1 /*
  2  * Copyright (c) 2002, 2023, 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 #ifndef SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP
 26 #define SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP
 27 
 28 #include "gc/parallel/psPromotionManager.hpp"
 29 
 30 #include "gc/parallel/parallelScavengeHeap.hpp"
 31 #include "gc/parallel/parMarkBitMap.inline.hpp"
 32 #include "gc/parallel/psOldGen.hpp"
 33 #include "gc/parallel/psPromotionLAB.inline.hpp"
 34 #include "gc/parallel/psScavenge.inline.hpp"
 35 #include "gc/parallel/psStringDedup.hpp"
 36 #include "gc/shared/continuationGCSupport.inline.hpp"
 37 #include "gc/shared/taskqueue.inline.hpp"
 38 #include "gc/shared/tlab_globals.hpp"
 39 #include "logging/log.hpp"
 40 #include "memory/iterator.inline.hpp"
 41 #include "oops/access.inline.hpp"
 42 #include "oops/oop.inline.hpp"
 43 #include "runtime/orderAccess.hpp"
 44 #include "runtime/prefetch.inline.hpp"
 45 #include "utilities/copy.hpp"
 46 
 47 inline PSPromotionManager* PSPromotionManager::manager_array(uint index) {
 48   assert(_manager_array != nullptr, "access of null manager_array");
 49   assert(index < ParallelGCThreads, "out of range manager_array access");
 50   return &_manager_array[index];
 51 }
 52 
 53 inline void PSPromotionManager::push_depth(ScannerTask task) {
 54   claimed_stack_depth()->push(task);
 55 }
 56 
 57 template <class T>
 58 inline void PSPromotionManager::claim_or_forward_depth(T* p) {
 59   assert(ParallelScavengeHeap::heap()->is_in(p), "pointer outside heap");
 60   T heap_oop = RawAccess<>::oop_load(p);
 61   if (PSScavenge::is_obj_in_young(heap_oop)) {
 62     oop obj = CompressedOops::decode_not_null(heap_oop);
 63     assert(!PSScavenge::is_obj_in_to_space(obj), "revisiting object?");
 64     Prefetch::write(obj->mark_addr(), 0);
 65     push_depth(ScannerTask(p));
 66   }
 67 }
 68 
 69 inline void PSPromotionManager::promotion_trace_event(oop new_obj, oop old_obj,
 70                                                       size_t obj_size,
 71                                                       uint age, bool tenured,
 72                                                       const PSPromotionLAB* lab) {
 73   // Skip if memory allocation failed
 74   if (new_obj != nullptr) {
 75     const ParallelScavengeTracer* gc_tracer = PSScavenge::gc_tracer();
 76 
 77     if (lab != nullptr) {
 78       // Promotion of object through newly allocated PLAB
 79       if (gc_tracer->should_report_promotion_in_new_plab_event()) {
 80         size_t obj_bytes = obj_size * HeapWordSize;
 81         size_t lab_size = lab->capacity();
 82         gc_tracer->report_promotion_in_new_plab_event(old_obj->klass(), obj_bytes,
 83                                                       age, tenured, lab_size);
 84       }
 85     } else {
 86       // Promotion of object directly to heap
 87       if (gc_tracer->should_report_promotion_outside_plab_event()) {
 88         size_t obj_bytes = obj_size * HeapWordSize;
 89         gc_tracer->report_promotion_outside_plab_event(old_obj->klass(), obj_bytes,
 90                                                        age, tenured);
 91       }
 92     }
 93   }
 94 }
 95 
 96 class PSPushContentsClosure: public BasicOopIterateClosure {
 97   PSPromotionManager* _pm;
 98  public:
 99   PSPushContentsClosure(PSPromotionManager* pm) : BasicOopIterateClosure(PSScavenge::reference_processor()), _pm(pm) {}
100 
101   template <typename T> void do_oop_work(T* p) {
102     _pm->claim_or_forward_depth(p);
103   }
104 
105   virtual void do_oop(oop* p)       { do_oop_work(p); }
106   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
107 };
108 
109 //
110 // This closure specialization will override the one that is defined in
111 // instanceRefKlass.inline.cpp. It swaps the order of oop_oop_iterate and
112 // oop_oop_iterate_ref_processing. Unfortunately G1 and Parallel behaves
113 // significantly better (especially in the Derby benchmark) using opposite
114 // order of these function calls.
115 //
116 template <>
117 inline void InstanceRefKlass::oop_oop_iterate_reverse<oop, PSPushContentsClosure>(oop obj, PSPushContentsClosure* closure) {
118   oop_oop_iterate_ref_processing<oop>(obj, closure);
119   InstanceKlass::oop_oop_iterate_reverse<oop>(obj, closure);
120 }
121 
122 template <>
123 inline void InstanceRefKlass::oop_oop_iterate_reverse<narrowOop, PSPushContentsClosure>(oop obj, PSPushContentsClosure* closure) {
124   oop_oop_iterate_ref_processing<narrowOop>(obj, closure);
125   InstanceKlass::oop_oop_iterate_reverse<narrowOop>(obj, closure);
126 }
127 
128 inline void PSPromotionManager::push_contents(oop obj) {
129   if (!obj->klass()->is_typeArray_klass()) {
130     PSPushContentsClosure pcc(this);
131     obj->oop_iterate_backwards(&pcc);
132   }
133 }
134 
135 inline void PSPromotionManager::push_contents_bounded(oop obj, HeapWord* left, HeapWord* right) {
136   PSPushContentsClosure pcc(this);
137   obj->oop_iterate(&pcc, MemRegion(left, right));
138 }
139 
140 template<bool promote_immediately>
141 inline oop PSPromotionManager::copy_to_survivor_space(oop o) {
142   assert(should_scavenge(&o), "Sanity");
143 
144   // NOTE! We must be very careful with any methods that access the mark
145   // in o. There may be multiple threads racing on it, and it may be forwarded
146   // at any time.
147   markWord m = o->mark();
148   if (!m.is_marked()) {
149     return copy_unmarked_to_survivor_space<promote_immediately>(o, m);
150   } else {
151     // Ensure any loads from the forwardee follow all changes that precede
152     // the release-cmpxchg that performed the forwarding, possibly in some
153     // other thread.
154     OrderAccess::acquire();
155     // Return the already installed forwardee.
156     return cast_to_oop(m.decode_pointer());
157   }
158 }
159 
160 //
161 // This method is pretty bulky. It would be nice to split it up
162 // into smaller submethods, but we need to be careful not to hurt
163 // performance.
164 //
165 template<bool promote_immediately>
166 inline oop PSPromotionManager::copy_unmarked_to_survivor_space(oop o,
167                                                                markWord test_mark) {
168   assert(should_scavenge(&o), "Sanity");
169 
170   oop new_obj = nullptr;
171   bool new_obj_is_tenured = false;
172   size_t new_obj_size = o->size();
173 
174   // Find the objects age, MT safe.
175   uint age = (test_mark.has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
176       test_mark.displaced_mark_helper().age() : test_mark.age();
177 
178   if (!promote_immediately) {
179     // Try allocating obj in to-space (unless too old)
180     if (age < PSScavenge::tenuring_threshold()) {
181       new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
182       if (new_obj == nullptr && !_young_gen_is_full) {
183         // Do we allocate directly, or flush and refill?
184         if (new_obj_size > (YoungPLABSize / 2)) {
185           // Allocate this object directly
186           new_obj = cast_to_oop(young_space()->cas_allocate(new_obj_size));
187           promotion_trace_event(new_obj, o, new_obj_size, age, false, nullptr);
188         } else {
189           // Flush and fill
190           _young_lab.flush();
191 
192           HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
193           if (lab_base != nullptr) {
194             _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
195             // Try the young lab allocation again.
196             new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
197             promotion_trace_event(new_obj, o, new_obj_size, age, false, &_young_lab);
198           } else {
199             _young_gen_is_full = true;
200           }
201         }
202       }
203     }
204   }
205 
206   // Otherwise try allocating obj tenured
207   if (new_obj == nullptr) {
208 #ifndef PRODUCT
209     if (ParallelScavengeHeap::heap()->promotion_should_fail()) {
210       return oop_promotion_failed(o, test_mark);
211     }
212 #endif  // #ifndef PRODUCT
213 
214     new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
215     new_obj_is_tenured = true;
216 
217     if (new_obj == nullptr) {
218       if (!_old_gen_is_full) {
219         // Do we allocate directly, or flush and refill?
220         if (new_obj_size > (OldPLABSize / 2)) {
221           // Allocate this object directly
222           new_obj = cast_to_oop(old_gen()->allocate(new_obj_size));
223           promotion_trace_event(new_obj, o, new_obj_size, age, true, nullptr);
224         } else {
225           // Flush and fill
226           _old_lab.flush();
227 
228           HeapWord* lab_base = old_gen()->allocate(OldPLABSize);
229           if(lab_base != nullptr) {
230             _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
231             // Try the old lab allocation again.
232             new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
233             promotion_trace_event(new_obj, o, new_obj_size, age, true, &_old_lab);
234           }
235         }
236       }
237 
238       // This is the promotion failed test, and code handling.
239       // The code belongs here for two reasons. It is slightly
240       // different than the code below, and cannot share the
241       // CAS testing code. Keeping the code here also minimizes
242       // the impact on the common case fast path code.
243 
244       if (new_obj == nullptr) {
245         _old_gen_is_full = true;
246         return oop_promotion_failed(o, test_mark);
247       }
248     }
249   }
250 
251   assert(new_obj != nullptr, "allocation should have succeeded");
252 
253   // Copy obj
254   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(o), cast_from_oop<HeapWord*>(new_obj), new_obj_size);
255 
256   // Parallel GC claims with a release - so other threads might access this object
257   // after claiming and they should see the "completed" object.
258   ContinuationGCSupport::transform_stack_chunk(new_obj);
259 
260   // Now we have to CAS in the header.
261   // Make copy visible to threads reading the forwardee.
262   oop forwardee = o->forward_to_atomic(new_obj, test_mark, memory_order_release);
263   if (forwardee == nullptr) {  // forwardee is null when forwarding is successful
264     // We won any races, we "own" this object.
265     assert(new_obj == o->forwardee(), "Sanity");
266 
267     // Increment age if obj still in new generation. Now that
268     // we're dealing with a markWord that cannot change, it is
269     // okay to use the non mt safe oop methods.
270     if (!new_obj_is_tenured) {
271       new_obj->incr_age();
272       assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
273     }
274 
275     // Do the size comparison first with new_obj_size, which we
276     // already have. Hopefully, only a few objects are larger than
277     // _min_array_size_for_chunking, and most of them will be arrays.
278     // So, the is->objArray() test would be very infrequent.
279     if (new_obj_size > _min_array_size_for_chunking &&
280         new_obj->is_objArray() &&
281         PSChunkLargeArrays) {
282       // we'll chunk it
283       push_depth(ScannerTask(PartialArrayScanTask(o)));
284       TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_array_chunk_pushes);
285     } else {
286       // we'll just push its contents
287       push_contents(new_obj);
288 
289       if (StringDedup::is_enabled() &&
290           java_lang_String::is_instance(new_obj) &&
291           psStringDedup::is_candidate_from_evacuation(new_obj, new_obj_is_tenured)) {
292         _string_dedup_requests.add(o);
293       }
294     }
295     return new_obj;
296   } else {
297     // We lost, someone else "owns" this object.
298     // Ensure loads from the forwardee follow all changes that preceded the
299     // release-cmpxchg that performed the forwarding in another thread.
300     OrderAccess::acquire();
301 
302     assert(o->is_forwarded(), "Object must be forwarded if the cas failed.");
303     assert(o->forwardee() == forwardee, "invariant");
304 
305     if (new_obj_is_tenured) {
306       _old_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
307     } else {
308       _young_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
309     }
310     return forwardee;
311   }
312 }
313 
314 // Attempt to "claim" oop at p via CAS, push the new obj if successful
315 template <bool promote_immediately, class T>
316 inline void PSPromotionManager::copy_and_push_safe_barrier(T* p) {
317   assert(ParallelScavengeHeap::heap()->is_in_reserved(p), "precondition");
318   assert(should_scavenge(p, true), "revisiting object?");
319 
320   oop o = RawAccess<IS_NOT_NULL>::oop_load(p);
321   oop new_obj = copy_to_survivor_space<promote_immediately>(o);
322   RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
323 
324   if (!PSScavenge::is_obj_in_young((HeapWord*)p) &&
325        PSScavenge::is_obj_in_young(new_obj)) {
326     PSScavenge::card_table()->inline_write_ref_field_gc(p);
327   }
328 }
329 
330 inline void PSPromotionManager::process_popped_location_depth(ScannerTask task) {
331   if (task.is_partial_array_task()) {
332     assert(PSChunkLargeArrays, "invariant");
333     process_array_chunk(task.to_partial_array_task());
334   } else {
335     if (task.is_narrow_oop_ptr()) {
336       assert(UseCompressedOops, "Error");
337       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_narrow_oop_ptr());
338     } else {
339       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_oop_ptr());
340     }
341   }
342 }
343 
344 inline bool PSPromotionManager::steal_depth(int queue_num, ScannerTask& t) {
345   return stack_array_depth()->steal(queue_num, t);
346 }
347 
348 #if TASKQUEUE_STATS
349 void PSPromotionManager::record_steal(ScannerTask task) {
350   if (task.is_partial_array_task()) {
351     ++_array_chunk_steals;
352   }
353 }
354 #endif // TASKQUEUE_STATS
355 
356 #endif // SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP