| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* ginget.c |
| 4 |
* fetch tuples from a GIN scan. |
| 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/gin/ginget.c |
| 12 |
*------------------------------------------------------------------------- |
| 13 |
*/ |
| 14 |
|
| 15 |
#include "postgres.h" |
| 16 |
|
| 17 |
#include "access/gin_private.h" |
| 18 |
#include "access/relscan.h" |
| 19 |
#include "common/pg_prng.h" |
| 20 |
#include "miscadmin.h" |
| 21 |
#include "storage/predicate.h" |
| 22 |
#include "utils/datum.h" |
| 23 |
#include "utils/memutils.h" |
| 24 |
#include "utils/rel.h" |
| 25 |
|
| 26 |
/* GUC parameter */ |
| 27 |
int GinFuzzySearchLimit = 0; |
| 28 |
|
| 29 |
typedef struct pendingPosition |
| 30 |
{ |
| 31 |
Buffer pendingBuffer; |
| 32 |
OffsetNumber firstOffset; |
| 33 |
OffsetNumber lastOffset; |
| 34 |
ItemPointerData item; |
| 35 |
bool *hasMatchKey; |
| 36 |
} pendingPosition; |
| 37 |
|
| 38 |
|
| 39 |
/* |
| 40 |
* Goes to the next page if current offset is outside of bounds |
| 41 |
*/ |
| 42 |
static bool |
| 43 |
moveRightIfItNeeded(GinBtreeData *btree, GinBtreeStack *stack, Snapshot snapshot) |
| 44 |
{ |
| 45 |
Page page = BufferGetPage(stack->buffer); |
| 46 |
|
| 47 |
if (stack->off > PageGetMaxOffsetNumber(page)) |
| 48 |
{ |
| 49 |
/* |
| 50 |
* We scanned the whole page, so we should take right page |
| 51 |
*/ |
| 52 |
if (GinPageRightMost(page)) |
| 53 |
return false; /* no more pages */ |
| 54 |
|
| 55 |
stack->buffer = ginStepRight(stack->buffer, btree->index, GIN_SHARE); |
| 56 |
stack->blkno = BufferGetBlockNumber(stack->buffer); |
| 57 |
stack->off = FirstOffsetNumber; |
| 58 |
PredicateLockPage(btree->index, stack->blkno, snapshot); |
| 59 |
} |
| 60 |
|
| 61 |
return true; |
| 62 |
} |
| 63 |
|
| 64 |
/* |
| 65 |
* Scan all pages of a posting tree and save all its heap ItemPointers |
| 66 |
* in scanEntry->matchBitmap |
| 67 |
*/ |
| 68 |
static void |
| 69 |
scanPostingTree(Relation index, GinScanEntry scanEntry, |
| 70 |
BlockNumber rootPostingTree) |
| 71 |
{ |
| 72 |
GinBtreeData btree; |
| 73 |
GinBtreeStack *stack; |
| 74 |
Buffer buffer; |
| 75 |
Page page; |
| 76 |
|
| 77 |
/* Descend to the leftmost leaf page */ |
| 78 |
stack = ginScanBeginPostingTree(&btree, index, rootPostingTree); |
| 79 |
buffer = stack->buffer; |
| 80 |
|
| 81 |
IncrBufferRefCount(buffer); /* prevent unpin in freeGinBtreeStack */ |
| 82 |
|
| 83 |
freeGinBtreeStack(stack); |
| 84 |
|
| 85 |
/* |
| 86 |
* Loop iterates through all leaf pages of posting tree |
| 87 |
*/ |
| 88 |
for (;;) |
| 89 |
{ |
| 90 |
page = BufferGetPage(buffer); |
| 91 |
if ((GinPageGetOpaque(page)->flags & GIN_DELETED) == 0) |
| 92 |
{ |
| 93 |
int n = GinDataLeafPageGetItemsToTbm(page, scanEntry->matchBitmap); |
| 94 |
|
| 95 |
scanEntry->predictNumberResult += n; |
| 96 |
} |
| 97 |
|
| 98 |
if (GinPageRightMost(page)) |
| 99 |
break; /* no more pages */ |
| 100 |
|
| 101 |
buffer = ginStepRight(buffer, index, GIN_SHARE); |
| 102 |
} |
| 103 |
|
| 104 |
UnlockReleaseBuffer(buffer); |
| 105 |
} |
| 106 |
|
| 107 |
/* |
| 108 |
* Collects TIDs into scanEntry->matchBitmap for all heap tuples that |
| 109 |
* match the search entry. This supports three different match modes: |
| 110 |
* |
| 111 |
* 1. Partial-match support: scan from current point until the |
| 112 |
* comparePartialFn says we're done. |
| 113 |
* 2. SEARCH_MODE_ALL: scan from current point (which should be first |
| 114 |
* key for the current attnum) until we hit null items or end of attnum |
| 115 |
* 3. SEARCH_MODE_EVERYTHING: scan from current point (which should be first |
| 116 |
* key for the current attnum) until we hit end of attnum |
| 117 |
* |
| 118 |
* Returns true if done, false if it's necessary to restart scan from scratch |
| 119 |
*/ |
| 120 |
static bool |
| 121 |
collectMatchBitmap(GinBtreeData *btree, GinBtreeStack *stack, |
| 122 |
GinScanEntry scanEntry, Snapshot snapshot) |
| 123 |
{ |
| 124 |
OffsetNumber attnum; |
| 125 |
CompactAttribute *attr; |
| 126 |
|
| 127 |
/* Initialize empty bitmap result */ |
| 128 |
scanEntry->matchBitmap = tbm_create(work_mem * (Size) 1024, NULL); |
| 129 |
|
| 130 |
/* Null query cannot partial-match anything */ |
| 131 |
if (scanEntry->isPartialMatch && |
| 132 |
scanEntry->queryCategory != GIN_CAT_NORM_KEY) |
| 133 |
return true; |
| 134 |
|
| 135 |
/* Locate tupdesc entry for key column (for attbyval/attlen data) */ |
| 136 |
attnum = scanEntry->attnum; |
| 137 |
attr = TupleDescCompactAttr(btree->ginstate->origTupdesc, attnum - 1); |
| 138 |
|
| 139 |
/* |
| 140 |
* Predicate lock entry leaf page, following pages will be locked by |
| 141 |
* moveRightIfItNeeded() |
| 142 |
*/ |
| 143 |
PredicateLockPage(btree->index, |
| 144 |
BufferGetBlockNumber(stack->buffer), |
| 145 |
snapshot); |
| 146 |
|
| 147 |
for (;;) |
| 148 |
{ |
| 149 |
Page page; |
| 150 |
IndexTuple itup; |
| 151 |
Datum idatum; |
| 152 |
GinNullCategory icategory; |
| 153 |
|
| 154 |
/* |
| 155 |
* stack->off points to the interested entry, buffer is already locked |
| 156 |
*/ |
| 157 |
if (moveRightIfItNeeded(btree, stack, snapshot) == false) |
| 158 |
return true; |
| 159 |
|
| 160 |
page = BufferGetPage(stack->buffer); |
| 161 |
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, stack->off)); |
| 162 |
|
| 163 |
/* |
| 164 |
* If tuple stores another attribute then stop scan |
| 165 |
*/ |
| 166 |
if (gintuple_get_attrnum(btree->ginstate, itup) != attnum) |
| 167 |
return true; |
| 168 |
|
| 169 |
/* Safe to fetch attribute value */ |
| 170 |
idatum = gintuple_get_key(btree->ginstate, itup, &icategory); |
| 171 |
|
| 172 |
/* |
| 173 |
* Check for appropriate scan stop conditions |
| 174 |
*/ |
| 175 |
if (scanEntry->isPartialMatch) |
| 176 |
{ |
| 177 |
int32 cmp; |
| 178 |
|
| 179 |
/* |
| 180 |
* In partial match, stop scan at any null (including |
| 181 |
* placeholders); partial matches never match nulls |
| 182 |
*/ |
| 183 |
if (icategory != GIN_CAT_NORM_KEY) |
| 184 |
return true; |
| 185 |
|
| 186 |
/*---------- |
| 187 |
* Check of partial match. |
| 188 |
* case cmp == 0 => match |
| 189 |
* case cmp > 0 => not match and finish scan |
| 190 |
* case cmp < 0 => not match and continue scan |
| 191 |
*---------- |
| 192 |
*/ |
| 193 |
cmp = DatumGetInt32(FunctionCall4Coll(&btree->ginstate->comparePartialFn[attnum - 1], |
| 194 |
btree->ginstate->supportCollation[attnum - 1], |
| 195 |
scanEntry->queryKey, |
| 196 |
idatum, |
| 197 |
UInt16GetDatum(scanEntry->strategy), |
| 198 |
PointerGetDatum(scanEntry->extra_data))); |
| 199 |
|
| 200 |
if (cmp > 0) |
| 201 |
return true; |
| 202 |
else if (cmp < 0) |
| 203 |
{ |
| 204 |
stack->off++; |
| 205 |
continue; |
| 206 |
} |
| 207 |
} |
| 208 |
else if (scanEntry->searchMode == GIN_SEARCH_MODE_ALL) |
| 209 |
{ |
| 210 |
/* |
| 211 |
* In ALL mode, we are not interested in null items, so we can |
| 212 |
* stop if we get to a null-item placeholder (which will be the |
| 213 |
* last entry for a given attnum). We do want to include NULL_KEY |
| 214 |
* and EMPTY_ITEM entries, though. |
| 215 |
*/ |
| 216 |
if (icategory == GIN_CAT_NULL_ITEM) |
| 217 |
return true; |
| 218 |
} |
| 219 |
|
| 220 |
/* |
| 221 |
* OK, we want to return the TIDs listed in this entry. |
| 222 |
*/ |
| 223 |
if (GinIsPostingTree(itup)) |
| 224 |
{ |
| 225 |
BlockNumber rootPostingTree = GinGetPostingTree(itup); |
| 226 |
|
| 227 |
/* |
| 228 |
* We should unlock current page (but not unpin) during tree scan |
| 229 |
* to prevent deadlock with vacuum processes. |
| 230 |
* |
| 231 |
* We save current entry value (idatum) to be able to re-find our |
| 232 |
* tuple after re-locking |
| 233 |
*/ |
| 234 |
if (icategory == GIN_CAT_NORM_KEY) |
| 235 |
idatum = datumCopy(idatum, attr->attbyval, attr->attlen); |
| 236 |
|
| 237 |
LockBuffer(stack->buffer, GIN_UNLOCK); |
| 238 |
|
| 239 |
/* |
| 240 |
* Acquire predicate lock on the posting tree. We already hold a |
| 241 |
* lock on the entry page, but insertions to the posting tree |
| 242 |
* don't check for conflicts on that level. |
| 243 |
*/ |
| 244 |
PredicateLockPage(btree->index, rootPostingTree, snapshot); |
| 245 |
|
| 246 |
/* Collect all the TIDs in this entry's posting tree */ |
| 247 |
scanPostingTree(btree->index, scanEntry, rootPostingTree); |
| 248 |
|
| 249 |
/* |
| 250 |
* We lock again the entry page and while it was unlocked insert |
| 251 |
* might have occurred, so we need to re-find our position. |
| 252 |
*/ |
| 253 |
LockBuffer(stack->buffer, GIN_SHARE); |
| 254 |
page = BufferGetPage(stack->buffer); |
| 255 |
if (!GinPageIsLeaf(page)) |
| 256 |
{ |
| 257 |
/* |
| 258 |
* Root page becomes non-leaf while we unlock it. We will |
| 259 |
* start again, this situation doesn't occur often - root can |
| 260 |
* became a non-leaf only once per life of index. |
| 261 |
*/ |
| 262 |
return false; |
| 263 |
} |
| 264 |
|
| 265 |
/* Search forward to re-find idatum */ |
| 266 |
for (;;) |
| 267 |
{ |
| 268 |
if (moveRightIfItNeeded(btree, stack, snapshot) == false) |
| 269 |
ereport(ERROR, |
| 270 |
(errcode(ERRCODE_INTERNAL_ERROR), |
| 271 |
errmsg("failed to re-find tuple within index \"%s\"", |
| 272 |
RelationGetRelationName(btree->index)))); |
| 273 |
|
| 274 |
page = BufferGetPage(stack->buffer); |
| 275 |
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, stack->off)); |
| 276 |
|
| 277 |
if (gintuple_get_attrnum(btree->ginstate, itup) == attnum) |
| 278 |
{ |
| 279 |
Datum newDatum; |
| 280 |
GinNullCategory newCategory; |
| 281 |
|
| 282 |
newDatum = gintuple_get_key(btree->ginstate, itup, |
| 283 |
&newCategory); |
| 284 |
|
| 285 |
if (ginCompareEntries(btree->ginstate, attnum, |
| 286 |
newDatum, newCategory, |
| 287 |
idatum, icategory) == 0) |
| 288 |
break; /* Found! */ |
| 289 |
} |
| 290 |
|
| 291 |
stack->off++; |
| 292 |
} |
| 293 |
|
| 294 |
if (icategory == GIN_CAT_NORM_KEY && !attr->attbyval) |
| 295 |
pfree(DatumGetPointer(idatum)); |
| 296 |
} |
| 297 |
else |
| 298 |
{ |
| 299 |
ItemPointer ipd; |
| 300 |
int nipd; |
| 301 |
|
| 302 |
ipd = ginReadTuple(btree->ginstate, scanEntry->attnum, itup, &nipd); |
| 303 |
tbm_add_tuples(scanEntry->matchBitmap, ipd, nipd, false); |
| 304 |
scanEntry->predictNumberResult += GinGetNPosting(itup); |
| 305 |
pfree(ipd); |
| 306 |
} |
| 307 |
|
| 308 |
/* |
| 309 |
* Done with this entry, go to the next |
| 310 |
*/ |
| 311 |
stack->off++; |
| 312 |
} |
| 313 |
} |
| 314 |
|
| 315 |
/* |
| 316 |
* Start* functions setup beginning state of searches: finds correct buffer and pins it. |
| 317 |
*/ |
| 318 |
static void |
| 319 |
startScanEntry(GinState *ginstate, GinScanEntry entry, Snapshot snapshot) |
| 320 |
{ |
| 321 |
GinBtreeData btreeEntry; |
| 322 |
GinBtreeStack *stackEntry; |
| 323 |
Page page; |
| 324 |
bool needUnlock; |
| 325 |
|
| 326 |
restartScanEntry: |
| 327 |
entry->buffer = InvalidBuffer; |
| 328 |
ItemPointerSetMin(&entry->curItem); |
| 329 |
entry->offset = InvalidOffsetNumber; |
| 330 |
if (entry->list) |
| 331 |
pfree(entry->list); |
| 332 |
entry->list = NULL; |
| 333 |
entry->nlist = 0; |
| 334 |
entry->matchBitmap = NULL; |
| 335 |
entry->matchNtuples = -1; |
| 336 |
entry->matchResult.blockno = InvalidBlockNumber; |
| 337 |
entry->reduceResult = false; |
| 338 |
entry->predictNumberResult = 0; |
| 339 |
|
| 340 |
/* |
| 341 |
* we should find entry, and begin scan of posting tree or just store |
| 342 |
* posting list in memory |
| 343 |
*/ |
| 344 |
ginPrepareEntryScan(&btreeEntry, entry->attnum, |
| 345 |
entry->queryKey, entry->queryCategory, |
| 346 |
ginstate); |
| 347 |
stackEntry = ginFindLeafPage(&btreeEntry, true, false); |
| 348 |
page = BufferGetPage(stackEntry->buffer); |
| 349 |
|
| 350 |
/* ginFindLeafPage() will have already checked snapshot age. */ |
| 351 |
needUnlock = true; |
| 352 |
|
| 353 |
entry->isFinished = true; |
| 354 |
|
| 355 |
if (entry->isPartialMatch || |
| 356 |
entry->queryCategory == GIN_CAT_EMPTY_QUERY) |
| 357 |
{ |
| 358 |
/* |
| 359 |
* btreeEntry.findItem locates the first item >= given search key. |
| 360 |
* (For GIN_CAT_EMPTY_QUERY, it will find the leftmost index item |
| 361 |
* because of the way the GIN_CAT_EMPTY_QUERY category code is |
| 362 |
* assigned.) We scan forward from there and collect all TIDs needed |
| 363 |
* for the entry type. |
| 364 |
*/ |
| 365 |
btreeEntry.findItem(&btreeEntry, stackEntry); |
| 366 |
if (collectMatchBitmap(&btreeEntry, stackEntry, entry, snapshot) |
| 367 |
== false) |
| 368 |
{ |
| 369 |
/* |
| 370 |
* GIN tree was seriously restructured, so we will cleanup all |
| 371 |
* found data and rescan. See comments near 'return false' in |
| 372 |
* collectMatchBitmap() |
| 373 |
*/ |
| 374 |
if (entry->matchBitmap) |
| 375 |
{ |
| 376 |
if (entry->matchIterator) |
| 377 |
tbm_end_private_iterate(entry->matchIterator); |
| 378 |
entry->matchIterator = NULL; |
| 379 |
tbm_free(entry->matchBitmap); |
| 380 |
entry->matchBitmap = NULL; |
| 381 |
} |
| 382 |
LockBuffer(stackEntry->buffer, GIN_UNLOCK); |
| 383 |
freeGinBtreeStack(stackEntry); |
| 384 |
goto restartScanEntry; |
| 385 |
} |
| 386 |
|
| 387 |
if (entry->matchBitmap && !tbm_is_empty(entry->matchBitmap)) |
| 388 |
{ |
| 389 |
entry->matchIterator = |
| 390 |
tbm_begin_private_iterate(entry->matchBitmap); |
| 391 |
entry->isFinished = false; |
| 392 |
} |
| 393 |
} |
| 394 |
else if (btreeEntry.findItem(&btreeEntry, stackEntry)) |
| 395 |
{ |
| 396 |
IndexTuple itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, stackEntry->off)); |
| 397 |
|
| 398 |
if (GinIsPostingTree(itup)) |
| 399 |
{ |
| 400 |
BlockNumber rootPostingTree = GinGetPostingTree(itup); |
| 401 |
GinBtreeStack *stack; |
| 402 |
Page entrypage; |
| 403 |
ItemPointerData minItem; |
| 404 |
|
| 405 |
/* |
| 406 |
* This is an equality scan, so lock the root of the posting tree. |
| 407 |
* It represents a lock on the exact key value, and covers all the |
| 408 |
* items in the posting tree. |
| 409 |
*/ |
| 410 |
PredicateLockPage(ginstate->index, rootPostingTree, snapshot); |
| 411 |
|
| 412 |
/* |
| 413 |
* We should unlock entry page before touching posting tree to |
| 414 |
* prevent deadlocks with vacuum processes. Because entry is never |
| 415 |
* deleted from page and posting tree is never reduced to the |
| 416 |
* posting list, we can unlock page after getting BlockNumber of |
| 417 |
* root of posting tree. |
| 418 |
*/ |
| 419 |
LockBuffer(stackEntry->buffer, GIN_UNLOCK); |
| 420 |
needUnlock = false; |
| 421 |
|
| 422 |
stack = ginScanBeginPostingTree(&entry->btree, ginstate->index, |
| 423 |
rootPostingTree); |
| 424 |
entry->buffer = stack->buffer; |
| 425 |
|
| 426 |
/* |
| 427 |
* We keep buffer pinned because we need to prevent deletion of |
| 428 |
* page during scan. See GIN's vacuum implementation. RefCount is |
| 429 |
* increased to keep buffer pinned after freeGinBtreeStack() call. |
| 430 |
*/ |
| 431 |
IncrBufferRefCount(entry->buffer); |
| 432 |
|
| 433 |
entrypage = BufferGetPage(entry->buffer); |
| 434 |
|
| 435 |
/* |
| 436 |
* Load the first page into memory. |
| 437 |
*/ |
| 438 |
ItemPointerSetMin(&minItem); |
| 439 |
entry->list = GinDataLeafPageGetItems(entrypage, &entry->nlist, minItem); |
| 440 |
|
| 441 |
entry->predictNumberResult = stack->predictNumber * entry->nlist; |
| 442 |
|
| 443 |
LockBuffer(entry->buffer, GIN_UNLOCK); |
| 444 |
freeGinBtreeStack(stack); |
| 445 |
entry->isFinished = false; |
| 446 |
} |
| 447 |
else |
| 448 |
{ |
| 449 |
/* |
| 450 |
* Lock the entry leaf page. This is more coarse-grained than |
| 451 |
* necessary, because it will conflict with any insertions that |
| 452 |
* land on the same leaf page, not only the exact key we searched |
| 453 |
* for. But locking an individual tuple would require updating |
| 454 |
* that lock whenever it moves because of insertions or vacuums, |
| 455 |
* which seems too complicated. |
| 456 |
*/ |
| 457 |
PredicateLockPage(ginstate->index, |
| 458 |
BufferGetBlockNumber(stackEntry->buffer), |
| 459 |
snapshot); |
| 460 |
if (GinGetNPosting(itup) > 0) |
| 461 |
{ |
| 462 |
entry->list = ginReadTuple(ginstate, entry->attnum, itup, |
| 463 |
&entry->nlist); |
| 464 |
entry->predictNumberResult = entry->nlist; |
| 465 |
|
| 466 |
entry->isFinished = false; |
| 467 |
} |
| 468 |
} |
| 469 |
} |
| 470 |
else |
| 471 |
{ |
| 472 |
/* |
| 473 |
* No entry found. Predicate lock the leaf page, to lock the place |
| 474 |
* where the entry would've been, had there been one. |
| 475 |
*/ |
| 476 |
PredicateLockPage(ginstate->index, |
| 477 |
BufferGetBlockNumber(stackEntry->buffer), snapshot); |
| 478 |
} |
| 479 |
|
| 480 |
if (needUnlock) |
| 481 |
LockBuffer(stackEntry->buffer, GIN_UNLOCK); |
| 482 |
freeGinBtreeStack(stackEntry); |
| 483 |
} |
| 484 |
|
| 485 |
/* |
| 486 |
* Comparison function for scan entry indexes. Sorts by predictNumberResult, |
| 487 |
* least frequent items first. |
| 488 |
*/ |
| 489 |
static int |
| 490 |
entryIndexByFrequencyCmp(const void *a1, const void *a2, void *arg) |
| 491 |
{ |
| 492 |
const GinScanKey key = (const GinScanKey) arg; |
| 493 |
int i1 = *(const int *) a1; |
| 494 |
int i2 = *(const int *) a2; |
| 495 |
uint32 n1 = key->scanEntry[i1]->predictNumberResult; |
| 496 |
uint32 n2 = key->scanEntry[i2]->predictNumberResult; |
| 497 |
|
| 498 |
if (n1 < n2) |
| 499 |
return -1; |
| 500 |
else if (n1 == n2) |
| 501 |
return 0; |
| 502 |
else |
| 503 |
return 1; |
| 504 |
} |
| 505 |
|
| 506 |
static void |
| 507 |
startScanKey(GinState *ginstate, GinScanOpaque so, GinScanKey key) |
| 508 |
{ |
| 509 |
MemoryContext oldCtx = CurrentMemoryContext; |
| 510 |
int i; |
| 511 |
int j; |
| 512 |
int *entryIndexes; |
| 513 |
|
| 514 |
ItemPointerSetMin(&key->curItem); |
| 515 |
key->curItemMatches = false; |
| 516 |
key->recheckCurItem = false; |
| 517 |
key->isFinished = false; |
| 518 |
|
| 519 |
/* |
| 520 |
* Divide the entries into two distinct sets: required and additional. |
| 521 |
* Additional entries are not enough for a match alone, without any items |
| 522 |
* from the required set, but are needed by the consistent function to |
| 523 |
* decide if an item matches. When scanning, we can skip over items from |
| 524 |
* additional entries that have no corresponding matches in any of the |
| 525 |
* required entries. That speeds up queries like "frequent & rare" |
| 526 |
* considerably, if the frequent term can be put in the additional set. |
| 527 |
* |
| 528 |
* There can be many legal ways to divide them entries into these two |
| 529 |
* sets. A conservative division is to just put everything in the required |
| 530 |
* set, but the more you can put in the additional set, the more you can |
| 531 |
* skip during the scan. To maximize skipping, we try to put as many |
| 532 |
* frequent items as possible into additional, and less frequent ones into |
| 533 |
* required. To do that, sort the entries by frequency |
| 534 |
* (predictNumberResult), and put entries into the required set in that |
| 535 |
* order, until the consistent function says that none of the remaining |
| 536 |
* entries can form a match, without any items from the required set. The |
| 537 |
* rest go to the additional set. |
| 538 |
* |
| 539 |
* Exclude-only scan keys are known to have no required entries. |
| 540 |
*/ |
| 541 |
if (key->excludeOnly) |
| 542 |
{ |
| 543 |
MemoryContextSwitchTo(so->keyCtx); |
| 544 |
|
| 545 |
key->nrequired = 0; |
| 546 |
key->nadditional = key->nentries; |
| 547 |
key->additionalEntries = palloc(key->nadditional * sizeof(GinScanEntry)); |
| 548 |
for (i = 0; i < key->nadditional; i++) |
| 549 |
key->additionalEntries[i] = key->scanEntry[i]; |
| 550 |
} |
| 551 |
else if (key->nentries > 1) |
| 552 |
{ |
| 553 |
MemoryContextSwitchTo(so->tempCtx); |
| 554 |
|
| 555 |
entryIndexes = (int *) palloc(sizeof(int) * key->nentries); |
| 556 |
for (i = 0; i < key->nentries; i++) |
| 557 |
entryIndexes[i] = i; |
| 558 |
qsort_arg(entryIndexes, key->nentries, sizeof(int), |
| 559 |
entryIndexByFrequencyCmp, key); |
| 560 |
|
| 561 |
for (i = 1; i < key->nentries; i++) |
| 562 |
key->entryRes[entryIndexes[i]] = GIN_MAYBE; |
| 563 |
for (i = 0; i < key->nentries - 1; i++) |
| 564 |
{ |
| 565 |
/* Pass all entries <= i as FALSE, and the rest as MAYBE */ |
| 566 |
key->entryRes[entryIndexes[i]] = GIN_FALSE; |
| 567 |
|
| 568 |
if (key->triConsistentFn(key) == GIN_FALSE) |
| 569 |
break; |
| 570 |
|
| 571 |
/* Make this loop interruptible in case there are many keys */ |
| 572 |
CHECK_FOR_INTERRUPTS(); |
| 573 |
} |
| 574 |
/* i is now the last required entry. */ |
| 575 |
|
| 576 |
MemoryContextSwitchTo(so->keyCtx); |
| 577 |
|
| 578 |
key->nrequired = i + 1; |
| 579 |
key->nadditional = key->nentries - key->nrequired; |
| 580 |
key->requiredEntries = palloc(key->nrequired * sizeof(GinScanEntry)); |
| 581 |
key->additionalEntries = palloc(key->nadditional * sizeof(GinScanEntry)); |
| 582 |
|
| 583 |
j = 0; |
| 584 |
for (i = 0; i < key->nrequired; i++) |
| 585 |
key->requiredEntries[i] = key->scanEntry[entryIndexes[j++]]; |
| 586 |
for (i = 0; i < key->nadditional; i++) |
| 587 |
key->additionalEntries[i] = key->scanEntry[entryIndexes[j++]]; |
| 588 |
|
| 589 |
/* clean up after consistentFn calls (also frees entryIndexes) */ |
| 590 |
MemoryContextReset(so->tempCtx); |
| 591 |
} |
| 592 |
else |
| 593 |
{ |
| 594 |
MemoryContextSwitchTo(so->keyCtx); |
| 595 |
|
| 596 |
key->nrequired = 1; |
| 597 |
key->nadditional = 0; |
| 598 |
key->requiredEntries = palloc(1 * sizeof(GinScanEntry)); |
| 599 |
key->requiredEntries[0] = key->scanEntry[0]; |
| 600 |
} |
| 601 |
MemoryContextSwitchTo(oldCtx); |
| 602 |
} |
| 603 |
|
| 604 |
static void |
| 605 |
startScan(IndexScanDesc scan) |
| 606 |
{ |
| 607 |
GinScanOpaque so = (GinScanOpaque) scan->opaque; |
| 608 |
GinState *ginstate = &so->ginstate; |
| 609 |
uint32 i; |
| 610 |
|
| 611 |
for (i = 0; i < so->totalentries; i++) |
| 612 |
startScanEntry(ginstate, so->entries[i], scan->xs_snapshot); |
| 613 |
|
| 614 |
if (GinFuzzySearchLimit > 0) |
| 615 |
{ |
| 616 |
/* |
| 617 |
* If all of keys more than threshold we will try to reduce result, we |
| 618 |
* hope (and only hope, for intersection operation of array our |
| 619 |
* supposition isn't true), that total result will not more than |
| 620 |
* minimal predictNumberResult. |
| 621 |
*/ |
| 622 |
bool reduce = true; |
| 623 |
|
| 624 |
for (i = 0; i < so->totalentries; i++) |
| 625 |
{ |
| 626 |
if (so->entries[i]->predictNumberResult <= so->totalentries * GinFuzzySearchLimit) |
| 627 |
{ |
| 628 |
reduce = false; |
| 629 |
break; |
| 630 |
} |
| 631 |
} |
| 632 |
if (reduce) |
| 633 |
{ |
| 634 |
for (i = 0; i < so->totalentries; i++) |
| 635 |
{ |
| 636 |
so->entries[i]->predictNumberResult /= so->totalentries; |
| 637 |
so->entries[i]->reduceResult = true; |
| 638 |
} |
| 639 |
} |
| 640 |
} |
| 641 |
|
| 642 |
/* |
| 643 |
* Now that we have the estimates for the entry frequencies, finish |
| 644 |
* initializing the scan keys. |
| 645 |
*/ |
| 646 |
for (i = 0; i < so->nkeys; i++) |
| 647 |
startScanKey(ginstate, so, so->keys + i); |
| 648 |
} |
| 649 |
|
| 650 |
/* |
| 651 |
* Load the next batch of item pointers from a posting tree. |
| 652 |
* |
| 653 |
* Note that we copy the page into GinScanEntry->list array and unlock it, but |
| 654 |
* keep it pinned to prevent interference with vacuum. |
| 655 |
*/ |
| 656 |
static void |
| 657 |
entryLoadMoreItems(GinState *ginstate, GinScanEntry entry, |
| 658 |
ItemPointerData advancePast) |
| 659 |
{ |
| 660 |
Page page; |
| 661 |
int i; |
| 662 |
bool stepright; |
| 663 |
|
| 664 |
if (!BufferIsValid(entry->buffer)) |
| 665 |
{ |
| 666 |
entry->isFinished = true; |
| 667 |
return; |
| 668 |
} |
| 669 |
|
| 670 |
/* |
| 671 |
* We have two strategies for finding the correct page: step right from |
| 672 |
* the current page, or descend the tree again from the root. If |
| 673 |
* advancePast equals the current item, the next matching item should be |
| 674 |
* on the next page, so we step right. Otherwise, descend from root. |
| 675 |
*/ |
| 676 |
if (ginCompareItemPointers(&entry->curItem, &advancePast) == 0) |
| 677 |
{ |
| 678 |
stepright = true; |
| 679 |
LockBuffer(entry->buffer, GIN_SHARE); |
| 680 |
} |
| 681 |
else |
| 682 |
{ |
| 683 |
GinBtreeStack *stack; |
| 684 |
|
| 685 |
ReleaseBuffer(entry->buffer); |
| 686 |
|
| 687 |
/* |
| 688 |
* Set the search key, and find the correct leaf page. |
| 689 |
*/ |
| 690 |
if (ItemPointerIsLossyPage(&advancePast)) |
| 691 |
{ |
| 692 |
ItemPointerSet(&entry->btree.itemptr, |
| 693 |
GinItemPointerGetBlockNumber(&advancePast) + 1, |
| 694 |
FirstOffsetNumber); |
| 695 |
} |
| 696 |
else |
| 697 |
{ |
| 698 |
ItemPointerSet(&entry->btree.itemptr, |
| 699 |
GinItemPointerGetBlockNumber(&advancePast), |
| 700 |
OffsetNumberNext(GinItemPointerGetOffsetNumber(&advancePast))); |
| 701 |
} |
| 702 |
entry->btree.fullScan = false; |
| 703 |
stack = ginFindLeafPage(&entry->btree, true, false); |
| 704 |
|
| 705 |
/* we don't need the stack, just the buffer. */ |
| 706 |
entry->buffer = stack->buffer; |
| 707 |
IncrBufferRefCount(entry->buffer); |
| 708 |
freeGinBtreeStack(stack); |
| 709 |
stepright = false; |
| 710 |
} |
| 711 |
|
| 712 |
elog(DEBUG2, "entryLoadMoreItems, %u/%u, skip: %d", |
| 713 |
GinItemPointerGetBlockNumber(&advancePast), |
| 714 |
GinItemPointerGetOffsetNumber(&advancePast), |
| 715 |
!stepright); |
| 716 |
|
| 717 |
page = BufferGetPage(entry->buffer); |
| 718 |
for (;;) |
| 719 |
{ |
| 720 |
entry->offset = InvalidOffsetNumber; |
| 721 |
if (entry->list) |
| 722 |
{ |
| 723 |
pfree(entry->list); |
| 724 |
entry->list = NULL; |
| 725 |
entry->nlist = 0; |
| 726 |
} |
| 727 |
|
| 728 |
if (stepright) |
| 729 |
{ |
| 730 |
/* |
| 731 |
* We've processed all the entries on this page. If it was the |
| 732 |
* last page in the tree, we're done. |
| 733 |
*/ |
| 734 |
if (GinPageRightMost(page)) |
| 735 |
{ |
| 736 |
UnlockReleaseBuffer(entry->buffer); |
| 737 |
entry->buffer = InvalidBuffer; |
| 738 |
entry->isFinished = true; |
| 739 |
return; |
| 740 |
} |
| 741 |
|
| 742 |
/* |
| 743 |
* Step to next page, following the right link. then find the |
| 744 |
* first ItemPointer greater than advancePast. |
| 745 |
*/ |
| 746 |
entry->buffer = ginStepRight(entry->buffer, |
| 747 |
ginstate->index, |
| 748 |
GIN_SHARE); |
| 749 |
page = BufferGetPage(entry->buffer); |
| 750 |
} |
| 751 |
stepright = true; |
| 752 |
|
| 753 |
if (GinPageGetOpaque(page)->flags & GIN_DELETED) |
| 754 |
continue; /* page was deleted by concurrent vacuum */ |
| 755 |
|
| 756 |
/* |
| 757 |
* The first item > advancePast might not be on this page, but |
| 758 |
* somewhere to the right, if the page was split, or a non-match from |
| 759 |
* another key in the query allowed us to skip some items from this |
| 760 |
* entry. Keep following the right-links until we re-find the correct |
| 761 |
* page. |
| 762 |
*/ |
| 763 |
if (!GinPageRightMost(page) && |
| 764 |
ginCompareItemPointers(&advancePast, GinDataPageGetRightBound(page)) >= 0) |
| 765 |
{ |
| 766 |
/* |
| 767 |
* the item we're looking is > the right bound of the page, so it |
| 768 |
* can't be on this page. |
| 769 |
*/ |
| 770 |
continue; |
| 771 |
} |
| 772 |
|
| 773 |
entry->list = GinDataLeafPageGetItems(page, &entry->nlist, advancePast); |
| 774 |
|
| 775 |
for (i = 0; i < entry->nlist; i++) |
| 776 |
{ |
| 777 |
if (ginCompareItemPointers(&advancePast, &entry->list[i]) < 0) |
| 778 |
{ |
| 779 |
entry->offset = i; |
| 780 |
|
| 781 |
if (GinPageRightMost(page)) |
| 782 |
{ |
| 783 |
/* after processing the copied items, we're done. */ |
| 784 |
UnlockReleaseBuffer(entry->buffer); |
| 785 |
entry->buffer = InvalidBuffer; |
| 786 |
} |
| 787 |
else |
| 788 |
LockBuffer(entry->buffer, GIN_UNLOCK); |
| 789 |
return; |
| 790 |
} |
| 791 |
} |
| 792 |
} |
| 793 |
} |
| 794 |
|
| 795 |
#define gin_rand() pg_prng_double(&pg_global_prng_state) |
| 796 |
#define dropItem(e) ( gin_rand() > ((double)GinFuzzySearchLimit)/((double)((e)->predictNumberResult)) ) |
| 797 |
|
| 798 |
/* |
| 799 |
* Sets entry->curItem to next heap item pointer > advancePast, for one entry |
| 800 |
* of one scan key, or sets entry->isFinished to true if there are no more. |
| 801 |
* |
| 802 |
* Item pointers are returned in ascending order. |
| 803 |
* |
| 804 |
* Note: this can return a "lossy page" item pointer, indicating that the |
| 805 |
* entry potentially matches all items on that heap page. However, it is |
| 806 |
* not allowed to return both a lossy page pointer and exact (regular) |
| 807 |
* item pointers for the same page. (Doing so would break the key-combination |
| 808 |
* logic in keyGetItem and scanGetItem; see comment in scanGetItem.) In the |
| 809 |
* current implementation this is guaranteed by the behavior of tidbitmaps. |
| 810 |
*/ |
| 811 |
static void |
| 812 |
entryGetItem(GinState *ginstate, GinScanEntry entry, |
| 813 |
ItemPointerData advancePast) |
| 814 |
{ |
| 815 |
Assert(!entry->isFinished); |
| 816 |
|
| 817 |
Assert(!ItemPointerIsValid(&entry->curItem) || |
| 818 |
ginCompareItemPointers(&entry->curItem, &advancePast) <= 0); |
| 819 |
|
| 820 |
if (entry->matchBitmap) |
| 821 |
{ |
| 822 |
/* A bitmap result */ |
| 823 |
BlockNumber advancePastBlk = GinItemPointerGetBlockNumber(&advancePast); |
| 824 |
OffsetNumber advancePastOff = GinItemPointerGetOffsetNumber(&advancePast); |
| 825 |
|
| 826 |
for (;;) |
| 827 |
{ |
| 828 |
/* |
| 829 |
* If we've exhausted all items on this block, move to next block |
| 830 |
* in the bitmap. tbm_private_iterate() sets matchResult.blockno |
| 831 |
* to InvalidBlockNumber when the bitmap is exhausted. |
| 832 |
*/ |
| 833 |
while ((!BlockNumberIsValid(entry->matchResult.blockno)) || |
| 834 |
(!entry->matchResult.lossy && |
| 835 |
entry->offset >= entry->matchNtuples) || |
| 836 |
entry->matchResult.blockno < advancePastBlk || |
| 837 |
(ItemPointerIsLossyPage(&advancePast) && |
| 838 |
entry->matchResult.blockno == advancePastBlk)) |
| 839 |
{ |
| 840 |
if (!tbm_private_iterate(entry->matchIterator, &entry->matchResult)) |
| 841 |
{ |
| 842 |
Assert(!BlockNumberIsValid(entry->matchResult.blockno)); |
| 843 |
ItemPointerSetInvalid(&entry->curItem); |
| 844 |
tbm_end_private_iterate(entry->matchIterator); |
| 845 |
entry->matchIterator = NULL; |
| 846 |
entry->isFinished = true; |
| 847 |
break; |
| 848 |
} |
| 849 |
|
| 850 |
/* Exact pages need their tuple offsets extracted. */ |
| 851 |
if (!entry->matchResult.lossy) |
| 852 |
entry->matchNtuples = tbm_extract_page_tuple(&entry->matchResult, |
| 853 |
entry->matchOffsets, |
| 854 |
TBM_MAX_TUPLES_PER_PAGE); |
| 855 |
|
| 856 |
/* |
| 857 |
* Reset counter to the beginning of entry->matchResult. Note: |
| 858 |
* entry->offset is still greater than matchResult.ntuples if |
| 859 |
* matchResult is lossy. So, on next call we will get next |
| 860 |
* result from TIDBitmap. |
| 861 |
*/ |
| 862 |
entry->offset = 0; |
| 863 |
} |
| 864 |
if (entry->isFinished) |
| 865 |
break; |
| 866 |
|
| 867 |
/* |
| 868 |
* We're now on the first page after advancePast which has any |
| 869 |
* items on it. If it's a lossy result, return that. |
| 870 |
*/ |
| 871 |
if (entry->matchResult.lossy) |
| 872 |
{ |
| 873 |
ItemPointerSetLossyPage(&entry->curItem, |
| 874 |
entry->matchResult.blockno); |
| 875 |
|
| 876 |
/* |
| 877 |
* We might as well fall out of the loop; we could not |
| 878 |
* estimate number of results on this page to support correct |
| 879 |
* reducing of result even if it's enabled. |
| 880 |
*/ |
| 881 |
break; |
| 882 |
} |
| 883 |
|
| 884 |
/* |
| 885 |
* Not a lossy page. If tuple offsets were extracted, |
| 886 |
* entry->matchNtuples must be > -1 |
| 887 |
*/ |
| 888 |
Assert(entry->matchNtuples > -1); |
| 889 |
|
| 890 |
/* Skip over any offsets <= advancePast, and return that. */ |
| 891 |
if (entry->matchResult.blockno == advancePastBlk) |
| 892 |
{ |
| 893 |
Assert(entry->matchNtuples > 0); |
| 894 |
|
| 895 |
/* |
| 896 |
* First, do a quick check against the last offset on the |
| 897 |
* page. If that's > advancePast, so are all the other |
| 898 |
* offsets, so just go back to the top to get the next page. |
| 899 |
*/ |
| 900 |
if (entry->matchOffsets[entry->matchNtuples - 1] <= advancePastOff) |
| 901 |
{ |
| 902 |
entry->offset = entry->matchNtuples; |
| 903 |
continue; |
| 904 |
} |
| 905 |
|
| 906 |
/* Otherwise scan to find the first item > advancePast */ |
| 907 |
while (entry->matchOffsets[entry->offset] <= advancePastOff) |
| 908 |
entry->offset++; |
| 909 |
} |
| 910 |
|
| 911 |
ItemPointerSet(&entry->curItem, |
| 912 |
entry->matchResult.blockno, |
| 913 |
entry->matchOffsets[entry->offset]); |
| 914 |
entry->offset++; |
| 915 |
|
| 916 |
/* Done unless we need to reduce the result */ |
| 917 |
if (!entry->reduceResult || !dropItem(entry)) |
| 918 |
break; |
| 919 |
} |
| 920 |
} |
| 921 |
else if (!BufferIsValid(entry->buffer)) |
| 922 |
{ |
| 923 |
/* |
| 924 |
* A posting list from an entry tuple, or the last page of a posting |
| 925 |
* tree. |
| 926 |
*/ |
| 927 |
for (;;) |
| 928 |
{ |
| 929 |
if (entry->offset >= entry->nlist) |
| 930 |
{ |
| 931 |
ItemPointerSetInvalid(&entry->curItem); |
| 932 |
entry->isFinished = true; |
| 933 |
break; |
| 934 |
} |
| 935 |
|
| 936 |
entry->curItem = entry->list[entry->offset++]; |
| 937 |
|
| 938 |
/* If we're not past advancePast, keep scanning */ |
| 939 |
if (ginCompareItemPointers(&entry->curItem, &advancePast) <= 0) |
| 940 |
continue; |
| 941 |
|
| 942 |
/* Done unless we need to reduce the result */ |
| 943 |
if (!entry->reduceResult || !dropItem(entry)) |
| 944 |
break; |
| 945 |
} |
| 946 |
} |
| 947 |
else |
| 948 |
{ |
| 949 |
/* A posting tree */ |
| 950 |
for (;;) |
| 951 |
{ |
| 952 |
/* If we've processed the current batch, load more items */ |
| 953 |
while (entry->offset >= entry->nlist) |
| 954 |
{ |
| 955 |
entryLoadMoreItems(ginstate, entry, advancePast); |
| 956 |
|
| 957 |
if (entry->isFinished) |
| 958 |
{ |
| 959 |
ItemPointerSetInvalid(&entry->curItem); |
| 960 |
return; |
| 961 |
} |
| 962 |
} |
| 963 |
|
| 964 |
entry->curItem = entry->list[entry->offset++]; |
| 965 |
|
| 966 |
/* If we're not past advancePast, keep scanning */ |
| 967 |
if (ginCompareItemPointers(&entry->curItem, &advancePast) <= 0) |
| 968 |
continue; |
| 969 |
|
| 970 |
/* Done unless we need to reduce the result */ |
| 971 |
if (!entry->reduceResult || !dropItem(entry)) |
| 972 |
break; |
| 973 |
|
| 974 |
/* |
| 975 |
* Advance advancePast (so that entryLoadMoreItems will load the |
| 976 |
* right data), and keep scanning |
| 977 |
*/ |
| 978 |
advancePast = entry->curItem; |
| 979 |
} |
| 980 |
} |
| 981 |
} |
| 982 |
|
| 983 |
/* |
| 984 |
* Identify the "current" item among the input entry streams for this scan key |
| 985 |
* that is greater than advancePast, and test whether it passes the scan key |
| 986 |
* qual condition. |
| 987 |
* |
| 988 |
* The current item is the smallest curItem among the inputs. key->curItem |
| 989 |
* is set to that value. key->curItemMatches is set to indicate whether that |
| 990 |
* TID passes the consistentFn test. If so, key->recheckCurItem is set true |
| 991 |
* iff recheck is needed for this item pointer (including the case where the |
| 992 |
* item pointer is a lossy page pointer). |
| 993 |
* |
| 994 |
* If all entry streams are exhausted, sets key->isFinished to true. |
| 995 |
* |
| 996 |
* Item pointers must be returned in ascending order. |
| 997 |
* |
| 998 |
* Note: this can return a "lossy page" item pointer, indicating that the |
| 999 |
* key potentially matches all items on that heap page. However, it is |
| 1000 |
* not allowed to return both a lossy page pointer and exact (regular) |
| 1001 |
* item pointers for the same page. (Doing so would break the key-combination |
| 1002 |
* logic in scanGetItem.) |
| 1003 |
*/ |
| 1004 |
static void |
| 1005 |
keyGetItem(GinState *ginstate, MemoryContext tempCtx, GinScanKey key, |
| 1006 |
ItemPointerData advancePast) |
| 1007 |
{ |
| 1008 |
ItemPointerData minItem; |
| 1009 |
ItemPointerData curPageLossy; |
| 1010 |
uint32 i; |
| 1011 |
bool haveLossyEntry; |
| 1012 |
GinScanEntry entry; |
| 1013 |
GinTernaryValue res; |
| 1014 |
MemoryContext oldCtx; |
| 1015 |
bool allFinished; |
| 1016 |
|
| 1017 |
Assert(!key->isFinished); |
| 1018 |
|
| 1019 |
/* |
| 1020 |
* We might have already tested this item; if so, no need to repeat work. |
| 1021 |
* (Note: the ">" case can happen, if advancePast is exact but we |
| 1022 |
* previously had to set curItem to a lossy-page pointer.) |
| 1023 |
*/ |
| 1024 |
if (ginCompareItemPointers(&key->curItem, &advancePast) > 0) |
| 1025 |
return; |
| 1026 |
|
| 1027 |
/* |
| 1028 |
* Find the minimum item > advancePast among the active entry streams. |
| 1029 |
* |
| 1030 |
* Note: a lossy-page entry is encoded by a ItemPointer with max value for |
| 1031 |
* offset (0xffff), so that it will sort after any exact entries for the |
| 1032 |
* same page. So we'll prefer to return exact pointers not lossy |
| 1033 |
* pointers, which is good. |
| 1034 |
*/ |
| 1035 |
ItemPointerSetMax(&minItem); |
| 1036 |
allFinished = true; |
| 1037 |
for (i = 0; i < key->nrequired; i++) |
| 1038 |
{ |
| 1039 |
entry = key->requiredEntries[i]; |
| 1040 |
|
| 1041 |
if (entry->isFinished) |
| 1042 |
continue; |
| 1043 |
|
| 1044 |
/* |
| 1045 |
* Advance this stream if necessary. |
| 1046 |
* |
| 1047 |
* In particular, since entry->curItem was initialized with |
| 1048 |
* ItemPointerSetMin, this ensures we fetch the first item for each |
| 1049 |
* entry on the first call. |
| 1050 |
*/ |
| 1051 |
if (ginCompareItemPointers(&entry->curItem, &advancePast) <= 0) |
| 1052 |
{ |
| 1053 |
entryGetItem(ginstate, entry, advancePast); |
| 1054 |
if (entry->isFinished) |
| 1055 |
continue; |
| 1056 |
} |
| 1057 |
|
| 1058 |
allFinished = false; |
| 1059 |
if (ginCompareItemPointers(&entry->curItem, &minItem) < 0) |
| 1060 |
minItem = entry->curItem; |
| 1061 |
} |
| 1062 |
|
| 1063 |
if (allFinished && !key->excludeOnly) |
| 1064 |
{ |
| 1065 |
/* all entries are finished */ |
| 1066 |
key->isFinished = true; |
| 1067 |
return; |
| 1068 |
} |
| 1069 |
|
| 1070 |
if (!key->excludeOnly) |
| 1071 |
{ |
| 1072 |
/* |
| 1073 |
* For a normal scan key, we now know there are no matches < minItem. |
| 1074 |
* |
| 1075 |
* If minItem is lossy, it means that there were no exact items on the |
| 1076 |
* page among requiredEntries, because lossy pointers sort after exact |
| 1077 |
* items. However, there might be exact items for the same page among |
| 1078 |
* additionalEntries, so we mustn't advance past them. |
| 1079 |
*/ |
| 1080 |
if (ItemPointerIsLossyPage(&minItem)) |
| 1081 |
{ |
| 1082 |
if (GinItemPointerGetBlockNumber(&advancePast) < |
| 1083 |
GinItemPointerGetBlockNumber(&minItem)) |
| 1084 |
{ |
| 1085 |
ItemPointerSet(&advancePast, |
| 1086 |
GinItemPointerGetBlockNumber(&minItem), |
| 1087 |
InvalidOffsetNumber); |
| 1088 |
} |
| 1089 |
} |
| 1090 |
else |
| 1091 |
{ |
| 1092 |
Assert(GinItemPointerGetOffsetNumber(&minItem) > 0); |
| 1093 |
ItemPointerSet(&advancePast, |
| 1094 |
GinItemPointerGetBlockNumber(&minItem), |
| 1095 |
OffsetNumberPrev(GinItemPointerGetOffsetNumber(&minItem))); |
| 1096 |
} |
| 1097 |
} |
| 1098 |
else |
| 1099 |
{ |
| 1100 |
/* |
| 1101 |
* excludeOnly scan keys don't have any entries that are necessarily |
| 1102 |
* present in matching items. So, we consider the item just after |
| 1103 |
* advancePast. |
| 1104 |
*/ |
| 1105 |
Assert(key->nrequired == 0); |
| 1106 |
ItemPointerSet(&minItem, |
| 1107 |
GinItemPointerGetBlockNumber(&advancePast), |
| 1108 |
OffsetNumberNext(GinItemPointerGetOffsetNumber(&advancePast))); |
| 1109 |
} |
| 1110 |
|
| 1111 |
/* |
| 1112 |
* We might not have loaded all the entry streams for this TID yet. We |
| 1113 |
* could call the consistent function, passing MAYBE for those entries, to |
| 1114 |
* see if it can decide if this TID matches based on the information we |
| 1115 |
* have. But if the consistent-function is expensive, and cannot in fact |
| 1116 |
* decide with partial information, that could be a big loss. So, load all |
| 1117 |
* the additional entries, before calling the consistent function. |
| 1118 |
*/ |
| 1119 |
for (i = 0; i < key->nadditional; i++) |
| 1120 |
{ |
| 1121 |
entry = key->additionalEntries[i]; |
| 1122 |
|
| 1123 |
if (entry->isFinished) |
| 1124 |
continue; |
| 1125 |
|
| 1126 |
if (ginCompareItemPointers(&entry->curItem, &advancePast) <= 0) |
| 1127 |
{ |
| 1128 |
entryGetItem(ginstate, entry, advancePast); |
| 1129 |
if (entry->isFinished) |
| 1130 |
continue; |
| 1131 |
} |
| 1132 |
|
| 1133 |
/* |
| 1134 |
* Normally, none of the items in additionalEntries can have a curItem |
| 1135 |
* larger than minItem. But if minItem is a lossy page, then there |
| 1136 |
* might be exact items on the same page among additionalEntries. |
| 1137 |
*/ |
| 1138 |
if (ginCompareItemPointers(&entry->curItem, &minItem) < 0) |
| 1139 |
{ |
| 1140 |
Assert(ItemPointerIsLossyPage(&minItem)); |
| 1141 |
minItem = entry->curItem; |
| 1142 |
} |
| 1143 |
} |
| 1144 |
|
| 1145 |
/* |
| 1146 |
* Ok, we've advanced all the entries up to minItem now. Set key->curItem, |
| 1147 |
* and perform consistentFn test. |
| 1148 |
* |
| 1149 |
* Lossy-page entries pose a problem, since we don't know the correct |
| 1150 |
* entryRes state to pass to the consistentFn, and we also don't know what |
| 1151 |
* its combining logic will be (could be AND, OR, or even NOT). If the |
| 1152 |
* logic is OR then the consistentFn might succeed for all items in the |
| 1153 |
* lossy page even when none of the other entries match. |
| 1154 |
* |
| 1155 |
* Our strategy is to call the tri-state consistent function, with the |
| 1156 |
* lossy-page entries set to MAYBE, and all the other entries FALSE. If it |
| 1157 |
* returns FALSE, none of the lossy items alone are enough for a match, so |
| 1158 |
* we don't need to return a lossy-page pointer. Otherwise, return a |
| 1159 |
* lossy-page pointer to indicate that the whole heap page must be |
| 1160 |
* checked. (On subsequent calls, we'll do nothing until minItem is past |
| 1161 |
* the page altogether, thus ensuring that we never return both regular |
| 1162 |
* and lossy pointers for the same page.) |
| 1163 |
* |
| 1164 |
* An exception is that it doesn't matter what we pass for lossy pointers |
| 1165 |
* in "hidden" entries, because the consistentFn's result can't depend on |
| 1166 |
* them. We could pass them as MAYBE as well, but if we're using the |
| 1167 |
* "shim" implementation of a tri-state consistent function (see |
| 1168 |
* ginlogic.c), it's better to pass as few MAYBEs as possible. So pass |
| 1169 |
* them as true. |
| 1170 |
* |
| 1171 |
* Note that only lossy-page entries pointing to the current item's page |
| 1172 |
* should trigger this processing; we might have future lossy pages in the |
| 1173 |
* entry array, but they aren't relevant yet. |
| 1174 |
*/ |
| 1175 |
key->curItem = minItem; |
| 1176 |
ItemPointerSetLossyPage(&curPageLossy, |
| 1177 |
GinItemPointerGetBlockNumber(&key->curItem)); |
| 1178 |
haveLossyEntry = false; |
| 1179 |
for (i = 0; i < key->nentries; i++) |
| 1180 |
{ |
| 1181 |
entry = key->scanEntry[i]; |
| 1182 |
if (entry->isFinished == false && |
| 1183 |
ginCompareItemPointers(&entry->curItem, &curPageLossy) == 0) |
| 1184 |
{ |
| 1185 |
if (i < key->nuserentries) |
| 1186 |
key->entryRes[i] = GIN_MAYBE; |
| 1187 |
else |
| 1188 |
key->entryRes[i] = GIN_TRUE; |
| 1189 |
haveLossyEntry = true; |
| 1190 |
} |
| 1191 |
else |
| 1192 |
key->entryRes[i] = GIN_FALSE; |
| 1193 |
} |
| 1194 |
|
| 1195 |
/* prepare for calling consistentFn in temp context */ |
| 1196 |
oldCtx = MemoryContextSwitchTo(tempCtx); |
| 1197 |
|
| 1198 |
if (haveLossyEntry) |
| 1199 |
{ |
| 1200 |
/* Have lossy-page entries, so see if whole page matches */ |
| 1201 |
res = key->triConsistentFn(key); |
| 1202 |
|
| 1203 |
if (res == GIN_TRUE || res == GIN_MAYBE) |
| 1204 |
{ |
| 1205 |
/* Yes, so clean up ... */ |
| 1206 |
MemoryContextSwitchTo(oldCtx); |
| 1207 |
MemoryContextReset(tempCtx); |
| 1208 |
|
| 1209 |
/* and return lossy pointer for whole page */ |
| 1210 |
key->curItem = curPageLossy; |
| 1211 |
key->curItemMatches = true; |
| 1212 |
key->recheckCurItem = true; |
| 1213 |
return; |
| 1214 |
} |
| 1215 |
} |
| 1216 |
|
| 1217 |
/* |
| 1218 |
* At this point we know that we don't need to return a lossy whole-page |
| 1219 |
* pointer, but we might have matches for individual exact item pointers, |
| 1220 |
* possibly in combination with a lossy pointer. Pass lossy pointers as |
| 1221 |
* MAYBE to the ternary consistent function, to let it decide if this |
| 1222 |
* tuple satisfies the overall key, even though we don't know if the lossy |
| 1223 |
* entries match. |
| 1224 |
* |
| 1225 |
* Prepare entryRes array to be passed to consistentFn. |
| 1226 |
*/ |
| 1227 |
for (i = 0; i < key->nentries; i++) |
| 1228 |
{ |
| 1229 |
entry = key->scanEntry[i]; |
| 1230 |
if (entry->isFinished) |
| 1231 |
key->entryRes[i] = GIN_FALSE; |
| 1232 |
#if 0 |
| 1233 |
|
| 1234 |
/* |
| 1235 |
* This case can't currently happen, because we loaded all the entries |
| 1236 |
* for this item earlier. |
| 1237 |
*/ |
| 1238 |
else if (ginCompareItemPointers(&entry->curItem, &advancePast) <= 0) |
| 1239 |
key->entryRes[i] = GIN_MAYBE; |
| 1240 |
#endif |
| 1241 |
else if (ginCompareItemPointers(&entry->curItem, &curPageLossy) == 0) |
| 1242 |
key->entryRes[i] = GIN_MAYBE; |
| 1243 |
else if (ginCompareItemPointers(&entry->curItem, &minItem) == 0) |
| 1244 |
key->entryRes[i] = GIN_TRUE; |
| 1245 |
else |
| 1246 |
key->entryRes[i] = GIN_FALSE; |
| 1247 |
} |
| 1248 |
|
| 1249 |
res = key->triConsistentFn(key); |
| 1250 |
|
| 1251 |
switch (res) |
| 1252 |
{ |
| 1253 |
case GIN_TRUE: |
| 1254 |
key->curItemMatches = true; |
| 1255 |
/* triConsistentFn set recheckCurItem */ |
| 1256 |
break; |
| 1257 |
|
| 1258 |
case GIN_FALSE: |
| 1259 |
key->curItemMatches = false; |
| 1260 |
break; |
| 1261 |
|
| 1262 |
case GIN_MAYBE: |
| 1263 |
key->curItemMatches = true; |
| 1264 |
key->recheckCurItem = true; |
| 1265 |
break; |
| 1266 |
|
| 1267 |
default: |
| 1268 |
|
| 1269 |
/* |
| 1270 |
* the 'default' case shouldn't happen, but if the consistent |
| 1271 |
* function returns something bogus, this is the safe result |
| 1272 |
*/ |
| 1273 |
key->curItemMatches = true; |
| 1274 |
key->recheckCurItem = true; |
| 1275 |
break; |
| 1276 |
} |
| 1277 |
|
| 1278 |
/* |
| 1279 |
* We have a tuple, and we know if it matches or not. If it's a non-match, |
| 1280 |
* we could continue to find the next matching tuple, but let's break out |
| 1281 |
* and give scanGetItem a chance to advance the other keys. They might be |
| 1282 |
* able to skip past to a much higher TID, allowing us to save work. |
| 1283 |
*/ |
| 1284 |
|
| 1285 |
/* clean up after consistentFn calls */ |
| 1286 |
MemoryContextSwitchTo(oldCtx); |
| 1287 |
MemoryContextReset(tempCtx); |
| 1288 |
} |
| 1289 |
|
| 1290 |
/* |
| 1291 |
* Get next heap item pointer (after advancePast) from scan. |
| 1292 |
* Returns true if anything found. |
| 1293 |
* On success, *item and *recheck are set. |
| 1294 |
* |
| 1295 |
* Note: this is very nearly the same logic as in keyGetItem(), except |
| 1296 |
* that we know the keys are to be combined with AND logic, whereas in |
| 1297 |
* keyGetItem() the combination logic is known only to the consistentFn. |
| 1298 |
*/ |
| 1299 |
static bool |
| 1300 |
scanGetItem(IndexScanDesc scan, ItemPointerData advancePast, |
| 1301 |
ItemPointerData *item, bool *recheck) |
| 1302 |
{ |
| 1303 |
GinScanOpaque so = (GinScanOpaque) scan->opaque; |
| 1304 |
uint32 i; |
| 1305 |
bool match; |
| 1306 |
|
| 1307 |
/*---------- |
| 1308 |
* Advance the scan keys in lock-step, until we find an item that matches |
| 1309 |
* all the keys. If any key reports isFinished, meaning its subset of the |
| 1310 |
* entries is exhausted, we can stop. Otherwise, set *item to the next |
| 1311 |
* matching item. |
| 1312 |
* |
| 1313 |
* This logic works only if a keyGetItem stream can never contain both |
| 1314 |
* exact and lossy pointers for the same page. Else we could have a |
| 1315 |
* case like |
| 1316 |
* |
| 1317 |
* stream 1 stream 2 |
| 1318 |
* ... ... |
| 1319 |
* 42/6 42/7 |
| 1320 |
* 50/1 42/0xffff |
| 1321 |
* ... ... |
| 1322 |
* |
| 1323 |
* We would conclude that 42/6 is not a match and advance stream 1, |
| 1324 |
* thus never detecting the match to the lossy pointer in stream 2. |
| 1325 |
* (keyGetItem has a similar problem versus entryGetItem.) |
| 1326 |
*---------- |
| 1327 |
*/ |
| 1328 |
do |
| 1329 |
{ |
| 1330 |
CHECK_FOR_INTERRUPTS(); |
| 1331 |
|
| 1332 |
ItemPointerSetMin(item); |
| 1333 |
match = true; |
| 1334 |
for (i = 0; i < so->nkeys && match; i++) |
| 1335 |
{ |
| 1336 |
GinScanKey key = so->keys + i; |
| 1337 |
|
| 1338 |
/* |
| 1339 |
* If we're considering a lossy page, skip excludeOnly keys. They |
| 1340 |
* can't exclude the whole page anyway. |
| 1341 |
*/ |
| 1342 |
if (ItemPointerIsLossyPage(item) && key->excludeOnly) |
| 1343 |
{ |
| 1344 |
/* |
| 1345 |
* ginNewScanKey() should never mark the first key as |
| 1346 |
* excludeOnly. |
| 1347 |
*/ |
| 1348 |
Assert(i > 0); |
| 1349 |
continue; |
| 1350 |
} |
| 1351 |
|
| 1352 |
/* Fetch the next item for this key that is > advancePast. */ |
| 1353 |
keyGetItem(&so->ginstate, so->tempCtx, key, advancePast); |
| 1354 |
|
| 1355 |
if (key->isFinished) |
| 1356 |
return false; |
| 1357 |
|
| 1358 |
/* |
| 1359 |
* If it's not a match, we can immediately conclude that nothing |
| 1360 |
* <= this item matches, without checking the rest of the keys. |
| 1361 |
*/ |
| 1362 |
if (!key->curItemMatches) |
| 1363 |
{ |
| 1364 |
advancePast = key->curItem; |
| 1365 |
match = false; |
| 1366 |
break; |
| 1367 |
} |
| 1368 |
|
| 1369 |
/* |
| 1370 |
* It's a match. We can conclude that nothing < matches, so the |
| 1371 |
* other key streams can skip to this item. |
| 1372 |
* |
| 1373 |
* Beware of lossy pointers, though; from a lossy pointer, we can |
| 1374 |
* only conclude that nothing smaller than this *block* matches. |
| 1375 |
*/ |
| 1376 |
if (ItemPointerIsLossyPage(&key->curItem)) |
| 1377 |
{ |
| 1378 |
if (GinItemPointerGetBlockNumber(&advancePast) < |
| 1379 |
GinItemPointerGetBlockNumber(&key->curItem)) |
| 1380 |
{ |
| 1381 |
ItemPointerSet(&advancePast, |
| 1382 |
GinItemPointerGetBlockNumber(&key->curItem), |
| 1383 |
InvalidOffsetNumber); |
| 1384 |
} |
| 1385 |
} |
| 1386 |
else |
| 1387 |
{ |
| 1388 |
Assert(GinItemPointerGetOffsetNumber(&key->curItem) > 0); |
| 1389 |
ItemPointerSet(&advancePast, |
| 1390 |
GinItemPointerGetBlockNumber(&key->curItem), |
| 1391 |
OffsetNumberPrev(GinItemPointerGetOffsetNumber(&key->curItem))); |
| 1392 |
} |
| 1393 |
|
| 1394 |
/* |
| 1395 |
* If this is the first key, remember this location as a potential |
| 1396 |
* match, and proceed to check the rest of the keys. |
| 1397 |
* |
| 1398 |
* Otherwise, check if this is the same item that we checked the |
| 1399 |
* previous keys for (or a lossy pointer for the same page). If |
| 1400 |
* not, loop back to check the previous keys for this item (we |
| 1401 |
* will check this key again too, but keyGetItem returns quickly |
| 1402 |
* for that) |
| 1403 |
*/ |
| 1404 |
if (i == 0) |
| 1405 |
{ |
| 1406 |
*item = key->curItem; |
| 1407 |
} |
| 1408 |
else |
| 1409 |
{ |
| 1410 |
if (ItemPointerIsLossyPage(&key->curItem) || |
| 1411 |
ItemPointerIsLossyPage(item)) |
| 1412 |
{ |
| 1413 |
Assert(GinItemPointerGetBlockNumber(&key->curItem) >= GinItemPointerGetBlockNumber(item)); |
| 1414 |
match = (GinItemPointerGetBlockNumber(&key->curItem) == |
| 1415 |
GinItemPointerGetBlockNumber(item)); |
| 1416 |
} |
| 1417 |
else |
| 1418 |
{ |
| 1419 |
Assert(ginCompareItemPointers(&key->curItem, item) >= 0); |
| 1420 |
match = (ginCompareItemPointers(&key->curItem, item) == 0); |
| 1421 |
} |
| 1422 |
} |
| 1423 |
} |
| 1424 |
} while (!match); |
| 1425 |
|
| 1426 |
Assert(!ItemPointerIsMin(item)); |
| 1427 |
|
| 1428 |
/* |
| 1429 |
* Now *item contains the first ItemPointer after previous result that |
| 1430 |
* satisfied all the keys for that exact TID, or a lossy reference to the |
| 1431 |
* same page. |
| 1432 |
* |
| 1433 |
* We must return recheck = true if any of the keys are marked recheck. |
| 1434 |
*/ |
| 1435 |
*recheck = false; |
| 1436 |
for (i = 0; i < so->nkeys; i++) |
| 1437 |
{ |
| 1438 |
GinScanKey key = so->keys + i; |
| 1439 |
|
| 1440 |
if (key->recheckCurItem) |
| 1441 |
{ |
| 1442 |
*recheck = true; |
| 1443 |
break; |
| 1444 |
} |
| 1445 |
} |
| 1446 |
|
| 1447 |
return true; |
| 1448 |
} |
| 1449 |
|
| 1450 |
|
| 1451 |
/* |
| 1452 |
* Functions for scanning the pending list |
| 1453 |
*/ |
| 1454 |
|
| 1455 |
|
| 1456 |
/* |
| 1457 |
* Get ItemPointer of next heap row to be checked from pending list. |
| 1458 |
* Returns false if there are no more. On pages with several heap rows |
| 1459 |
* it returns each row separately, on page with part of heap row returns |
| 1460 |
* per page data. pos->firstOffset and pos->lastOffset are set to identify |
| 1461 |
* the range of pending-list tuples belonging to this heap row. |
| 1462 |
* |
| 1463 |
* The pendingBuffer is presumed pinned and share-locked on entry, and is |
| 1464 |
* pinned and share-locked on success exit. On failure exit it's released. |
| 1465 |
*/ |
| 1466 |
static bool |
| 1467 |
scanGetCandidate(IndexScanDesc scan, pendingPosition *pos) |
| 1468 |
{ |
| 1469 |
OffsetNumber maxoff; |
| 1470 |
Page page; |
| 1471 |
IndexTuple itup; |
| 1472 |
|
| 1473 |
ItemPointerSetInvalid(&pos->item); |
| 1474 |
for (;;) |
| 1475 |
{ |
| 1476 |
page = BufferGetPage(pos->pendingBuffer); |
| 1477 |
|
| 1478 |
maxoff = PageGetMaxOffsetNumber(page); |
| 1479 |
if (pos->firstOffset > maxoff) |
| 1480 |
{ |
| 1481 |
BlockNumber blkno = GinPageGetOpaque(page)->rightlink; |
| 1482 |
|
| 1483 |
if (blkno == InvalidBlockNumber) |
| 1484 |
{ |
| 1485 |
UnlockReleaseBuffer(pos->pendingBuffer); |
| 1486 |
pos->pendingBuffer = InvalidBuffer; |
| 1487 |
|
| 1488 |
return false; |
| 1489 |
} |
| 1490 |
else |
| 1491 |
{ |
| 1492 |
/* |
| 1493 |
* Here we must prevent deletion of next page by insertcleanup |
| 1494 |
* process, which may be trying to obtain exclusive lock on |
| 1495 |
* current page. So, we lock next page before releasing the |
| 1496 |
* current one |
| 1497 |
*/ |
| 1498 |
Buffer tmpbuf = ReadBuffer(scan->indexRelation, blkno); |
| 1499 |
|
| 1500 |
LockBuffer(tmpbuf, GIN_SHARE); |
| 1501 |
UnlockReleaseBuffer(pos->pendingBuffer); |
| 1502 |
|
| 1503 |
pos->pendingBuffer = tmpbuf; |
| 1504 |
pos->firstOffset = FirstOffsetNumber; |
| 1505 |
} |
| 1506 |
} |
| 1507 |
else |
| 1508 |
{ |
| 1509 |
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, pos->firstOffset)); |
| 1510 |
pos->item = itup->t_tid; |
| 1511 |
if (GinPageHasFullRow(page)) |
| 1512 |
{ |
| 1513 |
/* |
| 1514 |
* find itempointer to the next row |
| 1515 |
*/ |
| 1516 |
for (pos->lastOffset = pos->firstOffset + 1; pos->lastOffset <= maxoff; pos->lastOffset++) |
| 1517 |
{ |
| 1518 |
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, pos->lastOffset)); |
| 1519 |
if (!ItemPointerEquals(&pos->item, &itup->t_tid)) |
| 1520 |
break; |
| 1521 |
} |
| 1522 |
} |
| 1523 |
else |
| 1524 |
{ |
| 1525 |
/* |
| 1526 |
* All itempointers are the same on this page |
| 1527 |
*/ |
| 1528 |
pos->lastOffset = maxoff + 1; |
| 1529 |
} |
| 1530 |
|
| 1531 |
/* |
| 1532 |
* Now pos->firstOffset points to the first tuple of current heap |
| 1533 |
* row, pos->lastOffset points to the first tuple of next heap row |
| 1534 |
* (or to the end of page) |
| 1535 |
*/ |
| 1536 |
break; |
| 1537 |
} |
| 1538 |
} |
| 1539 |
|
| 1540 |
return true; |
| 1541 |
} |
| 1542 |
|
| 1543 |
/* |
| 1544 |
* Scan pending-list page from current tuple (off) up till the first of: |
| 1545 |
* - match is found (then returns true) |
| 1546 |
* - no later match is possible |
| 1547 |
* - tuple's attribute number is not equal to entry's attrnum |
| 1548 |
* - reach end of page |
| 1549 |
* |
| 1550 |
* datum[]/category[]/datumExtracted[] arrays are used to cache the results |
| 1551 |
* of gintuple_get_key() on the current page. |
| 1552 |
*/ |
| 1553 |
static bool |
| 1554 |
matchPartialInPendingList(GinState *ginstate, Page page, |
| 1555 |
OffsetNumber off, OffsetNumber maxoff, |
| 1556 |
GinScanEntry entry, |
| 1557 |
Datum *datum, GinNullCategory *category, |
| 1558 |
bool *datumExtracted) |
| 1559 |
{ |
| 1560 |
IndexTuple itup; |
| 1561 |
int32 cmp; |
| 1562 |
|
| 1563 |
/* Partial match to a null is not possible */ |
| 1564 |
if (entry->queryCategory != GIN_CAT_NORM_KEY) |
| 1565 |
return false; |
| 1566 |
|
| 1567 |
while (off < maxoff) |
| 1568 |
{ |
| 1569 |
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, off)); |
| 1570 |
|
| 1571 |
if (gintuple_get_attrnum(ginstate, itup) != entry->attnum) |
| 1572 |
return false; |
| 1573 |
|
| 1574 |
if (datumExtracted[off - 1] == false) |
| 1575 |
{ |
| 1576 |
datum[off - 1] = gintuple_get_key(ginstate, itup, |
| 1577 |
&category[off - 1]); |
| 1578 |
datumExtracted[off - 1] = true; |
| 1579 |
} |
| 1580 |
|
| 1581 |
/* Once we hit nulls, no further match is possible */ |
| 1582 |
if (category[off - 1] != GIN_CAT_NORM_KEY) |
| 1583 |
return false; |
| 1584 |
|
| 1585 |
/*---------- |
| 1586 |
* Check partial match. |
| 1587 |
* case cmp == 0 => match |
| 1588 |
* case cmp > 0 => not match and end scan (no later match possible) |
| 1589 |
* case cmp < 0 => not match and continue scan |
| 1590 |
*---------- |
| 1591 |
*/ |
| 1592 |
cmp = DatumGetInt32(FunctionCall4Coll(&ginstate->comparePartialFn[entry->attnum - 1], |
| 1593 |
ginstate->supportCollation[entry->attnum - 1], |
| 1594 |
entry->queryKey, |
| 1595 |
datum[off - 1], |
| 1596 |
UInt16GetDatum(entry->strategy), |
| 1597 |
PointerGetDatum(entry->extra_data))); |
| 1598 |
if (cmp == 0) |
| 1599 |
return true; |
| 1600 |
else if (cmp > 0) |
| 1601 |
return false; |
| 1602 |
|
| 1603 |
off++; |
| 1604 |
} |
| 1605 |
|
| 1606 |
return false; |
| 1607 |
} |
| 1608 |
|
| 1609 |
/* |
| 1610 |
* Set up the entryRes array for each key by looking at |
| 1611 |
* every entry for current heap row in pending list. |
| 1612 |
* |
| 1613 |
* Returns true if each scan key has at least one entryRes match. |
| 1614 |
* This corresponds to the situations where the normal index search will |
| 1615 |
* try to apply the key's consistentFn. (A tuple not meeting that requirement |
| 1616 |
* cannot be returned by the normal search since no entry stream will |
| 1617 |
* source its TID.) |
| 1618 |
* |
| 1619 |
* The pendingBuffer is presumed pinned and share-locked on entry. |
| 1620 |
*/ |
| 1621 |
static bool |
| 1622 |
collectMatchesForHeapRow(IndexScanDesc scan, pendingPosition *pos) |
| 1623 |
{ |
| 1624 |
GinScanOpaque so = (GinScanOpaque) scan->opaque; |
| 1625 |
OffsetNumber attrnum; |
| 1626 |
Page page; |
| 1627 |
IndexTuple itup; |
| 1628 |
int i, |
| 1629 |
j; |
| 1630 |
|
| 1631 |
/* |
| 1632 |
* Reset all entryRes and hasMatchKey flags |
| 1633 |
*/ |
| 1634 |
for (i = 0; i < so->nkeys; i++) |
| 1635 |
{ |
| 1636 |
GinScanKey key = so->keys + i; |
| 1637 |
|
| 1638 |
memset(key->entryRes, GIN_FALSE, key->nentries); |
| 1639 |
} |
| 1640 |
memset(pos->hasMatchKey, false, so->nkeys); |
| 1641 |
|
| 1642 |
/* |
| 1643 |
* Outer loop iterates over multiple pending-list pages when a single heap |
| 1644 |
* row has entries spanning those pages. |
| 1645 |
*/ |
| 1646 |
for (;;) |
| 1647 |
{ |
| 1648 |
Datum datum[BLCKSZ / sizeof(IndexTupleData)]; |
| 1649 |
GinNullCategory category[BLCKSZ / sizeof(IndexTupleData)]; |
| 1650 |
bool datumExtracted[BLCKSZ / sizeof(IndexTupleData)]; |
| 1651 |
|
| 1652 |
Assert(pos->lastOffset > pos->firstOffset); |
| 1653 |
memset(datumExtracted + pos->firstOffset - 1, 0, |
| 1654 |
sizeof(bool) * (pos->lastOffset - pos->firstOffset)); |
| 1655 |
|
| 1656 |
page = BufferGetPage(pos->pendingBuffer); |
| 1657 |
|
| 1658 |
for (i = 0; i < so->nkeys; i++) |
| 1659 |
{ |
| 1660 |
GinScanKey key = so->keys + i; |
| 1661 |
|
| 1662 |
for (j = 0; j < key->nentries; j++) |
| 1663 |
{ |
| 1664 |
GinScanEntry entry = key->scanEntry[j]; |
| 1665 |
OffsetNumber StopLow = pos->firstOffset, |
| 1666 |
StopHigh = pos->lastOffset, |
| 1667 |
StopMiddle; |
| 1668 |
|
| 1669 |
/* If already matched on earlier page, do no extra work */ |
| 1670 |
if (key->entryRes[j]) |
| 1671 |
continue; |
| 1672 |
|
| 1673 |
/* |
| 1674 |
* Interesting tuples are from pos->firstOffset to |
| 1675 |
* pos->lastOffset and they are ordered by (attnum, Datum) as |
| 1676 |
* it's done in entry tree. So we can use binary search to |
| 1677 |
* avoid linear scanning. |
| 1678 |
*/ |
| 1679 |
while (StopLow < StopHigh) |
| 1680 |
{ |
| 1681 |
int res; |
| 1682 |
|
| 1683 |
StopMiddle = StopLow + ((StopHigh - StopLow) >> 1); |
| 1684 |
|
| 1685 |
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, StopMiddle)); |
| 1686 |
|
| 1687 |
attrnum = gintuple_get_attrnum(&so->ginstate, itup); |
| 1688 |
|
| 1689 |
if (key->attnum < attrnum) |
| 1690 |
{ |
| 1691 |
StopHigh = StopMiddle; |
| 1692 |
continue; |
| 1693 |
} |
| 1694 |
if (key->attnum > attrnum) |
| 1695 |
{ |
| 1696 |
StopLow = StopMiddle + 1; |
| 1697 |
continue; |
| 1698 |
} |
| 1699 |
|
| 1700 |
if (datumExtracted[StopMiddle - 1] == false) |
| 1701 |
{ |
| 1702 |
datum[StopMiddle - 1] = |
| 1703 |
gintuple_get_key(&so->ginstate, itup, |
| 1704 |
&category[StopMiddle - 1]); |
| 1705 |
datumExtracted[StopMiddle - 1] = true; |
| 1706 |
} |
| 1707 |
|
| 1708 |
if (entry->queryCategory == GIN_CAT_EMPTY_QUERY) |
| 1709 |
{ |
| 1710 |
/* special behavior depending on searchMode */ |
| 1711 |
if (entry->searchMode == GIN_SEARCH_MODE_ALL) |
| 1712 |
{ |
| 1713 |
/* match anything except NULL_ITEM */ |
| 1714 |
if (category[StopMiddle - 1] == GIN_CAT_NULL_ITEM) |
| 1715 |
res = -1; |
| 1716 |
else |
| 1717 |
res = 0; |
| 1718 |
} |
| 1719 |
else |
| 1720 |
{ |
| 1721 |
/* match everything */ |
| 1722 |
res = 0; |
| 1723 |
} |
| 1724 |
} |
| 1725 |
else |
| 1726 |
{ |
| 1727 |
res = ginCompareEntries(&so->ginstate, |
| 1728 |
entry->attnum, |
| 1729 |
entry->queryKey, |
| 1730 |
entry->queryCategory, |
| 1731 |
datum[StopMiddle - 1], |
| 1732 |
category[StopMiddle - 1]); |
| 1733 |
} |
| 1734 |
|
| 1735 |
if (res == 0) |
| 1736 |
{ |
| 1737 |
/* |
| 1738 |
* Found exact match (there can be only one, except in |
| 1739 |
* EMPTY_QUERY mode). |
| 1740 |
* |
| 1741 |
* If doing partial match, scan forward from here to |
| 1742 |
* end of page to check for matches. |
| 1743 |
* |
| 1744 |
* See comment above about tuple's ordering. |
| 1745 |
*/ |
| 1746 |
if (entry->isPartialMatch) |
| 1747 |
key->entryRes[j] = |
| 1748 |
matchPartialInPendingList(&so->ginstate, |
| 1749 |
page, |
| 1750 |
StopMiddle, |
| 1751 |
pos->lastOffset, |
| 1752 |
entry, |
| 1753 |
datum, |
| 1754 |
category, |
| 1755 |
datumExtracted); |
| 1756 |
else |
| 1757 |
key->entryRes[j] = true; |
| 1758 |
|
| 1759 |
/* done with binary search */ |
| 1760 |
break; |
| 1761 |
} |
| 1762 |
else if (res < 0) |
| 1763 |
StopHigh = StopMiddle; |
| 1764 |
else |
| 1765 |
StopLow = StopMiddle + 1; |
| 1766 |
} |
| 1767 |
|
| 1768 |
if (StopLow >= StopHigh && entry->isPartialMatch) |
| 1769 |
{ |
| 1770 |
/* |
| 1771 |
* No exact match on this page. If doing partial match, |
| 1772 |
* scan from the first tuple greater than target value to |
| 1773 |
* end of page. Note that since we don't remember whether |
| 1774 |
* the comparePartialFn told us to stop early on a |
| 1775 |
* previous page, we will uselessly apply comparePartialFn |
| 1776 |
* to the first tuple on each subsequent page. |
| 1777 |
*/ |
| 1778 |
key->entryRes[j] = |
| 1779 |
matchPartialInPendingList(&so->ginstate, |
| 1780 |
page, |
| 1781 |
StopHigh, |
| 1782 |
pos->lastOffset, |
| 1783 |
entry, |
| 1784 |
datum, |
| 1785 |
category, |
| 1786 |
datumExtracted); |
| 1787 |
} |
| 1788 |
|
| 1789 |
pos->hasMatchKey[i] |= key->entryRes[j]; |
| 1790 |
} |
| 1791 |
} |
| 1792 |
|
| 1793 |
/* Advance firstOffset over the scanned tuples */ |
| 1794 |
pos->firstOffset = pos->lastOffset; |
| 1795 |
|
| 1796 |
if (GinPageHasFullRow(page)) |
| 1797 |
{ |
| 1798 |
/* |
| 1799 |
* We have examined all pending entries for the current heap row. |
| 1800 |
* Break out of loop over pages. |
| 1801 |
*/ |
| 1802 |
break; |
| 1803 |
} |
| 1804 |
else |
| 1805 |
{ |
| 1806 |
/* |
| 1807 |
* Advance to next page of pending entries for the current heap |
| 1808 |
* row. Complain if there isn't one. |
| 1809 |
*/ |
| 1810 |
ItemPointerData item = pos->item; |
| 1811 |
|
| 1812 |
if (scanGetCandidate(scan, pos) == false || |
| 1813 |
!ItemPointerEquals(&pos->item, &item)) |
| 1814 |
elog(ERROR, "could not find additional pending pages for same heap tuple"); |
| 1815 |
} |
| 1816 |
} |
| 1817 |
|
| 1818 |
/* |
| 1819 |
* All scan keys except excludeOnly require at least one entry to match. |
| 1820 |
* excludeOnly keys are an exception, because their implied |
| 1821 |
* GIN_CAT_EMPTY_QUERY scanEntry always matches. So return "true" if all |
| 1822 |
* non-excludeOnly scan keys have at least one match. |
| 1823 |
*/ |
| 1824 |
for (i = 0; i < so->nkeys; i++) |
| 1825 |
{ |
| 1826 |
if (pos->hasMatchKey[i] == false && !so->keys[i].excludeOnly) |
| 1827 |
return false; |
| 1828 |
} |
| 1829 |
|
| 1830 |
return true; |
| 1831 |
} |
| 1832 |
|
| 1833 |
/* |
| 1834 |
* Collect all matched rows from pending list into bitmap. |
| 1835 |
*/ |
| 1836 |
static void |
| 1837 |
scanPendingInsert(IndexScanDesc scan, TIDBitmap *tbm, int64 *ntids) |
| 1838 |
{ |
| 1839 |
GinScanOpaque so = (GinScanOpaque) scan->opaque; |
| 1840 |
MemoryContext oldCtx; |
| 1841 |
bool recheck, |
| 1842 |
match; |
| 1843 |
int i; |
| 1844 |
pendingPosition pos; |
| 1845 |
Buffer metabuffer = ReadBuffer(scan->indexRelation, GIN_METAPAGE_BLKNO); |
| 1846 |
Page page; |
| 1847 |
BlockNumber blkno; |
| 1848 |
|
| 1849 |
*ntids = 0; |
| 1850 |
|
| 1851 |
/* |
| 1852 |
* Acquire predicate lock on the metapage, to conflict with any fastupdate |
| 1853 |
* insertions. |
| 1854 |
*/ |
| 1855 |
PredicateLockPage(scan->indexRelation, GIN_METAPAGE_BLKNO, scan->xs_snapshot); |
| 1856 |
|
| 1857 |
LockBuffer(metabuffer, GIN_SHARE); |
| 1858 |
page = BufferGetPage(metabuffer); |
| 1859 |
blkno = GinPageGetMeta(page)->head; |
| 1860 |
|
| 1861 |
/* |
| 1862 |
* fetch head of list before unlocking metapage. head page must be pinned |
| 1863 |
* to prevent deletion by vacuum process |
| 1864 |
*/ |
| 1865 |
if (blkno == InvalidBlockNumber) |
| 1866 |
{ |
| 1867 |
/* No pending list, so proceed with normal scan */ |
| 1868 |
UnlockReleaseBuffer(metabuffer); |
| 1869 |
return; |
| 1870 |
} |
| 1871 |
|
| 1872 |
pos.pendingBuffer = ReadBuffer(scan->indexRelation, blkno); |
| 1873 |
LockBuffer(pos.pendingBuffer, GIN_SHARE); |
| 1874 |
pos.firstOffset = FirstOffsetNumber; |
| 1875 |
UnlockReleaseBuffer(metabuffer); |
| 1876 |
pos.hasMatchKey = palloc(sizeof(bool) * so->nkeys); |
| 1877 |
|
| 1878 |
/* |
| 1879 |
* loop for each heap row. scanGetCandidate returns full row or row's |
| 1880 |
* tuples from first page. |
| 1881 |
*/ |
| 1882 |
while (scanGetCandidate(scan, &pos)) |
| 1883 |
{ |
| 1884 |
/* |
| 1885 |
* Check entries in tuple and set up entryRes array. |
| 1886 |
* |
| 1887 |
* If pending tuples belonging to the current heap row are spread |
| 1888 |
* across several pages, collectMatchesForHeapRow will read all of |
| 1889 |
* those pages. |
| 1890 |
*/ |
| 1891 |
if (!collectMatchesForHeapRow(scan, &pos)) |
| 1892 |
continue; |
| 1893 |
|
| 1894 |
/* |
| 1895 |
* Matching of entries of one row is finished, so check row using |
| 1896 |
* consistent functions. |
| 1897 |
*/ |
| 1898 |
oldCtx = MemoryContextSwitchTo(so->tempCtx); |
| 1899 |
recheck = false; |
| 1900 |
match = true; |
| 1901 |
|
| 1902 |
for (i = 0; i < so->nkeys; i++) |
| 1903 |
{ |
| 1904 |
GinScanKey key = so->keys + i; |
| 1905 |
|
| 1906 |
if (!key->boolConsistentFn(key)) |
| 1907 |
{ |
| 1908 |
match = false; |
| 1909 |
break; |
| 1910 |
} |
| 1911 |
recheck |= key->recheckCurItem; |
| 1912 |
} |
| 1913 |
|
| 1914 |
MemoryContextSwitchTo(oldCtx); |
| 1915 |
MemoryContextReset(so->tempCtx); |
| 1916 |
|
| 1917 |
if (match) |
| 1918 |
{ |
| 1919 |
tbm_add_tuples(tbm, &pos.item, 1, recheck); |
| 1920 |
(*ntids)++; |
| 1921 |
} |
| 1922 |
} |
| 1923 |
|
| 1924 |
pfree(pos.hasMatchKey); |
| 1925 |
} |
| 1926 |
|
| 1927 |
|
| 1928 |
#define GinIsVoidRes(s) ( ((GinScanOpaque) scan->opaque)->isVoidRes ) |
| 1929 |
|
| 1930 |
int64 |
| 1931 |
gingetbitmap(IndexScanDesc scan, TIDBitmap *tbm) |
| 1932 |
{ |
| 1933 |
GinScanOpaque so = (GinScanOpaque) scan->opaque; |
| 1934 |
int64 ntids; |
| 1935 |
ItemPointerData iptr; |
| 1936 |
bool recheck; |
| 1937 |
|
| 1938 |
/* |
| 1939 |
* Set up the scan keys, and check for unsatisfiable query. |
| 1940 |
*/ |
| 1941 |
ginFreeScanKeys(so); /* there should be no keys yet, but just to be |
| 1942 |
* sure */ |
| 1943 |
ginNewScanKey(scan); |
| 1944 |
|
| 1945 |
if (GinIsVoidRes(scan)) |
| 1946 |
return 0; |
| 1947 |
|
| 1948 |
ntids = 0; |
| 1949 |
|
| 1950 |
/* |
| 1951 |
* First, scan the pending list and collect any matching entries into the |
| 1952 |
* bitmap. After we scan a pending item, some other backend could post it |
| 1953 |
* into the main index, and so we might visit it a second time during the |
| 1954 |
* main scan. This is okay because we'll just re-set the same bit in the |
| 1955 |
* bitmap. (The possibility of duplicate visits is a major reason why GIN |
| 1956 |
* can't support the amgettuple API, however.) Note that it would not do |
| 1957 |
* to scan the main index before the pending list, since concurrent |
| 1958 |
* cleanup could then make us miss entries entirely. |
| 1959 |
*/ |
| 1960 |
scanPendingInsert(scan, tbm, &ntids); |
| 1961 |
|
| 1962 |
/* |
| 1963 |
* Now scan the main index. |
| 1964 |
*/ |
| 1965 |
startScan(scan); |
| 1966 |
|
| 1967 |
ItemPointerSetMin(&iptr); |
| 1968 |
|
| 1969 |
for (;;) |
| 1970 |
{ |
| 1971 |
if (!scanGetItem(scan, iptr, &iptr, &recheck)) |
| 1972 |
break; |
| 1973 |
|
| 1974 |
if (ItemPointerIsLossyPage(&iptr)) |
| 1975 |
tbm_add_page(tbm, ItemPointerGetBlockNumber(&iptr)); |
| 1976 |
else |
| 1977 |
tbm_add_tuples(tbm, &iptr, 1, recheck); |
| 1978 |
ntids++; |
| 1979 |
} |
| 1980 |
|
| 1981 |
return ntids; |
| 1982 |
} |