Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * pgstatapprox.c
4 : * Bloat estimation functions
5 : *
6 : * Copyright (c) 2014-2026, PostgreSQL Global Development Group
7 : *
8 : * IDENTIFICATION
9 : * contrib/pgstattuple/pgstatapprox.c
10 : *
11 : *-------------------------------------------------------------------------
12 : */
13 : #include "postgres.h"
14 :
15 : #include "access/heapam.h"
16 : #include "access/htup_details.h"
17 : #include "access/relation.h"
18 : #include "access/visibilitymap.h"
19 : #include "catalog/pg_am_d.h"
20 : #include "commands/vacuum.h"
21 : #include "funcapi.h"
22 : #include "miscadmin.h"
23 : #include "storage/bufmgr.h"
24 : #include "storage/freespace.h"
25 : #include "storage/procarray.h"
26 :
27 1 : PG_FUNCTION_INFO_V1(pgstattuple_approx);
28 2 : PG_FUNCTION_INFO_V1(pgstattuple_approx_v1_5);
29 :
30 : Datum pgstattuple_approx_internal(Oid relid, FunctionCallInfo fcinfo);
31 :
32 : typedef struct output_type
33 : {
34 : uint64 table_len;
35 : double scanned_percent;
36 : uint64 tuple_count;
37 : uint64 tuple_len;
38 : double tuple_percent;
39 : uint64 dead_tuple_count;
40 : uint64 dead_tuple_len;
41 : double dead_tuple_percent;
42 : uint64 free_space;
43 : double free_percent;
44 : } output_type;
45 :
46 : #define NUM_OUTPUT_COLUMNS 10
47 :
48 : /*
49 : * This function takes an already open relation and scans its pages,
50 : * skipping those that have the corresponding visibility map bit set.
51 : * For pages we skip, we find the free space from the free space map
52 : * and approximate tuple_len on that basis. For the others, we count
53 : * the exact number of dead tuples etc.
54 : *
55 : * This scan is loosely based on vacuumlazy.c:lazy_scan_heap(), but
56 : * we do not try to avoid skipping single pages.
57 : */
58 : static void
59 2 : statapprox_heap(Relation rel, output_type *stat)
60 : {
61 : BlockNumber scanned,
62 : nblocks,
63 : blkno;
64 2 : Buffer vmbuffer = InvalidBuffer;
65 : BufferAccessStrategy bstrategy;
66 : TransactionId OldestXmin;
67 :
68 2 : OldestXmin = GetOldestNonRemovableTransactionId(rel);
69 2 : bstrategy = GetAccessStrategy(BAS_BULKREAD);
70 :
71 2 : nblocks = RelationGetNumberOfBlocks(rel);
72 2 : scanned = 0;
73 :
74 2 : for (blkno = 0; blkno < nblocks; blkno++)
75 : {
76 : Buffer buf;
77 : Page page;
78 : OffsetNumber offnum,
79 : maxoff;
80 : Size freespace;
81 :
82 0 : CHECK_FOR_INTERRUPTS();
83 :
84 : /*
85 : * If the page has only visible tuples, then we can find out the free
86 : * space from the FSM and move on.
87 : */
88 0 : if (VM_ALL_VISIBLE(rel, blkno, &vmbuffer))
89 : {
90 0 : freespace = GetRecordedFreeSpace(rel, blkno);
91 0 : stat->tuple_len += BLCKSZ - freespace;
92 0 : stat->free_space += freespace;
93 0 : continue;
94 : }
95 :
96 0 : buf = ReadBufferExtended(rel, MAIN_FORKNUM, blkno,
97 : RBM_NORMAL, bstrategy);
98 :
99 0 : LockBuffer(buf, BUFFER_LOCK_SHARE);
100 :
101 0 : page = BufferGetPage(buf);
102 :
103 0 : stat->free_space += PageGetExactFreeSpace(page);
104 :
105 : /* We may count the page as scanned even if it's new/empty */
106 0 : scanned++;
107 :
108 0 : if (PageIsNew(page) || PageIsEmpty(page))
109 : {
110 0 : UnlockReleaseBuffer(buf);
111 0 : continue;
112 : }
113 :
114 : /*
115 : * Look at each tuple on the page and decide whether it's live or
116 : * dead, then count it and its size. Unlike lazy_scan_heap, we can
117 : * afford to ignore problems and special cases.
118 : */
119 0 : maxoff = PageGetMaxOffsetNumber(page);
120 :
121 0 : for (offnum = FirstOffsetNumber;
122 0 : offnum <= maxoff;
123 0 : offnum = OffsetNumberNext(offnum))
124 : {
125 : ItemId itemid;
126 : HeapTupleData tuple;
127 :
128 0 : itemid = PageGetItemId(page, offnum);
129 :
130 0 : if (!ItemIdIsUsed(itemid) || ItemIdIsRedirected(itemid) ||
131 0 : ItemIdIsDead(itemid))
132 : {
133 0 : continue;
134 : }
135 :
136 : Assert(ItemIdIsNormal(itemid));
137 :
138 0 : ItemPointerSet(&(tuple.t_self), blkno, offnum);
139 :
140 0 : tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
141 0 : tuple.t_len = ItemIdGetLength(itemid);
142 0 : tuple.t_tableOid = RelationGetRelid(rel);
143 :
144 : /*
145 : * We follow VACUUM's lead in counting INSERT_IN_PROGRESS tuples
146 : * as "dead" while DELETE_IN_PROGRESS tuples are "live". We don't
147 : * bother distinguishing tuples inserted/deleted by our own
148 : * transaction.
149 : */
150 0 : switch (HeapTupleSatisfiesVacuum(&tuple, OldestXmin, buf))
151 : {
152 0 : case HEAPTUPLE_LIVE:
153 : case HEAPTUPLE_DELETE_IN_PROGRESS:
154 0 : stat->tuple_len += tuple.t_len;
155 0 : stat->tuple_count++;
156 0 : break;
157 0 : case HEAPTUPLE_DEAD:
158 : case HEAPTUPLE_RECENTLY_DEAD:
159 : case HEAPTUPLE_INSERT_IN_PROGRESS:
160 0 : stat->dead_tuple_len += tuple.t_len;
161 0 : stat->dead_tuple_count++;
162 0 : break;
163 0 : default:
164 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
165 : break;
166 : }
167 : }
168 :
169 0 : UnlockReleaseBuffer(buf);
170 : }
171 :
172 2 : stat->table_len = (uint64) nblocks * BLCKSZ;
173 :
174 : /*
175 : * We don't know how many tuples are in the pages we didn't scan, so
176 : * extrapolate the live-tuple count to the whole table in the same way
177 : * that VACUUM does. (Like VACUUM, we're not taking a random sample, so
178 : * just extrapolating linearly seems unsafe.) There should be no dead
179 : * tuples in all-visible pages, so no correction is needed for that, and
180 : * we already accounted for the space in those pages, too.
181 : */
182 4 : stat->tuple_count = vac_estimate_reltuples(rel, nblocks, scanned,
183 2 : stat->tuple_count);
184 :
185 : /* It's not clear if we could get -1 here, but be safe. */
186 2 : stat->tuple_count = Max(stat->tuple_count, 0);
187 :
188 : /*
189 : * Calculate percentages if the relation has one or more pages.
190 : */
191 2 : if (nblocks != 0)
192 : {
193 0 : stat->scanned_percent = 100.0 * scanned / nblocks;
194 0 : stat->tuple_percent = 100.0 * stat->tuple_len / stat->table_len;
195 0 : stat->dead_tuple_percent = 100.0 * stat->dead_tuple_len / stat->table_len;
196 0 : stat->free_percent = 100.0 * stat->free_space / stat->table_len;
197 : }
198 :
199 2 : if (BufferIsValid(vmbuffer))
200 : {
201 0 : ReleaseBuffer(vmbuffer);
202 0 : vmbuffer = InvalidBuffer;
203 : }
204 2 : }
205 :
206 : /*
207 : * Returns estimated live/dead tuple statistics for the given relid.
208 : *
209 : * The superuser() check here must be kept as the library might be upgraded
210 : * without the extension being upgraded, meaning that in pre-1.5 installations
211 : * these functions could be called by any user.
212 : */
213 : Datum
214 0 : pgstattuple_approx(PG_FUNCTION_ARGS)
215 : {
216 0 : Oid relid = PG_GETARG_OID(0);
217 :
218 0 : if (!superuser())
219 0 : ereport(ERROR,
220 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
221 : errmsg("must be superuser to use pgstattuple functions")));
222 :
223 0 : PG_RETURN_DATUM(pgstattuple_approx_internal(relid, fcinfo));
224 : }
225 :
226 : /*
227 : * As of pgstattuple version 1.5, we no longer need to check if the user
228 : * is a superuser because we REVOKE EXECUTE on the SQL function from PUBLIC.
229 : * Users can then grant access to it based on their policies.
230 : *
231 : * Otherwise identical to pgstattuple_approx (above).
232 : */
233 : Datum
234 6 : pgstattuple_approx_v1_5(PG_FUNCTION_ARGS)
235 : {
236 6 : Oid relid = PG_GETARG_OID(0);
237 :
238 6 : PG_RETURN_DATUM(pgstattuple_approx_internal(relid, fcinfo));
239 : }
240 :
241 : Datum
242 6 : pgstattuple_approx_internal(Oid relid, FunctionCallInfo fcinfo)
243 : {
244 : Relation rel;
245 6 : output_type stat = {0};
246 : TupleDesc tupdesc;
247 : bool nulls[NUM_OUTPUT_COLUMNS];
248 : Datum values[NUM_OUTPUT_COLUMNS];
249 : HeapTuple ret;
250 6 : int i = 0;
251 :
252 6 : if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
253 0 : elog(ERROR, "return type must be a row type");
254 :
255 6 : if (tupdesc->natts != NUM_OUTPUT_COLUMNS)
256 0 : elog(ERROR, "incorrect number of output arguments");
257 :
258 6 : rel = relation_open(relid, AccessShareLock);
259 :
260 : /*
261 : * Reject attempts to read non-local temporary relations; we would be
262 : * likely to get wrong data since we have no visibility into the owning
263 : * session's local buffers.
264 : */
265 6 : if (RELATION_IS_OTHER_TEMP(rel))
266 0 : ereport(ERROR,
267 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
268 : errmsg("cannot access temporary tables of other sessions")));
269 :
270 : /*
271 : * We support only relation kinds with a visibility map and a free space
272 : * map.
273 : */
274 6 : if (!(rel->rd_rel->relkind == RELKIND_RELATION ||
275 5 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
276 5 : rel->rd_rel->relkind == RELKIND_TOASTVALUE))
277 4 : ereport(ERROR,
278 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
279 : errmsg("relation \"%s\" is of wrong relation kind",
280 : RelationGetRelationName(rel)),
281 : errdetail_relkind_not_supported(rel->rd_rel->relkind)));
282 :
283 2 : if (rel->rd_rel->relam != HEAP_TABLE_AM_OID)
284 0 : ereport(ERROR, (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
285 : errmsg("only heap AM is supported")));
286 :
287 2 : statapprox_heap(rel, &stat);
288 :
289 2 : relation_close(rel, AccessShareLock);
290 :
291 2 : memset(nulls, 0, sizeof(nulls));
292 :
293 2 : values[i++] = Int64GetDatum(stat.table_len);
294 2 : values[i++] = Float8GetDatum(stat.scanned_percent);
295 2 : values[i++] = Int64GetDatum(stat.tuple_count);
296 2 : values[i++] = Int64GetDatum(stat.tuple_len);
297 2 : values[i++] = Float8GetDatum(stat.tuple_percent);
298 2 : values[i++] = Int64GetDatum(stat.dead_tuple_count);
299 2 : values[i++] = Int64GetDatum(stat.dead_tuple_len);
300 2 : values[i++] = Float8GetDatum(stat.dead_tuple_percent);
301 2 : values[i++] = Int64GetDatum(stat.free_space);
302 2 : values[i++] = Float8GetDatum(stat.free_percent);
303 :
304 2 : ret = heap_form_tuple(tupdesc, values, nulls);
305 2 : return HeapTupleGetDatum(ret);
306 : }
|
