Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * pruneheap.c
4 : : * heap page pruning and HOT-chain management code
5 : : *
6 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/access/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/visibilitymap.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 "storage/freespace.h"
31 : : #include "utils/rel.h"
32 : : #include "utils/snapmgr.h"
33 : :
34 : : /* Working data for heap_page_prune_and_freeze() and subroutines */
35 : : typedef struct
36 : : {
37 : : /*-------------------------------------------------------
38 : : * Arguments passed to heap_page_prune_and_freeze()
39 : : *-------------------------------------------------------
40 : : */
41 : :
42 : : /* tuple visibility test, initialized for the relation */
43 : : GlobalVisState *vistest;
44 : : /* whether or not dead items can be set LP_UNUSED during pruning */
45 : : bool mark_unused_now;
46 : : /* whether to attempt freezing tuples */
47 : : bool attempt_freeze;
48 : : /* whether to attempt setting the VM */
49 : : bool attempt_set_vm;
50 : : struct VacuumCutoffs *cutoffs;
51 : : Relation relation;
52 : :
53 : : /*
54 : : * Keep the buffer, block, and page handy so that helpers needing to
55 : : * access them don't need to make repeated calls to BufferGetBlockNumber()
56 : : * and BufferGetPage().
57 : : */
58 : : BlockNumber block;
59 : : Buffer buffer;
60 : : Page page;
61 : :
62 : : /*-------------------------------------------------------
63 : : * Fields describing what to do to the page
64 : : *-------------------------------------------------------
65 : : */
66 : : TransactionId new_prune_xid; /* new prune hint value */
67 : : TransactionId latest_xid_removed;
68 : : int nredirected; /* numbers of entries in arrays below */
69 : : int ndead;
70 : : int nunused;
71 : : int nfrozen;
72 : : /* arrays that accumulate indexes of items to be changed */
73 : : OffsetNumber redirected[MaxHeapTuplesPerPage * 2];
74 : : OffsetNumber nowdead[MaxHeapTuplesPerPage];
75 : : OffsetNumber nowunused[MaxHeapTuplesPerPage];
76 : : HeapTupleFreeze frozen[MaxHeapTuplesPerPage];
77 : :
78 : : /*
79 : : * set_all_visible and set_all_frozen indicate if the all-visible and
80 : : * all-frozen bits in the visibility map can be set for this page after
81 : : * pruning. They are only tracked when the caller requests VM updates
82 : : * (attempt_set_vm); otherwise they remain false throughout.
83 : : *
84 : : * NOTE: set_all_visible and set_all_frozen initially don't include
85 : : * LP_DEAD items. That's convenient for heap_page_prune_and_freeze() to
86 : : * use them to decide whether to opportunistically freeze the page or not.
87 : : * The set_all_visible and set_all_frozen values ultimately used to set
88 : : * the VM are adjusted to include LP_DEAD items after we determine whether
89 : : * or not to opportunistically freeze.
90 : : */
91 : : bool set_all_visible;
92 : : bool set_all_frozen;
93 : :
94 : : /*-------------------------------------------------------
95 : : * Working state for HOT chain processing
96 : : *-------------------------------------------------------
97 : : */
98 : :
99 : : /*
100 : : * 'root_items' contains offsets of all LP_REDIRECT line pointers and
101 : : * normal non-HOT tuples. They can be stand-alone items or the first item
102 : : * in a HOT chain. 'heaponly_items' contains heap-only tuples which can
103 : : * only be removed as part of a HOT chain.
104 : : */
105 : : int nroot_items;
106 : : OffsetNumber root_items[MaxHeapTuplesPerPage];
107 : : int nheaponly_items;
108 : : OffsetNumber heaponly_items[MaxHeapTuplesPerPage];
109 : :
110 : : /*
111 : : * processed[offnum] is true if item at offnum has been processed.
112 : : *
113 : : * This needs to be MaxHeapTuplesPerPage + 1 long as FirstOffsetNumber is
114 : : * 1. Otherwise every access would need to subtract 1.
115 : : */
116 : : bool processed[MaxHeapTuplesPerPage + 1];
117 : :
118 : : /*
119 : : * Tuple visibility is only computed once for each tuple, for correctness
120 : : * and efficiency reasons; see comment in heap_page_prune_and_freeze() for
121 : : * details. This is of type int8[], instead of HTSV_Result[], so we can
122 : : * use -1 to indicate no visibility has been computed, e.g. for LP_DEAD
123 : : * items.
124 : : *
125 : : * This needs to be MaxHeapTuplesPerPage + 1 long as FirstOffsetNumber is
126 : : * 1. Otherwise every access would need to subtract 1.
127 : : */
128 : : int8 htsv[MaxHeapTuplesPerPage + 1];
129 : :
130 : : /*-------------------------------------------------------
131 : : * Working state for freezing
132 : : *-------------------------------------------------------
133 : : */
134 : : HeapPageFreeze pagefrz;
135 : :
136 : : /*-------------------------------------------------------
137 : : * Working state for visibility map processing
138 : : *-------------------------------------------------------
139 : : */
140 : :
141 : : /*
142 : : * Caller must provide a pinned vmbuffer corresponding to the heap block
143 : : * passed to heap_page_prune_and_freeze(). We will fix any corruption
144 : : * found in the VM and set the VM if the page is all-visible/all-frozen.
145 : : */
146 : : Buffer vmbuffer;
147 : :
148 : : /*
149 : : * The state of the VM bits at the beginning of pruning and the state they
150 : : * will be in at the end.
151 : : */
152 : : uint8 old_vmbits;
153 : : uint8 new_vmbits;
154 : :
155 : : /* The newest xmin of live tuples on the page */
156 : : TransactionId newest_live_xid;
157 : :
158 : : /*-------------------------------------------------------
159 : : * Information about what was done
160 : : *
161 : : * These fields are not used by pruning itself for the most part, but are
162 : : * used to collect information about what was pruned and what state the
163 : : * page is in after pruning, for the benefit of the caller. They are
164 : : * copied to the caller's PruneFreezeResult at the end.
165 : : * -------------------------------------------------------
166 : : */
167 : :
168 : : int ndeleted; /* Number of tuples deleted from the page */
169 : :
170 : : /* Number of live and recently dead tuples, after pruning */
171 : : int live_tuples;
172 : : int recently_dead_tuples;
173 : :
174 : : /* Whether or not the page makes rel truncation unsafe */
175 : : bool hastup;
176 : :
177 : : /*
178 : : * LP_DEAD items on the page after pruning. Includes existing LP_DEAD
179 : : * items
180 : : */
181 : : int lpdead_items; /* number of items in the array */
182 : : OffsetNumber *deadoffsets; /* points directly to presult->deadoffsets */
183 : : } PruneState;
184 : :
185 : : /*
186 : : * Type of visibility map corruption detected on a heap page and its
187 : : * associated VM page. Passed to heap_page_fix_vm_corruption() so the caller
188 : : * can specify what it found rather than having the function rederive the
189 : : * corruption from page state.
190 : : */
191 : : typedef enum VMCorruptionType
192 : : {
193 : : /* VM bits are set but the heap page-level PD_ALL_VISIBLE flag is not */
194 : : VM_CORRUPT_MISSING_PAGE_HINT,
195 : : /* LP_DEAD line pointers found on a page marked all-visible */
196 : : VM_CORRUPT_LPDEAD,
197 : : /* Tuple not visible to all transactions on a page marked all-visible */
198 : : VM_CORRUPT_TUPLE_VISIBILITY,
199 : : } VMCorruptionType;
200 : :
201 : : /* Local functions */
202 : : static void prune_freeze_setup(PruneFreezeParams *params,
203 : : TransactionId *new_relfrozen_xid,
204 : : MultiXactId *new_relmin_mxid,
205 : : PruneFreezeResult *presult,
206 : : PruneState *prstate);
207 : : static void heap_page_fix_vm_corruption(PruneState *prstate,
208 : : OffsetNumber offnum,
209 : : VMCorruptionType corruption_type);
210 : : static void prune_freeze_fast_path(PruneState *prstate,
211 : : PruneFreezeResult *presult);
212 : : static void prune_freeze_plan(PruneState *prstate,
213 : : OffsetNumber *off_loc);
214 : : static HTSV_Result heap_prune_satisfies_vacuum(PruneState *prstate,
215 : : HeapTuple tup);
216 : : static inline HTSV_Result htsv_get_valid_status(int status);
217 : : static void heap_prune_chain(OffsetNumber maxoff,
218 : : OffsetNumber rootoffnum, PruneState *prstate);
219 : : static void heap_prune_record_prunable(PruneState *prstate, TransactionId xid,
220 : : OffsetNumber offnum);
221 : : static void heap_prune_record_redirect(PruneState *prstate,
222 : : OffsetNumber offnum, OffsetNumber rdoffnum,
223 : : bool was_normal);
224 : : static void heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
225 : : bool was_normal);
226 : : static void heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
227 : : bool was_normal);
228 : : static void heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal);
229 : :
230 : : static void heap_prune_record_unchanged_lp_unused(PruneState *prstate, OffsetNumber offnum);
231 : : static void heap_prune_record_unchanged_lp_normal(PruneState *prstate, OffsetNumber offnum);
232 : : static void heap_prune_record_unchanged_lp_dead(PruneState *prstate, OffsetNumber offnum);
233 : : static void heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum);
234 : :
235 : : static void page_verify_redirects(Page page);
236 : :
237 : : static bool heap_page_will_freeze(bool did_tuple_hint_fpi, bool do_prune, bool do_hint_prune,
238 : : PruneState *prstate);
239 : : static bool heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
240 : : bool do_prune, bool do_freeze);
241 : :
242 : :
243 : : /*
244 : : * Optionally prune and repair fragmentation in the specified page.
245 : : *
246 : : * This is an opportunistic function. It will perform housekeeping
247 : : * only if the page heuristically looks like a candidate for pruning and we
248 : : * can acquire buffer cleanup lock without blocking.
249 : : *
250 : : * Note: this is called quite often. It's important that it fall out quickly
251 : : * if there's not any use in pruning.
252 : : *
253 : : * Caller must have pin on the buffer, and must *not* have a lock on it.
254 : : *
255 : : * This function may pin *vmbuffer. It's passed by reference so the caller can
256 : : * reuse the pin across calls, avoiding repeated pin/unpin cycles. If we find
257 : : * VM corruption during pruning, we will fix it. Caller is responsible for
258 : : * unpinning *vmbuffer.
259 : : *
260 : : * rel_read_only is true if we determined at plan time that the query does not
261 : : * modify the relation. It is counterproductive to set the VM if the query
262 : : * will immediately clear it.
263 : : *
264 : : * As noted in ScanRelIsReadOnly(), INSERT ... SELECT from the same table will
265 : : * report the scan relation as read-only. This is usually harmless in
266 : : * practice. It is useful to set scanned pages all-visible that won't be
267 : : * inserted into. Pages it does insert to will rarely meet the criteria for
268 : : * pruning, and those that do are likely to contain in-progress inserts which
269 : : * make the page not fully all-visible.
270 : : */
271 : : void
272 : 19817009 : heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer,
273 : : bool rel_read_only)
274 : : {
275 : 19817009 : Page page = BufferGetPage(buffer);
276 : : TransactionId prune_xid;
277 : : GlobalVisState *vistest;
278 : : Size minfree;
279 : :
280 : : /*
281 : : * We can't write WAL in recovery mode, so there's no point trying to
282 : : * clean the page. The primary will likely issue a cleaning WAL record
283 : : * soon anyway, so this is no particular loss.
284 : : */
285 [ + + ]: 19817009 : if (RecoveryInProgress())
286 : 264336 : return;
287 : :
288 : : /*
289 : : * First check whether there's any chance there's something to prune,
290 : : * determining the appropriate horizon is a waste if there's no prune_xid
291 : : * (i.e. no updates/deletes left potentially dead tuples around and no
292 : : * inserts inserted new tuples that may be visible to all).
293 : : */
294 : 19552673 : prune_xid = PageGetPruneXid(page);
295 [ + + ]: 19552673 : if (!TransactionIdIsValid(prune_xid))
296 : 8188940 : return;
297 : :
298 : : /*
299 : : * Check whether prune_xid indicates that there may be dead rows that can
300 : : * be cleaned up.
301 : : */
302 : 11363733 : vistest = GlobalVisTestFor(relation);
303 : :
304 [ + + ]: 11363733 : if (!GlobalVisTestIsRemovableXid(vistest, prune_xid, true))
305 : 8952586 : return;
306 : :
307 : : /*
308 : : * We prune when a previous UPDATE failed to find enough space on the page
309 : : * for a new tuple version, or when free space falls below the relation's
310 : : * fill-factor target (but not less than 10%).
311 : : *
312 : : * Checking free space here is questionable since we aren't holding any
313 : : * lock on the buffer; in the worst case we could get a bogus answer. It's
314 : : * unlikely to be *seriously* wrong, though, since reading either pd_lower
315 : : * or pd_upper is probably atomic. Avoiding taking a lock seems more
316 : : * important than sometimes getting a wrong answer in what is after all
317 : : * just a heuristic estimate.
318 : : */
319 [ + + ]: 2411147 : minfree = RelationGetTargetPageFreeSpace(relation,
320 : : HEAP_DEFAULT_FILLFACTOR);
321 : 2411147 : minfree = Max(minfree, BLCKSZ / 10);
322 : :
323 [ + + + + ]: 2411147 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
324 : : {
325 : 135883 : bool record_free_space = false;
326 : 135883 : Size freespace = 0;
327 : :
328 : : /*
329 : : * Pin the VM page before taking the heap cleanup lock. This may
330 : : * occasionally lead to an unnecessary pin when the buffer is
331 : : * contended, but the same VM page covers many heap pages, so there is
332 : : * a good chance for the work to be reusable.
333 : : */
334 : 135883 : visibilitymap_pin(relation, BufferGetBlockNumber(buffer), vmbuffer);
335 : :
336 : : /* OK, try to get exclusive buffer lock */
337 [ + + ]: 135883 : if (!ConditionalLockBufferForCleanup(buffer))
338 : 2038 : return;
339 : :
340 : : /*
341 : : * Now that we have buffer lock, get accurate information about the
342 : : * page's free space, and recheck the heuristic about whether to
343 : : * prune.
344 : : */
345 [ + + + + ]: 133845 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
346 : : {
347 : : OffsetNumber dummy_off_loc;
348 : : PruneFreezeResult presult;
349 : : PruneFreezeParams params;
350 : :
351 : 133844 : params.relation = relation;
352 : 133844 : params.buffer = buffer;
353 : 133844 : params.vmbuffer = *vmbuffer;
354 : 133844 : params.reason = PRUNE_ON_ACCESS;
355 : 133844 : params.vistest = vistest;
356 : 133844 : params.cutoffs = NULL;
357 : :
358 : : /*
359 : : * We don't pass the HEAP_PAGE_PRUNE_MARK_UNUSED_NOW option
360 : : * regardless of whether or not the relation has indexes, since we
361 : : * cannot safely determine that during on-access pruning with the
362 : : * current implementation.
363 : : */
364 : 133844 : params.options = HEAP_PAGE_PRUNE_ALLOW_FAST_PATH;
365 [ + + ]: 133844 : if (rel_read_only)
366 : 37786 : params.options |= HEAP_PAGE_PRUNE_SET_VM;
367 : :
368 : 133844 : heap_page_prune_and_freeze(¶ms, &presult, &dummy_off_loc,
369 : : NULL, NULL);
370 : :
371 : : /*
372 : : * Report the number of tuples reclaimed to pgstats. This is
373 : : * presult.ndeleted minus the number of newly-LP_DEAD-set items.
374 : : *
375 : : * We derive the number of dead tuples like this to avoid totally
376 : : * forgetting about items that were set to LP_DEAD, since they
377 : : * still need to be cleaned up by VACUUM. We only want to count
378 : : * heap-only tuples that just became LP_UNUSED in our report,
379 : : * which don't.
380 : : *
381 : : * VACUUM doesn't have to compensate in the same way when it
382 : : * tracks ndeleted, since it will set the same LP_DEAD items to
383 : : * LP_UNUSED separately.
384 : : */
385 [ + + ]: 133844 : if (presult.ndeleted > presult.nnewlpdead)
386 : 20492 : pgstat_update_heap_dead_tuples(relation,
387 : 20492 : presult.ndeleted - presult.nnewlpdead);
388 : :
389 : : /*
390 : : * If this prune newly set the page all-visible, VACUUM may later
391 : : * skip the page and not update the free space map (FSM) for it.
392 : : * Keep the FSM from going stale by recording it now. We do not
393 : : * want to update the freespace map otherwise, to reserve
394 : : * freespace on this page for HOT updates.
395 : : */
396 [ + + ]: 133844 : if (presult.newly_all_visible)
397 : : {
398 : 13908 : record_free_space = true;
399 : 13908 : freespace = PageGetHeapFreeSpace(page);
400 : : }
401 : : }
402 : :
403 : : /* And release buffer lock */
404 : 133845 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
405 : :
406 : : /*
407 : : * RecordPageWithFreeSpace() only dirties the FSM when the recorded
408 : : * free-space category actually changes. Note that vacuum will still
409 : : * do FreeSpaceMapVacuum() for ranges of pages that are skipped, so we
410 : : * don't have to worry about that here.
411 : : */
412 [ + + ]: 133845 : if (record_free_space)
413 : 13908 : RecordPageWithFreeSpace(relation, BufferGetBlockNumber(buffer), freespace);
414 : : }
415 : : }
416 : :
417 : : /*
418 : : * Helper for heap_page_prune_and_freeze() to initialize the PruneState using
419 : : * the provided parameters.
420 : : *
421 : : * params, new_relfrozen_xid, new_relmin_mxid, and presult are input
422 : : * parameters and are not modified by this function. Only prstate is modified.
423 : : */
424 : : static void
425 : 628666 : prune_freeze_setup(PruneFreezeParams *params,
426 : : TransactionId *new_relfrozen_xid,
427 : : MultiXactId *new_relmin_mxid,
428 : : PruneFreezeResult *presult,
429 : : PruneState *prstate)
430 : : {
431 : : /* Copy parameters to prstate */
432 : 628666 : prstate->vistest = params->vistest;
433 : 628666 : prstate->mark_unused_now =
434 : 628666 : (params->options & HEAP_PAGE_PRUNE_MARK_UNUSED_NOW) != 0;
435 : :
436 : : /* cutoffs must be provided if we will attempt freezing */
437 : : Assert(!(params->options & HEAP_PAGE_PRUNE_FREEZE) || params->cutoffs);
438 : 628666 : prstate->attempt_freeze = (params->options & HEAP_PAGE_PRUNE_FREEZE) != 0;
439 : 628666 : prstate->attempt_set_vm = (params->options & HEAP_PAGE_PRUNE_SET_VM) != 0;
440 : 628666 : prstate->cutoffs = params->cutoffs;
441 : 628666 : prstate->relation = params->relation;
442 : 628666 : prstate->block = BufferGetBlockNumber(params->buffer);
443 : 628666 : prstate->buffer = params->buffer;
444 : 628666 : prstate->page = BufferGetPage(params->buffer);
445 : :
446 : : Assert(BufferIsValid(params->vmbuffer));
447 : 628666 : prstate->vmbuffer = params->vmbuffer;
448 : 628666 : prstate->new_vmbits = 0;
449 : 628666 : prstate->old_vmbits = visibilitymap_get_status(prstate->relation,
450 : : prstate->block,
451 : : &prstate->vmbuffer);
452 : :
453 : : /*
454 : : * Our strategy is to scan the page and make lists of items to change,
455 : : * then apply the changes within a critical section. This keeps as much
456 : : * logic as possible out of the critical section, and also ensures that
457 : : * WAL replay will work the same as the normal case.
458 : : *
459 : : * First, initialize the new pd_prune_xid value to zero (indicating no
460 : : * prunable tuples). If we find any tuples which may soon become
461 : : * prunable, we will save the lowest relevant XID in new_prune_xid. Also
462 : : * initialize the rest of our working state.
463 : : */
464 : 628666 : prstate->new_prune_xid = InvalidTransactionId;
465 : 628666 : prstate->latest_xid_removed = InvalidTransactionId;
466 : 628666 : prstate->nredirected = prstate->ndead = prstate->nunused = 0;
467 : 628666 : prstate->nfrozen = 0;
468 : 628666 : prstate->nroot_items = 0;
469 : 628666 : prstate->nheaponly_items = 0;
470 : :
471 : : /* initialize page freezing working state */
472 : 628666 : prstate->pagefrz.freeze_required = false;
473 : 628666 : prstate->pagefrz.FreezePageConflictXid = InvalidTransactionId;
474 [ + + ]: 628666 : if (prstate->attempt_freeze)
475 : : {
476 : : Assert(new_relfrozen_xid && new_relmin_mxid);
477 : 494822 : prstate->pagefrz.FreezePageRelfrozenXid = *new_relfrozen_xid;
478 : 494822 : prstate->pagefrz.NoFreezePageRelfrozenXid = *new_relfrozen_xid;
479 : 494822 : prstate->pagefrz.FreezePageRelminMxid = *new_relmin_mxid;
480 : 494822 : prstate->pagefrz.NoFreezePageRelminMxid = *new_relmin_mxid;
481 : : }
482 : : else
483 : : {
484 : : Assert(!new_relfrozen_xid && !new_relmin_mxid);
485 : 133844 : prstate->pagefrz.FreezePageRelminMxid = InvalidMultiXactId;
486 : 133844 : prstate->pagefrz.NoFreezePageRelminMxid = InvalidMultiXactId;
487 : 133844 : prstate->pagefrz.FreezePageRelfrozenXid = InvalidTransactionId;
488 : 133844 : prstate->pagefrz.NoFreezePageRelfrozenXid = InvalidTransactionId;
489 : : }
490 : :
491 : 628666 : prstate->ndeleted = 0;
492 : 628666 : prstate->live_tuples = 0;
493 : 628666 : prstate->recently_dead_tuples = 0;
494 : 628666 : prstate->hastup = false;
495 : 628666 : prstate->lpdead_items = 0;
496 : :
497 : : /*
498 : : * deadoffsets are filled in during pruning but are only used to populate
499 : : * PruneFreezeResult->deadoffsets. To avoid needing two copies of the
500 : : * array, just save a pointer to the result offsets array in the
501 : : * PruneState.
502 : : */
503 : 628666 : prstate->deadoffsets = presult->deadoffsets;
504 : :
505 : : /*
506 : : * We track whether the page will be all-visible/all-frozen at the end of
507 : : * pruning and freezing. While examining tuple visibility, we'll set
508 : : * set_all_visible to false if there are tuples on the page not visible to
509 : : * all running and future transactions. If setting the VM is enabled for
510 : : * this scan, we will do so if the page ends up being all-visible.
511 : : *
512 : : * We also keep track of the newest live XID, which is used to calculate
513 : : * the snapshot conflict horizon for a WAL record setting the VM.
514 : : */
515 : 628666 : prstate->set_all_visible = prstate->attempt_set_vm;
516 : 628666 : prstate->newest_live_xid = InvalidTransactionId;
517 : :
518 : : /*
519 : : * Currently, only VACUUM performs freezing, but other callers may in the
520 : : * future. We must initialize set_all_frozen based on whether or not the
521 : : * caller passed HEAP_PAGE_PRUNE_FREEZE, because if they did not, we won't
522 : : * call heap_prepare_freeze_tuple() for each tuple, and set_all_frozen
523 : : * will never be cleared for tuples that need freezing. This would lead to
524 : : * incorrectly setting the visibility map all-frozen for this page. We
525 : : * can't set the page all-frozen in the VM if the caller didn't pass
526 : : * HEAP_PAGE_PRUNE_SET_VM.
527 : : *
528 : : * When freezing is not required (no XIDs/MXIDs older than the freeze
529 : : * cutoff), we may still choose to "opportunistically" freeze if doing so
530 : : * would make the page all-frozen.
531 : : *
532 : : * We will not be able to freeze the whole page at the end of vacuum if
533 : : * there are tuples present that are not visible to everyone or if there
534 : : * are dead tuples which will not be removable. However, dead tuples that
535 : : * will be removed by the end of vacuum should not prevent this
536 : : * opportunistic freezing.
537 : : *
538 : : * Therefore, we do not clear set_all_visible and set_all_frozen when we
539 : : * encounter LP_DEAD items. Instead, we correct them after deciding
540 : : * whether to freeze, but before updating the VM, to avoid setting the VM
541 : : * bits incorrectly.
542 : : */
543 [ + + + - ]: 628666 : prstate->set_all_frozen = prstate->attempt_freeze && prstate->attempt_set_vm;
544 : 628666 : }
545 : :
546 : : /*
547 : : * Helper for heap_page_prune_and_freeze(). Iterates over every tuple on the
548 : : * page, examines its visibility information, and determines the appropriate
549 : : * action for each tuple. All tuples are processed and classified during this
550 : : * phase, but no modifications are made to the page until the later execution
551 : : * stage.
552 : : *
553 : : * *off_loc is used for error callback and cleared before returning.
554 : : */
555 : : static void
556 : 431959 : prune_freeze_plan(PruneState *prstate, OffsetNumber *off_loc)
557 : : {
558 : 431959 : Page page = prstate->page;
559 : 431959 : BlockNumber blockno = prstate->block;
560 : 431959 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
561 : : OffsetNumber offnum;
562 : : HeapTupleData tup;
563 : :
564 : 431959 : tup.t_tableOid = RelationGetRelid(prstate->relation);
565 : :
566 : : /*
567 : : * Determine HTSV for all tuples, and queue them up for processing as HOT
568 : : * chain roots or as heap-only items.
569 : : *
570 : : * Determining HTSV only once for each tuple is required for correctness,
571 : : * to deal with cases where running HTSV twice could result in different
572 : : * results. For example, RECENTLY_DEAD can turn to DEAD if another
573 : : * checked item causes GlobalVisTestIsRemovableFullXid() to update the
574 : : * horizon, or INSERT_IN_PROGRESS can change to DEAD if the inserting
575 : : * transaction aborts.
576 : : *
577 : : * It's also good for performance. Most commonly tuples within a page are
578 : : * stored at decreasing offsets (while the items are stored at increasing
579 : : * offsets). When processing all tuples on a page this leads to reading
580 : : * memory at decreasing offsets within a page, with a variable stride.
581 : : * That's hard for CPU prefetchers to deal with. Processing the items in
582 : : * reverse order (and thus the tuples in increasing order) increases
583 : : * prefetching efficiency significantly / decreases the number of cache
584 : : * misses.
585 : : */
586 : 431959 : for (offnum = maxoff;
587 [ + + ]: 29496444 : offnum >= FirstOffsetNumber;
588 : 29064485 : offnum = OffsetNumberPrev(offnum))
589 : : {
590 : 29064485 : ItemId itemid = PageGetItemId(page, offnum);
591 : : HeapTupleHeader htup;
592 : :
593 : : /*
594 : : * Set the offset number so that we can display it along with any
595 : : * error that occurred while processing this tuple.
596 : : */
597 : 29064485 : *off_loc = offnum;
598 : :
599 : 29064485 : prstate->processed[offnum] = false;
600 : 29064485 : prstate->htsv[offnum] = -1;
601 : :
602 : : /* Nothing to do if slot doesn't contain a tuple */
603 [ + + ]: 29064485 : if (!ItemIdIsUsed(itemid))
604 : : {
605 : 240419 : heap_prune_record_unchanged_lp_unused(prstate, offnum);
606 : 240419 : continue;
607 : : }
608 : :
609 [ + + ]: 28824066 : if (ItemIdIsDead(itemid))
610 : : {
611 : : /*
612 : : * If the caller set mark_unused_now true, we can set dead line
613 : : * pointers LP_UNUSED now.
614 : : */
615 [ + + ]: 1735019 : if (unlikely(prstate->mark_unused_now))
616 : 1866 : heap_prune_record_unused(prstate, offnum, false);
617 : : else
618 : 1733153 : heap_prune_record_unchanged_lp_dead(prstate, offnum);
619 : 1735019 : continue;
620 : : }
621 : :
622 [ + + ]: 27089047 : if (ItemIdIsRedirected(itemid))
623 : : {
624 : : /* This is the start of a HOT chain */
625 : 198038 : prstate->root_items[prstate->nroot_items++] = offnum;
626 : 198038 : continue;
627 : : }
628 : :
629 : : Assert(ItemIdIsNormal(itemid));
630 : :
631 : : /*
632 : : * Get the tuple's visibility status and queue it up for processing.
633 : : */
634 : 26891009 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
635 : 26891009 : tup.t_data = htup;
636 : 26891009 : tup.t_len = ItemIdGetLength(itemid);
637 : 26891009 : ItemPointerSet(&tup.t_self, blockno, offnum);
638 : :
639 : 26891009 : prstate->htsv[offnum] = heap_prune_satisfies_vacuum(prstate, &tup);
640 : :
641 [ + + ]: 26891009 : if (!HeapTupleHeaderIsHeapOnly(htup))
642 : 26551964 : prstate->root_items[prstate->nroot_items++] = offnum;
643 : : else
644 : 339045 : prstate->heaponly_items[prstate->nheaponly_items++] = offnum;
645 : : }
646 : :
647 : : /*
648 : : * Process HOT chains.
649 : : *
650 : : * We added the items to the array starting from 'maxoff', so by
651 : : * processing the array in reverse order, we process the items in
652 : : * ascending offset number order. The order doesn't matter for
653 : : * correctness, but some quick micro-benchmarking suggests that this is
654 : : * faster. (Earlier PostgreSQL versions, which scanned all the items on
655 : : * the page instead of using the root_items array, also did it in
656 : : * ascending offset number order.)
657 : : */
658 [ + + ]: 27181961 : for (int i = prstate->nroot_items - 1; i >= 0; i--)
659 : : {
660 : 26750002 : offnum = prstate->root_items[i];
661 : :
662 : : /* Ignore items already processed as part of an earlier chain */
663 [ - + ]: 26750002 : if (prstate->processed[offnum])
664 : 0 : continue;
665 : :
666 : : /* see preceding loop */
667 : 26750002 : *off_loc = offnum;
668 : :
669 : : /* Process this item or chain of items */
670 : 26750002 : heap_prune_chain(maxoff, offnum, prstate);
671 : : }
672 : :
673 : : /*
674 : : * Process any heap-only tuples that were not already processed as part of
675 : : * a HOT chain.
676 : : */
677 [ + + ]: 771004 : for (int i = prstate->nheaponly_items - 1; i >= 0; i--)
678 : : {
679 : 339045 : offnum = prstate->heaponly_items[i];
680 : :
681 [ + + ]: 339045 : if (prstate->processed[offnum])
682 : 322477 : continue;
683 : :
684 : : /* see preceding loop */
685 : 16568 : *off_loc = offnum;
686 : :
687 : : /*
688 : : * If the tuple is DEAD and doesn't chain to anything else, mark it
689 : : * unused. (If it does chain, we can only remove it as part of
690 : : * pruning its chain.)
691 : : *
692 : : * We need this primarily to handle aborted HOT updates, that is,
693 : : * XMIN_INVALID heap-only tuples. Those might not be linked to by any
694 : : * chain, since the parent tuple might be re-updated before any
695 : : * pruning occurs. So we have to be able to reap them separately from
696 : : * chain-pruning. (Note that HeapTupleHeaderIsHotUpdated will never
697 : : * return true for an XMIN_INVALID tuple, so this code will work even
698 : : * when there were sequential updates within the aborted transaction.)
699 : : */
700 [ + + ]: 16568 : if (prstate->htsv[offnum] == HEAPTUPLE_DEAD)
701 : : {
702 : 3277 : ItemId itemid = PageGetItemId(page, offnum);
703 : 3277 : HeapTupleHeader htup = (HeapTupleHeader) PageGetItem(page, itemid);
704 : :
705 [ + - ]: 3277 : if (likely(!HeapTupleHeaderIsHotUpdated(htup)))
706 : : {
707 : 3277 : HeapTupleHeaderAdvanceConflictHorizon(htup,
708 : : &prstate->latest_xid_removed);
709 : 3277 : heap_prune_record_unused(prstate, offnum, true);
710 : : }
711 : : else
712 : : {
713 : : /*
714 : : * This tuple should've been processed and removed as part of
715 : : * a HOT chain, so something's wrong. To preserve evidence,
716 : : * we don't dare to remove it. We cannot leave behind a DEAD
717 : : * tuple either, because that will cause VACUUM to error out.
718 : : * Throwing an error with a distinct error message seems like
719 : : * the least bad option.
720 : : */
721 [ # # ]: 0 : elog(ERROR, "dead heap-only tuple (%u, %d) is not linked to from any HOT chain",
722 : : blockno, offnum);
723 : : }
724 : : }
725 : : else
726 : 13291 : heap_prune_record_unchanged_lp_normal(prstate, offnum);
727 : : }
728 : :
729 : : /* We should now have processed every tuple exactly once */
730 : : #ifdef USE_ASSERT_CHECKING
731 : : for (offnum = FirstOffsetNumber;
732 : : offnum <= maxoff;
733 : : offnum = OffsetNumberNext(offnum))
734 : : {
735 : : *off_loc = offnum;
736 : :
737 : : Assert(prstate->processed[offnum]);
738 : : }
739 : : #endif
740 : :
741 : : /* Clear the offset information once we have processed the given page. */
742 : 431959 : *off_loc = InvalidOffsetNumber;
743 : 431959 : }
744 : :
745 : : /*
746 : : * Decide whether to proceed with freezing according to the freeze plans
747 : : * prepared for the current heap buffer. If freezing is chosen, this function
748 : : * performs several pre-freeze checks.
749 : : *
750 : : * The values of do_prune, do_hint_prune, and did_tuple_hint_fpi must be
751 : : * determined before calling this function.
752 : : *
753 : : * prstate is both an input and output parameter.
754 : : *
755 : : * Returns true if we should apply the freeze plans and freeze tuples on the
756 : : * page, and false otherwise.
757 : : */
758 : : static bool
759 : 431959 : heap_page_will_freeze(bool did_tuple_hint_fpi,
760 : : bool do_prune,
761 : : bool do_hint_prune,
762 : : PruneState *prstate)
763 : : {
764 : 431959 : bool do_freeze = false;
765 : :
766 : : /*
767 : : * If the caller specified we should not attempt to freeze any tuples,
768 : : * validate that everything is in the right state and return.
769 : : */
770 [ + + ]: 431959 : if (!prstate->attempt_freeze)
771 : : {
772 : : Assert(!prstate->set_all_frozen && prstate->nfrozen == 0);
773 : 133746 : return false;
774 : : }
775 : :
776 [ + + ]: 298213 : if (prstate->pagefrz.freeze_required)
777 : : {
778 : : /*
779 : : * heap_prepare_freeze_tuple indicated that at least one XID/MXID from
780 : : * before FreezeLimit/MultiXactCutoff is present. Must freeze to
781 : : * advance relfrozenxid/relminmxid.
782 : : */
783 : 23096 : do_freeze = true;
784 : : }
785 : : else
786 : : {
787 : : /*
788 : : * Opportunistically freeze the page if we are generating an FPI
789 : : * anyway and if doing so means that we can set the page all-frozen
790 : : * afterwards (might not happen until VACUUM's final heap pass).
791 : : *
792 : : * XXX: Previously, we knew if pruning emitted an FPI by checking
793 : : * pgWalUsage.wal_fpi before and after pruning. Once the freeze and
794 : : * prune records were combined, this heuristic couldn't be used
795 : : * anymore. The opportunistic freeze heuristic must be improved;
796 : : * however, for now, try to approximate the old logic.
797 : : */
798 [ + + + + ]: 275117 : if (prstate->set_all_frozen && prstate->nfrozen > 0)
799 : : {
800 : : Assert(prstate->set_all_visible);
801 : :
802 : : /*
803 : : * Freezing would make the page all-frozen. Have already emitted
804 : : * an FPI or will do so anyway?
805 : : */
806 [ + + + + : 24129 : if (RelationNeedsWAL(prstate->relation))
+ - + - ]
807 : : {
808 [ + + ]: 22088 : if (did_tuple_hint_fpi)
809 : 1250 : do_freeze = true;
810 [ + + ]: 20838 : else if (do_prune)
811 : : {
812 [ + + ]: 2302 : if (XLogCheckBufferNeedsBackup(prstate->buffer))
813 : 766 : do_freeze = true;
814 : : }
815 [ + + ]: 18536 : else if (do_hint_prune)
816 : : {
817 [ + + + - : 22514 : if (XLogHintBitIsNeeded() &&
+ + ]
818 : 11257 : XLogCheckBufferNeedsBackup(prstate->buffer))
819 : 2027 : do_freeze = true;
820 : : }
821 : : }
822 : : }
823 : : }
824 : :
825 [ + + ]: 298213 : if (do_freeze)
826 : : {
827 : : /*
828 : : * Validate the tuples we will be freezing before entering the
829 : : * critical section.
830 : : */
831 : 27139 : heap_pre_freeze_checks(prstate->buffer, prstate->frozen, prstate->nfrozen);
832 : : Assert(TransactionIdPrecedes(prstate->pagefrz.FreezePageConflictXid,
833 : : prstate->cutoffs->OldestXmin));
834 : : }
835 [ + + ]: 271074 : else if (prstate->nfrozen > 0)
836 : : {
837 : : /*
838 : : * The page contained some tuples that were not already frozen, and we
839 : : * chose not to freeze them now. The page won't be all-frozen then.
840 : : */
841 : : Assert(!prstate->pagefrz.freeze_required);
842 : :
843 : 20951 : prstate->set_all_frozen = false;
844 : 20951 : prstate->nfrozen = 0; /* avoid miscounts in instrumentation */
845 : : }
846 : : else
847 : : {
848 : : /*
849 : : * We have no freeze plans to execute. The page might already be
850 : : * all-frozen (perhaps only following pruning), though. Such pages
851 : : * can be marked all-frozen in the VM by our caller, even though none
852 : : * of its tuples were newly frozen here.
853 : : */
854 : : }
855 : :
856 : 298213 : return do_freeze;
857 : : }
858 : :
859 : : /*
860 : : * Emit a warning about and fix visibility map corruption on the given page.
861 : : *
862 : : * The caller specifies the type of corruption it has already detected via
863 : : * corruption_type, so that we can emit the appropriate warning. All cases
864 : : * result in the VM bits being cleared; corruption types where PD_ALL_VISIBLE
865 : : * is incorrectly set also clear PD_ALL_VISIBLE.
866 : : *
867 : : * Must be called while holding an exclusive lock on the heap buffer. Dead
868 : : * items and not all-visible tuples must have been discovered under that same
869 : : * lock. Although we do not hold a lock on the VM buffer, it is pinned, and
870 : : * the heap buffer is exclusively locked, ensuring that no other backend can
871 : : * update the VM bits corresponding to this heap page.
872 : : *
873 : : * This function makes changes to the VM and, potentially, the heap page, but
874 : : * it does not need to be done in a critical section.
875 : : */
876 : : static void
877 : 0 : heap_page_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum,
878 : : VMCorruptionType corruption_type)
879 : : {
880 : 0 : const char *relname = RelationGetRelationName(prstate->relation);
881 : 0 : bool do_clear_vm = false;
882 : 0 : bool do_clear_heap = false;
883 : :
884 : : Assert(BufferIsLockedByMeInMode(prstate->buffer, BUFFER_LOCK_EXCLUSIVE));
885 : :
886 [ # # # # ]: 0 : switch (corruption_type)
887 : : {
888 : 0 : case VM_CORRUPT_LPDEAD:
889 [ # # ]: 0 : ereport(WARNING,
890 : : (errcode(ERRCODE_DATA_CORRUPTED),
891 : : errmsg("dead line pointer found on page marked all-visible"),
892 : : errcontext("relation \"%s\", page %u, tuple %u",
893 : : relname, prstate->block, offnum)));
894 : 0 : do_clear_vm = true;
895 : 0 : do_clear_heap = true;
896 : 0 : break;
897 : :
898 : 0 : case VM_CORRUPT_TUPLE_VISIBILITY:
899 : :
900 : : /*
901 : : * A HEAPTUPLE_LIVE tuple on an all-visible page can appear to not
902 : : * be visible to everyone when
903 : : * GetOldestNonRemovableTransactionId() returns a conservative
904 : : * value that's older than the real safe xmin. That is not
905 : : * corruption -- the PD_ALL_VISIBLE flag is still correct.
906 : : *
907 : : * However, dead tuple versions, in-progress inserts, and
908 : : * in-progress deletes should never appear on a page marked
909 : : * all-visible. That indicates real corruption. PD_ALL_VISIBLE
910 : : * should have been cleared by the DML operation that deleted or
911 : : * inserted the tuple.
912 : : */
913 [ # # ]: 0 : ereport(WARNING,
914 : : (errcode(ERRCODE_DATA_CORRUPTED),
915 : : errmsg("tuple not visible to all transactions found on page marked all-visible"),
916 : : errcontext("relation \"%s\", page %u, tuple %u",
917 : : relname, prstate->block, offnum)));
918 : 0 : do_clear_vm = true;
919 : 0 : do_clear_heap = true;
920 : 0 : break;
921 : :
922 : 0 : case VM_CORRUPT_MISSING_PAGE_HINT:
923 : :
924 : : /*
925 : : * As of PostgreSQL 9.2, the visibility map bit should never be
926 : : * set if the page-level bit is clear. However, for vacuum, it's
927 : : * possible that the bit got cleared after
928 : : * heap_vac_scan_next_block() was called, so we must recheck now
929 : : * that we have the buffer lock before concluding that the VM is
930 : : * corrupt.
931 : : */
932 : : Assert(!PageIsAllVisible(prstate->page));
933 : : Assert(prstate->old_vmbits & VISIBILITYMAP_VALID_BITS);
934 [ # # ]: 0 : ereport(WARNING,
935 : : (errcode(ERRCODE_DATA_CORRUPTED),
936 : : errmsg("page is not marked all-visible but visibility map bit is set"),
937 : : errcontext("relation \"%s\", page %u",
938 : : relname, prstate->block)));
939 : 0 : do_clear_vm = true;
940 : 0 : break;
941 : : }
942 : :
943 : : Assert(do_clear_heap || do_clear_vm);
944 : :
945 : : /* Avoid marking the buffer dirty if PD_ALL_VISIBLE is already clear */
946 [ # # ]: 0 : if (do_clear_heap)
947 : : {
948 : : Assert(PageIsAllVisible(prstate->page));
949 : 0 : PageClearAllVisible(prstate->page);
950 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
951 : : }
952 : :
953 [ # # ]: 0 : if (do_clear_vm)
954 : : {
955 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_EXCLUSIVE);
956 : : /* This VM clear is not WAL-logged, so its return value is not needed. */
957 : 0 : (void) visibilitymap_clear(prstate->relation->rd_locator,
958 : : prstate->block, prstate->vmbuffer,
959 : : VISIBILITYMAP_VALID_BITS);
960 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_UNLOCK);
961 : 0 : prstate->old_vmbits = 0;
962 : : }
963 : 0 : }
964 : :
965 : : /*
966 : : * Decide whether to set the visibility map bits (all-visible and all-frozen)
967 : : * for the current page using information from the PruneState and VM.
968 : : *
969 : : * This function does not actually set the VM bits or page-level visibility
970 : : * hint, PD_ALL_VISIBLE.
971 : : *
972 : : * This should be called only after do_freeze has been decided (and do_prune
973 : : * has been set), as these factor into our heuristic-based decision.
974 : : *
975 : : * Returns true if one or both VM bits should be set and false otherwise.
976 : : */
977 : : static bool
978 : 431959 : heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
979 : : bool do_prune, bool do_freeze)
980 : : {
981 [ + + ]: 431959 : if (!prstate->attempt_set_vm)
982 : 96058 : return false;
983 : :
984 [ + + ]: 335901 : if (!prstate->set_all_visible)
985 : 256420 : return false;
986 : :
987 : : /*
988 : : * If this is an on-access call and we're not actually pruning, avoid
989 : : * setting the visibility map if it would newly dirty the heap page or, if
990 : : * the page is already dirty, if doing so would require including a
991 : : * full-page image (FPI) of the heap page in the WAL.
992 : : */
993 [ + + + + : 79481 : if (reason == PRUNE_ON_ACCESS && !do_prune && !do_freeze &&
+ - ]
994 [ + + + + ]: 29498 : (!BufferIsDirty(prstate->buffer) || XLogCheckBufferNeedsBackup(prstate->buffer)))
995 : : {
996 : 15687 : prstate->set_all_visible = prstate->set_all_frozen = false;
997 : 15687 : return false;
998 : : }
999 : :
1000 : 63794 : prstate->new_vmbits = VISIBILITYMAP_ALL_VISIBLE;
1001 : :
1002 [ + + ]: 63794 : if (prstate->set_all_frozen)
1003 : 34050 : prstate->new_vmbits |= VISIBILITYMAP_ALL_FROZEN;
1004 : :
1005 [ + + ]: 63794 : if (prstate->new_vmbits == prstate->old_vmbits)
1006 : : {
1007 : 1729 : prstate->new_vmbits = 0;
1008 : 1729 : return false;
1009 : : }
1010 : :
1011 : 62065 : return true;
1012 : : }
1013 : :
1014 : : /*
1015 : : * If the page is already all-frozen, or already all-visible and freezing
1016 : : * won't be attempted, there is no remaining work and we can use the fast path
1017 : : * to avoid the expensive overhead of heap_page_prune_and_freeze().
1018 : : *
1019 : : * This can happen when the page has a stale prune hint, or if VACUUM is
1020 : : * scanning an already all-frozen page due to SKIP_PAGES_THRESHOLD.
1021 : : *
1022 : : * The caller must already have examined the visibility map and saved the
1023 : : * status of the page's VM bits in prstate->old_vmbits. Caller must hold a
1024 : : * content lock on the heap page since it will examine line pointers.
1025 : : *
1026 : : * Before calling prune_freeze_fast_path(), the caller should first
1027 : : * check for and fix any discrepancy between the page-level visibility hint
1028 : : * and the visibility map. Otherwise, the fast path will always prevent us
1029 : : * from getting them in sync. Note that if there are tuples on the page that
1030 : : * are not visible to all but the VM is incorrectly marked
1031 : : * all-visible/all-frozen, we will not get the chance to fix that corruption
1032 : : * when using the fast path.
1033 : : */
1034 : : static void
1035 : 196707 : prune_freeze_fast_path(PruneState *prstate, PruneFreezeResult *presult)
1036 : : {
1037 : 196707 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
1038 : 196707 : Page page = prstate->page;
1039 : :
1040 : : Assert((prstate->old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1041 : : ((prstate->old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1042 : : !prstate->attempt_freeze));
1043 : :
1044 : : /* We'll fill in presult for the caller */
1045 : 196707 : memset(presult, 0, sizeof(PruneFreezeResult));
1046 : :
1047 : : /* Clear any stale prune hint */
1048 [ + + ]: 196707 : if (TransactionIdIsValid(PageGetPruneXid(page)))
1049 : : {
1050 : 98 : PageClearPrunable(page);
1051 : 98 : MarkBufferDirtyHint(prstate->buffer, true);
1052 : : }
1053 : :
1054 [ - + ]: 196707 : if (PageIsEmpty(page))
1055 : 0 : return;
1056 : :
1057 : : /*
1058 : : * Since the page is all-visible, a count of the normal ItemIds on the
1059 : : * page should be sufficient for vacuum's live tuple count.
1060 : : */
1061 : 196707 : for (OffsetNumber off = FirstOffsetNumber;
1062 [ + + ]: 11279743 : off <= maxoff;
1063 : 11083036 : off = OffsetNumberNext(off))
1064 : : {
1065 : 11083036 : ItemId lp = PageGetItemId(page, off);
1066 : :
1067 [ + + ]: 11083036 : if (!ItemIdIsUsed(lp))
1068 : 262704 : continue;
1069 : :
1070 : 10820332 : presult->hastup = true;
1071 : :
1072 [ + + ]: 10820332 : if (ItemIdIsNormal(lp))
1073 : 10650565 : prstate->live_tuples++;
1074 : : }
1075 : :
1076 : 196707 : presult->live_tuples = prstate->live_tuples;
1077 : : }
1078 : :
1079 : : /*
1080 : : * Prune and repair fragmentation and potentially freeze tuples on the
1081 : : * specified page. If the page's visibility status has changed, update it in
1082 : : * the VM.
1083 : : *
1084 : : * Caller must have pin and buffer cleanup lock on the page. Note that we
1085 : : * don't update the FSM information for page on caller's behalf. Caller might
1086 : : * also need to account for a reduction in the length of the line pointer
1087 : : * array following array truncation by us.
1088 : : *
1089 : : * params contains the input parameters used to control freezing and pruning
1090 : : * behavior. See the definition of PruneFreezeParams for more on what each
1091 : : * parameter does.
1092 : : *
1093 : : * If the HEAP_PAGE_PRUNE_FREEZE option is set in params, we will freeze
1094 : : * tuples if it's required in order to advance relfrozenxid / relminmxid, or
1095 : : * if it's considered advantageous for overall system performance to do so
1096 : : * now. The 'params.cutoffs', 'presult', 'new_relfrozen_xid' and
1097 : : * 'new_relmin_mxid' arguments are required when freezing.
1098 : : *
1099 : : * A vmbuffer corresponding to the heap page is also passed and if the page is
1100 : : * found to be all-visible/all-frozen, we will set it in the VM.
1101 : : *
1102 : : * presult contains output parameters needed by callers, such as the number of
1103 : : * tuples removed and the offsets of dead items on the page after pruning.
1104 : : * heap_page_prune_and_freeze() is responsible for initializing it. Required
1105 : : * by all callers.
1106 : : *
1107 : : * off_loc is the offset location required by the caller to use in error
1108 : : * callback.
1109 : : *
1110 : : * new_relfrozen_xid and new_relmin_mxid must be provided by the caller if the
1111 : : * HEAP_PAGE_PRUNE_FREEZE option is set in params. On entry, they contain the
1112 : : * oldest XID and multi-XID seen on the relation so far. They will be updated
1113 : : * with the oldest values present on the page after pruning. After processing
1114 : : * the whole relation, VACUUM can use these values as the new
1115 : : * relfrozenxid/relminmxid for the relation.
1116 : : */
1117 : : void
1118 : 628666 : heap_page_prune_and_freeze(PruneFreezeParams *params,
1119 : : PruneFreezeResult *presult,
1120 : : OffsetNumber *off_loc,
1121 : : TransactionId *new_relfrozen_xid,
1122 : : MultiXactId *new_relmin_mxid)
1123 : : {
1124 : : PruneState prstate;
1125 : : bool do_freeze;
1126 : : bool do_prune;
1127 : : bool do_hint_prune;
1128 : : bool do_set_vm;
1129 : : bool did_tuple_hint_fpi;
1130 : 628666 : int64 fpi_before = pgWalUsage.wal_fpi;
1131 : : TransactionId conflict_xid;
1132 : :
1133 : : /* Initialize prstate */
1134 : 628666 : prune_freeze_setup(params,
1135 : : new_relfrozen_xid, new_relmin_mxid,
1136 : : presult, &prstate);
1137 : :
1138 : : /*
1139 : : * If the VM is set but PD_ALL_VISIBLE is clear, fix that corruption
1140 : : * before pruning and freezing so that the page and VM start out in a
1141 : : * consistent state.
1142 : : */
1143 [ + + ]: 628666 : if ((prstate.old_vmbits & VISIBILITYMAP_VALID_BITS) &&
1144 [ - + ]: 202006 : !PageIsAllVisible(prstate.page))
1145 : 0 : heap_page_fix_vm_corruption(&prstate, InvalidOffsetNumber,
1146 : : VM_CORRUPT_MISSING_PAGE_HINT);
1147 : :
1148 : : /*
1149 : : * If the page is already all-frozen, or already all-visible when freezing
1150 : : * is not being attempted, take the fast path, skipping pruning and
1151 : : * freezing code entirely. This must be done after fixing any discrepancy
1152 : : * between the page-level visibility hint and the VM, since that may have
1153 : : * cleared old_vmbits.
1154 : : */
1155 [ + + ]: 628666 : if ((params->options & HEAP_PAGE_PRUNE_ALLOW_FAST_PATH) != 0 &&
1156 [ + + ]: 627369 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1157 [ + + ]: 430662 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1158 [ - + ]: 4853 : !prstate.attempt_freeze)))
1159 : : {
1160 : 196707 : prune_freeze_fast_path(&prstate, presult);
1161 : 196707 : return;
1162 : : }
1163 : :
1164 : : /*
1165 : : * Examine all line pointers and tuple visibility information to determine
1166 : : * which line pointers should change state and which tuples may be frozen.
1167 : : * Prepare queue of state changes to later be executed in a critical
1168 : : * section.
1169 : : */
1170 : 431959 : prune_freeze_plan(&prstate, off_loc);
1171 : :
1172 : : /*
1173 : : * After processing all the live tuples on the page, if the newest xmin
1174 : : * amongst them may be considered running by any snapshot, the page cannot
1175 : : * be all-visible. This should be done before determining whether or not
1176 : : * to opportunistically freeze.
1177 : : */
1178 [ + + ]: 431959 : if (prstate.set_all_visible &&
1179 [ + + + + ]: 191900 : TransactionIdIsNormal(prstate.newest_live_xid) &&
1180 : 81775 : GlobalVisTestXidConsideredRunning(prstate.vistest,
1181 : : prstate.newest_live_xid,
1182 : : true))
1183 : 2853 : prstate.set_all_visible = prstate.set_all_frozen = false;
1184 : :
1185 : : /*
1186 : : * If checksums are enabled, calling heap_prune_satisfies_vacuum() while
1187 : : * checking tuple visibility information in prune_freeze_plan() may have
1188 : : * caused an FPI to be emitted.
1189 : : */
1190 : 431959 : did_tuple_hint_fpi = fpi_before != pgWalUsage.wal_fpi;
1191 : :
1192 : 1278219 : do_prune = prstate.nredirected > 0 ||
1193 [ + + + + ]: 792424 : prstate.ndead > 0 ||
1194 [ + + ]: 360465 : prstate.nunused > 0;
1195 : :
1196 : : /*
1197 : : * Even if we don't prune anything, if we found a new value for the
1198 : : * pd_prune_xid field or the page was marked full, we will update the hint
1199 : : * bit.
1200 : : */
1201 [ + + + + ]: 686660 : do_hint_prune = PageGetPruneXid(prstate.page) != prstate.new_prune_xid ||
1202 : 254701 : PageIsFull(prstate.page);
1203 : :
1204 : : /*
1205 : : * Decide if we want to go ahead with freezing according to the freeze
1206 : : * plans we prepared, or not.
1207 : : */
1208 : 431959 : do_freeze = heap_page_will_freeze(did_tuple_hint_fpi,
1209 : : do_prune,
1210 : : do_hint_prune,
1211 : : &prstate);
1212 : :
1213 : : /*
1214 : : * While scanning the line pointers, we did not clear
1215 : : * set_all_visible/set_all_frozen when encountering LP_DEAD items because
1216 : : * we wanted the decision whether or not to freeze the page to be
1217 : : * unaffected by the short-term presence of LP_DEAD items. These LP_DEAD
1218 : : * items are effectively assumed to be LP_UNUSED items in the making. It
1219 : : * doesn't matter which vacuum heap pass (initial pass or final pass) ends
1220 : : * up setting the page all-frozen, as long as the ongoing VACUUM does it.
1221 : : *
1222 : : * Now that we finished determining whether or not to freeze the page,
1223 : : * update set_all_visible and set_all_frozen so that they reflect the true
1224 : : * state of the page for setting PD_ALL_VISIBLE and VM bits.
1225 : : */
1226 [ + + ]: 431959 : if (prstate.lpdead_items > 0)
1227 : 75440 : prstate.set_all_visible = prstate.set_all_frozen = false;
1228 : :
1229 : : Assert(!prstate.set_all_frozen || prstate.set_all_visible);
1230 : : Assert(!prstate.set_all_visible || prstate.attempt_set_vm);
1231 : : Assert(!prstate.set_all_visible || (prstate.lpdead_items == 0));
1232 : :
1233 : 431959 : do_set_vm = heap_page_will_set_vm(&prstate, params->reason, do_prune, do_freeze);
1234 : :
1235 : : /*
1236 : : * new_vmbits should be 0 regardless of whether or not the page is
1237 : : * all-visible if we do not intend to set the VM.
1238 : : */
1239 : : Assert(do_set_vm || prstate.new_vmbits == 0);
1240 : :
1241 : : /*
1242 : : * The snapshot conflict horizon for the whole record is the most
1243 : : * conservative (newest) horizon required by any change in the record.
1244 : : */
1245 : 431959 : conflict_xid = InvalidTransactionId;
1246 [ + + ]: 431959 : if (do_set_vm)
1247 : 62065 : conflict_xid = prstate.newest_live_xid;
1248 [ + + + + ]: 431959 : if (do_freeze && TransactionIdFollows(prstate.pagefrz.FreezePageConflictXid, conflict_xid))
1249 : 4593 : conflict_xid = prstate.pagefrz.FreezePageConflictXid;
1250 [ + + + + ]: 431959 : if (do_prune && TransactionIdFollows(prstate.latest_xid_removed, conflict_xid))
1251 : 63493 : conflict_xid = prstate.latest_xid_removed;
1252 : :
1253 : : /* Lock vmbuffer before entering a critical section */
1254 [ + + ]: 431959 : if (do_set_vm)
1255 : 62065 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_EXCLUSIVE);
1256 : :
1257 : : /* Any error while applying the changes is critical */
1258 : 431959 : START_CRIT_SECTION();
1259 : :
1260 [ + + ]: 431959 : if (do_hint_prune)
1261 : : {
1262 : : /*
1263 : : * Update the page's pd_prune_xid field to either zero, or the lowest
1264 : : * XID of any soon-prunable tuple.
1265 : : */
1266 : 177334 : ((PageHeader) prstate.page)->pd_prune_xid = prstate.new_prune_xid;
1267 : :
1268 : : /*
1269 : : * Also clear the "page is full" flag, since there's no point in
1270 : : * repeating the prune/defrag process until something else happens to
1271 : : * the page.
1272 : : */
1273 : 177334 : PageClearFull(prstate.page);
1274 : :
1275 : : /*
1276 : : * If that's all we had to do to the page, this is a non-WAL-logged
1277 : : * hint. If we are going to freeze or prune the page or set
1278 : : * PD_ALL_VISIBLE, we will mark the buffer dirty below.
1279 : : *
1280 : : * Setting PD_ALL_VISIBLE is fully WAL-logged because it is forbidden
1281 : : * for the VM to be set and PD_ALL_VISIBLE to be clear.
1282 : : */
1283 [ + + + + : 177334 : if (!do_freeze && !do_prune && !do_set_vm)
+ + ]
1284 : 67230 : MarkBufferDirtyHint(prstate.buffer, true);
1285 : : }
1286 : :
1287 [ + + + + : 431959 : if (do_prune || do_freeze || do_set_vm)
+ + ]
1288 : : {
1289 : : /* Apply the planned item changes and repair page fragmentation. */
1290 [ + + ]: 135758 : if (do_prune)
1291 : : {
1292 : 71939 : heap_page_prune_execute(prstate.buffer, false,
1293 : : prstate.redirected, prstate.nredirected,
1294 : : prstate.nowdead, prstate.ndead,
1295 : : prstate.nowunused, prstate.nunused);
1296 : : }
1297 : :
1298 [ + + ]: 135758 : if (do_freeze)
1299 : 27139 : heap_freeze_prepared_tuples(prstate.buffer, prstate.frozen, prstate.nfrozen);
1300 : :
1301 : : /* Set the visibility map and page visibility hint */
1302 [ + + ]: 135758 : if (do_set_vm)
1303 : : {
1304 : : /*
1305 : : * While it is valid for PD_ALL_VISIBLE to be set when the
1306 : : * corresponding VM bit is clear, we strongly prefer to keep them
1307 : : * in sync.
1308 : : *
1309 : : * The heap buffer must be marked dirty before adding it to the
1310 : : * WAL chain when setting the VM. We don't worry about
1311 : : * unnecessarily dirtying the heap buffer if PD_ALL_VISIBLE is
1312 : : * already set, though. It is extremely rare to have a clean heap
1313 : : * buffer with PD_ALL_VISIBLE already set and the VM bits clear,
1314 : : * so there is no point in optimizing it.
1315 : : */
1316 : 62065 : PageSetAllVisible(prstate.page);
1317 : 62065 : PageClearPrunable(prstate.page);
1318 : 62065 : visibilitymap_set(prstate.block, prstate.vmbuffer, prstate.new_vmbits,
1319 : 62065 : prstate.relation->rd_locator);
1320 : : }
1321 : :
1322 : 135758 : MarkBufferDirty(prstate.buffer);
1323 : :
1324 : : /*
1325 : : * Emit a WAL XLOG_HEAP2_PRUNE* record showing what we did
1326 : : */
1327 [ + + + + : 135758 : if (RelationNeedsWAL(prstate.relation))
+ - + - ]
1328 : : {
1329 [ + + + + ]: 192582 : log_heap_prune_and_freeze(prstate.relation, prstate.buffer,
1330 : : do_set_vm ? prstate.vmbuffer : InvalidBuffer,
1331 : 60093 : do_set_vm ? prstate.new_vmbits : 0,
1332 : : conflict_xid,
1333 : : true, /* cleanup lock */
1334 : : params->reason,
1335 : : prstate.frozen, prstate.nfrozen,
1336 : : prstate.redirected, prstate.nredirected,
1337 : : prstate.nowdead, prstate.ndead,
1338 : : prstate.nowunused, prstate.nunused);
1339 : : }
1340 : : }
1341 : :
1342 : 431959 : END_CRIT_SECTION();
1343 : :
1344 [ + + ]: 431959 : if (do_set_vm)
1345 : 62065 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_UNLOCK);
1346 : :
1347 : : /*
1348 : : * During its second pass over the heap, VACUUM calls
1349 : : * heap_page_would_be_all_visible() to determine whether a page is
1350 : : * all-visible and all-frozen. The logic here is similar. After completing
1351 : : * pruning and freezing, use an assertion to verify that our results
1352 : : * remain consistent with heap_page_would_be_all_visible(). It's also a
1353 : : * valuable cross-check of the page state after pruning and freezing.
1354 : : */
1355 : : #ifdef USE_ASSERT_CHECKING
1356 : : if (prstate.set_all_visible)
1357 : : {
1358 : : TransactionId debug_cutoff;
1359 : : bool debug_all_frozen;
1360 : :
1361 : : Assert(prstate.lpdead_items == 0);
1362 : :
1363 : : Assert(heap_page_is_all_visible(prstate.relation, prstate.buffer,
1364 : : prstate.vistest,
1365 : : &debug_all_frozen,
1366 : : &debug_cutoff, off_loc));
1367 : :
1368 : : Assert(!TransactionIdIsValid(debug_cutoff) ||
1369 : : debug_cutoff == prstate.newest_live_xid);
1370 : :
1371 : : /*
1372 : : * It's possible the page is composed entirely of frozen tuples but is
1373 : : * not set all-frozen in the VM and did not pass
1374 : : * HEAP_PAGE_PRUNE_FREEZE. In this case, it's possible
1375 : : * heap_page_is_all_visible() finds the page completely frozen, even
1376 : : * though prstate.set_all_frozen is false.
1377 : : */
1378 : : Assert(!prstate.set_all_frozen || debug_all_frozen);
1379 : : }
1380 : : #endif
1381 : :
1382 : : /* Copy information back for caller */
1383 : 431959 : presult->ndeleted = prstate.ndeleted;
1384 : 431959 : presult->nnewlpdead = prstate.ndead;
1385 : 431959 : presult->nfrozen = prstate.nfrozen;
1386 : 431959 : presult->live_tuples = prstate.live_tuples;
1387 : 431959 : presult->recently_dead_tuples = prstate.recently_dead_tuples;
1388 : 431959 : presult->hastup = prstate.hastup;
1389 : :
1390 : 431959 : presult->lpdead_items = prstate.lpdead_items;
1391 : : /* the presult->deadoffsets array was already filled in */
1392 : :
1393 : 431959 : presult->newly_all_visible = false;
1394 : 431959 : presult->newly_all_frozen = false;
1395 : 431959 : presult->newly_all_visible_frozen = false;
1396 [ + + ]: 431959 : if (do_set_vm)
1397 : : {
1398 [ + + ]: 62065 : if ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
1399 : : {
1400 : 58495 : presult->newly_all_visible = true;
1401 [ + + ]: 58495 : if (prstate.set_all_frozen)
1402 : 30035 : presult->newly_all_visible_frozen = true;
1403 : : }
1404 [ + - ]: 3570 : else if ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
1405 [ + - ]: 3570 : prstate.set_all_frozen)
1406 : 3570 : presult->newly_all_frozen = true;
1407 : : }
1408 : :
1409 [ + + ]: 431959 : if (prstate.attempt_freeze)
1410 : : {
1411 [ + + ]: 298213 : if (presult->nfrozen > 0)
1412 : : {
1413 : 27139 : *new_relfrozen_xid = prstate.pagefrz.FreezePageRelfrozenXid;
1414 : 27139 : *new_relmin_mxid = prstate.pagefrz.FreezePageRelminMxid;
1415 : : }
1416 : : else
1417 : : {
1418 : 271074 : *new_relfrozen_xid = prstate.pagefrz.NoFreezePageRelfrozenXid;
1419 : 271074 : *new_relmin_mxid = prstate.pagefrz.NoFreezePageRelminMxid;
1420 : : }
1421 : : }
1422 : : }
1423 : :
1424 : :
1425 : : /*
1426 : : * Perform visibility checks for heap pruning.
1427 : : */
1428 : : static HTSV_Result
1429 : 26891009 : heap_prune_satisfies_vacuum(PruneState *prstate, HeapTuple tup)
1430 : : {
1431 : : HTSV_Result res;
1432 : : TransactionId dead_after;
1433 : :
1434 : 26891009 : res = HeapTupleSatisfiesVacuumHorizon(tup, prstate->buffer, &dead_after);
1435 : :
1436 [ + + ]: 26891009 : if (res != HEAPTUPLE_RECENTLY_DEAD)
1437 : 22618862 : return res;
1438 : :
1439 : : /*
1440 : : * For VACUUM, we must be sure to prune tuples with xmax older than
1441 : : * OldestXmin -- a visibility cutoff determined at the beginning of
1442 : : * vacuuming the relation. OldestXmin is used for freezing determination
1443 : : * and we cannot freeze dead tuples' xmaxes.
1444 : : */
1445 [ + + ]: 4272147 : if (prstate->cutoffs &&
1446 [ + - ]: 1295594 : TransactionIdIsValid(prstate->cutoffs->OldestXmin) &&
1447 [ + + ]: 1295594 : NormalTransactionIdPrecedes(dead_after, prstate->cutoffs->OldestXmin))
1448 : 964393 : return HEAPTUPLE_DEAD;
1449 : :
1450 : : /*
1451 : : * Determine whether or not the tuple is considered dead when compared
1452 : : * with the provided GlobalVisState. On-access pruning does not provide
1453 : : * VacuumCutoffs. And for vacuum, even if the tuple's xmax is not older
1454 : : * than OldestXmin, GlobalVisTestIsRemovableXid() could find the row dead
1455 : : * if the GlobalVisState has been updated since the beginning of vacuuming
1456 : : * the relation.
1457 : : */
1458 [ + + ]: 3307754 : if (GlobalVisTestIsRemovableXid(prstate->vistest, dead_after, true))
1459 : 2925308 : return HEAPTUPLE_DEAD;
1460 : :
1461 : 382446 : return res;
1462 : : }
1463 : :
1464 : :
1465 : : /*
1466 : : * Pruning calculates tuple visibility once and saves the results in an array
1467 : : * of int8. See PruneState.htsv for details. This helper function is meant
1468 : : * to guard against examining visibility status array members which have not
1469 : : * yet been computed.
1470 : : */
1471 : : static inline HTSV_Result
1472 : 26874441 : htsv_get_valid_status(int status)
1473 : : {
1474 : : Assert(status >= HEAPTUPLE_DEAD &&
1475 : : status <= HEAPTUPLE_DELETE_IN_PROGRESS);
1476 : 26874441 : return (HTSV_Result) status;
1477 : : }
1478 : :
1479 : : /*
1480 : : * Prune specified line pointer or a HOT chain originating at line pointer.
1481 : : *
1482 : : * Tuple visibility information is provided in prstate->htsv.
1483 : : *
1484 : : * If the item is an index-referenced tuple (i.e. not a heap-only tuple),
1485 : : * the HOT chain is pruned by removing all DEAD tuples at the start of the HOT
1486 : : * chain. We also prune any RECENTLY_DEAD tuples preceding a DEAD tuple.
1487 : : * This is OK because a RECENTLY_DEAD tuple preceding a DEAD tuple is really
1488 : : * DEAD, our visibility test is just too coarse to detect it.
1489 : : *
1490 : : * Pruning must never leave behind a DEAD tuple that still has tuple storage.
1491 : : * VACUUM isn't prepared to deal with that case.
1492 : : *
1493 : : * The root line pointer is redirected to the tuple immediately after the
1494 : : * latest DEAD tuple. If all tuples in the chain are DEAD, the root line
1495 : : * pointer is marked LP_DEAD. (This includes the case of a DEAD simple
1496 : : * tuple, which we treat as a chain of length 1.)
1497 : : *
1498 : : * We don't actually change the page here. We just add entries to the arrays in
1499 : : * prstate showing the changes to be made. Items to be redirected are added
1500 : : * to the redirected[] array (two entries per redirection); items to be set to
1501 : : * LP_DEAD state are added to nowdead[]; and items to be set to LP_UNUSED
1502 : : * state are added to nowunused[]. We perform bookkeeping of live tuples,
1503 : : * visibility etc. based on what the page will look like after the changes
1504 : : * applied. All that bookkeeping is performed in the heap_prune_record_*()
1505 : : * subroutines. The division of labor is that heap_prune_chain() decides the
1506 : : * fate of each tuple, ie. whether it's going to be removed, redirected or
1507 : : * left unchanged, and the heap_prune_record_*() subroutines update PruneState
1508 : : * based on that outcome.
1509 : : */
1510 : : static void
1511 : 26750002 : heap_prune_chain(OffsetNumber maxoff, OffsetNumber rootoffnum,
1512 : : PruneState *prstate)
1513 : : {
1514 : 26750002 : TransactionId priorXmax = InvalidTransactionId;
1515 : : ItemId rootlp;
1516 : : OffsetNumber offnum;
1517 : : OffsetNumber chainitems[MaxHeapTuplesPerPage];
1518 : 26750002 : Page page = prstate->page;
1519 : :
1520 : : /*
1521 : : * After traversing the HOT chain, ndeadchain is the index in chainitems
1522 : : * of the first live successor after the last dead item.
1523 : : */
1524 : 26750002 : int ndeadchain = 0,
1525 : 26750002 : nchain = 0;
1526 : :
1527 : 26750002 : rootlp = PageGetItemId(page, rootoffnum);
1528 : :
1529 : : /* Start from the root tuple */
1530 : 26750002 : offnum = rootoffnum;
1531 : :
1532 : : /* while not end of the chain */
1533 : : for (;;)
1534 : 322477 : {
1535 : : HeapTupleHeader htup;
1536 : : ItemId lp;
1537 : :
1538 : : /* Sanity check (pure paranoia) */
1539 [ - + ]: 27072479 : if (offnum < FirstOffsetNumber)
1540 : 0 : break;
1541 : :
1542 : : /*
1543 : : * An offset past the end of page's line pointer array is possible
1544 : : * when the array was truncated (original item must have been unused)
1545 : : */
1546 [ - + ]: 27072479 : if (offnum > maxoff)
1547 : 0 : break;
1548 : :
1549 : : /* If item is already processed, stop --- it must not be same chain */
1550 [ - + ]: 27072479 : if (prstate->processed[offnum])
1551 : 0 : break;
1552 : :
1553 : 27072479 : lp = PageGetItemId(page, offnum);
1554 : :
1555 : : /*
1556 : : * Unused item obviously isn't part of the chain. Likewise, a dead
1557 : : * line pointer can't be part of the chain. Both of those cases were
1558 : : * already marked as processed.
1559 : : */
1560 : : Assert(ItemIdIsUsed(lp));
1561 : : Assert(!ItemIdIsDead(lp));
1562 : :
1563 : : /*
1564 : : * If we are looking at the redirected root line pointer, jump to the
1565 : : * first normal tuple in the chain. If we find a redirect somewhere
1566 : : * else, stop --- it must not be same chain.
1567 : : */
1568 [ + + ]: 27072479 : if (ItemIdIsRedirected(lp))
1569 : : {
1570 [ - + ]: 198038 : if (nchain > 0)
1571 : 0 : break; /* not at start of chain */
1572 : 198038 : chainitems[nchain++] = offnum;
1573 : 198038 : offnum = ItemIdGetRedirect(rootlp);
1574 : 198038 : continue;
1575 : : }
1576 : :
1577 : : Assert(ItemIdIsNormal(lp));
1578 : :
1579 : 26874441 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1580 : :
1581 : : /*
1582 : : * Check the tuple XMIN against prior XMAX, if any
1583 : : */
1584 [ + + - + ]: 26998880 : if (TransactionIdIsValid(priorXmax) &&
1585 : 124439 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
1586 : 0 : break;
1587 : :
1588 : : /*
1589 : : * OK, this tuple is indeed a member of the chain.
1590 : : */
1591 : 26874441 : chainitems[nchain++] = offnum;
1592 : :
1593 [ + + + - ]: 26874441 : switch (htsv_get_valid_status(prstate->htsv[offnum]))
1594 : : {
1595 : 3961614 : case HEAPTUPLE_DEAD:
1596 : :
1597 : : /* Remember the last DEAD tuple seen */
1598 : 3961614 : ndeadchain = nchain;
1599 : 3961614 : HeapTupleHeaderAdvanceConflictHorizon(htup,
1600 : : &prstate->latest_xid_removed);
1601 : : /* Advance to next chain member */
1602 : 3961614 : break;
1603 : :
1604 : 382446 : case HEAPTUPLE_RECENTLY_DEAD:
1605 : :
1606 : : /*
1607 : : * We don't need to advance the conflict horizon for
1608 : : * RECENTLY_DEAD tuples, even if we are removing them. This
1609 : : * is because we only remove RECENTLY_DEAD tuples if they
1610 : : * precede a DEAD tuple, and the DEAD tuple must have been
1611 : : * inserted by a newer transaction than the RECENTLY_DEAD
1612 : : * tuple by virtue of being later in the chain. We will have
1613 : : * advanced the conflict horizon for the DEAD tuple.
1614 : : */
1615 : :
1616 : : /*
1617 : : * Advance past RECENTLY_DEAD tuples just in case there's a
1618 : : * DEAD one after them. We have to make sure that we don't
1619 : : * miss any DEAD tuples, since DEAD tuples that still have
1620 : : * tuple storage after pruning will confuse VACUUM.
1621 : : */
1622 : 382446 : break;
1623 : :
1624 : 22530381 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1625 : : case HEAPTUPLE_LIVE:
1626 : : case HEAPTUPLE_INSERT_IN_PROGRESS:
1627 : 22530381 : goto process_chain;
1628 : :
1629 : 0 : default:
1630 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1631 : : goto process_chain;
1632 : : }
1633 : :
1634 : : /*
1635 : : * If the tuple is not HOT-updated, then we are at the end of this
1636 : : * HOT-update chain.
1637 : : */
1638 [ + + ]: 4344060 : if (!HeapTupleHeaderIsHotUpdated(htup))
1639 : 4219621 : goto process_chain;
1640 : :
1641 : : /* HOT implies it can't have moved to different partition */
1642 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1643 : :
1644 : : /*
1645 : : * Advance to next chain member.
1646 : : */
1647 : : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == prstate->block);
1648 : 124439 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1649 : 124439 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1650 : : }
1651 : :
1652 [ # # # # ]: 0 : if (ItemIdIsRedirected(rootlp) && nchain < 2)
1653 : : {
1654 : : /*
1655 : : * We found a redirect item that doesn't point to a valid follow-on
1656 : : * item. This can happen if the loop in heap_page_prune_and_freeze()
1657 : : * caused us to visit the dead successor of a redirect item before
1658 : : * visiting the redirect item. We can clean up by setting the
1659 : : * redirect item to LP_DEAD state or LP_UNUSED if the caller
1660 : : * indicated.
1661 : : */
1662 : 0 : heap_prune_record_dead_or_unused(prstate, rootoffnum, false);
1663 : 0 : return;
1664 : : }
1665 : :
1666 : 0 : process_chain:
1667 : :
1668 [ + + ]: 26750002 : if (ndeadchain == 0)
1669 : : {
1670 : : /*
1671 : : * No DEAD tuple was found, so the chain is entirely composed of
1672 : : * normal, unchanged tuples. Leave it alone.
1673 : : */
1674 : 22833338 : int i = 0;
1675 : :
1676 [ + + ]: 22833338 : if (ItemIdIsRedirected(rootlp))
1677 : : {
1678 : 175972 : heap_prune_record_unchanged_lp_redirect(prstate, rootoffnum);
1679 : 175972 : i++;
1680 : : }
1681 [ + + ]: 45671813 : for (; i < nchain; i++)
1682 : 22838475 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1683 : : }
1684 [ + + ]: 3916664 : else if (ndeadchain == nchain)
1685 : : {
1686 : : /*
1687 : : * The entire chain is dead. Mark the root line pointer LP_DEAD, and
1688 : : * fully remove the other tuples in the chain.
1689 : : */
1690 : 3843974 : heap_prune_record_dead_or_unused(prstate, rootoffnum, ItemIdIsNormal(rootlp));
1691 [ + + ]: 3888812 : for (int i = 1; i < nchain; i++)
1692 : 44838 : heap_prune_record_unused(prstate, chainitems[i], true);
1693 : : }
1694 : : else
1695 : : {
1696 : : /*
1697 : : * We found a DEAD tuple in the chain. Redirect the root line pointer
1698 : : * to the first non-DEAD tuple, and mark as unused each intermediate
1699 : : * item that we are able to remove from the chain.
1700 : : */
1701 : 72690 : heap_prune_record_redirect(prstate, rootoffnum, chainitems[ndeadchain],
1702 : 72690 : ItemIdIsNormal(rootlp));
1703 [ + + ]: 94868 : for (int i = 1; i < ndeadchain; i++)
1704 : 22178 : heap_prune_record_unused(prstate, chainitems[i], true);
1705 : :
1706 : : /* the rest of tuples in the chain are normal, unchanged tuples */
1707 [ + + ]: 147042 : for (int i = ndeadchain; i < nchain; i++)
1708 : 74352 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1709 : : }
1710 : : }
1711 : :
1712 : : /* Record lowest soon-prunable XID */
1713 : : static void
1714 : 6077428 : heap_prune_record_prunable(PruneState *prstate, TransactionId xid,
1715 : : OffsetNumber offnum)
1716 : : {
1717 : : /*
1718 : : * This should exactly match the PageSetPrunable macro. We can't store
1719 : : * directly into the page header yet, so we update working state.
1720 : : */
1721 : : Assert(TransactionIdIsNormal(xid));
1722 [ + + + + ]: 11899835 : if (!TransactionIdIsValid(prstate->new_prune_xid) ||
1723 : 5822407 : TransactionIdPrecedes(xid, prstate->new_prune_xid))
1724 : 256496 : prstate->new_prune_xid = xid;
1725 : :
1726 : : /*
1727 : : * It's incorrect for a page to be marked all-visible if it contains
1728 : : * prunable items.
1729 : : */
1730 [ - + ]: 6077428 : if (PageIsAllVisible(prstate->page))
1731 : 0 : heap_page_fix_vm_corruption(prstate, offnum,
1732 : : VM_CORRUPT_TUPLE_VISIBILITY);
1733 : 6077428 : }
1734 : :
1735 : : /* Record line pointer to be redirected */
1736 : : static void
1737 : 72690 : heap_prune_record_redirect(PruneState *prstate,
1738 : : OffsetNumber offnum, OffsetNumber rdoffnum,
1739 : : bool was_normal)
1740 : : {
1741 : : Assert(!prstate->processed[offnum]);
1742 : 72690 : prstate->processed[offnum] = true;
1743 : :
1744 : : /*
1745 : : * Do not mark the redirect target here. It needs to be counted
1746 : : * separately as an unchanged tuple.
1747 : : */
1748 : :
1749 : : Assert(prstate->nredirected < MaxHeapTuplesPerPage);
1750 : 72690 : prstate->redirected[prstate->nredirected * 2] = offnum;
1751 : 72690 : prstate->redirected[prstate->nredirected * 2 + 1] = rdoffnum;
1752 : :
1753 : 72690 : prstate->nredirected++;
1754 : :
1755 : : /*
1756 : : * If the root entry had been a normal tuple, we are deleting it, so count
1757 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1758 : : * count.
1759 : : */
1760 [ + + ]: 72690 : if (was_normal)
1761 : 63685 : prstate->ndeleted++;
1762 : :
1763 : 72690 : prstate->hastup = true;
1764 : 72690 : }
1765 : :
1766 : : /* Record line pointer to be marked dead */
1767 : : static void
1768 : 3808950 : heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
1769 : : bool was_normal)
1770 : : {
1771 : : Assert(!prstate->processed[offnum]);
1772 : 3808950 : prstate->processed[offnum] = true;
1773 : :
1774 : : Assert(prstate->ndead < MaxHeapTuplesPerPage);
1775 : 3808950 : prstate->nowdead[prstate->ndead] = offnum;
1776 : 3808950 : prstate->ndead++;
1777 : :
1778 : : /*
1779 : : * Deliberately delay unsetting set_all_visible and set_all_frozen until
1780 : : * later during pruning. Removable dead tuples shouldn't preclude freezing
1781 : : * the page.
1782 : : */
1783 : :
1784 : : /* Record the dead offset for vacuum */
1785 : 3808950 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1786 : :
1787 : : /*
1788 : : * If the root entry had been a normal tuple, we are deleting it, so count
1789 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1790 : : * count.
1791 : : */
1792 [ + + ]: 3808950 : if (was_normal)
1793 : 3795889 : prstate->ndeleted++;
1794 : 3808950 : }
1795 : :
1796 : : /*
1797 : : * Depending on whether or not the caller set mark_unused_now to true, record that a
1798 : : * line pointer should be marked LP_DEAD or LP_UNUSED. There are other cases in
1799 : : * which we will mark line pointers LP_UNUSED, but we will not mark line
1800 : : * pointers LP_DEAD if mark_unused_now is true.
1801 : : */
1802 : : static void
1803 : 3843974 : heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
1804 : : bool was_normal)
1805 : : {
1806 : : /*
1807 : : * If the caller set mark_unused_now to true, we can remove dead tuples
1808 : : * during pruning instead of marking their line pointers dead. Set this
1809 : : * tuple's line pointer LP_UNUSED. We hint that this option is less
1810 : : * likely.
1811 : : */
1812 [ + + ]: 3843974 : if (unlikely(prstate->mark_unused_now))
1813 : 35024 : heap_prune_record_unused(prstate, offnum, was_normal);
1814 : : else
1815 : 3808950 : heap_prune_record_dead(prstate, offnum, was_normal);
1816 : :
1817 : : /*
1818 : : * It's incorrect for the page to be set all-visible if it contains dead
1819 : : * items. Fix that on the heap page and check the VM for corruption as
1820 : : * well. Do that here rather than in heap_prune_record_dead() so we also
1821 : : * cover tuples that are directly marked LP_UNUSED via mark_unused_now.
1822 : : */
1823 [ - + ]: 3843974 : if (PageIsAllVisible(prstate->page))
1824 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
1825 : 3843974 : }
1826 : :
1827 : : /* Record line pointer to be marked unused */
1828 : : static void
1829 : 107183 : heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal)
1830 : : {
1831 : : Assert(!prstate->processed[offnum]);
1832 : 107183 : prstate->processed[offnum] = true;
1833 : :
1834 : : Assert(prstate->nunused < MaxHeapTuplesPerPage);
1835 : 107183 : prstate->nowunused[prstate->nunused] = offnum;
1836 : 107183 : prstate->nunused++;
1837 : :
1838 : : /*
1839 : : * If the root entry had been a normal tuple, we are deleting it, so count
1840 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1841 : : * count.
1842 : : */
1843 [ + + ]: 107183 : if (was_normal)
1844 : 105317 : prstate->ndeleted++;
1845 : 107183 : }
1846 : :
1847 : : /*
1848 : : * Record an unused line pointer that is left unchanged.
1849 : : */
1850 : : static void
1851 : 240419 : heap_prune_record_unchanged_lp_unused(PruneState *prstate, OffsetNumber offnum)
1852 : : {
1853 : : Assert(!prstate->processed[offnum]);
1854 : 240419 : prstate->processed[offnum] = true;
1855 : 240419 : }
1856 : :
1857 : : /*
1858 : : * Record line pointer that is left unchanged. We consider freezing it, and
1859 : : * update bookkeeping of tuple counts and page visibility.
1860 : : */
1861 : : static void
1862 : 22926118 : heap_prune_record_unchanged_lp_normal(PruneState *prstate, OffsetNumber offnum)
1863 : : {
1864 : : HeapTupleHeader htup;
1865 : : TransactionId xmin;
1866 : 22926118 : Page page = prstate->page;
1867 : :
1868 : : Assert(!prstate->processed[offnum]);
1869 : 22926118 : prstate->processed[offnum] = true;
1870 : :
1871 : 22926118 : prstate->hastup = true; /* the page is not empty */
1872 : :
1873 : : /*
1874 : : * The criteria for counting a tuple as live in this block need to match
1875 : : * what analyze.c's acquire_sample_rows() does, otherwise VACUUM and
1876 : : * ANALYZE may produce wildly different reltuples values, e.g. when there
1877 : : * are many recently-dead tuples.
1878 : : *
1879 : : * The logic here is a bit simpler than acquire_sample_rows(), as VACUUM
1880 : : * can't run inside a transaction block, which makes some cases impossible
1881 : : * (e.g. in-progress insert from the same transaction).
1882 : : *
1883 : : * HEAPTUPLE_DEAD are handled by the other heap_prune_record_*()
1884 : : * subroutines. They don't count dead items like acquire_sample_rows()
1885 : : * does, because we assume that all dead items will become LP_UNUSED
1886 : : * before VACUUM finishes. This difference is only superficial. VACUUM
1887 : : * effectively agrees with ANALYZE about DEAD items, in the end. VACUUM
1888 : : * won't remember LP_DEAD items, but only because they're not supposed to
1889 : : * be left behind when it is done. (Cases where we bypass index vacuuming
1890 : : * will violate this optimistic assumption, but the overall impact of that
1891 : : * should be negligible.)
1892 : : */
1893 : 22926118 : htup = (HeapTupleHeader) PageGetItem(page, PageGetItemId(page, offnum));
1894 : :
1895 [ + + + + : 22926118 : switch (prstate->htsv[offnum])
- ]
1896 : : {
1897 : 16848690 : case HEAPTUPLE_LIVE:
1898 : :
1899 : : /*
1900 : : * Count it as live. Not only is this natural, but it's also what
1901 : : * acquire_sample_rows() does.
1902 : : */
1903 : 16848690 : prstate->live_tuples++;
1904 : :
1905 : : /*
1906 : : * Is the tuple definitely visible to all transactions?
1907 : : *
1908 : : * NB: Like with per-tuple hint bits, we can't set the
1909 : : * PD_ALL_VISIBLE flag if the inserter committed asynchronously.
1910 : : * See SetHintBits for more info. Check that the tuple is hinted
1911 : : * xmin-committed because of that.
1912 : : */
1913 [ + + ]: 16848690 : if (!HeapTupleHeaderXminCommitted(htup))
1914 : : {
1915 : 34192 : prstate->set_all_visible = false;
1916 : 34192 : prstate->set_all_frozen = false;
1917 : 34192 : break;
1918 : : }
1919 : :
1920 : : /*
1921 : : * The inserter definitely committed. But we don't know if it is
1922 : : * old enough that everyone sees it as committed. Later, after
1923 : : * processing all the tuples on the page, we'll check if there is
1924 : : * any snapshot that still considers the newest xid on the page to
1925 : : * be running. If so, we don't consider the page all-visible.
1926 : : */
1927 : 16814498 : xmin = HeapTupleHeaderGetXmin(htup);
1928 : :
1929 : : /* Track newest xmin on page. */
1930 [ + + + + ]: 16814498 : if (TransactionIdFollows(xmin, prstate->newest_live_xid) &&
1931 : : TransactionIdIsNormal(xmin))
1932 : 613765 : prstate->newest_live_xid = xmin;
1933 : :
1934 : 16814498 : break;
1935 : :
1936 : 382446 : case HEAPTUPLE_RECENTLY_DEAD:
1937 : 382446 : prstate->recently_dead_tuples++;
1938 : 382446 : prstate->set_all_visible = false;
1939 : 382446 : prstate->set_all_frozen = false;
1940 : :
1941 : : /*
1942 : : * This tuple will soon become DEAD. Update the hint field so
1943 : : * that the page is reconsidered for pruning in future.
1944 : : */
1945 : 382446 : heap_prune_record_prunable(prstate,
1946 : : HeapTupleHeaderGetUpdateXid(htup),
1947 : : offnum);
1948 : 382446 : break;
1949 : :
1950 : 139402 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1951 : :
1952 : : /*
1953 : : * We do not count these rows as live, because we expect the
1954 : : * inserting transaction to update the counters at commit, and we
1955 : : * assume that will happen only after we report our results. This
1956 : : * assumption is a bit shaky, but it is what acquire_sample_rows()
1957 : : * does, so be consistent.
1958 : : */
1959 : 139402 : prstate->set_all_visible = false;
1960 : 139402 : prstate->set_all_frozen = false;
1961 : :
1962 : : /*
1963 : : * Though there is nothing "prunable" on the page, we maintain
1964 : : * pd_prune_xid for inserts so that we have the opportunity to
1965 : : * mark them all-visible during the next round of pruning.
1966 : : */
1967 : 139402 : heap_prune_record_prunable(prstate,
1968 : : HeapTupleHeaderGetXmin(htup),
1969 : : offnum);
1970 : 139402 : break;
1971 : :
1972 : 5555580 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1973 : :
1974 : : /*
1975 : : * This an expected case during concurrent vacuum. Count such
1976 : : * rows as live. As above, we assume the deleting transaction
1977 : : * will commit and update the counters after we report.
1978 : : */
1979 : 5555580 : prstate->live_tuples++;
1980 : 5555580 : prstate->set_all_visible = false;
1981 : 5555580 : prstate->set_all_frozen = false;
1982 : :
1983 : : /*
1984 : : * This tuple may soon become DEAD. Update the hint field so that
1985 : : * the page is reconsidered for pruning in future.
1986 : : */
1987 : 5555580 : heap_prune_record_prunable(prstate,
1988 : : HeapTupleHeaderGetUpdateXid(htup),
1989 : : offnum);
1990 : 5555580 : break;
1991 : :
1992 : 0 : default:
1993 : :
1994 : : /*
1995 : : * DEAD tuples should've been passed to heap_prune_record_dead()
1996 : : * or heap_prune_record_unused() instead.
1997 : : */
1998 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result %d",
1999 : : prstate->htsv[offnum]);
2000 : : break;
2001 : : }
2002 : :
2003 : : /* Consider freezing any normal tuples which will not be removed */
2004 [ + + ]: 22926118 : if (prstate->attempt_freeze)
2005 : : {
2006 : : bool totally_frozen;
2007 : :
2008 [ + + ]: 12690325 : if ((heap_prepare_freeze_tuple(htup,
2009 : 12690325 : prstate->cutoffs,
2010 : : &prstate->pagefrz,
2011 : 12690325 : &prstate->frozen[prstate->nfrozen],
2012 : : &totally_frozen)))
2013 : : {
2014 : : /* Save prepared freeze plan for later */
2015 : 3539055 : prstate->frozen[prstate->nfrozen++].offset = offnum;
2016 : : }
2017 : :
2018 : : /*
2019 : : * If any tuple isn't either totally frozen already or eligible to
2020 : : * become totally frozen (according to its freeze plan), then the page
2021 : : * definitely cannot be set all-frozen in the visibility map later on.
2022 : : */
2023 [ + + ]: 12690325 : if (!totally_frozen)
2024 : 6281339 : prstate->set_all_frozen = false;
2025 : : }
2026 : 22926118 : }
2027 : :
2028 : :
2029 : : /*
2030 : : * Record line pointer that was already LP_DEAD and is left unchanged.
2031 : : */
2032 : : static void
2033 : 1733153 : heap_prune_record_unchanged_lp_dead(PruneState *prstate, OffsetNumber offnum)
2034 : : {
2035 : : Assert(!prstate->processed[offnum]);
2036 : 1733153 : prstate->processed[offnum] = true;
2037 : :
2038 : : /*
2039 : : * Deliberately don't set hastup for LP_DEAD items. We make the soft
2040 : : * assumption that any LP_DEAD items encountered here will become
2041 : : * LP_UNUSED later on, before count_nondeletable_pages is reached. If we
2042 : : * don't make this assumption then rel truncation will only happen every
2043 : : * other VACUUM, at most. Besides, VACUUM must treat
2044 : : * hastup/nonempty_pages as provisional no matter how LP_DEAD items are
2045 : : * handled (handled here, or handled later on).
2046 : : *
2047 : : * Similarly, don't unset set_all_visible and set_all_frozen until later,
2048 : : * at the end of heap_page_prune_and_freeze(). This will allow us to
2049 : : * attempt to freeze the page after pruning. As long as we unset it
2050 : : * before updating the visibility map, this will be correct.
2051 : : */
2052 : :
2053 : : /* Record the dead offset for vacuum */
2054 : 1733153 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
2055 : :
2056 : : /*
2057 : : * It's incorrect for a page to be marked all-visible if it contains dead
2058 : : * items.
2059 : : */
2060 [ - + ]: 1733153 : if (PageIsAllVisible(prstate->page))
2061 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
2062 : 1733153 : }
2063 : :
2064 : : /*
2065 : : * Record LP_REDIRECT that is left unchanged.
2066 : : */
2067 : : static void
2068 : 175972 : heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum)
2069 : : {
2070 : : /*
2071 : : * A redirect line pointer doesn't count as a live tuple.
2072 : : *
2073 : : * If we leave a redirect line pointer in place, there will be another
2074 : : * tuple on the page that it points to. We will do the bookkeeping for
2075 : : * that separately. So we have nothing to do here, except remember that
2076 : : * we processed this item.
2077 : : */
2078 : : Assert(!prstate->processed[offnum]);
2079 : 175972 : prstate->processed[offnum] = true;
2080 : 175972 : }
2081 : :
2082 : : /*
2083 : : * Perform the actual page changes needed by heap_page_prune_and_freeze().
2084 : : *
2085 : : * If 'lp_truncate_only' is set, we are merely marking LP_DEAD line pointers
2086 : : * as unused, not redirecting or removing anything else. The
2087 : : * PageRepairFragmentation() call is skipped in that case.
2088 : : *
2089 : : * If 'lp_truncate_only' is not set, the caller must hold a cleanup lock on
2090 : : * the buffer. If it is set, an ordinary exclusive lock suffices.
2091 : : */
2092 : : void
2093 : 82908 : heap_page_prune_execute(Buffer buffer, bool lp_truncate_only,
2094 : : OffsetNumber *redirected, int nredirected,
2095 : : OffsetNumber *nowdead, int ndead,
2096 : : OffsetNumber *nowunused, int nunused)
2097 : : {
2098 : 82908 : Page page = BufferGetPage(buffer);
2099 : : OffsetNumber *offnum;
2100 : : HeapTupleHeader htup PG_USED_FOR_ASSERTS_ONLY;
2101 : :
2102 : : /* Shouldn't be called unless there's something to do */
2103 : : Assert(nredirected > 0 || ndead > 0 || nunused > 0);
2104 : :
2105 : : /* If 'lp_truncate_only', we can only remove already-dead line pointers */
2106 : : Assert(!lp_truncate_only || (nredirected == 0 && ndead == 0));
2107 : :
2108 : : /* Update all redirected line pointers */
2109 : 82908 : offnum = redirected;
2110 [ + + ]: 174861 : for (int i = 0; i < nredirected; i++)
2111 : : {
2112 : 91953 : OffsetNumber fromoff = *offnum++;
2113 : 91953 : OffsetNumber tooff = *offnum++;
2114 : 91953 : ItemId fromlp = PageGetItemId(page, fromoff);
2115 : : ItemId tolp PG_USED_FOR_ASSERTS_ONLY;
2116 : :
2117 : : #ifdef USE_ASSERT_CHECKING
2118 : :
2119 : : /*
2120 : : * Any existing item that we set as an LP_REDIRECT (any 'from' item)
2121 : : * must be the first item from a HOT chain. If the item has tuple
2122 : : * storage then it can't be a heap-only tuple. Otherwise we are just
2123 : : * maintaining an existing LP_REDIRECT from an existing HOT chain that
2124 : : * has been pruned at least once before now.
2125 : : */
2126 : : if (!ItemIdIsRedirected(fromlp))
2127 : : {
2128 : : Assert(ItemIdHasStorage(fromlp) && ItemIdIsNormal(fromlp));
2129 : :
2130 : : htup = (HeapTupleHeader) PageGetItem(page, fromlp);
2131 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2132 : : }
2133 : : else
2134 : : {
2135 : : /* We shouldn't need to redundantly set the redirect */
2136 : : Assert(ItemIdGetRedirect(fromlp) != tooff);
2137 : : }
2138 : :
2139 : : /*
2140 : : * The item that we're about to set as an LP_REDIRECT (the 'from'
2141 : : * item) will point to an existing item (the 'to' item) that is
2142 : : * already a heap-only tuple. There can be at most one LP_REDIRECT
2143 : : * item per HOT chain.
2144 : : *
2145 : : * We need to keep around an LP_REDIRECT item (after original
2146 : : * non-heap-only root tuple gets pruned away) so that it's always
2147 : : * possible for VACUUM to easily figure out what TID to delete from
2148 : : * indexes when an entire HOT chain becomes dead. A heap-only tuple
2149 : : * can never become LP_DEAD; an LP_REDIRECT item or a regular heap
2150 : : * tuple can.
2151 : : *
2152 : : * This check may miss problems, e.g. the target of a redirect could
2153 : : * be marked as unused subsequently. The page_verify_redirects() check
2154 : : * below will catch such problems.
2155 : : */
2156 : : tolp = PageGetItemId(page, tooff);
2157 : : Assert(ItemIdHasStorage(tolp) && ItemIdIsNormal(tolp));
2158 : : htup = (HeapTupleHeader) PageGetItem(page, tolp);
2159 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2160 : : #endif
2161 : :
2162 : 91953 : ItemIdSetRedirect(fromlp, tooff);
2163 : : }
2164 : :
2165 : : /* Update all now-dead line pointers */
2166 : 82908 : offnum = nowdead;
2167 [ + + ]: 4174102 : for (int i = 0; i < ndead; i++)
2168 : : {
2169 : 4091194 : OffsetNumber off = *offnum++;
2170 : 4091194 : ItemId lp = PageGetItemId(page, off);
2171 : :
2172 : : #ifdef USE_ASSERT_CHECKING
2173 : :
2174 : : /*
2175 : : * An LP_DEAD line pointer must be left behind when the original item
2176 : : * (which is dead to everybody) could still be referenced by a TID in
2177 : : * an index. This should never be necessary with any individual
2178 : : * heap-only tuple item, though. (It's not clear how much of a problem
2179 : : * that would be, but there is no reason to allow it.)
2180 : : */
2181 : : if (ItemIdHasStorage(lp))
2182 : : {
2183 : : Assert(ItemIdIsNormal(lp));
2184 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2185 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2186 : : }
2187 : : else
2188 : : {
2189 : : /* Whole HOT chain becomes dead */
2190 : : Assert(ItemIdIsRedirected(lp));
2191 : : }
2192 : : #endif
2193 : :
2194 : 4091194 : ItemIdSetDead(lp);
2195 : : }
2196 : :
2197 : : /* Update all now-unused line pointers */
2198 : 82908 : offnum = nowunused;
2199 [ + + ]: 427979 : for (int i = 0; i < nunused; i++)
2200 : : {
2201 : 345071 : OffsetNumber off = *offnum++;
2202 : 345071 : ItemId lp = PageGetItemId(page, off);
2203 : :
2204 : : #ifdef USE_ASSERT_CHECKING
2205 : :
2206 : : if (lp_truncate_only)
2207 : : {
2208 : : /* Setting LP_DEAD to LP_UNUSED in vacuum's second pass */
2209 : : Assert(ItemIdIsDead(lp) && !ItemIdHasStorage(lp));
2210 : : }
2211 : : else
2212 : : {
2213 : : /*
2214 : : * When heap_page_prune_and_freeze() was called, mark_unused_now
2215 : : * may have been passed as true, which allows would-be LP_DEAD
2216 : : * items to be made LP_UNUSED instead. This is only possible if
2217 : : * the relation has no indexes. If there are any dead items, then
2218 : : * mark_unused_now was not true and every item being marked
2219 : : * LP_UNUSED must refer to a heap-only tuple.
2220 : : */
2221 : : if (ndead > 0)
2222 : : {
2223 : : Assert(ItemIdHasStorage(lp) && ItemIdIsNormal(lp));
2224 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2225 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2226 : : }
2227 : : else
2228 : : Assert(ItemIdIsUsed(lp));
2229 : : }
2230 : :
2231 : : #endif
2232 : :
2233 : 345071 : ItemIdSetUnused(lp);
2234 : : }
2235 : :
2236 [ + + ]: 82908 : if (lp_truncate_only)
2237 : 2505 : PageTruncateLinePointerArray(page);
2238 : : else
2239 : : {
2240 : : /*
2241 : : * Finally, repair any fragmentation, and update the page's hint bit
2242 : : * about whether it has free pointers.
2243 : : */
2244 : 80403 : PageRepairFragmentation(page);
2245 : :
2246 : : /*
2247 : : * Now that the page has been modified, assert that redirect items
2248 : : * still point to valid targets.
2249 : : */
2250 : 80403 : page_verify_redirects(page);
2251 : : }
2252 : 82908 : }
2253 : :
2254 : :
2255 : : /*
2256 : : * If built with assertions, verify that all LP_REDIRECT items point to a
2257 : : * valid item.
2258 : : *
2259 : : * One way that bugs related to HOT pruning show is redirect items pointing to
2260 : : * removed tuples. It's not trivial to reliably check that marking an item
2261 : : * unused will not orphan a redirect item during heap_prune_chain() /
2262 : : * heap_page_prune_execute(), so we additionally check the whole page after
2263 : : * pruning. Without this check such bugs would typically only cause asserts
2264 : : * later, potentially well after the corruption has been introduced.
2265 : : *
2266 : : * Also check comments in heap_page_prune_execute()'s redirection loop.
2267 : : */
2268 : : static void
2269 : 80403 : page_verify_redirects(Page page)
2270 : : {
2271 : : #ifdef USE_ASSERT_CHECKING
2272 : : OffsetNumber offnum;
2273 : : OffsetNumber maxoff;
2274 : :
2275 : : maxoff = PageGetMaxOffsetNumber(page);
2276 : : for (offnum = FirstOffsetNumber;
2277 : : offnum <= maxoff;
2278 : : offnum = OffsetNumberNext(offnum))
2279 : : {
2280 : : ItemId itemid = PageGetItemId(page, offnum);
2281 : : OffsetNumber targoff;
2282 : : ItemId targitem;
2283 : : HeapTupleHeader htup;
2284 : :
2285 : : if (!ItemIdIsRedirected(itemid))
2286 : : continue;
2287 : :
2288 : : targoff = ItemIdGetRedirect(itemid);
2289 : : targitem = PageGetItemId(page, targoff);
2290 : :
2291 : : Assert(ItemIdIsUsed(targitem));
2292 : : Assert(ItemIdIsNormal(targitem));
2293 : : Assert(ItemIdHasStorage(targitem));
2294 : : htup = (HeapTupleHeader) PageGetItem(page, targitem);
2295 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2296 : : }
2297 : : #endif
2298 : 80403 : }
2299 : :
2300 : :
2301 : : /*
2302 : : * For all items in this page, find their respective root line pointers.
2303 : : * If item k is part of a HOT-chain with root at item j, then we set
2304 : : * root_offsets[k - 1] = j.
2305 : : *
2306 : : * The passed-in root_offsets array must have MaxHeapTuplesPerPage entries.
2307 : : * Unused entries are filled with InvalidOffsetNumber (zero).
2308 : : *
2309 : : * The function must be called with at least share lock on the buffer, to
2310 : : * prevent concurrent prune operations.
2311 : : *
2312 : : * Note: The information collected here is valid only as long as the caller
2313 : : * holds a pin on the buffer. Once pin is released, a tuple might be pruned
2314 : : * and reused by a completely unrelated tuple.
2315 : : */
2316 : : void
2317 : 139117 : heap_get_root_tuples(Page page, OffsetNumber *root_offsets)
2318 : : {
2319 : : OffsetNumber offnum,
2320 : : maxoff;
2321 : :
2322 [ + - - + : 139117 : MemSet(root_offsets, InvalidOffsetNumber,
- - - - -
- ]
2323 : : MaxHeapTuplesPerPage * sizeof(OffsetNumber));
2324 : :
2325 : 139117 : maxoff = PageGetMaxOffsetNumber(page);
2326 [ + + ]: 11709563 : for (offnum = FirstOffsetNumber; offnum <= maxoff; offnum = OffsetNumberNext(offnum))
2327 : : {
2328 : 11570446 : ItemId lp = PageGetItemId(page, offnum);
2329 : : HeapTupleHeader htup;
2330 : : OffsetNumber nextoffnum;
2331 : : TransactionId priorXmax;
2332 : :
2333 : : /* skip unused and dead items */
2334 [ + + + + ]: 11570446 : if (!ItemIdIsUsed(lp) || ItemIdIsDead(lp))
2335 : 11673 : continue;
2336 : :
2337 [ + + ]: 11558773 : if (ItemIdIsNormal(lp))
2338 : : {
2339 : 11555524 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2340 : :
2341 : : /*
2342 : : * Check if this tuple is part of a HOT-chain rooted at some other
2343 : : * tuple. If so, skip it for now; we'll process it when we find
2344 : : * its root.
2345 : : */
2346 [ + + ]: 11555524 : if (HeapTupleHeaderIsHeapOnly(htup))
2347 : 3589 : continue;
2348 : :
2349 : : /*
2350 : : * This is either a plain tuple or the root of a HOT-chain.
2351 : : * Remember it in the mapping.
2352 : : */
2353 : 11551935 : root_offsets[offnum - 1] = offnum;
2354 : :
2355 : : /* If it's not the start of a HOT-chain, we're done with it */
2356 [ + + ]: 11551935 : if (!HeapTupleHeaderIsHotUpdated(htup))
2357 : 11551671 : continue;
2358 : :
2359 : : /* Set up to scan the HOT-chain */
2360 : 264 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2361 : 264 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2362 : : }
2363 : : else
2364 : : {
2365 : : /* Must be a redirect item. We do not set its root_offsets entry */
2366 : : Assert(ItemIdIsRedirected(lp));
2367 : : /* Set up to scan the HOT-chain */
2368 : 3249 : nextoffnum = ItemIdGetRedirect(lp);
2369 : 3249 : priorXmax = InvalidTransactionId;
2370 : : }
2371 : :
2372 : : /*
2373 : : * Now follow the HOT-chain and collect other tuples in the chain.
2374 : : *
2375 : : * Note: Even though this is a nested loop, the complexity of the
2376 : : * function is O(N) because a tuple in the page should be visited not
2377 : : * more than twice, once in the outer loop and once in HOT-chain
2378 : : * chases.
2379 : : */
2380 : : for (;;)
2381 : : {
2382 : : /* Sanity check (pure paranoia) */
2383 [ - + ]: 3585 : if (nextoffnum < FirstOffsetNumber)
2384 : 0 : break;
2385 : :
2386 : : /*
2387 : : * An offset past the end of page's line pointer array is possible
2388 : : * when the array was truncated
2389 : : */
2390 [ - + ]: 3585 : if (nextoffnum > maxoff)
2391 : 0 : break;
2392 : :
2393 : 3585 : lp = PageGetItemId(page, nextoffnum);
2394 : :
2395 : : /* Check for broken chains */
2396 [ - + ]: 3585 : if (!ItemIdIsNormal(lp))
2397 : 0 : break;
2398 : :
2399 : 3585 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2400 : :
2401 [ + + - + ]: 3921 : if (TransactionIdIsValid(priorXmax) &&
2402 : 336 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(htup)))
2403 : 0 : break;
2404 : :
2405 : : /* Remember the root line pointer for this item */
2406 : 3585 : root_offsets[nextoffnum - 1] = offnum;
2407 : :
2408 : : /* Advance to next chain member, if any */
2409 [ + + ]: 3585 : if (!HeapTupleHeaderIsHotUpdated(htup))
2410 : 3513 : break;
2411 : :
2412 : : /* HOT implies it can't have moved to different partition */
2413 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
2414 : :
2415 : 72 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2416 : 72 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2417 : : }
2418 : : }
2419 : 139117 : }
2420 : :
2421 : :
2422 : : /*
2423 : : * Compare fields that describe actions required to freeze tuple with caller's
2424 : : * open plan. If everything matches then the frz tuple plan is equivalent to
2425 : : * caller's plan.
2426 : : */
2427 : : static inline bool
2428 : 1344970 : heap_log_freeze_eq(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2429 : : {
2430 [ + + ]: 1344970 : if (plan->xmax == frz->xmax &&
2431 [ + + ]: 1343676 : plan->t_infomask2 == frz->t_infomask2 &&
2432 [ + + ]: 1342652 : plan->t_infomask == frz->t_infomask &&
2433 [ + - ]: 1339469 : plan->frzflags == frz->frzflags)
2434 : 1339469 : return true;
2435 : :
2436 : : /* Caller must call heap_log_freeze_new_plan again for frz */
2437 : 5501 : return false;
2438 : : }
2439 : :
2440 : : /*
2441 : : * Comparator used to deduplicate the freeze plans used in WAL records.
2442 : : */
2443 : : static int
2444 : 1814430 : heap_log_freeze_cmp(const void *arg1, const void *arg2)
2445 : : {
2446 : 1814430 : const HeapTupleFreeze *frz1 = arg1;
2447 : 1814430 : const HeapTupleFreeze *frz2 = arg2;
2448 : :
2449 [ + + ]: 1814430 : if (frz1->xmax < frz2->xmax)
2450 : 13228 : return -1;
2451 [ + + ]: 1801202 : else if (frz1->xmax > frz2->xmax)
2452 : 14282 : return 1;
2453 : :
2454 [ + + ]: 1786920 : if (frz1->t_infomask2 < frz2->t_infomask2)
2455 : 6105 : return -1;
2456 [ + + ]: 1780815 : else if (frz1->t_infomask2 > frz2->t_infomask2)
2457 : 6078 : return 1;
2458 : :
2459 [ + + ]: 1774737 : if (frz1->t_infomask < frz2->t_infomask)
2460 : 13521 : return -1;
2461 [ + + ]: 1761216 : else if (frz1->t_infomask > frz2->t_infomask)
2462 : 22571 : return 1;
2463 : :
2464 [ - + ]: 1738645 : if (frz1->frzflags < frz2->frzflags)
2465 : 0 : return -1;
2466 [ - + ]: 1738645 : else if (frz1->frzflags > frz2->frzflags)
2467 : 0 : return 1;
2468 : :
2469 : : /*
2470 : : * heap_log_freeze_eq would consider these tuple-wise plans to be equal.
2471 : : * (So the tuples will share a single canonical freeze plan.)
2472 : : *
2473 : : * We tiebreak on page offset number to keep each freeze plan's page
2474 : : * offset number array individually sorted. (Unnecessary, but be tidy.)
2475 : : */
2476 [ + + ]: 1738645 : if (frz1->offset < frz2->offset)
2477 : 1496149 : return -1;
2478 [ + - ]: 242496 : else if (frz1->offset > frz2->offset)
2479 : 242496 : return 1;
2480 : :
2481 : : Assert(false);
2482 : 0 : return 0;
2483 : : }
2484 : :
2485 : : /*
2486 : : * Start new plan initialized using tuple-level actions. At least one tuple
2487 : : * will have steps required to freeze described by caller's plan during REDO.
2488 : : */
2489 : : static inline void
2490 : 32637 : heap_log_freeze_new_plan(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2491 : : {
2492 : 32637 : plan->xmax = frz->xmax;
2493 : 32637 : plan->t_infomask2 = frz->t_infomask2;
2494 : 32637 : plan->t_infomask = frz->t_infomask;
2495 : 32637 : plan->frzflags = frz->frzflags;
2496 : 32637 : plan->ntuples = 1; /* for now */
2497 : 32637 : }
2498 : :
2499 : : /*
2500 : : * Deduplicate tuple-based freeze plans so that each distinct set of
2501 : : * processing steps is only stored once in the WAL record.
2502 : : * Called during original execution of freezing (for logged relations).
2503 : : *
2504 : : * Return value is number of plans set in *plans_out for caller. Also writes
2505 : : * an array of offset numbers into *offsets_out output argument for caller
2506 : : * (actually there is one array per freeze plan, but that's not of immediate
2507 : : * concern to our caller).
2508 : : */
2509 : : static int
2510 : 27136 : heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
2511 : : xlhp_freeze_plan *plans_out,
2512 : : OffsetNumber *offsets_out)
2513 : : {
2514 : 27136 : int nplans = 0;
2515 : :
2516 : : /* Sort tuple-based freeze plans in the order required to deduplicate */
2517 : 27136 : qsort(tuples, ntuples, sizeof(HeapTupleFreeze), heap_log_freeze_cmp);
2518 : :
2519 [ + + ]: 1399242 : for (int i = 0; i < ntuples; i++)
2520 : : {
2521 : 1372106 : HeapTupleFreeze *frz = tuples + i;
2522 : :
2523 [ + + ]: 1372106 : if (i == 0)
2524 : : {
2525 : : /* New canonical freeze plan starting with first tup */
2526 : 27136 : heap_log_freeze_new_plan(plans_out, frz);
2527 : 27136 : nplans++;
2528 : : }
2529 [ + + ]: 1344970 : else if (heap_log_freeze_eq(plans_out, frz))
2530 : : {
2531 : : /* tup matches open canonical plan -- include tup in it */
2532 : : Assert(offsets_out[i - 1] < frz->offset);
2533 : 1339469 : plans_out->ntuples++;
2534 : : }
2535 : : else
2536 : : {
2537 : : /* Tup doesn't match current plan -- done with it now */
2538 : 5501 : plans_out++;
2539 : :
2540 : : /* New canonical freeze plan starting with this tup */
2541 : 5501 : heap_log_freeze_new_plan(plans_out, frz);
2542 : 5501 : nplans++;
2543 : : }
2544 : :
2545 : : /*
2546 : : * Save page offset number in dedicated buffer in passing.
2547 : : *
2548 : : * REDO routine relies on the record's offset numbers array grouping
2549 : : * offset numbers by freeze plan. The sort order within each grouping
2550 : : * is ascending offset number order, just to keep things tidy.
2551 : : */
2552 : 1372106 : offsets_out[i] = frz->offset;
2553 : : }
2554 : :
2555 : : Assert(nplans > 0 && nplans <= ntuples);
2556 : :
2557 : 27136 : return nplans;
2558 : : }
2559 : :
2560 : : /*
2561 : : * Write an XLOG_HEAP2_PRUNE* WAL record
2562 : : *
2563 : : * This is used for several different page maintenance operations:
2564 : : *
2565 : : * - Page pruning, in VACUUM's 1st pass or on access: Some items are
2566 : : * redirected, some marked dead, and some removed altogether.
2567 : : *
2568 : : * - Freezing: Items are marked as 'frozen'.
2569 : : *
2570 : : * - Vacuum, 2nd pass: Items that are already LP_DEAD are marked as unused.
2571 : : *
2572 : : * They have enough commonalities that we use a single WAL record for them
2573 : : * all.
2574 : : *
2575 : : * If replaying the record requires a cleanup lock, pass cleanup_lock = true.
2576 : : * Replaying 'redirected' or 'dead' items always requires a cleanup lock, but
2577 : : * replaying 'unused' items depends on whether they were all previously marked
2578 : : * as dead.
2579 : : *
2580 : : * If the VM is being updated, vmflags will contain the bits to set. In this
2581 : : * case, vmbuffer should already have been updated and marked dirty and should
2582 : : * still be pinned and locked.
2583 : : *
2584 : : * Note: This function scribbles on the 'frozen' array.
2585 : : *
2586 : : * Note: This is called in a critical section, so careful what you do here.
2587 : : */
2588 : : void
2589 : 149177 : log_heap_prune_and_freeze(Relation relation, Buffer buffer,
2590 : : Buffer vmbuffer, uint8 vmflags,
2591 : : TransactionId conflict_xid,
2592 : : bool cleanup_lock,
2593 : : PruneReason reason,
2594 : : HeapTupleFreeze *frozen, int nfrozen,
2595 : : OffsetNumber *redirected, int nredirected,
2596 : : OffsetNumber *dead, int ndead,
2597 : : OffsetNumber *unused, int nunused)
2598 : : {
2599 : : xl_heap_prune xlrec;
2600 : : XLogRecPtr recptr;
2601 : : uint8 info;
2602 : : uint8 regbuf_flags_heap;
2603 : :
2604 : 149177 : Page heap_page = BufferGetPage(buffer);
2605 : :
2606 : : /* The following local variables hold data registered in the WAL record: */
2607 : : xlhp_freeze_plan plans[MaxHeapTuplesPerPage];
2608 : : xlhp_freeze_plans freeze_plans;
2609 : : xlhp_prune_items redirect_items;
2610 : : xlhp_prune_items dead_items;
2611 : : xlhp_prune_items unused_items;
2612 : : OffsetNumber frz_offsets[MaxHeapTuplesPerPage];
2613 [ + + + + : 149177 : bool do_prune = nredirected > 0 || ndead > 0 || nunused > 0;
+ + ]
2614 : 149177 : bool do_set_vm = vmflags & VISIBILITYMAP_VALID_BITS;
2615 : 149177 : bool heap_fpi_allowed = true;
2616 : :
2617 : : Assert((vmflags & VISIBILITYMAP_VALID_BITS) == vmflags);
2618 : :
2619 : 149177 : xlrec.flags = 0;
2620 : 149177 : regbuf_flags_heap = REGBUF_STANDARD;
2621 : :
2622 : : /*
2623 : : * We can avoid an FPI of the heap page if the only modification we are
2624 : : * making to it is to set PD_ALL_VISIBLE and checksums/wal_log_hints are
2625 : : * disabled.
2626 : : *
2627 : : * However, if the page has never been WAL-logged (LSN is invalid), we
2628 : : * must force an FPI regardless. This can happen when another backend
2629 : : * extends the heap, initializes the page, and then fails before WAL-
2630 : : * logging it. Since heap extension is not WAL-logged, recovery might try
2631 : : * to replay our record and find that the page isn't initialized, which
2632 : : * would cause a PANIC.
2633 : : */
2634 [ - + ]: 149177 : if (!XLogRecPtrIsValid(PageGetLSN(heap_page)))
2635 : 0 : regbuf_flags_heap |= REGBUF_FORCE_IMAGE;
2636 [ + + + + : 149177 : else if (!do_prune && nfrozen == 0 && (!do_set_vm || !XLogHintBitIsNeeded()))
+ - + + +
+ ]
2637 : : {
2638 : 2414 : regbuf_flags_heap |= REGBUF_NO_IMAGE;
2639 : 2414 : heap_fpi_allowed = false;
2640 : : }
2641 : :
2642 : : /*
2643 : : * Prepare data for the buffer. The arrays are not actually in the
2644 : : * buffer, but we pretend that they are. When XLogInsert stores a full
2645 : : * page image, the arrays can be omitted.
2646 : : */
2647 : 149177 : XLogBeginInsert();
2648 : 149177 : XLogRegisterBuffer(0, buffer, regbuf_flags_heap);
2649 : :
2650 [ + + ]: 149177 : if (do_set_vm)
2651 : 76672 : XLogRegisterBuffer(1, vmbuffer, 0);
2652 : :
2653 [ + + ]: 149177 : if (nfrozen > 0)
2654 : : {
2655 : : int nplans;
2656 : :
2657 : 27136 : xlrec.flags |= XLHP_HAS_FREEZE_PLANS;
2658 : :
2659 : : /*
2660 : : * Prepare deduplicated representation for use in the WAL record. This
2661 : : * destructively sorts frozen tuples array in-place.
2662 : : */
2663 : 27136 : nplans = heap_log_freeze_plan(frozen, nfrozen, plans, frz_offsets);
2664 : :
2665 : 27136 : freeze_plans.nplans = nplans;
2666 : 27136 : XLogRegisterBufData(0, &freeze_plans,
2667 : : offsetof(xlhp_freeze_plans, plans));
2668 : 27136 : XLogRegisterBufData(0, plans,
2669 : : sizeof(xlhp_freeze_plan) * nplans);
2670 : : }
2671 [ + + ]: 149177 : if (nredirected > 0)
2672 : : {
2673 : 17653 : xlrec.flags |= XLHP_HAS_REDIRECTIONS;
2674 : :
2675 : 17653 : redirect_items.ntargets = nredirected;
2676 : 17653 : XLogRegisterBufData(0, &redirect_items,
2677 : : offsetof(xlhp_prune_items, data));
2678 : 17653 : XLogRegisterBufData(0, redirected,
2679 : : sizeof(OffsetNumber[2]) * nredirected);
2680 : : }
2681 [ + + ]: 149177 : if (ndead > 0)
2682 : : {
2683 : 58120 : xlrec.flags |= XLHP_HAS_DEAD_ITEMS;
2684 : :
2685 : 58120 : dead_items.ntargets = ndead;
2686 : 58120 : XLogRegisterBufData(0, &dead_items,
2687 : : offsetof(xlhp_prune_items, data));
2688 : 58120 : XLogRegisterBufData(0, dead,
2689 : : sizeof(OffsetNumber) * ndead);
2690 : : }
2691 [ + + ]: 149177 : if (nunused > 0)
2692 : : {
2693 : 31057 : xlrec.flags |= XLHP_HAS_NOW_UNUSED_ITEMS;
2694 : :
2695 : 31057 : unused_items.ntargets = nunused;
2696 : 31057 : XLogRegisterBufData(0, &unused_items,
2697 : : offsetof(xlhp_prune_items, data));
2698 : 31057 : XLogRegisterBufData(0, unused,
2699 : : sizeof(OffsetNumber) * nunused);
2700 : : }
2701 [ + + ]: 149177 : if (nfrozen > 0)
2702 : 27136 : XLogRegisterBufData(0, frz_offsets,
2703 : : sizeof(OffsetNumber) * nfrozen);
2704 : :
2705 : : /*
2706 : : * Prepare the main xl_heap_prune record. We already set the XLHP_HAS_*
2707 : : * flag above.
2708 : : */
2709 [ + + ]: 149177 : if (vmflags & VISIBILITYMAP_ALL_VISIBLE)
2710 : : {
2711 : 76672 : xlrec.flags |= XLHP_VM_ALL_VISIBLE;
2712 [ + + ]: 76672 : if (vmflags & VISIBILITYMAP_ALL_FROZEN)
2713 : 46018 : xlrec.flags |= XLHP_VM_ALL_FROZEN;
2714 : : }
2715 [ + + + + : 149177 : if (RelationIsAccessibleInLogicalDecoding(relation))
+ - - + -
- - - + +
+ + - + -
- + - ]
2716 : 664 : xlrec.flags |= XLHP_IS_CATALOG_REL;
2717 [ + + ]: 149177 : if (TransactionIdIsValid(conflict_xid))
2718 : 118926 : xlrec.flags |= XLHP_HAS_CONFLICT_HORIZON;
2719 [ + + ]: 149177 : if (cleanup_lock)
2720 : 132489 : xlrec.flags |= XLHP_CLEANUP_LOCK;
2721 : : else
2722 : : {
2723 : : Assert(nredirected == 0 && ndead == 0);
2724 : : /* also, any items in 'unused' must've been LP_DEAD previously */
2725 : : }
2726 : 149177 : XLogRegisterData(&xlrec, SizeOfHeapPrune);
2727 [ + + ]: 149177 : if (TransactionIdIsValid(conflict_xid))
2728 : 118926 : XLogRegisterData(&conflict_xid, sizeof(TransactionId));
2729 : :
2730 [ + + + - ]: 149177 : switch (reason)
2731 : : {
2732 : 70288 : case PRUNE_ON_ACCESS:
2733 : 70288 : info = XLOG_HEAP2_PRUNE_ON_ACCESS;
2734 : 70288 : break;
2735 : 62201 : case PRUNE_VACUUM_SCAN:
2736 : 62201 : info = XLOG_HEAP2_PRUNE_VACUUM_SCAN;
2737 : 62201 : break;
2738 : 16688 : case PRUNE_VACUUM_CLEANUP:
2739 : 16688 : info = XLOG_HEAP2_PRUNE_VACUUM_CLEANUP;
2740 : 16688 : break;
2741 : 0 : default:
2742 [ # # ]: 0 : elog(ERROR, "unrecognized prune reason: %d", (int) reason);
2743 : : break;
2744 : : }
2745 : 149177 : recptr = XLogInsert(RM_HEAP2_ID, info);
2746 : :
2747 [ + + ]: 149177 : if (do_set_vm)
2748 : : {
2749 : : Assert(BufferIsDirty(vmbuffer));
2750 : 76672 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2751 : : }
2752 : :
2753 : : /*
2754 : : * If we explicitly skip an FPI, we must not stamp the heap page with this
2755 : : * record's LSN. Recovery skips records <= the stamped LSN, so this could
2756 : : * lead to skipping an earlier FPI needed to repair a torn page.
2757 : : */
2758 [ + + ]: 149177 : if (heap_fpi_allowed)
2759 : : {
2760 : : Assert(BufferIsDirty(buffer));
2761 : 146763 : PageSetLSN(heap_page, recptr);
2762 : : }
2763 : 149177 : }
|
