Print this page
rev 3708 : 8000244: G1: Ergonomically set MarkStackSize and use virtual space for global marking stack
Summary: Set the value of MarkStackSize to a value based on the number of parallel marking threads with a reasonable minimum. Expand the marking stack if we have to restart marking due to an overflow up to a reasonable maximum. Allocate the underlying space for the marking stack from virtual memory.
Reviewed-by: jmasa
rev 3709 : imported patch reuse-old-marking-stack
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, 2012, 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, mtGC> CMTaskQueue;
34 34 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> 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) : _g1(g1) { }
46 46
47 47 void do_object(oop obj) {
48 48 ShouldNotCallThis();
49 49 }
50 50 bool do_object_b(oop obj);
51 51 };
52 52
53 53 // A generic CM bit map. This is essentially a wrapper around the BitMap
54 54 // class, with one bit per (1<<_shifter) HeapWords.
55 55
↓ open down ↓ |
55 lines elided |
↑ open up ↑ |
56 56 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
57 57 protected:
58 58 HeapWord* _bmStartWord; // base address of range covered by map
59 59 size_t _bmWordSize; // map size (in #HeapWords covered)
60 60 const int _shifter; // map to char or bit
61 61 VirtualSpace _virtual_space; // underlying the bit map
62 62 BitMap _bm; // the bit map itself
63 63
64 64 public:
65 65 // constructor
66 - CMBitMapRO(ReservedSpace rs, int shifter);
66 + CMBitMapRO(int shifter);
67 67
68 68 enum { do_yield = true };
69 69
70 70 // inquiries
71 71 HeapWord* startWord() const { return _bmStartWord; }
72 72 size_t sizeInWords() const { return _bmWordSize; }
73 73 // the following is one past the last word in space
74 74 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
75 75
76 76 // read marks
77 77
78 78 bool isMarked(HeapWord* addr) const {
79 79 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
80 80 "outside underlying space?");
81 81 return _bm.at(heapWordToOffset(addr));
82 82 }
83 83
84 84 // iteration
85 85 inline bool iterate(BitMapClosure* cl, MemRegion mr);
86 86 inline bool iterate(BitMapClosure* cl);
87 87
88 88 // Return the address corresponding to the next marked bit at or after
89 89 // "addr", and before "limit", if "limit" is non-NULL. If there is no
90 90 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
91 91 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
92 92 HeapWord* limit = NULL) const;
93 93 // Return the address corresponding to the next unmarked bit at or after
94 94 // "addr", and before "limit", if "limit" is non-NULL. If there is no
95 95 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
96 96 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
97 97 HeapWord* limit = NULL) const;
98 98
99 99 // conversion utilities
100 100 // XXX Fix these so that offsets are size_t's...
101 101 HeapWord* offsetToHeapWord(size_t offset) const {
102 102 return _bmStartWord + (offset << _shifter);
103 103 }
104 104 size_t heapWordToOffset(HeapWord* addr) const {
105 105 return pointer_delta(addr, _bmStartWord) >> _shifter;
106 106 }
107 107 int heapWordDiffToOffsetDiff(size_t diff) const;
108 108 HeapWord* nextWord(HeapWord* addr) {
109 109 return offsetToHeapWord(heapWordToOffset(addr) + 1);
↓ open down ↓ |
33 lines elided |
↑ open up ↑ |
110 110 }
111 111
112 112 // debugging
113 113 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
114 114 };
115 115
116 116 class CMBitMap : public CMBitMapRO {
117 117
118 118 public:
119 119 // constructor
120 - CMBitMap(ReservedSpace rs, int shifter) :
121 - CMBitMapRO(rs, shifter) {}
120 + CMBitMap(int shifter) :
121 + CMBitMapRO(shifter) {}
122 +
123 + // Allocates the back store for the marking bitmap
124 + bool allocate(ReservedSpace heap_rs);
122 125
123 126 // write marks
124 127 void mark(HeapWord* addr) {
125 128 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
126 129 "outside underlying space?");
127 130 _bm.set_bit(heapWordToOffset(addr));
128 131 }
129 132 void clear(HeapWord* addr) {
130 133 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
131 134 "outside underlying space?");
132 135 _bm.clear_bit(heapWordToOffset(addr));
133 136 }
134 137 bool parMark(HeapWord* addr) {
135 138 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
136 139 "outside underlying space?");
137 140 return _bm.par_set_bit(heapWordToOffset(addr));
138 141 }
139 142 bool parClear(HeapWord* addr) {
140 143 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
141 144 "outside underlying space?");
142 145 return _bm.par_clear_bit(heapWordToOffset(addr));
143 146 }
144 147 void markRange(MemRegion mr);
145 148 void clearAll();
146 149 void clearRange(MemRegion mr);
147 150
↓ open down ↓ |
16 lines elided |
↑ open up ↑ |
148 151 // Starting at the bit corresponding to "addr" (inclusive), find the next
149 152 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
150 153 // the end of this run (stopping at "end_addr"). Return the MemRegion
151 154 // covering from the start of the region corresponding to the first bit
152 155 // of the run to the end of the region corresponding to the last bit of
153 156 // the run. If there is no "1" bit at or after "addr", return an empty
154 157 // MemRegion.
155 158 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
156 159 };
157 160
158 -// Represents a marking stack used by the CM collector.
159 -// Ideally this should be GrowableArray<> just like MSC's marking stack(s).
161 +// Represents a marking stack used by ConcurrentMarking in the G1 collector.
160 162 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
163 + ReservedSpace _rs;
164 + VirtualSpace _virtual_space; // Underlying backing store for actual stack
161 165 ConcurrentMark* _cm;
162 - oop* _base; // bottom of stack
163 - jint _index; // one more than last occupied index
164 - jint _capacity; // max #elements
165 - jint _saved_index; // value of _index saved at start of GC
166 - NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
166 + oop* _base; // bottom of stack
167 + jint _index; // one more than last occupied index
168 + jint _capacity; // max #elements
169 + jint _saved_index; // value of _index saved at start of GC
170 + NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
167 171
168 - bool _overflow;
172 + bool _overflow;
173 + bool _should_expand;
169 174 DEBUG_ONLY(bool _drain_in_progress;)
170 175 DEBUG_ONLY(bool _drain_in_progress_yields;)
171 176
172 177 public:
173 178 CMMarkStack(ConcurrentMark* cm);
174 179 ~CMMarkStack();
175 180
176 - void allocate(size_t size);
181 +#ifndef PRODUCT
182 + jint max_depth() const {
183 + return _max_depth;
184 + }
185 +#endif
186 +
187 + bool allocate(size_t capacity);
177 188
178 189 oop pop() {
179 190 if (!isEmpty()) {
180 191 return _base[--_index] ;
181 192 }
182 193 return NULL;
183 194 }
184 195
185 196 // If overflow happens, don't do the push, and record the overflow.
