< prev index next >

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

Print this page

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 = cast_to_oop(m.decode_pointer());
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_objArray(), "must be obj array");
224   assert(from_obj->is_forwarded(), "must be forwarded");
225 
226   oop to_obj = from_obj->forwardee();
227   assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
228   assert(to_obj->is_objArray(), "must be obj array");
229   objArrayOop to_array = objArrayOop(to_obj);
230 
231   PartialArrayTaskStepper::Step step
232     = _partial_array_stepper.next(objArrayOop(from_obj),
233                                   to_array,
234                                   _partial_objarray_chunk_size);
235   for (uint i = 0; i < step._ncreate; ++i) {
236     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
237   }
238 
239   G1HeapRegionAttr dest_attr = _g1h->region_attr(to_array);
240   G1SkipCardEnqueueSetter x(&_scanner, dest_attr.is_new_survivor());
241   // Process claimed task.  The length of to_array is not correct, but
242   // fortunately the iteration ignores the length field and just relies
243   // on start/end.
244   to_array->oop_iterate_range(&_scanner,
245                               step._index,
246                               step._index + _partial_objarray_chunk_size);
247 }
248 
249 MAYBE_INLINE_EVACUATION
250 void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
251                                                   oop from_obj,
252                                                   oop to_obj) {
253   assert(from_obj->is_objArray(), "precondition");
254   assert(from_obj->is_forwarded(), "precondition");
255   assert(from_obj->forwardee() == to_obj, "precondition");
256   assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
257   assert(to_obj->is_objArray(), "precondition");
258 
259   objArrayOop to_array = objArrayOop(to_obj);
260 
261   PartialArrayTaskStepper::Step step
262     = _partial_array_stepper.start(objArrayOop(from_obj),
263                                    to_array,
264                                    _partial_objarray_chunk_size);
265 
266   // Push any needed partial scan tasks.  Pushed before processing the
267   // initial chunk to allow other workers to steal while we're processing.
268   for (uint i = 0; i < step._ncreate; ++i) {
269     push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
270   }
271 
272   // Skip the card enqueue iff the object (to_array) is in survivor region.
273   // However, HeapRegion::is_survivor() is too expensive here.

360     // no other space to try.
361     return nullptr;
362   }
363 }
364 
365 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
366   assert(region_attr.is_young() || region_attr.is_old(), "must be either Young or Old");
367 
368   if (region_attr.is_young()) {
369     age = !m.has_displaced_mark_helper() ? m.age()
370                                          : m.displaced_mark_helper().age();
371     if (age < _tenuring_threshold) {
372       return region_attr;
373     }
374   }
375   // young-to-old (promotion) or old-to-old; destination is old in both cases.
376   return G1HeapRegionAttr::Old;
377 }
378 
379 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
380                                                   oop const old, size_t word_sz, uint age,
381                                                   HeapWord * const obj_ptr, uint node_index) const {
382   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
383   if (alloc_buf->contains(obj_ptr)) {
384     _g1h->gc_tracer_stw()->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
385                                                               dest_attr.type() == G1HeapRegionAttr::Old,
386                                                               alloc_buf->word_sz() * HeapWordSize);
387   } else {
388     _g1h->gc_tracer_stw()->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
389                                                                dest_attr.type() == G1HeapRegionAttr::Old);
390   }
391 }
392 
393 NOINLINE
394 HeapWord* G1ParScanThreadState::allocate_copy_slow(G1HeapRegionAttr* dest_attr,
395                                                    oop old,

396                                                    size_t word_sz,
397                                                    uint age,
398                                                    uint node_index) {
399   HeapWord* obj_ptr = nullptr;
400   // Try slow-path allocation unless we're allocating old and old is already full.
401   if (!(dest_attr->is_old() && _old_gen_is_full)) {
402     bool plab_refill_failed = false;
403     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(*dest_attr,
404                                                            word_sz,
405                                                            &plab_refill_failed,
406                                                            node_index);
407     if (obj_ptr == nullptr) {
408       obj_ptr = allocate_in_next_plab(dest_attr,
409                                       word_sz,
410                                       plab_refill_failed,
411                                       node_index);
412     }
413   }
414   if (obj_ptr != nullptr) {
415     update_numa_stats(node_index);
416     if (_g1h->gc_tracer_stw()->should_report_promotion_events()) {
417       // The events are checked individually as part of the actual commit
418       report_promotion_event(*dest_attr, old, word_sz, age, obj_ptr, node_index);
419     }
420   }
421   return obj_ptr;
422 }
423 
424 #if EVAC_FAILURE_INJECTOR
425 bool G1ParScanThreadState::inject_evacuation_failure(uint region_idx) {
426   return _g1h->evac_failure_injector()->evacuation_should_fail(_evac_failure_inject_counter, region_idx);
427 }
428 #endif
429 
430 NOINLINE
431 void G1ParScanThreadState::undo_allocation(G1HeapRegionAttr dest_attr,
432                                            HeapWord* obj_ptr,
433                                            size_t word_sz,
434                                            uint node_index) {
435   _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
436 }
437 
438 void G1ParScanThreadState::update_bot_after_copying(oop obj, size_t word_sz) {
439   HeapWord* obj_start = cast_from_oop<HeapWord*>(obj);
440   HeapRegion* region = _g1h->heap_region_containing(obj_start);
441   region->update_bot_for_obj(obj_start, word_sz);
442 }
443 
444 // Private inline function, for direct internal use and providing the
445 // implementation of the public not-inline function.
446 MAYBE_INLINE_EVACUATION
447 oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const region_attr,
448                                                     oop const old,
449                                                     markWord const old_mark) {
450   assert(region_attr.is_in_cset(),
451          "Unexpected region attr type: %s", region_attr.get_type_str());
452 




453   // Get the klass once.  We'll need it again later, and this avoids
454   // re-decoding when it's compressed.
455   Klass* klass = old->klass();








456   const size_t word_sz = old->size_given_klass(klass);
457 
458   uint age = 0;
459   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
460   HeapRegion* const from_region = _g1h->heap_region_containing(old);
461   uint node_index = from_region->node_index();
462 
463   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
464 
465   // PLAB allocations should succeed most of the time, so we'll
466   // normally check against null once and that's it.
467   if (obj_ptr == nullptr) {
468     obj_ptr = allocate_copy_slow(&dest_attr, old, word_sz, age, node_index);
469     if (obj_ptr == nullptr) {
470       // This will either forward-to-self, or detect that someone else has
471       // installed a forwarding pointer.
472       return handle_evacuation_failure_par(old, old_mark, word_sz);
473     }
474   }
475 
476   assert(obj_ptr != nullptr, "when we get here, allocation should have succeeded");
477   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
478 
479   // Should this evacuation fail?
480   if (inject_evacuation_failure(from_region->hrm_index())) {
481     // Doing this after all the allocation attempts also tests the
482     // undo_allocation() method too.
483     undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
484     return handle_evacuation_failure_par(old, old_mark, word_sz);
485   }
486 
487   // We're going to allocate linearly, so might as well prefetch ahead.
488   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);

603     delete pss;
604     _states[worker_id] = nullptr;
605   }
606   _flushed = true;
607 }
608 
609 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
610   for (uint worker_index = 0; worker_index < _num_workers; ++worker_index) {
611     G1ParScanThreadState* pss = _states[worker_index];
612     assert(pss != nullptr, "must be initialized");
613 
614     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
615     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
616   }
617 }
618 
619 NOINLINE
620 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m, size_t word_sz) {
621   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
622 
623   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
624   if (forward_ptr == nullptr) {
625     // Forward-to-self succeeded. We are the "owner" of the object.
626     HeapRegion* r = _g1h->heap_region_containing(old);
627 
628     if (_evac_failure_regions->record(r->hrm_index())) {
629       _g1h->hr_printer()->evac_failure(r);
630     }
631 
632     // Mark the failing object in the marking bitmap and later use the bitmap to handle
633     // evacuation failure recovery.
634     _g1h->mark_evac_failure_object(_worker_id, old, word_sz);
635 
636     _preserved_marks->push_if_necessary(old, m);
637 
638     ContinuationGCSupport::transform_stack_chunk(old);
639 
640     _evacuation_failed_info.register_copy_failure(word_sz);
641 
642     // For iterating objects that failed evacuation currently we can reuse the
643     // existing closure to scan evacuated objects because:

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.

358     // no other space to try.
359     return nullptr;
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 = nullptr;
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 == nullptr) {
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 != nullptr) {
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 == nullptr) {
479     obj_ptr = allocate_copy_slow(&dest_attr, old, klass, word_sz, age, node_index);
480     if (obj_ptr == nullptr) {
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 != nullptr, "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);

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