granicus.if.org Git - postgresql/blob - src/backend/commands/analyze.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * analyze.c
   4  *        the Postgres statistics generator
   5  *
   6  * Portions Copyright (c) 1996-2016, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/commands/analyze.c
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15 #include "postgres.h"
  16
  17 #include <math.h>
  18
  19 #include "access/multixact.h"
  20 #include "access/transam.h"
  21 #include "access/tupconvert.h"
  22 #include "access/tuptoaster.h"
  23 #include "access/visibilitymap.h"
  24 #include "access/xact.h"
  25 #include "catalog/catalog.h"
  26 #include "catalog/index.h"
  27 #include "catalog/indexing.h"
  28 #include "catalog/pg_collation.h"
  29 #include "catalog/pg_inherits_fn.h"
  30 #include "catalog/pg_namespace.h"
  31 #include "commands/dbcommands.h"
  32 #include "commands/tablecmds.h"
  33 #include "commands/vacuum.h"
  34 #include "executor/executor.h"
  35 #include "foreign/fdwapi.h"
  36 #include "miscadmin.h"
  37 #include "nodes/nodeFuncs.h"
  38 #include "parser/parse_oper.h"
  39 #include "parser/parse_relation.h"
  40 #include "pgstat.h"
  41 #include "postmaster/autovacuum.h"
  42 #include "storage/bufmgr.h"
  43 #include "storage/lmgr.h"
  44 #include "storage/proc.h"
  45 #include "storage/procarray.h"
  46 #include "utils/acl.h"
  47 #include "utils/attoptcache.h"
  48 #include "utils/datum.h"
  49 #include "utils/guc.h"
  50 #include "utils/lsyscache.h"
  51 #include "utils/memutils.h"
  52 #include "utils/pg_rusage.h"
  53 #include "utils/sampling.h"
  54 #include "utils/sortsupport.h"
  55 #include "utils/syscache.h"
  56 #include "utils/timestamp.h"
  57 #include "utils/tqual.h"
  58
  59
  60 /* Per-index data for ANALYZE */
  61 typedef struct AnlIndexData
  62 {
  63         IndexInfo  *indexInfo;          /* BuildIndexInfo result */
  64         double          tupleFract;             /* fraction of rows for partial index */
  65         VacAttrStats **vacattrstats;    /* index attrs to analyze */
  66         int                     attr_cnt;
  67 } AnlIndexData;
  68
  69
  70 /* Default statistics target (GUC parameter) */
  71 int                     default_statistics_target = 100;
  72
  73 /* A few variables that don't seem worth passing around as parameters */
  74 static MemoryContext anl_context = NULL;
  75 static BufferAccessStrategy vac_strategy;
  76
  77
  78 static void do_analyze_rel(Relation onerel, int options,
  79                            VacuumParams *params, List *va_cols,
  80                            AcquireSampleRowsFunc acquirefunc, BlockNumber relpages,
  81                            bool inh, bool in_outer_xact, int elevel);
  82 static void compute_index_stats(Relation onerel, double totalrows,
  83                                         AnlIndexData *indexdata, int nindexes,
  84                                         HeapTuple *rows, int numrows,
  85                                         MemoryContext col_context);
  86 static VacAttrStats *examine_attribute(Relation onerel, int attnum,
  87                                   Node *index_expr);
  88 static int acquire_sample_rows(Relation onerel, int elevel,
  89                                         HeapTuple *rows, int targrows,
  90                                         double *totalrows, double *totaldeadrows);
  91 static int      compare_rows(const void *a, const void *b);
  92 static int acquire_inherited_sample_rows(Relation onerel, int elevel,
  93                                                           HeapTuple *rows, int targrows,
  94                                                           double *totalrows, double *totaldeadrows);
  95 static void update_attstats(Oid relid, bool inh,
  96                                 int natts, VacAttrStats **vacattrstats);
  97 static Datum std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
  98 static Datum ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull);
  99
 100
 101 /*
 102  *      analyze_rel() -- analyze one relation
 103  */
 104 void
 105 analyze_rel(Oid relid, RangeVar *relation, int options,
 106                         VacuumParams *params, List *va_cols, bool in_outer_xact,
 107                         BufferAccessStrategy bstrategy)
 108 {
 109         Relation        onerel;
 110         int                     elevel;
 111         AcquireSampleRowsFunc acquirefunc = NULL;
 112         BlockNumber relpages = 0;
 113
 114         /* Select logging level */
 115         if (options & VACOPT_VERBOSE)
 116                 elevel = INFO;
 117         else
 118                 elevel = DEBUG2;
 119
 120         /* Set up static variables */
 121         vac_strategy = bstrategy;
 122
 123         /*
 124          * Check for user-requested abort.
 125          */
 126         CHECK_FOR_INTERRUPTS();
 127
 128         /*
 129          * Open the relation, getting ShareUpdateExclusiveLock to ensure that two
 130          * ANALYZEs don't run on it concurrently.  (This also locks out a
 131          * concurrent VACUUM, which doesn't matter much at the moment but might
 132          * matter if we ever try to accumulate stats on dead tuples.) If the rel
 133          * has been dropped since we last saw it, we don't need to process it.
 134          */
 135         if (!(options & VACOPT_NOWAIT))
 136                 onerel = try_relation_open(relid, ShareUpdateExclusiveLock);
 137         else if (ConditionalLockRelationOid(relid, ShareUpdateExclusiveLock))
 138                 onerel = try_relation_open(relid, NoLock);
 139         else
 140         {
 141                 onerel = NULL;
 142                 if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
 143                         ereport(LOG,
 144                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
 145                                   errmsg("skipping analyze of \"%s\" --- lock not available",
 146                                                  relation->relname)));
 147         }
 148         if (!onerel)
 149                 return;
 150
 151         /*
 152          * Check permissions --- this should match vacuum's check!
 153          */
 154         if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
 155                   (pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
 156         {
 157                 /* No need for a WARNING if we already complained during VACUUM */
 158                 if (!(options & VACOPT_VACUUM))
 159                 {
 160                         if (onerel->rd_rel->relisshared)
 161                                 ereport(WARNING,
 162                                  (errmsg("skipping \"%s\" --- only superuser can analyze it",
 163                                                  RelationGetRelationName(onerel))));
 164                         else if (onerel->rd_rel->relnamespace == PG_CATALOG_NAMESPACE)
 165                                 ereport(WARNING,
 166                                                 (errmsg("skipping \"%s\" --- only superuser or database owner can analyze it",
 167                                                                 RelationGetRelationName(onerel))));
 168                         else
 169                                 ereport(WARNING,
 170                                                 (errmsg("skipping \"%s\" --- only table or database owner can analyze it",
 171                                                                 RelationGetRelationName(onerel))));
 172                 }
 173                 relation_close(onerel, ShareUpdateExclusiveLock);
 174                 return;
 175         }
 176
 177         /*
 178          * Silently ignore tables that are temp tables of other backends ---
 179          * trying to analyze these is rather pointless, since their contents are
 180          * probably not up-to-date on disk.  (We don't throw a warning here; it
 181          * would just lead to chatter during a database-wide ANALYZE.)
 182          */
 183         if (RELATION_IS_OTHER_TEMP(onerel))
 184         {
 185                 relation_close(onerel, ShareUpdateExclusiveLock);
 186                 return;
 187         }
 188
 189         /*
 190          * We can ANALYZE any table except pg_statistic. See update_attstats
 191          */
 192         if (RelationGetRelid(onerel) == StatisticRelationId)
 193         {
 194                 relation_close(onerel, ShareUpdateExclusiveLock);
 195                 return;
 196         }
 197
 198         /*
 199          * Check that it's a plain table, materialized view, or foreign table; we
 200          * used to do this in get_rel_oids() but seems safer to check after we've
 201          * locked the relation.
 202          */
 203         if (onerel->rd_rel->relkind == RELKIND_RELATION ||
 204                 onerel->rd_rel->relkind == RELKIND_MATVIEW)
 205         {
 206                 /* Regular table, so we'll use the regular row acquisition function */
 207                 acquirefunc = acquire_sample_rows;
 208                 /* Also get regular table's size */
 209                 relpages = RelationGetNumberOfBlocks(onerel);
 210         }
 211         else if (onerel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
 212         {
 213                 /*
 214                  * For a foreign table, call the FDW's hook function to see whether it
 215                  * supports analysis.
 216                  */
 217                 FdwRoutine *fdwroutine;
 218                 bool            ok = false;
 219
 220                 fdwroutine = GetFdwRoutineForRelation(onerel, false);
 221
 222                 if (fdwroutine->AnalyzeForeignTable != NULL)
 223                         ok = fdwroutine->AnalyzeForeignTable(onerel,
 224                                                                                                  &acquirefunc,
 225                                                                                                  &relpages);
 226
 227                 if (!ok)
 228                 {
 229                         ereport(WARNING,
 230                          (errmsg("skipping \"%s\" --- cannot analyze this foreign table",
 231                                          RelationGetRelationName(onerel))));
 232                         relation_close(onerel, ShareUpdateExclusiveLock);
 233                         return;
 234                 }
 235         }
 236         else
 237         {
 238                 /* No need for a WARNING if we already complained during VACUUM */
 239                 if (!(options & VACOPT_VACUUM))
 240                         ereport(WARNING,
 241                                         (errmsg("skipping \"%s\" --- cannot analyze non-tables or special system tables",
 242                                                         RelationGetRelationName(onerel))));
 243                 relation_close(onerel, ShareUpdateExclusiveLock);
 244                 return;
 245         }
 246
 247         /*
 248          * OK, let's do it.  First let other backends know I'm in ANALYZE.
 249          */
 250         LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
 251         MyPgXact->vacuumFlags |= PROC_IN_ANALYZE;
 252         LWLockRelease(ProcArrayLock);
 253
 254         /*
 255          * Do the normal non-recursive ANALYZE.
 256          */
 257         do_analyze_rel(onerel, options, params, va_cols, acquirefunc, relpages,
 258                                    false, in_outer_xact, elevel);
 259
 260         /*
 261          * If there are child tables, do recursive ANALYZE.
 262          */
 263         if (onerel->rd_rel->relhassubclass)
 264                 do_analyze_rel(onerel, options, params, va_cols, acquirefunc, relpages,
 265                                            true, in_outer_xact, elevel);
 266
 267         /*
 268          * Close source relation now, but keep lock so that no one deletes it
 269          * before we commit.  (If someone did, they'd fail to clean up the entries
 270          * we made in pg_statistic.  Also, releasing the lock before commit would
 271          * expose us to concurrent-update failures in update_attstats.)
 272          */
 273         relation_close(onerel, NoLock);
 274
 275         /*
 276          * Reset my PGXACT flag.  Note: we need this here, and not in vacuum_rel,
 277          * because the vacuum flag is cleared by the end-of-xact code.
 278          */
 279         LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
 280         MyPgXact->vacuumFlags &= ~PROC_IN_ANALYZE;
 281         LWLockRelease(ProcArrayLock);
 282 }
 283
 284 /*
 285  *      do_analyze_rel() -- analyze one relation, recursively or not
 286  *
 287  * Note that "acquirefunc" is only relevant for the non-inherited case.
 288  * For the inherited case, acquire_inherited_sample_rows() determines the
 289  * appropriate acquirefunc for each child table.
 290  */
 291 static void
 292 do_analyze_rel(Relation onerel, int options, VacuumParams *params,
 293                            List *va_cols, AcquireSampleRowsFunc acquirefunc,
 294                            BlockNumber relpages, bool inh, bool in_outer_xact,
 295                            int elevel)
 296 {
 297         int                     attr_cnt,
 298                                 tcnt,
 299                                 i,
 300                                 ind;
 301         Relation   *Irel;
 302         int                     nindexes;
 303         bool            hasindex;
 304         VacAttrStats **vacattrstats;
 305         AnlIndexData *indexdata;
 306         int                     targrows,
 307                                 numrows;
 308         double          totalrows,
 309                                 totaldeadrows;
 310         HeapTuple  *rows;
 311         PGRUsage        ru0;
 312         TimestampTz starttime = 0;
 313         MemoryContext caller_context;
 314         Oid                     save_userid;
 315         int                     save_sec_context;
 316         int                     save_nestlevel;
 317
 318         if (inh)
 319                 ereport(elevel,
 320                                 (errmsg("analyzing \"%s.%s\" inheritance tree",
 321                                                 get_namespace_name(RelationGetNamespace(onerel)),
 322                                                 RelationGetRelationName(onerel))));
 323         else
 324                 ereport(elevel,
 325                                 (errmsg("analyzing \"%s.%s\"",
 326                                                 get_namespace_name(RelationGetNamespace(onerel)),
 327                                                 RelationGetRelationName(onerel))));
 328
 329         /*
 330          * Set up a working context so that we can easily free whatever junk gets
 331          * created.
 332          */
 333         anl_context = AllocSetContextCreate(CurrentMemoryContext,
 334                                                                                 "Analyze",
 335                                                                                 ALLOCSET_DEFAULT_MINSIZE,
 336                                                                                 ALLOCSET_DEFAULT_INITSIZE,
 337                                                                                 ALLOCSET_DEFAULT_MAXSIZE);
 338         caller_context = MemoryContextSwitchTo(anl_context);
 339
 340         /*
 341          * Switch to the table owner's userid, so that any index functions are run
 342          * as that user.  Also lock down security-restricted operations and
 343          * arrange to make GUC variable changes local to this command.
 344          */
 345         GetUserIdAndSecContext(&save_userid, &save_sec_context);
 346         SetUserIdAndSecContext(onerel->rd_rel->relowner,
 347                                                    save_sec_context | SECURITY_RESTRICTED_OPERATION);
 348         save_nestlevel = NewGUCNestLevel();
 349
 350         /* measure elapsed time iff autovacuum logging requires it */
 351         if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
 352         {
 353                 pg_rusage_init(&ru0);
 354                 if (params->log_min_duration > 0)
 355                         starttime = GetCurrentTimestamp();
 356         }
 357
 358         /*
 359          * Determine which columns to analyze
 360          *
 361          * Note that system attributes are never analyzed.
 362          */
 363         if (va_cols != NIL)
 364         {
 365                 ListCell   *le;
 366
 367                 vacattrstats = (VacAttrStats **) palloc(list_length(va_cols) *
 368                                                                                                 sizeof(VacAttrStats *));
 369                 tcnt = 0;
 370                 foreach(le, va_cols)
 371                 {
 372                         char       *col = strVal(lfirst(le));
 373
 374                         i = attnameAttNum(onerel, col, false);
 375                         if (i == InvalidAttrNumber)
 376                                 ereport(ERROR,
 377                                                 (errcode(ERRCODE_UNDEFINED_COLUMN),
 378                                         errmsg("column \"%s\" of relation \"%s\" does not exist",
 379                                                    col, RelationGetRelationName(onerel))));
 380                         vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
 381                         if (vacattrstats[tcnt] != NULL)
 382                                 tcnt++;
 383                 }
 384                 attr_cnt = tcnt;
 385         }
 386         else
 387         {
 388                 attr_cnt = onerel->rd_att->natts;
 389                 vacattrstats = (VacAttrStats **)
 390                         palloc(attr_cnt * sizeof(VacAttrStats *));
 391                 tcnt = 0;
 392                 for (i = 1; i <= attr_cnt; i++)
 393                 {
 394                         vacattrstats[tcnt] = examine_attribute(onerel, i, NULL);
 395                         if (vacattrstats[tcnt] != NULL)
 396                                 tcnt++;
 397                 }
 398                 attr_cnt = tcnt;
 399         }
 400
 401         /*
 402          * Open all indexes of the relation, and see if there are any analyzable
 403          * columns in the indexes.  We do not analyze index columns if there was
 404          * an explicit column list in the ANALYZE command, however.  If we are
 405          * doing a recursive scan, we don't want to touch the parent's indexes at
 406          * all.
 407          */
 408         if (!inh)
 409                 vac_open_indexes(onerel, AccessShareLock, &nindexes, &Irel);
 410         else
 411         {
 412                 Irel = NULL;
 413                 nindexes = 0;
 414         }
 415         hasindex = (nindexes > 0);
 416         indexdata = NULL;
 417         if (hasindex)
 418         {
 419                 indexdata = (AnlIndexData *) palloc0(nindexes * sizeof(AnlIndexData));
 420                 for (ind = 0; ind < nindexes; ind++)
 421                 {
 422                         AnlIndexData *thisdata = &indexdata[ind];
 423                         IndexInfo  *indexInfo;
 424
 425                         thisdata->indexInfo = indexInfo = BuildIndexInfo(Irel[ind]);
 426                         thisdata->tupleFract = 1.0; /* fix later if partial */
 427                         if (indexInfo->ii_Expressions != NIL && va_cols == NIL)
 428                         {
 429                                 ListCell   *indexpr_item = list_head(indexInfo->ii_Expressions);
 430
 431                                 thisdata->vacattrstats = (VacAttrStats **)
 432                                         palloc(indexInfo->ii_NumIndexAttrs * sizeof(VacAttrStats *));
 433                                 tcnt = 0;
 434                                 for (i = 0; i < indexInfo->ii_NumIndexAttrs; i++)
 435                                 {
 436                                         int                     keycol = indexInfo->ii_KeyAttrNumbers[i];
 437
 438                                         if (keycol == 0)
 439                                         {
 440                                                 /* Found an index expression */
 441                                                 Node       *indexkey;
 442
 443                                                 if (indexpr_item == NULL)               /* shouldn't happen */
 444                                                         elog(ERROR, "too few entries in indexprs list");
 445                                                 indexkey = (Node *) lfirst(indexpr_item);
 446                                                 indexpr_item = lnext(indexpr_item);
 447                                                 thisdata->vacattrstats[tcnt] =
 448                                                         examine_attribute(Irel[ind], i + 1, indexkey);
 449                                                 if (thisdata->vacattrstats[tcnt] != NULL)
 450                                                         tcnt++;
 451                                         }
 452                                 }
 453                                 thisdata->attr_cnt = tcnt;
 454                         }
 455                 }
 456         }
 457
 458         /*
 459          * Determine how many rows we need to sample, using the worst case from
 460          * all analyzable columns.  We use a lower bound of 100 rows to avoid
 461          * possible overflow in Vitter's algorithm.  (Note: that will also be the
 462          * target in the corner case where there are no analyzable columns.)
 463          */
 464         targrows = 100;
 465         for (i = 0; i < attr_cnt; i++)
 466         {
 467                 if (targrows < vacattrstats[i]->minrows)
 468                         targrows = vacattrstats[i]->minrows;
 469         }
 470         for (ind = 0; ind < nindexes; ind++)
 471         {
 472                 AnlIndexData *thisdata = &indexdata[ind];
 473
 474                 for (i = 0; i < thisdata->attr_cnt; i++)
 475                 {
 476                         if (targrows < thisdata->vacattrstats[i]->minrows)
 477                                 targrows = thisdata->vacattrstats[i]->minrows;
 478                 }
 479         }
 480
 481         /*
 482          * Acquire the sample rows
 483          */
 484         rows = (HeapTuple *) palloc(targrows * sizeof(HeapTuple));
 485         if (inh)
 486                 numrows = acquire_inherited_sample_rows(onerel, elevel,
 487                                                                                                 rows, targrows,
 488                                                                                                 &totalrows, &totaldeadrows);
 489         else
 490                 numrows = (*acquirefunc) (onerel, elevel,
 491                                                                   rows, targrows,
 492                                                                   &totalrows, &totaldeadrows);
 493
 494         /*
 495          * Compute the statistics.  Temporary results during the calculations for
 496          * each column are stored in a child context.  The calc routines are
 497          * responsible to make sure that whatever they store into the VacAttrStats
 498          * structure is allocated in anl_context.
 499          */
 500         if (numrows > 0)
 501         {
 502                 MemoryContext col_context,
 503                                         old_context;
 504
 505                 col_context = AllocSetContextCreate(anl_context,
 506                                                                                         "Analyze Column",
 507                                                                                         ALLOCSET_DEFAULT_MINSIZE,
 508                                                                                         ALLOCSET_DEFAULT_INITSIZE,
 509                                                                                         ALLOCSET_DEFAULT_MAXSIZE);
 510                 old_context = MemoryContextSwitchTo(col_context);
 511
 512                 for (i = 0; i < attr_cnt; i++)
 513                 {
 514                         VacAttrStats *stats = vacattrstats[i];
 515                         AttributeOpts *aopt;
 516
 517                         stats->rows = rows;
 518                         stats->tupDesc = onerel->rd_att;
 519                         (*stats->compute_stats) (stats,
 520                                                                          std_fetch_func,
 521                                                                          numrows,
 522                                                                          totalrows);
 523
 524                         /*
 525                          * If the appropriate flavor of the n_distinct option is
 526                          * specified, override with the corresponding value.
 527                          */
 528                         aopt = get_attribute_options(onerel->rd_id, stats->attr->attnum);
 529                         if (aopt != NULL)
 530                         {
 531                                 float8          n_distinct;
 532
 533                                 n_distinct = inh ? aopt->n_distinct_inherited : aopt->n_distinct;
 534                                 if (n_distinct != 0.0)
 535                                         stats->stadistinct = n_distinct;
 536                         }
 537
 538                         MemoryContextResetAndDeleteChildren(col_context);
 539                 }
 540
 541                 if (hasindex)
 542                         compute_index_stats(onerel, totalrows,
 543                                                                 indexdata, nindexes,
 544                                                                 rows, numrows,
 545                                                                 col_context);
 546
 547                 MemoryContextSwitchTo(old_context);
 548                 MemoryContextDelete(col_context);
 549
 550                 /*
 551                  * Emit the completed stats rows into pg_statistic, replacing any
 552                  * previous statistics for the target columns.  (If there are stats in
 553                  * pg_statistic for columns we didn't process, we leave them alone.)
 554                  */
 555                 update_attstats(RelationGetRelid(onerel), inh,
 556                                                 attr_cnt, vacattrstats);
 557
 558                 for (ind = 0; ind < nindexes; ind++)
 559                 {
 560                         AnlIndexData *thisdata = &indexdata[ind];
 561
 562                         update_attstats(RelationGetRelid(Irel[ind]), false,
 563                                                         thisdata->attr_cnt, thisdata->vacattrstats);
 564                 }
 565         }
 566
 567         /*
 568          * Update pages/tuples stats in pg_class ... but not if we're doing
 569          * inherited stats.
 570          */
 571         if (!inh)
 572         {
 573                 BlockNumber             relallvisible;
 574
 575                 visibilitymap_count(onerel, &relallvisible, NULL);
 576
 577                 vac_update_relstats(onerel,
 578                                                         relpages,
 579                                                         totalrows,
 580                                                         relallvisible,
 581                                                         hasindex,
 582                                                         InvalidTransactionId,
 583                                                         InvalidMultiXactId,
 584                                                         in_outer_xact);
 585         }
 586
 587         /*
 588          * Same for indexes. Vacuum always scans all indexes, so if we're part of
 589          * VACUUM ANALYZE, don't overwrite the accurate count already inserted by
 590          * VACUUM.
 591          */
 592         if (!inh && !(options & VACOPT_VACUUM))
 593         {
 594                 for (ind = 0; ind < nindexes; ind++)
 595                 {
 596                         AnlIndexData *thisdata = &indexdata[ind];
 597                         double          totalindexrows;
 598
 599                         totalindexrows = ceil(thisdata->tupleFract * totalrows);
 600                         vac_update_relstats(Irel[ind],
 601                                                                 RelationGetNumberOfBlocks(Irel[ind]),
 602                                                                 totalindexrows,
 603                                                                 0,
 604                                                                 false,
 605                                                                 InvalidTransactionId,
 606                                                                 InvalidMultiXactId,
 607                                                                 in_outer_xact);
 608                 }
 609         }
 610
 611         /*
 612          * Report ANALYZE to the stats collector, too.  However, if doing
 613          * inherited stats we shouldn't report, because the stats collector only
 614          * tracks per-table stats.
 615          */
 616         if (!inh)
 617                 pgstat_report_analyze(onerel, totalrows, totaldeadrows);
 618
 619         /* If this isn't part of VACUUM ANALYZE, let index AMs do cleanup */
 620         if (!(options & VACOPT_VACUUM))
 621         {
 622                 for (ind = 0; ind < nindexes; ind++)
 623                 {
 624                         IndexBulkDeleteResult *stats;
 625                         IndexVacuumInfo ivinfo;
 626
 627                         ivinfo.index = Irel[ind];
 628                         ivinfo.analyze_only = true;
 629                         ivinfo.estimated_count = true;
 630                         ivinfo.message_level = elevel;
 631                         ivinfo.num_heap_tuples = onerel->rd_rel->reltuples;
 632                         ivinfo.strategy = vac_strategy;
 633
 634                         stats = index_vacuum_cleanup(&ivinfo, NULL);
 635
 636                         if (stats)
 637                                 pfree(stats);
 638                 }
 639         }
 640
 641         /* Done with indexes */
 642         vac_close_indexes(nindexes, Irel, NoLock);
 643
 644         /* Log the action if appropriate */
 645         if (IsAutoVacuumWorkerProcess() && params->log_min_duration >= 0)
 646         {
 647                 if (params->log_min_duration == 0 ||
 648                         TimestampDifferenceExceeds(starttime, GetCurrentTimestamp(),
 649                                                                            params->log_min_duration))
 650                         ereport(LOG,
 651                                         (errmsg("automatic analyze of table \"%s.%s.%s\" system usage: %s",
 652                                                         get_database_name(MyDatabaseId),
 653                                                         get_namespace_name(RelationGetNamespace(onerel)),
 654                                                         RelationGetRelationName(onerel),
 655                                                         pg_rusage_show(&ru0))));
 656         }
 657
 658         /* Roll back any GUC changes executed by index functions */
 659         AtEOXact_GUC(false, save_nestlevel);
 660
 661         /* Restore userid and security context */
 662         SetUserIdAndSecContext(save_userid, save_sec_context);
 663
 664         /* Restore current context and release memory */
 665         MemoryContextSwitchTo(caller_context);
 666         MemoryContextDelete(anl_context);
 667         anl_context = NULL;
 668 }
 669
 670 /*
 671  * Compute statistics about indexes of a relation
 672  */
 673 static void
 674 compute_index_stats(Relation onerel, double totalrows,
 675                                         AnlIndexData *indexdata, int nindexes,
 676                                         HeapTuple *rows, int numrows,
 677                                         MemoryContext col_context)
 678 {
 679         MemoryContext ind_context,
 680                                 old_context;
 681         Datum           values[INDEX_MAX_KEYS];
 682         bool            isnull[INDEX_MAX_KEYS];
 683         int                     ind,
 684                                 i;
 685
 686         ind_context = AllocSetContextCreate(anl_context,
 687                                                                                 "Analyze Index",
 688                                                                                 ALLOCSET_DEFAULT_MINSIZE,
 689                                                                                 ALLOCSET_DEFAULT_INITSIZE,
 690                                                                                 ALLOCSET_DEFAULT_MAXSIZE);
 691         old_context = MemoryContextSwitchTo(ind_context);
 692
 693         for (ind = 0; ind < nindexes; ind++)
 694         {
 695                 AnlIndexData *thisdata = &indexdata[ind];
 696                 IndexInfo  *indexInfo = thisdata->indexInfo;
 697                 int                     attr_cnt = thisdata->attr_cnt;
 698                 TupleTableSlot *slot;
 699                 EState     *estate;
 700                 ExprContext *econtext;
 701                 List       *predicate;
 702                 Datum      *exprvals;
 703                 bool       *exprnulls;
 704                 int                     numindexrows,
 705                                         tcnt,
 706                                         rowno;
 707                 double          totalindexrows;
 708
 709                 /* Ignore index if no columns to analyze and not partial */
 710                 if (attr_cnt == 0 && indexInfo->ii_Predicate == NIL)
 711                         continue;
 712
 713                 /*
 714                  * Need an EState for evaluation of index expressions and
 715                  * partial-index predicates.  Create it in the per-index context to be
 716                  * sure it gets cleaned up at the bottom of the loop.
 717                  */
 718                 estate = CreateExecutorState();
 719                 econtext = GetPerTupleExprContext(estate);
 720                 /* Need a slot to hold the current heap tuple, too */
 721                 slot = MakeSingleTupleTableSlot(RelationGetDescr(onerel));
 722
 723                 /* Arrange for econtext's scan tuple to be the tuple under test */
 724                 econtext->ecxt_scantuple = slot;
 725
 726                 /* Set up execution state for predicate. */
 727                 predicate = (List *)
 728                         ExecPrepareExpr((Expr *) indexInfo->ii_Predicate,
 729                                                         estate);
 730
 731                 /* Compute and save index expression values */
 732                 exprvals = (Datum *) palloc(numrows * attr_cnt * sizeof(Datum));
 733                 exprnulls = (bool *) palloc(numrows * attr_cnt * sizeof(bool));
 734                 numindexrows = 0;
 735                 tcnt = 0;
 736                 for (rowno = 0; rowno < numrows; rowno++)
 737                 {
 738                         HeapTuple       heapTuple = rows[rowno];
 739
 740                         vacuum_delay_point();
 741
 742                         /*
 743                          * Reset the per-tuple context each time, to reclaim any cruft
 744                          * left behind by evaluating the predicate or index expressions.
 745                          */
 746                         ResetExprContext(econtext);
 747
 748                         /* Set up for predicate or expression evaluation */
 749                         ExecStoreTuple(heapTuple, slot, InvalidBuffer, false);
 750
 751                         /* If index is partial, check predicate */
 752                         if (predicate != NIL)
 753                         {
 754                                 if (!ExecQual(predicate, econtext, false))
 755                                         continue;
 756                         }
 757                         numindexrows++;
 758
 759                         if (attr_cnt > 0)
 760                         {
 761                                 /*
 762                                  * Evaluate the index row to compute expression values. We
 763                                  * could do this by hand, but FormIndexDatum is convenient.
 764                                  */
 765                                 FormIndexDatum(indexInfo,
 766                                                            slot,
 767                                                            estate,
 768                                                            values,
 769                                                            isnull);
 770
 771                                 /*
 772                                  * Save just the columns we care about.  We copy the values
 773                                  * into ind_context from the estate's per-tuple context.
 774                                  */
 775                                 for (i = 0; i < attr_cnt; i++)
 776                                 {
 777                                         VacAttrStats *stats = thisdata->vacattrstats[i];
 778                                         int                     attnum = stats->attr->attnum;
 779
 780                                         if (isnull[attnum - 1])
 781                                         {
 782                                                 exprvals[tcnt] = (Datum) 0;
 783                                                 exprnulls[tcnt] = true;
 784                                         }
 785                                         else
 786                                         {
 787                                                 exprvals[tcnt] = datumCopy(values[attnum - 1],
 788                                                                                                    stats->attrtype->typbyval,
 789                                                                                                    stats->attrtype->typlen);
 790                                                 exprnulls[tcnt] = false;
 791                                         }
 792                                         tcnt++;
 793                                 }
 794                         }
 795                 }
 796
 797                 /*
 798                  * Having counted the number of rows that pass the predicate in the
 799                  * sample, we can estimate the total number of rows in the index.
 800                  */
 801                 thisdata->tupleFract = (double) numindexrows / (double) numrows;
 802                 totalindexrows = ceil(thisdata->tupleFract * totalrows);
 803
 804                 /*
 805                  * Now we can compute the statistics for the expression columns.
 806                  */
 807                 if (numindexrows > 0)
 808                 {
 809                         MemoryContextSwitchTo(col_context);
 810                         for (i = 0; i < attr_cnt; i++)
 811                         {
 812                                 VacAttrStats *stats = thisdata->vacattrstats[i];
 813                                 AttributeOpts *aopt =
 814                                 get_attribute_options(stats->attr->attrelid,
 815                                                                           stats->attr->attnum);
 816
 817                                 stats->exprvals = exprvals + i;
 818                                 stats->exprnulls = exprnulls + i;
 819                                 stats->rowstride = attr_cnt;
 820                                 (*stats->compute_stats) (stats,
 821                                                                                  ind_fetch_func,
 822                                                                                  numindexrows,
 823                                                                                  totalindexrows);
 824
 825                                 /*
 826                                  * If the n_distinct option is specified, it overrides the
 827                                  * above computation.  For indices, we always use just
 828                                  * n_distinct, not n_distinct_inherited.
 829                                  */
 830                                 if (aopt != NULL && aopt->n_distinct != 0.0)
 831                                         stats->stadistinct = aopt->n_distinct;
 832
 833                                 MemoryContextResetAndDeleteChildren(col_context);
 834                         }
 835                 }
 836
 837                 /* And clean up */
 838                 MemoryContextSwitchTo(ind_context);
 839
 840                 ExecDropSingleTupleTableSlot(slot);
 841                 FreeExecutorState(estate);
 842                 MemoryContextResetAndDeleteChildren(ind_context);
 843         }
 844
 845         MemoryContextSwitchTo(old_context);
 846         MemoryContextDelete(ind_context);
 847 }
 848
 849 /*
 850  * examine_attribute -- pre-analysis of a single column
 851  *
 852  * Determine whether the column is analyzable; if so, create and initialize
 853  * a VacAttrStats struct for it.  If not, return NULL.
 854  *
 855  * If index_expr isn't NULL, then we're trying to analyze an expression index,
 856  * and index_expr is the expression tree representing the column's data.
 857  */
 858 static VacAttrStats *
 859 examine_attribute(Relation onerel, int attnum, Node *index_expr)
 860 {
 861         Form_pg_attribute attr = onerel->rd_att->attrs[attnum - 1];
 862         HeapTuple       typtuple;
 863         VacAttrStats *stats;
 864         int                     i;
 865         bool            ok;
 866
 867         /* Never analyze dropped columns */
 868         if (attr->attisdropped)
 869                 return NULL;
 870
 871         /* Don't analyze column if user has specified not to */
 872         if (attr->attstattarget == 0)
 873                 return NULL;
 874
 875         /*
 876          * Create the VacAttrStats struct.  Note that we only have a copy of the
 877          * fixed fields of the pg_attribute tuple.
 878          */
 879         stats = (VacAttrStats *) palloc0(sizeof(VacAttrStats));
 880         stats->attr = (Form_pg_attribute) palloc(ATTRIBUTE_FIXED_PART_SIZE);
 881         memcpy(stats->attr, attr, ATTRIBUTE_FIXED_PART_SIZE);
 882
 883         /*
 884          * When analyzing an expression index, believe the expression tree's type
 885          * not the column datatype --- the latter might be the opckeytype storage
 886          * type of the opclass, which is not interesting for our purposes.  (Note:
 887          * if we did anything with non-expression index columns, we'd need to
 888          * figure out where to get the correct type info from, but for now that's
 889          * not a problem.)      It's not clear whether anyone will care about the
 890          * typmod, but we store that too just in case.
 891          */
 892         if (index_expr)
 893         {
 894                 stats->attrtypid = exprType(index_expr);
 895                 stats->attrtypmod = exprTypmod(index_expr);
 896         }
 897         else
 898         {
 899                 stats->attrtypid = attr->atttypid;
 900                 stats->attrtypmod = attr->atttypmod;
 901         }
 902
 903         typtuple = SearchSysCacheCopy1(TYPEOID,
 904                                                                    ObjectIdGetDatum(stats->attrtypid));
 905         if (!HeapTupleIsValid(typtuple))
 906                 elog(ERROR, "cache lookup failed for type %u", stats->attrtypid);
 907         stats->attrtype = (Form_pg_type) GETSTRUCT(typtuple);
 908         stats->anl_context = anl_context;
 909         stats->tupattnum = attnum;
 910
 911         /*
 912          * The fields describing the stats->stavalues[n] element types default to
 913          * the type of the data being analyzed, but the type-specific typanalyze
 914          * function can change them if it wants to store something else.
 915          */
 916         for (i = 0; i < STATISTIC_NUM_SLOTS; i++)
 917         {
 918                 stats->statypid[i] = stats->attrtypid;
 919                 stats->statyplen[i] = stats->attrtype->typlen;
 920                 stats->statypbyval[i] = stats->attrtype->typbyval;
 921                 stats->statypalign[i] = stats->attrtype->typalign;
 922         }
 923
 924         /*
 925          * Call the type-specific typanalyze function.  If none is specified, use
 926          * std_typanalyze().
 927          */
 928         if (OidIsValid(stats->attrtype->typanalyze))
 929                 ok = DatumGetBool(OidFunctionCall1(stats->attrtype->typanalyze,
 930                                                                                    PointerGetDatum(stats)));
 931         else
 932                 ok = std_typanalyze(stats);
 933
 934         if (!ok || stats->compute_stats == NULL || stats->minrows <= 0)
 935         {
 936                 heap_freetuple(typtuple);
 937                 pfree(stats->attr);
 938                 pfree(stats);
 939                 return NULL;
 940         }
 941
 942         return stats;
 943 }
 944
 945 /*
 946  * acquire_sample_rows -- acquire a random sample of rows from the table
 947  *
 948  * Selected rows are returned in the caller-allocated array rows[], which
 949  * must have at least targrows entries.
 950  * The actual number of rows selected is returned as the function result.
 951  * We also estimate the total numbers of live and dead rows in the table,
 952  * and return them into *totalrows and *totaldeadrows, respectively.
 953  *
 954  * The returned list of tuples is in order by physical position in the table.
 955  * (We will rely on this later to derive correlation estimates.)
 956  *
 957  * As of May 2004 we use a new two-stage method:  Stage one selects up
 958  * to targrows random blocks (or all blocks, if there aren't so many).
 959  * Stage two scans these blocks and uses the Vitter algorithm to create
 960  * a random sample of targrows rows (or less, if there are less in the
 961  * sample of blocks).  The two stages are executed simultaneously: each
 962  * block is processed as soon as stage one returns its number and while
 963  * the rows are read stage two controls which ones are to be inserted
 964  * into the sample.
 965  *
 966  * Although every row has an equal chance of ending up in the final
 967  * sample, this sampling method is not perfect: not every possible
 968  * sample has an equal chance of being selected.  For large relations
 969  * the number of different blocks represented by the sample tends to be
 970  * too small.  We can live with that for now.  Improvements are welcome.
 971  *
 972  * An important property of this sampling method is that because we do
 973  * look at a statistically unbiased set of blocks, we should get
 974  * unbiased estimates of the average numbers of live and dead rows per
 975  * block.  The previous sampling method put too much credence in the row
 976  * density near the start of the table.
 977  */
 978 static int
 979 acquire_sample_rows(Relation onerel, int elevel,
 980                                         HeapTuple *rows, int targrows,
 981                                         double *totalrows, double *totaldeadrows)
 982 {
 983         int                     numrows = 0;    /* # rows now in reservoir */
 984         double          samplerows = 0; /* total # rows collected */
 985         double          liverows = 0;   /* # live rows seen */
 986         double          deadrows = 0;   /* # dead rows seen */
 987         double          rowstoskip = -1;        /* -1 means not set yet */
 988         BlockNumber totalblocks;
 989         TransactionId OldestXmin;
 990         BlockSamplerData bs;
 991         ReservoirStateData rstate;
 992
 993         Assert(targrows > 0);
 994
 995         totalblocks = RelationGetNumberOfBlocks(onerel);
 996
 997         /* Need a cutoff xmin for HeapTupleSatisfiesVacuum */
 998         OldestXmin = GetOldestXmin(onerel, true);
 999
1000         /* Prepare for sampling block numbers */
1001         BlockSampler_Init(&bs, totalblocks, targrows, random());
1002         /* Prepare for sampling rows */
1003         reservoir_init_selection_state(&rstate, targrows);
1004
1005         /* Outer loop over blocks to sample */
1006         while (BlockSampler_HasMore(&bs))
1007         {
1008                 BlockNumber targblock = BlockSampler_Next(&bs);
1009                 Buffer          targbuffer;
1010                 Page            targpage;
1011                 OffsetNumber targoffset,
1012                                         maxoffset;
1013
1014                 vacuum_delay_point();
1015
1016                 /*
1017                  * We must maintain a pin on the target page's buffer to ensure that
1018                  * the maxoffset value stays good (else concurrent VACUUM might delete
1019                  * tuples out from under us).  Hence, pin the page until we are done
1020                  * looking at it.  We also choose to hold sharelock on the buffer
1021                  * throughout --- we could release and re-acquire sharelock for each
1022                  * tuple, but since we aren't doing much work per tuple, the extra
1023                  * lock traffic is probably better avoided.
1024                  */
1025                 targbuffer = ReadBufferExtended(onerel, MAIN_FORKNUM, targblock,
1026                                                                                 RBM_NORMAL, vac_strategy);
1027                 LockBuffer(targbuffer, BUFFER_LOCK_SHARE);
1028                 targpage = BufferGetPage(targbuffer);
1029                 maxoffset = PageGetMaxOffsetNumber(targpage);
1030
1031                 /* Inner loop over all tuples on the selected page */
1032                 for (targoffset = FirstOffsetNumber; targoffset <= maxoffset; targoffset++)
1033                 {
1034                         ItemId          itemid;
1035                         HeapTupleData targtuple;
1036                         bool            sample_it = false;
1037
1038                         itemid = PageGetItemId(targpage, targoffset);
1039
1040                         /*
1041                          * We ignore unused and redirect line pointers.  DEAD line
1042                          * pointers should be counted as dead, because we need vacuum to
1043                          * run to get rid of them.  Note that this rule agrees with the
1044                          * way that heap_page_prune() counts things.
1045                          */
1046                         if (!ItemIdIsNormal(itemid))
1047                         {
1048                                 if (ItemIdIsDead(itemid))
1049                                         deadrows += 1;
1050                                 continue;
1051                         }
1052
1053                         ItemPointerSet(&targtuple.t_self, targblock, targoffset);
1054
1055                         targtuple.t_tableOid = RelationGetRelid(onerel);
1056                         targtuple.t_data = (HeapTupleHeader) PageGetItem(targpage, itemid);
1057                         targtuple.t_len = ItemIdGetLength(itemid);
1058
1059                         switch (HeapTupleSatisfiesVacuum(&targtuple,
1060                                                                                          OldestXmin,
1061                                                                                          targbuffer))
1062                         {
1063                                 case HEAPTUPLE_LIVE:
1064                                         sample_it = true;
1065                                         liverows += 1;
1066                                         break;
1067
1068                                 case HEAPTUPLE_DEAD:
1069                                 case HEAPTUPLE_RECENTLY_DEAD:
1070                                         /* Count dead and recently-dead rows */
1071                                         deadrows += 1;
1072                                         break;
1073
1074                                 case HEAPTUPLE_INSERT_IN_PROGRESS:
1075
1076                                         /*
1077                                          * Insert-in-progress rows are not counted.  We assume
1078                                          * that when the inserting transaction commits or aborts,
1079                                          * it will send a stats message to increment the proper
1080                                          * count.  This works right only if that transaction ends
1081                                          * after we finish analyzing the table; if things happen
1082                                          * in the other order, its stats update will be
1083                                          * overwritten by ours.  However, the error will be large
1084                                          * only if the other transaction runs long enough to
1085                                          * insert many tuples, so assuming it will finish after us
1086                                          * is the safer option.
1087                                          *
1088                                          * A special case is that the inserting transaction might
1089                                          * be our own.  In this case we should count and sample
1090                                          * the row, to accommodate users who load a table and
1091                                          * analyze it in one transaction.  (pgstat_report_analyze
1092                                          * has to adjust the numbers we send to the stats
1093                                          * collector to make this come out right.)
1094                                          */
1095                                         if (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetXmin(targtuple.t_data)))
1096                                         {
1097                                                 sample_it = true;
1098                                                 liverows += 1;
1099                                         }
1100                                         break;
1101
1102                                 case HEAPTUPLE_DELETE_IN_PROGRESS:
1103
1104                                         /*
1105                                          * We count delete-in-progress rows as still live, using
1106                                          * the same reasoning given above; but we don't bother to
1107                                          * include them in the sample.
1108                                          *
1109                                          * If the delete was done by our own transaction, however,
1110                                          * we must count the row as dead to make
1111                                          * pgstat_report_analyze's stats adjustments come out
1112                                          * right.  (Note: this works out properly when the row was
1113                                          * both inserted and deleted in our xact.)
1114                                          */
1115                                         if (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetUpdateXid(targtuple.t_data)))
1116                                                 deadrows += 1;
1117                                         else
1118                                                 liverows += 1;
1119                                         break;
1120
1121                                 default:
1122                                         elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1123                                         break;
1124                         }
1125
1126                         if (sample_it)
1127                         {
1128                                 /*
1129                                  * The first targrows sample rows are simply copied into the
1130                                  * reservoir. Then we start replacing tuples in the sample
1131                                  * until we reach the end of the relation.  This algorithm is
1132                                  * from Jeff Vitter's paper (see full citation below). It
1133                                  * works by repeatedly computing the number of tuples to skip
1134                                  * before selecting a tuple, which replaces a randomly chosen
1135                                  * element of the reservoir (current set of tuples).  At all
1136                                  * times the reservoir is a true random sample of the tuples
1137                                  * we've passed over so far, so when we fall off the end of
1138                                  * the relation we're done.
1139                                  */
1140                                 if (numrows < targrows)
1141                                         rows[numrows++] = heap_copytuple(&targtuple);
1142                                 else
1143                                 {
1144                                         /*
1145                                          * t in Vitter's paper is the number of records already
1146                                          * processed.  If we need to compute a new S value, we
1147                                          * must use the not-yet-incremented value of samplerows as
1148                                          * t.
1149                                          */
1150                                         if (rowstoskip < 0)
1151                                                 rowstoskip = reservoir_get_next_S(&rstate, samplerows, targrows);
1152
1153                                         if (rowstoskip <= 0)
1154                                         {
1155                                                 /*
1156                                                  * Found a suitable tuple, so save it, replacing one
1157                                                  * old tuple at random
1158                                                  */
1159                                                 int                     k = (int) (targrows * sampler_random_fract(rstate.randstate));
1160
1161                                                 Assert(k >= 0 && k < targrows);
1162                                                 heap_freetuple(rows[k]);
1163                                                 rows[k] = heap_copytuple(&targtuple);
1164                                         }
1165
1166                                         rowstoskip -= 1;
1167                                 }
1168
1169                                 samplerows += 1;
1170                         }
1171                 }
1172
1173                 /* Now release the lock and pin on the page */
1174                 UnlockReleaseBuffer(targbuffer);
1175         }
1176
1177         /*
1178          * If we didn't find as many tuples as we wanted then we're done. No sort
1179          * is needed, since they're already in order.
1180          *
1181          * Otherwise we need to sort the collected tuples by position
1182          * (itempointer). It's not worth worrying about corner cases where the
1183          * tuples are already sorted.
1184          */
1185         if (numrows == targrows)
1186                 qsort((void *) rows, numrows, sizeof(HeapTuple), compare_rows);
1187
1188         /*
1189          * Estimate total numbers of rows in relation.  For live rows, use
1190          * vac_estimate_reltuples; for dead rows, we have no source of old
1191          * information, so we have to assume the density is the same in unseen
1192          * pages as in the pages we scanned.
1193          */
1194         *totalrows = vac_estimate_reltuples(onerel, true,
1195                                                                                 totalblocks,
1196                                                                                 bs.m,
1197                                                                                 liverows);
1198         if (bs.m > 0)
1199                 *totaldeadrows = floor((deadrows / bs.m) * totalblocks + 0.5);
1200         else
1201                 *totaldeadrows = 0.0;
1202
1203         /*
1204          * Emit some interesting relation info
1205          */
1206         ereport(elevel,
1207                         (errmsg("\"%s\": scanned %d of %u pages, "
1208                                         "containing %.0f live rows and %.0f dead rows; "
1209                                         "%d rows in sample, %.0f estimated total rows",
1210                                         RelationGetRelationName(onerel),
1211                                         bs.m, totalblocks,
1212                                         liverows, deadrows,
1213                                         numrows, *totalrows)));
1214
1215         return numrows;
1216 }
1217
1218 /*
1219  * qsort comparator for sorting rows[] array
1220  */
1221 static int
1222 compare_rows(const void *a, const void *b)
1223 {
1224         HeapTuple       ha = *(const HeapTuple *) a;
1225         HeapTuple       hb = *(const HeapTuple *) b;
1226         BlockNumber ba = ItemPointerGetBlockNumber(&ha->t_self);
1227         OffsetNumber oa = ItemPointerGetOffsetNumber(&ha->t_self);
1228         BlockNumber bb = ItemPointerGetBlockNumber(&hb->t_self);
1229         OffsetNumber ob = ItemPointerGetOffsetNumber(&hb->t_self);
1230
1231         if (ba < bb)
1232                 return -1;
1233         if (ba > bb)
1234                 return 1;
1235         if (oa < ob)
1236                 return -1;
1237         if (oa > ob)
1238                 return 1;
1239         return 0;
1240 }
1241
1242
1243 /*
1244  * acquire_inherited_sample_rows -- acquire sample rows from inheritance tree
1245  *
1246  * This has the same API as acquire_sample_rows, except that rows are
1247  * collected from all inheritance children as well as the specified table.
1248  * We fail and return zero if there are no inheritance children, or if all
1249  * children are foreign tables that don't support ANALYZE.
1250  */
1251 static int
1252 acquire_inherited_sample_rows(Relation onerel, int elevel,
1253                                                           HeapTuple *rows, int targrows,
1254                                                           double *totalrows, double *totaldeadrows)
1255 {
1256         List       *tableOIDs;
1257         Relation   *rels;
1258         AcquireSampleRowsFunc *acquirefuncs;
1259         double     *relblocks;
1260         double          totalblocks;
1261         int                     numrows,
1262                                 nrels,
1263                                 i;
1264         ListCell   *lc;
1265
1266         /*
1267          * Find all members of inheritance set.  We only need AccessShareLock on
1268          * the children.
1269          */
1270         tableOIDs =
1271                 find_all_inheritors(RelationGetRelid(onerel), AccessShareLock, NULL);
1272
1273         /*
1274          * Check that there's at least one descendant, else fail.  This could
1275          * happen despite analyze_rel's relhassubclass check, if table once had a
1276          * child but no longer does.  In that case, we can clear the
1277          * relhassubclass field so as not to make the same mistake again later.
1278          * (This is safe because we hold ShareUpdateExclusiveLock.)
1279          */
1280         if (list_length(tableOIDs) < 2)
1281         {
1282                 /* CCI because we already updated the pg_class row in this command */
1283                 CommandCounterIncrement();
1284                 SetRelationHasSubclass(RelationGetRelid(onerel), false);
1285                 ereport(elevel,
1286                                 (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no child tables",
1287                                                 get_namespace_name(RelationGetNamespace(onerel)),
1288                                                 RelationGetRelationName(onerel))));
1289                 return 0;
1290         }
1291
1292         /*
1293          * Identify acquirefuncs to use, and count blocks in all the relations.
1294          * The result could overflow BlockNumber, so we use double arithmetic.
1295          */
1296         rels = (Relation *) palloc(list_length(tableOIDs) * sizeof(Relation));
1297         acquirefuncs = (AcquireSampleRowsFunc *)
1298                 palloc(list_length(tableOIDs) * sizeof(AcquireSampleRowsFunc));
1299         relblocks = (double *) palloc(list_length(tableOIDs) * sizeof(double));
1300         totalblocks = 0;
1301         nrels = 0;
1302         foreach(lc, tableOIDs)
1303         {
1304                 Oid                     childOID = lfirst_oid(lc);
1305                 Relation        childrel;
1306                 AcquireSampleRowsFunc acquirefunc = NULL;
1307                 BlockNumber relpages = 0;
1308
1309                 /* We already got the needed lock */
1310                 childrel = heap_open(childOID, NoLock);
1311
1312                 /* Ignore if temp table of another backend */
1313                 if (RELATION_IS_OTHER_TEMP(childrel))
1314                 {
1315                         /* ... but release the lock on it */
1316                         Assert(childrel != onerel);
1317                         heap_close(childrel, AccessShareLock);
1318                         continue;
1319                 }
1320
1321                 /* Check table type (MATVIEW can't happen, but might as well allow) */
1322                 if (childrel->rd_rel->relkind == RELKIND_RELATION ||
1323                         childrel->rd_rel->relkind == RELKIND_MATVIEW)
1324                 {
1325                         /* Regular table, so use the regular row acquisition function */
1326                         acquirefunc = acquire_sample_rows;
1327                         relpages = RelationGetNumberOfBlocks(childrel);
1328                 }
1329                 else if (childrel->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
1330                 {
1331                         /*
1332                          * For a foreign table, call the FDW's hook function to see
1333                          * whether it supports analysis.
1334                          */
1335                         FdwRoutine *fdwroutine;
1336                         bool            ok = false;
1337
1338                         fdwroutine = GetFdwRoutineForRelation(childrel, false);
1339
1340                         if (fdwroutine->AnalyzeForeignTable != NULL)
1341                                 ok = fdwroutine->AnalyzeForeignTable(childrel,
1342                                                                                                          &acquirefunc,
1343                                                                                                          &relpages);
1344
1345                         if (!ok)
1346                         {
1347                                 /* ignore, but release the lock on it */
1348                                 Assert(childrel != onerel);
1349                                 heap_close(childrel, AccessShareLock);
1350                                 continue;
1351                         }
1352                 }
1353                 else
1354                 {
1355                         /* ignore, but release the lock on it */
1356                         Assert(childrel != onerel);
1357                         heap_close(childrel, AccessShareLock);
1358                         continue;
1359                 }
1360
1361                 /* OK, we'll process this child */
1362                 rels[nrels] = childrel;
1363                 acquirefuncs[nrels] = acquirefunc;
1364                 relblocks[nrels] = (double) relpages;
1365                 totalblocks += (double) relpages;
1366                 nrels++;
1367         }
1368
1369         /*
1370          * If we don't have at least two tables to consider, fail.
1371          */
1372         if (nrels < 2)
1373         {
1374                 ereport(elevel,
1375                                 (errmsg("skipping analyze of \"%s.%s\" inheritance tree --- this inheritance tree contains no analyzable child tables",
1376                                                 get_namespace_name(RelationGetNamespace(onerel)),
1377                                                 RelationGetRelationName(onerel))));
1378                 return 0;
1379         }
1380
1381         /*
1382          * Now sample rows from each relation, proportionally to its fraction of
1383          * the total block count.  (This might be less than desirable if the child
1384          * rels have radically different free-space percentages, but it's not
1385          * clear that it's worth working harder.)
1386          */
1387         numrows = 0;
1388         *totalrows = 0;
1389         *totaldeadrows = 0;
1390         for (i = 0; i < nrels; i++)
1391         {
1392                 Relation        childrel = rels[i];
1393                 AcquireSampleRowsFunc acquirefunc = acquirefuncs[i];
1394                 double          childblocks = relblocks[i];
1395
1396                 if (childblocks > 0)
1397                 {
1398                         int                     childtargrows;
1399
1400                         childtargrows = (int) rint(targrows * childblocks / totalblocks);
1401                         /* Make sure we don't overrun due to roundoff error */
1402                         childtargrows = Min(childtargrows, targrows - numrows);
1403                         if (childtargrows > 0)
1404                         {
1405                                 int                     childrows;
1406                                 double          trows,
1407                                                         tdrows;
1408
1409                                 /* Fetch a random sample of the child's rows */
1410                                 childrows = (*acquirefunc) (childrel, elevel,
1411                                                                                         rows + numrows, childtargrows,
1412                                                                                         &trows, &tdrows);
1413
1414                                 /* We may need to convert from child's rowtype to parent's */
1415                                 if (childrows > 0 &&
1416                                         !equalTupleDescs(RelationGetDescr(childrel),
1417                                                                          RelationGetDescr(onerel)))
1418                                 {
1419                                         TupleConversionMap *map;
1420
1421                                         map = convert_tuples_by_name(RelationGetDescr(childrel),
1422                                                                                                  RelationGetDescr(onerel),
1423                                                                  gettext_noop("could not convert row type"));
1424                                         if (map != NULL)
1425                                         {
1426                                                 int                     j;
1427
1428                                                 for (j = 0; j < childrows; j++)
1429                                                 {
1430                                                         HeapTuple       newtup;
1431
1432                                                         newtup = do_convert_tuple(rows[numrows + j], map);
1433                                                         heap_freetuple(rows[numrows + j]);
1434                                                         rows[numrows + j] = newtup;
1435                                                 }
1436                                                 free_conversion_map(map);
1437                                         }
1438                                 }
1439
1440                                 /* And add to counts */
1441                                 numrows += childrows;
1442                                 *totalrows += trows;
1443                                 *totaldeadrows += tdrows;
1444                         }
1445                 }
1446
1447                 /*
1448                  * Note: we cannot release the child-table locks, since we may have
1449                  * pointers to their TOAST tables in the sampled rows.
1450                  */
1451                 heap_close(childrel, NoLock);
1452         }
1453
1454         return numrows;
1455 }
1456
1457
1458 /*
1459  *      update_attstats() -- update attribute statistics for one relation
1460  *
1461  *              Statistics are stored in several places: the pg_class row for the
1462  *              relation has stats about the whole relation, and there is a
1463  *              pg_statistic row for each (non-system) attribute that has ever
1464  *              been analyzed.  The pg_class values are updated by VACUUM, not here.
1465  *
1466  *              pg_statistic rows are just added or updated normally.  This means
1467  *              that pg_statistic will probably contain some deleted rows at the
1468  *              completion of a vacuum cycle, unless it happens to get vacuumed last.
1469  *
1470  *              To keep things simple, we punt for pg_statistic, and don't try
1471  *              to compute or store rows for pg_statistic itself in pg_statistic.
1472  *              This could possibly be made to work, but it's not worth the trouble.
1473  *              Note analyze_rel() has seen to it that we won't come here when
1474  *              vacuuming pg_statistic itself.
1475  *
1476  *              Note: there would be a race condition here if two backends could
1477  *              ANALYZE the same table concurrently.  Presently, we lock that out
1478  *              by taking a self-exclusive lock on the relation in analyze_rel().
1479  */
1480 static void
1481 update_attstats(Oid relid, bool inh, int natts, VacAttrStats **vacattrstats)
1482 {
1483         Relation        sd;
1484         int                     attno;
1485
1486         if (natts <= 0)
1487                 return;                                 /* nothing to do */
1488
1489         sd = heap_open(StatisticRelationId, RowExclusiveLock);
1490
1491         for (attno = 0; attno < natts; attno++)
1492         {
1493                 VacAttrStats *stats = vacattrstats[attno];
1494                 HeapTuple       stup,
1495                                         oldtup;
1496                 int                     i,
1497                                         k,
1498                                         n;
1499                 Datum           values[Natts_pg_statistic];
1500                 bool            nulls[Natts_pg_statistic];
1501                 bool            replaces[Natts_pg_statistic];
1502
1503                 /* Ignore attr if we weren't able to collect stats */
1504                 if (!stats->stats_valid)
1505                         continue;
1506
1507                 /*
1508                  * Construct a new pg_statistic tuple
1509                  */
1510                 for (i = 0; i < Natts_pg_statistic; ++i)
1511                 {
1512                         nulls[i] = false;
1513                         replaces[i] = true;
1514                 }
1515
1516                 values[Anum_pg_statistic_starelid - 1] = ObjectIdGetDatum(relid);
1517                 values[Anum_pg_statistic_staattnum - 1] = Int16GetDatum(stats->attr->attnum);
1518                 values[Anum_pg_statistic_stainherit - 1] = BoolGetDatum(inh);
1519                 values[Anum_pg_statistic_stanullfrac - 1] = Float4GetDatum(stats->stanullfrac);
1520                 values[Anum_pg_statistic_stawidth - 1] = Int32GetDatum(stats->stawidth);
1521                 values[Anum_pg_statistic_stadistinct - 1] = Float4GetDatum(stats->stadistinct);
1522                 i = Anum_pg_statistic_stakind1 - 1;
1523                 for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1524                 {
1525                         values[i++] = Int16GetDatum(stats->stakind[k]);         /* stakindN */
1526                 }
1527                 i = Anum_pg_statistic_staop1 - 1;
1528                 for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1529                 {
1530                         values[i++] = ObjectIdGetDatum(stats->staop[k]);        /* staopN */
1531                 }
1532                 i = Anum_pg_statistic_stanumbers1 - 1;
1533                 for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1534                 {
1535                         int                     nnum = stats->numnumbers[k];
1536
1537                         if (nnum > 0)
1538                         {
1539                                 Datum      *numdatums = (Datum *) palloc(nnum * sizeof(Datum));
1540                                 ArrayType  *arry;
1541
1542                                 for (n = 0; n < nnum; n++)
1543                                         numdatums[n] = Float4GetDatum(stats->stanumbers[k][n]);
1544                                 /* XXX knows more than it should about type float4: */
1545                                 arry = construct_array(numdatums, nnum,
1546                                                                            FLOAT4OID,
1547                                                                            sizeof(float4), FLOAT4PASSBYVAL, 'i');
1548                                 values[i++] = PointerGetDatum(arry);    /* stanumbersN */
1549                         }
1550                         else
1551                         {
1552                                 nulls[i] = true;
1553                                 values[i++] = (Datum) 0;
1554                         }
1555                 }
1556                 i = Anum_pg_statistic_stavalues1 - 1;
1557                 for (k = 0; k < STATISTIC_NUM_SLOTS; k++)
1558                 {
1559                         if (stats->numvalues[k] > 0)
1560                         {
1561                                 ArrayType  *arry;
1562
1563                                 arry = construct_array(stats->stavalues[k],
1564                                                                            stats->numvalues[k],
1565                                                                            stats->statypid[k],
1566                                                                            stats->statyplen[k],
1567                                                                            stats->statypbyval[k],
1568                                                                            stats->statypalign[k]);
1569                                 values[i++] = PointerGetDatum(arry);    /* stavaluesN */
1570                         }
1571                         else
1572                         {
1573                                 nulls[i] = true;
1574                                 values[i++] = (Datum) 0;
1575                         }
1576                 }
1577
1578                 /* Is there already a pg_statistic tuple for this attribute? */
1579                 oldtup = SearchSysCache3(STATRELATTINH,
1580                                                                  ObjectIdGetDatum(relid),
1581                                                                  Int16GetDatum(stats->attr->attnum),
1582                                                                  BoolGetDatum(inh));
1583
1584                 if (HeapTupleIsValid(oldtup))
1585                 {
1586                         /* Yes, replace it */
1587                         stup = heap_modify_tuple(oldtup,
1588                                                                          RelationGetDescr(sd),
1589                                                                          values,
1590                                                                          nulls,
1591                                                                          replaces);
1592                         ReleaseSysCache(oldtup);
1593                         simple_heap_update(sd, &stup->t_self, stup);
1594                 }
1595                 else
1596                 {
1597                         /* No, insert new tuple */
1598                         stup = heap_form_tuple(RelationGetDescr(sd), values, nulls);
1599                         simple_heap_insert(sd, stup);
1600                 }
1601
1602                 /* update indexes too */
1603                 CatalogUpdateIndexes(sd, stup);
1604
1605                 heap_freetuple(stup);
1606         }
1607
1608         heap_close(sd, RowExclusiveLock);
1609 }
1610
1611 /*
1612  * Standard fetch function for use by compute_stats subroutines.
1613  *
1614  * This exists to provide some insulation between compute_stats routines
1615  * and the actual storage of the sample data.
1616  */
1617 static Datum
1618 std_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1619 {
1620         int                     attnum = stats->tupattnum;
1621         HeapTuple       tuple = stats->rows[rownum];
1622         TupleDesc       tupDesc = stats->tupDesc;
1623
1624         return heap_getattr(tuple, attnum, tupDesc, isNull);
1625 }
1626
1627 /*
1628  * Fetch function for analyzing index expressions.
1629  *
1630  * We have not bothered to construct index tuples, instead the data is
1631  * just in Datum arrays.
1632  */
1633 static Datum
1634 ind_fetch_func(VacAttrStatsP stats, int rownum, bool *isNull)
1635 {
1636         int                     i;
1637
1638         /* exprvals and exprnulls are already offset for proper column */
1639         i = rownum * stats->rowstride;
1640         *isNull = stats->exprnulls[i];
1641         return stats->exprvals[i];
1642 }
1643
1644
1645 /*==========================================================================
1646  *
1647  * Code below this point represents the "standard" type-specific statistics
1648  * analysis algorithms.  This code can be replaced on a per-data-type basis
1649  * by setting a nonzero value in pg_type.typanalyze.
1650  *
1651  *==========================================================================
1652  */
1653
1654
1655 /*
1656  * To avoid consuming too much memory during analysis and/or too much space
1657  * in the resulting pg_statistic rows, we ignore varlena datums that are wider
1658  * than WIDTH_THRESHOLD (after detoasting!).  This is legitimate for MCV
1659  * and distinct-value calculations since a wide value is unlikely to be
1660  * duplicated at all, much less be a most-common value.  For the same reason,
1661  * ignoring wide values will not affect our estimates of histogram bin
1662  * boundaries very much.
1663  */
1664 #define WIDTH_THRESHOLD  1024
1665
1666 #define swapInt(a,b)    do {int _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1667 #define swapDatum(a,b)  do {Datum _tmp; _tmp=a; a=b; b=_tmp;} while(0)
1668
1669 /*
1670  * Extra information used by the default analysis routines
1671  */
1672 typedef struct
1673 {
1674         Oid                     eqopr;                  /* '=' operator for datatype, if any */
1675         Oid                     eqfunc;                 /* and associated function */
1676         Oid                     ltopr;                  /* '<' operator for datatype, if any */
1677 } StdAnalyzeData;
1678
1679 typedef struct
1680 {
1681         Datum           value;                  /* a data value */
1682         int                     tupno;                  /* position index for tuple it came from */
1683 } ScalarItem;
1684
1685 typedef struct
1686 {
1687         int                     count;                  /* # of duplicates */
1688         int                     first;                  /* values[] index of first occurrence */
1689 } ScalarMCVItem;
1690
1691 typedef struct
1692 {
1693         SortSupport ssup;
1694         int                *tupnoLink;
1695 } CompareScalarsContext;
1696
1697
1698 static void compute_trivial_stats(VacAttrStatsP stats,
1699                                           AnalyzeAttrFetchFunc fetchfunc,
1700                                           int samplerows,
1701                                           double totalrows);
1702 static void compute_distinct_stats(VacAttrStatsP stats,
1703                                            AnalyzeAttrFetchFunc fetchfunc,
1704                                            int samplerows,
1705                                            double totalrows);
1706 static void compute_scalar_stats(VacAttrStatsP stats,
1707                                          AnalyzeAttrFetchFunc fetchfunc,
1708                                          int samplerows,
1709                                          double totalrows);
1710 static int      compare_scalars(const void *a, const void *b, void *arg);
1711 static int      compare_mcvs(const void *a, const void *b);
1712
1713
1714 /*
1715  * std_typanalyze -- the default type-specific typanalyze function
1716  */
1717 bool
1718 std_typanalyze(VacAttrStats *stats)
1719 {
1720         Form_pg_attribute attr = stats->attr;
1721         Oid                     ltopr;
1722         Oid                     eqopr;
1723         StdAnalyzeData *mystats;
1724
1725         /* If the attstattarget column is negative, use the default value */
1726         /* NB: it is okay to scribble on stats->attr since it's a copy */
1727         if (attr->attstattarget < 0)
1728                 attr->attstattarget = default_statistics_target;
1729
1730         /* Look for default "<" and "=" operators for column's type */
1731         get_sort_group_operators(stats->attrtypid,
1732                                                          false, false, false,
1733                                                          &ltopr, &eqopr, NULL,
1734                                                          NULL);
1735
1736         /* Save the operator info for compute_stats routines */
1737         mystats = (StdAnalyzeData *) palloc(sizeof(StdAnalyzeData));
1738         mystats->eqopr = eqopr;
1739         mystats->eqfunc = OidIsValid(eqopr) ? get_opcode(eqopr) : InvalidOid;
1740         mystats->ltopr = ltopr;
1741         stats->extra_data = mystats;
1742
1743         /*
1744          * Determine which standard statistics algorithm to use
1745          */
1746         if (OidIsValid(eqopr) && OidIsValid(ltopr))
1747         {
1748                 /* Seems to be a scalar datatype */
1749                 stats->compute_stats = compute_scalar_stats;
1750                 /*--------------------
1751                  * The following choice of minrows is based on the paper
1752                  * "Random sampling for histogram construction: how much is enough?"
1753                  * by Surajit Chaudhuri, Rajeev Motwani and Vivek Narasayya, in
1754                  * Proceedings of ACM SIGMOD International Conference on Management
1755                  * of Data, 1998, Pages 436-447.  Their Corollary 1 to Theorem 5
1756                  * says that for table size n, histogram size k, maximum relative
1757                  * error in bin size f, and error probability gamma, the minimum
1758                  * random sample size is
1759                  *              r = 4 * k * ln(2*n/gamma) / f^2
1760                  * Taking f = 0.5, gamma = 0.01, n = 10^6 rows, we obtain
1761                  *              r = 305.82 * k
1762                  * Note that because of the log function, the dependence on n is
1763                  * quite weak; even at n = 10^12, a 300*k sample gives <= 0.66
1764                  * bin size error with probability 0.99.  So there's no real need to
1765                  * scale for n, which is a good thing because we don't necessarily
1766                  * know it at this point.
1767                  *--------------------
1768                  */
1769                 stats->minrows = 300 * attr->attstattarget;
1770         }
1771         else if (OidIsValid(eqopr))
1772         {
1773                 /* We can still recognize distinct values */
1774                 stats->compute_stats = compute_distinct_stats;
1775                 /* Might as well use the same minrows as above */
1776                 stats->minrows = 300 * attr->attstattarget;
1777         }
1778         else
1779         {
1780                 /* Can't do much but the trivial stuff */
1781                 stats->compute_stats = compute_trivial_stats;
1782                 /* Might as well use the same minrows as above */
1783                 stats->minrows = 300 * attr->attstattarget;
1784         }
1785
1786         return true;
1787 }
1788
1789
1790 /*
1791  *      compute_trivial_stats() -- compute very basic column statistics
1792  *
1793  *      We use this when we cannot find a hash "=" operator for the datatype.
1794  *
1795  *      We determine the fraction of non-null rows and the average datum width.
1796  */
1797 static void
1798 compute_trivial_stats(VacAttrStatsP stats,
1799                                           AnalyzeAttrFetchFunc fetchfunc,
1800                                           int samplerows,
1801                                           double totalrows)
1802 {
1803         int                     i;
1804         int                     null_cnt = 0;
1805         int                     nonnull_cnt = 0;
1806         double          total_width = 0;
1807         bool            is_varlena = (!stats->attrtype->typbyval &&
1808                                                           stats->attrtype->typlen == -1);
1809         bool            is_varwidth = (!stats->attrtype->typbyval &&
1810                                                            stats->attrtype->typlen < 0);
1811
1812         for (i = 0; i < samplerows; i++)
1813         {
1814                 Datum           value;
1815                 bool            isnull;
1816
1817                 vacuum_delay_point();
1818
1819                 value = fetchfunc(stats, i, &isnull);
1820
1821                 /* Check for null/nonnull */
1822                 if (isnull)
1823                 {
1824                         null_cnt++;
1825                         continue;
1826                 }
1827                 nonnull_cnt++;
1828
1829                 /*
1830                  * If it's a variable-width field, add up widths for average width
1831                  * calculation.  Note that if the value is toasted, we use the toasted
1832                  * width.  We don't bother with this calculation if it's a fixed-width
1833                  * type.
1834                  */
1835                 if (is_varlena)
1836                 {
1837                         total_width += VARSIZE_ANY(DatumGetPointer(value));
1838                 }
1839                 else if (is_varwidth)
1840                 {
1841                         /* must be cstring */
1842                         total_width += strlen(DatumGetCString(value)) + 1;
1843                 }
1844         }
1845
1846         /* We can only compute average width if we found some non-null values. */
1847         if (nonnull_cnt > 0)
1848         {
1849                 stats->stats_valid = true;
1850                 /* Do the simple null-frac and width stats */
1851                 stats->stanullfrac = (double) null_cnt / (double) samplerows;
1852                 if (is_varwidth)
1853                         stats->stawidth = total_width / (double) nonnull_cnt;
1854                 else
1855                         stats->stawidth = stats->attrtype->typlen;
1856                 stats->stadistinct = 0.0;               /* "unknown" */
1857         }
1858         else if (null_cnt > 0)
1859         {
1860                 /* We found only nulls; assume the column is entirely null */
1861                 stats->stats_valid = true;
1862                 stats->stanullfrac = 1.0;
1863                 if (is_varwidth)
1864                         stats->stawidth = 0;    /* "unknown" */
1865                 else
1866                         stats->stawidth = stats->attrtype->typlen;
1867                 stats->stadistinct = 0.0;               /* "unknown" */
1868         }
1869 }
1870
1871
1872 /*
1873  *      compute_distinct_stats() -- compute column statistics including ndistinct
1874  *
1875  *      We use this when we can find only an "=" operator for the datatype.
1876  *
1877  *      We determine the fraction of non-null rows, the average width, the
1878  *      most common values, and the (estimated) number of distinct values.
1879  *
1880  *      The most common values are determined by brute force: we keep a list
1881  *      of previously seen values, ordered by number of times seen, as we scan
1882  *      the samples.  A newly seen value is inserted just after the last
1883  *      multiply-seen value, causing the bottommost (oldest) singly-seen value
1884  *      to drop off the list.  The accuracy of this method, and also its cost,
1885  *      depend mainly on the length of the list we are willing to keep.
1886  */
1887 static void
1888 compute_distinct_stats(VacAttrStatsP stats,
1889                                            AnalyzeAttrFetchFunc fetchfunc,
1890                                            int samplerows,
1891                                            double totalrows)
1892 {
1893         int                     i;
1894         int                     null_cnt = 0;
1895         int                     nonnull_cnt = 0;
1896         int                     toowide_cnt = 0;
1897         double          total_width = 0;
1898         bool            is_varlena = (!stats->attrtype->typbyval &&
1899                                                           stats->attrtype->typlen == -1);
1900         bool            is_varwidth = (!stats->attrtype->typbyval &&
1901                                                            stats->attrtype->typlen < 0);
1902         FmgrInfo        f_cmpeq;
1903         typedef struct
1904         {
1905                 Datum           value;
1906                 int                     count;
1907         } TrackItem;
1908         TrackItem  *track;
1909         int                     track_cnt,
1910                                 track_max;
1911         int                     num_mcv = stats->attr->attstattarget;
1912         StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
1913
1914         /*
1915          * We track up to 2*n values for an n-element MCV list; but at least 10
1916          */
1917         track_max = 2 * num_mcv;
1918         if (track_max < 10)
1919                 track_max = 10;
1920         track = (TrackItem *) palloc(track_max * sizeof(TrackItem));
1921         track_cnt = 0;
1922
1923         fmgr_info(mystats->eqfunc, &f_cmpeq);
1924
1925         for (i = 0; i < samplerows; i++)
1926         {
1927                 Datum           value;
1928                 bool            isnull;
1929                 bool            match;
1930                 int                     firstcount1,
1931                                         j;
1932
1933                 vacuum_delay_point();
1934
1935                 value = fetchfunc(stats, i, &isnull);
1936
1937                 /* Check for null/nonnull */
1938                 if (isnull)
1939                 {
1940                         null_cnt++;
1941                         continue;
1942                 }
1943                 nonnull_cnt++;
1944
1945                 /*
1946                  * If it's a variable-width field, add up widths for average width
1947                  * calculation.  Note that if the value is toasted, we use the toasted
1948                  * width.  We don't bother with this calculation if it's a fixed-width
1949                  * type.
1950                  */
1951                 if (is_varlena)
1952                 {
1953                         total_width += VARSIZE_ANY(DatumGetPointer(value));
1954
1955                         /*
1956                          * If the value is toasted, we want to detoast it just once to
1957                          * avoid repeated detoastings and resultant excess memory usage
1958                          * during the comparisons.  Also, check to see if the value is
1959                          * excessively wide, and if so don't detoast at all --- just
1960                          * ignore the value.
1961                          */
1962                         if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
1963                         {
1964                                 toowide_cnt++;
1965                                 continue;
1966                         }
1967                         value = PointerGetDatum(PG_DETOAST_DATUM(value));
1968                 }
1969                 else if (is_varwidth)
1970                 {
1971                         /* must be cstring */
1972                         total_width += strlen(DatumGetCString(value)) + 1;
1973                 }
1974
1975                 /*
1976                  * See if the value matches anything we're already tracking.
1977                  */
1978                 match = false;
1979                 firstcount1 = track_cnt;
1980                 for (j = 0; j < track_cnt; j++)
1981                 {
1982                         /* We always use the default collation for statistics */
1983                         if (DatumGetBool(FunctionCall2Coll(&f_cmpeq,
1984                                                                                            DEFAULT_COLLATION_OID,
1985                                                                                            value, track[j].value)))
1986                         {
1987                                 match = true;
1988                                 break;
1989                         }
1990                         if (j < firstcount1 && track[j].count == 1)
1991                                 firstcount1 = j;
1992                 }
1993
1994                 if (match)
1995                 {
1996                         /* Found a match */
1997                         track[j].count++;
1998                         /* This value may now need to "bubble up" in the track list */
1999                         while (j > 0 && track[j].count > track[j - 1].count)
2000                         {
2001                                 swapDatum(track[j].value, track[j - 1].value);
2002                                 swapInt(track[j].count, track[j - 1].count);
2003                                 j--;
2004                         }
2005                 }
2006                 else
2007                 {
2008                         /* No match.  Insert at head of count-1 list */
2009                         if (track_cnt < track_max)
2010                                 track_cnt++;
2011                         for (j = track_cnt - 1; j > firstcount1; j--)
2012                         {
2013                                 track[j].value = track[j - 1].value;
2014                                 track[j].count = track[j - 1].count;
2015                         }
2016                         if (firstcount1 < track_cnt)
2017                         {
2018                                 track[firstcount1].value = value;
2019                                 track[firstcount1].count = 1;
2020                         }
2021                 }
2022         }
2023
2024         /* We can only compute real stats if we found some non-null values. */
2025         if (nonnull_cnt > 0)
2026         {
2027                 int                     nmultiple,
2028                                         summultiple;
2029
2030                 stats->stats_valid = true;
2031                 /* Do the simple null-frac and width stats */
2032                 stats->stanullfrac = (double) null_cnt / (double) samplerows;
2033                 if (is_varwidth)
2034                         stats->stawidth = total_width / (double) nonnull_cnt;
2035                 else
2036                         stats->stawidth = stats->attrtype->typlen;
2037
2038                 /* Count the number of values we found multiple times */
2039                 summultiple = 0;
2040                 for (nmultiple = 0; nmultiple < track_cnt; nmultiple++)
2041                 {
2042                         if (track[nmultiple].count == 1)
2043                                 break;
2044                         summultiple += track[nmultiple].count;
2045                 }
2046
2047                 if (nmultiple == 0)
2048                 {
2049                         /* If we found no repeated values, assume it's a unique column */
2050                         stats->stadistinct = -1.0;
2051                 }
2052                 else if (track_cnt < track_max && toowide_cnt == 0 &&
2053                                  nmultiple == track_cnt)
2054                 {
2055                         /*
2056                          * Our track list includes every value in the sample, and every
2057                          * value appeared more than once.  Assume the column has just
2058                          * these values.  (This case is meant to address columns with
2059                          * small, fixed sets of possible values, such as boolean or enum
2060                          * columns.  If there are any values that appear just once in the
2061                          * sample, including too-wide values, we should assume that that's
2062                          * not what we're dealing with.)
2063                          */
2064                         stats->stadistinct = track_cnt;
2065                 }
2066                 else
2067                 {
2068                         /*----------
2069                          * Estimate the number of distinct values using the estimator
2070                          * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2071                          *              n*d / (n - f1 + f1*n/N)
2072                          * where f1 is the number of distinct values that occurred
2073                          * exactly once in our sample of n rows (from a total of N),
2074                          * and d is the total number of distinct values in the sample.
2075                          * This is their Duj1 estimator; the other estimators they
2076                          * recommend are considerably more complex, and are numerically
2077                          * very unstable when n is much smaller than N.
2078                          *
2079                          * In this calculation, we consider only non-nulls.  We used to
2080                          * include rows with null values in the n and N counts, but that
2081                          * leads to inaccurate answers in columns with many nulls, and
2082                          * it's intuitively bogus anyway considering the desired result is
2083                          * the number of distinct non-null values.
2084                          *
2085                          * We assume (not very reliably!) that all the multiply-occurring
2086                          * values are reflected in the final track[] list, and the other
2087                          * nonnull values all appeared but once.  (XXX this usually
2088                          * results in a drastic overestimate of ndistinct.  Can we do
2089                          * any better?)
2090                          *----------
2091                          */
2092                         int                     f1 = nonnull_cnt - summultiple;
2093                         int                     d = f1 + nmultiple;
2094                         double          n = samplerows - null_cnt;
2095                         double          N = totalrows * (1.0 - stats->stanullfrac);
2096                         double          stadistinct;
2097
2098                         /* N == 0 shouldn't happen, but just in case ... */
2099                         if (N > 0)
2100                                 stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2101                         else
2102                                 stadistinct = 0;
2103
2104                         /* Clamp to sane range in case of roundoff error */
2105                         if (stadistinct < d)
2106                                 stadistinct = d;
2107                         if (stadistinct > N)
2108                                 stadistinct = N;
2109                         /* And round to integer */
2110                         stats->stadistinct = floor(stadistinct + 0.5);
2111                 }
2112
2113                 /*
2114                  * If we estimated the number of distinct values at more than 10% of
2115                  * the total row count (a very arbitrary limit), then assume that
2116                  * stadistinct should scale with the row count rather than be a fixed
2117                  * value.
2118                  */
2119                 if (stats->stadistinct > 0.1 * totalrows)
2120                         stats->stadistinct = -(stats->stadistinct / totalrows);
2121
2122                 /*
2123                  * Decide how many values are worth storing as most-common values. If
2124                  * we are able to generate a complete MCV list (all the values in the
2125                  * sample will fit, and we think these are all the ones in the table),
2126                  * then do so.  Otherwise, store only those values that are
2127                  * significantly more common than the (estimated) average. We set the
2128                  * threshold rather arbitrarily at 25% more than average, with at
2129                  * least 2 instances in the sample.
2130                  *
2131                  * Note: the first of these cases is meant to address columns with
2132                  * small, fixed sets of possible values, such as boolean or enum
2133                  * columns.  If we can *completely* represent the column population by
2134                  * an MCV list that will fit into the stats target, then we should do
2135                  * so and thus provide the planner with complete information.  But if
2136                  * the MCV list is not complete, it's generally worth being more
2137                  * selective, and not just filling it all the way up to the stats
2138                  * target.  So for an incomplete list, we try to take only MCVs that
2139                  * are significantly more common than average.
2140                  */
2141                 if (track_cnt < track_max && toowide_cnt == 0 &&
2142                         stats->stadistinct > 0 &&
2143                         track_cnt <= num_mcv)
2144                 {
2145                         /* Track list includes all values seen, and all will fit */
2146                         num_mcv = track_cnt;
2147                 }
2148                 else
2149                 {
2150                         double          ndistinct_table = stats->stadistinct;
2151                         double          avgcount,
2152                                                 mincount;
2153
2154                         /* Re-extract estimate of # distinct nonnull values in table */
2155                         if (ndistinct_table < 0)
2156                                 ndistinct_table = -ndistinct_table * totalrows;
2157                         /* estimate # occurrences in sample of a typical nonnull value */
2158                         avgcount = (double) nonnull_cnt / ndistinct_table;
2159                         /* set minimum threshold count to store a value */
2160                         mincount = avgcount * 1.25;
2161                         if (mincount < 2)
2162                                 mincount = 2;
2163                         if (num_mcv > track_cnt)
2164                                 num_mcv = track_cnt;
2165                         for (i = 0; i < num_mcv; i++)
2166                         {
2167                                 if (track[i].count < mincount)
2168                                 {
2169                                         num_mcv = i;
2170                                         break;
2171                                 }
2172                         }
2173                 }
2174
2175                 /* Generate MCV slot entry */
2176                 if (num_mcv > 0)
2177                 {
2178                         MemoryContext old_context;
2179                         Datum      *mcv_values;
2180                         float4     *mcv_freqs;
2181
2182                         /* Must copy the target values into anl_context */
2183                         old_context = MemoryContextSwitchTo(stats->anl_context);
2184                         mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2185                         mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2186                         for (i = 0; i < num_mcv; i++)
2187                         {
2188                                 mcv_values[i] = datumCopy(track[i].value,
2189                                                                                   stats->attrtype->typbyval,
2190                                                                                   stats->attrtype->typlen);
2191                                 mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2192                         }
2193                         MemoryContextSwitchTo(old_context);
2194
2195                         stats->stakind[0] = STATISTIC_KIND_MCV;
2196                         stats->staop[0] = mystats->eqopr;
2197                         stats->stanumbers[0] = mcv_freqs;
2198                         stats->numnumbers[0] = num_mcv;
2199                         stats->stavalues[0] = mcv_values;
2200                         stats->numvalues[0] = num_mcv;
2201
2202                         /*
2203                          * Accept the defaults for stats->statypid and others. They have
2204                          * been set before we were called (see vacuum.h)
2205                          */
2206                 }
2207         }
2208         else if (null_cnt > 0)
2209         {
2210                 /* We found only nulls; assume the column is entirely null */
2211                 stats->stats_valid = true;
2212                 stats->stanullfrac = 1.0;
2213                 if (is_varwidth)
2214                         stats->stawidth = 0;    /* "unknown" */
2215                 else
2216                         stats->stawidth = stats->attrtype->typlen;
2217                 stats->stadistinct = 0.0;               /* "unknown" */
2218         }
2219
2220         /* We don't need to bother cleaning up any of our temporary palloc's */
2221 }
2222
2223
2224 /*
2225  *      compute_scalar_stats() -- compute column statistics
2226  *
2227  *      We use this when we can find "=" and "<" operators for the datatype.
2228  *
2229  *      We determine the fraction of non-null rows, the average width, the
2230  *      most common values, the (estimated) number of distinct values, the
2231  *      distribution histogram, and the correlation of physical to logical order.
2232  *
2233  *      The desired stats can be determined fairly easily after sorting the
2234  *      data values into order.
2235  */
2236 static void
2237 compute_scalar_stats(VacAttrStatsP stats,
2238                                          AnalyzeAttrFetchFunc fetchfunc,
2239                                          int samplerows,
2240                                          double totalrows)
2241 {
2242         int                     i;
2243         int                     null_cnt = 0;
2244         int                     nonnull_cnt = 0;
2245         int                     toowide_cnt = 0;
2246         double          total_width = 0;
2247         bool            is_varlena = (!stats->attrtype->typbyval &&
2248                                                           stats->attrtype->typlen == -1);
2249         bool            is_varwidth = (!stats->attrtype->typbyval &&
2250                                                            stats->attrtype->typlen < 0);
2251         double          corr_xysum;
2252         SortSupportData ssup;
2253         ScalarItem *values;
2254         int                     values_cnt = 0;
2255         int                *tupnoLink;
2256         ScalarMCVItem *track;
2257         int                     track_cnt = 0;
2258         int                     num_mcv = stats->attr->attstattarget;
2259         int                     num_bins = stats->attr->attstattarget;
2260         StdAnalyzeData *mystats = (StdAnalyzeData *) stats->extra_data;
2261
2262         values = (ScalarItem *) palloc(samplerows * sizeof(ScalarItem));
2263         tupnoLink = (int *) palloc(samplerows * sizeof(int));
2264         track = (ScalarMCVItem *) palloc(num_mcv * sizeof(ScalarMCVItem));
2265
2266         memset(&ssup, 0, sizeof(ssup));
2267         ssup.ssup_cxt = CurrentMemoryContext;
2268         /* We always use the default collation for statistics */
2269         ssup.ssup_collation = DEFAULT_COLLATION_OID;
2270         ssup.ssup_nulls_first = false;
2271
2272         /*
2273          * For now, don't perform abbreviated key conversion, because full values
2274          * are required for MCV slot generation.  Supporting that optimization
2275          * would necessitate teaching compare_scalars() to call a tie-breaker.
2276          */
2277         ssup.abbreviate = false;
2278
2279         PrepareSortSupportFromOrderingOp(mystats->ltopr, &ssup);
2280
2281         /* Initial scan to find sortable values */
2282         for (i = 0; i < samplerows; i++)
2283         {
2284                 Datum           value;
2285                 bool            isnull;
2286
2287                 vacuum_delay_point();
2288
2289                 value = fetchfunc(stats, i, &isnull);
2290
2291                 /* Check for null/nonnull */
2292                 if (isnull)
2293                 {
2294                         null_cnt++;
2295                         continue;
2296                 }
2297                 nonnull_cnt++;
2298
2299                 /*
2300                  * If it's a variable-width field, add up widths for average width
2301                  * calculation.  Note that if the value is toasted, we use the toasted
2302                  * width.  We don't bother with this calculation if it's a fixed-width
2303                  * type.
2304                  */
2305                 if (is_varlena)
2306                 {
2307                         total_width += VARSIZE_ANY(DatumGetPointer(value));
2308
2309                         /*
2310                          * If the value is toasted, we want to detoast it just once to
2311                          * avoid repeated detoastings and resultant excess memory usage
2312                          * during the comparisons.  Also, check to see if the value is
2313                          * excessively wide, and if so don't detoast at all --- just
2314                          * ignore the value.
2315                          */
2316                         if (toast_raw_datum_size(value) > WIDTH_THRESHOLD)
2317                         {
2318                                 toowide_cnt++;
2319                                 continue;
2320                         }
2321                         value = PointerGetDatum(PG_DETOAST_DATUM(value));
2322                 }
2323                 else if (is_varwidth)
2324                 {
2325                         /* must be cstring */
2326                         total_width += strlen(DatumGetCString(value)) + 1;
2327                 }
2328
2329                 /* Add it to the list to be sorted */
2330                 values[values_cnt].value = value;
2331                 values[values_cnt].tupno = values_cnt;
2332                 tupnoLink[values_cnt] = values_cnt;
2333                 values_cnt++;
2334         }
2335
2336         /* We can only compute real stats if we found some sortable values. */
2337         if (values_cnt > 0)
2338         {
2339                 int                     ndistinct,      /* # distinct values in sample */
2340                                         nmultiple,      /* # that appear multiple times */
2341                                         num_hist,
2342                                         dups_cnt;
2343                 int                     slot_idx = 0;
2344                 CompareScalarsContext cxt;
2345
2346                 /* Sort the collected values */
2347                 cxt.ssup = &ssup;
2348                 cxt.tupnoLink = tupnoLink;
2349                 qsort_arg((void *) values, values_cnt, sizeof(ScalarItem),
2350                                   compare_scalars, (void *) &cxt);
2351
2352                 /*
2353                  * Now scan the values in order, find the most common ones, and also
2354                  * accumulate ordering-correlation statistics.
2355                  *
2356                  * To determine which are most common, we first have to count the
2357                  * number of duplicates of each value.  The duplicates are adjacent in
2358                  * the sorted list, so a brute-force approach is to compare successive
2359                  * datum values until we find two that are not equal. However, that
2360                  * requires N-1 invocations of the datum comparison routine, which are
2361                  * completely redundant with work that was done during the sort.  (The
2362                  * sort algorithm must at some point have compared each pair of items
2363                  * that are adjacent in the sorted order; otherwise it could not know
2364                  * that it's ordered the pair correctly.) We exploit this by having
2365                  * compare_scalars remember the highest tupno index that each
2366                  * ScalarItem has been found equal to.  At the end of the sort, a
2367                  * ScalarItem's tupnoLink will still point to itself if and only if it
2368                  * is the last item of its group of duplicates (since the group will
2369                  * be ordered by tupno).
2370                  */
2371                 corr_xysum = 0;
2372                 ndistinct = 0;
2373                 nmultiple = 0;
2374                 dups_cnt = 0;
2375                 for (i = 0; i < values_cnt; i++)
2376                 {
2377                         int                     tupno = values[i].tupno;
2378
2379                         corr_xysum += ((double) i) * ((double) tupno);
2380                         dups_cnt++;
2381                         if (tupnoLink[tupno] == tupno)
2382                         {
2383                                 /* Reached end of duplicates of this value */
2384                                 ndistinct++;
2385                                 if (dups_cnt > 1)
2386                                 {
2387                                         nmultiple++;
2388                                         if (track_cnt < num_mcv ||
2389                                                 dups_cnt > track[track_cnt - 1].count)
2390                                         {
2391                                                 /*
2392                                                  * Found a new item for the mcv list; find its
2393                                                  * position, bubbling down old items if needed. Loop
2394                                                  * invariant is that j points at an empty/ replaceable
2395                                                  * slot.
2396                                                  */
2397                                                 int                     j;
2398
2399                                                 if (track_cnt < num_mcv)
2400                                                         track_cnt++;
2401                                                 for (j = track_cnt - 1; j > 0; j--)
2402                                                 {
2403                                                         if (dups_cnt <= track[j - 1].count)
2404                                                                 break;
2405                                                         track[j].count = track[j - 1].count;
2406                                                         track[j].first = track[j - 1].first;
2407                                                 }
2408                                                 track[j].count = dups_cnt;
2409                                                 track[j].first = i + 1 - dups_cnt;
2410                                         }
2411                                 }
2412                                 dups_cnt = 0;
2413                         }
2414                 }
2415
2416                 stats->stats_valid = true;
2417                 /* Do the simple null-frac and width stats */
2418                 stats->stanullfrac = (double) null_cnt / (double) samplerows;
2419                 if (is_varwidth)
2420                         stats->stawidth = total_width / (double) nonnull_cnt;
2421                 else
2422                         stats->stawidth = stats->attrtype->typlen;
2423
2424                 if (nmultiple == 0)
2425                 {
2426                         /* If we found no repeated values, assume it's a unique column */
2427                         stats->stadistinct = -1.0;
2428                 }
2429                 else if (toowide_cnt == 0 && nmultiple == ndistinct)
2430                 {
2431                         /*
2432                          * Every value in the sample appeared more than once.  Assume the
2433                          * column has just these values.  (This case is meant to address
2434                          * columns with small, fixed sets of possible values, such as
2435                          * boolean or enum columns.  If there are any values that appear
2436                          * just once in the sample, including too-wide values, we should
2437                          * assume that that's not what we're dealing with.)
2438                          */
2439                         stats->stadistinct = ndistinct;
2440                 }
2441                 else
2442                 {
2443                         /*----------
2444                          * Estimate the number of distinct values using the estimator
2445                          * proposed by Haas and Stokes in IBM Research Report RJ 10025:
2446                          *              n*d / (n - f1 + f1*n/N)
2447                          * where f1 is the number of distinct values that occurred
2448                          * exactly once in our sample of n rows (from a total of N),
2449                          * and d is the total number of distinct values in the sample.
2450                          * This is their Duj1 estimator; the other estimators they
2451                          * recommend are considerably more complex, and are numerically
2452                          * very unstable when n is much smaller than N.
2453                          *
2454                          * In this calculation, we consider only non-nulls.  We used to
2455                          * include rows with null values in the n and N counts, but that
2456                          * leads to inaccurate answers in columns with many nulls, and
2457                          * it's intuitively bogus anyway considering the desired result is
2458                          * the number of distinct non-null values.
2459                          *
2460                          * Overwidth values are assumed to have been distinct.
2461                          *----------
2462                          */
2463                         int                     f1 = ndistinct - nmultiple + toowide_cnt;
2464                         int                     d = f1 + nmultiple;
2465                         double          n = samplerows - null_cnt;
2466                         double          N = totalrows * (1.0 - stats->stanullfrac);
2467                         double          stadistinct;
2468
2469                         /* N == 0 shouldn't happen, but just in case ... */
2470                         if (N > 0)
2471                                 stadistinct = (n * d) / ((n - f1) + f1 * n / N);
2472                         else
2473                                 stadistinct = 0;
2474
2475                         /* Clamp to sane range in case of roundoff error */
2476                         if (stadistinct < d)
2477                                 stadistinct = d;
2478                         if (stadistinct > N)
2479                                 stadistinct = N;
2480                         /* And round to integer */
2481                         stats->stadistinct = floor(stadistinct + 0.5);
2482                 }
2483
2484                 /*
2485                  * If we estimated the number of distinct values at more than 10% of
2486                  * the total row count (a very arbitrary limit), then assume that
2487                  * stadistinct should scale with the row count rather than be a fixed
2488                  * value.
2489                  */
2490                 if (stats->stadistinct > 0.1 * totalrows)
2491                         stats->stadistinct = -(stats->stadistinct / totalrows);
2492
2493                 /*
2494                  * Decide how many values are worth storing as most-common values. If
2495                  * we are able to generate a complete MCV list (all the values in the
2496                  * sample will fit, and we think these are all the ones in the table),
2497                  * then do so.  Otherwise, store only those values that are
2498                  * significantly more common than the (estimated) average. We set the
2499                  * threshold rather arbitrarily at 25% more than average, with at
2500                  * least 2 instances in the sample.  Also, we won't suppress values
2501                  * that have a frequency of at least 1/K where K is the intended
2502                  * number of histogram bins; such values might otherwise cause us to
2503                  * emit duplicate histogram bin boundaries.  (We might end up with
2504                  * duplicate histogram entries anyway, if the distribution is skewed;
2505                  * but we prefer to treat such values as MCVs if at all possible.)
2506                  *
2507                  * Note: the first of these cases is meant to address columns with
2508                  * small, fixed sets of possible values, such as boolean or enum
2509                  * columns.  If we can *completely* represent the column population by
2510                  * an MCV list that will fit into the stats target, then we should do
2511                  * so and thus provide the planner with complete information.  But if
2512                  * the MCV list is not complete, it's generally worth being more
2513                  * selective, and not just filling it all the way up to the stats
2514                  * target.  So for an incomplete list, we try to take only MCVs that
2515                  * are significantly more common than average.
2516                  */
2517                 if (track_cnt == ndistinct && toowide_cnt == 0 &&
2518                         stats->stadistinct > 0 &&
2519                         track_cnt <= num_mcv)
2520                 {
2521                         /* Track list includes all values seen, and all will fit */
2522                         num_mcv = track_cnt;
2523                 }
2524                 else
2525                 {
2526                         double          ndistinct_table = stats->stadistinct;
2527                         double          avgcount,
2528                                                 mincount,
2529                                                 maxmincount;
2530
2531                         /* Re-extract estimate of # distinct nonnull values in table */
2532                         if (ndistinct_table < 0)
2533                                 ndistinct_table = -ndistinct_table * totalrows;
2534                         /* estimate # occurrences in sample of a typical nonnull value */
2535                         avgcount = (double) nonnull_cnt / ndistinct_table;
2536                         /* set minimum threshold count to store a value */
2537                         mincount = avgcount * 1.25;
2538                         if (mincount < 2)
2539                                 mincount = 2;
2540                         /* don't let threshold exceed 1/K, however */
2541                         maxmincount = (double) values_cnt / (double) num_bins;
2542                         if (mincount > maxmincount)
2543                                 mincount = maxmincount;
2544                         if (num_mcv > track_cnt)
2545                                 num_mcv = track_cnt;
2546                         for (i = 0; i < num_mcv; i++)
2547                         {
2548                                 if (track[i].count < mincount)
2549                                 {
2550                                         num_mcv = i;
2551                                         break;
2552                                 }
2553                         }
2554                 }
2555
2556                 /* Generate MCV slot entry */
2557                 if (num_mcv > 0)
2558                 {
2559                         MemoryContext old_context;
2560                         Datum      *mcv_values;
2561                         float4     *mcv_freqs;
2562
2563                         /* Must copy the target values into anl_context */
2564                         old_context = MemoryContextSwitchTo(stats->anl_context);
2565                         mcv_values = (Datum *) palloc(num_mcv * sizeof(Datum));
2566                         mcv_freqs = (float4 *) palloc(num_mcv * sizeof(float4));
2567                         for (i = 0; i < num_mcv; i++)
2568                         {
2569                                 mcv_values[i] = datumCopy(values[track[i].first].value,
2570                                                                                   stats->attrtype->typbyval,
2571                                                                                   stats->attrtype->typlen);
2572                                 mcv_freqs[i] = (double) track[i].count / (double) samplerows;
2573                         }
2574                         MemoryContextSwitchTo(old_context);
2575
2576                         stats->stakind[slot_idx] = STATISTIC_KIND_MCV;
2577                         stats->staop[slot_idx] = mystats->eqopr;
2578                         stats->stanumbers[slot_idx] = mcv_freqs;
2579                         stats->numnumbers[slot_idx] = num_mcv;
2580                         stats->stavalues[slot_idx] = mcv_values;
2581                         stats->numvalues[slot_idx] = num_mcv;
2582
2583                         /*
2584                          * Accept the defaults for stats->statypid and others. They have
2585                          * been set before we were called (see vacuum.h)
2586                          */
2587                         slot_idx++;
2588                 }
2589
2590                 /*
2591                  * Generate a histogram slot entry if there are at least two distinct
2592                  * values not accounted for in the MCV list.  (This ensures the
2593                  * histogram won't collapse to empty or a singleton.)
2594                  */
2595                 num_hist = ndistinct - num_mcv;
2596                 if (num_hist > num_bins)
2597                         num_hist = num_bins + 1;
2598                 if (num_hist >= 2)
2599                 {
2600                         MemoryContext old_context;
2601                         Datum      *hist_values;
2602                         int                     nvals;
2603                         int                     pos,
2604                                                 posfrac,
2605                                                 delta,
2606                                                 deltafrac;
2607
2608                         /* Sort the MCV items into position order to speed next loop */
2609                         qsort((void *) track, num_mcv,
2610                                   sizeof(ScalarMCVItem), compare_mcvs);
2611
2612                         /*
2613                          * Collapse out the MCV items from the values[] array.
2614                          *
2615                          * Note we destroy the values[] array here... but we don't need it
2616                          * for anything more.  We do, however, still need values_cnt.
2617                          * nvals will be the number of remaining entries in values[].
2618                          */
2619                         if (num_mcv > 0)
2620                         {
2621                                 int                     src,
2622                                                         dest;
2623                                 int                     j;
2624
2625                                 src = dest = 0;
2626                                 j = 0;                  /* index of next interesting MCV item */
2627                                 while (src < values_cnt)
2628                                 {
2629                                         int                     ncopy;
2630
2631                                         if (j < num_mcv)
2632                                         {
2633                                                 int                     first = track[j].first;
2634
2635                                                 if (src >= first)
2636                                                 {
2637                                                         /* advance past this MCV item */
2638                                                         src = first + track[j].count;
2639                                                         j++;
2640                                                         continue;
2641                                                 }
2642                                                 ncopy = first - src;
2643                                         }
2644                                         else
2645                                                 ncopy = values_cnt - src;
2646                                         memmove(&values[dest], &values[src],
2647                                                         ncopy * sizeof(ScalarItem));
2648                                         src += ncopy;
2649                                         dest += ncopy;
2650                                 }
2651                                 nvals = dest;
2652                         }
2653                         else
2654                                 nvals = values_cnt;
2655                         Assert(nvals >= num_hist);
2656
2657                         /* Must copy the target values into anl_context */
2658                         old_context = MemoryContextSwitchTo(stats->anl_context);
2659                         hist_values = (Datum *) palloc(num_hist * sizeof(Datum));
2660
2661                         /*
2662                          * The object of this loop is to copy the first and last values[]
2663                          * entries along with evenly-spaced values in between.  So the
2664                          * i'th value is values[(i * (nvals - 1)) / (num_hist - 1)].  But
2665                          * computing that subscript directly risks integer overflow when
2666                          * the stats target is more than a couple thousand.  Instead we
2667                          * add (nvals - 1) / (num_hist - 1) to pos at each step, tracking
2668                          * the integral and fractional parts of the sum separately.
2669                          */
2670                         delta = (nvals - 1) / (num_hist - 1);
2671                         deltafrac = (nvals - 1) % (num_hist - 1);
2672                         pos = posfrac = 0;
2673
2674                         for (i = 0; i < num_hist; i++)
2675                         {
2676                                 hist_values[i] = datumCopy(values[pos].value,
2677                                                                                    stats->attrtype->typbyval,
2678                                                                                    stats->attrtype->typlen);
2679                                 pos += delta;
2680                                 posfrac += deltafrac;
2681                                 if (posfrac >= (num_hist - 1))
2682                                 {
2683                                         /* fractional part exceeds 1, carry to integer part */
2684                                         pos++;
2685                                         posfrac -= (num_hist - 1);
2686                                 }
2687                         }
2688
2689                         MemoryContextSwitchTo(old_context);
2690
2691                         stats->stakind[slot_idx] = STATISTIC_KIND_HISTOGRAM;
2692                         stats->staop[slot_idx] = mystats->ltopr;
2693                         stats->stavalues[slot_idx] = hist_values;
2694                         stats->numvalues[slot_idx] = num_hist;
2695
2696                         /*
2697                          * Accept the defaults for stats->statypid and others. They have
2698                          * been set before we were called (see vacuum.h)
2699                          */
2700                         slot_idx++;
2701                 }
2702
2703                 /* Generate a correlation entry if there are multiple values */
2704                 if (values_cnt > 1)
2705                 {
2706                         MemoryContext old_context;
2707                         float4     *corrs;
2708                         double          corr_xsum,
2709                                                 corr_x2sum;
2710
2711                         /* Must copy the target values into anl_context */
2712                         old_context = MemoryContextSwitchTo(stats->anl_context);
2713                         corrs = (float4 *) palloc(sizeof(float4));
2714                         MemoryContextSwitchTo(old_context);
2715
2716                         /*----------
2717                          * Since we know the x and y value sets are both
2718                          *              0, 1, ..., values_cnt-1
2719                          * we have sum(x) = sum(y) =
2720                          *              (values_cnt-1)*values_cnt / 2
2721                          * and sum(x^2) = sum(y^2) =
2722                          *              (values_cnt-1)*values_cnt*(2*values_cnt-1) / 6.
2723                          *----------
2724                          */
2725                         corr_xsum = ((double) (values_cnt - 1)) *
2726                                 ((double) values_cnt) / 2.0;
2727                         corr_x2sum = ((double) (values_cnt - 1)) *
2728                                 ((double) values_cnt) * (double) (2 * values_cnt - 1) / 6.0;
2729
2730                         /* And the correlation coefficient reduces to */
2731                         corrs[0] = (values_cnt * corr_xysum - corr_xsum * corr_xsum) /
2732                                 (values_cnt * corr_x2sum - corr_xsum * corr_xsum);
2733
2734                         stats->stakind[slot_idx] = STATISTIC_KIND_CORRELATION;
2735                         stats->staop[slot_idx] = mystats->ltopr;
2736                         stats->stanumbers[slot_idx] = corrs;
2737                         stats->numnumbers[slot_idx] = 1;
2738                         slot_idx++;
2739                 }
2740         }
2741         else if (nonnull_cnt > 0)
2742         {
2743                 /* We found some non-null values, but they were all too wide */
2744                 Assert(nonnull_cnt == toowide_cnt);
2745                 stats->stats_valid = true;
2746                 /* Do the simple null-frac and width stats */
2747                 stats->stanullfrac = (double) null_cnt / (double) samplerows;
2748                 if (is_varwidth)
2749                         stats->stawidth = total_width / (double) nonnull_cnt;
2750                 else
2751                         stats->stawidth = stats->attrtype->typlen;
2752                 /* Assume all too-wide values are distinct, so it's a unique column */
2753                 stats->stadistinct = -1.0;
2754         }
2755         else if (null_cnt > 0)
2756         {
2757                 /* We found only nulls; assume the column is entirely null */
2758                 stats->stats_valid = true;
2759                 stats->stanullfrac = 1.0;
2760                 if (is_varwidth)
2761                         stats->stawidth = 0;    /* "unknown" */
2762                 else
2763                         stats->stawidth = stats->attrtype->typlen;
2764                 stats->stadistinct = 0.0;               /* "unknown" */
2765         }
2766
2767         /* We don't need to bother cleaning up any of our temporary palloc's */
2768 }
2769
2770 /*
2771  * qsort_arg comparator for sorting ScalarItems
2772  *
2773  * Aside from sorting the items, we update the tupnoLink[] array
2774  * whenever two ScalarItems are found to contain equal datums.  The array
2775  * is indexed by tupno; for each ScalarItem, it contains the highest
2776  * tupno that that item's datum has been found to be equal to.  This allows
2777  * us to avoid additional comparisons in compute_scalar_stats().
2778  */
2779 static int
2780 compare_scalars(const void *a, const void *b, void *arg)
2781 {
2782         Datum           da = ((const ScalarItem *) a)->value;
2783         int                     ta = ((const ScalarItem *) a)->tupno;
2784         Datum           db = ((const ScalarItem *) b)->value;
2785         int                     tb = ((const ScalarItem *) b)->tupno;
2786         CompareScalarsContext *cxt = (CompareScalarsContext *) arg;
2787         int                     compare;
2788
2789         compare = ApplySortComparator(da, false, db, false, cxt->ssup);
2790         if (compare != 0)
2791                 return compare;
2792
2793         /*
2794          * The two datums are equal, so update cxt->tupnoLink[].
2795          */
2796         if (cxt->tupnoLink[ta] < tb)
2797                 cxt->tupnoLink[ta] = tb;
2798         if (cxt->tupnoLink[tb] < ta)
2799                 cxt->tupnoLink[tb] = ta;
2800
2801         /*
2802          * For equal datums, sort by tupno
2803          */
2804         return ta - tb;
2805 }
2806
2807 /*
2808  * qsort comparator for sorting ScalarMCVItems by position
2809  */
2810 static int
2811 compare_mcvs(const void *a, const void *b)
2812 {
2813         int                     da = ((const ScalarMCVItem *) a)->first;
2814         int                     db = ((const ScalarMCVItem *) b)->first;
2815
2816         return da - db;
2817 }