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(should_scavenge(p, true), "revisiting object?");
 60   assert(ParallelScavengeHeap::heap()->is_in(p), "pointer outside heap");
 61   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
 62   Prefetch::write(obj->mark_addr(), 0);
 63   push_depth(ScannerTask(p));
 64 }
 65 
 66 inline void PSPromotionManager::promotion_trace_event(oop new_obj, Klass* klass,
 67                                                       size_t obj_size,
 68                                                       uint age, bool tenured,
 69                                                       const PSPromotionLAB* lab) {
 70   // Skip if memory allocation failed
 71   if (new_obj != nullptr) {
 72     const ParallelScavengeTracer* gc_tracer = PSScavenge::gc_tracer();
 73 
 74     if (lab != nullptr) {
 75       // Promotion of object through newly allocated PLAB
 76       if (gc_tracer->should_report_promotion_in_new_plab_event()) {
 77         size_t obj_bytes = obj_size * HeapWordSize;
 78         size_t lab_size = lab->capacity();
 79         gc_tracer->report_promotion_in_new_plab_event(klass, obj_bytes,
 80                                                       age, tenured, lab_size);
 81       }
 82     } else {
 83       // Promotion of object directly to heap
 84       if (gc_tracer->should_report_promotion_outside_plab_event()) {
 85         size_t obj_bytes = obj_size * HeapWordSize;
 86         gc_tracer->report_promotion_outside_plab_event(klass, obj_bytes,
 87                                                        age, tenured);
 88       }
 89     }
 90   }
 91 }
 92 
 93 class PSPushContentsClosure: public BasicOopIterateClosure {
 94   PSPromotionManager* _pm;
 95  public:
 96   PSPushContentsClosure(PSPromotionManager* pm) : BasicOopIterateClosure(PSScavenge::reference_processor()), _pm(pm) {}
 97 
 98   template <typename T> void do_oop_nv(T* p) {
 99     if (PSScavenge::should_scavenge(p)) {
100       _pm->claim_or_forward_depth(p);
101     }
102   }
103 
104   virtual void do_oop(oop* p)       { do_oop_nv(p); }
105   virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
106 };
107 
108 //
109 // This closure specialization will override the one that is defined in
110 // instanceRefKlass.inline.cpp. It swaps the order of oop_oop_iterate and
111 // oop_oop_iterate_ref_processing. Unfortunately G1 and Parallel behaves
112 // significantly better (especially in the Derby benchmark) using opposite
113 // order of these function calls.
114 //
115 template <>
116 inline void InstanceRefKlass::oop_oop_iterate_reverse<oop, PSPushContentsClosure>(oop obj, PSPushContentsClosure* closure) {
117   oop_oop_iterate_ref_processing<oop>(obj, closure);
118   InstanceKlass::oop_oop_iterate_reverse<oop>(obj, closure);
119 }
120 
121 template <>
122 inline void InstanceRefKlass::oop_oop_iterate_reverse<narrowOop, PSPushContentsClosure>(oop obj, PSPushContentsClosure* closure) {
123   oop_oop_iterate_ref_processing<narrowOop>(obj, closure);
124   InstanceKlass::oop_oop_iterate_reverse<narrowOop>(obj, closure);
125 }
126 
127 inline void PSPromotionManager::push_contents(oop obj) {
128   if (!obj->klass()->is_typeArray_klass()) {
129     PSPushContentsClosure pcc(this);
130     obj->oop_iterate_backwards(&pcc);
131   }
132 }
133 
134 template<bool promote_immediately>
135 inline oop PSPromotionManager::copy_to_survivor_space(oop o) {
136   assert(should_scavenge(&o), "Sanity");
137 
138   // NOTE! We must be very careful with any methods that access the mark
139   // in o. There may be multiple threads racing on it, and it may be forwarded
140   // at any time.
141   markWord m = o->mark();
142   if (!m.is_marked()) {
143     return copy_unmarked_to_survivor_space<promote_immediately>(o, m);
144   } else {
145     // Ensure any loads from the forwardee follow all changes that precede
146     // the release-cmpxchg that performed the forwarding, possibly in some
147     // other thread.
148     OrderAccess::acquire();
149     // Return the already installed forwardee.
150     return o->forwardee(m);
151   }
152 }
153 
154 //
155 // This method is pretty bulky. It would be nice to split it up
156 // into smaller submethods, but we need to be careful not to hurt
157 // performance.
158 //
159 template<bool promote_immediately>
160 inline oop PSPromotionManager::copy_unmarked_to_survivor_space(oop o,
161                                                                markWord test_mark) {
162   assert(should_scavenge(&o), "Sanity");
163 
164   oop new_obj = nullptr;
165   bool new_obj_is_tenured = false;
166   // NOTE: With compact headers, it is not safe to load the Klass* from o, because
167   // that would access the mark-word, and the mark-word might change at any time by
168   // concurrent promotion. The promoted mark-word would point to the forwardee, which
169   // may not yet have completed copying. Therefore we must load the Klass* from
170   // the mark-word that we have already loaded. This is safe, because we have checked
171   // that this is not yet forwarded in the caller.
172   Klass* klass = o->forward_safe_klass(test_mark);
173   size_t new_obj_size = o->size_given_klass(klass);
174 
175   // Find the objects age, MT safe.
176   uint age = (test_mark.has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
177       test_mark.displaced_mark_helper().age() : test_mark.age();
178 
179   if (!promote_immediately) {
180     // Try allocating obj in to-space (unless too old)
181     if (age < PSScavenge::tenuring_threshold()) {
182       new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
183       if (new_obj == nullptr && !_young_gen_is_full) {
184         // Do we allocate directly, or flush and refill?
185         if (new_obj_size > (YoungPLABSize / 2)) {
186           // Allocate this object directly
187           new_obj = cast_to_oop(young_space()->cas_allocate(new_obj_size));
188           promotion_trace_event(new_obj, klass, new_obj_size, age, false, nullptr);
189         } else {
190           // Flush and fill
191           _young_lab.flush();
192 
193           HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
194           if (lab_base != nullptr) {
195             _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
196             // Try the young lab allocation again.
197             new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
198             promotion_trace_event(new_obj, klass, new_obj_size, age, false, &_young_lab);
199           } else {
200             _young_gen_is_full = true;
201           }
202         }
203       }
204     }
205   }
206 
207   // Otherwise try allocating obj tenured
208   if (new_obj == nullptr) {
209 #ifndef PRODUCT
210     if (ParallelScavengeHeap::heap()->promotion_should_fail()) {
211       return oop_promotion_failed(o, test_mark);
212     }
213 #endif  // #ifndef PRODUCT
214 
215     new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
216     new_obj_is_tenured = true;
217 
218     if (new_obj == nullptr) {
219       if (!_old_gen_is_full) {
220         // Do we allocate directly, or flush and refill?
221         if (new_obj_size > (OldPLABSize / 2)) {
222           // Allocate this object directly
223           new_obj = cast_to_oop(old_gen()->allocate(new_obj_size));
224           promotion_trace_event(new_obj, klass, new_obj_size, age, true, nullptr);
225         } else {
226           // Flush and fill
227           _old_lab.flush();
228 
229           HeapWord* lab_base = old_gen()->allocate(OldPLABSize);
230           if(lab_base != nullptr) {
231             _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
232             // Try the old lab allocation again.
233             new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
234             promotion_trace_event(new_obj, klass, new_obj_size, age, true, &_old_lab);
235           }
236         }
237       }
238 
239       // This is the promotion failed test, and code handling.
240       // The code belongs here for two reasons. It is slightly
241       // different than the code below, and cannot share the
242       // CAS testing code. Keeping the code here also minimizes
243       // the impact on the common case fast path code.
244 
245       if (new_obj == nullptr) {
246         _old_gen_is_full = true;
247         return oop_promotion_failed(o, test_mark);
248       }
249     }
250   }
251 
252   assert(new_obj != nullptr, "allocation should have succeeded");
253 
254   // Copy obj
255   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(o), cast_from_oop<HeapWord*>(new_obj), new_obj_size);
256 
257   // Parallel GC claims with a release - so other threads might access this object
258   // after claiming and they should see the "completed" object.
259   if (UseCompactObjectHeaders) {
260     // The copy above is not atomic. Make sure we have seen the proper mark
261     // and re-install it into the copy, so that Klass* is guaranteed to be correct.
262     markWord mark = o->mark();
263     if (!mark.is_marked()) {
264       new_obj->set_mark(mark);
265       ContinuationGCSupport::transform_stack_chunk(new_obj);
266     } else {
267       // If we copied a mark-word that indicates 'forwarded' state, the object
268       // installation would not succeed. We cannot access Klass* anymore either.
269       // Skip the transformation.
270     }
271   } else {
272     ContinuationGCSupport::transform_stack_chunk(new_obj);
273   }
274 
275   // Now we have to CAS in the header.
276   // Make copy visible to threads reading the forwardee.
277   oop forwardee = o->forward_to_atomic(new_obj, test_mark, memory_order_release);
278   if (forwardee == nullptr) {  // forwardee is null when forwarding is successful
279     // We won any races, we "own" this object.
280     assert(new_obj == o->forwardee(), "Sanity");
281 
282     // Increment age if obj still in new generation. Now that
283     // we're dealing with a markWord that cannot change, it is
284     // okay to use the non mt safe oop methods.
285     if (!new_obj_is_tenured) {
286       new_obj->incr_age();
287       assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
288     }
289 
290     // Do the size comparison first with new_obj_size, which we
291     // already have. Hopefully, only a few objects are larger than
292     // _min_array_size_for_chunking, and most of them will be arrays.
293     // So, the is->objArray() test would be very infrequent.
294     if (new_obj_size > _min_array_size_for_chunking &&
295         new_obj->is_objArray() &&
296         PSChunkLargeArrays) {
297       // we'll chunk it
298       push_depth(ScannerTask(PartialArrayScanTask(o)));
299       TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_array_chunk_pushes);
300     } else {
301       // we'll just push its contents
302       push_contents(new_obj);
303 
304       if (StringDedup::is_enabled() &&
305           java_lang_String::is_instance(new_obj) &&
306           psStringDedup::is_candidate_from_evacuation(new_obj, new_obj_is_tenured)) {
307         _string_dedup_requests.add(o);
308       }
309     }
310     return new_obj;
311   } else {
312     // We lost, someone else "owns" this object.
313     // Ensure loads from the forwardee follow all changes that preceded the
314     // release-cmpxchg that performed the forwarding in another thread.
315     OrderAccess::acquire();
316 
317     assert(o->is_forwarded(), "Object must be forwarded if the cas failed.");
318     assert(o->forwardee() == forwardee, "invariant");
319 
320     if (new_obj_is_tenured) {
321       _old_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
322     } else {
323       _young_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
324     }
325     return forwardee;
326   }
327 }
328 
329 // Attempt to "claim" oop at p via CAS, push the new obj if successful
330 template <bool promote_immediately, class T>
331 inline void PSPromotionManager::copy_and_push_safe_barrier(T* p) {
332   assert(ParallelScavengeHeap::heap()->is_in_reserved(p), "precondition");
333   assert(should_scavenge(p, true), "revisiting object?");
334 
335   oop o = RawAccess<IS_NOT_NULL>::oop_load(p);
336   oop new_obj = copy_to_survivor_space<promote_immediately>(o);
337   RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
338 
339   if (!PSScavenge::is_obj_in_young((HeapWord*)p) &&
340        PSScavenge::is_obj_in_young(new_obj)) {
341     PSScavenge::card_table()->inline_write_ref_field_gc(p);
342   }
343 }
344 
345 inline void PSPromotionManager::process_popped_location_depth(ScannerTask task) {
346   if (task.is_partial_array_task()) {
347     assert(PSChunkLargeArrays, "invariant");
348     process_array_chunk(task.to_partial_array_task());
349   } else {
350     if (task.is_narrow_oop_ptr()) {
351       assert(UseCompressedOops, "Error");
352       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_narrow_oop_ptr());
353     } else {
354       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_oop_ptr());
355     }
356   }
357 }
358 
359 inline bool PSPromotionManager::steal_depth(int queue_num, ScannerTask& t) {
360   return stack_array_depth()->steal(queue_num, t);
361 }
362 
363 #if TASKQUEUE_STATS
364 void PSPromotionManager::record_steal(ScannerTask task) {
365   if (task.is_partial_array_task()) {
366     ++_array_chunk_steals;
367   }
368 }
369 #endif // TASKQUEUE_STATS
370 
371 #endif // SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP