1 /*
  2  * Copyright (c) 2002, 2021, 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/shared/taskqueue.inline.hpp"
 36 #include "gc/shared/tlab_globals.hpp"
 37 #include "logging/log.hpp"
 38 #include "memory/iterator.inline.hpp"
 39 #include "oops/access.inline.hpp"
 40 #include "oops/oop.inline.hpp"
 41 #include "runtime/orderAccess.hpp"
 42 #include "runtime/prefetch.inline.hpp"
 43 
 44 inline PSPromotionManager* PSPromotionManager::manager_array(uint index) {
 45   assert(_manager_array != NULL, "access of NULL manager_array");
 46   assert(index <= ParallelGCThreads, "out of range manager_array access");
 47   return &_manager_array[index];
 48 }
 49 
 50 inline void PSPromotionManager::push_depth(ScannerTask task) {
 51   claimed_stack_depth()->push(task);
 52 }
 53 
 54 template <class T>
 55 inline void PSPromotionManager::claim_or_forward_depth(T* p) {
 56   assert(should_scavenge(p, true), "revisiting object?");
 57   assert(ParallelScavengeHeap::heap()->is_in(p), "pointer outside heap");
 58   oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
 59   Prefetch::write(obj->mark_addr(), 0);
 60   push_depth(ScannerTask(p));
 61 }
 62 
 63 inline void PSPromotionManager::promotion_trace_event(oop new_obj, oop old_obj,
 64                                                       Klass* klass, size_t obj_size,
 65                                                       uint age, bool tenured,
 66                                                       const PSPromotionLAB* lab) {
 67   // Skip if memory allocation failed
 68   if (new_obj != NULL) {
 69     const ParallelScavengeTracer* gc_tracer = PSScavenge::gc_tracer();
 70 
 71     if (lab != NULL) {
 72       // Promotion of object through newly allocated PLAB
 73       if (gc_tracer->should_report_promotion_in_new_plab_event()) {
 74         size_t obj_bytes = obj_size * HeapWordSize;
 75         size_t lab_size = lab->capacity();
 76         gc_tracer->report_promotion_in_new_plab_event(klass, obj_bytes,
 77                                                       age, tenured, lab_size);
 78       }
 79     } else {
 80       // Promotion of object directly to heap
 81       if (gc_tracer->should_report_promotion_outside_plab_event()) {
 82         size_t obj_bytes = obj_size * HeapWordSize;
 83         gc_tracer->report_promotion_outside_plab_event(klass, obj_bytes,
 84                                                        age, tenured);
 85       }
 86     }
 87   }
 88 }
 89 
 90 class PSPushContentsClosure: public BasicOopIterateClosure {
 91   PSPromotionManager* _pm;
 92  public:
 93   PSPushContentsClosure(PSPromotionManager* pm) : BasicOopIterateClosure(PSScavenge::reference_processor()), _pm(pm) {}
 94 
 95   template <typename T> void do_oop_nv(T* p) {
 96     if (PSScavenge::should_scavenge(p)) {
 97       _pm->claim_or_forward_depth(p);
 98     }
 99   }
100 
101   virtual void do_oop(oop* p)       { do_oop_nv(p); }
102   virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
103 };
104 
105 //
106 // This closure specialization will override the one that is defined in
107 // instanceRefKlass.inline.cpp. It swaps the order of oop_oop_iterate and
108 // oop_oop_iterate_ref_processing. Unfortunately G1 and Parallel behaves
109 // significantly better (especially in the Derby benchmark) using opposite
110 // order of these function calls.
111 //
112 template <>
113 inline void InstanceRefKlass::oop_oop_iterate_reverse<oop, PSPushContentsClosure>(oop obj, PSPushContentsClosure* closure) {
114   oop_oop_iterate_ref_processing<oop>(obj, closure);
115   InstanceKlass::oop_oop_iterate_reverse<oop>(obj, closure);
116 }
117 
118 template <>
119 inline void InstanceRefKlass::oop_oop_iterate_reverse<narrowOop, PSPushContentsClosure>(oop obj, PSPushContentsClosure* closure) {
120   oop_oop_iterate_ref_processing<narrowOop>(obj, closure);
121   InstanceKlass::oop_oop_iterate_reverse<narrowOop>(obj, closure);
122 }
123 
124 inline void PSPromotionManager::push_contents(oop obj) {
125   if (!obj->klass()->is_typeArray_klass()) {
126     PSPushContentsClosure pcc(this);
127     obj->oop_iterate_backwards(&pcc);
128   }
129 }
130 
131 template<bool promote_immediately>
132 inline oop PSPromotionManager::copy_to_survivor_space(oop o) {
133   assert(should_scavenge(&o), "Sanity");
134 
135   // NOTE! We must be very careful with any methods that access the mark
136   // in o. There may be multiple threads racing on it, and it may be forwarded
137   // at any time.
138   markWord m = o->mark();
139   if (!m.is_marked()) {
140     return copy_unmarked_to_survivor_space<promote_immediately>(o, m);
141   } else {
142     // Ensure any loads from the forwardee follow all changes that precede
143     // the release-cmpxchg that performed the forwarding, possibly in some
144     // other thread.
145     OrderAccess::acquire();
146     // Return the already installed forwardee.
147     return o->forwardee(m);
148   }
149 }
150 
151 //
152 // This method is pretty bulky. It would be nice to split it up
153 // into smaller submethods, but we need to be careful not to hurt
154 // performance.
155 //
156 template<bool promote_immediately>
157 inline oop PSPromotionManager::copy_unmarked_to_survivor_space(oop o,
158                                                                markWord test_mark) {
159   assert(should_scavenge(&o), "Sanity");
160 
161   oop new_obj = NULL;
162   bool new_obj_is_tenured = false;
163   // NOTE: With compact headers, it is not safe to load the Klass* from o, because
164   // that would access the mark-word, and the mark-word might change at any time by
165   // concurrent promotion. The promoted mark-word would point to the forwardee, which
166   // may not yet have completed copying. Therefore we must load the Klass* from
167   // the mark-word that we have already loaded. This is safe, because we have checked
168   // that this is not yet forwarded in the caller.
169   Klass* klass = o->forward_safe_klass(test_mark);
170   size_t new_obj_size = o->size_given_klass(klass);
171 
172   // Find the objects age, MT safe.
173   uint age = (test_mark.has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
174       test_mark.displaced_mark_helper().age() : test_mark.age();
175 
176   if (!promote_immediately) {
177     // Try allocating obj in to-space (unless too old)
178     if (age < PSScavenge::tenuring_threshold()) {
179       new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
180       if (new_obj == NULL && !_young_gen_is_full) {
181         // Do we allocate directly, or flush and refill?
182         if (new_obj_size > (YoungPLABSize / 2)) {
183           // Allocate this object directly
184           new_obj = cast_to_oop(young_space()->cas_allocate(new_obj_size));
185           promotion_trace_event(new_obj, o, klass, new_obj_size, age, false, NULL);
186         } else {
187           // Flush and fill
188           _young_lab.flush();
189 
190           HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
191           if (lab_base != NULL) {
192             _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
193             // Try the young lab allocation again.
194             new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
195             promotion_trace_event(new_obj, o, klass, new_obj_size, age, false, &_young_lab);
196           } else {
197             _young_gen_is_full = true;
198           }
199         }
200       }
201     }
202   }
203 
204   // Otherwise try allocating obj tenured
205   if (new_obj == NULL) {
206 #ifndef PRODUCT
207     if (ParallelScavengeHeap::heap()->promotion_should_fail()) {
208       return oop_promotion_failed(o, test_mark);
209     }
210 #endif  // #ifndef PRODUCT
211 
212     new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
213     new_obj_is_tenured = true;
214 
215     if (new_obj == NULL) {
216       if (!_old_gen_is_full) {
217         // Do we allocate directly, or flush and refill?
218         if (new_obj_size > (OldPLABSize / 2)) {
219           // Allocate this object directly
220           new_obj = cast_to_oop(old_gen()->allocate(new_obj_size));
221           promotion_trace_event(new_obj, o, klass, new_obj_size, age, true, NULL);
222         } else {
223           // Flush and fill
224           _old_lab.flush();
225 
226           HeapWord* lab_base = old_gen()->allocate(OldPLABSize);
227           if(lab_base != NULL) {
228 #ifdef ASSERT
229             // Delay the initialization of the promotion lab (plab).
230             // This exposes uninitialized plabs to card table processing.
231             if (GCWorkerDelayMillis > 0) {
232               os::naked_sleep(GCWorkerDelayMillis);
233             }
234 #endif
235             _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
236             // Try the old lab allocation again.
237             new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
238             promotion_trace_event(new_obj, o, klass, new_obj_size, age, true, &_old_lab);
239           }
240         }
241       }
242 
243       // This is the promotion failed test, and code handling.
244       // The code belongs here for two reasons. It is slightly
245       // different than the code below, and cannot share the
246       // CAS testing code. Keeping the code here also minimizes
247       // the impact on the common case fast path code.
248 
249       if (new_obj == NULL) {
250         _old_gen_is_full = true;
251         return oop_promotion_failed(o, test_mark);
252       }
253     }
254   }
255 
256   assert(new_obj != NULL, "allocation should have succeeded");
257 
258   // Copy obj
259   Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(o), cast_from_oop<HeapWord*>(new_obj), new_obj_size);
260 
261   // Now we have to CAS in the header.
262   // Make copy visible to threads reading the forwardee.
263   oop forwardee = o->forward_to_atomic(new_obj, test_mark, memory_order_release);
264   if (forwardee == NULL) {  // forwardee is NULL when forwarding is successful
265     // We won any races, we "own" this object.
266     assert(new_obj == o->forwardee(), "Sanity");
267 
268     // Increment age if obj still in new generation. Now that
269     // we're dealing with a markWord that cannot change, it is
270     // okay to use the non mt safe oop methods.
271     if (!new_obj_is_tenured) {
272       new_obj->incr_age();
273       assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
274     }
275 
276     log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
277                                     new_obj_is_tenured ? "copying" : "tenuring",
278                                     new_obj->klass()->internal_name(),
279                                     p2i((void *)o), p2i((void *)new_obj), new_obj->size());
280 
281     // Do the size comparison first with new_obj_size, which we
282     // already have. Hopefully, only a few objects are larger than
283     // _min_array_size_for_chunking, and most of them will be arrays.
284     // So, the is->objArray() test would be very infrequent.
285     if (new_obj_size > _min_array_size_for_chunking &&
286         new_obj->is_objArray() &&
287         PSChunkLargeArrays) {
288       // we'll chunk it
289       push_depth(ScannerTask(PartialArrayScanTask(o)));
290       TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_array_chunk_pushes);
291     } else {
292       // we'll just push its contents
293       push_contents(new_obj);
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 preceeded 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     // Try to deallocate the space.  If it was directly allocated we cannot
306     // deallocate it, so we have to test.  If the deallocation fails,
307     // overwrite with a filler object.
308     if (new_obj_is_tenured) {
309       if (!_old_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size)) {
310         CollectedHeap::fill_with_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
311       }
312     } else if (!_young_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size)) {
313       CollectedHeap::fill_with_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
314     }
315     return forwardee;
316   }
317 }
318 
319 // Attempt to "claim" oop at p via CAS, push the new obj if successful
320 // This version tests the oop* to make sure it is within the heap before
321 // attempting marking.
322 template <bool promote_immediately, class T>
323 inline void PSPromotionManager::copy_and_push_safe_barrier(T* p) {
324   assert(should_scavenge(p, true), "revisiting object?");
325 
326   oop o = RawAccess<IS_NOT_NULL>::oop_load(p);
327   oop new_obj = copy_to_survivor_space<promote_immediately>(o);
328   RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
329 
330   // We cannot mark without test, as some code passes us pointers
331   // that are outside the heap. These pointers are either from roots
332   // or from metadata.
333   if ((!PSScavenge::is_obj_in_young((HeapWord*)p)) &&
334       ParallelScavengeHeap::heap()->is_in_reserved(p)) {
335     if (PSScavenge::is_obj_in_young(new_obj)) {
336       PSScavenge::card_table()->inline_write_ref_field_gc(p, new_obj);
337     }
338   }
339 }
340 
341 inline void PSPromotionManager::process_popped_location_depth(ScannerTask task) {
342   if (task.is_partial_array_task()) {
343     assert(PSChunkLargeArrays, "invariant");
344     process_array_chunk(task.to_partial_array_task());
345   } else {
346     if (task.is_narrow_oop_ptr()) {
347       assert(UseCompressedOops, "Error");
348       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_narrow_oop_ptr());
349     } else {
350       copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_oop_ptr());
351     }
352   }
353 }
354 
355 inline bool PSPromotionManager::steal_depth(int queue_num, ScannerTask& t) {
356   return stack_array_depth()->steal(queue_num, t);
357 }
358 
359 #if TASKQUEUE_STATS
360 void PSPromotionManager::record_steal(ScannerTask task) {
361   if (task.is_partial_array_task()) {
362     ++_array_chunk_steals;
363   }
364 }
365 #endif // TASKQUEUE_STATS
366 
367 #endif // SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP