Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * vacuumlazy.c
4 : * Concurrent ("lazy") vacuuming.
5 : *
6 : * Heap relations are vacuumed in three main phases. In phase I, vacuum scans
7 : * relation pages, pruning and freezing tuples and saving dead tuples' TIDs in
8 : * a TID store. If that TID store fills up or vacuum finishes scanning the
9 : * relation, it progresses to phase II: index vacuuming. Index vacuuming
10 : * deletes the dead index entries referenced in the TID store. In phase III,
11 : * vacuum scans the blocks of the relation referred to by the TIDs in the TID
12 : * store and reaps the corresponding dead items, freeing that space for future
13 : * tuples.
14 : *
15 : * If there are no indexes or index scanning is disabled, phase II may be
16 : * skipped. If phase I identified very few dead index entries or if vacuum's
17 : * failsafe mechanism has triggered (to avoid transaction ID wraparound),
18 : * vacuum may skip phases II and III.
19 : *
20 : * If the TID store fills up in phase I, vacuum suspends phase I and proceeds
21 : * to phases II and III, cleaning up the dead tuples referenced in the current
22 : * TID store. This empties the TID store, allowing vacuum to resume phase I.
23 : *
24 : * In a way, the phases are more like states in a state machine, but they have
25 : * been referred to colloquially as phases for so long that they are referred
26 : * to as such here.
27 : *
28 : * Manually invoked VACUUMs may scan indexes during phase II in parallel. For
29 : * more information on this, see the comment at the top of vacuumparallel.c.
30 : *
31 : * In between phases, vacuum updates the freespace map (every
32 : * VACUUM_FSM_EVERY_PAGES).
33 : *
34 : * After completing all three phases, vacuum may truncate the relation if it
35 : * has emptied pages at the end. Finally, vacuum updates relation statistics
36 : * in pg_class and the cumulative statistics subsystem.
37 : *
38 : * Relation Scanning:
39 : *
40 : * Vacuum scans the heap relation, starting at the beginning and progressing
41 : * to the end, skipping pages as permitted by their visibility status, vacuum
42 : * options, and various other requirements.
43 : *
44 : * Vacuums are either aggressive or normal. Aggressive vacuums must scan every
45 : * unfrozen tuple in order to advance relfrozenxid and avoid transaction ID
46 : * wraparound. Normal vacuums may scan otherwise skippable pages for one of
47 : * two reasons:
48 : *
49 : * When page skipping is not disabled, a normal vacuum may scan pages that are
50 : * marked all-visible (and even all-frozen) in the visibility map if the range
51 : * of skippable pages is below SKIP_PAGES_THRESHOLD. This is primarily for the
52 : * benefit of kernel readahead (see comment in heap_vac_scan_next_block()).
53 : *
54 : * A normal vacuum may also scan skippable pages in an effort to freeze them
55 : * and decrease the backlog of all-visible but not all-frozen pages that have
56 : * to be processed by the next aggressive vacuum. These are referred to as
57 : * eagerly scanned pages. Pages scanned due to SKIP_PAGES_THRESHOLD do not
58 : * count as eagerly scanned pages.
59 : *
60 : * Eagerly scanned pages that are set all-frozen in the VM are successful
61 : * eager freezes and those not set all-frozen in the VM are failed eager
62 : * freezes.
63 : *
64 : * Because we want to amortize the overhead of freezing pages over multiple
65 : * vacuums, normal vacuums cap the number of successful eager freezes to
66 : * MAX_EAGER_FREEZE_SUCCESS_RATE of the number of all-visible but not
67 : * all-frozen pages at the beginning of the vacuum. Since eagerly frozen pages
68 : * may be unfrozen before the next aggressive vacuum, capping the number of
69 : * successful eager freezes also caps the downside of eager freezing:
70 : * potentially wasted work.
71 : *
72 : * Once the success cap has been hit, eager scanning is disabled for the
73 : * remainder of the vacuum of the relation.
74 : *
75 : * Success is capped globally because we don't want to limit our successes if
76 : * old data happens to be concentrated in a particular part of the table. This
77 : * is especially likely to happen for append-mostly workloads where the oldest
78 : * data is at the beginning of the unfrozen portion of the relation.
79 : *
80 : * On the assumption that different regions of the table are likely to contain
81 : * similarly aged data, normal vacuums use a localized eager freeze failure
82 : * cap. The failure count is reset for each region of the table -- comprised
83 : * of EAGER_SCAN_REGION_SIZE blocks. In each region, we tolerate
84 : * vacuum_max_eager_freeze_failure_rate of EAGER_SCAN_REGION_SIZE failures
85 : * before suspending eager scanning until the end of the region.
86 : * vacuum_max_eager_freeze_failure_rate is configurable both globally and per
87 : * table.
88 : *
89 : * Aggressive vacuums must examine every unfrozen tuple and thus are not
90 : * subject to any of the limits imposed by the eager scanning algorithm.
91 : *
92 : * Once vacuum has decided to scan a given block, it must read the block and
93 : * obtain a cleanup lock to prune tuples on the page. A non-aggressive vacuum
94 : * may choose to skip pruning and freezing if it cannot acquire a cleanup lock
95 : * on the buffer right away. In this case, it may miss cleaning up dead tuples
96 : * and their associated index entries (though it is free to reap any existing
97 : * dead items on the page).
98 : *
99 : * After pruning and freezing, pages that are newly all-visible and all-frozen
100 : * are marked as such in the visibility map.
101 : *
102 : * Dead TID Storage:
103 : *
104 : * The major space usage for vacuuming is storage for the dead tuple IDs that
105 : * are to be removed from indexes. We want to ensure we can vacuum even the
106 : * very largest relations with finite memory space usage. To do that, we set
107 : * upper bounds on the memory that can be used for keeping track of dead TIDs
108 : * at once.
109 : *
110 : * We are willing to use at most maintenance_work_mem (or perhaps
111 : * autovacuum_work_mem) memory space to keep track of dead TIDs. If the
112 : * TID store is full, we must call lazy_vacuum to vacuum indexes (and to vacuum
113 : * the pages that we've pruned). This frees up the memory space dedicated to
114 : * store dead TIDs.
115 : *
116 : * In practice VACUUM will often complete its initial pass over the target
117 : * heap relation without ever running out of space to store TIDs. This means
118 : * that there only needs to be one call to lazy_vacuum, after the initial pass
119 : * completes.
120 : *
121 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
122 : * Portions Copyright (c) 1994, Regents of the University of California
123 : *
124 : *
125 : * IDENTIFICATION
126 : * src/backend/access/heap/vacuumlazy.c
127 : *
128 : *-------------------------------------------------------------------------
129 : */
130 : #include "postgres.h"
131 :
132 : #include <math.h>
133 :
134 : #include "access/genam.h"
135 : #include "access/heapam.h"
136 : #include "access/htup_details.h"
137 : #include "access/multixact.h"
138 : #include "access/tidstore.h"
139 : #include "access/transam.h"
140 : #include "access/visibilitymap.h"
141 : #include "access/xloginsert.h"
142 : #include "catalog/storage.h"
143 : #include "commands/progress.h"
144 : #include "commands/vacuum.h"
145 : #include "common/int.h"
146 : #include "common/pg_prng.h"
147 : #include "executor/instrument.h"
148 : #include "miscadmin.h"
149 : #include "pgstat.h"
150 : #include "portability/instr_time.h"
151 : #include "postmaster/autovacuum.h"
152 : #include "storage/bufmgr.h"
153 : #include "storage/freespace.h"
154 : #include "storage/lmgr.h"
155 : #include "storage/read_stream.h"
156 : #include "utils/lsyscache.h"
157 : #include "utils/pg_rusage.h"
158 : #include "utils/timestamp.h"
159 :
160 :
161 : /*
162 : * Space/time tradeoff parameters: do these need to be user-tunable?
163 : *
164 : * To consider truncating the relation, we want there to be at least
165 : * REL_TRUNCATE_MINIMUM or (relsize / REL_TRUNCATE_FRACTION) (whichever
166 : * is less) potentially-freeable pages.
167 : */
168 : #define REL_TRUNCATE_MINIMUM 1000
169 : #define REL_TRUNCATE_FRACTION 16
170 :
171 : /*
172 : * Timing parameters for truncate locking heuristics.
173 : *
174 : * These were not exposed as user tunable GUC values because it didn't seem
175 : * that the potential for improvement was great enough to merit the cost of
176 : * supporting them.
177 : */
178 : #define VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL 20 /* ms */
179 : #define VACUUM_TRUNCATE_LOCK_WAIT_INTERVAL 50 /* ms */
180 : #define VACUUM_TRUNCATE_LOCK_TIMEOUT 5000 /* ms */
181 :
182 : /*
183 : * Threshold that controls whether we bypass index vacuuming and heap
184 : * vacuuming as an optimization
185 : */
186 : #define BYPASS_THRESHOLD_PAGES 0.02 /* i.e. 2% of rel_pages */
187 :
188 : /*
189 : * Perform a failsafe check each time we scan another 4GB of pages.
190 : * (Note that this is deliberately kept to a power-of-two, usually 2^19.)
191 : */
192 : #define FAILSAFE_EVERY_PAGES \
193 : ((BlockNumber) (((uint64) 4 * 1024 * 1024 * 1024) / BLCKSZ))
194 :
195 : /*
196 : * When a table has no indexes, vacuum the FSM after every 8GB, approximately
197 : * (it won't be exact because we only vacuum FSM after processing a heap page
198 : * that has some removable tuples). When there are indexes, this is ignored,
199 : * and we vacuum FSM after each index/heap cleaning pass.
200 : */
201 : #define VACUUM_FSM_EVERY_PAGES \
202 : ((BlockNumber) (((uint64) 8 * 1024 * 1024 * 1024) / BLCKSZ))
203 :
204 : /*
205 : * Before we consider skipping a page that's marked as clean in
206 : * visibility map, we must've seen at least this many clean pages.
207 : */
208 : #define SKIP_PAGES_THRESHOLD ((BlockNumber) 32)
209 :
210 : /*
211 : * Size of the prefetch window for lazy vacuum backwards truncation scan.
212 : * Needs to be a power of 2.
213 : */
214 : #define PREFETCH_SIZE ((BlockNumber) 32)
215 :
216 : /*
217 : * Macro to check if we are in a parallel vacuum. If true, we are in the
218 : * parallel mode and the DSM segment is initialized.
219 : */
220 : #define ParallelVacuumIsActive(vacrel) ((vacrel)->pvs != NULL)
221 :
222 : /* Phases of vacuum during which we report error context. */
223 : typedef enum
224 : {
225 : VACUUM_ERRCB_PHASE_UNKNOWN,
226 : VACUUM_ERRCB_PHASE_SCAN_HEAP,
227 : VACUUM_ERRCB_PHASE_VACUUM_INDEX,
228 : VACUUM_ERRCB_PHASE_VACUUM_HEAP,
229 : VACUUM_ERRCB_PHASE_INDEX_CLEANUP,
230 : VACUUM_ERRCB_PHASE_TRUNCATE,
231 : } VacErrPhase;
232 :
233 : /*
234 : * An eager scan of a page that is set all-frozen in the VM is considered
235 : * "successful". To spread out freezing overhead across multiple normal
236 : * vacuums, we limit the number of successful eager page freezes. The maximum
237 : * number of eager page freezes is calculated as a ratio of the all-visible
238 : * but not all-frozen pages at the beginning of the vacuum.
239 : */
240 : #define MAX_EAGER_FREEZE_SUCCESS_RATE 0.2
241 :
242 : /*
243 : * On the assumption that different regions of the table tend to have
244 : * similarly aged data, once vacuum fails to freeze
245 : * vacuum_max_eager_freeze_failure_rate of the blocks in a region of size
246 : * EAGER_SCAN_REGION_SIZE, it suspends eager scanning until it has progressed
247 : * to another region of the table with potentially older data.
248 : */
249 : #define EAGER_SCAN_REGION_SIZE 4096
250 :
251 : /*
252 : * heap_vac_scan_next_block() sets these flags to communicate information
253 : * about the block it read to the caller.
254 : */
255 : #define VAC_BLK_WAS_EAGER_SCANNED (1 << 0)
256 : #define VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM (1 << 1)
257 :
258 : typedef struct LVRelState
259 : {
260 : /* Target heap relation and its indexes */
261 : Relation rel;
262 : Relation *indrels;
263 : int nindexes;
264 :
265 : /* Buffer access strategy and parallel vacuum state */
266 : BufferAccessStrategy bstrategy;
267 : ParallelVacuumState *pvs;
268 :
269 : /* Aggressive VACUUM? (must set relfrozenxid >= FreezeLimit) */
270 : bool aggressive;
271 : /* Use visibility map to skip? (disabled by DISABLE_PAGE_SKIPPING) */
272 : bool skipwithvm;
273 : /* Consider index vacuuming bypass optimization? */
274 : bool consider_bypass_optimization;
275 :
276 : /* Doing index vacuuming, index cleanup, rel truncation? */
277 : bool do_index_vacuuming;
278 : bool do_index_cleanup;
279 : bool do_rel_truncate;
280 :
281 : /* VACUUM operation's cutoffs for freezing and pruning */
282 : struct VacuumCutoffs cutoffs;
283 : GlobalVisState *vistest;
284 : /* Tracks oldest extant XID/MXID for setting relfrozenxid/relminmxid */
285 : TransactionId NewRelfrozenXid;
286 : MultiXactId NewRelminMxid;
287 : bool skippedallvis;
288 :
289 : /* Error reporting state */
290 : char *dbname;
291 : char *relnamespace;
292 : char *relname;
293 : char *indname; /* Current index name */
294 : BlockNumber blkno; /* used only for heap operations */
295 : OffsetNumber offnum; /* used only for heap operations */
296 : VacErrPhase phase;
297 : bool verbose; /* VACUUM VERBOSE? */
298 :
299 : /*
300 : * dead_items stores TIDs whose index tuples are deleted by index
301 : * vacuuming. Each TID points to an LP_DEAD line pointer from a heap page
302 : * that has been processed by lazy_scan_prune. Also needed by
303 : * lazy_vacuum_heap_rel, which marks the same LP_DEAD line pointers as
304 : * LP_UNUSED during second heap pass.
305 : *
306 : * Both dead_items and dead_items_info are allocated in shared memory in
307 : * parallel vacuum cases.
308 : */
309 : TidStore *dead_items; /* TIDs whose index tuples we'll delete */
310 : VacDeadItemsInfo *dead_items_info;
311 :
312 : BlockNumber rel_pages; /* total number of pages */
313 : BlockNumber scanned_pages; /* # pages examined (not skipped via VM) */
314 :
315 : /*
316 : * Count of all-visible blocks eagerly scanned (for logging only). This
317 : * does not include skippable blocks scanned due to SKIP_PAGES_THRESHOLD.
318 : */
319 : BlockNumber eager_scanned_pages;
320 :
321 : BlockNumber removed_pages; /* # pages removed by relation truncation */
322 : BlockNumber new_frozen_tuple_pages; /* # pages with newly frozen tuples */
323 :
324 : /* # pages newly set all-visible in the VM */
325 : BlockNumber vm_new_visible_pages;
326 :
327 : /*
328 : * # pages newly set all-visible and all-frozen in the VM. This is a
329 : * subset of vm_new_visible_pages. That is, vm_new_visible_pages includes
330 : * all pages set all-visible, but vm_new_visible_frozen_pages includes
331 : * only those which were also set all-frozen.
332 : */
333 : BlockNumber vm_new_visible_frozen_pages;
334 :
335 : /* # all-visible pages newly set all-frozen in the VM */
336 : BlockNumber vm_new_frozen_pages;
337 :
338 : BlockNumber lpdead_item_pages; /* # pages with LP_DEAD items */
339 : BlockNumber missed_dead_pages; /* # pages with missed dead tuples */
340 : BlockNumber nonempty_pages; /* actually, last nonempty page + 1 */
341 :
342 : /* Statistics output by us, for table */
343 : double new_rel_tuples; /* new estimated total # of tuples */
344 : double new_live_tuples; /* new estimated total # of live tuples */
345 : /* Statistics output by index AMs */
346 : IndexBulkDeleteResult **indstats;
347 :
348 : /* Instrumentation counters */
349 : int num_index_scans;
350 : /* Counters that follow are only for scanned_pages */
351 : int64 tuples_deleted; /* # deleted from table */
352 : int64 tuples_frozen; /* # newly frozen */
353 : int64 lpdead_items; /* # deleted from indexes */
354 : int64 live_tuples; /* # live tuples remaining */
355 : int64 recently_dead_tuples; /* # dead, but not yet removable */
356 : int64 missed_dead_tuples; /* # removable, but not removed */
357 :
358 : /* State maintained by heap_vac_scan_next_block() */
359 : BlockNumber current_block; /* last block returned */
360 : BlockNumber next_unskippable_block; /* next unskippable block */
361 : bool next_unskippable_allvis; /* its visibility status */
362 : bool next_unskippable_eager_scanned; /* if it was eagerly scanned */
363 : Buffer next_unskippable_vmbuffer; /* buffer containing its VM bit */
364 :
365 : /* State related to managing eager scanning of all-visible pages */
366 :
367 : /*
368 : * A normal vacuum that has failed to freeze too many eagerly scanned
369 : * blocks in a region suspends eager scanning.
370 : * next_eager_scan_region_start is the block number of the first block
371 : * eligible for resumed eager scanning.
372 : *
373 : * When eager scanning is permanently disabled, either initially
374 : * (including for aggressive vacuum) or due to hitting the success cap,
375 : * this is set to InvalidBlockNumber.
376 : */
377 : BlockNumber next_eager_scan_region_start;
378 :
379 : /*
380 : * The remaining number of blocks a normal vacuum will consider eager
381 : * scanning when it is successful. When eager scanning is enabled, this is
382 : * initialized to MAX_EAGER_FREEZE_SUCCESS_RATE of the total number of
383 : * all-visible but not all-frozen pages. For each eager freeze success,
384 : * this is decremented. Once it hits 0, eager scanning is permanently
385 : * disabled. It is initialized to 0 if eager scanning starts out disabled
386 : * (including for aggressive vacuum).
387 : */
388 : BlockNumber eager_scan_remaining_successes;
389 :
390 : /*
391 : * The maximum number of blocks which may be eagerly scanned and not
392 : * frozen before eager scanning is temporarily suspended. This is
393 : * configurable both globally, via the
394 : * vacuum_max_eager_freeze_failure_rate GUC, and per table, with a table
395 : * storage parameter of the same name. It is calculated as
396 : * vacuum_max_eager_freeze_failure_rate of EAGER_SCAN_REGION_SIZE blocks.
397 : * It is 0 when eager scanning is disabled.
398 : */
399 : BlockNumber eager_scan_max_fails_per_region;
400 :
401 : /*
402 : * The number of eagerly scanned blocks vacuum failed to freeze (due to
403 : * age) in the current eager scan region. Vacuum resets it to
404 : * eager_scan_max_fails_per_region each time it enters a new region of the
405 : * relation. If eager_scan_remaining_fails hits 0, eager scanning is
406 : * suspended until the next region. It is also 0 if eager scanning has
407 : * been permanently disabled.
408 : */
409 : BlockNumber eager_scan_remaining_fails;
410 : } LVRelState;
411 :
412 :
413 : /* Struct for saving and restoring vacuum error information. */
414 : typedef struct LVSavedErrInfo
415 : {
416 : BlockNumber blkno;
417 : OffsetNumber offnum;
418 : VacErrPhase phase;
419 : } LVSavedErrInfo;
420 :
421 :
422 : /* non-export function prototypes */
423 : static void lazy_scan_heap(LVRelState *vacrel);
424 : static void heap_vacuum_eager_scan_setup(LVRelState *vacrel,
425 : const VacuumParams params);
426 : static BlockNumber heap_vac_scan_next_block(ReadStream *stream,
427 : void *callback_private_data,
428 : void *per_buffer_data);
429 : static void find_next_unskippable_block(LVRelState *vacrel, bool *skipsallvis);
430 : static bool lazy_scan_new_or_empty(LVRelState *vacrel, Buffer buf,
431 : BlockNumber blkno, Page page,
432 : bool sharelock, Buffer vmbuffer);
433 : static int lazy_scan_prune(LVRelState *vacrel, Buffer buf,
434 : BlockNumber blkno, Page page,
435 : Buffer vmbuffer, bool all_visible_according_to_vm,
436 : bool *has_lpdead_items, bool *vm_page_frozen);
437 : static bool lazy_scan_noprune(LVRelState *vacrel, Buffer buf,
438 : BlockNumber blkno, Page page,
439 : bool *has_lpdead_items);
440 : static void lazy_vacuum(LVRelState *vacrel);
441 : static bool lazy_vacuum_all_indexes(LVRelState *vacrel);
442 : static void lazy_vacuum_heap_rel(LVRelState *vacrel);
443 : static void lazy_vacuum_heap_page(LVRelState *vacrel, BlockNumber blkno,
444 : Buffer buffer, OffsetNumber *deadoffsets,
445 : int num_offsets, Buffer vmbuffer);
446 : static bool lazy_check_wraparound_failsafe(LVRelState *vacrel);
447 : static void lazy_cleanup_all_indexes(LVRelState *vacrel);
448 : static IndexBulkDeleteResult *lazy_vacuum_one_index(Relation indrel,
449 : IndexBulkDeleteResult *istat,
450 : double reltuples,
451 : LVRelState *vacrel);
452 : static IndexBulkDeleteResult *lazy_cleanup_one_index(Relation indrel,
453 : IndexBulkDeleteResult *istat,
454 : double reltuples,
455 : bool estimated_count,
456 : LVRelState *vacrel);
457 : static bool should_attempt_truncation(LVRelState *vacrel);
458 : static void lazy_truncate_heap(LVRelState *vacrel);
459 : static BlockNumber count_nondeletable_pages(LVRelState *vacrel,
460 : bool *lock_waiter_detected);
461 : static void dead_items_alloc(LVRelState *vacrel, int nworkers);
462 : static void dead_items_add(LVRelState *vacrel, BlockNumber blkno, OffsetNumber *offsets,
463 : int num_offsets);
464 : static void dead_items_reset(LVRelState *vacrel);
465 : static void dead_items_cleanup(LVRelState *vacrel);
466 :
467 : #ifdef USE_ASSERT_CHECKING
468 : static bool heap_page_is_all_visible(Relation rel, Buffer buf,
469 : TransactionId OldestXmin,
470 : bool *all_frozen,
471 : TransactionId *visibility_cutoff_xid,
472 : OffsetNumber *logging_offnum);
473 : #endif
474 : static bool heap_page_would_be_all_visible(Relation rel, Buffer buf,
475 : TransactionId OldestXmin,
476 : OffsetNumber *deadoffsets,
477 : int ndeadoffsets,
478 : bool *all_frozen,
479 : TransactionId *visibility_cutoff_xid,
480 : OffsetNumber *logging_offnum);
481 : static void update_relstats_all_indexes(LVRelState *vacrel);
482 : static void vacuum_error_callback(void *arg);
483 : static void update_vacuum_error_info(LVRelState *vacrel,
484 : LVSavedErrInfo *saved_vacrel,
485 : int phase, BlockNumber blkno,
486 : OffsetNumber offnum);
487 : static void restore_vacuum_error_info(LVRelState *vacrel,
488 : const LVSavedErrInfo *saved_vacrel);
489 :
490 :
491 :
492 : /*
493 : * Helper to set up the eager scanning state for vacuuming a single relation.
494 : * Initializes the eager scan management related members of the LVRelState.
495 : *
496 : * Caller provides whether or not an aggressive vacuum is required due to
497 : * vacuum options or for relfrozenxid/relminmxid advancement.
498 : */
499 : static void
500 257572 : heap_vacuum_eager_scan_setup(LVRelState *vacrel, const VacuumParams params)
501 : {
502 : uint32 randseed;
503 : BlockNumber allvisible;
504 : BlockNumber allfrozen;
505 : float first_region_ratio;
506 257572 : bool oldest_unfrozen_before_cutoff = false;
507 :
508 : /*
509 : * Initialize eager scan management fields to their disabled values.
510 : * Aggressive vacuums, normal vacuums of small tables, and normal vacuums
511 : * of tables without sufficiently old tuples disable eager scanning.
512 : */
513 257572 : vacrel->next_eager_scan_region_start = InvalidBlockNumber;
514 257572 : vacrel->eager_scan_max_fails_per_region = 0;
515 257572 : vacrel->eager_scan_remaining_fails = 0;
516 257572 : vacrel->eager_scan_remaining_successes = 0;
517 :
518 : /* If eager scanning is explicitly disabled, just return. */
519 257572 : if (params.max_eager_freeze_failure_rate == 0)
520 257572 : return;
521 :
522 : /*
523 : * The caller will have determined whether or not an aggressive vacuum is
524 : * required by either the vacuum parameters or the relative age of the
525 : * oldest unfrozen transaction IDs. An aggressive vacuum must scan every
526 : * all-visible page to safely advance the relfrozenxid and/or relminmxid,
527 : * so scans of all-visible pages are not considered eager.
528 : */
529 257572 : if (vacrel->aggressive)
530 246854 : return;
531 :
532 : /*
533 : * Aggressively vacuuming a small relation shouldn't take long, so it
534 : * isn't worth amortizing. We use two times the region size as the size
535 : * cutoff because the eager scan start block is a random spot somewhere in
536 : * the first region, making the second region the first to be eager
537 : * scanned normally.
538 : */
539 10718 : if (vacrel->rel_pages < 2 * EAGER_SCAN_REGION_SIZE)
540 10718 : return;
541 :
542 : /*
543 : * We only want to enable eager scanning if we are likely to be able to
544 : * freeze some of the pages in the relation.
545 : *
546 : * Tuples with XIDs older than OldestXmin or MXIDs older than OldestMxact
547 : * are technically freezable, but we won't freeze them unless the criteria
548 : * for opportunistic freezing is met. Only tuples with XIDs/MXIDs older
549 : * than the FreezeLimit/MultiXactCutoff are frozen in the common case.
550 : *
551 : * So, as a heuristic, we wait until the FreezeLimit has advanced past the
552 : * relfrozenxid or the MultiXactCutoff has advanced past the relminmxid to
553 : * enable eager scanning.
554 : */
555 0 : if (TransactionIdIsNormal(vacrel->cutoffs.relfrozenxid) &&
556 0 : TransactionIdPrecedes(vacrel->cutoffs.relfrozenxid,
557 : vacrel->cutoffs.FreezeLimit))
558 0 : oldest_unfrozen_before_cutoff = true;
559 :
560 0 : if (!oldest_unfrozen_before_cutoff &&
561 0 : MultiXactIdIsValid(vacrel->cutoffs.relminmxid) &&
562 0 : MultiXactIdPrecedes(vacrel->cutoffs.relminmxid,
563 : vacrel->cutoffs.MultiXactCutoff))
564 0 : oldest_unfrozen_before_cutoff = true;
565 :
566 0 : if (!oldest_unfrozen_before_cutoff)
567 0 : return;
568 :
569 : /* We have met the criteria to eagerly scan some pages. */
570 :
571 : /*
572 : * Our success cap is MAX_EAGER_FREEZE_SUCCESS_RATE of the number of
573 : * all-visible but not all-frozen blocks in the relation.
574 : */
575 0 : visibilitymap_count(vacrel->rel, &allvisible, &allfrozen);
576 :
577 0 : vacrel->eager_scan_remaining_successes =
578 0 : (BlockNumber) (MAX_EAGER_FREEZE_SUCCESS_RATE *
579 0 : (allvisible - allfrozen));
580 :
581 : /* If every all-visible page is frozen, eager scanning is disabled. */
582 0 : if (vacrel->eager_scan_remaining_successes == 0)
583 0 : return;
584 :
585 : /*
586 : * Now calculate the bounds of the first eager scan region. Its end block
587 : * will be a random spot somewhere in the first EAGER_SCAN_REGION_SIZE
588 : * blocks. This affects the bounds of all subsequent regions and avoids
589 : * eager scanning and failing to freeze the same blocks each vacuum of the
590 : * relation.
591 : */
592 0 : randseed = pg_prng_uint32(&pg_global_prng_state);
593 :
594 0 : vacrel->next_eager_scan_region_start = randseed % EAGER_SCAN_REGION_SIZE;
595 :
596 : Assert(params.max_eager_freeze_failure_rate > 0 &&
597 : params.max_eager_freeze_failure_rate <= 1);
598 :
599 0 : vacrel->eager_scan_max_fails_per_region =
600 0 : params.max_eager_freeze_failure_rate *
601 : EAGER_SCAN_REGION_SIZE;
602 :
603 : /*
604 : * The first region will be smaller than subsequent regions. As such,
605 : * adjust the eager freeze failures tolerated for this region.
606 : */
607 0 : first_region_ratio = 1 - (float) vacrel->next_eager_scan_region_start /
608 : EAGER_SCAN_REGION_SIZE;
609 :
610 0 : vacrel->eager_scan_remaining_fails =
611 0 : vacrel->eager_scan_max_fails_per_region *
612 : first_region_ratio;
613 : }
614 :
615 : /*
616 : * heap_vacuum_rel() -- perform VACUUM for one heap relation
617 : *
618 : * This routine sets things up for and then calls lazy_scan_heap, where
619 : * almost all work actually takes place. Finalizes everything after call
620 : * returns by managing relation truncation and updating rel's pg_class
621 : * entry. (Also updates pg_class entries for any indexes that need it.)
622 : *
623 : * At entry, we have already established a transaction and opened
624 : * and locked the relation.
625 : */
626 : void
627 257572 : heap_vacuum_rel(Relation rel, const VacuumParams params,
628 : BufferAccessStrategy bstrategy)
629 : {
630 : LVRelState *vacrel;
631 : bool verbose,
632 : instrument,
633 : skipwithvm,
634 : frozenxid_updated,
635 : minmulti_updated;
636 : BlockNumber orig_rel_pages,
637 : new_rel_pages,
638 : new_rel_allvisible,
639 : new_rel_allfrozen;
640 : PGRUsage ru0;
641 257572 : TimestampTz starttime = 0;
642 257572 : PgStat_Counter startreadtime = 0,
643 257572 : startwritetime = 0;
644 257572 : WalUsage startwalusage = pgWalUsage;
645 257572 : BufferUsage startbufferusage = pgBufferUsage;
646 : ErrorContextCallback errcallback;
647 257572 : char **indnames = NULL;
648 :
649 257572 : verbose = (params.options & VACOPT_VERBOSE) != 0;
650 489150 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
651 231578 : params.log_vacuum_min_duration >= 0));
652 257572 : if (instrument)
653 : {
654 231602 : pg_rusage_init(&ru0);
655 231602 : if (track_io_timing)
656 : {
657 0 : startreadtime = pgStatBlockReadTime;
658 0 : startwritetime = pgStatBlockWriteTime;
659 : }
660 : }
661 :
662 : /* Used for instrumentation and stats report */
663 257572 : starttime = GetCurrentTimestamp();
664 :
665 257572 : pgstat_progress_start_command(PROGRESS_COMMAND_VACUUM,
666 : RelationGetRelid(rel));
667 :
668 : /*
669 : * Setup error traceback support for ereport() first. The idea is to set
670 : * up an error context callback to display additional information on any
671 : * error during a vacuum. During different phases of vacuum, we update
672 : * the state so that the error context callback always display current
673 : * information.
674 : *
675 : * Copy the names of heap rel into local memory for error reporting
676 : * purposes, too. It isn't always safe to assume that we can get the name
677 : * of each rel. It's convenient for code in lazy_scan_heap to always use
678 : * these temp copies.
679 : */
680 257572 : vacrel = (LVRelState *) palloc0(sizeof(LVRelState));
681 257572 : vacrel->dbname = get_database_name(MyDatabaseId);
682 257572 : vacrel->relnamespace = get_namespace_name(RelationGetNamespace(rel));
683 257572 : vacrel->relname = pstrdup(RelationGetRelationName(rel));
684 257572 : vacrel->indname = NULL;
685 257572 : vacrel->phase = VACUUM_ERRCB_PHASE_UNKNOWN;
686 257572 : vacrel->verbose = verbose;
687 257572 : errcallback.callback = vacuum_error_callback;
688 257572 : errcallback.arg = vacrel;
689 257572 : errcallback.previous = error_context_stack;
690 257572 : error_context_stack = &errcallback;
691 :
692 : /* Set up high level stuff about rel and its indexes */
693 257572 : vacrel->rel = rel;
694 257572 : vac_open_indexes(vacrel->rel, RowExclusiveLock, &vacrel->nindexes,
695 : &vacrel->indrels);
696 257572 : vacrel->bstrategy = bstrategy;
697 257572 : if (instrument && vacrel->nindexes > 0)
698 : {
699 : /* Copy index names used by instrumentation (not error reporting) */
700 221520 : indnames = palloc(sizeof(char *) * vacrel->nindexes);
701 569992 : for (int i = 0; i < vacrel->nindexes; i++)
702 348472 : indnames[i] = pstrdup(RelationGetRelationName(vacrel->indrels[i]));
703 : }
704 :
705 : /*
706 : * The index_cleanup param either disables index vacuuming and cleanup or
707 : * forces it to go ahead when we would otherwise apply the index bypass
708 : * optimization. The default is 'auto', which leaves the final decision
709 : * up to lazy_vacuum().
710 : *
711 : * The truncate param allows user to avoid attempting relation truncation,
712 : * though it can't force truncation to happen.
713 : */
714 : Assert(params.index_cleanup != VACOPTVALUE_UNSPECIFIED);
715 : Assert(params.truncate != VACOPTVALUE_UNSPECIFIED &&
716 : params.truncate != VACOPTVALUE_AUTO);
717 :
718 : /*
719 : * While VacuumFailSafeActive is reset to false before calling this, we
720 : * still need to reset it here due to recursive calls.
721 : */
722 257572 : VacuumFailsafeActive = false;
723 257572 : vacrel->consider_bypass_optimization = true;
724 257572 : vacrel->do_index_vacuuming = true;
725 257572 : vacrel->do_index_cleanup = true;
726 257572 : vacrel->do_rel_truncate = (params.truncate != VACOPTVALUE_DISABLED);
727 257572 : if (params.index_cleanup == VACOPTVALUE_DISABLED)
728 : {
729 : /* Force disable index vacuuming up-front */
730 260 : vacrel->do_index_vacuuming = false;
731 260 : vacrel->do_index_cleanup = false;
732 : }
733 257312 : else if (params.index_cleanup == VACOPTVALUE_ENABLED)
734 : {
735 : /* Force index vacuuming. Note that failsafe can still bypass. */
736 30 : vacrel->consider_bypass_optimization = false;
737 : }
738 : else
739 : {
740 : /* Default/auto, make all decisions dynamically */
741 : Assert(params.index_cleanup == VACOPTVALUE_AUTO);
742 : }
743 :
744 : /* Initialize page counters explicitly (be tidy) */
745 257572 : vacrel->scanned_pages = 0;
746 257572 : vacrel->eager_scanned_pages = 0;
747 257572 : vacrel->removed_pages = 0;
748 257572 : vacrel->new_frozen_tuple_pages = 0;
749 257572 : vacrel->lpdead_item_pages = 0;
750 257572 : vacrel->missed_dead_pages = 0;
751 257572 : vacrel->nonempty_pages = 0;
752 : /* dead_items_alloc allocates vacrel->dead_items later on */
753 :
754 : /* Allocate/initialize output statistics state */
755 257572 : vacrel->new_rel_tuples = 0;
756 257572 : vacrel->new_live_tuples = 0;
757 257572 : vacrel->indstats = (IndexBulkDeleteResult **)
758 257572 : palloc0(vacrel->nindexes * sizeof(IndexBulkDeleteResult *));
759 :
760 : /* Initialize remaining counters (be tidy) */
761 257572 : vacrel->num_index_scans = 0;
762 257572 : vacrel->tuples_deleted = 0;
763 257572 : vacrel->tuples_frozen = 0;
764 257572 : vacrel->lpdead_items = 0;
765 257572 : vacrel->live_tuples = 0;
766 257572 : vacrel->recently_dead_tuples = 0;
767 257572 : vacrel->missed_dead_tuples = 0;
768 :
769 257572 : vacrel->vm_new_visible_pages = 0;
770 257572 : vacrel->vm_new_visible_frozen_pages = 0;
771 257572 : vacrel->vm_new_frozen_pages = 0;
772 :
773 : /*
774 : * Get cutoffs that determine which deleted tuples are considered DEAD,
775 : * not just RECENTLY_DEAD, and which XIDs/MXIDs to freeze. Then determine
776 : * the extent of the blocks that we'll scan in lazy_scan_heap. It has to
777 : * happen in this order to ensure that the OldestXmin cutoff field works
778 : * as an upper bound on the XIDs stored in the pages we'll actually scan
779 : * (NewRelfrozenXid tracking must never be allowed to miss unfrozen XIDs).
780 : *
781 : * Next acquire vistest, a related cutoff that's used in pruning. We use
782 : * vistest in combination with OldestXmin to ensure that
783 : * heap_page_prune_and_freeze() always removes any deleted tuple whose
784 : * xmax is < OldestXmin. lazy_scan_prune must never become confused about
785 : * whether a tuple should be frozen or removed. (In the future we might
786 : * want to teach lazy_scan_prune to recompute vistest from time to time,
787 : * to increase the number of dead tuples it can prune away.)
788 : */
789 257572 : vacrel->aggressive = vacuum_get_cutoffs(rel, params, &vacrel->cutoffs);
790 257572 : vacrel->rel_pages = orig_rel_pages = RelationGetNumberOfBlocks(rel);
791 257572 : vacrel->vistest = GlobalVisTestFor(rel);
792 :
793 : /* Initialize state used to track oldest extant XID/MXID */
794 257572 : vacrel->NewRelfrozenXid = vacrel->cutoffs.OldestXmin;
795 257572 : vacrel->NewRelminMxid = vacrel->cutoffs.OldestMxact;
796 :
797 : /*
798 : * Initialize state related to tracking all-visible page skipping. This is
799 : * very important to determine whether or not it is safe to advance the
800 : * relfrozenxid/relminmxid.
801 : */
802 257572 : vacrel->skippedallvis = false;
803 257572 : skipwithvm = true;
804 257572 : if (params.options & VACOPT_DISABLE_PAGE_SKIPPING)
805 : {
806 : /*
807 : * Force aggressive mode, and disable skipping blocks using the
808 : * visibility map (even those set all-frozen)
809 : */
810 344 : vacrel->aggressive = true;
811 344 : skipwithvm = false;
812 : }
813 :
814 257572 : vacrel->skipwithvm = skipwithvm;
815 :
816 : /*
817 : * Set up eager scan tracking state. This must happen after determining
818 : * whether or not the vacuum must be aggressive, because only normal
819 : * vacuums use the eager scan algorithm.
820 : */
821 257572 : heap_vacuum_eager_scan_setup(vacrel, params);
822 :
823 257572 : if (verbose)
824 : {
825 24 : if (vacrel->aggressive)
826 2 : ereport(INFO,
827 : (errmsg("aggressively vacuuming \"%s.%s.%s\"",
828 : vacrel->dbname, vacrel->relnamespace,
829 : vacrel->relname)));
830 : else
831 22 : ereport(INFO,
832 : (errmsg("vacuuming \"%s.%s.%s\"",
833 : vacrel->dbname, vacrel->relnamespace,
834 : vacrel->relname)));
835 : }
836 :
837 : /*
838 : * Allocate dead_items memory using dead_items_alloc. This handles
839 : * parallel VACUUM initialization as part of allocating shared memory
840 : * space used for dead_items. (But do a failsafe precheck first, to
841 : * ensure that parallel VACUUM won't be attempted at all when relfrozenxid
842 : * is already dangerously old.)
843 : */
844 257572 : lazy_check_wraparound_failsafe(vacrel);
845 257572 : dead_items_alloc(vacrel, params.nworkers);
846 :
847 : /*
848 : * Call lazy_scan_heap to perform all required heap pruning, index
849 : * vacuuming, and heap vacuuming (plus related processing)
850 : */
851 257572 : lazy_scan_heap(vacrel);
852 :
853 : /*
854 : * Free resources managed by dead_items_alloc. This ends parallel mode in
855 : * passing when necessary.
856 : */
857 257572 : dead_items_cleanup(vacrel);
858 : Assert(!IsInParallelMode());
859 :
860 : /*
861 : * Update pg_class entries for each of rel's indexes where appropriate.
862 : *
863 : * Unlike the later update to rel's pg_class entry, this is not critical.
864 : * Maintains relpages/reltuples statistics used by the planner only.
865 : */
866 257572 : if (vacrel->do_index_cleanup)
867 187374 : update_relstats_all_indexes(vacrel);
868 :
869 : /* Done with rel's indexes */
870 257572 : vac_close_indexes(vacrel->nindexes, vacrel->indrels, NoLock);
871 :
872 : /* Optionally truncate rel */
873 257572 : if (should_attempt_truncation(vacrel))
874 300 : lazy_truncate_heap(vacrel);
875 :
876 : /* Pop the error context stack */
877 257572 : error_context_stack = errcallback.previous;
878 :
879 : /* Report that we are now doing final cleanup */
880 257572 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
881 : PROGRESS_VACUUM_PHASE_FINAL_CLEANUP);
882 :
883 : /*
884 : * Prepare to update rel's pg_class entry.
885 : *
886 : * Aggressive VACUUMs must always be able to advance relfrozenxid to a
887 : * value >= FreezeLimit, and relminmxid to a value >= MultiXactCutoff.
888 : * Non-aggressive VACUUMs may advance them by any amount, or not at all.
889 : */
890 : Assert(vacrel->NewRelfrozenXid == vacrel->cutoffs.OldestXmin ||
891 : TransactionIdPrecedesOrEquals(vacrel->aggressive ? vacrel->cutoffs.FreezeLimit :
892 : vacrel->cutoffs.relfrozenxid,
893 : vacrel->NewRelfrozenXid));
894 : Assert(vacrel->NewRelminMxid == vacrel->cutoffs.OldestMxact ||
895 : MultiXactIdPrecedesOrEquals(vacrel->aggressive ? vacrel->cutoffs.MultiXactCutoff :
896 : vacrel->cutoffs.relminmxid,
897 : vacrel->NewRelminMxid));
898 257572 : if (vacrel->skippedallvis)
899 : {
900 : /*
901 : * Must keep original relfrozenxid in a non-aggressive VACUUM that
902 : * chose to skip an all-visible page range. The state that tracks new
903 : * values will have missed unfrozen XIDs from the pages we skipped.
904 : */
905 : Assert(!vacrel->aggressive);
906 60 : vacrel->NewRelfrozenXid = InvalidTransactionId;
907 60 : vacrel->NewRelminMxid = InvalidMultiXactId;
908 : }
909 :
910 : /*
911 : * For safety, clamp relallvisible to be not more than what we're setting
912 : * pg_class.relpages to
913 : */
914 257572 : new_rel_pages = vacrel->rel_pages; /* After possible rel truncation */
915 257572 : visibilitymap_count(rel, &new_rel_allvisible, &new_rel_allfrozen);
916 257572 : if (new_rel_allvisible > new_rel_pages)
917 0 : new_rel_allvisible = new_rel_pages;
918 :
919 : /*
920 : * An all-frozen block _must_ be all-visible. As such, clamp the count of
921 : * all-frozen blocks to the count of all-visible blocks. This matches the
922 : * clamping of relallvisible above.
923 : */
924 257572 : if (new_rel_allfrozen > new_rel_allvisible)
925 0 : new_rel_allfrozen = new_rel_allvisible;
926 :
927 : /*
928 : * Now actually update rel's pg_class entry.
929 : *
930 : * In principle new_live_tuples could be -1 indicating that we (still)
931 : * don't know the tuple count. In practice that can't happen, since we
932 : * scan every page that isn't skipped using the visibility map.
933 : */
934 257572 : vac_update_relstats(rel, new_rel_pages, vacrel->new_live_tuples,
935 : new_rel_allvisible, new_rel_allfrozen,
936 257572 : vacrel->nindexes > 0,
937 : vacrel->NewRelfrozenXid, vacrel->NewRelminMxid,
938 : &frozenxid_updated, &minmulti_updated, false);
939 :
940 : /*
941 : * Report results to the cumulative stats system, too.
942 : *
943 : * Deliberately avoid telling the stats system about LP_DEAD items that
944 : * remain in the table due to VACUUM bypassing index and heap vacuuming.
945 : * ANALYZE will consider the remaining LP_DEAD items to be dead "tuples".
946 : * It seems like a good idea to err on the side of not vacuuming again too
947 : * soon in cases where the failsafe prevented significant amounts of heap
948 : * vacuuming.
949 : */
950 154616 : pgstat_report_vacuum(RelationGetRelid(rel),
951 257572 : rel->rd_rel->relisshared,
952 102956 : Max(vacrel->new_live_tuples, 0),
953 257572 : vacrel->recently_dead_tuples +
954 257572 : vacrel->missed_dead_tuples,
955 : starttime);
956 257572 : pgstat_progress_end_command();
957 :
958 257572 : if (instrument)
959 : {
960 231602 : TimestampTz endtime = GetCurrentTimestamp();
961 :
962 231804 : if (verbose || params.log_vacuum_min_duration == 0 ||
963 202 : TimestampDifferenceExceeds(starttime, endtime,
964 202 : params.log_vacuum_min_duration))
965 : {
966 : long secs_dur;
967 : int usecs_dur;
968 : WalUsage walusage;
969 : BufferUsage bufferusage;
970 : StringInfoData buf;
971 : char *msgfmt;
972 : int32 diff;
973 231400 : double read_rate = 0,
974 231400 : write_rate = 0;
975 : int64 total_blks_hit;
976 : int64 total_blks_read;
977 : int64 total_blks_dirtied;
978 :
979 231400 : TimestampDifference(starttime, endtime, &secs_dur, &usecs_dur);
980 231400 : memset(&walusage, 0, sizeof(WalUsage));
981 231400 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
982 231400 : memset(&bufferusage, 0, sizeof(BufferUsage));
983 231400 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
984 :
985 231400 : total_blks_hit = bufferusage.shared_blks_hit +
986 231400 : bufferusage.local_blks_hit;
987 231400 : total_blks_read = bufferusage.shared_blks_read +
988 231400 : bufferusage.local_blks_read;
989 231400 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
990 231400 : bufferusage.local_blks_dirtied;
991 :
992 231400 : initStringInfo(&buf);
993 231400 : if (verbose)
994 : {
995 : /*
996 : * Aggressiveness already reported earlier, in dedicated
997 : * VACUUM VERBOSE ereport
998 : */
999 : Assert(!params.is_wraparound);
1000 24 : msgfmt = _("finished vacuuming \"%s.%s.%s\": index scans: %d\n");
1001 : }
1002 231376 : else if (params.is_wraparound)
1003 : {
1004 : /*
1005 : * While it's possible for a VACUUM to be both is_wraparound
1006 : * and !aggressive, that's just a corner-case -- is_wraparound
1007 : * implies aggressive. Produce distinct output for the corner
1008 : * case all the same, just in case.
1009 : */
1010 231336 : if (vacrel->aggressive)
1011 231320 : msgfmt = _("automatic aggressive vacuum to prevent wraparound of table \"%s.%s.%s\": index scans: %d\n");
1012 : else
1013 16 : msgfmt = _("automatic vacuum to prevent wraparound of table \"%s.%s.%s\": index scans: %d\n");
1014 : }
1015 : else
1016 : {
1017 40 : if (vacrel->aggressive)
1018 34 : msgfmt = _("automatic aggressive vacuum of table \"%s.%s.%s\": index scans: %d\n");
1019 : else
1020 6 : msgfmt = _("automatic vacuum of table \"%s.%s.%s\": index scans: %d\n");
1021 : }
1022 231400 : appendStringInfo(&buf, msgfmt,
1023 : vacrel->dbname,
1024 : vacrel->relnamespace,
1025 : vacrel->relname,
1026 : vacrel->num_index_scans);
1027 323466 : appendStringInfo(&buf, _("pages: %u removed, %u remain, %u scanned (%.2f%% of total), %u eagerly scanned\n"),
1028 : vacrel->removed_pages,
1029 : new_rel_pages,
1030 : vacrel->scanned_pages,
1031 : orig_rel_pages == 0 ? 100.0 :
1032 92066 : 100.0 * vacrel->scanned_pages /
1033 : orig_rel_pages,
1034 : vacrel->eager_scanned_pages);
1035 231400 : appendStringInfo(&buf,
1036 231400 : _("tuples: %" PRId64 " removed, %" PRId64 " remain, %" PRId64 " are dead but not yet removable\n"),
1037 : vacrel->tuples_deleted,
1038 231400 : (int64) vacrel->new_rel_tuples,
1039 : vacrel->recently_dead_tuples);
1040 231400 : if (vacrel->missed_dead_tuples > 0)
1041 0 : appendStringInfo(&buf,
1042 0 : _("tuples missed: %" PRId64 " dead from %u pages not removed due to cleanup lock contention\n"),
1043 : vacrel->missed_dead_tuples,
1044 : vacrel->missed_dead_pages);
1045 231400 : diff = (int32) (ReadNextTransactionId() -
1046 231400 : vacrel->cutoffs.OldestXmin);
1047 231400 : appendStringInfo(&buf,
1048 231400 : _("removable cutoff: %u, which was %d XIDs old when operation ended\n"),
1049 : vacrel->cutoffs.OldestXmin, diff);
1050 231400 : if (frozenxid_updated)
1051 : {
1052 34990 : diff = (int32) (vacrel->NewRelfrozenXid -
1053 34990 : vacrel->cutoffs.relfrozenxid);
1054 34990 : appendStringInfo(&buf,
1055 34990 : _("new relfrozenxid: %u, which is %d XIDs ahead of previous value\n"),
1056 : vacrel->NewRelfrozenXid, diff);
1057 : }
1058 231400 : if (minmulti_updated)
1059 : {
1060 10 : diff = (int32) (vacrel->NewRelminMxid -
1061 10 : vacrel->cutoffs.relminmxid);
1062 10 : appendStringInfo(&buf,
1063 10 : _("new relminmxid: %u, which is %d MXIDs ahead of previous value\n"),
1064 : vacrel->NewRelminMxid, diff);
1065 : }
1066 323466 : appendStringInfo(&buf, _("frozen: %u pages from table (%.2f%% of total) had %" PRId64 " tuples frozen\n"),
1067 : vacrel->new_frozen_tuple_pages,
1068 : orig_rel_pages == 0 ? 100.0 :
1069 92066 : 100.0 * vacrel->new_frozen_tuple_pages /
1070 : orig_rel_pages,
1071 : vacrel->tuples_frozen);
1072 :
1073 231400 : appendStringInfo(&buf,
1074 231400 : _("visibility map: %u pages set all-visible, %u pages set all-frozen (%u were all-visible)\n"),
1075 : vacrel->vm_new_visible_pages,
1076 231400 : vacrel->vm_new_visible_frozen_pages +
1077 231400 : vacrel->vm_new_frozen_pages,
1078 : vacrel->vm_new_frozen_pages);
1079 231400 : if (vacrel->do_index_vacuuming)
1080 : {
1081 161680 : if (vacrel->nindexes == 0 || vacrel->num_index_scans == 0)
1082 161650 : appendStringInfoString(&buf, _("index scan not needed: "));
1083 : else
1084 30 : appendStringInfoString(&buf, _("index scan needed: "));
1085 :
1086 161680 : msgfmt = _("%u pages from table (%.2f%% of total) had %" PRId64 " dead item identifiers removed\n");
1087 : }
1088 : else
1089 : {
1090 69720 : if (!VacuumFailsafeActive)
1091 0 : appendStringInfoString(&buf, _("index scan bypassed: "));
1092 : else
1093 69720 : appendStringInfoString(&buf, _("index scan bypassed by failsafe: "));
1094 :
1095 69720 : msgfmt = _("%u pages from table (%.2f%% of total) have %" PRId64 " dead item identifiers\n");
1096 : }
1097 323466 : appendStringInfo(&buf, msgfmt,
1098 : vacrel->lpdead_item_pages,
1099 : orig_rel_pages == 0 ? 100.0 :
1100 92066 : 100.0 * vacrel->lpdead_item_pages / orig_rel_pages,
1101 : vacrel->lpdead_items);
1102 579528 : for (int i = 0; i < vacrel->nindexes; i++)
1103 : {
1104 348128 : IndexBulkDeleteResult *istat = vacrel->indstats[i];
1105 :
1106 348128 : if (!istat)
1107 348078 : continue;
1108 :
1109 50 : appendStringInfo(&buf,
1110 50 : _("index \"%s\": pages: %u in total, %u newly deleted, %u currently deleted, %u reusable\n"),
1111 50 : indnames[i],
1112 : istat->num_pages,
1113 : istat->pages_newly_deleted,
1114 : istat->pages_deleted,
1115 : istat->pages_free);
1116 : }
1117 231400 : if (track_cost_delay_timing)
1118 : {
1119 : /*
1120 : * We bypass the changecount mechanism because this value is
1121 : * only updated by the calling process. We also rely on the
1122 : * above call to pgstat_progress_end_command() to not clear
1123 : * the st_progress_param array.
1124 : */
1125 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
1126 0 : (double) MyBEEntry->st_progress_param[PROGRESS_VACUUM_DELAY_TIME] / 1000000.0);
1127 : }
1128 231400 : if (track_io_timing)
1129 : {
1130 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
1131 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
1132 :
1133 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
1134 : read_ms, write_ms);
1135 : }
1136 231400 : if (secs_dur > 0 || usecs_dur > 0)
1137 : {
1138 231400 : read_rate = (double) BLCKSZ * total_blks_read /
1139 231400 : (1024 * 1024) / (secs_dur + usecs_dur / 1000000.0);
1140 231400 : write_rate = (double) BLCKSZ * total_blks_dirtied /
1141 231400 : (1024 * 1024) / (secs_dur + usecs_dur / 1000000.0);
1142 : }
1143 231400 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
1144 : read_rate, write_rate);
1145 231400 : appendStringInfo(&buf,
1146 231400 : _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
1147 : total_blks_hit,
1148 : total_blks_read,
1149 : total_blks_dirtied);
1150 231400 : appendStringInfo(&buf,
1151 231400 : _("WAL usage: %" PRId64 " records, %" PRId64 " full page images, %" PRIu64 " bytes, %" PRId64 " buffers full\n"),
1152 : walusage.wal_records,
1153 : walusage.wal_fpi,
1154 : walusage.wal_bytes,
1155 : walusage.wal_buffers_full);
1156 231400 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
1157 :
1158 231400 : ereport(verbose ? INFO : LOG,
1159 : (errmsg_internal("%s", buf.data)));
1160 231400 : pfree(buf.data);
1161 : }
1162 : }
1163 :
1164 : /* Cleanup index statistics and index names */
1165 643250 : for (int i = 0; i < vacrel->nindexes; i++)
1166 : {
1167 385678 : if (vacrel->indstats[i])
1168 2696 : pfree(vacrel->indstats[i]);
1169 :
1170 385678 : if (instrument)
1171 348472 : pfree(indnames[i]);
1172 : }
1173 257572 : }
1174 :
1175 : /*
1176 : * lazy_scan_heap() -- workhorse function for VACUUM
1177 : *
1178 : * This routine prunes each page in the heap, and considers the need to
1179 : * freeze remaining tuples with storage (not including pages that can be
1180 : * skipped using the visibility map). Also performs related maintenance
1181 : * of the FSM and visibility map. These steps all take place during an
1182 : * initial pass over the target heap relation.
1183 : *
1184 : * Also invokes lazy_vacuum_all_indexes to vacuum indexes, which largely
1185 : * consists of deleting index tuples that point to LP_DEAD items left in
1186 : * heap pages following pruning. Earlier initial pass over the heap will
1187 : * have collected the TIDs whose index tuples need to be removed.
1188 : *
1189 : * Finally, invokes lazy_vacuum_heap_rel to vacuum heap pages, which
1190 : * largely consists of marking LP_DEAD items (from vacrel->dead_items)
1191 : * as LP_UNUSED. This has to happen in a second, final pass over the
1192 : * heap, to preserve a basic invariant that all index AMs rely on: no
1193 : * extant index tuple can ever be allowed to contain a TID that points to
1194 : * an LP_UNUSED line pointer in the heap. We must disallow premature
1195 : * recycling of line pointers to avoid index scans that get confused
1196 : * about which TID points to which tuple immediately after recycling.
1197 : * (Actually, this isn't a concern when target heap relation happens to
1198 : * have no indexes, which allows us to safely apply the one-pass strategy
1199 : * as an optimization).
1200 : *
1201 : * In practice we often have enough space to fit all TIDs, and so won't
1202 : * need to call lazy_vacuum more than once, after our initial pass over
1203 : * the heap has totally finished. Otherwise things are slightly more
1204 : * complicated: our "initial pass" over the heap applies only to those
1205 : * pages that were pruned before we needed to call lazy_vacuum, and our
1206 : * "final pass" over the heap only vacuums these same heap pages.
1207 : * However, we process indexes in full every time lazy_vacuum is called,
1208 : * which makes index processing very inefficient when memory is in short
1209 : * supply.
1210 : */
1211 : static void
1212 257572 : lazy_scan_heap(LVRelState *vacrel)
1213 : {
1214 : ReadStream *stream;
1215 257572 : BlockNumber rel_pages = vacrel->rel_pages,
1216 257572 : blkno = 0,
1217 257572 : next_fsm_block_to_vacuum = 0;
1218 257572 : BlockNumber orig_eager_scan_success_limit =
1219 : vacrel->eager_scan_remaining_successes; /* for logging */
1220 257572 : Buffer vmbuffer = InvalidBuffer;
1221 257572 : const int initprog_index[] = {
1222 : PROGRESS_VACUUM_PHASE,
1223 : PROGRESS_VACUUM_TOTAL_HEAP_BLKS,
1224 : PROGRESS_VACUUM_MAX_DEAD_TUPLE_BYTES
1225 : };
1226 : int64 initprog_val[3];
1227 :
1228 : /* Report that we're scanning the heap, advertising total # of blocks */
1229 257572 : initprog_val[0] = PROGRESS_VACUUM_PHASE_SCAN_HEAP;
1230 257572 : initprog_val[1] = rel_pages;
1231 257572 : initprog_val[2] = vacrel->dead_items_info->max_bytes;
1232 257572 : pgstat_progress_update_multi_param(3, initprog_index, initprog_val);
1233 :
1234 : /* Initialize for the first heap_vac_scan_next_block() call */
1235 257572 : vacrel->current_block = InvalidBlockNumber;
1236 257572 : vacrel->next_unskippable_block = InvalidBlockNumber;
1237 257572 : vacrel->next_unskippable_allvis = false;
1238 257572 : vacrel->next_unskippable_eager_scanned = false;
1239 257572 : vacrel->next_unskippable_vmbuffer = InvalidBuffer;
1240 :
1241 : /*
1242 : * Set up the read stream for vacuum's first pass through the heap.
1243 : *
1244 : * This could be made safe for READ_STREAM_USE_BATCHING, but only with
1245 : * explicit work in heap_vac_scan_next_block.
1246 : */
1247 257572 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE,
1248 : vacrel->bstrategy,
1249 : vacrel->rel,
1250 : MAIN_FORKNUM,
1251 : heap_vac_scan_next_block,
1252 : vacrel,
1253 : sizeof(uint8));
1254 :
1255 : while (true)
1256 1160540 : {
1257 : Buffer buf;
1258 : Page page;
1259 1418112 : uint8 blk_info = 0;
1260 1418112 : int ndeleted = 0;
1261 : bool has_lpdead_items;
1262 1418112 : void *per_buffer_data = NULL;
1263 1418112 : bool vm_page_frozen = false;
1264 1418112 : bool got_cleanup_lock = false;
1265 :
1266 1418112 : vacuum_delay_point(false);
1267 :
1268 : /*
1269 : * Regularly check if wraparound failsafe should trigger.
1270 : *
1271 : * There is a similar check inside lazy_vacuum_all_indexes(), but
1272 : * relfrozenxid might start to look dangerously old before we reach
1273 : * that point. This check also provides failsafe coverage for the
1274 : * one-pass strategy, and the two-pass strategy with the index_cleanup
1275 : * param set to 'off'.
1276 : */
1277 1418112 : if (vacrel->scanned_pages > 0 &&
1278 1160540 : vacrel->scanned_pages % FAILSAFE_EVERY_PAGES == 0)
1279 0 : lazy_check_wraparound_failsafe(vacrel);
1280 :
1281 : /*
1282 : * Consider if we definitely have enough space to process TIDs on page
1283 : * already. If we are close to overrunning the available space for
1284 : * dead_items TIDs, pause and do a cycle of vacuuming before we tackle
1285 : * this page. However, let's force at least one page-worth of tuples
1286 : * to be stored as to ensure we do at least some work when the memory
1287 : * configured is so low that we run out before storing anything.
1288 : */
1289 1418112 : if (vacrel->dead_items_info->num_items > 0 &&
1290 48326 : TidStoreMemoryUsage(vacrel->dead_items) > vacrel->dead_items_info->max_bytes)
1291 : {
1292 : /*
1293 : * Before beginning index vacuuming, we release any pin we may
1294 : * hold on the visibility map page. This isn't necessary for
1295 : * correctness, but we do it anyway to avoid holding the pin
1296 : * across a lengthy, unrelated operation.
1297 : */
1298 14 : if (BufferIsValid(vmbuffer))
1299 : {
1300 14 : ReleaseBuffer(vmbuffer);
1301 14 : vmbuffer = InvalidBuffer;
1302 : }
1303 :
1304 : /* Perform a round of index and heap vacuuming */
1305 14 : vacrel->consider_bypass_optimization = false;
1306 14 : lazy_vacuum(vacrel);
1307 :
1308 : /*
1309 : * Vacuum the Free Space Map to make newly-freed space visible on
1310 : * upper-level FSM pages. Note that blkno is the previously
1311 : * processed block.
1312 : */
1313 14 : FreeSpaceMapVacuumRange(vacrel->rel, next_fsm_block_to_vacuum,
1314 : blkno + 1);
1315 14 : next_fsm_block_to_vacuum = blkno;
1316 :
1317 : /* Report that we are once again scanning the heap */
1318 14 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
1319 : PROGRESS_VACUUM_PHASE_SCAN_HEAP);
1320 : }
1321 :
1322 1418112 : buf = read_stream_next_buffer(stream, &per_buffer_data);
1323 :
1324 : /* The relation is exhausted. */
1325 1418112 : if (!BufferIsValid(buf))
1326 257572 : break;
1327 :
1328 1160540 : blk_info = *((uint8 *) per_buffer_data);
1329 1160540 : CheckBufferIsPinnedOnce(buf);
1330 1160540 : page = BufferGetPage(buf);
1331 1160540 : blkno = BufferGetBlockNumber(buf);
1332 :
1333 1160540 : vacrel->scanned_pages++;
1334 1160540 : if (blk_info & VAC_BLK_WAS_EAGER_SCANNED)
1335 0 : vacrel->eager_scanned_pages++;
1336 :
1337 : /* Report as block scanned, update error traceback information */
1338 1160540 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_SCANNED, blkno);
1339 1160540 : update_vacuum_error_info(vacrel, NULL, VACUUM_ERRCB_PHASE_SCAN_HEAP,
1340 : blkno, InvalidOffsetNumber);
1341 :
1342 : /*
1343 : * Pin the visibility map page in case we need to mark the page
1344 : * all-visible. In most cases this will be very cheap, because we'll
1345 : * already have the correct page pinned anyway.
1346 : */
1347 1160540 : visibilitymap_pin(vacrel->rel, blkno, &vmbuffer);
1348 :
1349 : /*
1350 : * We need a buffer cleanup lock to prune HOT chains and defragment
1351 : * the page in lazy_scan_prune. But when it's not possible to acquire
1352 : * a cleanup lock right away, we may be able to settle for reduced
1353 : * processing using lazy_scan_noprune.
1354 : */
1355 1160540 : got_cleanup_lock = ConditionalLockBufferForCleanup(buf);
1356 :
1357 1160540 : if (!got_cleanup_lock)
1358 312 : LockBuffer(buf, BUFFER_LOCK_SHARE);
1359 :
1360 : /* Check for new or empty pages before lazy_scan_[no]prune call */
1361 1160540 : if (lazy_scan_new_or_empty(vacrel, buf, blkno, page, !got_cleanup_lock,
1362 1160540 : vmbuffer))
1363 : {
1364 : /* Processed as new/empty page (lock and pin released) */
1365 1224 : continue;
1366 : }
1367 :
1368 : /*
1369 : * If we didn't get the cleanup lock, we can still collect LP_DEAD
1370 : * items in the dead_items area for later vacuuming, count live and
1371 : * recently dead tuples for vacuum logging, and determine if this
1372 : * block could later be truncated. If we encounter any xid/mxids that
1373 : * require advancing the relfrozenxid/relminxid, we'll have to wait
1374 : * for a cleanup lock and call lazy_scan_prune().
1375 : */
1376 1159316 : if (!got_cleanup_lock &&
1377 312 : !lazy_scan_noprune(vacrel, buf, blkno, page, &has_lpdead_items))
1378 : {
1379 : /*
1380 : * lazy_scan_noprune could not do all required processing. Wait
1381 : * for a cleanup lock, and call lazy_scan_prune in the usual way.
1382 : */
1383 : Assert(vacrel->aggressive);
1384 114 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1385 114 : LockBufferForCleanup(buf);
1386 114 : got_cleanup_lock = true;
1387 : }
1388 :
1389 : /*
1390 : * If we have a cleanup lock, we must now prune, freeze, and count
1391 : * tuples. We may have acquired the cleanup lock originally, or we may
1392 : * have gone back and acquired it after lazy_scan_noprune() returned
1393 : * false. Either way, the page hasn't been processed yet.
1394 : *
1395 : * Like lazy_scan_noprune(), lazy_scan_prune() will count
1396 : * recently_dead_tuples and live tuples for vacuum logging, determine
1397 : * if the block can later be truncated, and accumulate the details of
1398 : * remaining LP_DEAD line pointers on the page into dead_items. These
1399 : * dead items include those pruned by lazy_scan_prune() as well as
1400 : * line pointers previously marked LP_DEAD.
1401 : */
1402 1159316 : if (got_cleanup_lock)
1403 1159118 : ndeleted = lazy_scan_prune(vacrel, buf, blkno, page,
1404 : vmbuffer,
1405 1159118 : blk_info & VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM,
1406 : &has_lpdead_items, &vm_page_frozen);
1407 :
1408 : /*
1409 : * Count an eagerly scanned page as a failure or a success.
1410 : *
1411 : * Only lazy_scan_prune() freezes pages, so if we didn't get the
1412 : * cleanup lock, we won't have frozen the page. However, we only count
1413 : * pages that were too new to require freezing as eager freeze
1414 : * failures.
1415 : *
1416 : * We could gather more information from lazy_scan_noprune() about
1417 : * whether or not there were tuples with XIDs or MXIDs older than the
1418 : * FreezeLimit or MultiXactCutoff. However, for simplicity, we simply
1419 : * exclude pages skipped due to cleanup lock contention from eager
1420 : * freeze algorithm caps.
1421 : */
1422 1159316 : if (got_cleanup_lock &&
1423 1159118 : (blk_info & VAC_BLK_WAS_EAGER_SCANNED))
1424 : {
1425 : /* Aggressive vacuums do not eager scan. */
1426 : Assert(!vacrel->aggressive);
1427 :
1428 0 : if (vm_page_frozen)
1429 : {
1430 0 : if (vacrel->eager_scan_remaining_successes > 0)
1431 0 : vacrel->eager_scan_remaining_successes--;
1432 :
1433 0 : if (vacrel->eager_scan_remaining_successes == 0)
1434 : {
1435 : /*
1436 : * Report only once that we disabled eager scanning. We
1437 : * may eagerly read ahead blocks in excess of the success
1438 : * or failure caps before attempting to freeze them, so we
1439 : * could reach here even after disabling additional eager
1440 : * scanning.
1441 : */
1442 0 : if (vacrel->eager_scan_max_fails_per_region > 0)
1443 0 : ereport(vacrel->verbose ? INFO : DEBUG2,
1444 : (errmsg("disabling eager scanning after freezing %u eagerly scanned blocks of relation \"%s.%s.%s\"",
1445 : orig_eager_scan_success_limit,
1446 : vacrel->dbname, vacrel->relnamespace,
1447 : vacrel->relname)));
1448 :
1449 : /*
1450 : * If we hit our success cap, permanently disable eager
1451 : * scanning by setting the other eager scan management
1452 : * fields to their disabled values.
1453 : */
1454 0 : vacrel->eager_scan_remaining_fails = 0;
1455 0 : vacrel->next_eager_scan_region_start = InvalidBlockNumber;
1456 0 : vacrel->eager_scan_max_fails_per_region = 0;
1457 : }
1458 : }
1459 0 : else if (vacrel->eager_scan_remaining_fails > 0)
1460 0 : vacrel->eager_scan_remaining_fails--;
1461 : }
1462 :
1463 : /*
1464 : * Now drop the buffer lock and, potentially, update the FSM.
1465 : *
1466 : * Our goal is to update the freespace map the last time we touch the
1467 : * page. If we'll process a block in the second pass, we may free up
1468 : * additional space on the page, so it is better to update the FSM
1469 : * after the second pass. If the relation has no indexes, or if index
1470 : * vacuuming is disabled, there will be no second heap pass; if this
1471 : * particular page has no dead items, the second heap pass will not
1472 : * touch this page. So, in those cases, update the FSM now.
1473 : *
1474 : * Note: In corner cases, it's possible to miss updating the FSM
1475 : * entirely. If index vacuuming is currently enabled, we'll skip the
1476 : * FSM update now. But if failsafe mode is later activated, or there
1477 : * are so few dead tuples that index vacuuming is bypassed, there will
1478 : * also be no opportunity to update the FSM later, because we'll never
1479 : * revisit this page. Since updating the FSM is desirable but not
1480 : * absolutely required, that's OK.
1481 : */
1482 1159316 : if (vacrel->nindexes == 0
1483 1122372 : || !vacrel->do_index_vacuuming
1484 786422 : || !has_lpdead_items)
1485 1133518 : {
1486 1133518 : Size freespace = PageGetHeapFreeSpace(page);
1487 :
1488 1133518 : UnlockReleaseBuffer(buf);
1489 1133518 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
1490 :
1491 : /*
1492 : * Periodically perform FSM vacuuming to make newly-freed space
1493 : * visible on upper FSM pages. This is done after vacuuming if the
1494 : * table has indexes. There will only be newly-freed space if we
1495 : * held the cleanup lock and lazy_scan_prune() was called.
1496 : */
1497 1133518 : if (got_cleanup_lock && vacrel->nindexes == 0 && ndeleted > 0 &&
1498 914 : blkno - next_fsm_block_to_vacuum >= VACUUM_FSM_EVERY_PAGES)
1499 : {
1500 0 : FreeSpaceMapVacuumRange(vacrel->rel, next_fsm_block_to_vacuum,
1501 : blkno);
1502 0 : next_fsm_block_to_vacuum = blkno;
1503 : }
1504 : }
1505 : else
1506 25798 : UnlockReleaseBuffer(buf);
1507 : }
1508 :
1509 257572 : vacrel->blkno = InvalidBlockNumber;
1510 257572 : if (BufferIsValid(vmbuffer))
1511 103116 : ReleaseBuffer(vmbuffer);
1512 :
1513 : /*
1514 : * Report that everything is now scanned. We never skip scanning the last
1515 : * block in the relation, so we can pass rel_pages here.
1516 : */
1517 257572 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_SCANNED,
1518 : rel_pages);
1519 :
1520 : /* now we can compute the new value for pg_class.reltuples */
1521 515144 : vacrel->new_live_tuples = vac_estimate_reltuples(vacrel->rel, rel_pages,
1522 : vacrel->scanned_pages,
1523 257572 : vacrel->live_tuples);
1524 :
1525 : /*
1526 : * Also compute the total number of surviving heap entries. In the
1527 : * (unlikely) scenario that new_live_tuples is -1, take it as zero.
1528 : */
1529 257572 : vacrel->new_rel_tuples =
1530 257572 : Max(vacrel->new_live_tuples, 0) + vacrel->recently_dead_tuples +
1531 257572 : vacrel->missed_dead_tuples;
1532 :
1533 257572 : read_stream_end(stream);
1534 :
1535 : /*
1536 : * Do index vacuuming (call each index's ambulkdelete routine), then do
1537 : * related heap vacuuming
1538 : */
1539 257572 : if (vacrel->dead_items_info->num_items > 0)
1540 1268 : lazy_vacuum(vacrel);
1541 :
1542 : /*
1543 : * Vacuum the remainder of the Free Space Map. We must do this whether or
1544 : * not there were indexes, and whether or not we bypassed index vacuuming.
1545 : * We can pass rel_pages here because we never skip scanning the last
1546 : * block of the relation.
1547 : */
1548 257572 : if (rel_pages > next_fsm_block_to_vacuum)
1549 103118 : FreeSpaceMapVacuumRange(vacrel->rel, next_fsm_block_to_vacuum, rel_pages);
1550 :
1551 : /* report all blocks vacuumed */
1552 257572 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_VACUUMED, rel_pages);
1553 :
1554 : /* Do final index cleanup (call each index's amvacuumcleanup routine) */
1555 257572 : if (vacrel->nindexes > 0 && vacrel->do_index_cleanup)
1556 178248 : lazy_cleanup_all_indexes(vacrel);
1557 257572 : }
1558 :
1559 : /*
1560 : * heap_vac_scan_next_block() -- read stream callback to get the next block
1561 : * for vacuum to process
1562 : *
1563 : * Every time lazy_scan_heap() needs a new block to process during its first
1564 : * phase, it invokes read_stream_next_buffer() with a stream set up to call
1565 : * heap_vac_scan_next_block() to get the next block.
1566 : *
1567 : * heap_vac_scan_next_block() uses the visibility map, vacuum options, and
1568 : * various thresholds to skip blocks which do not need to be processed and
1569 : * returns the next block to process or InvalidBlockNumber if there are no
1570 : * remaining blocks.
1571 : *
1572 : * The visibility status of the next block to process and whether or not it
1573 : * was eager scanned is set in the per_buffer_data.
1574 : *
1575 : * callback_private_data contains a reference to the LVRelState, passed to the
1576 : * read stream API during stream setup. The LVRelState is an in/out parameter
1577 : * here (locally named `vacrel`). Vacuum options and information about the
1578 : * relation are read from it. vacrel->skippedallvis is set if we skip a block
1579 : * that's all-visible but not all-frozen (to ensure that we don't update
1580 : * relfrozenxid in that case). vacrel also holds information about the next
1581 : * unskippable block -- as bookkeeping for this function.
1582 : */
1583 : static BlockNumber
1584 1418112 : heap_vac_scan_next_block(ReadStream *stream,
1585 : void *callback_private_data,
1586 : void *per_buffer_data)
1587 : {
1588 : BlockNumber next_block;
1589 1418112 : LVRelState *vacrel = callback_private_data;
1590 1418112 : uint8 blk_info = 0;
1591 :
1592 : /* relies on InvalidBlockNumber + 1 overflowing to 0 on first call */
1593 1418112 : next_block = vacrel->current_block + 1;
1594 :
1595 : /* Have we reached the end of the relation? */
1596 1418112 : if (next_block >= vacrel->rel_pages)
1597 : {
1598 257572 : if (BufferIsValid(vacrel->next_unskippable_vmbuffer))
1599 : {
1600 100212 : ReleaseBuffer(vacrel->next_unskippable_vmbuffer);
1601 100212 : vacrel->next_unskippable_vmbuffer = InvalidBuffer;
1602 : }
1603 257572 : return InvalidBlockNumber;
1604 : }
1605 :
1606 : /*
1607 : * We must be in one of the three following states:
1608 : */
1609 1160540 : if (next_block > vacrel->next_unskippable_block ||
1610 415744 : vacrel->next_unskippable_block == InvalidBlockNumber)
1611 : {
1612 : /*
1613 : * 1. We have just processed an unskippable block (or we're at the
1614 : * beginning of the scan). Find the next unskippable block using the
1615 : * visibility map.
1616 : */
1617 : bool skipsallvis;
1618 :
1619 847914 : find_next_unskippable_block(vacrel, &skipsallvis);
1620 :
1621 : /*
1622 : * We now know the next block that we must process. It can be the
1623 : * next block after the one we just processed, or something further
1624 : * ahead. If it's further ahead, we can jump to it, but we choose to
1625 : * do so only if we can skip at least SKIP_PAGES_THRESHOLD consecutive
1626 : * pages. Since we're reading sequentially, the OS should be doing
1627 : * readahead for us, so there's no gain in skipping a page now and
1628 : * then. Skipping such a range might even discourage sequential
1629 : * detection.
1630 : *
1631 : * This test also enables more frequent relfrozenxid advancement
1632 : * during non-aggressive VACUUMs. If the range has any all-visible
1633 : * pages then skipping makes updating relfrozenxid unsafe, which is a
1634 : * real downside.
1635 : */
1636 847914 : if (vacrel->next_unskippable_block - next_block >= SKIP_PAGES_THRESHOLD)
1637 : {
1638 9258 : next_block = vacrel->next_unskippable_block;
1639 9258 : if (skipsallvis)
1640 64 : vacrel->skippedallvis = true;
1641 : }
1642 : }
1643 :
1644 : /* Now we must be in one of the two remaining states: */
1645 1160540 : if (next_block < vacrel->next_unskippable_block)
1646 : {
1647 : /*
1648 : * 2. We are processing a range of blocks that we could have skipped
1649 : * but chose not to. We know that they are all-visible in the VM,
1650 : * otherwise they would've been unskippable.
1651 : */
1652 312626 : vacrel->current_block = next_block;
1653 312626 : blk_info |= VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM;
1654 312626 : *((uint8 *) per_buffer_data) = blk_info;
1655 312626 : return vacrel->current_block;
1656 : }
1657 : else
1658 : {
1659 : /*
1660 : * 3. We reached the next unskippable block. Process it. On next
1661 : * iteration, we will be back in state 1.
1662 : */
1663 : Assert(next_block == vacrel->next_unskippable_block);
1664 :
1665 847914 : vacrel->current_block = next_block;
1666 847914 : if (vacrel->next_unskippable_allvis)
1667 94212 : blk_info |= VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM;
1668 847914 : if (vacrel->next_unskippable_eager_scanned)
1669 0 : blk_info |= VAC_BLK_WAS_EAGER_SCANNED;
1670 847914 : *((uint8 *) per_buffer_data) = blk_info;
1671 847914 : return vacrel->current_block;
1672 : }
1673 : }
1674 :
1675 : /*
1676 : * Find the next unskippable block in a vacuum scan using the visibility map.
1677 : * The next unskippable block and its visibility information is updated in
1678 : * vacrel.
1679 : *
1680 : * Note: our opinion of which blocks can be skipped can go stale immediately.
1681 : * It's okay if caller "misses" a page whose all-visible or all-frozen marking
1682 : * was concurrently cleared, though. All that matters is that caller scan all
1683 : * pages whose tuples might contain XIDs < OldestXmin, or MXIDs < OldestMxact.
1684 : * (Actually, non-aggressive VACUUMs can choose to skip all-visible pages with
1685 : * older XIDs/MXIDs. The *skippedallvis flag will be set here when the choice
1686 : * to skip such a range is actually made, making everything safe.)
1687 : */
1688 : static void
1689 847914 : find_next_unskippable_block(LVRelState *vacrel, bool *skipsallvis)
1690 : {
1691 847914 : BlockNumber rel_pages = vacrel->rel_pages;
1692 847914 : BlockNumber next_unskippable_block = vacrel->next_unskippable_block + 1;
1693 847914 : Buffer next_unskippable_vmbuffer = vacrel->next_unskippable_vmbuffer;
1694 847914 : bool next_unskippable_eager_scanned = false;
1695 : bool next_unskippable_allvis;
1696 :
1697 847914 : *skipsallvis = false;
1698 :
1699 990434 : for (;; next_unskippable_block++)
1700 990434 : {
1701 1838348 : uint8 mapbits = visibilitymap_get_status(vacrel->rel,
1702 : next_unskippable_block,
1703 : &next_unskippable_vmbuffer);
1704 :
1705 1838348 : next_unskippable_allvis = (mapbits & VISIBILITYMAP_ALL_VISIBLE) != 0;
1706 :
1707 : /*
1708 : * At the start of each eager scan region, normal vacuums with eager
1709 : * scanning enabled reset the failure counter, allowing vacuum to
1710 : * resume eager scanning if it had been suspended in the previous
1711 : * region.
1712 : */
1713 1838348 : if (next_unskippable_block >= vacrel->next_eager_scan_region_start)
1714 : {
1715 0 : vacrel->eager_scan_remaining_fails =
1716 0 : vacrel->eager_scan_max_fails_per_region;
1717 0 : vacrel->next_eager_scan_region_start += EAGER_SCAN_REGION_SIZE;
1718 : }
1719 :
1720 : /*
1721 : * A block is unskippable if it is not all visible according to the
1722 : * visibility map.
1723 : */
1724 1838348 : if (!next_unskippable_allvis)
1725 : {
1726 : Assert((mapbits & VISIBILITYMAP_ALL_FROZEN) == 0);
1727 753702 : break;
1728 : }
1729 :
1730 : /*
1731 : * Caller must scan the last page to determine whether it has tuples
1732 : * (caller must have the opportunity to set vacrel->nonempty_pages).
1733 : * This rule avoids having lazy_truncate_heap() take access-exclusive
1734 : * lock on rel to attempt a truncation that fails anyway, just because
1735 : * there are tuples on the last page (it is likely that there will be
1736 : * tuples on other nearby pages as well, but those can be skipped).
1737 : *
1738 : * Implement this by always treating the last block as unsafe to skip.
1739 : */
1740 1084646 : if (next_unskippable_block == rel_pages - 1)
1741 93402 : break;
1742 :
1743 : /* DISABLE_PAGE_SKIPPING makes all skipping unsafe */
1744 991244 : if (!vacrel->skipwithvm)
1745 810 : break;
1746 :
1747 : /*
1748 : * All-frozen pages cannot contain XIDs < OldestXmin (XIDs that aren't
1749 : * already frozen by now), so this page can be skipped.
1750 : */
1751 990434 : if ((mapbits & VISIBILITYMAP_ALL_FROZEN) != 0)
1752 985198 : continue;
1753 :
1754 : /*
1755 : * Aggressive vacuums cannot skip any all-visible pages that are not
1756 : * also all-frozen.
1757 : */
1758 5236 : if (vacrel->aggressive)
1759 0 : break;
1760 :
1761 : /*
1762 : * Normal vacuums with eager scanning enabled only skip all-visible
1763 : * but not all-frozen pages if they have hit the failure limit for the
1764 : * current eager scan region.
1765 : */
1766 5236 : if (vacrel->eager_scan_remaining_fails > 0)
1767 : {
1768 0 : next_unskippable_eager_scanned = true;
1769 0 : break;
1770 : }
1771 :
1772 : /*
1773 : * All-visible blocks are safe to skip in a normal vacuum. But
1774 : * remember that the final range contains such a block for later.
1775 : */
1776 5236 : *skipsallvis = true;
1777 : }
1778 :
1779 : /* write the local variables back to vacrel */
1780 847914 : vacrel->next_unskippable_block = next_unskippable_block;
1781 847914 : vacrel->next_unskippable_allvis = next_unskippable_allvis;
1782 847914 : vacrel->next_unskippable_eager_scanned = next_unskippable_eager_scanned;
1783 847914 : vacrel->next_unskippable_vmbuffer = next_unskippable_vmbuffer;
1784 847914 : }
1785 :
1786 : /*
1787 : * lazy_scan_new_or_empty() -- lazy_scan_heap() new/empty page handling.
1788 : *
1789 : * Must call here to handle both new and empty pages before calling
1790 : * lazy_scan_prune or lazy_scan_noprune, since they're not prepared to deal
1791 : * with new or empty pages.
1792 : *
1793 : * It's necessary to consider new pages as a special case, since the rules for
1794 : * maintaining the visibility map and FSM with empty pages are a little
1795 : * different (though new pages can be truncated away during rel truncation).
1796 : *
1797 : * Empty pages are not really a special case -- they're just heap pages that
1798 : * have no allocated tuples (including even LP_UNUSED items). You might
1799 : * wonder why we need to handle them here all the same. It's only necessary
1800 : * because of a corner-case involving a hard crash during heap relation
1801 : * extension. If we ever make relation-extension crash safe, then it should
1802 : * no longer be necessary to deal with empty pages here (or new pages, for
1803 : * that matter).
1804 : *
1805 : * Caller must hold at least a shared lock. We might need to escalate the
1806 : * lock in that case, so the type of lock caller holds needs to be specified
1807 : * using 'sharelock' argument.
1808 : *
1809 : * Returns false in common case where caller should go on to call
1810 : * lazy_scan_prune (or lazy_scan_noprune). Otherwise returns true, indicating
1811 : * that lazy_scan_heap is done processing the page, releasing lock on caller's
1812 : * behalf.
1813 : *
1814 : * No vm_page_frozen output parameter (like that passed to lazy_scan_prune())
1815 : * is passed here because neither empty nor new pages can be eagerly frozen.
1816 : * New pages are never frozen. Empty pages are always set frozen in the VM at
1817 : * the same time that they are set all-visible, and we don't eagerly scan
1818 : * frozen pages.
1819 : */
1820 : static bool
1821 1160540 : lazy_scan_new_or_empty(LVRelState *vacrel, Buffer buf, BlockNumber blkno,
1822 : Page page, bool sharelock, Buffer vmbuffer)
1823 : {
1824 : Size freespace;
1825 :
1826 1160540 : if (PageIsNew(page))
1827 : {
1828 : /*
1829 : * All-zeroes pages can be left over if either a backend extends the
1830 : * relation by a single page, but crashes before the newly initialized
1831 : * page has been written out, or when bulk-extending the relation
1832 : * (which creates a number of empty pages at the tail end of the
1833 : * relation), and then enters them into the FSM.
1834 : *
1835 : * Note we do not enter the page into the visibilitymap. That has the
1836 : * downside that we repeatedly visit this page in subsequent vacuums,
1837 : * but otherwise we'll never discover the space on a promoted standby.
1838 : * The harm of repeated checking ought to normally not be too bad. The
1839 : * space usually should be used at some point, otherwise there
1840 : * wouldn't be any regular vacuums.
1841 : *
1842 : * Make sure these pages are in the FSM, to ensure they can be reused.
1843 : * Do that by testing if there's any space recorded for the page. If
1844 : * not, enter it. We do so after releasing the lock on the heap page,
1845 : * the FSM is approximate, after all.
1846 : */
1847 1170 : UnlockReleaseBuffer(buf);
1848 :
1849 1170 : if (GetRecordedFreeSpace(vacrel->rel, blkno) == 0)
1850 : {
1851 858 : freespace = BLCKSZ - SizeOfPageHeaderData;
1852 :
1853 858 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
1854 : }
1855 :
1856 1170 : return true;
1857 : }
1858 :
1859 1159370 : if (PageIsEmpty(page))
1860 : {
1861 : /*
1862 : * It seems likely that caller will always be able to get a cleanup
1863 : * lock on an empty page. But don't take any chances -- escalate to
1864 : * an exclusive lock (still don't need a cleanup lock, though).
1865 : */
1866 54 : if (sharelock)
1867 : {
1868 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1869 0 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
1870 :
1871 0 : if (!PageIsEmpty(page))
1872 : {
1873 : /* page isn't new or empty -- keep lock and pin for now */
1874 0 : return false;
1875 : }
1876 : }
1877 : else
1878 : {
1879 : /* Already have a full cleanup lock (which is more than enough) */
1880 : }
1881 :
1882 : /*
1883 : * Unlike new pages, empty pages are always set all-visible and
1884 : * all-frozen.
1885 : */
1886 54 : if (!PageIsAllVisible(page))
1887 : {
1888 0 : START_CRIT_SECTION();
1889 :
1890 : /* mark buffer dirty before writing a WAL record */
1891 0 : MarkBufferDirty(buf);
1892 :
1893 : /*
1894 : * It's possible that another backend has extended the heap,
1895 : * initialized the page, and then failed to WAL-log the page due
1896 : * to an ERROR. Since heap extension is not WAL-logged, recovery
1897 : * might try to replay our record setting the page all-visible and
1898 : * find that the page isn't initialized, which will cause a PANIC.
1899 : * To prevent that, check whether the page has been previously
1900 : * WAL-logged, and if not, do that now.
1901 : */
1902 0 : if (RelationNeedsWAL(vacrel->rel) &&
1903 0 : PageGetLSN(page) == InvalidXLogRecPtr)
1904 0 : log_newpage_buffer(buf, true);
1905 :
1906 0 : PageSetAllVisible(page);
1907 0 : visibilitymap_set(vacrel->rel, blkno, buf,
1908 : InvalidXLogRecPtr,
1909 : vmbuffer, InvalidTransactionId,
1910 : VISIBILITYMAP_ALL_VISIBLE |
1911 : VISIBILITYMAP_ALL_FROZEN);
1912 0 : END_CRIT_SECTION();
1913 :
1914 : /* Count the newly all-frozen pages for logging */
1915 0 : vacrel->vm_new_visible_pages++;
1916 0 : vacrel->vm_new_visible_frozen_pages++;
1917 : }
1918 :
1919 54 : freespace = PageGetHeapFreeSpace(page);
1920 54 : UnlockReleaseBuffer(buf);
1921 54 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
1922 54 : return true;
1923 : }
1924 :
1925 : /* page isn't new or empty -- keep lock and pin */
1926 1159316 : return false;
1927 : }
1928 :
1929 : /* qsort comparator for sorting OffsetNumbers */
1930 : static int
1931 5801542 : cmpOffsetNumbers(const void *a, const void *b)
1932 : {
1933 5801542 : return pg_cmp_u16(*(const OffsetNumber *) a, *(const OffsetNumber *) b);
1934 : }
1935 :
1936 : /*
1937 : * lazy_scan_prune() -- lazy_scan_heap() pruning and freezing.
1938 : *
1939 : * Caller must hold pin and buffer cleanup lock on the buffer.
1940 : *
1941 : * vmbuffer is the buffer containing the VM block with visibility information
1942 : * for the heap block, blkno. all_visible_according_to_vm is the saved
1943 : * visibility status of the heap block looked up earlier by the caller. We
1944 : * won't rely entirely on this status, as it may be out of date.
1945 : *
1946 : * *has_lpdead_items is set to true or false depending on whether, upon return
1947 : * from this function, any LP_DEAD items are still present on the page.
1948 : *
1949 : * *vm_page_frozen is set to true if the page is newly set all-frozen in the
1950 : * VM. The caller currently only uses this for determining whether an eagerly
1951 : * scanned page was successfully set all-frozen.
1952 : *
1953 : * Returns the number of tuples deleted from the page during HOT pruning.
1954 : */
1955 : static int
1956 1159118 : lazy_scan_prune(LVRelState *vacrel,
1957 : Buffer buf,
1958 : BlockNumber blkno,
1959 : Page page,
1960 : Buffer vmbuffer,
1961 : bool all_visible_according_to_vm,
1962 : bool *has_lpdead_items,
1963 : bool *vm_page_frozen)
1964 : {
1965 1159118 : Relation rel = vacrel->rel;
1966 : PruneFreezeResult presult;
1967 1159118 : int prune_options = 0;
1968 :
1969 : Assert(BufferGetBlockNumber(buf) == blkno);
1970 :
1971 : /*
1972 : * Prune all HOT-update chains and potentially freeze tuples on this page.
1973 : *
1974 : * If the relation has no indexes, we can immediately mark would-be dead
1975 : * items LP_UNUSED.
1976 : *
1977 : * The number of tuples removed from the page is returned in
1978 : * presult.ndeleted. It should not be confused with presult.lpdead_items;
1979 : * presult.lpdead_items's final value can be thought of as the number of
1980 : * tuples that were deleted from indexes.
1981 : *
1982 : * We will update the VM after collecting LP_DEAD items and freezing
1983 : * tuples. Pruning will have determined whether or not the page is
1984 : * all-visible.
1985 : */
1986 1159118 : prune_options = HEAP_PAGE_PRUNE_FREEZE;
1987 1159118 : if (vacrel->nindexes == 0)
1988 36944 : prune_options |= HEAP_PAGE_PRUNE_MARK_UNUSED_NOW;
1989 :
1990 1159118 : heap_page_prune_and_freeze(rel, buf, vacrel->vistest, prune_options,
1991 : &vacrel->cutoffs, &presult, PRUNE_VACUUM_SCAN,
1992 : &vacrel->offnum,
1993 : &vacrel->NewRelfrozenXid, &vacrel->NewRelminMxid);
1994 :
1995 : Assert(MultiXactIdIsValid(vacrel->NewRelminMxid));
1996 : Assert(TransactionIdIsValid(vacrel->NewRelfrozenXid));
1997 :
1998 1159118 : if (presult.nfrozen > 0)
1999 : {
2000 : /*
2001 : * We don't increment the new_frozen_tuple_pages instrumentation
2002 : * counter when nfrozen == 0, since it only counts pages with newly
2003 : * frozen tuples (don't confuse that with pages newly set all-frozen
2004 : * in VM).
2005 : */
2006 45534 : vacrel->new_frozen_tuple_pages++;
2007 : }
2008 :
2009 : /*
2010 : * VACUUM will call heap_page_is_all_visible() during the second pass over
2011 : * the heap to determine all_visible and all_frozen for the page -- this
2012 : * is a specialized version of the logic from this function. Now that
2013 : * we've finished pruning and freezing, make sure that we're in total
2014 : * agreement with heap_page_is_all_visible() using an assertion.
2015 : */
2016 : #ifdef USE_ASSERT_CHECKING
2017 : /* Note that all_frozen value does not matter when !all_visible */
2018 : if (presult.all_visible)
2019 : {
2020 : TransactionId debug_cutoff;
2021 : bool debug_all_frozen;
2022 :
2023 : Assert(presult.lpdead_items == 0);
2024 :
2025 : if (!heap_page_is_all_visible(vacrel->rel, buf,
2026 : vacrel->cutoffs.OldestXmin, &debug_all_frozen,
2027 : &debug_cutoff, &vacrel->offnum))
2028 : Assert(false);
2029 :
2030 : Assert(presult.all_frozen == debug_all_frozen);
2031 :
2032 : Assert(!TransactionIdIsValid(debug_cutoff) ||
2033 : debug_cutoff == presult.vm_conflict_horizon);
2034 : }
2035 : #endif
2036 :
2037 : /*
2038 : * Now save details of the LP_DEAD items from the page in vacrel
2039 : */
2040 1159118 : if (presult.lpdead_items > 0)
2041 : {
2042 30358 : vacrel->lpdead_item_pages++;
2043 :
2044 : /*
2045 : * deadoffsets are collected incrementally in
2046 : * heap_page_prune_and_freeze() as each dead line pointer is recorded,
2047 : * with an indeterminate order, but dead_items_add requires them to be
2048 : * sorted.
2049 : */
2050 30358 : qsort(presult.deadoffsets, presult.lpdead_items, sizeof(OffsetNumber),
2051 : cmpOffsetNumbers);
2052 :
2053 30358 : dead_items_add(vacrel, blkno, presult.deadoffsets, presult.lpdead_items);
2054 : }
2055 :
2056 : /* Finally, add page-local counts to whole-VACUUM counts */
2057 1159118 : vacrel->tuples_deleted += presult.ndeleted;
2058 1159118 : vacrel->tuples_frozen += presult.nfrozen;
2059 1159118 : vacrel->lpdead_items += presult.lpdead_items;
2060 1159118 : vacrel->live_tuples += presult.live_tuples;
2061 1159118 : vacrel->recently_dead_tuples += presult.recently_dead_tuples;
2062 :
2063 : /* Can't truncate this page */
2064 1159118 : if (presult.hastup)
2065 1144324 : vacrel->nonempty_pages = blkno + 1;
2066 :
2067 : /* Did we find LP_DEAD items? */
2068 1159118 : *has_lpdead_items = (presult.lpdead_items > 0);
2069 :
2070 : Assert(!presult.all_visible || !(*has_lpdead_items));
2071 :
2072 : /*
2073 : * Handle setting visibility map bit based on information from the VM (as
2074 : * of last heap_vac_scan_next_block() call), and from all_visible and
2075 : * all_frozen variables
2076 : */
2077 1159118 : if (!all_visible_according_to_vm && presult.all_visible)
2078 71622 : {
2079 : uint8 old_vmbits;
2080 71622 : uint8 flags = VISIBILITYMAP_ALL_VISIBLE;
2081 :
2082 71622 : if (presult.all_frozen)
2083 : {
2084 : Assert(!TransactionIdIsValid(presult.vm_conflict_horizon));
2085 53406 : flags |= VISIBILITYMAP_ALL_FROZEN;
2086 : }
2087 :
2088 : /*
2089 : * It should never be the case that the visibility map page is set
2090 : * while the page-level bit is clear, but the reverse is allowed (if
2091 : * checksums are not enabled). Regardless, set both bits so that we
2092 : * get back in sync.
2093 : *
2094 : * NB: If the heap page is all-visible but the VM bit is not set, we
2095 : * don't need to dirty the heap page. However, if checksums are
2096 : * enabled, we do need to make sure that the heap page is dirtied
2097 : * before passing it to visibilitymap_set(), because it may be logged.
2098 : * Given that this situation should only happen in rare cases after a
2099 : * crash, it is not worth optimizing.
2100 : */
2101 71622 : PageSetAllVisible(page);
2102 71622 : MarkBufferDirty(buf);
2103 71622 : old_vmbits = visibilitymap_set(vacrel->rel, blkno, buf,
2104 : InvalidXLogRecPtr,
2105 : vmbuffer, presult.vm_conflict_horizon,
2106 : flags);
2107 :
2108 : /*
2109 : * If the page wasn't already set all-visible and/or all-frozen in the
2110 : * VM, count it as newly set for logging.
2111 : */
2112 71622 : if ((old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
2113 : {
2114 71622 : vacrel->vm_new_visible_pages++;
2115 71622 : if (presult.all_frozen)
2116 : {
2117 53406 : vacrel->vm_new_visible_frozen_pages++;
2118 53406 : *vm_page_frozen = true;
2119 : }
2120 : }
2121 0 : else if ((old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
2122 0 : presult.all_frozen)
2123 : {
2124 0 : vacrel->vm_new_frozen_pages++;
2125 0 : *vm_page_frozen = true;
2126 : }
2127 : }
2128 :
2129 : /*
2130 : * As of PostgreSQL 9.2, the visibility map bit should never be set if the
2131 : * page-level bit is clear. However, it's possible that the bit got
2132 : * cleared after heap_vac_scan_next_block() was called, so we must recheck
2133 : * with buffer lock before concluding that the VM is corrupt.
2134 : */
2135 1087496 : else if (all_visible_according_to_vm && !PageIsAllVisible(page) &&
2136 0 : visibilitymap_get_status(vacrel->rel, blkno, &vmbuffer) != 0)
2137 : {
2138 0 : ereport(WARNING,
2139 : (errcode(ERRCODE_DATA_CORRUPTED),
2140 : errmsg("page is not marked all-visible but visibility map bit is set in relation \"%s\" page %u",
2141 : vacrel->relname, blkno)));
2142 :
2143 0 : visibilitymap_clear(vacrel->rel, blkno, vmbuffer,
2144 : VISIBILITYMAP_VALID_BITS);
2145 : }
2146 :
2147 : /*
2148 : * It's possible for the value returned by
2149 : * GetOldestNonRemovableTransactionId() to move backwards, so it's not
2150 : * wrong for us to see tuples that appear to not be visible to everyone
2151 : * yet, while PD_ALL_VISIBLE is already set. The real safe xmin value
2152 : * never moves backwards, but GetOldestNonRemovableTransactionId() is
2153 : * conservative and sometimes returns a value that's unnecessarily small,
2154 : * so if we see that contradiction it just means that the tuples that we
2155 : * think are not visible to everyone yet actually are, and the
2156 : * PD_ALL_VISIBLE flag is correct.
2157 : *
2158 : * There should never be LP_DEAD items on a page with PD_ALL_VISIBLE set,
2159 : * however.
2160 : */
2161 1087496 : else if (presult.lpdead_items > 0 && PageIsAllVisible(page))
2162 : {
2163 0 : ereport(WARNING,
2164 : (errcode(ERRCODE_DATA_CORRUPTED),
2165 : errmsg("page containing LP_DEAD items is marked as all-visible in relation \"%s\" page %u",
2166 : vacrel->relname, blkno)));
2167 :
2168 0 : PageClearAllVisible(page);
2169 0 : MarkBufferDirty(buf);
2170 0 : visibilitymap_clear(vacrel->rel, blkno, vmbuffer,
2171 : VISIBILITYMAP_VALID_BITS);
2172 : }
2173 :
2174 : /*
2175 : * If the all-visible page is all-frozen but not marked as such yet, mark
2176 : * it as all-frozen. Note that all_frozen is only valid if all_visible is
2177 : * true, so we must check both all_visible and all_frozen.
2178 : */
2179 1087496 : else if (all_visible_according_to_vm && presult.all_visible &&
2180 406642 : presult.all_frozen && !VM_ALL_FROZEN(vacrel->rel, blkno, &vmbuffer))
2181 : {
2182 : uint8 old_vmbits;
2183 :
2184 : /*
2185 : * Avoid relying on all_visible_according_to_vm as a proxy for the
2186 : * page-level PD_ALL_VISIBLE bit being set, since it might have become
2187 : * stale -- even when all_visible is set
2188 : */
2189 42 : if (!PageIsAllVisible(page))
2190 : {
2191 0 : PageSetAllVisible(page);
2192 0 : MarkBufferDirty(buf);
2193 : }
2194 :
2195 : /*
2196 : * Set the page all-frozen (and all-visible) in the VM.
2197 : *
2198 : * We can pass InvalidTransactionId as our cutoff_xid, since a
2199 : * snapshotConflictHorizon sufficient to make everything safe for REDO
2200 : * was logged when the page's tuples were frozen.
2201 : */
2202 : Assert(!TransactionIdIsValid(presult.vm_conflict_horizon));
2203 42 : old_vmbits = visibilitymap_set(vacrel->rel, blkno, buf,
2204 : InvalidXLogRecPtr,
2205 : vmbuffer, InvalidTransactionId,
2206 : VISIBILITYMAP_ALL_VISIBLE |
2207 : VISIBILITYMAP_ALL_FROZEN);
2208 :
2209 : /*
2210 : * The page was likely already set all-visible in the VM. However,
2211 : * there is a small chance that it was modified sometime between
2212 : * setting all_visible_according_to_vm and checking the visibility
2213 : * during pruning. Check the return value of old_vmbits anyway to
2214 : * ensure the visibility map counters used for logging are accurate.
2215 : */
2216 42 : if ((old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
2217 : {
2218 0 : vacrel->vm_new_visible_pages++;
2219 0 : vacrel->vm_new_visible_frozen_pages++;
2220 0 : *vm_page_frozen = true;
2221 : }
2222 :
2223 : /*
2224 : * We already checked that the page was not set all-frozen in the VM
2225 : * above, so we don't need to test the value of old_vmbits.
2226 : */
2227 : else
2228 : {
2229 42 : vacrel->vm_new_frozen_pages++;
2230 42 : *vm_page_frozen = true;
2231 : }
2232 : }
2233 :
2234 1159118 : return presult.ndeleted;
2235 : }
2236 :
2237 : /*
2238 : * lazy_scan_noprune() -- lazy_scan_prune() without pruning or freezing
2239 : *
2240 : * Caller need only hold a pin and share lock on the buffer, unlike
2241 : * lazy_scan_prune, which requires a full cleanup lock. While pruning isn't
2242 : * performed here, it's quite possible that an earlier opportunistic pruning
2243 : * operation left LP_DEAD items behind. We'll at least collect any such items
2244 : * in dead_items for removal from indexes.
2245 : *
2246 : * For aggressive VACUUM callers, we may return false to indicate that a full
2247 : * cleanup lock is required for processing by lazy_scan_prune. This is only
2248 : * necessary when the aggressive VACUUM needs to freeze some tuple XIDs from
2249 : * one or more tuples on the page. We always return true for non-aggressive
2250 : * callers.
2251 : *
2252 : * If this function returns true, *has_lpdead_items gets set to true or false
2253 : * depending on whether, upon return from this function, any LP_DEAD items are
2254 : * present on the page. If this function returns false, *has_lpdead_items
2255 : * is not updated.
2256 : */
2257 : static bool
2258 312 : lazy_scan_noprune(LVRelState *vacrel,
2259 : Buffer buf,
2260 : BlockNumber blkno,
2261 : Page page,
2262 : bool *has_lpdead_items)
2263 : {
2264 : OffsetNumber offnum,
2265 : maxoff;
2266 : int lpdead_items,
2267 : live_tuples,
2268 : recently_dead_tuples,
2269 : missed_dead_tuples;
2270 : bool hastup;
2271 : HeapTupleHeader tupleheader;
2272 312 : TransactionId NoFreezePageRelfrozenXid = vacrel->NewRelfrozenXid;
2273 312 : MultiXactId NoFreezePageRelminMxid = vacrel->NewRelminMxid;
2274 : OffsetNumber deadoffsets[MaxHeapTuplesPerPage];
2275 :
2276 : Assert(BufferGetBlockNumber(buf) == blkno);
2277 :
2278 312 : hastup = false; /* for now */
2279 :
2280 312 : lpdead_items = 0;
2281 312 : live_tuples = 0;
2282 312 : recently_dead_tuples = 0;
2283 312 : missed_dead_tuples = 0;
2284 :
2285 312 : maxoff = PageGetMaxOffsetNumber(page);
2286 312 : for (offnum = FirstOffsetNumber;
2287 8296 : offnum <= maxoff;
2288 7984 : offnum = OffsetNumberNext(offnum))
2289 : {
2290 : ItemId itemid;
2291 : HeapTupleData tuple;
2292 :
2293 8098 : vacrel->offnum = offnum;
2294 8098 : itemid = PageGetItemId(page, offnum);
2295 :
2296 8098 : if (!ItemIdIsUsed(itemid))
2297 2198 : continue;
2298 :
2299 6836 : if (ItemIdIsRedirected(itemid))
2300 : {
2301 514 : hastup = true;
2302 514 : continue;
2303 : }
2304 :
2305 6322 : if (ItemIdIsDead(itemid))
2306 : {
2307 : /*
2308 : * Deliberately don't set hastup=true here. See same point in
2309 : * lazy_scan_prune for an explanation.
2310 : */
2311 422 : deadoffsets[lpdead_items++] = offnum;
2312 422 : continue;
2313 : }
2314 :
2315 5900 : hastup = true; /* page prevents rel truncation */
2316 5900 : tupleheader = (HeapTupleHeader) PageGetItem(page, itemid);
2317 5900 : if (heap_tuple_should_freeze(tupleheader, &vacrel->cutoffs,
2318 : &NoFreezePageRelfrozenXid,
2319 : &NoFreezePageRelminMxid))
2320 : {
2321 : /* Tuple with XID < FreezeLimit (or MXID < MultiXactCutoff) */
2322 242 : if (vacrel->aggressive)
2323 : {
2324 : /*
2325 : * Aggressive VACUUMs must always be able to advance rel's
2326 : * relfrozenxid to a value >= FreezeLimit (and be able to
2327 : * advance rel's relminmxid to a value >= MultiXactCutoff).
2328 : * The ongoing aggressive VACUUM won't be able to do that
2329 : * unless it can freeze an XID (or MXID) from this tuple now.
2330 : *
2331 : * The only safe option is to have caller perform processing
2332 : * of this page using lazy_scan_prune. Caller might have to
2333 : * wait a while for a cleanup lock, but it can't be helped.
2334 : */
2335 114 : vacrel->offnum = InvalidOffsetNumber;
2336 114 : return false;
2337 : }
2338 :
2339 : /*
2340 : * Non-aggressive VACUUMs are under no obligation to advance
2341 : * relfrozenxid (even by one XID). We can be much laxer here.
2342 : *
2343 : * Currently we always just accept an older final relfrozenxid
2344 : * and/or relminmxid value. We never make caller wait or work a
2345 : * little harder, even when it likely makes sense to do so.
2346 : */
2347 : }
2348 :
2349 5786 : ItemPointerSet(&(tuple.t_self), blkno, offnum);
2350 5786 : tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
2351 5786 : tuple.t_len = ItemIdGetLength(itemid);
2352 5786 : tuple.t_tableOid = RelationGetRelid(vacrel->rel);
2353 :
2354 5786 : switch (HeapTupleSatisfiesVacuum(&tuple, vacrel->cutoffs.OldestXmin,
2355 : buf))
2356 : {
2357 5586 : case HEAPTUPLE_DELETE_IN_PROGRESS:
2358 : case HEAPTUPLE_LIVE:
2359 :
2360 : /*
2361 : * Count both cases as live, just like lazy_scan_prune
2362 : */
2363 5586 : live_tuples++;
2364 :
2365 5586 : break;
2366 196 : case HEAPTUPLE_DEAD:
2367 :
2368 : /*
2369 : * There is some useful work for pruning to do, that won't be
2370 : * done due to failure to get a cleanup lock.
2371 : */
2372 196 : missed_dead_tuples++;
2373 196 : break;
2374 4 : case HEAPTUPLE_RECENTLY_DEAD:
2375 :
2376 : /*
2377 : * Count in recently_dead_tuples, just like lazy_scan_prune
2378 : */
2379 4 : recently_dead_tuples++;
2380 4 : break;
2381 0 : case HEAPTUPLE_INSERT_IN_PROGRESS:
2382 :
2383 : /*
2384 : * Do not count these rows as live, just like lazy_scan_prune
2385 : */
2386 0 : break;
2387 0 : default:
2388 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
2389 : break;
2390 : }
2391 : }
2392 :
2393 198 : vacrel->offnum = InvalidOffsetNumber;
2394 :
2395 : /*
2396 : * By here we know for sure that caller can put off freezing and pruning
2397 : * this particular page until the next VACUUM. Remember its details now.
2398 : * (lazy_scan_prune expects a clean slate, so we have to do this last.)
2399 : */
2400 198 : vacrel->NewRelfrozenXid = NoFreezePageRelfrozenXid;
2401 198 : vacrel->NewRelminMxid = NoFreezePageRelminMxid;
2402 :
2403 : /* Save any LP_DEAD items found on the page in dead_items */
2404 198 : if (vacrel->nindexes == 0)
2405 : {
2406 : /* Using one-pass strategy (since table has no indexes) */
2407 0 : if (lpdead_items > 0)
2408 : {
2409 : /*
2410 : * Perfunctory handling for the corner case where a single pass
2411 : * strategy VACUUM cannot get a cleanup lock, and it turns out
2412 : * that there is one or more LP_DEAD items: just count the LP_DEAD
2413 : * items as missed_dead_tuples instead. (This is a bit dishonest,
2414 : * but it beats having to maintain specialized heap vacuuming code
2415 : * forever, for vanishingly little benefit.)
2416 : */
2417 0 : hastup = true;
2418 0 : missed_dead_tuples += lpdead_items;
2419 : }
2420 : }
2421 198 : else if (lpdead_items > 0)
2422 : {
2423 : /*
2424 : * Page has LP_DEAD items, and so any references/TIDs that remain in
2425 : * indexes will be deleted during index vacuuming (and then marked
2426 : * LP_UNUSED in the heap)
2427 : */
2428 36 : vacrel->lpdead_item_pages++;
2429 :
2430 36 : dead_items_add(vacrel, blkno, deadoffsets, lpdead_items);
2431 :
2432 36 : vacrel->lpdead_items += lpdead_items;
2433 : }
2434 :
2435 : /*
2436 : * Finally, add relevant page-local counts to whole-VACUUM counts
2437 : */
2438 198 : vacrel->live_tuples += live_tuples;
2439 198 : vacrel->recently_dead_tuples += recently_dead_tuples;
2440 198 : vacrel->missed_dead_tuples += missed_dead_tuples;
2441 198 : if (missed_dead_tuples > 0)
2442 22 : vacrel->missed_dead_pages++;
2443 :
2444 : /* Can't truncate this page */
2445 198 : if (hastup)
2446 192 : vacrel->nonempty_pages = blkno + 1;
2447 :
2448 : /* Did we find LP_DEAD items? */
2449 198 : *has_lpdead_items = (lpdead_items > 0);
2450 :
2451 : /* Caller won't need to call lazy_scan_prune with same page */
2452 198 : return true;
2453 : }
2454 :
2455 : /*
2456 : * Main entry point for index vacuuming and heap vacuuming.
2457 : *
2458 : * Removes items collected in dead_items from table's indexes, then marks the
2459 : * same items LP_UNUSED in the heap. See the comments above lazy_scan_heap
2460 : * for full details.
2461 : *
2462 : * Also empties dead_items, freeing up space for later TIDs.
2463 : *
2464 : * We may choose to bypass index vacuuming at this point, though only when the
2465 : * ongoing VACUUM operation will definitely only have one index scan/round of
2466 : * index vacuuming.
2467 : */
2468 : static void
2469 1282 : lazy_vacuum(LVRelState *vacrel)
2470 : {
2471 : bool bypass;
2472 :
2473 : /* Should not end up here with no indexes */
2474 : Assert(vacrel->nindexes > 0);
2475 : Assert(vacrel->lpdead_item_pages > 0);
2476 :
2477 1282 : if (!vacrel->do_index_vacuuming)
2478 : {
2479 : Assert(!vacrel->do_index_cleanup);
2480 18 : dead_items_reset(vacrel);
2481 18 : return;
2482 : }
2483 :
2484 : /*
2485 : * Consider bypassing index vacuuming (and heap vacuuming) entirely.
2486 : *
2487 : * We currently only do this in cases where the number of LP_DEAD items
2488 : * for the entire VACUUM operation is close to zero. This avoids sharp
2489 : * discontinuities in the duration and overhead of successive VACUUM
2490 : * operations that run against the same table with a fixed workload.
2491 : * Ideally, successive VACUUM operations will behave as if there are
2492 : * exactly zero LP_DEAD items in cases where there are close to zero.
2493 : *
2494 : * This is likely to be helpful with a table that is continually affected
2495 : * by UPDATEs that can mostly apply the HOT optimization, but occasionally
2496 : * have small aberrations that lead to just a few heap pages retaining
2497 : * only one or two LP_DEAD items. This is pretty common; even when the
2498 : * DBA goes out of their way to make UPDATEs use HOT, it is practically
2499 : * impossible to predict whether HOT will be applied in 100% of cases.
2500 : * It's far easier to ensure that 99%+ of all UPDATEs against a table use
2501 : * HOT through careful tuning.
2502 : */
2503 1264 : bypass = false;
2504 1264 : if (vacrel->consider_bypass_optimization && vacrel->rel_pages > 0)
2505 : {
2506 : BlockNumber threshold;
2507 :
2508 : Assert(vacrel->num_index_scans == 0);
2509 : Assert(vacrel->lpdead_items == vacrel->dead_items_info->num_items);
2510 : Assert(vacrel->do_index_vacuuming);
2511 : Assert(vacrel->do_index_cleanup);
2512 :
2513 : /*
2514 : * This crossover point at which we'll start to do index vacuuming is
2515 : * expressed as a percentage of the total number of heap pages in the
2516 : * table that are known to have at least one LP_DEAD item. This is
2517 : * much more important than the total number of LP_DEAD items, since
2518 : * it's a proxy for the number of heap pages whose visibility map bits
2519 : * cannot be set on account of bypassing index and heap vacuuming.
2520 : *
2521 : * We apply one further precautionary test: the space currently used
2522 : * to store the TIDs (TIDs that now all point to LP_DEAD items) must
2523 : * not exceed 32MB. This limits the risk that we will bypass index
2524 : * vacuuming again and again until eventually there is a VACUUM whose
2525 : * dead_items space is not CPU cache resident.
2526 : *
2527 : * We don't take any special steps to remember the LP_DEAD items (such
2528 : * as counting them in our final update to the stats system) when the
2529 : * optimization is applied. Though the accounting used in analyze.c's
2530 : * acquire_sample_rows() will recognize the same LP_DEAD items as dead
2531 : * rows in its own stats report, that's okay. The discrepancy should
2532 : * be negligible. If this optimization is ever expanded to cover more
2533 : * cases then this may need to be reconsidered.
2534 : */
2535 1236 : threshold = (double) vacrel->rel_pages * BYPASS_THRESHOLD_PAGES;
2536 1238 : bypass = (vacrel->lpdead_item_pages < threshold &&
2537 2 : TidStoreMemoryUsage(vacrel->dead_items) < 32 * 1024 * 1024);
2538 : }
2539 :
2540 1264 : if (bypass)
2541 : {
2542 : /*
2543 : * There are almost zero TIDs. Behave as if there were precisely
2544 : * zero: bypass index vacuuming, but do index cleanup.
2545 : *
2546 : * We expect that the ongoing VACUUM operation will finish very
2547 : * quickly, so there is no point in considering speeding up as a
2548 : * failsafe against wraparound failure. (Index cleanup is expected to
2549 : * finish very quickly in cases where there were no ambulkdelete()
2550 : * calls.)
2551 : */
2552 2 : vacrel->do_index_vacuuming = false;
2553 : }
2554 1262 : else if (lazy_vacuum_all_indexes(vacrel))
2555 : {
2556 : /*
2557 : * We successfully completed a round of index vacuuming. Do related
2558 : * heap vacuuming now.
2559 : */
2560 1262 : lazy_vacuum_heap_rel(vacrel);
2561 : }
2562 : else
2563 : {
2564 : /*
2565 : * Failsafe case.
2566 : *
2567 : * We attempted index vacuuming, but didn't finish a full round/full
2568 : * index scan. This happens when relfrozenxid or relminmxid is too
2569 : * far in the past.
2570 : *
2571 : * From this point on the VACUUM operation will do no further index
2572 : * vacuuming or heap vacuuming. This VACUUM operation won't end up
2573 : * back here again.
2574 : */
2575 : Assert(VacuumFailsafeActive);
2576 : }
2577 :
2578 : /*
2579 : * Forget the LP_DEAD items that we just vacuumed (or just decided to not
2580 : * vacuum)
2581 : */
2582 1264 : dead_items_reset(vacrel);
2583 : }
2584 :
2585 : /*
2586 : * lazy_vacuum_all_indexes() -- Main entry for index vacuuming
2587 : *
2588 : * Returns true in the common case when all indexes were successfully
2589 : * vacuumed. Returns false in rare cases where we determined that the ongoing
2590 : * VACUUM operation is at risk of taking too long to finish, leading to
2591 : * wraparound failure.
2592 : */
2593 : static bool
2594 1262 : lazy_vacuum_all_indexes(LVRelState *vacrel)
2595 : {
2596 1262 : bool allindexes = true;
2597 1262 : double old_live_tuples = vacrel->rel->rd_rel->reltuples;
2598 1262 : const int progress_start_index[] = {
2599 : PROGRESS_VACUUM_PHASE,
2600 : PROGRESS_VACUUM_INDEXES_TOTAL
2601 : };
2602 1262 : const int progress_end_index[] = {
2603 : PROGRESS_VACUUM_INDEXES_TOTAL,
2604 : PROGRESS_VACUUM_INDEXES_PROCESSED,
2605 : PROGRESS_VACUUM_NUM_INDEX_VACUUMS
2606 : };
2607 : int64 progress_start_val[2];
2608 : int64 progress_end_val[3];
2609 :
2610 : Assert(vacrel->nindexes > 0);
2611 : Assert(vacrel->do_index_vacuuming);
2612 : Assert(vacrel->do_index_cleanup);
2613 :
2614 : /* Precheck for XID wraparound emergencies */
2615 1262 : if (lazy_check_wraparound_failsafe(vacrel))
2616 : {
2617 : /* Wraparound emergency -- don't even start an index scan */
2618 0 : return false;
2619 : }
2620 :
2621 : /*
2622 : * Report that we are now vacuuming indexes and the number of indexes to
2623 : * vacuum.
2624 : */
2625 1262 : progress_start_val[0] = PROGRESS_VACUUM_PHASE_VACUUM_INDEX;
2626 1262 : progress_start_val[1] = vacrel->nindexes;
2627 1262 : pgstat_progress_update_multi_param(2, progress_start_index, progress_start_val);
2628 :
2629 1262 : if (!ParallelVacuumIsActive(vacrel))
2630 : {
2631 3638 : for (int idx = 0; idx < vacrel->nindexes; idx++)
2632 : {
2633 2400 : Relation indrel = vacrel->indrels[idx];
2634 2400 : IndexBulkDeleteResult *istat = vacrel->indstats[idx];
2635 :
2636 2400 : vacrel->indstats[idx] = lazy_vacuum_one_index(indrel, istat,
2637 : old_live_tuples,
2638 : vacrel);
2639 :
2640 : /* Report the number of indexes vacuumed */
2641 2400 : pgstat_progress_update_param(PROGRESS_VACUUM_INDEXES_PROCESSED,
2642 2400 : idx + 1);
2643 :
2644 2400 : if (lazy_check_wraparound_failsafe(vacrel))
2645 : {
2646 : /* Wraparound emergency -- end current index scan */
2647 0 : allindexes = false;
2648 0 : break;
2649 : }
2650 : }
2651 : }
2652 : else
2653 : {
2654 : /* Outsource everything to parallel variant */
2655 24 : parallel_vacuum_bulkdel_all_indexes(vacrel->pvs, old_live_tuples,
2656 : vacrel->num_index_scans);
2657 :
2658 : /*
2659 : * Do a postcheck to consider applying wraparound failsafe now. Note
2660 : * that parallel VACUUM only gets the precheck and this postcheck.
2661 : */
2662 24 : if (lazy_check_wraparound_failsafe(vacrel))
2663 0 : allindexes = false;
2664 : }
2665 :
2666 : /*
2667 : * We delete all LP_DEAD items from the first heap pass in all indexes on
2668 : * each call here (except calls where we choose to do the failsafe). This
2669 : * makes the next call to lazy_vacuum_heap_rel() safe (except in the event
2670 : * of the failsafe triggering, which prevents the next call from taking
2671 : * place).
2672 : */
2673 : Assert(vacrel->num_index_scans > 0 ||
2674 : vacrel->dead_items_info->num_items == vacrel->lpdead_items);
2675 : Assert(allindexes || VacuumFailsafeActive);
2676 :
2677 : /*
2678 : * Increase and report the number of index scans. Also, we reset
2679 : * PROGRESS_VACUUM_INDEXES_TOTAL and PROGRESS_VACUUM_INDEXES_PROCESSED.
2680 : *
2681 : * We deliberately include the case where we started a round of bulk
2682 : * deletes that we weren't able to finish due to the failsafe triggering.
2683 : */
2684 1262 : vacrel->num_index_scans++;
2685 1262 : progress_end_val[0] = 0;
2686 1262 : progress_end_val[1] = 0;
2687 1262 : progress_end_val[2] = vacrel->num_index_scans;
2688 1262 : pgstat_progress_update_multi_param(3, progress_end_index, progress_end_val);
2689 :
2690 1262 : return allindexes;
2691 : }
2692 :
2693 : /*
2694 : * Read stream callback for vacuum's third phase (second pass over the heap).
2695 : * Gets the next block from the TID store and returns it or InvalidBlockNumber
2696 : * if there are no further blocks to vacuum.
2697 : *
2698 : * NB: Assumed to be safe to use with READ_STREAM_USE_BATCHING.
2699 : */
2700 : static BlockNumber
2701 27058 : vacuum_reap_lp_read_stream_next(ReadStream *stream,
2702 : void *callback_private_data,
2703 : void *per_buffer_data)
2704 : {
2705 27058 : TidStoreIter *iter = callback_private_data;
2706 : TidStoreIterResult *iter_result;
2707 :
2708 27058 : iter_result = TidStoreIterateNext(iter);
2709 27058 : if (iter_result == NULL)
2710 1262 : return InvalidBlockNumber;
2711 :
2712 : /*
2713 : * Save the TidStoreIterResult for later, so we can extract the offsets.
2714 : * It is safe to copy the result, according to TidStoreIterateNext().
2715 : */
2716 25796 : memcpy(per_buffer_data, iter_result, sizeof(*iter_result));
2717 :
2718 25796 : return iter_result->blkno;
2719 : }
2720 :
2721 : /*
2722 : * lazy_vacuum_heap_rel() -- second pass over the heap for two pass strategy
2723 : *
2724 : * This routine marks LP_DEAD items in vacrel->dead_items as LP_UNUSED. Pages
2725 : * that never had lazy_scan_prune record LP_DEAD items are not visited at all.
2726 : *
2727 : * We may also be able to truncate the line pointer array of the heap pages we
2728 : * visit. If there is a contiguous group of LP_UNUSED items at the end of the
2729 : * array, it can be reclaimed as free space. These LP_UNUSED items usually
2730 : * start out as LP_DEAD items recorded by lazy_scan_prune (we set items from
2731 : * each page to LP_UNUSED, and then consider if it's possible to truncate the
2732 : * page's line pointer array).
2733 : *
2734 : * Note: the reason for doing this as a second pass is we cannot remove the
2735 : * tuples until we've removed their index entries, and we want to process
2736 : * index entry removal in batches as large as possible.
2737 : */
2738 : static void
2739 1262 : lazy_vacuum_heap_rel(LVRelState *vacrel)
2740 : {
2741 : ReadStream *stream;
2742 1262 : BlockNumber vacuumed_pages = 0;
2743 1262 : Buffer vmbuffer = InvalidBuffer;
2744 : LVSavedErrInfo saved_err_info;
2745 : TidStoreIter *iter;
2746 :
2747 : Assert(vacrel->do_index_vacuuming);
2748 : Assert(vacrel->do_index_cleanup);
2749 : Assert(vacrel->num_index_scans > 0);
2750 :
2751 : /* Report that we are now vacuuming the heap */
2752 1262 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
2753 : PROGRESS_VACUUM_PHASE_VACUUM_HEAP);
2754 :
2755 : /* Update error traceback information */
2756 1262 : update_vacuum_error_info(vacrel, &saved_err_info,
2757 : VACUUM_ERRCB_PHASE_VACUUM_HEAP,
2758 : InvalidBlockNumber, InvalidOffsetNumber);
2759 :
2760 1262 : iter = TidStoreBeginIterate(vacrel->dead_items);
2761 :
2762 : /*
2763 : * Set up the read stream for vacuum's second pass through the heap.
2764 : *
2765 : * It is safe to use batchmode, as vacuum_reap_lp_read_stream_next() does
2766 : * not need to wait for IO and does not perform locking. Once we support
2767 : * parallelism it should still be fine, as presumably the holder of locks
2768 : * would never be blocked by IO while holding the lock.
2769 : */
2770 1262 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
2771 : READ_STREAM_USE_BATCHING,
2772 : vacrel->bstrategy,
2773 : vacrel->rel,
2774 : MAIN_FORKNUM,
2775 : vacuum_reap_lp_read_stream_next,
2776 : iter,
2777 : sizeof(TidStoreIterResult));
2778 :
2779 : while (true)
2780 25796 : {
2781 : BlockNumber blkno;
2782 : Buffer buf;
2783 : Page page;
2784 : TidStoreIterResult *iter_result;
2785 : Size freespace;
2786 : OffsetNumber offsets[MaxOffsetNumber];
2787 : int num_offsets;
2788 :
2789 27058 : vacuum_delay_point(false);
2790 :
2791 27058 : buf = read_stream_next_buffer(stream, (void **) &iter_result);
2792 :
2793 : /* The relation is exhausted */
2794 27058 : if (!BufferIsValid(buf))
2795 1262 : break;
2796 :
2797 25796 : vacrel->blkno = blkno = BufferGetBlockNumber(buf);
2798 :
2799 : Assert(iter_result);
2800 25796 : num_offsets = TidStoreGetBlockOffsets(iter_result, offsets, lengthof(offsets));
2801 : Assert(num_offsets <= lengthof(offsets));
2802 :
2803 : /*
2804 : * Pin the visibility map page in case we need to mark the page
2805 : * all-visible. In most cases this will be very cheap, because we'll
2806 : * already have the correct page pinned anyway.
2807 : */
2808 25796 : visibilitymap_pin(vacrel->rel, blkno, &vmbuffer);
2809 :
2810 : /* We need a non-cleanup exclusive lock to mark dead_items unused */
2811 25796 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2812 25796 : lazy_vacuum_heap_page(vacrel, blkno, buf, offsets,
2813 : num_offsets, vmbuffer);
2814 :
2815 : /* Now that we've vacuumed the page, record its available space */
2816 25796 : page = BufferGetPage(buf);
2817 25796 : freespace = PageGetHeapFreeSpace(page);
2818 :
2819 25796 : UnlockReleaseBuffer(buf);
2820 25796 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
2821 25796 : vacuumed_pages++;
2822 : }
2823 :
2824 1262 : read_stream_end(stream);
2825 1262 : TidStoreEndIterate(iter);
2826 :
2827 1262 : vacrel->blkno = InvalidBlockNumber;
2828 1262 : if (BufferIsValid(vmbuffer))
2829 1262 : ReleaseBuffer(vmbuffer);
2830 :
2831 : /*
2832 : * We set all LP_DEAD items from the first heap pass to LP_UNUSED during
2833 : * the second heap pass. No more, no less.
2834 : */
2835 : Assert(vacrel->num_index_scans > 1 ||
2836 : (vacrel->dead_items_info->num_items == vacrel->lpdead_items &&
2837 : vacuumed_pages == vacrel->lpdead_item_pages));
2838 :
2839 1262 : ereport(DEBUG2,
2840 : (errmsg("table \"%s\": removed %" PRId64 " dead item identifiers in %u pages",
2841 : vacrel->relname, vacrel->dead_items_info->num_items,
2842 : vacuumed_pages)));
2843 :
2844 : /* Revert to the previous phase information for error traceback */
2845 1262 : restore_vacuum_error_info(vacrel, &saved_err_info);
2846 1262 : }
2847 :
2848 : /*
2849 : * lazy_vacuum_heap_page() -- free page's LP_DEAD items listed in the
2850 : * vacrel->dead_items store.
2851 : *
2852 : * Caller must have an exclusive buffer lock on the buffer (though a full
2853 : * cleanup lock is also acceptable). vmbuffer must be valid and already have
2854 : * a pin on blkno's visibility map page.
2855 : */
2856 : static void
2857 25796 : lazy_vacuum_heap_page(LVRelState *vacrel, BlockNumber blkno, Buffer buffer,
2858 : OffsetNumber *deadoffsets, int num_offsets,
2859 : Buffer vmbuffer)
2860 : {
2861 25796 : Page page = BufferGetPage(buffer);
2862 : OffsetNumber unused[MaxHeapTuplesPerPage];
2863 25796 : int nunused = 0;
2864 : TransactionId visibility_cutoff_xid;
2865 25796 : TransactionId conflict_xid = InvalidTransactionId;
2866 : bool all_frozen;
2867 : LVSavedErrInfo saved_err_info;
2868 25796 : uint8 vmflags = 0;
2869 :
2870 : Assert(vacrel->do_index_vacuuming);
2871 :
2872 25796 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_VACUUMED, blkno);
2873 :
2874 : /* Update error traceback information */
2875 25796 : update_vacuum_error_info(vacrel, &saved_err_info,
2876 : VACUUM_ERRCB_PHASE_VACUUM_HEAP, blkno,
2877 : InvalidOffsetNumber);
2878 :
2879 : /*
2880 : * Before marking dead items unused, check whether the page will become
2881 : * all-visible once that change is applied. This lets us reap the tuples
2882 : * and mark the page all-visible within the same critical section,
2883 : * enabling both changes to be emitted in a single WAL record. Since the
2884 : * visibility checks may perform I/O and allocate memory, they must be
2885 : * done outside the critical section.
2886 : */
2887 25796 : if (heap_page_would_be_all_visible(vacrel->rel, buffer,
2888 : vacrel->cutoffs.OldestXmin,
2889 : deadoffsets, num_offsets,
2890 : &all_frozen, &visibility_cutoff_xid,
2891 : &vacrel->offnum))
2892 : {
2893 25480 : vmflags |= VISIBILITYMAP_ALL_VISIBLE;
2894 25480 : if (all_frozen)
2895 : {
2896 20316 : vmflags |= VISIBILITYMAP_ALL_FROZEN;
2897 : Assert(!TransactionIdIsValid(visibility_cutoff_xid));
2898 : }
2899 :
2900 : /*
2901 : * Take the lock on the vmbuffer before entering a critical section.
2902 : * The heap page lock must also be held while updating the VM to
2903 : * ensure consistency.
2904 : */
2905 25480 : LockBuffer(vmbuffer, BUFFER_LOCK_EXCLUSIVE);
2906 : }
2907 :
2908 25796 : START_CRIT_SECTION();
2909 :
2910 1614854 : for (int i = 0; i < num_offsets; i++)
2911 : {
2912 : ItemId itemid;
2913 1589058 : OffsetNumber toff = deadoffsets[i];
2914 :
2915 1589058 : itemid = PageGetItemId(page, toff);
2916 :
2917 : Assert(ItemIdIsDead(itemid) && !ItemIdHasStorage(itemid));
2918 1589058 : ItemIdSetUnused(itemid);
2919 1589058 : unused[nunused++] = toff;
2920 : }
2921 :
2922 : Assert(nunused > 0);
2923 :
2924 : /* Attempt to truncate line pointer array now */
2925 25796 : PageTruncateLinePointerArray(page);
2926 :
2927 25796 : if ((vmflags & VISIBILITYMAP_VALID_BITS) != 0)
2928 : {
2929 : /*
2930 : * The page is guaranteed to have had dead line pointers, so we always
2931 : * set PD_ALL_VISIBLE.
2932 : */
2933 25480 : PageSetAllVisible(page);
2934 25480 : visibilitymap_set_vmbits(blkno,
2935 : vmbuffer, vmflags,
2936 25480 : vacrel->rel->rd_locator);
2937 25480 : conflict_xid = visibility_cutoff_xid;
2938 : }
2939 :
2940 : /*
2941 : * Mark buffer dirty before we write WAL.
2942 : */
2943 25796 : MarkBufferDirty(buffer);
2944 :
2945 : /* XLOG stuff */
2946 25796 : if (RelationNeedsWAL(vacrel->rel))
2947 : {
2948 24098 : log_heap_prune_and_freeze(vacrel->rel, buffer,
2949 : vmflags != 0 ? vmbuffer : InvalidBuffer,
2950 : vmflags,
2951 : conflict_xid,
2952 : false, /* no cleanup lock required */
2953 : PRUNE_VACUUM_CLEANUP,
2954 : NULL, 0, /* frozen */
2955 : NULL, 0, /* redirected */
2956 : NULL, 0, /* dead */
2957 : unused, nunused);
2958 : }
2959 :
2960 25796 : END_CRIT_SECTION();
2961 :
2962 25796 : if ((vmflags & VISIBILITYMAP_ALL_VISIBLE) != 0)
2963 : {
2964 : /* Count the newly set VM page for logging */
2965 25480 : LockBuffer(vmbuffer, BUFFER_LOCK_UNLOCK);
2966 25480 : vacrel->vm_new_visible_pages++;
2967 25480 : if (all_frozen)
2968 20316 : vacrel->vm_new_visible_frozen_pages++;
2969 : }
2970 :
2971 : /* Revert to the previous phase information for error traceback */
2972 25796 : restore_vacuum_error_info(vacrel, &saved_err_info);
2973 25796 : }
2974 :
2975 : /*
2976 : * Trigger the failsafe to avoid wraparound failure when vacrel table has a
2977 : * relfrozenxid and/or relminmxid that is dangerously far in the past.
2978 : * Triggering the failsafe makes the ongoing VACUUM bypass any further index
2979 : * vacuuming and heap vacuuming. Truncating the heap is also bypassed.
2980 : *
2981 : * Any remaining work (work that VACUUM cannot just bypass) is typically sped
2982 : * up when the failsafe triggers. VACUUM stops applying any cost-based delay
2983 : * that it started out with.
2984 : *
2985 : * Returns true when failsafe has been triggered.
2986 : */
2987 : static bool
2988 261258 : lazy_check_wraparound_failsafe(LVRelState *vacrel)
2989 : {
2990 : /* Don't warn more than once per VACUUM */
2991 261258 : if (VacuumFailsafeActive)
2992 0 : return true;
2993 :
2994 261258 : if (unlikely(vacuum_xid_failsafe_check(&vacrel->cutoffs)))
2995 : {
2996 69938 : const int progress_index[] = {
2997 : PROGRESS_VACUUM_INDEXES_TOTAL,
2998 : PROGRESS_VACUUM_INDEXES_PROCESSED
2999 : };
3000 69938 : int64 progress_val[2] = {0, 0};
3001 :
3002 69938 : VacuumFailsafeActive = true;
3003 :
3004 : /*
3005 : * Abandon use of a buffer access strategy to allow use of all of
3006 : * shared buffers. We assume the caller who allocated the memory for
3007 : * the BufferAccessStrategy will free it.
3008 : */
3009 69938 : vacrel->bstrategy = NULL;
3010 :
3011 : /* Disable index vacuuming, index cleanup, and heap rel truncation */
3012 69938 : vacrel->do_index_vacuuming = false;
3013 69938 : vacrel->do_index_cleanup = false;
3014 69938 : vacrel->do_rel_truncate = false;
3015 :
3016 : /* Reset the progress counters */
3017 69938 : pgstat_progress_update_multi_param(2, progress_index, progress_val);
3018 :
3019 69938 : ereport(WARNING,
3020 : (errmsg("bypassing nonessential maintenance of table \"%s.%s.%s\" as a failsafe after %d index scans",
3021 : vacrel->dbname, vacrel->relnamespace, vacrel->relname,
3022 : vacrel->num_index_scans),
3023 : errdetail("The table's relfrozenxid or relminmxid is too far in the past."),
3024 : errhint("Consider increasing configuration parameter \"maintenance_work_mem\" or \"autovacuum_work_mem\".\n"
3025 : "You might also need to consider other ways for VACUUM to keep up with the allocation of transaction IDs.")));
3026 :
3027 : /* Stop applying cost limits from this point on */
3028 69938 : VacuumCostActive = false;
3029 69938 : VacuumCostBalance = 0;
3030 :
3031 69938 : return true;
3032 : }
3033 :
3034 191320 : return false;
3035 : }
3036 :
3037 : /*
3038 : * lazy_cleanup_all_indexes() -- cleanup all indexes of relation.
3039 : */
3040 : static void
3041 178248 : lazy_cleanup_all_indexes(LVRelState *vacrel)
3042 : {
3043 178248 : double reltuples = vacrel->new_rel_tuples;
3044 178248 : bool estimated_count = vacrel->scanned_pages < vacrel->rel_pages;
3045 178248 : const int progress_start_index[] = {
3046 : PROGRESS_VACUUM_PHASE,
3047 : PROGRESS_VACUUM_INDEXES_TOTAL
3048 : };
3049 178248 : const int progress_end_index[] = {
3050 : PROGRESS_VACUUM_INDEXES_TOTAL,
3051 : PROGRESS_VACUUM_INDEXES_PROCESSED
3052 : };
3053 : int64 progress_start_val[2];
3054 178248 : int64 progress_end_val[2] = {0, 0};
3055 :
3056 : Assert(vacrel->do_index_cleanup);
3057 : Assert(vacrel->nindexes > 0);
3058 :
3059 : /*
3060 : * Report that we are now cleaning up indexes and the number of indexes to
3061 : * cleanup.
3062 : */
3063 178248 : progress_start_val[0] = PROGRESS_VACUUM_PHASE_INDEX_CLEANUP;
3064 178248 : progress_start_val[1] = vacrel->nindexes;
3065 178248 : pgstat_progress_update_multi_param(2, progress_start_index, progress_start_val);
3066 :
3067 178248 : if (!ParallelVacuumIsActive(vacrel))
3068 : {
3069 458340 : for (int idx = 0; idx < vacrel->nindexes; idx++)
3070 : {
3071 280126 : Relation indrel = vacrel->indrels[idx];
3072 280126 : IndexBulkDeleteResult *istat = vacrel->indstats[idx];
3073 :
3074 560252 : vacrel->indstats[idx] =
3075 280126 : lazy_cleanup_one_index(indrel, istat, reltuples,
3076 : estimated_count, vacrel);
3077 :
3078 : /* Report the number of indexes cleaned up */
3079 280126 : pgstat_progress_update_param(PROGRESS_VACUUM_INDEXES_PROCESSED,
3080 280126 : idx + 1);
3081 : }
3082 : }
3083 : else
3084 : {
3085 : /* Outsource everything to parallel variant */
3086 34 : parallel_vacuum_cleanup_all_indexes(vacrel->pvs, reltuples,
3087 : vacrel->num_index_scans,
3088 : estimated_count);
3089 : }
3090 :
3091 : /* Reset the progress counters */
3092 178248 : pgstat_progress_update_multi_param(2, progress_end_index, progress_end_val);
3093 178248 : }
3094 :
3095 : /*
3096 : * lazy_vacuum_one_index() -- vacuum index relation.
3097 : *
3098 : * Delete all the index tuples containing a TID collected in
3099 : * vacrel->dead_items. Also update running statistics. Exact
3100 : * details depend on index AM's ambulkdelete routine.
3101 : *
3102 : * reltuples is the number of heap tuples to be passed to the
3103 : * bulkdelete callback. It's always assumed to be estimated.
3104 : * See indexam.sgml for more info.
3105 : *
3106 : * Returns bulk delete stats derived from input stats
3107 : */
3108 : static IndexBulkDeleteResult *
3109 2400 : lazy_vacuum_one_index(Relation indrel, IndexBulkDeleteResult *istat,
3110 : double reltuples, LVRelState *vacrel)
3111 : {
3112 : IndexVacuumInfo ivinfo;
3113 : LVSavedErrInfo saved_err_info;
3114 :
3115 2400 : ivinfo.index = indrel;
3116 2400 : ivinfo.heaprel = vacrel->rel;
3117 2400 : ivinfo.analyze_only = false;
3118 2400 : ivinfo.report_progress = false;
3119 2400 : ivinfo.estimated_count = true;
3120 2400 : ivinfo.message_level = DEBUG2;
3121 2400 : ivinfo.num_heap_tuples = reltuples;
3122 2400 : ivinfo.strategy = vacrel->bstrategy;
3123 :
3124 : /*
3125 : * Update error traceback information.
3126 : *
3127 : * The index name is saved during this phase and restored immediately
3128 : * after this phase. See vacuum_error_callback.
3129 : */
3130 : Assert(vacrel->indname == NULL);
3131 2400 : vacrel->indname = pstrdup(RelationGetRelationName(indrel));
3132 2400 : update_vacuum_error_info(vacrel, &saved_err_info,
3133 : VACUUM_ERRCB_PHASE_VACUUM_INDEX,
3134 : InvalidBlockNumber, InvalidOffsetNumber);
3135 :
3136 : /* Do bulk deletion */
3137 2400 : istat = vac_bulkdel_one_index(&ivinfo, istat, vacrel->dead_items,
3138 : vacrel->dead_items_info);
3139 :
3140 : /* Revert to the previous phase information for error traceback */
3141 2400 : restore_vacuum_error_info(vacrel, &saved_err_info);
3142 2400 : pfree(vacrel->indname);
3143 2400 : vacrel->indname = NULL;
3144 :
3145 2400 : return istat;
3146 : }
3147 :
3148 : /*
3149 : * lazy_cleanup_one_index() -- do post-vacuum cleanup for index relation.
3150 : *
3151 : * Calls index AM's amvacuumcleanup routine. reltuples is the number
3152 : * of heap tuples and estimated_count is true if reltuples is an
3153 : * estimated value. See indexam.sgml for more info.
3154 : *
3155 : * Returns bulk delete stats derived from input stats
3156 : */
3157 : static IndexBulkDeleteResult *
3158 280126 : lazy_cleanup_one_index(Relation indrel, IndexBulkDeleteResult *istat,
3159 : double reltuples, bool estimated_count,
3160 : LVRelState *vacrel)
3161 : {
3162 : IndexVacuumInfo ivinfo;
3163 : LVSavedErrInfo saved_err_info;
3164 :
3165 280126 : ivinfo.index = indrel;
3166 280126 : ivinfo.heaprel = vacrel->rel;
3167 280126 : ivinfo.analyze_only = false;
3168 280126 : ivinfo.report_progress = false;
3169 280126 : ivinfo.estimated_count = estimated_count;
3170 280126 : ivinfo.message_level = DEBUG2;
3171 :
3172 280126 : ivinfo.num_heap_tuples = reltuples;
3173 280126 : ivinfo.strategy = vacrel->bstrategy;
3174 :
3175 : /*
3176 : * Update error traceback information.
3177 : *
3178 : * The index name is saved during this phase and restored immediately
3179 : * after this phase. See vacuum_error_callback.
3180 : */
3181 : Assert(vacrel->indname == NULL);
3182 280126 : vacrel->indname = pstrdup(RelationGetRelationName(indrel));
3183 280126 : update_vacuum_error_info(vacrel, &saved_err_info,
3184 : VACUUM_ERRCB_PHASE_INDEX_CLEANUP,
3185 : InvalidBlockNumber, InvalidOffsetNumber);
3186 :
3187 280126 : istat = vac_cleanup_one_index(&ivinfo, istat);
3188 :
3189 : /* Revert to the previous phase information for error traceback */
3190 280126 : restore_vacuum_error_info(vacrel, &saved_err_info);
3191 280126 : pfree(vacrel->indname);
3192 280126 : vacrel->indname = NULL;
3193 :
3194 280126 : return istat;
3195 : }
3196 :
3197 : /*
3198 : * should_attempt_truncation - should we attempt to truncate the heap?
3199 : *
3200 : * Don't even think about it unless we have a shot at releasing a goodly
3201 : * number of pages. Otherwise, the time taken isn't worth it, mainly because
3202 : * an AccessExclusive lock must be replayed on any hot standby, where it can
3203 : * be particularly disruptive.
3204 : *
3205 : * Also don't attempt it if wraparound failsafe is in effect. The entire
3206 : * system might be refusing to allocate new XIDs at this point. The system
3207 : * definitely won't return to normal unless and until VACUUM actually advances
3208 : * the oldest relfrozenxid -- which hasn't happened for target rel just yet.
3209 : * If lazy_truncate_heap attempted to acquire an AccessExclusiveLock to
3210 : * truncate the table under these circumstances, an XID exhaustion error might
3211 : * make it impossible for VACUUM to fix the underlying XID exhaustion problem.
3212 : * There is very little chance of truncation working out when the failsafe is
3213 : * in effect in any case. lazy_scan_prune makes the optimistic assumption
3214 : * that any LP_DEAD items it encounters will always be LP_UNUSED by the time
3215 : * we're called.
3216 : */
3217 : static bool
3218 257572 : should_attempt_truncation(LVRelState *vacrel)
3219 : {
3220 : BlockNumber possibly_freeable;
3221 :
3222 257572 : if (!vacrel->do_rel_truncate || VacuumFailsafeActive)
3223 70228 : return false;
3224 :
3225 187344 : possibly_freeable = vacrel->rel_pages - vacrel->nonempty_pages;
3226 187344 : if (possibly_freeable > 0 &&
3227 318 : (possibly_freeable >= REL_TRUNCATE_MINIMUM ||
3228 318 : possibly_freeable >= vacrel->rel_pages / REL_TRUNCATE_FRACTION))
3229 300 : return true;
3230 :
3231 187044 : return false;
3232 : }
3233 :
3234 : /*
3235 : * lazy_truncate_heap - try to truncate off any empty pages at the end
3236 : */
3237 : static void
3238 300 : lazy_truncate_heap(LVRelState *vacrel)
3239 : {
3240 300 : BlockNumber orig_rel_pages = vacrel->rel_pages;
3241 : BlockNumber new_rel_pages;
3242 : bool lock_waiter_detected;
3243 : int lock_retry;
3244 :
3245 : /* Report that we are now truncating */
3246 300 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
3247 : PROGRESS_VACUUM_PHASE_TRUNCATE);
3248 :
3249 : /* Update error traceback information one last time */
3250 300 : update_vacuum_error_info(vacrel, NULL, VACUUM_ERRCB_PHASE_TRUNCATE,
3251 : vacrel->nonempty_pages, InvalidOffsetNumber);
3252 :
3253 : /*
3254 : * Loop until no more truncating can be done.
3255 : */
3256 : do
3257 : {
3258 : /*
3259 : * We need full exclusive lock on the relation in order to do
3260 : * truncation. If we can't get it, give up rather than waiting --- we
3261 : * don't want to block other backends, and we don't want to deadlock
3262 : * (which is quite possible considering we already hold a lower-grade
3263 : * lock).
3264 : */
3265 300 : lock_waiter_detected = false;
3266 300 : lock_retry = 0;
3267 : while (true)
3268 : {
3269 704 : if (ConditionalLockRelation(vacrel->rel, AccessExclusiveLock))
3270 296 : break;
3271 :
3272 : /*
3273 : * Check for interrupts while trying to (re-)acquire the exclusive
3274 : * lock.
3275 : */
3276 408 : CHECK_FOR_INTERRUPTS();
3277 :
3278 408 : if (++lock_retry > (VACUUM_TRUNCATE_LOCK_TIMEOUT /
3279 : VACUUM_TRUNCATE_LOCK_WAIT_INTERVAL))
3280 : {
3281 : /*
3282 : * We failed to establish the lock in the specified number of
3283 : * retries. This means we give up truncating.
3284 : */
3285 4 : ereport(vacrel->verbose ? INFO : DEBUG2,
3286 : (errmsg("\"%s\": stopping truncate due to conflicting lock request",
3287 : vacrel->relname)));
3288 6 : return;
3289 : }
3290 :
3291 404 : (void) WaitLatch(MyLatch,
3292 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
3293 : VACUUM_TRUNCATE_LOCK_WAIT_INTERVAL,
3294 : WAIT_EVENT_VACUUM_TRUNCATE);
3295 404 : ResetLatch(MyLatch);
3296 : }
3297 :
3298 : /*
3299 : * Now that we have exclusive lock, look to see if the rel has grown
3300 : * whilst we were vacuuming with non-exclusive lock. If so, give up;
3301 : * the newly added pages presumably contain non-deletable tuples.
3302 : */
3303 296 : new_rel_pages = RelationGetNumberOfBlocks(vacrel->rel);
3304 296 : if (new_rel_pages != orig_rel_pages)
3305 : {
3306 : /*
3307 : * Note: we intentionally don't update vacrel->rel_pages with the
3308 : * new rel size here. If we did, it would amount to assuming that
3309 : * the new pages are empty, which is unlikely. Leaving the numbers
3310 : * alone amounts to assuming that the new pages have the same
3311 : * tuple density as existing ones, which is less unlikely.
3312 : */
3313 0 : UnlockRelation(vacrel->rel, AccessExclusiveLock);
3314 0 : return;
3315 : }
3316 :
3317 : /*
3318 : * Scan backwards from the end to verify that the end pages actually
3319 : * contain no tuples. This is *necessary*, not optional, because
3320 : * other backends could have added tuples to these pages whilst we
3321 : * were vacuuming.
3322 : */
3323 296 : new_rel_pages = count_nondeletable_pages(vacrel, &lock_waiter_detected);
3324 296 : vacrel->blkno = new_rel_pages;
3325 :
3326 296 : if (new_rel_pages >= orig_rel_pages)
3327 : {
3328 : /* can't do anything after all */
3329 2 : UnlockRelation(vacrel->rel, AccessExclusiveLock);
3330 2 : return;
3331 : }
3332 :
3333 : /*
3334 : * Okay to truncate.
3335 : */
3336 294 : RelationTruncate(vacrel->rel, new_rel_pages);
3337 :
3338 : /*
3339 : * We can release the exclusive lock as soon as we have truncated.
3340 : * Other backends can't safely access the relation until they have
3341 : * processed the smgr invalidation that smgrtruncate sent out ... but
3342 : * that should happen as part of standard invalidation processing once
3343 : * they acquire lock on the relation.
3344 : */
3345 294 : UnlockRelation(vacrel->rel, AccessExclusiveLock);
3346 :
3347 : /*
3348 : * Update statistics. Here, it *is* correct to adjust rel_pages
3349 : * without also touching reltuples, since the tuple count wasn't
3350 : * changed by the truncation.
3351 : */
3352 294 : vacrel->removed_pages += orig_rel_pages - new_rel_pages;
3353 294 : vacrel->rel_pages = new_rel_pages;
3354 :
3355 294 : ereport(vacrel->verbose ? INFO : DEBUG2,
3356 : (errmsg("table \"%s\": truncated %u to %u pages",
3357 : vacrel->relname,
3358 : orig_rel_pages, new_rel_pages)));
3359 294 : orig_rel_pages = new_rel_pages;
3360 294 : } while (new_rel_pages > vacrel->nonempty_pages && lock_waiter_detected);
3361 : }
3362 :
3363 : /*
3364 : * Rescan end pages to verify that they are (still) empty of tuples.
3365 : *
3366 : * Returns number of nondeletable pages (last nonempty page + 1).
3367 : */
3368 : static BlockNumber
3369 296 : count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected)
3370 : {
3371 : BlockNumber blkno;
3372 : BlockNumber prefetchedUntil;
3373 : instr_time starttime;
3374 :
3375 : /* Initialize the starttime if we check for conflicting lock requests */
3376 296 : INSTR_TIME_SET_CURRENT(starttime);
3377 :
3378 : /*
3379 : * Start checking blocks at what we believe relation end to be and move
3380 : * backwards. (Strange coding of loop control is needed because blkno is
3381 : * unsigned.) To make the scan faster, we prefetch a few blocks at a time
3382 : * in forward direction, so that OS-level readahead can kick in.
3383 : */
3384 296 : blkno = vacrel->rel_pages;
3385 : StaticAssertStmt((PREFETCH_SIZE & (PREFETCH_SIZE - 1)) == 0,
3386 : "prefetch size must be power of 2");
3387 296 : prefetchedUntil = InvalidBlockNumber;
3388 4450 : while (blkno > vacrel->nonempty_pages)
3389 : {
3390 : Buffer buf;
3391 : Page page;
3392 : OffsetNumber offnum,
3393 : maxoff;
3394 : bool hastup;
3395 :
3396 : /*
3397 : * Check if another process requests a lock on our relation. We are
3398 : * holding an AccessExclusiveLock here, so they will be waiting. We
3399 : * only do this once per VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL, and we
3400 : * only check if that interval has elapsed once every 32 blocks to
3401 : * keep the number of system calls and actual shared lock table
3402 : * lookups to a minimum.
3403 : */
3404 4162 : if ((blkno % 32) == 0)
3405 : {
3406 : instr_time currenttime;
3407 : instr_time elapsed;
3408 :
3409 132 : INSTR_TIME_SET_CURRENT(currenttime);
3410 132 : elapsed = currenttime;
3411 132 : INSTR_TIME_SUBTRACT(elapsed, starttime);
3412 132 : if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000)
3413 : >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL)
3414 : {
3415 0 : if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock))
3416 : {
3417 0 : ereport(vacrel->verbose ? INFO : DEBUG2,
3418 : (errmsg("table \"%s\": suspending truncate due to conflicting lock request",
3419 : vacrel->relname)));
3420 :
3421 0 : *lock_waiter_detected = true;
3422 0 : return blkno;
3423 : }
3424 0 : starttime = currenttime;
3425 : }
3426 : }
3427 :
3428 : /*
3429 : * We don't insert a vacuum delay point here, because we have an
3430 : * exclusive lock on the table which we want to hold for as short a
3431 : * time as possible. We still need to check for interrupts however.
3432 : */
3433 4162 : CHECK_FOR_INTERRUPTS();
3434 :
3435 4162 : blkno--;
3436 :
3437 : /* If we haven't prefetched this lot yet, do so now. */
3438 4162 : if (prefetchedUntil > blkno)
3439 : {
3440 : BlockNumber prefetchStart;
3441 : BlockNumber pblkno;
3442 :
3443 390 : prefetchStart = blkno & ~(PREFETCH_SIZE - 1);
3444 6216 : for (pblkno = prefetchStart; pblkno <= blkno; pblkno++)
3445 : {
3446 5826 : PrefetchBuffer(vacrel->rel, MAIN_FORKNUM, pblkno);
3447 5826 : CHECK_FOR_INTERRUPTS();
3448 : }
3449 390 : prefetchedUntil = prefetchStart;
3450 : }
3451 :
3452 4162 : buf = ReadBufferExtended(vacrel->rel, MAIN_FORKNUM, blkno, RBM_NORMAL,
3453 : vacrel->bstrategy);
3454 :
3455 : /* In this phase we only need shared access to the buffer */
3456 4162 : LockBuffer(buf, BUFFER_LOCK_SHARE);
3457 :
3458 4162 : page = BufferGetPage(buf);
3459 :
3460 4162 : if (PageIsNew(page) || PageIsEmpty(page))
3461 : {
3462 1646 : UnlockReleaseBuffer(buf);
3463 1646 : continue;
3464 : }
3465 :
3466 2516 : hastup = false;
3467 2516 : maxoff = PageGetMaxOffsetNumber(page);
3468 2516 : for (offnum = FirstOffsetNumber;
3469 5474 : offnum <= maxoff;
3470 2958 : offnum = OffsetNumberNext(offnum))
3471 : {
3472 : ItemId itemid;
3473 :
3474 2966 : itemid = PageGetItemId(page, offnum);
3475 :
3476 : /*
3477 : * Note: any non-unused item should be taken as a reason to keep
3478 : * this page. Even an LP_DEAD item makes truncation unsafe, since
3479 : * we must not have cleaned out its index entries.
3480 : */
3481 2966 : if (ItemIdIsUsed(itemid))
3482 : {
3483 8 : hastup = true;
3484 8 : break; /* can stop scanning */
3485 : }
3486 : } /* scan along page */
3487 :
3488 2516 : UnlockReleaseBuffer(buf);
3489 :
3490 : /* Done scanning if we found a tuple here */
3491 2516 : if (hastup)
3492 8 : return blkno + 1;
3493 : }
3494 :
3495 : /*
3496 : * If we fall out of the loop, all the previously-thought-to-be-empty
3497 : * pages still are; we need not bother to look at the last known-nonempty
3498 : * page.
3499 : */
3500 288 : return vacrel->nonempty_pages;
3501 : }
3502 :
3503 : /*
3504 : * Allocate dead_items and dead_items_info (either using palloc, or in dynamic
3505 : * shared memory). Sets both in vacrel for caller.
3506 : *
3507 : * Also handles parallel initialization as part of allocating dead_items in
3508 : * DSM when required.
3509 : */
3510 : static void
3511 257572 : dead_items_alloc(LVRelState *vacrel, int nworkers)
3512 : {
3513 : VacDeadItemsInfo *dead_items_info;
3514 746722 : int vac_work_mem = AmAutoVacuumWorkerProcess() &&
3515 231578 : autovacuum_work_mem != -1 ?
3516 489150 : autovacuum_work_mem : maintenance_work_mem;
3517 :
3518 : /*
3519 : * Initialize state for a parallel vacuum. As of now, only one worker can
3520 : * be used for an index, so we invoke parallelism only if there are at
3521 : * least two indexes on a table.
3522 : */
3523 257572 : if (nworkers >= 0 && vacrel->nindexes > 1 && vacrel->do_index_vacuuming)
3524 : {
3525 : /*
3526 : * Since parallel workers cannot access data in temporary tables, we
3527 : * can't perform parallel vacuum on them.
3528 : */
3529 10296 : if (RelationUsesLocalBuffers(vacrel->rel))
3530 : {
3531 : /*
3532 : * Give warning only if the user explicitly tries to perform a
3533 : * parallel vacuum on the temporary table.
3534 : */
3535 6 : if (nworkers > 0)
3536 6 : ereport(WARNING,
3537 : (errmsg("disabling parallel option of vacuum on \"%s\" --- cannot vacuum temporary tables in parallel",
3538 : vacrel->relname)));
3539 : }
3540 : else
3541 10290 : vacrel->pvs = parallel_vacuum_init(vacrel->rel, vacrel->indrels,
3542 : vacrel->nindexes, nworkers,
3543 : vac_work_mem,
3544 10290 : vacrel->verbose ? INFO : DEBUG2,
3545 : vacrel->bstrategy);
3546 :
3547 : /*
3548 : * If parallel mode started, dead_items and dead_items_info spaces are
3549 : * allocated in DSM.
3550 : */
3551 10296 : if (ParallelVacuumIsActive(vacrel))
3552 : {
3553 34 : vacrel->dead_items = parallel_vacuum_get_dead_items(vacrel->pvs,
3554 : &vacrel->dead_items_info);
3555 34 : return;
3556 : }
3557 : }
3558 :
3559 : /*
3560 : * Serial VACUUM case. Allocate both dead_items and dead_items_info
3561 : * locally.
3562 : */
3563 :
3564 257538 : dead_items_info = (VacDeadItemsInfo *) palloc(sizeof(VacDeadItemsInfo));
3565 257538 : dead_items_info->max_bytes = vac_work_mem * (Size) 1024;
3566 257538 : dead_items_info->num_items = 0;
3567 257538 : vacrel->dead_items_info = dead_items_info;
3568 :
3569 257538 : vacrel->dead_items = TidStoreCreateLocal(dead_items_info->max_bytes, true);
3570 : }
3571 :
3572 : /*
3573 : * Add the given block number and offset numbers to dead_items.
3574 : */
3575 : static void
3576 30394 : dead_items_add(LVRelState *vacrel, BlockNumber blkno, OffsetNumber *offsets,
3577 : int num_offsets)
3578 : {
3579 30394 : const int prog_index[2] = {
3580 : PROGRESS_VACUUM_NUM_DEAD_ITEM_IDS,
3581 : PROGRESS_VACUUM_DEAD_TUPLE_BYTES
3582 : };
3583 : int64 prog_val[2];
3584 :
3585 30394 : TidStoreSetBlockOffsets(vacrel->dead_items, blkno, offsets, num_offsets);
3586 30394 : vacrel->dead_items_info->num_items += num_offsets;
3587 :
3588 : /* update the progress information */
3589 30394 : prog_val[0] = vacrel->dead_items_info->num_items;
3590 30394 : prog_val[1] = TidStoreMemoryUsage(vacrel->dead_items);
3591 30394 : pgstat_progress_update_multi_param(2, prog_index, prog_val);
3592 30394 : }
3593 :
3594 : /*
3595 : * Forget all collected dead items.
3596 : */
3597 : static void
3598 1282 : dead_items_reset(LVRelState *vacrel)
3599 : {
3600 1282 : if (ParallelVacuumIsActive(vacrel))
3601 : {
3602 24 : parallel_vacuum_reset_dead_items(vacrel->pvs);
3603 24 : vacrel->dead_items = parallel_vacuum_get_dead_items(vacrel->pvs,
3604 : &vacrel->dead_items_info);
3605 24 : return;
3606 : }
3607 :
3608 : /* Recreate the tidstore with the same max_bytes limitation */
3609 1258 : TidStoreDestroy(vacrel->dead_items);
3610 1258 : vacrel->dead_items = TidStoreCreateLocal(vacrel->dead_items_info->max_bytes, true);
3611 :
3612 : /* Reset the counter */
3613 1258 : vacrel->dead_items_info->num_items = 0;
3614 : }
3615 :
3616 : /*
3617 : * Perform cleanup for resources allocated in dead_items_alloc
3618 : */
3619 : static void
3620 257572 : dead_items_cleanup(LVRelState *vacrel)
3621 : {
3622 257572 : if (!ParallelVacuumIsActive(vacrel))
3623 : {
3624 : /* Don't bother with pfree here */
3625 257538 : return;
3626 : }
3627 :
3628 : /* End parallel mode */
3629 34 : parallel_vacuum_end(vacrel->pvs, vacrel->indstats);
3630 34 : vacrel->pvs = NULL;
3631 : }
3632 :
3633 : #ifdef USE_ASSERT_CHECKING
3634 :
3635 : /*
3636 : * Wrapper for heap_page_would_be_all_visible() which can be used for callers
3637 : * that expect no LP_DEAD on the page. Currently assert-only, but there is no
3638 : * reason not to use it outside of asserts.
3639 : */
3640 : static bool
3641 : heap_page_is_all_visible(Relation rel, Buffer buf,
3642 : TransactionId OldestXmin,
3643 : bool *all_frozen,
3644 : TransactionId *visibility_cutoff_xid,
3645 : OffsetNumber *logging_offnum)
3646 : {
3647 :
3648 : return heap_page_would_be_all_visible(rel, buf,
3649 : OldestXmin,
3650 : NULL, 0,
3651 : all_frozen,
3652 : visibility_cutoff_xid,
3653 : logging_offnum);
3654 : }
3655 : #endif
3656 :
3657 : /*
3658 : * Check whether the heap page in buf is all-visible except for the dead
3659 : * tuples referenced in the deadoffsets array.
3660 : *
3661 : * Vacuum uses this to check if a page would become all-visible after reaping
3662 : * known dead tuples. This function does not remove the dead items.
3663 : *
3664 : * This cannot be called in a critical section, as the visibility checks may
3665 : * perform IO and allocate memory.
3666 : *
3667 : * Returns true if the page is all-visible other than the provided
3668 : * deadoffsets and false otherwise.
3669 : *
3670 : * OldestXmin is used to determine visibility.
3671 : *
3672 : * Output parameters:
3673 : *
3674 : * - *all_frozen: true if every tuple on the page is frozen
3675 : * - *visibility_cutoff_xid: newest xmin; valid only if page is all-visible
3676 : * - *logging_offnum: OffsetNumber of current tuple being processed;
3677 : * used by vacuum's error callback system.
3678 : *
3679 : * Callers looking to verify that the page is already all-visible can call
3680 : * heap_page_is_all_visible().
3681 : *
3682 : * This logic is closely related to heap_prune_record_unchanged_lp_normal().
3683 : * If you modify this function, ensure consistency with that code. An
3684 : * assertion cross-checks that both remain in agreement. Do not introduce new
3685 : * side-effects.
3686 : */
3687 : static bool
3688 25796 : heap_page_would_be_all_visible(Relation rel, Buffer buf,
3689 : TransactionId OldestXmin,
3690 : OffsetNumber *deadoffsets,
3691 : int ndeadoffsets,
3692 : bool *all_frozen,
3693 : TransactionId *visibility_cutoff_xid,
3694 : OffsetNumber *logging_offnum)
3695 : {
3696 25796 : Page page = BufferGetPage(buf);
3697 25796 : BlockNumber blockno = BufferGetBlockNumber(buf);
3698 : OffsetNumber offnum,
3699 : maxoff;
3700 25796 : bool all_visible = true;
3701 25796 : int matched_dead_count = 0;
3702 :
3703 25796 : *visibility_cutoff_xid = InvalidTransactionId;
3704 25796 : *all_frozen = true;
3705 :
3706 : Assert(ndeadoffsets == 0 || deadoffsets);
3707 :
3708 : #ifdef USE_ASSERT_CHECKING
3709 : /* Confirm input deadoffsets[] is strictly sorted */
3710 : if (ndeadoffsets > 1)
3711 : {
3712 : for (int i = 1; i < ndeadoffsets; i++)
3713 : Assert(deadoffsets[i - 1] < deadoffsets[i]);
3714 : }
3715 : #endif
3716 :
3717 25796 : maxoff = PageGetMaxOffsetNumber(page);
3718 25796 : for (offnum = FirstOffsetNumber;
3719 2556668 : offnum <= maxoff && all_visible;
3720 2530872 : offnum = OffsetNumberNext(offnum))
3721 : {
3722 : ItemId itemid;
3723 : HeapTupleData tuple;
3724 :
3725 : /*
3726 : * Set the offset number so that we can display it along with any
3727 : * error that occurred while processing this tuple.
3728 : */
3729 2530874 : *logging_offnum = offnum;
3730 2530874 : itemid = PageGetItemId(page, offnum);
3731 :
3732 : /* Unused or redirect line pointers are of no interest */
3733 2530874 : if (!ItemIdIsUsed(itemid) || ItemIdIsRedirected(itemid))
3734 1647978 : continue;
3735 :
3736 2466508 : ItemPointerSet(&(tuple.t_self), blockno, offnum);
3737 :
3738 : /*
3739 : * Dead line pointers can have index pointers pointing to them. So
3740 : * they can't be treated as visible
3741 : */
3742 2466508 : if (ItemIdIsDead(itemid))
3743 : {
3744 1583614 : if (!deadoffsets ||
3745 1583612 : matched_dead_count >= ndeadoffsets ||
3746 1583612 : deadoffsets[matched_dead_count] != offnum)
3747 : {
3748 2 : *all_frozen = all_visible = false;
3749 2 : break;
3750 : }
3751 1583612 : matched_dead_count++;
3752 1583612 : continue;
3753 : }
3754 :
3755 : Assert(ItemIdIsNormal(itemid));
3756 :
3757 882894 : tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
3758 882894 : tuple.t_len = ItemIdGetLength(itemid);
3759 882894 : tuple.t_tableOid = RelationGetRelid(rel);
3760 :
3761 : /* Visibility checks may do IO or allocate memory */
3762 : Assert(CritSectionCount == 0);
3763 882894 : switch (HeapTupleSatisfiesVacuum(&tuple, OldestXmin, buf))
3764 : {
3765 882672 : case HEAPTUPLE_LIVE:
3766 : {
3767 : TransactionId xmin;
3768 :
3769 : /* Check comments in lazy_scan_prune. */
3770 882672 : if (!HeapTupleHeaderXminCommitted(tuple.t_data))
3771 : {
3772 0 : all_visible = false;
3773 0 : *all_frozen = false;
3774 0 : break;
3775 : }
3776 :
3777 : /*
3778 : * The inserter definitely committed. But is it old enough
3779 : * that everyone sees it as committed?
3780 : */
3781 882672 : xmin = HeapTupleHeaderGetXmin(tuple.t_data);
3782 882672 : if (!TransactionIdPrecedes(xmin, OldestXmin))
3783 : {
3784 92 : all_visible = false;
3785 92 : *all_frozen = false;
3786 92 : break;
3787 : }
3788 :
3789 : /* Track newest xmin on page. */
3790 882580 : if (TransactionIdFollows(xmin, *visibility_cutoff_xid) &&
3791 : TransactionIdIsNormal(xmin))
3792 18120 : *visibility_cutoff_xid = xmin;
3793 :
3794 : /* Check whether this tuple is already frozen or not */
3795 1138602 : if (all_visible && *all_frozen &&
3796 256022 : heap_tuple_needs_eventual_freeze(tuple.t_data))
3797 5216 : *all_frozen = false;
3798 : }
3799 882580 : break;
3800 :
3801 222 : case HEAPTUPLE_DEAD:
3802 : case HEAPTUPLE_RECENTLY_DEAD:
3803 : case HEAPTUPLE_INSERT_IN_PROGRESS:
3804 : case HEAPTUPLE_DELETE_IN_PROGRESS:
3805 : {
3806 222 : all_visible = false;
3807 222 : *all_frozen = false;
3808 222 : break;
3809 : }
3810 0 : default:
3811 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
3812 : break;
3813 : }
3814 : } /* scan along page */
3815 :
3816 : /* Clear the offset information once we have processed the given page. */
3817 25796 : *logging_offnum = InvalidOffsetNumber;
3818 :
3819 25796 : return all_visible;
3820 : }
3821 :
3822 : /*
3823 : * Update index statistics in pg_class if the statistics are accurate.
3824 : */
3825 : static void
3826 187374 : update_relstats_all_indexes(LVRelState *vacrel)
3827 : {
3828 187374 : Relation *indrels = vacrel->indrels;
3829 187374 : int nindexes = vacrel->nindexes;
3830 187374 : IndexBulkDeleteResult **indstats = vacrel->indstats;
3831 :
3832 : Assert(vacrel->do_index_cleanup);
3833 :
3834 467622 : for (int idx = 0; idx < nindexes; idx++)
3835 : {
3836 280248 : Relation indrel = indrels[idx];
3837 280248 : IndexBulkDeleteResult *istat = indstats[idx];
3838 :
3839 280248 : if (istat == NULL || istat->estimated_count)
3840 277566 : continue;
3841 :
3842 : /* Update index statistics */
3843 2682 : vac_update_relstats(indrel,
3844 : istat->num_pages,
3845 : istat->num_index_tuples,
3846 : 0, 0,
3847 : false,
3848 : InvalidTransactionId,
3849 : InvalidMultiXactId,
3850 : NULL, NULL, false);
3851 : }
3852 187374 : }
3853 :
3854 : /*
3855 : * Error context callback for errors occurring during vacuum. The error
3856 : * context messages for index phases should match the messages set in parallel
3857 : * vacuum. If you change this function for those phases, change
3858 : * parallel_vacuum_error_callback() as well.
3859 : */
3860 : static void
3861 267012 : vacuum_error_callback(void *arg)
3862 : {
3863 267012 : LVRelState *errinfo = arg;
3864 :
3865 267012 : switch (errinfo->phase)
3866 : {
3867 0 : case VACUUM_ERRCB_PHASE_SCAN_HEAP:
3868 0 : if (BlockNumberIsValid(errinfo->blkno))
3869 : {
3870 0 : if (OffsetNumberIsValid(errinfo->offnum))
3871 0 : errcontext("while scanning block %u offset %u of relation \"%s.%s\"",
3872 0 : errinfo->blkno, errinfo->offnum, errinfo->relnamespace, errinfo->relname);
3873 : else
3874 0 : errcontext("while scanning block %u of relation \"%s.%s\"",
3875 : errinfo->blkno, errinfo->relnamespace, errinfo->relname);
3876 : }
3877 : else
3878 0 : errcontext("while scanning relation \"%s.%s\"",
3879 : errinfo->relnamespace, errinfo->relname);
3880 0 : break;
3881 :
3882 0 : case VACUUM_ERRCB_PHASE_VACUUM_HEAP:
3883 0 : if (BlockNumberIsValid(errinfo->blkno))
3884 : {
3885 0 : if (OffsetNumberIsValid(errinfo->offnum))
3886 0 : errcontext("while vacuuming block %u offset %u of relation \"%s.%s\"",
3887 0 : errinfo->blkno, errinfo->offnum, errinfo->relnamespace, errinfo->relname);
3888 : else
3889 0 : errcontext("while vacuuming block %u of relation \"%s.%s\"",
3890 : errinfo->blkno, errinfo->relnamespace, errinfo->relname);
3891 : }
3892 : else
3893 0 : errcontext("while vacuuming relation \"%s.%s\"",
3894 : errinfo->relnamespace, errinfo->relname);
3895 0 : break;
3896 :
3897 0 : case VACUUM_ERRCB_PHASE_VACUUM_INDEX:
3898 0 : errcontext("while vacuuming index \"%s\" of relation \"%s.%s\"",
3899 : errinfo->indname, errinfo->relnamespace, errinfo->relname);
3900 0 : break;
3901 :
3902 0 : case VACUUM_ERRCB_PHASE_INDEX_CLEANUP:
3903 0 : errcontext("while cleaning up index \"%s\" of relation \"%s.%s\"",
3904 : errinfo->indname, errinfo->relnamespace, errinfo->relname);
3905 0 : break;
3906 :
3907 6 : case VACUUM_ERRCB_PHASE_TRUNCATE:
3908 6 : if (BlockNumberIsValid(errinfo->blkno))
3909 6 : errcontext("while truncating relation \"%s.%s\" to %u blocks",
3910 : errinfo->relnamespace, errinfo->relname, errinfo->blkno);
3911 6 : break;
3912 :
3913 267006 : case VACUUM_ERRCB_PHASE_UNKNOWN:
3914 : default:
3915 267006 : return; /* do nothing; the errinfo may not be
3916 : * initialized */
3917 : }
3918 : }
3919 :
3920 : /*
3921 : * Updates the information required for vacuum error callback. This also saves
3922 : * the current information which can be later restored via restore_vacuum_error_info.
3923 : */
3924 : static void
3925 1470424 : update_vacuum_error_info(LVRelState *vacrel, LVSavedErrInfo *saved_vacrel,
3926 : int phase, BlockNumber blkno, OffsetNumber offnum)
3927 : {
3928 1470424 : if (saved_vacrel)
3929 : {
3930 309584 : saved_vacrel->offnum = vacrel->offnum;
3931 309584 : saved_vacrel->blkno = vacrel->blkno;
3932 309584 : saved_vacrel->phase = vacrel->phase;
3933 : }
3934 :
3935 1470424 : vacrel->blkno = blkno;
3936 1470424 : vacrel->offnum = offnum;
3937 1470424 : vacrel->phase = phase;
3938 1470424 : }
3939 :
3940 : /*
3941 : * Restores the vacuum information saved via a prior call to update_vacuum_error_info.
3942 : */
3943 : static void
3944 309584 : restore_vacuum_error_info(LVRelState *vacrel,
3945 : const LVSavedErrInfo *saved_vacrel)
3946 : {
3947 309584 : vacrel->blkno = saved_vacrel->blkno;
3948 309584 : vacrel->offnum = saved_vacrel->offnum;
3949 309584 : vacrel->phase = saved_vacrel->phase;
3950 309584 : }
|
