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, Klass* klass,
 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(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(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_forwarded()) {
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 o->forwardee(m);
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   // NOTE: With compact headers, it is not safe to load the Klass* from o, because
173   // that would access the mark-word, and the mark-word might change at any time by
174   // concurrent promotion. The promoted mark-word would point to the forwardee, which
175   // may not yet have completed copying. Therefore we must load the Klass* from
176   // the mark-word that we have already loaded. This is safe, because we have checked
177   // that this is not yet forwarded in the caller.
178   Klass* klass = o->forward_safe_klass(test_mark);
179   size_t new_obj_size = o->size_given_klass(klass);
180 
181   // Find the objects age, MT safe.
182   uint age = (test_mark.has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
183       test_mark.displaced_mark_helper().age() : test_mark.age();
184 
185   if (!promote_immediately) {
186     // Try allocating obj in to-space (unless too old)
187     if (age < PSScavenge::tenuring_threshold()) {
188       new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
189       if (new_obj == nullptr && !_young_gen_is_full) {
190         // Do we allocate directly, or flush and refill?
191         if (new_obj_size > (YoungPLABSize / 2)) {
192           // Allocate this object directly
193           new_obj = cast_to_oop(young_space()->cas_allocate(new_obj_size));
194           promotion_trace_event(new_obj, klass, new_obj_size, age, false, nullptr);
195         } else {
196           // Flush and fill
197           _young_lab.flush();
198 
199           HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
200           if (lab_base != nullptr) {
201             _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
202             // Try the young lab allocation again.
203             new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
204             promotion_trace_event(new_obj, klass, new_obj_size, age, false, &_young_lab);
205           } else {
206             _young_gen_is_full = true;
207           }
208         }
209       }
210     }
211   }
212 
213   // Otherwise try allocating obj tenured
214   if (new_obj == nullptr) {
215 #ifndef PRODUCT
216     if (ParallelScavengeHeap::heap()->promotion_should_fail()) {
217       return oop_promotion_failed(o, test_mark);
218     }
219 #endif  // #ifndef PRODUCT
220 
221     new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
222     new_obj_is_tenured = true;
223 
224     if (new_obj == nullptr) {
225       if (!_old_gen_is_full) {
226         // Do we allocate directly, or flush and refill?
227         if (new_obj_size > (OldPLABSize / 2)) {
228           // Allocate this object directly
229           new_obj = cast_to_oop(old_gen()->allocate(new_obj_size));
230           promotion_trace_event(new_obj, klass, new_obj_size, age, true, nullptr);
231         } else {
232           // Flush and fill
233           _old_lab.flush();
234 
235           HeapWord* lab_base = old_gen()->allocate(OldPLABSize);
236           if(lab_base != nullptr) {
237             _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
238             // Try the old lab allocation again.
239             new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
240             promotion_trace_event(new_obj, klass, new_obj_size, age, true, &_old_lab);
241           }
242         }
243       }
244 
245       // This is the promotion failed test, and code handling.
246       // The code belongs here for two reasons. It is slightly
247       // different than the code below, and cannot share the
248       // CAS testing code. Keeping the code here also minimizes
249       // the impact on the common case fast path code.
250 
251       if (new_obj == nullptr) {
252         _old_gen_is_full = true;
253         return oop_promotion_failed(o, test_mark);
254       }
255     }
256   }
257 
258   assert(new_obj != nullptr, "allocation should have succeeded");
259 
260   // Copy obj
261   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(o), cast_from_oop<HeapWord*>(new_obj), new_obj_size);
262 
263   if (UseCompactObjectHeaders) {
264     // The copy above is not atomic. Make sure we have seen the proper mark
265     // and re-install it into the copy, so that Klass* is guaranteed to be correct.
266     markWord mark = o->mark();
267     if (!mark.is_forwarded()) {
268       new_obj->set_mark(mark);
269       ContinuationGCSupport::transform_stack_chunk(new_obj);
270     } else {
271       // If we copied a mark-word that indicates 'forwarded' state, the object
272       // installation would not succeed. We cannot access Klass* anymore either.
273       // Skip the transformation.
274     }
275   } else {
276     ContinuationGCSupport::transform_stack_chunk(new_obj);
277   }
278 
279   // Now we have to CAS in the header.
280   // Make copy visible to threads reading the forwardee.
281   oop forwardee = o->forward_to_atomic(new_obj, test_mark, memory_order_release);
282   if (forwardee == nullptr) {  // forwardee is null when forwarding is successful
283     // We won any races, we "own" this object.
284     assert(new_obj == o->forwardee(), "Sanity");
285 
286     // Increment age if obj still in new generation. Now that
287     // we're dealing with a markWord that cannot change, it is
288     // okay to use the non mt safe oop methods.
289     if (!new_obj_is_tenured) {
290       new_obj->incr_age();
291       assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
292     }
293 
294     // Do the size comparison first with new_obj_size, which we
295     // already have. Hopefully, only a few objects are larger than
296     // _min_array_size_for_chunking, and most of them will be arrays.
297     // So, the is->objArray() test would be very infrequent.
298     if (new_obj_size > _min_array_size_for_chunking &&
299         new_obj->is_objArray() &&
300         PSChunkLargeArrays) {
301       // we'll chunk it
302       push_depth(ScannerTask(PartialArrayScanTask(o)));
303       TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_array_chunk_pushes);
304     } else {
305       // we'll just push its contents
306       push_contents(new_obj);
307 
308       if (StringDedup::is_enabled() &&
309           java_lang_String::is_instance(new_obj) &&
310           psStringDedup::is_candidate_from_evacuation(new_obj, new_obj_is_tenured)) {
311         _string_dedup_requests.add(o);
312       }
313     }
314     return new_obj;
315   } else {
316     // We lost, someone else "owns" this object.
317     // Ensure loads from the forwardee follow all changes that preceded the
318     // release-cmpxchg that performed the forwarding in another thread.
319     OrderAccess::acquire();
320 
321     assert(o->is_forwarded(), "Object must be forwarded if the cas failed.");
322     assert(o->forwardee() == forwardee, "invariant");
323 
324     if (new_obj_is_tenured) {
325       _old_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
326     } else {
327       _young_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
328     }
329     return forwardee;
330   }
331 }
332 
333 // Attempt to "claim" oop at p via CAS, push the new obj if successful
334 template <bool promote_immediately, class T>
335 inline void PSPromotionManager::copy_and_push_safe_barrier(T* p) {
336   assert(ParallelScavengeHeap::heap()->is_in_reserved(p), "precondition");
337   assert(should_scavenge(p, true), "revisiting object?");
338 
339   oop o = RawAccess<IS_NOT_NULL>::oop_load(p);
340   oop new_obj = copy_to_survivor_space<promote_immediately>(o);
341   RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
342 
343   if (!PSScavenge::is_obj_in_young((HeapWord*)p) &&
344        PSScavenge::is_obj_in_young(new_obj)) {
345     PSScavenge::card_table()->inline_write_ref_field_gc(p);
346   }
347 }
348 
349 inline void PSPromotionManager::process_popped_location_depth(ScannerTask task) {
350   if (task.is_partial_array_task()) {
351     assert(PSChunkLargeArrays, "invariant");
352     process_array_chunk(task.to_partial_array_task());
353   } else {
354     if (task.is_narrow_oop_ptr()) {
355       assert(UseCompressedOops, "Error");
356       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_narrow_oop_ptr());
357     } else {
358       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_oop_ptr());
359     }
360   }
361 }
362 
363 inline bool PSPromotionManager::steal_depth(int queue_num, ScannerTask& t) {
364   return stack_array_depth()->steal(queue_num, t);
365 }
366 
367 #if TASKQUEUE_STATS
368 void PSPromotionManager::record_steal(ScannerTask task) {
369   if (task.is_partial_array_task()) {
370     ++_array_chunk_steals;
371   }
372 }
373 #endif // TASKQUEUE_STATS
374 
375 #endif // SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP