Branch data Line data Source code
1 : : /*----------------------------------------------------------------------
2 : : *
3 : : * tableam.c
4 : : * Table access method routines too big to be inline functions.
5 : : *
6 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/access/table/tableam.c
12 : : *
13 : : * NOTES
14 : : * Note that most functions in here are documented in tableam.h, rather than
15 : : * here. That's because there's a lot of inline functions in tableam.h and
16 : : * it'd be harder to understand if one constantly had to switch between files.
17 : : *
18 : : *----------------------------------------------------------------------
19 : : */
20 : : #include "postgres.h"
21 : :
22 : : #include <math.h>
23 : :
24 : : #include "access/syncscan.h"
25 : : #include "access/tableam.h"
26 : : #include "access/xact.h"
27 : : #include "optimizer/optimizer.h"
28 : : #include "optimizer/plancat.h"
29 : : #include "port/pg_bitutils.h"
30 : : #include "storage/bufmgr.h"
31 : : #include "storage/shmem.h"
32 : : #include "storage/smgr.h"
33 : :
34 : : /*
35 : : * Constants to control the behavior of block allocation to parallel workers
36 : : * during a parallel seqscan. Technically these values do not need to be
37 : : * powers of 2, but having them as powers of 2 makes the math more optimal
38 : : * and makes the ramp-down stepping more even.
39 : : */
40 : :
41 : : /* The number of I/O chunks we try to break a parallel seqscan down into */
42 : : #define PARALLEL_SEQSCAN_NCHUNKS 2048
43 : : /* Ramp down size of allocations when we've only this number of chunks left */
44 : : #define PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS 64
45 : : /* Cap the size of parallel I/O chunks to this number of blocks */
46 : : #define PARALLEL_SEQSCAN_MAX_CHUNK_SIZE 8192
47 : :
48 : : /* GUC variables */
49 : : char *default_table_access_method = DEFAULT_TABLE_ACCESS_METHOD;
50 : : bool synchronize_seqscans = true;
51 : :
52 : :
53 : : /* ----------------------------------------------------------------------------
54 : : * Slot functions.
55 : : * ----------------------------------------------------------------------------
56 : : */
57 : :
58 : : const TupleTableSlotOps *
59 : 18416675 : table_slot_callbacks(Relation relation)
60 : : {
61 : : const TupleTableSlotOps *tts_cb;
62 : :
63 [ + + ]: 18416675 : if (relation->rd_tableam)
64 : 18411610 : tts_cb = relation->rd_tableam->slot_callbacks(relation);
65 [ + + ]: 5065 : else if (relation->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
66 : : {
67 : : /*
68 : : * Historically FDWs expect to store heap tuples in slots. Continue
69 : : * handing them one, to make it less painful to adapt FDWs to new
70 : : * versions. The cost of a heap slot over a virtual slot is pretty
71 : : * small.
72 : : */
73 : 224 : tts_cb = &TTSOpsHeapTuple;
74 : : }
75 : : else
76 : : {
77 : : /*
78 : : * These need to be supported, as some parts of the code (like COPY)
79 : : * need to create slots for such relations too. It seems better to
80 : : * centralize the knowledge that a heap slot is the right thing in
81 : : * that case here.
82 : : */
83 : : Assert(relation->rd_rel->relkind == RELKIND_VIEW ||
84 : : relation->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
85 : 4841 : tts_cb = &TTSOpsVirtual;
86 : : }
87 : :
88 : 18416675 : return tts_cb;
89 : : }
90 : :
91 : : TupleTableSlot *
92 : 18108476 : table_slot_create(Relation relation, List **reglist)
93 : : {
94 : : const TupleTableSlotOps *tts_cb;
95 : : TupleTableSlot *slot;
96 : :
97 : 18108476 : tts_cb = table_slot_callbacks(relation);
98 : 18108476 : slot = MakeSingleTupleTableSlot(RelationGetDescr(relation), tts_cb);
99 : :
100 [ + + ]: 18108476 : if (reglist)
101 : 154531 : *reglist = lappend(*reglist, slot);
102 : :
103 : 18108476 : return slot;
104 : : }
105 : :
106 : :
107 : : /* ----------------------------------------------------------------------------
108 : : * Table scan functions.
109 : : * ----------------------------------------------------------------------------
110 : : */
111 : :
112 : : TableScanDesc
113 : 47718 : table_beginscan_catalog(Relation relation, int nkeys, ScanKeyData *key)
114 : : {
115 : 47718 : uint32 flags = SO_TYPE_SEQSCAN |
116 : : SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE | SO_TEMP_SNAPSHOT;
117 : 47718 : Oid relid = RelationGetRelid(relation);
118 : 47718 : Snapshot snapshot = RegisterSnapshot(GetCatalogSnapshot(relid));
119 : :
120 : 47718 : return table_beginscan_common(relation, snapshot, nkeys, key,
121 : : NULL, flags, SO_NONE);
122 : : }
123 : :
124 : :
125 : : /* ----------------------------------------------------------------------------
126 : : * Parallel table scan related functions.
127 : : * ----------------------------------------------------------------------------
128 : : */
129 : :
130 : : Size
131 : 1337 : table_parallelscan_estimate(Relation rel, Snapshot snapshot)
132 : : {
133 : 1337 : Size sz = 0;
134 : :
135 [ + + ]: 1337 : if (IsMVCCSnapshot(snapshot))
136 : 1210 : sz = add_size(sz, EstimateSnapshotSpace(snapshot));
137 : : else
138 : : Assert(snapshot == SnapshotAny);
139 : :
140 : 1337 : sz = add_size(sz, rel->rd_tableam->parallelscan_estimate(rel));
141 : :
142 : 1337 : return sz;
143 : : }
144 : :
145 : : void
146 : 1337 : table_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan,
147 : : Snapshot snapshot)
148 : : {
149 : 1337 : Size snapshot_off = rel->rd_tableam->parallelscan_initialize(rel, pscan);
150 : :
151 : 1337 : pscan->phs_snapshot_off = snapshot_off;
152 : :
153 [ + + ]: 1337 : if (IsMVCCSnapshot(snapshot))
154 : : {
155 : 1210 : SerializeSnapshot(snapshot, (char *) pscan + pscan->phs_snapshot_off);
156 : 1210 : pscan->phs_snapshot_any = false;
157 : : }
158 : : else
159 : : {
160 : : Assert(snapshot == SnapshotAny);
161 : 127 : pscan->phs_snapshot_any = true;
162 : : }
163 : 1337 : }
164 : :
165 : : TableScanDesc
166 : 4438 : table_beginscan_parallel(Relation relation, ParallelTableScanDesc pscan,
167 : : uint32 flags)
168 : : {
169 : : Snapshot snapshot;
170 : 4438 : uint32 internal_flags = SO_TYPE_SEQSCAN |
171 : : SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE;
172 : :
173 : : Assert(RelFileLocatorEquals(relation->rd_locator, pscan->phs_locator));
174 : :
175 [ + + ]: 4438 : if (!pscan->phs_snapshot_any)
176 : : {
177 : : /* Snapshot was serialized -- restore it */
178 : 4167 : snapshot = RestoreSnapshot((char *) pscan + pscan->phs_snapshot_off);
179 : 4167 : RegisterSnapshot(snapshot);
180 : 4167 : internal_flags |= SO_TEMP_SNAPSHOT;
181 : : }
182 : : else
183 : : {
184 : : /* SnapshotAny passed by caller (not serialized) */
185 : 271 : snapshot = SnapshotAny;
186 : : }
187 : :
188 : 4438 : return table_beginscan_common(relation, snapshot, 0, NULL,
189 : : pscan, internal_flags, flags);
190 : : }
191 : :
192 : : TableScanDesc
193 : 80 : table_beginscan_parallel_tidrange(Relation relation,
194 : : ParallelTableScanDesc pscan,
195 : : uint32 flags)
196 : : {
197 : : Snapshot snapshot;
198 : : TableScanDesc sscan;
199 : 80 : uint32 internal_flags = SO_TYPE_TIDRANGESCAN | SO_ALLOW_PAGEMODE;
200 : :
201 : : Assert(RelFileLocatorEquals(relation->rd_locator, pscan->phs_locator));
202 : :
203 : : /* disable syncscan in parallel tid range scan. */
204 : 80 : pscan->phs_syncscan = false;
205 : :
206 [ + - ]: 80 : if (!pscan->phs_snapshot_any)
207 : : {
208 : : /* Snapshot was serialized -- restore it */
209 : 80 : snapshot = RestoreSnapshot((char *) pscan + pscan->phs_snapshot_off);
210 : 80 : RegisterSnapshot(snapshot);
211 : 80 : internal_flags |= SO_TEMP_SNAPSHOT;
212 : : }
213 : : else
214 : : {
215 : : /* SnapshotAny passed by caller (not serialized) */
216 : 0 : snapshot = SnapshotAny;
217 : : }
218 : :
219 : 80 : sscan = table_beginscan_common(relation, snapshot, 0, NULL,
220 : : pscan, internal_flags, flags);
221 : 80 : return sscan;
222 : : }
223 : :
224 : :
225 : : /* ----------------------------------------------------------------------------
226 : : * Index scan related functions.
227 : : * ----------------------------------------------------------------------------
228 : : */
229 : :
230 : : /*
231 : : * To perform that check simply start an index scan, create the necessary
232 : : * slot, do the heap lookup, and shut everything down again. This could be
233 : : * optimized, but is unlikely to matter from a performance POV. If there
234 : : * frequently are live index pointers also matching a unique index key, the
235 : : * CPU overhead of this routine is unlikely to matter.
236 : : *
237 : : * Note that *tid may be modified when we return true if the AM supports
238 : : * storing multiple row versions reachable via a single index entry (like
239 : : * heap's HOT).
240 : : */
241 : : bool
242 : 7601687 : table_index_fetch_tuple_check(Relation rel,
243 : : ItemPointer tid,
244 : : Snapshot snapshot,
245 : : bool *all_dead)
246 : : {
247 : : IndexFetchTableData *scan;
248 : : TupleTableSlot *slot;
249 : 7601687 : bool call_again = false;
250 : : bool found;
251 : :
252 : 7601687 : slot = table_slot_create(rel, NULL);
253 : 7601687 : scan = table_index_fetch_begin(rel, SO_NONE);
254 : 7601687 : found = table_index_fetch_tuple(scan, tid, snapshot, slot, &call_again,
255 : : all_dead);
256 : 7601687 : table_index_fetch_end(scan);
257 : 7601687 : ExecDropSingleTupleTableSlot(slot);
258 : :
259 : 7601687 : return found;
260 : : }
261 : :
262 : :
263 : : /* ------------------------------------------------------------------------
264 : : * Functions for non-modifying operations on individual tuples
265 : : * ------------------------------------------------------------------------
266 : : */
267 : :
268 : : void
269 : 215 : table_tuple_get_latest_tid(TableScanDesc scan, ItemPointer tid)
270 : : {
271 : 215 : Relation rel = scan->rs_rd;
272 : 215 : const TableAmRoutine *tableam = rel->rd_tableam;
273 : :
274 : : /*
275 : : * Since this can be called with user-supplied TID, don't trust the input
276 : : * too much.
277 : : */
278 [ + + ]: 215 : if (!tableam->tuple_tid_valid(scan, tid))
279 [ + - ]: 8 : ereport(ERROR,
280 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
281 : : errmsg("tid (%u, %u) is not valid for relation \"%s\"",
282 : : ItemPointerGetBlockNumberNoCheck(tid),
283 : : ItemPointerGetOffsetNumberNoCheck(tid),
284 : : RelationGetRelationName(rel))));
285 : :
286 : 207 : tableam->tuple_get_latest_tid(scan, tid);
287 : 207 : }
288 : :
289 : :
290 : : /* ----------------------------------------------------------------------------
291 : : * Functions to make modifications a bit simpler.
292 : : * ----------------------------------------------------------------------------
293 : : */
294 : :
295 : : /*
296 : : * simple_table_tuple_insert - insert a tuple
297 : : *
298 : : * Currently, this routine differs from table_tuple_insert only in supplying a
299 : : * default command ID and not allowing access to the speedup options.
300 : : */
301 : : void
302 : 96458 : simple_table_tuple_insert(Relation rel, TupleTableSlot *slot)
303 : : {
304 : 96458 : table_tuple_insert(rel, slot, GetCurrentCommandId(true), 0, NULL);
305 : 96458 : }
306 : :
307 : : /*
308 : : * simple_table_tuple_delete - delete a tuple
309 : : *
310 : : * This routine may be used to delete a tuple when concurrent updates of
311 : : * the target tuple are not expected (for example, because we have a lock
312 : : * on the relation associated with the tuple). Any failure is reported
313 : : * via ereport().
314 : : */
315 : : void
316 : 40318 : simple_table_tuple_delete(Relation rel, ItemPointer tid, Snapshot snapshot)
317 : : {
318 : : TM_Result result;
319 : : TM_FailureData tmfd;
320 : :
321 : 40318 : result = table_tuple_delete(rel, tid,
322 : : GetCurrentCommandId(true),
323 : : 0, snapshot, InvalidSnapshot,
324 : : true /* wait for commit */ ,
325 : : &tmfd);
326 : :
327 [ - + - - : 40318 : switch (result)
- ]
328 : : {
329 : 0 : case TM_SelfModified:
330 : : /* Tuple was already updated in current command? */
331 [ # # ]: 0 : elog(ERROR, "tuple already updated by self");
332 : : break;
333 : :
334 : 40318 : case TM_Ok:
335 : : /* done successfully */
336 : 40318 : break;
337 : :
338 : 0 : case TM_Updated:
339 [ # # ]: 0 : elog(ERROR, "tuple concurrently updated");
340 : : break;
341 : :
342 : 0 : case TM_Deleted:
343 [ # # ]: 0 : elog(ERROR, "tuple concurrently deleted");
344 : : break;
345 : :
346 : 0 : default:
347 [ # # ]: 0 : elog(ERROR, "unrecognized table_tuple_delete status: %u", result);
348 : : break;
349 : : }
350 : 40318 : }
351 : :
352 : : /*
353 : : * simple_table_tuple_update - replace a tuple
354 : : *
355 : : * This routine may be used to update a tuple when concurrent updates of
356 : : * the target tuple are not expected (for example, because we have a lock
357 : : * on the relation associated with the tuple). Any failure is reported
358 : : * via ereport().
359 : : */
360 : : void
361 : 31924 : simple_table_tuple_update(Relation rel, ItemPointer otid,
362 : : TupleTableSlot *slot,
363 : : Snapshot snapshot,
364 : : TU_UpdateIndexes *update_indexes)
365 : : {
366 : : TM_Result result;
367 : : TM_FailureData tmfd;
368 : : LockTupleMode lockmode;
369 : :
370 : 31924 : result = table_tuple_update(rel, otid, slot,
371 : : GetCurrentCommandId(true),
372 : : 0, snapshot, InvalidSnapshot,
373 : : true /* wait for commit */ ,
374 : : &tmfd, &lockmode, update_indexes);
375 : :
376 [ - + - - : 31924 : switch (result)
- ]
377 : : {
378 : 0 : case TM_SelfModified:
379 : : /* Tuple was already updated in current command? */
380 [ # # ]: 0 : elog(ERROR, "tuple already updated by self");
381 : : break;
382 : :
383 : 31924 : case TM_Ok:
384 : : /* done successfully */
385 : 31924 : break;
386 : :
387 : 0 : case TM_Updated:
388 [ # # ]: 0 : elog(ERROR, "tuple concurrently updated");
389 : : break;
390 : :
391 : 0 : case TM_Deleted:
392 [ # # ]: 0 : elog(ERROR, "tuple concurrently deleted");
393 : : break;
394 : :
395 : 0 : default:
396 [ # # ]: 0 : elog(ERROR, "unrecognized table_tuple_update status: %u", result);
397 : : break;
398 : : }
399 : 31924 : }
400 : :
401 : :
402 : : /* ----------------------------------------------------------------------------
403 : : * Helper functions to implement parallel scans for block oriented AMs.
404 : : * ----------------------------------------------------------------------------
405 : : */
406 : :
407 : : Size
408 : 1337 : table_block_parallelscan_estimate(Relation rel)
409 : : {
410 : 1337 : return sizeof(ParallelBlockTableScanDescData);
411 : : }
412 : :
413 : : Size
414 : 1337 : table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
415 : : {
416 : 1337 : ParallelBlockTableScanDesc bpscan = (ParallelBlockTableScanDesc) pscan;
417 : :
418 : 1337 : bpscan->base.phs_locator = rel->rd_locator;
419 : 1337 : bpscan->phs_nblocks = RelationGetNumberOfBlocks(rel);
420 : : /* compare phs_syncscan initialization to similar logic in initscan */
421 : 3681 : bpscan->base.phs_syncscan = synchronize_seqscans &&
422 [ + + + - ]: 2344 : !RelationUsesLocalBuffers(rel) &&
423 [ + + ]: 1007 : bpscan->phs_nblocks > NBuffers / 4;
424 : 1337 : SpinLockInit(&bpscan->phs_mutex);
425 : 1337 : bpscan->phs_startblock = InvalidBlockNumber;
426 : 1337 : bpscan->phs_numblock = InvalidBlockNumber;
427 : 1337 : pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
428 : :
429 : 1337 : return sizeof(ParallelBlockTableScanDescData);
430 : : }
431 : :
432 : : void
433 : 152 : table_block_parallelscan_reinitialize(Relation rel, ParallelTableScanDesc pscan)
434 : : {
435 : 152 : ParallelBlockTableScanDesc bpscan = (ParallelBlockTableScanDesc) pscan;
436 : :
437 : 152 : pg_atomic_write_u64(&bpscan->phs_nallocated, 0);
438 : 152 : }
439 : :
440 : : /*
441 : : * find and set the scan's startblock
442 : : *
443 : : * Determine where the parallel seq scan should start. This function may be
444 : : * called many times, once by each parallel worker. We must be careful only
445 : : * to set the phs_startblock and phs_numblock fields once.
446 : : *
447 : : * Callers may optionally specify a non-InvalidBlockNumber value for
448 : : * 'startblock' to force the scan to start at the given page. Likewise,
449 : : * 'numblocks' can be specified as a non-InvalidBlockNumber to limit the
450 : : * number of blocks to scan to that many blocks.
451 : : */
452 : : void
453 : 2808 : table_block_parallelscan_startblock_init(Relation rel,
454 : : ParallelBlockTableScanWorker pbscanwork,
455 : : ParallelBlockTableScanDesc pbscan,
456 : : BlockNumber startblock,
457 : : BlockNumber numblocks)
458 : : {
459 : : StaticAssertDecl(MaxBlockNumber <= 0xFFFFFFFE,
460 : : "pg_nextpower2_32 may be too small for non-standard BlockNumber width");
461 : :
462 : 2808 : BlockNumber sync_startpage = InvalidBlockNumber;
463 : : BlockNumber scan_nblocks;
464 : :
465 : : /* Reset the state we use for controlling allocation size. */
466 : 2808 : memset(pbscanwork, 0, sizeof(*pbscanwork));
467 : :
468 : 2810 : retry:
469 : : /* Grab the spinlock. */
470 : 2810 : SpinLockAcquire(&pbscan->phs_mutex);
471 : :
472 : : /*
473 : : * When the caller specified a limit on the number of blocks to scan, set
474 : : * that in the ParallelBlockTableScanDesc, if it's not been done by
475 : : * another worker already.
476 : : */
477 [ + + ]: 2810 : if (numblocks != InvalidBlockNumber &&
478 [ + + ]: 80 : pbscan->phs_numblock == InvalidBlockNumber)
479 : : {
480 : 16 : pbscan->phs_numblock = numblocks;
481 : : }
482 : :
483 : : /*
484 : : * If the scan's phs_startblock has not yet been initialized, we must do
485 : : * so now. If a startblock was specified, start there, otherwise if this
486 : : * is not a synchronized scan, we just start at block 0, but if it is a
487 : : * synchronized scan, we must get the starting position from the
488 : : * synchronized scan machinery. We can't hold the spinlock while doing
489 : : * that, though, so release the spinlock, get the information we need, and
490 : : * retry. If nobody else has initialized the scan in the meantime, we'll
491 : : * fill in the value we fetched on the second time through.
492 : : */
493 [ + + ]: 2810 : if (pbscan->phs_startblock == InvalidBlockNumber)
494 : : {
495 [ + + ]: 1326 : if (startblock != InvalidBlockNumber)
496 : 16 : pbscan->phs_startblock = startblock;
497 [ + + ]: 1310 : else if (!pbscan->base.phs_syncscan)
498 : 1306 : pbscan->phs_startblock = 0;
499 [ + + ]: 4 : else if (sync_startpage != InvalidBlockNumber)
500 : 2 : pbscan->phs_startblock = sync_startpage;
501 : : else
502 : : {
503 : 2 : SpinLockRelease(&pbscan->phs_mutex);
504 : 2 : sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
505 : 2 : goto retry;
506 : : }
507 : : }
508 : 2808 : SpinLockRelease(&pbscan->phs_mutex);
509 : :
510 : : /*
511 : : * Figure out how many blocks we're going to scan; either all of them, or
512 : : * just phs_numblock's worth, if a limit has been imposed.
513 : : */
514 [ + + ]: 2808 : if (pbscan->phs_numblock == InvalidBlockNumber)
515 : 2728 : scan_nblocks = pbscan->phs_nblocks;
516 : : else
517 : 80 : scan_nblocks = pbscan->phs_numblock;
518 : :
519 : : /*
520 : : * We determine the chunk size based on scan_nblocks. First we split
521 : : * scan_nblocks into PARALLEL_SEQSCAN_NCHUNKS chunks then we calculate the
522 : : * next highest power of 2 number of the result. This means we split the
523 : : * blocks we're scanning into somewhere between PARALLEL_SEQSCAN_NCHUNKS
524 : : * and PARALLEL_SEQSCAN_NCHUNKS / 2 chunks.
525 : : */
526 [ + + ]: 2808 : pbscanwork->phsw_chunk_size = pg_nextpower2_32(Max(scan_nblocks /
527 : : PARALLEL_SEQSCAN_NCHUNKS, 1));
528 : :
529 : : /*
530 : : * Ensure we don't go over the maximum chunk size with larger tables. This
531 : : * means we may get much more than PARALLEL_SEQSCAN_NCHUNKS for larger
532 : : * tables. Too large a chunk size has been shown to be detrimental to
533 : : * sequential scan performance.
534 : : */
535 : 2808 : pbscanwork->phsw_chunk_size = Min(pbscanwork->phsw_chunk_size,
536 : : PARALLEL_SEQSCAN_MAX_CHUNK_SIZE);
537 : 2808 : }
538 : :
539 : : /*
540 : : * get the next page to scan
541 : : *
542 : : * Get the next page to scan. Even if there are no pages left to scan,
543 : : * another backend could have grabbed a page to scan and not yet finished
544 : : * looking at it, so it doesn't follow that the scan is done when the first
545 : : * backend gets an InvalidBlockNumber return.
546 : : */
547 : : BlockNumber
548 : 146285 : table_block_parallelscan_nextpage(Relation rel,
549 : : ParallelBlockTableScanWorker pbscanwork,
550 : : ParallelBlockTableScanDesc pbscan)
551 : : {
552 : : BlockNumber scan_nblocks;
553 : : BlockNumber page;
554 : : uint64 nallocated;
555 : :
556 : : /*
557 : : * The logic below allocates block numbers out to parallel workers in a
558 : : * way that each worker will receive a set of consecutive block numbers to
559 : : * scan. Earlier versions of this would allocate the next highest block
560 : : * number to the next worker to call this function. This would generally
561 : : * result in workers never receiving consecutive block numbers. Some
562 : : * operating systems would not detect the sequential I/O pattern due to
563 : : * each backend being a different process which could result in poor
564 : : * performance due to inefficient or no readahead. To work around this
565 : : * issue, we now allocate a range of block numbers for each worker and
566 : : * when they come back for another block, we give them the next one in
567 : : * that range until the range is complete. When the worker completes the
568 : : * range of blocks we then allocate another range for it and return the
569 : : * first block number from that range.
570 : : *
571 : : * Here we name these ranges of blocks "chunks". The initial size of
572 : : * these chunks is determined in table_block_parallelscan_startblock_init
573 : : * based on the number of blocks to scan. Towards the end of the scan, we
574 : : * start making reductions in the size of the chunks in order to attempt
575 : : * to divide the remaining work over all the workers as evenly as
576 : : * possible.
577 : : *
578 : : * Here pbscanwork is local worker memory. phsw_chunk_remaining tracks
579 : : * the number of blocks remaining in the chunk. When that reaches 0 then
580 : : * we must allocate a new chunk for the worker.
581 : : *
582 : : * phs_nallocated tracks how many blocks have been allocated to workers
583 : : * already. When phs_nallocated >= rs_nblocks, all blocks have been
584 : : * allocated.
585 : : *
586 : : * Because we use an atomic fetch-and-add to fetch the current value, the
587 : : * phs_nallocated counter will exceed rs_nblocks, because workers will
588 : : * still increment the value, when they try to allocate the next block but
589 : : * all blocks have been allocated already. The counter must be 64 bits
590 : : * wide because of that, to avoid wrapping around when scan_nblocks is
591 : : * close to 2^32.
592 : : *
593 : : * The actual block to return is calculated by adding the counter to the
594 : : * starting block number, modulo phs_nblocks.
595 : : */
596 : :
597 : : /* First, figure out how many blocks we're planning on scanning */
598 [ + + ]: 146285 : if (pbscan->phs_numblock == InvalidBlockNumber)
599 : 145873 : scan_nblocks = pbscan->phs_nblocks;
600 : : else
601 : 412 : scan_nblocks = pbscan->phs_numblock;
602 : :
603 : : /*
604 : : * Now check if we have any remaining blocks in a previous chunk for this
605 : : * worker. We must consume all of the blocks from that before we allocate
606 : : * a new chunk to the worker.
607 : : */
608 [ + + ]: 146285 : if (pbscanwork->phsw_chunk_remaining > 0)
609 : : {
610 : : /*
611 : : * Give them the next block in the range and update the remaining
612 : : * number of blocks.
613 : : */
614 : 17946 : nallocated = ++pbscanwork->phsw_nallocated;
615 : 17946 : pbscanwork->phsw_chunk_remaining--;
616 : : }
617 : : else
618 : : {
619 : : /*
620 : : * When we've only got PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS chunks
621 : : * remaining in the scan, we half the chunk size. Since we reduce the
622 : : * chunk size here, we'll hit this again after doing
623 : : * PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS at the new size. After a few
624 : : * iterations of this, we'll end up doing the last few blocks with the
625 : : * chunk size set to 1.
626 : : */
627 [ + + ]: 128339 : if (pbscanwork->phsw_chunk_size > 1 &&
628 : 3939 : pbscanwork->phsw_nallocated > scan_nblocks -
629 [ + + ]: 3939 : (pbscanwork->phsw_chunk_size * PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS))
630 : 13 : pbscanwork->phsw_chunk_size >>= 1;
631 : :
632 : 128339 : nallocated = pbscanwork->phsw_nallocated =
633 : 128339 : pg_atomic_fetch_add_u64(&pbscan->phs_nallocated,
634 : 128339 : pbscanwork->phsw_chunk_size);
635 : :
636 : : /*
637 : : * Set the remaining number of blocks in this chunk so that subsequent
638 : : * calls from this worker continue on with this chunk until it's done.
639 : : */
640 : 128339 : pbscanwork->phsw_chunk_remaining = pbscanwork->phsw_chunk_size - 1;
641 : : }
642 : :
643 : : /* Check if we've run out of blocks to scan */
644 [ + + ]: 146285 : if (nallocated >= scan_nblocks)
645 : 2808 : page = InvalidBlockNumber; /* all blocks have been allocated */
646 : : else
647 : 143477 : page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
648 : :
649 : : /*
650 : : * Report scan location. Normally, we report the current page number.
651 : : * When we reach the end of the scan, though, we report the starting page,
652 : : * not the ending page, just so the starting positions for later scans
653 : : * doesn't slew backwards. We only report the position at the end of the
654 : : * scan once, though: subsequent callers will report nothing.
655 : : */
656 [ + + ]: 146285 : if (pbscan->base.phs_syncscan)
657 : : {
658 [ + + ]: 22130 : if (page != InvalidBlockNumber)
659 : 22125 : ss_report_location(rel, page);
660 [ + + ]: 5 : else if (nallocated == pbscan->phs_nblocks)
661 : 2 : ss_report_location(rel, pbscan->phs_startblock);
662 : : }
663 : :
664 : 146285 : return page;
665 : : }
666 : :
667 : : /* ----------------------------------------------------------------------------
668 : : * Helper functions to implement relation sizing for block oriented AMs.
669 : : * ----------------------------------------------------------------------------
670 : : */
671 : :
672 : : /*
673 : : * table_block_relation_size
674 : : *
675 : : * If a table AM uses the various relation forks as the sole place where data
676 : : * is stored, and if it uses them in the expected manner (e.g. the actual data
677 : : * is in the main fork rather than some other), it can use this implementation
678 : : * of the relation_size callback rather than implementing its own.
679 : : */
680 : : uint64
681 : 1698001 : table_block_relation_size(Relation rel, ForkNumber forkNumber)
682 : : {
683 : 1698001 : uint64 nblocks = 0;
684 : :
685 : : /* InvalidForkNumber indicates returning the size for all forks */
686 [ - + ]: 1698001 : if (forkNumber == InvalidForkNumber)
687 : : {
688 [ # # ]: 0 : for (int i = 0; i < MAX_FORKNUM; i++)
689 : 0 : nblocks += smgrnblocks(RelationGetSmgr(rel), i);
690 : : }
691 : : else
692 : 1698001 : nblocks = smgrnblocks(RelationGetSmgr(rel), forkNumber);
693 : :
694 : 1697982 : return nblocks * BLCKSZ;
695 : : }
696 : :
697 : : /*
698 : : * table_block_relation_estimate_size
699 : : *
700 : : * This function can't be directly used as the implementation of the
701 : : * relation_estimate_size callback, because it has a few additional parameters.
702 : : * Instead, it is intended to be used as a helper function; the caller can
703 : : * pass through the arguments to its relation_estimate_size function plus the
704 : : * additional values required here.
705 : : *
706 : : * overhead_bytes_per_tuple should contain the approximate number of bytes
707 : : * of storage required to store a tuple above and beyond what is required for
708 : : * the tuple data proper. Typically, this would include things like the
709 : : * size of the tuple header and item pointer. This is only used for query
710 : : * planning, so a table AM where the value is not constant could choose to
711 : : * pass a "best guess".
712 : : *
713 : : * usable_bytes_per_page should contain the approximate number of bytes per
714 : : * page usable for tuple data, excluding the page header and any anticipated
715 : : * special space.
716 : : */
717 : : void
718 : 344974 : table_block_relation_estimate_size(Relation rel, int32 *attr_widths,
719 : : BlockNumber *pages, double *tuples,
720 : : double *allvisfrac,
721 : : Size overhead_bytes_per_tuple,
722 : : Size usable_bytes_per_page)
723 : : {
724 : : BlockNumber curpages;
725 : : BlockNumber relpages;
726 : : double reltuples;
727 : : BlockNumber relallvisible;
728 : : double density;
729 : :
730 : : /* it should have storage, so we can call the smgr */
731 : 344974 : curpages = RelationGetNumberOfBlocks(rel);
732 : :
733 : : /* coerce values in pg_class to more desirable types */
734 : 344974 : relpages = (BlockNumber) rel->rd_rel->relpages;
735 : 344974 : reltuples = (double) rel->rd_rel->reltuples;
736 : 344974 : relallvisible = (BlockNumber) rel->rd_rel->relallvisible;
737 : :
738 : : /*
739 : : * HACK: if the relation has never yet been vacuumed, use a minimum size
740 : : * estimate of 10 pages. The idea here is to avoid assuming a
741 : : * newly-created table is really small, even if it currently is, because
742 : : * that may not be true once some data gets loaded into it. Once a vacuum
743 : : * or analyze cycle has been done on it, it's more reasonable to believe
744 : : * the size is somewhat stable.
745 : : *
746 : : * (Note that this is only an issue if the plan gets cached and used again
747 : : * after the table has been filled. What we're trying to avoid is using a
748 : : * nestloop-type plan on a table that has grown substantially since the
749 : : * plan was made. Normally, autovacuum/autoanalyze will occur once enough
750 : : * inserts have happened and cause cached-plan invalidation; but that
751 : : * doesn't happen instantaneously, and it won't happen at all for cases
752 : : * such as temporary tables.)
753 : : *
754 : : * We test "never vacuumed" by seeing whether reltuples < 0.
755 : : *
756 : : * If the table has inheritance children, we don't apply this heuristic.
757 : : * Totally empty parent tables are quite common, so we should be willing
758 : : * to believe that they are empty.
759 : : */
760 [ + + + + ]: 344974 : if (curpages < 10 &&
761 : 90609 : reltuples < 0 &&
762 [ + + ]: 90609 : !rel->rd_rel->relhassubclass)
763 : 88370 : curpages = 10;
764 : :
765 : : /* report estimated # pages */
766 : 344974 : *pages = curpages;
767 : : /* quick exit if rel is clearly empty */
768 [ + + ]: 344974 : if (curpages == 0)
769 : : {
770 : 16413 : *tuples = 0;
771 : 16413 : *allvisfrac = 0;
772 : 16413 : return;
773 : : }
774 : :
775 : : /* estimate number of tuples from previous tuple density */
776 [ + + + + ]: 328561 : if (reltuples >= 0 && relpages > 0)
777 : 200873 : density = reltuples / (double) relpages;
778 : : else
779 : : {
780 : : /*
781 : : * When we have no data because the relation was never yet vacuumed,
782 : : * estimate tuple width from attribute datatypes. We assume here that
783 : : * the pages are completely full, which is OK for tables but is
784 : : * probably an overestimate for indexes. Fortunately
785 : : * get_relation_info() can clamp the overestimate to the parent
786 : : * table's size.
787 : : *
788 : : * Note: this code intentionally disregards alignment considerations,
789 : : * because (a) that would be gilding the lily considering how crude
790 : : * the estimate is, (b) it creates platform dependencies in the
791 : : * default plans which are kind of a headache for regression testing,
792 : : * and (c) different table AMs might use different padding schemes.
793 : : */
794 : : int32 tuple_width;
795 : : int fillfactor;
796 : :
797 : : /*
798 : : * Without reltuples/relpages, we also need to consider fillfactor.
799 : : * The other branch considers it implicitly by calculating density
800 : : * from actual relpages/reltuples statistics.
801 : : */
802 [ + + ]: 127688 : fillfactor = RelationGetFillFactor(rel, HEAP_DEFAULT_FILLFACTOR);
803 : :
804 : 127688 : tuple_width = get_rel_data_width(rel, attr_widths);
805 : 127688 : tuple_width += overhead_bytes_per_tuple;
806 : : /* note: integer division is intentional here */
807 : 127688 : density = (usable_bytes_per_page * fillfactor / 100) / tuple_width;
808 : : /* There's at least one row on the page, even with low fillfactor. */
809 : 127688 : density = clamp_row_est(density);
810 : : }
811 : 328561 : *tuples = rint(density * (double) curpages);
812 : :
813 : : /*
814 : : * We use relallvisible as-is, rather than scaling it up like we do for
815 : : * the pages and tuples counts, on the theory that any pages added since
816 : : * the last VACUUM are most likely not marked all-visible. But costsize.c
817 : : * wants it converted to a fraction.
818 : : */
819 [ + + - + ]: 328561 : if (relallvisible == 0 || curpages <= 0)
820 : 169237 : *allvisfrac = 0;
821 [ + + ]: 159324 : else if ((double) relallvisible >= curpages)
822 : 83032 : *allvisfrac = 1;
823 : : else
824 : 76292 : *allvisfrac = (double) relallvisible / curpages;
825 : : }
|
