Sat, 28 Mar 2009 15:47:29 -0700
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
Summary: When using compressed oops, rather than chaining the overflowed grey objects' pre-images through their klass words, we use GC-worker thread-local overflow stacks.
Reviewed-by: jcoomes, jmasa
1 /*
2 * Copyright 2001-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 class ChunkArray;
26 class ParScanWithoutBarrierClosure;
27 class ParScanWithBarrierClosure;
28 class ParRootScanWithoutBarrierClosure;
29 class ParRootScanWithBarrierTwoGensClosure;
30 class ParEvacuateFollowersClosure;
32 // It would be better if these types could be kept local to the .cpp file,
33 // but they must be here to allow ParScanClosure::do_oop_work to be defined
34 // in genOopClosures.inline.hpp.
36 typedef OopTaskQueue ObjToScanQueue;
37 typedef OopTaskQueueSet ObjToScanQueueSet;
39 // Enable this to get push/pop/steal stats.
40 const int PAR_STATS_ENABLED = 0;
42 class ParKeepAliveClosure: public DefNewGeneration::KeepAliveClosure {
43 private:
44 ParScanWeakRefClosure* _par_cl;
45 protected:
46 template <class T> void do_oop_work(T* p);
47 public:
48 ParKeepAliveClosure(ParScanWeakRefClosure* cl);
49 virtual void do_oop(oop* p);
50 virtual void do_oop(narrowOop* p);
51 };
53 // The state needed by thread performing parallel young-gen collection.
54 class ParScanThreadState {
55 friend class ParScanThreadStateSet;
56 private:
57 ObjToScanQueue *_work_queue;
58 GrowableArray<oop>* _overflow_stack;
60 ParGCAllocBuffer _to_space_alloc_buffer;
62 ParScanWithoutBarrierClosure _to_space_closure; // scan_without_gc_barrier
63 ParScanWithBarrierClosure _old_gen_closure; // scan_with_gc_barrier
64 ParRootScanWithoutBarrierClosure _to_space_root_closure; // scan_root_without_gc_barrier
65 // One of these two will be passed to process_strong_roots, which will
66 // set its generation. The first is for two-gen configs where the
67 // old gen collects the perm gen; the second is for arbitrary configs.
68 // The second isn't used right now (it used to be used for the train, an
69 // incremental collector) but the declaration has been left as a reminder.
70 ParRootScanWithBarrierTwoGensClosure _older_gen_closure;
71 // This closure will always be bound to the old gen; it will be used
72 // in evacuate_followers.
73 ParRootScanWithBarrierTwoGensClosure _old_gen_root_closure; // scan_old_root_with_gc_barrier
74 ParEvacuateFollowersClosure _evacuate_followers;
75 DefNewGeneration::IsAliveClosure _is_alive_closure;
76 ParScanWeakRefClosure _scan_weak_ref_closure;
77 ParKeepAliveClosure _keep_alive_closure;
80 Space* _to_space;
81 Space* to_space() { return _to_space; }
83 ParNewGeneration* _young_gen;
84 ParNewGeneration* young_gen() const { return _young_gen; }
86 Generation* _old_gen;
87 Generation* old_gen() { return _old_gen; }
89 HeapWord *_young_old_boundary;
91 int _hash_seed;
92 int _thread_num;
93 ageTable _ageTable;
95 bool _to_space_full;
97 int _pushes, _pops, _steals, _steal_attempts, _term_attempts;
98 int _overflow_pushes, _overflow_refills, _overflow_refill_objs;
100 // Timing numbers.
101 double _start;
102 double _start_strong_roots;
103 double _strong_roots_time;
104 double _start_term;
105 double _term_time;
107 // Helper for trim_queues. Scans subset of an array and makes
108 // remainder available for work stealing.
109 void scan_partial_array_and_push_remainder(oop obj);
111 // In support of CMS' parallel rescan of survivor space.
