agora inbox for [email protected]  
help / color / mirror / Atom feed
[PATCH v5 2/3] Allow CREATE INDEX CONCURRENTLY on partitioned table
9+ messages / 2 participants
[nested] [flat]

* [PATCH v5 2/3] Allow CREATE INDEX CONCURRENTLY on partitioned table
@ 2020-06-06 22:42  Justin Pryzby <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Justin Pryzby @ 2020-06-06 22:42 UTC (permalink / raw)

Note, this effectively reverts 050098b14, so take care to not reintroduce the
bug it fixed.
---
 doc/src/sgml/ref/create_index.sgml     |   9 ---
 src/backend/commands/indexcmds.c       | 104 +++++++++++++++++++------
 src/test/regress/expected/indexing.out |  57 +++++++++++++-
 src/test/regress/sql/indexing.sql      |  16 +++-
 4 files changed, 148 insertions(+), 38 deletions(-)

diff --git a/doc/src/sgml/ref/create_index.sgml b/doc/src/sgml/ref/create_index.sgml
index 33aa64e81d..c780dc9547 100644
--- a/doc/src/sgml/ref/create_index.sgml
+++ b/doc/src/sgml/ref/create_index.sgml
@@ -657,15 +657,6 @@ Indexes:
     cannot.
    </para>
 
-   <para>
-    Concurrent builds for indexes on partitioned tables are currently not
-    supported.  However, you may concurrently build the index on each
-    partition individually and then finally create the partitioned index
-    non-concurrently in order to reduce the time where writes to the
-    partitioned table will be locked out.  In this case, building the
-    partitioned index is a metadata only operation.
-   </para>
-
   </refsect2>
  </refsect1>
 
diff --git a/src/backend/commands/indexcmds.c b/src/backend/commands/indexcmds.c
index 68d75916ae..cdf31a47d8 100644
--- a/src/backend/commands/indexcmds.c
+++ b/src/backend/commands/indexcmds.c
@@ -655,17 +655,6 @@ DefineIndex(Oid relationId,
 	partitioned = rel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE;
 	if (partitioned)
 	{
-		/*
-		 * Note: we check 'stmt->concurrent' rather than 'concurrent', so that
-		 * the error is thrown also for temporary tables.  Seems better to be
-		 * consistent, even though we could do it on temporary table because
-		 * we're not actually doing it concurrently.
-		 */
-		if (stmt->concurrent)
-			ereport(ERROR,
-					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
-					 errmsg("cannot create index on partitioned table \"%s\" concurrently",
-							RelationGetRelationName(rel))));
 		if (stmt->excludeOpNames)
 			ereport(ERROR,
 					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
@@ -1100,6 +1089,9 @@ DefineIndex(Oid relationId,
 		if (pd->nparts != 0)
 			flags |= INDEX_CREATE_INVALID;
 	}
+	else if (concurrent && OidIsValid(parentIndexId))
+		/* Initial concurrent build of index partition is invalid */
+		flags |= INDEX_CREATE_INVALID;
 
 	if (stmt->deferrable)
 		constr_flags |= INDEX_CONSTR_CREATE_DEFERRABLE;
@@ -1152,8 +1144,17 @@ DefineIndex(Oid relationId,
 		 */
 		if (!stmt->relation || stmt->relation->inh)
 		{
+			/*
+			 * Need to close the relation before recursing into children, so
+			 * copy needed data into a longlived context.
+			 */
+
+			MemoryContext	ind_context = AllocSetContextCreate(PortalContext, "CREATE INDEX",
+					ALLOCSET_DEFAULT_SIZES);
+			MemoryContext	oldcontext = MemoryContextSwitchTo(ind_context);
 			PartitionDesc partdesc = RelationGetPartitionDesc(rel);
 			int			nparts = partdesc->nparts;
+			char		*relname = pstrdup(RelationGetRelationName(rel));
 			Oid		   *part_oids = palloc(sizeof(Oid) * nparts);
 			bool		invalidate_parent = false;
 			TupleDesc	parentDesc;
@@ -1163,8 +1164,10 @@ DefineIndex(Oid relationId,
 										 nparts);
 
 			memcpy(part_oids, partdesc->oids, sizeof(Oid) * nparts);
+			parentDesc = CreateTupleDescCopy(RelationGetDescr(rel));
+			table_close(rel, NoLock);
+			MemoryContextSwitchTo(oldcontext);
 
-			parentDesc = RelationGetDescr(rel);
 			opfamOids = palloc(sizeof(Oid) * numberOfKeyAttributes);
 			for (i = 0; i < numberOfKeyAttributes; i++)
 				opfamOids[i] = get_opclass_family(classObjectId[i]);
@@ -1199,18 +1202,20 @@ DefineIndex(Oid relationId,
 						ereport(ERROR,
 								(errcode(ERRCODE_WRONG_OBJECT_TYPE),
 								 errmsg("cannot create unique index on partitioned table \"%s\"",
-										RelationGetRelationName(rel)),
+										relname),
 								 errdetail("Table \"%s\" contains partitions that are foreign tables.",
-										   RelationGetRelationName(rel))));
+										   relname)));
 
 					table_close(childrel, lockmode);
 					continue;
 				}
 
+				oldcontext = MemoryContextSwitchTo(ind_context);
 				childidxs = RelationGetIndexList(childrel);
 				attmap =
 					build_attrmap_by_name(RelationGetDescr(childrel),
 										  parentDesc);
+				MemoryContextSwitchTo(oldcontext);
 
 				foreach(cell, childidxs)
 				{
@@ -1336,10 +1341,17 @@ DefineIndex(Oid relationId,
 								createdConstraintId,
 								is_alter_table, check_rights, check_not_in_use,
 								skip_build, quiet);
+					if (concurrent)
+					{
+						PushActiveSnapshot(GetTransactionSnapshot());
+						invalidate_parent = true;
+					}
 				}
 
-				pgstat_progress_update_param(PROGRESS_CREATEIDX_PARTITIONS_DONE,
-											 i + 1);
+				/* For concurrent build, this is a catalog-only stage */
+				if (!concurrent)
+					pgstat_progress_update_param(PROGRESS_CREATEIDX_PARTITIONS_DONE,
+												 i + 1);
 				free_attrmap(attmap);
 			}
 
@@ -1366,23 +1378,72 @@ DefineIndex(Oid relationId,
 				table_close(pg_index, RowExclusiveLock);
 				heap_freetuple(newtup);
 			}
+		} else
+			table_close(rel, NoLock);
+
+		if (concurrent)
+		{
+			List	*childs;
+			ListCell *lc;
+			int		npart = 0;
+
+			/* Reindex invalid child indexes created earlier */
+			MemoryContext	ind_context = AllocSetContextCreate(PortalContext, "CREATE INDEX",
+					ALLOCSET_DEFAULT_SIZES);
+			MemoryContext	oldcontext = MemoryContextSwitchTo(ind_context);
+			childs = find_inheritance_children(indexRelationId, NoLock);
+			MemoryContextSwitchTo(oldcontext);
+
+			/* Make the catalog changes visible to get_partition_parent */
+			CommandCounterIncrement();
+
+			foreach (lc, childs)
+			{
+				Oid			indexrelid = lfirst_oid(lc);
+				Relation	cldidx;
+				bool		isvalid;
+
+				if (!OidIsValid(parentIndexId))
+					pgstat_progress_update_param(PROGRESS_CREATEIDX_PARTITIONS_DONE,
+												 npart++);
+
+				cldidx = index_open(indexrelid, ShareUpdateExclusiveLock);
+				isvalid = cldidx->rd_index->indisvalid;
+				index_close(cldidx, NoLock);
+				if (isvalid)
+					continue;
+
+				/* This may be a partitioned index, which is fine too */
+				ReindexRelationConcurrently(indexrelid, 0);
+				PushActiveSnapshot(GetTransactionSnapshot());
+			}
+
+			PopActiveSnapshot();
+			CommitTransactionCommand();
+			StartTransactionCommand();
+
+			/*
+			 * CIC needs to mark a partitioned index as VALID, which itself
+			 * requires setting READY, which is unset for CIC (even though
+			 * it's meaningless for an index without storage).
+			 */
+			index_set_state_flags(indexRelationId, INDEX_CREATE_SET_READY);
+			CommandCounterIncrement();
+			index_set_state_flags(indexRelationId, INDEX_CREATE_SET_VALID);
 		}
 
 		/*
 		 * Indexes on partitioned tables are not themselves built, so we're
 		 * done here.
 		 */
-		table_close(rel, NoLock);
 		if (!OidIsValid(parentIndexId))
 			pgstat_progress_end_command();
 		return address;
 	}
 
+	table_close(rel, NoLock);
 	if (!concurrent)
 	{
-		/* Close the heap and we're done, in the non-concurrent case */
-		table_close(rel, NoLock);
-
 		/* If this is the top-level index, we're done. */
 		if (!OidIsValid(parentIndexId))
 			pgstat_progress_end_command();
@@ -1390,10 +1451,9 @@ DefineIndex(Oid relationId,
 		return address;
 	}
 
-	/* save lockrelid and locktag for below, then close rel */
+	/* save lockrelid and locktag for below */
 	heaprelid = rel->rd_lockInfo.lockRelId;
 	SET_LOCKTAG_RELATION(heaplocktag, heaprelid.dbId, heaprelid.relId);
-	table_close(rel, NoLock);
 
 	/*
 	 * For a concurrent build, it's important to make the catalog entries
diff --git a/src/test/regress/expected/indexing.out b/src/test/regress/expected/indexing.out
index f78865ef81..1ae58e110f 100644
--- a/src/test/regress/expected/indexing.out
+++ b/src/test/regress/expected/indexing.out
@@ -50,11 +50,60 @@ select relname, relkind, relhassubclass, inhparent::regclass
 (8 rows)
 
 drop table idxpart;
--- Some unsupported features
+-- CIC on partitioned table
 create table idxpart (a int, b int, c text) partition by range (a);
-create table idxpart1 partition of idxpart for values from (0) to (10);
-create index concurrently on idxpart (a);
-ERROR:  cannot create index on partitioned table "idxpart" concurrently
+create table idxpart1 partition of idxpart for values from (0) to (10) partition by range(a);
+create table idxpart11 partition of idxpart1 for values from (0) to (10) partition by range(a);
+create table idxpart2 partition of idxpart for values from (10) to (20);
+insert into idxpart2 values(10),(10); -- not unique
+create index concurrently on idxpart (a); -- partitioned
+create index concurrently on idxpart1 (a); -- partitioned and partition
+create index concurrently on idxpart11 (a); -- partitioned and partition, with no leaves
+create index concurrently on idxpart2 (a); -- leaf
+create unique index concurrently on idxpart (a); -- partitioned, unique failure
+ERROR:  could not create unique index "idxpart2_a_idx2"
+DETAIL:  Key (a)=(10) is duplicated.
+\d idxpart
+        Partitioned table "public.idxpart"
+ Column |  Type   | Collation | Nullable | Default 
+--------+---------+-----------+----------+---------
+ a      | integer |           |          | 
+ b      | integer |           |          | 
+ c      | text    |           |          | 
+Partition key: RANGE (a)
+Indexes:
+    "idxpart_a_idx" btree (a)
+    "idxpart_a_idx1" UNIQUE, btree (a) INVALID
+Number of partitions: 2 (Use \d+ to list them.)
+
+\d idxpart1
+        Partitioned table "public.idxpart1"
+ Column |  Type   | Collation | Nullable | Default 
+--------+---------+-----------+----------+---------
+ a      | integer |           |          | 
+ b      | integer |           |          | 
+ c      | text    |           |          | 
+Partition of: idxpart FOR VALUES FROM (0) TO (10)
+Partition key: RANGE (a)
+Indexes:
+    "idxpart1_a_idx" btree (a)
+    "idxpart1_a_idx1" btree (a)
+    "idxpart1_a_idx2" UNIQUE, btree (a)
+Number of partitions: 1 (Use \d+ to list them.)
+
+\d idxpart2
+              Table "public.idxpart2"
+ Column |  Type   | Collation | Nullable | Default 
+--------+---------+-----------+----------+---------
+ a      | integer |           |          | 
+ b      | integer |           |          | 
+ c      | text    |           |          | 
+Partition of: idxpart FOR VALUES FROM (10) TO (20)
+Indexes:
+    "idxpart2_a_idx" btree (a)
+    "idxpart2_a_idx1" btree (a)
+    "idxpart2_a_idx2" UNIQUE, btree (a) INVALID
+
 drop table idxpart;
 -- Verify bugfix with query on indexed partitioned table with no partitions
 -- https://postgr.es/m/[email protected]
diff --git a/src/test/regress/sql/indexing.sql b/src/test/regress/sql/indexing.sql
index 35d159f41b..d908ee6d17 100644
--- a/src/test/regress/sql/indexing.sql
+++ b/src/test/regress/sql/indexing.sql
@@ -29,10 +29,20 @@ select relname, relkind, relhassubclass, inhparent::regclass
 	where relname like 'idxpart%' order by relname;
 drop table idxpart;
 
--- Some unsupported features
+-- CIC on partitioned table
 create table idxpart (a int, b int, c text) partition by range (a);
-create table idxpart1 partition of idxpart for values from (0) to (10);
-create index concurrently on idxpart (a);
+create table idxpart1 partition of idxpart for values from (0) to (10) partition by range(a);
+create table idxpart11 partition of idxpart1 for values from (0) to (10) partition by range(a);
+create table idxpart2 partition of idxpart for values from (10) to (20);
+insert into idxpart2 values(10),(10); -- not unique
+create index concurrently on idxpart (a); -- partitioned
+create index concurrently on idxpart1 (a); -- partitioned and partition
+create index concurrently on idxpart11 (a); -- partitioned and partition, with no leaves
+create index concurrently on idxpart2 (a); -- leaf
+create unique index concurrently on idxpart (a); -- partitioned, unique failure
+\d idxpart
+\d idxpart1
+\d idxpart2
 drop table idxpart;
 
 -- Verify bugfix with query on indexed partitioned table with no partitions
-- 
2.17.0


--hYooF8G/hrfVAmum
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
 filename="v5-0003-Implement-CLUSTER-of-partitioned-table.patch"



^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH 2/3] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c           | 43 ++++++++++++++++-------
 src/backend/executor/nodeHashjoin.c       | 22 ++++++++----
 src/backend/utils/sort/sharedtuplestore.c |  8 +++++
 src/include/executor/hashjoin.h           | 30 ++++++++++------
 src/include/executor/nodeHashjoin.h       |  2 +-
 5 files changed, 75 insertions(+), 30 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 5fd1c5553b..e529f50a3e 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or spillCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->spillCxt = AllocSetContextCreate(hashtable->hashCxt,
+								 "HashSpillContext",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,12 +574,19 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldctx);
+
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
@@ -913,7 +924,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 	int			oldnbatch = hashtable->nbatch;
 	int			curbatch = hashtable->curbatch;
 	int			nbatch;
-	MemoryContext oldcxt;
 	long		ninmemory;
 	long		nfreed;
 	HashMemoryChunk oldchunks;
@@ -934,13 +944,16 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
-
 	if (hashtable->innerBatchFile == NULL)
 	{
+		MemoryContext oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
+
 		/* we had no file arrays before */
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldcxt);
+
 		/* time to establish the temp tablespaces, too */
 		PrepareTempTablespaces();
 	}
@@ -951,8 +964,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		hashtable->outerBatchFile = repalloc0_array(hashtable->outerBatchFile, BufFile *, oldnbatch, nbatch);
 	}
 
-	MemoryContextSwitchTo(oldcxt);
-
 	hashtable->nbatch = nbatch;
 
 	/*
@@ -1024,7 +1035,8 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 				Assert(batchno > curbatch);
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->spillCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1683,7 +1695,8 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		Assert(batchno > hashtable->curbatch);
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->spillCxt);
 	}
 
 	if (shouldFree)
@@ -2664,7 +2677,8 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->spillCxt);
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
@@ -3093,8 +3107,11 @@ ExecParallelHashJoinSetUpBatches(HashJoinTable hashtable, int nbatch)
 	pstate->nbatch = nbatch;
 	batches = dsa_get_address(hashtable->area, pstate->batches);
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/*
+	 * Use hash join spill memory context to allocate accessors and their
+	 * buffers
+	 */
+	oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = nbatch;
@@ -3196,8 +3213,8 @@ ExecParallelHashEnsureBatchAccessors(HashJoinTable hashtable)
 	 */
 	Assert(DsaPointerIsValid(pstate->batches));
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/* Use hash join spill memory context to allocate accessors */
+	oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = pstate->nbatch;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index a39164c15d..8618e0df40 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -491,7 +491,8 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->spillCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1313,21 +1314,30 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * If this is the first write to the batch file, this function first
+ * create it.  The associated BufFile buffer is allocated in the given
+ * context.  It is important to always give the HashSpillContext
+ * context.  First to avoid a shorter-lived context, else the temp file
+ * buffers will get messed up.  Second, for a better accounting of the
+ * spilling memory consumption.
+ *
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext cxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
-		/* First write to this batch file, so open it. */
+		MemoryContext oldctx;
+
+		oldctx = MemoryContextSwitchTo(cxt);
+
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/backend/utils/sort/sharedtuplestore.c b/src/backend/utils/sort/sharedtuplestore.c
index 0831249159..236be65f22 100644
--- a/src/backend/utils/sort/sharedtuplestore.c
+++ b/src/backend/utils/sort/sharedtuplestore.c
@@ -308,11 +308,15 @@ sts_puttuple(SharedTuplestoreAccessor *accessor, void *meta_data,
 	{
 		SharedTuplestoreParticipant *participant;
 		char		name[MAXPGPATH];
+		MemoryContext oldcxt;
 
 		/* Create one.  Only this backend will write into it. */
 		sts_filename(name, accessor, accessor->participant);
+
+		oldcxt = MemoryContextSwitchTo(accessor->context);
 		accessor->write_file =
 			BufFileCreateFileSet(&accessor->fileset->fs, name);
+		MemoryContextSwitchTo(oldcxt);
 
 		/* Set up the shared state for this backend's file. */
 		participant = &accessor->sts->participants[accessor->participant];
@@ -527,11 +531,15 @@ sts_parallel_scan_next(SharedTuplestoreAccessor *accessor, void *meta_data)
 			if (accessor->read_file == NULL)
 			{
 				char		name[MAXPGPATH];
+				MemoryContext oldcxt;
 
 				sts_filename(name, accessor, accessor->read_participant);
+
+				oldcxt = MemoryContextSwitchTo(accessor->context);
 				accessor->read_file =
 					BufFileOpenFileSet(&accessor->fileset->fs, name, O_RDONLY,
 									   false);
+				MemoryContextSwitchTo(oldcxt);
 			}
 
 			/* Seek and load the chunk header. */
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index 8ee59d2c71..09113f0192 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -23,22 +23,31 @@
 /* ----------------------------------------------------------------
  *				hash-join hash table structures
  *
- * Each active hashjoin has a HashJoinTable control block, which is
- * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
- * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * Each active hashjoin has a HashJoinTable structure, which is
+ * palloc'd in the executor's per-query context.  Other storages needed for
+ * each hashjoin is kept in child contexts, three for each hashjoin:
+ *   - HashTableContext (hashCxt): the parent hash table storage context
+ *     - HashSpillContext (spillCxt): storage for temp files buffers
+ *     - HashBatchContext (batchCxt): storage for a batch in serial hash join
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
  * aborted early by an error.  (Likewise, any temporary files we make will
  * be cleaned up by the virtual file manager in event of an error.)
  *
- * Storage that should live through the entire join is allocated from the
- * "hashCxt", while storage that is only wanted for the current batch is
- * allocated in the "batchCxt".  By resetting the batchCxt at the end of
- * each batch, we free all the per-batch storage reliably and without tedium.
+ * The "hashCxt" context makes it easy and fast to release the storage when we
+ * don't need it anymore. Storage that should live through the entire join is
+ * allocated from the "hashCxt" (mainly hash's meta datas).
+ *
+ * Data associated with temp files lives in the "spillCxt" context which lives
+ * during the entire join as temp files might need to survives batches. These
+ * files are explicitly destroyed by calling BufFileClose() when the code is
+ * done with them. The aim of this context is to help accounting the memory
+ * allocations dedicated to temp files and their buffers.
+ *
+ * Finaly, storage that is only wanted for the current batch is allocated in
+ * the "batchCxt".  By resetting the batchCxt at the end of each batch, we free
+ * all the per-batch storage reliably and without tedium.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -350,6 +359,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext spillCxt;		/* context for spilling to temp files */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..958df42240 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext spillcxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.40.1


--MP_/WHuN2qnG.U=I.v0AxCf+Cxe--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH 2/3] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c           | 43 ++++++++++++++++-------
 src/backend/executor/nodeHashjoin.c       | 18 +++++++---
 src/backend/utils/sort/sharedtuplestore.c |  8 +++++
 src/include/executor/hashjoin.h           | 15 ++++++--
 src/include/executor/nodeHashjoin.h       |  2 +-
 5 files changed, 64 insertions(+), 22 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 5fd1c5553b..a4fbf29301 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or fileCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->fileCxt = AllocSetContextCreate(hashtable->hashCxt,
+								 "HashBatchFiles",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,15 +574,21 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	MemoryContextSwitchTo(oldcxt);
@@ -913,7 +923,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 	int			oldnbatch = hashtable->nbatch;
 	int			curbatch = hashtable->curbatch;
 	int			nbatch;
-	MemoryContext oldcxt;
 	long		ninmemory;
 	long		nfreed;
 	HashMemoryChunk oldchunks;
@@ -934,13 +943,16 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
-
 	if (hashtable->innerBatchFile == NULL)
 	{
+		MemoryContext oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		/* we had no file arrays before */
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldcxt);
+
 		/* time to establish the temp tablespaces, too */
 		PrepareTempTablespaces();
 	}
@@ -951,8 +963,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		hashtable->outerBatchFile = repalloc0_array(hashtable->outerBatchFile, BufFile *, oldnbatch, nbatch);
 	}
 
-	MemoryContextSwitchTo(oldcxt);
-
 	hashtable->nbatch = nbatch;
 
 	/*
@@ -1022,9 +1032,11 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 			{
 				/* dump it out */
 				Assert(batchno > curbatch);
+
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->fileCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1681,9 +1693,11 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		 * put the tuple into a temp file for later batches
 		 */
 		Assert(batchno > hashtable->curbatch);
+
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->fileCxt);
 	}
 
 	if (shouldFree)
@@ -2663,8 +2677,11 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 		{
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
+
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->fileCxt);
+
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
@@ -3093,8 +3110,8 @@ ExecParallelHashJoinSetUpBatches(HashJoinTable hashtable, int nbatch)
 	pstate->nbatch = nbatch;
 	batches = dsa_get_address(hashtable->area, pstate->batches);
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/* Use hash join file memory context. */
+	oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = nbatch;
@@ -3196,8 +3213,8 @@ ExecParallelHashEnsureBatchAccessors(HashJoinTable hashtable)
 	 */
 	Assert(DsaPointerIsValid(pstate->batches));
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/* Use hash join file memory context. */
+	oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = pstate->nbatch;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index 920d1831c2..ac72fbfbb6 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -485,8 +485,10 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					 */
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
+
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->fileCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1308,21 +1310,27 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * Note: it is important always to call this in the HashBatchFiles context,
+ * not in a shorter-lived context; else the temp file buffers will get messed
+ * up.
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext filecxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
+		MemoryContext oldctx;
+
+		oldctx = MemoryContextSwitchTo(filecxt);
+
 		/* First write to this batch file, so open it. */
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/backend/utils/sort/sharedtuplestore.c b/src/backend/utils/sort/sharedtuplestore.c
index 0831249159..236be65f22 100644
--- a/src/backend/utils/sort/sharedtuplestore.c
+++ b/src/backend/utils/sort/sharedtuplestore.c
@@ -308,11 +308,15 @@ sts_puttuple(SharedTuplestoreAccessor *accessor, void *meta_data,
 	{
 		SharedTuplestoreParticipant *participant;
 		char		name[MAXPGPATH];
+		MemoryContext oldcxt;
 
 		/* Create one.  Only this backend will write into it. */
 		sts_filename(name, accessor, accessor->participant);
+
+		oldcxt = MemoryContextSwitchTo(accessor->context);
 		accessor->write_file =
 			BufFileCreateFileSet(&accessor->fileset->fs, name);
+		MemoryContextSwitchTo(oldcxt);
 
 		/* Set up the shared state for this backend's file. */
 		participant = &accessor->sts->participants[accessor->participant];
@@ -527,11 +531,15 @@ sts_parallel_scan_next(SharedTuplestoreAccessor *accessor, void *meta_data)
 			if (accessor->read_file == NULL)
 			{
 				char		name[MAXPGPATH];
+				MemoryContext oldcxt;
 
 				sts_filename(name, accessor, accessor->read_participant);
+
+				oldcxt = MemoryContextSwitchTo(accessor->context);
 				accessor->read_file =
 					BufFileOpenFileSet(&accessor->fileset->fs, name, O_RDONLY,
 									   false);
+				MemoryContextSwitchTo(oldcxt);
 			}
 
 			/* Seek and load the chunk header. */
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index 8ee59d2c71..74867c3e40 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -25,10 +25,14 @@
  *
  * Each active hashjoin has a HashJoinTable control block, which is
  * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
+ * for the hashjoin is kept in private memory contexts, three for each
+ * hashjoin:
+ * - HashTableContext (hashCxt): the control block associated to the hash table
+ * - HashBatchContext (batchCxt): storages for batches
+ * - HashBatchFiles (fileCxt): storage for temp files buffers
+ *
  * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * anymore.
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
@@ -39,6 +43,10 @@
  * "hashCxt", while storage that is only wanted for the current batch is
  * allocated in the "batchCxt".  By resetting the batchCxt at the end of
  * each batch, we free all the per-batch storage reliably and without tedium.
+ * Note that data associated with the temp files lives in the "fileCxt" context
+ * which lives during the entire join as temp files might need to survives
+ * batches. These files are explicitly destroyed by calling BufFileClose()
+ * when the code is done with them.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -350,6 +358,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext fileCxt;		/* context for the BufFile related storage */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..a8f9ae1989 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext filecxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.39.2


--MP_/w5PdZ2hoFAvEnG3QNPoMNWk--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH 2/3] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c           | 43 ++++++++++++++++-------
 src/backend/executor/nodeHashjoin.c       | 24 +++++++++----
 src/backend/utils/sort/sharedtuplestore.c |  8 +++++
 src/include/executor/hashjoin.h           | 30 ++++++++++------
 src/include/executor/nodeHashjoin.h       |  2 +-
 5 files changed, 76 insertions(+), 31 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 5fd1c5553b..e529f50a3e 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or spillCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->spillCxt = AllocSetContextCreate(hashtable->hashCxt,
+								 "HashSpillContext",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,12 +574,19 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldctx);
+
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
@@ -913,7 +924,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 	int			oldnbatch = hashtable->nbatch;
 	int			curbatch = hashtable->curbatch;
 	int			nbatch;
-	MemoryContext oldcxt;
 	long		ninmemory;
 	long		nfreed;
 	HashMemoryChunk oldchunks;
@@ -934,13 +944,16 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
-
 	if (hashtable->innerBatchFile == NULL)
 	{
+		MemoryContext oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
+
 		/* we had no file arrays before */
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldcxt);
+
 		/* time to establish the temp tablespaces, too */
 		PrepareTempTablespaces();
 	}
@@ -951,8 +964,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		hashtable->outerBatchFile = repalloc0_array(hashtable->outerBatchFile, BufFile *, oldnbatch, nbatch);
 	}
 
-	MemoryContextSwitchTo(oldcxt);
-
 	hashtable->nbatch = nbatch;
 
 	/*
@@ -1024,7 +1035,8 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 				Assert(batchno > curbatch);
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->spillCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1683,7 +1695,8 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		Assert(batchno > hashtable->curbatch);
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->spillCxt);
 	}
 
 	if (shouldFree)
@@ -2664,7 +2677,8 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->spillCxt);
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
@@ -3093,8 +3107,11 @@ ExecParallelHashJoinSetUpBatches(HashJoinTable hashtable, int nbatch)
 	pstate->nbatch = nbatch;
 	batches = dsa_get_address(hashtable->area, pstate->batches);
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/*
+	 * Use hash join spill memory context to allocate accessors and their
+	 * buffers
+	 */
+	oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = nbatch;
@@ -3196,8 +3213,8 @@ ExecParallelHashEnsureBatchAccessors(HashJoinTable hashtable)
 	 */
 	Assert(DsaPointerIsValid(pstate->batches));
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/* Use hash join spill memory context to allocate accessors */
+	oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = pstate->nbatch;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index 920d1831c2..3b903cc905 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -486,7 +486,8 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->spillCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1307,22 +1308,31 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  *
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
- *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext spillcxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
-		/* First write to this batch file, so open it. */
+		MemoryContext oldctx;
+
+		/*
+		 * First write to this batch file, so open it.
+		 *
+		 * It is important to always create BufFile'd temp files in the
+		 * HashSpillContext context. First to avoid a shorter-lived context,
+		 * else the temp file buffers will get messed up. Second, for a better
+		 * accounting of the spilling memory consumption.
+		 */
+		oldctx = MemoryContextSwitchTo(spillcxt);
+
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/backend/utils/sort/sharedtuplestore.c b/src/backend/utils/sort/sharedtuplestore.c
index 0831249159..236be65f22 100644
--- a/src/backend/utils/sort/sharedtuplestore.c
+++ b/src/backend/utils/sort/sharedtuplestore.c
@@ -308,11 +308,15 @@ sts_puttuple(SharedTuplestoreAccessor *accessor, void *meta_data,
 	{
 		SharedTuplestoreParticipant *participant;
 		char		name[MAXPGPATH];
+		MemoryContext oldcxt;
 
 		/* Create one.  Only this backend will write into it. */
 		sts_filename(name, accessor, accessor->participant);
+
+		oldcxt = MemoryContextSwitchTo(accessor->context);
 		accessor->write_file =
 			BufFileCreateFileSet(&accessor->fileset->fs, name);
+		MemoryContextSwitchTo(oldcxt);
 
 		/* Set up the shared state for this backend's file. */
 		participant = &accessor->sts->participants[accessor->participant];
@@ -527,11 +531,15 @@ sts_parallel_scan_next(SharedTuplestoreAccessor *accessor, void *meta_data)
 			if (accessor->read_file == NULL)
 			{
 				char		name[MAXPGPATH];
+				MemoryContext oldcxt;
 
 				sts_filename(name, accessor, accessor->read_participant);
+
+				oldcxt = MemoryContextSwitchTo(accessor->context);
 				accessor->read_file =
 					BufFileOpenFileSet(&accessor->fileset->fs, name, O_RDONLY,
 									   false);
+				MemoryContextSwitchTo(oldcxt);
 			}
 
 			/* Seek and load the chunk header. */
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index 8ee59d2c71..09113f0192 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -23,22 +23,31 @@
 /* ----------------------------------------------------------------
  *				hash-join hash table structures
  *
- * Each active hashjoin has a HashJoinTable control block, which is
- * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
- * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * Each active hashjoin has a HashJoinTable structure, which is
+ * palloc'd in the executor's per-query context.  Other storages needed for
+ * each hashjoin is kept in child contexts, three for each hashjoin:
+ *   - HashTableContext (hashCxt): the parent hash table storage context
+ *     - HashSpillContext (spillCxt): storage for temp files buffers
+ *     - HashBatchContext (batchCxt): storage for a batch in serial hash join
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
  * aborted early by an error.  (Likewise, any temporary files we make will
  * be cleaned up by the virtual file manager in event of an error.)
  *
- * Storage that should live through the entire join is allocated from the
- * "hashCxt", while storage that is only wanted for the current batch is
- * allocated in the "batchCxt".  By resetting the batchCxt at the end of
- * each batch, we free all the per-batch storage reliably and without tedium.
+ * The "hashCxt" context makes it easy and fast to release the storage when we
+ * don't need it anymore. Storage that should live through the entire join is
+ * allocated from the "hashCxt" (mainly hash's meta datas).
+ *
+ * Data associated with temp files lives in the "spillCxt" context which lives
+ * during the entire join as temp files might need to survives batches. These
+ * files are explicitly destroyed by calling BufFileClose() when the code is
+ * done with them. The aim of this context is to help accounting the memory
+ * allocations dedicated to temp files and their buffers.
+ *
+ * Finaly, storage that is only wanted for the current batch is allocated in
+ * the "batchCxt".  By resetting the batchCxt at the end of each batch, we free
+ * all the per-batch storage reliably and without tedium.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -350,6 +359,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext spillCxt;		/* context for spilling to temp files */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..958df42240 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext spillcxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.39.2


--MP_/=afXY/P8/zXs58RNNJ5YG6H--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c     | 27 ++++++++++++++++++++++-----
 src/backend/executor/nodeHashjoin.c | 18 +++++++++++++-----
 src/include/executor/hashjoin.h     | 15 ++++++++++++---
 src/include/executor/nodeHashjoin.h |  2 +-
 4 files changed, 48 insertions(+), 14 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 748c9b0024..0c0d5b4a3c 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or fileCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->fileCxt = AllocSetContextCreate(hashtable->hashCxt,
+								 "HashBatchFiles",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,15 +574,21 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	MemoryContextSwitchTo(oldcxt);
@@ -934,7 +944,7 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
 
 	if (hashtable->innerBatchFile == NULL)
 	{
@@ -1022,9 +1032,11 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 			{
 				/* dump it out */
 				Assert(batchno > curbatch);
+
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->fileCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1681,9 +1693,11 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		 * put the tuple into a temp file for later batches
 		 */
 		Assert(batchno > hashtable->curbatch);
+
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->fileCxt);
 	}
 
 	if (shouldFree)
@@ -2534,8 +2548,11 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 		{
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
+
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->fileCxt);
+
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index f189fb4d28..6055abde49 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -432,8 +432,10 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					 */
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
+
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->fileCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1234,21 +1236,27 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * Note: it is important always to call this in the HashBatchFiles context,
+ * not in a shorter-lived context; else the temp file buffers will get messed
+ * up.
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext filecxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
+		MemoryContext oldctx;
+
+		oldctx = MemoryContextSwitchTo(filecxt);
+
 		/* First write to this batch file, so open it. */
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index acb7592ca0..d759235d7f 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -25,10 +25,14 @@
  *
  * Each active hashjoin has a HashJoinTable control block, which is
  * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
+ * for the hashjoin is kept in private memory contexts, three for each
+ * hashjoin:
+ * - HashTableContext (hashCxt): the control block associated to the hash table
+ * - HashBatchContext (batchCxt): storages for batches
+ * - HashBatchFiles (fileCxt): storage for temp files buffers
+ *
  * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * anymore.
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
@@ -39,6 +43,10 @@
  * "hashCxt", while storage that is only wanted for the current batch is
  * allocated in the "batchCxt".  By resetting the batchCxt at the end of
  * each batch, we free all the per-batch storage reliably and without tedium.
+ * Note that data associated with the temp files lives in the "fileCxt" context
+ * which lives during the entire join as temp files might need to survives
+ * batches. These files are explicitly destroyed by calling BufFileClose()
+ * when the code is done with them.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -348,6 +356,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext fileCxt;		/* context for the BufFile related storage */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..a8f9ae1989 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext filecxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.39.2


--MP_/HNeUeUP7WKSzkiBqcPpwVN+--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH 2/4] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c     | 35 +++++++++++++++++++++--------
 src/backend/executor/nodeHashjoin.c | 18 ++++++++++-----
 src/include/executor/hashjoin.h     | 15 ++++++++++---
 src/include/executor/nodeHashjoin.h |  2 +-
 4 files changed, 52 insertions(+), 18 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index a45bd3a315..4544296391 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or fileCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->fileCxt = AllocSetContextCreate(hashtable->hashCxt,
+								 "HashBatchFiles",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,15 +574,21 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	MemoryContextSwitchTo(oldcxt);
@@ -913,7 +923,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 	int			oldnbatch = hashtable->nbatch;
 	int			curbatch = hashtable->curbatch;
 	int			nbatch;
-	MemoryContext oldcxt;
 	long		ninmemory;
 	long		nfreed;
 	HashMemoryChunk oldchunks;
@@ -934,13 +943,16 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
-
 	if (hashtable->innerBatchFile == NULL)
 	{
+		MemoryContext oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		/* we had no file arrays before */
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldcxt);
+
 		/* time to establish the temp tablespaces, too */
 		PrepareTempTablespaces();
 	}
@@ -951,8 +963,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		hashtable->outerBatchFile = repalloc0_array(hashtable->outerBatchFile, BufFile *, oldnbatch, nbatch);
 	}
 
-	MemoryContextSwitchTo(oldcxt);
-
 	hashtable->nbatch = nbatch;
 
 	/*
@@ -1022,9 +1032,11 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 			{
 				/* dump it out */
 				Assert(batchno > curbatch);
+
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->fileCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1681,9 +1693,11 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		 * put the tuple into a temp file for later batches
 		 */
 		Assert(batchno > hashtable->curbatch);
+
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->fileCxt);
 	}
 
 	if (shouldFree)
@@ -2663,8 +2677,11 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 		{
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
+
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->fileCxt);
+
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index 5454afbff2..5bc7f814c6 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -485,8 +485,10 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					 */
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
+
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->fileCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1297,21 +1299,27 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * Note: it is important always to call this in the HashBatchFiles context,
+ * not in a shorter-lived context; else the temp file buffers will get messed
+ * up.
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext filecxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
+		MemoryContext oldctx;
+
+		oldctx = MemoryContextSwitchTo(filecxt);
+
 		/* First write to this batch file, so open it. */
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index 8ee59d2c71..74867c3e40 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -25,10 +25,14 @@
  *
  * Each active hashjoin has a HashJoinTable control block, which is
  * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
+ * for the hashjoin is kept in private memory contexts, three for each
+ * hashjoin:
+ * - HashTableContext (hashCxt): the control block associated to the hash table
+ * - HashBatchContext (batchCxt): storages for batches
+ * - HashBatchFiles (fileCxt): storage for temp files buffers
+ *
  * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * anymore.
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
@@ -39,6 +43,10 @@
  * "hashCxt", while storage that is only wanted for the current batch is
  * allocated in the "batchCxt".  By resetting the batchCxt at the end of
  * each batch, we free all the per-batch storage reliably and without tedium.
+ * Note that data associated with the temp files lives in the "fileCxt" context
+ * which lives during the entire join as temp files might need to survives
+ * batches. These files are explicitly destroyed by calling BufFileClose()
+ * when the code is done with them.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -350,6 +358,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext fileCxt;		/* context for the BufFile related storage */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..a8f9ae1989 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext filecxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.39.2


--MP_/xtmrMo8rz+yGL.lLMmD=jly
Content-Type: text/x-patch
Content-Transfer-Encoding: 7bit
Content-Disposition: attachment;
 filename=0003-v4-Add-some-debug-and-metrics.patch



^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c     | 39 +++++++++++++++++++++++++++--
 src/backend/executor/nodeHashjoin.c | 14 ++++++++---
 src/include/executor/hashjoin.h     |  1 +
 3 files changed, 49 insertions(+), 5 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 748c9b0024..3f1ae9755e 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or fileCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->fileCxt = AllocSetContextCreate(CurrentMemoryContext,
+								 "HashBatchFiles",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,15 +574,21 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	MemoryContextSwitchTo(oldcxt);
@@ -934,7 +944,7 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
 
 	if (hashtable->innerBatchFile == NULL)
 	{
@@ -1020,12 +1030,19 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 			}
 			else
 			{
+				MemoryContext oldctx;
+
 				/* dump it out */
 				Assert(batchno > curbatch);
+
+				oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
 									  &hashtable->innerBatchFile[batchno]);
 
+				MemoryContextSwitchTo(oldctx);
+
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
 			}
@@ -1042,6 +1059,9 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		oldchunks = nextchunk;
 	}
 
+	if (hashtable->fileCxt->mem_allocated > hashtable->spaceAllowed)
+		elog(WARNING, "Growing number of hash batch is exhausting memory");
+
 #ifdef HJDEBUG
 	printf("Hashjoin %p: freed %ld of %ld tuples, space now %zu\n",
 		   hashtable, nfreed, ninmemory, hashtable->spaceUsed);
@@ -1677,13 +1697,20 @@ ExecHashTableInsert(HashJoinTable hashtable,
 	}
 	else
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * put the tuple into a temp file for later batches
 		 */
 		Assert(batchno > hashtable->curbatch);
+
+		oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
 							  &hashtable->innerBatchFile[batchno]);
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	if (shouldFree)
@@ -2532,10 +2559,18 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 		}
 		else
 		{
+			MemoryContext oldctx;
+
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
+
+			oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 			ExecHashJoinSaveTuple(tuple, hashvalue,
 								  &hashtable->innerBatchFile[batchno]);
+
+			MemoryContextSwitchTo(oldctx);
+
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index f189fb4d28..ba1f27c2c4 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -422,6 +422,7 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 				if (batchno != hashtable->curbatch &&
 					node->hj_CurSkewBucketNo == INVALID_SKEW_BUCKET_NO)
 				{
+					MemoryContext oldctx;
 					bool		shouldFree;
 					MinimalTuple mintuple = ExecFetchSlotMinimalTuple(outerTupleSlot,
 																	  &shouldFree);
@@ -432,9 +433,14 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					 */
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
+
+					oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
 										  &hashtable->outerBatchFile[batchno]);
 
+					MemoryContextSwitchTo(oldctx);
+
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
 
@@ -1234,9 +1240,9 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * Note: it is important always to call this in the HashBatchFiles context,
+ * not in a shorter-lived context; else the temp file buffers will get messed
+ * up.
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
@@ -1246,6 +1252,8 @@ ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
 
 	if (file == NULL)
 	{
+		Assert(strcmp(CurrentMemoryContext->name, "HashBatchFiles") == 0);
+
 		/* First write to this batch file, so open it. */
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index acb7592ca0..d36beb7229 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -348,6 +348,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext fileCxt;		/* context for the BufFile related storage */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
-- 
2.39.2


--MP_/GVa.HActNIsT4j_JhNzVDcp--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH] Allocate hash batches related BufFile in a dedicated context
@ 2023-03-27 13:54  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-03-27 13:54 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c     | 43 +++++++++++++++++++++++++----
 src/backend/executor/nodeHashjoin.c | 18 ++++++++----
 src/include/executor/hashjoin.h     | 15 ++++++++--
 src/include/executor/nodeHashjoin.h |  2 +-
 4 files changed, 64 insertions(+), 14 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 748c9b0024..3da83ac22a 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or fileCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->fileCxt = AllocSetContextCreate(CurrentMemoryContext,
+								 "HashBatchFiles",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,15 +574,21 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->fileCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	MemoryContextSwitchTo(oldcxt);
@@ -929,12 +939,18 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 	nbatch = oldnbatch * 2;
 	Assert(nbatch > 1);
 
+	elog(WARNING, "increasing number of batches from %d to %d", oldnbatch, nbatch);
+
+	elog(LOG, "ExecHashIncreaseNumBatches ======= context stats start =======");
+	MemoryContextStats(TopMemoryContext);
+
+
 #ifdef HJDEBUG
 	printf("Hashjoin %p: increasing nbatch to %d because space = %zu\n",
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	oldcxt = MemoryContextSwitchTo(hashtable->fileCxt);
 
 	if (hashtable->innerBatchFile == NULL)
 	{
@@ -1022,9 +1038,11 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 			{
 				/* dump it out */
 				Assert(batchno > curbatch);
+
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->fileCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1042,6 +1060,13 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		oldchunks = nextchunk;
 	}
 
+	if (hashtable->fileCxt->mem_allocated > hashtable->spaceAllowed)
+		elog(NOTICE,
+			 "Growing number of hash batch to %d is exhausting allowed memory (%ld > %ld)",
+			 nbatch,
+			 hashtable->fileCxt->mem_allocated,
+			 hashtable->spaceAllowed);
+
 #ifdef HJDEBUG
 	printf("Hashjoin %p: freed %ld of %ld tuples, space now %zu\n",
 		   hashtable, nfreed, ninmemory, hashtable->spaceUsed);
@@ -1063,6 +1088,9 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 			   hashtable);
 #endif
 	}
+
+	elog(LOG, "ExecHashIncreaseNumBatches ======= context stats end =======");
+	MemoryContextStats(TopMemoryContext);
 }
 
 /*
@@ -1681,9 +1709,11 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		 * put the tuple into a temp file for later batches
 		 */
 		Assert(batchno > hashtable->curbatch);
+
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->fileCxt);
 	}
 
 	if (shouldFree)
@@ -2534,8 +2564,11 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 		{
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
+
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->fileCxt);
+
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index f189fb4d28..6055abde49 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -432,8 +432,10 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					 */
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
+
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->fileCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1234,21 +1236,27 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * Note: it is important always to call this in the HashBatchFiles context,
+ * not in a shorter-lived context; else the temp file buffers will get messed
+ * up.
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext filecxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
+		MemoryContext oldctx;
+
+		oldctx = MemoryContextSwitchTo(filecxt);
+
 		/* First write to this batch file, so open it. */
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index acb7592ca0..d759235d7f 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -25,10 +25,14 @@
  *
  * Each active hashjoin has a HashJoinTable control block, which is
  * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
+ * for the hashjoin is kept in private memory contexts, three for each
+ * hashjoin:
+ * - HashTableContext (hashCxt): the control block associated to the hash table
+ * - HashBatchContext (batchCxt): storages for batches
+ * - HashBatchFiles (fileCxt): storage for temp files buffers
+ *
  * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * anymore.
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
@@ -39,6 +43,10 @@
  * "hashCxt", while storage that is only wanted for the current batch is
  * allocated in the "batchCxt".  By resetting the batchCxt at the end of
  * each batch, we free all the per-batch storage reliably and without tedium.
+ * Note that data associated with the temp files lives in the "fileCxt" context
+ * which lives during the entire join as temp files might need to survives
+ * batches. These files are explicitly destroyed by calling BufFileClose()
+ * when the code is done with them.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -348,6 +356,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext fileCxt;		/* context for the BufFile related storage */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..a8f9ae1989 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext filecxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.39.2


--MP_/cg+MWpd=CPm0_9SBk5uZwGx--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread

* [PATCH v8 2/3] Allocate hash batches related BufFile in a dedicated context
@ 2023-05-16 13:42  Jehan-Guillaume de Rorthais <[email protected]>
  0 siblings, 0 replies; 9+ messages in thread

From: Jehan-Guillaume de Rorthais @ 2023-05-16 13:42 UTC (permalink / raw)

---
 src/backend/executor/nodeHash.c           | 43 ++++++++++++++++-------
 src/backend/executor/nodeHashjoin.c       | 22 ++++++++----
 src/backend/utils/sort/sharedtuplestore.c |  8 +++++
 src/include/executor/hashjoin.h           | 29 ++++++++++-----
 src/include/executor/nodeHashjoin.h       |  2 +-
 5 files changed, 75 insertions(+), 29 deletions(-)

diff --git a/src/backend/executor/nodeHash.c b/src/backend/executor/nodeHash.c
index 5fd1c5553b..e529f50a3e 100644
--- a/src/backend/executor/nodeHash.c
+++ b/src/backend/executor/nodeHash.c
@@ -484,7 +484,7 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 	 *
 	 * The hashtable control block is just palloc'd from the executor's
 	 * per-query memory context.  Everything else should be kept inside the
-	 * subsidiary hashCxt or batchCxt.
+	 * subsidiary hashCxt, batchCxt or spillCxt.
 	 */
 	hashtable = palloc_object(HashJoinTableData);
 	hashtable->nbuckets = nbuckets;
@@ -538,6 +538,10 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 												"HashBatchContext",
 												ALLOCSET_DEFAULT_SIZES);
 
+	hashtable->spillCxt = AllocSetContextCreate(hashtable->hashCxt,
+								 "HashSpillContext",
+								 ALLOCSET_DEFAULT_SIZES);
+
 	/* Allocate data that will live for the life of the hashjoin */
 
 	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
@@ -570,12 +574,19 @@ ExecHashTableCreate(HashState *state, List *hashOperators, List *hashCollations,
 
 	if (nbatch > 1 && hashtable->parallel_state == NULL)
 	{
+		MemoryContext oldctx;
+
 		/*
 		 * allocate and initialize the file arrays in hashCxt (not needed for
 		 * parallel case which uses shared tuplestores instead of raw files)
 		 */
+		oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
+
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldctx);
+
 		/* The files will not be opened until needed... */
 		/* ... but make sure we have temp tablespaces established for them */
 		PrepareTempTablespaces();
@@ -913,7 +924,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 	int			oldnbatch = hashtable->nbatch;
 	int			curbatch = hashtable->curbatch;
 	int			nbatch;
-	MemoryContext oldcxt;
 	long		ninmemory;
 	long		nfreed;
 	HashMemoryChunk oldchunks;
@@ -934,13 +944,16 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		   hashtable, nbatch, hashtable->spaceUsed);
 #endif
 
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
-
 	if (hashtable->innerBatchFile == NULL)
 	{
+		MemoryContext oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
+
 		/* we had no file arrays before */
 		hashtable->innerBatchFile = palloc0_array(BufFile *, nbatch);
 		hashtable->outerBatchFile = palloc0_array(BufFile *, nbatch);
+
+		MemoryContextSwitchTo(oldcxt);
+
 		/* time to establish the temp tablespaces, too */
 		PrepareTempTablespaces();
 	}
@@ -951,8 +964,6 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 		hashtable->outerBatchFile = repalloc0_array(hashtable->outerBatchFile, BufFile *, oldnbatch, nbatch);
 	}
 
-	MemoryContextSwitchTo(oldcxt);
-
 	hashtable->nbatch = nbatch;
 
 	/*
@@ -1024,7 +1035,8 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
 				Assert(batchno > curbatch);
 				ExecHashJoinSaveTuple(HJTUPLE_MINTUPLE(hashTuple),
 									  hashTuple->hashvalue,
-									  &hashtable->innerBatchFile[batchno]);
+									  &hashtable->innerBatchFile[batchno],
+									  hashtable->spillCxt);
 
 				hashtable->spaceUsed -= hashTupleSize;
 				nfreed++;
@@ -1683,7 +1695,8 @@ ExecHashTableInsert(HashJoinTable hashtable,
 		Assert(batchno > hashtable->curbatch);
 		ExecHashJoinSaveTuple(tuple,
 							  hashvalue,
-							  &hashtable->innerBatchFile[batchno]);
+							  &hashtable->innerBatchFile[batchno],
+							  hashtable->spillCxt);
 	}
 
 	if (shouldFree)
@@ -2664,7 +2677,8 @@ ExecHashRemoveNextSkewBucket(HashJoinTable hashtable)
 			/* Put the tuple into a temp file for later batches */
 			Assert(batchno > hashtable->curbatch);
 			ExecHashJoinSaveTuple(tuple, hashvalue,
-								  &hashtable->innerBatchFile[batchno]);
+								  &hashtable->innerBatchFile[batchno],
+								  hashtable->spillCxt);
 			pfree(hashTuple);
 			hashtable->spaceUsed -= tupleSize;
 			hashtable->spaceUsedSkew -= tupleSize;
@@ -3093,8 +3107,11 @@ ExecParallelHashJoinSetUpBatches(HashJoinTable hashtable, int nbatch)
 	pstate->nbatch = nbatch;
 	batches = dsa_get_address(hashtable->area, pstate->batches);
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/*
+	 * Use hash join spill memory context to allocate accessors and their
+	 * buffers
+	 */
+	oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = nbatch;
@@ -3196,8 +3213,8 @@ ExecParallelHashEnsureBatchAccessors(HashJoinTable hashtable)
 	 */
 	Assert(DsaPointerIsValid(pstate->batches));
 
-	/* Use hash join memory context. */
-	oldcxt = MemoryContextSwitchTo(hashtable->hashCxt);
+	/* Use hash join spill memory context to allocate accessors */
+	oldcxt = MemoryContextSwitchTo(hashtable->spillCxt);
 
 	/* Allocate this backend's accessor array. */
 	hashtable->nbatch = pstate->nbatch;
diff --git a/src/backend/executor/nodeHashjoin.c b/src/backend/executor/nodeHashjoin.c
index 93a78f6f74..e9e58e9647 100644
--- a/src/backend/executor/nodeHashjoin.c
+++ b/src/backend/executor/nodeHashjoin.c
@@ -491,7 +491,8 @@ ExecHashJoinImpl(PlanState *pstate, bool parallel)
 					Assert(parallel_state == NULL);
 					Assert(batchno > hashtable->curbatch);
 					ExecHashJoinSaveTuple(mintuple, hashvalue,
-										  &hashtable->outerBatchFile[batchno]);
+										  &hashtable->outerBatchFile[batchno],
+										  hashtable->spillCxt);
 
 					if (shouldFree)
 						heap_free_minimal_tuple(mintuple);
@@ -1313,21 +1314,30 @@ ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
  * The data recorded in the file for each tuple is its hash value,
  * then the tuple in MinimalTuple format.
  *
- * Note: it is important always to call this in the regular executor
- * context, not in a shorter-lived context; else the temp file buffers
- * will get messed up.
+ * If this is the first write to the batch file, this function first
+ * create it.  The associated BufFile buffer is allocated in the given
+ * context.  It is important to always give the HashSpillContext
+ * context.  First to avoid a shorter-lived context, else the temp file
+ * buffers will get messed up.  Second, for a better accounting of the
+ * spilling memory consumption.
+ *
  */
 void
 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-					  BufFile **fileptr)
+					  BufFile **fileptr, MemoryContext cxt)
 {
 	BufFile    *file = *fileptr;
 
 	if (file == NULL)
 	{
-		/* First write to this batch file, so open it. */
+		MemoryContext oldctx;
+
+		oldctx = MemoryContextSwitchTo(cxt);
+
 		file = BufFileCreateTemp(false);
 		*fileptr = file;
+
+		MemoryContextSwitchTo(oldctx);
 	}
 
 	BufFileWrite(file, &hashvalue, sizeof(uint32));
diff --git a/src/backend/utils/sort/sharedtuplestore.c b/src/backend/utils/sort/sharedtuplestore.c
index 0831249159..236be65f22 100644
--- a/src/backend/utils/sort/sharedtuplestore.c
+++ b/src/backend/utils/sort/sharedtuplestore.c
@@ -308,11 +308,15 @@ sts_puttuple(SharedTuplestoreAccessor *accessor, void *meta_data,
 	{
 		SharedTuplestoreParticipant *participant;
 		char		name[MAXPGPATH];
+		MemoryContext oldcxt;
 
 		/* Create one.  Only this backend will write into it. */
 		sts_filename(name, accessor, accessor->participant);
+
+		oldcxt = MemoryContextSwitchTo(accessor->context);
 		accessor->write_file =
 			BufFileCreateFileSet(&accessor->fileset->fs, name);
+		MemoryContextSwitchTo(oldcxt);
 
 		/* Set up the shared state for this backend's file. */
 		participant = &accessor->sts->participants[accessor->participant];
@@ -527,11 +531,15 @@ sts_parallel_scan_next(SharedTuplestoreAccessor *accessor, void *meta_data)
 			if (accessor->read_file == NULL)
 			{
 				char		name[MAXPGPATH];
+				MemoryContext oldcxt;
 
 				sts_filename(name, accessor, accessor->read_participant);
+
+				oldcxt = MemoryContextSwitchTo(accessor->context);
 				accessor->read_file =
 					BufFileOpenFileSet(&accessor->fileset->fs, name, O_RDONLY,
 									   false);
+				MemoryContextSwitchTo(oldcxt);
 			}
 
 			/* Seek and load the chunk header. */
diff --git a/src/include/executor/hashjoin.h b/src/include/executor/hashjoin.h
index 8ee59d2c71..ac27222d18 100644
--- a/src/include/executor/hashjoin.h
+++ b/src/include/executor/hashjoin.h
@@ -23,12 +23,12 @@
 /* ----------------------------------------------------------------
  *				hash-join hash table structures
  *
- * Each active hashjoin has a HashJoinTable control block, which is
- * palloc'd in the executor's per-query context.  All other storage needed
- * for the hashjoin is kept in private memory contexts, two for each hashjoin.
- * This makes it easy and fast to release the storage when we don't need it
- * anymore.  (Exception: data associated with the temp files lives in the
- * per-query context too, since we always call buffile.c in that context.)
+ * Each active hashjoin has a HashJoinTable structure, which is
+ * palloc'd in the executor's per-query context.  Other storages needed for
+ * each hashjoin is kept in child contexts, three for each hashjoin:
+ *   - HashTableContext (hashCxt): the parent hash table storage context
+ *   - HashSpillContext (spillCxt): storage for temp files buffers
+ *   - HashBatchContext (batchCxt): storage for a batch in serial hash join
  *
  * The hashtable contexts are made children of the per-query context, ensuring
  * that they will be discarded at end of statement even if the join is
@@ -36,9 +36,19 @@
  * be cleaned up by the virtual file manager in event of an error.)
  *
  * Storage that should live through the entire join is allocated from the
- * "hashCxt", while storage that is only wanted for the current batch is
- * allocated in the "batchCxt".  By resetting the batchCxt at the end of
- * each batch, we free all the per-batch storage reliably and without tedium.
+ * "hashCxt" (mainly hash's meta datas). Also, the "hashCxt" context is the
+ * parent of "spillCxt" and "batchCxt". It makes it easy and fast to release
+ * the storage when we don't need it anymore.
+ *
+ * Data associated with temp files lives in the "spillCxt" context which lives
+ * during the entire join as temp files might need to survives batches. These
+ * files are explicitly destroyed by calling BufFileClose() when the code is
+ * done with them. The aim of this context is to help accounting the memory
+ * allocations dedicated to temp files and their buffers.
+ *
+ * Finaly, storage that is only wanted for the current batch is allocated in
+ * the "batchCxt".  By resetting the batchCxt at the end of each batch, we free
+ * all the per-batch storage reliably and without tedium.
  *
  * During first scan of inner relation, we get its tuples from executor.
  * If nbatch > 1 then tuples that don't belong in first batch get saved
@@ -350,6 +360,7 @@ typedef struct HashJoinTableData
 
 	MemoryContext hashCxt;		/* context for whole-hash-join storage */
 	MemoryContext batchCxt;		/* context for this-batch-only storage */
+	MemoryContext spillCxt;		/* context for spilling to temp files */
 
 	/* used for dense allocation of tuples (into linked chunks) */
 	HashMemoryChunk chunks;		/* one list for the whole batch */
diff --git a/src/include/executor/nodeHashjoin.h b/src/include/executor/nodeHashjoin.h
index d367070883..fb33994366 100644
--- a/src/include/executor/nodeHashjoin.h
+++ b/src/include/executor/nodeHashjoin.h
@@ -29,6 +29,6 @@ extern void ExecHashJoinInitializeWorker(HashJoinState *state,
 										 ParallelWorkerContext *pwcxt);
 
 extern void ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
-								  BufFile **fileptr);
+								  BufFile **fileptr, MemoryContext cxt);
 
 #endif							/* NODEHASHJOIN_H */
-- 
2.40.1


--MP_/KUKoyOpWagGRmSskvJu5/tn--





^ permalink  raw  reply  [nested|flat] 9+ messages in thread


end of thread, other threads:[~2023-05-16 13:42 UTC | newest]

Thread overview: 9+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2020-06-06 22:42 [PATCH v5 2/3] Allow CREATE INDEX CONCURRENTLY on partitioned table Justin Pryzby <[email protected]>
2023-03-27 13:54 [PATCH 2/3] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-03-27 13:54 [PATCH 2/3] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-03-27 13:54 [PATCH] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-03-27 13:54 [PATCH 2/3] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-03-27 13:54 [PATCH] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-03-27 13:54 [PATCH 2/4] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-03-27 13:54 [PATCH] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>
2023-05-16 13:42 [PATCH v8 2/3] Allocate hash batches related BufFile in a dedicated context Jehan-Guillaume de Rorthais <[email protected]>

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox