Mon, 05 Apr 2010 12:19:22 -0400
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
Summary: Calling the methods region_stack_push() and region_stack_pop() concurrent is not MT-safe. The assumption is that we will only call region_stack_push() during a GC pause and region_stack_pop() during marking. Unfortunately, we also call region_stack_push() during marking which seems to be introducing subtle marking failures. This change introduces lock-based methods for pushing / popping to be called during marking.
Reviewed-by: iveresov, johnc
ysr@777 | 1 | /* |
xdono@1014 | 2 | * Copyright 2001-2009 Sun Microsystems, Inc. All Rights Reserved. |
ysr@777 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
ysr@777 | 4 | * |
ysr@777 | 5 | * This code is free software; you can redistribute it and/or modify it |
ysr@777 | 6 | * under the terms of the GNU General Public License version 2 only, as |
ysr@777 | 7 | * published by the Free Software Foundation. |
ysr@777 | 8 | * |
ysr@777 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
ysr@777 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
ysr@777 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
ysr@777 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
ysr@777 | 13 | * accompanied this code). |
ysr@777 | 14 | * |
ysr@777 | 15 | * You should have received a copy of the GNU General Public License version |
ysr@777 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
ysr@777 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
ysr@777 | 18 | * |
ysr@777 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
ysr@777 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
ysr@777 | 21 | * have any questions. |
ysr@777 | 22 | * |
ysr@777 | 23 | */ |
ysr@777 | 24 | |
ysr@777 | 25 | class G1CollectedHeap; |
ysr@777 | 26 | class CMTask; |
jcoomes@1746 | 27 | typedef GenericTaskQueue<oop> CMTaskQueue; |
jcoomes@1746 | 28 | typedef GenericTaskQueueSet<CMTaskQueue> CMTaskQueueSet; |
ysr@777 | 29 | |
ysr@777 | 30 | // A generic CM bit map. This is essentially a wrapper around the BitMap |
ysr@777 | 31 | // class, with one bit per (1<<_shifter) HeapWords. |
ysr@777 | 32 | |
apetrusenko@984 | 33 | class CMBitMapRO VALUE_OBJ_CLASS_SPEC { |
ysr@777 | 34 | protected: |
ysr@777 | 35 | HeapWord* _bmStartWord; // base address of range covered by map |
ysr@777 | 36 | size_t _bmWordSize; // map size (in #HeapWords covered) |
ysr@777 | 37 | const int _shifter; // map to char or bit |
ysr@777 | 38 | VirtualSpace _virtual_space; // underlying the bit map |
ysr@777 | 39 | BitMap _bm; // the bit map itself |
ysr@777 | 40 | |
ysr@777 | 41 | public: |
ysr@777 | 42 | // constructor |
ysr@777 | 43 | CMBitMapRO(ReservedSpace rs, int shifter); |
ysr@777 | 44 | |
ysr@777 | 45 | enum { do_yield = true }; |
ysr@777 | 46 | |
ysr@777 | 47 | // inquiries |
ysr@777 | 48 | HeapWord* startWord() const { return _bmStartWord; } |
ysr@777 | 49 | size_t sizeInWords() const { return _bmWordSize; } |
ysr@777 | 50 | // the following is one past the last word in space |
ysr@777 | 51 | HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } |
ysr@777 | 52 | |
ysr@777 | 53 | // read marks |
ysr@777 | 54 | |
ysr@777 | 55 | bool isMarked(HeapWord* addr) const { |
ysr@777 | 56 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 57 | "outside underlying space?"); |
ysr@777 | 58 | return _bm.at(heapWordToOffset(addr)); |
ysr@777 | 59 | } |
ysr@777 | 60 | |
ysr@777 | 61 | // iteration |
ysr@777 | 62 | bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); } |
ysr@777 | 63 | bool iterate(BitMapClosure* cl, MemRegion mr); |
ysr@777 | 64 | |
ysr@777 | 65 | // Return the address corresponding to the next marked bit at or after |
ysr@777 | 66 | // "addr", and before "limit", if "limit" is non-NULL. If there is no |
ysr@777 | 67 | // such bit, returns "limit" if that is non-NULL, or else "endWord()". |
ysr@777 | 68 | HeapWord* getNextMarkedWordAddress(HeapWord* addr, |
ysr@777 | 69 | HeapWord* limit = NULL) const; |
ysr@777 | 70 | // Return the address corresponding to the next unmarked bit at or after |
ysr@777 | 71 | // "addr", and before "limit", if "limit" is non-NULL. If there is no |
ysr@777 | 72 | // such bit, returns "limit" if that is non-NULL, or else "endWord()". |
ysr@777 | 73 | HeapWord* getNextUnmarkedWordAddress(HeapWord* addr, |
ysr@777 | 74 | HeapWord* limit = NULL) const; |
ysr@777 | 75 | |
ysr@777 | 76 | // conversion utilities |
ysr@777 | 77 | // XXX Fix these so that offsets are size_t's... |
ysr@777 | 78 | HeapWord* offsetToHeapWord(size_t offset) const { |
ysr@777 | 79 | return _bmStartWord + (offset << _shifter); |
ysr@777 | 80 | } |
ysr@777 | 81 | size_t heapWordToOffset(HeapWord* addr) const { |
ysr@777 | 82 | return pointer_delta(addr, _bmStartWord) >> _shifter; |
ysr@777 | 83 | } |
ysr@777 | 84 | int heapWordDiffToOffsetDiff(size_t diff) const; |
ysr@777 | 85 | HeapWord* nextWord(HeapWord* addr) { |
ysr@777 | 86 | return offsetToHeapWord(heapWordToOffset(addr) + 1); |
ysr@777 | 87 | } |
ysr@777 | 88 | |
ysr@777 | 89 | void mostly_disjoint_range_union(BitMap* from_bitmap, |
ysr@777 | 90 | size_t from_start_index, |
ysr@777 | 91 | HeapWord* to_start_word, |
ysr@777 | 92 | size_t word_num); |
ysr@777 | 93 | |
ysr@777 | 94 | // debugging |
ysr@777 | 95 | NOT_PRODUCT(bool covers(ReservedSpace rs) const;) |
ysr@777 | 96 | }; |
ysr@777 | 97 | |
ysr@777 | 98 | class CMBitMap : public CMBitMapRO { |
ysr@777 | 99 | |
ysr@777 | 100 | public: |
ysr@777 | 101 | // constructor |
ysr@777 | 102 | CMBitMap(ReservedSpace rs, int shifter) : |
ysr@777 | 103 | CMBitMapRO(rs, shifter) {} |
ysr@777 | 104 | |
ysr@777 | 105 | // write marks |
ysr@777 | 106 | void mark(HeapWord* addr) { |
ysr@777 | 107 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 108 | "outside underlying space?"); |
ysr@777 | 109 | _bm.at_put(heapWordToOffset(addr), true); |
ysr@777 | 110 | } |
ysr@777 | 111 | void clear(HeapWord* addr) { |
ysr@777 | 112 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 113 | "outside underlying space?"); |
ysr@777 | 114 | _bm.at_put(heapWordToOffset(addr), false); |
ysr@777 | 115 | } |
ysr@777 | 116 | bool parMark(HeapWord* addr) { |
ysr@777 | 117 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 118 | "outside underlying space?"); |
ysr@777 | 119 | return _bm.par_at_put(heapWordToOffset(addr), true); |
ysr@777 | 120 | } |
ysr@777 | 121 | bool parClear(HeapWord* addr) { |
ysr@777 | 122 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 123 | "outside underlying space?"); |
ysr@777 | 124 | return _bm.par_at_put(heapWordToOffset(addr), false); |
ysr@777 | 125 | } |
ysr@777 | 126 | void markRange(MemRegion mr); |
ysr@777 | 127 | void clearAll(); |
ysr@777 | 128 | void clearRange(MemRegion mr); |
ysr@777 | 129 | |
ysr@777 | 130 | // Starting at the bit corresponding to "addr" (inclusive), find the next |