112 ChunkArray* _survivor_chunk_array;
113 ChunkArray* survivor_chunk_array() { return _survivor_chunk_array; }
115 void record_survivor_plab(HeapWord* plab_start, size_t plab_word_size);
117 ParScanThreadState(Space* to_space_, ParNewGeneration* gen_,
118 Generation* old_gen_, int thread_num_,
119 ObjToScanQueueSet* work_queue_set_, size_t desired_plab_sz_,
120 ParallelTaskTerminator& term_);
122 public:
123 ageTable* age_table() {return &_ageTable;}
125 ObjToScanQueue* work_queue() { return _work_queue; }
127 ParGCAllocBuffer* to_space_alloc_buffer() {
128 return &_to_space_alloc_buffer;
129 }
131 ParEvacuateFollowersClosure& evacuate_followers_closure() { return _evacuate_followers; }
132 DefNewGeneration::IsAliveClosure& is_alive_closure() { return _is_alive_closure; }
133 ParScanWeakRefClosure& scan_weak_ref_closure() { return _scan_weak_ref_closure; }
134 ParKeepAliveClosure& keep_alive_closure() { return _keep_alive_closure; }
135 ParScanClosure& older_gen_closure() { return _older_gen_closure; }
136 ParRootScanWithoutBarrierClosure& to_space_root_closure() { return _to_space_root_closure; };
138 // Decrease queue size below "max_size".
139 void trim_queues(int max_size);
141 // Private overflow stack usage
142 GrowableArray<oop>* overflow_stack() { return _overflow_stack; }
143 bool take_from_overflow_stack();
144 void push_on_overflow_stack(oop p);
146 // Is new_obj a candidate for scan_partial_array_and_push_remainder method.
147 inline bool should_be_partially_scanned(oop new_obj, oop old_obj) const;
149 int* hash_seed() { return &_hash_seed; }
150 int thread_num() { return _thread_num; }
152 // Allocate a to-space block of size "sz", or else return NULL.
153 HeapWord* alloc_in_to_space_slow(size_t word_sz);
155 HeapWord* alloc_in_to_space(size_t word_sz) {
156 HeapWord* obj = to_space_alloc_buffer()->allocate(word_sz);
157 if (obj != NULL) return obj;
158 else return alloc_in_to_space_slow(word_sz);
159 }
161 HeapWord* young_old_boundary() { return _young_old_boundary; }
163 void set_young_old_boundary(HeapWord *boundary) {
164 _young_old_boundary = boundary;
165 }
167 // Undo the most recent allocation ("obj", of "word_sz").
168 void undo_alloc_in_to_space(HeapWord* obj, size_t word_sz);
170 int pushes() { return _pushes; }
171 int pops() { return _pops; }
172 int steals() { return _steals; }
173 int steal_attempts() { return _steal_attempts; }
174 int term_attempts() { return _term_attempts; }
175 int overflow_pushes() { return _overflow_pushes; }
176 int overflow_refills() { return _overflow_refills; }
177 int overflow_refill_objs() { return _overflow_refill_objs; }
179 void note_push() { if (PAR_STATS_ENABLED) _pushes++; }
180 void note_pop() { if (PAR_STATS_ENABLED) _pops++; }
181 void note_steal() { if (PAR_STATS_ENABLED) _steals++; }
182 void note_steal_attempt() { if (PAR_STATS_ENABLED) _steal_attempts++; }
183 void note_term_attempt() { if (PAR_STATS_ENABLED) _term_attempts++; }
184 void note_overflow_push() { if (PAR_STATS_ENABLED) _overflow_pushes++; }
185 void note_overflow_refill(int objs) {
186 if (PAR_STATS_ENABLED) {
187 _overflow_refills++;
188 _overflow_refill_objs += objs;
189 }
190 }
192 void start_strong_roots() {
193 _start_strong_roots = os::elapsedTime();
194 }
195 void end_strong_roots() {
196 _strong_roots_time += (os::elapsedTime() - _start_strong_roots);
197 }
198 double strong_roots_time() { return _strong_roots_time; }
199 void start_term_time() {
200 note_term_attempt();
201 _start_term = os::elapsedTime();
202 }
203 void end_term_time() {
204 _term_time += (os::elapsedTime() - _start_term);
205 }
206 double term_time() { return _term_time; }
208 double elapsed() {
209 return os::elapsedTime() - _start;
210 }
211 };
213 class ParNewGenTask: public AbstractGangTask {
214 private:
215 ParNewGeneration* _gen;
216 Generation* _next_gen;
217 HeapWord* _young_old_boundary;
218 class ParScanThreadStateSet* _state_set;
220 public:
221 ParNewGenTask(ParNewGeneration* gen,
222 Generation* next_gen,
223 HeapWord* young_old_boundary,
224 ParScanThreadStateSet* state_set);
226 HeapWord* young_old_boundary() { return _young_old_boundary; }
228 void work(int i);
229 };
231 class KeepAliveClosure: public DefNewGeneration::KeepAliveClosure {
232 protected:
233 template <class T> void do_oop_work(T* p);
234 public:
235 KeepAliveClosure(ScanWeakRefClosure* cl);
236 virtual void do_oop(oop* p);
237 virtual void do_oop(narrowOop* p);
238 };
240 class EvacuateFollowersClosureGeneral: public VoidClosure {
241 private:
242 GenCollectedHeap* _gch;
243 int _level;
244 OopsInGenClosure* _scan_cur_or_nonheap;
245 OopsInGenClosure* _scan_older;
246 public:
247 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
248 OopsInGenClosure* cur,
249 OopsInGenClosure* older);
250 virtual void do_void();
251 };
253 // Closure for scanning ParNewGeneration.
254 // Same as ScanClosure, except does parallel GC barrier.
255 class ScanClosureWithParBarrier: public ScanClosure {
256 protected:
257 template <class T> void do_oop_work(T* p);
258 public:
259 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier);
260 virtual void do_oop(oop* p);
261 virtual void do_oop(narrowOop* p);
262 };
264 // Implements AbstractRefProcTaskExecutor for ParNew.
265 class ParNewRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
266 private:
267 ParNewGeneration& _generation;
268 ParScanThreadStateSet& _state_set;
269 public:
270 ParNewRefProcTaskExecutor(ParNewGeneration& generation,
271 ParScanThreadStateSet& state_set)
272 : _generation(generation), _state_set(state_set)
273 { }
275 // Executes a task using worker threads.
276 virtual void execute(ProcessTask& task);
277 virtual void execute(EnqueueTask& task);
278 // Switch to single threaded mode.
279 virtual void set_single_threaded_mode();
280 };
283 // A Generation that does parallel young-gen collection.
285 class ParNewGeneration: public DefNewGeneration {
286 friend class ParNewGenTask;
287 friend class ParNewRefProcTask;
288 friend class ParNewRefProcTaskExecutor;
289 friend class ParScanThreadStateSet;
290 friend class ParEvacuateFollowersClosure;
292 private:
293 // XXX use a global constant instead of 64!
294 struct ObjToScanQueuePadded {
295 ObjToScanQueue work_queue;
296 char pad[64 - sizeof(ObjToScanQueue)]; // prevent false sharing
297 };
299 // The per-thread work queues, available here for stealing.
300 ObjToScanQueueSet* _task_queues;
302 // Desired size of survivor space plab's
303 PLABStats _plab_stats;
305 // A list of from-space images of to-be-scanned objects, threaded through
306 // klass-pointers (klass information already copied to the forwarded
307 // image.) Manipulated with CAS.
308 oop _overflow_list;
309 NOT_PRODUCT(ssize_t _num_par_pushes;)
311 // If true, older generation does not support promotion undo, so avoid.
312 static bool _avoid_promotion_undo;
314 // This closure is used by the reference processor to filter out
315 // references to live referent.
316 DefNewGeneration::IsAliveClosure _is_alive_closure;
318 static oop real_forwardee_slow(oop obj);
319 static void waste_some_time();
321 // Preserve the mark of "obj", if necessary, in preparation for its mark
322 // word being overwritten with a self-forwarding-pointer.
323 void preserve_mark_if_necessary(oop obj, markOop m);
325 protected:
327 bool _survivor_overflow;
329 bool avoid_promotion_undo() { return _avoid_promotion_undo; }
330 void set_avoid_promotion_undo(bool v) { _avoid_promotion_undo = v; }
332 bool survivor_overflow() { return _survivor_overflow; }
333 void set_survivor_overflow(bool v) { _survivor_overflow = v; }
335 // Adjust the tenuring threshold. See the implementation for
336 // the details of the policy.
337 virtual void adjust_desired_tenuring_threshold();
339 public:
340 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level);
342 ~ParNewGeneration() {
343 for (uint i = 0; i < ParallelGCThreads; i++)
344 delete _task_queues->queue(i);
346 delete _task_queues;
347 }
349 virtual void ref_processor_init();
350 virtual Generation::Name kind() { return Generation::ParNew; }
351 virtual const char* name() const;
352 virtual const char* short_name() const { return "ParNew"; }
354 // override
355 virtual bool refs_discovery_is_mt() const {
356 assert(UseParNewGC, "ParNewGeneration only when UseParNewGC");
357 return ParallelGCThreads > 1;
358 }
360 // Make the collection virtual.
361 virtual void collect(bool full,
362 bool clear_all_soft_refs,
363 size_t size,
364 bool is_tlab);
366 // This needs to be visible to the closure function.
367 // "obj" is the object to be copied, "m" is a recent value of its mark
368 // that must not contain a forwarding pointer (though one might be
369 // inserted in "obj"s mark word by a parallel thread).
370 inline oop copy_to_survivor_space(ParScanThreadState* par_scan_state,
371 oop obj, size_t obj_sz, markOop m) {
372 if (_avoid_promotion_undo) {
373 return copy_to_survivor_space_avoiding_promotion_undo(par_scan_state,
374 obj, obj_sz, m);
375 }
377 return copy_to_survivor_space_with_undo(par_scan_state, obj, obj_sz, m);
378 }
380 oop copy_to_survivor_space_avoiding_promotion_undo(ParScanThreadState* par_scan_state,
381 oop obj, size_t obj_sz, markOop m);
383 oop copy_to_survivor_space_with_undo(ParScanThreadState* par_scan_state,
384 oop obj, size_t obj_sz, markOop m);
386 // in support of testing overflow code
387 NOT_PRODUCT(int _overflow_counter;)
388 NOT_PRODUCT(bool should_simulate_overflow();)
390 // Accessor for overflow list
391 oop overflow_list() { return _overflow_list; }
393 // Push the given (from-space) object on the global overflow list.
394 void push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state);
396 // If the global overflow list is non-empty, move some tasks from it
397 // onto "work_q" (which need not be empty). No more than 1/4 of the
398 // available space on "work_q" is used.
399 bool take_from_overflow_list(ParScanThreadState* par_scan_state);
400 bool take_from_overflow_list_work(ParScanThreadState* par_scan_state);
402 // The task queues to be used by parallel GC threads.
403 ObjToScanQueueSet* task_queues() {
404 return _task_queues;
405 }
407 PLABStats* plab_stats() {
408 return &_plab_stats;
409 }
411 size_t desired_plab_sz() {
412 return _plab_stats.desired_plab_sz();
413 }
415 static oop real_forwardee(oop obj);
417 DEBUG_ONLY(static bool is_legal_forward_ptr(oop p);)
418 };