annotate src/share/vm/gc_implementation/g1/concurrentMark.hpp @ 2981:2ace1c4ee8da

6888336: G1: avoid explicitly marking and pushing objects in survivor spaces Summary: This change simplifies the interaction between GC and concurrent marking. By disabling survivor spaces during the initial-mark pause we don't need to propagate marks of objects we copy during each GC (since we never need to copy an explicitly marked object). Reviewed-by: johnc, brutisso
author tonyp
date Tue, 10 Jan 2012 18:58:13 -0500
parents 441e946dc1af
children 2e966d967c5c
rev   line source
ysr@342 1 /*
tonyp@2981 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
ysr@342 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
ysr@342 4 *
ysr@342 5 * This code is free software; you can redistribute it and/or modify it
ysr@342 6 * under the terms of the GNU General Public License version 2 only, as
ysr@342 7 * published by the Free Software Foundation.
ysr@342 8 *
ysr@342 9 * This code is distributed in the hope that it will be useful, but WITHOUT
ysr@342 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
ysr@342 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
ysr@342 12 * version 2 for more details (a copy is included in the LICENSE file that
ysr@342 13 * accompanied this code).
ysr@342 14 *
ysr@342 15 * You should have received a copy of the GNU General Public License version
ysr@342 16 * 2 along with this work; if not, write to the Free Software Foundation,
ysr@342 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
ysr@342 18 *
trims@1472 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1472 20 * or visit www.oracle.com if you need additional information or have any
trims@1472 21 * questions.
ysr@342 22 *
ysr@342 23 */
ysr@342 24
stefank@1879 25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
stefank@1879 26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
stefank@1879 27
tonyp@2037 28 #include "gc_implementation/g1/heapRegionSets.hpp"
stefank@1879 29 #include "utilities/taskqueue.hpp"
stefank@1879 30
ysr@342 31 class G1CollectedHeap;
ysr@342 32 class CMTask;
jcoomes@1311 33 typedef GenericTaskQueue<oop> CMTaskQueue;
jcoomes@1311 34 typedef GenericTaskQueueSet<CMTaskQueue> CMTaskQueueSet;
ysr@342 35
johnc@1944 36 // Closure used by CM during concurrent reference discovery
johnc@1944 37 // and reference processing (during remarking) to determine
johnc@1944 38 // if a particular object is alive. It is primarily used
johnc@1944 39 // to determine if referents of discovered reference objects
johnc@1944 40 // are alive. An instance is also embedded into the
johnc@1944 41 // reference processor as the _is_alive_non_header field
johnc@1944 42 class G1CMIsAliveClosure: public BoolObjectClosure {
johnc@1944 43 G1CollectedHeap* _g1;
johnc@1944 44 public:
johnc@1944 45 G1CMIsAliveClosure(G1CollectedHeap* g1) :
johnc@1944 46 _g1(g1)
johnc@1944 47 {}
johnc@1944 48
johnc@1944 49 void do_object(oop obj) {
johnc@1944 50 ShouldNotCallThis();
johnc@1944 51 }
johnc@1944 52 bool do_object_b(oop obj);
johnc@1944 53 };
johnc@1944 54
ysr@342 55 // A generic CM bit map. This is essentially a wrapper around the BitMap
ysr@342 56 // class, with one bit per (1<<_shifter) HeapWords.
ysr@342 57
apetrusenko@549 58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
ysr@342 59 protected:
ysr@342 60 HeapWord* _bmStartWord; // base address of range covered by map
ysr@342 61 size_t _bmWordSize; // map size (in #HeapWords covered)
ysr@342 62 const int _shifter; // map to char or bit
ysr@342 63 VirtualSpace _virtual_space; // underlying the bit map
ysr@342 64 BitMap _bm; // the bit map itself
ysr@342 65
ysr@342 66 public:
ysr@342 67 // constructor
ysr@342 68 CMBitMapRO(ReservedSpace rs, int shifter);
ysr@342 69
ysr@342 70 enum { do_yield = true };
ysr@342 71
ysr@342 72 // inquiries
ysr@342 73 HeapWord* startWord() const { return _bmStartWord; }
ysr@342 74 size_t sizeInWords() const { return _bmWordSize; }
ysr@342 75 // the following is one past the last word in space
ysr@342 76 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
ysr@342 77
ysr@342 78 // read marks
ysr@342 79
ysr@342 80 bool isMarked(HeapWord* addr) const {
ysr@342 81 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@342 82 "outside underlying space?");
ysr@342 83 return _bm.at(heapWordToOffset(addr));
ysr@342 84 }
ysr@342 85
ysr@342 86 // iteration
ysr@342 87 bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }
ysr@342 88 bool iterate(BitMapClosure* cl, MemRegion mr);
ysr@342 89
ysr@342 90 // Return the address corresponding to the next marked bit at or after
ysr@342 91 // "addr", and before "limit", if "limit" is non-NULL. If there is no
ysr@342 92 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
ysr@342 93 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
ysr@342 94 HeapWord* limit = NULL) const;
ysr@342 95 // Return the address corresponding to the next unmarked bit at or after
ysr@342 96 // "addr", and before "limit", if "limit" is non-NULL. If there is no
ysr@342 97 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
ysr@342 98 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
ysr@342 99 HeapWord* limit = NULL) const;
ysr@342 100
ysr@342 101 // conversion utilities
ysr@342 102 // XXX Fix these so that offsets are size_t's...
ysr@342 103 HeapWord* offsetToHeapWord(size_t offset) const {
ysr@342 104 return _bmStartWord + (offset << _shifter);
ysr@342 105 }
ysr@342 106 size_t heapWordToOffset(HeapWord* addr) const {
ysr@342 107 return pointer_delta(addr, _bmStartWord) >> _shifter;
ysr@342 108 }
ysr@342 109 int heapWordDiffToOffsetDiff(size_t diff) const;
ysr@342 110 HeapWord* nextWord(HeapWord* addr) {
ysr@342 111 return offsetToHeapWord(heapWordToOffset(addr) + 1);
ysr@342 112 }
ysr@342 113
ysr@342 114 void mostly_disjoint_range_union(BitMap* from_bitmap,
ysr@342 115 size_t from_start_index,
ysr@342 116 HeapWord* to_start_word,
ysr@342 117 size_t word_num);
ysr@342 118
ysr@342 119 // debugging
ysr@342 120 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
ysr@342 121 };
ysr@342 122
ysr@342 123 class CMBitMap : public CMBitMapRO {
ysr@342 124
ysr@342 125 public:
ysr@342 126 // constructor
ysr@342 127 CMBitMap(ReservedSpace rs, int shifter) :
ysr@342 128 CMBitMapRO(rs, shifter) {}
ysr@342 129
ysr@342 130 // write marks
ysr@342 131 void mark(HeapWord* addr) {
ysr@342 132 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@342 133 "outside underlying space?");
tonyp@2533 134 _bm.set_bit(heapWordToOffset(addr));
ysr@342 135 }
ysr@342 136 void clear(HeapWord* addr) {
ysr@342 137 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@342 138 "outside underlying space?");
tonyp@2533 139 _bm.clear_bit(heapWordToOffset(addr));
ysr@342 140 }
ysr@342 141 bool parMark(HeapWord* addr) {
ysr@342 142 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@342 143 "outside underlying space?");
tonyp@2533 144 return _bm.par_set_bit(heapWordToOffset(addr));
ysr@342 145 }
ysr@342 146 bool parClear(HeapWord* addr) {
ysr@342 147 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
ysr@342 148 "outside underlying space?");
tonyp@2533 149 return _bm.par_clear_bit(heapWordToOffset(addr));
ysr@342 150 }
ysr@342 151 void markRange(MemRegion mr);
ysr@342 152 void clearAll();
ysr@342 153 void clearRange(MemRegion mr);
ysr@342 154
ysr@342 155 // Starting at the bit corresponding to "addr" (inclusive), find the next
ysr@342 156 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
ysr@342 157 // the end of this run (stopping at "end_addr"). Return the MemRegion
ysr@342 158 // covering from the start of the region corresponding to the first bit
ysr@342 159 // of the run to the end of the region corresponding to the last bit of
ysr@342 160 // the run. If there is no "1" bit at or after "addr", return an empty
ysr@342 161 // MemRegion.
ysr@342 162 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
ysr@342 163 };
ysr@342 164
ysr@342 165 // Represents a marking stack used by the CM collector.
ysr@342 166 // Ideally this should be GrowableArray<> just like MSC's marking stack(s).
apetrusenko@549 167 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
ysr@342 168 ConcurrentMark* _cm;
tonyp@2981 169 oop* _base; // bottom of stack
tonyp@2981 170 jint _index; // one more than last occupied index
tonyp@2981 171 jint _capacity; // max #elements
tonyp@2981 172 jint _saved_index; // value of _index saved at start of GC
ysr@342 173 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
ysr@342 174
ysr@342 175 bool _overflow;
ysr@342 176 DEBUG_ONLY(bool _drain_in_progress;)
ysr@342 177 DEBUG_ONLY(bool _drain_in_progress_yields;)
ysr@342 178
ysr@342 179 public:
ysr@342 180 CMMarkStack(ConcurrentMark* cm);
ysr@342 181 ~CMMarkStack();
ysr@342 182
ysr@342 183 void allocate(size_t size);
ysr@342 184
ysr@342 185 oop pop() {
ysr@342 186 if (!isEmpty()) {
ysr@342 187 return _base[--_index] ;
ysr@342 188 }
ysr@342 189 return NULL;
ysr@342 190 }
ysr@342 191
ysr@342 192 // If overflow happens, don't do the push, and record the overflow.
ysr@342 193 // *Requires* that "ptr" is already marked.
ysr@342 194 void push(oop ptr) {
ysr@342 195 if (isFull()) {
ysr@342 196 // Record overflow.
ysr@342 197 _overflow = true;
ysr@342 198 return;
ysr@342 199 } else {
ysr@342 200 _base[_index++] = ptr;
ysr@342 201 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
ysr@342 202 }
ysr@342 203 }
ysr@342 204 // Non-block impl. Note: concurrency is allowed only with other
ysr@342 205 // "par_push" operations, not with "pop" or "drain". We would need
ysr@342 206 // parallel versions of them if such concurrency was desired.
ysr@342 207 void par_push(oop ptr);
ysr@342 208
ysr@342 209 // Pushes the first "n" elements of "ptr_arr" on the stack.
ysr@342 210 // Non-block impl. Note: concurrency is allowed only with other
ysr@342 211 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
ysr@342 212 void par_adjoin_arr(oop* ptr_arr, int n);
ysr@342 213
ysr@342 214 // Pushes the first "n" elements of "ptr_arr" on the stack.
ysr@342 215 // Locking impl: concurrency is allowed only with
ysr@342 216 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
ysr@342 217 // locking strategy.
ysr@342 218 void par_push_arr(oop* ptr_arr, int n);
ysr@342 219
ysr@342 220 // If returns false, the array was empty. Otherwise, removes up to "max"
ysr@342 221 // elements from the stack, and transfers them to "ptr_arr" in an
ysr@342 222 // unspecified order. The actual number transferred is given in "n" ("n
ysr@342 223 // == 0" is deliberately redundant with the return value.) Locking impl:
ysr@342 224 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
ysr@342 225 // operations, which use the same locking strategy.
ysr@342 226 bool par_pop_arr(oop* ptr_arr, int max, int* n);
ysr@342 227
ysr@342 228 // Drain the mark stack, applying the given closure to all fields of
ysr@342 229 // objects on the stack. (That is, continue until the stack is empty,
ysr@342 230 // even if closure applications add entries to the stack.) The "bm"
ysr@342 231 // argument, if non-null, may be used to verify that only marked objects
ysr@342 232 // are on the mark stack. If "yield_after" is "true", then the
ysr@342 233 // concurrent marker performing the drain offers to yield after
ysr@342 234 // processing each object. If a yield occurs, stops the drain operation
ysr@342 235 // and returns false. Otherwise, returns true.
ysr@342 236 template<class OopClosureClass>
ysr@342 237 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
ysr@342 238
ysr@342 239 bool isEmpty() { return _index == 0; }
ysr@342 240 bool isFull() { return _index == _capacity; }
ysr@342 241 int maxElems() { return _capacity; }
ysr@342 242
ysr@342 243 bool overflow() { return _overflow; }
ysr@342 244 void clear_overflow() { _overflow = false; }
ysr@342 245
ysr@342 246 int size() { return _index; }
ysr@342 247
ysr@342 248 void setEmpty() { _index = 0; clear_overflow(); }
ysr@342 249
tonyp@2981 250 // Record the current index.
tonyp@2981 251 void note_start_of_gc();
tonyp@2981 252
tonyp@2981 253 // Make sure that we have not added any entries to the stack during GC.
tonyp@2981 254 void note_end_of_gc();
tonyp@2981 255
ysr@342 256 // iterate over the oops in the mark stack, up to the bound recorded via
ysr@342 257 // the call above.
ysr@342 258 void oops_do(OopClosure* f);
ysr@342 259 };
ysr@342 260
apetrusenko@549 261 class CMRegionStack VALUE_OBJ_CLASS_SPEC {
ysr@342 262 MemRegion* _base;
ysr@342 263 jint _capacity;
ysr@342 264 jint _index;
ysr@342 265 jint _oops_do_bound;
ysr@342 266 bool _overflow;
ysr@342 267 public:
ysr@342 268 CMRegionStack();
ysr@342 269 ~CMRegionStack();
ysr@342 270 void allocate(size_t size);
ysr@342 271
ysr@342 272 // This is lock-free; assumes that it will only be called in parallel
ysr@342 273 // with other "push" operations (no pops).
johnc@1755 274 void push_lock_free(MemRegion mr);
tonyp@1358 275
ysr@342 276 // Lock-free; assumes that it will only be called in parallel
ysr@342 277 // with other "pop" operations (no pushes).
johnc@1755 278 MemRegion pop_lock_free();
johnc@1755 279
johnc@1755 280 #if 0
johnc@1755 281 // The routines that manipulate the region stack with a lock are
johnc@1755 282 // not currently used. They should be retained, however, as a
johnc@1755 283 // diagnostic aid.
tonyp@1358 284
tonyp@1358 285 // These two are the implementations that use a lock. They can be
tonyp@1358 286 // called concurrently with each other but they should not be called
tonyp@1358 287 // concurrently with the lock-free versions (push() / pop()).
tonyp@1358 288 void push_with_lock(MemRegion mr);
tonyp@1358 289 MemRegion pop_with_lock();
johnc@1755 290 #endif
ysr@342 291
ysr@342 292 bool isEmpty() { return _index == 0; }
ysr@342 293 bool isFull() { return _index == _capacity; }
ysr@342 294
ysr@342 295 bool overflow() { return _overflow; }
ysr@342 296 void clear_overflow() { _overflow = false; }
ysr@342 297
ysr@342 298 int size() { return _index; }
ysr@342 299
ysr@342 300 // It iterates over the entries in the region stack and it
ysr@342 301 // invalidates (i.e. assigns MemRegion()) the ones that point to
ysr@342 302 // regions in the collection set.
ysr@342 303 bool invalidate_entries_into_cset();
ysr@342 304
ysr@342 305 // This gives an upper bound up to which the iteration in
ysr@342 306 // invalidate_entries_into_cset() will reach. This prevents
ysr@342 307 // newly-added entries to be unnecessarily scanned.
