1 /*
  2  * Copyright (c) 2014, 2022, Oracle and/or its affiliates. All rights reserved.
  3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  4  *
  5  * This code is free software; you can redistribute it and/or modify it
  6  * under the terms of the GNU General Public License version 2 only, as
  7  * published by the Free Software Foundation.
  8  *
  9  * This code is distributed in the hope that it will be useful, but WITHOUT
 10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 12  * version 2 for more details (a copy is included in the LICENSE file that
 13  * accompanied this code).
 14  *
 15  * You should have received a copy of the GNU General Public License version
 16  * 2 along with this work; if not, write to the Free Software Foundation,
 17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 18  *
 19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 20  * or visit www.oracle.com if you need additional information or have any
 21  * questions.
 22  *
 23  */
 24 
 25 #include "precompiled.hpp"
 26 #include "gc/g1/g1Allocator.inline.hpp"
 27 #include "gc/g1/g1CollectedHeap.inline.hpp"
 28 #include "gc/g1/g1CollectionSet.hpp"
 29 #include "gc/g1/g1EvacFailureRegions.inline.hpp"
 30 #include "gc/g1/g1OopClosures.inline.hpp"
 31 #include "gc/g1/g1ParScanThreadState.inline.hpp"
 32 #include "gc/g1/g1RootClosures.hpp"
 33 #include "gc/g1/g1StringDedup.hpp"
 34 #include "gc/g1/g1Trace.hpp"
 35 #include "gc/g1/g1YoungGCEvacFailureInjector.inline.hpp"
 36 #include "gc/shared/continuationGCSupport.inline.hpp"
 37 #include "gc/shared/partialArrayTaskStepper.inline.hpp"
 38 #include "gc/shared/preservedMarks.inline.hpp"
 39 #include "gc/shared/stringdedup/stringDedup.hpp"
 40 #include "gc/shared/taskqueue.inline.hpp"
 41 #include "memory/allocation.inline.hpp"
 42 #include "oops/access.inline.hpp"
 43 #include "oops/oop.inline.hpp"
 44 #include "runtime/atomic.hpp"
 45 #include "runtime/prefetch.inline.hpp"
 46 #include "utilities/globalDefinitions.hpp"
 47 #include "utilities/macros.hpp"
 48 
 49 // In fastdebug builds the code size can get out of hand, potentially
 50 // tripping over compiler limits (which may be bugs, but nevertheless
 51 // need to be taken into consideration).  A side benefit of limiting
 52 // inlining is that we get more call frames that might aid debugging.
 53 // And the fastdebug compile time for this file is much reduced.
 54 // Explicit NOINLINE to block ATTRIBUTE_FLATTENing.
 55 #define MAYBE_INLINE_EVACUATION NOT_DEBUG(inline) DEBUG_ONLY(NOINLINE)
 56 
 57 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
 58                                            G1RedirtyCardsQueueSet* rdcqs,
 59                                            PreservedMarks* preserved_marks,
 60                                            uint worker_id,
 61                                            uint n_workers,
 62                                            size_t young_cset_length,
 63                                            size_t optional_cset_length,
 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(NULL),
 70     _plab_allocator(NULL),
 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(NULL),
 80     _surviving_young_words(NULL),
 81     _surviving_words_length(young_cset_length + 1),
 82     _old_gen_is_full(false),
 83     _partial_objarray_chunk_size(ParGCArrayScanChunk),
 84     _partial_array_stepper(n_workers),
 85     _string_dedup_requests(),
 86     _num_optional_regions(optional_cset_length),
 87     _numa(g1h->numa()),
 88     _obj_alloc_stat(NULL),
 89     EVAC_FAILURE_INJECTOR_ONLY(_evac_failure_inject_counter(0) COMMA)
 90     _preserved_marks(preserved_marks),
 91     _evacuation_failed_info(),
 92     _evac_failure_regions(evac_failure_regions)
 93 {
 94   // We allocate number of young gen regions in the collection set plus one
 95   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
 96   // We also add a few elements at the beginning and at the end in
 97   // an attempt to eliminate cache contention
 98   const size_t padding_elem_num = (DEFAULT_CACHE_LINE_SIZE / sizeof(size_t));
 99   size_t array_length = padding_elem_num + _surviving_words_length + padding_elem_num;
100 
101   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
102   _surviving_young_words = _surviving_young_words_base + padding_elem_num;
103   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
104 
105   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
106 
107   _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
108 
109   _oops_into_optional_regions = new G1OopStarChunkedList[_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   _rdc_local_qset.flush();
116   flush_numa_stats();
117   // Update allocation statistics.
118   _plab_allocator->flush_and_retire_stats(num_workers);
119   _g1h->policy()->record_age_table(&_age_table);
120 
121   if (_evacuation_failed_info.has_failed()) {
122      _g1h->gc_tracer_stw()->report_evacuation_failed(_evacuation_failed_info);
123   }
124 
125   size_t sum = 0;
126   for (uint i = 0; i < _surviving_words_length; i++) {
127     surviving_young_words[i] += _surviving_young_words[i];
128     sum += _surviving_young_words[i];
129   }
130   return sum;
131 }
132 
133 G1ParScanThreadState::~G1ParScanThreadState() {
134   delete _plab_allocator;
135   delete _closures;
136   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
137   delete[] _oops_into_optional_regions;
138   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
139 }
140 
141 size_t G1ParScanThreadState::lab_waste_words() const {
142   return _plab_allocator->waste();
143 }
144 
145 size_t G1ParScanThreadState::lab_undo_waste_words() const {
146   return _plab_allocator->undo_waste();
147 }
148 
149 #ifdef ASSERT
150 void G1ParScanThreadState::verify_task(narrowOop* task) const {
151   assert(task != NULL, "invariant");
152   assert(UseCompressedOops, "sanity");
153   oop p = RawAccess<>::oop_load(task);
154   assert(_g1h->is_in_reserved(p),
155          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
156 }
157 
158 void G1ParScanThreadState::verify_task(oop* task) const {
159   assert(task != NULL, "invariant");
160   oop p = RawAccess<>::oop_load(task);
161   assert(_g1h->is_in_reserved(p),
162          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
163 }
164 
165 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
166   // Must be in the collection set--it's already been copied.
167   oop p = task.to_source_array();
168   assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
169 }
170 
171 void G1ParScanThreadState::verify_task(ScannerTask task) const {
172   if (task.is_narrow_oop_ptr()) {
173     verify_task(task.to_narrow_oop_ptr());
174   } else if (task.is_oop_ptr()) {
175     verify_task(task.to_oop_ptr());
176   } else if (task.is_partial_array_task()) {
177     verify_task(task.to_partial_array_task());
178   } else {
179     ShouldNotReachHere();
180   }
181 }
182 #endif // ASSERT
183 
184 template <class T>
185 MAYBE_INLINE_EVACUATION
186 void G1ParScanThreadState::do_oop_evac(T* p) {
187   // Reference should not be NULL here as such are never pushed to the task queue.
188   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
189 
190   // Although we never intentionally push references outside of the collection
191   // set, due to (benign) races in the claim mechanism during RSet scanning more
192   // than one thread might claim the same card. So the same card may be
193   // processed multiple times, and so we might get references into old gen here.
194   // So we need to redo this check.
195   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
196   // References pushed onto the work stack should never point to a humongous region
197   // as they are not added to the collection set due to above precondition.
198   assert(!region_attr.is_humongous_candidate(),
199          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
200          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
201 
202   if (!region_attr.is_in_cset()) {
203     // In this case somebody else already did all the work.
204     return;
205   }
206 
207   markWord m = obj->mark();
208   if (m.is_marked()) {
209     obj = obj->forwardee(m);
210   } else {
211     obj = do_copy_to_survivor_space(region_attr, obj, m);
212   }
213   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
214 
215   write_ref_field_post(p, obj);
216 }
217 
218 MAYBE_INLINE_EVACUATION
219 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
220   oop from_obj = task.to_source_array();
221 
222   assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
223   assert(from_obj->is_forwarded(), "must be forwarded");
224 
225   oop to_obj = from_obj->forwardee();
226   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
227   assert(to_obj->is_objArray(), "must be obj array");
228   objArrayOop to_array = objArrayOop(to_obj);
229 
230   PartialArrayTaskStepper::Step step
231     = _partial_array_stepper.next(objArrayOop(from_obj),
232                                   to_array,
233                                   _partial_objarray_chunk_size);
234   for (uint i = 0; i < step._ncreate; ++i) {
235     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
236   }
237 
238   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
239   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
240   // Process claimed task.  The length of to_array is not correct, but
241   // fortunately the iteration ignores the length field and just relies
242   // on start/end.
243   to_array->oop_iterate_range(&_scanner,
244                               step._index,
245                               step._index + _partial_objarray_chunk_size);
246 }
247 
248 MAYBE_INLINE_EVACUATION
249 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
250                                                   oop from_obj,
251                                                   oop to_obj) {
252   assert(from_obj->is_forwarded(), "precondition");
253   assert(from_obj->forwardee() == to_obj, "precondition");
254   assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
255   assert(to_obj->is_objArray(), "precondition");
256 
257   objArrayOop to_array = objArrayOop(to_obj);
258 
259   PartialArrayTaskStepper::Step step
260     = _partial_array_stepper.start(objArrayOop(from_obj),
261                                    to_array,
262                                    _partial_objarray_chunk_size);
263 
264   // Push any needed partial scan tasks.  Pushed before processing the
265   // initial chunk to allow other workers to steal while we're processing.
266   for (uint i = 0; i < step._ncreate; ++i) {
267     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
268   }
269 
270   // Skip the card enqueue iff the object (to_array) is in survivor region.
271   // However, HeapRegion::is_survivor() is too expensive here.
272   // Instead, we use dest_attr.is_young() because the two values are always
273   // equal: successfully allocated young regions must be survivor regions.
274   assert(dest_attr.is_young() == _g1h->heap_region_containing(to_array)->is_survivor(), "must be");
275   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
276   // Process the initial chunk.  No need to process the type in the
277   // klass, as it will already be handled by processing the built-in
278   // module. The length of to_array is not correct, but fortunately
279   // the iteration ignores that length field and relies on start/end.
280   to_array->oop_iterate_range(&_scanner, 0, step._index);
281 }
282 
283 MAYBE_INLINE_EVACUATION
284 void G1ParScanThreadState::dispatch_task(ScannerTask task) {
285   verify_task(task);
286   if (task.is_narrow_oop_ptr()) {
287     do_oop_evac(task.to_narrow_oop_ptr());
288   } else if (task.is_oop_ptr()) {
289     do_oop_evac(task.to_oop_ptr());
290   } else {
291     do_partial_array(task.to_partial_array_task());
292   }
293 }
294 
295 // Process tasks until overflow queue is empty and local queue
296 // contains no more than threshold entries.  NOINLINE to prevent
297 // inlining into steal_and_trim_queue.
298 ATTRIBUTE_FLATTEN NOINLINE
299 void G1ParScanThreadState::trim_queue_to_threshold(uint threshold) {
300   ScannerTask task;
301   do {
302     while (_task_queue->pop_overflow(task)) {
303       if (!_task_queue->try_push_to_taskqueue(task)) {
304         dispatch_task(task);
305       }
306     }
307     while (_task_queue->pop_local(task, threshold)) {
308       dispatch_task(task);
309     }
310   } while (!_task_queue->overflow_empty());
311 }
312 
313 ATTRIBUTE_FLATTEN
314 void G1ParScanThreadState::steal_and_trim_queue(G1ScannerTasksQueueSet* task_queues) {
315   ScannerTask stolen_task;
316   while (task_queues->steal(_worker_id, stolen_task)) {
317     dispatch_task(stolen_task);
318     // Processing stolen task may have added tasks to our queue.
319     trim_queue();
320   }
321 }
322 
323 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
324                                                       size_t word_sz,
325                                                       bool previous_plab_refill_failed,
326                                                       uint node_index) {
327 
328   assert(dest->is_in_cset_or_humongous_candidate(), "Unexpected dest: %s region attr", dest->get_type_str());
329 
330   // Right now we only have two types of regions (young / old) so
331   // let's keep the logic here simple. We can generalize it when necessary.
332   if (dest->is_young()) {
333     bool plab_refill_in_old_failed = false;
334     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
335                                                         word_sz,
336                                                         &plab_refill_in_old_failed,
337                                                         node_index);
338     // Make sure that we won't attempt to copy any other objects out
339     // of a survivor region (given that apparently we cannot allocate
340     // any new ones) to avoid coming into this slow path again and again.
341     // Only consider failed PLAB refill here: failed inline allocations are
342     // typically large, so not indicative of remaining space.
