Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * spgutils.c
4 : * various support functions for SP-GiST
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2024, 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 1238 : spghandler(PG_FUNCTION_ARGS)
45 : {
46 1238 : IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
47 :
48 1238 : amroutine->amstrategies = 0;
49 1238 : amroutine->amsupport = SPGISTNProc;
50 1238 : amroutine->amoptsprocnum = SPGIST_OPTIONS_PROC;
51 1238 : amroutine->amcanorder = false;
52 1238 : amroutine->amcanorderbyop = true;
53 1238 : amroutine->amcanbackward = false;
54 1238 : amroutine->amcanunique = false;
55 1238 : amroutine->amcanmulticol = false;
56 1238 : amroutine->amoptionalkey = true;
57 1238 : amroutine->amsearcharray = false;
58 1238 : amroutine->amsearchnulls = true;
59 1238 : amroutine->amstorage = true;
60 1238 : amroutine->amclusterable = false;
61 1238 : amroutine->ampredlocks = false;
62 1238 : amroutine->amcanparallel = false;
63 1238 : amroutine->amcanbuildparallel = false;
64 1238 : amroutine->amcaninclude = true;
65 1238 : amroutine->amusemaintenanceworkmem = false;
66 1238 : amroutine->amsummarizing = false;
67 1238 : amroutine->amparallelvacuumoptions =
68 : VACUUM_OPTION_PARALLEL_BULKDEL | VACUUM_OPTION_PARALLEL_COND_CLEANUP;
69 1238 : amroutine->amkeytype = InvalidOid;
70 :
71 1238 : amroutine->ambuild = spgbuild;
72 1238 : amroutine->ambuildempty = spgbuildempty;
73 1238 : amroutine->aminsert = spginsert;
74 1238 : amroutine->aminsertcleanup = NULL;
75 1238 : amroutine->ambulkdelete = spgbulkdelete;
76 1238 : amroutine->amvacuumcleanup = spgvacuumcleanup;
77 1238 : amroutine->amcanreturn = spgcanreturn;
78 1238 : amroutine->amcostestimate = spgcostestimate;
79 1238 : amroutine->amgettreeheight = NULL;
80 1238 : amroutine->amoptions = spgoptions;
81 1238 : amroutine->amproperty = spgproperty;
82 1238 : amroutine->ambuildphasename = NULL;
83 1238 : amroutine->amvalidate = spgvalidate;
84 1238 : amroutine->amadjustmembers = spgadjustmembers;
85 1238 : amroutine->ambeginscan = spgbeginscan;
86 1238 : amroutine->amrescan = spgrescan;
87 1238 : amroutine->amgettuple = spggettuple;
88 1238 : amroutine->amgetbitmap = spggetbitmap;
89 1238 : amroutine->amendscan = spgendscan;
90 1238 : amroutine->ammarkpos = NULL;
91 1238 : amroutine->amrestrpos = NULL;
92 1238 : amroutine->amestimateparallelscan = NULL;
93 1238 : amroutine->aminitparallelscan = NULL;
94 1238 : amroutine->amparallelrescan = NULL;
95 :
96 1238 : PG_RETURN_POINTER(amroutine);
97 : }
98 :
99 : /*
100 : * GetIndexInputType
101 : * Determine the nominal input data type for an index column
102 : *
103 : * We define the "nominal" input type as the associated opclass's opcintype,
104 : * or if that is a polymorphic type, the base type of the heap column or
105 : * expression that is the index's input. The reason for preferring the
106 : * opcintype is that non-polymorphic opclasses probably don't want to hear
107 : * about binary-compatible input types. For instance, if a text opclass
108 : * is being used with a varchar heap column, we want to report "text" not
109 : * "varchar". Likewise, opclasses don't want to hear about domain types,
110 : * so if we do consult the actual input type, we make sure to flatten domains.
111 : *
112 : * At some point maybe this should go somewhere else, but it's not clear
113 : * if any other index AMs have a use for it.
114 : */
115 : static Oid
116 404 : GetIndexInputType(Relation index, AttrNumber indexcol)
117 : {
118 : Oid opcintype;
119 : AttrNumber heapcol;
120 : List *indexprs;
121 : ListCell *indexpr_item;
122 :
123 : Assert(index->rd_index != NULL);
124 : Assert(indexcol > 0 && indexcol <= index->rd_index->indnkeyatts);
125 404 : opcintype = index->rd_opcintype[indexcol - 1];
126 404 : if (!IsPolymorphicType(opcintype))
127 300 : return opcintype;
128 104 : heapcol = index->rd_index->indkey.values[indexcol - 1];
129 104 : if (heapcol != 0) /* Simple index column? */
130 92 : return getBaseType(get_atttype(index->rd_index->indrelid, heapcol));
131 :
132 : /*
133 : * If the index expressions are already cached, skip calling
134 : * RelationGetIndexExpressions, as it will make a copy which is overkill.
135 : * We're not going to modify the trees, and we're not going to do anything
136 : * that would invalidate the relcache entry before we're done.
137 : */
138 12 : if (index->rd_indexprs)
139 0 : indexprs = index->rd_indexprs;
140 : else
141 12 : indexprs = RelationGetIndexExpressions(index);
142 12 : indexpr_item = list_head(indexprs);
143 12 : for (int i = 1; i <= index->rd_index->indnkeyatts; i++)
144 : {
145 12 : if (index->rd_index->indkey.values[i - 1] == 0)
146 : {
147 : /* expression column */
148 12 : if (indexpr_item == NULL)
149 0 : elog(ERROR, "wrong number of index expressions");
150 12 : if (i == indexcol)
151 12 : return getBaseType(exprType((Node *) lfirst(indexpr_item)));
152 0 : indexpr_item = lnext(indexprs, indexpr_item);
153 : }
154 : }
155 0 : elog(ERROR, "wrong number of index expressions");
156 : return InvalidOid; /* keep compiler quiet */
157 : }
158 :
159 : /* Fill in a SpGistTypeDesc struct with info about the specified data type */
160 : static void
161 1242 : fillTypeDesc(SpGistTypeDesc *desc, Oid type)
162 : {
163 : HeapTuple tp;
164 : Form_pg_type typtup;
165 :
166 1242 : desc->type = type;
167 1242 : tp = SearchSysCache1(TYPEOID, ObjectIdGetDatum(type));
168 1242 : if (!HeapTupleIsValid(tp))
169 0 : elog(ERROR, "cache lookup failed for type %u", type);
170 1242 : typtup = (Form_pg_type) GETSTRUCT(tp);
171 1242 : desc->attlen = typtup->typlen;
172 1242 : desc->attbyval = typtup->typbyval;
173 1242 : desc->attalign = typtup->typalign;
174 1242 : desc->attstorage = typtup->typstorage;
175 1242 : ReleaseSysCache(tp);
176 1242 : }
177 :
178 : /*
179 : * Fetch local cache of AM-specific info about the index, initializing it
180 : * if necessary
181 : */
182 : SpGistCache *
183 2670492 : spgGetCache(Relation index)
184 : {
185 : SpGistCache *cache;
186 :
187 2670492 : if (index->rd_amcache == NULL)
188 : {
189 : Oid atttype;
190 : spgConfigIn in;
191 : FmgrInfo *procinfo;
192 :
193 404 : cache = MemoryContextAllocZero(index->rd_indexcxt,
194 : sizeof(SpGistCache));
195 :
196 : /* SPGiST must have one key column and can also have INCLUDE columns */
197 : Assert(IndexRelationGetNumberOfKeyAttributes(index) == 1);
198 : Assert(IndexRelationGetNumberOfAttributes(index) <= INDEX_MAX_KEYS);
199 :
200 : /*
201 : * Get the actual (well, nominal) data type of the key column. We
202 : * pass this to the opclass config function so that polymorphic
203 : * opclasses are possible.
204 : */
205 404 : atttype = GetIndexInputType(index, spgKeyColumn + 1);
206 :
207 : /* Call the config function to get config info for the opclass */
208 404 : in.attType = atttype;
209 :
210 404 : procinfo = index_getprocinfo(index, 1, SPGIST_CONFIG_PROC);
211 404 : FunctionCall2Coll(procinfo,
212 404 : index->rd_indcollation[spgKeyColumn],
213 : PointerGetDatum(&in),
214 404 : PointerGetDatum(&cache->config));
215 :
216 : /*
217 : * If leafType isn't specified, use the declared index column type,
218 : * which index.c will have derived from the opclass's opcintype.
219 : * (Although we now make spgvalidate.c warn if these aren't the same,
220 : * old user-defined opclasses may not set the STORAGE parameter
221 : * correctly, so believe leafType if it's given.)
222 : */
223 404 : if (!OidIsValid(cache->config.leafType))
224 : {
225 374 : cache->config.leafType =
226 374 : TupleDescAttr(RelationGetDescr(index), spgKeyColumn)->atttypid;
227 :
228 : /*
229 : * If index column type is binary-coercible to atttype (for
230 : * example, it's a domain over atttype), treat it as plain atttype
231 : * to avoid thinking we need to compress.
232 : */
233 388 : if (cache->config.leafType != atttype &&
234 14 : IsBinaryCoercible(cache->config.leafType, atttype))
235 14 : cache->config.leafType = atttype;
236 : }
237 :
238 : /* Get the information we need about each relevant datatype */
239 404 : fillTypeDesc(&cache->attType, atttype);
240 :
241 404 : if (cache->config.leafType != atttype)
242 : {
243 30 : if (!OidIsValid(index_getprocid(index, 1, SPGIST_COMPRESS_PROC)))
244 0 : ereport(ERROR,
245 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
246 : errmsg("compress method must be defined when leaf type is different from input type")));
247 :
248 30 : fillTypeDesc(&cache->attLeafType, cache->config.leafType);
249 : }
250 : else
251 : {
252 : /* Save lookups in this common case */
253 374 : cache->attLeafType = cache->attType;
254 : }
255 :
256 404 : fillTypeDesc(&cache->attPrefixType, cache->config.prefixType);
257 404 : fillTypeDesc(&cache->attLabelType, cache->config.labelType);
258 :
259 : /*
260 : * Finally, if it's a real index (not a partitioned one), get the
261 : * lastUsedPages data from the metapage
262 : */
263 404 : if (index->rd_rel->relkind != RELKIND_PARTITIONED_INDEX)
264 : {
265 : Buffer metabuffer;
266 : SpGistMetaPageData *metadata;
267 :
268 398 : metabuffer = ReadBuffer(index, SPGIST_METAPAGE_BLKNO);
269 398 : LockBuffer(metabuffer, BUFFER_LOCK_SHARE);
270 :
271 398 : metadata = SpGistPageGetMeta(BufferGetPage(metabuffer));
272 :
273 398 : if (metadata->magicNumber != SPGIST_MAGIC_NUMBER)
274 0 : elog(ERROR, "index \"%s\" is not an SP-GiST index",
275 : RelationGetRelationName(index));
276 :
277 398 : cache->lastUsedPages = metadata->lastUsedPages;
278 :
279 398 : UnlockReleaseBuffer(metabuffer);
280 : }
281 :
282 404 : index->rd_amcache = (void *) cache;
283 : }
284 : else
285 : {
286 : /* assume it's up to date */
287 2670088 : cache = (SpGistCache *) index->rd_amcache;
288 : }
289 :
290 2670492 : return cache;
291 : }
292 :
293 : /*
294 : * Compute a tuple descriptor for leaf tuples or index-only-scan result tuples.
295 : *
296 : * We can use the relcache's tupdesc as-is in many cases, and it's always
297 : * OK so far as any INCLUDE columns are concerned. However, the entry for
298 : * the key column has to match leafType in the first case or attType in the
299 : * second case. While the relcache's tupdesc *should* show leafType, this
300 : * might not hold for legacy user-defined opclasses, since before v14 they
301 : * were not allowed to declare their true storage type in CREATE OPCLASS.
302 : * Also, attType can be different from what is in the relcache.
303 : *
304 : * This function gives back either a pointer to the relcache's tupdesc
305 : * if that is suitable, or a palloc'd copy that's been adjusted to match
306 : * the specified key column type. We can avoid doing any catalog lookups
307 : * here by insisting that the caller pass an SpGistTypeDesc not just an OID.
308 : */
309 : TupleDesc
310 244742 : getSpGistTupleDesc(Relation index, SpGistTypeDesc *keyType)
311 : {
312 : TupleDesc outTupDesc;
313 : Form_pg_attribute att;
314 :
315 244742 : if (keyType->type ==
316 244742 : TupleDescAttr(RelationGetDescr(index), spgKeyColumn)->atttypid)
317 244612 : outTupDesc = RelationGetDescr(index);
318 : else
319 : {
320 130 : outTupDesc = CreateTupleDescCopy(RelationGetDescr(index));
321 130 : att = TupleDescAttr(outTupDesc, spgKeyColumn);
322 : /* It's sufficient to update the type-dependent fields of the column */
323 130 : att->atttypid = keyType->type;
324 130 : att->atttypmod = -1;
325 130 : att->attlen = keyType->attlen;
326 130 : att->attbyval = keyType->attbyval;
327 130 : att->attalign = keyType->attalign;
328 130 : att->attstorage = keyType->attstorage;
329 : /* We shouldn't need to bother with making these valid: */
330 130 : att->attcompression = InvalidCompressionMethod;
331 130 : att->attcollation = InvalidOid;
332 : /* In case we changed typlen, we'd better reset following offsets */
333 146 : for (int i = spgFirstIncludeColumn; i < outTupDesc->natts; i++)
334 16 : TupleDescAttr(outTupDesc, i)->attcacheoff = -1;
335 : }
336 244742 : return outTupDesc;
337 : }
338 :
339 : /* Initialize SpGistState for working with the given index */
340 : void
341 243838 : initSpGistState(SpGistState *state, Relation index)
342 : {
343 : SpGistCache *cache;
344 :
345 243838 : state->index = index;
346 :
347 : /* Get cached static information about index */
348 243838 : cache = spgGetCache(index);
349 :
350 243838 : state->config = cache->config;
351 243838 : state->attType = cache->attType;
352 243838 : state->attLeafType = cache->attLeafType;
353 243838 : state->attPrefixType = cache->attPrefixType;
354 243838 : state->attLabelType = cache->attLabelType;
355 :
356 : /* Ensure we have a valid descriptor for leaf tuples */
357 243838 : state->leafTupDesc = getSpGistTupleDesc(state->index, &state->attLeafType);
358 :
359 : /* Make workspace for constructing dead tuples */
360 243838 : state->deadTupleStorage = palloc0(SGDTSIZE);
361 :
362 : /*
363 : * Set horizon XID to use in redirection tuples. Use our own XID if we
364 : * have one, else use InvalidTransactionId. The latter case can happen in
365 : * VACUUM or REINDEX CONCURRENTLY, and in neither case would it be okay to
366 : * force an XID to be assigned. VACUUM won't create any redirection
367 : * tuples anyway, but REINDEX CONCURRENTLY can. Fortunately, REINDEX
368 : * CONCURRENTLY doesn't mark the index valid until the end, so there could
369 : * never be any concurrent scans "in flight" to a redirection tuple it has
370 : * inserted. And it locks out VACUUM until the end, too. So it's okay
371 : * for VACUUM to immediately expire a redirection tuple that contains an
372 : * invalid xid.
373 : */
374 243838 : state->redirectXid = GetTopTransactionIdIfAny();
375 :
376 : /* Assume we're not in an index build (spgbuild will override) */
377 243838 : state->isBuild = false;
378 243838 : }
379 :
380 : /*
381 : * Allocate a new page (either by recycling, or by extending the index file).
382 : *
383 : * The returned buffer is already pinned and exclusive-locked.
384 : * Caller is responsible for initializing the page by calling SpGistInitBuffer.
385 : */
386 : Buffer
387 6546 : SpGistNewBuffer(Relation index)
388 : {
389 : Buffer buffer;
390 :
391 : /* First, try to get a page from FSM */
392 : for (;;)
393 0 : {
394 6546 : BlockNumber blkno = GetFreeIndexPage(index);
395 :
396 6546 : if (blkno == InvalidBlockNumber)
397 6538 : break; /* nothing known to FSM */
398 :
399 : /*
400 : * The fixed pages shouldn't ever be listed in FSM, but just in case
401 : * one is, ignore it.
402 : */
403 8 : if (SpGistBlockIsFixed(blkno))
404 0 : continue;
405 :
406 8 : buffer = ReadBuffer(index, blkno);
407 :
408 : /*
409 : * We have to guard against the possibility that someone else already
410 : * recycled this page; the buffer may be locked if so.
411 : */
412 8 : if (ConditionalLockBuffer(buffer))
413 : {
414 8 : Page page = BufferGetPage(buffer);
415 :
416 8 : if (PageIsNew(page))
417 0 : return buffer; /* OK to use, if never initialized */
418 :
419 8 : if (SpGistPageIsDeleted(page) || PageIsEmpty(page))
420 8 : return buffer; /* OK to use */
421 :
422 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
423 : }
424 :
425 : /* Can't use it, so release buffer and try again */
426 0 : ReleaseBuffer(buffer);
427 : }
428 :
429 6538 : buffer = ExtendBufferedRel(BMR_REL(index), MAIN_FORKNUM, NULL,
430 : EB_LOCK_FIRST);
431 :
432 6538 : return buffer;
433 : }
434 :
435 : /*
436 : * Update index metapage's lastUsedPages info from local cache, if possible
437 : *
438 : * Updating meta page isn't critical for index working, so
439 : * 1 use ConditionalLockBuffer to improve concurrency
440 : * 2 don't WAL-log metabuffer changes to decrease WAL traffic
441 : */
442 : void
443 242866 : SpGistUpdateMetaPage(Relation index)
444 : {
445 242866 : SpGistCache *cache = (SpGistCache *) index->rd_amcache;
446 :
447 242866 : if (cache != NULL)
448 : {
449 : Buffer metabuffer;
450 :
451 242866 : metabuffer = ReadBuffer(index, SPGIST_METAPAGE_BLKNO);
452 :
453 242866 : if (ConditionalLockBuffer(metabuffer))
454 : {
455 242858 : Page metapage = BufferGetPage(metabuffer);
456 242858 : SpGistMetaPageData *metadata = SpGistPageGetMeta(metapage);
457 :
458 242858 : metadata->lastUsedPages = cache->lastUsedPages;
459 :
460 : /*
461 : * Set pd_lower just past the end of the metadata. This is
462 : * essential, because without doing so, metadata will be lost if
463 : * xlog.c compresses the page. (We must do this here because
464 : * pre-v11 versions of PG did not set the metapage's pd_lower
465 : * correctly, so a pg_upgraded index might contain the wrong
466 : * value.)
467 : */
468 242858 : ((PageHeader) metapage)->pd_lower =
469 242858 : ((char *) metadata + sizeof(SpGistMetaPageData)) - (char *) metapage;
470 :
471 242858 : MarkBufferDirty(metabuffer);
472 242858 : UnlockReleaseBuffer(metabuffer);
473 : }
474 : else
475 : {
476 8 : ReleaseBuffer(metabuffer);
477 : }
478 : }
479 242866 : }
480 :
481 : /* Macro to select proper element of lastUsedPages cache depending on flags */
482 : /* Masking flags with SPGIST_CACHED_PAGES is just for paranoia's sake */
483 : #define GET_LUP(c, f) (&(c)->lastUsedPages.cachedPage[((unsigned int) (f)) % SPGIST_CACHED_PAGES])
484 :
485 : /*
486 : * Allocate and initialize a new buffer of the type and parity specified by
487 : * flags. The returned buffer is already pinned and exclusive-locked.
488 : *
489 : * When requesting an inner page, if we get one with the wrong parity,
490 : * we just release the buffer and try again. We will get a different page
491 : * because GetFreeIndexPage will have marked the page used in FSM. The page
492 : * is entered in our local lastUsedPages cache, so there's some hope of
493 : * making use of it later in this session, but otherwise we rely on VACUUM
494 : * to eventually re-enter the page in FSM, making it available for recycling.
495 : * Note that such a page does not get marked dirty here, so unless it's used
496 : * fairly soon, the buffer will just get discarded and the page will remain
497 : * as it was on disk.
