annotate src/share/vm/gc_implementation/g1/concurrentMark.hpp @ 1857:4805b9f4779e

6941395: G1: Use only lock-free versions of region stack push() and pop() Summary: Re-enable use of the lock-free versions of region stack push() and pop() by recording aborted regions in a thread-local structure, which are then processed when scanning of the region stack restarts. The previous locking versions of these routines are retained for diagnostic purposes. Reviewed-by: tonyp, ysr
author johnc
date Tue, 28 Sep 2010 09:51:37 -0700
parents c18cbe5936b8
children f95d63e2154a
rev   line source
ysr@345 1 /*
trims@1563 2 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
ysr@345 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
ysr@345 4 *
ysr@345 5 * This code is free software; you can redistribute it and/or modify it
ysr@345 6 * under the terms of the GNU General Public License version 2 only, as
ysr@345 7 * published by the Free Software Foundation.
ysr@345 8 *
ysr@345 9 * This code is distributed in the hope that it will be useful, but WITHOUT
ysr@345 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
ysr@345 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
ysr@345 12 * version 2 for more details (a copy is included in the LICENSE file that
ysr@345 13 * accompanied this code).
ysr@345 14 *
ysr@345 15 * You should have received a copy of the GNU General Public License version
ysr@345 16 * 2 along with this work; if not, write to the Free Software Foundation,
ysr@345 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
ysr@345 18 *
trims@1563 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1563 20 * or visit www.oracle.com if you need additional information or have any
trims@1563 21 * questions.
ysr@345 22 *
ysr@345 23 */
ysr@345 24
ysr@345 25 class G1CollectedHeap;
ysr@345 26 class CMTask;
jcoomes@1398 27 typedef GenericTaskQueue<oop> CMTaskQueue;
jcoomes@1398 28 typedef GenericTaskQueueSet<CMTaskQueue> CMTaskQueueSet;
ysr@345 29
ysr@345 30 // A generic CM bit map. This is essentially a wrapper around the BitMap
ysr@345 31 // class, with one bit per (1<<_shifter) HeapWords.
ysr@345 32
apetrusenko@581 33 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
ysr@345 34 protected:
ysr@345 35 HeapWord* _bmStartWord; // base address of range covered by map
ysr@345 36 size_t _bmWordSize; // map size (in #HeapWords covered)
ysr@345 37 const int _shifter; // map to char or bit
ysr@345 38 VirtualSpace _virtual_space; // underlying the bit map
ysr@345 39 BitMap _bm; // the bit map itself
ysr@345 40
ysr@345 41 public:
ysr@345 42 // constructor
ysr@345 43 CMBitMapRO(ReservedSpace rs, int shifter);
ysr@345 44
ysr@345 45 enum { do_yield = true };
ysr@345 46
ysr@345 47 // inquiries
ysr@345 48 HeapWord* startWord() const { return _bmStartWord; }
ysr@345 49 size_t sizeInWords() const { return _bmWordSize; }
ysr@345 50 // the following is one past the last word in space
ysr@345 51 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
ysr@345 52
ysr@345 53 // read marks
ysr@345 54
ysr@345 55 bool isMarked(HeapWord* addr) const {
ysr@345 56 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@345 57 "outside underlying space?");
ysr@345 58 return _bm.at(heapWordToOffset(addr));
ysr@345 59 }
ysr@345 60
ysr@345 61 // iteration
ysr@345 62 bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }
ysr@345 63 bool iterate(BitMapClosure* cl, MemRegion mr);
ysr@345 64
ysr@345 65 // Return the address corresponding to the next marked bit at or after
ysr@345 66 // "addr", and before "limit", if "limit" is non-NULL. If there is no
ysr@345 67 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
ysr@345 68 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
ysr@345 69 HeapWord* limit = NULL) const;
ysr@345 70 // Return the address corresponding to the next unmarked bit at or after
ysr@345 71 // "addr", and before "limit", if "limit" is non-NULL. If there is no
ysr@345 72 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
ysr@345 73 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
ysr@345 74 HeapWord* limit = NULL) const;
ysr@345 75
ysr@345 76 // conversion utilities
ysr@345 77 // XXX Fix these so that offsets are size_t's...
ysr@345 78 HeapWord* offsetToHeapWord(size_t offset) const {
ysr@345 79 return _bmStartWord + (offset << _shifter);
ysr@345 80 }
ysr@345 81 size_t heapWordToOffset(HeapWord* addr) const {
ysr@345 82 return pointer_delta(addr, _bmStartWord) >> _shifter;
ysr@345 83 }
ysr@345 84 int heapWordDiffToOffsetDiff(size_t diff) const;
ysr@345 85 HeapWord* nextWord(HeapWord* addr) {
ysr@345 86 return offsetToHeapWord(heapWordToOffset(addr) + 1);
ysr@345 87 }
ysr@345 88
ysr@345 89 void mostly_disjoint_range_union(BitMap* from_bitmap,
ysr@345 90 size_t from_start_index,
ysr@345 91 HeapWord* to_start_word,
ysr@345 92 size_t word_num);
ysr@345 93
ysr@345 94 // debugging
ysr@345 95 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
ysr@345 96 };
ysr@345 97
ysr@345 98 class CMBitMap : public CMBitMapRO {
ysr@345 99
ysr@345 100 public:
ysr@345 101 // constructor
ysr@345 102 CMBitMap(ReservedSpace rs, int shifter) :
ysr@345 103 CMBitMapRO(rs, shifter) {}
ysr@345 104
ysr@345 105 // write marks
ysr@345 106 void mark(HeapWord* addr) {
ysr@345 107 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@345 108 "outside underlying space?");
ysr@345 109 _bm.at_put(heapWordToOffset(addr), true);
ysr@345 110 }
ysr@345 111 void clear(HeapWord* addr) {
ysr@345 112 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@345 113 "outside underlying space?");
ysr@345 114 _bm.at_put(heapWordToOffset(addr), false);
ysr@345 115 }
ysr@345 116 bool parMark(HeapWord* addr) {
ysr@345 117 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@345 118 "outside underlying space?");
ysr@345 119 return _bm.par_at_put(heapWordToOffset(addr), true);
ysr@345 120 }
ysr@345 121 bool parClear(HeapWord* addr) {
ysr@345 122 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@345 123 "outside underlying space?");
ysr@345 124 return _bm.par_at_put(heapWordToOffset(addr), false);
ysr@345 125 }
ysr@345 126 void markRange(MemRegion mr);
ysr@345 127 void clearAll();
ysr@345 128 void clearRange(MemRegion mr);
ysr@345 129
ysr@345 130 // Starting at the bit corresponding to "addr" (inclusive), find the next
ysr@345 131 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
ysr@345 132 // the end of this run (stopping at "end_addr"). Return the MemRegion
ysr@345 133 // covering from the start of the region corresponding to the first bit
ysr@345 134 // of the run to the end of the region corresponding to the last bit of
ysr@345 135 // the run. If there is no "1" bit at or after "addr", return an empty
ysr@345 136 // MemRegion.
ysr@345 137 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
ysr@345 138 };
ysr@345 139
ysr@345 140 // Represents a marking stack used by the CM collector.
ysr@345 141 // Ideally this should be GrowableArray<> just like MSC's marking stack(s).
apetrusenko@581 142 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
ysr@345 143 ConcurrentMark* _cm;
ysr@345 144 oop* _base; // bottom of stack
ysr@345 145 jint _index; // one more than last occupied index
ysr@345 146 jint _capacity; // max #elements
ysr@345 147 jint _oops_do_bound; // Number of elements to include in next iteration.
ysr@345 148 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
ysr@345 149
ysr@345 150 bool _overflow;
ysr@345 151 DEBUG_ONLY(bool _drain_in_progress;)
ysr@345 152 DEBUG_ONLY(bool _drain_in_progress_yields;)
ysr@345 153
ysr@345 154 public:
ysr@345 155 CMMarkStack(ConcurrentMark* cm);
ysr@345 156 ~CMMarkStack();
ysr@345 157
ysr@345 158 void allocate(size_t size);
ysr@345 159
ysr@345 160 oop pop() {
ysr@345 161 if (!isEmpty()) {
ysr@345 162 return _base[--_index] ;
ysr@345 163 }
ysr@345 164 return NULL;
ysr@345 165 }
ysr@345 166
ysr@345 167 // If overflow happens, don't do the push, and record the overflow.
ysr@345 168 // *Requires* that "ptr" is already marked.
ysr@345 169 void push(oop ptr) {
ysr@345 170 if (isFull()) {
ysr@345 171 // Record overflow.
ysr@345 172 _overflow = true;
ysr@345 173 return;
ysr@345 174 } else {
ysr@345 175 _base[_index++] = ptr;
ysr@345 176 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
ysr@345 177 }
ysr@345 178 }
ysr@345 179 // Non-block impl. Note: concurrency is allowed only with other
ysr@345 180 // "par_push" operations, not with "pop" or "drain". We would need
ysr@345 181 // parallel versions of them if such concurrency was desired.
ysr@345 182 void par_push(oop ptr);
ysr@345 183
ysr@345 184 // Pushes the first "n" elements of "ptr_arr" on the stack.
ysr@345 185 // Non-block impl. Note: concurrency is allowed only with other
ysr@345 186 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
ysr@345 187 void par_adjoin_arr(oop* ptr_arr, int n);
ysr@345 188
ysr@345 189 // Pushes the first "n" elements of "ptr_arr" on the stack.
ysr@345 190 // Locking impl: concurrency is allowed only with
ysr@345 191 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
ysr@345 192 // locking strategy.
ysr@345 193 void par_push_arr(oop* ptr_arr, int n);
ysr@345 194
ysr@345 195 // If returns false, the array was empty. Otherwise, removes up to "max"
ysr@345 196 // elements from the stack, and transfers them to "ptr_arr" in an
ysr@345 197 // unspecified order. The actual number transferred is given in "n" ("n
ysr@345 198 // == 0" is deliberately redundant with the return value.) Locking impl:
ysr@345 199 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
ysr@345 200 // operations, which use the same locking strategy.
ysr@345 201 bool par_pop_arr(oop* ptr_arr, int max, int* n);
ysr@345 202
ysr@345 203 // Drain the mark stack, applying the given closure to all fields of
ysr@345 204 // objects on the stack. (That is, continue until the stack is empty,
ysr@345 205 // even if closure applications add entries to the stack.) The "bm"
ysr@345 206 // argument, if non-null, may be used to verify that only marked objects
ysr@345 207 // are on the mark stack. If "yield_after" is "true", then the
ysr@345 208 // concurrent marker performing the drain offers to yield after
ysr@345 209 // processing each object. If a yield occurs, stops the drain operation
ysr@345 210 // and returns false. Otherwise, returns true.
ysr@345 211 template<class OopClosureClass>
ysr@345 212 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
ysr@345 213
ysr@345 214 bool isEmpty() { return _index == 0; }
ysr@345 215 bool isFull() { return _index == _capacity; }
ysr@345 216 int maxElems() { return _capacity; }
ysr@345 217
ysr@345 218 bool overflow() { return _overflow; }
ysr@345 219 void clear_overflow() { _overflow = false; }
ysr@345 220
ysr@345 221 int size() { return _index; }
ysr@345 222
ysr@345 223 void setEmpty() { _index = 0; clear_overflow(); }
ysr@345 224
ysr@345 225 // Record the current size; a subsequent "oops_do" will iterate only over
ysr@345 226 // indices valid at the time of this call.
ysr@345 227 void set_oops_do_bound(jint bound = -1) {
ysr@345 228 if (bound == -1) {
ysr@345 229 _oops_do_bound = _index;
ysr@345 230 } else {
ysr@345 231 _oops_do_bound = bound;
ysr@345 232 }
ysr@345 233 }
ysr@345 234 jint oops_do_bound() { return _oops_do_bound; }
ysr@345 235 // iterate over the oops in the mark stack, up to the bound recorded via
ysr@345 236 // the call above.
ysr@345 237 void oops_do(OopClosure* f);
ysr@345 238 };
ysr@345 239
apetrusenko@581 240 class CMRegionStack VALUE_OBJ_CLASS_SPEC {
ysr@345 241 MemRegion* _base;
ysr@345 242 jint _capacity;
ysr@345 243 jint _index;
ysr@345 244 jint _oops_do_bound;
ysr@345 245 bool _overflow;
ysr@345 246 public:
ysr@345 247 CMRegionStack();
ysr@345 248 ~CMRegionStack();
ysr@345 249 void allocate(size_t size);
ysr@345 250
ysr@345 251 // This is lock-free; assumes that it will only be called in parallel
ysr@345 252 // with other "push" operations (no pops).
johnc@1857 253 void push_lock_free(MemRegion mr);
tonyp@1447 254
ysr@345 255 // Lock-free; assumes that it will only be called in parallel
ysr@345 256 // with other "pop" operations (no pushes).
johnc@1857 257 MemRegion pop_lock_free();
johnc@1857 258
johnc@1857 259 #if 0
johnc@1857 260 // The routines that manipulate the region stack with a lock are
johnc@1857 261 // not currently used. They should be retained, however, as a
johnc@1857 262 // diagnostic aid.
tonyp@1447 263
tonyp@1447 264 // These two are the implementations that use a lock. They can be
tonyp@1447 265 // called concurrently with each other but they should not be called
tonyp@1447 266 // concurrently with the lock-free versions (push() / pop()).
tonyp@1447 267 void push_with_lock(MemRegion mr);
tonyp@1447 268 MemRegion pop_with_lock();
johnc@1857 269 #endif
ysr@345 270
ysr@345 271 bool isEmpty() { return _index == 0; }
ysr@345 272 bool isFull() { return _index == _capacity; }
ysr@345 273
ysr@345 274 bool overflow() { return _overflow; }
ysr@345 275 void clear_overflow() { _overflow = false; }
ysr@345 276
ysr@345 277 int size() { return _index; }
ysr@345 278
ysr@345 279 // It iterates over the entries in the region stack and it
ysr@345 280 // invalidates (i.e. assigns MemRegion()) the ones that point to
ysr@345 281 // regions in the collection set.
ysr@345 282 bool invalidate_entries_into_cset();
ysr@345 283
ysr@345 284 // This gives an upper bound up to which the iteration in
ysr@345 285 // invalidate_entries_into_cset() will reach. This prevents
ysr@345 286 // newly-added entries to be unnecessarily scanned.
ysr@345 287 void set_oops_do_bound() {
ysr@345 288 _oops_do_bound = _index;
ysr@345 289 }
ysr@345 290
ysr@345 291 void setEmpty() { _index = 0; clear_overflow(); }
ysr@345 292 };
ysr@345 293
ysr@345 294 // this will enable a variety of different statistics per GC task
ysr@345 295 #define _MARKING_STATS_ 0
ysr@345 296 // this will enable the higher verbose levels
ysr@345 297 #define _MARKING_VERBOSE_ 0
ysr@345 298
ysr@345 299 #if _MARKING_STATS_
ysr@345 300 #define statsOnly(statement) \
ysr@345 301 do { \
ysr@345 302 statement ; \
ysr@345 303 } while (0)
ysr@345 304 #else // _MARKING_STATS_
ysr@345 305 #define statsOnly(statement) \
ysr@345 306 do { \
ysr@345 307 } while (0)
ysr@345 308 #endif // _MARKING_STATS_
ysr@345 309
ysr@345 310 typedef enum {
ysr@345 311 no_verbose = 0, // verbose turned off
ysr@345 312 stats_verbose, // only prints stats at the end of marking
ysr@345 313 low_verbose, // low verbose, mostly per region and per major event
ysr@345 314 medium_verbose, // a bit more detailed than low
ysr@345 315 high_verbose // per object verbose
ysr@345 316 } CMVerboseLevel;
ysr@345 317
ysr@345 318
ysr@345 319 class ConcurrentMarkThread;
ysr@345 320
apetrusenko@581 321 class ConcurrentMark: public CHeapObj {
ysr@345 322 friend class ConcurrentMarkThread;
ysr@345 323 friend class CMTask;
ysr@345 324 friend class CMBitMapClosure;
ysr@345 325 friend class CSMarkOopClosure;
ysr@345 326 friend class CMGlobalObjectClosure;
ysr@345 327 friend class CMRemarkTask;
ysr@345 328 friend class CMConcurrentMarkingTask;
ysr@345 329 friend class G1ParNoteEndTask;
ysr@345 330 friend class CalcLiveObjectsClosure;
ysr@345 331
ysr@345 332 protected:
ysr@345 333 ConcurrentMarkThread* _cmThread; // the thread doing the work
ysr@345 334 G1CollectedHeap* _g1h; // the heap.
ysr@345 335 size_t _parallel_marking_threads; // the number of marking
ysr@345 336 // threads we'll use
ysr@345 337 double _sleep_factor; // how much we have to sleep, with
ysr@345 338 // respect to the work we just did, to
ysr@345 339 // meet the marking overhead goal
ysr@345 340 double _marking_task_overhead; // marking target overhead for
ysr@345 341 // a single task
ysr@345 342
ysr@345 343 // same as the two above, but for the cleanup task
ysr@345 344 double _cleanup_sleep_factor;
ysr@345 345 double _cleanup_task_overhead;
ysr@345 346
ysr@345 347 // Stuff related to age cohort processing.
ysr@345 348 struct ParCleanupThreadState {
ysr@345 349 char _pre[64];
ysr@345 350 UncleanRegionList list;
ysr@345 351 char _post[64];
ysr@345 352 };
ysr@345 353 ParCleanupThreadState** _par_cleanup_thread_state;
ysr@345 354
ysr@345 355 // CMS marking support structures
ysr@345 356 CMBitMap _markBitMap1;
ysr@345 357 CMBitMap _markBitMap2;
ysr@345 358 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
ysr@345 359 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
ysr@345 360 bool _at_least_one_mark_complete;
ysr@345 361
ysr@345 362 BitMap _region_bm;
ysr@345 363 BitMap _card_bm;
ysr@345 364
ysr@345 365 // Heap bounds
ysr@345 366 HeapWord* _heap_start;
ysr@345 367 HeapWord* _heap_end;
ysr@345 368
ysr@345 369 // For gray objects
ysr@345 370 CMMarkStack _markStack; // Grey objects behind global finger.
ysr@345 371 CMRegionStack _regionStack; // Grey regions behind global finger.
ysr@345 372 HeapWord* volatile _finger; // the global finger, region aligned,
ysr@345 373 // always points to the end of the
ysr@345 374 // last claimed region
ysr@345 375
ysr@345 376 // marking tasks
ysr@345 377 size_t _max_task_num; // maximum task number
ysr@345 378 size_t _active_tasks; // task num currently active
ysr@345 379 CMTask** _tasks; // task queue array (max_task_num len)
ysr@345 380 CMTaskQueueSet* _task_queues; // task queue set
ysr@345 381 ParallelTaskTerminator _terminator; // for termination
ysr@345 382
ysr@345 383 // Two sync barriers that are used to synchronise tasks when an
ysr@345 384 // overflow occurs. The algorithm is the following. All tasks enter
ysr@345 385 // the first one to ensure that they have all stopped manipulating
ysr@345 386 // the global data structures. After they exit it, they re-initialise
ysr@345 387 // their data structures and task 0 re-initialises the global data
ysr@345 388 // structures. Then, they enter the second sync barrier. This
ysr@345 389 // ensure, that no task starts doing work before all data
ysr@345 390 // structures (local and global) have been re-initialised. When they
ysr@345 391 // exit it, they are free to start working again.
ysr@345 392 WorkGangBarrierSync _first_overflow_barrier_sync;
ysr@345 393 WorkGangBarrierSync _second_overflow_barrier_sync;
ysr@345 394
ysr@345 395
ysr@345 396 // this is set by any task, when an overflow on the global data
ysr@345 397 // structures is detected.
ysr@345 398 volatile bool _has_overflown;
ysr@345 399 // true: marking is concurrent, false: we're in remark
ysr@345 400 volatile bool _concurrent;
ysr@345 401 // set at the end of a Full GC so that marking aborts
ysr@345 402 volatile bool _has_aborted;
johnc@1857 403
ysr@345 404 // used when remark aborts due to an overflow to indicate that
ysr@345 405 // another concurrent marking phase should start
ysr@345 406 volatile bool _restart_for_overflow;
ysr@345 407
ysr@345 408 // This is true from the very start of concurrent marking until the
ysr@345 409 // point when all the tasks complete their work. It is really used
ysr@345 410 // to determine the points between the end of concurrent marking and
ysr@345 411 // time of remark.
ysr@345 412 volatile bool _concurrent_marking_in_progress;
ysr@345 413
ysr@345 414 // verbose level
ysr@345 415 CMVerboseLevel _verbose_level;
ysr@345 416
ysr@345 417 // These two fields are used to implement the optimisation that
ysr@345 418 // avoids pushing objects on the global/region stack if there are
ysr@345 419 // no collection set regions above the lowest finger.
ysr@345 420
ysr@345 421 // This is the lowest finger (among the global and local fingers),
ysr@345 422 // which is calculated before a new collection set is chosen.
ysr@345 423 HeapWord* _min_finger;
ysr@345 424 // If this flag is true, objects/regions that are marked below the
ysr@345 425 // finger should be pushed on the stack(s). If this is flag is
ysr@345 426 // false, it is safe not to push them on the stack(s).
ysr@345 427 bool _should_gray_objects;
ysr@345 428
ysr@345 429 // All of these times are in ms.
ysr@345 430 NumberSeq _init_times;
ysr@345 431 NumberSeq _remark_times;
ysr@345 432 NumberSeq _remark_mark_times;
ysr@345 433 NumberSeq _remark_weak_ref_times;
ysr@345 434 NumberSeq _cleanup_times;
ysr@345 435 double _total_counting_time;
ysr@345 436 double _total_rs_scrub_time;
ysr@345 437
ysr@345 438 double* _accum_task_vtime; // accumulated task vtime
ysr@345 439
ysr@345 440 WorkGang* _parallel_workers;
ysr@345 441
ysr@345 442 void weakRefsWork(bool clear_all_soft_refs);
ysr@345 443
ysr@345 444 void swapMarkBitMaps();
ysr@345 445
ysr@345 446 // It resets the global marking data structures, as well as the
ysr@345 447 // task local ones; should be called during initial mark.
ysr@345 448 void reset();
ysr@345 449 // It resets all the marking data structures.
ysr@345 450 void clear_marking_state();
ysr@345 451
ysr@345 452 // It should be called to indicate which phase we're in (concurrent
ysr@345 453 // mark or remark) and how many threads are currently active.
ysr@345 454 void set_phase(size_t active_tasks, bool concurrent);
ysr@345 455 // We do this after we're done with marking so that the marking data
ysr@345 456 // structures are initialised to a sensible and predictable state.
ysr@345 457 void set_non_marking_state();
ysr@345 458
ysr@345 459 // prints all gathered CM-related statistics
ysr@345 460 void print_stats();
ysr@345 461
ysr@345 462 // accessor methods
ysr@345 463 size_t parallel_marking_threads() { return _parallel_marking_threads; }
ysr@345 464 double sleep_factor() { return _sleep_factor; }
ysr@345 465 double marking_task_overhead() { return _marking_task_overhead;}
ysr@345 466 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
ysr@345 467 double cleanup_task_overhead() { return _cleanup_task_overhead;}
ysr@345 468
ysr@345 469 HeapWord* finger() { return _finger; }
ysr@345 470 bool concurrent() { return _concurrent; }
ysr@345 471 size_t active_tasks() { return _active_tasks; }
ysr@345 472 ParallelTaskTerminator* terminator() { return &_terminator; }
ysr@345 473
ysr@345 474 // It claims the next available region to be scanned by a marking
ysr@345 475 // task. It might return NULL if the next region is empty or we have
ysr@345 476 // run out of regions. In the latter case, out_of_regions()
ysr@345 477 // determines whether we've really run out of regions or the task
ysr@345 478 // should call claim_region() again. This might seem a bit
ysr@345 479 // awkward. Originally, the code was written so that claim_region()
ysr@345 480 // either successfully returned with a non-empty region or there
ysr@345 481 // were no more regions to be claimed. The problem with this was
ysr@345 482 // that, in certain circumstances, it iterated over large chunks of
ysr@345 483 // the heap finding only empty regions and, while it was working, it
ysr@345 484 // was preventing the calling task to call its regular clock
ysr@345 485 // method. So, this way, each task will spend very little time in
ysr@345 486 // claim_region() and is allowed to call the regular clock method
ysr@345 487 // frequently.
ysr@345 488 HeapRegion* claim_region(int task);
ysr@345 489
ysr@345 490 // It determines whether we've run out of regions to scan.
ysr@345 491 bool out_of_regions() { return _finger == _heap_end; }
ysr@345 492
ysr@345 493 // Returns the task with the given id
ysr@345 494 CMTask* task(int id) {
tonyp@1082 495 assert(0 <= id && id < (int) _active_tasks,
tonyp@1082 496 "task id not within active bounds");
ysr@345 497 return _tasks[id];
ysr@345 498 }
ysr@345 499
ysr@345 500 // Returns the task queue with the given id
ysr@345 501 CMTaskQueue* task_queue(int id) {
tonyp@1082 502 assert(0 <= id && id < (int) _active_tasks,
tonyp@1082 503 "task queue id not within active bounds");
ysr@345 504 return (CMTaskQueue*) _task_queues->queue(id);
ysr@345 505 }
ysr@345 506
ysr@345 507 // Returns the task queue set
ysr@345 508 CMTaskQueueSet* task_queues() { return _task_queues; }
ysr@345 509
ysr@345 510 // Access / manipulation of the overflow flag which is set to
ysr@345 511 // indicate that the global stack or region stack has overflown
ysr@345 512 bool has_overflown() { return _has_overflown; }
ysr@345 513 void set_has_overflown() { _has_overflown = true; }
ysr@345 514 void clear_has_overflown() { _has_overflown = false; }
ysr@345 515
ysr@345 516 bool has_aborted() { return _has_aborted; }
ysr@345 517 bool restart_for_overflow() { return _restart_for_overflow; }
ysr@345 518
ysr@345 519 // Methods to enter the two overflow sync barriers
ysr@345 520 void enter_first_sync_barrier(int task_num);
ysr@345 521 void enter_second_sync_barrier(int task_num);
ysr@345 522
ysr@345 523 public:
ysr@345 524 // Manipulation of the global mark stack.
ysr@345 525 // Notice that the first mark_stack_push is CAS-based, whereas the
ysr@345 526 // two below are Mutex-based. This is OK since the first one is only
ysr@345 527 // called during evacuation pauses and doesn't compete with the
ysr@345 528 // other two (which are called by the marking tasks during
ysr@345 529 // concurrent marking or remark).
ysr@345 530 bool mark_stack_push(oop p) {
ysr@345 531 _markStack.par_push(p);
ysr@345 532 if (_markStack.overflow()) {
ysr@345 533 set_has_overflown();
ysr@345 534 return false;
ysr@345 535 }
ysr@345 536 return true;
ysr@345 537 }
ysr@345 538 bool mark_stack_push(oop* arr, int n) {
ysr@345 539 _markStack.par_push_arr(arr, n);
ysr@345 540 if (_markStack.overflow()) {
ysr@345 541 set_has_overflown();
ysr@345 542 return false;
ysr@345 543 }
ysr@345 544 return true;
ysr@345 545 }
ysr@345 546 void mark_stack_pop(oop* arr, int max, int* n) {
ysr@345 547 _markStack.par_pop_arr(arr, max, n);
ysr@345 548 }
ysr@345 549 size_t mark_stack_size() { return _markStack.size(); }
ysr@345 550 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
ysr@345 551 bool mark_stack_overflow() { return _markStack.overflow(); }
ysr@345 552 bool mark_stack_empty() { return _markStack.isEmpty(); }
ysr@345 553
johnc@1857 554 // (Lock-free) Manipulation of the region stack
johnc@1857 555 bool region_stack_push_lock_free(MemRegion mr) {
tonyp@1447 556 // Currently we only call the lock-free version during evacuation
tonyp@1447 557 // pauses.
