< prev index next >

src/hotspot/share/gc/g1/g1ParScanThreadState.cpp

Print this page

195 
196   // Although we never intentionally push references outside of the collection
197   // set, due to (benign) races in the claim mechanism during RSet scanning more
198   // than one thread might claim the same card. So the same card may be
199   // processed multiple times, and so we might get references into old gen here.
200   // So we need to redo this check.
201   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
202   // References pushed onto the work stack should never point to a humongous region
203   // as they are not added to the collection set due to above precondition.
204   assert(!region_attr.is_humongous_candidate(),
205          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
206          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
207 
208   if (!region_attr.is_in_cset()) {
209     // In this case somebody else already did all the work.
210     return;
211   }
212 
213   markWord m = obj->mark();
214   if (m.is_marked()) {
215     obj = cast_to_oop(m.decode_pointer());
216   } else {
217     obj = do_copy_to_survivor_space(region_attr, obj, m);
218   }
219   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
220 
221   write_ref_field_post(p, obj);
222 }
223 
224 MAYBE_INLINE_EVACUATION
225 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
226   oop from_obj = task.to_source_array();
227 
228   assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
229   assert(from_obj->is_objArray(), "must be obj array");
230   assert(from_obj->is_forwarded(), "must be forwarded");
231 
232   oop to_obj = from_obj->forwardee();
233   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
234   assert(to_obj->is_objArray(), "must be obj array");
235   objArrayOop to_array = objArrayOop(to_obj);
236 
237   PartialArrayTaskStepper::Step step
238     = _partial_array_stepper.next(objArrayOop(from_obj),
239                                   to_array,
240                                   _partial_objarray_chunk_size);
241   for (uint i = 0; i < step._ncreate; ++i) {
242     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
243   }
244 
245   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
246   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
247   // Process claimed task.  The length of to_array is not correct, but
248   // fortunately the iteration ignores the length field and just relies
249   // on start/end.
250   to_array->oop_iterate_range(&_scanner,
251                               step._index,
252                               step._index + _partial_objarray_chunk_size);
253 }
254 
255 MAYBE_INLINE_EVACUATION
256 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
257                                                   oop from_obj,
258                                                   oop to_obj) {
259   assert(from_obj->is_objArray(), "precondition");
260   assert(from_obj->is_forwarded(), "precondition");
261   assert(from_obj->forwardee() == to_obj, "precondition");
262   assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
263   assert(to_obj->is_objArray(), "precondition");
264 
265   objArrayOop to_array = objArrayOop(to_obj);
266 
267   PartialArrayTaskStepper::Step step
268     = _partial_array_stepper.start(objArrayOop(from_obj),
269                                    to_array,
270                                    _partial_objarray_chunk_size);
271 
272   // Push any needed partial scan tasks.  Pushed before processing the
273   // initial chunk to allow other workers to steal while we're processing.
274   for (uint i = 0; i < step._ncreate; ++i) {
275     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
276   }
277 
278   // Skip the card enqueue iff the object (to_array) is in survivor region.
279   // However, HeapRegion::is_survivor() is too expensive here.

366     // no other space to try.
367     return nullptr;
368   }
369 }
370 
371 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
372   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
373 
374   if (region_attr.is_young()) {
375     age = !m.has_displaced_mark_helper() ? m.age()
376                                          : m.displaced_mark_helper().age();
377     if (age < _tenuring_threshold) {
378       return region_attr;
379     }
380   }
381   // young-to-old (promotion) or old-to-old; destination is old in both cases.
382   return G1HeapRegionAttr::Old;
383 }
384 
385 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
386                                                   oop const old, size_t word_sz, uint age,
387                                                   HeapWord * const obj_ptr, uint node_index) const {
388   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
389   if (alloc_buf->contains(obj_ptr)) {
390     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
391                                                               dest_attr.type() == G1HeapRegionAttr::Old,
392                                                               alloc_buf->word_sz() * HeapWordSize);
393   } else {
394     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
395                                                                dest_attr.type() == G1HeapRegionAttr::Old);
396   }
397 }
398 
399 NOINLINE
400 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
401                                                    oop old,
402                                                    size_t word_sz,
403                                                    uint age,
404                                                    uint node_index) {
405   HeapWord* obj_ptr = nullptr;
406   // Try slow-path allocation unless we're allocating old and old is already full.
407   if (!(dest_attr->is_old() && _old_gen_is_full)) {
408     bool plab_refill_failed = false;
409     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
410                                                            word_sz,
411                                                            &plab_refill_failed,
412                                                            node_index);
413     if (obj_ptr == nullptr) {
414       obj_ptr = allocate_in_next_plab(dest_attr,
415                                       word_sz,
416                                       plab_refill_failed,
417                                       node_index);
418     }
419   }
420   if (obj_ptr != nullptr) {
421     update_numa_stats(node_index);
422     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
423       // The events are checked individually as part of the actual commit
424       report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
425     }
426   }
427   return obj_ptr;
428 }
429 
430 #if EVAC_FAILURE_INJECTOR
431 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
432   return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
433 }
434 #endif
435 
436 NOINLINE
437 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
438                                            HeapWord* obj_ptr,
439                                            size_t word_sz,
440                                            uint node_index) {
441   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
442 }
443 
444 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
445   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
446   HeapRegion* region = _g1h->heap_region_containing(obj_start);
447   region->update_bot_for_obj(obj_start, word_sz);
448 }
449 
450 // Private inline function, for direct internal use and providing the
451 // implementation of the public not-inline function.
452 MAYBE_INLINE_EVACUATION
453 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
454                                                     oop const old,
455                                                     markWord const old_mark) {
456   assert(region_attr.is_in_cset(),
457          "Unexpected region attr type: %s", region_attr.get_type_str());
458 
459   // Get the klass once.  We'll need it again later, and this avoids
460   // re-decoding when it's compressed.
461   Klass* klass = old->klass();






462   const size_t word_sz = old->size_given_klass(klass);
463 
464   uint age = 0;
465   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
466   HeapRegion* const from_region = _g1h->heap_region_containing(old);
467   uint node_index = from_region->node_index();
468 
469   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
470 
471   // PLAB allocations should succeed most of the time, so we'll
472   // normally check against null once and that's it.
473   if (obj_ptr == nullptr) {
474     obj_ptr = allocate_copy_slow(&dest_attr, old, word_sz, age, node_index);
475     if (obj_ptr == nullptr) {
476       // This will either forward-to-self, or detect that someone else has
477       // installed a forwarding pointer.
478       return handle_evacuation_failure_par(old, old_mark, word_sz);
479     }
480   }
481 
482   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
483   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
484 
485   // Should this evacuation fail?
486   if (inject_evacuation_failure(from_region->hrm_index())) {
487     // Doing this after all the allocation attempts also tests the
488     // undo_allocation() method too.
489     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
490     return handle_evacuation_failure_par(old, old_mark, word_sz);
491   }
492 
493   // We're going to allocate linearly, so might as well prefetch ahead.
494   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);

610     delete pss;
611     _states[worker_id] = nullptr;
612   }
613   _flushed = true;
614 }
615 
616 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
617   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
618     G1ParScanThreadState* pss = _states[worker_index];
619     assert(pss != nullptr, "must be initialized");
620 
621     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
622     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
623   }
624 }
625 
626 NOINLINE
627 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
628   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
629 
630   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
631   if (forward_ptr == nullptr) {
632     // Forward-to-self succeeded. We are the "owner" of the object.
633     HeapRegion* r = _g1h->heap_region_containing(old);
634 
635     if (_evac_failure_regions->record(r->hrm_index())) {
636       _g1h->hr_printer()->evac_failure(r);
637     }
638 
639     // Mark the failing object in the marking bitmap and later use the bitmap to handle
640     // evacuation failure recovery.
641     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
642 
643     _preserved_marks->push_if_necessary(old, m);
644 
645     ContinuationGCSupport::transform_stack_chunk(old);
646 
647     _evacuation_failed_info.register_copy_failure(word_sz);
648 
649     // For iterating objects that failed evacuation currently we can reuse the
650     // existing closure to scan evacuated objects; since we are iterating from a

