Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * gistbuild.c
4 : * build algorithm for GiST indexes implementation.
5 : *
6 : * There are two different strategies:
7 : *
8 : * 1. Sort all input tuples, pack them into GiST leaf pages in the sorted
9 : * order, and create downlinks and internal pages as we go. This builds
10 : * the index from the bottom up, similar to how B-tree index build
11 : * works.
12 : *
13 : * 2. Start with an empty index, and insert all tuples one by one.
14 : *
15 : * The sorted method is used if the operator classes for all columns have
16 : * a 'sortsupport' defined. Otherwise, we resort to the second strategy.
17 : *
18 : * The second strategy can optionally use buffers at different levels of
19 : * the tree to reduce I/O, see "Buffering build algorithm" in the README
20 : * for a more detailed explanation. It initially calls insert over and
21 : * over, but switches to the buffered algorithm after a certain number of
22 : * tuples (unless buffering mode is disabled).
23 : *
24 : *
25 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
26 : * Portions Copyright (c) 1994, Regents of the University of California
27 : *
28 : * IDENTIFICATION
29 : * src/backend/access/gist/gistbuild.c
30 : *
31 : *-------------------------------------------------------------------------
32 : */
33 : #include "postgres.h"
34 :
35 : #include <math.h>
36 :
37 : #include "access/genam.h"
38 : #include "access/gist_private.h"
39 : #include "access/tableam.h"
40 : #include "access/xloginsert.h"
41 : #include "miscadmin.h"
42 : #include "nodes/execnodes.h"
43 : #include "optimizer/optimizer.h"
44 : #include "storage/bufmgr.h"
45 : #include "storage/bulk_write.h"
46 :
47 : #include "utils/memutils.h"
48 : #include "utils/rel.h"
49 : #include "utils/tuplesort.h"
50 :
51 : /* Step of index tuples for check whether to switch to buffering build mode */
52 : #define BUFFERING_MODE_SWITCH_CHECK_STEP 256
53 :
54 : /*
55 : * Number of tuples to process in the slow way before switching to buffering
56 : * mode, when buffering is explicitly turned on. Also, the number of tuples
57 : * to process between readjusting the buffer size parameter, while in
58 : * buffering mode.
59 : */
60 : #define BUFFERING_MODE_TUPLE_SIZE_STATS_TARGET 4096
61 :
62 : /*
63 : * Strategy used to build the index. It can change between the
64 : * GIST_BUFFERING_* modes on the fly, but if the Sorted method is used,
65 : * that needs to be decided up-front and cannot be changed afterwards.
66 : */
67 : typedef enum
68 : {
69 : GIST_SORTED_BUILD, /* bottom-up build by sorting */
70 : GIST_BUFFERING_DISABLED, /* in regular build mode and aren't going to
71 : * switch */
72 : GIST_BUFFERING_AUTO, /* in regular build mode, but will switch to
73 : * buffering build mode if the index grows too
74 : * big */
75 : GIST_BUFFERING_STATS, /* gathering statistics of index tuple size
76 : * before switching to the buffering build
77 : * mode */
78 : GIST_BUFFERING_ACTIVE, /* in buffering build mode */
79 : } GistBuildMode;
80 :
81 : /* Working state for gistbuild and its callback */
82 : typedef struct
83 : {
84 : Relation indexrel;
85 : Relation heaprel;
86 : GISTSTATE *giststate;
87 :
88 : Size freespace; /* amount of free space to leave on pages */
89 :
90 : GistBuildMode buildMode;
91 :
92 : int64 indtuples; /* number of tuples indexed */
93 :
94 : /*
95 : * Extra data structures used during a buffering build. 'gfbb' contains
96 : * information related to managing the build buffers. 'parentMap' is a
97 : * lookup table of the parent of each internal page.
98 : */
99 : int64 indtuplesSize; /* total size of all indexed tuples */
100 : GISTBuildBuffers *gfbb;
101 : HTAB *parentMap;
102 :
103 : /*
104 : * Extra data structures used during a sorting build.
105 : */
106 : Tuplesortstate *sortstate; /* state data for tuplesort.c */
107 :
108 : BlockNumber pages_allocated;
109 :
110 : BulkWriteState *bulkstate;
111 : } GISTBuildState;
112 :
113 : #define GIST_SORTED_BUILD_PAGE_NUM 4
114 :
115 : /*
116 : * In sorted build, we use a stack of these structs, one for each level,
117 : * to hold an in-memory buffer of last pages at the level.
118 : *
119 : * Sorting GiST build requires good linearization of the sort opclass. This is
120 : * not always the case in multidimensional data. To tackle the anomalies, we
121 : * buffer index tuples and apply picksplit that can be multidimension-aware.
122 : */
123 : typedef struct GistSortedBuildLevelState
124 : {
125 : int current_page;
126 : BlockNumber last_blkno;
127 : struct GistSortedBuildLevelState *parent; /* Upper level, if any */
128 : Page pages[GIST_SORTED_BUILD_PAGE_NUM];
129 : } GistSortedBuildLevelState;
130 :
131 : /* prototypes for private functions */
132 :
133 : static void gistSortedBuildCallback(Relation index, ItemPointer tid,
134 : Datum *values, bool *isnull,
135 : bool tupleIsAlive, void *state);
136 : static void gist_indexsortbuild(GISTBuildState *state);
137 : static void gist_indexsortbuild_levelstate_add(GISTBuildState *state,
138 : GistSortedBuildLevelState *levelstate,
139 : IndexTuple itup);
140 : static void gist_indexsortbuild_levelstate_flush(GISTBuildState *state,
141 : GistSortedBuildLevelState *levelstate);
142 :
143 : static void gistInitBuffering(GISTBuildState *buildstate);
144 : static int calculatePagesPerBuffer(GISTBuildState *buildstate, int levelStep);
145 : static void gistBuildCallback(Relation index,
146 : ItemPointer tid,
147 : Datum *values,
148 : bool *isnull,
149 : bool tupleIsAlive,
150 : void *state);
151 : static void gistBufferingBuildInsert(GISTBuildState *buildstate,
152 : IndexTuple itup);
153 : static bool gistProcessItup(GISTBuildState *buildstate, IndexTuple itup,
154 : BlockNumber startblkno, int startlevel);
155 : static BlockNumber gistbufferinginserttuples(GISTBuildState *buildstate,
156 : Buffer buffer, int level,
157 : IndexTuple *itup, int ntup, OffsetNumber oldoffnum,
158 : BlockNumber parentblk, OffsetNumber downlinkoffnum);
159 : static Buffer gistBufferingFindCorrectParent(GISTBuildState *buildstate,
160 : BlockNumber childblkno, int level,
161 : BlockNumber *parentblkno,
162 : OffsetNumber *downlinkoffnum);
163 : static void gistProcessEmptyingQueue(GISTBuildState *buildstate);
164 : static void gistEmptyAllBuffers(GISTBuildState *buildstate);
165 : static int gistGetMaxLevel(Relation index);
166 :
167 : static void gistInitParentMap(GISTBuildState *buildstate);
168 : static void gistMemorizeParent(GISTBuildState *buildstate, BlockNumber child,
169 : BlockNumber parent);
170 : static void gistMemorizeAllDownlinks(GISTBuildState *buildstate,
171 : Buffer parentbuf);
172 : static BlockNumber gistGetParent(GISTBuildState *buildstate, BlockNumber child);
173 :
174 :
175 : /*
176 : * Main entry point to GiST index build.
177 : */
178 : IndexBuildResult *
179 1554 : gistbuild(Relation heap, Relation index, IndexInfo *indexInfo)
180 : {
181 : IndexBuildResult *result;
182 : double reltuples;
183 : GISTBuildState buildstate;
184 1554 : MemoryContext oldcxt = CurrentMemoryContext;
185 : int fillfactor;
186 : Oid SortSupportFnOids[INDEX_MAX_KEYS];
187 1554 : GiSTOptions *options = (GiSTOptions *) index->rd_options;
188 :
189 : /*
190 : * We expect to be called exactly once for any index relation. If that's
191 : * not the case, big trouble's what we have.
192 : */
193 1554 : if (RelationGetNumberOfBlocks(index) != 0)
194 0 : elog(ERROR, "index \"%s\" already contains data",
195 : RelationGetRelationName(index));
196 :
197 1554 : buildstate.indexrel = index;
198 1554 : buildstate.heaprel = heap;
199 1554 : buildstate.sortstate = NULL;
200 1554 : buildstate.giststate = initGISTstate(index);
201 :
202 : /*
203 : * Create a temporary memory context that is reset once for each tuple
204 : * processed. (Note: we don't bother to make this a child of the
205 : * giststate's scanCxt, so we have to delete it separately at the end.)
206 : */
207 1554 : buildstate.giststate->tempCxt = createTempGistContext();
208 :
209 : /*
210 : * Choose build strategy. First check whether the user specified to use
211 : * buffering mode. (The use-case for that in the field is somewhat
212 : * questionable perhaps, but it's important for testing purposes.)
213 : */
214 1554 : if (options)
215 : {
216 32 : if (options->buffering_mode == GIST_OPTION_BUFFERING_ON)
217 12 : buildstate.buildMode = GIST_BUFFERING_STATS;
218 20 : else if (options->buffering_mode == GIST_OPTION_BUFFERING_OFF)
219 6 : buildstate.buildMode = GIST_BUFFERING_DISABLED;
220 : else /* must be "auto" */
221 14 : buildstate.buildMode = GIST_BUFFERING_AUTO;
222 : }
223 : else
224 : {
225 1522 : buildstate.buildMode = GIST_BUFFERING_AUTO;
226 : }
227 :
228 : /*
229 : * Unless buffering mode was forced, see if we can use sorting instead.
230 : */
231 1554 : if (buildstate.buildMode != GIST_BUFFERING_STATS)
232 : {
233 1542 : bool hasallsortsupports = true;
234 1542 : int keyscount = IndexRelationGetNumberOfKeyAttributes(index);
235 :
236 1698 : for (int i = 0; i < keyscount; i++)
237 : {
238 1548 : SortSupportFnOids[i] = index_getprocid(index, i + 1,
239 : GIST_SORTSUPPORT_PROC);
240 1548 : if (!OidIsValid(SortSupportFnOids[i]))
241 : {
242 1392 : hasallsortsupports = false;
243 1392 : break;
244 : }
245 : }
246 1542 : if (hasallsortsupports)
247 150 : buildstate.buildMode = GIST_SORTED_BUILD;
248 : }
249 :
250 : /*
251 : * Calculate target amount of free space to leave on pages.
252 : */
253 1554 : fillfactor = options ? options->fillfactor : GIST_DEFAULT_FILLFACTOR;
254 1554 : buildstate.freespace = BLCKSZ * (100 - fillfactor) / 100;
255 :
256 : /*
257 : * Build the index using the chosen strategy.
258 : */
259 1554 : buildstate.indtuples = 0;
260 1554 : buildstate.indtuplesSize = 0;
261 :
262 1554 : if (buildstate.buildMode == GIST_SORTED_BUILD)
263 : {
264 : /*
265 : * Sort all data, build the index from bottom up.
266 : */
267 150 : buildstate.sortstate = tuplesort_begin_index_gist(heap,
268 : index,
269 : maintenance_work_mem,
270 : NULL,
271 : TUPLESORT_NONE);
272 :
273 : /* Scan the table, adding all tuples to the tuplesort */
274 150 : reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
275 : gistSortedBuildCallback,
276 : (void *) &buildstate, NULL);
277 :
278 : /*
279 : * Perform the sort and build index pages.
280 : */
281 150 : tuplesort_performsort(buildstate.sortstate);
282 :
283 150 : gist_indexsortbuild(&buildstate);
284 :
285 150 : tuplesort_end(buildstate.sortstate);
286 : }
287 : else
288 : {
289 : /*
290 : * Initialize an empty index and insert all tuples, possibly using
291 : * buffers on intermediate levels.
292 : */
293 : Buffer buffer;
294 : Page page;
295 :
296 : /* initialize the root page */
297 1404 : buffer = gistNewBuffer(index, heap);
298 : Assert(BufferGetBlockNumber(buffer) == GIST_ROOT_BLKNO);
299 1404 : page = BufferGetPage(buffer);
300 :
301 1404 : START_CRIT_SECTION();
302 :
303 1404 : GISTInitBuffer(buffer, F_LEAF);
304 :
305 1404 : MarkBufferDirty(buffer);
306 1404 : PageSetLSN(page, GistBuildLSN);
307 :
308 1404 : UnlockReleaseBuffer(buffer);
309 :
310 1404 : END_CRIT_SECTION();
311 :
312 : /* Scan the table, inserting all the tuples to the index. */
313 1404 : reltuples = table_index_build_scan(heap, index, indexInfo, true, true,
314 : gistBuildCallback,
315 : (void *) &buildstate, NULL);
316 :
317 : /*
318 : * If buffering was used, flush out all the tuples that are still in
319 : * the buffers.
320 : */
321 1404 : if (buildstate.buildMode == GIST_BUFFERING_ACTIVE)
322 : {
323 6 : elog(DEBUG1, "all tuples processed, emptying buffers");
324 6 : gistEmptyAllBuffers(&buildstate);
325 6 : gistFreeBuildBuffers(buildstate.gfbb);
326 : }
327 :
328 : /*
329 : * We didn't write WAL records as we built the index, so if
330 : * WAL-logging is required, write all pages to the WAL now.
331 : */
332 1404 : if (RelationNeedsWAL(index))
333 : {
334 892 : log_newpage_range(index, MAIN_FORKNUM,
335 : 0, RelationGetNumberOfBlocks(index),
336 : true);
337 : }
338 : }
339 :
340 : /* okay, all heap tuples are indexed */
341 1554 : MemoryContextSwitchTo(oldcxt);
342 1554 : MemoryContextDelete(buildstate.giststate->tempCxt);
343 :
344 1554 : freeGISTstate(buildstate.giststate);
345 :
346 : /*
347 : * Return statistics
348 : */
349 1554 : result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
350 :
351 1554 : result->heap_tuples = reltuples;
352 1554 : result->index_tuples = (double) buildstate.indtuples;
353 :
354 1554 : return result;
355 : }
356 :
357 : /*-------------------------------------------------------------------------
358 : * Routines for sorted build
359 : *-------------------------------------------------------------------------
360 : */
361 :
362 : /*
363 : * Per-tuple callback for table_index_build_scan.
364 : */
365 : static void
366 196144 : gistSortedBuildCallback(Relation index,
367 : ItemPointer tid,
368 : Datum *values,
369 : bool *isnull,
370 : bool tupleIsAlive,
371 : void *state)
372 : {
373 196144 : GISTBuildState *buildstate = (GISTBuildState *) state;
374 : MemoryContext oldCtx;
375 : Datum compressed_values[INDEX_MAX_KEYS];
376 :
377 196144 : oldCtx = MemoryContextSwitchTo(buildstate->giststate->tempCxt);
378 :
379 : /* Form an index tuple and point it at the heap tuple */
380 196144 : gistCompressValues(buildstate->giststate, index,
381 : values, isnull,
382 : true, compressed_values);
383 :
384 196144 : tuplesort_putindextuplevalues(buildstate->sortstate,
385 : buildstate->indexrel,
386 : tid,
387 : compressed_values, isnull);
388 :
389 196144 : MemoryContextSwitchTo(oldCtx);
390 196144 : MemoryContextReset(buildstate->giststate->tempCxt);
391 :
392 : /* Update tuple count. */
393 196144 : buildstate->indtuples += 1;
394 196144 : }
395 :
396 : /*
397 : * Build GiST index from bottom up from pre-sorted tuples.
398 : */
399 : static void
400 150 : gist_indexsortbuild(GISTBuildState *state)
401 : {
402 : IndexTuple itup;
403 : GistSortedBuildLevelState *levelstate;
404 : BulkWriteBuffer rootbuf;
405 :
406 : /* Reserve block 0 for the root page */
407 150 : state->pages_allocated = 1;
408 :
409 150 : state->bulkstate = smgr_bulk_start_rel(state->indexrel, MAIN_FORKNUM);
410 :
411 : /* Allocate a temporary buffer for the first leaf page batch. */
412 150 : levelstate = palloc0(sizeof(GistSortedBuildLevelState));
413 150 : levelstate->pages[0] = palloc(BLCKSZ);
414 150 : levelstate->parent = NULL;
415 150 : gistinitpage(levelstate->pages[0], F_LEAF);
416 :
417 : /*
418 : * Fill index pages with tuples in the sorted order.
419 : */
420 196294 : while ((itup = tuplesort_getindextuple(state->sortstate, true)) != NULL)
421 : {
422 196144 : gist_indexsortbuild_levelstate_add(state, levelstate, itup);
423 196144 : MemoryContextReset(state->giststate->tempCxt);
424 : }
425 :
426 : /*
427 : * Write out the partially full non-root pages.
428 : *
429 : * Keep in mind that flush can build a new root. If number of pages is > 1
430 : * then new root is required.
431 : */
432 178 : while (levelstate->parent != NULL || levelstate->current_page != 0)
433 : {
434 : GistSortedBuildLevelState *parent;
435 :
436 28 : gist_indexsortbuild_levelstate_flush(state, levelstate);
437 28 : parent = levelstate->parent;
438 140 : for (int i = 0; i < GIST_SORTED_BUILD_PAGE_NUM; i++)
439 112 : if (levelstate->pages[i])
440 112 : pfree(levelstate->pages[i]);
441 28 : pfree(levelstate);
442 28 : levelstate = parent;
443 : }
444 :
445 : /* Write out the root */
446 150 : PageSetLSN(levelstate->pages[0], GistBuildLSN);
447 150 : rootbuf = smgr_bulk_get_buf(state->bulkstate);
448 150 : memcpy(rootbuf, levelstate->pages[0], BLCKSZ);
449 150 : smgr_bulk_write(state->bulkstate, GIST_ROOT_BLKNO, rootbuf, true);
450 :
451 150 : pfree(levelstate);
452 :
453 150 : smgr_bulk_finish(state->bulkstate);
454 150 : }
455 :
456 : /*
457 : * Add tuple to a page. If the pages are full, write them out and re-initialize
458 : * a new page first.
459 : */
460 : static void
461 197558 : gist_indexsortbuild_levelstate_add(GISTBuildState *state,
462 : GistSortedBuildLevelState *levelstate,
463 : IndexTuple itup)
464 : {
465 : Size sizeNeeded;
466 :
467 : /* Check if tuple can be added to the current page */
468 197558 : sizeNeeded = IndexTupleSize(itup) + sizeof(ItemIdData); /* fillfactor ignored */
469 197558 : if (PageGetFreeSpace(levelstate->pages[levelstate->current_page]) < sizeNeeded)
470 : {
471 : Page newPage;
472 1056 : Page old_page = levelstate->pages[levelstate->current_page];
473 1056 : uint16 old_page_flags = GistPageGetOpaque(old_page)->flags;
474 :
475 1056 : if (levelstate->current_page + 1 == GIST_SORTED_BUILD_PAGE_NUM)
476 : {
477 256 : gist_indexsortbuild_levelstate_flush(state, levelstate);
478 : }
479 : else
480 800 : levelstate->current_page++;
481 :
482 1056 : if (levelstate->pages[levelstate->current_page] == NULL)
483 84 : levelstate->pages[levelstate->current_page] = palloc0(BLCKSZ);
484 :
485 1056 : newPage = levelstate->pages[levelstate->current_page];
486 1056 : gistinitpage(newPage, old_page_flags);
487 : }
488 :
489 197558 : gistfillbuffer(levelstate->pages[levelstate->current_page], &itup, 1, InvalidOffsetNumber);
490 197558 : }
491 :
492 : static void
493 284 : gist_indexsortbuild_levelstate_flush(GISTBuildState *state,
494 : GistSortedBuildLevelState *levelstate)
495 : {
496 : GistSortedBuildLevelState *parent;
497 : BlockNumber blkno;
498 : MemoryContext oldCtx;
499 : IndexTuple union_tuple;
500 : SplitPageLayout *dist;
501 : IndexTuple *itvec;
502 : int vect_len;
503 284 : bool isleaf = GistPageIsLeaf(levelstate->pages[0]);
504 :
505 284 : CHECK_FOR_INTERRUPTS();
506 :
507 284 : oldCtx = MemoryContextSwitchTo(state->giststate->tempCxt);
508 :
509 : /* Get index tuples from first page */
510 284 : itvec = gistextractpage(levelstate->pages[0], &vect_len);
511 284 : if (levelstate->current_page > 0)
512 : {
513 : /* Append tuples from each page */
514 1078 : for (int i = 1; i < levelstate->current_page + 1; i++)
515 : {
516 : int len_local;
517 800 : IndexTuple *itvec_local = gistextractpage(levelstate->pages[i], &len_local);
518 :
519 800 : itvec = gistjoinvector(itvec, &vect_len, itvec_local, len_local);
520 800 : pfree(itvec_local);
521 : }
522 :
523 : /* Apply picksplit to list of all collected tuples */
524 278 : dist = gistSplit(state->indexrel, levelstate->pages[0], itvec, vect_len, state->giststate);
525 : }
526 : else
527 : {
528 : /* Create split layout from single page */
529 6 : dist = (SplitPageLayout *) palloc0(sizeof(SplitPageLayout));
530 6 : union_tuple = gistunion(state->indexrel, itvec, vect_len,
531 : state->giststate);
532 6 : dist->itup = union_tuple;
533 6 : dist->list = gistfillitupvec(itvec, vect_len, &(dist->lenlist));
534 6 : dist->block.num = vect_len;
535 : }
536 :
537 284 : MemoryContextSwitchTo(oldCtx);
538 :
539 : /* Reset page counter */
540 284 : levelstate->current_page = 0;
541 :
542 : /* Create pages for all partitions in split result */
543 1698 : for (; dist != NULL; dist = dist->next)
544 : {
545 : char *data;
546 : BulkWriteBuffer buf;
547 : Page target;
548 :
549 : /* check once per page */
550 1414 : CHECK_FOR_INTERRUPTS();
551 :
552 : /* Create page and copy data */
553 1414 : data = (char *) (dist->list);
554 1414 : buf = smgr_bulk_get_buf(state->bulkstate);
555 1414 : target = (Page) buf;
556 1414 : gistinitpage(target, isleaf ? F_LEAF : 0);
557 197426 : for (int i = 0; i < dist->block.num; i++)
558 : {
559 196012 : IndexTuple thistup = (IndexTuple) data;
560 :
561 196012 : if (PageAddItem(target, (Item) data, IndexTupleSize(thistup), i + FirstOffsetNumber, false, false) == InvalidOffsetNumber)
562 0 : elog(ERROR, "failed to add item to index page in \"%s\"", RelationGetRelationName(state->indexrel));
563 :
564 196012 : data += IndexTupleSize(thistup);
565 : }
566 1414 : union_tuple = dist->itup;
567 :
568 : /*
569 : * Set the right link to point to the previous page. This is just for
570 : * debugging purposes: GiST only follows the right link if a page is
571 : * split concurrently to a scan, and that cannot happen during index
572 : * build.
573 : *
574 : * It's a bit counterintuitive that we set the right link on the new
575 : * page to point to the previous page, not the other way around. But
576 : * GiST pages are not ordered like B-tree pages are, so as long as the
577 : * right-links form a chain through all the pages at the same level,
578 : * the order doesn't matter.
579 : */
580 1414 : if (levelstate->last_blkno)
581 1386 : GistPageGetOpaque(target)->rightlink = levelstate->last_blkno;
582 :
583 : /*
584 : * The page is now complete. Assign a block number to it, and pass it
585 : * to the bulk writer.
586 : */
587 1414 : blkno = state->pages_allocated++;
588 1414 : PageSetLSN(target, GistBuildLSN);
589 1414 : smgr_bulk_write(state->bulkstate, blkno, buf, true);
590 1414 : ItemPointerSetBlockNumber(&(union_tuple->t_tid), blkno);
591 1414 : levelstate->last_blkno = blkno;
592 :
593 : /*
594 : * Insert the downlink to the parent page. If this was the root,
595 : * create a new page as the parent, which becomes the new root.
596 : */
597 1414 : parent = levelstate->parent;
598 1414 : if (parent == NULL)
599 : {
600 28 : parent = palloc0(sizeof(GistSortedBuildLevelState));
601 28 : parent->pages[0] = palloc(BLCKSZ);
602 28 : parent->parent = NULL;
603 28 : gistinitpage(parent->pages[0], 0);
604 :
605 28 : levelstate->parent = parent;
606 : }
607 1414 : gist_indexsortbuild_levelstate_add(state, parent, union_tuple);
608 : }
609 284 : }
610 :
611 :
612 : /*-------------------------------------------------------------------------
613 : * Routines for non-sorted build
614 : *-------------------------------------------------------------------------
615 : */
616 :
617 : /*
618 : * Attempt to switch to buffering mode.
619 : *
620 : * If there is not enough memory for buffering build, sets bufferingMode
621 : * to GIST_BUFFERING_DISABLED, so that we don't bother to try the switch
622 : * anymore. Otherwise initializes the build buffers, and sets bufferingMode to
623 : * GIST_BUFFERING_ACTIVE.
624 : */
625 : static void
626 6 : gistInitBuffering(GISTBuildState *buildstate)
627 : {
628 6 : Relation index = buildstate->indexrel;
629 : int pagesPerBuffer;
630 : Size pageFreeSpace;
631 : Size itupAvgSize,
632 : itupMinSize;
633 : double avgIndexTuplesPerPage,
634 : maxIndexTuplesPerPage;
635 : int i;
636 : int levelStep;
637 :
638 : /* Calc space of index page which is available for index tuples */
639 6 : pageFreeSpace = BLCKSZ - SizeOfPageHeaderData - sizeof(GISTPageOpaqueData)
640 : - sizeof(ItemIdData)
641 6 : - buildstate->freespace;
642 :
643 : /*
644 : * Calculate average size of already inserted index tuples using gathered
645 : * statistics.
646 : */
647 6 : itupAvgSize = (double) buildstate->indtuplesSize /
648 6 : (double) buildstate->indtuples;
649 :
650 : /*
651 : * Calculate minimal possible size of index tuple by index metadata.
652 : * Minimal possible size of varlena is VARHDRSZ.
653 : *
654 : * XXX: that's not actually true, as a short varlen can be just 2 bytes.
655 : * And we should take padding into account here.
656 : */
657 6 : itupMinSize = (Size) MAXALIGN(sizeof(IndexTupleData));
658 12 : for (i = 0; i < index->rd_att->natts; i++)
659 : {
660 6 : if (TupleDescAttr(index->rd_att, i)->attlen < 0)
661 0 : itupMinSize += VARHDRSZ;
662 : else
663 6 : itupMinSize += TupleDescAttr(index->rd_att, i)->attlen;
664 : }
665 :
666 : /* Calculate average and maximal number of index tuples which fit to page */
667 6 : avgIndexTuplesPerPage = pageFreeSpace / itupAvgSize;
668 6 : maxIndexTuplesPerPage = pageFreeSpace / itupMinSize;
669 :
670 : /*
671 : * We need to calculate two parameters for the buffering algorithm:
672 : * levelStep and pagesPerBuffer.
673 : *
674 : * levelStep determines the size of subtree that we operate on, while
675 : * emptying a buffer. A higher value is better, as you need fewer buffer
676 : * emptying steps to build the index. However, if you set it too high, the
677 : * subtree doesn't fit in cache anymore, and you quickly lose the benefit
678 : * of the buffers.
679 : *
680 : * In Arge et al's paper, levelStep is chosen as logB(M/4B), where B is
681 : * the number of tuples on page (ie. fanout), and M is the amount of
682 : * internal memory available. Curiously, they doesn't explain *why* that
683 : * setting is optimal. We calculate it by taking the highest levelStep so
684 : * that a subtree still fits in cache. For a small B, our way of
685 : * calculating levelStep is very close to Arge et al's formula. For a
686 : * large B, our formula gives a value that is 2x higher.
687 : *
688 : * The average size (in pages) of a subtree of depth n can be calculated
689 : * as a geometric series:
690 : *
691 : * B^0 + B^1 + B^2 + ... + B^n = (1 - B^(n + 1)) / (1 - B)
692 : *
693 : * where B is the average number of index tuples on page. The subtree is
694 : * cached in the shared buffer cache and the OS cache, so we choose
695 : * levelStep so that the subtree size is comfortably smaller than
696 : * effective_cache_size, with a safety factor of 4.
697 : *
698 : * The estimate on the average number of index tuples on page is based on
699 : * average tuple sizes observed before switching to buffered build, so the
700 : * real subtree size can be somewhat larger. Also, it would selfish to
701 : * gobble the whole cache for our index build. The safety factor of 4
702 : * should account for those effects.
703 : *
704 : * The other limiting factor for setting levelStep is that while
705 : * processing a subtree, we need to hold one page for each buffer at the
706 : * next lower buffered level. The max. number of buffers needed for that
707 : * is maxIndexTuplesPerPage^levelStep. This is very conservative, but
708 : * hopefully maintenance_work_mem is set high enough that you're
709 : * constrained by effective_cache_size rather than maintenance_work_mem.
710 : *
711 : * XXX: the buffer hash table consumes a fair amount of memory too per
712 : * buffer, but that is not currently taken into account. That scales on
713 : * the total number of buffers used, ie. the index size and on levelStep.
714 : * Note that a higher levelStep *reduces* the amount of memory needed for
715 : * the hash table.
716 : */
717 6 : levelStep = 1;
718 : for (;;)
719 6 : {
720 : double subtreesize;
721 : double maxlowestlevelpages;
722 :
723 : /* size of an average subtree at this levelStep (in pages). */
724 12 : subtreesize =
725 12 : (1 - pow(avgIndexTuplesPerPage, (double) (levelStep + 1))) /
726 12 : (1 - avgIndexTuplesPerPage);
727 :
728 : /* max number of pages at the lowest level of a subtree */
729 12 : maxlowestlevelpages = pow(maxIndexTuplesPerPage, (double) levelStep);
730 :
731 : /* subtree must fit in cache (with safety factor of 4) */
732 12 : if (subtreesize > effective_cache_size / 4)
733 0 : break;
734 :
735 : /* each node in the lowest level of a subtree has one page in memory */
736 12 : if (maxlowestlevelpages > ((double) maintenance_work_mem * 1024) / BLCKSZ)
737 6 : break;
738 :
739 : /* Good, we can handle this levelStep. See if we can go one higher. */
740 6 : levelStep++;
741 : }
742 :
743 : /*
744 : * We just reached an unacceptable value of levelStep in previous loop.
745 : * So, decrease levelStep to get last acceptable value.
746 : */
747 6 : levelStep--;
748 :
749 : /*
750 : * If there's not enough cache or maintenance_work_mem, fall back to plain
751 : * inserts.
752 : */
753 6 : if (levelStep <= 0)
754 : {
755 0 : elog(DEBUG1, "failed to switch to buffered GiST build");
756 0 : buildstate->buildMode = GIST_BUFFERING_DISABLED;
757 0 : return;
758 : }
759 :
760 : /*
761 : * The second parameter to set is pagesPerBuffer, which determines the
762 : * size of each buffer. We adjust pagesPerBuffer also during the build,
763 : * which is why this calculation is in a separate function.
764 : */
765 6 : pagesPerBuffer = calculatePagesPerBuffer(buildstate, levelStep);
766 :
767 : /* Initialize GISTBuildBuffers with these parameters */
768 6 : buildstate->gfbb = gistInitBuildBuffers(pagesPerBuffer, levelStep,
769 : gistGetMaxLevel(index));
770 :
771 6 : gistInitParentMap(buildstate);
772 :
773 6 : buildstate->buildMode = GIST_BUFFERING_ACTIVE;
774 :
775 6 : elog(DEBUG1, "switched to buffered GiST build; level step = %d, pagesPerBuffer = %d",
776 : levelStep, pagesPerBuffer);
777 : }
778 :
779 : /*
780 : * Calculate pagesPerBuffer parameter for the buffering algorithm.
781 : *
782 : * Buffer size is chosen so that assuming that tuples are distributed
783 : * randomly, emptying half a buffer fills on average one page in every buffer
784 : * at the next lower level.
785 : */
786 : static int
787 12 : calculatePagesPerBuffer(GISTBuildState *buildstate, int levelStep)
788 : {
789 : double pagesPerBuffer;
790 : double avgIndexTuplesPerPage;
791 : double itupAvgSize;
792 : Size pageFreeSpace;
793 :
794 : /* Calc space of index page which is available for index tuples */
795 12 : pageFreeSpace = BLCKSZ - SizeOfPageHeaderData - sizeof(GISTPageOpaqueData)
796 : - sizeof(ItemIdData)
797 12 : - buildstate->freespace;
798 :
799 : /*
800 : * Calculate average size of already inserted index tuples using gathered
801 : * statistics.
802 : */
803 12 : itupAvgSize = (double) buildstate->indtuplesSize /
804 12 : (double) buildstate->indtuples;
805 :
806 12 : avgIndexTuplesPerPage = pageFreeSpace / itupAvgSize;
807 :
808 : /*
809 : * Recalculate required size of buffers.
810 : */
811 12 : pagesPerBuffer = 2 * pow(avgIndexTuplesPerPage, levelStep);
812 :
813 12 : return (int) rint(pagesPerBuffer);
814 : }
815 :
816 : /*
817 : * Per-tuple callback for table_index_build_scan.
818 : */
819 : static void
820 660486 : gistBuildCallback(Relation index,
821 : ItemPointer tid,
822 : Datum *values,
823 : bool *isnull,
824 : bool tupleIsAlive,
825 : void *state)
826 : {
827 660486 : GISTBuildState *buildstate = (GISTBuildState *) state;
828 : IndexTuple itup;
829 : MemoryContext oldCtx;
830 :
831 660486 : oldCtx = MemoryContextSwitchTo(buildstate->giststate->tempCxt);
832 :
833 : /* form an index tuple and point it at the heap tuple */
834 660486 : itup = gistFormTuple(buildstate->giststate, index,
835 : values, isnull,
836 : true);
837 660486 : itup->t_tid = *tid;
838 :
839 : /* Update tuple count and total size. */
840 660486 : buildstate->indtuples += 1;
841 660486 : buildstate->indtuplesSize += IndexTupleSize(itup);
842 :
843 : /*
844 : * XXX In buffering builds, the tempCxt is also reset down inside
845 : * gistProcessEmptyingQueue(). This is not great because it risks
846 : * confusion and possible use of dangling pointers (for example, itup
847 : * might be already freed when control returns here). It's generally
848 : * better that a memory context be "owned" by only one function. However,
849 : * currently this isn't causing issues so it doesn't seem worth the amount
850 : * of refactoring that would be needed to avoid it.
851 : */
852 660486 : if (buildstate->buildMode == GIST_BUFFERING_ACTIVE)
853 : {
854 : /* We have buffers, so use them. */
855 35430 : gistBufferingBuildInsert(buildstate, itup);
856 : }
857 : else
858 : {
859 : /*
860 : * There's no buffers (yet). Since we already have the index relation
861 : * locked, we call gistdoinsert directly.
862 : */
863 625056 : gistdoinsert(index, itup, buildstate->freespace,
864 : buildstate->giststate, buildstate->heaprel, true);
865 : }
866 :
867 660486 : MemoryContextSwitchTo(oldCtx);
868 660486 : MemoryContextReset(buildstate->giststate->tempCxt);
869 :
870 660486 : if (buildstate->buildMode == GIST_BUFFERING_ACTIVE &&
871 35430 : buildstate->indtuples % BUFFERING_MODE_TUPLE_SIZE_STATS_TARGET == 0)
872 : {
873 : /* Adjust the target buffer size now */
874 6 : buildstate->gfbb->pagesPerBuffer =
875 6 : calculatePagesPerBuffer(buildstate, buildstate->gfbb->levelStep);
876 : }
877 :
878 : /*
879 : * In 'auto' mode, check if the index has grown too large to fit in cache,
880 : * and switch to buffering mode if it has.
881 : *
882 : * To avoid excessive calls to smgrnblocks(), only check this every
883 : * BUFFERING_MODE_SWITCH_CHECK_STEP index tuples.
884 : *
885 : * In 'stats' state, switch as soon as we have seen enough tuples to have
886 : * some idea of the average tuple size.
887 : */
888 660486 : if ((buildstate->buildMode == GIST_BUFFERING_AUTO &&
889 600480 : buildstate->indtuples % BUFFERING_MODE_SWITCH_CHECK_STEP == 0 &&
890 2256 : effective_cache_size < smgrnblocks(RelationGetSmgr(index),
891 660486 : MAIN_FORKNUM)) ||
892 660486 : (buildstate->buildMode == GIST_BUFFERING_STATS &&
893 24576 : buildstate->indtuples >= BUFFERING_MODE_TUPLE_SIZE_STATS_TARGET))
894 : {
895 : /*
896 : * Index doesn't fit in effective cache anymore. Try to switch to
897 : * buffering build mode.
898 : */
899 6 : gistInitBuffering(buildstate);
900 : }
901 660486 : }
902 :
903 : /*
904 : * Insert function for buffering index build.
905 : */
906 : static void
907 35430 : gistBufferingBuildInsert(GISTBuildState *buildstate, IndexTuple itup)
908 : {
909 : /* Insert the tuple to buffers. */
910 35430 : gistProcessItup(buildstate, itup, 0, buildstate->gfbb->rootlevel);
911 :
912 : /* If we filled up (half of a) buffer, process buffer emptying. */
913 35430 : gistProcessEmptyingQueue(buildstate);
914 35430 : }
915 :
916 : /*
917 : * Process an index tuple. Runs the tuple down the tree until we reach a leaf
918 : * page or node buffer, and inserts the tuple there. Returns true if we have
919 : * to stop buffer emptying process (because one of child buffers can't take
920 : * index tuples anymore).
921 : */
922 : static bool
923 69762 : gistProcessItup(GISTBuildState *buildstate, IndexTuple itup,
924 : BlockNumber startblkno, int startlevel)
925 : {
926 69762 : GISTSTATE *giststate = buildstate->giststate;
927 69762 : GISTBuildBuffers *gfbb = buildstate->gfbb;
928 69762 : Relation indexrel = buildstate->indexrel;
929 : BlockNumber childblkno;
930 : Buffer buffer;
931 69762 : bool result = false;
932 : BlockNumber blkno;
933 : int level;
934 69762 : OffsetNumber downlinkoffnum = InvalidOffsetNumber;
935 69762 : BlockNumber parentblkno = InvalidBlockNumber;
936 :
937 69762 : CHECK_FOR_INTERRUPTS();
938 :
939 : /*
940 : * Loop until we reach a leaf page (level == 0) or a level with buffers
941 : * (not including the level we start at, because we would otherwise make
942 : * no progress).
943 : */
944 69762 : blkno = startblkno;
945 69762 : level = startlevel;
946 : for (;;)
947 69762 : {
948 : ItemId iid;
949 : IndexTuple idxtuple,
950 : newtup;
951 : Page page;
952 : OffsetNumber childoffnum;
953 :
954 : /* Have we reached a level with buffers? */
955 139524 : if (LEVEL_HAS_BUFFERS(level, gfbb) && level != startlevel)
956 34332 : break;
957 :
958 : /* Have we reached a leaf page? */
959 105192 : if (level == 0)
960 35430 : break;
961 :
962 : /*
963 : * Nope. Descend down to the next level then. Choose a child to
964 : * descend down to.
965 : */
966 :
967 69762 : buffer = ReadBuffer(indexrel, blkno);
968 69762 : LockBuffer(buffer, GIST_EXCLUSIVE);
969 :
970 69762 : page = (Page) BufferGetPage(buffer);
971 69762 : childoffnum = gistchoose(indexrel, page, itup, giststate);
972 69762 : iid = PageGetItemId(page, childoffnum);
973 69762 : idxtuple = (IndexTuple) PageGetItem(page, iid);
974 69762 : childblkno = ItemPointerGetBlockNumber(&(idxtuple->t_tid));
975 :
976 69762 : if (level > 1)
977 34332 : gistMemorizeParent(buildstate, childblkno, blkno);
978 :
979 : /*
980 : * Check that the key representing the target child node is consistent
981 : * with the key we're inserting. Update it if it's not.
982 : */
983 69762 : newtup = gistgetadjusted(indexrel, idxtuple, itup, giststate);
984 69762 : if (newtup)
985 : {
986 68946 : blkno = gistbufferinginserttuples(buildstate,
987 : buffer,
988 : level,
989 : &newtup,
990 : 1,
991 : childoffnum,
992 : InvalidBlockNumber,
993 : InvalidOffsetNumber);
994 : /* gistbufferinginserttuples() released the buffer */
995 : }
996 : else
997 816 : UnlockReleaseBuffer(buffer);
998 :
999 : /* Descend to the child */
1000 69762 : parentblkno = blkno;
1001 69762 : blkno = childblkno;
1002 69762 : downlinkoffnum = childoffnum;
1003 : Assert(level > 0);
1004 69762 : level--;
1005 : }
1006 :
1007 69762 : if (LEVEL_HAS_BUFFERS(level, gfbb))
1008 34332 : {
1009 : /*
1010 : * We've reached level with buffers. Place the index tuple to the
1011 : * buffer, and add the buffer to the emptying queue if it overflows.
1012 : */
1013 : GISTNodeBuffer *childNodeBuffer;
1014 :
1015 : /* Find the buffer or create a new one */
1016 34332 : childNodeBuffer = gistGetNodeBuffer(gfbb, giststate, blkno, level);
1017 :
1018 : /* Add index tuple to it */
1019 34332 : gistPushItupToNodeBuffer(gfbb, childNodeBuffer, itup);
1020 :
1021 34332 : if (BUFFER_OVERFLOWED(childNodeBuffer, gfbb))
1022 0 : result = true;
1023 : }
1024 : else
1025 : {
1026 : /*
1027 : * We've reached a leaf page. Place the tuple here.
1028 : */
1029 : Assert(level == 0);
1030 35430 : buffer = ReadBuffer(indexrel, blkno);
1031 35430 : LockBuffer(buffer, GIST_EXCLUSIVE);
1032 35430 : gistbufferinginserttuples(buildstate, buffer, level,
1033 : &itup, 1, InvalidOffsetNumber,
1034 : parentblkno, downlinkoffnum);
1035 : /* gistbufferinginserttuples() released the buffer */
1036 : }
1037 :
1038 69762 : return result;
1039 : }
1040 :
1041 : /*
1042 : * Insert tuples to a given page.
1043 : *
1044 : * This is analogous with gistinserttuples() in the regular insertion code.
1045 : *
1046 : * Returns the block number of the page where the (first) new or updated tuple
1047 : * was inserted. Usually that's the original page, but might be a sibling page
1048 : * if the original page was split.
1049 : *
1050 : * Caller should hold a lock on 'buffer' on entry. This function will unlock
1051 : * and unpin it.
1052 : */
1053 : static BlockNumber
1054 105144 : gistbufferinginserttuples(GISTBuildState *buildstate, Buffer buffer, int level,
1055 : IndexTuple *itup, int ntup, OffsetNumber oldoffnum,
1056 : BlockNumber parentblk, OffsetNumber downlinkoffnum)
1057 : {
1058 105144 : GISTBuildBuffers *gfbb = buildstate->gfbb;
1059 : List *splitinfo;
1060 : bool is_split;
1061 105144 : BlockNumber placed_to_blk = InvalidBlockNumber;
1062 :
1063 105144 : is_split = gistplacetopage(buildstate->indexrel,
1064 : buildstate->freespace,
1065 : buildstate->giststate,
1066 : buffer,
1067 : itup, ntup, oldoffnum, &placed_to_blk,
1068 : InvalidBuffer,
1069 : &splitinfo,
1070 : false,
1071 : buildstate->heaprel, true);
1072 :
1073 : /*
1074 : * If this is a root split, update the root path item kept in memory. This
1075 : * ensures that all path stacks are always complete, including all parent
1076 : * nodes up to the root. That simplifies the algorithm to re-find correct
1077 : * parent.
1078 : */
1079 105144 : if (is_split && BufferGetBlockNumber(buffer) == GIST_ROOT_BLKNO)
1080 : {
1081 6 : Page page = BufferGetPage(buffer);
1082 : OffsetNumber off;
1083 : OffsetNumber maxoff;
1084 :
1085 : Assert(level == gfbb->rootlevel);
1086 6 : gfbb->rootlevel++;
1087 :
1088 6 : elog(DEBUG2, "splitting GiST root page, now %d levels deep", gfbb->rootlevel);
1089 :
1090 : /*
1091 : * All the downlinks on the old root page are now on one of the child
1092 : * pages. Visit all the new child pages to memorize the parents of the
1093 : * grandchildren.
1094 : */
1095 6 : if (gfbb->rootlevel > 1)
1096 : {
1097 6 : maxoff = PageGetMaxOffsetNumber(page);
1098 18 : for (off = FirstOffsetNumber; off <= maxoff; off++)
1099 : {
1100 12 : ItemId iid = PageGetItemId(page, off);
1101 12 : IndexTuple idxtuple = (IndexTuple) PageGetItem(page, iid);
1102 12 : BlockNumber childblkno = ItemPointerGetBlockNumber(&(idxtuple->t_tid));
1103 12 : Buffer childbuf = ReadBuffer(buildstate->indexrel, childblkno);
1104 :
1105 12 : LockBuffer(childbuf, GIST_SHARE);
1106 12 : gistMemorizeAllDownlinks(buildstate, childbuf);
1107 12 : UnlockReleaseBuffer(childbuf);
1108 :
1109 : /*
1110 : * Also remember that the parent of the new child page is the
1111 : * root block.
1112 : */
1113 12 : gistMemorizeParent(buildstate, childblkno, GIST_ROOT_BLKNO);
1114 : }
1115 : }
1116 : }
1117 :
1118 105144 : if (splitinfo)
1119 : {
1120 : /*
1121 : * Insert the downlinks to the parent. This is analogous with
1122 : * gistfinishsplit() in the regular insertion code, but the locking is
1123 : * simpler, and we have to maintain the buffers on internal nodes and
1124 : * the parent map.
1125 : */
1126 : IndexTuple *downlinks;
1127 : int ndownlinks,
1128 : i;
1129 : Buffer parentBuffer;
1130 : ListCell *lc;
1131 :
1132 : /* Parent may have changed since we memorized this path. */
1133 : parentBuffer =
1134 768 : gistBufferingFindCorrectParent(buildstate,
1135 : BufferGetBlockNumber(buffer),
1136 : level,
1137 : &parentblk,
1138 : &downlinkoffnum);
1139 :
1140 : /*
1141 : * If there's a buffer associated with this page, that needs to be
1142 : * split too. gistRelocateBuildBuffersOnSplit() will also adjust the
1143 : * downlinks in 'splitinfo', to make sure they're consistent not only
1144 : * with the tuples already on the pages, but also the tuples in the
1145 : * buffers that will eventually be inserted to them.
1146 : */
1147 768 : gistRelocateBuildBuffersOnSplit(gfbb,
1148 : buildstate->giststate,
1149 : buildstate->indexrel,
1150 : level,
1151 : buffer, splitinfo);
1152 :
1153 : /* Create an array of all the downlink tuples */
1154 768 : ndownlinks = list_length(splitinfo);
1155 768 : downlinks = (IndexTuple *) palloc(sizeof(IndexTuple) * ndownlinks);
1156 768 : i = 0;
1157 2304 : foreach(lc, splitinfo)
1158 : {
1159 1536 : GISTPageSplitInfo *splitinfo = lfirst(lc);
1160 :
1161 : /*
1162 : * Remember the parent of each new child page in our parent map.
1163 : * This assumes that the downlinks fit on the parent page. If the
1164 : * parent page is split, too, when we recurse up to insert the
1165 : * downlinks, the recursive gistbufferinginserttuples() call will
1166 : * update the map again.
1167 : */
1168 1536 : if (level > 0)
1169 24 : gistMemorizeParent(buildstate,
1170 : BufferGetBlockNumber(splitinfo->buf),
1171 : BufferGetBlockNumber(parentBuffer));
1172 :
1173 : /*
1174 : * Also update the parent map for all the downlinks that got moved
1175 : * to a different page. (actually this also loops through the
1176 : * downlinks that stayed on the original page, but it does no
1177 : * harm).
1178 : */
1179 1536 : if (level > 1)
1180 0 : gistMemorizeAllDownlinks(buildstate, splitinfo->buf);
1181 :
1182 : /*
1183 : * Since there's no concurrent access, we can release the lower
1184 : * level buffers immediately. This includes the original page.
1185 : */
1186 1536 : UnlockReleaseBuffer(splitinfo->buf);
1187 1536 : downlinks[i++] = splitinfo->downlink;
1188 : }
1189 :
1190 : /* Insert them into parent. */
1191 768 : gistbufferinginserttuples(buildstate, parentBuffer, level + 1,
1192 : downlinks, ndownlinks, downlinkoffnum,
1193 : InvalidBlockNumber, InvalidOffsetNumber);
1194 :
1195 768 : list_free_deep(splitinfo); /* we don't need this anymore */
1196 : }
1197 : else
1198 104376 : UnlockReleaseBuffer(buffer);
1199 :
1200 105144 : return placed_to_blk;
1201 : }
1202 :
1203 : /*
1204 : * Find the downlink pointing to a child page.
1205 : *
1206 : * 'childblkno' indicates the child page to find the parent for. 'level' is
1207 : * the level of the child. On entry, *parentblkno and *downlinkoffnum can
1208 : * point to a location where the downlink used to be - we will check that
1209 : * location first, and save some cycles if it hasn't moved. The function
1210 : * returns a buffer containing the downlink, exclusively-locked, and
1211 : * *parentblkno and *downlinkoffnum are set to the real location of the
1212 : * downlink.
1213 : *
1214 : * If the child page is a leaf (level == 0), the caller must supply a correct
1215 : * parentblkno. Otherwise we use the parent map hash table to find the parent
1216 : * block.
1217 : *
1218 : * This function serves the same purpose as gistFindCorrectParent() during
1219 : * normal index inserts, but this is simpler because we don't need to deal
1220 : * with concurrent inserts.
1221 : */
1222 : static Buffer
1223 768 : gistBufferingFindCorrectParent(GISTBuildState *buildstate,
1224 : BlockNumber childblkno, int level,
1225 : BlockNumber *parentblkno,
1226 : OffsetNumber *downlinkoffnum)
1227 : {
1228 : BlockNumber parent;
1229 : Buffer buffer;
1230 : Page page;
1231 : OffsetNumber maxoff;
1232 : OffsetNumber off;
1233 :
1234 768 : if (level > 0)
1235 12 : parent = gistGetParent(buildstate, childblkno);
1236 : else
1237 : {
1238 : /*
1239 : * For a leaf page, the caller must supply a correct parent block
1240 : * number.
1241 : */
1242 756 : if (*parentblkno == InvalidBlockNumber)
1243 0 : elog(ERROR, "no parent buffer provided of child %u", childblkno);
1244 756 : parent = *parentblkno;
1245 : }
1246 :
1247 768 : buffer = ReadBuffer(buildstate->indexrel, parent);
1248 768 : page = BufferGetPage(buffer);
1249 768 : LockBuffer(buffer, GIST_EXCLUSIVE);
1250 768 : gistcheckpage(buildstate->indexrel, buffer);
1251 768 : maxoff = PageGetMaxOffsetNumber(page);
1252 :
1253 : /* Check if it was not moved */
1254 768 : if (parent == *parentblkno && *parentblkno != InvalidBlockNumber &&
1255 756 : *downlinkoffnum != InvalidOffsetNumber && *downlinkoffnum <= maxoff)
1256 : {
1257 756 : ItemId iid = PageGetItemId(page, *downlinkoffnum);
1258 756 : IndexTuple idxtuple = (IndexTuple) PageGetItem(page, iid);
1259 :
1260 756 : if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == childblkno)
1261 : {
1262 : /* Still there */
1263 756 : return buffer;
1264 : }
1265 : }
1266 :
1267 : /*
1268 : * Downlink was not at the offset where it used to be. Scan the page to
1269 : * find it. During normal gist insertions, it might've moved to another
1270 : * page, to the right, but during a buffering build, we keep track of the
1271 : * parent of each page in the lookup table so we should always know what
1272 : * page it's on.
1273 : */
1274 30 : for (off = FirstOffsetNumber; off <= maxoff; off = OffsetNumberNext(off))
1275 : {
1276 30 : ItemId iid = PageGetItemId(page, off);
1277 30 : IndexTuple idxtuple = (IndexTuple) PageGetItem(page, iid);
1278 :
1279 30 : if (ItemPointerGetBlockNumber(&(idxtuple->t_tid)) == childblkno)
1280 : {
1281 : /* yes!!, found it */
1282 12 : *downlinkoffnum = off;
1283 12 : return buffer;
1284 : }
1285 : }
1286 :
1287 0 : elog(ERROR, "failed to re-find parent for block %u", childblkno);
1288 : return InvalidBuffer; /* keep compiler quiet */
1289 : }
1290 :
1291 : /*
1292 : * Process buffers emptying stack. Emptying of one buffer can cause emptying
1293 : * of other buffers. This function iterates until this cascading emptying
1294 : * process finished, e.g. until buffers emptying stack is empty.
1295 : */
1296 : static void
1297 35448 : gistProcessEmptyingQueue(GISTBuildState *buildstate)
1298 : {
1299 35448 : GISTBuildBuffers *gfbb = buildstate->gfbb;
1300 :
1301 : /* Iterate while we have elements in buffers emptying stack. */
1302 35466 : while (gfbb->bufferEmptyingQueue != NIL)
1303 : {
1304 : GISTNodeBuffer *emptyingNodeBuffer;
1305 :
1306 : /* Get node buffer from emptying stack. */
1307 18 : emptyingNodeBuffer = (GISTNodeBuffer *) linitial(gfbb->bufferEmptyingQueue);
1308 18 : gfbb->bufferEmptyingQueue = list_delete_first(gfbb->bufferEmptyingQueue);
1309 18 : emptyingNodeBuffer->queuedForEmptying = false;
1310 :
1311 : /*
1312 : * We are going to load last pages of buffers where emptying will be
1313 : * to. So let's unload any previously loaded buffers.
1314 : */
1315 18 : gistUnloadNodeBuffers(gfbb);
1316 :
1317 : /*
1318 : * Pop tuples from the buffer and run them down to the buffers at
1319 : * lower level, or leaf pages. We continue until one of the lower
1320 : * level buffers fills up, or this buffer runs empty.
1321 : *
1322 : * In Arge et al's paper, the buffer emptying is stopped after
1323 : * processing 1/2 node buffer worth of tuples, to avoid overfilling
1324 : * any of the lower level buffers. However, it's more efficient to
1325 : * keep going until one of the lower level buffers actually fills up,
1326 : * so that's what we do. This doesn't need to be exact, if a buffer
1327 : * overfills by a few tuples, there's no harm done.
1328 : */
1329 : while (true)
1330 34332 : {
1331 : IndexTuple itup;
1332 :
1333 : /* Get next index tuple from the buffer */
1334 34350 : if (!gistPopItupFromNodeBuffer(gfbb, emptyingNodeBuffer, &itup))
1335 18 : break;
1336 :
1337 : /*
1338 : * Run it down to the underlying node buffer or leaf page.
1339 : *
1340 : * Note: it's possible that the buffer we're emptying splits as a
1341 : * result of this call. If that happens, our emptyingNodeBuffer
1342 : * points to the left half of the split. After split, it's very
1343 : * likely that the new left buffer is no longer over the half-full
1344 : * threshold, but we might as well keep flushing tuples from it
1345 : * until we fill a lower-level buffer.
1346 : */
1347 34332 : if (gistProcessItup(buildstate, itup, emptyingNodeBuffer->nodeBlocknum, emptyingNodeBuffer->level))
1348 : {
1349 : /*
1350 : * A lower level buffer filled up. Stop emptying this buffer,
1351 : * to avoid overflowing the lower level buffer.
1352 : */
1353 0 : break;
1354 : }
1355 :
1356 : /* Free all the memory allocated during index tuple processing */
1357 34332 : MemoryContextReset(buildstate->giststate->tempCxt);
1358 : }
1359 : }
1360 35448 : }
1361 :
1362 : /*
1363 : * Empty all node buffers, from top to bottom. This is done at the end of
1364 : * index build to flush all remaining tuples to the index.
1365 : *
1366 : * Note: This destroys the buffersOnLevels lists, so the buffers should not
1367 : * be inserted to after this call.
1368 : */
1369 : static void
1370 6 : gistEmptyAllBuffers(GISTBuildState *buildstate)
1371 : {
1372 6 : GISTBuildBuffers *gfbb = buildstate->gfbb;
1373 : MemoryContext oldCtx;
1374 : int i;
1375 :
1376 6 : oldCtx = MemoryContextSwitchTo(buildstate->giststate->tempCxt);
1377 :
1378 : /*
1379 : * Iterate through the levels from top to bottom.
1380 : */
1381 18 : for (i = gfbb->buffersOnLevelsLen - 1; i >= 0; i--)
1382 : {
1383 : /*
1384 : * Empty all buffers on this level. Note that new buffers can pop up
1385 : * in the list during the processing, as a result of page splits, so a
1386 : * simple walk through the list won't work. We remove buffers from the
1387 : * list when we see them empty; a buffer can't become non-empty once
1388 : * it's been fully emptied.
1389 : */
1390 48 : while (gfbb->buffersOnLevels[i] != NIL)
1391 : {
1392 : GISTNodeBuffer *nodeBuffer;
1393 :
1394 36 : nodeBuffer = (GISTNodeBuffer *) linitial(gfbb->buffersOnLevels[i]);
1395 :
1396 36 : if (nodeBuffer->blocksCount != 0)
1397 : {
1398 : /*
1399 : * Add this buffer to the emptying queue, and proceed to empty
1400 : * the queue.
1401 : */
1402 18 : if (!nodeBuffer->queuedForEmptying)
1403 : {
1404 18 : MemoryContextSwitchTo(gfbb->context);
1405 18 : nodeBuffer->queuedForEmptying = true;
1406 18 : gfbb->bufferEmptyingQueue =
1407 18 : lcons(nodeBuffer, gfbb->bufferEmptyingQueue);
1408 18 : MemoryContextSwitchTo(buildstate->giststate->tempCxt);
1409 : }
1410 18 : gistProcessEmptyingQueue(buildstate);
1411 : }
1412 : else
1413 18 : gfbb->buffersOnLevels[i] =
1414 18 : list_delete_first(gfbb->buffersOnLevels[i]);
1415 : }
1416 12 : elog(DEBUG2, "emptied all buffers at level %d", i);
1417 : }
1418 6 : MemoryContextSwitchTo(oldCtx);
1419 6 : }
1420 :
1421 : /*
1422 : * Get the depth of the GiST index.
1423 : */
1424 : static int
1425 6 : gistGetMaxLevel(Relation index)
1426 : {
1427 : int maxLevel;
1428 : BlockNumber blkno;
1429 :
1430 : /*
1431 : * Traverse down the tree, starting from the root, until we hit the leaf
1432 : * level.
1433 : */
1434 6 : maxLevel = 0;
1435 6 : blkno = GIST_ROOT_BLKNO;
1436 : while (true)
1437 6 : {
1438 : Buffer buffer;
1439 : Page page;
1440 : IndexTuple itup;
1441 :
1442 12 : buffer = ReadBuffer(index, blkno);
1443 :
1444 : /*
1445 : * There's no concurrent access during index build, so locking is just
1446 : * pro forma.
1447 : */
1448 12 : LockBuffer(buffer, GIST_SHARE);
1449 12 : page = (Page) BufferGetPage(buffer);
1450 :
1451 12 : if (GistPageIsLeaf(page))
1452 : {
1453 : /* We hit the bottom, so we're done. */
1454 6 : UnlockReleaseBuffer(buffer);
1455 6 : break;
1456 : }
1457 :
1458 : /*
1459 : * Pick the first downlink on the page, and follow it. It doesn't
1460 : * matter which downlink we choose, the tree has the same depth
1461 : * everywhere, so we just pick the first one.
1462 : */
1463 6 : itup = (IndexTuple) PageGetItem(page,
1464 : PageGetItemId(page, FirstOffsetNumber));
1465 6 : blkno = ItemPointerGetBlockNumber(&(itup->t_tid));
1466 6 : UnlockReleaseBuffer(buffer);
1467 :
1468 : /*
1469 : * We're going down on the tree. It means that there is yet one more
1470 : * level in the tree.
1471 : */
1472 6 : maxLevel++;
1473 : }
1474 6 : return maxLevel;
1475 : }
1476 :
1477 :
1478 : /*
1479 : * Routines for managing the parent map.
1480 : *
1481 : * Whenever a page is split, we need to insert the downlinks into the parent.
1482 : * We need to somehow find the parent page to do that. In normal insertions,
1483 : * we keep a stack of nodes visited when we descend the tree. However, in
1484 : * buffering build, we can start descending the tree from any internal node,
1485 : * when we empty a buffer by cascading tuples to its children. So we don't
1486 : * have a full stack up to the root available at that time.
1487 : *
1488 : * So instead, we maintain a hash table to track the parent of every internal
1489 : * page. We don't need to track the parents of leaf nodes, however. Whenever
1490 : * we insert to a leaf, we've just descended down from its parent, so we know
1491 : * its immediate parent already. This helps a lot to limit the memory used
1492 : * by this hash table.
1493 : *
1494 : * Whenever an internal node is split, the parent map needs to be updated.
1495 : * the parent of the new child page needs to be recorded, and also the
1496 : * entries for all page whose downlinks are moved to a new page at the split
1497 : * needs to be updated.
1498 : *
1499 : * We also update the parent map whenever we descend the tree. That might seem
1500 : * unnecessary, because we maintain the map whenever a downlink is moved or
1501 : * created, but it is needed because we switch to buffering mode after
1502 : * creating a tree with regular index inserts. Any pages created before
1503 : * switching to buffering mode will not be present in the parent map initially,
1504 : * but will be added there the first time we visit them.
1505 : */
1506 :
1507 : typedef struct
1508 : {
1509 : BlockNumber childblkno; /* hash key */
1510 : BlockNumber parentblkno;
1511 : } ParentMapEntry;
1512 :
1513 : static void
1514 6 : gistInitParentMap(GISTBuildState *buildstate)
1515 : {
1516 : HASHCTL hashCtl;
1517 :
1518 6 : hashCtl.keysize = sizeof(BlockNumber);
1519 6 : hashCtl.entrysize = sizeof(ParentMapEntry);
1520 6 : hashCtl.hcxt = CurrentMemoryContext;
1521 6 : buildstate->parentMap = hash_create("gistbuild parent map",
1522 : 1024,
1523 : &hashCtl,
1524 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1525 6 : }
1526 :
1527 : static void
1528 34926 : gistMemorizeParent(GISTBuildState *buildstate, BlockNumber child, BlockNumber parent)
1529 : {
1530 : ParentMapEntry *entry;
1531 : bool found;
1532 :
1533 34926 : entry = (ParentMapEntry *) hash_search(buildstate->parentMap,
1534 : &child,
1535 : HASH_ENTER,
1536 : &found);
1537 34926 : entry->parentblkno = parent;
1538 34926 : }
1539 :
1540 : /*
1541 : * Scan all downlinks on a page, and memorize their parent.
1542 : */
1543 : static void
1544 12 : gistMemorizeAllDownlinks(GISTBuildState *buildstate, Buffer parentbuf)
1545 : {
1546 : OffsetNumber maxoff;
1547 : OffsetNumber off;
1548 12 : BlockNumber parentblkno = BufferGetBlockNumber(parentbuf);
1549 12 : Page page = BufferGetPage(parentbuf);
1550 :
1551 : Assert(!GistPageIsLeaf(page));
1552 :
1553 12 : maxoff = PageGetMaxOffsetNumber(page);
1554 570 : for (off = FirstOffsetNumber; off <= maxoff; off++)
1555 : {
1556 558 : ItemId iid = PageGetItemId(page, off);
1557 558 : IndexTuple idxtuple = (IndexTuple) PageGetItem(page, iid);
1558 558 : BlockNumber childblkno = ItemPointerGetBlockNumber(&(idxtuple->t_tid));
1559 :
1560 558 : gistMemorizeParent(buildstate, childblkno, parentblkno);
1561 : }
1562 12 : }
1563 :
1564 : static BlockNumber
1565 12 : gistGetParent(GISTBuildState *buildstate, BlockNumber child)
1566 : {
1567 : ParentMapEntry *entry;
1568 : bool found;
1569 :
1570 : /* Find node buffer in hash table */
1571 12 : entry = (ParentMapEntry *) hash_search(buildstate->parentMap,
1572 : &child,
1573 : HASH_FIND,
1574 : &found);
1575 12 : if (!found)
1576 0 : elog(ERROR, "could not find parent of block %u in lookup table", child);
1577 :
1578 12 : return entry->parentblkno;
1579 : }
|