tonyp@1447 558 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
tonyp@1447 559
johnc@1857 560 _regionStack.push_lock_free(mr);
ysr@345 561 if (_regionStack.overflow()) {
ysr@345 562 set_has_overflown();
ysr@345 563 return false;
ysr@345 564 }
ysr@345 565 return true;
ysr@345 566 }
johnc@1857 567
johnc@1857 568 // Lock-free version of region-stack pop. Should only be
johnc@1857 569 // called in tandem with other lock-free pops.
johnc@1857 570 MemRegion region_stack_pop_lock_free() {
johnc@1857 571 return _regionStack.pop_lock_free();
johnc@1857 572 }
johnc@1857 573
tonyp@1447 574 #if 0
johnc@1857 575 // The routines that manipulate the region stack with a lock are
johnc@1857 576 // not currently used. They should be retained, however, as a
johnc@1857 577 // diagnostic aid.
tonyp@1447 578
tonyp@1447 579 bool region_stack_push_with_lock(MemRegion mr) {
tonyp@1447 580 // Currently we only call the lock-based version during either
tonyp@1447 581 // concurrent marking or remark.
tonyp@1447 582 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
tonyp@1447 583 "if we are at a safepoint it should be the remark safepoint");
tonyp@1447 584
tonyp@1447 585 _regionStack.push_with_lock(mr);
tonyp@1447 586 if (_regionStack.overflow()) {
tonyp@1447 587 set_has_overflown();
tonyp@1447 588 return false;
tonyp@1447 589 }
tonyp@1447 590 return true;
tonyp@1447 591 }
johnc@1857 592
tonyp@1447 593 MemRegion region_stack_pop_with_lock() {
tonyp@1447 594 // Currently we only call the lock-based version during either
tonyp@1447 595 // concurrent marking or remark.
tonyp@1447 596 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
tonyp@1447 597 "if we are at a safepoint it should be the remark safepoint");
tonyp@1447 598
tonyp@1447 599 return _regionStack.pop_with_lock();
tonyp@1447 600 }
johnc@1857 601 #endif
tonyp@1447 602
ysr@345 603 int region_stack_size() { return _regionStack.size(); }
ysr@345 604 bool region_stack_overflow() { return _regionStack.overflow(); }
ysr@345 605 bool region_stack_empty() { return _regionStack.isEmpty(); }
ysr@345 606
johnc@1857 607 // Iterate over any regions that were aborted while draining the
johnc@1857 608 // region stack (any such regions are saved in the corresponding
johnc@1857 609 // CMTask) and invalidate (i.e. assign to the empty MemRegion())
johnc@1857 610 // any regions that point into the collection set.
johnc@1857 611 bool invalidate_aborted_regions_in_cset();
johnc@1857 612
johnc@1857 613 // Returns true if there are any aborted memory regions.
johnc@1857 614 bool has_aborted_regions();
johnc@1857 615
ysr@345 616 bool concurrent_marking_in_progress() {
ysr@345 617 return _concurrent_marking_in_progress;
ysr@345 618 }
ysr@345 619 void set_concurrent_marking_in_progress() {
ysr@345 620 _concurrent_marking_in_progress = true;
ysr@345 621 }
ysr@345 622 void clear_concurrent_marking_in_progress() {
ysr@345 623 _concurrent_marking_in_progress = false;
ysr@345 624 }
ysr@345 625
ysr@345 626 void update_accum_task_vtime(int i, double vtime) {
ysr@345 627 _accum_task_vtime[i] += vtime;
ysr@345 628 }
ysr@345 629
ysr@345 630 double all_task_accum_vtime() {
ysr@345 631 double ret = 0.0;
ysr@345 632 for (int i = 0; i < (int)_max_task_num; ++i)
ysr@345 633 ret += _accum_task_vtime[i];
ysr@345 634 return ret;
ysr@345 635 }
ysr@345 636
ysr@345 637 // Attempts to steal an object from the task queues of other tasks
ysr@345 638 bool try_stealing(int task_num, int* hash_seed, oop& obj) {
ysr@345 639 return _task_queues->steal(task_num, hash_seed, obj);
ysr@345 640 }
ysr@345 641
ysr@345 642 // It grays an object by first marking it. Then, if it's behind the
ysr@345 643 // global finger, it also pushes it on the global stack.
ysr@345 644 void deal_with_reference(oop obj);
ysr@345 645
ysr@345 646 ConcurrentMark(ReservedSpace rs, int max_regions);
ysr@345 647 ~ConcurrentMark();
ysr@345 648 ConcurrentMarkThread* cmThread() { return _cmThread; }
ysr@345 649
ysr@345 650 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
ysr@345 651 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
ysr@345 652
ysr@345 653 // The following three are interaction between CM and
ysr@345 654 // G1CollectedHeap
ysr@345 655
ysr@345 656 // This notifies CM that a root during initial-mark needs to be
ysr@345 657 // grayed and it's MT-safe. Currently, we just mark it. But, in the
ysr@345 658 // future, we can experiment with pushing it on the stack and we can
ysr@345 659 // do this without changing G1CollectedHeap.
ysr@345 660 void grayRoot(oop p);
ysr@345 661 // It's used during evacuation pauses to gray a region, if
ysr@345 662 // necessary, and it's MT-safe. It assumes that the caller has
ysr@345 663 // marked any objects on that region. If _should_gray_objects is
ysr@345 664 // true and we're still doing concurrent marking, the region is
ysr@345 665 // pushed on the region stack, if it is located below the global
ysr@345 666 // finger, otherwise we do nothing.
ysr@345 667 void grayRegionIfNecessary(MemRegion mr);
ysr@345 668 // It's used during evacuation pauses to mark and, if necessary,
ysr@345 669 // gray a single object and it's MT-safe. It assumes the caller did
ysr@345 670 // not mark the object. If _should_gray_objects is true and we're
ysr@345 671 // still doing concurrent marking, the objects is pushed on the
ysr@345 672 // global stack, if it is located below the global finger, otherwise
ysr@345 673 // we do nothing.
ysr@345 674 void markAndGrayObjectIfNecessary(oop p);
ysr@345 675
tonyp@1477 676 // It iterates over the heap and for each object it comes across it
tonyp@1477 677 // will dump the contents of its reference fields, as well as
tonyp@1477 678 // liveness information for the object and its referents. The dump
tonyp@1477 679 // will be written to a file with the following name:
tonyp@1477 680 // G1PrintReachableBaseFile + "." + str. use_prev_marking decides
tonyp@1477 681 // whether the prev (use_prev_marking == true) or next
tonyp@1477 682 // (use_prev_marking == false) marking information will be used to
tonyp@1477 683 // determine the liveness of each object / referent. If all is true,
tonyp@1477 684 // all objects in the heap will be dumped, otherwise only the live
tonyp@1477 685 // ones. In the dump the following symbols / abbreviations are used:
tonyp@1477 686 // M : an explicitly live object (its bitmap bit is set)
tonyp@1477 687 // > : an implicitly live object (over tams)
tonyp@1477 688 // O : an object outside the G1 heap (typically: in the perm gen)
tonyp@1477 689 // NOT : a reference field whose referent is not live
tonyp@1477 690 // AND MARKED : indicates that an object is both explicitly and
tonyp@1477 691 // implicitly live (it should be one or the other, not both)
tonyp@1477 692 void print_reachable(const char* str,
tonyp@1477 693 bool use_prev_marking, bool all) PRODUCT_RETURN;
ysr@345 694
ysr@345 695 // Clear the next marking bitmap (will be called concurrently).
ysr@345 696 void clearNextBitmap();
ysr@345 697
ysr@345 698 // main CMS steps and related support
ysr@345 699 void checkpointRootsInitial();
ysr@345 700
ysr@345 701 // These two do the work that needs to be done before and after the
ysr@345 702 // initial root checkpoint. Since this checkpoint can be done at two
ysr@345 703 // different points (i.e. an explicit pause or piggy-backed on a
ysr@345 704 // young collection), then it's nice to be able to easily share the
ysr@345 705 // pre/post code. It might be the case that we can put everything in
ysr@345 706 // the post method. TP
ysr@345 707 void checkpointRootsInitialPre();
ysr@345 708 void checkpointRootsInitialPost();
ysr@345 709
ysr@345 710 // Do concurrent phase of marking, to a tentative transitive closure.
ysr@345 711 void markFromRoots();
ysr@345 712
ysr@345 713 // Process all unprocessed SATB buffers. It is called at the
ysr@345 714 // beginning of an evacuation pause.
ysr@345 715 void drainAllSATBBuffers();
ysr@345 716
ysr@345 717 void checkpointRootsFinal(bool clear_all_soft_refs);
ysr@345 718 void checkpointRootsFinalWork();
ysr@345 719 void calcDesiredRegions();
ysr@345 720 void cleanup();
ysr@345 721 void completeCleanup();
ysr@345 722
ysr@345 723 // Mark in the previous bitmap. NB: this is usually read-only, so use
ysr@345 724 // this carefully!
ysr@345 725 void markPrev(oop p);
ysr@345 726 void clear(oop p);
ysr@345 727 // Clears marks for all objects in the given range, for both prev and
ysr@345 728 // next bitmaps. NB: the previous bitmap is usually read-only, so use
ysr@345 729 // this carefully!
ysr@345 730 void clearRangeBothMaps(MemRegion mr);
ysr@345 731
ysr@345 732 // Record the current top of the mark and region stacks; a
ysr@345 733 // subsequent oops_do() on the mark stack and
ysr@345 734 // invalidate_entries_into_cset() on the region stack will iterate
ysr@345 735 // only over indices valid at the time of this call.
ysr@345 736 void set_oops_do_bound() {
ysr@345 737 _markStack.set_oops_do_bound();
ysr@345 738 _regionStack.set_oops_do_bound();
ysr@345 739 }
ysr@345 740 // Iterate over the oops in the mark stack and all local queues. It
ysr@345 741 // also calls invalidate_entries_into_cset() on the region stack.
ysr@345 742 void oops_do(OopClosure* f);
ysr@345 743 // It is called at the end of an evacuation pause during marking so
ysr@345 744 // that CM is notified of where the new end of the heap is. It
ysr@345 745 // doesn't do anything if concurrent_marking_in_progress() is false,
ysr@345 746 // unless the force parameter is true.
ysr@345 747 void update_g1_committed(bool force = false);
ysr@345 748
ysr@345 749 void complete_marking_in_collection_set();
ysr@345 750
ysr@345 751 // It indicates that a new collection set is being chosen.
ysr@345 752 void newCSet();
ysr@345 753 // It registers a collection set heap region with CM. This is used
ysr@345 754 // to determine whether any heap regions are located above the finger.
ysr@345 755 void registerCSetRegion(HeapRegion* hr);
ysr@345 756
johnc@1483 757 // Registers the maximum region-end associated with a set of
johnc@1483 758 // regions with CM. Again this is used to determine whether any
johnc@1483 759 // heap regions are located above the finger.
johnc@1483 760 void register_collection_set_finger(HeapWord* max_finger) {
johnc@1483 761 // max_finger is the highest heap region end of the regions currently
johnc@1483 762 // contained in the collection set. If this value is larger than
johnc@1483 763 // _min_finger then we need to gray objects.
johnc@1483 764 // This routine is like registerCSetRegion but for an entire
johnc@1483 765 // collection of regions.
johnc@1483 766 if (max_finger > _min_finger)
johnc@1483 767 _should_gray_objects = true;
johnc@1483 768 }
johnc@1483 769
ysr@345 770 // Returns "true" if at least one mark has been completed.
ysr@345 771 bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
ysr@345 772
ysr@345 773 bool isMarked(oop p) const {
ysr@345 774 assert(p != NULL && p->is_oop(), "expected an oop");
ysr@345 775 HeapWord* addr = (HeapWord*)p;
ysr@345 776 assert(addr >= _nextMarkBitMap->startWord() ||
ysr@345 777 addr < _nextMarkBitMap->endWord(), "in a region");
ysr@345 778
ysr@345 779 return _nextMarkBitMap->isMarked(addr);
ysr@345 780 }
ysr@345 781
ysr@345 782 inline bool not_yet_marked(oop p) const;
ysr@345 783
ysr@345 784 // XXX Debug code
ysr@345 785 bool containing_card_is_marked(void* p);
ysr@345 786 bool containing_cards_are_marked(void* start, void* last);
ysr@345 787
ysr@345 788 bool isPrevMarked(oop p) const {
ysr@345 789 assert(p != NULL && p->is_oop(), "expected an oop");
ysr@345 790 HeapWord* addr = (HeapWord*)p;
ysr@345 791 assert(addr >= _prevMarkBitMap->startWord() ||
ysr@345 792 addr < _prevMarkBitMap->endWord(), "in a region");
ysr@345 793
ysr@345 794 return _prevMarkBitMap->isMarked(addr);
ysr@345 795 }
ysr@345 796
ysr@345 797 inline bool do_yield_check(int worker_i = 0);
ysr@345 798 inline bool should_yield();
ysr@345 799
ysr@345 800 // Called to abort the marking cycle after a Full GC takes palce.
ysr@345 801 void abort();
ysr@345 802
ysr@345 803 // This prints the global/local fingers. It is used for debugging.
ysr@345 804 NOT_PRODUCT(void print_finger();)
ysr@345 805
ysr@345 806 void print_summary_info();
ysr@345 807
tonyp@1078 808 void print_worker_threads_on(outputStream* st) const;
tonyp@1078 809
ysr@345 810 // The following indicate whether a given verbose level has been
ysr@345 811 // set. Notice that anything above stats is conditional to
ysr@345 812 // _MARKING_VERBOSE_ having been set to 1
ysr@345 813 bool verbose_stats()
ysr@345 814 { return _verbose_level >= stats_verbose; }
ysr@345 815 bool verbose_low()
ysr@345 816 { return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; }
ysr@345 817 bool verbose_medium()
ysr@345 818 { return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; }
ysr@345 819 bool verbose_high()
ysr@345 820 { return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; }
ysr@345 821 };
ysr@345 822
ysr@345 823 // A class representing a marking task.
ysr@345 824 class CMTask : public TerminatorTerminator {
ysr@345 825 private:
ysr@345 826 enum PrivateConstants {
ysr@345 827 // the regular clock call is called once the scanned words reaches
ysr@345 828 // this limit
ysr@345 829 words_scanned_period = 12*1024,
ysr@345 830 // the regular clock call is called once the number of visited
ysr@345 831 // references reaches this limit
ysr@345 832 refs_reached_period = 384,
ysr@345 833 // initial value for the hash seed, used in the work stealing code
ysr@345 834 init_hash_seed = 17,
ysr@345 835 // how many entries will be transferred between global stack and
ysr@345 836 // local queues
ysr@345 837 global_stack_transfer_size = 16
ysr@345 838 };
ysr@345 839
ysr@345 840 int _task_id;
ysr@345 841 G1CollectedHeap* _g1h;
ysr@345 842 ConcurrentMark* _cm;
ysr@345 843 CMBitMap* _nextMarkBitMap;
ysr@345 844 // the task queue of this task
ysr@345 845 CMTaskQueue* _task_queue;
ysr@896 846 private:
ysr@345 847 // the task queue set---needed for stealing
ysr@345 848 CMTaskQueueSet* _task_queues;
ysr@345 849 // indicates whether the task has been claimed---this is only for
ysr@345 850 // debugging purposes
ysr@345 851 bool _claimed;
ysr@345 852
ysr@345 853 // number of calls to this task
ysr@345 854 int _calls;
ysr@345 855
ysr@345 856 // when the virtual timer reaches this time, the marking step should
ysr@345 857 // exit
ysr@345 858 double _time_target_ms;
ysr@345 859 // the start time of the current marking step
ysr@345 860 double _start_time_ms;
ysr@345 861
ysr@345 862 // the oop closure used for iterations over oops
ysr@345 863 OopClosure* _oop_closure;
ysr@345 864
ysr@345 865 // the region this task is scanning, NULL if we're not scanning any
ysr@345 866 HeapRegion* _curr_region;
ysr@345 867 // the local finger of this task, NULL if we're not scanning a region
ysr@345 868 HeapWord* _finger;
ysr@345 869 // limit of the region this task is scanning, NULL if we're not scanning one
ysr@345 870 HeapWord* _region_limit;
ysr@345 871
ysr@345 872 // This is used only when we scan regions popped from the region
ysr@345 873 // stack. It records what the last object on such a region we
ysr@345 874 // scanned was. It is used to ensure that, if we abort region
ysr@345 875 // iteration, we do not rescan the first part of the region. This
ysr@345 876 // should be NULL when we're not scanning a region from the region
ysr@345 877 // stack.
ysr@345 878 HeapWord* _region_finger;
ysr@345 879
johnc@1857 880 // If we abort while scanning a region we record the remaining
johnc@1857 881 // unscanned portion and check this field when marking restarts.
johnc@1857 882 // This avoids having to push on the region stack while other
johnc@1857 883 // marking threads may still be popping regions.
johnc@1857 884 // If we were to push the unscanned portion directly to the
johnc@1857 885 // region stack then we would need to using locking versions
johnc@1857 886 // of the push and pop operations.
johnc@1857 887 MemRegion _aborted_region;
johnc@1857 888
ysr@345 889 // the number of words this task has scanned
ysr@345 890 size_t _words_scanned;
ysr@345 891 // When _words_scanned reaches this limit, the regular clock is
ysr@345 892 // called. Notice that this might be decreased under certain
ysr@345 893 // circumstances (i.e. when we believe that we did an expensive
ysr@345 894 // operation).
ysr@345 895 size_t _words_scanned_limit;
ysr@345 896 // the initial value of _words_scanned_limit (i.e. what it was
ysr@345 897 // before it was decreased).
ysr@345 898 size_t _real_words_scanned_limit;
ysr@345 899
ysr@345 900 // the number of references this task has visited
ysr@345 901 size_t _refs_reached;
ysr@345 902 // When _refs_reached reaches this limit, the regular clock is
ysr@345 903 // called. Notice this this might be decreased under certain
ysr@345 904 // circumstances (i.e. when we believe that we did an expensive
ysr@345 905 // operation).
ysr@345 906 size_t _refs_reached_limit;
ysr@345 907 // the initial value of _refs_reached_limit (i.e. what it was before
ysr@345 908 // it was decreased).
ysr@345 909 size_t _real_refs_reached_limit;
ysr@345 910
ysr@345 911 // used by the work stealing stuff
ysr@345 912 int _hash_seed;
ysr@345 913 // if this is true, then the task has aborted for some reason
ysr@345 914 bool _has_aborted;
ysr@345 915 // set when the task aborts because it has met its time quota
ysr@345 916 bool _has_aborted_timed_out;
ysr@345 917 // true when we're draining SATB buffers; this avoids the task
ysr@345 918 // aborting due to SATB buffers being available (as we're already
ysr@345 919 // dealing with them)
ysr@345 920 bool _draining_satb_buffers;
ysr@345 921
ysr@345 922 // number sequence of past step times
ysr@345 923 NumberSeq _step_times_ms;
ysr@345 924 // elapsed time of this task
ysr@345 925 double _elapsed_time_ms;
ysr@345 926 // termination time of this task
ysr@345 927 double _termination_time_ms;
ysr@345 928 // when this task got into the termination protocol
ysr@345 929 double _termination_start_time_ms;
ysr@345 930
ysr@345 931 // true when the task is during a concurrent phase, false when it is
ysr@345 932 // in the remark phase (so, in the latter case, we do not have to
ysr@345 933 // check all the things that we have to check during the concurrent
ysr@345 934 // phase, i.e. SATB buffer availability...)
ysr@345 935 bool _concurrent;
ysr@345 936
ysr@345 937 TruncatedSeq _marking_step_diffs_ms;
ysr@345 938
ysr@345 939 // LOTS of statistics related with this task
ysr@345 940 #if _MARKING_STATS_
ysr@345 941 NumberSeq _all_clock_intervals_ms;
ysr@345 942 double _interval_start_time_ms;
ysr@345 943
ysr@345 944 int _aborted;
ysr@345 945 int _aborted_overflow;
ysr@345 946 int _aborted_cm_aborted;
ysr@345 947 int _aborted_yield;
ysr@345 948 int _aborted_timed_out;
ysr@345 949 int _aborted_satb;
ysr@345 950 int _aborted_termination;
ysr@345 951
ysr@345 952 int _steal_attempts;
ysr@345 953 int _steals;
ysr@345 954
ysr@345 955 int _clock_due_to_marking;
ysr@345 956 int _clock_due_to_scanning;
ysr@345 957
ysr@345 958 int _local_pushes;
ysr@345 959 int _local_pops;
ysr@345 960 int _local_max_size;
ysr@345 961 int _objs_scanned;
ysr@345 962
ysr@345 963 int _global_pushes;
ysr@345 964 int _global_pops;
ysr@345 965 int _global_max_size;
ysr@345 966
ysr@345 967 int _global_transfers_to;
ysr@345 968 int _global_transfers_from;
ysr@345 969
ysr@345 970 int _region_stack_pops;
ysr@345 971
ysr@345 972 int _regions_claimed;
ysr@345 973 int _objs_found_on_bitmap;
ysr@345 974
ysr@345 975 int _satb_buffers_processed;
ysr@345 976 #endif // _MARKING_STATS_
ysr@345 977
ysr@345 978 // it updates the local fields after this task has claimed
ysr@345 979 // a new region to scan
ysr@345 980 void setup_for_region(HeapRegion* hr);
ysr@345 981 // it brings up-to-date the limit of the region
ysr@345 982 void update_region_limit();
ysr@345 983 // it resets the local fields after a task has finished scanning a
ysr@345 984 // region
ysr@345 985 void giveup_current_region();
ysr@345 986
ysr@345 987 // called when either the words scanned or the refs visited limit
ysr@345 988 // has been reached
ysr@345 989 void reached_limit();
ysr@345 990 // recalculates the words scanned and refs visited limits
ysr@345 991 void recalculate_limits();
ysr@345 992 // decreases the words scanned and refs visited limits when we reach
ysr@345 993 // an expensive operation
ysr@345 994 void decrease_limits();
ysr@345 995 // it checks whether the words scanned or refs visited reached their
ysr@345 996 // respective limit and calls reached_limit() if they have
ysr@345 997 void check_limits() {
ysr@345 998 if (_words_scanned >= _words_scanned_limit ||
ysr@345 999 _refs_reached >= _refs_reached_limit)
ysr@345 1000 reached_limit();
ysr@345 1001 }
ysr@345 1002 // this is supposed to be called regularly during a marking step as
ysr@345 1003 // it checks a bunch of conditions that might cause the marking step
ysr@345 1004 // to abort
ysr@345 1005 void regular_clock_call();
ysr@345 1006 bool concurrent() { return _concurrent; }
ysr@345 1007
ysr@345 1008 public:
ysr@345 1009 // It resets the task; it should be called right at the beginning of
ysr@345 1010 // a marking phase.
ysr@345 1011 void reset(CMBitMap* _nextMarkBitMap);
ysr@345 1012 // it clears all the fields that correspond to a claimed region.
ysr@345 1013 void clear_region_fields();
ysr@345 1014
ysr@345 1015 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
ysr@345 1016
ysr@345 1017 // The main method of this class which performs a marking step
ysr@345 1018 // trying not to exceed the given duration. However, it might exit
ysr@345 1019 // prematurely, according to some conditions (i.e. SATB buffers are
ysr@345 1020 // available for processing).
ysr@345 1021 void do_marking_step(double target_ms);
ysr@345 1022
ysr@345 1023 // These two calls start and stop the timer
ysr@345 1024 void record_start_time() {
ysr@345 1025 _elapsed_time_ms = os::elapsedTime() * 1000.0;
ysr@345 1026 }
ysr@345 1027 void record_end_time() {
ysr@345 1028 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
ysr@345 1029 }
ysr@345 1030
ysr@345 1031 // returns the task ID
ysr@345 1032 int task_id() { return _task_id; }
ysr@345 1033
ysr@345 1034 // From TerminatorTerminator. It determines whether this task should
ysr@345 1035 // exit the termination protocol after it's entered it.
ysr@345 1036 virtual bool should_exit_termination();
ysr@345 1037
ysr@345 1038 HeapWord* finger() { return _finger; }
ysr@345 1039
ysr@345 1040 bool has_aborted() { return _has_aborted; }
ysr@345 1041 void set_has_aborted() { _has_aborted = true; }
ysr@345 1042 void clear_has_aborted() { _has_aborted = false; }
ysr@345 1043 bool claimed() { return _claimed; }
ysr@345 1044
johnc@1857 1045 // Support routines for the partially scanned region that may be
johnc@1857 1046 // recorded as a result of aborting while draining the CMRegionStack
johnc@1857 1047 MemRegion aborted_region() { return _aborted_region; }
johnc@1857 1048 void set_aborted_region(MemRegion mr)
johnc@1857 1049 { _aborted_region = mr; }
johnc@1857 1050
johnc@1857 1051 // Clears any recorded partially scanned region
johnc@1857 1052 void clear_aborted_region() { set_aborted_region(MemRegion()); }
johnc@1857 1053
ysr@345 1054 void set_oop_closure(OopClosure* oop_closure) {
ysr@345 1055 _oop_closure = oop_closure;
ysr@345 1056 }
ysr@345 1057
ysr@345 1058 // It grays the object by marking it and, if necessary, pushing it
ysr@345 1059 // on the local queue
ysr@345 1060 void deal_with_reference(oop obj);
ysr@345 1061
ysr@345 1062 // It scans an object and visits its children.
ysr@345 1063 void scan_object(oop obj) {
tonyp@1082 1064 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
ysr@345 1065
ysr@345 1066 if (_cm->verbose_high())
ysr@345 1067 gclog_or_tty->print_cr("[%d] we're scanning object "PTR_FORMAT,
ysr@345 1068 _task_id, (void*) obj);
ysr@345 1069
ysr@345 1070 size_t obj_size = obj->size();
ysr@345 1071 _words_scanned += obj_size;
ysr@345 1072
ysr@345 1073 obj->oop_iterate(_oop_closure);
ysr@345 1074 statsOnly( ++_objs_scanned );
ysr@345 1075 check_limits();
ysr@345 1076 }
ysr@345 1077
ysr@345 1078 // It pushes an object on the local queue.
ysr@345 1079 void push(oop obj);
ysr@345 1080
ysr@345 1081 // These two move entries to/from the global stack.
ysr@345 1082 void move_entries_to_global_stack();
ysr@345 1083 void get_entries_from_global_stack();
ysr@345 1084
ysr@345 1085 // It pops and scans objects from the local queue. If partially is
ysr@345 1086 // true, then it stops when the queue size is of a given limit. If
ysr@345 1087 // partially is false, then it stops when the queue is empty.
ysr@345 1088 void drain_local_queue(bool partially);
ysr@345 1089 // It moves entries from the global stack to the local queue and
ysr@345 1090 // drains the local queue. If partially is true, then it stops when
ysr@345 1091 // both the global stack and the local queue reach a given size. If
ysr@345 1092 // partially if false, it tries to empty them totally.
ysr@345 1093 void drain_global_stack(bool partially);
ysr@345 1094 // It keeps picking SATB buffers and processing them until no SATB
ysr@345 1095 // buffers are available.
ysr@345 1096 void drain_satb_buffers();
ysr@345 1097 // It keeps popping regions from the region stack and processing
ysr@345 1098 // them until the region stack is empty.
ysr@345 1099 void drain_region_stack(BitMapClosure* closure);
ysr@345 1100
ysr@345 1101 // moves the local finger to a new location
ysr@345 1102 inline void move_finger_to(HeapWord* new_finger) {
tonyp@1082 1103 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
ysr@345 1104 _finger = new_finger;
ysr@345 1105 }
ysr@345 1106
ysr@345 1107 // moves the region finger to a new location
ysr@345 1108 inline void move_region_finger_to(HeapWord* new_finger) {
tonyp@1082 1109 assert(new_finger < _cm->finger(), "invariant");
ysr@345 1110 _region_finger = new_finger;
ysr@345 1111 }
ysr@345 1112
ysr@345 1113 CMTask(int task_num, ConcurrentMark *cm,
ysr@345 1114 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
ysr@345 1115
ysr@345 1116 // it prints statistics associated with this task
ysr@345 1117 void print_stats();
ysr@345 1118
ysr@345 1119 #if _MARKING_STATS_
ysr@345 1120 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
ysr@345 1121 #endif // _MARKING_STATS_
ysr@345 1122 };