Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * vacuumparallel.c
4 : * Support routines for parallel vacuum execution.
5 : *
6 : * This file contains routines that are intended to support setting up, using,
7 : * and tearing down a ParallelVacuumState.
8 : *
9 : * In a parallel vacuum, we perform both index bulk deletion and index cleanup
10 : * with parallel worker processes. Individual indexes are processed by one
11 : * vacuum process. ParallelVacuumState contains shared information as well as
12 : * the memory space for storing dead items allocated in the DSA area. We
13 : * launch parallel worker processes at the start of parallel index
14 : * bulk-deletion and index cleanup and once all indexes are processed, the
15 : * parallel worker processes exit. Each time we process indexes in parallel,
16 : * the parallel context is re-initialized so that the same DSM can be used for
17 : * multiple passes of index bulk-deletion and index cleanup.
18 : *
19 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
20 : * Portions Copyright (c) 1994, Regents of the University of California
21 : *
22 : * IDENTIFICATION
23 : * src/backend/commands/vacuumparallel.c
24 : *
25 : *-------------------------------------------------------------------------
26 : */
27 : #include "postgres.h"
28 :
29 : #include "access/amapi.h"
30 : #include "access/table.h"
31 : #include "access/xact.h"
32 : #include "commands/progress.h"
33 : #include "commands/vacuum.h"
34 : #include "executor/instrument.h"
35 : #include "optimizer/paths.h"
36 : #include "pgstat.h"
37 : #include "storage/bufmgr.h"
38 : #include "tcop/tcopprot.h"
39 : #include "utils/lsyscache.h"
40 : #include "utils/rel.h"
41 :
42 : /*
43 : * DSM keys for parallel vacuum. Unlike other parallel execution code, since
44 : * we don't need to worry about DSM keys conflicting with plan_node_id we can
45 : * use small integers.
46 : */
47 : #define PARALLEL_VACUUM_KEY_SHARED 1
48 : #define PARALLEL_VACUUM_KEY_QUERY_TEXT 2
49 : #define PARALLEL_VACUUM_KEY_BUFFER_USAGE 3
50 : #define PARALLEL_VACUUM_KEY_WAL_USAGE 4
51 : #define PARALLEL_VACUUM_KEY_INDEX_STATS 5
52 :
53 : /*
54 : * Shared information among parallel workers. So this is allocated in the DSM
55 : * segment.
56 : */
57 : typedef struct PVShared
58 : {
59 : /*
60 : * Target table relid, log level (for messages about parallel workers
61 : * launched during VACUUM VERBOSE) and query ID. These fields are not
62 : * modified during the parallel vacuum.
63 : */
64 : Oid relid;
65 : int elevel;
66 : int64 queryid;
67 :
68 : /*
69 : * Fields for both index vacuum and cleanup.
70 : *
71 : * reltuples is the total number of input heap tuples. We set either old
72 : * live tuples in the index vacuum case or the new live tuples in the
73 : * index cleanup case.
74 : *
75 : * estimated_count is true if reltuples is an estimated value. (Note that
76 : * reltuples could be -1 in this case, indicating we have no idea.)
77 : */
78 : double reltuples;
79 : bool estimated_count;
80 :
81 : /*
82 : * In single process vacuum we could consume more memory during index
83 : * vacuuming or cleanup apart from the memory for heap scanning. In
84 : * parallel vacuum, since individual vacuum workers can consume memory
85 : * equal to maintenance_work_mem, the new maintenance_work_mem for each
86 : * worker is set such that the parallel operation doesn't consume more
87 : * memory than single process vacuum.
88 : */
89 : int maintenance_work_mem_worker;
90 :
91 : /*
92 : * The number of buffers each worker's Buffer Access Strategy ring should
93 : * contain.
94 : */
95 : int ring_nbuffers;
96 :
97 : /*
98 : * Shared vacuum cost balance. During parallel vacuum,
99 : * VacuumSharedCostBalance points to this value and it accumulates the
100 : * balance of each parallel vacuum worker.
101 : */
102 : pg_atomic_uint32 cost_balance;
103 :
104 : /*
105 : * Number of active parallel workers. This is used for computing the
106 : * minimum threshold of the vacuum cost balance before a worker sleeps for
107 : * cost-based delay.
108 : */
109 : pg_atomic_uint32 active_nworkers;
110 :
111 : /* Counter for vacuuming and cleanup */
112 : pg_atomic_uint32 idx;
113 :
114 : /* DSA handle where the TidStore lives */
115 : dsa_handle dead_items_dsa_handle;
116 :
117 : /* DSA pointer to the shared TidStore */
118 : dsa_pointer dead_items_handle;
119 :
120 : /* Statistics of shared dead items */
121 : VacDeadItemsInfo dead_items_info;
122 : } PVShared;
123 :
124 : /* Status used during parallel index vacuum or cleanup */
125 : typedef enum PVIndVacStatus
126 : {
127 : PARALLEL_INDVAC_STATUS_INITIAL = 0,
128 : PARALLEL_INDVAC_STATUS_NEED_BULKDELETE,
129 : PARALLEL_INDVAC_STATUS_NEED_CLEANUP,
130 : PARALLEL_INDVAC_STATUS_COMPLETED,
131 : } PVIndVacStatus;
132 :
133 : /*
134 : * Struct for index vacuum statistics of an index that is used for parallel vacuum.
135 : * This includes the status of parallel index vacuum as well as index statistics.
136 : */
137 : typedef struct PVIndStats
138 : {
139 : /*
140 : * The following two fields are set by leader process before executing
141 : * parallel index vacuum or parallel index cleanup. These fields are not
142 : * fixed for the entire VACUUM operation. They are only fixed for an
143 : * individual parallel index vacuum and cleanup.
144 : *
145 : * parallel_workers_can_process is true if both leader and worker can
146 : * process the index, otherwise only leader can process it.
147 : */
148 : PVIndVacStatus status;
149 : bool parallel_workers_can_process;
150 :
151 : /*
152 : * Individual worker or leader stores the result of index vacuum or
153 : * cleanup.
154 : */
155 : bool istat_updated; /* are the stats updated? */
156 : IndexBulkDeleteResult istat;
157 : } PVIndStats;
158 :
159 : /*
160 : * Struct for maintaining a parallel vacuum state. typedef appears in vacuum.h.
161 : */
162 : struct ParallelVacuumState
163 : {
164 : /* NULL for worker processes */
165 : ParallelContext *pcxt;
166 :
167 : /* Parent Heap Relation */
168 : Relation heaprel;
169 :
170 : /* Target indexes */
171 : Relation *indrels;
172 : int nindexes;
173 :
174 : /* Shared information among parallel vacuum workers */
175 : PVShared *shared;
176 :
177 : /*
178 : * Shared index statistics among parallel vacuum workers. The array
179 : * element is allocated for every index, even those indexes where parallel
180 : * index vacuuming is unsafe or not worthwhile (e.g.,
181 : * will_parallel_vacuum[] is false). During parallel vacuum,
182 : * IndexBulkDeleteResult of each index is kept in DSM and is copied into
183 : * local memory at the end of parallel vacuum.
184 : */
185 : PVIndStats *indstats;
186 :
187 : /* Shared dead items space among parallel vacuum workers */
188 : TidStore *dead_items;
189 :
190 : /* Points to buffer usage area in DSM */
191 : BufferUsage *buffer_usage;
192 :
193 : /* Points to WAL usage area in DSM */
194 : WalUsage *wal_usage;
195 :
196 : /*
197 : * False if the index is totally unsuitable target for all parallel
198 : * processing. For example, the index could be <
199 : * min_parallel_index_scan_size cutoff.
200 : */
201 : bool *will_parallel_vacuum;
202 :
203 : /*
204 : * The number of indexes that support parallel index bulk-deletion and
205 : * parallel index cleanup respectively.
206 : */
207 : int nindexes_parallel_bulkdel;
208 : int nindexes_parallel_cleanup;
209 : int nindexes_parallel_condcleanup;
210 :
211 : /* Buffer access strategy used by leader process */
212 : BufferAccessStrategy bstrategy;
213 :
214 : /*
215 : * Error reporting state. The error callback is set only for workers
216 : * processes during parallel index vacuum.
217 : */
218 : char *relnamespace;
219 : char *relname;
220 : char *indname;
221 : PVIndVacStatus status;
222 : };
223 :
224 : static int parallel_vacuum_compute_workers(Relation *indrels, int nindexes, int nrequested,
225 : bool *will_parallel_vacuum);
226 : static void parallel_vacuum_process_all_indexes(ParallelVacuumState *pvs, int num_index_scans,
227 : bool vacuum);
228 : static void parallel_vacuum_process_safe_indexes(ParallelVacuumState *pvs);
229 : static void parallel_vacuum_process_unsafe_indexes(ParallelVacuumState *pvs);
230 : static void parallel_vacuum_process_one_index(ParallelVacuumState *pvs, Relation indrel,
231 : PVIndStats *indstats);
232 : static bool parallel_vacuum_index_is_parallel_safe(Relation indrel, int num_index_scans,
233 : bool vacuum);
234 : static void parallel_vacuum_error_callback(void *arg);
235 :
236 : /*
237 : * Try to enter parallel mode and create a parallel context. Then initialize
238 : * shared memory state.
239 : *
240 : * On success, return parallel vacuum state. Otherwise return NULL.
241 : */
242 : ParallelVacuumState *
243 5530 : parallel_vacuum_init(Relation rel, Relation *indrels, int nindexes,
244 : int nrequested_workers, int vac_work_mem,
245 : int elevel, BufferAccessStrategy bstrategy)
246 : {
247 : ParallelVacuumState *pvs;
248 : ParallelContext *pcxt;
249 : PVShared *shared;
250 : TidStore *dead_items;
251 : PVIndStats *indstats;
252 : BufferUsage *buffer_usage;
253 : WalUsage *wal_usage;
254 : bool *will_parallel_vacuum;
255 : Size est_indstats_len;
256 : Size est_shared_len;
257 5530 : int nindexes_mwm = 0;
258 5530 : int parallel_workers = 0;
259 : int querylen;
260 :
261 : /*
262 : * A parallel vacuum must be requested and there must be indexes on the
263 : * relation
264 : */
265 : Assert(nrequested_workers >= 0);
266 : Assert(nindexes > 0);
267 :
268 : /*
269 : * Compute the number of parallel vacuum workers to launch
270 : */
271 5530 : will_parallel_vacuum = palloc0_array(bool, nindexes);
272 5530 : parallel_workers = parallel_vacuum_compute_workers(indrels, nindexes,
273 : nrequested_workers,
274 : will_parallel_vacuum);
275 5530 : if (parallel_workers <= 0)
276 : {
277 : /* Can't perform vacuum in parallel -- return NULL */
278 5513 : pfree(will_parallel_vacuum);
279 5513 : return NULL;
280 : }
281 :
282 17 : pvs = palloc0_object(ParallelVacuumState);
283 17 : pvs->indrels = indrels;
284 17 : pvs->nindexes = nindexes;
285 17 : pvs->will_parallel_vacuum = will_parallel_vacuum;
286 17 : pvs->bstrategy = bstrategy;
287 17 : pvs->heaprel = rel;
288 :
289 17 : EnterParallelMode();
290 17 : pcxt = CreateParallelContext("postgres", "parallel_vacuum_main",
291 : parallel_workers);
292 : Assert(pcxt->nworkers > 0);
293 17 : pvs->pcxt = pcxt;
294 :
295 : /* Estimate size for index vacuum stats -- PARALLEL_VACUUM_KEY_INDEX_STATS */
296 17 : est_indstats_len = mul_size(sizeof(PVIndStats), nindexes);
297 17 : shm_toc_estimate_chunk(&pcxt->estimator, est_indstats_len);
298 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
299 :
300 : /* Estimate size for shared information -- PARALLEL_VACUUM_KEY_SHARED */
301 17 : est_shared_len = sizeof(PVShared);
302 17 : shm_toc_estimate_chunk(&pcxt->estimator, est_shared_len);
303 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
304 :
305 : /*
306 : * Estimate space for BufferUsage and WalUsage --
307 : * PARALLEL_VACUUM_KEY_BUFFER_USAGE and PARALLEL_VACUUM_KEY_WAL_USAGE.
308 : *
309 : * If there are no extensions loaded that care, we could skip this. We
310 : * have no way of knowing whether anyone's looking at pgBufferUsage or
311 : * pgWalUsage, so do it unconditionally.
312 : */
313 17 : shm_toc_estimate_chunk(&pcxt->estimator,
314 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
315 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
316 17 : shm_toc_estimate_chunk(&pcxt->estimator,
317 : mul_size(sizeof(WalUsage), pcxt->nworkers));
318 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
319 :
320 : /* Finally, estimate PARALLEL_VACUUM_KEY_QUERY_TEXT space */
321 17 : if (debug_query_string)
322 : {
323 17 : querylen = strlen(debug_query_string);
324 17 : shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
325 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
326 : }
327 : else
328 0 : querylen = 0; /* keep compiler quiet */
329 :
330 17 : InitializeParallelDSM(pcxt);
331 :
332 : /* Prepare index vacuum stats */
333 17 : indstats = (PVIndStats *) shm_toc_allocate(pcxt->toc, est_indstats_len);
334 383 : MemSet(indstats, 0, est_indstats_len);
335 78 : for (int i = 0; i < nindexes; i++)
336 : {
337 61 : Relation indrel = indrels[i];
338 61 : uint8 vacoptions = indrel->rd_indam->amparallelvacuumoptions;
339 :
340 : /*
341 : * Cleanup option should be either disabled, always performing in
342 : * parallel or conditionally performing in parallel.
343 : */
344 : Assert(((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) == 0) ||
345 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) == 0));
346 : Assert(vacoptions <= VACUUM_OPTION_MAX_VALID_VALUE);
347 :
348 61 : if (!will_parallel_vacuum[i])
349 3 : continue;
350 :
351 58 : if (indrel->rd_indam->amusemaintenanceworkmem)
352 6 : nindexes_mwm++;
353 :
354 : /*
355 : * Remember the number of indexes that support parallel operation for
356 : * each phase.
357 : */
358 58 : if ((vacoptions & VACUUM_OPTION_PARALLEL_BULKDEL) != 0)
359 52 : pvs->nindexes_parallel_bulkdel++;
360 58 : if ((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) != 0)
361 12 : pvs->nindexes_parallel_cleanup++;
362 58 : if ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) != 0)
363 40 : pvs->nindexes_parallel_condcleanup++;
364 : }
365 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_INDEX_STATS, indstats);
366 17 : pvs->indstats = indstats;
367 :
368 : /* Prepare shared information */
369 17 : shared = (PVShared *) shm_toc_allocate(pcxt->toc, est_shared_len);
370 187 : MemSet(shared, 0, est_shared_len);
371 17 : shared->relid = RelationGetRelid(rel);
372 17 : shared->elevel = elevel;
373 17 : shared->queryid = pgstat_get_my_query_id();
374 17 : shared->maintenance_work_mem_worker =
375 : (nindexes_mwm > 0) ?
376 17 : maintenance_work_mem / Min(parallel_workers, nindexes_mwm) :
377 : maintenance_work_mem;
378 17 : shared->dead_items_info.max_bytes = vac_work_mem * (size_t) 1024;
379 :
380 : /* Prepare DSA space for dead items */
381 17 : dead_items = TidStoreCreateShared(shared->dead_items_info.max_bytes,
382 : LWTRANCHE_PARALLEL_VACUUM_DSA);
383 17 : pvs->dead_items = dead_items;
384 17 : shared->dead_items_handle = TidStoreGetHandle(dead_items);
385 17 : shared->dead_items_dsa_handle = dsa_get_handle(TidStoreGetDSA(dead_items));
386 :
387 : /* Use the same buffer size for all workers */
388 17 : shared->ring_nbuffers = GetAccessStrategyBufferCount(bstrategy);
389 :
390 17 : pg_atomic_init_u32(&(shared->cost_balance), 0);
391 17 : pg_atomic_init_u32(&(shared->active_nworkers), 0);
392 17 : pg_atomic_init_u32(&(shared->idx), 0);
393 :
394 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_SHARED, shared);
395 17 : pvs->shared = shared;
396 :
397 : /*
398 : * Allocate space for each worker's BufferUsage and WalUsage; no need to
399 : * initialize
400 : */
401 17 : buffer_usage = shm_toc_allocate(pcxt->toc,
402 17 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
403 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_BUFFER_USAGE, buffer_usage);
404 17 : pvs->buffer_usage = buffer_usage;
405 17 : wal_usage = shm_toc_allocate(pcxt->toc,
406 17 : mul_size(sizeof(WalUsage), pcxt->nworkers));
407 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_WAL_USAGE, wal_usage);
408 17 : pvs->wal_usage = wal_usage;
409 :
410 : /* Store query string for workers */
411 17 : if (debug_query_string)
412 : {
413 : char *sharedquery;
414 :
415 17 : sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
416 17 : memcpy(sharedquery, debug_query_string, querylen + 1);
417 17 : sharedquery[querylen] = '\0';
418 17 : shm_toc_insert(pcxt->toc,
419 : PARALLEL_VACUUM_KEY_QUERY_TEXT, sharedquery);
420 : }
421 :
422 : /* Success -- return parallel vacuum state */
423 17 : return pvs;
424 : }
425 :
426 : /*
427 : * Destroy the parallel context, and end parallel mode.
428 : *
429 : * Since writes are not allowed during parallel mode, copy the
430 : * updated index statistics from DSM into local memory and then later use that
431 : * to update the index statistics. One might think that we can exit from
432 : * parallel mode, update the index statistics and then destroy parallel
433 : * context, but that won't be safe (see ExitParallelMode).
434 : */
435 : void
436 17 : parallel_vacuum_end(ParallelVacuumState *pvs, IndexBulkDeleteResult **istats)
437 : {
438 : Assert(!IsParallelWorker());
439 :
440 : /* Copy the updated statistics */
441 78 : for (int i = 0; i < pvs->nindexes; i++)
442 : {
443 61 : PVIndStats *indstats = &(pvs->indstats[i]);
444 :
445 61 : if (indstats->istat_updated)
446 : {
447 41 : istats[i] = palloc0_object(IndexBulkDeleteResult);
448 41 : memcpy(istats[i], &indstats->istat, sizeof(IndexBulkDeleteResult));
449 : }
450 : else
451 20 : istats[i] = NULL;
452 : }
453 :
454 17 : TidStoreDestroy(pvs->dead_items);
455 :
456 17 : DestroyParallelContext(pvs->pcxt);
457 17 : ExitParallelMode();
458 :
459 17 : pfree(pvs->will_parallel_vacuum);
460 17 : pfree(pvs);
461 17 : }
462 :
463 : /*
464 : * Returns the dead items space and dead items information.
465 : */
466 : TidStore *
467 28 : parallel_vacuum_get_dead_items(ParallelVacuumState *pvs, VacDeadItemsInfo **dead_items_info_p)
468 : {
469 28 : *dead_items_info_p = &(pvs->shared->dead_items_info);
470 28 : return pvs->dead_items;
471 : }
472 :
473 : /* Forget all items in dead_items */
474 : void
475 11 : parallel_vacuum_reset_dead_items(ParallelVacuumState *pvs)
476 : {
477 11 : VacDeadItemsInfo *dead_items_info = &(pvs->shared->dead_items_info);
478 :
479 : /*
480 : * Free the current tidstore and return allocated DSA segments to the
481 : * operating system. Then we recreate the tidstore with the same max_bytes
482 : * limitation we just used.
483 : */
484 11 : TidStoreDestroy(pvs->dead_items);
485 11 : pvs->dead_items = TidStoreCreateShared(dead_items_info->max_bytes,
486 : LWTRANCHE_PARALLEL_VACUUM_DSA);
487 :
488 : /* Update the DSA pointer for dead_items to the new one */
489 11 : pvs->shared->dead_items_dsa_handle = dsa_get_handle(TidStoreGetDSA(pvs->dead_items));
490 11 : pvs->shared->dead_items_handle = TidStoreGetHandle(pvs->dead_items);
491 :
492 : /* Reset the counter */
493 11 : dead_items_info->num_items = 0;
494 11 : }
495 :
496 : /*
497 : * Do parallel index bulk-deletion with parallel workers.
498 : */
499 : void
500 11 : parallel_vacuum_bulkdel_all_indexes(ParallelVacuumState *pvs, long num_table_tuples,
501 : int num_index_scans)
502 : {
503 : Assert(!IsParallelWorker());
504 :
505 : /*
506 : * We can only provide an approximate value of num_heap_tuples, at least
507 : * for now.
508 : */
509 11 : pvs->shared->reltuples = num_table_tuples;
510 11 : pvs->shared->estimated_count = true;
511 :
512 11 : parallel_vacuum_process_all_indexes(pvs, num_index_scans, true);
513 11 : }
514 :
515 : /*
516 : * Do parallel index cleanup with parallel workers.
517 : */
518 : void
519 17 : parallel_vacuum_cleanup_all_indexes(ParallelVacuumState *pvs, long num_table_tuples,
520 : int num_index_scans, bool estimated_count)
521 : {
522 : Assert(!IsParallelWorker());
523 :
524 : /*
525 : * We can provide a better estimate of total number of surviving tuples
526 : * (we assume indexes are more interested in that than in the number of
527 : * nominally live tuples).
528 : */
529 17 : pvs->shared->reltuples = num_table_tuples;
530 17 : pvs->shared->estimated_count = estimated_count;
531 :
532 17 : parallel_vacuum_process_all_indexes(pvs, num_index_scans, false);
533 17 : }
534 :
535 : /*
536 : * Compute the number of parallel worker processes to request. Both index
537 : * vacuum and index cleanup can be executed with parallel workers.
538 : * The index is eligible for parallel vacuum iff its size is greater than
539 : * min_parallel_index_scan_size as invoking workers for very small indexes
540 : * can hurt performance.
541 : *
542 : * nrequested is the number of parallel workers that user requested. If
543 : * nrequested is 0, we compute the parallel degree based on nindexes, that is
544 : * the number of indexes that support parallel vacuum. This function also
545 : * sets will_parallel_vacuum to remember indexes that participate in parallel
546 : * vacuum.
547 : */
548 : static int
549 5530 : parallel_vacuum_compute_workers(Relation *indrels, int nindexes, int nrequested,
550 : bool *will_parallel_vacuum)
551 : {
552 5530 : int nindexes_parallel = 0;
553 5530 : int nindexes_parallel_bulkdel = 0;
554 5530 : int nindexes_parallel_cleanup = 0;
555 : int parallel_workers;
556 :
557 : /*
558 : * We don't allow performing parallel operation in standalone backend or
559 : * when parallelism is disabled.
560 : */
561 5530 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
562 2401 : return 0;
563 :
564 : /*
565 : * Compute the number of indexes that can participate in parallel vacuum.
566 : */
567 10205 : for (int i = 0; i < nindexes; i++)
568 : {
569 7076 : Relation indrel = indrels[i];
570 7076 : uint8 vacoptions = indrel->rd_indam->amparallelvacuumoptions;
571 :
572 : /* Skip index that is not a suitable target for parallel index vacuum */
573 7076 : if (vacoptions == VACUUM_OPTION_NO_PARALLEL ||
574 7076 : RelationGetNumberOfBlocks(indrel) < min_parallel_index_scan_size)
575 7011 : continue;
576 :
577 65 : will_parallel_vacuum[i] = true;
578 :
579 65 : if ((vacoptions & VACUUM_OPTION_PARALLEL_BULKDEL) != 0)
580 59 : nindexes_parallel_bulkdel++;
581 65 : if (((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) != 0) ||
582 53 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) != 0))
583 59 : nindexes_parallel_cleanup++;
584 : }
585 :
586 3129 : nindexes_parallel = Max(nindexes_parallel_bulkdel,
587 : nindexes_parallel_cleanup);
588 :
589 : /* The leader process takes one index */
590 3129 : nindexes_parallel--;
591 :
592 : /* No index supports parallel vacuum */
593 3129 : if (nindexes_parallel <= 0)
594 3112 : return 0;
595 :
596 : /* Compute the parallel degree */
597 17 : parallel_workers = (nrequested > 0) ?
598 17 : Min(nrequested, nindexes_parallel) : nindexes_parallel;
599 :
600 : /* Cap by max_parallel_maintenance_workers */
601 17 : parallel_workers = Min(parallel_workers, max_parallel_maintenance_workers);
602 :
603 17 : return parallel_workers;
604 : }
605 :
606 : /*
607 : * Perform index vacuum or index cleanup with parallel workers. This function
608 : * must be used by the parallel vacuum leader process.
609 : */
610 : static void
611 28 : parallel_vacuum_process_all_indexes(ParallelVacuumState *pvs, int num_index_scans,
612 : bool vacuum)
613 : {
614 : int nworkers;
615 : PVIndVacStatus new_status;
616 :
617 : Assert(!IsParallelWorker());
618 :
619 28 : if (vacuum)
620 : {
621 11 : new_status = PARALLEL_INDVAC_STATUS_NEED_BULKDELETE;
622 :
623 : /* Determine the number of parallel workers to launch */
624 11 : nworkers = pvs->nindexes_parallel_bulkdel;
625 : }
626 : else
627 : {
628 17 : new_status = PARALLEL_INDVAC_STATUS_NEED_CLEANUP;
629 :
630 : /* Determine the number of parallel workers to launch */
631 17 : nworkers = pvs->nindexes_parallel_cleanup;
632 :
633 : /* Add conditionally parallel-aware indexes if in the first time call */
634 17 : if (num_index_scans == 0)
635 10 : nworkers += pvs->nindexes_parallel_condcleanup;
636 : }
637 :
638 : /* The leader process will participate */
639 28 : nworkers--;
640 :
641 : /*
642 : * It is possible that parallel context is initialized with fewer workers
643 : * than the number of indexes that need a separate worker in the current
644 : * phase, so we need to consider it. See
645 : * parallel_vacuum_compute_workers().
646 : */
647 28 : nworkers = Min(nworkers, pvs->pcxt->nworkers);
648 :
649 : /*
650 : * Set index vacuum status and mark whether parallel vacuum worker can
651 : * process it.
652 : */
653 118 : for (int i = 0; i < pvs->nindexes; i++)
654 : {
655 90 : PVIndStats *indstats = &(pvs->indstats[i]);
656 :
657 : Assert(indstats->status == PARALLEL_INDVAC_STATUS_INITIAL);
658 90 : indstats->status = new_status;
659 90 : indstats->parallel_workers_can_process =
660 174 : (pvs->will_parallel_vacuum[i] &&
661 84 : parallel_vacuum_index_is_parallel_safe(pvs->indrels[i],
662 : num_index_scans,
663 90 : vacuum));
664 : }
665 :
666 : /* Reset the parallel index processing and progress counters */
667 28 : pg_atomic_write_u32(&(pvs->shared->idx), 0);
668 :
669 : /* Setup the shared cost-based vacuum delay and launch workers */
670 28 : if (nworkers > 0)
671 : {
672 : /* Reinitialize parallel context to relaunch parallel workers */
673 22 : if (num_index_scans > 0)
674 5 : ReinitializeParallelDSM(pvs->pcxt);
675 :
676 : /*
677 : * Set up shared cost balance and the number of active workers for
678 : * vacuum delay. We need to do this before launching workers as
679 : * otherwise, they might not see the updated values for these
680 : * parameters.
681 : */
682 22 : pg_atomic_write_u32(&(pvs->shared->cost_balance), VacuumCostBalance);
683 22 : pg_atomic_write_u32(&(pvs->shared->active_nworkers), 0);
684 :
685 : /*
686 : * The number of workers can vary between bulkdelete and cleanup
687 : * phase.
688 : */
689 22 : ReinitializeParallelWorkers(pvs->pcxt, nworkers);
690 :
691 22 : LaunchParallelWorkers(pvs->pcxt);
692 :
693 22 : if (pvs->pcxt->nworkers_launched > 0)
694 : {
695 : /*
696 : * Reset the local cost values for leader backend as we have
697 : * already accumulated the remaining balance of heap.
698 : */
699 22 : VacuumCostBalance = 0;
700 22 : VacuumCostBalanceLocal = 0;
701 :
702 : /* Enable shared cost balance for leader backend */
703 22 : VacuumSharedCostBalance = &(pvs->shared->cost_balance);
704 22 : VacuumActiveNWorkers = &(pvs->shared->active_nworkers);
705 : }
706 :
707 22 : if (vacuum)
708 11 : ereport(pvs->shared->elevel,
709 : (errmsg(ngettext("launched %d parallel vacuum worker for index vacuuming (planned: %d)",
710 : "launched %d parallel vacuum workers for index vacuuming (planned: %d)",
711 : pvs->pcxt->nworkers_launched),
712 : pvs->pcxt->nworkers_launched, nworkers)));
713 : else
714 11 : ereport(pvs->shared->elevel,
715 : (errmsg(ngettext("launched %d parallel vacuum worker for index cleanup (planned: %d)",
716 : "launched %d parallel vacuum workers for index cleanup (planned: %d)",
717 : pvs->pcxt->nworkers_launched),
718 : pvs->pcxt->nworkers_launched, nworkers)));
719 : }
720 :
721 : /* Vacuum the indexes that can be processed by only leader process */
722 28 : parallel_vacuum_process_unsafe_indexes(pvs);
723 :
724 : /*
725 : * Join as a parallel worker. The leader vacuums alone processes all
726 : * parallel-safe indexes in the case where no workers are launched.
727 : */
728 28 : parallel_vacuum_process_safe_indexes(pvs);
729 :
730 : /*
731 : * Next, accumulate buffer and WAL usage. (This must wait for the workers
732 : * to finish, or we might get incomplete data.)
733 : */
734 28 : if (nworkers > 0)
735 : {
736 : /* Wait for all vacuum workers to finish */
737 22 : WaitForParallelWorkersToFinish(pvs->pcxt);
738 :
739 50 : for (int i = 0; i < pvs->pcxt->nworkers_launched; i++)
740 28 : InstrAccumParallelQuery(&pvs->buffer_usage[i], &pvs->wal_usage[i]);
741 : }
742 :
743 : /*
744 : * Reset all index status back to initial (while checking that we have
745 : * vacuumed all indexes).
746 : */
747 118 : for (int i = 0; i < pvs->nindexes; i++)
748 : {
749 90 : PVIndStats *indstats = &(pvs->indstats[i]);
750 :
751 90 : if (indstats->status != PARALLEL_INDVAC_STATUS_COMPLETED)
752 0 : elog(ERROR, "parallel index vacuum on index \"%s\" is not completed",
753 : RelationGetRelationName(pvs->indrels[i]));
754 :
755 90 : indstats->status = PARALLEL_INDVAC_STATUS_INITIAL;
756 : }
757 :
758 : /*
759 : * Carry the shared balance value to heap scan and disable shared costing
760 : */
761 28 : if (VacuumSharedCostBalance)
762 : {
763 22 : VacuumCostBalance = pg_atomic_read_u32(VacuumSharedCostBalance);
764 22 : VacuumSharedCostBalance = NULL;
765 22 : VacuumActiveNWorkers = NULL;
766 : }
767 28 : }
768 :
769 : /*
770 : * Index vacuum/cleanup routine used by the leader process and parallel
771 : * vacuum worker processes to vacuum the indexes in parallel.
772 : */
773 : static void
774 56 : parallel_vacuum_process_safe_indexes(ParallelVacuumState *pvs)
775 : {
776 : /*
777 : * Increment the active worker count if we are able to launch any worker.
778 : */
779 56 : if (VacuumActiveNWorkers)
780 50 : pg_atomic_add_fetch_u32(VacuumActiveNWorkers, 1);
781 :
782 : /* Loop until all indexes are vacuumed */
783 : for (;;)
784 90 : {
785 : int idx;
786 : PVIndStats *indstats;
787 :
788 : /* Get an index number to process */
789 146 : idx = pg_atomic_fetch_add_u32(&(pvs->shared->idx), 1);
790 :
791 : /* Done for all indexes? */
792 146 : if (idx >= pvs->nindexes)
793 56 : break;
794 :
795 90 : indstats = &(pvs->indstats[idx]);
796 :
797 : /*
798 : * Skip vacuuming index that is unsafe for workers or has an
799 : * unsuitable target for parallel index vacuum (this is vacuumed in
800 : * parallel_vacuum_process_unsafe_indexes() by the leader).
801 : */
802 90 : if (!indstats->parallel_workers_can_process)
803 28 : continue;
804 :
805 : /* Do vacuum or cleanup of the index */
806 62 : parallel_vacuum_process_one_index(pvs, pvs->indrels[idx], indstats);
807 : }
808 :
809 : /*
810 : * We have completed the index vacuum so decrement the active worker
811 : * count.
812 : */
813 56 : if (VacuumActiveNWorkers)
814 50 : pg_atomic_sub_fetch_u32(VacuumActiveNWorkers, 1);
815 56 : }
816 :
817 : /*
818 : * Perform parallel vacuuming of indexes in leader process.
819 : *
820 : * Handles index vacuuming (or index cleanup) for indexes that are not
821 : * parallel safe. It's possible that this will vary for a given index, based
822 : * on details like whether we're performing index cleanup right now.
823 : *
824 : * Also performs vacuuming of smaller indexes that fell under the size cutoff
825 : * enforced by parallel_vacuum_compute_workers().
826 : */
827 : static void
828 28 : parallel_vacuum_process_unsafe_indexes(ParallelVacuumState *pvs)
829 : {
830 : Assert(!IsParallelWorker());
831 :
832 : /*
833 : * Increment the active worker count if we are able to launch any worker.
834 : */
835 28 : if (VacuumActiveNWorkers)
836 22 : pg_atomic_add_fetch_u32(VacuumActiveNWorkers, 1);
837 :
838 118 : for (int i = 0; i < pvs->nindexes; i++)
839 : {
840 90 : PVIndStats *indstats = &(pvs->indstats[i]);
841 :
842 : /* Skip, indexes that are safe for workers */
843 90 : if (indstats->parallel_workers_can_process)
844 62 : continue;
845 :
846 : /* Do vacuum or cleanup of the index */
847 28 : parallel_vacuum_process_one_index(pvs, pvs->indrels[i], indstats);
848 : }
849 :
850 : /*
851 : * We have completed the index vacuum so decrement the active worker
852 : * count.
853 : */
854 28 : if (VacuumActiveNWorkers)
855 22 : pg_atomic_sub_fetch_u32(VacuumActiveNWorkers, 1);
856 28 : }
857 :
858 : /*
859 : * Vacuum or cleanup index either by leader process or by one of the worker
860 : * process. After vacuuming the index this function copies the index
861 : * statistics returned from ambulkdelete and amvacuumcleanup to the DSM
862 : * segment.
863 : */
864 : static void
865 90 : parallel_vacuum_process_one_index(ParallelVacuumState *pvs, Relation indrel,
866 : PVIndStats *indstats)
867 : {
868 90 : IndexBulkDeleteResult *istat = NULL;
869 : IndexBulkDeleteResult *istat_res;
870 : IndexVacuumInfo ivinfo;
871 :
872 : /*
873 : * Update the pointer to the corresponding bulk-deletion result if someone
874 : * has already updated it
875 : */
876 90 : if (indstats->istat_updated)
877 29 : istat = &(indstats->istat);
878 :
879 90 : ivinfo.index = indrel;
880 90 : ivinfo.heaprel = pvs->heaprel;
881 90 : ivinfo.analyze_only = false;
882 90 : ivinfo.report_progress = false;
883 90 : ivinfo.message_level = DEBUG2;
884 90 : ivinfo.estimated_count = pvs->shared->estimated_count;
885 90 : ivinfo.num_heap_tuples = pvs->shared->reltuples;
886 90 : ivinfo.strategy = pvs->bstrategy;
887 :
888 : /* Update error traceback information */
889 90 : pvs->indname = pstrdup(RelationGetRelationName(indrel));
890 90 : pvs->status = indstats->status;
891 :
892 90 : switch (indstats->status)
893 : {
894 29 : case PARALLEL_INDVAC_STATUS_NEED_BULKDELETE:
895 29 : istat_res = vac_bulkdel_one_index(&ivinfo, istat, pvs->dead_items,
896 29 : &pvs->shared->dead_items_info);
897 29 : break;
898 61 : case PARALLEL_INDVAC_STATUS_NEED_CLEANUP:
899 61 : istat_res = vac_cleanup_one_index(&ivinfo, istat);
900 61 : break;
901 0 : default:
902 0 : elog(ERROR, "unexpected parallel vacuum index status %d for index \"%s\"",
903 : indstats->status,
904 : RelationGetRelationName(indrel));
905 : }
906 :
907 : /*
908 : * Copy the index bulk-deletion result returned from ambulkdelete and
909 : * amvacuumcleanup to the DSM segment if it's the first cycle because they
910 : * allocate locally and it's possible that an index will be vacuumed by a
911 : * different vacuum process the next cycle. Copying the result normally
912 : * happens only the first time an index is vacuumed. For any additional
913 : * vacuum pass, we directly point to the result on the DSM segment and
914 : * pass it to vacuum index APIs so that workers can update it directly.
915 : *
916 : * Since all vacuum workers write the bulk-deletion result at different
917 : * slots we can write them without locking.
918 : */
919 90 : if (!indstats->istat_updated && istat_res != NULL)
920 : {
921 41 : memcpy(&(indstats->istat), istat_res, sizeof(IndexBulkDeleteResult));
922 41 : indstats->istat_updated = true;
923 :
924 : /* Free the locally-allocated bulk-deletion result */
925 41 : pfree(istat_res);
926 : }
927 :
928 : /*
929 : * Update the status to completed. No need to lock here since each worker
930 : * touches different indexes.
931 : */
932 90 : indstats->status = PARALLEL_INDVAC_STATUS_COMPLETED;
933 :
934 : /* Reset error traceback information */
935 90 : pvs->status = PARALLEL_INDVAC_STATUS_COMPLETED;
936 90 : pfree(pvs->indname);
937 90 : pvs->indname = NULL;
938 :
939 : /*
940 : * Call the parallel variant of pgstat_progress_incr_param so workers can
941 : * report progress of index vacuum to the leader.
942 : */
943 90 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_INDEXES_PROCESSED, 1);
944 90 : }
945 :
946 : /*
947 : * Returns false, if the given index can't participate in the next execution of
948 : * parallel index vacuum or parallel index cleanup.
949 : */
950 : static bool
951 84 : parallel_vacuum_index_is_parallel_safe(Relation indrel, int num_index_scans,
952 : bool vacuum)
953 : {
954 : uint8 vacoptions;
955 :
956 84 : vacoptions = indrel->rd_indam->amparallelvacuumoptions;
957 :
958 : /* In parallel vacuum case, check if it supports parallel bulk-deletion */
959 84 : if (vacuum)
960 26 : return ((vacoptions & VACUUM_OPTION_PARALLEL_BULKDEL) != 0);
961 :
962 : /* Not safe, if the index does not support parallel cleanup */
963 58 : if (((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) == 0) &&
964 46 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) == 0))
965 6 : return false;
966 :
967 : /*
968 : * Not safe, if the index supports parallel cleanup conditionally, but we
969 : * have already processed the index (for bulkdelete). We do this to avoid
970 : * the need to invoke workers when parallel index cleanup doesn't need to
971 : * scan the index. See the comments for option
972 : * VACUUM_OPTION_PARALLEL_COND_CLEANUP to know when indexes support
973 : * parallel cleanup conditionally.
974 : */
975 52 : if (num_index_scans > 0 &&
976 17 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) != 0))
977 15 : return false;
978 :
979 37 : return true;
980 : }
981 :
982 : /*
983 : * Perform work within a launched parallel process.
984 : *
985 : * Since parallel vacuum workers perform only index vacuum or index cleanup,
986 : * we don't need to report progress information.
987 : */
988 : void
989 28 : parallel_vacuum_main(dsm_segment *seg, shm_toc *toc)
990 : {
991 : ParallelVacuumState pvs;
992 : Relation rel;
993 : Relation *indrels;
994 : PVIndStats *indstats;
995 : PVShared *shared;
996 : TidStore *dead_items;
997 : BufferUsage *buffer_usage;
998 : WalUsage *wal_usage;
999 : int nindexes;
1000 : char *sharedquery;
1001 : ErrorContextCallback errcallback;
1002 :
1003 : /*
1004 : * A parallel vacuum worker must have only PROC_IN_VACUUM flag since we
1005 : * don't support parallel vacuum for autovacuum as of now.
1006 : */
1007 : Assert(MyProc->statusFlags == PROC_IN_VACUUM);
1008 :
1009 28 : elog(DEBUG1, "starting parallel vacuum worker");
1010 :
1011 28 : shared = (PVShared *) shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_SHARED, false);
1012 :
1013 : /* Set debug_query_string for individual workers */
1014 28 : sharedquery = shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_QUERY_TEXT, true);
1015 28 : debug_query_string = sharedquery;
1016 28 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
1017 :
1018 : /* Track query ID */
1019 28 : pgstat_report_query_id(shared->queryid, false);
1020 :
1021 : /*
1022 : * Open table. The lock mode is the same as the leader process. It's
1023 : * okay because the lock mode does not conflict among the parallel
1024 : * workers.
1025 : */
1026 28 : rel = table_open(shared->relid, ShareUpdateExclusiveLock);
1027 :
1028 : /*
1029 : * Open all indexes. indrels are sorted in order by OID, which should be
1030 : * matched to the leader's one.
1031 : */
1032 28 : vac_open_indexes(rel, RowExclusiveLock, &nindexes, &indrels);
1033 : Assert(nindexes > 0);
1034 :
1035 : /*
1036 : * Apply the desired value of maintenance_work_mem within this process.
1037 : * Really we should use SetConfigOption() to change a GUC, but since we're
1038 : * already in parallel mode guc.c would complain about that. Fortunately,
1039 : * by the same token guc.c will not let any user-defined code change it.
1040 : * So just avert your eyes while we do this:
1041 : */
1042 28 : if (shared->maintenance_work_mem_worker > 0)
1043 28 : maintenance_work_mem = shared->maintenance_work_mem_worker;
1044 :
1045 : /* Set index statistics */
1046 28 : indstats = (PVIndStats *) shm_toc_lookup(toc,
1047 : PARALLEL_VACUUM_KEY_INDEX_STATS,
1048 : false);
1049 :
1050 : /* Find dead_items in shared memory */
1051 28 : dead_items = TidStoreAttach(shared->dead_items_dsa_handle,
1052 : shared->dead_items_handle);
1053 :
1054 : /* Set cost-based vacuum delay */
1055 28 : VacuumUpdateCosts();
1056 28 : VacuumCostBalance = 0;
1057 28 : VacuumCostBalanceLocal = 0;
1058 28 : VacuumSharedCostBalance = &(shared->cost_balance);
1059 28 : VacuumActiveNWorkers = &(shared->active_nworkers);
1060 :
1061 : /* Set parallel vacuum state */
1062 28 : pvs.indrels = indrels;
1063 28 : pvs.nindexes = nindexes;
1064 28 : pvs.indstats = indstats;
1065 28 : pvs.shared = shared;
1066 28 : pvs.dead_items = dead_items;
1067 28 : pvs.relnamespace = get_namespace_name(RelationGetNamespace(rel));
1068 28 : pvs.relname = pstrdup(RelationGetRelationName(rel));
1069 28 : pvs.heaprel = rel;
1070 :
1071 : /* These fields will be filled during index vacuum or cleanup */
1072 28 : pvs.indname = NULL;
1073 28 : pvs.status = PARALLEL_INDVAC_STATUS_INITIAL;
1074 :
1075 : /* Each parallel VACUUM worker gets its own access strategy. */
1076 56 : pvs.bstrategy = GetAccessStrategyWithSize(BAS_VACUUM,
1077 28 : shared->ring_nbuffers * (BLCKSZ / 1024));
1078 :
1079 : /* Setup error traceback support for ereport() */
1080 28 : errcallback.callback = parallel_vacuum_error_callback;
1081 28 : errcallback.arg = &pvs;
1082 28 : errcallback.previous = error_context_stack;
1083 28 : error_context_stack = &errcallback;
1084 :
1085 : /* Prepare to track buffer usage during parallel execution */
1086 28 : InstrStartParallelQuery();
1087 :
1088 : /* Process indexes to perform vacuum/cleanup */
1089 28 : parallel_vacuum_process_safe_indexes(&pvs);
1090 :
1091 : /* Report buffer/WAL usage during parallel execution */
1092 28 : buffer_usage = shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_BUFFER_USAGE, false);
1093 28 : wal_usage = shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_WAL_USAGE, false);
1094 28 : InstrEndParallelQuery(&buffer_usage[ParallelWorkerNumber],
1095 28 : &wal_usage[ParallelWorkerNumber]);
1096 :
1097 : /* Report any remaining cost-based vacuum delay time */
1098 28 : if (track_cost_delay_timing)
1099 0 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_DELAY_TIME,
1100 : parallel_vacuum_worker_delay_ns);
1101 :
1102 28 : TidStoreDetach(dead_items);
1103 :
1104 : /* Pop the error context stack */
1105 28 : error_context_stack = errcallback.previous;
1106 :
1107 28 : vac_close_indexes(nindexes, indrels, RowExclusiveLock);
1108 28 : table_close(rel, ShareUpdateExclusiveLock);
1109 28 : FreeAccessStrategy(pvs.bstrategy);
1110 28 : }
1111 :
1112 : /*
1113 : * Error context callback for errors occurring during parallel index vacuum.
1114 : * The error context messages should match the messages set in the lazy vacuum
1115 : * error context. If you change this function, change vacuum_error_callback()
1116 : * as well.
1117 : */
1118 : static void
1119 0 : parallel_vacuum_error_callback(void *arg)
1120 : {
1121 0 : ParallelVacuumState *errinfo = arg;
1122 :
1123 0 : switch (errinfo->status)
1124 : {
1125 0 : case PARALLEL_INDVAC_STATUS_NEED_BULKDELETE:
1126 0 : errcontext("while vacuuming index \"%s\" of relation \"%s.%s\"",
1127 : errinfo->indname,
1128 : errinfo->relnamespace,
1129 : errinfo->relname);
1130 0 : break;
1131 0 : case PARALLEL_INDVAC_STATUS_NEED_CLEANUP:
1132 0 : errcontext("while cleaning up index \"%s\" of relation \"%s.%s\"",
1133 : errinfo->indname,
1134 : errinfo->relnamespace,
1135 : errinfo->relname);
1136 0 : break;
1137 0 : case PARALLEL_INDVAC_STATUS_INITIAL:
1138 : case PARALLEL_INDVAC_STATUS_COMPLETED:
1139 : default:
1140 0 : return;
1141 : }
1142 : }
|
