Age Owner Branch data TLA 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
118 melanieplageman@gmai 272 :CBC 19452254 : heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer,
273 : : bool rel_read_only)
274 : : {
3749 kgrittn@postgresql.o 275 : 19452254 : 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 : : */
4528 rhaas@postgresql.org 285 [ + + ]: 19452254 : if (RecoveryInProgress())
286 : 262489 : 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 : : */
143 melanieplageman@gmai 294 : 19189765 : prune_xid = PageGetPruneXid(page);
2174 andres@anarazel.de 295 [ + + ]: 19189765 : if (!TransactionIdIsValid(prune_xid))
296 : 7662032 : return;
297 : :
298 : : /*
299 : : * Check whether prune_xid indicates that there may be dead rows that can
300 : : * be cleaned up.
301 : : */
302 : 11527733 : vistest = GlobalVisTestFor(relation);
303 : :
124 melanieplageman@gmai 304 [ + + ]: 11527733 : if (!GlobalVisTestIsRemovableXid(vistest, prune_xid, true))
1055 tmunro@postgresql.or 305 : 9245676 : 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 : : */
836 akorotkov@postgresql 319 [ + + ]: 2282057 : minfree = RelationGetTargetPageFreeSpace(relation,
320 : : HEAP_DEFAULT_FILLFACTOR);
6884 tgl@sss.pgh.pa.us 321 : 2282057 : minfree = Max(minfree, BLCKSZ / 10);
322 : :
6587 323 [ + + + + ]: 2282057 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
324 : : {
16 melanieplageman@gmai 325 : 133394 : bool record_free_space = false;
326 : 133394 : Size freespace = 0;
327 : :
328 : : /* OK, try to get exclusive buffer lock */
6884 tgl@sss.pgh.pa.us 329 [ + + ]: 133394 : if (!ConditionalLockBufferForCleanup(buffer))
330 : 2171 : return;
331 : :
332 : : /*
333 : : * Now that we have buffer lock, get accurate information about the
334 : : * page's free space, and recheck the heuristic about whether to
335 : : * prune.
336 : : */
6587 337 [ + + + - ]: 131223 : if (PageIsFull(page) || PageGetHeapFreeSpace(page) < minfree)
338 : : {
339 : : OffsetNumber dummy_off_loc;
340 : : PruneFreezeResult presult;
341 : : PruneFreezeParams params;
342 : :
126 melanieplageman@gmai 343 : 131223 : visibilitymap_pin(relation, BufferGetBlockNumber(buffer),
344 : : vmbuffer);
345 : :
346 : 131223 : params.relation = relation;
347 : 131223 : params.buffer = buffer;
348 : 131223 : params.vmbuffer = *vmbuffer;
349 : 131223 : params.reason = PRUNE_ON_ACCESS;
350 : 131223 : params.vistest = vistest;
351 : 131223 : params.cutoffs = NULL;
352 : :
353 : : /*
354 : : * We don't pass the HEAP_PAGE_PRUNE_MARK_UNUSED_NOW option
355 : : * regardless of whether or not the relation has indexes, since we
356 : : * cannot safely determine that during on-access pruning with the
357 : : * current implementation.
358 : : */
359 : 131223 : params.options = HEAP_PAGE_PRUNE_ALLOW_FAST_PATH;
118 360 [ + + ]: 131223 : if (rel_read_only)
361 : 45638 : params.options |= HEAP_PAGE_PRUNE_SET_VM;
362 : :
248 363 : 131223 : heap_page_prune_and_freeze(¶ms, &presult, &dummy_off_loc,
364 : : NULL, NULL);
365 : :
366 : : /*
367 : : * Report the number of tuples reclaimed to pgstats. This is
368 : : * presult.ndeleted minus the number of newly-LP_DEAD-set items.
369 : : *
370 : : * We derive the number of dead tuples like this to avoid totally
371 : : * forgetting about items that were set to LP_DEAD, since they
372 : : * still need to be cleaned up by VACUUM. We only want to count
373 : : * heap-only tuples that just became LP_UNUSED in our report,
374 : : * which don't.
375 : : *
376 : : * VACUUM doesn't have to compensate in the same way when it
377 : : * tracks ndeleted, since it will set the same LP_DEAD items to
378 : : * LP_UNUSED separately.
379 : : */
1032 rhaas@postgresql.org 380 [ + + ]: 131223 : if (presult.ndeleted > presult.nnewlpdead)
1717 pg@bowt.ie 381 : 21757 : pgstat_update_heap_dead_tuples(relation,
1032 rhaas@postgresql.org 382 : 21757 : presult.ndeleted - presult.nnewlpdead);
383 : :
384 : : /*
385 : : * If this prune newly set the page all-visible, VACUUM may later
386 : : * skip the page and not update the free space map (FSM) for it.
387 : : * Keep the FSM from going stale by recording it now. We do not
388 : : * want to update the freespace map otherwise, to reserve
389 : : * freespace on this page for HOT updates.
390 : : */
16 melanieplageman@gmai 391 [ + + ]: 131223 : if (presult.newly_all_visible)
392 : : {
393 : 13559 : record_free_space = true;
394 : 13559 : freespace = PageGetHeapFreeSpace(page);
395 : : }
396 : : }
397 : :
398 : : /* And release buffer lock */
6884 tgl@sss.pgh.pa.us 399 : 131223 : LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
400 : :
401 : : /*
402 : : * RecordPageWithFreeSpace() only dirties the FSM when the recorded
403 : : * free-space category actually changes. Note that vacuum will still
404 : : * do FreeSpaceMapVacuum() for ranges of pages that are skipped, so we
405 : : * don't have to worry about that here.
406 : : */
16 melanieplageman@gmai 407 [ + + ]: 131223 : if (record_free_space)
408 : 13559 : RecordPageWithFreeSpace(relation, BufferGetBlockNumber(buffer), freespace);
409 : : }
410 : : }
411 : :
412 : : /*
413 : : * Helper for heap_page_prune_and_freeze() to initialize the PruneState using
414 : : * the provided parameters.
415 : : *
416 : : * params, new_relfrozen_xid, new_relmin_mxid, and presult are input
417 : : * parameters and are not modified by this function. Only prstate is modified.
418 : : */
419 : : static void
242 420 : 238548 : prune_freeze_setup(PruneFreezeParams *params,
421 : : TransactionId *new_relfrozen_xid,
422 : : MultiXactId *new_relmin_mxid,
423 : : PruneFreezeResult *presult,
424 : : PruneState *prstate)
425 : : {
426 : : /* Copy parameters to prstate */
427 : 238548 : prstate->vistest = params->vistest;
428 : 238548 : prstate->mark_unused_now =
248 429 : 238548 : (params->options & HEAP_PAGE_PRUNE_MARK_UNUSED_NOW) != 0;
430 : :
431 : : /* cutoffs must be provided if we will attempt freezing */
243 432 [ + + - + ]: 238548 : Assert(!(params->options & HEAP_PAGE_PRUNE_FREEZE) || params->cutoffs);
242 433 : 238548 : prstate->attempt_freeze = (params->options & HEAP_PAGE_PRUNE_FREEZE) != 0;
118 434 : 238548 : prstate->attempt_set_vm = (params->options & HEAP_PAGE_PRUNE_SET_VM) != 0;
242 435 : 238548 : prstate->cutoffs = params->cutoffs;
143 436 : 238548 : prstate->relation = params->relation;
437 : 238548 : prstate->block = BufferGetBlockNumber(params->buffer);
438 : 238548 : prstate->buffer = params->buffer;
439 : 238548 : prstate->page = BufferGetPage(params->buffer);
440 : :
126 441 [ - + ]: 238548 : Assert(BufferIsValid(params->vmbuffer));
442 : 238548 : prstate->vmbuffer = params->vmbuffer;
124 443 : 238548 : prstate->new_vmbits = 0;
126 444 : 238548 : prstate->old_vmbits = visibilitymap_get_status(prstate->relation,
445 : : prstate->block,
446 : : &prstate->vmbuffer);
447 : :
448 : : /*
449 : : * Our strategy is to scan the page and make lists of items to change,
450 : : * then apply the changes within a critical section. This keeps as much
451 : : * logic as possible out of the critical section, and also ensures that
452 : : * WAL replay will work the same as the normal case.
453 : : *
454 : : * First, initialize the new pd_prune_xid value to zero (indicating no
455 : : * prunable tuples). If we find any tuples which may soon become
456 : : * prunable, we will save the lowest relevant XID in new_prune_xid. Also
457 : : * initialize the rest of our working state.
458 : : */
242 459 : 238548 : prstate->new_prune_xid = InvalidTransactionId;
460 : 238548 : prstate->latest_xid_removed = InvalidTransactionId;
461 : 238548 : prstate->nredirected = prstate->ndead = prstate->nunused = 0;
462 : 238548 : prstate->nfrozen = 0;
463 : 238548 : prstate->nroot_items = 0;
464 : 238548 : prstate->nheaponly_items = 0;
465 : :
466 : : /* initialize page freezing working state */
467 : 238548 : prstate->pagefrz.freeze_required = false;
138 468 : 238548 : prstate->pagefrz.FreezePageConflictXid = InvalidTransactionId;
242 469 [ + + ]: 238548 : if (prstate->attempt_freeze)
470 : : {
234 471 [ + - - + ]: 107325 : Assert(new_relfrozen_xid && new_relmin_mxid);
472 : 107325 : prstate->pagefrz.FreezePageRelfrozenXid = *new_relfrozen_xid;
473 : 107325 : prstate->pagefrz.NoFreezePageRelfrozenXid = *new_relfrozen_xid;
474 : 107325 : prstate->pagefrz.FreezePageRelminMxid = *new_relmin_mxid;
475 : 107325 : prstate->pagefrz.NoFreezePageRelminMxid = *new_relmin_mxid;
476 : : }
477 : : else
478 : : {
479 [ + - - + ]: 131223 : Assert(!new_relfrozen_xid && !new_relmin_mxid);
242 480 : 131223 : prstate->pagefrz.FreezePageRelminMxid = InvalidMultiXactId;
481 : 131223 : prstate->pagefrz.NoFreezePageRelminMxid = InvalidMultiXactId;
482 : 131223 : prstate->pagefrz.FreezePageRelfrozenXid = InvalidTransactionId;
483 : 131223 : prstate->pagefrz.NoFreezePageRelfrozenXid = InvalidTransactionId;
484 : : }
485 : :
486 : 238548 : prstate->ndeleted = 0;
487 : 238548 : prstate->live_tuples = 0;
488 : 238548 : prstate->recently_dead_tuples = 0;
489 : 238548 : prstate->hastup = false;
490 : 238548 : prstate->lpdead_items = 0;
491 : :
492 : : /*
493 : : * deadoffsets are filled in during pruning but are only used to populate
494 : : * PruneFreezeResult->deadoffsets. To avoid needing two copies of the
495 : : * array, just save a pointer to the result offsets array in the
496 : : * PruneState.
497 : : */
222 498 : 238548 : prstate->deadoffsets = presult->deadoffsets;
499 : :
500 : : /*
501 : : * We track whether the page will be all-visible/all-frozen at the end of
502 : : * pruning and freezing. While examining tuple visibility, we'll set
503 : : * set_all_visible to false if there are tuples on the page not visible to
504 : : * all running and future transactions. If setting the VM is enabled for
505 : : * this scan, we will do so if the page ends up being all-visible.
506 : : *
507 : : * We also keep track of the newest live XID, which is used to calculate
508 : : * the snapshot conflict horizon for a WAL record setting the VM.
509 : : */
118 510 : 238548 : prstate->set_all_visible = prstate->attempt_set_vm;
124 511 : 238548 : prstate->newest_live_xid = InvalidTransactionId;
512 : :
513 : : /*
514 : : * Currently, only VACUUM performs freezing, but other callers may in the
515 : : * future. We must initialize set_all_frozen based on whether or not the
516 : : * caller passed HEAP_PAGE_PRUNE_FREEZE, because if they did not, we won't
517 : : * call heap_prepare_freeze_tuple() for each tuple, and set_all_frozen
518 : : * will never be cleared for tuples that need freezing. This would lead to
519 : : * incorrectly setting the visibility map all-frozen for this page. We
520 : : * can't set the page all-frozen in the VM if the caller didn't pass
521 : : * HEAP_PAGE_PRUNE_SET_VM.
522 : : *
523 : : * When freezing is not required (no XIDs/MXIDs older than the freeze
524 : : * cutoff), we may still choose to "opportunistically" freeze if doing so
525 : : * would make the page all-frozen.
526 : : *
527 : : * We will not be able to freeze the whole page at the end of vacuum if
528 : : * there are tuples present that are not visible to everyone or if there
529 : : * are dead tuples which will not be removable. However, dead tuples that
530 : : * will be removed by the end of vacuum should not prevent this
531 : : * opportunistic freezing.
532 : : *
533 : : * Therefore, we do not clear set_all_visible and set_all_frozen when we
534 : : * encounter LP_DEAD items. Instead, we correct them after deciding
535 : : * whether to freeze, but before updating the VM, to avoid setting the VM
536 : : * bits incorrectly.
537 : : */
118 538 [ + + + - ]: 238548 : prstate->set_all_frozen = prstate->attempt_freeze && prstate->attempt_set_vm;
242 539 : 238548 : }
540 : :
541 : : /*
542 : : * Helper for heap_page_prune_and_freeze(). Iterates over every tuple on the
543 : : * page, examines its visibility information, and determines the appropriate
544 : : * action for each tuple. All tuples are processed and classified during this
545 : : * phase, but no modifications are made to the page until the later execution
546 : : * stage.
547 : : *
548 : : * *off_loc is used for error callback and cleared before returning.
549 : : */
550 : : static void
143 551 : 224834 : prune_freeze_plan(PruneState *prstate, OffsetNumber *off_loc)
552 : : {
553 : 224834 : Page page = prstate->page;
554 : 224834 : BlockNumber blockno = prstate->block;
555 : 224834 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
556 : : OffsetNumber offnum;
557 : : HeapTupleData tup;
558 : :
559 : 224834 : tup.t_tableOid = RelationGetRelid(prstate->relation);
560 : :
561 : : /*
562 : : * Determine HTSV for all tuples, and queue them up for processing as HOT
563 : : * chain roots or as heap-only items.
564 : : *
565 : : * Determining HTSV only once for each tuple is required for correctness,
566 : : * to deal with cases where running HTSV twice could result in different
567 : : * results. For example, RECENTLY_DEAD can turn to DEAD if another
568 : : * checked item causes GlobalVisTestIsRemovableFullXid() to update the
569 : : * horizon, or INSERT_IN_PROGRESS can change to DEAD if the inserting
570 : : * transaction aborts.
571 : : *
572 : : * It's also good for performance. Most commonly tuples within a page are
573 : : * stored at decreasing offsets (while the items are stored at increasing
574 : : * offsets). When processing all tuples on a page this leads to reading
575 : : * memory at decreasing offsets within a page, with a variable stride.
576 : : * That's hard for CPU prefetchers to deal with. Processing the items in
577 : : * reverse order (and thus the tuples in increasing order) increases
578 : : * prefetching efficiency significantly / decreases the number of cache
579 : : * misses.
580 : : */
1689 andres@anarazel.de 581 : 224834 : for (offnum = maxoff;
582 [ + + ]: 24250808 : offnum >= FirstOffsetNumber;
583 : 24025974 : offnum = OffsetNumberPrev(offnum))
584 : : {
585 : 24025974 : ItemId itemid = PageGetItemId(page, offnum);
586 : : HeapTupleHeader htup;
587 : :
588 : : /*
589 : : * Set the offset number so that we can display it along with any
590 : : * error that occurred while processing this tuple.
591 : : */
846 heikki.linnakangas@i 592 : 24025974 : *off_loc = offnum;
593 : :
242 melanieplageman@gmai 594 : 24025974 : prstate->processed[offnum] = false;
595 : 24025974 : prstate->htsv[offnum] = -1;
596 : :
597 : : /* Nothing to do if slot doesn't contain a tuple */
846 heikki.linnakangas@i 598 [ + + ]: 24025974 : if (!ItemIdIsUsed(itemid))
599 : : {
143 melanieplageman@gmai 600 : 376212 : heap_prune_record_unchanged_lp_unused(prstate, offnum);
1689 andres@anarazel.de 601 : 376212 : continue;
602 : : }
603 : :
846 heikki.linnakangas@i 604 [ + + ]: 23649762 : if (ItemIdIsDead(itemid))
605 : : {
606 : : /*
607 : : * If the caller set mark_unused_now true, we can set dead line
608 : : * pointers LP_UNUSED now.
609 : : */
242 melanieplageman@gmai 610 [ + + ]: 2605412 : if (unlikely(prstate->mark_unused_now))
611 : 1001690 : heap_prune_record_unused(prstate, offnum, false);
612 : : else
143 613 : 1603722 : heap_prune_record_unchanged_lp_dead(prstate, offnum);
846 heikki.linnakangas@i 614 : 2605412 : continue;
615 : : }
616 : :
617 [ + + ]: 21044350 : if (ItemIdIsRedirected(itemid))
618 : : {
619 : : /* This is the start of a HOT chain */
242 melanieplageman@gmai 620 : 226039 : prstate->root_items[prstate->nroot_items++] = offnum;
846 heikki.linnakangas@i 621 : 226039 : continue;
622 : : }
623 : :
624 [ - + ]: 20818311 : Assert(ItemIdIsNormal(itemid));
625 : :
626 : : /*
627 : : * Get the tuple's visibility status and queue it up for processing.
628 : : */
629 : 20818311 : htup = (HeapTupleHeader) PageGetItem(page, itemid);
630 : 20818311 : tup.t_data = htup;
631 : 20818311 : tup.t_len = ItemIdGetLength(itemid);
632 : 20818311 : ItemPointerSet(&tup.t_self, blockno, offnum);
633 : :
143 melanieplageman@gmai 634 : 20818311 : prstate->htsv[offnum] = heap_prune_satisfies_vacuum(prstate, &tup);
635 : :
846 heikki.linnakangas@i 636 [ + + ]: 20818311 : if (!HeapTupleHeaderIsHeapOnly(htup))
242 melanieplageman@gmai 637 : 20440853 : prstate->root_items[prstate->nroot_items++] = offnum;
638 : : else
639 : 377458 : prstate->heaponly_items[prstate->nheaponly_items++] = offnum;
640 : : }
641 : :
642 : : /*
643 : : * Process HOT chains.
644 : : *
645 : : * We added the items to the array starting from 'maxoff', so by
646 : : * processing the array in reverse order, we process the items in
647 : : * ascending offset number order. The order doesn't matter for
648 : : * correctness, but some quick micro-benchmarking suggests that this is
649 : : * faster. (Earlier PostgreSQL versions, which scanned all the items on
650 : : * the page instead of using the root_items array, also did it in
651 : : * ascending offset number order.)
652 : : */
653 [ + + ]: 20891726 : for (int i = prstate->nroot_items - 1; i >= 0; i--)
654 : : {
655 : 20666892 : offnum = prstate->root_items[i];
656 : :
657 : : /* Ignore items already processed as part of an earlier chain */
658 [ - + ]: 20666892 : if (prstate->processed[offnum])
6714 tgl@sss.pgh.pa.us 659 :UBC 0 : continue;
660 : :
661 : : /* see preceding loop */
846 heikki.linnakangas@i 662 :CBC 20666892 : *off_loc = offnum;
663 : :
664 : : /* Process this item or chain of items */
143 melanieplageman@gmai 665 : 20666892 : heap_prune_chain(maxoff, offnum, prstate);
666 : : }
667 : :
668 : : /*
669 : : * Process any heap-only tuples that were not already processed as part of
670 : : * a HOT chain.
671 : : */
242 672 [ + + ]: 602292 : for (int i = prstate->nheaponly_items - 1; i >= 0; i--)
673 : : {
674 : 377458 : offnum = prstate->heaponly_items[i];
675 : :
676 [ + + ]: 377458 : if (prstate->processed[offnum])
846 heikki.linnakangas@i 677 : 360823 : continue;
678 : :
679 : : /* see preceding loop */
680 : 16635 : *off_loc = offnum;
681 : :
682 : : /*
683 : : * If the tuple is DEAD and doesn't chain to anything else, mark it
684 : : * unused. (If it does chain, we can only remove it as part of
685 : : * pruning its chain.)
686 : : *
687 : : * We need this primarily to handle aborted HOT updates, that is,
688 : : * XMIN_INVALID heap-only tuples. Those might not be linked to by any
689 : : * chain, since the parent tuple might be re-updated before any
690 : : * pruning occurs. So we have to be able to reap them separately from
691 : : * chain-pruning. (Note that HeapTupleHeaderIsHotUpdated will never
692 : : * return true for an XMIN_INVALID tuple, so this code will work even
693 : : * when there were sequential updates within the aborted transaction.)
694 : : */
242 melanieplageman@gmai 695 [ + + ]: 16635 : if (prstate->htsv[offnum] == HEAPTUPLE_DEAD)
696 : : {
846 heikki.linnakangas@i 697 : 3388 : ItemId itemid = PageGetItemId(page, offnum);
698 : 3388 : HeapTupleHeader htup = (HeapTupleHeader) PageGetItem(page, itemid);
699 : :
700 [ + - ]: 3388 : if (likely(!HeapTupleHeaderIsHotUpdated(htup)))
701 : : {
702 : 3388 : HeapTupleHeaderAdvanceConflictHorizon(htup,
703 : : &prstate->latest_xid_removed);
242 melanieplageman@gmai 704 : 3388 : heap_prune_record_unused(prstate, offnum, true);
705 : : }
706 : : else
707 : : {
708 : : /*
709 : : * This tuple should've been processed and removed as part of
710 : : * a HOT chain, so something's wrong. To preserve evidence,
711 : : * we don't dare to remove it. We cannot leave behind a DEAD
712 : : * tuple either, because that will cause VACUUM to error out.
713 : : * Throwing an error with a distinct error message seems like
714 : : * the least bad option.
715 : : */
846 heikki.linnakangas@i 716 [ # # ]:UBC 0 : elog(ERROR, "dead heap-only tuple (%u, %d) is not linked to from any HOT chain",
717 : : blockno, offnum);
718 : : }
719 : : }
720 : : else
143 melanieplageman@gmai 721 :CBC 13247 : heap_prune_record_unchanged_lp_normal(prstate, offnum);
722 : : }
723 : :
724 : : /* We should now have processed every tuple exactly once */
725 : : #ifdef USE_ASSERT_CHECKING
846 heikki.linnakangas@i 726 : 224834 : for (offnum = FirstOffsetNumber;
727 [ + + ]: 24250808 : offnum <= maxoff;
728 : 24025974 : offnum = OffsetNumberNext(offnum))
729 : : {
730 : 24025974 : *off_loc = offnum;
731 : :
242 melanieplageman@gmai 732 [ - + ]: 24025974 : Assert(prstate->processed[offnum]);
733 : : }
734 : : #endif
735 : :
736 : : /* Clear the offset information once we have processed the given page. */
846 heikki.linnakangas@i 737 : 224834 : *off_loc = InvalidOffsetNumber;
242 melanieplageman@gmai 738 : 224834 : }
739 : :
740 : : /*
741 : : * Decide whether to proceed with freezing according to the freeze plans
742 : : * prepared for the current heap buffer. If freezing is chosen, this function
743 : : * performs several pre-freeze checks.
744 : : *
745 : : * The values of do_prune, do_hint_prune, and did_tuple_hint_fpi must be
746 : : * determined before calling this function.
747 : : *
748 : : * prstate is both an input and output parameter.
749 : : *
750 : : * Returns true if we should apply the freeze plans and freeze tuples on the
751 : : * page, and false otherwise.
752 : : */
753 : : static bool
143 754 : 224834 : heap_page_will_freeze(bool did_tuple_hint_fpi,
755 : : bool do_prune,
756 : : bool do_hint_prune,
757 : : PruneState *prstate)
758 : : {
242 759 : 224834 : bool do_freeze = false;
760 : :
761 : : /*
762 : : * If the caller specified we should not attempt to freeze any tuples,
763 : : * validate that everything is in the right state and return.
764 : : */
765 [ + + ]: 224834 : if (!prstate->attempt_freeze)
766 : : {
143 767 [ + - - + ]: 131125 : Assert(!prstate->set_all_frozen && prstate->nfrozen == 0);
242 768 : 131125 : return false;
769 : : }
770 : :
771 [ + + ]: 93709 : if (prstate->pagefrz.freeze_required)
772 : : {
773 : : /*
774 : : * heap_prepare_freeze_tuple indicated that at least one XID/MXID from
775 : : * before FreezeLimit/MultiXactCutoff is present. Must freeze to
776 : : * advance relfrozenxid/relminmxid.
777 : : */
778 : 20647 : do_freeze = true;
779 : : }
780 : : else
781 : : {
782 : : /*
783 : : * Opportunistically freeze the page if we are generating an FPI
784 : : * anyway and if doing so means that we can set the page all-frozen
785 : : * afterwards (might not happen until VACUUM's final heap pass).
786 : : *
787 : : * XXX: Previously, we knew if pruning emitted an FPI by checking
788 : : * pgWalUsage.wal_fpi before and after pruning. Once the freeze and
789 : : * prune records were combined, this heuristic couldn't be used
790 : : * anymore. The opportunistic freeze heuristic must be improved;
791 : : * however, for now, try to approximate the old logic.
792 : : */
143 793 [ + + + + ]: 73062 : if (prstate->set_all_frozen && prstate->nfrozen > 0)
794 : : {
795 [ - + ]: 28616 : Assert(prstate->set_all_visible);
796 : :
797 : : /*
798 : : * Freezing would make the page all-frozen. Have already emitted
799 : : * an FPI or will do so anyway?
800 : : */
801 [ + + + + : 28616 : if (RelationNeedsWAL(prstate->relation))
+ - + - ]
802 : : {
242 803 [ + + ]: 26955 : if (did_tuple_hint_fpi)
804 : 1611 : do_freeze = true;
805 [ + + ]: 25344 : else if (do_prune)
806 : : {
143 807 [ + + ]: 4232 : if (XLogCheckBufferNeedsBackup(prstate->buffer))
242 808 : 979 : do_freeze = true;
809 : : }
810 [ + + ]: 21112 : else if (do_hint_prune)
811 : : {
143 812 [ + + + - : 23972 : if (XLogHintBitIsNeeded() &&
+ + ]
813 : 11986 : XLogCheckBufferNeedsBackup(prstate->buffer))
242 814 : 1912 : do_freeze = true;
815 : : }
816 : : }
817 : : }
818 : : }
819 : :
820 [ + + ]: 93709 : if (do_freeze)
821 : : {
822 : : /*
823 : : * Validate the tuples we will be freezing before entering the
824 : : * critical section.
825 : : */
143 826 : 25149 : heap_pre_freeze_checks(prstate->buffer, prstate->frozen, prstate->nfrozen);
138 827 [ - + ]: 25149 : Assert(TransactionIdPrecedes(prstate->pagefrz.FreezePageConflictXid,
828 : : prstate->cutoffs->OldestXmin));
829 : : }
242 830 [ + + ]: 68560 : else if (prstate->nfrozen > 0)
831 : : {
832 : : /*
833 : : * The page contained some tuples that were not already frozen, and we
834 : : * chose not to freeze them now. The page won't be all-frozen then.
835 : : */
836 [ - + ]: 30442 : Assert(!prstate->pagefrz.freeze_required);
837 : :
143 838 : 30442 : prstate->set_all_frozen = false;
242 839 : 30442 : prstate->nfrozen = 0; /* avoid miscounts in instrumentation */
840 : : }
841 : : else
842 : : {
843 : : /*
844 : : * We have no freeze plans to execute. The page might already be
845 : : * all-frozen (perhaps only following pruning), though. Such pages
846 : : * can be marked all-frozen in the VM by our caller, even though none
847 : : * of its tuples were newly frozen here.
848 : : */
849 : : }
850 : :
851 : 93709 : return do_freeze;
852 : : }
853 : :
854 : : /*
855 : : * Emit a warning about and fix visibility map corruption on the given page.
856 : : *
857 : : * The caller specifies the type of corruption it has already detected via
858 : : * corruption_type, so that we can emit the appropriate warning. All cases
859 : : * result in the VM bits being cleared; corruption types where PD_ALL_VISIBLE
860 : : * is incorrectly set also clear PD_ALL_VISIBLE.
861 : : *
862 : : * Must be called while holding an exclusive lock on the heap buffer. Dead
863 : : * items and not all-visible tuples must have been discovered under that same
864 : : * lock. Although we do not hold a lock on the VM buffer, it is pinned, and
865 : : * the heap buffer is exclusively locked, ensuring that no other backend can
866 : : * update the VM bits corresponding to this heap page.
867 : : *
868 : : * This function makes changes to the VM and, potentially, the heap page, but
869 : : * it does not need to be done in a critical section.
870 : : */
871 : : static void
126 melanieplageman@gmai 872 :UBC 0 : heap_page_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum,
873 : : VMCorruptionType corruption_type)
874 : : {
875 : 0 : const char *relname = RelationGetRelationName(prstate->relation);
876 : 0 : bool do_clear_vm = false;
877 : 0 : bool do_clear_heap = false;
878 : :
879 [ # # ]: 0 : Assert(BufferIsLockedByMeInMode(prstate->buffer, BUFFER_LOCK_EXCLUSIVE));
880 : :
881 [ # # # # ]: 0 : switch (corruption_type)
882 : : {
883 : 0 : case VM_CORRUPT_LPDEAD:
884 [ # # ]: 0 : ereport(WARNING,
885 : : (errcode(ERRCODE_DATA_CORRUPTED),
886 : : errmsg("dead line pointer found on page marked all-visible"),
887 : : errcontext("relation \"%s\", page %u, tuple %u",
888 : : relname, prstate->block, offnum)));
889 : 0 : do_clear_vm = true;
890 : 0 : do_clear_heap = true;
891 : 0 : break;
892 : :
893 : 0 : case VM_CORRUPT_TUPLE_VISIBILITY:
894 : :
895 : : /*
896 : : * A HEAPTUPLE_LIVE tuple on an all-visible page can appear to not
897 : : * be visible to everyone when
898 : : * GetOldestNonRemovableTransactionId() returns a conservative
899 : : * value that's older than the real safe xmin. That is not
900 : : * corruption -- the PD_ALL_VISIBLE flag is still correct.
901 : : *
902 : : * However, dead tuple versions, in-progress inserts, and
903 : : * in-progress deletes should never appear on a page marked
904 : : * all-visible. That indicates real corruption. PD_ALL_VISIBLE
905 : : * should have been cleared by the DML operation that deleted or
906 : : * inserted the tuple.
907 : : */
908 [ # # ]: 0 : ereport(WARNING,
909 : : (errcode(ERRCODE_DATA_CORRUPTED),
910 : : errmsg("tuple not visible to all transactions found on page marked all-visible"),
911 : : errcontext("relation \"%s\", page %u, tuple %u",
912 : : relname, prstate->block, offnum)));
913 : 0 : do_clear_vm = true;
914 : 0 : do_clear_heap = true;
915 : 0 : break;
916 : :
917 : 0 : case VM_CORRUPT_MISSING_PAGE_HINT:
918 : :
919 : : /*
920 : : * As of PostgreSQL 9.2, the visibility map bit should never be
921 : : * set if the page-level bit is clear. However, for vacuum, it's
922 : : * possible that the bit got cleared after
923 : : * heap_vac_scan_next_block() was called, so we must recheck now
924 : : * that we have the buffer lock before concluding that the VM is
925 : : * corrupt.
926 : : */
927 [ # # ]: 0 : Assert(!PageIsAllVisible(prstate->page));
928 [ # # ]: 0 : Assert(prstate->old_vmbits & VISIBILITYMAP_VALID_BITS);
929 [ # # ]: 0 : ereport(WARNING,
930 : : (errcode(ERRCODE_DATA_CORRUPTED),
931 : : errmsg("page is not marked all-visible but visibility map bit is set"),
932 : : errcontext("relation \"%s\", page %u",
933 : : relname, prstate->block)));
934 : 0 : do_clear_vm = true;
935 : 0 : break;
936 : : }
937 : :
938 [ # # # # ]: 0 : Assert(do_clear_heap || do_clear_vm);
939 : :
940 : : /* Avoid marking the buffer dirty if PD_ALL_VISIBLE is already clear */
941 [ # # ]: 0 : if (do_clear_heap)
942 : : {
943 [ # # ]: 0 : Assert(PageIsAllVisible(prstate->page));
944 : 0 : PageClearAllVisible(prstate->page);
945 : 0 : MarkBufferDirtyHint(prstate->buffer, true);
946 : : }
947 : :
948 [ # # ]: 0 : if (do_clear_vm)
949 : : {
11 950 : 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_EXCLUSIVE);
11 melanieplageman@gmai 951 :UNC 0 : visibilitymap_clear(prstate->relation->rd_locator, prstate->block,
952 : : prstate->vmbuffer,
953 : : VISIBILITYMAP_VALID_BITS);
11 melanieplageman@gmai 954 :UBC 0 : LockBuffer(prstate->vmbuffer, BUFFER_LOCK_UNLOCK);
126 955 : 0 : prstate->old_vmbits = 0;
956 : : }
957 : 0 : }
958 : :
959 : : /*
960 : : * Decide whether to set the visibility map bits (all-visible and all-frozen)
961 : : * for the current page using information from the PruneState and VM.
962 : : *
963 : : * This function does not actually set the VM bits or page-level visibility
964 : : * hint, PD_ALL_VISIBLE.
965 : : *
966 : : * This should be called only after do_freeze has been decided (and do_prune
967 : : * has been set), as these factor into our heuristic-based decision.
968 : : *
969 : : * Returns true if one or both VM bits should be set and false otherwise.
970 : : */
971 : : static bool
118 melanieplageman@gmai 972 :CBC 224834 : heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
973 : : bool do_prune, bool do_freeze)
974 : : {
975 [ + + ]: 224834 : if (!prstate->attempt_set_vm)
124 976 : 85585 : return false;
977 : :
978 [ + + ]: 139249 : if (!prstate->set_all_visible)
979 : 50122 : return false;
980 : :
981 : : /*
982 : : * If this is an on-access call and we're not actually pruning, avoid
983 : : * setting the visibility map if it would newly dirty the heap page or, if
984 : : * the page is already dirty, if doing so would require including a
985 : : * full-page image (FPI) of the heap page in the WAL.
986 : : */
118 987 [ + + + + : 89127 : if (reason == PRUNE_ON_ACCESS && !do_prune && !do_freeze &&
+ - ]
988 [ + + + + ]: 33272 : (!BufferIsDirty(prstate->buffer) || XLogCheckBufferNeedsBackup(prstate->buffer)))
989 : : {
990 : 19814 : prstate->set_all_visible = prstate->set_all_frozen = false;
991 : 19814 : return false;
992 : : }
993 : :
124 994 : 69313 : prstate->new_vmbits = VISIBILITYMAP_ALL_VISIBLE;
995 : :
996 [ + + ]: 69313 : if (prstate->set_all_frozen)
997 : 38010 : prstate->new_vmbits |= VISIBILITYMAP_ALL_FROZEN;
998 : :
999 [ + + ]: 69313 : if (prstate->new_vmbits == prstate->old_vmbits)
1000 : : {
1001 : 2864 : prstate->new_vmbits = 0;
1002 : 2864 : return false;
1003 : : }
1004 : :
1005 : 66449 : return true;
1006 : : }
1007 : :
1008 : : /*
1009 : : * If the page is already all-frozen, or already all-visible and freezing
1010 : : * won't be attempted, there is no remaining work and we can use the fast path
1011 : : * to avoid the expensive overhead of heap_page_prune_and_freeze().
1012 : : *
1013 : : * This can happen when the page has a stale prune hint, or if VACUUM is
1014 : : * scanning an already all-frozen page due to SKIP_PAGES_THRESHOLD.
1015 : : *
1016 : : * The caller must already have examined the visibility map and saved the
1017 : : * status of the page's VM bits in prstate->old_vmbits. Caller must hold a
1018 : : * content lock on the heap page since it will examine line pointers.
1019 : : *
1020 : : * Before calling prune_freeze_fast_path(), the caller should first
1021 : : * check for and fix any discrepancy between the page-level visibility hint
1022 : : * and the visibility map. Otherwise, the fast path will always prevent us
1023 : : * from getting them in sync. Note that if there are tuples on the page that
1024 : : * are not visible to all but the VM is incorrectly marked
1025 : : * all-visible/all-frozen, we will not get the chance to fix that corruption
1026 : : * when using the fast path.
1027 : : */
1028 : : static void
126 1029 : 13714 : prune_freeze_fast_path(PruneState *prstate, PruneFreezeResult *presult)
1030 : : {
1031 : 13714 : OffsetNumber maxoff = PageGetMaxOffsetNumber(prstate->page);
1032 : 13714 : Page page = prstate->page;
1033 : :
1034 [ - + - - : 13714 : Assert((prstate->old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
- - ]
1035 : : ((prstate->old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1036 : : !prstate->attempt_freeze));
1037 : :
1038 : : /* We'll fill in presult for the caller */
1039 : 13714 : memset(presult, 0, sizeof(PruneFreezeResult));
1040 : :
1041 : : /* Clear any stale prune hint */
1042 [ + + ]: 13714 : if (TransactionIdIsValid(PageGetPruneXid(page)))
1043 : : {
1044 : 98 : PageClearPrunable(page);
1045 : 98 : MarkBufferDirtyHint(prstate->buffer, true);
1046 : : }
1047 : :
1048 [ - + ]: 13714 : if (PageIsEmpty(page))
126 melanieplageman@gmai 1049 :UBC 0 : return;
1050 : :
1051 : : /*
1052 : : * Since the page is all-visible, a count of the normal ItemIds on the
1053 : : * page should be sufficient for vacuum's live tuple count.
1054 : : */
126 melanieplageman@gmai 1055 :CBC 13714 : for (OffsetNumber off = FirstOffsetNumber;
1056 [ + + ]: 812096 : off <= maxoff;
1057 : 798382 : off = OffsetNumberNext(off))
1058 : : {
1059 : 798382 : ItemId lp = PageGetItemId(page, off);
1060 : :
1061 [ + + ]: 798382 : if (!ItemIdIsUsed(lp))
1062 : 19235 : continue;
1063 : :
1064 : 779147 : presult->hastup = true;
1065 : :
1066 [ + + ]: 779147 : if (ItemIdIsNormal(lp))
1067 : 768159 : prstate->live_tuples++;
1068 : : }
1069 : :
1070 : 13714 : presult->live_tuples = prstate->live_tuples;
1071 : : }
1072 : :
1073 : : /*
1074 : : * Prune and repair fragmentation and potentially freeze tuples on the
1075 : : * specified page. If the page's visibility status has changed, update it in
1076 : : * the VM.
1077 : : *
1078 : : * Caller must have pin and buffer cleanup lock on the page. Note that we
1079 : : * don't update the FSM information for page on caller's behalf. Caller might
1080 : : * also need to account for a reduction in the length of the line pointer
1081 : : * array following array truncation by us.
1082 : : *
1083 : : * params contains the input parameters used to control freezing and pruning
1084 : : * behavior. See the definition of PruneFreezeParams for more on what each
1085 : : * parameter does.
1086 : : *
1087 : : * If the HEAP_PAGE_PRUNE_FREEZE option is set in params, we will freeze
1088 : : * tuples if it's required in order to advance relfrozenxid / relminmxid, or
1089 : : * if it's considered advantageous for overall system performance to do so
1090 : : * now. The 'params.cutoffs', 'presult', 'new_relfrozen_xid' and
1091 : : * 'new_relmin_mxid' arguments are required when freezing.
1092 : : *
1093 : : * A vmbuffer corresponding to the heap page is also passed and if the page is
1094 : : * found to be all-visible/all-frozen, we will set it in the VM.
1095 : : *
1096 : : * presult contains output parameters needed by callers, such as the number of
1097 : : * tuples removed and the offsets of dead items on the page after pruning.
1098 : : * heap_page_prune_and_freeze() is responsible for initializing it. Required
1099 : : * by all callers.
1100 : : *
1101 : : * off_loc is the offset location required by the caller to use in error
1102 : : * callback.
1103 : : *
1104 : : * new_relfrozen_xid and new_relmin_mxid must be provided by the caller if the
1105 : : * HEAP_PAGE_PRUNE_FREEZE option is set in params. On entry, they contain the
1106 : : * oldest XID and multi-XID seen on the relation so far. They will be updated
1107 : : * with the oldest values present on the page after pruning. After processing
1108 : : * the whole relation, VACUUM can use these values as the new
1109 : : * relfrozenxid/relminmxid for the relation.
1110 : : */
1111 : : void
242 1112 : 238548 : heap_page_prune_and_freeze(PruneFreezeParams *params,
1113 : : PruneFreezeResult *presult,
1114 : : OffsetNumber *off_loc,
1115 : : TransactionId *new_relfrozen_xid,
1116 : : MultiXactId *new_relmin_mxid)
1117 : : {
1118 : : PruneState prstate;
1119 : : bool do_freeze;
1120 : : bool do_prune;
1121 : : bool do_hint_prune;
1122 : : bool do_set_vm;
1123 : : bool did_tuple_hint_fpi;
1124 : 238548 : int64 fpi_before = pgWalUsage.wal_fpi;
1125 : : TransactionId conflict_xid;
1126 : :
1127 : : /* Initialize prstate */
1128 : 238548 : prune_freeze_setup(params,
1129 : : new_relfrozen_xid, new_relmin_mxid,
1130 : : presult, &prstate);
1131 : :
1132 : : /*
1133 : : * If the VM is set but PD_ALL_VISIBLE is clear, fix that corruption
1134 : : * before pruning and freezing so that the page and VM start out in a
1135 : : * consistent state.
1136 : : */
126 1137 [ + + ]: 238548 : if ((prstate.old_vmbits & VISIBILITYMAP_VALID_BITS) &&
1138 [ - + ]: 20084 : !PageIsAllVisible(prstate.page))
126 melanieplageman@gmai 1139 :UBC 0 : heap_page_fix_vm_corruption(&prstate, InvalidOffsetNumber,
1140 : : VM_CORRUPT_MISSING_PAGE_HINT);
1141 : :
1142 : : /*
1143 : : * If the page is already all-frozen, or already all-visible when freezing
1144 : : * is not being attempted, take the fast path, skipping pruning and
1145 : : * freezing code entirely. This must be done after fixing any discrepancy
1146 : : * between the page-level visibility hint and the VM, since that may have
1147 : : * cleared old_vmbits.
1148 : : */
126 melanieplageman@gmai 1149 [ + + ]:CBC 238548 : if ((params->options & HEAP_PAGE_PRUNE_ALLOW_FAST_PATH) != 0 &&
1150 [ + + ]: 237234 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) ||
1151 [ + + ]: 223520 : ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) &&
1152 [ - + ]: 5911 : !prstate.attempt_freeze)))
1153 : : {
1154 : 13714 : prune_freeze_fast_path(&prstate, presult);
1155 : 13714 : return;
1156 : : }
1157 : :
1158 : : /*
1159 : : * Examine all line pointers and tuple visibility information to determine
1160 : : * which line pointers should change state and which tuples may be frozen.
1161 : : * Prepare queue of state changes to later be executed in a critical
1162 : : * section.
1163 : : */
143 1164 : 224834 : prune_freeze_plan(&prstate, off_loc);
1165 : :
1166 : : /*
1167 : : * After processing all the live tuples on the page, if the newest xmin
1168 : : * amongst them may be considered running by any snapshot, the page cannot
1169 : : * be all-visible. This should be done before determining whether or not
1170 : : * to opportunistically freeze.
1171 : : */
124 1172 [ + + ]: 224834 : if (prstate.set_all_visible &&
1173 [ + + + + ]: 216270 : TransactionIdIsNormal(prstate.newest_live_xid) &&
1174 : 89711 : GlobalVisTestXidConsideredRunning(prstate.vistest,
1175 : : prstate.newest_live_xid,
1176 : : true))
1177 : 3233 : prstate.set_all_visible = prstate.set_all_frozen = false;
1178 : :
1179 : : /*
1180 : : * If checksums are enabled, calling heap_prune_satisfies_vacuum() while
1181 : : * checking tuple visibility information in prune_freeze_plan() may have
1182 : : * caused an FPI to be emitted.
1183 : : */
242 1184 : 224834 : did_tuple_hint_fpi = fpi_before != pgWalUsage.wal_fpi;
1185 : :
844 heikki.linnakangas@i 1186 : 655619 : do_prune = prstate.nredirected > 0 ||
1187 [ + + + + ]: 375389 : prstate.ndead > 0 ||
1188 [ + + ]: 150555 : prstate.nunused > 0;
1189 : :
1190 : : /*
1191 : : * Even if we don't prune anything, if we found a new value for the
1192 : : * pd_prune_xid field or the page was marked full, we will update the hint
1193 : : * bit.
1194 : : */
143 melanieplageman@gmai 1195 [ + + + + ]: 271571 : do_hint_prune = PageGetPruneXid(prstate.page) != prstate.new_prune_xid ||
1196 : 46737 : PageIsFull(prstate.page);
1197 : :
1198 : : /*
1199 : : * Decide if we want to go ahead with freezing according to the freeze
1200 : : * plans we prepared, or not.
1201 : : */
1202 : 224834 : do_freeze = heap_page_will_freeze(did_tuple_hint_fpi,
1203 : : do_prune,
1204 : : do_hint_prune,
1205 : : &prstate);
1206 : :
1207 : : /*
1208 : : * While scanning the line pointers, we did not clear
1209 : : * set_all_visible/set_all_frozen when encountering LP_DEAD items because
1210 : : * we wanted the decision whether or not to freeze the page to be
1211 : : * unaffected by the short-term presence of LP_DEAD items. These LP_DEAD
1212 : : * items are effectively assumed to be LP_UNUSED items in the making. It
1213 : : * doesn't matter which vacuum heap pass (initial pass or final pass) ends
1214 : : * up setting the page all-frozen, as long as the ongoing VACUUM does it.
1215 : : *
1216 : : * Now that we finished determining whether or not to freeze the page,
1217 : : * update set_all_visible and set_all_frozen so that they reflect the true
1218 : : * state of the page for setting PD_ALL_VISIBLE and VM bits.
1219 : : */
248 1220 [ + + ]: 224834 : if (prstate.lpdead_items > 0)
143 1221 : 77722 : prstate.set_all_visible = prstate.set_all_frozen = false;
1222 : :
1223 [ + + - + ]: 224834 : Assert(!prstate.set_all_frozen || prstate.set_all_visible);
118 1224 [ + + - + ]: 224834 : Assert(!prstate.set_all_visible || prstate.attempt_set_vm);
124 1225 [ + + - + ]: 224834 : Assert(!prstate.set_all_visible || (prstate.lpdead_items == 0));
1226 : :
118 1227 : 224834 : do_set_vm = heap_page_will_set_vm(&prstate, params->reason, do_prune, do_freeze);
1228 : :
1229 : : /*
1230 : : * new_vmbits should be 0 regardless of whether or not the page is
1231 : : * all-visible if we do not intend to set the VM.
1232 : : */
124 1233 [ + + - + ]: 224834 : Assert(do_set_vm || prstate.new_vmbits == 0);
1234 : :
1235 : : /*
1236 : : * The snapshot conflict horizon for the whole record is the most
1237 : : * conservative (newest) horizon required by any change in the record.
1238 : : */
1239 : 224834 : conflict_xid = InvalidTransactionId;
1240 [ + + ]: 224834 : if (do_set_vm)
1241 : 66449 : conflict_xid = prstate.newest_live_xid;
1242 [ + + + + ]: 224834 : if (do_freeze && TransactionIdFollows(prstate.pagefrz.FreezePageConflictXid, conflict_xid))
1243 : 3013 : conflict_xid = prstate.pagefrz.FreezePageConflictXid;
1244 [ + + + + ]: 224834 : if (do_prune && TransactionIdFollows(prstate.latest_xid_removed, conflict_xid))
1245 : 66075 : conflict_xid = prstate.latest_xid_removed;
1246 : :
1247 : : /* Lock vmbuffer before entering a critical section */
1248 [ + + ]: 224834 : if (do_set_vm)
1249 : 66449 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_EXCLUSIVE);
1250 : :
1251 : : /* Any error while applying the changes is critical */
844 heikki.linnakangas@i 1252 : 224834 : START_CRIT_SECTION();
1253 : :
285 melanieplageman@gmai 1254 [ + + ]: 224834 : if (do_hint_prune)
1255 : : {
1256 : : /*
1257 : : * Update the page's pd_prune_xid field to either zero, or the lowest
1258 : : * XID of any soon-prunable tuple.
1259 : : */
143 1260 : 182257 : ((PageHeader) prstate.page)->pd_prune_xid = prstate.new_prune_xid;
1261 : :
1262 : : /*
1263 : : * Also clear the "page is full" flag, since there's no point in
1264 : : * repeating the prune/defrag process until something else happens to
1265 : : * the page.
1266 : : */
1267 : 182257 : PageClearFull(prstate.page);
1268 : :
1269 : : /*
1270 : : * If that's all we had to do to the page, this is a non-WAL-logged
1271 : : * hint. If we are going to freeze or prune the page or set
1272 : : * PD_ALL_VISIBLE, we will mark the buffer dirty below.
1273 : : *
1274 : : * Setting PD_ALL_VISIBLE is fully WAL-logged because it is forbidden
1275 : : * for the VM to be set and PD_ALL_VISIBLE to be clear.
1276 : : */
124 1277 [ + + + + : 182257 : if (!do_freeze && !do_prune && !do_set_vm)
+ + ]
143 1278 : 69190 : MarkBufferDirtyHint(prstate.buffer, true);
1279 : : }
1280 : :
124 1281 [ + + + + : 224834 : if (do_prune || do_freeze || do_set_vm)
+ + ]
1282 : : {
1283 : : /* Apply the planned item changes and repair page fragmentation. */
844 heikki.linnakangas@i 1284 [ + + ]: 140572 : if (do_prune)
1285 : : {
143 melanieplageman@gmai 1286 : 79167 : heap_page_prune_execute(prstate.buffer, false,
1287 : : prstate.redirected, prstate.nredirected,
1288 : : prstate.nowdead, prstate.ndead,
1289 : : prstate.nowunused, prstate.nunused);
1290 : : }
1291 : :
844 heikki.linnakangas@i 1292 [ + + ]: 140572 : if (do_freeze)
143 melanieplageman@gmai 1293 : 25149 : heap_freeze_prepared_tuples(prstate.buffer, prstate.frozen, prstate.nfrozen);
1294 : :
1295 : : /* Set the visibility map and page visibility hint */
124 1296 [ + + ]: 140572 : if (do_set_vm)
1297 : : {
1298 : : /*
1299 : : * While it is valid for PD_ALL_VISIBLE to be set when the
1300 : : * corresponding VM bit is clear, we strongly prefer to keep them
1301 : : * in sync.
1302 : : *
1303 : : * The heap buffer must be marked dirty before adding it to the
1304 : : * WAL chain when setting the VM. We don't worry about
1305 : : * unnecessarily dirtying the heap buffer if PD_ALL_VISIBLE is
1306 : : * already set, though. It is extremely rare to have a clean heap
1307 : : * buffer with PD_ALL_VISIBLE already set and the VM bits clear,
1308 : : * so there is no point in optimizing it.
1309 : : */
1310 : 66449 : PageSetAllVisible(prstate.page);
1311 : 66449 : PageClearPrunable(prstate.page);
1312 : 66449 : visibilitymap_set(prstate.block, prstate.vmbuffer, prstate.new_vmbits,
1313 : 66449 : prstate.relation->rd_locator);
1314 : : }
1315 : :
143 1316 : 140572 : MarkBufferDirty(prstate.buffer);
1317 : :
1318 : : /*
1319 : : * Emit a WAL XLOG_HEAP2_PRUNE* record showing what we did
1320 : : */
1321 [ + + + + : 140572 : if (RelationNeedsWAL(prstate.relation))
+ - + - ]
1322 : : {
1323 [ + + + + ]: 202721 : log_heap_prune_and_freeze(prstate.relation, prstate.buffer,
1324 : : do_set_vm ? prstate.vmbuffer : InvalidBuffer,
124 1325 : 64853 : do_set_vm ? prstate.new_vmbits : 0,
1326 : : conflict_xid,
1327 : : true, /* cleanup lock */
1328 : : params->reason,
1329 : : prstate.frozen, prstate.nfrozen,
1330 : : prstate.redirected, prstate.nredirected,
1331 : : prstate.nowdead, prstate.ndead,
1332 : : prstate.nowunused, prstate.nunused);
1333 : : }
1334 : : }
1335 : :
6884 tgl@sss.pgh.pa.us 1336 [ - + ]: 224834 : END_CRIT_SECTION();
1337 : :
124 melanieplageman@gmai 1338 [ + + ]: 224834 : if (do_set_vm)
1339 : 66449 : LockBuffer(prstate.vmbuffer, BUFFER_LOCK_UNLOCK);
1340 : :
1341 : : /*
1342 : : * During its second pass over the heap, VACUUM calls
1343 : : * heap_page_would_be_all_visible() to determine whether a page is
1344 : : * all-visible and all-frozen. The logic here is similar. After completing
1345 : : * pruning and freezing, use an assertion to verify that our results
1346 : : * remain consistent with heap_page_would_be_all_visible(). It's also a
1347 : : * valuable cross-check of the page state after pruning and freezing.
1348 : : */
1349 : : #ifdef USE_ASSERT_CHECKING
1350 [ + + ]: 224834 : if (prstate.set_all_visible)
1351 : : {
1352 : : TransactionId debug_cutoff;
1353 : : bool debug_all_frozen;
1354 : :
1355 [ - + ]: 69313 : Assert(prstate.lpdead_items == 0);
1356 : :
1357 [ - + ]: 69313 : Assert(heap_page_is_all_visible(prstate.relation, prstate.buffer,
1358 : : prstate.vistest,
1359 : : &debug_all_frozen,
1360 : : &debug_cutoff, off_loc));
1361 : :
1362 [ + + - + ]: 69313 : Assert(!TransactionIdIsValid(debug_cutoff) ||
1363 : : debug_cutoff == prstate.newest_live_xid);
1364 : :
1365 : : /*
1366 : : * It's possible the page is composed entirely of frozen tuples but is
1367 : : * not set all-frozen in the VM and did not pass
1368 : : * HEAP_PAGE_PRUNE_FREEZE. In this case, it's possible
1369 : : * heap_page_is_all_visible() finds the page completely frozen, even
1370 : : * though prstate.set_all_frozen is false.
1371 : : */
1372 [ + + - + ]: 69313 : Assert(!prstate.set_all_frozen || debug_all_frozen);
1373 : : }
1374 : : #endif
1375 : :
1376 : : /* Copy information back for caller */
846 heikki.linnakangas@i 1377 : 224834 : presult->ndeleted = prstate.ndeleted;
844 1378 : 224834 : presult->nnewlpdead = prstate.ndead;
1379 : 224834 : presult->nfrozen = prstate.nfrozen;
1380 : 224834 : presult->live_tuples = prstate.live_tuples;
1381 : 224834 : presult->recently_dead_tuples = prstate.recently_dead_tuples;
1382 : 224834 : presult->hastup = prstate.hastup;
1383 : :
1384 : 224834 : presult->lpdead_items = prstate.lpdead_items;
1385 : : /* the presult->deadoffsets array was already filled in */
1386 : :
124 melanieplageman@gmai 1387 : 224834 : presult->newly_all_visible = false;
1388 : 224834 : presult->newly_all_frozen = false;
1389 : 224834 : presult->newly_all_visible_frozen = false;
1390 [ + + ]: 224834 : if (do_set_vm)
1391 : : {
1392 [ + + ]: 66449 : if ((prstate.old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
1393 : : {
1394 : 62943 : presult->newly_all_visible = true;
1395 [ + + ]: 62943 : if (prstate.set_all_frozen)
1396 : 34046 : presult->newly_all_visible_frozen = true;
1397 : : }
1398 [ + - ]: 3506 : else if ((prstate.old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
1399 [ + - ]: 3506 : prstate.set_all_frozen)
1400 : 3506 : presult->newly_all_frozen = true;
1401 : : }
1402 : :
285 1403 [ + + ]: 224834 : if (prstate.attempt_freeze)
1404 : : {
844 heikki.linnakangas@i 1405 [ + + ]: 93709 : if (presult->nfrozen > 0)
1406 : : {
1407 : 25149 : *new_relfrozen_xid = prstate.pagefrz.FreezePageRelfrozenXid;
1408 : 25149 : *new_relmin_mxid = prstate.pagefrz.FreezePageRelminMxid;
1409 : : }
1410 : : else
1411 : : {
1412 : 68560 : *new_relfrozen_xid = prstate.pagefrz.NoFreezePageRelfrozenXid;
1413 : 68560 : *new_relmin_mxid = prstate.pagefrz.NoFreezePageRelminMxid;
1414 : : }
1415 : : }
1416 : : }
1417 : :
1418 : :
1419 : : /*
1420 : : * Perform visibility checks for heap pruning.
1421 : : */
1422 : : static HTSV_Result
143 melanieplageman@gmai 1423 : 20818311 : heap_prune_satisfies_vacuum(PruneState *prstate, HeapTuple tup)
1424 : : {
1425 : : HTSV_Result res;
1426 : : TransactionId dead_after;
1427 : :
1428 : 20818311 : res = HeapTupleSatisfiesVacuumHorizon(tup, prstate->buffer, &dead_after);
1429 : :
2174 andres@anarazel.de 1430 [ + + ]: 20818311 : if (res != HEAPTUPLE_RECENTLY_DEAD)
1431 : 16455687 : return res;
1432 : :
1433 : : /*
1434 : : * For VACUUM, we must be sure to prune tuples with xmax older than
1435 : : * OldestXmin -- a visibility cutoff determined at the beginning of
1436 : : * vacuuming the relation. OldestXmin is used for freezing determination
1437 : : * and we cannot freeze dead tuples' xmaxes.
1438 : : */
737 melanieplageman@gmai 1439 [ + + ]: 4362624 : if (prstate->cutoffs &&
1440 [ + - ]: 1405567 : TransactionIdIsValid(prstate->cutoffs->OldestXmin) &&
1441 [ + - - + : 1405567 : NormalTransactionIdPrecedes(dead_after, prstate->cutoffs->OldestXmin))
+ + ]
1442 : 970427 : return HEAPTUPLE_DEAD;
1443 : :
1444 : : /*
1445 : : * Determine whether or not the tuple is considered dead when compared
1446 : : * with the provided GlobalVisState. On-access pruning does not provide
1447 : : * VacuumCutoffs. And for vacuum, even if the tuple's xmax is not older
1448 : : * than OldestXmin, GlobalVisTestIsRemovableXid() could find the row dead
1449 : : * if the GlobalVisState has been updated since the beginning of vacuuming
1450 : : * the relation.
1451 : : */
124 1452 [ + + ]: 3392197 : if (GlobalVisTestIsRemovableXid(prstate->vistest, dead_after, true))
737 1453 : 2894831 : return HEAPTUPLE_DEAD;
1454 : :
2174 andres@anarazel.de 1455 : 497366 : return res;
1456 : : }
1457 : :
1458 : :
1459 : : /*
1460 : : * Pruning calculates tuple visibility once and saves the results in an array
1461 : : * of int8. See PruneState.htsv for details. This helper function is meant
1462 : : * to guard against examining visibility status array members which have not
1463 : : * yet been computed.
1464 : : */
1465 : : static inline HTSV_Result
844 heikki.linnakangas@i 1466 : 20801676 : htsv_get_valid_status(int status)
1467 : : {
1468 [ + - - + ]: 20801676 : Assert(status >= HEAPTUPLE_DEAD &&
1469 : : status <= HEAPTUPLE_DELETE_IN_PROGRESS);
1470 : 20801676 : return (HTSV_Result) status;
1471 : : }
1472 : :
1473 : : /*
1474 : : * Prune specified line pointer or a HOT chain originating at line pointer.
1475 : : *
1476 : : * Tuple visibility information is provided in prstate->htsv.
1477 : : *
1478 : : * If the item is an index-referenced tuple (i.e. not a heap-only tuple),
1479 : : * the HOT chain is pruned by removing all DEAD tuples at the start of the HOT
1480 : : * chain. We also prune any RECENTLY_DEAD tuples preceding a DEAD tuple.
1481 : : * This is OK because a RECENTLY_DEAD tuple preceding a DEAD tuple is really
1482 : : * DEAD, our visibility test is just too coarse to detect it.
1483 : : *
1484 : : * Pruning must never leave behind a DEAD tuple that still has tuple storage.
1485 : : * VACUUM isn't prepared to deal with that case.
1486 : : *
1487 : : * The root line pointer is redirected to the tuple immediately after the
1488 : : * latest DEAD tuple. If all tuples in the chain are DEAD, the root line
1489 : : * pointer is marked LP_DEAD. (This includes the case of a DEAD simple
1490 : : * tuple, which we treat as a chain of length 1.)
1491 : : *
1492 : : * We don't actually change the page here. We just add entries to the arrays in
1493 : : * prstate showing the changes to be made. Items to be redirected are added
1494 : : * to the redirected[] array (two entries per redirection); items to be set to
1495 : : * LP_DEAD state are added to nowdead[]; and items to be set to LP_UNUSED
1496 : : * state are added to nowunused[]. We perform bookkeeping of live tuples,
1497 : : * visibility etc. based on what the page will look like after the changes
1498 : : * applied. All that bookkeeping is performed in the heap_prune_record_*()
1499 : : * subroutines. The division of labor is that heap_prune_chain() decides the
1500 : : * fate of each tuple, ie. whether it's going to be removed, redirected or
1501 : : * left unchanged, and the heap_prune_record_*() subroutines update PruneState
1502 : : * based on that outcome.
1503 : : */
1504 : : static void
143 melanieplageman@gmai 1505 : 20666892 : heap_prune_chain(OffsetNumber maxoff, OffsetNumber rootoffnum,
1506 : : PruneState *prstate)
1507 : : {
6828 bruce@momjian.us 1508 : 20666892 : TransactionId priorXmax = InvalidTransactionId;
1509 : : ItemId rootlp;
1510 : : OffsetNumber offnum;
1511 : : OffsetNumber chainitems[MaxHeapTuplesPerPage];
143 melanieplageman@gmai 1512 : 20666892 : Page page = prstate->page;
1513 : :
1514 : : /*
1515 : : * After traversing the HOT chain, ndeadchain is the index in chainitems
1516 : : * of the first live successor after the last dead item.
1517 : : */
846 heikki.linnakangas@i 1518 : 20666892 : int ndeadchain = 0,
1519 : 20666892 : nchain = 0;
1520 : :
1521 : 20666892 : rootlp = PageGetItemId(page, rootoffnum);
1522 : :
1523 : : /* Start from the root tuple */
6884 tgl@sss.pgh.pa.us 1524 : 20666892 : offnum = rootoffnum;
1525 : :
1526 : : /* while not end of the chain */
1527 : : for (;;)
1528 : 360823 : {
1529 : : HeapTupleHeader htup;
1530 : : ItemId lp;
1531 : :
1532 : : /* Sanity check (pure paranoia) */
1768 pg@bowt.ie 1533 [ - + ]: 21027715 : if (offnum < FirstOffsetNumber)
1768 pg@bowt.ie 1534 :UBC 0 : break;
1535 : :
1536 : : /*
1537 : : * An offset past the end of page's line pointer array is possible
1538 : : * when the array was truncated (original item must have been unused)
1539 : : */
1768 pg@bowt.ie 1540 [ - + ]:CBC 21027715 : if (offnum > maxoff)
6884 tgl@sss.pgh.pa.us 1541 :UBC 0 : break;
1542 : :
1543 : : /* If item is already processed, stop --- it must not be same chain */
846 heikki.linnakangas@i 1544 [ - + ]:CBC 21027715 : if (prstate->processed[offnum])
6714 tgl@sss.pgh.pa.us 1545 :UBC 0 : break;
1546 : :
846 heikki.linnakangas@i 1547 :CBC 21027715 : lp = PageGetItemId(page, offnum);
1548 : :
1549 : : /*
1550 : : * Unused item obviously isn't part of the chain. Likewise, a dead
1551 : : * line pointer can't be part of the chain. Both of those cases were
1552 : : * already marked as processed.
1553 : : */
1554 [ - + ]: 21027715 : Assert(ItemIdIsUsed(lp));
1555 [ - + ]: 21027715 : Assert(!ItemIdIsDead(lp));
1556 : :
1557 : : /*
1558 : : * If we are looking at the redirected root line pointer, jump to the
1559 : : * first normal tuple in the chain. If we find a redirect somewhere
1560 : : * else, stop --- it must not be same chain.
1561 : : */
6884 tgl@sss.pgh.pa.us 1562 [ + + ]: 21027715 : if (ItemIdIsRedirected(lp))
1563 : : {
1564 [ - + ]: 226039 : if (nchain > 0)
6884 tgl@sss.pgh.pa.us 1565 :UBC 0 : break; /* not at start of chain */
6884 tgl@sss.pgh.pa.us 1566 :CBC 226039 : chainitems[nchain++] = offnum;
1567 : 226039 : offnum = ItemIdGetRedirect(rootlp);
1568 : 226039 : continue;
1569 : : }
1570 : :
1571 [ - + ]: 20801676 : Assert(ItemIdIsNormal(lp));
1572 : :
846 heikki.linnakangas@i 1573 : 20801676 : htup = (HeapTupleHeader) PageGetItem(page, lp);
1574 : :
1575 : : /*
1576 : : * Check the tuple XMIN against prior XMAX, if any
1577 : : */
6884 tgl@sss.pgh.pa.us 1578 [ + + - + ]: 20936460 : if (TransactionIdIsValid(priorXmax) &&
3188 alvherre@alvh.no-ip. 1579 : 134784 : !TransactionIdEquals(HeapTupleHeaderGetXmin(htup), priorXmax))
6884 tgl@sss.pgh.pa.us 1580 :UBC 0 : break;
1581 : :
1582 : : /*
1583 : : * OK, this tuple is indeed a member of the chain.
1584 : : */
6884 tgl@sss.pgh.pa.us 1585 :CBC 20801676 : chainitems[nchain++] = offnum;
1586 : :
844 heikki.linnakangas@i 1587 [ + + + - ]: 20801676 : switch (htsv_get_valid_status(prstate->htsv[offnum]))
1588 : : {
6884 tgl@sss.pgh.pa.us 1589 : 3937284 : case HEAPTUPLE_DEAD:
1590 : :
1591 : : /* Remember the last DEAD tuple seen */
846 heikki.linnakangas@i 1592 : 3937284 : ndeadchain = nchain;
1593 : 3937284 : HeapTupleHeaderAdvanceConflictHorizon(htup,
1594 : : &prstate->latest_xid_removed);
1595 : : /* Advance to next chain member */
6884 tgl@sss.pgh.pa.us 1596 : 3937284 : break;
1597 : :
1598 : 497366 : case HEAPTUPLE_RECENTLY_DEAD:
1599 : :
1600 : : /*
1601 : : * We don't need to advance the conflict horizon for
1602 : : * RECENTLY_DEAD tuples, even if we are removing them. This
1603 : : * is because we only remove RECENTLY_DEAD tuples if they
1604 : : * precede a DEAD tuple, and the DEAD tuple must have been
1605 : : * inserted by a newer transaction than the RECENTLY_DEAD
1606 : : * tuple by virtue of being later in the chain. We will have
1607 : : * advanced the conflict horizon for the DEAD tuple.
1608 : : */
1609 : :
1610 : : /*
1611 : : * Advance past RECENTLY_DEAD tuples just in case there's a
1612 : : * DEAD one after them. We have to make sure that we don't
1613 : : * miss any DEAD tuples, since DEAD tuples that still have
1614 : : * tuple storage after pruning will confuse VACUUM.
1615 : : */
1616 : 497366 : break;
1617 : :
1618 : 16367026 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1619 : : case HEAPTUPLE_LIVE:
1620 : : case HEAPTUPLE_INSERT_IN_PROGRESS:
846 heikki.linnakangas@i 1621 : 16367026 : goto process_chain;
1622 : :
6884 tgl@sss.pgh.pa.us 1623 :UBC 0 : default:
1624 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
1625 : : goto process_chain;
1626 : : }
1627 : :
1628 : : /*
1629 : : * If the tuple is not HOT-updated, then we are at the end of this
1630 : : * HOT-update chain.
1631 : : */
6884 tgl@sss.pgh.pa.us 1632 [ + + ]:CBC 4434650 : if (!HeapTupleHeaderIsHotUpdated(htup))
846 heikki.linnakangas@i 1633 : 4299866 : goto process_chain;
1634 : :
1635 : : /* HOT implies it can't have moved to different partition */
3032 andres@anarazel.de 1636 [ - + ]: 134784 : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
1637 : :
1638 : : /*
1639 : : * Advance to next chain member.
1640 : : */
143 melanieplageman@gmai 1641 [ - + ]: 134784 : Assert(ItemPointerGetBlockNumber(&htup->t_ctid) == prstate->block);
6884 tgl@sss.pgh.pa.us 1642 : 134784 : offnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
4932 alvherre@alvh.no-ip. 1643 : 134784 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
1644 : : }
1645 : :
846 heikki.linnakangas@i 1646 [ # # # # ]:UBC 0 : if (ItemIdIsRedirected(rootlp) && nchain < 2)
1647 : : {
1648 : : /*
1649 : : * We found a redirect item that doesn't point to a valid follow-on
1650 : : * item. This can happen if the loop in heap_page_prune_and_freeze()
1651 : : * caused us to visit the dead successor of a redirect item before
1652 : : * visiting the redirect item. We can clean up by setting the
1653 : : * redirect item to LP_DEAD state or LP_UNUSED if the caller
1654 : : * indicated.
1655 : : */
1656 : 0 : heap_prune_record_dead_or_unused(prstate, rootoffnum, false);
1657 : 0 : return;
1658 : : }
1659 : :
1660 : 0 : process_chain:
1661 : :
846 heikki.linnakangas@i 1662 [ + + ]:CBC 20666892 : if (ndeadchain == 0)
1663 : : {
1664 : : /*
1665 : : * No DEAD tuple was found, so the chain is entirely composed of
1666 : : * normal, unchanged tuples. Leave it alone.
1667 : : */
844 1668 : 16778933 : int i = 0;
1669 : :
1670 [ + + ]: 16778933 : if (ItemIdIsRedirected(rootlp))
1671 : : {
1672 : 203062 : heap_prune_record_unchanged_lp_redirect(prstate, rootoffnum);
1673 : 203062 : i++;
1674 : : }
1675 [ + + ]: 33566342 : for (; i < nchain; i++)
143 melanieplageman@gmai 1676 : 16787409 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1677 : : }
846 heikki.linnakangas@i 1678 [ + + ]: 3887959 : else if (ndeadchain == nchain)
1679 : : {
1680 : : /*
1681 : : * The entire chain is dead. Mark the root line pointer LP_DEAD, and
1682 : : * fully remove the other tuples in the chain.
1683 : : */
1684 : 3814090 : heap_prune_record_dead_or_unused(prstate, rootoffnum, ItemIdIsNormal(rootlp));
1685 [ + + ]: 3862052 : for (int i = 1; i < nchain; i++)
1686 : 47962 : heap_prune_record_unused(prstate, chainitems[i], true);
1687 : : }
1688 : : else
1689 : : {
1690 : : /*
1691 : : * We found a DEAD tuple in the chain. Redirect the root line pointer
1692 : : * to the first non-DEAD tuple, and mark as unused each intermediate
1693 : : * item that we are able to remove from the chain.
1694 : : */
1695 : 73869 : heap_prune_record_redirect(prstate, rootoffnum, chainitems[ndeadchain],
1696 : 73869 : ItemIdIsNormal(rootlp));
1697 [ + + ]: 98209 : for (int i = 1; i < ndeadchain; i++)
1698 : 24340 : heap_prune_record_unused(prstate, chainitems[i], true);
1699 : :
1700 : : /* the rest of tuples in the chain are normal, unchanged tuples */
1701 [ + + ]: 150852 : for (int i = ndeadchain; i < nchain; i++)
143 melanieplageman@gmai 1702 : 76983 : heap_prune_record_unchanged_lp_normal(prstate, chainitems[i]);
1703 : : }
1704 : : }
1705 : :
1706 : : /* Record lowest soon-prunable XID */
1707 : : static void
126 1708 : 2276346 : heap_prune_record_prunable(PruneState *prstate, TransactionId xid,
1709 : : OffsetNumber offnum)
1710 : : {
1711 : : /*
1712 : : * This should exactly match the PageSetPrunable macro. We can't store
1713 : : * directly into the page header yet, so we update working state.
1714 : : */
6714 tgl@sss.pgh.pa.us 1715 [ - + ]: 2276346 : Assert(TransactionIdIsNormal(xid));
1716 [ + + + + ]: 4510295 : if (!TransactionIdIsValid(prstate->new_prune_xid) ||
1717 : 2233949 : TransactionIdPrecedes(xid, prstate->new_prune_xid))
1718 : 44335 : prstate->new_prune_xid = xid;
1719 : :
1720 : : /*
1721 : : * It's incorrect for a page to be marked all-visible if it contains
1722 : : * prunable items.
1723 : : */
126 melanieplageman@gmai 1724 [ - + ]: 2276346 : if (PageIsAllVisible(prstate->page))
126 melanieplageman@gmai 1725 :UBC 0 : heap_page_fix_vm_corruption(prstate, offnum,
1726 : : VM_CORRUPT_TUPLE_VISIBILITY);
6714 tgl@sss.pgh.pa.us 1727 :CBC 2276346 : }
1728 : :
1729 : : /* Record line pointer to be redirected */
1730 : : static void
1731 : 73869 : heap_prune_record_redirect(PruneState *prstate,
1732 : : OffsetNumber offnum, OffsetNumber rdoffnum,
1733 : : bool was_normal)
1734 : : {
846 heikki.linnakangas@i 1735 [ - + ]: 73869 : Assert(!prstate->processed[offnum]);
1736 : 73869 : prstate->processed[offnum] = true;
1737 : :
1738 : : /*
1739 : : * Do not mark the redirect target here. It needs to be counted
1740 : : * separately as an unchanged tuple.
1741 : : */
1742 : :
6714 tgl@sss.pgh.pa.us 1743 [ - + ]: 73869 : Assert(prstate->nredirected < MaxHeapTuplesPerPage);
1744 : 73869 : prstate->redirected[prstate->nredirected * 2] = offnum;
1745 : 73869 : prstate->redirected[prstate->nredirected * 2 + 1] = rdoffnum;
1746 : :
1747 : 73869 : prstate->nredirected++;
1748 : :
1749 : : /*
1750 : : * If the root entry had been a normal tuple, we are deleting it, so count
1751 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1752 : : * count.
1753 : : */
846 heikki.linnakangas@i 1754 [ + + ]: 73869 : if (was_normal)
1755 : 63367 : prstate->ndeleted++;
1756 : :
844 1757 : 73869 : prstate->hastup = true;
6884 tgl@sss.pgh.pa.us 1758 : 73869 : }
1759 : :
1760 : : /* Record line pointer to be marked dead */
1761 : : static void
846 heikki.linnakangas@i 1762 : 3779888 : heap_prune_record_dead(PruneState *prstate, OffsetNumber offnum,
1763 : : bool was_normal)
1764 : : {
1765 [ - + ]: 3779888 : Assert(!prstate->processed[offnum]);
1766 : 3779888 : prstate->processed[offnum] = true;
1767 : :
6714 tgl@sss.pgh.pa.us 1768 [ - + ]: 3779888 : Assert(prstate->ndead < MaxHeapTuplesPerPage);
1769 : 3779888 : prstate->nowdead[prstate->ndead] = offnum;
1770 : 3779888 : prstate->ndead++;
1771 : :
1772 : : /*
1773 : : * Deliberately delay unsetting set_all_visible and set_all_frozen until
1774 : : * later during pruning. Removable dead tuples shouldn't preclude freezing
1775 : : * the page.
1776 : : */
1777 : :
1778 : : /* Record the dead offset for vacuum */
844 heikki.linnakangas@i 1779 : 3779888 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
1780 : :
1781 : : /*
1782 : : * If the root entry had been a normal tuple, we are deleting it, so count
1783 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1784 : : * count.
1785 : : */
846 1786 [ + + ]: 3779888 : if (was_normal)
1787 : 3767413 : prstate->ndeleted++;
6884 tgl@sss.pgh.pa.us 1788 : 3779888 : }
1789 : :
1790 : : /*
1791 : : * Depending on whether or not the caller set mark_unused_now to true, record that a
1792 : : * line pointer should be marked LP_DEAD or LP_UNUSED. There are other cases in
1793 : : * which we will mark line pointers LP_UNUSED, but we will not mark line
1794 : : * pointers LP_DEAD if mark_unused_now is true.
1795 : : */
1796 : : static void
846 heikki.linnakangas@i 1797 : 3814090 : heap_prune_record_dead_or_unused(PruneState *prstate, OffsetNumber offnum,
1798 : : bool was_normal)
1799 : : {
1800 : : /*
1801 : : * If the caller set mark_unused_now to true, we can remove dead tuples
1802 : : * during pruning instead of marking their line pointers dead. Set this
1803 : : * tuple's line pointer LP_UNUSED. We hint that this option is less
1804 : : * likely.
1805 : : */
920 rhaas@postgresql.org 1806 [ + + ]: 3814090 : if (unlikely(prstate->mark_unused_now))
846 heikki.linnakangas@i 1807 : 34202 : heap_prune_record_unused(prstate, offnum, was_normal);
1808 : : else
1809 : 3779888 : heap_prune_record_dead(prstate, offnum, was_normal);
1810 : :
1811 : : /*
1812 : : * It's incorrect for the page to be set all-visible if it contains dead
1813 : : * items. Fix that on the heap page and check the VM for corruption as
1814 : : * well. Do that here rather than in heap_prune_record_dead() so we also
1815 : : * cover tuples that are directly marked LP_UNUSED via mark_unused_now.
1816 : : */
126 melanieplageman@gmai 1817 [ - + ]: 3814090 : if (PageIsAllVisible(prstate->page))
126 melanieplageman@gmai 1818 :UBC 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
920 rhaas@postgresql.org 1819 :CBC 3814090 : }
1820 : :
1821 : : /* Record line pointer to be marked unused */
1822 : : static void
846 heikki.linnakangas@i 1823 : 1111582 : heap_prune_record_unused(PruneState *prstate, OffsetNumber offnum, bool was_normal)
1824 : : {
1825 [ - + ]: 1111582 : Assert(!prstate->processed[offnum]);
1826 : 1111582 : prstate->processed[offnum] = true;
1827 : :
6714 tgl@sss.pgh.pa.us 1828 [ - + ]: 1111582 : Assert(prstate->nunused < MaxHeapTuplesPerPage);
1829 : 1111582 : prstate->nowunused[prstate->nunused] = offnum;
1830 : 1111582 : prstate->nunused++;
1831 : :
1832 : : /*
1833 : : * If the root entry had been a normal tuple, we are deleting it, so count
1834 : : * it in the result. But changing a redirect (even to DEAD state) doesn't
1835 : : * count.
1836 : : */
846 heikki.linnakangas@i 1837 [ + + ]: 1111582 : if (was_normal)
1838 : 109892 : prstate->ndeleted++;
6714 tgl@sss.pgh.pa.us 1839 : 1111582 : }
1840 : :
1841 : : /*
1842 : : * Record an unused line pointer that is left unchanged.
1843 : : */
1844 : : static void
143 melanieplageman@gmai 1845 : 376212 : heap_prune_record_unchanged_lp_unused(PruneState *prstate, OffsetNumber offnum)
1846 : : {
844 heikki.linnakangas@i 1847 [ - + ]: 376212 : Assert(!prstate->processed[offnum]);
1848 : 376212 : prstate->processed[offnum] = true;
1849 : 376212 : }
1850 : :
1851 : : /*
1852 : : * Record line pointer that is left unchanged. We consider freezing it, and
1853 : : * update bookkeeping of tuple counts and page visibility.
1854 : : */
1855 : : static void
143 melanieplageman@gmai 1856 : 16877639 : heap_prune_record_unchanged_lp_normal(PruneState *prstate, OffsetNumber offnum)
1857 : : {
1858 : : HeapTupleHeader htup;
1859 : : TransactionId xmin;
1860 : 16877639 : Page page = prstate->page;
1861 : :
844 heikki.linnakangas@i 1862 [ - + ]: 16877639 : Assert(!prstate->processed[offnum]);
1863 : 16877639 : prstate->processed[offnum] = true;
1864 : :
1865 : 16877639 : prstate->hastup = true; /* the page is not empty */
1866 : :
1867 : : /*
1868 : : * The criteria for counting a tuple as live in this block need to match
1869 : : * what analyze.c's acquire_sample_rows() does, otherwise VACUUM and
1870 : : * ANALYZE may produce wildly different reltuples values, e.g. when there
1871 : : * are many recently-dead tuples.
1872 : : *
1873 : : * The logic here is a bit simpler than acquire_sample_rows(), as VACUUM
1874 : : * can't run inside a transaction block, which makes some cases impossible
1875 : : * (e.g. in-progress insert from the same transaction).
1876 : : *
1877 : : * HEAPTUPLE_DEAD are handled by the other heap_prune_record_*()
1878 : : * subroutines. They don't count dead items like acquire_sample_rows()
1879 : : * does, because we assume that all dead items will become LP_UNUSED
1880 : : * before VACUUM finishes. This difference is only superficial. VACUUM
1881 : : * effectively agrees with ANALYZE about DEAD items, in the end. VACUUM
1882 : : * won't remember LP_DEAD items, but only because they're not supposed to
1883 : : * be left behind when it is done. (Cases where we bypass index vacuuming
1884 : : * will violate this optimistic assumption, but the overall impact of that
1885 : : * should be negligible.)
1886 : : */
1887 : 16877639 : htup = (HeapTupleHeader) PageGetItem(page, PageGetItemId(page, offnum));
1888 : :
1889 [ + + + + : 16877639 : switch (prstate->htsv[offnum])
- ]
1890 : : {
1891 : 14601293 : case HEAPTUPLE_LIVE:
1892 : :
1893 : : /*
1894 : : * Count it as live. Not only is this natural, but it's also what
1895 : : * acquire_sample_rows() does.
1896 : : */
1897 : 14601293 : prstate->live_tuples++;
1898 : :
1899 : : /*
1900 : : * Is the tuple definitely visible to all transactions?
1901 : : *
1902 : : * NB: Like with per-tuple hint bits, we can't set the
1903 : : * PD_ALL_VISIBLE flag if the inserter committed asynchronously.
1904 : : * See SetHintBits for more info. Check that the tuple is hinted
1905 : : * xmin-committed because of that.
1906 : : */
124 melanieplageman@gmai 1907 [ + + ]: 14601293 : if (!HeapTupleHeaderXminCommitted(htup))
1908 : : {
1909 : 38582 : prstate->set_all_visible = false;
1910 : 38582 : prstate->set_all_frozen = false;
1911 : 38582 : break;
1912 : : }
1913 : :
1914 : : /*
1915 : : * The inserter definitely committed. But we don't know if it is
1916 : : * old enough that everyone sees it as committed. Later, after
1917 : : * processing all the tuples on the page, we'll check if there is
1918 : : * any snapshot that still considers the newest xid on the page to
1919 : : * be running. If so, we don't consider the page all-visible.
1920 : : */
1921 : 14562711 : xmin = HeapTupleHeaderGetXmin(htup);
1922 : :
1923 : : /* Track newest xmin on page. */
1924 [ + + + + ]: 14562711 : if (TransactionIdFollows(xmin, prstate->newest_live_xid) &&
1925 : : TransactionIdIsNormal(xmin))
1926 : 636389 : prstate->newest_live_xid = xmin;
1927 : :
844 heikki.linnakangas@i 1928 : 14562711 : break;
1929 : :
1930 : 497366 : case HEAPTUPLE_RECENTLY_DEAD:
1931 : 497366 : prstate->recently_dead_tuples++;
143 melanieplageman@gmai 1932 : 497366 : prstate->set_all_visible = false;
1933 : 497366 : prstate->set_all_frozen = false;
1934 : :
1935 : : /*
1936 : : * This tuple will soon become DEAD. Update the hint field so
1937 : : * that the page is reconsidered for pruning in future.
1938 : : */
844 heikki.linnakangas@i 1939 : 497366 : heap_prune_record_prunable(prstate,
1940 : : HeapTupleHeaderGetUpdateXid(htup),
1941 : : offnum);
1942 : 497366 : break;
1943 : :
1944 : 879958 : case HEAPTUPLE_INSERT_IN_PROGRESS:
1945 : :
1946 : : /*
1947 : : * We do not count these rows as live, because we expect the
1948 : : * inserting transaction to update the counters at commit, and we
1949 : : * assume that will happen only after we report our results. This
1950 : : * assumption is a bit shaky, but it is what acquire_sample_rows()
1951 : : * does, so be consistent.
1952 : : */
143 melanieplageman@gmai 1953 : 879958 : prstate->set_all_visible = false;
1954 : 879958 : prstate->set_all_frozen = false;
1955 : :
1956 : : /*
1957 : : * Though there is nothing "prunable" on the page, we maintain
1958 : : * pd_prune_xid for inserts so that we have the opportunity to
1959 : : * mark them all-visible during the next round of pruning.
1960 : : */
118 1961 : 879958 : heap_prune_record_prunable(prstate,
1962 : : HeapTupleHeaderGetXmin(htup),
1963 : : offnum);
844 heikki.linnakangas@i 1964 : 879958 : break;
1965 : :
1966 : 899022 : case HEAPTUPLE_DELETE_IN_PROGRESS:
1967 : :
1968 : : /*
1969 : : * This an expected case during concurrent vacuum. Count such
1970 : : * rows as live. As above, we assume the deleting transaction
1971 : : * will commit and update the counters after we report.
1972 : : */
1973 : 899022 : prstate->live_tuples++;
143 melanieplageman@gmai 1974 : 899022 : prstate->set_all_visible = false;
1975 : 899022 : prstate->set_all_frozen = false;
1976 : :
1977 : : /*
1978 : : * This tuple may soon become DEAD. Update the hint field so that
1979 : : * the page is reconsidered for pruning in future.
1980 : : */
844 heikki.linnakangas@i 1981 : 899022 : heap_prune_record_prunable(prstate,
1982 : : HeapTupleHeaderGetUpdateXid(htup),
1983 : : offnum);
1984 : 899022 : break;
1985 : :
844 heikki.linnakangas@i 1986 :UBC 0 : default:
1987 : :
1988 : : /*
1989 : : * DEAD tuples should've been passed to heap_prune_record_dead()
1990 : : * or heap_prune_record_unused() instead.
1991 : : */
1992 [ # # ]: 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result %d",
1993 : : prstate->htsv[offnum]);
1994 : : break;
1995 : : }
1996 : :
1997 : : /* Consider freezing any normal tuples which will not be removed */
285 melanieplageman@gmai 1998 [ + + ]:CBC 16877639 : if (prstate->attempt_freeze)
1999 : : {
2000 : : bool totally_frozen;
2001 : :
844 heikki.linnakangas@i 2002 [ + + ]: 6929052 : if ((heap_prepare_freeze_tuple(htup,
2003 : 6929052 : prstate->cutoffs,
2004 : : &prstate->pagefrz,
2005 : 6929052 : &prstate->frozen[prstate->nfrozen],
2006 : : &totally_frozen)))
2007 : : {
2008 : : /* Save prepared freeze plan for later */
2009 : 4615593 : prstate->frozen[prstate->nfrozen++].offset = offnum;
2010 : : }
2011 : :
2012 : : /*
2013 : : * If any tuple isn't either totally frozen already or eligible to
2014 : : * become totally frozen (according to its freeze plan), then the page
2015 : : * definitely cannot be set all-frozen in the visibility map later on.
2016 : : */
2017 [ + + ]: 6929052 : if (!totally_frozen)
143 melanieplageman@gmai 2018 : 2545745 : prstate->set_all_frozen = false;
2019 : : }
844 heikki.linnakangas@i 2020 : 16877639 : }
2021 : :
2022 : :
2023 : : /*
2024 : : * Record line pointer that was already LP_DEAD and is left unchanged.
2025 : : */
2026 : : static void
143 melanieplageman@gmai 2027 : 1603722 : heap_prune_record_unchanged_lp_dead(PruneState *prstate, OffsetNumber offnum)
2028 : : {
844 heikki.linnakangas@i 2029 [ - + ]: 1603722 : Assert(!prstate->processed[offnum]);
2030 : 1603722 : prstate->processed[offnum] = true;
2031 : :
2032 : : /*
2033 : : * Deliberately don't set hastup for LP_DEAD items. We make the soft
2034 : : * assumption that any LP_DEAD items encountered here will become
2035 : : * LP_UNUSED later on, before count_nondeletable_pages is reached. If we
2036 : : * don't make this assumption then rel truncation will only happen every
2037 : : * other VACUUM, at most. Besides, VACUUM must treat
2038 : : * hastup/nonempty_pages as provisional no matter how LP_DEAD items are
2039 : : * handled (handled here, or handled later on).
2040 : : *
2041 : : * Similarly, don't unset set_all_visible and set_all_frozen until later,
2042 : : * at the end of heap_page_prune_and_freeze(). This will allow us to
2043 : : * attempt to freeze the page after pruning. As long as we unset it
2044 : : * before updating the visibility map, this will be correct.
2045 : : */
2046 : :
2047 : : /* Record the dead offset for vacuum */
2048 : 1603722 : prstate->deadoffsets[prstate->lpdead_items++] = offnum;
2049 : :
2050 : : /*
2051 : : * It's incorrect for a page to be marked all-visible if it contains dead
2052 : : * items.
2053 : : */
126 melanieplageman@gmai 2054 [ - + ]: 1603722 : if (PageIsAllVisible(prstate->page))
126 melanieplageman@gmai 2055 :UBC 0 : heap_page_fix_vm_corruption(prstate, offnum, VM_CORRUPT_LPDEAD);
844 heikki.linnakangas@i 2056 :CBC 1603722 : }
2057 : :
2058 : : /*
2059 : : * Record LP_REDIRECT that is left unchanged.
2060 : : */
2061 : : static void
2062 : 203062 : heap_prune_record_unchanged_lp_redirect(PruneState *prstate, OffsetNumber offnum)
2063 : : {
2064 : : /*
2065 : : * A redirect line pointer doesn't count as a live tuple.
2066 : : *
2067 : : * If we leave a redirect line pointer in place, there will be another
2068 : : * tuple on the page that it points to. We will do the bookkeeping for
2069 : : * that separately. So we have nothing to do here, except remember that
2070 : : * we processed this item.
2071 : : */
846 2072 [ - + ]: 203062 : Assert(!prstate->processed[offnum]);
2073 : 203062 : prstate->processed[offnum] = true;
2074 : 203062 : }
2075 : :
2076 : : /*
2077 : : * Perform the actual page changes needed by heap_page_prune_and_freeze().
2078 : : *
2079 : : * If 'lp_truncate_only' is set, we are merely marking LP_DEAD line pointers
2080 : : * as unused, not redirecting or removing anything else. The
2081 : : * PageRepairFragmentation() call is skipped in that case.
2082 : : *
2083 : : * If 'lp_truncate_only' is not set, the caller must hold a cleanup lock on
2084 : : * the buffer. If it is set, an ordinary exclusive lock suffices.
2085 : : */
2086 : : void
853 2087 : 90855 : heap_page_prune_execute(Buffer buffer, bool lp_truncate_only,
2088 : : OffsetNumber *redirected, int nredirected,
2089 : : OffsetNumber *nowdead, int ndead,
2090 : : OffsetNumber *nowunused, int nunused)
2091 : : {
331 peter@eisentraut.org 2092 : 90855 : Page page = BufferGetPage(buffer);
2093 : : OffsetNumber *offnum;
2094 : : HeapTupleHeader htup PG_USED_FOR_ASSERTS_ONLY;
2095 : :
2096 : : /* Shouldn't be called unless there's something to do */
1937 pg@bowt.ie 2097 [ + + + + : 90855 : Assert(nredirected > 0 || ndead > 0 || nunused > 0);
- + ]
2098 : :
2099 : : /* If 'lp_truncate_only', we can only remove already-dead line pointers */
853 heikki.linnakangas@i 2100 [ + + + - : 90855 : Assert(!lp_truncate_only || (nredirected == 0 && ndead == 0));
- + ]
2101 : :
2102 : : /* Update all redirected line pointers */
6714 tgl@sss.pgh.pa.us 2103 : 90855 : offnum = redirected;
1725 pg@bowt.ie 2104 [ + + ]: 184045 : for (int i = 0; i < nredirected; i++)
2105 : : {
6714 tgl@sss.pgh.pa.us 2106 : 93190 : OffsetNumber fromoff = *offnum++;
2107 : 93190 : OffsetNumber tooff = *offnum++;
2108 : 93190 : ItemId fromlp = PageGetItemId(page, fromoff);
2109 : : ItemId tolp PG_USED_FOR_ASSERTS_ONLY;
2110 : :
2111 : : #ifdef USE_ASSERT_CHECKING
2112 : :
2113 : : /*
2114 : : * Any existing item that we set as an LP_REDIRECT (any 'from' item)
2115 : : * must be the first item from a HOT chain. If the item has tuple
2116 : : * storage then it can't be a heap-only tuple. Otherwise we are just
2117 : : * maintaining an existing LP_REDIRECT from an existing HOT chain that
2118 : : * has been pruned at least once before now.
2119 : : */
1725 pg@bowt.ie 2120 [ + + ]: 93190 : if (!ItemIdIsRedirected(fromlp))
2121 : : {
2122 [ + - - + ]: 82125 : Assert(ItemIdHasStorage(fromlp) && ItemIdIsNormal(fromlp));
2123 : :
2124 : 82125 : htup = (HeapTupleHeader) PageGetItem(page, fromlp);
2125 [ - + ]: 82125 : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2126 : : }
2127 : : else
2128 : : {
2129 : : /* We shouldn't need to redundantly set the redirect */
2130 [ - + ]: 11065 : Assert(ItemIdGetRedirect(fromlp) != tooff);
2131 : : }
2132 : :
2133 : : /*
2134 : : * The item that we're about to set as an LP_REDIRECT (the 'from'
2135 : : * item) will point to an existing item (the 'to' item) that is
2136 : : * already a heap-only tuple. There can be at most one LP_REDIRECT
2137 : : * item per HOT chain.
2138 : : *
2139 : : * We need to keep around an LP_REDIRECT item (after original
2140 : : * non-heap-only root tuple gets pruned away) so that it's always
2141 : : * possible for VACUUM to easily figure out what TID to delete from
2142 : : * indexes when an entire HOT chain becomes dead. A heap-only tuple
2143 : : * can never become LP_DEAD; an LP_REDIRECT item or a regular heap
2144 : : * tuple can.
2145 : : *
2146 : : * This check may miss problems, e.g. the target of a redirect could
2147 : : * be marked as unused subsequently. The page_verify_redirects() check
2148 : : * below will catch such problems.
2149 : : */
2150 : 93190 : tolp = PageGetItemId(page, tooff);
2151 [ + - - + ]: 93190 : Assert(ItemIdHasStorage(tolp) && ItemIdIsNormal(tolp));
2152 : 93190 : htup = (HeapTupleHeader) PageGetItem(page, tolp);
2153 [ - + ]: 93190 : Assert(HeapTupleHeaderIsHeapOnly(htup));
2154 : : #endif
2155 : :
6012 tgl@sss.pgh.pa.us 2156 : 93190 : ItemIdSetRedirect(fromlp, tooff);
2157 : : }
2158 : :
2159 : : /* Update all now-dead line pointers */
6714 2160 : 90855 : offnum = nowdead;
1725 pg@bowt.ie 2161 [ + + ]: 4150321 : for (int i = 0; i < ndead; i++)
2162 : : {
6714 tgl@sss.pgh.pa.us 2163 : 4059466 : OffsetNumber off = *offnum++;
2164 : 4059466 : ItemId lp = PageGetItemId(page, off);
2165 : :
2166 : : #ifdef USE_ASSERT_CHECKING
2167 : :
2168 : : /*
2169 : : * An LP_DEAD line pointer must be left behind when the original item
2170 : : * (which is dead to everybody) could still be referenced by a TID in
2171 : : * an index. This should never be necessary with any individual
2172 : : * heap-only tuple item, though. (It's not clear how much of a problem
2173 : : * that would be, but there is no reason to allow it.)
2174 : : */
1725 pg@bowt.ie 2175 [ + + ]: 4059466 : if (ItemIdHasStorage(lp))
2176 : : {
2177 [ - + ]: 4045486 : Assert(ItemIdIsNormal(lp));
2178 : 4045486 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2179 [ - + ]: 4045486 : Assert(!HeapTupleHeaderIsHeapOnly(htup));
2180 : : }
2181 : : else
2182 : : {
2183 : : /* Whole HOT chain becomes dead */
2184 [ - + ]: 13980 : Assert(ItemIdIsRedirected(lp));
2185 : : }
2186 : : #endif
2187 : :
6714 tgl@sss.pgh.pa.us 2188 : 4059466 : ItemIdSetDead(lp);
2189 : : }
2190 : :
2191 : : /* Update all now-unused line pointers */
2192 : 90855 : offnum = nowunused;
1725 pg@bowt.ie 2193 [ + + ]: 1425293 : for (int i = 0; i < nunused; i++)
2194 : : {
6714 tgl@sss.pgh.pa.us 2195 : 1334438 : OffsetNumber off = *offnum++;
2196 : 1334438 : ItemId lp = PageGetItemId(page, off);
2197 : :
2198 : : #ifdef USE_ASSERT_CHECKING
2199 : :
853 heikki.linnakangas@i 2200 [ + + ]: 1334438 : if (lp_truncate_only)
2201 : : {
2202 : : /* Setting LP_DEAD to LP_UNUSED in vacuum's second pass */
2203 [ + - - + ]: 199913 : Assert(ItemIdIsDead(lp) && !ItemIdHasStorage(lp));
2204 : : }
2205 : : else
2206 : : {
2207 : : /*
2208 : : * When heap_page_prune_and_freeze() was called, mark_unused_now
2209 : : * may have been passed as true, which allows would-be LP_DEAD
2210 : : * items to be made LP_UNUSED instead. This is only possible if
2211 : : * the relation has no indexes. If there are any dead items, then
2212 : : * mark_unused_now was not true and every item being marked
2213 : : * LP_UNUSED must refer to a heap-only tuple.
2214 : : */
2215 [ + + ]: 1134525 : if (ndead > 0)
2216 : : {
2217 [ + - - + ]: 65145 : Assert(ItemIdHasStorage(lp) && ItemIdIsNormal(lp));
2218 : 65145 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2219 [ - + ]: 65145 : Assert(HeapTupleHeaderIsHeapOnly(htup));
2220 : : }
2221 : : else
2222 [ - + ]: 1069380 : Assert(ItemIdIsUsed(lp));
2223 : : }
2224 : :
2225 : : #endif
2226 : :
6714 tgl@sss.pgh.pa.us 2227 : 1334438 : ItemIdSetUnused(lp);
2228 : : }
2229 : :
853 heikki.linnakangas@i 2230 [ + + ]: 90855 : if (lp_truncate_only)
2231 : 2903 : PageTruncateLinePointerArray(page);
2232 : : else
2233 : : {
2234 : : /*
2235 : : * Finally, repair any fragmentation, and update the page's hint bit
2236 : : * about whether it has free pointers.
2237 : : */
2238 : 87952 : PageRepairFragmentation(page);
2239 : :
2240 : : /*
2241 : : * Now that the page has been modified, assert that redirect items
2242 : : * still point to valid targets.
2243 : : */
2244 : 87952 : page_verify_redirects(page);
2245 : : }
1707 andres@anarazel.de 2246 : 90855 : }
2247 : :
2248 : :
2249 : : /*
2250 : : * If built with assertions, verify that all LP_REDIRECT items point to a
2251 : : * valid item.
2252 : : *
2253 : : * One way that bugs related to HOT pruning show is redirect items pointing to
2254 : : * removed tuples. It's not trivial to reliably check that marking an item
2255 : : * unused will not orphan a redirect item during heap_prune_chain() /
2256 : : * heap_page_prune_execute(), so we additionally check the whole page after
2257 : : * pruning. Without this check such bugs would typically only cause asserts
2258 : : * later, potentially well after the corruption has been introduced.
2259 : : *
2260 : : * Also check comments in heap_page_prune_execute()'s redirection loop.
2261 : : */
2262 : : static void
2263 : 87952 : page_verify_redirects(Page page)
2264 : : {
2265 : : #ifdef USE_ASSERT_CHECKING
2266 : : OffsetNumber offnum;
2267 : : OffsetNumber maxoff;
2268 : :
2269 : 87952 : maxoff = PageGetMaxOffsetNumber(page);
2270 : 87952 : for (offnum = FirstOffsetNumber;
2271 [ + + ]: 8324129 : offnum <= maxoff;
2272 : 8236177 : offnum = OffsetNumberNext(offnum))
2273 : : {
2274 : 8236177 : ItemId itemid = PageGetItemId(page, offnum);
2275 : : OffsetNumber targoff;
2276 : : ItemId targitem;
2277 : : HeapTupleHeader htup;
2278 : :
2279 [ + + ]: 8236177 : if (!ItemIdIsRedirected(itemid))
2280 : 7968272 : continue;
2281 : :
2282 : 267905 : targoff = ItemIdGetRedirect(itemid);
2283 : 267905 : targitem = PageGetItemId(page, targoff);
2284 : :
2285 [ - + ]: 267905 : Assert(ItemIdIsUsed(targitem));
2286 [ - + ]: 267905 : Assert(ItemIdIsNormal(targitem));
2287 [ - + ]: 267905 : Assert(ItemIdHasStorage(targitem));
2288 : 267905 : htup = (HeapTupleHeader) PageGetItem(page, targitem);
2289 [ - + ]: 267905 : Assert(HeapTupleHeaderIsHeapOnly(htup));
2290 : : }
2291 : : #endif
6884 tgl@sss.pgh.pa.us 2292 : 87952 : }
2293 : :
2294 : :
2295 : : /*
2296 : : * For all items in this page, find their respective root line pointers.
2297 : : * If item k is part of a HOT-chain with root at item j, then we set
2298 : : * root_offsets[k - 1] = j.
2299 : : *
2300 : : * The passed-in root_offsets array must have MaxHeapTuplesPerPage entries.
2301 : : * Unused entries are filled with InvalidOffsetNumber (zero).
2302 : : *
2303 : : * The function must be called with at least share lock on the buffer, to
2304 : : * prevent concurrent prune operations.
2305 : : *
2306 : : * Note: The information collected here is valid only as long as the caller
2307 : : * holds a pin on the buffer. Once pin is released, a tuple might be pruned
2308 : : * and reused by a completely unrelated tuple.
2309 : : */
2310 : : void
2311 : 138999 : heap_get_root_tuples(Page page, OffsetNumber *root_offsets)
2312 : : {
2313 : : OffsetNumber offnum,
2314 : : maxoff;
2315 : :
2173 alvherre@alvh.no-ip. 2316 [ + - - + : 138999 : MemSet(root_offsets, InvalidOffsetNumber,
- - - - -
- ]
2317 : : MaxHeapTuplesPerPage * sizeof(OffsetNumber));
2318 : :
6884 tgl@sss.pgh.pa.us 2319 : 138999 : maxoff = PageGetMaxOffsetNumber(page);
6648 bruce@momjian.us 2320 [ + + ]: 11636448 : for (offnum = FirstOffsetNumber; offnum <= maxoff; offnum = OffsetNumberNext(offnum))
2321 : : {
6828 2322 : 11497449 : ItemId lp = PageGetItemId(page, offnum);
2323 : : HeapTupleHeader htup;
2324 : : OffsetNumber nextoffnum;
2325 : : TransactionId priorXmax;
2326 : :
2327 : : /* skip unused and dead items */
6884 tgl@sss.pgh.pa.us 2328 [ + + + + ]: 11497449 : if (!ItemIdIsUsed(lp) || ItemIdIsDead(lp))
2329 : 9814 : continue;
2330 : :
2331 [ + + ]: 11487635 : if (ItemIdIsNormal(lp))
2332 : : {
2333 : 11483759 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2334 : :
2335 : : /*
2336 : : * Check if this tuple is part of a HOT-chain rooted at some other
2337 : : * tuple. If so, skip it for now; we'll process it when we find
2338 : : * its root.
2339 : : */
2340 [ + + ]: 11483759 : if (HeapTupleHeaderIsHeapOnly(htup))
2341 : 4348 : continue;
2342 : :
2343 : : /*
2344 : : * This is either a plain tuple or the root of a HOT-chain.
2345 : : * Remember it in the mapping.
2346 : : */
2347 : 11479411 : root_offsets[offnum - 1] = offnum;
2348 : :
2349 : : /* If it's not the start of a HOT-chain, we're done with it */
2350 [ + + ]: 11479411 : if (!HeapTupleHeaderIsHotUpdated(htup))
2351 : 11479087 : continue;
2352 : :
2353 : : /* Set up to scan the HOT-chain */
2354 : 324 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
4932 alvherre@alvh.no-ip. 2355 : 324 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2356 : : }
2357 : : else
2358 : : {
2359 : : /* Must be a redirect item. We do not set its root_offsets entry */
6884 tgl@sss.pgh.pa.us 2360 [ - + ]: 3876 : Assert(ItemIdIsRedirected(lp));
2361 : : /* Set up to scan the HOT-chain */
2362 : 3876 : nextoffnum = ItemIdGetRedirect(lp);
2363 : 3876 : priorXmax = InvalidTransactionId;
2364 : : }
2365 : :
2366 : : /*
2367 : : * Now follow the HOT-chain and collect other tuples in the chain.
2368 : : *
2369 : : * Note: Even though this is a nested loop, the complexity of the
2370 : : * function is O(N) because a tuple in the page should be visited not
2371 : : * more than twice, once in the outer loop and once in HOT-chain
2372 : : * chases.
2373 : : */
2374 : : for (;;)
2375 : : {
2376 : : /* Sanity check (pure paranoia) */
10 pg@bowt.ie 2377 [ - + ]: 4344 : if (nextoffnum < FirstOffsetNumber)
1768 pg@bowt.ie 2378 :UBC 0 : break;
2379 : :
2380 : : /*
2381 : : * An offset past the end of page's line pointer array is possible
2382 : : * when the array was truncated
2383 : : */
10 pg@bowt.ie 2384 [ - + ]:CBC 4344 : if (nextoffnum > maxoff)
1936 pg@bowt.ie 2385 :UBC 0 : break;
2386 : :
6884 tgl@sss.pgh.pa.us 2387 :CBC 4344 : lp = PageGetItemId(page, nextoffnum);
2388 : :
2389 : : /* Check for broken chains */
2390 [ - + ]: 4344 : if (!ItemIdIsNormal(lp))
6884 tgl@sss.pgh.pa.us 2391 :UBC 0 : break;
2392 : :
6884 tgl@sss.pgh.pa.us 2393 :CBC 4344 : htup = (HeapTupleHeader) PageGetItem(page, lp);
2394 : :
2395 [ + + - + ]: 4812 : if (TransactionIdIsValid(priorXmax) &&
3188 alvherre@alvh.no-ip. 2396 : 468 : !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(htup)))
6884 tgl@sss.pgh.pa.us 2397 :UBC 0 : break;
2398 : :
2399 : : /* Remember the root line pointer for this item */
6884 tgl@sss.pgh.pa.us 2400 :CBC 4344 : root_offsets[nextoffnum - 1] = offnum;
2401 : :
2402 : : /* Advance to next chain member, if any */
2403 [ + + ]: 4344 : if (!HeapTupleHeaderIsHotUpdated(htup))
2404 : 4200 : break;
2405 : :
2406 : : /* HOT implies it can't have moved to different partition */
3032 andres@anarazel.de 2407 [ - + ]: 144 : Assert(!HeapTupleHeaderIndicatesMovedPartitions(htup));
2408 : :
6884 tgl@sss.pgh.pa.us 2409 : 144 : nextoffnum = ItemPointerGetOffsetNumber(&htup->t_ctid);
4932 alvherre@alvh.no-ip. 2410 : 144 : priorXmax = HeapTupleHeaderGetUpdateXid(htup);
2411 : : }
2412 : : }
6884 tgl@sss.pgh.pa.us 2413 : 138999 : }
2414 : :
2415 : :
2416 : : /*
2417 : : * Compare fields that describe actions required to freeze tuple with caller's
2418 : : * open plan. If everything matches then the frz tuple plan is equivalent to
2419 : : * caller's plan.
2420 : : */
2421 : : static inline bool
853 heikki.linnakangas@i 2422 : 1267341 : heap_log_freeze_eq(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2423 : : {
2424 [ + + ]: 1267341 : if (plan->xmax == frz->xmax &&
2425 [ + + ]: 1267302 : plan->t_infomask2 == frz->t_infomask2 &&
2426 [ + + ]: 1266309 : plan->t_infomask == frz->t_infomask &&
2427 [ + - ]: 1263049 : plan->frzflags == frz->frzflags)
2428 : 1263049 : return true;
2429 : :
2430 : : /* Caller must call heap_log_freeze_new_plan again for frz */
2431 : 4292 : return false;
2432 : : }
2433 : :
2434 : : /*
2435 : : * Comparator used to deduplicate the freeze plans used in WAL records.
2436 : : */
2437 : : static int
2438 : 1604917 : heap_log_freeze_cmp(const void *arg1, const void *arg2)
2439 : : {
195 peter@eisentraut.org 2440 : 1604917 : const HeapTupleFreeze *frz1 = arg1;
2441 : 1604917 : const HeapTupleFreeze *frz2 = arg2;
2442 : :
853 heikki.linnakangas@i 2443 [ + + ]: 1604917 : if (frz1->xmax < frz2->xmax)
2444 : 56 : return -1;
2445 [ + + ]: 1604861 : else if (frz1->xmax > frz2->xmax)
2446 : 86 : return 1;
2447 : :
2448 [ + + ]: 1604775 : if (frz1->t_infomask2 < frz2->t_infomask2)
2449 : 4195 : return -1;
2450 [ + + ]: 1600580 : else if (frz1->t_infomask2 > frz2->t_infomask2)
2451 : 7046 : return 1;
2452 : :
2453 [ + + ]: 1593534 : if (frz1->t_infomask < frz2->t_infomask)
2454 : 13870 : return -1;
2455 [ + + ]: 1579664 : else if (frz1->t_infomask > frz2->t_infomask)
2456 : 19114 : return 1;
2457 : :
2458 [ - + ]: 1560550 : if (frz1->frzflags < frz2->frzflags)
853 heikki.linnakangas@i 2459 :UBC 0 : return -1;
853 heikki.linnakangas@i 2460 [ - + ]:CBC 1560550 : else if (frz1->frzflags > frz2->frzflags)
853 heikki.linnakangas@i 2461 :UBC 0 : return 1;
2462 : :
2463 : : /*
2464 : : * heap_log_freeze_eq would consider these tuple-wise plans to be equal.
2465 : : * (So the tuples will share a single canonical freeze plan.)
2466 : : *
2467 : : * We tiebreak on page offset number to keep each freeze plan's page
2468 : : * offset number array individually sorted. (Unnecessary, but be tidy.)
2469 : : */
853 heikki.linnakangas@i 2470 [ + + ]:CBC 1560550 : if (frz1->offset < frz2->offset)
2471 : 1412019 : return -1;
2472 [ + - ]: 148531 : else if (frz1->offset > frz2->offset)
2473 : 148531 : return 1;
2474 : :
853 heikki.linnakangas@i 2475 :UBC 0 : Assert(false);
2476 : : return 0;
2477 : : }
2478 : :
2479 : : /*
2480 : : * Start new plan initialized using tuple-level actions. At least one tuple
2481 : : * will have steps required to freeze described by caller's plan during REDO.
2482 : : */
2483 : : static inline void
853 heikki.linnakangas@i 2484 :CBC 29438 : heap_log_freeze_new_plan(xlhp_freeze_plan *plan, HeapTupleFreeze *frz)
2485 : : {
2486 : 29438 : plan->xmax = frz->xmax;
2487 : 29438 : plan->t_infomask2 = frz->t_infomask2;
2488 : 29438 : plan->t_infomask = frz->t_infomask;
2489 : 29438 : plan->frzflags = frz->frzflags;
2490 : 29438 : plan->ntuples = 1; /* for now */
2491 : 29438 : }
2492 : :
2493 : : /*
2494 : : * Deduplicate tuple-based freeze plans so that each distinct set of
2495 : : * processing steps is only stored once in the WAL record.
2496 : : * Called during original execution of freezing (for logged relations).
2497 : : *
2498 : : * Return value is number of plans set in *plans_out for caller. Also writes
2499 : : * an array of offset numbers into *offsets_out output argument for caller
2500 : : * (actually there is one array per freeze plan, but that's not of immediate
2501 : : * concern to our caller).
2502 : : */
2503 : : static int
2504 : 25146 : heap_log_freeze_plan(HeapTupleFreeze *tuples, int ntuples,
2505 : : xlhp_freeze_plan *plans_out,
2506 : : OffsetNumber *offsets_out)
2507 : : {
2508 : 25146 : int nplans = 0;
2509 : :
2510 : : /* Sort tuple-based freeze plans in the order required to deduplicate */
2511 : 25146 : qsort(tuples, ntuples, sizeof(HeapTupleFreeze), heap_log_freeze_cmp);
2512 : :
2513 [ + + ]: 1317633 : for (int i = 0; i < ntuples; i++)
2514 : : {
2515 : 1292487 : HeapTupleFreeze *frz = tuples + i;
2516 : :
2517 [ + + ]: 1292487 : if (i == 0)
2518 : : {
2519 : : /* New canonical freeze plan starting with first tup */
2520 : 25146 : heap_log_freeze_new_plan(plans_out, frz);
2521 : 25146 : nplans++;
2522 : : }
2523 [ + + ]: 1267341 : else if (heap_log_freeze_eq(plans_out, frz))
2524 : : {
2525 : : /* tup matches open canonical plan -- include tup in it */
2526 [ - + ]: 1263049 : Assert(offsets_out[i - 1] < frz->offset);
2527 : 1263049 : plans_out->ntuples++;
2528 : : }
2529 : : else
2530 : : {
2531 : : /* Tup doesn't match current plan -- done with it now */
2532 : 4292 : plans_out++;
2533 : :
2534 : : /* New canonical freeze plan starting with this tup */
2535 : 4292 : heap_log_freeze_new_plan(plans_out, frz);
2536 : 4292 : nplans++;
2537 : : }
2538 : :
2539 : : /*
2540 : : * Save page offset number in dedicated buffer in passing.
2541 : : *
2542 : : * REDO routine relies on the record's offset numbers array grouping
2543 : : * offset numbers by freeze plan. The sort order within each grouping
2544 : : * is ascending offset number order, just to keep things tidy.
2545 : : */
2546 : 1292487 : offsets_out[i] = frz->offset;
2547 : : }
2548 : :
2549 [ + - - + ]: 25146 : Assert(nplans > 0 && nplans <= ntuples);
2550 : :
2551 : 25146 : return nplans;
2552 : : }
2553 : :
2554 : : /*
2555 : : * Write an XLOG_HEAP2_PRUNE* WAL record
2556 : : *
2557 : : * This is used for several different page maintenance operations:
2558 : : *
2559 : : * - Page pruning, in VACUUM's 1st pass or on access: Some items are
2560 : : * redirected, some marked dead, and some removed altogether.
2561 : : *
2562 : : * - Freezing: Items are marked as 'frozen'.
2563 : : *
2564 : : * - Vacuum, 2nd pass: Items that are already LP_DEAD are marked as unused.
2565 : : *
2566 : : * They have enough commonalities that we use a single WAL record for them
2567 : : * all.
2568 : : *
2569 : : * If replaying the record requires a cleanup lock, pass cleanup_lock = true.
2570 : : * Replaying 'redirected' or 'dead' items always requires a cleanup lock, but
2571 : : * replaying 'unused' items depends on whether they were all previously marked
2572 : : * as dead.
2573 : : *
2574 : : * If the VM is being updated, vmflags will contain the bits to set. In this
2575 : : * case, vmbuffer should already have been updated and marked dirty and should
2576 : : * still be pinned and locked.
2577 : : *
2578 : : * Note: This function scribbles on the 'frozen' array.
2579 : : *
2580 : : * Note: This is called in a critical section, so careful what you do here.
2581 : : */
2582 : : void
2583 : 160094 : log_heap_prune_and_freeze(Relation relation, Buffer buffer,
2584 : : Buffer vmbuffer, uint8 vmflags,
2585 : : TransactionId conflict_xid,
2586 : : bool cleanup_lock,
2587 : : PruneReason reason,
2588 : : HeapTupleFreeze *frozen, int nfrozen,
2589 : : OffsetNumber *redirected, int nredirected,
2590 : : OffsetNumber *dead, int ndead,
2591 : : OffsetNumber *unused, int nunused)
2592 : : {
2593 : : xl_heap_prune xlrec;
2594 : : XLogRecPtr recptr;
2595 : : uint8 info;
2596 : : uint8 regbuf_flags_heap;
2597 : :
124 melanieplageman@gmai 2598 : 160094 : Page heap_page = BufferGetPage(buffer);
2599 : :
2600 : : /* The following local variables hold data registered in the WAL record: */
2601 : : xlhp_freeze_plan plans[MaxHeapTuplesPerPage];
2602 : : xlhp_freeze_plans freeze_plans;
2603 : : xlhp_prune_items redirect_items;
2604 : : xlhp_prune_items dead_items;
2605 : : xlhp_prune_items unused_items;
2606 : : OffsetNumber frz_offsets[MaxHeapTuplesPerPage];
286 2607 [ + + + + : 160094 : bool do_prune = nredirected > 0 || ndead > 0 || nunused > 0;
+ + ]
2608 : 160094 : bool do_set_vm = vmflags & VISIBILITYMAP_VALID_BITS;
124 2609 : 160094 : bool heap_fpi_allowed = true;
2610 : :
286 2611 [ - + ]: 160094 : Assert((vmflags & VISIBILITYMAP_VALID_BITS) == vmflags);
2612 : :
853 heikki.linnakangas@i 2613 : 160094 : xlrec.flags = 0;
286 melanieplageman@gmai 2614 : 160094 : regbuf_flags_heap = REGBUF_STANDARD;
2615 : :
2616 : : /*
2617 : : * We can avoid an FPI of the heap page if the only modification we are
2618 : : * making to it is to set PD_ALL_VISIBLE and checksums/wal_log_hints are
2619 : : * disabled.
2620 : : *
2621 : : * However, if the page has never been WAL-logged (LSN is invalid), we
2622 : : * must force an FPI regardless. This can happen when another backend
2623 : : * extends the heap, initializes the page, and then fails before WAL-
2624 : : * logging it. Since heap extension is not WAL-logged, recovery might try
2625 : : * to replay our record and find that the page isn't initialized, which
2626 : : * would cause a PANIC.
2627 : : */
124 2628 [ - + ]: 160094 : if (!XLogRecPtrIsValid(PageGetLSN(heap_page)))
124 melanieplageman@gmai 2629 :UBC 0 : regbuf_flags_heap |= REGBUF_FORCE_IMAGE;
124 melanieplageman@gmai 2630 [ + + + + :CBC 160094 : else if (!do_prune && nfrozen == 0 && (!do_set_vm || !XLogHintBitIsNeeded()))
+ - + + +
+ ]
2631 : : {
286 2632 : 2664 : regbuf_flags_heap |= REGBUF_NO_IMAGE;
124 2633 : 2664 : heap_fpi_allowed = false;
2634 : : }
2635 : :
2636 : : /*
2637 : : * Prepare data for the buffer. The arrays are not actually in the
2638 : : * buffer, but we pretend that they are. When XLogInsert stores a full
2639 : : * page image, the arrays can be omitted.
2640 : : */
853 heikki.linnakangas@i 2641 : 160094 : XLogBeginInsert();
286 melanieplageman@gmai 2642 : 160094 : XLogRegisterBuffer(0, buffer, regbuf_flags_heap);
2643 : :
2644 [ + + ]: 160094 : if (do_set_vm)
2645 : 86379 : XLogRegisterBuffer(1, vmbuffer, 0);
2646 : :
853 heikki.linnakangas@i 2647 [ + + ]: 160094 : if (nfrozen > 0)
2648 : : {
2649 : : int nplans;
2650 : :
2651 : 25146 : xlrec.flags |= XLHP_HAS_FREEZE_PLANS;
2652 : :
2653 : : /*
2654 : : * Prepare deduplicated representation for use in the WAL record. This
2655 : : * destructively sorts frozen tuples array in-place.
2656 : : */
2657 : 25146 : nplans = heap_log_freeze_plan(frozen, nfrozen, plans, frz_offsets);
2658 : :
2659 : 25146 : freeze_plans.nplans = nplans;
530 peter@eisentraut.org 2660 : 25146 : XLogRegisterBufData(0, &freeze_plans,
2661 : : offsetof(xlhp_freeze_plans, plans));
2662 : 25146 : XLogRegisterBufData(0, plans,
2663 : : sizeof(xlhp_freeze_plan) * nplans);
2664 : : }
853 heikki.linnakangas@i 2665 [ + + ]: 160094 : if (nredirected > 0)
2666 : : {
2667 : 18879 : xlrec.flags |= XLHP_HAS_REDIRECTIONS;
2668 : :
2669 : 18879 : redirect_items.ntargets = nredirected;
530 peter@eisentraut.org 2670 : 18879 : XLogRegisterBufData(0, &redirect_items,
2671 : : offsetof(xlhp_prune_items, data));
2672 : 18879 : XLogRegisterBufData(0, redirected,
2673 : : sizeof(OffsetNumber[2]) * nredirected);
2674 : : }
853 heikki.linnakangas@i 2675 [ + + ]: 160094 : if (ndead > 0)
2676 : : {
2677 : 59689 : xlrec.flags |= XLHP_HAS_DEAD_ITEMS;
2678 : :
2679 : 59689 : dead_items.ntargets = ndead;
530 peter@eisentraut.org 2680 : 59689 : XLogRegisterBufData(0, &dead_items,
2681 : : offsetof(xlhp_prune_items, data));
2682 : 59689 : XLogRegisterBufData(0, dead,
2683 : : sizeof(OffsetNumber) * ndead);
2684 : : }
853 heikki.linnakangas@i 2685 [ + + ]: 160094 : if (nunused > 0)
2686 : : {
2687 : 42629 : xlrec.flags |= XLHP_HAS_NOW_UNUSED_ITEMS;
2688 : :
2689 : 42629 : unused_items.ntargets = nunused;
530 peter@eisentraut.org 2690 : 42629 : XLogRegisterBufData(0, &unused_items,
2691 : : offsetof(xlhp_prune_items, data));
2692 : 42629 : XLogRegisterBufData(0, unused,
2693 : : sizeof(OffsetNumber) * nunused);
2694 : : }
853 heikki.linnakangas@i 2695 [ + + ]: 160094 : if (nfrozen > 0)
530 peter@eisentraut.org 2696 : 25146 : XLogRegisterBufData(0, frz_offsets,
2697 : : sizeof(OffsetNumber) * nfrozen);
2698 : :
2699 : : /*
2700 : : * Prepare the main xl_heap_prune record. We already set the XLHP_HAS_*
2701 : : * flag above.
2702 : : */
286 melanieplageman@gmai 2703 [ + + ]: 160094 : if (vmflags & VISIBILITYMAP_ALL_VISIBLE)
2704 : : {
2705 : 86379 : xlrec.flags |= XLHP_VM_ALL_VISIBLE;
2706 [ + + ]: 86379 : if (vmflags & VISIBILITYMAP_ALL_FROZEN)
2707 : 52805 : xlrec.flags |= XLHP_VM_ALL_FROZEN;
2708 : : }
853 heikki.linnakangas@i 2709 [ + + + + : 160094 : if (RelationIsAccessibleInLogicalDecoding(relation))
+ - - + -
- - - + +
+ + - + -
- + - ]
2710 : 622 : xlrec.flags |= XLHP_IS_CATALOG_REL;
2711 [ + + ]: 160094 : if (TransactionIdIsValid(conflict_xid))
2712 : 121953 : xlrec.flags |= XLHP_HAS_CONFLICT_HORIZON;
2713 [ + + ]: 160094 : if (cleanup_lock)
2714 : 137868 : xlrec.flags |= XLHP_CLEANUP_LOCK;
2715 : : else
2716 : : {
2717 [ + - - + ]: 22226 : Assert(nredirected == 0 && ndead == 0);
2718 : : /* also, any items in 'unused' must've been LP_DEAD previously */
2719 : : }
530 peter@eisentraut.org 2720 : 160094 : XLogRegisterData(&xlrec, SizeOfHeapPrune);
853 heikki.linnakangas@i 2721 [ + + ]: 160094 : if (TransactionIdIsValid(conflict_xid))
530 peter@eisentraut.org 2722 : 121953 : XLogRegisterData(&conflict_xid, sizeof(TransactionId));
2723 : :
853 heikki.linnakangas@i 2724 [ + + + - ]: 160094 : switch (reason)
2725 : : {
2726 : 70747 : case PRUNE_ON_ACCESS:
2727 : 70747 : info = XLOG_HEAP2_PRUNE_ON_ACCESS;
2728 : 70747 : break;
2729 : 67121 : case PRUNE_VACUUM_SCAN:
2730 : 67121 : info = XLOG_HEAP2_PRUNE_VACUUM_SCAN;
2731 : 67121 : break;
2732 : 22226 : case PRUNE_VACUUM_CLEANUP:
2733 : 22226 : info = XLOG_HEAP2_PRUNE_VACUUM_CLEANUP;
2734 : 22226 : break;
853 heikki.linnakangas@i 2735 :UBC 0 : default:
2736 [ # # ]: 0 : elog(ERROR, "unrecognized prune reason: %d", (int) reason);
2737 : : break;
2738 : : }
853 heikki.linnakangas@i 2739 :CBC 160094 : recptr = XLogInsert(RM_HEAP2_ID, info);
2740 : :
286 melanieplageman@gmai 2741 [ + + ]: 160094 : if (do_set_vm)
2742 : : {
2743 [ - + ]: 86379 : Assert(BufferIsDirty(vmbuffer));
2744 : 86379 : PageSetLSN(BufferGetPage(vmbuffer), recptr);
2745 : : }
2746 : :
2747 : : /*
2748 : : * If we explicitly skip an FPI, we must not stamp the heap page with this
2749 : : * record's LSN. Recovery skips records <= the stamped LSN, so this could
2750 : : * lead to skipping an earlier FPI needed to repair a torn page.
2751 : : */
124 2752 [ + + ]: 160094 : if (heap_fpi_allowed)
2753 : : {
286 2754 [ - + ]: 157430 : Assert(BufferIsDirty(buffer));
124 2755 : 157430 : PageSetLSN(heap_page, recptr);
2756 : : }
853 heikki.linnakangas@i 2757 : 160094 : }
|
