1 /*
  2  * Copyright (c) 2014, 2025, 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/atomic.hpp"
 46 #include "runtime/mutexLocker.hpp"
 47 #include "runtime/prefetch.inline.hpp"
 48 #include "utilities/globalDefinitions.hpp"
 49 #include "utilities/macros.hpp"
 50 
 51 // In fastdebug builds the code size can get out of hand, potentially
 52 // tripping over compiler limits (which may be bugs, but nevertheless
 53 // need to be taken into consideration).  A side benefit of limiting
 54 // inlining is that we get more call frames that might aid debugging.
 55 // And the fastdebug compile time for this file is much reduced.
 56 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
 57 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
 58 
 59 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
 60                                            G1RedirtyCardsQueueSet* rdcqs,
 61                                            uint worker_id,
 62                                            uint num_workers,
 63                                            G1CollectionSet* collection_set,
 64                                            G1EvacFailureRegions* evac_failure_regions)
 65   : _g1h(g1h),
 66     _task_queue(g1h->task_queue(worker_id)),
 67     _rdc_local_qset(rdcqs),
 68     _ct(g1h->card_table()),
 69     _closures(nullptr),
 70     _plab_allocator(nullptr),
 71     _age_table(false),
 72     _tenuring_threshold(g1h->policy()->tenuring_threshold()),
 73     _scanner(g1h, this),
 74     _worker_id(worker_id),
 75     _last_enqueued_card(SIZE_MAX),
 76     _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
 77     _stack_trim_lower_threshold(GCDrainStackTargetSize),
 78     _trim_ticks(),
 79     _surviving_young_words_base(nullptr),
 80     _surviving_young_words(nullptr),
 81     _surviving_words_length(collection_set->young_region_length() + 1),
 82     _old_gen_is_full(false),
 83     _partial_array_splitter(g1h->partial_array_state_manager(), 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     ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
 89     _evacuation_failed_info(),
 90     _evac_failure_regions(evac_failure_regions),
 91     _evac_failure_enqueued_cards(0)
 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_PADDING_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(_g1h,
107                                                              this,
108                                                              collection_set->only_contains_young_regions());
109 
110   _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
111 
112   initialize_numa_stats();
113 }
114 
115 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers, BufferNodeList* rdc_buffers) {
116   *rdc_buffers = _rdc_local_qset.flush();
117   flush_numa_stats();
118   // Update allocation statistics.
119   _plab_allocator->flush_and_retire_stats(num_workers);
120   _g1h->policy()->record_age_table(&_age_table);
121 
122   if (_evacuation_failed_info.has_failed()) {
123     _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
124   }
125 
126   size_t sum = 0;
127   for (uint i = 0; i < _surviving_words_length; i++) {
128     surviving_young_words[i] += _surviving_young_words[i];
129     sum += _surviving_young_words[i];
130   }
131   return sum;
132 }
133 
134 G1ParScanThreadState::~G1ParScanThreadState() {
135   delete _plab_allocator;
136   delete _closures;
137   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
138   delete[] _oops_into_optional_regions;
139   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
140 }
141 
142 size_t G1ParScanThreadState::lab_waste_words() const {
143   return _plab_allocator->waste();
144 }
145 
146 size_t G1ParScanThreadState::lab_undo_waste_words() const {
147   return _plab_allocator->undo_waste();
148 }
149 
150 size_t G1ParScanThreadState::evac_failure_enqueued_cards() const {
151   return _evac_failure_enqueued_cards;
152 }
153 
154 #ifdef ASSERT
155 void G1ParScanThreadState::verify_task(narrowOop* task) const {
156   assert(task != nullptr, "invariant");
157   assert(UseCompressedOops, "sanity");
158   oop p = RawAccess<>::oop_load(task);
159   assert(_g1h->is_in_reserved(p),
160          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
161 }
162 
163 void G1ParScanThreadState::verify_task(oop* task) const {
164   assert(task != nullptr, "invariant");
165   oop p = RawAccess<>::oop_load(task);
166   assert(_g1h->is_in_reserved(p),
167          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
168 }
169 
170 void G1ParScanThreadState::verify_task(PartialArrayState* task) const {
171   assert(task != nullptr, "invariant");
172   // Source isn't used for processing, so not recorded in task.
173   assert(task->source() == nullptr, "invariant");
174   oop p = task->destination();
175   assert(_g1h->is_in_reserved(p),
176          "task=" PTR_FORMAT " dest=" PTR_FORMAT, p2i(task), p2i(p));
177 }
178 
179 void G1ParScanThreadState::verify_task(ScannerTask task) const {
180   if (task.is_narrow_oop_ptr()) {
181     verify_task(task.to_narrow_oop_ptr());
182   } else if (task.is_oop_ptr()) {
183     verify_task(task.to_oop_ptr());
184   } else if (task.is_partial_array_state()) {
185     verify_task(task.to_partial_array_state());
186   } else {
187     ShouldNotReachHere();
188   }
189 }
190 #endif // ASSERT
191 
192 template <class T>
193 MAYBE_INLINE_EVACUATION
194 void G1ParScanThreadState::do_oop_evac(T* p) {
195   // Reference should not be null here as such are never pushed to the task queue.
196   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
197 
198   // Although we never intentionally push references outside of the collection
199   // set, due to (benign) races in the claim mechanism during RSet scanning more
200   // than one thread might claim the same card. So the same card may be
201   // processed multiple times, and so we might get references into old gen here.
202   // So we need to redo this check.
203   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
204   // References pushed onto the work stack should never point to a humongous region
205   // as they are not added to the collection set due to above precondition.
206   assert(!region_attr.is_humongous_candidate(),
207          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
208          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
209 
210   if (!region_attr.is_in_cset()) {
211     // In this case somebody else already did all the work.
212     return;
213   }
214 
215   markWord m = obj->mark();
216   if (m.is_forwarded()) {
217     obj = obj->forwardee(m);
218   } else {
219     obj = do_copy_to_survivor_space(region_attr, obj, m);
220   }
221   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
222 
223   write_ref_field_post(p, obj);
224 }
225 
226 MAYBE_INLINE_EVACUATION
227 void G1ParScanThreadState::do_partial_array(PartialArrayState* state, bool stolen) {
228   // Access state before release by claim().
229   objArrayOop to_array = objArrayOop(state->destination());
230   PartialArraySplitter::Claim claim =
231     _partial_array_splitter.claim(state, _task_queue, stolen);
232   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
233   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
234   // Process claimed task.
235   to_array->oop_iterate_range(&_scanner,
236                               checked_cast<int>(claim._start),
237                               checked_cast<int>(claim._end));
238 }
239 
240 MAYBE_INLINE_EVACUATION
241 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
242                                                   oop from_obj,
243                                                   oop to_obj) {
244   assert(from_obj->is_forwarded(), "precondition");
245   assert(from_obj->forwardee() == to_obj, "precondition");
246   assert(to_obj->is_objArray(), "precondition");
247 
248   objArrayOop to_array = objArrayOop(to_obj);
249   size_t array_length = to_array->length();
250   size_t initial_chunk_size =
251     // The source array is unused when processing states.
252     _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length);
253 
254   // Skip the card enqueue iff the object (to_array) is in survivor region.
255   // However, G1HeapRegion::is_survivor() is too expensive here.
256   // Instead, we use dest_attr.is_young() because the two values are always
257   // equal: successfully allocated young regions must be survivor regions.
258   assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
259   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
260   // Process the initial chunk.  No need to process the type in the
261   // klass, as it will already be handled by processing the built-in
262   // module.
263   to_array->oop_iterate_range(&_scanner, 0, checked_cast<int>(initial_chunk_size));
264 }
265 
266 MAYBE_INLINE_EVACUATION
267 void G1ParScanThreadState::dispatch_task(ScannerTask task, bool stolen) {
268   verify_task(task);
269   if (task.is_narrow_oop_ptr()) {
270     do_oop_evac(task.to_narrow_oop_ptr());
271   } else if (task.is_oop_ptr()) {
272     do_oop_evac(task.to_oop_ptr());
273   } else {
274     do_partial_array(task.to_partial_array_state(), stolen);
275   }
276 }
277 
278 // Process tasks until overflow queue is empty and local queue
279 // contains no more than threshold entries.  NOINLINE to prevent
280 // inlining into steal_and_trim_queue.
281 ATTRIBUTE_FLATTEN NOINLINE
282 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
283   ScannerTask task;
284   do {
285     while (_task_queue->pop_overflow(task)) {
286       if (!_task_queue->try_push_to_taskqueue(task)) {
287         dispatch_task(task, false);
288       }
289     }
290     while (_task_queue->pop_local(task, threshold)) {
291       dispatch_task(task, false);
292     }
293   } while (!_task_queue->overflow_empty());
294 }
295 
296 ATTRIBUTE_FLATTEN
297 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
298   ScannerTask stolen_task;
299   while (task_queues->steal(_worker_id, stolen_task)) {
300     dispatch_task(stolen_task, true);
301     // Processing stolen task may have added tasks to our queue.
302     trim_queue();
303   }
304 }
305 
306 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
307                                                       size_t word_sz,
308                                                       bool previous_plab_refill_failed,
309                                                       uint node_index) {
310 
311   assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
312 
313   // Right now we only have two types of regions (young / old) so
314   // let's keep the logic here simple. We can generalize it when necessary.
315   if (dest->is_young()) {
316     bool plab_refill_in_old_failed = false;
317     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
318                                                         word_sz,
319                                                         &plab_refill_in_old_failed,
320                                                         node_index);
321     // Make sure that we won't attempt to copy any other objects out
322     // of a survivor region (given that apparently we cannot allocate
323     // any new ones) to avoid coming into this slow path again and again.
324     // Only consider failed PLAB refill here: failed inline allocations are
325     // typically large, so not indicative of remaining space.
326     if (previous_plab_refill_failed) {
327       _tenuring_threshold = 0;
328     }
329 
330     if (obj_ptr != nullptr) {
331       dest->set_old();
332     } else {
333       // We just failed to allocate in old gen. The same idea as explained above
334       // for making survivor gen unavailable for allocation applies for old gen.
335       _old_gen_is_full = plab_refill_in_old_failed;
336     }
337     return obj_ptr;
338   } else {
339     _old_gen_is_full = previous_plab_refill_failed;
340     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
341     // no other space to try.
342     return nullptr;
343   }
344 }
345 
346 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
347   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
348 
349   if (region_attr.is_young()) {
350     age = !m.has_displaced_mark_helper() ? m.age()
351                                          : m.displaced_mark_helper().age();
352     if (age < _tenuring_threshold) {
353       return region_attr;
354     }
355   }
356   // young-to-old (promotion) or old-to-old; destination is old in both cases.
357   return G1HeapRegionAttr::Old;
358 }
359 
360 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
361                                                   Klass* klass, size_t word_sz, uint age,
362                                                   HeapWord * const obj_ptr, uint node_index) const {
363   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
364   if (alloc_buf->contains(obj_ptr)) {
365     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
366                                                               dest_attr.type() == G1HeapRegionAttr::Old,
367                                                               alloc_buf->word_sz() * HeapWordSize);
368   } else {
369     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
370                                                                dest_attr.type() == G1HeapRegionAttr::Old);
371   }
372 }
373 
374 NOINLINE
375 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
376                                                    Klass* klass,
377                                                    size_t word_sz,
378                                                    uint age,
379                                                    uint node_index) {
380   HeapWord* obj_ptr = nullptr;
381   // Try slow-path allocation unless we're allocating old and old is already full.
382   if (!(dest_attr->is_old() && _old_gen_is_full)) {
383     bool plab_refill_failed = false;
384     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
385                                                            word_sz,
386                                                            &plab_refill_failed,
387                                                            node_index);
388     if (obj_ptr == nullptr) {
389       obj_ptr = allocate_in_next_plab(dest_attr,
390                                       word_sz,
391                                       plab_refill_failed,
392                                       node_index);
393     }
394   }
395   if (obj_ptr != nullptr) {
396     update_numa_stats(node_index);
397     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
398       // The events are checked individually as part of the actual commit
399       report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
400     }
401   }
402   return obj_ptr;
403 }
404 
405 #if ALLOCATION_FAILURE_INJECTOR
406 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
407   return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
408 }
409 #endif
410 
411 NOINLINE
412 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
413                                            HeapWord* obj_ptr,
414                                            size_t word_sz,
415                                            uint node_index) {
416   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
417 }
418 
419 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
420   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
421   G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
422   region->update_bot_for_block(obj_start, obj_start + word_sz);
423 }
424 
425 // Private inline function, for direct internal use and providing the
426 // implementation of the public not-inline function.
427 MAYBE_INLINE_EVACUATION
428 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
429                                                     oop const old,
430                                                     markWord const old_mark) {
431   assert(region_attr.is_in_cset(),
432          "Unexpected region attr type: %s", region_attr.get_type_str());
433 
434   // NOTE: With compact headers, it is not safe to load the Klass* from old, because
435   // that would access the mark-word, that might change at any time by concurrent
436   // workers.
437   // This mark word would refer to a forwardee, which may not yet have completed
438   // copying. Therefore we must load the Klass* from the mark-word that we already
439   // loaded. This is safe, because we only enter here if not yet forwarded.
440   assert(!old_mark.is_forwarded(), "precondition");
441   Klass* klass = UseCompactObjectHeaders
442       ? old_mark.klass()
443       : old->klass();
444 
445   const size_t word_sz = old->size_given_klass(klass);
446 
447   // JNI only allows pinning of typeArrays, so we only need to keep those in place.
448   if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
449     return handle_evacuation_failure_par(old, old_mark, word_sz, true /* cause_pinned */);
450   }
451 
452   uint age = 0;
453   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
454   G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
455   uint node_index = from_region->node_index();
456 
457   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
458 
459   // PLAB allocations should succeed most of the time, so we'll
460   // normally check against null once and that's it.
461   if (obj_ptr == nullptr) {
462     obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
463     if (obj_ptr == nullptr) {
464       // This will either forward-to-self, or detect that someone else has
465       // installed a forwarding pointer.
466       return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
467     }
468   }
469 
470   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
471   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
472 
473   // Should this evacuation fail?
474   if (inject_allocation_failure(from_region->hrm_index())) {
475     // Doing this after all the allocation attempts also tests the
476     // undo_allocation() method too.
477     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
478     return handle_evacuation_failure_par(old, old_mark, word_sz, false /* cause_pinned */);
479   }
480 
481   // We're going to allocate linearly, so might as well prefetch ahead.
482   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
483   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
484 
485   const oop obj = cast_to_oop(obj_ptr);
486   // Because the forwarding is done with memory_order_relaxed there is no
487   // ordering with the above copy.  Clients that get the forwardee must not
488   // examine its contents without other synchronization, since the contents
489   // may not be up to date for them.
490   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
491   if (forward_ptr == nullptr) {
492 
493     {
494       const uint young_index = from_region->young_index_in_cset();
495       assert((from_region->is_young() && young_index >  0) ||
496              (!from_region->is_young() && young_index == 0), "invariant" );
497       _surviving_young_words[young_index] += word_sz;
498     }
499 
500     if (dest_attr.is_young()) {
501       if (age < markWord::max_age) {
502         age++;
503         obj->incr_age();
504       }
505       _age_table.add(age, word_sz);
506     } else {
507       update_bot_after_copying(obj, word_sz);
508     }
509 
510     // Most objects are not arrays, so do one array check rather than
511     // checking for each array category for each object.
512     if (klass->is_array_klass()) {
513       if (klass->is_objArray_klass()) {
514         start_partial_objarray(dest_attr, old, obj);
515       } else {
516         // Nothing needs to be done for typeArrays.  Body doesn't contain
517         // any oops to scan, and the type in the klass will already be handled
518         // by processing the built-in module.
519         assert(klass->is_typeArray_klass(), "invariant");
520       }
521       return obj;
522     }
523 
524     ContinuationGCSupport::transform_stack_chunk(obj);
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, G1HeapRegion::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 < _num_workers, "out of bounds access");
560   if (_states[worker_id] == nullptr) {
561     _states[worker_id] =
562       new G1ParScanThreadState(_g1h, rdcqs(),
563                                worker_id,
564                                _num_workers,
565                                _collection_set,
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_stats() {
577   assert(!_flushed, "thread local state from the per thread states should be flushed once");
578   for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
579     G1ParScanThreadState* pss = _states[worker_id];
580     assert(pss != nullptr, "must be initialized");
581 
582     G1GCPhaseTimes* p = _g1h->phase_times();
583 
584     // Need to get the following two before the call to G1ParThreadScanState::flush()
585     // because it resets the PLAB allocator where we get this info from.
586     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
587     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
588     size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers, &_rdc_buffers[worker_id]) * HeapWordSize;
589     size_t evac_fail_enqueued_cards = pss->evac_failure_enqueued_cards();
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     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_fail_enqueued_cards, G1GCPhaseTimes::MergePSSEvacFailExtra);
595 
596     delete pss;
597     _states[worker_id] = nullptr;
598   }
599 
600   G1DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
601   dcq.merge_bufferlists(rdcqs());
602   rdcqs()->verify_empty();
603 
604   _flushed = true;
605 }
606 
607 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
608   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
609     G1ParScanThreadState* pss = _states[worker_index];
610     assert(pss != nullptr, "must be initialized");
611 
612     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
613     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
614   }
615 }
616 
617 NOINLINE
618 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz, bool cause_pinned) {
619   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
620 
621   oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
622   if (forward_ptr == nullptr) {
623     // Forward-to-self succeeded. We are the "owner" of the object.
624     G1HeapRegion* r = _g1h->heap_region_containing(old);
625 
626     if (_evac_failure_regions->record(_worker_id, r->hrm_index(), cause_pinned)) {
627       G1HeapRegionPrinter::evac_failure(r);
628     }
629 
630     // Mark the failing object in the marking bitmap and later use the bitmap to handle
631     // evacuation failure recovery.
632     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
633 
634     ContinuationGCSupport::transform_stack_chunk(old);
635 
636     _evacuation_failed_info.register_copy_failure(word_sz);
637 
638     // For iterating objects that failed evacuation currently we can reuse the
639     // existing closure to scan evacuated objects; since we are iterating from a
640     // collection set region (i.e. never a Survivor region), we always need to
641     // gather cards for this case.
642     G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
643     old->oop_iterate_backwards(&_scanner);
644 
645     return old;
646   } else {
647     // Forward-to-self failed. Either someone else managed to allocate
648     // space for this object (old != forward_ptr) or they beat us in
649     // self-forwarding it (old == forward_ptr).
650     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
651            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
652            "should not be in the CSet",
653            p2i(old), p2i(forward_ptr));
654     return forward_ptr;
655   }
656 }
657 
658 void G1ParScanThreadState::initialize_numa_stats() {
659   if (_numa->is_enabled()) {
660     LogTarget(Info, gc, heap, numa) lt;
661 
662     if (lt.is_enabled()) {
663       uint num_nodes = _numa->num_active_nodes();
664       // Record only if there are multiple active nodes.
665       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
666       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
667     }
668   }
669 }
670 
671 void G1ParScanThreadState::flush_numa_stats() {
672   if (_obj_alloc_stat != nullptr) {
673     uint node_index = _numa->index_of_current_thread();
674     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
675   }
676 }
677 
678 void G1ParScanThreadState::update_numa_stats(uint node_index) {
679   if (_obj_alloc_stat != nullptr) {
680     _obj_alloc_stat[node_index]++;
681   }
682 }
683 
684 #if TASKQUEUE_STATS
685 
686 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
687   return _partial_array_splitter.stats();
688 }
689 
690 #endif // TASKQUEUE_STATS
691 
692 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
693                                                  uint num_workers,
694                                                  G1CollectionSet* collection_set,
695                                                  G1EvacFailureRegions* evac_failure_regions) :
696     _g1h(g1h),
697     _collection_set(collection_set),
698     _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
699     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
700     _rdc_buffers(NEW_C_HEAP_ARRAY(BufferNodeList, num_workers, mtGC)),
701     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
702     _num_workers(num_workers),
703     _flushed(false),
704     _evac_failure_regions(evac_failure_regions)
705 {
706   for (uint i = 0; i < num_workers; ++i) {
707     _states[i] = nullptr;
708     _rdc_buffers[i] = BufferNodeList();
709   }
710   memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
711 }
712 
713 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
714   assert(_flushed, "thread local state from the per thread states should have been flushed");
715   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
716   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
717   FREE_C_HEAP_ARRAY(BufferNodeList, _rdc_buffers);
718 }
719 
720 #if TASKQUEUE_STATS
721 
722 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
723   auto get_stats = [&](uint i) {
724     return state_for_worker(i)->partial_array_task_stats();
725   };
726   PartialArrayTaskStats::log_set(_num_workers, get_stats,
727                                  "Partial Array Task Stats");
728 }
729 
730 #endif // TASKQUEUE_STATS