Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * nbtutils.c
4 : * Utility code for Postgres btree implementation.
5 : *
6 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/nbtree/nbtutils.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 :
16 : #include "postgres.h"
17 :
18 : #include <time.h>
19 :
20 : #include "access/nbtree.h"
21 : #include "access/reloptions.h"
22 : #include "access/relscan.h"
23 : #include "commands/progress.h"
24 : #include "lib/qunique.h"
25 : #include "miscadmin.h"
26 : #include "utils/array.h"
27 : #include "utils/datum.h"
28 : #include "utils/lsyscache.h"
29 : #include "utils/memutils.h"
30 : #include "utils/rel.h"
31 :
32 : #define LOOK_AHEAD_REQUIRED_RECHECKS 3
33 : #define LOOK_AHEAD_DEFAULT_DISTANCE 5
34 :
35 : typedef struct BTSortArrayContext
36 : {
37 : FmgrInfo *sortproc;
38 : Oid collation;
39 : bool reverse;
40 : } BTSortArrayContext;
41 :
42 : typedef struct BTScanKeyPreproc
43 : {
44 : ScanKey skey;
45 : int ikey;
46 : int arrayidx;
47 : } BTScanKeyPreproc;
48 :
49 : static void _bt_setup_array_cmp(IndexScanDesc scan, ScanKey skey, Oid elemtype,
50 : FmgrInfo *orderproc, FmgrInfo **sortprocp);
51 : static Datum _bt_find_extreme_element(IndexScanDesc scan, ScanKey skey,
52 : Oid elemtype, StrategyNumber strat,
53 : Datum *elems, int nelems);
54 : static int _bt_sort_array_elements(ScanKey skey, FmgrInfo *sortproc,
55 : bool reverse, Datum *elems, int nelems);
56 : static bool _bt_merge_arrays(IndexScanDesc scan, ScanKey skey,
57 : FmgrInfo *sortproc, bool reverse,
58 : Oid origelemtype, Oid nextelemtype,
59 : Datum *elems_orig, int *nelems_orig,
60 : Datum *elems_next, int nelems_next);
61 : static bool _bt_compare_array_scankey_args(IndexScanDesc scan,
62 : ScanKey arraysk, ScanKey skey,
63 : FmgrInfo *orderproc, BTArrayKeyInfo *array,
64 : bool *qual_ok);
65 : static ScanKey _bt_preprocess_array_keys(IndexScanDesc scan);
66 : static void _bt_preprocess_array_keys_final(IndexScanDesc scan, int *keyDataMap);
67 : static int _bt_compare_array_elements(const void *a, const void *b, void *arg);
68 : static inline int32 _bt_compare_array_skey(FmgrInfo *orderproc,
69 : Datum tupdatum, bool tupnull,
70 : Datum arrdatum, ScanKey cur);
71 : static int _bt_binsrch_array_skey(FmgrInfo *orderproc,
72 : bool cur_elem_trig, ScanDirection dir,
73 : Datum tupdatum, bool tupnull,
74 : BTArrayKeyInfo *array, ScanKey cur,
75 : int32 *set_elem_result);
76 : static bool _bt_advance_array_keys_increment(IndexScanDesc scan, ScanDirection dir);
77 : static void _bt_rewind_nonrequired_arrays(IndexScanDesc scan, ScanDirection dir);
78 : static bool _bt_tuple_before_array_skeys(IndexScanDesc scan, ScanDirection dir,
79 : IndexTuple tuple, TupleDesc tupdesc, int tupnatts,
80 : bool readpagetup, int sktrig, bool *scanBehind);
81 : static bool _bt_advance_array_keys(IndexScanDesc scan, BTReadPageState *pstate,
82 : IndexTuple tuple, int tupnatts, TupleDesc tupdesc,
83 : int sktrig, bool sktrig_required);
84 : #ifdef USE_ASSERT_CHECKING
85 : static bool _bt_verify_arrays_bt_first(IndexScanDesc scan, ScanDirection dir);
86 : static bool _bt_verify_keys_with_arraykeys(IndexScanDesc scan);
87 : #endif
88 : static bool _bt_compare_scankey_args(IndexScanDesc scan, ScanKey op,
89 : ScanKey leftarg, ScanKey rightarg,
90 : BTArrayKeyInfo *array, FmgrInfo *orderproc,
91 : bool *result);
92 : static bool _bt_fix_scankey_strategy(ScanKey skey, int16 *indoption);
93 : static void _bt_mark_scankey_required(ScanKey skey);
94 : static bool _bt_check_compare(IndexScanDesc scan, ScanDirection dir,
95 : IndexTuple tuple, int tupnatts, TupleDesc tupdesc,
96 : bool advancenonrequired, bool prechecked, bool firstmatch,
97 : bool *continuescan, int *ikey);
98 : static bool _bt_check_rowcompare(ScanKey skey,
99 : IndexTuple tuple, int tupnatts, TupleDesc tupdesc,
100 : ScanDirection dir, bool *continuescan);
101 : static void _bt_checkkeys_look_ahead(IndexScanDesc scan, BTReadPageState *pstate,
102 : int tupnatts, TupleDesc tupdesc);
103 : static int _bt_keep_natts(Relation rel, IndexTuple lastleft,
104 : IndexTuple firstright, BTScanInsert itup_key);
105 :
106 :
107 : /*
108 : * _bt_mkscankey
109 : * Build an insertion scan key that contains comparison data from itup
110 : * as well as comparator routines appropriate to the key datatypes.
111 : *
112 : * The result is intended for use with _bt_compare() and _bt_truncate().
113 : * Callers that don't need to fill out the insertion scankey arguments
114 : * (e.g. they use an ad-hoc comparison routine, or only need a scankey
115 : * for _bt_truncate()) can pass a NULL index tuple. The scankey will
116 : * be initialized as if an "all truncated" pivot tuple was passed
117 : * instead.
118 : *
119 : * Note that we may occasionally have to share lock the metapage to
120 : * determine whether or not the keys in the index are expected to be
121 : * unique (i.e. if this is a "heapkeyspace" index). We assume a
122 : * heapkeyspace index when caller passes a NULL tuple, allowing index
123 : * build callers to avoid accessing the non-existent metapage. We
124 : * also assume that the index is _not_ allequalimage when a NULL tuple
125 : * is passed; CREATE INDEX callers call _bt_allequalimage() to set the
126 : * field themselves.
127 : */
128 : BTScanInsert
129 10804120 : _bt_mkscankey(Relation rel, IndexTuple itup)
130 : {
131 : BTScanInsert key;
132 : ScanKey skey;
133 : TupleDesc itupdesc;
134 : int indnkeyatts;
135 : int16 *indoption;
136 : int tupnatts;
137 : int i;
138 :
139 10804120 : itupdesc = RelationGetDescr(rel);
140 10804120 : indnkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
141 10804120 : indoption = rel->rd_indoption;
142 10804120 : tupnatts = itup ? BTreeTupleGetNAtts(itup, rel) : 0;
143 :
144 : Assert(tupnatts <= IndexRelationGetNumberOfAttributes(rel));
145 :
146 : /*
147 : * We'll execute search using scan key constructed on key columns.
148 : * Truncated attributes and non-key attributes are omitted from the final
149 : * scan key.
150 : */
151 10804120 : key = palloc(offsetof(BTScanInsertData, scankeys) +
152 10804120 : sizeof(ScanKeyData) * indnkeyatts);
153 10804120 : if (itup)
154 10675280 : _bt_metaversion(rel, &key->heapkeyspace, &key->allequalimage);
155 : else
156 : {
157 : /* Utility statement callers can set these fields themselves */
158 128840 : key->heapkeyspace = true;
159 128840 : key->allequalimage = false;
160 : }
161 10804120 : key->anynullkeys = false; /* initial assumption */
162 10804120 : key->nextkey = false; /* usual case, required by btinsert */
163 10804120 : key->backward = false; /* usual case, required by btinsert */
164 10804120 : key->keysz = Min(indnkeyatts, tupnatts);
165 10804120 : key->scantid = key->heapkeyspace && itup ?
166 21608240 : BTreeTupleGetHeapTID(itup) : NULL;
167 10804120 : skey = key->scankeys;
168 29022678 : for (i = 0; i < indnkeyatts; i++)
169 : {
170 : FmgrInfo *procinfo;
171 : Datum arg;
172 : bool null;
173 : int flags;
174 :
175 : /*
176 : * We can use the cached (default) support procs since no cross-type
177 : * comparison can be needed.
178 : */
179 18218558 : procinfo = index_getprocinfo(rel, i + 1, BTORDER_PROC);
180 :
181 : /*
182 : * Key arguments built from truncated attributes (or when caller
183 : * provides no tuple) are defensively represented as NULL values. They
184 : * should never be used.
185 : */
186 18218558 : if (i < tupnatts)
187 17988246 : arg = index_getattr(itup, i + 1, itupdesc, &null);
188 : else
189 : {
190 230312 : arg = (Datum) 0;
191 230312 : null = true;
192 : }
193 18218558 : flags = (null ? SK_ISNULL : 0) | (indoption[i] << SK_BT_INDOPTION_SHIFT);
194 18218558 : ScanKeyEntryInitializeWithInfo(&skey[i],
195 : flags,
196 18218558 : (AttrNumber) (i + 1),
197 : InvalidStrategy,
198 : InvalidOid,
199 18218558 : rel->rd_indcollation[i],
200 : procinfo,
201 : arg);
202 : /* Record if any key attribute is NULL (or truncated) */
203 18218558 : if (null)
204 250888 : key->anynullkeys = true;
205 : }
206 :
207 : /*
208 : * In NULLS NOT DISTINCT mode, we pretend that there are no null keys, so
209 : * that full uniqueness check is done.
210 : */
211 10804120 : if (rel->rd_index->indnullsnotdistinct)
212 168 : key->anynullkeys = false;
213 :
214 10804120 : return key;
215 : }
216 :
217 : /*
218 : * free a retracement stack made by _bt_search.
219 : */
220 : void
221 17625748 : _bt_freestack(BTStack stack)
222 : {
223 : BTStack ostack;
224 :
225 32353488 : while (stack != NULL)
226 : {
227 14727740 : ostack = stack;
228 14727740 : stack = stack->bts_parent;
229 14727740 : pfree(ostack);
230 : }
231 17625748 : }
232 :
233 :
234 : /*
235 : * _bt_preprocess_array_keys() -- Preprocess SK_SEARCHARRAY scan keys
236 : *
237 : * If there are any SK_SEARCHARRAY scan keys, deconstruct the array(s) and
238 : * set up BTArrayKeyInfo info for each one that is an equality-type key.
239 : * Returns modified scan keys as input for further, standard preprocessing.
240 : *
241 : * Currently we perform two kinds of preprocessing to deal with redundancies.
242 : * For inequality array keys, it's sufficient to find the extreme element
243 : * value and replace the whole array with that scalar value. This eliminates
244 : * all but one array element as redundant. Similarly, we are capable of
245 : * "merging together" multiple equality array keys (from two or more input
246 : * scan keys) into a single output scan key containing only the intersecting
247 : * array elements. This can eliminate many redundant array elements, as well
248 : * as eliminating whole array scan keys as redundant. It can also allow us to
249 : * detect contradictory quals.
250 : *
251 : * It is convenient for _bt_preprocess_keys caller to have to deal with no
252 : * more than one equality strategy array scan key per index attribute. We'll
253 : * always be able to set things up that way when complete opfamilies are used.
254 : * Eliminated array scan keys can be recognized as those that have had their
255 : * sk_strategy field set to InvalidStrategy here by us. Caller should avoid
256 : * including these in the scan's so->keyData[] output array.
257 : *
258 : * We set the scan key references from the scan's BTArrayKeyInfo info array to
259 : * offsets into the temp modified input array returned to caller. Scans that
260 : * have array keys should call _bt_preprocess_array_keys_final when standard
261 : * preprocessing steps are complete. This will convert the scan key offset
262 : * references into references to the scan's so->keyData[] output scan keys.
263 : *
264 : * Note: the reason we need to return a temp scan key array, rather than just
265 : * scribbling on scan->keyData, is that callers are permitted to call btrescan
266 : * without supplying a new set of scankey data.
267 : */
268 : static ScanKey
269 11819214 : _bt_preprocess_array_keys(IndexScanDesc scan)
270 : {
271 11819214 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
272 11819214 : Relation rel = scan->indexRelation;
273 11819214 : int numberOfKeys = scan->numberOfKeys;
274 11819214 : int16 *indoption = rel->rd_indoption;
275 : int numArrayKeys;
276 11819214 : int origarrayatt = InvalidAttrNumber,
277 11819214 : origarraykey = -1;
278 11819214 : Oid origelemtype = InvalidOid;
279 : ScanKey cur;
280 : MemoryContext oldContext;
281 : ScanKey arrayKeyData; /* modified copy of scan->keyData */
282 :
283 : Assert(numberOfKeys);
284 :
285 : /* Quick check to see if there are any array keys */
286 11819214 : numArrayKeys = 0;
287 31097502 : for (int i = 0; i < numberOfKeys; i++)
288 : {
289 19278288 : cur = &scan->keyData[i];
290 19278288 : if (cur->sk_flags & SK_SEARCHARRAY)
291 : {
292 982 : numArrayKeys++;
293 : Assert(!(cur->sk_flags & (SK_ROW_HEADER | SK_SEARCHNULL | SK_SEARCHNOTNULL)));
294 : /* If any arrays are null as a whole, we can quit right now. */
295 982 : if (cur->sk_flags & SK_ISNULL)
296 : {
297 0 : so->qual_ok = false;
298 0 : return NULL;
299 : }
300 : }
301 : }
302 :
303 : /* Quit if nothing to do. */
304 11819214 : if (numArrayKeys == 0)
305 11818382 : return NULL;
306 :
307 : /*
308 : * Make a scan-lifespan context to hold array-associated data, or reset it
309 : * if we already have one from a previous rescan cycle.
310 : */
311 832 : if (so->arrayContext == NULL)
312 832 : so->arrayContext = AllocSetContextCreate(CurrentMemoryContext,
313 : "BTree array context",
314 : ALLOCSET_SMALL_SIZES);
315 : else
316 0 : MemoryContextReset(so->arrayContext);
317 :
318 832 : oldContext = MemoryContextSwitchTo(so->arrayContext);
319 :
320 : /* Create modifiable copy of scan->keyData in the workspace context */
321 832 : arrayKeyData = (ScanKey) palloc(numberOfKeys * sizeof(ScanKeyData));
322 832 : memcpy(arrayKeyData, scan->keyData, numberOfKeys * sizeof(ScanKeyData));
323 :
324 : /* Allocate space for per-array data in the workspace context */
325 832 : so->arrayKeys = (BTArrayKeyInfo *) palloc(numArrayKeys * sizeof(BTArrayKeyInfo));
326 :
327 : /* Allocate space for ORDER procs used to help _bt_checkkeys */
328 832 : so->orderProcs = (FmgrInfo *) palloc(numberOfKeys * sizeof(FmgrInfo));
329 :
330 : /* Now process each array key */
331 832 : numArrayKeys = 0;
332 2102 : for (int i = 0; i < numberOfKeys; i++)
333 : {
334 : FmgrInfo sortproc;
335 1276 : FmgrInfo *sortprocp = &sortproc;
336 : Oid elemtype;
337 : bool reverse;
338 : ArrayType *arrayval;
339 : int16 elmlen;
340 : bool elmbyval;
341 : char elmalign;
342 : int num_elems;
343 : Datum *elem_values;
344 : bool *elem_nulls;
345 : int num_nonnulls;
346 : int j;
347 :
348 1276 : cur = &arrayKeyData[i];
349 1276 : if (!(cur->sk_flags & SK_SEARCHARRAY))
350 306 : continue;
351 :
352 : /*
353 : * First, deconstruct the array into elements. Anything allocated
354 : * here (including a possibly detoasted array value) is in the
355 : * workspace context.
356 : */
357 982 : arrayval = DatumGetArrayTypeP(cur->sk_argument);
358 : /* We could cache this data, but not clear it's worth it */
359 982 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
360 : &elmlen, &elmbyval, &elmalign);
361 982 : deconstruct_array(arrayval,
362 : ARR_ELEMTYPE(arrayval),
363 : elmlen, elmbyval, elmalign,
364 : &elem_values, &elem_nulls, &num_elems);
365 :
366 : /*
367 : * Compress out any null elements. We can ignore them since we assume
368 : * all btree operators are strict.
369 : */
370 982 : num_nonnulls = 0;
371 5234 : for (j = 0; j < num_elems; j++)
372 : {
373 4252 : if (!elem_nulls[j])
374 4252 : elem_values[num_nonnulls++] = elem_values[j];
375 : }
376 :
377 : /* We could pfree(elem_nulls) now, but not worth the cycles */
378 :
379 : /* If there's no non-nulls, the scan qual is unsatisfiable */
380 982 : if (num_nonnulls == 0)
381 : {
382 0 : so->qual_ok = false;
383 6 : break;
384 : }
385 :
386 : /*
387 : * Determine the nominal datatype of the array elements. We have to
388 : * support the convention that sk_subtype == InvalidOid means the
389 : * opclass input type; this is a hack to simplify life for
390 : * ScanKeyInit().
391 : */
392 982 : elemtype = cur->sk_subtype;
393 982 : if (elemtype == InvalidOid)
394 0 : elemtype = rel->rd_opcintype[cur->sk_attno - 1];
395 :
396 : /*
397 : * If the comparison operator is not equality, then the array qual
398 : * degenerates to a simple comparison against the smallest or largest
399 : * non-null array element, as appropriate.
400 : */
401 982 : switch (cur->sk_strategy)
402 : {
403 0 : case BTLessStrategyNumber:
404 : case BTLessEqualStrategyNumber:
405 0 : cur->sk_argument =
406 0 : _bt_find_extreme_element(scan, cur, elemtype,
407 : BTGreaterStrategyNumber,
408 : elem_values, num_nonnulls);
409 0 : continue;
410 976 : case BTEqualStrategyNumber:
411 : /* proceed with rest of loop */
412 976 : break;
413 6 : case BTGreaterEqualStrategyNumber:
414 : case BTGreaterStrategyNumber:
415 6 : cur->sk_argument =
416 6 : _bt_find_extreme_element(scan, cur, elemtype,
417 : BTLessStrategyNumber,
418 : elem_values, num_nonnulls);
419 6 : continue;
420 0 : default:
421 0 : elog(ERROR, "unrecognized StrategyNumber: %d",
422 : (int) cur->sk_strategy);
423 : break;
424 : }
425 :
426 : /*
427 : * We'll need a 3-way ORDER proc to perform binary searches for the
428 : * next matching array element. Set that up now.
429 : *
430 : * Array scan keys with cross-type equality operators will require a
431 : * separate same-type ORDER proc for sorting their array. Otherwise,
432 : * sortproc just points to the same proc used during binary searches.
433 : */
434 976 : _bt_setup_array_cmp(scan, cur, elemtype,
435 976 : &so->orderProcs[i], &sortprocp);
436 :
437 : /*
438 : * Sort the non-null elements and eliminate any duplicates. We must
439 : * sort in the same ordering used by the index column, so that the
440 : * arrays can be advanced in lockstep with the scan's progress through
441 : * the index's key space.
442 : */
443 976 : reverse = (indoption[cur->sk_attno - 1] & INDOPTION_DESC) != 0;
444 976 : num_elems = _bt_sort_array_elements(cur, sortprocp, reverse,
445 : elem_values, num_nonnulls);
446 :
447 976 : if (origarrayatt == cur->sk_attno)
448 : {
449 12 : BTArrayKeyInfo *orig = &so->arrayKeys[origarraykey];
450 :
451 : /*
452 : * This array scan key is redundant with a previous equality
453 : * operator array scan key. Merge the two arrays together to
454 : * eliminate contradictory non-intersecting elements (or try to).
455 : *
456 : * We merge this next array back into attribute's original array.
457 : */
458 : Assert(arrayKeyData[orig->scan_key].sk_attno == cur->sk_attno);
459 : Assert(arrayKeyData[orig->scan_key].sk_collation ==
460 : cur->sk_collation);
461 12 : if (_bt_merge_arrays(scan, cur, sortprocp, reverse,
462 : origelemtype, elemtype,
463 : orig->elem_values, &orig->num_elems,
464 : elem_values, num_elems))
465 : {
466 : /* Successfully eliminated this array */
467 12 : pfree(elem_values);
468 :
469 : /*
470 : * If no intersecting elements remain in the original array,
471 : * the scan qual is unsatisfiable
472 : */
473 12 : if (orig->num_elems == 0)
474 : {
475 6 : so->qual_ok = false;
476 6 : break;
477 : }
478 :
479 : /*
480 : * Indicate to _bt_preprocess_keys caller that it must ignore
481 : * this scan key
482 : */
483 6 : cur->sk_strategy = InvalidStrategy;
484 6 : continue;
485 : }
486 :
487 : /*
488 : * Unable to merge this array with previous array due to a lack of
489 : * suitable cross-type opfamily support. Will need to keep both
490 : * scan keys/arrays.
491 : */
492 : }
493 : else
494 : {
495 : /*
496 : * This array is the first for current index attribute.
497 : *
498 : * If it turns out to not be the last array (that is, if the next
499 : * array is redundantly applied to this same index attribute),
500 : * we'll then treat this array as the attribute's "original" array
501 : * when merging.
502 : */
503 964 : origarrayatt = cur->sk_attno;
504 964 : origarraykey = numArrayKeys;
505 964 : origelemtype = elemtype;
506 : }
507 :
508 : /*
509 : * And set up the BTArrayKeyInfo data.
510 : *
511 : * Note: _bt_preprocess_array_keys_final will fix-up each array's
512 : * scan_key field later on, after so->keyData[] has been finalized.
513 : */
514 964 : so->arrayKeys[numArrayKeys].scan_key = i;
515 964 : so->arrayKeys[numArrayKeys].num_elems = num_elems;
516 964 : so->arrayKeys[numArrayKeys].elem_values = elem_values;
517 964 : numArrayKeys++;
518 : }
519 :
520 832 : so->numArrayKeys = numArrayKeys;
521 :
522 832 : MemoryContextSwitchTo(oldContext);
523 :
524 832 : return arrayKeyData;
525 : }
526 :
527 : /*
528 : * _bt_preprocess_array_keys_final() -- fix up array scan key references
529 : *
530 : * When _bt_preprocess_array_keys performed initial array preprocessing, it
531 : * set each array's array->scan_key to the array's arrayKeys[] entry offset
532 : * (that also work as references into the original scan->keyData[] array).
533 : * This function handles translation of the scan key references from the
534 : * BTArrayKeyInfo info array, from input scan key references (to the keys in
535 : * scan->keyData[]), into output references (to the keys in so->keyData[]).
536 : * Caller's keyDataMap[] array tells us how to perform this remapping.
537 : *
538 : * Also finalizes so->orderProcs[] for the scan. Arrays already have an ORDER
539 : * proc, which might need to be repositioned to its so->keyData[]-wise offset
540 : * (very much like the remapping that we apply to array->scan_key references).
541 : * Non-array equality strategy scan keys (that survived preprocessing) don't
542 : * yet have an so->orderProcs[] entry, so we set one for them here.
543 : *
544 : * Also converts single-element array scan keys into equivalent non-array
545 : * equality scan keys, which decrements so->numArrayKeys. It's possible that
546 : * this will leave this new btrescan without any arrays at all. This isn't
547 : * necessary for correctness; it's just an optimization. Non-array equality
548 : * scan keys are slightly faster than equivalent array scan keys at runtime.
549 : */
550 : static void
551 384 : _bt_preprocess_array_keys_final(IndexScanDesc scan, int *keyDataMap)
552 : {
553 384 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
554 384 : Relation rel = scan->indexRelation;
555 384 : int arrayidx = 0;
556 384 : int last_equal_output_ikey PG_USED_FOR_ASSERTS_ONLY = -1;
557 :
558 : Assert(so->qual_ok);
559 :
560 : /*
561 : * Nothing for us to do when _bt_preprocess_array_keys only had to deal
562 : * with array inequalities
563 : */
564 384 : if (so->numArrayKeys == 0)
565 0 : return;
566 :
567 1152 : for (int output_ikey = 0; output_ikey < so->numberOfKeys; output_ikey++)
568 : {
569 786 : ScanKey outkey = so->keyData + output_ikey;
570 : int input_ikey;
571 786 : bool found PG_USED_FOR_ASSERTS_ONLY = false;
572 :
573 : Assert(outkey->sk_strategy != InvalidStrategy);
574 :
575 786 : if (outkey->sk_strategy != BTEqualStrategyNumber)
576 18 : continue;
577 :
578 768 : input_ikey = keyDataMap[output_ikey];
579 :
580 : Assert(last_equal_output_ikey < output_ikey);
581 : Assert(last_equal_output_ikey < input_ikey);
582 768 : last_equal_output_ikey = output_ikey;
583 :
584 : /*
585 : * We're lazy about looking up ORDER procs for non-array keys, since
586 : * not all input keys become output keys. Take care of it now.
587 : */
588 768 : if (!(outkey->sk_flags & SK_SEARCHARRAY))
589 : {
590 : Oid elemtype;
591 :
592 : /* No need for an ORDER proc given an IS NULL scan key */
593 246 : if (outkey->sk_flags & SK_SEARCHNULL)
594 6 : continue;
595 :
596 : /*
597 : * A non-required scan key doesn't need an ORDER proc, either
598 : * (unless it's associated with an array, which this one isn't)
599 : */
600 240 : if (!(outkey->sk_flags & SK_BT_REQFWD))
601 90 : continue;
602 :
603 150 : elemtype = outkey->sk_subtype;
604 150 : if (elemtype == InvalidOid)
605 0 : elemtype = rel->rd_opcintype[outkey->sk_attno - 1];
606 :
607 150 : _bt_setup_array_cmp(scan, outkey, elemtype,
608 150 : &so->orderProcs[output_ikey], NULL);
609 150 : continue;
610 : }
611 :
612 : /*
613 : * Reorder existing array scan key so->orderProcs[] entries.
614 : *
615 : * Doing this in-place is safe because preprocessing is required to
616 : * output all equality strategy scan keys in original input order
617 : * (among each group of entries against the same index attribute).
618 : * This is also the order that the arrays themselves appear in.
619 : */
620 522 : so->orderProcs[output_ikey] = so->orderProcs[input_ikey];
621 :
622 : /* Fix-up array->scan_key references for arrays */
623 522 : for (; arrayidx < so->numArrayKeys; arrayidx++)
624 : {
625 522 : BTArrayKeyInfo *array = &so->arrayKeys[arrayidx];
626 :
627 : Assert(array->num_elems > 0);
628 :
629 522 : if (array->scan_key == input_ikey)
630 : {
631 : /* found it */
632 522 : array->scan_key = output_ikey;
633 522 : found = true;
634 :
635 : /*
636 : * Transform array scan keys that have exactly 1 element
637 : * remaining (following all prior preprocessing) into
638 : * equivalent non-array scan keys.
639 : */
640 522 : if (array->num_elems == 1)
641 : {
642 24 : outkey->sk_flags &= ~SK_SEARCHARRAY;
643 24 : outkey->sk_argument = array->elem_values[0];
644 24 : so->numArrayKeys--;
645 :
646 : /* If we're out of array keys, we can quit right away */
647 24 : if (so->numArrayKeys == 0)
648 18 : return;
649 :
650 : /* Shift other arrays forward */
651 6 : memmove(array, array + 1,
652 : sizeof(BTArrayKeyInfo) *
653 6 : (so->numArrayKeys - arrayidx));
654 :
655 : /*
656 : * Don't increment arrayidx (there was an entry that was
657 : * just shifted forward to the offset at arrayidx, which
658 : * will still need to be matched)
659 : */
660 : }
661 : else
662 : {
663 : /* Match found, so done with this array */
664 498 : arrayidx++;
665 : }
666 :
667 504 : break;
668 : }
669 : }
670 :
671 : Assert(found);
672 : }
673 :
674 : /*
675 : * Parallel index scans require space in shared memory to store the
676 : * current array elements (for arrays kept by preprocessing) to schedule
677 : * the next primitive index scan. The underlying structure is protected
678 : * using a spinlock, so defensively limit its size. In practice this can
679 : * only affect parallel scans that use an incomplete opfamily.
