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 267340 : 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 267340 : 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 267340 : vacrel->next_eager_scan_region_start = InvalidBlockNumber;
514 267340 : vacrel->eager_scan_max_fails_per_region = 0;
515 267340 : vacrel->eager_scan_remaining_fails = 0;
516 267340 : vacrel->eager_scan_remaining_successes = 0;
517 :
518 : /* If eager scanning is explicitly disabled, just return. */
519 267340 : if (params.max_eager_freeze_failure_rate == 0)
520 267340 : 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 267340 : if (vacrel->aggressive)
530 256116 : 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 11224 : if (vacrel->rel_pages < 2 * EAGER_SCAN_REGION_SIZE)
540 11224 : 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 267340 : 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 267340 : TimestampTz starttime = 0;
642 267340 : PgStat_Counter startreadtime = 0,
643 267340 : startwritetime = 0;
644 267340 : WalUsage startwalusage = pgWalUsage;
645 267340 : BufferUsage startbufferusage = pgBufferUsage;
646 : ErrorContextCallback errcallback;
647 267340 : char **indnames = NULL;
648 :
649 267340 : verbose = (params.options & VACOPT_VERBOSE) != 0;
650 507192 : instrument = (verbose || (AmAutoVacuumWorkerProcess() &&
651 239852 : params.log_vacuum_min_duration >= 0));
652 267340 : if (instrument)
653 : {
654 239876 : pg_rusage_init(&ru0);
655 239876 : if (track_io_timing)
656 : {
657 0 : startreadtime = pgStatBlockReadTime;
658 0 : startwritetime = pgStatBlockWriteTime;
659 : }
660 : }
661 :
662 : /* Used for instrumentation and stats report */
663 267340 : starttime = GetCurrentTimestamp();
664 :
665 267340 : pgstat_progress_start_command(PROGRESS_COMMAND_VACUUM,
666 : RelationGetRelid(rel));
667 267340 : if (AmAutoVacuumWorkerProcess())
668 239852 : pgstat_progress_update_param(PROGRESS_VACUUM_STARTED_BY,
669 239852 : params.is_wraparound
670 : ? PROGRESS_VACUUM_STARTED_BY_AUTOVACUUM_WRAPAROUND
671 : : PROGRESS_VACUUM_STARTED_BY_AUTOVACUUM);
672 : else
673 27488 : pgstat_progress_update_param(PROGRESS_VACUUM_STARTED_BY,
674 : PROGRESS_VACUUM_STARTED_BY_MANUAL);
675 :
676 : /*
677 : * Setup error traceback support for ereport() first. The idea is to set
678 : * up an error context callback to display additional information on any
679 : * error during a vacuum. During different phases of vacuum, we update
680 : * the state so that the error context callback always display current
681 : * information.
682 : *
683 : * Copy the names of heap rel into local memory for error reporting
684 : * purposes, too. It isn't always safe to assume that we can get the name
685 : * of each rel. It's convenient for code in lazy_scan_heap to always use
686 : * these temp copies.
687 : */
688 267340 : vacrel = palloc0_object(LVRelState);
689 267340 : vacrel->dbname = get_database_name(MyDatabaseId);
690 267340 : vacrel->relnamespace = get_namespace_name(RelationGetNamespace(rel));
691 267340 : vacrel->relname = pstrdup(RelationGetRelationName(rel));
692 267340 : vacrel->indname = NULL;
693 267340 : vacrel->phase = VACUUM_ERRCB_PHASE_UNKNOWN;
694 267340 : vacrel->verbose = verbose;
695 267340 : errcallback.callback = vacuum_error_callback;
696 267340 : errcallback.arg = vacrel;
697 267340 : errcallback.previous = error_context_stack;
698 267340 : error_context_stack = &errcallback;
699 :
700 : /* Set up high level stuff about rel and its indexes */
701 267340 : vacrel->rel = rel;
702 267340 : vac_open_indexes(vacrel->rel, RowExclusiveLock, &vacrel->nindexes,
703 : &vacrel->indrels);
704 267340 : vacrel->bstrategy = bstrategy;
705 267340 : if (instrument && vacrel->nindexes > 0)
706 : {
707 : /* Copy index names used by instrumentation (not error reporting) */
708 229466 : indnames = palloc_array(char *, vacrel->nindexes);
709 590384 : for (int i = 0; i < vacrel->nindexes; i++)
710 360918 : indnames[i] = pstrdup(RelationGetRelationName(vacrel->indrels[i]));
711 : }
712 :
713 : /*
714 : * The index_cleanup param either disables index vacuuming and cleanup or
715 : * forces it to go ahead when we would otherwise apply the index bypass
716 : * optimization. The default is 'auto', which leaves the final decision
717 : * up to lazy_vacuum().
718 : *
719 : * The truncate param allows user to avoid attempting relation truncation,
720 : * though it can't force truncation to happen.
721 : */
722 : Assert(params.index_cleanup != VACOPTVALUE_UNSPECIFIED);
723 : Assert(params.truncate != VACOPTVALUE_UNSPECIFIED &&
724 : params.truncate != VACOPTVALUE_AUTO);
725 :
726 : /*
727 : * While VacuumFailSafeActive is reset to false before calling this, we
728 : * still need to reset it here due to recursive calls.
729 : */
730 267340 : VacuumFailsafeActive = false;
731 267340 : vacrel->consider_bypass_optimization = true;
732 267340 : vacrel->do_index_vacuuming = true;
733 267340 : vacrel->do_index_cleanup = true;
734 267340 : vacrel->do_rel_truncate = (params.truncate != VACOPTVALUE_DISABLED);
735 267340 : if (params.index_cleanup == VACOPTVALUE_DISABLED)
736 : {
737 : /* Force disable index vacuuming up-front */
738 260 : vacrel->do_index_vacuuming = false;
739 260 : vacrel->do_index_cleanup = false;
740 : }
741 267080 : else if (params.index_cleanup == VACOPTVALUE_ENABLED)
742 : {
743 : /* Force index vacuuming. Note that failsafe can still bypass. */
744 30 : vacrel->consider_bypass_optimization = false;
745 : }
746 : else
747 : {
748 : /* Default/auto, make all decisions dynamically */
749 : Assert(params.index_cleanup == VACOPTVALUE_AUTO);
750 : }
751 :
752 : /* Initialize page counters explicitly (be tidy) */
753 267340 : vacrel->scanned_pages = 0;
754 267340 : vacrel->eager_scanned_pages = 0;
755 267340 : vacrel->removed_pages = 0;
756 267340 : vacrel->new_frozen_tuple_pages = 0;
757 267340 : vacrel->lpdead_item_pages = 0;
758 267340 : vacrel->missed_dead_pages = 0;
759 267340 : vacrel->nonempty_pages = 0;
760 : /* dead_items_alloc allocates vacrel->dead_items later on */
761 :
762 : /* Allocate/initialize output statistics state */
763 267340 : vacrel->new_rel_tuples = 0;
764 267340 : vacrel->new_live_tuples = 0;
765 267340 : vacrel->indstats = (IndexBulkDeleteResult **)
766 267340 : palloc0(vacrel->nindexes * sizeof(IndexBulkDeleteResult *));
767 :
768 : /* Initialize remaining counters (be tidy) */
769 267340 : vacrel->num_index_scans = 0;
770 267340 : vacrel->tuples_deleted = 0;
771 267340 : vacrel->tuples_frozen = 0;
772 267340 : vacrel->lpdead_items = 0;
773 267340 : vacrel->live_tuples = 0;
774 267340 : vacrel->recently_dead_tuples = 0;
775 267340 : vacrel->missed_dead_tuples = 0;
776 :
777 267340 : vacrel->vm_new_visible_pages = 0;
778 267340 : vacrel->vm_new_visible_frozen_pages = 0;
779 267340 : vacrel->vm_new_frozen_pages = 0;
780 :
781 : /*
782 : * Get cutoffs that determine which deleted tuples are considered DEAD,
783 : * not just RECENTLY_DEAD, and which XIDs/MXIDs to freeze. Then determine
784 : * the extent of the blocks that we'll scan in lazy_scan_heap. It has to
785 : * happen in this order to ensure that the OldestXmin cutoff field works
786 : * as an upper bound on the XIDs stored in the pages we'll actually scan
787 : * (NewRelfrozenXid tracking must never be allowed to miss unfrozen XIDs).
788 : *
789 : * Next acquire vistest, a related cutoff that's used in pruning. We use
790 : * vistest in combination with OldestXmin to ensure that
791 : * heap_page_prune_and_freeze() always removes any deleted tuple whose
792 : * xmax is < OldestXmin. lazy_scan_prune must never become confused about
793 : * whether a tuple should be frozen or removed. (In the future we might
794 : * want to teach lazy_scan_prune to recompute vistest from time to time,
795 : * to increase the number of dead tuples it can prune away.)
796 : */
797 267340 : vacrel->aggressive = vacuum_get_cutoffs(rel, params, &vacrel->cutoffs);
798 267340 : vacrel->rel_pages = orig_rel_pages = RelationGetNumberOfBlocks(rel);
799 267340 : vacrel->vistest = GlobalVisTestFor(rel);
800 :
801 : /* Initialize state used to track oldest extant XID/MXID */
802 267340 : vacrel->NewRelfrozenXid = vacrel->cutoffs.OldestXmin;
803 267340 : vacrel->NewRelminMxid = vacrel->cutoffs.OldestMxact;
804 :
805 : /*
806 : * Initialize state related to tracking all-visible page skipping. This is
807 : * very important to determine whether or not it is safe to advance the
808 : * relfrozenxid/relminmxid.
809 : */
810 267340 : vacrel->skippedallvis = false;
811 267340 : skipwithvm = true;
812 267340 : if (params.options & VACOPT_DISABLE_PAGE_SKIPPING)
813 : {
814 : /*
815 : * Force aggressive mode, and disable skipping blocks using the
816 : * visibility map (even those set all-frozen)
817 : */
818 344 : vacrel->aggressive = true;
819 344 : skipwithvm = false;
820 : }
821 :
822 267340 : vacrel->skipwithvm = skipwithvm;
823 :
824 : /*
825 : * Set up eager scan tracking state. This must happen after determining
826 : * whether or not the vacuum must be aggressive, because only normal
827 : * vacuums use the eager scan algorithm.
828 : */
829 267340 : heap_vacuum_eager_scan_setup(vacrel, params);
830 :
831 : /* Report the vacuum mode: 'normal' or 'aggressive' */
832 267340 : pgstat_progress_update_param(PROGRESS_VACUUM_MODE,
833 267340 : vacrel->aggressive
834 : ? PROGRESS_VACUUM_MODE_AGGRESSIVE
835 : : PROGRESS_VACUUM_MODE_NORMAL);
836 :
837 267340 : if (verbose)
838 : {
839 24 : if (vacrel->aggressive)
840 2 : ereport(INFO,
841 : (errmsg("aggressively vacuuming \"%s.%s.%s\"",
842 : vacrel->dbname, vacrel->relnamespace,
843 : vacrel->relname)));
844 : else
845 22 : ereport(INFO,
846 : (errmsg("vacuuming \"%s.%s.%s\"",
847 : vacrel->dbname, vacrel->relnamespace,
848 : vacrel->relname)));
849 : }
850 :
851 : /*
852 : * Allocate dead_items memory using dead_items_alloc. This handles
853 : * parallel VACUUM initialization as part of allocating shared memory
854 : * space used for dead_items. (But do a failsafe precheck first, to
855 : * ensure that parallel VACUUM won't be attempted at all when relfrozenxid
856 : * is already dangerously old.)
857 : */
858 267340 : lazy_check_wraparound_failsafe(vacrel);
859 267340 : dead_items_alloc(vacrel, params.nworkers);
860 :
861 : /*
862 : * Call lazy_scan_heap to perform all required heap pruning, index
863 : * vacuuming, and heap vacuuming (plus related processing)
864 : */
865 267340 : lazy_scan_heap(vacrel);
866 :
867 : /*
868 : * Free resources managed by dead_items_alloc. This ends parallel mode in
869 : * passing when necessary.
870 : */
871 267340 : dead_items_cleanup(vacrel);
872 : Assert(!IsInParallelMode());
873 :
874 : /*
875 : * Update pg_class entries for each of rel's indexes where appropriate.
876 : *
877 : * Unlike the later update to rel's pg_class entry, this is not critical.
878 : * Maintains relpages/reltuples statistics used by the planner only.
879 : */
880 267340 : if (vacrel->do_index_cleanup)
881 192158 : update_relstats_all_indexes(vacrel);
882 :
883 : /* Done with rel's indexes */
884 267340 : vac_close_indexes(vacrel->nindexes, vacrel->indrels, NoLock);
885 :
886 : /* Optionally truncate rel */
887 267340 : if (should_attempt_truncation(vacrel))
888 316 : lazy_truncate_heap(vacrel);
889 :
890 : /* Pop the error context stack */
891 267340 : error_context_stack = errcallback.previous;
892 :
893 : /* Report that we are now doing final cleanup */
894 267340 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
895 : PROGRESS_VACUUM_PHASE_FINAL_CLEANUP);
896 :
897 : /*
898 : * Prepare to update rel's pg_class entry.
899 : *
900 : * Aggressive VACUUMs must always be able to advance relfrozenxid to a
901 : * value >= FreezeLimit, and relminmxid to a value >= MultiXactCutoff.
902 : * Non-aggressive VACUUMs may advance them by any amount, or not at all.
903 : */
904 : Assert(vacrel->NewRelfrozenXid == vacrel->cutoffs.OldestXmin ||
905 : TransactionIdPrecedesOrEquals(vacrel->aggressive ? vacrel->cutoffs.FreezeLimit :
906 : vacrel->cutoffs.relfrozenxid,
907 : vacrel->NewRelfrozenXid));
908 : Assert(vacrel->NewRelminMxid == vacrel->cutoffs.OldestMxact ||
909 : MultiXactIdPrecedesOrEquals(vacrel->aggressive ? vacrel->cutoffs.MultiXactCutoff :
910 : vacrel->cutoffs.relminmxid,
911 : vacrel->NewRelminMxid));
912 267340 : if (vacrel->skippedallvis)
913 : {
914 : /*
915 : * Must keep original relfrozenxid in a non-aggressive VACUUM that
916 : * chose to skip an all-visible page range. The state that tracks new
917 : * values will have missed unfrozen XIDs from the pages we skipped.
918 : */
919 : Assert(!vacrel->aggressive);
920 70 : vacrel->NewRelfrozenXid = InvalidTransactionId;
921 70 : vacrel->NewRelminMxid = InvalidMultiXactId;
922 : }
923 :
924 : /*
925 : * For safety, clamp relallvisible to be not more than what we're setting
926 : * pg_class.relpages to
927 : */
928 267340 : new_rel_pages = vacrel->rel_pages; /* After possible rel truncation */
929 267340 : visibilitymap_count(rel, &new_rel_allvisible, &new_rel_allfrozen);
930 267340 : if (new_rel_allvisible > new_rel_pages)
931 0 : new_rel_allvisible = new_rel_pages;
932 :
933 : /*
934 : * An all-frozen block _must_ be all-visible. As such, clamp the count of
935 : * all-frozen blocks to the count of all-visible blocks. This matches the
936 : * clamping of relallvisible above.
937 : */
938 267340 : if (new_rel_allfrozen > new_rel_allvisible)
939 0 : new_rel_allfrozen = new_rel_allvisible;
940 :
941 : /*
942 : * Now actually update rel's pg_class entry.
943 : *
944 : * In principle new_live_tuples could be -1 indicating that we (still)
945 : * don't know the tuple count. In practice that can't happen, since we
946 : * scan every page that isn't skipped using the visibility map.
947 : */
948 267340 : vac_update_relstats(rel, new_rel_pages, vacrel->new_live_tuples,
949 : new_rel_allvisible, new_rel_allfrozen,
950 267340 : vacrel->nindexes > 0,
951 : vacrel->NewRelfrozenXid, vacrel->NewRelminMxid,
952 : &frozenxid_updated, &minmulti_updated, false);
953 :
954 : /*
955 : * Report results to the cumulative stats system, too.
956 : *
957 : * Deliberately avoid telling the stats system about LP_DEAD items that
958 : * remain in the table due to VACUUM bypassing index and heap vacuuming.
959 : * ANALYZE will consider the remaining LP_DEAD items to be dead "tuples".
960 : * It seems like a good idea to err on the side of not vacuuming again too
961 : * soon in cases where the failsafe prevented significant amounts of heap
962 : * vacuuming.
963 : */
964 160402 : pgstat_report_vacuum(RelationGetRelid(rel),
965 267336 : rel->rd_rel->relisshared,
966 106934 : Max(vacrel->new_live_tuples, 0),
967 267336 : vacrel->recently_dead_tuples +
968 267336 : vacrel->missed_dead_tuples,
969 : starttime);
970 267336 : pgstat_progress_end_command();
971 :
972 267336 : if (instrument)
973 : {
974 239876 : TimestampTz endtime = GetCurrentTimestamp();
975 :
976 240154 : if (verbose || params.log_vacuum_min_duration == 0 ||
977 278 : TimestampDifferenceExceeds(starttime, endtime,
978 278 : params.log_vacuum_min_duration))
979 : {
980 : long secs_dur;
981 : int usecs_dur;
982 : WalUsage walusage;
983 : BufferUsage bufferusage;
984 : StringInfoData buf;
985 : char *msgfmt;
986 : int32 diff;
987 239598 : double read_rate = 0,
988 239598 : write_rate = 0;
989 : int64 total_blks_hit;
990 : int64 total_blks_read;
991 : int64 total_blks_dirtied;
992 :
993 239598 : TimestampDifference(starttime, endtime, &secs_dur, &usecs_dur);
994 239598 : memset(&walusage, 0, sizeof(WalUsage));
995 239598 : WalUsageAccumDiff(&walusage, &pgWalUsage, &startwalusage);
996 239598 : memset(&bufferusage, 0, sizeof(BufferUsage));
997 239598 : BufferUsageAccumDiff(&bufferusage, &pgBufferUsage, &startbufferusage);
998 :
999 239598 : total_blks_hit = bufferusage.shared_blks_hit +
1000 239598 : bufferusage.local_blks_hit;
1001 239598 : total_blks_read = bufferusage.shared_blks_read +
1002 239598 : bufferusage.local_blks_read;
1003 239598 : total_blks_dirtied = bufferusage.shared_blks_dirtied +
1004 239598 : bufferusage.local_blks_dirtied;
1005 :
1006 239598 : initStringInfo(&buf);
1007 239598 : if (verbose)
1008 : {
1009 : /*
1010 : * Aggressiveness already reported earlier, in dedicated
1011 : * VACUUM VERBOSE ereport
1012 : */
1013 : Assert(!params.is_wraparound);
1014 24 : msgfmt = _("finished vacuuming \"%s.%s.%s\": index scans: %d\n");
1015 : }
1016 239574 : else if (params.is_wraparound)
1017 : {
1018 : /*
1019 : * While it's possible for a VACUUM to be both is_wraparound
1020 : * and !aggressive, that's just a corner-case -- is_wraparound
1021 : * implies aggressive. Produce distinct output for the corner
1022 : * case all the same, just in case.
1023 : */
1024 239518 : if (vacrel->aggressive)
1025 239506 : msgfmt = _("automatic aggressive vacuum to prevent wraparound of table \"%s.%s.%s\": index scans: %d\n");
1026 : else
1027 12 : msgfmt = _("automatic vacuum to prevent wraparound of table \"%s.%s.%s\": index scans: %d\n");
1028 : }
1029 : else
1030 : {
1031 56 : if (vacrel->aggressive)
1032 40 : msgfmt = _("automatic aggressive vacuum of table \"%s.%s.%s\": index scans: %d\n");
1033 : else
1034 16 : msgfmt = _("automatic vacuum of table \"%s.%s.%s\": index scans: %d\n");
1035 : }
1036 239598 : appendStringInfo(&buf, msgfmt,
1037 : vacrel->dbname,
1038 : vacrel->relnamespace,
1039 : vacrel->relname,
1040 : vacrel->num_index_scans);
1041 334976 : appendStringInfo(&buf, _("pages: %u removed, %u remain, %u scanned (%.2f%% of total), %u eagerly scanned\n"),
1042 : vacrel->removed_pages,
1043 : new_rel_pages,
1044 : vacrel->scanned_pages,
1045 : orig_rel_pages == 0 ? 100.0 :
1046 95378 : 100.0 * vacrel->scanned_pages /
1047 : orig_rel_pages,
1048 : vacrel->eager_scanned_pages);
1049 239598 : appendStringInfo(&buf,
1050 239598 : _("tuples: %" PRId64 " removed, %" PRId64 " remain, %" PRId64 " are dead but not yet removable\n"),
1051 : vacrel->tuples_deleted,
1052 239598 : (int64) vacrel->new_rel_tuples,
1053 : vacrel->recently_dead_tuples);
1054 239598 : if (vacrel->missed_dead_tuples > 0)
1055 0 : appendStringInfo(&buf,
1056 0 : _("tuples missed: %" PRId64 " dead from %u pages not removed due to cleanup lock contention\n"),
1057 : vacrel->missed_dead_tuples,
1058 : vacrel->missed_dead_pages);
1059 239598 : diff = (int32) (ReadNextTransactionId() -
1060 239598 : vacrel->cutoffs.OldestXmin);
1061 239598 : appendStringInfo(&buf,
1062 239598 : _("removable cutoff: %u, which was %d XIDs old when operation ended\n"),
1063 : vacrel->cutoffs.OldestXmin, diff);
1064 239598 : if (frozenxid_updated)
1065 : {
1066 35102 : diff = (int32) (vacrel->NewRelfrozenXid -
1067 35102 : vacrel->cutoffs.relfrozenxid);
1068 35102 : appendStringInfo(&buf,
1069 35102 : _("new relfrozenxid: %u, which is %d XIDs ahead of previous value\n"),
1070 : vacrel->NewRelfrozenXid, diff);
1071 : }
1072 239598 : if (minmulti_updated)
1073 : {
1074 24 : diff = (int32) (vacrel->NewRelminMxid -
1075 24 : vacrel->cutoffs.relminmxid);
1076 24 : appendStringInfo(&buf,
1077 24 : _("new relminmxid: %u, which is %d MXIDs ahead of previous value\n"),
1078 : vacrel->NewRelminMxid, diff);
1079 : }
1080 334976 : appendStringInfo(&buf, _("frozen: %u pages from table (%.2f%% of total) had %" PRId64 " tuples frozen\n"),
1081 : vacrel->new_frozen_tuple_pages,
1082 : orig_rel_pages == 0 ? 100.0 :
1083 95378 : 100.0 * vacrel->new_frozen_tuple_pages /
1084 : orig_rel_pages,
1085 : vacrel->tuples_frozen);
1086 :
1087 239598 : appendStringInfo(&buf,
1088 239598 : _("visibility map: %u pages set all-visible, %u pages set all-frozen (%u were all-visible)\n"),
1089 : vacrel->vm_new_visible_pages,
1090 239598 : vacrel->vm_new_visible_frozen_pages +
1091 239598 : vacrel->vm_new_frozen_pages,
1092 : vacrel->vm_new_frozen_pages);
1093 239598 : if (vacrel->do_index_vacuuming)
1094 : {
1095 164894 : if (vacrel->nindexes == 0 || vacrel->num_index_scans == 0)
1096 164852 : appendStringInfoString(&buf, _("index scan not needed: "));
1097 : else
1098 42 : appendStringInfoString(&buf, _("index scan needed: "));
1099 :
1100 164894 : msgfmt = _("%u pages from table (%.2f%% of total) had %" PRId64 " dead item identifiers removed\n");
1101 : }
1102 : else
1103 : {
1104 74704 : if (!VacuumFailsafeActive)
1105 0 : appendStringInfoString(&buf, _("index scan bypassed: "));
1106 : else
1107 74704 : appendStringInfoString(&buf, _("index scan bypassed by failsafe: "));
1108 :
1109 74704 : msgfmt = _("%u pages from table (%.2f%% of total) have %" PRId64 " dead item identifiers\n");
1110 : }
1111 334976 : appendStringInfo(&buf, msgfmt,
1112 : vacrel->lpdead_item_pages,
1113 : orig_rel_pages == 0 ? 100.0 :
1114 95378 : 100.0 * vacrel->lpdead_item_pages / orig_rel_pages,
1115 : vacrel->lpdead_items);
1116 600024 : for (int i = 0; i < vacrel->nindexes; i++)
1117 : {
1118 360426 : IndexBulkDeleteResult *istat = vacrel->indstats[i];
1119 :
1120 360426 : if (!istat)
1121 360348 : continue;
1122 :
1123 78 : appendStringInfo(&buf,
1124 78 : _("index \"%s\": pages: %u in total, %u newly deleted, %u currently deleted, %u reusable\n"),
1125 78 : indnames[i],
1126 : istat->num_pages,
1127 : istat->pages_newly_deleted,
1128 : istat->pages_deleted,
1129 : istat->pages_free);
1130 : }
1131 239598 : if (track_cost_delay_timing)
1132 : {
1133 : /*
1134 : * We bypass the changecount mechanism because this value is
1135 : * only updated by the calling process. We also rely on the
1136 : * above call to pgstat_progress_end_command() to not clear
1137 : * the st_progress_param array.
1138 : */
1139 0 : appendStringInfo(&buf, _("delay time: %.3f ms\n"),
1140 0 : (double) MyBEEntry->st_progress_param[PROGRESS_VACUUM_DELAY_TIME] / 1000000.0);
1141 : }
1142 239598 : if (track_io_timing)
1143 : {
1144 0 : double read_ms = (double) (pgStatBlockReadTime - startreadtime) / 1000;
1145 0 : double write_ms = (double) (pgStatBlockWriteTime - startwritetime) / 1000;
1146 :
1147 0 : appendStringInfo(&buf, _("I/O timings: read: %.3f ms, write: %.3f ms\n"),
1148 : read_ms, write_ms);
1149 : }
1150 239598 : if (secs_dur > 0 || usecs_dur > 0)
1151 : {
1152 239598 : read_rate = (double) BLCKSZ * total_blks_read /
1153 239598 : (1024 * 1024) / (secs_dur + usecs_dur / 1000000.0);
1154 239598 : write_rate = (double) BLCKSZ * total_blks_dirtied /
1155 239598 : (1024 * 1024) / (secs_dur + usecs_dur / 1000000.0);
1156 : }
1157 239598 : appendStringInfo(&buf, _("avg read rate: %.3f MB/s, avg write rate: %.3f MB/s\n"),
1158 : read_rate, write_rate);
1159 239598 : appendStringInfo(&buf,
1160 239598 : _("buffer usage: %" PRId64 " hits, %" PRId64 " reads, %" PRId64 " dirtied\n"),
1161 : total_blks_hit,
1162 : total_blks_read,
1163 : total_blks_dirtied);
1164 239598 : appendStringInfo(&buf,
1165 239598 : _("WAL usage: %" PRId64 " records, %" PRId64 " full page images, %" PRIu64 " bytes, %" PRIu64 " full page image bytes, %" PRId64 " buffers full\n"),
1166 : walusage.wal_records,
1167 : walusage.wal_fpi,
1168 : walusage.wal_bytes,
1169 : walusage.wal_fpi_bytes,
1170 : walusage.wal_buffers_full);
1171 239598 : appendStringInfo(&buf, _("system usage: %s"), pg_rusage_show(&ru0));
1172 :
1173 239598 : ereport(verbose ? INFO : LOG,
1174 : (errmsg_internal("%s", buf.data)));
1175 239598 : pfree(buf.data);
1176 : }
1177 : }
1178 :
1179 : /* Cleanup index statistics and index names */
1180 667748 : for (int i = 0; i < vacrel->nindexes; i++)
1181 : {
1182 400412 : if (vacrel->indstats[i])
1183 2864 : pfree(vacrel->indstats[i]);
1184 :
1185 400412 : if (instrument)
1186 360918 : pfree(indnames[i]);
1187 : }
1188 267336 : }
1189 :
1190 : /*
1191 : * lazy_scan_heap() -- workhorse function for VACUUM
1192 : *
1193 : * This routine prunes each page in the heap, and considers the need to
1194 : * freeze remaining tuples with storage (not including pages that can be
1195 : * skipped using the visibility map). Also performs related maintenance
1196 : * of the FSM and visibility map. These steps all take place during an
1197 : * initial pass over the target heap relation.
1198 : *
1199 : * Also invokes lazy_vacuum_all_indexes to vacuum indexes, which largely
1200 : * consists of deleting index tuples that point to LP_DEAD items left in
1201 : * heap pages following pruning. Earlier initial pass over the heap will
1202 : * have collected the TIDs whose index tuples need to be removed.
1203 : *
1204 : * Finally, invokes lazy_vacuum_heap_rel to vacuum heap pages, which
1205 : * largely consists of marking LP_DEAD items (from vacrel->dead_items)
1206 : * as LP_UNUSED. This has to happen in a second, final pass over the
1207 : * heap, to preserve a basic invariant that all index AMs rely on: no
1208 : * extant index tuple can ever be allowed to contain a TID that points to
1209 : * an LP_UNUSED line pointer in the heap. We must disallow premature
1210 : * recycling of line pointers to avoid index scans that get confused
1211 : * about which TID points to which tuple immediately after recycling.
1212 : * (Actually, this isn't a concern when target heap relation happens to
1213 : * have no indexes, which allows us to safely apply the one-pass strategy
1214 : * as an optimization).
1215 : *
1216 : * In practice we often have enough space to fit all TIDs, and so won't
1217 : * need to call lazy_vacuum more than once, after our initial pass over
1218 : * the heap has totally finished. Otherwise things are slightly more
1219 : * complicated: our "initial pass" over the heap applies only to those
1220 : * pages that were pruned before we needed to call lazy_vacuum, and our
1221 : * "final pass" over the heap only vacuums these same heap pages.
1222 : * However, we process indexes in full every time lazy_vacuum is called,
1223 : * which makes index processing very inefficient when memory is in short
1224 : * supply.
1225 : */
1226 : static void
1227 267340 : lazy_scan_heap(LVRelState *vacrel)
1228 : {
1229 : ReadStream *stream;
1230 267340 : BlockNumber rel_pages = vacrel->rel_pages,
1231 267340 : blkno = 0,
1232 267340 : next_fsm_block_to_vacuum = 0;
1233 267340 : BlockNumber orig_eager_scan_success_limit =
1234 : vacrel->eager_scan_remaining_successes; /* for logging */
1235 267340 : Buffer vmbuffer = InvalidBuffer;
1236 267340 : const int initprog_index[] = {
1237 : PROGRESS_VACUUM_PHASE,
1238 : PROGRESS_VACUUM_TOTAL_HEAP_BLKS,
1239 : PROGRESS_VACUUM_MAX_DEAD_TUPLE_BYTES
1240 : };
1241 : int64 initprog_val[3];
1242 :
1243 : /* Report that we're scanning the heap, advertising total # of blocks */
1244 267340 : initprog_val[0] = PROGRESS_VACUUM_PHASE_SCAN_HEAP;
1245 267340 : initprog_val[1] = rel_pages;
1246 267340 : initprog_val[2] = vacrel->dead_items_info->max_bytes;
1247 267340 : pgstat_progress_update_multi_param(3, initprog_index, initprog_val);
1248 :
1249 : /* Initialize for the first heap_vac_scan_next_block() call */
1250 267340 : vacrel->current_block = InvalidBlockNumber;
1251 267340 : vacrel->next_unskippable_block = InvalidBlockNumber;
1252 267340 : vacrel->next_unskippable_allvis = false;
1253 267340 : vacrel->next_unskippable_eager_scanned = false;
1254 267340 : vacrel->next_unskippable_vmbuffer = InvalidBuffer;
1255 :
1256 : /*
1257 : * Set up the read stream for vacuum's first pass through the heap.
1258 : *
1259 : * This could be made safe for READ_STREAM_USE_BATCHING, but only with
1260 : * explicit work in heap_vac_scan_next_block.
1261 : */
1262 267340 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE,
1263 : vacrel->bstrategy,
1264 : vacrel->rel,
1265 : MAIN_FORKNUM,
1266 : heap_vac_scan_next_block,
1267 : vacrel,
1268 : sizeof(uint8));
1269 :
1270 : while (true)
1271 1245168 : {
1272 : Buffer buf;
1273 : Page page;
1274 1512508 : uint8 blk_info = 0;
1275 1512508 : int ndeleted = 0;
1276 : bool has_lpdead_items;
1277 1512508 : void *per_buffer_data = NULL;
1278 1512508 : bool vm_page_frozen = false;
1279 1512508 : bool got_cleanup_lock = false;
1280 :
1281 1512508 : vacuum_delay_point(false);
1282 :
1283 : /*
1284 : * Regularly check if wraparound failsafe should trigger.
1285 : *
1286 : * There is a similar check inside lazy_vacuum_all_indexes(), but
1287 : * relfrozenxid might start to look dangerously old before we reach
1288 : * that point. This check also provides failsafe coverage for the
1289 : * one-pass strategy, and the two-pass strategy with the index_cleanup
1290 : * param set to 'off'.
1291 : */
1292 1512508 : if (vacrel->scanned_pages > 0 &&
1293 1245168 : vacrel->scanned_pages % FAILSAFE_EVERY_PAGES == 0)
1294 0 : lazy_check_wraparound_failsafe(vacrel);
1295 :
1296 : /*
1297 : * Consider if we definitely have enough space to process TIDs on page
1298 : * already. If we are close to overrunning the available space for
1299 : * dead_items TIDs, pause and do a cycle of vacuuming before we tackle
1300 : * this page. However, let's force at least one page-worth of tuples
1301 : * to be stored as to ensure we do at least some work when the memory
1302 : * configured is so low that we run out before storing anything.
1303 : */
1304 1512508 : if (vacrel->dead_items_info->num_items > 0 &&
1305 56088 : TidStoreMemoryUsage(vacrel->dead_items) > vacrel->dead_items_info->max_bytes)
1306 : {
1307 : /*
1308 : * Before beginning index vacuuming, we release any pin we may
1309 : * hold on the visibility map page. This isn't necessary for
1310 : * correctness, but we do it anyway to avoid holding the pin
1311 : * across a lengthy, unrelated operation.
1312 : */
1313 4 : if (BufferIsValid(vmbuffer))
1314 : {
1315 4 : ReleaseBuffer(vmbuffer);
1316 4 : vmbuffer = InvalidBuffer;
1317 : }
1318 :
1319 : /* Perform a round of index and heap vacuuming */
1320 4 : vacrel->consider_bypass_optimization = false;
1321 4 : lazy_vacuum(vacrel);
1322 :
1323 : /*
1324 : * Vacuum the Free Space Map to make newly-freed space visible on
1325 : * upper-level FSM pages. Note that blkno is the previously
1326 : * processed block.
1327 : */
1328 4 : FreeSpaceMapVacuumRange(vacrel->rel, next_fsm_block_to_vacuum,
1329 : blkno + 1);
1330 4 : next_fsm_block_to_vacuum = blkno;
1331 :
1332 : /* Report that we are once again scanning the heap */
1333 4 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
1334 : PROGRESS_VACUUM_PHASE_SCAN_HEAP);
1335 : }
1336 :
1337 1512508 : buf = read_stream_next_buffer(stream, &per_buffer_data);
1338 :
1339 : /* The relation is exhausted. */
1340 1512508 : if (!BufferIsValid(buf))
1341 267340 : break;
1342 :
1343 1245168 : blk_info = *((uint8 *) per_buffer_data);
1344 1245168 : CheckBufferIsPinnedOnce(buf);
1345 1245168 : page = BufferGetPage(buf);
1346 1245168 : blkno = BufferGetBlockNumber(buf);
1347 :
1348 1245168 : vacrel->scanned_pages++;
1349 1245168 : if (blk_info & VAC_BLK_WAS_EAGER_SCANNED)
1350 0 : vacrel->eager_scanned_pages++;
1351 :
1352 : /* Report as block scanned, update error traceback information */
1353 1245168 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_SCANNED, blkno);
1354 1245168 : update_vacuum_error_info(vacrel, NULL, VACUUM_ERRCB_PHASE_SCAN_HEAP,
1355 : blkno, InvalidOffsetNumber);
1356 :
1357 : /*
1358 : * Pin the visibility map page in case we need to mark the page
1359 : * all-visible. In most cases this will be very cheap, because we'll
1360 : * already have the correct page pinned anyway.
1361 : */
1362 1245168 : visibilitymap_pin(vacrel->rel, blkno, &vmbuffer);
1363 :
1364 : /*
1365 : * We need a buffer cleanup lock to prune HOT chains and defragment
1366 : * the page in lazy_scan_prune. But when it's not possible to acquire
1367 : * a cleanup lock right away, we may be able to settle for reduced
1368 : * processing using lazy_scan_noprune.
1369 : */
1370 1245168 : got_cleanup_lock = ConditionalLockBufferForCleanup(buf);
1371 :
1372 1245168 : if (!got_cleanup_lock)
1373 290 : LockBuffer(buf, BUFFER_LOCK_SHARE);
1374 :
1375 : /* Check for new or empty pages before lazy_scan_[no]prune call */
1376 1245168 : if (lazy_scan_new_or_empty(vacrel, buf, blkno, page, !got_cleanup_lock,
1377 1245168 : vmbuffer))
1378 : {
1379 : /* Processed as new/empty page (lock and pin released) */
1380 1960 : continue;
1381 : }
1382 :
1383 : /*
1384 : * If we didn't get the cleanup lock, we can still collect LP_DEAD
1385 : * items in the dead_items area for later vacuuming, count live and
1386 : * recently dead tuples for vacuum logging, and determine if this
1387 : * block could later be truncated. If we encounter any xid/mxids that
1388 : * require advancing the relfrozenxid/relminxid, we'll have to wait
1389 : * for a cleanup lock and call lazy_scan_prune().
1390 : */
1391 1243208 : if (!got_cleanup_lock &&
1392 290 : !lazy_scan_noprune(vacrel, buf, blkno, page, &has_lpdead_items))
1393 : {
1394 : /*
1395 : * lazy_scan_noprune could not do all required processing. Wait
1396 : * for a cleanup lock, and call lazy_scan_prune in the usual way.
1397 : */
1398 : Assert(vacrel->aggressive);
1399 110 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1400 110 : LockBufferForCleanup(buf);
1401 110 : got_cleanup_lock = true;
1402 : }
1403 :
1404 : /*
1405 : * If we have a cleanup lock, we must now prune, freeze, and count
1406 : * tuples. We may have acquired the cleanup lock originally, or we may
1407 : * have gone back and acquired it after lazy_scan_noprune() returned
1408 : * false. Either way, the page hasn't been processed yet.
1409 : *
1410 : * Like lazy_scan_noprune(), lazy_scan_prune() will count
1411 : * recently_dead_tuples and live tuples for vacuum logging, determine
1412 : * if the block can later be truncated, and accumulate the details of
1413 : * remaining LP_DEAD line pointers on the page into dead_items. These
1414 : * dead items include those pruned by lazy_scan_prune() as well as
1415 : * line pointers previously marked LP_DEAD.
1416 : */
1417 1243208 : if (got_cleanup_lock)
1418 1243028 : ndeleted = lazy_scan_prune(vacrel, buf, blkno, page,
1419 : vmbuffer,
1420 1243028 : blk_info & VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM,
1421 : &has_lpdead_items, &vm_page_frozen);
1422 :
1423 : /*
1424 : * Count an eagerly scanned page as a failure or a success.
1425 : *
1426 : * Only lazy_scan_prune() freezes pages, so if we didn't get the
1427 : * cleanup lock, we won't have frozen the page. However, we only count
1428 : * pages that were too new to require freezing as eager freeze
1429 : * failures.
1430 : *
1431 : * We could gather more information from lazy_scan_noprune() about
1432 : * whether or not there were tuples with XIDs or MXIDs older than the
1433 : * FreezeLimit or MultiXactCutoff. However, for simplicity, we simply
1434 : * exclude pages skipped due to cleanup lock contention from eager
1435 : * freeze algorithm caps.
1436 : */
1437 1243208 : if (got_cleanup_lock &&
1438 1243028 : (blk_info & VAC_BLK_WAS_EAGER_SCANNED))
1439 : {
1440 : /* Aggressive vacuums do not eager scan. */
1441 : Assert(!vacrel->aggressive);
1442 :
1443 0 : if (vm_page_frozen)
1444 : {
1445 0 : if (vacrel->eager_scan_remaining_successes > 0)
1446 0 : vacrel->eager_scan_remaining_successes--;
1447 :
1448 0 : if (vacrel->eager_scan_remaining_successes == 0)
1449 : {
1450 : /*
1451 : * Report only once that we disabled eager scanning. We
1452 : * may eagerly read ahead blocks in excess of the success
1453 : * or failure caps before attempting to freeze them, so we
1454 : * could reach here even after disabling additional eager
1455 : * scanning.
1456 : */
1457 0 : if (vacrel->eager_scan_max_fails_per_region > 0)
1458 0 : ereport(vacrel->verbose ? INFO : DEBUG2,
1459 : (errmsg("disabling eager scanning after freezing %u eagerly scanned blocks of relation \"%s.%s.%s\"",
1460 : orig_eager_scan_success_limit,
1461 : vacrel->dbname, vacrel->relnamespace,
1462 : vacrel->relname)));
1463 :
1464 : /*
1465 : * If we hit our success cap, permanently disable eager
1466 : * scanning by setting the other eager scan management
1467 : * fields to their disabled values.
1468 : */
1469 0 : vacrel->eager_scan_remaining_fails = 0;
1470 0 : vacrel->next_eager_scan_region_start = InvalidBlockNumber;
1471 0 : vacrel->eager_scan_max_fails_per_region = 0;
1472 : }
1473 : }
1474 0 : else if (vacrel->eager_scan_remaining_fails > 0)
1475 0 : vacrel->eager_scan_remaining_fails--;
1476 : }
1477 :
1478 : /*
1479 : * Now drop the buffer lock and, potentially, update the FSM.
1480 : *
1481 : * Our goal is to update the freespace map the last time we touch the
1482 : * page. If we'll process a block in the second pass, we may free up
1483 : * additional space on the page, so it is better to update the FSM
1484 : * after the second pass. If the relation has no indexes, or if index
1485 : * vacuuming is disabled, there will be no second heap pass; if this
1486 : * particular page has no dead items, the second heap pass will not
1487 : * touch this page. So, in those cases, update the FSM now.
1488 : *
1489 : * Note: In corner cases, it's possible to miss updating the FSM
1490 : * entirely. If index vacuuming is currently enabled, we'll skip the
1491 : * FSM update now. But if failsafe mode is later activated, or there
1492 : * are so few dead tuples that index vacuuming is bypassed, there will
1493 : * also be no opportunity to update the FSM later, because we'll never
1494 : * revisit this page. Since updating the FSM is desirable but not
1495 : * absolutely required, that's OK.
1496 : */
1497 1243208 : if (vacrel->nindexes == 0
1498 1205348 : || !vacrel->do_index_vacuuming
1499 852550 : || !has_lpdead_items)
1500 1211508 : {
1501 1211508 : Size freespace = PageGetHeapFreeSpace(page);
1502 :
1503 1211508 : UnlockReleaseBuffer(buf);
1504 1211508 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
1505 :
1506 : /*
1507 : * Periodically perform FSM vacuuming to make newly-freed space
1508 : * visible on upper FSM pages. This is done after vacuuming if the
1509 : * table has indexes. There will only be newly-freed space if we
1510 : * held the cleanup lock and lazy_scan_prune() was called.
1511 : */
1512 1211508 : if (got_cleanup_lock && vacrel->nindexes == 0 && ndeleted > 0 &&
1513 912 : blkno - next_fsm_block_to_vacuum >= VACUUM_FSM_EVERY_PAGES)
1514 : {
1515 0 : FreeSpaceMapVacuumRange(vacrel->rel, next_fsm_block_to_vacuum,
1516 : blkno);
1517 0 : next_fsm_block_to_vacuum = blkno;
1518 : }
1519 : }
1520 : else
1521 31700 : UnlockReleaseBuffer(buf);
1522 : }
1523 :
1524 267340 : vacrel->blkno = InvalidBlockNumber;
1525 267340 : if (BufferIsValid(vmbuffer))
1526 107096 : ReleaseBuffer(vmbuffer);
1527 :
1528 : /*
1529 : * Report that everything is now scanned. We never skip scanning the last
1530 : * block in the relation, so we can pass rel_pages here.
1531 : */
1532 267340 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_SCANNED,
1533 : rel_pages);
1534 :
1535 : /* now we can compute the new value for pg_class.reltuples */
1536 534680 : vacrel->new_live_tuples = vac_estimate_reltuples(vacrel->rel, rel_pages,
1537 : vacrel->scanned_pages,
1538 267340 : vacrel->live_tuples);
1539 :
1540 : /*
1541 : * Also compute the total number of surviving heap entries. In the
1542 : * (unlikely) scenario that new_live_tuples is -1, take it as zero.
1543 : */
1544 267340 : vacrel->new_rel_tuples =
1545 267340 : Max(vacrel->new_live_tuples, 0) + vacrel->recently_dead_tuples +
1546 267340 : vacrel->missed_dead_tuples;
1547 :
1548 267340 : read_stream_end(stream);
1549 :
1550 : /*
1551 : * Do index vacuuming (call each index's ambulkdelete routine), then do
1552 : * related heap vacuuming
1553 : */
1554 267340 : if (vacrel->dead_items_info->num_items > 0)
1555 1358 : lazy_vacuum(vacrel);
1556 :
1557 : /*
1558 : * Vacuum the remainder of the Free Space Map. We must do this whether or
1559 : * not there were indexes, and whether or not we bypassed index vacuuming.
1560 : * We can pass rel_pages here because we never skip scanning the last
1561 : * block of the relation.
1562 : */
1563 267340 : if (rel_pages > next_fsm_block_to_vacuum)
1564 107096 : FreeSpaceMapVacuumRange(vacrel->rel, next_fsm_block_to_vacuum, rel_pages);
1565 :
1566 : /* report all blocks vacuumed */
1567 267340 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_VACUUMED, rel_pages);
1568 :
1569 : /* Do final index cleanup (call each index's amvacuumcleanup routine) */
1570 267340 : if (vacrel->nindexes > 0 && vacrel->do_index_cleanup)
1571 182870 : lazy_cleanup_all_indexes(vacrel);
1572 267340 : }
1573 :
1574 : /*
1575 : * heap_vac_scan_next_block() -- read stream callback to get the next block
1576 : * for vacuum to process
1577 : *
1578 : * Every time lazy_scan_heap() needs a new block to process during its first
1579 : * phase, it invokes read_stream_next_buffer() with a stream set up to call
1580 : * heap_vac_scan_next_block() to get the next block.
1581 : *
1582 : * heap_vac_scan_next_block() uses the visibility map, vacuum options, and
1583 : * various thresholds to skip blocks which do not need to be processed and
1584 : * returns the next block to process or InvalidBlockNumber if there are no
1585 : * remaining blocks.
1586 : *
1587 : * The visibility status of the next block to process and whether or not it
1588 : * was eager scanned is set in the per_buffer_data.
1589 : *
1590 : * callback_private_data contains a reference to the LVRelState, passed to the
1591 : * read stream API during stream setup. The LVRelState is an in/out parameter
1592 : * here (locally named `vacrel`). Vacuum options and information about the
1593 : * relation are read from it. vacrel->skippedallvis is set if we skip a block
1594 : * that's all-visible but not all-frozen (to ensure that we don't update
1595 : * relfrozenxid in that case). vacrel also holds information about the next
1596 : * unskippable block -- as bookkeeping for this function.
1597 : */
1598 : static BlockNumber
1599 1512508 : heap_vac_scan_next_block(ReadStream *stream,
1600 : void *callback_private_data,
1601 : void *per_buffer_data)
1602 : {
1603 : BlockNumber next_block;
1604 1512508 : LVRelState *vacrel = callback_private_data;
1605 1512508 : uint8 blk_info = 0;
1606 :
1607 : /* relies on InvalidBlockNumber + 1 overflowing to 0 on first call */
1608 1512508 : next_block = vacrel->current_block + 1;
1609 :
1610 : /* Have we reached the end of the relation? */
1611 1512508 : if (next_block >= vacrel->rel_pages)
1612 : {
1613 267340 : if (BufferIsValid(vacrel->next_unskippable_vmbuffer))
1614 : {
1615 104146 : ReleaseBuffer(vacrel->next_unskippable_vmbuffer);
1616 104146 : vacrel->next_unskippable_vmbuffer = InvalidBuffer;
1617 : }
1618 267340 : return InvalidBlockNumber;
1619 : }
1620 :
1621 : /*
1622 : * We must be in one of the three following states:
1623 : */
1624 1245168 : if (next_block > vacrel->next_unskippable_block ||
1625 432388 : vacrel->next_unskippable_block == InvalidBlockNumber)
1626 : {
1627 : /*
1628 : * 1. We have just processed an unskippable block (or we're at the
1629 : * beginning of the scan). Find the next unskippable block using the
1630 : * visibility map.
1631 : */
1632 : bool skipsallvis;
1633 :
1634 919876 : find_next_unskippable_block(vacrel, &skipsallvis);
1635 :
1636 : /*
1637 : * We now know the next block that we must process. It can be the
1638 : * next block after the one we just processed, or something further
1639 : * ahead. If it's further ahead, we can jump to it, but we choose to
1640 : * do so only if we can skip at least SKIP_PAGES_THRESHOLD consecutive
1641 : * pages. Since we're reading sequentially, the OS should be doing
1642 : * readahead for us, so there's no gain in skipping a page now and
1643 : * then. Skipping such a range might even discourage sequential
1644 : * detection.
1645 : *
1646 : * This test also enables more frequent relfrozenxid advancement
1647 : * during non-aggressive VACUUMs. If the range has any all-visible
1648 : * pages then skipping makes updating relfrozenxid unsafe, which is a
1649 : * real downside.
1650 : */
1651 919876 : if (vacrel->next_unskippable_block - next_block >= SKIP_PAGES_THRESHOLD)
1652 : {
1653 9650 : next_block = vacrel->next_unskippable_block;
1654 9650 : if (skipsallvis)
1655 74 : vacrel->skippedallvis = true;
1656 : }
1657 : }
1658 :
1659 : /* Now we must be in one of the two remaining states: */
1660 1245168 : if (next_block < vacrel->next_unskippable_block)
1661 : {
1662 : /*
1663 : * 2. We are processing a range of blocks that we could have skipped
1664 : * but chose not to. We know that they are all-visible in the VM,
1665 : * otherwise they would've been unskippable.
1666 : */
1667 325292 : vacrel->current_block = next_block;
1668 325292 : blk_info |= VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM;
1669 325292 : *((uint8 *) per_buffer_data) = blk_info;
1670 325292 : return vacrel->current_block;
1671 : }
1672 : else
1673 : {
1674 : /*
1675 : * 3. We reached the next unskippable block. Process it. On next
1676 : * iteration, we will be back in state 1.
1677 : */
1678 : Assert(next_block == vacrel->next_unskippable_block);
1679 :
1680 919876 : vacrel->current_block = next_block;
1681 919876 : if (vacrel->next_unskippable_allvis)
1682 97828 : blk_info |= VAC_BLK_ALL_VISIBLE_ACCORDING_TO_VM;
1683 919876 : if (vacrel->next_unskippable_eager_scanned)
1684 0 : blk_info |= VAC_BLK_WAS_EAGER_SCANNED;
1685 919876 : *((uint8 *) per_buffer_data) = blk_info;
1686 919876 : return vacrel->current_block;
1687 : }
1688 : }
1689 :
1690 : /*
1691 : * Find the next unskippable block in a vacuum scan using the visibility map.
1692 : * The next unskippable block and its visibility information is updated in
1693 : * vacrel.
1694 : *
1695 : * Note: our opinion of which blocks can be skipped can go stale immediately.
1696 : * It's okay if caller "misses" a page whose all-visible or all-frozen marking
1697 : * was concurrently cleared, though. All that matters is that caller scan all
1698 : * pages whose tuples might contain XIDs < OldestXmin, or MXIDs < OldestMxact.
1699 : * (Actually, non-aggressive VACUUMs can choose to skip all-visible pages with
1700 : * older XIDs/MXIDs. The *skippedallvis flag will be set here when the choice
1701 : * to skip such a range is actually made, making everything safe.)
1702 : */
1703 : static void
1704 919876 : find_next_unskippable_block(LVRelState *vacrel, bool *skipsallvis)
1705 : {
1706 919876 : BlockNumber rel_pages = vacrel->rel_pages;
1707 919876 : BlockNumber next_unskippable_block = vacrel->next_unskippable_block + 1;
1708 919876 : Buffer next_unskippable_vmbuffer = vacrel->next_unskippable_vmbuffer;
1709 919876 : bool next_unskippable_eager_scanned = false;
1710 : bool next_unskippable_allvis;
1711 :
1712 919876 : *skipsallvis = false;
1713 :
1714 1035222 : for (;; next_unskippable_block++)
1715 1035222 : {
1716 1955098 : uint8 mapbits = visibilitymap_get_status(vacrel->rel,
1717 : next_unskippable_block,
1718 : &next_unskippable_vmbuffer);
1719 :
1720 1955098 : next_unskippable_allvis = (mapbits & VISIBILITYMAP_ALL_VISIBLE) != 0;
1721 :
1722 : /*
1723 : * At the start of each eager scan region, normal vacuums with eager
1724 : * scanning enabled reset the failure counter, allowing vacuum to
1725 : * resume eager scanning if it had been suspended in the previous
1726 : * region.
1727 : */
1728 1955098 : if (next_unskippable_block >= vacrel->next_eager_scan_region_start)
1729 : {
1730 0 : vacrel->eager_scan_remaining_fails =
1731 0 : vacrel->eager_scan_max_fails_per_region;
1732 0 : vacrel->next_eager_scan_region_start += EAGER_SCAN_REGION_SIZE;
1733 : }
1734 :
1735 : /*
1736 : * A block is unskippable if it is not all visible according to the
1737 : * visibility map.
1738 : */
1739 1955098 : if (!next_unskippable_allvis)
1740 : {
1741 : Assert((mapbits & VISIBILITYMAP_ALL_FROZEN) == 0);
1742 822048 : break;
1743 : }
1744 :
1745 : /*
1746 : * Caller must scan the last page to determine whether it has tuples
1747 : * (caller must have the opportunity to set vacrel->nonempty_pages).
1748 : * This rule avoids having lazy_truncate_heap() take access-exclusive
1749 : * lock on rel to attempt a truncation that fails anyway, just because
1750 : * there are tuples on the last page (it is likely that there will be
1751 : * tuples on other nearby pages as well, but those can be skipped).
1752 : *
1753 : * Implement this by always treating the last block as unsafe to skip.
1754 : */
1755 1133050 : if (next_unskippable_block == rel_pages - 1)
1756 97016 : break;
1757 :
1758 : /* DISABLE_PAGE_SKIPPING makes all skipping unsafe */
1759 1036034 : if (!vacrel->skipwithvm)
1760 810 : break;
1761 :
1762 : /*
1763 : * All-frozen pages cannot contain XIDs < OldestXmin (XIDs that aren't
1764 : * already frozen by now), so this page can be skipped.
1765 : */
1766 1035224 : if ((mapbits & VISIBILITYMAP_ALL_FROZEN) != 0)
1767 1029662 : continue;
1768 :
1769 : /*
1770 : * Aggressive vacuums cannot skip any all-visible pages that are not
1771 : * also all-frozen.
1772 : */
1773 5562 : if (vacrel->aggressive)
1774 2 : break;
1775 :
1776 : /*
1777 : * Normal vacuums with eager scanning enabled only skip all-visible
1778 : * but not all-frozen pages if they have hit the failure limit for the
1779 : * current eager scan region.
1780 : */
1781 5560 : if (vacrel->eager_scan_remaining_fails > 0)
1782 : {
1783 0 : next_unskippable_eager_scanned = true;
1784 0 : break;
1785 : }
1786 :
1787 : /*
1788 : * All-visible blocks are safe to skip in a normal vacuum. But
1789 : * remember that the final range contains such a block for later.
1790 : */
1791 5560 : *skipsallvis = true;
1792 : }
1793 :
1794 : /* write the local variables back to vacrel */
1795 919876 : vacrel->next_unskippable_block = next_unskippable_block;
1796 919876 : vacrel->next_unskippable_allvis = next_unskippable_allvis;
1797 919876 : vacrel->next_unskippable_eager_scanned = next_unskippable_eager_scanned;
1798 919876 : vacrel->next_unskippable_vmbuffer = next_unskippable_vmbuffer;
1799 919876 : }
1800 :
1801 : /*
1802 : * lazy_scan_new_or_empty() -- lazy_scan_heap() new/empty page handling.
1803 : *
1804 : * Must call here to handle both new and empty pages before calling
1805 : * lazy_scan_prune or lazy_scan_noprune, since they're not prepared to deal
1806 : * with new or empty pages.
1807 : *
1808 : * It's necessary to consider new pages as a special case, since the rules for
1809 : * maintaining the visibility map and FSM with empty pages are a little
1810 : * different (though new pages can be truncated away during rel truncation).
1811 : *
1812 : * Empty pages are not really a special case -- they're just heap pages that
1813 : * have no allocated tuples (including even LP_UNUSED items). You might
1814 : * wonder why we need to handle them here all the same. It's only necessary
1815 : * because of a corner-case involving a hard crash during heap relation
1816 : * extension. If we ever make relation-extension crash safe, then it should
1817 : * no longer be necessary to deal with empty pages here (or new pages, for
1818 : * that matter).
1819 : *
1820 : * Caller must hold at least a shared lock. We might need to escalate the
1821 : * lock in that case, so the type of lock caller holds needs to be specified
1822 : * using 'sharelock' argument.
1823 : *
1824 : * Returns false in common case where caller should go on to call
1825 : * lazy_scan_prune (or lazy_scan_noprune). Otherwise returns true, indicating
1826 : * that lazy_scan_heap is done processing the page, releasing lock on caller's
1827 : * behalf.
1828 : *
1829 : * No vm_page_frozen output parameter (like that passed to lazy_scan_prune())
1830 : * is passed here because neither empty nor new pages can be eagerly frozen.
1831 : * New pages are never frozen. Empty pages are always set frozen in the VM at
1832 : * the same time that they are set all-visible, and we don't eagerly scan
1833 : * frozen pages.
1834 : */
1835 : static bool
1836 1245168 : lazy_scan_new_or_empty(LVRelState *vacrel, Buffer buf, BlockNumber blkno,
1837 : Page page, bool sharelock, Buffer vmbuffer)
1838 : {
1839 : Size freespace;
1840 :
1841 1245168 : if (PageIsNew(page))
1842 : {
1843 : /*
1844 : * All-zeroes pages can be left over if either a backend extends the
1845 : * relation by a single page, but crashes before the newly initialized
1846 : * page has been written out, or when bulk-extending the relation
1847 : * (which creates a number of empty pages at the tail end of the
1848 : * relation), and then enters them into the FSM.
1849 : *
1850 : * Note we do not enter the page into the visibilitymap. That has the
1851 : * downside that we repeatedly visit this page in subsequent vacuums,
1852 : * but otherwise we'll never discover the space on a promoted standby.
1853 : * The harm of repeated checking ought to normally not be too bad. The
1854 : * space usually should be used at some point, otherwise there
1855 : * wouldn't be any regular vacuums.
1856 : *
1857 : * Make sure these pages are in the FSM, to ensure they can be reused.
1858 : * Do that by testing if there's any space recorded for the page. If
1859 : * not, enter it. We do so after releasing the lock on the heap page,
1860 : * the FSM is approximate, after all.
1861 : */
1862 1906 : UnlockReleaseBuffer(buf);
1863 :
1864 1906 : if (GetRecordedFreeSpace(vacrel->rel, blkno) == 0)
1865 : {
1866 950 : freespace = BLCKSZ - SizeOfPageHeaderData;
1867 :
1868 950 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
1869 : }
1870 :
1871 1906 : return true;
1872 : }
1873 :
1874 1243262 : if (PageIsEmpty(page))
1875 : {
1876 : /*
1877 : * It seems likely that caller will always be able to get a cleanup
1878 : * lock on an empty page. But don't take any chances -- escalate to
1879 : * an exclusive lock (still don't need a cleanup lock, though).
1880 : */
1881 54 : if (sharelock)
1882 : {
1883 0 : LockBuffer(buf, BUFFER_LOCK_UNLOCK);
1884 0 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
1885 :
1886 0 : if (!PageIsEmpty(page))
1887 : {
1888 : /* page isn't new or empty -- keep lock and pin for now */
1889 0 : return false;
1890 : }
1891 : }
1892 : else
1893 : {
1894 : /* Already have a full cleanup lock (which is more than enough) */
1895 : }
1896 :
1897 : /*
1898 : * Unlike new pages, empty pages are always set all-visible and
1899 : * all-frozen.
1900 : */
1901 54 : if (!PageIsAllVisible(page))
1902 : {
1903 0 : START_CRIT_SECTION();
1904 :
1905 : /* mark buffer dirty before writing a WAL record */
1906 0 : MarkBufferDirty(buf);
1907 :
1908 : /*
1909 : * It's possible that another backend has extended the heap,
1910 : * initialized the page, and then failed to WAL-log the page due
1911 : * to an ERROR. Since heap extension is not WAL-logged, recovery
1912 : * might try to replay our record setting the page all-visible and
1913 : * find that the page isn't initialized, which will cause a PANIC.
1914 : * To prevent that, check whether the page has been previously
1915 : * WAL-logged, and if not, do that now.
1916 : */
1917 0 : if (RelationNeedsWAL(vacrel->rel) &&
1918 0 : !XLogRecPtrIsValid(PageGetLSN(page)))
1919 0 : log_newpage_buffer(buf, true);
1920 :
1921 0 : PageSetAllVisible(page);
1922 0 : visibilitymap_set(vacrel->rel, blkno, buf,
1923 : InvalidXLogRecPtr,
1924 : vmbuffer, InvalidTransactionId,
1925 : VISIBILITYMAP_ALL_VISIBLE |
1926 : VISIBILITYMAP_ALL_FROZEN);
1927 0 : END_CRIT_SECTION();
1928 :
1929 : /* Count the newly all-frozen pages for logging */
1930 0 : vacrel->vm_new_visible_pages++;
1931 0 : vacrel->vm_new_visible_frozen_pages++;
1932 : }
1933 :
1934 54 : freespace = PageGetHeapFreeSpace(page);
1935 54 : UnlockReleaseBuffer(buf);
1936 54 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
1937 54 : return true;
1938 : }
1939 :
1940 : /* page isn't new or empty -- keep lock and pin */
1941 1243208 : return false;
1942 : }
1943 :
1944 : /* qsort comparator for sorting OffsetNumbers */
1945 : static int
1946 7459716 : cmpOffsetNumbers(const void *a, const void *b)
1947 : {
1948 7459716 : return pg_cmp_u16(*(const OffsetNumber *) a, *(const OffsetNumber *) b);
1949 : }
1950 :
1951 : /*
1952 : * lazy_scan_prune() -- lazy_scan_heap() pruning and freezing.
1953 : *
1954 : * Caller must hold pin and buffer cleanup lock on the buffer.
1955 : *
1956 : * vmbuffer is the buffer containing the VM block with visibility information
1957 : * for the heap block, blkno. all_visible_according_to_vm is the saved
1958 : * visibility status of the heap block looked up earlier by the caller. We
1959 : * won't rely entirely on this status, as it may be out of date.
1960 : *
1961 : * *has_lpdead_items is set to true or false depending on whether, upon return
1962 : * from this function, any LP_DEAD items are still present on the page.
1963 : *
1964 : * *vm_page_frozen is set to true if the page is newly set all-frozen in the
1965 : * VM. The caller currently only uses this for determining whether an eagerly
1966 : * scanned page was successfully set all-frozen.
1967 : *
1968 : * Returns the number of tuples deleted from the page during HOT pruning.
1969 : */
1970 : static int
1971 1243028 : lazy_scan_prune(LVRelState *vacrel,
1972 : Buffer buf,
1973 : BlockNumber blkno,
1974 : Page page,
1975 : Buffer vmbuffer,
1976 : bool all_visible_according_to_vm,
1977 : bool *has_lpdead_items,
1978 : bool *vm_page_frozen)
1979 : {
1980 1243028 : Relation rel = vacrel->rel;
1981 : PruneFreezeResult presult;
1982 1243028 : PruneFreezeParams params = {
1983 : .relation = rel,
1984 : .buffer = buf,
1985 : .reason = PRUNE_VACUUM_SCAN,
1986 : .options = HEAP_PAGE_PRUNE_FREEZE,
1987 1243028 : .vistest = vacrel->vistest,
1988 1243028 : .cutoffs = &vacrel->cutoffs,
1989 : };
1990 :
1991 : Assert(BufferGetBlockNumber(buf) == blkno);
1992 :
1993 : /*
1994 : * Prune all HOT-update chains and potentially freeze tuples on this page.
1995 : *
1996 : * If the relation has no indexes, we can immediately mark would-be dead
1997 : * items LP_UNUSED.
1998 : *
1999 : * The number of tuples removed from the page is returned in
2000 : * presult.ndeleted. It should not be confused with presult.lpdead_items;
2001 : * presult.lpdead_items's final value can be thought of as the number of
2002 : * tuples that were deleted from indexes.
2003 : *
2004 : * We will update the VM after collecting LP_DEAD items and freezing
2005 : * tuples. Pruning will have determined whether or not the page is
2006 : * all-visible.
2007 : */
2008 1243028 : if (vacrel->nindexes == 0)
2009 37860 : params.options |= HEAP_PAGE_PRUNE_MARK_UNUSED_NOW;
2010 :
2011 1243028 : heap_page_prune_and_freeze(¶ms,
2012 : &presult,
2013 : &vacrel->offnum,
2014 : &vacrel->NewRelfrozenXid, &vacrel->NewRelminMxid);
2015 :
2016 : Assert(MultiXactIdIsValid(vacrel->NewRelminMxid));
2017 : Assert(TransactionIdIsValid(vacrel->NewRelfrozenXid));
2018 :
2019 1243028 : if (presult.nfrozen > 0)
2020 : {
2021 : /*
2022 : * We don't increment the new_frozen_tuple_pages instrumentation
2023 : * counter when nfrozen == 0, since it only counts pages with newly
2024 : * frozen tuples (don't confuse that with pages newly set all-frozen
2025 : * in VM).
2026 : */
2027 47448 : vacrel->new_frozen_tuple_pages++;
2028 : }
2029 :
2030 : /*
2031 : * VACUUM will call heap_page_is_all_visible() during the second pass over
2032 : * the heap to determine all_visible and all_frozen for the page -- this
2033 : * is a specialized version of the logic from this function. Now that
2034 : * we've finished pruning and freezing, make sure that we're in total
2035 : * agreement with heap_page_is_all_visible() using an assertion.
2036 : */
2037 : #ifdef USE_ASSERT_CHECKING
2038 : if (presult.all_visible)
2039 : {
2040 : TransactionId debug_cutoff;
2041 : bool debug_all_frozen;
2042 :
2043 : Assert(presult.lpdead_items == 0);
2044 :
2045 : Assert(heap_page_is_all_visible(vacrel->rel, buf,
2046 : vacrel->cutoffs.OldestXmin, &debug_all_frozen,
2047 : &debug_cutoff, &vacrel->offnum));
2048 :
2049 : Assert(presult.all_frozen == debug_all_frozen);
2050 :
2051 : Assert(!TransactionIdIsValid(debug_cutoff) ||
2052 : debug_cutoff == presult.vm_conflict_horizon);
2053 : }
2054 : #endif
2055 :
2056 : /*
2057 : * Now save details of the LP_DEAD items from the page in vacrel
2058 : */
2059 1243028 : if (presult.lpdead_items > 0)
2060 : {
2061 36304 : vacrel->lpdead_item_pages++;
2062 :
2063 : /*
2064 : * deadoffsets are collected incrementally in
2065 : * heap_page_prune_and_freeze() as each dead line pointer is recorded,
2066 : * with an indeterminate order, but dead_items_add requires them to be
2067 : * sorted.
2068 : */
2069 36304 : qsort(presult.deadoffsets, presult.lpdead_items, sizeof(OffsetNumber),
2070 : cmpOffsetNumbers);
2071 :
2072 36304 : dead_items_add(vacrel, blkno, presult.deadoffsets, presult.lpdead_items);
2073 : }
2074 :
2075 : /* Finally, add page-local counts to whole-VACUUM counts */
2076 1243028 : vacrel->tuples_deleted += presult.ndeleted;
2077 1243028 : vacrel->tuples_frozen += presult.nfrozen;
2078 1243028 : vacrel->lpdead_items += presult.lpdead_items;
2079 1243028 : vacrel->live_tuples += presult.live_tuples;
2080 1243028 : vacrel->recently_dead_tuples += presult.recently_dead_tuples;
2081 :
2082 : /* Can't truncate this page */
2083 1243028 : if (presult.hastup)
2084 1225124 : vacrel->nonempty_pages = blkno + 1;
2085 :
2086 : /* Did we find LP_DEAD items? */
2087 1243028 : *has_lpdead_items = (presult.lpdead_items > 0);
2088 :
2089 : Assert(!presult.all_visible || !(*has_lpdead_items));
2090 : Assert(!presult.all_frozen || presult.all_visible);
2091 :
2092 : /*
2093 : * Handle setting visibility map bit based on information from the VM (as
2094 : * of last heap_vac_scan_next_block() call), and from all_visible and
2095 : * all_frozen variables
2096 : */
2097 1243028 : if (!all_visible_according_to_vm && presult.all_visible)
2098 74130 : {
2099 : uint8 old_vmbits;
2100 74130 : uint8 flags = VISIBILITYMAP_ALL_VISIBLE;
2101 :
2102 74130 : if (presult.all_frozen)
2103 : {
2104 : Assert(!TransactionIdIsValid(presult.vm_conflict_horizon));
2105 54366 : flags |= VISIBILITYMAP_ALL_FROZEN;
2106 : }
2107 :
2108 : /*
2109 : * It should never be the case that the visibility map page is set
2110 : * while the page-level bit is clear, but the reverse is allowed (if
2111 : * checksums are not enabled). Regardless, set both bits so that we
2112 : * get back in sync.
2113 : *
2114 : * NB: If the heap page is all-visible but the VM bit is not set, we
2115 : * don't need to dirty the heap page. However, if checksums are
2116 : * enabled, we do need to make sure that the heap page is dirtied
2117 : * before passing it to visibilitymap_set(), because it may be logged.
2118 : * Given that this situation should only happen in rare cases after a
2119 : * crash, it is not worth optimizing.
2120 : */
2121 74130 : PageSetAllVisible(page);
2122 74130 : MarkBufferDirty(buf);
2123 74130 : old_vmbits = visibilitymap_set(vacrel->rel, blkno, buf,
2124 : InvalidXLogRecPtr,
2125 : vmbuffer, presult.vm_conflict_horizon,
2126 : flags);
2127 :
2128 : /*
2129 : * If the page wasn't already set all-visible and/or all-frozen in the
2130 : * VM, count it as newly set for logging.
2131 : */
2132 74130 : if ((old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
2133 : {
2134 74130 : vacrel->vm_new_visible_pages++;
2135 74130 : if (presult.all_frozen)
2136 : {
2137 54366 : vacrel->vm_new_visible_frozen_pages++;
2138 54366 : *vm_page_frozen = true;
2139 : }
2140 : }
2141 0 : else if ((old_vmbits & VISIBILITYMAP_ALL_FROZEN) == 0 &&
2142 0 : presult.all_frozen)
2143 : {
2144 0 : vacrel->vm_new_frozen_pages++;
2145 0 : *vm_page_frozen = true;
2146 : }
2147 : }
2148 :
2149 : /*
2150 : * As of PostgreSQL 9.2, the visibility map bit should never be set if the
2151 : * page-level bit is clear. However, it's possible that the bit got
2152 : * cleared after heap_vac_scan_next_block() was called, so we must recheck
2153 : * with buffer lock before concluding that the VM is corrupt.
2154 : */
2155 1168898 : else if (all_visible_according_to_vm && !PageIsAllVisible(page) &&
2156 0 : visibilitymap_get_status(vacrel->rel, blkno, &vmbuffer) != 0)
2157 : {
2158 0 : ereport(WARNING,
2159 : (errcode(ERRCODE_DATA_CORRUPTED),
2160 : errmsg("page is not marked all-visible but visibility map bit is set in relation \"%s\" page %u",
2161 : vacrel->relname, blkno)));
2162 :
2163 0 : visibilitymap_clear(vacrel->rel, blkno, vmbuffer,
2164 : VISIBILITYMAP_VALID_BITS);
2165 : }
2166 :
2167 : /*
2168 : * It's possible for the value returned by
2169 : * GetOldestNonRemovableTransactionId() to move backwards, so it's not
2170 : * wrong for us to see tuples that appear to not be visible to everyone
2171 : * yet, while PD_ALL_VISIBLE is already set. The real safe xmin value
2172 : * never moves backwards, but GetOldestNonRemovableTransactionId() is
2173 : * conservative and sometimes returns a value that's unnecessarily small,
2174 : * so if we see that contradiction it just means that the tuples that we
2175 : * think are not visible to everyone yet actually are, and the
2176 : * PD_ALL_VISIBLE flag is correct.
2177 : *
2178 : * There should never be LP_DEAD items on a page with PD_ALL_VISIBLE set,
2179 : * however.
2180 : */
2181 1168898 : else if (presult.lpdead_items > 0 && PageIsAllVisible(page))
2182 : {
2183 0 : ereport(WARNING,
2184 : (errcode(ERRCODE_DATA_CORRUPTED),
2185 : errmsg("page containing LP_DEAD items is marked as all-visible in relation \"%s\" page %u",
2186 : vacrel->relname, blkno)));
2187 :
2188 0 : PageClearAllVisible(page);
2189 0 : MarkBufferDirty(buf);
2190 0 : visibilitymap_clear(vacrel->rel, blkno, vmbuffer,
2191 : VISIBILITYMAP_VALID_BITS);
2192 : }
2193 :
2194 : /*
2195 : * If the all-visible page is all-frozen but not marked as such yet, mark
2196 : * it as all-frozen.
2197 : */
2198 1168898 : else if (all_visible_according_to_vm && presult.all_frozen &&
2199 421628 : !VM_ALL_FROZEN(vacrel->rel, blkno, &vmbuffer))
2200 : {
2201 : uint8 old_vmbits;
2202 :
2203 : /*
2204 : * Avoid relying on all_visible_according_to_vm as a proxy for the
2205 : * page-level PD_ALL_VISIBLE bit being set, since it might have become
2206 : * stale -- even when all_visible is set
2207 : */
2208 50 : if (!PageIsAllVisible(page))
2209 : {
2210 0 : PageSetAllVisible(page);
2211 0 : MarkBufferDirty(buf);
2212 : }
2213 :
2214 : /*
2215 : * Set the page all-frozen (and all-visible) in the VM.
2216 : *
2217 : * We can pass InvalidTransactionId as our cutoff_xid, since a
2218 : * snapshotConflictHorizon sufficient to make everything safe for REDO
2219 : * was logged when the page's tuples were frozen.
