Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * spgutils.c
4 : * various support functions for SP-GiST
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2025, 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 1324 : spghandler(PG_FUNCTION_ARGS)
45 : {
46 1324 : IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
47 :
48 1324 : amroutine->amstrategies = 0;
49 1324 : amroutine->amsupport = SPGISTNProc;
50 1324 : amroutine->amoptsprocnum = SPGIST_OPTIONS_PROC;
51 1324 : amroutine->amcanorder = false;
52 1324 : amroutine->amcanorderbyop = true;
53 1324 : amroutine->amcanhash = false;
54 1324 : amroutine->amconsistentequality = false;
55 1324 : amroutine->amconsistentordering = false;
56 1324 : amroutine->amcanbackward = false;
57 1324 : amroutine->amcanunique = false;
58 1324 : amroutine->amcanmulticol = false;
59 1324 : amroutine->amoptionalkey = true;
60 1324 : amroutine->amsearcharray = false;
61 1324 : amroutine->amsearchnulls = true;
62 1324 : amroutine->amstorage = true;
63 1324 : amroutine->amclusterable = false;
64 1324 : amroutine->ampredlocks = false;
65 1324 : amroutine->amcanparallel = false;
66 1324 : amroutine->amcanbuildparallel = false;
67 1324 : amroutine->amcaninclude = true;
68 1324 : amroutine->amusemaintenanceworkmem = false;
69 1324 : amroutine->amsummarizing = false;
70 1324 : amroutine->amparallelvacuumoptions =
71 : VACUUM_OPTION_PARALLEL_BULKDEL | VACUUM_OPTION_PARALLEL_COND_CLEANUP;
72 1324 : amroutine->amkeytype = InvalidOid;
73 :
74 1324 : amroutine->ambuild = spgbuild;
75 1324 : amroutine->ambuildempty = spgbuildempty;
76 1324 : amroutine->aminsert = spginsert;
77 1324 : amroutine->aminsertcleanup = NULL;
78 1324 : amroutine->ambulkdelete = spgbulkdelete;
79 1324 : amroutine->amvacuumcleanup = spgvacuumcleanup;
80 1324 : amroutine->amcanreturn = spgcanreturn;
81 1324 : amroutine->amcostestimate = spgcostestimate;
82 1324 : amroutine->amgettreeheight = NULL;
83 1324 : amroutine->amoptions = spgoptions;
84 1324 : amroutine->amproperty = spgproperty;
85 1324 : amroutine->ambuildphasename = NULL;
86 1324 : amroutine->amvalidate = spgvalidate;
87 1324 : amroutine->amadjustmembers = spgadjustmembers;
88 1324 : amroutine->ambeginscan = spgbeginscan;
89 1324 : amroutine->amrescan = spgrescan;
90 1324 : amroutine->amgettuple = spggettuple;
91 1324 : amroutine->amgetbitmap = spggetbitmap;
92 1324 : amroutine->amendscan = spgendscan;
93 1324 : amroutine->ammarkpos = NULL;
94 1324 : amroutine->amrestrpos = NULL;
95 1324 : amroutine->amestimateparallelscan = NULL;
96 1324 : amroutine->aminitparallelscan = NULL;
97 1324 : amroutine->amparallelrescan = NULL;
98 1324 : amroutine->amtranslatestrategy = NULL;
99 1324 : amroutine->amtranslatecmptype = NULL;
100 :
101 1324 : PG_RETURN_POINTER(amroutine);
102 : }
103 :
104 : /*
105 : * GetIndexInputType
106 : * Determine the nominal input data type for an index column
107 : *
108 : * We define the "nominal" input type as the associated opclass's opcintype,
109 : * or if that is a polymorphic type, the base type of the heap column or
110 : * expression that is the index's input. The reason for preferring the
111 : * opcintype is that non-polymorphic opclasses probably don't want to hear
112 : * about binary-compatible input types. For instance, if a text opclass
113 : * is being used with a varchar heap column, we want to report "text" not
114 : * "varchar". Likewise, opclasses don't want to hear about domain types,
115 : * so if we do consult the actual input type, we make sure to flatten domains.
116 : *
117 : * At some point maybe this should go somewhere else, but it's not clear
118 : * if any other index AMs have a use for it.
119 : */
120 : static Oid
121 412 : GetIndexInputType(Relation index, AttrNumber indexcol)
122 : {
123 : Oid opcintype;
124 : AttrNumber heapcol;
125 : List *indexprs;
126 : ListCell *indexpr_item;
127 :
128 : Assert(index->rd_index != NULL);
129 : Assert(indexcol > 0 && indexcol <= index->rd_index->indnkeyatts);
130 412 : opcintype = index->rd_opcintype[indexcol - 1];
131 412 : if (!IsPolymorphicType(opcintype))
132 308 : return opcintype;
133 104 : heapcol = index->rd_index->indkey.values[indexcol - 1];
134 104 : if (heapcol != 0) /* Simple index column? */
135 92 : return getBaseType(get_atttype(index->rd_index->indrelid, heapcol));
136 :
137 : /*
138 : * If the index expressions are already cached, skip calling
139 : * RelationGetIndexExpressions, as it will make a copy which is overkill.
140 : * We're not going to modify the trees, and we're not going to do anything
141 : * that would invalidate the relcache entry before we're done.
142 : */
143 12 : if (index->rd_indexprs)
144 0 : indexprs = index->rd_indexprs;
145 : else
146 12 : indexprs = RelationGetIndexExpressions(index);
147 12 : indexpr_item = list_head(indexprs);
148 12 : for (int i = 1; i <= index->rd_index->indnkeyatts; i++)
149 : {
150 12 : if (index->rd_index->indkey.values[i - 1] == 0)
151 : {
152 : /* expression column */
153 12 : if (indexpr_item == NULL)
154 0 : elog(ERROR, "wrong number of index expressions");
155 12 : if (i == indexcol)
156 12 : return getBaseType(exprType((Node *) lfirst(indexpr_item)));
157 0 : indexpr_item = lnext(indexprs, indexpr_item);
158 : }
159 : }
160 0 : elog(ERROR, "wrong number of index expressions");
161 : return InvalidOid; /* keep compiler quiet */
162 : }
163 :
164 : /* Fill in a SpGistTypeDesc struct with info about the specified data type */
165 : static void
166 1266 : fillTypeDesc(SpGistTypeDesc *desc, Oid type)
167 : {
168 : HeapTuple tp;
169 : Form_pg_type typtup;
170 :
171 1266 : desc->type = type;
172 1266 : tp = SearchSysCache1(TYPEOID, ObjectIdGetDatum(type));
173 1266 : if (!HeapTupleIsValid(tp))
174 0 : elog(ERROR, "cache lookup failed for type %u", type);
175 1266 : typtup = (Form_pg_type) GETSTRUCT(tp);
176 1266 : desc->attlen = typtup->typlen;
177 1266 : desc->attbyval = typtup->typbyval;
178 1266 : desc->attalign = typtup->typalign;
179 1266 : desc->attstorage = typtup->typstorage;
180 1266 : ReleaseSysCache(tp);
181 1266 : }
182 :
183 : /*
184 : * Fetch local cache of AM-specific info about the index, initializing it
185 : * if necessary
186 : */
187 : SpGistCache *
188 2653872 : spgGetCache(Relation index)
189 : {
190 : SpGistCache *cache;
191 :
192 2653872 : if (index->rd_amcache == NULL)
193 : {
194 : Oid atttype;
195 : spgConfigIn in;
196 : FmgrInfo *procinfo;
197 :
198 412 : cache = MemoryContextAllocZero(index->rd_indexcxt,
199 : sizeof(SpGistCache));
200 :
201 : /* SPGiST must have one key column and can also have INCLUDE columns */
202 : Assert(IndexRelationGetNumberOfKeyAttributes(index) == 1);
203 : Assert(IndexRelationGetNumberOfAttributes(index) <= INDEX_MAX_KEYS);
204 :
205 : /*
206 : * Get the actual (well, nominal) data type of the key column. We
207 : * pass this to the opclass config function so that polymorphic
208 : * opclasses are possible.
209 : */
210 412 : atttype = GetIndexInputType(index, spgKeyColumn + 1);
211 :
212 : /* Call the config function to get config info for the opclass */
213 412 : in.attType = atttype;
214 :
215 412 : procinfo = index_getprocinfo(index, 1, SPGIST_CONFIG_PROC);
216 412 : FunctionCall2Coll(procinfo,
217 412 : index->rd_indcollation[spgKeyColumn],
218 : PointerGetDatum(&in),
219 412 : PointerGetDatum(&cache->config));
220 :
221 : /*
222 : * If leafType isn't specified, use the declared index column type,
223 : * which index.c will have derived from the opclass's opcintype.
224 : * (Although we now make spgvalidate.c warn if these aren't the same,
225 : * old user-defined opclasses may not set the STORAGE parameter
226 : * correctly, so believe leafType if it's given.)
227 : */
228 412 : if (!OidIsValid(cache->config.leafType))
229 : {
230 382 : cache->config.leafType =
231 382 : TupleDescAttr(RelationGetDescr(index), spgKeyColumn)->atttypid;
232 :
233 : /*
234 : * If index column type is binary-coercible to atttype (for
235 : * example, it's a domain over atttype), treat it as plain atttype
236 : * to avoid thinking we need to compress.
237 : */
238 396 : if (cache->config.leafType != atttype &&
239 14 : IsBinaryCoercible(cache->config.leafType, atttype))
240 14 : cache->config.leafType = atttype;
241 : }
242 :
243 : /* Get the information we need about each relevant datatype */
244 412 : fillTypeDesc(&cache->attType, atttype);
245 :
246 412 : if (cache->config.leafType != atttype)
247 : {
248 30 : if (!OidIsValid(index_getprocid(index, 1, SPGIST_COMPRESS_PROC)))
249 0 : ereport(ERROR,
250 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
251 : errmsg("compress method must be defined when leaf type is different from input type")));
252 :
253 30 : fillTypeDesc(&cache->attLeafType, cache->config.leafType);
254 : }
255 : else
256 : {
257 : /* Save lookups in this common case */
258 382 : cache->attLeafType = cache->attType;
259 : }
260 :
261 412 : fillTypeDesc(&cache->attPrefixType, cache->config.prefixType);
262 412 : fillTypeDesc(&cache->attLabelType, cache->config.labelType);
263 :
264 : /*
265 : * Finally, if it's a real index (not a partitioned one), get the
266 : * lastUsedPages data from the metapage
267 : */
268 412 : if (index->rd_rel->relkind != RELKIND_PARTITIONED_INDEX)
269 : {
270 : Buffer metabuffer;
271 : SpGistMetaPageData *metadata;
272 :
273 406 : metabuffer = ReadBuffer(index, SPGIST_METAPAGE_BLKNO);
274 406 : LockBuffer(metabuffer, BUFFER_LOCK_SHARE);
275 :
276 406 : metadata = SpGistPageGetMeta(BufferGetPage(metabuffer));
277 :
278 406 : if (metadata->magicNumber != SPGIST_MAGIC_NUMBER)
279 0 : elog(ERROR, "index \"%s\" is not an SP-GiST index",
280 : RelationGetRelationName(index));
281 :
282 406 : cache->lastUsedPages = metadata->lastUsedPages;
283 :
284 406 : UnlockReleaseBuffer(metabuffer);
285 : }
286 :
287 412 : index->rd_amcache = cache;
288 : }
289 : else
290 : {
291 : /* assume it's up to date */
292 2653460 : cache = (SpGistCache *) index->rd_amcache;
293 : }
294 :
295 2653872 : return cache;
296 : }
297 :
298 : /*
299 : * Compute a tuple descriptor for leaf tuples or index-only-scan result tuples.
300 : *
301 : * We can use the relcache's tupdesc as-is in many cases, and it's always
302 : * OK so far as any INCLUDE columns are concerned. However, the entry for
303 : * the key column has to match leafType in the first case or attType in the
304 : * second case. While the relcache's tupdesc *should* show leafType, this
305 : * might not hold for legacy user-defined opclasses, since before v14 they
306 : * were not allowed to declare their true storage type in CREATE OPCLASS.
307 : * Also, attType can be different from what is in the relcache.
308 : *
309 : * This function gives back either a pointer to the relcache's tupdesc
310 : * if that is suitable, or a palloc'd copy that's been adjusted to match
311 : * the specified key column type. We can avoid doing any catalog lookups
312 : * here by insisting that the caller pass an SpGistTypeDesc not just an OID.
313 : */
314 : TupleDesc
315 244896 : getSpGistTupleDesc(Relation index, SpGistTypeDesc *keyType)
316 : {
317 : TupleDesc outTupDesc;
318 : Form_pg_attribute att;
319 :
320 489792 : if (keyType->type ==
321 244896 : TupleDescAttr(RelationGetDescr(index), spgKeyColumn)->atttypid)
322 244766 : outTupDesc = RelationGetDescr(index);
323 : else
324 : {
325 130 : outTupDesc = CreateTupleDescCopy(RelationGetDescr(index));
326 130 : att = TupleDescAttr(outTupDesc, spgKeyColumn);
327 : /* It's sufficient to update the type-dependent fields of the column */
328 130 : att->atttypid = keyType->type;
329 130 : att->atttypmod = -1;
330 130 : att->attlen = keyType->attlen;
331 130 : att->attbyval = keyType->attbyval;
332 130 : att->attalign = keyType->attalign;
333 130 : att->attstorage = keyType->attstorage;
334 : /* We shouldn't need to bother with making these valid: */
335 130 : att->attcompression = InvalidCompressionMethod;
336 130 : att->attcollation = InvalidOid;
337 : /* In case we changed typlen, we'd better reset following offsets */
338 146 : for (int i = spgFirstIncludeColumn; i < outTupDesc->natts; i++)
339 16 : TupleDescCompactAttr(outTupDesc, i)->attcacheoff = -1;
340 :
341 130 : populate_compact_attribute(outTupDesc, spgKeyColumn);
342 : }
343 244896 : return outTupDesc;
344 : }
345 :
346 : /* Initialize SpGistState for working with the given index */
347 : void
348 243992 : initSpGistState(SpGistState *state, Relation index)
349 : {
350 : SpGistCache *cache;
351 :
352 243992 : state->index = index;
353 :
354 : /* Get cached static information about index */
355 243992 : cache = spgGetCache(index);
356 :
357 243992 : state->config = cache->config;
358 243992 : state->attType = cache->attType;
359 243992 : state->attLeafType = cache->attLeafType;
360 243992 : state->attPrefixType = cache->attPrefixType;
361 243992 : state->attLabelType = cache->attLabelType;
362 :
363 : /* Ensure we have a valid descriptor for leaf tuples */
364 243992 : state->leafTupDesc = getSpGistTupleDesc(state->index, &state->attLeafType);
365 :
366 : /* Make workspace for constructing dead tuples */
367 243992 : state->deadTupleStorage = palloc0(SGDTSIZE);
368 :
369 : /*
370 : * Set horizon XID to use in redirection tuples. Use our own XID if we
371 : * have one, else use InvalidTransactionId. The latter case can happen in
372 : * VACUUM or REINDEX CONCURRENTLY, and in neither case would it be okay to
373 : * force an XID to be assigned. VACUUM won't create any redirection
374 : * tuples anyway, but REINDEX CONCURRENTLY can. Fortunately, REINDEX
375 : * CONCURRENTLY doesn't mark the index valid until the end, so there could
376 : * never be any concurrent scans "in flight" to a redirection tuple it has
377 : * inserted. And it locks out VACUUM until the end, too. So it's okay
378 : * for VACUUM to immediately expire a redirection tuple that contains an
379 : * invalid xid.
380 : */
381 243992 : state->redirectXid = GetTopTransactionIdIfAny();
382 :
383 : /* Assume we're not in an index build (spgbuild will override) */
384 243992 : state->isBuild = false;
385 243992 : }
386 :
387 : /*
388 : * Allocate a new page (either by recycling, or by extending the index file).
389 : *
390 : * The returned buffer is already pinned and exclusive-locked.
391 : * Caller is responsible for initializing the page by calling SpGistInitBuffer.
392 : */
393 : Buffer
394 6542 : SpGistNewBuffer(Relation index)
395 : {
396 : Buffer buffer;
397 :
398 : /* First, try to get a page from FSM */
399 : for (;;)
400 0 : {
401 6542 : BlockNumber blkno = GetFreeIndexPage(index);
402 :
403 6542 : if (blkno == InvalidBlockNumber)
404 6534 : break; /* nothing known to FSM */
405 :
406 : /*
407 : * The fixed pages shouldn't ever be listed in FSM, but just in case
408 : * one is, ignore it.
409 : */
410 8 : if (SpGistBlockIsFixed(blkno))
411 0 : continue;
412 :
413 8 : buffer = ReadBuffer(index, blkno);
414 :
415 : /*
416 : * We have to guard against the possibility that someone else already
417 : * recycled this page; the buffer may be locked if so.
418 : */
419 8 : if (ConditionalLockBuffer(buffer))
420 : {
421 8 : Page page = BufferGetPage(buffer);
422 :
423 8 : if (PageIsNew(page))
424 2 : return buffer; /* OK to use, if never initialized */
425 :
426 6 : if (SpGistPageIsDeleted(page) || PageIsEmpty(page))
427 6 : return buffer; /* OK to use */
428 :
429 0 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
430 : }
431 :
432 : /* Can't use it, so release buffer and try again */
433 0 : ReleaseBuffer(buffer);
434 : }
435 :
436 6534 : buffer = ExtendBufferedRel(BMR_REL(index), MAIN_FORKNUM, NULL,
437 : EB_LOCK_FIRST);
438 :
439 6534 : return buffer;
440 : }
441 :
442 : /*
443 : * Update index metapage's lastUsedPages info from local cache, if possible
444 : *
445 : * Updating meta page isn't critical for index working, so
446 : * 1 use ConditionalLockBuffer to improve concurrency
447 : * 2 don't WAL-log metabuffer changes to decrease WAL traffic
448 : */
449 : void
450 242920 : SpGistUpdateMetaPage(Relation index)
451 : {
452 242920 : SpGistCache *cache = (SpGistCache *) index->rd_amcache;
453 :
454 242920 : if (cache != NULL)
455 : {
456 : Buffer metabuffer;
457 :
458 242920 : metabuffer = ReadBuffer(index, SPGIST_METAPAGE_BLKNO);
459 :
460 242920 : if (ConditionalLockBuffer(metabuffer))
461 : {
462 242912 : Page metapage = BufferGetPage(metabuffer);
463 242912 : SpGistMetaPageData *metadata = SpGistPageGetMeta(metapage);
464 :
465 242912 : metadata->lastUsedPages = cache->lastUsedPages;
466 :
467 : /*
468 : * Set pd_lower just past the end of the metadata. This is
469 : * essential, because without doing so, metadata will be lost if
470 : * xlog.c compresses the page. (We must do this here because
471 : * pre-v11 versions of PG did not set the metapage's pd_lower
472 : * correctly, so a pg_upgraded index might contain the wrong
473 : * value.)
474 : */
475 242912 : ((PageHeader) metapage)->pd_lower =
476 242912 : ((char *) metadata + sizeof(SpGistMetaPageData)) - (char *) metapage;
477 :
478 242912 : MarkBufferDirty(metabuffer);
479 242912 : UnlockReleaseBuffer(metabuffer);
480 : }
481 : else
482 : {
483 8 : ReleaseBuffer(metabuffer);
484 : }
485 : }
486 242920 : }
487 :
488 : /* Macro to select proper element of lastUsedPages cache depending on flags */
489 : /* Masking flags with SPGIST_CACHED_PAGES is just for paranoia's sake */
490 : #define GET_LUP(c, f) (&(c)->lastUsedPages.cachedPage[((unsigned int) (f)) % SPGIST_CACHED_PAGES])
491 :
492 : /*
493 : * Allocate and initialize a new buffer of the type and parity specified by
494 : * flags. The returned buffer is already pinned and exclusive-locked.
495 : *
496 : * When requesting an inner page, if we get one with the wrong parity,
497 : * we just release the buffer and try again. We will get a different page
498 : * because GetFreeIndexPage will have marked the page used in FSM. The page
499 : * is entered in our local lastUsedPages cache, so there's some hope of
500 : * making use of it later in this session, but otherwise we rely on VACUUM
501 : * to eventually re-enter the page in FSM, making it available for recycling.
502 : * Note that such a page does not get marked dirty here, so unless it's used
503 : * fairly soon, the buffer will just get discarded and the page will remain
504 : * as it was on disk.
505 : *
506 : * When we return a buffer to the caller, the page is *not* entered into
507 : * the lastUsedPages cache; we expect the caller will do so after it's taken
508 : * whatever space it will use. This is because after the caller has used up
509 : * some space, the page might have less space than whatever was cached already
510 : * so we'd rather not trash the old cache entry.
511 : */
512 : static Buffer
513 5832 : allocNewBuffer(Relation index, int flags)
514 : {
515 5832 : SpGistCache *cache = spgGetCache(index);
516 5832 : uint16 pageflags = 0;
517 :
518 5832 : if (GBUF_REQ_LEAF(flags))
519 5726 : pageflags |= SPGIST_LEAF;
520 5832 : if (GBUF_REQ_NULLS(flags))
521 0 : pageflags |= SPGIST_NULLS;
522 :
523 : for (;;)
524 86 : {
525 : Buffer buffer;
526 :
527 5918 : buffer = SpGistNewBuffer(index);
528 5918 : SpGistInitBuffer(buffer, pageflags);
529 :
530 5918 : if (pageflags & SPGIST_LEAF)
531 : {
532 : /* Leaf pages have no parity concerns, so just use it */
533 5726 : return buffer;
534 : }
535 : else
536 : {
537 192 : BlockNumber blkno = BufferGetBlockNumber(buffer);
538 192 : int blkFlags = GBUF_INNER_PARITY(blkno);
539 :
540 192 : if ((flags & GBUF_PARITY_MASK) == blkFlags)
541 : {
542 : /* Page has right parity, use it */
543 106 : return buffer;
544 : }
545 : else
546 : {
547 : /* Page has wrong parity, record it in cache and try again */
548 86 : if (pageflags & SPGIST_NULLS)
549 0 : blkFlags |= GBUF_NULLS;
550 86 : cache->lastUsedPages.cachedPage[blkFlags].blkno = blkno;
551 86 : cache->lastUsedPages.cachedPage[blkFlags].freeSpace =
552 86 : PageGetExactFreeSpace(BufferGetPage(buffer));
553 86 : UnlockReleaseBuffer(buffer);
554 : }
555 : }
556 : }
557 : }
558 :
559 : /*
560 : * Get a buffer of the type and parity specified by flags, having at least
561 : * as much free space as indicated by needSpace. We use the lastUsedPages
562 : * cache to assign the same buffer previously requested when possible.
563 : * The returned buffer is already pinned and exclusive-locked.
564 : *
565 : * *isNew is set true if the page was initialized here, false if it was
566 : * already valid.
567 : */
568 : Buffer
569 10906 : SpGistGetBuffer(Relation index, int flags, int needSpace, bool *isNew)
570 : {
571 10906 : SpGistCache *cache = spgGetCache(index);
572 : SpGistLastUsedPage *lup;
573 :
574 : /* Bail out if even an empty page wouldn't meet the demand */
575 10906 : if (needSpace > SPGIST_PAGE_CAPACITY)
576 0 : elog(ERROR, "desired SPGiST tuple size is too big");
577 :
578 : /*
579 : * If possible, increase the space request to include relation's
580 : * fillfactor. This ensures that when we add unrelated tuples to a page,
581 : * we try to keep 100-fillfactor% available for adding tuples that are
582 : * related to the ones already on it. But fillfactor mustn't cause an
583 : * error for requests that would otherwise be legal.
584 : */
585 10906 : needSpace += SpGistGetTargetPageFreeSpace(index);
586 10906 : needSpace = Min(needSpace, SPGIST_PAGE_CAPACITY);
587 :
588 : /* Get the cache entry for this flags setting */
589 10906 : lup = GET_LUP(cache, flags);
590 :
591 : /* If we have nothing cached, just turn it over to allocNewBuffer */
592 10906 : if (lup->blkno == InvalidBlockNumber)
593 : {
594 182 : *isNew = true;
595 182 : return allocNewBuffer(index, flags);
596 : }
597 :
598 : /* fixed pages should never be in cache */
599 : Assert(!SpGistBlockIsFixed(lup->blkno));
600 :
601 : /* If cached freeSpace isn't enough, don't bother looking at the page */
602 10724 : if (lup->freeSpace >= needSpace)
603 : {
604 : Buffer buffer;
605 : Page page;
606 :
607 5076 : buffer = ReadBuffer(index, lup->blkno);
608 :
609 5076 : if (!ConditionalLockBuffer(buffer))
610 : {
611 : /*
612 : * buffer is locked by another process, so return a new buffer
613 : */
614 2 : ReleaseBuffer(buffer);
615 2 : *isNew = true;
616 2 : return allocNewBuffer(index, flags);
617 : }
618 :
619 5074 : page = BufferGetPage(buffer);
620 :
621 5074 : if (PageIsNew(page) || SpGistPageIsDeleted(page) || PageIsEmpty(page))
622 : {
623 : /* OK to initialize the page */
624 188 : uint16 pageflags = 0;
625 :
626 188 : if (GBUF_REQ_LEAF(flags))
627 182 : pageflags |= SPGIST_LEAF;
628 188 : if (GBUF_REQ_NULLS(flags))
629 0 : pageflags |= SPGIST_NULLS;
630 188 : SpGistInitBuffer(buffer, pageflags);
631 188 : lup->freeSpace = PageGetExactFreeSpace(page) - needSpace;
632 188 : *isNew = true;
633 188 : return buffer;
634 : }
635 :
636 : /*
637 : * Check that page is of right type and has enough space. We must
638 : * recheck this since our cache isn't necessarily up to date.
639 : */
640 9772 : if ((GBUF_REQ_LEAF(flags) ? SpGistPageIsLeaf(page) : !SpGistPageIsLeaf(page)) &&
641 4886 : (GBUF_REQ_NULLS(flags) ? SpGistPageStoresNulls(page) : !SpGistPageStoresNulls(page)))
642 : {
643 4886 : int freeSpace = PageGetExactFreeSpace(page);
644 :
645 4886 : if (freeSpace >= needSpace)
646 : {
647 : /* Success, update freespace info and return the buffer */
648 4886 : lup->freeSpace = freeSpace - needSpace;
649 4886 : *isNew = false;
650 4886 : return buffer;
651 : }
652 : }
653 :
654 : /*
655 : * fallback to allocation of new buffer
656 : */
657 0 : UnlockReleaseBuffer(buffer);
658 : }
659 :
660 : /* No success with cache, so return a new buffer */
661 5648 : *isNew = true;
662 5648 : return allocNewBuffer(index, flags);
663 : }
664 :
665 : /*
666 : * Update lastUsedPages cache when done modifying a page.
667 : *
668 : * We update the appropriate cache entry if it already contained this page
669 : * (its freeSpace is likely obsolete), or if this page has more space than
670 : * whatever we had cached.
671 : */
672 : void
673 2391316 : SpGistSetLastUsedPage(Relation index, Buffer buffer)
674 : {
675 2391316 : SpGistCache *cache = spgGetCache(index);
676 : SpGistLastUsedPage *lup;
677 : int freeSpace;
678 2391316 : Page page = BufferGetPage(buffer);
679 2391316 : BlockNumber blkno = BufferGetBlockNumber(buffer);
680 : int flags;
681 :
682 : /* Never enter fixed pages (root pages) in cache, though */
683 2391316 : if (SpGistBlockIsFixed(blkno))
684 805560 : return;
685 :
686 1585756 : if (SpGistPageIsLeaf(page))
687 800854 : flags = GBUF_LEAF;
688 : else
689 784902 : flags = GBUF_INNER_PARITY(blkno);
690 1585756 : if (SpGistPageStoresNulls(page))
691 0 : flags |= GBUF_NULLS;
692 :
693 1585756 : lup = GET_LUP(cache, flags);
694 :
695 1585756 : freeSpace = PageGetExactFreeSpace(page);
696 1585756 : if (lup->blkno == InvalidBlockNumber || lup->blkno == blkno ||
697 420076 : lup->freeSpace < freeSpace)
698 : {
699 1174328 : lup->blkno = blkno;
700 1174328 : lup->freeSpace = freeSpace;
701 : }
702 : }
703 :
704 : /*
705 : * Initialize an SPGiST page to empty, with specified flags
706 : */
707 : void
708 7626 : SpGistInitPage(Page page, uint16 f)
709 : {
710 : SpGistPageOpaque opaque;
711 :
712 7626 : PageInit(page, BLCKSZ, sizeof(SpGistPageOpaqueData));
713 7626 : opaque = SpGistPageGetOpaque(page);
714 7626 : opaque->flags = f;
715 7626 : opaque->spgist_page_id = SPGIST_PAGE_ID;
716 7626 : }
717 :
718 : /*
719 : * Initialize a buffer's page to empty, with specified flags
720 : */
721 : void
722 7394 : SpGistInitBuffer(Buffer b, uint16 f)
723 : {
724 : Assert(BufferGetPageSize(b) == BLCKSZ);
725 7394 : SpGistInitPage(BufferGetPage(b), f);
726 7394 : }
727 :
728 : /*
729 : * Initialize metadata page
730 : */
731 : void
732 216 : SpGistInitMetapage(Page page)
733 : {
734 : SpGistMetaPageData *metadata;
735 : int i;
736 :
737 216 : SpGistInitPage(page, SPGIST_META);
738 216 : metadata = SpGistPageGetMeta(page);
739 216 : memset(metadata, 0, sizeof(SpGistMetaPageData));
740 216 : metadata->magicNumber = SPGIST_MAGIC_NUMBER;
741 :
742 : /* initialize last-used-page cache to empty */
743 1944 : for (i = 0; i < SPGIST_CACHED_PAGES; i++)
744 1728 : metadata->lastUsedPages.cachedPage[i].blkno = InvalidBlockNumber;
745 :
746 : /*
747 : * Set pd_lower just past the end of the metadata. This is essential,
748 : * because without doing so, metadata will be lost if xlog.c compresses
749 : * the page.
750 : */
751 216 : ((PageHeader) page)->pd_lower =
752 216 : ((char *) metadata + sizeof(SpGistMetaPageData)) - (char *) page;
753 216 : }
754 :
755 : /*
756 : * reloptions processing for SPGiST
757 : */
758 : bytea *
759 128 : spgoptions(Datum reloptions, bool validate)
760 : {
761 : static const relopt_parse_elt tab[] = {
762 : {"fillfactor", RELOPT_TYPE_INT, offsetof(SpGistOptions, fillfactor)},
763 : };
764 :
765 128 : return (bytea *) build_reloptions(reloptions, validate,
766 : RELOPT_KIND_SPGIST,
767 : sizeof(SpGistOptions),
768 : tab, lengthof(tab));
769 : }
770 :
771 : /*
772 : * Get the space needed to store a non-null datum of the indicated type
773 : * in an inner tuple (that is, as a prefix or node label).
774 : * Note the result is already rounded up to a MAXALIGN boundary.
775 : * Here we follow the convention that pass-by-val types are just stored
776 : * in their Datum representation (compare memcpyInnerDatum).
777 : */
778 : unsigned int
779 12330 : SpGistGetInnerTypeSize(SpGistTypeDesc *att, Datum datum)
780 : {
781 : unsigned int size;
782 :
783 12330 : if (att->attbyval)
784 6484 : size = sizeof(Datum);
785 5846 : else if (att->attlen > 0)
786 3962 : size = att->attlen;
787 : else
788 1884 : size = VARSIZE_ANY(datum);
789 :
790 12330 : return MAXALIGN(size);
791 : }
792 :
793 : /*
794 : * Copy the given non-null datum to *target, in the inner-tuple case
795 : */
796 : static void
797 12330 : memcpyInnerDatum(void *target, SpGistTypeDesc *att, Datum datum)
798 : {
799 : unsigned int size;
800 :
801 12330 : if (att->attbyval)
802 : {
803 6484 : memcpy(target, &datum, sizeof(Datum));
804 : }
805 : else
806 : {
807 5846 : size = (att->attlen > 0) ? att->attlen : VARSIZE_ANY(datum);
808 5846 : memcpy(target, DatumGetPointer(datum), size);
809 : }
810 12330 : }
811 :
812 : /*
813 : * Compute space required for a leaf tuple holding the given data.
814 : *
815 : * This must match the size-calculation portion of spgFormLeafTuple.
816 : */
817 : Size
818 19455914 : SpGistGetLeafTupleSize(TupleDesc tupleDescriptor,
819 : const Datum *datums, const bool *isnulls)
820 : {
821 : Size size;
822 : Size data_size;
823 19455914 : bool needs_null_mask = false;
824 19455914 : int natts = tupleDescriptor->natts;
825 :
826 : /*
827 : * Decide whether we need a nulls bitmask.
828 : *
829 : * If there is only a key attribute (natts == 1), never use a bitmask, for
830 : * compatibility with the pre-v14 layout of leaf tuples. Otherwise, we
831 : * need one if any attribute is null.
832 : */
833 19455914 : if (natts > 1)
834 : {
835 993914 : for (int i = 0; i < natts; i++)
836 : {
837 680744 : if (isnulls[i])
838 : {
839 18318 : needs_null_mask = true;
840 18318 : break;
841 : }
842 : }
843 : }
844 :
845 : /*
846 : * Calculate size of the data part; same as for heap tuples.
847 : */
848 19455914 : data_size = heap_compute_data_size(tupleDescriptor, datums, isnulls);
849 :
850 : /*
851 : * Compute total size.
852 : */
853 19455914 : size = SGLTHDRSZ(needs_null_mask);
854 19455914 : size += data_size;
855 19455914 : size = MAXALIGN(size);
856 :
857 : /*
858 : * Ensure that we can replace the tuple with a dead tuple later. This test
859 : * is unnecessary when there are any non-null attributes, but be safe.
860 : */
861 19455914 : if (size < SGDTSIZE)
862 0 : size = SGDTSIZE;
863 :
864 19455914 : return size;
865 : }
866 :
867 : /*
868 : * Construct a leaf tuple containing the given heap TID and datum values
869 : */
870 : SpGistLeafTuple
871 1527250 : spgFormLeafTuple(SpGistState *state, ItemPointer heapPtr,
872 : const Datum *datums, const bool *isnulls)
873 : {
874 : SpGistLeafTuple tup;
875 1527250 : TupleDesc tupleDescriptor = state->leafTupDesc;
876 : Size size;
877 : Size hoff;
878 : Size data_size;
879 1527250 : bool needs_null_mask = false;
880 1527250 : int natts = tupleDescriptor->natts;
881 : char *tp; /* ptr to tuple data */
882 1527250 : uint16 tupmask = 0; /* unused heap_fill_tuple output */
883 :
884 : /*
885 : * Decide whether we need a nulls bitmask.
886 : *
887 : * If there is only a key attribute (natts == 1), never use a bitmask, for
888 : * compatibility with the pre-v14 layout of leaf tuples. Otherwise, we
889 : * need one if any attribute is null.
890 : */
891 1527250 : if (natts > 1)
892 : {
893 420754 : for (int i = 0; i < natts; i++)
894 : {
895 292402 : if (isnulls[i])
896 : {
897 12064 : needs_null_mask = true;
898 12064 : break;
899 : }
900 : }
901 : }
902 :
903 : /*
904 : * Calculate size of the data part; same as for heap tuples.
905 : */
906 1527250 : data_size = heap_compute_data_size(tupleDescriptor, datums, isnulls);
907 :
908 : /*
909 : * Compute total size.
910 : */
911 1527250 : hoff = SGLTHDRSZ(needs_null_mask);
912 1527250 : size = hoff + data_size;
913 1527250 : size = MAXALIGN(size);
914 :
915 : /*
916 : * Ensure that we can replace the tuple with a dead tuple later. This test
917 : * is unnecessary when there are any non-null attributes, but be safe.
918 : */
919 1527250 : if (size < SGDTSIZE)
920 0 : size = SGDTSIZE;
921 :
922 : /* OK, form the tuple */
923 1527250 : tup = (SpGistLeafTuple) palloc0(size);
924 :
925 1527250 : tup->size = size;
926 1527250 : SGLT_SET_NEXTOFFSET(tup, InvalidOffsetNumber);
927 1527250 : tup->heapPtr = *heapPtr;
928 :
929 1527250 : tp = (char *) tup + hoff;
930 :
931 1527250 : if (needs_null_mask)
932 : {
933 : bits8 *bp; /* ptr to null bitmap in tuple */
934 :
935 : /* Set nullmask presence bit in SpGistLeafTuple header */
936 12064 : SGLT_SET_HASNULLMASK(tup, true);
937 : /* Fill the data area and null mask */
938 12064 : bp = (bits8 *) ((char *) tup + sizeof(SpGistLeafTupleData));
939 12064 : heap_fill_tuple(tupleDescriptor, datums, isnulls, tp, data_size,
940 : &tupmask, bp);
941 : }
942 1515186 : else if (natts > 1 || !isnulls[spgKeyColumn])
943 : {
944 : /* Fill data area only */
945 1515114 : heap_fill_tuple(tupleDescriptor, datums, isnulls, tp, data_size,
946 : &tupmask, (bits8 *) NULL);
947 : }
948 : /* otherwise we have no data, nor a bitmap, to fill */
949 :
950 1527250 : return tup;
951 : }
952 :
953 : /*
954 : * Construct a node (to go into an inner tuple) containing the given label
955 : *
956 : * Note that the node's downlink is just set invalid here. Caller will fill
957 : * it in later.
958 : */
959 : SpGistNodeTuple
960 40554 : spgFormNodeTuple(SpGistState *state, Datum label, bool isnull)
961 : {
962 : SpGistNodeTuple tup;
963 : unsigned int size;
964 40554 : unsigned short infomask = 0;
965 :
966 : /* compute space needed (note result is already maxaligned) */
967 40554 : size = SGNTHDRSZ;
968 40554 : if (!isnull)
969 5794 : size += SpGistGetInnerTypeSize(&state->attLabelType, label);
970 :
971 : /*
972 : * Here we make sure that the size will fit in the field reserved for it
973 : * in t_info.
974 : */
975 40554 : if ((size & INDEX_SIZE_MASK) != size)
976 0 : ereport(ERROR,
977 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
978 : errmsg("index row requires %zu bytes, maximum size is %zu",
979 : (Size) size, (Size) INDEX_SIZE_MASK)));
980 :
981 40554 : tup = (SpGistNodeTuple) palloc0(size);
982 :
983 40554 : if (isnull)
984 34760 : infomask |= INDEX_NULL_MASK;
985 : /* we don't bother setting the INDEX_VAR_MASK bit */
986 40554 : infomask |= size;
987 40554 : tup->t_info = infomask;
988 :
989 : /* The TID field will be filled in later */
990 40554 : ItemPointerSetInvalid(&tup->t_tid);
991 :
992 40554 : if (!isnull)
993 5794 : memcpyInnerDatum(SGNTDATAPTR(tup), &state->attLabelType, label);
994 :
995 40554 : return tup;
996 : }
997 :
998 : /*
999 : * Construct an inner tuple containing the given prefix and node array
1000 : */
1001 : SpGistInnerTuple
1002 8530 : spgFormInnerTuple(SpGistState *state, bool hasPrefix, Datum prefix,
1003 : int nNodes, SpGistNodeTuple *nodes)
1004 : {
1005 : SpGistInnerTuple tup;
1006 : unsigned int size;
1007 : unsigned int prefixSize;
1008 : int i;
1009 : char *ptr;
1010 :
1011 : /* Compute size needed */
1012 8530 : if (hasPrefix)
1013 6536 : prefixSize = SpGistGetInnerTypeSize(&state->attPrefixType, prefix);
1014 : else
1015 1994 : prefixSize = 0;
1016 :
1017 8530 : size = SGITHDRSZ + prefixSize;
1018 :
1019 : /* Note: we rely on node tuple sizes to be maxaligned already */
1020 59374 : for (i = 0; i < nNodes; i++)
1021 50844 : size += IndexTupleSize(nodes[i]);
1022 :
1023 : /*
1024 : * Ensure that we can replace the tuple with a dead tuple later. This
1025 : * test is unnecessary given current tuple layouts, but let's be safe.
1026 : */
1027 8530 : if (size < SGDTSIZE)
1028 0 : size = SGDTSIZE;
1029 :
1030 : /*
1031 : * Inner tuple should be small enough to fit on a page
1032 : */
1033 8530 : if (size > SPGIST_PAGE_CAPACITY - sizeof(ItemIdData))
1034 0 : ereport(ERROR,
1035 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1036 : errmsg("SP-GiST inner tuple size %zu exceeds maximum %zu",
1037 : (Size) size,
1038 : SPGIST_PAGE_CAPACITY - sizeof(ItemIdData)),
1039 : errhint("Values larger than a buffer page cannot be indexed.")));
1040 :
1041 : /*
1042 : * Check for overflow of header fields --- probably can't fail if the
1043 : * above succeeded, but let's be paranoid
1044 : */
1045 8530 : if (size > SGITMAXSIZE ||
1046 8530 : prefixSize > SGITMAXPREFIXSIZE ||
1047 : nNodes > SGITMAXNNODES)
1048 0 : elog(ERROR, "SPGiST inner tuple header field is too small");
1049 :
1050 : /* OK, form the tuple */
1051 8530 : tup = (SpGistInnerTuple) palloc0(size);
1052 :
1053 8530 : tup->nNodes = nNodes;
1054 8530 : tup->prefixSize = prefixSize;
1055 8530 : tup->size = size;
1056 :
1057 8530 : if (hasPrefix)
1058 6536 : memcpyInnerDatum(SGITDATAPTR(tup), &state->attPrefixType, prefix);
1059 :
1060 8530 : ptr = (char *) SGITNODEPTR(tup);
1061 :
1062 59374 : for (i = 0; i < nNodes; i++)
1063 : {
1064 50844 : SpGistNodeTuple node = nodes[i];
1065 :
1066 50844 : memcpy(ptr, node, IndexTupleSize(node));
1067 50844 : ptr += IndexTupleSize(node);
1068 : }
1069 :
1070 8530 : return tup;
1071 : }
1072 :
1073 : /*
1074 : * Construct a "dead" tuple to replace a tuple being deleted.
1075 : *
1076 : * The state can be SPGIST_REDIRECT, SPGIST_DEAD, or SPGIST_PLACEHOLDER.
1077 : * For a REDIRECT tuple, a pointer (blkno+offset) must be supplied, and
1078 : * the xid field is filled in automatically.
1079 : *
1080 : * This is called in critical sections, so we don't use palloc; the tuple
1081 : * is built in preallocated storage. It should be copied before another
1082 : * call with different parameters can occur.
1083 : */
1084 : SpGistDeadTuple
1085 13540 : spgFormDeadTuple(SpGistState *state, int tupstate,
1086 : BlockNumber blkno, OffsetNumber offnum)
1087 : {
1088 13540 : SpGistDeadTuple tuple = (SpGistDeadTuple) state->deadTupleStorage;
1089 :
1090 13540 : tuple->tupstate = tupstate;
1091 13540 : tuple->size = SGDTSIZE;
1092 13540 : SGLT_SET_NEXTOFFSET(tuple, InvalidOffsetNumber);
1093 :
1094 13540 : if (tupstate == SPGIST_REDIRECT)
1095 : {
1096 2384 : ItemPointerSet(&tuple->pointer, blkno, offnum);
1097 2384 : tuple->xid = state->redirectXid;
1098 : }
1099 : else
1100 : {
1101 11156 : ItemPointerSetInvalid(&tuple->pointer);
1102 11156 : tuple->xid = InvalidTransactionId;
1103 : }
1104 :
1105 13540 : return tuple;
1106 : }
1107 :
1108 : /*
1109 : * Convert an SPGiST leaf tuple into Datum/isnull arrays.
1110 : *
1111 : * The caller must allocate sufficient storage for the output arrays.
1112 : * (INDEX_MAX_KEYS entries should be enough.)
1113 : */
1114 : void
1115 60848 : spgDeformLeafTuple(SpGistLeafTuple tup, TupleDesc tupleDescriptor,
1116 : Datum *datums, bool *isnulls, bool keyColumnIsNull)
1117 : {
1118 60848 : bool hasNullsMask = SGLT_GET_HASNULLMASK(tup);
1119 : char *tp; /* ptr to tuple data */
1120 : bits8 *bp; /* ptr to null bitmap in tuple */
1121 :
1122 60848 : if (keyColumnIsNull && tupleDescriptor->natts == 1)
1123 : {
1124 : /*
1125 : * Trivial case: there is only the key attribute and we're in a nulls
1126 : * tree. The hasNullsMask bit in the tuple header should not be set
1127 : * (and thus we can't use index_deform_tuple_internal), but
1128 : * nonetheless the result is NULL.
1129 : *
1130 : * Note: currently this is dead code, because noplace calls this when
1131 : * there is only the key attribute. But we should cover the case.
1132 : */
1133 : Assert(!hasNullsMask);
1134 :
1135 0 : datums[spgKeyColumn] = (Datum) 0;
1136 0 : isnulls[spgKeyColumn] = true;
1137 0 : return;
1138 : }
1139 :
1140 60848 : tp = (char *) tup + SGLTHDRSZ(hasNullsMask);
1141 60848 : bp = (bits8 *) ((char *) tup + sizeof(SpGistLeafTupleData));
1142 :
1143 60848 : index_deform_tuple_internal(tupleDescriptor,
1144 : datums, isnulls,
1145 : tp, bp, hasNullsMask);
1146 :
1147 : /*
1148 : * Key column isnull value from the tuple should be consistent with
1149 : * keyColumnIsNull flag from the caller.
1150 : */
1151 : Assert(keyColumnIsNull == isnulls[spgKeyColumn]);
1152 : }
1153 :
1154 : /*
1155 : * Extract the label datums of the nodes within innerTuple
1156 : *
1157 : * Returns NULL if label datums are NULLs
1158 : */
1159 : Datum *
1160 18676896 : spgExtractNodeLabels(SpGistState *state, SpGistInnerTuple innerTuple)
1161 : {
1162 : Datum *nodeLabels;
1163 : int i;
1164 : SpGistNodeTuple node;
1165 :
1166 : /* Either all the labels must be NULL, or none. */
1167 18676896 : node = SGITNODEPTR(innerTuple);
1168 18676896 : if (IndexTupleHasNulls(node))
1169 : {
1170 100677996 : SGITITERATE(innerTuple, i, node)
1171 : {
1172 82234966 : if (!IndexTupleHasNulls(node))
1173 0 : elog(ERROR, "some but not all node labels are null in SPGiST inner tuple");
1174 : }
1175 : /* They're all null, so just return NULL */
1176 18443030 : return NULL;
1177 : }
1178 : else
1179 : {
1180 233866 : nodeLabels = (Datum *) palloc(sizeof(Datum) * innerTuple->nNodes);
1181 2683416 : SGITITERATE(innerTuple, i, node)
1182 : {
1183 2449550 : if (IndexTupleHasNulls(node))
1184 0 : elog(ERROR, "some but not all node labels are null in SPGiST inner tuple");
1185 2449550 : nodeLabels[i] = SGNTDATUM(node, state);
1186 : }
1187 233866 : return nodeLabels;
1188 : }
1189 : }
1190 :
1191 : /*
1192 : * Add a new item to the page, replacing a PLACEHOLDER item if possible.
1193 : * Return the location it's inserted at, or InvalidOffsetNumber on failure.
1194 : *
1195 : * If startOffset isn't NULL, we start searching for placeholders at
1196 : * *startOffset, and update that to the next place to search. This is just
1197 : * an optimization for repeated insertions.
1198 : *
1199 : * If errorOK is false, we throw error when there's not enough room,
1200 : * rather than returning InvalidOffsetNumber.
1201 : */
1202 : OffsetNumber
1203 1618528 : SpGistPageAddNewItem(SpGistState *state, Page page, Item item, Size size,
1204 : OffsetNumber *startOffset, bool errorOK)
1205 : {
1206 1618528 : SpGistPageOpaque opaque = SpGistPageGetOpaque(page);
1207 : OffsetNumber i,
1208 : maxoff,
1209 : offnum;
1210 :
1211 1618528 : if (opaque->nPlaceholder > 0 &&
1212 461796 : PageGetExactFreeSpace(page) + SGDTSIZE >= MAXALIGN(size))
1213 : {
1214 : /* Try to replace a placeholder */
1215 461796 : maxoff = PageGetMaxOffsetNumber(page);
1216 461796 : offnum = InvalidOffsetNumber;
1217 :
1218 : for (;;)
1219 : {
1220 461796 : if (startOffset && *startOffset != InvalidOffsetNumber)
1221 114926 : i = *startOffset;
1222 : else
1223 346870 : i = FirstOffsetNumber;
1224 31587354 : for (; i <= maxoff; i++)
1225 : {
1226 31587354 : SpGistDeadTuple it = (SpGistDeadTuple) PageGetItem(page,
1227 31587354 : PageGetItemId(page, i));
1228 :
1229 31587354 : if (it->tupstate == SPGIST_PLACEHOLDER)
1230 : {
1231 461796 : offnum = i;
1232 461796 : break;
1233 : }
1234 : }
1235 :
1236 : /* Done if we found a placeholder */
1237 461796 : if (offnum != InvalidOffsetNumber)
1238 461796 : break;
1239 :
1240 0 : if (startOffset && *startOffset != InvalidOffsetNumber)
1241 : {
1242 : /* Hint was no good, re-search from beginning */
1243 0 : *startOffset = InvalidOffsetNumber;
1244 0 : continue;
1245 : }
1246 :
1247 : /* Hmm, no placeholder found? */
1248 0 : opaque->nPlaceholder = 0;
1249 0 : break;
1250 : }
1251 :
1252 461796 : if (offnum != InvalidOffsetNumber)
1253 : {
1254 : /* Replace the placeholder tuple */
1255 461796 : PageIndexTupleDelete(page, offnum);
1256 :
1257 461796 : offnum = PageAddItem(page, item, size, offnum, false, false);
1258 :
1259 : /*
1260 : * We should not have failed given the size check at the top of
1261 : * the function, but test anyway. If we did fail, we must PANIC
1262 : * because we've already deleted the placeholder tuple, and
1263 : * there's no other way to keep the damage from getting to disk.
1264 : */
1265 461796 : if (offnum != InvalidOffsetNumber)
1266 : {
1267 : Assert(opaque->nPlaceholder > 0);
1268 461796 : opaque->nPlaceholder--;
1269 461796 : if (startOffset)
1270 117560 : *startOffset = offnum + 1;
1271 : }
1272 : else
1273 0 : elog(PANIC, "failed to add item of size %zu to SPGiST index page",
1274 : size);
1275 :
1276 461796 : return offnum;
1277 : }
1278 : }
1279 :
1280 : /* No luck in replacing a placeholder, so just add it to the page */
1281 1156732 : offnum = PageAddItem(page, item, size,
1282 : InvalidOffsetNumber, false, false);
1283 :
1284 1156732 : if (offnum == InvalidOffsetNumber && !errorOK)
1285 0 : elog(ERROR, "failed to add item of size %zu to SPGiST index page",
1286 : size);
1287 :
1288 1156732 : return offnum;
1289 : }
1290 :
1291 : /*
1292 : * spgproperty() -- Check boolean properties of indexes.
1293 : *
1294 : * This is optional for most AMs, but is required for SP-GiST because the core
1295 : * property code doesn't support AMPROP_DISTANCE_ORDERABLE.
1296 : */
1297 : bool
1298 186 : spgproperty(Oid index_oid, int attno,
1299 : IndexAMProperty prop, const char *propname,
1300 : bool *res, bool *isnull)
1301 : {
1302 : Oid opclass,
1303 : opfamily,
1304 : opcintype;
1305 : CatCList *catlist;
1306 : int i;
1307 :
1308 : /* Only answer column-level inquiries */
1309 186 : if (attno == 0)
1310 66 : return false;
1311 :
1312 120 : switch (prop)
1313 : {
1314 12 : case AMPROP_DISTANCE_ORDERABLE:
1315 12 : break;
1316 108 : default:
1317 108 : return false;
1318 : }
1319 :
1320 : /*
1321 : * Currently, SP-GiST distance-ordered scans require that there be a
1322 : * distance operator in the opclass with the default types. So we assume
1323 : * that if such an operator exists, then there's a reason for it.
1324 : */
1325 :
1326 : /* First we need to know the column's opclass. */
1327 12 : opclass = get_index_column_opclass(index_oid, attno);
1328 12 : if (!OidIsValid(opclass))
1329 : {
1330 0 : *isnull = true;
1331 0 : return true;
1332 : }
1333 :
1334 : /* Now look up the opclass family and input datatype. */
1335 12 : if (!get_opclass_opfamily_and_input_type(opclass, &opfamily, &opcintype))
1336 : {
1337 0 : *isnull = true;
1338 0 : return true;
1339 : }
1340 :
1341 : /* And now we can check whether the operator is provided. */
1342 12 : catlist = SearchSysCacheList1(AMOPSTRATEGY,
1343 : ObjectIdGetDatum(opfamily));
1344 :
1345 12 : *res = false;
1346 :
1347 102 : for (i = 0; i < catlist->n_members; i++)
1348 : {
1349 96 : HeapTuple amoptup = &catlist->members[i]->tuple;
1350 96 : Form_pg_amop amopform = (Form_pg_amop) GETSTRUCT(amoptup);
1351 :
1352 96 : if (amopform->amoppurpose == AMOP_ORDER &&
1353 6 : (amopform->amoplefttype == opcintype ||
1354 6 : amopform->amoprighttype == opcintype) &&
1355 6 : opfamily_can_sort_type(amopform->amopsortfamily,
1356 : get_op_rettype(amopform->amopopr)))
1357 : {
1358 6 : *res = true;
1359 6 : break;
1360 : }
1361 : }
1362 :
1363 12 : ReleaseSysCacheList(catlist);
1364 :
1365 12 : *isnull = false;
1366 :
1367 12 : return true;
1368 : }
|
