1 /*
2 * Copyright (c) 2002, 2024, 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 // Return the already installed forwardee.
152 return o->forwardee(m);
153 }
154 }
155
156 //
157 // This method is pretty bulky. It would be nice to split it up
158 // into smaller submethods, but we need to be careful not to hurt
159 // performance.
160 //
161 template<bool promote_immediately>
162 inline oop PSPromotionManager::copy_unmarked_to_survivor_space(oop o,
163 markWord test_mark) {
164 assert(should_scavenge(&o), "Sanity");
165
166 oop new_obj = nullptr;
167 bool new_obj_is_tenured = false;
168
169 // NOTE: With compact headers, it is not safe to load the Klass* from old, because
170 // that would access the mark-word, that might change at any time by concurrent
171 // workers.
172 // This mark word would refer to a forwardee, which may not yet have completed
173 // copying. Therefore we must load the Klass* from the mark-word that we already
174 // loaded. This is safe, because we only enter here if not yet forwarded.
175 assert(!test_mark.is_forwarded(), "precondition");
176 Klass* klass = UseCompactObjectHeaders
177 ? test_mark.klass()
178 : o->klass();
179
180 size_t old_obj_size = o->size_given_mark_and_klass(test_mark, klass);
181 size_t new_obj_size = o->copy_size(old_obj_size, test_mark);
182
183 // Find the objects age, MT safe.
184 uint age = (test_mark.has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
185 test_mark.displaced_mark_helper().age() : test_mark.age();
186
187 if (!promote_immediately) {
188 // Try allocating obj in to-space (unless too old)
189 if (age < PSScavenge::tenuring_threshold()) {
190 new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
191 if (new_obj == nullptr && !_young_gen_is_full) {
192 // Do we allocate directly, or flush and refill?
193 if (new_obj_size > (YoungPLABSize / 2)) {
194 // Allocate this object directly
195 new_obj = cast_to_oop(young_space()->cas_allocate(new_obj_size));
196 promotion_trace_event(new_obj, klass, new_obj_size, age, false, nullptr);
197 } else {
198 // Flush and fill
199 _young_lab.flush();
200
201 HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
202 if (lab_base != nullptr) {
203 _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
204 // Try the young lab allocation again.
205 new_obj = cast_to_oop(_young_lab.allocate(new_obj_size));
206 promotion_trace_event(new_obj, klass, new_obj_size, age, false, &_young_lab);
207 } else {
208 _young_gen_is_full = true;
209 }
210 }
211 }
212 }
213 }
214
215 // Otherwise try allocating obj tenured
216 if (new_obj == nullptr) {
217 #ifndef PRODUCT
218 if (ParallelScavengeHeap::heap()->promotion_should_fail()) {
219 return oop_promotion_failed(o, test_mark);
220 }
221 #endif // #ifndef PRODUCT
222
223 new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
224 new_obj_is_tenured = true;
225
226 if (new_obj == nullptr) {
227 if (!_old_gen_is_full) {
228 // Do we allocate directly, or flush and refill?
229 if (new_obj_size > (OldPLABSize / 2)) {
230 // Allocate this object directly
231 new_obj = cast_to_oop(old_gen()->allocate(new_obj_size));
232 promotion_trace_event(new_obj, klass, new_obj_size, age, true, nullptr);
233 } else {
234 // Flush and fill
235 _old_lab.flush();
236
237 HeapWord* lab_base = old_gen()->allocate(OldPLABSize);
238 if(lab_base != nullptr) {
239 _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
240 // Try the old lab allocation again.
241 new_obj = cast_to_oop(_old_lab.allocate(new_obj_size));
242 promotion_trace_event(new_obj, klass, new_obj_size, age, true, &_old_lab);
243 }
244 }
245 }
246
247 // This is the promotion failed test, and code handling.
248 // The code belongs here for two reasons. It is slightly
249 // different than the code below, and cannot share the
250 // CAS testing code. Keeping the code here also minimizes
251 // the impact on the common case fast path code.
252
253 if (new_obj == nullptr) {
254 _old_gen_is_full = true;
255 return oop_promotion_failed(o, test_mark);
256 }
257 }
258 }
259
260 assert(new_obj != nullptr, "allocation should have succeeded");
261
262 // Copy obj
263 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(o), cast_from_oop<HeapWord*>(new_obj), new_obj_size);
264
265 // Now we have to CAS in the header.
266 // Because the forwarding is done with memory_order_relaxed there is no
267 // ordering with the above copy. Clients that get the forwardee must not
268 // examine its contents without other synchronization, since the contents
269 // may not be up to date for them.
270 oop forwardee = o->forward_to_atomic(new_obj, test_mark, memory_order_relaxed);
271 if (forwardee == nullptr) { // forwardee is null when forwarding is successful
272 // We won any races, we "own" this object.
273 assert(new_obj == o->forwardee(), "Sanity");
274
275 // Increment age if obj still in new generation. Now that
276 // we're dealing with a markWord that cannot change, it is
277 // okay to use the non mt safe oop methods.
278 if (!new_obj_is_tenured) {
279 new_obj->incr_age();
280 assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
281 }
282
283 ContinuationGCSupport::transform_stack_chunk(new_obj);
284
285 // Do the size comparison first with new_obj_size, which we
286 // already have. Hopefully, only a few objects are larger than
287 // _min_array_size_for_chunking, and most of them will be arrays.
288 // So, the is->objArray() test would be very infrequent.
289 if (new_obj_size > _min_array_size_for_chunking &&
290 new_obj->is_objArray() &&
291 PSChunkLargeArrays) {
292 push_objArray(o, new_obj);
293 } else {
294 // we'll just push its contents
295 push_contents(new_obj);
296
297 if (StringDedup::is_enabled() &&
298 java_lang_String::is_instance(new_obj) &&
299 psStringDedup::is_candidate_from_evacuation(new_obj, new_obj_is_tenured)) {
300 _string_dedup_requests.add(o);
301 }
302 }
303 return new_obj;
304 } else {
305 // We lost, someone else "owns" this object.
306
307 assert(o->is_forwarded(), "Object must be forwarded if the cas failed.");
308 assert(o->forwardee() == forwardee, "invariant");
309
310 if (new_obj_is_tenured) {
311 _old_lab.unallocate_object(cast_from_oop<HeapWord*>(new_obj), new_obj_size);
312 } else {
313 _young_lab.unallocate_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 template <bool promote_immediately, class T>
321 inline void PSPromotionManager::copy_and_push_safe_barrier(T* p) {
322 assert(ParallelScavengeHeap::heap()->is_in_reserved(p), "precondition");
323 assert(should_scavenge(p, true), "revisiting object?");
324
325 oop o = RawAccess<IS_NOT_NULL>::oop_load(p);
326 oop new_obj = copy_to_survivor_space<promote_immediately>(o);
327 RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
328
329 if (!PSScavenge::is_obj_in_young((HeapWord*)p) &&
330 PSScavenge::is_obj_in_young(new_obj)) {
331 PSScavenge::card_table()->inline_write_ref_field_gc(p);
332 }
333 }
334
335 inline void PSPromotionManager::process_popped_location_depth(ScannerTask task) {
336 if (task.is_partial_array_state()) {
337 assert(PSChunkLargeArrays, "invariant");
338 process_array_chunk(task.to_partial_array_state());
339 } else {
340 if (task.is_narrow_oop_ptr()) {
341 assert(UseCompressedOops, "Error");
342 copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_narrow_oop_ptr());
343 } else {
344 copy_and_push_safe_barrier</*promote_immediately=*/false>(task.to_oop_ptr());
345 }
346 }
347 }
348
349 inline bool PSPromotionManager::steal_depth(int queue_num, ScannerTask& t) {
350 return stack_array_depth()->steal(queue_num, t);
351 }
352
353 #if TASKQUEUE_STATS
354 void PSPromotionManager::record_steal(ScannerTask task) {
355 if (task.is_partial_array_state()) {
356 ++_array_chunk_steals;
357 }
358 }
359 #endif // TASKQUEUE_STATS
360
361 #endif // SHARE_GC_PARALLEL_PSPROMOTIONMANAGER_INLINE_HPP