1 /*
  2  * Copyright (c) 2014, 2026, Oracle and/or its affiliates. All rights reserved.
  3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  4  *
  5  * This code is free software; you can redistribute it and/or modify it
  6  * under the terms of the GNU General Public License version 2 only, as
  7  * published by the Free Software Foundation.
  8  *
  9  * This code is distributed in the hope that it will be useful, but WITHOUT
 10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 12  * version 2 for more details (a copy is included in the LICENSE file that
 13  * accompanied this code).
 14  *
 15  * You should have received a copy of the GNU General Public License version
 16  * 2 along with this work; if not, write to the Free Software Foundation,
 17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 18  *
 19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 20  * or visit www.oracle.com if you need additional information or have any
 21  * questions.
 22  *
 23  */
 24 
 25 #include "gc/g1/g1Allocator.inline.hpp"
 26 #include "gc/g1/g1CollectedHeap.inline.hpp"
 27 #include "gc/g1/g1CollectionSet.hpp"
 28 #include "gc/g1/g1EvacFailureRegions.inline.hpp"
 29 #include "gc/g1/g1HeapRegionPrinter.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/g1YoungGCAllocationFailureInjector.inline.hpp"
 36 #include "gc/shared/continuationGCSupport.inline.hpp"
 37 #include "gc/shared/partialArraySplitter.inline.hpp"
 38 #include "gc/shared/partialArrayState.hpp"
 39 #include "gc/shared/partialArrayTaskStats.hpp"
 40 #include "gc/shared/stringdedup/stringDedup.hpp"
 41 #include "gc/shared/taskqueue.inline.hpp"
 42 #include "memory/allocation.inline.hpp"
 43 #include "oops/access.inline.hpp"
 44 #include "oops/oop.inline.hpp"
 45 #include "runtime/mutexLocker.hpp"
 46 #include "runtime/prefetch.inline.hpp"
 47 #include "utilities/globalDefinitions.hpp"
 48 #include "utilities/macros.hpp"
 49 
 50 // In fastdebug builds the code size can get out of hand, potentially
 51 // tripping over compiler limits (which may be bugs, but nevertheless
 52 // need to be taken into consideration).  A side benefit of limiting
 53 // inlining is that we get more call frames that might aid debugging.
 54 // And the fastdebug compile time for this file is much reduced.
 55 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
 56 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
 57 
 58 // Good estimate for the initial table size.
 59 static uint initial_nmethod_table_size(G1CollectedHeap* g1h) {
 60   // The +1 is both to consider the retained old region likely to be added, and avoid zero-sized initial tables.
 61   return MIN3(g1h->collection_set()->num_regions(), g1h->max_num_regions() / 2, g1h->num_available_regions()) + 1;
 62 }
 63 
 64 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
 65                                            G1ParScanThreadStateSet* per_thread_states,
 66                                            uint worker_id,
 67                                            uint num_workers,
 68                                            G1CollectionSet* collection_set,
 69                                            G1EvacFailureRegions* evac_failure_regions)
 70   : _g1h(g1h),
 71     _per_thread_states(per_thread_states),
 72     _task_queue(g1h->task_queue(worker_id)),
 73     _ct(g1h->refinement_table()),
 74     _closures(nullptr),
 75     _plab_allocator(nullptr),
 76     _age_table(false),
 77     _tenuring_threshold(g1h->policy()->tenuring_threshold()),
 78     _scanner(g1h, this),
 79     _worker_id(worker_id),
 80     _num_cards_marked_dirty(0),
 81     _num_cards_marked_to_cset(0),
 82     _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
 83     _stack_trim_lower_threshold(GCDrainStackTargetSize),
 84     _trim_ticks(),
 85     _surviving_young_words_base(nullptr),
 86     _surviving_young_words(nullptr),
 87     _surviving_words_length(collection_set->num_young_regions() + 1),
 88     _old_gen_is_full(false),
 89     _partial_array_splitter(g1h->partial_array_state_manager(), num_workers),
 90     _string_dedup_requests(),
 91     _max_num_optional_regions(collection_set->num_optional_regions()),
 92     _numa(g1h->numa()),
 93     _obj_alloc_stat(nullptr),
 94     // The initial size estimate is relatively conservative, assuming that all regions
 95     // in the collection set get evacuated into the same amount of new regions.
 96     _nmethods_to_add(initial_nmethod_table_size(g1h),
 97                      MAX2(initial_nmethod_table_size(g1h), _g1h->max_num_regions() / 2)),
 98     ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
 99     _evacuation_failed_info(),
100     _evac_failure_regions(evac_failure_regions),
101     _num_cards_from_evac_failure(0)
102 {
103   // We allocate number of young gen regions in the collection set plus one
104   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
105   // We also add a few elements at the beginning and at the end in
106   // an attempt to eliminate cache contention
107   const size_t padding_elem_num = (DEFAULT_PADDING_SIZE / sizeof(size_t));
108   size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
109 
110   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
111   _surviving_young_words = _surviving_young_words_base + padding_elem_num;
112   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
113 
114   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
115 
116   _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
117                                                              this,
118                                                              collection_set->only_contains_young_regions());
119 
120   _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
121 
122   initialize_numa_stats();
123 }
124 
125 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
126   flush_numa_stats();
127   // Update allocation statistics.
128   _plab_allocator->flush_and_retire_stats(num_workers);
129   _g1h->policy()->record_age_table(&_age_table);
130 
131   if (_evacuation_failed_info.has_failed()) {
132     _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
133   }
134 
135   size_t sum = 0;
136   for (uint i = 0; i < _surviving_words_length; i++) {
137     surviving_young_words[i] += _surviving_young_words[i];
138     sum += _surviving_young_words[i];
139   }
140   return sum;
141 }
142 
143 G1ParScanThreadState::~G1ParScanThreadState() {
144   auto delete_all = [&] (uint region, G1NmethodSet* nmethods) -> bool {
145     delete nmethods;
146     return true;
147   };
148   _nmethods_to_add.iterate(delete_all);
149 
150   delete _plab_allocator;
151   delete _closures;
152   FREE_C_HEAP_ARRAY(_surviving_young_words_base);
153   delete[] _oops_into_optional_regions;
154   FREE_C_HEAP_ARRAY(_obj_alloc_stat);
155 }
156 
157 size_t G1ParScanThreadState::lab_waste_words() const {
158   return _plab_allocator->waste();
159 }
160 
161 size_t G1ParScanThreadState::lab_undo_waste_words() const {
162   return _plab_allocator->undo_waste();
163 }
164 
165 size_t G1ParScanThreadState::num_cards_pending() const {
166   return _num_cards_marked_dirty + _num_cards_from_evac_failure;
167 }
168 
169 size_t G1ParScanThreadState::num_cards_marked() const {
170   return num_cards_pending() + _num_cards_marked_to_cset;
171 }
172 
173 size_t G1ParScanThreadState::num_cards_from_evac_failure() const {
174   return _num_cards_from_evac_failure;
175 }
176 
177 #ifdef ASSERT
178 void G1ParScanThreadState::verify_task(narrowOop* task) const {
179   assert(task != nullptr, "invariant");
180   assert(UseCompressedOops, "sanity");
181   oop p = RawAccess<>::oop_load(task);
182   assert(_g1h->is_in_reserved(p),
183          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
184 }
185 
186 void G1ParScanThreadState::verify_task(oop* task) const {
187   assert(task != nullptr, "invariant");
188   oop p = RawAccess<>::oop_load(task);
189   assert(_g1h->is_in_reserved(p),
190          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
191 }
192 
193 void G1ParScanThreadState::verify_task(PartialArrayState* task) const {
194   assert(task != nullptr, "invariant");
195   // Source isn't used for processing, so not recorded in task.
196   assert(task->source() == nullptr, "invariant");
197   oop p = task->destination();
198   assert(_g1h->is_in_reserved(p),
199          "task=" PTR_FORMAT " dest=" PTR_FORMAT, p2i(task), p2i(p));
200 }
201 
202 void G1ParScanThreadState::verify_task(ScannerTask task) const {
203   if (task.is_narrow_oop_ptr()) {
204     verify_task(task.to_narrow_oop_ptr());
205   } else if (task.is_oop_ptr()) {
206     verify_task(task.to_oop_ptr());
207   } else if (task.is_partial_array_state()) {
208     verify_task(task.to_partial_array_state());
209   } else {
210     ShouldNotReachHere();
211   }
212 }
213 #endif // ASSERT
214 
215 template <class T>
216 MAYBE_INLINE_EVACUATION
217 void G1ParScanThreadState::do_oop_evac(T* p) {
218   // Reference should not be null here as such are never pushed to the task queue.
219   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
220 
221   // Although we never intentionally push references outside of the collection
222   // set, due to (benign) races in the claim mechanism during RSet scanning more
223   // than one thread might claim the same card. So the same card may be
224   // processed multiple times, and so we might get references into old gen here.
225   // So we need to redo this check.
226   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
227   // References pushed onto the work stack should never point to a humongous region
228   // as they are not added to the collection set due to above precondition.
229   assert(!region_attr.is_humongous_candidate(),
230          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
231          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
232 
233   if (!region_attr.is_in_cset()) {
234     // In this case somebody else already did all the work.
235     return;
236   }
237 
238   markWord m = obj->mark();
239   if (m.is_forwarded()) {
240     obj = obj->forwardee(m);
241   } else {
242     obj = do_copy_to_survivor_space(region_attr, obj, m);
243   }
244   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
245 
246   write_ref_field_post(p, obj);
247 }
248 
249 MAYBE_INLINE_EVACUATION
250 void G1ParScanThreadState::do_partial_array(PartialArrayState* state, bool stolen) {
251   // Access state before release by claim().
252   objArrayOop to_array = objArrayOop(state->destination());
253   PartialArraySplitter::Claim claim =
254     _partial_array_splitter.claim(state, _task_queue, stolen);
255   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
256   G1SkipCardMarkSetter x(&_scanner, dest_attr.is_new_survivor());
257   // Process claimed task.
258   to_array->oop_iterate_elements_range(&_scanner,
259                                        checked_cast<int>(claim._start),
260                                        checked_cast<int>(claim._end));
261 }
262 
263 MAYBE_INLINE_EVACUATION
264 void G1ParScanThreadState::start_partial_objarray(oop from_obj,
265                                                   oop to_obj) {
266   assert(from_obj->is_forwarded(), "precondition");
267   assert(from_obj->forwardee() == to_obj, "precondition");
268   assert(to_obj->is_objArray(), "precondition");
269 
270   objArrayOop to_array = objArrayOop(to_obj);
271   size_t array_length = to_array->length();
272   size_t initial_chunk_size =
273     // The source array is unused when processing states.
274     _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length, ParGCArrayScanChunk);
275 
276   assert(_scanner.skip_card_mark_set(), "must be");
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.
280   to_array->oop_iterate_elements_range(&_scanner, 0, checked_cast<int>(initial_chunk_size));
281 }
282 
283 MAYBE_INLINE_EVACUATION
284 void G1ParScanThreadState::dispatch_task(ScannerTask task, bool stolen) {
285   verify_task(task);
286   if (task.is_narrow_oop_ptr()) {
287     do_oop_evac(task.to_narrow_oop_ptr());
288   } else if (task.is_oop_ptr()) {
289     do_oop_evac(task.to_oop_ptr());
290   } else {
291     do_partial_array(task.to_partial_array_state(), stolen);
292   }
293 }
294 
295 // Process tasks until overflow queue is empty and local queue
296 // contains no more than threshold entries.  NOINLINE to prevent
297 // inlining into steal_and_trim_queue.
298 ATTRIBUTE_FLATTEN NOINLINE
299 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
300   ScannerTask task;
301   do {
302     while (_task_queue->pop_overflow(task)) {
303       if (!_task_queue->try_push_to_taskqueue(task)) {
304         dispatch_task(task, false);
305       }
306     }
307     while (_task_queue->pop_local(task, threshold)) {
308       dispatch_task(task, false);
309     }
310   } while (!_task_queue->overflow_empty());
311 }
312 
313 ATTRIBUTE_FLATTEN
314 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
315   ScannerTask stolen_task;
316   while (task_queues->steal(_worker_id, stolen_task)) {
317     dispatch_task(stolen_task, true);
318     // Processing stolen task may have added tasks to our queue.
319     trim_queue();
320   }
321 }
322 
323 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
324                                                       size_t word_sz,
325                                                       bool previous_plab_refill_failed,
326                                                       uint node_index) {
327 
328   assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
329 
330   // Right now we only have two types of regions (young / old) so
331   // let's keep the logic here simple. We can generalize it when necessary.
332   if (dest->is_young()) {
333     bool plab_refill_in_old_failed = false;
334     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
335                                                         word_sz,
336                                                         &plab_refill_in_old_failed,
337                                                         node_index);
338     // Make sure that we won't attempt to copy any other objects out
339     // of a survivor region (given that apparently we cannot allocate
340     // any new ones) to avoid coming into this slow path again and again.
341     // Only consider failed PLAB refill here: failed inline allocations are
342     // typically large, so not indicative of remaining space.
343     if (previous_plab_refill_failed) {
344       _tenuring_threshold = 0;
345     }
346 
347     if (obj_ptr != nullptr) {
348       dest->set_old();
349     } else {
350       // We just failed to allocate in old gen. The same idea as explained above
351       // for making survivor gen unavailable for allocation applies for old gen.
352       _old_gen_is_full = plab_refill_in_old_failed;
353     }
354     return obj_ptr;
355   } else {
356     _old_gen_is_full = previous_plab_refill_failed;
357     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
358     // no other space to try.
359     return nullptr;
360   }
361 }
362 
363 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
364   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
365 
366   if (region_attr.is_young()) {
367     age = !m.has_displaced_mark_helper() ? m.age()
368                                          : m.displaced_mark_helper().age();
369     if (age < _tenuring_threshold) {
370       return region_attr;
371     }
372   }
373   // young-to-old (promotion) or old-to-old; destination is old in both cases.
374   return G1HeapRegionAttr::Old;
375 }
376 
377 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
378                                                   Klass* klass, size_t word_sz, uint age,
379                                                   HeapWord * const obj_ptr, uint node_index) const {
380   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
381   if (alloc_buf->contains(obj_ptr)) {
382     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
383                                                               dest_attr.type() == G1HeapRegionAttr::Old,
384                                                               alloc_buf->word_sz() * HeapWordSize);
385   } else {
386     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
387                                                                dest_attr.type() == G1HeapRegionAttr::Old);
388   }
389 }
390 
391 NOINLINE
392 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
393                                                    Klass* klass,
394                                                    size_t word_sz,
395                                                    uint age,
396                                                    uint node_index) {
397   HeapWord* obj_ptr = nullptr;
398   // Try slow-path allocation unless we're allocating old and old is already full.
399   if (!(dest_attr->is_old() && _old_gen_is_full)) {
400     bool plab_refill_failed = false;
401     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
402                                                            word_sz,
403                                                            &plab_refill_failed,
404                                                            node_index);
405     if (obj_ptr == nullptr) {
406       obj_ptr = allocate_in_next_plab(dest_attr,
407                                       word_sz,
408                                       plab_refill_failed,
409                                       node_index);
410     }
411   }
412   if (obj_ptr != nullptr) {
413     update_numa_stats(node_index);
414     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
415       // The events are checked individually as part of the actual commit
416       report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
417     }
418   }
419   return obj_ptr;
420 }
421 
422 #if ALLOCATION_FAILURE_INJECTOR
423 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
424   return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
425 }
426 #endif
427 
428 NOINLINE
429 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
430                                            HeapWord* obj_ptr,
431                                            size_t word_sz,
432                                            uint node_index) {
433   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
434 }
435 
436 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
437   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
438   G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
439   region->update_bot_for_block(obj_start, obj_start + word_sz);
440 }
441 
442 ALWAYSINLINE
443 void G1ParScanThreadState::do_iterate_object(oop const obj,
444                                              oop const old,
445                                              Klass* const klass,
446                                              G1HeapRegionAttr const region_attr,
447                                              G1HeapRegionAttr const dest_attr,
448                                              uint age) {
449     // Most objects are not arrays, so do one array check rather than
450     // checking for each array category for each object.
451     if (klass->is_array_klass()) {
452       assert(!klass->is_stack_chunk_instance_klass(), "must be");
453 
454       if (klass->is_objArray_klass()) {
455         start_partial_objarray(old, obj);
456       } else {
457         // Nothing needs to be done for typeArrays.  Body doesn't contain
458         // any oops to scan, and the type in the klass will already be handled
459         // by processing the built-in module.
460         assert(klass->is_typeArray_klass(), "invariant");
461       }
462       return;
463     }
464 
465     ContinuationGCSupport::transform_stack_chunk(obj, klass);
466 
467     // Check for deduplicating young Strings.
468     if (G1StringDedup::is_candidate_from_evacuation(klass,
469                                                     region_attr,
470                                                     dest_attr,
471                                                     age)) {
472       // Record old; request adds a new weak reference, which reference
473       // processing expects to refer to a from-space object.
474       _string_dedup_requests.add(old);
475     }
476 
477     assert(_scanner.skip_card_mark_set(), "must be");
478     obj->oop_iterate_backwards(&_scanner, klass);
479 }
480 
481 // Private inline function, for direct internal use and providing the
482 // implementation of the public not-inline function.
483 MAYBE_INLINE_EVACUATION
484 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
485                                                     oop const old,
486                                                     markWord const old_mark) {
487   assert(region_attr.is_in_cset(),
488          "Unexpected region attr type: %s", region_attr.get_type_str());
489 
490   // NOTE: With compact headers, it is not safe to load the Klass* from old, because
491   // that would access the mark-word, that might change at any time by concurrent
492   // workers.
493   // This mark word would refer to a forwardee, which may not yet have completed
494   // copying. Therefore we must load the Klass* from the mark-word that we already
495   // loaded. This is safe, because we only enter here if not yet forwarded.
496   assert(!old_mark.is_forwarded(), "precondition");
497   Klass* klass = UseCompactObjectHeaders
498       ? old_mark.klass()
499       : old->klass();
500 
501   const size_t word_sz = old->size_given_klass(klass);
502 
503   // JNI only allows pinning of typeArrays, so we only need to keep those in place.
504   if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
505     return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, true /* cause_pinned */);
506   }
507 
508   uint age = 0;
509   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
510   G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
511   uint node_index = from_region->node_index();
512 
513   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
514 
515   // PLAB allocations should succeed most of the time, so we'll
516   // normally check against null once and that's it.
517   if (obj_ptr == nullptr) {
518     obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
519     if (obj_ptr == nullptr) {
520       // This will either forward-to-self, or detect that someone else has
521       // installed a forwarding pointer.
522       return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
523     }
524   }
525 
526   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
527   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
528 
529   // Should this evacuation fail?
530   if (inject_allocation_failure(from_region->hrm_index())) {
531     // Doing this after all the allocation attempts also tests the
532     // undo_allocation() method too.
533     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
534     return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
535   }
536 
537   // We're going to allocate linearly, so might as well prefetch ahead.
538   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
539   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
540 
541   const oop obj = cast_to_oop(obj_ptr);
542   // Because the forwarding is done with memory_order_relaxed there is no
543   // ordering with the above copy.  Clients that get the forwardee must not
544   // examine its contents without other synchronization, since the contents
545   // may not be up to date for them.
546   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
547   if (forward_ptr == nullptr) {
548 
549     {
550       const uint young_index = from_region->young_index_in_cset();
551       assert((from_region->is_young() && young_index >  0) ||
552              (!from_region->is_young() && young_index == 0), "invariant" );
553       _surviving_young_words[young_index] += word_sz;
554     }
555 
556     if (dest_attr.is_young()) {
557       if (age < markWord::max_age) {
558         age++;
559         obj->incr_age();
560       }
561       _age_table.add(age, word_sz);
562     } else {
563       update_bot_after_copying(obj, word_sz);
564     }
565 
566     {
567       // Skip the card enqueue iff the object (obj) is in survivor region.
568       // However, G1HeapRegion::is_survivor() is too expensive here.
569       // Instead, we use dest_attr.is_young() because the two values are always
570       // equal: successfully allocated young regions must be survivor regions.
571       assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
572       G1SkipCardMarkSetter x(&_scanner, dest_attr.is_young());
573       do_iterate_object(obj, old, klass, region_attr, dest_attr, age);
574     }
575 
576     return obj;
577   } else {
578     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
579     return forward_ptr;
580   }
581 }
582 
583 // Public not-inline entry point.
584 ATTRIBUTE_FLATTEN
585 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
586                                                  oop old,
587                                                  markWord old_mark) {
588   return do_copy_to_survivor_space(region_attr, old, old_mark);
589 }
590 
591 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
592   assert(worker_id < _num_workers, "out of bounds access");
593   if (_states[worker_id] == nullptr) {
594     _states[worker_id] =
595       new G1ParScanThreadState(_g1h,
596                                this,
597                                worker_id,
598                                _num_workers,
599                                _collection_set,
600                                _evac_failure_regions);
601   }
602   return _states[worker_id];
603 }
604 
605 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
606   assert(_flushed, "thread local state from the per thread states should have been flushed");
607   return _surviving_young_words_total;
608 }
609 
610 void G1ParScanThreadStateSet::flush_stats() {
611   assert(!_flushed, "thread local state from the per thread states should be flushed once");
612   for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
613     G1ParScanThreadState* pss = _states[worker_id];
614     assert(pss != nullptr, "must be initialized");
615 
616     G1GCPhaseTimes* p = _g1h->phase_times();
617 
618     // Need to get the following two before the call to G1ParThreadScanState::flush()
619     // because it resets the PLAB allocator where we get this info from.
620     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
621     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
622     size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers) * HeapWordSize;
623     size_t pending_cards = pss->num_cards_pending();
624     size_t to_young_gen_cards = pss->num_cards_marked() - pss->num_cards_pending();
625     size_t evac_failure_cards = pss->num_cards_from_evac_failure();
626     size_t marked_cards = pss->num_cards_marked();
627 
628     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, copied_bytes, G1GCPhaseTimes::FlushPSSCopiedBytes);
629     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::FlushPSSLABWasteBytes);
630     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::FlushPSSLABUndoWasteBytes);
631     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, pending_cards, G1GCPhaseTimes::FlushPSSPendingCards);
632     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, to_young_gen_cards, G1GCPhaseTimes::FlushPSSToYoungGenCards);
633     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, evac_failure_cards, G1GCPhaseTimes::FlushPSSEvacFail);
634     p->record_or_add_thread_work_item(G1GCPhaseTimes::FlushPSS, worker_id, marked_cards, G1GCPhaseTimes::FlushPSSMarked);
635   }
636 
637   _flushed = true;
638 }
639 
640 void G1ParScanThreadStateSet::destroy_worker_states() {
641   assert(_flushed, "statistics must already be flushed");
642   for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
643     delete _states[worker_id];
644     _states[worker_id] = nullptr;
645   }
646 }
647 
648 void G1ParScanThreadStateSet::update_nmethod_regions_to_add(G1NmethodsToAdd* nmethods) {
649   if (nmethods->number_of_entries() == 0) {
650     return;
651   }
652 
653   // Take the key set, look which are not yet in the global set, and update the necessary ones.
654   ResourceMark rm;
655   GrowableArray<uint> regions_to_add = GrowableArray<uint>(nmethods->table_size());
656 
657   nmethods->iterate_all([&] (uint& region, void*) {
658     if (_has_nmethods_to_add.par_set_bit(region, memory_order_relaxed)) {
659       regions_to_add.push(region);
660     }
661   });
662 
663   uint num_regions_to_add = (uint)regions_to_add.length();
664 
665   if (num_regions_to_add == 0) {
666     return;
667   }
668 
669   uint first_index = _num_nmethod_regions_to_add.fetch_then_add(num_regions_to_add, memory_order_relaxed);
670   guarantee(first_index + num_regions_to_add <= _g1h->max_num_regions(), "must be");
671 
672   memcpy(&_nmethod_regions_to_add[first_index], regions_to_add.adr_at(0), num_regions_to_add * sizeof(uint));
673 }
674 
675 void G1ParScanThreadStateSet::par_iterate_nmethod_regions_to_add(G1HeapRegionClosure* cl,
676                                                                  G1HeapRegionClaimer* claimer,
677                                                                  uint worker_id) {
678   _g1h->par_iterate_regions_array(cl, claimer, _nmethod_regions_to_add, num_nmethod_regions_to_add(), worker_id);
679 }
680 
681 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
682   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
683     G1ParScanThreadState* pss = _states[worker_index];
684     assert(pss != nullptr, "must be initialized");
685 
686     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
687     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
688   }
689 }
690 
691 void G1ParScanThreadState::record_evacuation_failed_region(G1HeapRegion* r, uint worker_id, bool cause_pinned) {
692   if (_evac_failure_regions->record(worker_id, r->hrm_index(), cause_pinned)) {
693     G1HeapRegionPrinter::evac_failure(r);
694   }
695 }
696 
697 NOINLINE
698 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, Klass* klass, G1HeapRegionAttr attr, size_t word_sz, bool cause_pinned) {
699   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
700 
701   oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
702   if (forward_ptr == nullptr) {
703     // Forward-to-self succeeded. We are the "owner" of the object.
704     G1HeapRegion* r = _g1h->heap_region_containing(old);
705 
706     record_evacuation_failed_region(r, _worker_id, cause_pinned);
707 
708     // Mark the failing object in the marking bitmap and later use the bitmap to handle
709     // evacuation failure recovery.
710     _g1h->mark_evac_failure_object(old);
711 
712     _evacuation_failed_info.register_copy_failure(word_sz);
713 
714     {
715       // For iterating objects that failed evacuation currently we can reuse the
716       // existing closure to scan evacuated objects; since we are iterating from a
717       // collection set region (i.e. never a Survivor region), we always need to
718       // gather cards for this case.
719       G1SkipCardMarkSetter x(&_scanner, false /* skip_card_mark */);
720       do_iterate_object(old, old, klass, attr, attr, m.age());
721     }
722 
723     return old;
724   } else {
725     // Forward-to-self failed. Either someone else managed to allocate
726     // space for this object (old != forward_ptr) or they beat us in
727     // self-forwarding it (old == forward_ptr).
728     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
729            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
730            "should not be in the CSet",
731            p2i(old), p2i(forward_ptr));
732     return forward_ptr;
733   }
734 }
735 
736 void G1ParScanThreadState::update_nmethod_regions_to_add() {
737   _per_thread_states->update_nmethod_regions_to_add(&_nmethods_to_add);
738 }
739 
740 void G1ParScanThreadState::initialize_numa_stats() {
741   if (_numa->is_enabled()) {
742     LogTarget(Info, gc, heap, numa) lt;
743 
744     if (lt.is_enabled()) {
745       uint num_nodes = _numa->num_active_nodes();
746       // Record only if there are multiple active nodes.
747       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
748       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
749     }
750   }
751 }
752 
753 void G1ParScanThreadState::flush_numa_stats() {
754   if (_obj_alloc_stat != nullptr) {
755     uint node_index = _numa->index_of_current_thread();
756     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
757   }
758 }
759 
760 void G1ParScanThreadState::update_numa_stats(uint node_index) {
761   if (_obj_alloc_stat != nullptr) {
762     _obj_alloc_stat[node_index]++;
763   }
764 }
765 
766 #if TASKQUEUE_STATS
767 
768 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
769   return _partial_array_splitter.stats();
770 }
771 
772 #endif // TASKQUEUE_STATS
773 
774 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
775                                                  uint num_workers,
776                                                  G1CollectionSet* collection_set,
777                                                  G1EvacFailureRegions* evac_failure_regions) :
778     _g1h(g1h),
779     _collection_set(collection_set),
780     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
781     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->num_young_regions() + 1, mtGC)),
782     _num_workers(num_workers),
783     _flushed(false),
784     _evac_failure_regions(evac_failure_regions),
785     _has_nmethods_to_add(g1h->max_num_regions(), mtGC),
786     _num_nmethod_regions_to_add(0),
787     _nmethod_regions_to_add(NEW_C_HEAP_ARRAY(uint, g1h->max_num_regions(), mtGC)) // Conservative length estimation.
788 {
789   for (uint i = 0; i < num_workers; ++i) {
790     _states[i] = nullptr;
791   }
792   memset(_surviving_young_words_total, 0, (collection_set->num_young_regions() + 1) * sizeof(size_t));
793 }
794 
795 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
796   for (uint i = 0; i < _num_workers; i++) {
797     assert(_states[i] == nullptr, "must be");
798   }
799   FREE_C_HEAP_ARRAY(_nmethod_regions_to_add);
800   FREE_C_HEAP_ARRAY(_states);
801   FREE_C_HEAP_ARRAY(_surviving_young_words_total);
802 }
803 
804 #if TASKQUEUE_STATS
805 
806 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
807   auto get_stats = [&](uint i) {
808     return state_for_worker(i)->partial_array_task_stats();
809   };
810   PartialArrayTaskStats::log_set(_num_workers, get_stats,
811                                  "Young GC Partial Array");
812 }
813 
814 #endif // TASKQUEUE_STATS