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/atomicAccess.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                                            uint worker_id,
 61                                            uint num_workers,
 62                                            G1CollectionSet* collection_set,
 63                                            G1EvacFailureRegions* evac_failure_regions)
 64   : _g1h(g1h),
 65     _task_queue(g1h->task_queue(worker_id)),
 66     _ct(g1h->refinement_table()),
 67     _closures(nullptr),
 68     _plab_allocator(nullptr),
 69     _age_table(false),
 70     _tenuring_threshold(g1h->policy()->tenuring_threshold()),
 71     _scanner(g1h, this),
 72     _worker_id(worker_id),
 73     _num_cards_marked_dirty(0),
 74     _num_cards_marked_to_cset(0),
 75     _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
 76     _stack_trim_lower_threshold(GCDrainStackTargetSize),
 77     _trim_ticks(),
 78     _surviving_young_words_base(nullptr),
 79     _surviving_young_words(nullptr),
 80     _surviving_words_length(collection_set->young_region_length() + 1),
 81     _old_gen_is_full(false),
 82     _partial_array_splitter(g1h->partial_array_state_manager(), num_workers, ParGCArrayScanChunk),
 83     _string_dedup_requests(),
 84     _max_num_optional_regions(collection_set->num_optional_regions()),
 85     _numa(g1h->numa()),
 86     _obj_alloc_stat(nullptr),
 87     ALLOCATION_FAILURE_INJECTOR_ONLY(_allocation_failure_inject_counter(0) COMMA)
 88     _evacuation_failed_info(),
 89     _evac_failure_regions(evac_failure_regions),
 90     _num_cards_from_evac_failure(0)
 91 {
 92   // We allocate number of young gen regions in the collection set plus one
 93   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
 94   // We also add a few elements at the beginning and at the end in
 95   // an attempt to eliminate cache contention
 96   const size_t padding_elem_num = (DEFAULT_PADDING_SIZE / sizeof(size_t));
 97   size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
 98 
 99   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
100   _surviving_young_words = _surviving_young_words_base + padding_elem_num;
101   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
102 
103   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
104 
105   _closures = G1EvacuationRootClosures::create_root_closures(_g1h,
106                                                              this,
107                                                              collection_set->only_contains_young_regions());
108 
109   _oops_into_optional_regions = new G1OopStarChunkedList[_max_num_optional_regions];
110 
111   initialize_numa_stats();
112 }
113 
114 size_t G1ParScanThreadState::flush_stats(size_t* surviving_young_words, uint num_workers) {
115   flush_numa_stats();
116   // Update allocation statistics.
117   _plab_allocator->flush_and_retire_stats(num_workers);
118   _g1h->policy()->record_age_table(&_age_table);
119 
120   if (_evacuation_failed_info.has_failed()) {
121     _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
122   }
123 
124   size_t sum = 0;
125   for (uint i = 0; i < _surviving_words_length; i++) {
126     surviving_young_words[i] += _surviving_young_words[i];
127     sum += _surviving_young_words[i];
128   }
129   return sum;
130 }
131 
132 G1ParScanThreadState::~G1ParScanThreadState() {
133   delete _plab_allocator;
134   delete _closures;
135   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
136   delete[] _oops_into_optional_regions;
137   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
138 }
139 
140 size_t G1ParScanThreadState::lab_waste_words() const {
141   return _plab_allocator->waste();
142 }
143 
144 size_t G1ParScanThreadState::lab_undo_waste_words() const {
145   return _plab_allocator->undo_waste();
146 }
147 
148 size_t G1ParScanThreadState::num_cards_pending() const {
149   return _num_cards_marked_dirty + _num_cards_from_evac_failure;
150 }
151 
152 size_t G1ParScanThreadState::num_cards_marked() const {
153   return num_cards_pending() + _num_cards_marked_to_cset;
154 }
155 
156 size_t G1ParScanThreadState::num_cards_from_evac_failure() const {
157   return _num_cards_from_evac_failure;
158 }
159 
160 #ifdef ASSERT
161 void G1ParScanThreadState::verify_task(narrowOop* task) const {
162   assert(task != nullptr, "invariant");
163   assert(UseCompressedOops, "sanity");
164   oop p = RawAccess<>::oop_load(task);
165   assert(_g1h->is_in_reserved(p),
166          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
167 }
168 
169 void G1ParScanThreadState::verify_task(oop* task) const {
170   assert(task != nullptr, "invariant");
171   oop p = RawAccess<>::oop_load(task);
172   assert(_g1h->is_in_reserved(p),
173          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
174 }
175 
176 void G1ParScanThreadState::verify_task(PartialArrayState* task) const {
177   assert(task != nullptr, "invariant");
178   // Source isn't used for processing, so not recorded in task.
179   assert(task->source() == nullptr, "invariant");
180   oop p = task->destination();
181   assert(_g1h->is_in_reserved(p),
182          "task=" PTR_FORMAT " dest=" PTR_FORMAT, p2i(task), p2i(p));
183 }
184 
185 void G1ParScanThreadState::verify_task(ScannerTask task) const {
186   if (task.is_narrow_oop_ptr()) {
187     verify_task(task.to_narrow_oop_ptr());
188   } else if (task.is_oop_ptr()) {
189     verify_task(task.to_oop_ptr());
190   } else if (task.is_partial_array_state()) {
191     verify_task(task.to_partial_array_state());
192   } else {
193     ShouldNotReachHere();
194   }
195 }
196 #endif // ASSERT
197 
198 template <class T>
199 MAYBE_INLINE_EVACUATION
200 void G1ParScanThreadState::do_oop_evac(T* p) {
201   // Reference should not be null here as such are never pushed to the task queue.
202   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
203 
204   // Although we never intentionally push references outside of the collection
205   // set, due to (benign) races in the claim mechanism during RSet scanning more
206   // than one thread might claim the same card. So the same card may be
207   // processed multiple times, and so we might get references into old gen here.
208   // So we need to redo this check.
209   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
210   // References pushed onto the work stack should never point to a humongous region
211   // as they are not added to the collection set due to above precondition.
212   assert(!region_attr.is_humongous_candidate(),
213          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
214          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
215 
216   if (!region_attr.is_in_cset()) {
217     // In this case somebody else already did all the work.
218     return;
219   }
220 
221   markWord m = obj->mark();
222   if (m.is_forwarded()) {
223     obj = obj->forwardee(m);
224   } else {
225     obj = do_copy_to_survivor_space(region_attr, obj, m);
226   }
227   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
228 
229   write_ref_field_post(p, obj);
230 }
231 
232 MAYBE_INLINE_EVACUATION
233 void G1ParScanThreadState::do_partial_array(PartialArrayState* state, bool stolen) {
234   // Access state before release by claim().
235   objArrayOop to_array = objArrayOop(state->destination());
236   PartialArraySplitter::Claim claim =
237     _partial_array_splitter.claim(state, _task_queue, stolen);
238   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
239   G1SkipCardMarkSetter x(&_scanner, dest_attr.is_new_survivor());
240   // Process claimed task.
241   to_array->oop_iterate_range(&_scanner,
242                               checked_cast<int>(claim._start),
243                               checked_cast<int>(claim._end));
244 }
245 
246 MAYBE_INLINE_EVACUATION
247 void G1ParScanThreadState::start_partial_objarray(oop from_obj,
248                                                   oop to_obj) {
249   assert(from_obj->is_forwarded(), "precondition");
250   assert(from_obj->forwardee() == to_obj, "precondition");
251   assert(to_obj->is_objArray(), "precondition");
252 
253   objArrayOop to_array = objArrayOop(to_obj);
254   size_t array_length = to_array->length();
255   size_t initial_chunk_size =
256     // The source array is unused when processing states.
257     _partial_array_splitter.start(_task_queue, nullptr, to_array, array_length);
258 
259   assert(_scanner.skip_card_mark_set(), "must be");
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 ALWAYSINLINE
426 void G1ParScanThreadState::do_iterate_object(oop const obj,
427                                              oop const old,
428                                              Klass* const klass,
429                                              G1HeapRegionAttr const region_attr,
430                                              G1HeapRegionAttr const dest_attr,
431                                              uint age) {
432     // Most objects are not arrays, so do one array check rather than
433     // checking for each array category for each object.
434     if (klass->is_array_klass()) {
435       assert(!klass->is_stack_chunk_instance_klass(), "must be");
436 
437       if (klass->is_objArray_klass()) {
438         start_partial_objarray(old, obj);
439       } else {
440         // Nothing needs to be done for typeArrays.  Body doesn't contain
441         // any oops to scan, and the type in the klass will already be handled
442         // by processing the built-in module.
443         assert(klass->is_typeArray_klass(), "invariant");
444       }
445       return;
446     }
447 
448     ContinuationGCSupport::transform_stack_chunk(obj);
449 
450     // Check for deduplicating young Strings.
451     if (G1StringDedup::is_candidate_from_evacuation(klass,
452                                                     region_attr,
453                                                     dest_attr,
454                                                     age)) {
455       // Record old; request adds a new weak reference, which reference
456       // processing expects to refer to a from-space object.
457       _string_dedup_requests.add(old);
458     }
459 
460     assert(_scanner.skip_card_mark_set(), "must be");
461     obj->oop_iterate_backwards(&_scanner, klass);
462 }
463 
464 // Private inline function, for direct internal use and providing the
465 // implementation of the public not-inline function.
466 MAYBE_INLINE_EVACUATION
467 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
468                                                     oop const old,
469                                                     markWord const old_mark) {
470   assert(region_attr.is_in_cset(),
471          "Unexpected region attr type: %s", region_attr.get_type_str());
472 
473   // NOTE: With compact headers, it is not safe to load the Klass* from old, because
474   // that would access the mark-word, that might change at any time by concurrent
475   // workers.
476   // This mark word would refer to a forwardee, which may not yet have completed
477   // copying. Therefore we must load the Klass* from the mark-word that we already
478   // loaded. This is safe, because we only enter here if not yet forwarded.
479   assert(!old_mark.is_forwarded(), "precondition");
480   Klass* klass = UseCompactObjectHeaders
481       ? old_mark.klass()
482       : old->klass();
483 
484   const size_t word_sz = old->size_given_klass(klass);
485 
486   // JNI only allows pinning of typeArrays, so we only need to keep those in place.
487   if (region_attr.is_pinned() && klass->is_typeArray_klass()) {
488     return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, true /* cause_pinned */);
489   }
490 
491   uint age = 0;
492   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
493   G1HeapRegion* const from_region = _g1h->heap_region_containing(old);
494   uint node_index = from_region->node_index();
495 
496   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
497 
498   // PLAB allocations should succeed most of the time, so we'll
499   // normally check against null once and that's it.
500   if (obj_ptr == nullptr) {
501     obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
502     if (obj_ptr == nullptr) {
503       // This will either forward-to-self, or detect that someone else has
504       // installed a forwarding pointer.
505       return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
506     }
507   }
508 
509   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
510   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
511 
512   // Should this evacuation fail?
513   if (inject_allocation_failure(from_region->hrm_index())) {
514     // Doing this after all the allocation attempts also tests the
515     // undo_allocation() method too.
516     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
517     return handle_evacuation_failure_par(old, old_mark, klass, region_attr, word_sz, false /* cause_pinned */);
518   }
519 
520   // We're going to allocate linearly, so might as well prefetch ahead.
521   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
522   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
523 
524   const oop obj = cast_to_oop(obj_ptr);
525   // Because the forwarding is done with memory_order_relaxed there is no
526   // ordering with the above copy.  Clients that get the forwardee must not
527   // examine its contents without other synchronization, since the contents
528   // may not be up to date for them.
529   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
530   if (forward_ptr == nullptr) {
531 
532     {
533       const uint young_index = from_region->young_index_in_cset();
534       assert((from_region->is_young() && young_index >  0) ||
535              (!from_region->is_young() && young_index == 0), "invariant" );
536       _surviving_young_words[young_index] += word_sz;
537     }
538 
539     if (dest_attr.is_young()) {
540       if (age < markWord::max_age) {
541         age++;
542         obj->incr_age();
543       }
544       _age_table.add(age, word_sz);
545     } else {
546       update_bot_after_copying(obj, word_sz);
547     }
548 
549     {
550       // Skip the card enqueue iff the object (obj) is in survivor region.
551       // However, G1HeapRegion::is_survivor() is too expensive here.
552       // Instead, we use dest_attr.is_young() because the two values are always
553       // equal: successfully allocated young regions must be survivor regions.
554       assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
555       G1SkipCardMarkSetter x(&_scanner, dest_attr.is_young());
556       do_iterate_object(obj, old, klass, region_attr, dest_attr, age);
557     }
558 
559     return obj;
560   } else {
561     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
562     return forward_ptr;
563   }
564 }
565 
566 // Public not-inline entry point.
567 ATTRIBUTE_FLATTEN
568 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
569                                                  oop old,
570                                                  markWord old_mark) {
571   return do_copy_to_survivor_space(region_attr, old, old_mark);
572 }
573 
574 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
575   assert(worker_id < _num_workers, "out of bounds access");
576   if (_states[worker_id] == nullptr) {
577     _states[worker_id] =
578       new G1ParScanThreadState(_g1h,
579                                worker_id,
580                                _num_workers,
581                                _collection_set,
582                                _evac_failure_regions);
583   }
584   return _states[worker_id];
585 }
586 
587 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
588   assert(_flushed, "thread local state from the per thread states should have been flushed");
589   return _surviving_young_words_total;
590 }
591 
592 void G1ParScanThreadStateSet::flush_stats() {
593   assert(!_flushed, "thread local state from the per thread states should be flushed once");
594   for (uint worker_id = 0; worker_id < _num_workers; ++worker_id) {
595     G1ParScanThreadState* pss = _states[worker_id];
596     assert(pss != nullptr, "must be initialized");
597 
598     G1GCPhaseTimes* p = _g1h->phase_times();
599 
600     // Need to get the following two before the call to G1ParThreadScanState::flush()
601     // because it resets the PLAB allocator where we get this info from.
602     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
603     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
604     size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _num_workers) * HeapWordSize;
605     size_t pending_cards = pss->num_cards_pending();
606     size_t to_young_gen_cards = pss->num_cards_marked() - pss->num_cards_pending();
607     size_t evac_failure_cards = pss->num_cards_from_evac_failure();
608     size_t marked_cards = pss->num_cards_marked();
609 
610     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
611     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
612     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
613     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, pending_cards, G1GCPhaseTimes::MergePSSPendingCards);
614     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, to_young_gen_cards, G1GCPhaseTimes::MergePSSToYoungGenCards);
615     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, evac_failure_cards, G1GCPhaseTimes::MergePSSEvacFail);
616     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, marked_cards, G1GCPhaseTimes::MergePSSMarked);
617 
618     delete pss;
619     _states[worker_id] = nullptr;
620   }
621 
622   _flushed = true;
623 }
624 
625 void G1ParScanThreadStateSet::record_unused_optional_region(G1HeapRegion* hr) {
626   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
627     G1ParScanThreadState* pss = _states[worker_index];
628     assert(pss != nullptr, "must be initialized");
629 
630     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
631     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
632   }
633 }
634 
635 void G1ParScanThreadState::record_evacuation_failed_region(G1HeapRegion* r, uint worker_id, bool cause_pinned) {
636   if (_evac_failure_regions->record(worker_id, r->hrm_index(), cause_pinned)) {
637     G1HeapRegionPrinter::evac_failure(r);
638   }
639 }
640 
641 NOINLINE
642 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, Klass* klass, G1HeapRegionAttr attr, size_t word_sz, bool cause_pinned) {
643   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
644 
645   oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
646   if (forward_ptr == nullptr) {
647     // Forward-to-self succeeded. We are the "owner" of the object.
648     G1HeapRegion* r = _g1h->heap_region_containing(old);
649 
650     record_evacuation_failed_region(r, _worker_id, cause_pinned);
651 
652     // Mark the failing object in the marking bitmap and later use the bitmap to handle
653     // evacuation failure recovery.
654     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
655 
656     _evacuation_failed_info.register_copy_failure(word_sz);
657 
658     {
659       // For iterating objects that failed evacuation currently we can reuse the
660       // existing closure to scan evacuated objects; since we are iterating from a
661       // collection set region (i.e. never a Survivor region), we always need to
662       // gather cards for this case.
663       G1SkipCardMarkSetter x(&_scanner, false /* skip_card_mark */);
664       do_iterate_object(old, old, klass, attr, attr, m.age());
665     }
666 
667     return old;
668   } else {
669     // Forward-to-self failed. Either someone else managed to allocate
670     // space for this object (old != forward_ptr) or they beat us in
671     // self-forwarding it (old == forward_ptr).
672     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
673            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
674            "should not be in the CSet",
675            p2i(old), p2i(forward_ptr));
676     return forward_ptr;
677   }
678 }
679 
680 void G1ParScanThreadState::initialize_numa_stats() {
681   if (_numa->is_enabled()) {
682     LogTarget(Info, gc, heap, numa) lt;
683 
684     if (lt.is_enabled()) {
685       uint num_nodes = _numa->num_active_nodes();
686       // Record only if there are multiple active nodes.
687       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
688       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
689     }
690   }
691 }
692 
693 void G1ParScanThreadState::flush_numa_stats() {
694   if (_obj_alloc_stat != nullptr) {
695     uint node_index = _numa->index_of_current_thread();
696     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
697   }
698 }
699 
700 void G1ParScanThreadState::update_numa_stats(uint node_index) {
701   if (_obj_alloc_stat != nullptr) {
702     _obj_alloc_stat[node_index]++;
703   }
704 }
705 
706 #if TASKQUEUE_STATS
707 
708 PartialArrayTaskStats* G1ParScanThreadState::partial_array_task_stats() {
709   return _partial_array_splitter.stats();
710 }
711 
712 #endif // TASKQUEUE_STATS
713 
714 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
715                                                  uint num_workers,
716                                                  G1CollectionSet* collection_set,
717                                                  G1EvacFailureRegions* evac_failure_regions) :
718     _g1h(g1h),
719     _collection_set(collection_set),
720     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, num_workers, mtGC)),
721     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, collection_set->young_region_length() + 1, mtGC)),
722     _num_workers(num_workers),
723     _flushed(false),
724     _evac_failure_regions(evac_failure_regions)
725 {
726   for (uint i = 0; i < num_workers; ++i) {
727     _states[i] = nullptr;
728   }
729   memset(_surviving_young_words_total, 0, (collection_set->young_region_length() + 1) * sizeof(size_t));
730 }
731 
732 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
733   assert(_flushed, "thread local state from the per thread states should have been flushed");
734   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
735   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
736 }
737 
738 #if TASKQUEUE_STATS
739 
740 void G1ParScanThreadStateSet::print_partial_array_task_stats() {
741   auto get_stats = [&](uint i) {
742     return state_for_worker(i)->partial_array_task_stats();
743   };
744   PartialArrayTaskStats::log_set(_num_workers, get_stats,
745                                  "Partial Array Task Stats");
746 }
747 
748 #endif // TASKQUEUE_STATS