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 : 12271387 : heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer,
273 : : bool rel_read_only)
274 : : {
275 : 12271387 : 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 [ + + ]: 12271387 : if (RecoveryInProgress())
286 : 269303 : 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 : 12002084 : prune_xid = PageGetPruneXid(page);
295 [ + + ]: 12002084 : if (!TransactionIdIsValid(prune_xid))
296 : 8247314 : return;
297 : :
298 : : /*
299 : : * Check whether prune_xid indicates that there may be dead rows that can
300 : : * be cleaned up.
301 : : */
302 : 3754770 : vistest = GlobalVisTestFor(relation);
303 : :
304 [ + + ]: 3754770 : if (!GlobalVisTestIsRemovableXid(vistest, prune_xid, true))
305 : 1403759 : 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 [ + + ]: 2351011 : minfree = RelationGetTargetPageFreeSpace(relation,
320 : : HEAP_DEFAULT_FILLFACTOR);
321 : 2351011 : minfree = Max(minfree, BLCKSZ / 10);
322 : :
323 [ + + + + ]: 2351011 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
324 : : {
325 : 136495 : bool record_free_space = false;
326 : 136495 : 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 : 136495 : visibilitymap_pin(relation, BufferGetBlockNumber(buffer), vmbuffer);
335 : :
336 : : /* OK, try to get exclusive buffer lock */
337 [ + + ]: 136495 : if (!ConditionalLockBufferForCleanup(buffer))
338 : 1807 : 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 [ + + + + ]: 134688 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
346 : : {
347 : : OffsetNumber dummy_off_loc;
348 : : PruneFreezeResult presult;
349 : : PruneFreezeParams params;
350 : :
351 : 134685 : params.relation = relation;
352 : 134685 : params.buffer = buffer;
353 : 134685 : params.vmbuffer = *vmbuffer;
354 : 134685 : params.reason = PRUNE_ON_ACCESS;
355 : 134685 : params.vistest = vistest;
356 : 134685 : 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 : 134685 : params.options = HEAP_PAGE_PRUNE_ALLOW_FAST_PATH;
365 [ + + ]: 134685 : if (rel_read_only)
366 : 37448 : params.options |= HEAP_PAGE_PRUNE_SET_VM;
367 : :
368 : 134685 : 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 [ + + ]: 134685 : if (presult.ndeleted > presult.nnewlpdead)
386 : 20651 : pgstat_update_heap_dead_tuples(relation,
387 : 20651 : 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 [ + + ]: 134685 : if (presult.newly_all_visible)
397 : : {
398 : 14353 : record_free_space = true;
399 : 14353 : freespace = PageGetHeapFreeSpace(page);
400 : : }
401 : : }
402 : :
403 : : /* And release buffer lock */
404 : 134688 : 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 [ + + ]: 134688 : if (record_free_space)
413 : 14353 : 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 : 643908 : 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 : 643908 : prstate->vistest = params->vistest;
433 : 643908 : prstate->mark_unused_now =
434 : 643908 : (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 : 643908 : prstate->attempt_freeze = (params->options & HEAP_PAGE_PRUNE_FREEZE) != 0;
439 : 643908 : prstate->attempt_set_vm = (params->options & HEAP_PAGE_PRUNE_SET_VM) != 0;
440 : 643908 : prstate->cutoffs = params->cutoffs;
441 : 643908 : prstate->relation = params->relation;
442 : 643908 : prstate->block = BufferGetBlockNumber(params->buffer);
443 : 643908 : prstate->buffer = params->buffer;
444 : 643908 : prstate->page = BufferGetPage(params->buffer);
445 : :
446 : : /*
447 : : * The caller must have pinned the VM page covering this heap block. If it
448 : : * doesn't have the correct page pinned, visibilitymap_get_status() will
449 : : * silently release the caller's pin and take its own, leaving the caller
450 : : * holding a stale buffer and leaking ours.
451 : : */
452 : : Assert(visibilitymap_pin_ok(prstate->block, params->vmbuffer));
453 : 643908 : prstate->vmbuffer = params->vmbuffer;
454 : 643908 : prstate->new_vmbits = 0;
455 : 643908 : prstate->old_vmbits = visibilitymap_get_status(prstate->relation,
456 : : prstate->block,
457 : : &prstate->vmbuffer);
458 : :
459 : : /*
460 : : * Our strategy is to scan the page and make lists of items to change,
461 : : * then apply the changes within a critical section. This keeps as much
462 : : * logic as possible out of the critical section, and also ensures that
463 : : * WAL replay will work the same as the normal case.
464 : : *
465 : : * First, initialize the new pd_prune_xid value to zero (indicating no
466 : : * prunable tuples). If we find any tuples which may soon become
467 : : * prunable, we will save the lowest relevant XID in new_prune_xid. Also
468 : : * initialize the rest of our working state.
469 : : */
470 : 643908 : prstate->new_prune_xid = InvalidTransactionId;
471 : 643908 : prstate->latest_xid_removed = InvalidTransactionId;
472 : 643908 : prstate->nredirected = prstate->ndead = prstate->nunused = 0;
473 : 643908 : prstate->nfrozen = 0;
474 : 643908 : prstate->nroot_items = 0;
475 : 643908 : prstate->nheaponly_items = 0;
476 : :
477 : : /* initialize page freezing working state */
478 : 643908 : prstate->pagefrz.freeze_required = false;
479 : 643908 : prstate->pagefrz.FreezePageConflictXid = InvalidTransactionId;
480 [ + + ]: 643908 : if (prstate->attempt_freeze)
481 : : {
482 : : Assert(new_relfrozen_xid && new_relmin_mxid);
483 : 509223 : prstate->pagefrz.FreezePageRelfrozenXid = *new_relfrozen_xid;
484 : 509223 : prstate->pagefrz.NoFreezePageRelfrozenXid = *new_relfrozen_xid;
485 : 509223 : prstate->pagefrz.FreezePageRelminMxid = *new_relmin_mxid;
486 : 509223 : prstate->pagefrz.NoFreezePageRelminMxid = *new_relmin_mxid;
487 : : }
488 : : else
489 : : {
490 : : Assert(!new_relfrozen_xid && !new_relmin_mxid);
491 : 134685 : prstate->pagefrz.FreezePageRelminMxid = InvalidMultiXactId;
492 : 134685 : prstate->pagefrz.NoFreezePageRelminMxid = InvalidMultiXactId;
493 : 134685 : prstate->pagefrz.FreezePageRelfrozenXid = InvalidTransactionId;
494 : 134685 : prstate->pagefrz.NoFreezePageRelfrozenXid = InvalidTransactionId;
495 : : }
496 : :
497 : 643908 : prstate->ndeleted = 0;
498 : 643908 : prstate->live_tuples = 0;
499 : 643908 : prstate->recently_dead_tuples = 0;
500 : 643908 : prstate->hastup = false;
501 : 643908 : prstate->lpdead_items = 0;
502 : :
503 : : /*
504 : : * deadoffsets are filled in during pruning but are only used to populate
505 : : * PruneFreezeResult->deadoffsets. To avoid needing two copies of the
506 : : * array, just save a pointer to the result offsets array in the
507 : : * PruneState.
508 : : */
509 : 643908 : prstate->deadoffsets = presult->deadoffsets;
510 : :
511 : : /*
512 : : * We track whether the page will be all-visible/all-frozen at the end of
513 : : * pruning and freezing. While examining tuple visibility, we'll set
514 : : * set_all_visible to false if there are tuples on the page not visible to
515 : : * all running and future transactions. If setting the VM is enabled for
516 : : * this scan, we will do so if the page ends up being all-visible.
517 : : *
518 : : * We also keep track of the newest live XID, which is used to calculate
519 : : * the snapshot conflict horizon for a WAL record setting the VM.
520 : : */
521 : 643908 : prstate->set_all_visible = prstate->attempt_set_vm;
522 : 643908 : prstate->newest_live_xid = InvalidTransactionId;
523 : :
524 : : /*
525 : : * Currently, only VACUUM performs freezing, but other callers may in the
526 : : * future. We must initialize set_all_frozen based on whether or not the
527 : : * caller passed HEAP_PAGE_PRUNE_FREEZE, because if they did not, we won't
528 : : * call heap_prepare_freeze_tuple() for each tuple, and set_all_frozen
529 : : * will never be cleared for tuples that need freezing. This would lead to
530 : : * incorrectly setting the visibility map all-frozen for this page. We
531 : : * can't set the page all-frozen in the VM if the caller didn't pass
532 : : * HEAP_PAGE_PRUNE_SET_VM.
533 : : *
534 : : * When freezing is not required (no XIDs/MXIDs older than the freeze
535 : : * cutoff), we may still choose to "opportunistically" freeze if doing so
536 : : * would make the page all-frozen.
537 : : *
538 : : * We will not be able to freeze the whole page at the end of vacuum if
539 : : * there are tuples present that are not visible to everyone or if there
540 : : * are dead tuples which will not be removable. However, dead tuples that
541 : : * will be removed by the end of vacuum should not prevent this
542 : : * opportunistic freezing.
543 : : *
544 : : * Therefore, we do not clear set_all_visible and set_all_frozen when we
545 : : * encounter LP_DEAD items. Instead, we correct them after deciding
546 : : * whether to freeze, but before updating the VM, to avoid setting the VM
547 : : * bits incorrectly.
548 : : */
549 [ + + + - ]: 643908 : prstate->set_all_frozen = prstate->attempt_freeze && prstate->attempt_set_vm;
550 : 643908 : }
551 : :
552 : : /*
553 : : * Helper for heap_page_prune_and_freeze(). Iterates over every tuple on the
554 : : * page, examines its visibility information, and determines the appropriate
555 : : * action for each tuple. All tuples are processed and classified during this
556 : : * phase, but no modifications are made to the page until the later execution
557 : : * stage.
558 : : *
559 : : * *off_loc is used for error callback and cleared before returning.
560 : : */
561 : : static void
562 : 450086 : prune_freeze_plan(PruneState *prstate, OffsetNumber *off_loc)
563 : : {
564 : 450086 : Page page = prstate->page;
565 : 450086 : BlockNumber blockno = prstate->block;
566 : 450086 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
567 : : OffsetNumber offnum;
568 : : HeapTupleData tup;
569 : :
570 : 450086 : tup.t_tableOid = RelationGetRelid(prstate->relation);
571 : :
572 : : /*
573 : : * Determine HTSV for all tuples, and queue them up for processing as HOT
574 : : * chain roots or as heap-only items.
575 : : *
576 : : * Determining HTSV only once for each tuple is required for correctness,
577 : : * to deal with cases where running HTSV twice could result in different
578 : : * results. For example, RECENTLY_DEAD can turn to DEAD if another
579 : : * checked item causes GlobalVisTestIsRemovableFullXid() to update the
580 : : * horizon, or INSERT_IN_PROGRESS can change to DEAD if the inserting
581 : : * transaction aborts.
582 : : *
583 : : * It's also good for performance. Most commonly tuples within a page are
584 : : * stored at decreasing offsets (while the items are stored at increasing
585 : : * offsets). When processing all tuples on a page this leads to reading
586 : : * memory at decreasing offsets within a page, with a variable stride.
587 : : * That's hard for CPU prefetchers to deal with. Processing the items in
588 : : * reverse order (and thus the tuples in increasing order) increases
589 : : * prefetching efficiency significantly / decreases the number of cache
590 : : * misses.
591 : : */
592 : 450086 : for (offnum = maxoff;
593 [ + + ]: 30012212 : offnum >= FirstOffsetNumber;
594 : 29562126 : offnum = OffsetNumberPrev(offnum))
595 : : {
596 : 29562126 : ItemId itemid = PageGetItemId(page, offnum);
597 : : HeapTupleHeader htup;
598 : :
599 : : /*
600 : : * Set the offset number so that we can display it along with any
601 : : * error that occurred while processing this tuple.
602 : : */
603 : 29562126 : *off_loc = offnum;
604 : :
605 : 29562126 : prstate->processed[offnum] = false;
606 : 29562126 : prstate->htsv[offnum] = -1;
607 : :
608 : : /* Nothing to do if slot doesn't contain a tuple */
609 [ + + ]: 29562126 : if (!ItemIdIsUsed(itemid))
610 : : {
611 : 217451 : heap_prune_record_unchanged_lp_unused(prstate, offnum);
612 : 217451 : continue;
613 : : }
614 : :
615 [ + + ]: 29344675 : if (ItemIdIsDead(itemid))
616 : : {
617 : : /*
618 : : * If the caller set mark_unused_now true, we can set dead line
619 : : * pointers LP_UNUSED now.
620 : : */
621 [ + + ]: 1688334 : if (unlikely(prstate->mark_unused_now))
622 : 2243 : heap_prune_record_unused(prstate, offnum, false);
623 : : else
624 : 1686091 : heap_prune_record_unchanged_lp_dead(prstate, offnum);
625 : 1688334 : continue;
626 : : }
627 : :
628 [ + + ]: 27656341 : if (ItemIdIsRedirected(itemid))
629 : : {
630 : : /* This is the start of a HOT chain */
631 : 200356 : prstate->root_items[prstate->nroot_items++] = offnum;
632 : 200356 : continue;
633 : : }
634 : :
635 : : Assert(ItemIdIsNormal(itemid));
636 : :
637 : : /*
638 : : * Get the tuple's visibility status and queue it up for processing.
639 : : */
640 : 27455985 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
641 : 27455985 : tup.t_data = htup;
642 : 27455985 : tup.t_len = ItemIdGetLength(itemid);
643 : 27455985 : ItemPointerSet(&tup.t_self, blockno, offnum);
644 : :
645 : 27455985 : prstate->htsv[offnum] = heap_prune_satisfies_vacuum(prstate, &tup);
646 : :
647 [ + + ]: 27455985 : if (!HeapTupleHeaderIsHeapOnly(htup))
648 : 27113810 : prstate->root_items[prstate->nroot_items++] = offnum;
649 : : else
650 : 342175 : prstate->heaponly_items[prstate->nheaponly_items++] = offnum;
651 : : }
652 : :
653 : : /*
654 : : * Process HOT chains.
655 : : *
656 : : * We added the items to the array starting from 'maxoff', so by
657 : : * processing the array in reverse order, we process the items in
658 : : * ascending offset number order. The order doesn't matter for
659 : : * correctness, but some quick micro-benchmarking suggests that this is
660 : : * faster. (Earlier PostgreSQL versions, which scanned all the items on
661 : : * the page instead of using the root_items array, also did it in
662 : : * ascending offset number order.)
663 : : */
664 [ + + ]: 27764252 : for (int i = prstate->nroot_items - 1; i >= 0; i--)
665 : : {
666 : 27314166 : offnum = prstate->root_items[i];
667 : :
668 : : /* Ignore items already processed as part of an earlier chain */
669 [ - + ]: 27314166 : if (prstate->processed[offnum])
670 : 0 : continue;
671 : :
672 : : /* see preceding loop */
673 : 27314166 : *off_loc = offnum;
674 : :
675 : : /* Process this item or chain of items */
676 : 27314166 : heap_prune_chain(maxoff, offnum, prstate);
677 : : }
678 : :
679 : : /*
680 : : * Process any heap-only tuples that were not already processed as part of
681 : : * a HOT chain.
682 : : */
683 [ + + ]: 792261 : for (int i = prstate->nheaponly_items - 1; i >= 0; i--)
684 : : {
685 : 342175 : offnum = prstate->heaponly_items[i];
686 : :
687 [ + + ]: 342175 : if (prstate->processed[offnum])
688 : 325717 : continue;
689 : :
690 : : /* see preceding loop */
691 : 16458 : *off_loc = offnum;
692 : :
693 : : /*
694 : : * If the tuple is DEAD and doesn't chain to anything else, mark it
695 : : * unused. (If it does chain, we can only remove it as part of
696 : : * pruning its chain.)
697 : : *
698 : : * We need this primarily to handle aborted HOT updates, that is,
699 : : * XMIN_INVALID heap-only tuples. Those might not be linked to by any
700 : : * chain, since the parent tuple might be re-updated before any
701 : : * pruning occurs. So we have to be able to reap them separately from
702 : : * chain-pruning. (Note that HeapTupleHeaderIsHotUpdated will never
703 : : * return true for an XMIN_INVALID tuple, so this code will work even
704 : : * when there were sequential updates within the aborted transaction.)
705 : : */
706 [ + + ]: 16458 : if (prstate->htsv[offnum] == HEAPTUPLE_DEAD)
707 : : {
708 : 3262 : ItemId itemid = PageGetItemId(page, offnum);
709 : 3262 : HeapTupleHeader htup = (HeapTupleHeader) PageGetItem(page, itemid);
710 : :
711 [ + - ]: 3262 : if (likely(!HeapTupleHeaderIsHotUpdated(htup)))
712 : : {
713 : 3262 : HeapTupleHeaderAdvanceConflictHorizon(htup,
714 : : &prstate->latest_xid_removed);
715 : 3262 : heap_prune_record_unused(prstate, offnum, true);
716 : : }
717 : : else
718 : : {
719 : : /*
720 : : * This tuple should've been processed and removed as part of
721 : : * a HOT chain, so something's wrong. To preserve evidence,
722 : : * we don't dare to remove it. We cannot leave behind a DEAD
723 : : * tuple either, because that will cause VACUUM to error out.
724 : : * Throwing an error with a distinct error message seems like
725 : : * the least bad option.
726 : : */
727 [ # # ]: 0 : elog(ERROR, "dead heap-only tuple (%u, %d) is not linked to from any HOT chain",
728 : : blockno, offnum);
729 : : }
730 : : }
731 : : else
732 : 13196 : heap_prune_record_unchanged_lp_normal(prstate, offnum);
733 : : }
734 : :
735 : : /* We should now have processed every tuple exactly once */
736 : : #ifdef USE_ASSERT_CHECKING
737 : : for (offnum = FirstOffsetNumber;
738 : : offnum <= maxoff;
739 : : offnum = OffsetNumberNext(offnum))
740 : : {
741 : : *off_loc = offnum;
742 : :
743 : : Assert(prstate->processed[offnum]);
744 : : }
745 : : #endif
746 : :
747 : : /* Clear the offset information once we have processed the given page. */
748 : 450086 : *off_loc = InvalidOffsetNumber;
749 : 450086 : }
750 : :
751 : : /*
752 : : * Decide whether to proceed with freezing according to the freeze plans
753 : : * prepared for the current heap buffer. If freezing is chosen, this function
754 : : * performs several pre-freeze checks.
755 : : *
756 : : * The values of do_prune, do_hint_prune, and did_tuple_hint_fpi must be
757 : : * determined before calling this function.
758 : : *
759 : : * prstate is both an input and output parameter.
760 : : *
761 : : * Returns true if we should apply the freeze plans and freeze tuples on the
762 : : * page, and false otherwise.
763 : : */
764 : : static bool
765 : 450086 : heap_page_will_freeze(bool did_tuple_hint_fpi,
766 : : bool do_prune,
767 : : bool do_hint_prune,
768 : : PruneState *prstate)
769 : : {
770 : 450086 : bool do_freeze = false;
771 : :
772 : : /*
773 : : * If the caller specified we should not attempt to freeze any tuples,
774 : : * validate that everything is in the right state and return.
775 : : */
776 [ + + ]: 450086 : if (!prstate->attempt_freeze)
777 : : {
778 : : Assert(!prstate->set_all_frozen && prstate->nfrozen == 0);
779 : 134685 : return false;
780 : : }
781 : :
782 [ + + ]: 315401 : if (prstate->pagefrz.freeze_required)
783 : : {
784 : : /*
785 : : * heap_prepare_freeze_tuple indicated that at least one XID/MXID from
786 : : * before FreezeLimit/MultiXactCutoff is present. Must freeze to
787 : : * advance relfrozenxid/relminmxid.
788 : : */
789 : 23513 : do_freeze = true;
790 : : }
791 : : else
792 : : {
793 : : /*
794 : : * Opportunistically freeze the page if we are generating an FPI
795 : : * anyway and if doing so means that we can set the page all-frozen
796 : : * afterwards (might not happen until VACUUM's final heap pass).
797 : : *
798 : : * XXX: Previously, we knew if pruning emitted an FPI by checking
799 : : * pgWalUsage.wal_fpi before and after pruning. Once the freeze and
800 : : * prune records were combined, this heuristic couldn't be used
801 : : * anymore. The opportunistic freeze heuristic must be improved;
802 : : * however, for now, try to approximate the old logic.
803 : : */
804 [ + + + + ]: 291888 : if (prstate->set_all_frozen && prstate->nfrozen > 0)
805 : : {
806 : : Assert(prstate->set_all_visible);
807 : :
808 : : /*
809 : : * Freezing would make the page all-frozen. Have already emitted
810 : : * an FPI or will do so anyway?
811 : : */
812 [ + + + + : 23745 : if (RelationNeedsWAL(prstate->relation))
+ - + - ]
813 : : {
814 [ + + ]: 21704 : if (did_tuple_hint_fpi)
815 : 1330 : do_freeze = true;
816 [ + + ]: 20374 : else if (do_prune)
817 : : {
818 [ + + ]: 2261 : if (XLogCheckBufferNeedsBackup(prstate->buffer))
819 : 749 : do_freeze = true;
820 : : }
821 [ + + ]: 18113 : else if (do_hint_prune)
822 : : {
823 [ + + + - : 22588 : if (XLogHintBitIsNeeded() &&
+ + ]
824 : 11294 : XLogCheckBufferNeedsBackup(prstate->buffer))
825 : 1897 : do_freeze = true;
826 : : }
827 : : }
828 : : }
829 : : }
830 : :
831 [ + + ]: 315401 : if (do_freeze)
832 : : {
833 : : /*
834 : : * Validate the tuples we will be freezing before entering the
835 : : * critical section.
836 : : */
837 : 27489 : heap_pre_freeze_checks(prstate->buffer, prstate->frozen, prstate->nfrozen);
838 : : Assert(TransactionIdPrecedes(prstate->pagefrz.FreezePageConflictXid,
839 : : prstate->cutoffs->OldestXmin));
840 : : }
841 [ + + ]: 287912 : else if (prstate->nfrozen > 0)
842 : : {
843 : : /*
844 : : * The page contained some tuples that were not already frozen, and we
845 : : * chose not to freeze them now. The page won't be all-frozen then.
846 : : */
847 : : Assert(!prstate->pagefrz.freeze_required);
848 : :
849 : 20164 : prstate->set_all_frozen = false;
850 : 20164 : prstate->nfrozen = 0; /* avoid miscounts in instrumentation */
851 : : }
852 : : else
853 : : {
854 : : /*
855 : : * We have no freeze plans to execute. The page might already be
856 : : * all-frozen (perhaps only following pruning), though. Such pages
857 : : * can be marked all-frozen in the VM by our caller, even though none
858 : : * of its tuples were newly frozen here.
859 : : */
860 : : }
861 : :
862 : 315401 : return do_freeze;
863 : : }
864 : :
865 : : /*
866 : : * Emit a warning about and fix visibility map corruption on the given page.
867 : : *
868 : : * The caller specifies the type of corruption it has already detected via
869 : : * corruption_type, so that we can emit the appropriate warning. All cases
870 : : * result in the VM bits being cleared; corruption types where PD_ALL_VISIBLE
871 : : * is incorrectly set also clear PD_ALL_VISIBLE.
872 : : *
873 : : * Must be called while holding an exclusive lock on the heap buffer. Dead
874 : : * items and not all-visible tuples must have been discovered under that same
875 : : * lock. Although we do not hold a lock on the VM buffer, it is pinned, and
876 : : * the heap buffer is exclusively locked, ensuring that no other backend can
877 : : * update the VM bits corresponding to this heap page.
878 : : *
879 : : * This function makes changes to the VM and, potentially, the heap page, but
880 : : * it does not need to be done in a critical section.
881 : : */
882 : : static void
883 : 0 : heap_page_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum,
884 : : VMCorruptionType corruption_type)
885 : : {
886 : 0 : const char *relname = RelationGetRelationName(prstate->relation);
887 : 0 : bool do_clear_vm = false;
888 : 0 : bool do_clear_heap = false;
889 : :
890 : : Assert(BufferIsLockedByMeInMode(prstate->buffer, BUFFER_LOCK_EXCLUSIVE));
891 : :
892 [ # # # # ]: 0 : switch (corruption_type)
893 : : {
894 : 0 : case VM_CORRUPT_LPDEAD:
895 [ # # ]: 0 : ereport(WARNING,
896 : : (errcode(ERRCODE_DATA_CORRUPTED),
897 : : errmsg("dead line pointer found on page marked all-visible"),
898 : : errcontext("relation \"%s\", page %u, tuple %u",
899 : : relname, prstate->block, offnum)));
900 : 0 : do_clear_vm = true;
901 : 0 : do_clear_heap = true;
902 : 0 : break;
903 : :
904 : 0 : case VM_CORRUPT_TUPLE_VISIBILITY:
905 : :
906 : : /*
907 : : * A HEAPTUPLE_LIVE tuple on an all-visible page can appear to not
908 : : * be visible to everyone when
909 : : * GetOldestNonRemovableTransactionId() returns a conservative
910 : : * value that's older than the real safe xmin. That is not
911 : : * corruption -- the PD_ALL_VISIBLE flag is still correct.
912 : : *
913 : : * However, dead tuple versions, in-progress inserts, and
914 : : * in-progress deletes should never appear on a page marked
915 : : * all-visible. That indicates real corruption. PD_ALL_VISIBLE
916 : : * should have been cleared by the DML operation that deleted or
917 : : * inserted the tuple.
918 : : */
919 [ # # ]: 0 : ereport(WARNING,
920 : : (errcode(ERRCODE_DATA_CORRUPTED),
921 : : errmsg("tuple not visible to all transactions found on page marked all-visible"),
922 : : errcontext("relation \"%s\", page %u, tuple %u",
923 : : relname, prstate->block, offnum)));
924 : 0 : do_clear_vm = true;
925 : 0 : do_clear_heap = true;
926 : 0 : break;
927 : :
928 : 0 : case VM_CORRUPT_MISSING_PAGE_HINT:
929 : :
930 : : /*
931 : : * As of PostgreSQL 9.2, the visibility map bit should never be
932 : : * set if the page-level bit is clear. However, for vacuum, it's
933 : : * possible that the bit got cleared after
934 : : * heap_vac_scan_next_block() was called, so we must recheck now
935 : : * that we have the buffer lock before concluding that the VM is
936 : : * corrupt.
937 : : */
938 : : Assert(!PageIsAllVisible(prstate->page));
939 : : Assert(prstate->old_vmbits & VISIBILITYMAP_VALID_BITS);
940 [ # # ]: 0 : ereport(WARNING,
941 : : (errcode(ERRCODE_DATA_CORRUPTED),
942 : : errmsg("page is not marked all-visible but visibility map bit is set"),
943 : : errcontext("relation \"%s\", page %u",
944 : : relname, prstate->block)));
945 : 0 : do_clear_vm = true;
946 : 0 : break;
947 : : }
948 : :
949 : : Assert(do_clear_heap || do_clear_vm);
950 : :
951 : : /* Avoid marking the buffer dirty if PD_ALL_VISIBLE is already clear */
952 [ # # ]: 0 : if (do_clear_heap)
953 : : {
954 : : Assert(PageIsAllVisible(prstate->page));
955 : 0 : PageClearAllVisible(prstate->page);
956 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
957 : : }
958 : :
959 [ # # ]: 0 : if (do_clear_vm)
960 : : {
961 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_EXCLUSIVE);
962 : : /* This VM clear is not WAL-logged, so its return value is not needed. */
963 : 0 : (void) visibilitymap_clear(prstate->relation->rd_locator,
964 : : prstate->block, prstate->vmbuffer,
965 : : VISIBILITYMAP_VALID_BITS);
966 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_UNLOCK);
967 : 0 : prstate->old_vmbits = 0;
968 : : }
969 : 0 : }
970 : :
971 : : /*
972 : : * Decide whether to set the visibility map bits (all-visible and all-frozen)
973 : : * for the current page using information from the PruneState and VM.
974 : : *
975 : : * This function does not actually set the VM bits or page-level visibility
976 : : * hint, PD_ALL_VISIBLE.
977 : : *
978 : : * This should be called only after do_freeze has been decided (and do_prune
979 : : * has been set), as these factor into our heuristic-based decision.
980 : : *
981 : : * Returns true if one or both VM bits should be set and false otherwise.
982 : : */
983 : : static bool
984 : 450086 : heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
985 : : bool do_prune, bool do_freeze)
986 : : {
987 [ + + ]: 450086 : if (!prstate->attempt_set_vm)
988 : 97237 : return false;
989 : :
990 [ + + ]: 352849 : if (!prstate->set_all_visible)
991 : 272978 : return false;
992 : :
993 : : /*
994 : : * If this is an on-access call and we're not actually pruning, avoid
995 : : * setting the visibility map if it would newly dirty the heap page or, if
996 : : * the page is already dirty, if doing so would require including a
997 : : * full-page image (FPI) of the heap page in the WAL.
998 : : */
999 [ + + + + : 79871 : if (reason == PRUNE_ON_ACCESS && !do_prune && !do_freeze &&
+ - ]
1000 [ + + + + ]: 29358 : (!BufferIsDirty(prstate->buffer) || XLogCheckBufferNeedsBackup(prstate->buffer)))
1001 : : {
1002 : 15115 : prstate->set_all_visible = prstate->set_all_frozen = false;
1003 : 15115 : return false;
1004 : : }
1005 : :
1006 : 64756 : prstate->new_vmbits = VISIBILITYMAP_ALL_VISIBLE;
1007 : :
1008 [ + + ]: 64756 : if (prstate->set_all_frozen)
1009 : 34835 : prstate->new_vmbits |= VISIBILITYMAP_ALL_FROZEN;
1010 : :
1011 [ + + ]: 64756 : if (prstate->new_vmbits == prstate->old_vmbits)
1012 : : {
1013 : 1820 : prstate->new_vmbits = 0;
1014 : 1820 : return false;
1015 : : }
1016 : :
1017 : 62936 : return true;
1018 : : }
1019 : :
1020 : : /*
1021 : : * If the page is already all-frozen, or already all-visible and freezing
1022 : : * won't be attempted, there is no remaining work and we can use the fast path
1023 : : * to avoid the expensive overhead of heap_page_prune_and_freeze().
1024 : : *
1025 : : * This can happen when the page has a stale prune hint, or if VACUUM is
1026 : : * scanning an already all-frozen page due to SKIP_PAGES_THRESHOLD.
1027 : : *
1028 : : * The caller must already have examined the visibility map and saved the
1029 : : * status of the page's VM bits in prstate->old_vmbits. Caller must hold a
1030 : : * content lock on the heap page since it will examine line pointers.
1031 : : *
1032 : : * Before calling prune_freeze_fast_path(), the caller should first
1033 : : * check for and fix any discrepancy between the page-level visibility hint
1034 : : * and the visibility map. Otherwise, the fast path will always prevent us
1035 : : * from getting them in sync. Note that if there are tuples on the page that
1036 : : * are not visible to all but the VM is incorrectly marked
1037 : : * all-visible/all-frozen, we will not get the chance to fix that corruption
1038 : : * when using the fast path.
1039 : : */
1040 : : static void
1041 : 193822 : prune_freeze_fast_path(PruneState *prstate, PruneFreezeResult *presult)
1042 : : {
1043 : 193822 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
1044 : 193822 : Page page = prstate->page;
1045 : :
1046 : : Assert((prstate->old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1047 : : ((prstate->old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1048 : : !prstate->attempt_freeze));
1049 : :
1050 : : /* We'll fill in presult for the caller */
1051 : 193822 : memset(presult, 0, sizeof(PruneFreezeResult));
1052 : :
1053 : : /* Clear any stale prune hint */
1054 [ - + ]: 193822 : if (TransactionIdIsValid(PageGetPruneXid(page)))
1055 : : {
1056 : 0 : PageClearPrunable(page);
1057 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
1058 : : }
1059 : :
1060 [ - + ]: 193822 : if (PageIsEmpty(page))
1061 : 0 : return;
1062 : :
1063 : : /*
1064 : : * Since the page is all-visible, a count of the normal ItemIds on the
1065 : : * page should be sufficient for vacuum's live tuple count.
1066 : : */
1067 : 193822 : for (OffsetNumber off = FirstOffsetNumber;
1068 [ + + ]: 11131296 : off <= maxoff;
1069 : 10937474 : off = OffsetNumberNext(off))
1070 : : {
1071 : 10937474 : ItemId lp = PageGetItemId(page, off);
1072 : :
1073 [ + + ]: 10937474 : if (!ItemIdIsUsed(lp))
1074 : 260268 : continue;
1075 : :
1076 : 10677206 : presult->hastup = true;
1077 : :
1078 [ + + ]: 10677206 : if (ItemIdIsNormal(lp))
1079 : 10509358 : prstate->live_tuples++;
1080 : : }
1081 : :
1082 : 193822 : presult->live_tuples = prstate->live_tuples;
1083 : : }
1084 : :
1085 : : /*
1086 : : * Prune and repair fragmentation and potentially freeze tuples on the
1087 : : * specified page. If the page's visibility status has changed, update it in
1088 : : * the VM.
1089 : : *
1090 : : * Caller must have pin and buffer cleanup lock on the page. Note that we
1091 : : * don't update the FSM information for page on caller's behalf. Caller might
1092 : : * also need to account for a reduction in the length of the line pointer
1093 : : * array following array truncation by us.
1094 : : *
1095 : : * params contains the input parameters used to control freezing and pruning
1096 : : * behavior. See the definition of PruneFreezeParams for more on what each
1097 : : * parameter does.
1098 : : *
1099 : : * If the HEAP_PAGE_PRUNE_FREEZE option is set in params, we will freeze
1100 : : * tuples if it's required in order to advance relfrozenxid / relminmxid, or
1101 : : * if it's considered advantageous for overall system performance to do so
1102 : : * now. The 'params.cutoffs', 'presult', 'new_relfrozen_xid' and
1103 : : * 'new_relmin_mxid' arguments are required when freezing.
1104 : : *
1105 : : * A vmbuffer corresponding to the heap page is also passed and if the page is
1106 : : * found to be all-visible/all-frozen, we will set it in the VM.
1107 : : *
1108 : : * presult contains output parameters needed by callers, such as the number of
1109 : : * tuples removed and the offsets of dead items on the page after pruning.
1110 : : * heap_page_prune_and_freeze() is responsible for initializing it. Required
1111 : : * by all callers.
1112 : : *
1113 : : * off_loc is the offset location required by the caller to use in error
1114 : : * callback.
1115 : : *
1116 : : * new_relfrozen_xid and new_relmin_mxid must be provided by the caller if the
1117 : : * HEAP_PAGE_PRUNE_FREEZE option is set in params. On entry, they contain the
1118 : : * oldest XID and multi-XID seen on the relation so far. They will be updated
1119 : : * with the oldest values present on the page after pruning. After processing
1120 : : * the whole relation, VACUUM can use these values as the new
1121 : : * relfrozenxid/relminmxid for the relation.
1122 : : */
1123 : : void
1124 : 643908 : heap_page_prune_and_freeze(PruneFreezeParams *params,
1125 : : PruneFreezeResult *presult,
1126 : : OffsetNumber *off_loc,
1127 : : TransactionId *new_relfrozen_xid,
1128 : : MultiXactId *new_relmin_mxid)
1129 : : {
1130 : : PruneState prstate;
1131 : : bool do_freeze;
1132 : : bool do_prune;
1133 : : bool do_hint_prune;
1134 : : bool do_set_vm;
1135 : : bool did_tuple_hint_fpi;
1136 : 643908 : int64 fpi_before = pgWalUsage.wal_fpi;
1137 : : TransactionId conflict_xid;
1138 : :
1139 : : /* Initialize prstate */
1140 : 643908 : prune_freeze_setup(params,
1141 : : new_relfrozen_xid, new_relmin_mxid,
1142 : : presult, &prstate);
1143 : :
1144 : : /*
1145 : : * If the VM is set but PD_ALL_VISIBLE is clear, fix that corruption
1146 : : * before pruning and freezing so that the page and VM start out in a
1147 : : * consistent state.
1148 : : */
1149 [ + + ]: 643908 : if ((prstate.old_vmbits & VISIBILITYMAP_VALID_BITS) &&
1150 [ - + ]: 199283 : !PageIsAllVisible(prstate.page))
1151 : 0 : heap_page_fix_vm_corruption(&prstate, InvalidOffsetNumber,
1152 : : VM_CORRUPT_MISSING_PAGE_HINT);
1153 : :
1154 : : /*
1155 : : * If the page is already all-frozen, or already all-visible when freezing
1156 : : * is not being attempted, take the fast path, skipping pruning and
1157 : : * freezing code entirely. This must be done after fixing any discrepancy
1158 : : * between the page-level visibility hint and the VM, since that may have
1159 : : * cleared old_vmbits.
1160 : : */
1161 [ + + ]: 643908 : if ((params->options & HEAP_PAGE_PRUNE_ALLOW_FAST_PATH) != 0 &&
1162 [ + + ]: 642616 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1163 [ + + ]: 448794 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1164 [ - + ]: 5016 : !prstate.attempt_freeze)))
1165 : : {
1166 : 193822 : prune_freeze_fast_path(&prstate, presult);
1167 : 193822 : return;
1168 : : }
1169 : :
1170 : : /*
1171 : : * Examine all line pointers and tuple visibility information to determine
1172 : : * which line pointers should change state and which tuples may be frozen.
1173 : : * Prepare queue of state changes to later be executed in a critical
1174 : : * section.
1175 : : */
1176 : 450086 : prune_freeze_plan(&prstate, off_loc);
1177 : :
1178 : : /*
1179 : : * After processing all the live tuples on the page, if the newest xmin
1180 : : * amongst them may be considered running by any snapshot, the page cannot
1181 : : * be all-visible. This should be done before determining whether or not
1182 : : * to opportunistically freeze.
1183 : : */
1184 [ + + ]: 450086 : if (prstate.set_all_visible &&
1185 [ + + + + ]: 192154 : TransactionIdIsNormal(prstate.newest_live_xid) &&
1186 : 81734 : GlobalVisTestXidConsideredRunning(prstate.vistest,
1187 : : prstate.newest_live_xid,
1188 : : true))
1189 : 2877 : prstate.set_all_visible = prstate.set_all_frozen = false;
1190 : :
1191 : : /*
1192 : : * If checksums are enabled, calling heap_prune_satisfies_vacuum() while
1193 : : * checking tuple visibility information in prune_freeze_plan() may have
1194 : : * caused an FPI to be emitted.
1195 : : */
1196 : 450086 : did_tuple_hint_fpi = fpi_before != pgWalUsage.wal_fpi;
1197 : :
1198 : 1332433 : do_prune = prstate.nredirected > 0 ||
1199 [ + + + + ]: 828959 : prstate.ndead > 0 ||
1200 [ + + ]: 378873 : prstate.nunused > 0;
1201 : :
1202 : : /*
1203 : : * Even if we don't prune anything, if we found a new value for the
1204 : : * pd_prune_xid field or the page was marked full, we will update the hint
1205 : : * bit.
1206 : : */
1207 [ + + + + ]: 721498 : do_hint_prune = PageGetPruneXid(prstate.page) != prstate.new_prune_xid ||
1208 : 271412 : PageIsFull(prstate.page);
1209 : :
1210 : : /*
1211 : : * Decide if we want to go ahead with freezing according to the freeze
1212 : : * plans we prepared, or not.
1213 : : */
1214 : 450086 : do_freeze = heap_page_will_freeze(did_tuple_hint_fpi,
1215 : : do_prune,
1216 : : do_hint_prune,
1217 : : &prstate);
1218 : :
1219 : : /*
1220 : : * While scanning the line pointers, we did not clear
1221 : : * set_all_visible/set_all_frozen when encountering LP_DEAD items because
1222 : : * we wanted the decision whether or not to freeze the page to be
1223 : : * unaffected by the short-term presence of LP_DEAD items. These LP_DEAD
1224 : : * items are effectively assumed to be LP_UNUSED items in the making. It
1225 : : * doesn't matter which vacuum heap pass (initial pass or final pass) ends
1226 : : * up setting the page all-frozen, as long as the ongoing VACUUM does it.
1227 : : *
1228 : : * Now that we finished determining whether or not to freeze the page,
1229 : : * update set_all_visible and set_all_frozen so that they reflect the true
1230 : : * state of the page for setting PD_ALL_VISIBLE and VM bits.
1231 : : */
1232 [ + + ]: 450086 : if (prstate.lpdead_items > 0)
1233 : 74954 : prstate.set_all_visible = prstate.set_all_frozen = false;
1234 : :
1235 : : Assert(!prstate.set_all_frozen || prstate.set_all_visible);
1236 : : Assert(!prstate.set_all_visible || prstate.attempt_set_vm);
1237 : : Assert(!prstate.set_all_visible || (prstate.lpdead_items == 0));
1238 : :
1239 : 450086 : do_set_vm = heap_page_will_set_vm(&prstate, params->reason, do_prune, do_freeze);
1240 : :
1241 : : /*
1242 : : * new_vmbits should be 0 regardless of whether or not the page is
1243 : : * all-visible if we do not intend to set the VM.
1244 : : */
1245 : : Assert(do_set_vm || prstate.new_vmbits == 0);
1246 : :
1247 : : /*
1248 : : * The snapshot conflict horizon for the whole record is the most
1249 : : * conservative (newest) horizon required by any change in the record.
1250 : : */
1251 : 450086 : conflict_xid = InvalidTransactionId;
1252 [ + + ]: 450086 : if (do_set_vm)
1253 : 62936 : conflict_xid = prstate.newest_live_xid;
1254 [ + + + + ]: 450086 : if (do_freeze && TransactionIdFollows(prstate.pagefrz.FreezePageConflictXid, conflict_xid))
1255 : 4536 : conflict_xid = prstate.pagefrz.FreezePageConflictXid;
1256 [ + + + + ]: 450086 : if (do_prune && TransactionIdFollows(prstate.latest_xid_removed, conflict_xid))
1257 : 63242 : conflict_xid = prstate.latest_xid_removed;
1258 : :
1259 : : /* Lock vmbuffer before entering a critical section */
1260 [ + + ]: 450086 : if (do_set_vm)
1261 : 62936 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_EXCLUSIVE);
1262 : :
1263 : : /* Any error while applying the changes is critical */
1264 : 450086 : START_CRIT_SECTION();
1265 : :
1266 [ + + ]: 450086 : if (do_hint_prune)
1267 : : {
1268 : : /*
1269 : : * Update the page's pd_prune_xid field to either zero, or the lowest
1270 : : * XID of any soon-prunable tuple.
1271 : : */
1272 : 178753 : ((PageHeader) prstate.page)->pd_prune_xid = prstate.new_prune_xid;
1273 : :
1274 : : /*
1275 : : * Also clear the "page is full" flag, since there's no point in
1276 : : * repeating the prune/defrag process until something else happens to
1277 : : * the page.
1278 : : */
1279 : 178753 : PageClearFull(prstate.page);
1280 : :
1281 : : /*
1282 : : * If that's all we had to do to the page, this is a non-WAL-logged
1283 : : * hint. If we are going to freeze or prune the page or set
1284 : : * PD_ALL_VISIBLE, we will mark the buffer dirty below.
1285 : : *
1286 : : * Setting PD_ALL_VISIBLE is fully WAL-logged because it is forbidden
1287 : : * for the VM to be set and PD_ALL_VISIBLE to be clear.
1288 : : */
1289 [ + + + + : 178753 : if (!do_freeze && !do_prune && !do_set_vm)
+ + ]
1290 : 68129 : MarkBufferDirtyHint(prstate.buffer, true);
1291 : : }
1292 : :
1293 [ + + + + : 450086 : if (do_prune || do_freeze || do_set_vm)
+ + ]
1294 : : {
1295 : : /* Apply the planned item changes and repair page fragmentation. */
1296 [ + + ]: 136335 : if (do_prune)
1297 : : {
1298 : 71649 : heap_page_prune_execute(prstate.buffer, false,
1299 : : prstate.redirected, prstate.nredirected,
1300 : : prstate.nowdead, prstate.ndead,
1301 : : prstate.nowunused, prstate.nunused);
1302 : : }
1303 : :
1304 [ + + ]: 136335 : if (do_freeze)
1305 : 27489 : heap_freeze_prepared_tuples(prstate.buffer, prstate.frozen, prstate.nfrozen);
1306 : :
1307 : : /* Set the visibility map and page visibility hint */
1308 [ + + ]: 136335 : if (do_set_vm)
1309 : : {
1310 : : /*
1311 : : * While it is valid for PD_ALL_VISIBLE to be set when the
1312 : : * corresponding VM bit is clear, we strongly prefer to keep them
1313 : : * in sync.
1314 : : *
1315 : : * The heap buffer must be marked dirty before adding it to the
1316 : : * WAL chain when setting the VM. We don't worry about
1317 : : * unnecessarily dirtying the heap buffer if PD_ALL_VISIBLE is
1318 : : * already set, though. It is extremely rare to have a clean heap
1319 : : * buffer with PD_ALL_VISIBLE already set and the VM bits clear,
1320 : : * so there is no point in optimizing it.
1321 : : */
1322 : 62936 : PageSetAllVisible(prstate.page);
1323 : 62936 : PageClearPrunable(prstate.page);
1324 : 62936 : (void) visibilitymap_set(prstate.block, prstate.vmbuffer,
1325 : 62936 : prstate.new_vmbits,
1326 : 62936 : prstate.relation->rd_locator);
1327 : : }
1328 : :
1329 : 136335 : MarkBufferDirty(prstate.buffer);
1330 : :
1331 : : /*
1332 : : * Emit a WAL XLOG_HEAP2_PRUNE* record showing what we did
1333 : : */
1334 [ + + + + : 136335 : if (RelationNeedsWAL(prstate.relation))
+ - + - ]
1335 : : {
1336 [ + + + + ]: 194028 : log_heap_prune_and_freeze(prstate.relation, prstate.buffer,
1337 : : do_set_vm ? prstate.vmbuffer : InvalidBuffer,
1338 : 60963 : do_set_vm ? prstate.new_vmbits : 0,
1339 : : conflict_xid,
1340 : : do_prune, /* cleanup lock */
1341 : : params->reason,
1342 : : prstate.frozen, prstate.nfrozen,
1343 : : prstate.redirected, prstate.nredirected,
1344 : : prstate.nowdead, prstate.ndead,
1345 : : prstate.nowunused, prstate.nunused);
1346 : : }
1347 : : }
1348 : :
1349 : 450086 : END_CRIT_SECTION();
1350 : :
1351 [ + + ]: 450086 : if (do_set_vm)
1352 : 62936 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_UNLOCK);
1353 : :
1354 : : /*
1355 : : * During its second pass over the heap, VACUUM calls
1356 : : * heap_page_would_be_all_visible() to determine whether a page is
1357 : : * all-visible and all-frozen. The logic here is similar. After completing
1358 : : * pruning and freezing, use an assertion to verify that our results
1359 : : * remain consistent with heap_page_would_be_all_visible(). It's also a
1360 : : * valuable cross-check of the page state after pruning and freezing.
1361 : : */
1362 : : #ifdef USE_ASSERT_CHECKING
1363 : : if (prstate.set_all_visible)
1364 : : {
1365 : : TransactionId debug_cutoff;
1366 : : bool debug_all_frozen;
1367 : :
1368 : : Assert(prstate.lpdead_items == 0);
1369 : :
1370 : : Assert(heap_page_is_all_visible(prstate.relation, prstate.buffer,
1371 : : prstate.vistest,
1372 : : &debug_all_frozen,
1373 : : &debug_cutoff, off_loc));
1374 : :
1375 : : Assert(!TransactionIdIsValid(debug_cutoff) ||
1376 : : debug_cutoff == prstate.newest_live_xid);
1377 : :
1378 : : /*
1379 : : * It's possible the page is composed entirely of frozen tuples but is
1380 : : * not set all-frozen in the VM and did not pass
1381 : : * HEAP_PAGE_PRUNE_FREEZE. In this case, it's possible
1382 : : * heap_page_is_all_visible() finds the page completely frozen, even
1383 : : * though prstate.set_all_frozen is false.
1384 : : */
1385 : : Assert(!prstate.set_all_frozen || debug_all_frozen);
1386 : : }
1387 : : #endif
1388 : :
1389 : : /* Copy information back for caller */
1390 : 450086 : presult->ndeleted = prstate.ndeleted;
1391 : 450086 : presult->nnewlpdead = prstate.ndead;
1392 : 450086 : presult->nfrozen = prstate.nfrozen;
1393 : 450086 : presult->live_tuples = prstate.live_tuples;
1394 : 450086 : presult->recently_dead_tuples = prstate.recently_dead_tuples;
1395 : 450086 : presult->hastup = prstate.hastup;
1396 : :
1397 : 450086 : presult->lpdead_items = prstate.lpdead_items;
1398 : : /* the presult->deadoffsets array was already filled in */
1399 : :
1400 : 450086 : presult->newly_all_visible = false;
1401 : 450086 : presult->newly_all_frozen = false;
1402 : 450086 : presult->newly_all_visible_frozen = false;
1403 [ + + ]: 450086 : if (do_set_vm)
1404 : : {
1405 [ + + ]: 62936 : if ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
1406 : : {
1407 : 59295 : presult->newly_all_visible = true;
1408 [ + + ]: 59295 : if (prstate.set_all_frozen)
1409 : 30750 : presult->newly_all_visible_frozen = true;
1410 : : }
1411 [ + - ]: 3641 : else if ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
1412 [ + - ]: 3641 : prstate.set_all_frozen)
1413 : 3641 : presult->newly_all_frozen = true;
1414 : : }
1415 : :
1416 [ + + ]: 450086 : if (prstate.attempt_freeze)
1417 : : {
1418 [ + + ]: 315401 : if (presult->nfrozen > 0)
1419 : : {
1420 : 27489 : *new_relfrozen_xid = prstate.pagefrz.FreezePageRelfrozenXid;
1421 : 27489 : *new_relmin_mxid = prstate.pagefrz.FreezePageRelminMxid;
1422 : : }
1423 : : else
1424 : : {
1425 : 287912 : *new_relfrozen_xid = prstate.pagefrz.NoFreezePageRelfrozenXid;
1426 : 287912 : *new_relmin_mxid = prstate.pagefrz.NoFreezePageRelminMxid;
1427 : : }
1428 : : }
1429 : : }
1430 : :
1431 : :
1432 : : /*
1433 : : * Perform visibility checks for heap pruning.
1434 : : */
1435 : : static HTSV_Result
1436 : 27455985 : heap_prune_satisfies_vacuum(PruneState *prstate, HeapTuple tup)
1437 : : {
1438 : : HTSV_Result res;
1439 : : TransactionId dead_after;
1440 : :
1441 : 27455985 : res = HeapTupleSatisfiesVacuumHorizon(tup, prstate->buffer, &dead_after);
1442 : :
1443 [ + + ]: 27455985 : if (res != HEAPTUPLE_RECENTLY_DEAD)
1444 : 23215513 : return res;
1445 : :
1446 : : /*
1447 : : * For VACUUM, we must be sure to prune tuples with xmax older than
1448 : : * OldestXmin -- a visibility cutoff determined at the beginning of
1449 : : * vacuuming the relation. OldestXmin is used for freezing determination
1450 : : * and we cannot freeze dead tuples' xmaxes.
1451 : : */
1452 [ + + ]: 4240472 : if (prstate->cutoffs &&
1453 [ + - ]: 1300702 : TransactionIdIsValid(prstate->cutoffs->OldestXmin) &&
1454 [ + + ]: 1300702 : NormalTransactionIdPrecedes(dead_after, prstate->cutoffs->OldestXmin))
1455 : 962511 : return HEAPTUPLE_DEAD;
1456 : :
1457 : : /*
1458 : : * Determine whether or not the tuple is considered dead when compared
1459 : : * with the provided GlobalVisState. On-access pruning does not provide
1460 : : * VacuumCutoffs. And for vacuum, even if the tuple's xmax is not older
1461 : : * than OldestXmin, GlobalVisTestIsRemovableXid() could find the row dead
1462 : : * if the GlobalVisState has been updated since the beginning of vacuuming
1463 : : * the relation.
1464 : : */
1465 [ + + ]: 3277961 : if (GlobalVisTestIsRemovableXid(prstate->vistest, dead_after, true))
1466 : 2889579 : return HEAPTUPLE_DEAD;
1467 : :
1468 : 388382 : return res;
1469 : : }
1470 : :
1471 : :
1472 : : /*
1473 : : * Pruning calculates tuple visibility once and saves the results in an array
1474 : : * of int8. See PruneState.htsv for details. This helper function is meant
1475 : : * to guard against examining visibility status array members which have not
1476 : : * yet been computed.
1477 : : */
1478 : : static inline HTSV_Result
1479 : 27439527 : htsv_get_valid_status(int status)
1480 : : {
1481 : : Assert(status >= HEAPTUPLE_DEAD &&
1482 : : status <= HEAPTUPLE_DELETE_IN_PROGRESS);
1483 : 27439527 : return (HTSV_Result) status;
1484 : : }
1485 : :
1486 : : /*
1487 : : * Prune specified line pointer or a HOT chain originating at line pointer.
1488 : : *
1489 : : * Tuple visibility information is provided in prstate->htsv.
1490 : : *
1491 : : * If the item is an index-referenced tuple (i.e. not a heap-only tuple),
1492 : : * the HOT chain is pruned by removing all DEAD tuples at the start of the HOT
1493 : : * chain. We also prune any RECENTLY_DEAD tuples preceding a DEAD tuple.
1494 : : * This is OK because a RECENTLY_DEAD tuple preceding a DEAD tuple is really
1495 : : * DEAD, our visibility test is just too coarse to detect it.
1496 : : *
1497 : : * Pruning must never leave behind a DEAD tuple that still has tuple storage.
1498 : : * VACUUM isn't prepared to deal with that case.
1499 : : *
1500 : : * The root line pointer is redirected to the tuple immediately after the
1501 : : * latest DEAD tuple. If all tuples in the chain are DEAD, the root line
1502 : : * pointer is marked LP_DEAD. (This includes the case of a DEAD simple
1503 : : * tuple, which we treat as a chain of length 1.)
1504 : : *
1505 : : * We don't actually change the page here. We just add entries to the arrays in
1506 : : * prstate showing the changes to be made. Items to be redirected are added
1507 : : * to the redirected[] array (two entries per redirection); items to be set to
1508 : : * LP_DEAD state are added to nowdead[]; and items to be set to LP_UNUSED
1509 : : * state are added to nowunused[]. We perform bookkeeping of live tuples,
1510 : : * visibility etc. based on what the page will look like after the changes
1511 : : * applied. All that bookkeeping is performed in the heap_prune_record_*()
1512 : : * subroutines. The division of labor is that heap_prune_chain() decides the
1513 : : * fate of each tuple, ie. whether it's going to be removed, redirected or
1514 : : * left unchanged, and the heap_prune_record_*() subroutines update PruneState
1515 : : * based on that outcome.
1516 : : */
1517 : : static void
1518 : 27314166 : heap_prune_chain(OffsetNumber maxoff, OffsetNumber rootoffnum,
1519 : : PruneState *prstate)
1520 : : {
1521 : 27314166 : TransactionId priorXmax = InvalidTransactionId;
1522 : : ItemId rootlp;
1523 : : OffsetNumber offnum;
1524 : : OffsetNumber chainitems[MaxHeapTuplesPerPage];
1525 : 27314166 : Page page = prstate->page;
1526 : :
1527 : : /*
1528 : : * After traversing the HOT chain, ndeadchain is the index in chainitems
1529 : : * of the first live successor after the last dead item.
1530 : : */
1531 : 27314166 : int ndeadchain = 0,
1532 : 27314166 : nchain = 0;
1533 : :
1534 : 27314166 : rootlp = PageGetItemId(page, rootoffnum);
1535 : :
1536 : : /* Start from the root tuple */
1537 : 27314166 : offnum = rootoffnum;
1538 : :
1539 : : /* while not end of the chain */
1540 : : for (;;)
1541 : 325717 : {
1542 : : HeapTupleHeader htup;
1543 : : ItemId lp;
1544 : :
1545 : : /* Sanity check (pure paranoia) */
1546 [ - + ]: 27639883 : if (offnum < FirstOffsetNumber)
1547 : 0 : break;
1548 : :
1549 : : /*
1550 : : * An offset past the end of page's line pointer array is possible
1551 : : * when the array was truncated (original item must have been unused)
1552 : : */
1553 [ - + ]: 27639883 : if (offnum > maxoff)
1554 : 0 : break;
1555 : :
1556 : : /* If item is already processed, stop --- it must not be same chain */
1557 [ - + ]: 27639883 : if (prstate->processed[offnum])
1558 : 0 : break;
1559 : :
1560 : 27639883 : lp = PageGetItemId(page, offnum);
1561 : :
1562 : : /*
1563 : : * Unused item obviously isn't part of the chain. Likewise, a dead
1564 : : * line pointer can't be part of the chain. Both of those cases were
1565 : : * already marked as processed.
1566 : : */
1567 : : Assert(ItemIdIsUsed(lp));
1568 : : Assert(!ItemIdIsDead(lp));
1569 : :
1570 : : /*
1571 : : * If we are looking at the redirected root line pointer, jump to the
1572 : : * first normal tuple in the chain. If we find a redirect somewhere
1573 : : * else, stop --- it must not be same chain.
1574 : : */
1575 [ + + ]: 27639883 : if (ItemIdIsRedirected(lp))
1576 : : {
1577 [ - + ]: 200356 : if (nchain > 0)
1578 : 0 : break; /* not at start of chain */
1579 : 200356 : chainitems[nchain++] = offnum;
1580 : 200356 : offnum = ItemIdGetRedirect(rootlp);
1581 : 200356 : continue;
1582 : : }
1583 : :
1584 : : Assert(ItemIdIsNormal(lp));
1585 : :
1586 : 27439527 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1587 : :
1588 : : /*
1589 : : * Check the tuple XMIN against prior XMAX, if any
1590 : : */
1591 [ + + - + ]: 27564888 : if (TransactionIdIsValid(priorXmax) &&
1592 : 125361 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
1593 : 0 : break;
1594 : :
1595 : : /*
1596 : : * OK, this tuple is indeed a member of the chain.
1597 : : */
1598 : 27439527 : chainitems[nchain++] = offnum;
1599 : :
1600 [ + + + - ]: 27439527 : switch (htsv_get_valid_status(prstate->htsv[offnum]))
1601 : : {
1602 : 3924081 : case HEAPTUPLE_DEAD:
1603 : :
1604 : : /* Remember the last DEAD tuple seen */
1605 : 3924081 : ndeadchain = nchain;
1606 : 3924081 : HeapTupleHeaderAdvanceConflictHorizon(htup,
1607 : : &prstate->latest_xid_removed);
1608 : : /* Advance to next chain member */
1609 : 3924081 : break;
1610 : :
1611 : 388382 : case HEAPTUPLE_RECENTLY_DEAD:
1612 : :
1613 : : /*
1614 : : * We don't need to advance the conflict horizon for
1615 : : * RECENTLY_DEAD tuples, even if we are removing them. This
1616 : : * is because we only remove RECENTLY_DEAD tuples if they
1617 : : * precede a DEAD tuple, and the DEAD tuple must have been
1618 : : * inserted by a newer transaction than the RECENTLY_DEAD
1619 : : * tuple by virtue of being later in the chain. We will have
1620 : : * advanced the conflict horizon for the DEAD tuple.
1621 : : */
1622 : :
1623 : : /*
1624 : : * Advance past RECENTLY_DEAD tuples just in case there's a
1625 : : * DEAD one after them. We have to make sure that we don't
1626 : : * miss any DEAD tuples, since DEAD tuples that still have
1627 : : * tuple storage after pruning will confuse VACUUM.
1628 : : */
1629 : 388382 : break;
1630 : :
1631 : 23127064 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1632 : : case HEAPTUPLE_LIVE:
1633 : : case HEAPTUPLE_INSERT_IN_PROGRESS:
1634 : 23127064 : goto process_chain;
1635 : :
1636 : 0 : default:
1637 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1638 : : goto process_chain;
1639 : : }
1640 : :
1641 : : /*
1642 : : * If the tuple is not HOT-updated, then we are at the end of this
1643 : : * HOT-update chain.
1644 : : */
1645 [ + + ]: 4312463 : if (!HeapTupleHeaderIsHotUpdated(htup))
1646 : 4187102 : goto process_chain;
1647 : :
1648 : : /* HOT implies it can't have moved to different partition */
1649 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1650 : :
1651 : : /*
1652 : : * Advance to next chain member.
1653 : : */
1654 : : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == prstate->block);
1655 : 125361 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1656 : 125361 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1657 : : }
1658 : :
1659 [ # # # # ]: 0 : if (ItemIdIsRedirected(rootlp) && nchain < 2)
1660 : : {
1661 : : /*
1662 : : * We found a redirect item that doesn't point to a valid follow-on
1663 : : * item. This can happen if the loop in heap_page_prune_and_freeze()
1664 : : * caused us to visit the dead successor of a redirect item before
1665 : : * visiting the redirect item. We can clean up by setting the
1666 : : * redirect item to LP_DEAD state or LP_UNUSED if the caller
1667 : : * indicated.
1668 : : */
1669 : 0 : heap_prune_record_dead_or_unused(prstate, rootoffnum, false);
1670 : 0 : return;
1671 : : }
1672 : :
1673 : 0 : process_chain:
1674 : :
1675 [ + + ]: 27314166 : if (ndeadchain == 0)
1676 : : {
1677 : : /*
1678 : : * No DEAD tuple was found, so the chain is entirely composed of
1679 : : * normal, unchanged tuples. Leave it alone.
1680 : : */
1681 : 23434805 : int i = 0;
1682 : :
1683 [ + + ]: 23434805 : if (ItemIdIsRedirected(rootlp))
1684 : : {
1685 : 178122 : heap_prune_record_unchanged_lp_redirect(prstate, rootoffnum);
1686 : 178122 : i++;
1687 : : }
1688 [ + + ]: 46874938 : for (; i < nchain; i++)
1689 : 23440133 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1690 : : }
1691 [ + + ]: 3879361 : else if (ndeadchain == nchain)
1692 : : {
1693 : : /*
1694 : : * The entire chain is dead. Mark the root line pointer LP_DEAD, and
1695 : : * fully remove the other tuples in the chain.
1696 : : */
1697 : 3805917 : heap_prune_record_dead_or_unused(prstate, rootoffnum, ItemIdIsNormal(rootlp));
1698 [ + + ]: 3850832 : for (int i = 1; i < nchain; i++)
1699 : 44915 : heap_prune_record_unused(prstate, chainitems[i], true);
1700 : : }
1701 : : else
1702 : : {
1703 : : /*
1704 : : * We found a DEAD tuple in the chain. Redirect the root line pointer
1705 : : * to the first non-DEAD tuple, and mark as unused each intermediate
1706 : : * item that we are able to remove from the chain.
1707 : : */
1708 : 73444 : heap_prune_record_redirect(prstate, rootoffnum, chainitems[ndeadchain],
1709 : 73444 : ItemIdIsNormal(rootlp));
1710 [ + + ]: 95483 : for (int i = 1; i < ndeadchain; i++)
1711 : 22039 : heap_prune_record_unused(prstate, chainitems[i], true);
1712 : :
1713 : : /* the rest of tuples in the chain are normal, unchanged tuples */
1714 [ + + ]: 148757 : for (int i = ndeadchain; i < nchain; i++)
1715 : 75313 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1716 : : }
1717 : : }
1718 : :
1719 : : /* Record lowest soon-prunable XID */
1720 : : static void
1721 : 6478520 : heap_prune_record_prunable(PruneState *prstate, TransactionId xid,
1722 : : OffsetNumber offnum)
1723 : : {
1724 : : /*
1725 : : * This should exactly match the PageSetPrunable macro. We can't store
1726 : : * directly into the page header yet, so we update working state.
1727 : : */
1728 : : Assert(TransactionIdIsNormal(xid));
1729 [ + + + + ]: 12685645 : if (!TransactionIdIsValid(prstate->new_prune_xid) ||
1730 : 6207125 : TransactionIdPrecedes(xid, prstate->new_prune_xid))
1731 : 272795 : prstate->new_prune_xid = xid;
1732 : :
1733 : : /*
1734 : : * It's incorrect for a page to be marked all-visible if it contains
1735 : : * prunable items.
1736 : : */
1737 [ - + ]: 6478520 : if (PageIsAllVisible(prstate->page))
1738 : 0 : heap_page_fix_vm_corruption(prstate, offnum,
1739 : : VM_CORRUPT_TUPLE_VISIBILITY);
1740 : 6478520 : }
1741 : :
1742 : : /* Record line pointer to be redirected */
1743 : : static void
1744 : 73444 : heap_prune_record_redirect(PruneState *prstate,
1745 : : OffsetNumber offnum, OffsetNumber rdoffnum,
1746 : : bool was_normal)
1747 : : {
1748 : : Assert(!prstate->processed[offnum]);
1749 : 73444 : prstate->processed[offnum] = true;
1750 : :
1751 : : /*
1752 : : * Do not mark the redirect target here. It needs to be counted
1753 : : * separately as an unchanged tuple.
1754 : : */
1755 : :
1756 : : Assert(prstate->nredirected < MaxHeapTuplesPerPage);
1757 : 73444 : prstate->redirected[prstate->nredirected * 2] = offnum;
1758 : 73444 : prstate->redirected[prstate->nredirected * 2 + 1] = rdoffnum;
1759 : :
1760 : 73444 : prstate->nredirected++;
1761 : :
1762 : : /*
1763 : : * If the root entry had been a normal tuple, we are deleting it, so count
1764 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1765 : : * count.
1766 : : */
1767 [ + + ]: 73444 : if (was_normal)
1768 : 64292 : prstate->ndeleted++;
1769 : :
1770 : 73444 : prstate->hastup = true;
1771 : 73444 : }
1772 : :
1773 : : /* Record line pointer to be marked dead */
1774 : : static void
1775 : 3771276 : heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
1776 : : bool was_normal)
1777 : : {
1778 : : Assert(!prstate->processed[offnum]);
1779 : 3771276 : prstate->processed[offnum] = true;
1780 : :
1781 : : Assert(prstate->ndead < MaxHeapTuplesPerPage);
1782 : 3771276 : prstate->nowdead[prstate->ndead] = offnum;
1783 : 3771276 : prstate->ndead++;
1784 : :
1785 : : /*
1786 : : * Deliberately delay unsetting set_all_visible and set_all_frozen until
1787 : : * later during pruning. Removable dead tuples shouldn't preclude freezing
1788 : : * the page.
1789 : : */
1790 : :
1791 : : /* Record the dead offset for vacuum */
1792 : 3771276 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1793 : :
1794 : : /*
1795 : : * If the root entry had been a normal tuple, we are deleting it, so count
1796 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1797 : : * count.
1798 : : */
1799 [ + + ]: 3771276 : if (was_normal)
1800 : 3758194 : prstate->ndeleted++;
1801 : 3771276 : }
1802 : :
1803 : : /*
1804 : : * Depending on whether or not the caller set mark_unused_now to true, record that a
1805 : : * line pointer should be marked LP_DEAD or LP_UNUSED. There are other cases in
1806 : : * which we will mark line pointers LP_UNUSED, but we will not mark line
1807 : : * pointers LP_DEAD if mark_unused_now is true.
1808 : : */
1809 : : static void
1810 : 3805917 : heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
1811 : : bool was_normal)
1812 : : {
1813 : : /*
1814 : : * If the caller set mark_unused_now to true, we can remove dead tuples
1815 : : * during pruning instead of marking their line pointers dead. Set this
1816 : : * tuple's line pointer LP_UNUSED. We hint that this option is less
1817 : : * likely.
1818 : : */
1819 [ + + ]: 3805917 : if (unlikely(prstate->mark_unused_now))
1820 : 34641 : heap_prune_record_unused(prstate, offnum, was_normal);
1821 : : else
1822 : 3771276 : heap_prune_record_dead(prstate, offnum, was_normal);
1823 : :
1824 : : /*
1825 : : * It's incorrect for the page to be set all-visible if it contains dead
1826 : : * items. Fix that on the heap page and check the VM for corruption as
1827 : : * well. Do that here rather than in heap_prune_record_dead() so we also
1828 : : * cover tuples that are directly marked LP_UNUSED via mark_unused_now.
1829 : : */
1830 [ - + ]: 3805917 : if (PageIsAllVisible(prstate->page))
1831 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
1832 : 3805917 : }
1833 : :
1834 : : /* Record line pointer to be marked unused */
1835 : : static void
1836 : 107100 : heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal)
1837 : : {
1838 : : Assert(!prstate->processed[offnum]);
1839 : 107100 : prstate->processed[offnum] = true;
1840 : :
1841 : : Assert(prstate->nunused < MaxHeapTuplesPerPage);
1842 : 107100 : prstate->nowunused[prstate->nunused] = offnum;
1843 : 107100 : prstate->nunused++;
1844 : :
1845 : : /*
1846 : : * If the root entry had been a normal tuple, we are deleting it, so count
1847 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1848 : : * count.
1849 : : */
1850 [ + + ]: 107100 : if (was_normal)
1851 : 104857 : prstate->ndeleted++;
1852 : 107100 : }
1853 : :
1854 : : /*
1855 : : * Record an unused line pointer that is left unchanged.
1856 : : */
1857 : : static void
1858 : 217451 : heap_prune_record_unchanged_lp_unused(PruneState *prstate, OffsetNumber offnum)
1859 : : {
1860 : : Assert(!prstate->processed[offnum]);
1861 : 217451 : prstate->processed[offnum] = true;
1862 : 217451 : }
1863 : :
1864 : : /*
1865 : : * Record line pointer that is left unchanged. We consider freezing it, and
1866 : : * update bookkeeping of tuple counts and page visibility.
1867 : : */
1868 : : static void
1869 : 23528642 : heap_prune_record_unchanged_lp_normal(PruneState *prstate, OffsetNumber offnum)
1870 : : {
1871 : : HeapTupleHeader htup;
1872 : : TransactionId xmin;
1873 : 23528642 : Page page = prstate->page;
1874 : :
1875 : : Assert(!prstate->processed[offnum]);
1876 : 23528642 : prstate->processed[offnum] = true;
1877 : :
1878 : 23528642 : prstate->hastup = true; /* the page is not empty */
1879 : :
1880 : : /*
1881 : : * The criteria for counting a tuple as live in this block need to match
1882 : : * what analyze.c's acquire_sample_rows() does, otherwise VACUUM and
1883 : : * ANALYZE may produce wildly different reltuples values, e.g. when there
1884 : : * are many recently-dead tuples.
1885 : : *
1886 : : * The logic here is a bit simpler than acquire_sample_rows(), as VACUUM
1887 : : * can't run inside a transaction block, which makes some cases impossible
1888 : : * (e.g. in-progress insert from the same transaction).
1889 : : *
1890 : : * HEAPTUPLE_DEAD are handled by the other heap_prune_record_*()
1891 : : * subroutines. They don't count dead items like acquire_sample_rows()
1892 : : * does, because we assume that all dead items will become LP_UNUSED
1893 : : * before VACUUM finishes. This difference is only superficial. VACUUM
1894 : : * effectively agrees with ANALYZE about DEAD items, in the end. VACUUM
1895 : : * won't remember LP_DEAD items, but only because they're not supposed to
1896 : : * be left behind when it is done. (Cases where we bypass index vacuuming
1897 : : * will violate this optimistic assumption, but the overall impact of that
1898 : : * should be negligible.)
1899 : : */
1900 : 23528642 : htup = (HeapTupleHeader) PageGetItem(page, PageGetItemId(page, offnum));
1901 : :
1902 [ + + + + : 23528642 : switch (prstate->htsv[offnum])
- ]
1903 : : {
1904 : 17050122 : case HEAPTUPLE_LIVE:
1905 : :
1906 : : /*
1907 : : * Count it as live. Not only is this natural, but it's also what
1908 : : * acquire_sample_rows() does.
1909 : : */
1910 : 17050122 : prstate->live_tuples++;
1911 : :
1912 : : /*
1913 : : * Is the tuple definitely visible to all transactions?
1914 : : *
1915 : : * NB: Like with per-tuple hint bits, we can't set the
1916 : : * PD_ALL_VISIBLE flag if the inserter committed asynchronously.
1917 : : * See SetHintBits for more info. Check that the tuple is hinted
1918 : : * xmin-committed because of that.
1919 : : */
1920 [ + + ]: 17050122 : if (!HeapTupleHeaderXminCommitted(htup))
1921 : : {
1922 : 29634 : prstate->set_all_visible = false;
1923 : 29634 : prstate->set_all_frozen = false;
1924 : 29634 : break;
1925 : : }
1926 : :
1927 : : /*
1928 : : * The inserter definitely committed. But we don't know if it is
1929 : : * old enough that everyone sees it as committed. Later, after
1930 : : * processing all the tuples on the page, we'll check if there is
1931 : : * any snapshot that still considers the newest xid on the page to
1932 : : * be running. If so, we don't consider the page all-visible.
1933 : : */
1934 : 17020488 : xmin = HeapTupleHeaderGetXmin(htup);
1935 : :
1936 : : /* Track newest xmin on page. */
1937 [ + + + + ]: 17020488 : if (TransactionIdFollows(xmin, prstate->newest_live_xid) &&
1938 : : TransactionIdIsNormal(xmin))
1939 : 622908 : prstate->newest_live_xid = xmin;
1940 : :
1941 : 17020488 : break;
1942 : :
1943 : 388382 : case HEAPTUPLE_RECENTLY_DEAD:
1944 : 388382 : prstate->recently_dead_tuples++;
1945 : 388382 : prstate->set_all_visible = false;
1946 : 388382 : prstate->set_all_frozen = false;
1947 : :
1948 : : /*
1949 : : * This tuple will soon become DEAD. Update the hint field so
1950 : : * that the page is reconsidered for pruning in future.
1951 : : */
1952 : 388382 : heap_prune_record_prunable(prstate,
1953 : : HeapTupleHeaderGetUpdateXid(htup),
1954 : : offnum);
1955 : 388382 : break;
1956 : :
1957 : 138122 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1958 : :
1959 : : /*
1960 : : * We do not count these rows as live, because we expect the
1961 : : * inserting transaction to update the counters at commit, and we
1962 : : * assume that will happen only after we report our results. This
1963 : : * assumption is a bit shaky, but it is what acquire_sample_rows()
1964 : : * does, so be consistent.
1965 : : */
1966 : 138122 : prstate->set_all_visible = false;
1967 : 138122 : prstate->set_all_frozen = false;
1968 : :
1969 : : /*
1970 : : * Though there is nothing "prunable" on the page, we maintain
1971 : : * pd_prune_xid for inserts so that we have the opportunity to
1972 : : * mark them all-visible during the next round of pruning.
1973 : : */
1974 : 138122 : heap_prune_record_prunable(prstate,
1975 : : HeapTupleHeaderGetXmin(htup),
1976 : : offnum);
1977 : 138122 : break;
1978 : :
1979 : 5952016 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1980 : :
1981 : : /*
1982 : : * This an expected case during concurrent vacuum. Count such
1983 : : * rows as live. As above, we assume the deleting transaction
1984 : : * will commit and update the counters after we report.
1985 : : */
1986 : 5952016 : prstate->live_tuples++;
1987 : 5952016 : prstate->set_all_visible = false;
1988 : 5952016 : prstate->set_all_frozen = false;
1989 : :
1990 : : /*
1991 : : * This tuple may soon become DEAD. Update the hint field so that
1992 : : * the page is reconsidered for pruning in future.
1993 : : */
1994 : 5952016 : heap_prune_record_prunable(prstate,
1995 : : HeapTupleHeaderGetUpdateXid(htup),
1996 : : offnum);
1997 : 5952016 : break;
1998 : :
1999 : 0 : default:
2000 : :
2001 : : /*
2002 : : * DEAD tuples should've been passed to heap_prune_record_dead()
2003 : : * or heap_prune_record_unused() instead.
2004 : : */
2005 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result %d",
2006 : : prstate->htsv[offnum]);
2007 : : break;
2008 : : }
2009 : :
2010 : : /* Consider freezing any normal tuples which will not be removed */
2011 [ + + ]: 23528642 : if (prstate->attempt_freeze)
2012 : : {
2013 : : bool totally_frozen;
2014 : :
2015 [ + + ]: 13277502 : if ((heap_prepare_freeze_tuple(htup,
2016 : 13277502 : prstate->cutoffs,
2017 : : &prstate->pagefrz,
2018 : 13277502 : &prstate->frozen[prstate->nfrozen],
2019 : : &totally_frozen)))
2020 : : {
2021 : : /* Save prepared freeze plan for later */
2022 : 3428671 : prstate->frozen[prstate->nfrozen++].offset = offnum;
2023 : : }
2024 : :
2025 : : /*
2026 : : * If any tuple isn't either totally frozen already or eligible to
2027 : : * become totally frozen (according to its freeze plan), then the page
2028 : : * definitely cannot be set all-frozen in the visibility map later on.
2029 : : */
2030 [ + + ]: 13277502 : if (!totally_frozen)
2031 : 6707449 : prstate->set_all_frozen = false;
2032 : : }
2033 : 23528642 : }
2034 : :
2035 : :
2036 : : /*
2037 : : * Record line pointer that was already LP_DEAD and is left unchanged.
2038 : : */
2039 : : static void
2040 : 1686091 : heap_prune_record_unchanged_lp_dead(PruneState *prstate, OffsetNumber offnum)
2041 : : {
2042 : : Assert(!prstate->processed[offnum]);
2043 : 1686091 : prstate->processed[offnum] = true;
2044 : :
2045 : : /*
2046 : : * Deliberately don't set hastup for LP_DEAD items. We make the soft
2047 : : * assumption that any LP_DEAD items encountered here will become
2048 : : * LP_UNUSED later on, before count_nondeletable_pages is reached. If we
2049 : : * don't make this assumption then rel truncation will only happen every
2050 : : * other VACUUM, at most. Besides, VACUUM must treat
2051 : : * hastup/nonempty_pages as provisional no matter how LP_DEAD items are
2052 : : * handled (handled here, or handled later on).
2053 : : *
2054 : : * Similarly, don't unset set_all_visible and set_all_frozen until later,
2055 : : * at the end of heap_page_prune_and_freeze(). This will allow us to
2056 : : * attempt to freeze the page after pruning. As long as we unset it
2057 : : * before updating the visibility map, this will be correct.
2058 : : */
2059 : :
2060 : : /* Record the dead offset for vacuum */
2061 : 1686091 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
2062 : :
2063 : : /*
2064 : : * It's incorrect for a page to be marked all-visible if it contains dead
2065 : : * items.
2066 : : */
2067 [ - + ]: 1686091 : if (PageIsAllVisible(prstate->page))
2068 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
2069 : 1686091 : }
2070 : :
2071 : : /*
2072 : : * Record LP_REDIRECT that is left unchanged.
2073 : : */
2074 : : static void
2075 : 178122 : heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum)
2076 : : {
2077 : : /*
2078 : : * A redirect line pointer doesn't count as a live tuple.
2079 : : *
2080 : : * If we leave a redirect line pointer in place, there will be another
2081 : : * tuple on the page that it points to. We will do the bookkeeping for
2082 : : * that separately. So we have nothing to do here, except remember that
2083 : : * we processed this item.
2084 : : */
2085 : : Assert(!prstate->processed[offnum]);
2086 : 178122 : prstate->processed[offnum] = true;
2087 : 178122 : }
2088 : :
2089 : : /*
2090 : : * Perform the actual page changes needed by heap_page_prune_and_freeze().
2091 : : *
2092 : : * If 'lp_truncate_only' is set, we are merely marking LP_DEAD line pointers
2093 : : * as unused, not redirecting or removing anything else. The
2094 : : * PageRepairFragmentation() call is skipped in that case.
2095 : : *
2096 : : * If 'lp_truncate_only' is not set, the caller must hold a cleanup lock on
2097 : : * the buffer. If it is set, an ordinary exclusive lock suffices.
2098 : : */
2099 : : void
2100 : 82787 : heap_page_prune_execute(Buffer buffer, bool lp_truncate_only,
2101 : : OffsetNumber *redirected, int nredirected,
2102 : : OffsetNumber *nowdead, int ndead,
2103 : : OffsetNumber *nowunused, int nunused)
2104 : : {
2105 : 82787 : Page page = BufferGetPage(buffer);
2106 : : OffsetNumber *offnum;
2107 : : HeapTupleHeader htup PG_USED_FOR_ASSERTS_ONLY;
2108 : :
2109 : : /* Shouldn't be called unless there's something to do */
2110 : : Assert(nredirected > 0 || ndead > 0 || nunused > 0);
2111 : :
2112 : : /* If 'lp_truncate_only', we can only remove already-dead line pointers */
2113 : : Assert(!lp_truncate_only || (nredirected == 0 && ndead == 0));
2114 : :
2115 : : /* Update all redirected line pointers */
2116 : 82787 : offnum = redirected;
2117 [ + + ]: 176471 : for (int i = 0; i < nredirected; i++)
2118 : : {
2119 : 93684 : OffsetNumber fromoff = *offnum++;
2120 : 93684 : OffsetNumber tooff = *offnum++;
2121 : 93684 : ItemId fromlp = PageGetItemId(page, fromoff);
2122 : : ItemId tolp PG_USED_FOR_ASSERTS_ONLY;
2123 : :
2124 : : #ifdef USE_ASSERT_CHECKING
2125 : :
2126 : : /*
2127 : : * Any existing item that we set as an LP_REDIRECT (any 'from' item)
2128 : : * must be the first item from a HOT chain. If the item has tuple
2129 : : * storage then it can't be a heap-only tuple. Otherwise we are just
2130 : : * maintaining an existing LP_REDIRECT from an existing HOT chain that
2131 : : * has been pruned at least once before now.
2132 : : */
2133 : : if (!ItemIdIsRedirected(fromlp))
2134 : : {
2135 : : Assert(ItemIdHasStorage(fromlp) && ItemIdIsNormal(fromlp));
2136 : :
2137 : : htup = (HeapTupleHeader) PageGetItem(page, fromlp);
2138 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2139 : : }
2140 : : else
2141 : : {
2142 : : /* We shouldn't need to redundantly set the redirect */
2143 : : Assert(ItemIdGetRedirect(fromlp) != tooff);
2144 : : }
2145 : :
2146 : : /*
2147 : : * The item that we're about to set as an LP_REDIRECT (the 'from'
2148 : : * item) will point to an existing item (the 'to' item) that is
2149 : : * already a heap-only tuple. There can be at most one LP_REDIRECT
2150 : : * item per HOT chain.
2151 : : *
2152 : : * We need to keep around an LP_REDIRECT item (after original
2153 : : * non-heap-only root tuple gets pruned away) so that it's always
2154 : : * possible for VACUUM to easily figure out what TID to delete from
2155 : : * indexes when an entire HOT chain becomes dead. A heap-only tuple
2156 : : * can never become LP_DEAD; an LP_REDIRECT item or a regular heap
2157 : : * tuple can.
2158 : : *
2159 : : * This check may miss problems, e.g. the target of a redirect could
2160 : : * be marked as unused subsequently. The page_verify_redirects() check
2161 : : * below will catch such problems.
2162 : : */
2163 : : tolp = PageGetItemId(page, tooff);
2164 : : Assert(ItemIdHasStorage(tolp) && ItemIdIsNormal(tolp));
2165 : : htup = (HeapTupleHeader) PageGetItem(page, tolp);
2166 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2167 : : #endif
2168 : :
2169 : 93684 : ItemIdSetRedirect(fromlp, tooff);
2170 : : }
2171 : :
2172 : : /* Update all now-dead line pointers */
2173 : 82787 : offnum = nowdead;
2174 [ + + ]: 4137707 : for (int i = 0; i < ndead; i++)
2175 : : {
2176 : 4054920 : OffsetNumber off = *offnum++;
2177 : 4054920 : ItemId lp = PageGetItemId(page, off);
2178 : :
2179 : : #ifdef USE_ASSERT_CHECKING
2180 : :
2181 : : /*
2182 : : * An LP_DEAD line pointer must be left behind when the original item
2183 : : * (which is dead to everybody) could still be referenced by a TID in
2184 : : * an index. This should never be necessary with any individual
2185 : : * heap-only tuple item, though. (It's not clear how much of a problem
2186 : : * that would be, but there is no reason to allow it.)
2187 : : */
2188 : : if (ItemIdHasStorage(lp))
2189 : : {
2190 : : Assert(ItemIdIsNormal(lp));
2191 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2192 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2193 : : }
2194 : : else
2195 : : {
2196 : : /* Whole HOT chain becomes dead */
2197 : : Assert(ItemIdIsRedirected(lp));
2198 : : }
2199 : : #endif
2200 : :
2201 : 4054920 : ItemIdSetDead(lp);
2202 : : }
2203 : :
2204 : : /* Update all now-unused line pointers */
2205 : 82787 : offnum = nowunused;
2206 [ + + ]: 428774 : for (int i = 0; i < nunused; i++)
2207 : : {
2208 : 345987 : OffsetNumber off = *offnum++;
2209 : 345987 : ItemId lp = PageGetItemId(page, off);
2210 : :
2211 : : #ifdef USE_ASSERT_CHECKING
2212 : :
2213 : : if (lp_truncate_only)
2214 : : {
2215 : : /* Setting LP_DEAD to LP_UNUSED in vacuum's second pass */
2216 : : Assert(ItemIdIsDead(lp) && !ItemIdHasStorage(lp));
2217 : : }
2218 : : else
2219 : : {
2220 : : /*
2221 : : * When heap_page_prune_and_freeze() was called, mark_unused_now
2222 : : * may have been passed as true, which allows would-be LP_DEAD
2223 : : * items to be made LP_UNUSED instead. This is only possible if
2224 : : * the relation has no indexes. If there are any dead items, then
2225 : : * mark_unused_now was not true and every item being marked
2226 : : * LP_UNUSED must refer to a heap-only tuple.
2227 : : */
2228 : : if (ndead > 0)
2229 : : {
2230 : : Assert(ItemIdHasStorage(lp) && ItemIdIsNormal(lp));
2231 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2232 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2233 : : }
2234 : : else
2235 : : Assert(ItemIdIsUsed(lp));
2236 : : }
2237 : :
2238 : : #endif
2239 : :
2240 : 345987 : ItemIdSetUnused(lp);
2241 : : }
2242 : :
2243 [ + + ]: 82787 : if (lp_truncate_only)
2244 : 2518 : PageTruncateLinePointerArray(page);
2245 : : else
2246 : : {
2247 : : /*
2248 : : * Finally, repair any fragmentation, and update the page's hint bit
2249 : : * about whether it has free pointers.
2250 : : */
2251 : 80269 : PageRepairFragmentation(page);
2252 : :
2253 : : /*
2254 : : * Now that the page has been modified, assert that redirect items
2255 : : * still point to valid targets.
2256 : : */
2257 : 80269 : page_verify_redirects(page);
2258 : : }
2259 : 82787 : }
2260 : :
2261 : :
2262 : : /*
2263 : : * If built with assertions, verify that all LP_REDIRECT items point to a
2264 : : * valid item.
2265 : : *
2266 : : * One way that bugs related to HOT pruning show is redirect items pointing to
2267 : : * removed tuples. It's not trivial to reliably check that marking an item
2268 : : * unused will not orphan a redirect item during heap_prune_chain() /
2269 : : * heap_page_prune_execute(), so we additionally check the whole page after
2270 : : * pruning. Without this check such bugs would typically only cause asserts
2271 : : * later, potentially well after the corruption has been introduced.
2272 : : *
2273 : : * Also check comments in heap_page_prune_execute()'s redirection loop.
2274 : : */
2275 : : static void
2276 : 80269 : page_verify_redirects(Page page)
2277 : : {
2278 : : #ifdef USE_ASSERT_CHECKING
2279 : : OffsetNumber offnum;
2280 : : OffsetNumber maxoff;
2281 : :
2282 : : maxoff = PageGetMaxOffsetNumber(page);
2283 : : for (offnum = FirstOffsetNumber;
2284 : : offnum <= maxoff;
2285 : : offnum = OffsetNumberNext(offnum))
2286 : : {
2287 : : ItemId itemid = PageGetItemId(page, offnum);
2288 : : OffsetNumber targoff;
2289 : : ItemId targitem;
2290 : : HeapTupleHeader htup;
2291 : :
2292 : : if (!ItemIdIsRedirected(itemid))
2293 : : continue;
2294 : :
2295 : : targoff = ItemIdGetRedirect(itemid);
2296 : : targitem = PageGetItemId(page, targoff);
2297 : :
2298 : : Assert(ItemIdIsUsed(targitem));
2299 : : Assert(ItemIdIsNormal(targitem));
2300 : : Assert(ItemIdHasStorage(targitem));
2301 : : htup = (HeapTupleHeader) PageGetItem(page, targitem);
2302 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2303 : : }
2304 : : #endif
2305 : 80269 : }
2306 : :
2307 : :
2308 : : /*
2309 : : * For all items in this page, find their respective root line pointers.
2310 : : * If item k is part of a HOT-chain with root at item j, then we set
2311 : : * root_offsets[k - 1] = j.
2312 : : *
2313 : : * The passed-in root_offsets array must have MaxHeapTuplesPerPage entries.
2314 : : * Unused entries are filled with InvalidOffsetNumber (zero).
2315 : : *
2316 : : * The function must be called with at least share lock on the buffer, to
2317 : : * prevent concurrent prune operations.
2318 : : *
2319 : : * Note: The information collected here is valid only as long as the caller
2320 : : * holds a pin on the buffer. Once pin is released, a tuple might be pruned
2321 : : * and reused by a completely unrelated tuple.
2322 : : */
2323 : : void
2324 : 140101 : heap_get_root_tuples(Page page, OffsetNumber *root_offsets)
2325 : : {
2326 : : OffsetNumber offnum,
2327 : : maxoff;
2328 : :
2329 [ + - - + : 140101 : MemSet(root_offsets, InvalidOffsetNumber,
- - - - -
- ]
2330 : : MaxHeapTuplesPerPage * sizeof(OffsetNumber));
2331 : :
2332 : 140101 : maxoff = PageGetMaxOffsetNumber(page);
2333 [ + + ]: 11752292 : for (offnum = FirstOffsetNumber; offnum <= maxoff; offnum = OffsetNumberNext(offnum))
2334 : : {
2335 : 11612191 : ItemId lp = PageGetItemId(page, offnum);
2336 : : HeapTupleHeader htup;
2337 : : OffsetNumber nextoffnum;
2338 : : TransactionId priorXmax;
2339 : :
2340 : : /* skip unused and dead items */
2341 [ + + + + ]: 11612191 : if (!ItemIdIsUsed(lp) || ItemIdIsDead(lp))
2342 : 11514 : continue;
2343 : :
2344 [ + + ]: 11600677 : if (ItemIdIsNormal(lp))
2345 : : {
2346 : 11597430 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2347 : :
2348 : : /*
2349 : : * Check if this tuple is part of a HOT-chain rooted at some other
2350 : : * tuple. If so, skip it for now; we'll process it when we find
2351 : : * its root.
2352 : : */
2353 [ + + ]: 11597430 : if (HeapTupleHeaderIsHeapOnly(htup))
2354 : 3641 : continue;
2355 : :
2356 : : /*
2357 : : * This is either a plain tuple or the root of a HOT-chain.
2358 : : * Remember it in the mapping.
2359 : : */
2360 : 11593789 : root_offsets[offnum - 1] = offnum;
2361 : :
2362 : : /* If it's not the start of a HOT-chain, we're done with it */
2363 [ + + ]: 11593789 : if (!HeapTupleHeaderIsHotUpdated(htup))
2364 : 11593507 : continue;
2365 : :
2366 : : /* Set up to scan the HOT-chain */
2367 : 282 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2368 : 282 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2369 : : }
2370 : : else
2371 : : {
2372 : : /* Must be a redirect item. We do not set its root_offsets entry */
2373 : : Assert(ItemIdIsRedirected(lp));
2374 : : /* Set up to scan the HOT-chain */
2375 : 3247 : nextoffnum = ItemIdGetRedirect(lp);
2376 : 3247 : priorXmax = InvalidTransactionId;
2377 : : }
2378 : :
2379 : : /*
2380 : : * Now follow the HOT-chain and collect other tuples in the chain.
2381 : : *
2382 : : * Note: Even though this is a nested loop, the complexity of the
2383 : : * function is O(N) because a tuple in the page should be visited not
2384 : : * more than twice, once in the outer loop and once in HOT-chain
2385 : : * chases.
2386 : : */
2387 : : for (;;)
2388 : : {
2389 : : /* Sanity check (pure paranoia) */
2390 [ - + ]: 3637 : if (nextoffnum < FirstOffsetNumber)
2391 : 0 : break;
2392 : :
2393 : : /*
2394 : : * An offset past the end of page's line pointer array is possible
2395 : : * when the array was truncated
2396 : : */
2397 [ - + ]: 3637 : if (nextoffnum > maxoff)
2398 : 0 : break;
2399 : :
2400 : 3637 : lp = PageGetItemId(page, nextoffnum);
2401 : :
2402 : : /* Check for broken chains */
2403 [ - + ]: 3637 : if (!ItemIdIsNormal(lp))
2404 : 0 : break;
2405 : :
2406 : 3637 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2407 : :
2408 [ + + - + ]: 4027 : if (TransactionIdIsValid(priorXmax) &&
2409 : 390 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(htup)))
2410 : 0 : break;
2411 : :
2412 : : /* Remember the root line pointer for this item */
2413 : 3637 : root_offsets[nextoffnum - 1] = offnum;
2414 : :
2415 : : /* Advance to next chain member, if any */
2416 [ + + ]: 3637 : if (!HeapTupleHeaderIsHotUpdated(htup))
2417 : 3529 : break;
2418 : :
2419 : : /* HOT implies it can't have moved to different partition */
2420 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
2421 : :
2422 : 108 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2423 : 108 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2424 : : }
2425 : : }
2426 : 140101 : }
2427 : :
2428 : :
2429 : : /*
2430 : : * Compare fields that describe actions required to freeze tuple with caller's
2431 : : * open plan. If everything matches then the frz tuple plan is equivalent to
2432 : : * caller's plan.
2433 : : */
2434 : : static inline bool
2435 : 1338731 : heap_log_freeze_eq(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2436 : : {
2437 [ + + ]: 1338731 : if (plan->xmax == frz->xmax &&
2438 [ + + ]: 1337437 : plan->t_infomask2 == frz->t_infomask2 &&
2439 [ + + ]: 1336404 : plan->t_infomask == frz->t_infomask &&
2440 [ + - ]: 1333034 : plan->frzflags == frz->frzflags)
2441 : 1333034 : return true;
2442 : :
2443 : : /* Caller must call heap_log_freeze_new_plan again for frz */
2444 : 5697 : return false;
2445 : : }
2446 : :
2447 : : /*
2448 : : * Comparator used to deduplicate the freeze plans used in WAL records.
2449 : : */
2450 : : static int
2451 : 1816300 : heap_log_freeze_cmp(const void *arg1, const void *arg2)
2452 : : {
2453 : 1816300 : const HeapTupleFreeze *frz1 = arg1;
2454 : 1816300 : const HeapTupleFreeze *frz2 = arg2;
2455 : :
2456 [ + + ]: 1816300 : if (frz1->xmax < frz2->xmax)
2457 : 13140 : return -1;
2458 [ + + ]: 1803160 : else if (frz1->xmax > frz2->xmax)
2459 : 14572 : return 1;
2460 : :
2461 [ + + ]: 1788588 : if (frz1->t_infomask2 < frz2->t_infomask2)
2462 : 6207 : return -1;
2463 [ + + ]: 1782381 : else if (frz1->t_infomask2 > frz2->t_infomask2)
2464 : 6115 : return 1;
2465 : :
2466 [ + + ]: 1776266 : if (frz1->t_infomask < frz2->t_infomask)
2467 : 12837 : return -1;
2468 [ + + ]: 1763429 : else if (frz1->t_infomask > frz2->t_infomask)
2469 : 23544 : return 1;
2470 : :
2471 [ - + ]: 1739885 : if (frz1->frzflags < frz2->frzflags)
2472 : 0 : return -1;
2473 [ - + ]: 1739885 : else if (frz1->frzflags > frz2->frzflags)
2474 : 0 : return 1;
2475 : :
2476 : : /*
2477 : : * heap_log_freeze_eq would consider these tuple-wise plans to be equal.
2478 : : * (So the tuples will share a single canonical freeze plan.)
2479 : : *
2480 : : * We tiebreak on page offset number to keep each freeze plan's page
2481 : : * offset number array individually sorted. (Unnecessary, but be tidy.)
2482 : : */
2483 [ + + ]: 1739885 : if (frz1->offset < frz2->offset)
2484 : 1492473 : return -1;
2485 [ + - ]: 247412 : else if (frz1->offset > frz2->offset)
2486 : 247412 : return 1;
2487 : :
2488 : : Assert(false);
2489 : 0 : return 0;
2490 : : }
2491 : :
2492 : : /*
2493 : : * Start new plan initialized using tuple-level actions. At least one tuple
2494 : : * will have steps required to freeze described by caller's plan during REDO.
2495 : : */
2496 : : static inline void
2497 : 33183 : heap_log_freeze_new_plan(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2498 : : {
2499 : 33183 : plan->xmax = frz->xmax;
2500 : 33183 : plan->t_infomask2 = frz->t_infomask2;
2501 : 33183 : plan->t_infomask = frz->t_infomask;
2502 : 33183 : plan->frzflags = frz->frzflags;
2503 : 33183 : plan->ntuples = 1; /* for now */
2504 : 33183 : }
2505 : :
2506 : : /*
2507 : : * Deduplicate tuple-based freeze plans so that each distinct set of
2508 : : * processing steps is only stored once in the WAL record.
2509 : : * Called during original execution of freezing (for logged relations).
2510 : : *
2511 : : * Return value is number of plans set in *plans_out for caller. Also writes
2512 : : * an array of offset numbers into *offsets_out output argument for caller
2513 : : * (actually there is one array per freeze plan, but that's not of immediate
2514 : : * concern to our caller).
2515 : : */
2516 : : static int
2517 : 27486 : heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
2518 : : xlhp_freeze_plan *plans_out,
2519 : : OffsetNumber *offsets_out)
2520 : : {
2521 : 27486 : int nplans = 0;
2522 : :
2523 : : /* Sort tuple-based freeze plans in the order required to deduplicate */
2524 : 27486 : qsort(tuples, ntuples, sizeof(HeapTupleFreeze), heap_log_freeze_cmp);
2525 : :
2526 [ + + ]: 1393703 : for (int i = 0; i < ntuples; i++)
2527 : : {
2528 : 1366217 : HeapTupleFreeze *frz = tuples + i;
2529 : :
2530 [ + + ]: 1366217 : if (i == 0)
2531 : : {
2532 : : /* New canonical freeze plan starting with first tup */
2533 : 27486 : heap_log_freeze_new_plan(plans_out, frz);
2534 : 27486 : nplans++;
2535 : : }
2536 [ + + ]: 1338731 : else if (heap_log_freeze_eq(plans_out, frz))
2537 : : {
2538 : : /* tup matches open canonical plan -- include tup in it */
2539 : : Assert(offsets_out[i - 1] < frz->offset);
2540 : 1333034 : plans_out->ntuples++;
2541 : : }
2542 : : else
2543 : : {
2544 : : /* Tup doesn't match current plan -- done with it now */
2545 : 5697 : plans_out++;
2546 : :
2547 : : /* New canonical freeze plan starting with this tup */
2548 : 5697 : heap_log_freeze_new_plan(plans_out, frz);
2549 : 5697 : nplans++;
2550 : : }
2551 : :
2552 : : /*
2553 : : * Save page offset number in dedicated buffer in passing.
2554 : : *
2555 : : * REDO routine relies on the record's offset numbers array grouping
2556 : : * offset numbers by freeze plan. The sort order within each grouping
2557 : : * is ascending offset number order, just to keep things tidy.
2558 : : */
2559 : 1366217 : offsets_out[i] = frz->offset;
2560 : : }
2561 : :
2562 : : Assert(nplans > 0 && nplans <= ntuples);
2563 : :
2564 : 27486 : return nplans;
2565 : : }
2566 : :
2567 : : /*
2568 : : * Write an XLOG_HEAP2_PRUNE* WAL record
2569 : : *
2570 : : * This is used for several different page maintenance operations:
2571 : : *
2572 : : * - Page pruning, in VACUUM's 1st pass or on access: Some items are
2573 : : * redirected, some marked dead, and some removed altogether.
2574 : : *
2575 : : * - Freezing: Items are marked as 'frozen'.
2576 : : *
2577 : : * - Vacuum, 2nd pass: Items that are already LP_DEAD are marked as unused.
2578 : : *
2579 : : * They have enough commonalities that we use a single WAL record for them
2580 : : * all.
2581 : : *
2582 : : * If replaying the record requires a cleanup lock, pass cleanup_lock = true.
2583 : : * Replaying 'redirected' or 'dead' items always requires a cleanup lock, but
2584 : : * replaying 'unused' items depends on whether they were all previously marked
2585 : : * as dead.
2586 : : *
2587 : : * If the VM is being updated, vmflags will contain the bits to set. In this
2588 : : * case, vmbuffer should already have been updated and marked dirty and should
2589 : : * still be pinned and locked.
2590 : : *
2591 : : * Note: This function scribbles on the 'frozen' array.
2592 : : *
2593 : : * Note: This is called in a critical section, so careful what you do here.
2594 : : */
2595 : : void
2596 : 149749 : log_heap_prune_and_freeze(Relation relation, Buffer buffer,
2597 : : Buffer vmbuffer, uint8 vmflags,
2598 : : TransactionId conflict_xid,
2599 : : bool cleanup_lock,
2600 : : PruneReason reason,
2601 : : HeapTupleFreeze *frozen, int nfrozen,
2602 : : OffsetNumber *redirected, int nredirected,
2603 : : OffsetNumber *dead, int ndead,
2604 : : OffsetNumber *unused, int nunused)
2605 : : {
2606 : : xl_heap_prune xlrec;
2607 : : XLogRecPtr recptr;
2608 : : uint8 info;
2609 : : uint8 regbuf_flags_heap;
2610 : :
2611 : 149749 : Page heap_page = BufferGetPage(buffer);
2612 : :
2613 : : /* The following local variables hold data registered in the WAL record: */
2614 : : xlhp_freeze_plan plans[MaxHeapTuplesPerPage];
2615 : : xlhp_freeze_plans freeze_plans;
2616 : : xlhp_prune_items redirect_items;
2617 : : xlhp_prune_items dead_items;
2618 : : xlhp_prune_items unused_items;
2619 : : OffsetNumber frz_offsets[MaxHeapTuplesPerPage];
2620 [ + + + + : 149749 : bool do_prune = nredirected > 0 || ndead > 0 || nunused > 0;
+ + ]
2621 : 149749 : bool do_set_vm = vmflags & VISIBILITYMAP_VALID_BITS;
2622 : 149749 : bool heap_fpi_allowed = true;
2623 : :
2624 : : Assert((vmflags & VISIBILITYMAP_VALID_BITS) == vmflags);
2625 : :
2626 : 149749 : xlrec.flags = 0;
2627 : 149749 : regbuf_flags_heap = REGBUF_STANDARD;
2628 : :
2629 : : /*
2630 : : * We can avoid an FPI of the heap page if the only modification we are
2631 : : * making to it is to set PD_ALL_VISIBLE and checksums/wal_log_hints are
2632 : : * disabled.
2633 : : *
2634 : : * However, if the page has never been WAL-logged (LSN is invalid), we
2635 : : * must force an FPI regardless. This can happen when another backend
2636 : : * extends the heap, initializes the page, and then fails before WAL-
2637 : : * logging it. Since heap extension is not WAL-logged, recovery might try
2638 : : * to replay our record and find that the page isn't initialized, which
2639 : : * would cause a PANIC.
2640 : : */
2641 [ - + ]: 149749 : if (!XLogRecPtrIsValid(PageGetLSN(heap_page)))
2642 : 0 : regbuf_flags_heap |= REGBUF_FORCE_IMAGE;
2643 [ + + + + : 149749 : else if (!do_prune && nfrozen == 0 && (!do_set_vm || !XLogHintBitIsNeeded()))
+ - + + +
+ ]
2644 : : {
2645 : 2852 : regbuf_flags_heap |= REGBUF_NO_IMAGE;
2646 : 2852 : heap_fpi_allowed = false;
2647 : : }
2648 : :
2649 : : /*
2650 : : * Prepare data for the buffer. The arrays are not actually in the
2651 : : * buffer, but we pretend that they are. When XLogInsert stores a full
2652 : : * page image, the arrays can be omitted.
2653 : : */
2654 : 149749 : XLogBeginInsert();
2655 : 149749 : XLogRegisterBuffer(0, buffer, regbuf_flags_heap);
2656 : :
2657 [ + + ]: 149749 : if (do_set_vm)
2658 : 77506 : XLogRegisterBuffer(1, vmbuffer, 0);
2659 : :
2660 [ + + ]: 149749 : if (nfrozen > 0)
2661 : : {
2662 : : int nplans;
2663 : :
2664 : 27486 : xlrec.flags |= XLHP_HAS_FREEZE_PLANS;
2665 : :
2666 : : /*
2667 : : * Prepare deduplicated representation for use in the WAL record. This
2668 : : * destructively sorts frozen tuples array in-place.
2669 : : */
2670 : 27486 : nplans = heap_log_freeze_plan(frozen, nfrozen, plans, frz_offsets);
2671 : :
2672 : 27486 : freeze_plans.nplans = nplans;
2673 : 27486 : XLogRegisterBufData(0, &freeze_plans,
2674 : : offsetof(xlhp_freeze_plans, plans));
2675 : 27486 : XLogRegisterBufData(0, plans,
2676 : : sizeof(xlhp_freeze_plan) * nplans);
2677 : : }
2678 [ + + ]: 149749 : if (nredirected > 0)
2679 : : {
2680 : 17818 : xlrec.flags |= XLHP_HAS_REDIRECTIONS;
2681 : :
2682 : 17818 : redirect_items.ntargets = nredirected;
2683 : 17818 : XLogRegisterBufData(0, &redirect_items,
2684 : : offsetof(xlhp_prune_items, data));
2685 : 17818 : XLogRegisterBufData(0, redirected,
2686 : : sizeof(OffsetNumber[2]) * nredirected);
2687 : : }
2688 [ + + ]: 149749 : if (ndead > 0)
2689 : : {
2690 : 57717 : xlrec.flags |= XLHP_HAS_DEAD_ITEMS;
2691 : :
2692 : 57717 : dead_items.ntargets = ndead;
2693 : 57717 : XLogRegisterBufData(0, &dead_items,
2694 : : offsetof(xlhp_prune_items, data));
2695 : 57717 : XLogRegisterBufData(0, dead,
2696 : : sizeof(OffsetNumber) * ndead);
2697 : : }
2698 [ + + ]: 149749 : if (nunused > 0)
2699 : : {
2700 : 31008 : xlrec.flags |= XLHP_HAS_NOW_UNUSED_ITEMS;
2701 : :
2702 : 31008 : unused_items.ntargets = nunused;
2703 : 31008 : XLogRegisterBufData(0, &unused_items,
2704 : : offsetof(xlhp_prune_items, data));
2705 : 31008 : XLogRegisterBufData(0, unused,
2706 : : sizeof(OffsetNumber) * nunused);
2707 : : }
2708 [ + + ]: 149749 : if (nfrozen > 0)
2709 : 27486 : XLogRegisterBufData(0, frz_offsets,
2710 : : sizeof(OffsetNumber) * nfrozen);
2711 : :
2712 : : /*
2713 : : * Prepare the main xl_heap_prune record. We already set the XLHP_HAS_*
2714 : : * flag above.
2715 : : */
2716 [ + + ]: 149749 : if (vmflags & VISIBILITYMAP_ALL_VISIBLE)
2717 : : {
2718 : 77506 : xlrec.flags |= XLHP_VM_ALL_VISIBLE;
2719 [ + + ]: 77506 : if (vmflags & VISIBILITYMAP_ALL_FROZEN)
2720 : 46809 : xlrec.flags |= XLHP_VM_ALL_FROZEN;
2721 : : }
2722 [ + + + + : 149749 : if (RelationIsAccessibleInLogicalDecoding(relation))
+ - - + -
- - - + +
+ + - + -
- + - ]
2723 : 662 : xlrec.flags |= XLHP_IS_CATALOG_REL;
2724 [ + + ]: 149749 : if (TransactionIdIsValid(conflict_xid))
2725 : 119122 : xlrec.flags |= XLHP_HAS_CONFLICT_HORIZON;
2726 [ + + ]: 149749 : if (cleanup_lock)
2727 : 70334 : xlrec.flags |= XLHP_CLEANUP_LOCK;
2728 : : else
2729 : : {
2730 : : Assert(nredirected == 0 && ndead == 0);
2731 : : /* also, any items in 'unused' must've been LP_DEAD previously */
2732 : : }
2733 : 149749 : XLogRegisterData(&xlrec, SizeOfHeapPrune);
2734 [ + + ]: 149749 : if (TransactionIdIsValid(conflict_xid))
2735 : 119122 : XLogRegisterData(&conflict_xid, sizeof(TransactionId));
2736 : :
2737 [ + + + - ]: 149749 : switch (reason)
2738 : : {
2739 : 70526 : case PRUNE_ON_ACCESS:
2740 : 70526 : info = XLOG_HEAP2_PRUNE_ON_ACCESS;
2741 : 70526 : break;
2742 : 62539 : case PRUNE_VACUUM_SCAN:
2743 : 62539 : info = XLOG_HEAP2_PRUNE_VACUUM_SCAN;
2744 : 62539 : break;
2745 : 16684 : case PRUNE_VACUUM_CLEANUP:
2746 : 16684 : info = XLOG_HEAP2_PRUNE_VACUUM_CLEANUP;
2747 : 16684 : break;
2748 : 0 : default:
2749 [ # # ]: 0 : elog(ERROR, "unrecognized prune reason: %d", (int) reason);
2750 : : break;
2751 : : }
2752 : 149749 : recptr = XLogInsert(RM_HEAP2_ID, info);
2753 : :
2754 [ + + ]: 149749 : if (do_set_vm)
2755 : : {
2756 : : Assert(BufferIsDirty(vmbuffer));
2757 : 77506 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2758 : : }
2759 : :
2760 : : /*
2761 : : * If we explicitly skip an FPI, we must not stamp the heap page with this
2762 : : * record's LSN. Recovery skips records <= the stamped LSN, so this could
2763 : : * lead to skipping an earlier FPI needed to repair a torn page.
2764 : : */
2765 [ + + ]: 149749 : if (heap_fpi_allowed)
2766 : : {
2767 : : Assert(BufferIsDirty(buffer));
2768 : 146897 : PageSetLSN(heap_page, recptr);
2769 : : }
2770 : 149749 : }
|
