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 : 12166611 : heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer,
273 : : bool rel_read_only)
274 : : {
275 : 12166611 : 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 [ + + ]: 12166611 : if (RecoveryInProgress())
286 : 265764 : 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 : 11900847 : prune_xid = PageGetPruneXid(page);
295 [ + + ]: 11900847 : if (!TransactionIdIsValid(prune_xid))
296 : 8146921 : return;
297 : :
298 : : /*
299 : : * Check whether prune_xid indicates that there may be dead rows that can
300 : : * be cleaned up.
301 : : */
302 : 3753926 : vistest = GlobalVisTestFor(relation);
303 : :
304 [ + + ]: 3753926 : if (!GlobalVisTestIsRemovableXid(vistest, prune_xid, true))
305 : 1410753 : 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 [ + + ]: 2343173 : minfree = RelationGetTargetPageFreeSpace(relation,
320 : : HEAP_DEFAULT_FILLFACTOR);
321 : 2343173 : minfree = Max(minfree, BLCKSZ / 10);
322 : :
323 [ + + + + ]: 2343173 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
324 : : {
325 : 136812 : bool record_free_space = false;
326 : 136812 : 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 : 136812 : visibilitymap_pin(relation, BufferGetBlockNumber(buffer), vmbuffer);
335 : :
336 : : /* OK, try to get exclusive buffer lock */
337 [ + + ]: 136812 : if (!ConditionalLockBufferForCleanup(buffer))
338 : 1891 : 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 [ + + + + ]: 134921 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
346 : : {
347 : : OffsetNumber dummy_off_loc;
348 : : PruneFreezeResult presult;
349 : : PruneFreezeParams params;
350 : :
351 : 134919 : params.relation = relation;
352 : 134919 : params.buffer = buffer;
353 : 134919 : params.vmbuffer = *vmbuffer;
354 : 134919 : params.reason = PRUNE_ON_ACCESS;
355 : 134919 : params.vistest = vistest;
356 : 134919 : 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 : 134919 : params.options = HEAP_PAGE_PRUNE_ALLOW_FAST_PATH;
365 [ + + ]: 134919 : if (rel_read_only)
366 : 38613 : params.options |= HEAP_PAGE_PRUNE_SET_VM;
367 : :
368 : 134919 : 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 [ + + ]: 134919 : if (presult.ndeleted > presult.nnewlpdead)
386 : 20446 : pgstat_update_heap_dead_tuples(relation,
387 : 20446 : 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 [ + + ]: 134919 : if (presult.newly_all_visible)
397 : : {
398 : 14979 : record_free_space = true;
399 : 14979 : freespace = PageGetHeapFreeSpace(page);
400 : : }
401 : : }
402 : :
403 : : /* And release buffer lock */
404 : 134921 : 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 [ + + ]: 134921 : if (record_free_space)
413 : 14979 : 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 : 621164 : 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 : 621164 : prstate->vistest = params->vistest;
433 : 621164 : prstate->mark_unused_now =
434 : 621164 : (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 : 621164 : prstate->attempt_freeze = (params->options & HEAP_PAGE_PRUNE_FREEZE) != 0;
439 : 621164 : prstate->attempt_set_vm = (params->options & HEAP_PAGE_PRUNE_SET_VM) != 0;
440 : 621164 : prstate->cutoffs = params->cutoffs;
441 : 621164 : prstate->relation = params->relation;
442 : 621164 : prstate->block = BufferGetBlockNumber(params->buffer);
443 : 621164 : prstate->buffer = params->buffer;
444 : 621164 : prstate->page = BufferGetPage(params->buffer);
445 : :
446 : : Assert(BufferIsValid(params->vmbuffer));
447 : 621164 : prstate->vmbuffer = params->vmbuffer;
448 : 621164 : prstate->new_vmbits = 0;
449 : 621164 : prstate->old_vmbits = visibilitymap_get_status(prstate->relation,
450 : : prstate->block,
451 : : &prstate->vmbuffer);
452 : :
453 : : /*
454 : : * Our strategy is to scan the page and make lists of items to change,
455 : : * then apply the changes within a critical section. This keeps as much
456 : : * logic as possible out of the critical section, and also ensures that
457 : : * WAL replay will work the same as the normal case.
458 : : *
459 : : * First, initialize the new pd_prune_xid value to zero (indicating no
460 : : * prunable tuples). If we find any tuples which may soon become
461 : : * prunable, we will save the lowest relevant XID in new_prune_xid. Also
462 : : * initialize the rest of our working state.
463 : : */
464 : 621164 : prstate->new_prune_xid = InvalidTransactionId;
465 : 621164 : prstate->latest_xid_removed = InvalidTransactionId;
466 : 621164 : prstate->nredirected = prstate->ndead = prstate->nunused = 0;
467 : 621164 : prstate->nfrozen = 0;
468 : 621164 : prstate->nroot_items = 0;
469 : 621164 : prstate->nheaponly_items = 0;
470 : :
471 : : /* initialize page freezing working state */
472 : 621164 : prstate->pagefrz.freeze_required = false;
473 : 621164 : prstate->pagefrz.FreezePageConflictXid = InvalidTransactionId;
474 [ + + ]: 621164 : if (prstate->attempt_freeze)
475 : : {
476 : : Assert(new_relfrozen_xid && new_relmin_mxid);
477 : 486245 : prstate->pagefrz.FreezePageRelfrozenXid = *new_relfrozen_xid;
478 : 486245 : prstate->pagefrz.NoFreezePageRelfrozenXid = *new_relfrozen_xid;
479 : 486245 : prstate->pagefrz.FreezePageRelminMxid = *new_relmin_mxid;
480 : 486245 : prstate->pagefrz.NoFreezePageRelminMxid = *new_relmin_mxid;
481 : : }
482 : : else
483 : : {
484 : : Assert(!new_relfrozen_xid && !new_relmin_mxid);
485 : 134919 : prstate->pagefrz.FreezePageRelminMxid = InvalidMultiXactId;
486 : 134919 : prstate->pagefrz.NoFreezePageRelminMxid = InvalidMultiXactId;
487 : 134919 : prstate->pagefrz.FreezePageRelfrozenXid = InvalidTransactionId;
488 : 134919 : prstate->pagefrz.NoFreezePageRelfrozenXid = InvalidTransactionId;
489 : : }
490 : :
491 : 621164 : prstate->ndeleted = 0;
492 : 621164 : prstate->live_tuples = 0;
493 : 621164 : prstate->recently_dead_tuples = 0;
494 : 621164 : prstate->hastup = false;
495 : 621164 : prstate->lpdead_items = 0;
496 : :
497 : : /*
498 : : * deadoffsets are filled in during pruning but are only used to populate
499 : : * PruneFreezeResult->deadoffsets. To avoid needing two copies of the
500 : : * array, just save a pointer to the result offsets array in the
501 : : * PruneState.
502 : : */
503 : 621164 : prstate->deadoffsets = presult->deadoffsets;
504 : :
505 : : /*
506 : : * We track whether the page will be all-visible/all-frozen at the end of
507 : : * pruning and freezing. While examining tuple visibility, we'll set
508 : : * set_all_visible to false if there are tuples on the page not visible to
509 : : * all running and future transactions. If setting the VM is enabled for
510 : : * this scan, we will do so if the page ends up being all-visible.
511 : : *
512 : : * We also keep track of the newest live XID, which is used to calculate
513 : : * the snapshot conflict horizon for a WAL record setting the VM.
514 : : */
515 : 621164 : prstate->set_all_visible = prstate->attempt_set_vm;
516 : 621164 : prstate->newest_live_xid = InvalidTransactionId;
517 : :
518 : : /*
519 : : * Currently, only VACUUM performs freezing, but other callers may in the
520 : : * future. We must initialize set_all_frozen based on whether or not the
521 : : * caller passed HEAP_PAGE_PRUNE_FREEZE, because if they did not, we won't
522 : : * call heap_prepare_freeze_tuple() for each tuple, and set_all_frozen
523 : : * will never be cleared for tuples that need freezing. This would lead to
524 : : * incorrectly setting the visibility map all-frozen for this page. We
525 : : * can't set the page all-frozen in the VM if the caller didn't pass
526 : : * HEAP_PAGE_PRUNE_SET_VM.
527 : : *
528 : : * When freezing is not required (no XIDs/MXIDs older than the freeze
529 : : * cutoff), we may still choose to "opportunistically" freeze if doing so
530 : : * would make the page all-frozen.
531 : : *
532 : : * We will not be able to freeze the whole page at the end of vacuum if
533 : : * there are tuples present that are not visible to everyone or if there
534 : : * are dead tuples which will not be removable. However, dead tuples that
535 : : * will be removed by the end of vacuum should not prevent this
536 : : * opportunistic freezing.
537 : : *
538 : : * Therefore, we do not clear set_all_visible and set_all_frozen when we
539 : : * encounter LP_DEAD items. Instead, we correct them after deciding
540 : : * whether to freeze, but before updating the VM, to avoid setting the VM
541 : : * bits incorrectly.
542 : : */
543 [ + + + - ]: 621164 : prstate->set_all_frozen = prstate->attempt_freeze && prstate->attempt_set_vm;
544 : 621164 : }
545 : :
546 : : /*
547 : : * Helper for heap_page_prune_and_freeze(). Iterates over every tuple on the
548 : : * page, examines its visibility information, and determines the appropriate
549 : : * action for each tuple. All tuples are processed and classified during this
550 : : * phase, but no modifications are made to the page until the later execution
551 : : * stage.
552 : : *
553 : : * *off_loc is used for error callback and cleared before returning.
554 : : */
555 : : static void
556 : 435560 : prune_freeze_plan(PruneState *prstate, OffsetNumber *off_loc)
557 : : {
558 : 435560 : Page page = prstate->page;
559 : 435560 : BlockNumber blockno = prstate->block;
560 : 435560 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
561 : : OffsetNumber offnum;
562 : : HeapTupleData tup;
563 : :
564 : 435560 : tup.t_tableOid = RelationGetRelid(prstate->relation);
565 : :
566 : : /*
567 : : * Determine HTSV for all tuples, and queue them up for processing as HOT
568 : : * chain roots or as heap-only items.
569 : : *
570 : : * Determining HTSV only once for each tuple is required for correctness,
571 : : * to deal with cases where running HTSV twice could result in different
572 : : * results. For example, RECENTLY_DEAD can turn to DEAD if another
573 : : * checked item causes GlobalVisTestIsRemovableFullXid() to update the
574 : : * horizon, or INSERT_IN_PROGRESS can change to DEAD if the inserting
575 : : * transaction aborts.
576 : : *
577 : : * It's also good for performance. Most commonly tuples within a page are
578 : : * stored at decreasing offsets (while the items are stored at increasing
579 : : * offsets). When processing all tuples on a page this leads to reading
580 : : * memory at decreasing offsets within a page, with a variable stride.
581 : : * That's hard for CPU prefetchers to deal with. Processing the items in
582 : : * reverse order (and thus the tuples in increasing order) increases
583 : : * prefetching efficiency significantly / decreases the number of cache
584 : : * misses.
585 : : */
586 : 435560 : for (offnum = maxoff;
587 [ + + ]: 29397911 : offnum >= FirstOffsetNumber;
588 : 28962351 : offnum = OffsetNumberPrev(offnum))
589 : : {
590 : 28962351 : ItemId itemid = PageGetItemId(page, offnum);
591 : : HeapTupleHeader htup;
592 : :
593 : : /*
594 : : * Set the offset number so that we can display it along with any
595 : : * error that occurred while processing this tuple.
596 : : */
597 : 28962351 : *off_loc = offnum;
598 : :
599 : 28962351 : prstate->processed[offnum] = false;
600 : 28962351 : prstate->htsv[offnum] = -1;
601 : :
602 : : /* Nothing to do if slot doesn't contain a tuple */
603 [ + + ]: 28962351 : if (!ItemIdIsUsed(itemid))
604 : : {
605 : 205084 : heap_prune_record_unchanged_lp_unused(prstate, offnum);
606 : 205084 : continue;
607 : : }
608 : :
609 [ + + ]: 28757267 : if (ItemIdIsDead(itemid))
610 : : {
611 : : /*
612 : : * If the caller set mark_unused_now true, we can set dead line
613 : : * pointers LP_UNUSED now.
614 : : */
615 [ + + ]: 1630910 : if (unlikely(prstate->mark_unused_now))
616 : 1852 : heap_prune_record_unused(prstate, offnum, false);
617 : : else
618 : 1629058 : heap_prune_record_unchanged_lp_dead(prstate, offnum);
619 : 1630910 : continue;
620 : : }
621 : :
622 [ + + ]: 27126357 : if (ItemIdIsRedirected(itemid))
623 : : {
624 : : /* This is the start of a HOT chain */
625 : 197842 : prstate->root_items[prstate->nroot_items++] = offnum;
626 : 197842 : continue;
627 : : }
628 : :
629 : : Assert(ItemIdIsNormal(itemid));
630 : :
631 : : /*
632 : : * Get the tuple's visibility status and queue it up for processing.
633 : : */
634 : 26928515 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
635 : 26928515 : tup.t_data = htup;
636 : 26928515 : tup.t_len = ItemIdGetLength(itemid);
637 : 26928515 : ItemPointerSet(&tup.t_self, blockno, offnum);
638 : :
639 : 26928515 : prstate->htsv[offnum] = heap_prune_satisfies_vacuum(prstate, &tup);
640 : :
641 [ + + ]: 26928515 : if (!HeapTupleHeaderIsHeapOnly(htup))
642 : 26588589 : prstate->root_items[prstate->nroot_items++] = offnum;
643 : : else
644 : 339926 : prstate->heaponly_items[prstate->nheaponly_items++] = offnum;
645 : : }
646 : :
647 : : /*
648 : : * Process HOT chains.
649 : : *
650 : : * We added the items to the array starting from 'maxoff', so by
651 : : * processing the array in reverse order, we process the items in
652 : : * ascending offset number order. The order doesn't matter for
653 : : * correctness, but some quick micro-benchmarking suggests that this is
654 : : * faster. (Earlier PostgreSQL versions, which scanned all the items on
655 : : * the page instead of using the root_items array, also did it in
656 : : * ascending offset number order.)
657 : : */
658 [ + + ]: 27221991 : for (int i = prstate->nroot_items - 1; i >= 0; i--)
659 : : {
660 : 26786431 : offnum = prstate->root_items[i];
661 : :
662 : : /* Ignore items already processed as part of an earlier chain */
663 [ - + ]: 26786431 : if (prstate->processed[offnum])
664 : 0 : continue;
665 : :
666 : : /* see preceding loop */
667 : 26786431 : *off_loc = offnum;
668 : :
669 : : /* Process this item or chain of items */
670 : 26786431 : heap_prune_chain(maxoff, offnum, prstate);
671 : : }
672 : :
673 : : /*
674 : : * Process any heap-only tuples that were not already processed as part of
675 : : * a HOT chain.
676 : : */
677 [ + + ]: 775486 : for (int i = prstate->nheaponly_items - 1; i >= 0; i--)
678 : : {
679 : 339926 : offnum = prstate->heaponly_items[i];
680 : :
681 [ + + ]: 339926 : if (prstate->processed[offnum])
682 : 323260 : continue;
683 : :
684 : : /* see preceding loop */
685 : 16666 : *off_loc = offnum;
686 : :
687 : : /*
688 : : * If the tuple is DEAD and doesn't chain to anything else, mark it
689 : : * unused. (If it does chain, we can only remove it as part of
690 : : * pruning its chain.)
691 : : *
692 : : * We need this primarily to handle aborted HOT updates, that is,
693 : : * XMIN_INVALID heap-only tuples. Those might not be linked to by any
694 : : * chain, since the parent tuple might be re-updated before any
695 : : * pruning occurs. So we have to be able to reap them separately from
696 : : * chain-pruning. (Note that HeapTupleHeaderIsHotUpdated will never
697 : : * return true for an XMIN_INVALID tuple, so this code will work even
698 : : * when there were sequential updates within the aborted transaction.)
699 : : */
700 [ + + ]: 16666 : if (prstate->htsv[offnum] == HEAPTUPLE_DEAD)
701 : : {
702 : 3268 : ItemId itemid = PageGetItemId(page, offnum);
703 : 3268 : HeapTupleHeader htup = (HeapTupleHeader) PageGetItem(page, itemid);
704 : :
705 [ + - ]: 3268 : if (likely(!HeapTupleHeaderIsHotUpdated(htup)))
706 : : {
707 : 3268 : HeapTupleHeaderAdvanceConflictHorizon(htup,
708 : : &prstate->latest_xid_removed);
709 : 3268 : heap_prune_record_unused(prstate, offnum, true);
710 : : }
711 : : else
712 : : {
713 : : /*
714 : : * This tuple should've been processed and removed as part of
715 : : * a HOT chain, so something's wrong. To preserve evidence,
716 : : * we don't dare to remove it. We cannot leave behind a DEAD
717 : : * tuple either, because that will cause VACUUM to error out.
718 : : * Throwing an error with a distinct error message seems like
719 : : * the least bad option.
720 : : */
721 [ # # ]: 0 : elog(ERROR, "dead heap-only tuple (%u, %d) is not linked to from any HOT chain",
722 : : blockno, offnum);
723 : : }
724 : : }
725 : : else
726 : 13398 : heap_prune_record_unchanged_lp_normal(prstate, offnum);
727 : : }
728 : :
729 : : /* We should now have processed every tuple exactly once */
730 : : #ifdef USE_ASSERT_CHECKING
731 : : for (offnum = FirstOffsetNumber;
732 : : offnum <= maxoff;
733 : : offnum = OffsetNumberNext(offnum))
734 : : {
735 : : *off_loc = offnum;
736 : :
737 : : Assert(prstate->processed[offnum]);
738 : : }
739 : : #endif
740 : :
741 : : /* Clear the offset information once we have processed the given page. */
742 : 435560 : *off_loc = InvalidOffsetNumber;
743 : 435560 : }
744 : :
745 : : /*
746 : : * Decide whether to proceed with freezing according to the freeze plans
747 : : * prepared for the current heap buffer. If freezing is chosen, this function
748 : : * performs several pre-freeze checks.
749 : : *
750 : : * The values of do_prune, do_hint_prune, and did_tuple_hint_fpi must be
751 : : * determined before calling this function.
752 : : *
753 : : * prstate is both an input and output parameter.
754 : : *
755 : : * Returns true if we should apply the freeze plans and freeze tuples on the
756 : : * page, and false otherwise.
757 : : */
758 : : static bool
759 : 435560 : heap_page_will_freeze(bool did_tuple_hint_fpi,
760 : : bool do_prune,
761 : : bool do_hint_prune,
762 : : PruneState *prstate)
763 : : {
764 : 435560 : bool do_freeze = false;
765 : :
766 : : /*
767 : : * If the caller specified we should not attempt to freeze any tuples,
768 : : * validate that everything is in the right state and return.
769 : : */
770 [ + + ]: 435560 : if (!prstate->attempt_freeze)
771 : : {
772 : : Assert(!prstate->set_all_frozen && prstate->nfrozen == 0);
773 : 134919 : return false;
774 : : }
775 : :
776 [ + + ]: 300641 : if (prstate->pagefrz.freeze_required)
777 : : {
778 : : /*
779 : : * heap_prepare_freeze_tuple indicated that at least one XID/MXID from
780 : : * before FreezeLimit/MultiXactCutoff is present. Must freeze to
781 : : * advance relfrozenxid/relminmxid.
782 : : */
783 : 23513 : do_freeze = true;
784 : : }
785 : : else
786 : : {
787 : : /*
788 : : * Opportunistically freeze the page if we are generating an FPI
789 : : * anyway and if doing so means that we can set the page all-frozen
790 : : * afterwards (might not happen until VACUUM's final heap pass).
791 : : *
792 : : * XXX: Previously, we knew if pruning emitted an FPI by checking
793 : : * pgWalUsage.wal_fpi before and after pruning. Once the freeze and
794 : : * prune records were combined, this heuristic couldn't be used
795 : : * anymore. The opportunistic freeze heuristic must be improved;
796 : : * however, for now, try to approximate the old logic.
797 : : */
798 [ + + + + ]: 277128 : if (prstate->set_all_frozen && prstate->nfrozen > 0)
799 : : {
800 : : Assert(prstate->set_all_visible);
801 : :
802 : : /*
803 : : * Freezing would make the page all-frozen. Have already emitted
804 : : * an FPI or will do so anyway?
805 : : */
806 [ + + + + : 23489 : if (RelationNeedsWAL(prstate->relation))
+ - + - ]
807 : : {
808 [ + + ]: 21449 : if (did_tuple_hint_fpi)
809 : 1185 : do_freeze = true;
810 [ + + ]: 20264 : else if (do_prune)
811 : : {
812 [ + + ]: 2276 : if (XLogCheckBufferNeedsBackup(prstate->buffer))
813 : 738 : do_freeze = true;
814 : : }
815 [ + + ]: 17988 : else if (do_hint_prune)
816 : : {
817 [ + + + - : 22360 : if (XLogHintBitIsNeeded() &&
+ + ]
818 : 11180 : XLogCheckBufferNeedsBackup(prstate->buffer))
819 : 1908 : do_freeze = true;
820 : : }
821 : : }
822 : : }
823 : : }
824 : :
825 [ + + ]: 300641 : if (do_freeze)
826 : : {
827 : : /*
828 : : * Validate the tuples we will be freezing before entering the
829 : : * critical section.
830 : : */
831 : 27344 : heap_pre_freeze_checks(prstate->buffer, prstate->frozen, prstate->nfrozen);
832 : : Assert(TransactionIdPrecedes(prstate->pagefrz.FreezePageConflictXid,
833 : : prstate->cutoffs->OldestXmin));
834 : : }
835 [ + + ]: 273297 : else if (prstate->nfrozen > 0)
836 : : {
837 : : /*
838 : : * The page contained some tuples that were not already frozen, and we
839 : : * chose not to freeze them now. The page won't be all-frozen then.
840 : : */
841 : : Assert(!prstate->pagefrz.freeze_required);
842 : :
843 : 20036 : prstate->set_all_frozen = false;
844 : 20036 : prstate->nfrozen = 0; /* avoid miscounts in instrumentation */
845 : : }
846 : : else
847 : : {
848 : : /*
849 : : * We have no freeze plans to execute. The page might already be
850 : : * all-frozen (perhaps only following pruning), though. Such pages
851 : : * can be marked all-frozen in the VM by our caller, even though none
852 : : * of its tuples were newly frozen here.
853 : : */
854 : : }
855 : :
856 : 300641 : return do_freeze;
857 : : }
858 : :
859 : : /*
860 : : * Emit a warning about and fix visibility map corruption on the given page.
861 : : *
862 : : * The caller specifies the type of corruption it has already detected via
863 : : * corruption_type, so that we can emit the appropriate warning. All cases
864 : : * result in the VM bits being cleared; corruption types where PD_ALL_VISIBLE
865 : : * is incorrectly set also clear PD_ALL_VISIBLE.
866 : : *
867 : : * Must be called while holding an exclusive lock on the heap buffer. Dead
868 : : * items and not all-visible tuples must have been discovered under that same
869 : : * lock. Although we do not hold a lock on the VM buffer, it is pinned, and
870 : : * the heap buffer is exclusively locked, ensuring that no other backend can
871 : : * update the VM bits corresponding to this heap page.
872 : : *
873 : : * This function makes changes to the VM and, potentially, the heap page, but
874 : : * it does not need to be done in a critical section.
875 : : */
876 : : static void
877 : 0 : heap_page_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum,
878 : : VMCorruptionType corruption_type)
879 : : {
880 : 0 : const char *relname = RelationGetRelationName(prstate->relation);
881 : 0 : bool do_clear_vm = false;
882 : 0 : bool do_clear_heap = false;
883 : :
884 : : Assert(BufferIsLockedByMeInMode(prstate->buffer, BUFFER_LOCK_EXCLUSIVE));
885 : :
886 [ # # # # ]: 0 : switch (corruption_type)
887 : : {
888 : 0 : case VM_CORRUPT_LPDEAD:
889 [ # # ]: 0 : ereport(WARNING,
890 : : (errcode(ERRCODE_DATA_CORRUPTED),
891 : : errmsg("dead line pointer found on page marked all-visible"),
892 : : errcontext("relation \"%s\", page %u, tuple %u",
893 : : relname, prstate->block, offnum)));
894 : 0 : do_clear_vm = true;
895 : 0 : do_clear_heap = true;
896 : 0 : break;
897 : :
898 : 0 : case VM_CORRUPT_TUPLE_VISIBILITY:
899 : :
900 : : /*
901 : : * A HEAPTUPLE_LIVE tuple on an all-visible page can appear to not
902 : : * be visible to everyone when
903 : : * GetOldestNonRemovableTransactionId() returns a conservative
904 : : * value that's older than the real safe xmin. That is not
905 : : * corruption -- the PD_ALL_VISIBLE flag is still correct.
906 : : *
907 : : * However, dead tuple versions, in-progress inserts, and
908 : : * in-progress deletes should never appear on a page marked
909 : : * all-visible. That indicates real corruption. PD_ALL_VISIBLE
910 : : * should have been cleared by the DML operation that deleted or
911 : : * inserted the tuple.
912 : : */
913 [ # # ]: 0 : ereport(WARNING,
914 : : (errcode(ERRCODE_DATA_CORRUPTED),
915 : : errmsg("tuple not visible to all transactions found on page marked all-visible"),
916 : : errcontext("relation \"%s\", page %u, tuple %u",
917 : : relname, prstate->block, offnum)));
918 : 0 : do_clear_vm = true;
919 : 0 : do_clear_heap = true;
920 : 0 : break;
921 : :
922 : 0 : case VM_CORRUPT_MISSING_PAGE_HINT:
923 : :
924 : : /*
925 : : * As of PostgreSQL 9.2, the visibility map bit should never be
926 : : * set if the page-level bit is clear. However, for vacuum, it's
927 : : * possible that the bit got cleared after
928 : : * heap_vac_scan_next_block() was called, so we must recheck now
929 : : * that we have the buffer lock before concluding that the VM is
930 : : * corrupt.
931 : : */
932 : : Assert(!PageIsAllVisible(prstate->page));
933 : : Assert(prstate->old_vmbits & VISIBILITYMAP_VALID_BITS);
934 [ # # ]: 0 : ereport(WARNING,
935 : : (errcode(ERRCODE_DATA_CORRUPTED),
936 : : errmsg("page is not marked all-visible but visibility map bit is set"),
937 : : errcontext("relation \"%s\", page %u",
938 : : relname, prstate->block)));
939 : 0 : do_clear_vm = true;
940 : 0 : break;
941 : : }
942 : :
943 : : Assert(do_clear_heap || do_clear_vm);
944 : :
945 : : /* Avoid marking the buffer dirty if PD_ALL_VISIBLE is already clear */
946 [ # # ]: 0 : if (do_clear_heap)
947 : : {
948 : : Assert(PageIsAllVisible(prstate->page));
949 : 0 : PageClearAllVisible(prstate->page);
950 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
951 : : }
952 : :
953 [ # # ]: 0 : if (do_clear_vm)
954 : : {
955 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_EXCLUSIVE);
956 : : /* This VM clear is not WAL-logged, so its return value is not needed. */
957 : 0 : (void) visibilitymap_clear(prstate->relation->rd_locator,
958 : : prstate->block, prstate->vmbuffer,
959 : : VISIBILITYMAP_VALID_BITS);
960 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_UNLOCK);
961 : 0 : prstate->old_vmbits = 0;
962 : : }
963 : 0 : }
964 : :
965 : : /*
966 : : * Decide whether to set the visibility map bits (all-visible and all-frozen)
967 : : * for the current page using information from the PruneState and VM.
968 : : *
969 : : * This function does not actually set the VM bits or page-level visibility
970 : : * hint, PD_ALL_VISIBLE.
971 : : *
972 : : * This should be called only after do_freeze has been decided (and do_prune
973 : : * has been set), as these factor into our heuristic-based decision.
974 : : *
975 : : * Returns true if one or both VM bits should be set and false otherwise.
976 : : */
977 : : static bool
978 : 435560 : heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
979 : : bool do_prune, bool do_freeze)
980 : : {
981 [ + + ]: 435560 : if (!prstate->attempt_set_vm)
982 : 96306 : return false;
983 : :
984 [ + + ]: 339254 : if (!prstate->set_all_visible)
985 : 258218 : return false;
986 : :
987 : : /*
988 : : * If this is an on-access call and we're not actually pruning, avoid
989 : : * setting the visibility map if it would newly dirty the heap page or, if
990 : : * the page is already dirty, if doing so would require including a
991 : : * full-page image (FPI) of the heap page in the WAL.
992 : : */
993 [ + + + + : 81036 : if (reason == PRUNE_ON_ACCESS && !do_prune && !do_freeze &&
+ - ]
994 [ + + + + ]: 30625 : (!BufferIsDirty(prstate->buffer) || XLogCheckBufferNeedsBackup(prstate->buffer)))
995 : : {
996 : 15757 : prstate->set_all_visible = prstate->set_all_frozen = false;
997 : 15757 : return false;
998 : : }
999 : :
1000 : 65279 : prstate->new_vmbits = VISIBILITYMAP_ALL_VISIBLE;
1001 : :
1002 [ + + ]: 65279 : if (prstate->set_all_frozen)
1003 : 34764 : prstate->new_vmbits |= VISIBILITYMAP_ALL_FROZEN;
1004 : :
1005 [ + + ]: 65279 : if (prstate->new_vmbits == prstate->old_vmbits)
1006 : : {
1007 : 1720 : prstate->new_vmbits = 0;
1008 : 1720 : return false;
1009 : : }
1010 : :
1011 : 63559 : return true;
1012 : : }
1013 : :
1014 : : /*
1015 : : * If the page is already all-frozen, or already all-visible and freezing
1016 : : * won't be attempted, there is no remaining work and we can use the fast path
1017 : : * to avoid the expensive overhead of heap_page_prune_and_freeze().
1018 : : *
1019 : : * This can happen when the page has a stale prune hint, or if VACUUM is
1020 : : * scanning an already all-frozen page due to SKIP_PAGES_THRESHOLD.
1021 : : *
1022 : : * The caller must already have examined the visibility map and saved the
1023 : : * status of the page's VM bits in prstate->old_vmbits. Caller must hold a
1024 : : * content lock on the heap page since it will examine line pointers.
1025 : : *
1026 : : * Before calling prune_freeze_fast_path(), the caller should first
1027 : : * check for and fix any discrepancy between the page-level visibility hint
1028 : : * and the visibility map. Otherwise, the fast path will always prevent us
1029 : : * from getting them in sync. Note that if there are tuples on the page that
1030 : : * are not visible to all but the VM is incorrectly marked
1031 : : * all-visible/all-frozen, we will not get the chance to fix that corruption
1032 : : * when using the fast path.
1033 : : */
1034 : : static void
1035 : 185604 : prune_freeze_fast_path(PruneState *prstate, PruneFreezeResult *presult)
1036 : : {
1037 : 185604 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
1038 : 185604 : Page page = prstate->page;
1039 : :
1040 : : Assert((prstate->old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1041 : : ((prstate->old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1042 : : !prstate->attempt_freeze));
1043 : :
1044 : : /* We'll fill in presult for the caller */
1045 : 185604 : memset(presult, 0, sizeof(PruneFreezeResult));
1046 : :
1047 : : /* Clear any stale prune hint */
1048 [ - + ]: 185604 : if (TransactionIdIsValid(PageGetPruneXid(page)))
1049 : : {
1050 : 0 : PageClearPrunable(page);
1051 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
1052 : : }
1053 : :
1054 [ - + ]: 185604 : if (PageIsEmpty(page))
1055 : 0 : return;
1056 : :
1057 : : /*
1058 : : * Since the page is all-visible, a count of the normal ItemIds on the
1059 : : * page should be sufficient for vacuum's live tuple count.
1060 : : */
1061 : 185604 : for (OffsetNumber off = FirstOffsetNumber;
1062 [ + + ]: 10692294 : off <= maxoff;
1063 : 10506690 : off = OffsetNumberNext(off))
1064 : : {
1065 : 10506690 : ItemId lp = PageGetItemId(page, off);
1066 : :
1067 [ + + ]: 10506690 : if (!ItemIdIsUsed(lp))
1068 : 251204 : continue;
1069 : :
1070 : 10255486 : presult->hastup = true;
1071 : :
1072 [ + + ]: 10255486 : if (ItemIdIsNormal(lp))
1073 : 10093625 : prstate->live_tuples++;
1074 : : }
1075 : :
1076 : 185604 : presult->live_tuples = prstate->live_tuples;
1077 : : }
1078 : :
1079 : : /*
1080 : : * Prune and repair fragmentation and potentially freeze tuples on the
1081 : : * specified page. If the page's visibility status has changed, update it in
1082 : : * the VM.
1083 : : *
1084 : : * Caller must have pin and buffer cleanup lock on the page. Note that we
1085 : : * don't update the FSM information for page on caller's behalf. Caller might
1086 : : * also need to account for a reduction in the length of the line pointer
1087 : : * array following array truncation by us.
1088 : : *
1089 : : * params contains the input parameters used to control freezing and pruning
1090 : : * behavior. See the definition of PruneFreezeParams for more on what each
1091 : : * parameter does.
1092 : : *
1093 : : * If the HEAP_PAGE_PRUNE_FREEZE option is set in params, we will freeze
1094 : : * tuples if it's required in order to advance relfrozenxid / relminmxid, or
1095 : : * if it's considered advantageous for overall system performance to do so
1096 : : * now. The 'params.cutoffs', 'presult', 'new_relfrozen_xid' and
1097 : : * 'new_relmin_mxid' arguments are required when freezing.
1098 : : *
1099 : : * A vmbuffer corresponding to the heap page is also passed and if the page is
1100 : : * found to be all-visible/all-frozen, we will set it in the VM.
1101 : : *
1102 : : * presult contains output parameters needed by callers, such as the number of
1103 : : * tuples removed and the offsets of dead items on the page after pruning.
1104 : : * heap_page_prune_and_freeze() is responsible for initializing it. Required
1105 : : * by all callers.
1106 : : *
1107 : : * off_loc is the offset location required by the caller to use in error
1108 : : * callback.
1109 : : *
1110 : : * new_relfrozen_xid and new_relmin_mxid must be provided by the caller if the
1111 : : * HEAP_PAGE_PRUNE_FREEZE option is set in params. On entry, they contain the
1112 : : * oldest XID and multi-XID seen on the relation so far. They will be updated
1113 : : * with the oldest values present on the page after pruning. After processing
1114 : : * the whole relation, VACUUM can use these values as the new
1115 : : * relfrozenxid/relminmxid for the relation.
1116 : : */
1117 : : void
1118 : 621164 : heap_page_prune_and_freeze(PruneFreezeParams *params,
1119 : : PruneFreezeResult *presult,
1120 : : OffsetNumber *off_loc,
1121 : : TransactionId *new_relfrozen_xid,
1122 : : MultiXactId *new_relmin_mxid)
1123 : : {
1124 : : PruneState prstate;
1125 : : bool do_freeze;
1126 : : bool do_prune;
1127 : : bool do_hint_prune;
1128 : : bool do_set_vm;
1129 : : bool did_tuple_hint_fpi;
1130 : 621164 : int64 fpi_before = pgWalUsage.wal_fpi;
1131 : : TransactionId conflict_xid;
1132 : :
1133 : : /* Initialize prstate */
1134 : 621164 : prune_freeze_setup(params,
1135 : : new_relfrozen_xid, new_relmin_mxid,
1136 : : presult, &prstate);
1137 : :
1138 : : /*
1139 : : * If the VM is set but PD_ALL_VISIBLE is clear, fix that corruption
1140 : : * before pruning and freezing so that the page and VM start out in a
1141 : : * consistent state.
1142 : : */
1143 [ + + ]: 621164 : if ((prstate.old_vmbits & VISIBILITYMAP_VALID_BITS) &&
1144 [ - + ]: 190967 : !PageIsAllVisible(prstate.page))
1145 : 0 : heap_page_fix_vm_corruption(&prstate, InvalidOffsetNumber,
1146 : : VM_CORRUPT_MISSING_PAGE_HINT);
1147 : :
1148 : : /*
1149 : : * If the page is already all-frozen, or already all-visible when freezing
1150 : : * is not being attempted, take the fast path, skipping pruning and
1151 : : * freezing code entirely. This must be done after fixing any discrepancy
1152 : : * between the page-level visibility hint and the VM, since that may have
1153 : : * cleared old_vmbits.
1154 : : */
1155 [ + + ]: 621164 : if ((params->options & HEAP_PAGE_PRUNE_ALLOW_FAST_PATH) != 0 &&
1156 [ + + ]: 619874 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1157 [ + + ]: 434270 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1158 [ - + ]: 4918 : !prstate.attempt_freeze)))
1159 : : {
1160 : 185604 : prune_freeze_fast_path(&prstate, presult);
1161 : 185604 : return;
1162 : : }
1163 : :
1164 : : /*
1165 : : * Examine all line pointers and tuple visibility information to determine
1166 : : * which line pointers should change state and which tuples may be frozen.
1167 : : * Prepare queue of state changes to later be executed in a critical
1168 : : * section.
1169 : : */
1170 : 435560 : prune_freeze_plan(&prstate, off_loc);
1171 : :
1172 : : /*
1173 : : * After processing all the live tuples on the page, if the newest xmin
1174 : : * amongst them may be considered running by any snapshot, the page cannot
1175 : : * be all-visible. This should be done before determining whether or not
1176 : : * to opportunistically freeze.
1177 : : */
1178 [ + + ]: 435560 : if (prstate.set_all_visible &&
1179 [ + + + + ]: 193762 : TransactionIdIsNormal(prstate.newest_live_xid) &&
1180 : 82485 : GlobalVisTestXidConsideredRunning(prstate.vistest,
1181 : : prstate.newest_live_xid,
1182 : : true))
1183 : 2833 : prstate.set_all_visible = prstate.set_all_frozen = false;
1184 : :
1185 : : /*
1186 : : * If checksums are enabled, calling heap_prune_satisfies_vacuum() while
1187 : : * checking tuple visibility information in prune_freeze_plan() may have
1188 : : * caused an FPI to be emitted.
1189 : : */
1190 : 435560 : did_tuple_hint_fpi = fpi_before != pgWalUsage.wal_fpi;
1191 : :
1192 : 1288891 : do_prune = prstate.nredirected > 0 ||
1193 [ + + + + ]: 800252 : prstate.ndead > 0 ||
1194 [ + + ]: 364692 : prstate.nunused > 0;
1195 : :
1196 : : /*
1197 : : * Even if we don't prune anything, if we found a new value for the
1198 : : * pd_prune_xid field or the page was marked full, we will update the hint
1199 : : * bit.
1200 : : */
1201 [ + + + + ]: 692174 : do_hint_prune = PageGetPruneXid(prstate.page) != prstate.new_prune_xid ||
1202 : 256614 : PageIsFull(prstate.page);
1203 : :
1204 : : /*
1205 : : * Decide if we want to go ahead with freezing according to the freeze
1206 : : * plans we prepared, or not.
1207 : : */
1208 : 435560 : do_freeze = heap_page_will_freeze(did_tuple_hint_fpi,
1209 : : do_prune,
1210 : : do_hint_prune,
1211 : : &prstate);
1212 : :
1213 : : /*
1214 : : * While scanning the line pointers, we did not clear
1215 : : * set_all_visible/set_all_frozen when encountering LP_DEAD items because
1216 : : * we wanted the decision whether or not to freeze the page to be
1217 : : * unaffected by the short-term presence of LP_DEAD items. These LP_DEAD
1218 : : * items are effectively assumed to be LP_UNUSED items in the making. It
1219 : : * doesn't matter which vacuum heap pass (initial pass or final pass) ends
1220 : : * up setting the page all-frozen, as long as the ongoing VACUUM does it.
1221 : : *
1222 : : * Now that we finished determining whether or not to freeze the page,
1223 : : * update set_all_visible and set_all_frozen so that they reflect the true
1224 : : * state of the page for setting PD_ALL_VISIBLE and VM bits.
1225 : : */
1226 [ + + ]: 435560 : if (prstate.lpdead_items > 0)
1227 : 74102 : prstate.set_all_visible = prstate.set_all_frozen = false;
1228 : :
1229 : : Assert(!prstate.set_all_frozen || prstate.set_all_visible);
1230 : : Assert(!prstate.set_all_visible || prstate.attempt_set_vm);
1231 : : Assert(!prstate.set_all_visible || (prstate.lpdead_items == 0));
1232 : :
1233 : 435560 : do_set_vm = heap_page_will_set_vm(&prstate, params->reason, do_prune, do_freeze);
1234 : :
1235 : : /*
1236 : : * new_vmbits should be 0 regardless of whether or not the page is
1237 : : * all-visible if we do not intend to set the VM.
1238 : : */
1239 : : Assert(do_set_vm || prstate.new_vmbits == 0);
1240 : :
1241 : : /*
1242 : : * The snapshot conflict horizon for the whole record is the most
1243 : : * conservative (newest) horizon required by any change in the record.
1244 : : */
1245 : 435560 : conflict_xid = InvalidTransactionId;
1246 [ + + ]: 435560 : if (do_set_vm)
1247 : 63559 : conflict_xid = prstate.newest_live_xid;
1248 [ + + + + ]: 435560 : if (do_freeze && TransactionIdFollows(prstate.pagefrz.FreezePageConflictXid, conflict_xid))
1249 : 4460 : conflict_xid = prstate.pagefrz.FreezePageConflictXid;
1250 [ + + + + ]: 435560 : if (do_prune && TransactionIdFollows(prstate.latest_xid_removed, conflict_xid))
1251 : 62900 : conflict_xid = prstate.latest_xid_removed;
1252 : :
1253 : : /* Lock vmbuffer before entering a critical section */
1254 [ + + ]: 435560 : if (do_set_vm)
1255 : 63559 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_EXCLUSIVE);
1256 : :
1257 : : /* Any error while applying the changes is critical */
1258 : 435560 : START_CRIT_SECTION();
1259 : :
1260 [ + + ]: 435560 : if (do_hint_prune)
1261 : : {
1262 : : /*
1263 : : * Update the page's pd_prune_xid field to either zero, or the lowest
1264 : : * XID of any soon-prunable tuple.
1265 : : */
1266 : 178987 : ((PageHeader) prstate.page)->pd_prune_xid = prstate.new_prune_xid;
1267 : :
1268 : : /*
1269 : : * Also clear the "page is full" flag, since there's no point in
1270 : : * repeating the prune/defrag process until something else happens to
1271 : : * the page.
1272 : : */
1273 : 178987 : PageClearFull(prstate.page);
1274 : :
1275 : : /*
1276 : : * If that's all we had to do to the page, this is a non-WAL-logged
1277 : : * hint. If we are going to freeze or prune the page or set
1278 : : * PD_ALL_VISIBLE, we will mark the buffer dirty below.
1279 : : *
1280 : : * Setting PD_ALL_VISIBLE is fully WAL-logged because it is forbidden
1281 : : * for the VM to be set and PD_ALL_VISIBLE to be clear.
1282 : : */
1283 [ + + + + : 178987 : if (!do_freeze && !do_prune && !do_set_vm)
+ + ]
1284 : 68268 : MarkBufferDirtyHint(prstate.buffer, true);
1285 : : }
1286 : :
1287 [ + + + + : 435560 : if (do_prune || do_freeze || do_set_vm)
+ + ]
1288 : : {
1289 : : /* Apply the planned item changes and repair page fragmentation. */
1290 [ + + ]: 136594 : if (do_prune)
1291 : : {
1292 : 71308 : heap_page_prune_execute(prstate.buffer, false,
1293 : : prstate.redirected, prstate.nredirected,
1294 : : prstate.nowdead, prstate.ndead,
1295 : : prstate.nowunused, prstate.nunused);
1296 : : }
1297 : :
1298 [ + + ]: 136594 : if (do_freeze)
1299 : 27344 : heap_freeze_prepared_tuples(prstate.buffer, prstate.frozen, prstate.nfrozen);
1300 : :
1301 : : /* Set the visibility map and page visibility hint */
1302 [ + + ]: 136594 : if (do_set_vm)
1303 : : {
1304 : : /*
1305 : : * While it is valid for PD_ALL_VISIBLE to be set when the
1306 : : * corresponding VM bit is clear, we strongly prefer to keep them
1307 : : * in sync.
1308 : : *
1309 : : * The heap buffer must be marked dirty before adding it to the
1310 : : * WAL chain when setting the VM. We don't worry about
1311 : : * unnecessarily dirtying the heap buffer if PD_ALL_VISIBLE is
1312 : : * already set, though. It is extremely rare to have a clean heap
1313 : : * buffer with PD_ALL_VISIBLE already set and the VM bits clear,
1314 : : * so there is no point in optimizing it.
1315 : : */
1316 : 63559 : PageSetAllVisible(prstate.page);
1317 : 63559 : PageClearPrunable(prstate.page);
1318 : 63559 : (void) visibilitymap_set(prstate.block, prstate.vmbuffer,
1319 : 63559 : prstate.new_vmbits,
1320 : 63559 : prstate.relation->rd_locator);
1321 : : }
1322 : :
1323 : 136594 : MarkBufferDirty(prstate.buffer);
1324 : :
1325 : : /*
1326 : : * Emit a WAL XLOG_HEAP2_PRUNE* record showing what we did
1327 : : */
1328 [ + + + + : 136594 : if (RelationNeedsWAL(prstate.relation))
+ - + - ]
1329 : : {
1330 [ + + + + ]: 194914 : log_heap_prune_and_freeze(prstate.relation, prstate.buffer,
1331 : : do_set_vm ? prstate.vmbuffer : InvalidBuffer,
1332 : 61588 : do_set_vm ? prstate.new_vmbits : 0,
1333 : : conflict_xid,
1334 : : true, /* cleanup lock */
1335 : : params->reason,
1336 : : prstate.frozen, prstate.nfrozen,
1337 : : prstate.redirected, prstate.nredirected,
1338 : : prstate.nowdead, prstate.ndead,
1339 : : prstate.nowunused, prstate.nunused);
1340 : : }
1341 : : }
1342 : :
1343 : 435560 : END_CRIT_SECTION();
1344 : :
1345 [ + + ]: 435560 : if (do_set_vm)
1346 : 63559 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_UNLOCK);
1347 : :
1348 : : /*
1349 : : * During its second pass over the heap, VACUUM calls
1350 : : * heap_page_would_be_all_visible() to determine whether a page is
1351 : : * all-visible and all-frozen. The logic here is similar. After completing
1352 : : * pruning and freezing, use an assertion to verify that our results
1353 : : * remain consistent with heap_page_would_be_all_visible(). It's also a
1354 : : * valuable cross-check of the page state after pruning and freezing.
1355 : : */
1356 : : #ifdef USE_ASSERT_CHECKING
1357 : : if (prstate.set_all_visible)
1358 : : {
1359 : : TransactionId debug_cutoff;
1360 : : bool debug_all_frozen;
1361 : :
1362 : : Assert(prstate.lpdead_items == 0);
1363 : :
1364 : : Assert(heap_page_is_all_visible(prstate.relation, prstate.buffer,
1365 : : prstate.vistest,
1366 : : &debug_all_frozen,
1367 : : &debug_cutoff, off_loc));
1368 : :
1369 : : Assert(!TransactionIdIsValid(debug_cutoff) ||
1370 : : debug_cutoff == prstate.newest_live_xid);
1371 : :
1372 : : /*
1373 : : * It's possible the page is composed entirely of frozen tuples but is
1374 : : * not set all-frozen in the VM and did not pass
1375 : : * HEAP_PAGE_PRUNE_FREEZE. In this case, it's possible
1376 : : * heap_page_is_all_visible() finds the page completely frozen, even
1377 : : * though prstate.set_all_frozen is false.
1378 : : */
1379 : : Assert(!prstate.set_all_frozen || debug_all_frozen);
1380 : : }
1381 : : #endif
1382 : :
1383 : : /* Copy information back for caller */
1384 : 435560 : presult->ndeleted = prstate.ndeleted;
1385 : 435560 : presult->nnewlpdead = prstate.ndead;
1386 : 435560 : presult->nfrozen = prstate.nfrozen;
1387 : 435560 : presult->live_tuples = prstate.live_tuples;
1388 : 435560 : presult->recently_dead_tuples = prstate.recently_dead_tuples;
1389 : 435560 : presult->hastup = prstate.hastup;
1390 : :
1391 : 435560 : presult->lpdead_items = prstate.lpdead_items;
1392 : : /* the presult->deadoffsets array was already filled in */
1393 : :
1394 : 435560 : presult->newly_all_visible = false;
1395 : 435560 : presult->newly_all_frozen = false;
1396 : 435560 : presult->newly_all_visible_frozen = false;
1397 [ + + ]: 435560 : if (do_set_vm)
1398 : : {
1399 [ + + ]: 63559 : if ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
1400 : : {
1401 : 59916 : presult->newly_all_visible = true;
1402 [ + + ]: 59916 : if (prstate.set_all_frozen)
1403 : 30677 : presult->newly_all_visible_frozen = true;
1404 : : }
1405 [ + - ]: 3643 : else if ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
1406 [ + - ]: 3643 : prstate.set_all_frozen)
1407 : 3643 : presult->newly_all_frozen = true;
1408 : : }
1409 : :
1410 [ + + ]: 435560 : if (prstate.attempt_freeze)
1411 : : {
1412 [ + + ]: 300641 : if (presult->nfrozen > 0)
1413 : : {
1414 : 27344 : *new_relfrozen_xid = prstate.pagefrz.FreezePageRelfrozenXid;
1415 : 27344 : *new_relmin_mxid = prstate.pagefrz.FreezePageRelminMxid;
1416 : : }
1417 : : else
1418 : : {
1419 : 273297 : *new_relfrozen_xid = prstate.pagefrz.NoFreezePageRelfrozenXid;
1420 : 273297 : *new_relmin_mxid = prstate.pagefrz.NoFreezePageRelminMxid;
1421 : : }
1422 : : }
1423 : : }
1424 : :
1425 : :
1426 : : /*
1427 : : * Perform visibility checks for heap pruning.
1428 : : */
1429 : : static HTSV_Result
1430 : 26928515 : heap_prune_satisfies_vacuum(PruneState *prstate, HeapTuple tup)
1431 : : {
1432 : : HTSV_Result res;
1433 : : TransactionId dead_after;
1434 : :
1435 : 26928515 : res = HeapTupleSatisfiesVacuumHorizon(tup, prstate->buffer, &dead_after);
1436 : :
1437 [ + + ]: 26928515 : if (res != HEAPTUPLE_RECENTLY_DEAD)
1438 : 22663208 : return res;
1439 : :
1440 : : /*
1441 : : * For VACUUM, we must be sure to prune tuples with xmax older than
1442 : : * OldestXmin -- a visibility cutoff determined at the beginning of
1443 : : * vacuuming the relation. OldestXmin is used for freezing determination
1444 : : * and we cannot freeze dead tuples' xmaxes.
1445 : : */
1446 [ + + ]: 4265307 : if (prstate->cutoffs &&
1447 [ + - ]: 1298408 : TransactionIdIsValid(prstate->cutoffs->OldestXmin) &&
1448 [ + + ]: 1298408 : NormalTransactionIdPrecedes(dead_after, prstate->cutoffs->OldestXmin))
1449 : 968276 : return HEAPTUPLE_DEAD;
1450 : :
1451 : : /*
1452 : : * Determine whether or not the tuple is considered dead when compared
1453 : : * with the provided GlobalVisState. On-access pruning does not provide
1454 : : * VacuumCutoffs. And for vacuum, even if the tuple's xmax is not older
1455 : : * than OldestXmin, GlobalVisTestIsRemovableXid() could find the row dead
1456 : : * if the GlobalVisState has been updated since the beginning of vacuuming
1457 : : * the relation.
1458 : : */
1459 [ + + ]: 3297031 : if (GlobalVisTestIsRemovableXid(prstate->vistest, dead_after, true))
1460 : 2920612 : return HEAPTUPLE_DEAD;
1461 : :
1462 : 376419 : return res;
1463 : : }
1464 : :
1465 : :
1466 : : /*
1467 : : * Pruning calculates tuple visibility once and saves the results in an array
1468 : : * of int8. See PruneState.htsv for details. This helper function is meant
1469 : : * to guard against examining visibility status array members which have not
1470 : : * yet been computed.
1471 : : */
1472 : : static inline HTSV_Result
1473 : 26911849 : htsv_get_valid_status(int status)
1474 : : {
1475 : : Assert(status >= HEAPTUPLE_DEAD &&
1476 : : status <= HEAPTUPLE_DELETE_IN_PROGRESS);
1477 : 26911849 : return (HTSV_Result) status;
1478 : : }
1479 : :
1480 : : /*
1481 : : * Prune specified line pointer or a HOT chain originating at line pointer.
1482 : : *
1483 : : * Tuple visibility information is provided in prstate->htsv.
1484 : : *
1485 : : * If the item is an index-referenced tuple (i.e. not a heap-only tuple),
1486 : : * the HOT chain is pruned by removing all DEAD tuples at the start of the HOT
1487 : : * chain. We also prune any RECENTLY_DEAD tuples preceding a DEAD tuple.
1488 : : * This is OK because a RECENTLY_DEAD tuple preceding a DEAD tuple is really
1489 : : * DEAD, our visibility test is just too coarse to detect it.
1490 : : *
1491 : : * Pruning must never leave behind a DEAD tuple that still has tuple storage.
1492 : : * VACUUM isn't prepared to deal with that case.
1493 : : *
1494 : : * The root line pointer is redirected to the tuple immediately after the
1495 : : * latest DEAD tuple. If all tuples in the chain are DEAD, the root line
1496 : : * pointer is marked LP_DEAD. (This includes the case of a DEAD simple
1497 : : * tuple, which we treat as a chain of length 1.)
1498 : : *
1499 : : * We don't actually change the page here. We just add entries to the arrays in
1500 : : * prstate showing the changes to be made. Items to be redirected are added
1501 : : * to the redirected[] array (two entries per redirection); items to be set to
1502 : : * LP_DEAD state are added to nowdead[]; and items to be set to LP_UNUSED
1503 : : * state are added to nowunused[]. We perform bookkeeping of live tuples,
1504 : : * visibility etc. based on what the page will look like after the changes
1505 : : * applied. All that bookkeeping is performed in the heap_prune_record_*()
1506 : : * subroutines. The division of labor is that heap_prune_chain() decides the
1507 : : * fate of each tuple, ie. whether it's going to be removed, redirected or
1508 : : * left unchanged, and the heap_prune_record_*() subroutines update PruneState
1509 : : * based on that outcome.
1510 : : */
1511 : : static void
1512 : 26786431 : heap_prune_chain(OffsetNumber maxoff, OffsetNumber rootoffnum,
1513 : : PruneState *prstate)
1514 : : {
1515 : 26786431 : TransactionId priorXmax = InvalidTransactionId;
1516 : : ItemId rootlp;
1517 : : OffsetNumber offnum;
1518 : : OffsetNumber chainitems[MaxHeapTuplesPerPage];
1519 : 26786431 : Page page = prstate->page;
1520 : :
1521 : : /*
1522 : : * After traversing the HOT chain, ndeadchain is the index in chainitems
1523 : : * of the first live successor after the last dead item.
1524 : : */
1525 : 26786431 : int ndeadchain = 0,
1526 : 26786431 : nchain = 0;
1527 : :
1528 : 26786431 : rootlp = PageGetItemId(page, rootoffnum);
1529 : :
1530 : : /* Start from the root tuple */
1531 : 26786431 : offnum = rootoffnum;
1532 : :
1533 : : /* while not end of the chain */
1534 : : for (;;)
1535 : 323260 : {
1536 : : HeapTupleHeader htup;
1537 : : ItemId lp;
1538 : :
1539 : : /* Sanity check (pure paranoia) */
1540 [ - + ]: 27109691 : if (offnum < FirstOffsetNumber)
1541 : 0 : break;
1542 : :
1543 : : /*
1544 : : * An offset past the end of page's line pointer array is possible
1545 : : * when the array was truncated (original item must have been unused)
1546 : : */
1547 [ - + ]: 27109691 : if (offnum > maxoff)
1548 : 0 : break;
1549 : :
1550 : : /* If item is already processed, stop --- it must not be same chain */
1551 [ - + ]: 27109691 : if (prstate->processed[offnum])
1552 : 0 : break;
1553 : :
1554 : 27109691 : lp = PageGetItemId(page, offnum);
1555 : :
1556 : : /*
1557 : : * Unused item obviously isn't part of the chain. Likewise, a dead
1558 : : * line pointer can't be part of the chain. Both of those cases were
1559 : : * already marked as processed.
1560 : : */
1561 : : Assert(ItemIdIsUsed(lp));
1562 : : Assert(!ItemIdIsDead(lp));
1563 : :
1564 : : /*
1565 : : * If we are looking at the redirected root line pointer, jump to the
1566 : : * first normal tuple in the chain. If we find a redirect somewhere
1567 : : * else, stop --- it must not be same chain.
1568 : : */
1569 [ + + ]: 27109691 : if (ItemIdIsRedirected(lp))
1570 : : {
1571 [ - + ]: 197842 : if (nchain > 0)
1572 : 0 : break; /* not at start of chain */
1573 : 197842 : chainitems[nchain++] = offnum;
1574 : 197842 : offnum = ItemIdGetRedirect(rootlp);
1575 : 197842 : continue;
1576 : : }
1577 : :
1578 : : Assert(ItemIdIsNormal(lp));
1579 : :
1580 : 26911849 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1581 : :
1582 : : /*
1583 : : * Check the tuple XMIN against prior XMAX, if any
1584 : : */
1585 [ + + - + ]: 27037267 : if (TransactionIdIsValid(priorXmax) &&
1586 : 125418 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
1587 : 0 : break;
1588 : :
1589 : : /*
1590 : : * OK, this tuple is indeed a member of the chain.
1591 : : */
1592 : 26911849 : chainitems[nchain++] = offnum;
1593 : :
1594 [ + + + - ]: 26911849 : switch (htsv_get_valid_status(prstate->htsv[offnum]))
1595 : : {
1596 : 3960649 : case HEAPTUPLE_DEAD:
1597 : :
1598 : : /* Remember the last DEAD tuple seen */
1599 : 3960649 : ndeadchain = nchain;
1600 : 3960649 : HeapTupleHeaderAdvanceConflictHorizon(htup,
1601 : : &prstate->latest_xid_removed);
1602 : : /* Advance to next chain member */
1603 : 3960649 : break;
1604 : :
1605 : 376419 : case HEAPTUPLE_RECENTLY_DEAD:
1606 : :
1607 : : /*
1608 : : * We don't need to advance the conflict horizon for
1609 : : * RECENTLY_DEAD tuples, even if we are removing them. This
1610 : : * is because we only remove RECENTLY_DEAD tuples if they
1611 : : * precede a DEAD tuple, and the DEAD tuple must have been
1612 : : * inserted by a newer transaction than the RECENTLY_DEAD
1613 : : * tuple by virtue of being later in the chain. We will have
1614 : : * advanced the conflict horizon for the DEAD tuple.
1615 : : */
1616 : :
1617 : : /*
1618 : : * Advance past RECENTLY_DEAD tuples just in case there's a
1619 : : * DEAD one after them. We have to make sure that we don't
1620 : : * miss any DEAD tuples, since DEAD tuples that still have
1621 : : * tuple storage after pruning will confuse VACUUM.
1622 : : */
1623 : 376419 : break;
1624 : :
1625 : 22574781 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1626 : : case HEAPTUPLE_LIVE:
1627 : : case HEAPTUPLE_INSERT_IN_PROGRESS:
1628 : 22574781 : goto process_chain;
1629 : :
1630 : 0 : default:
1631 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1632 : : goto process_chain;
1633 : : }
1634 : :
1635 : : /*
1636 : : * If the tuple is not HOT-updated, then we are at the end of this
1637 : : * HOT-update chain.
1638 : : */
1639 [ + + ]: 4337068 : if (!HeapTupleHeaderIsHotUpdated(htup))
1640 : 4211650 : goto process_chain;
1641 : :
1642 : : /* HOT implies it can't have moved to different partition */
1643 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1644 : :
1645 : : /*
1646 : : * Advance to next chain member.
1647 : : */
1648 : : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == prstate->block);
1649 : 125418 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1650 : 125418 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1651 : : }
1652 : :
1653 [ # # # # ]: 0 : if (ItemIdIsRedirected(rootlp) && nchain < 2)
1654 : : {
1655 : : /*
1656 : : * We found a redirect item that doesn't point to a valid follow-on
1657 : : * item. This can happen if the loop in heap_page_prune_and_freeze()
1658 : : * caused us to visit the dead successor of a redirect item before
1659 : : * visiting the redirect item. We can clean up by setting the
1660 : : * redirect item to LP_DEAD state or LP_UNUSED if the caller
1661 : : * indicated.
1662 : : */
1663 : 0 : heap_prune_record_dead_or_unused(prstate, rootoffnum, false);
1664 : 0 : return;
1665 : : }
1666 : :
1667 : 0 : process_chain:
1668 : :
1669 [ + + ]: 26786431 : if (ndeadchain == 0)
1670 : : {
1671 : : /*
1672 : : * No DEAD tuple was found, so the chain is entirely composed of
1673 : : * normal, unchanged tuples. Leave it alone.
1674 : : */
1675 : 22870813 : int i = 0;
1676 : :
1677 [ + + ]: 22870813 : if (ItemIdIsRedirected(rootlp))
1678 : : {
1679 : 175952 : heap_prune_record_unchanged_lp_redirect(prstate, rootoffnum);
1680 : 175952 : i++;
1681 : : }
1682 [ + + ]: 45745903 : for (; i < nchain; i++)
1683 : 22875090 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1684 : : }
1685 [ + + ]: 3915618 : else if (ndeadchain == nchain)
1686 : : {
1687 : : /*
1688 : : * The entire chain is dead. Mark the root line pointer LP_DEAD, and
1689 : : * fully remove the other tuples in the chain.
1690 : : */
1691 : 3840505 : heap_prune_record_dead_or_unused(prstate, rootoffnum, ItemIdIsNormal(rootlp));
1692 [ + + ]: 3884806 : for (int i = 1; i < nchain; i++)
1693 : 44301 : heap_prune_record_unused(prstate, chainitems[i], true);
1694 : : }
1695 : : else
1696 : : {
1697 : : /*
1698 : : * We found a DEAD tuple in the chain. Redirect the root line pointer
1699 : : * to the first non-DEAD tuple, and mark as unused each intermediate
1700 : : * item that we are able to remove from the chain.
1701 : : */
1702 : 75113 : heap_prune_record_redirect(prstate, rootoffnum, chainitems[ndeadchain],
1703 : 75113 : ItemIdIsNormal(rootlp));
1704 [ + + ]: 97733 : for (int i = 1; i < ndeadchain; i++)
1705 : 22620 : heap_prune_record_unused(prstate, chainitems[i], true);
1706 : :
1707 : : /* the rest of tuples in the chain are normal, unchanged tuples */
1708 [ + + ]: 151223 : for (int i = ndeadchain; i < nchain; i++)
1709 : 76110 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1710 : : }
1711 : : }
1712 : :
1713 : : /* Record lowest soon-prunable XID */
1714 : : static void
1715 : 6118983 : heap_prune_record_prunable(PruneState *prstate, TransactionId xid,
1716 : : OffsetNumber offnum)
1717 : : {
1718 : : /*
1719 : : * This should exactly match the PageSetPrunable macro. We can't store
1720 : : * directly into the page header yet, so we update working state.
1721 : : */
1722 : : Assert(TransactionIdIsNormal(xid));
1723 [ + + + + ]: 11981816 : if (!TransactionIdIsValid(prstate->new_prune_xid) ||
1724 : 5862833 : TransactionIdPrecedes(xid, prstate->new_prune_xid))
1725 : 257551 : prstate->new_prune_xid = xid;
1726 : :
1727 : : /*
1728 : : * It's incorrect for a page to be marked all-visible if it contains
1729 : : * prunable items.
1730 : : */
1731 [ - + ]: 6118983 : if (PageIsAllVisible(prstate->page))
1732 : 0 : heap_page_fix_vm_corruption(prstate, offnum,
1733 : : VM_CORRUPT_TUPLE_VISIBILITY);
1734 : 6118983 : }
1735 : :
1736 : : /* Record line pointer to be redirected */
1737 : : static void
1738 : 75113 : heap_prune_record_redirect(PruneState *prstate,
1739 : : OffsetNumber offnum, OffsetNumber rdoffnum,
1740 : : bool was_normal)
1741 : : {
1742 : : Assert(!prstate->processed[offnum]);
1743 : 75113 : prstate->processed[offnum] = true;
1744 : :
1745 : : /*
1746 : : * Do not mark the redirect target here. It needs to be counted
1747 : : * separately as an unchanged tuple.
1748 : : */
1749 : :
1750 : : Assert(prstate->nredirected < MaxHeapTuplesPerPage);
1751 : 75113 : prstate->redirected[prstate->nredirected * 2] = offnum;
1752 : 75113 : prstate->redirected[prstate->nredirected * 2 + 1] = rdoffnum;
1753 : :
1754 : 75113 : prstate->nredirected++;
1755 : :
1756 : : /*
1757 : : * If the root entry had been a normal tuple, we are deleting it, so count
1758 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1759 : : * count.
1760 : : */
1761 [ + + ]: 75113 : if (was_normal)
1762 : 66131 : prstate->ndeleted++;
1763 : :
1764 : 75113 : prstate->hastup = true;
1765 : 75113 : }
1766 : :
1767 : : /* Record line pointer to be marked dead */
1768 : : static void
1769 : 3805471 : heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
1770 : : bool was_normal)
1771 : : {
1772 : : Assert(!prstate->processed[offnum]);
1773 : 3805471 : prstate->processed[offnum] = true;
1774 : :
1775 : : Assert(prstate->ndead < MaxHeapTuplesPerPage);
1776 : 3805471 : prstate->nowdead[prstate->ndead] = offnum;
1777 : 3805471 : prstate->ndead++;
1778 : :
1779 : : /*
1780 : : * Deliberately delay unsetting set_all_visible and set_all_frozen until
1781 : : * later during pruning. Removable dead tuples shouldn't preclude freezing
1782 : : * the page.
1783 : : */
1784 : :
1785 : : /* Record the dead offset for vacuum */
1786 : 3805471 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1787 : :
1788 : : /*
1789 : : * If the root entry had been a normal tuple, we are deleting it, so count
1790 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1791 : : * count.
1792 : : */
1793 [ + + ]: 3805471 : if (was_normal)
1794 : 3792563 : prstate->ndeleted++;
1795 : 3805471 : }
1796 : :
1797 : : /*
1798 : : * Depending on whether or not the caller set mark_unused_now to true, record that a
1799 : : * line pointer should be marked LP_DEAD or LP_UNUSED. There are other cases in
1800 : : * which we will mark line pointers LP_UNUSED, but we will not mark line
1801 : : * pointers LP_DEAD if mark_unused_now is true.
1802 : : */
1803 : : static void
1804 : 3840505 : heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
1805 : : bool was_normal)
1806 : : {
1807 : : /*
1808 : : * If the caller set mark_unused_now to true, we can remove dead tuples
1809 : : * during pruning instead of marking their line pointers dead. Set this
1810 : : * tuple's line pointer LP_UNUSED. We hint that this option is less
1811 : : * likely.
1812 : : */
1813 [ + + ]: 3840505 : if (unlikely(prstate->mark_unused_now))
1814 : 35034 : heap_prune_record_unused(prstate, offnum, was_normal);
1815 : : else
1816 : 3805471 : heap_prune_record_dead(prstate, offnum, was_normal);
1817 : :
1818 : : /*
1819 : : * It's incorrect for the page to be set all-visible if it contains dead
1820 : : * items. Fix that on the heap page and check the VM for corruption as
1821 : : * well. Do that here rather than in heap_prune_record_dead() so we also
1822 : : * cover tuples that are directly marked LP_UNUSED via mark_unused_now.
1823 : : */
1824 [ - + ]: 3840505 : if (PageIsAllVisible(prstate->page))
1825 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
1826 : 3840505 : }
1827 : :
1828 : : /* Record line pointer to be marked unused */
1829 : : static void
1830 : 107075 : heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal)
1831 : : {
1832 : : Assert(!prstate->processed[offnum]);
1833 : 107075 : prstate->processed[offnum] = true;
1834 : :
1835 : : Assert(prstate->nunused < MaxHeapTuplesPerPage);
1836 : 107075 : prstate->nowunused[prstate->nunused] = offnum;
1837 : 107075 : prstate->nunused++;
1838 : :
1839 : : /*
1840 : : * If the root entry had been a normal tuple, we are deleting it, so count
1841 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1842 : : * count.
1843 : : */
1844 [ + + ]: 107075 : if (was_normal)
1845 : 105223 : prstate->ndeleted++;
1846 : 107075 : }
1847 : :
1848 : : /*
1849 : : * Record an unused line pointer that is left unchanged.
1850 : : */
1851 : : static void
1852 : 205084 : heap_prune_record_unchanged_lp_unused(PruneState *prstate, OffsetNumber offnum)
1853 : : {
1854 : : Assert(!prstate->processed[offnum]);
1855 : 205084 : prstate->processed[offnum] = true;
1856 : 205084 : }
1857 : :
1858 : : /*
1859 : : * Record line pointer that is left unchanged. We consider freezing it, and
1860 : : * update bookkeeping of tuple counts and page visibility.
1861 : : */
1862 : : static void
1863 : 22964598 : heap_prune_record_unchanged_lp_normal(PruneState *prstate, OffsetNumber offnum)
1864 : : {
1865 : : HeapTupleHeader htup;
1866 : : TransactionId xmin;
1867 : 22964598 : Page page = prstate->page;
1868 : :
1869 : : Assert(!prstate->processed[offnum]);
1870 : 22964598 : prstate->processed[offnum] = true;
1871 : :
1872 : 22964598 : prstate->hastup = true; /* the page is not empty */
1873 : :
1874 : : /*
1875 : : * The criteria for counting a tuple as live in this block need to match
1876 : : * what analyze.c's acquire_sample_rows() does, otherwise VACUUM and
1877 : : * ANALYZE may produce wildly different reltuples values, e.g. when there
1878 : : * are many recently-dead tuples.
1879 : : *
1880 : : * The logic here is a bit simpler than acquire_sample_rows(), as VACUUM
1881 : : * can't run inside a transaction block, which makes some cases impossible
1882 : : * (e.g. in-progress insert from the same transaction).
1883 : : *
1884 : : * HEAPTUPLE_DEAD are handled by the other heap_prune_record_*()
1885 : : * subroutines. They don't count dead items like acquire_sample_rows()
1886 : : * does, because we assume that all dead items will become LP_UNUSED
1887 : : * before VACUUM finishes. This difference is only superficial. VACUUM
1888 : : * effectively agrees with ANALYZE about DEAD items, in the end. VACUUM
1889 : : * won't remember LP_DEAD items, but only because they're not supposed to
1890 : : * be left behind when it is done. (Cases where we bypass index vacuuming
1891 : : * will violate this optimistic assumption, but the overall impact of that
1892 : : * should be negligible.)
1893 : : */
1894 : 22964598 : htup = (HeapTupleHeader) PageGetItem(page, PageGetItemId(page, offnum));
1895 : :
1896 [ + + + + : 22964598 : switch (prstate->htsv[offnum])
- ]
1897 : : {
1898 : 16845615 : case HEAPTUPLE_LIVE:
1899 : :
1900 : : /*
1901 : : * Count it as live. Not only is this natural, but it's also what
1902 : : * acquire_sample_rows() does.
1903 : : */
1904 : 16845615 : prstate->live_tuples++;
1905 : :
1906 : : /*
1907 : : * Is the tuple definitely visible to all transactions?
1908 : : *
1909 : : * NB: Like with per-tuple hint bits, we can't set the
1910 : : * PD_ALL_VISIBLE flag if the inserter committed asynchronously.
1911 : : * See SetHintBits for more info. Check that the tuple is hinted
1912 : : * xmin-committed because of that.
1913 : : */
1914 [ + + ]: 16845615 : if (!HeapTupleHeaderXminCommitted(htup))
1915 : : {
1916 : 44030 : prstate->set_all_visible = false;
1917 : 44030 : prstate->set_all_frozen = false;
1918 : 44030 : break;
1919 : : }
1920 : :
1921 : : /*
1922 : : * The inserter definitely committed. But we don't know if it is
1923 : : * old enough that everyone sees it as committed. Later, after
1924 : : * processing all the tuples on the page, we'll check if there is
1925 : : * any snapshot that still considers the newest xid on the page to
1926 : : * be running. If so, we don't consider the page all-visible.
1927 : : */
1928 : 16801585 : xmin = HeapTupleHeaderGetXmin(htup);
1929 : :
1930 : : /* Track newest xmin on page. */
1931 [ + + + + ]: 16801585 : if (TransactionIdFollows(xmin, prstate->newest_live_xid) &&
1932 : : TransactionIdIsNormal(xmin))
1933 : 615217 : prstate->newest_live_xid = xmin;
1934 : :
1935 : 16801585 : break;
1936 : :
1937 : 376419 : case HEAPTUPLE_RECENTLY_DEAD:
1938 : 376419 : prstate->recently_dead_tuples++;
1939 : 376419 : prstate->set_all_visible = false;
1940 : 376419 : prstate->set_all_frozen = false;
1941 : :
1942 : : /*
1943 : : * This tuple will soon become DEAD. Update the hint field so
1944 : : * that the page is reconsidered for pruning in future.
1945 : : */
1946 : 376419 : heap_prune_record_prunable(prstate,
1947 : : HeapTupleHeaderGetUpdateXid(htup),
1948 : : offnum);
1949 : 376419 : break;
1950 : :
1951 : 137385 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1952 : :
1953 : : /*
1954 : : * We do not count these rows as live, because we expect the
1955 : : * inserting transaction to update the counters at commit, and we
1956 : : * assume that will happen only after we report our results. This
1957 : : * assumption is a bit shaky, but it is what acquire_sample_rows()
1958 : : * does, so be consistent.
1959 : : */
1960 : 137385 : prstate->set_all_visible = false;
1961 : 137385 : prstate->set_all_frozen = false;
1962 : :
1963 : : /*
1964 : : * Though there is nothing "prunable" on the page, we maintain
1965 : : * pd_prune_xid for inserts so that we have the opportunity to
1966 : : * mark them all-visible during the next round of pruning.
1967 : : */
1968 : 137385 : heap_prune_record_prunable(prstate,
1969 : : HeapTupleHeaderGetXmin(htup),
1970 : : offnum);
1971 : 137385 : break;
1972 : :
1973 : 5605179 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1974 : :
1975 : : /*
1976 : : * This an expected case during concurrent vacuum. Count such
1977 : : * rows as live. As above, we assume the deleting transaction
1978 : : * will commit and update the counters after we report.
1979 : : */
1980 : 5605179 : prstate->live_tuples++;
1981 : 5605179 : prstate->set_all_visible = false;
1982 : 5605179 : prstate->set_all_frozen = false;
1983 : :
1984 : : /*
1985 : : * This tuple may soon become DEAD. Update the hint field so that
1986 : : * the page is reconsidered for pruning in future.
1987 : : */
1988 : 5605179 : heap_prune_record_prunable(prstate,
1989 : : HeapTupleHeaderGetUpdateXid(htup),
1990 : : offnum);
1991 : 5605179 : break;
1992 : :
1993 : 0 : default:
1994 : :
1995 : : /*
1996 : : * DEAD tuples should've been passed to heap_prune_record_dead()
1997 : : * or heap_prune_record_unused() instead.
1998 : : */
1999 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result %d",
2000 : : prstate->htsv[offnum]);
2001 : : break;
2002 : : }
2003 : :
2004 : : /* Consider freezing any normal tuples which will not be removed */
2005 [ + + ]: 22964598 : if (prstate->attempt_freeze)
2006 : : {
2007 : : bool totally_frozen;
2008 : :
2009 [ + + ]: 12686798 : if ((heap_prepare_freeze_tuple(htup,
2010 : 12686798 : prstate->cutoffs,
2011 : : &prstate->pagefrz,
2012 : 12686798 : &prstate->frozen[prstate->nfrozen],
2013 : : &totally_frozen)))
2014 : : {
2015 : : /* Save prepared freeze plan for later */
2016 : 3352623 : prstate->frozen[prstate->nfrozen++].offset = offnum;
2017 : : }
2018 : :
2019 : : /*
2020 : : * If any tuple isn't either totally frozen already or eligible to
2021 : : * become totally frozen (according to its freeze plan), then the page
2022 : : * definitely cannot be set all-frozen in the visibility map later on.
2023 : : */
2024 [ + + ]: 12686798 : if (!totally_frozen)
2025 : 6349888 : prstate->set_all_frozen = false;
2026 : : }
2027 : 22964598 : }
2028 : :
2029 : :
2030 : : /*
2031 : : * Record line pointer that was already LP_DEAD and is left unchanged.
2032 : : */
2033 : : static void
2034 : 1629058 : heap_prune_record_unchanged_lp_dead(PruneState *prstate, OffsetNumber offnum)
2035 : : {
2036 : : Assert(!prstate->processed[offnum]);
2037 : 1629058 : prstate->processed[offnum] = true;
2038 : :
2039 : : /*
2040 : : * Deliberately don't set hastup for LP_DEAD items. We make the soft
2041 : : * assumption that any LP_DEAD items encountered here will become
2042 : : * LP_UNUSED later on, before count_nondeletable_pages is reached. If we
2043 : : * don't make this assumption then rel truncation will only happen every
2044 : : * other VACUUM, at most. Besides, VACUUM must treat
2045 : : * hastup/nonempty_pages as provisional no matter how LP_DEAD items are
2046 : : * handled (handled here, or handled later on).
2047 : : *
2048 : : * Similarly, don't unset set_all_visible and set_all_frozen until later,
2049 : : * at the end of heap_page_prune_and_freeze(). This will allow us to
2050 : : * attempt to freeze the page after pruning. As long as we unset it
2051 : : * before updating the visibility map, this will be correct.
2052 : : */
2053 : :
2054 : : /* Record the dead offset for vacuum */
2055 : 1629058 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
2056 : :
2057 : : /*
2058 : : * It's incorrect for a page to be marked all-visible if it contains dead
2059 : : * items.
2060 : : */
2061 [ - + ]: 1629058 : if (PageIsAllVisible(prstate->page))
2062 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
2063 : 1629058 : }
2064 : :
2065 : : /*
2066 : : * Record LP_REDIRECT that is left unchanged.
2067 : : */
2068 : : static void
2069 : 175952 : heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum)
2070 : : {
2071 : : /*
2072 : : * A redirect line pointer doesn't count as a live tuple.
2073 : : *
2074 : : * If we leave a redirect line pointer in place, there will be another
2075 : : * tuple on the page that it points to. We will do the bookkeeping for
2076 : : * that separately. So we have nothing to do here, except remember that
2077 : : * we processed this item.
2078 : : */
2079 : : Assert(!prstate->processed[offnum]);
2080 : 175952 : prstate->processed[offnum] = true;
2081 : 175952 : }
2082 : :
2083 : : /*
2084 : : * Perform the actual page changes needed by heap_page_prune_and_freeze().
2085 : : *
2086 : : * If 'lp_truncate_only' is set, we are merely marking LP_DEAD line pointers
2087 : : * as unused, not redirecting or removing anything else. The
2088 : : * PageRepairFragmentation() call is skipped in that case.
2089 : : *
2090 : : * If 'lp_truncate_only' is not set, the caller must hold a cleanup lock on
2091 : : * the buffer. If it is set, an ordinary exclusive lock suffices.
2092 : : */
2093 : : void
2094 : 81354 : heap_page_prune_execute(Buffer buffer, bool lp_truncate_only,
2095 : : OffsetNumber *redirected, int nredirected,
2096 : : OffsetNumber *nowdead, int ndead,
2097 : : OffsetNumber *nowunused, int nunused)
2098 : : {
2099 : 81354 : Page page = BufferGetPage(buffer);
2100 : : OffsetNumber *offnum;
2101 : : HeapTupleHeader htup PG_USED_FOR_ASSERTS_ONLY;
2102 : :
2103 : : /* Shouldn't be called unless there's something to do */
2104 : : Assert(nredirected > 0 || ndead > 0 || nunused > 0);
2105 : :
2106 : : /* If 'lp_truncate_only', we can only remove already-dead line pointers */
2107 : : Assert(!lp_truncate_only || (nredirected == 0 && ndead == 0));
2108 : :
2109 : : /* Update all redirected line pointers */
2110 : 81354 : offnum = redirected;
2111 [ + + ]: 176564 : for (int i = 0; i < nredirected; i++)
2112 : : {
2113 : 95210 : OffsetNumber fromoff = *offnum++;
2114 : 95210 : OffsetNumber tooff = *offnum++;
2115 : 95210 : ItemId fromlp = PageGetItemId(page, fromoff);
2116 : : ItemId tolp PG_USED_FOR_ASSERTS_ONLY;
2117 : :
2118 : : #ifdef USE_ASSERT_CHECKING
2119 : :
2120 : : /*
2121 : : * Any existing item that we set as an LP_REDIRECT (any 'from' item)
2122 : : * must be the first item from a HOT chain. If the item has tuple
2123 : : * storage then it can't be a heap-only tuple. Otherwise we are just
2124 : : * maintaining an existing LP_REDIRECT from an existing HOT chain that
2125 : : * has been pruned at least once before now.
2126 : : */
2127 : : if (!ItemIdIsRedirected(fromlp))
2128 : : {
2129 : : Assert(ItemIdHasStorage(fromlp) && ItemIdIsNormal(fromlp));
2130 : :
2131 : : htup = (HeapTupleHeader) PageGetItem(page, fromlp);
2132 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2133 : : }
2134 : : else
2135 : : {
2136 : : /* We shouldn't need to redundantly set the redirect */
2137 : : Assert(ItemIdGetRedirect(fromlp) != tooff);
2138 : : }
2139 : :
2140 : : /*
2141 : : * The item that we're about to set as an LP_REDIRECT (the 'from'
2142 : : * item) will point to an existing item (the 'to' item) that is
2143 : : * already a heap-only tuple. There can be at most one LP_REDIRECT
2144 : : * item per HOT chain.
2145 : : *
2146 : : * We need to keep around an LP_REDIRECT item (after original
2147 : : * non-heap-only root tuple gets pruned away) so that it's always
2148 : : * possible for VACUUM to easily figure out what TID to delete from
2149 : : * indexes when an entire HOT chain becomes dead. A heap-only tuple
2150 : : * can never become LP_DEAD; an LP_REDIRECT item or a regular heap
2151 : : * tuple can.
2152 : : *
2153 : : * This check may miss problems, e.g. the target of a redirect could
2154 : : * be marked as unused subsequently. The page_verify_redirects() check
2155 : : * below will catch such problems.
2156 : : */
2157 : : tolp = PageGetItemId(page, tooff);
2158 : : Assert(ItemIdHasStorage(tolp) && ItemIdIsNormal(tolp));
2159 : : htup = (HeapTupleHeader) PageGetItem(page, tolp);
2160 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2161 : : #endif
2162 : :
2163 : 95210 : ItemIdSetRedirect(fromlp, tooff);
2164 : : }
2165 : :
2166 : : /* Update all now-dead line pointers */
2167 : 81354 : offnum = nowdead;
2168 [ + + ]: 4163176 : for (int i = 0; i < ndead; i++)
2169 : : {
2170 : 4081822 : OffsetNumber off = *offnum++;
2171 : 4081822 : ItemId lp = PageGetItemId(page, off);
2172 : :
2173 : : #ifdef USE_ASSERT_CHECKING
2174 : :
2175 : : /*
2176 : : * An LP_DEAD line pointer must be left behind when the original item
2177 : : * (which is dead to everybody) could still be referenced by a TID in
2178 : : * an index. This should never be necessary with any individual
2179 : : * heap-only tuple item, though. (It's not clear how much of a problem
2180 : : * that would be, but there is no reason to allow it.)
2181 : : */
2182 : : if (ItemIdHasStorage(lp))
2183 : : {
2184 : : Assert(ItemIdIsNormal(lp));
2185 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2186 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2187 : : }
2188 : : else
2189 : : {
2190 : : /* Whole HOT chain becomes dead */
2191 : : Assert(ItemIdIsRedirected(lp));
2192 : : }
2193 : : #endif
2194 : :
2195 : 4081822 : ItemIdSetDead(lp);
2196 : : }
2197 : :
2198 : : /* Update all now-unused line pointers */
2199 : 81354 : offnum = nowunused;
2200 [ + + ]: 327947 : for (int i = 0; i < nunused; i++)
2201 : : {
2202 : 246593 : OffsetNumber off = *offnum++;
2203 : 246593 : ItemId lp = PageGetItemId(page, off);
2204 : :
2205 : : #ifdef USE_ASSERT_CHECKING
2206 : :
2207 : : if (lp_truncate_only)
2208 : : {
2209 : : /* Setting LP_DEAD to LP_UNUSED in vacuum's second pass */
2210 : : Assert(ItemIdIsDead(lp) && !ItemIdHasStorage(lp));
2211 : : }
2212 : : else
2213 : : {
2214 : : /*
2215 : : * When heap_page_prune_and_freeze() was called, mark_unused_now
2216 : : * may have been passed as true, which allows would-be LP_DEAD
2217 : : * items to be made LP_UNUSED instead. This is only possible if
2218 : : * the relation has no indexes. If there are any dead items, then
2219 : : * mark_unused_now was not true and every item being marked
2220 : : * LP_UNUSED must refer to a heap-only tuple.
2221 : : */
2222 : : if (ndead > 0)
2223 : : {
2224 : : Assert(ItemIdHasStorage(lp) && ItemIdIsNormal(lp));
2225 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2226 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2227 : : }
2228 : : else
2229 : : Assert(ItemIdIsUsed(lp));
2230 : : }
2231 : :
2232 : : #endif
2233 : :
2234 : 246593 : ItemIdSetUnused(lp);
2235 : : }
2236 : :
2237 [ + + ]: 81354 : if (lp_truncate_only)
2238 : 1686 : PageTruncateLinePointerArray(page);
2239 : : else
2240 : : {
2241 : : /*
2242 : : * Finally, repair any fragmentation, and update the page's hint bit
2243 : : * about whether it has free pointers.
2244 : : */
2245 : 79668 : PageRepairFragmentation(page);
2246 : :
2247 : : /*
2248 : : * Now that the page has been modified, assert that redirect items
2249 : : * still point to valid targets.
2250 : : */
2251 : 79668 : page_verify_redirects(page);
2252 : : }
2253 : 81354 : }
2254 : :
2255 : :
2256 : : /*
2257 : : * If built with assertions, verify that all LP_REDIRECT items point to a
2258 : : * valid item.
2259 : : *
2260 : : * One way that bugs related to HOT pruning show is redirect items pointing to
2261 : : * removed tuples. It's not trivial to reliably check that marking an item
2262 : : * unused will not orphan a redirect item during heap_prune_chain() /
2263 : : * heap_page_prune_execute(), so we additionally check the whole page after
2264 : : * pruning. Without this check such bugs would typically only cause asserts
2265 : : * later, potentially well after the corruption has been introduced.
2266 : : *
2267 : : * Also check comments in heap_page_prune_execute()'s redirection loop.
2268 : : */
2269 : : static void
2270 : 79668 : page_verify_redirects(Page page)
2271 : : {
2272 : : #ifdef USE_ASSERT_CHECKING
2273 : : OffsetNumber offnum;
2274 : : OffsetNumber maxoff;
2275 : :
2276 : : maxoff = PageGetMaxOffsetNumber(page);
2277 : : for (offnum = FirstOffsetNumber;
2278 : : offnum <= maxoff;
2279 : : offnum = OffsetNumberNext(offnum))
2280 : : {
2281 : : ItemId itemid = PageGetItemId(page, offnum);
2282 : : OffsetNumber targoff;
2283 : : ItemId targitem;
2284 : : HeapTupleHeader htup;
2285 : :
2286 : : if (!ItemIdIsRedirected(itemid))
2287 : : continue;
2288 : :
2289 : : targoff = ItemIdGetRedirect(itemid);
2290 : : targitem = PageGetItemId(page, targoff);
2291 : :
2292 : : Assert(ItemIdIsUsed(targitem));
2293 : : Assert(ItemIdIsNormal(targitem));
2294 : : Assert(ItemIdHasStorage(targitem));
2295 : : htup = (HeapTupleHeader) PageGetItem(page, targitem);
2296 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2297 : : }
2298 : : #endif
2299 : 79668 : }
2300 : :
2301 : :
2302 : : /*
2303 : : * For all items in this page, find their respective root line pointers.
2304 : : * If item k is part of a HOT-chain with root at item j, then we set
2305 : : * root_offsets[k - 1] = j.
2306 : : *
2307 : : * The passed-in root_offsets array must have MaxHeapTuplesPerPage entries.
2308 : : * Unused entries are filled with InvalidOffsetNumber (zero).
2309 : : *
2310 : : * The function must be called with at least share lock on the buffer, to
2311 : : * prevent concurrent prune operations.
2312 : : *
2313 : : * Note: The information collected here is valid only as long as the caller
2314 : : * holds a pin on the buffer. Once pin is released, a tuple might be pruned
2315 : : * and reused by a completely unrelated tuple.
2316 : : */
2317 : : void
2318 : 140230 : heap_get_root_tuples(Page page, OffsetNumber *root_offsets)
2319 : : {
2320 : : OffsetNumber offnum,
2321 : : maxoff;
2322 : :
2323 [ + - - + : 140230 : MemSet(root_offsets, InvalidOffsetNumber,
- - - - -
- ]
2324 : : MaxHeapTuplesPerPage * sizeof(OffsetNumber));
2325 : :
2326 : 140230 : maxoff = PageGetMaxOffsetNumber(page);
2327 [ + + ]: 11764004 : for (offnum = FirstOffsetNumber; offnum <= maxoff; offnum = OffsetNumberNext(offnum))
2328 : : {
2329 : 11623774 : ItemId lp = PageGetItemId(page, offnum);
2330 : : HeapTupleHeader htup;
2331 : : OffsetNumber nextoffnum;
2332 : : TransactionId priorXmax;
2333 : :
2334 : : /* skip unused and dead items */
2335 [ + + + + ]: 11623774 : if (!ItemIdIsUsed(lp) || ItemIdIsDead(lp))
2336 : 11541 : continue;
2337 : :
2338 [ + + ]: 11612233 : if (ItemIdIsNormal(lp))
2339 : : {
2340 : 11608972 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2341 : :
2342 : : /*
2343 : : * Check if this tuple is part of a HOT-chain rooted at some other
2344 : : * tuple. If so, skip it for now; we'll process it when we find
2345 : : * its root.
2346 : : */
2347 [ + + ]: 11608972 : if (HeapTupleHeaderIsHeapOnly(htup))
2348 : 3577 : continue;
2349 : :
2350 : : /*
2351 : : * This is either a plain tuple or the root of a HOT-chain.
2352 : : * Remember it in the mapping.
2353 : : */
2354 : 11605395 : root_offsets[offnum - 1] = offnum;
2355 : :
2356 : : /* If it's not the start of a HOT-chain, we're done with it */
2357 [ + + ]: 11605395 : if (!HeapTupleHeaderIsHotUpdated(htup))
2358 : 11605143 : continue;
2359 : :
2360 : : /* Set up to scan the HOT-chain */
2361 : 252 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2362 : 252 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2363 : : }
2364 : : else
2365 : : {
2366 : : /* Must be a redirect item. We do not set its root_offsets entry */
2367 : : Assert(ItemIdIsRedirected(lp));
2368 : : /* Set up to scan the HOT-chain */
2369 : 3261 : nextoffnum = ItemIdGetRedirect(lp);
2370 : 3261 : priorXmax = InvalidTransactionId;
2371 : : }
2372 : :
2373 : : /*
2374 : : * Now follow the HOT-chain and collect other tuples in the chain.
2375 : : *
2376 : : * Note: Even though this is a nested loop, the complexity of the
2377 : : * function is O(N) because a tuple in the page should be visited not
2378 : : * more than twice, once in the outer loop and once in HOT-chain
2379 : : * chases.
2380 : : */
2381 : : for (;;)
2382 : : {
2383 : : /* Sanity check (pure paranoia) */
2384 [ - + ]: 3573 : if (nextoffnum < FirstOffsetNumber)
2385 : 0 : break;
2386 : :
2387 : : /*
2388 : : * An offset past the end of page's line pointer array is possible
2389 : : * when the array was truncated
2390 : : */
2391 [ - + ]: 3573 : if (nextoffnum > maxoff)
2392 : 0 : break;
2393 : :
2394 : 3573 : lp = PageGetItemId(page, nextoffnum);
2395 : :
2396 : : /* Check for broken chains */
2397 [ - + ]: 3573 : if (!ItemIdIsNormal(lp))
2398 : 0 : break;
2399 : :
2400 : 3573 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2401 : :
2402 [ + + - + ]: 3885 : if (TransactionIdIsValid(priorXmax) &&
2403 : 312 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(htup)))
2404 : 0 : break;
2405 : :
2406 : : /* Remember the root line pointer for this item */
2407 : 3573 : root_offsets[nextoffnum - 1] = offnum;
2408 : :
2409 : : /* Advance to next chain member, if any */
2410 [ + + ]: 3573 : if (!HeapTupleHeaderIsHotUpdated(htup))
2411 : 3513 : break;
2412 : :
2413 : : /* HOT implies it can't have moved to different partition */
2414 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
2415 : :
2416 : 60 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2417 : 60 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2418 : : }
2419 : : }
2420 : 140230 : }
2421 : :
2422 : :
2423 : : /*
2424 : : * Compare fields that describe actions required to freeze tuple with caller's
2425 : : * open plan. If everything matches then the frz tuple plan is equivalent to
2426 : : * caller's plan.
2427 : : */
2428 : : static inline bool
2429 : 1327181 : heap_log_freeze_eq(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2430 : : {
2431 [ + + ]: 1327181 : if (plan->xmax == frz->xmax &&
2432 [ + + ]: 1325885 : plan->t_infomask2 == frz->t_infomask2 &&
2433 [ + + ]: 1324846 : plan->t_infomask == frz->t_infomask &&
2434 [ + - ]: 1321535 : plan->frzflags == frz->frzflags)
2435 : 1321535 : return true;
2436 : :
2437 : : /* Caller must call heap_log_freeze_new_plan again for frz */
2438 : 5646 : return false;
2439 : : }
2440 : :
2441 : : /*
2442 : : * Comparator used to deduplicate the freeze plans used in WAL records.
2443 : : */
2444 : : static int
2445 : 1783566 : heap_log_freeze_cmp(const void *arg1, const void *arg2)
2446 : : {
2447 : 1783566 : const HeapTupleFreeze *frz1 = arg1;
2448 : 1783566 : const HeapTupleFreeze *frz2 = arg2;
2449 : :
2450 [ + + ]: 1783566 : if (frz1->xmax < frz2->xmax)
2451 : 13175 : return -1;
2452 [ + + ]: 1770391 : else if (frz1->xmax > frz2->xmax)
2453 : 14158 : return 1;
2454 : :
2455 [ + + ]: 1756233 : if (frz1->t_infomask2 < frz2->t_infomask2)
2456 : 6117 : return -1;
2457 [ + + ]: 1750116 : else if (frz1->t_infomask2 > frz2->t_infomask2)
2458 : 6126 : return 1;
2459 : :
2460 [ + + ]: 1743990 : if (frz1->t_infomask < frz2->t_infomask)
2461 : 12767 : return -1;
2462 [ + + ]: 1731223 : else if (frz1->t_infomask > frz2->t_infomask)
2463 : 22493 : return 1;
2464 : :
2465 [ - + ]: 1708730 : if (frz1->frzflags < frz2->frzflags)
2466 : 0 : return -1;
2467 [ - + ]: 1708730 : else if (frz1->frzflags > frz2->frzflags)
2468 : 0 : return 1;
2469 : :
2470 : : /*
2471 : : * heap_log_freeze_eq would consider these tuple-wise plans to be equal.
2472 : : * (So the tuples will share a single canonical freeze plan.)
2473 : : *
2474 : : * We tiebreak on page offset number to keep each freeze plan's page
2475 : : * offset number array individually sorted. (Unnecessary, but be tidy.)
2476 : : */
2477 [ + + ]: 1708730 : if (frz1->offset < frz2->offset)
2478 : 1469860 : return -1;
2479 [ + - ]: 238870 : else if (frz1->offset > frz2->offset)
2480 : 238870 : return 1;
2481 : :
2482 : : Assert(false);
2483 : 0 : return 0;
2484 : : }
2485 : :
2486 : : /*
2487 : : * Start new plan initialized using tuple-level actions. At least one tuple
2488 : : * will have steps required to freeze described by caller's plan during REDO.
2489 : : */
2490 : : static inline void
2491 : 32987 : heap_log_freeze_new_plan(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2492 : : {
2493 : 32987 : plan->xmax = frz->xmax;
2494 : 32987 : plan->t_infomask2 = frz->t_infomask2;
2495 : 32987 : plan->t_infomask = frz->t_infomask;
2496 : 32987 : plan->frzflags = frz->frzflags;
2497 : 32987 : plan->ntuples = 1; /* for now */
2498 : 32987 : }
2499 : :
2500 : : /*
2501 : : * Deduplicate tuple-based freeze plans so that each distinct set of
2502 : : * processing steps is only stored once in the WAL record.
2503 : : * Called during original execution of freezing (for logged relations).
2504 : : *
2505 : : * Return value is number of plans set in *plans_out for caller. Also writes
2506 : : * an array of offset numbers into *offsets_out output argument for caller
2507 : : * (actually there is one array per freeze plan, but that's not of immediate
2508 : : * concern to our caller).
2509 : : */
2510 : : static int
2511 : 27341 : heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
2512 : : xlhp_freeze_plan *plans_out,
2513 : : OffsetNumber *offsets_out)
2514 : : {
2515 : 27341 : int nplans = 0;
2516 : :
2517 : : /* Sort tuple-based freeze plans in the order required to deduplicate */
2518 : 27341 : qsort(tuples, ntuples, sizeof(HeapTupleFreeze), heap_log_freeze_cmp);
2519 : :
2520 [ + + ]: 1381863 : for (int i = 0; i < ntuples; i++)
2521 : : {
2522 : 1354522 : HeapTupleFreeze *frz = tuples + i;
2523 : :
2524 [ + + ]: 1354522 : if (i == 0)
2525 : : {
2526 : : /* New canonical freeze plan starting with first tup */
2527 : 27341 : heap_log_freeze_new_plan(plans_out, frz);
2528 : 27341 : nplans++;
2529 : : }
2530 [ + + ]: 1327181 : else if (heap_log_freeze_eq(plans_out, frz))
2531 : : {
2532 : : /* tup matches open canonical plan -- include tup in it */
2533 : : Assert(offsets_out[i - 1] < frz->offset);
2534 : 1321535 : plans_out->ntuples++;
2535 : : }
2536 : : else
2537 : : {
2538 : : /* Tup doesn't match current plan -- done with it now */
2539 : 5646 : plans_out++;
2540 : :
2541 : : /* New canonical freeze plan starting with this tup */
2542 : 5646 : heap_log_freeze_new_plan(plans_out, frz);
2543 : 5646 : nplans++;
2544 : : }
2545 : :
2546 : : /*
2547 : : * Save page offset number in dedicated buffer in passing.
2548 : : *
2549 : : * REDO routine relies on the record's offset numbers array grouping
2550 : : * offset numbers by freeze plan. The sort order within each grouping
2551 : : * is ascending offset number order, just to keep things tidy.
2552 : : */
2553 : 1354522 : offsets_out[i] = frz->offset;
2554 : : }
2555 : :
2556 : : Assert(nplans > 0 && nplans <= ntuples);
2557 : :
2558 : 27341 : return nplans;
2559 : : }
2560 : :
2561 : : /*
2562 : : * Write an XLOG_HEAP2_PRUNE* WAL record
2563 : : *
2564 : : * This is used for several different page maintenance operations:
2565 : : *
2566 : : * - Page pruning, in VACUUM's 1st pass or on access: Some items are
2567 : : * redirected, some marked dead, and some removed altogether.
2568 : : *
2569 : : * - Freezing: Items are marked as 'frozen'.
2570 : : *
2571 : : * - Vacuum, 2nd pass: Items that are already LP_DEAD are marked as unused.
2572 : : *
2573 : : * They have enough commonalities that we use a single WAL record for them
2574 : : * all.
2575 : : *
2576 : : * If replaying the record requires a cleanup lock, pass cleanup_lock = true.
2577 : : * Replaying 'redirected' or 'dead' items always requires a cleanup lock, but
2578 : : * replaying 'unused' items depends on whether they were all previously marked
2579 : : * as dead.
2580 : : *
2581 : : * If the VM is being updated, vmflags will contain the bits to set. In this
2582 : : * case, vmbuffer should already have been updated and marked dirty and should
2583 : : * still be pinned and locked.
2584 : : *
2585 : : * Note: This function scribbles on the 'frozen' array.
2586 : : *
2587 : : * Note: This is called in a critical section, so careful what you do here.
2588 : : */
2589 : : void
2590 : 149893 : log_heap_prune_and_freeze(Relation relation, Buffer buffer,
2591 : : Buffer vmbuffer, uint8 vmflags,
2592 : : TransactionId conflict_xid,
2593 : : bool cleanup_lock,
2594 : : PruneReason reason,
2595 : : HeapTupleFreeze *frozen, int nfrozen,
2596 : : OffsetNumber *redirected, int nredirected,
2597 : : OffsetNumber *dead, int ndead,
2598 : : OffsetNumber *unused, int nunused)
2599 : : {
2600 : : xl_heap_prune xlrec;
2601 : : XLogRecPtr recptr;
2602 : : uint8 info;
2603 : : uint8 regbuf_flags_heap;
2604 : :
2605 : 149893 : Page heap_page = BufferGetPage(buffer);
2606 : :
2607 : : /* The following local variables hold data registered in the WAL record: */
2608 : : xlhp_freeze_plan plans[MaxHeapTuplesPerPage];
2609 : : xlhp_freeze_plans freeze_plans;
2610 : : xlhp_prune_items redirect_items;
2611 : : xlhp_prune_items dead_items;
2612 : : xlhp_prune_items unused_items;
2613 : : OffsetNumber frz_offsets[MaxHeapTuplesPerPage];
2614 [ + + + + : 149893 : bool do_prune = nredirected > 0 || ndead > 0 || nunused > 0;
+ + ]
2615 : 149893 : bool do_set_vm = vmflags & VISIBILITYMAP_VALID_BITS;
2616 : 149893 : bool heap_fpi_allowed = true;
2617 : :
2618 : : Assert((vmflags & VISIBILITYMAP_VALID_BITS) == vmflags);
2619 : :
2620 : 149893 : xlrec.flags = 0;
2621 : 149893 : regbuf_flags_heap = REGBUF_STANDARD;
2622 : :
2623 : : /*
2624 : : * We can avoid an FPI of the heap page if the only modification we are
2625 : : * making to it is to set PD_ALL_VISIBLE and checksums/wal_log_hints are
2626 : : * disabled.
2627 : : *
2628 : : * However, if the page has never been WAL-logged (LSN is invalid), we
2629 : : * must force an FPI regardless. This can happen when another backend
2630 : : * extends the heap, initializes the page, and then fails before WAL-
2631 : : * logging it. Since heap extension is not WAL-logged, recovery might try
2632 : : * to replay our record and find that the page isn't initialized, which
2633 : : * would cause a PANIC.
2634 : : */
2635 [ - + ]: 149893 : if (!XLogRecPtrIsValid(PageGetLSN(heap_page)))
2636 : 0 : regbuf_flags_heap |= REGBUF_FORCE_IMAGE;
2637 [ + + + + : 149893 : else if (!do_prune && nfrozen == 0 && (!do_set_vm || !XLogHintBitIsNeeded()))
+ - + + +
+ ]
2638 : : {
2639 : 2852 : regbuf_flags_heap |= REGBUF_NO_IMAGE;
2640 : 2852 : heap_fpi_allowed = false;
2641 : : }
2642 : :
2643 : : /*
2644 : : * Prepare data for the buffer. The arrays are not actually in the
2645 : : * buffer, but we pretend that they are. When XLogInsert stores a full
2646 : : * page image, the arrays can be omitted.
2647 : : */
2648 : 149893 : XLogBeginInsert();
2649 : 149893 : XLogRegisterBuffer(0, buffer, regbuf_flags_heap);
2650 : :
2651 [ + + ]: 149893 : if (do_set_vm)
2652 : 78028 : XLogRegisterBuffer(1, vmbuffer, 0);
2653 : :
2654 [ + + ]: 149893 : if (nfrozen > 0)
2655 : : {
2656 : : int nplans;
2657 : :
2658 : 27341 : xlrec.flags |= XLHP_HAS_FREEZE_PLANS;
2659 : :
2660 : : /*
2661 : : * Prepare deduplicated representation for use in the WAL record. This
2662 : : * destructively sorts frozen tuples array in-place.
2663 : : */
2664 : 27341 : nplans = heap_log_freeze_plan(frozen, nfrozen, plans, frz_offsets);
2665 : :
2666 : 27341 : freeze_plans.nplans = nplans;
2667 : 27341 : XLogRegisterBufData(0, &freeze_plans,
2668 : : offsetof(xlhp_freeze_plans, plans));
2669 : 27341 : XLogRegisterBufData(0, plans,
2670 : : sizeof(xlhp_freeze_plan) * nplans);
2671 : : }
2672 [ + + ]: 149893 : if (nredirected > 0)
2673 : : {
2674 : 17784 : xlrec.flags |= XLHP_HAS_REDIRECTIONS;
2675 : :
2676 : 17784 : redirect_items.ntargets = nredirected;
2677 : 17784 : XLogRegisterBufData(0, &redirect_items,
2678 : : offsetof(xlhp_prune_items, data));
2679 : 17784 : XLogRegisterBufData(0, redirected,
2680 : : sizeof(OffsetNumber[2]) * nredirected);
2681 : : }
2682 [ + + ]: 149893 : if (ndead > 0)
2683 : : {
2684 : 57425 : xlrec.flags |= XLHP_HAS_DEAD_ITEMS;
2685 : :
2686 : 57425 : dead_items.ntargets = ndead;
2687 : 57425 : XLogRegisterBufData(0, &dead_items,
2688 : : offsetof(xlhp_prune_items, data));
2689 : 57425 : XLogRegisterBufData(0, dead,
2690 : : sizeof(OffsetNumber) * ndead);
2691 : : }
2692 [ + + ]: 149893 : if (nunused > 0)
2693 : : {
2694 : 30655 : xlrec.flags |= XLHP_HAS_NOW_UNUSED_ITEMS;
2695 : :
2696 : 30655 : unused_items.ntargets = nunused;
2697 : 30655 : XLogRegisterBufData(0, &unused_items,
2698 : : offsetof(xlhp_prune_items, data));
2699 : 30655 : XLogRegisterBufData(0, unused,
2700 : : sizeof(OffsetNumber) * nunused);
2701 : : }
2702 [ + + ]: 149893 : if (nfrozen > 0)
2703 : 27341 : XLogRegisterBufData(0, frz_offsets,
2704 : : sizeof(OffsetNumber) * nfrozen);
2705 : :
2706 : : /*
2707 : : * Prepare the main xl_heap_prune record. We already set the XLHP_HAS_*
2708 : : * flag above.
2709 : : */
2710 [ + + ]: 149893 : if (vmflags & VISIBILITYMAP_ALL_VISIBLE)
2711 : : {
2712 : 78028 : xlrec.flags |= XLHP_VM_ALL_VISIBLE;
2713 [ + + ]: 78028 : if (vmflags & VISIBILITYMAP_ALL_FROZEN)
2714 : 46732 : xlrec.flags |= XLHP_VM_ALL_FROZEN;
2715 : : }
2716 [ + + + + : 149893 : if (RelationIsAccessibleInLogicalDecoding(relation))
+ - - + -
- - - + +
+ + - + -
- + - ]
2717 : 662 : xlrec.flags |= XLHP_IS_CATALOG_REL;
2718 [ + + ]: 149893 : if (TransactionIdIsValid(conflict_xid))
2719 : 119310 : xlrec.flags |= XLHP_HAS_CONFLICT_HORIZON;
2720 [ + + ]: 149893 : if (cleanup_lock)
2721 : 133326 : xlrec.flags |= XLHP_CLEANUP_LOCK;
2722 : : else
2723 : : {
2724 : : Assert(nredirected == 0 && ndead == 0);
2725 : : /* also, any items in 'unused' must've been LP_DEAD previously */
2726 : : }
2727 : 149893 : XLogRegisterData(&xlrec, SizeOfHeapPrune);
2728 [ + + ]: 149893 : if (TransactionIdIsValid(conflict_xid))
2729 : 119310 : XLogRegisterData(&conflict_xid, sizeof(TransactionId));
2730 : :
2731 [ + + + - ]: 149893 : switch (reason)
2732 : : {
2733 : 70541 : case PRUNE_ON_ACCESS:
2734 : 70541 : info = XLOG_HEAP2_PRUNE_ON_ACCESS;
2735 : 70541 : break;
2736 : 62785 : case PRUNE_VACUUM_SCAN:
2737 : 62785 : info = XLOG_HEAP2_PRUNE_VACUUM_SCAN;
2738 : 62785 : break;
2739 : 16567 : case PRUNE_VACUUM_CLEANUP:
2740 : 16567 : info = XLOG_HEAP2_PRUNE_VACUUM_CLEANUP;
2741 : 16567 : break;
2742 : 0 : default:
2743 [ # # ]: 0 : elog(ERROR, "unrecognized prune reason: %d", (int) reason);
2744 : : break;
2745 : : }
2746 : 149893 : recptr = XLogInsert(RM_HEAP2_ID, info);
2747 : :
2748 [ + + ]: 149893 : if (do_set_vm)
2749 : : {
2750 : : Assert(BufferIsDirty(vmbuffer));
2751 : 78028 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2752 : : }
2753 : :
2754 : : /*
2755 : : * If we explicitly skip an FPI, we must not stamp the heap page with this
2756 : : * record's LSN. Recovery skips records <= the stamped LSN, so this could
2757 : : * lead to skipping an earlier FPI needed to repair a torn page.
2758 : : */
2759 [ + + ]: 149893 : if (heap_fpi_allowed)
2760 : : {
2761 : : Assert(BufferIsDirty(buffer));
2762 : 147041 : PageSetLSN(heap_page, recptr);
2763 : : }
2764 : 149893 : }
|
