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 #include "precompiled.hpp"
26 #include "classfile/javaClasses.inline.hpp"
27 #include "gc/parallel/mutableSpace.hpp"
28 #include "gc/parallel/parallelScavengeHeap.hpp"
29 #include "gc/parallel/psOldGen.hpp"
30 #include "gc/parallel/psPromotionManager.inline.hpp"
31 #include "gc/parallel/psScavenge.inline.hpp"
32 #include "gc/shared/continuationGCSupport.inline.hpp"
33 #include "gc/shared/gcTrace.hpp"
34 #include "gc/shared/preservedMarks.inline.hpp"
35 #include "gc/shared/taskqueue.inline.hpp"
36 #include "logging/log.hpp"
37 #include "logging/logStream.hpp"
38 #include "memory/allocation.inline.hpp"
39 #include "memory/iterator.inline.hpp"
40 #include "memory/memRegion.hpp"
41 #include "memory/padded.inline.hpp"
42 #include "memory/resourceArea.hpp"
43 #include "oops/access.inline.hpp"
44 #include "oops/compressedOops.inline.hpp"
45
46 PaddedEnd<PSPromotionManager>* PSPromotionManager::_manager_array = nullptr;
47 PSPromotionManager::PSScannerTasksQueueSet* PSPromotionManager::_stack_array_depth = nullptr;
48 PreservedMarksSet* PSPromotionManager::_preserved_marks_set = nullptr;
49 PSOldGen* PSPromotionManager::_old_gen = nullptr;
50 MutableSpace* PSPromotionManager::_young_space = nullptr;
51
52 void PSPromotionManager::initialize() {
53 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
54
55 _old_gen = heap->old_gen();
56 _young_space = heap->young_gen()->to_space();
57
58 const uint promotion_manager_num = ParallelGCThreads;
59
60 // To prevent false sharing, we pad the PSPromotionManagers
61 // and make sure that the first instance starts at a cache line.
62 assert(_manager_array == nullptr, "Attempt to initialize twice");
63 _manager_array = PaddedArray<PSPromotionManager, mtGC>::create_unfreeable(promotion_manager_num);
64
65 _stack_array_depth = new PSScannerTasksQueueSet(ParallelGCThreads);
66
67 // Create and register the PSPromotionManager(s) for the worker threads.
68 for(uint i=0; i<ParallelGCThreads; i++) {
69 stack_array_depth()->register_queue(i, _manager_array[i].claimed_stack_depth());
70 }
71 // The VMThread gets its own PSPromotionManager, which is not available
72 // for work stealing.
73
74 assert(_preserved_marks_set == nullptr, "Attempt to initialize twice");
75 _preserved_marks_set = new PreservedMarksSet(true /* in_c_heap */);
76 _preserved_marks_set->init(promotion_manager_num);
77 for (uint i = 0; i < promotion_manager_num; i += 1) {
78 _manager_array[i].register_preserved_marks(_preserved_marks_set->get(i));
79 }
80 }
81
82 // Helper functions to get around the circular dependency between
83 // psScavenge.inline.hpp and psPromotionManager.inline.hpp.
84 bool PSPromotionManager::should_scavenge(oop* p, bool check_to_space) {
85 return PSScavenge::should_scavenge(p, check_to_space);
86 }
87 bool PSPromotionManager::should_scavenge(narrowOop* p, bool check_to_space) {
88 return PSScavenge::should_scavenge(p, check_to_space);
89 }
90
91 PSPromotionManager* PSPromotionManager::gc_thread_promotion_manager(uint index) {
92 assert(index < ParallelGCThreads, "index out of range");
93 assert(_manager_array != nullptr, "Sanity");
94 return &_manager_array[index];
95 }
96
97 PSPromotionManager* PSPromotionManager::vm_thread_promotion_manager() {
98 assert(_manager_array != nullptr, "Sanity");
99 return &_manager_array[0];
100 }
101
102 void PSPromotionManager::pre_scavenge() {
103 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
104
105 _preserved_marks_set->assert_empty();
106 _young_space = heap->young_gen()->to_space();
107
108 for(uint i=0; i<ParallelGCThreads; i++) {
109 manager_array(i)->reset();
110 }
111 }
112
113 bool PSPromotionManager::post_scavenge(YoungGCTracer& gc_tracer) {
114 bool promotion_failure_occurred = false;
115
116 TASKQUEUE_STATS_ONLY(print_taskqueue_stats());
117 for (uint i = 0; i < ParallelGCThreads; i++) {
118 PSPromotionManager* manager = manager_array(i);
119 assert(manager->claimed_stack_depth()->is_empty(), "should be empty");
120 if (manager->_promotion_failed_info.has_failed()) {
121 gc_tracer.report_promotion_failed(manager->_promotion_failed_info);
122 promotion_failure_occurred = true;
123 }
124 manager->flush_labs();
125 manager->flush_string_dedup_requests();
126 }
127 if (!promotion_failure_occurred) {
128 // If there was no promotion failure, the preserved mark stacks
129 // should be empty.
130 _preserved_marks_set->assert_empty();
131 }
132 return promotion_failure_occurred;
133 }
134
135 #if TASKQUEUE_STATS
136 void
137 PSPromotionManager::print_local_stats(outputStream* const out, uint i) const {
138 #define FMT " " SIZE_FORMAT_W(10)
139 out->print_cr("%3u" FMT FMT FMT FMT,
140 i, _array_chunk_pushes, _array_chunk_steals,
141 _arrays_chunked, _array_chunks_processed);
142 #undef FMT
143 }
144
145 static const char* const pm_stats_hdr[] = {
146 " ----partial array---- arrays array",
147 "thr push steal chunked chunks",
148 "--- ---------- ---------- ---------- ----------"
149 };
150
151 void PSPromotionManager::print_taskqueue_stats() {
152 if (!log_is_enabled(Trace, gc, task, stats)) {
153 return;
154 }
155 Log(gc, task, stats) log;
156 ResourceMark rm;
157 LogStream ls(log.trace());
158
159 stack_array_depth()->print_taskqueue_stats(&ls, "Oop Queue");
160
161 const uint hlines = sizeof(pm_stats_hdr) / sizeof(pm_stats_hdr[0]);
162 for (uint i = 0; i < hlines; ++i) ls.print_cr("%s", pm_stats_hdr[i]);
163 for (uint i = 0; i < ParallelGCThreads; ++i) {
164 manager_array(i)->print_local_stats(&ls, i);
165 }
166 }
167
168 void PSPromotionManager::reset_stats() {
169 claimed_stack_depth()->stats.reset();
170 _array_chunk_pushes = _array_chunk_steals = 0;
171 _arrays_chunked = _array_chunks_processed = 0;
172 }
173 #endif // TASKQUEUE_STATS
174
175 PSPromotionManager::PSPromotionManager() {
176 ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
177
178 // We set the old lab's start array.
179 _old_lab.set_start_array(old_gen()->start_array());
180
181 uint queue_size;
182 queue_size = claimed_stack_depth()->max_elems();
183
184 if (ParallelGCThreads == 1) {
185 _target_stack_size = 0;
186 } else {
187 // don't let the target stack size to be more than 1/4 of the entries
188 _target_stack_size = (uint) MIN2((uint) GCDrainStackTargetSize,
189 (uint) (queue_size / 4));
190 }
191
192 _array_chunk_size = ParGCArrayScanChunk;
193 // let's choose 1.5x the chunk size
194 _min_array_size_for_chunking = 3 * _array_chunk_size / 2;
195
196 _preserved_marks = nullptr;
197
198 reset();
199 }
200
201 void PSPromotionManager::reset() {
202 assert(stacks_empty(), "reset of non-empty stack");
203
204 // We need to get an assert in here to make sure the labs are always flushed.
205
206 // Do not prefill the LAB's, save heap wastage!
207 HeapWord* lab_base = young_space()->top();
208 _young_lab.initialize(MemRegion(lab_base, (size_t)0));
209 _young_gen_is_full = false;
210
211 lab_base = old_gen()->object_space()->top();
212 _old_lab.initialize(MemRegion(lab_base, (size_t)0));
213 _old_gen_is_full = false;
214
215 _promotion_failed_info.reset();
216
217 TASKQUEUE_STATS_ONLY(reset_stats());
218 }
219
220 void PSPromotionManager::register_preserved_marks(PreservedMarks* preserved_marks) {
221 assert(_preserved_marks == nullptr, "do not set it twice");
222 _preserved_marks = preserved_marks;
223 }
224
225 void PSPromotionManager::restore_preserved_marks() {
226 _preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
227 }
228
229 void PSPromotionManager::drain_stacks_depth(bool totally_drain) {
230 const uint threshold = totally_drain ? 0
231 : _target_stack_size;
232
233 PSScannerTasksQueue* const tq = claimed_stack_depth();
234 do {
235 ScannerTask task;
236
237 // Drain overflow stack first, so other threads can steal from
238 // claimed stack while we work.
239 while (tq->pop_overflow(task)) {
240 // In PSCardTable::scavenge_contents_parallel(), when work is distributed
241 // among different workers, an object is never split between multiple workers.
242 // Therefore, if a worker gets owned a large objArray, it may accumulate
243 // many tasks (corresponding to every element in this array) in its
244 // task queue. When there are too many overflow tasks, publishing them
245 // (via try_push_to_taskqueue()) can incur noticeable overhead in Young GC
246 // pause, so it is better to process them locally until large-objArray-splitting is implemented.
247 process_popped_location_depth(task);
248 }
249
250 while (tq->pop_local(task, threshold)) {
251 process_popped_location_depth(task);
252 }
253 } while (!tq->overflow_empty());
254
255 assert(!totally_drain || tq->taskqueue_empty(), "Sanity");
256 assert(totally_drain || tq->size() <= _target_stack_size, "Sanity");
257 assert(tq->overflow_empty(), "Sanity");
258 }
259
260 void PSPromotionManager::flush_labs() {
261 assert(stacks_empty(), "Attempt to flush lab with live stack");
262
263 // If either promotion lab fills up, we can flush the
264 // lab but not refill it, so check first.
265 assert(!_young_lab.is_flushed() || _young_gen_is_full, "Sanity");
266 if (!_young_lab.is_flushed())
267 _young_lab.flush();
268
269 assert(!_old_lab.is_flushed() || _old_gen_is_full, "Sanity");
270 if (!_old_lab.is_flushed())
271 _old_lab.flush();
272
273 // Let PSScavenge know if we overflowed
274 if (_young_gen_is_full) {
275 PSScavenge::set_survivor_overflow(true);
276 }
277 }
278
279 template <class T> void PSPromotionManager::process_array_chunk_work(
280 oop obj,
281 int start, int end) {
282 assert(start <= end, "invariant");
283 T* const base = (T*)objArrayOop(obj)->base();
284 T* p = base + start;
285 T* const chunk_end = base + end;
286 while (p < chunk_end) {
287 if (PSScavenge::should_scavenge(p)) {
288 claim_or_forward_depth(p);
289 }
290 ++p;
291 }
292 }
293
294 void PSPromotionManager::process_array_chunk(PartialArrayScanTask task) {
295 assert(PSChunkLargeArrays, "invariant");
296
297 oop old = task.to_source_array();
298 assert(old->forward_safe_klass()->is_objArray_klass(), "invariant");
299 assert(old->is_forwarded(), "invariant");
300
301 TASKQUEUE_STATS_ONLY(++_array_chunks_processed);
302
303 oop const obj = old->forwardee();
304
305 int start;
306 int const end = arrayOop(old)->length();
307 if (end > (int) _min_array_size_for_chunking) {
308 // we'll chunk more
309 start = end - _array_chunk_size;
310 assert(start > 0, "invariant");
311 arrayOop(old)->set_length(start);
312 push_depth(ScannerTask(PartialArrayScanTask(old)));
313 TASKQUEUE_STATS_ONLY(++_array_chunk_pushes);
314 } else {
315 // this is the final chunk for this array
316 start = 0;
317 int const actual_length = arrayOop(obj)->length();
318 arrayOop(old)->set_length(actual_length);
319 }
320
321 if (UseCompressedOops) {
322 process_array_chunk_work<narrowOop>(obj, start, end);
323 } else {
324 process_array_chunk_work<oop>(obj, start, end);
325 }
326 }
327
328 oop PSPromotionManager::oop_promotion_failed(oop obj, markWord obj_mark) {
329 assert(_old_gen_is_full || PromotionFailureALot, "Sanity");
330
331 // Attempt to CAS in the header.
332 // This tests if the header is still the same as when
333 // this started. If it is the same (i.e., no forwarding
334 // pointer has been installed), then this thread owns
335 // it.
336 if (obj->forward_to_self_atomic(obj_mark) == nullptr) {
337 // We won any races, we "own" this object.
338 assert(obj == obj->forwardee(), "Sanity");
339
340 _promotion_failed_info.register_copy_failure(obj->size());
341
342 ContinuationGCSupport::transform_stack_chunk(obj);
343
344 push_contents(obj);
345
346 // Save the markWord of promotion-failed objs in _preserved_marks for later
347 // restoration. This way we don't have to walk the young-gen to locate
348 // these promotion-failed objs.
349 _preserved_marks->push_always(obj, obj_mark);
350 } else {
351 // We lost, someone else "owns" this object
352 guarantee(obj->is_forwarded(), "Object must be forwarded if the cas failed.");
353
354 // No unallocation to worry about.
355 obj = obj->forwardee();
356 }
357
358 return obj;
359 }