Print this page
rev 2896 : 7121547: G1: High number mispredicted branches while iterating over the marking bitmap
Summary: There is a high number of mispredicted branches associated with calling BitMap::iteratate() from within CMBitMapRO::iterate(). Implement a version of CMBitMapRO::iterate() directly using inline-able routines.
Reviewed-by:
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/gc_implementation/g1/concurrentMark.hpp
+++ new/src/share/vm/gc_implementation/g1/concurrentMark.hpp
1 1 /*
2 2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
26 26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
27 27
28 28 #include "gc_implementation/g1/heapRegionSets.hpp"
29 29 #include "utilities/taskqueue.hpp"
30 30
31 31 class G1CollectedHeap;
32 32 class CMTask;
33 33 typedef GenericTaskQueue<oop> CMTaskQueue;
34 34 typedef GenericTaskQueueSet<CMTaskQueue> CMTaskQueueSet;
35 35
36 36 // Closure used by CM during concurrent reference discovery
37 37 // and reference processing (during remarking) to determine
38 38 // if a particular object is alive. It is primarily used
39 39 // to determine if referents of discovered reference objects
40 40 // are alive. An instance is also embedded into the
41 41 // reference processor as the _is_alive_non_header field
42 42 class G1CMIsAliveClosure: public BoolObjectClosure {
43 43 G1CollectedHeap* _g1;
44 44 public:
45 45 G1CMIsAliveClosure(G1CollectedHeap* g1) :
46 46 _g1(g1)
47 47 {}
48 48
49 49 void do_object(oop obj) {
50 50 ShouldNotCallThis();
51 51 }
52 52 bool do_object_b(oop obj);
53 53 };
54 54
55 55 // A generic CM bit map. This is essentially a wrapper around the BitMap
56 56 // class, with one bit per (1<<_shifter) HeapWords.
57 57
58 58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
59 59 protected:
60 60 HeapWord* _bmStartWord; // base address of range covered by map
61 61 size_t _bmWordSize; // map size (in #HeapWords covered)
62 62 const int _shifter; // map to char or bit
63 63 VirtualSpace _virtual_space; // underlying the bit map
64 64 BitMap _bm; // the bit map itself
65 65
66 66 public:
67 67 // constructor
68 68 CMBitMapRO(ReservedSpace rs, int shifter);
69 69
70 70 enum { do_yield = true };
71 71
72 72 // inquiries
73 73 HeapWord* startWord() const { return _bmStartWord; }
74 74 size_t sizeInWords() const { return _bmWordSize; }
75 75 // the following is one past the last word in space
76 76 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
↓ open down ↓ |
76 lines elided |
↑ open up ↑ |
77 77
78 78 // read marks
79 79
80 80 bool isMarked(HeapWord* addr) const {
81 81 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
82 82 "outside underlying space?");
83 83 return _bm.at(heapWordToOffset(addr));
84 84 }
85 85
86 86 // iteration
87 - bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }
88 - bool iterate(BitMapClosure* cl, MemRegion mr);
87 + inline bool iterate(BitMapClosure* cl, MemRegion mr);
88 + inline bool iterate(BitMapClosure* cl);
89 89
90 90 // Return the address corresponding to the next marked bit at or after
91 91 // "addr", and before "limit", if "limit" is non-NULL. If there is no
92 92 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
93 93 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
94 94 HeapWord* limit = NULL) const;
95 95 // Return the address corresponding to the next unmarked bit at or after
96 96 // "addr", and before "limit", if "limit" is non-NULL. If there is no
97 97 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
98 98 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
99 99 HeapWord* limit = NULL) const;
100 100
101 101 // conversion utilities
102 102 // XXX Fix these so that offsets are size_t's...
103 103 HeapWord* offsetToHeapWord(size_t offset) const {
104 104 return _bmStartWord + (offset << _shifter);
105 105 }
106 106 size_t heapWordToOffset(HeapWord* addr) const {
107 107 return pointer_delta(addr, _bmStartWord) >> _shifter;
108 108 }
109 109 int heapWordDiffToOffsetDiff(size_t diff) const;
110 110 HeapWord* nextWord(HeapWord* addr) {
111 111 return offsetToHeapWord(heapWordToOffset(addr) + 1);
112 112 }
113 113
114 114 void mostly_disjoint_range_union(BitMap* from_bitmap,
115 115 size_t from_start_index,
116 116 HeapWord* to_start_word,
117 117 size_t word_num);
118 118
119 119 // debugging
120 120 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
121 121 };
122 122
123 123 class CMBitMap : public CMBitMapRO {
124 124
125 125 public:
126 126 // constructor
127 127 CMBitMap(ReservedSpace rs, int shifter) :
128 128 CMBitMapRO(rs, shifter) {}
129 129
130 130 // write marks
131 131 void mark(HeapWord* addr) {
132 132 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
133 133 "outside underlying space?");
134 134 _bm.set_bit(heapWordToOffset(addr));
135 135 }
136 136 void clear(HeapWord* addr) {
137 137 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
138 138 "outside underlying space?");
139 139 _bm.clear_bit(heapWordToOffset(addr));
140 140 }
141 141 bool parMark(HeapWord* addr) {
142 142 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
143 143 "outside underlying space?");
144 144 return _bm.par_set_bit(heapWordToOffset(addr));
145 145 }
146 146 bool parClear(HeapWord* addr) {
147 147 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
148 148 "outside underlying space?");
149 149 return _bm.par_clear_bit(heapWordToOffset(addr));
150 150 }
151 151 void markRange(MemRegion mr);
152 152 void clearAll();
153 153 void clearRange(MemRegion mr);
154 154
155 155 // Starting at the bit corresponding to "addr" (inclusive), find the next
156 156 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
157 157 // the end of this run (stopping at "end_addr"). Return the MemRegion
158 158 // covering from the start of the region corresponding to the first bit
159 159 // of the run to the end of the region corresponding to the last bit of
160 160 // the run. If there is no "1" bit at or after "addr", return an empty
161 161 // MemRegion.
162 162 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
163 163 };
164 164
165 165 // Represents a marking stack used by the CM collector.
166 166 // Ideally this should be GrowableArray<> just like MSC's marking stack(s).
167 167 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
168 168 ConcurrentMark* _cm;
169 169 oop* _base; // bottom of stack
170 170 jint _index; // one more than last occupied index
171 171 jint _capacity; // max #elements
172 172 jint _oops_do_bound; // Number of elements to include in next iteration.
173 173 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
174 174
175 175 bool _overflow;
176 176 DEBUG_ONLY(bool _drain_in_progress;)
177 177 DEBUG_ONLY(bool _drain_in_progress_yields;)
178 178
179 179 public:
180 180 CMMarkStack(ConcurrentMark* cm);
181 181 ~CMMarkStack();
182 182
183 183 void allocate(size_t size);
184 184
185 185 oop pop() {
186 186 if (!isEmpty()) {
187 187 return _base[--_index] ;
188 188 }
189 189 return NULL;
190 190 }
191 191
192 192 // If overflow happens, don't do the push, and record the overflow.
193 193 // *Requires* that "ptr" is already marked.
194 194 void push(oop ptr) {
195 195 if (isFull()) {
196 196 // Record overflow.
197 197 _overflow = true;
198 198 return;
199 199 } else {
200 200 _base[_index++] = ptr;
201 201 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
202 202 }
203 203 }
204 204 // Non-block impl. Note: concurrency is allowed only with other
205 205 // "par_push" operations, not with "pop" or "drain". We would need
206 206 // parallel versions of them if such concurrency was desired.
207 207 void par_push(oop ptr);
208 208
209 209 // Pushes the first "n" elements of "ptr_arr" on the stack.
210 210 // Non-block impl. Note: concurrency is allowed only with other
211 211 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
212 212 void par_adjoin_arr(oop* ptr_arr, int n);
213 213
214 214 // Pushes the first "n" elements of "ptr_arr" on the stack.
215 215 // Locking impl: concurrency is allowed only with
216 216 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
217 217 // locking strategy.
218 218 void par_push_arr(oop* ptr_arr, int n);
219 219
220 220 // If returns false, the array was empty. Otherwise, removes up to "max"
221 221 // elements from the stack, and transfers them to "ptr_arr" in an
222 222 // unspecified order. The actual number transferred is given in "n" ("n
223 223 // == 0" is deliberately redundant with the return value.) Locking impl:
224 224 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
225 225 // operations, which use the same locking strategy.
226 226 bool par_pop_arr(oop* ptr_arr, int max, int* n);
227 227
228 228 // Drain the mark stack, applying the given closure to all fields of
229 229 // objects on the stack. (That is, continue until the stack is empty,
230 230 // even if closure applications add entries to the stack.) The "bm"
231 231 // argument, if non-null, may be used to verify that only marked objects
232 232 // are on the mark stack. If "yield_after" is "true", then the
233 233 // concurrent marker performing the drain offers to yield after
234 234 // processing each object. If a yield occurs, stops the drain operation
235 235 // and returns false. Otherwise, returns true.
236 236 template<class OopClosureClass>
237 237 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
238 238
239 239 bool isEmpty() { return _index == 0; }
240 240 bool isFull() { return _index == _capacity; }
241 241 int maxElems() { return _capacity; }
242 242
243 243 bool overflow() { return _overflow; }
244 244 void clear_overflow() { _overflow = false; }
245 245
246 246 int size() { return _index; }
247 247
248 248 void setEmpty() { _index = 0; clear_overflow(); }
249 249
250 250 // Record the current size; a subsequent "oops_do" will iterate only over
251 251 // indices valid at the time of this call.
252 252 void set_oops_do_bound(jint bound = -1) {
253 253 if (bound == -1) {
254 254 _oops_do_bound = _index;
255 255 } else {
256 256 _oops_do_bound = bound;
257 257 }
258 258 }
259 259 jint oops_do_bound() { return _oops_do_bound; }
260 260 // iterate over the oops in the mark stack, up to the bound recorded via
261 261 // the call above.
262 262 void oops_do(OopClosure* f);
263 263 };
264 264
265 265 class CMRegionStack VALUE_OBJ_CLASS_SPEC {
266 266 MemRegion* _base;
267 267 jint _capacity;
268 268 jint _index;
269 269 jint _oops_do_bound;
270 270 bool _overflow;
271 271 public:
272 272 CMRegionStack();
273 273 ~CMRegionStack();
274 274 void allocate(size_t size);
275 275
276 276 // This is lock-free; assumes that it will only be called in parallel
277 277 // with other "push" operations (no pops).
278 278 void push_lock_free(MemRegion mr);
279 279
280 280 // Lock-free; assumes that it will only be called in parallel
281 281 // with other "pop" operations (no pushes).
282 282 MemRegion pop_lock_free();
283 283
284 284 #if 0
285 285 // The routines that manipulate the region stack with a lock are
286 286 // not currently used. They should be retained, however, as a
287 287 // diagnostic aid.
288 288
289 289 // These two are the implementations that use a lock. They can be
290 290 // called concurrently with each other but they should not be called
291 291 // concurrently with the lock-free versions (push() / pop()).
292 292 void push_with_lock(MemRegion mr);
293 293 MemRegion pop_with_lock();
294 294 #endif
295 295
296 296 bool isEmpty() { return _index == 0; }
297 297 bool isFull() { return _index == _capacity; }
298 298
299 299 bool overflow() { return _overflow; }
300 300 void clear_overflow() { _overflow = false; }
301 301
302 302 int size() { return _index; }
303 303
304 304 // It iterates over the entries in the region stack and it
305 305 // invalidates (i.e. assigns MemRegion()) the ones that point to
306 306 // regions in the collection set.
307 307 bool invalidate_entries_into_cset();
308 308
309 309 // This gives an upper bound up to which the iteration in
310 310 // invalidate_entries_into_cset() will reach. This prevents
311 311 // newly-added entries to be unnecessarily scanned.