195 
196   // Although we never intentionally push references outside of the collection
197   // set, due to (benign) races in the claim mechanism during RSet scanning more
198   // than one thread might claim the same card. So the same card may be
199   // processed multiple times, and so we might get references into old gen here.
200   // So we need to redo this check.
201   const G1HeapRegionAttr region_attr = _g1h->region_attr(obj);
202   // References pushed onto the work stack should never point to a humongous region
203   // as they are not added to the collection set due to above precondition.
204   assert(!region_attr.is_humongous_candidate(),
205          "Obj " PTR_FORMAT " should not refer to humongous region %u from " PTR_FORMAT,
206          p2i(obj), _g1h->addr_to_region(obj), p2i(p));
207 
208   if (!region_attr.is_in_cset()) {
209     // In this case somebody else already did all the work.
210     return;
211   }
212 
213   markWord m = obj->mark();
214   if (m.is_marked()) {
215     obj = obj->forwardee(m);
216   } else {
217     obj = do_copy_to_survivor_space(region_attr, obj, m);
218   }
219   RawAccess<IS_NOT_NULL>::oop_store(p, obj);
220 
221   write_ref_field_post(p, obj);
222 }
223 
224 MAYBE_INLINE_EVACUATION
225 void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {
226   oop from_obj = task.to_source_array();
227 
228   assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
229   assert(from_obj->forward_safe_klass()->is_objArray_klass(), "must be obj array");
230   assert(from_obj->is_forwarded(), "must be forwarded");
231 
232   oop to_obj = from_obj->forwardee();
233   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
234   assert(to_obj->is_objArray(), "must be obj array");
235   objArrayOop to_array = objArrayOop(to_obj);
236 
237   PartialArrayTaskStepper::Step step
238     = _partial_array_stepper.next(objArrayOop(from_obj),
239                                   to_array,
240                                   _partial_objarray_chunk_size);
241   for (uint i = 0; i < step._ncreate; ++i) {
242     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
243   }
244 
245   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
246   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
247   // Process claimed task.  The length of to_array is not correct, but
248   // fortunately the iteration ignores the length field and just relies
249   // on start/end.
250   to_array->oop_iterate_range(&_scanner,
251                               step._index,
252                               step._index + _partial_objarray_chunk_size);
253 }
254 
255 MAYBE_INLINE_EVACUATION
256 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
257                                                   oop from_obj,
258                                                   oop to_obj) {
259   assert(from_obj->forward_safe_klass()->is_objArray_klass(), "precondition");
260   assert(from_obj->is_forwarded(), "precondition");
261   assert(from_obj->forwardee() == to_obj, "precondition");
262   assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
263   assert(to_obj->is_objArray(), "precondition");
264 
265   objArrayOop to_array = objArrayOop(to_obj);
266 
267   PartialArrayTaskStepper::Step step
268     = _partial_array_stepper.start(objArrayOop(from_obj),
269                                    to_array,
270                                    _partial_objarray_chunk_size);
271 
272   // Push any needed partial scan tasks.  Pushed before processing the
273   // initial chunk to allow other workers to steal while we're processing.
274   for (uint i = 0; i < step._ncreate; ++i) {
275     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
276   }
277 
278   // Skip the card enqueue iff the object (to_array) is in survivor region.
279   // However, HeapRegion::is_survivor() is too expensive here.

366     // no other space to try.
367     return nullptr;
368   }
369 }
370 
371 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
372   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
373 
374   if (region_attr.is_young()) {
375     age = !m.has_displaced_mark_helper() ? m.age()
376                                          : m.displaced_mark_helper().age();
377     if (age < _tenuring_threshold) {
378       return region_attr;
379     }
380   }
381   // young-to-old (promotion) or old-to-old; destination is old in both cases.
382   return G1HeapRegionAttr::Old;
383 }
384 
385 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
386                                                   Klass* klass, size_t word_sz, uint age,
387                                                   HeapWord * const obj_ptr, uint node_index) const {
388   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
389   if (alloc_buf->contains(obj_ptr)) {
390     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(klass, word_sz * HeapWordSize, age,
391                                                               dest_attr.type() == G1HeapRegionAttr::Old,
392                                                               alloc_buf->word_sz() * HeapWordSize);
393   } else {
394     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(klass, word_sz * HeapWordSize, age,
395                                                                dest_attr.type() == G1HeapRegionAttr::Old);
396   }
397 }
398 
399 NOINLINE
400 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
401                                                    Klass* klass,
402                                                    size_t word_sz,
403                                                    uint age,
404                                                    uint node_index) {
405   HeapWord* obj_ptr = nullptr;
406   // Try slow-path allocation unless we're allocating old and old is already full.
407   if (!(dest_attr->is_old() && _old_gen_is_full)) {
408     bool plab_refill_failed = false;
409     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
410                                                            word_sz,
411                                                            &plab_refill_failed,
412                                                            node_index);
413     if (obj_ptr == nullptr) {
414       obj_ptr = allocate_in_next_plab(dest_attr,
415                                       word_sz,
416                                       plab_refill_failed,
417                                       node_index);
418     }
419   }
420   if (obj_ptr != nullptr) {
421     update_numa_stats(node_index);
422     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
423       // The events are checked individually as part of the actual commit
424       report_promotion_event(*dest_attr, klass, word_sz, age, obj_ptr, node_index);
425     }
426   }
427   return obj_ptr;
428 }
429 
430 #if EVAC_FAILURE_INJECTOR
431 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
432   return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
433 }
434 #endif
435 
436 NOINLINE
437 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
438                                            HeapWord* obj_ptr,
439                                            size_t word_sz,
440                                            uint node_index) {
441   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
442 }
443 
444 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
445   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
446   HeapRegion* region = _g1h->heap_region_containing(obj_start);
447   region->update_bot_for_obj(obj_start, word_sz);
448 }
449 
450 // Private inline function, for direct internal use and providing the
451 // implementation of the public not-inline function.
452 MAYBE_INLINE_EVACUATION
453 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
454                                                     oop const old,
455                                                     markWord const old_mark) {
456   assert(region_attr.is_in_cset(),
457          "Unexpected region attr type: %s", region_attr.get_type_str());
458 
459   // Get the klass once.  We'll need it again later, and this avoids
460   // re-decoding when it's compressed.
461   // NOTE: With compact headers, it is not safe to load the Klass* from o, because
462   // that would access the mark-word, and the mark-word might change at any time by
463   // concurrent promotion. The promoted mark-word would point to the forwardee, which
464   // may not yet have completed copying. Therefore we must load the Klass* from
465   // the mark-word that we have already loaded. This is safe, because we have checked
466   // that this is not yet forwarded in the caller.
467   Klass* klass = old->forward_safe_klass(old_mark);
468   const size_t word_sz = old->size_given_klass(klass);
469 
470   uint age = 0;
471   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
472   HeapRegion* const from_region = _g1h->heap_region_containing(old);
473   uint node_index = from_region->node_index();
474 
475   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
476 
477   // PLAB allocations should succeed most of the time, so we'll
478   // normally check against null once and that's it.
479   if (obj_ptr == nullptr) {
480     obj_ptr = allocate_copy_slow(&dest_attr, klass, word_sz, age, node_index);
481     if (obj_ptr == nullptr) {
482       // This will either forward-to-self, or detect that someone else has
483       // installed a forwarding pointer.
484       return handle_evacuation_failure_par(old, old_mark, word_sz);
485     }
486   }
487 
488   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
489   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
490 
491   // Should this evacuation fail?
492   if (inject_evacuation_failure(from_region->hrm_index())) {
493     // Doing this after all the allocation attempts also tests the
494     // undo_allocation() method too.
495     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
496     return handle_evacuation_failure_par(old, old_mark, word_sz);
497   }
498 
499   // We're going to allocate linearly, so might as well prefetch ahead.
500   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);

616     delete pss;
617     _states[worker_id] = nullptr;
618   }
619   _flushed = true;
620 }
621 
622 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
623   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
624     G1ParScanThreadState* pss = _states[worker_index];
625     assert(pss != nullptr, "must be initialized");
626 
627     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
628     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
629   }
630 }
631 
632 NOINLINE
633 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
634   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
635 
636   oop forward_ptr = old->forward_to_self_atomic(m, memory_order_relaxed);
637   if (forward_ptr == nullptr) {
638     // Forward-to-self succeeded. We are the "owner" of the object.
639     HeapRegion* r = _g1h->heap_region_containing(old);
640 
641     if (_evac_failure_regions->record(r->hrm_index())) {
642       _g1h->hr_printer()->evac_failure(r);
643     }
644 
645     // Mark the failing object in the marking bitmap and later use the bitmap to handle
646     // evacuation failure recovery.
647     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
648 
649     _preserved_marks->push_if_necessary(old, m);
650 
651     ContinuationGCSupport::transform_stack_chunk(old);
652 
653     _evacuation_failed_info.register_copy_failure(word_sz);
654 
655     // For iterating objects that failed evacuation currently we can reuse the
656     // existing closure to scan evacuated objects; since we are iterating from a
< prev index next >