680 : */
681 366 : if (scan->parallel_scan && so->numArrayKeys > INDEX_MAX_KEYS)
682 0 : ereport(ERROR,
683 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
684 : errmsg_internal("number of array scan keys left by preprocessing (%d) exceeds the maximum allowed by parallel btree index scans (%d)",
685 : so->numArrayKeys, INDEX_MAX_KEYS)));
686 : }
687 :
688 : /*
689 : * _bt_setup_array_cmp() -- Set up array comparison functions
690 : *
691 : * Sets ORDER proc in caller's orderproc argument, which is used during binary
692 : * searches of arrays during the index scan. Also sets a same-type ORDER proc
693 : * in caller's *sortprocp argument, which is used when sorting the array.
694 : *
695 : * Preprocessing calls here with all equality strategy scan keys (when scan
696 : * uses equality array keys), including those not associated with any array.
697 : * See _bt_advance_array_keys for an explanation of why it'll need to treat
698 : * simple scalar equality scan keys as degenerate single element arrays.
699 : *
700 : * Caller should pass an orderproc pointing to space that'll store the ORDER
701 : * proc for the scan, and a *sortprocp pointing to its own separate space.
702 : * When calling here for a non-array scan key, sortprocp arg should be NULL.
703 : *
704 : * In the common case where we don't need to deal with cross-type operators,
705 : * only one ORDER proc is actually required by caller. We'll set *sortprocp
706 : * to point to the same memory that caller's orderproc continues to point to.
707 : * Otherwise, *sortprocp will continue to point to caller's own space. Either
708 : * way, *sortprocp will point to a same-type ORDER proc (since that's the only
709 : * safe way to sort/deduplicate the array associated with caller's scan key).
710 : */
711 : static void
712 1126 : _bt_setup_array_cmp(IndexScanDesc scan, ScanKey skey, Oid elemtype,
713 : FmgrInfo *orderproc, FmgrInfo **sortprocp)
714 : {
715 1126 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
716 1126 : Relation rel = scan->indexRelation;
717 : RegProcedure cmp_proc;
718 1126 : Oid opcintype = rel->rd_opcintype[skey->sk_attno - 1];
719 :
720 : Assert(skey->sk_strategy == BTEqualStrategyNumber);
721 : Assert(OidIsValid(elemtype));
722 :
723 : /*
724 : * If scankey operator is not a cross-type comparison, we can use the
725 : * cached comparison function; otherwise gotta look it up in the catalogs
726 : */
727 1126 : if (elemtype == opcintype)
728 : {
729 : /* Set same-type ORDER procs for caller */
730 1102 : *orderproc = *index_getprocinfo(rel, skey->sk_attno, BTORDER_PROC);
731 1102 : if (sortprocp)
732 970 : *sortprocp = orderproc;
733 :
734 1102 : return;
735 : }
736 :
737 : /*
738 : * Look up the appropriate cross-type comparison function in the opfamily.
739 : *
740 : * Use the opclass input type as the left hand arg type, and the array
741 : * element type as the right hand arg type (since binary searches use an
742 : * index tuple's attribute value to search for a matching array element).
743 : *
744 : * Note: it's possible that this would fail, if the opfamily is
745 : * incomplete, but only in cases where it's quite likely that _bt_first
746 : * would fail in just the same way (had we not failed before it could).
747 : */
748 24 : cmp_proc = get_opfamily_proc(rel->rd_opfamily[skey->sk_attno - 1],
749 : opcintype, elemtype, BTORDER_PROC);
750 24 : if (!RegProcedureIsValid(cmp_proc))
751 0 : elog(ERROR, "missing support function %d(%u,%u) for attribute %d of index \"%s\"",
752 : BTORDER_PROC, opcintype, elemtype, skey->sk_attno,
753 : RelationGetRelationName(rel));
754 :
755 : /* Set cross-type ORDER proc for caller */
756 24 : fmgr_info_cxt(cmp_proc, orderproc, so->arrayContext);
757 :
758 : /* Done if caller doesn't actually have an array they'll need to sort */
759 24 : if (!sortprocp)
760 18 : return;
761 :
762 : /*
763 : * Look up the appropriate same-type comparison function in the opfamily.
764 : *
765 : * Note: it's possible that this would fail, if the opfamily is
766 : * incomplete, but it seems quite unlikely that an opfamily would omit
767 : * non-cross-type comparison procs for any datatype that it supports at
768 : * all.
769 : */
770 6 : cmp_proc = get_opfamily_proc(rel->rd_opfamily[skey->sk_attno - 1],
771 : elemtype, elemtype, BTORDER_PROC);
772 6 : if (!RegProcedureIsValid(cmp_proc))
773 0 : elog(ERROR, "missing support function %d(%u,%u) for attribute %d of index \"%s\"",
774 : BTORDER_PROC, elemtype, elemtype,
775 : skey->sk_attno, RelationGetRelationName(rel));
776 :
777 : /* Set same-type ORDER proc for caller */
778 6 : fmgr_info_cxt(cmp_proc, *sortprocp, so->arrayContext);
779 : }
780 :
781 : /*
782 : * _bt_find_extreme_element() -- get least or greatest array element
783 : *
784 : * scan and skey identify the index column, whose opfamily determines the
785 : * comparison semantics. strat should be BTLessStrategyNumber to get the
786 : * least element, or BTGreaterStrategyNumber to get the greatest.
787 : */
788 : static Datum
789 6 : _bt_find_extreme_element(IndexScanDesc scan, ScanKey skey, Oid elemtype,
790 : StrategyNumber strat,
791 : Datum *elems, int nelems)
792 : {
793 6 : Relation rel = scan->indexRelation;
794 : Oid cmp_op;
795 : RegProcedure cmp_proc;
796 : FmgrInfo flinfo;
797 : Datum result;
798 : int i;
799 :
800 : /*
801 : * Look up the appropriate comparison operator in the opfamily.
802 : *
803 : * Note: it's possible that this would fail, if the opfamily is
804 : * incomplete, but it seems quite unlikely that an opfamily would omit
805 : * non-cross-type comparison operators for any datatype that it supports
806 : * at all.
807 : */
808 : Assert(skey->sk_strategy != BTEqualStrategyNumber);
809 : Assert(OidIsValid(elemtype));
810 6 : cmp_op = get_opfamily_member(rel->rd_opfamily[skey->sk_attno - 1],
811 : elemtype,
812 : elemtype,
813 : strat);
814 6 : if (!OidIsValid(cmp_op))
815 0 : elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
816 : strat, elemtype, elemtype,
817 : rel->rd_opfamily[skey->sk_attno - 1]);
818 6 : cmp_proc = get_opcode(cmp_op);
819 6 : if (!RegProcedureIsValid(cmp_proc))
820 0 : elog(ERROR, "missing oprcode for operator %u", cmp_op);
821 :
822 6 : fmgr_info(cmp_proc, &flinfo);
823 :
824 : Assert(nelems > 0);
825 6 : result = elems[0];
826 12 : for (i = 1; i < nelems; i++)
827 : {
828 6 : if (DatumGetBool(FunctionCall2Coll(&flinfo,
829 : skey->sk_collation,
830 6 : elems[i],
831 : result)))
832 0 : result = elems[i];
833 : }
834 :
835 6 : return result;
836 : }
837 :
838 : /*
839 : * _bt_sort_array_elements() -- sort and de-dup array elements
840 : *
841 : * The array elements are sorted in-place, and the new number of elements
842 : * after duplicate removal is returned.
843 : *
844 : * skey identifies the index column whose opfamily determines the comparison
845 : * semantics, and sortproc is a corresponding ORDER proc. If reverse is true,
846 : * we sort in descending order.
847 : */
848 : static int
849 976 : _bt_sort_array_elements(ScanKey skey, FmgrInfo *sortproc, bool reverse,
850 : Datum *elems, int nelems)
851 : {
852 : BTSortArrayContext cxt;
853 :
854 976 : if (nelems <= 1)
855 12 : return nelems; /* no work to do */
856 :
857 : /* Sort the array elements */
858 964 : cxt.sortproc = sortproc;
859 964 : cxt.collation = skey->sk_collation;
860 964 : cxt.reverse = reverse;
861 964 : qsort_arg(elems, nelems, sizeof(Datum),
862 : _bt_compare_array_elements, &cxt);
863 :
864 : /* Now scan the sorted elements and remove duplicates */
865 964 : return qunique_arg(elems, nelems, sizeof(Datum),
866 : _bt_compare_array_elements, &cxt);
867 : }
868 :
869 : /*
870 : * _bt_merge_arrays() -- merge next array's elements into an original array
871 : *
872 : * Called when preprocessing encounters a pair of array equality scan keys,
873 : * both against the same index attribute (during initial array preprocessing).
874 : * Merging reorganizes caller's original array (the left hand arg) in-place,
875 : * without ever copying elements from one array into the other. (Mixing the
876 : * elements together like this would be wrong, since they don't necessarily
877 : * use the same underlying element type, despite all the other similarities.)
878 : *
879 : * Both arrays must have already been sorted and deduplicated by calling
880 : * _bt_sort_array_elements. sortproc is the same-type ORDER proc that was
881 : * just used to sort and deduplicate caller's "next" array. We'll usually be
882 : * able to reuse that order PROC to merge the arrays together now. If not,
883 : * then we'll perform a separate ORDER proc lookup.
884 : *
885 : * If the opfamily doesn't supply a complete set of cross-type ORDER procs we
886 : * may not be able to determine which elements are contradictory. If we have
887 : * the required ORDER proc then we return true (and validly set *nelems_orig),
888 : * guaranteeing that at least the next array can be considered redundant. We
889 : * return false if the required comparisons cannot not be made (caller must
890 : * keep both arrays when this happens).
891 : */
892 : static bool
893 12 : _bt_merge_arrays(IndexScanDesc scan, ScanKey skey, FmgrInfo *sortproc,
894 : bool reverse, Oid origelemtype, Oid nextelemtype,
895 : Datum *elems_orig, int *nelems_orig,
896 : Datum *elems_next, int nelems_next)
897 : {
898 12 : Relation rel = scan->indexRelation;
899 12 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
900 : BTSortArrayContext cxt;
901 12 : int nelems_orig_start = *nelems_orig,
902 12 : nelems_orig_merged = 0;
903 12 : FmgrInfo *mergeproc = sortproc;
904 : FmgrInfo crosstypeproc;
905 :
906 : Assert(skey->sk_strategy == BTEqualStrategyNumber);
907 : Assert(OidIsValid(origelemtype) && OidIsValid(nextelemtype));
908 :
909 12 : if (origelemtype != nextelemtype)
910 : {
911 : RegProcedure cmp_proc;
912 :
913 : /*
914 : * Cross-array-element-type merging is required, so can't just reuse
915 : * sortproc when merging
916 : */
917 6 : cmp_proc = get_opfamily_proc(rel->rd_opfamily[skey->sk_attno - 1],
918 : origelemtype, nextelemtype, BTORDER_PROC);
919 6 : if (!RegProcedureIsValid(cmp_proc))
920 : {
921 : /* Can't make the required comparisons */
922 0 : return false;
923 : }
924 :
925 : /* We have all we need to determine redundancy/contradictoriness */
926 6 : mergeproc = &crosstypeproc;
927 6 : fmgr_info_cxt(cmp_proc, mergeproc, so->arrayContext);
928 : }
929 :
930 12 : cxt.sortproc = mergeproc;
931 12 : cxt.collation = skey->sk_collation;
932 12 : cxt.reverse = reverse;
933 :
934 54 : for (int i = 0, j = 0; i < nelems_orig_start && j < nelems_next;)
935 : {
936 42 : Datum *oelem = elems_orig + i,
937 42 : *nelem = elems_next + j;
938 42 : int res = _bt_compare_array_elements(oelem, nelem, &cxt);
939 :
940 42 : if (res == 0)
941 : {
942 6 : elems_orig[nelems_orig_merged++] = *oelem;
943 6 : i++;
944 6 : j++;
945 : }
946 36 : else if (res < 0)
947 24 : i++;
948 : else /* res > 0 */
949 12 : j++;
950 : }
951 :
952 12 : *nelems_orig = nelems_orig_merged;
953 :
954 12 : return true;
955 : }
956 :
957 : /*
958 : * Compare an array scan key to a scalar scan key, eliminating contradictory
959 : * array elements such that the scalar scan key becomes redundant.
960 : *
961 : * Array elements can be eliminated as contradictory when excluded by some
962 : * other operator on the same attribute. For example, with an index scan qual
963 : * "WHERE a IN (1, 2, 3) AND a < 2", all array elements except the value "1"
964 : * are eliminated, and the < scan key is eliminated as redundant. Cases where
965 : * every array element is eliminated by a redundant scalar scan key have an
966 : * unsatisfiable qual, which we handle by setting *qual_ok=false for caller.
967 : *
968 : * If the opfamily doesn't supply a complete set of cross-type ORDER procs we
969 : * may not be able to determine which elements are contradictory. If we have
970 : * the required ORDER proc then we return true (and validly set *qual_ok),
971 : * guaranteeing that at least the scalar scan key can be considered redundant.
972 : * We return false if the comparison could not be made (caller must keep both
973 : * scan keys when this happens).
974 : */
975 : static bool
976 30 : _bt_compare_array_scankey_args(IndexScanDesc scan, ScanKey arraysk, ScanKey skey,
977 : FmgrInfo *orderproc, BTArrayKeyInfo *array,
978 : bool *qual_ok)
979 : {
980 30 : Relation rel = scan->indexRelation;
981 30 : Oid opcintype = rel->rd_opcintype[arraysk->sk_attno - 1];
982 30 : int cmpresult = 0,
983 30 : cmpexact = 0,
984 : matchelem,
985 30 : new_nelems = 0;
986 : FmgrInfo crosstypeproc;
987 30 : FmgrInfo *orderprocp = orderproc;
988 :
989 : Assert(arraysk->sk_attno == skey->sk_attno);
990 : Assert(array->num_elems > 0);
991 : Assert(!(arraysk->sk_flags & (SK_ISNULL | SK_ROW_HEADER | SK_ROW_MEMBER)));
992 : Assert((arraysk->sk_flags & SK_SEARCHARRAY) &&
993 : arraysk->sk_strategy == BTEqualStrategyNumber);
994 : Assert(!(skey->sk_flags & (SK_ISNULL | SK_ROW_HEADER | SK_ROW_MEMBER)));
995 : Assert(!(skey->sk_flags & SK_SEARCHARRAY) ||
996 : skey->sk_strategy != BTEqualStrategyNumber);
997 :
998 : /*
999 : * _bt_binsrch_array_skey searches an array for the entry best matching a
1000 : * datum of opclass input type for the index's attribute (on-disk type).
1001 : * We can reuse the array's ORDER proc whenever the non-array scan key's
1002 : * type is a match for the corresponding attribute's input opclass type.
1003 : * Otherwise, we have to do another ORDER proc lookup so that our call to
1004 : * _bt_binsrch_array_skey applies the correct comparator.
1005 : *
1006 : * Note: we have to support the convention that sk_subtype == InvalidOid
1007 : * means the opclass input type; this is a hack to simplify life for
1008 : * ScanKeyInit().
1009 : */
1010 30 : if (skey->sk_subtype != opcintype && skey->sk_subtype != InvalidOid)
1011 : {
1012 : RegProcedure cmp_proc;
1013 : Oid arraysk_elemtype;
1014 :
1015 : /*
1016 : * Need an ORDER proc lookup to detect redundancy/contradictoriness
1017 : * with this pair of scankeys.
1018 : *
1019 : * Scalar scan key's argument will be passed to _bt_compare_array_skey
1020 : * as its tupdatum/lefthand argument (rhs arg is for array elements).
1021 : */
1022 6 : arraysk_elemtype = arraysk->sk_subtype;
1023 6 : if (arraysk_elemtype == InvalidOid)
1024 0 : arraysk_elemtype = rel->rd_opcintype[arraysk->sk_attno - 1];
1025 6 : cmp_proc = get_opfamily_proc(rel->rd_opfamily[arraysk->sk_attno - 1],
1026 : skey->sk_subtype, arraysk_elemtype,
1027 : BTORDER_PROC);
1028 6 : if (!RegProcedureIsValid(cmp_proc))
1029 : {
1030 : /* Can't make the comparison */
1031 0 : *qual_ok = false; /* suppress compiler warnings */
1032 0 : return false;
1033 : }
1034 :
1035 : /* We have all we need to determine redundancy/contradictoriness */
1036 6 : orderprocp = &crosstypeproc;
1037 6 : fmgr_info(cmp_proc, orderprocp);
1038 : }
1039 :
1040 30 : matchelem = _bt_binsrch_array_skey(orderprocp, false,
1041 : NoMovementScanDirection,
1042 : skey->sk_argument, false, array,
1043 : arraysk, &cmpresult);
1044 :
1045 30 : switch (skey->sk_strategy)
1046 : {
1047 6 : case BTLessStrategyNumber:
1048 6 : cmpexact = 1; /* exclude exact match, if any */
1049 : /* FALL THRU */
1050 6 : case BTLessEqualStrategyNumber:
1051 6 : if (cmpresult >= cmpexact)
1052 0 : matchelem++;
1053 : /* Resize, keeping elements from the start of the array */
1054 6 : new_nelems = matchelem;
1055 6 : break;
1056 12 : case BTEqualStrategyNumber:
1057 12 : if (cmpresult != 0)
1058 : {
1059 : /* qual is unsatisfiable */
1060 6 : new_nelems = 0;
1061 : }
1062 : else
1063 : {
1064 : /* Shift matching element to the start of the array, resize */
1065 6 : array->elem_values[0] = array->elem_values[matchelem];
1066 6 : new_nelems = 1;
1067 : }
1068 12 : break;
1069 6 : case BTGreaterEqualStrategyNumber:
1070 6 : cmpexact = 1; /* include exact match, if any */
1071 : /* FALL THRU */
1072 12 : case BTGreaterStrategyNumber:
1073 12 : if (cmpresult >= cmpexact)
1074 6 : matchelem++;
1075 : /* Shift matching elements to the start of the array, resize */
1076 12 : new_nelems = array->num_elems - matchelem;
1077 12 : memmove(array->elem_values, array->elem_values + matchelem,
1078 : sizeof(Datum) * new_nelems);
1079 12 : break;
1080 0 : default:
1081 0 : elog(ERROR, "unrecognized StrategyNumber: %d",
1082 : (int) skey->sk_strategy);
1083 : break;
1084 : }
1085 :
1086 : Assert(new_nelems >= 0);
1087 : Assert(new_nelems <= array->num_elems);
1088 :
1089 30 : array->num_elems = new_nelems;
1090 30 : *qual_ok = new_nelems > 0;
1091 :
1092 30 : return true;
1093 : }
1094 :
1095 : /*
1096 : * qsort_arg comparator for sorting array elements
1097 : */
1098 : static int
1099 8694 : _bt_compare_array_elements(const void *a, const void *b, void *arg)
1100 : {
1101 8694 : Datum da = *((const Datum *) a);
1102 8694 : Datum db = *((const Datum *) b);
1103 8694 : BTSortArrayContext *cxt = (BTSortArrayContext *) arg;
1104 : int32 compare;
1105 :
1106 8694 : compare = DatumGetInt32(FunctionCall2Coll(cxt->sortproc,
1107 : cxt->collation,
1108 : da, db));
1109 8694 : if (cxt->reverse)
1110 30 : INVERT_COMPARE_RESULT(compare);
1111 8694 : return compare;
1112 : }
1113 :
1114 : /*
1115 : * _bt_compare_array_skey() -- apply array comparison function
1116 : *
1117 : * Compares caller's tuple attribute value to a scan key/array element.
1118 : * Helper function used during binary searches of SK_SEARCHARRAY arrays.
1119 : *
1120 : * This routine returns:
1121 : * <0 if tupdatum < arrdatum;
1122 : * 0 if tupdatum == arrdatum;
1123 : * >0 if tupdatum > arrdatum.
1124 : *
1125 : * This is essentially the same interface as _bt_compare: both functions
1126 : * compare the value that they're searching for to a binary search pivot.
1127 : * However, unlike _bt_compare, this function's "tuple argument" comes first,
1128 : * while its "array/scankey argument" comes second.
1129 : */
1130 : static inline int32
1131 12372 : _bt_compare_array_skey(FmgrInfo *orderproc,
1132 : Datum tupdatum, bool tupnull,
1133 : Datum arrdatum, ScanKey cur)
1134 : {
1135 12372 : int32 result = 0;
1136 :
1137 : Assert(cur->sk_strategy == BTEqualStrategyNumber);
1138 :
1139 12372 : if (tupnull) /* NULL tupdatum */
1140 : {
1141 6 : if (cur->sk_flags & SK_ISNULL)
1142 6 : result = 0; /* NULL "=" NULL */
1143 0 : else if (cur->sk_flags & SK_BT_NULLS_FIRST)
1144 0 : result = -1; /* NULL "<" NOT_NULL */
1145 : else
1146 0 : result = 1; /* NULL ">" NOT_NULL */
1147 : }
1148 12366 : else if (cur->sk_flags & SK_ISNULL) /* NOT_NULL tupdatum, NULL arrdatum */
1149 : {
1150 6 : if (cur->sk_flags & SK_BT_NULLS_FIRST)
1151 0 : result = 1; /* NOT_NULL ">" NULL */
1152 : else
1153 6 : result = -1; /* NOT_NULL "<" NULL */
1154 : }
1155 : else
1156 : {
1157 : /*
1158 : * Like _bt_compare, we need to be careful of cross-type comparisons,
1159 : * so the left value has to be the value that came from an index tuple
1160 : */
1161 12360 : result = DatumGetInt32(FunctionCall2Coll(orderproc, cur->sk_collation,
1162 : tupdatum, arrdatum));
1163 :
1164 : /*
1165 : * We flip the sign by following the obvious rule: flip whenever the
1166 : * column is a DESC column.
1167 : *
1168 : * _bt_compare does it the wrong way around (flip when *ASC*) in order
1169 : * to compensate for passing its orderproc arguments backwards. We
1170 : * don't need to play these games because we find it natural to pass
1171 : * tupdatum as the left value (and arrdatum as the right value).
1172 : */
1173 12360 : if (cur->sk_flags & SK_BT_DESC)
1174 24 : INVERT_COMPARE_RESULT(result);
1175 : }
1176 :
1177 12372 : return result;
1178 : }
1179 :
1180 : /*
1181 : * _bt_binsrch_array_skey() -- Binary search for next matching array key
1182 : *
1183 : * Returns an index to the first array element >= caller's tupdatum argument.
1184 : * This convention is more natural for forwards scan callers, but that can't
1185 : * really matter to backwards scan callers. Both callers require handling for
1186 : * the case where the match we return is < tupdatum, and symmetric handling
1187 : * for the case where our best match is > tupdatum.
1188 : *
1189 : * Also sets *set_elem_result to the result _bt_compare_array_skey returned
1190 : * when we used it to compare the matching array element to tupdatum/tupnull.
1191 : *
1192 : * cur_elem_trig indicates if array advancement was triggered by this array's
1193 : * scan key, and that the array is for a required scan key. We can apply this
1194 : * information to find the next matching array element in the current scan
1195 : * direction using far fewer comparisons (fewer on average, compared to naive
1196 : * binary search). This scheme takes advantage of an important property of
1197 : * required arrays: required arrays always advance in lockstep with the index
1198 : * scan's progress through the index's key space.
1199 : */
1200 : static int
1201 3948 : _bt_binsrch_array_skey(FmgrInfo *orderproc,
1202 : bool cur_elem_trig, ScanDirection dir,
1203 : Datum tupdatum, bool tupnull,
1204 : BTArrayKeyInfo *array, ScanKey cur,
1205 : int32 *set_elem_result)
1206 : {
1207 3948 : int low_elem = 0,
1208 3948 : mid_elem = -1,
1209 3948 : high_elem = array->num_elems - 1,
1210 3948 : result = 0;
1211 : Datum arrdatum;
1212 :
1213 : Assert(cur->sk_flags & SK_SEARCHARRAY);
1214 : Assert(cur->sk_strategy == BTEqualStrategyNumber);
1215 :
1216 3948 : if (cur_elem_trig)
1217 : {
1218 : Assert(!ScanDirectionIsNoMovement(dir));
1219 : Assert(cur->sk_flags & SK_BT_REQFWD);
1220 :
1221 : /*
1222 : * When the scan key that triggered array advancement is a required
1223 : * array scan key, it is now certain that the current array element
1224 : * (plus all prior elements relative to the current scan direction)
1225 : * cannot possibly be at or ahead of the corresponding tuple value.
1226 : * (_bt_checkkeys must have called _bt_tuple_before_array_skeys, which
1227 : * makes sure this is true as a condition of advancing the arrays.)
1228 : *
1229 : * This makes it safe to exclude array elements up to and including
1230 : * the former-current array element from our search.
1231 : *
1232 : * Separately, when array advancement was triggered by a required scan
1233 : * key, the array element immediately after the former-current element
1234 : * is often either an exact tupdatum match, or a "close by" near-match
1235 : * (a near-match tupdatum is one whose key space falls _between_ the
1236 : * former-current and new-current array elements). We'll detect both
1237 : * cases via an optimistic comparison of the new search lower bound
1238 : * (or new search upper bound in the case of backwards scans).
1239 : */
1240 3462 : if (ScanDirectionIsForward(dir))
1241 : {
1242 3438 : low_elem = array->cur_elem + 1; /* old cur_elem exhausted */
1243 :
1244 : /* Compare prospective new cur_elem (also the new lower bound) */
1245 3438 : if (high_elem >= low_elem)
1246 : {
1247 2874 : arrdatum = array->elem_values[low_elem];
1248 2874 : result = _bt_compare_array_skey(orderproc, tupdatum, tupnull,
1249 : arrdatum, cur);
1250 :
1251 2874 : if (result <= 0)
1252 : {
1253 : /* Optimistic comparison optimization worked out */
1254 2812 : *set_elem_result = result;
1255 2812 : return low_elem;
1256 : }
1257 62 : mid_elem = low_elem;
1258 62 : low_elem++; /* this cur_elem exhausted, too */
1259 : }
1260 :
1261 626 : if (high_elem < low_elem)
1262 : {
1263 : /* Caller needs to perform "beyond end" array advancement */
1264 564 : *set_elem_result = 1;
1265 564 : return high_elem;
1266 : }
1267 : }
1268 : else
1269 : {
1270 24 : high_elem = array->cur_elem - 1; /* old cur_elem exhausted */
1271 :
1272 : /* Compare prospective new cur_elem (also the new upper bound) */
1273 24 : if (high_elem >= low_elem)
1274 : {
1275 18 : arrdatum = array->elem_values[high_elem];
1276 18 : result = _bt_compare_array_skey(orderproc, tupdatum, tupnull,
1277 : arrdatum, cur);
1278 :
1279 18 : if (result >= 0)
1280 : {
1281 : /* Optimistic comparison optimization worked out */
1282 18 : *set_elem_result = result;
1283 18 : return high_elem;
1284 : }
1285 0 : mid_elem = high_elem;
1286 0 : high_elem--; /* this cur_elem exhausted, too */
1287 : }
1288 :
1289 6 : if (high_elem < low_elem)
1290 : {
1291 : /* Caller needs to perform "beyond end" array advancement */
1292 6 : *set_elem_result = -1;
1293 6 : return low_elem;
1294 : }
1295 : }
1296 : }
1297 :
1298 1004 : while (high_elem > low_elem)
1299 : {
1300 600 : mid_elem = low_elem + ((high_elem - low_elem) / 2);
1301 600 : arrdatum = array->elem_values[mid_elem];
1302 :
1303 600 : result = _bt_compare_array_skey(orderproc, tupdatum, tupnull,
1304 : arrdatum, cur);
1305 :
1306 600 : if (result == 0)
1307 : {
1308 : /*
1309 : * It's safe to quit as soon as we see an equal array element.
1310 : * This often saves an extra comparison or two...
1311 : */
1312 144 : low_elem = mid_elem;
1313 144 : break;
1314 : }
1315 :
1316 456 : if (result > 0)
1317 402 : low_elem = mid_elem + 1;
1318 : else
1319 54 : high_elem = mid_elem;
1320 : }
1321 :
1322 : /*
1323 : * ...but our caller also cares about how its searched-for tuple datum
1324 : * compares to the low_elem datum. Must always set *set_elem_result with
1325 : * the result of that comparison specifically.
1326 : */
1327 548 : if (low_elem != mid_elem)
1328 362 : result = _bt_compare_array_skey(orderproc, tupdatum, tupnull,
1329 362 : array->elem_values[low_elem], cur);
1330 :
1331 548 : *set_elem_result = result;
1332 :
1333 548 : return low_elem;
1334 : }
1335 :
1336 : /*
1337 : * _bt_start_array_keys() -- Initialize array keys at start of a scan
1338 : *
1339 : * Set up the cur_elem counters and fill in the first sk_argument value for
1340 : * each array scankey.
1341 : */
1342 : void
1343 1402 : _bt_start_array_keys(IndexScanDesc scan, ScanDirection dir)
1344 : {
1345 1402 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
1346 : int i;
1347 :
1348 : Assert(so->numArrayKeys);
1349 : Assert(so->qual_ok);
1350 :
1351 3068 : for (i = 0; i < so->numArrayKeys; i++)
1352 : {
1353 1666 : BTArrayKeyInfo *curArrayKey = &so->arrayKeys[i];
1354 1666 : ScanKey skey = &so->keyData[curArrayKey->scan_key];
1355 :
1356 : Assert(curArrayKey->num_elems > 0);
1357 : Assert(skey->sk_flags & SK_SEARCHARRAY);
1358 :
1359 1666 : if (ScanDirectionIsBackward(dir))
1360 756 : curArrayKey->cur_elem = curArrayKey->num_elems - 1;
1361 : else
1362 910 : curArrayKey->cur_elem = 0;
1363 1666 : skey->sk_argument = curArrayKey->elem_values[curArrayKey->cur_elem];
1364 : }
1365 1402 : so->scanBehind = false;
1366 1402 : }
1367 :
1368 : /*
1369 : * _bt_advance_array_keys_increment() -- Advance to next set of array elements
1370 : *
1371 : * Advances the array keys by a single increment in the current scan
1372 : * direction. When there are multiple array keys this can roll over from the
1373 : * lowest order array to higher order arrays.
1374 : *
1375 : * Returns true if there is another set of values to consider, false if not.
1376 : * On true result, the scankeys are initialized with the next set of values.
1377 : * On false result, the scankeys stay the same, and the array keys are not
1378 : * advanced (every array remains at its final element for scan direction).
1379 : */
1380 : static bool
1381 632 : _bt_advance_array_keys_increment(IndexScanDesc scan, ScanDirection dir)
1382 : {
1383 632 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
1384 :
1385 : /*
1386 : * We must advance the last array key most quickly, since it will
1387 : * correspond to the lowest-order index column among the available
1388 : * qualifications
1389 : */
1390 1382 : for (int i = so->numArrayKeys - 1; i >= 0; i--)
1391 : {
1392 764 : BTArrayKeyInfo *curArrayKey = &so->arrayKeys[i];
1393 764 : ScanKey skey = &so->keyData[curArrayKey->scan_key];
1394 764 : int cur_elem = curArrayKey->cur_elem;
1395 764 : int num_elems = curArrayKey->num_elems;
1396 764 : bool rolled = false;
1397 :
1398 764 : if (ScanDirectionIsForward(dir) && ++cur_elem >= num_elems)
1399 : {
1400 744 : cur_elem = 0;
1401 744 : rolled = true;
1402 : }
1403 20 : else if (ScanDirectionIsBackward(dir) && --cur_elem < 0)
1404 : {
1405 6 : cur_elem = num_elems - 1;
1406 6 : rolled = true;
1407 : }
1408 :
1409 764 : curArrayKey->cur_elem = cur_elem;
1410 764 : skey->sk_argument = curArrayKey->elem_values[cur_elem];
1411 764 : if (!rolled)
1412 14 : return true;
1413 :
1414 : /* Need to advance next array key, if any */
1415 : }
1416 :
1417 : /*
1418 : * The array keys are now exhausted. (There isn't actually a distinct
1419 : * state that represents array exhaustion, since index scans don't always
1420 : * end after btgettuple returns "false".)
1421 : *
1422 : * Restore the array keys to the state they were in immediately before we
1423 : * were called. This ensures that the arrays only ever ratchet in the
1424 : * current scan direction. Without this, scans would overlook matching
1425 : * tuples if and when the scan's direction was subsequently reversed.
1426 : */
1427 618 : _bt_start_array_keys(scan, -dir);
1428 :
1429 618 : return false;
1430 : }
1431 :
1432 : /*
1433 : * _bt_rewind_nonrequired_arrays() -- Rewind non-required arrays
1434 : *
1435 : * Called when _bt_advance_array_keys decides to start a new primitive index
1436 : * scan on the basis of the current scan position being before the position
1437 : * that _bt_first is capable of repositioning the scan to by applying an
1438 : * inequality operator required in the opposite-to-scan direction only.
1439 : *
1440 : * Although equality strategy scan keys (for both arrays and non-arrays alike)
1441 : * are either marked required in both directions or in neither direction,
1442 : * there is a sense in which non-required arrays behave like required arrays.
1443 : * With a qual such as "WHERE a IN (100, 200) AND b >= 3 AND c IN (5, 6, 7)",
1444 : * the scan key on "c" is non-required, but nevertheless enables positioning
1445 : * the scan at the first tuple >= "(100, 3, 5)" on the leaf level during the
1446 : * first descent of the tree by _bt_first. Later on, there could also be a
1447 : * second descent, that places the scan right before tuples >= "(200, 3, 5)".
1448 : * _bt_first must never be allowed to build an insertion scan key whose "c"
1449 : * entry is set to a value other than 5, the "c" array's first element/value.
1450 : * (Actually, it's the first in the current scan direction. This example uses
1451 : * a forward scan.)
1452 : *
1453 : * Calling here resets the array scan key elements for the scan's non-required
1454 : * arrays. This is strictly necessary for correctness in a subset of cases
1455 : * involving "required in opposite direction"-triggered primitive index scans.
1456 : * Not all callers are at risk of _bt_first using a non-required array like
1457 : * this, but advancement always resets the arrays when another primitive scan
1458 : * is scheduled, just to keep things simple. Array advancement even makes
1459 : * sure to reset non-required arrays during scans that have no inequalities.
1460 : * (Advancement still won't call here when there are no inequalities, though
1461 : * that's just because it's all handled indirectly instead.)
1462 : *
1463 : * Note: _bt_verify_arrays_bt_first is called by an assertion to enforce that
1464 : * everybody got this right.
1465 : */
1466 : static void
1467 0 : _bt_rewind_nonrequired_arrays(IndexScanDesc scan, ScanDirection dir)
1468 : {
1469 0 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
1470 0 : int arrayidx = 0;
1471 :
1472 0 : for (int ikey = 0; ikey < so->numberOfKeys; ikey++)
1473 : {
1474 0 : ScanKey cur = so->keyData + ikey;
1475 0 : BTArrayKeyInfo *array = NULL;
1476 : int first_elem_dir;
1477 :
1478 0 : if (!(cur->sk_flags & SK_SEARCHARRAY) ||
1479 0 : cur->sk_strategy != BTEqualStrategyNumber)
1480 0 : continue;
1481 :
1482 0 : array = &so->arrayKeys[arrayidx++];
1483 : Assert(array->scan_key == ikey);
1484 :
1485 0 : if ((cur->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD)))
1486 0 : continue;
1487 :
1488 0 : if (ScanDirectionIsForward(dir))
1489 0 : first_elem_dir = 0;
1490 : else
1491 0 : first_elem_dir = array->num_elems - 1;
1492 :
1493 0 : if (array->cur_elem != first_elem_dir)
1494 : {
1495 0 : array->cur_elem = first_elem_dir;
1496 0 : cur->sk_argument = array->elem_values[first_elem_dir];
1497 : }
1498 : }
1499 0 : }
1500 :
1501 : /*
1502 : * _bt_tuple_before_array_skeys() -- too early to advance required arrays?
1503 : *
1504 : * We always compare the tuple using the current array keys (which we assume
1505 : * are already set in so->keyData[]). readpagetup indicates if tuple is the
1506 : * scan's current _bt_readpage-wise tuple.
1507 : *
1508 : * readpagetup callers must only call here when _bt_check_compare already set
1509 : * continuescan=false. We help these callers deal with _bt_check_compare's
1510 : * inability to distinguishing between the < and > cases (it uses equality
1511 : * operator scan keys, whereas we use 3-way ORDER procs). These callers pass
1512 : * a _bt_check_compare-set sktrig value that indicates which scan key
1513 : * triggered the call (!readpagetup callers just pass us sktrig=0 instead).
1514 : * This information allows us to avoid wastefully checking earlier scan keys
1515 : * that were already deemed to have been satisfied inside _bt_check_compare.
1516 : *
1517 : * Returns false when caller's tuple is >= the current required equality scan
1518 : * keys (or <=, in the case of backwards scans). This happens to readpagetup
1519 : * callers when the scan has reached the point of needing its array keys
1520 : * advanced; caller will need to advance required and non-required arrays at
1521 : * scan key offsets >= sktrig, plus scan keys < sktrig iff sktrig rolls over.
1522 : * (When we return false to readpagetup callers, tuple can only be == current
1523 : * required equality scan keys when caller's sktrig indicates that the arrays
1524 : * need to be advanced due to an unsatisfied required inequality key trigger.)
1525 : *
1526 : * Returns true when caller passes a tuple that is < the current set of
1527 : * equality keys for the most significant non-equal required scan key/column
1528 : * (or > the keys, during backwards scans). This happens to readpagetup
1529 : * callers when tuple is still before the start of matches for the scan's
1530 : * required equality strategy scan keys. (sktrig can't have indicated that an
1531 : * inequality strategy scan key wasn't satisfied in _bt_check_compare when we
1532 : * return true. In fact, we automatically return false when passed such an
1533 : * inequality sktrig by readpagetup callers -- _bt_check_compare's initial
1534 : * continuescan=false doesn't really need to be confirmed here by us.)
1535 : *
1536 : * !readpagetup callers optionally pass us *scanBehind, which tracks whether
1537 : * any missing truncated attributes might have affected array advancement
1538 : * (compared to what would happen if it was shown the first non-pivot tuple on
1539 : * the page to the right of caller's finaltup/high key tuple instead). It's
1540 : * only possible that we'll set *scanBehind to true when caller passes us a
1541 : * pivot tuple (with truncated -inf attributes) that we return false for.
1542 : */
1543 : static bool
1544 8096 : _bt_tuple_before_array_skeys(IndexScanDesc scan, ScanDirection dir,
1545 : IndexTuple tuple, TupleDesc tupdesc, int tupnatts,
1546 : bool readpagetup, int sktrig, bool *scanBehind)
1547 : {
1548 8096 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
1549 :
1550 : Assert(so->numArrayKeys);
1551 : Assert(so->numberOfKeys);
1552 : Assert(sktrig == 0 || readpagetup);
1553 : Assert(!readpagetup || scanBehind == NULL);
1554 :
1555 8096 : if (scanBehind)
1556 1040 : *scanBehind = false;
1557 :
1558 8210 : for (int ikey = sktrig; ikey < so->numberOfKeys; ikey++)
1559 : {
1560 8166 : ScanKey cur = so->keyData + ikey;
1561 : Datum tupdatum;
1562 : bool tupnull;
1563 : int32 result;
1564 :
1565 : /* readpagetup calls require one ORDER proc comparison (at most) */
1566 : Assert(!readpagetup || ikey == sktrig);
1567 :
1568 : /*
1569 : * Once we reach a non-required scan key, we're completely done.
1570 : *
1571 : * Note: we deliberately don't consider the scan direction here.
1572 : * _bt_advance_array_keys caller requires that we track *scanBehind
1573 : * without concern for scan direction.
1574 : */
1575 8166 : if ((cur->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD)) == 0)
1576 : {
1577 : Assert(!readpagetup);
1578 : Assert(ikey > sktrig || ikey == 0);
1579 8052 : return false;
1580 : }
1581 :
1582 8166 : if (cur->sk_attno > tupnatts)
1583 : {
1584 : Assert(!readpagetup);
1585 :
1586 : /*
1587 : * When we reach a high key's truncated attribute, assume that the
1588 : * tuple attribute's value is >= the scan's equality constraint
1589 : * scan keys (but set *scanBehind to let interested callers know
1590 : * that a truncated attribute might have affected our answer).
1591 : */
1592 6 : if (scanBehind)
1593 6 : *scanBehind = true;
1594 :
1595 6 : return false;
1596 : }
1597 :
1598 : /*
1599 : * Deal with inequality strategy scan keys that _bt_check_compare set
1600 : * continuescan=false for
1601 : */
1602 8160 : if (cur->sk_strategy != BTEqualStrategyNumber)
1603 : {
1604 : /*
1605 : * When _bt_check_compare indicated that a required inequality
1606 : * scan key wasn't satisfied, there's no need to verify anything;
1607 : * caller always calls _bt_advance_array_keys with this sktrig.
1608 : */
1609 6 : if (readpagetup)
1610 6 : return false;
1611 :
1612 : /*
1613 : * Otherwise we can't give up, since we must check all required
1614 : * scan keys (required in either direction) in order to correctly
1615 : * track *scanBehind for caller
1616 : */
1617 0 : continue;
1618 : }
1619 :
1620 8154 : tupdatum = index_getattr(tuple, cur->sk_attno, tupdesc, &tupnull);
1621 :
1622 8154 : result = _bt_compare_array_skey(&so->orderProcs[ikey],
1623 : tupdatum, tupnull,
1624 : cur->sk_argument, cur);
1625 :
1626 : /*
1627 : * Does this comparison indicate that caller must _not_ advance the
1628 : * scan's arrays just yet?
1629 : */
1630 8154 : if ((ScanDirectionIsForward(dir) && result < 0) ||
1631 60 : (ScanDirectionIsBackward(dir) && result > 0))
1632 3600 : return true;
1633 :
1634 : /*
1635 : * Does this comparison indicate that caller should now advance the
1636 : * scan's arrays? (Must be if we get here during a readpagetup call.)
1637 : */
1638 4554 : if (readpagetup || result != 0)
1639 : {
1640 : Assert(result != 0);
1641 4440 : return false;
1642 : }
1643 :
1644 : /*
1645 : * Inconclusive -- need to check later scan keys, too.
1646 : *
1647 : * This must be a finaltup precheck, or a call made from an assertion.
1648 : */
1649 : Assert(result == 0);
1650 : }
1651 :
1652 : Assert(!readpagetup);
1653 :
1654 44 : return false;
1655 : }
1656 :
1657 : /*
1658 : * _bt_start_prim_scan() -- start scheduled primitive index scan?
1659 : *
1660 : * Returns true if _bt_checkkeys scheduled another primitive index scan, just
1661 : * as the last one ended. Otherwise returns false, indicating that the array
1662 : * keys are now fully exhausted.
1663 : *
1664 : * Only call here during scans with one or more equality type array scan keys,
1665 : * after _bt_first or _bt_next return false.
1666 : */
1667 : bool
1668 1364 : _bt_start_prim_scan(IndexScanDesc scan, ScanDirection dir)
1669 : {
1670 1364 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
1671 :
1672 : Assert(so->numArrayKeys);
1673 :
1674 : /* scanBehind flag doesn't persist across primitive index scans - reset */
1675 1364 : so->scanBehind = false;
1676 :
1677 : /*
1678 : * Array keys are advanced within _bt_checkkeys when the scan reaches the
1679 : * leaf level (more precisely, they're advanced when the scan reaches the
1680 : * end of each distinct set of array elements). This process avoids
1681 : * repeat access to leaf pages (across multiple primitive index scans) by
1682 : * advancing the scan's array keys when it allows the primitive index scan
1683 : * to find nearby matching tuples (or when it eliminates ranges of array
1684 : * key space that can't possibly be satisfied by any index tuple).
1685 : *
1686 : * _bt_checkkeys sets a simple flag variable to schedule another primitive
1687 : * index scan. The flag tells us what to do.
1688 : *
1689 : * We cannot rely on _bt_first always reaching _bt_checkkeys. There are
1690 : * various cases where that won't happen. For example, if the index is
1691 : * completely empty, then _bt_first won't call _bt_readpage/_bt_checkkeys.
1692 : * We also don't expect a call to _bt_checkkeys during searches for a
1693 : * non-existent value that happens to be lower/higher than any existing
1694 : * value in the index.
1695 : *
1696 : * We don't require special handling for these cases -- we don't need to
1697 : * be explicitly instructed to _not_ perform another primitive index scan.
1698 : * It's up to code under the control of _bt_first to always set the flag
1699 : * when another primitive index scan will be required.
1700 : *
1701 : * This works correctly, even with the tricky cases listed above, which
1702 : * all involve access to leaf pages "near the boundaries of the key space"
1703 : * (whether it's from a leftmost/rightmost page, or an imaginary empty
1704 : * leaf root page). If _bt_checkkeys cannot be reached by a primitive
1705 : * index scan for one set of array keys, then it also won't be reached for
1706 : * any later set ("later" in terms of the direction that we scan the index
1707 : * and advance the arrays). The array keys won't have advanced in these
1708 : * cases, but that's the correct behavior (even _bt_advance_array_keys
1709 : * won't always advance the arrays at the point they become "exhausted").
1710 : */
1711 1364 : if (so->needPrimScan)
1712 : {
1713 : Assert(_bt_verify_arrays_bt_first(scan, dir));
1714 :
1715 : /*
1716 : * Flag was set -- must call _bt_first again, which will reset the
1717 : * scan's needPrimScan flag
1718 : */
1719 562 : return true;
1720 : }
1721 :
1722 : /* The top-level index scan ran out of tuples in this scan direction */
1723 802 : if (scan->parallel_scan != NULL)
1724 30 : _bt_parallel_done(scan);
1725 :
1726 802 : return false;
1727 : }
1728 :
1729 : /*
1730 : * _bt_advance_array_keys() -- Advance array elements using a tuple
1731 : *
1732 : * The scan always gets a new qual as a consequence of calling here (except
1733 : * when we determine that the top-level scan has run out of matching tuples).
1734 : * All later _bt_check_compare calls also use the same new qual that was first
1735 : * used here (at least until the next call here advances the keys once again).
1736 : * It's convenient to structure _bt_check_compare rechecks of caller's tuple
1737 : * (using the new qual) as one the steps of advancing the scan's array keys,
1738 : * so this function works as a wrapper around _bt_check_compare.
1739 : *
1740 : * Like _bt_check_compare, we'll set pstate.continuescan on behalf of the
1741 : * caller, and return a boolean indicating if caller's tuple satisfies the
1742 : * scan's new qual. But unlike _bt_check_compare, we set so->needPrimScan
1743 : * when we set continuescan=false, indicating if a new primitive index scan
1744 : * has been scheduled (otherwise, the top-level scan has run out of tuples in
1745 : * the current scan direction).
1746 : *
1747 : * Caller must use _bt_tuple_before_array_skeys to determine if the current
1748 : * place in the scan is >= the current array keys _before_ calling here.
1749 : * We're responsible for ensuring that caller's tuple is <= the newly advanced
1750 : * required array keys once we return. We try to find an exact match, but
1751 : * failing that we'll advance the array keys to whatever set of array elements
1752 : * comes next in the key space for the current scan direction. Required array
1753 : * keys "ratchet forwards" (or backwards). They can only advance as the scan
1754 : * itself advances through the index/key space.
1755 : *
1756 : * (The rules are the same for backwards scans, except that the operators are
1757 : * flipped: just replace the precondition's >= operator with a <=, and the
1758 : * postcondition's <= operator with a >=. In other words, just swap the
1759 : * precondition with the postcondition.)
1760 : *
1761 : * We also deal with "advancing" non-required arrays here. Callers whose
1762 : * sktrig scan key is non-required specify sktrig_required=false. These calls
1763 : * are the only exception to the general rule about always advancing the
1764 : * required array keys (the scan may not even have a required array). These
1765 : * callers should just pass a NULL pstate (since there is never any question
1766 : * of stopping the scan). No call to _bt_tuple_before_array_skeys is required
1767 : * ahead of these calls (it's already clear that any required scan keys must
1768 : * be satisfied by caller's tuple).
1769 : *
1770 : * Note that we deal with non-array required equality strategy scan keys as
1771 : * degenerate single element arrays here. Obviously, they can never really
1772 : * advance in the way that real arrays can, but they must still affect how we
1773 : * advance real array scan keys (exactly like true array equality scan keys).
1774 : * We have to keep around a 3-way ORDER proc for these (using the "=" operator
1775 : * won't do), since in general whether the tuple is < or > _any_ unsatisfied
1776 : * required equality key influences how the scan's real arrays must advance.
1777 : *
1778 : * Note also that we may sometimes need to advance the array keys when the
1779 : * existing required array keys (and other required equality keys) are already
1780 : * an exact match for every corresponding value from caller's tuple. We must
1781 : * do this for inequalities that _bt_check_compare set continuescan=false for.
1782 : * They'll advance the array keys here, just like any other scan key that
1783 : * _bt_check_compare stops on. (This can even happen _after_ we advance the
1784 : * array keys, in which case we'll advance the array keys a second time. That
1785 : * way _bt_checkkeys caller always has its required arrays advance to the
1786 : * maximum possible extent that its tuple will allow.)
1787 : */
1788 : static bool
1789 3740 : _bt_advance_array_keys(IndexScanDesc scan, BTReadPageState *pstate,
1790 : IndexTuple tuple, int tupnatts, TupleDesc tupdesc,
1791 : int sktrig, bool sktrig_required)
1792 : {
1793 3740 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
1794 3740 : Relation rel = scan->indexRelation;
1795 3740 : ScanDirection dir = pstate ? pstate->dir : ForwardScanDirection;
1796 3740 : int arrayidx = 0;
1797 3740 : bool beyond_end_advance = false,
1798 3740 : has_required_opposite_direction_only = false,
1799 3740 : oppodir_inequality_sktrig = false,
1800 3740 : all_required_satisfied = true,
1801 3740 : all_satisfied = true;
1802 :
1803 3740 : if (sktrig_required)
1804 : {
1805 : /*
1806 : * Precondition array state assertion
1807 : */
1808 : Assert(!_bt_tuple_before_array_skeys(scan, dir, tuple, tupdesc,
1809 : tupnatts, false, 0, NULL));
1810 :
1811 3500 : so->scanBehind = false; /* reset */
1812 :
1813 : /*
1814 : * Required scan key wasn't satisfied, so required arrays will have to
1815 : * advance. Invalidate page-level state that tracks whether the
1816 : * scan's required-in-opposite-direction-only keys are known to be
1817 : * satisfied by page's remaining tuples.
1818 : */
1819 3500 : pstate->firstmatch = false;
1820 :
1821 : /* Shouldn't have to invalidate 'prechecked', though */
1822 : Assert(!pstate->prechecked);
1823 :
1824 : /*
1825 : * Once we return we'll have a new set of required array keys, so
1826 : * reset state used by "look ahead" optimization
1827 : */
1828 3500 : pstate->rechecks = 0;
1829 3500 : pstate->targetdistance = 0;
1830 : }
1831 :
1832 : Assert(_bt_verify_keys_with_arraykeys(scan));
1833 :
1834 10272 : for (int ikey = 0; ikey < so->numberOfKeys; ikey++)
1835 : {
1836 6748 : ScanKey cur = so->keyData + ikey;
1837 6748 : BTArrayKeyInfo *array = NULL;
1838 : Datum tupdatum;
1839 6748 : bool required = false,
1840 6748 : required_opposite_direction_only = false,
1841 : tupnull;
1842 : int32 result;
1843 6748 : int set_elem = 0;
1844 :
1845 6748 : if (cur->sk_strategy == BTEqualStrategyNumber)
1846 : {
1847 : /* Manage array state */
1848 6484 : if (cur->sk_flags & SK_SEARCHARRAY)
1849 : {
1850 4532 : array = &so->arrayKeys[arrayidx++];
1851 : Assert(array->scan_key == ikey);
1852 : }
1853 : }
1854 : else
1855 : {
1856 : /*
1857 : * Are any inequalities required in the opposite direction only
1858 : * present here?
1859 : */
1860 264 : if (((ScanDirectionIsForward(dir) &&
1861 264 : (cur->sk_flags & (SK_BT_REQBKWD))) ||
1862 0 : (ScanDirectionIsBackward(dir) &&
1863 0 : (cur->sk_flags & (SK_BT_REQFWD)))))
1864 18 : has_required_opposite_direction_only =
1865 18 : required_opposite_direction_only = true;
1866 : }
1867 :
1868 : /* Optimization: skip over known-satisfied scan keys */
1869 6748 : if (ikey < sktrig)
1870 2466 : continue;
1871 :
1872 6206 : if (cur->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD))
1873 : {
1874 : Assert(sktrig_required);
1875 :
1876 4790 : required = true;
1877 :
1878 4790 : if (cur->sk_attno > tupnatts)
1879 : {
1880 : /* Set this just like _bt_tuple_before_array_skeys */
1881 : Assert(sktrig < ikey);
1882 0 : so->scanBehind = true;
1883 : }
1884 : }
1885 :
1886 : /*
1887 : * Handle a required non-array scan key that the initial call to
1888 : * _bt_check_compare indicated triggered array advancement, if any.
1889 : *
1890 : * The non-array scan key's strategy will be <, <=, or = during a
1891 : * forwards scan (or any one of =, >=, or > during a backwards scan).
1892 : * It follows that the corresponding tuple attribute's value must now
1893 : * be either > or >= the scan key value (for backwards scans it must
1894 : * be either < or <= that value).
1895 : *
1896 : * If this is a required equality strategy scan key, this is just an
1897 : * optimization; _bt_tuple_before_array_skeys already confirmed that
1898 : * this scan key places us ahead of caller's tuple. There's no need
1899 : * to repeat that work now. (The same underlying principle also gets
1900 : * applied by the cur_elem_trig optimization used to speed up searches
1901 : * for the next array element.)
1902 : *
1903 : * If this is a required inequality strategy scan key, we _must_ rely
1904 : * on _bt_check_compare like this; we aren't capable of directly
1905 : * evaluating required inequality strategy scan keys here, on our own.
1906 : */
1907 6206 : if (ikey == sktrig && !array)
1908 : {
1909 : Assert(sktrig_required && required && all_required_satisfied);
1910 :
1911 : /* Use "beyond end" advancement. See below for an explanation. */
1912 38 : beyond_end_advance = true;
1913 38 : all_satisfied = all_required_satisfied = false;
1914 :
1915 : /*
1916 : * Set a flag that remembers that this was an inequality required
1917 : * in the opposite scan direction only, that nevertheless
1918 : * triggered the call here.
1919 : *
1920 : * This only happens when an inequality operator (which must be
1921 : * strict) encounters a group of NULLs that indicate the end of
1922 : * non-NULL values for tuples in the current scan direction.
1923 : */
1924 38 : if (unlikely(required_opposite_direction_only))
1925 0 : oppodir_inequality_sktrig = true;
1926 :
1927 38 : continue;
1928 : }
1929 :
1930 : /*
1931 : * Nothing more for us to do with an inequality strategy scan key that
1932 : * wasn't the one that _bt_check_compare stopped on, though.
1933 : *
1934 : * Note: if our later call to _bt_check_compare (to recheck caller's
1935 : * tuple) sets continuescan=false due to finding this same inequality
1936 : * unsatisfied (possible when it's required in the scan direction),
1937 : * we'll deal with it via a recursive "second pass" call.
1938 : */
1939 6168 : else if (cur->sk_strategy != BTEqualStrategyNumber)
1940 18 : continue;
1941 :
1942 : /*
1943 : * Nothing for us to do with an equality strategy scan key that isn't
1944 : * marked required, either -- unless it's a non-required array
1945 : */
1946 6150 : else if (!required && !array)
1947 1170 : continue;
1948 :
1949 : /*
1950 : * Here we perform steps for all array scan keys after a required
1951 : * array scan key whose binary search triggered "beyond end of array
1952 : * element" array advancement due to encountering a tuple attribute
1953 : * value > the closest matching array key (or < for backwards scans).
1954 : */
1955 4980 : if (beyond_end_advance)
1956 : {
1957 : int final_elem_dir;
1958 :
1959 270 : if (ScanDirectionIsBackward(dir) || !array)
1960 102 : final_elem_dir = 0;
1961 : else
1962 168 : final_elem_dir = array->num_elems - 1;
1963 :
1964 270 : if (array && array->cur_elem != final_elem_dir)
1965 : {
1966 48 : array->cur_elem = final_elem_dir;
1967 48 : cur->sk_argument = array->elem_values[final_elem_dir];
1968 : }
1969 :
1970 270 : continue;
1971 : }
1972 :
1973 : /*
1974 : * Here we perform steps for all array scan keys after a required
1975 : * array scan key whose tuple attribute was < the closest matching
1976 : * array key when we dealt with it (or > for backwards scans).
1977 : *
1978 : * This earlier required array key already puts us ahead of caller's
1979 : * tuple in the key space (for the current scan direction). We must
1980 : * make sure that subsequent lower-order array keys do not put us too
1981 : * far ahead (ahead of tuples that have yet to be seen by our caller).
1982 : * For example, when a tuple "(a, b) = (42, 5)" advances the array
1983 : * keys on "a" from 40 to 45, we must also set "b" to whatever the
1984 : * first array element for "b" is. It would be wrong to allow "b" to
1985 : * be set based on the tuple value.
1986 : *
1987 : * Perform the same steps with truncated high key attributes. You can
1988 : * think of this as a "binary search" for the element closest to the
1989 : * value -inf. Again, the arrays must never get ahead of the scan.
1990 : */
1991 4710 : if (!all_required_satisfied || cur->sk_attno > tupnatts)
1992 : {
1993 : int first_elem_dir;
1994 :
1995 428 : if (ScanDirectionIsForward(dir) || !array)
1996 428 : first_elem_dir = 0;
1997 : else
1998 0 : first_elem_dir = array->num_elems - 1;
1999 :
2000 428 : if (array && array->cur_elem != first_elem_dir)
2001 : {
2002 180 : array->cur_elem = first_elem_dir;
2003 180 : cur->sk_argument = array->elem_values[first_elem_dir];
2004 : }
2005 :
2006 428 : continue;
2007 : }
2008 :
2009 : /*
2010 : * Search in scankey's array for the corresponding tuple attribute
2011 : * value from caller's tuple
2012 : */
2013 4282 : tupdatum = index_getattr(tuple, cur->sk_attno, tupdesc, &tupnull);
2014 :
2015 4282 : if (array)
2016 : {
2017 3918 : bool cur_elem_trig = (sktrig_required && ikey == sktrig);
2018 :
2019 : /*
2020 : * Binary search for closest match that's available from the array
2021 : */
2022 3918 : set_elem = _bt_binsrch_array_skey(&so->orderProcs[ikey],
2023 : cur_elem_trig, dir,
2024 : tupdatum, tupnull, array, cur,
2025 : &result);
2026 :
2027 : Assert(set_elem >= 0 && set_elem < array->num_elems);
2028 : }
2029 : else
2030 : {
2031 : Assert(sktrig_required && required);
2032 :
2033 : /*
2034 : * This is a required non-array equality strategy scan key, which
2035 : * we'll treat as a degenerate single element array.
2036 : *
2037 : * This scan key's imaginary "array" can't really advance, but it
2038 : * can still roll over like any other array. (Actually, this is
2039 : * no different to real single value arrays, which never advance
2040 : * without rolling over -- they can never truly advance, either.)
2041 : */
2042 364 : result = _bt_compare_array_skey(&so->orderProcs[ikey],
2043 : tupdatum, tupnull,
2044 : cur->sk_argument, cur);
2045 : }
2046 :
2047 : /*
2048 : * Consider "beyond end of array element" array advancement.
2049 : *
2050 : * When the tuple attribute value is > the closest matching array key
2051 : * (or < in the backwards scan case), we need to ratchet this array
2052 : * forward (backward) by one increment, so that caller's tuple ends up
2053 : * being < final array value instead (or > final array value instead).
2054 : * This process has to work for all of the arrays, not just this one:
2055 : * it must "carry" to higher-order arrays when the set_elem that we
2056 : * just found happens to be the final one for the scan's direction.
2057 : * Incrementing (decrementing) set_elem itself isn't good enough.
2058 : *
2059 : * Our approach is to provisionally use set_elem as if it was an exact
2060 : * match now, then set each later/less significant array to whatever
2061 : * its final element is. Once outside the loop we'll then "increment
2062 : * this array's set_elem" by calling _bt_advance_array_keys_increment.
2063 : * That way the process rolls over to higher order arrays as needed.
2064 : *
2065 : * Under this scheme any required arrays only ever ratchet forwards
2066 : * (or backwards), and always do so to the maximum possible extent
2067 : * that we can know will be safe without seeing the scan's next tuple.
2068 : * We don't need any special handling for required scan keys that lack
2069 : * a real array to advance, nor for redundant scan keys that couldn't
2070 : * be eliminated by _bt_preprocess_keys. It won't matter if some of
2071 : * our "true" array scan keys (or even all of them) are non-required.
2072 : */
2073 4282 : if (required &&
2074 4042 : ((ScanDirectionIsForward(dir) && result > 0) ||
2075 24 : (ScanDirectionIsBackward(dir) && result < 0)))
2076 594 : beyond_end_advance = true;
2077 :
2078 : Assert(all_required_satisfied && all_satisfied);
2079 4282 : if (result != 0)
2080 : {
2081 : /*
2082 : * Track whether caller's tuple satisfies our new post-advancement
2083 : * qual, for required scan keys, as well as for the entire set of
2084 : * interesting scan keys (all required scan keys plus non-required
2085 : * array scan keys are considered interesting.)
2086 : */
2087 1890 : all_satisfied = false;
2088 1890 : if (required)
2089 1674 : all_required_satisfied = false;
2090 : else
2091 : {
2092 : /*
2093 : * There's no need to advance the arrays using the best
2094 : * available match for a non-required array. Give up now.
2095 : * (Though note that sktrig_required calls still have to do
2096 : * all the usual post-advancement steps, including the recheck
2097 : * call to _bt_check_compare.)
2098 : */
2099 216 : break;
2100 : }
2101 : }
2102 :
2103 : /* Advance array keys, even when set_elem isn't an exact match */
2104 4066 : if (array && array->cur_elem != set_elem)
2105 : {
2106 3108 : array->cur_elem = set_elem;
2107 3108 : cur->sk_argument = array->elem_values[set_elem];
2108 : }
2109 : }
2110 :
2111 : /*
2112 : * Advance the array keys incrementally whenever "beyond end of array
2113 : * element" array advancement happens, so that advancement will carry to
2114 : * higher-order arrays (might exhaust all the scan's arrays instead, which
2115 : * ends the top-level scan).
2116 : */
2117 3740 : if (beyond_end_advance && !_bt_advance_array_keys_increment(scan, dir))
2118 618 : goto end_toplevel_scan;
2119 :
2120 : Assert(_bt_verify_keys_with_arraykeys(scan));
2121 :
2122 : /*
2123 : * Does tuple now satisfy our new qual? Recheck with _bt_check_compare.
2124 : *
2125 : * Calls triggered by an unsatisfied required scan key, whose tuple now
2126 : * satisfies all required scan keys, but not all nonrequired array keys,
2127 : * will still require a recheck call to _bt_check_compare. They'll still
2128 : * need its "second pass" handling of required inequality scan keys.
2129 : * (Might have missed a still-unsatisfied required inequality scan key
2130 : * that caller didn't detect as the sktrig scan key during its initial
2131 : * _bt_check_compare call that used the old/original qual.)
2132 : *
2133 : * Calls triggered by an unsatisfied nonrequired array scan key never need
2134 : * "second pass" handling of required inequalities (nor any other handling
2135 : * of any required scan key). All that matters is whether caller's tuple
2136 : * satisfies the new qual, so it's safe to just skip the _bt_check_compare
2137 : * recheck when we've already determined that it can only return 'false'.
2138 : */
2139 3122 : if ((sktrig_required && all_required_satisfied) ||
2140 1334 : (!sktrig_required && all_satisfied))
2141 : {
2142 1812 : int nsktrig = sktrig + 1;
2143 : bool continuescan;
2144 :
2145 : Assert(all_required_satisfied);
2146 :
2147 : /* Recheck _bt_check_compare on behalf of caller */
2148 1812 : if (_bt_check_compare(scan, dir, tuple, tupnatts, tupdesc,
2149 : false, false, false,
2150 1806 : &continuescan, &nsktrig) &&
2151 1806 : !so->scanBehind)
2152 : {
2153 : /* This tuple satisfies the new qual */
2154 : Assert(all_satisfied && continuescan);
2155 :
2156 1806 : if (pstate)
2157 1782 : pstate->continuescan = true;
2158 :
2159 1806 : return true;
2160 : }
2161 :
2162 : /*
2163 : * Consider "second pass" handling of required inequalities.
2164 : *
2165 : * It's possible that our _bt_check_compare call indicated that the
2166 : * scan should end due to some unsatisfied inequality that wasn't
2167 : * initially recognized as such by us. Handle this by calling
2168 : * ourselves recursively, this time indicating that the trigger is the
2169 : * inequality that we missed first time around (and using a set of
2170 : * required array/equality keys that are now exact matches for tuple).
2171 : *
2172 : * We make a strong, general guarantee that every _bt_checkkeys call
2173 : * here will advance the array keys to the maximum possible extent
2174 : * that we can know to be safe based on caller's tuple alone. If we
2175 : * didn't perform this step, then that guarantee wouldn't quite hold.
2176 : */
2177 6 : if (unlikely(!continuescan))
2178 : {
2179 : bool satisfied PG_USED_FOR_ASSERTS_ONLY;
2180 :
2181 : Assert(sktrig_required);
2182 : Assert(so->keyData[nsktrig].sk_strategy != BTEqualStrategyNumber);
2183 :
2184 : /*
2185 : * The tuple must use "beyond end" advancement during the
2186 : * recursive call, so we cannot possibly end up back here when
2187 : * recursing. We'll consume a small, fixed amount of stack space.
2188 : */
2189 : Assert(!beyond_end_advance);
2190 :
2191 : /* Advance the array keys a second time using same tuple */
2192 0 : satisfied = _bt_advance_array_keys(scan, pstate, tuple, tupnatts,
2193 : tupdesc, nsktrig, true);
2194 :
2195 : /* This tuple doesn't satisfy the inequality */
2196 : Assert(!satisfied);
2197 0 : return false;
2198 : }
2199 :
2200 : /*
2201 : * Some non-required scan key (from new qual) still not satisfied.
2202 : *
2203 : * All scan keys required in the current scan direction must still be
2204 : * satisfied, though, so we can trust all_required_satisfied below.
2205 : */
2206 : }
2207 :
2208 : /*
2209 : * When we were called just to deal with "advancing" non-required arrays,
2210 : * this is as far as we can go (cannot stop the scan for these callers)
2211 : */
2212 1316 : if (!sktrig_required)
2213 : {
2214 : /* Caller's tuple doesn't match any qual */
2215 216 : return false;
2216 : }
2217 :
2218 : /*
2219 : * Postcondition array state assertion (for still-unsatisfied tuples).
2220 : *
2221 : * By here we have established that the scan's required arrays (scan must
2222 : * have at least one required array) advanced, without becoming exhausted.
2223 : *
2224 : * Caller's tuple is now < the newly advanced array keys (or > when this
2225 : * is a backwards scan), except in the case where we only got this far due
2226 : * to an unsatisfied non-required scan key. Verify that with an assert.
2227 : *
2228 : * Note: we don't just quit at this point when all required scan keys were
2229 : * found to be satisfied because we need to consider edge-cases involving
2230 : * scan keys required in the opposite direction only; those aren't tracked
2231 : * by all_required_satisfied. (Actually, oppodir_inequality_sktrig trigger
2232 : * scan keys are tracked by all_required_satisfied, since it's convenient
2233 : * for _bt_check_compare to behave as if they are required in the current
2234 : * scan direction to deal with NULLs. We'll account for that separately.)
2235 : */
2236 : Assert(_bt_tuple_before_array_skeys(scan, dir, tuple, tupdesc, tupnatts,
2237 : false, 0, NULL) ==
2238 : !all_required_satisfied);
2239 :
2240 : /*
2241 : * We generally permit primitive index scans to continue onto the next
2242 : * sibling page when the page's finaltup satisfies all required scan keys
2243 : * at the point where we're between pages.
2244 : *
2245 : * If caller's tuple is also the page's finaltup, and we see that required
2246 : * scan keys still aren't satisfied, start a new primitive index scan.
2247 : */
2248 1100 : if (!all_required_satisfied && pstate->finaltup == tuple)
2249 0 : goto new_prim_scan;
2250 :
2251 : /*
2252 : * Proactively check finaltup (don't wait until finaltup is reached by the
2253 : * scan) when it might well turn out to not be satisfied later on.
2254 : *
2255 : * Note: if so->scanBehind hasn't already been set for finaltup by us,
2256 : * it'll be set during this call to _bt_tuple_before_array_skeys. Either
2257 : * way, it'll be set correctly (for the whole page) after this point.
2258 : */
2259 2140 : if (!all_required_satisfied && pstate->finaltup &&
2260 2080 : _bt_tuple_before_array_skeys(scan, dir, pstate->finaltup, tupdesc,
2261 2080 : BTreeTupleGetNAtts(pstate->finaltup, rel),
2262 : false, 0, &so->scanBehind))
2263 562 : goto new_prim_scan;
2264 :
2265 : /*
2266 : * When we encounter a truncated finaltup high key attribute, we're
2267 : * optimistic about the chances of its corresponding required scan key
2268 : * being satisfied when we go on to check it against tuples from this
2269 : * page's right sibling leaf page. We consider truncated attributes to be
2270 : * satisfied by required scan keys, which allows the primitive index scan
2271 : * to continue to the next leaf page. We must set so->scanBehind to true
2272 : * to remember that the last page's finaltup had "satisfied" required scan
2273 : * keys for one or more truncated attribute values (scan keys required in
2274 : * _either_ scan direction).
2275 : *
2276 : * There is a chance that _bt_checkkeys (which checks so->scanBehind) will
2277 : * find that even the sibling leaf page's finaltup is < the new array
2278 : * keys. When that happens, our optimistic policy will have incurred a
2279 : * single extra leaf page access that could have been avoided.
2280 : *
2281 : * A pessimistic policy would give backward scans a gratuitous advantage
2282 : * over forward scans. We'd punish forward scans for applying more
2283 : * accurate information from the high key, rather than just using the
2284 : * final non-pivot tuple as finaltup, in the style of backward scans.
2285 : * Being pessimistic would also give some scans with non-required arrays a
2286 : * perverse advantage over similar scans that use required arrays instead.
2287 : *
2288 : * You can think of this as a speculative bet on what the scan is likely
2289 : * to find on the next page. It's not much of a gamble, though, since the
2290 : * untruncated prefix of attributes must strictly satisfy the new qual
2291 : * (though it's okay if any non-required scan keys fail to be satisfied).
2292 : */
2293 538 : if (so->scanBehind && has_required_opposite_direction_only)
2294 : {
2295 : /*
2296 : * However, we avoid this behavior whenever the scan involves a scan
2297 : * key required in the opposite direction to the scan only, along with
2298 : * a finaltup with at least one truncated attribute that's associated
2299 : * with a scan key marked required (required in either direction).
2300 : *
2301 : * _bt_check_compare simply won't stop the scan for a scan key that's
2302 : * marked required in the opposite scan direction only. That leaves
2303 : * us without any reliable way of reconsidering any opposite-direction
2304 : * inequalities if it turns out that starting a new primitive index
2305 : * scan will allow _bt_first to skip ahead by a great many leaf pages
2306 : * (see next section for details of how that works).
2307 : */
2308 0 : goto new_prim_scan;
2309 : }
2310 :
2311 : /*
2312 : * Handle inequalities marked required in the opposite scan direction.
2313 : * They can also signal that we should start a new primitive index scan.
2314 : *
2315 : * It's possible that the scan is now positioned where "matching" tuples
2316 : * begin, and that caller's tuple satisfies all scan keys required in the
2317 : * current scan direction. But if caller's tuple still doesn't satisfy
2318 : * other scan keys that are required in the opposite scan direction only
2319 : * (e.g., a required >= strategy scan key when scan direction is forward),
2320 : * it's still possible that there are many leaf pages before the page that
2321 : * _bt_first could skip straight to. Groveling through all those pages
2322 : * will always give correct answers, but it can be very inefficient. We
2323 : * must avoid needlessly scanning extra pages.
2324 : *
2325 : * Separately, it's possible that _bt_check_compare set continuescan=false
2326 : * for a scan key that's required in the opposite direction only. This is
2327 : * a special case, that happens only when _bt_check_compare sees that the
2328 : * inequality encountered a NULL value. This signals the end of non-NULL
2329 : * values in the current scan direction, which is reason enough to end the
2330 : * (primitive) scan. If this happens at the start of a large group of
2331 : * NULL values, then we shouldn't expect to be called again until after
2332 : * the scan has already read indefinitely-many leaf pages full of tuples
2333 : * with NULL suffix values. We need a separate test for this case so that
2334 : * we don't miss our only opportunity to skip over such a group of pages.
2335 : * (_bt_first is expected to skip over the group of NULLs by applying a
2336 : * similar "deduce NOT NULL" rule, where it finishes its insertion scan
2337 : * key by consing up an explicit SK_SEARCHNOTNULL key.)
2338 : *
2339 : * Apply a test against finaltup to detect and recover from these problem:
2340 : * if even finaltup doesn't satisfy such an inequality, we just skip by
2341 : * starting a new primitive index scan. When we skip, we know for sure
2342 : * that all of the tuples on the current page following caller's tuple are
2343 : * also before the _bt_first-wise start of tuples for our new qual. That
2344 : * at least suggests many more skippable pages beyond the current page.
2345 : */
2346 538 : if (has_required_opposite_direction_only && pstate->finaltup &&
2347 0 : (all_required_satisfied || oppodir_inequality_sktrig))
2348 : {
2349 0 : int nfinaltupatts = BTreeTupleGetNAtts(pstate->finaltup, rel);
2350 : ScanDirection flipped;
2351 : bool continuescanflip;
2352 : int opsktrig;
2353 :
2354 : /*
2355 : * We're checking finaltup (which is usually not caller's tuple), so
2356 : * cannot reuse work from caller's earlier _bt_check_compare call.
2357 : *
2358 : * Flip the scan direction when calling _bt_check_compare this time,
2359 : * so that it will set continuescanflip=false when it encounters an
2360 : * inequality required in the opposite scan direction.
2361 : */
2362 : Assert(!so->scanBehind);
2363 0 : opsktrig = 0;
2364 0 : flipped = -dir;
2365 0 : _bt_check_compare(scan, flipped,
2366 : pstate->finaltup, nfinaltupatts, tupdesc,
2367 : false, false, false,
2368 : &continuescanflip, &opsktrig);
2369 :
2370 : /*
2371 : * If we ended up here due to the all_required_satisfied criteria,
2372 : * test opsktrig in a way that ensures that finaltup contains the same
2373 : * prefix of key columns as caller's tuple (a prefix that satisfies
2374 : * earlier required-in-current-direction scan keys).
2375 : *
2376 : * If we ended up here due to the oppodir_inequality_sktrig criteria,
2377 : * test opsktrig in a way that ensures that the same scan key that our
2378 : * caller found to be unsatisfied (by the scan's tuple) was also the
2379 : * one unsatisfied just now (by finaltup). That way we'll only start
2380 : * a new primitive scan when we're sure that both tuples _don't_ share
2381 : * the same prefix of satisfied equality-constrained attribute values,
2382 : * and that finaltup has a non-NULL attribute value indicated by the
2383 : * unsatisfied scan key at offset opsktrig/sktrig. (This depends on
2384 : * _bt_check_compare not caring about the direction that inequalities
2385 : * are required in whenever NULL attribute values are unsatisfied. It
2386 : * only cares about the scan direction, and its relationship to
2387 : * whether NULLs are stored first or last relative to non-NULLs.)
2388 : */
2389 : Assert(all_required_satisfied != oppodir_inequality_sktrig);
2390 0 : if (unlikely(!continuescanflip &&
2391 : ((all_required_satisfied && opsktrig > sktrig) ||
2392 : (oppodir_inequality_sktrig && opsktrig >= sktrig))))
2393 : {
2394 : Assert(so->keyData[opsktrig].sk_strategy != BTEqualStrategyNumber);
2395 :
2396 : /*
2397 : * Make sure that any non-required arrays are set to the first
2398 : * array element for the current scan direction
2399 : */
2400 0 : _bt_rewind_nonrequired_arrays(scan, dir);
2401 :
2402 0 : goto new_prim_scan;
2403 : }
2404 : }
2405 :
2406 : /*
2407 : * Stick with the ongoing primitive index scan for now.
2408 : *
2409 : * It's possible that later tuples will also turn out to have values that
2410 : * are still < the now-current array keys (or > the current array keys).
2411 : * Our caller will handle this by performing what amounts to a linear
2412 : * search of the page, implemented by calling _bt_check_compare and then
2413 : * _bt_tuple_before_array_skeys for each tuple.
2414 : *
2415 : * This approach has various advantages over a binary search of the page.
2416 : * Repeated binary searches of the page (one binary search for every array
2417 : * advancement) won't outperform a continuous linear search. While there
2418 : * are workloads that a naive linear search won't handle well, our caller
2419 : * has a "look ahead" fallback mechanism to deal with that problem.
2420 : */
2421 538 : pstate->continuescan = true; /* Override _bt_check_compare */
2422 538 : so->needPrimScan = false; /* _bt_readpage has more tuples to check */
2423 :
2424 538 : if (so->scanBehind)
2425 : {
2426 : /* Optimization: skip by setting "look ahead" mechanism's offnum */
2427 : Assert(ScanDirectionIsForward(dir));
2428 6 : pstate->skip = pstate->maxoff + 1;
2429 : }
2430 :
2431 : /* Caller's tuple doesn't match the new qual */
2432 538 : return false;
2433 :
2434 562 : new_prim_scan:
2435 :
2436 : /*
2437 : * End this primitive index scan, but schedule another.
2438 : *
2439 : * Note: If the scan direction happens to change, this scheduled primitive
2440 : * index scan won't go ahead after all.
2441 : */
2442 562 : pstate->continuescan = false; /* Tell _bt_readpage we're done... */
2443 562 : so->needPrimScan = true; /* ...but call _bt_first again */
2444 :
2445 562 : if (scan->parallel_scan)
2446 36 : _bt_parallel_primscan_schedule(scan, pstate->prev_scan_page);
2447 :
2448 : /* Caller's tuple doesn't match the new qual */
2449 562 : return false;
2450 :
2451 618 : end_toplevel_scan:
2452 :
2453 : /*
2454 : * End the current primitive index scan, but don't schedule another.
2455 : *
2456 : * This ends the entire top-level scan in the current scan direction.
2457 : *
2458 : * Note: The scan's arrays (including any non-required arrays) are now in
2459 : * their final positions for the current scan direction. If the scan
2460 : * direction happens to change, then the arrays will already be in their
2461 : * first positions for what will then be the current scan direction.
2462 : */
2463 618 : pstate->continuescan = false; /* Tell _bt_readpage we're done... */
2464 618 : so->needPrimScan = false; /* ...don't call _bt_first again, though */
2465 :
2466 : /* Caller's tuple doesn't match any qual */
2467 618 : return false;
2468 : }
2469 :
2470 : /*
2471 : * _bt_preprocess_keys() -- Preprocess scan keys
2472 : *
2473 : * The given search-type keys (taken from scan->keyData[])
2474 : * are copied to so->keyData[] with possible transformation.
2475 : * scan->numberOfKeys is the number of input keys, so->numberOfKeys gets
2476 : * the number of output keys (possibly less, never greater).
2477 : *
2478 : * The output keys are marked with additional sk_flags bits beyond the
2479 : * system-standard bits supplied by the caller. The DESC and NULLS_FIRST
2480 : * indoption bits for the relevant index attribute are copied into the flags.
2481 : * Also, for a DESC column, we commute (flip) all the sk_strategy numbers
2482 : * so that the index sorts in the desired direction.
2483 : *
2484 : * One key purpose of this routine is to discover which scan keys must be
2485 : * satisfied to continue the scan. It also attempts to eliminate redundant
2486 : * keys and detect contradictory keys. (If the index opfamily provides
2487 : * incomplete sets of cross-type operators, we may fail to detect redundant
2488 : * or contradictory keys, but we can survive that.)
2489 : *
2490 : * The output keys must be sorted by index attribute. Presently we expect
2491 : * (but verify) that the input keys are already so sorted --- this is done
2492 : * by match_clauses_to_index() in indxpath.c. Some reordering of the keys
2493 : * within each attribute may be done as a byproduct of the processing here.
2494 : * That process must leave array scan keys (within an attribute) in the same
2495 : * order as corresponding entries from the scan's BTArrayKeyInfo array info.
2496 : *
2497 : * The output keys are marked with flags SK_BT_REQFWD and/or SK_BT_REQBKWD
2498 : * if they must be satisfied in order to continue the scan forward or backward
2499 : * respectively. _bt_checkkeys uses these flags. For example, if the quals
2500 : * are "x = 1 AND y < 4 AND z < 5", then _bt_checkkeys will reject a tuple
2501 : * (1,2,7), but we must continue the scan in case there are tuples (1,3,z).
2502 : * But once we reach tuples like (1,4,z) we can stop scanning because no
2503 : * later tuples could match. This is reflected by marking the x and y keys,
2504 : * but not the z key, with SK_BT_REQFWD. In general, the keys for leading
2505 : * attributes with "=" keys are marked both SK_BT_REQFWD and SK_BT_REQBKWD.
2506 : * For the first attribute without an "=" key, any "<" and "<=" keys are
2507 : * marked SK_BT_REQFWD while any ">" and ">=" keys are marked SK_BT_REQBKWD.
2508 : * This can be seen to be correct by considering the above example. Note
2509 : * in particular that if there are no keys for a given attribute, the keys for
2510 : * subsequent attributes can never be required; for instance "WHERE y = 4"
2511 : * requires a full-index scan.
2512 : *
2513 : * If possible, redundant keys are eliminated: we keep only the tightest
2514 : * >/>= bound and the tightest </<= bound, and if there's an = key then
2515 : * that's the only one returned. (So, we return either a single = key,
2516 : * or one or two boundary-condition keys for each attr.) However, if we
2517 : * cannot compare two keys for lack of a suitable cross-type operator,
2518 : * we cannot eliminate either. If there are two such keys of the same
2519 : * operator strategy, the second one is just pushed into the output array
2520 : * without further processing here. We may also emit both >/>= or both
2521 : * </<= keys if we can't compare them. The logic about required keys still
2522 : * works if we don't eliminate redundant keys.
2523 : *
2524 : * Note that one reason we need direction-sensitive required-key flags is
2525 : * precisely that we may not be able to eliminate redundant keys. Suppose
2526 : * we have "x > 4::int AND x > 10::bigint", and we are unable to determine
2527 : * which key is more restrictive for lack of a suitable cross-type operator.
2528 : * _bt_first will arbitrarily pick one of the keys to do the initial
2529 : * positioning with. If it picks x > 4, then the x > 10 condition will fail
2530 : * until we reach index entries > 10; but we can't stop the scan just because
2531 : * x > 10 is failing. On the other hand, if we are scanning backwards, then
2532 : * failure of either key is indeed enough to stop the scan. (In general, when
2533 : * inequality keys are present, the initial-positioning code only promises to
2534 : * position before the first possible match, not exactly at the first match,
2535 : * for a forward scan; or after the last match for a backward scan.)
2536 : *
2537 : * As a byproduct of this work, we can detect contradictory quals such
2538 : * as "x = 1 AND x > 2". If we see that, we return so->qual_ok = false,
2539 : * indicating the scan need not be run at all since no tuples can match.
2540 : * (In this case we do not bother completing the output key array!)
2541 : * Again, missing cross-type operators might cause us to fail to prove the
2542 : * quals contradictory when they really are, but the scan will work correctly.
2543 : *
2544 : * Row comparison keys are currently also treated without any smarts:
2545 : * we just transfer them into the preprocessed array without any
2546 : * editorialization. We can treat them the same as an ordinary inequality
2547 : * comparison on the row's first index column, for the purposes of the logic
2548 : * about required keys.
2549 : *
2550 : * Note: the reason we have to copy the preprocessed scan keys into private
2551 : * storage is that we are modifying the array based on comparisons of the
2552 : * key argument values, which could change on a rescan. Therefore we can't
2553 : * overwrite the source data.
2554 : */
2555 : void
2556 11830462 : _bt_preprocess_keys(IndexScanDesc scan)
2557 : {
2558 11830462 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
2559 11830462 : int numberOfKeys = scan->numberOfKeys;
2560 11830462 : int16 *indoption = scan->indexRelation->rd_indoption;
2561 : int new_numberOfKeys;
2562 : int numberOfEqualCols;
2563 : ScanKey inkeys;
2564 : ScanKey outkeys;
2565 : ScanKey cur;
2566 : BTScanKeyPreproc xform[BTMaxStrategyNumber];
2567 : bool test_result;
2568 : int i,
2569 : j;
2570 : AttrNumber attno;
2571 : ScanKey arrayKeyData;
2572 11830462 : int *keyDataMap = NULL;
2573 11830462 : int arrayidx = 0;
2574 :
2575 11830462 : if (so->numberOfKeys > 0)
2576 : {
2577 : /*
2578 : * Only need to do preprocessing once per btrescan, at most. All
2579 : * calls after the first are handled as no-ops.
2580 : *
2581 : * If there are array scan keys in so->keyData[], then the now-current
2582 : * array elements must already be present in each array's scan key.
2583 : * Verify that that happened using an assertion.
2584 : */
2585 : Assert(_bt_verify_keys_with_arraykeys(scan));
2586 5859096 : return;
2587 : }
2588 :
2589 : /* initialize result variables */
2590 11829900 : so->qual_ok = true;
2591 11829900 : so->numberOfKeys = 0;
2592 :
2593 11829900 : if (numberOfKeys < 1)
2594 10686 : return; /* done if qual-less scan */
2595 :
2596 : /* If any keys are SK_SEARCHARRAY type, set up array-key info */
2597 11819214 : arrayKeyData = _bt_preprocess_array_keys(scan);
2598 11819214 : if (!so->qual_ok)
2599 : {
2600 : /* unmatchable array, so give up */
2601 6 : return;
2602 : }
2603 :
2604 : /*
2605 : * Treat arrayKeyData[] (a partially preprocessed copy of scan->keyData[])
2606 : * as our input if _bt_preprocess_array_keys just allocated it, else just
2607 : * use scan->keyData[]
2608 : */
2609 11819208 : if (arrayKeyData)
2610 : {
2611 826 : inkeys = arrayKeyData;
2612 :
2613 : /* Also maintain keyDataMap for remapping so->orderProc[] later */
2614 826 : keyDataMap = MemoryContextAlloc(so->arrayContext,
2615 : numberOfKeys * sizeof(int));
2616 : }
2617 : else
2618 11818382 : inkeys = scan->keyData;
2619 :
2620 11819208 : outkeys = so->keyData;
2621 11819208 : cur = &inkeys[0];
2622 : /* we check that input keys are correctly ordered */
2623 11819208 : if (cur->sk_attno < 1)
2624 0 : elog(ERROR, "btree index keys must be ordered by attribute");
2625 :
2626 : /* We can short-circuit most of the work if there's just one key */
2627 11819208 : if (numberOfKeys == 1)
2628 : {
2629 : /* Apply indoption to scankey (might change sk_strategy!) */
2630 5847782 : if (!_bt_fix_scankey_strategy(cur, indoption))
2631 2642 : so->qual_ok = false;
2632 5847782 : memcpy(outkeys, cur, sizeof(ScanKeyData));
2633 5847782 : so->numberOfKeys = 1;
2634 : /* We can mark the qual as required if it's for first index col */
2635 5847782 : if (cur->sk_attno == 1)
2636 5845126 : _bt_mark_scankey_required(outkeys);
2637 : if (arrayKeyData)
2638 : {
2639 : /*
2640 : * Don't call _bt_preprocess_array_keys_final in this fast path
2641 : * (we'll miss out on the single value array transformation, but
2642 : * that's not nearly as important when there's only one scan key)
2643 : */
2644 : Assert(cur->sk_flags & SK_SEARCHARRAY);
2645 : Assert(cur->sk_strategy != BTEqualStrategyNumber ||
2646 : (so->arrayKeys[0].scan_key == 0 &&
2647 : OidIsValid(so->orderProcs[0].fn_oid)));
2648 : }
2649 :
2650 5847782 : return;
2651 : }
2652 :
2653 : /*
2654 : * Otherwise, do the full set of pushups.
2655 : */
2656 5971426 : new_numberOfKeys = 0;
2657 5971426 : numberOfEqualCols = 0;
2658 :
2659 : /*
2660 : * Initialize for processing of keys for attr 1.
2661 : *
2662 : * xform[i] points to the currently best scan key of strategy type i+1; it
2663 : * is NULL if we haven't yet found such a key for this attr.
2664 : */
2665 5971426 : attno = 1;
2666 5971426 : memset(xform, 0, sizeof(xform));
2667 :
2668 : /*
2669 : * Loop iterates from 0 to numberOfKeys inclusive; we use the last pass to
2670 : * handle after-last-key processing. Actual exit from the loop is at the
2671 : * "break" statement below.
2672 : */
2673 19401896 : for (i = 0;; cur++, i++)
2674 : {
2675 19401896 : if (i < numberOfKeys)
2676 : {
2677 : /* Apply indoption to scankey (might change sk_strategy!) */
2678 13430494 : if (!_bt_fix_scankey_strategy(cur, indoption))
2679 : {
2680 : /* NULL can't be matched, so give up */
2681 18 : so->qual_ok = false;
2682 18 : return;
2683 : }
2684 : }
2685 :
2686 : /*
2687 : * If we are at the end of the keys for a particular attr, finish up
2688 : * processing and emit the cleaned-up keys.
2689 : */
2690 19401878 : if (i == numberOfKeys || cur->sk_attno != attno)
2691 : {
2692 13428480 : int priorNumberOfEqualCols = numberOfEqualCols;
2693 :
2694 : /* check input keys are correctly ordered */
2695 13428480 : if (i < numberOfKeys && cur->sk_attno < attno)
2696 0 : elog(ERROR, "btree index keys must be ordered by attribute");
2697 :
2698 : /*
2699 : * If = has been specified, all other keys can be eliminated as
2700 : * redundant. If we have a case like key = 1 AND key > 2, we can
2701 : * set qual_ok to false and abandon further processing.
2702 : *
2703 : * We also have to deal with the case of "key IS NULL", which is
2704 : * unsatisfiable in combination with any other index condition. By
2705 : * the time we get here, that's been classified as an equality
2706 : * check, and we've rejected any combination of it with a regular
2707 : * equality condition; but not with other types of conditions.
2708 : */
2709 13428480 : if (xform[BTEqualStrategyNumber - 1].skey)
2710 : {
2711 12383460 : ScanKey eq = xform[BTEqualStrategyNumber - 1].skey;
2712 12383460 : BTArrayKeyInfo *array = NULL;
2713 12383460 : FmgrInfo *orderproc = NULL;
2714 :
2715 12383460 : if (arrayKeyData && (eq->sk_flags & SK_SEARCHARRAY))
2716 : {
2717 : int eq_in_ikey,
2718 : eq_arrayidx;
2719 :
2720 534 : eq_in_ikey = xform[BTEqualStrategyNumber - 1].ikey;
2721 534 : eq_arrayidx = xform[BTEqualStrategyNumber - 1].arrayidx;
2722 534 : array = &so->arrayKeys[eq_arrayidx - 1];
2723 534 : orderproc = so->orderProcs + eq_in_ikey;
2724 :
2725 : Assert(array->scan_key == eq_in_ikey);
2726 : Assert(OidIsValid(orderproc->fn_oid));
2727 : }
2728 :
2729 74300604 : for (j = BTMaxStrategyNumber; --j >= 0;)
2730 : {
2731 61917180 : ScanKey chk = xform[j].skey;
2732 :
2733 61917180 : if (!chk || j == (BTEqualStrategyNumber - 1))
2734 61917070 : continue;
2735 :
2736 110 : if (eq->sk_flags & SK_SEARCHNULL)
2737 : {
2738 : /* IS NULL is contradictory to anything else */
2739 24 : so->qual_ok = false;
2740 24 : return;
2741 : }
2742 :
2743 86 : if (_bt_compare_scankey_args(scan, chk, eq, chk,
2744 : array, orderproc,
2745 : &test_result))
2746 : {
2747 86 : if (!test_result)
2748 : {
2749 : /* keys proven mutually contradictory */
2750 12 : so->qual_ok = false;
2751 12 : return;
2752 : }
2753 : /* else discard the redundant non-equality key */
2754 : Assert(!array || array->num_elems > 0);
2755 74 : xform[j].skey = NULL;
2756 74 : xform[j].ikey = -1;
2757 : }
2758 : /* else, cannot determine redundancy, keep both keys */
2759 : }
2760 : /* track number of attrs for which we have "=" keys */
2761 12383424 : numberOfEqualCols++;
2762 : }
2763 :
2764 : /* try to keep only one of <, <= */
2765 13428444 : if (xform[BTLessStrategyNumber - 1].skey
2766 1964 : && xform[BTLessEqualStrategyNumber - 1].skey)
2767 : {
2768 6 : ScanKey lt = xform[BTLessStrategyNumber - 1].skey;
2769 6 : ScanKey le = xform[BTLessEqualStrategyNumber - 1].skey;
2770 :
2771 6 : if (_bt_compare_scankey_args(scan, le, lt, le, NULL, NULL,
2772 : &test_result))
2773 : {
2774 6 : if (test_result)
2775 6 : xform[BTLessEqualStrategyNumber - 1].skey = NULL;
2776 : else
2777 0 : xform[BTLessStrategyNumber - 1].skey = NULL;
2778 : }
2779 : }
2780 :
2781 : /* try to keep only one of >, >= */
2782 13428444 : if (xform[BTGreaterStrategyNumber - 1].skey
2783 1040678 : && xform[BTGreaterEqualStrategyNumber - 1].skey)
2784 : {
2785 6 : ScanKey gt = xform[BTGreaterStrategyNumber - 1].skey;
2786 6 : ScanKey ge = xform[BTGreaterEqualStrategyNumber - 1].skey;
2787 :
2788 6 : if (_bt_compare_scankey_args(scan, ge, gt, ge, NULL, NULL,
2789 : &test_result))
2790 : {
2791 6 : if (test_result)
2792 0 : xform[BTGreaterEqualStrategyNumber - 1].skey = NULL;
2793 : else
2794 6 : xform[BTGreaterStrategyNumber - 1].skey = NULL;
2795 : }
2796 : }
2797 :
2798 : /*
2799 : * Emit the cleaned-up keys into the outkeys[] array, and then
2800 : * mark them if they are required. They are required (possibly
2801 : * only in one direction) if all attrs before this one had "=".
2802 : */
2803 80570664 : for (j = BTMaxStrategyNumber; --j >= 0;)
2804 : {
2805 67142220 : if (xform[j].skey)
2806 : {
2807 13430240 : ScanKey outkey = &outkeys[new_numberOfKeys++];
2808 :
2809 13430240 : memcpy(outkey, xform[j].skey, sizeof(ScanKeyData));
2810 13430240 : if (arrayKeyData)
2811 786 : keyDataMap[new_numberOfKeys - 1] = xform[j].ikey;
2812 13430240 : if (priorNumberOfEqualCols == attno - 1)
2813 13429620 : _bt_mark_scankey_required(outkey);
2814 : }
2815 : }
2816 :
2817 : /*
2818 : * Exit loop here if done.
2819 : */
2820 13428444 : if (i == numberOfKeys)
2821 5971366 : break;
2822 :
2823 : /* Re-initialize for new attno */
2824 7457078 : attno = cur->sk_attno;
2825 7457078 : memset(xform, 0, sizeof(xform));
2826 : }
2827 :
2828 : /* check strategy this key's operator corresponds to */
2829 13430476 : j = cur->sk_strategy - 1;
2830 :
2831 : /* if row comparison, push it directly to the output array */
2832 13430476 : if (cur->sk_flags & SK_ROW_HEADER)
2833 : {
2834 0 : ScanKey outkey = &outkeys[new_numberOfKeys++];
2835 :
2836 0 : memcpy(outkey, cur, sizeof(ScanKeyData));
2837 0 : if (arrayKeyData)
2838 0 : keyDataMap[new_numberOfKeys - 1] = i;
2839 0 : if (numberOfEqualCols == attno - 1)
2840 0 : _bt_mark_scankey_required(outkey);
2841 :
2842 : /*
2843 : * We don't support RowCompare using equality; such a qual would
2844 : * mess up the numberOfEqualCols tracking.
2845 : */
2846 : Assert(j != (BTEqualStrategyNumber - 1));
2847 0 : continue;
2848 : }
2849 :
2850 : /*
2851 : * Does this input scan key require further processing as an array?
2852 : */
2853 13430476 : if (cur->sk_strategy == InvalidStrategy)
2854 : {
2855 : /* _bt_preprocess_array_keys marked this array key redundant */
2856 : Assert(arrayKeyData);
2857 : Assert(cur->sk_flags & SK_SEARCHARRAY);
2858 6 : continue;
2859 : }
2860 :
2861 13430470 : if (cur->sk_strategy == BTEqualStrategyNumber &&
2862 12383502 : (cur->sk_flags & SK_SEARCHARRAY))
2863 : {
2864 : /* _bt_preprocess_array_keys kept this array key */
2865 : Assert(arrayKeyData);
2866 540 : arrayidx++;
2867 : }
2868 :
2869 : /*
2870 : * have we seen a scan key for this same attribute and using this same
2871 : * operator strategy before now?
2872 : */
2873 13430470 : if (xform[j].skey == NULL)
2874 : {
2875 : /* nope, so this scan key wins by default (at least for now) */
2876 13430422 : xform[j].skey = cur;
2877 13430422 : xform[j].ikey = i;
2878 13430422 : xform[j].arrayidx = arrayidx;
2879 : }
2880 : else
2881 : {
2882 48 : FmgrInfo *orderproc = NULL;
2883 48 : BTArrayKeyInfo *array = NULL;
2884 :
2885 : /*
2886 : * Seen one of these before, so keep only the more restrictive key
2887 : * if possible
2888 : */
2889 48 : if (j == (BTEqualStrategyNumber - 1) && arrayKeyData)
2890 : {
2891 : /*
2892 : * Have to set up array keys
2893 : */
2894 12 : if ((cur->sk_flags & SK_SEARCHARRAY))
2895 : {
2896 0 : array = &so->arrayKeys[arrayidx - 1];
2897 0 : orderproc = so->orderProcs + i;
2898 :
2899 : Assert(array->scan_key == i);
2900 : Assert(OidIsValid(orderproc->fn_oid));
2901 : }
2902 12 : else if ((xform[j].skey->sk_flags & SK_SEARCHARRAY))
2903 : {
2904 12 : array = &so->arrayKeys[xform[j].arrayidx - 1];
2905 12 : orderproc = so->orderProcs + xform[j].ikey;
2906 :
2907 : Assert(array->scan_key == xform[j].ikey);
2908 : Assert(OidIsValid(orderproc->fn_oid));
2909 : }
2910 :
2911 : /*
2912 : * Both scan keys might have arrays, in which case we'll
2913 : * arbitrarily pass only one of the arrays. That won't
2914 : * matter, since _bt_compare_scankey_args is aware that two
2915 : * SEARCHARRAY scan keys mean that _bt_preprocess_array_keys
2916 : * failed to eliminate redundant arrays through array merging.
2917 : * _bt_compare_scankey_args just returns false when it sees
2918 : * this; it won't even try to examine either array.
2919 : */
2920 : }
2921 :
2922 48 : if (_bt_compare_scankey_args(scan, cur, cur, xform[j].skey,
2923 : array, orderproc, &test_result))
2924 : {
2925 : /* Have all we need to determine redundancy */
2926 48 : if (test_result)
2927 : {
2928 : Assert(!array || array->num_elems > 0);
2929 :
2930 : /*
2931 : * New key is more restrictive, and so replaces old key...
2932 : */
2933 42 : if (j != (BTEqualStrategyNumber - 1) ||
2934 12 : !(xform[j].skey->sk_flags & SK_SEARCHARRAY))
2935 : {
2936 36 : xform[j].skey = cur;
2937 36 : xform[j].ikey = i;
2938 36 : xform[j].arrayidx = arrayidx;
2939 : }
2940 : else
2941 : {
2942 : /*
2943 : * ...unless we have to keep the old key because it's
2944 : * an array that rendered the new key redundant. We
2945 : * need to make sure that we don't throw away an array
2946 : * scan key. _bt_compare_scankey_args expects us to
2947 : * always keep arrays (and discard non-arrays).
2948 : */
2949 : Assert(!(cur->sk_flags & SK_SEARCHARRAY));
2950 : }
2951 : }
2952 6 : else if (j == (BTEqualStrategyNumber - 1))
2953 : {
2954 : /* key == a && key == b, but a != b */
2955 6 : so->qual_ok = false;
2956 6 : return;
2957 : }
2958 : /* else old key is more restrictive, keep it */
2959 : }
2960 : else
2961 : {
2962 : /*
2963 : * We can't determine which key is more restrictive. Push
2964 : * xform[j] directly to the output array, then set xform[j] to
2965 : * the new scan key.
2966 : *
2967 : * Note: We do things this way around so that our arrays are
2968 : * always in the same order as their corresponding scan keys,
2969 : * even with incomplete opfamilies. _bt_advance_array_keys
2970 : * depends on this.
2971 : */
2972 0 : ScanKey outkey = &outkeys[new_numberOfKeys++];
2973 :
2974 0 : memcpy(outkey, xform[j].skey, sizeof(ScanKeyData));
2975 0 : if (arrayKeyData)
2976 0 : keyDataMap[new_numberOfKeys - 1] = xform[j].ikey;
2977 0 : if (numberOfEqualCols == attno - 1)
2978 0 : _bt_mark_scankey_required(outkey);
2979 0 : xform[j].skey = cur;
2980 0 : xform[j].ikey = i;
2981 0 : xform[j].arrayidx = arrayidx;
2982 : }
2983 : }
2984 : }
2985 :
2986 5971366 : so->numberOfKeys = new_numberOfKeys;
2987 :
2988 : /*
2989 : * Now that we've built a temporary mapping from so->keyData[] (output
2990 : * scan keys) to scan->keyData[] (input scan keys), fix array->scan_key
2991 : * references. Also consolidate the so->orderProc[] array such that it
2992 : * can be subscripted using so->keyData[]-wise offsets.
2993 : */
2994 5971366 : if (arrayKeyData)
2995 384 : _bt_preprocess_array_keys_final(scan, keyDataMap);
2996 :
2997 : /* Could pfree arrayKeyData/keyDataMap now, but not worth the cycles */
2998 : }
2999 :
3000 : #ifdef USE_ASSERT_CHECKING
3001 : /*
3002 : * Verify that the scan's qual state matches what we expect at the point that
3003 : * _bt_start_prim_scan is about to start a just-scheduled new primitive scan.
3004 : *
3005 : * We enforce a rule against non-required array scan keys: they must start out
3006 : * with whatever element is the first for the scan's current scan direction.
3007 : * See _bt_rewind_nonrequired_arrays comments for an explanation.
3008 : */
3009 : static bool
3010 : _bt_verify_arrays_bt_first(IndexScanDesc scan, ScanDirection dir)
3011 : {
3012 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
3013 : int arrayidx = 0;
3014 :
3015 : for (int ikey = 0; ikey < so->numberOfKeys; ikey++)
3016 : {
3017 : ScanKey cur = so->keyData + ikey;
3018 : BTArrayKeyInfo *array = NULL;
3019 : int first_elem_dir;
3020 :
3021 : if (!(cur->sk_flags & SK_SEARCHARRAY) ||
3022 : cur->sk_strategy != BTEqualStrategyNumber)
3023 : continue;
3024 :
3025 : array = &so->arrayKeys[arrayidx++];
3026 :
3027 : if (((cur->sk_flags & SK_BT_REQFWD) && ScanDirectionIsForward(dir)) ||
3028 : ((cur->sk_flags & SK_BT_REQBKWD) && ScanDirectionIsBackward(dir)))
3029 : continue;
3030 :
3031 : if (ScanDirectionIsForward(dir))
3032 : first_elem_dir = 0;
3033 : else
3034 : first_elem_dir = array->num_elems - 1;
3035 :
3036 : if (array->cur_elem != first_elem_dir)
3037 : return false;
3038 : }
3039 :
3040 : return _bt_verify_keys_with_arraykeys(scan);
3041 : }
3042 :
3043 : /*
3044 : * Verify that the scan's "so->keyData[]" scan keys are in agreement with
3045 : * its array key state
3046 : */
3047 : static bool
3048 : _bt_verify_keys_with_arraykeys(IndexScanDesc scan)
3049 : {
3050 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
3051 : int last_sk_attno = InvalidAttrNumber,
3052 : arrayidx = 0;
3053 :
3054 : if (!so->qual_ok)
3055 : return false;
3056 :
3057 : for (int ikey = 0; ikey < so->numberOfKeys; ikey++)
3058 : {
3059 : ScanKey cur = so->keyData + ikey;
3060 : BTArrayKeyInfo *array;
3061 :
3062 : if (cur->sk_strategy != BTEqualStrategyNumber ||
3063 : !(cur->sk_flags & SK_SEARCHARRAY))
3064 : continue;
3065 :
3066 : array = &so->arrayKeys[arrayidx++];
3067 : if (array->scan_key != ikey)
3068 : return false;
3069 :
3070 : if (array->num_elems <= 0)
3071 : return false;
3072 :
3073 : if (cur->sk_argument != array->elem_values[array->cur_elem])
3074 : return false;
3075 : if (last_sk_attno > cur->sk_attno)
3076 : return false;
3077 : last_sk_attno = cur->sk_attno;
3078 : }
3079 :
3080 : if (arrayidx != so->numArrayKeys)
3081 : return false;
3082 :
3083 : return true;
3084 : }
3085 : #endif
3086 :
3087 : /*
3088 : * Compare two scankey values using a specified operator.
3089 : *
3090 : * The test we want to perform is logically "leftarg op rightarg", where
3091 : * leftarg and rightarg are the sk_argument values in those ScanKeys, and
3092 : * the comparison operator is the one in the op ScanKey. However, in
3093 : * cross-data-type situations we may need to look up the correct operator in
3094 : * the index's opfamily: it is the one having amopstrategy = op->sk_strategy
3095 : * and amoplefttype/amoprighttype equal to the two argument datatypes.
3096 : *
3097 : * If the opfamily doesn't supply a complete set of cross-type operators we
3098 : * may not be able to make the comparison. If we can make the comparison
3099 : * we store the operator result in *result and return true. We return false
3100 : * if the comparison could not be made.
3101 : *
3102 : * If either leftarg or rightarg are an array, we'll apply array-specific
3103 : * rules to determine which array elements are redundant on behalf of caller.
3104 : * It is up to our caller to save whichever of the two scan keys is the array,
3105 : * and discard the non-array scan key (the non-array scan key is guaranteed to
3106 : * be redundant with any complete opfamily). Caller isn't expected to call
3107 : * here with a pair of array scan keys provided we're dealing with a complete
3108 : * opfamily (_bt_preprocess_array_keys will merge array keys together to make
3109 : * sure of that).
3110 : *
3111 : * Note: we'll also shrink caller's array as needed to eliminate redundant
3112 : * array elements. One reason why caller should prefer to discard non-array
3113 : * scan keys is so that we'll have the opportunity to shrink the array
3114 : * multiple times, in multiple calls (for each of several other scan keys on
3115 : * the same index attribute).
3116 : *
3117 : * Note: op always points at the same ScanKey as either leftarg or rightarg.
3118 : * Since we don't scribble on the scankeys themselves, this aliasing should
3119 : * cause no trouble.
3120 : *
3121 : * Note: this routine needs to be insensitive to any DESC option applied
3122 : * to the index column. For example, "x < 4" is a tighter constraint than
3123 : * "x < 5" regardless of which way the index is sorted.
3124 : */
3125 : static bool
3126 146 : _bt_compare_scankey_args(IndexScanDesc scan, ScanKey op,
3127 : ScanKey leftarg, ScanKey rightarg,
3128 : BTArrayKeyInfo *array, FmgrInfo *orderproc,
3129 : bool *result)
3130 : {
3131 146 : Relation rel = scan->indexRelation;
3132 : Oid lefttype,
3133 : righttype,
3134 : optype,
3135 : opcintype,
3136 : cmp_op;
3137 : StrategyNumber strat;
3138 :
3139 : /*
3140 : * First, deal with cases where one or both args are NULL. This should
3141 : * only happen when the scankeys represent IS NULL/NOT NULL conditions.
3142 : */
3143 146 : if ((leftarg->sk_flags | rightarg->sk_flags) & SK_ISNULL)
3144 : {
3145 : bool leftnull,
3146 : rightnull;
3147 :
3148 12 : if (leftarg->sk_flags & SK_ISNULL)
3149 : {
3150 : Assert(leftarg->sk_flags & (SK_SEARCHNULL | SK_SEARCHNOTNULL));
3151 0 : leftnull = true;
3152 : }
3153 : else
3154 12 : leftnull = false;
3155 12 : if (rightarg->sk_flags & SK_ISNULL)
3156 : {
3157 : Assert(rightarg->sk_flags & (SK_SEARCHNULL | SK_SEARCHNOTNULL));
3158 12 : rightnull = true;
3159 : }
3160 : else
3161 0 : rightnull = false;
3162 :
3163 : /*
3164 : * We treat NULL as either greater than or less than all other values.
3165 : * Since true > false, the tests below work correctly for NULLS LAST
3166 : * logic. If the index is NULLS FIRST, we need to flip the strategy.
3167 : */
3168 12 : strat = op->sk_strategy;
3169 12 : if (op->sk_flags & SK_BT_NULLS_FIRST)
3170 0 : strat = BTCommuteStrategyNumber(strat);
3171 :
3172 12 : switch (strat)
3173 : {
3174 12 : case BTLessStrategyNumber:
3175 12 : *result = (leftnull < rightnull);
3176 12 : break;
3177 0 : case BTLessEqualStrategyNumber:
3178 0 : *result = (leftnull <= rightnull);
3179 0 : break;
3180 0 : case BTEqualStrategyNumber:
3181 0 : *result = (leftnull == rightnull);
3182 0 : break;
3183 0 : case BTGreaterEqualStrategyNumber:
3184 0 : *result = (leftnull >= rightnull);
3185 0 : break;
3186 0 : case BTGreaterStrategyNumber:
3187 0 : *result = (leftnull > rightnull);
3188 0 : break;
3189 0 : default:
3190 0 : elog(ERROR, "unrecognized StrategyNumber: %d", (int) strat);
3191 : *result = false; /* keep compiler quiet */
3192 : break;
3193 : }
3194 12 : return true;
3195 : }
3196 :
3197 : /*
3198 : * If either leftarg or rightarg are equality-type array scankeys, we need
3199 : * specialized handling (since by now we know that IS NULL wasn't used)
3200 : */
3201 134 : if (array)
3202 : {
3203 : bool leftarray,
3204 : rightarray;
3205 :
3206 48 : leftarray = ((leftarg->sk_flags & SK_SEARCHARRAY) &&
3207 18 : leftarg->sk_strategy == BTEqualStrategyNumber);
3208 42 : rightarray = ((rightarg->sk_flags & SK_SEARCHARRAY) &&
3209 12 : rightarg->sk_strategy == BTEqualStrategyNumber);
3210 :
3211 : /*
3212 : * _bt_preprocess_array_keys is responsible for merging together array
3213 : * scan keys, and will do so whenever the opfamily has the required
3214 : * cross-type support. If it failed to do that, we handle it just
3215 : * like the case where we can't make the comparison ourselves.
3216 : */
3217 30 : if (leftarray && rightarray)
3218 : {
3219 : /* Can't make the comparison */
3220 0 : *result = false; /* suppress compiler warnings */
3221 0 : return false;
3222 : }
3223 :
3224 : /*
3225 : * Otherwise we need to determine if either one of leftarg or rightarg
3226 : * uses an array, then pass this through to a dedicated helper
3227 : * function.
3228 : */
3229 30 : if (leftarray)
3230 18 : return _bt_compare_array_scankey_args(scan, leftarg, rightarg,
3231 : orderproc, array, result);
3232 12 : else if (rightarray)
3233 12 : return _bt_compare_array_scankey_args(scan, rightarg, leftarg,
3234 : orderproc, array, result);
3235 :
3236 : /* FALL THRU */
3237 : }
3238 :
3239 : /*
3240 : * The opfamily we need to worry about is identified by the index column.
3241 : */
3242 : Assert(leftarg->sk_attno == rightarg->sk_attno);
3243 :
3244 104 : opcintype = rel->rd_opcintype[leftarg->sk_attno - 1];
3245 :
3246 : /*
3247 : * Determine the actual datatypes of the ScanKey arguments. We have to
3248 : * support the convention that sk_subtype == InvalidOid means the opclass
3249 : * input type; this is a hack to simplify life for ScanKeyInit().
3250 : */
3251 104 : lefttype = leftarg->sk_subtype;
3252 104 : if (lefttype == InvalidOid)
3253 0 : lefttype = opcintype;
3254 104 : righttype = rightarg->sk_subtype;
3255 104 : if (righttype == InvalidOid)
3256 0 : righttype = opcintype;
3257 104 : optype = op->sk_subtype;
3258 104 : if (optype == InvalidOid)
3259 0 : optype = opcintype;
3260 :
3261 : /*
3262 : * If leftarg and rightarg match the types expected for the "op" scankey,
3263 : * we can use its already-looked-up comparison function.
3264 : */
3265 104 : if (lefttype == opcintype && righttype == optype)
3266 : {
3267 98 : *result = DatumGetBool(FunctionCall2Coll(&op->sk_func,
3268 : op->sk_collation,
3269 : leftarg->sk_argument,
3270 : rightarg->sk_argument));
3271 98 : return true;
3272 : }
3273 :
3274 : /*
3275 : * Otherwise, we need to go to the syscache to find the appropriate
3276 : * operator. (This cannot result in infinite recursion, since no
3277 : * indexscan initiated by syscache lookup will use cross-data-type
3278 : * operators.)
3279 : *
3280 : * If the sk_strategy was flipped by _bt_fix_scankey_strategy, we have to
3281 : * un-flip it to get the correct opfamily member.
3282 : */
3283 6 : strat = op->sk_strategy;
3284 6 : if (op->sk_flags & SK_BT_DESC)
3285 0 : strat = BTCommuteStrategyNumber(strat);
3286 :
3287 6 : cmp_op = get_opfamily_member(rel->rd_opfamily[leftarg->sk_attno - 1],
3288 : lefttype,
3289 : righttype,
3290 : strat);
3291 6 : if (OidIsValid(cmp_op))
3292 : {
3293 6 : RegProcedure cmp_proc = get_opcode(cmp_op);
3294 :
3295 6 : if (RegProcedureIsValid(cmp_proc))
3296 : {
3297 6 : *result = DatumGetBool(OidFunctionCall2Coll(cmp_proc,
3298 : op->sk_collation,
3299 : leftarg->sk_argument,
3300 : rightarg->sk_argument));
3301 6 : return true;
3302 : }
3303 : }
3304 :
3305 : /* Can't make the comparison */
3306 0 : *result = false; /* suppress compiler warnings */
3307 0 : return false;
3308 : }
3309 :
3310 : /*
3311 : * Adjust a scankey's strategy and flags setting as needed for indoptions.
3312 : *
3313 : * We copy the appropriate indoption value into the scankey sk_flags
3314 : * (shifting to avoid clobbering system-defined flag bits). Also, if
3315 : * the DESC option is set, commute (flip) the operator strategy number.
3316 : *
3317 : * A secondary purpose is to check for IS NULL/NOT NULL scankeys and set up
3318 : * the strategy field correctly for them.
3319 : *
3320 : * Lastly, for ordinary scankeys (not IS NULL/NOT NULL), we check for a
3321 : * NULL comparison value. Since all btree operators are assumed strict,
3322 : * a NULL means that the qual cannot be satisfied. We return true if the
3323 : * comparison value isn't NULL, or false if the scan should be abandoned.
3324 : *
3325 : * This function is applied to the *input* scankey structure; therefore
3326 : * on a rescan we will be looking at already-processed scankeys. Hence
3327 : * we have to be careful not to re-commute the strategy if we already did it.
3328 : * It's a bit ugly to modify the caller's copy of the scankey but in practice
3329 : * there shouldn't be any problem, since the index's indoptions are certainly
3330 : * not going to change while the scankey survives.
3331 : */
3332 : static bool
3333 19278276 : _bt_fix_scankey_strategy(ScanKey skey, int16 *indoption)
3334 : {
3335 : int addflags;
3336 :
3337 19278276 : addflags = indoption[skey->sk_attno - 1] << SK_BT_INDOPTION_SHIFT;
3338 :
3339 : /*
3340 : * We treat all btree operators as strict (even if they're not so marked
3341 : * in pg_proc). This means that it is impossible for an operator condition
3342 : * with a NULL comparison constant to succeed, and we can reject it right
3343 : * away.
3344 : *
3345 : * However, we now also support "x IS NULL" clauses as search conditions,
3346 : * so in that case keep going. The planner has not filled in any
3347 : * particular strategy in this case, so set it to BTEqualStrategyNumber
3348 : * --- we can treat IS NULL as an equality operator for purposes of search
3349 : * strategy.
3350 : *
3351 : * Likewise, "x IS NOT NULL" is supported. We treat that as either "less
3352 : * than NULL" in a NULLS LAST index, or "greater than NULL" in a NULLS
3353 : * FIRST index.
3354 : *
3355 : * Note: someday we might have to fill in sk_collation from the index
3356 : * column's collation. At the moment this is a non-issue because we'll
3357 : * never actually call the comparison operator on a NULL.
3358 : */
3359 19278276 : if (skey->sk_flags & SK_ISNULL)
3360 : {
3361 : /* SK_ISNULL shouldn't be set in a row header scankey */
3362 : Assert(!(skey->sk_flags & SK_ROW_HEADER));
3363 :
3364 : /* Set indoption flags in scankey (might be done already) */
3365 107266 : skey->sk_flags |= addflags;
3366 :
3367 : /* Set correct strategy for IS NULL or NOT NULL search */
3368 107266 : if (skey->sk_flags & SK_SEARCHNULL)
3369 : {
3370 140 : skey->sk_strategy = BTEqualStrategyNumber;
3371 140 : skey->sk_subtype = InvalidOid;
3372 140 : skey->sk_collation = InvalidOid;
3373 : }
3374 107126 : else if (skey->sk_flags & SK_SEARCHNOTNULL)
3375 : {
3376 104466 : if (skey->sk_flags & SK_BT_NULLS_FIRST)
3377 36 : skey->sk_strategy = BTGreaterStrategyNumber;
3378 : else
3379 104430 : skey->sk_strategy = BTLessStrategyNumber;
3380 104466 : skey->sk_subtype = InvalidOid;
3381 104466 : skey->sk_collation = InvalidOid;
3382 : }
3383 : else
3384 : {
3385 : /* regular qual, so it cannot be satisfied */
3386 2660 : return false;
3387 : }
3388 :
3389 : /* Needn't do the rest */
3390 104606 : return true;
3391 : }
3392 :
3393 19171010 : if (skey->sk_strategy == InvalidStrategy)
3394 : {
3395 : /* Already-eliminated array scan key; don't need to fix anything */
3396 : Assert(skey->sk_flags & SK_SEARCHARRAY);
3397 6 : return true;
3398 : }
3399 :
3400 : /* Adjust strategy for DESC, if we didn't already */
3401 19171004 : if ((addflags & SK_BT_DESC) && !(skey->sk_flags & SK_BT_DESC))
3402 6 : skey->sk_strategy = BTCommuteStrategyNumber(skey->sk_strategy);
3403 19171004 : skey->sk_flags |= addflags;
3404 :
3405 : /* If it's a row header, fix row member flags and strategies similarly */
3406 19171004 : if (skey->sk_flags & SK_ROW_HEADER)
3407 : {
3408 36 : ScanKey subkey = (ScanKey) DatumGetPointer(skey->sk_argument);
3409 :
3410 : for (;;)
3411 : {
3412 36 : Assert(subkey->sk_flags & SK_ROW_MEMBER);
3413 72 : addflags = indoption[subkey->sk_attno - 1] << SK_BT_INDOPTION_SHIFT;
3414 72 : if ((addflags & SK_BT_DESC) && !(subkey->sk_flags & SK_BT_DESC))
3415 0 : subkey->sk_strategy = BTCommuteStrategyNumber(subkey->sk_strategy);
3416 72 : subkey->sk_flags |= addflags;
3417 72 : if (subkey->sk_flags & SK_ROW_END)
3418 36 : break;
3419 36 : subkey++;
3420 : }
3421 : }
3422 :
3423 19171004 : return true;
3424 : }
3425 :
3426 : /*
3427 : * Mark a scankey as "required to continue the scan".
3428 : *
3429 : * Depending on the operator type, the key may be required for both scan
3430 : * directions or just one. Also, if the key is a row comparison header,
3431 : * we have to mark its first subsidiary ScanKey as required. (Subsequent
3432 : * subsidiary ScanKeys are normally for lower-order columns, and thus
3433 : * cannot be required, since they're after the first non-equality scankey.)
3434 : *
3435 : * Note: when we set required-key flag bits in a subsidiary scankey, we are
3436 : * scribbling on a data structure belonging to the index AM's caller, not on
3437 : * our private copy. This should be OK because the marking will not change
3438 : * from scan to scan within a query, and so we'd just re-mark the same way
3439 : * anyway on a rescan. Something to keep an eye on though.
3440 : */
3441 : static void
3442 19274746 : _bt_mark_scankey_required(ScanKey skey)
3443 : {
3444 : int addflags;
3445 :
3446 19274746 : switch (skey->sk_strategy)
3447 : {
3448 107428 : case BTLessStrategyNumber:
3449 : case BTLessEqualStrategyNumber:
3450 107428 : addflags = SK_BT_REQFWD;
3451 107428 : break;
3452 18121432 : case BTEqualStrategyNumber:
3453 18121432 : addflags = SK_BT_REQFWD | SK_BT_REQBKWD;
3454 18121432 : break;
3455 1045886 : case BTGreaterEqualStrategyNumber:
3456 : case BTGreaterStrategyNumber:
3457 1045886 : addflags = SK_BT_REQBKWD;
3458 1045886 : break;
3459 0 : default:
3460 0 : elog(ERROR, "unrecognized StrategyNumber: %d",
3461 : (int) skey->sk_strategy);
3462 : addflags = 0; /* keep compiler quiet */
3463 : break;
3464 : }
3465 :
3466 19274746 : skey->sk_flags |= addflags;
3467 :
3468 19274746 : if (skey->sk_flags & SK_ROW_HEADER)
3469 : {
3470 36 : ScanKey subkey = (ScanKey) DatumGetPointer(skey->sk_argument);
3471 :
3472 : /* First subkey should be same column/operator as the header */
3473 : Assert(subkey->sk_flags & SK_ROW_MEMBER);
3474 : Assert(subkey->sk_attno == skey->sk_attno);
3475 : Assert(subkey->sk_strategy == skey->sk_strategy);
3476 36 : subkey->sk_flags |= addflags;
3477 : }
3478 19274746 : }
3479 :
3480 : /*
3481 : * Test whether an indextuple satisfies all the scankey conditions.
3482 : *
3483 : * Return true if so, false if not. If the tuple fails to pass the qual,
3484 : * we also determine whether there's any need to continue the scan beyond
3485 : * this tuple, and set pstate.continuescan accordingly. See comments for
3486 : * _bt_preprocess_keys(), above, about how this is done.
3487 : *
3488 : * Forward scan callers can pass a high key tuple in the hopes of having
3489 : * us set *continuescan to false, and avoiding an unnecessary visit to
3490 : * the page to the right.
3491 : *
3492 : * Advances the scan's array keys when necessary for arrayKeys=true callers.
3493 : * Caller can avoid all array related side-effects when calling just to do a
3494 : * page continuescan precheck -- pass arrayKeys=false for that. Scans without
3495 : * any arrays keys must always pass arrayKeys=false.
3496 : *
3497 : * Also stops and starts primitive index scans for arrayKeys=true callers.
3498 : * Scans with array keys are required to set up page state that helps us with
3499 : * this. The page's finaltup tuple (the page high key for a forward scan, or
3500 : * the page's first non-pivot tuple for a backward scan) must be set in
3501 : * pstate.finaltup ahead of the first call here for the page (or possibly the
3502 : * first call after an initial continuescan-setting page precheck call). Set
3503 : * this to NULL for rightmost page (or the leftmost page for backwards scans).
3504 : *
3505 : * scan: index scan descriptor (containing a search-type scankey)
3506 : * pstate: page level input and output parameters
3507 : * arrayKeys: should we advance the scan's array keys if necessary?
3508 : * tuple: index tuple to test
3509 : * tupnatts: number of attributes in tupnatts (high key may be truncated)
3510 : */
3511 : bool
3512 48604634 : _bt_checkkeys(IndexScanDesc scan, BTReadPageState *pstate, bool arrayKeys,
3513 : IndexTuple tuple, int tupnatts)
3514 : {
3515 48604634 : TupleDesc tupdesc = RelationGetDescr(scan->indexRelation);
3516 48604634 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
3517 48604634 : ScanDirection dir = pstate->dir;
3518 48604634 : int ikey = 0;
3519 : bool res;
3520 :
3521 : Assert(BTreeTupleGetNAtts(tuple, scan->indexRelation) == tupnatts);
3522 :
3523 48604634 : res = _bt_check_compare(scan, dir, tuple, tupnatts, tupdesc,
3524 48604634 : arrayKeys, pstate->prechecked, pstate->firstmatch,
3525 : &pstate->continuescan, &ikey);
3526 :
3527 : #ifdef USE_ASSERT_CHECKING
3528 : if (!arrayKeys && so->numArrayKeys)
3529 : {
3530 : /*
3531 : * This is a continuescan precheck call for a scan with array keys.
3532 : *
3533 : * Assert that the scan isn't in danger of becoming confused.
3534 : */
3535 : Assert(!so->scanBehind && !pstate->prechecked && !pstate->firstmatch);
3536 : Assert(!_bt_tuple_before_array_skeys(scan, dir, tuple, tupdesc,
3537 : tupnatts, false, 0, NULL));
3538 : }
3539 : if (pstate->prechecked || pstate->firstmatch)
3540 : {
3541 : bool dcontinuescan;
3542 : int dikey = 0;
3543 :
3544 : /*
3545 : * Call relied on continuescan/firstmatch prechecks -- assert that we
3546 : * get the same answer without those optimizations
3547 : */
3548 : Assert(res == _bt_check_compare(scan, dir, tuple, tupnatts, tupdesc,
3549 : false, false, false,
3550 : &dcontinuescan, &dikey));
3551 : Assert(pstate->continuescan == dcontinuescan);
3552 : }
3553 : #endif
3554 :
3555 : /*
3556 : * Only one _bt_check_compare call is required in the common case where
3557 : * there are no equality strategy array scan keys. Otherwise we can only
3558 : * accept _bt_check_compare's answer unreservedly when it didn't set
3559 : * pstate.continuescan=false.
3560 : */
3561 48604634 : if (!arrayKeys || pstate->continuescan)
3562 48598594 : return res;
3563 :
3564 : /*
3565 : * _bt_check_compare call set continuescan=false in the presence of
3566 : * equality type array keys. This could mean that the tuple is just past
3567 : * the end of matches for the current array keys.
3568 : *
3569 : * It's also possible that the scan is still _before_ the _start_ of
3570 : * tuples matching the current set of array keys. Check for that first.
3571 : */
3572 6040 : if (_bt_tuple_before_array_skeys(scan, dir, tuple, tupdesc, tupnatts, true,
3573 : ikey, NULL))
3574 : {
3575 : /*
3576 : * Tuple is still before the start of matches according to the scan's
3577 : * required array keys (according to _all_ of its required equality
3578 : * strategy keys, actually).
3579 : *
3580 : * _bt_advance_array_keys occasionally sets so->scanBehind to signal
3581 : * that the scan's current position/tuples might be significantly
3582 : * behind (multiple pages behind) its current array keys. When this
3583 : * happens, we need to be prepared to recover by starting a new
3584 : * primitive index scan here, on our own.
3585 : */
3586 : Assert(!so->scanBehind ||
3587 : so->keyData[ikey].sk_strategy == BTEqualStrategyNumber);
3588 2570 : if (unlikely(so->scanBehind) && pstate->finaltup &&
3589 60 : _bt_tuple_before_array_skeys(scan, dir, pstate->finaltup, tupdesc,
3590 60 : BTreeTupleGetNAtts(pstate->finaltup,
3591 : scan->indexRelation),
3592 : false, 0, NULL))
3593 : {
3594 : /* Cut our losses -- start a new primitive index scan now */
3595 0 : pstate->continuescan = false;
3596 0 : so->needPrimScan = true;
3597 : }
3598 : else
3599 : {
3600 : /* Override _bt_check_compare, continue primitive scan */
3601 2540 : pstate->continuescan = true;
3602 :
3603 : /*
3604 : * We will end up here repeatedly given a group of tuples > the
3605 : * previous array keys and < the now-current keys (for a backwards
3606 : * scan it's just the same, though the operators swap positions).
3607 : *
3608 : * We must avoid allowing this linear search process to scan very
3609 : * many tuples from well before the start of tuples matching the
3610 : * current array keys (or from well before the point where we'll
3611 : * once again have to advance the scan's array keys).
3612 : *
3613 : * We keep the overhead under control by speculatively "looking
3614 : * ahead" to later still-unscanned items from this same leaf page.
3615 : * We'll only attempt this once the number of tuples that the
3616 : * linear search process has examined starts to get out of hand.
3617 : */
3618 2540 : pstate->rechecks++;
3619 2540 : if (pstate->rechecks >= LOOK_AHEAD_REQUIRED_RECHECKS)
3620 : {
3621 : /* See if we should skip ahead within the current leaf page */
3622 986 : _bt_checkkeys_look_ahead(scan, pstate, tupnatts, tupdesc);
3623 :
3624 : /*
3625 : * Might have set pstate.skip to a later page offset. When
3626 : * that happens then _bt_readpage caller will inexpensively
3627 : * skip ahead to a later tuple from the same page (the one
3628 : * just after the tuple we successfully "looked ahead" to).
3629 : */
3630 : }
3631 : }
3632 :
3633 : /* This indextuple doesn't match the current qual, in any case */
3634 2540 : return false;
3635 : }
3636 :
3637 : /*
3638 : * Caller's tuple is >= the current set of array keys and other equality
3639 : * constraint scan keys (or <= if this is a backwards scan). It's now
3640 : * clear that we _must_ advance any required array keys in lockstep with
3641 : * the scan.
3642 : */
3643 3500 : return _bt_advance_array_keys(scan, pstate, tuple, tupnatts, tupdesc,
3644 : ikey, true);
3645 : }
3646 :
3647 : /*
3648 : * Test whether an indextuple satisfies current scan condition.
3649 : *
3650 : * Return true if so, false if not. If not, also sets *continuescan to false
3651 : * when it's also not possible for any later tuples to pass the current qual
3652 : * (with the scan's current set of array keys, in the current scan direction),
3653 : * in addition to setting *ikey to the so->keyData[] subscript/offset for the
3654 : * unsatisfied scan key (needed when caller must consider advancing the scan's
3655 : * array keys).
3656 : *
3657 : * This is a subroutine for _bt_checkkeys. We provisionally assume that
3658 : * reaching the end of the current set of required keys (in particular the
3659 : * current required array keys) ends the ongoing (primitive) index scan.
3660 : * Callers without array keys should just end the scan right away when they
3661 : * find that continuescan has been set to false here by us. Things are more
3662 : * complicated for callers with array keys.
3663 : *
3664 : * Callers with array keys must first consider advancing the arrays when
3665 : * continuescan has been set to false here by us. They must then consider if
3666 : * it really does make sense to end the current (primitive) index scan, in
3667 : * light of everything that is known at that point. (In general when we set
3668 : * continuescan=false for these callers it must be treated as provisional.)
3669 : *
3670 : * We deal with advancing unsatisfied non-required arrays directly, though.
3671 : * This is safe, since by definition non-required keys can't end the scan.
3672 : * This is just how we determine if non-required arrays are just unsatisfied
3673 : * by the current array key, or if they're truly unsatisfied (that is, if
3674 : * they're unsatisfied by every possible array key).
3675 : *
3676 : * Though we advance non-required array keys on our own, that shouldn't have
3677 : * any lasting consequences for the scan. By definition, non-required arrays
3678 : * have no fixed relationship with the scan's progress. (There are delicate
3679 : * considerations for non-required arrays when the arrays need to be advanced
3680 : * following our setting continuescan to false, but that doesn't concern us.)
3681 : *
3682 : * Pass advancenonrequired=false to avoid all array related side effects.
3683 : * This allows _bt_advance_array_keys caller to avoid infinite recursion.
3684 : */
3685 : static bool
3686 48606446 : _bt_check_compare(IndexScanDesc scan, ScanDirection dir,
3687 : IndexTuple tuple, int tupnatts, TupleDesc tupdesc,
3688 : bool advancenonrequired, bool prechecked, bool firstmatch,
3689 : bool *continuescan, int *ikey)
3690 : {
3691 48606446 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
3692 :
3693 48606446 : *continuescan = true; /* default assumption */
3694 :
3695 94424502 : for (; *ikey < so->numberOfKeys; (*ikey)++)
3696 : {
3697 54944864 : ScanKey key = so->keyData + *ikey;
3698 : Datum datum;
3699 : bool isNull;
3700 54944864 : bool requiredSameDir = false,
3701 54944864 : requiredOppositeDirOnly = false;
3702 :
3703 : /*
3704 : * Check if the key is required in the current scan direction, in the
3705 : * opposite scan direction _only_, or in neither direction
3706 : */
3707 54944864 : if (((key->sk_flags & SK_BT_REQFWD) && ScanDirectionIsForward(dir)) ||
3708 11762040 : ((key->sk_flags & SK_BT_REQBKWD) && ScanDirectionIsBackward(dir)))
3709 43200138 : requiredSameDir = true;
3710 11744726 : else if (((key->sk_flags & SK_BT_REQFWD) && ScanDirectionIsBackward(dir)) ||
3711 5354210 : ((key->sk_flags & SK_BT_REQBKWD) && ScanDirectionIsForward(dir)))
3712 11252378 : requiredOppositeDirOnly = true;
3713 :
3714 : /*
3715 : * If the caller told us the *continuescan flag is known to be true
3716 : * for the last item on the page, then we know the keys required for
3717 : * the current direction scan should be matched. Otherwise, the
3718 : * *continuescan flag would be set for the current item and
3719 : * subsequently the last item on the page accordingly.
3720 : *
3721 : * If the key is required for the opposite direction scan, we can skip
3722 : * the check if the caller tells us there was already at least one
3723 : * matching item on the page. Also, we require the *continuescan flag
3724 : * to be true for the last item on the page to know there are no
3725 : * NULLs.
3726 : *
3727 : * Both cases above work except for the row keys, where NULLs could be
3728 : * found in the middle of matching values.
3729 : */
3730 54944864 : if (prechecked &&
3731 1447150 : (requiredSameDir || (requiredOppositeDirOnly && firstmatch)) &&
3732 1405142 : !(key->sk_flags & SK_ROW_HEADER))
3733 18455402 : continue;
3734 :
3735 53539722 : if (key->sk_attno > tupnatts)
3736 : {
3737 : /*
3738 : * This attribute is truncated (must be high key). The value for
3739 : * this attribute in the first non-pivot tuple on the page to the
3740 : * right could be any possible value. Assume that truncated
3741 : * attribute passes the qual.
3742 : */
3743 : Assert(BTreeTupleIsPivot(tuple));
3744 2282 : continue;
3745 : }
3746 :
3747 : /* row-comparison keys need special processing */
3748 53537440 : if (key->sk_flags & SK_ROW_HEADER)
3749 : {
3750 1980 : if (_bt_check_rowcompare(key, tuple, tupnatts, tupdesc, dir,
3751 : continuescan))
3752 1926 : continue;
3753 9126808 : return false;
3754 : }
3755 :
3756 53535460 : datum = index_getattr(tuple,
3757 53535460 : key->sk_attno,
3758 : tupdesc,
3759 : &isNull);
3760 :
3761 53535460 : if (key->sk_flags & SK_ISNULL)
3762 : {
3763 : /* Handle IS NULL/NOT NULL tests */
3764 17094214 : if (key->sk_flags & SK_SEARCHNULL)
3765 : {
3766 48236 : if (isNull)
3767 164 : continue; /* tuple satisfies this qual */
3768 : }
3769 : else
3770 : {
3771 : Assert(key->sk_flags & SK_SEARCHNOTNULL);
3772 17045978 : if (!isNull)
3773 17045888 : continue; /* tuple satisfies this qual */
3774 : }
3775 :
3776 : /*
3777 : * Tuple fails this qual. If it's a required qual for the current
3778 : * scan direction, then we can conclude no further tuples will
3779 : * pass, either.
3780 : */
3781 48162 : if (requiredSameDir)
3782 36 : *continuescan = false;
3783 :
3784 : /*
3785 : * In any case, this indextuple doesn't match the qual.
3786 : */
3787 48162 : return false;
3788 : }
3789 :
3790 36441246 : if (isNull)
3791 : {
3792 150 : if (key->sk_flags & SK_BT_NULLS_FIRST)
3793 : {
3794 : /*
3795 : * Since NULLs are sorted before non-NULLs, we know we have
3796 : * reached the lower limit of the range of values for this
3797 : * index attr. On a backward scan, we can stop if this qual
3798 : * is one of the "must match" subset. We can stop regardless
3799 : * of whether the qual is > or <, so long as it's required,
3800 : * because it's not possible for any future tuples to pass. On
3801 : * a forward scan, however, we must keep going, because we may
3802 : * have initially positioned to the start of the index.
3803 : * (_bt_advance_array_keys also relies on this behavior during
3804 : * forward scans.)
3805 : */
3806 0 : if ((key->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD)) &&
3807 : ScanDirectionIsBackward(dir))
3808 0 : *continuescan = false;
3809 : }
3810 : else
3811 : {
3812 : /*
3813 : * Since NULLs are sorted after non-NULLs, we know we have
3814 : * reached the upper limit of the range of values for this
3815 : * index attr. On a forward scan, we can stop if this qual is
3816 : * one of the "must match" subset. We can stop regardless of
3817 : * whether the qual is > or <, so long as it's required,
3818 : * because it's not possible for any future tuples to pass. On
3819 : * a backward scan, however, we must keep going, because we
3820 : * may have initially positioned to the end of the index.
3821 : * (_bt_advance_array_keys also relies on this behavior during
3822 : * backward scans.)
3823 : */
3824 150 : if ((key->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD)) &&
3825 : ScanDirectionIsForward(dir))
3826 84 : *continuescan = false;
3827 : }
3828 :
3829 : /*
3830 : * In any case, this indextuple doesn't match the qual.
3831 : */
3832 150 : return false;
3833 : }
3834 :
3835 : /*
3836 : * Apply the key-checking function, though only if we must.
3837 : *
3838 : * When a key is required in the opposite-of-scan direction _only_,
3839 : * then it must already be satisfied if firstmatch=true indicates that
3840 : * an earlier tuple from this same page satisfied it earlier on.
3841 : */
3842 36441096 : if (!(requiredOppositeDirOnly && firstmatch) &&
3843 33685882 : !DatumGetBool(FunctionCall2Coll(&key->sk_func, key->sk_collation,
3844 : datum, key->sk_argument)))
3845 : {
3846 : /*
3847 : * Tuple fails this qual. If it's a required qual for the current
3848 : * scan direction, then we can conclude no further tuples will
3849 : * pass, either.
3850 : *
3851 : * Note: because we stop the scan as soon as any required equality
3852 : * qual fails, it is critical that equality quals be used for the
3853 : * initial positioning in _bt_first() when they are available. See
3854 : * comments in _bt_first().
3855 : */
3856 9078442 : if (requiredSameDir)
3857 8721088 : *continuescan = false;
3858 :
3859 : /*
3860 : * If this is a non-required equality-type array key, the tuple
3861 : * needs to be checked against every possible array key. Handle
3862 : * this by "advancing" the scan key's array to a matching value
3863 : * (if we're successful then the tuple might match the qual).
3864 : */
3865 357354 : else if (advancenonrequired &&
3866 306 : key->sk_strategy == BTEqualStrategyNumber &&
3867 240 : (key->sk_flags & SK_SEARCHARRAY))
3868 240 : return _bt_advance_array_keys(scan, NULL, tuple, tupnatts,
3869 : tupdesc, *ikey, false);
3870 :
3871 : /*
3872 : * This indextuple doesn't match the qual.
3873 : */
3874 9078202 : return false;
3875 : }
3876 : }
3877 :
3878 : /* If we get here, the tuple passes all index quals. */
3879 39479638 : return true;
3880 : }
3881 :
3882 : /*
3883 : * Test whether an indextuple satisfies a row-comparison scan condition.
3884 : *
3885 : * Return true if so, false if not. If not, also clear *continuescan if
3886 : * it's not possible for any future tuples in the current scan direction
3887 : * to pass the qual.
3888 : *
3889 : * This is a subroutine for _bt_checkkeys/_bt_check_compare.
3890 : */
3891 : static bool
3892 1980 : _bt_check_rowcompare(ScanKey skey, IndexTuple tuple, int tupnatts,
3893 : TupleDesc tupdesc, ScanDirection dir, bool *continuescan)
3894 : {
3895 1980 : ScanKey subkey = (ScanKey) DatumGetPointer(skey->sk_argument);
3896 1980 : int32 cmpresult = 0;
3897 : bool result;
3898 :
3899 : /* First subkey should be same as the header says */
3900 : Assert(subkey->sk_attno == skey->sk_attno);
3901 :
3902 : /* Loop over columns of the row condition */
3903 : for (;;)
3904 156 : {
3905 : Datum datum;
3906 : bool isNull;
3907 :
3908 : Assert(subkey->sk_flags & SK_ROW_MEMBER);
3909 :
3910 2136 : if (subkey->sk_attno > tupnatts)
3911 : {
3912 : /*
3913 : * This attribute is truncated (must be high key). The value for
3914 : * this attribute in the first non-pivot tuple on the page to the
3915 : * right could be any possible value. Assume that truncated
3916 : * attribute passes the qual.
3917 : */
3918 : Assert(BTreeTupleIsPivot(tuple));
3919 6 : cmpresult = 0;
3920 6 : if (subkey->sk_flags & SK_ROW_END)
3921 6 : break;
3922 0 : subkey++;
3923 0 : continue;
3924 : }
3925 :
3926 2130 : datum = index_getattr(tuple,
3927 2130 : subkey->sk_attno,
3928 : tupdesc,
3929 : &isNull);
3930 :
3931 2130 : if (isNull)
3932 : {
3933 48 : if (subkey->sk_flags & SK_BT_NULLS_FIRST)
3934 : {
3935 : /*
3936 : * Since NULLs are sorted before non-NULLs, we know we have
3937 : * reached the lower limit of the range of values for this
3938 : * index attr. On a backward scan, we can stop if this qual
3939 : * is one of the "must match" subset. We can stop regardless
3940 : * of whether the qual is > or <, so long as it's required,
3941 : * because it's not possible for any future tuples to pass. On
3942 : * a forward scan, however, we must keep going, because we may
3943 : * have initially positioned to the start of the index.
3944 : * (_bt_advance_array_keys also relies on this behavior during
3945 : * forward scans.)
3946 : */
3947 0 : if ((subkey->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD)) &&
3948 : ScanDirectionIsBackward(dir))
3949 0 : *continuescan = false;
3950 : }
3951 : else
3952 : {
3953 : /*
3954 : * Since NULLs are sorted after non-NULLs, we know we have
3955 : * reached the upper limit of the range of values for this
3956 : * index attr. On a forward scan, we can stop if this qual is
3957 : * one of the "must match" subset. We can stop regardless of
3958 : * whether the qual is > or <, so long as it's required,
3959 : * because it's not possible for any future tuples to pass. On
3960 : * a backward scan, however, we must keep going, because we
3961 : * may have initially positioned to the end of the index.
3962 : * (_bt_advance_array_keys also relies on this behavior during
3963 : * backward scans.)
3964 : */
3965 48 : if ((subkey->sk_flags & (SK_BT_REQFWD | SK_BT_REQBKWD)) &&
3966 : ScanDirectionIsForward(dir))
3967 0 : *continuescan = false;
3968 : }
3969 :
3970 : /*
3971 : * In any case, this indextuple doesn't match the qual.
3972 : */
3973 48 : return false;
3974 : }
3975 :
3976 2082 : if (subkey->sk_flags & SK_ISNULL)
3977 : {
3978 : /*
3979 : * Unlike the simple-scankey case, this isn't a disallowed case.
3980 : * But it can never match. If all the earlier row comparison
3981 : * columns are required for the scan direction, we can stop the
3982 : * scan, because there can't be another tuple that will succeed.
3983 : */
3984 0 : if (subkey != (ScanKey) DatumGetPointer(skey->sk_argument))
3985 0 : subkey--;
3986 0 : if ((subkey->sk_flags & SK_BT_REQFWD) &&
3987 : ScanDirectionIsForward(dir))
3988 0 : *continuescan = false;
3989 0 : else if ((subkey->sk_flags & SK_BT_REQBKWD) &&
3990 : ScanDirectionIsBackward(dir))
3991 0 : *continuescan = false;
3992 0 : return false;
3993 : }
3994 :
3995 : /* Perform the test --- three-way comparison not bool operator */
3996 2082 : cmpresult = DatumGetInt32(FunctionCall2Coll(&subkey->sk_func,
3997 : subkey->sk_collation,
3998 : datum,
3999 : subkey->sk_argument));
4000 :
4001 2082 : if (subkey->sk_flags & SK_BT_DESC)
4002 0 : INVERT_COMPARE_RESULT(cmpresult);
4003 :
4004 : /* Done comparing if unequal, else advance to next column */
4005 2082 : if (cmpresult != 0)
4006 1926 : break;
4007 :
4008 156 : if (subkey->sk_flags & SK_ROW_END)
4009 0 : break;
4010 156 : subkey++;
4011 : }
4012 :
4013 : /*
4014 : * At this point cmpresult indicates the overall result of the row
4015 : * comparison, and subkey points to the deciding column (or the last
4016 : * column if the result is "=").
4017 : */
4018 1932 : switch (subkey->sk_strategy)
4019 : {
4020 : /* EQ and NE cases aren't allowed here */
4021 0 : case BTLessStrategyNumber:
4022 0 : result = (cmpresult < 0);
4023 0 : break;
4024 1590 : case BTLessEqualStrategyNumber:
4025 1590 : result = (cmpresult <= 0);
4026 1590 : break;
4027 240 : case BTGreaterEqualStrategyNumber:
4028 240 : result = (cmpresult >= 0);
4029 240 : break;
4030 102 : case BTGreaterStrategyNumber:
4031 102 : result = (cmpresult > 0);
4032 102 : break;
4033 0 : default:
4034 0 : elog(ERROR, "unrecognized RowCompareType: %d",
4035 : (int) subkey->sk_strategy);
4036 : result = 0; /* keep compiler quiet */
4037 : break;
4038 : }
4039 :
4040 1932 : if (!result)
4041 : {
4042 : /*
4043 : * Tuple fails this qual. If it's a required qual for the current
4044 : * scan direction, then we can conclude no further tuples will pass,
4045 : * either. Note we have to look at the deciding column, not
4046 : * necessarily the first or last column of the row condition.
4047 : */
4048 6 : if ((subkey->sk_flags & SK_BT_REQFWD) &&
4049 : ScanDirectionIsForward(dir))
4050 6 : *continuescan = false;
4051 0 : else if ((subkey->sk_flags & SK_BT_REQBKWD) &&
4052 : ScanDirectionIsBackward(dir))
4053 0 : *continuescan = false;
4054 : }
4055 :
4056 1932 : return result;
4057 : }
4058 :
4059 : /*
4060 : * Determine if a scan with array keys should skip over uninteresting tuples.
4061 : *
4062 : * This is a subroutine for _bt_checkkeys. Called when _bt_readpage's linear
4063 : * search process (started after it finishes reading an initial group of
4064 : * matching tuples, used to locate the start of the next group of tuples
4065 : * matching the next set of required array keys) has already scanned an
4066 : * excessive number of tuples whose key space is "between arrays".
4067 : *
4068 : * When we perform look ahead successfully, we'll sets pstate.skip, which
4069 : * instructs _bt_readpage to skip ahead to that tuple next (could be past the
4070 : * end of the scan's leaf page). Pages where the optimization is effective
4071 : * will generally still need to skip several times. Each call here performs
4072 : * only a single "look ahead" comparison of a later tuple, whose distance from
4073 : * the current tuple's offset number is determined by applying heuristics.
4074 : */
4075 : static void
4076 986 : _bt_checkkeys_look_ahead(IndexScanDesc scan, BTReadPageState *pstate,
4077 : int tupnatts, TupleDesc tupdesc)
4078 : {
4079 986 : ScanDirection dir = pstate->dir;
4080 : OffsetNumber aheadoffnum;
4081 : IndexTuple ahead;
4082 :
4083 : /* Avoid looking ahead when comparing the page high key */
4084 986 : if (pstate->offnum < pstate->minoff)
4085 0 : return;
4086 :
4087 : /*
4088 : * Don't look ahead when there aren't enough tuples remaining on the page
4089 : * (in the current scan direction) for it to be worth our while
4090 : */
4091 986 : if (ScanDirectionIsForward(dir) &&
4092 980 : pstate->offnum >= pstate->maxoff - LOOK_AHEAD_DEFAULT_DISTANCE)
4093 0 : return;
4094 986 : else if (ScanDirectionIsBackward(dir) &&
4095 6 : pstate->offnum <= pstate->minoff + LOOK_AHEAD_DEFAULT_DISTANCE)
4096 0 : return;
4097 :
4098 : /*
4099 : * The look ahead distance starts small, and ramps up as each call here
4100 : * allows _bt_readpage to skip over more tuples
4101 : */
4102 986 : if (!pstate->targetdistance)
4103 372 : pstate->targetdistance = LOOK_AHEAD_DEFAULT_DISTANCE;
4104 : else
4105 614 : pstate->targetdistance *= 2;
4106 :
4107 : /* Don't read past the end (or before the start) of the page, though */
4108 986 : if (ScanDirectionIsForward(dir))
4109 980 : aheadoffnum = Min((int) pstate->maxoff,
4110 : (int) pstate->offnum + pstate->targetdistance);
4111 : else
4112 6 : aheadoffnum = Max((int) pstate->minoff,
4113 : (int) pstate->offnum - pstate->targetdistance);
4114 :
4115 986 : ahead = (IndexTuple) PageGetItem(pstate->page,
4116 : PageGetItemId(pstate->page, aheadoffnum));
4117 986 : if (_bt_tuple_before_array_skeys(scan, dir, ahead, tupdesc, tupnatts,
4118 : false, 0, NULL))
4119 : {
4120 : /*
4121 : * Success -- instruct _bt_readpage to skip ahead to very next tuple
4122 : * after the one we determined was still before the current array keys
4123 : */
4124 498 : if (ScanDirectionIsForward(dir))
4125 492 : pstate->skip = aheadoffnum + 1;
4126 : else
4127 6 : pstate->skip = aheadoffnum - 1;
4128 : }
4129 : else
4130 : {
4131 : /*
4132 : * Failure -- "ahead" tuple is too far ahead (we were too aggressive).
4133 : *
4134 : * Reset the number of rechecks, and aggressively reduce the target
4135 : * distance (we're much more aggressive here than we were when the
4136 : * distance was initially ramped up).
4137 : */
4138 488 : pstate->rechecks = 0;
4139 488 : pstate->targetdistance = Max(pstate->targetdistance / 8, 1);
4140 : }
4141 : }
4142 :
4143 : /*
4144 : * _bt_killitems - set LP_DEAD state for items an indexscan caller has
4145 : * told us were killed
4146 : *
4147 : * scan->opaque, referenced locally through so, contains information about the
4148 : * current page and killed tuples thereon (generally, this should only be
4149 : * called if so->numKilled > 0).
4150 : *
4151 : * The caller does not have a lock on the page and may or may not have the
4152 : * page pinned in a buffer. Note that read-lock is sufficient for setting
4153 : * LP_DEAD status (which is only a hint).
4154 : *
4155 : * We match items by heap TID before assuming they are the right ones to
4156 : * delete. We cope with cases where items have moved right due to insertions.
4157 : * If an item has moved off the current page due to a split, we'll fail to
4158 : * find it and do nothing (this is not an error case --- we assume the item
4159 : * will eventually get marked in a future indexscan).
4160 : *
4161 : * Note that if we hold a pin on the target page continuously from initially
4162 : * reading the items until applying this function, VACUUM cannot have deleted
4163 : * any items from the page, and so there is no need to search left from the
4164 : * recorded offset. (This observation also guarantees that the item is still
4165 : * the right one to delete, which might otherwise be questionable since heap
4166 : * TIDs can get recycled.) This holds true even if the page has been modified
4167 : * by inserts and page splits, so there is no need to consult the LSN.
4168 : *
4169 : * If the pin was released after reading the page, then we re-read it. If it
4170 : * has been modified since we read it (as determined by the LSN), we dare not
4171 : * flag any entries because it is possible that the old entry was vacuumed
4172 : * away and the TID was re-used by a completely different heap tuple.
4173 : */
4174 : void
4175 137956 : _bt_killitems(IndexScanDesc scan)
4176 : {
4177 137956 : BTScanOpaque so = (BTScanOpaque) scan->opaque;
4178 : Page page;
4179 : BTPageOpaque opaque;
4180 : OffsetNumber minoff;
4181 : OffsetNumber maxoff;
4182 : int i;
4183 137956 : int numKilled = so->numKilled;
4184 137956 : bool killedsomething = false;
4185 : bool droppedpin PG_USED_FOR_ASSERTS_ONLY;
4186 :
4187 : Assert(BTScanPosIsValid(so->currPos));
4188 :
4189 : /*
4190 : * Always reset the scan state, so we don't look for same items on other
4191 : * pages.
4192 : */
4193 137956 : so->numKilled = 0;
4194 :
4195 137956 : if (BTScanPosIsPinned(so->currPos))
4196 : {
4197 : /*
4198 : * We have held the pin on this page since we read the index tuples,
4199 : * so all we need to do is lock it. The pin will have prevented
4200 : * re-use of any TID on the page, so there is no need to check the
4201 : * LSN.
4202 : */
4203 34432 : droppedpin = false;
4204 34432 : _bt_lockbuf(scan->indexRelation, so->currPos.buf, BT_READ);
4205 :
4206 34432 : page = BufferGetPage(so->currPos.buf);
4207 : }
4208 : else
4209 : {
4210 : Buffer buf;
4211 :
4212 103524 : droppedpin = true;
4213 : /* Attempt to re-read the buffer, getting pin and lock. */
4214 103524 : buf = _bt_getbuf(scan->indexRelation, so->currPos.currPage, BT_READ);
4215 :
4216 103524 : page = BufferGetPage(buf);
4217 103524 : if (BufferGetLSNAtomic(buf) == so->currPos.lsn)
4218 103444 : so->currPos.buf = buf;
4219 : else
4220 : {
4221 : /* Modified while not pinned means hinting is not safe. */
4222 80 : _bt_relbuf(scan->indexRelation, buf);
4223 80 : return;
4224 : }
4225 : }
4226 :
4227 137876 : opaque = BTPageGetOpaque(page);
4228 137876 : minoff = P_FIRSTDATAKEY(opaque);
4229 137876 : maxoff = PageGetMaxOffsetNumber(page);
4230 :
4231 552404 : for (i = 0; i < numKilled; i++)
4232 : {
4233 414528 : int itemIndex = so->killedItems[i];
4234 414528 : BTScanPosItem *kitem = &so->currPos.items[itemIndex];
4235 414528 : OffsetNumber offnum = kitem->indexOffset;
4236 :
4237 : Assert(itemIndex >= so->currPos.firstItem &&
4238 : itemIndex <= so->currPos.lastItem);
4239 414528 : if (offnum < minoff)
4240 0 : continue; /* pure paranoia */
4241 6333048 : while (offnum <= maxoff)
4242 : {
4243 6275348 : ItemId iid = PageGetItemId(page, offnum);
4244 6275348 : IndexTuple ituple = (IndexTuple) PageGetItem(page, iid);
4245 6275348 : bool killtuple = false;
4246 :
4247 6275348 : if (BTreeTupleIsPosting(ituple))
4248 : {
4249 2456376 : int pi = i + 1;
4250 2456376 : int nposting = BTreeTupleGetNPosting(ituple);
4251 : int j;
4252 :
4253 : /*
4254 : * We rely on the convention that heap TIDs in the scanpos
4255 : * items array are stored in ascending heap TID order for a
4256 : * group of TIDs that originally came from a posting list
4257 : * tuple. This convention even applies during backwards
4258 : * scans, where returning the TIDs in descending order might
4259 : * seem more natural. This is about effectiveness, not
4260 : * correctness.
4261 : *
4262 : * Note that the page may have been modified in almost any way
4263 : * since we first read it (in the !droppedpin case), so it's
4264 : * possible that this posting list tuple wasn't a posting list
4265 : * tuple when we first encountered its heap TIDs.
4266 : */
4267 2520786 : for (j = 0; j < nposting; j++)
4268 : {
4269 2518048 : ItemPointer item = BTreeTupleGetPostingN(ituple, j);
4270 :
4271 2518048 : if (!ItemPointerEquals(item, &kitem->heapTid))
4272 2453638 : break; /* out of posting list loop */
4273 :
4274 : /*
4275 : * kitem must have matching offnum when heap TIDs match,
4276 : * though only in the common case where the page can't
4277 : * have been concurrently modified
4278 : */
4279 : Assert(kitem->indexOffset == offnum || !droppedpin);
4280 :
4281 : /*
4282 : * Read-ahead to later kitems here.
4283 : *
4284 : * We rely on the assumption that not advancing kitem here
4285 : * will prevent us from considering the posting list tuple
4286 : * fully dead by not matching its next heap TID in next
4287 : * loop iteration.
4288 : *
4289 : * If, on the other hand, this is the final heap TID in
4290 : * the posting list tuple, then tuple gets killed
4291 : * regardless (i.e. we handle the case where the last
4292 : * kitem is also the last heap TID in the last index tuple
4293 : * correctly -- posting tuple still gets killed).
4294 : */
4295 64410 : if (pi < numKilled)
4296 44566 : kitem = &so->currPos.items[so->killedItems[pi++]];
4297 : }
4298 :
4299 : /*
4300 : * Don't bother advancing the outermost loop's int iterator to
4301 : * avoid processing killed items that relate to the same
4302 : * offnum/posting list tuple. This micro-optimization hardly
4303 : * seems worth it. (Further iterations of the outermost loop
4304 : * will fail to match on this same posting list's first heap
4305 : * TID instead, so we'll advance to the next offnum/index
4306 : * tuple pretty quickly.)
4307 : */
4308 2456376 : if (j == nposting)
4309 2738 : killtuple = true;
4310 : }
4311 3818972 : else if (ItemPointerEquals(&ituple->t_tid, &kitem->heapTid))
4312 354624 : killtuple = true;
4313 :
4314 : /*
4315 : * Mark index item as dead, if it isn't already. Since this
4316 : * happens while holding a buffer lock possibly in shared mode,
4317 : * it's possible that multiple processes attempt to do this
4318 : * simultaneously, leading to multiple full-page images being sent
4319 : * to WAL (if wal_log_hints or data checksums are enabled), which
4320 : * is undesirable.
4321 : */
4322 6275348 : if (killtuple && !ItemIdIsDead(iid))
4323 : {
4324 : /* found the item/all posting list items */
4325 356828 : ItemIdMarkDead(iid);
4326 356828 : killedsomething = true;
4327 356828 : break; /* out of inner search loop */
4328 : }
4329 5918520 : offnum = OffsetNumberNext(offnum);
4330 : }
4331 : }
4332 :
4333 : /*
4334 : * Since this can be redone later if needed, mark as dirty hint.
4335 : *
4336 : * Whenever we mark anything LP_DEAD, we also set the page's
4337 : * BTP_HAS_GARBAGE flag, which is likewise just a hint. (Note that we
4338 : * only rely on the page-level flag in !heapkeyspace indexes.)
4339 : */
4340 137876 : if (killedsomething)
4341 : {
4342 117128 : opaque->btpo_flags |= BTP_HAS_GARBAGE;
4343 117128 : MarkBufferDirtyHint(so->currPos.buf, true);
4344 : }
4345 :
4346 137876 : _bt_unlockbuf(scan->indexRelation, so->currPos.buf);
4347 : }
4348 :
4349 :
4350 : /*
4351 : * The following routines manage a shared-memory area in which we track
4352 : * assignment of "vacuum cycle IDs" to currently-active btree vacuuming
4353 : * operations. There is a single counter which increments each time we
4354 : * start a vacuum to assign it a cycle ID. Since multiple vacuums could
4355 : * be active concurrently, we have to track the cycle ID for each active
4356 : * vacuum; this requires at most MaxBackends entries (usually far fewer).
4357 : * We assume at most one vacuum can be active for a given index.
4358 : *
4359 : * Access to the shared memory area is controlled by BtreeVacuumLock.
4360 : * In principle we could use a separate lmgr locktag for each index,
4361 : * but a single LWLock is much cheaper, and given the short time that
4362 : * the lock is ever held, the concurrency hit should be minimal.
4363 : */
4364 :
4365 : typedef struct BTOneVacInfo
4366 : {
4367 : LockRelId relid; /* global identifier of an index */
4368 : BTCycleId cycleid; /* cycle ID for its active VACUUM */
4369 : } BTOneVacInfo;
4370 :
4371 : typedef struct BTVacInfo
4372 : {
4373 : BTCycleId cycle_ctr; /* cycle ID most recently assigned */
4374 : int num_vacuums; /* number of currently active VACUUMs */
4375 : int max_vacuums; /* allocated length of vacuums[] array */
4376 : BTOneVacInfo vacuums[FLEXIBLE_ARRAY_MEMBER];
4377 : } BTVacInfo;
4378 :
4379 : static BTVacInfo *btvacinfo;
4380 :
4381 :
4382 : /*
4383 : * _bt_vacuum_cycleid --- get the active vacuum cycle ID for an index,
4384 : * or zero if there is no active VACUUM
4385 : *
4386 : * Note: for correct interlocking, the caller must already hold pin and
4387 : * exclusive lock on each buffer it will store the cycle ID into. This
4388 : * ensures that even if a VACUUM starts immediately afterwards, it cannot
4389 : * process those pages until the page split is complete.
4390 : */
4391 : BTCycleId
4392 19944 : _bt_vacuum_cycleid(Relation rel)
4393 : {
4394 19944 : BTCycleId result = 0;
4395 : int i;
4396 :
4397 : /* Share lock is enough since this is a read-only operation */
4398 19944 : LWLockAcquire(BtreeVacuumLock, LW_SHARED);
4399 :
4400 19944 : for (i = 0; i < btvacinfo->num_vacuums; i++)
4401 : {
4402 0 : BTOneVacInfo *vac = &btvacinfo->vacuums[i];
4403 :
4404 0 : if (vac->relid.relId == rel->rd_lockInfo.lockRelId.relId &&
4405 0 : vac->relid.dbId == rel->rd_lockInfo.lockRelId.dbId)
4406 : {
4407 0 : result = vac->cycleid;
4408 0 : break;
4409 : }
4410 : }
4411 :
4412 19944 : LWLockRelease(BtreeVacuumLock);
4413 19944 : return result;
4414 : }
4415 :
4416 : /*
4417 : * _bt_start_vacuum --- assign a cycle ID to a just-starting VACUUM operation
4418 : *
4419 : * Note: the caller must guarantee that it will eventually call
4420 : * _bt_end_vacuum, else we'll permanently leak an array slot. To ensure
4421 : * that this happens even in elog(FATAL) scenarios, the appropriate coding
4422 : * is not just a PG_TRY, but
4423 : * PG_ENSURE_ERROR_CLEANUP(_bt_end_vacuum_callback, PointerGetDatum(rel))
4424 : */
4425 : BTCycleId
4426 2156 : _bt_start_vacuum(Relation rel)
4427 : {
4428 : BTCycleId result;
4429 : int i;
4430 : BTOneVacInfo *vac;
4431 :
4432 2156 : LWLockAcquire(BtreeVacuumLock, LW_EXCLUSIVE);
4433 :
4434 : /*
4435 : * Assign the next cycle ID, being careful to avoid zero as well as the
4436 : * reserved high values.
4437 : */
4438 2156 : result = ++(btvacinfo->cycle_ctr);
4439 2156 : if (result == 0 || result > MAX_BT_CYCLE_ID)
4440 0 : result = btvacinfo->cycle_ctr = 1;
4441 :
4442 : /* Let's just make sure there's no entry already for this index */
4443 2156 : for (i = 0; i < btvacinfo->num_vacuums; i++)
4444 : {
4445 0 : vac = &btvacinfo->vacuums[i];
4446 0 : if (vac->relid.relId == rel->rd_lockInfo.lockRelId.relId &&
4447 0 : vac->relid.dbId == rel->rd_lockInfo.lockRelId.dbId)
4448 : {
4449 : /*
4450 : * Unlike most places in the backend, we have to explicitly
4451 : * release our LWLock before throwing an error. This is because
4452 : * we expect _bt_end_vacuum() to be called before transaction
4453 : * abort cleanup can run to release LWLocks.
4454 : */
4455 0 : LWLockRelease(BtreeVacuumLock);
4456 0 : elog(ERROR, "multiple active vacuums for index \"%s\"",
4457 : RelationGetRelationName(rel));
4458 : }
4459 : }
4460 :
4461 : /* OK, add an entry */
4462 2156 : if (btvacinfo->num_vacuums >= btvacinfo->max_vacuums)
4463 : {
4464 0 : LWLockRelease(BtreeVacuumLock);
4465 0 : elog(ERROR, "out of btvacinfo slots");
4466 : }
4467 2156 : vac = &btvacinfo->vacuums[btvacinfo->num_vacuums];
4468 2156 : vac->relid = rel->rd_lockInfo.lockRelId;
4469 2156 : vac->cycleid = result;
4470 2156 : btvacinfo->num_vacuums++;
4471 :
4472 2156 : LWLockRelease(BtreeVacuumLock);
4473 2156 : return result;
4474 : }
4475 :
4476 : /*
4477 : * _bt_end_vacuum --- mark a btree VACUUM operation as done
4478 : *
4479 : * Note: this is deliberately coded not to complain if no entry is found;
4480 : * this allows the caller to put PG_TRY around the start_vacuum operation.
4481 : */
4482 : void
4483 2156 : _bt_end_vacuum(Relation rel)
4484 : {
4485 : int i;
4486 :
4487 2156 : LWLockAcquire(BtreeVacuumLock, LW_EXCLUSIVE);
4488 :
4489 : /* Find the array entry */
4490 2156 : for (i = 0; i < btvacinfo->num_vacuums; i++)
4491 : {
4492 2156 : BTOneVacInfo *vac = &btvacinfo->vacuums[i];
4493 :
4494 2156 : if (vac->relid.relId == rel->rd_lockInfo.lockRelId.relId &&
4495 2156 : vac->relid.dbId == rel->rd_lockInfo.lockRelId.dbId)
4496 : {
4497 : /* Remove it by shifting down the last entry */
4498 2156 : *vac = btvacinfo->vacuums[btvacinfo->num_vacuums - 1];
4499 2156 : btvacinfo->num_vacuums--;
4500 2156 : break;
4501 : }
4502 : }
4503 :
4504 2156 : LWLockRelease(BtreeVacuumLock);
4505 2156 : }
4506 :
4507 : /*
4508 : * _bt_end_vacuum wrapped as an on_shmem_exit callback function
4509 : */
4510 : void
4511 0 : _bt_end_vacuum_callback(int code, Datum arg)
4512 : {
4513 0 : _bt_end_vacuum((Relation) DatumGetPointer(arg));
4514 0 : }
4515 :
4516 : /*
4517 : * BTreeShmemSize --- report amount of shared memory space needed
4518 : */
4519 : Size
4520 5066 : BTreeShmemSize(void)
4521 : {
4522 : Size size;
4523 :
4524 5066 : size = offsetof(BTVacInfo, vacuums);
4525 5066 : size = add_size(size, mul_size(MaxBackends, sizeof(BTOneVacInfo)));
4526 5066 : return size;
4527 : }
4528 :
4529 : /*
4530 : * BTreeShmemInit --- initialize this module's shared memory
4531 : */
4532 : void
4533 1768 : BTreeShmemInit(void)
4534 : {
4535 : bool found;
4536 :
4537 1768 : btvacinfo = (BTVacInfo *) ShmemInitStruct("BTree Vacuum State",
4538 : BTreeShmemSize(),
4539 : &found);
4540 :
4541 1768 : if (!IsUnderPostmaster)
4542 : {
4543 : /* Initialize shared memory area */
4544 : Assert(!found);
4545 :
4546 : /*
4547 : * It doesn't really matter what the cycle counter starts at, but
4548 : * having it always start the same doesn't seem good. Seed with
4549 : * low-order bits of time() instead.
4550 : */
4551 1768 : btvacinfo->cycle_ctr = (BTCycleId) time(NULL);
4552 :
4553 1768 : btvacinfo->num_vacuums = 0;
4554 1768 : btvacinfo->max_vacuums = MaxBackends;
4555 : }
4556 : else
4557 : Assert(found);
4558 1768 : }
4559 :
4560 : bytea *
4561 270 : btoptions(Datum reloptions, bool validate)
4562 : {
4563 : static const relopt_parse_elt tab[] = {
4564 : {"fillfactor", RELOPT_TYPE_INT, offsetof(BTOptions, fillfactor)},
4565 : {"vacuum_cleanup_index_scale_factor", RELOPT_TYPE_REAL,
4566 : offsetof(BTOptions, vacuum_cleanup_index_scale_factor)},
4567 : {"deduplicate_items", RELOPT_TYPE_BOOL,
4568 : offsetof(BTOptions, deduplicate_items)}
4569 : };
4570 :
4571 270 : return (bytea *) build_reloptions(reloptions, validate,
4572 : RELOPT_KIND_BTREE,
4573 : sizeof(BTOptions),
4574 : tab, lengthof(tab));
4575 : }
4576 :
4577 : /*
4578 : * btproperty() -- Check boolean properties of indexes.
4579 : *
4580 : * This is optional, but handling AMPROP_RETURNABLE here saves opening the rel
4581 : * to call btcanreturn.
4582 : */
4583 : bool
4584 756 : btproperty(Oid index_oid, int attno,
4585 : IndexAMProperty prop, const char *propname,
4586 : bool *res, bool *isnull)
4587 : {
4588 756 : switch (prop)
4589 : {
4590 42 : case AMPROP_RETURNABLE:
4591 : /* answer only for columns, not AM or whole index */
4592 42 : if (attno == 0)
4593 12 : return false;
4594 : /* otherwise, btree can always return data */
4595 30 : *res = true;
4596 30 : return true;
4597 :
4598 714 : default:
4599 714 : return false; /* punt to generic code */
4600 : }
4601 : }
4602 :
4603 : /*
4604 : * btbuildphasename() -- Return name of index build phase.
4605 : */
4606 : char *
4607 0 : btbuildphasename(int64 phasenum)
4608 : {
4609 0 : switch (phasenum)
4610 : {
4611 0 : case PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE:
4612 0 : return "initializing";
4613 0 : case PROGRESS_BTREE_PHASE_INDEXBUILD_TABLESCAN:
4614 0 : return "scanning table";
4615 0 : case PROGRESS_BTREE_PHASE_PERFORMSORT_1:
4616 0 : return "sorting live tuples";
4617 0 : case PROGRESS_BTREE_PHASE_PERFORMSORT_2:
4618 0 : return "sorting dead tuples";
4619 0 : case PROGRESS_BTREE_PHASE_LEAF_LOAD:
4620 0 : return "loading tuples in tree";
4621 0 : default:
4622 0 : return NULL;
4623 : }
4624 : }
4625 :
4626 : /*
4627 : * _bt_truncate() -- create tuple without unneeded suffix attributes.
4628 : *
4629 : * Returns truncated pivot index tuple allocated in caller's memory context,
4630 : * with key attributes copied from caller's firstright argument. If rel is
4631 : * an INCLUDE index, non-key attributes will definitely be truncated away,
4632 : * since they're not part of the key space. More aggressive suffix
4633 : * truncation can take place when it's clear that the returned tuple does not
4634 : * need one or more suffix key attributes. We only need to keep firstright
4635 : * attributes up to and including the first non-lastleft-equal attribute.
4636 : * Caller's insertion scankey is used to compare the tuples; the scankey's
4637 : * argument values are not considered here.
4638 : *
4639 : * Note that returned tuple's t_tid offset will hold the number of attributes
4640 : * present, so the original item pointer offset is not represented. Caller
4641 : * should only change truncated tuple's downlink. Note also that truncated
4642 : * key attributes are treated as containing "minus infinity" values by
4643 : * _bt_compare().
4644 : *
4645 : * In the worst case (when a heap TID must be appended to distinguish lastleft
4646 : * from firstright), the size of the returned tuple is the size of firstright
4647 : * plus the size of an additional MAXALIGN()'d item pointer. This guarantee
4648 : * is important, since callers need to stay under the 1/3 of a page
4649 : * restriction on tuple size. If this routine is ever taught to truncate
4650 : * within an attribute/datum, it will need to avoid returning an enlarged
4651 : * tuple to caller when truncation + TOAST compression ends up enlarging the
4652 : * final datum.
4653 : */
4654 : IndexTuple
4655 56998 : _bt_truncate(Relation rel, IndexTuple lastleft, IndexTuple firstright,
4656 : BTScanInsert itup_key)
4657 : {
4658 56998 : TupleDesc itupdesc = RelationGetDescr(rel);
4659 56998 : int16 nkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
4660 : int keepnatts;
4661 : IndexTuple pivot;
4662 : IndexTuple tidpivot;
4663 : ItemPointer pivotheaptid;
4664 : Size newsize;
4665 :
4666 : /*
4667 : * We should only ever truncate non-pivot tuples from leaf pages. It's
4668 : * never okay to truncate when splitting an internal page.
4669 : */
4670 : Assert(!BTreeTupleIsPivot(lastleft) && !BTreeTupleIsPivot(firstright));
4671 :
4672 : /* Determine how many attributes must be kept in truncated tuple */
4673 56998 : keepnatts = _bt_keep_natts(rel, lastleft, firstright, itup_key);
4674 :
4675 : #ifdef DEBUG_NO_TRUNCATE
4676 : /* Force truncation to be ineffective for testing purposes */
4677 : keepnatts = nkeyatts + 1;
4678 : #endif
4679 :
4680 56998 : pivot = index_truncate_tuple(itupdesc, firstright,
4681 : Min(keepnatts, nkeyatts));
4682 :
4683 56998 : if (BTreeTupleIsPosting(pivot))
4684 : {
4685 : /*
4686 : * index_truncate_tuple() just returns a straight copy of firstright
4687 : * when it has no attributes to truncate. When that happens, we may
4688 : * need to truncate away a posting list here instead.
4689 : */
4690 : Assert(keepnatts == nkeyatts || keepnatts == nkeyatts + 1);
4691 : Assert(IndexRelationGetNumberOfAttributes(rel) == nkeyatts);
4692 1282 : pivot->t_info &= ~INDEX_SIZE_MASK;
4693 1282 : pivot->t_info |= MAXALIGN(BTreeTupleGetPostingOffset(firstright));
4694 : }
4695 :
4696 : /*
4697 : * If there is a distinguishing key attribute within pivot tuple, we're
4698 : * done
4699 : */
4700 56998 : if (keepnatts <= nkeyatts)
4701 : {
4702 55910 : BTreeTupleSetNAtts(pivot, keepnatts, false);
4703 55910 : return pivot;
4704 : }
4705 :
4706 : /*
4707 : * We have to store a heap TID in the new pivot tuple, since no non-TID
4708 : * key attribute value in firstright distinguishes the right side of the
4709 : * split from the left side. nbtree conceptualizes this case as an
4710 : * inability to truncate away any key attributes, since heap TID is
4711 : * treated as just another key attribute (despite lacking a pg_attribute
4712 : * entry).
4713 : *
4714 : * Use enlarged space that holds a copy of pivot. We need the extra space
4715 : * to store a heap TID at the end (using the special pivot tuple
4716 : * representation). Note that the original pivot already has firstright's
4717 : * possible posting list/non-key attribute values removed at this point.
4718 : */
4719 1088 : newsize = MAXALIGN(IndexTupleSize(pivot)) + MAXALIGN(sizeof(ItemPointerData));
4720 1088 : tidpivot = palloc0(newsize);
4721 1088 : memcpy(tidpivot, pivot, MAXALIGN(IndexTupleSize(pivot)));
4722 : /* Cannot leak memory here */
4723 1088 : pfree(pivot);
4724 :
4725 : /*
4726 : * Store all of firstright's key attribute values plus a tiebreaker heap
4727 : * TID value in enlarged pivot tuple
4728 : */
4729 1088 : tidpivot->t_info &= ~INDEX_SIZE_MASK;
4730 1088 : tidpivot->t_info |= newsize;
4731 1088 : BTreeTupleSetNAtts(tidpivot, nkeyatts, true);
4732 1088 : pivotheaptid = BTreeTupleGetHeapTID(tidpivot);
4733 :
4734 : /*
4735 : * Lehman & Yao use lastleft as the leaf high key in all cases, but don't
4736 : * consider suffix truncation. It seems like a good idea to follow that
4737 : * example in cases where no truncation takes place -- use lastleft's heap
4738 : * TID. (This is also the closest value to negative infinity that's
4739 : * legally usable.)
4740 : */
4741 1088 : ItemPointerCopy(BTreeTupleGetMaxHeapTID(lastleft), pivotheaptid);
4742 :
4743 : /*
4744 : * We're done. Assert() that heap TID invariants hold before returning.
4745 : *
4746 : * Lehman and Yao require that the downlink to the right page, which is to
4747 : * be inserted into the parent page in the second phase of a page split be
4748 : * a strict lower bound on items on the right page, and a non-strict upper
4749 : * bound for items on the left page. Assert that heap TIDs follow these
4750 : * invariants, since a heap TID value is apparently needed as a
4751 : * tiebreaker.
4752 : */
4753 : #ifndef DEBUG_NO_TRUNCATE
4754 : Assert(ItemPointerCompare(BTreeTupleGetMaxHeapTID(lastleft),
4755 : BTreeTupleGetHeapTID(firstright)) < 0);
4756 : Assert(ItemPointerCompare(pivotheaptid,
4757 : BTreeTupleGetHeapTID(lastleft)) >= 0);
4758 : Assert(ItemPointerCompare(pivotheaptid,
4759 : BTreeTupleGetHeapTID(firstright)) < 0);
4760 : #else
4761 :
4762 : /*
4763 : * Those invariants aren't guaranteed to hold for lastleft + firstright
4764 : * heap TID attribute values when they're considered here only because
4765 : * DEBUG_NO_TRUNCATE is defined (a heap TID is probably not actually
4766 : * needed as a tiebreaker). DEBUG_NO_TRUNCATE must therefore use a heap
4767 : * TID value that always works as a strict lower bound for items to the
4768 : * right. In particular, it must avoid using firstright's leading key
4769 : * attribute values along with lastleft's heap TID value when lastleft's
4770 : * TID happens to be greater than firstright's TID.
4771 : */
4772 : ItemPointerCopy(BTreeTupleGetHeapTID(firstright), pivotheaptid);
4773 :
4774 : /*
4775 : * Pivot heap TID should never be fully equal to firstright. Note that
4776 : * the pivot heap TID will still end up equal to lastleft's heap TID when
4777 : * that's the only usable value.
4778 : */
4779 : ItemPointerSetOffsetNumber(pivotheaptid,
4780 : OffsetNumberPrev(ItemPointerGetOffsetNumber(pivotheaptid)));
4781 : Assert(ItemPointerCompare(pivotheaptid,
4782 : BTreeTupleGetHeapTID(firstright)) < 0);
4783 : #endif
4784 :
4785 1088 : return tidpivot;
4786 : }
4787 :
4788 : /*
4789 : * _bt_keep_natts - how many key attributes to keep when truncating.
4790 : *
4791 : * Caller provides two tuples that enclose a split point. Caller's insertion
4792 : * scankey is used to compare the tuples; the scankey's argument values are
4793 : * not considered here.
4794 : *
4795 : * This can return a number of attributes that is one greater than the
4796 : * number of key attributes for the index relation. This indicates that the
4797 : * caller must use a heap TID as a unique-ifier in new pivot tuple.
4798 : */
4799 : static int
4800 56998 : _bt_keep_natts(Relation rel, IndexTuple lastleft, IndexTuple firstright,
4801 : BTScanInsert itup_key)
4802 : {
4803 56998 : int nkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
4804 56998 : TupleDesc itupdesc = RelationGetDescr(rel);
4805 : int keepnatts;
4806 : ScanKey scankey;
4807 :
4808 : /*
4809 : * _bt_compare() treats truncated key attributes as having the value minus
4810 : * infinity, which would break searches within !heapkeyspace indexes. We
4811 : * must still truncate away non-key attribute values, though.
4812 : */
4813 56998 : if (!itup_key->heapkeyspace)
4814 0 : return nkeyatts;
4815 :
4816 56998 : scankey = itup_key->scankeys;
4817 56998 : keepnatts = 1;
4818 69130 : for (int attnum = 1; attnum <= nkeyatts; attnum++, scankey++)
4819 : {
4820 : Datum datum1,
4821 : datum2;
4822 : bool isNull1,
4823 : isNull2;
4824 :
4825 68042 : datum1 = index_getattr(lastleft, attnum, itupdesc, &isNull1);
4826 68042 : datum2 = index_getattr(firstright, attnum, itupdesc, &isNull2);
4827 :
4828 68042 : if (isNull1 != isNull2)
4829 55910 : break;
4830 :
4831 136054 : if (!isNull1 &&
4832 68012 : DatumGetInt32(FunctionCall2Coll(&scankey->sk_func,
4833 : scankey->sk_collation,
4834 : datum1,
4835 : datum2)) != 0)
4836 55910 : break;
4837 :
4838 12132 : keepnatts++;
4839 : }
4840 :
4841 : /*
4842 : * Assert that _bt_keep_natts_fast() agrees with us in passing. This is
4843 : * expected in an allequalimage index.
4844 : */
4845 : Assert(!itup_key->allequalimage ||
4846 : keepnatts == _bt_keep_natts_fast(rel, lastleft, firstright));
4847 :
4848 56998 : return keepnatts;
4849 : }
4850 :
4851 : /*
4852 : * _bt_keep_natts_fast - fast bitwise variant of _bt_keep_natts.
4853 : *
4854 : * This is exported so that a candidate split point can have its effect on
4855 : * suffix truncation inexpensively evaluated ahead of time when finding a
4856 : * split location. A naive bitwise approach to datum comparisons is used to
4857 : * save cycles.
4858 : *
4859 : * The approach taken here usually provides the same answer as _bt_keep_natts
4860 : * will (for the same pair of tuples from a heapkeyspace index), since the
4861 : * majority of btree opclasses can never indicate that two datums are equal
4862 : * unless they're bitwise equal after detoasting. When an index only has
4863 : * "equal image" columns, routine is guaranteed to give the same result as
4864 : * _bt_keep_natts would.
4865 : *
4866 : * Callers can rely on the fact that attributes considered equal here are
4867 : * definitely also equal according to _bt_keep_natts, even when the index uses
4868 : * an opclass or collation that is not "allequalimage"/deduplication-safe.
4869 : * This weaker guarantee is good enough for nbtsplitloc.c caller, since false
4870 : * negatives generally only have the effect of making leaf page splits use a
4871 : * more balanced split point.
4872 : */
4873 : int
4874 12445386 : _bt_keep_natts_fast(Relation rel, IndexTuple lastleft, IndexTuple firstright)
4875 : {
4876 12445386 : TupleDesc itupdesc = RelationGetDescr(rel);
4877 12445386 : int keysz = IndexRelationGetNumberOfKeyAttributes(rel);
4878 : int keepnatts;
4879 :
4880 12445386 : keepnatts = 1;
4881 20538064 : for (int attnum = 1; attnum <= keysz; attnum++)
4882 : {
4883 : Datum datum1,
4884 : datum2;
4885 : bool isNull1,
4886 : isNull2;
4887 : Form_pg_attribute att;
4888 :
4889 18356132 : datum1 = index_getattr(lastleft, attnum, itupdesc, &isNull1);
4890 18356132 : datum2 = index_getattr(firstright, attnum, itupdesc, &isNull2);
4891 18356132 : att = TupleDescAttr(itupdesc, attnum - 1);
4892 :
4893 18356132 : if (isNull1 != isNull2)
4894 10263454 : break;
4895 :
4896 18355988 : if (!isNull1 &&
4897 18308932 : !datum_image_eq(datum1, datum2, att->attbyval, att->attlen))
4898 10263310 : break;
4899 :
4900 8092678 : keepnatts++;
4901 : }
4902 :
4903 12445386 : return keepnatts;
4904 : }
4905 :
4906 : /*
4907 : * _bt_check_natts() -- Verify tuple has expected number of attributes.
4908 : *
4909 : * Returns value indicating if the expected number of attributes were found
4910 : * for a particular offset on page. This can be used as a general purpose
4911 : * sanity check.
4912 : *
4913 : * Testing a tuple directly with BTreeTupleGetNAtts() should generally be
4914 : * preferred to calling here. That's usually more convenient, and is always
4915 : * more explicit. Call here instead when offnum's tuple may be a negative
4916 : * infinity tuple that uses the pre-v11 on-disk representation, or when a low
4917 : * context check is appropriate. This routine is as strict as possible about
4918 : * what is expected on each version of btree.
4919 : */
4920 : bool
4921 3993638 : _bt_check_natts(Relation rel, bool heapkeyspace, Page page, OffsetNumber offnum)
4922 : {
4923 3993638 : int16 natts = IndexRelationGetNumberOfAttributes(rel);
4924 3993638 : int16 nkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
4925 3993638 : BTPageOpaque opaque = BTPageGetOpaque(page);
4926 : IndexTuple itup;
4927 : int tupnatts;
4928 :
4929 : /*
4930 : * We cannot reliably test a deleted or half-dead page, since they have
4931 : * dummy high keys
4932 : */
4933 3993638 : if (P_IGNORE(opaque))
4934 0 : return true;
4935 :
4936 : Assert(offnum >= FirstOffsetNumber &&
4937 : offnum <= PageGetMaxOffsetNumber(page));
4938 :
4939 3993638 : itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offnum));
4940 3993638 : tupnatts = BTreeTupleGetNAtts(itup, rel);
4941 :
4942 : /* !heapkeyspace indexes do not support deduplication */
4943 3993638 : if (!heapkeyspace && BTreeTupleIsPosting(itup))
4944 0 : return false;
4945 :
4946 : /* Posting list tuples should never have "pivot heap TID" bit set */
4947 3993638 : if (BTreeTupleIsPosting(itup) &&
4948 20518 : (ItemPointerGetOffsetNumberNoCheck(&itup->t_tid) &
4949 : BT_PIVOT_HEAP_TID_ATTR) != 0)
4950 0 : return false;
4951 :
4952 : /* INCLUDE indexes do not support deduplication */
4953 3993638 : if (natts != nkeyatts && BTreeTupleIsPosting(itup))
4954 0 : return false;
4955 :
4956 3993638 : if (P_ISLEAF(opaque))
4957 : {
4958 3979704 : if (offnum >= P_FIRSTDATAKEY(opaque))
4959 : {
4960 : /*
4961 : * Non-pivot tuple should never be explicitly marked as a pivot
4962 : * tuple
4963 : */
4964 3966806 : if (BTreeTupleIsPivot(itup))
4965 0 : return false;
4966 :
4967 : /*
4968 : * Leaf tuples that are not the page high key (non-pivot tuples)
4969 : * should never be truncated. (Note that tupnatts must have been
4970 : * inferred, even with a posting list tuple, because only pivot
4971 : * tuples store tupnatts directly.)
4972 : */
4973 3966806 : return tupnatts == natts;
4974 : }
4975 : else
4976 : {
4977 : /*
4978 : * Rightmost page doesn't contain a page high key, so tuple was
4979 : * checked above as ordinary leaf tuple
4980 : */
4981 : Assert(!P_RIGHTMOST(opaque));
4982 :
4983 : /*
4984 : * !heapkeyspace high key tuple contains only key attributes. Note
4985 : * that tupnatts will only have been explicitly represented in
4986 : * !heapkeyspace indexes that happen to have non-key attributes.
4987 : */
4988 12898 : if (!heapkeyspace)
4989 0 : return tupnatts == nkeyatts;
4990 :
4991 : /* Use generic heapkeyspace pivot tuple handling */
4992 : }
4993 : }
4994 : else /* !P_ISLEAF(opaque) */
4995 : {
4996 13934 : if (offnum == P_FIRSTDATAKEY(opaque))
4997 : {
4998 : /*
4999 : * The first tuple on any internal page (possibly the first after
5000 : * its high key) is its negative infinity tuple. Negative
5001 : * infinity tuples are always truncated to zero attributes. They
5002 : * are a particular kind of pivot tuple.
5003 : */
5004 1030 : if (heapkeyspace)
5005 1030 : return tupnatts == 0;
5006 :
5007 : /*
5008 : * The number of attributes won't be explicitly represented if the
5009 : * negative infinity tuple was generated during a page split that
5010 : * occurred with a version of Postgres before v11. There must be
5011 : * a problem when there is an explicit representation that is
5012 : * non-zero, or when there is no explicit representation and the
5013 : * tuple is evidently not a pre-pg_upgrade tuple.
5014 : *
5015 : * Prior to v11, downlinks always had P_HIKEY as their offset.
5016 : * Accept that as an alternative indication of a valid
5017 : * !heapkeyspace negative infinity tuple.
5018 : */
5019 0 : return tupnatts == 0 ||
5020 0 : ItemPointerGetOffsetNumber(&(itup->t_tid)) == P_HIKEY;
5021 : }
5022 : else
5023 : {
5024 : /*
5025 : * !heapkeyspace downlink tuple with separator key contains only
5026 : * key attributes. Note that tupnatts will only have been
5027 : * explicitly represented in !heapkeyspace indexes that happen to
5028 : * have non-key attributes.
5029 : */
5030 12904 : if (!heapkeyspace)
5031 0 : return tupnatts == nkeyatts;
5032 :
5033 : /* Use generic heapkeyspace pivot tuple handling */
5034 : }
5035 : }
5036 :
5037 : /* Handle heapkeyspace pivot tuples (excluding minus infinity items) */
5038 : Assert(heapkeyspace);
5039 :
5040 : /*
5041 : * Explicit representation of the number of attributes is mandatory with
5042 : * heapkeyspace index pivot tuples, regardless of whether or not there are
5043 : * non-key attributes.
5044 : */
5045 25802 : if (!BTreeTupleIsPivot(itup))
5046 0 : return false;
5047 :
5048 : /* Pivot tuple should not use posting list representation (redundant) */
5049 25802 : if (BTreeTupleIsPosting(itup))
5050 0 : return false;
5051 :
5052 : /*
5053 : * Heap TID is a tiebreaker key attribute, so it cannot be untruncated
5054 : * when any other key attribute is truncated
5055 : */
5056 25802 : if (BTreeTupleGetHeapTID(itup) != NULL && tupnatts != nkeyatts)
5057 0 : return false;
5058 :
5059 : /*
5060 : * Pivot tuple must have at least one untruncated key attribute (minus
5061 : * infinity pivot tuples are the only exception). Pivot tuples can never
5062 : * represent that there is a value present for a key attribute that
5063 : * exceeds pg_index.indnkeyatts for the index.
5064 : */
5065 25802 : return tupnatts > 0 && tupnatts <= nkeyatts;
5066 : }
5067 :
5068 : /*
5069 : *
5070 : * _bt_check_third_page() -- check whether tuple fits on a btree page at all.
5071 : *
5072 : * We actually need to be able to fit three items on every page, so restrict
5073 : * any one item to 1/3 the per-page available space. Note that itemsz should
5074 : * not include the ItemId overhead.
5075 : *
5076 : * It might be useful to apply TOAST methods rather than throw an error here.
5077 : * Using out of line storage would break assumptions made by suffix truncation
5078 : * and by contrib/amcheck, though.
5079 : */
5080 : void
5081 264 : _bt_check_third_page(Relation rel, Relation heap, bool needheaptidspace,
5082 : Page page, IndexTuple newtup)
5083 : {
5084 : Size itemsz;
5085 : BTPageOpaque opaque;
5086 :
5087 264 : itemsz = MAXALIGN(IndexTupleSize(newtup));
5088 :
5089 : /* Double check item size against limit */
5090 264 : if (itemsz <= BTMaxItemSize(page))
5091 0 : return;
5092 :
5093 : /*
5094 : * Tuple is probably too large to fit on page, but it's possible that the
5095 : * index uses version 2 or version 3, or that page is an internal page, in
5096 : * which case a slightly higher limit applies.
5097 : */
5098 264 : if (!needheaptidspace && itemsz <= BTMaxItemSizeNoHeapTid(page))
5099 264 : return;
5100 :
5101 : /*
5102 : * Internal page insertions cannot fail here, because that would mean that
5103 : * an earlier leaf level insertion that should have failed didn't
5104 : */
5105 0 : opaque = BTPageGetOpaque(page);
5106 0 : if (!P_ISLEAF(opaque))
5107 0 : elog(ERROR, "cannot insert oversized tuple of size %zu on internal page of index \"%s\"",
5108 : itemsz, RelationGetRelationName(rel));
5109 :
5110 0 : ereport(ERROR,
5111 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
5112 : errmsg("index row size %zu exceeds btree version %u maximum %zu for index \"%s\"",
5113 : itemsz,
5114 : needheaptidspace ? BTREE_VERSION : BTREE_NOVAC_VERSION,
5115 : needheaptidspace ? BTMaxItemSize(page) :
5116 : BTMaxItemSizeNoHeapTid(page),
5117 : RelationGetRelationName(rel)),
5118 : errdetail("Index row references tuple (%u,%u) in relation \"%s\".",
5119 : ItemPointerGetBlockNumber(BTreeTupleGetHeapTID(newtup)),
5120 : ItemPointerGetOffsetNumber(BTreeTupleGetHeapTID(newtup)),
5121 : RelationGetRelationName(heap)),
5122 : errhint("Values larger than 1/3 of a buffer page cannot be indexed.\n"
5123 : "Consider a function index of an MD5 hash of the value, "
5124 : "or use full text indexing."),
5125 : errtableconstraint(heap, RelationGetRelationName(rel))));
5126 : }
5127 :
5128 : /*
5129 : * Are all attributes in rel "equality is image equality" attributes?
5130 : *
5131 : * We use each attribute's BTEQUALIMAGE_PROC opclass procedure. If any
5132 : * opclass either lacks a BTEQUALIMAGE_PROC procedure or returns false, we
5133 : * return false; otherwise we return true.
5134 : *
5135 : * Returned boolean value is stored in index metapage during index builds.
5136 : * Deduplication can only be used when we return true.
5137 : */
5138 : bool
5139 53648 : _bt_allequalimage(Relation rel, bool debugmessage)
5140 : {
5141 53648 : bool allequalimage = true;
5142 :
5143 : /* INCLUDE indexes can never support deduplication */
5144 53648 : if (IndexRelationGetNumberOfAttributes(rel) !=
5145 53648 : IndexRelationGetNumberOfKeyAttributes(rel))
5146 266 : return false;
5147 :
5148 140488 : for (int i = 0; i < IndexRelationGetNumberOfKeyAttributes(rel); i++)
5149 : {
5150 87564 : Oid opfamily = rel->rd_opfamily[i];
5151 87564 : Oid opcintype = rel->rd_opcintype[i];
5152 87564 : Oid collation = rel->rd_indcollation[i];
5153 : Oid equalimageproc;
5154 :
5155 87564 : equalimageproc = get_opfamily_proc(opfamily, opcintype, opcintype,
5156 : BTEQUALIMAGE_PROC);
5157 :
5158 : /*
5159 : * If there is no BTEQUALIMAGE_PROC then deduplication is assumed to
5160 : * be unsafe. Otherwise, actually call proc and see what it says.
5161 : */
5162 87564 : if (!OidIsValid(equalimageproc) ||
5163 87126 : !DatumGetBool(OidFunctionCall1Coll(equalimageproc, collation,
5164 : ObjectIdGetDatum(opcintype))))
5165 : {
5166 458 : allequalimage = false;
5167 458 : break;
5168 : }
5169 : }
5170 :
5171 53382 : if (debugmessage)
5172 : {
5173 45308 : if (allequalimage)
5174 44850 : elog(DEBUG1, "index \"%s\" can safely use deduplication",
5175 : RelationGetRelationName(rel));
5176 : else
5177 458 : elog(DEBUG1, "index \"%s\" cannot use deduplication",
5178 : RelationGetRelationName(rel));
5179 : }
5180 :
5181 53382 : return allequalimage;
5182 : }
|