comparison src/share/vm/gc_implementation/g1/concurrentG1Refine.hpp @ 890:6cb8e9df7174

6819077: G1: first GC thread coming late into the GC. Summary: The first worker thread is delayed when entering the GC because it clears the card count table that is used in identifying hot cards. Replace the card count table with a dynamically sized evicting hash table that includes an epoch based counter. Reviewed-by: iveresov, tonyp
author johnc
date Tue, 04 Aug 2009 16:00:17 -0700
parents 15c5903cf9e1
children 035d2e036a9b
comparison
equal deleted inserted replaced
889:15c5903cf9e1 890:6cb8e9df7174
27 class G1RemSet; 27 class G1RemSet;
28 28
29 class ConcurrentG1Refine: public CHeapObj { 29 class ConcurrentG1Refine: public CHeapObj {
30 ConcurrentG1RefineThread** _threads; 30 ConcurrentG1RefineThread** _threads;
31 int _n_threads; 31 int _n_threads;
32
32 // The cache for card refinement. 33 // The cache for card refinement.
33 bool _use_cache; 34 bool _use_cache;
34 bool _def_use_cache; 35 bool _def_use_cache;
35 size_t _n_periods;
36 size_t _total_cards;
37 size_t _total_travs;
38 36
39 unsigned char* _card_counts; 37 size_t _n_periods; // Used as clearing epoch
40 unsigned _n_card_counts; 38
41 const jbyte* _ct_bot; 39 // An evicting cache of the number of times each card
42 unsigned* _cur_card_count_histo; 40 // is accessed. Reduces, but does not eliminate, the amount
43 unsigned* _cum_card_count_histo; 41 // of duplicated processing of dirty cards.
42
43 enum SomePrivateConstants {
44 epoch_bits = 32,
45 card_num_shift = epoch_bits,
46 epoch_mask = AllBits,
47 card_num_mask = AllBits,
48
49 // The initial cache size is approximately this fraction
50 // of a maximal cache (i.e. the size needed for all cards
51 // in the heap)
52 InitialCacheFraction = 512
53 };
54
55 const static julong card_num_mask_in_place =
56 (julong) card_num_mask << card_num_shift;
57
58 typedef struct {
59 julong _value; // | card_num | epoch |
60 } CardEpochCacheEntry;
61
62 julong make_epoch_entry(unsigned int card_num, unsigned int epoch) {
63 assert(0 <= card_num && card_num < _max_n_card_counts, "Bounds");
64 assert(0 <= epoch && epoch <= _n_periods, "must be");
65
66 return ((julong) card_num << card_num_shift) | epoch;
67 }
68
69 unsigned int extract_epoch(julong v) {
70 return (v & epoch_mask);
71 }
72
73 unsigned int extract_card_num(julong v) {
74 return (v & card_num_mask_in_place) >> card_num_shift;
75 }
76
77 typedef struct {
78 unsigned char _count;
79 unsigned char _evict_count;
80 } CardCountCacheEntry;
81
82 CardCountCacheEntry* _card_counts;
83 CardEpochCacheEntry* _card_epochs;
84
85 // The current number of buckets in the card count cache
86 unsigned _n_card_counts;
87
88 // The max number of buckets required for the number of
89 // cards for the entire reserved heap
90 unsigned _max_n_card_counts;
91
92 // Possible sizes of the cache: odd primes that roughly double in size.
93 // (See jvmtiTagMap.cpp).
94 static int _cc_cache_sizes[];
95
96 // The index in _cc_cache_sizes corresponding to the size of
97 // _card_counts.
98 int _cache_size_index;
99
100 bool _expand_card_counts;
101
102 const jbyte* _ct_bot;
44 103
45 jbyte** _hot_cache; 104 jbyte** _hot_cache;
46 int _hot_cache_size; 105 int _hot_cache_size;
47 int _n_hot; 106 int _n_hot;
48 int _hot_cache_idx; 107 int _hot_cache_idx;
49 108
50 int _hot_cache_par_chunk_size; 109 int _hot_cache_par_chunk_size;
51 volatile int _hot_cache_par_claimed_idx; 110 volatile int _hot_cache_par_claimed_idx;
52 111
112 // Needed to workaround 6817995
113 CardTableModRefBS* _ct_bs;
114 G1CollectedHeap* _g1h;
115
116 // Expands the array that holds the card counts to the next size up
117 void expand_card_count_cache();
118
119 // hash a given key (index of card_ptr) with the specified size
120 static unsigned int hash(size_t key, int size) {
121 return (unsigned int) key % size;
122 }
123
124 // hash a given key (index of card_ptr)
125 unsigned int hash(size_t key) {
126 return hash(key, _n_card_counts);
127 }
128
129 unsigned ptr_2_card_num(jbyte* card_ptr) {
130 return (unsigned) (card_ptr - _ct_bot);
131 }
132
133 jbyte* card_num_2_ptr(unsigned card_num) {
134 return (jbyte*) (_ct_bot + card_num);
135 }
136
53 // Returns the count of this card after incrementing it. 137 // Returns the count of this card after incrementing it.
54 int add_card_count(jbyte* card_ptr); 138 jbyte* add_card_count(jbyte* card_ptr, int* count, bool* defer);
55 139
56 void print_card_count_histo_range(unsigned* histo, int from, int to, 140 // Returns true if this card is in a young region
57 float& cum_card_pct, 141 bool is_young_card(jbyte* card_ptr);
58 float& cum_travs_pct); 142
59 public: 143 public:
60 ConcurrentG1Refine(); 144 ConcurrentG1Refine();
61 ~ConcurrentG1Refine(); 145 ~ConcurrentG1Refine();
62 146
63 void init(); // Accomplish some initialization that has to wait. 147 void init(); // Accomplish some initialization that has to wait.
67 void threads_do(ThreadClosure *tc); 151 void threads_do(ThreadClosure *tc);
68 152
69 // If this is the first entry for the slot, writes into the cache and 153 // If this is the first entry for the slot, writes into the cache and
70 // returns NULL. If it causes an eviction, returns the evicted pointer. 154 // returns NULL. If it causes an eviction, returns the evicted pointer.
71 // Otherwise, its a cache hit, and returns NULL. 155 // Otherwise, its a cache hit, and returns NULL.
72 jbyte* cache_insert(jbyte* card_ptr); 156 jbyte* cache_insert(jbyte* card_ptr, bool* defer);
73 157
74 // Process the cached entries. 158 // Process the cached entries.
75 void clean_up_cache(int worker_i, G1RemSet* g1rs); 159 void clean_up_cache(int worker_i, G1RemSet* g1rs);
76 160
77 // Set up for parallel processing of the cards in the hot cache 161 // Set up for parallel processing of the cards in the hot cache
91 if (b) _use_cache = _def_use_cache; 175 if (b) _use_cache = _def_use_cache;
92 else _use_cache = false; 176 else _use_cache = false;
93 } 177 }
94 178
95 void clear_and_record_card_counts(); 179 void clear_and_record_card_counts();
96 void print_final_card_counts();
97 180
98 static size_t thread_num(); 181 static size_t thread_num();
99 }; 182 };