ysr@777 | 131 | // "1" bit, if any. This bit starts some run of consecutive "1"'s; find |
ysr@777 | 132 | // the end of this run (stopping at "end_addr"). Return the MemRegion |
ysr@777 | 133 | // covering from the start of the region corresponding to the first bit |
ysr@777 | 134 | // of the run to the end of the region corresponding to the last bit of |
ysr@777 | 135 | // the run. If there is no "1" bit at or after "addr", return an empty |
ysr@777 | 136 | // MemRegion. |
ysr@777 | 137 | MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr); |
ysr@777 | 138 | }; |
ysr@777 | 139 | |
ysr@777 | 140 | // Represents a marking stack used by the CM collector. |
ysr@777 | 141 | // Ideally this should be GrowableArray<> just like MSC's marking stack(s). |
apetrusenko@984 | 142 | class CMMarkStack VALUE_OBJ_CLASS_SPEC { |
ysr@777 | 143 | ConcurrentMark* _cm; |
ysr@777 | 144 | oop* _base; // bottom of stack |
ysr@777 | 145 | jint _index; // one more than last occupied index |
ysr@777 | 146 | jint _capacity; // max #elements |
ysr@777 | 147 | jint _oops_do_bound; // Number of elements to include in next iteration. |
ysr@777 | 148 | NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run |
ysr@777 | 149 | |
ysr@777 | 150 | bool _overflow; |
ysr@777 | 151 | DEBUG_ONLY(bool _drain_in_progress;) |
ysr@777 | 152 | DEBUG_ONLY(bool _drain_in_progress_yields;) |
ysr@777 | 153 | |
ysr@777 | 154 | public: |
ysr@777 | 155 | CMMarkStack(ConcurrentMark* cm); |
ysr@777 | 156 | ~CMMarkStack(); |
ysr@777 | 157 | |
ysr@777 | 158 | void allocate(size_t size); |
ysr@777 | 159 | |
ysr@777 | 160 | oop pop() { |
ysr@777 | 161 | if (!isEmpty()) { |
ysr@777 | 162 | return _base[--_index] ; |
ysr@777 | 163 | } |
ysr@777 | 164 | return NULL; |
ysr@777 | 165 | } |
ysr@777 | 166 | |
ysr@777 | 167 | // If overflow happens, don't do the push, and record the overflow. |
ysr@777 | 168 | // *Requires* that "ptr" is already marked. |
ysr@777 | 169 | void push(oop ptr) { |
ysr@777 | 170 | if (isFull()) { |
ysr@777 | 171 | // Record overflow. |
ysr@777 | 172 | _overflow = true; |
ysr@777 | 173 | return; |
ysr@777 | 174 | } else { |
ysr@777 | 175 | _base[_index++] = ptr; |
ysr@777 | 176 | NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index)); |
ysr@777 | 177 | } |
ysr@777 | 178 | } |
ysr@777 | 179 | // Non-block impl. Note: concurrency is allowed only with other |
ysr@777 | 180 | // "par_push" operations, not with "pop" or "drain". We would need |
ysr@777 | 181 | // parallel versions of them if such concurrency was desired. |
ysr@777 | 182 | void par_push(oop ptr); |
ysr@777 | 183 | |
ysr@777 | 184 | // Pushes the first "n" elements of "ptr_arr" on the stack. |
ysr@777 | 185 | // Non-block impl. Note: concurrency is allowed only with other |
ysr@777 | 186 | // "par_adjoin_arr" or "push" operations, not with "pop" or "drain". |
ysr@777 | 187 | void par_adjoin_arr(oop* ptr_arr, int n); |
ysr@777 | 188 | |
ysr@777 | 189 | // Pushes the first "n" elements of "ptr_arr" on the stack. |
ysr@777 | 190 | // Locking impl: concurrency is allowed only with |
ysr@777 | 191 | // "par_push_arr" and/or "par_pop_arr" operations, which use the same |
ysr@777 | 192 | // locking strategy. |
ysr@777 | 193 | void par_push_arr(oop* ptr_arr, int n); |
ysr@777 | 194 | |
ysr@777 | 195 | // If returns false, the array was empty. Otherwise, removes up to "max" |
ysr@777 | 196 | // elements from the stack, and transfers them to "ptr_arr" in an |
ysr@777 | 197 | // unspecified order. The actual number transferred is given in "n" ("n |
ysr@777 | 198 | // == 0" is deliberately redundant with the return value.) Locking impl: |
ysr@777 | 199 | // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" |
ysr@777 | 200 | // operations, which use the same locking strategy. |
ysr@777 | 201 | bool par_pop_arr(oop* ptr_arr, int max, int* n); |
ysr@777 | 202 | |
ysr@777 | 203 | // Drain the mark stack, applying the given closure to all fields of |
ysr@777 | 204 | // objects on the stack. (That is, continue until the stack is empty, |
ysr@777 | 205 | // even if closure applications add entries to the stack.) The "bm" |
ysr@777 | 206 | // argument, if non-null, may be used to verify that only marked objects |
ysr@777 | 207 | // are on the mark stack. If "yield_after" is "true", then the |
ysr@777 | 208 | // concurrent marker performing the drain offers to yield after |
ysr@777 | 209 | // processing each object. If a yield occurs, stops the drain operation |
ysr@777 | 210 | // and returns false. Otherwise, returns true. |
ysr@777 | 211 | template<class OopClosureClass> |
ysr@777 | 212 | bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false); |
ysr@777 | 213 | |
ysr@777 | 214 | bool isEmpty() { return _index == 0; } |
ysr@777 | 215 | bool isFull() { return _index == _capacity; } |
ysr@777 | 216 | int maxElems() { return _capacity; } |
ysr@777 | 217 | |
ysr@777 | 218 | bool overflow() { return _overflow; } |
ysr@777 | 219 | void clear_overflow() { _overflow = false; } |
ysr@777 | 220 | |
ysr@777 | 221 | int size() { return _index; } |
ysr@777 | 222 | |
ysr@777 | 223 | void setEmpty() { _index = 0; clear_overflow(); } |
ysr@777 | 224 | |
ysr@777 | 225 | // Record the current size; a subsequent "oops_do" will iterate only over |
ysr@777 | 226 | // indices valid at the time of this call. |
ysr@777 | 227 | void set_oops_do_bound(jint bound = -1) { |
ysr@777 | 228 | if (bound == -1) { |
ysr@777 | 229 | _oops_do_bound = _index; |
ysr@777 | 230 | } else { |
ysr@777 | 231 | _oops_do_bound = bound; |
ysr@777 | 232 | } |
ysr@777 | 233 | } |
ysr@777 | 234 | jint oops_do_bound() { return _oops_do_bound; } |
ysr@777 | 235 | // iterate over the oops in the mark stack, up to the bound recorded via |
ysr@777 | 236 | // the call above. |
ysr@777 | 237 | void oops_do(OopClosure* f); |
ysr@777 | 238 | }; |
ysr@777 | 239 | |
apetrusenko@984 | 240 | class CMRegionStack VALUE_OBJ_CLASS_SPEC { |
ysr@777 | 241 | MemRegion* _base; |
ysr@777 | 242 | jint _capacity; |
ysr@777 | 243 | jint _index; |
ysr@777 | 244 | jint _oops_do_bound; |
ysr@777 | 245 | bool _overflow; |
ysr@777 | 246 | public: |
ysr@777 | 247 | CMRegionStack(); |
ysr@777 | 248 | ~CMRegionStack(); |
ysr@777 | 249 | void allocate(size_t size); |
ysr@777 | 250 | |
ysr@777 | 251 | // This is lock-free; assumes that it will only be called in parallel |
ysr@777 | 252 | // with other "push" operations (no pops). |
ysr@777 | 253 | void push(MemRegion mr); |
ysr@777 | 254 | |
tonyp@1793 | 255 | #if 0 |
tonyp@1793 | 256 | // This is currently not used. See the comment in the .cpp file. |
tonyp@1793 | 257 | |
ysr@777 | 258 | // Lock-free; assumes that it will only be called in parallel |
ysr@777 | 259 | // with other "pop" operations (no pushes). |
ysr@777 | 260 | MemRegion pop(); |
tonyp@1793 | 261 | #endif // 0 |
tonyp@1793 | 262 | |
tonyp@1793 | 263 | // These two are the implementations that use a lock. They can be |
tonyp@1793 | 264 | // called concurrently with each other but they should not be called |
tonyp@1793 | 265 | // concurrently with the lock-free versions (push() / pop()). |
tonyp@1793 | 266 | void push_with_lock(MemRegion mr); |
tonyp@1793 | 267 | MemRegion pop_with_lock(); |
ysr@777 | 268 | |
ysr@777 | 269 | bool isEmpty() { return _index == 0; } |
ysr@777 | 270 | bool isFull() { return _index == _capacity; } |
ysr@777 | 271 | |
ysr@777 | 272 | bool overflow() { return _overflow; } |
ysr@777 | 273 | void clear_overflow() { _overflow = false; } |
ysr@777 | 274 | |
ysr@777 | 275 | int size() { return _index; } |
ysr@777 | 276 | |
ysr@777 | 277 | // It iterates over the entries in the region stack and it |
ysr@777 | 278 | // invalidates (i.e. assigns MemRegion()) the ones that point to |
ysr@777 | 279 | // regions in the collection set. |
ysr@777 | 280 | bool invalidate_entries_into_cset(); |
ysr@777 | 281 | |
ysr@777 | 282 | // This gives an upper bound up to which the iteration in |
ysr@777 | 283 | // invalidate_entries_into_cset() will reach. This prevents |
ysr@777 | 284 | // newly-added entries to be unnecessarily scanned. |
ysr@777 | 285 | void set_oops_do_bound() { |
ysr@777 | 286 | _oops_do_bound = _index; |
ysr@777 | 287 | } |
ysr@777 | 288 | |
ysr@777 | 289 | void setEmpty() { _index = 0; clear_overflow(); } |
ysr@777 | 290 | }; |
ysr@777 | 291 | |
ysr@777 | 292 | // this will enable a variety of different statistics per GC task |
ysr@777 | 293 | #define _MARKING_STATS_ 0 |
ysr@777 | 294 | // this will enable the higher verbose levels |
ysr@777 | 295 | #define _MARKING_VERBOSE_ 0 |
ysr@777 | 296 | |
ysr@777 | 297 | #if _MARKING_STATS_ |
ysr@777 | 298 | #define statsOnly(statement) \ |
ysr@777 | 299 | do { \ |
ysr@777 | 300 | statement ; \ |
ysr@777 | 301 | } while (0) |
ysr@777 | 302 | #else // _MARKING_STATS_ |
ysr@777 | 303 | #define statsOnly(statement) \ |
ysr@777 | 304 | do { \ |
ysr@777 | 305 | } while (0) |
ysr@777 | 306 | #endif // _MARKING_STATS_ |
ysr@777 | 307 | |
ysr@777 | 308 | typedef enum { |
ysr@777 | 309 | no_verbose = 0, // verbose turned off |
ysr@777 | 310 | stats_verbose, // only prints stats at the end of marking |
ysr@777 | 311 | low_verbose, // low verbose, mostly per region and per major event |
ysr@777 | 312 | medium_verbose, // a bit more detailed than low |
ysr@777 | 313 | high_verbose // per object verbose |
ysr@777 | 314 | } CMVerboseLevel; |
ysr@777 | 315 | |
ysr@777 | 316 | |
ysr@777 | 317 | class ConcurrentMarkThread; |
ysr@777 | 318 | |
apetrusenko@984 | 319 | class ConcurrentMark: public CHeapObj { |
ysr@777 | 320 | friend class ConcurrentMarkThread; |
ysr@777 | 321 | friend class CMTask; |
ysr@777 | 322 | friend class CMBitMapClosure; |
ysr@777 | 323 | friend class CSMarkOopClosure; |
ysr@777 | 324 | friend class CMGlobalObjectClosure; |
ysr@777 | 325 | friend class CMRemarkTask; |
ysr@777 | 326 | friend class CMConcurrentMarkingTask; |
ysr@777 | 327 | friend class G1ParNoteEndTask; |
ysr@777 | 328 | friend class CalcLiveObjectsClosure; |
ysr@777 | 329 | |
ysr@777 | 330 | protected: |
ysr@777 | 331 | ConcurrentMarkThread* _cmThread; // the thread doing the work |
ysr@777 | 332 | G1CollectedHeap* _g1h; // the heap. |
ysr@777 | 333 | size_t _parallel_marking_threads; // the number of marking |
ysr@777 | 334 | // threads we'll use |
ysr@777 | 335 | double _sleep_factor; // how much we have to sleep, with |
ysr@777 | 336 | // respect to the work we just did, to |
ysr@777 | 337 | // meet the marking overhead goal |
ysr@777 | 338 | double _marking_task_overhead; // marking target overhead for |
ysr@777 | 339 | // a single task |
ysr@777 | 340 | |
ysr@777 | 341 | // same as the two above, but for the cleanup task |
ysr@777 | 342 | double _cleanup_sleep_factor; |
ysr@777 | 343 | double _cleanup_task_overhead; |
ysr@777 | 344 | |
ysr@777 | 345 | // Stuff related to age cohort processing. |
ysr@777 | 346 | struct ParCleanupThreadState { |
ysr@777 | 347 | char _pre[64]; |
ysr@777 | 348 | UncleanRegionList list; |
ysr@777 | 349 | char _post[64]; |
ysr@777 | 350 | }; |
ysr@777 | 351 | ParCleanupThreadState** _par_cleanup_thread_state; |
ysr@777 | 352 | |
ysr@777 | 353 | // CMS marking support structures |
ysr@777 | 354 | CMBitMap _markBitMap1; |
ysr@777 | 355 | CMBitMap _markBitMap2; |
ysr@777 | 356 | CMBitMapRO* _prevMarkBitMap; // completed mark bitmap |
ysr@777 | 357 | CMBitMap* _nextMarkBitMap; // under-construction mark bitmap |
ysr@777 | 358 | bool _at_least_one_mark_complete; |
ysr@777 | 359 | |
ysr@777 | 360 | BitMap _region_bm; |
ysr@777 | 361 | BitMap _card_bm; |
ysr@777 | 362 | |
ysr@777 | 363 | // Heap bounds |
ysr@777 | 364 | HeapWord* _heap_start; |
ysr@777 | 365 | HeapWord* _heap_end; |
ysr@777 | 366 | |
ysr@777 | 367 | // For gray objects |
ysr@777 | 368 | CMMarkStack _markStack; // Grey objects behind global finger. |
ysr@777 | 369 | CMRegionStack _regionStack; // Grey regions behind global finger. |
ysr@777 | 370 | HeapWord* volatile _finger; // the global finger, region aligned, |
ysr@777 | 371 | // always points to the end of the |
ysr@777 | 372 | // last claimed region |
ysr@777 | 373 | |
ysr@777 | 374 | // marking tasks |
ysr@777 | 375 | size_t _max_task_num; // maximum task number |
ysr@777 | 376 | size_t _active_tasks; // task num currently active |
ysr@777 | 377 | CMTask** _tasks; // task queue array (max_task_num len) |
ysr@777 | 378 | CMTaskQueueSet* _task_queues; // task queue set |
ysr@777 | 379 | ParallelTaskTerminator _terminator; // for termination |
ysr@777 | 380 | |
ysr@777 | 381 | // Two sync barriers that are used to synchronise tasks when an |
ysr@777 | 382 | // overflow occurs. The algorithm is the following. All tasks enter |
ysr@777 | 383 | // the first one to ensure that they have all stopped manipulating |
ysr@777 | 384 | // the global data structures. After they exit it, they re-initialise |
ysr@777 | 385 | // their data structures and task 0 re-initialises the global data |
ysr@777 | 386 | // structures. Then, they enter the second sync barrier. This |
ysr@777 | 387 | // ensure, that no task starts doing work before all data |
ysr@777 | 388 | // structures (local and global) have been re-initialised. When they |
ysr@777 | 389 | // exit it, they are free to start working again. |
ysr@777 | 390 | WorkGangBarrierSync _first_overflow_barrier_sync; |
ysr@777 | 391 | WorkGangBarrierSync _second_overflow_barrier_sync; |
ysr@777 | 392 | |
ysr@777 | 393 | |
ysr@777 | 394 | // this is set by any task, when an overflow on the global data |
ysr@777 | 395 | // structures is detected. |
ysr@777 | 396 | volatile bool _has_overflown; |
ysr@777 | 397 | // true: marking is concurrent, false: we're in remark |
ysr@777 | 398 | volatile bool _concurrent; |
ysr@777 | 399 | // set at the end of a Full GC so that marking aborts |
ysr@777 | 400 | volatile bool _has_aborted; |
ysr@777 | 401 | // used when remark aborts due to an overflow to indicate that |
ysr@777 | 402 | // another concurrent marking phase should start |
ysr@777 | 403 | volatile bool _restart_for_overflow; |
ysr@777 | 404 | |
ysr@777 | 405 | // This is true from the very start of concurrent marking until the |
ysr@777 | 406 | // point when all the tasks complete their work. It is really used |
ysr@777 | 407 | // to determine the points between the end of concurrent marking and |
ysr@777 | 408 | // time of remark. |
ysr@777 | 409 | volatile bool _concurrent_marking_in_progress; |
ysr@777 | 410 | |
ysr@777 | 411 | // verbose level |
ysr@777 | 412 | CMVerboseLevel _verbose_level; |
ysr@777 | 413 | |
ysr@777 | 414 | // These two fields are used to implement the optimisation that |
ysr@777 | 415 | // avoids pushing objects on the global/region stack if there are |
ysr@777 | 416 | // no collection set regions above the lowest finger. |
ysr@777 | 417 | |
ysr@777 | 418 | // This is the lowest finger (among the global and local fingers), |
ysr@777 | 419 | // which is calculated before a new collection set is chosen. |
ysr@777 | 420 | HeapWord* _min_finger; |
ysr@777 | 421 | // If this flag is true, objects/regions that are marked below the |
ysr@777 | 422 | // finger should be pushed on the stack(s). If this is flag is |
ysr@777 | 423 | // false, it is safe not to push them on the stack(s). |
ysr@777 | 424 | bool _should_gray_objects; |
ysr@777 | 425 | |
ysr@777 | 426 | // All of these times are in ms. |
ysr@777 | 427 | NumberSeq _init_times; |
ysr@777 | 428 | NumberSeq _remark_times; |
ysr@777 | 429 | NumberSeq _remark_mark_times; |
ysr@777 | 430 | NumberSeq _remark_weak_ref_times; |
ysr@777 | 431 | NumberSeq _cleanup_times; |
ysr@777 | 432 | double _total_counting_time; |
ysr@777 | 433 | double _total_rs_scrub_time; |
ysr@777 | 434 | |
ysr@777 | 435 | double* _accum_task_vtime; // accumulated task vtime |
ysr@777 | 436 | |
ysr@777 | 437 | WorkGang* _parallel_workers; |
ysr@777 | 438 | |
ysr@777 | 439 | void weakRefsWork(bool clear_all_soft_refs); |
ysr@777 | 440 | |
ysr@777 | 441 | void swapMarkBitMaps(); |
ysr@777 | 442 | |
ysr@777 | 443 | // It resets the global marking data structures, as well as the |
ysr@777 | 444 | // task local ones; should be called during initial mark. |
ysr@777 | 445 | void reset(); |
ysr@777 | 446 | // It resets all the marking data structures. |
ysr@777 | 447 | void clear_marking_state(); |
ysr@777 | 448 | |
ysr@777 | 449 | // It should be called to indicate which phase we're in (concurrent |
ysr@777 | 450 | // mark or remark) and how many threads are currently active. |
ysr@777 | 451 | void set_phase(size_t active_tasks, bool concurrent); |
ysr@777 | 452 | // We do this after we're done with marking so that the marking data |
ysr@777 | 453 | // structures are initialised to a sensible and predictable state. |
ysr@777 | 454 | void set_non_marking_state(); |
ysr@777 | 455 | |
ysr@777 | 456 | // prints all gathered CM-related statistics |
ysr@777 | 457 | void print_stats(); |
ysr@777 | 458 | |
ysr@777 | 459 | // accessor methods |
ysr@777 | 460 | size_t parallel_marking_threads() { return _parallel_marking_threads; } |
ysr@777 | 461 | double sleep_factor() { return _sleep_factor; } |
ysr@777 | 462 | double marking_task_overhead() { return _marking_task_overhead;} |
ysr@777 | 463 | double cleanup_sleep_factor() { return _cleanup_sleep_factor; } |
ysr@777 | 464 | double cleanup_task_overhead() { return _cleanup_task_overhead;} |
ysr@777 | 465 | |
ysr@777 | 466 | HeapWord* finger() { return _finger; } |
ysr@777 | 467 | bool concurrent() { return _concurrent; } |
ysr@777 | 468 | size_t active_tasks() { return _active_tasks; } |
ysr@777 | 469 | ParallelTaskTerminator* terminator() { return &_terminator; } |
ysr@777 | 470 | |
ysr@777 | 471 | // It claims the next available region to be scanned by a marking |
ysr@777 | 472 | // task. It might return NULL if the next region is empty or we have |
ysr@777 | 473 | // run out of regions. In the latter case, out_of_regions() |
ysr@777 | 474 | // determines whether we've really run out of regions or the task |
ysr@777 | 475 | // should call claim_region() again. This might seem a bit |
ysr@777 | 476 | // awkward. Originally, the code was written so that claim_region() |
ysr@777 | 477 | // either successfully returned with a non-empty region or there |
ysr@777 | 478 | // were no more regions to be claimed. The problem with this was |
ysr@777 | 479 | // that, in certain circumstances, it iterated over large chunks of |
ysr@777 | 480 | // the heap finding only empty regions and, while it was working, it |
ysr@777 | 481 | // was preventing the calling task to call its regular clock |
ysr@777 | 482 | // method. So, this way, each task will spend very little time in |
ysr@777 | 483 | // claim_region() and is allowed to call the regular clock method |
ysr@777 | 484 | // frequently. |
ysr@777 | 485 | HeapRegion* claim_region(int task); |
ysr@777 | 486 | |
ysr@777 | 487 | // It determines whether we've run out of regions to scan. |
ysr@777 | 488 | bool out_of_regions() { return _finger == _heap_end; } |
ysr@777 | 489 | |
ysr@777 | 490 | // Returns the task with the given id |
ysr@777 | 491 | CMTask* task(int id) { |
tonyp@1458 | 492 | assert(0 <= id && id < (int) _active_tasks, |
tonyp@1458 | 493 | "task id not within active bounds"); |
ysr@777 | 494 | return _tasks[id]; |
ysr@777 | 495 | } |
ysr@777 | 496 | |
ysr@777 | 497 | // Returns the task queue with the given id |
ysr@777 | 498 | CMTaskQueue* task_queue(int id) { |
tonyp@1458 | 499 | assert(0 <= id && id < (int) _active_tasks, |
tonyp@1458 | 500 | "task queue id not within active bounds"); |
ysr@777 | 501 | return (CMTaskQueue*) _task_queues->queue(id); |
ysr@777 | 502 | } |
ysr@777 | 503 | |
ysr@777 | 504 | // Returns the task queue set |
ysr@777 | 505 | CMTaskQueueSet* task_queues() { return _task_queues; } |
ysr@777 | 506 | |
ysr@777 | 507 | // Access / manipulation of the overflow flag which is set to |
ysr@777 | 508 | // indicate that the global stack or region stack has overflown |
ysr@777 | 509 | bool has_overflown() { return _has_overflown; } |
ysr@777 | 510 | void set_has_overflown() { _has_overflown = true; } |
ysr@777 | 511 | void clear_has_overflown() { _has_overflown = false; } |
ysr@777 | 512 | |
ysr@777 | 513 | bool has_aborted() { return _has_aborted; } |
ysr@777 | 514 | bool restart_for_overflow() { return _restart_for_overflow; } |
ysr@777 | 515 | |
ysr@777 | 516 | // Methods to enter the two overflow sync barriers |
ysr@777 | 517 | void enter_first_sync_barrier(int task_num); |
ysr@777 | 518 | void enter_second_sync_barrier(int task_num); |
ysr@777 | 519 | |
ysr@777 | 520 | public: |
ysr@777 | 521 | // Manipulation of the global mark stack. |
ysr@777 | 522 | // Notice that the first mark_stack_push is CAS-based, whereas the |
ysr@777 | 523 | // two below are Mutex-based. This is OK since the first one is only |
ysr@777 | 524 | // called during evacuation pauses and doesn't compete with the |
ysr@777 | 525 | // other two (which are called by the marking tasks during |
ysr@777 | 526 | // concurrent marking or remark). |
ysr@777 | 527 | bool mark_stack_push(oop p) { |
ysr@777 | 528 | _markStack.par_push(p); |
ysr@777 | 529 | if (_markStack.overflow()) { |
ysr@777 | 530 | set_has_overflown(); |
ysr@777 | 531 | return false; |
ysr@777 | 532 | } |
ysr@777 | 533 | return true; |
ysr@777 | 534 | } |
ysr@777 | 535 | bool mark_stack_push(oop* arr, int n) { |
ysr@777 | 536 | _markStack.par_push_arr(arr, n); |
ysr@777 | 537 | if (_markStack.overflow()) { |
ysr@777 | 538 | set_has_overflown(); |
ysr@777 | 539 | return false; |
ysr@777 | 540 | } |
ysr@777 | 541 | return true; |
ysr@777 | 542 | } |
ysr@777 | 543 | void mark_stack_pop(oop* arr, int max, int* n) { |
ysr@777 | 544 | _markStack.par_pop_arr(arr, max, n); |
ysr@777 | 545 | } |
ysr@777 | 546 | size_t mark_stack_size() { return _markStack.size(); } |
ysr@777 | 547 | size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } |
ysr@777 | 548 | bool mark_stack_overflow() { return _markStack.overflow(); } |
ysr@777 | 549 | bool mark_stack_empty() { return _markStack.isEmpty(); } |
ysr@777 | 550 | |
ysr@777 | 551 | // Manipulation of the region stack |
ysr@777 | 552 | bool region_stack_push(MemRegion mr) { |
tonyp@1793 | 553 | // Currently we only call the lock-free version during evacuation |
tonyp@1793 | 554 | // pauses. |
tonyp@1793 | 555 | assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); |
tonyp@1793 | 556 | |
ysr@777 | 557 | _regionStack.push(mr); |
ysr@777 | 558 | if (_regionStack.overflow()) { |
ysr@777 | 559 | set_has_overflown(); |
ysr@777 | 560 | return false; |
ysr@777 | 561 | } |
ysr@777 | 562 | return true; |
ysr@777 | 563 | } |
tonyp@1793 | 564 | #if 0 |
tonyp@1793 | 565 | // Currently this is not used. See the comment in the .cpp file. |
tonyp@1793 | 566 | MemRegion region_stack_pop() { return _regionStack.pop(); } |
tonyp@1793 | 567 | #endif // 0 |
tonyp@1793 | 568 | |
tonyp@1793 | 569 | bool region_stack_push_with_lock(MemRegion mr) { |
tonyp@1793 | 570 | // Currently we only call the lock-based version during either |
tonyp@1793 | 571 | // concurrent marking or remark. |
tonyp@1793 | 572 | assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(), |
tonyp@1793 | 573 | "if we are at a safepoint it should be the remark safepoint"); |
tonyp@1793 | 574 | |
tonyp@1793 | 575 | _regionStack.push_with_lock(mr); |
tonyp@1793 | 576 | if (_regionStack.overflow()) { |
tonyp@1793 | 577 | set_has_overflown(); |
tonyp@1793 | 578 | return false; |
tonyp@1793 | 579 | } |
tonyp@1793 | 580 | return true; |
tonyp@1793 | 581 | } |
tonyp@1793 | 582 | MemRegion region_stack_pop_with_lock() { |
tonyp@1793 | 583 | // Currently we only call the lock-based version during either |
tonyp@1793 | 584 | // concurrent marking or remark. |
tonyp@1793 | 585 | assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(), |
tonyp@1793 | 586 | "if we are at a safepoint it should be the remark safepoint"); |
tonyp@1793 | 587 | |
tonyp@1793 | 588 | return _regionStack.pop_with_lock(); |
tonyp@1793 | 589 | } |
tonyp@1793 | 590 | |
ysr@777 | 591 | int region_stack_size() { return _regionStack.size(); } |
ysr@777 | 592 | bool region_stack_overflow() { return _regionStack.overflow(); } |
ysr@777 | 593 | bool region_stack_empty() { return _regionStack.isEmpty(); } |
ysr@777 | 594 | |
ysr@777 | 595 | bool concurrent_marking_in_progress() { |
ysr@777 | 596 | return _concurrent_marking_in_progress; |
ysr@777 | 597 | } |
ysr@777 | 598 | void set_concurrent_marking_in_progress() { |
ysr@777 | 599 | _concurrent_marking_in_progress = true; |
ysr@777 | 600 | } |
ysr@777 | 601 | void clear_concurrent_marking_in_progress() { |
ysr@777 | 602 | _concurrent_marking_in_progress = false; |
ysr@777 | 603 | } |
ysr@777 | 604 | |
ysr@777 | 605 | void update_accum_task_vtime(int i, double vtime) { |
ysr@777 | 606 | _accum_task_vtime[i] += vtime; |
ysr@777 | 607 | } |
ysr@777 | 608 | |
ysr@777 | 609 | double all_task_accum_vtime() { |
ysr@777 | 610 | double ret = 0.0; |
ysr@777 | 611 | for (int i = 0; i < (int)_max_task_num; ++i) |
ysr@777 | 612 | ret += _accum_task_vtime[i]; |
ysr@777 | 613 | return ret; |
ysr@777 | 614 | } |
ysr@777 | 615 | |
ysr@777 | 616 | // Attempts to steal an object from the task queues of other tasks |
ysr@777 | 617 | bool try_stealing(int task_num, int* hash_seed, oop& obj) { |
ysr@777 | 618 | return _task_queues->steal(task_num, hash_seed, obj); |
ysr@777 | 619 | } |
ysr@777 | 620 | |
ysr@777 | 621 | // It grays an object by first marking it. Then, if it's behind the |
ysr@777 | 622 | // global finger, it also pushes it on the global stack. |
ysr@777 | 623 | void deal_with_reference(oop obj); |
ysr@777 | 624 | |
ysr@777 | 625 | ConcurrentMark(ReservedSpace rs, int max_regions); |
ysr@777 | 626 | ~ConcurrentMark(); |
ysr@777 | 627 | ConcurrentMarkThread* cmThread() { return _cmThread; } |
ysr@777 | 628 | |
ysr@777 | 629 | CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } |
ysr@777 | 630 | CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } |
ysr@777 | 631 | |
ysr@777 | 632 | // The following three are interaction between CM and |
ysr@777 | 633 | // G1CollectedHeap |
ysr@777 | 634 | |
ysr@777 | 635 | // This notifies CM that a root during initial-mark needs to be |
ysr@777 | 636 | // grayed and it's MT-safe. Currently, we just mark it. But, in the |
ysr@777 | 637 | // future, we can experiment with pushing it on the stack and we can |
ysr@777 | 638 | // do this without changing G1CollectedHeap. |
ysr@777 | 639 | void grayRoot(oop p); |
ysr@777 | 640 | // It's used during evacuation pauses to gray a region, if |
ysr@777 | 641 | // necessary, and it's MT-safe. It assumes that the caller has |
ysr@777 | 642 | // marked any objects on that region. If _should_gray_objects is |
ysr@777 | 643 | // true and we're still doing concurrent marking, the region is |
ysr@777 | 644 | // pushed on the region stack, if it is located below the global |
ysr@777 | 645 | // finger, otherwise we do nothing. |
ysr@777 | 646 | void grayRegionIfNecessary(MemRegion mr); |
ysr@777 | 647 | // It's used during evacuation pauses to mark and, if necessary, |
ysr@777 | 648 | // gray a single object and it's MT-safe. It assumes the caller did |
ysr@777 | 649 | // not mark the object. If _should_gray_objects is true and we're |
ysr@777 | 650 | // still doing concurrent marking, the objects is pushed on the |
ysr@777 | 651 | // global stack, if it is located below the global finger, otherwise |
ysr@777 | 652 | // we do nothing. |
ysr@777 | 653 | void markAndGrayObjectIfNecessary(oop p); |
ysr@777 | 654 | |
tonyp@1479 | 655 | // This iterates over the marking bitmap (either prev or next) and |
tonyp@1479 | 656 | // prints out all objects that are marked on the bitmap and indicates |
tonyp@1479 | 657 | // whether what they point to is also marked or not. It also iterates |
tonyp@1479 | 658 | // the objects over TAMS (either prev or next). |
tonyp@1479 | 659 | void print_reachable(bool use_prev_marking, const char* str); |
ysr@777 | 660 | |
ysr@777 | 661 | // Clear the next marking bitmap (will be called concurrently). |
ysr@777 | 662 | void clearNextBitmap(); |
ysr@777 | 663 | |
ysr@777 | 664 | // main CMS steps and related support |
ysr@777 | 665 | void checkpointRootsInitial(); |
ysr@777 | 666 | |
ysr@777 | 667 | // These two do the work that needs to be done before and after the |
ysr@777 | 668 | // initial root checkpoint. Since this checkpoint can be done at two |
ysr@777 | 669 | // different points (i.e. an explicit pause or piggy-backed on a |
ysr@777 | 670 | // young collection), then it's nice to be able to easily share the |
ysr@777 | 671 | // pre/post code. It might be the case that we can put everything in |
ysr@777 | 672 | // the post method. TP |
ysr@777 | 673 | void checkpointRootsInitialPre(); |
ysr@777 | 674 | void checkpointRootsInitialPost(); |
ysr@777 | 675 | |
ysr@777 | 676 | // Do concurrent phase of marking, to a tentative transitive closure. |
ysr@777 | 677 | void markFromRoots(); |
ysr@777 | 678 | |
ysr@777 | 679 | // Process all unprocessed SATB buffers. It is called at the |
ysr@777 | 680 | // beginning of an evacuation pause. |
ysr@777 | 681 | void drainAllSATBBuffers(); |
ysr@777 | 682 | |
ysr@777 | 683 | void checkpointRootsFinal(bool clear_all_soft_refs); |
ysr@777 | 684 | void checkpointRootsFinalWork(); |
ysr@777 | 685 | void calcDesiredRegions(); |
ysr@777 | 686 | void cleanup(); |
ysr@777 | 687 | void completeCleanup(); |
ysr@777 | 688 | |
ysr@777 | 689 | // Mark in the previous bitmap. NB: this is usually read-only, so use |
ysr@777 | 690 | // this carefully! |
ysr@777 | 691 | void markPrev(oop p); |
ysr@777 | 692 | void clear(oop p); |
ysr@777 | 693 | // Clears marks for all objects in the given range, for both prev and |
ysr@777 | 694 | // next bitmaps. NB: the previous bitmap is usually read-only, so use |
ysr@777 | 695 | // this carefully! |
ysr@777 | 696 | void clearRangeBothMaps(MemRegion mr); |
ysr@777 | 697 | |
ysr@777 | 698 | // Record the current top of the mark and region stacks; a |
ysr@777 | 699 | // subsequent oops_do() on the mark stack and |
ysr@777 | 700 | // invalidate_entries_into_cset() on the region stack will iterate |
ysr@777 | 701 | // only over indices valid at the time of this call. |
ysr@777 | 702 | void set_oops_do_bound() { |
ysr@777 | 703 | _markStack.set_oops_do_bound(); |
ysr@777 | 704 | _regionStack.set_oops_do_bound(); |
ysr@777 | 705 | } |
ysr@777 | 706 | // Iterate over the oops in the mark stack and all local queues. It |
ysr@777 | 707 | // also calls invalidate_entries_into_cset() on the region stack. |
ysr@777 | 708 | void oops_do(OopClosure* f); |
ysr@777 | 709 | // It is called at the end of an evacuation pause during marking so |
ysr@777 | 710 | // that CM is notified of where the new end of the heap is. It |
ysr@777 | 711 | // doesn't do anything if concurrent_marking_in_progress() is false, |
ysr@777 | 712 | // unless the force parameter is true. |
ysr@777 | 713 | void update_g1_committed(bool force = false); |
ysr@777 | 714 | |
ysr@777 | 715 | void complete_marking_in_collection_set(); |
ysr@777 | 716 | |
ysr@777 | 717 | // It indicates that a new collection set is being chosen. |
ysr@777 | 718 | void newCSet(); |
ysr@777 | 719 | // It registers a collection set heap region with CM. This is used |
ysr@777 | 720 | // to determine whether any heap regions are located above the finger. |
ysr@777 | 721 | void registerCSetRegion(HeapRegion* hr); |
ysr@777 | 722 | |
ysr@777 | 723 | // Returns "true" if at least one mark has been completed. |
ysr@777 | 724 | bool at_least_one_mark_complete() { return _at_least_one_mark_complete; } |
ysr@777 | 725 | |
ysr@777 | 726 | bool isMarked(oop p) const { |
ysr@777 | 727 | assert(p != NULL && p->is_oop(), "expected an oop"); |
ysr@777 | 728 | HeapWord* addr = (HeapWord*)p; |
ysr@777 | 729 | assert(addr >= _nextMarkBitMap->startWord() || |
ysr@777 | 730 | addr < _nextMarkBitMap->endWord(), "in a region"); |
ysr@777 | 731 | |
ysr@777 | 732 | return _nextMarkBitMap->isMarked(addr); |
ysr@777 | 733 | } |
ysr@777 | 734 | |
ysr@777 | 735 | inline bool not_yet_marked(oop p) const; |
ysr@777 | 736 | |
ysr@777 | 737 | // XXX Debug code |
ysr@777 | 738 | bool containing_card_is_marked(void* p); |
ysr@777 | 739 | bool containing_cards_are_marked(void* start, void* last); |
ysr@777 | 740 | |
ysr@777 | 741 | bool isPrevMarked(oop p) const { |
ysr@777 | 742 | assert(p != NULL && p->is_oop(), "expected an oop"); |
ysr@777 | 743 | HeapWord* addr = (HeapWord*)p; |
ysr@777 | 744 | assert(addr >= _prevMarkBitMap->startWord() || |
ysr@777 | 745 | addr < _prevMarkBitMap->endWord(), "in a region"); |
ysr@777 | 746 | |
ysr@777 | 747 | return _prevMarkBitMap->isMarked(addr); |
ysr@777 | 748 | } |
ysr@777 | 749 | |
ysr@777 | 750 | inline bool do_yield_check(int worker_i = 0); |
ysr@777 | 751 | inline bool should_yield(); |
ysr@777 | 752 | |
ysr@777 | 753 | // Called to abort the marking cycle after a Full GC takes palce. |
ysr@777 | 754 | void abort(); |
ysr@777 | 755 | |
ysr@777 | 756 | // This prints the global/local fingers. It is used for debugging. |
ysr@777 | 757 | NOT_PRODUCT(void print_finger();) |
ysr@777 | 758 | |
ysr@777 | 759 | void print_summary_info(); |
ysr@777 | 760 | |
tonyp@1454 | 761 | void print_worker_threads_on(outputStream* st) const; |
tonyp@1454 | 762 | |
ysr@777 | 763 | // The following indicate whether a given verbose level has been |
ysr@777 | 764 | // set. Notice that anything above stats is conditional to |
ysr@777 | 765 | // _MARKING_VERBOSE_ having been set to 1 |
ysr@777 | 766 | bool verbose_stats() |
ysr@777 | 767 | { return _verbose_level >= stats_verbose; } |
ysr@777 | 768 | bool verbose_low() |
ysr@777 | 769 | { return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; } |
ysr@777 | 770 | bool verbose_medium() |
ysr@777 | 771 | { return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; } |
ysr@777 | 772 | bool verbose_high() |
ysr@777 | 773 | { return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; } |
ysr@777 | 774 | }; |
ysr@777 | 775 | |
ysr@777 | 776 | // A class representing a marking task. |
ysr@777 | 777 | class CMTask : public TerminatorTerminator { |
ysr@777 | 778 | private: |
ysr@777 | 779 | enum PrivateConstants { |
ysr@777 | 780 | // the regular clock call is called once the scanned words reaches |
ysr@777 | 781 | // this limit |
ysr@777 | 782 | words_scanned_period = 12*1024, |
ysr@777 | 783 | // the regular clock call is called once the number of visited |
ysr@777 | 784 | // references reaches this limit |
ysr@777 | 785 | refs_reached_period = 384, |
ysr@777 | 786 | // initial value for the hash seed, used in the work stealing code |
ysr@777 | 787 | init_hash_seed = 17, |
ysr@777 | 788 | // how many entries will be transferred between global stack and |
ysr@777 | 789 | // local queues |
ysr@777 | 790 | global_stack_transfer_size = 16 |
ysr@777 | 791 | }; |
ysr@777 | 792 | |
ysr@777 | 793 | int _task_id; |
ysr@777 | 794 | G1CollectedHeap* _g1h; |
ysr@777 | 795 | ConcurrentMark* _cm; |
ysr@777 | 796 | CMBitMap* _nextMarkBitMap; |
ysr@777 | 797 | // the task queue of this task |
ysr@777 | 798 | CMTaskQueue* _task_queue; |
ysr@1280 | 799 | private: |
ysr@777 | 800 | // the task queue set---needed for stealing |
ysr@777 | 801 | CMTaskQueueSet* _task_queues; |
ysr@777 | 802 | // indicates whether the task has been claimed---this is only for |
ysr@777 | 803 | // debugging purposes |
ysr@777 | 804 | bool _claimed; |
ysr@777 | 805 | |
ysr@777 | 806 | // number of calls to this task |
ysr@777 | 807 | int _calls; |
ysr@777 | 808 | |
ysr@777 | 809 | // when the virtual timer reaches this time, the marking step should |
ysr@777 | 810 | // exit |
ysr@777 | 811 | double _time_target_ms; |
ysr@777 | 812 | // the start time of the current marking step |
ysr@777 | 813 | double _start_time_ms; |
ysr@777 | 814 | |
ysr@777 | 815 | // the oop closure used for iterations over oops |
ysr@777 | 816 | OopClosure* _oop_closure; |
ysr@777 | 817 | |
ysr@777 | 818 | // the region this task is scanning, NULL if we're not scanning any |
ysr@777 | 819 | HeapRegion* _curr_region; |
ysr@777 | 820 | // the local finger of this task, NULL if we're not scanning a region |
ysr@777 | 821 | HeapWord* _finger; |
ysr@777 | 822 | // limit of the region this task is scanning, NULL if we're not scanning one |
ysr@777 | 823 | HeapWord* _region_limit; |
ysr@777 | 824 | |
ysr@777 | 825 | // This is used only when we scan regions popped from the region |
ysr@777 | 826 | // stack. It records what the last object on such a region we |
ysr@777 | 827 | // scanned was. It is used to ensure that, if we abort region |
ysr@777 | 828 | // iteration, we do not rescan the first part of the region. This |
ysr@777 | 829 | // should be NULL when we're not scanning a region from the region |
ysr@777 | 830 | // stack. |
ysr@777 | 831 | HeapWord* _region_finger; |
ysr@777 | 832 | |
ysr@777 | 833 | // the number of words this task has scanned |
ysr@777 | 834 | size_t _words_scanned; |
ysr@777 | 835 | // When _words_scanned reaches this limit, the regular clock is |
ysr@777 | 836 | // called. Notice that this might be decreased under certain |
ysr@777 | 837 | // circumstances (i.e. when we believe that we did an expensive |
ysr@777 | 838 | // operation). |
ysr@777 | 839 | size_t _words_scanned_limit; |
ysr@777 | 840 | // the initial value of _words_scanned_limit (i.e. what it was |
ysr@777 | 841 | // before it was decreased). |
ysr@777 | 842 | size_t _real_words_scanned_limit; |
ysr@777 | 843 | |
ysr@777 | 844 | // the number of references this task has visited |
ysr@777 | 845 | size_t _refs_reached; |
ysr@777 | 846 | // When _refs_reached reaches this limit, the regular clock is |
ysr@777 | 847 | // called. Notice this this might be decreased under certain |
ysr@777 | 848 | // circumstances (i.e. when we believe that we did an expensive |
ysr@777 | 849 | // operation). |
ysr@777 | 850 | size_t _refs_reached_limit; |
ysr@777 | 851 | // the initial value of _refs_reached_limit (i.e. what it was before |
ysr@777 | 852 | // it was decreased). |
ysr@777 | 853 | size_t _real_refs_reached_limit; |
ysr@777 | 854 | |
ysr@777 | 855 | // used by the work stealing stuff |
ysr@777 | 856 | int _hash_seed; |
ysr@777 | 857 | // if this is true, then the task has aborted for some reason |
ysr@777 | 858 | bool _has_aborted; |
ysr@777 | 859 | // set when the task aborts because it has met its time quota |
ysr@777 | 860 | bool _has_aborted_timed_out; |
ysr@777 | 861 | // true when we're draining SATB buffers; this avoids the task |
ysr@777 | 862 | // aborting due to SATB buffers being available (as we're already |
ysr@777 | 863 | // dealing with them) |
ysr@777 | 864 | bool _draining_satb_buffers; |
ysr@777 | 865 | |
ysr@777 | 866 | // number sequence of past step times |
ysr@777 | 867 | NumberSeq _step_times_ms; |
ysr@777 | 868 | // elapsed time of this task |
ysr@777 | 869 | double _elapsed_time_ms; |
ysr@777 | 870 | // termination time of this task |
ysr@777 | 871 | double _termination_time_ms; |
ysr@777 | 872 | // when this task got into the termination protocol |
ysr@777 | 873 | double _termination_start_time_ms; |
ysr@777 | 874 | |
ysr@777 | 875 | // true when the task is during a concurrent phase, false when it is |
ysr@777 | 876 | // in the remark phase (so, in the latter case, we do not have to |
ysr@777 | 877 | // check all the things that we have to check during the concurrent |
ysr@777 | 878 | // phase, i.e. SATB buffer availability...) |
ysr@777 | 879 | bool _concurrent; |
ysr@777 | 880 | |
ysr@777 | 881 | TruncatedSeq _marking_step_diffs_ms; |
ysr@777 | 882 | |
ysr@777 | 883 | // LOTS of statistics related with this task |
ysr@777 | 884 | #if _MARKING_STATS_ |
ysr@777 | 885 | NumberSeq _all_clock_intervals_ms; |
ysr@777 | 886 | double _interval_start_time_ms; |
ysr@777 | 887 | |
ysr@777 | 888 | int _aborted; |
ysr@777 | 889 | int _aborted_overflow; |
ysr@777 | 890 | int _aborted_cm_aborted; |
ysr@777 | 891 | int _aborted_yield; |
ysr@777 | 892 | int _aborted_timed_out; |
ysr@777 | 893 | int _aborted_satb; |
ysr@777 | 894 | int _aborted_termination; |
ysr@777 | 895 | |
ysr@777 | 896 | int _steal_attempts; |
ysr@777 | 897 | int _steals; |
ysr@777 | 898 | |
ysr@777 | 899 | int _clock_due_to_marking; |
ysr@777 | 900 | int _clock_due_to_scanning; |
ysr@777 | 901 | |
ysr@777 | 902 | int _local_pushes; |
ysr@777 | 903 | int _local_pops; |
ysr@777 | 904 | int _local_max_size; |
ysr@777 | 905 | int _objs_scanned; |
ysr@777 | 906 | |
ysr@777 | 907 | int _global_pushes; |
ysr@777 | 908 | int _global_pops; |
ysr@777 | 909 | int _global_max_size; |
ysr@777 | 910 | |
ysr@777 | 911 | int _global_transfers_to; |
ysr@777 | 912 | int _global_transfers_from; |
ysr@777 | 913 | |
ysr@777 | 914 | int _region_stack_pops; |
ysr@777 | 915 | |
ysr@777 | 916 | int _regions_claimed; |
ysr@777 | 917 | int _objs_found_on_bitmap; |
ysr@777 | 918 | |
ysr@777 | 919 | int _satb_buffers_processed; |
ysr@777 | 920 | #endif // _MARKING_STATS_ |
ysr@777 | 921 | |
ysr@777 | 922 | // it updates the local fields after this task has claimed |
ysr@777 | 923 | // a new region to scan |
ysr@777 | 924 | void setup_for_region(HeapRegion* hr); |
ysr@777 | 925 | // it brings up-to-date the limit of the region |
ysr@777 | 926 | void update_region_limit(); |
ysr@777 | 927 | // it resets the local fields after a task has finished scanning a |
ysr@777 | 928 | // region |
ysr@777 | 929 | void giveup_current_region(); |
ysr@777 | 930 | |
ysr@777 | 931 | // called when either the words scanned or the refs visited limit |
ysr@777 | 932 | // has been reached |
ysr@777 | 933 | void reached_limit(); |
ysr@777 | 934 | // recalculates the words scanned and refs visited limits |
ysr@777 | 935 | void recalculate_limits(); |
ysr@777 | 936 | // decreases the words scanned and refs visited limits when we reach |
ysr@777 | 937 | // an expensive operation |
ysr@777 | 938 | void decrease_limits(); |
ysr@777 | 939 | // it checks whether the words scanned or refs visited reached their |
ysr@777 | 940 | // respective limit and calls reached_limit() if they have |
ysr@777 | 941 | void check_limits() { |
ysr@777 | 942 | if (_words_scanned >= _words_scanned_limit || |
ysr@777 | 943 | _refs_reached >= _refs_reached_limit) |
ysr@777 | 944 | reached_limit(); |
ysr@777 | 945 | } |
ysr@777 | 946 | // this is supposed to be called regularly during a marking step as |
ysr@777 | 947 | // it checks a bunch of conditions that might cause the marking step |
ysr@777 | 948 | // to abort |
ysr@777 | 949 | void regular_clock_call(); |
ysr@777 | 950 | bool concurrent() { return _concurrent; } |
ysr@777 | 951 | |
ysr@777 | 952 | public: |
ysr@777 | 953 | // It resets the task; it should be called right at the beginning of |
ysr@777 | 954 | // a marking phase. |
ysr@777 | 955 | void reset(CMBitMap* _nextMarkBitMap); |
ysr@777 | 956 | // it clears all the fields that correspond to a claimed region. |
ysr@777 | 957 | void clear_region_fields(); |
ysr@777 | 958 | |
ysr@777 | 959 | void set_concurrent(bool concurrent) { _concurrent = concurrent; } |
ysr@777 | 960 | |
ysr@777 | 961 | // The main method of this class which performs a marking step |
ysr@777 | 962 | // trying not to exceed the given duration. However, it might exit |
ysr@777 | 963 | // prematurely, according to some conditions (i.e. SATB buffers are |
ysr@777 | 964 | // available for processing). |
ysr@777 | 965 | void do_marking_step(double target_ms); |
ysr@777 | 966 | |
ysr@777 | 967 | // These two calls start and stop the timer |
ysr@777 | 968 | void record_start_time() { |
ysr@777 | 969 | _elapsed_time_ms = os::elapsedTime() * 1000.0; |
ysr@777 | 970 | } |
ysr@777 | 971 | void record_end_time() { |
ysr@777 | 972 | _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; |
ysr@777 | 973 | } |
ysr@777 | 974 | |
ysr@777 | 975 | // returns the task ID |
ysr@777 | 976 | int task_id() { return _task_id; } |
ysr@777 | 977 | |
ysr@777 | 978 | // From TerminatorTerminator. It determines whether this task should |
ysr@777 | 979 | // exit the termination protocol after it's entered it. |
ysr@777 | 980 | virtual bool should_exit_termination(); |
ysr@777 | 981 | |
ysr@777 | 982 | HeapWord* finger() { return _finger; } |
ysr@777 | 983 | |
ysr@777 | 984 | bool has_aborted() { return _has_aborted; } |
ysr@777 | 985 | void set_has_aborted() { _has_aborted = true; } |
ysr@777 | 986 | void clear_has_aborted() { _has_aborted = false; } |
ysr@777 | 987 | bool claimed() { return _claimed; } |
ysr@777 | 988 | |
ysr@777 | 989 | void set_oop_closure(OopClosure* oop_closure) { |
ysr@777 | 990 | _oop_closure = oop_closure; |
ysr@777 | 991 | } |
ysr@777 | 992 | |
ysr@777 | 993 | // It grays the object by marking it and, if necessary, pushing it |
ysr@777 | 994 | // on the local queue |
ysr@777 | 995 | void deal_with_reference(oop obj); |
ysr@777 | 996 | |
ysr@777 | 997 | // It scans an object and visits its children. |
ysr@777 | 998 | void scan_object(oop obj) { |
tonyp@1458 | 999 | assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant"); |
ysr@777 | 1000 | |
ysr@777 | 1001 | if (_cm->verbose_high()) |
ysr@777 | 1002 | gclog_or_tty->print_cr("[%d] we're scanning object "PTR_FORMAT, |
ysr@777 | 1003 | _task_id, (void*) obj); |
ysr@777 | 1004 | |
ysr@777 | 1005 | size_t obj_size = obj->size(); |
ysr@777 | 1006 | _words_scanned += obj_size; |
ysr@777 | 1007 | |
ysr@777 | 1008 | obj->oop_iterate(_oop_closure); |
ysr@777 | 1009 | statsOnly( ++_objs_scanned ); |
ysr@777 | 1010 | check_limits(); |
ysr@777 | 1011 | } |
ysr@777 | 1012 | |
ysr@777 | 1013 | // It pushes an object on the local queue. |
ysr@777 | 1014 | void push(oop obj); |
ysr@777 | 1015 | |
ysr@777 | 1016 | // These two move entries to/from the global stack. |
ysr@777 | 1017 | void move_entries_to_global_stack(); |
ysr@777 | 1018 | void get_entries_from_global_stack(); |
ysr@777 | 1019 | |
ysr@777 | 1020 | // It pops and scans objects from the local queue. If partially is |
ysr@777 | 1021 | // true, then it stops when the queue size is of a given limit. If |
ysr@777 | 1022 | // partially is false, then it stops when the queue is empty. |
ysr@777 | 1023 | void drain_local_queue(bool partially); |
ysr@777 | 1024 | // It moves entries from the global stack to the local queue and |
ysr@777 | 1025 | // drains the local queue. If partially is true, then it stops when |
ysr@777 | 1026 | // both the global stack and the local queue reach a given size. If |
ysr@777 | 1027 | // partially if false, it tries to empty them totally. |
ysr@777 | 1028 | void drain_global_stack(bool partially); |
ysr@777 | 1029 | // It keeps picking SATB buffers and processing them until no SATB |
ysr@777 | 1030 | // buffers are available. |
ysr@777 | 1031 | void drain_satb_buffers(); |
ysr@777 | 1032 | // It keeps popping regions from the region stack and processing |
ysr@777 | 1033 | // them until the region stack is empty. |
ysr@777 | 1034 | void drain_region_stack(BitMapClosure* closure); |
ysr@777 | 1035 | |
ysr@777 | 1036 | // moves the local finger to a new location |
ysr@777 | 1037 | inline void move_finger_to(HeapWord* new_finger) { |
tonyp@1458 | 1038 | assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); |
ysr@777 | 1039 | _finger = new_finger; |
ysr@777 | 1040 | } |
ysr@777 | 1041 | |
ysr@777 | 1042 | // moves the region finger to a new location |
ysr@777 | 1043 | inline void move_region_finger_to(HeapWord* new_finger) { |
tonyp@1458 | 1044 | assert(new_finger < _cm->finger(), "invariant"); |
ysr@777 | 1045 | _region_finger = new_finger; |
ysr@777 | 1046 | } |
ysr@777 | 1047 | |
ysr@777 | 1048 | CMTask(int task_num, ConcurrentMark *cm, |
ysr@777 | 1049 | CMTaskQueue* task_queue, CMTaskQueueSet* task_queues); |
ysr@777 | 1050 | |
ysr@777 | 1051 | // it prints statistics associated with this task |
ysr@777 | 1052 | void print_stats(); |
ysr@777 | 1053 | |
ysr@777 | 1054 | #if _MARKING_STATS_ |
ysr@777 | 1055 | void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; } |
ysr@777 | 1056 | #endif // _MARKING_STATS_ |
ysr@777 | 1057 | }; |