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 : 12259450 : heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer,
273 : : bool rel_read_only)
274 : : {
275 : 12259450 : 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 [ + + ]: 12259450 : if (RecoveryInProgress())
286 : 271203 : 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 : 11988247 : prune_xid = PageGetPruneXid(page);
295 [ + + ]: 11988247 : if (!TransactionIdIsValid(prune_xid))
296 : 8217018 : return;
297 : :
298 : : /*
299 : : * Check whether prune_xid indicates that there may be dead rows that can
300 : : * be cleaned up.
301 : : */
302 : 3771229 : vistest = GlobalVisTestFor(relation);
303 : :
304 [ + + ]: 3771229 : if (!GlobalVisTestIsRemovableXid(vistest, prune_xid, true))
305 : 1436643 : 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 [ + + ]: 2334586 : minfree = RelationGetTargetPageFreeSpace(relation,
320 : : HEAP_DEFAULT_FILLFACTOR);
321 : 2334586 : minfree = Max(minfree, BLCKSZ / 10);
322 : :
323 [ + + + + ]: 2334586 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
324 : : {
325 : 137324 : bool record_free_space = false;
326 : 137324 : 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 : 137324 : visibilitymap_pin(relation, BufferGetBlockNumber(buffer), vmbuffer);
335 : :
336 : : /* OK, try to get exclusive buffer lock */
337 [ + + ]: 137324 : if (!ConditionalLockBufferForCleanup(buffer))
338 : 1810 : 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 [ + + + + ]: 135514 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
346 : : {
347 : : OffsetNumber dummy_off_loc;
348 : : PruneFreezeResult presult;
349 : : PruneFreezeParams params;
350 : :
351 : 135513 : params.relation = relation;
352 : 135513 : params.buffer = buffer;
353 : 135513 : params.vmbuffer = *vmbuffer;
354 : 135513 : params.reason = PRUNE_ON_ACCESS;
355 : 135513 : params.vistest = vistest;
356 : 135513 : 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 : 135513 : params.options = HEAP_PAGE_PRUNE_ALLOW_FAST_PATH;
365 [ + + ]: 135513 : if (rel_read_only)
366 : 38570 : params.options |= HEAP_PAGE_PRUNE_SET_VM;
367 : :
368 : 135513 : 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 [ + + ]: 135513 : if (presult.ndeleted > presult.nnewlpdead)
386 : 20582 : pgstat_update_heap_dead_tuples(relation,
387 : 20582 : 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 [ + + ]: 135513 : if (presult.newly_all_visible)
397 : : {
398 : 29796 : record_free_space = true;
399 : 29796 : freespace = PageGetHeapFreeSpace(page);
400 : : }
401 : : }
402 : :
403 : : /* And release buffer lock */
404 : 135514 : 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 [ + + ]: 135514 : if (record_free_space)
413 : 29796 : 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 : 640228 : 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 : 640228 : prstate->vistest = params->vistest;
433 : 640228 : prstate->mark_unused_now =
434 : 640228 : (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 : 640228 : prstate->attempt_freeze = (params->options & HEAP_PAGE_PRUNE_FREEZE) != 0;
439 : 640228 : prstate->attempt_set_vm = (params->options & HEAP_PAGE_PRUNE_SET_VM) != 0;
440 : 640228 : prstate->cutoffs = params->cutoffs;
441 : 640228 : prstate->relation = params->relation;
442 : 640228 : prstate->block = BufferGetBlockNumber(params->buffer);
443 : 640228 : prstate->buffer = params->buffer;
444 : 640228 : prstate->page = BufferGetPage(params->buffer);
445 : :
446 : : /*
447 : : * The caller must have pinned the VM page covering this heap block. If it
448 : : * doesn't have the correct page pinned, visibilitymap_get_status() will
449 : : * silently release the caller's pin and take its own, leaving the caller
450 : : * holding a stale buffer and leaking ours.
451 : : */
452 : : Assert(visibilitymap_pin_ok(prstate->block, params->vmbuffer));
453 : 640228 : prstate->vmbuffer = params->vmbuffer;
454 : 640228 : prstate->new_vmbits = 0;
455 : 640228 : prstate->old_vmbits = visibilitymap_get_status(prstate->relation,
456 : : prstate->block,
457 : : &prstate->vmbuffer);
458 : :
459 : : /*
460 : : * Our strategy is to scan the page and make lists of items to change,
461 : : * then apply the changes within a critical section. This keeps as much
462 : : * logic as possible out of the critical section, and also ensures that
463 : : * WAL replay will work the same as the normal case.
464 : : *
465 : : * First, initialize the new pd_prune_xid value to zero (indicating no
466 : : * prunable tuples). If we find any tuples which may soon become
467 : : * prunable, we will save the lowest relevant XID in new_prune_xid. Also
468 : : * initialize the rest of our working state.
469 : : */
470 : 640228 : prstate->new_prune_xid = InvalidTransactionId;
471 : 640228 : prstate->latest_xid_removed = InvalidTransactionId;
472 : 640228 : prstate->nredirected = prstate->ndead = prstate->nunused = 0;
473 : 640228 : prstate->nfrozen = 0;
474 : 640228 : prstate->nroot_items = 0;
475 : 640228 : prstate->nheaponly_items = 0;
476 : :
477 : : /* initialize page freezing working state */
478 : 640228 : prstate->pagefrz.freeze_required = false;
479 : 640228 : prstate->pagefrz.FreezePageConflictXid = InvalidTransactionId;
480 [ + + ]: 640228 : if (prstate->attempt_freeze)
481 : : {
482 : : Assert(new_relfrozen_xid && new_relmin_mxid);
483 : 504715 : prstate->pagefrz.FreezePageRelfrozenXid = *new_relfrozen_xid;
484 : 504715 : prstate->pagefrz.NoFreezePageRelfrozenXid = *new_relfrozen_xid;
485 : 504715 : prstate->pagefrz.FreezePageRelminMxid = *new_relmin_mxid;
486 : 504715 : prstate->pagefrz.NoFreezePageRelminMxid = *new_relmin_mxid;
487 : : }
488 : : else
489 : : {
490 : : Assert(!new_relfrozen_xid && !new_relmin_mxid);
491 : 135513 : prstate->pagefrz.FreezePageRelminMxid = InvalidMultiXactId;
492 : 135513 : prstate->pagefrz.NoFreezePageRelminMxid = InvalidMultiXactId;
493 : 135513 : prstate->pagefrz.FreezePageRelfrozenXid = InvalidTransactionId;
494 : 135513 : prstate->pagefrz.NoFreezePageRelfrozenXid = InvalidTransactionId;
495 : : }
496 : :
497 : 640228 : prstate->ndeleted = 0;
498 : 640228 : prstate->live_tuples = 0;
499 : 640228 : prstate->recently_dead_tuples = 0;
500 : 640228 : prstate->hastup = false;
501 : 640228 : prstate->lpdead_items = 0;
502 : :
503 : : /*
504 : : * deadoffsets are filled in during pruning but are only used to populate
505 : : * PruneFreezeResult->deadoffsets. To avoid needing two copies of the
506 : : * array, just save a pointer to the result offsets array in the
507 : : * PruneState.
508 : : */
509 : 640228 : prstate->deadoffsets = presult->deadoffsets;
510 : :
511 : : /*
512 : : * We track whether the page will be all-visible/all-frozen at the end of
513 : : * pruning and freezing. While examining tuple visibility, we'll set
514 : : * set_all_visible to false if there are tuples on the page not visible to
515 : : * all running and future transactions. If setting the VM is enabled for
516 : : * this scan, we will do so if the page ends up being all-visible.
517 : : *
518 : : * We also keep track of the newest live XID, which is used to calculate
519 : : * the snapshot conflict horizon for a WAL record setting the VM.
520 : : */
521 : 640228 : prstate->set_all_visible = prstate->attempt_set_vm;
522 : 640228 : prstate->newest_live_xid = InvalidTransactionId;
523 : :
524 : : /*
525 : : * Currently, only VACUUM performs freezing, but other callers may in the
526 : : * future. We must initialize set_all_frozen based on whether or not the
527 : : * caller passed HEAP_PAGE_PRUNE_FREEZE, because if they did not, we won't
528 : : * call heap_prepare_freeze_tuple() for each tuple, and set_all_frozen
529 : : * will never be cleared for tuples that need freezing. This would lead to
530 : : * incorrectly setting the visibility map all-frozen for this page. We
531 : : * can't set the page all-frozen in the VM if the caller didn't pass
532 : : * HEAP_PAGE_PRUNE_SET_VM.
533 : : *
534 : : * When freezing is not required (no XIDs/MXIDs older than the freeze
535 : : * cutoff), we may still choose to "opportunistically" freeze if doing so
536 : : * would make the page all-frozen.
537 : : *
538 : : * We will not be able to freeze the whole page at the end of vacuum if
539 : : * there are tuples present that are not visible to everyone or if there
540 : : * are dead tuples which will not be removable. However, dead tuples that
541 : : * will be removed by the end of vacuum should not prevent this
542 : : * opportunistic freezing.
543 : : *
544 : : * Therefore, we do not clear set_all_visible and set_all_frozen when we
545 : : * encounter LP_DEAD items. Instead, we correct them after deciding
546 : : * whether to freeze, but before updating the VM, to avoid setting the VM
547 : : * bits incorrectly.
548 : : */
549 [ + + + - ]: 640228 : prstate->set_all_frozen = prstate->attempt_freeze && prstate->attempt_set_vm;
550 : 640228 : }
551 : :
552 : : /*
553 : : * Helper for heap_page_prune_and_freeze(). Iterates over every tuple on the
554 : : * page, examines its visibility information, and determines the appropriate
555 : : * action for each tuple. All tuples are processed and classified during this
556 : : * phase, but no modifications are made to the page until the later execution
557 : : * stage.
558 : : *
559 : : * *off_loc is used for error callback and cleared before returning.
560 : : */
561 : : static void
562 : 441507 : prune_freeze_plan(PruneState *prstate, OffsetNumber *off_loc)
563 : : {
564 : 441507 : Page page = prstate->page;
565 : 441507 : BlockNumber blockno = prstate->block;
566 : 441507 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
567 : : OffsetNumber offnum;
568 : : HeapTupleData tup;
569 : :
570 : 441507 : tup.t_tableOid = RelationGetRelid(prstate->relation);
571 : :
572 : : /*
573 : : * Determine HTSV for all tuples, and queue them up for processing as HOT
574 : : * chain roots or as heap-only items.
575 : : *
576 : : * Determining HTSV only once for each tuple is required for correctness,
577 : : * to deal with cases where running HTSV twice could result in different
578 : : * results. For example, RECENTLY_DEAD can turn to DEAD if another
579 : : * checked item causes GlobalVisTestIsRemovableFullXid() to update the
580 : : * horizon, or INSERT_IN_PROGRESS can change to DEAD if the inserting
581 : : * transaction aborts.
582 : : *
583 : : * It's also good for performance. Most commonly tuples within a page are
584 : : * stored at decreasing offsets (while the items are stored at increasing
585 : : * offsets). When processing all tuples on a page this leads to reading
586 : : * memory at decreasing offsets within a page, with a variable stride.
587 : : * That's hard for CPU prefetchers to deal with. Processing the items in
588 : : * reverse order (and thus the tuples in increasing order) increases
589 : : * prefetching efficiency significantly / decreases the number of cache
590 : : * misses.
591 : : */
592 : 441507 : for (offnum = maxoff;
593 [ + + ]: 29620160 : offnum >= FirstOffsetNumber;
594 : 29178653 : offnum = OffsetNumberPrev(offnum))
595 : : {
596 : 29178653 : ItemId itemid = PageGetItemId(page, offnum);
597 : : HeapTupleHeader htup;
598 : :
599 : : /*
600 : : * Set the offset number so that we can display it along with any
601 : : * error that occurred while processing this tuple.
602 : : */
603 : 29178653 : *off_loc = offnum;
604 : :
605 : 29178653 : prstate->processed[offnum] = false;
606 : 29178653 : prstate->htsv[offnum] = -1;
607 : :
608 : : /* Nothing to do if slot doesn't contain a tuple */
609 [ + + ]: 29178653 : if (!ItemIdIsUsed(itemid))
610 : : {
611 : 197890 : heap_prune_record_unchanged_lp_unused(prstate, offnum);
612 : 197890 : continue;
613 : : }
614 : :
615 [ + + ]: 28980763 : if (ItemIdIsDead(itemid))
616 : : {
617 : : /*
618 : : * If the caller set mark_unused_now true, we can set dead line
619 : : * pointers LP_UNUSED now.
620 : : */
621 [ + + ]: 1700175 : if (unlikely(prstate->mark_unused_now))
622 : 1592 : heap_prune_record_unused(prstate, offnum, false);
623 : : else
624 : 1698583 : heap_prune_record_unchanged_lp_dead(prstate, offnum);
625 : 1700175 : continue;
626 : : }
627 : :
628 [ + + ]: 27280588 : if (ItemIdIsRedirected(itemid))
629 : : {
630 : : /* This is the start of a HOT chain */
631 : 200000 : prstate->root_items[prstate->nroot_items++] = offnum;
632 : 200000 : continue;
633 : : }
634 : :
635 : : Assert(ItemIdIsNormal(itemid));
636 : :
637 : : /*
638 : : * Get the tuple's visibility status and queue it up for processing.
639 : : */
640 : 27080588 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
641 : 27080588 : tup.t_data = htup;
642 : 27080588 : tup.t_len = ItemIdGetLength(itemid);
643 : 27080588 : ItemPointerSet(&tup.t_self, blockno, offnum);
644 : :
645 : 27080588 : prstate->htsv[offnum] = heap_prune_satisfies_vacuum(prstate, &tup);
646 : :
647 [ + + ]: 27080588 : if (!HeapTupleHeaderIsHeapOnly(htup))
648 : 26738733 : prstate->root_items[prstate->nroot_items++] = offnum;
649 : : else
650 : 341855 : prstate->heaponly_items[prstate->nheaponly_items++] = offnum;
651 : : }
652 : :
653 : : /*
654 : : * Process HOT chains.
655 : : *
656 : : * We added the items to the array starting from 'maxoff', so by
657 : : * processing the array in reverse order, we process the items in
658 : : * ascending offset number order. The order doesn't matter for
659 : : * correctness, but some quick micro-benchmarking suggests that this is
660 : : * faster. (Earlier PostgreSQL versions, which scanned all the items on
661 : : * the page instead of using the root_items array, also did it in
662 : : * ascending offset number order.)
663 : : */
664 [ + + ]: 27380240 : for (int i = prstate->nroot_items - 1; i >= 0; i--)
665 : : {
666 : 26938733 : offnum = prstate->root_items[i];
667 : :
668 : : /* Ignore items already processed as part of an earlier chain */
669 [ - + ]: 26938733 : if (prstate->processed[offnum])
670 : 0 : continue;
671 : :
672 : : /* see preceding loop */
673 : 26938733 : *off_loc = offnum;
674 : :
675 : : /* Process this item or chain of items */
676 : 26938733 : heap_prune_chain(maxoff, offnum, prstate);
677 : : }
678 : :
679 : : /*
680 : : * Process any heap-only tuples that were not already processed as part of
681 : : * a HOT chain.
682 : : */
683 [ + + ]: 783362 : for (int i = prstate->nheaponly_items - 1; i >= 0; i--)
684 : : {
685 : 341855 : offnum = prstate->heaponly_items[i];
686 : :
687 [ + + ]: 341855 : if (prstate->processed[offnum])
688 : 325080 : continue;
689 : :
690 : : /* see preceding loop */
691 : 16775 : *off_loc = offnum;
692 : :
693 : : /*
694 : : * If the tuple is DEAD and doesn't chain to anything else, mark it
695 : : * unused. (If it does chain, we can only remove it as part of
696 : : * pruning its chain.)
697 : : *
698 : : * We need this primarily to handle aborted HOT updates, that is,
699 : : * XMIN_INVALID heap-only tuples. Those might not be linked to by any
700 : : * chain, since the parent tuple might be re-updated before any
701 : : * pruning occurs. So we have to be able to reap them separately from
702 : : * chain-pruning. (Note that HeapTupleHeaderIsHotUpdated will never
703 : : * return true for an XMIN_INVALID tuple, so this code will work even
704 : : * when there were sequential updates within the aborted transaction.)
705 : : */
706 [ + + ]: 16775 : if (prstate->htsv[offnum] == HEAPTUPLE_DEAD)
707 : : {
708 : 3505 : ItemId itemid = PageGetItemId(page, offnum);
709 : 3505 : HeapTupleHeader htup = (HeapTupleHeader) PageGetItem(page, itemid);
710 : :
711 [ + - ]: 3505 : if (likely(!HeapTupleHeaderIsHotUpdated(htup)))
712 : : {
713 : 3505 : HeapTupleHeaderAdvanceConflictHorizon(htup,
714 : : &prstate->latest_xid_removed);
715 : 3505 : heap_prune_record_unused(prstate, offnum, true);
716 : : }
717 : : else
718 : : {
719 : : /*
720 : : * This tuple should've been processed and removed as part of
721 : : * a HOT chain, so something's wrong. To preserve evidence,
722 : : * we don't dare to remove it. We cannot leave behind a DEAD
723 : : * tuple either, because that will cause VACUUM to error out.
724 : : * Throwing an error with a distinct error message seems like
725 : : * the least bad option.
726 : : */
727 [ # # ]: 0 : elog(ERROR, "dead heap-only tuple (%u, %d) is not linked to from any HOT chain",
728 : : blockno, offnum);
729 : : }
730 : : }
731 : : else
732 : 13270 : heap_prune_record_unchanged_lp_normal(prstate, offnum);
733 : : }
734 : :
735 : : /* We should now have processed every tuple exactly once */
736 : : #ifdef USE_ASSERT_CHECKING
737 : : for (offnum = FirstOffsetNumber;
738 : : offnum <= maxoff;
739 : : offnum = OffsetNumberNext(offnum))
740 : : {
741 : : *off_loc = offnum;
742 : :
743 : : Assert(prstate->processed[offnum]);
744 : : }
745 : : #endif
746 : :
747 : : /* Clear the offset information once we have processed the given page. */
748 : 441507 : *off_loc = InvalidOffsetNumber;
749 : 441507 : }
750 : :
751 : : /*
752 : : * Decide whether to proceed with freezing according to the freeze plans
753 : : * prepared for the current heap buffer. If freezing is chosen, this function
754 : : * performs several pre-freeze checks.
755 : : *
756 : : * The values of do_prune, do_hint_prune, and did_tuple_hint_fpi must be
757 : : * determined before calling this function.
758 : : *
759 : : * prstate is both an input and output parameter.
760 : : *
761 : : * Returns true if we should apply the freeze plans and freeze tuples on the
762 : : * page, and false otherwise.
763 : : */
764 : : static bool
765 : 441507 : heap_page_will_freeze(bool did_tuple_hint_fpi,
766 : : bool do_prune,
767 : : bool do_hint_prune,
768 : : PruneState *prstate)
769 : : {
770 : 441507 : bool do_freeze = false;
771 : :
772 : : /*
773 : : * If the caller specified we should not attempt to freeze any tuples,
774 : : * validate that everything is in the right state and return.
775 : : */
776 [ + + ]: 441507 : if (!prstate->attempt_freeze)
777 : : {
778 : : Assert(!prstate->set_all_frozen && prstate->nfrozen == 0);
779 : 135513 : return false;
780 : : }
781 : :
782 [ + + ]: 305994 : if (prstate->pagefrz.freeze_required)
783 : : {
784 : : /*
785 : : * heap_prepare_freeze_tuple indicated that at least one XID/MXID from
786 : : * before FreezeLimit/MultiXactCutoff is present. Must freeze to
787 : : * advance relfrozenxid/relminmxid.
788 : : */
789 : 23513 : do_freeze = true;
790 : : }
791 : : else
792 : : {
793 : : /*
794 : : * Opportunistically freeze the page if we are generating an FPI
795 : : * anyway and if doing so means that we can set the page all-frozen
796 : : * afterwards (might not happen until VACUUM's final heap pass).
797 : : *
798 : : * XXX: Previously, we knew if pruning emitted an FPI by checking
799 : : * pgWalUsage.wal_fpi before and after pruning. Once the freeze and
800 : : * prune records were combined, this heuristic couldn't be used
801 : : * anymore. The opportunistic freeze heuristic must be improved;
802 : : * however, for now, try to approximate the old logic.
803 : : */
804 [ + + + + ]: 282481 : if (prstate->set_all_frozen && prstate->nfrozen > 0)
805 : : {
806 : : Assert(prstate->set_all_visible);
807 : :
808 : : /*
809 : : * Freezing would make the page all-frozen. Have already emitted
810 : : * an FPI or will do so anyway?
811 : : */
812 [ + + + + : 22964 : if (RelationNeedsWAL(prstate->relation))
+ - + - ]
813 : : {
814 [ + + ]: 21474 : if (did_tuple_hint_fpi)
815 : 1451 : do_freeze = true;
816 [ + + ]: 20023 : else if (do_prune)
817 : : {
818 [ + + ]: 2295 : if (XLogCheckBufferNeedsBackup(prstate->buffer))
819 : 753 : do_freeze = true;
820 : : }
821 [ + + ]: 17728 : else if (do_hint_prune)
822 : : {
823 [ + + + - : 22318 : if (XLogHintBitIsNeeded() &&
+ + ]
824 : 11159 : XLogCheckBufferNeedsBackup(prstate->buffer))
825 : 1911 : do_freeze = true;
826 : : }
827 : : }
828 : : }
829 : : }
830 : :
831 [ + + ]: 305994 : if (do_freeze)
832 : : {
833 : : /*
834 : : * Validate the tuples we will be freezing before entering the
835 : : * critical section.
836 : : */
837 : 27628 : heap_pre_freeze_checks(prstate->buffer, prstate->frozen, prstate->nfrozen);
838 : : Assert(TransactionIdPrecedes(prstate->pagefrz.FreezePageConflictXid,
839 : : prstate->cutoffs->OldestXmin));
840 : : }
841 [ + + ]: 278366 : else if (prstate->nfrozen > 0)
842 : : {
843 : : /*
844 : : * The page contained some tuples that were not already frozen, and we
845 : : * chose not to freeze them now. The page won't be all-frozen then.
846 : : */
847 : : Assert(!prstate->pagefrz.freeze_required);
848 : :
849 : 19259 : prstate->set_all_frozen = false;
850 : 19259 : prstate->nfrozen = 0; /* avoid miscounts in instrumentation */
851 : : }
852 : : else
853 : : {
854 : : /*
855 : : * We have no freeze plans to execute. The page might already be
856 : : * all-frozen (perhaps only following pruning), though. Such pages
857 : : * can be marked all-frozen in the VM by our caller, even though none
858 : : * of its tuples were newly frozen here.
859 : : */
860 : : }
861 : :
862 : 305994 : return do_freeze;
863 : : }
864 : :
865 : : /*
866 : : * Emit a warning about and fix visibility map corruption on the given page.
867 : : *
868 : : * The caller specifies the type of corruption it has already detected via
869 : : * corruption_type, so that we can emit the appropriate warning. All cases
870 : : * result in the VM bits being cleared; corruption types where PD_ALL_VISIBLE
871 : : * is incorrectly set also clear PD_ALL_VISIBLE.
872 : : *
873 : : * Must be called while holding an exclusive lock on the heap buffer. Dead
874 : : * items and not all-visible tuples must have been discovered under that same
875 : : * lock. Although we do not hold a lock on the VM buffer, it is pinned, and
876 : : * the heap buffer is exclusively locked, ensuring that no other backend can
877 : : * update the VM bits corresponding to this heap page.
878 : : *
879 : : * This function makes changes to the VM and, potentially, the heap page, but
880 : : * it does not need to be done in a critical section.
881 : : */
882 : : static void
883 : 0 : heap_page_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum,
884 : : VMCorruptionType corruption_type)
885 : : {
886 : 0 : const char *relname = RelationGetRelationName(prstate->relation);
887 : 0 : bool do_clear_vm = false;
888 : 0 : bool do_clear_heap = false;
889 : :
890 : : Assert(BufferIsLockedByMeInMode(prstate->buffer, BUFFER_LOCK_EXCLUSIVE));
891 : :
892 [ # # # # ]: 0 : switch (corruption_type)
893 : : {
894 : 0 : case VM_CORRUPT_LPDEAD:
895 [ # # ]: 0 : ereport(WARNING,
896 : : (errcode(ERRCODE_DATA_CORRUPTED),
897 : : errmsg("dead line pointer found on page marked all-visible"),
898 : : errcontext("relation \"%s\", page %u, tuple %u",
899 : : relname, prstate->block, offnum)));
900 : 0 : do_clear_vm = true;
901 : 0 : do_clear_heap = true;
902 : 0 : break;
903 : :
904 : 0 : case VM_CORRUPT_TUPLE_VISIBILITY:
905 : :
906 : : /*
907 : : * A HEAPTUPLE_LIVE tuple on an all-visible page can appear to not
908 : : * be visible to everyone when
909 : : * GetOldestNonRemovableTransactionId() returns a conservative
910 : : * value that's older than the real safe xmin. That is not
911 : : * corruption -- the PD_ALL_VISIBLE flag is still correct.
912 : : *
913 : : * However, dead tuple versions, in-progress inserts, and
914 : : * in-progress deletes should never appear on a page marked
915 : : * all-visible. That indicates real corruption. PD_ALL_VISIBLE
916 : : * should have been cleared by the DML operation that deleted or
917 : : * inserted the tuple.
918 : : */
919 [ # # ]: 0 : ereport(WARNING,
920 : : (errcode(ERRCODE_DATA_CORRUPTED),
921 : : errmsg("tuple not visible to all transactions found on page marked all-visible"),
922 : : errcontext("relation \"%s\", page %u, tuple %u",
923 : : relname, prstate->block, offnum)));
924 : 0 : do_clear_vm = true;
925 : 0 : do_clear_heap = true;
926 : 0 : break;
927 : :
928 : 0 : case VM_CORRUPT_MISSING_PAGE_HINT:
929 : :
930 : : /*
931 : : * As of PostgreSQL 9.2, the visibility map bit should never be
932 : : * set if the page-level bit is clear. However, for vacuum, it's
933 : : * possible that the bit got cleared after
934 : : * heap_vac_scan_next_block() was called, so we must recheck now
935 : : * that we have the buffer lock before concluding that the VM is
936 : : * corrupt.
937 : : */
938 : : Assert(!PageIsAllVisible(prstate->page));
939 : : Assert(prstate->old_vmbits & VISIBILITYMAP_VALID_BITS);
940 [ # # ]: 0 : ereport(WARNING,
941 : : (errcode(ERRCODE_DATA_CORRUPTED),
942 : : errmsg("page is not marked all-visible but visibility map bit is set"),
943 : : errcontext("relation \"%s\", page %u",
944 : : relname, prstate->block)));
945 : 0 : do_clear_vm = true;
946 : 0 : break;
947 : : }
948 : :
949 : : Assert(do_clear_heap || do_clear_vm);
950 : :
951 : : /* Avoid marking the buffer dirty if PD_ALL_VISIBLE is already clear */
952 [ # # ]: 0 : if (do_clear_heap)
953 : : {
954 : : Assert(PageIsAllVisible(prstate->page));
955 : 0 : PageClearAllVisible(prstate->page);
956 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
957 : : }
958 : :
959 [ # # ]: 0 : if (do_clear_vm)
960 : : {
961 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_EXCLUSIVE);
962 : : /* This VM clear is not WAL-logged, so its return value is not needed. */
963 : 0 : (void) visibilitymap_clear(prstate->relation->rd_locator,
964 : : prstate->block, prstate->vmbuffer,
965 : : VISIBILITYMAP_VALID_BITS);
966 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_UNLOCK);
967 : 0 : prstate->old_vmbits = 0;
968 : : }
969 : 0 : }
970 : :
971 : : /*
972 : : * Decide whether to set the visibility map bits (all-visible and all-frozen)
973 : : * for the current page using information from the PruneState and VM.
974 : : *
975 : : * This function does not actually set the VM bits or page-level visibility
976 : : * hint, PD_ALL_VISIBLE.
977 : : *
978 : : * This should be called only after do_freeze has been decided (and do_prune
979 : : * has been set), as these factor into our heuristic-based decision.
980 : : *
981 : : * Returns true if one or both VM bits should be set and false otherwise.
982 : : */
983 : : static bool
984 : 441507 : heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
985 : : bool do_prune, bool do_freeze)
986 : : {
987 [ + + ]: 441507 : if (!prstate->attempt_set_vm)
988 : 96943 : return false;
989 : :
990 [ + + ]: 344564 : if (!prstate->set_all_visible)
991 : 264609 : return false;
992 : :
993 : : /*
994 : : * If this is an on-access call and we're not actually pruning or
995 : : * freezing, consider whether setting the VM would cost us an additional
996 : : * heap page FPI. If the relation isn't WAL-logged, or if hint bits are
997 : : * not WAL-logged, setting the VM won't include a heap page FPI (the
998 : : * latter passes REGBUF_NO_IMAGE for the heap page), apart from a page
999 : : * that has never been WAL-logged, which we don't bother about here.
1000 : : */
1001 [ + + + + : 79955 : if (reason == PRUNE_ON_ACCESS && !do_prune && !do_freeze &&
+ - ]
1002 [ + + + + : 30370 : RelationNeedsWAL(prstate->relation) && XLogHintBitIsNeeded())
+ - + - +
+ + + ]
1003 : : {
1004 : : /*
1005 : : * Because the page is known to be all-visible, we will clear
1006 : : * pd_prune_xid regardless of whether we actually set the page
1007 : : * all-visible in the VM. That clear is a hint update which is not
1008 : : * WAL-logged, other than an FPI for torn-page protection, so in some
1009 : : * cases we want to avoid setting the VM if doing so would cost us a
1010 : : * heap page FPI that clearing pd_prune_xid wouldn't have.
1011 : : *
1012 : : * Since hint bits are WAL-logged, if the buffer is clean, clearing
1013 : : * pd_prune_xid will already emit a heap page FPI if one is needed, so
1014 : : * there's no reason to avoid setting the VM.
1015 : : *
1016 : : * However, if the heap buffer is already dirty, clearing pd_prune_xid
1017 : : * will never emit an FPI. So avoid setting the VM if the page hasn't
1018 : : * been WAL-logged since the current checkpoint began, as the record
1019 : : * setting the VM would then include a heap page FPI.
1020 : : */
1021 [ + + + + ]: 32005 : if (BufferIsDirty(prstate->buffer) &&
1022 : 14952 : XLogCheckBufferNeedsBackup(prstate->buffer))
1023 : : {
1024 : 692 : prstate->set_all_visible = prstate->set_all_frozen = false;
1025 : 692 : return false;
1026 : : }
1027 : : }
1028 : :
1029 : 79263 : prstate->new_vmbits = VISIBILITYMAP_ALL_VISIBLE;
1030 : :
1031 [ + + ]: 79263 : if (prstate->set_all_frozen)
1032 : 34936 : prstate->new_vmbits |= VISIBILITYMAP_ALL_FROZEN;
1033 : :
1034 [ + + ]: 79263 : if (prstate->new_vmbits == prstate->old_vmbits)
1035 : : {
1036 : 2859 : prstate->new_vmbits = 0;
1037 : 2859 : return false;
1038 : : }
1039 : :
1040 : 76404 : return true;
1041 : : }
1042 : :
1043 : : /*
1044 : : * If the page is already all-frozen, or already all-visible and freezing
1045 : : * won't be attempted, there is no remaining work and we can use the fast path
1046 : : * to avoid the expensive overhead of heap_page_prune_and_freeze().
1047 : : *
1048 : : * This can happen when the page has a stale prune hint, or if VACUUM is
1049 : : * scanning an already all-frozen page due to SKIP_PAGES_THRESHOLD.
1050 : : *
1051 : : * The caller must already have examined the visibility map and saved the
1052 : : * status of the page's VM bits in prstate->old_vmbits. Caller must hold a
1053 : : * content lock on the heap page since it will examine line pointers.
1054 : : *
1055 : : * Before calling prune_freeze_fast_path(), the caller should first
1056 : : * check for and fix any discrepancy between the page-level visibility hint
1057 : : * and the visibility map. Otherwise, the fast path will always prevent us
1058 : : * from getting them in sync. Note that if there are tuples on the page that
1059 : : * are not visible to all but the VM is incorrectly marked
1060 : : * all-visible/all-frozen, we will not get the chance to fix that corruption
1061 : : * when using the fast path.
1062 : : */
1063 : : static void
1064 : 198721 : prune_freeze_fast_path(PruneState *prstate, PruneFreezeResult *presult)
1065 : : {
1066 : 198721 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
1067 : 198721 : Page page = prstate->page;
1068 : :
1069 : : Assert((prstate->old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1070 : : ((prstate->old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1071 : : !prstate->attempt_freeze));
1072 : :
1073 : : /* We'll fill in presult for the caller */
1074 : 198721 : memset(presult, 0, sizeof(PruneFreezeResult));
1075 : :
1076 : : /* Clear any stale prune hint */
1077 [ - + ]: 198721 : if (TransactionIdIsValid(PageGetPruneXid(page)))
1078 : : {
1079 : 0 : PageClearPrunable(page);
1080 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
1081 : : }
1082 : :
1083 [ - + ]: 198721 : if (PageIsEmpty(page))
1084 : 0 : return;
1085 : :
1086 : : /*
1087 : : * Since the page is all-visible, a count of the normal ItemIds on the
1088 : : * page should be sufficient for vacuum's live tuple count.
1089 : : */
1090 : 198721 : for (OffsetNumber off = FirstOffsetNumber;
1091 [ + + ]: 11387432 : off <= maxoff;
1092 : 11188711 : off = OffsetNumberNext(off))
1093 : : {
1094 : 11188711 : ItemId lp = PageGetItemId(page, off);
1095 : :
1096 [ + + ]: 11188711 : if (!ItemIdIsUsed(lp))
1097 : 265771 : continue;
1098 : :
1099 : 10922940 : presult->hastup = true;
1100 : :
1101 [ + + ]: 10922940 : if (ItemIdIsNormal(lp))
1102 : 10751356 : prstate->live_tuples++;
1103 : : }
1104 : :
1105 : 198721 : presult->live_tuples = prstate->live_tuples;
1106 : : }
1107 : :
1108 : : /*
1109 : : * Prune and repair fragmentation and potentially freeze tuples on the
1110 : : * specified page. If the page's visibility status has changed, update it in
1111 : : * the VM.
1112 : : *
1113 : : * Caller must have pin and buffer cleanup lock on the page. Note that we
1114 : : * don't update the FSM information for page on caller's behalf. Caller might
1115 : : * also need to account for a reduction in the length of the line pointer
1116 : : * array following array truncation by us.
1117 : : *
1118 : : * params contains the input parameters used to control freezing and pruning
1119 : : * behavior. See the definition of PruneFreezeParams for more on what each
1120 : : * parameter does.
1121 : : *
1122 : : * If the HEAP_PAGE_PRUNE_FREEZE option is set in params, we will freeze
1123 : : * tuples if it's required in order to advance relfrozenxid / relminmxid, or
1124 : : * if it's considered advantageous for overall system performance to do so
1125 : : * now. The 'params.cutoffs', 'presult', 'new_relfrozen_xid' and
1126 : : * 'new_relmin_mxid' arguments are required when freezing.
1127 : : *
1128 : : * A vmbuffer corresponding to the heap page is also passed and if the page is
1129 : : * found to be all-visible/all-frozen, we will set it in the VM.
1130 : : *
1131 : : * presult contains output parameters needed by callers, such as the number of
1132 : : * tuples removed and the offsets of dead items on the page after pruning.
1133 : : * heap_page_prune_and_freeze() is responsible for initializing it. Required
1134 : : * by all callers.
1135 : : *
1136 : : * off_loc is the offset location required by the caller to use in error
1137 : : * callback.
1138 : : *
1139 : : * new_relfrozen_xid and new_relmin_mxid must be provided by the caller if the
1140 : : * HEAP_PAGE_PRUNE_FREEZE option is set in params. On entry, they contain the
1141 : : * oldest XID and multi-XID seen on the relation so far. They will be updated
1142 : : * with the oldest values present on the page after pruning. After processing
1143 : : * the whole relation, VACUUM can use these values as the new
1144 : : * relfrozenxid/relminmxid for the relation.
1145 : : */
1146 : : void
1147 : 640228 : heap_page_prune_and_freeze(PruneFreezeParams *params,
1148 : : PruneFreezeResult *presult,
1149 : : OffsetNumber *off_loc,
1150 : : TransactionId *new_relfrozen_xid,
1151 : : MultiXactId *new_relmin_mxid)
1152 : : {
1153 : : PruneState prstate;
1154 : : bool do_freeze;
1155 : : bool do_prune;
1156 : : bool do_hint_prune;
1157 : : bool do_set_vm;
1158 : : bool did_tuple_hint_fpi;
1159 : 640228 : int64 fpi_before = pgWalUsage.wal_fpi;
1160 : : TransactionId conflict_xid;
1161 : :
1162 : : /* Initialize prstate */
1163 : 640228 : prune_freeze_setup(params,
1164 : : new_relfrozen_xid, new_relmin_mxid,
1165 : : presult, &prstate);
1166 : :
1167 : : /*
1168 : : * If the VM is set but PD_ALL_VISIBLE is clear, fix that corruption
1169 : : * before pruning and freezing so that the page and VM start out in a
1170 : : * consistent state.
1171 : : */
1172 [ + + ]: 640228 : if ((prstate.old_vmbits & VISIBILITYMAP_VALID_BITS) &&
1173 [ - + ]: 205749 : !PageIsAllVisible(prstate.page))
1174 : 0 : heap_page_fix_vm_corruption(&prstate, InvalidOffsetNumber,
1175 : : VM_CORRUPT_MISSING_PAGE_HINT);
1176 : :
1177 : : /*
1178 : : * If the page is already all-frozen, or already all-visible when freezing
1179 : : * is not being attempted, take the fast path, skipping pruning and
1180 : : * freezing code entirely. This must be done after fixing any discrepancy
1181 : : * between the page-level visibility hint and the VM, since that may have
1182 : : * cleared old_vmbits.
1183 : : */
1184 [ + + ]: 640228 : if ((params->options & HEAP_PAGE_PRUNE_ALLOW_FAST_PATH) != 0 &&
1185 [ + + ]: 638932 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1186 [ + + ]: 440211 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1187 [ - + ]: 6583 : !prstate.attempt_freeze)))
1188 : : {
1189 : 198721 : prune_freeze_fast_path(&prstate, presult);
1190 : 198721 : return;
1191 : : }
1192 : :
1193 : : /*
1194 : : * Examine all line pointers and tuple visibility information to determine
1195 : : * which line pointers should change state and which tuples may be frozen.
1196 : : * Prepare queue of state changes to later be executed in a critical
1197 : : * section.
1198 : : */
1199 : 441507 : prune_freeze_plan(&prstate, off_loc);
1200 : :
1201 : : /*
1202 : : * After processing all the live tuples on the page, if the newest xmin
1203 : : * amongst them may be considered running by any snapshot, the page cannot
1204 : : * be all-visible. This should be done before determining whether or not
1205 : : * to opportunistically freeze.
1206 : : */
1207 [ + + ]: 441507 : if (prstate.set_all_visible &&
1208 [ + + + + ]: 192971 : TransactionIdIsNormal(prstate.newest_live_xid) &&
1209 : 82020 : GlobalVisTestXidConsideredRunning(prstate.vistest,
1210 : : prstate.newest_live_xid,
1211 : : true))
1212 : 2970 : prstate.set_all_visible = prstate.set_all_frozen = false;
1213 : :
1214 : : /*
1215 : : * If checksums are enabled, calling heap_prune_satisfies_vacuum() while
1216 : : * checking tuple visibility information in prune_freeze_plan() may have
1217 : : * caused an FPI to be emitted.
1218 : : */
1219 : 441507 : did_tuple_hint_fpi = fpi_before != pgWalUsage.wal_fpi;
1220 : :
1221 : 1306673 : do_prune = prstate.nredirected > 0 ||
1222 [ + + + + ]: 811607 : prstate.ndead > 0 ||
1223 [ + + ]: 370100 : prstate.nunused > 0;
1224 : :
1225 : : /*
1226 : : * Even if we don't prune anything, if we found a new value for the
1227 : : * pd_prune_xid field or the page was marked full, we will update the hint
1228 : : * bit.
1229 : : */
1230 [ + + + + ]: 703212 : do_hint_prune = PageGetPruneXid(prstate.page) != prstate.new_prune_xid ||
1231 : 261705 : PageIsFull(prstate.page);
1232 : :
1233 : : /*
1234 : : * Decide if we want to go ahead with freezing according to the freeze
1235 : : * plans we prepared, or not.
1236 : : */
1237 : 441507 : do_freeze = heap_page_will_freeze(did_tuple_hint_fpi,
1238 : : do_prune,
1239 : : do_hint_prune,
1240 : : &prstate);
1241 : :
1242 : : /*
1243 : : * While scanning the line pointers, we did not clear
1244 : : * set_all_visible/set_all_frozen when encountering LP_DEAD items because
1245 : : * we wanted the decision whether or not to freeze the page to be
1246 : : * unaffected by the short-term presence of LP_DEAD items. These LP_DEAD
1247 : : * items are effectively assumed to be LP_UNUSED items in the making. It
1248 : : * doesn't matter which vacuum heap pass (initial pass or final pass) ends
1249 : : * up setting the page all-frozen, as long as the ongoing VACUUM does it.
1250 : : *
1251 : : * Now that we finished determining whether or not to freeze the page,
1252 : : * update set_all_visible and set_all_frozen so that they reflect the true
1253 : : * state of the page for setting PD_ALL_VISIBLE and VM bits.
1254 : : */
1255 [ + + ]: 441507 : if (prstate.lpdead_items > 0)
1256 : 75247 : prstate.set_all_visible = prstate.set_all_frozen = false;
1257 : :
1258 : : Assert(!prstate.set_all_frozen || prstate.set_all_visible);
1259 : : Assert(!prstate.set_all_visible || prstate.attempt_set_vm);
1260 : : Assert(!prstate.set_all_visible || (prstate.lpdead_items == 0));
1261 : :
1262 : 441507 : do_set_vm = heap_page_will_set_vm(&prstate, params->reason, do_prune, do_freeze);
1263 : :
1264 : : /*
1265 : : * new_vmbits should be 0 regardless of whether or not the page is
1266 : : * all-visible if we do not intend to set the VM.
1267 : : */
1268 : : Assert(do_set_vm || prstate.new_vmbits == 0);
1269 : :
1270 : : /*
1271 : : * The snapshot conflict horizon for the whole record is the most
1272 : : * conservative (newest) horizon required by any change in the record.
1273 : : */
1274 : 441507 : conflict_xid = InvalidTransactionId;
1275 [ + + ]: 441507 : if (do_set_vm)
1276 : 76404 : conflict_xid = prstate.newest_live_xid;
1277 [ + + + + ]: 441507 : if (do_freeze && TransactionIdFollows(prstate.pagefrz.FreezePageConflictXid, conflict_xid))
1278 : 4570 : conflict_xid = prstate.pagefrz.FreezePageConflictXid;
1279 [ + + + + ]: 441507 : if (do_prune && TransactionIdFollows(prstate.latest_xid_removed, conflict_xid))
1280 : 63386 : conflict_xid = prstate.latest_xid_removed;
1281 : :
1282 : : /* Lock vmbuffer before entering a critical section */
1283 [ + + ]: 441507 : if (do_set_vm)
1284 : 76404 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_EXCLUSIVE);
1285 : :
1286 : : /* Any error while applying the changes is critical */
1287 : 441507 : START_CRIT_SECTION();
1288 : :
1289 [ + + ]: 441507 : if (do_hint_prune)
1290 : : {
1291 : : /*
1292 : : * Update the page's pd_prune_xid field to either zero, or the lowest
1293 : : * XID of any soon-prunable tuple.
1294 : : */
1295 : 179901 : ((PageHeader) prstate.page)->pd_prune_xid = prstate.new_prune_xid;
1296 : :
1297 : : /*
1298 : : * Also clear the "page is full" flag, since there's no point in
1299 : : * repeating the prune/defrag process until something else happens to
1300 : : * the page.
1301 : : */
1302 : 179901 : PageClearFull(prstate.page);
1303 : :
1304 : : /*
1305 : : * If that's all we had to do to the page, this is a non-WAL-logged
1306 : : * hint. If we are going to freeze or prune the page or set
1307 : : * PD_ALL_VISIBLE, we will mark the buffer dirty below.
1308 : : *
1309 : : * Setting PD_ALL_VISIBLE is fully WAL-logged because it is forbidden
1310 : : * for the VM to be set and PD_ALL_VISIBLE to be clear.
1311 : : */
1312 [ + + + + : 179901 : if (!do_freeze && !do_prune && !do_set_vm)
+ + ]
1313 : 53645 : MarkBufferDirtyHint(prstate.buffer, true);
1314 : : }
1315 : :
1316 [ + + + + : 441507 : if (do_prune || do_freeze || do_set_vm)
+ + ]
1317 : : {
1318 : : /* Apply the planned item changes and repair page fragmentation. */
1319 [ + + ]: 150011 : if (do_prune)
1320 : : {
1321 : 71858 : heap_page_prune_execute(prstate.buffer, false,
1322 : : prstate.redirected, prstate.nredirected,
1323 : : prstate.nowdead, prstate.ndead,
1324 : : prstate.nowunused, prstate.nunused);
1325 : : }
1326 : :
1327 [ + + ]: 150011 : if (do_freeze)
1328 : 27628 : heap_freeze_prepared_tuples(prstate.buffer, prstate.frozen, prstate.nfrozen);
1329 : :
1330 : : /* Set the visibility map and page visibility hint */
1331 [ + + ]: 150011 : if (do_set_vm)
1332 : : {
1333 : : /*
1334 : : * While it is valid for PD_ALL_VISIBLE to be set when the
1335 : : * corresponding VM bit is clear, we strongly prefer to keep them
1336 : : * in sync.
1337 : : *
1338 : : * The heap buffer must be marked dirty before adding it to the
1339 : : * WAL chain when setting the VM. We don't worry about
1340 : : * unnecessarily dirtying the heap buffer if PD_ALL_VISIBLE is
1341 : : * already set, though. It is extremely rare to have a clean heap
1342 : : * buffer with PD_ALL_VISIBLE already set and the VM bits clear,
1343 : : * so there is no point in optimizing it.
1344 : : */
1345 : 76404 : PageSetAllVisible(prstate.page);
1346 : 76404 : PageClearPrunable(prstate.page);
1347 : 76404 : (void) visibilitymap_set(prstate.block, prstate.vmbuffer,
1348 : 76404 : prstate.new_vmbits,
1349 : 76404 : prstate.relation->rd_locator);
1350 : : }
1351 : :
1352 : 150011 : MarkBufferDirty(prstate.buffer);
1353 : :
1354 : : /*
1355 : : * Emit a WAL XLOG_HEAP2_PRUNE* record showing what we did
1356 : : */
1357 [ + + + + : 150011 : if (RelationNeedsWAL(prstate.relation))
+ - + - ]
1358 : : {
1359 [ + + + + ]: 198525 : log_heap_prune_and_freeze(prstate.relation, prstate.buffer,
1360 : : do_set_vm ? prstate.vmbuffer : InvalidBuffer,
1361 : 63112 : do_set_vm ? prstate.new_vmbits : 0,
1362 : : conflict_xid,
1363 : : do_prune, /* cleanup lock */
1364 : : params->reason,
1365 : : prstate.frozen, prstate.nfrozen,
1366 : : prstate.redirected, prstate.nredirected,
1367 : : prstate.nowdead, prstate.ndead,
1368 : : prstate.nowunused, prstate.nunused);
1369 : : }
1370 : : }
1371 : :
1372 : 441507 : END_CRIT_SECTION();
1373 : :
1374 [ + + ]: 441507 : if (do_set_vm)
1375 : 76404 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_UNLOCK);
1376 : :
1377 : : /*
1378 : : * During its second pass over the heap, VACUUM calls
1379 : : * heap_page_would_be_all_visible() to determine whether a page is
1380 : : * all-visible and all-frozen. The logic here is similar. After completing
1381 : : * pruning and freezing, use an assertion to verify that our results
1382 : : * remain consistent with heap_page_would_be_all_visible(). It's also a
1383 : : * valuable cross-check of the page state after pruning and freezing.
1384 : : */
1385 : : #ifdef USE_ASSERT_CHECKING
1386 : : if (prstate.set_all_visible)
1387 : : {
1388 : : TransactionId debug_cutoff;
1389 : : bool debug_all_frozen;
1390 : :
1391 : : Assert(prstate.lpdead_items == 0);
1392 : :
1393 : : Assert(heap_page_is_all_visible(prstate.relation, prstate.buffer,
1394 : : prstate.vistest,
1395 : : &debug_all_frozen,
1396 : : &debug_cutoff, off_loc));
1397 : :
1398 : : Assert(!TransactionIdIsValid(debug_cutoff) ||
1399 : : debug_cutoff == prstate.newest_live_xid);
1400 : :
1401 : : /*
1402 : : * It's possible the page is composed entirely of frozen tuples but is
1403 : : * not set all-frozen in the VM and did not pass
1404 : : * HEAP_PAGE_PRUNE_FREEZE. In this case, it's possible
1405 : : * heap_page_is_all_visible() finds the page completely frozen, even
1406 : : * though prstate.set_all_frozen is false.
1407 : : */
1408 : : Assert(!prstate.set_all_frozen || debug_all_frozen);
1409 : : }
1410 : : #endif
1411 : :
1412 : : /* Copy information back for caller */
1413 : 441507 : presult->ndeleted = prstate.ndeleted;
1414 : 441507 : presult->nnewlpdead = prstate.ndead;
1415 : 441507 : presult->nfrozen = prstate.nfrozen;
1416 : 441507 : presult->live_tuples = prstate.live_tuples;
1417 : 441507 : presult->recently_dead_tuples = prstate.recently_dead_tuples;
1418 : 441507 : presult->hastup = prstate.hastup;
1419 : :
1420 : 441507 : presult->lpdead_items = prstate.lpdead_items;
1421 : : /* the presult->deadoffsets array was already filled in */
1422 : :
1423 : 441507 : presult->newly_all_visible = false;
1424 : 441507 : presult->newly_all_frozen = false;
1425 : 441507 : presult->newly_all_visible_frozen = false;
1426 [ + + ]: 441507 : if (do_set_vm)
1427 : : {
1428 [ + + ]: 76404 : if ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
1429 : : {
1430 : 72235 : presult->newly_all_visible = true;
1431 [ + + ]: 72235 : if (prstate.set_all_frozen)
1432 : 30323 : presult->newly_all_visible_frozen = true;
1433 : : }
1434 [ + - ]: 4169 : else if ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
1435 [ + - ]: 4169 : prstate.set_all_frozen)
1436 : 4169 : presult->newly_all_frozen = true;
1437 : : }
1438 : :
1439 [ + + ]: 441507 : if (prstate.attempt_freeze)
1440 : : {
1441 [ + + ]: 305994 : if (presult->nfrozen > 0)
1442 : : {
1443 : 27628 : *new_relfrozen_xid = prstate.pagefrz.FreezePageRelfrozenXid;
1444 : 27628 : *new_relmin_mxid = prstate.pagefrz.FreezePageRelminMxid;
1445 : : }
1446 : : else
1447 : : {
1448 : 278366 : *new_relfrozen_xid = prstate.pagefrz.NoFreezePageRelfrozenXid;
1449 : 278366 : *new_relmin_mxid = prstate.pagefrz.NoFreezePageRelminMxid;
1450 : : }
1451 : : }
1452 : : }
1453 : :
1454 : :
1455 : : /*
1456 : : * Perform visibility checks for heap pruning.
1457 : : */
1458 : : static HTSV_Result
1459 : 27080588 : heap_prune_satisfies_vacuum(PruneState *prstate, HeapTuple tup)
1460 : : {
1461 : : HTSV_Result res;
1462 : : TransactionId dead_after;
1463 : :
1464 : 27080588 : res = HeapTupleSatisfiesVacuumHorizon(tup, prstate->buffer, &dead_after);
1465 : :
1466 [ + + ]: 27080588 : if (res != HEAPTUPLE_RECENTLY_DEAD)
1467 : 22822684 : return res;
1468 : :
1469 : : /*
1470 : : * For VACUUM, we must be sure to prune tuples with xmax older than
1471 : : * OldestXmin -- a visibility cutoff determined at the beginning of
1472 : : * vacuuming the relation. OldestXmin is used for freezing determination
1473 : : * and we cannot freeze dead tuples' xmaxes.
1474 : : */
1475 [ + + ]: 4257904 : if (prstate->cutoffs &&
1476 [ + - ]: 1303478 : TransactionIdIsValid(prstate->cutoffs->OldestXmin) &&
1477 [ + + ]: 1303478 : NormalTransactionIdPrecedes(dead_after, prstate->cutoffs->OldestXmin))
1478 : 969918 : return HEAPTUPLE_DEAD;
1479 : :
1480 : : /*
1481 : : * Determine whether or not the tuple is considered dead when compared
1482 : : * with the provided GlobalVisState. On-access pruning does not provide
1483 : : * VacuumCutoffs. And for vacuum, even if the tuple's xmax is not older
1484 : : * than OldestXmin, GlobalVisTestIsRemovableXid() could find the row dead
1485 : : * if the GlobalVisState has been updated since the beginning of vacuuming
1486 : : * the relation.
1487 : : */
1488 [ + + ]: 3287986 : if (GlobalVisTestIsRemovableXid(prstate->vistest, dead_after, true))
1489 : 2906304 : return HEAPTUPLE_DEAD;
1490 : :
1491 : 381682 : return res;
1492 : : }
1493 : :
1494 : :
1495 : : /*
1496 : : * Pruning calculates tuple visibility once and saves the results in an array
1497 : : * of int8. See PruneState.htsv for details. This helper function is meant
1498 : : * to guard against examining visibility status array members which have not
1499 : : * yet been computed.
1500 : : */
1501 : : static inline HTSV_Result
1502 : 27063813 : htsv_get_valid_status(int status)
1503 : : {
1504 : : Assert(status >= HEAPTUPLE_DEAD &&
1505 : : status <= HEAPTUPLE_DELETE_IN_PROGRESS);
1506 : 27063813 : return (HTSV_Result) status;
1507 : : }
1508 : :
1509 : : /*
1510 : : * Prune specified line pointer or a HOT chain originating at line pointer.
1511 : : *
1512 : : * Tuple visibility information is provided in prstate->htsv.
1513 : : *
1514 : : * If the item is an index-referenced tuple (i.e. not a heap-only tuple),
1515 : : * the HOT chain is pruned by removing all DEAD tuples at the start of the HOT
1516 : : * chain. We also prune any RECENTLY_DEAD tuples preceding a DEAD tuple.
1517 : : * This is OK because a RECENTLY_DEAD tuple preceding a DEAD tuple is really
1518 : : * DEAD, our visibility test is just too coarse to detect it.
1519 : : *
1520 : : * Pruning must never leave behind a DEAD tuple that still has tuple storage.
1521 : : * VACUUM isn't prepared to deal with that case.
1522 : : *
1523 : : * The root line pointer is redirected to the tuple immediately after the
1524 : : * latest DEAD tuple. If all tuples in the chain are DEAD, the root line
1525 : : * pointer is marked LP_DEAD. (This includes the case of a DEAD simple
1526 : : * tuple, which we treat as a chain of length 1.)
1527 : : *
1528 : : * We don't actually change the page here. We just add entries to the arrays in
1529 : : * prstate showing the changes to be made. Items to be redirected are added
1530 : : * to the redirected[] array (two entries per redirection); items to be set to
1531 : : * LP_DEAD state are added to nowdead[]; and items to be set to LP_UNUSED
1532 : : * state are added to nowunused[]. We perform bookkeeping of live tuples,
1533 : : * visibility etc. based on what the page will look like after the changes
1534 : : * applied. All that bookkeeping is performed in the heap_prune_record_*()
1535 : : * subroutines. The division of labor is that heap_prune_chain() decides the
1536 : : * fate of each tuple, ie. whether it's going to be removed, redirected or
1537 : : * left unchanged, and the heap_prune_record_*() subroutines update PruneState
1538 : : * based on that outcome.
1539 : : */
1540 : : static void
1541 : 26938733 : heap_prune_chain(OffsetNumber maxoff, OffsetNumber rootoffnum,
1542 : : PruneState *prstate)
1543 : : {
1544 : 26938733 : TransactionId priorXmax = InvalidTransactionId;
1545 : : ItemId rootlp;
1546 : : OffsetNumber offnum;
1547 : : OffsetNumber chainitems[MaxHeapTuplesPerPage];
1548 : 26938733 : Page page = prstate->page;
1549 : :
1550 : : /*
1551 : : * After traversing the HOT chain, ndeadchain is the index in chainitems
1552 : : * of the first live successor after the last dead item.
1553 : : */
1554 : 26938733 : int ndeadchain = 0,
1555 : 26938733 : nchain = 0;
1556 : :
1557 : 26938733 : rootlp = PageGetItemId(page, rootoffnum);
1558 : :
1559 : : /* Start from the root tuple */
1560 : 26938733 : offnum = rootoffnum;
1561 : :
1562 : : /* while not end of the chain */
1563 : : for (;;)
1564 : 325080 : {
1565 : : HeapTupleHeader htup;
1566 : : ItemId lp;
1567 : :
1568 : : /* Sanity check (pure paranoia) */
1569 [ - + ]: 27263813 : if (offnum < FirstOffsetNumber)
1570 : 0 : break;
1571 : :
1572 : : /*
1573 : : * An offset past the end of page's line pointer array is possible
1574 : : * when the array was truncated (original item must have been unused)
1575 : : */
1576 [ - + ]: 27263813 : if (offnum > maxoff)
1577 : 0 : break;
1578 : :
1579 : : /* If item is already processed, stop --- it must not be same chain */
1580 [ - + ]: 27263813 : if (prstate->processed[offnum])
1581 : 0 : break;
1582 : :
1583 : 27263813 : lp = PageGetItemId(page, offnum);
1584 : :
1585 : : /*
1586 : : * Unused item obviously isn't part of the chain. Likewise, a dead
1587 : : * line pointer can't be part of the chain. Both of those cases were
1588 : : * already marked as processed.
1589 : : */
1590 : : Assert(ItemIdIsUsed(lp));
1591 : : Assert(!ItemIdIsDead(lp));
1592 : :
1593 : : /*
1594 : : * If we are looking at the redirected root line pointer, jump to the
1595 : : * first normal tuple in the chain. If we find a redirect somewhere
1596 : : * else, stop --- it must not be same chain.
1597 : : */
1598 [ + + ]: 27263813 : if (ItemIdIsRedirected(lp))
1599 : : {
1600 [ - + ]: 200000 : if (nchain > 0)
1601 : 0 : break; /* not at start of chain */
1602 : 200000 : chainitems[nchain++] = offnum;
1603 : 200000 : offnum = ItemIdGetRedirect(rootlp);
1604 : 200000 : continue;
1605 : : }
1606 : :
1607 : : Assert(ItemIdIsNormal(lp));
1608 : :
1609 : 27063813 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1610 : :
1611 : : /*
1612 : : * Check the tuple XMIN against prior XMAX, if any
1613 : : */
1614 [ + + - + ]: 27188893 : if (TransactionIdIsValid(priorXmax) &&
1615 : 125080 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
1616 : 0 : break;
1617 : :
1618 : : /*
1619 : : * OK, this tuple is indeed a member of the chain.
1620 : : */
1621 : 27063813 : chainitems[nchain++] = offnum;
1622 : :
1623 [ + + + - ]: 27063813 : switch (htsv_get_valid_status(prstate->htsv[offnum]))
1624 : : {
1625 : 3949267 : case HEAPTUPLE_DEAD:
1626 : :
1627 : : /* Remember the last DEAD tuple seen */
1628 : 3949267 : ndeadchain = nchain;
1629 : 3949267 : HeapTupleHeaderAdvanceConflictHorizon(htup,
1630 : : &prstate->latest_xid_removed);
1631 : : /* Advance to next chain member */
1632 : 3949267 : break;
1633 : :
1634 : 381682 : case HEAPTUPLE_RECENTLY_DEAD:
1635 : :
1636 : : /*
1637 : : * We don't need to advance the conflict horizon for
1638 : : * RECENTLY_DEAD tuples, even if we are removing them. This
1639 : : * is because we only remove RECENTLY_DEAD tuples if they
1640 : : * precede a DEAD tuple, and the DEAD tuple must have been
1641 : : * inserted by a newer transaction than the RECENTLY_DEAD
1642 : : * tuple by virtue of being later in the chain. We will have
1643 : : * advanced the conflict horizon for the DEAD tuple.
1644 : : */
1645 : :
1646 : : /*
1647 : : * Advance past RECENTLY_DEAD tuples just in case there's a
1648 : : * DEAD one after them. We have to make sure that we don't
1649 : : * miss any DEAD tuples, since DEAD tuples that still have
1650 : : * tuple storage after pruning will confuse VACUUM.
1651 : : */
1652 : 381682 : break;
1653 : :
1654 : 22732864 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1655 : : case HEAPTUPLE_LIVE:
1656 : : case HEAPTUPLE_INSERT_IN_PROGRESS:
1657 : 22732864 : goto process_chain;
1658 : :
1659 : 0 : default:
1660 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1661 : : goto process_chain;
1662 : : }
1663 : :
1664 : : /*
1665 : : * If the tuple is not HOT-updated, then we are at the end of this
1666 : : * HOT-update chain.
1667 : : */
1668 [ + + ]: 4330949 : if (!HeapTupleHeaderIsHotUpdated(htup))
1669 : 4205869 : goto process_chain;
1670 : :
1671 : : /* HOT implies it can't have moved to different partition */
1672 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1673 : :
1674 : : /*
1675 : : * Advance to next chain member.
1676 : : */
1677 : : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == prstate->block);
1678 : 125080 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
1679 : 125080 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1680 : : }
1681 : :
1682 [ # # # # ]: 0 : if (ItemIdIsRedirected(rootlp) && nchain < 2)
1683 : : {
1684 : : /*
1685 : : * We found a redirect item that doesn't point to a valid follow-on
1686 : : * item. This can happen if the loop in heap_page_prune_and_freeze()
1687 : : * caused us to visit the dead successor of a redirect item before
1688 : : * visiting the redirect item. We can clean up by setting the
1689 : : * redirect item to LP_DEAD state or LP_UNUSED if the caller
1690 : : * indicated.
1691 : : */
1692 : 0 : heap_prune_record_dead_or_unused(prstate, rootoffnum, false);
1693 : 0 : return;
1694 : : }
1695 : :
1696 : 0 : process_chain:
1697 : :
1698 [ + + ]: 26938733 : if (ndeadchain == 0)
1699 : : {
1700 : : /*
1701 : : * No DEAD tuple was found, so the chain is entirely composed of
1702 : : * normal, unchanged tuples. Leave it alone.
1703 : : */
1704 : 23034873 : int i = 0;
1705 : :
1706 [ + + ]: 23034873 : if (ItemIdIsRedirected(rootlp))
1707 : : {
1708 : 178064 : heap_prune_record_unchanged_lp_redirect(prstate, rootoffnum);
1709 : 178064 : i++;
1710 : : }
1711 [ + + ]: 46074723 : for (; i < nchain; i++)
1712 : 23039850 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1713 : : }
1714 [ + + ]: 3903860 : else if (ndeadchain == nchain)
1715 : : {
1716 : : /*
1717 : : * The entire chain is dead. Mark the root line pointer LP_DEAD, and
1718 : : * fully remove the other tuples in the chain.
1719 : : */
1720 : 3830836 : heap_prune_record_dead_or_unused(prstate, rootoffnum, ItemIdIsNormal(rootlp));
1721 [ + + ]: 3876015 : for (int i = 1; i < nchain; i++)
1722 : 45179 : heap_prune_record_unused(prstate, chainitems[i], true);
1723 : : }
1724 : : else
1725 : : {
1726 : : /*
1727 : : * We found a DEAD tuple in the chain. Redirect the root line pointer
1728 : : * to the first non-DEAD tuple, and mark as unused each intermediate
1729 : : * item that we are able to remove from the chain.
1730 : : */
1731 : 73024 : heap_prune_record_redirect(prstate, rootoffnum, chainitems[ndeadchain],
1732 : 73024 : ItemIdIsNormal(rootlp));
1733 [ + + ]: 95188 : for (int i = 1; i < ndeadchain; i++)
1734 : 22164 : heap_prune_record_unused(prstate, chainitems[i], true);
1735 : :
1736 : : /* the rest of tuples in the chain are normal, unchanged tuples */
1737 [ + + ]: 147720 : for (int i = ndeadchain; i < nchain; i++)
1738 : 74696 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1739 : : }
1740 : : }
1741 : :
1742 : : /* Record lowest soon-prunable XID */
1743 : : static void
1744 : 6251788 : heap_prune_record_prunable(PruneState *prstate, TransactionId xid,
1745 : : OffsetNumber offnum)
1746 : : {
1747 : : /*
1748 : : * This should exactly match the PageSetPrunable macro. We can't store
1749 : : * directly into the page header yet, so we update working state.
1750 : : */
1751 : : Assert(TransactionIdIsNormal(xid));
1752 [ + + + + ]: 12241368 : if (!TransactionIdIsValid(prstate->new_prune_xid) ||
1753 : 5989580 : TransactionIdPrecedes(xid, prstate->new_prune_xid))
1754 : 263591 : prstate->new_prune_xid = xid;
1755 : :
1756 : : /*
1757 : : * It's incorrect for a page to be marked all-visible if it contains
1758 : : * prunable items.
1759 : : */
1760 [ - + ]: 6251788 : if (PageIsAllVisible(prstate->page))
1761 : 0 : heap_page_fix_vm_corruption(prstate, offnum,
1762 : : VM_CORRUPT_TUPLE_VISIBILITY);
1763 : 6251788 : }
1764 : :
1765 : : /* Record line pointer to be redirected */
1766 : : static void
1767 : 73024 : heap_prune_record_redirect(PruneState *prstate,
1768 : : OffsetNumber offnum, OffsetNumber rdoffnum,
1769 : : bool was_normal)
1770 : : {
1771 : : Assert(!prstate->processed[offnum]);
1772 : 73024 : prstate->processed[offnum] = true;
1773 : :
1774 : : /*
1775 : : * Do not mark the redirect target here. It needs to be counted
1776 : : * separately as an unchanged tuple.
1777 : : */
1778 : :
1779 : : Assert(prstate->nredirected < MaxHeapTuplesPerPage);
1780 : 73024 : prstate->redirected[prstate->nredirected * 2] = offnum;
1781 : 73024 : prstate->redirected[prstate->nredirected * 2 + 1] = rdoffnum;
1782 : :
1783 : 73024 : prstate->nredirected++;
1784 : :
1785 : : /*
1786 : : * If the root entry had been a normal tuple, we are deleting it, so count
1787 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1788 : : * count.
1789 : : */
1790 [ + + ]: 73024 : if (was_normal)
1791 : 63990 : prstate->ndeleted++;
1792 : :
1793 : 73024 : prstate->hastup = true;
1794 : 73024 : }
1795 : :
1796 : : /* Record line pointer to be marked dead */
1797 : : static void
1798 : 3795543 : heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
1799 : : bool was_normal)
1800 : : {
1801 : : Assert(!prstate->processed[offnum]);
1802 : 3795543 : prstate->processed[offnum] = true;
1803 : :
1804 : : Assert(prstate->ndead < MaxHeapTuplesPerPage);
1805 : 3795543 : prstate->nowdead[prstate->ndead] = offnum;
1806 : 3795543 : prstate->ndead++;
1807 : :
1808 : : /*
1809 : : * Deliberately delay unsetting set_all_visible and set_all_frozen until
1810 : : * later during pruning. Removable dead tuples shouldn't preclude freezing
1811 : : * the page.
1812 : : */
1813 : :
1814 : : /* Record the dead offset for vacuum */
1815 : 3795543 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1816 : :
1817 : : /*
1818 : : * If the root entry had been a normal tuple, we are deleting it, so count
1819 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1820 : : * count.
1821 : : */
1822 [ + + ]: 3795543 : if (was_normal)
1823 : 3782641 : prstate->ndeleted++;
1824 : 3795543 : }
1825 : :
1826 : : /*
1827 : : * Depending on whether or not the caller set mark_unused_now to true, record that a
1828 : : * line pointer should be marked LP_DEAD or LP_UNUSED. There are other cases in
1829 : : * which we will mark line pointers LP_UNUSED, but we will not mark line
1830 : : * pointers LP_DEAD if mark_unused_now is true.
1831 : : */
1832 : : static void
1833 : 3830836 : heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
1834 : : bool was_normal)
1835 : : {
1836 : : /*
1837 : : * If the caller set mark_unused_now to true, we can remove dead tuples
1838 : : * during pruning instead of marking their line pointers dead. Set this
1839 : : * tuple's line pointer LP_UNUSED. We hint that this option is less
1840 : : * likely.
1841 : : */
1842 [ + + ]: 3830836 : if (unlikely(prstate->mark_unused_now))
1843 : 35293 : heap_prune_record_unused(prstate, offnum, was_normal);
1844 : : else
1845 : 3795543 : heap_prune_record_dead(prstate, offnum, was_normal);
1846 : :
1847 : : /*
1848 : : * It's incorrect for the page to be set all-visible if it contains dead
1849 : : * items. Fix that on the heap page and check the VM for corruption as
1850 : : * well. Do that here rather than in heap_prune_record_dead() so we also
1851 : : * cover tuples that are directly marked LP_UNUSED via mark_unused_now.
1852 : : */
1853 [ - + ]: 3830836 : if (PageIsAllVisible(prstate->page))
1854 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
1855 : 3830836 : }
1856 : :
1857 : : /* Record line pointer to be marked unused */
1858 : : static void
1859 : 107733 : heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal)
1860 : : {
1861 : : Assert(!prstate->processed[offnum]);
1862 : 107733 : prstate->processed[offnum] = true;
1863 : :
1864 : : Assert(prstate->nunused < MaxHeapTuplesPerPage);
1865 : 107733 : prstate->nowunused[prstate->nunused] = offnum;
1866 : 107733 : prstate->nunused++;
1867 : :
1868 : : /*
1869 : : * If the root entry had been a normal tuple, we are deleting it, so count
1870 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1871 : : * count.
1872 : : */
1873 [ + + ]: 107733 : if (was_normal)
1874 : 106141 : prstate->ndeleted++;
1875 : 107733 : }
1876 : :
1877 : : /*
1878 : : * Record an unused line pointer that is left unchanged.
1879 : : */
1880 : : static void
1881 : 197890 : heap_prune_record_unchanged_lp_unused(PruneState *prstate, OffsetNumber offnum)
1882 : : {
1883 : : Assert(!prstate->processed[offnum]);
1884 : 197890 : prstate->processed[offnum] = true;
1885 : 197890 : }
1886 : :
1887 : : /*
1888 : : * Record line pointer that is left unchanged. We consider freezing it, and
1889 : : * update bookkeeping of tuple counts and page visibility.
1890 : : */
1891 : : static void
1892 : 23127816 : heap_prune_record_unchanged_lp_normal(PruneState *prstate, OffsetNumber offnum)
1893 : : {
1894 : : HeapTupleHeader htup;
1895 : : TransactionId xmin;
1896 : 23127816 : Page page = prstate->page;
1897 : :
1898 : : Assert(!prstate->processed[offnum]);
1899 : 23127816 : prstate->processed[offnum] = true;
1900 : :
1901 : 23127816 : prstate->hastup = true; /* the page is not empty */
1902 : :
1903 : : /*
1904 : : * The criteria for counting a tuple as live in this block need to match
1905 : : * what analyze.c's acquire_sample_rows() does, otherwise VACUUM and
1906 : : * ANALYZE may produce wildly different reltuples values, e.g. when there
1907 : : * are many recently-dead tuples.
1908 : : *
1909 : : * The logic here is a bit simpler than acquire_sample_rows(), as VACUUM
1910 : : * can't run inside a transaction block, which makes some cases impossible
1911 : : * (e.g. in-progress insert from the same transaction).
1912 : : *
1913 : : * HEAPTUPLE_DEAD are handled by the other heap_prune_record_*()
1914 : : * subroutines. They don't count dead items like acquire_sample_rows()
1915 : : * does, because we assume that all dead items will become LP_UNUSED
1916 : : * before VACUUM finishes. This difference is only superficial. VACUUM
1917 : : * effectively agrees with ANALYZE about DEAD items, in the end. VACUUM
1918 : : * won't remember LP_DEAD items, but only because they're not supposed to
1919 : : * be left behind when it is done. (Cases where we bypass index vacuuming
1920 : : * will violate this optimistic assumption, but the overall impact of that
1921 : : * should be negligible.)
1922 : : */
1923 : 23127816 : htup = (HeapTupleHeader) PageGetItem(page, PageGetItemId(page, offnum));
1924 : :
1925 [ + + + + : 23127816 : switch (prstate->htsv[offnum])
- ]
1926 : : {
1927 : 16876028 : case HEAPTUPLE_LIVE:
1928 : :
1929 : : /*
1930 : : * Count it as live. Not only is this natural, but it's also what
1931 : : * acquire_sample_rows() does.
1932 : : */
1933 : 16876028 : prstate->live_tuples++;
1934 : :
1935 : : /*
1936 : : * Is the tuple definitely visible to all transactions?
1937 : : *
1938 : : * NB: Like with per-tuple hint bits, we can't set the
1939 : : * PD_ALL_VISIBLE flag if the inserter committed asynchronously.
1940 : : * See SetHintBits for more info. Check that the tuple is hinted
1941 : : * xmin-committed because of that.
1942 : : */
1943 [ + + ]: 16876028 : if (!HeapTupleHeaderXminCommitted(htup))
1944 : : {
1945 : 36077 : prstate->set_all_visible = false;
1946 : 36077 : prstate->set_all_frozen = false;
1947 : 36077 : break;
1948 : : }
1949 : :
1950 : : /*
1951 : : * The inserter definitely committed. But we don't know if it is
1952 : : * old enough that everyone sees it as committed. Later, after
1953 : : * processing all the tuples on the page, we'll check if there is
1954 : : * any snapshot that still considers the newest xid on the page to
1955 : : * be running. If so, we don't consider the page all-visible.
1956 : : */
1957 : 16839951 : xmin = HeapTupleHeaderGetXmin(htup);
1958 : :
1959 : : /* Track newest xmin on page. */
1960 [ + + + + ]: 16839951 : if (TransactionIdFollows(xmin, prstate->newest_live_xid) &&
1961 : : TransactionIdIsNormal(xmin))
1962 : 623737 : prstate->newest_live_xid = xmin;
1963 : :
1964 : 16839951 : break;
1965 : :
1966 : 381682 : case HEAPTUPLE_RECENTLY_DEAD:
1967 : 381682 : prstate->recently_dead_tuples++;
1968 : 381682 : prstate->set_all_visible = false;
1969 : 381682 : prstate->set_all_frozen = false;
1970 : :
1971 : : /*
1972 : : * This tuple will soon become DEAD. Update the hint field so
1973 : : * that the page is reconsidered for pruning in future.
1974 : : */
1975 : 381682 : heap_prune_record_prunable(prstate,
1976 : : HeapTupleHeaderGetUpdateXid(htup),
1977 : : offnum);
1978 : 381682 : break;
1979 : :
1980 : 139104 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1981 : :
1982 : : /*
1983 : : * We do not count these rows as live, because we expect the
1984 : : * inserting transaction to update the counters at commit, and we
1985 : : * assume that will happen only after we report our results. This
1986 : : * assumption is a bit shaky, but it is what acquire_sample_rows()
1987 : : * does, so be consistent.
1988 : : */
1989 : 139104 : prstate->set_all_visible = false;
1990 : 139104 : prstate->set_all_frozen = false;
1991 : :
1992 : : /*
1993 : : * Though there is nothing "prunable" on the page, we maintain
1994 : : * pd_prune_xid for inserts so that we have the opportunity to
1995 : : * mark them all-visible during the next round of pruning.
1996 : : */
1997 : 139104 : heap_prune_record_prunable(prstate,
1998 : : HeapTupleHeaderGetXmin(htup),
1999 : : offnum);
2000 : 139104 : break;
2001 : :
2002 : 5731002 : case HEAPTUPLE_DELETE_IN_PROGRESS:
2003 : :
2004 : : /*
2005 : : * This an expected case during concurrent vacuum. Count such
2006 : : * rows as live. As above, we assume the deleting transaction
2007 : : * will commit and update the counters after we report.
2008 : : */
2009 : 5731002 : prstate->live_tuples++;
2010 : 5731002 : prstate->set_all_visible = false;
2011 : 5731002 : prstate->set_all_frozen = false;
2012 : :
2013 : : /*
2014 : : * This tuple may soon become DEAD. Update the hint field so that
2015 : : * the page is reconsidered for pruning in future.
2016 : : */
2017 : 5731002 : heap_prune_record_prunable(prstate,
2018 : : HeapTupleHeaderGetUpdateXid(htup),
2019 : : offnum);
2020 : 5731002 : break;
2021 : :
2022 : 0 : default:
2023 : :
2024 : : /*
2025 : : * DEAD tuples should've been passed to heap_prune_record_dead()
2026 : : * or heap_prune_record_unused() instead.
2027 : : */
2028 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result %d",
2029 : : prstate->htsv[offnum]);
2030 : : break;
2031 : : }
2032 : :
2033 : : /* Consider freezing any normal tuples which will not be removed */
2034 [ + + ]: 23127816 : if (prstate->attempt_freeze)
2035 : : {
2036 : : bool totally_frozen;
2037 : :
2038 [ + + ]: 12880135 : if ((heap_prepare_freeze_tuple(htup,
2039 : 12880135 : prstate->cutoffs,
2040 : : &prstate->pagefrz,
2041 : 12880135 : &prstate->frozen[prstate->nfrozen],
2042 : : &totally_frozen)))
2043 : : {
2044 : : /* Save prepared freeze plan for later */
2045 : 3371689 : prstate->frozen[prstate->nfrozen++].offset = offnum;
2046 : : }
2047 : :
2048 : : /*
2049 : : * If any tuple isn't either totally frozen already or eligible to
2050 : : * become totally frozen (according to its freeze plan), then the page
2051 : : * definitely cannot be set all-frozen in the visibility map later on.
2052 : : */
2053 [ + + ]: 12880135 : if (!totally_frozen)
2054 : 6456408 : prstate->set_all_frozen = false;
2055 : : }
2056 : 23127816 : }
2057 : :
2058 : :
2059 : : /*
2060 : : * Record line pointer that was already LP_DEAD and is left unchanged.
2061 : : */
2062 : : static void
2063 : 1698583 : heap_prune_record_unchanged_lp_dead(PruneState *prstate, OffsetNumber offnum)
2064 : : {
2065 : : Assert(!prstate->processed[offnum]);
2066 : 1698583 : prstate->processed[offnum] = true;
2067 : :
2068 : : /*
2069 : : * Deliberately don't set hastup for LP_DEAD items. We make the soft
2070 : : * assumption that any LP_DEAD items encountered here will become
2071 : : * LP_UNUSED later on, before count_nondeletable_pages is reached. If we
2072 : : * don't make this assumption then rel truncation will only happen every
2073 : : * other VACUUM, at most. Besides, VACUUM must treat
2074 : : * hastup/nonempty_pages as provisional no matter how LP_DEAD items are
2075 : : * handled (handled here, or handled later on).
2076 : : *
2077 : : * Similarly, don't unset set_all_visible and set_all_frozen until later,
2078 : : * at the end of heap_page_prune_and_freeze(). This will allow us to
2079 : : * attempt to freeze the page after pruning. As long as we unset it
2080 : : * before updating the visibility map, this will be correct.
2081 : : */
2082 : :
2083 : : /* Record the dead offset for vacuum */
2084 : 1698583 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
2085 : :
2086 : : /*
2087 : : * It's incorrect for a page to be marked all-visible if it contains dead
2088 : : * items.
2089 : : */
2090 [ - + ]: 1698583 : if (PageIsAllVisible(prstate->page))
2091 : 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
2092 : 1698583 : }
2093 : :
2094 : : /*
2095 : : * Record LP_REDIRECT that is left unchanged.
2096 : : */
2097 : : static void
2098 : 178064 : heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum)
2099 : : {
2100 : : /*
2101 : : * A redirect line pointer doesn't count as a live tuple.
2102 : : *
2103 : : * If we leave a redirect line pointer in place, there will be another
2104 : : * tuple on the page that it points to. We will do the bookkeeping for
2105 : : * that separately. So we have nothing to do here, except remember that
2106 : : * we processed this item.
2107 : : */
2108 : : Assert(!prstate->processed[offnum]);
2109 : 178064 : prstate->processed[offnum] = true;
2110 : 178064 : }
2111 : :
2112 : : /*
2113 : : * Perform the actual page changes needed by heap_page_prune_and_freeze().
2114 : : *
2115 : : * If 'lp_truncate_only' is set, we are merely marking LP_DEAD line pointers
2116 : : * as unused, not redirecting or removing anything else. The
2117 : : * PageRepairFragmentation() call is skipped in that case.
2118 : : *
2119 : : * If 'lp_truncate_only' is not set, the caller must hold a cleanup lock on
2120 : : * the buffer. If it is set, an ordinary exclusive lock suffices.
2121 : : */
2122 : : void
2123 : 82714 : heap_page_prune_execute(Buffer buffer, bool lp_truncate_only,
2124 : : OffsetNumber *redirected, int nredirected,
2125 : : OffsetNumber *nowdead, int ndead,
2126 : : OffsetNumber *nowunused, int nunused)
2127 : : {
2128 : 82714 : Page page = BufferGetPage(buffer);
2129 : : OffsetNumber *offnum;
2130 : : HeapTupleHeader htup PG_USED_FOR_ASSERTS_ONLY;
2131 : :
2132 : : /* Shouldn't be called unless there's something to do */
2133 : : Assert(nredirected > 0 || ndead > 0 || nunused > 0);
2134 : :
2135 : : /* If 'lp_truncate_only', we can only remove already-dead line pointers */
2136 : : Assert(!lp_truncate_only || (nredirected == 0 && ndead == 0));
2137 : :
2138 : : /* Update all redirected line pointers */
2139 : 82714 : offnum = redirected;
2140 [ + + ]: 175908 : for (int i = 0; i < nredirected; i++)
2141 : : {
2142 : 93194 : OffsetNumber fromoff = *offnum++;
2143 : 93194 : OffsetNumber tooff = *offnum++;
2144 : 93194 : ItemId fromlp = PageGetItemId(page, fromoff);
2145 : : ItemId tolp PG_USED_FOR_ASSERTS_ONLY;
2146 : :
2147 : : #ifdef USE_ASSERT_CHECKING
2148 : :
2149 : : /*
2150 : : * Any existing item that we set as an LP_REDIRECT (any 'from' item)
2151 : : * must be the first item from a HOT chain. If the item has tuple
2152 : : * storage then it can't be a heap-only tuple. Otherwise we are just
2153 : : * maintaining an existing LP_REDIRECT from an existing HOT chain that
2154 : : * has been pruned at least once before now.
2155 : : */
2156 : : if (!ItemIdIsRedirected(fromlp))
2157 : : {
2158 : : Assert(ItemIdHasStorage(fromlp) && ItemIdIsNormal(fromlp));
2159 : :
2160 : : htup = (HeapTupleHeader) PageGetItem(page, fromlp);
2161 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2162 : : }
2163 : : else
2164 : : {
2165 : : /* We shouldn't need to redundantly set the redirect */
2166 : : Assert(ItemIdGetRedirect(fromlp) != tooff);
2167 : : }
2168 : :
2169 : : /*
2170 : : * The item that we're about to set as an LP_REDIRECT (the 'from'
2171 : : * item) will point to an existing item (the 'to' item) that is
2172 : : * already a heap-only tuple. There can be at most one LP_REDIRECT
2173 : : * item per HOT chain.
2174 : : *
2175 : : * We need to keep around an LP_REDIRECT item (after original
2176 : : * non-heap-only root tuple gets pruned away) so that it's always
2177 : : * possible for VACUUM to easily figure out what TID to delete from
2178 : : * indexes when an entire HOT chain becomes dead. A heap-only tuple
2179 : : * can never become LP_DEAD; an LP_REDIRECT item or a regular heap
2180 : : * tuple can.
2181 : : *
2182 : : * This check may miss problems, e.g. the target of a redirect could
2183 : : * be marked as unused subsequently. The page_verify_redirects() check
2184 : : * below will catch such problems.
2185 : : */
2186 : : tolp = PageGetItemId(page, tooff);
2187 : : Assert(ItemIdHasStorage(tolp) && ItemIdIsNormal(tolp));
2188 : : htup = (HeapTupleHeader) PageGetItem(page, tolp);
2189 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2190 : : #endif
2191 : :
2192 : 93194 : ItemIdSetRedirect(fromlp, tooff);
2193 : : }
2194 : :
2195 : : /* Update all now-dead line pointers */
2196 : 82714 : offnum = nowdead;
2197 [ + + ]: 4163272 : for (int i = 0; i < ndead; i++)
2198 : : {
2199 : 4080558 : OffsetNumber off = *offnum++;
2200 : 4080558 : ItemId lp = PageGetItemId(page, off);
2201 : :
2202 : : #ifdef USE_ASSERT_CHECKING
2203 : :
2204 : : /*
2205 : : * An LP_DEAD line pointer must be left behind when the original item
2206 : : * (which is dead to everybody) could still be referenced by a TID in
2207 : : * an index. This should never be necessary with any individual
2208 : : * heap-only tuple item, though. (It's not clear how much of a problem
2209 : : * that would be, but there is no reason to allow it.)
2210 : : */
2211 : : if (ItemIdHasStorage(lp))
2212 : : {
2213 : : Assert(ItemIdIsNormal(lp));
2214 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2215 : : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2216 : : }
2217 : : else
2218 : : {
2219 : : /* Whole HOT chain becomes dead */
2220 : : Assert(ItemIdIsRedirected(lp));
2221 : : }
2222 : : #endif
2223 : :
2224 : 4080558 : ItemIdSetDead(lp);
2225 : : }
2226 : :
2227 : : /* Update all now-unused line pointers */
2228 : 82714 : offnum = nowunused;
2229 [ + + ]: 415332 : for (int i = 0; i < nunused; i++)
2230 : : {
2231 : 332618 : OffsetNumber off = *offnum++;
2232 : 332618 : ItemId lp = PageGetItemId(page, off);
2233 : :
2234 : : #ifdef USE_ASSERT_CHECKING
2235 : :
2236 : : if (lp_truncate_only)
2237 : : {
2238 : : /* Setting LP_DEAD to LP_UNUSED in vacuum's second pass */
2239 : : Assert(ItemIdIsDead(lp) && !ItemIdHasStorage(lp));
2240 : : }
2241 : : else
2242 : : {
2243 : : /*
2244 : : * When heap_page_prune_and_freeze() was called, mark_unused_now
2245 : : * may have been passed as true, which allows would-be LP_DEAD
2246 : : * items to be made LP_UNUSED instead. This is only possible if
2247 : : * the relation has no indexes. If there are any dead items, then
2248 : : * mark_unused_now was not true and every item being marked
2249 : : * LP_UNUSED must refer to a heap-only tuple.
2250 : : */
2251 : : if (ndead > 0)
2252 : : {
2253 : : Assert(ItemIdHasStorage(lp) && ItemIdIsNormal(lp));
2254 : : htup = (HeapTupleHeader) PageGetItem(page, lp);
2255 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2256 : : }
2257 : : else
2258 : : Assert(ItemIdIsUsed(lp));
2259 : : }
2260 : :
2261 : : #endif
2262 : :
2263 : 332618 : ItemIdSetUnused(lp);
2264 : : }
2265 : :
2266 [ + + ]: 82714 : if (lp_truncate_only)
2267 : 2341 : PageTruncateLinePointerArray(page);
2268 : : else
2269 : : {
2270 : : /*
2271 : : * Finally, repair any fragmentation, and update the page's hint bit
2272 : : * about whether it has free pointers.
2273 : : */
2274 : 80373 : PageRepairFragmentation(page);
2275 : :
2276 : : /*
2277 : : * Now that the page has been modified, assert that redirect items
2278 : : * still point to valid targets.
2279 : : */
2280 : 80373 : page_verify_redirects(page);
2281 : : }
2282 : 82714 : }
2283 : :
2284 : :
2285 : : /*
2286 : : * If built with assertions, verify that all LP_REDIRECT items point to a
2287 : : * valid item.
2288 : : *
2289 : : * One way that bugs related to HOT pruning show is redirect items pointing to
2290 : : * removed tuples. It's not trivial to reliably check that marking an item
2291 : : * unused will not orphan a redirect item during heap_prune_chain() /
2292 : : * heap_page_prune_execute(), so we additionally check the whole page after
2293 : : * pruning. Without this check such bugs would typically only cause asserts
2294 : : * later, potentially well after the corruption has been introduced.
2295 : : *
2296 : : * Also check comments in heap_page_prune_execute()'s redirection loop.
2297 : : */
2298 : : static void
2299 : 80373 : page_verify_redirects(Page page)
2300 : : {
2301 : : #ifdef USE_ASSERT_CHECKING
2302 : : OffsetNumber offnum;
2303 : : OffsetNumber maxoff;
2304 : :
2305 : : maxoff = PageGetMaxOffsetNumber(page);
2306 : : for (offnum = FirstOffsetNumber;
2307 : : offnum <= maxoff;
2308 : : offnum = OffsetNumberNext(offnum))
2309 : : {
2310 : : ItemId itemid = PageGetItemId(page, offnum);
2311 : : OffsetNumber targoff;
2312 : : ItemId targitem;
2313 : : HeapTupleHeader htup;
2314 : :
2315 : : if (!ItemIdIsRedirected(itemid))
2316 : : continue;
2317 : :
2318 : : targoff = ItemIdGetRedirect(itemid);
2319 : : targitem = PageGetItemId(page, targoff);
2320 : :
2321 : : Assert(ItemIdIsUsed(targitem));
2322 : : Assert(ItemIdIsNormal(targitem));
2323 : : Assert(ItemIdHasStorage(targitem));
2324 : : htup = (HeapTupleHeader) PageGetItem(page, targitem);
2325 : : Assert(HeapTupleHeaderIsHeapOnly(htup));
2326 : : }
2327 : : #endif
2328 : 80373 : }
2329 : :
2330 : :
2331 : : /*
2332 : : * For all items in this page, find their respective root line pointers.
2333 : : * If item k is part of a HOT-chain with root at item j, then we set
2334 : : * root_offsets[k - 1] = j.
2335 : : *
2336 : : * The passed-in root_offsets array must have MaxHeapTuplesPerPage entries.
2337 : : * Unused entries are filled with InvalidOffsetNumber (zero).
2338 : : *
2339 : : * The function must be called with at least share lock on the buffer, to
2340 : : * prevent concurrent prune operations.
2341 : : *
2342 : : * Note: The information collected here is valid only as long as the caller
2343 : : * holds a pin on the buffer. Once pin is released, a tuple might be pruned
2344 : : * and reused by a completely unrelated tuple.
2345 : : */
2346 : : void
2347 : 139800 : heap_get_root_tuples(Page page, OffsetNumber *root_offsets)
2348 : : {
2349 : : OffsetNumber offnum,
2350 : : maxoff;
2351 : :
2352 [ + - - + : 139800 : MemSet(root_offsets, InvalidOffsetNumber,
- - - - -
- ]
2353 : : MaxHeapTuplesPerPage * sizeof(OffsetNumber));
2354 : :
2355 : 139800 : maxoff = PageGetMaxOffsetNumber(page);
2356 [ + + ]: 11742815 : for (offnum = FirstOffsetNumber; offnum <= maxoff; offnum = OffsetNumberNext(offnum))
2357 : : {
2358 : 11603015 : ItemId lp = PageGetItemId(page, offnum);
2359 : : HeapTupleHeader htup;
2360 : : OffsetNumber nextoffnum;
2361 : : TransactionId priorXmax;
2362 : :
2363 : : /* skip unused and dead items */
2364 [ + + + + ]: 11603015 : if (!ItemIdIsUsed(lp) || ItemIdIsDead(lp))
2365 : 11601 : continue;
2366 : :
2367 [ + + ]: 11591414 : if (ItemIdIsNormal(lp))
2368 : : {
2369 : 11588151 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2370 : :
2371 : : /*
2372 : : * Check if this tuple is part of a HOT-chain rooted at some other
2373 : : * tuple. If so, skip it for now; we'll process it when we find
2374 : : * its root.
2375 : : */
2376 [ + + ]: 11588151 : if (HeapTupleHeaderIsHeapOnly(htup))
2377 : 3579 : continue;
2378 : :
2379 : : /*
2380 : : * This is either a plain tuple or the root of a HOT-chain.
2381 : : * Remember it in the mapping.
2382 : : */
2383 : 11584572 : root_offsets[offnum - 1] = offnum;
2384 : :
2385 : : /* If it's not the start of a HOT-chain, we're done with it */
2386 [ + + ]: 11584572 : if (!HeapTupleHeaderIsHotUpdated(htup))
2387 : 11584320 : continue;
2388 : :
2389 : : /* Set up to scan the HOT-chain */
2390 : 252 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2391 : 252 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2392 : : }
2393 : : else
2394 : : {
2395 : : /* Must be a redirect item. We do not set its root_offsets entry */
2396 : : Assert(ItemIdIsRedirected(lp));
2397 : : /* Set up to scan the HOT-chain */
2398 : 3263 : nextoffnum = ItemIdGetRedirect(lp);
2399 : 3263 : priorXmax = InvalidTransactionId;
2400 : : }
2401 : :
2402 : : /*
2403 : : * Now follow the HOT-chain and collect other tuples in the chain.
2404 : : *
2405 : : * Note: Even though this is a nested loop, the complexity of the
2406 : : * function is O(N) because a tuple in the page should be visited not
2407 : : * more than twice, once in the outer loop and once in HOT-chain
2408 : : * chases.
2409 : : */
2410 : : for (;;)
2411 : : {
2412 : : /* Sanity check (pure paranoia) */
2413 [ - + ]: 3575 : if (nextoffnum < FirstOffsetNumber)
2414 : 0 : break;
2415 : :
2416 : : /*
2417 : : * An offset past the end of page's line pointer array is possible
2418 : : * when the array was truncated
2419 : : */
2420 [ - + ]: 3575 : if (nextoffnum > maxoff)
2421 : 0 : break;
2422 : :
2423 : 3575 : lp = PageGetItemId(page, nextoffnum);
2424 : :
2425 : : /* Check for broken chains */
2426 [ - + ]: 3575 : if (!ItemIdIsNormal(lp))
2427 : 0 : break;
2428 : :
2429 : 3575 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2430 : :
2431 [ + + - + ]: 3887 : if (TransactionIdIsValid(priorXmax) &&
2432 : 312 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(htup)))
2433 : 0 : break;
2434 : :
2435 : : /* Remember the root line pointer for this item */
2436 : 3575 : root_offsets[nextoffnum - 1] = offnum;
2437 : :
2438 : : /* Advance to next chain member, if any */
2439 [ + + ]: 3575 : if (!HeapTupleHeaderIsHotUpdated(htup))
2440 : 3515 : break;
2441 : :
2442 : : /* HOT implies it can't have moved to different partition */
2443 : : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
2444 : :
2445 : 60 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
2446 : 60 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2447 : : }
2448 : : }
2449 : 139800 : }
2450 : :
2451 : :
2452 : : /*
2453 : : * Compare fields that describe actions required to freeze tuple with caller's
2454 : : * open plan. If everything matches then the frz tuple plan is equivalent to
2455 : : * caller's plan.
2456 : : */
2457 : : static inline bool
2458 : 1360855 : heap_log_freeze_eq(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2459 : : {
2460 [ + + ]: 1360855 : if (plan->xmax == frz->xmax &&
2461 [ + + ]: 1359556 : plan->t_infomask2 == frz->t_infomask2 &&
2462 [ + + ]: 1358516 : plan->t_infomask == frz->t_infomask &&
2463 [ + - ]: 1355166 : plan->frzflags == frz->frzflags)
2464 : 1355166 : return true;
2465 : :
2466 : : /* Caller must call heap_log_freeze_new_plan again for frz */
2467 : 5689 : return false;
2468 : : }
2469 : :
2470 : : /*
2471 : : * Comparator used to deduplicate the freeze plans used in WAL records.
2472 : : */
2473 : : static int
2474 : 1840128 : heap_log_freeze_cmp(const void *arg1, const void *arg2)
2475 : : {
2476 : 1840128 : const HeapTupleFreeze *frz1 = arg1;
2477 : 1840128 : const HeapTupleFreeze *frz2 = arg2;
2478 : :
2479 [ + + ]: 1840128 : if (frz1->xmax < frz2->xmax)
2480 : 13214 : return -1;
2481 [ + + ]: 1826914 : else if (frz1->xmax > frz2->xmax)
2482 : 14031 : return 1;
2483 : :
2484 [ + + ]: 1812883 : if (frz1->t_infomask2 < frz2->t_infomask2)
2485 : 6118 : return -1;
2486 [ + + ]: 1806765 : else if (frz1->t_infomask2 > frz2->t_infomask2)
2487 : 6130 : return 1;
2488 : :
2489 [ + + ]: 1800635 : if (frz1->t_infomask < frz2->t_infomask)
2490 : 12855 : return -1;
2491 [ + + ]: 1787780 : else if (frz1->t_infomask > frz2->t_infomask)
2492 : 23456 : return 1;
2493 : :
2494 [ - + ]: 1764324 : if (frz1->frzflags < frz2->frzflags)
2495 : 0 : return -1;
2496 [ - + ]: 1764324 : else if (frz1->frzflags > frz2->frzflags)
2497 : 0 : return 1;
2498 : :
2499 : : /*
2500 : : * heap_log_freeze_eq would consider these tuple-wise plans to be equal.
2501 : : * (So the tuples will share a single canonical freeze plan.)
2502 : : *
2503 : : * We tiebreak on page offset number to keep each freeze plan's page
2504 : : * offset number array individually sorted. (Unnecessary, but be tidy.)
2505 : : */
2506 [ + + ]: 1764324 : if (frz1->offset < frz2->offset)
2507 : 1515525 : return -1;
2508 [ + - ]: 248799 : else if (frz1->offset > frz2->offset)
2509 : 248799 : return 1;
2510 : :
2511 : : Assert(false);
2512 : 0 : return 0;
2513 : : }
2514 : :
2515 : : /*
2516 : : * Start new plan initialized using tuple-level actions. At least one tuple
2517 : : * will have steps required to freeze described by caller's plan during REDO.
2518 : : */
2519 : : static inline void
2520 : 33314 : heap_log_freeze_new_plan(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2521 : : {
2522 : 33314 : plan->xmax = frz->xmax;
2523 : 33314 : plan->t_infomask2 = frz->t_infomask2;
2524 : 33314 : plan->t_infomask = frz->t_infomask;
2525 : 33314 : plan->frzflags = frz->frzflags;
2526 : 33314 : plan->ntuples = 1; /* for now */
2527 : 33314 : }
2528 : :
2529 : : /*
2530 : : * Deduplicate tuple-based freeze plans so that each distinct set of
2531 : : * processing steps is only stored once in the WAL record.
2532 : : * Called during original execution of freezing (for logged relations).
2533 : : *
2534 : : * Return value is number of plans set in *plans_out for caller. Also writes
2535 : : * an array of offset numbers into *offsets_out output argument for caller
2536 : : * (actually there is one array per freeze plan, but that's not of immediate
2537 : : * concern to our caller).
2538 : : */
2539 : : static int
2540 : 27625 : heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
2541 : : xlhp_freeze_plan *plans_out,
2542 : : OffsetNumber *offsets_out)
2543 : : {
2544 : 27625 : int nplans = 0;
2545 : :
2546 : : /* Sort tuple-based freeze plans in the order required to deduplicate */
2547 : 27625 : qsort(tuples, ntuples, sizeof(HeapTupleFreeze), heap_log_freeze_cmp);
2548 : :
2549 [ + + ]: 1416105 : for (int i = 0; i < ntuples; i++)
2550 : : {
2551 : 1388480 : HeapTupleFreeze *frz = tuples + i;
2552 : :
2553 [ + + ]: 1388480 : if (i == 0)
2554 : : {
2555 : : /* New canonical freeze plan starting with first tup */
2556 : 27625 : heap_log_freeze_new_plan(plans_out, frz);
2557 : 27625 : nplans++;
2558 : : }
2559 [ + + ]: 1360855 : else if (heap_log_freeze_eq(plans_out, frz))
2560 : : {
2561 : : /* tup matches open canonical plan -- include tup in it */
2562 : : Assert(offsets_out[i - 1] < frz->offset);
2563 : 1355166 : plans_out->ntuples++;
2564 : : }
2565 : : else
2566 : : {
2567 : : /* Tup doesn't match current plan -- done with it now */
2568 : 5689 : plans_out++;
2569 : :
2570 : : /* New canonical freeze plan starting with this tup */
2571 : 5689 : heap_log_freeze_new_plan(plans_out, frz);
2572 : 5689 : nplans++;
2573 : : }
2574 : :
2575 : : /*
2576 : : * Save page offset number in dedicated buffer in passing.
2577 : : *
2578 : : * REDO routine relies on the record's offset numbers array grouping
2579 : : * offset numbers by freeze plan. The sort order within each grouping
2580 : : * is ascending offset number order, just to keep things tidy.
2581 : : */
2582 : 1388480 : offsets_out[i] = frz->offset;
2583 : : }
2584 : :
2585 : : Assert(nplans > 0 && nplans <= ntuples);
2586 : :
2587 : 27625 : return nplans;
2588 : : }
2589 : :
2590 : : /*
2591 : : * Write an XLOG_HEAP2_PRUNE* WAL record
2592 : : *
2593 : : * This is used for several different page maintenance operations:
2594 : : *
2595 : : * - Page pruning, in VACUUM's 1st pass or on access: Some items are
2596 : : * redirected, some marked dead, and some removed altogether.
2597 : : *
2598 : : * - Freezing: Items are marked as 'frozen'.
2599 : : *
2600 : : * - Vacuum, 2nd pass: Items that are already LP_DEAD are marked as unused.
2601 : : *
2602 : : * They have enough commonalities that we use a single WAL record for them
2603 : : * all.
2604 : : *
2605 : : * If replaying the record requires a cleanup lock, pass cleanup_lock = true.
2606 : : * Replaying 'redirected' or 'dead' items always requires a cleanup lock, but
2607 : : * replaying 'unused' items depends on whether they were all previously marked
2608 : : * as dead.
2609 : : *
2610 : : * If the VM is being updated, vmflags will contain the bits to set. In this
2611 : : * case, vmbuffer should already have been updated and marked dirty and should
2612 : : * still be pinned and locked.
2613 : : *
2614 : : * Note: This function scribbles on the 'frozen' array.
2615 : : *
2616 : : * Note: This is called in a critical section, so careful what you do here.
2617 : : */
2618 : : void
2619 : 152395 : log_heap_prune_and_freeze(Relation relation, Buffer buffer,
2620 : : Buffer vmbuffer, uint8 vmflags,
2621 : : TransactionId conflict_xid,
2622 : : bool cleanup_lock,
2623 : : PruneReason reason,
2624 : : HeapTupleFreeze *frozen, int nfrozen,
2625 : : OffsetNumber *redirected, int nredirected,
2626 : : OffsetNumber *dead, int ndead,
2627 : : OffsetNumber *unused, int nunused)
2628 : : {
2629 : : xl_heap_prune xlrec;
2630 : : XLogRecPtr recptr;
2631 : : uint8 info;
2632 : : uint8 regbuf_flags_heap;
2633 : :
2634 : 152395 : Page heap_page = BufferGetPage(buffer);
2635 : :
2636 : : /* The following local variables hold data registered in the WAL record: */
2637 : : xlhp_freeze_plan plans[MaxHeapTuplesPerPage];
2638 : : xlhp_freeze_plans freeze_plans;
2639 : : xlhp_prune_items redirect_items;
2640 : : xlhp_prune_items dead_items;
2641 : : xlhp_prune_items unused_items;
2642 : : OffsetNumber frz_offsets[MaxHeapTuplesPerPage];
2643 [ + + + + : 152395 : bool do_prune = nredirected > 0 || ndead > 0 || nunused > 0;
+ + ]
2644 : 152395 : bool do_set_vm = vmflags & VISIBILITYMAP_VALID_BITS;
2645 : 152395 : bool heap_fpi_allowed = true;
2646 : :
2647 : : Assert((vmflags & VISIBILITYMAP_VALID_BITS) == vmflags);
2648 : :
2649 : 152395 : xlrec.flags = 0;
2650 : 152395 : regbuf_flags_heap = REGBUF_STANDARD;
2651 : :
2652 : : /*
2653 : : * We can avoid an FPI of the heap page if the only modification we are
2654 : : * making to it is to set PD_ALL_VISIBLE and checksums/wal_log_hints are
2655 : : * disabled.
2656 : : *
2657 : : * However, if the page has never been WAL-logged (LSN is invalid), we
2658 : : * must force an FPI regardless. This can happen when another backend
2659 : : * extends the heap, initializes the page, and then fails before WAL-
2660 : : * logging it. Since heap extension is not WAL-logged, recovery might try
2661 : : * to replay our record and find that the page isn't initialized, which
2662 : : * would cause a PANIC.
2663 : : */
2664 [ - + ]: 152395 : if (!XLogRecPtrIsValid(PageGetLSN(heap_page)))
2665 : 0 : regbuf_flags_heap |= REGBUF_FORCE_IMAGE;
2666 [ + + + + : 152395 : else if (!do_prune && nfrozen == 0 && (!do_set_vm || !XLogHintBitIsNeeded()))
+ - + + +
+ ]
2667 : : {
2668 : 2852 : regbuf_flags_heap |= REGBUF_NO_IMAGE;
2669 : 2852 : heap_fpi_allowed = false;
2670 : : }
2671 : :
2672 : : /*
2673 : : * Prepare data for the buffer. The arrays are not actually in the
2674 : : * buffer, but we pretend that they are. When XLogInsert stores a full
2675 : : * page image, the arrays can be omitted.
2676 : : */
2677 : 152395 : XLogBeginInsert();
2678 : 152395 : XLogRegisterBuffer(0, buffer, regbuf_flags_heap);
2679 : :
2680 [ + + ]: 152395 : if (do_set_vm)
2681 : 79961 : XLogRegisterBuffer(1, vmbuffer, 0);
2682 : :
2683 [ + + ]: 152395 : if (nfrozen > 0)
2684 : : {
2685 : : int nplans;
2686 : :
2687 : 27625 : xlrec.flags |= XLHP_HAS_FREEZE_PLANS;
2688 : :
2689 : : /*
2690 : : * Prepare deduplicated representation for use in the WAL record. This
2691 : : * destructively sorts frozen tuples array in-place.
2692 : : */
2693 : 27625 : nplans = heap_log_freeze_plan(frozen, nfrozen, plans, frz_offsets);
2694 : :
2695 : 27625 : freeze_plans.nplans = nplans;
2696 : 27625 : XLogRegisterBufData(0, &freeze_plans,
2697 : : offsetof(xlhp_freeze_plans, plans));
2698 : 27625 : XLogRegisterBufData(0, plans,
2699 : : sizeof(xlhp_freeze_plan) * nplans);
2700 : : }
2701 [ + + ]: 152395 : if (nredirected > 0)
2702 : : {
2703 : 17824 : xlrec.flags |= XLHP_HAS_REDIRECTIONS;
2704 : :
2705 : 17824 : redirect_items.ntargets = nredirected;
2706 : 17824 : XLogRegisterBufData(0, &redirect_items,
2707 : : offsetof(xlhp_prune_items, data));
2708 : 17824 : XLogRegisterBufData(0, redirected,
2709 : : sizeof(OffsetNumber[2]) * nredirected);
2710 : : }
2711 [ + + ]: 152395 : if (ndead > 0)
2712 : : {
2713 : 57848 : xlrec.flags |= XLHP_HAS_DEAD_ITEMS;
2714 : :
2715 : 57848 : dead_items.ntargets = ndead;
2716 : 57848 : XLogRegisterBufData(0, &dead_items,
2717 : : offsetof(xlhp_prune_items, data));
2718 : 57848 : XLogRegisterBufData(0, dead,
2719 : : sizeof(OffsetNumber) * ndead);
2720 : : }
2721 [ + + ]: 152395 : if (nunused > 0)
2722 : : {
2723 : 31326 : xlrec.flags |= XLHP_HAS_NOW_UNUSED_ITEMS;
2724 : :
2725 : 31326 : unused_items.ntargets = nunused;
2726 : 31326 : XLogRegisterBufData(0, &unused_items,
2727 : : offsetof(xlhp_prune_items, data));
2728 : 31326 : XLogRegisterBufData(0, unused,
2729 : : sizeof(OffsetNumber) * nunused);
2730 : : }
2731 [ + + ]: 152395 : if (nfrozen > 0)
2732 : 27625 : XLogRegisterBufData(0, frz_offsets,
2733 : : sizeof(OffsetNumber) * nfrozen);
2734 : :
2735 : : /*
2736 : : * Prepare the main xl_heap_prune record. We already set the XLHP_HAS_*
2737 : : * flag above.
2738 : : */
2739 [ + + ]: 152395 : if (vmflags & VISIBILITYMAP_ALL_VISIBLE)
2740 : : {
2741 : 79961 : xlrec.flags |= XLHP_VM_ALL_VISIBLE;
2742 [ + + ]: 79961 : if (vmflags & VISIBILITYMAP_ALL_FROZEN)
2743 : 47103 : xlrec.flags |= XLHP_VM_ALL_FROZEN;
2744 : : }
2745 [ + + + + : 152395 : if (RelationIsAccessibleInLogicalDecoding(relation))
+ - - + -
- - - + +
+ + - + -
- + - ]
2746 : 666 : xlrec.flags |= XLHP_IS_CATALOG_REL;
2747 [ + + ]: 152395 : if (TransactionIdIsValid(conflict_xid))
2748 : 121554 : xlrec.flags |= XLHP_HAS_CONFLICT_HORIZON;
2749 [ + + ]: 152395 : if (cleanup_lock)
2750 : 70525 : xlrec.flags |= XLHP_CLEANUP_LOCK;
2751 : : else
2752 : : {
2753 : : Assert(nredirected == 0 && ndead == 0);
2754 : : /* also, any items in 'unused' must've been LP_DEAD previously */
2755 : : }
2756 : 152395 : XLogRegisterData(&xlrec, SizeOfHeapPrune);
2757 [ + + ]: 152395 : if (TransactionIdIsValid(conflict_xid))
2758 : 121554 : XLogRegisterData(&conflict_xid, sizeof(TransactionId));
2759 : :
2760 [ + + + - ]: 152395 : switch (reason)
2761 : : {
2762 : 73098 : case PRUNE_ON_ACCESS:
2763 : 73098 : info = XLOG_HEAP2_PRUNE_ON_ACCESS;
2764 : 73098 : break;
2765 : 62315 : case PRUNE_VACUUM_SCAN:
2766 : 62315 : info = XLOG_HEAP2_PRUNE_VACUUM_SCAN;
2767 : 62315 : break;
2768 : 16982 : case PRUNE_VACUUM_CLEANUP:
2769 : 16982 : info = XLOG_HEAP2_PRUNE_VACUUM_CLEANUP;
2770 : 16982 : break;
2771 : 0 : default:
2772 [ # # ]: 0 : elog(ERROR, "unrecognized prune reason: %d", (int) reason);
2773 : : break;
2774 : : }
2775 : 152395 : recptr = XLogInsert(RM_HEAP2_ID, info);
2776 : :
2777 [ + + ]: 152395 : if (do_set_vm)
2778 : : {
2779 : : Assert(BufferIsDirty(vmbuffer));
2780 : 79961 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2781 : : }
2782 : :
2783 : : /*
2784 : : * If we explicitly skip an FPI, we must not stamp the heap page with this
2785 : : * record's LSN. Recovery skips records <= the stamped LSN, so this could
2786 : : * lead to skipping an earlier FPI needed to repair a torn page.
2787 : : */
2788 [ + + ]: 152395 : if (heap_fpi_allowed)
2789 : : {
2790 : : Assert(BufferIsDirty(buffer));
2791 : 149543 : PageSetLSN(heap_page, recptr);
2792 : : }
2793 : 152395 : }
|
