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 "storage/proc.h"
39 : #include "tcop/tcopprot.h"
40 : #include "utils/lsyscache.h"
41 : #include "utils/rel.h"
42 :
43 : /*
44 : * DSM keys for parallel vacuum. Unlike other parallel execution code, since
45 : * we don't need to worry about DSM keys conflicting with plan_node_id we can
46 : * use small integers.
47 : */
48 : #define PARALLEL_VACUUM_KEY_SHARED 1
49 : #define PARALLEL_VACUUM_KEY_QUERY_TEXT 2
50 : #define PARALLEL_VACUUM_KEY_BUFFER_USAGE 3
51 : #define PARALLEL_VACUUM_KEY_WAL_USAGE 4
52 : #define PARALLEL_VACUUM_KEY_INDEX_STATS 5
53 :
54 : /*
55 : * Shared information among parallel workers. So this is allocated in the DSM
56 : * segment.
57 : */
58 : typedef struct PVShared
59 : {
60 : /*
61 : * Target table relid, log level (for messages about parallel workers
62 : * launched during VACUUM VERBOSE) and query ID. These fields are not
63 : * modified during the parallel vacuum.
64 : */
65 : Oid relid;
66 : int elevel;
67 : int64 queryid;
68 :
69 : /*
70 : * Fields for both index vacuum and cleanup.
71 : *
72 : * reltuples is the total number of input heap tuples. We set either old
73 : * live tuples in the index vacuum case or the new live tuples in the
74 : * index cleanup case.
75 : *
76 : * estimated_count is true if reltuples is an estimated value. (Note that
77 : * reltuples could be -1 in this case, indicating we have no idea.)
78 : */
79 : double reltuples;
80 : bool estimated_count;
81 :
82 : /*
83 : * In single process vacuum we could consume more memory during index
84 : * vacuuming or cleanup apart from the memory for heap scanning. In
85 : * parallel vacuum, since individual vacuum workers can consume memory
86 : * equal to maintenance_work_mem, the new maintenance_work_mem for each
87 : * worker is set such that the parallel operation doesn't consume more
88 : * memory than single process vacuum.
89 : */
90 : int maintenance_work_mem_worker;
91 :
92 : /*
93 : * The number of buffers each worker's Buffer Access Strategy ring should
94 : * contain.
95 : */
96 : int ring_nbuffers;
97 :
98 : /*
99 : * Shared vacuum cost balance. During parallel vacuum,
100 : * VacuumSharedCostBalance points to this value and it accumulates the
101 : * balance of each parallel vacuum worker.
102 : */
103 : pg_atomic_uint32 cost_balance;
104 :
105 : /*
106 : * Number of active parallel workers. This is used for computing the
107 : * minimum threshold of the vacuum cost balance before a worker sleeps for
108 : * cost-based delay.
109 : */
110 : pg_atomic_uint32 active_nworkers;
111 :
112 : /* Counter for vacuuming and cleanup */
113 : pg_atomic_uint32 idx;
114 :
115 : /* DSA handle where the TidStore lives */
116 : dsa_handle dead_items_dsa_handle;
117 :
118 : /* DSA pointer to the shared TidStore */
119 : dsa_pointer dead_items_handle;
120 :
121 : /* Statistics of shared dead items */
122 : VacDeadItemsInfo dead_items_info;
123 : } PVShared;
124 :
125 : /* Status used during parallel index vacuum or cleanup */
126 : typedef enum PVIndVacStatus
127 : {
128 : PARALLEL_INDVAC_STATUS_INITIAL = 0,
129 : PARALLEL_INDVAC_STATUS_NEED_BULKDELETE,
130 : PARALLEL_INDVAC_STATUS_NEED_CLEANUP,
131 : PARALLEL_INDVAC_STATUS_COMPLETED,
132 : } PVIndVacStatus;
133 :
134 : /*
135 : * Struct for index vacuum statistics of an index that is used for parallel vacuum.
136 : * This includes the status of parallel index vacuum as well as index statistics.
137 : */
138 : typedef struct PVIndStats
139 : {
140 : /*
141 : * The following two fields are set by leader process before executing
142 : * parallel index vacuum or parallel index cleanup. These fields are not
143 : * fixed for the entire VACUUM operation. They are only fixed for an
144 : * individual parallel index vacuum and cleanup.
145 : *
146 : * parallel_workers_can_process is true if both leader and worker can
147 : * process the index, otherwise only leader can process it.
148 : */
149 : PVIndVacStatus status;
150 : bool parallel_workers_can_process;
151 :
152 : /*
153 : * Individual worker or leader stores the result of index vacuum or
154 : * cleanup.
155 : */
156 : bool istat_updated; /* are the stats updated? */
157 : IndexBulkDeleteResult istat;
158 : } PVIndStats;
159 :
160 : /*
161 : * Struct for maintaining a parallel vacuum state. typedef appears in vacuum.h.
162 : */
163 : struct ParallelVacuumState
164 : {
165 : /* NULL for worker processes */
166 : ParallelContext *pcxt;
167 :
168 : /* Parent Heap Relation */
169 : Relation heaprel;
170 :
171 : /* Target indexes */
172 : Relation *indrels;
173 : int nindexes;
174 :
175 : /* Shared information among parallel vacuum workers */
176 : PVShared *shared;
177 :
178 : /*
179 : * Shared index statistics among parallel vacuum workers. The array
180 : * element is allocated for every index, even those indexes where parallel
181 : * index vacuuming is unsafe or not worthwhile (e.g.,
182 : * will_parallel_vacuum[] is false). During parallel vacuum,
183 : * IndexBulkDeleteResult of each index is kept in DSM and is copied into
184 : * local memory at the end of parallel vacuum.
185 : */
186 : PVIndStats *indstats;
187 :
188 : /* Shared dead items space among parallel vacuum workers */
189 : TidStore *dead_items;
190 :
191 : /* Points to buffer usage area in DSM */
192 : BufferUsage *buffer_usage;
193 :
194 : /* Points to WAL usage area in DSM */
195 : WalUsage *wal_usage;
196 :
197 : /*
198 : * False if the index is totally unsuitable target for all parallel
199 : * processing. For example, the index could be <
200 : * min_parallel_index_scan_size cutoff.
201 : */
202 : bool *will_parallel_vacuum;
203 :
204 : /*
205 : * The number of indexes that support parallel index bulk-deletion and
206 : * parallel index cleanup respectively.
207 : */
208 : int nindexes_parallel_bulkdel;
209 : int nindexes_parallel_cleanup;
210 : int nindexes_parallel_condcleanup;
211 :
212 : /* Buffer access strategy used by leader process */
213 : BufferAccessStrategy bstrategy;
214 :
215 : /*
216 : * Error reporting state. The error callback is set only for workers
217 : * processes during parallel index vacuum.
218 : */
219 : char *relnamespace;
220 : char *relname;
221 : char *indname;
222 : PVIndVacStatus status;
223 : };
224 :
225 : static int parallel_vacuum_compute_workers(Relation *indrels, int nindexes, int nrequested,
226 : bool *will_parallel_vacuum);
227 : static void parallel_vacuum_process_all_indexes(ParallelVacuumState *pvs, int num_index_scans,
228 : bool vacuum);
229 : static void parallel_vacuum_process_safe_indexes(ParallelVacuumState *pvs);
230 : static void parallel_vacuum_process_unsafe_indexes(ParallelVacuumState *pvs);
231 : static void parallel_vacuum_process_one_index(ParallelVacuumState *pvs, Relation indrel,
232 : PVIndStats *indstats);
233 : static bool parallel_vacuum_index_is_parallel_safe(Relation indrel, int num_index_scans,
234 : bool vacuum);
235 : static void parallel_vacuum_error_callback(void *arg);
236 :
237 : /*
238 : * Try to enter parallel mode and create a parallel context. Then initialize
239 : * shared memory state.
240 : *
241 : * On success, return parallel vacuum state. Otherwise return NULL.
242 : */
243 : ParallelVacuumState *
244 5530 : parallel_vacuum_init(Relation rel, Relation *indrels, int nindexes,
245 : int nrequested_workers, int vac_work_mem,
246 : int elevel, BufferAccessStrategy bstrategy)
247 : {
248 : ParallelVacuumState *pvs;
249 : ParallelContext *pcxt;
250 : PVShared *shared;
251 : TidStore *dead_items;
252 : PVIndStats *indstats;
253 : BufferUsage *buffer_usage;
254 : WalUsage *wal_usage;
255 : bool *will_parallel_vacuum;
256 : Size est_indstats_len;
257 : Size est_shared_len;
258 5530 : int nindexes_mwm = 0;
259 5530 : int parallel_workers = 0;
260 : int querylen;
261 :
262 : /*
263 : * A parallel vacuum must be requested and there must be indexes on the
264 : * relation
265 : */
266 : Assert(nrequested_workers >= 0);
267 : Assert(nindexes > 0);
268 :
269 : /*
270 : * Compute the number of parallel vacuum workers to launch
271 : */
272 5530 : will_parallel_vacuum = palloc0_array(bool, nindexes);
273 5530 : parallel_workers = parallel_vacuum_compute_workers(indrels, nindexes,
274 : nrequested_workers,
275 : will_parallel_vacuum);
276 5530 : if (parallel_workers <= 0)
277 : {
278 : /* Can't perform vacuum in parallel -- return NULL */
279 5513 : pfree(will_parallel_vacuum);
280 5513 : return NULL;
281 : }
282 :
283 17 : pvs = palloc0_object(ParallelVacuumState);
284 17 : pvs->indrels = indrels;
285 17 : pvs->nindexes = nindexes;
286 17 : pvs->will_parallel_vacuum = will_parallel_vacuum;
287 17 : pvs->bstrategy = bstrategy;
288 17 : pvs->heaprel = rel;
289 :
290 17 : EnterParallelMode();
291 17 : pcxt = CreateParallelContext("postgres", "parallel_vacuum_main",
292 : parallel_workers);
293 : Assert(pcxt->nworkers > 0);
294 17 : pvs->pcxt = pcxt;
295 :
296 : /* Estimate size for index vacuum stats -- PARALLEL_VACUUM_KEY_INDEX_STATS */
297 17 : est_indstats_len = mul_size(sizeof(PVIndStats), nindexes);
298 17 : shm_toc_estimate_chunk(&pcxt->estimator, est_indstats_len);
299 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
300 :
301 : /* Estimate size for shared information -- PARALLEL_VACUUM_KEY_SHARED */
302 17 : est_shared_len = sizeof(PVShared);
303 17 : shm_toc_estimate_chunk(&pcxt->estimator, est_shared_len);
304 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
305 :
306 : /*
307 : * Estimate space for BufferUsage and WalUsage --
308 : * PARALLEL_VACUUM_KEY_BUFFER_USAGE and PARALLEL_VACUUM_KEY_WAL_USAGE.
309 : *
310 : * If there are no extensions loaded that care, we could skip this. We
311 : * have no way of knowing whether anyone's looking at pgBufferUsage or
312 : * pgWalUsage, so do it unconditionally.
313 : */
314 17 : shm_toc_estimate_chunk(&pcxt->estimator,
315 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
316 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
317 17 : shm_toc_estimate_chunk(&pcxt->estimator,
318 : mul_size(sizeof(WalUsage), pcxt->nworkers));
319 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
320 :
321 : /* Finally, estimate PARALLEL_VACUUM_KEY_QUERY_TEXT space */
322 17 : if (debug_query_string)
323 : {
324 17 : querylen = strlen(debug_query_string);
325 17 : shm_toc_estimate_chunk(&pcxt->estimator, querylen + 1);
326 17 : shm_toc_estimate_keys(&pcxt->estimator, 1);
327 : }
328 : else
329 0 : querylen = 0; /* keep compiler quiet */
330 :
331 17 : InitializeParallelDSM(pcxt);
332 :
333 : /* Prepare index vacuum stats */
334 17 : indstats = (PVIndStats *) shm_toc_allocate(pcxt->toc, est_indstats_len);
335 383 : MemSet(indstats, 0, est_indstats_len);
336 78 : for (int i = 0; i < nindexes; i++)
337 : {
338 61 : Relation indrel = indrels[i];
339 61 : uint8 vacoptions = indrel->rd_indam->amparallelvacuumoptions;
340 :
341 : /*
342 : * Cleanup option should be either disabled, always performing in
343 : * parallel or conditionally performing in parallel.
344 : */
345 : Assert(((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) == 0) ||
346 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) == 0));
347 : Assert(vacoptions <= VACUUM_OPTION_MAX_VALID_VALUE);
348 :
349 61 : if (!will_parallel_vacuum[i])
350 3 : continue;
351 :
352 58 : if (indrel->rd_indam->amusemaintenanceworkmem)
353 6 : nindexes_mwm++;
354 :
355 : /*
356 : * Remember the number of indexes that support parallel operation for
357 : * each phase.
358 : */
359 58 : if ((vacoptions & VACUUM_OPTION_PARALLEL_BULKDEL) != 0)
360 52 : pvs->nindexes_parallel_bulkdel++;
361 58 : if ((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) != 0)
362 12 : pvs->nindexes_parallel_cleanup++;
363 58 : if ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) != 0)
364 40 : pvs->nindexes_parallel_condcleanup++;
365 : }
366 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_INDEX_STATS, indstats);
367 17 : pvs->indstats = indstats;
368 :
369 : /* Prepare shared information */
370 17 : shared = (PVShared *) shm_toc_allocate(pcxt->toc, est_shared_len);
371 187 : MemSet(shared, 0, est_shared_len);
372 17 : shared->relid = RelationGetRelid(rel);
373 17 : shared->elevel = elevel;
374 17 : shared->queryid = pgstat_get_my_query_id();
375 17 : shared->maintenance_work_mem_worker =
376 : (nindexes_mwm > 0) ?
377 17 : maintenance_work_mem / Min(parallel_workers, nindexes_mwm) :
378 : maintenance_work_mem;
379 17 : shared->dead_items_info.max_bytes = vac_work_mem * (size_t) 1024;
380 :
381 : /* Prepare DSA space for dead items */
382 17 : dead_items = TidStoreCreateShared(shared->dead_items_info.max_bytes,
383 : LWTRANCHE_PARALLEL_VACUUM_DSA);
384 17 : pvs->dead_items = dead_items;
385 17 : shared->dead_items_handle = TidStoreGetHandle(dead_items);
386 17 : shared->dead_items_dsa_handle = dsa_get_handle(TidStoreGetDSA(dead_items));
387 :
388 : /* Use the same buffer size for all workers */
389 17 : shared->ring_nbuffers = GetAccessStrategyBufferCount(bstrategy);
390 :
391 17 : pg_atomic_init_u32(&(shared->cost_balance), 0);
392 17 : pg_atomic_init_u32(&(shared->active_nworkers), 0);
393 17 : pg_atomic_init_u32(&(shared->idx), 0);
394 :
395 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_SHARED, shared);
396 17 : pvs->shared = shared;
397 :
398 : /*
399 : * Allocate space for each worker's BufferUsage and WalUsage; no need to
400 : * initialize
401 : */
402 17 : buffer_usage = shm_toc_allocate(pcxt->toc,
403 17 : mul_size(sizeof(BufferUsage), pcxt->nworkers));
404 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_BUFFER_USAGE, buffer_usage);
405 17 : pvs->buffer_usage = buffer_usage;
406 17 : wal_usage = shm_toc_allocate(pcxt->toc,
407 17 : mul_size(sizeof(WalUsage), pcxt->nworkers));
408 17 : shm_toc_insert(pcxt->toc, PARALLEL_VACUUM_KEY_WAL_USAGE, wal_usage);
409 17 : pvs->wal_usage = wal_usage;
410 :
411 : /* Store query string for workers */
412 17 : if (debug_query_string)
413 : {
414 : char *sharedquery;
415 :
416 17 : sharedquery = (char *) shm_toc_allocate(pcxt->toc, querylen + 1);
417 17 : memcpy(sharedquery, debug_query_string, querylen + 1);
418 17 : sharedquery[querylen] = '\0';
419 17 : shm_toc_insert(pcxt->toc,
420 : PARALLEL_VACUUM_KEY_QUERY_TEXT, sharedquery);
421 : }
422 :
423 : /* Success -- return parallel vacuum state */
424 17 : return pvs;
425 : }
426 :
427 : /*
428 : * Destroy the parallel context, and end parallel mode.
429 : *
430 : * Since writes are not allowed during parallel mode, copy the
431 : * updated index statistics from DSM into local memory and then later use that
432 : * to update the index statistics. One might think that we can exit from
433 : * parallel mode, update the index statistics and then destroy parallel
434 : * context, but that won't be safe (see ExitParallelMode).
435 : */
436 : void
437 17 : parallel_vacuum_end(ParallelVacuumState *pvs, IndexBulkDeleteResult **istats)
438 : {
439 : Assert(!IsParallelWorker());
440 :
441 : /* Copy the updated statistics */
442 78 : for (int i = 0; i < pvs->nindexes; i++)
443 : {
444 61 : PVIndStats *indstats = &(pvs->indstats[i]);
445 :
446 61 : if (indstats->istat_updated)
447 : {
448 40 : istats[i] = palloc0_object(IndexBulkDeleteResult);
449 40 : memcpy(istats[i], &indstats->istat, sizeof(IndexBulkDeleteResult));
450 : }
451 : else
452 21 : istats[i] = NULL;
453 : }
454 :
455 17 : TidStoreDestroy(pvs->dead_items);
456 :
457 17 : DestroyParallelContext(pvs->pcxt);
458 17 : ExitParallelMode();
459 :
460 17 : pfree(pvs->will_parallel_vacuum);
461 17 : pfree(pvs);
462 17 : }
463 :
464 : /*
465 : * Returns the dead items space and dead items information.
466 : */
467 : TidStore *
468 32 : parallel_vacuum_get_dead_items(ParallelVacuumState *pvs, VacDeadItemsInfo **dead_items_info_p)
469 : {
470 32 : *dead_items_info_p = &(pvs->shared->dead_items_info);
471 32 : return pvs->dead_items;
472 : }
473 :
474 : /* Forget all items in dead_items */
475 : void
476 15 : parallel_vacuum_reset_dead_items(ParallelVacuumState *pvs)
477 : {
478 15 : VacDeadItemsInfo *dead_items_info = &(pvs->shared->dead_items_info);
479 :
480 : /*
481 : * Free the current tidstore and return allocated DSA segments to the
482 : * operating system. Then we recreate the tidstore with the same max_bytes
483 : * limitation we just used.
484 : */
485 15 : TidStoreDestroy(pvs->dead_items);
486 15 : pvs->dead_items = TidStoreCreateShared(dead_items_info->max_bytes,
487 : LWTRANCHE_PARALLEL_VACUUM_DSA);
488 :
489 : /* Update the DSA pointer for dead_items to the new one */
490 15 : pvs->shared->dead_items_dsa_handle = dsa_get_handle(TidStoreGetDSA(pvs->dead_items));
491 15 : pvs->shared->dead_items_handle = TidStoreGetHandle(pvs->dead_items);
492 :
493 : /* Reset the counter */
494 15 : dead_items_info->num_items = 0;
495 15 : }
496 :
497 : /*
498 : * Do parallel index bulk-deletion with parallel workers.
499 : */
500 : void
501 15 : parallel_vacuum_bulkdel_all_indexes(ParallelVacuumState *pvs, long num_table_tuples,
502 : int num_index_scans)
503 : {
504 : Assert(!IsParallelWorker());
505 :
506 : /*
507 : * We can only provide an approximate value of num_heap_tuples, at least
508 : * for now.
509 : */
510 15 : pvs->shared->reltuples = num_table_tuples;
511 15 : pvs->shared->estimated_count = true;
512 :
513 15 : parallel_vacuum_process_all_indexes(pvs, num_index_scans, true);
514 15 : }
515 :
516 : /*
517 : * Do parallel index cleanup with parallel workers.
518 : */
519 : void
520 17 : parallel_vacuum_cleanup_all_indexes(ParallelVacuumState *pvs, long num_table_tuples,
521 : int num_index_scans, bool estimated_count)
522 : {
523 : Assert(!IsParallelWorker());
524 :
525 : /*
526 : * We can provide a better estimate of total number of surviving tuples
527 : * (we assume indexes are more interested in that than in the number of
528 : * nominally live tuples).
529 : */
530 17 : pvs->shared->reltuples = num_table_tuples;
531 17 : pvs->shared->estimated_count = estimated_count;
532 :
533 17 : parallel_vacuum_process_all_indexes(pvs, num_index_scans, false);
534 17 : }
535 :
536 : /*
537 : * Compute the number of parallel worker processes to request. Both index
538 : * vacuum and index cleanup can be executed with parallel workers.
539 : * The index is eligible for parallel vacuum iff its size is greater than
540 : * min_parallel_index_scan_size as invoking workers for very small indexes
541 : * can hurt performance.
542 : *
543 : * nrequested is the number of parallel workers that user requested. If
544 : * nrequested is 0, we compute the parallel degree based on nindexes, that is
545 : * the number of indexes that support parallel vacuum. This function also
546 : * sets will_parallel_vacuum to remember indexes that participate in parallel
547 : * vacuum.
548 : */
549 : static int
550 5530 : parallel_vacuum_compute_workers(Relation *indrels, int nindexes, int nrequested,
551 : bool *will_parallel_vacuum)
552 : {
553 5530 : int nindexes_parallel = 0;
554 5530 : int nindexes_parallel_bulkdel = 0;
555 5530 : int nindexes_parallel_cleanup = 0;
556 : int parallel_workers;
557 :
558 : /*
559 : * We don't allow performing parallel operation in standalone backend or
560 : * when parallelism is disabled.
561 : */
562 5530 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
563 2401 : return 0;
564 :
565 : /*
566 : * Compute the number of indexes that can participate in parallel vacuum.
567 : */
568 10205 : for (int i = 0; i < nindexes; i++)
569 : {
570 7076 : Relation indrel = indrels[i];
571 7076 : uint8 vacoptions = indrel->rd_indam->amparallelvacuumoptions;
572 :
573 : /* Skip index that is not a suitable target for parallel index vacuum */
574 7076 : if (vacoptions == VACUUM_OPTION_NO_PARALLEL ||
575 7076 : RelationGetNumberOfBlocks(indrel) < min_parallel_index_scan_size)
576 7009 : continue;
577 :
578 67 : will_parallel_vacuum[i] = true;
579 :
580 67 : if ((vacoptions & VACUUM_OPTION_PARALLEL_BULKDEL) != 0)
581 61 : nindexes_parallel_bulkdel++;
582 67 : if (((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) != 0) ||
583 55 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) != 0))
584 61 : nindexes_parallel_cleanup++;
585 : }
586 :
587 3129 : nindexes_parallel = Max(nindexes_parallel_bulkdel,
588 : nindexes_parallel_cleanup);
589 :
590 : /* The leader process takes one index */
591 3129 : nindexes_parallel--;
592 :
593 : /* No index supports parallel vacuum */
594 3129 : if (nindexes_parallel <= 0)
595 3112 : return 0;
596 :
597 : /* Compute the parallel degree */
598 17 : parallel_workers = (nrequested > 0) ?
599 17 : Min(nrequested, nindexes_parallel) : nindexes_parallel;
600 :
601 : /* Cap by max_parallel_maintenance_workers */
602 17 : parallel_workers = Min(parallel_workers, max_parallel_maintenance_workers);
603 :
604 17 : return parallel_workers;
605 : }
606 :
607 : /*
608 : * Perform index vacuum or index cleanup with parallel workers. This function
609 : * must be used by the parallel vacuum leader process.
610 : */
611 : static void
612 32 : parallel_vacuum_process_all_indexes(ParallelVacuumState *pvs, int num_index_scans,
613 : bool vacuum)
614 : {
615 : int nworkers;
616 : PVIndVacStatus new_status;
617 :
618 : Assert(!IsParallelWorker());
619 :
620 32 : if (vacuum)
621 : {
622 15 : new_status = PARALLEL_INDVAC_STATUS_NEED_BULKDELETE;
623 :
624 : /* Determine the number of parallel workers to launch */
625 15 : nworkers = pvs->nindexes_parallel_bulkdel;
626 : }
627 : else
628 : {
629 17 : new_status = PARALLEL_INDVAC_STATUS_NEED_CLEANUP;
630 :
631 : /* Determine the number of parallel workers to launch */
632 17 : nworkers = pvs->nindexes_parallel_cleanup;
633 :
634 : /* Add conditionally parallel-aware indexes if in the first time call */
635 17 : if (num_index_scans == 0)
636 10 : nworkers += pvs->nindexes_parallel_condcleanup;
637 : }
638 :
639 : /* The leader process will participate */
640 32 : nworkers--;
641 :
642 : /*
643 : * It is possible that parallel context is initialized with fewer workers
644 : * than the number of indexes that need a separate worker in the current
645 : * phase, so we need to consider it. See
646 : * parallel_vacuum_compute_workers().
647 : */
648 32 : nworkers = Min(nworkers, pvs->pcxt->nworkers);
649 :
650 : /*
651 : * Set index vacuum status and mark whether parallel vacuum worker can
652 : * process it.
653 : */
654 129 : for (int i = 0; i < pvs->nindexes; i++)
655 : {
656 97 : PVIndStats *indstats = &(pvs->indstats[i]);
657 :
658 : Assert(indstats->status == PARALLEL_INDVAC_STATUS_INITIAL);
659 97 : indstats->status = new_status;
660 97 : indstats->parallel_workers_can_process =
661 189 : (pvs->will_parallel_vacuum[i] &&
662 92 : parallel_vacuum_index_is_parallel_safe(pvs->indrels[i],
663 : num_index_scans,
664 97 : vacuum));
665 : }
666 :
667 : /* Reset the parallel index processing and progress counters */
668 32 : pg_atomic_write_u32(&(pvs->shared->idx), 0);
669 :
670 : /* Setup the shared cost-based vacuum delay and launch workers */
671 32 : if (nworkers > 0)
672 : {
673 : /* Reinitialize parallel context to relaunch parallel workers */
674 26 : if (num_index_scans > 0)
675 9 : ReinitializeParallelDSM(pvs->pcxt);
676 :
677 : /*
678 : * Set up shared cost balance and the number of active workers for
679 : * vacuum delay. We need to do this before launching workers as
680 : * otherwise, they might not see the updated values for these
681 : * parameters.
682 : */
683 26 : pg_atomic_write_u32(&(pvs->shared->cost_balance), VacuumCostBalance);
684 26 : pg_atomic_write_u32(&(pvs->shared->active_nworkers), 0);
685 :
686 : /*
687 : * The number of workers can vary between bulkdelete and cleanup
688 : * phase.
689 : */
690 26 : ReinitializeParallelWorkers(pvs->pcxt, nworkers);
691 :
692 26 : LaunchParallelWorkers(pvs->pcxt);
693 :
694 26 : if (pvs->pcxt->nworkers_launched > 0)
695 : {
696 : /*
697 : * Reset the local cost values for leader backend as we have
698 : * already accumulated the remaining balance of heap.
699 : */
700 26 : VacuumCostBalance = 0;
701 26 : VacuumCostBalanceLocal = 0;
702 :
703 : /* Enable shared cost balance for leader backend */
704 26 : VacuumSharedCostBalance = &(pvs->shared->cost_balance);
705 26 : VacuumActiveNWorkers = &(pvs->shared->active_nworkers);
706 : }
707 :
708 26 : if (vacuum)
709 15 : ereport(pvs->shared->elevel,
710 : (errmsg(ngettext("launched %d parallel vacuum worker for index vacuuming (planned: %d)",
711 : "launched %d parallel vacuum workers for index vacuuming (planned: %d)",
712 : pvs->pcxt->nworkers_launched),
713 : pvs->pcxt->nworkers_launched, nworkers)));
714 : else
715 11 : ereport(pvs->shared->elevel,
716 : (errmsg(ngettext("launched %d parallel vacuum worker for index cleanup (planned: %d)",
717 : "launched %d parallel vacuum workers for index cleanup (planned: %d)",
718 : pvs->pcxt->nworkers_launched),
719 : pvs->pcxt->nworkers_launched, nworkers)));
720 : }
721 :
722 : /* Vacuum the indexes that can be processed by only leader process */
723 32 : parallel_vacuum_process_unsafe_indexes(pvs);
724 :
725 : /*
726 : * Join as a parallel worker. The leader vacuums alone processes all
727 : * parallel-safe indexes in the case where no workers are launched.
728 : */
729 32 : parallel_vacuum_process_safe_indexes(pvs);
730 :
731 : /*
732 : * Next, accumulate buffer and WAL usage. (This must wait for the workers
733 : * to finish, or we might get incomplete data.)
734 : */
735 32 : if (nworkers > 0)
736 : {
737 : /* Wait for all vacuum workers to finish */
738 26 : WaitForParallelWorkersToFinish(pvs->pcxt);
739 :
740 58 : for (int i = 0; i < pvs->pcxt->nworkers_launched; i++)
741 32 : InstrAccumParallelQuery(&pvs->buffer_usage[i], &pvs->wal_usage[i]);
742 : }
743 :
744 : /*
745 : * Reset all index status back to initial (while checking that we have
746 : * vacuumed all indexes).
747 : */
748 129 : for (int i = 0; i < pvs->nindexes; i++)
749 : {
750 97 : PVIndStats *indstats = &(pvs->indstats[i]);
751 :
752 97 : if (indstats->status != PARALLEL_INDVAC_STATUS_COMPLETED)
753 0 : elog(ERROR, "parallel index vacuum on index \"%s\" is not completed",
754 : RelationGetRelationName(pvs->indrels[i]));
755 :
756 97 : indstats->status = PARALLEL_INDVAC_STATUS_INITIAL;
757 : }
758 :
759 : /*
760 : * Carry the shared balance value to heap scan and disable shared costing
761 : */
762 32 : if (VacuumSharedCostBalance)
763 : {
764 26 : VacuumCostBalance = pg_atomic_read_u32(VacuumSharedCostBalance);
765 26 : VacuumSharedCostBalance = NULL;
766 26 : VacuumActiveNWorkers = NULL;
767 : }
768 32 : }
769 :
770 : /*
771 : * Index vacuum/cleanup routine used by the leader process and parallel
772 : * vacuum worker processes to vacuum the indexes in parallel.
773 : */
774 : static void
775 64 : parallel_vacuum_process_safe_indexes(ParallelVacuumState *pvs)
776 : {
777 : /*
778 : * Increment the active worker count if we are able to launch any worker.
779 : */
780 64 : if (VacuumActiveNWorkers)
781 58 : pg_atomic_add_fetch_u32(VacuumActiveNWorkers, 1);
782 :
783 : /* Loop until all indexes are vacuumed */
784 : for (;;)
785 97 : {
786 : int idx;
787 : PVIndStats *indstats;
788 :
789 : /* Get an index number to process */
790 161 : idx = pg_atomic_fetch_add_u32(&(pvs->shared->idx), 1);
791 :
792 : /* Done for all indexes? */
793 161 : if (idx >= pvs->nindexes)
794 64 : break;
795 :
796 97 : indstats = &(pvs->indstats[idx]);
797 :
798 : /*
799 : * Skip vacuuming index that is unsafe for workers or has an
800 : * unsuitable target for parallel index vacuum (this is vacuumed in
801 : * parallel_vacuum_process_unsafe_indexes() by the leader).
802 : */
803 97 : if (!indstats->parallel_workers_can_process)
804 27 : continue;
805 :
806 : /* Do vacuum or cleanup of the index */
807 70 : parallel_vacuum_process_one_index(pvs, pvs->indrels[idx], indstats);
808 : }
809 :
810 : /*
811 : * We have completed the index vacuum so decrement the active worker
812 : * count.
813 : */
814 64 : if (VacuumActiveNWorkers)
815 58 : pg_atomic_sub_fetch_u32(VacuumActiveNWorkers, 1);
816 64 : }
817 :
818 : /*
819 : * Perform parallel vacuuming of indexes in leader process.
820 : *
821 : * Handles index vacuuming (or index cleanup) for indexes that are not
822 : * parallel safe. It's possible that this will vary for a given index, based
823 : * on details like whether we're performing index cleanup right now.
824 : *
825 : * Also performs vacuuming of smaller indexes that fell under the size cutoff
826 : * enforced by parallel_vacuum_compute_workers().
827 : */
828 : static void
829 32 : parallel_vacuum_process_unsafe_indexes(ParallelVacuumState *pvs)
830 : {
831 : Assert(!IsParallelWorker());
832 :
833 : /*
834 : * Increment the active worker count if we are able to launch any worker.
835 : */
836 32 : if (VacuumActiveNWorkers)
837 26 : pg_atomic_add_fetch_u32(VacuumActiveNWorkers, 1);
838 :
839 129 : for (int i = 0; i < pvs->nindexes; i++)
840 : {
841 97 : PVIndStats *indstats = &(pvs->indstats[i]);
842 :
843 : /* Skip, indexes that are safe for workers */
844 97 : if (indstats->parallel_workers_can_process)
845 70 : continue;
846 :
847 : /* Do vacuum or cleanup of the index */
848 27 : parallel_vacuum_process_one_index(pvs, pvs->indrels[i], indstats);
849 : }
850 :
851 : /*
852 : * We have completed the index vacuum so decrement the active worker
853 : * count.
854 : */
855 32 : if (VacuumActiveNWorkers)
856 26 : pg_atomic_sub_fetch_u32(VacuumActiveNWorkers, 1);
857 32 : }
858 :
859 : /*
860 : * Vacuum or cleanup index either by leader process or by one of the worker
861 : * process. After vacuuming the index this function copies the index
862 : * statistics returned from ambulkdelete and amvacuumcleanup to the DSM
863 : * segment.
864 : */
865 : static void
866 97 : parallel_vacuum_process_one_index(ParallelVacuumState *pvs, Relation indrel,
867 : PVIndStats *indstats)
868 : {
869 97 : IndexBulkDeleteResult *istat = NULL;
870 : IndexBulkDeleteResult *istat_res;
871 : IndexVacuumInfo ivinfo;
872 :
873 : /*
874 : * Update the pointer to the corresponding bulk-deletion result if someone
875 : * has already updated it
876 : */
877 97 : if (indstats->istat_updated)
878 36 : istat = &(indstats->istat);
879 :
880 97 : ivinfo.index = indrel;
881 97 : ivinfo.heaprel = pvs->heaprel;
882 97 : ivinfo.analyze_only = false;
883 97 : ivinfo.report_progress = false;
884 97 : ivinfo.message_level = DEBUG2;
885 97 : ivinfo.estimated_count = pvs->shared->estimated_count;
886 97 : ivinfo.num_heap_tuples = pvs->shared->reltuples;
887 97 : ivinfo.strategy = pvs->bstrategy;
888 :
889 : /* Update error traceback information */
890 97 : pvs->indname = pstrdup(RelationGetRelationName(indrel));
891 97 : pvs->status = indstats->status;
892 :
893 97 : switch (indstats->status)
894 : {
895 36 : case PARALLEL_INDVAC_STATUS_NEED_BULKDELETE:
896 36 : istat_res = vac_bulkdel_one_index(&ivinfo, istat, pvs->dead_items,
897 36 : &pvs->shared->dead_items_info);
898 36 : break;
899 61 : case PARALLEL_INDVAC_STATUS_NEED_CLEANUP:
900 61 : istat_res = vac_cleanup_one_index(&ivinfo, istat);
901 61 : break;
902 0 : default:
903 0 : elog(ERROR, "unexpected parallel vacuum index status %d for index \"%s\"",
904 : indstats->status,
905 : RelationGetRelationName(indrel));
906 : }
907 :
908 : /*
909 : * Copy the index bulk-deletion result returned from ambulkdelete and
910 : * amvacuumcleanup to the DSM segment if it's the first cycle because they
911 : * allocate locally and it's possible that an index will be vacuumed by a
912 : * different vacuum process the next cycle. Copying the result normally
913 : * happens only the first time an index is vacuumed. For any additional
914 : * vacuum pass, we directly point to the result on the DSM segment and
915 : * pass it to vacuum index APIs so that workers can update it directly.
916 : *
917 : * Since all vacuum workers write the bulk-deletion result at different
918 : * slots we can write them without locking.
919 : */
920 97 : if (!indstats->istat_updated && istat_res != NULL)
921 : {
922 40 : memcpy(&(indstats->istat), istat_res, sizeof(IndexBulkDeleteResult));
923 40 : indstats->istat_updated = true;
924 :
925 : /* Free the locally-allocated bulk-deletion result */
926 40 : pfree(istat_res);
927 : }
928 :
929 : /*
930 : * Update the status to completed. No need to lock here since each worker
931 : * touches different indexes.
932 : */
933 97 : indstats->status = PARALLEL_INDVAC_STATUS_COMPLETED;
934 :
935 : /* Reset error traceback information */
936 97 : pvs->status = PARALLEL_INDVAC_STATUS_COMPLETED;
937 97 : pfree(pvs->indname);
938 97 : pvs->indname = NULL;
939 :
940 : /*
941 : * Call the parallel variant of pgstat_progress_incr_param so workers can
942 : * report progress of index vacuum to the leader.
943 : */
944 97 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_INDEXES_PROCESSED, 1);
945 97 : }
946 :
947 : /*
948 : * Returns false, if the given index can't participate in the next execution of
949 : * parallel index vacuum or parallel index cleanup.
950 : */
951 : static bool
952 92 : parallel_vacuum_index_is_parallel_safe(Relation indrel, int num_index_scans,
953 : bool vacuum)
954 : {
955 : uint8 vacoptions;
956 :
957 92 : vacoptions = indrel->rd_indam->amparallelvacuumoptions;
958 :
959 : /* In parallel vacuum case, check if it supports parallel bulk-deletion */
960 92 : if (vacuum)
961 34 : return ((vacoptions & VACUUM_OPTION_PARALLEL_BULKDEL) != 0);
962 :
963 : /* Not safe, if the index does not support parallel cleanup */
964 58 : if (((vacoptions & VACUUM_OPTION_PARALLEL_CLEANUP) == 0) &&
965 46 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) == 0))
966 6 : return false;
967 :
968 : /*
969 : * Not safe, if the index supports parallel cleanup conditionally, but we
970 : * have already processed the index (for bulkdelete). We do this to avoid
971 : * the need to invoke workers when parallel index cleanup doesn't need to
972 : * scan the index. See the comments for option
973 : * VACUUM_OPTION_PARALLEL_COND_CLEANUP to know when indexes support
974 : * parallel cleanup conditionally.
975 : */
976 52 : if (num_index_scans > 0 &&
977 17 : ((vacoptions & VACUUM_OPTION_PARALLEL_COND_CLEANUP) != 0))
978 15 : return false;
979 :
980 37 : return true;
981 : }
982 :
983 : /*
984 : * Perform work within a launched parallel process.
985 : *
986 : * Since parallel vacuum workers perform only index vacuum or index cleanup,
987 : * we don't need to report progress information.
988 : */
989 : void
990 32 : parallel_vacuum_main(dsm_segment *seg, shm_toc *toc)
991 : {
992 : ParallelVacuumState pvs;
993 : Relation rel;
994 : Relation *indrels;
995 : PVIndStats *indstats;
996 : PVShared *shared;
997 : TidStore *dead_items;
998 : BufferUsage *buffer_usage;
999 : WalUsage *wal_usage;
1000 : int nindexes;
1001 : char *sharedquery;
1002 : ErrorContextCallback errcallback;
1003 :
1004 : /*
1005 : * A parallel vacuum worker must have only PROC_IN_VACUUM flag since we
1006 : * don't support parallel vacuum for autovacuum as of now.
1007 : */
1008 : Assert(MyProc->statusFlags == PROC_IN_VACUUM);
1009 :
1010 32 : elog(DEBUG1, "starting parallel vacuum worker");
1011 :
1012 32 : shared = (PVShared *) shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_SHARED, false);
1013 :
1014 : /* Set debug_query_string for individual workers */
1015 32 : sharedquery = shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_QUERY_TEXT, true);
1016 32 : debug_query_string = sharedquery;
1017 32 : pgstat_report_activity(STATE_RUNNING, debug_query_string);
1018 :
1019 : /* Track query ID */
1020 32 : pgstat_report_query_id(shared->queryid, false);
1021 :
1022 : /*
1023 : * Open table. The lock mode is the same as the leader process. It's
1024 : * okay because the lock mode does not conflict among the parallel
1025 : * workers.
1026 : */
1027 32 : rel = table_open(shared->relid, ShareUpdateExclusiveLock);
1028 :
1029 : /*
1030 : * Open all indexes. indrels are sorted in order by OID, which should be
1031 : * matched to the leader's one.
1032 : */
1033 32 : vac_open_indexes(rel, RowExclusiveLock, &nindexes, &indrels);
1034 : Assert(nindexes > 0);
1035 :
1036 : /*
1037 : * Apply the desired value of maintenance_work_mem within this process.
1038 : * Really we should use SetConfigOption() to change a GUC, but since we're
1039 : * already in parallel mode guc.c would complain about that. Fortunately,
1040 : * by the same token guc.c will not let any user-defined code change it.
1041 : * So just avert your eyes while we do this:
1042 : */
1043 32 : if (shared->maintenance_work_mem_worker > 0)
1044 32 : maintenance_work_mem = shared->maintenance_work_mem_worker;
1045 :
1046 : /* Set index statistics */
1047 32 : indstats = (PVIndStats *) shm_toc_lookup(toc,
1048 : PARALLEL_VACUUM_KEY_INDEX_STATS,
1049 : false);
1050 :
1051 : /* Find dead_items in shared memory */
1052 32 : dead_items = TidStoreAttach(shared->dead_items_dsa_handle,
1053 : shared->dead_items_handle);
1054 :
1055 : /* Set cost-based vacuum delay */
1056 32 : VacuumUpdateCosts();
1057 32 : VacuumCostBalance = 0;
1058 32 : VacuumCostBalanceLocal = 0;
1059 32 : VacuumSharedCostBalance = &(shared->cost_balance);
1060 32 : VacuumActiveNWorkers = &(shared->active_nworkers);
1061 :
1062 : /* Set parallel vacuum state */
1063 32 : pvs.indrels = indrels;
1064 32 : pvs.nindexes = nindexes;
1065 32 : pvs.indstats = indstats;
1066 32 : pvs.shared = shared;
1067 32 : pvs.dead_items = dead_items;
1068 32 : pvs.relnamespace = get_namespace_name(RelationGetNamespace(rel));
1069 32 : pvs.relname = pstrdup(RelationGetRelationName(rel));
1070 32 : pvs.heaprel = rel;
1071 :
1072 : /* These fields will be filled during index vacuum or cleanup */
1073 32 : pvs.indname = NULL;
1074 32 : pvs.status = PARALLEL_INDVAC_STATUS_INITIAL;
1075 :
1076 : /* Each parallel VACUUM worker gets its own access strategy. */
1077 64 : pvs.bstrategy = GetAccessStrategyWithSize(BAS_VACUUM,
1078 32 : shared->ring_nbuffers * (BLCKSZ / 1024));
1079 :
1080 : /* Setup error traceback support for ereport() */
1081 32 : errcallback.callback = parallel_vacuum_error_callback;
1082 32 : errcallback.arg = &pvs;
1083 32 : errcallback.previous = error_context_stack;
1084 32 : error_context_stack = &errcallback;
1085 :
1086 : /* Prepare to track buffer usage during parallel execution */
1087 32 : InstrStartParallelQuery();
1088 :
1089 : /* Process indexes to perform vacuum/cleanup */
1090 32 : parallel_vacuum_process_safe_indexes(&pvs);
1091 :
1092 : /* Report buffer/WAL usage during parallel execution */
1093 32 : buffer_usage = shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_BUFFER_USAGE, false);
1094 32 : wal_usage = shm_toc_lookup(toc, PARALLEL_VACUUM_KEY_WAL_USAGE, false);
1095 32 : InstrEndParallelQuery(&buffer_usage[ParallelWorkerNumber],
1096 32 : &wal_usage[ParallelWorkerNumber]);
1097 :
1098 : /* Report any remaining cost-based vacuum delay time */
1099 32 : if (track_cost_delay_timing)
1100 0 : pgstat_progress_parallel_incr_param(PROGRESS_VACUUM_DELAY_TIME,
1101 : parallel_vacuum_worker_delay_ns);
1102 :
1103 32 : TidStoreDetach(dead_items);
1104 :
1105 : /* Pop the error context stack */
1106 32 : error_context_stack = errcallback.previous;
1107 :
1108 32 : vac_close_indexes(nindexes, indrels, RowExclusiveLock);
1109 32 : table_close(rel, ShareUpdateExclusiveLock);
1110 32 : FreeAccessStrategy(pvs.bstrategy);
1111 32 : }
1112 :
1113 : /*
1114 : * Error context callback for errors occurring during parallel index vacuum.
1115 : * The error context messages should match the messages set in the lazy vacuum
1116 : * error context. If you change this function, change vacuum_error_callback()
1117 : * as well.
1118 : */
1119 : static void
1120 0 : parallel_vacuum_error_callback(void *arg)
1121 : {
1122 0 : ParallelVacuumState *errinfo = arg;
1123 :
1124 0 : switch (errinfo->status)
1125 : {
1126 0 : case PARALLEL_INDVAC_STATUS_NEED_BULKDELETE:
1127 0 : errcontext("while vacuuming index \"%s\" of relation \"%s.%s\"",
1128 : errinfo->indname,
1129 : errinfo->relnamespace,
1130 : errinfo->relname);
1131 0 : break;
1132 0 : case PARALLEL_INDVAC_STATUS_NEED_CLEANUP:
1133 0 : errcontext("while cleaning up index \"%s\" of relation \"%s.%s\"",
1134 : errinfo->indname,
1135 : errinfo->relnamespace,
1136 : errinfo->relname);
1137 0 : break;
1138 0 : case PARALLEL_INDVAC_STATUS_INITIAL:
1139 : case PARALLEL_INDVAC_STATUS_COMPLETED:
1140 : default:
1141 0 : return;
1142 : }
1143 : }
|
