Print this page
rev 2896 : 6484965: G1: piggy-back liveness accounting phase on marking
Summary: Remove the separate counting phase of concurrent marking by tracking the amount of marked bytes and the cards spanned by marked objects in marking task/worker thread local data structures, which are updated as individual objects are marked.
Reviewed-by: brutisso
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; }
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 87 bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }
88 88 bool iterate(BitMapClosure* cl, MemRegion mr);
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
↓ open down ↓ |
422 lines elided |
↑ open up ↑ |
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 -
434 433 // this is set by any task, when an overflow on the global data
435 434 // structures is detected.
436 435 volatile bool _has_overflown;
437 436 // true: marking is concurrent, false: we're in remark
438 437 volatile bool _concurrent;
439 438 // set at the end of a Full GC so that marking aborts
440 439 volatile bool _has_aborted;
441 440
442 441 // used when remark aborts due to an overflow to indicate that
443 442 // another concurrent marking phase should start
444 443 volatile bool _restart_for_overflow;
445 444
446 445 // This is true from the very start of concurrent marking until the
447 446 // point when all the tasks complete their work. It is really used
448 447 // to determine the points between the end of concurrent marking and
449 448 // time of remark.
450 449 volatile bool _concurrent_marking_in_progress;
451 450
452 451 // verbose level
453 452 CMVerboseLevel _verbose_level;
454 453
455 454 // These two fields are used to implement the optimisation that
456 455 // avoids pushing objects on the global/region stack if there are
457 456 // no collection set regions above the lowest finger.
458 457
459 458 // This is the lowest finger (among the global and local fingers),
460 459 // which is calculated before a new collection set is chosen.
461 460 HeapWord* _min_finger;
462 461 // If this flag is true, objects/regions that are marked below the
463 462 // finger should be pushed on the stack(s). If this is flag is
464 463 // false, it is safe not to push them on the stack(s).
465 464 bool _should_gray_objects;
466 465
467 466 // All of these times are in ms.
468 467 NumberSeq _init_times;
469 468 NumberSeq _remark_times;
470 469 NumberSeq _remark_mark_times;
471 470 NumberSeq _remark_weak_ref_times;
472 471 NumberSeq _cleanup_times;
473 472 double _total_counting_time;
474 473 double _total_rs_scrub_time;
475 474
476 475 double* _accum_task_vtime; // accumulated task vtime
477 476
478 477 FlexibleWorkGang* _parallel_workers;
479 478
480 479 ForceOverflowSettings _force_overflow_conc;
481 480 ForceOverflowSettings _force_overflow_stw;
482 481
483 482 void weakRefsWork(bool clear_all_soft_refs);
484 483
485 484 void swapMarkBitMaps();
486 485
487 486 // It resets the global marking data structures, as well as the
488 487 // task local ones; should be called during initial mark.
489 488 void reset();
490 489 // It resets all the marking data structures.
491 490 void clear_marking_state(bool clear_overflow = true);
492 491
493 492 // It should be called to indicate which phase we're in (concurrent
494 493 // mark or remark) and how many threads are currently active.
495 494 void set_phase(size_t active_tasks, bool concurrent);
496 495 // We do this after we're done with marking so that the marking data
497 496 // structures are initialised to a sensible and predictable state.
498 497 void set_non_marking_state();
499 498
500 499 // prints all gathered CM-related statistics
501 500 void print_stats();
502 501
503 502 bool cleanup_list_is_empty() {
504 503 return _cleanup_list.is_empty();
505 504 }
506 505
507 506 // accessor methods
508 507 size_t parallel_marking_threads() { return _parallel_marking_threads; }
509 508 size_t max_parallel_marking_threads() { return _max_parallel_marking_threads;}
510 509 double sleep_factor() { return _sleep_factor; }
511 510 double marking_task_overhead() { return _marking_task_overhead;}
512 511 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
513 512 double cleanup_task_overhead() { return _cleanup_task_overhead;}
514 513
515 514 HeapWord* finger() { return _finger; }
516 515 bool concurrent() { return _concurrent; }
517 516 size_t active_tasks() { return _active_tasks; }
518 517 ParallelTaskTerminator* terminator() { return &_terminator; }
519 518
520 519 // It claims the next available region to be scanned by a marking
521 520 // task. It might return NULL if the next region is empty or we have
522 521 // run out of regions. In the latter case, out_of_regions()
523 522 // determines whether we've really run out of regions or the task
524 523 // should call claim_region() again. This might seem a bit
525 524 // awkward. Originally, the code was written so that claim_region()
526 525 // either successfully returned with a non-empty region or there
527 526 // were no more regions to be claimed. The problem with this was
528 527 // that, in certain circumstances, it iterated over large chunks of
529 528 // the heap finding only empty regions and, while it was working, it
530 529 // was preventing the calling task to call its regular clock
531 530 // method. So, this way, each task will spend very little time in
532 531 // claim_region() and is allowed to call the regular clock method
533 532 // frequently.
534 533 HeapRegion* claim_region(int task);
535 534
536 535 // It determines whether we've run out of regions to scan.
537 536 bool out_of_regions() { return _finger == _heap_end; }
538 537
539 538 // Returns the task with the given id
540 539 CMTask* task(int id) {
541 540 assert(0 <= id && id < (int) _active_tasks,
542 541 "task id not within active bounds");
543 542 return _tasks[id];
544 543 }
545 544
546 545 // Returns the task queue with the given id
547 546 CMTaskQueue* task_queue(int id) {
548 547 assert(0 <= id && id < (int) _active_tasks,
549 548 "task queue id not within active bounds");
550 549 return (CMTaskQueue*) _task_queues->queue(id);
551 550 }
552 551
553 552 // Returns the task queue set
554 553 CMTaskQueueSet* task_queues() { return _task_queues; }
555 554
556 555 // Access / manipulation of the overflow flag which is set to
557 556 // indicate that the global stack or region stack has overflown
558 557 bool has_overflown() { return _has_overflown; }
559 558 void set_has_overflown() { _has_overflown = true; }
560 559 void clear_has_overflown() { _has_overflown = false; }
561 560
562 561 bool has_aborted() { return _has_aborted; }
563 562 bool restart_for_overflow() { return _restart_for_overflow; }
564 563
565 564 // Methods to enter the two overflow sync barriers
566 565 void enter_first_sync_barrier(int task_num);
567 566 void enter_second_sync_barrier(int task_num);
568 567
569 568 ForceOverflowSettings* force_overflow_conc() {
570 569 return &_force_overflow_conc;
571 570 }
572 571
573 572 ForceOverflowSettings* force_overflow_stw() {
574 573 return &_force_overflow_stw;
↓ open down ↓ |
131 lines elided |
↑ open up ↑ |
575 574 }
576 575
577 576 ForceOverflowSettings* force_overflow() {
578 577 if (concurrent()) {
579 578 return force_overflow_conc();
580 579 } else {
581 580 return force_overflow_stw();
582 581 }
583 582 }
584 583
584 + // Live Data Counting data structures...
585 + // These data structures are initialized at the start of
586 + // marking. They are written to while marking is active.
587 + // They are aggregated during remark; the aggregated values
588 + // are then used to populate the _region_bm, _card_bm, and
589 + // the total live bytes, which are then subsequently updated
590 + // during cleanup.
591 +
592 + // An array of bitmaps (one bit map per task). Each bitmap
593 + // is used to record the cards spanned by the live objects
594 + // marked by that task/worker.
595 + BitMap* _count_card_bitmaps;
596 +
597 + // Used to record the number of marked live bytes
598 + // (for each region, by worker thread).
599 + size_t** _count_marked_bytes;
600 +
601 + // Card index of the bottom of the G1 heap. Used for biasing indices into
602 + // the card bitmaps.
603 + intptr_t _heap_bottom_card_num;
604 +
585 605 public:
586 606 // Manipulation of the global mark stack.
587 607 // Notice that the first mark_stack_push is CAS-based, whereas the
588 608 // two below are Mutex-based. This is OK since the first one is only
589 609 // called during evacuation pauses and doesn't compete with the
590 610 // other two (which are called by the marking tasks during
591 611 // concurrent marking or remark).
592 612 bool mark_stack_push(oop p) {
593 613 _markStack.par_push(p);
594 614 if (_markStack.overflow()) {
595 615 set_has_overflown();
596 616 return false;
597 617 }
598 618 return true;
599 619 }
600 620 bool mark_stack_push(oop* arr, int n) {
601 621 _markStack.par_push_arr(arr, n);
602 622 if (_markStack.overflow()) {
603 623 set_has_overflown();
604 624 return false;
605 625 }
606 626 return true;
607 627 }
608 628 void mark_stack_pop(oop* arr, int max, int* n) {
609 629 _markStack.par_pop_arr(arr, max, n);
610 630 }
611 631 size_t mark_stack_size() { return _markStack.size(); }
612 632 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
613 633 bool mark_stack_overflow() { return _markStack.overflow(); }
614 634 bool mark_stack_empty() { return _markStack.isEmpty(); }
615 635
616 636 // (Lock-free) Manipulation of the region stack
617 637 bool region_stack_push_lock_free(MemRegion mr) {
618 638 // Currently we only call the lock-free version during evacuation
619 639 // pauses.
620 640 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
621 641
622 642 _regionStack.push_lock_free(mr);
623 643 if (_regionStack.overflow()) {
624 644 set_has_overflown();
625 645 return false;
626 646 }
627 647 return true;
628 648 }
629 649
630 650 // Lock-free version of region-stack pop. Should only be
631 651 // called in tandem with other lock-free pops.
632 652 MemRegion region_stack_pop_lock_free() {
633 653 return _regionStack.pop_lock_free();
634 654 }
635 655
636 656 #if 0
637 657 // The routines that manipulate the region stack with a lock are
638 658 // not currently used. They should be retained, however, as a
639 659 // diagnostic aid.
640 660
641 661 bool region_stack_push_with_lock(MemRegion mr) {
642 662 // Currently we only call the lock-based version during either
643 663 // concurrent marking or remark.
644 664 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
645 665 "if we are at a safepoint it should be the remark safepoint");
646 666
647 667 _regionStack.push_with_lock(mr);
648 668 if (_regionStack.overflow()) {
649 669 set_has_overflown();
650 670 return false;
651 671 }
652 672 return true;
653 673 }
654 674
655 675 MemRegion region_stack_pop_with_lock() {
656 676 // Currently we only call the lock-based version during either
657 677 // concurrent marking or remark.
658 678 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
659 679 "if we are at a safepoint it should be the remark safepoint");
660 680
661 681 return _regionStack.pop_with_lock();
662 682 }
663 683 #endif
664 684
665 685 int region_stack_size() { return _regionStack.size(); }
666 686 bool region_stack_overflow() { return _regionStack.overflow(); }
667 687 bool region_stack_empty() { return _regionStack.isEmpty(); }
668 688
669 689 // Iterate over any regions that were aborted while draining the
670 690 // region stack (any such regions are saved in the corresponding
671 691 // CMTask) and invalidate (i.e. assign to the empty MemRegion())
672 692 // any regions that point into the collection set.
673 693 bool invalidate_aborted_regions_in_cset();
674 694
675 695 // Returns true if there are any aborted memory regions.
676 696 bool has_aborted_regions();
677 697
678 698 bool concurrent_marking_in_progress() {
679 699 return _concurrent_marking_in_progress;
680 700 }
681 701 void set_concurrent_marking_in_progress() {
682 702 _concurrent_marking_in_progress = true;
683 703 }
684 704 void clear_concurrent_marking_in_progress() {
685 705 _concurrent_marking_in_progress = false;
686 706 }
687 707
688 708 void update_accum_task_vtime(int i, double vtime) {
689 709 _accum_task_vtime[i] += vtime;
690 710 }
691 711
692 712 double all_task_accum_vtime() {
693 713 double ret = 0.0;
694 714 for (int i = 0; i < (int)_max_task_num; ++i)
695 715 ret += _accum_task_vtime[i];
↓ open down ↓ |
101 lines elided |
↑ open up ↑ |
696 716 return ret;
697 717 }
698 718
699 719 // Attempts to steal an object from the task queues of other tasks
700 720 bool try_stealing(int task_num, int* hash_seed, oop& obj) {
701 721 return _task_queues->steal(task_num, hash_seed, obj);
702 722 }
703 723
704 724 // It grays an object by first marking it. Then, if it's behind the
705 725 // global finger, it also pushes it on the global stack.
706 - void deal_with_reference(oop obj);
726 + void deal_with_reference(oop obj, int worker_i);
707 727
708 728 ConcurrentMark(ReservedSpace rs, int max_regions);
709 - ~ConcurrentMark();
729 +
710 730 ConcurrentMarkThread* cmThread() { return _cmThread; }
711 731
712 732 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
713 733 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
714 734
715 735 // Returns the number of GC threads to be used in a concurrent
716 736 // phase based on the number of GC threads being used in a STW
717 737 // phase.
718 738 size_t scale_parallel_threads(size_t n_par_threads);
719 739
720 740 // Calculates the number of GC threads to be used in a concurrent phase.
721 741 size_t calc_parallel_marking_threads();
722 742
723 743 // The following three are interaction between CM and
724 744 // G1CollectedHeap
725 745
726 746 // This notifies CM that a root during initial-mark needs to be
727 747 // grayed and it's MT-safe. Currently, we just mark it. But, in the
728 748 // future, we can experiment with pushing it on the stack and we can
729 749 // do this without changing G1CollectedHeap.
730 - void grayRoot(oop p);
750 + void grayRoot(oop p, int worker_i);
751 +
731 752 // It's used during evacuation pauses to gray a region, if
732 753 // necessary, and it's MT-safe. It assumes that the caller has
733 754 // marked any objects on that region. If _should_gray_objects is
734 755 // true and we're still doing concurrent marking, the region is
735 756 // pushed on the region stack, if it is located below the global
736 757 // finger, otherwise we do nothing.
737 758 void grayRegionIfNecessary(MemRegion mr);
759 +
738 760 // It's used during evacuation pauses to mark and, if necessary,
739 761 // gray a single object and it's MT-safe. It assumes the caller did
740 762 // not mark the object. If _should_gray_objects is true and we're
741 763 // still doing concurrent marking, the objects is pushed on the
742 764 // global stack, if it is located below the global finger, otherwise
743 765 // we do nothing.
744 - void markAndGrayObjectIfNecessary(oop p);
766 + void markAndGrayObjectIfNecessary(oop p, int worker_i);
745 767
746 768 // It iterates over the heap and for each object it comes across it
747 769 // will dump the contents of its reference fields, as well as
748 770 // liveness information for the object and its referents. The dump
749 771 // will be written to a file with the following name:
750 772 // G1PrintReachableBaseFile + "." + str.
751 773 // vo decides whether the prev (vo == UsePrevMarking), the next
752 774 // (vo == UseNextMarking) marking information, or the mark word
753 775 // (vo == UseMarkWord) will be used to determine the liveness of
754 776 // each object / referent.
755 777 // If all is true, all objects in the heap will be dumped, otherwise
756 778 // only the live ones. In the dump the following symbols / breviations
757 779 // are used:
758 780 // M : an explicitly live object (its bitmap bit is set)
759 781 // > : an implicitly live object (over tams)
760 782 // O : an object outside the G1 heap (typically: in the perm gen)
761 783 // NOT : a reference field whose referent is not live
762 784 // AND MARKED : indicates that an object is both explicitly and
763 785 // implicitly live (it should be one or the other, not both)
764 786 void print_reachable(const char* str,
765 787 VerifyOption vo, bool all) PRODUCT_RETURN;
766 788
767 789 // Clear the next marking bitmap (will be called concurrently).
768 790 void clearNextBitmap();
769 791
770 792 // These two do the work that needs to be done before and after the
771 793 // initial root checkpoint. Since this checkpoint can be done at two
772 794 // different points (i.e. an explicit pause or piggy-backed on a
773 795 // young collection), then it's nice to be able to easily share the
774 796 // pre/post code. It might be the case that we can put everything in
775 797 // the post method. TP
776 798 void checkpointRootsInitialPre();
777 799 void checkpointRootsInitialPost();
↓ open down ↓ |
23 lines elided |
↑ open up ↑ |
778 800
779 801 // Do concurrent phase of marking, to a tentative transitive closure.
780 802 void markFromRoots();
781 803
782 804 // Process all unprocessed SATB buffers. It is called at the
783 805 // beginning of an evacuation pause.
784 806 void drainAllSATBBuffers();
785 807
786 808 void checkpointRootsFinal(bool clear_all_soft_refs);
787 809 void checkpointRootsFinalWork();
788 - void calcDesiredRegions();
789 810 void cleanup();
790 811 void completeCleanup();
791 812
792 813 // Mark in the previous bitmap. NB: this is usually read-only, so use
793 814 // this carefully!
794 815 void markPrev(oop p);
795 - void clear(oop p);
816 +
817 + // Clears the mark in the next bitmap for the given object.
818 + void clear_mark(oop p);
819 +
796 820 // Clears marks for all objects in the given range, for both prev and
797 821 // next bitmaps. NB: the previous bitmap is usually read-only, so use
798 822 // this carefully!
799 823 void clearRangeBothMaps(MemRegion mr);
800 824
801 825 // Record the current top of the mark and region stacks; a
802 826 // subsequent oops_do() on the mark stack and
803 827 // invalidate_entries_into_cset() on the region stack will iterate
804 828 // only over indices valid at the time of this call.
805 829 void set_oops_do_bound() {
806 830 _markStack.set_oops_do_bound();
807 831 _regionStack.set_oops_do_bound();
808 832 }
809 833 // Iterate over the oops in the mark stack and all local queues. It
810 834 // also calls invalidate_entries_into_cset() on the region stack.
811 835 void oops_do(OopClosure* f);
812 836 // It is called at the end of an evacuation pause during marking so
813 837 // that CM is notified of where the new end of the heap is. It
814 838 // doesn't do anything if concurrent_marking_in_progress() is false,
815 839 // unless the force parameter is true.
816 840 void update_g1_committed(bool force = false);
817 841
818 842 void complete_marking_in_collection_set();
819 843
820 844 // It indicates that a new collection set is being chosen.
821 845 void newCSet();
822 846
823 847 // It registers a collection set heap region with CM. This is used
824 848 // to determine whether any heap regions are located above the finger.
825 849 void registerCSetRegion(HeapRegion* hr);
826 850
827 851 // Resets the region fields of any active CMTask whose region fields
828 852 // are in the collection set (i.e. the region currently claimed by
829 853 // the CMTask will be evacuated and may be used, subsequently, as
830 854 // an alloc region). When this happens the region fields in the CMTask
831 855 // are stale and, hence, should be cleared causing the worker thread
832 856 // to claim a new region.
833 857 void reset_active_task_region_fields_in_cset();
834 858
835 859 // Registers the maximum region-end associated with a set of
836 860 // regions with CM. Again this is used to determine whether any
837 861 // heap regions are located above the finger.
838 862 void register_collection_set_finger(HeapWord* max_finger) {
839 863 // max_finger is the highest heap region end of the regions currently
840 864 // contained in the collection set. If this value is larger than
841 865 // _min_finger then we need to gray objects.
842 866 // This routine is like registerCSetRegion but for an entire
843 867 // collection of regions.
844 868 if (max_finger > _min_finger) {
845 869 _should_gray_objects = true;
846 870 }
847 871 }
848 872
849 873 // Returns "true" if at least one mark has been completed.
850 874 bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
851 875
852 876 bool isMarked(oop p) const {
853 877 assert(p != NULL && p->is_oop(), "expected an oop");
854 878 HeapWord* addr = (HeapWord*)p;
855 879 assert(addr >= _nextMarkBitMap->startWord() ||
856 880 addr < _nextMarkBitMap->endWord(), "in a region");
857 881
858 882 return _nextMarkBitMap->isMarked(addr);
859 883 }
860 884
861 885 inline bool not_yet_marked(oop p) const;
862 886
863 887 // XXX Debug code
864 888 bool containing_card_is_marked(void* p);
865 889 bool containing_cards_are_marked(void* start, void* last);
866 890
867 891 bool isPrevMarked(oop p) const {
868 892 assert(p != NULL && p->is_oop(), "expected an oop");
869 893 HeapWord* addr = (HeapWord*)p;
870 894 assert(addr >= _prevMarkBitMap->startWord() ||
871 895 addr < _prevMarkBitMap->endWord(), "in a region");
872 896
873 897 return _prevMarkBitMap->isMarked(addr);
874 898 }
875 899
876 900 inline bool do_yield_check(int worker_i = 0);
877 901 inline bool should_yield();
878 902
879 903 // Called to abort the marking cycle after a Full GC takes palce.
880 904 void abort();
881 905
882 906 // This prints the global/local fingers. It is used for debugging.
883 907 NOT_PRODUCT(void print_finger();)
884 908
885 909 void print_summary_info();
886 910
887 911 void print_worker_threads_on(outputStream* st) const;
888 912
889 913 // The following indicate whether a given verbose level has been
890 914 // set. Notice that anything above stats is conditional to
891 915 // _MARKING_VERBOSE_ having been set to 1
892 916 bool verbose_stats() {
893 917 return _verbose_level >= stats_verbose;
↓ open down ↓ |
88 lines elided |
↑ open up ↑ |
894 918 }
895 919 bool verbose_low() {
896 920 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
897 921 }
898 922 bool verbose_medium() {
899 923 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
900 924 }
901 925 bool verbose_high() {
902 926 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
903 927 }
928 +
929 + // Counting data structure accessors
930 +
931 + // Returns the card number of the bottom of the G1 heap.
932 + // Used in biasing indices into accounting card bitmaps.
933 + intptr_t heap_bottom_card_num() const {
934 + return _heap_bottom_card_num;
935 + }
936 +
937 + // Returns the card bitmap for a given task or worker id.
938 + BitMap* count_card_bitmap_for(int worker_i) {
939 + assert(0 <= worker_i && (size_t) worker_i < _max_task_num, "oob");
940 + assert(_count_card_bitmaps != NULL, "uninitialized");
941 + BitMap* task_card_bm = &_count_card_bitmaps[worker_i];
942 + assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
943 + return task_card_bm;
944 + }
945 +
946 + // Returns the array containing the marked bytes for each region,
947 + // for the given worker or task id.
948 + size_t* count_marked_bytes_array_for(int worker_i) {
949 + assert(0 <= worker_i && (size_t) worker_i < _max_task_num, "oob");
950 + assert(_count_marked_bytes != NULL, "uninitialized");
951 + size_t* marked_bytes_array = _count_marked_bytes[worker_i];
952 + assert(marked_bytes_array != NULL, "uninitialized");
953 + return marked_bytes_array;
954 + }
955 +
956 + // Counts the size of the given memory region in the the given
957 + // marked_bytes array slot for the given HeapRegion.
958 + // Sets the bits in the given card bitmap that are associated with the
959 + // cards that are spanned by the memory region.
960 + inline void count_region(MemRegion mr, HeapRegion* hr,
961 + size_t* marked_bytes_array,
962 + BitMap* task_card_bm);
963 +
964 + // Counts the given memory region in the ask/worker counting
965 + // data structures for the given worker id.
966 + inline void count_region(MemRegion mr, int worker_i);
967 +
968 + // Counts the given object in the given task/worker counting
969 + // data structures.
970 + inline void count_object(oop obj, HeapRegion* hr,
971 + size_t* marked_bytes_array,
972 + BitMap* task_card_bm);
973 +
974 + // Counts the given object in the task/worker counting data
975 + // structures for the given worker id.
976 + inline void count_object(oop obj, HeapRegion* hr, int worker_i);
977 +
978 + // Attempts to mark the given object and, if successful, counts
979 + // the object in the given task/worker counting structures.
980 + inline bool par_mark_and_count(oop obj, HeapRegion* hr,
981 + size_t* marked_bytes_array,
982 + BitMap* task_card_bm);
983 +
984 + // Attempts to mark the given object and, if successful, counts
985 + // the object in the task/worker counting structures for the
986 + // given worker id.
987 + inline bool par_mark_and_count(oop obj, HeapRegion* hr, int worker_i);
988 +
989 + // Similar to the above routine but we don't know the heap region that
990 + // contains the object to be marked/counted, which this routine looks up.
991 + inline bool par_mark_and_count(oop obj, int worker_i);
992 +
993 + // Unconditionally mark the given object, and unconditinally count
994 + // the object in the counting structures for worker id 0.
995 + // Should *not* be called from parallel code.
996 + inline bool mark_and_count(oop obj, HeapRegion* hr);
997 +
998 + // Similar to the above routine but we don't know the heap region that
999 + // contains the object to be marked/counted, which this routine looks up.
1000 + // Should *not* be called from parallel code.
1001 + inline bool mark_and_count(oop obj);
1002 +
1003 + // Clears the count data for the given region from _all_ of
1004 + // the per-task counting data structures.
1005 + void clear_count_data_for_heap_region(HeapRegion* hr);
1006 +
1007 +protected:
1008 + // Clear all the per-task bitmaps and arrays used to store the
1009 + // counting data.
1010 + void clear_all_count_data();
1011 +
1012 + // Aggregates the counting data for each worker/task
1013 + // that was constructed while marking. Also sets
1014 + // the amount of marked bytes for each region and
1015 + // the top at concurrent mark count.
1016 + void aggregate_and_clear_count_data();
1017 +
1018 + // Verification routine
1019 + void verify_count_data();
904 1020 };
905 1021
906 1022 // A class representing a marking task.
907 1023 class CMTask : public TerminatorTerminator {
908 1024 private:
909 1025 enum PrivateConstants {
910 1026 // the regular clock call is called once the scanned words reaches
911 1027 // this limit
912 1028 words_scanned_period = 12*1024,
913 1029 // the regular clock call is called once the number of visited
914 1030 // references reaches this limit
915 1031 refs_reached_period = 384,
916 1032 // initial value for the hash seed, used in the work stealing code
917 1033 init_hash_seed = 17,
918 1034 // how many entries will be transferred between global stack and
919 1035 // local queues
920 1036 global_stack_transfer_size = 16
921 1037 };
922 1038
923 1039 int _task_id;
924 1040 G1CollectedHeap* _g1h;
925 1041 ConcurrentMark* _cm;
926 1042 CMBitMap* _nextMarkBitMap;
927 1043 // the task queue of this task
928 1044 CMTaskQueue* _task_queue;
929 1045 private:
930 1046 // the task queue set---needed for stealing
931 1047 CMTaskQueueSet* _task_queues;
932 1048 // indicates whether the task has been claimed---this is only for
933 1049 // debugging purposes
934 1050 bool _claimed;
935 1051
936 1052 // number of calls to this task
937 1053 int _calls;
938 1054
939 1055 // when the virtual timer reaches this time, the marking step should
940 1056 // exit
941 1057 double _time_target_ms;
942 1058 // the start time of the current marking step
943 1059 double _start_time_ms;
944 1060
945 1061 // the oop closure used for iterations over oops
946 1062 G1CMOopClosure* _cm_oop_closure;
947 1063
948 1064 // the region this task is scanning, NULL if we're not scanning any
949 1065 HeapRegion* _curr_region;
950 1066 // the local finger of this task, NULL if we're not scanning a region
951 1067 HeapWord* _finger;
952 1068 // limit of the region this task is scanning, NULL if we're not scanning one
953 1069 HeapWord* _region_limit;
954 1070
955 1071 // This is used only when we scan regions popped from the region
956 1072 // stack. It records what the last object on such a region we
957 1073 // scanned was. It is used to ensure that, if we abort region
958 1074 // iteration, we do not rescan the first part of the region. This
959 1075 // should be NULL when we're not scanning a region from the region
960 1076 // stack.
961 1077 HeapWord* _region_finger;
962 1078
963 1079 // If we abort while scanning a region we record the remaining
964 1080 // unscanned portion and check this field when marking restarts.
965 1081 // This avoids having to push on the region stack while other
966 1082 // marking threads may still be popping regions.
967 1083 // If we were to push the unscanned portion directly to the
968 1084 // region stack then we would need to using locking versions
969 1085 // of the push and pop operations.
970 1086 MemRegion _aborted_region;
971 1087
972 1088 // the number of words this task has scanned
973 1089 size_t _words_scanned;
974 1090 // When _words_scanned reaches this limit, the regular clock is
975 1091 // called. Notice that this might be decreased under certain
976 1092 // circumstances (i.e. when we believe that we did an expensive
977 1093 // operation).
978 1094 size_t _words_scanned_limit;
979 1095 // the initial value of _words_scanned_limit (i.e. what it was
980 1096 // before it was decreased).
981 1097 size_t _real_words_scanned_limit;
982 1098
983 1099 // the number of references this task has visited
984 1100 size_t _refs_reached;
985 1101 // When _refs_reached reaches this limit, the regular clock is
986 1102 // called. Notice this this might be decreased under certain
987 1103 // circumstances (i.e. when we believe that we did an expensive
988 1104 // operation).
989 1105 size_t _refs_reached_limit;
990 1106 // the initial value of _refs_reached_limit (i.e. what it was before
991 1107 // it was decreased).
992 1108 size_t _real_refs_reached_limit;
993 1109
994 1110 // used by the work stealing stuff
995 1111 int _hash_seed;
996 1112 // if this is true, then the task has aborted for some reason
997 1113 bool _has_aborted;
998 1114 // set when the task aborts because it has met its time quota
999 1115 bool _has_timed_out;
1000 1116 // true when we're draining SATB buffers; this avoids the task
1001 1117 // aborting due to SATB buffers being available (as we're already
1002 1118 // dealing with them)
1003 1119 bool _draining_satb_buffers;
1004 1120
1005 1121 // number sequence of past step times
1006 1122 NumberSeq _step_times_ms;
1007 1123 // elapsed time of this task
1008 1124 double _elapsed_time_ms;
1009 1125 // termination time of this task
1010 1126 double _termination_time_ms;
1011 1127 // when this task got into the termination protocol
↓ open down ↓ |
98 lines elided |
↑ open up ↑ |
1012 1128 double _termination_start_time_ms;
1013 1129
1014 1130 // true when the task is during a concurrent phase, false when it is
1015 1131 // in the remark phase (so, in the latter case, we do not have to
1016 1132 // check all the things that we have to check during the concurrent
1017 1133 // phase, i.e. SATB buffer availability...)
1018 1134 bool _concurrent;
1019 1135
1020 1136 TruncatedSeq _marking_step_diffs_ms;
1021 1137
1138 + // Counting data structures. Embedding the task's marked_bytes_array
1139 + // and card bitmap into the actual task saves having to go through
1140 + // the ConcurrentMark object.
1141 + size_t* _marked_bytes_array;
1142 + BitMap* _card_bm;
1143 +
1022 1144 // LOTS of statistics related with this task
1023 1145 #if _MARKING_STATS_
1024 1146 NumberSeq _all_clock_intervals_ms;
1025 1147 double _interval_start_time_ms;
1026 1148
1027 1149 int _aborted;
1028 1150 int _aborted_overflow;
1029 1151 int _aborted_cm_aborted;
1030 1152 int _aborted_yield;
1031 1153 int _aborted_timed_out;
1032 1154 int _aborted_satb;
1033 1155 int _aborted_termination;
1034 1156
1035 1157 int _steal_attempts;
1036 1158 int _steals;
1037 1159
1038 1160 int _clock_due_to_marking;
1039 1161 int _clock_due_to_scanning;
1040 1162
1041 1163 int _local_pushes;
1042 1164 int _local_pops;
1043 1165 int _local_max_size;
1044 1166 int _objs_scanned;
1045 1167
1046 1168 int _global_pushes;
1047 1169 int _global_pops;
1048 1170 int _global_max_size;
1049 1171
1050 1172 int _global_transfers_to;
1051 1173 int _global_transfers_from;
1052 1174
1053 1175 int _region_stack_pops;
1054 1176
1055 1177 int _regions_claimed;
1056 1178 int _objs_found_on_bitmap;
1057 1179
1058 1180 int _satb_buffers_processed;
1059 1181 #endif // _MARKING_STATS_
1060 1182
1061 1183 // it updates the local fields after this task has claimed
1062 1184 // a new region to scan
1063 1185 void setup_for_region(HeapRegion* hr);
1064 1186 // it brings up-to-date the limit of the region
1065 1187 void update_region_limit();
1066 1188
1067 1189 // called when either the words scanned or the refs visited limit
1068 1190 // has been reached
1069 1191 void reached_limit();
1070 1192 // recalculates the words scanned and refs visited limits
1071 1193 void recalculate_limits();
1072 1194 // decreases the words scanned and refs visited limits when we reach
1073 1195 // an expensive operation
1074 1196 void decrease_limits();
1075 1197 // it checks whether the words scanned or refs visited reached their
1076 1198 // respective limit and calls reached_limit() if they have
1077 1199 void check_limits() {
1078 1200 if (_words_scanned >= _words_scanned_limit ||
1079 1201 _refs_reached >= _refs_reached_limit) {
1080 1202 reached_limit();
1081 1203 }
1082 1204 }
1083 1205 // this is supposed to be called regularly during a marking step as
1084 1206 // it checks a bunch of conditions that might cause the marking step
1085 1207 // to abort
1086 1208 void regular_clock_call();
1087 1209 bool concurrent() { return _concurrent; }
1088 1210
1089 1211 public:
1090 1212 // It resets the task; it should be called right at the beginning of
1091 1213 // a marking phase.
1092 1214 void reset(CMBitMap* _nextMarkBitMap);
1093 1215 // it clears all the fields that correspond to a claimed region.
1094 1216 void clear_region_fields();
1095 1217
1096 1218 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1097 1219
1098 1220 // The main method of this class which performs a marking step
1099 1221 // trying not to exceed the given duration. However, it might exit
1100 1222 // prematurely, according to some conditions (i.e. SATB buffers are
1101 1223 // available for processing).
1102 1224 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1103 1225
1104 1226 // These two calls start and stop the timer
1105 1227 void record_start_time() {
1106 1228 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1107 1229 }
1108 1230 void record_end_time() {
1109 1231 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1110 1232 }
1111 1233
1112 1234 // returns the task ID
1113 1235 int task_id() { return _task_id; }
1114 1236
1115 1237 // From TerminatorTerminator. It determines whether this task should
1116 1238 // exit the termination protocol after it's entered it.
1117 1239 virtual bool should_exit_termination();
1118 1240
1119 1241 // Resets the local region fields after a task has finished scanning a
1120 1242 // region; or when they have become stale as a result of the region
1121 1243 // being evacuated.
1122 1244 void giveup_current_region();
1123 1245
1124 1246 HeapWord* finger() { return _finger; }
1125 1247
1126 1248 bool has_aborted() { return _has_aborted; }
1127 1249 void set_has_aborted() { _has_aborted = true; }
1128 1250 void clear_has_aborted() { _has_aborted = false; }
1129 1251 bool has_timed_out() { return _has_timed_out; }
1130 1252 bool claimed() { return _claimed; }
1131 1253
1132 1254 // Support routines for the partially scanned region that may be
1133 1255 // recorded as a result of aborting while draining the CMRegionStack
1134 1256 MemRegion aborted_region() { return _aborted_region; }
1135 1257 void set_aborted_region(MemRegion mr)
1136 1258 { _aborted_region = mr; }
1137 1259
1138 1260 // Clears any recorded partially scanned region
1139 1261 void clear_aborted_region() { set_aborted_region(MemRegion()); }
1140 1262
1141 1263 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1142 1264
1143 1265 // It grays the object by marking it and, if necessary, pushing it
1144 1266 // on the local queue
1145 1267 inline void deal_with_reference(oop obj);
1146 1268
1147 1269 // It scans an object and visits its children.
1148 1270 void scan_object(oop obj);
1149 1271
1150 1272 // It pushes an object on the local queue.
1151 1273 inline void push(oop obj);
1152 1274
1153 1275 // These two move entries to/from the global stack.
1154 1276 void move_entries_to_global_stack();
1155 1277 void get_entries_from_global_stack();
1156 1278
1157 1279 // It pops and scans objects from the local queue. If partially is
1158 1280 // true, then it stops when the queue size is of a given limit. If
1159 1281 // partially is false, then it stops when the queue is empty.
1160 1282 void drain_local_queue(bool partially);
1161 1283 // It moves entries from the global stack to the local queue and
1162 1284 // drains the local queue. If partially is true, then it stops when
1163 1285 // both the global stack and the local queue reach a given size. If
1164 1286 // partially if false, it tries to empty them totally.
1165 1287 void drain_global_stack(bool partially);
1166 1288 // It keeps picking SATB buffers and processing them until no SATB
1167 1289 // buffers are available.
1168 1290 void drain_satb_buffers();
1169 1291 // It keeps popping regions from the region stack and processing
1170 1292 // them until the region stack is empty.
1171 1293 void drain_region_stack(BitMapClosure* closure);
1172 1294
1173 1295 // moves the local finger to a new location
1174 1296 inline void move_finger_to(HeapWord* new_finger) {
1175 1297 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
↓ open down ↓ |
144 lines elided |
↑ open up ↑ |
1176 1298 _finger = new_finger;
1177 1299 }
1178 1300
1179 1301 // moves the region finger to a new location
1180 1302 inline void move_region_finger_to(HeapWord* new_finger) {
1181 1303 assert(new_finger < _cm->finger(), "invariant");
1182 1304 _region_finger = new_finger;
1183 1305 }
1184 1306
1185 1307 CMTask(int task_num, ConcurrentMark *cm,
1308 + size_t* marked_bytes, BitMap* card_bm,
1186 1309 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1187 1310
1188 1311 // it prints statistics associated with this task
1189 1312 void print_stats();
1190 1313
1191 1314 #if _MARKING_STATS_
1192 1315 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1193 1316 #endif // _MARKING_STATS_
1194 1317 };
1195 1318
1196 1319 // Class that's used to to print out per-region liveness
1197 1320 // information. It's currently used at the end of marking and also
1198 1321 // after we sort the old regions at the end of the cleanup operation.
1199 1322 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1200 1323 private:
1201 1324 outputStream* _out;
1202 1325
1203 1326 // Accumulators for these values.
1204 1327 size_t _total_used_bytes;
1205 1328 size_t _total_capacity_bytes;
1206 1329 size_t _total_prev_live_bytes;
1207 1330 size_t _total_next_live_bytes;
1208 1331
1209 1332 // These are set up when we come across a "stars humongous" region
1210 1333 // (as this is where most of this information is stored, not in the
1211 1334 // subsequent "continues humongous" regions). After that, for every
1212 1335 // region in a given humongous region series we deduce the right
1213 1336 // values for it by simply subtracting the appropriate amount from
1214 1337 // these fields. All these values should reach 0 after we've visited
1215 1338 // the last region in the series.
1216 1339 size_t _hum_used_bytes;
1217 1340 size_t _hum_capacity_bytes;
1218 1341 size_t _hum_prev_live_bytes;
1219 1342 size_t _hum_next_live_bytes;
1220 1343
1221 1344 static double perc(size_t val, size_t total) {
1222 1345 if (total == 0) {
1223 1346 return 0.0;
1224 1347 } else {
1225 1348 return 100.0 * ((double) val / (double) total);
1226 1349 }
1227 1350 }
1228 1351
1229 1352 static double bytes_to_mb(size_t val) {
1230 1353 return (double) val / (double) M;
1231 1354 }
1232 1355
1233 1356 // See the .cpp file.
1234 1357 size_t get_hum_bytes(size_t* hum_bytes);
1235 1358 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1236 1359 size_t* prev_live_bytes, size_t* next_live_bytes);
1237 1360
1238 1361 public:
1239 1362 // The header and footer are printed in the constructor and
1240 1363 // destructor respectively.
1241 1364 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1242 1365 virtual bool doHeapRegion(HeapRegion* r);
1243 1366 ~G1PrintRegionLivenessInfoClosure();
1244 1367 };
1245 1368
1246 1369 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
↓ open down ↓ |
51 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX