Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * pruneheap.c
4 : * heap page pruning and HOT-chain management code
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/access/heap/pruneheap.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include "access/heapam.h"
18 : #include "access/heapam_xlog.h"
19 : #include "access/htup_details.h"
20 : #include "access/multixact.h"
21 : #include "access/transam.h"
22 : #include "access/visibilitymapdefs.h"
23 : #include "access/xlog.h"
24 : #include "access/xloginsert.h"
25 : #include "commands/vacuum.h"
26 : #include "executor/instrument.h"
27 : #include "miscadmin.h"
28 : #include "pgstat.h"
29 : #include "storage/bufmgr.h"
30 : #include "utils/rel.h"
31 : #include "utils/snapmgr.h"
32 :
33 : /* Working data for heap_page_prune_and_freeze() and subroutines */
34 : typedef struct
35 : {
36 : /*-------------------------------------------------------
37 : * Arguments passed to heap_page_prune_and_freeze()
38 : *-------------------------------------------------------
39 : */
40 :
41 : /* tuple visibility test, initialized for the relation */
42 : GlobalVisState *vistest;
43 : /* whether or not dead items can be set LP_UNUSED during pruning */
44 : bool mark_unused_now;
45 : /* whether to attempt freezing tuples */
46 : bool attempt_freeze;
47 : struct VacuumCutoffs *cutoffs;
48 :
49 : /*-------------------------------------------------------
50 : * Fields describing what to do to the page
51 : *-------------------------------------------------------
52 : */
53 : TransactionId new_prune_xid; /* new prune hint value */
54 : TransactionId latest_xid_removed;
55 : int nredirected; /* numbers of entries in arrays below */
56 : int ndead;
57 : int nunused;
58 : int nfrozen;
59 : /* arrays that accumulate indexes of items to be changed */
60 : OffsetNumber redirected[MaxHeapTuplesPerPage * 2];
61 : OffsetNumber nowdead[MaxHeapTuplesPerPage];
62 : OffsetNumber nowunused[MaxHeapTuplesPerPage];
63 : HeapTupleFreeze frozen[MaxHeapTuplesPerPage];
64 :
65 : /*-------------------------------------------------------
66 : * Working state for HOT chain processing
67 : *-------------------------------------------------------
68 : */
69 :
70 : /*
71 : * 'root_items' contains offsets of all LP_REDIRECT line pointers and
72 : * normal non-HOT tuples. They can be stand-alone items or the first item
73 : * in a HOT chain. 'heaponly_items' contains heap-only tuples which can
74 : * only be removed as part of a HOT chain.
75 : */
76 : int nroot_items;
77 : OffsetNumber root_items[MaxHeapTuplesPerPage];
78 : int nheaponly_items;
79 : OffsetNumber heaponly_items[MaxHeapTuplesPerPage];
80 :
81 : /*
82 : * processed[offnum] is true if item at offnum has been processed.
83 : *
84 : * This needs to be MaxHeapTuplesPerPage + 1 long as FirstOffsetNumber is
85 : * 1. Otherwise every access would need to subtract 1.
86 : */
87 : bool processed[MaxHeapTuplesPerPage + 1];
88 :
89 : /*
90 : * Tuple visibility is only computed once for each tuple, for correctness
91 : * and efficiency reasons; see comment in heap_page_prune_and_freeze() for
92 : * details. This is of type int8[], instead of HTSV_Result[], so we can
93 : * use -1 to indicate no visibility has been computed, e.g. for LP_DEAD
94 : * items.
95 : *
96 : * This needs to be MaxHeapTuplesPerPage + 1 long as FirstOffsetNumber is
97 : * 1. Otherwise every access would need to subtract 1.
98 : */
99 : int8 htsv[MaxHeapTuplesPerPage + 1];
100 :
101 : /*
102 : * Freezing-related state.
103 : */
104 : HeapPageFreeze pagefrz;
105 :
106 : /*-------------------------------------------------------
107 : * Information about what was done
108 : *
109 : * These fields are not used by pruning itself for the most part, but are
110 : * used to collect information about what was pruned and what state the
111 : * page is in after pruning, for the benefit of the caller. They are
112 : * copied to the caller's PruneFreezeResult at the end.
113 : * -------------------------------------------------------
114 : */
115 :
116 : int ndeleted; /* Number of tuples deleted from the page */
117 :
118 : /* Number of live and recently dead tuples, after pruning */
119 : int live_tuples;
120 : int recently_dead_tuples;
121 :
122 : /* Whether or not the page makes rel truncation unsafe */
123 : bool hastup;
124 :
125 : /*
126 : * LP_DEAD items on the page after pruning. Includes existing LP_DEAD
127 : * items
128 : */
129 : int lpdead_items; /* number of items in the array */
130 : OffsetNumber *deadoffsets; /* points directly to presult->deadoffsets */
131 :
132 : /*
133 : * all_visible and all_frozen indicate if the all-visible and all-frozen
134 : * bits in the visibility map can be set for this page after pruning.
135 : *
136 : * visibility_cutoff_xid is the newest xmin of live tuples on the page.
137 : * The caller can use it as the conflict horizon, when setting the VM
138 : * bits. It is only valid if we froze some tuples, and all_frozen is
139 : * true.
140 : *
141 : * NOTE: all_visible and all_frozen don't include LP_DEAD items. That's
142 : * convenient for heap_page_prune_and_freeze(), to use them to decide
143 : * whether to freeze the page or not. The all_visible and all_frozen
144 : * values returned to the caller are adjusted to include LP_DEAD items at
145 : * the end.
146 : *
147 : * all_frozen should only be considered valid if all_visible is also set;
148 : * we don't bother to clear the all_frozen flag every time we clear the
149 : * all_visible flag.
150 : */
151 : bool all_visible;
152 : bool all_frozen;
153 : TransactionId visibility_cutoff_xid;
154 : } PruneState;
155 :
156 : /* Local functions */
157 : static HTSV_Result heap_prune_satisfies_vacuum(PruneState *prstate,
158 : HeapTuple tup,
159 : Buffer buffer);
160 : static inline HTSV_Result htsv_get_valid_status(int status);
161 : static void heap_prune_chain(Page page, BlockNumber blockno, OffsetNumber maxoff,
162 : OffsetNumber rootoffnum, PruneState *prstate);
163 : static void heap_prune_record_prunable(PruneState *prstate, TransactionId xid);
164 : static void heap_prune_record_redirect(PruneState *prstate,
165 : OffsetNumber offnum, OffsetNumber rdoffnum,
166 : bool was_normal);
167 : static void heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
168 : bool was_normal);
169 : static void heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
170 : bool was_normal);
171 : static void heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal);
172 :
173 : static void heap_prune_record_unchanged_lp_unused(Page page, PruneState *prstate, OffsetNumber offnum);
174 : static void heap_prune_record_unchanged_lp_normal(Page page, PruneState *prstate, OffsetNumber offnum);
175 : static void heap_prune_record_unchanged_lp_dead(Page page, PruneState *prstate, OffsetNumber offnum);
176 : static void heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum);
177 :
178 : static void page_verify_redirects(Page page);
179 :
180 : static bool heap_page_will_freeze(Relation relation, Buffer buffer,
181 : bool did_tuple_hint_fpi, bool do_prune, bool do_hint_prune,
182 : PruneState *prstate);
183 :
184 :
185 : /*
186 : * Optionally prune and repair fragmentation in the specified page.
187 : *
188 : * This is an opportunistic function. It will perform housekeeping
189 : * only if the page heuristically looks like a candidate for pruning and we
190 : * can acquire buffer cleanup lock without blocking.
191 : *
192 : * Note: this is called quite often. It's important that it fall out quickly
193 : * if there's not any use in pruning.
194 : *
195 : * Caller must have pin on the buffer, and must *not* have a lock on it.
196 : */
197 : void
198 32580806 : heap_page_prune_opt(Relation relation, Buffer buffer)
199 : {
200 32580806 : Page page = BufferGetPage(buffer);
201 : TransactionId prune_xid;
202 : GlobalVisState *vistest;
203 : Size minfree;
204 :
205 : /*
206 : * We can't write WAL in recovery mode, so there's no point trying to
207 : * clean the page. The primary will likely issue a cleaning WAL record
208 : * soon anyway, so this is no particular loss.
209 : */
210 32580806 : if (RecoveryInProgress())
211 424012 : return;
212 :
213 : /*
214 : * First check whether there's any chance there's something to prune,
215 : * determining the appropriate horizon is a waste if there's no prune_xid
216 : * (i.e. no updates/deletes left potentially dead tuples around).
217 : */
218 32156794 : prune_xid = ((PageHeader) page)->pd_prune_xid;
219 32156794 : if (!TransactionIdIsValid(prune_xid))
220 16659586 : return;
221 :
222 : /*
223 : * Check whether prune_xid indicates that there may be dead rows that can
224 : * be cleaned up.
225 : */
226 15497208 : vistest = GlobalVisTestFor(relation);
227 :
228 15497208 : if (!GlobalVisTestIsRemovableXid(vistest, prune_xid))
229 12999882 : return;
230 :
231 : /*
232 : * We prune when a previous UPDATE failed to find enough space on the page
233 : * for a new tuple version, or when free space falls below the relation's
234 : * fill-factor target (but not less than 10%).
235 : *
236 : * Checking free space here is questionable since we aren't holding any
237 : * lock on the buffer; in the worst case we could get a bogus answer. It's
238 : * unlikely to be *seriously* wrong, though, since reading either pd_lower
239 : * or pd_upper is probably atomic. Avoiding taking a lock seems more
240 : * important than sometimes getting a wrong answer in what is after all
241 : * just a heuristic estimate.
242 : */
243 2497326 : minfree = RelationGetTargetPageFreeSpace(relation,
244 : HEAP_DEFAULT_FILLFACTOR);
245 2497326 : minfree = Max(minfree, BLCKSZ / 10);
246 :
247 2497326 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
248 : {
249 : /* OK, try to get exclusive buffer lock */
250 81722 : if (!ConditionalLockBufferForCleanup(buffer))
251 1028 : return;
252 :
253 : /*
254 : * Now that we have buffer lock, get accurate information about the
255 : * page's free space, and recheck the heuristic about whether to
256 : * prune.
257 : */
258 80694 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
259 : {
260 : OffsetNumber dummy_off_loc;
261 : PruneFreezeResult presult;
262 :
263 : /*
264 : * For now, pass mark_unused_now as false regardless of whether or
265 : * not the relation has indexes, since we cannot safely determine
266 : * that during on-access pruning with the current implementation.
267 : */
268 80694 : heap_page_prune_and_freeze(relation, buffer, vistest, 0,
269 : NULL, &presult, PRUNE_ON_ACCESS, &dummy_off_loc, NULL, NULL);
270 :
271 : /*
272 : * Report the number of tuples reclaimed to pgstats. This is
273 : * presult.ndeleted minus the number of newly-LP_DEAD-set items.
274 : *
275 : * We derive the number of dead tuples like this to avoid totally
276 : * forgetting about items that were set to LP_DEAD, since they
277 : * still need to be cleaned up by VACUUM. We only want to count
278 : * heap-only tuples that just became LP_UNUSED in our report,
279 : * which don't.
280 : *
281 : * VACUUM doesn't have to compensate in the same way when it
282 : * tracks ndeleted, since it will set the same LP_DEAD items to
283 : * LP_UNUSED separately.
284 : */
285 80694 : if (presult.ndeleted > presult.nnewlpdead)
286 37260 : pgstat_update_heap_dead_tuples(relation,
287 37260 : presult.ndeleted - presult.nnewlpdead);
288 : }
289 :
290 : /* And release buffer lock */
291 80694 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
292 :
293 : /*
294 : * We avoid reuse of any free space created on the page by unrelated
295 : * UPDATEs/INSERTs by opting to not update the FSM at this point. The
296 : * free space should be reused by UPDATEs to *this* page.
297 : */
298 : }
299 : }
300 :
301 : /*
302 : * Decide whether to proceed with freezing according to the freeze plans
303 : * prepared for the given heap buffer. If freezing is chosen, this function
304 : * performs several pre-freeze checks.
305 : *
306 : * The values of do_prune, do_hint_prune, and did_tuple_hint_fpi must be
307 : * determined before calling this function.
308 : *
309 : * prstate is both an input and output parameter.
310 : *
311 : * Returns true if we should apply the freeze plans and freeze tuples on the
312 : * page, and false otherwise.
313 : */
314 : static bool
315 1197090 : heap_page_will_freeze(Relation relation, Buffer buffer,
316 : bool did_tuple_hint_fpi,
317 : bool do_prune,
318 : bool do_hint_prune,
319 : PruneState *prstate)
320 : {
321 1197090 : bool do_freeze = false;
322 :
323 : /*
324 : * If the caller specified we should not attempt to freeze any tuples,
325 : * validate that everything is in the right state and return.
326 : */
327 1197090 : if (!prstate->attempt_freeze)
328 : {
329 : Assert(!prstate->all_frozen && prstate->nfrozen == 0);
330 : Assert(prstate->lpdead_items == 0 || !prstate->all_visible);
331 80694 : return false;
332 : }
333 :
334 1116396 : if (prstate->pagefrz.freeze_required)
335 : {
336 : /*
337 : * heap_prepare_freeze_tuple indicated that at least one XID/MXID from
338 : * before FreezeLimit/MultiXactCutoff is present. Must freeze to
339 : * advance relfrozenxid/relminmxid.
340 : */
341 41630 : do_freeze = true;
342 : }
343 : else
344 : {
345 : /*
346 : * Opportunistically freeze the page if we are generating an FPI
347 : * anyway and if doing so means that we can set the page all-frozen
348 : * afterwards (might not happen until VACUUM's final heap pass).
349 : *
350 : * XXX: Previously, we knew if pruning emitted an FPI by checking
351 : * pgWalUsage.wal_fpi before and after pruning. Once the freeze and
352 : * prune records were combined, this heuristic couldn't be used
353 : * anymore. The opportunistic freeze heuristic must be improved;
354 : * however, for now, try to approximate the old logic.
355 : */
356 1074766 : if (prstate->all_visible && prstate->all_frozen && prstate->nfrozen > 0)
357 : {
358 : /*
359 : * Freezing would make the page all-frozen. Have already emitted
360 : * an FPI or will do so anyway?
361 : */
362 29916 : if (RelationNeedsWAL(relation))
363 : {
364 26606 : if (did_tuple_hint_fpi)
365 3248 : do_freeze = true;
366 23358 : else if (do_prune)
367 : {
368 2454 : if (XLogCheckBufferNeedsBackup(buffer))
369 1488 : do_freeze = true;
370 : }
371 20904 : else if (do_hint_prune)
372 : {
373 14 : if (XLogHintBitIsNeeded() && XLogCheckBufferNeedsBackup(buffer))
374 10 : do_freeze = true;
375 : }
376 : }
377 : }
378 : }
379 :
380 1116396 : if (do_freeze)
381 : {
382 : /*
383 : * Validate the tuples we will be freezing before entering the
384 : * critical section.
385 : */
386 46376 : heap_pre_freeze_checks(buffer, prstate->frozen, prstate->nfrozen);
387 : }
388 1070020 : else if (prstate->nfrozen > 0)
389 : {
390 : /*
391 : * The page contained some tuples that were not already frozen, and we
392 : * chose not to freeze them now. The page won't be all-frozen then.
393 : */
394 : Assert(!prstate->pagefrz.freeze_required);
395 :
396 26374 : prstate->all_frozen = false;
397 26374 : prstate->nfrozen = 0; /* avoid miscounts in instrumentation */
398 : }
399 : else
400 : {
401 : /*
402 : * We have no freeze plans to execute. The page might already be
403 : * all-frozen (perhaps only following pruning), though. Such pages
404 : * can be marked all-frozen in the VM by our caller, even though none
405 : * of its tuples were newly frozen here.
406 : */
407 : }
408 :
409 1116396 : return do_freeze;
410 : }
411 :
412 :
413 : /*
414 : * Prune and repair fragmentation and potentially freeze tuples on the
415 : * specified page.
416 : *
417 : * Caller must have pin and buffer cleanup lock on the page. Note that we
418 : * don't update the FSM information for page on caller's behalf. Caller might
419 : * also need to account for a reduction in the length of the line pointer
420 : * array following array truncation by us.
421 : *
422 : * If the HEAP_PRUNE_FREEZE option is set, we will also freeze tuples if it's
423 : * required in order to advance relfrozenxid / relminmxid, or if it's
424 : * considered advantageous for overall system performance to do so now. The
425 : * 'cutoffs', 'presult', 'new_relfrozen_xid' and 'new_relmin_mxid' arguments
426 : * are required when freezing. When HEAP_PRUNE_FREEZE option is set, we also
427 : * set presult->all_visible and presult->all_frozen on exit, to indicate if
428 : * the VM bits can be set. They are always set to false when the
429 : * HEAP_PRUNE_FREEZE option is not set, because at the moment only callers
430 : * that also freeze need that information.
431 : *
432 : * vistest is used to distinguish whether tuples are DEAD or RECENTLY_DEAD
433 : * (see heap_prune_satisfies_vacuum).
434 : *
435 : * options:
436 : * MARK_UNUSED_NOW indicates that dead items can be set LP_UNUSED during
437 : * pruning.
438 : *
439 : * FREEZE indicates that we will also freeze tuples, and will return
440 : * 'all_visible', 'all_frozen' flags to the caller.
441 : *
442 : * cutoffs contains the freeze cutoffs, established by VACUUM at the beginning
443 : * of vacuuming the relation. Required if HEAP_PRUNE_FREEZE option is set.
444 : * cutoffs->OldestXmin is also used to determine if dead tuples are
445 : * HEAPTUPLE_RECENTLY_DEAD or HEAPTUPLE_DEAD.
446 : *
447 : * presult contains output parameters needed by callers, such as the number of
448 : * tuples removed and the offsets of dead items on the page after pruning.
449 : * heap_page_prune_and_freeze() is responsible for initializing it. Required
450 : * by all callers.
451 : *
452 : * reason indicates why the pruning is performed. It is included in the WAL
453 : * record for debugging and analysis purposes, but otherwise has no effect.
454 : *
455 : * off_loc is the offset location required by the caller to use in error
456 : * callback.
457 : *
458 : * new_relfrozen_xid and new_relmin_mxid must provided by the caller if the
459 : * HEAP_PRUNE_FREEZE option is set. On entry, they contain the oldest XID and
460 : * multi-XID seen on the relation so far. They will be updated with oldest
461 : * values present on the page after pruning. After processing the whole
462 : * relation, VACUUM can use these values as the new relfrozenxid/relminmxid
463 : * for the relation.
464 : */
465 : void
466 1197090 : heap_page_prune_and_freeze(Relation relation, Buffer buffer,
467 : GlobalVisState *vistest,
468 : int options,
469 : struct VacuumCutoffs *cutoffs,
470 : PruneFreezeResult *presult,
471 : PruneReason reason,
472 : OffsetNumber *off_loc,
473 : TransactionId *new_relfrozen_xid,
474 : MultiXactId *new_relmin_mxid)
475 : {
476 1197090 : Page page = BufferGetPage(buffer);
477 1197090 : BlockNumber blockno = BufferGetBlockNumber(buffer);
478 : OffsetNumber offnum,
479 : maxoff;
480 : PruneState prstate;
481 : HeapTupleData tup;
482 : bool do_freeze;
483 : bool do_prune;
484 : bool do_hint_prune;
485 : bool did_tuple_hint_fpi;
486 1197090 : int64 fpi_before = pgWalUsage.wal_fpi;
487 :
488 : /* Copy parameters to prstate */
489 1197090 : prstate.vistest = vistest;
490 1197090 : prstate.mark_unused_now = (options & HEAP_PAGE_PRUNE_MARK_UNUSED_NOW) != 0;
491 1197090 : prstate.attempt_freeze = (options & HEAP_PAGE_PRUNE_FREEZE) != 0;
492 1197090 : prstate.cutoffs = cutoffs;
493 :
494 : /*
495 : * Our strategy is to scan the page and make lists of items to change,
496 : * then apply the changes within a critical section. This keeps as much
497 : * logic as possible out of the critical section, and also ensures that
498 : * WAL replay will work the same as the normal case.
499 : *
500 : * First, initialize the new pd_prune_xid value to zero (indicating no
501 : * prunable tuples). If we find any tuples which may soon become
502 : * prunable, we will save the lowest relevant XID in new_prune_xid. Also
503 : * initialize the rest of our working state.
504 : */
505 1197090 : prstate.new_prune_xid = InvalidTransactionId;
506 1197090 : prstate.latest_xid_removed = InvalidTransactionId;
507 1197090 : prstate.nredirected = prstate.ndead = prstate.nunused = prstate.nfrozen = 0;
508 1197090 : prstate.nroot_items = 0;
509 1197090 : prstate.nheaponly_items = 0;
510 :
511 : /* initialize page freezing working state */
512 1197090 : prstate.pagefrz.freeze_required = false;
513 1197090 : if (prstate.attempt_freeze)
514 : {
515 : Assert(new_relfrozen_xid && new_relmin_mxid);
516 1116396 : prstate.pagefrz.FreezePageRelfrozenXid = *new_relfrozen_xid;
517 1116396 : prstate.pagefrz.NoFreezePageRelfrozenXid = *new_relfrozen_xid;
518 1116396 : prstate.pagefrz.FreezePageRelminMxid = *new_relmin_mxid;
519 1116396 : prstate.pagefrz.NoFreezePageRelminMxid = *new_relmin_mxid;
520 : }
521 : else
522 : {
523 : Assert(new_relfrozen_xid == NULL && new_relmin_mxid == NULL);
524 80694 : prstate.pagefrz.FreezePageRelminMxid = InvalidMultiXactId;
525 80694 : prstate.pagefrz.NoFreezePageRelminMxid = InvalidMultiXactId;
526 80694 : prstate.pagefrz.FreezePageRelfrozenXid = InvalidTransactionId;
527 80694 : prstate.pagefrz.NoFreezePageRelfrozenXid = InvalidTransactionId;
528 : }
529 :
530 1197090 : prstate.ndeleted = 0;
531 1197090 : prstate.live_tuples = 0;
532 1197090 : prstate.recently_dead_tuples = 0;
533 1197090 : prstate.hastup = false;
534 1197090 : prstate.lpdead_items = 0;
535 1197090 : prstate.deadoffsets = presult->deadoffsets;
536 :
537 : /*
538 : * Caller may update the VM after we're done. We can keep track of
539 : * whether the page will be all-visible and all-frozen after pruning and
540 : * freezing to help the caller to do that.
541 : *
542 : * Currently, only VACUUM sets the VM bits. To save the effort, only do
543 : * the bookkeeping if the caller needs it. Currently, that's tied to
544 : * HEAP_PAGE_PRUNE_FREEZE, but it could be a separate flag if you wanted
545 : * to update the VM bits without also freezing or freeze without also
546 : * setting the VM bits.
547 : *
548 : * In addition to telling the caller whether it can set the VM bit, we
549 : * also use 'all_visible' and 'all_frozen' for our own decision-making. If
550 : * the whole page would become frozen, we consider opportunistically
551 : * freezing tuples. We will not be able to freeze the whole page if there
552 : * are tuples present that are not visible to everyone or if there are
553 : * dead tuples which are not yet removable. However, dead tuples which
554 : * will be removed by the end of vacuuming should not preclude us from
555 : * opportunistically freezing. Because of that, we do not clear
556 : * all_visible when we see LP_DEAD items. We fix that at the end of the
557 : * function, when we return the value to the caller, so that the caller
558 : * doesn't set the VM bit incorrectly.
559 : */
560 1197090 : if (prstate.attempt_freeze)
561 : {
562 1116396 : prstate.all_visible = true;
563 1116396 : prstate.all_frozen = true;
564 : }
565 : else
566 : {
567 : /*
568 : * Initializing to false allows skipping the work to update them in
569 : * heap_prune_record_unchanged_lp_normal().
570 : */
571 80694 : prstate.all_visible = false;
572 80694 : prstate.all_frozen = false;
573 : }
574 :
575 : /*
576 : * The visibility cutoff xid is the newest xmin of live tuples on the
577 : * page. In the common case, this will be set as the conflict horizon the
578 : * caller can use for updating the VM. If, at the end of freezing and
579 : * pruning, the page is all-frozen, there is no possibility that any
580 : * running transaction on the standby does not see tuples on the page as
581 : * all-visible, so the conflict horizon remains InvalidTransactionId.
582 : */
583 1197090 : prstate.visibility_cutoff_xid = InvalidTransactionId;
584 :
585 1197090 : maxoff = PageGetMaxOffsetNumber(page);
586 1197090 : tup.t_tableOid = RelationGetRelid(relation);
587 :
588 : /*
589 : * Determine HTSV for all tuples, and queue them up for processing as HOT
590 : * chain roots or as heap-only items.
591 : *
592 : * Determining HTSV only once for each tuple is required for correctness,
593 : * to deal with cases where running HTSV twice could result in different
594 : * results. For example, RECENTLY_DEAD can turn to DEAD if another
595 : * checked item causes GlobalVisTestIsRemovableFullXid() to update the
596 : * horizon, or INSERT_IN_PROGRESS can change to DEAD if the inserting
597 : * transaction aborts.
598 : *
599 : * It's also good for performance. Most commonly tuples within a page are
600 : * stored at decreasing offsets (while the items are stored at increasing
601 : * offsets). When processing all tuples on a page this leads to reading
602 : * memory at decreasing offsets within a page, with a variable stride.
603 : * That's hard for CPU prefetchers to deal with. Processing the items in
604 : * reverse order (and thus the tuples in increasing order) increases
605 : * prefetching efficiency significantly / decreases the number of cache
606 : * misses.
607 : */
608 1197090 : for (offnum = maxoff;
609 59068268 : offnum >= FirstOffsetNumber;
610 57871178 : offnum = OffsetNumberPrev(offnum))
611 : {
612 57871178 : ItemId itemid = PageGetItemId(page, offnum);
613 : HeapTupleHeader htup;
614 :
615 : /*
616 : * Set the offset number so that we can display it along with any
617 : * error that occurred while processing this tuple.
618 : */
619 57871178 : *off_loc = offnum;
620 :
621 57871178 : prstate.processed[offnum] = false;
622 57871178 : prstate.htsv[offnum] = -1;
623 :
624 : /* Nothing to do if slot doesn't contain a tuple */
625 57871178 : if (!ItemIdIsUsed(itemid))
626 : {
627 618730 : heap_prune_record_unchanged_lp_unused(page, &prstate, offnum);
628 618730 : continue;
629 : }
630 :
631 57252448 : if (ItemIdIsDead(itemid))
632 : {
633 : /*
634 : * If the caller set mark_unused_now true, we can set dead line
635 : * pointers LP_UNUSED now.
636 : */
637 1931388 : if (unlikely(prstate.mark_unused_now))
638 1476 : heap_prune_record_unused(&prstate, offnum, false);
639 : else
640 1929912 : heap_prune_record_unchanged_lp_dead(page, &prstate, offnum);
641 1931388 : continue;
642 : }
643 :
644 55321060 : if (ItemIdIsRedirected(itemid))
645 : {
646 : /* This is the start of a HOT chain */
647 647284 : prstate.root_items[prstate.nroot_items++] = offnum;
648 647284 : continue;
649 : }
650 :
651 : Assert(ItemIdIsNormal(itemid));
652 :
653 : /*
654 : * Get the tuple's visibility status and queue it up for processing.
655 : */
656 54673776 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
657 54673776 : tup.t_data = htup;
658 54673776 : tup.t_len = ItemIdGetLength(itemid);
659 54673776 : ItemPointerSet(&tup.t_self, blockno, offnum);
660 :
661 54673776 : prstate.htsv[offnum] = heap_prune_satisfies_vacuum(&prstate, &tup,
662 : buffer);
663 :
664 54673776 : if (!HeapTupleHeaderIsHeapOnly(htup))
665 53790140 : prstate.root_items[prstate.nroot_items++] = offnum;
666 : else
667 883636 : prstate.heaponly_items[prstate.nheaponly_items++] = offnum;
668 : }
669 :
670 : /*
671 : * If checksums are enabled, heap_prune_satisfies_vacuum() may have caused
672 : * an FPI to be emitted.
673 : */
674 1197090 : did_tuple_hint_fpi = fpi_before != pgWalUsage.wal_fpi;
675 :
676 : /*
677 : * Process HOT chains.
678 : *
679 : * We added the items to the array starting from 'maxoff', so by
680 : * processing the array in reverse order, we process the items in
681 : * ascending offset number order. The order doesn't matter for
682 : * correctness, but some quick micro-benchmarking suggests that this is
683 : * faster. (Earlier PostgreSQL versions, which scanned all the items on
684 : * the page instead of using the root_items array, also did it in
685 : * ascending offset number order.)
686 : */
687 55634514 : for (int i = prstate.nroot_items - 1; i >= 0; i--)
688 : {
689 54437424 : offnum = prstate.root_items[i];
690 :
691 : /* Ignore items already processed as part of an earlier chain */
692 54437424 : if (prstate.processed[offnum])
693 0 : continue;
694 :
695 : /* see preceding loop */
696 54437424 : *off_loc = offnum;
697 :
698 : /* Process this item or chain of items */
699 54437424 : heap_prune_chain(page, blockno, maxoff, offnum, &prstate);
700 : }
701 :
702 : /*
703 : * Process any heap-only tuples that were not already processed as part of
704 : * a HOT chain.
705 : */
706 2080726 : for (int i = prstate.nheaponly_items - 1; i >= 0; i--)
707 : {
708 883636 : offnum = prstate.heaponly_items[i];
709 :
710 883636 : if (prstate.processed[offnum])
711 857098 : continue;
712 :
713 : /* see preceding loop */
714 26538 : *off_loc = offnum;
715 :
716 : /*
717 : * If the tuple is DEAD and doesn't chain to anything else, mark it
718 : * unused. (If it does chain, we can only remove it as part of
719 : * pruning its chain.)
720 : *
721 : * We need this primarily to handle aborted HOT updates, that is,
722 : * XMIN_INVALID heap-only tuples. Those might not be linked to by any
723 : * chain, since the parent tuple might be re-updated before any
724 : * pruning occurs. So we have to be able to reap them separately from
725 : * chain-pruning. (Note that HeapTupleHeaderIsHotUpdated will never
726 : * return true for an XMIN_INVALID tuple, so this code will work even
727 : * when there were sequential updates within the aborted transaction.)
728 : */
729 26538 : if (prstate.htsv[offnum] == HEAPTUPLE_DEAD)
730 : {
731 4078 : ItemId itemid = PageGetItemId(page, offnum);
732 4078 : HeapTupleHeader htup = (HeapTupleHeader) PageGetItem(page, itemid);
733 :
734 4078 : if (likely(!HeapTupleHeaderIsHotUpdated(htup)))
735 : {
736 4078 : HeapTupleHeaderAdvanceConflictHorizon(htup,
737 : &prstate.latest_xid_removed);
738 4078 : heap_prune_record_unused(&prstate, offnum, true);
739 : }
740 : else
741 : {
742 : /*
743 : * This tuple should've been processed and removed as part of
744 : * a HOT chain, so something's wrong. To preserve evidence,
745 : * we don't dare to remove it. We cannot leave behind a DEAD
746 : * tuple either, because that will cause VACUUM to error out.
747 : * Throwing an error with a distinct error message seems like
748 : * the least bad option.
749 : */
750 0 : elog(ERROR, "dead heap-only tuple (%u, %d) is not linked to from any HOT chain",
751 : blockno, offnum);
752 : }
753 : }
754 : else
755 22460 : heap_prune_record_unchanged_lp_normal(page, &prstate, offnum);
756 : }
757 :
758 : /* We should now have processed every tuple exactly once */
759 : #ifdef USE_ASSERT_CHECKING
760 : for (offnum = FirstOffsetNumber;
761 : offnum <= maxoff;
762 : offnum = OffsetNumberNext(offnum))
763 : {
764 : *off_loc = offnum;
765 :
766 : Assert(prstate.processed[offnum]);
767 : }
768 : #endif
769 :
770 : /* Clear the offset information once we have processed the given page. */
771 1197090 : *off_loc = InvalidOffsetNumber;
772 :
773 3558074 : do_prune = prstate.nredirected > 0 ||
774 2291106 : prstate.ndead > 0 ||
775 1094016 : prstate.nunused > 0;
776 :
777 : /*
778 : * Even if we don't prune anything, if we found a new value for the
779 : * pd_prune_xid field or the page was marked full, we will update the hint
780 : * bit.
781 : */
782 2290488 : do_hint_prune = ((PageHeader) page)->pd_prune_xid != prstate.new_prune_xid ||
783 1093398 : PageIsFull(page);
784 :
785 : /*
786 : * Decide if we want to go ahead with freezing according to the freeze
787 : * plans we prepared, or not.
788 : */
789 1197090 : do_freeze = heap_page_will_freeze(relation, buffer,
790 : did_tuple_hint_fpi,
791 : do_prune,
792 : do_hint_prune,
793 : &prstate);
794 :
795 : /* Any error while applying the changes is critical */
796 1197090 : START_CRIT_SECTION();
797 :
798 1197090 : if (do_hint_prune)
799 : {
800 : /*
801 : * Update the page's pd_prune_xid field to either zero, or the lowest
802 : * XID of any soon-prunable tuple.
803 : */
804 103878 : ((PageHeader) page)->pd_prune_xid = prstate.new_prune_xid;
805 :
806 : /*
807 : * Also clear the "page is full" flag, since there's no point in
808 : * repeating the prune/defrag process until something else happens to
809 : * the page.
810 : */
811 103878 : PageClearFull(page);
812 :
813 : /*
814 : * If that's all we had to do to the page, this is a non-WAL-logged
815 : * hint. If we are going to freeze or prune the page, we will mark
816 : * the buffer dirty below.
817 : */
818 103878 : if (!do_freeze && !do_prune)
819 402 : MarkBufferDirtyHint(buffer, true);
820 : }
821 :
822 1197090 : if (do_prune || do_freeze)
823 : {
824 : /* Apply the planned item changes and repair page fragmentation. */
825 146096 : if (do_prune)
826 : {
827 103852 : heap_page_prune_execute(buffer, false,
828 : prstate.redirected, prstate.nredirected,
829 : prstate.nowdead, prstate.ndead,
830 : prstate.nowunused, prstate.nunused);
831 : }
832 :
833 146096 : if (do_freeze)
834 46376 : heap_freeze_prepared_tuples(buffer, prstate.frozen, prstate.nfrozen);
835 :
836 146096 : MarkBufferDirty(buffer);
837 :
838 : /*
839 : * Emit a WAL XLOG_HEAP2_PRUNE* record showing what we did
840 : */
841 146096 : if (RelationNeedsWAL(relation))
842 : {
843 : /*
844 : * The snapshotConflictHorizon for the whole record should be the
845 : * most conservative of all the horizons calculated for any of the
846 : * possible modifications. If this record will prune tuples, any
847 : * transactions on the standby older than the youngest xmax of the
848 : * most recently removed tuple this record will prune will
849 : * conflict. If this record will freeze tuples, any transactions
850 : * on the standby with xids older than the youngest tuple this
851 : * record will freeze will conflict.
852 : */
853 144342 : TransactionId frz_conflict_horizon = InvalidTransactionId;
854 : TransactionId conflict_xid;
855 :
856 : /*
857 : * We can use the visibility_cutoff_xid as our cutoff for
858 : * conflicts when the whole page is eligible to become all-frozen
859 : * in the VM once we're done with it. Otherwise we generate a
860 : * conservative cutoff by stepping back from OldestXmin.
861 : */
862 144342 : if (do_freeze)
863 : {
864 46372 : if (prstate.all_visible && prstate.all_frozen)
865 41832 : frz_conflict_horizon = prstate.visibility_cutoff_xid;
866 : else
867 : {
868 : /* Avoids false conflicts when hot_standby_feedback in use */
869 4540 : frz_conflict_horizon = prstate.cutoffs->OldestXmin;
870 4540 : TransactionIdRetreat(frz_conflict_horizon);
871 : }
872 : }
873 :
874 144342 : if (TransactionIdFollows(frz_conflict_horizon, prstate.latest_xid_removed))
875 42922 : conflict_xid = frz_conflict_horizon;
876 : else
877 101420 : conflict_xid = prstate.latest_xid_removed;
878 :
879 144342 : log_heap_prune_and_freeze(relation, buffer,
880 : InvalidBuffer, /* vmbuffer */
881 : 0, /* vmflags */
882 : conflict_xid,
883 : true, reason,
884 : prstate.frozen, prstate.nfrozen,
885 : prstate.redirected, prstate.nredirected,
886 : prstate.nowdead, prstate.ndead,
887 : prstate.nowunused, prstate.nunused);
888 : }
889 : }
890 :
891 1197090 : END_CRIT_SECTION();
892 :
893 : /* Copy information back for caller */
894 1197090 : presult->ndeleted = prstate.ndeleted;
895 1197090 : presult->nnewlpdead = prstate.ndead;
896 1197090 : presult->nfrozen = prstate.nfrozen;
897 1197090 : presult->live_tuples = prstate.live_tuples;
898 1197090 : presult->recently_dead_tuples = prstate.recently_dead_tuples;
899 :
900 : /*
901 : * It was convenient to ignore LP_DEAD items in all_visible earlier on to
902 : * make the choice of whether or not to freeze the page unaffected by the
903 : * short-term presence of LP_DEAD items. These LP_DEAD items were
904 : * effectively assumed to be LP_UNUSED items in the making. It doesn't
905 : * matter which vacuum heap pass (initial pass or final pass) ends up
906 : * setting the page all-frozen, as long as the ongoing VACUUM does it.
907 : *
908 : * Now that freezing has been finalized, unset all_visible if there are
909 : * any LP_DEAD items on the page. It needs to reflect the present state
910 : * of the page, as expected by our caller.
911 : */
912 1197090 : if (prstate.all_visible && prstate.lpdead_items == 0)
913 : {
914 466826 : presult->all_visible = prstate.all_visible;
915 466826 : presult->all_frozen = prstate.all_frozen;
916 : }
917 : else
918 : {
919 730264 : presult->all_visible = false;
920 730264 : presult->all_frozen = false;
921 : }
922 :
923 1197090 : presult->hastup = prstate.hastup;
924 :
925 : /*
926 : * For callers planning to update the visibility map, the conflict horizon
927 : * for that record must be the newest xmin on the page. However, if the
928 : * page is completely frozen, there can be no conflict and the
929 : * vm_conflict_horizon should remain InvalidTransactionId. This includes
930 : * the case that we just froze all the tuples; the prune-freeze record
931 : * included the conflict XID already so the caller doesn't need it.
932 : */
933 1197090 : if (presult->all_frozen)
934 447002 : presult->vm_conflict_horizon = InvalidTransactionId;
935 : else
936 750088 : presult->vm_conflict_horizon = prstate.visibility_cutoff_xid;
937 :
938 1197090 : presult->lpdead_items = prstate.lpdead_items;
939 : /* the presult->deadoffsets array was already filled in */
940 :
941 1197090 : if (prstate.attempt_freeze)
942 : {
943 1116396 : if (presult->nfrozen > 0)
944 : {
945 46376 : *new_relfrozen_xid = prstate.pagefrz.FreezePageRelfrozenXid;
946 46376 : *new_relmin_mxid = prstate.pagefrz.FreezePageRelminMxid;
947 : }
948 : else
949 : {
950 1070020 : *new_relfrozen_xid = prstate.pagefrz.NoFreezePageRelfrozenXid;
951 1070020 : *new_relmin_mxid = prstate.pagefrz.NoFreezePageRelminMxid;
952 : }
953 : }
954 1197090 : }
955 :
956 :
957 : /*
958 : * Perform visibility checks for heap pruning.
959 : */
960 : static HTSV_Result
961 54673776 : heap_prune_satisfies_vacuum(PruneState *prstate, HeapTuple tup, Buffer buffer)
962 : {
963 : HTSV_Result res;
964 : TransactionId dead_after;
965 :
966 54673776 : res = HeapTupleSatisfiesVacuumHorizon(tup, buffer, &dead_after);
967 :
968 54673776 : if (res != HEAPTUPLE_RECENTLY_DEAD)
969 51426912 : return res;
970 :
971 : /*
972 : * For VACUUM, we must be sure to prune tuples with xmax older than
973 : * OldestXmin -- a visibility cutoff determined at the beginning of
974 : * vacuuming the relation. OldestXmin is used for freezing determination
975 : * and we cannot freeze dead tuples' xmaxes.
976 : */
977 3246864 : if (prstate->cutoffs &&
978 1852486 : TransactionIdIsValid(prstate->cutoffs->OldestXmin) &&
979 1852486 : NormalTransactionIdPrecedes(dead_after, prstate->cutoffs->OldestXmin))
980 1405306 : return HEAPTUPLE_DEAD;
981 :
982 : /*
983 : * Determine whether or not the tuple is considered dead when compared
984 : * with the provided GlobalVisState. On-access pruning does not provide
985 : * VacuumCutoffs. And for vacuum, even if the tuple's xmax is not older
986 : * than OldestXmin, GlobalVisTestIsRemovableXid() could find the row dead
987 : * if the GlobalVisState has been updated since the beginning of vacuuming
988 : * the relation.
989 : */
990 1841558 : if (GlobalVisTestIsRemovableXid(prstate->vistest, dead_after))
991 1333752 : return HEAPTUPLE_DEAD;
992 :
993 507806 : return res;
994 : }
995 :
996 :
997 : /*
998 : * Pruning calculates tuple visibility once and saves the results in an array
999 : * of int8. See PruneState.htsv for details. This helper function is meant
1000 : * to guard against examining visibility status array members which have not
1001 : * yet been computed.
1002 : */
1003 : static inline HTSV_Result
1004 54647238 : htsv_get_valid_status(int status)
1005 : {
1006 : Assert(status >= HEAPTUPLE_DEAD &&
1007 : status <= HEAPTUPLE_DELETE_IN_PROGRESS);
1008 54647238 : return (HTSV_Result) status;
1009 : }
1010 :
1011 : /*
1012 : * Prune specified line pointer or a HOT chain originating at line pointer.
1013 : *
1014 : * Tuple visibility information is provided in prstate->htsv.
1015 : *
1016 : * If the item is an index-referenced tuple (i.e. not a heap-only tuple),
1017 : * the HOT chain is pruned by removing all DEAD tuples at the start of the HOT
1018 : * chain. We also prune any RECENTLY_DEAD tuples preceding a DEAD tuple.
1019 : * This is OK because a RECENTLY_DEAD tuple preceding a DEAD tuple is really
1020 : * DEAD, our visibility test is just too coarse to detect it.
1021 : *
1022 : * Pruning must never leave behind a DEAD tuple that still has tuple storage.
1023 : * VACUUM isn't prepared to deal with that case.
1024 : *
1025 : * The root line pointer is redirected to the tuple immediately after the
1026 : * latest DEAD tuple. If all tuples in the chain are DEAD, the root line
1027 : * pointer is marked LP_DEAD. (This includes the case of a DEAD simple
1028 : * tuple, which we treat as a chain of length 1.)
1029 : *
1030 : * We don't actually change the page here. We just add entries to the arrays in
1031 : * prstate showing the changes to be made. Items to be redirected are added
1032 : * to the redirected[] array (two entries per redirection); items to be set to
1033 : * LP_DEAD state are added to nowdead[]; and items to be set to LP_UNUSED
1034 : * state are added to nowunused[]. We perform bookkeeping of live tuples,
1035 : * visibility etc. based on what the page will look like after the changes
1036 : * applied. All that bookkeeping is performed in the heap_prune_record_*()
1037 : * subroutines. The division of labor is that heap_prune_chain() decides the
1038 : * fate of each tuple, ie. whether it's going to be removed, redirected or
1039 : * left unchanged, and the heap_prune_record_*() subroutines update PruneState
1040 : * based on that outcome.
1041 : */
1042 : static void
1043 54437424 : heap_prune_chain(Page page, BlockNumber blockno, OffsetNumber maxoff,
1044 : OffsetNumber rootoffnum, PruneState *prstate)
1045 : {
1046 54437424 : TransactionId priorXmax = InvalidTransactionId;
1047 : ItemId rootlp;
1048 : OffsetNumber offnum;
1049 : OffsetNumber chainitems[MaxHeapTuplesPerPage];
1050 :
1051 : /*
1052 : * After traversing the HOT chain, ndeadchain is the index in chainitems
1053 : * of the first live successor after the last dead item.
1054 : */
1055 54437424 : int ndeadchain = 0,
1056 54437424 : nchain = 0;
1057 :
1058 54437424 : rootlp = PageGetItemId(page, rootoffnum);
1059 :
1060 : /* Start from the root tuple */
1061 54437424 : offnum = rootoffnum;
1062 :
1063 : /* while not end of the chain */
1064 : for (;;)
1065 857098 : {
1066 : HeapTupleHeader htup;
1067 : ItemId lp;
1068 :
1069 : /* Sanity check (pure paranoia) */
1070 55294522 : if (offnum < FirstOffsetNumber)
1071 0 : break;
1072 :
1073 : /*
1074 : * An offset past the end of page's line pointer array is possible
1075 : * when the array was truncated (original item must have been unused)
1076 : */
1077 55294522 : if (offnum > maxoff)
1078 0 : break;
1079 :
1080 : /* If item is already processed, stop --- it must not be same chain */
1081 55294522 : if (prstate->processed[offnum])
1082 0 : break;
1083 :
1084 55294522 : lp = PageGetItemId(page, offnum);
1085 :
1086 : /*
1087 : * Unused item obviously isn't part of the chain. Likewise, a dead
1088 : * line pointer can't be part of the chain. Both of those cases were
1089 : * already marked as processed.
1090 : */
1091 : Assert(ItemIdIsUsed(lp));
1092 : Assert(!ItemIdIsDead(lp));
1093 :
1094 : /*
1095 : * If we are looking at the redirected root line pointer, jump to the
1096 : * first normal tuple in the chain. If we find a redirect somewhere
1097 : * else, stop --- it must not be same chain.
1098 : */
1099 55294522 : if (ItemIdIsRedirected(lp))
1100 : {
1101 647284 : if (nchain > 0)
1102 0 : break; /* not at start of chain */
1103 647284 : chainitems[nchain++] = offnum;
1104 647284 : offnum = ItemIdGetRedirect(rootlp);
1105 647284 : continue;
1106 : }
1107 :
1108 : Assert(ItemIdIsNormal(lp));
1109 :
1110 54647238 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1111 :
1112 : /*
1113 : * Check the tuple XMIN against prior XMAX, if any
1114 : */
1115 54857052 : if (TransactionIdIsValid(priorXmax) &&
1116 209814 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
1117 0 : break;
1118 :
1119 : /*
1120 : * OK, this tuple is indeed a member of the chain.
1121 : */
1122 54647238 : chainitems[nchain++] = offnum;
1123 :
1124 54647238 : switch (htsv_get_valid_status(prstate->htsv[offnum]))
1125 : {
1126 2821436 : case HEAPTUPLE_DEAD:
1127 :
1128 : /* Remember the last DEAD tuple seen */
1129 2821436 : ndeadchain = nchain;
1130 2821436 : HeapTupleHeaderAdvanceConflictHorizon(htup,
1131 : &prstate->latest_xid_removed);
1132 : /* Advance to next chain member */
1133 2821436 : break;
1134 :
1135 507806 : case HEAPTUPLE_RECENTLY_DEAD:
1136 :
1137 : /*
1138 : * We don't need to advance the conflict horizon for
1139 : * RECENTLY_DEAD tuples, even if we are removing them. This
1140 : * is because we only remove RECENTLY_DEAD tuples if they
1141 : * precede a DEAD tuple, and the DEAD tuple must have been
1142 : * inserted by a newer transaction than the RECENTLY_DEAD
1143 : * tuple by virtue of being later in the chain. We will have
1144 : * advanced the conflict horizon for the DEAD tuple.
1145 : */
1146 :
1147 : /*
1148 : * Advance past RECENTLY_DEAD tuples just in case there's a
1149 : * DEAD one after them. We have to make sure that we don't
1150 : * miss any DEAD tuples, since DEAD tuples that still have
1151 : * tuple storage after pruning will confuse VACUUM.
1152 : */
1153 507806 : break;
1154 :
1155 51317996 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1156 : case HEAPTUPLE_LIVE:
1157 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1158 51317996 : goto process_chain;
1159 :
1160 0 : default:
1161 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1162 : goto process_chain;
1163 : }
1164 :
1165 : /*
1166 : * If the tuple is not HOT-updated, then we are at the end of this
1167 : * HOT-update chain.
1168 : */
1169 3329242 : if (!HeapTupleHeaderIsHotUpdated(htup))
1170 3119428 : goto process_chain;
1171 :
1172 : /* HOT implies it can't have moved to different partition */
1173 : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1174 :
1175 : /*
1176 : * Advance to next chain member.
1177 : */
1178 : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == blockno);
1179 209814 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1180 209814 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1181 : }
1182 :
1183 0 : if (ItemIdIsRedirected(rootlp) && nchain < 2)
1184 : {
1185 : /*
1186 : * We found a redirect item that doesn't point to a valid follow-on
1187 : * item. This can happen if the loop in heap_page_prune_and_freeze()
1188 : * caused us to visit the dead successor of a redirect item before
1189 : * visiting the redirect item. We can clean up by setting the
1190 : * redirect item to LP_DEAD state or LP_UNUSED if the caller
1191 : * indicated.
1192 : */
1193 0 : heap_prune_record_dead_or_unused(prstate, rootoffnum, false);
1194 0 : return;
1195 : }
1196 :
1197 0 : process_chain:
1198 :
1199 54437424 : if (ndeadchain == 0)
1200 : {
1201 : /*
1202 : * No DEAD tuple was found, so the chain is entirely composed of
1203 : * normal, unchanged tuples. Leave it alone.
1204 : */
1205 51684132 : int i = 0;
1206 :
1207 51684132 : if (ItemIdIsRedirected(rootlp))
1208 : {
1209 613892 : heap_prune_record_unchanged_lp_redirect(prstate, rootoffnum);
1210 613892 : i++;
1211 : }
1212 103376362 : for (; i < nchain; i++)
1213 51692230 : heap_prune_record_unchanged_lp_normal(page, prstate, chainitems[i]);
1214 : }
1215 2753292 : else if (ndeadchain == nchain)
1216 : {
1217 : /*
1218 : * The entire chain is dead. Mark the root line pointer LP_DEAD, and
1219 : * fully remove the other tuples in the chain.
1220 : */
1221 2623846 : heap_prune_record_dead_or_unused(prstate, rootoffnum, ItemIdIsNormal(rootlp));
1222 2689284 : for (int i = 1; i < nchain; i++)
1223 65438 : heap_prune_record_unused(prstate, chainitems[i], true);
1224 : }
1225 : else
1226 : {
1227 : /*
1228 : * We found a DEAD tuple in the chain. Redirect the root line pointer
1229 : * to the first non-DEAD tuple, and mark as unused each intermediate
1230 : * item that we are able to remove from the chain.
1231 : */
1232 129446 : heap_prune_record_redirect(prstate, rootoffnum, chainitems[ndeadchain],
1233 129446 : ItemIdIsNormal(rootlp));
1234 165544 : for (int i = 1; i < ndeadchain; i++)
1235 36098 : heap_prune_record_unused(prstate, chainitems[i], true);
1236 :
1237 : /* the rest of tuples in the chain are normal, unchanged tuples */
1238 263018 : for (int i = ndeadchain; i < nchain; i++)
1239 133572 : heap_prune_record_unchanged_lp_normal(page, prstate, chainitems[i]);
1240 : }
1241 : }
1242 :
1243 : /* Record lowest soon-prunable XID */
1244 : static void
1245 15679162 : heap_prune_record_prunable(PruneState *prstate, TransactionId xid)
1246 : {
1247 : /*
1248 : * This should exactly match the PageSetPrunable macro. We can't store
1249 : * directly into the page header yet, so we update working state.
1250 : */
1251 : Assert(TransactionIdIsNormal(xid));
1252 30737548 : if (!TransactionIdIsValid(prstate->new_prune_xid) ||
1253 15058386 : TransactionIdPrecedes(xid, prstate->new_prune_xid))
1254 622782 : prstate->new_prune_xid = xid;
1255 15679162 : }
1256 :
1257 : /* Record line pointer to be redirected */
1258 : static void
1259 129446 : heap_prune_record_redirect(PruneState *prstate,
1260 : OffsetNumber offnum, OffsetNumber rdoffnum,
1261 : bool was_normal)
1262 : {
1263 : Assert(!prstate->processed[offnum]);
1264 129446 : prstate->processed[offnum] = true;
1265 :
1266 : /*
1267 : * Do not mark the redirect target here. It needs to be counted
1268 : * separately as an unchanged tuple.
1269 : */
1270 :
1271 : Assert(prstate->nredirected < MaxHeapTuplesPerPage);
1272 129446 : prstate->redirected[prstate->nredirected * 2] = offnum;
1273 129446 : prstate->redirected[prstate->nredirected * 2 + 1] = rdoffnum;
1274 :
1275 129446 : prstate->nredirected++;
1276 :
1277 : /*
1278 : * If the root entry had been a normal tuple, we are deleting it, so count
1279 : * it in the result. But changing a redirect (even to DEAD state) doesn't
1280 : * count.
1281 : */
1282 129446 : if (was_normal)
1283 114498 : prstate->ndeleted++;
1284 :
1285 129446 : prstate->hastup = true;
1286 129446 : }
1287 :
1288 : /* Record line pointer to be marked dead */
1289 : static void
1290 2555194 : heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
1291 : bool was_normal)
1292 : {
1293 : Assert(!prstate->processed[offnum]);
1294 2555194 : prstate->processed[offnum] = true;
1295 :
1296 : Assert(prstate->ndead < MaxHeapTuplesPerPage);
1297 2555194 : prstate->nowdead[prstate->ndead] = offnum;
1298 2555194 : prstate->ndead++;
1299 :
1300 : /*
1301 : * Deliberately delay unsetting all_visible until later during pruning.
1302 : * Removable dead tuples shouldn't preclude freezing the page.
1303 : */
1304 :
1305 : /* Record the dead offset for vacuum */
1306 2555194 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1307 :
1308 : /*
1309 : * If the root entry had been a normal tuple, we are deleting it, so count
1310 : * it in the result. But changing a redirect (even to DEAD state) doesn't
1311 : * count.
1312 : */
1313 2555194 : if (was_normal)
1314 2536750 : prstate->ndeleted++;
1315 2555194 : }
1316 :
1317 : /*
1318 : * Depending on whether or not the caller set mark_unused_now to true, record that a
1319 : * line pointer should be marked LP_DEAD or LP_UNUSED. There are other cases in
1320 : * which we will mark line pointers LP_UNUSED, but we will not mark line
1321 : * pointers LP_DEAD if mark_unused_now is true.
1322 : */
1323 : static void
1324 2623846 : heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
1325 : bool was_normal)
1326 : {
1327 : /*
1328 : * If the caller set mark_unused_now to true, we can remove dead tuples
1329 : * during pruning instead of marking their line pointers dead. Set this
1330 : * tuple's line pointer LP_UNUSED. We hint that this option is less
1331 : * likely.
1332 : */
1333 2623846 : if (unlikely(prstate->mark_unused_now))
1334 68652 : heap_prune_record_unused(prstate, offnum, was_normal);
1335 : else
1336 2555194 : heap_prune_record_dead(prstate, offnum, was_normal);
1337 2623846 : }
1338 :
1339 : /* Record line pointer to be marked unused */
1340 : static void
1341 175742 : heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal)
1342 : {
1343 : Assert(!prstate->processed[offnum]);
1344 175742 : prstate->processed[offnum] = true;
1345 :
1346 : Assert(prstate->nunused < MaxHeapTuplesPerPage);
1347 175742 : prstate->nowunused[prstate->nunused] = offnum;
1348 175742 : prstate->nunused++;
1349 :
1350 : /*
1351 : * If the root entry had been a normal tuple, we are deleting it, so count
1352 : * it in the result. But changing a redirect (even to DEAD state) doesn't
1353 : * count.
1354 : */
1355 175742 : if (was_normal)
1356 174266 : prstate->ndeleted++;
1357 175742 : }
1358 :
1359 : /*
1360 : * Record an unused line pointer that is left unchanged.
1361 : */
1362 : static void
1363 618730 : heap_prune_record_unchanged_lp_unused(Page page, PruneState *prstate, OffsetNumber offnum)
1364 : {
1365 : Assert(!prstate->processed[offnum]);
1366 618730 : prstate->processed[offnum] = true;
1367 618730 : }
1368 :
1369 : /*
1370 : * Record line pointer that is left unchanged. We consider freezing it, and
1371 : * update bookkeeping of tuple counts and page visibility.
1372 : */
1373 : static void
1374 51848262 : heap_prune_record_unchanged_lp_normal(Page page, PruneState *prstate, OffsetNumber offnum)
1375 : {
1376 : HeapTupleHeader htup;
1377 :
1378 : Assert(!prstate->processed[offnum]);
1379 51848262 : prstate->processed[offnum] = true;
1380 :
1381 51848262 : prstate->hastup = true; /* the page is not empty */
1382 :
1383 : /*
1384 : * The criteria for counting a tuple as live in this block need to match
1385 : * what analyze.c's acquire_sample_rows() does, otherwise VACUUM and
1386 : * ANALYZE may produce wildly different reltuples values, e.g. when there
1387 : * are many recently-dead tuples.
1388 : *
1389 : * The logic here is a bit simpler than acquire_sample_rows(), as VACUUM
1390 : * can't run inside a transaction block, which makes some cases impossible
1391 : * (e.g. in-progress insert from the same transaction).
1392 : *
1393 : * HEAPTUPLE_DEAD are handled by the other heap_prune_record_*()
1394 : * subroutines. They don't count dead items like acquire_sample_rows()
1395 : * does, because we assume that all dead items will become LP_UNUSED
1396 : * before VACUUM finishes. This difference is only superficial. VACUUM
1397 : * effectively agrees with ANALYZE about DEAD items, in the end. VACUUM
1398 : * won't remember LP_DEAD items, but only because they're not supposed to
1399 : * be left behind when it is done. (Cases where we bypass index vacuuming
1400 : * will violate this optimistic assumption, but the overall impact of that
1401 : * should be negligible.)
1402 : */
1403 51848262 : htup = (HeapTupleHeader) PageGetItem(page, PageGetItemId(page, offnum));
1404 :
1405 51848262 : switch (prstate->htsv[offnum])
1406 : {
1407 36046142 : case HEAPTUPLE_LIVE:
1408 :
1409 : /*
1410 : * Count it as live. Not only is this natural, but it's also what
1411 : * acquire_sample_rows() does.
1412 : */
1413 36046142 : prstate->live_tuples++;
1414 :
1415 : /*
1416 : * Is the tuple definitely visible to all transactions?
1417 : *
1418 : * NB: Like with per-tuple hint bits, we can't set the
1419 : * PD_ALL_VISIBLE flag if the inserter committed asynchronously.
1420 : * See SetHintBits for more info. Check that the tuple is hinted
1421 : * xmin-committed because of that.
1422 : */
1423 36046142 : if (prstate->all_visible)
1424 : {
1425 : TransactionId xmin;
1426 :
1427 26834528 : if (!HeapTupleHeaderXminCommitted(htup))
1428 : {
1429 330 : prstate->all_visible = false;
1430 330 : break;
1431 : }
1432 :
1433 : /*
1434 : * The inserter definitely committed. But is it old enough
1435 : * that everyone sees it as committed? A FrozenTransactionId
1436 : * is seen as committed to everyone. Otherwise, we check if
1437 : * there is a snapshot that considers this xid to still be
1438 : * running, and if so, we don't consider the page all-visible.
1439 : */
1440 26834198 : xmin = HeapTupleHeaderGetXmin(htup);
1441 :
1442 : /*
1443 : * For now always use prstate->cutoffs for this test, because
1444 : * we only update 'all_visible' when freezing is requested. We
1445 : * could use GlobalVisTestIsRemovableXid instead, if a
1446 : * non-freezing caller wanted to set the VM bit.
1447 : */
1448 : Assert(prstate->cutoffs);
1449 26834198 : if (!TransactionIdPrecedes(xmin, prstate->cutoffs->OldestXmin))
1450 : {
1451 6056 : prstate->all_visible = false;
1452 6056 : break;
1453 : }
1454 :
1455 : /* Track newest xmin on page. */
1456 26828142 : if (TransactionIdFollows(xmin, prstate->visibility_cutoff_xid) &&
1457 : TransactionIdIsNormal(xmin))
1458 217660 : prstate->visibility_cutoff_xid = xmin;
1459 : }
1460 36039756 : break;
1461 :
1462 507806 : case HEAPTUPLE_RECENTLY_DEAD:
1463 507806 : prstate->recently_dead_tuples++;
1464 507806 : prstate->all_visible = false;
1465 :
1466 : /*
1467 : * This tuple will soon become DEAD. Update the hint field so
1468 : * that the page is reconsidered for pruning in future.
1469 : */
1470 507806 : heap_prune_record_prunable(prstate,
1471 : HeapTupleHeaderGetUpdateXid(htup));
1472 507806 : break;
1473 :
1474 122958 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1475 :
1476 : /*
1477 : * We do not count these rows as live, because we expect the
1478 : * inserting transaction to update the counters at commit, and we
1479 : * assume that will happen only after we report our results. This
1480 : * assumption is a bit shaky, but it is what acquire_sample_rows()
1481 : * does, so be consistent.
1482 : */
1483 122958 : prstate->all_visible = false;
1484 :
1485 : /*
1486 : * If we wanted to optimize for aborts, we might consider marking
1487 : * the page prunable when we see INSERT_IN_PROGRESS. But we
1488 : * don't. See related decisions about when to mark the page
1489 : * prunable in heapam.c.
1490 : */
1491 122958 : break;
1492 :
1493 15171356 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1494 :
1495 : /*
1496 : * This an expected case during concurrent vacuum. Count such
1497 : * rows as live. As above, we assume the deleting transaction
1498 : * will commit and update the counters after we report.
1499 : */
1500 15171356 : prstate->live_tuples++;
1501 15171356 : prstate->all_visible = false;
1502 :
1503 : /*
1504 : * This tuple may soon become DEAD. Update the hint field so that
1505 : * the page is reconsidered for pruning in future.
1506 : */
1507 15171356 : heap_prune_record_prunable(prstate,
1508 : HeapTupleHeaderGetUpdateXid(htup));
1509 15171356 : break;
1510 :
1511 0 : default:
1512 :
1513 : /*
1514 : * DEAD tuples should've been passed to heap_prune_record_dead()
1515 : * or heap_prune_record_unused() instead.
1516 : */
1517 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result %d",
1518 : prstate->htsv[offnum]);
1519 : break;
1520 : }
1521 :
1522 : /* Consider freezing any normal tuples which will not be removed */
1523 51848262 : if (prstate->attempt_freeze)
1524 : {
1525 : bool totally_frozen;
1526 :
1527 48949516 : if ((heap_prepare_freeze_tuple(htup,
1528 48949516 : prstate->cutoffs,
1529 : &prstate->pagefrz,
1530 48949516 : &prstate->frozen[prstate->nfrozen],
1531 : &totally_frozen)))
1532 : {
1533 : /* Save prepared freeze plan for later */
1534 4360548 : prstate->frozen[prstate->nfrozen++].offset = offnum;
1535 : }
1536 :
1537 : /*
1538 : * If any tuple isn't either totally frozen already or eligible to
1539 : * become totally frozen (according to its freeze plan), then the page
1540 : * definitely cannot be set all-frozen in the visibility map later on.
1541 : */
1542 48949516 : if (!totally_frozen)
1543 16252756 : prstate->all_frozen = false;
1544 : }
1545 51848262 : }
1546 :
1547 :
1548 : /*
1549 : * Record line pointer that was already LP_DEAD and is left unchanged.
1550 : */
1551 : static void
1552 1929912 : heap_prune_record_unchanged_lp_dead(Page page, PruneState *prstate, OffsetNumber offnum)
1553 : {
1554 : Assert(!prstate->processed[offnum]);
1555 1929912 : prstate->processed[offnum] = true;
1556 :
1557 : /*
1558 : * Deliberately don't set hastup for LP_DEAD items. We make the soft
1559 : * assumption that any LP_DEAD items encountered here will become
1560 : * LP_UNUSED later on, before count_nondeletable_pages is reached. If we
1561 : * don't make this assumption then rel truncation will only happen every
1562 : * other VACUUM, at most. Besides, VACUUM must treat
1563 : * hastup/nonempty_pages as provisional no matter how LP_DEAD items are
1564 : * handled (handled here, or handled later on).
1565 : *
1566 : * Similarly, don't unset all_visible until later, at the end of
1567 : * heap_page_prune_and_freeze(). This will allow us to attempt to freeze
1568 : * the page after pruning. As long as we unset it before updating the
1569 : * visibility map, this will be correct.
1570 : */
1571 :
1572 : /* Record the dead offset for vacuum */
1573 1929912 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1574 1929912 : }
1575 :
1576 : /*
1577 : * Record LP_REDIRECT that is left unchanged.
1578 : */
1579 : static void
1580 613892 : heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum)
1581 : {
1582 : /*
1583 : * A redirect line pointer doesn't count as a live tuple.
1584 : *
1585 : * If we leave a redirect line pointer in place, there will be another
1586 : * tuple on the page that it points to. We will do the bookkeeping for
1587 : * that separately. So we have nothing to do here, except remember that
1588 : * we processed this item.
1589 : */
1590 : Assert(!prstate->processed[offnum]);
1591 613892 : prstate->processed[offnum] = true;
1592 613892 : }
1593 :
1594 : /*
1595 : * Perform the actual page changes needed by heap_page_prune_and_freeze().
1596 : *
1597 : * If 'lp_truncate_only' is set, we are merely marking LP_DEAD line pointers
1598 : * as unused, not redirecting or removing anything else. The
1599 : * PageRepairFragmentation() call is skipped in that case.
1600 : *
1601 : * If 'lp_truncate_only' is not set, the caller must hold a cleanup lock on
1602 : * the buffer. If it is set, an ordinary exclusive lock suffices.
1603 : */
1604 : void
1605 122134 : heap_page_prune_execute(Buffer buffer, bool lp_truncate_only,
1606 : OffsetNumber *redirected, int nredirected,
1607 : OffsetNumber *nowdead, int ndead,
1608 : OffsetNumber *nowunused, int nunused)
1609 : {
1610 122134 : Page page = BufferGetPage(buffer);
1611 : OffsetNumber *offnum;
1612 : HeapTupleHeader htup PG_USED_FOR_ASSERTS_ONLY;
1613 :
1614 : /* Shouldn't be called unless there's something to do */
1615 : Assert(nredirected > 0 || ndead > 0 || nunused > 0);
1616 :
1617 : /* If 'lp_truncate_only', we can only remove already-dead line pointers */
1618 : Assert(!lp_truncate_only || (nredirected == 0 && ndead == 0));
1619 :
1620 : /* Update all redirected line pointers */
1621 122134 : offnum = redirected;
1622 288608 : for (int i = 0; i < nredirected; i++)
1623 : {
1624 166474 : OffsetNumber fromoff = *offnum++;
1625 166474 : OffsetNumber tooff = *offnum++;
1626 166474 : ItemId fromlp = PageGetItemId(page, fromoff);
1627 : ItemId tolp PG_USED_FOR_ASSERTS_ONLY;
1628 :
1629 : #ifdef USE_ASSERT_CHECKING
1630 :
1631 : /*
1632 : * Any existing item that we set as an LP_REDIRECT (any 'from' item)
1633 : * must be the first item from a HOT chain. If the item has tuple
1634 : * storage then it can't be a heap-only tuple. Otherwise we are just
1635 : * maintaining an existing LP_REDIRECT from an existing HOT chain that
1636 : * has been pruned at least once before now.
1637 : */
1638 : if (!ItemIdIsRedirected(fromlp))
1639 : {
1640 : Assert(ItemIdHasStorage(fromlp) && ItemIdIsNormal(fromlp));
1641 :
1642 : htup = (HeapTupleHeader) PageGetItem(page, fromlp);
1643 : Assert(!HeapTupleHeaderIsHeapOnly(htup));
1644 : }
1645 : else
1646 : {
1647 : /* We shouldn't need to redundantly set the redirect */
1648 : Assert(ItemIdGetRedirect(fromlp) != tooff);
1649 : }
1650 :
1651 : /*
1652 : * The item that we're about to set as an LP_REDIRECT (the 'from'
1653 : * item) will point to an existing item (the 'to' item) that is
1654 : * already a heap-only tuple. There can be at most one LP_REDIRECT
1655 : * item per HOT chain.
1656 : *
1657 : * We need to keep around an LP_REDIRECT item (after original
1658 : * non-heap-only root tuple gets pruned away) so that it's always
1659 : * possible for VACUUM to easily figure out what TID to delete from
1660 : * indexes when an entire HOT chain becomes dead. A heap-only tuple
1661 : * can never become LP_DEAD; an LP_REDIRECT item or a regular heap
1662 : * tuple can.
1663 : *
1664 : * This check may miss problems, e.g. the target of a redirect could
1665 : * be marked as unused subsequently. The page_verify_redirects() check
1666 : * below will catch such problems.
1667 : */
1668 : tolp = PageGetItemId(page, tooff);
1669 : Assert(ItemIdHasStorage(tolp) && ItemIdIsNormal(tolp));
1670 : htup = (HeapTupleHeader) PageGetItem(page, tolp);
1671 : Assert(HeapTupleHeaderIsHeapOnly(htup));
1672 : #endif
1673 :
1674 166474 : ItemIdSetRedirect(fromlp, tooff);
1675 : }
1676 :
1677 : /* Update all now-dead line pointers */
1678 122134 : offnum = nowdead;
1679 3225094 : for (int i = 0; i < ndead; i++)
1680 : {
1681 3102960 : OffsetNumber off = *offnum++;
1682 3102960 : ItemId lp = PageGetItemId(page, off);
1683 :
1684 : #ifdef USE_ASSERT_CHECKING
1685 :
1686 : /*
1687 : * An LP_DEAD line pointer must be left behind when the original item
1688 : * (which is dead to everybody) could still be referenced by a TID in
1689 : * an index. This should never be necessary with any individual
1690 : * heap-only tuple item, though. (It's not clear how much of a problem
1691 : * that would be, but there is no reason to allow it.)
1692 : */
1693 : if (ItemIdHasStorage(lp))
1694 : {
1695 : Assert(ItemIdIsNormal(lp));
1696 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1697 : Assert(!HeapTupleHeaderIsHeapOnly(htup));
1698 : }
1699 : else
1700 : {
1701 : /* Whole HOT chain becomes dead */
1702 : Assert(ItemIdIsRedirected(lp));
1703 : }
1704 : #endif
1705 :
1706 3102960 : ItemIdSetDead(lp);
1707 : }
1708 :
1709 : /* Update all now-unused line pointers */
1710 122134 : offnum = nowunused;
1711 719730 : for (int i = 0; i < nunused; i++)
1712 : {
1713 597596 : OffsetNumber off = *offnum++;
1714 597596 : ItemId lp = PageGetItemId(page, off);
1715 :
1716 : #ifdef USE_ASSERT_CHECKING
1717 :
1718 : if (lp_truncate_only)
1719 : {
1720 : /* Setting LP_DEAD to LP_UNUSED in vacuum's second pass */
1721 : Assert(ItemIdIsDead(lp) && !ItemIdHasStorage(lp));
1722 : }
1723 : else
1724 : {
1725 : /*
1726 : * When heap_page_prune_and_freeze() was called, mark_unused_now
1727 : * may have been passed as true, which allows would-be LP_DEAD
1728 : * items to be made LP_UNUSED instead. This is only possible if
1729 : * the relation has no indexes. If there are any dead items, then
1730 : * mark_unused_now was not true and every item being marked
1731 : * LP_UNUSED must refer to a heap-only tuple.
1732 : */
1733 : if (ndead > 0)
1734 : {
1735 : Assert(ItemIdHasStorage(lp) && ItemIdIsNormal(lp));
1736 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1737 : Assert(HeapTupleHeaderIsHeapOnly(htup));
1738 : }
1739 : else
1740 : Assert(ItemIdIsUsed(lp));
1741 : }
1742 :
1743 : #endif
1744 :
1745 597596 : ItemIdSetUnused(lp);
1746 : }
1747 :
1748 122134 : if (lp_truncate_only)
1749 4202 : PageTruncateLinePointerArray(page);
1750 : else
1751 : {
1752 : /*
1753 : * Finally, repair any fragmentation, and update the page's hint bit
1754 : * about whether it has free pointers.
1755 : */
1756 117932 : PageRepairFragmentation(page);
1757 :
1758 : /*
1759 : * Now that the page has been modified, assert that redirect items
1760 : * still point to valid targets.
1761 : */
1762 117932 : page_verify_redirects(page);
1763 : }
1764 122134 : }
1765 :
1766 :
1767 : /*
1768 : * If built with assertions, verify that all LP_REDIRECT items point to a
1769 : * valid item.
1770 : *
1771 : * One way that bugs related to HOT pruning show is redirect items pointing to
1772 : * removed tuples. It's not trivial to reliably check that marking an item
1773 : * unused will not orphan a redirect item during heap_prune_chain() /
1774 : * heap_page_prune_execute(), so we additionally check the whole page after
1775 : * pruning. Without this check such bugs would typically only cause asserts
1776 : * later, potentially well after the corruption has been introduced.
1777 : *
1778 : * Also check comments in heap_page_prune_execute()'s redirection loop.
1779 : */
1780 : static void
1781 117932 : page_verify_redirects(Page page)
1782 : {
1783 : #ifdef USE_ASSERT_CHECKING
1784 : OffsetNumber offnum;
1785 : OffsetNumber maxoff;
1786 :
1787 : maxoff = PageGetMaxOffsetNumber(page);
1788 : for (offnum = FirstOffsetNumber;
1789 : offnum <= maxoff;
1790 : offnum = OffsetNumberNext(offnum))
1791 : {
1792 : ItemId itemid = PageGetItemId(page, offnum);
1793 : OffsetNumber targoff;
1794 : ItemId targitem;
1795 : HeapTupleHeader htup;
1796 :
1797 : if (!ItemIdIsRedirected(itemid))
1798 : continue;
1799 :
1800 : targoff = ItemIdGetRedirect(itemid);
1801 : targitem = PageGetItemId(page, targoff);
1802 :
1803 : Assert(ItemIdIsUsed(targitem));
1804 : Assert(ItemIdIsNormal(targitem));
1805 : Assert(ItemIdHasStorage(targitem));
1806 : htup = (HeapTupleHeader) PageGetItem(page, targitem);
1807 : Assert(HeapTupleHeaderIsHeapOnly(htup));
1808 : }
1809 : #endif
1810 117932 : }
1811 :
1812 :
1813 : /*
1814 : * For all items in this page, find their respective root line pointers.
1815 : * If item k is part of a HOT-chain with root at item j, then we set
1816 : * root_offsets[k - 1] = j.
1817 : *
1818 : * The passed-in root_offsets array must have MaxHeapTuplesPerPage entries.
1819 : * Unused entries are filled with InvalidOffsetNumber (zero).
1820 : *
1821 : * The function must be called with at least share lock on the buffer, to
1822 : * prevent concurrent prune operations.
1823 : *
1824 : * Note: The information collected here is valid only as long as the caller
1825 : * holds a pin on the buffer. Once pin is released, a tuple might be pruned
1826 : * and reused by a completely unrelated tuple.
1827 : */
1828 : void
1829 223186 : heap_get_root_tuples(Page page, OffsetNumber *root_offsets)
1830 : {
1831 : OffsetNumber offnum,
1832 : maxoff;
1833 :
1834 223186 : MemSet(root_offsets, InvalidOffsetNumber,
1835 : MaxHeapTuplesPerPage * sizeof(OffsetNumber));
1836 :
1837 223186 : maxoff = PageGetMaxOffsetNumber(page);
1838 17999426 : for (offnum = FirstOffsetNumber; offnum <= maxoff; offnum = OffsetNumberNext(offnum))
1839 : {
1840 17776240 : ItemId lp = PageGetItemId(page, offnum);
1841 : HeapTupleHeader htup;
1842 : OffsetNumber nextoffnum;
1843 : TransactionId priorXmax;
1844 :
1845 : /* skip unused and dead items */
1846 17776240 : if (!ItemIdIsUsed(lp) || ItemIdIsDead(lp))
1847 22468 : continue;
1848 :
1849 17753772 : if (ItemIdIsNormal(lp))
1850 : {
1851 17745358 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1852 :
1853 : /*
1854 : * Check if this tuple is part of a HOT-chain rooted at some other
1855 : * tuple. If so, skip it for now; we'll process it when we find
1856 : * its root.
1857 : */
1858 17745358 : if (HeapTupleHeaderIsHeapOnly(htup))
1859 9134 : continue;
1860 :
1861 : /*
1862 : * This is either a plain tuple or the root of a HOT-chain.
1863 : * Remember it in the mapping.
1864 : */
1865 17736224 : root_offsets[offnum - 1] = offnum;
1866 :
1867 : /* If it's not the start of a HOT-chain, we're done with it */
1868 17736224 : if (!HeapTupleHeaderIsHotUpdated(htup))
1869 17735722 : continue;
1870 :
1871 : /* Set up to scan the HOT-chain */
1872 502 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1873 502 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1874 : }
1875 : else
1876 : {
1877 : /* Must be a redirect item. We do not set its root_offsets entry */
1878 : Assert(ItemIdIsRedirected(lp));
1879 : /* Set up to scan the HOT-chain */
1880 8414 : nextoffnum = ItemIdGetRedirect(lp);
1881 8414 : priorXmax = InvalidTransactionId;
1882 : }
1883 :
1884 : /*
1885 : * Now follow the HOT-chain and collect other tuples in the chain.
1886 : *
1887 : * Note: Even though this is a nested loop, the complexity of the
1888 : * function is O(N) because a tuple in the page should be visited not
1889 : * more than twice, once in the outer loop and once in HOT-chain
1890 : * chases.
1891 : */
1892 : for (;;)
1893 : {
1894 : /* Sanity check (pure paranoia) */
1895 9128 : if (offnum < FirstOffsetNumber)
1896 0 : break;
1897 :
1898 : /*
1899 : * An offset past the end of page's line pointer array is possible
1900 : * when the array was truncated
1901 : */
1902 9128 : if (offnum > maxoff)
1903 0 : break;
1904 :
1905 9128 : lp = PageGetItemId(page, nextoffnum);
1906 :
1907 : /* Check for broken chains */
1908 9128 : if (!ItemIdIsNormal(lp))
1909 0 : break;
1910 :
1911 9128 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1912 :
1913 9842 : if (TransactionIdIsValid(priorXmax) &&
1914 714 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(htup)))
1915 0 : break;
1916 :
1917 : /* Remember the root line pointer for this item */
1918 9128 : root_offsets[nextoffnum - 1] = offnum;
1919 :
1920 : /* Advance to next chain member, if any */
1921 9128 : if (!HeapTupleHeaderIsHotUpdated(htup))
1922 8916 : break;
1923 :
1924 : /* HOT implies it can't have moved to different partition */
1925 : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1926 :
1927 212 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1928 212 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1929 : }
1930 : }
1931 223186 : }
1932 :
1933 :
1934 : /*
1935 : * Compare fields that describe actions required to freeze tuple with caller's
1936 : * open plan. If everything matches then the frz tuple plan is equivalent to
1937 : * caller's plan.
1938 : */
1939 : static inline bool
1940 1905326 : heap_log_freeze_eq(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
1941 : {
1942 1905326 : if (plan->xmax == frz->xmax &&
1943 1902744 : plan->t_infomask2 == frz->t_infomask2 &&
1944 1900950 : plan->t_infomask == frz->t_infomask &&
1945 1895870 : plan->frzflags == frz->frzflags)
1946 1895870 : return true;
1947 :
1948 : /* Caller must call heap_log_freeze_new_plan again for frz */
1949 9456 : return false;
1950 : }
1951 :
1952 : /*
1953 : * Comparator used to deduplicate the freeze plans used in WAL records.
1954 : */
1955 : static int
1956 2640838 : heap_log_freeze_cmp(const void *arg1, const void *arg2)
1957 : {
1958 2640838 : HeapTupleFreeze *frz1 = (HeapTupleFreeze *) arg1;
1959 2640838 : HeapTupleFreeze *frz2 = (HeapTupleFreeze *) arg2;
1960 :
1961 2640838 : if (frz1->xmax < frz2->xmax)
1962 26490 : return -1;
1963 2614348 : else if (frz1->xmax > frz2->xmax)
1964 27094 : return 1;
1965 :
1966 2587254 : if (frz1->t_infomask2 < frz2->t_infomask2)
1967 8888 : return -1;
1968 2578366 : else if (frz1->t_infomask2 > frz2->t_infomask2)
1969 9506 : return 1;
1970 :
1971 2568860 : if (frz1->t_infomask < frz2->t_infomask)
1972 22398 : return -1;
1973 2546462 : else if (frz1->t_infomask > frz2->t_infomask)
1974 31290 : return 1;
1975 :
1976 2515172 : if (frz1->frzflags < frz2->frzflags)
1977 0 : return -1;
1978 2515172 : else if (frz1->frzflags > frz2->frzflags)
1979 0 : return 1;
1980 :
1981 : /*
1982 : * heap_log_freeze_eq would consider these tuple-wise plans to be equal.
1983 : * (So the tuples will share a single canonical freeze plan.)
1984 : *
1985 : * We tiebreak on page offset number to keep each freeze plan's page
1986 : * offset number array individually sorted. (Unnecessary, but be tidy.)
1987 : */
1988 2515172 : if (frz1->offset < frz2->offset)
1989 2123142 : return -1;
1990 392030 : else if (frz1->offset > frz2->offset)
1991 392030 : return 1;
1992 :
1993 : Assert(false);
1994 0 : return 0;
1995 : }
1996 :
1997 : /*
1998 : * Start new plan initialized using tuple-level actions. At least one tuple
1999 : * will have steps required to freeze described by caller's plan during REDO.
2000 : */
2001 : static inline void
2002 55828 : heap_log_freeze_new_plan(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2003 : {
2004 55828 : plan->xmax = frz->xmax;
2005 55828 : plan->t_infomask2 = frz->t_infomask2;
2006 55828 : plan->t_infomask = frz->t_infomask;
2007 55828 : plan->frzflags = frz->frzflags;
2008 55828 : plan->ntuples = 1; /* for now */
2009 55828 : }
2010 :
2011 : /*
2012 : * Deduplicate tuple-based freeze plans so that each distinct set of
2013 : * processing steps is only stored once in the WAL record.
2014 : * Called during original execution of freezing (for logged relations).
2015 : *
2016 : * Return value is number of plans set in *plans_out for caller. Also writes
2017 : * an array of offset numbers into *offsets_out output argument for caller
2018 : * (actually there is one array per freeze plan, but that's not of immediate
2019 : * concern to our caller).
2020 : */
2021 : static int
2022 46372 : heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
2023 : xlhp_freeze_plan *plans_out,
2024 : OffsetNumber *offsets_out)
2025 : {
2026 46372 : int nplans = 0;
2027 :
2028 : /* Sort tuple-based freeze plans in the order required to deduplicate */
2029 46372 : qsort(tuples, ntuples, sizeof(HeapTupleFreeze), heap_log_freeze_cmp);
2030 :
2031 1998070 : for (int i = 0; i < ntuples; i++)
2032 : {
2033 1951698 : HeapTupleFreeze *frz = tuples + i;
2034 :
2035 1951698 : if (i == 0)
2036 : {
2037 : /* New canonical freeze plan starting with first tup */
2038 46372 : heap_log_freeze_new_plan(plans_out, frz);
2039 46372 : nplans++;
2040 : }
2041 1905326 : else if (heap_log_freeze_eq(plans_out, frz))
2042 : {
2043 : /* tup matches open canonical plan -- include tup in it */
2044 : Assert(offsets_out[i - 1] < frz->offset);
2045 1895870 : plans_out->ntuples++;
2046 : }
2047 : else
2048 : {
2049 : /* Tup doesn't match current plan -- done with it now */
2050 9456 : plans_out++;
2051 :
2052 : /* New canonical freeze plan starting with this tup */
2053 9456 : heap_log_freeze_new_plan(plans_out, frz);
2054 9456 : nplans++;
2055 : }
2056 :
2057 : /*
2058 : * Save page offset number in dedicated buffer in passing.
2059 : *
2060 : * REDO routine relies on the record's offset numbers array grouping
2061 : * offset numbers by freeze plan. The sort order within each grouping
2062 : * is ascending offset number order, just to keep things tidy.
2063 : */
2064 1951698 : offsets_out[i] = frz->offset;
2065 : }
2066 :
2067 : Assert(nplans > 0 && nplans <= ntuples);
2068 :
2069 46372 : return nplans;
2070 : }
2071 :
2072 : /*
2073 : * Write an XLOG_HEAP2_PRUNE* WAL record
2074 : *
2075 : * This is used for several different page maintenance operations:
2076 : *
2077 : * - Page pruning, in VACUUM's 1st pass or on access: Some items are
2078 : * redirected, some marked dead, and some removed altogether.
2079 : *
2080 : * - Freezing: Items are marked as 'frozen'.
2081 : *
2082 : * - Vacuum, 2nd pass: Items that are already LP_DEAD are marked as unused.
2083 : *
2084 : * They have enough commonalities that we use a single WAL record for them
2085 : * all.
2086 : *
2087 : * If replaying the record requires a cleanup lock, pass cleanup_lock = true.
2088 : * Replaying 'redirected' or 'dead' items always requires a cleanup lock, but
2089 : * replaying 'unused' items depends on whether they were all previously marked
2090 : * as dead.
2091 : *
2092 : * If the VM is being updated, vmflags will contain the bits to set. In this
2093 : * case, vmbuffer should already have been updated and marked dirty and should
2094 : * still be pinned and locked.
2095 : *
2096 : * Note: This function scribbles on the 'frozen' array.
2097 : *
2098 : * Note: This is called in a critical section, so careful what you do here.
2099 : */
2100 : void
2101 170754 : log_heap_prune_and_freeze(Relation relation, Buffer buffer,
2102 : Buffer vmbuffer, uint8 vmflags,
2103 : TransactionId conflict_xid,
2104 : bool cleanup_lock,
2105 : PruneReason reason,
2106 : HeapTupleFreeze *frozen, int nfrozen,
2107 : OffsetNumber *redirected, int nredirected,
2108 : OffsetNumber *dead, int ndead,
2109 : OffsetNumber *unused, int nunused)
2110 : {
2111 : xl_heap_prune xlrec;
2112 : XLogRecPtr recptr;
2113 : uint8 info;
2114 : uint8 regbuf_flags_heap;
2115 :
2116 : /* The following local variables hold data registered in the WAL record: */
2117 : xlhp_freeze_plan plans[MaxHeapTuplesPerPage];
2118 : xlhp_freeze_plans freeze_plans;
2119 : xlhp_prune_items redirect_items;
2120 : xlhp_prune_items dead_items;
2121 : xlhp_prune_items unused_items;
2122 : OffsetNumber frz_offsets[MaxHeapTuplesPerPage];
2123 170754 : bool do_prune = nredirected > 0 || ndead > 0 || nunused > 0;
2124 170754 : bool do_set_vm = vmflags & VISIBILITYMAP_VALID_BITS;
2125 :
2126 : Assert((vmflags & VISIBILITYMAP_VALID_BITS) == vmflags);
2127 :
2128 170754 : xlrec.flags = 0;
2129 170754 : regbuf_flags_heap = REGBUF_STANDARD;
2130 :
2131 : /*
2132 : * We can avoid an FPI of the heap page if the only modification we are
2133 : * making to it is to set PD_ALL_VISIBLE and checksums/wal_log_hints are
2134 : * disabled. Note that if we explicitly skip an FPI, we must not stamp the
2135 : * heap page with this record's LSN. Recovery skips records <= the stamped
2136 : * LSN, so this could lead to skipping an earlier FPI needed to repair a
2137 : * torn page.
2138 : */
2139 170754 : if (!do_prune &&
2140 0 : nfrozen == 0 &&
2141 0 : (!do_set_vm || !XLogHintBitIsNeeded()))
2142 0 : regbuf_flags_heap |= REGBUF_NO_IMAGE;
2143 :
2144 : /*
2145 : * Prepare data for the buffer. The arrays are not actually in the
2146 : * buffer, but we pretend that they are. When XLogInsert stores a full
2147 : * page image, the arrays can be omitted.
2148 : */
2149 170754 : XLogBeginInsert();
2150 170754 : XLogRegisterBuffer(0, buffer, regbuf_flags_heap);
2151 :
2152 170754 : if (do_set_vm)
2153 26212 : XLogRegisterBuffer(1, vmbuffer, 0);
2154 :
2155 170754 : if (nfrozen > 0)
2156 : {
2157 : int nplans;
2158 :
2159 46372 : xlrec.flags |= XLHP_HAS_FREEZE_PLANS;
2160 :
2161 : /*
2162 : * Prepare deduplicated representation for use in the WAL record. This
2163 : * destructively sorts frozen tuples array in-place.
2164 : */
2165 46372 : nplans = heap_log_freeze_plan(frozen, nfrozen, plans, frz_offsets);
2166 :
2167 46372 : freeze_plans.nplans = nplans;
2168 46372 : XLogRegisterBufData(0, &freeze_plans,
2169 : offsetof(xlhp_freeze_plans, plans));
2170 46372 : XLogRegisterBufData(0, plans,
2171 : sizeof(xlhp_freeze_plan) * nplans);
2172 : }
2173 170754 : if (nredirected > 0)
2174 : {
2175 33186 : xlrec.flags |= XLHP_HAS_REDIRECTIONS;
2176 :
2177 33186 : redirect_items.ntargets = nredirected;
2178 33186 : XLogRegisterBufData(0, &redirect_items,
2179 : offsetof(xlhp_prune_items, data));
2180 33186 : XLogRegisterBufData(0, redirected,
2181 : sizeof(OffsetNumber[2]) * nredirected);
2182 : }
2183 170754 : if (ndead > 0)
2184 : {
2185 76344 : xlrec.flags |= XLHP_HAS_DEAD_ITEMS;
2186 :
2187 76344 : dead_items.ntargets = ndead;
2188 76344 : XLogRegisterBufData(0, &dead_items,
2189 : offsetof(xlhp_prune_items, data));
2190 76344 : XLogRegisterBufData(0, dead,
2191 : sizeof(OffsetNumber) * ndead);
2192 : }
2193 170754 : if (nunused > 0)
2194 : {
2195 48210 : xlrec.flags |= XLHP_HAS_NOW_UNUSED_ITEMS;
2196 :
2197 48210 : unused_items.ntargets = nunused;
2198 48210 : XLogRegisterBufData(0, &unused_items,
2199 : offsetof(xlhp_prune_items, data));
2200 48210 : XLogRegisterBufData(0, unused,
2201 : sizeof(OffsetNumber) * nunused);
2202 : }
2203 170754 : if (nfrozen > 0)
2204 46372 : XLogRegisterBufData(0, frz_offsets,
2205 : sizeof(OffsetNumber) * nfrozen);
2206 :
2207 : /*
2208 : * Prepare the main xl_heap_prune record. We already set the XLHP_HAS_*
2209 : * flag above.
2210 : */
2211 170754 : if (vmflags & VISIBILITYMAP_ALL_VISIBLE)
2212 : {
2213 26212 : xlrec.flags |= XLHP_VM_ALL_VISIBLE;
2214 26212 : if (vmflags & VISIBILITYMAP_ALL_FROZEN)
2215 21128 : xlrec.flags |= XLHP_VM_ALL_FROZEN;
2216 : }
2217 170754 : if (RelationIsAccessibleInLogicalDecoding(relation))
2218 1280 : xlrec.flags |= XLHP_IS_CATALOG_REL;
2219 170754 : if (TransactionIdIsValid(conflict_xid))
2220 139762 : xlrec.flags |= XLHP_HAS_CONFLICT_HORIZON;
2221 170754 : if (cleanup_lock)
2222 144342 : xlrec.flags |= XLHP_CLEANUP_LOCK;
2223 : else
2224 : {
2225 : Assert(nredirected == 0 && ndead == 0);
2226 : /* also, any items in 'unused' must've been LP_DEAD previously */
2227 : }
2228 170754 : XLogRegisterData(&xlrec, SizeOfHeapPrune);
2229 170754 : if (TransactionIdIsValid(conflict_xid))
2230 139762 : XLogRegisterData(&conflict_xid, sizeof(TransactionId));
2231 :
2232 170754 : switch (reason)
2233 : {
2234 80246 : case PRUNE_ON_ACCESS:
2235 80246 : info = XLOG_HEAP2_PRUNE_ON_ACCESS;
2236 80246 : break;
2237 64096 : case PRUNE_VACUUM_SCAN:
2238 64096 : info = XLOG_HEAP2_PRUNE_VACUUM_SCAN;
2239 64096 : break;
2240 26412 : case PRUNE_VACUUM_CLEANUP:
2241 26412 : info = XLOG_HEAP2_PRUNE_VACUUM_CLEANUP;
2242 26412 : break;
2243 0 : default:
2244 0 : elog(ERROR, "unrecognized prune reason: %d", (int) reason);
2245 : break;
2246 : }
2247 170754 : recptr = XLogInsert(RM_HEAP2_ID, info);
2248 :
2249 170754 : if (do_set_vm)
2250 : {
2251 : Assert(BufferIsDirty(vmbuffer));
2252 26212 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2253 : }
2254 :
2255 : /*
2256 : * See comment at the top of the function about regbuf_flags_heap for
2257 : * details on when we can advance the page LSN.
2258 : */
2259 170754 : if (do_prune || nfrozen > 0 || (do_set_vm && XLogHintBitIsNeeded()))
2260 : {
2261 : Assert(BufferIsDirty(buffer));
2262 170754 : PageSetLSN(BufferGetPage(buffer), recptr);
2263 : }
2264 170754 : }
|