498 : *
499 : * When we return a buffer to the caller, the page is *not* entered into
500 : * the lastUsedPages cache; we expect the caller will do so after it's taken
501 : * whatever space it will use. This is because after the caller has used up
502 : * some space, the page might have less space than whatever was cached already
503 : * so we'd rather not trash the old cache entry.
504 : */
505 : static Buffer
506 5836 : allocNewBuffer(Relation index, int flags)
507 : {
508 5836 : SpGistCache *cache = spgGetCache(index);
509 5836 : uint16 pageflags = 0;
510 :
511 5836 : if (GBUF_REQ_LEAF(flags))
512 5730 : pageflags |= SPGIST_LEAF;
513 5836 : if (GBUF_REQ_NULLS(flags))
514 0 : pageflags |= SPGIST_NULLS;
515 :
516 : for (;;)
517 86 : {
518 : Buffer buffer;
519 :
520 5922 : buffer = SpGistNewBuffer(index);
521 5922 : SpGistInitBuffer(buffer, pageflags);
522 :
523 5922 : if (pageflags & SPGIST_LEAF)
524 : {
525 : /* Leaf pages have no parity concerns, so just use it */
526 5730 : return buffer;
527 : }
528 : else
529 : {
530 192 : BlockNumber blkno = BufferGetBlockNumber(buffer);
531 192 : int blkFlags = GBUF_INNER_PARITY(blkno);
532 :
533 192 : if ((flags & GBUF_PARITY_MASK) == blkFlags)
534 : {
535 : /* Page has right parity, use it */
536 106 : return buffer;
537 : }
538 : else
539 : {
540 : /* Page has wrong parity, record it in cache and try again */
541 86 : if (pageflags & SPGIST_NULLS)
542 0 : blkFlags |= GBUF_NULLS;
543 86 : cache->lastUsedPages.cachedPage[blkFlags].blkno = blkno;
544 86 : cache->lastUsedPages.cachedPage[blkFlags].freeSpace =
545 86 : PageGetExactFreeSpace(BufferGetPage(buffer));
546 86 : UnlockReleaseBuffer(buffer);
547 : }
548 : }
549 : }
550 : }
551 :
552 : /*
553 : * Get a buffer of the type and parity specified by flags, having at least
554 : * as much free space as indicated by needSpace. We use the lastUsedPages
555 : * cache to assign the same buffer previously requested when possible.
556 : * The returned buffer is already pinned and exclusive-locked.
557 : *
558 : * *isNew is set true if the page was initialized here, false if it was
559 : * already valid.
560 : */
561 : Buffer
562 10892 : SpGistGetBuffer(Relation index, int flags, int needSpace, bool *isNew)
563 : {
564 10892 : SpGistCache *cache = spgGetCache(index);
565 : SpGistLastUsedPage *lup;
566 :
567 : /* Bail out if even an empty page wouldn't meet the demand */
568 10892 : if (needSpace > SPGIST_PAGE_CAPACITY)
569 0 : elog(ERROR, "desired SPGiST tuple size is too big");
570 :
571 : /*
572 : * If possible, increase the space request to include relation's
573 : * fillfactor. This ensures that when we add unrelated tuples to a page,
574 : * we try to keep 100-fillfactor% available for adding tuples that are
575 : * related to the ones already on it. But fillfactor mustn't cause an
576 : * error for requests that would otherwise be legal.
577 : */
578 10892 : needSpace += SpGistGetTargetPageFreeSpace(index);
579 10892 : needSpace = Min(needSpace, SPGIST_PAGE_CAPACITY);
580 :
581 : /* Get the cache entry for this flags setting */
582 10892 : lup = GET_LUP(cache, flags);
583 :
584 : /* If we have nothing cached, just turn it over to allocNewBuffer */
585 10892 : if (lup->blkno == InvalidBlockNumber)
586 : {
587 182 : *isNew = true;
588 182 : return allocNewBuffer(index, flags);
589 : }
590 :
591 : /* fixed pages should never be in cache */
592 : Assert(!SpGistBlockIsFixed(lup->blkno));
593 :
594 : /* If cached freeSpace isn't enough, don't bother looking at the page */
595 10710 : if (lup->freeSpace >= needSpace)
596 : {
597 : Buffer buffer;
598 : Page page;
599 :
600 5056 : buffer = ReadBuffer(index, lup->blkno);
601 :
602 5056 : if (!ConditionalLockBuffer(buffer))
603 : {
604 : /*
605 : * buffer is locked by another process, so return a new buffer
606 : */
607 0 : ReleaseBuffer(buffer);
608 0 : *isNew = true;
609 0 : return allocNewBuffer(index, flags);
610 : }
611 :
612 5056 : page = BufferGetPage(buffer);
613 :
614 5056 : if (PageIsNew(page) || SpGistPageIsDeleted(page) || PageIsEmpty(page))
615 : {
616 : /* OK to initialize the page */
617 186 : uint16 pageflags = 0;
618 :
619 186 : if (GBUF_REQ_LEAF(flags))
620 180 : pageflags |= SPGIST_LEAF;
621 186 : if (GBUF_REQ_NULLS(flags))
622 0 : pageflags |= SPGIST_NULLS;
623 186 : SpGistInitBuffer(buffer, pageflags);
624 186 : lup->freeSpace = PageGetExactFreeSpace(page) - needSpace;
625 186 : *isNew = true;
626 186 : return buffer;
627 : }
628 :
629 : /*
630 : * Check that page is of right type and has enough space. We must
631 : * recheck this since our cache isn't necessarily up to date.
632 : */
633 9740 : if ((GBUF_REQ_LEAF(flags) ? SpGistPageIsLeaf(page) : !SpGistPageIsLeaf(page)) &&
634 4870 : (GBUF_REQ_NULLS(flags) ? SpGistPageStoresNulls(page) : !SpGistPageStoresNulls(page)))
635 : {
636 4870 : int freeSpace = PageGetExactFreeSpace(page);
637 :
638 4870 : if (freeSpace >= needSpace)
639 : {
640 : /* Success, update freespace info and return the buffer */
641 4870 : lup->freeSpace = freeSpace - needSpace;
642 4870 : *isNew = false;
643 4870 : return buffer;
644 : }
645 : }
646 :
647 : /*
648 : * fallback to allocation of new buffer
649 : */
650 0 : UnlockReleaseBuffer(buffer);
651 : }
652 :
653 : /* No success with cache, so return a new buffer */
654 5654 : *isNew = true;
655 5654 : return allocNewBuffer(index, flags);
656 : }
657 :
658 : /*
659 : * Update lastUsedPages cache when done modifying a page.
660 : *
661 : * We update the appropriate cache entry if it already contained this page
662 : * (its freeSpace is likely obsolete), or if this page has more space than
663 : * whatever we had cached.
664 : */
665 : void
666 2408100 : SpGistSetLastUsedPage(Relation index, Buffer buffer)
667 : {
668 2408100 : SpGistCache *cache = spgGetCache(index);
669 : SpGistLastUsedPage *lup;
670 : int freeSpace;
671 2408100 : Page page = BufferGetPage(buffer);
672 2408100 : BlockNumber blkno = BufferGetBlockNumber(buffer);
673 : int flags;
674 :
675 : /* Never enter fixed pages (root pages) in cache, though */
676 2408100 : if (SpGistBlockIsFixed(blkno))
677 805444 : return;
678 :
679 1602656 : if (SpGistPageIsLeaf(page))
680 817594 : flags = GBUF_LEAF;
681 : else
682 785062 : flags = GBUF_INNER_PARITY(blkno);
683 1602656 : if (SpGistPageStoresNulls(page))
684 0 : flags |= GBUF_NULLS;
685 :
686 1602656 : lup = GET_LUP(cache, flags);
687 :
688 1602656 : freeSpace = PageGetExactFreeSpace(page);
689 1602656 : if (lup->blkno == InvalidBlockNumber || lup->blkno == blkno ||
690 436600 : lup->freeSpace < freeSpace)
691 : {
692 1174722 : lup->blkno = blkno;
693 1174722 : lup->freeSpace = freeSpace;
694 : }
695 : }
696 :
697 : /*
698 : * Initialize an SPGiST page to empty, with specified flags
699 : */
700 : void
701 7602 : SpGistInitPage(Page page, uint16 f)
702 : {
703 : SpGistPageOpaque opaque;
704 :
705 7602 : PageInit(page, BLCKSZ, sizeof(SpGistPageOpaqueData));
706 7602 : opaque = SpGistPageGetOpaque(page);
707 7602 : opaque->flags = f;
708 7602 : opaque->spgist_page_id = SPGIST_PAGE_ID;
709 7602 : }
710 :
711 : /*
712 : * Initialize a buffer's page to empty, with specified flags
713 : */
714 : void
715 7370 : SpGistInitBuffer(Buffer b, uint16 f)
716 : {
717 : Assert(BufferGetPageSize(b) == BLCKSZ);
718 7370 : SpGistInitPage(BufferGetPage(b), f);
719 7370 : }
720 :
721 : /*
722 : * Initialize metadata page
723 : */
724 : void
725 216 : SpGistInitMetapage(Page page)
726 : {
727 : SpGistMetaPageData *metadata;
728 : int i;
729 :
730 216 : SpGistInitPage(page, SPGIST_META);
731 216 : metadata = SpGistPageGetMeta(page);
732 216 : memset(metadata, 0, sizeof(SpGistMetaPageData));
733 216 : metadata->magicNumber = SPGIST_MAGIC_NUMBER;
734 :
735 : /* initialize last-used-page cache to empty */
736 1944 : for (i = 0; i < SPGIST_CACHED_PAGES; i++)
737 1728 : metadata->lastUsedPages.cachedPage[i].blkno = InvalidBlockNumber;
738 :
739 : /*
740 : * Set pd_lower just past the end of the metadata. This is essential,
741 : * because without doing so, metadata will be lost if xlog.c compresses
742 : * the page.
743 : */
744 216 : ((PageHeader) page)->pd_lower =
745 216 : ((char *) metadata + sizeof(SpGistMetaPageData)) - (char *) page;
746 216 : }
747 :
748 : /*
749 : * reloptions processing for SPGiST
750 : */
751 : bytea *
752 118 : spgoptions(Datum reloptions, bool validate)
753 : {
754 : static const relopt_parse_elt tab[] = {
755 : {"fillfactor", RELOPT_TYPE_INT, offsetof(SpGistOptions, fillfactor)},
756 : };
757 :
758 118 : return (bytea *) build_reloptions(reloptions, validate,
759 : RELOPT_KIND_SPGIST,
760 : sizeof(SpGistOptions),
761 : tab, lengthof(tab));
762 : }
763 :
764 : /*
765 : * Get the space needed to store a non-null datum of the indicated type
766 : * in an inner tuple (that is, as a prefix or node label).
767 : * Note the result is already rounded up to a MAXALIGN boundary.
768 : * Here we follow the convention that pass-by-val types are just stored
769 : * in their Datum representation (compare memcpyInnerDatum).
770 : */
771 : unsigned int
772 12314 : SpGistGetInnerTypeSize(SpGistTypeDesc *att, Datum datum)
773 : {
774 : unsigned int size;
775 :
776 12314 : if (att->attbyval)
777 6460 : size = sizeof(Datum);
778 5854 : else if (att->attlen > 0)
779 3968 : size = att->attlen;
780 : else
781 1886 : size = VARSIZE_ANY(datum);
782 :
783 12314 : return MAXALIGN(size);
784 : }
785 :
786 : /*
787 : * Copy the given non-null datum to *target, in the inner-tuple case
788 : */
789 : static void
790 12314 : memcpyInnerDatum(void *target, SpGistTypeDesc *att, Datum datum)
791 : {
792 : unsigned int size;
793 :
794 12314 : if (att->attbyval)
795 : {
796 6460 : memcpy(target, &datum, sizeof(Datum));
797 : }
798 : else
799 : {
800 5854 : size = (att->attlen > 0) ? att->attlen : VARSIZE_ANY(datum);
801 5854 : memcpy(target, DatumGetPointer(datum), size);
802 : }
803 12314 : }
804 :
805 : /*
806 : * Compute space required for a leaf tuple holding the given data.
807 : *
808 : * This must match the size-calculation portion of spgFormLeafTuple.
809 : */
810 : Size
811 19458994 : SpGistGetLeafTupleSize(TupleDesc tupleDescriptor,
812 : const Datum *datums, const bool *isnulls)
813 : {
814 : Size size;
815 : Size data_size;
816 19458994 : bool needs_null_mask = false;
817 19458994 : int natts = tupleDescriptor->natts;
818 :
819 : /*
820 : * Decide whether we need a nulls bitmask.
821 : *
822 : * If there is only a key attribute (natts == 1), never use a bitmask, for
823 : * compatibility with the pre-v14 layout of leaf tuples. Otherwise, we
824 : * need one if any attribute is null.
825 : */
826 19458994 : if (natts > 1)
827 : {
828 1004074 : for (int i = 0; i < natts; i++)
829 : {
830 687438 : if (isnulls[i])
831 : {
832 18254 : needs_null_mask = true;
833 18254 : break;
834 : }
835 : }
836 : }
837 :
838 : /*
839 : * Calculate size of the data part; same as for heap tuples.
840 : */
841 19458994 : data_size = heap_compute_data_size(tupleDescriptor, datums, isnulls);
842 :
843 : /*
844 : * Compute total size.
845 : */
846 19458994 : size = SGLTHDRSZ(needs_null_mask);
847 19458994 : size += data_size;
848 19458994 : size = MAXALIGN(size);
849 :
850 : /*
851 : * Ensure that we can replace the tuple with a dead tuple later. This test
852 : * is unnecessary when there are any non-null attributes, but be safe.
853 : */
854 19458994 : if (size < SGDTSIZE)
855 0 : size = SGDTSIZE;
856 :
857 19458994 : return size;
858 : }
859 :
860 : /*
861 : * Construct a leaf tuple containing the given heap TID and datum values
862 : */
863 : SpGistLeafTuple
864 1527694 : spgFormLeafTuple(SpGistState *state, ItemPointer heapPtr,
865 : const Datum *datums, const bool *isnulls)
866 : {
867 : SpGistLeafTuple tup;
868 1527694 : TupleDesc tupleDescriptor = state->leafTupDesc;
869 : Size size;
870 : Size hoff;
871 : Size data_size;
872 1527694 : bool needs_null_mask = false;
873 1527694 : int natts = tupleDescriptor->natts;
874 : char *tp; /* ptr to tuple data */
875 1527694 : uint16 tupmask = 0; /* unused heap_fill_tuple output */
876 :
877 : /*
878 : * Decide whether we need a nulls bitmask.
879 : *
880 : * If there is only a key attribute (natts == 1), never use a bitmask, for
881 : * compatibility with the pre-v14 layout of leaf tuples. Otherwise, we
882 : * need one if any attribute is null.
883 : */
884 1527694 : if (natts > 1)
885 : {
886 425884 : for (int i = 0; i < natts; i++)
887 : {
888 295466 : if (isnulls[i])
889 : {
890 11662 : needs_null_mask = true;
891 11662 : break;
892 : }
893 : }
894 : }
895 :
896 : /*
897 : * Calculate size of the data part; same as for heap tuples.
898 : */
899 1527694 : data_size = heap_compute_data_size(tupleDescriptor, datums, isnulls);
900 :
901 : /*
902 : * Compute total size.
903 : */
904 1527694 : hoff = SGLTHDRSZ(needs_null_mask);
905 1527694 : size = hoff + data_size;
906 1527694 : size = MAXALIGN(size);
907 :
908 : /*
909 : * Ensure that we can replace the tuple with a dead tuple later. This test
910 : * is unnecessary when there are any non-null attributes, but be safe.
911 : */
912 1527694 : if (size < SGDTSIZE)
913 0 : size = SGDTSIZE;
914 :
915 : /* OK, form the tuple */
916 1527694 : tup = (SpGistLeafTuple) palloc0(size);
917 :
918 1527694 : tup->size = size;
919 1527694 : SGLT_SET_NEXTOFFSET(tup, InvalidOffsetNumber);
920 1527694 : tup->heapPtr = *heapPtr;
921 :
922 1527694 : tp = (char *) tup + hoff;
923 :
924 1527694 : if (needs_null_mask)
925 : {
926 : bits8 *bp; /* ptr to null bitmap in tuple */
927 :
928 : /* Set nullmask presence bit in SpGistLeafTuple header */
929 11662 : SGLT_SET_HASNULLMASK(tup, true);
930 : /* Fill the data area and null mask */
931 11662 : bp = (bits8 *) ((char *) tup + sizeof(SpGistLeafTupleData));
932 11662 : heap_fill_tuple(tupleDescriptor, datums, isnulls, tp, data_size,
933 : &tupmask, bp);
934 : }
935 1516032 : else if (natts > 1 || !isnulls[spgKeyColumn])
936 : {
937 : /* Fill data area only */
938 1515960 : heap_fill_tuple(tupleDescriptor, datums, isnulls, tp, data_size,
939 : &tupmask, (bits8 *) NULL);
940 : }
941 : /* otherwise we have no data, nor a bitmap, to fill */
942 :
943 1527694 : return tup;
944 : }
945 :
946 : /*
947 : * Construct a node (to go into an inner tuple) containing the given label
948 : *
949 : * Note that the node's downlink is just set invalid here. Caller will fill
950 : * it in later.
951 : */
952 : SpGistNodeTuple
953 40498 : spgFormNodeTuple(SpGistState *state, Datum label, bool isnull)
954 : {
955 : SpGistNodeTuple tup;
956 : unsigned int size;
957 40498 : unsigned short infomask = 0;
958 :
959 : /* compute space needed (note result is already maxaligned) */
960 40498 : size = SGNTHDRSZ;
961 40498 : if (!isnull)
962 5770 : size += SpGistGetInnerTypeSize(&state->attLabelType, label);
963 :
964 : /*
965 : * Here we make sure that the size will fit in the field reserved for it
966 : * in t_info.
967 : */
968 40498 : if ((size & INDEX_SIZE_MASK) != size)
969 0 : ereport(ERROR,
970 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
971 : errmsg("index row requires %zu bytes, maximum size is %zu",
972 : (Size) size, (Size) INDEX_SIZE_MASK)));
973 :
974 40498 : tup = (SpGistNodeTuple) palloc0(size);
975 :
976 40498 : if (isnull)
977 34728 : infomask |= INDEX_NULL_MASK;
978 : /* we don't bother setting the INDEX_VAR_MASK bit */
979 40498 : infomask |= size;
980 40498 : tup->t_info = infomask;
981 :
982 : /* The TID field will be filled in later */
983 40498 : ItemPointerSetInvalid(&tup->t_tid);
984 :
985 40498 : if (!isnull)
986 5770 : memcpyInnerDatum(SGNTDATAPTR(tup), &state->attLabelType, label);
987 :
988 40498 : return tup;
989 : }
990 :
991 : /*
992 : * Construct an inner tuple containing the given prefix and node array
993 : */
994 : SpGistInnerTuple
995 8538 : spgFormInnerTuple(SpGistState *state, bool hasPrefix, Datum prefix,
996 : int nNodes, SpGistNodeTuple *nodes)
997 : {
998 : SpGistInnerTuple tup;
999 : unsigned int size;
1000 : unsigned int prefixSize;
1001 : int i;
1002 : char *ptr;
1003 :
1004 : /* Compute size needed */
1005 8538 : if (hasPrefix)
1006 6544 : prefixSize = SpGistGetInnerTypeSize(&state->attPrefixType, prefix);
1007 : else
1008 1994 : prefixSize = 0;
1009 :
1010 8538 : size = SGITHDRSZ + prefixSize;
1011 :
1012 : /* Note: we rely on node tuple sizes to be maxaligned already */
1013 59308 : for (i = 0; i < nNodes; i++)
1014 50770 : size += IndexTupleSize(nodes[i]);
1015 :
1016 : /*
1017 : * Ensure that we can replace the tuple with a dead tuple later. This
1018 : * test is unnecessary given current tuple layouts, but let's be safe.
1019 : */
1020 8538 : if (size < SGDTSIZE)
1021 0 : size = SGDTSIZE;
1022 :
1023 : /*
1024 : * Inner tuple should be small enough to fit on a page
1025 : */
1026 8538 : if (size > SPGIST_PAGE_CAPACITY - sizeof(ItemIdData))
1027 0 : ereport(ERROR,
1028 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1029 : errmsg("SP-GiST inner tuple size %zu exceeds maximum %zu",
1030 : (Size) size,
1031 : SPGIST_PAGE_CAPACITY - sizeof(ItemIdData)),
1032 : errhint("Values larger than a buffer page cannot be indexed.")));
1033 :
1034 : /*
1035 : * Check for overflow of header fields --- probably can't fail if the
1036 : * above succeeded, but let's be paranoid
1037 : */
1038 8538 : if (size > SGITMAXSIZE ||
1039 8538 : prefixSize > SGITMAXPREFIXSIZE ||
1040 : nNodes > SGITMAXNNODES)
1041 0 : elog(ERROR, "SPGiST inner tuple header field is too small");
1042 :
1043 : /* OK, form the tuple */
1044 8538 : tup = (SpGistInnerTuple) palloc0(size);
1045 :
1046 8538 : tup->nNodes = nNodes;
1047 8538 : tup->prefixSize = prefixSize;
1048 8538 : tup->size = size;
1049 :
1050 8538 : if (hasPrefix)
1051 6544 : memcpyInnerDatum(SGITDATAPTR(tup), &state->attPrefixType, prefix);
1052 :
1053 8538 : ptr = (char *) SGITNODEPTR(tup);
1054 :
1055 59308 : for (i = 0; i < nNodes; i++)
1056 : {
1057 50770 : SpGistNodeTuple node = nodes[i];
1058 :
1059 50770 : memcpy(ptr, node, IndexTupleSize(node));
1060 50770 : ptr += IndexTupleSize(node);
1061 : }
1062 :
1063 8538 : return tup;
1064 : }
1065 :
1066 : /*
1067 : * Construct a "dead" tuple to replace a tuple being deleted.
1068 : *
1069 : * The state can be SPGIST_REDIRECT, SPGIST_DEAD, or SPGIST_PLACEHOLDER.
1070 : * For a REDIRECT tuple, a pointer (blkno+offset) must be supplied, and
1071 : * the xid field is filled in automatically.
1072 : *
1073 : * This is called in critical sections, so we don't use palloc; the tuple
1074 : * is built in preallocated storage. It should be copied before another
1075 : * call with different parameters can occur.
1076 : */
1077 : SpGistDeadTuple
1078 12868 : spgFormDeadTuple(SpGistState *state, int tupstate,
1079 : BlockNumber blkno, OffsetNumber offnum)
1080 : {
1081 12868 : SpGistDeadTuple tuple = (SpGistDeadTuple) state->deadTupleStorage;
1082 :
1083 12868 : tuple->tupstate = tupstate;
1084 12868 : tuple->size = SGDTSIZE;
1085 12868 : SGLT_SET_NEXTOFFSET(tuple, InvalidOffsetNumber);
1086 :
1087 12868 : if (tupstate == SPGIST_REDIRECT)
1088 : {
1089 2352 : ItemPointerSet(&tuple->pointer, blkno, offnum);
1090 2352 : tuple->xid = state->redirectXid;
1091 : }
1092 : else
1093 : {
1094 10516 : ItemPointerSetInvalid(&tuple->pointer);
1095 10516 : tuple->xid = InvalidTransactionId;
1096 : }
1097 :
1098 12868 : return tuple;
1099 : }
1100 :
1101 : /*
1102 : * Convert an SPGiST leaf tuple into Datum/isnull arrays.
1103 : *
1104 : * The caller must allocate sufficient storage for the output arrays.
1105 : * (INDEX_MAX_KEYS entries should be enough.)
1106 : */
1107 : void
1108 62608 : spgDeformLeafTuple(SpGistLeafTuple tup, TupleDesc tupleDescriptor,
1109 : Datum *datums, bool *isnulls, bool keyColumnIsNull)
1110 : {
1111 62608 : bool hasNullsMask = SGLT_GET_HASNULLMASK(tup);
1112 : char *tp; /* ptr to tuple data */
1113 : bits8 *bp; /* ptr to null bitmap in tuple */
1114 :
1115 62608 : if (keyColumnIsNull && tupleDescriptor->natts == 1)
1116 : {
1117 : /*
1118 : * Trivial case: there is only the key attribute and we're in a nulls
1119 : * tree. The hasNullsMask bit in the tuple header should not be set
1120 : * (and thus we can't use index_deform_tuple_internal), but
1121 : * nonetheless the result is NULL.
1122 : *
1123 : * Note: currently this is dead code, because noplace calls this when
1124 : * there is only the key attribute. But we should cover the case.
1125 : */
1126 : Assert(!hasNullsMask);
1127 :
1128 0 : datums[spgKeyColumn] = (Datum) 0;
1129 0 : isnulls[spgKeyColumn] = true;
1130 0 : return;
1131 : }
1132 :
1133 62608 : tp = (char *) tup + SGLTHDRSZ(hasNullsMask);
1134 62608 : bp = (bits8 *) ((char *) tup + sizeof(SpGistLeafTupleData));
1135 :
1136 62608 : index_deform_tuple_internal(tupleDescriptor,
1137 : datums, isnulls,
1138 : tp, bp, hasNullsMask);
1139 :
1140 : /*
1141 : * Key column isnull value from the tuple should be consistent with
1142 : * keyColumnIsNull flag from the caller.
1143 : */
1144 : Assert(keyColumnIsNull == isnulls[spgKeyColumn]);
1145 : }
1146 :
1147 : /*
1148 : * Extract the label datums of the nodes within innerTuple
1149 : *
1150 : * Returns NULL if label datums are NULLs
1151 : */
1152 : Datum *
1153 18680134 : spgExtractNodeLabels(SpGistState *state, SpGistInnerTuple innerTuple)
1154 : {
1155 : Datum *nodeLabels;
1156 : int i;
1157 : SpGistNodeTuple node;
1158 :
1159 : /* Either all the labels must be NULL, or none. */
1160 18680134 : node = SGITNODEPTR(innerTuple);
1161 18680134 : if (IndexTupleHasNulls(node))
1162 : {
1163 100684004 : SGITITERATE(innerTuple, i, node)
1164 : {
1165 82237518 : if (!IndexTupleHasNulls(node))
1166 0 : elog(ERROR, "some but not all node labels are null in SPGiST inner tuple");
1167 : }
1168 : /* They're all null, so just return NULL */
1169 18446486 : return NULL;
1170 : }
1171 : else
1172 : {
1173 233648 : nodeLabels = (Datum *) palloc(sizeof(Datum) * innerTuple->nNodes);
1174 2678410 : SGITITERATE(innerTuple, i, node)
1175 : {
1176 2444762 : if (IndexTupleHasNulls(node))
1177 0 : elog(ERROR, "some but not all node labels are null in SPGiST inner tuple");
1178 2444762 : nodeLabels[i] = SGNTDATUM(node, state);
1179 : }
1180 233648 : return nodeLabels;
1181 : }
1182 : }
1183 :
1184 : /*
1185 : * Add a new item to the page, replacing a PLACEHOLDER item if possible.
1186 : * Return the location it's inserted at, or InvalidOffsetNumber on failure.
1187 : *
1188 : * If startOffset isn't NULL, we start searching for placeholders at
1189 : * *startOffset, and update that to the next place to search. This is just
1190 : * an optimization for repeated insertions.
1191 : *
1192 : * If errorOK is false, we throw error when there's not enough room,
1193 : * rather than returning InvalidOffsetNumber.
1194 : */
1195 : OffsetNumber
1196 1618148 : SpGistPageAddNewItem(SpGistState *state, Page page, Item item, Size size,
1197 : OffsetNumber *startOffset, bool errorOK)
1198 : {
1199 1618148 : SpGistPageOpaque opaque = SpGistPageGetOpaque(page);
1200 : OffsetNumber i,
1201 : maxoff,
1202 : offnum;
1203 :
1204 1618148 : if (opaque->nPlaceholder > 0 &&
1205 462278 : PageGetExactFreeSpace(page) + SGDTSIZE >= MAXALIGN(size))
1206 : {
1207 : /* Try to replace a placeholder */
1208 462278 : maxoff = PageGetMaxOffsetNumber(page);
1209 462278 : offnum = InvalidOffsetNumber;
1210 :
1211 : for (;;)
1212 : {
1213 462278 : if (startOffset && *startOffset != InvalidOffsetNumber)
1214 115026 : i = *startOffset;
1215 : else
1216 347252 : i = FirstOffsetNumber;
1217 31712578 : for (; i <= maxoff; i++)
1218 : {
1219 31712578 : SpGistDeadTuple it = (SpGistDeadTuple) PageGetItem(page,
1220 : PageGetItemId(page, i));
1221 :
1222 31712578 : if (it->tupstate == SPGIST_PLACEHOLDER)
1223 : {
1224 462278 : offnum = i;
1225 462278 : break;
1226 : }
1227 : }
1228 :
1229 : /* Done if we found a placeholder */
1230 462278 : if (offnum != InvalidOffsetNumber)
1231 462278 : break;
1232 :
1233 0 : if (startOffset && *startOffset != InvalidOffsetNumber)
1234 : {
1235 : /* Hint was no good, re-search from beginning */
1236 0 : *startOffset = InvalidOffsetNumber;
1237 0 : continue;
1238 : }
1239 :
1240 : /* Hmm, no placeholder found? */
1241 0 : opaque->nPlaceholder = 0;
1242 0 : break;
1243 : }
1244 :
1245 462278 : if (offnum != InvalidOffsetNumber)
1246 : {
1247 : /* Replace the placeholder tuple */
1248 462278 : PageIndexTupleDelete(page, offnum);
1249 :
1250 462278 : offnum = PageAddItem(page, item, size, offnum, false, false);
1251 :
1252 : /*
1253 : * We should not have failed given the size check at the top of
1254 : * the function, but test anyway. If we did fail, we must PANIC
1255 : * because we've already deleted the placeholder tuple, and
1256 : * there's no other way to keep the damage from getting to disk.
1257 : */
1258 462278 : if (offnum != InvalidOffsetNumber)
1259 : {
1260 : Assert(opaque->nPlaceholder > 0);
1261 462278 : opaque->nPlaceholder--;
1262 462278 : if (startOffset)
1263 117666 : *startOffset = offnum + 1;
1264 : }
1265 : else
1266 0 : elog(PANIC, "failed to add item of size %zu to SPGiST index page",
1267 : size);
1268 :
1269 462278 : return offnum;
1270 : }
1271 : }
1272 :
1273 : /* No luck in replacing a placeholder, so just add it to the page */
1274 1155870 : offnum = PageAddItem(page, item, size,
1275 : InvalidOffsetNumber, false, false);
1276 :
1277 1155870 : if (offnum == InvalidOffsetNumber && !errorOK)
1278 0 : elog(ERROR, "failed to add item of size %zu to SPGiST index page",
1279 : size);
1280 :
1281 1155870 : return offnum;
1282 : }
1283 :
1284 : /*
1285 : * spgproperty() -- Check boolean properties of indexes.
1286 : *
1287 : * This is optional for most AMs, but is required for SP-GiST because the core
1288 : * property code doesn't support AMPROP_DISTANCE_ORDERABLE.
1289 : */
1290 : bool
1291 186 : spgproperty(Oid index_oid, int attno,
1292 : IndexAMProperty prop, const char *propname,
1293 : bool *res, bool *isnull)
1294 : {
1295 : Oid opclass,
1296 : opfamily,
1297 : opcintype;
1298 : CatCList *catlist;
1299 : int i;
1300 :
1301 : /* Only answer column-level inquiries */
1302 186 : if (attno == 0)
1303 66 : return false;
1304 :
1305 120 : switch (prop)
1306 : {
1307 12 : case AMPROP_DISTANCE_ORDERABLE:
1308 12 : break;
1309 108 : default:
1310 108 : return false;
1311 : }
1312 :
1313 : /*
1314 : * Currently, SP-GiST distance-ordered scans require that there be a
1315 : * distance operator in the opclass with the default types. So we assume
1316 : * that if such an operator exists, then there's a reason for it.
1317 : */
1318 :
1319 : /* First we need to know the column's opclass. */
1320 12 : opclass = get_index_column_opclass(index_oid, attno);
1321 12 : if (!OidIsValid(opclass))
1322 : {
1323 0 : *isnull = true;
1324 0 : return true;
1325 : }
1326 :
1327 : /* Now look up the opclass family and input datatype. */
1328 12 : if (!get_opclass_opfamily_and_input_type(opclass, &opfamily, &opcintype))
1329 : {
1330 0 : *isnull = true;
1331 0 : return true;
1332 : }
1333 :
1334 : /* And now we can check whether the operator is provided. */
1335 12 : catlist = SearchSysCacheList1(AMOPSTRATEGY,
1336 : ObjectIdGetDatum(opfamily));
1337 :
1338 12 : *res = false;
1339 :
1340 102 : for (i = 0; i < catlist->n_members; i++)
1341 : {
1342 96 : HeapTuple amoptup = &catlist->members[i]->tuple;
1343 96 : Form_pg_amop amopform = (Form_pg_amop) GETSTRUCT(amoptup);
1344 :
1345 96 : if (amopform->amoppurpose == AMOP_ORDER &&
1346 6 : (amopform->amoplefttype == opcintype ||
1347 6 : amopform->amoprighttype == opcintype) &&
1348 6 : opfamily_can_sort_type(amopform->amopsortfamily,
1349 : get_op_rettype(amopform->amopopr)))
1350 : {
1351 6 : *res = true;
1352 6 : break;
1353 : }
1354 : }
1355 :
1356 12 : ReleaseSysCacheList(catlist);
1357 :
1358 12 : *isnull = false;
1359 :
1360 12 : return true;
1361 : }
|