186 197 // *Requires* that "ptr" is already marked.
187 198 void push(oop ptr) {
188 199 if (isFull()) {
189 200 // Record overflow.
190 201 _overflow = true;
191 202 return;
192 203 } else {
193 204 _base[_index++] = ptr;
194 205 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
195 206 }
196 207 }
197 208 // Non-block impl. Note: concurrency is allowed only with other
198 209 // "par_push" operations, not with "pop" or "drain". We would need
199 210 // parallel versions of them if such concurrency was desired.
200 211 void par_push(oop ptr);
201 212
202 213 // Pushes the first "n" elements of "ptr_arr" on the stack.
203 214 // Non-block impl. Note: concurrency is allowed only with other
204 215 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
205 216 void par_adjoin_arr(oop* ptr_arr, int n);
206 217
207 218 // Pushes the first "n" elements of "ptr_arr" on the stack.
208 219 // Locking impl: concurrency is allowed only with
209 220 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
210 221 // locking strategy.
211 222 void par_push_arr(oop* ptr_arr, int n);
212 223
213 224 // If returns false, the array was empty. Otherwise, removes up to "max"
214 225 // elements from the stack, and transfers them to "ptr_arr" in an
215 226 // unspecified order. The actual number transferred is given in "n" ("n
216 227 // == 0" is deliberately redundant with the return value.) Locking impl:
217 228 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
218 229 // operations, which use the same locking strategy.
219 230 bool par_pop_arr(oop* ptr_arr, int max, int* n);
220 231
221 232 // Drain the mark stack, applying the given closure to all fields of
222 233 // objects on the stack. (That is, continue until the stack is empty,
223 234 // even if closure applications add entries to the stack.) The "bm"
↓ open down ↓ |
37 lines elided |
↑ open up ↑ |
224 235 // argument, if non-null, may be used to verify that only marked objects
225 236 // are on the mark stack. If "yield_after" is "true", then the
226 237 // concurrent marker performing the drain offers to yield after
227 238 // processing each object. If a yield occurs, stops the drain operation
228 239 // and returns false. Otherwise, returns true.
229 240 template<class OopClosureClass>
230 241 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
231 242
232 243 bool isEmpty() { return _index == 0; }
233 244 bool isFull() { return _index == _capacity; }
234 - int maxElems() { return _capacity; }
245 + int maxElems() { return _capacity; }
235 246
236 247 bool overflow() { return _overflow; }
237 248 void clear_overflow() { _overflow = false; }
238 249
250 + bool should_expand() const { return _should_expand; }
251 + void set_should_expand();
252 +
253 + // Expand the stack, typically in response to an overflow condition
254 + void expand();
255 +
239 256 int size() { return _index; }
240 257
241 258 void setEmpty() { _index = 0; clear_overflow(); }
242 259
243 260 // Record the current index.
244 261 void note_start_of_gc();
245 262
246 263 // Make sure that we have not added any entries to the stack during GC.
247 264 void note_end_of_gc();
248 265
249 266 // iterate over the oops in the mark stack, up to the bound recorded via
250 267 // the call above.
251 268 void oops_do(OopClosure* f);
252 269 };
253 270
254 271 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
255 272 private:
256 273 #ifndef PRODUCT
257 274 uintx _num_remaining;
258 275 bool _force;
259 276 #endif // !defined(PRODUCT)
260 277
261 278 public:
262 279 void init() PRODUCT_RETURN;
263 280 void update() PRODUCT_RETURN;
264 281 bool should_force() PRODUCT_RETURN_( return false; );
265 282 };
266 283
267 284 // this will enable a variety of different statistics per GC task
268 285 #define _MARKING_STATS_ 0
269 286 // this will enable the higher verbose levels
270 287 #define _MARKING_VERBOSE_ 0
271 288
272 289 #if _MARKING_STATS_
273 290 #define statsOnly(statement) \
274 291 do { \
275 292 statement ; \
276 293 } while (0)
277 294 #else // _MARKING_STATS_
278 295 #define statsOnly(statement) \
279 296 do { \
280 297 } while (0)
281 298 #endif // _MARKING_STATS_
282 299
283 300 typedef enum {
284 301 no_verbose = 0, // verbose turned off
285 302 stats_verbose, // only prints stats at the end of marking
286 303 low_verbose, // low verbose, mostly per region and per major event
287 304 medium_verbose, // a bit more detailed than low
288 305 high_verbose // per object verbose
289 306 } CMVerboseLevel;
290 307
291 308 class YoungList;
292 309
293 310 // Root Regions are regions that are not empty at the beginning of a
294 311 // marking cycle and which we might collect during an evacuation pause
295 312 // while the cycle is active. Given that, during evacuation pauses, we
296 313 // do not copy objects that are explicitly marked, what we have to do
297 314 // for the root regions is to scan them and mark all objects reachable
298 315 // from them. According to the SATB assumptions, we only need to visit
299 316 // each object once during marking. So, as long as we finish this scan
300 317 // before the next evacuation pause, we can copy the objects from the
301 318 // root regions without having to mark them or do anything else to them.
302 319 //
303 320 // Currently, we only support root region scanning once (at the start
304 321 // of the marking cycle) and the root regions are all the survivor
305 322 // regions populated during the initial-mark pause.
306 323 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
307 324 private:
308 325 YoungList* _young_list;
309 326 ConcurrentMark* _cm;
310 327
311 328 volatile bool _scan_in_progress;
312 329 volatile bool _should_abort;
313 330 HeapRegion* volatile _next_survivor;
314 331
315 332 public:
316 333 CMRootRegions();
317 334 // We actually do most of the initialization in this method.
318 335 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
319 336
320 337 // Reset the claiming / scanning of the root regions.
321 338 void prepare_for_scan();
322 339
323 340 // Forces get_next() to return NULL so that the iteration aborts early.
324 341 void abort() { _should_abort = true; }
325 342
326 343 // Return true if the CM thread are actively scanning root regions,
327 344 // false otherwise.
328 345 bool scan_in_progress() { return _scan_in_progress; }
329 346
330 347 // Claim the next root region to scan atomically, or return NULL if
331 348 // all have been claimed.
332 349 HeapRegion* claim_next();
333 350
334 351 // Flag that we're done with root region scanning and notify anyone
335 352 // who's waiting on it. If aborted is false, assume that all regions
336 353 // have been claimed.
↓ open down ↓ |
88 lines elided |
↑ open up ↑ |
337 354 void scan_finished();
338 355
339 356 // If CM threads are still scanning root regions, wait until they
340 357 // are done. Return true if we had to wait, false otherwise.
341 358 bool wait_until_scan_finished();
342 359 };
343 360
344 361 class ConcurrentMarkThread;
345 362
346 363 class ConcurrentMark: public CHeapObj<mtGC> {
364 + friend class CMMarkStack;
347 365 friend class ConcurrentMarkThread;
348 366 friend class CMTask;
349 367 friend class CMBitMapClosure;
350 368 friend class CMGlobalObjectClosure;
351 369 friend class CMRemarkTask;
352 370 friend class CMConcurrentMarkingTask;
353 371 friend class G1ParNoteEndTask;
354 372 friend class CalcLiveObjectsClosure;
355 373 friend class G1CMRefProcTaskProxy;
356 374 friend class G1CMRefProcTaskExecutor;
357 375 friend class G1CMParKeepAliveAndDrainClosure;
358 376 friend class G1CMParDrainMarkingStackClosure;
359 377
360 378 protected:
361 379 ConcurrentMarkThread* _cmThread; // the thread doing the work
362 380 G1CollectedHeap* _g1h; // the heap.
363 381 uint _parallel_marking_threads; // the number of marking
364 382 // threads we're use
365 383 uint _max_parallel_marking_threads; // max number of marking
366 384 // threads we'll ever use
367 385 double _sleep_factor; // how much we have to sleep, with
368 386 // respect to the work we just did, to
369 387 // meet the marking overhead goal
370 388 double _marking_task_overhead; // marking target overhead for
371 389 // a single task
372 390
373 391 // same as the two above, but for the cleanup task
374 392 double _cleanup_sleep_factor;
375 393 double _cleanup_task_overhead;
376 394
377 395 FreeRegionList _cleanup_list;
378 396
379 397 // Concurrent marking support structures
380 398 CMBitMap _markBitMap1;
381 399 CMBitMap _markBitMap2;
382 400 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
383 401 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
384 402
385 403 BitMap _region_bm;
386 404 BitMap _card_bm;
387 405
388 406 // Heap bounds
389 407 HeapWord* _heap_start;
390 408 HeapWord* _heap_end;
391 409
392 410 // Root region tracking and claiming.
393 411 CMRootRegions _root_regions;
394 412
395 413 // For gray objects
396 414 CMMarkStack _markStack; // Grey objects behind global finger.
397 415 HeapWord* volatile _finger; // the global finger, region aligned,
398 416 // always points to the end of the
399 417 // last claimed region
400 418
401 419 // marking tasks
402 420 uint _max_worker_id;// maximum worker id
403 421 uint _active_tasks; // task num currently active
404 422 CMTask** _tasks; // task queue array (max_worker_id len)
405 423 CMTaskQueueSet* _task_queues; // task queue set
406 424 ParallelTaskTerminator _terminator; // for termination
407 425
408 426 // Two sync barriers that are used to synchronise tasks when an
409 427 // overflow occurs. The algorithm is the following. All tasks enter
410 428 // the first one to ensure that they have all stopped manipulating
411 429 // the global data structures. After they exit it, they re-initialise
412 430 // their data structures and task 0 re-initialises the global data
413 431 // structures. Then, they enter the second sync barrier. This
414 432 // ensure, that no task starts doing work before all data
415 433 // structures (local and global) have been re-initialised. When they
416 434 // exit it, they are free to start working again.
417 435 WorkGangBarrierSync _first_overflow_barrier_sync;
418 436 WorkGangBarrierSync _second_overflow_barrier_sync;
419 437
420 438 // this is set by any task, when an overflow on the global data
421 439 // structures is detected.
422 440 volatile bool _has_overflown;
423 441 // true: marking is concurrent, false: we're in remark
424 442 volatile bool _concurrent;
425 443 // set at the end of a Full GC so that marking aborts
426 444 volatile bool _has_aborted;
427 445
428 446 // used when remark aborts due to an overflow to indicate that
429 447 // another concurrent marking phase should start
430 448 volatile bool _restart_for_overflow;
431 449
432 450 // This is true from the very start of concurrent marking until the
433 451 // point when all the tasks complete their work. It is really used
434 452 // to determine the points between the end of concurrent marking and
435 453 // time of remark.
436 454 volatile bool _concurrent_marking_in_progress;
437 455
438 456 // verbose level
439 457 CMVerboseLevel _verbose_level;
440 458
441 459 // All of these times are in ms.
442 460 NumberSeq _init_times;
443 461 NumberSeq _remark_times;
444 462 NumberSeq _remark_mark_times;
445 463 NumberSeq _remark_weak_ref_times;
446 464 NumberSeq _cleanup_times;
447 465 double _total_counting_time;
448 466 double _total_rs_scrub_time;
449 467
450 468 double* _accum_task_vtime; // accumulated task vtime
451 469
452 470 FlexibleWorkGang* _parallel_workers;
453 471
454 472 ForceOverflowSettings _force_overflow_conc;
455 473 ForceOverflowSettings _force_overflow_stw;
456 474
457 475 void weakRefsWork(bool clear_all_soft_refs);
458 476
459 477 void swapMarkBitMaps();
460 478
461 479 // It resets the global marking data structures, as well as the
462 480 // task local ones; should be called during initial mark.
463 481 void reset();
464 482 // It resets all the marking data structures.
465 483 void clear_marking_state(bool clear_overflow = true);
466 484
467 485 // It should be called to indicate which phase we're in (concurrent
468 486 // mark or remark) and how many threads are currently active.
469 487 void set_phase(uint active_tasks, bool concurrent);
470 488 // We do this after we're done with marking so that the marking data
471 489 // structures are initialised to a sensible and predictable state.
472 490 void set_non_marking_state();
473 491
474 492 // prints all gathered CM-related statistics
475 493 void print_stats();
476 494
477 495 bool cleanup_list_is_empty() {
478 496 return _cleanup_list.is_empty();
479 497 }
480 498
481 499 // accessor methods
482 500 uint parallel_marking_threads() { return _parallel_marking_threads; }
483 501 uint max_parallel_marking_threads() { return _max_parallel_marking_threads;}
484 502 double sleep_factor() { return _sleep_factor; }
485 503 double marking_task_overhead() { return _marking_task_overhead;}
486 504 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
487 505 double cleanup_task_overhead() { return _cleanup_task_overhead;}
488 506
489 507 HeapWord* finger() { return _finger; }
490 508 bool concurrent() { return _concurrent; }
491 509 uint active_tasks() { return _active_tasks; }
492 510 ParallelTaskTerminator* terminator() { return &_terminator; }
493 511
494 512 // It claims the next available region to be scanned by a marking
495 513 // task/thread. It might return NULL if the next region is empty or
496 514 // we have run out of regions. In the latter case, out_of_regions()
497 515 // determines whether we've really run out of regions or the task
498 516 // should call claim_region() again. This might seem a bit
499 517 // awkward. Originally, the code was written so that claim_region()
500 518 // either successfully returned with a non-empty region or there
501 519 // were no more regions to be claimed. The problem with this was
502 520 // that, in certain circumstances, it iterated over large chunks of
503 521 // the heap finding only empty regions and, while it was working, it
504 522 // was preventing the calling task to call its regular clock
505 523 // method. So, this way, each task will spend very little time in
506 524 // claim_region() and is allowed to call the regular clock method
507 525 // frequently.
508 526 HeapRegion* claim_region(uint worker_id);
509 527
510 528 // It determines whether we've run out of regions to scan.
511 529 bool out_of_regions() { return _finger == _heap_end; }
512 530
513 531 // Returns the task with the given id
514 532 CMTask* task(int id) {
515 533 assert(0 <= id && id < (int) _active_tasks,
516 534 "task id not within active bounds");
517 535 return _tasks[id];
518 536 }
519 537
520 538 // Returns the task queue with the given id
521 539 CMTaskQueue* task_queue(int id) {
522 540 assert(0 <= id && id < (int) _active_tasks,
523 541 "task queue id not within active bounds");
524 542 return (CMTaskQueue*) _task_queues->queue(id);
525 543 }
526 544
527 545 // Returns the task queue set
528 546 CMTaskQueueSet* task_queues() { return _task_queues; }
529 547
530 548 // Access / manipulation of the overflow flag which is set to
531 549 // indicate that the global stack has overflown
532 550 bool has_overflown() { return _has_overflown; }
533 551 void set_has_overflown() { _has_overflown = true; }
534 552 void clear_has_overflown() { _has_overflown = false; }
535 553 bool restart_for_overflow() { return _restart_for_overflow; }
536 554
537 555 bool has_aborted() { return _has_aborted; }
538 556
539 557 // Methods to enter the two overflow sync barriers
540 558 void enter_first_sync_barrier(uint worker_id);
541 559 void enter_second_sync_barrier(uint worker_id);
542 560
543 561 ForceOverflowSettings* force_overflow_conc() {
544 562 return &_force_overflow_conc;
545 563 }
546 564
547 565 ForceOverflowSettings* force_overflow_stw() {
548 566 return &_force_overflow_stw;
549 567 }
550 568
551 569 ForceOverflowSettings* force_overflow() {
552 570 if (concurrent()) {
553 571 return force_overflow_conc();
554 572 } else {
555 573 return force_overflow_stw();
556 574 }
557 575 }
558 576
559 577 // Live Data Counting data structures...
560 578 // These data structures are initialized at the start of
561 579 // marking. They are written to while marking is active.
562 580 // They are aggregated during remark; the aggregated values
563 581 // are then used to populate the _region_bm, _card_bm, and
564 582 // the total live bytes, which are then subsequently updated
565 583 // during cleanup.
566 584
567 585 // An array of bitmaps (one bit map per task). Each bitmap
568 586 // is used to record the cards spanned by the live objects
569 587 // marked by that task/worker.
↓ open down ↓ |
213 lines elided |
↑ open up ↑ |
570 588 BitMap* _count_card_bitmaps;
571 589
572 590 // Used to record the number of marked live bytes
573 591 // (for each region, by worker thread).
574 592 size_t** _count_marked_bytes;
575 593
576 594 // Card index of the bottom of the G1 heap. Used for biasing indices into
577 595 // the card bitmaps.
578 596 intptr_t _heap_bottom_card_num;
579 597
598 + // Set to true when initialization is complete
599 + bool _completed_initialization;
600 +
580 601 public:
581 602 // Manipulation of the global mark stack.
582 603 // Notice that the first mark_stack_push is CAS-based, whereas the
583 604 // two below are Mutex-based. This is OK since the first one is only
584 605 // called during evacuation pauses and doesn't compete with the
585 606 // other two (which are called by the marking tasks during
586 607 // concurrent marking or remark).
587 608 bool mark_stack_push(oop p) {
588 609 _markStack.par_push(p);
589 610 if (_markStack.overflow()) {
590 611 set_has_overflown();
591 612 return false;
592 613 }
593 614 return true;
594 615 }
595 616 bool mark_stack_push(oop* arr, int n) {
596 617 _markStack.par_push_arr(arr, n);
597 618 if (_markStack.overflow()) {
598 619 set_has_overflown();
599 620 return false;
600 621 }
601 622 return true;
602 623 }
603 624 void mark_stack_pop(oop* arr, int max, int* n) {
604 625 _markStack.par_pop_arr(arr, max, n);
605 626 }
606 627 size_t mark_stack_size() { return _markStack.size(); }
607 628 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
608 629 bool mark_stack_overflow() { return _markStack.overflow(); }
609 630 bool mark_stack_empty() { return _markStack.isEmpty(); }
610 631
611 632 CMRootRegions* root_regions() { return &_root_regions; }
612 633
613 634 bool concurrent_marking_in_progress() {
614 635 return _concurrent_marking_in_progress;
615 636 }
616 637 void set_concurrent_marking_in_progress() {
617 638 _concurrent_marking_in_progress = true;
618 639 }
619 640 void clear_concurrent_marking_in_progress() {
620 641 _concurrent_marking_in_progress = false;
621 642 }
622 643
623 644 void update_accum_task_vtime(int i, double vtime) {
624 645 _accum_task_vtime[i] += vtime;
625 646 }
626 647
627 648 double all_task_accum_vtime() {
628 649 double ret = 0.0;
↓ open down ↓ |
39 lines elided |
↑ open up ↑ |
629 650 for (uint i = 0; i < _max_worker_id; ++i)
630 651 ret += _accum_task_vtime[i];
631 652 return ret;
632 653 }
633 654
634 655 // Attempts to steal an object from the task queues of other tasks
635 656 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
636 657 return _task_queues->steal(worker_id, hash_seed, obj);
637 658 }
638 659
639 - ConcurrentMark(ReservedSpace rs, uint max_regions);
660 + ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
640 661 ~ConcurrentMark();
641 662
642 663 ConcurrentMarkThread* cmThread() { return _cmThread; }
643 664
644 665 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
645 666 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
646 667
647 668 // Returns the number of GC threads to be used in a concurrent
648 669 // phase based on the number of GC threads being used in a STW
649 670 // phase.
650 671 uint scale_parallel_threads(uint n_par_threads);
651 672
652 673 // Calculates the number of GC threads to be used in a concurrent phase.
653 674 uint calc_parallel_marking_threads();
654 675
655 676 // The following three are interaction between CM and
656 677 // G1CollectedHeap
657 678
658 679 // This notifies CM that a root during initial-mark needs to be
659 680 // grayed. It is MT-safe. word_size is the size of the object in
660 681 // words. It is passed explicitly as sometimes we cannot calculate
661 682 // it from the given object because it might be in an inconsistent
662 683 // state (e.g., in to-space and being copied). So the caller is
663 684 // responsible for dealing with this issue (e.g., get the size from
664 685 // the from-space image when the to-space image might be
665 686 // inconsistent) and always passing the size. hr is the region that
666 687 // contains the object and it's passed optionally from callers who
667 688 // might already have it (no point in recalculating it).
668 689 inline void grayRoot(oop obj, size_t word_size,
669 690 uint worker_id, HeapRegion* hr = NULL);
670 691
671 692 // It iterates over the heap and for each object it comes across it
672 693 // will dump the contents of its reference fields, as well as
673 694 // liveness information for the object and its referents. The dump
674 695 // will be written to a file with the following name:
675 696 // G1PrintReachableBaseFile + "." + str.
676 697 // vo decides whether the prev (vo == UsePrevMarking), the next
677 698 // (vo == UseNextMarking) marking information, or the mark word
678 699 // (vo == UseMarkWord) will be used to determine the liveness of
679 700 // each object / referent.
680 701 // If all is true, all objects in the heap will be dumped, otherwise
681 702 // only the live ones. In the dump the following symbols / breviations
682 703 // are used:
683 704 // M : an explicitly live object (its bitmap bit is set)
684 705 // > : an implicitly live object (over tams)
685 706 // O : an object outside the G1 heap (typically: in the perm gen)
686 707 // NOT : a reference field whose referent is not live
687 708 // AND MARKED : indicates that an object is both explicitly and
688 709 // implicitly live (it should be one or the other, not both)
689 710 void print_reachable(const char* str,
690 711 VerifyOption vo, bool all) PRODUCT_RETURN;
691 712
692 713 // Clear the next marking bitmap (will be called concurrently).
693 714 void clearNextBitmap();
694 715
695 716 // These two do the work that needs to be done before and after the
696 717 // initial root checkpoint. Since this checkpoint can be done at two
697 718 // different points (i.e. an explicit pause or piggy-backed on a
698 719 // young collection), then it's nice to be able to easily share the
699 720 // pre/post code. It might be the case that we can put everything in
700 721 // the post method. TP
701 722 void checkpointRootsInitialPre();
702 723 void checkpointRootsInitialPost();
703 724
704 725 // Scan all the root regions and mark everything reachable from
705 726 // them.
706 727 void scanRootRegions();
707 728
708 729 // Scan a single root region and mark everything reachable from it.
709 730 void scanRootRegion(HeapRegion* hr, uint worker_id);
710 731
711 732 // Do concurrent phase of marking, to a tentative transitive closure.
712 733 void markFromRoots();
713 734
714 735 void checkpointRootsFinal(bool clear_all_soft_refs);
715 736 void checkpointRootsFinalWork();
716 737 void cleanup();
717 738 void completeCleanup();
718 739
719 740 // Mark in the previous bitmap. NB: this is usually read-only, so use
720 741 // this carefully!
721 742 inline void markPrev(oop p);
722 743
723 744 // Clears marks for all objects in the given range, for the prev,
724 745 // next, or both bitmaps. NB: the previous bitmap is usually
725 746 // read-only, so use this carefully!
726 747 void clearRangePrevBitmap(MemRegion mr);
727 748 void clearRangeNextBitmap(MemRegion mr);
728 749 void clearRangeBothBitmaps(MemRegion mr);
729 750
730 751 // Notify data structures that a GC has started.
731 752 void note_start_of_gc() {
732 753 _markStack.note_start_of_gc();
733 754 }
734 755
735 756 // Notify data structures that a GC is finished.
736 757 void note_end_of_gc() {
737 758 _markStack.note_end_of_gc();
738 759 }
739 760
740 761 // Verify that there are no CSet oops on the stacks (taskqueues /
741 762 // global mark stack), enqueued SATB buffers, per-thread SATB
742 763 // buffers, and fingers (global / per-task). The boolean parameters
743 764 // decide which of the above data structures to verify. If marking
744 765 // is not in progress, it's a no-op.
745 766 void verify_no_cset_oops(bool verify_stacks,
746 767 bool verify_enqueued_buffers,
747 768 bool verify_thread_buffers,
748 769 bool verify_fingers) PRODUCT_RETURN;
749 770
750 771 // It is called at the end of an evacuation pause during marking so
751 772 // that CM is notified of where the new end of the heap is. It
752 773 // doesn't do anything if concurrent_marking_in_progress() is false,
753 774 // unless the force parameter is true.
754 775 void update_g1_committed(bool force = false);
755 776
756 777 bool isMarked(oop p) const {
757 778 assert(p != NULL && p->is_oop(), "expected an oop");
758 779 HeapWord* addr = (HeapWord*)p;
759 780 assert(addr >= _nextMarkBitMap->startWord() ||
760 781 addr < _nextMarkBitMap->endWord(), "in a region");
761 782
762 783 return _nextMarkBitMap->isMarked(addr);
763 784 }
764 785
765 786 inline bool not_yet_marked(oop p) const;
766 787
767 788 // XXX Debug code
768 789 bool containing_card_is_marked(void* p);
769 790 bool containing_cards_are_marked(void* start, void* last);
770 791
771 792 bool isPrevMarked(oop p) const {
772 793 assert(p != NULL && p->is_oop(), "expected an oop");
773 794 HeapWord* addr = (HeapWord*)p;
774 795 assert(addr >= _prevMarkBitMap->startWord() ||
775 796 addr < _prevMarkBitMap->endWord(), "in a region");
776 797
777 798 return _prevMarkBitMap->isMarked(addr);
778 799 }
779 800
780 801 inline bool do_yield_check(uint worker_i = 0);
781 802 inline bool should_yield();
782 803
783 804 // Called to abort the marking cycle after a Full GC takes palce.
784 805 void abort();
785 806
786 807 // This prints the global/local fingers. It is used for debugging.
787 808 NOT_PRODUCT(void print_finger();)
788 809
789 810 void print_summary_info();
790 811
791 812 void print_worker_threads_on(outputStream* st) const;
792 813
793 814 // The following indicate whether a given verbose level has been
794 815 // set. Notice that anything above stats is conditional to
795 816 // _MARKING_VERBOSE_ having been set to 1
796 817 bool verbose_stats() {
797 818 return _verbose_level >= stats_verbose;
798 819 }
799 820 bool verbose_low() {
800 821 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
801 822 }
802 823 bool verbose_medium() {
803 824 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
804 825 }
805 826 bool verbose_high() {
806 827 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
807 828 }
808 829
809 830 // Liveness counting
810 831
811 832 // Utility routine to set an exclusive range of cards on the given
812 833 // card liveness bitmap
813 834 inline void set_card_bitmap_range(BitMap* card_bm,
814 835 BitMap::idx_t start_idx,
815 836 BitMap::idx_t end_idx,
816 837 bool is_par);
817 838
818 839 // Returns the card number of the bottom of the G1 heap.
819 840 // Used in biasing indices into accounting card bitmaps.
820 841 intptr_t heap_bottom_card_num() const {
821 842 return _heap_bottom_card_num;
822 843 }
823 844
824 845 // Returns the card bitmap for a given task or worker id.
825 846 BitMap* count_card_bitmap_for(uint worker_id) {
826 847 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
827 848 assert(_count_card_bitmaps != NULL, "uninitialized");
828 849 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
829 850 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
830 851 return task_card_bm;
831 852 }
832 853
833 854 // Returns the array containing the marked bytes for each region,
834 855 // for the given worker or task id.
835 856 size_t* count_marked_bytes_array_for(uint worker_id) {
836 857 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
837 858 assert(_count_marked_bytes != NULL, "uninitialized");
838 859 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
839 860 assert(marked_bytes_array != NULL, "uninitialized");
840 861 return marked_bytes_array;
841 862 }
842 863
843 864 // Returns the index in the liveness accounting card table bitmap
844 865 // for the given address
845 866 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
846 867
847 868 // Counts the size of the given memory region in the the given
848 869 // marked_bytes array slot for the given HeapRegion.
849 870 // Sets the bits in the given card bitmap that are associated with the
850 871 // cards that are spanned by the memory region.
851 872 inline void count_region(MemRegion mr, HeapRegion* hr,
852 873 size_t* marked_bytes_array,
853 874 BitMap* task_card_bm);
854 875
855 876 // Counts the given memory region in the task/worker counting
856 877 // data structures for the given worker id.
857 878 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
858 879
859 880 // Counts the given memory region in the task/worker counting
860 881 // data structures for the given worker id.
861 882 inline void count_region(MemRegion mr, uint worker_id);
862 883
863 884 // Counts the given object in the given task/worker counting
864 885 // data structures.
865 886 inline void count_object(oop obj, HeapRegion* hr,
866 887 size_t* marked_bytes_array,
867 888 BitMap* task_card_bm);
868 889
869 890 // Counts the given object in the task/worker counting data
870 891 // structures for the given worker id.
871 892 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
872 893
873 894 // Attempts to mark the given object and, if successful, counts
874 895 // the object in the given task/worker counting structures.
875 896 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
876 897 size_t* marked_bytes_array,
877 898 BitMap* task_card_bm);
878 899
879 900 // Attempts to mark the given object and, if successful, counts
880 901 // the object in the task/worker counting structures for the
881 902 // given worker id.
882 903 inline bool par_mark_and_count(oop obj, size_t word_size,
883 904 HeapRegion* hr, uint worker_id);
884 905
885 906 // Attempts to mark the given object and, if successful, counts
886 907 // the object in the task/worker counting structures for the
887 908 // given worker id.
888 909 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
889 910
890 911 // Similar to the above routine but we don't know the heap region that
891 912 // contains the object to be marked/counted, which this routine looks up.
892 913 inline bool par_mark_and_count(oop obj, uint worker_id);
893 914
894 915 // Similar to the above routine but there are times when we cannot
895 916 // safely calculate the size of obj due to races and we, therefore,
896 917 // pass the size in as a parameter. It is the caller's reponsibility
897 918 // to ensure that the size passed in for obj is valid.
898 919 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
899 920
↓ open down ↓ |
250 lines elided |
↑ open up ↑ |
900 921 // Unconditionally mark the given object, and unconditinally count
901 922 // the object in the counting structures for worker id 0.
902 923 // Should *not* be called from parallel code.
903 924 inline bool mark_and_count(oop obj, HeapRegion* hr);
904 925
905 926 // Similar to the above routine but we don't know the heap region that
906 927 // contains the object to be marked/counted, which this routine looks up.
907 928 // Should *not* be called from parallel code.
908 929 inline bool mark_and_count(oop obj);
909 930
931 + // Returns true if initialization was successfully completed.
932 + bool completed_initialization() const {
933 + return _completed_initialization;
934 + }
935 +
910 936 protected:
911 937 // Clear all the per-task bitmaps and arrays used to store the
912 938 // counting data.
913 939 void clear_all_count_data();
914 940
915 941 // Aggregates the counting data for each worker/task
916 942 // that was constructed while marking. Also sets
917 943 // the amount of marked bytes for each region and
918 944 // the top at concurrent mark count.
919 945 void aggregate_count_data();
920 946
921 947 // Verification routine
922 948 void verify_count_data();
923 949 };
924 950
925 951 // A class representing a marking task.
926 952 class CMTask : public TerminatorTerminator {
927 953 private:
928 954 enum PrivateConstants {
929 955 // the regular clock call is called once the scanned words reaches
930 956 // this limit
931 957 words_scanned_period = 12*1024,
932 958 // the regular clock call is called once the number of visited
933 959 // references reaches this limit
934 960 refs_reached_period = 384,
935 961 // initial value for the hash seed, used in the work stealing code
936 962 init_hash_seed = 17,
937 963 // how many entries will be transferred between global stack and
938 964 // local queues
939 965 global_stack_transfer_size = 16
940 966 };
941 967
942 968 uint _worker_id;
943 969 G1CollectedHeap* _g1h;
944 970 ConcurrentMark* _cm;
945 971 CMBitMap* _nextMarkBitMap;
946 972 // the task queue of this task
947 973 CMTaskQueue* _task_queue;
948 974 private:
949 975 // the task queue set---needed for stealing
950 976 CMTaskQueueSet* _task_queues;
951 977 // indicates whether the task has been claimed---this is only for
952 978 // debugging purposes
953 979 bool _claimed;
954 980
955 981 // number of calls to this task
956 982 int _calls;
957 983
958 984 // when the virtual timer reaches this time, the marking step should
959 985 // exit
960 986 double _time_target_ms;
961 987 // the start time of the current marking step
962 988 double _start_time_ms;
963 989
964 990 // the oop closure used for iterations over oops
965 991 G1CMOopClosure* _cm_oop_closure;
966 992
967 993 // the region this task is scanning, NULL if we're not scanning any
968 994 HeapRegion* _curr_region;
969 995 // the local finger of this task, NULL if we're not scanning a region
970 996 HeapWord* _finger;
971 997 // limit of the region this task is scanning, NULL if we're not scanning one
972 998 HeapWord* _region_limit;
973 999
974 1000 // the number of words this task has scanned
975 1001 size_t _words_scanned;
976 1002 // When _words_scanned reaches this limit, the regular clock is
977 1003 // called. Notice that this might be decreased under certain
978 1004 // circumstances (i.e. when we believe that we did an expensive
979 1005 // operation).
980 1006 size_t _words_scanned_limit;
981 1007 // the initial value of _words_scanned_limit (i.e. what it was
982 1008 // before it was decreased).
983 1009 size_t _real_words_scanned_limit;
984 1010
985 1011 // the number of references this task has visited
986 1012 size_t _refs_reached;
987 1013 // When _refs_reached reaches this limit, the regular clock is
988 1014 // called. Notice this this might be decreased under certain
989 1015 // circumstances (i.e. when we believe that we did an expensive
990 1016 // operation).
991 1017 size_t _refs_reached_limit;
992 1018 // the initial value of _refs_reached_limit (i.e. what it was before
993 1019 // it was decreased).
994 1020 size_t _real_refs_reached_limit;
995 1021
996 1022 // used by the work stealing stuff
997 1023 int _hash_seed;
998 1024 // if this is true, then the task has aborted for some reason
999 1025 bool _has_aborted;
1000 1026 // set when the task aborts because it has met its time quota
1001 1027 bool _has_timed_out;
1002 1028 // true when we're draining SATB buffers; this avoids the task
1003 1029 // aborting due to SATB buffers being available (as we're already
1004 1030 // dealing with them)
1005 1031 bool _draining_satb_buffers;
1006 1032
1007 1033 // number sequence of past step times
1008 1034 NumberSeq _step_times_ms;
1009 1035 // elapsed time of this task
1010 1036 double _elapsed_time_ms;
1011 1037 // termination time of this task
1012 1038 double _termination_time_ms;
1013 1039 // when this task got into the termination protocol
1014 1040 double _termination_start_time_ms;
1015 1041
1016 1042 // true when the task is during a concurrent phase, false when it is
1017 1043 // in the remark phase (so, in the latter case, we do not have to
1018 1044 // check all the things that we have to check during the concurrent
1019 1045 // phase, i.e. SATB buffer availability...)
1020 1046 bool _concurrent;
1021 1047
1022 1048 TruncatedSeq _marking_step_diffs_ms;
1023 1049
1024 1050 // Counting data structures. Embedding the task's marked_bytes_array
1025 1051 // and card bitmap into the actual task saves having to go through
1026 1052 // the ConcurrentMark object.
1027 1053 size_t* _marked_bytes_array;
1028 1054 BitMap* _card_bm;
1029 1055
1030 1056 // LOTS of statistics related with this task
1031 1057 #if _MARKING_STATS_
1032 1058 NumberSeq _all_clock_intervals_ms;
1033 1059 double _interval_start_time_ms;
1034 1060
1035 1061 int _aborted;
1036 1062 int _aborted_overflow;
1037 1063 int _aborted_cm_aborted;
1038 1064 int _aborted_yield;
1039 1065 int _aborted_timed_out;
1040 1066 int _aborted_satb;
1041 1067 int _aborted_termination;
1042 1068
1043 1069 int _steal_attempts;
1044 1070 int _steals;
1045 1071
1046 1072 int _clock_due_to_marking;
1047 1073 int _clock_due_to_scanning;
1048 1074
1049 1075 int _local_pushes;
1050 1076 int _local_pops;
1051 1077 int _local_max_size;
1052 1078 int _objs_scanned;
1053 1079
1054 1080 int _global_pushes;
1055 1081 int _global_pops;
1056 1082 int _global_max_size;
1057 1083
1058 1084 int _global_transfers_to;
1059 1085 int _global_transfers_from;
1060 1086
1061 1087 int _regions_claimed;
1062 1088 int _objs_found_on_bitmap;
1063 1089
1064 1090 int _satb_buffers_processed;
1065 1091 #endif // _MARKING_STATS_
1066 1092
1067 1093 // it updates the local fields after this task has claimed
1068 1094 // a new region to scan
1069 1095 void setup_for_region(HeapRegion* hr);
1070 1096 // it brings up-to-date the limit of the region
1071 1097 void update_region_limit();
1072 1098
1073 1099 // called when either the words scanned or the refs visited limit
1074 1100 // has been reached
1075 1101 void reached_limit();
1076 1102 // recalculates the words scanned and refs visited limits
1077 1103 void recalculate_limits();
1078 1104 // decreases the words scanned and refs visited limits when we reach
1079 1105 // an expensive operation
1080 1106 void decrease_limits();
1081 1107 // it checks whether the words scanned or refs visited reached their
1082 1108 // respective limit and calls reached_limit() if they have
1083 1109 void check_limits() {
1084 1110 if (_words_scanned >= _words_scanned_limit ||
1085 1111 _refs_reached >= _refs_reached_limit) {
1086 1112 reached_limit();
1087 1113 }
1088 1114 }
1089 1115 // this is supposed to be called regularly during a marking step as
1090 1116 // it checks a bunch of conditions that might cause the marking step
1091 1117 // to abort
1092 1118 void regular_clock_call();
1093 1119 bool concurrent() { return _concurrent; }
1094 1120
1095 1121 public:
1096 1122 // It resets the task; it should be called right at the beginning of
1097 1123 // a marking phase.
1098 1124 void reset(CMBitMap* _nextMarkBitMap);
1099 1125 // it clears all the fields that correspond to a claimed region.
1100 1126 void clear_region_fields();
1101 1127
1102 1128 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1103 1129
1104 1130 // The main method of this class which performs a marking step
1105 1131 // trying not to exceed the given duration. However, it might exit
1106 1132 // prematurely, according to some conditions (i.e. SATB buffers are
1107 1133 // available for processing).
1108 1134 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1109 1135
1110 1136 // These two calls start and stop the timer
1111 1137 void record_start_time() {
1112 1138 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1113 1139 }
1114 1140 void record_end_time() {
1115 1141 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1116 1142 }
1117 1143
1118 1144 // returns the worker ID associated with this task.
1119 1145 uint worker_id() { return _worker_id; }
1120 1146
1121 1147 // From TerminatorTerminator. It determines whether this task should
1122 1148 // exit the termination protocol after it's entered it.
1123 1149 virtual bool should_exit_termination();
1124 1150
1125 1151 // Resets the local region fields after a task has finished scanning a
1126 1152 // region; or when they have become stale as a result of the region
1127 1153 // being evacuated.
1128 1154 void giveup_current_region();
1129 1155
1130 1156 HeapWord* finger() { return _finger; }
1131 1157
1132 1158 bool has_aborted() { return _has_aborted; }
1133 1159 void set_has_aborted() { _has_aborted = true; }
1134 1160 void clear_has_aborted() { _has_aborted = false; }
1135 1161 bool has_timed_out() { return _has_timed_out; }
1136 1162 bool claimed() { return _claimed; }
1137 1163
1138 1164 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1139 1165
1140 1166 // It grays the object by marking it and, if necessary, pushing it
1141 1167 // on the local queue
1142 1168 inline void deal_with_reference(oop obj);
1143 1169
1144 1170 // It scans an object and visits its children.
1145 1171 void scan_object(oop obj);
1146 1172
1147 1173 // It pushes an object on the local queue.
1148 1174 inline void push(oop obj);
1149 1175
1150 1176 // These two move entries to/from the global stack.
1151 1177 void move_entries_to_global_stack();
1152 1178 void get_entries_from_global_stack();
1153 1179
1154 1180 // It pops and scans objects from the local queue. If partially is
1155 1181 // true, then it stops when the queue size is of a given limit. If
1156 1182 // partially is false, then it stops when the queue is empty.
1157 1183 void drain_local_queue(bool partially);
1158 1184 // It moves entries from the global stack to the local queue and
1159 1185 // drains the local queue. If partially is true, then it stops when
1160 1186 // both the global stack and the local queue reach a given size. If
1161 1187 // partially if false, it tries to empty them totally.
1162 1188 void drain_global_stack(bool partially);
1163 1189 // It keeps picking SATB buffers and processing them until no SATB
1164 1190 // buffers are available.
1165 1191 void drain_satb_buffers();
1166 1192
1167 1193 // moves the local finger to a new location
1168 1194 inline void move_finger_to(HeapWord* new_finger) {
1169 1195 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1170 1196 _finger = new_finger;
1171 1197 }
1172 1198
1173 1199 CMTask(uint worker_id, ConcurrentMark *cm,
1174 1200 size_t* marked_bytes, BitMap* card_bm,
1175 1201 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1176 1202
1177 1203 // it prints statistics associated with this task
1178 1204 void print_stats();
1179 1205
1180 1206 #if _MARKING_STATS_
1181 1207 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1182 1208 #endif // _MARKING_STATS_
1183 1209 };
1184 1210
1185 1211 // Class that's used to to print out per-region liveness
1186 1212 // information. It's currently used at the end of marking and also
1187 1213 // after we sort the old regions at the end of the cleanup operation.
1188 1214 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1189 1215 private:
1190 1216 outputStream* _out;
1191 1217
1192 1218 // Accumulators for these values.
1193 1219 size_t _total_used_bytes;
1194 1220 size_t _total_capacity_bytes;
1195 1221 size_t _total_prev_live_bytes;
1196 1222 size_t _total_next_live_bytes;
1197 1223
1198 1224 // These are set up when we come across a "stars humongous" region
1199 1225 // (as this is where most of this information is stored, not in the
1200 1226 // subsequent "continues humongous" regions). After that, for every
1201 1227 // region in a given humongous region series we deduce the right
1202 1228 // values for it by simply subtracting the appropriate amount from
1203 1229 // these fields. All these values should reach 0 after we've visited
1204 1230 // the last region in the series.
1205 1231 size_t _hum_used_bytes;
1206 1232 size_t _hum_capacity_bytes;
1207 1233 size_t _hum_prev_live_bytes;
1208 1234 size_t _hum_next_live_bytes;
1209 1235
1210 1236 static double perc(size_t val, size_t total) {
1211 1237 if (total == 0) {
1212 1238 return 0.0;
1213 1239 } else {
1214 1240 return 100.0 * ((double) val / (double) total);
1215 1241 }
1216 1242 }
1217 1243
1218 1244 static double bytes_to_mb(size_t val) {
1219 1245 return (double) val / (double) M;
1220 1246 }
1221 1247
1222 1248 // See the .cpp file.
1223 1249 size_t get_hum_bytes(size_t* hum_bytes);
1224 1250 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1225 1251 size_t* prev_live_bytes, size_t* next_live_bytes);
1226 1252
1227 1253 public:
1228 1254 // The header and footer are printed in the constructor and
1229 1255 // destructor respectively.
1230 1256 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1231 1257 virtual bool doHeapRegion(HeapRegion* r);
1232 1258 ~G1PrintRegionLivenessInfoClosure();
1233 1259 };
1234 1260
1235 1261 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
↓ open down ↓ |
316 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX