1 | /*------------------------------------------------------------------------- |
2 | * |
3 | * hashsort.c |
4 | * Sort tuples for insertion into a new hash index. |
5 | * |
6 | * When building a very large hash index, we pre-sort the tuples by bucket |
7 | * number to improve locality of access to the index, and thereby avoid |
8 | * thrashing. We use tuplesort.c to sort the given index tuples into order. |
9 | * |
10 | * Note: if the number of rows in the table has been underestimated, |
11 | * bucket splits may occur during the index build. In that case we'd |
12 | * be inserting into two or more buckets for each possible masked-off |
13 | * hash code value. That's no big problem though, since we'll still have |
14 | * plenty of locality of access. |
15 | * |
16 | * |
17 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
18 | * Portions Copyright (c) 1994, Regents of the University of California |
19 | * |
20 | * IDENTIFICATION |
21 | * src/backend/access/hash/hashsort.c |
22 | * |
23 | *------------------------------------------------------------------------- |
24 | */ |
25 | |
26 | #include "postgres.h" |
27 | |
28 | #include "access/hash.h" |
29 | #include "commands/progress.h" |
30 | #include "miscadmin.h" |
31 | #include "pgstat.h" |
32 | #include "utils/tuplesort.h" |
33 | |
34 | |
35 | /* |
36 | * Status record for spooling/sorting phase. |
37 | */ |
38 | struct HSpool |
39 | { |
40 | Tuplesortstate *sortstate; /* state data for tuplesort.c */ |
41 | Relation index; |
42 | |
43 | /* |
44 | * We sort the hash keys based on the buckets they belong to. Below masks |
45 | * are used in _hash_hashkey2bucket to determine the bucket of given hash |
46 | * key. |
47 | */ |
48 | uint32 high_mask; |
49 | uint32 low_mask; |
50 | uint32 max_buckets; |
51 | }; |
52 | |
53 | |
54 | /* |
55 | * create and initialize a spool structure |
56 | */ |
57 | HSpool * |
58 | _h_spoolinit(Relation heap, Relation index, uint32 num_buckets) |
59 | { |
60 | HSpool *hspool = (HSpool *) palloc0(sizeof(HSpool)); |
61 | |
62 | hspool->index = index; |
63 | |
64 | /* |
65 | * Determine the bitmask for hash code values. Since there are currently |
66 | * num_buckets buckets in the index, the appropriate mask can be computed |
67 | * as follows. |
68 | * |
69 | * NOTE : This hash mask calculation should be in sync with similar |
70 | * calculation in _hash_init_metabuffer. |
71 | */ |
72 | hspool->high_mask = (((uint32) 1) << _hash_log2(num_buckets + 1)) - 1; |
73 | hspool->low_mask = (hspool->high_mask >> 1); |
74 | hspool->max_buckets = num_buckets - 1; |
75 | |
76 | /* |
77 | * We size the sort area as maintenance_work_mem rather than work_mem to |
78 | * speed index creation. This should be OK since a single backend can't |
79 | * run multiple index creations in parallel. |
80 | */ |
81 | hspool->sortstate = tuplesort_begin_index_hash(heap, |
82 | index, |
83 | hspool->high_mask, |
84 | hspool->low_mask, |
85 | hspool->max_buckets, |
86 | maintenance_work_mem, |
87 | NULL, |
88 | false); |
89 | |
90 | return hspool; |
91 | } |
92 | |
93 | /* |
94 | * clean up a spool structure and its substructures. |
95 | */ |
96 | void |
97 | _h_spooldestroy(HSpool *hspool) |
98 | { |
99 | tuplesort_end(hspool->sortstate); |
100 | pfree(hspool); |
101 | } |
102 | |
103 | /* |
104 | * spool an index entry into the sort file. |
105 | */ |
106 | void |
107 | _h_spool(HSpool *hspool, ItemPointer self, Datum *values, bool *isnull) |
108 | { |
109 | tuplesort_putindextuplevalues(hspool->sortstate, hspool->index, |
110 | self, values, isnull); |
111 | } |
112 | |
113 | /* |
114 | * given a spool loaded by successive calls to _h_spool, |
115 | * create an entire index. |
116 | */ |
117 | void |
118 | _h_indexbuild(HSpool *hspool, Relation heapRel) |
119 | { |
120 | IndexTuple itup; |
121 | int64 tups_done = 0; |
122 | #ifdef USE_ASSERT_CHECKING |
123 | uint32 hashkey = 0; |
124 | #endif |
125 | |
126 | tuplesort_performsort(hspool->sortstate); |
127 | |
128 | while ((itup = tuplesort_getindextuple(hspool->sortstate, true)) != NULL) |
129 | { |
130 | /* |
131 | * Technically, it isn't critical that hash keys be found in sorted |
132 | * order, since this sorting is only used to increase locality of |
133 | * access as a performance optimization. It still seems like a good |
134 | * idea to test tuplesort.c's handling of hash index tuple sorts |
135 | * through an assertion, though. |
136 | */ |
137 | #ifdef USE_ASSERT_CHECKING |
138 | uint32 lasthashkey = hashkey; |
139 | |
140 | hashkey = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup), |
141 | hspool->max_buckets, hspool->high_mask, |
142 | hspool->low_mask); |
143 | Assert(hashkey >= lasthashkey); |
144 | #endif |
145 | |
146 | _hash_doinsert(hspool->index, itup, heapRel); |
147 | |
148 | pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE, |
149 | ++tups_done); |
150 | } |
151 | } |
152 | |