Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * gininsert.c
4 : : * insert routines for the postgres inverted index access method.
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/gininsert.c
12 : : *-------------------------------------------------------------------------
13 : : */
14 : :
15 : : #include "postgres.h"
16 : :
17 : : #include "access/gin_private.h"
18 : : #include "access/gin_tuple.h"
19 : : #include "access/parallel.h"
20 : : #include "access/table.h"
21 : : #include "access/tableam.h"
22 : : #include "access/xloginsert.h"
23 : : #include "catalog/index.h"
24 : : #include "catalog/pg_collation.h"
25 : : #include "commands/progress.h"
26 : : #include "miscadmin.h"
27 : : #include "nodes/execnodes.h"
28 : : #include "pgstat.h"
29 : : #include "storage/bufmgr.h"
30 : : #include "storage/predicate.h"
31 : : #include "tcop/tcopprot.h"
32 : : #include "utils/datum.h"
33 : : #include "utils/memutils.h"
34 : : #include "utils/rel.h"
35 : : #include "utils/builtins.h"
36 : :
37 : :
38 : : /* Magic numbers for parallel state sharing */
39 : : #define PARALLEL_KEY_GIN_SHARED UINT64CONST(0xB000000000000001)
40 : : #define PARALLEL_KEY_TUPLESORT UINT64CONST(0xB000000000000002)
41 : : #define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xB000000000000003)
42 : : #define PARALLEL_KEY_WAL_USAGE UINT64CONST(0xB000000000000004)
43 : : #define PARALLEL_KEY_BUFFER_USAGE UINT64CONST(0xB000000000000005)
44 : :
45 : : /*
46 : : * Status for index builds performed in parallel. This is allocated in a
47 : : * dynamic shared memory segment.
48 : : */
49 : : typedef struct GinBuildShared
50 : : {
51 : : /*
52 : : * These fields are not modified during the build. They primarily exist
53 : : * for the benefit of worker processes that need to create state
54 : : * corresponding to that used by the leader.
55 : : */
56 : : Oid heaprelid;
57 : : Oid indexrelid;
58 : : bool isconcurrent;
59 : : int scantuplesortstates;
60 : :
61 : : /*
62 : : * workersdonecv is used to monitor the progress of workers. All parallel
63 : : * participants must indicate that they are done before leader can use
64 : : * results built by the workers (and before leader can write the data into
65 : : * the index).
66 : : */
67 : : ConditionVariable workersdonecv;
68 : :
69 : : /*
70 : : * mutex protects all following fields
71 : : *
72 : : * These fields contain status information of interest to GIN index builds
73 : : * that must work just the same when an index is built in parallel.
74 : : */
75 : : slock_t mutex;
76 : :
77 : : /*
78 : : * Mutable state that is maintained by workers, and reported back to
79 : : * leader at end of the scans.
80 : : *
81 : : * nparticipantsdone is number of worker processes finished.
82 : : *
83 : : * reltuples is the total number of input heap tuples.
84 : : *
85 : : * indtuples is the total number of tuples that made it into the index.
86 : : */
87 : : int nparticipantsdone;
88 : : double reltuples;
89 : : double indtuples;
90 : :
91 : : /*
92 : : * ParallelTableScanDescData data follows. Can't directly embed here, as
93 : : * implementations of the parallel table scan desc interface might need
94 : : * stronger alignment.
95 : : */
96 : : } GinBuildShared;
97 : :
98 : : /*
99 : : * Return pointer to a GinBuildShared's parallel table scan.
100 : : *
101 : : * c.f. shm_toc_allocate as to why BUFFERALIGN is used, rather than just
102 : : * MAXALIGN.
103 : : */
104 : : #define ParallelTableScanFromGinBuildShared(shared) \
105 : : (ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(GinBuildShared)))
106 : :
107 : : /*
108 : : * Status for leader in parallel index build.
109 : : */
110 : : typedef struct GinLeader
111 : : {
112 : : /* parallel context itself */
113 : : ParallelContext *pcxt;
114 : :
115 : : /*
116 : : * nparticipanttuplesorts is the exact number of worker processes
117 : : * successfully launched, plus one leader process if it participates as a
118 : : * worker (only DISABLE_LEADER_PARTICIPATION builds avoid leader
119 : : * participating as a worker).
120 : : */
121 : : int nparticipanttuplesorts;
122 : :
123 : : /*
124 : : * Leader process convenience pointers to shared state (leader avoids TOC
125 : : * lookups).
126 : : *
127 : : * GinBuildShared is the shared state for entire build. sharedsort is the
128 : : * shared, tuplesort-managed state passed to each process tuplesort.
129 : : * snapshot is the snapshot used by the scan iff an MVCC snapshot is
130 : : * required.
131 : : */
132 : : GinBuildShared *ginshared;
133 : : Sharedsort *sharedsort;
134 : : Snapshot snapshot;
135 : : WalUsage *walusage;
136 : : BufferUsage *bufferusage;
137 : : } GinLeader;
138 : :
139 : : typedef struct
140 : : {
141 : : GinState ginstate;
142 : : double indtuples;
143 : : GinStatsData buildStats;
144 : : MemoryContext tmpCtx;
145 : : MemoryContext funcCtx;
146 : : BuildAccumulator accum;
147 : : ItemPointerData tid;
148 : : int work_mem;
149 : :
150 : : /*
151 : : * bs_leader is only present when a parallel index build is performed, and
152 : : * only in the leader process.
153 : : */
154 : : GinLeader *bs_leader;
155 : : int bs_worker_id;
156 : :
157 : : /* used to pass information from workers to leader */
158 : : double bs_numtuples;
159 : : double bs_reltuples;
160 : :
161 : : /*
162 : : * The sortstate is used by workers (including the leader). It has to be
163 : : * part of the build state, because that's the only thing passed to the
164 : : * build callback etc.
165 : : */
166 : : Tuplesortstate *bs_sortstate;
167 : :
168 : : /*
169 : : * The sortstate used only within a single worker for the first merge pass
170 : : * happening there. In principle it doesn't need to be part of the build
171 : : * state and we could pass it around directly, but it's more convenient
172 : : * this way. And it's part of the build state, after all.
173 : : */
174 : : Tuplesortstate *bs_worker_sort;
175 : : } GinBuildState;
176 : :
177 : :
178 : : /* parallel index builds */
179 : : static void _gin_begin_parallel(GinBuildState *buildstate, Relation heap, Relation index,
180 : : bool isconcurrent, int request);
181 : : static void _gin_end_parallel(GinLeader *ginleader, GinBuildState *state);
182 : : static Size _gin_parallel_estimate_shared(Relation heap, Snapshot snapshot);
183 : : static double _gin_parallel_heapscan(GinBuildState *state);
184 : : static double _gin_parallel_merge(GinBuildState *state);
185 : : static void _gin_leader_participate_as_worker(GinBuildState *buildstate,
186 : : Relation heap, Relation index);
187 : : static void _gin_parallel_scan_and_build(GinBuildState *state,
188 : : GinBuildShared *ginshared,
189 : : Sharedsort *sharedsort,
190 : : Relation heap, Relation index,
191 : : int sortmem, bool progress);
192 : :
193 : : static ItemPointer _gin_parse_tuple_items(GinTuple *a);
194 : : static Datum _gin_parse_tuple_key(GinTuple *a);
195 : :
196 : : static GinTuple *_gin_build_tuple(OffsetNumber attrnum, unsigned char category,
197 : : Datum key, int16 typlen, bool typbyval,
198 : : ItemPointerData *items, uint32 nitems,
199 : : Size *len);
200 : :
201 : : /*
202 : : * Adds array of item pointers to tuple's posting list, or
203 : : * creates posting tree and tuple pointing to tree in case
204 : : * of not enough space. Max size of tuple is defined in
205 : : * GinFormTuple(). Returns a new, modified index tuple.
206 : : * items[] must be in sorted order with no duplicates.
207 : : */
208 : : static IndexTuple
5408 tgl@sss.pgh.pa.us 209 :CBC 114420 : addItemPointersToLeafTuple(GinState *ginstate,
210 : : IndexTuple old,
211 : : ItemPointerData *items, uint32 nitem,
212 : : GinStatsData *buildStats, Buffer buffer)
213 : : {
214 : : OffsetNumber attnum;
215 : : Datum key;
216 : : GinNullCategory category;
217 : : IndexTuple res;
218 : : ItemPointerData *newItems,
219 : : *oldItems;
220 : : int oldNPosting,
221 : : newNPosting,
222 : : nwritten;
223 : : GinPostingList *compressedList;
224 : :
225 [ - + ]: 114420 : Assert(!GinIsPostingTree(old));
226 : :
227 : 114420 : attnum = gintuple_get_attrnum(ginstate, old);
228 : 114420 : key = gintuple_get_key(ginstate, old, &category);
229 : :
230 : : /* merge the old and new posting lists */
4297 heikki.linnakangas@i 231 : 114420 : oldItems = ginReadTuple(ginstate, attnum, old, &oldNPosting);
232 : :
4236 233 : 114420 : newItems = ginMergeItemPointers(items, nitem,
234 : : oldItems, oldNPosting,
235 : : &newNPosting);
236 : :
237 : : /* Compress the posting list, and try to a build tuple with room for it */
4297 238 : 114420 : res = NULL;
22 msawada@postgresql.o 239 :GNC 114420 : compressedList = ginCompressPostingList(newItems, newNPosting, GinMaxItemSize, &nwritten);
240 [ + + ]: 114420 : if (nwritten == newNPosting)
241 : : {
4297 heikki.linnakangas@i 242 :CBC 114414 : res = GinFormTuple(ginstate, attnum, key, category,
243 : : (char *) compressedList,
244 : 114414 : SizeOfGinPostingList(compressedList),
245 : : newNPosting,
246 : : false);
247 : : }
248 : :
22 msawada@postgresql.o 249 :GNC 114420 : pfree(newItems);
250 : 114420 : pfree(compressedList);
251 : :
4297 heikki.linnakangas@i 252 [ + + ]:CBC 114420 : if (!res)
253 : : {
254 : : /* posting list would be too big, convert to posting tree */
255 : : BlockNumber postingRoot;
256 : :
257 : : /*
258 : : * Initialize posting tree with the old tuple's posting list. It's
259 : : * surely small enough to fit on one posting-tree page, and should
260 : : * already be in order with no duplicates.
261 : : */
5408 tgl@sss.pgh.pa.us 262 : 11 : postingRoot = createPostingTree(ginstate->index,
263 : : oldItems,
264 : : oldNPosting,
265 : : buildStats,
266 : : buffer);
267 : :
268 : : /* Now insert the TIDs-to-be-added into the posting tree */
4360 heikki.linnakangas@i 269 : 11 : ginInsertItemPointers(ginstate->index, postingRoot,
270 : : items, nitem,
271 : : buildStats);
272 : :
273 : : /* And build a new posting-tree-only result tuple */
4297 274 : 11 : res = GinFormTuple(ginstate, attnum, key, category, NULL, 0, 0, true);
5408 tgl@sss.pgh.pa.us 275 : 11 : GinSetPostingTree(res, postingRoot);
276 : : }
4297 heikki.linnakangas@i 277 : 114420 : pfree(oldItems);
278 : :
5408 tgl@sss.pgh.pa.us 279 : 114420 : return res;
280 : : }
281 : :
282 : : /*
283 : : * Build a fresh leaf tuple, either posting-list or posting-tree format
284 : : * depending on whether the given items list will fit.
285 : : * items[] must be in sorted order with no duplicates.
286 : : *
287 : : * This is basically the same logic as in addItemPointersToLeafTuple,
288 : : * but working from slightly different input.
289 : : */
290 : : static IndexTuple
291 : 256799 : buildFreshLeafTuple(GinState *ginstate,
292 : : OffsetNumber attnum, Datum key, GinNullCategory category,
293 : : ItemPointerData *items, uint32 nitem,
294 : : GinStatsData *buildStats, Buffer buffer)
295 : : {
4297 heikki.linnakangas@i 296 : 256799 : IndexTuple res = NULL;
297 : : GinPostingList *compressedList;
298 : : int nwritten;
299 : :
300 : : /* try to build a posting list tuple with all the items */
22 msawada@postgresql.o 301 :GNC 256799 : compressedList = ginCompressPostingList(items, nitem, GinMaxItemSize, &nwritten);
302 [ + + ]: 256799 : if (nwritten == nitem)
303 : : {
4297 heikki.linnakangas@i 304 :CBC 256748 : res = GinFormTuple(ginstate, attnum, key, category,
305 : : (char *) compressedList,
306 : 256748 : SizeOfGinPostingList(compressedList),
307 : : nitem, false);
308 : : }
22 msawada@postgresql.o 309 :GNC 256799 : pfree(compressedList);
310 : :
5408 tgl@sss.pgh.pa.us 311 [ + + ]:CBC 256799 : if (!res)
312 : : {
313 : : /* posting list would be too big, build posting tree */
314 : : BlockNumber postingRoot;
315 : :
316 : : /*
317 : : * Build posting-tree-only result tuple. We do this first so as to
318 : : * fail quickly if the key is too big.
319 : : */
4297 heikki.linnakangas@i 320 : 51 : res = GinFormTuple(ginstate, attnum, key, category, NULL, 0, 0, true);
321 : :
322 : : /*
323 : : * Initialize a new posting tree with the TIDs.
324 : : */
4374 325 : 51 : postingRoot = createPostingTree(ginstate->index, items, nitem,
326 : : buildStats, buffer);
327 : :
328 : : /* And save the root link in the result tuple */
5408 tgl@sss.pgh.pa.us 329 : 51 : GinSetPostingTree(res, postingRoot);
330 : : }
331 : :
7119 teodor@sigaev.ru 332 : 256799 : return res;
333 : : }
334 : :
335 : : /*
336 : : * Insert one or more heap TIDs associated with the given key value.
337 : : * This will either add a single key entry, or enlarge a pre-existing entry.
338 : : *
339 : : * During an index build, buildStats is non-null and the counters
340 : : * it contains should be incremented as needed.
341 : : */
342 : : void
5408 tgl@sss.pgh.pa.us 343 : 395921 : ginEntryInsert(GinState *ginstate,
344 : : OffsetNumber attnum, Datum key, GinNullCategory category,
345 : : ItemPointerData *items, uint32 nitem,
346 : : GinStatsData *buildStats)
347 : : {
348 : : GinBtreeData btree;
349 : : GinBtreeEntryInsertData insertdata;
350 : : GinBtreeStack *stack;
351 : : IndexTuple itup;
352 : : Page page;
353 : :
2995 peter_e@gmx.net 354 : 395921 : insertdata.isDelete = false;
355 : :
5408 tgl@sss.pgh.pa.us 356 : 395921 : ginPrepareEntryScan(&btree, attnum, key, category, ginstate);
2400 heikki.linnakangas@i 357 : 395921 : btree.isBuild = (buildStats != NULL);
358 : :
781 tmunro@postgresql.or 359 : 395921 : stack = ginFindLeafPage(&btree, false, false);
3478 kgrittn@postgresql.o 360 : 395921 : page = BufferGetPage(stack->buffer);
361 : :
6964 bruce@momjian.us 362 [ + + ]: 395921 : if (btree.findItem(&btree, stack))
363 : : {
364 : : /* found pre-existing entry */
7119 teodor@sigaev.ru 365 : 139120 : itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, stack->off));
366 : :
6964 bruce@momjian.us 367 [ + + ]: 139120 : if (GinIsPostingTree(itup))
368 : : {
369 : : /* add entries to existing posting tree */
370 : 24696 : BlockNumber rootPostingTree = GinGetPostingTree(itup);
371 : :
372 : : /* release all stack */
7119 teodor@sigaev.ru 373 : 24696 : LockBuffer(stack->buffer, GIN_UNLOCK);
6964 bruce@momjian.us 374 : 24696 : freeGinBtreeStack(stack);
375 : :
376 : : /* insert into posting tree */
4360 heikki.linnakangas@i 377 : 24696 : ginInsertItemPointers(ginstate->index, rootPostingTree,
378 : : items, nitem,
379 : : buildStats);
7119 teodor@sigaev.ru 380 : 24694 : return;
381 : : }
382 : :
2100 tmunro@postgresql.or 383 : 114424 : CheckForSerializableConflictIn(ginstate->index, NULL,
384 : : BufferGetBlockNumber(stack->buffer));
385 : : /* modify an existing leaf entry */
5408 tgl@sss.pgh.pa.us 386 : 114420 : itup = addItemPointersToLeafTuple(ginstate, itup,
387 : : items, nitem, buildStats, stack->buffer);
388 : :
2995 peter_e@gmx.net 389 : 114420 : insertdata.isDelete = true;
390 : : }
391 : : else
392 : : {
2100 tmunro@postgresql.or 393 : 256801 : CheckForSerializableConflictIn(ginstate->index, NULL,
394 : : BufferGetBlockNumber(stack->buffer));
395 : : /* no match, so construct a new leaf entry */
5408 tgl@sss.pgh.pa.us 396 : 256799 : itup = buildFreshLeafTuple(ginstate, attnum, key, category,
397 : : items, nitem, buildStats, stack->buffer);
398 : :
399 : : /*
400 : : * nEntries counts leaf tuples, so increment it only when we make a
401 : : * new one.
402 : : */
2185 403 [ + + ]: 256799 : if (buildStats)
404 : 76472 : buildStats->nEntries++;
405 : : }
406 : :
407 : : /* Insert the new or modified leaf tuple */
4353 heikki.linnakangas@i 408 : 371219 : insertdata.entry = itup;
409 : 371219 : ginInsertValue(&btree, stack, &insertdata, buildStats);
6964 bruce@momjian.us 410 : 371219 : pfree(itup);
411 : : }
412 : :
413 : : /*
414 : : * Extract index entries for a single indexable item, and add them to the
415 : : * BuildAccumulator's state.
416 : : *
417 : : * This function is used only during initial index creation.
418 : : */
419 : : static void
5408 tgl@sss.pgh.pa.us 420 : 475179 : ginHeapTupleBulkInsert(GinBuildState *buildstate, OffsetNumber attnum,
421 : : Datum value, bool isNull,
422 : : ItemPointer heapptr)
423 : : {
424 : : Datum *entries;
425 : : GinNullCategory *categories;
426 : : int32 nentries;
427 : : MemoryContext oldCtx;
428 : :
7049 teodor@sigaev.ru 429 : 475179 : oldCtx = MemoryContextSwitchTo(buildstate->funcCtx);
5408 tgl@sss.pgh.pa.us 430 : 475179 : entries = ginExtractEntries(buildstate->accum.ginstate, attnum,
431 : : value, isNull,
432 : : &nentries, &categories);
7049 teodor@sigaev.ru 433 : 475179 : MemoryContextSwitchTo(oldCtx);
434 : :
5408 tgl@sss.pgh.pa.us 435 : 475179 : ginInsertBAEntries(&buildstate->accum, heapptr, attnum,
436 : : entries, categories, nentries);
437 : :
438 : 475179 : buildstate->indtuples += nentries;
439 : :
7049 teodor@sigaev.ru 440 : 475179 : MemoryContextReset(buildstate->funcCtx);
7119 441 : 475179 : }
442 : :
443 : : static void
2181 andres@anarazel.de 444 : 474868 : ginBuildCallback(Relation index, ItemPointer tid, Datum *values,
445 : : bool *isnull, bool tupleIsAlive, void *state)
446 : : {
6964 bruce@momjian.us 447 : 474868 : GinBuildState *buildstate = (GinBuildState *) state;
448 : : MemoryContext oldCtx;
449 : : int i;
450 : :
7119 teodor@sigaev.ru 451 : 474868 : oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
452 : :
5983 bruce@momjian.us 453 [ + + ]: 950047 : for (i = 0; i < buildstate->ginstate.origTupdesc->natts; i++)
5408 tgl@sss.pgh.pa.us 454 : 475179 : ginHeapTupleBulkInsert(buildstate, (OffsetNumber) (i + 1),
2181 andres@anarazel.de 455 : 475179 : values[i], isnull[i], tid);
456 : :
457 : : /* If we've maxed out our available memory, dump everything to the index */
270 tgl@sss.pgh.pa.us 458 [ - + ]: 474868 : if (buildstate->accum.allocatedMemory >= maintenance_work_mem * (Size) 1024)
459 : : {
460 : : ItemPointerData *list;
461 : : Datum key;
462 : : GinNullCategory category;
463 : : uint32 nlist;
464 : : OffsetNumber attnum;
465 : :
5567 tgl@sss.pgh.pa.us 466 :UBC 0 : ginBeginBAScan(&buildstate->accum);
5408 467 : 0 : while ((list = ginGetBAEntry(&buildstate->accum,
3051 468 [ # # ]: 0 : &attnum, &key, &category, &nlist)) != NULL)
469 : : {
470 : : /* there could be many entries, so be willing to abort here */
6374 471 [ # # ]: 0 : CHECK_FOR_INTERRUPTS();
5408 472 : 0 : ginEntryInsert(&buildstate->ginstate, attnum, key, category,
473 : : list, nlist, &buildstate->buildStats);
474 : : }
475 : :
7119 teodor@sigaev.ru 476 : 0 : MemoryContextReset(buildstate->tmpCtx);
477 : 0 : ginInitBA(&buildstate->accum);
478 : : }
479 : :
7119 teodor@sigaev.ru 480 :CBC 474868 : MemoryContextSwitchTo(oldCtx);
481 : 474868 : }
482 : :
483 : : /*
484 : : * ginFlushBuildState
485 : : * Write all data from BuildAccumulator into the tuplesort.
486 : : */
487 : : static void
239 tomas.vondra@postgre 488 :UBC 0 : ginFlushBuildState(GinBuildState *buildstate, Relation index)
489 : : {
490 : : ItemPointerData *list;
491 : : Datum key;
492 : : GinNullCategory category;
493 : : uint32 nlist;
494 : : OffsetNumber attnum;
495 : 0 : TupleDesc tdesc = RelationGetDescr(index);
496 : :
497 : 0 : ginBeginBAScan(&buildstate->accum);
498 : 0 : while ((list = ginGetBAEntry(&buildstate->accum,
499 [ # # ]: 0 : &attnum, &key, &category, &nlist)) != NULL)
500 : : {
501 : : /* information about the key */
6 drowley@postgresql.o 502 :UNC 0 : CompactAttribute *attr = TupleDescCompactAttr(tdesc, (attnum - 1));
503 : :
504 : : /* GIN tuple and tuple length */
505 : : GinTuple *tup;
506 : : Size tuplen;
507 : :
508 : : /* there could be many entries, so be willing to abort here */
239 tomas.vondra@postgre 509 [ # # ]:UBC 0 : CHECK_FOR_INTERRUPTS();
510 : :
511 : 0 : tup = _gin_build_tuple(attnum, category,
512 : 0 : key, attr->attlen, attr->attbyval,
513 : : list, nlist, &tuplen);
514 : :
515 : 0 : tuplesort_putgintuple(buildstate->bs_worker_sort, tup, tuplen);
516 : :
517 : 0 : pfree(tup);
518 : : }
519 : :
520 : 0 : MemoryContextReset(buildstate->tmpCtx);
521 : 0 : ginInitBA(&buildstate->accum);
522 : 0 : }
523 : :
524 : : /*
525 : : * ginBuildCallbackParallel
526 : : * Callback for the parallel index build.
527 : : *
528 : : * This is similar to the serial build callback ginBuildCallback, but
529 : : * instead of writing the accumulated entries into the index, each worker
530 : : * writes them into a (local) tuplesort.
531 : : *
532 : : * The worker then sorts and combines these entries, before writing them
533 : : * into a shared tuplesort for the leader (see _gin_parallel_scan_and_build
534 : : * for the whole process).
535 : : */
536 : : static void
537 : 0 : ginBuildCallbackParallel(Relation index, ItemPointer tid, Datum *values,
538 : : bool *isnull, bool tupleIsAlive, void *state)
539 : : {
540 : 0 : GinBuildState *buildstate = (GinBuildState *) state;
541 : : MemoryContext oldCtx;
542 : : int i;
543 : :
544 : 0 : oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
545 : :
546 : : /*
547 : : * if scan wrapped around - flush accumulated entries and start anew
548 : : *
549 : : * With parallel scans, we don't have a guarantee the scan does not start
550 : : * half-way through the relation (serial builds disable sync scans and
551 : : * always start from block 0, parallel scans require allow_sync=true).
552 : : *
553 : : * Building the posting lists assumes the TIDs are monotonic and never go
554 : : * back, and the wrap around would break that. We handle that by detecting
555 : : * the wraparound, and flushing all entries. This means we'll later see
556 : : * two separate entries with non-overlapping TID lists (which can be
557 : : * combined by merge sort).
558 : : *
559 : : * To detect a wraparound, we remember the last TID seen by each worker
560 : : * (for any key). If the next TID seen by the worker is lower, the scan
561 : : * must have wrapped around.
562 : : */
563 [ # # ]: 0 : if (ItemPointerCompare(tid, &buildstate->tid) < 0)
564 : 0 : ginFlushBuildState(buildstate, index);
565 : :
566 : : /* remember the TID we're about to process */
567 : 0 : buildstate->tid = *tid;
568 : :
569 [ # # ]: 0 : for (i = 0; i < buildstate->ginstate.origTupdesc->natts; i++)
570 : 0 : ginHeapTupleBulkInsert(buildstate, (OffsetNumber) (i + 1),
571 : 0 : values[i], isnull[i], tid);
572 : :
573 : : /*
574 : : * If we've maxed out our available memory, dump everything to the
575 : : * tuplesort. We use half the per-worker fraction of maintenance_work_mem,
576 : : * the other half is used for the tuplesort.
577 : : */
578 [ # # ]: 0 : if (buildstate->accum.allocatedMemory >= buildstate->work_mem * (Size) 1024)
579 : 0 : ginFlushBuildState(buildstate, index);
580 : :
581 : 0 : MemoryContextSwitchTo(oldCtx);
582 : 0 : }
583 : :
584 : : IndexBuildResult *
3572 tgl@sss.pgh.pa.us 585 :CBC 214 : ginbuild(Relation heap, Relation index, IndexInfo *indexInfo)
586 : : {
587 : : IndexBuildResult *result;
588 : : double reltuples;
589 : : GinBuildState buildstate;
239 tomas.vondra@postgre 590 : 214 : GinBuildState *state = &buildstate;
591 : : Buffer RootBuffer,
592 : : MetaBuffer;
593 : : ItemPointerData *list;
594 : : Datum key;
595 : : GinNullCategory category;
596 : : uint32 nlist;
597 : : MemoryContext oldCtx;
598 : : OffsetNumber attnum;
599 : :
7119 teodor@sigaev.ru 600 [ - + ]: 214 : if (RelationGetNumberOfBlocks(index) != 0)
7119 teodor@sigaev.ru 601 [ # # ]:UBC 0 : elog(ERROR, "index \"%s\" already contains data",
602 : : RelationGetRelationName(index));
603 : :
7119 teodor@sigaev.ru 604 :CBC 214 : initGinState(&buildstate.ginstate, index);
5490 tgl@sss.pgh.pa.us 605 : 214 : buildstate.indtuples = 0;
606 : 214 : memset(&buildstate.buildStats, 0, sizeof(GinStatsData));
607 : :
608 : : /* Initialize fields for parallel build too. */
239 tomas.vondra@postgre 609 : 214 : buildstate.bs_numtuples = 0;
610 : 214 : buildstate.bs_reltuples = 0;
611 : 214 : buildstate.bs_leader = NULL;
612 : 214 : memset(&buildstate.tid, 0, sizeof(ItemPointerData));
613 : :
614 : : /* initialize the meta page */
6062 tgl@sss.pgh.pa.us 615 : 214 : MetaBuffer = GinNewBuffer(index);
616 : :
617 : : /* initialize the root page */
618 : 214 : RootBuffer = GinNewBuffer(index);
619 : :
7119 teodor@sigaev.ru 620 : 214 : START_CRIT_SECTION();
6062 tgl@sss.pgh.pa.us 621 : 214 : GinInitMetabuffer(MetaBuffer);
622 : 214 : MarkBufferDirty(MetaBuffer);
623 : 214 : GinInitBuffer(RootBuffer, GIN_LEAF);
624 : 214 : MarkBufferDirty(RootBuffer);
625 : :
626 : :
627 : 214 : UnlockReleaseBuffer(MetaBuffer);
628 : 214 : UnlockReleaseBuffer(RootBuffer);
7119 teodor@sigaev.ru 629 [ - + ]: 214 : END_CRIT_SECTION();
630 : :
631 : : /* count the root as first entry page */
5490 tgl@sss.pgh.pa.us 632 : 214 : buildstate.buildStats.nEntryPages++;
633 : :
634 : : /*
635 : : * create a temporary memory context that is used to hold data not yet
636 : : * dumped out to the index
637 : : */
7119 teodor@sigaev.ru 638 : 214 : buildstate.tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
639 : : "Gin build temporary context",
640 : : ALLOCSET_DEFAULT_SIZES);
641 : :
642 : : /*
643 : : * create a temporary memory context that is used for calling
644 : : * ginExtractEntries(), and can be reset after each tuple
645 : : */
3866 tgl@sss.pgh.pa.us 646 : 214 : buildstate.funcCtx = AllocSetContextCreate(CurrentMemoryContext,
647 : : "Gin build temporary context for user-defined function",
648 : : ALLOCSET_DEFAULT_SIZES);
649 : :
7119 teodor@sigaev.ru 650 : 214 : buildstate.accum.ginstate = &buildstate.ginstate;
6964 bruce@momjian.us 651 : 214 : ginInitBA(&buildstate.accum);
652 : :
653 : : /* Report table scan phase started */
239 tomas.vondra@postgre 654 : 214 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
655 : : PROGRESS_GIN_PHASE_INDEXBUILD_TABLESCAN);
656 : :
657 : : /*
658 : : * Attempt to launch parallel worker scan when required
659 : : *
660 : : * XXX plan_create_index_workers makes the number of workers dependent on
661 : : * maintenance_work_mem, requiring 32MB for each worker. For GIN that's
662 : : * reasonable too, because we sort the data just like btree. It does
663 : : * ignore the memory used to accumulate data in memory (set by work_mem),
664 : : * but there is no way to communicate that to plan_create_index_workers.
665 : : */
666 [ - + ]: 214 : if (indexInfo->ii_ParallelWorkers > 0)
239 tomas.vondra@postgre 667 :UBC 0 : _gin_begin_parallel(state, heap, index, indexInfo->ii_Concurrent,
668 : : indexInfo->ii_ParallelWorkers);
669 : :
670 : : /*
671 : : * If parallel build requested and at least one worker process was
672 : : * successfully launched, set up coordination state, wait for workers to
673 : : * complete. Then read all tuples from the shared tuplesort and insert
674 : : * them into the index.
675 : : *
676 : : * In serial mode, simply scan the table and build the index one index
677 : : * tuple at a time.
678 : : */
239 tomas.vondra@postgre 679 [ - + ]:CBC 214 : if (state->bs_leader)
680 : : {
681 : : SortCoordinate coordinate;
682 : :
239 tomas.vondra@postgre 683 :UBC 0 : coordinate = (SortCoordinate) palloc0(sizeof(SortCoordinateData));
684 : 0 : coordinate->isWorker = false;
685 : 0 : coordinate->nParticipants =
686 : 0 : state->bs_leader->nparticipanttuplesorts;
687 : 0 : coordinate->sharedsort = state->bs_leader->sharedsort;
688 : :
689 : : /*
690 : : * Begin leader tuplesort.
691 : : *
692 : : * In cases where parallelism is involved, the leader receives the
693 : : * same share of maintenance_work_mem as a serial sort (it is
694 : : * generally treated in the same way as a serial sort once we return).
695 : : * Parallel worker Tuplesortstates will have received only a fraction
696 : : * of maintenance_work_mem, though.
697 : : *
698 : : * We rely on the lifetime of the Leader Tuplesortstate almost not
699 : : * overlapping with any worker Tuplesortstate's lifetime. There may
700 : : * be some small overlap, but that's okay because we rely on leader
701 : : * Tuplesortstate only allocating a small, fixed amount of memory
702 : : * here. When its tuplesort_performsort() is called (by our caller),
703 : : * and significant amounts of memory are likely to be used, all
704 : : * workers must have already freed almost all memory held by their
705 : : * Tuplesortstates (they are about to go away completely, too). The
706 : : * overall effect is that maintenance_work_mem always represents an
707 : : * absolute high watermark on the amount of memory used by a CREATE
708 : : * INDEX operation, regardless of the use of parallelism or any other
709 : : * factor.
710 : : */
711 : 0 : state->bs_sortstate =
712 : 0 : tuplesort_begin_index_gin(heap, index,
713 : : maintenance_work_mem, coordinate,
714 : : TUPLESORT_NONE);
715 : :
716 : : /* scan the relation in parallel and merge per-worker results */
717 : 0 : reltuples = _gin_parallel_merge(state);
718 : :
719 : 0 : _gin_end_parallel(state->bs_leader, state);
720 : : }
721 : : else /* no parallel index build */
722 : : {
723 : : /*
724 : : * Do the heap scan. We disallow sync scan here because
725 : : * dataPlaceToPage prefers to receive tuples in TID order.
726 : : */
239 tomas.vondra@postgre 727 :CBC 214 : reltuples = table_index_build_scan(heap, index, indexInfo, false, true,
728 : : ginBuildCallback, &buildstate, NULL);
729 : :
730 : : /* dump remaining entries to the index */
731 : 214 : oldCtx = MemoryContextSwitchTo(buildstate.tmpCtx);
732 : 214 : ginBeginBAScan(&buildstate.accum);
733 : 76686 : while ((list = ginGetBAEntry(&buildstate.accum,
734 [ + + ]: 76686 : &attnum, &key, &category, &nlist)) != NULL)
735 : : {
736 : : /* there could be many entries, so be willing to abort here */
737 [ - + ]: 76472 : CHECK_FOR_INTERRUPTS();
738 : 76472 : ginEntryInsert(&buildstate.ginstate, attnum, key, category,
739 : : list, nlist, &buildstate.buildStats);
740 : : }
741 : 214 : MemoryContextSwitchTo(oldCtx);
742 : : }
743 : :
3866 tgl@sss.pgh.pa.us 744 : 214 : MemoryContextDelete(buildstate.funcCtx);
7119 teodor@sigaev.ru 745 : 214 : MemoryContextDelete(buildstate.tmpCtx);
746 : :
747 : : /*
748 : : * Update metapage stats
749 : : */
5490 tgl@sss.pgh.pa.us 750 : 214 : buildstate.buildStats.nTotalPages = RelationGetNumberOfBlocks(index);
2400 heikki.linnakangas@i 751 : 214 : ginUpdateStats(index, &buildstate.buildStats, true);
752 : :
753 : : /*
754 : : * We didn't write WAL records as we built the index, so if WAL-logging is
755 : : * required, write all pages to the WAL now.
756 : : */
757 [ + + + + : 214 : if (RelationNeedsWAL(index))
+ + + + ]
758 : : {
759 : 143 : log_newpage_range(index, MAIN_FORKNUM,
760 : : 0, RelationGetNumberOfBlocks(index),
761 : : true);
762 : : }
763 : :
764 : : /*
765 : : * Return statistics
766 : : */
7111 tgl@sss.pgh.pa.us 767 : 214 : result = (IndexBuildResult *) palloc(sizeof(IndexBuildResult));
768 : :
769 : 214 : result->heap_tuples = reltuples;
770 : 214 : result->index_tuples = buildstate.indtuples;
771 : :
3572 772 : 214 : return result;
773 : : }
774 : :
775 : : /*
776 : : * ginbuildempty() -- build an empty gin index in the initialization fork
777 : : */
778 : : void
779 : 3 : ginbuildempty(Relation index)
780 : : {
781 : : Buffer RootBuffer,
782 : : MetaBuffer;
783 : :
784 : : /* An empty GIN index has two pages. */
797 tmunro@postgresql.or 785 : 3 : MetaBuffer = ExtendBufferedRel(BMR_REL(index), INIT_FORKNUM, NULL,
786 : : EB_LOCK_FIRST | EB_SKIP_EXTENSION_LOCK);
787 : 3 : RootBuffer = ExtendBufferedRel(BMR_REL(index), INIT_FORKNUM, NULL,
788 : : EB_LOCK_FIRST | EB_SKIP_EXTENSION_LOCK);
789 : :
790 : : /* Initialize and xlog metabuffer and root buffer. */
5417 rhaas@postgresql.org 791 : 3 : START_CRIT_SECTION();
792 : 3 : GinInitMetabuffer(MetaBuffer);
793 : 3 : MarkBufferDirty(MetaBuffer);
2917 tgl@sss.pgh.pa.us 794 : 3 : log_newpage_buffer(MetaBuffer, true);
5417 rhaas@postgresql.org 795 : 3 : GinInitBuffer(RootBuffer, GIN_LEAF);
796 : 3 : MarkBufferDirty(RootBuffer);
4346 heikki.linnakangas@i 797 : 3 : log_newpage_buffer(RootBuffer, false);
5417 rhaas@postgresql.org 798 [ - + ]: 3 : END_CRIT_SECTION();
799 : :
800 : : /* Unlock and release the buffers. */
801 : 3 : UnlockReleaseBuffer(MetaBuffer);
802 : 3 : UnlockReleaseBuffer(RootBuffer);
803 : 3 : }
804 : :
805 : : /*
806 : : * Insert index entries for a single indexable item during "normal"
807 : : * (non-fast-update) insertion
808 : : */
809 : : static void
5408 tgl@sss.pgh.pa.us 810 : 26127 : ginHeapTupleInsert(GinState *ginstate, OffsetNumber attnum,
811 : : Datum value, bool isNull,
812 : : ItemPointer item)
813 : : {
814 : : Datum *entries;
815 : : GinNullCategory *categories;
816 : : int32 i,
817 : : nentries;
818 : :
819 : 26127 : entries = ginExtractEntries(ginstate, attnum, value, isNull,
820 : : &nentries, &categories);
821 : :
6964 bruce@momjian.us 822 [ + + ]: 162477 : for (i = 0; i < nentries; i++)
5408 tgl@sss.pgh.pa.us 823 : 136358 : ginEntryInsert(ginstate, attnum, entries[i], categories[i],
824 : : item, 1, NULL);
7119 teodor@sigaev.ru 825 : 26119 : }
826 : :
827 : : bool
3572 tgl@sss.pgh.pa.us 828 : 159023 : gininsert(Relation index, Datum *values, bool *isnull,
829 : : ItemPointer ht_ctid, Relation heapRel,
830 : : IndexUniqueCheck checkUnique,
831 : : bool indexUnchanged,
832 : : IndexInfo *indexInfo)
833 : : {
3183 834 : 159023 : GinState *ginstate = (GinState *) indexInfo->ii_AmCache;
835 : : MemoryContext oldCtx;
836 : : MemoryContext insertCtx;
837 : : int i;
838 : :
839 : : /* Initialize GinState cache if first call in this statement */
840 [ + + ]: 159023 : if (ginstate == NULL)
841 : : {
842 : 523 : oldCtx = MemoryContextSwitchTo(indexInfo->ii_Context);
843 : 523 : ginstate = (GinState *) palloc(sizeof(GinState));
844 : 523 : initGinState(ginstate, index);
334 peter@eisentraut.org 845 : 523 : indexInfo->ii_AmCache = ginstate;
3183 tgl@sss.pgh.pa.us 846 : 523 : MemoryContextSwitchTo(oldCtx);
847 : : }
848 : :
7119 teodor@sigaev.ru 849 : 159023 : insertCtx = AllocSetContextCreate(CurrentMemoryContext,
850 : : "Gin insert temporary context",
851 : : ALLOCSET_DEFAULT_SIZES);
852 : :
853 : 159023 : oldCtx = MemoryContextSwitchTo(insertCtx);
854 : :
5983 bruce@momjian.us 855 [ + - - + : 159023 : if (GinGetUseFastUpdate(index))
+ + + + ]
6062 tgl@sss.pgh.pa.us 856 : 132893 : {
857 : : GinTupleCollector collector;
858 : :
859 : 132896 : memset(&collector, 0, sizeof(GinTupleCollector));
860 : :
3183 861 [ + + ]: 325831 : for (i = 0; i < ginstate->origTupdesc->natts; i++)
862 : 192935 : ginHeapTupleFastCollect(ginstate, &collector,
5408 863 : 192935 : (OffsetNumber) (i + 1),
864 : 192935 : values[i], isnull[i],
865 : : ht_ctid);
866 : :
3183 867 : 132896 : ginHeapTupleFastInsert(ginstate, &collector);
868 : : }
869 : : else
870 : : {
871 [ + + ]: 52246 : for (i = 0; i < ginstate->origTupdesc->natts; i++)
872 : 26127 : ginHeapTupleInsert(ginstate, (OffsetNumber) (i + 1),
5408 873 : 26127 : values[i], isnull[i],
874 : : ht_ctid);
875 : : }
876 : :
7119 teodor@sigaev.ru 877 : 159012 : MemoryContextSwitchTo(oldCtx);
878 : 159012 : MemoryContextDelete(insertCtx);
879 : :
3572 tgl@sss.pgh.pa.us 880 : 159012 : return false;
881 : : }
882 : :
883 : : /*
884 : : * Create parallel context, and launch workers for leader.
885 : : *
886 : : * buildstate argument should be initialized (with the exception of the
887 : : * tuplesort states, which may later be created based on shared
888 : : * state initially set up here).
889 : : *
890 : : * isconcurrent indicates if operation is CREATE INDEX CONCURRENTLY.
891 : : *
892 : : * request is the target number of parallel worker processes to launch.
893 : : *
894 : : * Sets buildstate's GinLeader, which caller must use to shut down parallel
895 : : * mode by passing it to _gin_end_parallel() at the very end of its index
896 : : * build. If not even a single worker process can be launched, this is
897 : : * never set, and caller should proceed with a serial index build.
898 : : */
899 : : static void
239 tomas.vondra@postgre 900 :UBC 0 : _gin_begin_parallel(GinBuildState *buildstate, Relation heap, Relation index,
901 : : bool isconcurrent, int request)
902 : : {
903 : : ParallelContext *pcxt;
904 : : int scantuplesortstates;
905 : : Snapshot snapshot;
906 : : Size estginshared;
907 : : Size estsort;
908 : : GinBuildShared *ginshared;
909 : : Sharedsort *sharedsort;
910 : 0 : GinLeader *ginleader = (GinLeader *) palloc0(sizeof(GinLeader));
911 : : WalUsage *walusage;
912 : : BufferUsage *bufferusage;
913 : 0 : bool leaderparticipates = true;
914 : : int querylen;
915 : :
916 : : #ifdef DISABLE_LEADER_PARTICIPATION
917 : : leaderparticipates = false;
918 : : #endif
919 : :
920 : : /*
921 : : * Enter parallel mode, and create context for parallel build of gin index
922 : : */
923 : 0 : EnterParallelMode();
924 [ # # ]: 0 : Assert(request > 0);
925 : 0 : pcxt = CreateParallelContext("postgres", "_gin_parallel_build_main",
926 : : request);
927 : :
928 [ # # ]: 0 : scantuplesortstates = leaderparticipates ? request + 1 : request;
929 : :
930 : : /*
931 : : * Prepare for scan of the base relation. In a normal index build, we use
932 : : * SnapshotAny because we must retrieve all tuples and do our own time
933 : : * qual checks (because we have to index RECENTLY_DEAD tuples). In a
934 : : * concurrent build, we take a regular MVCC snapshot and index whatever's
935 : : * live according to that.
936 : : */
937 [ # # ]: 0 : if (!isconcurrent)
938 : 0 : snapshot = SnapshotAny;
939 : : else
940 : 0 : snapshot = RegisterSnapshot(GetTransactionSnapshot());
941 : :
942 : : /*
943 : : * Estimate size for our own PARALLEL_KEY_GIN_SHARED workspace.
944 : : */
945 : 0 : estginshared = _gin_parallel_estimate_shared(heap, snapshot);
946 : 0 : shm_toc_estimate_chunk(&pcxt->estimator, estginshared);
947 : 0 : estsort = tuplesort_estimate_shared(scantuplesortstates);
948 : 0 : shm_toc_estimate_chunk(&pcxt->estimator, estsort);
949 : :
950 : 0 : shm_toc_estimate_keys(&pcxt->estimator, 2);
951 : :
952 : : /*
953 : : * Estimate space for WalUsage and BufferUsage -- PARALLEL_KEY_WAL_USAGE
954 : : * and PARALLEL_KEY_BUFFER_USAGE.
955 : : *
956 : : * If there are no extensions loaded that care, we could skip this. We
957 : : * have no way of knowing whether anyone's looking at pgWalUsage or
958 : : * pgBufferUsage, so do it unconditionally.
959 : : */
960 : 0 : shm_toc_estimate_chunk(&pcxt->estimator,
961 : : mul_size(sizeof(WalUsage), pcxt->nworkers));
962 : 0 : shm_toc_estimate_keys(&pcxt->estimator, 1);
963 : 0 : shm_toc_estimate_chunk(&pcxt->estimator,
964 : : mul_size(sizeof(BufferUsage), pcxt->nworkers));
965 : 0 : shm_toc_estimate_keys(&pcxt->estimator, 1);
966 : :
967 : : /* Finally, estimate PARALLEL_KEY_QUERY_TEXT space */
968 [ # # ]: 0 : if (debug_query_string)
969 : : {
970 : 0 : querylen = strlen(debug_query_string);
971 : 0 : shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
972 : 0 : shm_toc_estimate_keys(&pcxt->estimator, 1);
973 : : }
974 : : else
975 : 0 : querylen = 0; /* keep compiler quiet */
976 : :
977 : : /* Everyone's had a chance to ask for space, so now create the DSM */
978 : 0 : InitializeParallelDSM(pcxt);
979 : :
980 : : /* If no DSM segment was available, back out (do serial build) */
981 [ # # ]: 0 : if (pcxt->seg == NULL)
982 : : {
983 [ # # # # ]: 0 : if (IsMVCCSnapshot(snapshot))
984 : 0 : UnregisterSnapshot(snapshot);
985 : 0 : DestroyParallelContext(pcxt);
986 : 0 : ExitParallelMode();
987 : 0 : return;
988 : : }
989 : :
990 : : /* Store shared build state, for which we reserved space */
991 : 0 : ginshared = (GinBuildShared *) shm_toc_allocate(pcxt->toc, estginshared);
992 : : /* Initialize immutable state */
993 : 0 : ginshared->heaprelid = RelationGetRelid(heap);
994 : 0 : ginshared->indexrelid = RelationGetRelid(index);
995 : 0 : ginshared->isconcurrent = isconcurrent;
996 : 0 : ginshared->scantuplesortstates = scantuplesortstates;
997 : :
998 : 0 : ConditionVariableInit(&ginshared->workersdonecv);
999 : 0 : SpinLockInit(&ginshared->mutex);
1000 : :
1001 : : /* Initialize mutable state */
1002 : 0 : ginshared->nparticipantsdone = 0;
1003 : 0 : ginshared->reltuples = 0.0;
1004 : 0 : ginshared->indtuples = 0.0;
1005 : :
1006 : 0 : table_parallelscan_initialize(heap,
1007 : : ParallelTableScanFromGinBuildShared(ginshared),
1008 : : snapshot);
1009 : :
1010 : : /*
1011 : : * Store shared tuplesort-private state, for which we reserved space.
1012 : : * Then, initialize opaque state using tuplesort routine.
1013 : : */
1014 : 0 : sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
1015 : 0 : tuplesort_initialize_shared(sharedsort, scantuplesortstates,
1016 : : pcxt->seg);
1017 : :
1018 : 0 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_GIN_SHARED, ginshared);
1019 : 0 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_TUPLESORT, sharedsort);
1020 : :
1021 : : /* Store query string for workers */
1022 [ # # ]: 0 : if (debug_query_string)
1023 : : {
1024 : : char *sharedquery;
1025 : :
1026 : 0 : sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
1027 : 0 : memcpy(sharedquery, debug_query_string, querylen + 1);
1028 : 0 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_QUERY_TEXT, sharedquery);
1029 : : }
1030 : :
1031 : : /*
1032 : : * Allocate space for each worker's WalUsage and BufferUsage; no need to
1033 : : * initialize.
1034 : : */
1035 : 0 : walusage = shm_toc_allocate(pcxt->toc,
1036 : 0 : mul_size(sizeof(WalUsage), pcxt->nworkers));
1037 : 0 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_WAL_USAGE, walusage);
1038 : 0 : bufferusage = shm_toc_allocate(pcxt->toc,
1039 : 0 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
1040 : 0 : shm_toc_insert(pcxt->toc, PARALLEL_KEY_BUFFER_USAGE, bufferusage);
1041 : :
1042 : : /* Launch workers, saving status for leader/caller */
1043 : 0 : LaunchParallelWorkers(pcxt);
1044 : 0 : ginleader->pcxt = pcxt;
1045 : 0 : ginleader->nparticipanttuplesorts = pcxt->nworkers_launched;
1046 [ # # ]: 0 : if (leaderparticipates)
1047 : 0 : ginleader->nparticipanttuplesorts++;
1048 : 0 : ginleader->ginshared = ginshared;
1049 : 0 : ginleader->sharedsort = sharedsort;
1050 : 0 : ginleader->snapshot = snapshot;
1051 : 0 : ginleader->walusage = walusage;
1052 : 0 : ginleader->bufferusage = bufferusage;
1053 : :
1054 : : /* If no workers were successfully launched, back out (do serial build) */
1055 [ # # ]: 0 : if (pcxt->nworkers_launched == 0)
1056 : : {
1057 : 0 : _gin_end_parallel(ginleader, NULL);
1058 : 0 : return;
1059 : : }
1060 : :
1061 : : /* Save leader state now that it's clear build will be parallel */
1062 : 0 : buildstate->bs_leader = ginleader;
1063 : :
1064 : : /* Join heap scan ourselves */
1065 [ # # ]: 0 : if (leaderparticipates)
1066 : 0 : _gin_leader_participate_as_worker(buildstate, heap, index);
1067 : :
1068 : : /*
1069 : : * Caller needs to wait for all launched workers when we return. Make
1070 : : * sure that the failure-to-start case will not hang forever.
1071 : : */
1072 : 0 : WaitForParallelWorkersToAttach(pcxt);
1073 : : }
1074 : :
1075 : : /*
1076 : : * Shut down workers, destroy parallel context, and end parallel mode.
1077 : : */
1078 : : static void
1079 : 0 : _gin_end_parallel(GinLeader *ginleader, GinBuildState *state)
1080 : : {
1081 : : int i;
1082 : :
1083 : : /* Shutdown worker processes */
1084 : 0 : WaitForParallelWorkersToFinish(ginleader->pcxt);
1085 : :
1086 : : /*
1087 : : * Next, accumulate WAL usage. (This must wait for the workers to finish,
1088 : : * or we might get incomplete data.)
1089 : : */
1090 [ # # ]: 0 : for (i = 0; i < ginleader->pcxt->nworkers_launched; i++)
1091 : 0 : InstrAccumParallelQuery(&ginleader->bufferusage[i], &ginleader->walusage[i]);
1092 : :
1093 : : /* Free last reference to MVCC snapshot, if one was used */
1094 [ # # # # ]: 0 : if (IsMVCCSnapshot(ginleader->snapshot))
1095 : 0 : UnregisterSnapshot(ginleader->snapshot);
1096 : 0 : DestroyParallelContext(ginleader->pcxt);
1097 : 0 : ExitParallelMode();
1098 : 0 : }
1099 : :
1100 : : /*
1101 : : * Within leader, wait for end of heap scan.
1102 : : *
1103 : : * When called, parallel heap scan started by _gin_begin_parallel() will
1104 : : * already be underway within worker processes (when leader participates
1105 : : * as a worker, we should end up here just as workers are finishing).
1106 : : *
1107 : : * Returns the total number of heap tuples scanned.
1108 : : */
1109 : : static double
1110 : 0 : _gin_parallel_heapscan(GinBuildState *state)
1111 : : {
1112 : 0 : GinBuildShared *ginshared = state->bs_leader->ginshared;
1113 : : int nparticipanttuplesorts;
1114 : :
1115 : 0 : nparticipanttuplesorts = state->bs_leader->nparticipanttuplesorts;
1116 : : for (;;)
1117 : : {
1118 [ # # ]: 0 : SpinLockAcquire(&ginshared->mutex);
1119 [ # # ]: 0 : if (ginshared->nparticipantsdone == nparticipanttuplesorts)
1120 : : {
1121 : : /* copy the data into leader state */
1122 : 0 : state->bs_reltuples = ginshared->reltuples;
1123 : 0 : state->bs_numtuples = ginshared->indtuples;
1124 : :
1125 : 0 : SpinLockRelease(&ginshared->mutex);
1126 : 0 : break;
1127 : : }
1128 : 0 : SpinLockRelease(&ginshared->mutex);
1129 : :
1130 : 0 : ConditionVariableSleep(&ginshared->workersdonecv,
1131 : : WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN);
1132 : : }
1133 : :
1134 : 0 : ConditionVariableCancelSleep();
1135 : :
1136 : 0 : return state->bs_reltuples;
1137 : : }
1138 : :
1139 : : /*
1140 : : * Buffer used to accumulate TIDs from multiple GinTuples for the same key
1141 : : * (we read these from the tuplesort, sorted by the key).
1142 : : *
1143 : : * This is similar to BuildAccumulator in that it's used to collect TIDs
1144 : : * in memory before inserting them into the index, but it's much simpler
1145 : : * as it only deals with a single index key at a time.
1146 : : *
1147 : : * When adding TIDs to the buffer, we make sure to keep them sorted, both
1148 : : * during the initial table scan (and detecting when the scan wraps around),
1149 : : * and during merging (where we do mergesort).
1150 : : */
1151 : : typedef struct GinBuffer
1152 : : {
1153 : : OffsetNumber attnum;
1154 : : GinNullCategory category;
1155 : : Datum key; /* 0 if no key (and keylen == 0) */
1156 : : Size keylen; /* number of bytes (not typlen) */
1157 : :
1158 : : /* type info */
1159 : : int16 typlen;
1160 : : bool typbyval;
1161 : :
1162 : : /* Number of TIDs to collect before attempt to write some out. */
1163 : : int maxitems;
1164 : :
1165 : : /* array of TID values */
1166 : : int nitems;
1167 : : int nfrozen;
1168 : : SortSupport ssup; /* for sorting/comparing keys */
1169 : : ItemPointerData *items;
1170 : : } GinBuffer;
1171 : :
1172 : : /*
1173 : : * Check that TID array contains valid values, and that it's sorted (if we
1174 : : * expect it to be).
1175 : : */
1176 : : static void
1177 : 0 : AssertCheckItemPointers(GinBuffer *buffer)
1178 : : {
1179 : : #ifdef USE_ASSERT_CHECKING
1180 : : /* we should not have a buffer with no TIDs to sort */
1181 [ # # ]: 0 : Assert(buffer->items != NULL);
1182 [ # # ]: 0 : Assert(buffer->nitems > 0);
1183 : :
1184 [ # # ]: 0 : for (int i = 0; i < buffer->nitems; i++)
1185 : : {
1186 [ # # ]: 0 : Assert(ItemPointerIsValid(&buffer->items[i]));
1187 : :
1188 : : /* don't check ordering for the first TID item */
1189 [ # # ]: 0 : if (i == 0)
1190 : 0 : continue;
1191 : :
1192 [ # # ]: 0 : Assert(ItemPointerCompare(&buffer->items[i - 1], &buffer->items[i]) < 0);
1193 : : }
1194 : : #endif
1195 : 0 : }
1196 : :
1197 : : /*
1198 : : * GinBuffer checks
1199 : : *
1200 : : * Make sure the nitems/items fields are consistent (either the array is empty
1201 : : * or not empty, the fields need to agree). If there are items, check ordering.
1202 : : */
1203 : : static void
1204 : 0 : AssertCheckGinBuffer(GinBuffer *buffer)
1205 : : {
1206 : : #ifdef USE_ASSERT_CHECKING
1207 : : /* if we have any items, the array must exist */
1208 [ # # # # ]: 0 : Assert(!((buffer->nitems > 0) && (buffer->items == NULL)));
1209 : :
1210 : : /*
1211 : : * The buffer may be empty, in which case we must not call the check of
1212 : : * item pointers, because that assumes non-emptiness.
1213 : : */
1214 [ # # ]: 0 : if (buffer->nitems == 0)
1215 : 0 : return;
1216 : :
1217 : : /* Make sure the item pointers are valid and sorted. */
1218 : 0 : AssertCheckItemPointers(buffer);
1219 : : #endif
1220 : : }
1221 : :
1222 : : /*
1223 : : * GinBufferInit
1224 : : * Initialize buffer to store tuples for a GIN index.
1225 : : *
1226 : : * Initialize the buffer used to accumulate TID for a single key at a time
1227 : : * (we process the data sorted), so we know when we received all data for
1228 : : * a given key.
1229 : : *
1230 : : * Initializes sort support procedures for all index attributes.
1231 : : */
1232 : : static GinBuffer *
1233 : 0 : GinBufferInit(Relation index)
1234 : : {
1235 : 0 : GinBuffer *buffer = palloc0(sizeof(GinBuffer));
1236 : : int i,
1237 : : nKeys;
1238 : 0 : TupleDesc desc = RelationGetDescr(index);
1239 : :
1240 : : /*
1241 : : * How many items can we fit into the memory limit? We don't want to end
1242 : : * with too many TIDs. and 64kB seems more than enough. But maybe this
1243 : : * should be tied to maintenance_work_mem or something like that?
1244 : : */
238 1245 : 0 : buffer->maxitems = (64 * 1024L) / sizeof(ItemPointerData);
1246 : :
239 1247 : 0 : nKeys = IndexRelationGetNumberOfKeyAttributes(index);
1248 : :
1249 : 0 : buffer->ssup = palloc0(sizeof(SortSupportData) * nKeys);
1250 : :
1251 : : /*
1252 : : * Lookup ordering operator for the index key data type, and initialize
1253 : : * the sort support function.
1254 : : */
1255 [ # # ]: 0 : for (i = 0; i < nKeys; i++)
1256 : : {
1257 : : Oid cmpFunc;
1258 : 0 : SortSupport sortKey = &buffer->ssup[i];
1259 : 0 : Form_pg_attribute att = TupleDescAttr(desc, i);
1260 : :
1261 : 0 : sortKey->ssup_cxt = CurrentMemoryContext;
1262 : 0 : sortKey->ssup_collation = index->rd_indcollation[i];
1263 : :
1264 [ # # ]: 0 : if (!OidIsValid(sortKey->ssup_collation))
1265 : 0 : sortKey->ssup_collation = DEFAULT_COLLATION_OID;
1266 : :
1267 : 0 : sortKey->ssup_nulls_first = false;
1268 : 0 : sortKey->ssup_attno = i + 1;
1269 : 0 : sortKey->abbreviate = false;
1270 : :
1271 [ # # ]: 0 : Assert(sortKey->ssup_attno != 0);
1272 : :
1273 : : /*
1274 : : * If the compare proc isn't specified in the opclass definition, look
1275 : : * up the index key type's default btree comparator.
1276 : : */
1277 : 0 : cmpFunc = index_getprocid(index, i + 1, GIN_COMPARE_PROC);
1278 [ # # ]: 0 : if (cmpFunc == InvalidOid)
1279 : : {
1280 : : TypeCacheEntry *typentry;
1281 : :
1282 : 0 : typentry = lookup_type_cache(att->atttypid,
1283 : : TYPECACHE_CMP_PROC_FINFO);
1284 [ # # ]: 0 : if (!OidIsValid(typentry->cmp_proc_finfo.fn_oid))
1285 [ # # ]: 0 : ereport(ERROR,
1286 : : (errcode(ERRCODE_UNDEFINED_FUNCTION),
1287 : : errmsg("could not identify a comparison function for type %s",
1288 : : format_type_be(att->atttypid))));
1289 : :
1290 : 0 : cmpFunc = typentry->cmp_proc_finfo.fn_oid;
1291 : : }
1292 : :
1293 : 0 : PrepareSortSupportComparisonShim(cmpFunc, sortKey);
1294 : : }
1295 : :
1296 : 0 : return buffer;
1297 : : }
1298 : :
1299 : : /* Is the buffer empty, i.e. has no TID values in the array? */
1300 : : static bool
1301 : 0 : GinBufferIsEmpty(GinBuffer *buffer)
1302 : : {
1303 : 0 : return (buffer->nitems == 0);
1304 : : }
1305 : :
1306 : : /*
1307 : : * GinBufferKeyEquals
1308 : : * Can the buffer store TIDs for the provided GIN tuple (same key)?
1309 : : *
1310 : : * Compare if the tuple matches the already accumulated data in the GIN
1311 : : * buffer. Compare scalar fields first, before the actual key.
1312 : : *
1313 : : * Returns true if the key matches, and the TID belongs to the buffer, or
1314 : : * false if the key does not match.
1315 : : */
1316 : : static bool
1317 : 0 : GinBufferKeyEquals(GinBuffer *buffer, GinTuple *tup)
1318 : : {
1319 : : int r;
1320 : : Datum tupkey;
1321 : :
1322 : 0 : AssertCheckGinBuffer(buffer);
1323 : :
1324 [ # # ]: 0 : if (tup->attrnum != buffer->attnum)
1325 : 0 : return false;
1326 : :
1327 : : /* same attribute should have the same type info */
1328 [ # # ]: 0 : Assert(tup->typbyval == buffer->typbyval);
1329 [ # # ]: 0 : Assert(tup->typlen == buffer->typlen);
1330 : :
1331 [ # # ]: 0 : if (tup->category != buffer->category)
1332 : 0 : return false;
1333 : :
1334 : : /*
1335 : : * For NULL/empty keys, this means equality, for normal keys we need to
1336 : : * compare the actual key value.
1337 : : */
1338 [ # # ]: 0 : if (buffer->category != GIN_CAT_NORM_KEY)
1339 : 0 : return true;
1340 : :
1341 : : /*
1342 : : * For the tuple, get either the first sizeof(Datum) bytes for byval
1343 : : * types, or a pointer to the beginning of the data array.
1344 : : */
1345 [ # # ]: 0 : tupkey = (buffer->typbyval) ? *(Datum *) tup->data : PointerGetDatum(tup->data);
1346 : :
1347 : 0 : r = ApplySortComparator(buffer->key, false,
1348 : : tupkey, false,
1349 : 0 : &buffer->ssup[buffer->attnum - 1]);
1350 : :
1351 : 0 : return (r == 0);
1352 : : }
1353 : :
1354 : : /*
1355 : : * GinBufferShouldTrim
1356 : : * Should we trim the list of item pointers?
1357 : : *
1358 : : * By trimming we understand writing out and removing the tuple IDs that
1359 : : * we know can't change by future merges. We can deduce the TID up to which
1360 : : * this is guaranteed from the "first" TID in each GIN tuple, which provides
1361 : : * a "horizon" (for a given key) thanks to the sort.
1362 : : *
1363 : : * We don't want to do this too often - compressing longer TID lists is more
1364 : : * efficient. But we also don't want to accumulate too many TIDs, for two
1365 : : * reasons. First, it consumes memory and we might exceed maintenance_work_mem
1366 : : * (or whatever limit applies), even if that's unlikely because TIDs are very
1367 : : * small so we can fit a lot of them. Second, and more importantly, long TID
1368 : : * lists are an issue if the scan wraps around, because a key may get a very
1369 : : * wide list (with min/max TID for that key), forcing "full" mergesorts for
1370 : : * every list merged into it (instead of the efficient append).
1371 : : *
1372 : : * So we look at two things when deciding if to trim - if the resulting list
1373 : : * (after adding TIDs from the new tuple) would be too long, and if there is
1374 : : * enough TIDs to trim (with values less than "first" TID from the new tuple),
1375 : : * we do the trim. By enough we mean at least 128 TIDs (mostly an arbitrary
1376 : : * number).
1377 : : */
1378 : : static bool
238 1379 : 0 : GinBufferShouldTrim(GinBuffer *buffer, GinTuple *tup)
1380 : : {
1381 : : /* not enough TIDs to trim (1024 is somewhat arbitrary number) */
1382 [ # # ]: 0 : if (buffer->nfrozen < 1024)
1383 : 0 : return false;
1384 : :
1385 : : /* no need to trim if we have not hit the memory limit yet */
1386 [ # # ]: 0 : if ((buffer->nitems + tup->nitems) < buffer->maxitems)
1387 : 0 : return false;
1388 : :
1389 : : /*
1390 : : * OK, we have enough frozen TIDs to flush, and we have hit the memory
1391 : : * limit, so it's time to write it out.
1392 : : */
1393 : 0 : return true;
1394 : : }
1395 : :
1396 : : /*
1397 : : * GinBufferStoreTuple
1398 : : * Add data (especially TID list) from a GIN tuple to the buffer.
1399 : : *
1400 : : * The buffer is expected to be empty (in which case it's initialized), or
1401 : : * having the same key. The TID values from the tuple are combined with the
1402 : : * stored values using a merge sort.
1403 : : *
1404 : : * The tuples (for the same key) are expected to be sorted by first TID. But
1405 : : * this does not guarantee the lists do not overlap, especially in the leader,
1406 : : * because the workers process interleaving data. There should be no overlaps
1407 : : * in a single worker - it could happen when the parallel scan wraps around,
1408 : : * but we detect that and flush the data (see ginBuildCallbackParallel).
1409 : : *
1410 : : * By sorting the GinTuple not only by key, but also by the first TID, we make
1411 : : * it more less likely the lists will overlap during merge. We merge them using
1412 : : * mergesort, but it's cheaper to just append one list to the other.
1413 : : *
1414 : : * How often can the lists overlap? There should be no overlaps in workers,
1415 : : * and in the leader we can see overlaps between lists built by different
1416 : : * workers. But the workers merge the items as much as possible, so there
1417 : : * should not be too many.
1418 : : */
1419 : : static void
239 1420 : 0 : GinBufferStoreTuple(GinBuffer *buffer, GinTuple *tup)
1421 : : {
1422 : : ItemPointerData *items;
1423 : : Datum key;
1424 : :
1425 : 0 : AssertCheckGinBuffer(buffer);
1426 : :
238 1427 : 0 : key = _gin_parse_tuple_key(tup);
1428 : 0 : items = _gin_parse_tuple_items(tup);
1429 : :
1430 : : /* if the buffer is empty, set the fields (and copy the key) */
239 1431 [ # # ]: 0 : if (GinBufferIsEmpty(buffer))
1432 : : {
1433 : 0 : buffer->category = tup->category;
1434 : 0 : buffer->keylen = tup->keylen;
1435 : 0 : buffer->attnum = tup->attrnum;
1436 : :
1437 : 0 : buffer->typlen = tup->typlen;
1438 : 0 : buffer->typbyval = tup->typbyval;
1439 : :
1440 [ # # ]: 0 : if (tup->category == GIN_CAT_NORM_KEY)
1441 : 0 : buffer->key = datumCopy(key, buffer->typbyval, buffer->typlen);
1442 : : else
1443 : 0 : buffer->key = (Datum) 0;
1444 : : }
1445 : :
1446 : : /*
1447 : : * Try freeze TIDs at the beginning of the list, i.e. exclude them from
1448 : : * the mergesort. We can do that with TIDs before the first TID in the new
1449 : : * tuple we're about to add into the buffer.
1450 : : *
1451 : : * We do this incrementally when adding data into the in-memory buffer,
1452 : : * and not later (e.g. when hitting a memory limit), because it allows us
1453 : : * to skip the frozen data during the mergesort, making it cheaper.
1454 : : */
1455 : :
1456 : : /*
1457 : : * Check if the last TID in the current list is frozen. This is the case
1458 : : * when merging non-overlapping lists, e.g. in each parallel worker.
1459 : : */
238 1460 [ # # # # ]: 0 : if ((buffer->nitems > 0) &&
1461 : 0 : (ItemPointerCompare(&buffer->items[buffer->nitems - 1],
1462 : : GinTupleGetFirst(tup)) == 0))
1463 : 0 : buffer->nfrozen = buffer->nitems;
1464 : :
1465 : : /*
1466 : : * Now find the last TID we know to be frozen, i.e. the last TID right
1467 : : * before the new GIN tuple.
1468 : : *
1469 : : * Start with the first not-yet-frozen tuple, and walk until we find the
1470 : : * first TID that's higher. If we already know the whole list is frozen
1471 : : * (i.e. nfrozen == nitems), this does nothing.
1472 : : *
1473 : : * XXX This might do a binary search for sufficiently long lists, but it
1474 : : * does not seem worth the complexity. Overlapping lists should be rare
1475 : : * common, TID comparisons are cheap, and we should quickly freeze most of
1476 : : * the list.
1477 : : */
1478 [ # # ]: 0 : for (int i = buffer->nfrozen; i < buffer->nitems; i++)
1479 : : {
1480 : : /* Is the TID after the first TID of the new tuple? Can't freeze. */
1481 [ # # ]: 0 : if (ItemPointerCompare(&buffer->items[i],
1482 : : GinTupleGetFirst(tup)) > 0)
1483 : 0 : break;
1484 : :
1485 : 0 : buffer->nfrozen++;
1486 : : }
1487 : :
1488 : : /* add the new TIDs into the buffer, combine using merge-sort */
1489 : : {
1490 : : int nnew;
1491 : : ItemPointer new;
1492 : :
1493 : : /*
1494 : : * Resize the array - we do this first, because we'll dereference the
1495 : : * first unfrozen TID, which would fail if the array is NULL. We'll
1496 : : * still pass 0 as number of elements in that array though.
1497 : : */
1498 [ # # ]: 0 : if (buffer->items == NULL)
1499 : 0 : buffer->items = palloc((buffer->nitems + tup->nitems) * sizeof(ItemPointerData));
1500 : : else
1501 : 0 : buffer->items = repalloc(buffer->items,
1502 : 0 : (buffer->nitems + tup->nitems) * sizeof(ItemPointerData));
1503 : :
192 michael@paquier.xyz 1504 : 0 : new = ginMergeItemPointers(&buffer->items[buffer->nfrozen], /* first unfrozen */
238 tomas.vondra@postgre 1505 : 0 : (buffer->nitems - buffer->nfrozen), /* num of unfrozen */
239 1506 : 0 : items, tup->nitems, &nnew);
1507 : :
238 1508 [ # # ]: 0 : Assert(nnew == (tup->nitems + (buffer->nitems - buffer->nfrozen)));
1509 : :
1510 : 0 : memcpy(&buffer->items[buffer->nfrozen], new,
1511 : : nnew * sizeof(ItemPointerData));
1512 : :
1513 : 0 : pfree(new);
1514 : :
1515 : 0 : buffer->nitems += tup->nitems;
1516 : :
239 1517 : 0 : AssertCheckItemPointers(buffer);
1518 : : }
1519 : :
1520 : : /* free the decompressed TID list */
238 1521 : 0 : pfree(items);
239 1522 : 0 : }
1523 : :
1524 : : /*
1525 : : * GinBufferReset
1526 : : * Reset the buffer into a state as if it contains no data.
1527 : : */
1528 : : static void
1529 : 0 : GinBufferReset(GinBuffer *buffer)
1530 : : {
1531 [ # # ]: 0 : Assert(!GinBufferIsEmpty(buffer));
1532 : :
1533 : : /* release byref values, do nothing for by-val ones */
1534 [ # # # # ]: 0 : if ((buffer->category == GIN_CAT_NORM_KEY) && !buffer->typbyval)
1535 : 0 : pfree(DatumGetPointer(buffer->key));
1536 : :
1537 : : /*
1538 : : * Not required, but makes it more likely to trigger NULL dereference if
1539 : : * using the value incorrectly, etc.
1540 : : */
1541 : 0 : buffer->key = (Datum) 0;
1542 : :
1543 : 0 : buffer->attnum = 0;
1544 : 0 : buffer->category = 0;
1545 : 0 : buffer->keylen = 0;
1546 : 0 : buffer->nitems = 0;
238 1547 : 0 : buffer->nfrozen = 0;
1548 : :
239 1549 : 0 : buffer->typlen = 0;
1550 : 0 : buffer->typbyval = 0;
1551 : 0 : }
1552 : :
1553 : : /*
1554 : : * GinBufferTrim
1555 : : * Discard the "frozen" part of the TID list (which should have been
1556 : : * written to disk/index before this call).
1557 : : */
1558 : : static void
238 1559 : 0 : GinBufferTrim(GinBuffer *buffer)
1560 : : {
1561 [ # # # # ]: 0 : Assert((buffer->nfrozen > 0) && (buffer->nfrozen <= buffer->nitems));
1562 : :
1563 : 0 : memmove(&buffer->items[0], &buffer->items[buffer->nfrozen],
1564 : 0 : sizeof(ItemPointerData) * (buffer->nitems - buffer->nfrozen));
1565 : :
1566 : 0 : buffer->nitems -= buffer->nfrozen;
1567 : 0 : buffer->nfrozen = 0;
1568 : 0 : }
1569 : :
1570 : : /*
1571 : : * GinBufferFree
1572 : : * Release memory associated with the GinBuffer (including TID array).
1573 : : */
1574 : : static void
239 1575 : 0 : GinBufferFree(GinBuffer *buffer)
1576 : : {
1577 [ # # ]: 0 : if (buffer->items)
1578 : 0 : pfree(buffer->items);
1579 : :
1580 : : /* release byref values, do nothing for by-val ones */
1581 [ # # ]: 0 : if (!GinBufferIsEmpty(buffer) &&
1582 [ # # # # ]: 0 : (buffer->category == GIN_CAT_NORM_KEY) && !buffer->typbyval)
1583 : 0 : pfree(DatumGetPointer(buffer->key));
1584 : :
1585 : 0 : pfree(buffer);
1586 : 0 : }
1587 : :
1588 : : /*
1589 : : * GinBufferCanAddKey
1590 : : * Check if a given GIN tuple can be added to the current buffer.
1591 : : *
1592 : : * Returns true if the buffer is either empty or for the same index key.
1593 : : */
1594 : : static bool
1595 : 0 : GinBufferCanAddKey(GinBuffer *buffer, GinTuple *tup)
1596 : : {
1597 : : /* empty buffer can accept data for any key */
1598 [ # # ]: 0 : if (GinBufferIsEmpty(buffer))
1599 : 0 : return true;
1600 : :
1601 : : /* otherwise just data for the same key */
1602 : 0 : return GinBufferKeyEquals(buffer, tup);
1603 : : }
1604 : :
1605 : : /*
1606 : : * Within leader, wait for end of heap scan and merge per-worker results.
1607 : : *
1608 : : * After waiting for all workers to finish, merge the per-worker results into
1609 : : * the complete index. The results from each worker are sorted by block number
1610 : : * (start of the page range). While combining the per-worker results we merge
1611 : : * summaries for the same page range, and also fill-in empty summaries for
1612 : : * ranges without any tuples.
1613 : : *
1614 : : * Returns the total number of heap tuples scanned.
1615 : : */
1616 : : static double
1617 : 0 : _gin_parallel_merge(GinBuildState *state)
1618 : : {
1619 : : GinTuple *tup;
1620 : : Size tuplen;
1621 : 0 : double reltuples = 0;
1622 : : GinBuffer *buffer;
1623 : :
1624 : : /* GIN tuples from workers, merged by leader */
1625 : 0 : double numtuples = 0;
1626 : :
1627 : : /* wait for workers to scan table and produce partial results */
1628 : 0 : reltuples = _gin_parallel_heapscan(state);
1629 : :
1630 : : /* Execute the sort */
1631 : 0 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
1632 : : PROGRESS_GIN_PHASE_PERFORMSORT_2);
1633 : :
1634 : : /* do the actual sort in the leader */
1635 : 0 : tuplesort_performsort(state->bs_sortstate);
1636 : :
1637 : : /*
1638 : : * Initialize buffer to combine entries for the same key.
1639 : : *
1640 : : * The leader is allowed to use the whole maintenance_work_mem buffer to
1641 : : * combine data. The parallel workers already completed.
1642 : : */
1643 : 0 : buffer = GinBufferInit(state->ginstate.index);
1644 : :
1645 : : /*
1646 : : * Set the progress target for the next phase. Reset the block number
1647 : : * values set by table_index_build_scan
1648 : : */
1649 : : {
1650 : 0 : const int progress_index[] = {
1651 : : PROGRESS_CREATEIDX_SUBPHASE,
1652 : : PROGRESS_CREATEIDX_TUPLES_TOTAL,
1653 : : PROGRESS_SCAN_BLOCKS_TOTAL,
1654 : : PROGRESS_SCAN_BLOCKS_DONE
1655 : : };
1656 : 0 : const int64 progress_vals[] = {
1657 : : PROGRESS_GIN_PHASE_MERGE_2,
1658 : 0 : state->bs_numtuples,
1659 : : 0, 0
1660 : : };
1661 : :
1662 : 0 : pgstat_progress_update_multi_param(4, progress_index, progress_vals);
1663 : : }
1664 : :
1665 : : /*
1666 : : * Read the GIN tuples from the shared tuplesort, sorted by category and
1667 : : * key. That probably gives us order matching how data is organized in the
1668 : : * index.
1669 : : *
1670 : : * We don't insert the GIN tuples right away, but instead accumulate as
1671 : : * many TIDs for the same key as possible, and then insert that at once.
1672 : : * This way we don't need to decompress/recompress the posting lists, etc.
1673 : : */
1674 [ # # ]: 0 : while ((tup = tuplesort_getgintuple(state->bs_sortstate, &tuplen, true)) != NULL)
1675 : : {
1676 : : MemoryContext oldCtx;
1677 : :
1678 [ # # ]: 0 : CHECK_FOR_INTERRUPTS();
1679 : :
1680 : : /*
1681 : : * If the buffer can accept the new GIN tuple, just store it there and
1682 : : * we're done. If it's a different key (or maybe too much data) flush
1683 : : * the current contents into the index first.
1684 : : */
1685 [ # # ]: 0 : if (!GinBufferCanAddKey(buffer, tup))
1686 : : {
1687 : : /*
1688 : : * Buffer is not empty and it's storing a different key - flush
1689 : : * the data into the insert, and start a new entry for current
1690 : : * GinTuple.
1691 : : */
1692 : 0 : AssertCheckItemPointers(buffer);
1693 : :
179 1694 : 0 : oldCtx = MemoryContextSwitchTo(state->tmpCtx);
1695 : :
239 1696 : 0 : ginEntryInsert(&state->ginstate,
1697 : 0 : buffer->attnum, buffer->key, buffer->category,
1698 : 0 : buffer->items, buffer->nitems, &state->buildStats);
1699 : :
179 1700 : 0 : MemoryContextSwitchTo(oldCtx);
1701 : 0 : MemoryContextReset(state->tmpCtx);
1702 : :
1703 : : /* discard the existing data */
239 1704 : 0 : GinBufferReset(buffer);
1705 : : }
1706 : :
1707 : : /*
1708 : : * We're about to add a GIN tuple to the buffer - check the memory
1709 : : * limit first, and maybe write out some of the data into the index
1710 : : * first, if needed (and possible). We only flush the part of the TID
1711 : : * list that we know won't change, and only if there's enough data for
1712 : : * compression to work well.
1713 : : */
238 1714 [ # # ]: 0 : if (GinBufferShouldTrim(buffer, tup))
1715 : : {
1716 [ # # ]: 0 : Assert(buffer->nfrozen > 0);
1717 : :
1718 : : /*
1719 : : * Buffer is not empty and it's storing a different key - flush
1720 : : * the data into the insert, and start a new entry for current
1721 : : * GinTuple.
1722 : : */
1723 : 0 : AssertCheckItemPointers(buffer);
1724 : :
179 1725 : 0 : oldCtx = MemoryContextSwitchTo(state->tmpCtx);
1726 : :
238 1727 : 0 : ginEntryInsert(&state->ginstate,
1728 : 0 : buffer->attnum, buffer->key, buffer->category,
1729 : 0 : buffer->items, buffer->nfrozen, &state->buildStats);
1730 : :
179 1731 : 0 : MemoryContextSwitchTo(oldCtx);
1732 : 0 : MemoryContextReset(state->tmpCtx);
1733 : :
1734 : : /* truncate the data we've just discarded */
238 1735 : 0 : GinBufferTrim(buffer);
1736 : : }
1737 : :
1738 : : /*
1739 : : * Remember data for the current tuple (either remember the new key,
1740 : : * or append if to the existing data).
1741 : : */
239 1742 : 0 : GinBufferStoreTuple(buffer, tup);
1743 : :
1744 : : /* Report progress */
1745 : 0 : pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE,
1746 : : ++numtuples);
1747 : : }
1748 : :
1749 : : /* flush data remaining in the buffer (for the last key) */
1750 [ # # ]: 0 : if (!GinBufferIsEmpty(buffer))
1751 : : {
1752 : 0 : AssertCheckItemPointers(buffer);
1753 : :
1754 : 0 : ginEntryInsert(&state->ginstate,
1755 : 0 : buffer->attnum, buffer->key, buffer->category,
1756 : 0 : buffer->items, buffer->nitems, &state->buildStats);
1757 : :
1758 : : /* discard the existing data */
1759 : 0 : GinBufferReset(buffer);
1760 : :
1761 : : /* Report progress */
1762 : 0 : pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE,
1763 : : ++numtuples);
1764 : : }
1765 : :
1766 : : /* relase all the memory */
1767 : 0 : GinBufferFree(buffer);
1768 : :
1769 : 0 : tuplesort_end(state->bs_sortstate);
1770 : :
1771 : 0 : return reltuples;
1772 : : }
1773 : :
1774 : : /*
1775 : : * Returns size of shared memory required to store state for a parallel
1776 : : * gin index build based on the snapshot its parallel scan will use.
1777 : : */
1778 : : static Size
1779 : 0 : _gin_parallel_estimate_shared(Relation heap, Snapshot snapshot)
1780 : : {
1781 : : /* c.f. shm_toc_allocate as to why BUFFERALIGN is used */
1782 : 0 : return add_size(BUFFERALIGN(sizeof(GinBuildShared)),
1783 : : table_parallelscan_estimate(heap, snapshot));
1784 : : }
1785 : :
1786 : : /*
1787 : : * Within leader, participate as a parallel worker.
1788 : : */
1789 : : static void
1790 : 0 : _gin_leader_participate_as_worker(GinBuildState *buildstate, Relation heap, Relation index)
1791 : : {
1792 : 0 : GinLeader *ginleader = buildstate->bs_leader;
1793 : : int sortmem;
1794 : :
1795 : : /*
1796 : : * Might as well use reliable figure when doling out maintenance_work_mem
1797 : : * (when requested number of workers were not launched, this will be
1798 : : * somewhat higher than it is for other workers).
1799 : : */
1800 : 0 : sortmem = maintenance_work_mem / ginleader->nparticipanttuplesorts;
1801 : :
1802 : : /* Perform work common to all participants */
1803 : 0 : _gin_parallel_scan_and_build(buildstate, ginleader->ginshared,
1804 : : ginleader->sharedsort, heap, index,
1805 : : sortmem, true);
1806 : 0 : }
1807 : :
1808 : : /*
1809 : : * _gin_process_worker_data
1810 : : * First phase of the key merging, happening in the worker.
1811 : : *
1812 : : * Depending on the number of distinct keys, the TID lists produced by the
1813 : : * callback may be very short (due to frequent evictions in the callback).
1814 : : * But combining many tiny lists is expensive, so we try to do as much as
1815 : : * possible in the workers and only then pass the results to the leader.
1816 : : *
1817 : : * We read the tuples sorted by the key, and merge them into larger lists.
1818 : : * At the moment there's no memory limit, so this will just produce one
1819 : : * huge (sorted) list per key in each worker. Which means the leader will
1820 : : * do a very limited number of mergesorts, which is good.
1821 : : */
1822 : : static void
1823 : 0 : _gin_process_worker_data(GinBuildState *state, Tuplesortstate *worker_sort,
1824 : : bool progress)
1825 : : {
1826 : : GinTuple *tup;
1827 : : Size tuplen;
1828 : :
1829 : : GinBuffer *buffer;
1830 : :
1831 : : /*
1832 : : * Initialize buffer to combine entries for the same key.
1833 : : *
1834 : : * The workers are limited to the same amount of memory as during the sort
1835 : : * in ginBuildCallbackParallel. But this probably should be the 32MB used
1836 : : * during planning, just like there.
1837 : : */
1838 : 0 : buffer = GinBufferInit(state->ginstate.index);
1839 : :
1840 : : /* sort the raw per-worker data */
1841 [ # # ]: 0 : if (progress)
1842 : 0 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
1843 : : PROGRESS_GIN_PHASE_PERFORMSORT_1);
1844 : :
1845 : 0 : tuplesort_performsort(state->bs_worker_sort);
1846 : :
1847 : : /* reset the number of GIN tuples produced by this worker */
1848 : 0 : state->bs_numtuples = 0;
1849 : :
1850 [ # # ]: 0 : if (progress)
1851 : 0 : pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE,
1852 : : PROGRESS_GIN_PHASE_MERGE_1);
1853 : :
1854 : : /*
1855 : : * Read the GIN tuples from the shared tuplesort, sorted by the key, and
1856 : : * merge them into larger chunks for the leader to combine.
1857 : : */
1858 [ # # ]: 0 : while ((tup = tuplesort_getgintuple(worker_sort, &tuplen, true)) != NULL)
1859 : : {
1860 : :
1861 [ # # ]: 0 : CHECK_FOR_INTERRUPTS();
1862 : :
1863 : : /*
1864 : : * If the buffer can accept the new GIN tuple, just store it there and
1865 : : * we're done. If it's a different key (or maybe too much data) flush
1866 : : * the current contents into the index first.
1867 : : */
1868 [ # # ]: 0 : if (!GinBufferCanAddKey(buffer, tup))
1869 : : {
1870 : : GinTuple *ntup;
1871 : : Size ntuplen;
1872 : :
1873 : : /*
1874 : : * Buffer is not empty and it's storing a different key - flush
1875 : : * the data into the insert, and start a new entry for current
1876 : : * GinTuple.
1877 : : */
1878 : 0 : AssertCheckItemPointers(buffer);
1879 : :
1880 : 0 : ntup = _gin_build_tuple(buffer->attnum, buffer->category,
1881 : 0 : buffer->key, buffer->typlen, buffer->typbyval,
1882 : 0 : buffer->items, buffer->nitems, &ntuplen);
1883 : :
1884 : 0 : tuplesort_putgintuple(state->bs_sortstate, ntup, ntuplen);
1885 : 0 : state->bs_numtuples++;
1886 : :
1887 : 0 : pfree(ntup);
1888 : :
1889 : : /* discard the existing data */
1890 : 0 : GinBufferReset(buffer);
1891 : : }
1892 : :
1893 : : /*
1894 : : * We're about to add a GIN tuple to the buffer - check the memory
1895 : : * limit first, and maybe write out some of the data into the index
1896 : : * first, if needed (and possible). We only flush the part of the TID
1897 : : * list that we know won't change, and only if there's enough data for
1898 : : * compression to work well.
1899 : : */
238 1900 [ # # ]: 0 : if (GinBufferShouldTrim(buffer, tup))
1901 : : {
1902 : : GinTuple *ntup;
1903 : : Size ntuplen;
1904 : :
1905 [ # # ]: 0 : Assert(buffer->nfrozen > 0);
1906 : :
1907 : : /*
1908 : : * Buffer is not empty and it's storing a different key - flush
1909 : : * the data into the insert, and start a new entry for current
1910 : : * GinTuple.
1911 : : */
1912 : 0 : AssertCheckItemPointers(buffer);
1913 : :
1914 : 0 : ntup = _gin_build_tuple(buffer->attnum, buffer->category,
1915 : 0 : buffer->key, buffer->typlen, buffer->typbyval,
1916 : 0 : buffer->items, buffer->nfrozen, &ntuplen);
1917 : :
1918 : 0 : tuplesort_putgintuple(state->bs_sortstate, ntup, ntuplen);
1919 : :
1920 : 0 : pfree(ntup);
1921 : :
1922 : : /* truncate the data we've just discarded */
1923 : 0 : GinBufferTrim(buffer);
1924 : : }
1925 : :
1926 : : /*
1927 : : * Remember data for the current tuple (either remember the new key,
1928 : : * or append if to the existing data).
1929 : : */
239 1930 : 0 : GinBufferStoreTuple(buffer, tup);
1931 : : }
1932 : :
1933 : : /* flush data remaining in the buffer (for the last key) */
1934 [ # # ]: 0 : if (!GinBufferIsEmpty(buffer))
1935 : : {
1936 : : GinTuple *ntup;
1937 : : Size ntuplen;
1938 : :
1939 : 0 : AssertCheckItemPointers(buffer);
1940 : :
1941 : 0 : ntup = _gin_build_tuple(buffer->attnum, buffer->category,
1942 : 0 : buffer->key, buffer->typlen, buffer->typbyval,
1943 : 0 : buffer->items, buffer->nitems, &ntuplen);
1944 : :
1945 : 0 : tuplesort_putgintuple(state->bs_sortstate, ntup, ntuplen);
1946 : 0 : state->bs_numtuples++;
1947 : :
1948 : 0 : pfree(ntup);
1949 : :
1950 : : /* discard the existing data */
1951 : 0 : GinBufferReset(buffer);
1952 : : }
1953 : :
1954 : : /* relase all the memory */
1955 : 0 : GinBufferFree(buffer);
1956 : :
1957 : 0 : tuplesort_end(worker_sort);
1958 : 0 : }
1959 : :
1960 : : /*
1961 : : * Perform a worker's portion of a parallel GIN index build sort.
1962 : : *
1963 : : * This generates a tuplesort for the worker portion of the table.
1964 : : *
1965 : : * sortmem is the amount of working memory to use within each worker,
1966 : : * expressed in KBs.
1967 : : *
1968 : : * When this returns, workers are done, and need only release resources.
1969 : : *
1970 : : * Before feeding data into a shared tuplesort (for the leader process),
1971 : : * the workers process data in two phases.
1972 : : *
1973 : : * 1) A worker reads a portion of rows from the table, accumulates entries
1974 : : * in memory, and flushes them into a private tuplesort (e.g. because of
1975 : : * using too much memory).
1976 : : *
1977 : : * 2) The private tuplesort gets sorted (by key and TID), the worker reads
1978 : : * the data again, and combines the entries as much as possible. This has
1979 : : * to happen eventually, and this way it's done in workers in parallel.
1980 : : *
1981 : : * Finally, the combined entries are written into the shared tuplesort, so
1982 : : * that the leader can process them.
1983 : : *
1984 : : * How well this works (compared to just writing entries into the shared
1985 : : * tuplesort) depends on the data set. For large tables with many distinct
1986 : : * keys this helps a lot. With many distinct keys it's likely the buffers has
1987 : : * to be flushed often, generating many entries with the same key and short
1988 : : * TID lists. These entries need to be sorted and merged at some point,
1989 : : * before writing them to the index. The merging is quite expensive, it can
1990 : : * easily be ~50% of a serial build, and doing as much of it in the workers
1991 : : * means it's parallelized. The leader still has to merge results from the
1992 : : * workers, but it's much more efficient to merge few large entries than
1993 : : * many tiny ones.
1994 : : *
1995 : : * This also reduces the amount of data the workers pass to the leader through
1996 : : * the shared tuplesort. OTOH the workers need more space for the private sort,
1997 : : * possibly up to 2x of the data, if no entries be merged in a worker. But this
1998 : : * is very unlikely, and the only consequence is inefficiency, so we ignore it.
1999 : : */
2000 : : static void
2001 : 0 : _gin_parallel_scan_and_build(GinBuildState *state,
2002 : : GinBuildShared *ginshared, Sharedsort *sharedsort,
2003 : : Relation heap, Relation index,
2004 : : int sortmem, bool progress)
2005 : : {
2006 : : SortCoordinate coordinate;
2007 : : TableScanDesc scan;
2008 : : double reltuples;
2009 : : IndexInfo *indexInfo;
2010 : :
2011 : : /* Initialize local tuplesort coordination state */
2012 : 0 : coordinate = palloc0(sizeof(SortCoordinateData));
2013 : 0 : coordinate->isWorker = true;
2014 : 0 : coordinate->nParticipants = -1;
2015 : 0 : coordinate->sharedsort = sharedsort;
2016 : :
2017 : : /* remember how much space is allowed for the accumulated entries */
2018 : 0 : state->work_mem = (sortmem / 2);
2019 : :
2020 : : /* Begin "partial" tuplesort */
2021 : 0 : state->bs_sortstate = tuplesort_begin_index_gin(heap, index,
2022 : : state->work_mem,
2023 : : coordinate,
2024 : : TUPLESORT_NONE);
2025 : :
2026 : : /* Local per-worker sort of raw-data */
2027 : 0 : state->bs_worker_sort = tuplesort_begin_index_gin(heap, index,
2028 : : state->work_mem,
2029 : : NULL,
2030 : : TUPLESORT_NONE);
2031 : :
2032 : : /* Join parallel scan */
2033 : 0 : indexInfo = BuildIndexInfo(index);
2034 : 0 : indexInfo->ii_Concurrent = ginshared->isconcurrent;
2035 : :
2036 : 0 : scan = table_beginscan_parallel(heap,
2037 : : ParallelTableScanFromGinBuildShared(ginshared));
2038 : :
2039 : 0 : reltuples = table_index_build_scan(heap, index, indexInfo, true, progress,
2040 : : ginBuildCallbackParallel, state, scan);
2041 : :
2042 : : /* write remaining accumulated entries */
2043 : 0 : ginFlushBuildState(state, index);
2044 : :
2045 : : /*
2046 : : * Do the first phase of in-worker processing - sort the data produced by
2047 : : * the callback, and combine them into much larger chunks and place that
2048 : : * into the shared tuplestore for leader to process.
2049 : : */
2050 : 0 : _gin_process_worker_data(state, state->bs_worker_sort, progress);
2051 : :
2052 : : /* sort the GIN tuples built by this worker */
2053 : 0 : tuplesort_performsort(state->bs_sortstate);
2054 : :
2055 : 0 : state->bs_reltuples += reltuples;
2056 : :
2057 : : /*
2058 : : * Done. Record ambuild statistics.
2059 : : */
2060 [ # # ]: 0 : SpinLockAcquire(&ginshared->mutex);
2061 : 0 : ginshared->nparticipantsdone++;
2062 : 0 : ginshared->reltuples += state->bs_reltuples;
2063 : 0 : ginshared->indtuples += state->bs_numtuples;
2064 : 0 : SpinLockRelease(&ginshared->mutex);
2065 : :
2066 : : /* Notify leader */
2067 : 0 : ConditionVariableSignal(&ginshared->workersdonecv);
2068 : :
2069 : 0 : tuplesort_end(state->bs_sortstate);
2070 : 0 : }
2071 : :
2072 : : /*
2073 : : * Perform work within a launched parallel process.
2074 : : */
2075 : : void
2076 : 0 : _gin_parallel_build_main(dsm_segment *seg, shm_toc *toc)
2077 : : {
2078 : : char *sharedquery;
2079 : : GinBuildShared *ginshared;
2080 : : Sharedsort *sharedsort;
2081 : : GinBuildState buildstate;
2082 : : Relation heapRel;
2083 : : Relation indexRel;
2084 : : LOCKMODE heapLockmode;
2085 : : LOCKMODE indexLockmode;
2086 : : WalUsage *walusage;
2087 : : BufferUsage *bufferusage;
2088 : : int sortmem;
2089 : :
2090 : : /*
2091 : : * The only possible status flag that can be set to the parallel worker is
2092 : : * PROC_IN_SAFE_IC.
2093 : : */
2094 [ # # # # ]: 0 : Assert((MyProc->statusFlags == 0) ||
2095 : : (MyProc->statusFlags == PROC_IN_SAFE_IC));
2096 : :
2097 : : /* Set debug_query_string for individual workers first */
2098 : 0 : sharedquery = shm_toc_lookup(toc, PARALLEL_KEY_QUERY_TEXT, true);
2099 : 0 : debug_query_string = sharedquery;
2100 : :
2101 : : /* Report the query string from leader */
2102 : 0 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
2103 : :
2104 : : /* Look up gin shared state */
2105 : 0 : ginshared = shm_toc_lookup(toc, PARALLEL_KEY_GIN_SHARED, false);
2106 : :
2107 : : /* Open relations using lock modes known to be obtained by index.c */
2108 [ # # ]: 0 : if (!ginshared->isconcurrent)
2109 : : {
2110 : 0 : heapLockmode = ShareLock;
2111 : 0 : indexLockmode = AccessExclusiveLock;
2112 : : }
2113 : : else
2114 : : {
2115 : 0 : heapLockmode = ShareUpdateExclusiveLock;
2116 : 0 : indexLockmode = RowExclusiveLock;
2117 : : }
2118 : :
2119 : : /* Open relations within worker */
2120 : 0 : heapRel = table_open(ginshared->heaprelid, heapLockmode);
2121 : 0 : indexRel = index_open(ginshared->indexrelid, indexLockmode);
2122 : :
2123 : : /* initialize the GIN build state */
2124 : 0 : initGinState(&buildstate.ginstate, indexRel);
2125 : 0 : buildstate.indtuples = 0;
2126 : 0 : memset(&buildstate.buildStats, 0, sizeof(GinStatsData));
2127 : 0 : memset(&buildstate.tid, 0, sizeof(ItemPointerData));
2128 : :
2129 : : /*
2130 : : * create a temporary memory context that is used to hold data not yet
2131 : : * dumped out to the index
2132 : : */
2133 : 0 : buildstate.tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
2134 : : "Gin build temporary context",
2135 : : ALLOCSET_DEFAULT_SIZES);
2136 : :
2137 : : /*
2138 : : * create a temporary memory context that is used for calling
2139 : : * ginExtractEntries(), and can be reset after each tuple
2140 : : */
2141 : 0 : buildstate.funcCtx = AllocSetContextCreate(CurrentMemoryContext,
2142 : : "Gin build temporary context for user-defined function",
2143 : : ALLOCSET_DEFAULT_SIZES);
2144 : :
2145 : 0 : buildstate.accum.ginstate = &buildstate.ginstate;
2146 : 0 : ginInitBA(&buildstate.accum);
2147 : :
2148 : :
2149 : : /* Look up shared state private to tuplesort.c */
2150 : 0 : sharedsort = shm_toc_lookup(toc, PARALLEL_KEY_TUPLESORT, false);
2151 : 0 : tuplesort_attach_shared(sharedsort, seg);
2152 : :
2153 : : /* Prepare to track buffer usage during parallel execution */
2154 : 0 : InstrStartParallelQuery();
2155 : :
2156 : : /*
2157 : : * Might as well use reliable figure when doling out maintenance_work_mem
2158 : : * (when requested number of workers were not launched, this will be
2159 : : * somewhat higher than it is for other workers).
2160 : : */
2161 : 0 : sortmem = maintenance_work_mem / ginshared->scantuplesortstates;
2162 : :
2163 : 0 : _gin_parallel_scan_and_build(&buildstate, ginshared, sharedsort,
2164 : : heapRel, indexRel, sortmem, false);
2165 : :
2166 : : /* Report WAL/buffer usage during parallel execution */
2167 : 0 : bufferusage = shm_toc_lookup(toc, PARALLEL_KEY_BUFFER_USAGE, false);
2168 : 0 : walusage = shm_toc_lookup(toc, PARALLEL_KEY_WAL_USAGE, false);
2169 : 0 : InstrEndParallelQuery(&bufferusage[ParallelWorkerNumber],
2170 : 0 : &walusage[ParallelWorkerNumber]);
2171 : :
2172 : 0 : index_close(indexRel, indexLockmode);
2173 : 0 : table_close(heapRel, heapLockmode);
2174 : 0 : }
2175 : :
2176 : : /*
2177 : : * Used to keep track of compressed TID lists when building a GIN tuple.
2178 : : */
2179 : : typedef struct
2180 : : {
2181 : : dlist_node node; /* linked list pointers */
2182 : : GinPostingList *seg;
2183 : : } GinSegmentInfo;
2184 : :
2185 : : /*
2186 : : * _gin_build_tuple
2187 : : * Serialize the state for an index key into a tuple for tuplesort.
2188 : : *
2189 : : * The tuple has a number of scalar fields (mostly matching the build state),
2190 : : * and then a data array that stores the key first, and then the TID list.
2191 : : *
2192 : : * For by-reference data types, we store the actual data. For by-val types
2193 : : * we simply copy the whole Datum, so that we don't have to care about stuff
2194 : : * like endianess etc. We could make it a little bit smaller, but it's not
2195 : : * worth it - it's a tiny fraction of the data, and we need to MAXALIGN the
2196 : : * start of the TID list anyway. So we wouldn't save anything. (This would
2197 : : * not be a good idea for the permanent in-index data, since we'd prefer
2198 : : * that that not depend on sizeof(Datum). But this is just a transient
2199 : : * representation to use while sorting the data.)
2200 : : *
2201 : : * The TID list is serialized as compressed - it's highly compressible, and
2202 : : * we already have ginCompressPostingList for this purpose. The list may be
2203 : : * pretty long, so we compress it into multiple segments and then copy all
2204 : : * of that into the GIN tuple.
2205 : : */
2206 : : static GinTuple *
2207 : 0 : _gin_build_tuple(OffsetNumber attrnum, unsigned char category,
2208 : : Datum key, int16 typlen, bool typbyval,
2209 : : ItemPointerData *items, uint32 nitems,
2210 : : Size *len)
2211 : : {
2212 : : GinTuple *tuple;
2213 : : char *ptr;
2214 : :
2215 : : Size tuplen;
2216 : : int keylen;
2217 : :
2218 : : dlist_mutable_iter iter;
2219 : : dlist_head segments;
2220 : : int ncompressed;
2221 : : Size compresslen;
2222 : :
2223 : : /*
2224 : : * Calculate how long is the key value. Only keys with GIN_CAT_NORM_KEY
2225 : : * have actual non-empty key. We include varlena headers and \0 bytes for
2226 : : * strings, to make it easier to access the data in-line.
2227 : : *
2228 : : * For byval types we simply copy the whole Datum. We could store just the
2229 : : * necessary bytes, but this is simpler to work with and not worth the
2230 : : * extra complexity. Moreover we still need to do the MAXALIGN to allow
2231 : : * direct access to items pointers.
2232 : : *
2233 : : * XXX Note that for byval types we store the whole datum, no matter what
2234 : : * the typlen value is.
2235 : : */
2236 [ # # ]: 0 : if (category != GIN_CAT_NORM_KEY)
2237 : 0 : keylen = 0;
2238 [ # # ]: 0 : else if (typbyval)
2239 : 0 : keylen = sizeof(Datum);
2240 [ # # ]: 0 : else if (typlen > 0)
2241 : 0 : keylen = typlen;
2242 [ # # ]: 0 : else if (typlen == -1)
84 peter@eisentraut.org 2243 :UNC 0 : keylen = VARSIZE_ANY(DatumGetPointer(key));
239 tomas.vondra@postgre 2244 [ # # ]:UBC 0 : else if (typlen == -2)
2245 : 0 : keylen = strlen(DatumGetPointer(key)) + 1;
2246 : : else
2247 [ # # ]: 0 : elog(ERROR, "unexpected typlen value (%d)", typlen);
2248 : :
2249 : : /* compress the item pointers */
238 2250 : 0 : ncompressed = 0;
2251 : 0 : compresslen = 0;
2252 : 0 : dlist_init(&segments);
2253 : :
2254 : : /* generate compressed segments of TID list chunks */
2255 [ # # ]: 0 : while (ncompressed < nitems)
2256 : : {
2257 : : int cnt;
2258 : 0 : GinSegmentInfo *seginfo = palloc(sizeof(GinSegmentInfo));
2259 : :
2260 : 0 : seginfo->seg = ginCompressPostingList(&items[ncompressed],
2261 : 0 : (nitems - ncompressed),
2262 : : UINT16_MAX,
2263 : : &cnt);
2264 : :
2265 : 0 : ncompressed += cnt;
2266 : 0 : compresslen += SizeOfGinPostingList(seginfo->seg);
2267 : :
2268 : 0 : dlist_push_tail(&segments, &seginfo->node);
2269 : : }
2270 : :
2271 : : /*
2272 : : * Determine GIN tuple length with all the data included. Be careful about
2273 : : * alignment, to allow direct access to compressed segments (those require
2274 : : * only SHORTALIGN).
2275 : : */
2276 : 0 : tuplen = SHORTALIGN(offsetof(GinTuple, data) + keylen) + compresslen;
2277 : :
239 2278 : 0 : *len = tuplen;
2279 : :
2280 : : /*
2281 : : * Allocate space for the whole GIN tuple.
2282 : : *
2283 : : * The palloc0 is needed - writetup_index_gin will write the whole tuple
2284 : : * to disk, so we need to make sure the padding bytes are defined
2285 : : * (otherwise valgrind would report this).
2286 : : */
2287 : 0 : tuple = palloc0(tuplen);
2288 : :
2289 : 0 : tuple->tuplen = tuplen;
2290 : 0 : tuple->attrnum = attrnum;
2291 : 0 : tuple->category = category;
2292 : 0 : tuple->keylen = keylen;
2293 : 0 : tuple->nitems = nitems;
2294 : :
2295 : : /* key type info */
2296 : 0 : tuple->typlen = typlen;
2297 : 0 : tuple->typbyval = typbyval;
2298 : :
2299 : : /*
2300 : : * Copy the key and items into the tuple. First the key value, which we
2301 : : * can simply copy right at the beginning of the data array.
2302 : : */
2303 [ # # ]: 0 : if (category == GIN_CAT_NORM_KEY)
2304 : : {
2305 [ # # ]: 0 : if (typbyval)
2306 : : {
2307 : 0 : memcpy(tuple->data, &key, sizeof(Datum));
2308 : : }
2309 [ # # ]: 0 : else if (typlen > 0) /* byref, fixed length */
2310 : : {
2311 : 0 : memcpy(tuple->data, DatumGetPointer(key), typlen);
2312 : : }
2313 [ # # ]: 0 : else if (typlen == -1)
2314 : : {
2315 : 0 : memcpy(tuple->data, DatumGetPointer(key), keylen);
2316 : : }
2317 [ # # ]: 0 : else if (typlen == -2)
2318 : : {
2319 : 0 : memcpy(tuple->data, DatumGetPointer(key), keylen);
2320 : : }
2321 : : }
2322 : :
2323 : : /* finally, copy the TIDs into the array */
2324 : 0 : ptr = (char *) tuple + SHORTALIGN(offsetof(GinTuple, data) + keylen);
2325 : :
2326 : : /* copy in the compressed data, and free the segments */
238 2327 [ # # # # ]: 0 : dlist_foreach_modify(iter, &segments)
2328 : : {
2329 : 0 : GinSegmentInfo *seginfo = dlist_container(GinSegmentInfo, node, iter.cur);
2330 : :
2331 : 0 : memcpy(ptr, seginfo->seg, SizeOfGinPostingList(seginfo->seg));
2332 : :
2333 : 0 : ptr += SizeOfGinPostingList(seginfo->seg);
2334 : :
2335 : 0 : dlist_delete(&seginfo->node);
2336 : :
2337 : 0 : pfree(seginfo->seg);
2338 : 0 : pfree(seginfo);
2339 : : }
2340 : :
239 2341 : 0 : return tuple;
2342 : : }
2343 : :
2344 : : /*
2345 : : * _gin_parse_tuple_key
2346 : : * Return a Datum representing the key stored in the tuple.
2347 : : *
2348 : : * Most of the tuple fields are directly accessible, the only thing that
2349 : : * needs more care is the key and the TID list.
2350 : : *
2351 : : * For the key, this returns a regular Datum representing it. It's either the
2352 : : * actual key value, or a pointer to the beginning of the data array (which is
2353 : : * where the data was copied by _gin_build_tuple).
2354 : : */
2355 : : static Datum
238 2356 : 0 : _gin_parse_tuple_key(GinTuple *a)
2357 : : {
2358 : : Datum key;
2359 : :
239 2360 [ # # ]: 0 : if (a->category != GIN_CAT_NORM_KEY)
2361 : 0 : return (Datum) 0;
2362 : :
2363 [ # # ]: 0 : if (a->typbyval)
2364 : : {
2365 : 0 : memcpy(&key, a->data, a->keylen);
2366 : 0 : return key;
2367 : : }
2368 : :
2369 : 0 : return PointerGetDatum(a->data);
2370 : : }
2371 : :
2372 : : /*
2373 : : * _gin_parse_tuple_items
2374 : : * Return a pointer to a palloc'd array of decompressed TID array.
2375 : : */
2376 : : static ItemPointer
238 2377 : 0 : _gin_parse_tuple_items(GinTuple *a)
2378 : : {
2379 : : int len;
2380 : : char *ptr;
2381 : : int ndecoded;
2382 : : ItemPointer items;
2383 : :
2384 : 0 : len = a->tuplen - SHORTALIGN(offsetof(GinTuple, data) + a->keylen);
2385 : 0 : ptr = (char *) a + SHORTALIGN(offsetof(GinTuple, data) + a->keylen);
2386 : :
2387 : 0 : items = ginPostingListDecodeAllSegments((GinPostingList *) ptr, len, &ndecoded);
2388 : :
2389 [ # # ]: 0 : Assert(ndecoded == a->nitems);
2390 : :
2391 : 0 : return (ItemPointer) items;
2392 : : }
2393 : :
2394 : : /*
2395 : : * _gin_compare_tuples
2396 : : * Compare GIN tuples, used by tuplesort during parallel index build.
2397 : : *
2398 : : * The scalar fields (attrnum, category) are compared first, the key value is
2399 : : * compared last. The comparisons are done using type-specific sort support
2400 : : * functions.
2401 : : *
2402 : : * If the key value matches, we compare the first TID value in the TID list,
2403 : : * which means the tuples are merged in an order in which they are most
2404 : : * likely to be simply concatenated. (This "first" TID will also allow us
2405 : : * to determine a point up to which the list is fully determined and can be
2406 : : * written into the index to enforce a memory limit etc.)
2407 : : */
2408 : : int
239 2409 : 0 : _gin_compare_tuples(GinTuple *a, GinTuple *b, SortSupport ssup)
2410 : : {
2411 : : int r;
2412 : : Datum keya,
2413 : : keyb;
2414 : :
2415 [ # # ]: 0 : if (a->attrnum < b->attrnum)
2416 : 0 : return -1;
2417 : :
2418 [ # # ]: 0 : if (a->attrnum > b->attrnum)
2419 : 0 : return 1;
2420 : :
2421 [ # # ]: 0 : if (a->category < b->category)
2422 : 0 : return -1;
2423 : :
2424 [ # # ]: 0 : if (a->category > b->category)
2425 : 0 : return 1;
2426 : :
2427 [ # # ]: 0 : if (a->category == GIN_CAT_NORM_KEY)
2428 : : {
238 2429 : 0 : keya = _gin_parse_tuple_key(a);
2430 : 0 : keyb = _gin_parse_tuple_key(b);
2431 : :
239 2432 : 0 : r = ApplySortComparator(keya, false,
2433 : : keyb, false,
2434 : 0 : &ssup[a->attrnum - 1]);
2435 : :
2436 : : /* if the key is the same, consider the first TID in the array */
2437 [ # # ]: 0 : return (r != 0) ? r : ItemPointerCompare(GinTupleGetFirst(a),
2438 : : GinTupleGetFirst(b));
2439 : : }
2440 : :
2441 : 0 : return ItemPointerCompare(GinTupleGetFirst(a),
2442 : : GinTupleGetFirst(b));
2443 : : }
|