343     if (previous_plab_refill_failed) {
344       _tenuring_threshold = 0;
345     }
346 
347     if (obj_ptr != NULL) {
348       dest->set_old();
349     } else {
350       // We just failed to allocate in old gen. The same idea as explained above
351       // for making survivor gen unavailable for allocation applies for old gen.
352       _old_gen_is_full = plab_refill_in_old_failed;
353     }
354     return obj_ptr;
355   } else {
356     _old_gen_is_full = previous_plab_refill_failed;
357     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
358     // no other space to try.
359     return NULL;
360   }
361 }
362 
363 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
364   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
365 
366   if (region_attr.is_young()) {
367     age = !m.has_displaced_mark_helper() ? m.age()
368                                          : m.displaced_mark_helper().age();
369     if (age < _tenuring_threshold) {
370       return region_attr;
371     }
372   }
373   // young-to-old (promotion) or old-to-old; destination is old in both cases.
374   return G1HeapRegionAttr::Old;
375 }
376 
377 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
378                                                   oop const old, Klass* klass, size_t word_sz, uint age,
379                                                   HeapWord * const obj_ptr, uint node_index) const {
380   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
381   if (alloc_buf->contains(obj_ptr)) {
382     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
383                                                               dest_attr.type() == G1HeapRegionAttr::Old,
384                                                               alloc_buf->word_sz() * HeapWordSize);
385   } else {
386     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
387                                                                dest_attr.type() == G1HeapRegionAttr::Old);
388   }
389 }
390 
391 NOINLINE
392 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
393                                                    oop old,
394                                                    Klass* klass,
395                                                    size_t word_sz,
396                                                    uint age,
397                                                    uint node_index) {
398   HeapWord* obj_ptr = NULL;
399   // Try slow-path allocation unless we're allocating old and old is already full.
400   if (!(dest_attr->is_old() && _old_gen_is_full)) {
401     bool plab_refill_failed = false;
402     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
403                                                            word_sz,
404                                                            &plab_refill_failed,
405                                                            node_index);
406     if (obj_ptr == NULL) {
407       obj_ptr = allocate_in_next_plab(dest_attr,
408                                       word_sz,
409                                       plab_refill_failed,
410                                       node_index);
411     }
412   }
413   if (obj_ptr != NULL) {
414     update_numa_stats(node_index);
415     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
416       // The events are checked individually as part of the actual commit
417       report_promotion_event(*dest_attr, old, klass, word_sz, age, obj_ptr, node_index);
418     }
419   }
420   return obj_ptr;
421 }
422 
423 #if EVAC_FAILURE_INJECTOR
424 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
425   return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
426 }
427 #endif
428 
429 NOINLINE
430 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
431                                            HeapWord* obj_ptr,
432                                            size_t word_sz,
433                                            uint node_index) {
434   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
435 }
436 
437 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
438   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
439   HeapRegion* region = _g1h->heap_region_containing(obj_start);
440   region->update_bot_for_obj(obj_start, word_sz);
441 }
442 
443 // Private inline function, for direct internal use and providing the
444 // implementation of the public not-inline function.
445 MAYBE_INLINE_EVACUATION
446 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
447                                                     oop const old,
448                                                     markWord const old_mark) {
449   assert(region_attr.is_in_cset(),
450          "Unexpected region attr type: %s", region_attr.get_type_str());
451 
452   if (old_mark.is_marked()) {
453     // Already forwarded by somebody else, return forwardee.
454     return old->forwardee(old_mark);
455   }
456   // Get the klass once.  We'll need it again later, and this avoids
457   // re-decoding when it's compressed.
458   Klass* klass;
459 #ifdef _LP64
460   if (UseCompactObjectHeaders) {
461     klass = old_mark.safe_klass();
462   } else
463 #endif
464   {
465     klass = old->klass();
466   }
467   const size_t word_sz = old->size_given_klass(klass);
468 
469   uint age = 0;
470   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
471   HeapRegion* const from_region = _g1h->heap_region_containing(old);
472   uint node_index = from_region->node_index();
473 
474   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
475 
476   // PLAB allocations should succeed most of the time, so we'll
477   // normally check against NULL once and that's it.
478   if (obj_ptr == NULL) {
479     obj_ptr = allocate_copy_slow(&dest_attr, old, klass, word_sz, age, node_index);
480     if (obj_ptr == NULL) {
481       // This will either forward-to-self, or detect that someone else has
482       // installed a forwarding pointer.
483       return handle_evacuation_failure_par(old, old_mark, word_sz);
484     }
485   }
486 
487   assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
488   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
489 
490   // Should this evacuation fail?
491   if (inject_evacuation_failure(from_region->hrm_index())) {
492     // Doing this after all the allocation attempts also tests the
493     // undo_allocation() method too.
494     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
495     return handle_evacuation_failure_par(old, old_mark, word_sz);
496   }
497 
498   // We're going to allocate linearly, so might as well prefetch ahead.
499   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
500   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
501 
502   const oop obj = cast_to_oop(obj_ptr);
503   // Because the forwarding is done with memory_order_relaxed there is no
504   // ordering with the above copy.  Clients that get the forwardee must not
505   // examine its contents without other synchronization, since the contents
506   // may not be up to date for them.
507   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
508   if (forward_ptr == NULL) {
509 
510     {
511       const uint young_index = from_region->young_index_in_cset();
512       assert((from_region->is_young() && young_index >  0) ||
513              (!from_region->is_young() && young_index == 0), "invariant" );
514       _surviving_young_words[young_index] += word_sz;
515     }
516 
517     if (dest_attr.is_young()) {
518       if (age < markWord::max_age) {
519         age++;
520         obj->incr_age();
521       }
522       _age_table.add(age, word_sz);
523     } else {
524       update_bot_after_copying(obj, word_sz);
525     }
526 
527     // Most objects are not arrays, so do one array check rather than
528     // checking for each array category for each object.
529     if (klass->is_array_klass()) {
530       if (klass->is_objArray_klass()) {
531         start_partial_objarray(dest_attr, old, obj);
532       } else {
533         // Nothing needs to be done for typeArrays.  Body doesn't contain
534         // any oops to scan, and the type in the klass will already be handled
535         // by processing the built-in module.
536         assert(klass->is_typeArray_klass(), "invariant");
537       }
538       return obj;
539     }
540 
541     ContinuationGCSupport::transform_stack_chunk(obj);
542 
543     // Check for deduplicating young Strings.
544     if (G1StringDedup::is_candidate_from_evacuation(klass,
545                                                     region_attr,
546                                                     dest_attr,
547                                                     age)) {
548       // Record old; request adds a new weak reference, which reference
549       // processing expects to refer to a from-space object.
550       _string_dedup_requests.add(old);
551     }
552 
553     // Skip the card enqueue iff the object (obj) is in survivor region.
554     // However, HeapRegion::is_survivor() is too expensive here.
555     // Instead, we use dest_attr.is_young() because the two values are always
556     // equal: successfully allocated young regions must be survivor regions.
557     assert(dest_attr.is_young() == _g1h->heap_region_containing(obj)->is_survivor(), "must be");
558     G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_young());
559     obj->oop_iterate_backwards(&_scanner, klass);
560     return obj;
561   } else {
562     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
563     return forward_ptr;
564   }
565 }
566 
567 // Public not-inline entry point.
568 ATTRIBUTE_FLATTEN
569 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr region_attr,
570                                                  oop old,
571                                                  markWord old_mark) {
572   return do_copy_to_survivor_space(region_attr, old, old_mark);
573 }
574 
575 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
576   assert(worker_id < _n_workers, "out of bounds access");
577   if (_states[worker_id] == NULL) {
578     _states[worker_id] =
579       new G1ParScanThreadState(_g1h, rdcqs(),
580                                _preserved_marks_set.get(worker_id),
581                                worker_id, _n_workers,
582                                _young_cset_length, _optional_cset_length,
583                                _evac_failure_regions);
584   }
585   return _states[worker_id];
586 }
587 
588 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
589   assert(_flushed, "thread local state from the per thread states should have been flushed");
590   return _surviving_young_words_total;
591 }
592 
593 void G1ParScanThreadStateSet::flush_stats() {
594   assert(!_flushed, "thread local state from the per thread states should be flushed once");
595 
596   for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
597     G1ParScanThreadState* pss = _states[worker_id];
598     assert(pss != nullptr, "must be initialized");
599 
600     G1GCPhaseTimes* p = _g1h->phase_times();
601 
602     // Need to get the following two before the call to G1ParThreadScanState::flush()
603     // because it resets the PLAB allocator where we get this info from.
604     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
605     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
606     size_t copied_bytes = pss->flush_stats(_surviving_young_words_total, _n_workers) * HeapWordSize;
607 
608     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
609     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
610     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
611 
612     delete pss;
613     _states[worker_id] = NULL;
614   }
615   _flushed = true;
616 }
617 
618 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
619   for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
620     G1ParScanThreadState* pss = _states[worker_index];
621     assert(pss != nullptr, "must be initialized");
622 
623     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
624     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
625   }
626 }
627 
628 NOINLINE
629 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
630   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
631 
632   oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
633   if (forward_ptr == NULL) {
634     // Forward-to-self succeeded. We are the "owner" of the object.
635     HeapRegion* r = _g1h->heap_region_containing(old);
636 
637     if (_evac_failure_regions->record(r->hrm_index())) {
638       _g1h->hr_printer()->evac_failure(r);
639     }
640 
641     // Mark the failing object in the marking bitmap and later use the bitmap to handle
642     // evacuation failure recovery.
643     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
644 
645     _preserved_marks->push_if_necessary(old, m);
646 
647     ContinuationGCSupport::transform_stack_chunk(old);
648 
649     _evacuation_failed_info.register_copy_failure(word_sz);
650 
651     // For iterating objects that failed evacuation currently we can reuse the
652     // existing closure to scan evacuated objects because:
653     // - for objects referring into the collection set we do not need to gather
654     // cards at this time. The regions they are in will be unconditionally turned
655     // to old regions without remembered sets.
656     // - since we are iterating from a collection set region (i.e. never a Survivor
657     // region), we always need to gather cards for this case.
658     G1SkipCardEnqueueSetter x(&_scanner, false /* skip_card_enqueue */);
659     old->oop_iterate_backwards(&_scanner);
660 
661     return old;
662   } else {
663     // Forward-to-self failed. Either someone else managed to allocate
664     // space for this object (old != forward_ptr) or they beat us in
665     // self-forwarding it (old == forward_ptr).
666     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
667            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
668            "should not be in the CSet",
669            p2i(old), p2i(forward_ptr));
670     return forward_ptr;
671   }
672 }
673 
674 void G1ParScanThreadState::initialize_numa_stats() {
675   if (_numa->is_enabled()) {
676     LogTarget(Info, gc, heap, numa) lt;
677 
678     if (lt.is_enabled()) {
679       uint num_nodes = _numa->num_active_nodes();
680       // Record only if there are multiple active nodes.
681       _obj_alloc_stat = NEW_C_HEAP_ARRAY(size_t, num_nodes, mtGC);
682       memset(_obj_alloc_stat, 0, sizeof(size_t) * num_nodes);
683     }
684   }
685 }
686 
687 void G1ParScanThreadState::flush_numa_stats() {
688   if (_obj_alloc_stat != NULL) {
689     uint node_index = _numa->index_of_current_thread();
690     _numa->copy_statistics(G1NUMAStats::LocalObjProcessAtCopyToSurv, node_index, _obj_alloc_stat);
691   }
692 }
693 
694 void G1ParScanThreadState::update_numa_stats(uint node_index) {
695   if (_obj_alloc_stat != NULL) {
696     _obj_alloc_stat[node_index]++;
697   }
698 }
699 
700 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
701                                                  uint n_workers,
702                                                  size_t young_cset_length,
703                                                  size_t optional_cset_length,
704                                                  G1EvacFailureRegions* evac_failure_regions) :
705     _g1h(g1h),
706     _rdcqs(G1BarrierSet::dirty_card_queue_set().allocator()),
707     _preserved_marks_set(true /* in_c_heap */),
708     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
709     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
710     _young_cset_length(young_cset_length),
711     _optional_cset_length(optional_cset_length),
712     _n_workers(n_workers),
713     _flushed(false),
714     _evac_failure_regions(evac_failure_regions) {
715   _preserved_marks_set.init(n_workers);
716   for (uint i = 0; i < n_workers; ++i) {
717     _states[i] = NULL;
718   }
719   memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
720 }
721 
722 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
723   assert(_flushed, "thread local state from the per thread states should have been flushed");
724   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
725   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
726   _preserved_marks_set.assert_empty();
727   _preserved_marks_set.reclaim();
728 }