Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * spgutils.c
4 : * various support functions for SP-GiST
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/spgist/spgutils.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 :
16 : #include "postgres.h"
17 :
18 : #include "access/amvalidate.h"
19 : #include "access/htup_details.h"
20 : #include "access/reloptions.h"
21 : #include "access/spgist_private.h"
22 : #include "access/toast_compression.h"
23 : #include "access/transam.h"
24 : #include "access/xact.h"
25 : #include "catalog/pg_amop.h"
26 : #include "commands/vacuum.h"
27 : #include "nodes/nodeFuncs.h"
28 : #include "parser/parse_coerce.h"
29 : #include "storage/bufmgr.h"
30 : #include "storage/indexfsm.h"
31 : #include "utils/catcache.h"
32 : #include "utils/fmgrprotos.h"
33 : #include "utils/index_selfuncs.h"
34 : #include "utils/lsyscache.h"
35 : #include "utils/rel.h"
36 : #include "utils/syscache.h"
37 :
38 :
39 : /*
40 : * SP-GiST handler function: return IndexAmRoutine with access method parameters
41 : * and callbacks.
42 : */
43 : Datum
44 1354 : spghandler(PG_FUNCTION_ARGS)
45 : {
46 : static const IndexAmRoutine amroutine = {
47 : .type = T_IndexAmRoutine,
48 : .amstrategies = 0,
49 : .amsupport = SPGISTNProc,
50 : .amoptsprocnum = SPGIST_OPTIONS_PROC,
51 : .amcanorder = false,
52 : .amcanorderbyop = true,
53 : .amcanhash = false,
54 : .amconsistentequality = false,
55 : .amconsistentordering = false,
56 : .amcanbackward = false,
57 : .amcanunique = false,
58 : .amcanmulticol = false,
59 : .amoptionalkey = true,
60 : .amsearcharray = false,
61 : .amsearchnulls = true,
62 : .amstorage = true,
63 : .amclusterable = false,
64 : .ampredlocks = false,
65 : .amcanparallel = false,
66 : .amcanbuildparallel = false,
67 : .amcaninclude = true,
68 : .amusemaintenanceworkmem = false,
69 : .amsummarizing = false,
70 : .amparallelvacuumoptions =
71 : VACUUM_OPTION_PARALLEL_BULKDEL | VACUUM_OPTION_PARALLEL_COND_CLEANUP,
72 : .amkeytype = InvalidOid,
73 :
74 : .ambuild = spgbuild,
75 : .ambuildempty = spgbuildempty,
76 : .aminsert = spginsert,
77 : .aminsertcleanup = NULL,
78 : .ambulkdelete = spgbulkdelete,
79 : .amvacuumcleanup = spgvacuumcleanup,
80 : .amcanreturn = spgcanreturn,
81 : .amcostestimate = spgcostestimate,
82 : .amgettreeheight = NULL,
83 : .amoptions = spgoptions,
84 : .amproperty = spgproperty,
85 : .ambuildphasename = NULL,
86 : .amvalidate = spgvalidate,
87 : .amadjustmembers = spgadjustmembers,
88 : .ambeginscan = spgbeginscan,
89 : .amrescan = spgrescan,
90 : .amgettuple = spggettuple,
91 : .amgetbitmap = spggetbitmap,
92 : .amendscan = spgendscan,
93 : .ammarkpos = NULL,
94 : .amrestrpos = NULL,
95 : .amestimateparallelscan = NULL,
96 : .aminitparallelscan = NULL,
97 : .amparallelrescan = NULL,
98 : .amtranslatestrategy = NULL,
99 : .amtranslatecmptype = NULL,
100 : };
101 :
102 1354 : PG_RETURN_POINTER(&amroutine);
103 : }
104 :
105 : /*
106 : * GetIndexInputType
107 : * Determine the nominal input data type for an index column
108 : *
109 : * We define the "nominal" input type as the associated opclass's opcintype,
110 : * or if that is a polymorphic type, the base type of the heap column or
111 : * expression that is the index's input. The reason for preferring the
112 : * opcintype is that non-polymorphic opclasses probably don't want to hear
113 : * about binary-compatible input types. For instance, if a text opclass
114 : * is being used with a varchar heap column, we want to report "text" not
115 : * "varchar". Likewise, opclasses don't want to hear about domain types,
116 : * so if we do consult the actual input type, we make sure to flatten domains.
117 : *
118 : * At some point maybe this should go somewhere else, but it's not clear
119 : * if any other index AMs have a use for it.
120 : */
121 : static Oid
122 418 : GetIndexInputType(Relation index, AttrNumber indexcol)
123 : {
124 : Oid opcintype;
125 : AttrNumber heapcol;
126 : List *indexprs;
127 : ListCell *indexpr_item;
128 :
129 : Assert(index->rd_index != NULL);
130 : Assert(indexcol > 0 && indexcol <= index->rd_index->indnkeyatts);
131 418 : opcintype = index->rd_opcintype[indexcol - 1];
132 418 : if (!IsPolymorphicType(opcintype))
133 314 : return opcintype;
134 104 : heapcol = index->rd_index->indkey.values[indexcol - 1];
135 104 : if (heapcol != 0) /* Simple index column? */
136 92 : return getBaseType(get_atttype(index->rd_index->indrelid, heapcol));
137 :
138 : /*
139 : * If the index expressions are already cached, skip calling
140 : * RelationGetIndexExpressions, as it will make a copy which is overkill.
141 : * We're not going to modify the trees, and we're not going to do anything
142 : * that would invalidate the relcache entry before we're done.
143 : */
144 12 : if (index->rd_indexprs)
145 0 : indexprs = index->rd_indexprs;
146 : else
147 12 : indexprs = RelationGetIndexExpressions(index);
148 12 : indexpr_item = list_head(indexprs);
149 12 : for (int i = 1; i <= index->rd_index->indnkeyatts; i++)
150 : {
151 12 : if (index->rd_index->indkey.values[i - 1] == 0)
152 : {
153 : /* expression column */
154 12 : if (indexpr_item == NULL)
155 0 : elog(ERROR, "wrong number of index expressions");
156 12 : if (i == indexcol)
157 12 : return getBaseType(exprType((Node *) lfirst(indexpr_item)));
158 0 : indexpr_item = lnext(indexprs, indexpr_item);
159 : }
160 : }
161 0 : elog(ERROR, "wrong number of index expressions");
162 : return InvalidOid; /* keep compiler quiet */
163 : }
164 :
165 : /* Fill in a SpGistTypeDesc struct with info about the specified data type */
166 : static void
167 1284 : fillTypeDesc(SpGistTypeDesc *desc, Oid type)
168 : {
169 : HeapTuple tp;
170 : Form_pg_type typtup;
171 :
172 1284 : desc->type = type;
173 1284 : tp = SearchSysCache1(TYPEOID, ObjectIdGetDatum(type));
174 1284 : if (!HeapTupleIsValid(tp))
175 0 : elog(ERROR, "cache lookup failed for type %u", type);
176 1284 : typtup = (Form_pg_type) GETSTRUCT(tp);
177 1284 : desc->attlen = typtup->typlen;
178 1284 : desc->attbyval = typtup->typbyval;
179 1284 : desc->attalign = typtup->typalign;
180 1284 : desc->attstorage = typtup->typstorage;
181 1284 : ReleaseSysCache(tp);
182 1284 : }
183 :
184 : /*
185 : * Fetch local cache of AM-specific info about the index, initializing it
186 : * if necessary
187 : */
188 : SpGistCache *
189 2711354 : spgGetCache(Relation index)
190 : {
191 : SpGistCache *cache;
192 :
193 2711354 : if (index->rd_amcache == NULL)
194 : {
195 : Oid atttype;
196 : spgConfigIn in;
197 : FmgrInfo *procinfo;
198 :
199 418 : cache = MemoryContextAllocZero(index->rd_indexcxt,
200 : sizeof(SpGistCache));
201 :
202 : /* SPGiST must have one key column and can also have INCLUDE columns */
203 : Assert(IndexRelationGetNumberOfKeyAttributes(index) == 1);
204 : Assert(IndexRelationGetNumberOfAttributes(index) <= INDEX_MAX_KEYS);
205 :
206 : /*
207 : * Get the actual (well, nominal) data type of the key column. We
208 : * pass this to the opclass config function so that polymorphic
209 : * opclasses are possible.
210 : */
211 418 : atttype = GetIndexInputType(index, spgKeyColumn + 1);
212 :
213 : /* Call the config function to get config info for the opclass */
214 418 : in.attType = atttype;
215 :
216 418 : procinfo = index_getprocinfo(index, 1, SPGIST_CONFIG_PROC);
217 418 : FunctionCall2Coll(procinfo,
218 418 : index->rd_indcollation[spgKeyColumn],
219 : PointerGetDatum(&in),
220 418 : PointerGetDatum(&cache->config));
221 :
222 : /*
223 : * If leafType isn't specified, use the declared index column type,
224 : * which index.c will have derived from the opclass's opcintype.
225 : * (Although we now make spgvalidate.c warn if these aren't the same,
226 : * old user-defined opclasses may not set the STORAGE parameter
227 : * correctly, so believe leafType if it's given.)
228 : */
229 418 : if (!OidIsValid(cache->config.leafType))
230 : {
231 388 : cache->config.leafType =
232 388 : TupleDescAttr(RelationGetDescr(index), spgKeyColumn)->atttypid;
233 :
234 : /*
235 : * If index column type is binary-coercible to atttype (for
236 : * example, it's a domain over atttype), treat it as plain atttype
237 : * to avoid thinking we need to compress.
238 : */
239 402 : if (cache->config.leafType != atttype &&
240 14 : IsBinaryCoercible(cache->config.leafType, atttype))
241 14 : cache->config.leafType = atttype;
242 : }
243 :
244 : /* Get the information we need about each relevant datatype */
245 418 : fillTypeDesc(&cache->attType, atttype);
246 :
247 418 : if (cache->config.leafType != atttype)
248 : {
249 30 : if (!OidIsValid(index_getprocid(index, 1, SPGIST_COMPRESS_PROC)))
250 0 : ereport(ERROR,
251 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
252 : errmsg("compress method must be defined when leaf type is different from input type")));
253 :
254 30 : fillTypeDesc(&cache->attLeafType, cache->config.leafType);
255 : }
256 : else
257 : {
258 : /* Save lookups in this common case */
259 388 : cache->attLeafType = cache->attType;
260 : }
261 :
262 418 : fillTypeDesc(&cache->attPrefixType, cache->config.prefixType);
263 418 : fillTypeDesc(&cache->attLabelType, cache->config.labelType);
264 :
265 : /*
266 : * Finally, if it's a real index (not a partitioned one), get the
267 : * lastUsedPages data from the metapage
268 : */
269 418 : if (index->rd_rel->relkind != RELKIND_PARTITIONED_INDEX)
270 : {
271 : Buffer metabuffer;
272 : SpGistMetaPageData *metadata;
273 :
274 412 : metabuffer = ReadBuffer(index, SPGIST_METAPAGE_BLKNO);
275 412 : LockBuffer(metabuffer, BUFFER_LOCK_SHARE);
276 :
277 412 : metadata = SpGistPageGetMeta(BufferGetPage(metabuffer));
278 :
279 412 : if (metadata->magicNumber != SPGIST_MAGIC_NUMBER)
280 0 : elog(ERROR, "index \"%s\" is not an SP-GiST index",
281 : RelationGetRelationName(index));
282 :
283 412 : cache->lastUsedPages = metadata->lastUsedPages;
284 :
285 412 : UnlockReleaseBuffer(metabuffer);
286 : }
287 :
288 418 : index->rd_amcache = cache;
289 : }
290 : else
291 : {
292 : /* assume it's up to date */
293 2710936 : cache = (SpGistCache *) index->rd_amcache;
294 : }
295 :
296 2711354 : return cache;
297 : }
298 :
299 : /*
300 : * Compute a tuple descriptor for leaf tuples or index-only-scan result tuples.
301 : *
302 : * We can use the relcache's tupdesc as-is in many cases, and it's always
303 : * OK so far as any INCLUDE columns are concerned. However, the entry for
304 : * the key column has to match leafType in the first case or attType in the
305 : * second case. While the relcache's tupdesc *should* show leafType, this
306 : * might not hold for legacy user-defined opclasses, since before v14 they
307 : * were not allowed to declare their true storage type in CREATE OPCLASS.
308 : * Also, attType can be different from what is in the relcache.
309 : *
310 : * This function gives back either a pointer to the relcache's tupdesc
311 : * if that is suitable, or a palloc'd copy that's been adjusted to match
312 : * the specified key column type. We can avoid doing any catalog lookups
313 : * here by insisting that the caller pass an SpGistTypeDesc not just an OID.
314 : */
315 : TupleDesc
316 245520 : getSpGistTupleDesc(Relation index, SpGistTypeDesc *keyType)
317 : {
318 : TupleDesc outTupDesc;
319 : Form_pg_attribute att;
320 :
321 491040 : if (keyType->type ==
322 245520 : TupleDescAttr(RelationGetDescr(index), spgKeyColumn)->atttypid)
323 245390 : outTupDesc = RelationGetDescr(index);
324 : else
325 : {
326 130 : outTupDesc = CreateTupleDescCopy(RelationGetDescr(index));
327 130 : att = TupleDescAttr(outTupDesc, spgKeyColumn);
328 : /* It's sufficient to update the type-dependent fields of the column */
329 130 : att->atttypid = keyType->type;
330 130 : att->atttypmod = -1;
331 130 : att->attlen = keyType->attlen;
332 130 : att->attbyval = keyType->attbyval;
333 130 : att->attalign = keyType->attalign;
334 130 : att->attstorage = keyType->attstorage;
335 : /* We shouldn't need to bother with making these valid: */
336 130 : att->attcompression = InvalidCompressionMethod;
337 130 : att->attcollation = InvalidOid;
338 : /* In case we changed typlen, we'd better reset following offsets */
339 146 : for (int i = spgFirstIncludeColumn; i < outTupDesc->natts; i++)
340 16 : TupleDescCompactAttr(outTupDesc, i)->attcacheoff = -1;
341 :
342 130 : populate_compact_attribute(outTupDesc, spgKeyColumn);
343 : }
344 245520 : return outTupDesc;
345 : }
346 :
347 : /* Initialize SpGistState for working with the given index */
348 : void
349 244616 : initSpGistState(SpGistState *state, Relation index)
350 : {
351 : SpGistCache *cache;
352 :
353 244616 : state->index = index;
354 :
355 : /* Get cached static information about index */
356 244616 : cache = spgGetCache(index);
357 :
358 244616 : state->config = cache->config;
359 244616 : state->attType = cache->attType;
360 244616 : state->attLeafType = cache->attLeafType;
361 244616 : state->attPrefixType = cache->attPrefixType;
362 244616 : state->attLabelType = cache->attLabelType;
363 :
364 : /* Ensure we have a valid descriptor for leaf tuples */
365 244616 : state->leafTupDesc = getSpGistTupleDesc(state->index, &state->attLeafType);
366 :
367 : /* Make workspace for constructing dead tuples */
368 244616 : state->deadTupleStorage = palloc0(SGDTSIZE);
369 :
370 : /*
371 : * Set horizon XID to use in redirection tuples. Use our own XID if we
372 : * have one, else use InvalidTransactionId. The latter case can happen in
373 : * VACUUM or REINDEX CONCURRENTLY, and in neither case would it be okay to
374 : * force an XID to be assigned. VACUUM won't create any redirection
375 : * tuples anyway, but REINDEX CONCURRENTLY can. Fortunately, REINDEX
376 : * CONCURRENTLY doesn't mark the index valid until the end, so there could
377 : * never be any concurrent scans "in flight" to a redirection tuple it has
378 : * inserted. And it locks out VACUUM until the end, too. So it's okay
379 : * for VACUUM to immediately expire a redirection tuple that contains an
380 : * invalid xid.
381 : */
382 244616 : state->redirectXid = GetTopTransactionIdIfAny();
383 :
384 : /* Assume we're not in an index build (spgbuild will override) */
385 244616 : state->isBuild = false;
386 244616 : }
387 :
388 : /*
389 : * Allocate a new page (either by recycling, or by extending the index file).
390 : *
391 : * The returned buffer is already pinned and exclusive-locked.
392 : * Caller is responsible for initializing the page by calling SpGistInitBuffer.
393 : */
394 : Buffer
395 6600 : SpGistNewBuffer(Relation index)
396 : {
397 : Buffer buffer;
398 :
399 : /* First, try to get a page from FSM */
400 : for (;;)
401 0 : {
402 6600 : BlockNumber blkno = GetFreeIndexPage(index);
403 :
404 6600 : if (blkno == InvalidBlockNumber)
405 6592 : break; /* nothing known to FSM */
406 :
407 : /*
408 : * The fixed pages shouldn't ever be listed in FSM, but just in case
409 : * one is, ignore it.
410 : */
411 8 : if (SpGistBlockIsFixed(blkno))
412 0 : continue;
413 :
414 8 : buffer = ReadBuffer(index, blkno);
415 :
416 : /*
417 : * We have to guard against the possibility that someone else already
418 : * recycled this page; the buffer may be locked if so.
419 : */
420 8 : if (ConditionalLockBuffer(buffer))
421 : {
422 8 : Page page = BufferGetPage(buffer);
423 :
424 8 : if (PageIsNew(page))
425 0 : return buffer; /* OK to use, if never initialized */
426 :
427 8 : if (SpGistPageIsDeleted(page) || PageIsEmpty(page))
428 8 : return buffer; /* OK to use */
429 :
430 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
431 : }
432 :
433 : /* Can't use it, so release buffer and try again */
434 0 : ReleaseBuffer(buffer);
435 : }
436 :
437 6592 : buffer = ExtendBufferedRel(BMR_REL(index), MAIN_FORKNUM, NULL,
438 : EB_LOCK_FIRST);
439 :
440 6592 : return buffer;
441 : }
442 :
443 : /*
444 : * Update index metapage's lastUsedPages info from local cache, if possible
445 : *
446 : * Updating meta page isn't critical for index working, so
447 : * 1 use ConditionalLockBuffer to improve concurrency
448 : * 2 don't WAL-log metabuffer changes to decrease WAL traffic
449 : */
450 : void
451 243152 : SpGistUpdateMetaPage(Relation index)
452 : {
453 243152 : SpGistCache *cache = (SpGistCache *) index->rd_amcache;
454 :
455 243152 : if (cache != NULL)
456 : {
457 : Buffer metabuffer;
458 :
459 243152 : metabuffer = ReadBuffer(index, SPGIST_METAPAGE_BLKNO);
460 :
461 243152 : if (ConditionalLockBuffer(metabuffer))
462 : {
463 243152 : Page metapage = BufferGetPage(metabuffer);
464 243152 : SpGistMetaPageData *metadata = SpGistPageGetMeta(metapage);
465 :
466 243152 : metadata->lastUsedPages = cache->lastUsedPages;
467 :
468 : /*
469 : * Set pd_lower just past the end of the metadata. This is
470 : * essential, because without doing so, metadata will be lost if
471 : * xlog.c compresses the page. (We must do this here because
472 : * pre-v11 versions of PG did not set the metapage's pd_lower
473 : * correctly, so a pg_upgraded index might contain the wrong
474 : * value.)
475 : */
476 243152 : ((PageHeader) metapage)->pd_lower =
477 243152 : ((char *) metadata + sizeof(SpGistMetaPageData)) - (char *) metapage;
478 :
479 243152 : MarkBufferDirty(metabuffer);
480 243152 : UnlockReleaseBuffer(metabuffer);
481 : }
482 : else
483 : {
484 0 : ReleaseBuffer(metabuffer);
485 : }
486 : }
487 243152 : }
488 :
489 : /* Macro to select proper element of lastUsedPages cache depending on flags */
490 : /* Masking flags with SPGIST_CACHED_PAGES is just for paranoia's sake */
491 : #define GET_LUP(c, f) (&(c)->lastUsedPages.cachedPage[((unsigned int) (f)) % SPGIST_CACHED_PAGES])
492 :
493 : /*
494 : * Allocate and initialize a new buffer of the type and parity specified by
495 : * flags. The returned buffer is already pinned and exclusive-locked.
496 : *
497 : * When requesting an inner page, if we get one with the wrong parity,
498 : * we just release the buffer and try again. We will get a different page
499 : * because GetFreeIndexPage will have marked the page used in FSM. The page
500 : * is entered in our local lastUsedPages cache, so there's some hope of
501 : * making use of it later in this session, but otherwise we rely on VACUUM
502 : * to eventually re-enter the page in FSM, making it available for recycling.
503 : * Note that such a page does not get marked dirty here, so unless it's used
504 : * fairly soon, the buffer will just get discarded and the page will remain
505 : * as it was on disk.
506 : *
507 : * When we return a buffer to the caller, the page is *not* entered into
508 : * the lastUsedPages cache; we expect the caller will do so after it's taken
509 : * whatever space it will use. This is because after the caller has used up
510 : * some space, the page might have less space than whatever was cached already
511 : * so we'd rather not trash the old cache entry.
512 : */
513 : static Buffer
514 5890 : allocNewBuffer(Relation index, int flags)
515 : {
516 5890 : SpGistCache *cache = spgGetCache(index);
517 5890 : uint16 pageflags = 0;
518 :
519 5890 : if (GBUF_REQ_LEAF(flags))
520 5784 : pageflags |= SPGIST_LEAF;
521 5890 : if (GBUF_REQ_NULLS(flags))
522 0 : pageflags |= SPGIST_NULLS;
523 :
524 : for (;;)
525 86 : {
526 : Buffer buffer;
527 :
528 5976 : buffer = SpGistNewBuffer(index);
529 5976 : SpGistInitBuffer(buffer, pageflags);
530 :
531 5976 : if (pageflags & SPGIST_LEAF)
532 : {
533 : /* Leaf pages have no parity concerns, so just use it */
534 5784 : return buffer;
535 : }
536 : else
537 : {
538 192 : BlockNumber blkno = BufferGetBlockNumber(buffer);
539 192 : int blkFlags = GBUF_INNER_PARITY(blkno);
540 :
541 192 : if ((flags & GBUF_PARITY_MASK) == blkFlags)
542 : {
543 : /* Page has right parity, use it */
544 106 : return buffer;
545 : }
546 : else
547 : {
548 : /* Page has wrong parity, record it in cache and try again */
549 86 : if (pageflags & SPGIST_NULLS)
550 0 : blkFlags |= GBUF_NULLS;
551 86 : cache->lastUsedPages.cachedPage[blkFlags].blkno = blkno;
552 86 : cache->lastUsedPages.cachedPage[blkFlags].freeSpace =
553 86 : PageGetExactFreeSpace(BufferGetPage(buffer));
554 86 : UnlockReleaseBuffer(buffer);
555 : }
556 : }
557 : }
558 : }
559 :
560 : /*
561 : * Get a buffer of the type and parity specified by flags, having at least
562 : * as much free space as indicated by needSpace. We use the lastUsedPages
563 : * cache to assign the same buffer previously requested when possible.
564 : * The returned buffer is already pinned and exclusive-locked.
565 : *
566 : * *isNew is set true if the page was initialized here, false if it was
567 : * already valid.
568 : */
569 : Buffer
570 10956 : SpGistGetBuffer(Relation index, int flags, int needSpace, bool *isNew)
571 : {
572 10956 : SpGistCache *cache = spgGetCache(index);
573 : SpGistLastUsedPage *lup;
574 :
575 : /* Bail out if even an empty page wouldn't meet the demand */
576 10956 : if (needSpace > SPGIST_PAGE_CAPACITY)
577 0 : elog(ERROR, "desired SPGiST tuple size is too big");
578 :
579 : /*
580 : * If possible, increase the space request to include relation's
581 : * fillfactor. This ensures that when we add unrelated tuples to a page,
582 : * we try to keep 100-fillfactor% available for adding tuples that are
583 : * related to the ones already on it. But fillfactor mustn't cause an
584 : * error for requests that would otherwise be legal.
585 : */
586 10956 : needSpace += SpGistGetTargetPageFreeSpace(index);
587 10956 : needSpace = Min(needSpace, SPGIST_PAGE_CAPACITY);
588 :
589 : /* Get the cache entry for this flags setting */
590 10956 : lup = GET_LUP(cache, flags);
591 :
592 : /* If we have nothing cached, just turn it over to allocNewBuffer */
593 10956 : if (lup->blkno == InvalidBlockNumber)
594 : {
595 182 : *isNew = true;
596 182 : return allocNewBuffer(index, flags);
597 : }
598 :
599 : /* fixed pages should never be in cache */
600 : Assert(!SpGistBlockIsFixed(lup->blkno));
601 :
602 : /* If cached freeSpace isn't enough, don't bother looking at the page */
603 10774 : if (lup->freeSpace >= needSpace)
604 : {
605 : Buffer buffer;
606 : Page page;
607 :
608 5066 : buffer = ReadBuffer(index, lup->blkno);
609 :
610 5066 : if (!ConditionalLockBuffer(buffer))
611 : {
612 : /*
613 : * buffer is locked by another process, so return a new buffer
614 : */
615 0 : ReleaseBuffer(buffer);
616 0 : *isNew = true;
617 0 : return allocNewBuffer(index, flags);
618 : }
619 :
620 5066 : page = BufferGetPage(buffer);
621 :
622 5066 : if (PageIsNew(page) || SpGistPageIsDeleted(page) || PageIsEmpty(page))
623 : {
624 : /* OK to initialize the page */
625 188 : uint16 pageflags = 0;
626 :
627 188 : if (GBUF_REQ_LEAF(flags))
628 182 : pageflags |= SPGIST_LEAF;
629 188 : if (GBUF_REQ_NULLS(flags))
630 0 : pageflags |= SPGIST_NULLS;
631 188 : SpGistInitBuffer(buffer, pageflags);
632 188 : lup->freeSpace = PageGetExactFreeSpace(page) - needSpace;
633 188 : *isNew = true;
634 188 : return buffer;
635 : }
636 :
637 : /*
638 : * Check that page is of right type and has enough space. We must
639 : * recheck this since our cache isn't necessarily up to date.
640 : */
641 9756 : if ((GBUF_REQ_LEAF(flags) ? SpGistPageIsLeaf(page) : !SpGistPageIsLeaf(page)) &&
642 4878 : (GBUF_REQ_NULLS(flags) ? SpGistPageStoresNulls(page) : !SpGistPageStoresNulls(page)))
643 : {
644 4878 : int freeSpace = PageGetExactFreeSpace(page);
645 :
646 4878 : if (freeSpace >= needSpace)
647 : {
648 : /* Success, update freespace info and return the buffer */
649 4878 : lup->freeSpace = freeSpace - needSpace;
650 4878 : *isNew = false;
651 4878 : return buffer;
652 : }
653 : }
654 :
655 : /*
656 : * fallback to allocation of new buffer
657 : */
658 0 : UnlockReleaseBuffer(buffer);
659 : }
660 :
661 : /* No success with cache, so return a new buffer */
662 5708 : *isNew = true;
663 5708 : return allocNewBuffer(index, flags);
664 : }
665 :
666 : /*
667 : * Update lastUsedPages cache when done modifying a page.
668 : *
669 : * We update the appropriate cache entry if it already contained this page
670 : * (its freeSpace is likely obsolete), or if this page has more space than
671 : * whatever we had cached.
672 : */
673 : void
674 2448066 : SpGistSetLastUsedPage(Relation index, Buffer buffer)
675 : {
676 2448066 : SpGistCache *cache = spgGetCache(index);
677 : SpGistLastUsedPage *lup;
678 : int freeSpace;
679 2448066 : Page page = BufferGetPage(buffer);
680 2448066 : BlockNumber blkno = BufferGetBlockNumber(buffer);
681 : int flags;
682 :
683 : /* Never enter fixed pages (root pages) in cache, though */
684 2448066 : if (SpGistBlockIsFixed(blkno))
685 809136 : return;
686 :
687 1638930 : if (SpGistPageIsLeaf(page))
688 853604 : flags = GBUF_LEAF;
689 : else
690 785326 : flags = GBUF_INNER_PARITY(blkno);
691 1638930 : if (SpGistPageStoresNulls(page))
692 0 : flags |= GBUF_NULLS;
693 :
694 1638930 : lup = GET_LUP(cache, flags);
695 :
696 1638930 : freeSpace = PageGetExactFreeSpace(page);
697 1638930 : if (lup->blkno == InvalidBlockNumber || lup->blkno == blkno ||
698 475464 : lup->freeSpace < freeSpace)
699 : {
700 1172034 : lup->blkno = blkno;
701 1172034 : lup->freeSpace = freeSpace;
702 : }
703 : }
704 :
705 : /*
706 : * Initialize an SPGiST page to empty, with specified flags
707 : */
708 : void
709 7686 : SpGistInitPage(Page page, uint16 f)
710 : {
711 : SpGistPageOpaque opaque;
712 :
713 7686 : PageInit(page, BLCKSZ, sizeof(SpGistPageOpaqueData));
714 7686 : opaque = SpGistPageGetOpaque(page);
715 7686 : opaque->flags = f;
716 7686 : opaque->spgist_page_id = SPGIST_PAGE_ID;
717 7686 : }
718 :
719 : /*
720 : * Initialize a buffer's page to empty, with specified flags
721 : */
722 : void
723 7454 : SpGistInitBuffer(Buffer b, uint16 f)
724 : {
725 : Assert(BufferGetPageSize(b) == BLCKSZ);
726 7454 : SpGistInitPage(BufferGetPage(b), f);
727 7454 : }
728 :
729 : /*
730 : * Initialize metadata page
731 : */
732 : void
733 216 : SpGistInitMetapage(Page page)
734 : {
735 : SpGistMetaPageData *metadata;
736 : int i;
737 :
738 216 : SpGistInitPage(page, SPGIST_META);
739 216 : metadata = SpGistPageGetMeta(page);
740 216 : memset(metadata, 0, sizeof(SpGistMetaPageData));
741 216 : metadata->magicNumber = SPGIST_MAGIC_NUMBER;
742 :
743 : /* initialize last-used-page cache to empty */
744 1944 : for (i = 0; i < SPGIST_CACHED_PAGES; i++)
745 1728 : metadata->lastUsedPages.cachedPage[i].blkno = InvalidBlockNumber;
746 :
747 : /*
748 : * Set pd_lower just past the end of the metadata. This is essential,
749 : * because without doing so, metadata will be lost if xlog.c compresses
750 : * the page.
751 : */
752 216 : ((PageHeader) page)->pd_lower =
753 216 : ((char *) metadata + sizeof(SpGistMetaPageData)) - (char *) page;
754 216 : }
755 :
756 : /*
757 : * reloptions processing for SPGiST
758 : */
759 : bytea *
760 132 : spgoptions(Datum reloptions, bool validate)
761 : {
762 : static const relopt_parse_elt tab[] = {
763 : {"fillfactor", RELOPT_TYPE_INT, offsetof(SpGistOptions, fillfactor)},
764 : };
765 :
766 132 : return (bytea *) build_reloptions(reloptions, validate,
767 : RELOPT_KIND_SPGIST,
768 : sizeof(SpGistOptions),
769 : tab, lengthof(tab));
770 : }
771 :
772 : /*
773 : * Get the space needed to store a non-null datum of the indicated type
774 : * in an inner tuple (that is, as a prefix or node label).
775 : * Note the result is already rounded up to a MAXALIGN boundary.
776 : * Here we follow the convention that pass-by-val types are just stored
777 : * in their Datum representation (compare memcpyInnerDatum).
778 : */
779 : unsigned int
780 12350 : SpGistGetInnerTypeSize(SpGistTypeDesc *att, Datum datum)
781 : {
782 : unsigned int size;
783 :
784 12350 : if (att->attbyval)
785 6452 : size = sizeof(Datum);
786 5898 : else if (att->attlen > 0)
787 4016 : size = att->attlen;
788 : else
789 1882 : size = VARSIZE_ANY(DatumGetPointer(datum));
790 :
791 12350 : return MAXALIGN(size);
792 : }
793 :
794 : /*
795 : * Copy the given non-null datum to *target, in the inner-tuple case
796 : */
797 : static void
798 12350 : memcpyInnerDatum(void *target, SpGistTypeDesc *att, Datum datum)
799 : {
800 : unsigned int size;
801 :
802 12350 : if (att->attbyval)
803 : {
804 6452 : memcpy(target, &datum, sizeof(Datum));
805 : }
806 : else
807 : {
808 5898 : size = (att->attlen > 0) ? att->attlen : VARSIZE_ANY(DatumGetPointer(datum));
809 5898 : memcpy(target, DatumGetPointer(datum), size);
810 : }
811 12350 : }
812 :
813 : /*
814 : * Compute space required for a leaf tuple holding the given data.
815 : *
816 : * This must match the size-calculation portion of spgFormLeafTuple.
817 : */
818 : Size
819 19517326 : SpGistGetLeafTupleSize(TupleDesc tupleDescriptor,
820 : const Datum *datums, const bool *isnulls)
821 : {
822 : Size size;
823 : Size data_size;
824 19517326 : bool needs_null_mask = false;
825 19517326 : int natts = tupleDescriptor->natts;
826 :
827 : /*
828 : * Decide whether we need a nulls bitmask.
829 : *
830 : * If there is only a key attribute (natts == 1), never use a bitmask, for
831 : * compatibility with the pre-v14 layout of leaf tuples. Otherwise, we
832 : * need one if any attribute is null.
833 : */
834 19517326 : if (natts > 1)
835 : {
836 978386 : for (int i = 0; i < natts; i++)
837 : {
838 670472 : if (isnulls[i])
839 : {
840 18422 : needs_null_mask = true;
841 18422 : break;
842 : }
843 : }
844 : }
845 :
846 : /*
847 : * Calculate size of the data part; same as for heap tuples.
848 : */
849 19517326 : data_size = heap_compute_data_size(tupleDescriptor, datums, isnulls);
850 :
851 : /*
852 : * Compute total size.
853 : */
854 19517326 : size = SGLTHDRSZ(needs_null_mask);
855 19517326 : size += data_size;
856 19517326 : size = MAXALIGN(size);
857 :
858 : /*
859 : * Ensure that we can replace the tuple with a dead tuple later. This test
860 : * is unnecessary when there are any non-null attributes, but be safe.
861 : */
862 19517326 : if (size < SGDTSIZE)
863 0 : size = SGDTSIZE;
864 :
865 19517326 : return size;
866 : }
867 :
868 : /*
869 : * Construct a leaf tuple containing the given heap TID and datum values
870 : */
871 : SpGistLeafTuple
872 1536018 : spgFormLeafTuple(SpGistState *state, const ItemPointerData *heapPtr,
873 : const Datum *datums, const bool *isnulls)
874 : {
875 : SpGistLeafTuple tup;
876 1536018 : TupleDesc tupleDescriptor = state->leafTupDesc;
877 : Size size;
878 : Size hoff;
879 : Size data_size;
880 1536018 : bool needs_null_mask = false;
881 1536018 : int natts = tupleDescriptor->natts;
882 : char *tp; /* ptr to tuple data */
883 1536018 : uint16 tupmask = 0; /* unused heap_fill_tuple output */
884 :
885 : /*
886 : * Decide whether we need a nulls bitmask.
887 : *
888 : * If there is only a key attribute (natts == 1), never use a bitmask, for
889 : * compatibility with the pre-v14 layout of leaf tuples. Otherwise, we
890 : * need one if any attribute is null.
891 : */
892 1536018 : if (natts > 1)
893 : {
894 423072 : for (int i = 0; i < natts; i++)
895 : {
896 293846 : if (isnulls[i])
897 : {
898 11922 : needs_null_mask = true;
899 11922 : break;
900 : }
901 : }
902 : }
903 :
904 : /*
905 : * Calculate size of the data part; same as for heap tuples.
906 : */
907 1536018 : data_size = heap_compute_data_size(tupleDescriptor, datums, isnulls);
908 :
909 : /*
910 : * Compute total size.
911 : */
912 1536018 : hoff = SGLTHDRSZ(needs_null_mask);
913 1536018 : size = hoff + data_size;
914 1536018 : size = MAXALIGN(size);
915 :
916 : /*
917 : * Ensure that we can replace the tuple with a dead tuple later. This test
918 : * is unnecessary when there are any non-null attributes, but be safe.
919 : */
920 1536018 : if (size < SGDTSIZE)
921 0 : size = SGDTSIZE;
922 :
923 : /* OK, form the tuple */
924 1536018 : tup = (SpGistLeafTuple) palloc0(size);
925 :
926 1536018 : tup->size = size;
927 1536018 : SGLT_SET_NEXTOFFSET(tup, InvalidOffsetNumber);
928 1536018 : tup->heapPtr = *heapPtr;
929 :
930 1536018 : tp = (char *) tup + hoff;
931 :
932 1536018 : if (needs_null_mask)
933 : {
934 : bits8 *bp; /* ptr to null bitmap in tuple */
935 :
936 : /* Set nullmask presence bit in SpGistLeafTuple header */
937 11922 : SGLT_SET_HASNULLMASK(tup, true);
938 : /* Fill the data area and null mask */
939 11922 : bp = (bits8 *) ((char *) tup + sizeof(SpGistLeafTupleData));
940 11922 : heap_fill_tuple(tupleDescriptor, datums, isnulls, tp, data_size,
941 : &tupmask, bp);
942 : }
943 1524096 : else if (natts > 1 || !isnulls[spgKeyColumn])
944 : {
945 : /* Fill data area only */
946 1524024 : heap_fill_tuple(tupleDescriptor, datums, isnulls, tp, data_size,
947 : &tupmask, (bits8 *) NULL);
948 : }
949 : /* otherwise we have no data, nor a bitmap, to fill */
950 :
951 1536018 : return tup;
952 : }
953 :
954 : /*
955 : * Construct a node (to go into an inner tuple) containing the given label
956 : *
957 : * Note that the node's downlink is just set invalid here. Caller will fill
958 : * it in later.
959 : */
960 : SpGistNodeTuple
961 40762 : spgFormNodeTuple(SpGistState *state, Datum label, bool isnull)
962 : {
963 : SpGistNodeTuple tup;
964 : unsigned int size;
965 40762 : unsigned short infomask = 0;
966 :
967 : /* compute space needed (note result is already maxaligned) */
968 40762 : size = SGNTHDRSZ;
969 40762 : if (!isnull)
970 5762 : size += SpGistGetInnerTypeSize(&state->attLabelType, label);
971 :
972 : /*
973 : * Here we make sure that the size will fit in the field reserved for it
974 : * in t_info.
975 : */
976 40762 : if ((size & INDEX_SIZE_MASK) != size)
977 0 : ereport(ERROR,
978 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
979 : errmsg("index row requires %zu bytes, maximum size is %zu",
980 : (Size) size, (Size) INDEX_SIZE_MASK)));
981 :
982 40762 : tup = (SpGistNodeTuple) palloc0(size);
983 :
984 40762 : if (isnull)
985 35000 : infomask |= INDEX_NULL_MASK;
986 : /* we don't bother setting the INDEX_VAR_MASK bit */
987 40762 : infomask |= size;
988 40762 : tup->t_info = infomask;
989 :
990 : /* The TID field will be filled in later */
991 40762 : ItemPointerSetInvalid(&tup->t_tid);
992 :
993 40762 : if (!isnull)
994 5762 : memcpyInnerDatum(SGNTDATAPTR(tup), &state->attLabelType, label);
995 :
996 40762 : return tup;
997 : }
998 :
999 : /*
1000 : * Construct an inner tuple containing the given prefix and node array
1001 : */
1002 : SpGistInnerTuple
1003 8576 : spgFormInnerTuple(SpGistState *state, bool hasPrefix, Datum prefix,
1004 : int nNodes, SpGistNodeTuple *nodes)
1005 : {
1006 : SpGistInnerTuple tup;
1007 : unsigned int size;
1008 : unsigned int prefixSize;
1009 : int i;
1010 : char *ptr;
1011 :
1012 : /* Compute size needed */
1013 8576 : if (hasPrefix)
1014 6588 : prefixSize = SpGistGetInnerTypeSize(&state->attPrefixType, prefix);
1015 : else
1016 1988 : prefixSize = 0;
1017 :
1018 8576 : size = SGITHDRSZ + prefixSize;
1019 :
1020 : /* Note: we rely on node tuple sizes to be maxaligned already */
1021 59590 : for (i = 0; i < nNodes; i++)
1022 51014 : size += IndexTupleSize(nodes[i]);
1023 :
1024 : /*
1025 : * Ensure that we can replace the tuple with a dead tuple later. This
1026 : * test is unnecessary given current tuple layouts, but let's be safe.
1027 : */
1028 8576 : if (size < SGDTSIZE)
1029 0 : size = SGDTSIZE;
1030 :
1031 : /*
1032 : * Inner tuple should be small enough to fit on a page
1033 : */
1034 8576 : if (size > SPGIST_PAGE_CAPACITY - sizeof(ItemIdData))
1035 0 : ereport(ERROR,
1036 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1037 : errmsg("SP-GiST inner tuple size %zu exceeds maximum %zu",
1038 : (Size) size,
1039 : SPGIST_PAGE_CAPACITY - sizeof(ItemIdData)),
1040 : errhint("Values larger than a buffer page cannot be indexed.")));
1041 :
1042 : /*
1043 : * Check for overflow of header fields --- probably can't fail if the
1044 : * above succeeded, but let's be paranoid
1045 : */
1046 8576 : if (size > SGITMAXSIZE ||
1047 8576 : prefixSize > SGITMAXPREFIXSIZE ||
1048 : nNodes > SGITMAXNNODES)
1049 0 : elog(ERROR, "SPGiST inner tuple header field is too small");
1050 :
1051 : /* OK, form the tuple */
1052 8576 : tup = (SpGistInnerTuple) palloc0(size);
1053 :
1054 8576 : tup->nNodes = nNodes;
1055 8576 : tup->prefixSize = prefixSize;
1056 8576 : tup->size = size;
1057 :
1058 8576 : if (hasPrefix)
1059 6588 : memcpyInnerDatum(SGITDATAPTR(tup), &state->attPrefixType, prefix);
1060 :
1061 8576 : ptr = (char *) SGITNODEPTR(tup);
1062 :
1063 59590 : for (i = 0; i < nNodes; i++)
1064 : {
1065 51014 : SpGistNodeTuple node = nodes[i];
1066 :
1067 51014 : memcpy(ptr, node, IndexTupleSize(node));
1068 51014 : ptr += IndexTupleSize(node);
1069 : }
1070 :
1071 8576 : return tup;
1072 : }
1073 :
1074 : /*
1075 : * Construct a "dead" tuple to replace a tuple being deleted.
1076 : *
1077 : * The state can be SPGIST_REDIRECT, SPGIST_DEAD, or SPGIST_PLACEHOLDER.
1078 : * For a REDIRECT tuple, a pointer (blkno+offset) must be supplied, and
1079 : * the xid field is filled in automatically.
1080 : *
1081 : * This is called in critical sections, so we don't use palloc; the tuple
1082 : * is built in preallocated storage. It should be copied before another
1083 : * call with different parameters can occur.
1084 : */
1085 : SpGistDeadTuple
1086 14146 : spgFormDeadTuple(SpGistState *state, int tupstate,
1087 : BlockNumber blkno, OffsetNumber offnum)
1088 : {
1089 14146 : SpGistDeadTuple tuple = (SpGistDeadTuple) state->deadTupleStorage;
1090 :
1091 14146 : tuple->tupstate = tupstate;
1092 14146 : tuple->size = SGDTSIZE;
1093 14146 : SGLT_SET_NEXTOFFSET(tuple, InvalidOffsetNumber);
1094 :
1095 14146 : if (tupstate == SPGIST_REDIRECT)
1096 : {
1097 2436 : ItemPointerSet(&tuple->pointer, blkno, offnum);
1098 2436 : tuple->xid = state->redirectXid;
1099 : }
1100 : else
1101 : {
1102 11710 : ItemPointerSetInvalid(&tuple->pointer);
1103 11710 : tuple->xid = InvalidTransactionId;
1104 : }
1105 :
1106 14146 : return tuple;
1107 : }
1108 :
1109 : /*
1110 : * Convert an SPGiST leaf tuple into Datum/isnull arrays.
1111 : *
1112 : * The caller must allocate sufficient storage for the output arrays.
1113 : * (INDEX_MAX_KEYS entries should be enough.)
1114 : */
1115 : void
1116 61448 : spgDeformLeafTuple(SpGistLeafTuple tup, TupleDesc tupleDescriptor,
1117 : Datum *datums, bool *isnulls, bool keyColumnIsNull)
1118 : {
1119 61448 : bool hasNullsMask = SGLT_GET_HASNULLMASK(tup);
1120 : char *tp; /* ptr to tuple data */
1121 : bits8 *bp; /* ptr to null bitmap in tuple */
1122 :
1123 61448 : if (keyColumnIsNull && tupleDescriptor->natts == 1)
1124 : {
1125 : /*
1126 : * Trivial case: there is only the key attribute and we're in a nulls
1127 : * tree. The hasNullsMask bit in the tuple header should not be set
1128 : * (and thus we can't use index_deform_tuple_internal), but
1129 : * nonetheless the result is NULL.
1130 : *
1131 : * Note: currently this is dead code, because noplace calls this when
1132 : * there is only the key attribute. But we should cover the case.
1133 : */
1134 : Assert(!hasNullsMask);
1135 :
1136 0 : datums[spgKeyColumn] = (Datum) 0;
1137 0 : isnulls[spgKeyColumn] = true;
1138 0 : return;
1139 : }
1140 :
1141 61448 : tp = (char *) tup + SGLTHDRSZ(hasNullsMask);
1142 61448 : bp = (bits8 *) ((char *) tup + sizeof(SpGistLeafTupleData));
1143 :
1144 61448 : index_deform_tuple_internal(tupleDescriptor,
1145 : datums, isnulls,
1146 : tp, bp, hasNullsMask);
1147 :
1148 : /*
1149 : * Key column isnull value from the tuple should be consistent with
1150 : * keyColumnIsNull flag from the caller.
1151 : */
1152 : Assert(keyColumnIsNull == isnulls[spgKeyColumn]);
1153 : }
1154 :
1155 : /*
1156 : * Extract the label datums of the nodes within innerTuple
1157 : *
1158 : * Returns NULL if label datums are NULLs
1159 : */
1160 : Datum *
1161 18734864 : spgExtractNodeLabels(SpGistState *state, SpGistInnerTuple innerTuple)
1162 : {
1163 : Datum *nodeLabels;
1164 : int i;
1165 : SpGistNodeTuple node;
1166 :
1167 : /* Either all the labels must be NULL, or none. */
1168 18734864 : node = SGITNODEPTR(innerTuple);
1169 18734864 : if (IndexTupleHasNulls(node))
1170 : {
1171 100977116 : SGITITERATE(innerTuple, i, node)
1172 : {
1173 82476038 : if (!IndexTupleHasNulls(node))
1174 0 : elog(ERROR, "some but not all node labels are null in SPGiST inner tuple");
1175 : }
1176 : /* They're all null, so just return NULL */
1177 18501078 : return NULL;
1178 : }
1179 : else
1180 : {
1181 233786 : nodeLabels = palloc_array(Datum, innerTuple->nNodes);
1182 2685092 : SGITITERATE(innerTuple, i, node)
1183 : {
1184 2451306 : if (IndexTupleHasNulls(node))
1185 0 : elog(ERROR, "some but not all node labels are null in SPGiST inner tuple");
1186 2451306 : nodeLabels[i] = SGNTDATUM(node, state);
1187 : }
1188 233786 : return nodeLabels;
1189 : }
1190 : }
1191 :
1192 : /*
1193 : * Add a new item to the page, replacing a PLACEHOLDER item if possible.
1194 : * Return the location it's inserted at, or InvalidOffsetNumber on failure.
1195 : *
1196 : * If startOffset isn't NULL, we start searching for placeholders at
1197 : * *startOffset, and update that to the next place to search. This is just
1198 : * an optimization for repeated insertions.
1199 : *
1200 : * If errorOK is false, we throw error when there's not enough room,
1201 : * rather than returning InvalidOffsetNumber.
1202 : */
1203 : OffsetNumber
1204 1627058 : SpGistPageAddNewItem(SpGistState *state, Page page, const void *item, Size size,
1205 : OffsetNumber *startOffset, bool errorOK)
1206 : {
1207 1627058 : SpGistPageOpaque opaque = SpGistPageGetOpaque(page);
1208 : OffsetNumber i,
1209 : maxoff,
1210 : offnum;
1211 :
1212 1627058 : if (opaque->nPlaceholder > 0 &&
1213 458992 : PageGetExactFreeSpace(page) + SGDTSIZE >= MAXALIGN(size))
1214 : {
1215 : /* Try to replace a placeholder */
1216 458992 : maxoff = PageGetMaxOffsetNumber(page);
1217 458992 : offnum = InvalidOffsetNumber;
1218 :
1219 : for (;;)
1220 : {
1221 458992 : if (startOffset && *startOffset != InvalidOffsetNumber)
1222 115246 : i = *startOffset;
1223 : else
1224 343746 : i = FirstOffsetNumber;
1225 31281330 : for (; i <= maxoff; i++)
1226 : {
1227 31281330 : SpGistDeadTuple it = (SpGistDeadTuple) PageGetItem(page,
1228 31281330 : PageGetItemId(page, i));
1229 :
1230 31281330 : if (it->tupstate == SPGIST_PLACEHOLDER)
1231 : {
1232 458992 : offnum = i;
1233 458992 : break;
1234 : }
1235 : }
1236 :
1237 : /* Done if we found a placeholder */
1238 458992 : if (offnum != InvalidOffsetNumber)
1239 458992 : break;
1240 :
1241 0 : if (startOffset && *startOffset != InvalidOffsetNumber)
1242 : {
1243 : /* Hint was no good, re-search from beginning */
1244 0 : *startOffset = InvalidOffsetNumber;
1245 0 : continue;
1246 : }
1247 :
1248 : /* Hmm, no placeholder found? */
1249 0 : opaque->nPlaceholder = 0;
1250 0 : break;
1251 : }
1252 :
1253 458992 : if (offnum != InvalidOffsetNumber)
1254 : {
1255 : /* Replace the placeholder tuple */
1256 458992 : PageIndexTupleDelete(page, offnum);
1257 :
1258 458992 : offnum = PageAddItem(page, item, size, offnum, false, false);
1259 :
1260 : /*
1261 : * We should not have failed given the size check at the top of
1262 : * the function, but test anyway. If we did fail, we must PANIC
1263 : * because we've already deleted the placeholder tuple, and
1264 : * there's no other way to keep the damage from getting to disk.
1265 : */
1266 458992 : if (offnum != InvalidOffsetNumber)
1267 : {
1268 : Assert(opaque->nPlaceholder > 0);
1269 458992 : opaque->nPlaceholder--;
1270 458992 : if (startOffset)
1271 117888 : *startOffset = offnum + 1;
1272 : }
1273 : else
1274 0 : elog(PANIC, "failed to add item of size %zu to SPGiST index page",
1275 : size);
1276 :
1277 458992 : return offnum;
1278 : }
1279 : }
1280 :
1281 : /* No luck in replacing a placeholder, so just add it to the page */
1282 1168066 : offnum = PageAddItem(page, item, size,
1283 : InvalidOffsetNumber, false, false);
1284 :
1285 1168066 : if (offnum == InvalidOffsetNumber && !errorOK)
1286 0 : elog(ERROR, "failed to add item of size %zu to SPGiST index page",
1287 : size);
1288 :
1289 1168066 : return offnum;
1290 : }
1291 :
1292 : /*
1293 : * spgproperty() -- Check boolean properties of indexes.
1294 : *
1295 : * This is optional for most AMs, but is required for SP-GiST because the core
1296 : * property code doesn't support AMPROP_DISTANCE_ORDERABLE.
1297 : */
1298 : bool
1299 186 : spgproperty(Oid index_oid, int attno,
1300 : IndexAMProperty prop, const char *propname,
1301 : bool *res, bool *isnull)
1302 : {
1303 : Oid opclass,
1304 : opfamily,
1305 : opcintype;
1306 : CatCList *catlist;
1307 : int i;
1308 :
1309 : /* Only answer column-level inquiries */
1310 186 : if (attno == 0)
1311 66 : return false;
1312 :
1313 120 : switch (prop)
1314 : {
1315 12 : case AMPROP_DISTANCE_ORDERABLE:
1316 12 : break;
1317 108 : default:
1318 108 : return false;
1319 : }
1320 :
1321 : /*
1322 : * Currently, SP-GiST distance-ordered scans require that there be a
1323 : * distance operator in the opclass with the default types. So we assume
1324 : * that if such an operator exists, then there's a reason for it.
1325 : */
1326 :
1327 : /* First we need to know the column's opclass. */
1328 12 : opclass = get_index_column_opclass(index_oid, attno);
1329 12 : if (!OidIsValid(opclass))
1330 : {
1331 0 : *isnull = true;
1332 0 : return true;
1333 : }
1334 :
1335 : /* Now look up the opclass family and input datatype. */
1336 12 : if (!get_opclass_opfamily_and_input_type(opclass, &opfamily, &opcintype))
1337 : {
1338 0 : *isnull = true;
1339 0 : return true;
1340 : }
1341 :
1342 : /* And now we can check whether the operator is provided. */
1343 12 : catlist = SearchSysCacheList1(AMOPSTRATEGY,
1344 : ObjectIdGetDatum(opfamily));
1345 :
1346 12 : *res = false;
1347 :
1348 102 : for (i = 0; i < catlist->n_members; i++)
1349 : {
1350 96 : HeapTuple amoptup = &catlist->members[i]->tuple;
1351 96 : Form_pg_amop amopform = (Form_pg_amop) GETSTRUCT(amoptup);
1352 :
1353 96 : if (amopform->amoppurpose == AMOP_ORDER &&
1354 6 : (amopform->amoplefttype == opcintype ||
1355 6 : amopform->amoprighttype == opcintype) &&
1356 6 : opfamily_can_sort_type(amopform->amopsortfamily,
1357 : get_op_rettype(amopform->amopopr)))
1358 : {
1359 6 : *res = true;
1360 6 : break;
1361 : }
1362 : }
1363 :
1364 12 : ReleaseSysCacheList(catlist);
1365 :
1366 12 : *isnull = false;
1367 :
1368 12 : return true;
1369 : }
|