312 312 void set_oops_do_bound() {
313 313 _oops_do_bound = _index;
314 314 }
315 315
316 316 void setEmpty() { _index = 0; clear_overflow(); }
317 317 };
318 318
319 319 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
320 320 private:
321 321 #ifndef PRODUCT
322 322 uintx _num_remaining;
323 323 bool _force;
324 324 #endif // !defined(PRODUCT)
325 325
326 326 public:
327 327 void init() PRODUCT_RETURN;
328 328 void update() PRODUCT_RETURN;
329 329 bool should_force() PRODUCT_RETURN_( return false; );
330 330 };
331 331
332 332 // this will enable a variety of different statistics per GC task
333 333 #define _MARKING_STATS_ 0
334 334 // this will enable the higher verbose levels
335 335 #define _MARKING_VERBOSE_ 0
336 336
337 337 #if _MARKING_STATS_
338 338 #define statsOnly(statement) \
339 339 do { \
340 340 statement ; \
341 341 } while (0)
342 342 #else // _MARKING_STATS_
343 343 #define statsOnly(statement) \
344 344 do { \
345 345 } while (0)
346 346 #endif // _MARKING_STATS_
347 347
348 348 typedef enum {
349 349 no_verbose = 0, // verbose turned off
350 350 stats_verbose, // only prints stats at the end of marking
351 351 low_verbose, // low verbose, mostly per region and per major event
352 352 medium_verbose, // a bit more detailed than low
353 353 high_verbose // per object verbose
354 354 } CMVerboseLevel;
355 355
356 356
357 357 class ConcurrentMarkThread;
358 358
359 359 class ConcurrentMark: public CHeapObj {
360 360 friend class ConcurrentMarkThread;
361 361 friend class CMTask;
362 362 friend class CMBitMapClosure;
363 363 friend class CSetMarkOopClosure;
364 364 friend class CMGlobalObjectClosure;
365 365 friend class CMRemarkTask;
366 366 friend class CMConcurrentMarkingTask;
367 367 friend class G1ParNoteEndTask;
368 368 friend class CalcLiveObjectsClosure;
369 369 friend class G1CMRefProcTaskProxy;
370 370 friend class G1CMRefProcTaskExecutor;
371 371 friend class G1CMParKeepAliveAndDrainClosure;
372 372 friend class G1CMParDrainMarkingStackClosure;
373 373
374 374 protected:
375 375 ConcurrentMarkThread* _cmThread; // the thread doing the work
376 376 G1CollectedHeap* _g1h; // the heap.
377 377 size_t _parallel_marking_threads; // the number of marking
378 378 // threads we're use
379 379 size_t _max_parallel_marking_threads; // max number of marking
380 380 // threads we'll ever use
381 381 double _sleep_factor; // how much we have to sleep, with
382 382 // respect to the work we just did, to
383 383 // meet the marking overhead goal
384 384 double _marking_task_overhead; // marking target overhead for
385 385 // a single task
386 386
387 387 // same as the two above, but for the cleanup task
388 388 double _cleanup_sleep_factor;
389 389 double _cleanup_task_overhead;
390 390
391 391 FreeRegionList _cleanup_list;
392 392
393 393 // CMS marking support structures
394 394 CMBitMap _markBitMap1;
395 395 CMBitMap _markBitMap2;
396 396 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
397 397 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
398 398 bool _at_least_one_mark_complete;
399 399
400 400 BitMap _region_bm;
401 401 BitMap _card_bm;
402 402
403 403 // Heap bounds
404 404 HeapWord* _heap_start;
405 405 HeapWord* _heap_end;
406 406
407 407 // For gray objects
408 408 CMMarkStack _markStack; // Grey objects behind global finger.
409 409 CMRegionStack _regionStack; // Grey regions behind global finger.
410 410 HeapWord* volatile _finger; // the global finger, region aligned,
411 411 // always points to the end of the
412 412 // last claimed region
413 413
414 414 // marking tasks
415 415 size_t _max_task_num; // maximum task number
416 416 size_t _active_tasks; // task num currently active
417 417 CMTask** _tasks; // task queue array (max_task_num len)
418 418 CMTaskQueueSet* _task_queues; // task queue set
419 419 ParallelTaskTerminator _terminator; // for termination
420 420
421 421 // Two sync barriers that are used to synchronise tasks when an
422 422 // overflow occurs. The algorithm is the following. All tasks enter
423 423 // the first one to ensure that they have all stopped manipulating
424 424 // the global data structures. After they exit it, they re-initialise
425 425 // their data structures and task 0 re-initialises the global data
426 426 // structures. Then, they enter the second sync barrier. This
427 427 // ensure, that no task starts doing work before all data
428 428 // structures (local and global) have been re-initialised. When they
429 429 // exit it, they are free to start working again.
430 430 WorkGangBarrierSync _first_overflow_barrier_sync;
431 431 WorkGangBarrierSync _second_overflow_barrier_sync;
432 432
433 433
434 434 // this is set by any task, when an overflow on the global data
435 435 // structures is detected.
436 436 volatile bool _has_overflown;
437 437 // true: marking is concurrent, false: we're in remark
438 438 volatile bool _concurrent;
439 439 // set at the end of a Full GC so that marking aborts
440 440 volatile bool _has_aborted;
441 441
442 442 // used when remark aborts due to an overflow to indicate that
443 443 // another concurrent marking phase should start
444 444 volatile bool _restart_for_overflow;
445 445
446 446 // This is true from the very start of concurrent marking until the
447 447 // point when all the tasks complete their work. It is really used
448 448 // to determine the points between the end of concurrent marking and
449 449 // time of remark.
450 450 volatile bool _concurrent_marking_in_progress;
451 451
452 452 // verbose level
453 453 CMVerboseLevel _verbose_level;
454 454
455 455 // These two fields are used to implement the optimisation that
456 456 // avoids pushing objects on the global/region stack if there are
457 457 // no collection set regions above the lowest finger.
458 458
459 459 // This is the lowest finger (among the global and local fingers),
460 460 // which is calculated before a new collection set is chosen.
461 461 HeapWord* _min_finger;
462 462 // If this flag is true, objects/regions that are marked below the
463 463 // finger should be pushed on the stack(s). If this is flag is
464 464 // false, it is safe not to push them on the stack(s).
465 465 bool _should_gray_objects;
466 466
467 467 // All of these times are in ms.
468 468 NumberSeq _init_times;
469 469 NumberSeq _remark_times;
470 470 NumberSeq _remark_mark_times;
471 471 NumberSeq _remark_weak_ref_times;
472 472 NumberSeq _cleanup_times;
473 473 double _total_counting_time;
474 474 double _total_rs_scrub_time;
475 475
476 476 double* _accum_task_vtime; // accumulated task vtime
477 477
478 478 FlexibleWorkGang* _parallel_workers;
479 479
480 480 ForceOverflowSettings _force_overflow_conc;
481 481 ForceOverflowSettings _force_overflow_stw;
482 482
483 483 void weakRefsWork(bool clear_all_soft_refs);
484 484
485 485 void swapMarkBitMaps();
486 486
487 487 // It resets the global marking data structures, as well as the
488 488 // task local ones; should be called during initial mark.
489 489 void reset();
490 490 // It resets all the marking data structures.
491 491 void clear_marking_state(bool clear_overflow = true);
492 492
493 493 // It should be called to indicate which phase we're in (concurrent
494 494 // mark or remark) and how many threads are currently active.
495 495 void set_phase(size_t active_tasks, bool concurrent);
496 496 // We do this after we're done with marking so that the marking data
497 497 // structures are initialised to a sensible and predictable state.
498 498 void set_non_marking_state();
499 499
500 500 // prints all gathered CM-related statistics
501 501 void print_stats();
502 502
503 503 bool cleanup_list_is_empty() {
504 504 return _cleanup_list.is_empty();
505 505 }
506 506
507 507 // accessor methods
508 508 size_t parallel_marking_threads() { return _parallel_marking_threads; }
509 509 size_t max_parallel_marking_threads() { return _max_parallel_marking_threads;}
510 510 double sleep_factor() { return _sleep_factor; }
511 511 double marking_task_overhead() { return _marking_task_overhead;}
512 512 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
513 513 double cleanup_task_overhead() { return _cleanup_task_overhead;}
514 514
515 515 HeapWord* finger() { return _finger; }
516 516 bool concurrent() { return _concurrent; }
517 517 size_t active_tasks() { return _active_tasks; }
518 518 ParallelTaskTerminator* terminator() { return &_terminator; }
519 519
520 520 // It claims the next available region to be scanned by a marking
521 521 // task. It might return NULL if the next region is empty or we have
522 522 // run out of regions. In the latter case, out_of_regions()
523 523 // determines whether we've really run out of regions or the task
524 524 // should call claim_region() again. This might seem a bit
525 525 // awkward. Originally, the code was written so that claim_region()
526 526 // either successfully returned with a non-empty region or there
527 527 // were no more regions to be claimed. The problem with this was
528 528 // that, in certain circumstances, it iterated over large chunks of
529 529 // the heap finding only empty regions and, while it was working, it
530 530 // was preventing the calling task to call its regular clock
531 531 // method. So, this way, each task will spend very little time in
532 532 // claim_region() and is allowed to call the regular clock method
533 533 // frequently.
534 534 HeapRegion* claim_region(int task);
535 535
536 536 // It determines whether we've run out of regions to scan.
537 537 bool out_of_regions() { return _finger == _heap_end; }
538 538
539 539 // Returns the task with the given id
540 540 CMTask* task(int id) {
541 541 assert(0 <= id && id < (int) _active_tasks,
542 542 "task id not within active bounds");
543 543 return _tasks[id];
544 544 }
545 545
546 546 // Returns the task queue with the given id
547 547 CMTaskQueue* task_queue(int id) {
548 548 assert(0 <= id && id < (int) _active_tasks,
549 549 "task queue id not within active bounds");
550 550 return (CMTaskQueue*) _task_queues->queue(id);
551 551 }
552 552
553 553 // Returns the task queue set
554 554 CMTaskQueueSet* task_queues() { return _task_queues; }
555 555
556 556 // Access / manipulation of the overflow flag which is set to
557 557 // indicate that the global stack or region stack has overflown
558 558 bool has_overflown() { return _has_overflown; }
559 559 void set_has_overflown() { _has_overflown = true; }
560 560 void clear_has_overflown() { _has_overflown = false; }
561 561
562 562 bool has_aborted() { return _has_aborted; }
563 563 bool restart_for_overflow() { return _restart_for_overflow; }
564 564
565 565 // Methods to enter the two overflow sync barriers
566 566 void enter_first_sync_barrier(int task_num);
567 567 void enter_second_sync_barrier(int task_num);
568 568
569 569 ForceOverflowSettings* force_overflow_conc() {
570 570 return &_force_overflow_conc;
571 571 }
572 572
573 573 ForceOverflowSettings* force_overflow_stw() {
574 574 return &_force_overflow_stw;
575 575 }
576 576
577 577 ForceOverflowSettings* force_overflow() {
578 578 if (concurrent()) {
579 579 return force_overflow_conc();
580 580 } else {
581 581 return force_overflow_stw();
582 582 }
583 583 }
584 584
585 585 public:
586 586 // Manipulation of the global mark stack.
587 587 // Notice that the first mark_stack_push is CAS-based, whereas the
588 588 // two below are Mutex-based. This is OK since the first one is only
589 589 // called during evacuation pauses and doesn't compete with the
590 590 // other two (which are called by the marking tasks during
591 591 // concurrent marking or remark).
592 592 bool mark_stack_push(oop p) {
593 593 _markStack.par_push(p);
594 594 if (_markStack.overflow()) {
595 595 set_has_overflown();
596 596 return false;
597 597 }
598 598 return true;
599 599 }
600 600 bool mark_stack_push(oop* arr, int n) {
601 601 _markStack.par_push_arr(arr, n);
602 602 if (_markStack.overflow()) {
603 603 set_has_overflown();
604 604 return false;
605 605 }
606 606 return true;
607 607 }
608 608 void mark_stack_pop(oop* arr, int max, int* n) {
609 609 _markStack.par_pop_arr(arr, max, n);
610 610 }
611 611 size_t mark_stack_size() { return _markStack.size(); }
612 612 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
613 613 bool mark_stack_overflow() { return _markStack.overflow(); }
614 614 bool mark_stack_empty() { return _markStack.isEmpty(); }
615 615
616 616 // (Lock-free) Manipulation of the region stack
617 617 bool region_stack_push_lock_free(MemRegion mr) {
618 618 // Currently we only call the lock-free version during evacuation
619 619 // pauses.
620 620 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
621 621
622 622 _regionStack.push_lock_free(mr);
623 623 if (_regionStack.overflow()) {
624 624 set_has_overflown();
625 625 return false;
626 626 }
627 627 return true;
628 628 }
629 629
630 630 // Lock-free version of region-stack pop. Should only be
631 631 // called in tandem with other lock-free pops.
632 632 MemRegion region_stack_pop_lock_free() {
633 633 return _regionStack.pop_lock_free();
634 634 }
635 635
636 636 #if 0
637 637 // The routines that manipulate the region stack with a lock are
638 638 // not currently used. They should be retained, however, as a
639 639 // diagnostic aid.
640 640
641 641 bool region_stack_push_with_lock(MemRegion mr) {
642 642 // Currently we only call the lock-based version during either
643 643 // concurrent marking or remark.
644 644 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
645 645 "if we are at a safepoint it should be the remark safepoint");
646 646
647 647 _regionStack.push_with_lock(mr);
648 648 if (_regionStack.overflow()) {
649 649 set_has_overflown();
650 650 return false;
651 651 }
652 652 return true;
653 653 }
654 654
655 655 MemRegion region_stack_pop_with_lock() {
656 656 // Currently we only call the lock-based version during either
657 657 // concurrent marking or remark.
658 658 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
659 659 "if we are at a safepoint it should be the remark safepoint");
660 660
661 661 return _regionStack.pop_with_lock();
662 662 }
663 663 #endif
664 664
665 665 int region_stack_size() { return _regionStack.size(); }
666 666 bool region_stack_overflow() { return _regionStack.overflow(); }
667 667 bool region_stack_empty() { return _regionStack.isEmpty(); }
668 668
669 669 // Iterate over any regions that were aborted while draining the
670 670 // region stack (any such regions are saved in the corresponding
671 671 // CMTask) and invalidate (i.e. assign to the empty MemRegion())
672 672 // any regions that point into the collection set.
673 673 bool invalidate_aborted_regions_in_cset();
674 674
675 675 // Returns true if there are any aborted memory regions.
676 676 bool has_aborted_regions();
677 677
678 678 bool concurrent_marking_in_progress() {
679 679 return _concurrent_marking_in_progress;
680 680 }
681 681 void set_concurrent_marking_in_progress() {
682 682 _concurrent_marking_in_progress = true;
683 683 }
684 684 void clear_concurrent_marking_in_progress() {
685 685 _concurrent_marking_in_progress = false;
686 686 }
687 687
688 688 void update_accum_task_vtime(int i, double vtime) {
689 689 _accum_task_vtime[i] += vtime;
690 690 }
691 691
692 692 double all_task_accum_vtime() {
693 693 double ret = 0.0;
694 694 for (int i = 0; i < (int)_max_task_num; ++i)
695 695 ret += _accum_task_vtime[i];
696 696 return ret;
697 697 }
698 698
699 699 // Attempts to steal an object from the task queues of other tasks
700 700 bool try_stealing(int task_num, int* hash_seed, oop& obj) {
701 701 return _task_queues->steal(task_num, hash_seed, obj);
702 702 }
703 703
704 704 // It grays an object by first marking it. Then, if it's behind the
705 705 // global finger, it also pushes it on the global stack.
706 706 void deal_with_reference(oop obj);
707 707
708 708 ConcurrentMark(ReservedSpace rs, int max_regions);
709 709 ~ConcurrentMark();
710 710 ConcurrentMarkThread* cmThread() { return _cmThread; }
711 711
712 712 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
713 713 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
714 714
715 715 // Returns the number of GC threads to be used in a concurrent
716 716 // phase based on the number of GC threads being used in a STW
717 717 // phase.
718 718 size_t scale_parallel_threads(size_t n_par_threads);
719 719
720 720 // Calculates the number of GC threads to be used in a concurrent phase.
721 721 size_t calc_parallel_marking_threads();
722 722
723 723 // The following three are interaction between CM and
724 724 // G1CollectedHeap
725 725
726 726 // This notifies CM that a root during initial-mark needs to be
727 727 // grayed and it's MT-safe. Currently, we just mark it. But, in the
728 728 // future, we can experiment with pushing it on the stack and we can
729 729 // do this without changing G1CollectedHeap.
730 730 void grayRoot(oop p);
731 731 // It's used during evacuation pauses to gray a region, if
732 732 // necessary, and it's MT-safe. It assumes that the caller has
733 733 // marked any objects on that region. If _should_gray_objects is
734 734 // true and we're still doing concurrent marking, the region is
735 735 // pushed on the region stack, if it is located below the global
736 736 // finger, otherwise we do nothing.
737 737 void grayRegionIfNecessary(MemRegion mr);
738 738 // It's used during evacuation pauses to mark and, if necessary,
739 739 // gray a single object and it's MT-safe. It assumes the caller did
740 740 // not mark the object. If _should_gray_objects is true and we're
741 741 // still doing concurrent marking, the objects is pushed on the
742 742 // global stack, if it is located below the global finger, otherwise
743 743 // we do nothing.
744 744 void markAndGrayObjectIfNecessary(oop p);
745 745
746 746 // It iterates over the heap and for each object it comes across it
747 747 // will dump the contents of its reference fields, as well as
748 748 // liveness information for the object and its referents. The dump
749 749 // will be written to a file with the following name:
750 750 // G1PrintReachableBaseFile + "." + str.
751 751 // vo decides whether the prev (vo == UsePrevMarking), the next
752 752 // (vo == UseNextMarking) marking information, or the mark word
753 753 // (vo == UseMarkWord) will be used to determine the liveness of
754 754 // each object / referent.
755 755 // If all is true, all objects in the heap will be dumped, otherwise
756 756 // only the live ones. In the dump the following symbols / breviations
757 757 // are used:
758 758 // M : an explicitly live object (its bitmap bit is set)
759 759 // > : an implicitly live object (over tams)
760 760 // O : an object outside the G1 heap (typically: in the perm gen)
761 761 // NOT : a reference field whose referent is not live
762 762 // AND MARKED : indicates that an object is both explicitly and
763 763 // implicitly live (it should be one or the other, not both)
764 764 void print_reachable(const char* str,
765 765 VerifyOption vo, bool all) PRODUCT_RETURN;
766 766
767 767 // Clear the next marking bitmap (will be called concurrently).
768 768 void clearNextBitmap();
769 769
770 770 // These two do the work that needs to be done before and after the
771 771 // initial root checkpoint. Since this checkpoint can be done at two
772 772 // different points (i.e. an explicit pause or piggy-backed on a
773 773 // young collection), then it's nice to be able to easily share the
774 774 // pre/post code. It might be the case that we can put everything in
775 775 // the post method. TP
776 776 void checkpointRootsInitialPre();
777 777 void checkpointRootsInitialPost();
778 778
779 779 // Do concurrent phase of marking, to a tentative transitive closure.
780 780 void markFromRoots();
781 781
782 782 // Process all unprocessed SATB buffers. It is called at the
783 783 // beginning of an evacuation pause.