ysr@342 308 void set_oops_do_bound() {
ysr@342 309 _oops_do_bound = _index;
ysr@342 310 }
ysr@342 311
ysr@342 312 void setEmpty() { _index = 0; clear_overflow(); }
ysr@342 313 };
ysr@342 314
tonyp@2413 315 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
tonyp@2413 316 private:
tonyp@2413 317 #ifndef PRODUCT
tonyp@2413 318 uintx _num_remaining;
tonyp@2413 319 bool _force;
tonyp@2413 320 #endif // !defined(PRODUCT)
tonyp@2413 321
tonyp@2413 322 public:
tonyp@2413 323 void init() PRODUCT_RETURN;
tonyp@2413 324 void update() PRODUCT_RETURN;
tonyp@2413 325 bool should_force() PRODUCT_RETURN_( return false; );
tonyp@2413 326 };
tonyp@2413 327
ysr@342 328 // this will enable a variety of different statistics per GC task
ysr@342 329 #define _MARKING_STATS_ 0
ysr@342 330 // this will enable the higher verbose levels
ysr@342 331 #define _MARKING_VERBOSE_ 0
ysr@342 332
ysr@342 333 #if _MARKING_STATS_
ysr@342 334 #define statsOnly(statement) \
ysr@342 335 do { \
ysr@342 336 statement ; \
ysr@342 337 } while (0)
ysr@342 338 #else // _MARKING_STATS_
ysr@342 339 #define statsOnly(statement) \
ysr@342 340 do { \
ysr@342 341 } while (0)
ysr@342 342 #endif // _MARKING_STATS_
ysr@342 343
ysr@342 344 typedef enum {
ysr@342 345 no_verbose = 0, // verbose turned off
ysr@342 346 stats_verbose, // only prints stats at the end of marking
ysr@342 347 low_verbose, // low verbose, mostly per region and per major event
ysr@342 348 medium_verbose, // a bit more detailed than low
ysr@342 349 high_verbose // per object verbose
ysr@342 350 } CMVerboseLevel;
ysr@342 351
ysr@342 352
ysr@342 353 class ConcurrentMarkThread;
ysr@342 354
apetrusenko@549 355 class ConcurrentMark: public CHeapObj {
ysr@342 356 friend class ConcurrentMarkThread;
ysr@342 357 friend class CMTask;
ysr@342 358 friend class CMBitMapClosure;
johnc@2861 359 friend class CSetMarkOopClosure;
ysr@342 360 friend class CMGlobalObjectClosure;
ysr@342 361 friend class CMRemarkTask;
ysr@342 362 friend class CMConcurrentMarkingTask;
ysr@342 363 friend class G1ParNoteEndTask;
ysr@342 364 friend class CalcLiveObjectsClosure;
johnc@2740 365 friend class G1CMRefProcTaskProxy;
johnc@2740 366 friend class G1CMRefProcTaskExecutor;
johnc@2059 367 friend class G1CMParKeepAliveAndDrainClosure;
johnc@2059 368 friend class G1CMParDrainMarkingStackClosure;
ysr@342 369
ysr@342 370 protected:
ysr@342 371 ConcurrentMarkThread* _cmThread; // the thread doing the work
ysr@342 372 G1CollectedHeap* _g1h; // the heap.
jmasa@2922 373 uint _parallel_marking_threads; // the number of marking
jmasa@2859 374 // threads we're use
jmasa@2922 375 uint _max_parallel_marking_threads; // max number of marking
jmasa@2859 376 // threads we'll ever use
ysr@342 377 double _sleep_factor; // how much we have to sleep, with
ysr@342 378 // respect to the work we just did, to
ysr@342 379 // meet the marking overhead goal
ysr@342 380 double _marking_task_overhead; // marking target overhead for
ysr@342 381 // a single task
ysr@342 382
ysr@342 383 // same as the two above, but for the cleanup task
ysr@342 384 double _cleanup_sleep_factor;
ysr@342 385 double _cleanup_task_overhead;
ysr@342 386
tonyp@2037 387 FreeRegionList _cleanup_list;
ysr@342 388
ysr@342 389 // CMS marking support structures
ysr@342 390 CMBitMap _markBitMap1;
ysr@342 391 CMBitMap _markBitMap2;
ysr@342 392 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
ysr@342 393 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
ysr@342 394 bool _at_least_one_mark_complete;
ysr@342 395
ysr@342 396 BitMap _region_bm;
ysr@342 397 BitMap _card_bm;
ysr@342 398
ysr@342 399 // Heap bounds
ysr@342 400 HeapWord* _heap_start;
ysr@342 401 HeapWord* _heap_end;
ysr@342 402
ysr@342 403 // For gray objects
ysr@342 404 CMMarkStack _markStack; // Grey objects behind global finger.
ysr@342 405 CMRegionStack _regionStack; // Grey regions behind global finger.
ysr@342 406 HeapWord* volatile _finger; // the global finger, region aligned,
ysr@342 407 // always points to the end of the
ysr@342 408 // last claimed region
ysr@342 409
ysr@342 410 // marking tasks
jmasa@2922 411 uint _max_task_num; // maximum task number
jmasa@2922 412 uint _active_tasks; // task num currently active
ysr@342 413 CMTask** _tasks; // task queue array (max_task_num len)
ysr@342 414 CMTaskQueueSet* _task_queues; // task queue set
ysr@342 415 ParallelTaskTerminator _terminator; // for termination
ysr@342 416
ysr@342 417 // Two sync barriers that are used to synchronise tasks when an
ysr@342 418 // overflow occurs. The algorithm is the following. All tasks enter
ysr@342 419 // the first one to ensure that they have all stopped manipulating
ysr@342 420 // the global data structures. After they exit it, they re-initialise
ysr@342 421 // their data structures and task 0 re-initialises the global data
ysr@342 422 // structures. Then, they enter the second sync barrier. This
ysr@342 423 // ensure, that no task starts doing work before all data
ysr@342 424 // structures (local and global) have been re-initialised. When they
ysr@342 425 // exit it, they are free to start working again.
ysr@342 426 WorkGangBarrierSync _first_overflow_barrier_sync;
ysr@342 427 WorkGangBarrierSync _second_overflow_barrier_sync;
ysr@342 428
ysr@342 429
ysr@342 430 // this is set by any task, when an overflow on the global data
ysr@342 431 // structures is detected.
ysr@342 432 volatile bool _has_overflown;
ysr@342 433 // true: marking is concurrent, false: we're in remark
ysr@342 434 volatile bool _concurrent;
ysr@342 435 // set at the end of a Full GC so that marking aborts
ysr@342 436 volatile bool _has_aborted;
johnc@1755 437
ysr@342 438 // used when remark aborts due to an overflow to indicate that
ysr@342 439 // another concurrent marking phase should start
ysr@342 440 volatile bool _restart_for_overflow;
ysr@342 441
ysr@342 442 // This is true from the very start of concurrent marking until the
ysr@342 443 // point when all the tasks complete their work. It is really used
ysr@342 444 // to determine the points between the end of concurrent marking and
ysr@342 445 // time of remark.
ysr@342 446 volatile bool _concurrent_marking_in_progress;
ysr@342 447
ysr@342 448 // verbose level
ysr@342 449 CMVerboseLevel _verbose_level;
ysr@342 450
ysr@342 451 // These two fields are used to implement the optimisation that
ysr@342 452 // avoids pushing objects on the global/region stack if there are
ysr@342 453 // no collection set regions above the lowest finger.
ysr@342 454
ysr@342 455 // This is the lowest finger (among the global and local fingers),
ysr@342 456 // which is calculated before a new collection set is chosen.
ysr@342 457 HeapWord* _min_finger;
ysr@342 458 // If this flag is true, objects/regions that are marked below the
ysr@342 459 // finger should be pushed on the stack(s). If this is flag is
ysr@342 460 // false, it is safe not to push them on the stack(s).
ysr@342 461 bool _should_gray_objects;
ysr@342 462
ysr@342 463 // All of these times are in ms.
ysr@342 464 NumberSeq _init_times;
ysr@342 465 NumberSeq _remark_times;
ysr@342 466 NumberSeq _remark_mark_times;
ysr@342 467 NumberSeq _remark_weak_ref_times;
ysr@342 468 NumberSeq _cleanup_times;
ysr@342 469 double _total_counting_time;
ysr@342 470 double _total_rs_scrub_time;
ysr@342 471
ysr@342 472 double* _accum_task_vtime; // accumulated task vtime
ysr@342 473
jmasa@2859 474 FlexibleWorkGang* _parallel_workers;
ysr@342 475
tonyp@2413 476 ForceOverflowSettings _force_overflow_conc;
tonyp@2413 477 ForceOverflowSettings _force_overflow_stw;
tonyp@2413 478
ysr@342 479 void weakRefsWork(bool clear_all_soft_refs);
ysr@342 480
ysr@342 481 void swapMarkBitMaps();
ysr@342 482
ysr@342 483 // It resets the global marking data structures, as well as the
ysr@342 484 // task local ones; should be called during initial mark.
ysr@342 485 void reset();
ysr@342 486 // It resets all the marking data structures.
tonyp@2413 487 void clear_marking_state(bool clear_overflow = true);
ysr@342 488
ysr@342 489 // It should be called to indicate which phase we're in (concurrent
ysr@342 490 // mark or remark) and how many threads are currently active.
jmasa@2922 491 void set_phase(uint active_tasks, bool concurrent);
ysr@342 492 // We do this after we're done with marking so that the marking data
ysr@342 493 // structures are initialised to a sensible and predictable state.
ysr@342 494 void set_non_marking_state();
ysr@342 495
ysr@342 496 // prints all gathered CM-related statistics
ysr@342 497 void print_stats();
ysr@342 498
tonyp@2037 499 bool cleanup_list_is_empty() {
tonyp@2037 500 return _cleanup_list.is_empty();
tonyp@2037 501 }
tonyp@2037 502
ysr@342 503 // accessor methods
jmasa@2922 504 uint parallel_marking_threads() { return _parallel_marking_threads; }
jmasa@2922 505 uint max_parallel_marking_threads() { return _max_parallel_marking_threads;}
ysr@342 506 double sleep_factor() { return _sleep_factor; }
ysr@342 507 double marking_task_overhead() { return _marking_task_overhead;}
ysr@342 508 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
ysr@342 509 double cleanup_task_overhead() { return _cleanup_task_overhead;}
ysr@342 510
ysr@342 511 HeapWord* finger() { return _finger; }
ysr@342 512 bool concurrent() { return _concurrent; }
jmasa@2922 513 uint active_tasks() { return _active_tasks; }
ysr@342 514 ParallelTaskTerminator* terminator() { return &_terminator; }
ysr@342 515
ysr@342 516 // It claims the next available region to be scanned by a marking
ysr@342 517 // task. It might return NULL if the next region is empty or we have
ysr@342 518 // run out of regions. In the latter case, out_of_regions()
ysr@342 519 // determines whether we've really run out of regions or the task
ysr@342 520 // should call claim_region() again. This might seem a bit
ysr@342 521 // awkward. Originally, the code was written so that claim_region()
ysr@342 522 // either successfully returned with a non-empty region or there
ysr@342 523 // were no more regions to be claimed. The problem with this was
ysr@342 524 // that, in certain circumstances, it iterated over large chunks of
ysr@342 525 // the heap finding only empty regions and, while it was working, it
ysr@342 526 // was preventing the calling task to call its regular clock
ysr@342 527 // method. So, this way, each task will spend very little time in
ysr@342 528 // claim_region() and is allowed to call the regular clock method
ysr@342 529 // frequently.
ysr@342 530 HeapRegion* claim_region(int task);
ysr@342 531
ysr@342 532 // It determines whether we've run out of regions to scan.
ysr@342 533 bool out_of_regions() { return _finger == _heap_end; }
ysr@342 534
ysr@342 535 // Returns the task with the given id
ysr@342 536 CMTask* task(int id) {
tonyp@1023 537 assert(0 <= id && id < (int) _active_tasks,
tonyp@1023 538 "task id not within active bounds");
ysr@342 539 return _tasks[id];
ysr@342 540 }
ysr@342 541
ysr@342 542 // Returns the task queue with the given id
ysr@342 543 CMTaskQueue* task_queue(int id) {
tonyp@1023 544 assert(0 <= id && id < (int) _active_tasks,
tonyp@1023 545 "task queue id not within active bounds");
ysr@342 546 return (CMTaskQueue*) _task_queues->queue(id);
ysr@342 547 }
ysr@342 548
ysr@342 549 // Returns the task queue set
ysr@342 550 CMTaskQueueSet* task_queues() { return _task_queues; }
ysr@342 551
ysr@342 552 // Access / manipulation of the overflow flag which is set to
ysr@342 553 // indicate that the global stack or region stack has overflown
ysr@342 554 bool has_overflown() { return _has_overflown; }
ysr@342 555 void set_has_overflown() { _has_overflown = true; }
ysr@342 556 void clear_has_overflown() { _has_overflown = false; }
ysr@342 557
ysr@342 558 bool has_aborted() { return _has_aborted; }
ysr@342 559 bool restart_for_overflow() { return _restart_for_overflow; }
ysr@342 560
ysr@342 561 // Methods to enter the two overflow sync barriers
ysr@342 562 void enter_first_sync_barrier(int task_num);
ysr@342 563 void enter_second_sync_barrier(int task_num);
ysr@342 564
tonyp@2413 565 ForceOverflowSettings* force_overflow_conc() {
tonyp@2413 566 return &_force_overflow_conc;
tonyp@2413 567 }
tonyp@2413 568
tonyp@2413 569 ForceOverflowSettings* force_overflow_stw() {
tonyp@2413 570 return &_force_overflow_stw;
tonyp@2413 571 }
tonyp@2413 572
tonyp@2413 573 ForceOverflowSettings* force_overflow() {
tonyp@2413 574 if (concurrent()) {
tonyp@2413 575 return force_overflow_conc();
tonyp@2413 576 } else {
tonyp@2413 577 return force_overflow_stw();
tonyp@2413 578 }
tonyp@2413 579 }
tonyp@2413 580
ysr@342 581 public:
ysr@342 582 // Manipulation of the global mark stack.
ysr@342 583 // Notice that the first mark_stack_push is CAS-based, whereas the
ysr@342 584 // two below are Mutex-based. This is OK since the first one is only
ysr@342 585 // called during evacuation pauses and doesn't compete with the
ysr@342 586 // other two (which are called by the marking tasks during
ysr@342 587 // concurrent marking or remark).
ysr@342 588 bool mark_stack_push(oop p) {
ysr@342 589 _markStack.par_push(p);
ysr@342 590 if (_markStack.overflow()) {
ysr@342 591 set_has_overflown();
ysr@342 592 return false;
ysr@342 593 }
ysr@342 594 return true;
ysr@342 595 }
ysr@342 596 bool mark_stack_push(oop* arr, int n) {
ysr@342 597 _markStack.par_push_arr(arr, n);
ysr@342 598 if (_markStack.overflow()) {
ysr@342 599 set_has_overflown();
ysr@342 600 return false;
ysr@342 601 }
ysr@342 602 return true;
ysr@342 603 }
ysr@342 604 void mark_stack_pop(oop* arr, int max, int* n) {
ysr@342 605 _markStack.par_pop_arr(arr, max, n);
ysr@342 606 }
tonyp@2538 607 size_t mark_stack_size() { return _markStack.size(); }
ysr@342 608 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
tonyp@2538 609 bool mark_stack_overflow() { return _markStack.overflow(); }
tonyp@2538 610 bool mark_stack_empty() { return _markStack.isEmpty(); }
ysr@342 611
johnc@1755 612 // (Lock-free) Manipulation of the region stack
johnc@1755 613 bool region_stack_push_lock_free(MemRegion mr) {
tonyp@1358 614 // Currently we only call the lock-free version during evacuation
tonyp@1358 615 // pauses.
tonyp@1358 616 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
tonyp@1358 617
johnc@1755 618 _regionStack.push_lock_free(mr);
ysr@342 619 if (_regionStack.overflow()) {
ysr@342 620 set_has_overflown();
ysr@342 621 return false;
ysr@342 622 }
ysr@342 623 return true;
ysr@342 624 }
johnc@1755 625
johnc@1755 626 // Lock-free version of region-stack pop. Should only be
johnc@1755 627 // called in tandem with other lock-free pops.
johnc@1755 628 MemRegion region_stack_pop_lock_free() {
johnc@1755 629 return _regionStack.pop_lock_free();
johnc@1755 630 }
johnc@1755 631
tonyp@1358 632 #if 0
johnc@1755 633 // The routines that manipulate the region stack with a lock are
johnc@1755 634 // not currently used. They should be retained, however, as a
johnc@1755 635 // diagnostic aid.
tonyp@1358 636
tonyp@1358 637 bool region_stack_push_with_lock(MemRegion mr) {
tonyp@1358 638 // Currently we only call the lock-based version during either
tonyp@1358 639 // concurrent marking or remark.
tonyp@1358 640 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
tonyp@1358 641 "if we are at a safepoint it should be the remark safepoint");
tonyp@1358 642
tonyp@1358 643 _regionStack.push_with_lock(mr);
tonyp@1358 644 if (_regionStack.overflow()) {
tonyp@1358 645 set_has_overflown();
tonyp@1358 646 return false;
tonyp@1358 647 }
tonyp@1358 648 return true;
tonyp@1358 649 }
johnc@1755 650
tonyp@1358 651 MemRegion region_stack_pop_with_lock() {
tonyp@1358 652 // Currently we only call the lock-based version during either
tonyp@1358 653 // concurrent marking or remark.
tonyp@1358 654 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
tonyp@1358 655 "if we are at a safepoint it should be the remark safepoint");
tonyp@1358 656
tonyp@1358 657 return _regionStack.pop_with_lock();
tonyp@1358 658 }
johnc@1755 659 #endif
tonyp@1358 660
ysr@342 661 int region_stack_size() { return _regionStack.size(); }
ysr@342 662 bool region_stack_overflow() { return _regionStack.overflow(); }
ysr@342 663 bool region_stack_empty() { return _regionStack.isEmpty(); }
ysr@342 664
johnc@1755 665 // Iterate over any regions that were aborted while draining the
johnc@1755 666 // region stack (any such regions are saved in the corresponding
johnc@1755 667 // CMTask) and invalidate (i.e. assign to the empty MemRegion())
johnc@1755 668 // any regions that point into the collection set.
johnc@1755 669 bool invalidate_aborted_regions_in_cset();
johnc@1755 670
johnc@1755 671 // Returns true if there are any aborted memory regions.
johnc@1755 672 bool has_aborted_regions();
johnc@1755 673
ysr@342 674 bool concurrent_marking_in_progress() {
ysr@342 675 return _concurrent_marking_in_progress;
ysr@342 676 }
ysr@342 677 void set_concurrent_marking_in_progress() {
ysr@342 678 _concurrent_marking_in_progress = true;
ysr@342 679 }
ysr@342 680 void clear_concurrent_marking_in_progress() {
ysr@342 681 _concurrent_marking_in_progress = false;
ysr@342 682 }
ysr@342 683
ysr@342 684 void update_accum_task_vtime(int i, double vtime) {
ysr@342 685 _accum_task_vtime[i] += vtime;
ysr@342 686 }
ysr@342 687
ysr@342 688 double all_task_accum_vtime() {
ysr@342 689 double ret = 0.0;
ysr@342 690 for (int i = 0; i < (int)_max_task_num; ++i)
ysr@342 691 ret += _accum_task_vtime[i];
ysr@342 692 return ret;
ysr@342 693 }
ysr@342 694
ysr@342 695 // Attempts to steal an object from the task queues of other tasks
ysr@342 696 bool try_stealing(int task_num, int* hash_seed, oop& obj) {
ysr@342 697 return _task_queues->steal(task_num, hash_seed, obj);
ysr@342 698 }
ysr@342 699
ysr@342 700 // It grays an object by first marking it. Then, if it's behind the
ysr@342 701 // global finger, it also pushes it on the global stack.
ysr@342 702 void deal_with_reference(oop obj);
ysr@342 703
ysr@342 704 ConcurrentMark(ReservedSpace rs, int max_regions);
ysr@342 705 ~ConcurrentMark();
ysr@342 706 ConcurrentMarkThread* cmThread() { return _cmThread; }
ysr@342 707
ysr@342 708 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
ysr@342 709 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
ysr@342 710
jmasa@2859 711 // Returns the number of GC threads to be used in a concurrent
jmasa@2859 712 // phase based on the number of GC threads being used in a STW
jmasa@2859 713 // phase.
jmasa@2922 714 uint scale_parallel_threads(uint n_par_threads);
jmasa@2859 715
jmasa@2859 716 // Calculates the number of GC threads to be used in a concurrent phase.
jmasa@2922 717 uint calc_parallel_marking_threads();
jmasa@2859 718
ysr@342 719 // The following three are interaction between CM and
ysr@342 720 // G1CollectedHeap
ysr@342 721
ysr@342 722 // This notifies CM that a root during initial-mark needs to be
tonyp@2981 723 // grayed. It is MT-safe.
tonyp@2981 724 inline void grayRoot(oop obj, size_t word_size);
tonyp@2981 725
ysr@342 726 // It's used during evacuation pauses to gray a region, if
ysr@342 727 // necessary, and it's MT-safe. It assumes that the caller has
ysr@342 728 // marked any objects on that region. If _should_gray_objects is
ysr@342 729 // true and we're still doing concurrent marking, the region is
ysr@342 730 // pushed on the region stack, if it is located below the global
ysr@342 731 // finger, otherwise we do nothing.
ysr@342 732 void grayRegionIfNecessary(MemRegion mr);
tonyp@2981 733
ysr@342 734 // It's used during evacuation pauses to mark and, if necessary,
ysr@342 735 // gray a single object and it's MT-safe. It assumes the caller did
ysr@342 736 // not mark the object. If _should_gray_objects is true and we're
ysr@342 737 // still doing concurrent marking, the objects is pushed on the
ysr@342 738 // global stack, if it is located below the global finger, otherwise
ysr@342 739 // we do nothing.
ysr@342 740 void markAndGrayObjectIfNecessary(oop p);
ysr@342 741
tonyp@1388 742 // It iterates over the heap and for each object it comes across it
tonyp@1388 743 // will dump the contents of its reference fields, as well as
tonyp@1388 744 // liveness information for the object and its referents. The dump
tonyp@1388 745 // will be written to a file with the following name:
johnc@2534 746 // G1PrintReachableBaseFile + "." + str.
johnc@2534 747 // vo decides whether the prev (vo == UsePrevMarking), the next
johnc@2534 748 // (vo == UseNextMarking) marking information, or the mark word
johnc@2534 749 // (vo == UseMarkWord) will be used to determine the liveness of
johnc@2534 750 // each object / referent.
johnc@2534 751 // If all is true, all objects in the heap will be dumped, otherwise
johnc@2534 752 // only the live ones. In the dump the following symbols / breviations
johnc@2534 753 // are used:
tonyp@1388 754 // M : an explicitly live object (its bitmap bit is set)
tonyp@1388 755 // > : an implicitly live object (over tams)
tonyp@1388 756 // O : an object outside the G1 heap (typically: in the perm gen)
tonyp@1388 757 // NOT : a reference field whose referent is not live
tonyp@1388 758 // AND MARKED : indicates that an object is both explicitly and
tonyp@1388 759 // implicitly live (it should be one or the other, not both)
tonyp@1388 760 void print_reachable(const char* str,
johnc@2534 761 VerifyOption vo, bool all) PRODUCT_RETURN;
ysr@342 762
ysr@342 763 // Clear the next marking bitmap (will be called concurrently).
ysr@342 764 void clearNextBitmap();
ysr@342 765
ysr@342 766 // These two do the work that needs to be done before and after the
ysr@342 767 // initial root checkpoint. Since this checkpoint can be done at two
ysr@342 768 // different points (i.e. an explicit pause or piggy-backed on a
ysr@342 769 // young collection), then it's nice to be able to easily share the
ysr@342 770 // pre/post code. It might be the case that we can put everything in
ysr@342 771 // the post method. TP
ysr@342 772 void checkpointRootsInitialPre();
ysr@342 773 void checkpointRootsInitialPost();
ysr@342 774
ysr@342 775 // Do concurrent phase of marking, to a tentative transitive closure.
ysr@342 776 void markFromRoots();
ysr@342 777
ysr@342 778 // Process all unprocessed SATB buffers. It is called at the
ysr@342 779 // beginning of an evacuation pause.
ysr@342 780 void drainAllSATBBuffers();
ysr@342 781
ysr@342 782 void checkpointRootsFinal(bool clear_all_soft_refs);
ysr@342 783 void checkpointRootsFinalWork();
ysr@342 784 void calcDesiredRegions();
ysr@342 785 void cleanup();
ysr@342 786 void completeCleanup();
ysr@342 787
ysr@342 788 // Mark in the previous bitmap. NB: this is usually read-only, so use
ysr@342 789 // this carefully!
tonyp@2981 790 inline void markPrev(oop p);
tonyp@2981 791 inline void markNext(oop p);
ysr@342 792 void clear(oop p);
tonyp@2981 793 // Clears marks for all objects in the given range, for the prev,
tonyp@2981 794 // next, or both bitmaps. NB: the previous bitmap is usually
tonyp@2981 795 // read-only, so use this carefully!
tonyp@2981 796 void clearRangePrevBitmap(MemRegion mr);
tonyp@2981 797 void clearRangeNextBitmap(MemRegion mr);
tonyp@2981 798 void clearRangeBothBitmaps(MemRegion mr);
ysr@342 799
tonyp@2981 800 // Notify data structures that a GC has started.
tonyp@2981 801 void note_start_of_gc() {
tonyp@2981 802 _markStack.note_start_of_gc();
ysr@342 803 }
tonyp@2981 804
tonyp@2981 805 // Notify data structures that a GC is finished.
tonyp@2981 806 void note_end_of_gc() {
tonyp@2981 807 _markStack.note_end_of_gc();
tonyp@2981 808 }
tonyp@2981 809
ysr@342 810 // Iterate over the oops in the mark stack and all local queues. It
ysr@342 811 // also calls invalidate_entries_into_cset() on the region stack.
ysr@342 812 void oops_do(OopClosure* f);
tonyp@2981 813
tonyp@2981 814 // Verify that there are no CSet oops on the stacks (taskqueues /
tonyp@2981 815 // global mark stack), enqueued SATB buffers, per-thread SATB
tonyp@2981 816 // buffers, and fingers (global / per-task). The boolean parameters
tonyp@2981 817 // decide which of the above data structures to verify. If marking
tonyp@2981 818 // is not in progress, it's a no-op.
tonyp@2981 819 void verify_no_cset_oops(bool verify_stacks,
tonyp@2981 820 bool verify_enqueued_buffers,
tonyp@2981 821 bool verify_thread_buffers,
tonyp@2981 822 bool verify_fingers) PRODUCT_RETURN;
tonyp@2981 823
ysr@342 824 // It is called at the end of an evacuation pause during marking so
ysr@342 825 // that CM is notified of where the new end of the heap is. It
ysr@342 826 // doesn't do anything if concurrent_marking_in_progress() is false,
ysr@342 827 // unless the force parameter is true.
ysr@342 828 void update_g1_committed(bool force = false);
ysr@342 829
ysr@342 830 void complete_marking_in_collection_set();
ysr@342 831
ysr@342 832 // It indicates that a new collection set is being chosen.
ysr@342 833 void newCSet();
johnc@2475 834
ysr@342 835 // It registers a collection set heap region with CM. This is used
ysr@342 836 // to determine whether any heap regions are located above the finger.
ysr@342 837 void registerCSetRegion(HeapRegion* hr);
ysr@342 838
johnc@2475 839 // Resets the region fields of any active CMTask whose region fields
johnc@2475 840 // are in the collection set (i.e. the region currently claimed by
johnc@2475 841 // the CMTask will be evacuated and may be used, subsequently, as
johnc@2475 842 // an alloc region). When this happens the region fields in the CMTask
johnc@2475 843 // are stale and, hence, should be cleared causing the worker thread
johnc@2475 844 // to claim a new region.
johnc@2475 845 void reset_active_task_region_fields_in_cset();
johnc@2475 846
johnc@1394 847 // Registers the maximum region-end associated with a set of
johnc@1394 848 // regions with CM. Again this is used to determine whether any
johnc@1394 849 // heap regions are located above the finger.
johnc@1394 850 void register_collection_set_finger(HeapWord* max_finger) {
johnc@1394 851 // max_finger is the highest heap region end of the regions currently
johnc@1394 852 // contained in the collection set. If this value is larger than
johnc@1394 853 // _min_finger then we need to gray objects.
johnc@1394 854 // This routine is like registerCSetRegion but for an entire
johnc@1394 855 // collection of regions.
tonyp@2538 856 if (max_finger > _min_finger) {
johnc@1394 857 _should_gray_objects = true;
tonyp@2538 858 }
johnc@1394 859 }
johnc@1394 860
ysr@342 861 // Returns "true" if at least one mark has been completed.
ysr@342 862 bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
ysr@342 863
ysr@342 864 bool isMarked(oop p) const {
ysr@342 865 assert(p != NULL && p->is_oop(), "expected an oop");
ysr@342 866 HeapWord* addr = (HeapWord*)p;
ysr@342 867 assert(addr >= _nextMarkBitMap->startWord() ||
ysr@342 868 addr < _nextMarkBitMap->endWord(), "in a region");
ysr@342 869
ysr@342 870 return _nextMarkBitMap->isMarked(addr);
ysr@342 871 }
ysr@342 872
ysr@342 873 inline bool not_yet_marked(oop p) const;
ysr@342 874
ysr@342 875 // XXX Debug code
ysr@342 876 bool containing_card_is_marked(void* p);
ysr@342 877 bool containing_cards_are_marked(void* start, void* last);
ysr@342 878
ysr@342 879 bool isPrevMarked(oop p) const {
ysr@342 880 assert(p != NULL && p->is_oop(), "expected an oop");
ysr@342 881 HeapWord* addr = (HeapWord*)p;
ysr@342 882 assert(addr >= _prevMarkBitMap->startWord() ||
ysr@342 883 addr < _prevMarkBitMap->endWord(), "in a region");
ysr@342 884
ysr@342 885 return _prevMarkBitMap->isMarked(addr);
ysr@342 886 }
ysr@342 887
jmasa@2922 888 inline bool do_yield_check(uint worker_i = 0);
ysr@342 889 inline bool should_yield();
ysr@342 890
ysr@342 891 // Called to abort the marking cycle after a Full GC takes palce.
ysr@342 892 void abort();
ysr@342 893
ysr@342 894 // This prints the global/local fingers. It is used for debugging.
ysr@342 895 NOT_PRODUCT(void print_finger();)
ysr@342 896
ysr@342 897 void print_summary_info();
ysr@342 898
tonyp@1019 899 void print_worker_threads_on(outputStream* st) const;
tonyp@1019 900
ysr@342 901 // The following indicate whether a given verbose level has been
ysr@342 902 // set. Notice that anything above stats is conditional to
ysr@342 903 // _MARKING_VERBOSE_ having been set to 1
tonyp@2538 904 bool verbose_stats() {
tonyp@2538 905 return _verbose_level >= stats_verbose;
tonyp@2538 906 }
tonyp@2538 907 bool verbose_low() {
tonyp@2538 908 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
tonyp@2538 909 }
tonyp@2538 910 bool verbose_medium() {
tonyp@2538 911 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
tonyp@2538 912 }
tonyp@2538 913 bool verbose_high() {
tonyp@2538 914 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
tonyp@2538 915 }
ysr@342 916 };
ysr@342 917
ysr@342 918 // A class representing a marking task.
ysr@342 919 class CMTask : public TerminatorTerminator {
ysr@342 920 private:
ysr@342 921 enum PrivateConstants {
ysr@342 922 // the regular clock call is called once the scanned words reaches
ysr@342 923 // this limit
ysr@342 924 words_scanned_period = 12*1024,
ysr@342 925 // the regular clock call is called once the number of visited
ysr@342 926 // references reaches this limit
ysr@342 927 refs_reached_period = 384,
ysr@342 928 // initial value for the hash seed, used in the work stealing code
ysr@342 929 init_hash_seed = 17,
ysr@342 930 // how many entries will be transferred between global stack and
ysr@342 931 // local queues
ysr@342 932 global_stack_transfer_size = 16
ysr@342 933 };
ysr@342 934
ysr@342 935 int _task_id;
ysr@342 936 G1CollectedHeap* _g1h;
ysr@342 937 ConcurrentMark* _cm;
ysr@342 938 CMBitMap* _nextMarkBitMap;
ysr@342 939 // the task queue of this task
ysr@342 940 CMTaskQueue* _task_queue;
ysr@845 941 private:
ysr@342 942 // the task queue set---needed for stealing
ysr@342 943 CMTaskQueueSet* _task_queues;
ysr@342 944 // indicates whether the task has been claimed---this is only for
ysr@342 945 // debugging purposes
ysr@342 946 bool _claimed;
ysr@342 947
ysr@342 948 // number of calls to this task
ysr@342 949 int _calls;
ysr@342 950
ysr@342 951 // when the virtual timer reaches this time, the marking step should
ysr@342 952 // exit
ysr@342 953 double _time_target_ms;
ysr@342 954 // the start time of the current marking step
ysr@342 955 double _start_time_ms;
ysr@342 956
ysr@342 957 // the oop closure used for iterations over oops
tonyp@2533 958 G1CMOopClosure* _cm_oop_closure;
ysr@342 959
ysr@342 960 // the region this task is scanning, NULL if we're not scanning any
ysr@342 961 HeapRegion* _curr_region;
ysr@342 962 // the local finger of this task, NULL if we're not scanning a region
ysr@342 963 HeapWord* _finger;
ysr@342 964 // limit of the region this task is scanning, NULL if we're not scanning one
ysr@342 965 HeapWord* _region_limit;
ysr@342 966
ysr@342 967 // This is used only when we scan regions popped from the region
ysr@342 968 // stack. It records what the last object on such a region we
ysr@342 969 // scanned was. It is used to ensure that, if we abort region
ysr@342 970 // iteration, we do not rescan the first part of the region. This
ysr@342 971 // should be NULL when we're not scanning a region from the region
ysr@342 972 // stack.
ysr@342 973 HeapWord* _region_finger;
ysr@342 974
johnc@1755 975 // If we abort while scanning a region we record the remaining
johnc@1755 976 // unscanned portion and check this field when marking restarts.
johnc@1755 977 // This avoids having to push on the region stack while other
johnc@1755 978 // marking threads may still be popping regions.
johnc@1755 979 // If we were to push the unscanned portion directly to the
johnc@1755 980 // region stack then we would need to using locking versions
johnc@1755 981 // of the push and pop operations.
johnc@1755 982 MemRegion _aborted_region;
johnc@1755 983
ysr@342 984 // the number of words this task has scanned
ysr@342 985 size_t _words_scanned;
ysr@342 986 // When _words_scanned reaches this limit, the regular clock is
ysr@342 987 // called. Notice that this might be decreased under certain
ysr@342 988 // circumstances (i.e. when we believe that we did an expensive
ysr@342 989 // operation).
ysr@342 990 size_t _words_scanned_limit;
ysr@342 991 // the initial value of _words_scanned_limit (i.e. what it was
ysr@342 992 // before it was decreased).
ysr@342 993 size_t _real_words_scanned_limit;
ysr@342 994
ysr@342 995 // the number of references this task has visited
ysr@342 996 size_t _refs_reached;
ysr@342 997 // When _refs_reached reaches this limit, the regular clock is
ysr@342 998 // called. Notice this this might be decreased under certain
ysr@342 999 // circumstances (i.e. when we believe that we did an expensive
ysr@342 1000 // operation).
ysr@342 1001 size_t _refs_reached_limit;
ysr@342 1002 // the initial value of _refs_reached_limit (i.e. what it was before
ysr@342 1003 // it was decreased).
ysr@342 1004 size_t _real_refs_reached_limit;
ysr@342 1005
ysr@342 1006 // used by the work stealing stuff
ysr@342 1007 int _hash_seed;
ysr@342 1008 // if this is true, then the task has aborted for some reason
ysr@342 1009 bool _has_aborted;
ysr@342 1010 // set when the task aborts because it has met its time quota
johnc@2059 1011 bool _has_timed_out;
ysr@342 1012 // true when we're draining SATB buffers; this avoids the task
ysr@342 1013 // aborting due to SATB buffers being available (as we're already
ysr@342 1014 // dealing with them)
ysr@342 1015 bool _draining_satb_buffers;
ysr@342 1016
ysr@342 1017 // number sequence of past step times
ysr@342 1018 NumberSeq _step_times_ms;
ysr@342 1019 // elapsed time of this task
ysr@342 1020 double _elapsed_time_ms;
ysr@342 1021 // termination time of this task
ysr@342 1022 double _termination_time_ms;
ysr@342 1023 // when this task got into the termination protocol
ysr@342 1024 double _termination_start_time_ms;
ysr@342 1025
ysr@342 1026 // true when the task is during a concurrent phase, false when it is
ysr@342 1027 // in the remark phase (so, in the latter case, we do not have to
ysr@342 1028 // check all the things that we have to check during the concurrent
ysr@342 1029 // phase, i.e. SATB buffer availability...)
ysr@342 1030 bool _concurrent;
ysr@342 1031
ysr@342 1032 TruncatedSeq _marking_step_diffs_ms;
ysr@342 1033
ysr@342 1034 // LOTS of statistics related with this task
ysr@342 1035 #if _MARKING_STATS_
ysr@342 1036 NumberSeq _all_clock_intervals_ms;
ysr@342 1037 double _interval_start_time_ms;
ysr@342 1038
ysr@342 1039 int _aborted;
ysr@342 1040 int _aborted_overflow;
ysr@342 1041 int _aborted_cm_aborted;
ysr@342 1042 int _aborted_yield;
ysr@342 1043 int _aborted_timed_out;
ysr@342 1044 int _aborted_satb;
ysr@342 1045 int _aborted_termination;
ysr@342 1046
ysr@342 1047 int _steal_attempts;
ysr@342 1048 int _steals;
ysr@342 1049
ysr@342 1050 int _clock_due_to_marking;
ysr@342 1051 int _clock_due_to_scanning;
ysr@342 1052
ysr@342 1053 int _local_pushes;
ysr@342 1054 int _local_pops;
ysr@342 1055 int _local_max_size;
ysr@342 1056 int _objs_scanned;
ysr@342 1057
ysr@342 1058 int _global_pushes;
ysr@342 1059 int _global_pops;
ysr@342 1060 int _global_max_size;
ysr@342 1061
ysr@342 1062 int _global_transfers_to;
ysr@342 1063 int _global_transfers_from;
ysr@342 1064
ysr@342 1065 int _region_stack_pops;
ysr@342 1066
ysr@342 1067 int _regions_claimed;
ysr@342 1068 int _objs_found_on_bitmap;
ysr@342 1069
ysr@342 1070 int _satb_buffers_processed;
ysr@342 1071 #endif // _MARKING_STATS_
ysr@342 1072
ysr@342 1073 // it updates the local fields after this task has claimed
ysr@342 1074 // a new region to scan
ysr@342 1075 void setup_for_region(HeapRegion* hr);
ysr@342 1076 // it brings up-to-date the limit of the region
ysr@342 1077 void update_region_limit();
ysr@342 1078
ysr@342 1079 // called when either the words scanned or the refs visited limit
ysr@342 1080 // has been reached
ysr@342 1081 void reached_limit();
ysr@342 1082 // recalculates the words scanned and refs visited limits
ysr@342 1083 void recalculate_limits();
ysr@342 1084 // decreases the words scanned and refs visited limits when we reach
ysr@342 1085 // an expensive operation
ysr@342 1086 void decrease_limits();
ysr@342 1087 // it checks whether the words scanned or refs visited reached their
ysr@342 1088 // respective limit and calls reached_limit() if they have
ysr@342 1089 void check_limits() {
ysr@342 1090 if (_words_scanned >= _words_scanned_limit ||
tonyp@2538 1091 _refs_reached >= _refs_reached_limit) {
ysr@342 1092 reached_limit();
tonyp@2538 1093 }
ysr@342 1094 }
ysr@342 1095 // this is supposed to be called regularly during a marking step as
ysr@342 1096 // it checks a bunch of conditions that might cause the marking step
ysr@342 1097 // to abort
ysr@342 1098 void regular_clock_call();
ysr@342 1099 bool concurrent() { return _concurrent; }
ysr@342 1100
ysr@342 1101 public:
ysr@342 1102 // It resets the task; it should be called right at the beginning of
ysr@342 1103 // a marking phase.
ysr@342 1104 void reset(CMBitMap* _nextMarkBitMap);
ysr@342 1105 // it clears all the fields that correspond to a claimed region.
ysr@342 1106 void clear_region_fields();
ysr@342 1107
ysr@342 1108 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
ysr@342 1109
ysr@342 1110 // The main method of this class which performs a marking step
ysr@342 1111 // trying not to exceed the given duration. However, it might exit
ysr@342 1112 // prematurely, according to some conditions (i.e. SATB buffers are
ysr@342 1113 // available for processing).
johnc@2059 1114 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
ysr@342 1115
ysr@342 1116 // These two calls start and stop the timer
ysr@342 1117 void record_start_time() {
ysr@342 1118 _elapsed_time_ms = os::elapsedTime() * 1000.0;
ysr@342 1119 }
ysr@342 1120 void record_end_time() {
ysr@342 1121 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
ysr@342 1122 }
ysr@342 1123
ysr@342 1124 // returns the task ID
ysr@342 1125 int task_id() { return _task_id; }
ysr@342 1126
ysr@342 1127 // From TerminatorTerminator. It determines whether this task should
ysr@342 1128 // exit the termination protocol after it's entered it.
ysr@342 1129 virtual bool should_exit_termination();
ysr@342 1130
johnc@2475 1131 // Resets the local region fields after a task has finished scanning a
johnc@2475 1132 // region; or when they have become stale as a result of the region
johnc@2475 1133 // being evacuated.
johnc@2475 1134 void giveup_current_region();
johnc@2475 1135
ysr@342 1136 HeapWord* finger() { return _finger; }
ysr@342 1137
ysr@342 1138 bool has_aborted() { return _has_aborted; }
ysr@342 1139 void set_has_aborted() { _has_aborted = true; }
ysr@342 1140 void clear_has_aborted() { _has_aborted = false; }
johnc@2059 1141 bool has_timed_out() { return _has_timed_out; }
johnc@2059 1142 bool claimed() { return _claimed; }
ysr@342 1143
johnc@1755 1144 // Support routines for the partially scanned region that may be
johnc@1755 1145 // recorded as a result of aborting while draining the CMRegionStack
johnc@1755 1146 MemRegion aborted_region() { return _aborted_region; }
johnc@1755 1147 void set_aborted_region(MemRegion mr)
johnc@1755 1148 { _aborted_region = mr; }
johnc@1755 1149
johnc@1755 1150 // Clears any recorded partially scanned region
johnc@1755 1151 void clear_aborted_region() { set_aborted_region(MemRegion()); }
johnc@1755 1152
tonyp@2533 1153 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
ysr@342 1154
ysr@342 1155 // It grays the object by marking it and, if necessary, pushing it
ysr@342 1156 // on the local queue
tonyp@2533 1157 inline void deal_with_reference(oop obj);
ysr@342 1158
ysr@342 1159 // It scans an object and visits its children.
tonyp@2533 1160 void scan_object(oop obj);
ysr@342 1161
ysr@342 1162 // It pushes an object on the local queue.
tonyp@2533 1163 inline void push(oop obj);
ysr@342 1164
ysr@342 1165 // These two move entries to/from the global stack.
ysr@342 1166 void move_entries_to_global_stack();
ysr@342 1167 void get_entries_from_global_stack();
ysr@342 1168
ysr@342 1169 // It pops and scans objects from the local queue. If partially is
ysr@342 1170 // true, then it stops when the queue size is of a given limit. If
ysr@342 1171 // partially is false, then it stops when the queue is empty.
ysr@342 1172 void drain_local_queue(bool partially);
ysr@342 1173 // It moves entries from the global stack to the local queue and
ysr@342 1174 // drains the local queue. If partially is true, then it stops when
ysr@342 1175 // both the global stack and the local queue reach a given size. If
ysr@342 1176 // partially if false, it tries to empty them totally.
ysr@342 1177 void drain_global_stack(bool partially);
ysr@342 1178 // It keeps picking SATB buffers and processing them until no SATB
ysr@342 1179 // buffers are available.
ysr@342 1180 void drain_satb_buffers();
tonyp@2981 1181
ysr@342 1182 // It keeps popping regions from the region stack and processing
ysr@342 1183 // them until the region stack is empty.
ysr@342 1184 void drain_region_stack(BitMapClosure* closure);
ysr@342 1185
ysr@342 1186 // moves the local finger to a new location
ysr@342 1187 inline void move_finger_to(HeapWord* new_finger) {
tonyp@1023 1188 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
ysr@342 1189 _finger = new_finger;
ysr@342 1190 }
ysr@342 1191
ysr@342 1192 // moves the region finger to a new location
ysr@342 1193 inline void move_region_finger_to(HeapWord* new_finger) {
tonyp@1023 1194 assert(new_finger < _cm->finger(), "invariant");
ysr@342 1195 _region_finger = new_finger;
ysr@342 1196 }
ysr@342 1197
ysr@342 1198 CMTask(int task_num, ConcurrentMark *cm,
ysr@342 1199 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
ysr@342 1200
ysr@342 1201 // it prints statistics associated with this task
ysr@342 1202 void print_stats();
ysr@342 1203
ysr@342 1204 #if _MARKING_STATS_
ysr@342 1205 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
ysr@342 1206 #endif // _MARKING_STATS_
ysr@342 1207 };
stefank@1879 1208
tonyp@2282 1209 // Class that's used to to print out per-region liveness
tonyp@2282 1210 // information. It's currently used at the end of marking and also
tonyp@2282 1211 // after we sort the old regions at the end of the cleanup operation.
tonyp@2282 1212 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
tonyp@2282 1213 private:
tonyp@2282 1214 outputStream* _out;
tonyp@2282 1215
tonyp@2282 1216 // Accumulators for these values.
tonyp@2282 1217 size_t _total_used_bytes;
tonyp@2282 1218 size_t _total_capacity_bytes;
tonyp@2282 1219 size_t _total_prev_live_bytes;
tonyp@2282 1220 size_t _total_next_live_bytes;
tonyp@2282 1221
tonyp@2282 1222 // These are set up when we come across a "stars humongous" region
tonyp@2282 1223 // (as this is where most of this information is stored, not in the
tonyp@2282 1224 // subsequent "continues humongous" regions). After that, for every
tonyp@2282 1225 // region in a given humongous region series we deduce the right
tonyp@2282 1226 // values for it by simply subtracting the appropriate amount from
tonyp@2282 1227 // these fields. All these values should reach 0 after we've visited
tonyp@2282 1228 // the last region in the series.
tonyp@2282 1229 size_t _hum_used_bytes;
tonyp@2282 1230 size_t _hum_capacity_bytes;
tonyp@2282 1231 size_t _hum_prev_live_bytes;
tonyp@2282 1232 size_t _hum_next_live_bytes;
tonyp@2282 1233
tonyp@2282 1234 static double perc(size_t val, size_t total) {
tonyp@2282 1235 if (total == 0) {
tonyp@2282 1236 return 0.0;
tonyp@2282 1237 } else {
tonyp@2282 1238 return 100.0 * ((double) val / (double) total);
tonyp@2282 1239 }
tonyp@2282 1240 }
tonyp@2282 1241
tonyp@2282 1242 static double bytes_to_mb(size_t val) {
tonyp@2282 1243 return (double) val / (double) M;
tonyp@2282 1244 }
tonyp@2282 1245
tonyp@2282 1246 // See the .cpp file.
tonyp@2282 1247 size_t get_hum_bytes(size_t* hum_bytes);
tonyp@2282 1248 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
tonyp@2282 1249 size_t* prev_live_bytes, size_t* next_live_bytes);
tonyp@2282 1250
tonyp@2282 1251 public:
tonyp@2282 1252 // The header and footer are printed in the constructor and
tonyp@2282 1253 // destructor respectively.
tonyp@2282 1254 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
tonyp@2282 1255 virtual bool doHeapRegion(HeapRegion* r);
tonyp@2282 1256 ~G1PrintRegionLivenessInfoClosure();
tonyp@2282 1257 };
tonyp@2282 1258
stefank@1879 1259 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP