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->num_optional_regions()),
 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(oop from_obj,
242                                                   oop to_obj) {
243   assert(from_obj->is_forwarded(), "precondition");
244   assert(from_obj->forwardee() == to_obj, "precondition");
245   assert(to_obj->is_objArray(), "precondition");
246 
247   objArrayOop to_array = objArrayOop(to_obj);
248   size_t array_length = to_array->length();
249   size_t initial_chunk_size =
250     // The source array is unused when processing states.
251     _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length);
252 
253   assert(_scanner.skip_card_enqueue_set(), "must be");
254   // Process the initial chunk.  No need to process the type in the
255   // klass, as it will already be handled by processing the built-in
256   // module.
257   to_array->oop_iterate_range(&_scanner, 0, checked_cast<int>(initial_chunk_size));
258 }
259 
260 MAYBE_INLINE_EVACUATION
261 void G1ParScanThreadState::dispatch_task(ScannerTask task, bool stolen) {
262   verify_task(task);
263   if (task.is_narrow_oop_ptr()) {
264     do_oop_evac(task.to_narrow_oop_ptr());
265   } else if (task.is_oop_ptr()) {
266     do_oop_evac(task.to_oop_ptr());
267   } else {
268     do_partial_array(task.to_partial_array_state(), stolen);
269   }
270 }
271 
272 // Process tasks until overflow queue is empty and local queue
273 // contains no more than threshold entries.  NOINLINE to prevent
274 // inlining into steal_and_trim_queue.
275 ATTRIBUTE_FLATTEN NOINLINE
276 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
277   ScannerTask task;
278   do {
279     while (_task_queue->pop_overflow(task)) {
280       if (!_task_queue->try_push_to_taskqueue(task)) {
281         dispatch_task(task, false);
282       }
283     }
284     while (_task_queue->pop_local(task, threshold)) {
285       dispatch_task(task, false);
286     }
287   } while (!_task_queue->overflow_empty());
288 }
289 
290 ATTRIBUTE_FLATTEN
291 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
292   ScannerTask stolen_task;
293   while (task_queues->steal(_worker_id, stolen_task)) {
294     dispatch_task(stolen_task, true);
295     // Processing stolen task may have added tasks to our queue.
296     trim_queue();
297   }
298 }
299 
300 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
301                                                       size_t word_sz,
302                                                       bool previous_plab_refill_failed,
303                                                       uint node_index) {
304 
305   assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
306 
307   // Right now we only have two types of regions (young / old) so
308   // let's keep the logic here simple. We can generalize it when necessary.
309   if (dest->is_young()) {
310     bool plab_refill_in_old_failed = false;
311     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
312                                                         word_sz,
313                                                         &plab_refill_in_old_failed,
314                                                         node_index);
315     // Make sure that we won't attempt to copy any other objects out
316     // of a survivor region (given that apparently we cannot allocate
317     // any new ones) to avoid coming into this slow path again and again.
318     // Only consider failed PLAB refill here: failed inline allocations are
319     // typically large, so not indicative of remaining space.
320     if (previous_plab_refill_failed) {
321       _tenuring_threshold = 0;
322     }
323 
324     if (obj_ptr != nullptr) {
325       dest->set_old();
326     } else {
327       // We just failed to allocate in old gen. The same idea as explained above
328       // for making survivor gen unavailable for allocation applies for old gen.
329       _old_gen_is_full = plab_refill_in_old_failed;
330     }
331     return obj_ptr;
332   } else {
333     _old_gen_is_full = previous_plab_refill_failed;
334     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
335     // no other space to try.
336     return nullptr;
337   }
338 }
339 
340 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
341   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
342 
343   if (region_attr.is_young()) {
344     age = !m.has_displaced_mark_helper() ? m.age()
345                                          : m.displaced_mark_helper().age();
346     if (age < _tenuring_threshold) {
347       return region_attr;
348     }
349   }
350   // young-to-old (promotion) or old-to-old; destination is old in both cases.
351   return G1HeapRegionAttr::Old;
352 }
353 
354 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
355                                                   Klass* klass, size_t word_sz, uint age,
356                                                   HeapWord * const obj_ptr, uint node_index) const {
357   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
358   if (alloc_buf->contains(obj_ptr)) {
359     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
360                                                               dest_attr.type() == G1HeapRegionAttr::Old,
361                                                               alloc_buf->word_sz() * HeapWordSize);
362   } else {
363     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
364                                                                dest_attr.type() == G1HeapRegionAttr::Old);
365   }
366 }
367 
368 NOINLINE
369 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
370                                                    Klass* klass,
371                                                    size_t word_sz,
372                                                    uint age,
373                                                    uint node_index) {
374   HeapWord* obj_ptr = nullptr;
375   // Try slow-path allocation unless we're allocating old and old is already full.
376   if (!(dest_attr->is_old() && _old_gen_is_full)) {
377     bool plab_refill_failed = false;
378     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
379                                                            word_sz,
380                                                            &plab_refill_failed,
381                                                            node_index);
382     if (obj_ptr == nullptr) {
383       obj_ptr = allocate_in_next_plab(dest_attr,
384                                       word_sz,
385                                       plab_refill_failed,
386                                       node_index);
387     }
388   }
389   if (obj_ptr != nullptr) {
390     update_numa_stats(node_index);
391     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
392       // The events are checked individually as part of the actual commit
393       report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
394     }
395   }
396   return obj_ptr;
397 }
398 
399 #if ALLOCATION_FAILURE_INJECTOR
400 bool G1ParScanThreadState::inject_allocation_failure(uint region_idx) {
401   return _g1h->allocation_failure_injector()->allocation_should_fail(_allocation_failure_inject_counter, region_idx);
402 }
403 #endif
404 
405 NOINLINE
406 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
407                                            HeapWord* obj_ptr,
408                                            size_t word_sz,
409                                            uint node_index) {
410   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
411 }
412 
413 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
414   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
415   G1HeapRegion* region = _g1h->heap_region_containing(obj_start);
416   region->update_bot_for_block(obj_start, obj_start + word_sz);
417 }
418 
419 ALWAYSINLINE
420 void G1ParScanThreadState::do_iterate_object(oop const obj,
421                                              oop const old,
422                                              Klass* const klass,
423                                              G1HeapRegionAttr const region_attr,
424                                              G1HeapRegionAttr const dest_attr,
425                                              uint age) {
426     // Most objects are not arrays, so do one array check rather than
427     // checking for each array category for each object.
428     if (klass->is_array_klass()) {
429       assert(!klass->is_stack_chunk_instance_klass(), "must be");
430 
431       if (klass->is_objArray_klass()) {
432         start_partial_objarray(old, obj);
433       } else {
434         // Nothing needs to be done for typeArrays.  Body doesn't contain
435         // any oops to scan, and the type in the klass will already be handled
436         // by processing the built-in module.
437         assert(klass->is_typeArray_klass(), "invariant");
438       }
439       return;
440     }
441 
442     ContinuationGCSupport::transform_stack_chunk(obj);
443 
444     // Check for deduplicating young Strings.
445     if (G1StringDedup::is_candidate_from_evacuation(klass,
446                                                     region_attr,
447                                                     dest_attr,
448                                                     age)) {
449       // Record old; request adds a new weak reference, which reference
450       // processing expects to refer to a from-space object.
451       _string_dedup_requests.add(old);
452     }
453 
454     assert(_scanner.skip_card_enqueue_set(), "must be");
455     obj->oop_iterate_backwards(&_scanner, klass);
456 }
457 
458 // Private inline function, for direct internal use and providing the
459 // implementation of the public not-inline function.
460 MAYBE_INLINE_EVACUATION
461 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
462                                                     oop const old,
463                                                     markWord const old_mark) {
464   assert(region_attr.is_in_cset(),
465          "Unexpected region attr type: %s", region_attr.get_type_str());
466 
467   // NOTE: With compact headers, it is not safe to load the Klass* from old, because
468   // that would access the mark-word, that might change at any time by concurrent
469   // workers.
470   // This mark word would refer to a forwardee, which may not yet have completed
471   // copying. Therefore we must load the Klass* from the mark-word that we already
472   // loaded. This is safe, because we only enter here if not yet forwarded.
473   assert(!old_mark.is_forwarded(), "precondition");
474   Klass* klass = UseCompactObjectHeaders
475       ? old_mark.klass()
476       : old->klass();
477 
478   const size_t word_sz = old->size_given_klass(klass);
479 
480   // JNI only allows pinning of typeArrays, so we only need to keep those in place.
481   if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
482     return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, true /* cause_pinned */);
483   }
484 
485   uint age = 0;
486   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
487   G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
488   uint node_index = from_region->node_index();
489 
490   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
491 
492   // PLAB allocations should succeed most of the time, so we'll
493   // normally check against null once and that's it.
494   if (obj_ptr == nullptr) {
495     obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
496     if (obj_ptr == nullptr) {
497       // This will either forward-to-self, or detect that someone else has
498       // installed a forwarding pointer.
499       return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
500     }
501   }
502 
503   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
504   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
505 
506   // Should this evacuation fail?
507   if (inject_allocation_failure(from_region->hrm_index())) {
508     // Doing this after all the allocation attempts also tests the
509     // undo_allocation() method too.
510     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
511     return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
512   }
513 
514   // We're going to allocate linearly, so might as well prefetch ahead.
515   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
516   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
517 
518   const oop obj = cast_to_oop(obj_ptr);
519   // Because the forwarding is done with memory_order_relaxed there is no
520   // ordering with the above copy.  Clients that get the forwardee must not
521   // examine its contents without other synchronization, since the contents
522   // may not be up to date for them.
523   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
524   if (forward_ptr == nullptr) {
525 
526     {
527       const uint young_index = from_region->young_index_in_cset();
528       assert((from_region->is_young() && young_index >  0) ||
529              (!from_region->is_young() && young_index == 0), "invariant" );
530       _surviving_young_words[young_index] += word_sz;
531     }
532 
533     if (dest_attr.is_young()) {
534       if (age < markWord::max_age) {
535         age++;
536         obj->incr_age();
537       }
538       _age_table.add(age, word_sz);
539     } else {
540       update_bot_after_copying(obj, word_sz);
541     }
542 
543     {
544       // Skip the card enqueue iff the object (obj) is in survivor region.
545       // However, G1HeapRegion::is_survivor() is too expensive here.
546       // Instead, we use dest_attr.is_young() because the two values are always
547       // equal: successfully allocated young regions must be survivor regions.
548       assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
549       G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
550       do_iterate_object(obj, old, klass, region_attr, dest_attr, age);
551     }
552 
553     return obj;
554   } else {
555     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
556     return forward_ptr;
557   }
558 }
559 
560 // Public not-inline entry point.
561 ATTRIBUTE_FLATTEN
562 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
563                                                  oop old,
564                                                  markWord old_mark) {
565   return do_copy_to_survivor_space(region_attr, old, old_mark);
566 }
567 
568 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
569   assert(worker_id < _num_workers, "out of bounds access");
570   if (_states[worker_id] == nullptr) {
571     _states[worker_id] =
572       new G1ParScanThreadState(_g1h, rdcqs(),
573                                worker_id,
574                                _num_workers,
575                                _collection_set,
576                                _evac_failure_regions);
577   }
578   return _states[worker_id];
579 }
580 
581 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
582   assert(_flushed, "thread local state from the per thread states should have been flushed");
583   return _surviving_young_words_total;
584 }
585 
586 void G1ParScanThreadStateSet::flush_stats() {
587   assert(!_flushed, "thread local state from the per thread states should be flushed once");
588   for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
589     G1ParScanThreadState* pss = _states[worker_id];
590     assert(pss != nullptr, "must be initialized");
591 
592     G1GCPhaseTimes* p = _g1h->phase_times();
593 
594     // Need to get the following two before the call to G1ParThreadScanState::flush()
595     // because it resets the PLAB allocator where we get this info from.
596     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
597     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
598     size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers, &_rdc_buffers[worker_id]) * HeapWordSize;
599     size_t evac_fail_enqueued_cards = pss->evac_failure_enqueued_cards();
600 
601     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
602     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
603     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
604     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_fail_enqueued_cards, G1GCPhaseTimes::MergePSSEvacFailExtra);
605 
606     delete pss;
607     _states[worker_id] = nullptr;
608   }
609 
610   G1DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
611   dcq.merge_bufferlists(rdcqs());
612   rdcqs()->verify_empty();
613 
614   _flushed = true;
615 }
616 
617 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* 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 void G1ParScanThreadState::record_evacuation_failed_region(G1HeapRegion* r, uint worker_id, bool cause_pinned) {
628   if (_evac_failure_regions->record(worker_id, r->hrm_index(), cause_pinned)) {
629     G1HeapRegionPrinter::evac_failure(r);
630   }
631 }
632 
633 NOINLINE
634 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, Klass* klass, G1HeapRegionAttr attr, size_t word_sz, bool cause_pinned) {
635   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
636 
637   oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
638   if (forward_ptr == nullptr) {
639     // Forward-to-self succeeded. We are the "owner" of the object.
640     G1HeapRegion* r = _g1h->heap_region_containing(old);
641 
642     record_evacuation_failed_region(r, _worker_id, cause_pinned);
643 
644     // Mark the failing object in the marking bitmap and later use the bitmap to handle
645     // evacuation failure recovery.
646     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
647 
648     _evacuation_failed_info.register_copy_failure(word_sz);
649 
650     {
651       // For iterating objects that failed evacuation currently we can reuse the
652       // existing closure to scan evacuated objects; since we are iterating from a
653       // collection set region (i.e. never a Survivor region), we always need to
654       // gather cards for this case.
655       G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
656       do_iterate_object(old, old, klass, attr, attr, m.age());
657     }
658 
659     return old;
660   } else {
661     // Forward-to-self failed. Either someone else managed to allocate
662     // space for this object (old != forward_ptr) or they beat us in
663     // self-forwarding it (old == forward_ptr).
664     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
665            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
666            "should not be in the CSet",
667            p2i(old), p2i(forward_ptr));
668     return forward_ptr;
669   }
670 }
671 
672 void G1ParScanThreadState::initialize_numa_stats() {
673   if (_numa->is_enabled()) {
674     LogTarget(Info, gc, heap, numa) lt;
675 
676     if (lt.is_enabled()) {
677       uint num_nodes = _numa->num_active_nodes();
678       // Record only if there are multiple active nodes.
679       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
680       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
681     }
682   }
683 }
684 
685 void G1ParScanThreadState::flush_numa_stats() {
686   if (_obj_alloc_stat != nullptr) {
687     uint node_index = _numa->index_of_current_thread();
688     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
689   }
690 }
691 
692 void G1ParScanThreadState::update_numa_stats(uint node_index) {
693   if (_obj_alloc_stat != nullptr) {
694     _obj_alloc_stat[node_index]++;
695   }
696 }
697 
698 #if TASKQUEUE_STATS
699 
700 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
701   return _partial_array_splitter.stats();
702 }
703 
704 #endif // TASKQUEUE_STATS
705 
706 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
707                                                  uint num_workers,
708                                                  G1CollectionSet* collection_set,
709                                                  G1EvacFailureRegions* evac_failure_regions) :
710     _g1h(g1h),
711     _collection_set(collection_set),
712     _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
713     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
714     _rdc_buffers(NEW_C_HEAP_ARRAY(BufferNodeList, num_workers, mtGC)),
715     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
716     _num_workers(num_workers),
717     _flushed(false),
718     _evac_failure_regions(evac_failure_regions)
719 {
720   for (uint i = 0; i < num_workers; ++i) {
721     _states[i] = nullptr;
722     _rdc_buffers[i] = BufferNodeList();
723   }
724   memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
725 }
726 
727 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
728   assert(_flushed, "thread local state from the per thread states should have been flushed");
729   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
730   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
731   FREE_C_HEAP_ARRAY(BufferNodeList, _rdc_buffers);
732 }
733 
734 #if TASKQUEUE_STATS
735 
736 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
737   auto get_stats = [&](uint i) {
738     return state_for_worker(i)->partial_array_task_stats();
739   };
740   PartialArrayTaskStats::log_set(_num_workers, get_stats,
741                                  "Partial Array Task Stats");
742 }
743 
744 #endif // TASKQUEUE_STATS