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, Klass* klass,
64 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 Klass* klass;164 #ifdef _LP64165 if (UseCompactObjectHeaders) {166 klass = test_mark.safe_klass();167 } else168 #endif169 {170 klass = o->klass();171 }172 size_t new_obj_size = o->size_given_klass(klass);
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 == NULL && !_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, klass, new_obj_size, age, false, NULL);
188 } else {
189 // Flush and fill
190 _young_lab.flush();
191
192 HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
193 if (lab_base != NULL) {
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, klass, 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 == NULL) {
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 == NULL) {
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, klass, new_obj_size, age, true, NULL);
224 } else {
225 // Flush and fill
226 _old_lab.flush();
227
228 HeapWord* lab_base = old_gen()->allocate(OldPLABSize);
229 if(lab_base != NULL) {
230 #ifdef ASSERT
231 // Delay the initialization of the promotion lab (plab).
232 // This exposes uninitialized plabs to card table processing.
233 if (GCWorkerDelayMillis > 0) {
234 os::naked_sleep(GCWorkerDelayMillis);
235 }
236 #endif
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, o, 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 == NULL) {
252 _old_gen_is_full = true;
253 return oop_promotion_failed(o, test_mark);
254 }
255 }
256 }
257
258 assert(new_obj != NULL, "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 // Now we have to CAS in the header.
264 // Make copy visible to threads reading the forwardee.
265 oop forwardee = o->forward_to_atomic(new_obj, test_mark, memory_order_release);
266 if (forwardee == NULL) { // forwardee is NULL when forwarding is successful
267 // We won any races, we "own" this object.
268 assert(new_obj == o->forwardee(), "Sanity");
269
270 // Increment age if obj still in new generation. Now that
271 // we're dealing with a markWord that cannot change, it is
272 // okay to use the non mt safe oop methods.
273 if (!new_obj_is_tenured) {
274 new_obj->incr_age();
275 assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
276 }
277
278 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
279 new_obj_is_tenured ? "copying" : "tenuring",
280 new_obj->klass()->internal_name(),
281 p2i((void *)o), p2i((void *)new_obj), new_obj->size());
282
283 // Do the size comparison first with new_obj_size, which we
284 // already have. Hopefully, only a few objects are larger than
285 // _min_array_size_for_chunking, and most of them will be arrays.
286 // So, the is->objArray() test would be very infrequent.
287 if (new_obj_size > _min_array_size_for_chunking &&
288 new_obj->is_objArray() &&
289 PSChunkLargeArrays) {
290 // we'll chunk it
291 push_depth(ScannerTask(PartialArrayScanTask(o)));
292 TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_array_chunk_pushes);
293 } else {
294 // we'll just push its contents
295 push_contents(new_obj);
296 }
297 return new_obj;
298 } else {
299 // We lost, someone else "owns" this object.
300 // Ensure loads from the forwardee follow all changes that preceeded the
301 // release-cmpxchg that performed the forwarding in another thread.
302 OrderAccess::acquire();
303
304 assert(o->is_forwarded(), "Object must be forwarded if the cas failed.");
305 assert(o->forwardee() == forwardee, "invariant");
306
307 // Try to deallocate the space. If it was directly allocated we cannot
308 // deallocate it, so we have to test. If the deallocation fails,
309 // overwrite with a filler object.
310 if (new_obj_is_tenured) {
311 if (!_old_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size)) {
312 CollectedHeap::fill_with_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
313 }
314 } else if (!_young_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size)) {
315 CollectedHeap::fill_with_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
316 }
317 return forwardee;
318 }
319 }
320
321 // Attempt to "claim" oop at p via CAS, push the new obj if successful
322 // This version tests the oop* to make sure it is within the heap before
323 // attempting marking.
324 template <bool promote_immediately, class T>
325 inline void PSPromotionManager::copy_and_push_safe_barrier(T* p) {
326 assert(should_scavenge(p, true), "revisiting object?");
327
328 oop o = RawAccess<IS_NOT_NULL>::oop_load(p);
329 oop new_obj = copy_to_survivor_space<promote_immediately>(o);
330 RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
331
332 // We cannot mark without test, as some code passes us pointers
333 // that are outside the heap. These pointers are either from roots
334 // or from metadata.
335 if ((!PSScavenge::is_obj_in_young((HeapWord*)p)) &&
336 ParallelScavengeHeap::heap()->is_in_reserved(p)) {
337 if (PSScavenge::is_obj_in_young(new_obj)) {
338 PSScavenge::card_table()->inline_write_ref_field_gc(p, new_obj);
339 }
340 }
341 }
342
343 inline void PSPromotionManager::process_popped_location_depth(ScannerTask task) {
344 if (task.is_partial_array_task()) {
345 assert(PSChunkLargeArrays, "invariant");
346 process_array_chunk(task.to_partial_array_task());
347 } else {
348 if (task.is_narrow_oop_ptr()) {
349 assert(UseCompressedOops, "Error");
350 copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_narrow_oop_ptr());
351 } else {
352 copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_oop_ptr());
353 }
354 }
355 }
356
357 inline bool PSPromotionManager::steal_depth(int queue_num, ScannerTask& t) {
358 return stack_array_depth()->steal(queue_num, t);
359 }
360
361 #if TASKQUEUE_STATS
362 void PSPromotionManager::record_steal(ScannerTask task) {
363 if (task.is_partial_array_task()) {
364 ++_array_chunk_steals;
365 }
366 }
367 #endif // TASKQUEUE_STATS
368
369 #endif // SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP
--- EOF ---