2220 : */
2221 : Assert(!TransactionIdIsValid(presult.vm_conflict_horizon));
2222 50 : old_vmbits = visibilitymap_set(vacrel->rel, blkno, buf,
2223 : InvalidXLogRecPtr,
2224 : vmbuffer, InvalidTransactionId,
2225 : VISIBILITYMAP_ALL_VISIBLE |
2226 : VISIBILITYMAP_ALL_FROZEN);
2227 :
2228 : /*
2229 : * The page was likely already set all-visible in the VM. However,
2230 : * there is a small chance that it was modified sometime between
2231 : * setting all_visible_according_to_vm and checking the visibility
2232 : * during pruning. Check the return value of old_vmbits anyway to
2233 : * ensure the visibility map counters used for logging are accurate.
2234 : */
2235 50 : if ((old_vmbits & VISIBILITYMAP_ALL_VISIBLE) == 0)
2236 : {
2237 0 : vacrel->vm_new_visible_pages++;
2238 0 : vacrel->vm_new_visible_frozen_pages++;
2239 0 : *vm_page_frozen = true;
2240 : }
2241 :
2242 : /*
2243 : * We already checked that the page was not set all-frozen in the VM
2244 : * above, so we don't need to test the value of old_vmbits.
2245 : */
2246 : else
2247 : {
2248 50 : vacrel->vm_new_frozen_pages++;
2249 50 : *vm_page_frozen = true;
2250 : }
2251 : }
2252 :
2253 1243028 : return presult.ndeleted;
2254 : }
2255 :
2256 : /*
2257 : * lazy_scan_noprune() -- lazy_scan_prune() without pruning or freezing
2258 : *
2259 : * Caller need only hold a pin and share lock on the buffer, unlike
2260 : * lazy_scan_prune, which requires a full cleanup lock. While pruning isn't
2261 : * performed here, it's quite possible that an earlier opportunistic pruning
2262 : * operation left LP_DEAD items behind. We'll at least collect any such items
2263 : * in dead_items for removal from indexes.
2264 : *
2265 : * For aggressive VACUUM callers, we may return false to indicate that a full
2266 : * cleanup lock is required for processing by lazy_scan_prune. This is only
2267 : * necessary when the aggressive VACUUM needs to freeze some tuple XIDs from
2268 : * one or more tuples on the page. We always return true for non-aggressive
2269 : * callers.
2270 : *
2271 : * If this function returns true, *has_lpdead_items gets set to true or false
2272 : * depending on whether, upon return from this function, any LP_DEAD items are
2273 : * present on the page. If this function returns false, *has_lpdead_items
2274 : * is not updated.
2275 : */
2276 : static bool
2277 290 : lazy_scan_noprune(LVRelState *vacrel,
2278 : Buffer buf,
2279 : BlockNumber blkno,
2280 : Page page,
2281 : bool *has_lpdead_items)
2282 : {
2283 : OffsetNumber offnum,
2284 : maxoff;
2285 : int lpdead_items,
2286 : live_tuples,
2287 : recently_dead_tuples,
2288 : missed_dead_tuples;
2289 : bool hastup;
2290 : HeapTupleHeader tupleheader;
2291 290 : TransactionId NoFreezePageRelfrozenXid = vacrel->NewRelfrozenXid;
2292 290 : MultiXactId NoFreezePageRelminMxid = vacrel->NewRelminMxid;
2293 : OffsetNumber deadoffsets[MaxHeapTuplesPerPage];
2294 :
2295 : Assert(BufferGetBlockNumber(buf) == blkno);
2296 :
2297 290 : hastup = false; /* for now */
2298 :
2299 290 : lpdead_items = 0;
2300 290 : live_tuples = 0;
2301 290 : recently_dead_tuples = 0;
2302 290 : missed_dead_tuples = 0;
2303 :
2304 290 : maxoff = PageGetMaxOffsetNumber(page);
2305 290 : for (offnum = FirstOffsetNumber;
2306 8342 : offnum <= maxoff;
2307 8052 : offnum = OffsetNumberNext(offnum))
2308 : {
2309 : ItemId itemid;
2310 : HeapTupleData tuple;
2311 :
2312 8162 : vacrel->offnum = offnum;
2313 8162 : itemid = PageGetItemId(page, offnum);
2314 :
2315 8162 : if (!ItemIdIsUsed(itemid))
2316 1952 : continue;
2317 :
2318 6880 : if (ItemIdIsRedirected(itemid))
2319 : {
2320 670 : hastup = true;
2321 670 : continue;
2322 : }
2323 :
2324 6210 : if (ItemIdIsDead(itemid))
2325 : {
2326 : /*
2327 : * Deliberately don't set hastup=true here. See same point in
2328 : * lazy_scan_prune for an explanation.
2329 : */
2330 0 : deadoffsets[lpdead_items++] = offnum;
2331 0 : continue;
2332 : }
2333 :
2334 6210 : hastup = true; /* page prevents rel truncation */
2335 6210 : tupleheader = (HeapTupleHeader) PageGetItem(page, itemid);
2336 6210 : if (heap_tuple_should_freeze(tupleheader, &vacrel->cutoffs,
2337 : &NoFreezePageRelfrozenXid,
2338 : &NoFreezePageRelminMxid))
2339 : {
2340 : /* Tuple with XID < FreezeLimit (or MXID < MultiXactCutoff) */
2341 238 : if (vacrel->aggressive)
2342 : {
2343 : /*
2344 : * Aggressive VACUUMs must always be able to advance rel's
2345 : * relfrozenxid to a value >= FreezeLimit (and be able to
2346 : * advance rel's relminmxid to a value >= MultiXactCutoff).
2347 : * The ongoing aggressive VACUUM won't be able to do that
2348 : * unless it can freeze an XID (or MXID) from this tuple now.
2349 : *
2350 : * The only safe option is to have caller perform processing
2351 : * of this page using lazy_scan_prune. Caller might have to
2352 : * wait a while for a cleanup lock, but it can't be helped.
2353 : */
2354 110 : vacrel->offnum = InvalidOffsetNumber;
2355 110 : return false;
2356 : }
2357 :
2358 : /*
2359 : * Non-aggressive VACUUMs are under no obligation to advance
2360 : * relfrozenxid (even by one XID). We can be much laxer here.
2361 : *
2362 : * Currently we always just accept an older final relfrozenxid
2363 : * and/or relminmxid value. We never make caller wait or work a
2364 : * little harder, even when it likely makes sense to do so.
2365 : */
2366 : }
2367 :
2368 6100 : ItemPointerSet(&(tuple.t_self), blkno, offnum);
2369 6100 : tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
2370 6100 : tuple.t_len = ItemIdGetLength(itemid);
2371 6100 : tuple.t_tableOid = RelationGetRelid(vacrel->rel);
2372 :
2373 6100 : switch (HeapTupleSatisfiesVacuum(&tuple, vacrel->cutoffs.OldestXmin,
2374 : buf))
2375 : {
2376 6092 : case HEAPTUPLE_DELETE_IN_PROGRESS:
2377 : case HEAPTUPLE_LIVE:
2378 :
2379 : /*
2380 : * Count both cases as live, just like lazy_scan_prune
2381 : */
2382 6092 : live_tuples++;
2383 :
2384 6092 : break;
2385 4 : case HEAPTUPLE_DEAD:
2386 :
2387 : /*
2388 : * There is some useful work for pruning to do, that won't be
2389 : * done due to failure to get a cleanup lock.
2390 : */
2391 4 : missed_dead_tuples++;
2392 4 : break;
2393 4 : case HEAPTUPLE_RECENTLY_DEAD:
2394 :
2395 : /*
2396 : * Count in recently_dead_tuples, just like lazy_scan_prune
2397 : */
2398 4 : recently_dead_tuples++;
2399 4 : break;
2400 0 : case HEAPTUPLE_INSERT_IN_PROGRESS:
2401 :
2402 : /*
2403 : * Do not count these rows as live, just like lazy_scan_prune
2404 : */
2405 0 : break;
2406 0 : default:
2407 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
2408 : break;
2409 : }
2410 : }
2411 :
2412 180 : vacrel->offnum = InvalidOffsetNumber;
2413 :
2414 : /*
2415 : * By here we know for sure that caller can put off freezing and pruning
2416 : * this particular page until the next VACUUM. Remember its details now.
2417 : * (lazy_scan_prune expects a clean slate, so we have to do this last.)
2418 : */
2419 180 : vacrel->NewRelfrozenXid = NoFreezePageRelfrozenXid;
2420 180 : vacrel->NewRelminMxid = NoFreezePageRelminMxid;
2421 :
2422 : /* Save any LP_DEAD items found on the page in dead_items */
2423 180 : if (vacrel->nindexes == 0)
2424 : {
2425 : /* Using one-pass strategy (since table has no indexes) */
2426 0 : if (lpdead_items > 0)
2427 : {
2428 : /*
2429 : * Perfunctory handling for the corner case where a single pass
2430 : * strategy VACUUM cannot get a cleanup lock, and it turns out
2431 : * that there is one or more LP_DEAD items: just count the LP_DEAD
2432 : * items as missed_dead_tuples instead. (This is a bit dishonest,
2433 : * but it beats having to maintain specialized heap vacuuming code
2434 : * forever, for vanishingly little benefit.)
2435 : */
2436 0 : hastup = true;
2437 0 : missed_dead_tuples += lpdead_items;
2438 : }
2439 : }
2440 180 : else if (lpdead_items > 0)
2441 : {
2442 : /*
2443 : * Page has LP_DEAD items, and so any references/TIDs that remain in
2444 : * indexes will be deleted during index vacuuming (and then marked
2445 : * LP_UNUSED in the heap)
2446 : */
2447 0 : vacrel->lpdead_item_pages++;
2448 :
2449 0 : dead_items_add(vacrel, blkno, deadoffsets, lpdead_items);
2450 :
2451 0 : vacrel->lpdead_items += lpdead_items;
2452 : }
2453 :
2454 : /*
2455 : * Finally, add relevant page-local counts to whole-VACUUM counts
2456 : */
2457 180 : vacrel->live_tuples += live_tuples;
2458 180 : vacrel->recently_dead_tuples += recently_dead_tuples;
2459 180 : vacrel->missed_dead_tuples += missed_dead_tuples;
2460 180 : if (missed_dead_tuples > 0)
2461 4 : vacrel->missed_dead_pages++;
2462 :
2463 : /* Can't truncate this page */
2464 180 : if (hastup)
2465 180 : vacrel->nonempty_pages = blkno + 1;
2466 :
2467 : /* Did we find LP_DEAD items? */
2468 180 : *has_lpdead_items = (lpdead_items > 0);
2469 :
2470 : /* Caller won't need to call lazy_scan_prune with same page */
2471 180 : return true;
2472 : }
2473 :
2474 : /*
2475 : * Main entry point for index vacuuming and heap vacuuming.
2476 : *
2477 : * Removes items collected in dead_items from table's indexes, then marks the
2478 : * same items LP_UNUSED in the heap. See the comments above lazy_scan_heap
2479 : * for full details.
2480 : *
2481 : * Also empties dead_items, freeing up space for later TIDs.
2482 : *
2483 : * We may choose to bypass index vacuuming at this point, though only when the
2484 : * ongoing VACUUM operation will definitely only have one index scan/round of
2485 : * index vacuuming.
2486 : */
2487 : static void
2488 1362 : lazy_vacuum(LVRelState *vacrel)
2489 : {
2490 : bool bypass;
2491 :
2492 : /* Should not end up here with no indexes */
2493 : Assert(vacrel->nindexes > 0);
2494 : Assert(vacrel->lpdead_item_pages > 0);
2495 :
2496 1362 : if (!vacrel->do_index_vacuuming)
2497 : {
2498 : Assert(!vacrel->do_index_cleanup);
2499 26 : dead_items_reset(vacrel);
2500 26 : return;
2501 : }
2502 :
2503 : /*
2504 : * Consider bypassing index vacuuming (and heap vacuuming) entirely.
2505 : *
2506 : * We currently only do this in cases where the number of LP_DEAD items
2507 : * for the entire VACUUM operation is close to zero. This avoids sharp
2508 : * discontinuities in the duration and overhead of successive VACUUM
2509 : * operations that run against the same table with a fixed workload.
2510 : * Ideally, successive VACUUM operations will behave as if there are
2511 : * exactly zero LP_DEAD items in cases where there are close to zero.
2512 : *
2513 : * This is likely to be helpful with a table that is continually affected
2514 : * by UPDATEs that can mostly apply the HOT optimization, but occasionally
2515 : * have small aberrations that lead to just a few heap pages retaining
2516 : * only one or two LP_DEAD items. This is pretty common; even when the
2517 : * DBA goes out of their way to make UPDATEs use HOT, it is practically
2518 : * impossible to predict whether HOT will be applied in 100% of cases.
2519 : * It's far easier to ensure that 99%+ of all UPDATEs against a table use
2520 : * HOT through careful tuning.
2521 : */
2522 1336 : bypass = false;
2523 1336 : if (vacrel->consider_bypass_optimization && vacrel->rel_pages > 0)
2524 : {
2525 : BlockNumber threshold;
2526 :
2527 : Assert(vacrel->num_index_scans == 0);
2528 : Assert(vacrel->lpdead_items == vacrel->dead_items_info->num_items);
2529 : Assert(vacrel->do_index_vacuuming);
2530 : Assert(vacrel->do_index_cleanup);
2531 :
2532 : /*
2533 : * This crossover point at which we'll start to do index vacuuming is
2534 : * expressed as a percentage of the total number of heap pages in the
2535 : * table that are known to have at least one LP_DEAD item. This is
2536 : * much more important than the total number of LP_DEAD items, since
2537 : * it's a proxy for the number of heap pages whose visibility map bits
2538 : * cannot be set on account of bypassing index and heap vacuuming.
2539 : *
2540 : * We apply one further precautionary test: the space currently used
2541 : * to store the TIDs (TIDs that now all point to LP_DEAD items) must
2542 : * not exceed 32MB. This limits the risk that we will bypass index
2543 : * vacuuming again and again until eventually there is a VACUUM whose
2544 : * dead_items space is not CPU cache resident.
2545 : *
2546 : * We don't take any special steps to remember the LP_DEAD items (such
2547 : * as counting them in our final update to the stats system) when the
2548 : * optimization is applied. Though the accounting used in analyze.c's
2549 : * acquire_sample_rows() will recognize the same LP_DEAD items as dead
2550 : * rows in its own stats report, that's okay. The discrepancy should
2551 : * be negligible. If this optimization is ever expanded to cover more
2552 : * cases then this may need to be reconsidered.
2553 : */
2554 1310 : threshold = (double) vacrel->rel_pages * BYPASS_THRESHOLD_PAGES;
2555 1318 : bypass = (vacrel->lpdead_item_pages < threshold &&
2556 8 : TidStoreMemoryUsage(vacrel->dead_items) < 32 * 1024 * 1024);
2557 : }
2558 :
2559 1336 : if (bypass)
2560 : {
2561 : /*
2562 : * There are almost zero TIDs. Behave as if there were precisely
2563 : * zero: bypass index vacuuming, but do index cleanup.
2564 : *
2565 : * We expect that the ongoing VACUUM operation will finish very
2566 : * quickly, so there is no point in considering speeding up as a
2567 : * failsafe against wraparound failure. (Index cleanup is expected to
2568 : * finish very quickly in cases where there were no ambulkdelete()
2569 : * calls.)
2570 : */
2571 8 : vacrel->do_index_vacuuming = false;
2572 : }
2573 1328 : else if (lazy_vacuum_all_indexes(vacrel))
2574 : {
2575 : /*
2576 : * We successfully completed a round of index vacuuming. Do related
2577 : * heap vacuuming now.
2578 : */
2579 1328 : lazy_vacuum_heap_rel(vacrel);
2580 : }
2581 : else
2582 : {
2583 : /*
2584 : * Failsafe case.
2585 : *
2586 : * We attempted index vacuuming, but didn't finish a full round/full
2587 : * index scan. This happens when relfrozenxid or relminmxid is too
2588 : * far in the past.
2589 : *
2590 : * From this point on the VACUUM operation will do no further index
2591 : * vacuuming or heap vacuuming. This VACUUM operation won't end up
2592 : * back here again.
2593 : */
2594 : Assert(VacuumFailsafeActive);
2595 : }
2596 :
2597 : /*
2598 : * Forget the LP_DEAD items that we just vacuumed (or just decided to not
2599 : * vacuum)
2600 : */
2601 1336 : dead_items_reset(vacrel);
2602 : }
2603 :
2604 : /*
2605 : * lazy_vacuum_all_indexes() -- Main entry for index vacuuming
2606 : *
2607 : * Returns true in the common case when all indexes were successfully
2608 : * vacuumed. Returns false in rare cases where we determined that the ongoing
2609 : * VACUUM operation is at risk of taking too long to finish, leading to
2610 : * wraparound failure.
2611 : */
2612 : static bool
2613 1328 : lazy_vacuum_all_indexes(LVRelState *vacrel)
2614 : {
2615 1328 : bool allindexes = true;
2616 1328 : double old_live_tuples = vacrel->rel->rd_rel->reltuples;
2617 1328 : const int progress_start_index[] = {
2618 : PROGRESS_VACUUM_PHASE,
2619 : PROGRESS_VACUUM_INDEXES_TOTAL
2620 : };
2621 1328 : const int progress_end_index[] = {
2622 : PROGRESS_VACUUM_INDEXES_TOTAL,
2623 : PROGRESS_VACUUM_INDEXES_PROCESSED,
2624 : PROGRESS_VACUUM_NUM_INDEX_VACUUMS
2625 : };
2626 : int64 progress_start_val[2];
2627 : int64 progress_end_val[3];
2628 :
2629 : Assert(vacrel->nindexes > 0);
2630 : Assert(vacrel->do_index_vacuuming);
2631 : Assert(vacrel->do_index_cleanup);
2632 :
2633 : /* Precheck for XID wraparound emergencies */
2634 1328 : if (lazy_check_wraparound_failsafe(vacrel))
2635 : {
2636 : /* Wraparound emergency -- don't even start an index scan */
2637 0 : return false;
2638 : }
2639 :
2640 : /*
2641 : * Report that we are now vacuuming indexes and the number of indexes to
2642 : * vacuum.
2643 : */
2644 1328 : progress_start_val[0] = PROGRESS_VACUUM_PHASE_VACUUM_INDEX;
2645 1328 : progress_start_val[1] = vacrel->nindexes;
2646 1328 : pgstat_progress_update_multi_param(2, progress_start_index, progress_start_val);
2647 :
2648 1328 : if (!ParallelVacuumIsActive(vacrel))
2649 : {
2650 3868 : for (int idx = 0; idx < vacrel->nindexes; idx++)
2651 : {
2652 2552 : Relation indrel = vacrel->indrels[idx];
2653 2552 : IndexBulkDeleteResult *istat = vacrel->indstats[idx];
2654 :
2655 2552 : vacrel->indstats[idx] = lazy_vacuum_one_index(indrel, istat,
2656 : old_live_tuples,
2657 : vacrel);
2658 :
2659 : /* Report the number of indexes vacuumed */
2660 2552 : pgstat_progress_update_param(PROGRESS_VACUUM_INDEXES_PROCESSED,
2661 2552 : idx + 1);
2662 :
2663 2552 : if (lazy_check_wraparound_failsafe(vacrel))
2664 : {
2665 : /* Wraparound emergency -- end current index scan */
2666 0 : allindexes = false;
2667 0 : break;
2668 : }
2669 : }
2670 : }
2671 : else
2672 : {
2673 : /* Outsource everything to parallel variant */
2674 12 : parallel_vacuum_bulkdel_all_indexes(vacrel->pvs, old_live_tuples,
2675 : vacrel->num_index_scans);
2676 :
2677 : /*
2678 : * Do a postcheck to consider applying wraparound failsafe now. Note
2679 : * that parallel VACUUM only gets the precheck and this postcheck.
2680 : */
2681 12 : if (lazy_check_wraparound_failsafe(vacrel))
2682 0 : allindexes = false;
2683 : }
2684 :
2685 : /*
2686 : * We delete all LP_DEAD items from the first heap pass in all indexes on
2687 : * each call here (except calls where we choose to do the failsafe). This
2688 : * makes the next call to lazy_vacuum_heap_rel() safe (except in the event
2689 : * of the failsafe triggering, which prevents the next call from taking
2690 : * place).
2691 : */
2692 : Assert(vacrel->num_index_scans > 0 ||
2693 : vacrel->dead_items_info->num_items == vacrel->lpdead_items);
2694 : Assert(allindexes || VacuumFailsafeActive);
2695 :
2696 : /*
2697 : * Increase and report the number of index scans. Also, we reset
2698 : * PROGRESS_VACUUM_INDEXES_TOTAL and PROGRESS_VACUUM_INDEXES_PROCESSED.
2699 : *
2700 : * We deliberately include the case where we started a round of bulk
2701 : * deletes that we weren't able to finish due to the failsafe triggering.
2702 : */
2703 1328 : vacrel->num_index_scans++;
2704 1328 : progress_end_val[0] = 0;
2705 1328 : progress_end_val[1] = 0;
2706 1328 : progress_end_val[2] = vacrel->num_index_scans;
2707 1328 : pgstat_progress_update_multi_param(3, progress_end_index, progress_end_val);
2708 :
2709 1328 : return allindexes;
2710 : }
2711 :
2712 : /*
2713 : * Read stream callback for vacuum's third phase (second pass over the heap).
2714 : * Gets the next block from the TID store and returns it or InvalidBlockNumber
2715 : * if there are no further blocks to vacuum.
2716 : *
2717 : * NB: Assumed to be safe to use with READ_STREAM_USE_BATCHING.
2718 : */
2719 : static BlockNumber
2720 33020 : vacuum_reap_lp_read_stream_next(ReadStream *stream,
2721 : void *callback_private_data,
2722 : void *per_buffer_data)
2723 : {
2724 33020 : TidStoreIter *iter = callback_private_data;
2725 : TidStoreIterResult *iter_result;
2726 :
2727 33020 : iter_result = TidStoreIterateNext(iter);
2728 33020 : if (iter_result == NULL)
2729 1328 : return InvalidBlockNumber;
2730 :
2731 : /*
2732 : * Save the TidStoreIterResult for later, so we can extract the offsets.
2733 : * It is safe to copy the result, according to TidStoreIterateNext().
2734 : */
2735 31692 : memcpy(per_buffer_data, iter_result, sizeof(*iter_result));
2736 :
2737 31692 : return iter_result->blkno;
2738 : }
2739 :
2740 : /*
2741 : * lazy_vacuum_heap_rel() -- second pass over the heap for two pass strategy
2742 : *
2743 : * This routine marks LP_DEAD items in vacrel->dead_items as LP_UNUSED. Pages
2744 : * that never had lazy_scan_prune record LP_DEAD items are not visited at all.
2745 : *
2746 : * We may also be able to truncate the line pointer array of the heap pages we
2747 : * visit. If there is a contiguous group of LP_UNUSED items at the end of the
2748 : * array, it can be reclaimed as free space. These LP_UNUSED items usually
2749 : * start out as LP_DEAD items recorded by lazy_scan_prune (we set items from
2750 : * each page to LP_UNUSED, and then consider if it's possible to truncate the
2751 : * page's line pointer array).
2752 : *
2753 : * Note: the reason for doing this as a second pass is we cannot remove the
2754 : * tuples until we've removed their index entries, and we want to process
2755 : * index entry removal in batches as large as possible.
2756 : */
2757 : static void
2758 1328 : lazy_vacuum_heap_rel(LVRelState *vacrel)
2759 : {
2760 : ReadStream *stream;
2761 1328 : BlockNumber vacuumed_pages = 0;
2762 1328 : Buffer vmbuffer = InvalidBuffer;
2763 : LVSavedErrInfo saved_err_info;
2764 : TidStoreIter *iter;
2765 :
2766 : Assert(vacrel->do_index_vacuuming);
2767 : Assert(vacrel->do_index_cleanup);
2768 : Assert(vacrel->num_index_scans > 0);
2769 :
2770 : /* Report that we are now vacuuming the heap */
2771 1328 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
2772 : PROGRESS_VACUUM_PHASE_VACUUM_HEAP);
2773 :
2774 : /* Update error traceback information */
2775 1328 : update_vacuum_error_info(vacrel, &saved_err_info,
2776 : VACUUM_ERRCB_PHASE_VACUUM_HEAP,
2777 : InvalidBlockNumber, InvalidOffsetNumber);
2778 :
2779 1328 : iter = TidStoreBeginIterate(vacrel->dead_items);
2780 :
2781 : /*
2782 : * Set up the read stream for vacuum's second pass through the heap.
2783 : *
2784 : * It is safe to use batchmode, as vacuum_reap_lp_read_stream_next() does
2785 : * not need to wait for IO and does not perform locking. Once we support
2786 : * parallelism it should still be fine, as presumably the holder of locks
2787 : * would never be blocked by IO while holding the lock.
2788 : */
2789 1328 : stream = read_stream_begin_relation(READ_STREAM_MAINTENANCE |
2790 : READ_STREAM_USE_BATCHING,
2791 : vacrel->bstrategy,
2792 : vacrel->rel,
2793 : MAIN_FORKNUM,
2794 : vacuum_reap_lp_read_stream_next,
2795 : iter,
2796 : sizeof(TidStoreIterResult));
2797 :
2798 : while (true)
2799 31692 : {
2800 : BlockNumber blkno;
2801 : Buffer buf;
2802 : Page page;
2803 : TidStoreIterResult *iter_result;
2804 : Size freespace;
2805 : OffsetNumber offsets[MaxOffsetNumber];
2806 : int num_offsets;
2807 :
2808 33020 : vacuum_delay_point(false);
2809 :
2810 33020 : buf = read_stream_next_buffer(stream, (void **) &iter_result);
2811 :
2812 : /* The relation is exhausted */
2813 33020 : if (!BufferIsValid(buf))
2814 1328 : break;
2815 :
2816 31692 : vacrel->blkno = blkno = BufferGetBlockNumber(buf);
2817 :
2818 : Assert(iter_result);
2819 31692 : num_offsets = TidStoreGetBlockOffsets(iter_result, offsets, lengthof(offsets));
2820 : Assert(num_offsets <= lengthof(offsets));
2821 :
2822 : /*
2823 : * Pin the visibility map page in case we need to mark the page
2824 : * all-visible. In most cases this will be very cheap, because we'll
2825 : * already have the correct page pinned anyway.
2826 : */
2827 31692 : visibilitymap_pin(vacrel->rel, blkno, &vmbuffer);
2828 :
2829 : /* We need a non-cleanup exclusive lock to mark dead_items unused */
2830 31692 : LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
2831 31692 : lazy_vacuum_heap_page(vacrel, blkno, buf, offsets,
2832 : num_offsets, vmbuffer);
2833 :
2834 : /* Now that we've vacuumed the page, record its available space */
2835 31692 : page = BufferGetPage(buf);
2836 31692 : freespace = PageGetHeapFreeSpace(page);
2837 :
2838 31692 : UnlockReleaseBuffer(buf);
2839 31692 : RecordPageWithFreeSpace(vacrel->rel, blkno, freespace);
2840 31692 : vacuumed_pages++;
2841 : }
2842 :
2843 1328 : read_stream_end(stream);
2844 1328 : TidStoreEndIterate(iter);
2845 :
2846 1328 : vacrel->blkno = InvalidBlockNumber;
2847 1328 : if (BufferIsValid(vmbuffer))
2848 1328 : ReleaseBuffer(vmbuffer);
2849 :
2850 : /*
2851 : * We set all LP_DEAD items from the first heap pass to LP_UNUSED during
2852 : * the second heap pass. No more, no less.
2853 : */
2854 : Assert(vacrel->num_index_scans > 1 ||
2855 : (vacrel->dead_items_info->num_items == vacrel->lpdead_items &&
2856 : vacuumed_pages == vacrel->lpdead_item_pages));
2857 :
2858 1328 : ereport(DEBUG2,
2859 : (errmsg("table \"%s\": removed %" PRId64 " dead item identifiers in %u pages",
2860 : vacrel->relname, vacrel->dead_items_info->num_items,
2861 : vacuumed_pages)));
2862 :
2863 : /* Revert to the previous phase information for error traceback */
2864 1328 : restore_vacuum_error_info(vacrel, &saved_err_info);
2865 1328 : }
2866 :
2867 : /*
2868 : * lazy_vacuum_heap_page() -- free page's LP_DEAD items listed in the
2869 : * vacrel->dead_items store.
2870 : *
2871 : * Caller must have an exclusive buffer lock on the buffer (though a full
2872 : * cleanup lock is also acceptable). vmbuffer must be valid and already have
2873 : * a pin on blkno's visibility map page.
2874 : */
2875 : static void
2876 31692 : lazy_vacuum_heap_page(LVRelState *vacrel, BlockNumber blkno, Buffer buffer,
2877 : OffsetNumber *deadoffsets, int num_offsets,
2878 : Buffer vmbuffer)
2879 : {
2880 31692 : Page page = BufferGetPage(buffer);
2881 : OffsetNumber unused[MaxHeapTuplesPerPage];
2882 31692 : int nunused = 0;
2883 : TransactionId visibility_cutoff_xid;
2884 31692 : TransactionId conflict_xid = InvalidTransactionId;
2885 : bool all_frozen;
2886 : LVSavedErrInfo saved_err_info;
2887 31692 : uint8 vmflags = 0;
2888 :
2889 : Assert(vacrel->do_index_vacuuming);
2890 :
2891 31692 : pgstat_progress_update_param(PROGRESS_VACUUM_HEAP_BLKS_VACUUMED, blkno);
2892 :
2893 : /* Update error traceback information */
2894 31692 : update_vacuum_error_info(vacrel, &saved_err_info,
2895 : VACUUM_ERRCB_PHASE_VACUUM_HEAP, blkno,
2896 : InvalidOffsetNumber);
2897 :
2898 : /*
2899 : * Before marking dead items unused, check whether the page will become
2900 : * all-visible once that change is applied. This lets us reap the tuples
2901 : * and mark the page all-visible within the same critical section,
2902 : * enabling both changes to be emitted in a single WAL record. Since the
2903 : * visibility checks may perform I/O and allocate memory, they must be
2904 : * done outside the critical section.
2905 : */
2906 31692 : if (heap_page_would_be_all_visible(vacrel->rel, buffer,
2907 : vacrel->cutoffs.OldestXmin,
2908 : deadoffsets, num_offsets,
2909 : &all_frozen, &visibility_cutoff_xid,
2910 : &vacrel->offnum))
2911 : {
2912 31354 : vmflags |= VISIBILITYMAP_ALL_VISIBLE;
2913 31354 : if (all_frozen)
2914 : {
2915 24702 : vmflags |= VISIBILITYMAP_ALL_FROZEN;
2916 : Assert(!TransactionIdIsValid(visibility_cutoff_xid));
2917 : }
2918 :
2919 : /*
2920 : * Take the lock on the vmbuffer before entering a critical section.
2921 : * The heap page lock must also be held while updating the VM to
2922 : * ensure consistency.
2923 : */
2924 31354 : LockBuffer(vmbuffer, BUFFER_LOCK_EXCLUSIVE);
2925 : }
2926 :
2927 31692 : START_CRIT_SECTION();
2928 :
2929 1896670 : for (int i = 0; i < num_offsets; i++)
2930 : {
2931 : ItemId itemid;
2932 1864978 : OffsetNumber toff = deadoffsets[i];
2933 :
2934 1864978 : itemid = PageGetItemId(page, toff);
2935 :
2936 : Assert(ItemIdIsDead(itemid) && !ItemIdHasStorage(itemid));
2937 1864978 : ItemIdSetUnused(itemid);
2938 1864978 : unused[nunused++] = toff;
2939 : }
2940 :
2941 : Assert(nunused > 0);
2942 :
2943 : /* Attempt to truncate line pointer array now */
2944 31692 : PageTruncateLinePointerArray(page);
2945 :
2946 31692 : if ((vmflags & VISIBILITYMAP_VALID_BITS) != 0)
2947 : {
2948 : /*
2949 : * The page is guaranteed to have had dead line pointers, so we always
2950 : * set PD_ALL_VISIBLE.
2951 : */
2952 31354 : PageSetAllVisible(page);
2953 31354 : visibilitymap_set_vmbits(blkno,
2954 : vmbuffer, vmflags,
2955 31354 : vacrel->rel->rd_locator);
2956 31354 : conflict_xid = visibility_cutoff_xid;
2957 : }
2958 :
2959 : /*
2960 : * Mark buffer dirty before we write WAL.
2961 : */
2962 31692 : MarkBufferDirty(buffer);
2963 :
2964 : /* XLOG stuff */
2965 31692 : if (RelationNeedsWAL(vacrel->rel))
2966 : {
2967 29998 : log_heap_prune_and_freeze(vacrel->rel, buffer,
2968 : vmflags != 0 ? vmbuffer : InvalidBuffer,
2969 : vmflags,
2970 : conflict_xid,
2971 : false, /* no cleanup lock required */
2972 : PRUNE_VACUUM_CLEANUP,
2973 : NULL, 0, /* frozen */
2974 : NULL, 0, /* redirected */
2975 : NULL, 0, /* dead */
2976 : unused, nunused);
2977 : }
2978 :
2979 31692 : END_CRIT_SECTION();
2980 :
2981 31692 : if ((vmflags & VISIBILITYMAP_ALL_VISIBLE) != 0)
2982 : {
2983 : /* Count the newly set VM page for logging */
2984 31354 : LockBuffer(vmbuffer, BUFFER_LOCK_UNLOCK);
2985 31354 : vacrel->vm_new_visible_pages++;
2986 31354 : if (all_frozen)
2987 24702 : vacrel->vm_new_visible_frozen_pages++;
2988 : }
2989 :
2990 : /* Revert to the previous phase information for error traceback */
2991 31692 : restore_vacuum_error_info(vacrel, &saved_err_info);
2992 31692 : }
2993 :
2994 : /*
2995 : * Trigger the failsafe to avoid wraparound failure when vacrel table has a
2996 : * relfrozenxid and/or relminmxid that is dangerously far in the past.
2997 : * Triggering the failsafe makes the ongoing VACUUM bypass any further index
2998 : * vacuuming and heap vacuuming. Truncating the heap is also bypassed.
2999 : *
3000 : * Any remaining work (work that VACUUM cannot just bypass) is typically sped
3001 : * up when the failsafe triggers. VACUUM stops applying any cost-based delay
3002 : * that it started out with.
3003 : *
3004 : * Returns true when failsafe has been triggered.
3005 : */
3006 : static bool
3007 271232 : lazy_check_wraparound_failsafe(LVRelState *vacrel)
3008 : {
3009 : /* Don't warn more than once per VACUUM */
3010 271232 : if (VacuumFailsafeActive)
3011 0 : return true;
3012 :
3013 271232 : if (unlikely(vacuum_xid_failsafe_check(&vacrel->cutoffs)))
3014 : {
3015 74922 : const int progress_index[] = {
3016 : PROGRESS_VACUUM_INDEXES_TOTAL,
3017 : PROGRESS_VACUUM_INDEXES_PROCESSED,
3018 : PROGRESS_VACUUM_MODE
3019 : };
3020 74922 : int64 progress_val[3] = {0, 0, PROGRESS_VACUUM_MODE_FAILSAFE};
3021 :
3022 74922 : VacuumFailsafeActive = true;
3023 :
3024 : /*
3025 : * Abandon use of a buffer access strategy to allow use of all of
3026 : * shared buffers. We assume the caller who allocated the memory for
3027 : * the BufferAccessStrategy will free it.
3028 : */
3029 74922 : vacrel->bstrategy = NULL;
3030 :
3031 : /* Disable index vacuuming, index cleanup, and heap rel truncation */
3032 74922 : vacrel->do_index_vacuuming = false;
3033 74922 : vacrel->do_index_cleanup = false;
3034 74922 : vacrel->do_rel_truncate = false;
3035 :
3036 : /* Reset the progress counters and set the failsafe mode */
3037 74922 : pgstat_progress_update_multi_param(3, progress_index, progress_val);
3038 :
3039 74922 : ereport(WARNING,
3040 : (errmsg("bypassing nonessential maintenance of table \"%s.%s.%s\" as a failsafe after %d index scans",
3041 : vacrel->dbname, vacrel->relnamespace, vacrel->relname,
3042 : vacrel->num_index_scans),
3043 : errdetail("The table's relfrozenxid or relminmxid is too far in the past."),
3044 : errhint("Consider increasing configuration parameter \"maintenance_work_mem\" or \"autovacuum_work_mem\".\n"
3045 : "You might also need to consider other ways for VACUUM to keep up with the allocation of transaction IDs.")));
3046 :
3047 : /* Stop applying cost limits from this point on */
3048 74922 : VacuumCostActive = false;
3049 74922 : VacuumCostBalance = 0;
3050 :
3051 74922 : return true;
3052 : }
3053 :
3054 196310 : return false;
3055 : }
3056 :
3057 : /*
3058 : * lazy_cleanup_all_indexes() -- cleanup all indexes of relation.
3059 : */
3060 : static void
3061 182870 : lazy_cleanup_all_indexes(LVRelState *vacrel)
3062 : {
3063 182870 : double reltuples = vacrel->new_rel_tuples;
3064 182870 : bool estimated_count = vacrel->scanned_pages < vacrel->rel_pages;
3065 182870 : const int progress_start_index[] = {
3066 : PROGRESS_VACUUM_PHASE,
3067 : PROGRESS_VACUUM_INDEXES_TOTAL
3068 : };
3069 182870 : const int progress_end_index[] = {
3070 : PROGRESS_VACUUM_INDEXES_TOTAL,
3071 : PROGRESS_VACUUM_INDEXES_PROCESSED
3072 : };
3073 : int64 progress_start_val[2];
3074 182870 : int64 progress_end_val[2] = {0, 0};
3075 :
3076 : Assert(vacrel->do_index_cleanup);
3077 : Assert(vacrel->nindexes > 0);
3078 :
3079 : /*
3080 : * Report that we are now cleaning up indexes and the number of indexes to
3081 : * cleanup.
3082 : */
3083 182870 : progress_start_val[0] = PROGRESS_VACUUM_PHASE_INDEX_CLEANUP;
3084 182870 : progress_start_val[1] = vacrel->nindexes;
3085 182870 : pgstat_progress_update_multi_param(2, progress_start_index, progress_start_val);
3086 :
3087 182870 : if (!ParallelVacuumIsActive(vacrel))
3088 : {
3089 470120 : for (int idx = 0; idx < vacrel->nindexes; idx++)
3090 : {
3091 287284 : Relation indrel = vacrel->indrels[idx];
3092 287284 : IndexBulkDeleteResult *istat = vacrel->indstats[idx];
3093 :
3094 574568 : vacrel->indstats[idx] =
3095 287284 : lazy_cleanup_one_index(indrel, istat, reltuples,
3096 : estimated_count, vacrel);
3097 :
3098 : /* Report the number of indexes cleaned up */
3099 287284 : pgstat_progress_update_param(PROGRESS_VACUUM_INDEXES_PROCESSED,
3100 287284 : idx + 1);
3101 : }
3102 : }
3103 : else
3104 : {
3105 : /* Outsource everything to parallel variant */
3106 34 : parallel_vacuum_cleanup_all_indexes(vacrel->pvs, reltuples,
3107 : vacrel->num_index_scans,
3108 : estimated_count);
3109 : }
3110 :
3111 : /* Reset the progress counters */
3112 182870 : pgstat_progress_update_multi_param(2, progress_end_index, progress_end_val);
3113 182870 : }
3114 :
3115 : /*
3116 : * lazy_vacuum_one_index() -- vacuum index relation.
3117 : *
3118 : * Delete all the index tuples containing a TID collected in
3119 : * vacrel->dead_items. Also update running statistics. Exact
3120 : * details depend on index AM's ambulkdelete routine.
3121 : *
3122 : * reltuples is the number of heap tuples to be passed to the
3123 : * bulkdelete callback. It's always assumed to be estimated.
3124 : * See indexam.sgml for more info.
3125 : *
3126 : * Returns bulk delete stats derived from input stats
3127 : */
3128 : static IndexBulkDeleteResult *
3129 2552 : lazy_vacuum_one_index(Relation indrel, IndexBulkDeleteResult *istat,
3130 : double reltuples, LVRelState *vacrel)
3131 : {
3132 : IndexVacuumInfo ivinfo;
3133 : LVSavedErrInfo saved_err_info;
3134 :
3135 2552 : ivinfo.index = indrel;
3136 2552 : ivinfo.heaprel = vacrel->rel;
3137 2552 : ivinfo.analyze_only = false;
3138 2552 : ivinfo.report_progress = false;
3139 2552 : ivinfo.estimated_count = true;
3140 2552 : ivinfo.message_level = DEBUG2;
3141 2552 : ivinfo.num_heap_tuples = reltuples;
3142 2552 : ivinfo.strategy = vacrel->bstrategy;
3143 :
3144 : /*
3145 : * Update error traceback information.
3146 : *
3147 : * The index name is saved during this phase and restored immediately
3148 : * after this phase. See vacuum_error_callback.
3149 : */
3150 : Assert(vacrel->indname == NULL);
3151 2552 : vacrel->indname = pstrdup(RelationGetRelationName(indrel));
3152 2552 : update_vacuum_error_info(vacrel, &saved_err_info,
3153 : VACUUM_ERRCB_PHASE_VACUUM_INDEX,
3154 : InvalidBlockNumber, InvalidOffsetNumber);
3155 :
3156 : /* Do bulk deletion */
3157 2552 : istat = vac_bulkdel_one_index(&ivinfo, istat, vacrel->dead_items,
3158 : vacrel->dead_items_info);
3159 :
3160 : /* Revert to the previous phase information for error traceback */
3161 2552 : restore_vacuum_error_info(vacrel, &saved_err_info);
3162 2552 : pfree(vacrel->indname);
3163 2552 : vacrel->indname = NULL;
3164 :
3165 2552 : return istat;
3166 : }
3167 :
3168 : /*
3169 : * lazy_cleanup_one_index() -- do post-vacuum cleanup for index relation.
3170 : *
3171 : * Calls index AM's amvacuumcleanup routine. reltuples is the number
3172 : * of heap tuples and estimated_count is true if reltuples is an
3173 : * estimated value. See indexam.sgml for more info.
3174 : *
3175 : * Returns bulk delete stats derived from input stats
3176 : */
3177 : static IndexBulkDeleteResult *
3178 287284 : lazy_cleanup_one_index(Relation indrel, IndexBulkDeleteResult *istat,
3179 : double reltuples, bool estimated_count,
3180 : LVRelState *vacrel)
3181 : {
3182 : IndexVacuumInfo ivinfo;
3183 : LVSavedErrInfo saved_err_info;
3184 :
3185 287284 : ivinfo.index = indrel;
3186 287284 : ivinfo.heaprel = vacrel->rel;
3187 287284 : ivinfo.analyze_only = false;
3188 287284 : ivinfo.report_progress = false;
3189 287284 : ivinfo.estimated_count = estimated_count;
3190 287284 : ivinfo.message_level = DEBUG2;
3191 :
3192 287284 : ivinfo.num_heap_tuples = reltuples;
3193 287284 : ivinfo.strategy = vacrel->bstrategy;
3194 :
3195 : /*
3196 : * Update error traceback information.
3197 : *
3198 : * The index name is saved during this phase and restored immediately
3199 : * after this phase. See vacuum_error_callback.
3200 : */
3201 : Assert(vacrel->indname == NULL);
3202 287284 : vacrel->indname = pstrdup(RelationGetRelationName(indrel));
3203 287284 : update_vacuum_error_info(vacrel, &saved_err_info,
3204 : VACUUM_ERRCB_PHASE_INDEX_CLEANUP,
3205 : InvalidBlockNumber, InvalidOffsetNumber);
3206 :
3207 287284 : istat = vac_cleanup_one_index(&ivinfo, istat);
3208 :
3209 : /* Revert to the previous phase information for error traceback */
3210 287284 : restore_vacuum_error_info(vacrel, &saved_err_info);
3211 287284 : pfree(vacrel->indname);
3212 287284 : vacrel->indname = NULL;
3213 :
3214 287284 : return istat;
3215 : }
3216 :
3217 : /*
3218 : * should_attempt_truncation - should we attempt to truncate the heap?
3219 : *
3220 : * Don't even think about it unless we have a shot at releasing a goodly
3221 : * number of pages. Otherwise, the time taken isn't worth it, mainly because
3222 : * an AccessExclusive lock must be replayed on any hot standby, where it can
3223 : * be particularly disruptive.
3224 : *
3225 : * Also don't attempt it if wraparound failsafe is in effect. The entire
3226 : * system might be refusing to allocate new XIDs at this point. The system
3227 : * definitely won't return to normal unless and until VACUUM actually advances
3228 : * the oldest relfrozenxid -- which hasn't happened for target rel just yet.
3229 : * If lazy_truncate_heap attempted to acquire an AccessExclusiveLock to
3230 : * truncate the table under these circumstances, an XID exhaustion error might
3231 : * make it impossible for VACUUM to fix the underlying XID exhaustion problem.
3232 : * There is very little chance of truncation working out when the failsafe is
3233 : * in effect in any case. lazy_scan_prune makes the optimistic assumption
3234 : * that any LP_DEAD items it encounters will always be LP_UNUSED by the time
3235 : * we're called.
3236 : */
3237 : static bool
3238 267340 : should_attempt_truncation(LVRelState *vacrel)
3239 : {
3240 : BlockNumber possibly_freeable;
3241 :
3242 267340 : if (!vacrel->do_rel_truncate || VacuumFailsafeActive)
3243 75212 : return false;
3244 :
3245 192128 : possibly_freeable = vacrel->rel_pages - vacrel->nonempty_pages;
3246 192128 : if (possibly_freeable > 0 &&
3247 338 : (possibly_freeable >= REL_TRUNCATE_MINIMUM ||
3248 338 : possibly_freeable >= vacrel->rel_pages / REL_TRUNCATE_FRACTION))
3249 316 : return true;
3250 :
3251 191812 : return false;
3252 : }
3253 :
3254 : /*
3255 : * lazy_truncate_heap - try to truncate off any empty pages at the end
3256 : */
3257 : static void
3258 316 : lazy_truncate_heap(LVRelState *vacrel)
3259 : {
3260 316 : BlockNumber orig_rel_pages = vacrel->rel_pages;
3261 : BlockNumber new_rel_pages;
3262 : bool lock_waiter_detected;
3263 : int lock_retry;
3264 :
3265 : /* Report that we are now truncating */
3266 316 : pgstat_progress_update_param(PROGRESS_VACUUM_PHASE,
3267 : PROGRESS_VACUUM_PHASE_TRUNCATE);
3268 :
3269 : /* Update error traceback information one last time */
3270 316 : update_vacuum_error_info(vacrel, NULL, VACUUM_ERRCB_PHASE_TRUNCATE,
3271 : vacrel->nonempty_pages, InvalidOffsetNumber);
3272 :
3273 : /*
3274 : * Loop until no more truncating can be done.
3275 : */
3276 : do
3277 : {
3278 : /*
3279 : * We need full exclusive lock on the relation in order to do
3280 : * truncation. If we can't get it, give up rather than waiting --- we
3281 : * don't want to block other backends, and we don't want to deadlock
3282 : * (which is quite possible considering we already hold a lower-grade
3283 : * lock).
3284 : */
3285 316 : lock_waiter_detected = false;
3286 316 : lock_retry = 0;
3287 : while (true)
3288 : {
3289 716 : if (ConditionalLockRelation(vacrel->rel, AccessExclusiveLock))
3290 312 : break;
3291 :
3292 : /*
3293 : * Check for interrupts while trying to (re-)acquire the exclusive
3294 : * lock.
3295 : */
3296 404 : CHECK_FOR_INTERRUPTS();
3297 :
3298 404 : if (++lock_retry > (VACUUM_TRUNCATE_LOCK_TIMEOUT /
3299 : VACUUM_TRUNCATE_LOCK_WAIT_INTERVAL))
3300 : {
3301 : /*
3302 : * We failed to establish the lock in the specified number of
3303 : * retries. This means we give up truncating.
3304 : */
3305 4 : ereport(vacrel->verbose ? INFO : DEBUG2,
3306 : (errmsg("\"%s\": stopping truncate due to conflicting lock request",
3307 : vacrel->relname)));
3308 6 : return;
3309 : }
3310 :
3311 400 : (void) WaitLatch(MyLatch,
3312 : WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
3313 : VACUUM_TRUNCATE_LOCK_WAIT_INTERVAL,
3314 : WAIT_EVENT_VACUUM_TRUNCATE);
3315 400 : ResetLatch(MyLatch);
3316 : }
3317 :
3318 : /*
3319 : * Now that we have exclusive lock, look to see if the rel has grown
3320 : * whilst we were vacuuming with non-exclusive lock. If so, give up;
3321 : * the newly added pages presumably contain non-deletable tuples.
3322 : */
3323 312 : new_rel_pages = RelationGetNumberOfBlocks(vacrel->rel);
3324 312 : if (new_rel_pages != orig_rel_pages)
3325 : {
3326 : /*
3327 : * Note: we intentionally don't update vacrel->rel_pages with the
3328 : * new rel size here. If we did, it would amount to assuming that
3329 : * the new pages are empty, which is unlikely. Leaving the numbers
3330 : * alone amounts to assuming that the new pages have the same
3331 : * tuple density as existing ones, which is less unlikely.
3332 : */
3333 0 : UnlockRelation(vacrel->rel, AccessExclusiveLock);
3334 0 : return;
3335 : }
3336 :
3337 : /*
3338 : * Scan backwards from the end to verify that the end pages actually
3339 : * contain no tuples. This is *necessary*, not optional, because
3340 : * other backends could have added tuples to these pages whilst we
3341 : * were vacuuming.
3342 : */
3343 312 : new_rel_pages = count_nondeletable_pages(vacrel, &lock_waiter_detected);
3344 312 : vacrel->blkno = new_rel_pages;
3345 :
3346 312 : if (new_rel_pages >= orig_rel_pages)
3347 : {
3348 : /* can't do anything after all */
3349 2 : UnlockRelation(vacrel->rel, AccessExclusiveLock);
3350 2 : return;
3351 : }
3352 :
3353 : /*
3354 : * Okay to truncate.
3355 : */
3356 310 : RelationTruncate(vacrel->rel, new_rel_pages);
3357 :
3358 : /*
3359 : * We can release the exclusive lock as soon as we have truncated.
3360 : * Other backends can't safely access the relation until they have
3361 : * processed the smgr invalidation that smgrtruncate sent out ... but
3362 : * that should happen as part of standard invalidation processing once
3363 : * they acquire lock on the relation.
3364 : */
3365 310 : UnlockRelation(vacrel->rel, AccessExclusiveLock);
3366 :
3367 : /*
3368 : * Update statistics. Here, it *is* correct to adjust rel_pages
3369 : * without also touching reltuples, since the tuple count wasn't
3370 : * changed by the truncation.
3371 : */
3372 310 : vacrel->removed_pages += orig_rel_pages - new_rel_pages;
3373 310 : vacrel->rel_pages = new_rel_pages;
3374 :
3375 310 : ereport(vacrel->verbose ? INFO : DEBUG2,
3376 : (errmsg("table \"%s\": truncated %u to %u pages",
3377 : vacrel->relname,
3378 : orig_rel_pages, new_rel_pages)));
3379 310 : orig_rel_pages = new_rel_pages;
3380 310 : } while (new_rel_pages > vacrel->nonempty_pages && lock_waiter_detected);
3381 : }
3382 :
3383 : /*
3384 : * Rescan end pages to verify that they are (still) empty of tuples.
3385 : *
3386 : * Returns number of nondeletable pages (last nonempty page + 1).
3387 : */
3388 : static BlockNumber
3389 312 : count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected)
3390 : {
3391 : StaticAssertDecl((PREFETCH_SIZE & (PREFETCH_SIZE - 1)) == 0,
3392 : "prefetch size must be power of 2");
3393 :
3394 : BlockNumber blkno;
3395 : BlockNumber prefetchedUntil;
3396 : instr_time starttime;
3397 :
3398 : /* Initialize the starttime if we check for conflicting lock requests */
3399 312 : INSTR_TIME_SET_CURRENT(starttime);
3400 :
3401 : /*
3402 : * Start checking blocks at what we believe relation end to be and move
3403 : * backwards. (Strange coding of loop control is needed because blkno is
3404 : * unsigned.) To make the scan faster, we prefetch a few blocks at a time
3405 : * in forward direction, so that OS-level readahead can kick in.
3406 : */
3407 312 : blkno = vacrel->rel_pages;
3408 312 : prefetchedUntil = InvalidBlockNumber;
3409 5364 : while (blkno > vacrel->nonempty_pages)
3410 : {
3411 : Buffer buf;
3412 : Page page;
3413 : OffsetNumber offnum,
3414 : maxoff;
3415 : bool hastup;
3416 :
3417 : /*
3418 : * Check if another process requests a lock on our relation. We are
3419 : * holding an AccessExclusiveLock here, so they will be waiting. We
3420 : * only do this once per VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL, and we
3421 : * only check if that interval has elapsed once every 32 blocks to
3422 : * keep the number of system calls and actual shared lock table
3423 : * lookups to a minimum.
3424 : */
3425 5060 : if ((blkno % 32) == 0)
3426 : {
3427 : instr_time currenttime;
3428 : instr_time elapsed;
3429 :
3430 168 : INSTR_TIME_SET_CURRENT(currenttime);
3431 168 : elapsed = currenttime;
3432 168 : INSTR_TIME_SUBTRACT(elapsed, starttime);
3433 168 : if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000)
3434 : >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL)
3435 : {
3436 0 : if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock))
3437 : {
3438 0 : ereport(vacrel->verbose ? INFO : DEBUG2,
3439 : (errmsg("table \"%s\": suspending truncate due to conflicting lock request",
3440 : vacrel->relname)));
3441 :
3442 0 : *lock_waiter_detected = true;
3443 0 : return blkno;
3444 : }
3445 0 : starttime = currenttime;
3446 : }
3447 : }
3448 :
3449 : /*
3450 : * We don't insert a vacuum delay point here, because we have an
3451 : * exclusive lock on the table which we want to hold for as short a
3452 : * time as possible. We still need to check for interrupts however.
3453 : */
3454 5060 : CHECK_FOR_INTERRUPTS();
3455 :
3456 5060 : blkno--;
3457 :
3458 : /* If we haven't prefetched this lot yet, do so now. */
3459 5060 : if (prefetchedUntil > blkno)
3460 : {
3461 : BlockNumber prefetchStart;
3462 : BlockNumber pblkno;
3463 :
3464 432 : prefetchStart = blkno & ~(PREFETCH_SIZE - 1);
3465 7494 : for (pblkno = prefetchStart; pblkno <= blkno; pblkno++)
3466 : {
3467 7062 : PrefetchBuffer(vacrel->rel, MAIN_FORKNUM, pblkno);
3468 7062 : CHECK_FOR_INTERRUPTS();
3469 : }
3470 432 : prefetchedUntil = prefetchStart;
3471 : }
3472 :
3473 5060 : buf = ReadBufferExtended(vacrel->rel, MAIN_FORKNUM, blkno, RBM_NORMAL,
3474 : vacrel->bstrategy);
3475 :
3476 : /* In this phase we only need shared access to the buffer */
3477 5060 : LockBuffer(buf, BUFFER_LOCK_SHARE);
3478 :
3479 5060 : page = BufferGetPage(buf);
3480 :
3481 5060 : if (PageIsNew(page) || PageIsEmpty(page))
3482 : {
3483 2382 : UnlockReleaseBuffer(buf);
3484 2382 : continue;
3485 : }
3486 :
3487 2678 : hastup = false;
3488 2678 : maxoff = PageGetMaxOffsetNumber(page);
3489 2678 : for (offnum = FirstOffsetNumber;
3490 5954 : offnum <= maxoff;
3491 3276 : offnum = OffsetNumberNext(offnum))
3492 : {
3493 : ItemId itemid;
3494 :
3495 3284 : itemid = PageGetItemId(page, offnum);
3496 :
3497 : /*
3498 : * Note: any non-unused item should be taken as a reason to keep
3499 : * this page. Even an LP_DEAD item makes truncation unsafe, since
3500 : * we must not have cleaned out its index entries.
3501 : */
3502 3284 : if (ItemIdIsUsed(itemid))
3503 : {
3504 8 : hastup = true;
3505 8 : break; /* can stop scanning */
3506 : }
3507 : } /* scan along page */
3508 :
3509 2678 : UnlockReleaseBuffer(buf);
3510 :
3511 : /* Done scanning if we found a tuple here */
3512 2678 : if (hastup)
3513 8 : return blkno + 1;
3514 : }
3515 :
3516 : /*
3517 : * If we fall out of the loop, all the previously-thought-to-be-empty
3518 : * pages still are; we need not bother to look at the last known-nonempty
3519 : * page.
3520 : */
3521 304 : return vacrel->nonempty_pages;
3522 : }
3523 :
3524 : /*
3525 : * Allocate dead_items and dead_items_info (either using palloc, or in dynamic
3526 : * shared memory). Sets both in vacrel for caller.
3527 : *
3528 : * Also handles parallel initialization as part of allocating dead_items in
3529 : * DSM when required.
3530 : */
3531 : static void
3532 267340 : dead_items_alloc(LVRelState *vacrel, int nworkers)
3533 : {
3534 : VacDeadItemsInfo *dead_items_info;
3535 774532 : int vac_work_mem = AmAutoVacuumWorkerProcess() &&
3536 239852 : autovacuum_work_mem != -1 ?
3537 507192 : autovacuum_work_mem : maintenance_work_mem;
3538 :
3539 : /*
3540 : * Initialize state for a parallel vacuum. As of now, only one worker can
3541 : * be used for an index, so we invoke parallelism only if there are at
3542 : * least two indexes on a table.
3543 : */
3544 267340 : if (nworkers >= 0 && vacrel->nindexes > 1 && vacrel->do_index_vacuuming)
3545 : {
3546 : /*
3547 : * Since parallel workers cannot access data in temporary tables, we
3548 : * can't perform parallel vacuum on them.
3549 : */
3550 10970 : if (RelationUsesLocalBuffers(vacrel->rel))
3551 : {
3552 : /*
3553 : * Give warning only if the user explicitly tries to perform a
3554 : * parallel vacuum on the temporary table.
3555 : */
3556 6 : if (nworkers > 0)
3557 6 : ereport(WARNING,
3558 : (errmsg("disabling parallel option of vacuum on \"%s\" --- cannot vacuum temporary tables in parallel",
3559 : vacrel->relname)));
3560 : }
3561 : else
3562 10964 : vacrel->pvs = parallel_vacuum_init(vacrel->rel, vacrel->indrels,
3563 : vacrel->nindexes, nworkers,
3564 : vac_work_mem,
3565 10964 : vacrel->verbose ? INFO : DEBUG2,
3566 : vacrel->bstrategy);
3567 :
3568 : /*
3569 : * If parallel mode started, dead_items and dead_items_info spaces are
3570 : * allocated in DSM.
3571 : */
3572 10970 : if (ParallelVacuumIsActive(vacrel))
3573 : {
3574 34 : vacrel->dead_items = parallel_vacuum_get_dead_items(vacrel->pvs,
3575 : &vacrel->dead_items_info);
3576 34 : return;
3577 : }
3578 : }
3579 :
3580 : /*
3581 : * Serial VACUUM case. Allocate both dead_items and dead_items_info
3582 : * locally.
3583 : */
3584 :
3585 267306 : dead_items_info = palloc_object(VacDeadItemsInfo);
3586 267306 : dead_items_info->max_bytes = vac_work_mem * (Size) 1024;
3587 267306 : dead_items_info->num_items = 0;
3588 267306 : vacrel->dead_items_info = dead_items_info;
3589 :
3590 267306 : vacrel->dead_items = TidStoreCreateLocal(dead_items_info->max_bytes, true);
3591 : }
3592 :
3593 : /*
3594 : * Add the given block number and offset numbers to dead_items.
3595 : */
3596 : static void
3597 36304 : dead_items_add(LVRelState *vacrel, BlockNumber blkno, OffsetNumber *offsets,
3598 : int num_offsets)
3599 : {
3600 36304 : const int prog_index[2] = {
3601 : PROGRESS_VACUUM_NUM_DEAD_ITEM_IDS,
3602 : PROGRESS_VACUUM_DEAD_TUPLE_BYTES
3603 : };
3604 : int64 prog_val[2];
3605 :
3606 36304 : TidStoreSetBlockOffsets(vacrel->dead_items, blkno, offsets, num_offsets);
3607 36304 : vacrel->dead_items_info->num_items += num_offsets;
3608 :
3609 : /* update the progress information */
3610 36304 : prog_val[0] = vacrel->dead_items_info->num_items;
3611 36304 : prog_val[1] = TidStoreMemoryUsage(vacrel->dead_items);
3612 36304 : pgstat_progress_update_multi_param(2, prog_index, prog_val);
3613 36304 : }
3614 :
3615 : /*
3616 : * Forget all collected dead items.
3617 : */
3618 : static void
3619 1362 : dead_items_reset(LVRelState *vacrel)
3620 : {
3621 1362 : if (ParallelVacuumIsActive(vacrel))
3622 : {
3623 12 : parallel_vacuum_reset_dead_items(vacrel->pvs);
3624 12 : vacrel->dead_items = parallel_vacuum_get_dead_items(vacrel->pvs,
3625 : &vacrel->dead_items_info);
3626 12 : return;
3627 : }
3628 :
3629 : /* Recreate the tidstore with the same max_bytes limitation */
3630 1350 : TidStoreDestroy(vacrel->dead_items);
3631 1350 : vacrel->dead_items = TidStoreCreateLocal(vacrel->dead_items_info->max_bytes, true);
3632 :
3633 : /* Reset the counter */
3634 1350 : vacrel->dead_items_info->num_items = 0;
3635 : }
3636 :
3637 : /*
3638 : * Perform cleanup for resources allocated in dead_items_alloc
3639 : */
3640 : static void
3641 267340 : dead_items_cleanup(LVRelState *vacrel)
3642 : {
3643 267340 : if (!ParallelVacuumIsActive(vacrel))
3644 : {
3645 : /* Don't bother with pfree here */
3646 267306 : return;
3647 : }
3648 :
3649 : /* End parallel mode */
3650 34 : parallel_vacuum_end(vacrel->pvs, vacrel->indstats);
3651 34 : vacrel->pvs = NULL;
3652 : }
3653 :
3654 : #ifdef USE_ASSERT_CHECKING
3655 :
3656 : /*
3657 : * Wrapper for heap_page_would_be_all_visible() which can be used for callers
3658 : * that expect no LP_DEAD on the page. Currently assert-only, but there is no
3659 : * reason not to use it outside of asserts.
3660 : */
3661 : static bool
3662 : heap_page_is_all_visible(Relation rel, Buffer buf,
3663 : TransactionId OldestXmin,
3664 : bool *all_frozen,
3665 : TransactionId *visibility_cutoff_xid,
3666 : OffsetNumber *logging_offnum)
3667 : {
3668 :
3669 : return heap_page_would_be_all_visible(rel, buf,
3670 : OldestXmin,
3671 : NULL, 0,
3672 : all_frozen,
3673 : visibility_cutoff_xid,
3674 : logging_offnum);
3675 : }
3676 : #endif
3677 :
3678 : /*
3679 : * Check whether the heap page in buf is all-visible except for the dead
3680 : * tuples referenced in the deadoffsets array.
3681 : *
3682 : * Vacuum uses this to check if a page would become all-visible after reaping
3683 : * known dead tuples. This function does not remove the dead items.
3684 : *
3685 : * This cannot be called in a critical section, as the visibility checks may
3686 : * perform IO and allocate memory.
3687 : *
3688 : * Returns true if the page is all-visible other than the provided
3689 : * deadoffsets and false otherwise.
3690 : *
3691 : * OldestXmin is used to determine visibility.
3692 : *
3693 : * Output parameters:
3694 : *
3695 : * - *all_frozen: true if every tuple on the page is frozen
3696 : * - *visibility_cutoff_xid: newest xmin; valid only if page is all-visible
3697 : * - *logging_offnum: OffsetNumber of current tuple being processed;
3698 : * used by vacuum's error callback system.
3699 : *
3700 : * Callers looking to verify that the page is already all-visible can call
3701 : * heap_page_is_all_visible().
3702 : *
3703 : * This logic is closely related to heap_prune_record_unchanged_lp_normal().
3704 : * If you modify this function, ensure consistency with that code. An
3705 : * assertion cross-checks that both remain in agreement. Do not introduce new
3706 : * side-effects.
3707 : */
3708 : static bool
3709 31692 : heap_page_would_be_all_visible(Relation rel, Buffer buf,
3710 : TransactionId OldestXmin,
3711 : OffsetNumber *deadoffsets,
3712 : int ndeadoffsets,
3713 : bool *all_frozen,
3714 : TransactionId *visibility_cutoff_xid,
3715 : OffsetNumber *logging_offnum)
3716 : {
3717 31692 : Page page = BufferGetPage(buf);
3718 31692 : BlockNumber blockno = BufferGetBlockNumber(buf);
3719 : OffsetNumber offnum,
3720 : maxoff;
3721 31692 : bool all_visible = true;
3722 31692 : int matched_dead_count = 0;
3723 :
3724 31692 : *visibility_cutoff_xid = InvalidTransactionId;
3725 31692 : *all_frozen = true;
3726 :
3727 : Assert(ndeadoffsets == 0 || deadoffsets);
3728 :
3729 : #ifdef USE_ASSERT_CHECKING
3730 : /* Confirm input deadoffsets[] is strictly sorted */
3731 : if (ndeadoffsets > 1)
3732 : {
3733 : for (int i = 1; i < ndeadoffsets; i++)
3734 : Assert(deadoffsets[i - 1] < deadoffsets[i]);
3735 : }
3736 : #endif
3737 :
3738 31692 : maxoff = PageGetMaxOffsetNumber(page);
3739 31692 : for (offnum = FirstOffsetNumber;
3740 2926832 : offnum <= maxoff && all_visible;
3741 2895140 : offnum = OffsetNumberNext(offnum))
3742 : {
3743 : ItemId itemid;
3744 : HeapTupleData tuple;
3745 :
3746 : /*
3747 : * Set the offset number so that we can display it along with any
3748 : * error that occurred while processing this tuple.
3749 : */
3750 2895142 : *logging_offnum = offnum;
3751 2895142 : itemid = PageGetItemId(page, offnum);
3752 :
3753 : /* Unused or redirect line pointers are of no interest */
3754 2895142 : if (!ItemIdIsUsed(itemid) || ItemIdIsRedirected(itemid))
3755 1940516 : continue;
3756 :
3757 2814352 : ItemPointerSet(&(tuple.t_self), blockno, offnum);
3758 :
3759 : /*
3760 : * Dead line pointers can have index pointers pointing to them. So
3761 : * they can't be treated as visible
3762 : */
3763 2814352 : if (ItemIdIsDead(itemid))
3764 : {
3765 1859728 : if (!deadoffsets ||
3766 1859728 : matched_dead_count >= ndeadoffsets ||
3767 1859728 : deadoffsets[matched_dead_count] != offnum)
3768 : {
3769 2 : *all_frozen = all_visible = false;
3770 2 : break;
3771 : }
3772 1859726 : matched_dead_count++;
3773 1859726 : continue;
3774 : }
3775 :
3776 : Assert(ItemIdIsNormal(itemid));
3777 :
3778 954624 : tuple.t_data = (HeapTupleHeader) PageGetItem(page, itemid);
3779 954624 : tuple.t_len = ItemIdGetLength(itemid);
3780 954624 : tuple.t_tableOid = RelationGetRelid(rel);
3781 :
3782 : /* Visibility checks may do IO or allocate memory */
3783 : Assert(CritSectionCount == 0);
3784 954624 : switch (HeapTupleSatisfiesVacuum(&tuple, OldestXmin, buf))
3785 : {
3786 954400 : case HEAPTUPLE_LIVE:
3787 : {
3788 : TransactionId xmin;
3789 :
3790 : /* Check comments in lazy_scan_prune. */
3791 954400 : if (!HeapTupleHeaderXminCommitted(tuple.t_data))
3792 : {
3793 0 : all_visible = false;
3794 0 : *all_frozen = false;
3795 0 : break;
3796 : }
3797 :
3798 : /*
3799 : * The inserter definitely committed. But is it old enough
3800 : * that everyone sees it as committed?
3801 : */
3802 954400 : xmin = HeapTupleHeaderGetXmin(tuple.t_data);
3803 954400 : if (!TransactionIdPrecedes(xmin, OldestXmin))
3804 : {
3805 112 : all_visible = false;
3806 112 : *all_frozen = false;
3807 112 : break;
3808 : }
3809 :
3810 : /* Track newest xmin on page. */
3811 954288 : if (TransactionIdFollows(xmin, *visibility_cutoff_xid) &&
3812 : TransactionIdIsNormal(xmin))
3813 26186 : *visibility_cutoff_xid = xmin;
3814 :
3815 : /* Check whether this tuple is already frozen or not */
3816 1250332 : if (all_visible && *all_frozen &&
3817 296044 : heap_tuple_needs_eventual_freeze(tuple.t_data))
3818 6696 : *all_frozen = false;
3819 : }
3820 954288 : break;
3821 :
3822 224 : case HEAPTUPLE_DEAD:
3823 : case HEAPTUPLE_RECENTLY_DEAD:
3824 : case HEAPTUPLE_INSERT_IN_PROGRESS:
3825 : case HEAPTUPLE_DELETE_IN_PROGRESS:
3826 : {
3827 224 : all_visible = false;
3828 224 : *all_frozen = false;
3829 224 : break;
3830 : }
3831 0 : default:
3832 0 : elog(ERROR, "unexpected HeapTupleSatisfiesVacuum result");
3833 : break;
3834 : }
3835 : } /* scan along page */
3836 :
3837 : /* Clear the offset information once we have processed the given page. */
3838 31692 : *logging_offnum = InvalidOffsetNumber;
3839 :
3840 31692 : return all_visible;
3841 : }
3842 :
3843 : /*
3844 : * Update index statistics in pg_class if the statistics are accurate.
3845 : */
3846 : static void
3847 192158 : update_relstats_all_indexes(LVRelState *vacrel)
3848 : {
3849 192158 : Relation *indrels = vacrel->indrels;
3850 192158 : int nindexes = vacrel->nindexes;
3851 192158 : IndexBulkDeleteResult **indstats = vacrel->indstats;
3852 :
3853 : Assert(vacrel->do_index_cleanup);
3854 :
3855 479564 : for (int idx = 0; idx < nindexes; idx++)
3856 : {
3857 287406 : Relation indrel = indrels[idx];
3858 287406 : IndexBulkDeleteResult *istat = indstats[idx];
3859 :
3860 287406 : if (istat == NULL || istat->estimated_count)
3861 284558 : continue;
3862 :
3863 : /* Update index statistics */
3864 2848 : vac_update_relstats(indrel,
3865 : istat->num_pages,
3866 : istat->num_index_tuples,
3867 : 0, 0,
3868 : false,
3869 : InvalidTransactionId,
3870 : InvalidMultiXactId,
3871 : NULL, NULL, false);
3872 : }
3873 192158 : }
3874 :
3875 : /*
3876 : * Error context callback for errors occurring during vacuum. The error
3877 : * context messages for index phases should match the messages set in parallel
3878 : * vacuum. If you change this function for those phases, change
3879 : * parallel_vacuum_error_callback() as well.
3880 : */
3881 : static void
3882 280030 : vacuum_error_callback(void *arg)
3883 : {
3884 280030 : LVRelState *errinfo = arg;
3885 :
3886 280030 : switch (errinfo->phase)
3887 : {
3888 0 : case VACUUM_ERRCB_PHASE_SCAN_HEAP:
3889 0 : if (BlockNumberIsValid(errinfo->blkno))
3890 : {
3891 0 : if (OffsetNumberIsValid(errinfo->offnum))
3892 0 : errcontext("while scanning block %u offset %u of relation \"%s.%s\"",
3893 0 : errinfo->blkno, errinfo->offnum, errinfo->relnamespace, errinfo->relname);
3894 : else
3895 0 : errcontext("while scanning block %u of relation \"%s.%s\"",
3896 : errinfo->blkno, errinfo->relnamespace, errinfo->relname);
3897 : }
3898 : else
3899 0 : errcontext("while scanning relation \"%s.%s\"",
3900 : errinfo->relnamespace, errinfo->relname);
3901 0 : break;
3902 :
3903 0 : case VACUUM_ERRCB_PHASE_VACUUM_HEAP:
3904 0 : if (BlockNumberIsValid(errinfo->blkno))
3905 : {
3906 0 : if (OffsetNumberIsValid(errinfo->offnum))
3907 0 : errcontext("while vacuuming block %u offset %u of relation \"%s.%s\"",
3908 0 : errinfo->blkno, errinfo->offnum, errinfo->relnamespace, errinfo->relname);
3909 : else
3910 0 : errcontext("while vacuuming block %u of relation \"%s.%s\"",
3911 : errinfo->blkno, errinfo->relnamespace, errinfo->relname);
3912 : }
3913 : else
3914 0 : errcontext("while vacuuming relation \"%s.%s\"",
3915 : errinfo->relnamespace, errinfo->relname);
3916 0 : break;
3917 :
3918 0 : case VACUUM_ERRCB_PHASE_VACUUM_INDEX:
3919 0 : errcontext("while vacuuming index \"%s\" of relation \"%s.%s\"",
3920 : errinfo->indname, errinfo->relnamespace, errinfo->relname);
3921 0 : break;
3922 :
3923 0 : case VACUUM_ERRCB_PHASE_INDEX_CLEANUP:
3924 0 : errcontext("while cleaning up index \"%s\" of relation \"%s.%s\"",
3925 : errinfo->indname, errinfo->relnamespace, errinfo->relname);
3926 0 : break;
3927 :
3928 6 : case VACUUM_ERRCB_PHASE_TRUNCATE:
3929 6 : if (BlockNumberIsValid(errinfo->blkno))
3930 6 : errcontext("while truncating relation \"%s.%s\" to %u blocks",
3931 : errinfo->relnamespace, errinfo->relname, errinfo->blkno);
3932 6 : break;
3933 :
3934 280024 : case VACUUM_ERRCB_PHASE_UNKNOWN:
3935 : default:
3936 280024 : return; /* do nothing; the errinfo may not be
3937 : * initialized */
3938 : }
3939 : }
3940 :
3941 : /*
3942 : * Updates the information required for vacuum error callback. This also saves
3943 : * the current information which can be later restored via restore_vacuum_error_info.
3944 : */
3945 : static void
3946 1568340 : update_vacuum_error_info(LVRelState *vacrel, LVSavedErrInfo *saved_vacrel,
3947 : int phase, BlockNumber blkno, OffsetNumber offnum)
3948 : {
3949 1568340 : if (saved_vacrel)
3950 : {
3951 322856 : saved_vacrel->offnum = vacrel->offnum;
3952 322856 : saved_vacrel->blkno = vacrel->blkno;
3953 322856 : saved_vacrel->phase = vacrel->phase;
3954 : }
3955 :
3956 1568340 : vacrel->blkno = blkno;
3957 1568340 : vacrel->offnum = offnum;
3958 1568340 : vacrel->phase = phase;
3959 1568340 : }
3960 :
3961 : /*
3962 : * Restores the vacuum information saved via a prior call to update_vacuum_error_info.
3963 : */
3964 : static void
3965 322856 : restore_vacuum_error_info(LVRelState *vacrel,
3966 : const LVSavedErrInfo *saved_vacrel)
3967 : {
3968 322856 : vacrel->blkno = saved_vacrel->blkno;
3969 322856 : vacrel->offnum = saved_vacrel->offnum;
3970 322856 : vacrel->phase = saved_vacrel->phase;
3971 322856 : }
|
