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) 1996-2024, 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 function 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/plancat.h"
28 : #include "port/pg_bitutils.h"
29 : #include "storage/bufmgr.h"
30 : #include "storage/shmem.h"
31 : #include "storage/smgr.h"
32 :
33 : /*
34 : * Constants to control the behavior of block allocation to parallel workers
35 : * during a parallel seqscan. Technically these values do not need to be
36 : * powers of 2, but having them as powers of 2 makes the math more optimal
37 : * and makes the ramp-down stepping more even.
38 : */
39 :
40 : /* The number of I/O chunks we try to break a parallel seqscan down into */
41 : #define PARALLEL_SEQSCAN_NCHUNKS 2048
42 : /* Ramp down size of allocations when we've only this number of chunks left */
43 : #define PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS 64
44 : /* Cap the size of parallel I/O chunks to this number of blocks */
45 : #define PARALLEL_SEQSCAN_MAX_CHUNK_SIZE 8192
46 :
47 : /* GUC variables */
48 : char *default_table_access_method = DEFAULT_TABLE_ACCESS_METHOD;
49 : bool synchronize_seqscans = true;
50 :
51 :
52 : /* ----------------------------------------------------------------------------
53 : * Slot functions.
54 : * ----------------------------------------------------------------------------
55 : */
56 :
57 : const TupleTableSlotOps *
58 23662766 : table_slot_callbacks(Relation relation)
59 : {
60 : const TupleTableSlotOps *tts_cb;
61 :
62 23662766 : if (relation->rd_tableam)
63 23652988 : tts_cb = relation->rd_tableam->slot_callbacks(relation);
64 9778 : else if (relation->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
65 : {
66 : /*
67 : * Historically FDWs expect to store heap tuples in slots. Continue
68 : * handing them one, to make it less painful to adapt FDWs to new
69 : * versions. The cost of a heap slot over a virtual slot is pretty
70 : * small.
71 : */
72 414 : tts_cb = &TTSOpsHeapTuple;
73 : }
74 : else
75 : {
76 : /*
77 : * These need to be supported, as some parts of the code (like COPY)
78 : * need to create slots for such relations too. It seems better to
79 : * centralize the knowledge that a heap slot is the right thing in
80 : * that case here.
81 : */
82 : Assert(relation->rd_rel->relkind == RELKIND_VIEW ||
83 : relation->rd_rel->relkind == RELKIND_PARTITIONED_TABLE);
84 9364 : tts_cb = &TTSOpsVirtual;
85 : }
86 :
87 23662766 : return tts_cb;
88 : }
89 :
90 : TupleTableSlot *
91 23265128 : table_slot_create(Relation relation, List **reglist)
92 : {
93 : const TupleTableSlotOps *tts_cb;
94 : TupleTableSlot *slot;
95 :
96 23265128 : tts_cb = table_slot_callbacks(relation);
97 23265128 : slot = MakeSingleTupleTableSlot(RelationGetDescr(relation), tts_cb);
98 :
99 23265128 : if (reglist)
100 272760 : *reglist = lappend(*reglist, slot);
101 :
102 23265128 : return slot;
103 : }
104 :
105 :
106 : /* ----------------------------------------------------------------------------
107 : * Table scan functions.
108 : * ----------------------------------------------------------------------------
109 : */
110 :
111 : TableScanDesc
112 78926 : table_beginscan_catalog(Relation relation, int nkeys, struct ScanKeyData *key)
113 : {
114 78926 : uint32 flags = SO_TYPE_SEQSCAN |
115 : SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE | SO_TEMP_SNAPSHOT;
116 78926 : Oid relid = RelationGetRelid(relation);
117 78926 : Snapshot snapshot = RegisterSnapshot(GetCatalogSnapshot(relid));
118 :
119 78926 : return relation->rd_tableam->scan_begin(relation, snapshot, nkeys, key,
120 : NULL, flags);
121 : }
122 :
123 :
124 : /* ----------------------------------------------------------------------------
125 : * Parallel table scan related functions.
126 : * ----------------------------------------------------------------------------
127 : */
128 :
129 : Size
130 1044 : table_parallelscan_estimate(Relation rel, Snapshot snapshot)
131 : {
132 1044 : Size sz = 0;
133 :
134 1044 : if (IsMVCCSnapshot(snapshot))
135 898 : sz = add_size(sz, EstimateSnapshotSpace(snapshot));
136 : else
137 : Assert(snapshot == SnapshotAny);
138 :
139 1044 : sz = add_size(sz, rel->rd_tableam->parallelscan_estimate(rel));
140 :
141 1044 : return sz;
142 : }
143 :
144 : void
145 1044 : table_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan,
146 : Snapshot snapshot)
147 : {
148 1044 : Size snapshot_off = rel->rd_tableam->parallelscan_initialize(rel, pscan);
149 :
150 1044 : pscan->phs_snapshot_off = snapshot_off;
151 :
152 1044 : if (IsMVCCSnapshot(snapshot))
153 : {
154 898 : SerializeSnapshot(snapshot, (char *) pscan + pscan->phs_snapshot_off);
155 898 : pscan->phs_snapshot_any = false;
156 : }
157 : else
158 : {
159 : Assert(snapshot == SnapshotAny);
160 146 : pscan->phs_snapshot_any = true;
161 : }
162 1044 : }
163 :
164 : TableScanDesc
165 3848 : table_beginscan_parallel(Relation relation, ParallelTableScanDesc pscan)
166 : {
167 : Snapshot snapshot;
168 3848 : uint32 flags = SO_TYPE_SEQSCAN |
169 : SO_ALLOW_STRAT | SO_ALLOW_SYNC | SO_ALLOW_PAGEMODE;
170 :
171 : Assert(RelationGetRelid(relation) == pscan->phs_relid);
172 :
173 3848 : if (!pscan->phs_snapshot_any)
174 : {
175 : /* Snapshot was serialized -- restore it */
176 3556 : snapshot = RestoreSnapshot((char *) pscan + pscan->phs_snapshot_off);
177 3556 : RegisterSnapshot(snapshot);
178 3556 : flags |= SO_TEMP_SNAPSHOT;
179 : }
180 : else
181 : {
182 : /* SnapshotAny passed by caller (not serialized) */
183 292 : snapshot = SnapshotAny;
184 : }
185 :
186 3848 : return relation->rd_tableam->scan_begin(relation, snapshot, 0, NULL,
187 : pscan, flags);
188 : }
189 :
190 :
191 : /* ----------------------------------------------------------------------------
192 : * Index scan related functions.
193 : * ----------------------------------------------------------------------------
194 : */
195 :
196 : /*
197 : * To perform that check simply start an index scan, create the necessary
198 : * slot, do the heap lookup, and shut everything down again. This could be
199 : * optimized, but is unlikely to matter from a performance POV. If there
200 : * frequently are live index pointers also matching a unique index key, the
201 : * CPU overhead of this routine is unlikely to matter.
202 : *
203 : * Note that *tid may be modified when we return true if the AM supports
204 : * storing multiple row versions reachable via a single index entry (like
205 : * heap's HOT).
206 : */
207 : bool
208 11422224 : table_index_fetch_tuple_check(Relation rel,
209 : ItemPointer tid,
210 : Snapshot snapshot,
211 : bool *all_dead)
212 : {
213 : IndexFetchTableData *scan;
214 : TupleTableSlot *slot;
215 11422224 : bool call_again = false;
216 : bool found;
217 :
218 11422224 : slot = table_slot_create(rel, NULL);
219 11422224 : scan = table_index_fetch_begin(rel);
220 11422224 : found = table_index_fetch_tuple(scan, tid, snapshot, slot, &call_again,
221 : all_dead);
222 11422224 : table_index_fetch_end(scan);
223 11422224 : ExecDropSingleTupleTableSlot(slot);
224 :
225 11422224 : return found;
226 : }
227 :
228 :
229 : /* ------------------------------------------------------------------------
230 : * Functions for non-modifying operations on individual tuples
231 : * ------------------------------------------------------------------------
232 : */
233 :
234 : void
235 312 : table_tuple_get_latest_tid(TableScanDesc scan, ItemPointer tid)
236 : {
237 312 : Relation rel = scan->rs_rd;
238 312 : const TableAmRoutine *tableam = rel->rd_tableam;
239 :
240 : /*
241 : * We don't expect direct calls to table_tuple_get_latest_tid with valid
242 : * CheckXidAlive for catalog or regular tables. See detailed comments in
243 : * xact.c where these variables are declared.
244 : */
245 312 : if (unlikely(TransactionIdIsValid(CheckXidAlive) && !bsysscan))
246 0 : elog(ERROR, "unexpected table_tuple_get_latest_tid call during logical decoding");
247 :
248 : /*
249 : * Since this can be called with user-supplied TID, don't trust the input
250 : * too much.
251 : */
252 312 : if (!tableam->tuple_tid_valid(scan, tid))
253 12 : ereport(ERROR,
254 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
255 : errmsg("tid (%u, %u) is not valid for relation \"%s\"",
256 : ItemPointerGetBlockNumberNoCheck(tid),
257 : ItemPointerGetOffsetNumberNoCheck(tid),
258 : RelationGetRelationName(rel))));
259 :
260 300 : tableam->tuple_get_latest_tid(scan, tid);
261 300 : }
262 :
263 :
264 : /* ----------------------------------------------------------------------------
265 : * Functions to make modifications a bit simpler.
266 : * ----------------------------------------------------------------------------
267 : */
268 :
269 : /*
270 : * simple_table_tuple_insert - insert a tuple
271 : *
272 : * Currently, this routine differs from table_tuple_insert only in supplying a
273 : * default command ID and not allowing access to the speedup options.
274 : */
275 : void
276 153026 : simple_table_tuple_insert(Relation rel, TupleTableSlot *slot)
277 : {
278 153026 : table_tuple_insert(rel, slot, GetCurrentCommandId(true), 0, NULL);
279 153026 : }
280 :
281 : /*
282 : * simple_table_tuple_delete - delete a tuple
283 : *
284 : * This routine may be used to delete a tuple when concurrent updates of
285 : * the target tuple are not expected (for example, because we have a lock
286 : * on the relation associated with the tuple). Any failure is reported
287 : * via ereport().
288 : */
289 : void
290 80602 : simple_table_tuple_delete(Relation rel, ItemPointer tid, Snapshot snapshot)
291 : {
292 : TM_Result result;
293 : TM_FailureData tmfd;
294 :
295 80602 : result = table_tuple_delete(rel, tid,
296 : GetCurrentCommandId(true),
297 : snapshot, InvalidSnapshot,
298 : true /* wait for commit */ ,
299 : &tmfd, false /* changingPart */ );
300 :
301 80602 : switch (result)
302 : {
303 0 : case TM_SelfModified:
304 : /* Tuple was already updated in current command? */
305 0 : elog(ERROR, "tuple already updated by self");
306 : break;
307 :
308 80602 : case TM_Ok:
309 : /* done successfully */
310 80602 : break;
311 :
312 0 : case TM_Updated:
313 0 : elog(ERROR, "tuple concurrently updated");
314 : break;
315 :
316 0 : case TM_Deleted:
317 0 : elog(ERROR, "tuple concurrently deleted");
318 : break;
319 :
320 0 : default:
321 0 : elog(ERROR, "unrecognized table_tuple_delete status: %u", result);
322 : break;
323 : }
324 80602 : }
325 :
326 : /*
327 : * simple_table_tuple_update - replace a tuple
328 : *
329 : * This routine may be used to update a tuple when concurrent updates of
330 : * the target tuple are not expected (for example, because we have a lock
331 : * on the relation associated with the tuple). Any failure is reported
332 : * via ereport().
333 : */
334 : void
335 63822 : simple_table_tuple_update(Relation rel, ItemPointer otid,
336 : TupleTableSlot *slot,
337 : Snapshot snapshot,
338 : TU_UpdateIndexes *update_indexes)
339 : {
340 : TM_Result result;
341 : TM_FailureData tmfd;
342 : LockTupleMode lockmode;
343 :
344 63822 : result = table_tuple_update(rel, otid, slot,
345 : GetCurrentCommandId(true),
346 : snapshot, InvalidSnapshot,
347 : true /* wait for commit */ ,
348 : &tmfd, &lockmode, update_indexes);
349 :
350 63822 : switch (result)
351 : {
352 0 : case TM_SelfModified:
353 : /* Tuple was already updated in current command? */
354 0 : elog(ERROR, "tuple already updated by self");
355 : break;
356 :
357 63822 : case TM_Ok:
358 : /* done successfully */
359 63822 : break;
360 :
361 0 : case TM_Updated:
362 0 : elog(ERROR, "tuple concurrently updated");
363 : break;
364 :
365 0 : case TM_Deleted:
366 0 : elog(ERROR, "tuple concurrently deleted");
367 : break;
368 :
369 0 : default:
370 0 : elog(ERROR, "unrecognized table_tuple_update status: %u", result);
371 : break;
372 : }
373 63822 : }
374 :
375 :
376 : /* ----------------------------------------------------------------------------
377 : * Helper functions to implement parallel scans for block oriented AMs.
378 : * ----------------------------------------------------------------------------
379 : */
380 :
381 : Size
382 1044 : table_block_parallelscan_estimate(Relation rel)
383 : {
384 1044 : return sizeof(ParallelBlockTableScanDescData);
385 : }
386 :
387 : Size
388 1044 : table_block_parallelscan_initialize(Relation rel, ParallelTableScanDesc pscan)
389 : {
390 1044 : ParallelBlockTableScanDesc bpscan = (ParallelBlockTableScanDesc) pscan;
391 :
392 1044 : bpscan->base.phs_relid = RelationGetRelid(rel);
393 1044 : bpscan->phs_nblocks = RelationGetNumberOfBlocks(rel);
394 : /* compare phs_syncscan initialization to similar logic in initscan */
395 2792 : bpscan->base.phs_syncscan = synchronize_seqscans &&
396 1748 : !RelationUsesLocalBuffers(rel) &&
397 704 : bpscan->phs_nblocks > NBuffers / 4;
398 1044 : SpinLockInit(&bpscan->phs_mutex);
399 1044 : bpscan->phs_startblock = InvalidBlockNumber;
400 1044 : pg_atomic_init_u64(&bpscan->phs_nallocated, 0);
401 :
402 1044 : return sizeof(ParallelBlockTableScanDescData);
403 : }
404 :
405 : void
406 228 : table_block_parallelscan_reinitialize(Relation rel, ParallelTableScanDesc pscan)
407 : {
408 228 : ParallelBlockTableScanDesc bpscan = (ParallelBlockTableScanDesc) pscan;
409 :
410 228 : pg_atomic_write_u64(&bpscan->phs_nallocated, 0);
411 228 : }
412 :
413 : /*
414 : * find and set the scan's startblock
415 : *
416 : * Determine where the parallel seq scan should start. This function may be
417 : * called many times, once by each parallel worker. We must be careful only
418 : * to set the startblock once.
419 : */
420 : void
421 2782 : table_block_parallelscan_startblock_init(Relation rel,
422 : ParallelBlockTableScanWorker pbscanwork,
423 : ParallelBlockTableScanDesc pbscan)
424 : {
425 2782 : BlockNumber sync_startpage = InvalidBlockNumber;
426 :
427 : /* Reset the state we use for controlling allocation size. */
428 2782 : memset(pbscanwork, 0, sizeof(*pbscanwork));
429 :
430 : StaticAssertStmt(MaxBlockNumber <= 0xFFFFFFFE,
431 : "pg_nextpower2_32 may be too small for non-standard BlockNumber width");
432 :
433 : /*
434 : * We determine the chunk size based on the size of the relation. First we
435 : * split the relation into PARALLEL_SEQSCAN_NCHUNKS chunks but we then
436 : * take the next highest power of 2 number of the chunk size. This means
437 : * we split the relation into somewhere between PARALLEL_SEQSCAN_NCHUNKS
438 : * and PARALLEL_SEQSCAN_NCHUNKS / 2 chunks.
439 : */
440 2782 : pbscanwork->phsw_chunk_size = pg_nextpower2_32(Max(pbscan->phs_nblocks /
441 : PARALLEL_SEQSCAN_NCHUNKS, 1));
442 :
443 : /*
444 : * Ensure we don't go over the maximum chunk size with larger tables. This
445 : * means we may get much more than PARALLEL_SEQSCAN_NCHUNKS for larger
446 : * tables. Too large a chunk size has been shown to be detrimental to
447 : * synchronous scan performance.
448 : */
449 2782 : pbscanwork->phsw_chunk_size = Min(pbscanwork->phsw_chunk_size,
450 : PARALLEL_SEQSCAN_MAX_CHUNK_SIZE);
451 :
452 2784 : retry:
453 : /* Grab the spinlock. */
454 2784 : SpinLockAcquire(&pbscan->phs_mutex);
455 :
456 : /*
457 : * If the scan's startblock has not yet been initialized, we must do so
458 : * now. If this is not a synchronized scan, we just start at block 0, but
459 : * if it is a synchronized scan, we must get the starting position from
460 : * the synchronized scan machinery. We can't hold the spinlock while
461 : * doing that, though, so release the spinlock, get the information we
462 : * need, and retry. If nobody else has initialized the scan in the
463 : * meantime, we'll fill in the value we fetched on the second time
464 : * through.
465 : */
466 2784 : if (pbscan->phs_startblock == InvalidBlockNumber)
467 : {
468 1024 : if (!pbscan->base.phs_syncscan)
469 1020 : pbscan->phs_startblock = 0;
470 4 : else if (sync_startpage != InvalidBlockNumber)
471 2 : pbscan->phs_startblock = sync_startpage;
472 : else
473 : {
474 2 : SpinLockRelease(&pbscan->phs_mutex);
475 2 : sync_startpage = ss_get_location(rel, pbscan->phs_nblocks);
476 2 : goto retry;
477 : }
478 : }
479 2782 : SpinLockRelease(&pbscan->phs_mutex);
480 2782 : }
481 :
482 : /*
483 : * get the next page to scan
484 : *
485 : * Get the next page to scan. Even if there are no pages left to scan,
486 : * another backend could have grabbed a page to scan and not yet finished
487 : * looking at it, so it doesn't follow that the scan is done when the first
488 : * backend gets an InvalidBlockNumber return.
489 : */
490 : BlockNumber
491 200950 : table_block_parallelscan_nextpage(Relation rel,
492 : ParallelBlockTableScanWorker pbscanwork,
493 : ParallelBlockTableScanDesc pbscan)
494 : {
495 : BlockNumber page;
496 : uint64 nallocated;
497 :
498 : /*
499 : * The logic below allocates block numbers out to parallel workers in a
500 : * way that each worker will receive a set of consecutive block numbers to
501 : * scan. Earlier versions of this would allocate the next highest block
502 : * number to the next worker to call this function. This would generally
503 : * result in workers never receiving consecutive block numbers. Some
504 : * operating systems would not detect the sequential I/O pattern due to
505 : * each backend being a different process which could result in poor
506 : * performance due to inefficient or no readahead. To work around this
507 : * issue, we now allocate a range of block numbers for each worker and
508 : * when they come back for another block, we give them the next one in
509 : * that range until the range is complete. When the worker completes the
510 : * range of blocks we then allocate another range for it and return the
511 : * first block number from that range.
512 : *
513 : * Here we name these ranges of blocks "chunks". The initial size of
514 : * these chunks is determined in table_block_parallelscan_startblock_init
515 : * based on the size of the relation. Towards the end of the scan, we
516 : * start making reductions in the size of the chunks in order to attempt
517 : * to divide the remaining work over all the workers as evenly as
518 : * possible.
519 : *
520 : * Here pbscanwork is local worker memory. phsw_chunk_remaining tracks
521 : * the number of blocks remaining in the chunk. When that reaches 0 then
522 : * we must allocate a new chunk for the worker.
523 : *
524 : * phs_nallocated tracks how many blocks have been allocated to workers
525 : * already. When phs_nallocated >= rs_nblocks, all blocks have been
526 : * allocated.
527 : *
528 : * Because we use an atomic fetch-and-add to fetch the current value, the
529 : * phs_nallocated counter will exceed rs_nblocks, because workers will
530 : * still increment the value, when they try to allocate the next block but
531 : * all blocks have been allocated already. The counter must be 64 bits
532 : * wide because of that, to avoid wrapping around when rs_nblocks is close
533 : * to 2^32.
534 : *
535 : * The actual block to return is calculated by adding the counter to the
536 : * starting block number, modulo nblocks.
537 : */
538 :
539 : /*
540 : * First check if we have any remaining blocks in a previous chunk for
541 : * this worker. We must consume all of the blocks from that before we
542 : * allocate a new chunk to the worker.
543 : */
544 200950 : if (pbscanwork->phsw_chunk_remaining > 0)
545 : {
546 : /*
547 : * Give them the next block in the range and update the remaining
548 : * number of blocks.
549 : */
550 13026 : nallocated = ++pbscanwork->phsw_nallocated;
551 13026 : pbscanwork->phsw_chunk_remaining--;
552 : }
553 : else
554 : {
555 : /*
556 : * When we've only got PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS chunks
557 : * remaining in the scan, we half the chunk size. Since we reduce the
558 : * chunk size here, we'll hit this again after doing
559 : * PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS at the new size. After a few
560 : * iterations of this, we'll end up doing the last few blocks with the
561 : * chunk size set to 1.
562 : */
563 187924 : if (pbscanwork->phsw_chunk_size > 1 &&
564 4430 : pbscanwork->phsw_nallocated > pbscan->phs_nblocks -
565 4430 : (pbscanwork->phsw_chunk_size * PARALLEL_SEQSCAN_RAMPDOWN_CHUNKS))
566 8 : pbscanwork->phsw_chunk_size >>= 1;
567 :
568 187924 : nallocated = pbscanwork->phsw_nallocated =
569 187924 : pg_atomic_fetch_add_u64(&pbscan->phs_nallocated,
570 187924 : pbscanwork->phsw_chunk_size);
571 :
572 : /*
573 : * Set the remaining number of blocks in this chunk so that subsequent
574 : * calls from this worker continue on with this chunk until it's done.
575 : */
576 187924 : pbscanwork->phsw_chunk_remaining = pbscanwork->phsw_chunk_size - 1;
577 : }
578 :
579 200950 : if (nallocated >= pbscan->phs_nblocks)
580 2782 : page = InvalidBlockNumber; /* all blocks have been allocated */
581 : else
582 198168 : page = (nallocated + pbscan->phs_startblock) % pbscan->phs_nblocks;
583 :
584 : /*
585 : * Report scan location. Normally, we report the current page number.
586 : * When we reach the end of the scan, though, we report the starting page,
587 : * not the ending page, just so the starting positions for later scans
588 : * doesn't slew backwards. We only report the position at the end of the
589 : * scan once, though: subsequent callers will report nothing.
590 : */
591 200950 : if (pbscan->base.phs_syncscan)
592 : {
593 17704 : if (page != InvalidBlockNumber)
594 17700 : ss_report_location(rel, page);
595 4 : else if (nallocated == pbscan->phs_nblocks)
596 2 : ss_report_location(rel, pbscan->phs_startblock);
597 : }
598 :
599 200950 : return page;
600 : }
601 :
602 : /* ----------------------------------------------------------------------------
603 : * Helper functions to implement relation sizing for block oriented AMs.
604 : * ----------------------------------------------------------------------------
605 : */
606 :
607 : /*
608 : * table_block_relation_size
609 : *
610 : * If a table AM uses the various relation forks as the sole place where data
611 : * is stored, and if it uses them in the expected manner (e.g. the actual data
612 : * is in the main fork rather than some other), it can use this implementation
613 : * of the relation_size callback rather than implementing its own.
614 : */
615 : uint64
616 2125032 : table_block_relation_size(Relation rel, ForkNumber forkNumber)
617 : {
618 2125032 : uint64 nblocks = 0;
619 :
620 : /* InvalidForkNumber indicates returning the size for all forks */
621 2125032 : if (forkNumber == InvalidForkNumber)
622 : {
623 0 : for (int i = 0; i < MAX_FORKNUM; i++)
624 0 : nblocks += smgrnblocks(RelationGetSmgr(rel), i);
625 : }
626 : else
627 2125032 : nblocks = smgrnblocks(RelationGetSmgr(rel), forkNumber);
628 :
629 2124994 : return nblocks * BLCKSZ;
630 : }
631 :
632 : /*
633 : * table_block_relation_estimate_size
634 : *
635 : * This function can't be directly used as the implementation of the
636 : * relation_estimate_size callback, because it has a few additional parameters.
637 : * Instead, it is intended to be used as a helper function; the caller can
638 : * pass through the arguments to its relation_estimate_size function plus the
639 : * additional values required here.
640 : *
641 : * overhead_bytes_per_tuple should contain the approximate number of bytes
642 : * of storage required to store a tuple above and beyond what is required for
643 : * the tuple data proper. Typically, this would include things like the
644 : * size of the tuple header and item pointer. This is only used for query
645 : * planning, so a table AM where the value is not constant could choose to
646 : * pass a "best guess".
647 : *
648 : * usable_bytes_per_page should contain the approximate number of bytes per
649 : * page usable for tuple data, excluding the page header and any anticipated
650 : * special space.
651 : */
652 : void
653 388914 : table_block_relation_estimate_size(Relation rel, int32 *attr_widths,
654 : BlockNumber *pages, double *tuples,
655 : double *allvisfrac,
656 : Size overhead_bytes_per_tuple,
657 : Size usable_bytes_per_page)
658 : {
659 : BlockNumber curpages;
660 : BlockNumber relpages;
661 : double reltuples;
662 : BlockNumber relallvisible;
663 : double density;
664 :
665 : /* it should have storage, so we can call the smgr */
666 388914 : curpages = RelationGetNumberOfBlocks(rel);
667 :
668 : /* coerce values in pg_class to more desirable types */
669 388914 : relpages = (BlockNumber) rel->rd_rel->relpages;
670 388914 : reltuples = (double) rel->rd_rel->reltuples;
671 388914 : relallvisible = (BlockNumber) rel->rd_rel->relallvisible;
672 :
673 : /*
674 : * HACK: if the relation has never yet been vacuumed, use a minimum size
675 : * estimate of 10 pages. The idea here is to avoid assuming a
676 : * newly-created table is really small, even if it currently is, because
677 : * that may not be true once some data gets loaded into it. Once a vacuum
678 : * or analyze cycle has been done on it, it's more reasonable to believe
679 : * the size is somewhat stable.
680 : *
681 : * (Note that this is only an issue if the plan gets cached and used again
682 : * after the table has been filled. What we're trying to avoid is using a
683 : * nestloop-type plan on a table that has grown substantially since the
684 : * plan was made. Normally, autovacuum/autoanalyze will occur once enough
685 : * inserts have happened and cause cached-plan invalidation; but that
686 : * doesn't happen instantaneously, and it won't happen at all for cases
687 : * such as temporary tables.)
688 : *
689 : * We test "never vacuumed" by seeing whether reltuples < 0.
690 : *
691 : * If the table has inheritance children, we don't apply this heuristic.
692 : * Totally empty parent tables are quite common, so we should be willing
693 : * to believe that they are empty.
694 : */
695 388914 : if (curpages < 10 &&
696 104030 : reltuples < 0 &&
697 104030 : !rel->rd_rel->relhassubclass)
698 101550 : curpages = 10;
699 :
700 : /* report estimated # pages */
701 388914 : *pages = curpages;
702 : /* quick exit if rel is clearly empty */
703 388914 : if (curpages == 0)
704 : {
705 14744 : *tuples = 0;
706 14744 : *allvisfrac = 0;
707 14744 : return;
708 : }
709 :
710 : /* estimate number of tuples from previous tuple density */
711 374170 : if (reltuples >= 0 && relpages > 0)
712 241120 : density = reltuples / (double) relpages;
713 : else
714 : {
715 : /*
716 : * When we have no data because the relation was never yet vacuumed,
717 : * estimate tuple width from attribute datatypes. We assume here that
718 : * the pages are completely full, which is OK for tables but is
719 : * probably an overestimate for indexes. Fortunately
720 : * get_relation_info() can clamp the overestimate to the parent
721 : * table's size.
722 : *
723 : * Note: this code intentionally disregards alignment considerations,
724 : * because (a) that would be gilding the lily considering how crude
725 : * the estimate is, (b) it creates platform dependencies in the
726 : * default plans which are kind of a headache for regression testing,
727 : * and (c) different table AMs might use different padding schemes.
728 : */
729 : int32 tuple_width;
730 : int fillfactor;
731 :
732 : /*
733 : * Without reltuples/relpages, we also need to consider fillfactor.
734 : * The other branch considers it implicitly by calculating density
735 : * from actual relpages/reltuples statistics.
736 : */
737 133050 : fillfactor = RelationGetFillFactor(rel, HEAP_DEFAULT_FILLFACTOR);
738 :
739 133050 : tuple_width = get_rel_data_width(rel, attr_widths);
740 133050 : tuple_width += overhead_bytes_per_tuple;
741 : /* note: integer division is intentional here */
742 133050 : density = (usable_bytes_per_page * fillfactor / 100) / tuple_width;
743 : }
744 374170 : *tuples = rint(density * (double) curpages);
745 :
746 : /*
747 : * We use relallvisible as-is, rather than scaling it up like we do for
748 : * the pages and tuples counts, on the theory that any pages added since
749 : * the last VACUUM are most likely not marked all-visible. But costsize.c
750 : * wants it converted to a fraction.
751 : */
752 374170 : if (relallvisible == 0 || curpages <= 0)
753 185204 : *allvisfrac = 0;
754 188966 : else if ((double) relallvisible >= curpages)
755 93162 : *allvisfrac = 1;
756 : else
757 95804 : *allvisfrac = (double) relallvisible / curpages;
758 : }
|