784 784 void drainAllSATBBuffers();
785 785
786 786 void checkpointRootsFinal(bool clear_all_soft_refs);
787 787 void checkpointRootsFinalWork();
788 788 void calcDesiredRegions();
789 789 void cleanup();
790 790 void completeCleanup();
791 791
792 792 // Mark in the previous bitmap. NB: this is usually read-only, so use
793 793 // this carefully!
794 794 void markPrev(oop p);
795 795 void clear(oop p);
796 796 // Clears marks for all objects in the given range, for both prev and
797 797 // next bitmaps. NB: the previous bitmap is usually read-only, so use
798 798 // this carefully!
799 799 void clearRangeBothMaps(MemRegion mr);
800 800
801 801 // Record the current top of the mark and region stacks; a
802 802 // subsequent oops_do() on the mark stack and
803 803 // invalidate_entries_into_cset() on the region stack will iterate
804 804 // only over indices valid at the time of this call.
805 805 void set_oops_do_bound() {
806 806 _markStack.set_oops_do_bound();
807 807 _regionStack.set_oops_do_bound();
808 808 }
809 809 // Iterate over the oops in the mark stack and all local queues. It
810 810 // also calls invalidate_entries_into_cset() on the region stack.
811 811 void oops_do(OopClosure* f);
812 812 // It is called at the end of an evacuation pause during marking so
813 813 // that CM is notified of where the new end of the heap is. It
814 814 // doesn't do anything if concurrent_marking_in_progress() is false,
815 815 // unless the force parameter is true.
816 816 void update_g1_committed(bool force = false);
817 817
818 818 void complete_marking_in_collection_set();
819 819
820 820 // It indicates that a new collection set is being chosen.
821 821 void newCSet();
822 822
823 823 // It registers a collection set heap region with CM. This is used
824 824 // to determine whether any heap regions are located above the finger.
825 825 void registerCSetRegion(HeapRegion* hr);
826 826
827 827 // Resets the region fields of any active CMTask whose region fields
828 828 // are in the collection set (i.e. the region currently claimed by
829 829 // the CMTask will be evacuated and may be used, subsequently, as
830 830 // an alloc region). When this happens the region fields in the CMTask
831 831 // are stale and, hence, should be cleared causing the worker thread
832 832 // to claim a new region.
833 833 void reset_active_task_region_fields_in_cset();
834 834
835 835 // Registers the maximum region-end associated with a set of
836 836 // regions with CM. Again this is used to determine whether any
837 837 // heap regions are located above the finger.
838 838 void register_collection_set_finger(HeapWord* max_finger) {
839 839 // max_finger is the highest heap region end of the regions currently
840 840 // contained in the collection set. If this value is larger than
841 841 // _min_finger then we need to gray objects.
842 842 // This routine is like registerCSetRegion but for an entire
843 843 // collection of regions.
844 844 if (max_finger > _min_finger) {
845 845 _should_gray_objects = true;
846 846 }
847 847 }
848 848
849 849 // Returns "true" if at least one mark has been completed.
850 850 bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
851 851
852 852 bool isMarked(oop p) const {
853 853 assert(p != NULL && p->is_oop(), "expected an oop");
854 854 HeapWord* addr = (HeapWord*)p;
855 855 assert(addr >= _nextMarkBitMap->startWord() ||
856 856 addr < _nextMarkBitMap->endWord(), "in a region");
857 857
858 858 return _nextMarkBitMap->isMarked(addr);
859 859 }
860 860
861 861 inline bool not_yet_marked(oop p) const;
862 862
863 863 // XXX Debug code
864 864 bool containing_card_is_marked(void* p);
865 865 bool containing_cards_are_marked(void* start, void* last);
866 866
867 867 bool isPrevMarked(oop p) const {
868 868 assert(p != NULL && p->is_oop(), "expected an oop");
869 869 HeapWord* addr = (HeapWord*)p;
870 870 assert(addr >= _prevMarkBitMap->startWord() ||
871 871 addr < _prevMarkBitMap->endWord(), "in a region");
872 872
873 873 return _prevMarkBitMap->isMarked(addr);
874 874 }
875 875
876 876 inline bool do_yield_check(int worker_i = 0);
877 877 inline bool should_yield();
878 878
879 879 // Called to abort the marking cycle after a Full GC takes palce.
880 880 void abort();
881 881
882 882 // This prints the global/local fingers. It is used for debugging.
883 883 NOT_PRODUCT(void print_finger();)
884 884
885 885 void print_summary_info();
886 886
887 887 void print_worker_threads_on(outputStream* st) const;
888 888
889 889 // The following indicate whether a given verbose level has been
890 890 // set. Notice that anything above stats is conditional to
891 891 // _MARKING_VERBOSE_ having been set to 1
892 892 bool verbose_stats() {
893 893 return _verbose_level >= stats_verbose;
894 894 }
895 895 bool verbose_low() {
896 896 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
897 897 }
898 898 bool verbose_medium() {
899 899 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
900 900 }
901 901 bool verbose_high() {
902 902 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
903 903 }
904 904 };
905 905
906 906 // A class representing a marking task.
907 907 class CMTask : public TerminatorTerminator {
908 908 private:
909 909 enum PrivateConstants {
910 910 // the regular clock call is called once the scanned words reaches
911 911 // this limit
912 912 words_scanned_period = 12*1024,
913 913 // the regular clock call is called once the number of visited
914 914 // references reaches this limit
915 915 refs_reached_period = 384,
916 916 // initial value for the hash seed, used in the work stealing code
917 917 init_hash_seed = 17,
918 918 // how many entries will be transferred between global stack and
919 919 // local queues
920 920 global_stack_transfer_size = 16
921 921 };
922 922
923 923 int _task_id;
924 924 G1CollectedHeap* _g1h;
925 925 ConcurrentMark* _cm;
926 926 CMBitMap* _nextMarkBitMap;
927 927 // the task queue of this task
928 928 CMTaskQueue* _task_queue;
929 929 private:
930 930 // the task queue set---needed for stealing
931 931 CMTaskQueueSet* _task_queues;
932 932 // indicates whether the task has been claimed---this is only for
933 933 // debugging purposes
934 934 bool _claimed;
935 935
936 936 // number of calls to this task
937 937 int _calls;
938 938
939 939 // when the virtual timer reaches this time, the marking step should
940 940 // exit
941 941 double _time_target_ms;
942 942 // the start time of the current marking step
943 943 double _start_time_ms;
944 944
945 945 // the oop closure used for iterations over oops
946 946 G1CMOopClosure* _cm_oop_closure;
947 947
948 948 // the region this task is scanning, NULL if we're not scanning any
949 949 HeapRegion* _curr_region;
950 950 // the local finger of this task, NULL if we're not scanning a region
951 951 HeapWord* _finger;
952 952 // limit of the region this task is scanning, NULL if we're not scanning one
953 953 HeapWord* _region_limit;
954 954
955 955 // This is used only when we scan regions popped from the region
956 956 // stack. It records what the last object on such a region we
957 957 // scanned was. It is used to ensure that, if we abort region
958 958 // iteration, we do not rescan the first part of the region. This
959 959 // should be NULL when we're not scanning a region from the region
960 960 // stack.
961 961 HeapWord* _region_finger;
962 962
963 963 // If we abort while scanning a region we record the remaining
964 964 // unscanned portion and check this field when marking restarts.
965 965 // This avoids having to push on the region stack while other
966 966 // marking threads may still be popping regions.
967 967 // If we were to push the unscanned portion directly to the
968 968 // region stack then we would need to using locking versions
969 969 // of the push and pop operations.
970 970 MemRegion _aborted_region;
971 971
972 972 // the number of words this task has scanned
973 973 size_t _words_scanned;
974 974 // When _words_scanned reaches this limit, the regular clock is
975 975 // called. Notice that this might be decreased under certain
976 976 // circumstances (i.e. when we believe that we did an expensive
977 977 // operation).
978 978 size_t _words_scanned_limit;
979 979 // the initial value of _words_scanned_limit (i.e. what it was
980 980 // before it was decreased).
981 981 size_t _real_words_scanned_limit;
982 982
983 983 // the number of references this task has visited
984 984 size_t _refs_reached;
985 985 // When _refs_reached reaches this limit, the regular clock is
986 986 // called. Notice this this might be decreased under certain
987 987 // circumstances (i.e. when we believe that we did an expensive
988 988 // operation).
989 989 size_t _refs_reached_limit;
990 990 // the initial value of _refs_reached_limit (i.e. what it was before
991 991 // it was decreased).
992 992 size_t _real_refs_reached_limit;
993 993
994 994 // used by the work stealing stuff
995 995 int _hash_seed;
996 996 // if this is true, then the task has aborted for some reason
997 997 bool _has_aborted;
998 998 // set when the task aborts because it has met its time quota
999 999 bool _has_timed_out;
1000 1000 // true when we're draining SATB buffers; this avoids the task
1001 1001 // aborting due to SATB buffers being available (as we're already
1002 1002 // dealing with them)
1003 1003 bool _draining_satb_buffers;
1004 1004
1005 1005 // number sequence of past step times
1006 1006 NumberSeq _step_times_ms;
1007 1007 // elapsed time of this task
1008 1008 double _elapsed_time_ms;
1009 1009 // termination time of this task
1010 1010 double _termination_time_ms;
1011 1011 // when this task got into the termination protocol
1012 1012 double _termination_start_time_ms;
1013 1013
1014 1014 // true when the task is during a concurrent phase, false when it is
1015 1015 // in the remark phase (so, in the latter case, we do not have to
1016 1016 // check all the things that we have to check during the concurrent
1017 1017 // phase, i.e. SATB buffer availability...)
1018 1018 bool _concurrent;
1019 1019
1020 1020 TruncatedSeq _marking_step_diffs_ms;
1021 1021
1022 1022 // LOTS of statistics related with this task
1023 1023 #if _MARKING_STATS_
1024 1024 NumberSeq _all_clock_intervals_ms;
1025 1025 double _interval_start_time_ms;
1026 1026
1027 1027 int _aborted;
1028 1028 int _aborted_overflow;
1029 1029 int _aborted_cm_aborted;
1030 1030 int _aborted_yield;
1031 1031 int _aborted_timed_out;
1032 1032 int _aborted_satb;
1033 1033 int _aborted_termination;
1034 1034
1035 1035 int _steal_attempts;
1036 1036 int _steals;
1037 1037
1038 1038 int _clock_due_to_marking;
1039 1039 int _clock_due_to_scanning;
1040 1040
1041 1041 int _local_pushes;
1042 1042 int _local_pops;
1043 1043 int _local_max_size;
1044 1044 int _objs_scanned;
1045 1045
1046 1046 int _global_pushes;
1047 1047 int _global_pops;
1048 1048 int _global_max_size;
1049 1049
1050 1050 int _global_transfers_to;
1051 1051 int _global_transfers_from;
1052 1052
1053 1053 int _region_stack_pops;
1054 1054
1055 1055 int _regions_claimed;
1056 1056 int _objs_found_on_bitmap;
1057 1057
1058 1058 int _satb_buffers_processed;
1059 1059 #endif // _MARKING_STATS_
1060 1060
1061 1061 // it updates the local fields after this task has claimed
1062 1062 // a new region to scan
1063 1063 void setup_for_region(HeapRegion* hr);
1064 1064 // it brings up-to-date the limit of the region
1065 1065 void update_region_limit();
1066 1066
1067 1067 // called when either the words scanned or the refs visited limit
1068 1068 // has been reached
1069 1069 void reached_limit();
1070 1070 // recalculates the words scanned and refs visited limits
1071 1071 void recalculate_limits();
1072 1072 // decreases the words scanned and refs visited limits when we reach
1073 1073 // an expensive operation
1074 1074 void decrease_limits();
1075 1075 // it checks whether the words scanned or refs visited reached their
1076 1076 // respective limit and calls reached_limit() if they have
1077 1077 void check_limits() {
1078 1078 if (_words_scanned >= _words_scanned_limit ||
1079 1079 _refs_reached >= _refs_reached_limit) {
1080 1080 reached_limit();
1081 1081 }
1082 1082 }
1083 1083 // this is supposed to be called regularly during a marking step as
1084 1084 // it checks a bunch of conditions that might cause the marking step
1085 1085 // to abort
1086 1086 void regular_clock_call();
1087 1087 bool concurrent() { return _concurrent; }
1088 1088
1089 1089 public:
1090 1090 // It resets the task; it should be called right at the beginning of
1091 1091 // a marking phase.
1092 1092 void reset(CMBitMap* _nextMarkBitMap);
1093 1093 // it clears all the fields that correspond to a claimed region.
1094 1094 void clear_region_fields();
1095 1095
1096 1096 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1097 1097
1098 1098 // The main method of this class which performs a marking step
1099 1099 // trying not to exceed the given duration. However, it might exit
1100 1100 // prematurely, according to some conditions (i.e. SATB buffers are
1101 1101 // available for processing).
1102 1102 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1103 1103
1104 1104 // These two calls start and stop the timer
1105 1105 void record_start_time() {
1106 1106 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1107 1107 }
1108 1108 void record_end_time() {
1109 1109 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1110 1110 }
1111 1111
1112 1112 // returns the task ID
1113 1113 int task_id() { return _task_id; }
1114 1114
1115 1115 // From TerminatorTerminator. It determines whether this task should
1116 1116 // exit the termination protocol after it's entered it.
1117 1117 virtual bool should_exit_termination();
1118 1118
1119 1119 // Resets the local region fields after a task has finished scanning a
1120 1120 // region; or when they have become stale as a result of the region
1121 1121 // being evacuated.
1122 1122 void giveup_current_region();
1123 1123
1124 1124 HeapWord* finger() { return _finger; }
1125 1125
1126 1126 bool has_aborted() { return _has_aborted; }
1127 1127 void set_has_aborted() { _has_aborted = true; }
1128 1128 void clear_has_aborted() { _has_aborted = false; }
1129 1129 bool has_timed_out() { return _has_timed_out; }
1130 1130 bool claimed() { return _claimed; }
1131 1131
1132 1132 // Support routines for the partially scanned region that may be
1133 1133 // recorded as a result of aborting while draining the CMRegionStack
1134 1134 MemRegion aborted_region() { return _aborted_region; }
1135 1135 void set_aborted_region(MemRegion mr)
1136 1136 { _aborted_region = mr; }
1137 1137
1138 1138 // Clears any recorded partially scanned region
1139 1139 void clear_aborted_region() { set_aborted_region(MemRegion()); }
1140 1140
1141 1141 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1142 1142
1143 1143 // It grays the object by marking it and, if necessary, pushing it
1144 1144 // on the local queue
1145 1145 inline void deal_with_reference(oop obj);
1146 1146
1147 1147 // It scans an object and visits its children.
1148 1148 void scan_object(oop obj);
1149 1149
1150 1150 // It pushes an object on the local queue.
1151 1151 inline void push(oop obj);
1152 1152
1153 1153 // These two move entries to/from the global stack.
1154 1154 void move_entries_to_global_stack();
1155 1155 void get_entries_from_global_stack();
1156 1156
1157 1157 // It pops and scans objects from the local queue. If partially is
1158 1158 // true, then it stops when the queue size is of a given limit. If
1159 1159 // partially is false, then it stops when the queue is empty.
1160 1160 void drain_local_queue(bool partially);
1161 1161 // It moves entries from the global stack to the local queue and
1162 1162 // drains the local queue. If partially is true, then it stops when
1163 1163 // both the global stack and the local queue reach a given size. If
1164 1164 // partially if false, it tries to empty them totally.
1165 1165 void drain_global_stack(bool partially);
1166 1166 // It keeps picking SATB buffers and processing them until no SATB
1167 1167 // buffers are available.
1168 1168 void drain_satb_buffers();
1169 1169 // It keeps popping regions from the region stack and processing
1170 1170 // them until the region stack is empty.
1171 1171 void drain_region_stack(BitMapClosure* closure);
1172 1172
1173 1173 // moves the local finger to a new location
1174 1174 inline void move_finger_to(HeapWord* new_finger) {
1175 1175 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1176 1176 _finger = new_finger;
1177 1177 }
1178 1178
1179 1179 // moves the region finger to a new location
1180 1180 inline void move_region_finger_to(HeapWord* new_finger) {
1181 1181 assert(new_finger < _cm->finger(), "invariant");
1182 1182 _region_finger = new_finger;
1183 1183 }
1184 1184
1185 1185 CMTask(int task_num, ConcurrentMark *cm,
1186 1186 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1187 1187
1188 1188 // it prints statistics associated with this task
1189 1189 void print_stats();
1190 1190
1191 1191 #if _MARKING_STATS_
1192 1192 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1193 1193 #endif // _MARKING_STATS_
1194 1194 };
1195 1195
1196 1196 // Class that's used to to print out per-region liveness
1197 1197 // information. It's currently used at the end of marking and also
1198 1198 // after we sort the old regions at the end of the cleanup operation.
1199 1199 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1200 1200 private:
1201 1201 outputStream* _out;
1202 1202
1203 1203 // Accumulators for these values.
1204 1204 size_t _total_used_bytes;
1205 1205 size_t _total_capacity_bytes;
1206 1206 size_t _total_prev_live_bytes;
1207 1207 size_t _total_next_live_bytes;
1208 1208
1209 1209 // These are set up when we come across a "stars humongous" region
1210 1210 // (as this is where most of this information is stored, not in the
1211 1211 // subsequent "continues humongous" regions). After that, for every
1212 1212 // region in a given humongous region series we deduce the right
1213 1213 // values for it by simply subtracting the appropriate amount from
1214 1214 // these fields. All these values should reach 0 after we've visited
1215 1215 // the last region in the series.
1216 1216 size_t _hum_used_bytes;
1217 1217 size_t _hum_capacity_bytes;
1218 1218 size_t _hum_prev_live_bytes;
1219 1219 size_t _hum_next_live_bytes;
1220 1220
1221 1221 static double perc(size_t val, size_t total) {
1222 1222 if (total == 0) {
1223 1223 return 0.0;
1224 1224 } else {
1225 1225 return 100.0 * ((double) val / (double) total);
1226 1226 }
1227 1227 }
1228 1228
1229 1229 static double bytes_to_mb(size_t val) {
1230 1230 return (double) val / (double) M;
1231 1231 }
1232 1232
1233 1233 // See the .cpp file.
1234 1234 size_t get_hum_bytes(size_t* hum_bytes);
1235 1235 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1236 1236 size_t* prev_live_bytes, size_t* next_live_bytes);
1237 1237
1238 1238 public:
1239 1239 // The header and footer are printed in the constructor and
1240 1240 // destructor respectively.
1241 1241 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1242 1242 virtual bool doHeapRegion(HeapRegion* r);
1243 1243 ~G1PrintRegionLivenessInfoClosure();
1244 1244 };
1245 1245
1246 1246 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
↓ open down ↓ |
1148 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX