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) 1996-2020, 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 "port/pg_bitutils.h"
33 #include "utils/tuplesort.h"
34 
35 
36 /*
37  * Status record for spooling/sorting phase.
38  */
39 struct HSpool
40 {
41 	Tuplesortstate *sortstate;	/* state data for tuplesort.c */
42 	Relation	index;
43 
44 	/*
45 	 * We sort the hash keys based on the buckets they belong to. Below masks
46 	 * are used in _hash_hashkey2bucket to determine the bucket of given hash
47 	 * key.
48 	 */
49 	uint32		high_mask;
50 	uint32		low_mask;
51 	uint32		max_buckets;
52 };
53 
54 
55 /*
56  * create and initialize a spool structure
57  */
58 HSpool *
_h_spoolinit(Relation heap,Relation index,uint32 num_buckets)59 _h_spoolinit(Relation heap, Relation index, uint32 num_buckets)
60 {
61 	HSpool	   *hspool = (HSpool *) palloc0(sizeof(HSpool));
62 
63 	hspool->index = index;
64 
65 	/*
66 	 * Determine the bitmask for hash code values.  Since there are currently
67 	 * num_buckets buckets in the index, the appropriate mask can be computed
68 	 * as follows.
69 	 *
70 	 * NOTE : This hash mask calculation should be in sync with similar
71 	 * calculation in _hash_init_metabuffer.
72 	 */
73 	hspool->high_mask = pg_nextpower2_32(num_buckets + 1) - 1;
74 	hspool->low_mask = (hspool->high_mask >> 1);
75 	hspool->max_buckets = num_buckets - 1;
76 
77 	/*
78 	 * We size the sort area as maintenance_work_mem rather than work_mem to
79 	 * speed index creation.  This should be OK since a single backend can't
80 	 * run multiple index creations in parallel.
81 	 */
82 	hspool->sortstate = tuplesort_begin_index_hash(heap,
83 												   index,
84 												   hspool->high_mask,
85 												   hspool->low_mask,
86 												   hspool->max_buckets,
87 												   maintenance_work_mem,
88 												   NULL,
89 												   false);
90 
91 	return hspool;
92 }
93 
94 /*
95  * clean up a spool structure and its substructures.
96  */
97 void
_h_spooldestroy(HSpool * hspool)98 _h_spooldestroy(HSpool *hspool)
99 {
100 	tuplesort_end(hspool->sortstate);
101 	pfree(hspool);
102 }
103 
104 /*
105  * spool an index entry into the sort file.
106  */
107 void
_h_spool(HSpool * hspool,ItemPointer self,Datum * values,bool * isnull)108 _h_spool(HSpool *hspool, ItemPointer self, Datum *values, bool *isnull)
109 {
110 	tuplesort_putindextuplevalues(hspool->sortstate, hspool->index,
111 								  self, values, isnull);
112 }
113 
114 /*
115  * given a spool loaded by successive calls to _h_spool,
116  * create an entire index.
117  */
118 void
_h_indexbuild(HSpool * hspool,Relation heapRel)119 _h_indexbuild(HSpool *hspool, Relation heapRel)
120 {
121 	IndexTuple	itup;
122 	int64		tups_done = 0;
123 #ifdef USE_ASSERT_CHECKING
124 	uint32		hashkey = 0;
125 #endif
126 
127 	tuplesort_performsort(hspool->sortstate);
128 
129 	while ((itup = tuplesort_getindextuple(hspool->sortstate, true)) != NULL)
130 	{
131 		/*
132 		 * Technically, it isn't critical that hash keys be found in sorted
133 		 * order, since this sorting is only used to increase locality of
134 		 * access as a performance optimization.  It still seems like a good
135 		 * idea to test tuplesort.c's handling of hash index tuple sorts
136 		 * through an assertion, though.
137 		 */
138 #ifdef USE_ASSERT_CHECKING
139 		uint32		lasthashkey = hashkey;
140 
141 		hashkey = _hash_hashkey2bucket(_hash_get_indextuple_hashkey(itup),
142 									   hspool->max_buckets, hspool->high_mask,
143 									   hspool->low_mask);
144 		Assert(hashkey >= lasthashkey);
145 #endif
146 
147 		_hash_doinsert(hspool->index, itup, heapRel);
148 
149 		pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE,
150 									 ++tups_done);
151 	}
152 }
153