ysr@777: /* johnc@4386: * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. ysr@777: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. ysr@777: * ysr@777: * This code is free software; you can redistribute it and/or modify it ysr@777: * under the terms of the GNU General Public License version 2 only, as ysr@777: * published by the Free Software Foundation. ysr@777: * ysr@777: * This code is distributed in the hope that it will be useful, but WITHOUT ysr@777: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or ysr@777: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License ysr@777: * version 2 for more details (a copy is included in the LICENSE file that ysr@777: * accompanied this code). ysr@777: * ysr@777: * You should have received a copy of the GNU General Public License version ysr@777: * 2 along with this work; if not, write to the Free Software Foundation, ysr@777: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. ysr@777: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. ysr@777: * ysr@777: */ ysr@777: stefank@2314: #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP stefank@2314: #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP stefank@2314: stefank@6992: #include "classfile/javaClasses.hpp" brutisso@6385: #include "gc_implementation/g1/heapRegionSet.hpp" brutisso@6904: #include "gc_implementation/shared/gcId.hpp" stefank@2314: #include "utilities/taskqueue.hpp" stefank@2314: ysr@777: class G1CollectedHeap; ysr@777: class CMTask; zgu@3900: typedef GenericTaskQueue CMTaskQueue; zgu@3900: typedef GenericTaskQueueSet CMTaskQueueSet; ysr@777: johnc@2379: // Closure used by CM during concurrent reference discovery johnc@2379: // and reference processing (during remarking) to determine johnc@2379: // if a particular object is alive. It is primarily used johnc@2379: // to determine if referents of discovered reference objects johnc@2379: // are alive. An instance is also embedded into the johnc@2379: // reference processor as the _is_alive_non_header field johnc@2379: class G1CMIsAliveClosure: public BoolObjectClosure { johnc@2379: G1CollectedHeap* _g1; johnc@2379: public: tonyp@3691: G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { } johnc@2379: johnc@2379: bool do_object_b(oop obj); johnc@2379: }; johnc@2379: ysr@777: // A generic CM bit map. This is essentially a wrapper around the BitMap ysr@777: // class, with one bit per (1<<_shifter) HeapWords. ysr@777: apetrusenko@984: class CMBitMapRO VALUE_OBJ_CLASS_SPEC { ysr@777: protected: ysr@777: HeapWord* _bmStartWord; // base address of range covered by map ysr@777: size_t _bmWordSize; // map size (in #HeapWords covered) ysr@777: const int _shifter; // map to char or bit ysr@777: VirtualSpace _virtual_space; // underlying the bit map ysr@777: BitMap _bm; // the bit map itself ysr@777: ysr@777: public: ysr@777: // constructor johnc@4333: CMBitMapRO(int shifter); ysr@777: ysr@777: enum { do_yield = true }; ysr@777: ysr@777: // inquiries ysr@777: HeapWord* startWord() const { return _bmStartWord; } ysr@777: size_t sizeInWords() const { return _bmWordSize; } ysr@777: // the following is one past the last word in space ysr@777: HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } ysr@777: ysr@777: // read marks ysr@777: ysr@777: bool isMarked(HeapWord* addr) const { ysr@777: assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), ysr@777: "outside underlying space?"); ysr@777: return _bm.at(heapWordToOffset(addr)); ysr@777: } ysr@777: ysr@777: // iteration johnc@3454: inline bool iterate(BitMapClosure* cl, MemRegion mr); johnc@3454: inline bool iterate(BitMapClosure* cl); ysr@777: ysr@777: // Return the address corresponding to the next marked bit at or after ysr@777: // "addr", and before "limit", if "limit" is non-NULL. If there is no ysr@777: // such bit, returns "limit" if that is non-NULL, or else "endWord()". stefank@6992: HeapWord* getNextMarkedWordAddress(const HeapWord* addr, stefank@6992: const HeapWord* limit = NULL) const; ysr@777: // Return the address corresponding to the next unmarked bit at or after ysr@777: // "addr", and before "limit", if "limit" is non-NULL. If there is no ysr@777: // such bit, returns "limit" if that is non-NULL, or else "endWord()". stefank@6992: HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr, stefank@6992: const HeapWord* limit = NULL) const; ysr@777: ysr@777: // conversion utilities ysr@777: HeapWord* offsetToHeapWord(size_t offset) const { ysr@777: return _bmStartWord + (offset << _shifter); ysr@777: } stefank@6992: size_t heapWordToOffset(const HeapWord* addr) const { ysr@777: return pointer_delta(addr, _bmStartWord) >> _shifter; ysr@777: } ysr@777: int heapWordDiffToOffsetDiff(size_t diff) const; tamao@4733: tamao@4733: // The argument addr should be the start address of a valid object tamao@4733: HeapWord* nextObject(HeapWord* addr) { tamao@4733: oop obj = (oop) addr; tamao@4733: HeapWord* res = addr + obj->size(); tamao@4733: assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity"); tamao@4733: return res; ysr@777: } ysr@777: stefank@4904: void print_on_error(outputStream* st, const char* prefix) const; stefank@4904: ysr@777: // debugging ysr@777: NOT_PRODUCT(bool covers(ReservedSpace rs) const;) ysr@777: }; ysr@777: ysr@777: class CMBitMap : public CMBitMapRO { ysr@777: ysr@777: public: ysr@777: // constructor johnc@4333: CMBitMap(int shifter) : johnc@4333: CMBitMapRO(shifter) {} johnc@4333: johnc@4333: // Allocates the back store for the marking bitmap johnc@4333: bool allocate(ReservedSpace heap_rs); ysr@777: ysr@777: // write marks ysr@777: void mark(HeapWord* addr) { ysr@777: assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), ysr@777: "outside underlying space?"); tonyp@2968: _bm.set_bit(heapWordToOffset(addr)); ysr@777: } ysr@777: void clear(HeapWord* addr) { ysr@777: assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), ysr@777: "outside underlying space?"); tonyp@2968: _bm.clear_bit(heapWordToOffset(addr)); ysr@777: } ysr@777: bool parMark(HeapWord* addr) { ysr@777: assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), ysr@777: "outside underlying space?"); tonyp@2968: return _bm.par_set_bit(heapWordToOffset(addr)); ysr@777: } ysr@777: bool parClear(HeapWord* addr) { ysr@777: assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), ysr@777: "outside underlying space?"); tonyp@2968: return _bm.par_clear_bit(heapWordToOffset(addr)); ysr@777: } ysr@777: void markRange(MemRegion mr); ysr@777: void clearAll(); ysr@777: void clearRange(MemRegion mr); ysr@777: ysr@777: // Starting at the bit corresponding to "addr" (inclusive), find the next ysr@777: // "1" bit, if any. This bit starts some run of consecutive "1"'s; find ysr@777: // the end of this run (stopping at "end_addr"). Return the MemRegion ysr@777: // covering from the start of the region corresponding to the first bit ysr@777: // of the run to the end of the region corresponding to the last bit of ysr@777: // the run. If there is no "1" bit at or after "addr", return an empty ysr@777: // MemRegion. ysr@777: MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr); ysr@777: }; ysr@777: johnc@4333: // Represents a marking stack used by ConcurrentMarking in the G1 collector. apetrusenko@984: class CMMarkStack VALUE_OBJ_CLASS_SPEC { johnc@4333: VirtualSpace _virtual_space; // Underlying backing store for actual stack ysr@777: ConcurrentMark* _cm; johnc@4787: oop* _base; // bottom of stack johnc@4333: jint _index; // one more than last occupied index johnc@4333: jint _capacity; // max #elements johnc@4333: jint _saved_index; // value of _index saved at start of GC johnc@4333: NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run ysr@777: johnc@4333: bool _overflow; johnc@4333: bool _should_expand; ysr@777: DEBUG_ONLY(bool _drain_in_progress;) ysr@777: DEBUG_ONLY(bool _drain_in_progress_yields;) ysr@777: ysr@777: public: ysr@777: CMMarkStack(ConcurrentMark* cm); ysr@777: ~CMMarkStack(); ysr@777: johnc@4333: #ifndef PRODUCT johnc@4333: jint max_depth() const { johnc@4333: return _max_depth; johnc@4333: } johnc@4333: #endif johnc@4333: johnc@4333: bool allocate(size_t capacity); ysr@777: ysr@777: oop pop() { ysr@777: if (!isEmpty()) { ysr@777: return _base[--_index] ; ysr@777: } ysr@777: return NULL; ysr@777: } ysr@777: ysr@777: // If overflow happens, don't do the push, and record the overflow. ysr@777: // *Requires* that "ptr" is already marked. ysr@777: void push(oop ptr) { ysr@777: if (isFull()) { ysr@777: // Record overflow. ysr@777: _overflow = true; ysr@777: return; ysr@777: } else { ysr@777: _base[_index++] = ptr; ysr@777: NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index)); ysr@777: } ysr@777: } ysr@777: // Non-block impl. Note: concurrency is allowed only with other ysr@777: // "par_push" operations, not with "pop" or "drain". We would need ysr@777: // parallel versions of them if such concurrency was desired. ysr@777: void par_push(oop ptr); ysr@777: ysr@777: // Pushes the first "n" elements of "ptr_arr" on the stack. ysr@777: // Non-block impl. Note: concurrency is allowed only with other ysr@777: // "par_adjoin_arr" or "push" operations, not with "pop" or "drain". ysr@777: void par_adjoin_arr(oop* ptr_arr, int n); ysr@777: ysr@777: // Pushes the first "n" elements of "ptr_arr" on the stack. ysr@777: // Locking impl: concurrency is allowed only with ysr@777: // "par_push_arr" and/or "par_pop_arr" operations, which use the same ysr@777: // locking strategy. ysr@777: void par_push_arr(oop* ptr_arr, int n); ysr@777: ysr@777: // If returns false, the array was empty. Otherwise, removes up to "max" ysr@777: // elements from the stack, and transfers them to "ptr_arr" in an ysr@777: // unspecified order. The actual number transferred is given in "n" ("n ysr@777: // == 0" is deliberately redundant with the return value.) Locking impl: ysr@777: // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" ysr@777: // operations, which use the same locking strategy. ysr@777: bool par_pop_arr(oop* ptr_arr, int max, int* n); ysr@777: ysr@777: // Drain the mark stack, applying the given closure to all fields of ysr@777: // objects on the stack. (That is, continue until the stack is empty, ysr@777: // even if closure applications add entries to the stack.) The "bm" ysr@777: // argument, if non-null, may be used to verify that only marked objects ysr@777: // are on the mark stack. If "yield_after" is "true", then the ysr@777: // concurrent marker performing the drain offers to yield after ysr@777: // processing each object. If a yield occurs, stops the drain operation ysr@777: // and returns false. Otherwise, returns true. ysr@777: template ysr@777: bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false); ysr@777: ysr@777: bool isEmpty() { return _index == 0; } ysr@777: bool isFull() { return _index == _capacity; } johnc@4333: int maxElems() { return _capacity; } ysr@777: ysr@777: bool overflow() { return _overflow; } ysr@777: void clear_overflow() { _overflow = false; } ysr@777: johnc@4333: bool should_expand() const { return _should_expand; } johnc@4333: void set_should_expand(); johnc@4333: johnc@4333: // Expand the stack, typically in response to an overflow condition johnc@4333: void expand(); johnc@4333: ysr@777: int size() { return _index; } ysr@777: ysr@777: void setEmpty() { _index = 0; clear_overflow(); } ysr@777: tonyp@3416: // Record the current index. tonyp@3416: void note_start_of_gc(); tonyp@3416: tonyp@3416: // Make sure that we have not added any entries to the stack during GC. tonyp@3416: void note_end_of_gc(); tonyp@3416: ysr@777: // iterate over the oops in the mark stack, up to the bound recorded via ysr@777: // the call above. ysr@777: void oops_do(OopClosure* f); ysr@777: }; ysr@777: tonyp@2848: class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC { tonyp@2848: private: tonyp@2848: #ifndef PRODUCT tonyp@2848: uintx _num_remaining; tonyp@2848: bool _force; tonyp@2848: #endif // !defined(PRODUCT) tonyp@2848: tonyp@2848: public: tonyp@2848: void init() PRODUCT_RETURN; tonyp@2848: void update() PRODUCT_RETURN; tonyp@2848: bool should_force() PRODUCT_RETURN_( return false; ); tonyp@2848: }; tonyp@2848: ysr@777: // this will enable a variety of different statistics per GC task ysr@777: #define _MARKING_STATS_ 0 ysr@777: // this will enable the higher verbose levels ysr@777: #define _MARKING_VERBOSE_ 0 ysr@777: ysr@777: #if _MARKING_STATS_ ysr@777: #define statsOnly(statement) \ ysr@777: do { \ ysr@777: statement ; \ ysr@777: } while (0) ysr@777: #else // _MARKING_STATS_ ysr@777: #define statsOnly(statement) \ ysr@777: do { \ ysr@777: } while (0) ysr@777: #endif // _MARKING_STATS_ ysr@777: ysr@777: typedef enum { ysr@777: no_verbose = 0, // verbose turned off ysr@777: stats_verbose, // only prints stats at the end of marking ysr@777: low_verbose, // low verbose, mostly per region and per major event ysr@777: medium_verbose, // a bit more detailed than low ysr@777: high_verbose // per object verbose ysr@777: } CMVerboseLevel; ysr@777: tonyp@3464: class YoungList; tonyp@3464: tonyp@3464: // Root Regions are regions that are not empty at the beginning of a tonyp@3464: // marking cycle and which we might collect during an evacuation pause tonyp@3464: // while the cycle is active. Given that, during evacuation pauses, we tonyp@3464: // do not copy objects that are explicitly marked, what we have to do tonyp@3464: // for the root regions is to scan them and mark all objects reachable tonyp@3464: // from them. According to the SATB assumptions, we only need to visit tonyp@3464: // each object once during marking. So, as long as we finish this scan tonyp@3464: // before the next evacuation pause, we can copy the objects from the tonyp@3464: // root regions without having to mark them or do anything else to them. tonyp@3464: // tonyp@3464: // Currently, we only support root region scanning once (at the start tonyp@3464: // of the marking cycle) and the root regions are all the survivor tonyp@3464: // regions populated during the initial-mark pause. tonyp@3464: class CMRootRegions VALUE_OBJ_CLASS_SPEC { tonyp@3464: private: tonyp@3464: YoungList* _young_list; tonyp@3464: ConcurrentMark* _cm; tonyp@3464: tonyp@3464: volatile bool _scan_in_progress; tonyp@3464: volatile bool _should_abort; tonyp@3464: HeapRegion* volatile _next_survivor; tonyp@3464: tonyp@3464: public: tonyp@3464: CMRootRegions(); tonyp@3464: // We actually do most of the initialization in this method. tonyp@3464: void init(G1CollectedHeap* g1h, ConcurrentMark* cm); tonyp@3464: tonyp@3464: // Reset the claiming / scanning of the root regions. tonyp@3464: void prepare_for_scan(); tonyp@3464: tonyp@3464: // Forces get_next() to return NULL so that the iteration aborts early. tonyp@3464: void abort() { _should_abort = true; } tonyp@3464: tonyp@3464: // Return true if the CM thread are actively scanning root regions, tonyp@3464: // false otherwise. tonyp@3464: bool scan_in_progress() { return _scan_in_progress; } tonyp@3464: tonyp@3464: // Claim the next root region to scan atomically, or return NULL if tonyp@3464: // all have been claimed. tonyp@3464: HeapRegion* claim_next(); tonyp@3464: tonyp@3464: // Flag that we're done with root region scanning and notify anyone tonyp@3464: // who's waiting on it. If aborted is false, assume that all regions tonyp@3464: // have been claimed. tonyp@3464: void scan_finished(); tonyp@3464: tonyp@3464: // If CM threads are still scanning root regions, wait until they tonyp@3464: // are done. Return true if we had to wait, false otherwise. tonyp@3464: bool wait_until_scan_finished(); tonyp@3464: }; ysr@777: ysr@777: class ConcurrentMarkThread; ysr@777: zgu@3900: class ConcurrentMark: public CHeapObj { johnc@4333: friend class CMMarkStack; ysr@777: friend class ConcurrentMarkThread; ysr@777: friend class CMTask; ysr@777: friend class CMBitMapClosure; ysr@777: friend class CMGlobalObjectClosure; ysr@777: friend class CMRemarkTask; ysr@777: friend class CMConcurrentMarkingTask; ysr@777: friend class G1ParNoteEndTask; ysr@777: friend class CalcLiveObjectsClosure; johnc@3175: friend class G1CMRefProcTaskProxy; johnc@3175: friend class G1CMRefProcTaskExecutor; johnc@4555: friend class G1CMKeepAliveAndDrainClosure; johnc@4555: friend class G1CMDrainMarkingStackClosure; ysr@777: ysr@777: protected: ysr@777: ConcurrentMarkThread* _cmThread; // the thread doing the work ysr@777: G1CollectedHeap* _g1h; // the heap. jmasa@3357: uint _parallel_marking_threads; // the number of marking jmasa@3294: // threads we're use jmasa@3357: uint _max_parallel_marking_threads; // max number of marking jmasa@3294: // threads we'll ever use ysr@777: double _sleep_factor; // how much we have to sleep, with ysr@777: // respect to the work we just did, to ysr@777: // meet the marking overhead goal ysr@777: double _marking_task_overhead; // marking target overhead for ysr@777: // a single task ysr@777: ysr@777: // same as the two above, but for the cleanup task ysr@777: double _cleanup_sleep_factor; ysr@777: double _cleanup_task_overhead; ysr@777: tonyp@2472: FreeRegionList _cleanup_list; ysr@777: brutisso@3455: // Concurrent marking support structures ysr@777: CMBitMap _markBitMap1; ysr@777: CMBitMap _markBitMap2; ysr@777: CMBitMapRO* _prevMarkBitMap; // completed mark bitmap ysr@777: CMBitMap* _nextMarkBitMap; // under-construction mark bitmap ysr@777: ysr@777: BitMap _region_bm; ysr@777: BitMap _card_bm; ysr@777: ysr@777: // Heap bounds ysr@777: HeapWord* _heap_start; ysr@777: HeapWord* _heap_end; ysr@777: tonyp@3464: // Root region tracking and claiming. tonyp@3464: CMRootRegions _root_regions; tonyp@3464: ysr@777: // For gray objects ysr@777: CMMarkStack _markStack; // Grey objects behind global finger. ysr@777: HeapWord* volatile _finger; // the global finger, region aligned, ysr@777: // always points to the end of the ysr@777: // last claimed region ysr@777: ysr@777: // marking tasks johnc@4173: uint _max_worker_id;// maximum worker id jmasa@3357: uint _active_tasks; // task num currently active johnc@4173: CMTask** _tasks; // task queue array (max_worker_id len) ysr@777: CMTaskQueueSet* _task_queues; // task queue set ysr@777: ParallelTaskTerminator _terminator; // for termination ysr@777: ysr@777: // Two sync barriers that are used to synchronise tasks when an ysr@777: // overflow occurs. The algorithm is the following. All tasks enter ysr@777: // the first one to ensure that they have all stopped manipulating ysr@777: // the global data structures. After they exit it, they re-initialise ysr@777: // their data structures and task 0 re-initialises the global data ysr@777: // structures. Then, they enter the second sync barrier. This ysr@777: // ensure, that no task starts doing work before all data ysr@777: // structures (local and global) have been re-initialised. When they ysr@777: // exit it, they are free to start working again. ysr@777: WorkGangBarrierSync _first_overflow_barrier_sync; ysr@777: WorkGangBarrierSync _second_overflow_barrier_sync; ysr@777: ysr@777: // this is set by any task, when an overflow on the global data ysr@777: // structures is detected. ysr@777: volatile bool _has_overflown; ysr@777: // true: marking is concurrent, false: we're in remark ysr@777: volatile bool _concurrent; ysr@777: // set at the end of a Full GC so that marking aborts ysr@777: volatile bool _has_aborted; brutisso@6904: GCId _aborted_gc_id; johnc@2190: ysr@777: // used when remark aborts due to an overflow to indicate that ysr@777: // another concurrent marking phase should start ysr@777: volatile bool _restart_for_overflow; ysr@777: ysr@777: // This is true from the very start of concurrent marking until the ysr@777: // point when all the tasks complete their work. It is really used ysr@777: // to determine the points between the end of concurrent marking and ysr@777: // time of remark. ysr@777: volatile bool _concurrent_marking_in_progress; ysr@777: ysr@777: // verbose level ysr@777: CMVerboseLevel _verbose_level; ysr@777: ysr@777: // All of these times are in ms. ysr@777: NumberSeq _init_times; ysr@777: NumberSeq _remark_times; ysr@777: NumberSeq _remark_mark_times; ysr@777: NumberSeq _remark_weak_ref_times; ysr@777: NumberSeq _cleanup_times; ysr@777: double _total_counting_time; ysr@777: double _total_rs_scrub_time; ysr@777: ysr@777: double* _accum_task_vtime; // accumulated task vtime ysr@777: jmasa@3294: FlexibleWorkGang* _parallel_workers; ysr@777: tonyp@2848: ForceOverflowSettings _force_overflow_conc; tonyp@2848: ForceOverflowSettings _force_overflow_stw; tonyp@2848: stefank@6992: void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes); ysr@777: void weakRefsWork(bool clear_all_soft_refs); ysr@777: ysr@777: void swapMarkBitMaps(); ysr@777: ysr@777: // It resets the global marking data structures, as well as the ysr@777: // task local ones; should be called during initial mark. ysr@777: void reset(); johnc@4386: johnc@4386: // Resets all the marking data structures. Called when we have to restart johnc@4386: // marking or when marking completes (via set_non_marking_state below). johnc@4386: void reset_marking_state(bool clear_overflow = true); johnc@4386: johnc@4386: // We do this after we're done with marking so that the marking data johnc@4386: // structures are initialised to a sensible and predictable state. johnc@4386: void set_non_marking_state(); ysr@777: johnc@4788: // Called to indicate how many threads are currently active. johnc@4788: void set_concurrency(uint active_tasks); johnc@4788: ysr@777: // It should be called to indicate which phase we're in (concurrent ysr@777: // mark or remark) and how many threads are currently active. johnc@4788: void set_concurrency_and_phase(uint active_tasks, bool concurrent); ysr@777: ysr@777: // prints all gathered CM-related statistics ysr@777: void print_stats(); ysr@777: tonyp@2472: bool cleanup_list_is_empty() { tonyp@2472: return _cleanup_list.is_empty(); tonyp@2472: } tonyp@2472: ysr@777: // accessor methods johnc@4549: uint parallel_marking_threads() const { return _parallel_marking_threads; } johnc@4549: uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;} johnc@4549: double sleep_factor() { return _sleep_factor; } johnc@4549: double marking_task_overhead() { return _marking_task_overhead;} johnc@4549: double cleanup_sleep_factor() { return _cleanup_sleep_factor; } johnc@4549: double cleanup_task_overhead() { return _cleanup_task_overhead;} ysr@777: johnc@4549: bool use_parallel_marking_threads() const { johnc@4549: assert(parallel_marking_threads() <= johnc@4549: max_parallel_marking_threads(), "sanity"); johnc@4549: assert((_parallel_workers == NULL && parallel_marking_threads() == 0) || johnc@4549: parallel_marking_threads() > 0, johnc@4549: "parallel workers not set up correctly"); johnc@4549: return _parallel_workers != NULL; johnc@4549: } johnc@4549: johnc@4549: HeapWord* finger() { return _finger; } johnc@4549: bool concurrent() { return _concurrent; } johnc@4549: uint active_tasks() { return _active_tasks; } johnc@4549: ParallelTaskTerminator* terminator() { return &_terminator; } ysr@777: ysr@777: // It claims the next available region to be scanned by a marking johnc@4173: // task/thread. It might return NULL if the next region is empty or johnc@4173: // we have run out of regions. In the latter case, out_of_regions() ysr@777: // determines whether we've really run out of regions or the task johnc@4173: // should call claim_region() again. This might seem a bit ysr@777: // awkward. Originally, the code was written so that claim_region() ysr@777: // either successfully returned with a non-empty region or there ysr@777: // were no more regions to be claimed. The problem with this was ysr@777: // that, in certain circumstances, it iterated over large chunks of ysr@777: // the heap finding only empty regions and, while it was working, it ysr@777: // was preventing the calling task to call its regular clock ysr@777: // method. So, this way, each task will spend very little time in ysr@777: // claim_region() and is allowed to call the regular clock method ysr@777: // frequently. johnc@4173: HeapRegion* claim_region(uint worker_id); ysr@777: pliden@6693: // It determines whether we've run out of regions to scan. Note that pliden@6693: // the finger can point past the heap end in case the heap was expanded pliden@6693: // to satisfy an allocation without doing a GC. This is fine, because all pliden@6693: // objects in those regions will be considered live anyway because of pliden@6693: // SATB guarantees (i.e. their TAMS will be equal to bottom). pliden@6693: bool out_of_regions() { return _finger >= _heap_end; } ysr@777: ysr@777: // Returns the task with the given id ysr@777: CMTask* task(int id) { tonyp@1458: assert(0 <= id && id < (int) _active_tasks, tonyp@1458: "task id not within active bounds"); ysr@777: return _tasks[id]; ysr@777: } ysr@777: ysr@777: // Returns the task queue with the given id ysr@777: CMTaskQueue* task_queue(int id) { tonyp@1458: assert(0 <= id && id < (int) _active_tasks, tonyp@1458: "task queue id not within active bounds"); ysr@777: return (CMTaskQueue*) _task_queues->queue(id); ysr@777: } ysr@777: ysr@777: // Returns the task queue set ysr@777: CMTaskQueueSet* task_queues() { return _task_queues; } ysr@777: ysr@777: // Access / manipulation of the overflow flag which is set to tonyp@3691: // indicate that the global stack has overflown ysr@777: bool has_overflown() { return _has_overflown; } ysr@777: void set_has_overflown() { _has_overflown = true; } ysr@777: void clear_has_overflown() { _has_overflown = false; } tonyp@3464: bool restart_for_overflow() { return _restart_for_overflow; } ysr@777: ysr@777: // Methods to enter the two overflow sync barriers johnc@4173: void enter_first_sync_barrier(uint worker_id); johnc@4173: void enter_second_sync_barrier(uint worker_id); ysr@777: tonyp@2848: ForceOverflowSettings* force_overflow_conc() { tonyp@2848: return &_force_overflow_conc; tonyp@2848: } tonyp@2848: tonyp@2848: ForceOverflowSettings* force_overflow_stw() { tonyp@2848: return &_force_overflow_stw; tonyp@2848: } tonyp@2848: tonyp@2848: ForceOverflowSettings* force_overflow() { tonyp@2848: if (concurrent()) { tonyp@2848: return force_overflow_conc(); tonyp@2848: } else { tonyp@2848: return force_overflow_stw(); tonyp@2848: } tonyp@2848: } tonyp@2848: johnc@3463: // Live Data Counting data structures... johnc@3463: // These data structures are initialized at the start of johnc@3463: // marking. They are written to while marking is active. johnc@3463: // They are aggregated during remark; the aggregated values johnc@3463: // are then used to populate the _region_bm, _card_bm, and johnc@3463: // the total live bytes, which are then subsequently updated johnc@3463: // during cleanup. johnc@3463: johnc@3463: // An array of bitmaps (one bit map per task). Each bitmap johnc@3463: // is used to record the cards spanned by the live objects johnc@3463: // marked by that task/worker. johnc@3463: BitMap* _count_card_bitmaps; johnc@3463: johnc@3463: // Used to record the number of marked live bytes johnc@3463: // (for each region, by worker thread). johnc@3463: size_t** _count_marked_bytes; johnc@3463: johnc@3463: // Card index of the bottom of the G1 heap. Used for biasing indices into johnc@3463: // the card bitmaps. johnc@3463: intptr_t _heap_bottom_card_num; johnc@3463: johnc@4333: // Set to true when initialization is complete johnc@4333: bool _completed_initialization; johnc@4333: ysr@777: public: ysr@777: // Manipulation of the global mark stack. ysr@777: // Notice that the first mark_stack_push is CAS-based, whereas the ysr@777: // two below are Mutex-based. This is OK since the first one is only ysr@777: // called during evacuation pauses and doesn't compete with the ysr@777: // other two (which are called by the marking tasks during ysr@777: // concurrent marking or remark). ysr@777: bool mark_stack_push(oop p) { ysr@777: _markStack.par_push(p); ysr@777: if (_markStack.overflow()) { ysr@777: set_has_overflown(); ysr@777: return false; ysr@777: } ysr@777: return true; ysr@777: } ysr@777: bool mark_stack_push(oop* arr, int n) { ysr@777: _markStack.par_push_arr(arr, n); ysr@777: if (_markStack.overflow()) { ysr@777: set_has_overflown(); ysr@777: return false; ysr@777: } ysr@777: return true; ysr@777: } ysr@777: void mark_stack_pop(oop* arr, int max, int* n) { ysr@777: _markStack.par_pop_arr(arr, max, n); ysr@777: } tonyp@2973: size_t mark_stack_size() { return _markStack.size(); } ysr@777: size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } tonyp@2973: bool mark_stack_overflow() { return _markStack.overflow(); } tonyp@2973: bool mark_stack_empty() { return _markStack.isEmpty(); } ysr@777: tonyp@3464: CMRootRegions* root_regions() { return &_root_regions; } tonyp@3464: ysr@777: bool concurrent_marking_in_progress() { ysr@777: return _concurrent_marking_in_progress; ysr@777: } ysr@777: void set_concurrent_marking_in_progress() { ysr@777: _concurrent_marking_in_progress = true; ysr@777: } ysr@777: void clear_concurrent_marking_in_progress() { ysr@777: _concurrent_marking_in_progress = false; ysr@777: } ysr@777: ysr@777: void update_accum_task_vtime(int i, double vtime) { ysr@777: _accum_task_vtime[i] += vtime; ysr@777: } ysr@777: ysr@777: double all_task_accum_vtime() { ysr@777: double ret = 0.0; johnc@4173: for (uint i = 0; i < _max_worker_id; ++i) ysr@777: ret += _accum_task_vtime[i]; ysr@777: return ret; ysr@777: } ysr@777: ysr@777: // Attempts to steal an object from the task queues of other tasks johnc@4173: bool try_stealing(uint worker_id, int* hash_seed, oop& obj) { johnc@4173: return _task_queues->steal(worker_id, hash_seed, obj); ysr@777: } ysr@777: johnc@4333: ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs); ysr@777: ~ConcurrentMark(); johnc@3463: ysr@777: ConcurrentMarkThread* cmThread() { return _cmThread; } ysr@777: ysr@777: CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } ysr@777: CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } ysr@777: jmasa@3294: // Returns the number of GC threads to be used in a concurrent jmasa@3294: // phase based on the number of GC threads being used in a STW jmasa@3294: // phase. jmasa@3357: uint scale_parallel_threads(uint n_par_threads); jmasa@3294: jmasa@3294: // Calculates the number of GC threads to be used in a concurrent phase. jmasa@3357: uint calc_parallel_marking_threads(); jmasa@3294: ysr@777: // The following three are interaction between CM and ysr@777: // G1CollectedHeap ysr@777: ysr@777: // This notifies CM that a root during initial-mark needs to be tonyp@3464: // grayed. It is MT-safe. word_size is the size of the object in tonyp@3464: // words. It is passed explicitly as sometimes we cannot calculate tonyp@3464: // it from the given object because it might be in an inconsistent tonyp@3464: // state (e.g., in to-space and being copied). So the caller is tonyp@3464: // responsible for dealing with this issue (e.g., get the size from tonyp@3464: // the from-space image when the to-space image might be tonyp@3464: // inconsistent) and always passing the size. hr is the region that tonyp@3464: // contains the object and it's passed optionally from callers who tonyp@3464: // might already have it (no point in recalculating it). tonyp@3464: inline void grayRoot(oop obj, size_t word_size, tonyp@3464: uint worker_id, HeapRegion* hr = NULL); tonyp@3416: tonyp@1823: // It iterates over the heap and for each object it comes across it tonyp@1823: // will dump the contents of its reference fields, as well as tonyp@1823: // liveness information for the object and its referents. The dump tonyp@1823: // will be written to a file with the following name: johnc@2969: // G1PrintReachableBaseFile + "." + str. johnc@2969: // vo decides whether the prev (vo == UsePrevMarking), the next johnc@2969: // (vo == UseNextMarking) marking information, or the mark word johnc@2969: // (vo == UseMarkWord) will be used to determine the liveness of johnc@2969: // each object / referent. johnc@2969: // If all is true, all objects in the heap will be dumped, otherwise johnc@2969: // only the live ones. In the dump the following symbols / breviations johnc@2969: // are used: tonyp@1823: // M : an explicitly live object (its bitmap bit is set) tonyp@1823: // > : an implicitly live object (over tams) tonyp@1823: // O : an object outside the G1 heap (typically: in the perm gen) tonyp@1823: // NOT : a reference field whose referent is not live tonyp@1823: // AND MARKED : indicates that an object is both explicitly and tonyp@1823: // implicitly live (it should be one or the other, not both) tonyp@1823: void print_reachable(const char* str, johnc@2969: VerifyOption vo, bool all) PRODUCT_RETURN; ysr@777: ysr@777: // Clear the next marking bitmap (will be called concurrently). ysr@777: void clearNextBitmap(); ysr@777: ysr@777: // These two do the work that needs to be done before and after the ysr@777: // initial root checkpoint. Since this checkpoint can be done at two ysr@777: // different points (i.e. an explicit pause or piggy-backed on a ysr@777: // young collection), then it's nice to be able to easily share the ysr@777: // pre/post code. It might be the case that we can put everything in ysr@777: // the post method. TP ysr@777: void checkpointRootsInitialPre(); ysr@777: void checkpointRootsInitialPost(); ysr@777: tonyp@3464: // Scan all the root regions and mark everything reachable from tonyp@3464: // them. tonyp@3464: void scanRootRegions(); tonyp@3464: tonyp@3464: // Scan a single root region and mark everything reachable from it. tonyp@3464: void scanRootRegion(HeapRegion* hr, uint worker_id); tonyp@3464: ysr@777: // Do concurrent phase of marking, to a tentative transitive closure. ysr@777: void markFromRoots(); ysr@777: ysr@777: void checkpointRootsFinal(bool clear_all_soft_refs); ysr@777: void checkpointRootsFinalWork(); ysr@777: void cleanup(); ysr@777: void completeCleanup(); ysr@777: ysr@777: // Mark in the previous bitmap. NB: this is usually read-only, so use ysr@777: // this carefully! tonyp@3416: inline void markPrev(oop p); johnc@3463: tonyp@3416: // Clears marks for all objects in the given range, for the prev, tonyp@3416: // next, or both bitmaps. NB: the previous bitmap is usually tonyp@3416: // read-only, so use this carefully! tonyp@3416: void clearRangePrevBitmap(MemRegion mr); tonyp@3416: void clearRangeNextBitmap(MemRegion mr); tonyp@3416: void clearRangeBothBitmaps(MemRegion mr); ysr@777: tonyp@3416: // Notify data structures that a GC has started. tonyp@3416: void note_start_of_gc() { tonyp@3416: _markStack.note_start_of_gc(); ysr@777: } tonyp@3416: tonyp@3416: // Notify data structures that a GC is finished. tonyp@3416: void note_end_of_gc() { tonyp@3416: _markStack.note_end_of_gc(); tonyp@3416: } tonyp@3416: tonyp@3416: // Verify that there are no CSet oops on the stacks (taskqueues / tonyp@3416: // global mark stack), enqueued SATB buffers, per-thread SATB tonyp@3416: // buffers, and fingers (global / per-task). The boolean parameters tonyp@3416: // decide which of the above data structures to verify. If marking tonyp@3416: // is not in progress, it's a no-op. tonyp@3416: void verify_no_cset_oops(bool verify_stacks, tonyp@3416: bool verify_enqueued_buffers, tonyp@3416: bool verify_thread_buffers, tonyp@3416: bool verify_fingers) PRODUCT_RETURN; tonyp@3416: ysr@777: // It is called at the end of an evacuation pause during marking so ysr@777: // that CM is notified of where the new end of the heap is. It ysr@777: // doesn't do anything if concurrent_marking_in_progress() is false, ysr@777: // unless the force parameter is true. ysr@777: void update_g1_committed(bool force = false); ysr@777: ysr@777: bool isMarked(oop p) const { ysr@777: assert(p != NULL && p->is_oop(), "expected an oop"); ysr@777: HeapWord* addr = (HeapWord*)p; ysr@777: assert(addr >= _nextMarkBitMap->startWord() || ysr@777: addr < _nextMarkBitMap->endWord(), "in a region"); ysr@777: ysr@777: return _nextMarkBitMap->isMarked(addr); ysr@777: } ysr@777: ysr@777: inline bool not_yet_marked(oop p) const; ysr@777: ysr@777: // XXX Debug code ysr@777: bool containing_card_is_marked(void* p); ysr@777: bool containing_cards_are_marked(void* start, void* last); ysr@777: ysr@777: bool isPrevMarked(oop p) const { ysr@777: assert(p != NULL && p->is_oop(), "expected an oop"); ysr@777: HeapWord* addr = (HeapWord*)p; ysr@777: assert(addr >= _prevMarkBitMap->startWord() || ysr@777: addr < _prevMarkBitMap->endWord(), "in a region"); ysr@777: ysr@777: return _prevMarkBitMap->isMarked(addr); ysr@777: } ysr@777: jmasa@3357: inline bool do_yield_check(uint worker_i = 0); ysr@777: ysr@777: // Called to abort the marking cycle after a Full GC takes palce. ysr@777: void abort(); ysr@777: sla@5237: bool has_aborted() { return _has_aborted; } sla@5237: brutisso@6904: const GCId& concurrent_gc_id(); brutisso@6904: ysr@777: // This prints the global/local fingers. It is used for debugging. ysr@777: NOT_PRODUCT(void print_finger();) ysr@777: ysr@777: void print_summary_info(); ysr@777: tonyp@1454: void print_worker_threads_on(outputStream* st) const; tonyp@1454: stefank@4904: void print_on_error(outputStream* st) const; stefank@4904: ysr@777: // The following indicate whether a given verbose level has been ysr@777: // set. Notice that anything above stats is conditional to ysr@777: // _MARKING_VERBOSE_ having been set to 1 tonyp@2973: bool verbose_stats() { tonyp@2973: return _verbose_level >= stats_verbose; tonyp@2973: } tonyp@2973: bool verbose_low() { tonyp@2973: return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; tonyp@2973: } tonyp@2973: bool verbose_medium() { tonyp@2973: return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; tonyp@2973: } tonyp@2973: bool verbose_high() { tonyp@2973: return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; tonyp@2973: } johnc@3463: johnc@4123: // Liveness counting johnc@4123: johnc@4123: // Utility routine to set an exclusive range of cards on the given johnc@4123: // card liveness bitmap johnc@4123: inline void set_card_bitmap_range(BitMap* card_bm, johnc@4123: BitMap::idx_t start_idx, johnc@4123: BitMap::idx_t end_idx, johnc@4123: bool is_par); johnc@3463: johnc@3463: // Returns the card number of the bottom of the G1 heap. johnc@3463: // Used in biasing indices into accounting card bitmaps. johnc@3463: intptr_t heap_bottom_card_num() const { johnc@3463: return _heap_bottom_card_num; johnc@3463: } johnc@3463: johnc@3463: // Returns the card bitmap for a given task or worker id. johnc@3463: BitMap* count_card_bitmap_for(uint worker_id) { johnc@4173: assert(0 <= worker_id && worker_id < _max_worker_id, "oob"); johnc@3463: assert(_count_card_bitmaps != NULL, "uninitialized"); johnc@3463: BitMap* task_card_bm = &_count_card_bitmaps[worker_id]; johnc@3463: assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); johnc@3463: return task_card_bm; johnc@3463: } johnc@3463: johnc@3463: // Returns the array containing the marked bytes for each region, johnc@3463: // for the given worker or task id. johnc@3463: size_t* count_marked_bytes_array_for(uint worker_id) { johnc@4173: assert(0 <= worker_id && worker_id < _max_worker_id, "oob"); johnc@3463: assert(_count_marked_bytes != NULL, "uninitialized"); johnc@3463: size_t* marked_bytes_array = _count_marked_bytes[worker_id]; johnc@3463: assert(marked_bytes_array != NULL, "uninitialized"); johnc@3463: return marked_bytes_array; johnc@3463: } johnc@3463: johnc@3463: // Returns the index in the liveness accounting card table bitmap johnc@3463: // for the given address johnc@3463: inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr); johnc@3463: johnc@3463: // Counts the size of the given memory region in the the given johnc@3463: // marked_bytes array slot for the given HeapRegion. johnc@3463: // Sets the bits in the given card bitmap that are associated with the johnc@3463: // cards that are spanned by the memory region. johnc@3463: inline void count_region(MemRegion mr, HeapRegion* hr, johnc@3463: size_t* marked_bytes_array, johnc@3463: BitMap* task_card_bm); johnc@3463: johnc@3463: // Counts the given memory region in the task/worker counting johnc@3463: // data structures for the given worker id. tonyp@3464: inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id); tonyp@3464: tonyp@3464: // Counts the given memory region in the task/worker counting tonyp@3464: // data structures for the given worker id. johnc@3463: inline void count_region(MemRegion mr, uint worker_id); johnc@3463: johnc@3463: // Counts the given object in the given task/worker counting johnc@3463: // data structures. johnc@3463: inline void count_object(oop obj, HeapRegion* hr, johnc@3463: size_t* marked_bytes_array, johnc@3463: BitMap* task_card_bm); johnc@3463: johnc@3463: // Counts the given object in the task/worker counting data johnc@3463: // structures for the given worker id. johnc@3463: inline void count_object(oop obj, HeapRegion* hr, uint worker_id); johnc@3463: johnc@3463: // Attempts to mark the given object and, if successful, counts johnc@3463: // the object in the given task/worker counting structures. johnc@3463: inline bool par_mark_and_count(oop obj, HeapRegion* hr, johnc@3463: size_t* marked_bytes_array, johnc@3463: BitMap* task_card_bm); johnc@3463: johnc@3463: // Attempts to mark the given object and, if successful, counts johnc@3463: // the object in the task/worker counting structures for the johnc@3463: // given worker id. tonyp@3464: inline bool par_mark_and_count(oop obj, size_t word_size, tonyp@3464: HeapRegion* hr, uint worker_id); tonyp@3464: tonyp@3464: // Attempts to mark the given object and, if successful, counts tonyp@3464: // the object in the task/worker counting structures for the tonyp@3464: // given worker id. johnc@3463: inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id); johnc@3463: johnc@3463: // Similar to the above routine but we don't know the heap region that johnc@3463: // contains the object to be marked/counted, which this routine looks up. johnc@3463: inline bool par_mark_and_count(oop obj, uint worker_id); johnc@3463: johnc@3463: // Similar to the above routine but there are times when we cannot johnc@3463: // safely calculate the size of obj due to races and we, therefore, johnc@3463: // pass the size in as a parameter. It is the caller's reponsibility johnc@3463: // to ensure that the size passed in for obj is valid. johnc@3463: inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id); johnc@3463: johnc@3463: // Unconditionally mark the given object, and unconditinally count johnc@3463: // the object in the counting structures for worker id 0. johnc@3463: // Should *not* be called from parallel code. johnc@3463: inline bool mark_and_count(oop obj, HeapRegion* hr); johnc@3463: johnc@3463: // Similar to the above routine but we don't know the heap region that johnc@3463: // contains the object to be marked/counted, which this routine looks up. johnc@3463: // Should *not* be called from parallel code. johnc@3463: inline bool mark_and_count(oop obj); johnc@3463: johnc@4333: // Returns true if initialization was successfully completed. johnc@4333: bool completed_initialization() const { johnc@4333: return _completed_initialization; johnc@4333: } johnc@4333: johnc@3463: protected: johnc@3463: // Clear all the per-task bitmaps and arrays used to store the johnc@3463: // counting data. johnc@3463: void clear_all_count_data(); johnc@3463: johnc@3463: // Aggregates the counting data for each worker/task johnc@3463: // that was constructed while marking. Also sets johnc@3463: // the amount of marked bytes for each region and johnc@3463: // the top at concurrent mark count. johnc@3463: void aggregate_count_data(); johnc@3463: johnc@3463: // Verification routine johnc@3463: void verify_count_data(); ysr@777: }; ysr@777: ysr@777: // A class representing a marking task. ysr@777: class CMTask : public TerminatorTerminator { ysr@777: private: ysr@777: enum PrivateConstants { ysr@777: // the regular clock call is called once the scanned words reaches ysr@777: // this limit ysr@777: words_scanned_period = 12*1024, ysr@777: // the regular clock call is called once the number of visited ysr@777: // references reaches this limit ysr@777: refs_reached_period = 384, ysr@777: // initial value for the hash seed, used in the work stealing code ysr@777: init_hash_seed = 17, ysr@777: // how many entries will be transferred between global stack and ysr@777: // local queues ysr@777: global_stack_transfer_size = 16 ysr@777: }; ysr@777: johnc@4173: uint _worker_id; ysr@777: G1CollectedHeap* _g1h; ysr@777: ConcurrentMark* _cm; ysr@777: CMBitMap* _nextMarkBitMap; ysr@777: // the task queue of this task ysr@777: CMTaskQueue* _task_queue; ysr@1280: private: ysr@777: // the task queue set---needed for stealing ysr@777: CMTaskQueueSet* _task_queues; ysr@777: // indicates whether the task has been claimed---this is only for ysr@777: // debugging purposes ysr@777: bool _claimed; ysr@777: ysr@777: // number of calls to this task ysr@777: int _calls; ysr@777: ysr@777: // when the virtual timer reaches this time, the marking step should ysr@777: // exit ysr@777: double _time_target_ms; ysr@777: // the start time of the current marking step ysr@777: double _start_time_ms; ysr@777: ysr@777: // the oop closure used for iterations over oops tonyp@2968: G1CMOopClosure* _cm_oop_closure; ysr@777: ysr@777: // the region this task is scanning, NULL if we're not scanning any ysr@777: HeapRegion* _curr_region; ysr@777: // the local finger of this task, NULL if we're not scanning a region ysr@777: HeapWord* _finger; ysr@777: // limit of the region this task is scanning, NULL if we're not scanning one ysr@777: HeapWord* _region_limit; ysr@777: ysr@777: // the number of words this task has scanned ysr@777: size_t _words_scanned; ysr@777: // When _words_scanned reaches this limit, the regular clock is ysr@777: // called. Notice that this might be decreased under certain ysr@777: // circumstances (i.e. when we believe that we did an expensive ysr@777: // operation). ysr@777: size_t _words_scanned_limit; ysr@777: // the initial value of _words_scanned_limit (i.e. what it was ysr@777: // before it was decreased). ysr@777: size_t _real_words_scanned_limit; ysr@777: ysr@777: // the number of references this task has visited ysr@777: size_t _refs_reached; ysr@777: // When _refs_reached reaches this limit, the regular clock is ysr@777: // called. Notice this this might be decreased under certain ysr@777: // circumstances (i.e. when we believe that we did an expensive ysr@777: // operation). ysr@777: size_t _refs_reached_limit; ysr@777: // the initial value of _refs_reached_limit (i.e. what it was before ysr@777: // it was decreased). ysr@777: size_t _real_refs_reached_limit; ysr@777: ysr@777: // used by the work stealing stuff ysr@777: int _hash_seed; ysr@777: // if this is true, then the task has aborted for some reason ysr@777: bool _has_aborted; ysr@777: // set when the task aborts because it has met its time quota johnc@2494: bool _has_timed_out; ysr@777: // true when we're draining SATB buffers; this avoids the task ysr@777: // aborting due to SATB buffers being available (as we're already ysr@777: // dealing with them) ysr@777: bool _draining_satb_buffers; ysr@777: ysr@777: // number sequence of past step times ysr@777: NumberSeq _step_times_ms; ysr@777: // elapsed time of this task ysr@777: double _elapsed_time_ms; ysr@777: // termination time of this task ysr@777: double _termination_time_ms; ysr@777: // when this task got into the termination protocol ysr@777: double _termination_start_time_ms; ysr@777: ysr@777: // true when the task is during a concurrent phase, false when it is ysr@777: // in the remark phase (so, in the latter case, we do not have to ysr@777: // check all the things that we have to check during the concurrent ysr@777: // phase, i.e. SATB buffer availability...) ysr@777: bool _concurrent; ysr@777: ysr@777: TruncatedSeq _marking_step_diffs_ms; ysr@777: johnc@3463: // Counting data structures. Embedding the task's marked_bytes_array johnc@3463: // and card bitmap into the actual task saves having to go through johnc@3463: // the ConcurrentMark object. johnc@3463: size_t* _marked_bytes_array; johnc@3463: BitMap* _card_bm; johnc@3463: ysr@777: // LOTS of statistics related with this task ysr@777: #if _MARKING_STATS_ ysr@777: NumberSeq _all_clock_intervals_ms; ysr@777: double _interval_start_time_ms; ysr@777: ysr@777: int _aborted; ysr@777: int _aborted_overflow; ysr@777: int _aborted_cm_aborted; ysr@777: int _aborted_yield; ysr@777: int _aborted_timed_out; ysr@777: int _aborted_satb; ysr@777: int _aborted_termination; ysr@777: ysr@777: int _steal_attempts; ysr@777: int _steals; ysr@777: ysr@777: int _clock_due_to_marking; ysr@777: int _clock_due_to_scanning; ysr@777: ysr@777: int _local_pushes; ysr@777: int _local_pops; ysr@777: int _local_max_size; ysr@777: int _objs_scanned; ysr@777: ysr@777: int _global_pushes; ysr@777: int _global_pops; ysr@777: int _global_max_size; ysr@777: ysr@777: int _global_transfers_to; ysr@777: int _global_transfers_from; ysr@777: ysr@777: int _regions_claimed; ysr@777: int _objs_found_on_bitmap; ysr@777: ysr@777: int _satb_buffers_processed; ysr@777: #endif // _MARKING_STATS_ ysr@777: ysr@777: // it updates the local fields after this task has claimed ysr@777: // a new region to scan ysr@777: void setup_for_region(HeapRegion* hr); ysr@777: // it brings up-to-date the limit of the region ysr@777: void update_region_limit(); ysr@777: ysr@777: // called when either the words scanned or the refs visited limit ysr@777: // has been reached ysr@777: void reached_limit(); ysr@777: // recalculates the words scanned and refs visited limits ysr@777: void recalculate_limits(); ysr@777: // decreases the words scanned and refs visited limits when we reach ysr@777: // an expensive operation ysr@777: void decrease_limits(); ysr@777: // it checks whether the words scanned or refs visited reached their ysr@777: // respective limit and calls reached_limit() if they have ysr@777: void check_limits() { ysr@777: if (_words_scanned >= _words_scanned_limit || tonyp@2973: _refs_reached >= _refs_reached_limit) { ysr@777: reached_limit(); tonyp@2973: } ysr@777: } ysr@777: // this is supposed to be called regularly during a marking step as ysr@777: // it checks a bunch of conditions that might cause the marking step ysr@777: // to abort ysr@777: void regular_clock_call(); ysr@777: bool concurrent() { return _concurrent; } ysr@777: ysr@777: public: ysr@777: // It resets the task; it should be called right at the beginning of ysr@777: // a marking phase. ysr@777: void reset(CMBitMap* _nextMarkBitMap); ysr@777: // it clears all the fields that correspond to a claimed region. ysr@777: void clear_region_fields(); ysr@777: ysr@777: void set_concurrent(bool concurrent) { _concurrent = concurrent; } ysr@777: ysr@777: // The main method of this class which performs a marking step ysr@777: // trying not to exceed the given duration. However, it might exit ysr@777: // prematurely, according to some conditions (i.e. SATB buffers are ysr@777: // available for processing). johnc@4787: void do_marking_step(double target_ms, johnc@4787: bool do_termination, johnc@4787: bool is_serial); ysr@777: ysr@777: // These two calls start and stop the timer ysr@777: void record_start_time() { ysr@777: _elapsed_time_ms = os::elapsedTime() * 1000.0; ysr@777: } ysr@777: void record_end_time() { ysr@777: _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; ysr@777: } ysr@777: johnc@4173: // returns the worker ID associated with this task. johnc@4173: uint worker_id() { return _worker_id; } ysr@777: ysr@777: // From TerminatorTerminator. It determines whether this task should ysr@777: // exit the termination protocol after it's entered it. ysr@777: virtual bool should_exit_termination(); ysr@777: johnc@2910: // Resets the local region fields after a task has finished scanning a johnc@2910: // region; or when they have become stale as a result of the region johnc@2910: // being evacuated. johnc@2910: void giveup_current_region(); johnc@2910: ysr@777: HeapWord* finger() { return _finger; } ysr@777: ysr@777: bool has_aborted() { return _has_aborted; } ysr@777: void set_has_aborted() { _has_aborted = true; } ysr@777: void clear_has_aborted() { _has_aborted = false; } johnc@2494: bool has_timed_out() { return _has_timed_out; } johnc@2494: bool claimed() { return _claimed; } ysr@777: tonyp@2968: void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); ysr@777: ysr@777: // It grays the object by marking it and, if necessary, pushing it ysr@777: // on the local queue tonyp@2968: inline void deal_with_reference(oop obj); ysr@777: ysr@777: // It scans an object and visits its children. tonyp@2968: void scan_object(oop obj); ysr@777: ysr@777: // It pushes an object on the local queue. tonyp@2968: inline void push(oop obj); ysr@777: ysr@777: // These two move entries to/from the global stack. ysr@777: void move_entries_to_global_stack(); ysr@777: void get_entries_from_global_stack(); ysr@777: ysr@777: // It pops and scans objects from the local queue. If partially is ysr@777: // true, then it stops when the queue size is of a given limit. If ysr@777: // partially is false, then it stops when the queue is empty. ysr@777: void drain_local_queue(bool partially); ysr@777: // It moves entries from the global stack to the local queue and ysr@777: // drains the local queue. If partially is true, then it stops when ysr@777: // both the global stack and the local queue reach a given size. If ysr@777: // partially if false, it tries to empty them totally. ysr@777: void drain_global_stack(bool partially); ysr@777: // It keeps picking SATB buffers and processing them until no SATB ysr@777: // buffers are available. ysr@777: void drain_satb_buffers(); tonyp@3416: ysr@777: // moves the local finger to a new location ysr@777: inline void move_finger_to(HeapWord* new_finger) { tonyp@1458: assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); ysr@777: _finger = new_finger; ysr@777: } ysr@777: johnc@4173: CMTask(uint worker_id, ConcurrentMark *cm, johnc@3463: size_t* marked_bytes, BitMap* card_bm, ysr@777: CMTaskQueue* task_queue, CMTaskQueueSet* task_queues); ysr@777: ysr@777: // it prints statistics associated with this task ysr@777: void print_stats(); ysr@777: ysr@777: #if _MARKING_STATS_ ysr@777: void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; } ysr@777: #endif // _MARKING_STATS_ ysr@777: }; stefank@2314: tonyp@2717: // Class that's used to to print out per-region liveness tonyp@2717: // information. It's currently used at the end of marking and also tonyp@2717: // after we sort the old regions at the end of the cleanup operation. tonyp@2717: class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { tonyp@2717: private: tonyp@2717: outputStream* _out; tonyp@2717: tonyp@2717: // Accumulators for these values. tonyp@2717: size_t _total_used_bytes; tonyp@2717: size_t _total_capacity_bytes; tonyp@2717: size_t _total_prev_live_bytes; tonyp@2717: size_t _total_next_live_bytes; tonyp@2717: tonyp@2717: // These are set up when we come across a "stars humongous" region tonyp@2717: // (as this is where most of this information is stored, not in the tonyp@2717: // subsequent "continues humongous" regions). After that, for every tonyp@2717: // region in a given humongous region series we deduce the right tonyp@2717: // values for it by simply subtracting the appropriate amount from tonyp@2717: // these fields. All these values should reach 0 after we've visited tonyp@2717: // the last region in the series. tonyp@2717: size_t _hum_used_bytes; tonyp@2717: size_t _hum_capacity_bytes; tonyp@2717: size_t _hum_prev_live_bytes; tonyp@2717: size_t _hum_next_live_bytes; tonyp@2717: tschatzl@5122: // Accumulator for the remembered set size tschatzl@5122: size_t _total_remset_bytes; tschatzl@5122: johnc@5548: // Accumulator for strong code roots memory size johnc@5548: size_t _total_strong_code_roots_bytes; johnc@5548: tonyp@2717: static double perc(size_t val, size_t total) { tonyp@2717: if (total == 0) { tonyp@2717: return 0.0; tonyp@2717: } else { tonyp@2717: return 100.0 * ((double) val / (double) total); tonyp@2717: } tonyp@2717: } tonyp@2717: tonyp@2717: static double bytes_to_mb(size_t val) { tonyp@2717: return (double) val / (double) M; tonyp@2717: } tonyp@2717: tonyp@2717: // See the .cpp file. tonyp@2717: size_t get_hum_bytes(size_t* hum_bytes); tonyp@2717: void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes, tonyp@2717: size_t* prev_live_bytes, size_t* next_live_bytes); tonyp@2717: tonyp@2717: public: tonyp@2717: // The header and footer are printed in the constructor and tonyp@2717: // destructor respectively. tonyp@2717: G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name); tonyp@2717: virtual bool doHeapRegion(HeapRegion* r); tonyp@2717: ~G1PrintRegionLivenessInfoClosure(); tonyp@2717: }; tonyp@2717: stefank@2314: #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP