Mon, 12 Mar 2012 14:59:00 -0700
7147724: G1: hang in SurrogateLockerThread::manipulatePLL
Summary: Attempting to initiate a marking cycle when allocating a humongous object can, if a marking cycle is successfully initiated by another thread, result in the allocating thread spinning until the marking cycle is complete. Eliminate a deadlock between the main ConcurrentMarkThread, the SurrogateLocker thread, the VM thread, and a mutator thread waiting on the SecondaryFreeList_lock (while free regions are going to become available) by not manipulating the pending list lock during the prologue and epilogue of the cleanup pause.
Reviewed-by: brutisso, jcoomes, tonyp
ysr@777 | 1 | /* |
tonyp@3416 | 2 | * Copyright (c) 2001, 2012, Oracle and/or its affiliates. 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 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
ysr@777 | 22 | * |
ysr@777 | 23 | */ |
ysr@777 | 24 | |
stefank@2314 | 25 | #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP |
stefank@2314 | 26 | #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP |
stefank@2314 | 27 | |
tonyp@2472 | 28 | #include "gc_implementation/g1/heapRegionSets.hpp" |
stefank@2314 | 29 | #include "utilities/taskqueue.hpp" |
stefank@2314 | 30 | |
ysr@777 | 31 | class G1CollectedHeap; |
ysr@777 | 32 | class CMTask; |
jcoomes@1746 | 33 | typedef GenericTaskQueue<oop> CMTaskQueue; |
jcoomes@1746 | 34 | typedef GenericTaskQueueSet<CMTaskQueue> CMTaskQueueSet; |
ysr@777 | 35 | |
johnc@2379 | 36 | // Closure used by CM during concurrent reference discovery |
johnc@2379 | 37 | // and reference processing (during remarking) to determine |
johnc@2379 | 38 | // if a particular object is alive. It is primarily used |
johnc@2379 | 39 | // to determine if referents of discovered reference objects |
johnc@2379 | 40 | // are alive. An instance is also embedded into the |
johnc@2379 | 41 | // reference processor as the _is_alive_non_header field |
johnc@2379 | 42 | class G1CMIsAliveClosure: public BoolObjectClosure { |
johnc@2379 | 43 | G1CollectedHeap* _g1; |
johnc@2379 | 44 | public: |
johnc@2379 | 45 | G1CMIsAliveClosure(G1CollectedHeap* g1) : |
johnc@2379 | 46 | _g1(g1) |
johnc@2379 | 47 | {} |
johnc@2379 | 48 | |
johnc@2379 | 49 | void do_object(oop obj) { |
johnc@2379 | 50 | ShouldNotCallThis(); |
johnc@2379 | 51 | } |
johnc@2379 | 52 | bool do_object_b(oop obj); |
johnc@2379 | 53 | }; |
johnc@2379 | 54 | |
ysr@777 | 55 | // A generic CM bit map. This is essentially a wrapper around the BitMap |
ysr@777 | 56 | // class, with one bit per (1<<_shifter) HeapWords. |
ysr@777 | 57 | |
apetrusenko@984 | 58 | class CMBitMapRO VALUE_OBJ_CLASS_SPEC { |
ysr@777 | 59 | protected: |
ysr@777 | 60 | HeapWord* _bmStartWord; // base address of range covered by map |
ysr@777 | 61 | size_t _bmWordSize; // map size (in #HeapWords covered) |
ysr@777 | 62 | const int _shifter; // map to char or bit |
ysr@777 | 63 | VirtualSpace _virtual_space; // underlying the bit map |
ysr@777 | 64 | BitMap _bm; // the bit map itself |
ysr@777 | 65 | |
ysr@777 | 66 | public: |
ysr@777 | 67 | // constructor |
ysr@777 | 68 | CMBitMapRO(ReservedSpace rs, int shifter); |
ysr@777 | 69 | |
ysr@777 | 70 | enum { do_yield = true }; |
ysr@777 | 71 | |
ysr@777 | 72 | // inquiries |
ysr@777 | 73 | HeapWord* startWord() const { return _bmStartWord; } |
ysr@777 | 74 | size_t sizeInWords() const { return _bmWordSize; } |
ysr@777 | 75 | // the following is one past the last word in space |
ysr@777 | 76 | HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } |
ysr@777 | 77 | |
ysr@777 | 78 | // read marks |
ysr@777 | 79 | |
ysr@777 | 80 | bool isMarked(HeapWord* addr) const { |
ysr@777 | 81 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 82 | "outside underlying space?"); |
ysr@777 | 83 | return _bm.at(heapWordToOffset(addr)); |
ysr@777 | 84 | } |
ysr@777 | 85 | |
ysr@777 | 86 | // iteration |
johnc@3454 | 87 | inline bool iterate(BitMapClosure* cl, MemRegion mr); |
johnc@3454 | 88 | inline bool iterate(BitMapClosure* cl); |
ysr@777 | 89 | |
ysr@777 | 90 | // Return the address corresponding to the next marked bit at or after |
ysr@777 | 91 | // "addr", and before "limit", if "limit" is non-NULL. If there is no |
ysr@777 | 92 | // such bit, returns "limit" if that is non-NULL, or else "endWord()". |
ysr@777 | 93 | HeapWord* getNextMarkedWordAddress(HeapWord* addr, |
ysr@777 | 94 | HeapWord* limit = NULL) const; |
ysr@777 | 95 | // Return the address corresponding to the next unmarked bit at or after |
ysr@777 | 96 | // "addr", and before "limit", if "limit" is non-NULL. If there is no |
ysr@777 | 97 | // such bit, returns "limit" if that is non-NULL, or else "endWord()". |
ysr@777 | 98 | HeapWord* getNextUnmarkedWordAddress(HeapWord* addr, |
ysr@777 | 99 | HeapWord* limit = NULL) const; |
ysr@777 | 100 | |
ysr@777 | 101 | // conversion utilities |
ysr@777 | 102 | // XXX Fix these so that offsets are size_t's... |
ysr@777 | 103 | HeapWord* offsetToHeapWord(size_t offset) const { |
ysr@777 | 104 | return _bmStartWord + (offset << _shifter); |
ysr@777 | 105 | } |
ysr@777 | 106 | size_t heapWordToOffset(HeapWord* addr) const { |
ysr@777 | 107 | return pointer_delta(addr, _bmStartWord) >> _shifter; |
ysr@777 | 108 | } |
ysr@777 | 109 | int heapWordDiffToOffsetDiff(size_t diff) const; |
ysr@777 | 110 | HeapWord* nextWord(HeapWord* addr) { |
ysr@777 | 111 | return offsetToHeapWord(heapWordToOffset(addr) + 1); |
ysr@777 | 112 | } |
ysr@777 | 113 | |
ysr@777 | 114 | void mostly_disjoint_range_union(BitMap* from_bitmap, |
ysr@777 | 115 | size_t from_start_index, |
ysr@777 | 116 | HeapWord* to_start_word, |
ysr@777 | 117 | size_t word_num); |
ysr@777 | 118 | |
ysr@777 | 119 | // debugging |
ysr@777 | 120 | NOT_PRODUCT(bool covers(ReservedSpace rs) const;) |
ysr@777 | 121 | }; |
ysr@777 | 122 | |
ysr@777 | 123 | class CMBitMap : public CMBitMapRO { |
ysr@777 | 124 | |
ysr@777 | 125 | public: |
ysr@777 | 126 | // constructor |
ysr@777 | 127 | CMBitMap(ReservedSpace rs, int shifter) : |
ysr@777 | 128 | CMBitMapRO(rs, shifter) {} |
ysr@777 | 129 | |
ysr@777 | 130 | // write marks |
ysr@777 | 131 | void mark(HeapWord* addr) { |
ysr@777 | 132 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 133 | "outside underlying space?"); |
tonyp@2968 | 134 | _bm.set_bit(heapWordToOffset(addr)); |
ysr@777 | 135 | } |
ysr@777 | 136 | void clear(HeapWord* addr) { |
ysr@777 | 137 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 138 | "outside underlying space?"); |
tonyp@2968 | 139 | _bm.clear_bit(heapWordToOffset(addr)); |
ysr@777 | 140 | } |
ysr@777 | 141 | bool parMark(HeapWord* addr) { |
ysr@777 | 142 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 143 | "outside underlying space?"); |
tonyp@2968 | 144 | return _bm.par_set_bit(heapWordToOffset(addr)); |
ysr@777 | 145 | } |
ysr@777 | 146 | bool parClear(HeapWord* addr) { |
ysr@777 | 147 | assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), |
ysr@777 | 148 | "outside underlying space?"); |
tonyp@2968 | 149 | return _bm.par_clear_bit(heapWordToOffset(addr)); |
ysr@777 | 150 | } |
ysr@777 | 151 | void markRange(MemRegion mr); |
ysr@777 | 152 | void clearAll(); |
ysr@777 | 153 | void clearRange(MemRegion mr); |
ysr@777 | 154 | |
ysr@777 | 155 | // Starting at the bit corresponding to "addr" (inclusive), find the next |
ysr@777 | 156 | // "1" bit, if any. This bit starts some run of consecutive "1"'s; find |
ysr@777 | 157 | // the end of this run (stopping at "end_addr"). Return the MemRegion |
ysr@777 | 158 | // covering from the start of the region corresponding to the first bit |
ysr@777 | 159 | // of the run to the end of the region corresponding to the last bit of |
ysr@777 | 160 | // the run. If there is no "1" bit at or after "addr", return an empty |
ysr@777 | 161 | // MemRegion. |
ysr@777 | 162 | MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr); |
ysr@777 | 163 | }; |
ysr@777 | 164 | |
ysr@777 | 165 | // Represents a marking stack used by the CM collector. |
ysr@777 | 166 | // Ideally this should be GrowableArray<> just like MSC's marking stack(s). |
apetrusenko@984 | 167 | class CMMarkStack VALUE_OBJ_CLASS_SPEC { |
ysr@777 | 168 | ConcurrentMark* _cm; |
tonyp@3416 | 169 | oop* _base; // bottom of stack |
tonyp@3416 | 170 | jint _index; // one more than last occupied index |
tonyp@3416 | 171 | jint _capacity; // max #elements |
tonyp@3416 | 172 | jint _saved_index; // value of _index saved at start of GC |
ysr@777 | 173 | NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run |
ysr@777 | 174 | |
ysr@777 | 175 | bool _overflow; |
ysr@777 | 176 | DEBUG_ONLY(bool _drain_in_progress;) |
ysr@777 | 177 | DEBUG_ONLY(bool _drain_in_progress_yields;) |
ysr@777 | 178 | |
ysr@777 | 179 | public: |
ysr@777 | 180 | CMMarkStack(ConcurrentMark* cm); |
ysr@777 | 181 | ~CMMarkStack(); |
ysr@777 | 182 | |
ysr@777 | 183 | void allocate(size_t size); |
ysr@777 | 184 | |
ysr@777 | 185 | oop pop() { |
ysr@777 | 186 | if (!isEmpty()) { |
ysr@777 | 187 | return _base[--_index] ; |
ysr@777 | 188 | } |
ysr@777 | 189 | return NULL; |
ysr@777 | 190 | } |
ysr@777 | 191 | |
ysr@777 | 192 | // If overflow happens, don't do the push, and record the overflow. |
ysr@777 | 193 | // *Requires* that "ptr" is already marked. |
ysr@777 | 194 | void push(oop ptr) { |
ysr@777 | 195 | if (isFull()) { |
ysr@777 | 196 | // Record overflow. |
ysr@777 | 197 | _overflow = true; |
ysr@777 | 198 | return; |
ysr@777 | 199 | } else { |
ysr@777 | 200 | _base[_index++] = ptr; |
ysr@777 | 201 | NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index)); |
ysr@777 | 202 | } |
ysr@777 | 203 | } |
ysr@777 | 204 | // Non-block impl. Note: concurrency is allowed only with other |
ysr@777 | 205 | // "par_push" operations, not with "pop" or "drain". We would need |
ysr@777 | 206 | // parallel versions of them if such concurrency was desired. |
ysr@777 | 207 | void par_push(oop ptr); |
ysr@777 | 208 | |
ysr@777 | 209 | // Pushes the first "n" elements of "ptr_arr" on the stack. |
ysr@777 | 210 | // Non-block impl. Note: concurrency is allowed only with other |
ysr@777 | 211 | // "par_adjoin_arr" or "push" operations, not with "pop" or "drain". |
ysr@777 | 212 | void par_adjoin_arr(oop* ptr_arr, int n); |
ysr@777 | 213 | |
ysr@777 | 214 | // Pushes the first "n" elements of "ptr_arr" on the stack. |
ysr@777 | 215 | // Locking impl: concurrency is allowed only with |
ysr@777 | 216 | // "par_push_arr" and/or "par_pop_arr" operations, which use the same |
ysr@777 | 217 | // locking strategy. |
ysr@777 | 218 | void par_push_arr(oop* ptr_arr, int n); |
ysr@777 | 219 | |
ysr@777 | 220 | // If returns false, the array was empty. Otherwise, removes up to "max" |
ysr@777 | 221 | // elements from the stack, and transfers them to "ptr_arr" in an |
ysr@777 | 222 | // unspecified order. The actual number transferred is given in "n" ("n |
ysr@777 | 223 | // == 0" is deliberately redundant with the return value.) Locking impl: |
ysr@777 | 224 | // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" |
ysr@777 | 225 | // operations, which use the same locking strategy. |
ysr@777 | 226 | bool par_pop_arr(oop* ptr_arr, int max, int* n); |
ysr@777 | 227 | |
ysr@777 | 228 | // Drain the mark stack, applying the given closure to all fields of |
ysr@777 | 229 | // objects on the stack. (That is, continue until the stack is empty, |
ysr@777 | 230 | // even if closure applications add entries to the stack.) The "bm" |
ysr@777 | 231 | // argument, if non-null, may be used to verify that only marked objects |
ysr@777 | 232 | // are on the mark stack. If "yield_after" is "true", then the |
ysr@777 | 233 | // concurrent marker performing the drain offers to yield after |
ysr@777 | 234 | // processing each object. If a yield occurs, stops the drain operation |
ysr@777 | 235 | // and returns false. Otherwise, returns true. |
ysr@777 | 236 | template<class OopClosureClass> |
ysr@777 | 237 | bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false); |
ysr@777 | 238 | |
ysr@777 | 239 | bool isEmpty() { return _index == 0; } |
ysr@777 | 240 | bool isFull() { return _index == _capacity; } |
ysr@777 | 241 | int maxElems() { return _capacity; } |
ysr@777 | 242 | |
ysr@777 | 243 | bool overflow() { return _overflow; } |
ysr@777 | 244 | void clear_overflow() { _overflow = false; } |
ysr@777 | 245 | |
ysr@777 | 246 | int size() { return _index; } |
ysr@777 | 247 | |
ysr@777 | 248 | void setEmpty() { _index = 0; clear_overflow(); } |
ysr@777 | 249 | |
tonyp@3416 | 250 | // Record the current index. |
tonyp@3416 | 251 | void note_start_of_gc(); |
tonyp@3416 | 252 | |
tonyp@3416 | 253 | // Make sure that we have not added any entries to the stack during GC. |
tonyp@3416 | 254 | void note_end_of_gc(); |
tonyp@3416 | 255 | |
ysr@777 | 256 | // iterate over the oops in the mark stack, up to the bound recorded via |
ysr@777 | 257 | // the call above. |
ysr@777 | 258 | void oops_do(OopClosure* f); |
ysr@777 | 259 | }; |
ysr@777 | 260 | |
apetrusenko@984 | 261 | class CMRegionStack VALUE_OBJ_CLASS_SPEC { |
ysr@777 | 262 | MemRegion* _base; |
ysr@777 | 263 | jint _capacity; |
ysr@777 | 264 | jint _index; |
ysr@777 | 265 | jint _oops_do_bound; |
ysr@777 | 266 | bool _overflow; |
ysr@777 | 267 | public: |
ysr@777 | 268 | CMRegionStack(); |
ysr@777 | 269 | ~CMRegionStack(); |
ysr@777 | 270 | void allocate(size_t size); |
ysr@777 | 271 | |
ysr@777 | 272 | // This is lock-free; assumes that it will only be called in parallel |
ysr@777 | 273 | // with other "push" operations (no pops). |
johnc@2190 | 274 | void push_lock_free(MemRegion mr); |
tonyp@1793 | 275 | |
ysr@777 | 276 | // Lock-free; assumes that it will only be called in parallel |
ysr@777 | 277 | // with other "pop" operations (no pushes). |
johnc@2190 | 278 | MemRegion pop_lock_free(); |
johnc@2190 | 279 | |
johnc@2190 | 280 | #if 0 |
johnc@2190 | 281 | // The routines that manipulate the region stack with a lock are |
johnc@2190 | 282 | // not currently used. They should be retained, however, as a |
johnc@2190 | 283 | // diagnostic aid. |
tonyp@1793 | 284 | |
tonyp@1793 | 285 | // These two are the implementations that use a lock. They can be |
tonyp@1793 | 286 | // called concurrently with each other but they should not be called |
tonyp@1793 | 287 | // concurrently with the lock-free versions (push() / pop()). |
tonyp@1793 | 288 | void push_with_lock(MemRegion mr); |
tonyp@1793 | 289 | MemRegion pop_with_lock(); |
johnc@2190 | 290 | #endif |
ysr@777 | 291 | |
ysr@777 | 292 | bool isEmpty() { return _index == 0; } |
ysr@777 | 293 | bool isFull() { return _index == _capacity; } |
ysr@777 | 294 | |
ysr@777 | 295 | bool overflow() { return _overflow; } |
ysr@777 | 296 | void clear_overflow() { _overflow = false; } |
ysr@777 | 297 | |
ysr@777 | 298 | int size() { return _index; } |
ysr@777 | 299 | |
ysr@777 | 300 | // It iterates over the entries in the region stack and it |
ysr@777 | 301 | // invalidates (i.e. assigns MemRegion()) the ones that point to |
ysr@777 | 302 | // regions in the collection set. |
ysr@777 | 303 | bool invalidate_entries_into_cset(); |
ysr@777 | 304 | |
ysr@777 | 305 | // This gives an upper bound up to which the iteration in |
ysr@777 | 306 | // invalidate_entries_into_cset() will reach. This prevents |
ysr@777 | 307 | // newly-added entries to be unnecessarily scanned. |
ysr@777 | 308 | void set_oops_do_bound() { |
ysr@777 | 309 | _oops_do_bound = _index; |
ysr@777 | 310 | } |
ysr@777 | 311 | |
ysr@777 | 312 | void setEmpty() { _index = 0; clear_overflow(); } |
ysr@777 | 313 | }; |
ysr@777 | 314 | |
tonyp@2848 | 315 | class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC { |
tonyp@2848 | 316 | private: |
tonyp@2848 | 317 | #ifndef PRODUCT |
tonyp@2848 | 318 | uintx _num_remaining; |
tonyp@2848 | 319 | bool _force; |
tonyp@2848 | 320 | #endif // !defined(PRODUCT) |
tonyp@2848 | 321 | |
tonyp@2848 | 322 | public: |
tonyp@2848 | 323 | void init() PRODUCT_RETURN; |
tonyp@2848 | 324 | void update() PRODUCT_RETURN; |
tonyp@2848 | 325 | bool should_force() PRODUCT_RETURN_( return false; ); |
tonyp@2848 | 326 | }; |
tonyp@2848 | 327 | |
ysr@777 | 328 | // this will enable a variety of different statistics per GC task |
ysr@777 | 329 | #define _MARKING_STATS_ 0 |
ysr@777 | 330 | // this will enable the higher verbose levels |
ysr@777 | 331 | #define _MARKING_VERBOSE_ 0 |
ysr@777 | 332 | |
ysr@777 | 333 | #if _MARKING_STATS_ |
ysr@777 | 334 | #define statsOnly(statement) \ |
ysr@777 | 335 | do { \ |
ysr@777 | 336 | statement ; \ |
ysr@777 | 337 | } while (0) |
ysr@777 | 338 | #else // _MARKING_STATS_ |
ysr@777 | 339 | #define statsOnly(statement) \ |
ysr@777 | 340 | do { \ |
ysr@777 | 341 | } while (0) |
ysr@777 | 342 | #endif // _MARKING_STATS_ |
ysr@777 | 343 | |
ysr@777 | 344 | typedef enum { |
ysr@777 | 345 | no_verbose = 0, // verbose turned off |
ysr@777 | 346 | stats_verbose, // only prints stats at the end of marking |
ysr@777 | 347 | low_verbose, // low verbose, mostly per region and per major event |
ysr@777 | 348 | medium_verbose, // a bit more detailed than low |
ysr@777 | 349 | high_verbose // per object verbose |
ysr@777 | 350 | } CMVerboseLevel; |
ysr@777 | 351 | |
tonyp@3464 | 352 | class YoungList; |
tonyp@3464 | 353 | |
tonyp@3464 | 354 | // Root Regions are regions that are not empty at the beginning of a |
tonyp@3464 | 355 | // marking cycle and which we might collect during an evacuation pause |
tonyp@3464 | 356 | // while the cycle is active. Given that, during evacuation pauses, we |
tonyp@3464 | 357 | // do not copy objects that are explicitly marked, what we have to do |
tonyp@3464 | 358 | // for the root regions is to scan them and mark all objects reachable |
tonyp@3464 | 359 | // from them. According to the SATB assumptions, we only need to visit |
tonyp@3464 | 360 | // each object once during marking. So, as long as we finish this scan |
tonyp@3464 | 361 | // before the next evacuation pause, we can copy the objects from the |
tonyp@3464 | 362 | // root regions without having to mark them or do anything else to them. |
tonyp@3464 | 363 | // |
tonyp@3464 | 364 | // Currently, we only support root region scanning once (at the start |
tonyp@3464 | 365 | // of the marking cycle) and the root regions are all the survivor |
tonyp@3464 | 366 | // regions populated during the initial-mark pause. |
tonyp@3464 | 367 | class CMRootRegions VALUE_OBJ_CLASS_SPEC { |
tonyp@3464 | 368 | private: |
tonyp@3464 | 369 | YoungList* _young_list; |
tonyp@3464 | 370 | ConcurrentMark* _cm; |
tonyp@3464 | 371 | |
tonyp@3464 | 372 | volatile bool _scan_in_progress; |
tonyp@3464 | 373 | volatile bool _should_abort; |
tonyp@3464 | 374 | HeapRegion* volatile _next_survivor; |
tonyp@3464 | 375 | |
tonyp@3464 | 376 | public: |
tonyp@3464 | 377 | CMRootRegions(); |
tonyp@3464 | 378 | // We actually do most of the initialization in this method. |
tonyp@3464 | 379 | void init(G1CollectedHeap* g1h, ConcurrentMark* cm); |
tonyp@3464 | 380 | |
tonyp@3464 | 381 | // Reset the claiming / scanning of the root regions. |
tonyp@3464 | 382 | void prepare_for_scan(); |
tonyp@3464 | 383 | |
tonyp@3464 | 384 | // Forces get_next() to return NULL so that the iteration aborts early. |
tonyp@3464 | 385 | void abort() { _should_abort = true; } |
tonyp@3464 | 386 | |
tonyp@3464 | 387 | // Return true if the CM thread are actively scanning root regions, |
tonyp@3464 | 388 | // false otherwise. |
tonyp@3464 | 389 | bool scan_in_progress() { return _scan_in_progress; } |
tonyp@3464 | 390 | |
tonyp@3464 | 391 | // Claim the next root region to scan atomically, or return NULL if |
tonyp@3464 | 392 | // all have been claimed. |
tonyp@3464 | 393 | HeapRegion* claim_next(); |
tonyp@3464 | 394 | |
tonyp@3464 | 395 | // Flag that we're done with root region scanning and notify anyone |
tonyp@3464 | 396 | // who's waiting on it. If aborted is false, assume that all regions |
tonyp@3464 | 397 | // have been claimed. |
tonyp@3464 | 398 | void scan_finished(); |
tonyp@3464 | 399 | |
tonyp@3464 | 400 | // If CM threads are still scanning root regions, wait until they |
tonyp@3464 | 401 | // are done. Return true if we had to wait, false otherwise. |
tonyp@3464 | 402 | bool wait_until_scan_finished(); |
tonyp@3464 | 403 | }; |
ysr@777 | 404 | |
ysr@777 | 405 | class ConcurrentMarkThread; |
ysr@777 | 406 | |
tonyp@3464 | 407 | class ConcurrentMark : public CHeapObj { |
ysr@777 | 408 | friend class ConcurrentMarkThread; |
ysr@777 | 409 | friend class CMTask; |
ysr@777 | 410 | friend class CMBitMapClosure; |
johnc@3296 | 411 | friend class CSetMarkOopClosure; |
ysr@777 | 412 | friend class CMGlobalObjectClosure; |
ysr@777 | 413 | friend class CMRemarkTask; |
ysr@777 | 414 | friend class CMConcurrentMarkingTask; |
ysr@777 | 415 | friend class G1ParNoteEndTask; |
ysr@777 | 416 | friend class CalcLiveObjectsClosure; |
johnc@3175 | 417 | friend class G1CMRefProcTaskProxy; |
johnc@3175 | 418 | friend class G1CMRefProcTaskExecutor; |
johnc@2494 | 419 | friend class G1CMParKeepAliveAndDrainClosure; |
johnc@2494 | 420 | friend class G1CMParDrainMarkingStackClosure; |
ysr@777 | 421 | |
ysr@777 | 422 | protected: |
ysr@777 | 423 | ConcurrentMarkThread* _cmThread; // the thread doing the work |
ysr@777 | 424 | G1CollectedHeap* _g1h; // the heap. |
jmasa@3357 | 425 | uint _parallel_marking_threads; // the number of marking |
jmasa@3294 | 426 | // threads we're use |
jmasa@3357 | 427 | uint _max_parallel_marking_threads; // max number of marking |
jmasa@3294 | 428 | // threads we'll ever use |
ysr@777 | 429 | double _sleep_factor; // how much we have to sleep, with |
ysr@777 | 430 | // respect to the work we just did, to |
ysr@777 | 431 | // meet the marking overhead goal |
ysr@777 | 432 | double _marking_task_overhead; // marking target overhead for |
ysr@777 | 433 | // a single task |
ysr@777 | 434 | |
ysr@777 | 435 | // same as the two above, but for the cleanup task |
ysr@777 | 436 | double _cleanup_sleep_factor; |
ysr@777 | 437 | double _cleanup_task_overhead; |
ysr@777 | 438 | |
tonyp@2472 | 439 | FreeRegionList _cleanup_list; |
ysr@777 | 440 | |
brutisso@3455 | 441 | // Concurrent marking support structures |
ysr@777 | 442 | CMBitMap _markBitMap1; |
ysr@777 | 443 | CMBitMap _markBitMap2; |
ysr@777 | 444 | CMBitMapRO* _prevMarkBitMap; // completed mark bitmap |
ysr@777 | 445 | CMBitMap* _nextMarkBitMap; // under-construction mark bitmap |
ysr@777 | 446 | bool _at_least_one_mark_complete; |
ysr@777 | 447 | |
ysr@777 | 448 | BitMap _region_bm; |
ysr@777 | 449 | BitMap _card_bm; |
ysr@777 | 450 | |
ysr@777 | 451 | // Heap bounds |
ysr@777 | 452 | HeapWord* _heap_start; |
ysr@777 | 453 | HeapWord* _heap_end; |
ysr@777 | 454 | |
tonyp@3464 | 455 | // Root region tracking and claiming. |
tonyp@3464 | 456 | CMRootRegions _root_regions; |
tonyp@3464 | 457 | |
ysr@777 | 458 | // For gray objects |
ysr@777 | 459 | CMMarkStack _markStack; // Grey objects behind global finger. |
ysr@777 | 460 | CMRegionStack _regionStack; // Grey regions behind global finger. |
ysr@777 | 461 | HeapWord* volatile _finger; // the global finger, region aligned, |
ysr@777 | 462 | // always points to the end of the |
ysr@777 | 463 | // last claimed region |
ysr@777 | 464 | |
ysr@777 | 465 | // marking tasks |
jmasa@3357 | 466 | uint _max_task_num; // maximum task number |
jmasa@3357 | 467 | uint _active_tasks; // task num currently active |
ysr@777 | 468 | CMTask** _tasks; // task queue array (max_task_num len) |
ysr@777 | 469 | CMTaskQueueSet* _task_queues; // task queue set |
ysr@777 | 470 | ParallelTaskTerminator _terminator; // for termination |
ysr@777 | 471 | |
ysr@777 | 472 | // Two sync barriers that are used to synchronise tasks when an |
ysr@777 | 473 | // overflow occurs. The algorithm is the following. All tasks enter |
ysr@777 | 474 | // the first one to ensure that they have all stopped manipulating |
ysr@777 | 475 | // the global data structures. After they exit it, they re-initialise |
ysr@777 | 476 | // their data structures and task 0 re-initialises the global data |
ysr@777 | 477 | // structures. Then, they enter the second sync barrier. This |
ysr@777 | 478 | // ensure, that no task starts doing work before all data |
ysr@777 | 479 | // structures (local and global) have been re-initialised. When they |
ysr@777 | 480 | // exit it, they are free to start working again. |
ysr@777 | 481 | WorkGangBarrierSync _first_overflow_barrier_sync; |
ysr@777 | 482 | WorkGangBarrierSync _second_overflow_barrier_sync; |
ysr@777 | 483 | |
ysr@777 | 484 | // this is set by any task, when an overflow on the global data |
ysr@777 | 485 | // structures is detected. |
ysr@777 | 486 | volatile bool _has_overflown; |
ysr@777 | 487 | // true: marking is concurrent, false: we're in remark |
ysr@777 | 488 | volatile bool _concurrent; |
ysr@777 | 489 | // set at the end of a Full GC so that marking aborts |
ysr@777 | 490 | volatile bool _has_aborted; |
johnc@2190 | 491 | |
ysr@777 | 492 | // used when remark aborts due to an overflow to indicate that |
ysr@777 | 493 | // another concurrent marking phase should start |
ysr@777 | 494 | volatile bool _restart_for_overflow; |
ysr@777 | 495 | |
ysr@777 | 496 | // This is true from the very start of concurrent marking until the |
ysr@777 | 497 | // point when all the tasks complete their work. It is really used |
ysr@777 | 498 | // to determine the points between the end of concurrent marking and |
ysr@777 | 499 | // time of remark. |
ysr@777 | 500 | volatile bool _concurrent_marking_in_progress; |
ysr@777 | 501 | |
ysr@777 | 502 | // verbose level |
ysr@777 | 503 | CMVerboseLevel _verbose_level; |
ysr@777 | 504 | |
ysr@777 | 505 | // These two fields are used to implement the optimisation that |
ysr@777 | 506 | // avoids pushing objects on the global/region stack if there are |
ysr@777 | 507 | // no collection set regions above the lowest finger. |
ysr@777 | 508 | |
ysr@777 | 509 | // This is the lowest finger (among the global and local fingers), |
ysr@777 | 510 | // which is calculated before a new collection set is chosen. |
ysr@777 | 511 | HeapWord* _min_finger; |
ysr@777 | 512 | // If this flag is true, objects/regions that are marked below the |
ysr@777 | 513 | // finger should be pushed on the stack(s). If this is flag is |
ysr@777 | 514 | // false, it is safe not to push them on the stack(s). |
ysr@777 | 515 | bool _should_gray_objects; |
ysr@777 | 516 | |
ysr@777 | 517 | // All of these times are in ms. |
ysr@777 | 518 | NumberSeq _init_times; |
ysr@777 | 519 | NumberSeq _remark_times; |
ysr@777 | 520 | NumberSeq _remark_mark_times; |
ysr@777 | 521 | NumberSeq _remark_weak_ref_times; |
ysr@777 | 522 | NumberSeq _cleanup_times; |
ysr@777 | 523 | double _total_counting_time; |
ysr@777 | 524 | double _total_rs_scrub_time; |
ysr@777 | 525 | |
ysr@777 | 526 | double* _accum_task_vtime; // accumulated task vtime |
ysr@777 | 527 | |
jmasa@3294 | 528 | FlexibleWorkGang* _parallel_workers; |
ysr@777 | 529 | |
tonyp@2848 | 530 | ForceOverflowSettings _force_overflow_conc; |
tonyp@2848 | 531 | ForceOverflowSettings _force_overflow_stw; |
tonyp@2848 | 532 | |
ysr@777 | 533 | void weakRefsWork(bool clear_all_soft_refs); |
ysr@777 | 534 | |
ysr@777 | 535 | void swapMarkBitMaps(); |
ysr@777 | 536 | |
ysr@777 | 537 | // It resets the global marking data structures, as well as the |
ysr@777 | 538 | // task local ones; should be called during initial mark. |
ysr@777 | 539 | void reset(); |
ysr@777 | 540 | // It resets all the marking data structures. |
tonyp@2848 | 541 | void clear_marking_state(bool clear_overflow = true); |
ysr@777 | 542 | |
ysr@777 | 543 | // It should be called to indicate which phase we're in (concurrent |
ysr@777 | 544 | // mark or remark) and how many threads are currently active. |
jmasa@3357 | 545 | void set_phase(uint active_tasks, bool concurrent); |
ysr@777 | 546 | // We do this after we're done with marking so that the marking data |
ysr@777 | 547 | // structures are initialised to a sensible and predictable state. |
ysr@777 | 548 | void set_non_marking_state(); |
ysr@777 | 549 | |
ysr@777 | 550 | // prints all gathered CM-related statistics |
ysr@777 | 551 | void print_stats(); |
ysr@777 | 552 | |
tonyp@2472 | 553 | bool cleanup_list_is_empty() { |
tonyp@2472 | 554 | return _cleanup_list.is_empty(); |
tonyp@2472 | 555 | } |
tonyp@2472 | 556 | |
ysr@777 | 557 | // accessor methods |
jmasa@3357 | 558 | uint parallel_marking_threads() { return _parallel_marking_threads; } |
jmasa@3357 | 559 | uint max_parallel_marking_threads() { return _max_parallel_marking_threads;} |
ysr@777 | 560 | double sleep_factor() { return _sleep_factor; } |
ysr@777 | 561 | double marking_task_overhead() { return _marking_task_overhead;} |
ysr@777 | 562 | double cleanup_sleep_factor() { return _cleanup_sleep_factor; } |
ysr@777 | 563 | double cleanup_task_overhead() { return _cleanup_task_overhead;} |
ysr@777 | 564 | |
ysr@777 | 565 | HeapWord* finger() { return _finger; } |
ysr@777 | 566 | bool concurrent() { return _concurrent; } |
jmasa@3357 | 567 | uint active_tasks() { return _active_tasks; } |
ysr@777 | 568 | ParallelTaskTerminator* terminator() { return &_terminator; } |
ysr@777 | 569 | |
ysr@777 | 570 | // It claims the next available region to be scanned by a marking |
ysr@777 | 571 | // task. It might return NULL if the next region is empty or we have |
ysr@777 | 572 | // run out of regions. In the latter case, out_of_regions() |
ysr@777 | 573 | // determines whether we've really run out of regions or the task |
ysr@777 | 574 | // should call claim_region() again. This might seem a bit |
ysr@777 | 575 | // awkward. Originally, the code was written so that claim_region() |
ysr@777 | 576 | // either successfully returned with a non-empty region or there |
ysr@777 | 577 | // were no more regions to be claimed. The problem with this was |
ysr@777 | 578 | // that, in certain circumstances, it iterated over large chunks of |
ysr@777 | 579 | // the heap finding only empty regions and, while it was working, it |
ysr@777 | 580 | // was preventing the calling task to call its regular clock |
ysr@777 | 581 | // method. So, this way, each task will spend very little time in |
ysr@777 | 582 | // claim_region() and is allowed to call the regular clock method |
ysr@777 | 583 | // frequently. |
ysr@777 | 584 | HeapRegion* claim_region(int task); |
ysr@777 | 585 | |
ysr@777 | 586 | // It determines whether we've run out of regions to scan. |
ysr@777 | 587 | bool out_of_regions() { return _finger == _heap_end; } |
ysr@777 | 588 | |
ysr@777 | 589 | // Returns the task with the given id |
ysr@777 | 590 | CMTask* task(int id) { |
tonyp@1458 | 591 | assert(0 <= id && id < (int) _active_tasks, |
tonyp@1458 | 592 | "task id not within active bounds"); |
ysr@777 | 593 | return _tasks[id]; |
ysr@777 | 594 | } |
ysr@777 | 595 | |
ysr@777 | 596 | // Returns the task queue with the given id |
ysr@777 | 597 | CMTaskQueue* task_queue(int id) { |
tonyp@1458 | 598 | assert(0 <= id && id < (int) _active_tasks, |
tonyp@1458 | 599 | "task queue id not within active bounds"); |
ysr@777 | 600 | return (CMTaskQueue*) _task_queues->queue(id); |
ysr@777 | 601 | } |
ysr@777 | 602 | |
ysr@777 | 603 | // Returns the task queue set |
ysr@777 | 604 | CMTaskQueueSet* task_queues() { return _task_queues; } |
ysr@777 | 605 | |
ysr@777 | 606 | // Access / manipulation of the overflow flag which is set to |
ysr@777 | 607 | // indicate that the global stack or region stack has overflown |
ysr@777 | 608 | bool has_overflown() { return _has_overflown; } |
ysr@777 | 609 | void set_has_overflown() { _has_overflown = true; } |
ysr@777 | 610 | void clear_has_overflown() { _has_overflown = false; } |
tonyp@3464 | 611 | bool restart_for_overflow() { return _restart_for_overflow; } |
ysr@777 | 612 | |
ysr@777 | 613 | bool has_aborted() { return _has_aborted; } |
ysr@777 | 614 | |
ysr@777 | 615 | // Methods to enter the two overflow sync barriers |
ysr@777 | 616 | void enter_first_sync_barrier(int task_num); |
ysr@777 | 617 | void enter_second_sync_barrier(int task_num); |
ysr@777 | 618 | |
tonyp@2848 | 619 | ForceOverflowSettings* force_overflow_conc() { |
tonyp@2848 | 620 | return &_force_overflow_conc; |
tonyp@2848 | 621 | } |
tonyp@2848 | 622 | |
tonyp@2848 | 623 | ForceOverflowSettings* force_overflow_stw() { |
tonyp@2848 | 624 | return &_force_overflow_stw; |
tonyp@2848 | 625 | } |
tonyp@2848 | 626 | |
tonyp@2848 | 627 | ForceOverflowSettings* force_overflow() { |
tonyp@2848 | 628 | if (concurrent()) { |
tonyp@2848 | 629 | return force_overflow_conc(); |
tonyp@2848 | 630 | } else { |
tonyp@2848 | 631 | return force_overflow_stw(); |
tonyp@2848 | 632 | } |
tonyp@2848 | 633 | } |
tonyp@2848 | 634 | |
johnc@3463 | 635 | // Live Data Counting data structures... |
johnc@3463 | 636 | // These data structures are initialized at the start of |
johnc@3463 | 637 | // marking. They are written to while marking is active. |
johnc@3463 | 638 | // They are aggregated during remark; the aggregated values |
johnc@3463 | 639 | // are then used to populate the _region_bm, _card_bm, and |
johnc@3463 | 640 | // the total live bytes, which are then subsequently updated |
johnc@3463 | 641 | // during cleanup. |
johnc@3463 | 642 | |
johnc@3463 | 643 | // An array of bitmaps (one bit map per task). Each bitmap |
johnc@3463 | 644 | // is used to record the cards spanned by the live objects |
johnc@3463 | 645 | // marked by that task/worker. |
johnc@3463 | 646 | BitMap* _count_card_bitmaps; |
johnc@3463 | 647 | |
johnc@3463 | 648 | // Used to record the number of marked live bytes |
johnc@3463 | 649 | // (for each region, by worker thread). |
johnc@3463 | 650 | size_t** _count_marked_bytes; |
johnc@3463 | 651 | |
johnc@3463 | 652 | // Card index of the bottom of the G1 heap. Used for biasing indices into |
johnc@3463 | 653 | // the card bitmaps. |
johnc@3463 | 654 | intptr_t _heap_bottom_card_num; |
johnc@3463 | 655 | |
ysr@777 | 656 | public: |
ysr@777 | 657 | // Manipulation of the global mark stack. |
ysr@777 | 658 | // Notice that the first mark_stack_push is CAS-based, whereas the |
ysr@777 | 659 | // two below are Mutex-based. This is OK since the first one is only |
ysr@777 | 660 | // called during evacuation pauses and doesn't compete with the |
ysr@777 | 661 | // other two (which are called by the marking tasks during |
ysr@777 | 662 | // concurrent marking or remark). |
ysr@777 | 663 | bool mark_stack_push(oop p) { |
ysr@777 | 664 | _markStack.par_push(p); |
ysr@777 | 665 | if (_markStack.overflow()) { |
ysr@777 | 666 | set_has_overflown(); |
ysr@777 | 667 | return false; |
ysr@777 | 668 | } |
ysr@777 | 669 | return true; |
ysr@777 | 670 | } |
ysr@777 | 671 | bool mark_stack_push(oop* arr, int n) { |
ysr@777 | 672 | _markStack.par_push_arr(arr, n); |
ysr@777 | 673 | if (_markStack.overflow()) { |
ysr@777 | 674 | set_has_overflown(); |
ysr@777 | 675 | return false; |
ysr@777 | 676 | } |
ysr@777 | 677 | return true; |
ysr@777 | 678 | } |
ysr@777 | 679 | void mark_stack_pop(oop* arr, int max, int* n) { |
ysr@777 | 680 | _markStack.par_pop_arr(arr, max, n); |
ysr@777 | 681 | } |
tonyp@2973 | 682 | size_t mark_stack_size() { return _markStack.size(); } |
ysr@777 | 683 | size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } |
tonyp@2973 | 684 | bool mark_stack_overflow() { return _markStack.overflow(); } |
tonyp@2973 | 685 | bool mark_stack_empty() { return _markStack.isEmpty(); } |
ysr@777 | 686 | |
johnc@2190 | 687 | // (Lock-free) Manipulation of the region stack |
johnc@2190 | 688 | bool region_stack_push_lock_free(MemRegion mr) { |
tonyp@1793 | 689 | // Currently we only call the lock-free version during evacuation |
tonyp@1793 | 690 | // pauses. |
tonyp@1793 | 691 | assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); |
tonyp@1793 | 692 | |
johnc@2190 | 693 | _regionStack.push_lock_free(mr); |
ysr@777 | 694 | if (_regionStack.overflow()) { |
ysr@777 | 695 | set_has_overflown(); |
ysr@777 | 696 | return false; |
ysr@777 | 697 | } |
ysr@777 | 698 | return true; |
ysr@777 | 699 | } |
johnc@2190 | 700 | |
johnc@2190 | 701 | // Lock-free version of region-stack pop. Should only be |
johnc@2190 | 702 | // called in tandem with other lock-free pops. |
johnc@2190 | 703 | MemRegion region_stack_pop_lock_free() { |
johnc@2190 | 704 | return _regionStack.pop_lock_free(); |
johnc@2190 | 705 | } |
johnc@2190 | 706 | |
tonyp@1793 | 707 | #if 0 |
johnc@2190 | 708 | // The routines that manipulate the region stack with a lock are |
johnc@2190 | 709 | // not currently used. They should be retained, however, as a |
johnc@2190 | 710 | // diagnostic aid. |
tonyp@1793 | 711 | |
tonyp@1793 | 712 | bool region_stack_push_with_lock(MemRegion mr) { |
tonyp@1793 | 713 | // Currently we only call the lock-based version during either |
tonyp@1793 | 714 | // concurrent marking or remark. |
tonyp@1793 | 715 | assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(), |
tonyp@1793 | 716 | "if we are at a safepoint it should be the remark safepoint"); |
tonyp@1793 | 717 | |
tonyp@1793 | 718 | _regionStack.push_with_lock(mr); |
tonyp@1793 | 719 | if (_regionStack.overflow()) { |
tonyp@1793 | 720 | set_has_overflown(); |
tonyp@1793 | 721 | return false; |
tonyp@1793 | 722 | } |
tonyp@1793 | 723 | return true; |
tonyp@1793 | 724 | } |
johnc@2190 | 725 | |
tonyp@1793 | 726 | MemRegion region_stack_pop_with_lock() { |
tonyp@1793 | 727 | // Currently we only call the lock-based version during either |
tonyp@1793 | 728 | // concurrent marking or remark. |
tonyp@1793 | 729 | assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(), |
tonyp@1793 | 730 | "if we are at a safepoint it should be the remark safepoint"); |
tonyp@1793 | 731 | |
tonyp@1793 | 732 | return _regionStack.pop_with_lock(); |
tonyp@1793 | 733 | } |
johnc@2190 | 734 | #endif |
tonyp@1793 | 735 | |
ysr@777 | 736 | int region_stack_size() { return _regionStack.size(); } |
ysr@777 | 737 | bool region_stack_overflow() { return _regionStack.overflow(); } |
ysr@777 | 738 | bool region_stack_empty() { return _regionStack.isEmpty(); } |
ysr@777 | 739 | |
johnc@2190 | 740 | // Iterate over any regions that were aborted while draining the |
johnc@2190 | 741 | // region stack (any such regions are saved in the corresponding |
johnc@2190 | 742 | // CMTask) and invalidate (i.e. assign to the empty MemRegion()) |
johnc@2190 | 743 | // any regions that point into the collection set. |
johnc@2190 | 744 | bool invalidate_aborted_regions_in_cset(); |
johnc@2190 | 745 | |
johnc@2190 | 746 | // Returns true if there are any aborted memory regions. |
johnc@2190 | 747 | bool has_aborted_regions(); |
johnc@2190 | 748 | |
tonyp@3464 | 749 | CMRootRegions* root_regions() { return &_root_regions; } |
tonyp@3464 | 750 | |
ysr@777 | 751 | bool concurrent_marking_in_progress() { |
ysr@777 | 752 | return _concurrent_marking_in_progress; |
ysr@777 | 753 | } |
ysr@777 | 754 | void set_concurrent_marking_in_progress() { |
ysr@777 | 755 | _concurrent_marking_in_progress = true; |
ysr@777 | 756 | } |
ysr@777 | 757 | void clear_concurrent_marking_in_progress() { |
ysr@777 | 758 | _concurrent_marking_in_progress = false; |
ysr@777 | 759 | } |
ysr@777 | 760 | |
ysr@777 | 761 | void update_accum_task_vtime(int i, double vtime) { |
ysr@777 | 762 | _accum_task_vtime[i] += vtime; |
ysr@777 | 763 | } |
ysr@777 | 764 | |
ysr@777 | 765 | double all_task_accum_vtime() { |
ysr@777 | 766 | double ret = 0.0; |
ysr@777 | 767 | for (int i = 0; i < (int)_max_task_num; ++i) |
ysr@777 | 768 | ret += _accum_task_vtime[i]; |
ysr@777 | 769 | return ret; |
ysr@777 | 770 | } |
ysr@777 | 771 | |
ysr@777 | 772 | // Attempts to steal an object from the task queues of other tasks |
ysr@777 | 773 | bool try_stealing(int task_num, int* hash_seed, oop& obj) { |
ysr@777 | 774 | return _task_queues->steal(task_num, hash_seed, obj); |
ysr@777 | 775 | } |
ysr@777 | 776 | |
ysr@777 | 777 | // It grays an object by first marking it. Then, if it's behind the |
ysr@777 | 778 | // global finger, it also pushes it on the global stack. |
ysr@777 | 779 | void deal_with_reference(oop obj); |
ysr@777 | 780 | |
ysr@777 | 781 | ConcurrentMark(ReservedSpace rs, int max_regions); |
ysr@777 | 782 | ~ConcurrentMark(); |
johnc@3463 | 783 | |
ysr@777 | 784 | ConcurrentMarkThread* cmThread() { return _cmThread; } |
ysr@777 | 785 | |
ysr@777 | 786 | CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } |
ysr@777 | 787 | CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } |
ysr@777 | 788 | |
jmasa@3294 | 789 | // Returns the number of GC threads to be used in a concurrent |
jmasa@3294 | 790 | // phase based on the number of GC threads being used in a STW |
jmasa@3294 | 791 | // phase. |
jmasa@3357 | 792 | uint scale_parallel_threads(uint n_par_threads); |
jmasa@3294 | 793 | |
jmasa@3294 | 794 | // Calculates the number of GC threads to be used in a concurrent phase. |
jmasa@3357 | 795 | uint calc_parallel_marking_threads(); |
jmasa@3294 | 796 | |
ysr@777 | 797 | // The following three are interaction between CM and |
ysr@777 | 798 | // G1CollectedHeap |
ysr@777 | 799 | |
ysr@777 | 800 | // This notifies CM that a root during initial-mark needs to be |
tonyp@3464 | 801 | // grayed. It is MT-safe. word_size is the size of the object in |
tonyp@3464 | 802 | // words. It is passed explicitly as sometimes we cannot calculate |
tonyp@3464 | 803 | // it from the given object because it might be in an inconsistent |
tonyp@3464 | 804 | // state (e.g., in to-space and being copied). So the caller is |
tonyp@3464 | 805 | // responsible for dealing with this issue (e.g., get the size from |
tonyp@3464 | 806 | // the from-space image when the to-space image might be |
tonyp@3464 | 807 | // inconsistent) and always passing the size. hr is the region that |
tonyp@3464 | 808 | // contains the object and it's passed optionally from callers who |
tonyp@3464 | 809 | // might already have it (no point in recalculating it). |
tonyp@3464 | 810 | inline void grayRoot(oop obj, size_t word_size, |
tonyp@3464 | 811 | uint worker_id, HeapRegion* hr = NULL); |
tonyp@3416 | 812 | |
ysr@777 | 813 | // It's used during evacuation pauses to gray a region, if |
ysr@777 | 814 | // necessary, and it's MT-safe. It assumes that the caller has |
ysr@777 | 815 | // marked any objects on that region. If _should_gray_objects is |
ysr@777 | 816 | // true and we're still doing concurrent marking, the region is |
ysr@777 | 817 | // pushed on the region stack, if it is located below the global |
ysr@777 | 818 | // finger, otherwise we do nothing. |
ysr@777 | 819 | void grayRegionIfNecessary(MemRegion mr); |
tonyp@3416 | 820 | |
ysr@777 | 821 | // It's used during evacuation pauses to mark and, if necessary, |
ysr@777 | 822 | // gray a single object and it's MT-safe. It assumes the caller did |
ysr@777 | 823 | // not mark the object. If _should_gray_objects is true and we're |
ysr@777 | 824 | // still doing concurrent marking, the objects is pushed on the |
ysr@777 | 825 | // global stack, if it is located below the global finger, otherwise |
ysr@777 | 826 | // we do nothing. |
ysr@777 | 827 | void markAndGrayObjectIfNecessary(oop p); |
ysr@777 | 828 | |
tonyp@1823 | 829 | // It iterates over the heap and for each object it comes across it |
tonyp@1823 | 830 | // will dump the contents of its reference fields, as well as |
tonyp@1823 | 831 | // liveness information for the object and its referents. The dump |
tonyp@1823 | 832 | // will be written to a file with the following name: |
johnc@2969 | 833 | // G1PrintReachableBaseFile + "." + str. |
johnc@2969 | 834 | // vo decides whether the prev (vo == UsePrevMarking), the next |
johnc@2969 | 835 | // (vo == UseNextMarking) marking information, or the mark word |
johnc@2969 | 836 | // (vo == UseMarkWord) will be used to determine the liveness of |
johnc@2969 | 837 | // each object / referent. |
johnc@2969 | 838 | // If all is true, all objects in the heap will be dumped, otherwise |
johnc@2969 | 839 | // only the live ones. In the dump the following symbols / breviations |
johnc@2969 | 840 | // are used: |
tonyp@1823 | 841 | // M : an explicitly live object (its bitmap bit is set) |
tonyp@1823 | 842 | // > : an implicitly live object (over tams) |
tonyp@1823 | 843 | // O : an object outside the G1 heap (typically: in the perm gen) |
tonyp@1823 | 844 | // NOT : a reference field whose referent is not live |
tonyp@1823 | 845 | // AND MARKED : indicates that an object is both explicitly and |
tonyp@1823 | 846 | // implicitly live (it should be one or the other, not both) |
tonyp@1823 | 847 | void print_reachable(const char* str, |
johnc@2969 | 848 | VerifyOption vo, bool all) PRODUCT_RETURN; |
ysr@777 | 849 | |
ysr@777 | 850 | // Clear the next marking bitmap (will be called concurrently). |
ysr@777 | 851 | void clearNextBitmap(); |
ysr@777 | 852 | |
ysr@777 | 853 | // These two do the work that needs to be done before and after the |
ysr@777 | 854 | // initial root checkpoint. Since this checkpoint can be done at two |
ysr@777 | 855 | // different points (i.e. an explicit pause or piggy-backed on a |
ysr@777 | 856 | // young collection), then it's nice to be able to easily share the |
ysr@777 | 857 | // pre/post code. It might be the case that we can put everything in |
ysr@777 | 858 | // the post method. TP |
ysr@777 | 859 | void checkpointRootsInitialPre(); |
ysr@777 | 860 | void checkpointRootsInitialPost(); |
ysr@777 | 861 | |
tonyp@3464 | 862 | // Scan all the root regions and mark everything reachable from |
tonyp@3464 | 863 | // them. |
tonyp@3464 | 864 | void scanRootRegions(); |
tonyp@3464 | 865 | |
tonyp@3464 | 866 | // Scan a single root region and mark everything reachable from it. |
tonyp@3464 | 867 | void scanRootRegion(HeapRegion* hr, uint worker_id); |
tonyp@3464 | 868 | |
ysr@777 | 869 | // Do concurrent phase of marking, to a tentative transitive closure. |
ysr@777 | 870 | void markFromRoots(); |
ysr@777 | 871 | |
ysr@777 | 872 | // Process all unprocessed SATB buffers. It is called at the |
ysr@777 | 873 | // beginning of an evacuation pause. |
ysr@777 | 874 | void drainAllSATBBuffers(); |
ysr@777 | 875 | |
ysr@777 | 876 | void checkpointRootsFinal(bool clear_all_soft_refs); |
ysr@777 | 877 | void checkpointRootsFinalWork(); |
ysr@777 | 878 | void cleanup(); |
ysr@777 | 879 | void completeCleanup(); |
ysr@777 | 880 | |
ysr@777 | 881 | // Mark in the previous bitmap. NB: this is usually read-only, so use |
ysr@777 | 882 | // this carefully! |
tonyp@3416 | 883 | inline void markPrev(oop p); |
johnc@3463 | 884 | |
tonyp@3416 | 885 | // Clears marks for all objects in the given range, for the prev, |
tonyp@3416 | 886 | // next, or both bitmaps. NB: the previous bitmap is usually |
tonyp@3416 | 887 | // read-only, so use this carefully! |
tonyp@3416 | 888 | void clearRangePrevBitmap(MemRegion mr); |
tonyp@3416 | 889 | void clearRangeNextBitmap(MemRegion mr); |
tonyp@3416 | 890 | void clearRangeBothBitmaps(MemRegion mr); |
ysr@777 | 891 | |
tonyp@3416 | 892 | // Notify data structures that a GC has started. |
tonyp@3416 | 893 | void note_start_of_gc() { |
tonyp@3416 | 894 | _markStack.note_start_of_gc(); |
ysr@777 | 895 | } |
tonyp@3416 | 896 | |
tonyp@3416 | 897 | // Notify data structures that a GC is finished. |
tonyp@3416 | 898 | void note_end_of_gc() { |
tonyp@3416 | 899 | _markStack.note_end_of_gc(); |
tonyp@3416 | 900 | } |
tonyp@3416 | 901 | |
ysr@777 | 902 | // Iterate over the oops in the mark stack and all local queues. It |
ysr@777 | 903 | // also calls invalidate_entries_into_cset() on the region stack. |
ysr@777 | 904 | void oops_do(OopClosure* f); |
tonyp@3416 | 905 | |
tonyp@3416 | 906 | // Verify that there are no CSet oops on the stacks (taskqueues / |
tonyp@3416 | 907 | // global mark stack), enqueued SATB buffers, per-thread SATB |
tonyp@3416 | 908 | // buffers, and fingers (global / per-task). The boolean parameters |
tonyp@3416 | 909 | // decide which of the above data structures to verify. If marking |
tonyp@3416 | 910 | // is not in progress, it's a no-op. |
tonyp@3416 | 911 | void verify_no_cset_oops(bool verify_stacks, |
tonyp@3416 | 912 | bool verify_enqueued_buffers, |
tonyp@3416 | 913 | bool verify_thread_buffers, |
tonyp@3416 | 914 | bool verify_fingers) PRODUCT_RETURN; |
tonyp@3416 | 915 | |
ysr@777 | 916 | // It is called at the end of an evacuation pause during marking so |
ysr@777 | 917 | // that CM is notified of where the new end of the heap is. It |
ysr@777 | 918 | // doesn't do anything if concurrent_marking_in_progress() is false, |
ysr@777 | 919 | // unless the force parameter is true. |
ysr@777 | 920 | void update_g1_committed(bool force = false); |
ysr@777 | 921 | |
ysr@777 | 922 | void complete_marking_in_collection_set(); |
ysr@777 | 923 | |
ysr@777 | 924 | // It indicates that a new collection set is being chosen. |
ysr@777 | 925 | void newCSet(); |
johnc@2910 | 926 | |
ysr@777 | 927 | // It registers a collection set heap region with CM. This is used |
ysr@777 | 928 | // to determine whether any heap regions are located above the finger. |
ysr@777 | 929 | void registerCSetRegion(HeapRegion* hr); |
ysr@777 | 930 | |
johnc@2910 | 931 | // Resets the region fields of any active CMTask whose region fields |
johnc@2910 | 932 | // are in the collection set (i.e. the region currently claimed by |
johnc@2910 | 933 | // the CMTask will be evacuated and may be used, subsequently, as |
johnc@2910 | 934 | // an alloc region). When this happens the region fields in the CMTask |
johnc@2910 | 935 | // are stale and, hence, should be cleared causing the worker thread |
johnc@2910 | 936 | // to claim a new region. |
johnc@2910 | 937 | void reset_active_task_region_fields_in_cset(); |
johnc@2910 | 938 | |
johnc@1829 | 939 | // Registers the maximum region-end associated with a set of |
johnc@1829 | 940 | // regions with CM. Again this is used to determine whether any |
johnc@1829 | 941 | // heap regions are located above the finger. |
johnc@1829 | 942 | void register_collection_set_finger(HeapWord* max_finger) { |
johnc@1829 | 943 | // max_finger is the highest heap region end of the regions currently |
johnc@1829 | 944 | // contained in the collection set. If this value is larger than |
johnc@1829 | 945 | // _min_finger then we need to gray objects. |
johnc@1829 | 946 | // This routine is like registerCSetRegion but for an entire |
johnc@1829 | 947 | // collection of regions. |
tonyp@2973 | 948 | if (max_finger > _min_finger) { |
johnc@1829 | 949 | _should_gray_objects = true; |
tonyp@2973 | 950 | } |
johnc@1829 | 951 | } |
johnc@1829 | 952 | |
ysr@777 | 953 | // Returns "true" if at least one mark has been completed. |
ysr@777 | 954 | bool at_least_one_mark_complete() { return _at_least_one_mark_complete; } |
ysr@777 | 955 | |
ysr@777 | 956 | bool isMarked(oop p) const { |
ysr@777 | 957 | assert(p != NULL && p->is_oop(), "expected an oop"); |
ysr@777 | 958 | HeapWord* addr = (HeapWord*)p; |
ysr@777 | 959 | assert(addr >= _nextMarkBitMap->startWord() || |
ysr@777 | 960 | addr < _nextMarkBitMap->endWord(), "in a region"); |
ysr@777 | 961 | |
ysr@777 | 962 | return _nextMarkBitMap->isMarked(addr); |
ysr@777 | 963 | } |
ysr@777 | 964 | |
ysr@777 | 965 | inline bool not_yet_marked(oop p) const; |
ysr@777 | 966 | |
ysr@777 | 967 | // XXX Debug code |
ysr@777 | 968 | bool containing_card_is_marked(void* p); |
ysr@777 | 969 | bool containing_cards_are_marked(void* start, void* last); |
ysr@777 | 970 | |
ysr@777 | 971 | bool isPrevMarked(oop p) const { |
ysr@777 | 972 | assert(p != NULL && p->is_oop(), "expected an oop"); |
ysr@777 | 973 | HeapWord* addr = (HeapWord*)p; |
ysr@777 | 974 | assert(addr >= _prevMarkBitMap->startWord() || |
ysr@777 | 975 | addr < _prevMarkBitMap->endWord(), "in a region"); |
ysr@777 | 976 | |
ysr@777 | 977 | return _prevMarkBitMap->isMarked(addr); |
ysr@777 | 978 | } |
ysr@777 | 979 | |
jmasa@3357 | 980 | inline bool do_yield_check(uint worker_i = 0); |
ysr@777 | 981 | inline bool should_yield(); |
ysr@777 | 982 | |
ysr@777 | 983 | // Called to abort the marking cycle after a Full GC takes palce. |
ysr@777 | 984 | void abort(); |
ysr@777 | 985 | |
ysr@777 | 986 | // This prints the global/local fingers. It is used for debugging. |
ysr@777 | 987 | NOT_PRODUCT(void print_finger();) |
ysr@777 | 988 | |
ysr@777 | 989 | void print_summary_info(); |
ysr@777 | 990 | |
tonyp@1454 | 991 | void print_worker_threads_on(outputStream* st) const; |
tonyp@1454 | 992 | |
ysr@777 | 993 | // The following indicate whether a given verbose level has been |
ysr@777 | 994 | // set. Notice that anything above stats is conditional to |
ysr@777 | 995 | // _MARKING_VERBOSE_ having been set to 1 |
tonyp@2973 | 996 | bool verbose_stats() { |
tonyp@2973 | 997 | return _verbose_level >= stats_verbose; |
tonyp@2973 | 998 | } |
tonyp@2973 | 999 | bool verbose_low() { |
tonyp@2973 | 1000 | return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; |
tonyp@2973 | 1001 | } |
tonyp@2973 | 1002 | bool verbose_medium() { |
tonyp@2973 | 1003 | return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; |
tonyp@2973 | 1004 | } |
tonyp@2973 | 1005 | bool verbose_high() { |
tonyp@2973 | 1006 | return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; |
tonyp@2973 | 1007 | } |
johnc@3463 | 1008 | |
johnc@3463 | 1009 | // Counting data structure accessors |
johnc@3463 | 1010 | |
johnc@3463 | 1011 | // Returns the card number of the bottom of the G1 heap. |
johnc@3463 | 1012 | // Used in biasing indices into accounting card bitmaps. |
johnc@3463 | 1013 | intptr_t heap_bottom_card_num() const { |
johnc@3463 | 1014 | return _heap_bottom_card_num; |
johnc@3463 | 1015 | } |
johnc@3463 | 1016 | |
johnc@3463 | 1017 | // Returns the card bitmap for a given task or worker id. |
johnc@3463 | 1018 | BitMap* count_card_bitmap_for(uint worker_id) { |
johnc@3463 | 1019 | assert(0 <= worker_id && worker_id < _max_task_num, "oob"); |
johnc@3463 | 1020 | assert(_count_card_bitmaps != NULL, "uninitialized"); |
johnc@3463 | 1021 | BitMap* task_card_bm = &_count_card_bitmaps[worker_id]; |
johnc@3463 | 1022 | assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); |
johnc@3463 | 1023 | return task_card_bm; |
johnc@3463 | 1024 | } |
johnc@3463 | 1025 | |
johnc@3463 | 1026 | // Returns the array containing the marked bytes for each region, |
johnc@3463 | 1027 | // for the given worker or task id. |
johnc@3463 | 1028 | size_t* count_marked_bytes_array_for(uint worker_id) { |
johnc@3463 | 1029 | assert(0 <= worker_id && worker_id < _max_task_num, "oob"); |
johnc@3463 | 1030 | assert(_count_marked_bytes != NULL, "uninitialized"); |
johnc@3463 | 1031 | size_t* marked_bytes_array = _count_marked_bytes[worker_id]; |
johnc@3463 | 1032 | assert(marked_bytes_array != NULL, "uninitialized"); |
johnc@3463 | 1033 | return marked_bytes_array; |
johnc@3463 | 1034 | } |
johnc@3463 | 1035 | |
johnc@3463 | 1036 | // Returns the index in the liveness accounting card table bitmap |
johnc@3463 | 1037 | // for the given address |
johnc@3463 | 1038 | inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr); |
johnc@3463 | 1039 | |
johnc@3463 | 1040 | // Counts the size of the given memory region in the the given |
johnc@3463 | 1041 | // marked_bytes array slot for the given HeapRegion. |
johnc@3463 | 1042 | // Sets the bits in the given card bitmap that are associated with the |
johnc@3463 | 1043 | // cards that are spanned by the memory region. |
johnc@3463 | 1044 | inline void count_region(MemRegion mr, HeapRegion* hr, |
johnc@3463 | 1045 | size_t* marked_bytes_array, |
johnc@3463 | 1046 | BitMap* task_card_bm); |
johnc@3463 | 1047 | |
johnc@3463 | 1048 | // Counts the given memory region in the task/worker counting |
johnc@3463 | 1049 | // data structures for the given worker id. |
tonyp@3464 | 1050 | inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id); |
tonyp@3464 | 1051 | |
tonyp@3464 | 1052 | // Counts the given memory region in the task/worker counting |
tonyp@3464 | 1053 | // data structures for the given worker id. |
johnc@3463 | 1054 | inline void count_region(MemRegion mr, uint worker_id); |
johnc@3463 | 1055 | |
johnc@3463 | 1056 | // Counts the given object in the given task/worker counting |
johnc@3463 | 1057 | // data structures. |
johnc@3463 | 1058 | inline void count_object(oop obj, HeapRegion* hr, |
johnc@3463 | 1059 | size_t* marked_bytes_array, |
johnc@3463 | 1060 | BitMap* task_card_bm); |
johnc@3463 | 1061 | |
johnc@3463 | 1062 | // Counts the given object in the task/worker counting data |
johnc@3463 | 1063 | // structures for the given worker id. |
johnc@3463 | 1064 | inline void count_object(oop obj, HeapRegion* hr, uint worker_id); |
johnc@3463 | 1065 | |
johnc@3463 | 1066 | // Attempts to mark the given object and, if successful, counts |
johnc@3463 | 1067 | // the object in the given task/worker counting structures. |
johnc@3463 | 1068 | inline bool par_mark_and_count(oop obj, HeapRegion* hr, |
johnc@3463 | 1069 | size_t* marked_bytes_array, |
johnc@3463 | 1070 | BitMap* task_card_bm); |
johnc@3463 | 1071 | |
johnc@3463 | 1072 | // Attempts to mark the given object and, if successful, counts |
johnc@3463 | 1073 | // the object in the task/worker counting structures for the |
johnc@3463 | 1074 | // given worker id. |
tonyp@3464 | 1075 | inline bool par_mark_and_count(oop obj, size_t word_size, |
tonyp@3464 | 1076 | HeapRegion* hr, uint worker_id); |
tonyp@3464 | 1077 | |
tonyp@3464 | 1078 | // Attempts to mark the given object and, if successful, counts |
tonyp@3464 | 1079 | // the object in the task/worker counting structures for the |
tonyp@3464 | 1080 | // given worker id. |
johnc@3463 | 1081 | inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id); |
johnc@3463 | 1082 | |
johnc@3463 | 1083 | // Similar to the above routine but we don't know the heap region that |
johnc@3463 | 1084 | // contains the object to be marked/counted, which this routine looks up. |
johnc@3463 | 1085 | inline bool par_mark_and_count(oop obj, uint worker_id); |
johnc@3463 | 1086 | |
johnc@3463 | 1087 | // Similar to the above routine but there are times when we cannot |
johnc@3463 | 1088 | // safely calculate the size of obj due to races and we, therefore, |
johnc@3463 | 1089 | // pass the size in as a parameter. It is the caller's reponsibility |
johnc@3463 | 1090 | // to ensure that the size passed in for obj is valid. |
johnc@3463 | 1091 | inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id); |
johnc@3463 | 1092 | |
johnc@3463 | 1093 | // Unconditionally mark the given object, and unconditinally count |
johnc@3463 | 1094 | // the object in the counting structures for worker id 0. |
johnc@3463 | 1095 | // Should *not* be called from parallel code. |
johnc@3463 | 1096 | inline bool mark_and_count(oop obj, HeapRegion* hr); |
johnc@3463 | 1097 | |
johnc@3463 | 1098 | // Similar to the above routine but we don't know the heap region that |
johnc@3463 | 1099 | // contains the object to be marked/counted, which this routine looks up. |
johnc@3463 | 1100 | // Should *not* be called from parallel code. |
johnc@3463 | 1101 | inline bool mark_and_count(oop obj); |
johnc@3463 | 1102 | |
johnc@3463 | 1103 | protected: |
johnc@3463 | 1104 | // Clear all the per-task bitmaps and arrays used to store the |
johnc@3463 | 1105 | // counting data. |
johnc@3463 | 1106 | void clear_all_count_data(); |
johnc@3463 | 1107 | |
johnc@3463 | 1108 | // Aggregates the counting data for each worker/task |
johnc@3463 | 1109 | // that was constructed while marking. Also sets |
johnc@3463 | 1110 | // the amount of marked bytes for each region and |
johnc@3463 | 1111 | // the top at concurrent mark count. |
johnc@3463 | 1112 | void aggregate_count_data(); |
johnc@3463 | 1113 | |
johnc@3463 | 1114 | // Verification routine |
johnc@3463 | 1115 | void verify_count_data(); |
ysr@777 | 1116 | }; |
ysr@777 | 1117 | |
ysr@777 | 1118 | // A class representing a marking task. |
ysr@777 | 1119 | class CMTask : public TerminatorTerminator { |
ysr@777 | 1120 | private: |
ysr@777 | 1121 | enum PrivateConstants { |
ysr@777 | 1122 | // the regular clock call is called once the scanned words reaches |
ysr@777 | 1123 | // this limit |
ysr@777 | 1124 | words_scanned_period = 12*1024, |
ysr@777 | 1125 | // the regular clock call is called once the number of visited |
ysr@777 | 1126 | // references reaches this limit |
ysr@777 | 1127 | refs_reached_period = 384, |
ysr@777 | 1128 | // initial value for the hash seed, used in the work stealing code |
ysr@777 | 1129 | init_hash_seed = 17, |
ysr@777 | 1130 | // how many entries will be transferred between global stack and |
ysr@777 | 1131 | // local queues |
ysr@777 | 1132 | global_stack_transfer_size = 16 |
ysr@777 | 1133 | }; |
ysr@777 | 1134 | |
ysr@777 | 1135 | int _task_id; |
ysr@777 | 1136 | G1CollectedHeap* _g1h; |
ysr@777 | 1137 | ConcurrentMark* _cm; |
ysr@777 | 1138 | CMBitMap* _nextMarkBitMap; |
ysr@777 | 1139 | // the task queue of this task |
ysr@777 | 1140 | CMTaskQueue* _task_queue; |
ysr@1280 | 1141 | private: |
ysr@777 | 1142 | // the task queue set---needed for stealing |
ysr@777 | 1143 | CMTaskQueueSet* _task_queues; |
ysr@777 | 1144 | // indicates whether the task has been claimed---this is only for |
ysr@777 | 1145 | // debugging purposes |
ysr@777 | 1146 | bool _claimed; |
ysr@777 | 1147 | |
ysr@777 | 1148 | // number of calls to this task |
ysr@777 | 1149 | int _calls; |
ysr@777 | 1150 | |
ysr@777 | 1151 | // when the virtual timer reaches this time, the marking step should |
ysr@777 | 1152 | // exit |
ysr@777 | 1153 | double _time_target_ms; |
ysr@777 | 1154 | // the start time of the current marking step |
ysr@777 | 1155 | double _start_time_ms; |
ysr@777 | 1156 | |
ysr@777 | 1157 | // the oop closure used for iterations over oops |
tonyp@2968 | 1158 | G1CMOopClosure* _cm_oop_closure; |
ysr@777 | 1159 | |
ysr@777 | 1160 | // the region this task is scanning, NULL if we're not scanning any |
ysr@777 | 1161 | HeapRegion* _curr_region; |
ysr@777 | 1162 | // the local finger of this task, NULL if we're not scanning a region |
ysr@777 | 1163 | HeapWord* _finger; |
ysr@777 | 1164 | // limit of the region this task is scanning, NULL if we're not scanning one |
ysr@777 | 1165 | HeapWord* _region_limit; |
ysr@777 | 1166 | |
ysr@777 | 1167 | // This is used only when we scan regions popped from the region |
ysr@777 | 1168 | // stack. It records what the last object on such a region we |
ysr@777 | 1169 | // scanned was. It is used to ensure that, if we abort region |
ysr@777 | 1170 | // iteration, we do not rescan the first part of the region. This |
ysr@777 | 1171 | // should be NULL when we're not scanning a region from the region |
ysr@777 | 1172 | // stack. |
ysr@777 | 1173 | HeapWord* _region_finger; |
ysr@777 | 1174 | |
johnc@2190 | 1175 | // If we abort while scanning a region we record the remaining |
johnc@2190 | 1176 | // unscanned portion and check this field when marking restarts. |
johnc@2190 | 1177 | // This avoids having to push on the region stack while other |
johnc@2190 | 1178 | // marking threads may still be popping regions. |
johnc@2190 | 1179 | // If we were to push the unscanned portion directly to the |
johnc@2190 | 1180 | // region stack then we would need to using locking versions |
johnc@2190 | 1181 | // of the push and pop operations. |
johnc@2190 | 1182 | MemRegion _aborted_region; |
johnc@2190 | 1183 | |
ysr@777 | 1184 | // the number of words this task has scanned |
ysr@777 | 1185 | size_t _words_scanned; |
ysr@777 | 1186 | // When _words_scanned reaches this limit, the regular clock is |
ysr@777 | 1187 | // called. Notice that this might be decreased under certain |
ysr@777 | 1188 | // circumstances (i.e. when we believe that we did an expensive |
ysr@777 | 1189 | // operation). |
ysr@777 | 1190 | size_t _words_scanned_limit; |
ysr@777 | 1191 | // the initial value of _words_scanned_limit (i.e. what it was |
ysr@777 | 1192 | // before it was decreased). |
ysr@777 | 1193 | size_t _real_words_scanned_limit; |
ysr@777 | 1194 | |
ysr@777 | 1195 | // the number of references this task has visited |
ysr@777 | 1196 | size_t _refs_reached; |
ysr@777 | 1197 | // When _refs_reached reaches this limit, the regular clock is |
ysr@777 | 1198 | // called. Notice this this might be decreased under certain |
ysr@777 | 1199 | // circumstances (i.e. when we believe that we did an expensive |
ysr@777 | 1200 | // operation). |
ysr@777 | 1201 | size_t _refs_reached_limit; |
ysr@777 | 1202 | // the initial value of _refs_reached_limit (i.e. what it was before |
ysr@777 | 1203 | // it was decreased). |
ysr@777 | 1204 | size_t _real_refs_reached_limit; |
ysr@777 | 1205 | |
ysr@777 | 1206 | // used by the work stealing stuff |
ysr@777 | 1207 | int _hash_seed; |
ysr@777 | 1208 | // if this is true, then the task has aborted for some reason |
ysr@777 | 1209 | bool _has_aborted; |
ysr@777 | 1210 | // set when the task aborts because it has met its time quota |
johnc@2494 | 1211 | bool _has_timed_out; |
ysr@777 | 1212 | // true when we're draining SATB buffers; this avoids the task |
ysr@777 | 1213 | // aborting due to SATB buffers being available (as we're already |
ysr@777 | 1214 | // dealing with them) |
ysr@777 | 1215 | bool _draining_satb_buffers; |
ysr@777 | 1216 | |
ysr@777 | 1217 | // number sequence of past step times |
ysr@777 | 1218 | NumberSeq _step_times_ms; |
ysr@777 | 1219 | // elapsed time of this task |
ysr@777 | 1220 | double _elapsed_time_ms; |
ysr@777 | 1221 | // termination time of this task |
ysr@777 | 1222 | double _termination_time_ms; |
ysr@777 | 1223 | // when this task got into the termination protocol |
ysr@777 | 1224 | double _termination_start_time_ms; |
ysr@777 | 1225 | |
ysr@777 | 1226 | // true when the task is during a concurrent phase, false when it is |
ysr@777 | 1227 | // in the remark phase (so, in the latter case, we do not have to |
ysr@777 | 1228 | // check all the things that we have to check during the concurrent |
ysr@777 | 1229 | // phase, i.e. SATB buffer availability...) |
ysr@777 | 1230 | bool _concurrent; |
ysr@777 | 1231 | |
ysr@777 | 1232 | TruncatedSeq _marking_step_diffs_ms; |
ysr@777 | 1233 | |
johnc@3463 | 1234 | // Counting data structures. Embedding the task's marked_bytes_array |
johnc@3463 | 1235 | // and card bitmap into the actual task saves having to go through |
johnc@3463 | 1236 | // the ConcurrentMark object. |
johnc@3463 | 1237 | size_t* _marked_bytes_array; |
johnc@3463 | 1238 | BitMap* _card_bm; |
johnc@3463 | 1239 | |
ysr@777 | 1240 | // LOTS of statistics related with this task |
ysr@777 | 1241 | #if _MARKING_STATS_ |
ysr@777 | 1242 | NumberSeq _all_clock_intervals_ms; |
ysr@777 | 1243 | double _interval_start_time_ms; |
ysr@777 | 1244 | |
ysr@777 | 1245 | int _aborted; |
ysr@777 | 1246 | int _aborted_overflow; |
ysr@777 | 1247 | int _aborted_cm_aborted; |
ysr@777 | 1248 | int _aborted_yield; |
ysr@777 | 1249 | int _aborted_timed_out; |
ysr@777 | 1250 | int _aborted_satb; |
ysr@777 | 1251 | int _aborted_termination; |
ysr@777 | 1252 | |
ysr@777 | 1253 | int _steal_attempts; |
ysr@777 | 1254 | int _steals; |
ysr@777 | 1255 | |
ysr@777 | 1256 | int _clock_due_to_marking; |
ysr@777 | 1257 | int _clock_due_to_scanning; |
ysr@777 | 1258 | |
ysr@777 | 1259 | int _local_pushes; |
ysr@777 | 1260 | int _local_pops; |
ysr@777 | 1261 | int _local_max_size; |
ysr@777 | 1262 | int _objs_scanned; |
ysr@777 | 1263 | |
ysr@777 | 1264 | int _global_pushes; |
ysr@777 | 1265 | int _global_pops; |
ysr@777 | 1266 | int _global_max_size; |
ysr@777 | 1267 | |
ysr@777 | 1268 | int _global_transfers_to; |
ysr@777 | 1269 | int _global_transfers_from; |
ysr@777 | 1270 | |
ysr@777 | 1271 | int _region_stack_pops; |
ysr@777 | 1272 | |
ysr@777 | 1273 | int _regions_claimed; |
ysr@777 | 1274 | int _objs_found_on_bitmap; |
ysr@777 | 1275 | |
ysr@777 | 1276 | int _satb_buffers_processed; |
ysr@777 | 1277 | #endif // _MARKING_STATS_ |
ysr@777 | 1278 | |
ysr@777 | 1279 | // it updates the local fields after this task has claimed |
ysr@777 | 1280 | // a new region to scan |
ysr@777 | 1281 | void setup_for_region(HeapRegion* hr); |
ysr@777 | 1282 | // it brings up-to-date the limit of the region |
ysr@777 | 1283 | void update_region_limit(); |
ysr@777 | 1284 | |
ysr@777 | 1285 | // called when either the words scanned or the refs visited limit |
ysr@777 | 1286 | // has been reached |
ysr@777 | 1287 | void reached_limit(); |
ysr@777 | 1288 | // recalculates the words scanned and refs visited limits |
ysr@777 | 1289 | void recalculate_limits(); |
ysr@777 | 1290 | // decreases the words scanned and refs visited limits when we reach |
ysr@777 | 1291 | // an expensive operation |
ysr@777 | 1292 | void decrease_limits(); |
ysr@777 | 1293 | // it checks whether the words scanned or refs visited reached their |
ysr@777 | 1294 | // respective limit and calls reached_limit() if they have |
ysr@777 | 1295 | void check_limits() { |
ysr@777 | 1296 | if (_words_scanned >= _words_scanned_limit || |
tonyp@2973 | 1297 | _refs_reached >= _refs_reached_limit) { |
ysr@777 | 1298 | reached_limit(); |
tonyp@2973 | 1299 | } |
ysr@777 | 1300 | } |
ysr@777 | 1301 | // this is supposed to be called regularly during a marking step as |
ysr@777 | 1302 | // it checks a bunch of conditions that might cause the marking step |
ysr@777 | 1303 | // to abort |
ysr@777 | 1304 | void regular_clock_call(); |
ysr@777 | 1305 | bool concurrent() { return _concurrent; } |
ysr@777 | 1306 | |
ysr@777 | 1307 | public: |
ysr@777 | 1308 | // It resets the task; it should be called right at the beginning of |
ysr@777 | 1309 | // a marking phase. |
ysr@777 | 1310 | void reset(CMBitMap* _nextMarkBitMap); |
ysr@777 | 1311 | // it clears all the fields that correspond to a claimed region. |
ysr@777 | 1312 | void clear_region_fields(); |
ysr@777 | 1313 | |
ysr@777 | 1314 | void set_concurrent(bool concurrent) { _concurrent = concurrent; } |
ysr@777 | 1315 | |
ysr@777 | 1316 | // The main method of this class which performs a marking step |
ysr@777 | 1317 | // trying not to exceed the given duration. However, it might exit |
ysr@777 | 1318 | // prematurely, according to some conditions (i.e. SATB buffers are |
ysr@777 | 1319 | // available for processing). |
johnc@2494 | 1320 | void do_marking_step(double target_ms, bool do_stealing, bool do_termination); |
ysr@777 | 1321 | |
ysr@777 | 1322 | // These two calls start and stop the timer |
ysr@777 | 1323 | void record_start_time() { |
ysr@777 | 1324 | _elapsed_time_ms = os::elapsedTime() * 1000.0; |
ysr@777 | 1325 | } |
ysr@777 | 1326 | void record_end_time() { |
ysr@777 | 1327 | _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; |
ysr@777 | 1328 | } |
ysr@777 | 1329 | |
ysr@777 | 1330 | // returns the task ID |
ysr@777 | 1331 | int task_id() { return _task_id; } |
ysr@777 | 1332 | |
ysr@777 | 1333 | // From TerminatorTerminator. It determines whether this task should |
ysr@777 | 1334 | // exit the termination protocol after it's entered it. |
ysr@777 | 1335 | virtual bool should_exit_termination(); |
ysr@777 | 1336 | |
johnc@2910 | 1337 | // Resets the local region fields after a task has finished scanning a |
johnc@2910 | 1338 | // region; or when they have become stale as a result of the region |
johnc@2910 | 1339 | // being evacuated. |
johnc@2910 | 1340 | void giveup_current_region(); |
johnc@2910 | 1341 | |
ysr@777 | 1342 | HeapWord* finger() { return _finger; } |
ysr@777 | 1343 | |
ysr@777 | 1344 | bool has_aborted() { return _has_aborted; } |
ysr@777 | 1345 | void set_has_aborted() { _has_aborted = true; } |
ysr@777 | 1346 | void clear_has_aborted() { _has_aborted = false; } |
johnc@2494 | 1347 | bool has_timed_out() { return _has_timed_out; } |
johnc@2494 | 1348 | bool claimed() { return _claimed; } |
ysr@777 | 1349 | |
johnc@2190 | 1350 | // Support routines for the partially scanned region that may be |
johnc@2190 | 1351 | // recorded as a result of aborting while draining the CMRegionStack |
johnc@2190 | 1352 | MemRegion aborted_region() { return _aborted_region; } |
johnc@2190 | 1353 | void set_aborted_region(MemRegion mr) |
johnc@2190 | 1354 | { _aborted_region = mr; } |
johnc@2190 | 1355 | |
johnc@2190 | 1356 | // Clears any recorded partially scanned region |
johnc@2190 | 1357 | void clear_aborted_region() { set_aborted_region(MemRegion()); } |
johnc@2190 | 1358 | |
tonyp@2968 | 1359 | void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); |
ysr@777 | 1360 | |
ysr@777 | 1361 | // It grays the object by marking it and, if necessary, pushing it |
ysr@777 | 1362 | // on the local queue |
tonyp@2968 | 1363 | inline void deal_with_reference(oop obj); |
ysr@777 | 1364 | |
ysr@777 | 1365 | // It scans an object and visits its children. |
tonyp@2968 | 1366 | void scan_object(oop obj); |
ysr@777 | 1367 | |
ysr@777 | 1368 | // It pushes an object on the local queue. |
tonyp@2968 | 1369 | inline void push(oop obj); |
ysr@777 | 1370 | |
ysr@777 | 1371 | // These two move entries to/from the global stack. |
ysr@777 | 1372 | void move_entries_to_global_stack(); |
ysr@777 | 1373 | void get_entries_from_global_stack(); |
ysr@777 | 1374 | |
ysr@777 | 1375 | // It pops and scans objects from the local queue. If partially is |
ysr@777 | 1376 | // true, then it stops when the queue size is of a given limit. If |
ysr@777 | 1377 | // partially is false, then it stops when the queue is empty. |
ysr@777 | 1378 | void drain_local_queue(bool partially); |
ysr@777 | 1379 | // It moves entries from the global stack to the local queue and |
ysr@777 | 1380 | // drains the local queue. If partially is true, then it stops when |
ysr@777 | 1381 | // both the global stack and the local queue reach a given size. If |
ysr@777 | 1382 | // partially if false, it tries to empty them totally. |
ysr@777 | 1383 | void drain_global_stack(bool partially); |
ysr@777 | 1384 | // It keeps picking SATB buffers and processing them until no SATB |
ysr@777 | 1385 | // buffers are available. |
ysr@777 | 1386 | void drain_satb_buffers(); |
tonyp@3416 | 1387 | |
ysr@777 | 1388 | // It keeps popping regions from the region stack and processing |
ysr@777 | 1389 | // them until the region stack is empty. |
ysr@777 | 1390 | void drain_region_stack(BitMapClosure* closure); |
ysr@777 | 1391 | |
ysr@777 | 1392 | // moves the local finger to a new location |
ysr@777 | 1393 | inline void move_finger_to(HeapWord* new_finger) { |
tonyp@1458 | 1394 | assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); |
ysr@777 | 1395 | _finger = new_finger; |
ysr@777 | 1396 | } |
ysr@777 | 1397 | |
ysr@777 | 1398 | // moves the region finger to a new location |
ysr@777 | 1399 | inline void move_region_finger_to(HeapWord* new_finger) { |
tonyp@1458 | 1400 | assert(new_finger < _cm->finger(), "invariant"); |
ysr@777 | 1401 | _region_finger = new_finger; |
ysr@777 | 1402 | } |
ysr@777 | 1403 | |
ysr@777 | 1404 | CMTask(int task_num, ConcurrentMark *cm, |
johnc@3463 | 1405 | size_t* marked_bytes, BitMap* card_bm, |
ysr@777 | 1406 | CMTaskQueue* task_queue, CMTaskQueueSet* task_queues); |
ysr@777 | 1407 | |
ysr@777 | 1408 | // it prints statistics associated with this task |
ysr@777 | 1409 | void print_stats(); |
ysr@777 | 1410 | |
ysr@777 | 1411 | #if _MARKING_STATS_ |
ysr@777 | 1412 | void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; } |
ysr@777 | 1413 | #endif // _MARKING_STATS_ |
ysr@777 | 1414 | }; |
stefank@2314 | 1415 | |
tonyp@2717 | 1416 | // Class that's used to to print out per-region liveness |
tonyp@2717 | 1417 | // information. It's currently used at the end of marking and also |
tonyp@2717 | 1418 | // after we sort the old regions at the end of the cleanup operation. |
tonyp@2717 | 1419 | class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { |
tonyp@2717 | 1420 | private: |
tonyp@2717 | 1421 | outputStream* _out; |
tonyp@2717 | 1422 | |
tonyp@2717 | 1423 | // Accumulators for these values. |
tonyp@2717 | 1424 | size_t _total_used_bytes; |
tonyp@2717 | 1425 | size_t _total_capacity_bytes; |
tonyp@2717 | 1426 | size_t _total_prev_live_bytes; |
tonyp@2717 | 1427 | size_t _total_next_live_bytes; |
tonyp@2717 | 1428 | |
tonyp@2717 | 1429 | // These are set up when we come across a "stars humongous" region |
tonyp@2717 | 1430 | // (as this is where most of this information is stored, not in the |
tonyp@2717 | 1431 | // subsequent "continues humongous" regions). After that, for every |
tonyp@2717 | 1432 | // region in a given humongous region series we deduce the right |
tonyp@2717 | 1433 | // values for it by simply subtracting the appropriate amount from |
tonyp@2717 | 1434 | // these fields. All these values should reach 0 after we've visited |
tonyp@2717 | 1435 | // the last region in the series. |
tonyp@2717 | 1436 | size_t _hum_used_bytes; |
tonyp@2717 | 1437 | size_t _hum_capacity_bytes; |
tonyp@2717 | 1438 | size_t _hum_prev_live_bytes; |
tonyp@2717 | 1439 | size_t _hum_next_live_bytes; |
tonyp@2717 | 1440 | |
tonyp@2717 | 1441 | static double perc(size_t val, size_t total) { |
tonyp@2717 | 1442 | if (total == 0) { |
tonyp@2717 | 1443 | return 0.0; |
tonyp@2717 | 1444 | } else { |
tonyp@2717 | 1445 | return 100.0 * ((double) val / (double) total); |
tonyp@2717 | 1446 | } |
tonyp@2717 | 1447 | } |
tonyp@2717 | 1448 | |
tonyp@2717 | 1449 | static double bytes_to_mb(size_t val) { |
tonyp@2717 | 1450 | return (double) val / (double) M; |
tonyp@2717 | 1451 | } |
tonyp@2717 | 1452 | |
tonyp@2717 | 1453 | // See the .cpp file. |
tonyp@2717 | 1454 | size_t get_hum_bytes(size_t* hum_bytes); |
tonyp@2717 | 1455 | void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes, |
tonyp@2717 | 1456 | size_t* prev_live_bytes, size_t* next_live_bytes); |
tonyp@2717 | 1457 | |
tonyp@2717 | 1458 | public: |
tonyp@2717 | 1459 | // The header and footer are printed in the constructor and |
tonyp@2717 | 1460 | // destructor respectively. |
tonyp@2717 | 1461 | G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name); |
tonyp@2717 | 1462 | virtual bool doHeapRegion(HeapRegion* r); |
tonyp@2717 | 1463 | ~G1PrintRegionLivenessInfoClosure(); |
tonyp@2717 | 1464 | }; |
tonyp@2717 | 1465 | |
stefank@2314 | 1466 | #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP |