1 /*-------------------------------------------------------------------------
4 * Scan a plan in multiple workers, and do order-preserving merge.
6 * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * src/backend/executor/nodeGatherMerge.c
12 *-------------------------------------------------------------------------
17 #include "executor/executor.h"
18 #include "executor/execParallel.h"
19 #include "executor/nodeGatherMerge.h"
20 #include "executor/tqueue.h"
21 #include "lib/binaryheap.h"
22 #include "miscadmin.h"
23 #include "optimizer/optimizer.h"
26 * When we read tuples from workers, it's a good idea to read several at once
27 * for efficiency when possible: this minimizes context-switching overhead.
28 * But reading too many at a time wastes memory without improving performance.
29 * We'll read up to MAX_TUPLE_STORE tuples (in addition to the first one).
31 #define MAX_TUPLE_STORE 10
34 * Pending-tuple array for each worker. This holds additional tuples that
35 * we were able to fetch from the worker, but can't process yet. In addition,
36 * this struct holds the "done" flag indicating the worker is known to have
37 * no more tuples. (We do not use this struct for the leader; we don't keep
38 * any pending tuples for the leader, and the need_to_scan_locally flag serves
39 * as its "done" indicator.)
41 typedef struct GMReaderTupleBuffer
43 MinimalTuple
*tuple
; /* array of length MAX_TUPLE_STORE */
44 int nTuples
; /* number of tuples currently stored */
45 int readCounter
; /* index of next tuple to extract */
46 bool done
; /* true if reader is known exhausted */
47 } GMReaderTupleBuffer
;
49 static TupleTableSlot
*ExecGatherMerge(PlanState
*pstate
);
50 static int32
heap_compare_slots(Datum a
, Datum b
, void *arg
);
51 static TupleTableSlot
*gather_merge_getnext(GatherMergeState
*gm_state
);
52 static MinimalTuple
gm_readnext_tuple(GatherMergeState
*gm_state
, int nreader
,
53 bool nowait
, bool *done
);
54 static void ExecShutdownGatherMergeWorkers(GatherMergeState
*node
);
55 static void gather_merge_setup(GatherMergeState
*gm_state
);
56 static void gather_merge_init(GatherMergeState
*gm_state
);
57 static void gather_merge_clear_tuples(GatherMergeState
*gm_state
);
58 static bool gather_merge_readnext(GatherMergeState
*gm_state
, int reader
,
60 static void load_tuple_array(GatherMergeState
*gm_state
, int reader
);
62 /* ----------------------------------------------------------------
64 * ----------------------------------------------------------------
67 ExecInitGatherMerge(GatherMerge
*node
, EState
*estate
, int eflags
)
69 GatherMergeState
*gm_state
;
73 /* Gather merge node doesn't have innerPlan node. */
74 Assert(innerPlan(node
) == NULL
);
77 * create state structure
79 gm_state
= makeNode(GatherMergeState
);
80 gm_state
->ps
.plan
= (Plan
*) node
;
81 gm_state
->ps
.state
= estate
;
82 gm_state
->ps
.ExecProcNode
= ExecGatherMerge
;
84 gm_state
->initialized
= false;
85 gm_state
->gm_initialized
= false;
86 gm_state
->tuples_needed
= -1;
89 * Miscellaneous initialization
91 * create expression context for node
93 ExecAssignExprContext(estate
, &gm_state
->ps
);
96 * GatherMerge doesn't support checking a qual (it's always more efficient
97 * to do it in the child node).
99 Assert(!node
->plan
.qual
);
102 * now initialize outer plan
104 outerNode
= outerPlan(node
);
105 outerPlanState(gm_state
) = ExecInitNode(outerNode
, estate
, eflags
);
108 * Leader may access ExecProcNode result directly (if
109 * need_to_scan_locally), or from workers via tuple queue. So we can't
110 * trivially rely on the slot type being fixed for expressions evaluated
113 gm_state
->ps
.outeropsset
= true;
114 gm_state
->ps
.outeropsfixed
= false;
117 * Store the tuple descriptor into gather merge state, so we can use it
118 * while initializing the gather merge slots.
120 tupDesc
= ExecGetResultType(outerPlanState(gm_state
));
121 gm_state
->tupDesc
= tupDesc
;
124 * Initialize result type and projection.
126 ExecInitResultTypeTL(&gm_state
->ps
);
127 ExecConditionalAssignProjectionInfo(&gm_state
->ps
, tupDesc
, OUTER_VAR
);
130 * Without projections result slot type is not trivially known, see
133 if (gm_state
->ps
.ps_ProjInfo
== NULL
)
135 gm_state
->ps
.resultopsset
= true;
136 gm_state
->ps
.resultopsfixed
= false;
140 * initialize sort-key information
146 gm_state
->gm_nkeys
= node
->numCols
;
147 gm_state
->gm_sortkeys
=
148 palloc0(sizeof(SortSupportData
) * node
->numCols
);
150 for (i
= 0; i
< node
->numCols
; i
++)
152 SortSupport sortKey
= gm_state
->gm_sortkeys
+ i
;
154 sortKey
->ssup_cxt
= CurrentMemoryContext
;
155 sortKey
->ssup_collation
= node
->collations
[i
];
156 sortKey
->ssup_nulls_first
= node
->nullsFirst
[i
];
157 sortKey
->ssup_attno
= node
->sortColIdx
[i
];
160 * We don't perform abbreviated key conversion here, for the same
161 * reasons that it isn't used in MergeAppend
163 sortKey
->abbreviate
= false;
165 PrepareSortSupportFromOrderingOp(node
->sortOperators
[i
], sortKey
);
169 /* Now allocate the workspace for gather merge */
170 gather_merge_setup(gm_state
);
175 /* ----------------------------------------------------------------
176 * ExecGatherMerge(node)
178 * Scans the relation via multiple workers and returns
179 * the next qualifying tuple.
180 * ----------------------------------------------------------------
182 static TupleTableSlot
*
183 ExecGatherMerge(PlanState
*pstate
)
185 GatherMergeState
*node
= castNode(GatherMergeState
, pstate
);
186 TupleTableSlot
*slot
;
187 ExprContext
*econtext
;
189 CHECK_FOR_INTERRUPTS();
192 * As with Gather, we don't launch workers until this node is actually
195 if (!node
->initialized
)
197 EState
*estate
= node
->ps
.state
;
198 GatherMerge
*gm
= castNode(GatherMerge
, node
->ps
.plan
);
201 * Sometimes we might have to run without parallelism; but if parallel
202 * mode is active then we can try to fire up some workers.
204 if (gm
->num_workers
> 0 && estate
->es_use_parallel_mode
)
206 ParallelContext
*pcxt
;
208 /* Initialize, or re-initialize, shared state needed by workers. */
210 node
->pei
= ExecInitParallelPlan(outerPlanState(node
),
214 node
->tuples_needed
);
216 ExecParallelReinitialize(outerPlanState(node
),
220 /* Try to launch workers. */
221 pcxt
= node
->pei
->pcxt
;
222 LaunchParallelWorkers(pcxt
);
223 /* We save # workers launched for the benefit of EXPLAIN */
224 node
->nworkers_launched
= pcxt
->nworkers_launched
;
227 * Count number of workers originally wanted and actually
230 estate
->es_parallel_workers_to_launch
+= pcxt
->nworkers_to_launch
;
231 estate
->es_parallel_workers_launched
+= pcxt
->nworkers_launched
;
233 /* Set up tuple queue readers to read the results. */
234 if (pcxt
->nworkers_launched
> 0)
236 ExecParallelCreateReaders(node
->pei
);
237 /* Make a working array showing the active readers */
238 node
->nreaders
= pcxt
->nworkers_launched
;
239 node
->reader
= (TupleQueueReader
**)
240 palloc(node
->nreaders
* sizeof(TupleQueueReader
*));
241 memcpy(node
->reader
, node
->pei
->reader
,
242 node
->nreaders
* sizeof(TupleQueueReader
*));
246 /* No workers? Then never mind. */
252 /* allow leader to participate if enabled or no choice */
253 if (parallel_leader_participation
|| node
->nreaders
== 0)
254 node
->need_to_scan_locally
= true;
255 node
->initialized
= true;
259 * Reset per-tuple memory context to free any expression evaluation
260 * storage allocated in the previous tuple cycle.
262 econtext
= node
->ps
.ps_ExprContext
;
263 ResetExprContext(econtext
);
266 * Get next tuple, either from one of our workers, or by running the plan
269 slot
= gather_merge_getnext(node
);
273 /* If no projection is required, we're done. */
274 if (node
->ps
.ps_ProjInfo
== NULL
)
278 * Form the result tuple using ExecProject(), and return it.
280 econtext
->ecxt_outertuple
= slot
;
281 return ExecProject(node
->ps
.ps_ProjInfo
);
284 /* ----------------------------------------------------------------
287 * frees any storage allocated through C routines.
288 * ----------------------------------------------------------------
291 ExecEndGatherMerge(GatherMergeState
*node
)
293 ExecEndNode(outerPlanState(node
)); /* let children clean up first */
294 ExecShutdownGatherMerge(node
);
297 /* ----------------------------------------------------------------
298 * ExecShutdownGatherMerge
300 * Destroy the setup for parallel workers including parallel context.
301 * ----------------------------------------------------------------
304 ExecShutdownGatherMerge(GatherMergeState
*node
)
306 ExecShutdownGatherMergeWorkers(node
);
308 /* Now destroy the parallel context. */
309 if (node
->pei
!= NULL
)
311 ExecParallelCleanup(node
->pei
);
316 /* ----------------------------------------------------------------
317 * ExecShutdownGatherMergeWorkers
319 * Stop all the parallel workers.
320 * ----------------------------------------------------------------
323 ExecShutdownGatherMergeWorkers(GatherMergeState
*node
)
325 if (node
->pei
!= NULL
)
326 ExecParallelFinish(node
->pei
);
328 /* Flush local copy of reader array */
334 /* ----------------------------------------------------------------
335 * ExecReScanGatherMerge
337 * Prepare to re-scan the result of a GatherMerge.
338 * ----------------------------------------------------------------
341 ExecReScanGatherMerge(GatherMergeState
*node
)
343 GatherMerge
*gm
= (GatherMerge
*) node
->ps
.plan
;
344 PlanState
*outerPlan
= outerPlanState(node
);
346 /* Make sure any existing workers are gracefully shut down */
347 ExecShutdownGatherMergeWorkers(node
);
349 /* Free any unused tuples, so we don't leak memory across rescans */
350 gather_merge_clear_tuples(node
);
352 /* Mark node so that shared state will be rebuilt at next call */
353 node
->initialized
= false;
354 node
->gm_initialized
= false;
357 * Set child node's chgParam to tell it that the next scan might deliver a
358 * different set of rows within the leader process. (The overall rowset
359 * shouldn't change, but the leader process's subset might; hence nodes
360 * between here and the parallel table scan node mustn't optimize on the
361 * assumption of an unchanging rowset.)
363 if (gm
->rescan_param
>= 0)
364 outerPlan
->chgParam
= bms_add_member(outerPlan
->chgParam
,
368 * If chgParam of subnode is not null then plan will be re-scanned by
369 * first ExecProcNode. Note: because this does nothing if we have a
370 * rescan_param, it's currently guaranteed that parallel-aware child nodes
371 * will not see a ReScan call until after they get a ReInitializeDSM call.
372 * That ordering might not be something to rely on, though. A good rule
373 * of thumb is that ReInitializeDSM should reset only shared state, ReScan
374 * should reset only local state, and anything that depends on both of
375 * those steps being finished must wait until the first ExecProcNode call.
377 if (outerPlan
->chgParam
== NULL
)
378 ExecReScan(outerPlan
);
382 * Set up the data structures that we'll need for Gather Merge.
384 * We allocate these once on the basis of gm->num_workers, which is an
385 * upper bound for the number of workers we'll actually have. During
386 * a rescan, we reset the structures to empty. This approach simplifies
387 * not leaking memory across rescans.
389 * In the gm_slots[] array, index 0 is for the leader, and indexes 1 to n
390 * are for workers. The values placed into gm_heap correspond to indexes
391 * in gm_slots[]. The gm_tuple_buffers[] array, however, is indexed from
392 * 0 to n-1; it has no entry for the leader.
395 gather_merge_setup(GatherMergeState
*gm_state
)
397 GatherMerge
*gm
= castNode(GatherMerge
, gm_state
->ps
.plan
);
398 int nreaders
= gm
->num_workers
;
402 * Allocate gm_slots for the number of workers + one more slot for leader.
403 * Slot 0 is always for the leader. Leader always calls ExecProcNode() to
404 * read the tuple, and then stores it directly into its gm_slots entry.
405 * For other slots, code below will call ExecInitExtraTupleSlot() to
406 * create a slot for the worker's results. Note that during any single
407 * scan, we might have fewer than num_workers available workers, in which
408 * case the extra array entries go unused.
410 gm_state
->gm_slots
= (TupleTableSlot
**)
411 palloc0((nreaders
+ 1) * sizeof(TupleTableSlot
*));
413 /* Allocate the tuple slot and tuple array for each worker */
414 gm_state
->gm_tuple_buffers
= (GMReaderTupleBuffer
*)
415 palloc0(nreaders
* sizeof(GMReaderTupleBuffer
));
417 for (i
= 0; i
< nreaders
; i
++)
419 /* Allocate the tuple array with length MAX_TUPLE_STORE */
420 gm_state
->gm_tuple_buffers
[i
].tuple
=
421 (MinimalTuple
*) palloc0(sizeof(MinimalTuple
) * MAX_TUPLE_STORE
);
423 /* Initialize tuple slot for worker */
424 gm_state
->gm_slots
[i
+ 1] =
425 ExecInitExtraTupleSlot(gm_state
->ps
.state
, gm_state
->tupDesc
,
426 &TTSOpsMinimalTuple
);
429 /* Allocate the resources for the merge */
430 gm_state
->gm_heap
= binaryheap_allocate(nreaders
+ 1,
436 * Initialize the Gather Merge.
438 * Reset data structures to ensure they're empty. Then pull at least one
439 * tuple from leader + each worker (or set its "done" indicator), and set up
443 gather_merge_init(GatherMergeState
*gm_state
)
445 int nreaders
= gm_state
->nreaders
;
449 /* Assert that gather_merge_setup made enough space */
450 Assert(nreaders
<= castNode(GatherMerge
, gm_state
->ps
.plan
)->num_workers
);
452 /* Reset leader's tuple slot to empty */
453 gm_state
->gm_slots
[0] = NULL
;
455 /* Reset the tuple slot and tuple array for each worker */
456 for (i
= 0; i
< nreaders
; i
++)
458 /* Reset tuple array to empty */
459 gm_state
->gm_tuple_buffers
[i
].nTuples
= 0;
460 gm_state
->gm_tuple_buffers
[i
].readCounter
= 0;
461 /* Reset done flag to not-done */
462 gm_state
->gm_tuple_buffers
[i
].done
= false;
463 /* Ensure output slot is empty */
464 ExecClearTuple(gm_state
->gm_slots
[i
+ 1]);
467 /* Reset binary heap to empty */
468 binaryheap_reset(gm_state
->gm_heap
);
471 * First, try to read a tuple from each worker (including leader) in
472 * nowait mode. After this, if not all workers were able to produce a
473 * tuple (or a "done" indication), then re-read from remaining workers,
474 * this time using wait mode. Add all live readers (those producing at
475 * least one tuple) to the heap.
478 for (i
= 0; i
<= nreaders
; i
++)
480 CHECK_FOR_INTERRUPTS();
482 /* skip this source if already known done */
483 if ((i
== 0) ? gm_state
->need_to_scan_locally
:
484 !gm_state
->gm_tuple_buffers
[i
- 1].done
)
486 if (TupIsNull(gm_state
->gm_slots
[i
]))
488 /* Don't have a tuple yet, try to get one */
489 if (gather_merge_readnext(gm_state
, i
, nowait
))
490 binaryheap_add_unordered(gm_state
->gm_heap
,
496 * We already got at least one tuple from this worker, but
497 * might as well see if it has any more ready by now.
499 load_tuple_array(gm_state
, i
);
504 /* need not recheck leader, since nowait doesn't matter for it */
505 for (i
= 1; i
<= nreaders
; i
++)
507 if (!gm_state
->gm_tuple_buffers
[i
- 1].done
&&
508 TupIsNull(gm_state
->gm_slots
[i
]))
515 /* Now heapify the heap. */
516 binaryheap_build(gm_state
->gm_heap
);
518 gm_state
->gm_initialized
= true;
522 * Clear out the tuple table slot, and any unused pending tuples,
523 * for each gather merge input.
526 gather_merge_clear_tuples(GatherMergeState
*gm_state
)
530 for (i
= 0; i
< gm_state
->nreaders
; i
++)
532 GMReaderTupleBuffer
*tuple_buffer
= &gm_state
->gm_tuple_buffers
[i
];
534 while (tuple_buffer
->readCounter
< tuple_buffer
->nTuples
)
535 pfree(tuple_buffer
->tuple
[tuple_buffer
->readCounter
++]);
537 ExecClearTuple(gm_state
->gm_slots
[i
+ 1]);
542 * Read the next tuple for gather merge.
544 * Fetch the sorted tuple out of the heap.
546 static TupleTableSlot
*
547 gather_merge_getnext(GatherMergeState
*gm_state
)
551 if (!gm_state
->gm_initialized
)
554 * First time through: pull the first tuple from each participant, and
557 gather_merge_init(gm_state
);
562 * Otherwise, pull the next tuple from whichever participant we
563 * returned from last time, and reinsert that participant's index into
564 * the heap, because it might now compare differently against the
565 * other elements of the heap.
567 i
= DatumGetInt32(binaryheap_first(gm_state
->gm_heap
));
569 if (gather_merge_readnext(gm_state
, i
, false))
570 binaryheap_replace_first(gm_state
->gm_heap
, Int32GetDatum(i
));
573 /* reader exhausted, remove it from heap */
574 (void) binaryheap_remove_first(gm_state
->gm_heap
);
578 if (binaryheap_empty(gm_state
->gm_heap
))
580 /* All the queues are exhausted, and so is the heap */
581 gather_merge_clear_tuples(gm_state
);
586 /* Return next tuple from whichever participant has the leading one */
587 i
= DatumGetInt32(binaryheap_first(gm_state
->gm_heap
));
588 return gm_state
->gm_slots
[i
];
593 * Read tuple(s) for given reader in nowait mode, and load into its tuple
594 * array, until we have MAX_TUPLE_STORE of them or would have to block.
597 load_tuple_array(GatherMergeState
*gm_state
, int reader
)
599 GMReaderTupleBuffer
*tuple_buffer
;
602 /* Don't do anything if this is the leader. */
606 tuple_buffer
= &gm_state
->gm_tuple_buffers
[reader
- 1];
608 /* If there's nothing in the array, reset the counters to zero. */
609 if (tuple_buffer
->nTuples
== tuple_buffer
->readCounter
)
610 tuple_buffer
->nTuples
= tuple_buffer
->readCounter
= 0;
612 /* Try to fill additional slots in the array. */
613 for (i
= tuple_buffer
->nTuples
; i
< MAX_TUPLE_STORE
; i
++)
617 tuple
= gm_readnext_tuple(gm_state
,
620 &tuple_buffer
->done
);
623 tuple_buffer
->tuple
[i
] = tuple
;
624 tuple_buffer
->nTuples
++;
629 * Store the next tuple for a given reader into the appropriate slot.
631 * Returns true if successful, false if not (either reader is exhausted,
632 * or we didn't want to wait for a tuple). Sets done flag if reader
633 * is found to be exhausted.
636 gather_merge_readnext(GatherMergeState
*gm_state
, int reader
, bool nowait
)
638 GMReaderTupleBuffer
*tuple_buffer
;
642 * If we're being asked to generate a tuple from the leader, then we just
643 * call ExecProcNode as normal to produce one.
647 if (gm_state
->need_to_scan_locally
)
649 PlanState
*outerPlan
= outerPlanState(gm_state
);
650 TupleTableSlot
*outerTupleSlot
;
651 EState
*estate
= gm_state
->ps
.state
;
653 /* Install our DSA area while executing the plan. */
654 estate
->es_query_dsa
= gm_state
->pei
? gm_state
->pei
->area
: NULL
;
655 outerTupleSlot
= ExecProcNode(outerPlan
);
656 estate
->es_query_dsa
= NULL
;
658 if (!TupIsNull(outerTupleSlot
))
660 gm_state
->gm_slots
[0] = outerTupleSlot
;
663 /* need_to_scan_locally serves as "done" flag for leader */
664 gm_state
->need_to_scan_locally
= false;
669 /* Otherwise, check the state of the relevant tuple buffer. */
670 tuple_buffer
= &gm_state
->gm_tuple_buffers
[reader
- 1];
672 if (tuple_buffer
->nTuples
> tuple_buffer
->readCounter
)
674 /* Return any tuple previously read that is still buffered. */
675 tup
= tuple_buffer
->tuple
[tuple_buffer
->readCounter
++];
677 else if (tuple_buffer
->done
)
679 /* Reader is known to be exhausted. */
684 /* Read and buffer next tuple. */
685 tup
= gm_readnext_tuple(gm_state
,
688 &tuple_buffer
->done
);
693 * Attempt to read more tuples in nowait mode and store them in the
694 * pending-tuple array for the reader.
696 load_tuple_array(gm_state
, reader
);
701 /* Build the TupleTableSlot for the given tuple */
702 ExecStoreMinimalTuple(tup
, /* tuple to store */
703 gm_state
->gm_slots
[reader
], /* slot in which to
705 true); /* pfree tuple when done with it */
711 * Attempt to read a tuple from given worker.
714 gm_readnext_tuple(GatherMergeState
*gm_state
, int nreader
, bool nowait
,
717 TupleQueueReader
*reader
;
720 /* Check for async events, particularly messages from workers. */
721 CHECK_FOR_INTERRUPTS();
724 * Attempt to read a tuple.
726 * Note that TupleQueueReaderNext will just return NULL for a worker which
727 * fails to initialize. We'll treat that worker as having produced no
728 * tuples; WaitForParallelWorkersToFinish will error out when we get
731 reader
= gm_state
->reader
[nreader
- 1];
732 tup
= TupleQueueReaderNext(reader
, nowait
, done
);
735 * Since we'll be buffering these across multiple calls, we need to make a
738 return tup
? heap_copy_minimal_tuple(tup
) : NULL
;
742 * We have one slot for each item in the heap array. We use SlotNumber
743 * to store slot indexes. This doesn't actually provide any formal
744 * type-safety, but it makes the code more self-documenting.
746 typedef int32 SlotNumber
;
749 * Compare the tuples in the two given slots.
752 heap_compare_slots(Datum a
, Datum b
, void *arg
)
754 GatherMergeState
*node
= (GatherMergeState
*) arg
;
755 SlotNumber slot1
= DatumGetInt32(a
);
756 SlotNumber slot2
= DatumGetInt32(b
);
758 TupleTableSlot
*s1
= node
->gm_slots
[slot1
];
759 TupleTableSlot
*s2
= node
->gm_slots
[slot2
];
762 Assert(!TupIsNull(s1
));
763 Assert(!TupIsNull(s2
));
765 for (nkey
= 0; nkey
< node
->gm_nkeys
; nkey
++)
767 SortSupport sortKey
= node
->gm_sortkeys
+ nkey
;
768 AttrNumber attno
= sortKey
->ssup_attno
;
775 datum1
= slot_getattr(s1
, attno
, &isNull1
);
776 datum2
= slot_getattr(s2
, attno
, &isNull2
);
778 compare
= ApplySortComparator(datum1
, isNull1
,
783 INVERT_COMPARE_RESULT(compare
);