->is_enabled()) {
672     LogTarget(Info, gc, heap, numa) lt;
673 
674     if (lt.is_enabled()) {
675       uint num_nodes = _numa->num_active_nodes();
676       // Record only if there are multiple active nodes.
677       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
678       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
679     }
680   }
681 }
682 
683 void G1ParScanThreadState::flush_numa_stats() {
684   if (_obj_alloc_stat != nullptr) {
685     uint node_index = _numa->index_of_current_thread();
686     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
687   }
688 }
689 
690 void G1ParScanThreadState::update_numa_stats(uint node_index) {
691   if (_obj_alloc_stat != nullptr) {
692     _obj_alloc_stat[node_index]++;
693   }
694 }
695 
696 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
697                                                  uint num_workers,
698                                                  G1CollectionSet* collection_set,
699                                                  G1EvacFailureRegions* evac_failure_regions) :
700     _g1h(g1h),
701     _collection_set(collection_set),
702     _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
703     _preserved_marks_set(true /* in_c_heap */),
704     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
705     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
706     _num_workers(num_workers),
707     _flushed(false),
708     _evac_failure_regions(evac_failure_regions) {
709   _preserved_marks_set.init(num_workers);
710   for (uint i = 0; i < num_workers; ++i) {
711     _states[i] = nullptr;
712   }
713   memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
714 }
715 
716 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
717   assert(_flushed, "thread local state from the per thread states should have been flushed");
718   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
719   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
720   _preserved_marks_set.reclaim();
721 }